S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e

SOCIAL SCIENCES Copyright © 2024 The

Authors, some rights
Addressing climate change with behavioral science: A reserved; exclusive
licensee American
global intervention tournament in 63 countries Association for the
Advancement of
Science. No claim to
Madalina Vlasceanu1*†, Kimberly C. Doell1,2*†, Joseph B. Bak-­Coleman3,4†, Boryana Todorova2, original U.S.
Michael M. Berkebile-­Weinberg1, Samantha J. Grayson5‡, Yash Patel1, Danielle Goldwert1, Government Works.
Yifei Pei1, Alek Chakroff6, Ekaterina Pronizius2, Karlijn L. van den Broek7, Denisa Vlasceanu8, Distributed under a
Sara Constantino9,10, Michael J. Morais11, Philipp Schumann12, Steve Rathje1, Ke Fang1, Creative Commons
Salvatore Maria Aglioti13,14, Mark Alfano15, Andy J. Alvarado-­Yepez16, Angélica Andersen17, Attribution License 4.0
(CC BY).
Frederik Anseel18, Matthew A. J. Apps19, Chillar Asadli20, Fonda Jane Awuor21, Flavio Azevedo22,
Piero Basaglia23, Jocelyn J. Bélanger24, Sebastian Berger25, Paul Bertin26,27, Michał Białek28,
Olga Bialobrzeska29, Michelle Blaya-­Burgo30, Daniëlle N. M. Bleize31, Simen Bø32, Lea Boecker33,
Paulo S. Boggio34, Sylvie Borau35, Björn Bos36, Ayoub Bouguettaya37, Markus Brauer38,
Cameron Brick39,40, Tymofii Brik41, Roman Briker42, Tobias Brosch43, Ondrej Buchel44,
Daniel Buonauro45, Radhika Butalia46, Héctor Carvacho47, Sarah A. E. Chamberlain48,
Hang-­Yee Chan49, Dawn Chow50, Dongil Chung51, Luca Cian52, Noa Cohen-­Eick53,54,
Luis Sebastian Contreras-­Huerta19,55, Davide Contu56, Vladimir Cristea57, Jo Cutler19,
Silvana D'Ottone58, Jonas De Keersmaecker59,60, Sarah Delcourt61, Sylvain Delouvée62,
Kathi Diel63, Benjamin D. Douglas38, Moritz A. Drupp23,64, Shreya Dubey65, Jānis Ekmanis66,
Christian T. Elbaek67, Mahmoud Elsherif68,69, Iris M. Engelhard70, Yannik A. Escher71,
Tom W. Etienne57,72, Laura Farage73, Ana Rita Farias74, Stefan Feuerriegel75, Andrej Findor76,
Lucia Freira77, Malte Friese63, Neil Philip Gains78, Albina Gallyamova79, Sandra J. Geiger80,
Oliver Genschow81, Biljana Gjoneska82, Theofilos Gkinopoulos83, Beth Goldberg84,
Amit Goldenberg85,86,87, Sarah Gradidge88, Simone Grassini89,90, Kurt Gray91, Sonja Grelle92,
Siobhán M. Griffin93, Lusine Grigoryan94, Ani Grigoryan95, Dmitry Grigoryev79, June Gruber96,
Johnrev Guilaran97, Britt Hadar98, Ulf J.J. Hahnel99, Eran Halperin53, Annelie J. Harvey88,
Christian A. P. Haugestad100, Aleksandra M. Herman101,102, Hal E. Hershfield103,
Toshiyuki Himichi104, Donald W. Hine105, Wilhelm Hofmann92, Lauren Howe106,
Enma T. Huaman-­Chulluncuy107, Guanxiong Huang108, Tatsunori Ishii109, Ayahito Ito110,
Fanli Jia111, John T. Jost1, Veljko Jovanović112, Dominika Jurgiel113, Ondřej Kácha114,
Reeta Kankaanpää115,116, Jaroslaw Kantorowicz117, Elena Kantorowicz-­Reznichenko118,
Keren Kaplan Mintz119,120, Ilker Kaya121, Ozgur Kaya121, Narine Khachatryan95, Anna Klas122,
Colin Klein123, Christian A. Klöckner124, Lina Koppel125, Alexandra I. Kosachenko126,
Emily J. Kothe122, Ruth Krebs127, Amy R. Krosch128, Andre P.M. Krouwel129, Yara Kyrychenko130,
Maria Lagomarsino131, Claus Lamm2, Florian Lange61, Julia Lee Cunningham132, Jeffrey Lees133,134,
Tak Yan Leung135, Neil Levy136, Patricia L. Lockwood19, Chiara Longoni137,
Alberto López Ortega138, David D. Loschelder139, Jackson G. Lu140, Yu Luo141, Joseph Luomba142,
Annika E. Lutz143, Johann M. Majer144, Ezra Markowitz145, Abigail A. Marsh146,
Karen Louise Mascarenhas147,148, Bwambale Mbilingi149, Winfred Mbungu150, Cillian McHugh151,
Marijn H.C. Meijers152, Hugo Mercier153, Fenant Laurent Mhagama154, Katerina Michalakis155,
Nace Mikus156,157, Sarah Milliron128, Panagiotis Mitkidis67, Fredy S. Monge-­Rodríguez158,
Youri L. Mora159,160, David Moreau161, Kosuke Motoki162, Manuel Moyano163, Mathilde Mus164,
Joaquin Navajas165,166, Tam Luong Nguyen167, Dung Minh Nguyen168, Trieu Nguyen168,
Laura Niemi169, Sari R. R. Nijssen170, Gustav Nilsonne171,172, Jonas P. Nitschke2, Laila Nockur173,
Ritah Okura149, Sezin Öner174, Asil Ali Özdoğru175,176, Helena Palumbo177,
Costas Panagopoulos178, Maria Serena Panasiti14,179, Philip Pärnamets180,
Mariola Paruzel-­Czachura181,182, Yuri G. Pavlov183, César Payán-­Gómez184, Adam R. Pearson45,
Leonor Pereira da Costa185, Hannes M. Petrowsky186, Stefan Pfattheicher173, Nhat Tan Pham187,
Vladimir Ponizovskiy92, Clara Pretus188, Gabriel G. Rêgo189, Ritsaart Reimann136,
Shawn A. Rhoads190,191, Julian Riano-­Moreno192, Isabell Richter193, Jan Philipp Röer194,
Jahred Rosa-­Sullivan195, Robert M. Ross136, Anandita Sabherwal196, Toshiki Saito197,198,
Oriane Sarrasin199, Nicolas Say200, Katharina Schmid201, Michael T. Schmitt143,

Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024 1 of 19

S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e

Philipp Schoenegger202,203, Christin Scholz204, Mariah G. Schug205, Stefan Schulreich206,207,

Ganga Shreedhar196, Eric Shuman1,208, Smadar Sivan209, Hallgeir Sjåstad32, Meikel Soliman210,
Katia Soud211,212, Tobia Spampatti213,214, Gregg Sparkman215, Ognen Spasovski216,217,
Samantha K. Stanley218, Jessica A. Stern219, Noel Strahm25, Yasushi Suko220, Sunhae Sul221,
Stylianos Syropoulos222, Neil C. Taylor223, Elisa Tedaldi224, Gustav Tinghög125,
Luu Duc Toan Huynh225, Giovanni Antonio Travaglino226, Manos Tsakiris155, İlayda Tüter227,
Michael Tyrala228, Özden Melis Uluğ229, Arkadiusz Urbanek230, Danila Valko231,232,
Sander van der Linden233, Kevin van Schie234, Aart van Stekelenburg235, Edmunds Vanags236,
Daniel Västfjäll237, Stepan Vesely124, Jáchym Vintr114, Marek Vranka238, Patrick Otuo Wanguche239,
Robb Willer240, Adrian Dominik Wojcik241, Rachel Xu84, Anjali Yadav242,243, Magdalena Zawisza88,
Xian Zhao244, Jiaying Zhao141,245, Dawid Żuk246, Jay J. Van Bavel1,32,247

Effectively reducing climate change requires marked, global behavior change. However, it is unclear which strate-
gies are most likely to motivate people to change their climate beliefs and behaviors. Here, we tested 11 expert-­
crowdsourced interventions on four climate mitigation outcomes: beliefs, policy support, information sharing
intention, and an effortful tree-­planting behavioral task. Across 59,440 participants from 63 countries, the inter-
ventions’ effectiveness was small, largely limited to nonclimate skeptics, and differed across outcomes: Beliefs
were strengthened mostly by decreasing psychological distance (by 2.3%), policy support by writing a letter to a
future-­generation member (2.6%), information sharing by negative emotion induction (12.1%), and no interven-
tion increased the more effortful behavior—several interventions even reduced tree planting. Last, the effects of
each intervention differed depending on people’s initial climate beliefs. These findings suggest that the impact of
behavioral climate interventions varies across audiences and target behaviors.

INTRODUCTION While many of these theories, as well as their corresponding in-

The climate crisis is one of humanity’s most consequential and chal-
lenging problems (1). Successfully rising to the challenge depends
on both “top-­down” structural changes (e.g., regulation and invest-
ment) and “bottom-­up” changes (e.g., individuals’ and collectives’
beliefs and behaviors). These bottom-­up processes require wide-
spread belief in climate change, support for climate change policy,
and willingness to engage in climate action (2–4). The behavioral
sciences have been seen as a crucial component in promoting
bottom-­up change, through the development of large-­scale inter-
ventions that can shift public opinion and enable and support top-­
down governmental climate policies (5–7). However, it is unclear
which strategies are most likely to motivate people to change their
climate change beliefs and climate mitigation behaviors. Here, we
assess the effectiveness of expert-­crowdsourced, theoretically de-
rived interventions at promoting a range of climate change mitiga-
tion behaviors in a large and diverse global sample.
A growing body of research across the behavioral sciences has
been investigating intervention strategies aimed at boosting sus-
tainable intentions and behaviors such as recycling, public trans-
portation use, and household energy saving (3, 8, 9). For instance,
communications aimed at reducing the psychological distance of
climate change, by making it feel more geographically, socially, and
temporally close, were effective at increasing climate concern and
amplifying self-­reported intentions to engage in mitigating behav-
iors, such as reducing energy consumption (10). Similarly, norma-
tive appeals that include an invitation to work together and “join
in” were found effective at influencing behaviors such as charitable
giving (11). These are only two examples in a growing list of behav-
ioral interventions designed to mitigate climate change. Hence,
there are numerous competing theories in the behavioral sciences
about how to stimulate climate change beliefs and proenviron-
mental behaviors.
See last page for author affiliations.
terventions, are promising, they have been tested independently
with different samples and, on separate outcomes, making it impos-
sible to directly compare their effectiveness. In addition, assessing
interventions on a single outcome renders it difficult to understand
their effects on multiple facets of climate mitigation, which are all
necessary to substantially reduce climate change (e.g., support for
climate mitigation policy and sustainable behavior). These limita-
tions are a major barrier to resolving theoretical debates within the
scientific community (12, 13) and to translating scientific findings
into impactful policies (14, 15). Moreover, traditional attempts to
compare interventions (e.g., meta-­analyses) (16) are limited by dif-
ferences in experimental protocols, outcome variables, samples, and
operationalizations (17, 18, 19). These differences hinder evaluations
of the relative effectiveness of different theories and interventions
(15). To address these concerns, we used the megastudy approach—
an experimental paradigm similar to a randomized controlled trial
but designed to evaluate the efficacy of many interventions on sev-
eral outcome variables, in the same large-­scale experiment (18). This
provides a rigorous direct comparison of competing approaches to
climate change mitigation.
Another challenge is that most prior work across the behavioral
sciences (including the megastudy approach) has been mainly con-
ducted on Western, educated samples from industrialized, rich, and
developed countries (i.e., WEIRD) (20). Results from these samples
may not generalize to other nations, restricting the ability to apply
findings beyond WEIRD populations. This is a particular problem
for a topic like climate change where the social and political dynam-
ics, as well as exposure to the impacts of climate change, vary across
countries (21, 22). While wealthier nations are disproportionately
responsible for causing climate change (23), it is still important to
understand which interventions work across a diversity of cultures
since the most effective mitigation strategies will likely require
global cooperation. Accordingly, we leveraged the many labs approach,
in which the same study is being conducted by many research

Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024 2 of 19

S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
laboratories around the world, aggregating the results in the same
international dataset (17, 24).
In this global megastudy, we crowdsourced interventions previ-
ously found to stimulate climate mitigation from behavioral science
experts (fig. S5). We used a crowdsourcing approach to determine
which interventions to test, given recent evidence that crowdsourc-
ing can improve the quality of scientific investigations by promoting
ideation, inclusiveness, transparency, rigor, and reliability (25). This
resulted in the identification of 11 behavioral interventions based on
competing theoretical frameworks in the behavioral sciences (Fig. 1).
We tested these interventions in a global tournament spanning
63 countries on four outcome variables, which were also crowdsourced
and selected on the basis of their theoretical and practical relevance
to climate mitigation. The first outcome on which we assessed each
intervention was belief in climate change (four items; e.g., “Climate
change poses a serious threat to humanity”). Given that belief is a key
antecedent of proenvironmental intentions, behavior, and policy
support (26), we examined how the interventions would affect these
outcomes for different people along the belief continuum ranging
from skeptics to true believers.
The second outcome was support for climate change mitigation
policy (nine items; e.g., “I support raising carbon taxes on gas/fossil
fuels/coal”). Given that successful climate mitigation requires large-­
scale policy reform (1) and the public’s support for climate policies
is the top predictor of policy adoption (27), this outcome variable
reflects the importance of impactful systemic change, rather than
private mitigation efforts based on individual decision-­ making
(28–30). Recent work argues that individual-­level behaviors should
be targeted alongside structural changes (31), especially since fram-
ing climate change as an individual level problem can backfire,
leading to feelings of helplessness and concerns about free riding
(32, 33).
To target more ecologically valid behavior and climate activism
(34), the third outcome was willingness to share climate mitigation
information on social media (i.e., “Did you know that removing
meat and dairy for only two out of three meals per day could de-
crease food-­related carbon emissions by 60%?”). While this behav-
ior is relatively low effort, recent work suggests climate information
sharing with one’s community as an essential step in addressing the
climate crisis (35).
Last, given the large gap between self-­reported measures and ob-
jective proenvironmental behavior (36), the fourth outcome we tar-
geted was a more effortful behavior of contributing to a real tree
planting initiative by engaging in a cognitively demanding task (i.e.,
a modified version of the work for environmental protection task or
WEPT) (37). The WEPT is a multitrial, web-­based procedure in
which participants choose to exert voluntary effort screening stimu-
li for specific numerical combinations (i.e., an even first digit and
odd second digit) in exchange for donations to a tree-­planting envi-
ronmental organization. Thus, they had the opportunity to produce
actual environmental benefits at actual behavioral costs, mimicking
classic sustainable behavior trade-­offs (38–40).
Participants (N = 59,440, from 63 counties; Table 1) were most-
ly recruited through online data collection platforms (80.8%) or via
convenience/snowball sampling (19.1%; Table 1). They were ran-
domly assigned to 1 of 11 experimental interventions (Fig. 1) or a
no intervention control condition in which they read a passage
from a literary text. Then, in a randomized order, participants indi-
cated their climate beliefs, climate policy support, and willingness
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
to share climate-­related information on social media. Last, partici-
pants were able to opt into completing up to eight pages of a tree-­
planting task, each completed page resulting in the real planting of
a tree through a donation to The Eden Reforestation Project. As a
result of participants’ behavior, our team actually planted 333,333
trees. Assuming that the average fully grown tree absorbs between
10 and 40 kg of carbon dioxide per year in 5 to 10 years when all
trees are fully grown, the efforts from this project will result in
~9,999,990 kg of carbon dioxide sequestered per year, which is the
equivalent amount of carbon dioxide used to produce energy for
1260 U.S. homes.
RESULTS
Main effects of intervention
First, we examined the effect of each intervention on each of the four
outcomes, estimated using a series of Bayesian regressions (see Ma-
terials and Methods). As the goal of this study is to estimate the rela-
tive effectiveness of treatments, in contrast to establishing non-­null
effects or differences, Bayesian estimation is preferable to classical
null hypothesis significance testing. Bayesian techniques produce
posterior distributions for parameters (here, treatment effects) that
characterize their magnitude and associated uncertainty. We sum-
marize this distribution in Fig. 2 using a point estimate correspond-
ing to the mean and a 94% credible region, which differs from a
confidence interval in which it indicates a region with a 94% chance
of containing the unobserved parameter value (41). Moreover, we
also conducted similar frequentist analyses (hierarchical mixed
models) and found converging results (see the Supplementary Ma-
terials for details).
We began by assessing the main intervention effects on each out-
come. For belief in climate change (measured on a scale from 0 to
100), the top performing intervention, decreasing psychological dis-
tance, increased beliefs by an absolute effect size of 2.3% (1.6 to 2.9)
(94% credible region) compared to the control condition. Consis-
tent with prior work (10), some interventions slightly increased be-
liefs. However, other interventions had near-­zero effect, suggesting
that findings of some prior research did not extend to this context
(Fig. 2A) (11).
For climate policy support (measured on a scale from 0 to 100),
the intervention with the largest average effect was writing a letter to
a member of the future generation, which increased policy support
by 2.6% (2.0 to 3.2). Similar to belief, all interventions produced ei-
ther more policy support or no discernible differences from the con-
trol condition (Fig. 2B).
For willingness to share climate change information on social
media (measured as a binary choice), all interventions generally in-
creased intentions to share. The largest gains were exhibited in the
negative emotion induction condition, which led to 12.1% (9.8 to
14.6) more sharing compared to the control condition (Fig. 2C).
For the number of pages completed on the WEPT tree-­planting
task (from 0 to 8), no intervention was better than the control
condition, and some interventions (i.e., decreasing psychological
distance, inducing negative emotions, work-­together normative
appeals, and writing a letter to a future-­generation member)
appeared to reduce tree planting (Fig. 2D). These results held re-
gardless of the operationalization of a tree planted as participants’
confirmation that they wanted to complete another WEPT page
or their accuracy in the task (table S24).
3 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
The interventions that produced negative effects on the WEPT
were also those that took the most time to complete (see the Supple-
mentary Materials). Assuming that participants have a limited bud-
get of time for completing surveys and given that the tree-­planting
task requires time, it is expected that we observed a trade-­off be-
tween the time spent on the intervention and on the outcome task.
Therefore, in an exploratory analysis (tables S22 and S23), we as-
sessed the effects of the interventions when adjusting for the time
spent on each intervention. While we still observed the negative ef-
fects of some interventions on tree planting, we now also observed
positive effects of five interventions. That is, when controlling for
intervention length, binding moral foundations, scientific consen-
sus, dynamic norms, pluralistic ignorance, and system justification
all increased the number of trees planted compared to the control
condition. Thus, in the absence of time constraints, these interven-
tions might increase proenvironmental behavior. However, the de-
gree to which these findings actually generalize to proenvironmental
behaviors that do not hinge on time (e.g., donations) should be as-
sessed in future studies.
For further assessing the average effects of each intervention on
each outcome within any subsample of interest varying along demo-
graphics such as nationality, political ideology, age, gender, education,
or income level, we provide an easy to use and disseminate web tool:
https://climate-­interventions.shinyapps.io/climate-­interventions/.
Heterogeneous intervention effects along initial
belief continuum
We found a high level of belief in climate change [i.e., 85.7% (85.2 to
86.2), an estimate computed using the ratings of belief in the control
participants and estimated preintervention levels of belief from all
other participants]. This could raise two potential concerns when
evaluating the main effects of the interventions mentioned above:
On the one hand, at this high level of belief, participants may be
particularly receptive to interventions. As a result, average effects
may tend to overestimate the effectiveness of interventions in ap-
plied contexts where the aim is to increase belief or policy support in
skeptical participants that do not already believe in climate change.
On the other hand, as our outcomes are bounded, these high levels
of belief may lead to ceiling effects in the estimation of the average
effects, which may undervalue the true effectiveness of the interven-
tions. To address this concern, we conducted an additional analysis
where we modeled heterogeneous effects as a function of unob-
served preintervention belief (see Materials and Methods and the
Supplementary Materials). This analysis allowed us to visualize how
effective interventions were across the continuum from climate
change skeptics (i.e., those with initial beliefs less than 35%) to true
believers (i.e., those with initial beliefs higher than 65%; Fig. 3).
For the impact of interventions on belief (Fig. 3A), we found clear
indications of ceiling effects with many interventions being maxi-
mally impactful among uncertain participants, even those with low
to moderate levels of initial belief. Even in participants with low lev-
els of preexisting climate change belief (i.e., less than 35%), interven-
tions such as reducing psychological distance, future self-­continuity,
and effective collective action are all viable ways to increase belief in
climate change.
For policy support, a different pattern emerged. Interventions
such as writing a letter to a member of the future generation, collec-
tive action efficacy, future self-­continuity, and decreasing psycho-
logical distance all increased support for climate policy (Fig. 3B).
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
Those same interventions appear to function well on individuals
with modest to high levels of initial climate change belief (i.e., at ~35
to 90%; Fig. 3B). However, they were relatively ineffectual among
those that were low in initial belief (i.e., climate skeptics). The main
exception is in writing a letter to a member of the future-­generation
intervention, which worked across nearly the entire spectrum of ini-
tial belief. In addition, for those that were very low to moderate (i.e.,
0 to 65%) on initial belief, the negative emotion intervention ap-
peared to backfire, reducing support for climate change policies.
Similar to belief, the work-­together normative appeal also slightly
backfired in participants with moderate levels of initial belief.
Regarding social media sharing, nearly all interventions (i.e., 9 of 11)
increased willingness to share even at moderate levels of initial belief
(i.e., those greater than ~35 to 60%). Moreover, the increase in willing-
ness to share by inducing negative emotions extended into individuals
who generally do not believe in climate change. Last, the work-­together
normative appeal intervention backfired among those who are very low
on initial belief (i.e., ~0 to 15%), reducing their willingness to share in-
formation on social media by up to 12%. Last, for the tree-­planting task,
more than half of the interventions decreased the number of pages
completed on the WEPT across all levels of initial belief (Fig. 3D).
Country-­level main effects
Last, we examined the country-­level main effects for each of our key
outcome variables. We found that average belief in climate change,
across all countries surveyed, was high [85.7% (85.2 to 86.2); this
includes both ratings of belief in the control participants and esti-
mated preintervention levels of belief from all other participants].
There was a very little variation between countries (Fig. 4A, fig. S4A,
and table S5) indicating a clear majority belief in climate change.
Similar patterns were observed for policy support (Fig. 4B), with all
countries indicating clear majority support for a variety of climate
change policies [72.2% (71.6 to 72.8)]. These results suggest that
there is clear and consistent global consensus regarding the dangers
posed by climate change and the importance of enacting climate
change mitigation.
Other outcome variables exhibited larger variation across coun-
tries. Willingness to share climate change–related information on
social media was more modest [56.9 (56.4 to 57.5)] and variable,
ranging from low in Latvia of 17.6% (14.3 to 21.4) to high of 93.3%
(90.4 to 95.7) in Kenya (Fig. 4C). These results suggest that observa-
tions of climate change discussion online may not accurately reflect
global sentiments about the reality of climate change but rather dif-
ferent local norms. Last, half of all participants (50.7% of total sam-
ple and 53.1% of control condition sample) completed all eight
pages of the WEPT, earning the maximum number of trees possible,
with an overall average of 5.2 (5.1 to 5.3) pages completed (Fig. 4D).
DISCUSSION
In a global megastudy conducted on a sample of 59,440 people
from 63 countries, we empirically assessed the relative effectiveness
of 11 expert-­crowdsourced, theoretically derived behavioral inter-
ventions at stimulating climate mitigation beliefs and behaviors
(i.e., climate change beliefs, policy support, willingness to share in-
formation, and tree-­planting contributions). We found that differ-
ent interventions tended to have small global effects, which varied
across outcomes and largely affected nonskeptics, emphasizing the
importance of examining the impact of climate interventions on a
4 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
Fig. 1. Interventions, theoretical frameworks, and brief descriptions.
range of outcomes before drawing conclusions regarding their
overarching relative efficacy. These findings suggest that the impact
of behavioral climate interventions varies across audiences’ charac-
teristics and target behaviors.
Here, climate change beliefs were strengthened most by decreas-
ing the psychological distance of climate change. Support for cli-
mate change mitigation policy was increased mostly by writing a
letter to be read in the future by a socially close child, describing
one’s current climate change mitigation actions. Willingness to share
climate change information on social media was increased most by
inducing negative emotions through “doom and gloom”–styled
messaging about the consequences of climate change. Last, while
half of the tested interventions had no effect on the effortful tree-­
planting behavior, the other half of the interventions reduced the
number of trees participants planted. Beyond revealing the utility of
harnessing a multioutcome approach, these results also highlight
the need for tailoring interventions to target outcomes.
Our findings extend prior work and are theoretically informative
in several ways. Notably, these findings help reconcile several theo-
retical debates in the literature. For example, some have argued in
favor of using a doom and gloom messaging style in climate com-
munications (i.e., induce negative emotions) as a way to stimulate
climate mitigation behaviors (42). For instance, recent work found
that online news consumption is largely driven by the negative con-
tent of the news (43). However, others have warned that this mes-
saging may have no impact on behavior (44) or, worse, that it may
depress and demoralize the public into inaction (45). Here, we
found empirical support for both accounts on different outcomes:
While negative emotion messaging was highly effective at
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
stimulating climate information–sharing intentions (a relatively
low-­effort behavior), it decreased tree-­planting efforts. Further, the
negative emotion induction intervention appeared to backfire on
policy support among participants with low initial climate beliefs.
These results suggest that climate scientists should carefully consid-
er the differential effects of the prevalent fear-­inducing writing styles
on different proclimate outcomes. Moreover, it suggests that theo-
retical models need to explain divergent patterns across outcomes.
The results also indicate that the impact of the interventions on
each outcome depends on peoples’ preexisting belief in climate
change, supporting the claim that interventions need to be tailored
to the characteristics of their audience (44, 45). For belief, the effec-
tiveness of several interventions (e.g., decreasing the psychological
distance and collective action efficacy) was maximized among the
uncertain, with lesser effects among believers and skeptics. For pol-
icy support, however, interventions were generally only effective
among those with high initial levels of belief, with negative emo-
tions backfiring among skeptics. Similarly, the robust increases in
willingness to share on social media were largely restricted to people
who already believed in climate change—with negative emotions
increasing sharing intentions even among skeptics. For the higher
effort behavior, however, interventions appeared to uniformly re-
duce tree planting across all levels of initial belief.
Given the heterogeneity of these results across outcomes, we cre-
ated a web tool resource (https://climate-­interventions.shinyapps.io/
climate-­interventions/) that can easily and rapidly assess intervention
efficacy across each of the four outcomes and across a range of vari-
ables, including country, political ideology, gender, age, socioeco-
nomic status, income, and education. While we caution that users
5 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e

Table 1. Variables on which the samples in each country were matched to the population. Countries in which no demographic variable was census
matched are marked as “N/A” in the “matched variables” column. SES, socioeconomic status.
Sample Matched variables N Sample Matched variables N

Algeria N/A 528 Philippines N/A 145

Armenia N/A 492 Poland_1 Age, gender, education 1883
Australia Gender 979 Poland_2 N/A 463
Austria Age, gender 502 Portugal N/A 499
Belgium_1 Age, gender 522 Romania N/A 411
Belgium_2 Age, gender 512 Russia_1 N/A 718
Brazil Age, gender, education 1261 Russia_2 Region, ethnicity 395
Bulgaria Age, gender 778 Russia_3 N/A 322
Canada_1 N/A 858 Saudi Arabia N/A 489
Canada_2 Age, gender 303 Serbia N/A 337
Chile Age, gender, region, 1992 Singapore N/A 500
SES
China N/A 896 Slovakia Age, gender, region, 1027
municipality size
Czechia N/A 547 Slovenia Age, gender 501
Denmark Age, gender, region 792 South Africa Age, gender 496
Ecuador Age, gender, region 679 South Korea Age, gender 639
Finland Age, gender 625 Spain_1 N/A 110
France Age, gender 1480 Spain_2 Age, gender, region 434
Gambia N/A 527 Sri Lanka N/A 413
Germany Age, gender, region 1545 Sudan Age, gender 623
Ghana Age, gender 522 Sweden Age, gender 2393
Greece Age, gender 597 Switzerland_1 Age, gender 512
India N/A 688 Switzerland_2 Age, gender 531
Ireland N/A 753 Taiwan N/A 206
Israel Age, gender, region, 1384 Tanzania Age, gender 104
ethnicity
Italy_1 Age, gender, region 591 Thailand N/A 586
Italy_2 Gender 993 Turkey_1 N/A 359
Japan_1 N/A 653 Turkey_2 Age, gender 347
Japan_2 Income, education, 802 Uganda Age, gender 476
region, ethnicity
Kenya Age, gender 409 UK_1 N/A 220
Latvia Income, education, 485 UK_2 Age, gender 952
ethnicity
Mexico Age, gender 490 UK_3 N/A 234
Morocco Age, gender 474 UK_4 gender 501
Netherlands_1 Age, gender 854 Ukraine N/A 496
Netherlands_2 Age, gender 510 UAE Broadly representative 554
Netherlands_3 N/A 500 Uruguay N/A 838
New Zealand Gender 1005 USA_1 Age, gender 2360
Nigeria Age, gender 1513 USA_2 Age, gender, region, 5055
ethnicity
North Macedonia N/A 878 USA_3 Age, gender 497
Norway Age, gender, ethnicity 997 Venezuela N/A 110
Peru Age, gender 405 Vietnam N/A 383

Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024 6 of 19

S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
must take into account the sample sizes when exploring subsamples
of the data and the fact that they are looking at percentage of change
compared to the control condition, this web tool can be used as a
rapid and intuitive way to query intervention efficacy within subsam-
ples of interest. For example, for highly educated conservatives in the
United States, the top intervention to increase climate policy support
was the future self-­continuity intervention, increasing support by
18%. This intervention also increased climate beliefs in Russian par-
ticipants by 9%. The scientific consensus intervention increased cli-
mate policy support by 9% in Romania but decreased it by 5% in
Canada. The binding moral foundations intervention increased the
number of trees planted by Australians under the age of 40 by 40%,
and by Gambians by 35%, but this intervention decreased the number
of trees planted by wealthy Japanese participants by 24%. These results
can inform the development of local intervention strategies, which
should then be empirically validated. Critically, these results also bol-
ster the message that interventions need to be tailored to the charac-
teristics of the target audience, nationality being an important factor.
The accompanying data exploration web tool and the open-­source
raw dataset contribute to the data-­as-­public-­good trend emerging in
the spirit of open science, thus facilitating the testing of additional
hypotheses and advancement of science.
In a linked forecasting experiment (46), academics (e.g., behav-
ioral scientists) and the general public were asked to predict how
each intervention would affect belief, policy support, and the tree-­
planting behavior in a subset of participants from this study (i.e.,
those from the United States). While academics were better than the
general public at predicting the efficacy of these interventions on
beliefs and policy support, when compared to statistical models us-
ing simple heuristics such as “interventions would have no effect,”
no group was able to accurately predict how interventions would
affect behavior. These results suggest that our findings here reflect an
important departure from the expectations within the academic
community.
There are also several limitations and future directions that
should be emphasized. First, the sampling procedures differed be-
tween countries (e.g., the U.S. and Israel samples matched the census
on age, gender, region, ethnicity; and the Norway sample matched
on age, gender, ethnicity; etc.; Table 1). It should be noted that 73.6%
of the entire sample was matched for at least one variable. However,
despite these differences, recent work has found that representative
samples are not required to obtain generalizable estimates of effect
sizes within countries (47, 48). Various analyses have highlighted
A Belief
Psychological distance
Collective action
Future self-continuity
Letter to future gen.
System justification
Scientific consensus
Binding moral found.
Dynamic soc. norm
Negative emotions
Pluralistic ignorance
Work- t ogether norm
82 83 84 85 86 87 88 89 90
Belief in climate change (%)
Fig. 2. Average effects (i.e., posterior estimates using Bayesian regressions) by intervention for each outcome. Dots indicate the mean, with error bars indicating
the 94% credible region. Thicker error bars indicate the interquartile range. Vertical lines indicate control average. (A) Belief, (B) support for policy, (C) willingness to share
climate change information on social media, and (D) number of trees planted in the WEPT. Estimates are reported in tables S1 to S4.
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
that convenience samples are adequate for estimating treatment ef-
fects (49, 50). Hence, given that our paper is primarily concerned
with the effects of these interventions rather than with estimating
levels of opinion within each country, our sampling procedures were
appropriate for the analyses and conclusions drawn here. However,
while realizing that it will be a challenge, we encourage future work
to examine these processes using larger, more representative samples
from an even broader sample of countries.
Second, we leveraged an online survey–based approach, which
means that we were able to capture a limited set of contextual factors
that may have influenced our results. This approach was the most
effective way to measure and compare intervention efficacy in such
a diverse global sample. However, one important and potentially
impactful avenue for future research could be to leverage these find-
ings to conduct local field experimentation in targeted samples.
One of the major strengths of our tournament was testing 11 dif-
ferent interventions simultaneously in a large global sample across
multiple outcomes. Given the heterogeneity in the effectiveness of the
interventions across the outcomes, future work should likewise pri-
oritize testing promising interventions on even more climate-­relevant
antecedents and outcomes for a more comprehensive assessment of
climate interventions and their underlying theoretical frameworks.
One constraint we faced when attempting to test additional theories
was the decision to not use deception in our interventions. For ex-
ample, descriptive or injunctive norm–based interventions would
have needed to be based on deception to be included in and deployed
at this global scale, given the unavailability of the empirical informa-
tion critical to creating these interventions. We hope that the current
dataset can provide this information for future research in interna-
tional contexts. Future work should also investigate additional
proenvironmental behaviors, such as investment decisions, activism,
advocacy, or civic participation, critical to climate change mitigation.
Future research should also assess the processes behind the nega-
tive effects we observed on the tree-­planting task. Here, we find evi-
dence for a trade-­off between time spent on the intervention and in
the behavioral task, but additional processes may also be at play. For
instance, the negative effects observed might suggest a negative
spillover process, by which increasing some mitigation actions (e.g.,
policy support, social media sharing, etc.) could have decreased
other mitigation actions (e.g., contributing to tree planting). Given
that the tree-­planting task was also the last outcome variable com-
pleted by participants, such a process could be plausible. However,
each of the first three outcomes (i.e., climate belief, climate policy
B Policy C Share D Action
91 92 67 68 69 70 71 72 73 74 75 76 77 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66 68 4.0 4.4 4.8 5.2 5.6 6.0
Climate policy support (%) Willingness to share climate info (%) Number of trees planted
7 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e

support, and information-­sharing willingness) was positively asso-
ciated with the last outcome (i.e., WEPT; fig. S2 and tables S13 to
S15). These positive associations at the study level also held within
each of the 12 conditions (tables S16 to S18). That is, the more a
participant supported climate policy, the more trees they planted, a
pattern found under each condition (table S17). Similarly, partici-
pants who were willing to share climate information on social media
also planted more trees, again a pattern found under each condition
(table S18). These positive associations are more consistent with a
positive spillover.
An alternative explanation for the intervention effects on the
tree-­planting task could be that current behavioral science theories
and their corresponding interventions are more effective at target-
ing conceptual processes compared to more effortful and time-­
consuming behavioral signatures, especially in such a heterogeneous
global sample. However, another explanation could be that inter-
ventions that made the negative consequences of climate change
more salient (e.g., negative emotions, decreasing in psychological
distance, and future self-­continuity), triggered the perception that
individual-­level solutions (e.g., planting trees) may be futile in the
face of such an insurmountable phenomenon, in line with the
learned helplessness hypothesis (45). On the other hand, perhaps, a
combination of these explanations gave rise to the effects observed.
Future research is needed to clarify these processes and identify in-
terventions that increase more effortful climate actions around the
world, as well as actions that are more effective solutions to the cli-
mate crisis (30).
Last, while, in this global study, we tested the effects of several
theoretically derived behavioral interventions on people’s beliefs and
actions in the context of climate change, our findings provide mean-
ingful insights to the broader fields of social and behavioral sciences.
For instance, the average global effects of the interventions tested
ranged from effectively zero to very small in the conceptual out-
comes (beliefs and policy support) and near zero to negative in the
behavioral outcome (tree planting). These findings point to critical
limitations in these theories’ utility and generalizability beyond the
contexts in which they were developed. The most extreme example is
the correcting pluralistic ignorance intervention, which had no

A B
15 Change in policy support (%) 15

10 10
Change in belief (%)

5 5

0 0

5 5
Intervention
Intervention Collective action Letter future gen.
10 System justif. Psych distance 10 Dyn. soc. norm Psych distance
Letter future gen. Work-together norm System justif. Future self-cont.
Future self-cont. Collective action Work- t ogether norm Neg. emotions
15 15
0 20 40 60 80 100 0 20 40 60 80 100
Initial belief (%) Initial belief (%)

C D
15 1.00
Intervention
0.75 Collective action Letter future gen.
10
Change in trees planted

Plural ignorance Psych distance

Change in sharing (%)

0.50 Future self-cont. Work- t ogether norm

Neg. emotions
5
0.25

0 0.00

Intervention 0.25
5
Collective action Letter future gen.
Sci. consensus Dyn. soc. norm 0.50
10 Psych distance System justif.
Future self-cont. Work- t ogether norm 0.75
Binding moral Neg. emotions
15 1.00
0 20 40 60 80 100 0 20 40 60 80 100
Initial belief (%) Initial belief (%)

Fig. 3. Marginal effects (as the difference between interventions and control) as a function of estimated preintervention belief in climate change. Lines indicate
the average effect size, with shaded regions indicating the 94% credible region. For visual clarity, regions in which the 94% credible region overlap zero are omitted from
the figure. Where interventions have positive and negative effects that meet these criteria, a dashed line is used to connect these regions. The dashed vertical line indi-
cates average belief, where effects in Fig. 1 are estimated. (A) Climate change belief, (B) policy support, (C) sharing information on social media, and (D) trees planted via
the WEPT.

Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024 8 of 19

S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
effect on beliefs, policy support, or willingness to share information
on social media and even reduced tree-­planting efforts. Theories are
often tested and evaluated mainly on their ability to account for de-
contextualized patterns of data in laboratory settings, rather than
their ability to help solve societal problems (51). In response to this
limitation, researchers have recently proposed reverting the scien-
tific paradigm to an impact-­oriented theoretical and empirical re-
search agenda (30).
The small effect sizes we observed in this global sample might
also be partly interpreted through the lens of recent work reporting
that over 60% of studies in the most prestigious journals in psychol-
ogy have only focused on 11% of the world’s population (52). In our
data collected in the United States or other WEIRD nations, the
effects of the top interventions on belief and policy support were
much stronger than at the global level. The skewed representation
in the field may pose another notable obstacle in addressing societal
problems that depend on global cooperation and a diversity of solu-
tions for different cultural contexts, as is the case in climate change
among numerous others global crises. One promising solution to
these generalizability and practicality limitations in the behavioral
sciences relies on embracing international collaborative science.
Large global scientific projects can benefit from access to not only a
wider range of populations but also from a diversity of scientific
perspectives. For example, crowdsourcing has been found to
Fig. 4. Country-­level means of each outcome variable. Countries without available data are shown in gray. Statistics are shown in tables S5 to S8. (A) Climate change
belief, (B) policy support, (C) sharing information on social media, and (D) trees planted via the WEPT.
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
improve the quality of scientific investigations by promoting ide-
ation, inclusiveness, transparency, rigor, and reliability among oth-
er factors (25). Thus, crowdsourcing decisions related to the
experimental design from experts more widely representative of the
global scientific community might increase the impact and general-
izability of scientific investigations. For example, the crowdsourc-
ing of the theories tested from our large international team has led
us to include less established interventions, such as “letter to future
generation,” which ended up being one of the top interventions
tested. Future work could also consider extending this crowdsourc-
ing paradigm to include nonexperts (e.g., lay audiences), as recent
work suggests that there may be unique benefits (e.g., increased in-
terdisciplinarity), sometimes even producing research questions
that outperform experts’ suggestions (53). Last, combining this
“many labs” approach (24) with the megastudy approach (18)
promises to push the limits of conventional scientific practices and
overcome some of the main barriers of science generalization and
implementation (17).
Overall, we tested the effectiveness of 11 expert-­crowdsourced
behavioral interventions at increasing climate awareness and action
in 63 countries. Our findings provide theoretical support for many
of the tested interventions. However, variation in effectiveness
across outcomes, between countries, and along the spectrum of cli-
mate beliefs, suggests substantial gaps in our current theoretical
9 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
understanding of climate change behavior. Moreover, the high pre-
existing levels of belief and policy support, alongside the small ef-
fect sizes observed here, raise critical questions about the practical
capacity to facilitate bottom-­up change at a global level, suggesting
that top-­down change might need to be prioritized to achieve the
emissions reduction necessary to stay within safe planetary limits
for human civilization. Practically, these findings provide critical
information to policymakers considering climate solution imple-
mentations, streamlining the behavioral sciences’ response to the
climate crisis.
MATERIALS AND METHODS
Participants
The data were collected between July 2022 and July 2023 (see “note
added in proof ”). A total of 83,927 completed the study. Of them,
59,440 participants (Mage = 39.13, SDage = 15.76; 50% women and
46% men) from 63 countries (Fig. 5 and Table 1) who passed the two
attention checks (i.e., “Please select the color “purple” from the list
below” and “To indicate you are reading this paragraph, please type
the word sixty in the text box below”) and correctly completed the
WEPT demo were included in the analyses. Although removing
participants who failed these preregistered attention checks risks
contributing to a selection bias in the sample (54), we a priori deter-
mined we would screen participants according to these criteria to
ensure data quality.
Ethics approval was obtained independently by each data collec-
tion team from their corresponding Institutional Review Board.
Only datasets submitted along with Institutional Review Board ap-
proval were included in the analysis.
Collaboration procedure
Following procedures from Van Bavel and colleagues (24), the orga-
nizational team submitted a call for collaboration (https://manylab-
sclimate.wordpress.com/call-­for-­collaboration/) in November 2021
on social media (i.e., Twitter), via personal networks and by posting
on various mailing lists. We asked researchers from around the world
to join our project by contributing in one of three ways: (i) collecting
data (i.e., >500 responses) from a country in which data had not al-
ready been collected, (ii) propose an intervention that becomes in-
cluded in the final study, and/or (iii) fund data collection (i.e., >500
responses) from a country in which data had not already been col-
lected and support a local team who lacks funding. The collaborators
who proposed an intervention were asked to keep in mind time con-
straints (i.e., each intervention had to take on average at most 5 min)
and the targeted outcome variables (i.e., climate beliefs, policy sup-
port, social media sharing, and tree-­planting contributions). We re-
ceived a total of 36 proposed interventions, which were coded by the
first authors (who were blinded to the intervention authors). The cod-
ing procedure involved screening the proposed interventions for fea-
sibility in an international context, relevance for the dependent
variables, and theoretical support from prior work (quantified by pre-
viously reported effect sizes). We also aggregated similar interven-
tions and duplicates. Following this procedure, we identified 11
unique and feasible interventions, which we then asked all collabora-
tors to rate on perceived efficacy (practical support) and theoretical
value (theoretical support), initially aiming to select the top five inter-
ventions. We obtained 188 responses from our collaborators in Janu-
ary 2022 (fig. S5). Given high support for all interventions, we
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
decided to test all 11 interventions in the main study. We then con-
tacted the collaborators whose interventions had been selected to
be included in the main study, to coordinate the intervention imple-
mentation and programming on the Qualtrics survey platform (www.
qualtrics.com/). After obtaining the programmed interventions, we
gave our collaborators feedback on their submissions and allowed
them time to address our comments. After receiving the revised
interventions, we contacted expert researchers who had published
theoretical work relevant to each intervention, asking them to criti-
cally review each intervention’s implementation. For example, Jost
(55) reviewed the system justification intervention, and van der
Linden et al. (56) reviewed the scientific consensus intervention.
This process was iterated for each of the 11 interventions. After
receiving critical suggestions from these experts, we engaged in
another round of revisions. Last, in an attempt to reduce American-­
centric researcher biases, we asked all collaborators from around
the world for additional feedback on the entire survey, including
all interventions, demographics, and independent variables. This
process lasted until the end of May 2022, when we started piloting
the final version of the study, on a sample of 723 participants
(Mage = 43.6; SDage = 15.7; 52% women and 46% men, <2% nonbi-
nary), collected in the United States. Using the pilot data, we wrote
our analysis scripts and the preregistration (available at https://
aspredicted.org/blind.php?x=W83_WTL). After the piloting was
completed (July 2022), we sent our collaborators the final version of
the study in Qualtrics along with an in-­depth instructions manual
(https://osf.io/ytf89/files/osfstorage/6454f8e3b30b49156cb9
dd79/) on how to translate and adapt the study to each country. We
also instructed our collaborators to obtain ethics approval from
their International Review Boards before launching data collection.
All collaborators were given 10 months (until May 2023) to submit
their data.
Experimental design
Participants signing up to complete the study (expected to take
15 min to complete) were first asked to read and sign the informed
consent. They were then exposed to the first attention check (“Please
select the color “purple” from the list below. We would like to make
sure that you are reading these questions carefully.”), which removed
from the experiment any participants choosing an incorrect answer.
Then, participants were then given a definition of climate change:
“Climate change is the phenomenon describing the fact that the
world’s average temperature has been increasing over the past
150 years and will likely be increasing more in the future.” After read-
ing this definition, participants were randomly assigned to 1 of 12
conditions: 11 experimental interventions (Fig. 1) or a no interven-
tion control condition, in a between-­subjects design. Participants un-
der the control condition were then exposed to a short, thematically
unrelated text from the novel Great Expectations by Charles Dickens,
while participants under the experimental conditions were exposed
to an intervention. Then, all participants were directed to the out-
come variable phase, in which they rated (in random order) their (i)
climate beliefs, (ii) climate policy support, and (iii) willingness to
share climate information on social media. Last, participants were
given the chance to contribute to tree-­planting efforts by completing
the WEPT. Then, participants under the control condition were
asked to complete an additional set of variables. Last, all participants
were asked to fill out a series of demographic variables, which in-
cluded another attention check (“In the previous section, you viewed
10 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
some information about climate change. To indicate you are reading
this paragraph, please type the word sixty in the text box below.”).
Notably, participants filled out the entire survey in the primary lan-
guage of their country of residence.
Outcome variables
Climate beliefs
Climate beliefs were measured by participants’ answer to the ques-
tion “How accurate do you think these statements are?” from
0 = not at all accurate to 100 = extremely accurate. The four state-
ments were as follows: “Taking action to fight climate change is nec-
essary to avoid a global catastrophe,” “Human activities are causing
climate change,” “Climate change poses a serious threat to humanity,”
and “Climate change is a global emergency.” The Cronbach’s alpha
measure of internal consistency of this four-­item scale in this data-
set was 0.934.
Climate policy support
This dependent variable consisted of participants’ level of agreement
from 0 = not at all to 100 = very much so, with the following nine
statements: “I support raising carbon taxes on gas/fossil fuels/coal?,”
“I support significantly expanding infrastructure for public trans-
portation,” “I support increasing the number of charging stations for
electric vehicles,” “I support increasing the use of sustainable energy
such as wind and solar energy,” “I support increasing taxes on airline
companies to offset carbon emissions,” “I support protecting forest-
ed and land areas,” “I support investing more in green jobs and busi-
nesses,” “I support introducing laws to keep waterways and oceans
clean,” and “I support increasing taxes on carbon intense foods (for
example meat and dairy).” The Cronbach’s alpha measure of internal
consistency of this nine-­item scale in this dataset was 0.876.
Social media sharing
Participants were first presented with the text, “Did you know that
removing meat and dairy for only two of three meals per day could
decrease food-­related carbon emissions by 60%? It is an easy way to
fight #ClimateChange #ManyLabsClimate${e://Field/cond} source:
https://econ.st/3qjvOnn” (where “{e://Field/cond}” was replaced
with the condition code for each group). Participants were then
asked “Are you willing to share this information on your social me-
dia?,” the answer options being “Yes, I am willing to share this infor-
mation,” “I am not willing to share this information,” and “I do not
use social media.” Participants who indicated that they do not use
social media were excluded from this analysis (i.e., a third of the
sample). Moreover, participants were asked to indicate the platform
(e.g., Facebook, Twitter, and Instagram) on which they posted the
information.
WEPT tree-­planting efforts
To measure an action with a real-­world impact performed at an ac-
tual cost to participants, we used a modified version of the WEPT
(37). This task is a multitrial web-­based procedure that detects con-
sequential proenvironmental behavior by allowing participants the
opportunity of engaging in voluntary cognitive effort (i.e., screen
numerical stimuli) in exchange for donations to an environmental
organization. This measure has been validated and has been found
to correlate with well-­established scales for the assessing proenvi-
ronmental behavioral intentions (e.g., general ecological behavior
scale) (57) and with direct donation behaviors (e.g., the donation of
a part of their payment to an environmental organization) (39).
Participants were first exposed to a demonstration of the WEPT,
in which they were instructed to identify all target numbers for
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
which the first digit is even and the second digit is odd (4 of 18 num-
bers were target numbers on the demonstration page). Participants
were not allowed to advance the page until they correctly completed
the WEPT demonstration. Then, they were told that planting trees is
one of the best ways to combat climate change and that they would
have the opportunity to plant up to eight trees if they chose to en-
gage in additional pages of the item identification task (one tree per
page of WEPT completed). These pages contained 60 numbers per
page, where participants had to screen for target numbers. Along-
side these instructions, participants were shown a pictogram of
eight trees, one of which was colored green to mark their progress in
the task. Participants were allowed to exit the task at any point with
no penalty.
Demographics
Participants were asked to indicate their gender, age, education lev-
el, political orientation for economic and social issues, and house-
hold income.
Experimental conditions (interventions)
Working-­together norms
The working-­together norms intervention was submitted by
M. Vlasceanu and J. Van Bavel. This intervention was adapted from
Howe et al. (11), and it combines referencing a social norm with an
invitation to work with others toward a common goal. This working-­
together normative appeal invites people to join in and “do it to-
gether” and has been found to increase interest in and actual
charitable giving, reduce paper-­towel use in public restrooms, and
increase interest in reducing personal carbon emissions (11). Me-
diation analyses in prior work also suggested that working-­together
normative appeals are effective because they foster a feeling in par-
ticipants that they are working together with others, which can in-
crease motivation while reducing social pressure. Participants under
this condition were exposed to a flier adapted from Howe and col-
leagues (11), after which they were asked 15 questions about the fli-
er, serving as manipulation checks that were also meant to reinforce
the manipulation [e.g., “If you are taking steps toward reducing your
carbon footprint, to what extent would you feel like you are doing so
together with other Americans (or participants’ group, adapted for
each country)?” on a scale from 0 = not at all to 100 = extremely or
“How strongly do you identify with your fellow Americans (or par-
ticipants’ group, adapted for each country)?” on a scale from 0 = not
at all to 100 = extremely].
System justification
SThe sstem justification intervention was submitted by O. Buchel,
M. Tyrala, and A. Findor. This intervention is situated at the inter-
section of social identity, collective narcissism, and system justifica-
tion approaches [based on (58)] and consists of framing climate
change as uniquely threatening the way of life of participants’ na-
tionality (e.g., the American way of life). Participants were asked to
read a text emphasizing the importance of nature and the environ-
ment to one’s life [e.g., “(..) the food you eat, the sports you enjoy, the
customs you observe, how you spend your free time, or even how
you imagine growing old, all are likely impacted by where you live”],
followed by examples of the effect of climate change on the local
environment of participants’ nation [e.g., “(..) we can already see the
consequences of climate change in the United States. For example,
floods are becoming more and more frequent, putting a quarter of
Americans at risk of losing their homes. Similarly, wildfires are be-
coming more frequent and more intense, threatening millions of
11 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
Americans.”]. The text ends with an appeal to being proenviron-
mental as a patriotic gesture that will protect one’s way of life (e.g.,
“Being proenvironmental allows us to protect and preserve the
American way of life. It is patriotic to conserve the country’s natural
resources. It is important to protect and preserve our environment
so that the United States remains the United States.”). This narrative
was also intertwined with representative images of participants’
country of residence.
Binding moral foundations
The binding moral foundations intervention were submitted by
B. Douglas and M. Brauer. This intervention relies on evoking
ingroup-­loyalty and authority moral foundations, which has been
shown to increase support for proenvironmental behavior and atti-
tudes (59, 60). Participants were asked to read the following text
“We are Americans (or participants’ nationality, adapted for each
country). This means we can rise to any challenge that faces our
country. From scientists to experts in the military, there is near uni-
versal agreement that climate change is real. The time to act is now.
Using clean energy will help to keep our air, water, and land pure. It
is the American (or participants’ nationality, adapted for each coun-
try) solution to the climate crisis,” after which they were exposed to
an image of a person holding the national flag of participants’ coun-
try of residence.
Exposure to effective collective action
The exposure to effective collection action intervention was submit-
ted by E. Shuman and A. Goldenberg. This intervention features
examples of successful collective action that have had meaningful
effects on climate policies, building on prior work showing that ex-
posure to nonviolent action can increase willingness to join and
maintain support (61, 62). In addition, prior work also found that
highlighting the possibility of making real concrete changes through
collective action can increase hope, efficacy, and collective action
(61). Participants were exposed to a text explaining that the impact
people’s actions can have on curbing the effects of climate change,
citing research indicating there is still “a window of opportunity” to
make a difference. Then, participants were informed that the effec-
tiveness of people’s actions to fight climate change depends on their
ability to “come together and demand systemic change.” Participants
were then exposed to several successful examples in which people
solved global issues, such as the restoration of the ozone layer in
Fig. 5. The number of participants in each of the 63 countries represented in
the sample (Ntotal = 59,440).
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
1987. Then, participants were exposed to examples of climate activ-
ism initialized by individual people and leading to large-­scale move-
ments or policy implementation (e.g., protests by locals from the
American Midwest against fossil fuels pressured the governors of
Illinois, Indiana, Michigan, Minnesota, and Wisconsin to build a
new network for charging electric vehicles.). Images of concepts de-
scribed in the text were displayed throughout.
Future self-­continuity
The future self-­continuity intervention was submitted by V. Ponizovskiy,
L. Grigoryan, S. Grelle, and W. Hofmann. This intervention consists of
emphasizing the future self that has been found in prior work to moti-
vate future-­oriented behaviors, such as academic performance, ethical
decision making, and proenvironmental behavior (63–65). Participants
were asked to read a text emphasizing the importance of engaging in
climate action [i.e., “If no changes are made, the average temperature
can increase by up to 6.5°C (12 F) by the year 2100 (IPCC, 2022). This
would be extremely dangerous as super hurricanes, gigantic wildfires,
and extreme food, and water shortages would become commonplace.”].
They were then presented with a series of causes for this phenomenon
(i.e., “Human behaviors like energy production from fossil fuels, exces-
sive meat consumption, and car driving increase the concentrations of
greenhouse gasses in Earth’s atmosphere. Over 90% of the increase in
the world’s temperature is caused by human activity.”). Last, partici-
pants were asked to imagine that their 2030 self is writing a letter to
their present self, in which their future self is describing the actions they
would have wanted to take regarding climate change [i.e., “Please put
yourself in the year 2030—8 years from now. Take a few moments to
imagine your life in that future. Imagine how you will look, where you
will be, and who you are with. In the year 2030, it will be clear whether
keeping climate change under 2°C is still possible. It will be clear wheth-
er the necessary change occurred fast enough to match the speed of the
changing climate. As the Earth’s atmosphere continues to heat up, the
effects of climate change will be more apparent: The “highest observed
temperature” records will keep being updated, heatwaves and the
draughts will become more common, species will continue to become
extinct. Now please write yourself a “letter from the future.” This should
be a letter you are writing in the year 2030, to your past self. As the
person that you will be in 2030, what role would you think would be
appropriate for you in respect to climate change? What would you want
to tell yourself in the past? What would you like your past self to do?
Please spend a bit of time on this task and try to write at least 100 words
(five sentences), or more, if possible.”].
Scientific consensus
The scientific consensus intervention was submitted by A. van Stekelenburg,
C. Klöckner, S. Vesely, and D. Bleize. This intervention consists of
a message suggesting that climate scientists are in agreement with
each other that climate change is real and primarily caused by hu-
man action. This messaging has been found to increase people’s be-
lief in climate change and support for climate mitigation policy (56,
66). Participants were exposed to the following text “Did you know
that 99% of expert climate scientists agree that the Earth is warming
and climate change is happening, mainly because of human activity
(for example, burning fossil fuels)? [Myers et al. (67), Environmen-
tal Research Letters; Lynas et al. (68), Environmental Research Let-
ters; Doran and Zimmerman (69), EOS]”. The text was accompanied
by a pie chart with 99% of the surface area shaded.
Decreasing psychological distance
The decreasing psychological distance intervention was designed by
S. Chamberlain, D. Hine, and G. Huang. This intervention is based
12 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
on prior work finding that many perceive climate change as psycho-
logically distant (i.e., “as a set of uncertain events that may occur far
in the future, impacting distant places and affecting people dissimi-
lar to themselves”) (10). Thus, framing climate change as a psycho-
logically proximal risk issue (e.g., geographic) is expected to reduce
psychological distance and increase public engagement. Participants
were exposed to a paragraph emphasizing the impact of climate
change (i.e., “There is no doubt that humans are the main driver of
climate change. Human influence has warmed the atmosphere,
ocean, and land. Climate change is already affecting every region
across the world. It has resulted in more frequent and intense ex-
treme weather events, causing widespread harm and damage to
people, wildlife, and ecosystems. Human systems are being pushed
beyond their ability to cope and adapt.”). They were then exposed to
two examples of recent natural disasters caused by climate change in
participants’ region (e.g., U.S. participants will be exposed to infor-
mation about the 2021 record-­breaking heat wave in North America
causing the Lytton wildlife and to information about the 2017 Hur-
ricane Harvey in Texas and Hurricane Irma in Florida, killing 232
people and causing $175 billion in damage). Participants were then
asked to select the aspects of their lives affected by climate change
from a list including: food production, farming and crop produc-
tion, health and wellbeing, infectious disease, heat related harm and
deaths, lack of, mental health issues, flooding and storms, changed
land, freshwater and ocean environments, damaged infrastructure,
and economy. After making the selections, participants were pro-
vided the correct answers based on current scientific estimates (i.e.,
all the possible options). Last, participants were asked to write about
how climate change will affect them and their community (i.e.,
“Please write in a few sentences: How those climate consequences
will affect you, your friends and family, and your community. Try to
imagine these things happening today so you can be specific and
describe what it will be like.”).
Dynamic social norms
The dynamic social norms intervention were submitted by
O. Genschow, D. Loschelder, G. Sparkman, and K. C. Doell). This
intervention is based on work showing that dynamic norms (i.e.,
how other people’s behavior is changing over time) are even more im-
pactful at changing behavior than static social norms (70). Participants
in this intervention first read a paragraph emphasizing that “People in
the United States and around the world are changing: More and more
people are concerned about climate change and are now taking action
across multiple fronts,” accompanied by an image featuring relevant
data in support of this claim. Then, participants were given examples of
actions people are starting to take to mitigate the changing climate [i.e.,
“Since 2013, concerns about climate change have increased in most
countries surveyed. What kinds of actions are people taking right now?
More than ever before, people are making changes to their lifestyles,
supporting policies to address climate change, and are giving the issue
more time and attention. For example, more and more people from
around the world are now cutting back on personal consumption, es-
pecially meat and dairy products, spending time, effort, and money on
initiatives to mitigate climate change (for example, planting trees, off-
setting carbon emissions), switching to low carbon modes of transpor-
tation (for example, taking bicycles). There’s also been a notable increase
in support for climate change mitigation policy—some of the most
popular policies include attempting to conserve forests and land, tran-
sitioning to solar, wind, and other renewable energy sources, creating/
raising carbon taxes on fossil fuels, coal, gas, etc.”].
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
Correcting pluralistic ignorance
The correcting pluralistic ignorance intervention was submitted by
M. Schmitt, A. Lutz, and J. Lees. This intervention builds on work
reporting that people substantially underestimate the climate change
concern of others, a phenomenon labeled as “pluralistic ignorance”
(71). Accordingly, collective action might be limited by people’s
misperception that not many people are concerned. This interven-
tion presented real public opinion data, which show that majorities
around the world are concerned about climate change. Participants
were first asked to predict the percent of people in their country who
hold the belief that climate change is a global emergency [i.e., Re-
searchers recently conducted the “People’s Climate Vote,” which is
the World’s largest survey of public opinion on climate change
(“global warming”). A total of 1.2 million people completed the sur-
vey from 50 different countries around the globe. The survey in-
cluded people from the United States. Think for a moment about
Americans and their views on climate change. How many Ameri-
cans do you think would agree with the statement “Climate change
is a global emergency”?]. After providing a prediction, participants
were shown the actual percentage of people in their country who
hold the belief in question, according to the Peoples’ Climate Vote
(72). For example, participants in the United States will be told that
“The People’s Climate Vote found that 65% of Americans agree that
climate change is a global emergency”. For countries where the Peo-
ple’s Climate Vote does not report national level results, participants
were presented with the climate opinion of people in their region.
Letter to future generations
The letter to future generations intervention was submitted by
S. Syropoulos and E. Markowitz. This intervention involves writing
a letter to a member of the future generation, which has been shown
to reduce the psychological distance between one’s current choices
and their consequences on future generations (73, 74). Participants
were asked to write a letter to a child who will read it in the future
[i.e., “Please think of a child that is currently less than 5 years old (..)
Now imagine that child is a 30-­year-­old adult. It is approximately
the year 2055, they have started a family of their own, and they are
finding their own way in the world. Whether they recognize it or
not, they live in a world that is powerfully shaped by the decisions
we are all making now, in 2022. One day, (..) they find a letter writ-
ten today, in 2022, which is a message from you.”]. In this letter,
participants are encouraged to write about their actions toward en-
suring an inhabitable plant (i.e., “In it, you tell this family about all
of the things you have done and want to do in the future to ensure
that they will inherit a healthy, inhabitable planet. You tell them
about your own personal efforts—however small or large—to con-
front the complex environmental problems of your time, from
habitat loss to water pollution to climate change. In this letter, you
also tell this family in 2055 about how you want to be remem-
bered by them and future generations as someone who did their
best to ensure a safe, flourishing world.”). Participants were al-
lowed to write for 3 min and encouraged to write at least 100
words or 5 sentences.
Negative emotion
The negative emotion intervention was submitted by K. Doell and
C. Pretus. This intervention involves exposure to scientific facts re-
garding the impacts of climate change in a doom and gloom mes-
saging style typically used by climate communicators to induce
negative emotions as a way of stimulating mitigation behaviors (45).
Participants were first asked to report their baseline levels of emotions
13 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
related to climate change, (e.g., hopeful, anxious, depressed, scared,
indifferent, angry, helpless, and guilty). They were then exposed to
information about the consequences of climate change alongside
representative images [e.g., “Climate change is happening much
more quickly and will have a much greater impact than climate sci-
entists previously thought, according to the latest report by the In-
tergovernmental Panel on Climate Change (IPCC, 2022). If your
anxiety about climate change is dominated by fears of starving polar
bears, glaciers melting, and sea levels rising, you are barely scratch-
ing the surface of what terrors are possible, even within the lifetime
of a young adult today. And yet the swelling seas—and the cities
they will drown—have so dominated the picture of climate change/
global warming that they have blinded us to other threats, many
much closer at hand and much more catastrophic (...)”]. Last, par-
ticipants were asked to report their levels of emotions related to cli-
mate change again.
Control condition
Participants in the control condition were instructed to read a text
retrieved from the novel Great Expectations by Charles Dickens [i.e.,
“As soon as the great black velvet pall outside my little window was
shot with gray, I got up and went downstairs; every board upon the
way and every crack in every board calling after me (…) I took it in
the hope that it was not intended for early use and would not be
missed for some time.”]. Participants were required to spend at least
10 s reading this text. This was to ensure that participants exerted
some level of cognitive effort before being exposed to the dependent
variable phase, to mirror the experience of participants in the ex-
perimental conditions. We chose a fiction text to prevent priming
participants in any relevant way that could influence the dependent
variables. After reading the excerpt, participants under the control
condition were directed to the dependent variable phase, followed
by the demographics phase. Last, participants under the control
condition were also directed to an additional independent variable
phase, exclusive to participants under this condition.
Additional variables collected
These variables were only displayed to participants under the control
condition, after they completed all dependent variables. First, partici-
pants were asked to rate the competence of climate scientists (“On aver-
age, how competent are climate change research scientists?” on a scale
from 0 = not at all to 100 = very much so), their trust in scientific re-
search about climate change (“On average, how much do you trust sci-
entific research about climate change?” on a scale from 0 = not at all to
100 = very much so), their trust in their government (“On average, how
much do you trust your government?” on a scale from 0 = not at all to
100 = very much so), their attitudes toward human welfare (“To what
degree do you see yourself as someone who cares about human wel-
fare?” on a scale from 0 = not at all to 100 = very much so), their global
citizenship identity (“To what degree do you see yourself as a global
citizen?” on a scale from 0 = not at all to 100 = very much so), their
environmental identification (e.g., “To what degree do you see yourself
as someone who cares about the natural environment?” on a scale from
0 = not at all to 100 = very much so), and their extrinsic environmental
motivation (e.g., “Because of today’s politically correct standards, I try
to appear proenvironmental.” on a scale from 0 = strongly disagree to
100 = strongly agree). Then, they were asked to estimate the percentage
of people in their country who believe that climate change is a global
emergency.
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
Statistical methods
Our dependent variables have distributional properties (fig. S6) that
preclude unbiased estimation with common, off-­the-­shelf, regres-
sion tools (such as the preregistered analyses). To address this, esti-
mates presented in Fig. 2 relied on Bayesian methods and custom
likelihood functions. Full mathematical descriptions of all models
can be found in the supplied code (https://github.com/josephbb/
ManyLabsClimate). Additional analyses can be found at https://
github.com/mvlasceanu/ClimateTournament.
Belief was estimated using a hierarchical Zero-­One-­Inflated Beta
(ZOIB) model. This model was further used to derive adjusted
participant-­level estimates of preintervention belief, to avoid post­
intervention bias in subsequent models. Sharing on social media
was evaluated with a logistic regression. For WEPT, we used a geo-
metric regression with a customized likelihood function to account
for truncation and overinflation for the maximum number of trees
planted. Priors were selected using prior-­predictive simulation,
with model structure iteratively developed through analysis of the
prior predictive distribution and validated through model compari-
son using posterior predictive simulation. Posteriors were sampled
using a No-­U-­Turn Sampler implemented on a graphics process-
ing unit (GPU) with PyMC/NumPyro.
We note that these modeling choices are different from our
preregistered analysis, which specified linear (belief and policy),
ordinal (WEPT), and logistic (sharing) mixed-­effects models. Plots
of residuals from preregistered models suggested moderate to
severe violations of distributional assumptions. For this reason,
P values and estimates of effect sizes for these models may be unre-
liable. Despite these issues, we note that the findings from preregis-
tered analyses are qualitatively similar to those from the Bayesian
analyses. Overall, similarities between the preregistered and Bayes-
ian analyses suggest effects that are remarkably robust to analy-
sis decisions.
For completeness, we include the results as preregistered in
tables S9 to S12 and fig. S1. Belief and policy support were mod-
eled using a linear mixed-­effects model with climate policy sup-
port as the dependent variable, condition as the fixed effect,
including item (nine policies), participant, and country as ran-
dom effects. WEPT was modeled using an ordinal mixed-­effects
model with climate action (WEPT) as the dependent variable
and condition as the fixed effect, including country as random
effects. Sharing was modeled using an ordinal mixed-­effects
model with climate action (WEPT) as the dependent variable
and condition as the fixed effect, including country as random
effects.
To develop and evaluate our Bayesian models, we adapted an
established Principle Bayesian Workflow (75). This process begins
by identifying inference goals, domain knowledge, and features of
the dataset. Candidate statistical models are proposed, with prior
predictive checks that are used to identify reasonable priors. Data
are simulated from the prior predictive distribution, and the sta-
tistical model is fit to this simulated data. This allows for evalua-
tion of computational properties of the model, tuning of the
sampler, adjustment of the model or priors, and refinement. Key
insight was gained through visual inspection of the posterior z-­
score versus posterior contraction, which can indicate issues with
overfit, underfit, bad prior models, or poorly identified model
specification. This process was iterated on until a suitable candi-
date model and priors were identified. Last, posterior predictive
14 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
checks were used to verify that models adequately reconstructed
broad properties of the data without regard to the estimands of
interest (i.e., country/treatment effects). Failures here lead to
adjustment of the underlying model. Once all model develop-
ment criteria were satisfied, final analysis of the dataset was
used to generate estimates of treatment and country level effects
as well as all relevant figures. We note that priors for similar
parameters across models may differ as a result of this iterative
process, owing to distinct link functions and differing computa-
tional constraints. However, the impact of the prior on posterior
samples is unlikely to be meaningful, given the volume of data.
We fit the selected model to the study data using PYMC (76)
with a No U-­Turn Sampler implemented on the GPU in NumPy-
ro. We evaluated the model fit, ensuring the absence of diver-
gent transitions, sufficient mixing of the (four) Markov chains, a
large enough effective sample size, and an acceptable Estimated
Bayesian Fraction of Missing Information. Last, data were simu-
lated from the posterior distribution and visual inspection of
these posterior retrodictive checks that were used to assess
model fit. Sampling parameters were largely default and can be
found in the supplied code.
Belief
Belief was indicated for four items on a scale from 0 to 100, in-
clusive. We scaled the outcome variable for each item to 0 to 1 to
facilitate the use of common bound distributions. However, as
both 0 and 1 were possible values, our likelihood function need-
ed to account for possible inflation. Hence, we implemented a
hierarchical ZOIB regression. We developed a generative model
in which participants were estimated to have an unobserved
preintervention belief, defined by their observed belief minus
the estimated preintervention effect for their level of belief (i.e.,
as though they had been in the control condition) that was par-
tially pooled by country, which, in turn, was partially pooled via
a hyperparameter for average belief. Interventions were mod-
eled with an intercept, corresponding to the average effect, and
an effect of the estimated preintervention belief. The interven-
tion effect and intercept for the control condition were fixed at
zero. Otherwise, we modeled intervention effects using a multi-
variate normal distribution to account for covariance between
intercepts and interventions. Further, we included partially
pooled intercepts for item-­specific effects. Where necessary,
noncentered parameterizations were used to improve model fit.
Last, we extracted the posterior average preintervention be-
lief for each participant to use in modeling policy support, so-
cial media sharing, and WEPT. This reflects the observed level
of belief after adjusting for intervention effects on belief. As the
treatment effects are small, these adjustments are minimal. Ide-
ally, one would jointly model belief and other outcomes; how-
ever, the large sample sizes inherent to a megastudy impose
computational constraints, a particular issue with model devel-
opment and evaluation. Extracting intervention-­adjusted esti-
mates of initial belief enables us to examine heterogeneous
intervention effects for each of these outcomes, at a tractable
degree of model complexity. We chose to focus on belief for
evaluating heterogeneous intervention effects under the as-
sumption that belief is more likely to be a cause of support for
policy, social media sharing, and investment in tree-­planting
activities than a consequence. Full mathematical descriptions of
the model can be found in the supplied code.
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
Policy support
Support for policy was indicated for nine items on a scale from 0 to
100, inclusive. Because of computational constraints with the full data-
set, we examined the average of these items. As with belief, this out-
come was scaled from 0 to 1, and a ZOIB was used to model the data.
Policy support was modeled with an intercept, an effect of adjusted
belief, with intercept and belief effects modeled for interventions and
countries. Intervention and country effects were modeled as separate
zero-­centered normal distributions.
Social media sharing
Sharing was a binary outcome, restricted to users who used social
media. To analyze the impact on sharing, we relied on a Bayesian
logistic regression. The probability of sharing was modeled with an
intercept, an effect of adjusted belief, with intercept and belief effects
modeled for interventions and countries. Intervention and country
effects were modeled as separate zero-­centered normal distributions.
Work for environmental protection task
Participants were able to plant between one and eight trees. We be-
gan by modeling this as a truncated geometric distribution, assum-
ing that participants have a per–time step chance of giving up and
are forced to stop at 8. However, we noticed an overabundance of
planting eight trees consistent with some participants committing to
planting all eight. Accordingly, we modified our likelihood to in-
clude inflation at eight trees. Posterior predictive fits confirmed ad-
equate model fit. With this likelihood, we constructed a Bayesian
hierarchical with an intercept, an effect of adjusted belief, and inter-
cepts and belief effects modeled for interventions and countries.
Note added in proof: After this manuscript was accepted for pub-
lication, the authors alerted the editorial office to a paper they re-
cently finalized that includes data used in this paper. This data can
be found at: K. C. Doell, et al. The International Climate Psychology
Collaboration: Climate change-­related data collected from 63 coun-
tries. (2024). https://doi.org/10.31234/osf.io/7fy2g
Supplementary Materials
This PDF file includes:
Figs. S1 to S6
Tables S1 to S25
REFERENCES AND NOTES
1. IPCC, Climate Change 2023: Synthesis Report. A Report of the Intergovernmental Panel on
Climate Change. Contribution of Working Groups I, II and III to the Sixth Assessment Report of
the Intergovernmental Panel on Climate Change (IPCC, 2023).
2. D. Ivanova, J. Barrett, D. Wiedenhofer, B. Macura, M. Callaghan, F. Creutzig, Quantifying
the potential for climate change mitigation of consumption options. Environ. Res. Lett.
15, 093001 (2020).
3. R. Gifford, C. Kormos, A. McIntyre, Behavioral dimensions of climate change: Drivers,
responses, barriers, and interventions. WIREs Climate Change 2, 801–827 (2011).
4. I. M. Otto, J. F. Donges, R. Cremades, A. Bhowmik, R. J. Hewitt, W. Lucht, J. Rockström,
F. Allerberger, M. McCaffrey, S. S. P. Doe, A. Lenferna, N. Morán, D. P. van Vuuren,
H. J. Schellnhuber, Social tipping dynamics for stabilizing Earth’s climate by 2050. Proc.
Natl. Acad. Sci. U.S.A. 117, 2354–2365 (2020).
5. H. Allcott, S. Mullainathan, Behavior and energy policy. Science 327, 1204–1205 (2010).
6. S. Carattini, S. Kallbekken, A. Orlov, How to win public support for a global carbon tax.
Nature 565, 289–291 (2019).
7. M. A. Jenny, C. Betsch, Large-­scale behavioural data are key to climate policy. Nat. Hum.
Behav. 6, 1444–1447 (2022).
8. J. W. Composto, E. U. Weber, Effectiveness of behavioural interventions to reduce
household energy demand: A scoping review. Environ. Res. Lett. 17, 063005 (2022).
9. C. F. Nisa, J. J. Bélanger, B. M. Schumpe, D. G. Faller, Meta-­analysis of randomised
controlled trials testing behavioural interventions to promote household action on
climate change. Nat. Commun. 10, 4545 (2019).
15 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
10. C. Jones, D. W. Hine, A. D. G. Marks, The future is now: Reducing psychological distance to
increase public engagement with climate change. Risk Anal. 37, 331–341 (2017).
11. L. C. Howe, P. B. Carr, G. M. Walton, Normative appeals motivate people to contribute to
collective action problems more when they invite people to work together toward a
common goal. J. Pers. Soc. Psychol. 121, 215–238 (2021).
12. C. J. Clark, T. Costello, G. Mitchell, P. E. Tetlock, Keep your enemies close: Adversarial
collaborations will improve behavioral science. J. Appl. Res. Mem. Cogn. 11, 1–18 (2022).
13. D. Kahneman, Experiences of collaborative research. Am. Psychol. 58, 723–730 (2003).
14. S. Benartzi, J. Beshears, K. L. Milkman, C. R. Sunstein, R. H. Thaler, M. Shankar,
W. Tucker-­Ray, W. J. Congdon, S. Galing, Should governments invest more in nudging?
Psychol. Sci. 28, 1041–1055 (2017).
15. A. L. Duckworth, K. L. Milkman, A guide to megastudies. PNAS Nexus 1, pgac214 (2022).
16. M. Bergquist, M. Thiel, M. H. Goldberg, S. van der Linden, Field interventions for climate
change mitigation behaviors: A second-­order meta-­analysis. Proc. Natl. Acad. Sci. U.S.A.
120, e2214851120 (2023).
17. K. C. Doell, Megastudies to test the efficacy of behavioural interventions. Nat. Rev.
Psychol. 2, 263 (2023).
18. K. L. Milkman, D. Gromet, H. Ho, J. S. Kay, T. W. Lee, P. Pandiloski, Y. Park, A. Rai,
M. Bazerman, J. Beshears, L. Bonacorsi, C. Camerer, E. Chang, G. Chapman, R. Cialdini,
H. Dai, L. Eskreis-­Winkler, A. Fishbach, J. J. Gross, S. Horn, A. Hubbard, S. J. Jones, D. Karlan,
T. Kautz, E. Kirgios, J. Klusowski, A. Kristal, R. Ladhania, G. Loewenstein, J. Ludwig,
B. Mellers, S. Mullainathan, S. Saccardo, J. Spiess, G. Suri, J. H. Talloen, J. Taxer, Y. Trope,
L. Ungar, K. G. Volpp, A. Whillans, J. Zinman, A. L. Duckworth, Megastudies improve the
impact of applied behavioural science. Nature 600, 478–483 (2021).
19. A. M. van Valkengoed, W. Abrahamse, L. Steg, To select effective interventions for
pro-­environmental behaviour change, we need to consider determinants of behaviour.
Nat. Hum. Behav. 6, 1482–1492 (2022).
20. J. Henrich, S. J. Heine, A. Norenzayan, Most people are not WEIRD. Nature 466, 29 (2010).
21. K. C. Doell, P. Pärnamets, E. A. Harris, L. M. Hackel, J. J. Van Bavel, Understanding the
effects of partisan identity on climate change. Curr. Opin. Behav. Sci. 42, 54–59 (2021).
22. M. J. Hornsey, E. A. Harris, K. S. Fielding, Relationships among conspiratorial beliefs,
conservatism and climate scepticism across nations. Nat. Clim. Change 8, 614–620 (2018).
23. L. Chancel, Global carbon inequality over 1990–2019. Nat. Sustain. 5, 931–938 (2022).
24. J. J. Van Bavel, A. Cichocka, V. Capraro, H. Sjåstad, J. B. Nezlek, T. Pavlović, M. Alfano,
M. J. Gelfand, F. Azevedo, M. D. Birtel, A. Cislak, P. L. Lockwood, R. M. Ross, K. Abts,
E. Agadullina, J. J. B. Aruta, S. N. Besharati, A. Bor, B. L. Choma, C. D. Crabtree,
W. A. Cunningham, K. De, W. Ejaz, C. T. Elbaek, A. Findor, D. Flichtentrei, R. Franc,
B. Gjoneska, J. Gruber, E. Gualda, Y. Horiuchi, T. L. D. Huynh, A. Ibanez, M. A. Imran,
J. Israelashvili, K. Jasko, J. Kantorowicz, E. Kantorowicz-­Reznichenko, A. Krouwel,
M. Laakasuo, C. Lamm, C. Leygue, M.-­J. Lin, M. S. Mansoor, A. Marie, L. Mayiwar,
H. Mazepus, C. McHugh, J. P. Minda, P. Mitkidis, A. Olsson, T. Otterbring, D. J. Packer,
A. Perry, M. B. Petersen, A. Puthillam, J. C. Riaño-­Moreno, T. Rothmund,
H. Santamaría-­García, P. C. Schmid, D. Stoyanov, S. Tewari, B. Todosijević, M. Tsakiris,
H. H. Tung, R. G. Umbreș, E. Vanags, M. Vlasceanu, A. Vonasch, M. Yucel, Y. Zhang, M. Abad,
E. Adler, N. Akrawi, H. A. Mdarhri, H. Amara, D. M. Amodio, B. G. Antazo, M. Apps, F. C. Ay,
M. H. Ba, S. Barbosa, B. Bastian, A. Berg, M. P. Bernal-­Zárate, M. Bernstein, M. Białek,
E. Bilancini, N. Bogatyreva, L. Boncinelli, J. E. Booth, S. Borau, O. Buchel, C. D. Cameron,
C. F. Carvalho, T. Celadin, C. Cerami, H. N. Chalise, X. Cheng, L. Cian, K. Cockcroft,
J. Conway, M. A. Córdoba-­Delgado, C. Crespi, M. Crouzevialle, J. Cutler, M. Cypryańska,
J. Dabrowska, M. A. Daniels, V. H. Davis, P. N. Dayley, S. Delouvee, O. Denkovski,
G. Dezecache, N. A. Dhaliwal, A. B. Diato, R. Di Paolo, M. Drosinou, U. Dulleck, J. Ekmanis,
A. S. Ertan, T. W. Etienne, H. H. Farhana, F. Farkhari, H. Farmer, A. Fenwick, K. Fidanovski,
T. Flew, S. Fraser, R. B. Frempong, J. A. Fugelsang, J. Gale, E. B. Garcia-­Navarro,
P. Garladinne, O. Ghajjou, T. Gkinopoulos, K. Gray, S. M. Griffin, B. Gronfeldt, M. Gümren,
R. L. Gurung, E. Halperin, E. Harris, V. Herzon, M. Hruška, G. Huang, M. F. C. Hudecek,
O. Isler, S. Jangard, F. J. Jørgensen, F. Kachanoff, J. Kahn, A. K. Dangol, O. Keudel, L. Koppel,
M. Koverola, E. Kubin, A. Kunnari, Y. Kutiyski, O. Laguna, J. Leota, E. Lermer, J. Levy, N. Levy,
C. Li, E. U. Long, C. Longoni, M. Maglić, D. McCashin, A. L. Metcalf, I. Mikloušić,
S. El Mimouni, A. Miura, J. Molina-­Paredes, C. Monroy-­Fonseca, E. Morales-­Marente,
D. Moreau, R. Muda, A. Myer, K. Nash, T. Nesh-­Nash, J. P. Nitschke, M. S. Nurse, Y. Ohtsubo,
V. Oldemburgo de Mello, C. O’Madagain, M. Onderco, M. S. Palacios-­Galvez, J. Palomäki,
Y. Pan, Z. Papp, P. Pärnamets, M. Paruzel-­Czachura, Z. Pavlović, C. Payán-­Gómez,
S. Perander, M. M. Pitman, R. Prasad, J. Pyrkosz-­Pacyna, S. Rathje, A. Raza,
G. G. Rêgo, K. Rhee, C. E. Robertson, I. Rodríguez-­Pascual, T. Saikkonen,
O. Salvador-­Ginez, W. M. Sampaio, G. C. Santi, N. Santiago-­Tovar, D. Savage, J. A. Scheffer,
P. Schönegger, D. T. Schultner, E. M. Schutte, A. Scott, M. Sharma, P. Sharma, A. Skali,
D. Stadelmann, C. A. Stafford, D. Stanojević, A. Stefaniak, A. Sternisko, A. Stoica,
K. K. Stoyanova, B. Strickland, J. Sundvall, J. P. Thomas, G. Tinghög, B. Torgler, I. J. Traast,
R. Tucciarelli, M. Tyrala, N. D. Ungson, M. S. Uysal, P. A. M. Van Lange, J.-­W. van Prooijen,
D. van Rooy, D. Västfjäll, P. Verkoeijen, J. B. Vieira, C. von Sikorski, A. C. Walker,
J. Watermeyer, E. Wetter, A. Whillans, R. Willardt, M. J. A. Wohl, A. D. Wójcik, K. Wu,
Y. Yamada, O. Yilmaz, K. Yogeeswaran, C.-­T. Ziemer, R. A. Zwaan, P. S. Boggio, National
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
identity predicts public health support during a global pandemic. Nat. Commun. 13, 517
(2022).
25. E. L. Uhlmann, C. R. Ebersole, C. R. Chartier, T. M. Errington, M. C. Kidwell, C. K. Lai,
R. J. McCarthy, A. Riegelman, R. Silberzahn, B. A. Nosek, Scientific utopia III:
Crowdsourcing science. Perspect. Psychol. Sci. 14, 711–733 (2019).
26. M. J. Hornsey, E. A. Harris, P. G. Bain, K. S. Fielding, Meta-­analyses of the determinants and
outcomes of belief in climate change. Nat. Clim. Change 6, 622–626 (2016).
27. A. J. Yeganeh, A. P. McCoy, T. Schenk, Determinants of climate change policy adoption: A
meta-­analysis. Urban Clim. 31, 100547 (2020).
28. N. Chater, G. Loewenstein, The i-­frame and the s-­frame: How focusing on the individual-­
level solutions has led behavioral public policy astray. Behav. Brain Sci. 46, e147 (2023).
29. M. Fairbrother, Public opinion about climate policies: A review and call for more studies
of what people want. PLOS Clim. 1, e0000030 (2022).
30. K. S. Nielsen, V. Cologna, F. Lange, C. Brick, P. C. Stern, The case for impact-­focused
environmental psychology. J. Environ. Psychol. 74, 101559 (2021).
31. F. Creutzig, L. Niamir, X. Bai, M. Callaghan, J. Cullen, J. Díaz-­José, M. Figueroa, A. Grubler,
W. F. Lamb, A. Leip, E. Masanet, É. Mata, L. Mattauch, J. C. Minx, S. Mirasgedis,
Y. Mulugetta, S. B. Nugroho, M. Pathak, P. Perkins, J. Roy, S. de la Rue du Can, Y. Saheb,
S. Some, L. Steg, J. Steinberger, D. Ürge-­Vorsatz, Demand-­side solutions to climate change
mitigation consistent with high levels of well-­being. Nat. Clim. Change 12, 36–46 (2022).
32. E. Salomon, J. L. Preston, M. B. Tannenbaum, Climate change helplessness and the (de)
moralization of individual energy behavior. J. Exp. Psychol.-­Appl. 23, 15–28 (2017).
33. B. Simpson, Social identity and cooperation in social dilemmas. Ration. Soc. 18,
10.1177/1043463106066381 (2006).
34. W. J. Brady, J. A. Wills, J. T. Jost, J. A. Tucker, J. J. Van Bavel, Emotion shapes the diffusion of
moralized content in social networks. Proc. Natl. Acad. Sci. U.S.A. 114, 7313–7318 (2017).
35. K. Hayhoe, Saving Us (Atria/One Signal Publishers, 2022).
36. C. Kormos, R. Gifford, The validity of self-­report measures of proenvironmental behavior:
A meta-­analytic review. J. Environ. Psychol. 40, 359–371 (2014).
37. F. Lange, S. Dewitte, The work for environmental protection task: A consequential
web-­based procedure for studying pro-­environmental behavior. Behav. Res. Methods 54,
133–145 (2022).
38. L. R. Glasman, D. Albarracín, Forming attitudes that predict future behavior: A
meta-­analysis of the attitude-­behavior relation. Psychol. Bull. 132, 778–822 (2006).
39. F. Lange, Behavioral paradigms for studying pro-­environmental behavior: A systematic
review. Behav. Res. Methods 55, 600–622 (2023).
40. P. Sheeran, T. L. Webb, The intention–behavior gap. Soc. Personal. Psychol. Compass 10,
503–518 (2016).
41. L. Hespanhol, C. S. Vallio, L. M. Costa, B. T. Saragiotto, Understanding and interpreting confidence
and credible intervals around effect estimates. Braz. J. Phys. Ther. 23, 290–301 (2019).
42. M. E. Mann, S. J. Hassol, T. Toles, “Opinion | Doomsday scenarios are as harmful as climate
change denial,” Washington Post, 12 July 2017; www.washingtonpost.com/opinions/
doomsday-­scenarios-­are-­as-­harmful-­as-­climate-­change-­denial/2017/07/12/880ed002-­
6714-­11e7-­a1d7-­9a32c91c6f40_story.html.
43. C. E. Robertson, N. Pröllochs, K. Schwarzenegger, P. Pärnamets, J. J. Van Bavel,
S. Feuerriegel, Negativity drives online news consumption. Nat. Hum. Behav. 7, 812–822 (2023).
44. K. C. Doell, B. Conte, T. Brosch, Interindividual differences in environmentally relevant
positive trait affect impacts sustainable behavior in everyday life. Sci. Rep. 11, 20423 (2021).
45. D. A. Chapman, B. Lickel, E. M. Markowitz, Reassessing emotion in climate change
communication. Nat. Clim. Change 7, 850–852 (2017).
46. K. C. Doell, L. Lengersdorff, S. A. Rhoads, B. Todorova, J. P. Nitschke, J. Druckman,
M. Vlasceanu, Manylabs Climate Consortia, C. Lamm, J. J. Van Bavel, Academics are more
accurate at predicting the efficacy of climate action interventions for outcomes with
more recent scientific investigation.
47. A. Coppock, T. J. Leeper, K. J. Mullinix, Generalizability of heterogeneous treatment
effect estimates across samples. Proc. Natl. Acad. Sci. U.S.A. 115, 12441–12446 (2018).
48. A. Coppock, Persuasion in Parallel: How Information Changes Minds About Politics (The
University of Chicago Press, 2022).
49. K. J. Mullinix, T. J. Leeper, J. N. Druckman, J. Freese, The generalizability of survey
experiments. J. Exp. Pol. Sci. 2, 109–138 (2015).
50. J. D. Weinberg, J. Freese, D. McElhattan, Comparing data characteristics and results of an
online factorial survey between a population-­based and a crowdsource-­recruited
sample. Sociol. Sci. 1, 292–310 (2014).
51. E. T. Berkman, S. M. Wilson, So useful as a good theory? The practicality crisis in (social)
psychological theory. Perspect. Psychol. Sci. 16, 864–874 (2021).
52. A. G. Thalmayer, C. Toscanelli, J. J. Arnett, The neglected 95% revisited: Is American
psychology becoming less American? Am. Psychol. 76, 116–129 (2021).
53. S. Beck, T. Brasseur, M. Poetz, H. Sauermann, Crowdsourcing research questions in
science. Res. Policy 51, 104491 (2022).
54. A. J. Berinsky, M. F. Margolis, M. W. Sances, Separating the shirkers from the workers?
Making sure respondents pay attention on self-­administered surveys. Am. J. Polit. Sci. 58,
739–753 (2014).
16 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
55. J. T. Jost, A Theory of System Justification (Harvard Univ. Press, 2020).
56. S. L. van der Linden, A. A. Leiserowitz, G. D. Feinberg, E. W. Maibach, The scientific consensus on
climate change as a gateway belief: Experimental evidence. PLOS ONE 10, e0118489 (2015).
57. F. G. Kaiser, M. Wilson, Goal-­directed conservation behavior: The specific composition of a
general performance. Personal. Individ. Differ. 36, 1531–1544 (2004).
58. I. Feygina, J. T. Jost, R. E. Goldsmith, System justification, the denial of global warming, and
the possibility of “system-­sanctioned change”. Pers. Soc. Psychol. Bull. 36, 326–338 (2010).
59. C. Wolsko, H. Ariceaga, J. Seiden, Red, white, and blue enough to be green: Effects of
moral framing on climate change attitudes and conservation behaviors. J. Exp. Soc.
Psychol. 65, 7–19 (2016).
60. M. Feinberg, R. Willer, Moral reframing: A technique for effective and persuasive
communication across political divides. Soc. Personal. Psychol. Compass 13, e12501 (2019).
61. A. Goldenberg, S. Cohen-­Chen, J. P. Goyer, C. S. Dweck, J. J. Gross, E. Halperin, Testing the
impact and durability of a group malleability intervention in the context of the
Israeli–Palestinian conflict. Proc. Natl. Acad. Sci. U.S.A 115, 696–701 (2018).
62. E. Shuman, T. Saguy, M. van Zomeren, E. Halperin, Disrupting the system constructively:
Testing the effectiveness of nonnormative nonviolent collective action. J. Pers. Soc.
Psychol. 121, 819–841 (2021).
63. H. E. Hershfield, T. R. Cohen, L. Thompson, Short horizons and tempting situations: Lack
of continuity to our future selves leads to unethical decision making and behavior.
Organ. Behav. Hum. Decis. Process. 117, 298–310 (2012).
64. R. M. Adelman, S. D. Herrmann, J. E. Bodford, J. E. Barbour, O. Graudejus, M. A. Okun,
V. S. Y. Kwan, Feeling closer to the future self and doing better: Temporal psychological
mechanisms underlying academic performance. J. Pers. 85, 398–408 (2017).
65. C. Nurra, D. Oyserman, From future self to current action: An identity-­based motivation
perspective. Self Identity 17, 343–364 (2018).
66. J. B. Rode, S. Iqbal, B. J. Butler, P. H. Ditto, Using a news article to convey climate science
consensus information. Sci. Commun. 43, 651–673 (2021).
67. K. F. Myers, P. T. Doran, J. Cook, T. A. Myers, Consensus revisited: Quantifying scientific
agreement on climate change and climate expertise among Earth scientists 10 years
later. Environ. Res. Lett. 16, 104030 (2021).
68. M. Lynas, B. Z. Houlton, S. Perry, Greater than 99% consensus on human caused climate
change in the peer-­reviewed scientific literature. Environ. Res. Lett. 16, 114005 (2021).
69. P. T. Doran, M. K. Zimmerman, Examining the scientific consensus on climate change.
Environ. Res. Lett. 90, 22–23 (2009).
70. G. Sparkman, G. M. Walton, Dynamic norms promote sustainable behavior, even if it is
counternormative. Psychol. Sci. 28, 1663–1674 (2017).
71. N. Geiger, J. K. Swim, Climate of silence: Pluralistic ignorance as a barrier to climate
change discussion. J. Environ. Psychol. 47, 79–90 (2016).
72. “The peoples’ climate vote | United Nations development programme,” UNDP; www.
undp.org/publications/peoples-­climate-­vote.
73. T. R. Shrum, The salience of future impacts and the willingness to pay for climate change
mitigation: An experiment in intergenerational framing. Clim. Change 165, 18 (2021).
74. R. H. Wickersham, L. Zaval, N. A. Pachana, M. A. Smyer, The impact of place and legacy
framing on climate action: A lifespan approach. PLOS ONE 15, e0228963 (2020).
75. D. J. Schad, M. Betancourt, S. Vasishth, Toward a principled Bayesian workflow in
cognitive science. Psychol. Methods 26, 103–126 (2021).
76. J. Salvatier, T. V. Wiecki, C. Fonnesbeck, Probabilistic programming in Python using
PyMC3. PeerJ Comput. Sci. 2, e55 (2016).
77. M. Lizzio-­Wilson, E. F. Thomas, W. R. Louis, B. Wilcockson, C. E. Amiot, F. M. Moghaddam,
C. McGarty, How collective-­action failure shapes group heterogeneity and engagement
in conventional and radical action over time. Psychol. Sci. 32, 519–535 (2021).
78. S. van der Linden, The gateway belief model (GBM): A review and research agenda for
communicating the scientific consensus on climate change. Curr. Opin. Psychol. 42, 7–12 (2021).
Acknowledgments
Funding: This work was supported by Google Jigsaw grant (to M.Vl., K.C.D., and J.J.V.B.), Swiss
National Science Foundation P400PS_190997 (to K.C.D.), Dutch Research Council grant 7934
(to K.L.v.d.B.), European Union grant no. ID 776608 (to K.L.v.d.B.), John Templeton Foundation
grant 61378 (to M.A.), The National Council for Scientific and Technological Development grant
(to A.A.), Christ Church College Research Centre grant (to M.A.J.A.), David Phillips Fellowship
grant BB/R010668/2 (to M.A.J.A.), Jacobs Foundation Fellowship (to M.A.J.A.), DFG grant project
no. 390683824 (to M.A.D., P.Ba., and B.B.), NYUAD research funds (to J.J.B.), The Swiss Federal
Office of Energy through the Energy, Economy, and Society program grant number:
SI/502093-­01 (to S.Be.), The Belgian National Fund for Scientific Research (FRS-­FNRS) PDR
0253.19 (to P.Be.), Fund for scientific development at the Faculty of Psychology at SWPS
University in Warsaw (to O.Bi.), Radboud University Behavioral Science Institute (to D.N.M.B.),
Leuphana University Lüneburg research fund (to D.D.L., L.B., Y.A.E., H.M.P., and M.S.), University
of Birmingham Start up Seed Grant (to A.B.), Prime-­Pump Fund from University of Birmingham
(to A.B. and M.E.), University of Geneva Faculty Seed Funding (to T.Bro.), Pomona College Hirsch
Research Initiation Grant (to A.R.P.), Center for Social Conflict and Cohesion Studies grant ANID/
FONDAP #15130009 (to H.C. and S.D’.), Center for Intercultural and Indigenous Research grant
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
ANID/FONDAP #15110006 (to H.C. and S.D’.), National Research Foundation of Korea
NRF-­2020S1A3A2A02097375 (to D.Chu. and S.Su.), Darden School of Business (to L.C.),
Kieskompas–Election Compass (to T.W.E., A.P.M.K., V.C., and A.L.O.), The National Agency of
Research and Development, National Doctoral Scholarship 24210087 (to S.D’.), Dutch Science
Foundation (NWO) grant VI.Veni.201S.075 (to M.H.C.M.), The Netherlands Organization for
Scientific Research (NWO) Vici grant 453-­15-­005 (to I.M.E.), Foundation for Science and
Technology–FCT (Portuguese Ministry of Science, Technology and Higher Education) grant
UIDB/05380/2020 (to A.R.F.), The Slovak Research and Development Agency (APVV) contract
no. APVV-­21-­0114 (to A.F.), The James McDonnell Foundation 21st Century Science Initiative in
Understanding Human Cognition—scholar award grant 220020334 (to L.Fr. and J.N.),
Sponsored Research Agreement between Meta and Fundación Universidad Torcuato Di Tella
grant INB2376941 (to L.Fr. and J.N.), Thammasat University Fast Track Research Fund (TUFT)
12/2566 (to N.P.G.), HSE University Basic Research Program (to D.Gr. and A.Ga.), ARU Centre for
Societies and Groups Research Centre Development Funds (to S.Grad., A.J.H., and M.Z.),
University of Stavanger faculty of Social Science research activities grant (to S.Gras.), Center for
the Science of Moral Understanding (to K.G.), University of Colorado Boulder Faculty research
fund (to J.Gr.), Swiss National Science Foundation grant 203283 (to U.J.J.H.), Kochi University of
Technology Research Funds (to T.H.), RUB appointment funds (to W.H.), Dean’s Office, College of
Arts and Sciences at Seton Hall University (to F.J.), Nicolaus Copernicus University (NCU) budget
(to D.J. and A.D.W.), Sectorplan Social Sciences and Humanities, The Netherlands (to E.K.-­R.),
Erasmus Centre of Empirical Legal Studies (ECELS), Erasmus School of Law, Erasmus University
Rotterdam, The Netherlands (to E.K.-­R.), American University of Sharjah Faculty Research Grant
2020 FRG20-­M-­B134 (to O.Ka. and I.K.), Centre for Social and Early Emotional Development
SEED grant (to A.K. and E.J.K.), ANU Futures Grant (to C.K.), Research Council of Norway through
Centres of Excellence Scheme, FAIR project no. 262675 (to H.S. and S. Bø), Aarhus University
Research Foundation grant AUFF-­E-­2021-­7-­16 (to R.Kr. and L.No.), Social Perception and
Intergroup Inequality Lab at Cornell University (to A.R.K.), COVID-­19 Rapid Response grant,
University of Vienna (to C.La.), Austrian Science Fund FWF I3381 (to C.La.), FWO Postdoctoral
Fellowship 12U1221N (to F.L.), National Geographic Society (to J.L.C.), University of Michigan
Ross School of Business Faculty Research Funds (to J.L.C.), The Clemson University Media
Forensics Hub (to J.Le.), Norwegian Retailers’ Environment Fund, Poster Competition Grant
2022 (to I.R.), John Templeton Foundation grant 62631 (to N.L. and R.M.R.), ARC Discovery
Project DP180102384 (to N.L.), Medical Research Council Fellowship grant MR/P014097/1 (to
P.L.L.), Medical Research Council Fellowship grant MR/P014097/2 (to P.L.L.), Jacobs Foundation
(to P.L.L.), Wellcome Trust and the Royal Society Sir Henry Dale Fellowship grant 223264/Z/21/Z
(to P.L.L.), JFRAP grant (to J.G.L.), Social Sciences and Humanities Research Council (SSHRC)
Doctoral Fellowship (to Y.L.), Simon Fraser University Psychology Department Research Grant
(to A.E.L. and M.T.S.), GU internal funding (to A.A.M. and S.A.R.), FAPESP 2014/50279-­4 (to
K.L.M.), FAPESP 2020/15230-­5 (to K.L.M.), Shell Brasil (to K.L.M.), Brazil’s National Oil, Natural Gas
and Biofuels Agency (ANP) through the R&D levy regulation (to K.L.M.), ANR grant SCALUP,
ANR-­21-­CE28-­0016-­01 (to H.M.), NOMIS Foundation grant for the Centre for the Politics of
Feelings (to K.Mi. and M.Ts.), Applied Moral Psychology Lab at Cornell University (to S.M., L.Ni.,
and M.Z.), Universidad Peruana Cayetano Heredia Project 209465 (to F.S.M.-­R.), Belgian National
Fund for Scientific Research (FRS-­FNRS) grant PDR 0253.19 (to Y.L.M.), Riksbankens
Jubileumsfond grant P21-­0384 (to G.N.), European Research Council funded by the UKRI Grant
EP/X02170X/1 (to M.S.P. and G.A.T.), Statutory Funding of Institute of Psychology, University of
Silesia in Katowice (to M.P.-­C.), Aarhus University Research Foundation AUFF-­E-­2018-­7-­13 (to
S.P.), São Paulo Research Foundation (FAPESP) grant 2019/26665-­5 (to G.G.R.), Mistletoe
Unfettered Research Grant, National Science Foundation GRFP Award 1937959 (to S.A.R.),
Japan Society for the Promotion of Science grant 21 J01224 (to T.Sa.), Institute of Psychology &
the Faculty of Social and Political Sciences, University of Lausanne (to O.S.), Universitat Ramon
Llull, Esade Business School (to K.Sc.), University of St Andrews (to P.S.), Dutch Science
Foundation (NWO) VI.Veni.191G.034 (to C.S.), Universität Hamburg (to S.Sc.), Faculty of Health
PhD fellowship, Aarhus University (to K.So.), School of Medicine and Psychology, Australian
National University (to S.K.S.), Swedish Research Council grant 2018-­01755 (to G.T.), Russian
Federation Government grant project 075-­15-­2021-­611 (to D.Va.), Swedish Research Council (to
D.Vä.), Cooperatio Program MCOM (to M.Vr.), Stanford Center on Philanthropy and Civil Society
(to R.W.), and Canada Research Chairs program (to J.Z.). Author contributions:
Conceptualization: M.Vl., K.C.D., and J.J.V.B. Methodology: M.Vl., K.C.D., J.B.B.-­C., and J.J.V.B.
Investigation: All authors. Visualization: J.B.B.-­C. and M.Vl. Funding acquisition: M.Vl., K.C.D.,
J.J.V.B., and A.C. Project administration: M.Vl., K.C.D., S.Gr., Y.Pe., E.P., and D.Vl. Supervision: M.Vl.,
K.C.D., and J.J.V.B. Writing (original draft): M.Vl., K.C.D., J.J.V.B., J.B.B.-­C., and S.C. Writing (review
and editing): All authors. Competing interests: A.P.M.K. (Departments of Political Science and
Communication Science at Vrije Universiteit Amsterdam) is founder and stockholder of
Kieskompas (data collection service) but has not financially benefited from this data collection
or study. The authors declare that they have no other competing interests. Data and materials
availability: The data can be downloaded from https://zenodo.org/records/10345806. All data
and code can also be found on GitHub (https://github.com/josephbb/ManyLabsClimate and
https://github.com/mvlasceanu/ClimateTournament). The interventions (in each language) can
be accessed as qsf files (to be imported in Qualtrics): https://osf.io/ytf89/files/
osfstorage/6454f8d771778511d9b0f48f. A web tool for rapidly assessing which intervention is
most likely to be effective at increasing climate change beliefs, policy support, information
17 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
sharing, and tree-­planting efforts, for any subsample target of interest, varying along
demographics such as nationality, political ideology, age, gender, education, or income level
can be found at https://climate-­interventions.shinyapps.io/climate-­interventions/. All other
data needed to evaluate the conclusions in this paper are present in the paper and/or the
Supplementary Materials.
1
Department of Psychology, New York University, New York, NY 10003, USA. 2Depart-
ment of Cognition, Emotion, and Methods in Psychology, Faculty of Psychology,
University of Vienna, Vienna 1010, Austria. 3Craig Newmark Center for Journalism
Ethics and Security, Columbia University, New York, NY 10018, USA. 4Institute for
Rebooting Social Media, Harvard University, Cambridge, MA 02138, USA. 5Depart-
ment of Psychology, Stanford University, Stanford, CA 94305, USA. 6San Luis Obispo, CA
93405, USA. 7Copernicus Institute of Sustainable Development, Utrecht University,
Utrecht, 3584 CB, Netherlands. 8Department of Psychology, University of Pitts-
burgh, Pittsburgh, PA 15260, USA. 9School of Public Policy and Urban Affairs,
Northeastern University, Boston, MA 02115, USA. 10Department of Psychology,
Northeastern University, Boston, MA 02115, USA. 11Amazon, Seattle, WA 98109, USA.
12
Department of Psychology, Carl von Ossietzky University of Oldenburg, Olden-
burg 26129, Germany. 13Santa Lucia Foundation, IRCCS, Rome 179, Italy. 14Depart-
ment of Psychology, Sapienza University of Rome, Rome 185, Italy. 15Department
of Philosophy, Macquarie University, Sydney, NSW 2000, Australia. 16Universidad
Peruana Cayetano Heredia, San Martín de Porres 15102, Peru. 17Post-­Graduation
Program in Linguistics, Federal University of Paraná, Curitiba 80060150, Brasil.
18
UNSW Business School, University of New South Wales, Sydney, NSW 2052, Australia.
19
Centre for Human Brain Health, School of Psychology, University of Birmingham,
Birmingham B15 2TT, UK. 20Psychology Scientific Research Institute, Baku, Azerbai-
jan. 21Kenya Marine and Fisheries Research Institute, Kisumu 1881-­40100, Kenya.
22
Department of Psychology, University of Groningen, Groningen 9712TS, Netherlands.
23
Department of Economics, University of Hamburg, Hamburg 20146, Germany. 24De­
partment of Psychology, New York University Abu Dhabi, Abu Dhabi 129188,
United Arab Emirates. 25Department of Sociology, University of Bern, Bern 3012,
Switzerland. 26LAPCOS, Université Côte d’Azur, Nice 6357, France. 27Center for Social
and Cultural Psychology, Université libre de Bruxelles, Brussels 1050, Belgium. 28Insti-
tute of Psychology, Faculty of Historical and Pedagogical Sciences, University of
Wroclaw, Wroclaw 50-­120, Poland. 29Institute of Psychology, SWPS University, Warsaw
03-­815, Poland. 30Department of Psychology, Division of Behavioral & Organiza-
tional Sciences, Claremont Graduate University, Claremont, NH 91711, USA. 31Be-
havioural Science Institute, Radboud University, Nijmegen, 6500 HE, Netherlands.
32
Department of Strategy and Management, Norwegian School of Economics, Ber-
gen 5045, Norway. 33Department of Economic Psychology, Social Psychology and
Experimental Methods, Leuphana University Lüneburg, Lüneburg 21335, Germa-
ny. 34Social and Cognitive Neuroscience Laboratory, Mackenzie Presbyterian Uni-
versity, Sao Paulo 1241001, Brazil. 35Toulouse Business School, Institute for
Advanced Study in Toulouse. Toulouse, 31000, France. 36Department of Economics,
University of Hamburg, Hamburg 20146, Hamburg. 37School of Psychology, Univer-
sity of Birmingham, Birmingham B15 2TT, UK. 38Department of Psychology, Univer-
sity of Wisconsin–Madison, Madison, WI 53706, USA. 39Department of Psychology,
University of Amsterdam, Amsterdam, 1018 WT, Netherlands. 40Department of
Psychology, Inland Norway University of Applied Sciences, Elverum 2418, Norway.
41
Policy Research Department, Kyiv School of Economics, Kyiv 2000, Ukraine. 42De-
partment of Organisation, Strategy, and Entrepreneurship, School of Business and
Economics, Maastricht University, Maastricht 6211 LK, Netherlands. 43Department
of Psychology and Swiss Center for Affective Sciences, University of Geneva, Ge-
neva 1205, Switzerland. 44Institute for Sociology of the Slovak Academy of Scienc-
es, Slovak Academy of Sciences, Bratislava 81364, Slovakia. 45Psychological Science,
Pomona College, Claremont, CA 91711, USA. 46Department of Movement Sciences,
KU Leuven, Leuven 3001, Belgium. 47Escuela de Psicología, Pontificia Universidad
Católica de Chile, Santiago, Chile. 48School of Psychology, Speech, and Hearing,
University of Canterbury, University of Canterbury, Christchurch 8051, New Zealand.
49
Department of Marketing, King’s Business School, King’s College London, London
WC2B 4BG, UK. 50Department of Management and Marketing, University of Mel-
bourne, Melbourne, VIC 3010, Australia. 51Department of Biomedical Engineering,
Ulsan National Institute of Science and Technology, Ulsan 44919, Republic of Korea.
52
Department of Marketing, University of Virginia, Darden School of Business, Char-
lottesville, VA 22903, USA. 53Department of Psychology, The Hebrew University,
Jerusalem 9190501, Israel. 54Department of Psychology, University of Groningen,
Groningen 9712 CP, Netherlands. 55Center for Social and Cognitive Neuroscience
(CSCN), School of Psychology, Universidad Adolfo Ibáñez, Viña del Mar, Chile. 56Fac-
ulty of Management, Canadian University Dubai, Dubai 117781, United Arab Emir-
ates. 57Kieskompas–Election Compass, Amsterdam 1052XH, Netherlands. 58Escuela
de Psicología, Pontificia Universidad Católica de Chile, Santiago 8331150, Chile.
59
Department of Developmental, Personality and Social Psychology, Ghent Univer-
sity, Ghent 9000, Belgium. 60Department of People Management and Organiza-
tion, Esade Business School, Universitat Ramon Llull, Barcelona 8034, Spain.
61
Behavioral Economics and Engineering Group, KU Leuven, Leuven 3000, Bel-
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
gium. 62LP3C, Université Rennes 2, Rennes 35000, France. 63Department of Psychol-
ogy, Saarland University, Saarbrücken 66123, Germany. 64Center for Earth System
Research and Sustainability (CEN), University of Hamburg, Hamburg 20146, Ger-
many. 65Amsterdam School of Communication Research, University of Amsterdam,
Amsterdam 1018WV, Netherlands. 66Department of Psychology, University of Lat-
via, Riga, Latvia. 67Department of Management, Aarhus University, Aarhus 8210,
Denmark. 68Department of Psychology, University of Birmingham, Birmingham
B15 2TT, UK. 69Department of Vision Science, University of Leicester, Leicester LE1
7RH, UK. 70Department of Clinical Psychology, Utrecht University, Utrecht 3508 TC,
Netherlands. 71Institute of Management & Organization, Leuphana University
Lüneburg, Lüneburg 21335, Germany. 72Department of Political Science & Annen-
berg School for Communication, University of Pennsylvania, Philadelphia, PA
19104, USA. 73Department of Psychology, University of Salzburg, Salzburg 5020,
Salzburg. 74HEI-­Lab: Digital Human-­Environment Interaction Labs, Lusófona Uni-
versity, Lisbon 1700, Portugal. 75School of Management, Ludwig Maximilian Uni-
versity of Munich, Munich 80539, Germany. 76Institute of European Studies and
International Relations, Faculty of Social and Economic Sciences, Comenius Univer-
sity Bratislava, Bratislava 82105, Slovakia. 77Laboratorio de Neurociencia, Escuela
de Negocios, Universidad Torcuato Di Tella, Buenos Aires C1428, Argentina.
78
School of Global Studies, Thammasat University, Bangkok 12121, Thailand. 79Cen-
ter for Sociocultural Research, HSE University, Moscow 101000, Russia. 80Environ-
mental Psychology, Department of Cognition, Emotion, and Methods in
Psychology, Faculty of Psychology, University of Vienna, Vienna A-­1010, Austria.
81
Institute for Management and Organization, Leuphana University Lüneburg,
Lüneburg 21335, Germany. 82Macedonian Academy of Sciences and Arts, Skopje
1000, North Macedonia. 83Faculty of Philosophy, Institute of Psychology, Faculty of
Philosophy, Jagiellonian University, Krakow 30-­060, Poland. 84Jigsaw, Google, New
York, NY 10011, USA. 85Harvard Business School, Harvard University, Boston, MA
2163, USA. 86Department of Psychology, Harvard University, Cambridge, MA 2138,
USA. 87Digital Data and Design Institute at Harvard, Harvard University, Allston,
Boston, MA 2134, USA. 88School of Psychology and Sport Science, Anglia Ruskin
University, Cambridge CB1 1PT, UK. 89Psychosocial Science, University of Bergen,
Bergen 5007, Norway. 90Cognitive and Behavioral Neuroscience Laboratory, Uni-
versity of Stavanger, Stavanger 4021, Norway. 91Department of Psychology and
Neuroscience, University of North Carolina, Chapel Hill, Chapel Hill, NC 27599, USA.
92
Department of Psychology, Ruhr University Bochum, Bochum 44801, Germany.
93
Department of Psychology, University of Limerick, Limerick V94T9PX, Ireland.
94
Department of Psychology, University of York, York YO10 5DD, UK. 95Department
of Personality Psychology, Yerevan State University, Yerevan 0025, Armenia. 96De-
partment of Psychology and Neuroscience, University of Colorado Boulder, Boul-
der, CO 80309, USA. 97Division of Social Sciences, University of the Philippines
Visayas, Miagao 5023, Philippines. 98Baruch Ivcher School of Psychology, Reichman
University, Herzliya 4610101, Israel. 99Faculty of Psychology, University of Basel, Ba-
sel 4055, Switzerland. 100Department of Psychology, University of Oslo, Oslo 373,
Norway. 101Nencki Institute of Experimental Biology, Polish Academy of Sciences,
Warsaw 02-­093, Poland. 102School of Psychology, University of Sussex, Falmer BN1
9RH, UK. 103Anderson School of Management, University of California, Los Angeles,
Los Angeles, CA 90095, USA. 104School of Economics & Management, Kochi Univer-
sity of Technology, Kami City 782-­8502, Japan. 105School of Psychology, Speech and
Hearing, University of Canterbury, Christchurch 8051, New Zealand. 106Department
of Business Administration, University of Zurich, Zurich 8032, Switzerland. 107De-
partment of Psychology, Universidad Nacional de San Antonio Abad del Cusco,
Cusco 800, Peru. 108Department of Media and Communication, City University of
Hong Kong, Kowloon 999077, Hong Kong, China. 109Department of Psychology,
Japan Women’s University, Tokyo 1128681, Japan. 110Graduate School of Education,
Tohoku University, Sendai 9808576, Japan. 111Department of Psychology, Seton
Hall University, South Orange, NJ 7079, USA. 112Department of Psychology, Faculty
of Philosophy, University of Novi Sad, Novi Sad 21000, Serbia. 113Doctoral School of
Social Sciences, Nicolaus Copernicus University, Toruń 87-­100, Poland. 114Green
Dock, Hostivice 25301, Czech Republic. 115Faculty of Social Sciences, Tampere Uni-
versity, Tampere 33100, Finland. 116INVEST Research Flagship, University of Turku,
Turku 20014, Finland. 117Institute of Security and Global Affairs, Leiden University,
The Hague 2511DP, Netherlands. 118Erasmus School of Law, Erasmus University
Rotterdam, Rotterdam 3062PA, Netherlands. 119Shamir Research Institute, Univer-
sity of Haifa, Haifa 3498838, Israel. 120Department of Learning and Instructional Sci-
ences, University of Haifa, Haifa, 3498838, Israel. 121Deparment of Economics,
American University of Sharjah, Sharjah 26666, United Arab Emirates. 122School of
Psychology, Deakin University, Geelong, VIC 3216, Australia. 123School of Philoso-
phy, Australian National University, Canberra, ACT 2600, Australia. 124Department of
Psychology, Norwegian University of Science and Technology, Trondheim 7049,
Norway. 125Department of Management and Engineering, Linköping University,
Linköping 58183, Sweden. 126Academic and Research Laboratory of Neurotechnol-
ogy, Ural Federal University, Ekaterinburg 620075, Russia. 127Department of Experi-
mental Psychology, Ghent University, Ghent 9000, Belgium. 128Department of
Psychology, Cornell University, Ithaca, NY 14850, USA. 129Departments of Political
Science and Communication Science, Vrije Universiteit Amsterdam, Amsterdam
18 of 19
S c i e n c e A d va n c e s | R e s e ar c h A r t i c l e
1081 HV, Netherlands. 130Department of Psychology, University of Cambridge,
Cambridge CB2 3EL, UK. 131Psychology of Sustainability and Behavior Change, Uni-
versity of Basel, Basel 4055, Switzerland. 132Management & Organizations, Stephen
M. Ross School of Business, University of Michigan, Ann Arbor, MI 48105, USA.
133
John E. Walker Department of Economics, Clemson University, Clemson, SC
29634, USA. 134Andlinger Center for Energy and the Environment, Princeton Uni-
versity, Princeton, NJ 8544, USA. 135School of Business and Creative Industries, Uni-
versity of the Sunshine Coast, Queensland, BNE 4556, Australia. 136Department of
Philosophy, Macquarie University, Sydney, NSW 2109, Australia. 137Department of
Marketing, Bocconi University, Milan 20136, Italy. 138Department of Communica-
tion Science, Vrije Universiteit Amsterdam, Amsterdam 1081 HV, Netherlands. 139In-
stitute of Management and Organization, Leuphana University of Lüneburg,
Lueneburg 21337, Germany. 140MIT Sloan School of Management, Massachusetts In-
stitute of Technology, Cambridge, MA 2139, USA. 141Department of Psychology, Uni-
versity of British Columbia, Vancouver, BC V6T 1Z4, Canada. 142Tanzanian Fisheries
Research Institute, Mwanza, Tanzania. 143Department of Psychology, Simon Fraser
University, Burnaby, BC V5A 1S6, Canada. 144Department of Social, Organizational, &
Economic Psychology, University of Hildesheim, Hildesheim 31141, Germany.
145
Department of Environmental Conservation, University of Massachusetts Am-
herst, Amherst, MA 1003, USA. 146Department of Psychology, Georgetown Univer-
sity, Washington, DC 20057, USA. 147Research Centre for Greenhouse Gas
Innovation (RCGI), University of São Paulo, São Paulo 05508-­030, Brazil. 148Depart-
ment of Social Psychology, Institute of Psychology, University of São Paulo, São
Paulo 05508-­030, Brazil. 149National Fisheries Resources Research Institute, Jinja,
Uganda. 150Department of Civil and Water Resources Engineering School of Engi-
neering and Technology, Sokoine University of Agriculture, Morogoro, Tanzania.
151
Department of Psychology, University of Limerick, Limerick V94 T9PX, Ireland.
152
Department of Communication Science, University of Amsterdam, Amsterdam
1001 NG, Netherlands. 153Institut Jean Nicod, Département d’études cognitives,
ENS, EHESS, PSL University, CNRS, Paris 75005, France. 154Tanzania Fisheries Research
Institute, Mwanza, Tanzania. 155Royal Holloway, University of London, Egham
TW200EX, UK. 156School of Culture and Society–Interacting Minds Centre, Aarhus
University, Aarhus 8000, Denmark. 157Neuropsychopharmacology and Biopsychol-
ogy Unit, University of Vienna, Vienna 1010, Austria. 158Department of Psychology,
Universidad Peruana Cayetano Heredia, Lima 2002, Peru. 159Fonds de la Recherche
Scientifique, Brussels 1050, Belgium. 160Center for Social and Cultural Psychology,
Université libre de Bruxelles, Brussels 1312, Belgium. 161School of Psychology, Uni-
versity of Auckland, Auckland 1010, New Zealand. 162Department of Management,
The University of Tokyo, Tokyo 113-­8654, Japan. 163Department of Psychology, Uni-
versity of Cordoba, Cordoba 14071, Spain. 164Département d’études cognitives, Insti-
tut Jean Nicod ENS, EHESS, PSL University, CNRS, Tokyo 113-­ 8654, Japan.
165
Comisión Nacional de Investigaciones Científicas y Técnicas (CONICET), Buenos
166
Aires, Argentina. Laboratorio de Neurociencia, Escuela de Negocios, Universidad
Torcuato Di Tella, Buenos Aires C1428 CABA, Argentina. 167University of Economics
HCMC (UEH), Ho Chi Minh City, Vietnam. 168College of Management, National Kaoh-
siung University of Science and Technology, Kaohsiung 800, Taiwan. 169Department
of Psychology and Dyson School of Applied Economics and Management, Cornell
University, Ithaca, NY 14850, USA. 170Environmental Psychology, Department of
Cognition, Emotion, and Methods in Psychology, Faculty of Psychology, University
of Vienna, Vienna 1010, Austria. 171Department of Clinical Neuroscience, Karolinska
Institutet, Stockholm 17177, Sweden. 172Department of Psychology, Stockholm
University, Stockholm 11419, Sweden. 173Department of Psychology and Behav-
ioural Sciences, Aarhus University, Aarhus 8000, Denmark. 174Department of Psy-
chology, Kadir Has University, İstanbul 34083, Turkey. 175Department of Psychology,
Marmara University, İstanbul 34722, Turkey. 176Department of Psychology, Üsküdar
University, İstanbul 34662, Turkey. 177Department of Economics and Business, Uni-
versitat Pompeu Fabra, Barcelona 8005, Spain. 178Department of Political Science,
Northeastern University, Boston, MA 2115, USA. 179IRCCS, Santa Lucia Foundation,
Rome 142, Italy. 180Department of Clinical Neuroscience, Division of Psychology,
Karolinska Institutet, Stockholm 171 77, Sweden. 181Penn Center for Neuroaesthet-
ics, University of Pennsylvania, Philadelphia, PA 19104, USA. 182Institute of Psychol-
ogy, University of Silesia in Katowice, Katowice 40-­007, Poland. 183Institute of
Medical Psychology and Behavioral Neurobiology, University of Tuebingen, Tue-
bingen 72076, Germany. 184Dirección Académica, Universidad Nacional de Colom-
bia, Sede de La Paz, Cesar, Colombia. 185HEI-­Lab: Digital Human-­Environment
Interaction Labs, Lusófona University, Lisbon, Portugal. 186Institute of Management
and Organization, Leuphana University Lueneburg, Lueneburg 21337, Germany.
187
School of Business, International University, Vietnam National University HCMC,
Ho Chi Minh City 700000, Vietnam. 188Department of Psychobioloogy and Meth-
odology of Heath Sciences, Universitat Autònima de Barcelona, Barcelona 8193,
Spain. 189Center for Health and Biological Sciences, Mackenzie Presbyterian Uni-
versity, São Paulo 01221-­040, Brazil. 190Department of Psychology, Georgetown
University, Washington DC, 20057, USA. 191Center for Computational Psychiatry,
Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA. 192Faculty of
Medicine, Universidad Cooperativa de Colombia, Villavicencio, Colombia. 193De-
partment of Psychology, Faculty for Social and Educational Sciences, Norwegian
Vlasceanu et al., Sci. Adv. 10, eadj5778 (2024) 7 February 2024
University of Science and Technology, Trondheim 7491, Norway. 194Department of
Psychology and Psychotherapy, Witten/Herdecke University, Witten 58455, Ger-
many. 195Department of Psychology, University of California, Los Angeles, Los An-
geles, CA 90095, USA. 196Department of Psychological and Behavioural Science,
London School of Economics and Political Science, London WC2A 2AE, UK. 197Ja-
pan Society for the Promotion of Science, Tokyo 1020083, Japan. 198Faculty of Sci-
ence and Engineering, Waseda University, Tokyo 1658555, Japan. 199Institute of
Psychology, University of Lausanne, Lausanne 1015, Switzerland. 200Department of
Management, Prague University of Economics and Business, Prague 13067, Czech
Republic. 201Department of People Management and Organization, Universitat Ra-
mon Llull, Esade Business School, Barcelona 8034, Spain. 202School of Economics &
Finance, University of St Andrews, St Andrews KY16 9AJ, UK. 203School of Philo-
sophical, Anthropological and Film Studies, University of St Andrews, St Andrews
KY16 9AJ, UK. 204Department of Communication, Amsterdam School of Communi-
cation Research, University of Amsterdam, Amsterdam 1018WV, Netherlands.
205
Department of Psychology, Widener University, Chester 19013, USA. 206Depart-
ment of Cognitive Psychology, Universität Hamburg, Hamburg 20146, Germany.
207
Department of Nutritional Sciences, University of Vienna, Vienna 1090, Austria.
208
Harvard Business School, Harvard University, Boston, MA 2163, USA. 209Depart-
ment of Social Psychology, Reichman University (RUNI), Herzliya 4610101, Israel.
210
Research Center for Digital Transformation, Leuphana University Lüneburg,
Lüneburg 21335, Germany. 211Department of Biomedicine, Aarhus University, Aar-
hus 8000, Denmark. 212Danish Research Institute of Translational Neuroscience
(DANDRITE), Aarhus University, Aarhus 8000, Denmark. 213Faculty of Psychology
and Educational Sciences, University of Geneva, Geneva 1205, Switzerland. 214Swiss
Center for Affective Sciences, University of Geneva, Geneva 1205, Switzerland.
215
Department of Psychology and Neuroscience, Boston College, Chestnut Hill, MA
2467, USA. 216Faculty of Philosophy, Ss. Cyril and Methodius University in Skopje,
Skopje 1000, Republic of North Macedonia. 217Faculty of Philosophy, University of
Ss. Cyril and Methodius in Trnava, Trnava 917 01, Slovakia. 218School of Medicine
and Psychology, Australian National University, Canberra, ACT 200, Australia. 219De-
partment of Psychology, University of Virginia, Charlottesville, VA 22902, USA.
220
Faculty of Social Sciences/Psychology, Tampere University, Tampere FI-­33014,
Finland. 221Department of Psychology, Pusan National University, Busan 46241, Re-
public of Korea. 222Psychology and Neuroscience, Schiller Institute for Integrated
Science and Society, Boston College, Brighton, MA 2135, USA. 223UQ Business
School, University of Queensland, Brisbane, QLD 4067, Australia. 224Department of
Developmental Psychology and Socialisation, University of Padova, Padua 35131,
Italy. 225School of Business and Management, Queen Mary University of London,
London E1 4NS, UK. 226Institute for the Study of Power, Crime, and Society | Depart-
ment of Law & Criminology, Royal Holloway, University of London, Egham
TW200EX, UK. 227Department of Psychology, Üsküdar University, Istanbul 34664,
Turkey. 228Department of Public and International Affairs, City University of Hong
Kong, Kowloon 999077, Hong Kong. 229School of Psychology, University of Sussex,
Falmer BN19QH, UK. 230Institute of Pedagogy, Faculty of Historical and Pedagogical
Sciences, University of Wroclaw, Wroclaw 50-­120, Poland. 231Research Department,
The South Ural University of Technology, Chelyabinsk 454052, Russia. 232Laborato-
ry of Interdisciplinary Space Studies, School for Environmental and Social Studies,
Tyumen State University, Tyumen 625003, Russia. 233Department of Psycholgy,
Cambridge University, Cambridge CB2 3 EB, UK. 234Department of Medical and
Clinical Psychology, Tilburg University, Tilburg 5037 AB, Netherlands. 235Behav-
ioural Science Institute, Radboud University, Nijmegen 6500 HE, Netherlands.
236
Department of Psychology, University of Latvia, Riga 1083, Latvia. 237Division of
Psychology, Linköping University, Linköping 58183, Sweden. 238Department of
Marketing Communication and Public Relations, Charles University, Prague 11000,
Czech Republic. 239Kenya Marine and Fisheries Research Institute, Kisumu 1881-­
40100, Kenya. 240Department of Sociology, Stanford University, Stanford, CA 94305,
USA. 241Faculty of Philosophy and Social Sciences, Nicolaus Copernicus University,
Toruń 87-­100, Poland. 242Department of Humanities and Social Sciences, Climate
and Energy Policy Research Lab, Indian Institute of Technology Kanpur, Kanpur
208016, India. 243School of Computing, Engineering and Mathematical Sciences, La
Trobe University Melbourne, Melbourne, VIC 3086, Australia. 244Kellogg School of
Management, Northwestern University, Evanston, IL 60208, USA. 245Institute for
Resources, Environment and Sustainability, University of British Columbia, Vancouver,
BC V6T 1Z4, Canada. 246Faculty of Psychology, University of Warsaw, Warsaw 00-­183,
Poland. 247Center for Neural Science, New York University, New York, NY 10003, USA.
*Corresponding author. Email: vlasceanu@​nyu.​edu (M.V.); kimberlycdoell@​gmail.​
com (K.C.D.)
†These authors contributed equally to this work.
‡Present address: Columbia Business School, New York, NY 10027, USA.
Submitted 4 July 2023
Accepted 4 January 2024
Published 7 February 2024
10.1126/sciadv.adj5778
19 of 19
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

D E V E L O P M E N TA L N E U R O S C I E N C E Copyright © 2024 The

Authors, some rights
RNF220-­mediated K63-­linked polyubiquitination reserved; exclusive
licensee American
stabilizes Olig proteins during oligodendroglial Association for the
Advancement of
development and myelination Science. No claim to
original U.S.
Government Works.
Yuwei Li1,2, Li Pear Wan1,2,3, Ning-­Ning Song4, Yu-­Qiang Ding4,5, Shuhua Zhao6, Jianqin Niu7, Distributed under a
Bingyu Mao1,8*, Nengyin Sheng1,3,8*, Pengcheng Ma1* Creative Commons
Attribution License 4.0
Maldevelopment of oligodendroglia underlies neural developmental disorders such as leukodystrophy. Precise (CC BY).
regulation of the activity of specific transcription factors (TFs) by various posttranslational modifications (PTMs) is
required to ensure proper oligodendroglial development and myelination. However, the role of ubiquitination of
these TFs during oligodendroglial development is yet unexplored. Here, we find that RNF220, a known leukodystrophy-­
related E3 ubiquitin ligase, is required for oligodendroglial development. RNF220 depletion in oligodendrocyte
lineage cells impedes oligodendrocyte progenitor cell proliferation, differentiation, and (re)myelination, which con-
sequently leads to learning and memory defects. Mechanistically, RNF220 targets Olig1/2 for K63-­linked polyubiquitination
and stabilization during oligodendroglial development. Furthermore, in a knock-­in mouse model of leukodystrophy-­
related RNF220R365Q mutation, the ubiquitination and stabilization of Olig proteins are deregulated in oligodendroglial
cells. This results in pathomimetic oligodendroglial developmental defects, impaired myelination, and abnormal
behaviors. Together, our evidence provides an alternative insight into PTMs of oligodendroglial TFs and how this
essential process may be implicated in the etiology of leukodystrophy.

INTRODUCTION
Myelination by mature oligodendrocytes (OLs) enables saltatory con-
duction of action potentials and provides long-­term trophic support
for neuronal axons, maintaining neural integrity throughout the central
nervous system (CNS) (1). During development, neural stem cells–
derived OL progenitor cells (OPCs) proliferate to fill up the CNS, differen-
tiate into OLs, and terminally wrap around axons for proper myelination
(2). During these processes, defects in oligodendroglial development
and myelination can cause various neurological disorders, such as
leukodystrophy, which is also a rare genetic disorder (3). Patients with
leukodystrophy can feature cognitive symptoms, intellectual disabilities,
or social dysfunction (4–6). However, knowledge about the etiological
mechanism of leukodystrophy is lacking, especially with regards to
the regulatory factors needed for oligodendroglial development.
Precise and dynamic regulation of transcription factor (TF)
activities is critical for proper oligodendroglial development (7).
Posttranslational modifications (PTMs) are involved in modulating
TF subcellular localization and also the protein interacting com-
plexes. Through this, PTMs are able to regulate specific processes of
oligodendroglial development and regeneration (7). For instance, it
has been reported that acetylation and phosphorylation are required
1
State Key Laboratory of Genetic Resources and Evolution, Kunming Institute of
Zoology, Chinese Academy of Sciences, Kunming 650223, China. 2Kunming College
of Life Science, University of Chinese Academy of Sciences, Kunming 650223, China.
3
Key Laboratory of Animal Models and Human Disease Mechanisms of Yunnan
Province, Kunming Institute of Zoology, Chinese Academy of Sciences, Kunming,
Yunnan 650223, China. 4Department of Laboratory Animal Science, Fudan Uni-
versity, Shanghai 200032, China. 5State Key Laboratory of Medical Neurobiology
and MOE Frontiers Center for Brain Science, Institutes of Brain Science, Fudan Uni-
versity, Shanghai 200032, China. 6First Affiliated Hospital of Kunming Medical Uni-
versity, Kunming 650032, China. 7Department of Histology and Embryology, Third
Military Medical University, Chongqing 400038, China. 8Center for Excellence in
Animal Evolution and Genetics, Chinese Academy of Sciences, Kunming 650223,
China.
*Corresponding author. Email: kunmapch@​mail.​kiz.​ac.​cn (P.M.); shengnengyin@​
mail.​kiz.​ac.​cn (N.S.); mao@​mail.​kiz.​ac.​cn (B.M.)
for Olig1 translocation from the nucleus to the cytoplasm during
oligodendroglial differentiation (8, 9); and dephosphorylation of
Ser147 primes Olig2 for the function of OPC specification via inter-
acting specific cofactors (10). It is known that E3 ubiquitin ligase–
mediated protein homeostasis is crucial for neural development by
influencing the protein stabilities of various master TFs that govern
neurogenesis such as Pax6 and Ascl1 (11, 12), as well as astroglio-
genesis like Ets variant transcriptional factors (ETVs) (13). Hence,
activities of E3 ubiquitin ligases should be precisely regulated during
neural developmental processes. However, it remains elusive whether
and how regulatory TFs are modulated by ubiquitination during
oligodendroglial development.
Among the estimated 600 to 700 human E3 ubiquitin ligase–encoded
genes, 83 genes are mutated and implicated in 70 different types of
neurological diseases (14). From these diseases, one typical neuro-
developmental disorder is hypomyelinating leukodystrophy (HLD)
which is characterized by a primary lack of myelin deposition (3).
Among the 26 types of established HLD recorded in the Online
Mendelian Inheritance in Man (OMIM) database, HLD-­23 is caused
by mutations of the gene RNF220 (RING finger protein 220) (15),
which encodes a RING-­type E3 ubiquitin ligase. Individuals harboring
homozygous R363Q or R365Q mutation of RNF220 have symptoms
of leukodystrophy and corpus callosum agenesis, as well as intellectual
disability (16, 17). However, the physiological roles of RNF220 during
oligodendroglial development and its ability to regulate TFs during
leukodystrophy onset are unclear.
In this study, we found that the E3 ubiquitin ligase RNF220 is
required for oligodendroglial development, and its specific depletion
in OL lineage cells impairs OPC proliferation, differentiation, sub-
sequent myelination, as well as remyelination after demyelinating
injury in the adult mouse brains. Furthermore, we showed that
Olig1 and Olig2 are two direct ubiquitination substrates of RNF220,
and the protein stabilities of these two TFs are maintained by
RNF220 through K63-­linked polyubiquitination. We were able to

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 1 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

replicate leukodystrophy-­like symptoms in a pathomimetic RNF220R365Q referred to as RNF220-­WT), RNF220flox/wt;Olig1-­Cre+/− (RNF220-­cHet),

knock-­in mouse model, which represents deregulated K63 ubiquiti-
nation of Olig proteins and impediments of oligodendroglial differ-
entiation and myelination.
RESULTS
RNF220 is required for oligodendroglial development
and myelination
To examine the expression of RNF220 in OL lineage cells, OPCs
and differentiated OLs were purified with either platelet-­derived
growth factor receptor α–positive (PDGFRα+) or oligodendrocyte
cell surface antigen 4 (O4+) beads, by using astrocyte cell surface
antigen-­2 (ASCA2+) beads purification of astrocytes as a control
(fig. S1A). Quantitative polymerase chain reaction (PCR) (fig. S1B)
and Western blot (WB) (fig. S1C) results indicated that the expres-
sion levels of RNF220 mRNA and protein were relatively constant
in OPC and OL cells. Moreover, immunofluorescence staining on
corpus callosum sections showed that RNF220 was mainly observed in
the cytoplasm of PDGFRα+ OPCs (fig. S1D) and CC1+ OLs
(fig. S1E) while the intensity was much lower in the nucleus.
To analyze the function of RNF220 during oligodendroglial
development, we depleted RNF220 in oligodendroglial lineage cells
by crossing conditional RNF220flox/flox mice (18) with Olig1-­Cre
mice, which is a pan-­OL lineage Cre mouse line (19), and the resultant
three mouse lines, namely, RNF220wt/wt;Olig1-­Cre+/− (hereafter
and RNF220flox/flox;Olig1-­Cre+/− (RNF220-­cKO), were used for subse-
quent studies. The deletion of RNF220 in these mice was confirmed
through checking relative mRNA and protein levels (fig. S2, A and B).
Immunofluorescence staining analyses further showed that RNF220
expression in Sox10+ OL lineage cells of corpus callosum on postnatal
day 21 (P21) was also reduced (fig. S2C). Moreover, all the three lines
of both sexes were born at Mendelian ratios and there was no signifi-
cant difference of survival rate, general appearance, body weight, or
brain weight at adult age (fig. S2, D to F).
As RNF220 mutation is regarded as a pathological cause of leuko-
dystrophy, T2-­weight magnetic resonance imaging (MRI) analysis
was used to examine white matter volume of corpus callosum in the
above three mouse lines. The results showed that the signal intensity
was specifically and significantly increased in adult RNF220-­cKO
mice as compared to RNF220-­WT and RNF220-­cHet, suggesting that
RNF220 deficiency may lead to myelination defects in mouse brains
(Fig. 1A). Furthermore, black-­gold (BG) staining was further used
to examine the axonal fibers of corpus callosum in these mice, where
consistently severe hypomyelination was only found in RNF220-­
cKO mice (Fig. 1B). When transmission electron microscopy (TEM)
imaging was applied to analyze the myelin ultrastructure, although
the number of total axons in the corpus callosum was not changed
(WT: 77.85 ± 16.39 × 104/mm2; cHet: 76.80 ± 15.86 × 104/mm2;
cKO: 73.11 ± 9.05 × 104/mm2; Fig. 1C), the number (WT:
62.34 ± 12.77 × 104/mm2; cHet: 69.58 ± 16.04 × 104/mm2; cKO:

Fig. 1. RNF220 knockout leads to myelination defects in mice. (A) Representative T2-­weighted mouse brain MRI scanning of P60 RNF220-­WT (n = 3), RNF220-­cHet (n = 3),
and RNF220-­cKO (n = 3) mice. The arrows indicate the corpus callosum. (B) Representative images showing myelin staining (BG) in corpus callosum regions of P60 RNF220-­
WT (n = 3), RNF220-­cHet (n = 3), and RNF220-­cKO (n = 3) mice. (C) Electron microscopy images of the corpus callosum transverse sections from P60 RNF220-­WT (n = 8),
RNF220-­cHet (n = 8), and RNF220-­cKO (n = 8) mice with indicated magnifications. Scale bars, 5 μm for upper panels and 2 μm for lower panels. Bar graphs (mean ± SD) show
quantification of the total fiber number and the number or percentage of myelinated axons, and scatterplots show g-­ratios relative to axon diameter. (D) WB analyses of
protein expression of RNF220, MBP, PLP, MAG, and MOG in the cerebral cortex of P21 RNF220-­WT (n = 3), RNF220-­cHet (n = 3), and RNF220-­cKO (n = 3) mice. α-­Tubulin was
used as the internal controls. Bar graphs (mean ± SD) show the relative levels normalized against indicated proteins expression in the respective wild-­type (WT) controls. IB,
immunoblotting. Statistical analyses are compared to respective control with Mann-­Whitney U test with Bonferroni correction. n.s. (not significant), P > 0.05; **P < 0.01.

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 2 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

31.24 ± 10.56 × 104/mm2; Fig. 1C) and percentage (WT: 77.20 ± 5.20%;
cHet: 73.80 ± 7.33%; cKO: 37.90 ± 7.33%; Fig. 1C) of myelinated
axons, and myelin thickness (g-­ratios: WT: 0.68 ± 0.07; cHet:
0.65 ± 0.09; cKO: 0.74 ± 0.06; Fig. 1C) were specifically and
markedly reduced in the corpus callosum of adult RNF220-­cKO
mice. Accordingly, both the mRNA (fig. S2G) and protein levels
(RNF220: WT: 1.00 ± 0.12, cHet: 0.65 ± 0.06, cKO: 0.36 ± 0.05;
myelin basic protein (MBP): WT: 1.00 ± 0.19, cHet: 1.06 ± 0.14,
cKO: 0.35 ± 0.03; proteolipid protein (PLP): WT: 1.00 ± 0.03, cHet:
0.92 ± 0.09, cKO: 0.37 ± 0.06; myelin-­associated glycoprotein (MAG):
WT: 1.00 ± 0.08, cHet: 1.08 ± 0.09, cKO: 0.32 ± 0.06; myelin oligo-
dendrocyte glycoprotein (MOG): WT: 1.00 ± 0.05, cHet: 1.04 ± 0.12,
cKO: 0.38 ± 0.10; Fig. 1D) of myelin-­related molecules, including
MAG, MOG, PLP, and MBP, were significantly decreased in RNF220-­
cKO adult brains. In addition, the myelin ultrastructure in the spinal
cord in these mice was also analyzed by TEM imaging analyses, and
similar results were obtained to show specific impairment of the
axon myelination in adult RNF220-­cKO mice (fig. S3). Together,
these results indicate that RNF220 is required for myelination
in mice.
As RNF220 is highly expressed in OL lineage cells (fig. S1), we
wondered whether the corpus callosum agenesis in RNF220-­cKO
mice can be attributed to the maldevelopment of oligodendroglia
cells. To this end, we first examined OPC proliferation during develop-
ment and found that both 5-­bromo-­2′-­deoxyuridine (BrdU) labeling
and Ki67 staining in the corpus callosum of P3 mice were signifi-
cantly reduced after RNF220 knockout (BrdU+Sox10+ cells: WT:
0.81 ± 0.28 × 102/mm2; cHet: 0.64 ± 0.13 × 102/mm2; cKO:
0.42 ± 0.23 × 102/mm2; Ki67+PDGFRα+ cells: WT: 2.51 ± 0.67 × 102/
mm2; cHet: 2.64 ± 0.95 × 102/mm2; cKO: 1.80 ± 0.55 × 102/mm2;
Fig. 2A). Correspondingly, RNF220 knockdown in MOPC cells
(fig. S4A), which is a mouse OPC cell line, also exhibited a decreased
proliferating rate by analyzing 5-­ethynyl-­2-­deoxyuridine (EdU) in-
corporation (fig. S4B) and the expression of OPC markers (fig. S4C).
Coimmunostaining of the OPC marker Sox10 and the OL marker
CC1 was used to analyze OL differentiation, and we further found
that the number of CC1+Sox10+ cells was significantly less in the
corpus callosum of P21 RNF220-­cKO mice (WT: 6.28 ± 1.32 × 102/
mm2; cHet: 6.80 ± 1.57 × 102/mm2; cKO: 4.94 ± 1.02 × 102/mm2;
Fig. 2B). In addition, when MOPC cells were induced to differentiate
using T3 supplements (fig. S4D), the expression levels of all the
examined myelin-­related genes were down-­regulated after RNF220
was knocked down while were restored by small interfering RNA
(siRNA)–resistant RNF220 coexpression (fig. S4E). It is known that
there is a chronological procedure of olilgodendrolial development
in postnatal mouse brains: before P7 OPCs are mainly proliferated
and after that OL cells are differentiated (20). Therefore, to examine
whether the impaired OL differentiation after RNF220 depletion is a
secondary effect to the early deregulated proliferation, RNF220flox/
flox
mice were crossed with the PDGFRα-­CreER transgenic mouse

Fig. 2. RNF220 knockout impairs oligodendroglial development defects in mice. (A) Immunofluorescence staining assays of BrdU and Sox10, or Ki67 and PDGFRα in
the corpus callosum regions of RNF220-­W T (n = 3), RNF220-­cHet (n = 3), and RNF220-­cKO (n = 3) mice on P3. Bar graphs (mean ± SD) show the quantification of BrdU+Sox10+
and Ki67+PDGFRα+ cells. (B) Immunofluorescence staining assays of CC1 and Sox10 in the corpus callosum regions of RNF220-­W T (n = 3), RNF220-­cHet (n = 3), and RNF220-­
cKO (n = 3) mice on P21. Bar graphs (mean ± SD) show the quantification of CC1+Sox10+ cells. (C) Immunofluorescence costaining of BrdU and Sox10, or Ki67 and PDGRFα
in the corpus callosum of RNF220-­iWT (n = 3) and RNF220-­icKO (n = 3) on P3, and bar graphs (mean ± SD) show the quantification of BrdU+Sox10+ cells, and Ki67+
PDGRFα+ cells. (D) Immunofluorescence staining of CC1 and Sox10 in the corpus callosum of RNF220-­iWT (n = 3) and RNF220-­icKO (n = 3) on P21, and bar graphs
(mean ± SD) show the quantification of CC1+Sox10+ cells. Scale bars, 50 μm. Statistical analyses are compared to respective control with Mann-­Whitney U test with Bonferroni
correction. n.s. (not significant), P > 0.05; *P < 0.05; **P < 0.01.

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 3 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
line, an OPC-­specific tamoxifen-­inducible Cre line (21), to obtain
the three mouse lines (RNF220wt/wt;PDGFRα-­CreER, RNF220-­iWT;
RNF220flox/wt;PDGFRα-­CreER, RNF220-­icHet; RNF220flox/flox;PDGFRα-­
CreER, RNF220-­icKO). The RNF220-­icKO mice, as well as the two
controls, were treated with tamoxifen to induce RNF220 deletion at
different developmental stages. At first, the RNF220-­icKO mice were
treated with tamoxifen on P0 and P1 to induce RNF220 deletion from
birth, and the brain samples were harvested on P3 after BrdU injection
for 2 hours (fig. S5A). Double immunostaining confirmed RNF220
loss in Sox10+ cells in the corpus callosum (fig. S5A), and it was found
that the number of proliferated Sox10+BrdU+ (iWT: 0.74 ± 0.34 × 102/
mm2; icKO: 0.36 ± 0.22 × 102/mm2) or PDGFRα+Ki67+ (iWT:
3.08 ± 0.43 × 102/mm2; icKO: 1.91 ± 0.38 × 102/mm2) OPCs was se-
verely decreased in RNF220-­icKO mice (Fig. 2C), suggesting that
RNF220 is involved in regulating OPC proliferation at the early
developmental stage. Then tamoxifen was administrated to the
RNF220-­icKO mice from P8 to P10, and the brain samples were har-
vested, and OPC proliferation and OL differentiation were examined
on P21 (Fig. 2D and fig. S5B). It was shown that there was no signifi-
cant difference of the number of PDGFRα+Ki67+ OPCs between
the RNF220-­iWT and RNF220-­icKO mice (fig. S5B), while the number
of CC1+Sox10+ OLs was significantly less in the corpus callosum of
RNF220-­icKO brains (iWT: 11.33 ± 1.65 × 102/mm2; icKO:
9.72 ± 1.60 × 102/mm2; Fig. 2D), suggesting a direct differentiation
defect during this later stage. Together, all these results indicate that
RNF220 is involved in both two oligodendroglial developmental
procedures of OPC proliferation and OL differentiation. Together,
these results indicate that RNF220 is involved in regulating both
OPC proliferation and differentiation to OL, and its deficiency impairs
oligodendroglial development and leads to hypomyelination in
mouse brains.
OL lineage–specific RNF220 deficiency impairs learning and
memory functions
It is known that myelin dysfunction has a profound effect on neuro-
logical functions including processing information during high cog-
nition (5), and patients with RNF220 mutations have symptoms of
intellectual disability (16, 17). We next examined whether the RNF220
depletion–induced hypomyelination in RNF220-­cKO mice would
affect neural behaviors, particularly learning and memory. In the novel
object recognition test, RNF220-­cKO mice showed less preference to
the new object as compared to RNF220-­WT and RNF220-­cHet litter-
mates (preference for new object: WT: 67.36 ± 14.75%; cHet:
75.19 ± 12.98%; cKO: 51.09 ± 16.90%; discrimination index: WT:
1.15 ± 1.01; cHet: 1.97 ± 1.62; cKO: 0.02 ± 1.61; Fig. 3A), suggesting
that nonspatial memory is impaired with RNF220 deficiency. During
the three-­chamber test that examines social interaction and memo-
ry, it was found that RNF220-­cKO mice showed a significant preference
for the animated stranger over the inanimate ball to the same extent
as RNF220-­W T and RNF220-­cHet controls (preference: WT:
29.82 ± 14.93% for ball and 70.18 ± 14.93% for stranger 1; cHet:
28.17 ± 10.19% for ball and 71.83 ± 10.19% for stranger 1; cKO:
38.31 ± 18.48% for ball and 61.69 ± 18.48% for stranger 1; social
ratio: WT: 1.38 ± 1.12; cHet: 1.42 ± 0.72; cKO, 0.80 ± 1.25; Fig. 3B).
However, RNF220-­cKO mice had a significant lower preference for
the new stranger over the familiar stranger compared to the two
control littermates (preference: WT: 42.31 ± 15.84% for stranger 1
and 57.69 ± 15.84% for stranger 2; cHet: 36.25 ± 15.35% for stranger
1 and 63.75 ± 15.35% for stranger 2; cKO: 45.61 ± 14.62% for
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
stranger 1 and 54.39 ± 14.62% for stranger 2; social ratio: WT:
0.85 ± 0.86; cHet: 0.75 ± 0.64; cKO: −0.50 ± 0.80; Fig. 3C), suggesting
an impairment of social memory.
The Morris water maze test was used to examine spatial learning
and memory, and it was found that RNF220-­cKO mice took more
time to find the hidden platform on the second to sixth days during
the 8-­day training (escape latency: day 1: WT: 57.20 ± 4.12 s, cHet:
53.27 ± 7.34 s, cKO: 52.44 ± 10.42 s; day 2: WT: 36.80 ± 4.14 s, cHet:
38.66 ± 11.70 s, cKO: 46.58 ± 7.60 s; day 3: WT: 31.48 ± 2.99 s,
cHet: 30.21 ± 6.49 s, cKO: 42.19 ± 5.76 s; day 4: WT: 27.45 ± 1.62 s, cHet:
23.93 ± 4.61 s, cKO: 40.37 ± 6.61 s; day 5: WT: 18.52 ± 2.04 s, cHet:
24.63 ± 4.20 s, cKO: 39.63 ± 3.68 s; day 6: WT: 15.91 ± 4.80 s,
cHet: 23.33 ± 2.98 s, cKO: 29.21 ± 3.86 s; day 7: WT: 14.38 ± 2.96 s,
cHet: 18.71 ± 1.77 s, cKO: 21.00 ± 6.03 s; day 8: WT: 10.94 ± 4.70 s, cHet:
17.11 ± 2.50 s, cKO: 18.06 ± 3.59 s; Fig. 3D), suggesting a higher
escape latency and impairment of spatial learning. One day after the
training trails, spatial memory was tested and it was shown that the
total movement distance of RNF220-­cKO mice was not shorter, even
longer, than the RNF220-­WT or RNF220-­cHet control mice (WT:
534.20 ± 131.90 cm; cHet: 590.70 ± 184.00 cm; cKO: 711.30 ± 101.50 cm;
Fig. 3E), suggesting that RNF220 knockout in OL lineage cells had
no damaging effect on motor function. However, RNF220-­cKO
mice spent significantly less duration in both target quadrants
(WT: 41.35 ± 9.05 s; cHet: 42.56 ± 7.22 s; cKO: 30.70 ± 7.82 s) and
zones (WT: 8.91 ± 2.47 s; cHet: 7.71 ± 1.40 s; cKO: 4.69 ± 1.88 s)
and had a severely decreased cross for target zone (WT: 6.92 ± 2.50;
cHet: 6.71 ± 2.23; cKO: 4.00 ± 1.71) (Fig. 3E). Together, these re-
sults indicate that spatial learning and memory are impaired in the
RNF220-­cKO mice.
We next examined contextual fear memory using the fear condi-
tioning test and found that RNF220-­cKO mice showed reduced
freezing performance in the postshock periods of the second to fifth
trials during conditioning (freezing: preshock: WT: 1.38 ± 0.18%,
cHet: 1.91 ± 0.68%, cKO: 1.83 ± 0.54%; first shock: WT: 1.93 ± 0.53%,
cHet: 2.73 ± 0.22%, cKO: 2.11 ± 0.57%; second shock: WT:
19.19 ± 7.41%, cHet: 24.90 ± 6.89%, cKO: 7.65 ± 3.07%; third shock:
WT: 42.01 ± 6.90%, cHet: 44.59 ± 9.92%, cKO: 30.31 ± 9.03%;
fourth shock: WT: 50.94 ± 10.45%, cHet: 57.83 ± 12.25%, cKO:
41.74 ± 10.69%; fifth shock: WT: 54.78 ± 9.89%, cHet: 56.16 ± 9.64%,
cKO, 45.91 ± 10.48%; Fig. 3F). Moreover, significant reduced freezing
levels were observed in RNF220-­cKO mice from short-­term (1 hour)
to long-­term (6 days) after conditioning (60 min: WT: 62.38 ± 29.94%;
cHet: 61.24 ± 22.54%; cKO: 46.86 ± 18.54%; 1 day: WT: 67.07 ± 21.17%;
cHet: 57.70 ± 22.94%; cKO: 30.24 ± 17.97%; 3 days: WT: 51.66 ± 16.74%;
cHet: 46.06 ± 24.04%; cKO: 24.67 ± 17.73%; and 6 days: WT:
43.47 ± 29.41%; cHet: 36.93 ± 22.11%; cKO: 12.72 ± 12.26%;
Fig. 3G), suggesting that both short-­term and remote memory are
impaired. Collectively, these results indicate that RNF220 deficiency
in OL lineage cells leads to leukodystrophy-­like malfunction of
learning and memory behaviors.
RNF220 is essential for remyelination in adult brains
Although the main signaling cascades and factors involved in the
OL program of developmental myelination are shared by the remye-
lination program in adult (22), our recent work has identified an
OPC-­specific ubiquitin ligase that is only involved in remyelination
(23). We next investigated whether RNF220 is required for OL
regeneration and remyelination in adult mouse brains after demye-
linating injury. To deplete RNF220 in adult OPCs, RNF220-­icKO
4 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. RNF220 deficiency in OL lineage cells causes impairment of learning and memory behaviors. Behavioral tests of novel object recognition (A), three-­chamber
sociability and social novelty (B and C), Morris water maze (D and E) and contextual fear conditioning (F and G) for P60 RNF220-­W T (n = 15), RNF220-­cHet (n = 15), and
RNF220-­cKO (n = 15) mice. (A) Bar graphs (mean ± SD) show the percentage of spending time close to new object and logarithm of discrimination index. (B) Bar graphs
(mean ± SD) show the percentage of spending time close to inanimate ball and animated stranger mouse and logarithm of social ratio. (C) Bar graphs (mean ± SD) show
the percentage of spending time close to new and familiar animated strangers and logarithm of social ratio. (D) and (E) Escape latencies (mean ± SD) to find the platform
throughout the 8-­day learning trials. (E) Spatial memory retrieval of these mice used in (D) was examined when the platform was removed, and bar graphs (mean ± SD)
show total movement distances, time in target quadrant and zone, and cross for target zone. (F) and (G) Percentage of freezing behavior (mean ± SD) across fear conditioning
sessions. (G) Fear memory of these mice used in (F) was tested by exposure to the environment only. Bar graphs (mean ± SD) show percentage of freezing behavior at 60
min, 1 day, 3 days, and 6 days after the contextual fear conditioning. Str, stranger. Statistical analyses are compared to respective control with Mann-­Whitney U test with
Bonferroni correction or unpaired Student’s t test. n.s. (not significant), P > 0.05; *P < 0.05; **P < 0.01.

mice were administered with tamoxifen for 7 days to fully induce
the expected recombination (Fig. 4A), which was reflected by a specific
and marked decrease of RNF220 levels in the adult OPCs of RNF220-­
icKO mice (RNF220+Sox10+ cells: iWT: 6.45 ± 1.28 × 102/mm2;
icHet: 6.27 ± 1.08 × 102/mm2; icKO: 1.78 ± 0.16 × 102/mm2; Fig. 4B).
On the fourth day, l-­α-­lysophosphatidylcholine (LPC) was injected
into the corpus callosum for lesion induction, and the samples were
harvested and examined on the day post lesion 7 (DPL 7) and DPL
21 (Fig. 4A). We first examined the population of OL lineage cells at
lesion sites and found that the numbers of both Sox10+ OPCs on
DPL 7 (iWT: 4.91 ± 5.06 × 102/mm2; icHet: 5.77 ± 1.26 × 102/mm2;
icKO: 4.12 ± 0.46 × 102/mm2; Fig. 4C) and CC1+ matured OLs on
DPL 21 (iWT: 9.50 ± 1.58 × 102/mm2; icHet: 11.44 ± 1.38 × 102/
mm2; icKO: 4.87 ± 1.62 × 102/mm2; Fig. 4C) were decreased signifi-
cantly in the brains of RNF220-­icKO mice, compared to RNF220-­
iWT and RNF220-­icHet littermates. Consistently, it was shown that
on DPL 21 the number (iWT: 76.12 ± 11.30 × 104/mm2; icHet:
68.05 ± 14.31 × 104/mm2; icKO: 31.24 ± 8.05 × 104/mm2; Fig. 4D)
and percentage of remyelinated axons (iWT: 73.47 ± 8.05%; icHet:
71.85 ± 7.84%; icKO: 38.59 ± 6.44%; Fig. 4D) and the thickness of
axonal myelin sheath (g-­ratios: iWT: 0.65 ± 0.07; icHet: 0.65 ± 0.08;
icKO: 0.76 ± 0.08; Fig. 4D) were severely reduced in the corpus callo-
sum of RNF220-­icKO mice relative to mice of the other two geno-
types (Fig. 4D), suggesting an impairment of remyelination process.

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 5 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. RNF220 knockout in OL lineage cells impairs remyelination in adult mouse brains. (A) Schematic diagram of l-­α-­lysophosphatidylcholine (LPC)–induced lesioning
in the corpus callosum regions of RNF220-­iWT, RNF220-­icHet, and RNF220-­icKO mice. (B) Immunofluorescence staining of RNF220 and Sox10 in the corpus callosum of RNF220-­
iWT (n = 3), RNF220-­icHet (n = 3), and RNF220-­icKO (n = 3) mice on DPL 7, and bar graphs (mean ± SD) show the quantification of RNF220+Sox10+ cells. Scale bars, 50 μm.
(C) Immunofluorescence staining of Sox10 (on DPL 7) and CC1 (on DPL 21) in lesion areas of RNF220-­iWT (n = 3), RNF220-­icHet (n = 3), and RNF220-­icKO (n = 3) mice brains,
and bar graphs (mean ± SD) show the quantification of Sox10+ cells and CC1+ cells. Scale bars, 50 μm. (D) Electron microscopy images show the myelinated axons in lesion
areas of RNF220-­iWT (n = 4), RNF220-­icHet (n = 4), and RNF220-­icKO (n = 4) mice brains on DPL 21. Scale bars, 5 μm for upper panels and 2 μm for lower panels. Bar graphs
(mean ± SD) show the quantification of number and percentage of myelinated axons, and scatterplots show g-­ratio relative to axon diameters. Statistical analyses are com-
pared to respective control with Mann-­Whitney U test with Bonferroni correction. n.s. (not significant), P > 0.05; **P < 0.01.

Together, these data indicate that RNF220 plays a crucial role for OL
lineage cell regeneration in the context of white matter injury.
RNF220 targets Olig1 and Olig2 for regulating their
K63-­linked ubiquitination
To identify the underlying mechanism by which RNF220 regulates
OPC proliferation and OL differentiation, we decided to examine
whether and how RNF220 could possibly modulate master TFs of
transcriptional regulatory network during oligodendroglial develop-
ment. To this end, coimmunoprecipitation (co-­IP) assays in human
embryonic kidney (HEK) 293 cells were used to test whether
RNF220 directly interacted with key TFs, including Olig1, Olig2,
Sox10, Nkx2-­2, Hes5, Id2, Id4, Sox5, Sox6, and Ascl1 (24). Our results
showed that among these TFs, only Ascl1, Olig1 and Olig2 were
found to be interacted with RNF220 (fig. S6, A to J). Because neither
the protein level nor the ubiquitination of Ascl1 was affected by
RNF220 depletion in OPCs (fig. S6, K and L), we decided to exclude
Ascl1 in subsequent investigations.
Interaction between RNF220 and the Olig proteins was further
confirmed by the reverse in vitro co-­IP analyses (fig. S6M). In the
forebrain lysates of P7 mice, it was found that RNF220 and Olig proteins
were endogenously interacting (Fig. 5A). These results suggest that
Olig1 and Olig2 are putative regulatory substrates of RNF220 during
oligodendroglial development. As RNF220 is an E3 ubiquitin ligase
involved in neural development (25), we then examined whether
the ubiquitination of Olig proteins is regulated by RNF220. Through
in vitro ubiquitination assays in HEK293 cells, it was found that the
levels of polyubiquitinated Olig1 and Olig2 were both increased
significantly by coexpression of RNF220 (fig. S7A), and this regula-
tion depends on RNF220’s E3 ubiquitin ligase activity because either

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 6 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
RNF220ΔRING or RNF220W539R, which lacked the RING domain or
had the essential site mutation respectively, failed to promote the
polyubiquitination of Olig1 or Olig2 (fig. S7A). Moreover, when the
endogenous Olig proteins were immunoprecipitated from P7 mice
forebrains, their polyubiquitination levels in RNF220-­cKO mice were
markedly decreased compared to RNF220-­WT and RNF220-­cHet
(Olig1: WT: 1.00 ± 0.12, cKO: 0.50 ± 0.11; Olig2: WT: 1.00 ± 0.10,
cKO: 0.48 ± 0.10; Fig. 5B). In addition, the ubiquitinated Olig1 and
Olig2 in adult forebrains were also significantly reduced when
RNF220-­icKO mice were treated with tamoxifen to deplete RNF220
[Olig1: iWT (TAM−): 1.00 ± 0.02, iWT (TAM+): 1.01 ± 0.04, icKO
(TAM−): 0.99 ± 0.03, icKO (TAM+): 0.38 ± 0.04; Olig2: iWT
(TAM−): 1.00 ± 0.03, iWT (TAM+): 0.98 ± 0.03, icKO (TAM−):
0.96 ± 0.06, icKO (TAM+): 0.36 ± 0.05; Fig. 5C]. Together, these
results suggest that Olig1 and Olig2 are direct ubiquitination sub-
strates for the E3 ubiquitin ligase RNF220 in OL lineage cells.
Various types of ubiquitination linkages have distinct regulatory
effects on protein stability and activity (26). To determine the type of
ubiquitin chain added by RNF220 on Olig proteins, different and
specific ubiquitin mutants were used in the ubiquitination assays in
HEK293 cells. It was shown that both Olig1 and Olig2 were only
ubiquitinated by RNF220 when K63-­type ubiquitin was present
(fig. S7B). Consistent with this, when the K63R mutated ubiquitin
was coexpressed, the RNF220 induced polyubiquitination levels of
Olig proteins were fully diminished (Fig. 5D). When an in vitro
ubiquitination assay was carried out using purified proteins, it was
found that RNF220 promoted polyubiquitination of the Olig pro-
teins efficiently in the presence of wild-­type (WT) or K63 ubiquitin
(Fig. 5E). Lysine residues in substrate act as the stereotypical ubiqui-
tination site(s), and there are 8 and 14 lysine residues in mouse
Olig1 and Olig2 proteins, respectively (Fig. 5F). To determine the
responsible ubiquitination site(s), we individually mutated all these
lysines into arginines and then examined RNF220-­mediated ubiqui-
tination of each mutant. It was found that K59/K64 and K69/K262
were required for the polyubiquitination of Olig1 and Olig2, respectively
(fig. S7, C and D). Furthermore, when these lysine residues were
simultaneously mutated into arginines, RNF220 failed to enhance
the polyubiquitination levels of the resulted Olig12KR and Olig22KR
mutants (Fig. 5G), suggesting that these lysine residues are direct
ubiquitination sites. Together, these results indicate that RNF220
regulates Olig1 and Olig2 ubiquitination by adding K63-­linked polyu-
biquitin chains at the lysine sites of K59/K64 and K69/K262, re-
spectively.
RNF220-­mediated ubiquitination of Olig1 and Olig2
maintains the protein stability
After determining that RNF220 targets Olig proteins for ubiquitination
during oligodendroglial development, we next examined whether
RNF220 is involved in controlling their protein quality. Unexpectedly,
it was found that the protein levels of both Olig1 and Olig2 were
increased markedly by coexpression of RNF220 in HEK293 cells
(Fig. 6A and fig. S8A). Moreover, this regulation was diminished
when either RNF220 mutants lacking the E3 ligase activity, or Olig1
and Olig2 mutants lacking the essential ubiquitination residues,
were used for the analyses (Fig. 6A and fig. S8A). Similarly, RNF220
knockdown significantly down-­regulated Olig1 and Olig2 protein
levels in MOPC cells, while their mRNA expression was not changed
by RNF220 siRNAs (fig. S8, B and C). Consistent with these results,
the protein levels of both Olig1 and Olig2 were significantly reduced
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
in the respective isolated OPCs or OLs when RNF220 was depleted
in RNF220-­cKO (OPC: RNF220: WT: 1.00 ± 0.18, cHet: 1.07 ± 0.14,
cKO: 0.22 ± 0.07; Olig1: WT: 1.00 ± 0.14, cHet: 0.99 ± 0.11, cKO:
0.46 ± 0.07; Olig2: WT: 1.00 ± 0.19, cHet: 0.98 ± 0.16, cKO,
0.44 ± 0.11; OL: RNF220: WT: 1.00 ± 0.12, cHet: 1.14 ± 0.09, cKO:
0.23 ± 0.06; Olig1: WT: 1.00 ± 0.09, cHet: 1.03 ± 0.10, cKO:
0.38 ± 0.04; Olig2: WT: 1.00 ± 0.11, cHet: 1.03 ± 0.10, cKO:
0.39 ± 0.09; Fig. 6B) or RNF220-­icKO mice [RNF220: iWT (TAM−):
1.00 ± 0.03, iWT (TAM+): 1.01 ± 0.03, icKO (TAM−): 0.97 ± 0.11,
icKO (TAM+): 0.33 ± 0.03; Olig1: iWT (TAM−): 1.00 ± 0.03, iWT
(TAM+): 0.97 ± 0.08, icKO (TAM−): 1.00 ± 0.05, icKO (TAM+):
0.35 ± 0.03; Olig2: iWT (TAM−): 1.00 ± 0.01, iWT (TAM+):
0.95 ± 0.05, icKO (TAM−): 1.04 ± 0.06, icKO (TAM+): 0.34 ± 0.06;
Fig. 6C]. All these results suggest that this RNF220-­mediated K63-­
linked polyubiquitination of Olig1 and Olig2 is involved in promoting
their protein stability. To further confirm this conclusion, the protein
degradation kinetics of endogenous Olig1 and Olig2 were examined
in MOPC cells through cycloheximide chase analysis. It was shown
that overexpressing RNF220 increased the stabilities of Olig proteins
and this modulation was depended on its E3 ubiquitin ligase activity,
as the ligase-­dead mutant lost the regulatory capability (Olig1: Control:
0 hours: 1.00 ± 0.10, 3 hours: 0.87 ± 0.04, 6 hours: 0.72 ± 0.06,
9 hours: 0.41 ± 0.01; RNF220WT: 0 hours: 1.00 ± 0.02, 3 hours:
0.86 ± 0.04, 6 hours: 0.83 ± 0.05, 9 hours, 0.80 ± 0.03; RNF220W539R:
0 hours: 1.00 ± 0.03, 3 hours: 0.79 ± 0.04, 6 hours: 0.58 ± 0.01,
9 hours: 0.47 ± 0.06; Olig2: Control: 0 hours: 1.00 ± 0.03, 3 hours:
0.87 ± 0.04, 6 hours: 0.64 ± 0.02, 9 hours: 0.46 ± 0.05; RNF220WT:
0 hours: 1.00 ± 0.04, 3 hours: 0.91 ± 0.03, 6 hours: 0.82 ± 0.06,
9 hours: 0.76 ± 0.09; RNF220W539R: 0 hours: 1.00 ± 0.03, 3 hours:
0.84 ± 0.06, 6 hours: 0.62 ± 0.05, 9 hours: 0.42 ± 0.01; Fig. 6D).
Conversely, RNF220 knockdown decreased the stabilities of both
Olig1 and Olig2 proteins (fig. S8D). Together, these results indicate
that RNF220 maintains the stability of Olig proteins by regulating
their K63-­linked polyubiquitination in OL lineage cells. To further
examine the RNF220 regulation of Olig1 and Olig2 during oligoden-
droglial development, we coexpressed Olig1 or Olig2 with RNF220
siRNA in MOPC cells and found that Olig2 or Olig1 overexpression
rescued the expression levels of OPC markers (fig. S9A) or OL
markers after T3-­ induced differentiation respectively (fig. S9B),
while it had no effect on RNF220 expression, suggesting that Oligo
proteins could reverse RNF220 loss-­caused impairment of the pro-
liferation and differentiation of oligodendroglial cells. This indicates
that they are the main downstream effectors of RNF220 during oligo-
dendroglial development.
Leukodystrophy-­related mutations impair RNF220
regulation on Olig proteins stabilization, oligodendroglial
development, and myelination
As the phenotypes observed in RNF220-­cKO mice, including corpus
callosum agenesis and impairment of learning and memory, are closely
related to clinical symptoms of patients harboring RNF220 missense
mutations (16, 17), we wondered whether our findings may demonstrate
an underlying pathological mechanism of leukodystrophy. Toward this
goal, we first tried to investigate the effect of leukodystrophy-­related
mutations, RNF220R363Q and RNF220R365Q, on Olig protein regulation.
Co-­IP analyses results showed that the two RNF220 mutations im-
paired its binding affinity with Olig1 and Olig2 (fig. S10A).
Moreover, the capability to promote the polyubiquitination of Olig
proteins was markedly reduced by either of these two missense variants
7 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. RNF220 regulates K63-­linked polyubiquitination of Olig1 and Olig2. (A) Co-­IP analyses of the interaction between endogenous RNF220 and Olig1 or Olig2 in
forebrains with the indicated antibodies. IgG, immunoglobulin G. (B) Ubiquitination analyses of polyubiquitinated Olig1 and Olig2 in the forebrains of P7 RNF220-­W T
(n = 3) and RNF220-­cKO (n = 3) mice, and bar graphs (mean ± SD) show normalized polyubiquitination levels against respective expression in the WT controls. (C) Analyses
of ubiquitination levels of Olig1 and Olig2 in the forebrains of P60 RNF220-­iWT (n = 3) and RNF220-­icKO (n = 3) mice with tamoxifen treatment or not, and bar graphs
(mean ± SD) show normalized polyubiquitination levels against respective expression in the WT controls without tamoxifen treatment. (D) Ubiquitination analyses of
RNF220-­mediated polyubiquitination of Olig1 and Olig2 in the presence of WT or indicated ubiquitin mutants in HEK293 cells. HA, hemagglutinin. (E) Ubiquitination
analyses of RNF220-­mediated polyubiquitination of Olig1 and Olig2 with purified proteins as indicated in vitro. (F) Schematic diagram showing all lysine sites in mouse
Olig1 and Olig2 proteins and the lysine sites targeted by RNF220 are highlighted in red. (G) Ubiquitination analyses of RNF220-­mediated polyubiquitination of WT or
Olig1/2 mutants in HEK293 cells as indicated. IP, immunoprecipitation; WCL, whole-­cell lysate; TAM, tamoxifen. Statistical analyses are compared to respective control with
Mann-­Whitney U test with Bonferroni correction in (D) and unpaired Student’s t test in (F). n.s. (not significant), P > 0.05; **P < 0.01.

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 8 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. RNF220 maintains Olig1 and Olig2 proteins stabilization through regulating their ubiquitination. (A) WB analyses of protein expression of WT or Olig1/2
mutants when coexpressed with WT RNF220 or mutants lacking ubiquitin E3 ligase activity in HEK293 cells. (B) WB analyses of endogenous protein expression of RNF220,
Olig1, and Olig2 in OPC and OL cells, isolated as in fig. S1A, from the brains of P7 RNF220-­W T (n = 3), RNF220-­cHet (n = 3), and RNF220-­cKO (n = 3) mice, and bar graphs
(mean ± SD) show normalized levels against respective protein expression in the WT controls. (C) WB analyses of endogenous protein expression of RNF220, Olig1, and
Olig2 in isolated OPC from the brains of P60 RNF220-­iWT (n = 3) and RNF220-­icKO (n = 3) mice with tamoxifen treatment or not, and bar graphs (mean ± SD) show normalized
levels against respective protein expression from the WT control without tamoxifen treatment. (D) Cycloheximide chase analyses of protein half-­lives of endogenous Olig1
and Olig2 in MOPC cells overexpressing WT RNF220 or RNF220W539R mutant lacking E3 ligase activity, and broken line graphs (mean ± SD) show normalized levels against
respective protein expression from the control without cycloheximide treatment. ΔR, ΔRING; CHX, cycloheximide. Statistical analyses are compared to respective control
with Mann-­Whitney U test with Bonferroni correction. n.s. (not significant), P > 0.05; *P < 0.05; **P < 0.01.

(Fig. 7A and fig. S10B). Furthermore, both the two mutants failed to
increase protein levels of Olig1 and Olig2 when coexpressed in HEK293
cells (fig. S10C). In addition, when overexpressed in MOPC cells, both
RNF220R363Q and RNF220R365Q lost the ability to increase the pro-
tein levels of endogenous Olig1 and Olig2 (Fig. 7B and fig. S10D).
Together, these results indicate that leukodystrophy-­related RNF220
mutations impair its regulation of ubiquitination of Olig1 and Olig2
and their protein stability.
To study the RNF220 neuropathological mutations on oligodendroglial
differentiation and myelination, we generated a RNF220R365Q knock-­in
mouse model through CRISPR-­Cas9 technology (fig. S11A). The cross
resulted in WT, heterozygous, and homozygous mice, which we will
hereafter refer to as RNF220-­RR, RNF220-­RQ, and RNF220-­QQ, respec-
tively. It was found that both the mRNA and protein levels of RNF220 in
forebrain OPC and OL cells were comparable among the three mouse
lines (fig. S11, B and C). There was no difference in fertility and survival
rate among these three genotypes. Moreover, they showed no significant
difference in appearance, body weight, brain weight, and structures at
adult age (fig. S11, D to F).
T2-­weighted MRI imaging analysis was first used to study the
myelin deposition in adult brains of the three mice, and it was shown
that the signal intensity of the corpus callosum in RNF220-­QQ mice,
but not in RNF220-­RQ, was higher than that in the RNF220-­RR con-
trols (Fig. 7C), suggesting that there is corpus callosum agenesis in

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 9 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Leukodystrophy-­related mutations impair RNF220 regulation on Olig protein stabilization, oligodendroglial differentiation, and myelination. (A) Ubiq-
uitination analyses of polyubiquitinated Olig1 and Olig2 when coexpressed with WT RNF220 or leukodystrophy-­related missense mutants, RNF220R363Q and RNF220R365Q,
in HEK293 cells. (B) WB analyses of endogenous Olig1 and Olig2 expression in MOPC cells overexpressing WT RNF220 or leukodystrophy-­related missense mutants, RN-
F220R363Q and RNF220R365Q. (C) Representative T2-­weighted mouse brain MRI scanning of P60 RNF220-­RR (n = 3), RNF220-­RQ (n = 3), and RNF220-­QQ (n = 3) mice. The
arrows indicate the corpus callosum. (D) Representative images showing myelin staining (BG) in corpus callosum regions of P60 RNF220-­RR (n = 3), RNF220-­RQ (n = 3), and
RNF220-­QQ (n = 3) mice. (E) Electron microscopy images of the corpus callosum transverse sections from P60 RNF220-­RR (n = 4), RNF220-­RQ (n = 4), and RNF220-­QQ (n = 4)
mice. Scale bars, 2 μm. Bar graphs (mean ± SD) show quantification of the number and percentage of myelinated axons, and scatterplots show g-­ratio relative to axon
diameter. (F) Immunofluorescence staining assays of BrdU and Sox10 in the corpus callosum regions of P3 RNF220-­RR (n = 3), RNF220-­RQ (n = 3), and RNF220-­QQ (n = 3)
mice. Scale bars, 50 μm. (G) Bar graphs (mean ± SD) show the quantification of BrdU+Sox10+ cells, Ki67+PDGFRα+ cells on P3, and CC1+Sox10+ cells on P21. Statistical
analyses are compared to respective control with Mann-­Whitney U test with Bonferroni correction or unpaired Student’s t test. n.s. (not significant), P > 0.05; **P < 0.01.

this disease mouse model. Consistent with this finding, BG staining
and TEM analyses showed that the axonal fibers of corpus callosum
in RNF220-­Q Q mice were severely hypomyelinated (the number
of myelinated axons: RR: 65.92 ± 11.07 × 104/mm2; RQ: 81.05 ± 14.27
× 104/mm2; QQ: 33.01 ± 8.44 × 104/mm2; the percentage of myelinated
axons: RR: 84.39 ± 5.52%; RQ: 85.01 ± 6.39%; QQ: 45.44 ± 12.01%;
g-­ratios: RR: 0.67 ± 0.06; RQ: 0.65 ± 0.08; QQ: 0.76 ± 0.07; Fig. 7, D
and E, and fig. S12, A and B). Furthermore, the mRNA and protein
levels of myelin-­related molecules, including MBP, PLP, MAG, and
MOG, were significantly down-­regulated in the forebrains of RNF220-­
QQ mice (fig. S12, C to E). When proliferation of OPC was examined
in the corpus callosum of P3 mice, it was found that the number of
either BrdU+Sox10+ (RR: 0.82 ± 0.25 × 102/mm2; RQ: 0.84 ± 0.33 × 102/
mm2; QQ: 0.54 ± 0.21 × 102/mm2; Fig. 7G) or Ki67+PDGFRα+ (RR:
2.52 ± 0.51 × 102/mm2; RQ: 2.37 ± 0.69 × 102/mm2; QQ: 2.03 ± 0.41
× 102/mm2; Fig. 7G) cells was decreased in the homozygous knock-­in
mice, compared to heterozygous and WT mice (Fig. 7, F and G, and
fig. S12F). Moreover, in P21 mouse corpus callosum, the number of
CC1+Sox10+ mature OLs (RR: 8.13 ± 1.41 × 102/mm2; RQ:
9.21 ± 1.80 × 102/mm2; QQ: 2.36 ± 0.75 × 102/mm2; Fig. 7G) was
also reduced in the RNF220-­QQ mice (Fig. 7G and fig. S12F), implying
an OL differentiation defect. Together, these results indicate that the
oligodendroglial development and myelination are impaired by the
leukodystrophy-­related RNF220 missense mutation.
We then used the above neural behavioral tests to examine the
learning and memory abilities of these three mouse lines. Similar to
RNF220-­cKO mice, RNF220-­QQ mice showed abnormal performances
in tests of new object recognition (preference for new object: RR:
65.10 ± 6.65%; RQ: 69.35 ± 6.24%; QQ: 46.60 ± 7.60%; discrimination
index: RR: 0.92 ± 0.42; RQ: 1.21 ± 0.47; QQ: −0.21 ± 0.47; Fig. 8A),
three-­chamber social interaction (preference: RR: 34.59 ± 6.38% for
ball and 65.41 ± 6.38% for stranger 1; RQ: 37.36 ± 8.61% for ball and

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 10 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

62.64 ± 8.61% for stranger 1; QQ: 40.26 ± 13.51% for ball and
59.74 ± 13.51% for stranger 1; social ratio: RR: 0.94 ± 0.41; RQ:
0.78 ± 0.60; QQ: 0.62 ± 0.86; Fig. 8B and preference: RR:
36.69 ± 9.19% for stranger 1 and 63.31 ± 9.19% for stranger 2; RQ:
30.44 ± 14.71% for stranger 1 and 69.56 ± 14.71% for stranger 2;
QQ: 48.09 ± 8.51% for stranger 1 and 51.91 ± 8.51% for stranger 2;
social ratio: RR: 0.82 ± 0.61; RQ: 1.34 ± 1.01; QQ: 0.12 ± 0.50;
Fig. 8C), Morris water maze (fig. S12, G and H) and fear condition-
ing (fig. S12, I and J), suggesting that the homozygous RNF220R365Q
mouse model has neuropathological cognition deficits. Last, when
the levels of Olig1 and Olig2 were examined in RNF220-­QQ mouse
brains, we found that the polyubiquitination levels of Olig proteins
were remarkably decreased in their forebrains compared with the
WT littermates (Olig1: RR: 1.00 ± 0.23, QQ: 0.35 ± 0.09; Olig2: RR:
1.00 ± 0.18, QQ: 0.35 ± 0.12; Fig. 8D). Consistently, the proteins
levels of both Olig1 and Olig2 in the isolated OPC and OL cells from
P7 RNF220-­QQ mouse forebrains were much lower than that of the
controls (OPC: Olig1: RR: 1.00 ± 0.11, RQ: 1.04 ± 0.14, QQ:
0.33 ± 0.03; Olig2: RR: 1.00 ± 0.13, RQ: 1.09 ± 0.18, QQ: 0.37 ± 0.10;
OL: Olig1: RR: 1.00 ± 0.07, RQ, 0.96 ± 0.14, QQ, 0.35 ± 0.07; Olig2:
RR: 1.00 ± 0.16, RQ: 1.11 ± 0.15, QQ: 0.38 ± 0.17; Fig. 8E), suggest-
ing dysfunction of the maintenance of Olig proteins stability. To-
gether, all these results suggest that RNF220 mutations impair the
regulation of Olig proteins stability and oligodendroglial develop-
ment, which can serve as a potential etiological mechanism for leu-
kodystrophy pathogenesis.
DISCUSSION
In this study, we found that RNF220 promotes K63-­linked polyubiq-
uitination of Olig1 and Olig2, two master regulatory TFs involved in
oligodendroglial development. RNF220 regulated K63-­linked poly-
ubiquitination to maintain Olig1 and Olig2 protein stability, and this
modulation is required for oligodendroglial development and myelin-
ation, as well as remyelination. Moreover, in the pathomimetic
RNF220R365Q knock-­in mouse model, deregulation of ubiquitination
and stabilization of the Olig proteins leads to maldevelopment of
oligodendroglia, corpus callosum agenesis, and leukodystrophy-­
like symptoms.
Olig1 and Olig2 are key TFs of the gene regulatory network that
control oligodendroglial development and myelination in the
brain (24), and their stage-­specific functions heavily depends on
posttranslational modifications, such as phosphorylation and
acetylation. During neural development, Olig2 phosphorylation is
dynamically regulated, and phosphorylation of the triple serine res-
idues at the N terminus is required for neural progenitor prolifera-
tion (27) while its phosphorylation at Ser147 in the basic helix-­loop-­helix
(bHLH) domain is crucial for motor neuron differentiation in spinal
cord (10). Moreover, Ser147 dephosphorylation of Olig2 is required
for later OPC fate determination (10). During OL maturation, Olig1
protein is regulated by acetylation and phosphorylation, which are
essential for its translocation from the nuclear to the cytoplasm and
subsequent developmental processes (8, 9). Here, we find that in
OL lineage cells, Olig proteins have undergone K63-­linked

Fig. 8. Leukodystrophy-­related RNF220 mutation knock-­in mice show impaired regulation of learning and memory behaviors and of Olig proteins. (A to
C) Behavioral tests of novel object recognition (A), three-­chamber sociability, and social novelty [(B) and (C)], for P60 RNF220-­RR (n = 15), RNF220-­RQ (n = 15), and RNF220QQ
(n = 15) mice. (A) Bar graphs (mean ± SD) show the percentage of spending time close to new object and logarithm of discrimination index. (B) Bar graphs (mean ± SD) show
the percentage of spending time close to inanimate ball and animated stranger mouse and logarithm of social ratio. (C) Bar graphs (mean ± SD) show the percentage of
spending time close to new and familiar animated strangers and logarithm of social ratio. (D) Ubiquitination analyses of polyubiquitinated Olig1 and Olig2 in the forebrains
of P7 RNF220-­RR (n = 3) and RNF220-­QQ (n = 3) mice, and bar graphs (mean ± SD) show normalized polyubiquitination levels against respective expression in the WT
controls. (E) WB analyses of endogenous protein expression of Olig1 and Olig2 in OPC and OL cells, isolated as in fig. S1A, from the brains of P7 RNF220-­RR (n = 3), RNF220-­RQ
(n = 3), and RNF220QQ (n = 3) mice, and bar graphs (mean ± SD) show normalized levels against respective protein expression in the WT controls. Statistical analyses are
compared to respective control with Mann-­Whitney U test with Bonferroni correction or unpaired Student’s t test. n.s. (not significant), P > 0.05; **P < 0.01.

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 11 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
polyubiquitination for stabilization, which is involved in regulating
both OPC proliferation and OL differentiation. Moreover, we found
that this ubiquitination has no effect on Olig1 subcellular localiza-
tion as we found that the percentage of HEK293 cells with Olig1
cytoplasm distribution was not changed by overexpressing WT
RNF220 or its E3 ubiquitin ligase–dead mutants (fig. S13). There-
fore, the RNF220-­mediated ubiquitination of Olig1 is an indepen-
dent regulatory process from those modifications of acetylation or
phosphorylation. Although we provide evidence that the protein
levels of Olig1 and Olig2 are enhanced by this K63-­linked ubiquiti-
nation, we cannot be certain whether it is also involved in regulating
the interactome and transcriptional activities of Olig proteins, which
the phosphorylation and SUMOylation for Olig2 is involved in (10,
27, 28). It would be interesting to further study whether there is any
cross-­talk among these PTMs to regulate Olig proteins activities dur-
ing oligodendroglial development and (re)myelination.
There are seven internal lysine residues in the ubiquitin protein,
and accordingly, seven different kinds of polyubiquitin chains can
form, thereby conferring differential fates to a single substrate (29,
30). Both K48 ubiquitination–induced degradation and nonproteo-
lytic K63 ubiquitination play crucial roles during neural development
and for modulating brain functions, as well as in the pathology neuro-
logical diseases (31, 32). As an E3 ubiquitin ligase, RNF220 is capable
of adding such distinct polyubiquitin chains to exert its biological
functions. Our previous studies have found that it modulates K63-­
linked ubiquitination of Gli protein to fine-­tune Shh signaling activity
during spinal cord patterning (18) as well as K48-­linked ubiquitination
of α-­amino-­3-­hydroxy-­5-­methyl-­4-­isoxazole propionic acid type
glutamate receptors (AMPARs) to regulate synaptic transmission
and plasticity in forebrain excitatory neurons (33). Here, we further
uncover a novel role in OL lineage cells where RNF220 regulates
K63-­linked ubiquitination of Olig proteins to promote their stabiliza-
tion for (re)myelination. However, it is unknown how RNF220
achieves such specificity in different neural cells. As the Olig protein
levels are decreased markedly when RNF220 is lacking, further
efforts are also needed to find out the underlying E3 ubiquitin
ligase(s) regulating their K48-­linked ubiquitination and protein
degradation. Moreover, it would be interesting to determine how
the different activities of these E3 ubiquitin ligases are balanced to
bidirectionally modulate Olig1 and Olig2 expression levels for en-
suring normal oligodendroglial development and (re)myelination.
The E3 ubiquitin ligase–mediated ubiquitination is a fundamental
PTM and participates in various processes of neural development,
from neural progenitor proliferation and differentiation to synapse
formation and circuit connection (32, 34). Most of our knowledge
about the pathological roles of ubiquitination-­controlled proteins
homeostasis is based on the identified E3 ubiquitin ligase mutations
in neurological disorders. Of the known and putative E3 ubiquitin
ligase genes in human genome, around 13% have been found to be
mutated in common and rare neurological disorders, several of
which are attributed to deregulation of neural development, includ-
ing autism spectrum disorder, Angelman syndrome, and Gordon
Holmes syndrome (14). However, among these mutated E3 ubiqui-
tin ligases, only a few have been extensively studied to understand
their physiological roles and pathological mechanisms, such as in
UBE3A, PARKIN, etc. (14). HLD is a developmental disease caused
by deregulation of oligodendroglial development and myelination.
Among the identified HLD-­related genes in the OMIM database,
only RNF220 and VPS11 may encode putative E3 ubiquitin ligases
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
(15). VPS11 is a subunit of the HOPS/CORVET complex for endo-
somal fusion, and recently, it has been reported to function as an E3
ubiquitin ligase involved in fine-­tuning various signaling pathways,
including Wnt, estrogen receptor α, and nuclear factor κB (35). Al-
though a recent study has found that VPS11 is expressed in mature
OLs and may be involved in myelination (36), more efforts are needed
to investigate whether the VPS11-­related HLD depends on its ubiq-
uitin ligase activity. In this study, we present ample evidence that
dysregulation of RNF220-­regulated ubiquitination and stabilization
of Olig proteins are implicated in the pathology of leukodystrophy.
Furthermore, we and other groups have recently found several E3
ubiquitin ligases, including Cbl, Nedd4, and RNF43, that are in-
volved in oligodendroglial differentiation and (re)myelination (23,
37, 38), though genetic epidemiology studies have so far not
found any pathological mutation in these genes in patients with
HLD. The prevalence of E3 ubiquitin ligase–mediated proteos-
tasis in OPC and OL cells suggests that uncovering the underly-
ing regulatory network of ubiquitination will be helpful in
furthering our understanding of the etiology of leukodystrophy.
Although the RNF220-­QQ mouse is a global knock-­in model
that faithfully mimics the scenario in patients, RNF220-­QQ animals
do not recapitulate all the features in patient homozygous for RNF220
mutations. For example, RNF220-­QQ mice showed no significant
difference in survival and walking performance during our 12-­month
observation, while most affected patients suffered from severe ataxia
during their early teenage years and died at their second teenage
years (16, 17). As it was shown that the total swimming distance in
RNF220-­cKO (Fig. 3E) or RNF220-­QQ mice (fig. S12H) was not
shorter than their control mice during the Morris water maze tests,
we think that RNF220 might not be required for movement regulation.
However, in our following rotarod test, unexpectedly, it was found
that the RNF220-­cKO and RNF220-­QQ mice exhibited shorter latency
to fall from a uniformly accelerating apparatus compared with the
respective control littermates (fig. S14), suggesting that RNF220
depletion or mutation in oligodendroglial cells indeed impairs mice
motor ability. Therefore, RNF220-­mediated myelination is involved
in regulating fine-­tune motor but not general movement. The differ-
ences in neuropathology between rodent and human, as well as the
complicated motor behaviors we have found, indicate that more
accurate disease mice models are required to dissect the distinct
pathogenic reasons for the complex neuropsychiatry and neuro-
developmental disorders. Moreover, it would be helpful by con-
structing and investigating larger animal model, such as no-­human
primate, of which brain is more to our human being.
MATERIALS AND METHODS
Cell culture and transfection
Human HEK293 cells were grown in Dulbecco’s modified Eagle’s
medium (DMEM) (Gibco, 11965092) supplemented with 10%
fetal bovine serum (Gibico, A5669701), and penicillin-­streptomycin
(100 mg/ml) (Biological Industries, 03-­031-­1B). The Mouse OL pre-
cursor cell (MOPC) line was generated by immortalization of
the primary mouse OPC cells isolated from mouse forebrains
and we obtained it from the Conservation Genetics of CAS Kunming
Cell Bank. MOPCs were cultured as previously described with
minor modifications (39). MOPCs were grown in the basic growth
medium [DMEM plus 0.1% BSA (Solarbio, A8020), apo-­transferrin
(50 μg/ml; Sigma-­Aldrich, T1147), 1 mM sodium pyruvate (Gibico,
12 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
11360070), 30 nM sodium selenite (Sigma-­Aldrich, S5261), 10 nM
D-­
biotin (Sigma-­ Aldrich, 2031), 10 nM hydrocortisone (Sigma-­
Aldrich, 3867), PDGF-­AA (10 ng/ml; Peprotech, 100-­13A), basic fi-
broblast growth factor (10 ng/ml; Proteintech, 100-­18B-­100), and
2% N2 (Life Technologies, 17502-­048)]. For differentiation, T3 (10 ng/
ml; MedChemExpress, HY-­A0070) was included in the growth me-
dium. Both HEK293 and MOPC cells were transfected using
Lipofectamine 2000 (Invitrogen, 11668019) according to the
manufacturer’s instructions.
Animal procedures
All mice with C57BL/6 background were maintained and handled ac-
cording to guidelines approved by the Animal Care and Use Commit-
tee of the Kunming Institute of Zoology, Chinese Academy of Sciences
(IACUC-­PA-­2022-­02-­022). The conditional RNF220 knockout mouse
model RNF220flox/flox was used as previously described (18). Olig1-­Cre
(19) and PDGFRα-­CreER (21) mice were used for the generation of
OL lineage conditional and inducible knockout mouse model, respec-
tively. RNF220R365Q knock-­in founder mouse lines (strain no. T054604)
were generated by NRCMM (National Resource Center for Mutant
Mice，Nanjing University) using CRISPR-­Cas9 technology. In brief,
the sequences 5′-­GATACGGGCCACCACACTCC-­3′ and 5′-­GAAG
CCACCTTCCAGGAGTG-­3′ were chosen as the single guide RNA
targeting genomic DNA sequences. A donor vector containing the
mutation of RNF220R365Q flanked by sequences upstream and down-
stream of the genomic region was used to insert the specific mutation
by homologous recombination.
All genotypes described were confirmed by PCR using genomic
DNA prepared from tail tips. RNF220 floxed allele was genotyped as
previously described (18). The Cre allele was identified by the following
primers: forward, 5′-­GCCTGCATTACCGGTCGATGC-­3′ and re-
verse, 5’-­CAGGGTGTTATAAGCAATCCC-­3′. RNF220 KI allele was
genotyped by real-­time PCR using the following primers: forward,
5′-­ATTGAGTGTATTTCACGGGAGCC-­3′ and reverse, 5′-­CAGC
CTCACAATTCCATTTCCC-­3′, supplemented with taqman probes:
WT: 5′-­(VIC)-­CGGATACGGGCCACCACACTCCTG-­(BHQ1)-­3′,
and KI: 5′-­(FAM)-­CGGATACAGGCCACCACATTGCTG-­(BHQ1)
-­3′.
To induce RNF220 knockout in adult RNF220flox/flox;PDGFRα-­
CreER mice, intragastrical administration of tamoxifen (Sigma-­
Aldrich, T5648) at a dose of 200 mg/kg body weight was given daily
for 7 days continuously. The model of adult injury and remyelination
was performed as described previously (40). LPC (1 μl; Sigma-­
Aldrich, L4219) of 1% solution was injected into the corpus callosum,
at the coordinate: 0.8 mm lateral, 0.8 mm rostral to bregma, 1.2 mm
deep to brain surface, to induce lesion on the fourth day of tamoxifen-­
administration. For BrdU pulse labeling, animals were injected
intraperitoneally with BrdU (500 mg/kg; Sigma-­Aldrich, B5002) per
body weight at 2 hours before sacrifice.
Plasmids and reagents
Ubiquitin constructs were gifts from Ceshi Chen’s lab. Olig2 expression
plasmid was a gift from Y. Li’s laboratory (Shanghai Jiao Tong University
School of Medicine). The construct containing mouse Olig1 cDNA
was purchased from OriGene (MR203390). RNF220 constructs were
used as previously described (18).
To prepare the Olig1 and Olig2 KR mutation constructs, site-­directed
mutagenesis was carried out by PCR-­driven overlap extension using
Pfu DNA polymerase (TianGen, EP101-­02) as previously described
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
(41). The following siRNAs were used to knock down RNF220 in
MOPC cells: control siRNA: siControl, 5’-­TTCTCCGAACGTGTC
ACGT-­3′; siRNAs for RNF220: siRNA-­1#: 5’-­CCTGCAAGAACA
GCGACATTG-­3′, siRNA-­2#: 5’-­AGACTGAAGCACATGTAATA
T-­3′, and siRNA-­3#: 5’-­GGAGTATGGGAAACCACAATA-­3′.
RNA isolation, cDNA synthesis, and quantitive real-­time PCR
Total RNA was isolated using TRIzol reagent (TianGen, DP-­424)
according to the manufacturer’s instructions. Typically, 1 μg of total
RNA was reversely transcribed to cDNA using a first-­strand cDNA
synthesis kit (Fermantas, K1632). Gene expressions were quantified
using LightCycler480 SYBR Green I Master (Roche, 4707516001)
on a LightCycler480 system (Roche). All reactions were run in rep-
licates of at least three samples. The following primers were used:
mouse Actin: forward, 5′-­GCCAACCGTGAAAAGATGAC-­3′ and
reverse, 5′-­GAGGCATACAGGGACAGCAC-­3′; mouse RNF220:
forward, 5′-­GTCTCAGTAGACAAGGACGTTCACA-­3′ and re-
verse, 5′-­GGGGTGGAGGTGTAGTAAGGAAG-­3′; mouse Olig1:
forward, 5′-­TCTTCCACCGCATCCCTTCT-­3′ and reverse, 5′-­CC
GAGTAGGGTAGGATAACTTCG-­3′; mouse Olig2: forward, 5′-­TC
CCCAGAACCCGATGATCTT-­3′ and reverse, 5′-­CGTGGACGAG
GACACAGTC-­3′; mouse MBP: forward, 5′-­GACCATCCAAGAA
GACCCCAC-­3′ and reverse, 5′-­GCCATAATGGGTAGTTCTCG
TGT-­3′; mouse PDGFRα: forward, 5′-­ACACGTTTGAGCTGTC
AACC-­3′ and reverse, 5′-­CCCGACCACACAAGAACAGG-­3′; mouse
GFAP: forward, 5′-­CGGAGACGCATCACCTCTG-­3′ and reverse,
5′-­AGGGAGTGGAGGAGTCATTCG-­3′; mouse PSD95: forward,
5′-­CAAGGATGGCAGGTTGCAGATCGGA-­3′ and reverse, 5′-­TCC
TCATGCATGACATCCTCTAG-­3′; mouse PLP: forward, 5′-­CCAG
AATGTATGGTGTTCTCCC-­3′ and reverse, 5′-­GGCCCATGAGTTT
AAGGACG-­3′; mouse MAG: forward, 5′-­CTGCCGCTGTTTTGGA
TAATGA-­3′ and reverse, 5′-­CATCGGGGAAGTCGAAACGG-­3′;
mouse MOG: forward, 5′-­AGCTGCTTCCTCTCCCTTCTC-­3′ and
reverse, 5′-­ACTAAAGCCCGGATGGGATAC-­3′.
Immunofluorescence analysis and myelin staining
Transverse brain sections of 12-­μm thickness were prepared and
HEK293 cells were seeded on LAB-­TEK chamber slides (Ther-
mo Fisher Scientific, 154453) for immunoflorescence assays as
previously described (18, 42). The following antibodies were
used: anti-­RNF220 (1:200; Sigma-­Aldrich, HPA027578), anti-­
PDGFRα (1:400; R&D, AF1026), anti-­C C1 (1:200; Millipore,
OP80), anti-­S ox10 (1:300; Oasis Biofarm, OB-­PGP001), anti-­
BrdU (1:200; Bio-­R ad, MCA6144), anti-­Ki67 (1:200; Abcam,
ab15580), and fluorescence-­conjugated secondary antibodies,
Alexa Fluor 488/555/594 donkey/goat anti-­ rabbit/mouse/rat/
guinea pig immunoglobulin G (1:400; Invitrogen; A11076, A11055,
A21202, A48269, A21206, A31572). Images were captured us-
ing a light microscope (Olympus, VS120) and analyzed with ImageJ
software.
For myelination analysis, 20-­μm brain coronal sections were pre-
pared with a cryostat microtome (Leica, CM1850UV) and stained
with the Black-­Gold II kit (Biosensis, TR-­100-­BG) according to the
manufacturers’ instructions. Images were captured using a light micro-
scope (Olympus, VS120) and analyzed with ImageJ software.
Electron microscopy
Mice were deeply anesthetized, perfused with cold PBS followed
by 4% paraformaldehyde (Sigma-­Aldrich, 158172). The corpus
13 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

callosum tissues were dissected on ice and cut into small pieces. Mouse behavior tests
Then, pieces of corpus callosum were postfixed with 2.5% glutaral- All the neural behavioral experiments were approved by the Animal
dehyde (Sigma-­Aldrich, G5882) overnight at 4°C; treated with 1% Care and Use Committee of Kunming Institute of Zoology, Chinese
osmium tetroxide (Sigma-­A ldrich, 1.24505), ethanol (Sigma-­ Academy of Sciences. All behavioral experiments were performed in
Aldrich, 459828) dehydrated, and acetone (Sigma-­Aldrich, 270725) the light phase (9:00 a.m. to 6:00 p.m.) in a sound proof room with a
transition; and embedded into epoxy resins (Structure Probe, SPI-­ neutral environment, and individual tests were performed in relatively
Pon 812). Ultrathin sections (60 nm) were stained with 2% uranyl fixed time. Eight-­week-­old mice were used for all tests. The mice were
acetate (Electron Microscopy Sciences, 22400-­2) and lead citrate first given a 60-­min habituation after transferring to the rooms for be-
(Electron Microscopy Sciences, 22410) for electron microscopy im- havioral tests. The experimenter was blind to the group identity of the
aging using JEM-­1400 plus (JEOL). The extent of axonal myelina- tested mice, and the inner surfaces of instruments were cleaned with
tion was quantified by calculating the g-­ratio (the ratio between the 75% alcohol after each test. Mice with different genotypes were con-
inner and the outer diameter of the myelin sheath). ImageJ software ducted in random order.
was used for measurement of the axonal caliber and axonal counting. Novel object recognition test
The novel object recognition test consisted of three phases: habitua-
MRI data acquisition and analysis tion, training, and test. In the habituation phase, the animals were al-
MRI data were obtained on a 7.0-­T Bruker MR system (BioSpec 70/20 lowed to explore an empty arena for 5 min. Then, 24 hours later, during
USR, Bruker). The mice were anaesthetized by inhalation of 3% iso- the training trial, each mouse was individually placed into the arena
flurane (RWD, R510-­22-­10) before scanning, and physiological pa- containing two identical objects (A1 and A2), equidistant from each
rameters were monitored and kept constant during the experiment. other, and allowed to explore the objects for 10 min. After 1 hour, dur-
Tooth and ear bars were used to restrain the mice for imaging. A two-­ ing the test phase, one copy of the familiar object (A3) and a novel
dimensional rapid acquisition with relaxation enhancement sequence object (B) was placed in the same location as during the training trial.
was applied to obtain T2-­weighted images with the following param- The exploring time was recorded when the mouse touched the object
eters: repetition time, 3500 ms; echo time, 45 ms; slice width, 0.6 mm. with the tip of its nose or the front paws in a total time of 5 min. The
discrimination index was calculated as the difference in time exploring
WB, co-­IP, and ubiquitination assays the novel and the familiar objects, expressed as the percentage ratio of
WB, in vitro and in vivo ubiquitination, and co-­IP assays were carried the total time spent exploring both objects.
out as previously described (18, 33, 41). The following primary anti- Three-­chamber sociability and social novelty test
bodies were used for immunoblotting: anti-­Flag (1:5000; Sigma-­ Sociability and social novelty tests were performed on mice as previ-
Aldrich, F7425), anti-­Myc (1:5000; Proteintech, 16286-­1-­AP), ously described with minor modifications (33). Both strangers used
anti-­hemagglutinin (1:5000; Sigma-­Aldrich, H3663), anti-­RNF220 were WT C57BL/6 mice with matched age, body weight, and sex of the
(1:1000; Sigma-­Aldrich, HPA027578), anti-­MBP (1:1000; Sigma-­ mice being tested. The social test apparatus was made of a clear glass box
Aldrich, ab9348), anti-­PDGFRα (1:1000; BD Biosciences, 558774), (60 cm by 40 cm by 30 cm) with three equally divided chambers
anti–glial fibrillary acidic protein (GFAP) (1:1000; Proteintech, (40 cm by 30 cm each). The chambers were interconnected with 5 cm by
60190-­1-­Ig), anti-­PLP (1:1000; Cell Signaling Technology, 85971S), 5 cm openings, which could be opened or closed manually. The inverted
anti-­MOG (1:1000; Proteintech, 12690-­1-­AP), anti-­MAG (1:1000; cylindrical wire cups, which contain the stranger mouse or an object
Proteintech, 14386-­1-­AP), anti-­Olig1 (1:1000; Santa Cruz Biotech- (ball), were 10 cm in height and contained a 10-­cm floor with the metal
nology, SC-­166257), anti-­Olig2 (1:1000; Millipore, MABN50), anti-­ bars spaced 0.8 cm apart. The day before the test, each of the stranger
ubiquitin (1:1000; Santa Cruz Biotechnology, SC-­8017), and mice was habituated inside the inverted wire cups, and each of the test
anti–α-­tubulin (1:5000; Proteintech, 66031-­1-­Ig). Horseradish per- mice was habituated to the apparatus with two empty wire cups inside
oxidase–coupled goat anti-­mouse or rabbit antibodies (Pierce, the box for 15 min. On the test day, during the habituation phase, an
31430 and 31460) were used as secondary antibodies. Chemilu- empty wire cup was placed into the left and right chamber, and the tested
minescence detection was conducted using a chemiluminescent mouse was placed into the center chamber and allowed to explore for
protein detection kit (Thermo Fisher Scientific, 34580) accord- 5 min, with all doors open between chambers. During the sociability test
ing to the manufacturer’s instructions. Immunoblot signals were phase, an unfamiliar mouse (S1) was placed inside the inverted wire cup
quantified by ImageJ software. in one of the side chambers, an object (O) was placed inside the inverted
wire cup on another side chamber, and the test mouse was introduced to
Cell sorting the center chamber with the doors to both side chambers closed. Then,
Isolation of OLs and astrocytes was performed as described previ- the doors between chambers were lifted simultaneously, and the test
ously (43). Mice were deeply anesthetized and then forebrains mouse was allowed to explore all three chambers for 10 min. During the
were dissected and minced. Then, the tissues were incubated social novelty test phase, the test mouse was placed in the central cham-
with PBS containing papain (Worthington, LS003119) and de- ber with all doors closed between chambers. After a novel mouse (S2)
oxyribonuclease I (Thermo Fisher Scientific, EN0521) for 30 min was introduced in the inverted wired cup, replacing the object (O) in one
at 25°C. After digestion, the tissues were triturated using a dounce of the side chambers, the doors between chambers were lifted simultane-
homogenizer and then filtered through a 40-­μm cell strainer. ously, and the test mouse was allowed to explore all three chambers for
Single-­ c ell suspensions were incubated with anti-­ P DGFRα an additional 10 min. Time spent in close proximity to the empty cup
(Miltenyi, 130-­1 01-­5 02), anti-­O 4 (Miltenyi, 130-­0 96-­6 70), or (E1 and E2) or the stranger mice (S1 and S2) or object (O) was analyzed.
anti-­ASCA2 microbeads (Miltenyi, 130-­097-­679) at 4°C for 4 to 6 hours. Morris water maze test
Cells bound to the beads were harvested for total RNA or pro- To assess spatial learning and memory, the Morris water maze
tein extraction. test was performed as described previously (44). Briefly, a blue

Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024 14 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
circular pool of 120 cm in diameter and was divided into four
equal quadrants with two hypothetical crossed lines. The hidden
circular platform located in the middle of the target quadrant
was 10 cm in diameter and submerged 1 cm below the water sur-
face. During the training period, mice were trained to find the
hidden platform over eight consecutive days with three trials
everyday using a semirandom set of start locations, with the
restriction that only one trial each day started from each of the
three different starting positions. If a mouse failed to find the
platform within 60 s, it was picked up and placed on the platform
for 20 s. On the ninth day, 24 hours after the last training session,
the platform was removed, and 60 s was given to each mouse to
search for the platform in the pool, with the starting location op-
posite to the previous position. The movement of the mice was
monitored using Noldus software (EthoVision XT 8.0, Noldus
Technology). Escape latency to find the platform, total distance
moved, average velocity, total distance to platform, and duration
in platform zone were automatically analyzed by the software.
Contextual fear conditioning test
The contextual fear conditioning test was conducted as previ-
ously described with minor modifications (44). Mice were placed
in a box and received five foot shocks (1.2 mA, 2 s) with 2-­min
intershock intervals by FreezeFrame (Coulbourn Instruments).
Freezing behavior was measured by the amount of time mice ex-
hibited freezing behavior during each intershock interval. Then,
mice were placed back in the box (fear context) 30 min, 1 day,
3 days, and 6 days after fear conditioning, and their contextual
freezing behavior was measured for 11 min without any foot
shocks applied.
Rotarod test
Motor coordination was assessed with a rotarod apparatus (Econo-
mex, Columbus Instruments) as described previously (45). Briefly,
for each mouse, testing was done at least three times and the
intratrial interval was about 45 to 60 min. The latency (time and
speed) to fall of each mouse was recorded, and the results from
three trials were averaged as the indicator of motor performance.
Quantification and statistical analysis
Data analysis was carried out using Excel (Microsoft) and Graph-
Pad Prism (GraphPad Software) and data are presented by
mean ± SD. For biochemical analyses and immunostaining as-
says, all experiments were repeated at least three times indepen-
dently. Statistical analyses are compared to respective control
with Mann-­Whitney U test with Bonferroni correction. There is
no sample or data point excluded or omitted for the analysis re-
ported in this study.
Supplementary Materials
This PDF file includes:
Figs. S1 to S14
REFERENCES AND NOTES
1. K. A. Nave, H. B. Werner, Myelination of the nervous system: mechanisms and functions.
Annu. Rev. Cell Dev. Biol. 30, 503–533 (2014).
2. B. Emery, Regulation of oligodendrocyte differentiation and myelination. Science 330,
779–782 (2010).
3. N. I. Wolf, C. Ffrench-­Constant, M. S. van der Knaap, Hypomyelinating leukodystrophies
-­unravelling myelin biology. Nat. Rev. Neurol. 17, 88–103 (2021).
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
4. K. Berry, J. Wang, Q. R. Lu, Epigenetic regulation of oligodendrocyte myelination in
developmental disorders and neurodegenerative diseases. F1000Res. 9, F1000 Faculty
Rev-­105 (2020).
5. W. Xin, J. R. Chan, Myelin plasticity: Sculpting circuits in learning and memory. Nat. Rev.
Neurosci. 21, 682–694 (2020).
6. B. Zhou, Z. Zhu, B. R. Ransom, X. Tong, Oligodendrocyte lineage cells and depression.
Mol. Psychiatry 26, 103–117 (2021).
7. B. Elbaz, B. Popko, Molecular Control of Oligodendrocyte Development. Trends Neurosci.
42, 263–277 (2019).
8. J. Niu, F. Mei, L. Wang, S. Liu, Y. Tian, W. Mo, H. Li, Q. R. Lu, L. Xiao, Phosphorylated olig1
localizes to the cytosol of oligodendrocytes and promotes membrane expansion and
maturation. Glia 60, 1427–1436 (2012).
9. J. Dai, K. K. Bercury, W. Jin, W. B. Macklin, Olig1 acetylation and nuclear export mediate
oligodendrocyte development. J. Neurosci. 35, 15875–15893 (2015).
10. H. Li, J. P. de Faria, P. Andrew, J. Nitarska, W. D. Richardson, Phosphorylation regulates
OLIG2 cofactor choice and the motor neuron-­oligodendrocyte fate switch. Neuron 69,
918–929 (2011).
11. T. C. Tuoc, A. Stoykova, Trim11 modulates the function of neurogenic transcription factor
Pax6 through ubiquitin-­proteosome system. Genes Dev. 22, 1972–1986 (2008).
12. N. Urban, D. L. van den Berg, A. Forget, J. Andersen, J. A. Demmers, C. Hunt, O. Ayrault,
F. Guillemot, Return to quiescence of mouse neural stem cells by degradation of a
proactivation protein. Science 353, 292–295 (2016).
13. K. Newton, D. L. Dugger, A. Sengupta-­Ghosh, R. E. Ferrando, F. Chu, J. Tao, W. Lam,
S. Haller, S. Chan, S. Sa, D. Dunlap, J. Eastham-­Anderson, H. Ngu, J. Hung,
D. M. French, J. D. Webster, B. Bolon, J. Liu, R. Reja, S. Kummerfeld, Y. J. Chen, Z.
Modrusan, J. W. Lewcock, V. M. Dixit, Ubiquitin ligase COP1 coordinates transcriptional
programs that control cell type specification in the developing mouse brain. Proc. Natl.
Acad. Sci. U.S.A. 115, 11244–11249 (2018).
14. A. J. George, Y. C. Hoffiz, A. J. Charles, Y. Zhu, A. M. Mabb, A Comprehensive Atlas of E3
Ubiquitin Ligase Mutations in Neurological Disorders. Front. Genet. 9, 29 (2018).
15. T. Torii, J. Yamauchi, Molecular Pathogenic Mechanisms of Hypomyelinating
Leukodystrophies (HLDs). Neurol. Int. 15, 1155–1173 (2023).
16. H. Yan, T. Kubisiak, K. Gao, J. Xiao, J. Wang, Y. Zhang, Y. Wu, Y. Jiang, M. Burmeister, The
recurrent mutation in RNF220 also causes hypomyelination in China and is a CpG hot
spot. Brain 145, e122–e124 (2022).
17. A. Sferra, P. Fortugno, M. Motta, C. Aiello, S. Petrini, A. Ciolfi, F. Cipressa, I. Moroni,
V. Leuzzi, L. Pieroni, F. Marini, O. Boespflug Tanguy, E. Eymard-­Pierre, F. R. Danti, C.
Compagnucci, G. Zambruno, A. Brusco, F. M. Santorelli, L. Chiapparini, P. Francalanci,
A. L. Loizzo, M. Tartaglia, G. Cestra, E. Bertini, Biallelic mutations in RNF220 cause
laminopathies featuring leukodystrophy, ataxia and deafness. Brain 144, 3020–3035 (2021).
18. P. Ma, N. N. Song, Y. Li, Q. Zhang, L. Zhang, Q. Kong, L. Ma, X. Yang, B. Ren, C. Li, X. Zhao,
Y. Xu, X. Gao, Y. Q. Ding, B. Mao, Fine-­Tuning of Shh/Gli Signaling Gradient by
Non-­proteolytic Ubiquitination during Neural Patterning. Cell Rep. 28, 541–553.e4 (2019).
19. Q. R. Lu, T. Sun, Z. Zhu, N. Ma, M. Garcia, C. D. Stiles, D. H. Rowitch, Common
developmental requirement for Olig function indicates a motor neuron/oligodendrocyte
connection. Cell 109, 75–86 (2002).
20. C. D. Cristobal, H. K. Lee, Development of myelinating glia: An overview. Glia 70,
2237–2259 (2022).
21. S. H. Kang, M. Fukaya, J. K. Yang, J. D. Rothstein, D. E. Bergles, NG2+ CNS glial progenitors
remain committed to the oligodendrocyte lineage in postnatal life and following
neurodegeneration. Neuron 68, 668–681 (2010).
22. V. Gallo, B. Deneen, Glial development: the crossroads of regeneration and repair in the
CNS. Neuron 83, 283–308 (2014).
23. J. Niu, G. Yu, X. Wang, W. Xia, Y. Wang, K. K. Hoi, F. Mei, L. Xiao, J. R. Chan, S. P. J. Fancy,
Oligodendroglial ring finger protein Rnf43 is an essential injury-­specific regulator of
oligodendrocyte maturation. Neuron 109, 3104–3118.e6 (2021).
24. B. Emery, Q. R. Lu, Transcriptional and Epigenetic Regulation of Oligodendrocyte
Development and Myelination in the Central Nervous System. Cold Spring Harb. Perspect.
Biol. 7, a020461 (2015).
25. P. Ma, B. Mao, The many faces of the E3 ubiquitin ligase, RNF220, in neural development
and beyond. Dev. Growth Differ. 64, 98–105 (2022).
26. K. N. Swatek, D. Komander, Ubiquitin modifications. Cell Res. 26, 399–422 (2016).
27. Y. Sun, D. H. Meijer, J. A. Alberta, S. Mehta, M. F. Kane, A. C. Tien, H. Fu, M. A. Petryniak,
G. B. Potter, Z. Liu, J. F. Powers, I. S. Runquist, D. H. Rowitch, C. D. Stiles, Phosphorylation state of
Olig2 regulates proliferation of neural progenitors. Neuron 69, 906–917
(2011).
28. H. Liu, W. Weng, R. Guo, J. Zhou, J. Xue, S. Zhong, J. Cheng, M. X. Zhu, S. J. Pan, Y. Li, Olig2
SUMOylation protects against genotoxic damage response by antagonizing p53 gene
targeting. Cell Death Differ. 27, 3146–3161 (2020).
29. Y. Kulathu, D. Komander, Atypical ubiquitylation -­the unexplored world of polyubiquitin
beyond Lys48 and Lys63 linkages. Nat. Rev. Mol. Cell Biol. 13, 508–523 (2012).
15 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
30. K. Husnjak, I. Dikic, Ubiquitin-­binding proteins: decoders of ubiquitin-­mediated cellular
functions. Annu. Rev. Biochem. 81, 291–322 (2012).
31. A. Zajicek, W. D. Yao, Remodeling without destruction: Non-­proteolytic ubiquitin chains
in neural function and brain disorders. Mol. Psychiatry 26, 247–264 (2021).
32. A. Upadhyay, V. Joshi, A. Amanullah, R. Mishra, N. Arora, A. Prasad, A. Mishra, E3 Ubiquitin
Ligases Neurobiological Mechanisms: Development to Degeneration. Front. Mol.
Neurosci. 10, 151 (2017).
33. P. Ma, L. P. Wan, Y. Li, C. H. He, N. N. Song, S. Zhao, H. Wang, Y. Q. Ding, B. Mao, N. Sheng,
RNF220 is an E3 ubiquitin ligase for AMPA receptors to regulate synaptic transmission.
Sci. Adv. 8, eabq4736 (2022).
34. H. Kawabe, N. Brose, The role of ubiquitylation in nerve cell development. Nat. Rev.
Neurosci. 12, 251–268 (2011).
35. G. Segala, M. A. Bennesch, N. M. Ghahhari, D. P. Pandey, P. C. Echeverria, F. Karch,
R. K. Maeda, D. Picard, Vps11 and Vps18 of Vps-­C membrane traffic complexes are E3
ubiquitin ligases and fine-­tune signalling. Nat. Commun. 10, 1833 (2019).
36. R. P. Skoff, D. Bessert, S. Banerjee, X. Luo, R. Thummel, Characterization of the Expression
of Vacuolar Protein Sorting 11 (Vps11) in Mammalian Oligodendrocytes. ASN Neuro 13,
175909142110098 (2021).
37. X. Ding, J. Jo, C. Y. Wang, C. D. Cristobal, Z. Zuo, Q. Ye, M. Wirianto, A. Lindeke-­M yers,
J. M. Choi, C. A. Mohila, H. Kawabe, S. Y. Jung, H. J. Bellen, S. H. Yoo, H. K. Lee, The
Daam2-­VHL-­Nedd4 axis governs developmental and regenerative oligodendrocyte
differentiation. Genes Dev. 34, 1177–1189 (2020).
38. S. Calabretta, G. Vogel, Z. Yu, K. Choquet, L. Darbelli, T. B. Nicholson, C. L. Kleinman,
S. Richard, Loss of PRMT5 Promotes PDGFRα Degradation during Oligodendrocyte
Differentiation and Myelination. Dev. Cell 46, 426–440.e5 (2018).
39. R. Wu, A. Li, B. Sun, J. G. Sun, J. Zhang, T. Zhang, Y. Chen, Y. Xiao, Y. Gao, Q. Zhang, J. Ma,
X. Yang, Y. Liao, W. Y. Lai, X. Qi, S. Wang, Y. Shu, H. L. Wang, F. Wang, Y. G. Yang, Z. Yuan,
A novel m6 A reader Prrc2a controls oligodendroglial specification and myelination. Cell
Res. 29, 23–41 (2019).
40. M. Zhang, J. Wang, K. Zhang, G. Lu, Y. Liu, K. Ren, W. Wang, D. Xin, L. Xu, H. Mao, J. Xing,
X. Gao, W. Jin, K. Berry, K. Mikoshiba, S. Wu, Q. R. Lu, X. Zhao, Ten-­eleven translocation 1
mediated-­DNA hydroxymethylation is required for myelination and remyelination in the
mouse brain. Nat. Commun. 12, 5091 (2021).
41. P. Ma, X. Yang, Q. Kong, C. Li, S. Yang, Y. Li, B. Mao, The ubiquitin ligase RNF220 enhances
canonical Wnt signaling through USP7-­mediated deubiquitination of β-­Catenin. Mol. Cell.
Biol. 34, 4355–4366 (2014).
42. P. Ma, N. N. Song, X. Cheng, L. Zhu, Q. Zhang, L. L. Zhang, X. Yang, H. Wang, Q. Kong, D. Shi,
Y. Q. Ding, B. Mao, ZC4H2 stabilizes RNF220 to pattern ventral spinal cord through
modulating Shh/Gli signaling. J. Mol. Cell Biol. 12, 337–344 (2020).
Li et al., Sci. Adv. 10, eadk3931 (2024) 7 February 2024
43. D. B. Swartzlander, N. E. Propson, E. R. Roy, T. Saito, T. Saido, B. Wang, H. Zheng, Concurrent
cell type-­specific isolation and profiling of mouse brains in inflammation and Alzheimer's
disease. JCI Insight 3, e121109 (2018).
44. J. X. Dai, H. L. Han, M. Tian, J. Cao, J. B. Xiu, N. N. Song, Y. Huang, T. L. Xu, Y. Q. Ding,
L. Xu, Enhanced contextual fear memory in central serotonin-­deficient mice. Proc. Natl.
Acad. Sci. U.S.A. 105, 11981–11986 (2008).
45. Q. Cui, H. Bi, Z. Lv, Q. Wu, J. Hua, B. Gu, C. Huo, M. Tang, Y. Chen, C. Chen, S. Chen,
X. Zhang, Z. Wu, Z. Lao, N. Sheng, C. Shen, Y. Zhang, Z. Y. Wu, Z. Jin, P. Yang, H. Liu,
J. Li, G. Bai, Diverse CMT2 neuropathies are linked to aberrant G3BP interactions in stress
granules. Cell 186, 803–820.e25 (2023).
Acknowledgments: We thank C. Yi for discussion of and comments on the manuscript. We are
grateful to all members of the Sheng and Mao laboratories for discussion of this work. We would like
to thank the Institutional Center for Shared Technologies and Facilities of Kunming Institute of
Zoology (KIZ), Chinese Academy of Sciences (CAS) for providing us with VS120 FV1000 confocal and
JEM-­1400 plus electron microscopes and the Service Center for Bioactivity Screening of Kunming
Institute of Botany (KIB), CAS for providing us with CryoStar NX50 freezing microtome. We thank Y.
Guo and C. Li (Institutional Center for Shared Technologies and Facilities of KIZ, CAS), and G. Zhao
(Genome Center of Biodiversity, KIZ, CAS) for technical supports. Funding: This work was supported
by the National Key Research and Development Program of China (2023YFA1800500 to N.S. and P.M.
and 2021YFF0702700 to B.M.), the National Natural Science Foundation of China (32170965 to P.M.,
82060292 to S.Z., and 32371017 to N.S.), Yunnan Fundamental Research Projects (202205 AC160065,
202201AW070009, and 202301AS070059 to P.M. and 202001AS070036 to B.M.), and Science and
Technology Department of Yunnan Province (202305AH340007). P.M. was supported by the Youth
Innovation Promotion Association of Chinese Academy of Sciences. Author contributions:
Conceptualization: P.M., N.S., Y.-­Q.D., B.M., and Y.L. Methodology: P.M., S.Z., J.N., N.S., B.M., and Y.L.
Investigation: P.M., Y.L., and L.P.W. Supervision: P.M., N.S., Y.-­Q.D., B.M., and Y.L. Writing—original draft:
P.M., N.S., Y.-­Q.D., and Y.L. Writing—review and editing: P.M., J.N., N.S., Y.-­Q.D., B.M., and Y.L. Resources:
P.M., J.N., N.S., B.M., Y.L., and L.P.W. Funding acquisition: P.M., S.Z., N.S., and B.M. Validation: P.M., N.S.,
B.M., N.-­N.S., Y.L., and L.P.W. Formal analysis: P.M., N.S., N.-­N.S., and Y.L. Project administration: P.M., N.S.,
B.M., and Y.L. Visualization: P.M., N.S., B.M., and Y.L. Data curation: N.S., B.M., Y.L., and L.P.W. Competing
interests: The authors declare that they have no competing interests. Data and materials
availability: All data needed to evaluate the conclusions in the paper are present in the paper and/
or the Supplementary Materials.
Submitted 18 August 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk3931
16 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

GEOPHYSICS Copyright © 2024 The

Authors, some rights
Toward a near real-­time magma ascent monitoring by reserved; exclusive
licensee American
combined fluid inclusion barometry and Association for the
Advancement of
ongoing seismicity Science. No claim to
original U.S.
Government Works.
Vittorio Zanon1*, Luca D’Auria2,3, Federica Schiavi4, Klaudia Cyrzan1, Matthew J. Pankhurst2,3 Distributed under a
Creative Commons
Fluid inclusion microthermometry on olivines, clinopyroxenes, and amphiboles was used during a volcanic erup- Attribution
tion, in combination with real-­time seismic data and rapid petrographic observations, for petrological monitoring License 4.0 (CC BY).
purposes. By applying this approach to the study of 18 volcanic samples collected during the eruption of Tajogaite
volcano on La Palma Island (Canary Islands) in 2021, changes in the magma system were identified over time and
space. Magma batches with distinct petrographic and geochemical characteristics emerged from source zones
whose depth progressively increased from 27 to 31 kilometers. The rise of magma of deeper origin is attested by
fluid inclusions made of N2 and CO, markers of mantle outgassing. Magma accumulation occurred over different
durations at depths of 22 to 27 and 4 to 16 kilometers. Time-­integrated magma ascent velocities (including pond-
ing times) were estimated at between 0.01 and 0.1 meters per second. This method is cost-­effective and quickly
identifies changes in the magma system during an eruption, enhancing petrological monitoring procedures.

INTRODUCTION subsequent improvements (13), has an uncertainty range of up to

Geophysical monitoring of an active magma system informs us of
the depths at which the magma is moving (1, 2) but does not provide
information about the source of the magmas and/or where they
were stored. Moreover, the geophysical data do not reveal the in-
volvement of multiple magmas of different compositions.
Integrating geophysical information with data from erupted
products is an effective way to address these challenges (3, 4, 5),
tracking the paths of magmas from the source zones to the surface.
To identify the arrival of fresh magma into a magmatic system, fre-
quent analysis of collected volcanic fragments or small lava clots via
rapid microchemical analysis is essential. This protocol is particu-
larly useful for frequently active volcanoes and helps monitor fun-
damental compositional changes in the magmatic system (6). In
contrast, global-­scale investigations aiming to provide detailed tem-
poral and spatial resolution of magmatic processes before and dur-
ing an eruption require more detailed geochemical characterization
(major and trace elements) of whole rock, glassy tephra samples,
and silicate melt inclusions, as well as mineral chemistry and isoto-
pic analyses (5, 7, 8, 9). These conventional methods are unsuitable
for effective monitoring because the generation and interpretation
of petrological and geochemical data demands considerable (pro-
cessing) time. Efforts have recently been made to speed up the ana-
lytical procedures and integrate data from geophysics or gas
geochemistry. This approach has been used during the 2018 Kilauea
eruption (10), the 2021 Fagradalsfjall eruption in the Reykjanes
peninsula of Iceland (11), and the 2021 Tajogaite eruption (5, 12).
However, considerable uncertainties persist in determining magma
ponding depths because mineral-­melt geothermobarometry, spe-
cifically the pyroxene barometer in its original formulation and
1
Instituto de Investigação em Vulcanologia e Avaliação de Riscos (IVAR), Universi-
dade dos Açores, Rua Mãe de Deus, 9500-­123 Ponta Delgada, Portugal. 2Instituto
Tecnológico y de Energías Renovables (ITER), 38600 Granadilla de Abona, Tenerife,
Canary Islands, Spain. 3Instituto Volcanológico de Canarias (INVOLCAN), 38400
Puerto de la Cruz, Tenerife, Canary Islands, Spain. 4Laboratoire Magmas et Volcans,
CNRS, IRD, OPGC, Université Clermont Auvergne, F-­ 63000 Clermont-­ Ferrand,
France.
*Corresponding author. Email: Vittorio.​VZ.​Zanon@​azores.​gov.​pt
140 MPa for calibrated data and up to 370 MPa for uncalibrated data
(14). These pressures correspond to estimated depths of ±4.6 to 6.1
and ±12.2 to 16.1 km, respectively, in a density range from the up-
per crust (~2350 kg m−3) to the mantle (~3100 kg m−3). We evalu-
ated a cost-­effective, rapid approach to identify the movement of
magma beneath the surface during the 2021 Tajogaite eruption at
the Cumbre Vieja volcanic system, La Palma, the northernmost is-
land of the Canary Archipelago situated in the Eastern Atlantic
Ocean. This approach integrates fluid inclusion (FI) microther-
mometry in olivines, clinopyroxenes, and amphiboles with daily
syn-­eruptive seismicity and quick qualitative petrographic observa-
tions on samples erupted from 19 September to 13 December 2021.
FI are droplets of fluid trapped inside crystals either during growth
or later as a result of recrystallization along healed micro-­fractures. A
microthermometric study consists in determining the temperatures
of phase changes during heating and cooling of FI to obtain the baro-
metric state of crystallization and re-­equilibration specific to the crys-
tal under study. This study, when applied to many crystals, allows
constraints to be placed on the dynamics and depth of magma ascent
and ponding. FI microthermometry is often more accurate than
clinopyroxene melt barometry because it can analyze various inclu-
sions in a mineral assemblage from a single sample with a calculated
uncertainty range of 30 to 60 MPa (i.e., ±1.2 to 2.7 km), over a density
range from the Earth’s crust to the mantle, and requires little sample
preparation. FI barometry has been used to model the magmatic sys-
tem of numerous volcanoes (15–23) but has never been applied dur-
ing an eruption. An attempt in this direction was made using Raman
microspectroscopy on a smaller set of FI-­hosting samples collected
after the Tajogaite eruption, showing good correlation with tectonic
signals (24). Microthermometry and Raman spectroscopy comple-
ment each other in characterizing the density and composition of
FI. The accuracy of a microthermometric inquiry is notably higher for
high-­density inclusions (25), even if Raman spectrometers are cor-
rectly calibrated (26). Nonetheless, Raman spectroscopy enables full
characterization of the composition of an inclusion.
In this context, data are extracted over time, with the microther-
mometric data for each sample reflecting integrated information

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 1 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
that relies on ponding depths and ascent velocities of the respective
magma parcel. In principle, it would be possible to access the pre-
eruptive magma storage depths and the dynamics of the magma as-
cent path in real time by collecting microthermometric information
from multiple snapshots throughout the eruption (27).
Eruption, seismicity, and sample description
Eleven distinct seismic swarms of low-­magnitude events (ML < 2)
occurred from October 2017 to early September 2021 (28), affecting
the entire northern and central sectors of the Cumbre Vieja volcanic
system. The hypocentral depths for each seismic swarm generally
varied between 15 and 25 km (Fig. 1A). Contextually, the first two
seismic swarms were accompanied by geochemical anomalies in the
gases emitted at the surface (29). All these elements indicated a pro-
gressive refilling of the system by multiple magma intrusions at dif-
ferent depths. Starting on 11 September, seismic activity continued
with hypocenters initially at a depth of 8 to 9 km and gradually ris-
ing at a constant rate up to about 5 km (Fig. 1B). Approximately 1 day
before the beginning of the eruption, hypocentral depths migrated
rapidly to shallower levels. The latter seismic activity was hydrother-
mal and subsequently vanished or was obstructed by volcanic trem-
ors (30). The eruption began at 14:12 UTC on 19 September and
since then syn-­eruptive seismicity occurred within two different
depth ranges: one at ~18 to 27.5 km (with a mode at 22 km) and
another at ~4 to 14 km (with a mode at 9 km) below the volcano
(Fig. 1, B and C). Deep seismic activity began to decrease in early
December, ~13 days before the cessation of magma emission.
Several volcanic vents opened along a north northwest–south
southeast trending fissure that stretched for 500 m on the western
flank of the Cumbre Vieja ridge at ~950 m above sea level in the area
known as “Cabeza de Vaca.” These vents produced a cinder cone that
gave rise to a composite lava flow for 3 months (Fig. 1D). Addi-
tional information on the volcanological description of the eruption
can be found in (12, 31, 32). A systematic sampling of tephra and
lava was conducted during the eruption for petrological monitoring
purposes (Fig. 1E). The first magma erupted during the event was a
basanite (fig. S1), which contained zoned clinopyroxene, amphibole,
titanomagnetite, and rare olivine. This magma displayed a fine-­
grained groundmass with acicular feldspar. The tephra ejected until
19 October had a tephritic composition (SiO2 = 45 to 47 wt % and
total alkali = 6.6 to 7.9 wt %) and were composed of glassy frag-
ments. The early samples exhibit euhedral amphibole crystals mea-
suring up to 2.5 mm in diameter with reduced breakdown rims
(Fig. 2A), indicating both rapid ascent and equilibrium conditions
with the host magma. Aggregates composed of both amphibole and
clinopyroxene are sometimes observed. As the eruption progressed,
the size of amphibole crystals reduced, and larger reaction rims were
formed. After a period of 7 to 10 days, amphibole crystals progres-
sively disappeared from the mineral assemblage and rarely reap-
peared. Euhedral/subhedral clinopyroxene crystals often several
millimeters in size and with zoned rims are common as isolated
crystals and in polycrystalline aggregates. These crystals also exhibit
notable embayments, irregularly shaped internal cavities, and cor-
roded rims that are indicative of disequilibrium pressure and tem-
perature conditions. Large crystals forming aggregates show
oscillatory or patchy zoning, are twinned, and contain oxides. The
olivine crystals have a euhedral to subhedral shape and are rarely
larger than 1 mm in size (Fig. 2B). Initially, they were scarce, but
their abundance increased as the eruption progressed.
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
On 24 September, there was a decrease in shallow seismic activ-
ity, and, on 27 September, magma emission and volcanic tremors
stopped for about half a day. Following the resumption of magma
emission, more fluid basanites containing zoned clinopyroxene
(Fig. 2C) and olivine in a matrix hosting acicular plagioclase were
emitted. This clinopyroxene is smaller than earlier crystals, shows
oscillatory or patchy zoning, and contains oxides. Skeletal micro-
phenocrysts exhibit normal zoning. The final lava emitted is a ba-
sanite at the boundary with trachybasalt (fig. S1A). These lavas are
chemically similar to those erupted in 1949 from the Llano del Banco
fissure (fig. S1B), aligned with and located only 1.5 km southeast of
the Tajogaite cone vents.
RESULTS
Fluid inclusions
FI are present in all the examined samples. They were found more
frequently in olivine (N = 1355), less frequently in clinopyroxene
(N = 247), and rarely in amphibole (N = 59). Early olivines do not
contain FI.
Trails of FI crossing the crystals in multiple directions represent
the most common texture in all host minerals (Fig. 3A). In olivines,
these trails consist of either rounded or negative-­crystal shaped in-
clusions, which are typically 2 to 4 μm in diameter and lack obvious
evidence of re-­equilibration such as a black rim around the main
cavity or tiny radial cracks (33–35). However, accurate observation
is prevented by the small size of the inclusions, so the possible oc-
currence of a minimal degree of re-­equilibration cannot be ruled
out. These inclusions were trapped after the host mineral formed
during magma ponding events [secondary or late-­stage FI, based on
textural criteria; (36)]. In clinopyroxenes and amphiboles, these
trails consist of inclusions with a small variability in size (10 to 15 μm
across) in each trail. The inclusions have a rounded to slightly elliptical
shape, suggesting some degree of re-­equilibration.
In all host minerals, isolated or clustered inclusions of variable
size (10 to 40 μm across) are less common (Fig. 3B). These inclu-
sions, trapped during crystallization of the host [primary or early-­
stage FI, based on textural criteria; (36)], often show elongated
shapes or a halo of tiny bubbles (<0.5 μm across) surrounding the
main cavity, which indicates the occurrence of re-­equilibration.
FI microthermometry
At room temperature, FI are single-­phase (either liquid or vapor)
or two-­phase (liquid + vapor) (Fig. 3B). Amphibole, clinopyroxene,
and most olivines contain FI with pure CO2 composition and
melting temperature of −­56.6° ± 0.1°C. Liquid water was not
clearly detected either optically or by Raman spectroscopy at
room temperature; however, the presence of a few moles cannot
be excluded (37). Raman spectroscopy revealed the frequent
presence of very small crystals of magnesite in FI found in oliv-
ines. Its presence suggests that some water was originally present
before reacting with the host olivine and fluid CO2 to form a
carbonate (38). It was estimated that the likely original amount
of water dissolved in magmas from intraplate settings is around
10 mol % (39, 40). Thermodynamic modeling suggests that,
upon cooling and at 0.1 GPa, talc would have formed in FI along-
side carbonates from the reaction between forsterite and CO2-­
H2O fluids if XH2O > 0.1 (25). However, talc has not been
observed in Raman spectra.
2 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Seismicity, sampling locations, and timeline of volcanic events of the 2021 eruption. Seismic activity before and during the eruption is illustrated in (A and
B), obtained from the open catalog (28). The change in hypocenter location is evident just before the eruption commences (B). (C) The frequency of seismic events for the
two depth regions. The beginning of the eruption is marked by the blue dashed line, while the two depth ranges discussed within the paper, 4 to 14 km and 14 to 30 km,
are marked by horizontal red lines. Depths are from the sea level. (D) A digital elevation model of La Palma Island highlighting the distribution of the 2021, 1949, and 1585
lavas. The geography of the Canary Islands is presented in the inset. (E) The chronogram of the Tajogaite eruption, illustrating the sample collection (indicated by black
arrows) and the change in magma composition. The red labels indicate lava samples that were used for whole rock analysis, whereas the blue labels represent tephra
samples. On 27 September, the temporary cessation of volcanic activity is marked by a red vertical dashed line.

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 3 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 2. Mineral textures in selected lava samples. (A) Zoned amphibole from
CAN-­LLP-­0045 lava from 25 September, showing poorly developed breakdown
rims. (B) Euhedral/subhedral olivines, also forming an aggregate with small pyrox-
ene and large and partially fragmented clinopyroxene from CAN-­LLP-­0 045.
(C) Zoned euhedral clinopyroxene with FI from CAN-­LLP-­0046 lava from 29 October.
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
Approximately 5% of the FI hosted in olivines that erupted after
21 October had a melting temperature ranging from −­57.4° to
−­57.0° ± 0.1°C. Some of these inclusions got entrapped alongside
chromite (Fig. 3C). The composition of these inclusions, expressed
in mole percentage, consists of ~85 to 88% CO2, ~10% H2O, ~3 to
5% N2, and occasionally 0.9% CO (Fig. 3D; details of the molar pro-
portion quantification procedure are given in Materials and Meth-
ods). Comparable inclusions were discovered in a spinel-­bearing
dunite from Lanzarote (41).
Final inclusion homogenization occurred to the liquid phase
(ThL) in all minerals and to the vapor phase (ThV) in a few olivines.
Table 1 provides a summary of the homogenization temperatures,
densities, and pressures present in the microthermometric database.
The density distribution histograms (Fig. 4) and the textural
characteristics of FI overall reveal a main trapping event of fluids at
depth, followed by a single re-­equilibration event during magma
ponding at a shallower level. It is assumed that no further re-­
equilibration occurred during ascent in the conduit. FI populations
of fluids trapped in amphiboles and pyroxenes are both early and
late-­stage and define unimodal or slightly skewed distributions,
which can be attributed to inclusion stretching in response to over-
pressure, developed during rapid and quasi-­isothermal ascent (42).
Although relatively scarce in quantity, the data are coherent with
each other. Nearly all FI found in amphiboles (ρr = 522 to 735 kg
m−3) are texturally late-­stage, except for a few early-­stage FI with a
density of 636 to 694 kg m−3. Similarly, clinopyroxenes from the ba-
sanites erupted from mid-­October host both early and late-­stage FI
(ρr = 514 to 733 kg m−3).
The fluids trapped in olivine illustrate polymodal density distri-
butions, with two or three primary modes observed on most sam-
pling days (Fig. 4). Most data range between ~550 and 1045 kg m−3.
In this density range, small-­sized late-­stage inclusions with no ap-
parent signs of re-­equilibration were found, and skewed distribu-
tions are limited to plastic deformation and fluid leakage (43).
Although there is a possibility of a minimal extent of re-­equilibration,
these inclusions are still considered good proxies for the original
trapping condition. On the other hand, several larger FI exhibiting
isolated density peaks show signs of re-­equilibration and reduced
densities, ranging between 230 and 500 kg m−3.
Until 19 October, the FI population with densities ~625 to 730 kg
m−3 and the olivine-­hosted FI population with densities of ~850 to
960 kg m−3 have been present, suggesting a well-­established mag-
ma ascent path in terms of conduits, ponding stages, and ascent rate.
The populations of high-­density FI gradually increased over time,
and an intermittently appearing population of very high-­density
inclusions (~995 to 1045 kg m−3) emerged starting from 21 October.
Scattered low-­density peaks at around 240, 300, and 370 kg m−3
appeared intermittently in olivine-­hosted FI from 8 October to
27 November.
The studied samples show no evidence of post-­ eruptive re-­
equilibration in slow cooling lavas. For instance, the density histo-
grams of FI in pyroxenes from 25 September lava and 26 September
tephra look very similar and exhibit a single mode between 650 and
700 kg m−3. The scattered low-­density peaks, ranging from 240 to
370 kg m−3 and associated with early-­stage inclusions, are found in
both lavas and tephra samples emitted from October to late Novem-
ber. The comparison between lava and tephra sampled on the last
day of the eruption shows a reduced number of late-­stage FI re-­
equilibrated to low pressure in the lava sample. This population
4 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. FI textures and Raman data. (A) Late-­stage FI showing negative crystal shape in olivine from tephra erupted on 5 October. (B) Early-­stage FI hosted by olivine
showing the coexistence of vapor and liquid phases at room temperature. (C) Trail of late-­stage FI hosted by olivine from 1 December showing the simultaneous trapping
of fluid and chromite. (D) Raman spectrum of the largest FI shown in (C) containing CO2, N2, and CO. Raman peaks of O2 and N2 from the atmosphere are also shown.

characterizes only the FI present in one olivine from the six ana- high to very high pressure populations of FI in November and De-
lyzed lava samples, suggesting the capture of a remnant of a prior cember olivines contain ∼3 to 5 mol % N2 and occasionally 0.9 mol
magma pulse. % CO (Fig. 3D). In the same samples, other olivines contain FI with
only CO2 + H2O composition and reach a maximum pressure of
740 MPa. Nitrogen is a volatile species that is poorly soluble in an
DISCUSSION oxidized mantle, such as that beneath the Canary Islands (45, 46).
The architecture of the Cumbre Vieja magmatic system Under such conditions, early N2 degassing from a silicate melt is
The barometric information on the magma system was derived enhanced at depth (47–49), leaving silicate melts saturated with CO2
from the isochore distribution in P-­T space. It was assumed that all and H2O. Thus, FI assemblages could track the degassing path of the
FI in clinopyroxenes and amphiboles were either trapped or re-­ magma, with the N2-­bearing population trapped early at depth and
equilibrated at 1075°C, and those in olivines at 1150°C (the reader is simple CO2 + H2O FI trapped at shallower depths.
referred to Materials and Methods for more details). The magma The FI populations trapped or partially re-­equilibrated at ~300 to
ascent was considered rapid enough to be treated as isothermal. It 500 MPa are observable until 27 November (fig. S2), and their
was required to create a stratigraphic model beneath the volcano to modes occasionally do not match. The FI found in amphiboles and
translate the barometric data into depths. Thus, assumptions were pyroxenes, along with some olivines erupted until 24 November, re-
made about the density of the rock bodies and the depth of the tran- turn pressures ranging from 200 to 400 MPa. Last, there are some
sition from crust to mantle (please refer to the Supplementary Ma- sharp modes, scattered between 75 and 180 MPa, which are associ-
terials for more details). ated with early-­stage FI and suggest trapping during brief shallow-­
In early to mid-­October, two stable modes of FI populations were level magma ponding periods.
observed at the ranges of 325 to 420 MPa and 600 to 750 MPa Previous microthermometry-­based studies on FI at Cumbre Vieja
(fig. S2). The high-­pressure range corresponds to the same baromet- volcano found similar barometric intervals (17). This correspon-
ric conditions of amphibole and clinopyroxene crystallization and dence indicates that the method is reliable and reproducible. Our
fractionation found in the magmas from the 1949 eruption (44). At microthermometric database, which is based on a single eruption,
the end of November, the high-­pressure range widened slightly at extends to 865 MPa, which is, however, lower than the value of 1140 MPa
590 to 790 MPa, and a population of very high-­pressure FI (755 to obtained using clinopyroxene-­melt barometry for the 1971 Teneguia
865 MPa) emerged intermittently starting from 21 October. The eruption (50). It is lower than the range of 1.04 to 1.17 GPa found for

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 5 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Table 1. Micrometric database. Data correspond to the day of eruption. Crystals analyzed are olivines (ol), clinopyroxenes (cpx), and amphiboles (amph), and
the total number of crystals analyzed is given between parentheses. Homogenization temperatures to the vapor phase are shown in italics. The reported density
has been corrected for the possible presence of 10 mol % H2O in the inclusion using the method proposed in (40).
ThL range (°C)
Crystals ana-
Sample Data lyzed No. measures ThV range (°C) ρ (kg m−­3) Ttrapping (°C) Pmax (MPa)

CAN-­LLP-­0045 25 September cpx (2) 122 25.6–28.8 664–733 1075 345–424

CAN-­TLP-­0015 26 September cpx (1) 56 28.1–30.3 605–683 1075 286–366
CAN-­TLP-­0022 29 September ol (2) 12 3.8–16.2 855–945 1150 610–729
LPA-­16 4 October amph (2) 3 23.3–30.9 546–768 1075 240–473
cpx (1) 2 30.1–30.9 514–615 1075 215–296
ol (3) 34 13.9–23.7 763–868 1150 493–628
CAN-­TLP-­0037 4 October amph (1) 3 26.1–28.7 665–725 1075 349–419
cpx (1) 21 28.9–30.9 543–661 1075 235–342
ol (1) 29 6.5–11.4 889–926 1150 656–704
CAN-­TLP-­0039 5 October ol (3) 166 1.1–30.9 514–963 1150 227–752
CAN-­TLP-­0052 8 October ol (4) 113 1.0–30.9 514–963 1150 227–752
30.3–30.4 382–388 159–167
CAN-­TLP-­0053 8 October ol (1) 1 14.2 865 1150 624
CAN-­TLP-­0128 19 October amph (2) 53 23.3–30.8 554–768 1075 246–473
cpx (1) 46 27.2–30.9 514–703 1075 215–388
ol (2) 8 12.1–15.9 850–883 1150 603–648
CAN-­LLP-­0043 21 October ol (3) 127 −13.0–30.9 496–1043 1150 650–847
23.0–30.9 230–433 91–179
CAN-­LLP-­0046 29 October ol (4) 210 20.5–30.8 536–803 1150 263–540
24.9–30.8 252–423 101–174
CAN-­TLP-­0214 6 November ol (2) 80 −9.7–30.9 496–1023 1150 215–816
30.8 416 171
CAN-­TLP-­0312 24 November ol (5) 143 1.3–0.8 561–961 1150 264–749
CAN-­TLP-­0340 27 November ol (9) 103 −8.3–24.3 701–1018 1150 409–818
27.7–28.4 269–311 122–129
CAN-­TLP-­0370 2 December ol (3) 78 −3.9–19.8 811–993 1150 551–789
LPA-­17 10 December ol (6) 153 −13.4–30.8 562–1045 1150 264–865
27.8–30.9 298–471 123–199
LPA-­18 10 December ol (1) 98 12.5–21.3 794–880 1150 488–644

the old and eroded volcanic edifices of Cumbre Nueva and Taburiente
(51). This barometric interval recorded by the FI is horizontally
distributed under the entire Cumbre Vieja volcano and also extends
below the nearby island of El Hierro, highlighting its regional im-
portance (17, 52, 53).
FI barometry is turned into depths according to the conceptual
stratigraphic model presented in fig. S3. These data, in good agree-
ment with the geophysical depths recorded before and during the
eruption, indicate that ascending magmas ponded for a longer time
at two depth intervals, specifically from −­8 to −­16 km and from −­22
to −­27 km, and discontinuously at −­4 and −7 km (Fig. 1, B and C, and
fig. S4). Both main seismic sources identified sub-­vertical volumes
that were almost coaxial with a displacement of ~4 km, located near
the eruptive fracture (28, 54) and connected by a dike that dips 19°
to 20° northwest. These elements exclude horizontal magma propa-
gation of considerable magnitude, resembling the 2011–2012 El Hi-
erro eruption (55). Ascending magmas were temporarily retained
and accumulated before their final ascent.
The deepest magma accumulation zone at a depth range of 20.5
to 27.5 km probably consists of mafic to ultramafic cumulate layers
(50, 56) with a ρ = 3115 kg m−3, and its lower limit probably marks
the transition to the lithospheric mantle (ρ = 3390 kg m−3). The
lower limit of the intermediate accumulation zone at a depth of 10
to 15 km would mark the transition from rocks with a density of
2655 kg m−3 to deeper rocks with a density of 3060 kg m−3 (fig. S3).
Furthermore, the magmas ponded briefly and intermittently at
depths of ~4 km (only in late October) and ~7 km (from October to
early November and late December).
Magma ascent dynamics and velocity
The 2021 Tajogaite eruption ejected multiple batches of mantle-­
derived magma that ascended through the volcano’s magma system
at different rates, as evidenced by the different degrees of re-­
equilibration (43, 57) and by the simultaneous presence in the same
sample of olivines hosting CO2 (+ H2O) fluids and olivines hosting
CO2 + N2 (± CO) fluids. The survival of nitrogen-­bearing FI

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 6 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Histograms of FI density data. Data from clinopyroxenes are shown in blue; data from amphiboles are shown in black. Data from CO2 + H2O–bearing FI in olivines are shown
in red; cyan bars are data from CO2 + H2O + CO + N2–bearing FI. Vertical pale yellow stripes represent density ranges from literature data for the whole Cumbre Vieja volcano (17).

depended on the duration of magma ponding at depths near the
transition to mantle lithologies at the base of the deeper magma ac-
cumulation zone.
Given the architecture of the magma system, the preeruptive
seismicity was related to magma refilling of existing structures at
depth. Regarding syn-­eruptive seismicity, two well-­separated clus-
ters were observed (28) and justified with the readjustment of the
crust and upper mantle due to the emptying of these two magmatic
reservoirs. Figure 5A clearly shows the temporal relationship be-
tween the deep and intermediate seismicity, confirming some kind
of hydraulic connection between these two reservoirs. This hypoth-
esis is also consistent with the results of (5). This process may have
induced a downward piston effect (28) that temporarily halted mag-
ma withdrawal during periods of compression and enhanced brief

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 7 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

magma ponding. This piston effect could be responsible for the
deepening of the magma source over time, as revealed by FI barom-
etry. A similar process was already observed in the recent Fagradals-
fjall eruption (11) and was accompanied by an increase in the
flow rate.
The similarity of the earthquake frequency curves generated in
the two storage areas (Fig. 5A), along with the temporal shift of
similar frequency peaks, can provide estimates of the total number
of magma pulses that occurred (method 1). These observations
may additionally provide an estimate of the time-­integrated mag-
ma ascent velocity between the two main ponding zones, which
includes the residence time, as indicated in Table 2. To avoid any
possible misinterpretation, we only considered peaks whose am-
plitude clearly stands out from the surrounding values and that
show a similar shape in the two curves. Using this conservative
approach is essential to ensure the accuracy of our qualitative esti-
mates. Moreover, the correlation between intermittent magma
ponding at depths of ~4 and ~7 to 8 km with peaks of high tremor
amplitude in Fig. 5B can facilitate the calculation of the final mag-
ma velocity (method 2). In this diagram, the very long period
component of tremor, which seems to be related to the volume of
gas involved, has been used. We hypothesize that the shallow
ponding mentioned in this context could be related to the accu-
mulation of gas, which ultimately fueled the energetic episodes of
lava fountains.
The velocities obtained by these two independent methods are
summarized in Table 2 and range from 0.01 to 0.04 m s−1 from the
deep to the intermediate magma accumulation zone and from 0.05
to 0.1 m s−1 from the shallower, intermittently active ponding zone
at 4-­to 7-­km depth. The velocities listed above are minimum values
because it is difficult to determine the precise timing of tephra emis-
sion and the exact depth of seismicity corresponding to each peak.
The variability of these estimates could be partially explained by the
presence of volumetrically different pulses of magma ascending
through a magma system, which changes with time, by the dynam-
ics of magma extraction that experiences varying degrees of pres-
surization and decompression, and by different ponding times.
These estimates are consistent with the absence of dense ultramafic
xenoliths in the erupted products, indicating that an ascent rate
greater than 0.2 m s−1 is required for a bubble-­free melt, as suggest-
ed by (58). The estimated time-­integrated velocities mentioned
above are an order of magnitude lower compared to the ascent ve-
locity calculated for xenolith-­bearing basanites during the 1949
eruption (59).

Fig. 5. Magma velocity derived from geophysical data. (A) The frequency curves of deep (red curve) and shallow (blue curve) earthquakes obtained by a 1-­day moving
average. On the ordinates, the daily number of earthquakes occurring at different depths is reported. The boundary between the two depth ranges is arbitrary. Velocity is
estimated as the time difference between two corresponding frequency peaks as the magma moves from the top of the deep reservoir to the upper reservoir. The codes
are explained in Table 2. (B) The tremor amplitude pattern (blue dots) and the interpolation of the mean values (red curve). Dashed lines mark the beginning and end of
magma emission. Velocity is estimated as the time difference between high-­amplitude tremor events (black arrows) and the day of sampling, when FI indicated a short
and very shallow ponding stage (between 4.3 and 7 km). Relevant data are presented in Table 2.

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 8 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Table 2. Time-­integrated magma ascent velocity between seismic zones. Seismic references are shown in Fig. 5. ∆X of 13.5 km is the distance between the
tops of the two magma accumulation zones, whereas ∆X is the distance between the shallowest magma ponding zone and the sea level in the remaining part
of the table.
Possible interpreta-
Seismic reference Date Estimated time tion ∆X (m) vav (m s−­1)

I0 25 September 3 days, 20 hours The decompression 13,500 0.04

D0 29 September (331,200 s) caused by magma
withdrawal from 4 to 14
km depth (I0) causes the
mobilization of deeper
magma batch from
the deeper reservoir
(D0) to compensate the
pressure gradient.

D1 9 October 14 days, 22 hours At the end of deep 13,500 0.01

I1 24 October (1,288,800 s) magma withdrawal
(D1), possible grav-
itative adjustments
caused a negative
pressure gradient
propagating upward,
affecting the conduit
system at 4-­to 14-­km
depth (I1).

D2 10 November 8 days At the end of deep 13,500 0.02

I2 16 November (691,200 s) magma extraction (D2),
possible gravitative
adjustments caused
a negative pressure
gradient propagating
upward, affecting the
conduit system at 4-­to
14-­km depth (I2).

D3 22 November 6 days, 1 hour At the end of deep 13,500 0.03

I3 30 November (522,000 s) magma extraction (D3),
possible gravitative
adjustments caused
a negative pressure
gradient propagating
upward, affecting the
conduit system at 4-­to
14-­km depth (I3).

D4 7 December 4 days, 23 hours Following the last event 13,500 0.03

I4 12 December (428,400 s) of magma withdrawal
(D4), possible grav-
itative adjustments
caused a negative
pressure gradient
propagating upward,
affecting the conduit
system at 4-­to 14-­km
depth (I4).

Tremor amplitude 27 September 2 days, 8 hours Temporary end of 15,000 0.07

tremor
I0 25 September (201,600 s) Beginning of decom-
pression of the interme-
diate reservoir

(Continued)

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 9 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

(Continued)

Possible interpreta-
Seismic reference Date Estimated time tion ∆X (m) vav (m s−1­ )
Tremor amplitude 20 October 12 hours Peak of tremor ampli- 4,300 0.1
tude
Sample collection day* 21 October (43,200 s) FI recording shallow
ponding stage

Tremor amplitude 27 October 24 hours Peak of tremor ampli- 4,300 0.05

tude
Sample collection day* 29 October (86,400 s) FI recording shallow
ponding stage

Tremor amplitude 4 November 1 days, 10 hours Peak of tremor ampli- 7,000 0.06
tude
Sample collection day* 6 November (122,400 s) FI recording shallow
ponding stage

Tremor amplitude 25 November 16 hours Peak of tremor ampli- 5,000 0.09

tude
Sample collection day* 27 November (57,600 s) FI recording shallow
ponding stage

*A delay of 12 hours was applied between tephra emission and sample collection.

Near real-­time magma ascent monitoring
The histograms of magma ponding depths (fig. S4) were decon-
structed as a function of the various ascent rates between the two
principal zones of magma accumulation to describe how the magma
system’s dynamics evolved during the eruption as multiple pulses of
magma ascended (Fig. 6).
Phase 1: From 19 to 27 September (Fig. 6E). Preeruptive activity
included a series of low-­magnitude seismic swarms lasting a few
days and ranging in depth from 15 to 25 km (Fig. 1A), highlighting
multiple magma intrusions. During its ascent, the 11 September in-
trusion displaced a cool and partially evolved basanite residing be-
tween ~13-­and 16.5-­km depth. The latter magma, the first to be
erupted (batch#1), was either a remnant of the 1949 Llano del Banco
eruption or, more likely, one of the earliest intrusions in 2017, whose
amphibole and clinopyroxene had sufficient time to fully re-­
equilibrate their FI populations. The calculated ascent rate of 0.04 m
s−1 (Table 2) allows to define that all olivines sampled on 29 Septem-
ber belonged to the 11 September magma (batch#2) and showed a
nonstop ascent from a depth of 27 km, as confirmed by the absence
of any re-­equilibration at shallower depths. Seismicity during this
period occurred within the shallower 4 km and in the intermediate
seismic zone (Figs. 1A and 5A) and was associated with the empty-
ing of the intermediate magma accumulation zone at a rate of 0.04 m
s−1. The cessation of seismicity at 4-­to 14-­km depth on 24 Septem-
ber, followed by the temporary cessation of magma emission and
seismic tremor, suggested that the last 14 km of the magma system
had been emptied.
Phase 2: From 27 September to 24 October (Fig. 6B). In the early
afternoon of 27 September, the eruption resumed with the emission
of a hotter basanite (batch#2) whose mineralogy was characterized
by a progressive reduction in pyroxene and amphibole crystals and
an increase in olivines. In early October, the deeper seismic source
was activated, indicating magma ponding at depths of 25 to 27 km.
FI barometry still confirms the source of pyroxenes and amphiboles
at a depth range of ~13 to 16.5 km (Fig. 6), while olivine rose from a
depth of ~22 to 27 km (Fig. 4), but with partial re-­equilibration in
the intermediate storage zone. The duration of these re-­equilibration
events was responsible for the estimated slow velocity between the
two magma accumulation zones (0.01 m s−1).
Phase 3: From 24 October to 19 November (Fig. 6C). During this
period, the magma system of Tajogaite volcano was partially occu-
pied by a new pulse of basanite (batch#3), characterized by olivines
hosting CO2 + H2O + CO + N2–bearing FI and ascending with an
integrated velocity of 0.02 m s−1. The presence of olivines hosting
CO-­N2–free FI at the same time suggests that remaining amounts of
the prior magma (batch #2) with slow mobility were still present in
the conduit (see Fig. 5A). Noteworthy, the contemporaneous pres-
ence of these two kinds of magma inside the volcano is confirmed by
the decoupling of Sr isotope compositions in matrix melt (5) and in
the split in Os isotope compositions in whole rock (12). The distinc-
tive depths of the two magma pondings and the elevated seismic
activity in this period suggest the possible activation of a secondary
pathway, which would allow the two magmas to ascend inde-
pendently.
Phase 4: From 19 November to 13 December (Fig. 6D). From
mid-­November, seismicity began to decrease at all depths (Fig. 5A).
The amplitude of tremor was sustained until early December. FI ba-
rometry shows active magma ponding in the intermediate magma
accumulation zone only in late November. N2-­bearing FI and N2-­
free FI in olivines continued to be erupted, possibly through the

Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024 10 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 6. Cartoon resuming the main periods of magma ascent monitoring under
the Tajogaite volcano, in 2021. (A to D) Histograms of the frequency of depths
have been normalized to 100%. Green bars represent CO2 + H2O FI (hosted in clino-
pyroxene, amphibole, and olivine); cyan bars represent olivine-­hosted CO + N2 (+
CO2 + H2O)–bearing FI. Light-­violet horizontal stripes represent seismicity (28).
These stripes, when dashed, indicate decreasing seismicity. The estimated ascent
time is shown in blue. The cartoons on the right show the conceptual model with
the different ascending magma batches (in different colors) and their ponding
stages.
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
dual ascent path. The time-­integrated ascent velocity during this
phase was calculated to be 0.03 m s−1, in agreement with the absence
of magma ponding in the intermediate accumulation zone. The
magma erupted toward the end of the eruption is an alkali basalt
(more silicic, but with the same alkali content as the basanites),
which forms the batch#4 and ascended from a depth of 31 km.
Starting from 1 December, the deep seismic activity came to an
end and the intermediate seismic activity continued to decrease.
This marked the end of magma supply from the mantle, along
with the progressive emptying of the entire magma system of the
volcano.
In conclusion, FI microthermometry combined with basic petro-
graphic observations in tephra and lava samples, along with seis-
micity recorded during the 2021 eruption of Cumbre Vieja, has
enabled near real-­time monitoring of the ascent of various magma
pulses. The consistency of our database, based on 18 time-­
constrained samples, has highlighted the rise of different magma
pulses over 3 months. The geobarometric information on magma
ponding from FI analysis correlated precisely with the location of
magma extraction from seismogenic sources and indicated the pro-
gressive deepening of the magma source (from 27-­to 31-­km depth)
during the eruption.
This combined approach facilitated the reconstruction of the
syn-­eruptive changes in the magma system of the volcano and pro-
vided estimates of the time-­integrated magma velocities (from 0.01
to 0.1 m s−1), which include pauses in the magma accumulation
zones. This information is crucial to comprehend the overall dy-
namics of magma ascent and to interpret geophysical signals accu-
rately. The high resolution of the data provided in this study is
related to the investigation of time-­constrained samples and the
high number of inclusions analyzed per sample, regardless of the
mineral host.
Similar conclusions have been provided in other works, per-
forming Raman spectroscopy on FI (24) and a chemical study
on clinopyroxene combined with clinopyroxene-­melt barometry
(5), indicating that the study of time-­constrained samples is the
way for a more efficient petrological monitoring. However, the ability to
trace magma ponding events in a short time, due to the reduced
time required for sample preparation and analysis, which can be
performed in near real-­t ime (27), makes FI microthermometry
a rapid, concise, and informative method. Its combination with
geophysical monitoring makes its use highly recommended for
cost-­effective petrological monitoring of ongoing magmatic
processes during an eruption and represents an advance for the
improvement of monitoring strategies, especially in institutions
where large analytical platforms are not feasible.
MATERIALS AND METHODS
Seismicity
Both the Instituto Geográfico Nacional (IGN) and the Instituto
Volcanológico de Canarias (INVOLCAN) recorded seismic moni-
toring data during the eruption. This work uses the catalog pro-
duced by INVOLCAN [publicly available dataset; (28)]. The
catalog contains hypocenters that have been relocated in real-­time
during the eruption in a three-­dimensional tomographic model,
resulting from the joint analysis of seismic phases from both IGN
and INVOLCAN. This leads to more reliable depths of the
earthquakes.
11 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Sample description
A total of 13 tephra and five lava samples were collected during the
joint routine monitoring of the Tajogaite eruption by the INVOL-
CAN and the Instituto de Investigação em Vulcanologia e Avaliação
de Riscos. A brief description and coordinates are given in table S1.
Whole-­rock and mineral chemistry
Major element compositions of five lava samples differing in petro-
graphic characteristics were analyzed at Actlabs (Activation Labora-
tories, Canada) using a PerkinElmer 9000 inductively coupled
plasma–mass spectrometer and an Agilent 735 inductively coupled
plasma–atomic emission spectrometer. Alkaline dissolution with
lithium metaborate/tetraborate followed by nitric acid was used on
1 g of rock powder before melting in an induction furnace. The melt
was poured into a 5% nitric acid solution containing cadmium as an
internal standard and stirred until complete dissolution. The result-
ing analytical accuracy is better than 6% for all major elements.
Seven international rock standards were used to calibrate the two
methods. Compositional data for the major elements are given
in table S2.
The major element compositions of olivines and clinopyro
xenes were measured using a Cameca SXFive electron microprobe
(Camparis, Paris, France) at 15 keV and a 20-­nA focused beam.
Counting times were 30 s on the peak and 10 s on each background.
The San Carlos olivine and the Puy de Dôme clinopyroxene, used as
standards, were calibrated to within 2% at 2σ. Raw data were cor-
rected using a Phi-­Rho-­Z quantitative analysis program. The typical
detection limit for each element is 0.01%. Relative errors are better
than 6% for NiO, alkali, and MnO and better than 2% for all other
major elements.
Eruptive temperatures of early basanite magmas were calculated
from the MgO content of glass (60). Glassy fragments of lapilli eject-
ed in October were analyzed using a Cameca SXFive electron
microprobe (Laboratoire Magmas et Volcans, University of Clermont-­
Auvergne, France) at 15 keV and a 10-­nA beam with a defocused
spot size of 10 μm. ALV-­98I and CH98-­DR11 international standard
glasses were used to check the errors and reproducibility. Relative
errors are better than 3% for most major elements, 19% for MnO,
and 25% for K2O. Microprobe averages, number of measurements,
and SD are given in table S3.
Fluid inclusions
FI were identified in 18 samples of lava and tephra. The lava samples
were initially crushed roughly using a jaw crusher. The crushed lavas
and tephra were then sieved to isolate different crystal size popula-
tions. For every sample, around 100 olivines (diameters ranging
from 0.35 to 0.65 cm), 50 clinopyroxenes, and amphiboles (up to 1.0 cm
long) were separated. These separated crystals were then thinned to
a thickness of 60 to 80 μm, double polished, and examined under a
light microscope to search for FI. The inclusions were identified in
all of the samples. Microthermometry was conducted on a Linkam
MDSG600 heating-­cooling stage, which was calibrated with syn-
thetic FI standards consisting of pure CO2 and H2O. Reproducible
melting and homogenization temperatures to within ±0.1°C were
obtained at heating rates of 0.2° to 0.5°C/min.
The density values of the CO2 fluid were computed according to
equations 3.14 and 3.15 of (61). Isochores were obtained for a pure
CO2 fluid using the equation of state for carbon dioxide of (62),
which is valid up to at least 2000 K and 10 GPa. High density values
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
for pure CO2 fluids in olivine were corrected for a probable pristine
presence of ~10% water [H2O:CO2 = 1:9; (40)] and then isochores
for these H2O-­CO2 fluids were calculated (63). When recording the
presence of N2 and CO, isochores were computed only for inclu-
sions having negative homogenization temperature using VX plots
(64). On the basis of this study’s findings, at positive homogeniza-
tion temperatures, the impact of increasing bulk density in FI up to
5% N2 in the mixture is trivial.
The eruption temperatures for early magmas have been calcu-
lated using chemical geothermometry (60) modified by (13), which
correlates the temperature of olivine crystallization with the MgO
concentration of the magma. The temperatures ranged from 1073°
to 1084°C (table S3). The temperature of 1075°C was applied to cal-
culate the isochores of FI that were trapped in clinopyroxenes and
amphiboles. For those that were trapped in olivines, a temperature
of 1150°C was used in compliance with direct field measurements
and experimental petrology data for magmas erupted in December
(12, 32, 65). Uncertainty on the eruptive temperature calculation has
a little effect on the final pressure calculation: Considering an FI
with ρ = 1000 kg m−3, an uncertainty of ±20°C in the estimation of
the eruptive temperature (e.g., between 1130° and 1150°C) gener-
ates a maximum error of ±10 MPa.
Table 1 presents microthermometric data, while fig. S2 shows the
histograms of trapping/re-­equilibration pressures. Using the strati-
graphic scheme described in the Supplementary Materials and
shown in fig. S3, pressures were converted into depths. In this mod-
el, we defined major stratigraphic changes through both the inter-
pretation of the seismic velocity model in (28, 54) and the FI
ponding stage scheme (this work). After measurements using an
electronic densimeter, we assigned rock density values to shallow
lavas (ρ = 2350 kg m−3), dense gabbroic xenoliths (ρ = 3655 kg
m−3), and mantle lithologies (ρ = 3115 to 3390 kg m−3), while we
assumed the density values for other rock bodies.
Raman microspectroscopy
Raman microspectroscopic analysis was performed on two olivine
crystals from a lava sample collected at the end of the eruption (10
December) that contain FI with melting temperatures ranging from
-­56.8° to -­57.3°C, indicating the presence of volatile species other
than CO2 and H2O. FI were analyzed using a Renishaw inVia confo-
cal Raman microspectrometer, equipped with a 532.1 ± 0.3–nm
diode-­ pulsed solid-­ state laser (~180-­ mW output power), a
1040 × 256 pixel charge-­coupled device detector, a Rayleigh rejec-
tion edge filter, and a Leica DM 2500 M optical microscope with a
motorized XYZ stage at the Laboratoire Magmas et Volcans
(France). The spectra were acquired in backscattered geometry, in
both standard and high confocality modes (slit aperture of 65 and
20 μm), with a 50× microscope objective and a grating of 2400 lines/
mm. Each acquisition consisted of two accumulations of 20 s, and
the laser power was set to 10 or 50% (i.e., ~12 and 60 mW). Spectra
were recorded in extended mode from 60 to 4500 cm−1 using WiRE
4.4 software. Daily calibration of the spectrometer was performed
using a silicon standard (520.5 cm−1 peak) and several neon lines.
Analyses were carried out at a constant temperature of ~20.5°C.
Following the method described in (66), the molar proportions
of the different components present in the FI (CO2, CO, and N2)
were calculated from Raman peaks areas (normalized for laser pow-
er and exciting time); instrumental efficiencies of 1, 1, and 0.5 for
N2, CO, and CO2, respectively; and wavelength-­dependent Raman
12 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
scattering efficiencies. The wavelength-­dependent Raman scattering
efficiencies for specific Raman shifts were calculated with respect to
the scattering efficiency of N2, based on equation 1 of (66). CO2 con-
centrations were calculated using the sum of the areas of the two
Fermi diad peaks and the sum of their scattering efficiencies.
Supplementary Materials
This PDF file includes:
Supplementary text
Figs. S1 to S4
Tables S1 to S4
References
REFERENCES AND NOTES
1. A. García, A. Fernández-­Ros, M. Berrocoso, J. M. Marrero, G. Prates, S. De la Cruz-­Reyna,
R. Ortiz, Magma displacements under insular volcanic fields, applications to
eruption forecasting: El Hierro, Canary Islands, 2011–2013. Geophys. J. Int. 197, 322–334
(2014).
2. Z. Duputel, O. Lengliné, V. Ferrazzini, Constraining spatiotemporal characteristics of
magma migration at Piton de la Fournaise volcano from pre-­eruptive seismicity. Geophys.
Res. Lett. 46, 119–127 (2019).
3. M. J. Pankhurst, D. J. Morgan, T. Thordarson, S. C. Loughlin, Magmatic crystal records in
time, space, and process, causatively linked with volcanic unrest. Earth Planet. Sci. Lett.
493, 231–241 (2018).
4. D. J. Rasmussen, T. A. Plank, D. C. Roman, J. A. Power, R. J. Bodnar, E. H. Hauri, When does
eruption run-­up begin? Multidisciplinary insight from the 1999 eruption of Shishaldin
volcano. Earth Planet. Sci. Lett. 486, 1–14 (2018).
5. T. Ubide, Á. Márquez, E. Ancochea, M. J. Huertas, R. Herrera, J. J. Coello-­Bravo,
D. Sanz-­Mangas, J. Mulder, A. MacDonald, I. Galindo, Discrete magma injections drive the
2021 La Palma eruption. Sci. Adv. 9, eadg4813 (2023).
6. R. Corsaro, L. Miraglia, V. Zanon, Petrologic monitoring of glasses in the pyroclastites
erupted in February 2004 by the Stromboli Volcano, Aeolian Islands, Southern Italy. J.
Volcanol. Geotherm. Res. 139, 339–343 (2005).
7. S. A. Halldórsson, E. Bali, M. E. Hartley, D. A. Neave, D. W. Peate, G. H. Guðfinnsson,
I. Bindeman, M. J. Whitehouse, M. S. Riishuus, G. B. M. Pedersen, S. Jakobsson, R. Askew,
C. R. Gallagher, E. R. Guðmundsdóttir, J. Gudnason, W. M. Moreland, B. W. Óskarsson,
P. Nikkola, H. I. Reynolds, J. Schmith, T. Thordarson, Petrology and geochemistry of the
2014–2015 Holuhraun eruption, central Iceland: Compositional and mineralogical
characteristics, temporal variability and magma storage. Contrib. Mineral. Petrol. 173,
1–25 (2018).
8. P. Landi, R. A. Corsaro, L. Francalanci, L. Civetta, L. Miraglia, M. Pompilio, R. Tesoro, Magma
dynamics during the 2007 Stromboli eruption (Aeolian Islands, Italy): Mineralogical,
geochemical and isotopic data. J. Volcanol. Geotherm. Res. 182, 255–268 (2009).
9. N. Spilliaert, P. Allard, N. Métrich, A. V. Sobolev, Melt inclusion record of the conditions of
ascent, degassing, and extrusion of volatile-­rich alkali basalt during the powerful 2002
flank eruption of Mount Etna (Italy). J. Geophys. Res. Solid 111, B04203 (2006).
10. C. Gansecki, R. L. Lee, T. Shea, S. P. Lundblad, K. Hon, C. Parcheta, The tangled tale of
Kīlauea’s 2018 eruption as told by geochemical monitoring. Science 366, eaaz0147
(2019).
11. S. A. Halldórsson, E. W. Marshall, A. Caracciolo, S. Matthews, E. Bali, M. B. Rasmussen,
E. Ranta, J. Gunnarsson-­Robin, G. H. Guðfinnsson, O. Sigmarsson, J. Maclennan,
M. G. Jackson, M. J. Whitehouse, H. Jeon, Q. H. A. van der Meer, G. K. Mibei,
M. H. Kalliokoski, M. M. Repczynska, R. H. Rúnarsdóttir, G. Sigurðsson, M. A. Pfeffer,
S. W. Scott, R. Kjartansdóttir, B. I. Kleine, C. Oppenheimer, A. Aiuppa, E. Ilyinskaya,
M. Bitetto, G. Giudice, A. Stefánsson, Rapid shifting of a deep magmatic source at
Fagradalsfjall volcano, Iceland. Nature 609, 529–534 (2022).
12. J. M. Day, V. R. Troll, M. Aulinas, F. M. Deegan, H. Geiger, J. C. Carracedo, G. Gisbert Pinto,
F. J. Perez-­Torrado, Mantle source characteristics and magmatic processes during the
2021 La Palma eruption. Earth Planet. Sci. Lett. 597, 117793 (2022).
13. K. D. Putirka, “Thermometers and barometers for volcanic systems” in Minerals, Inclusions
and Volcanic Processes, vol. 69 of Reviews in Mineralogy Geochemistry, K. D. Putirka,
F. Tepley, Eds. (Mineralogical Society of America, 2008), pp. 61–120.
14. P. E. Wieser, A. J. F. Kent, C. B. Till, J. Donovan, D. A. Neave, D. L. Blatter, M. J. Krawczynski,
Barometers behaving badly: Assessing the influence of analytical and experimental
uncertainty on clinopyroxene thermobarometry calculations at crustal conditions. J.
Petrol. 64, egac126 (2022).
15. T. H. Hansteen, A. Klügel, H.-­U. Schmincke, Multi-­stage magma ascent beneath the
Canary Islands: Evidence from fluid inclusions. Contrib. Mineral. Petrol. 132, 48–64
(1998).
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
16. E. Hildner, A. Klügel, T. H. Hansteen, Barometry of lavas from the 1951 eruption of Fogo,
Cape Verde Islands: Implications for historic and prehistoric magma plumbing systems.
J. Volcanol. Geotherm. Res. 217–218, 73–90 (2012).
17. A. Klügel, T. H. Hansteen, K. Galipp, Magma storage and underplating beneath Cumbre
Vieja volcano, La Palma (Canary Islands). Earth Planet. Sci. Lett. 236, 211–226 (2005).
18. G. Boudoire, Y-­A. Brugier, A. Di Muro, G. Wörner, I. Arienzo, N. Metrich, V. Zanon,
N. Braukmüller, A. Kronz, Y. Le Moigne, M. Michon, Eruptive activity on the western flank
of Piton de la Fournaise (La Réunion Island, Indian Ocean): Insights on magma transfer,
storage and evolution at an oceanic volcanic island. J. Petrol. 60, 1717–1752 (2019).
19. V. Zanon, M. L. Frezzotti, A. Peccerillo, Magmatic feeding system and crustal magma
accumulation beneath Vulcano Island (Italy): Evidence from CO2 fluid inclusions in quartz
xenoliths. J. Geophys. Res. Solid Earth 108, 2298 (2003).
20. V. Zanon, I. Nikogosian, Evidence of crustal melting events below the island of Salina
(Aeolian arc, southern Italy). Geol. Mag. 141, 525–540 (2004).
21. V. Zanon, A. Pimentel, M. Auxerre, G. Marchini, F. M. Stuart, Unravelling the magma
feeding system of a young basaltic oceanic volcano. Lithos 352-­353, 105325 (2020).
22. V. Zanon, “Conditions for mafic magma storage beneath fissure zones at oceanic islands.
The case of São Miguel island (Azores archipelago)” in Chemical, Physical and Temporal
Evolution of Magmatic Systems, vol. 422, L. Caricchi, J. D. Blundy Eds. (The Geological
Society of London, 2015), pp. 85–104.
23. F. M. Lo Forte, A. Aiuppa, S. G. Rotolo, V. Zanon, Temporal evolution of the fogo volcano
magma storage system (cape verde archipelago): A fluid inclusions perspective. J.
Volcanol. Geotherm. Res. 433, 107730 (2023).
24. K. Dayton, E. Gazel, P. Wieser, V. R. Troll, J. C. Carracedo, H. La Madrid, D. C. Roman, J. Ward,
M. Aulinas, H. Geiger, F. M. Deegan, G. Gisbert, F. J. Perez-­Torrado, Deep magma storage
during the 2021 La Palma eruption. Sci. Adv. 9, eade7641 (2023).
25. M. L. Frezzotti, S. Ferrando, F. Tecce, D. Castelli, Water content and nature of solutes in
shallow-­mantle fluids from fluid inclusions. Earth Planet. Sci. Lett. 351-­352, 70–83 (2012).
26. R. J. Bakker, The perfection of Raman spectroscopic gas densimeters. J. Raman Spectrosc.
52, 1923–1948 (2021).
27. V. Zanon, K. Cyrzan, L. D'Auria, M. J. Pankhurst, F. Rodríguez, B. Coldwell,
A. Martín-­Lorenzo, “The magma ascent path during the 2021 eruption of Cumbre Vieja
(La Palma Island, Canary archipelago) highlighted by fluid inclusions and seismicity” in
EGU General Assembly Conference Abstracts (Copernicus Publications, 2022)
p. EGU22-10203. doi: 10.5194/egusphere-egu22-1020.
28. L. D’Auria, I. Koulakov, J. Prudencio, I. Cabrera-­Pérez, J. M. Ibáñez, J. Barrancos,
R. García-­Hernández, D. M. van Dorth, G. D. Padilla, M. Przeor, V. Ortega, P. Hernández,
N. M. Peréz, Rapid magma ascent beneath La Palma revealed by seismic tomography. Sci.
Rep. 12, 17654 (2022).
29. P. A. Torres-­González, N. Luengo-­Oroz, H. Lamolda, W. D'Alessandro, H. Albert, I. Iribarren,
D. Moure-­García, V. Soler, Unrest signals after 46 years of quiescence at Cumbre Vieja,
La Palma, Canary Islands. J. Volcanol. Geotherm. Res. 392, 106757 (2020).
30. I. Cabrera-­Pérez, L. D'Auria, J. Soubestre, M. Przeor, J. Barrancos, R. García-­Hernández,
J. M. Ibáñez, I. Koulakov, D. M. van Dorth, V. Ortega, G. D. Padilla, T. Sagiya, N. Pérez,
Spatio-­temporal velocity variations observed during the pre-­eruptive episode of La Palma
2021 eruption inferred from ambient noise interferometry. Sci. Rep. 13, 12039 (2023).
31. J. E. Romero, M. Burton, F. Cáceres, J. Taddeucci, R. Civico, T. Ricci, M. J. Pankhurst,
P. A. Hernández, C. Bonadonna, E. W. Llewellin, M. Pistolesi, M. Polacci, C. Solana,
L. D'Auria, F. Arzilli, D. Andronico, F. Rodríguez, M. Asensio-­Ramos, A. Martín-­Lorenzo,
C. Hayera, P. Scarlato, N. M. Perez, The initial phase of the 2021 Cumbre Vieja ridge
eruption (Canary Islands): Products and dynamics controlling edifice growth and
collapse. J. Volcanol. Geotherm. Res. 431, 107642 (2022).
32. J. C. Carracedo, V. R. Troll, J. M. D. Day, H. Geiger, M. Aulinas, V. Soler, F. M. Deegan,
F. J. Perez-­Torrado, G. Gisbert, E. Gazel, A. Rodriguez-­Gonzalez, H. Albert, The 2021 eruption of
the Cumbre Vieja volcanic ridge on La Palma, Canary Islands. Geol. Today 38, 94–107 (2022).
33. S. M. Sterner, R. J. Bodnar, Synthetic fluid inclusions -­VII. Re-­equilibration of fluid
inclusions in quartz during laboratory-­simulated metamorphic burial and uplift. J.
Metamorph. Geol. 7, 243–260 (1989).
34. M. O. Vityk, R. J. Bodnar, C. S. Schmidt, Fluid inclusions as tectonothermobarometers:
Relation between pressure-­temperature history and reequilibration morphology during
crustal thickening. Geology 22, 731–734 (1994).
35. M. O. Vityk, R. J. Bodnar, Textural evolution of synthetic fluid inclusions in quartz during
reequilibration, with applications to tectonic reconstruction. Contrib. Mineral. Petrol. 121,
309–323 (1995).
36. V. Zanon, M. L. Frezzotti, Magma storage and ascent conditions beneath Pico and Faial
islands (Azores archipelago): A study on fluid inclusions. Geochem. Geophys. Geosystems
14, 3494–3514 (2013).
37. M. Berkesi, K. Hidas, T. Guzmics, J. Dubessy, R. J. Bodnar, C. Szabo, B. Vajna, T. Tsunogae,
Detection of small amounts of H2O in CO2-­rich fluid inclusions using Raman
spectroscopy. J Raman Spectrosc. 40, 1461–1463 (2009).
38. E. Sendula, H. M. Lamadrid, J. D. Rimstidt, M. Steele-­MacInnis, D. M. Sublett, L. E. Aradi,
C. Szabó, M. J. Caddick, Z. Zajacz, R. J. Bodnar, Synthetic Fluid Inclusions XXIV. In situ
13 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
monitoring of the carbonation of olivine under conditions relevant to carbon capture
and storage using synthetic fluid inclusion micro-­reactors: Determination of reaction
rates. Front. clim. 3, 722447 (2021).
39. J. E. Dixon, E. M. Stolper, An experimental study of water and carbon dioxide solubilities
in mid-­ocean ridge basaltic liquids, Part II: Applications to degassing. J. Petrol. 36,
1633–1646 (1995).
40. T. H. Hansteen, A. Klügel, Fluid inclusion thermobarometry as a tracer for magmatic
processes. Rev. Mineral. Geochem. 69, 143–177 (2008).
41. T. Andersen, E. A. J. Burke, E. R. Neumann, Nitrogen-­rich fluid in the upper mantle: Fluid
inclusions in spinel dunite from Lanzarote, Canary Islands. Contrib. Mineral. Petrol. 120,
20–28 (1995).
42. B. J. Wanamaker, B. Evans, Mechanical re-­equilibration of fluid inclusions in San Carlos
olivine by power-­law creep. Contrib. Mineral. Petrol. 102, 102–111 (1989).
43. M. O. Vityk, R. J. Bodnar, Statistical microthermometry of synthetic fluid inclusions in quartz
during decompression reequilibration. Contrib. Mineral. Petrol. 132, 149–162 (1998).
44. A. Klügel, K. A. Hoernle, H. U. Schmincke, J. D. White, The chemically zoned 1949 eruption
on La Palma (Canary Islands): Petrologic evolution and magma supply dynamics of a rift
zone eruption. Geophys. Res. Solid Earth 105, 5997–6016 (2000).
45. Y. Moussallam, M. A. Longpré, C. McCammon, A. Gomez-­Ulla, E. F. Rose-­Koga, B. Scaillet,
N. Peters, E. Gennaro, R. Paris, C. Oppenheimer, Mantle plumes are oxidised. Earth Planet.
Sci. Lett. 527, 115798 (2019).
46. R. W. Nicklas, R. K. Hahn, L. N. Willhite, M. G. Jackson, V. Zanon, R. Arevalo Jr., J. M. Day, Oxidized
mantle sources of HIMU-­and EM-­type ocean island basalts. Chem. Geol. 602, 120901 (2022).
47. G. Libourel, B. Marty, F. Humbert, Nitrogen solubility in basaltic melt. Part I. Effect of
oxygen fugacity. Geochim. Cosmochim. Acta 67, 4123–4135 (2003).
48. J. Boulliung, E. Füri, C. Dalou, L. Tissandier, L. Zimmermann, Y. Marrocchi, Oxygen fugacity
and melt composition controls on nitrogen solubility in silicate melts. Geochim.
Cosmochim. Acta 284, 120–133 (2020).
49. H. Keppler, L. Cialdella, F. Couffignal, M. Wiedenbeck, The solubility of N2 in silicate melts
and nitrogen partitioning between upper mantle minerals and basalt. Contrib. Mineral.
Petrol. 177, 83 (2022).
50. A. K. Barker, V. R. Troll, J. C. Carracedo, P. A. Nicholls, The magma plumbing system for the
1971 Teneguía eruption on La Palma, Canary Islands. Contrib. Mineral. Petrol. 170, 54 (2015).
51. K. Galipp, A. Klügel, T. H. Hansteen, Changing depths of magma fractionation and
stagnation during the evolution of an oceanic island volcano: La Palma (Canary Islands).
J. Volcanol. Geotherm. Res. 155, 285–306 (2006).
52. A. Klügel, E. Albers, T. H. Hansteen, Mantle and crustal xenoliths in a tephriphonolite from
La Palma (Canary Islands): Implications for phonolite formation at oceanic island
volcanoes. Earth Sci. 10, (2022).
53. M.-­A. Longpré, A. Klügel, A. Diehl, J. Stix, Mixing in mantle magma reservoirs prior
to and during the 2011–2012 eruption at El Hierro, Canary Islands. Geology 42, 315–318
(2014).
54. C. Del Fresno, S. Cesca, A. Klügel, I. Domínguez Cerdeña, E. A. Díaz-­Suárez, T. Dahm,
L. García-­Cañada, S. Meletlidis, C. Milkereit, C. Valenzuela-­Malebrán, Magmatic plumbing
and dynamic evolution of the 2021 La Palma eruption. Communications 14, 358 (2023).
55. A. Klügel, M.-­A. Longpré, L. García-­Cañada, J. Stix, Deep intrusions, lateral magma transport
and related uplift at ocean island volcanoes. Earth Planet. Sci. Lett. 431, 140–149 (2015).
56. A. Klügel, H.-­U. Schmincke, J. D. L. White, K. A. Hoernle, Chronology and volcanology of
the 1949 multi-­vent rift-­zone eruption on La Palma (Canary Islands). J. Volcanol.
Geotherm. Res. 94, 267–282 (1999).
57. C. Szabó, R. J. Bodnar, Changing magma ascent rates in the Nógrád-­Gömör volcanic field,
northern Hungary/southern Slovakia: Evidence from CO2-­rich fluid inclusions in
metasomatized upper mantle xenoliths. Petrology 4, 221–230 (1996).
58. F. J. Spera, Carbon dioxide in petrogenesis III: Role of volatiles in the ascent of alkaline
magma with special reference to xenolith-­bearing mafic lavas. Contrib. Mineral. Petrol. 88,
217–232 (1984).
59. A. Klügel, Reactions between mantle xenoliths and host magma beneath La Palma
(Canary Islands): Constraints on magma ascent rates and crustal reservoirs. Contrib.
Mineral. Petrol. 131, 237–257 (1998).
60. R. T. Helz, C. R. Thornber, Geothermometry of Kilauea Iki lava lake, Hawaii. Bull. Volcanol.
49, 651–668 (1987).
61. R. Span, W. Wagner, A new equation of state for carbon dioxide covering the fluid region
from the triple point temperature to 1100 K at pressures up to 800 MPa. JPCRD 25,
1509–1596 (1996).
62. S. M. Sterner, K. S. Pitzer, An equation of state for carbon dioxide valid from zero to
extreme pressures. Contrib. Mineral. Petrol. 117, 362–374 (1994).
63. S. M. Sterner, R. J. Bodnar, Synthetic fluid inclusions; X, Experimental determination of
P-­V-­T-­X properties in the CO 2 -­H 2 O system to 6 kb and 700 degrees C. Am. J. Sci. 291,
1–54 (1991).
64. R. Thiery, J. Vidal, J. Dubessy, Phase equilibria modelling applied to fluid inclusions:
Liquid-­vapour equilibria and calculation of the molar volume in the CO2 CH4 N2 system.
Geochim. Cosmochim. Acta 58, 1073–1082 (1994).
Zanon et al., Sci. Adv. 10, eadi4300 (2024) 7 February 2024
65. J. M. Castro, Y. Feisel, Eruption of ultralow-­viscosity basanite magma at Cumbre Vieja, La
Palma, Canary Islands. Nat. Commun. 13, 3174 (2022).
66. E. A. J. Burke, Raman microspectrometry of fluid inclusions. Lithos 55, 139–158 (2001).
67. H. A. Roeser, “Magnetic anomalies in the magnetic quiet zone off Morocco” in Geology of
the Northwest African Continental Margin, U. Rad, K. Hinz, M. Sarnthein, E. Seibold, Eds.
(Springer,1982) pp. 61–68.
68. J. C. Carracedo, E. R. Badiola, H. Guillou, J. De La Nuez, F. J. Perez Torrado, 2001, Geology
and volcanology of La Palma and El Hierro, western Canaries. Estudios Geologicos 57,
175–273 (1982).
69. K. D. Klitgord, H. Schouten, “Plate kinematics of the Central Atlantic” in The Geology of
North America, v. M: The Western North Atlantic Region, P. R. Vogt, B. E. Tucholke, Eds.
(Geological Society of America, 1986), pp. 351–378.
70. N.-­O. Prægel, P. M. Holm, Lithospheric contributions to high-­MgO basanites from the
Cumbre Vieja volcano, La Palma, Canary Islands, and evidence for temporal variation in
plume influence. J. Volcanol. Geotherm. Res. 149, 213–239 (2006).
71. K. V. Cashman, R. S. J. Sparks, J. D. Blundy, Vertically extensive and unstable magmatic
systems: A unified view of igneous processes. Science 355, (2017).
72. R. S. J. Sparks, C. Annen, J. D. Blundy, K. V. Cashman, A. C. Rust, M. D. Jackson, Formation
and dynamics of magma reservoirs. Phil. Trans. Royal Soc. A 377, 20180019 (2019).
73. A. Lodge, S. E. J. Nippress, A. Rietbrock, A. García-­Yeguas, J. M. Ibáñez, Evidence for
magmatic underplating and partial melt beneath the Canary Islands derived using
teleseismic receiver functions. Phys. Earth Planet. Inter. 212, 44–54 (2012).
74. C. Martinez-­Arevalo, F. de Lis Mancilla, G. Helffrich, A. Garcia, Seismic evidence of a
regional sublithospheric low velocity layer beneath the Canary Islands. Tectonophysics
608, 586–599 (2013).
75. E. Banda, J. J. Dan, E. Surin, J. Ansorge, Features of crustal structure under the Canary
Islands. Earth Planet. Sci. Lett. 55, 11–24 (1981).
76. L. Matias, N. A. Dias, I. Morais, D. Vales, F. Carrilho, J. Madeira, J. L. Gaspar, L. Senos,
A. B. Silveira, The 9th of July 1998 Faial Island (Azores, North Atlantic) seismic sequence. J.
Seismol. 11, 275–298 (2007).
77. J. Pim, C. Peirce, A. B. Watts, I. Grevemeyer, A. Krabbenhöft, Crustal structure and origin of
the Cape Verde Rise. Earth Planet. Sci. Lett. 272, 422–428 (2008).
78. C. R. Ranero, M. Torné, E. Banda, Gravity and multichannel seismic reflection constraints
on the lithospheric structure of the Canary Swell. Mar. Geophys. Res. 17, 519–534 (1995).
79. C. J. Lissenberg, C. J. MacLeod, E. N. Bennett, Consequences of a crystal mush-­dominated
magma plumbing system: A mid-­ocean ridge perspective. Phil. Trans. R. Soc. A. 377,
20180014 (2019).
80. K. Benn, A. Nicolas, I. Reuber, Mantle – crust transition zone and origin of wehrlitic
magmas: Evidence from the Oman Ophiolite. Tectonophysics 151, 75–85 (1988).
81. F. Boudier, A. Nicolas, Nature of the Moho transition zone in the Oman ophiolite. J. Petrol.
36, 777–796 (1995).
82. D. E. James, F. Niu, J. Rokosky, Crustal structure of the Kaapvaal craton and its significance
for early crustal evolution. Lithos 71, 413–429 (2003).
83. B. Ghosh, T. Morishita, B. S. Gupta, A. Tamura, S. Arai, D. Bandyopadhyay, Moho transition
zone in the Cretaceous Andaman ophiolite, India: A passage from the mantle to the
crust. Lithos 198-­199, 117–128 (2014).
84. J. A. Crisp, Rates of magma emplacement and volcanic output. J. Volcanol. Geotherm. Res.
20, 177–211 (1984).
Acknowledgments: The Dirección General de Seguridad y Emergencias del Gobierno de
Canarias is acknowledged for allowing access to the exclusion zone during the eruption. We
thank the reviewers for their comments on an earlier version of this manuscript. Funding: This
work was partially funded by the Fundação para a Ciência e Tecnologia (FCT), Portugal,
through the MAGAT project (ref. CIRCNA/OCT/2016/2019). It was also partially supported by
the projects VOLRISKMAC (MAC/3.5b/124) and VOLRISKMAC II (MAC2/3.5b/328), cofinanced by
the INTERREG V A Spain-­Portugal MAC 2014-­2020 program, European Commission and by the
project “Cumbre Vieja Emergencia,” financed by the Spanish Ministry of Science and
Innovation. V.Z. is a visiting researcher at the Laboratoire Magmas et Volcans (Clermont
Ferrand, France), funded by ClerVolc program. This is contribution no. 635 of the ClerVolc
program of the International Research Center for Disaster Sciences and Sustainable
Development of the University of Clermont Auvergne. Author contributions:
Conceptualization: V.Z. and L.D. Investigation: V.Z., F.S., L.D., K.C., and M.J.P. Visualization: V.Z.,
F.S., and L.D. Funding acquisition: V.Z., F.S., and L.D. Writing—original draft: V.Z. Writing—
review and editing: V.Z., F.S., L.D., and K.C. Validation: V.Z., F.S., and L.D. Competing interests:
The authors declare that they have no competing interests. Data and materials availability:
All data needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials.
Submitted 25 April 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adi4300
14 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

AT M O S P H E R I C S C I E N C E Copyright © 2024 The

Authors, some rights
Climate-­invariant machine learning reserved; exclusive
licensee American
Association for the
Tom Beucler1,2*, Pierre Gentine3, Janni Yuval4, Ankitesh Gupta2, Liran Peng2, Jerry Lin2, Advancement of
Sungduk Yu2, Stephan Rasp5, Fiaz Ahmed6, Paul A. O’Gorman4, J. David Neelin6, Science. No claim to
Nicholas J. Lutsko7, Michael Pritchard2,8 original U.S.
Government Works.
Projecting climate change is a generalization problem: We extrapolate the recent past using physical models Distributed under a
across past, present, and future climates. Current climate models require representations of processes that occur Creative Commons
at scales smaller than model grid size, which have been the main source of model projection uncertainty. Recent Attribution License 4.0
machine learning (ML) algorithms hold promise to improve such process representations but tend to extrapolate (CC BY).
poorly to climate regimes that they were not trained on. To get the best of the physical and statistical worlds, we
propose a framework, termed “climate-­invariant” ML, incorporating knowledge of climate processes into ML algo-
rithms, and show that it can maintain high offline accuracy across a wide range of climate conditions and configu-
rations in three distinct atmospheric models. Our results suggest that explicitly incorporating physical knowledge
into data-­driven models of Earth system processes can improve their consistency, data efficiency, and generaliz-
ability across climate regimes.

INTRODUCTION
Background
Following its success in computer vision and natural language pro-
cessing, machine learning (ML) is rapidly percolating through climate
science [e.g., reviews by (1–5)]. We use the term ML here to broadly
describe algorithms that learn a task from data without being explic-
itly programmed for that task. Applications of ML in atmospheric sci-
ence include the emulation of radiative transfer algorithms [e.g.,
(6–9)], momentum fluxes [e.g., (10–13)] and microphysical schemes
[e.g., (14–16)], the bias correction of climate predictions [e.g., (17,
18)], the detection and classification of clouds and storms [e.g., (19–
22)], and the development of subgrid-­scale “closures” (i.e., representa-
tion based on coarse-­ scale processes only) from high-­ resolution
simulation data [e.g., (23–25)], which is the main application dis-
cussed here.
ML algorithms typically optimize an objective on a training
dataset and make implicit assumptions when extrapolating. Here, ex-
trapolation refers to predictions outside of the training data range,
henceforth referred to as out-­of-­distribution predictions. As an ex-
ample, multiple linear regressions (MLR) assume that the linear rela-
tionship that best describes the training set is valid outside of that
training set. Alternatively, when confronted with out-­of-­distribution
inputs, random forests (RFs) (26) find the closest inputs in their train-
ing sets and assign the corresponding outputs regardless of the out-­of-­
distribution input values. Neural networks (NNs), which are powerful
nonlinear regression and classification tools, rely on nonlinear activa-
tion functions and fitted weights to extrapolate. Except in specific sit-
uations (e.g., samples in the close neighborhood of the training set or
described by the same nonlinear mapping as the training set), there is
no reason why NNs should generalize well far outside of their training
1
Faculty of Geosciences and Environment, University of Lausanne, Lausanne, VD
1015, Switzerland. 2Department of Earth System Science, University of California,
Irvine, CA 92697, USA. 3Department of Earth and Environmental Engineering, Columbia
University, New York, NY 10027, USA. 4Department of Earth, Atmospheric, and
Planetary Sciences, Massachusetts Institute of Technology, Cambridge, MA 02139,
USA. 5Google Research, Mountain View, CA 94043, USA. 6Department of Atmo-
spheric and Oceanic Sciences, University of California, Los Angeles, Los Angeles, CA
90095, USA. 7Scripps Institution of Oceanography, University of California, San Diego,
La Jolla, CA 92037, USA. 8NVIDIA, Santa Clara, CA 95050, USA.
*Corresponding author. Email: tom.​beucler@​gmail.​com
sets. We show later that different NN training approaches on the same
data can lead to drastically different out-­of-­distribution predictions,
highlighting the uncertainty associated with such predictions.
In climate applications, this extrapolation issue means that ML al-
gorithms typically fail when exposed to dynamic, thermodynamic, or
radiative conditions that differ substantially from the range of condi-
tions that they were trained on. Examples include O’Gorman and
Dwyer (27), who showed that an RF-­based moist convection scheme
generalizes poorly in the tropics of a climate 6.5 K warmer than the
training climate, and Hernanz et al. (28), who showed that NNs and
support vector machines downscaling surface air temperature made
substantial extrapolation errors when exposed to temperatures 2-­3 K
warmer than in the training set. Rasp et al. (29) showed that an NN-­
based thermodynamic subgrid-­scale closure generalizes well to cli-
mates 1 to 2 K warmer than the training one but makes large errors as
soon as the test climate is 4 K warmer than the training one.
Beucler et al. (30) confirmed that these generalization errors remain
even when the NN subgrid closure is modified to enforce conserva-
tion laws to within machine precision. This has led several studies to
recommend training ML models in multiple climates if possible (31,
32). Both Guillaumin and Zanna (33) who trained an NN parameter-
ization for subgrid oceanic momentum transport and Molina et al.
(34) who trained convolutional NNs (35) to classify thunderstorms in
high-­resolution model outputs found that their ML models general-
ized well to a warmer climate. While this may be because both models
relied heavily on velocity inputs and their gradients, whose distribu-
tions changed only slightly when the climate warmed, Molina et al.
noted that using two types of ML layers, namely, batch normalization
(BN) (36) followed by dropout (DP) (37), was key to this successful
generalization.
DP and BN are two examples of a larger set of methods that help
NNs generalize and avoid overfitting, broadly referred to as “regular-
izations” (38). Most empirical regularization methods (e.g., L1 regu-
larization) rely on the parsimony principle, i.e., that simpler models,
accurately describing the training set with fewer fitted parameters, are
preferable to more complex models and generalize better to unseen
conditions. More systematic approaches to regularization have been
developed to use ML models in out-­of-­distribution situations that still
require the same inputs/outputs, referred to as domain adaptation

Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024 1 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
[e.g., (39–41)], a particular case of transfer learning [e.g., (42)]. While
not all domain adaptation approaches (sample-­based, statistics-­based,
ensemble-­based, domain-­invariant feature learning, domain map-
ping, etc.) need supervision (43), they usually require at least a few
samples in the generalization domain.
Without dismissing existing domain adaptation methods, we here
focus on physically informed methods that do not require samples in
the generalization domain for three reasons: (i) one of the climate sci-
ence community’s long-­term goals is to train ML models that rely on
historical observations only as we cannot, by definition, observe the
future climate; (ii) as shown later, even if we have access to simulation
data across climates, ML models that intrinsically generalize to cli-
mates that they have not been trained on tend to be more data-­efficient
and robust to other changes (e.g., configuration changes); and (iii)
physically informed methods can be readily combined with existing
domain adaptation and regularization methods. Motivated by these
challenges, we ask: How can we enhance ML algorithms with physical
knowledge to make accurate predictions in climate conditions that, in
standard variables, lie far outside of the training set?
Problem definition
Our scientific contribution is to transform a mapping constructed us-
ing the original data’s features, henceforth referred to as “raw-­data”
mapping, into a mapping that remains nearly constant across cli-
mates, here referred to as “climate-­invariant.” Inspired by invariants in
physics and self-­similarity in fluid mechanics (44), we make the ML-­
emulated mapping climate-­invariant by transforming the input and
output vectors so that their distributions shift minimally across differ-
ent climates (see Fig. 1). We demonstrate this framework’s utility by
adapting ML closures of subgrid atmospheric thermodynamics (i.e.,
coarse-­scale thermodynamic tendencies resulting from subgrid con-
vection, radiation, gravity waves, and turbulence) so that they gener-
alize better across climates.
The motivation for this application is twofold. First, purely physi-
cally based subgrid closures remain one of the largest sources of un-
certainties in Earth system models (45–47). While ML-­based closures
have emerged as a promising alternative to traditional semiempirical
models (48), they lack robustness (49, 50) and, as discussed earlier,
usually fail to generalize across climates (27, 29, 30). Second, atmo-
spheric thermodynamic processes are directly affected by global tem-
perature changes, e.g., in response to anthropogenically forced climate
change (51). Therefore, predicting subgrid thermodynamics in a
warm climate with an ML model trained in a cold climate leads to
very apparent failure modes (30) that we can transparently tackle.
In mathematical terms, our goal is to build a climate-­invariant
mapping between the input vector x representing the large-­scale
(≈100 km) climate state and the output vector y grouping large-­scale
thermodynamic tendencies due to explicitly resolved convection and
parameterized radiative transfer and turbulent mixing at the ∼1-­km
scale [see section SB1 for details]. We keep the overall structure of
the mapping x ↦ y fixed throughout the manuscript and aim to pre-
dict y as accurately as possible in training and generalization climates
(out-­of-­distribution prediction). Note that this mapping makes some
implicit assumptions based on successful past work (29, 52), includ-
ing locality in horizontal space and time (outputs only depend on
inputs in the same atmospheric column at the same time step) and
determinism (only one possible output vector for a given input vec-
tor). We include cloud radiative effects in all heating terms (total
heating Ṫ , longwave heating lw, and shortwave heating sw) but, for
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
simplicity, do not predict changes in cloud liquid water and ice and
exclude cloud water and greenhouse gases other than water vapor qv
from the input vector x.
After introducing the climate simulations and training/validation/
test split (see Data), we define the climate-­invariant mapping and fea-
ture transformations (see Theory) and demonstrate and explain their
ability to generalize (see Results) before concluding. We refer the
reader to the Supplementary Materials for data availability (section
SA), additional derivations and descriptions of the mapping and
physical transformations (section SB), the implementation of our ML
framework (section SC), and additional results (section SD).
DATA
To test the robustness of our framework across model formulations
and configurations, we use three distinct storm-­resolving climate
models and experimental setups: aquaplanet simulations using the
Super-­Parameterized Community Atmosphere Model version 3.0
(SPCAM3), Earth-­like simulations (i.e., with continents) using the
Super-­Parameterized Community Earth System Model version 2
(SPCESM2), and quasi-­global aquaplanet hypohydrostatic simula-
tions using the System for Atmospheric Modeling version 6.3
(SAM). SPCAM3 and SPCESM2 assume a strict scale separation
between the resolved coarse scales and subgrid processes, making
them ideal testbeds to machine learn local subgrid closures (29,
53). In contrast, SAM does not assume scale separation as a global
storm-­resolving model. This improves realism but requires coarse-­
graining SAM’s output for ML parameterization purposes (54, 55).
For each climate model, we run three simulations with three differ-
ent prescribed surface temperature distributions: (i) (+0 K) a refer-
ence simulation with a temperature range analogous to the present
climate, (ii) (−4 K) a cold simulation with surface temperatures 4 K
cooler than the (+0 K) simulation, and (iii) (+4 K) a warm simula-
tion with surface temperatures 4 K warmer than the (+0 K) simula-
tion, with the exception of SAM for which only the (−4 K) and (+0 K)
simulations are available. By prescribing surface temperature, we
focus on ML’s ability to consistently predict the atmospheric re-
sponse to climate change across configurations. Projecting climate
change involves a broader range of processes and is beyond this
work’s scope. We summarize the simulations and indicate their spa-
tiotemporal resolutions in table S1. Figure 2 gives a visualization of
surface temperatures in each model, and fig. S1 provides snapshots
of mid-tropospheric subgrid heating, which is one of our ML mod-
els’ outputs.
Super-­parameterized aquaplanet simulations
We use data from 2-­year SPCAM3 (56) climate simulations in an
aquaplanet configuration (57), with zonally symmetric surface tem-
peratures fixed to a realistic meridionally asymmetric profile (58). The
insolation is fixed to boreal summer conditions with a full diurnal
cycle. A two-­dimensional storm-­resolving model is embedded in
each grid cell of SPCAM3, namely, eight SAM atmospheric columns
using a spatiotemporal resolution of 4 km × 30 levels × 25 s and the
default one-­moment microphysical scheme (59). SPCAM3 combines
a spectral primitive equation solver with a semi-­Lagrangian dynami-
cal core for advection (57). The (+0 K) SPCAM3 simulation was first
presented in (53) and subsequently used to train ML subgrid closures
in (29, 50, 60). Inspired by the generalization experiment of (27), the
(+4 K) simulation was introduced in (29), and we ran the (−4 K)
2 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. By transforming inputs x and outputs y to match their probability density functions across climates, the algorithms can learn a transformed mapping 𝛟̃
that holds across climates. To illustrate this, we show the marginal distributions of inputs and outputs in two different climates using blue and red lines, before (top) and
after (bottom) the physical transformation.

simulation for the work presented here to increase the surface tem- simulation is similar to that in (52), which showed the potential of ML
perature generalization gap from 4 to 8 K. for subgrid closures in Earth-­like conditions.

Super-­parameterized Earth-­like simulations Quasi-­global aquaplanet hypohydrostatic simulations

We run three 2-­year SPCESM2 (61) climate simulations in an Earth-­
like configuration with realistic surface boundary conditions, includ-
ing a land surface model, seasonality, aerosol conditions representative
of the year 2000, and a zonally asymmetric annual climatology of sea
surface temperatures derived from the “HadOIB1” dataset (62). We
use CESM v2.1.3 to couple CAM v4.0 with the Community Land
Model version 4.0 and similarly embed 32 SAM columns in each
atmospheric grid cell to explicitly represent deep convection. Our (+0 K)
While super-­parameterization is well adapted to statistically learning
subgrid closures due to its explicit scale separation, this scale separa-
tion comes at the cost of distorted mesoscale systems and momentum
fluxes (63). Furthermore, most ML subgrid closures are based on
coarse-­graining high-­resolution simulations [e.g., (64, 65)]. This mo-
tivates us to also test the climate-­invariant framework in hypohydro-
static SAM simulations in which the dynamics are not affected by a
prescribed scale separation. Computational expense is reduced through

Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024 3 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

training set, and evaluate the final model on samples from a separate
test set. We split each of the eight simulations into training/validation/
test sets by using noncontiguous 3-­month periods (reported in
table S1) to avoid high temporal correlations between training/
validation/test set samples (71). Following (29), the normalization
procedure involves subtracting the mean value of each input vari-
able at each vertical level and dividing by the maximum range of
that variable across the entire atmospheric column.
To understand which solutions are most promising for helping ML
algorithms generalize to unseen conditions, we design two generaliza-
tion experiments: (i) training and validating ML models on cold sim-
ulations (−4 K) and testing them on warm simulations (+4 K for
SPCAM3/SPCESM2 and +0 K for SAM); and (ii) training and vali-
dating ML models on aquaplanet simulations (SPCAM3) and testing
them on Earth-­like simulations with continents (SPCESM2).
Fig. 2. Surface temperatures in the three used atmospheric models. Prescribed
surface temperature (in kelvin) for (left) the aquaplanet SPCAM3 model and (right)
the hypohydrostatic SAM model. (Center) Annual-­mean, near-­surface air tempera-
tures in the Earth-­like SPCESM2 model. THEORY
We formally define a climate-­invariant mapping as a mapping that is
unchanged across climates. In practice, it is difficult to find mappings
that are exactly invariant, and we will use the terminology climate-­
hypohydrostatic scaling, which multiplies the vertical acceleration in invariant for any mapping that remains approximately constant across
the equations of motion by a factor of 16 to increase the horizontal climates. To achieve climate invariance, we introduce physically based
scale of convection without overly affecting the larger-­scale flow (66, feature transformations, defined as physically informed functions that
67). While these simulations use idealized settings, such as aquaplanet map the inputs/outputs to different inputs/outputs whose distribu-
configurations, an anelastic dynamical core, a quasi-­global equatorial tions vary little across climates. We deem the physical transformation
beta plane domain, and perpetual equinox without a diurnal cycle, to be climate-­invariant if it is successful at limiting distributions varia-
O’Gorman et al. (68) showed that they produce tropical rainfall inten- tions of the inputs/outputs across climates. Note that climate-­invariant
sity and cluster-­area distributions that are close to satellite observa- transformations are distinct from nondimensionalization in dimen-
tions. The prescribed surface temperature distribution in the control sional analysis, as nondimensionalization does not necessarily alter
simulation of (68) is designed to be close to zonal-­mean observations distribution shape while climate-­invariant transformations may yield
(69), and its maximum value is roughly 2 K colder than that of the variables that have physical units.
distribution used for the (+0 K) SPCAM3 simulation. To better match Throughout the following section, we compare two transforma-
the SPCAM3 maxima of distributions of upper-­level temperatures tion options for each input, whose univariate Probability Density
and humidities, we choose to treat this SAM control simulation as the Functions (PDFs) are depicted for all three atmospheric models in
(−4 K) SAM simulation and the warm simulation of (68) as the (+0 K) Fig. 3: no transformation (top) and our most successful transforma-
SAM simulation. We refer the reader interested in the details of the tion (bottom). All transformations are derived in section SB2. Our
simulations and the coarse-­graining (here by a factor of 8) to (54). comparison relies on the Hellinger and Jensen-­Shannon PDF distance
Differences in climate model formulation and ML parameterization metrics defined and calculated in Materials and Methods and section
design lead to key differences in the mappings learned for SAM as SD1. To prevent information leaks from generalization test sets into
compared to SPCAM3/SPCESM2, which we summarize below: (i) the physically informed ML framework, we take two precautions: (i)
The input vector does not contain specific humidity, surface pres- the physical transformations are fixed, meaning that their structure
sure, sensible heat fluxes, or latent heat fluxes (LHFs) but instead and parameters are non-­trainable; and (ii) transformations are ranked
contains the total non-precipitating water concentration and uses on the basis of their generalization from (−4 K) to (+0 K) in SPCAM3.
distance to the equator as a proxy for solar insolation. (ii) The output Our (+4 K) results across models and configurations independently
vector includes the subgrid total non-precipitating water tendency confirm this ranking.
instead of the subgrid specific humidity tendency and the subgrid
liquid/ice static energy tendency instead of the subgrid temperature Specific humidity
tendency. (iii) The output vector does not contain subgrid longwave Without any transformation, the PDF of specific humidity q (Fig. 3A,
and shortwave heating. (iv) SAM uses a height-based vertical coor- top) extends through a considerably larger range as the climate
dinate rather than a pressure-based one. (v) the generalization ex- warms. This is because, barring supersaturation, q has a theoretical
periment is from (−4 K) to (+0 K) [unavailable (+4 K) simulation]. upper bound in a given climate, namely, the saturation specific hu-
midity, which increases quasi-­exponentially with temperature through
Normalization and training, validation, and test split the Clausius-­Clapeyron relation [e.g., (72, 73)]. The relative humidity
Both generalization experiments expose ML models to out-­ of-­ (RH) transformation q̃ RH (Fig. 3A, bottom) normalizes specific hu-
distribution inputs that they have not been trained on. Following midity by its saturation value. As a result, most of the RH PDF lies
best ML practices (70), we use the training set to optimize the ML within [0,1], except for a few atmospheric columns exhibiting
model’s trainable parameters, save the trainable parameters that led supersaturation in SPCAM, and that PDF changes little as the climate
to the best performance on the validation set to avoid overfitting the warms (74). In addition to capturing grid-­scale saturation, q̃ RH helps

Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024 4 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Physical transformations can align distributions across climates. We show the univariate distributions of selected raw inputs x: (A) 600-­hPa specific humidity;
(B) 850-­hPa temperature; and (C) latent heat flux (LHF) in the cold (blue), reference (gray), and warm (red) simulations of each model (SPCAM3, SPCESM2, and SAM). For
each variable, we also show the PDFs of the transformed inputs x̃ as discussed in the Theory section. From top to bottom, the variables are q (grams per kilogram), relative
humidity (RH), T (kelvin), Bplume (meters per square second), LHF (watts per square meter), and LHFΔq (kilograms per square meter per second). For a given variable and
transformation, we use the same vertical logarithmic scale across models.

predict the subgrid effects of dry-­air entrainment, known to regulate
tropical convection (75, 76, 77) (see section SB2b for details of RH
calculations).
Temperature
The PDF of temperature T (Fig. 3B, top) shifts quasi-­linearly as the
climate warms. To address this shift without compromising the ap-
proximate invariance of tropopause temperatures with warming (78,
79, 80), we derive a temperature transformation directly relevant for
moist convection: the buoyancy of a non-­entraining, moist static
energy-­conserving plume T̃ buoyancy (Fig. 3B, bottom, see section SB2c
for this buoyancy’s derivation). This transformation is inspired by re-
cently introduced lower-­tropospheric buoyancy measures (81, 82), but
with an extension to the full troposphere (83). While T̃ buoyancy does
not explicitly include entrainment effects, the mapping of T̃ buoyancy (p)
and q̃ RH (p) to heating and moisture sink will implicitly include these.
This transformation captures leading order effects needed to yield
approximate climate invariance (Fig. 3B). T̃ buoyancy increases physical
interpretability by linking the vertical temperature structure and
near-­surface humidity changes to a metric that correlates well with
deep convective activity (84). T̃ buoyancy also captures the role of near-­
surface humidity relative to the temperature structure aloft in contrib-
uting to moist convective instability in the tropics.
Latent heat flux
The last input whose distribution changes visibly with warming
is the LHF (Fig. 3C, top; the remaining inputs, sensible heat
flux and surface pressure, change less with warming and are
discussed in section SB2d). Similar to specific humidity, the in-
crease of LHFs with warming is directly linked to the Clausius-­
Clapeyron relationship [e.g., (85)]. To address this shift, we
leverage the bulk aerodynamic formula to represent surface
fluxes and to provide a physics-­motivated transformation of
LHF using the near-­surface saturation deficit (Fig. 3C, bottom).

Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024 5 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

This transforms LHF, a thermodynamic variable, into L HF ̃ Δq , Benefits of incremental input transformations
approximately proportional to the magnitude of near-­surface In this section, we demonstrate that incrementally transforming the
horizontal winds and density [e.g., (85)], whose distributions inputs of NNs progressively improves their generalization abilities
vary less with warming. Note that this scaling is less effective from the cold (−4 K) aquaplanet (SPCAM3) simulation to the warm
over land (e.g., in SPCESM2) where evapotranspiration changes (+4 K) aquaplanet simulation. The largest surface temperature jump
do not follow a Clausius-­Clapeyron scaling. tested in this study is between the cold aquaplanet simulation and the
We now show that all three input transformations [̃qRH (p), tropics of the warm aquaplanet simulation (“warm tropics” for short),
̃ Δq] lead to statistically significant improve- defined as the regions with out-­of-­distribution surface temperatures,
T̃ buoyancy (p), and L HF
whose latitudes are between −15°S and 23°N (approximately the red
ments in the ML models’ ability to generalize.
regions in top-­left subplots in Fig. 2). To expose the failure modes of
the “brute force” model and the benefits of progressively transforming
RESULTS the inputs, we first trained several NNs on the cold aquaplanet until
The results are organized as follows. After demonstrating the bene- they reached high accuracy (Fig. 4A) before testing their out-­of-­
fits of progressively transforming the ML models’ inputs (Fig. 4), we distribution generalizability in the warm tropics (Fig. 4B).
show how climate-­invariant models learn subgrid closures across In the cold tropics, the vertical profiles of the mean-­squared error
climates and configurations during training (Fig. 5 and fig. S7). We (MSE) are nearly indistinguishable for all types of NNs and roughly
then discuss the global skill of different models after training (Fig. 6 follow the vertical structure of subgrid variance, as discussed in (53,
and figs. S8 and S9). Last, we investigate the structure of climate-­ 60). When evaluated in the warm tropics, the MSE of the brute force
2 −4
invariant mappings to understand why they generalize better across NN (blue line) increases by a factor of ≈10 and peaks above 100 W m ,
climates (Fig. 7), even when data from multiple climates are avail- underlining how raw-­data NNs fail to generalize across climates. As
able (Fig. 8). discussed in Theory, we progressively transformed the inputs starting
with specific humidity, which is transformed to RH (orange line). This

A B

Fig. 4. All neural networks (NNs), trained in the cold climate, exhibit low error in the cold climate’s test set, but much larger error in the warm climate’s test set.
(A) Low error in the cold climate’s test set. (B) Larger error in the warm climate’s test set. This generalization error decreases as inputs are incrementally transformed: first
no transformation (blue), then the vertical profile of specific humidity (orange), then the vertical profile of temperature (green), and lastly LHFs (red). For reference, the
purple line depicts an NN trained in the warm climate. We depict the tendencies’ mean-­squared error (MSE) versus pressure, horizontally averaged over the tropics of
SPCAM3 aquaplanet simulations, for the four model outputs: total moistening (q̇ ), total heating (Ṫ ), longwave heating (lw), and shortwave heating (sw). Given that the
raw-­data NN’s generalization error (blue line) greatly exceeds that of the transformed NNs, we zoom in on each panel to facilitate visualization.

Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024 6 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 5. Model error across temperatures and configurations. MSE (in W2
m−4) of six models trained in three simulations (first column) and evaluated
over the training or validation set of the same and two other simulations (last
four columns). The models (second column) are raw-­data (RD) or climate-­
invariant (CI), and MLRs or neural nets (NN), and sometimes include DP layers
preceded by a BN layer (DN). The models are trained for 20 epochs. We first
provide the MSE corresponding to the epoch of minimal validation loss, then
the MSE averaged over the five epochs with lowest validation losses (in paren-
theses), and lastly the MSE divided by the baseline MSE, where we use the
raw-­data MLR as baseline. Note that “different temperature” refers to (+4 K) for
(−4 K) training sets and vice versa. In each application case, we highlight the
best model’s error using bold font.
BA
Fig. 6. Climate-invariant NNs address the raw-data NNs’ generalization problems in the warm tropics. This is demonstrated by the 500-hPa subgrid heating’s coef-
ficient of determination R2 calculated over the test set for the raw-data (A) and climate-invariant (B) NNs. We train NNs using the cold (−4 K) training set of each model
(SPCAM3, SPCESM2, and SAM). We note that these NNs do not use DP nor BN, and we refer the readers to fig. S8 for latitude-pressure cross sections.
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
first transformation decreases the MSE so much (by a factor of 5 to
10) that we need to zoom in on each panel to distinguish the general-
ization abilities of additional NNs. Adding the transformations of
temperature to plume buoyancy (green line) improves the generaliza-
tion MSE for all variables. Adding the LHF to L HF ̃ Δq transformation
(red line) further decreases generalization MSE, except for shortwave
heating where the MSE improves in the cold but not the warm cli-
mate. Impressively, transforming all three inputs decreases the MSE
so much that the resulting climate-­invariant NN’s MSE (red line) is
within ≈25% of the MSE of a raw-­data NN that was directly trained in
the warm climate (purple line).
Hereafter, we use climate-­invariant to refer to models for which all
three inputs (q, T, and LHF), but no outputs were transformed, solely based
on physical principles. After demonstrating their success in the aquaplanet
case, we are now ready to investigate how these climate-­invariant models
learn in the other climates and simulations introduced in Data.
Learning across climates and configurations
In this section, we show that climate-­invariant models learn mappings
that are valid across climates and configurations and that their efficacy
improves when used in conjunction with ML regularization tech-
niques like BN and DP layers.
Figure 5 shows the MSE of ML models trained in three different
datasets and evaluated over their training and validation sets and
over test sets of different temperature and configuration. As discussed
7 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
previously, climate-­invariant NNs (NN CI) generalize better to warm-
er climates than raw-­data NNs (NN RD). We go one step further
by examining learning curves, defined as the MSE of an ML model at
the end of each epoch during training (one epoch corresponds to the
ML model being fed the entire training set once). Impressively, the
learning curve of the climate-­invariant NN trained in the cold aqua-
planet but tested in the warm aquaplanet (starred blue line in fig. S7A)
is mostly decreasing, supporting this manuscript’s key result: Climate-­
invariant NNs continuously learn about subgrid thermodynamics in
A B
Fig. 7. Explainable artificial intelligence suggests that climate-­invariant map-
pings are more spatially local. We depict ℳ for the (A) raw-­data and (B) climate-­
invariant NNs trained in the SPCAM3 (+4 K) warm aquaplanet simulation. The x
axes indicate the inputs’ vertical levels, from the surface (left, 103 hPa) to the top of
the atmosphere (right, 0 hPa), while the y axes indicate the outputs’ vertical levels,
from the surface (bottom, 103 hPa) to the top of the atmosphere (top, 0 hPa). We
additionally indicate the 200-­hPa vertical level with black dashed lines.
Fig. 8. Climate-­invariant (CI) NNs trained on datasets containing both cold (−4 K) and warm (+4 K) samples outperform raw-­data (RD) models offline in ≈95% of
cases, with less sensitivity to the data partition used for training. Dots on the left represent the median RD error from a 10-­fold cross-­validation without replacement,
with horizontal ticks indicating the first and ninth deciles. Stars on the right correspond to the median CI error. Ticks denote the majority of cases, for which CI models
outperform RD models, even when data from both climates are available; crosses indicate the rare exceptions. We use a logarithmic scale for both axes.
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
the warm aquaplanet as they are trained in the cold aquaplanet. In
contrast, the raw-­data NN trained in the cold aquaplanet but tested in
the warm aquaplanet makes extremely large generalization errors,
which worsen as the model is trained in the cold aquaplanet (see sec-
tion SD2 for details).
Climate-­invariant NNs also facilitate learning across configura-
tions, i.e., from the aquaplanet to the Earth-­like simulations and vice
versa (see NN CI rows in Fig. 5). Climate-­invariant transformations
additionally improve the MLR baseline’s generalization ability, albeit
less markedly. This smaller improvement in MLR’s generalization
abilities is linked to its relatively small number of trainable parame-
ters, resulting in (i) raw-­data MLRs generalizing better than raw-­data
NNs; and (ii) MLRs having lower descriptiveness and fitting their
training sets less well, limiting the maximal accuracy of climate-­
invariant MLRs on test sets.
There are a few cases in which transforming inputs does not fully
solve the generalization problem, e.g., when trying to generalize from
the aquaplanet to the Earth-­like simulation. In that case, we leverage
the fact that input transformations can easily be combined with stan-
dard techniques to improve generalization, such as DP layers before
each activation function and a single BN layer before the first DP
layer (36). As DP layers randomly drop a fixed proportion of the
trainable parameters during training, NNs with DP fit their training
set less well (see NN CI + DN row of Fig. 5). However, they improve
generalization in difficult cases (e.g., between cold aquaplanet and
Earth-­like simulations) and do not overly deteriorate generalization
in cases where the input transformations work particularly well (e.g.,
from cold to warm aquaplanet). Our results suggest that combining
physics-­guided generalization methods (e.g., physical transformation
of the inputs/outputs) with regularization methods (e.g., DP) is ad-
vantageous and deserves further investigation. After analyzing the
overall MSE during training, we now turn to the spatial characteristics
of our ML models’ skill after training.
8 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Global performance after training
In this section, we first compare the spatial characteristics of the brute
force and climate-­invariant NNs’ skill across climates of different tem-
peratures to further establish the advantages of our climate-­invariant
transformation. These advantages are clearly visible in Fig. 6, where
the raw-­data models struggle to generalize to the warm tropics for all
simulations despite fitting the cold training set well (Fig. 6A). We can
trace these generalization errors to warm temperature and moist at-
mospheric conditions the NNs were not exposed to during training,
visible when comparing Fig. 6A with Fig. 2A. In contrast, the climate-­
invariant models fit the warm climate almost as well as the cold cli-
mate that they were trained in (Fig. 6B). Note that Fig. 6 focuses on
the horizontal map of a single output, i.e., the total subgrid heating at
500 hPa, but that the horizontal maps of other outputs, such as the
near-­surface subgrid heating (see fig. S9), all exhibit the same pattern
of raw-­data models failing in the warm tropics and the climate-­
invariant models mostly correcting these generalization errors. Last,
the spatial distribution of the skill in the training set (e.g., middle col-
umn of fig. S8) is reassuringly consistent with the skill map of highly
tuned NNs trained in similar conditions [e.g., (52)]. This confirms
that the raw-­data models, representative of state-­of-­the-­art ML sub-
grid closures, fail to generalize. This failure is confirmed in the
latitude-­pressure map of the subgrid heating at all vertical levels
shown in fig. S8 and discussed in section SD3.
To fully compare ML models across climate and configurations,
we evaluate their overall MSEs in the training, validation, and both
generalization test sets in Fig. 5. In addition to the MSE of minimal
validation loss, we show the MSE averaged over the five epochs of
minimal validation loss in parentheses to confirm that our models
have converged. Consistent with the learning curves in fig. S7, climate-­
invariant NNs with DP and BN layers often demonstrate the highest
level of generalizability (two rightmost columns of each row’s NN
CI + DN models).
While they fit their training sets less well, raw-­data MLRs gener-
alize better than raw-­data NNs because they have fewer trainable
parameters (see MLR RD and NN RD models). In Fig. 5, we also
show that, while DP and BN layers generally increase the generaliza-
tion performance of raw-­data NNs (NN RD + DN models), we can
systematically improve these standard ML regularization methods
by combining them with input transformations (NN CI + DN
models).
Understanding climate invariance
To interpret our NN results, we use a game theory–based explainable
ML approach, called SHapley Additive exPlanations (SHAP) (86, 87),
to dissect climate-­invariant mappings and provide insight into why
they generalize better across climates and configurations. Note that
MLRs are interpretable by construction, and we can draw preliminary
conclusions by visualizing MLR weights without the need for explain-
able ML libraries (see section SD4). Similarly, we can directly visualize
the NNs’ linear responses (80, 88–90) by calculating their Jacobians
(gradients) via automatic differentiation (49). However, the difference
between RD and CI MLRs is small and the Jacobians (88) cannot al-
ways be reliably used to explain nonlinear NN predictions as they are
first-­order derivatives calculated over a sample (91).
Therefore, as climate-­invariant NNs have shown superior general-
izability from cold to warm climates (see NN CI errors in Fig. 5), we
use SHAP’s KernelExplainer to elucidate the NNs’ generalizability. We
choose this attribution method for its versatility, as it can be used for
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
any multi-­input/output ML model. SHAP estimates the impact of a
particular input value xi on each output yj of our model. It is designed
with a local accuracy property, ensuring that the sum of the effects of
individual inputs equals the difference between yj and its average val-
ue in the training set
def
∑
SHAP(xi , yj ) = yj� (1)
i
where we have introduced the deviation yj′, defined as the difference
def
between yj and its training set average yi = yi − ⟨yi ⟩ℰ . We use these
�
“Shapley values” to build a nonlinear feature matrix M capturing the
influence of an input xi on an output yj
def
ℳ ij = ⟨sign(xi� ) × SHAP(xi , yj )⟩ℰ (2)
where we use the sign of the input deviation xi′ to make ℳij positive
if xi′ and yj′ have the same sign, e.g., if a positive input deviation leads
to a positive output deviation. In the particular case of the MLR de-
fined in eq. S25, the nonlinear feature matrix becomes the regres-
sion weight matrix multiplied by the absolute value of the input
deviations: ℳ ij = ⟨Aij∣xi�∣⟩ℰ , confirming that the feature matrix ℳ is
a nonlinear extension of the Jacobian (A in the MLR case).
In Fig. 7, each panel depicts the SHAP feature matrix M for a given
model: raw-­data (A) and climate-­invariant (B). Each model’s inputs
(e.g., specific humidity q and temperature T) are organized on the x
axis from the surface to the top of the atmosphere. Each model’s out-
puts [subgrid moistening q̇ and subgrid heating Ṫ ; see fig. S14 and
fig. S15 for subgrid longwave heating (lw) and subgrid shortwave
heating (sw)] are organized on the y axis, from the surface to the top
of the atmosphere. Following a horizontal line shows how different
inputs contribute to a given output, while following a vertical line
shows how a given input influences different outputs.
Figure 7 contains a wealth of information about subgrid closures
trained in aquaplanet simulations; we focus here on visualizing how
the climate-­invariant NNs (B) operate in ways that generalize better
than their raw-­data mapping counterpart (A). Consider the row for
subgrid heating Ṫ . In the raw-­data case (A), M has large coefficients in
most of the troposphere (in the entire square below the dashed lines
depicting the approximate tropopause level). This means that specific
humidity and temperature deviations at all levels affect subgrid heat-
ing at a given level, i.e., there are large nonlocal relations in the verti-
cal. Some nonlocal relations are physically plausible for convection
because buoyant plumes tend to rise from the surface, and near-­
surface T and q influence Ṫ through the entire troposphere. However,
in this model, moisture at higher altitudes appears to influence q̇ at
lower altitudes, raising suspicions that some of the raw-­data NN’s
nonlocalities are not causal but rather due to high autocorrelations
within the input’s vertical profile, as Brenowitz et al. (49) showed
could happen. Temperature variations are observed to have strong
vertical correlations (92) in part because of deep convective effects.
Because temperature affects the saturation threshold for moisture, the
RD NN will have to correctly capture the effects of both temperature
and moisture wherever either has influence. In contrast, in the Bplume-­
RH climate-­invariant case (B), leading nonlocal effects between the
boundary layer and the free troposphere have already been taken into
account in the buoyancy formulation, and the temperature-­dependent
saturation threshold is built into RH. Thus, M for Ṫ tends to be con-
centrated near the red diagonal, meaning that positive deviations of
9 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
plume buoyancy and RH increase subgrid heating near the same ver-
tical level. The use of domain knowledge has effectively reduced the
effects that must be estimated by the NN for the climate-­invariant
models. This tends to yield differences between the models trained in
the cold and the warm climates that are much smaller than for the
raw-­data models (see last column of fig. S15).
Advantages of climate invariance with multi-­climate data
It is natural to wonder whether the benefits of climate invariance
carry over to training scenarios that entail data from multiple simu-
lations spanning diverse temperatures. In contrast, until now, we
have only trained ML models on single-­climate simulations. To find
out, we examine the benefits of our climate-­invariant transformation
approach under the ideal scenario where we have access to data from
multiple climates. We conduct experiments where we train NNs on
both the cold (−4 K) and warm (+4 K) aquaplanet simulations. We
progressively increase the amount of training data to assess data ef-
ficiency. Throughout the experiment, we use eight batches and sys-
tematically increase the batch size by powers of 2, starting with a
batch size of 4. Note that we obtain similar results when increasing
the number of batches while keeping the batch size fixed (not shown).
To obtain well-­defined uncertainty estimates, we use a 10-­fold cross-­
validation procedure via random sampling without replacement.
Our findings, depicted in Fig. 8, demonstrate that CI NNs (i) consis-
tently outperform RD NNs, particularly in data-­rich scenarios; and
(ii) exhibit lower sensitivity to the training data partition, resulting in
more reliable offline performance with reduced variability in test
errors. This confirms that our climate-­invariant mapping enhances
data efficiency, performance, and fit reproducibility across different
climates, even when training data from multiple climates are available.
DISCUSSION
In the context of climate change, we hypothesized that ML models
emulating climate-­invariant mappings (Fig. 1), for which the inputs/
outputs distributions change little across climates (Fig. 3), generalize
much better than ML models emulating raw-­data mappings, for
which the inputs/outputs distributions change substantially across cli-
mates. Tested on a suite of storm-­resolving atmospheric simulations
with different surface temperatures in three atmospheric models with
distinct configurations (Fig. 2), physically transformed NNs general-
ize better as their inputs are progressively transformed (Fig. 4).
Climate-­invariant NNs whose inputs have all been transformed learn
mappings that are robust to temperature and configuration changes
(Fig. 5) and hence exhibit superior generalization skill almost every-
where on the globe (Fig. 6), including when data from multiple cli-
mates are available (Fig. 8). Last, attribution maps reveal that in
addition to providing control on the features’ distributions, climate-­
invariant NNs learn more spatially local mappings that facilitate gen-
eralization across climates and configurations (Fig. 7).
From a computational perspective, incorporating physical knowl-
edge, here of climate change, into an ML framework to improve its
generalization skill is a successful example of using domain knowl-
edge to extract more informative predictors, informally referred to as
“feature engineering” [e.g., (93)]. This also aids interpretability of the
mapping. From a climate science perspective, requiring that a nonlin-
ear statistical model of the atmosphere generalize across climate is
a stringent test that helped us discover new mappings. This climate-­
invariant mapping is more robust to climate and configuration changes
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
and is more advantageous than directly using model and observation-
al outputs (e.g., specific humidity and temperature), even when data
are available in various climate regimes. In the particular case of sub-
grid thermodynamics, our generalization results suggest the possibil-
ity of NN-­powered closures that could work in Earth-­like settings,
even in vastly different climate conditions. Last, the attribution maps
suggest the possibility of new analytic representations of convection
from data, facilitated by the more local climate-­invariant representa-
tion of subgrid thermodynamics. Our strategy paves the way for the
successful use of ML models for climate change studies.
MATERIALS AND METHODS
This section outlines how to find feature transformations yielding cli-
mate invariance. Figure 9 illustrates our proposed workflow for find-
ing robust input/output transformations that transform the initial
raw-­data mapping into a climate-­invariant mapping when combined.
Note that this workflow assumes that we cannot or do not want to
retrain ML algorithms in the target climate, which excludes automati-
cally finding a transformation by training a model. This limitation
could arise because the data in the target climate are insufficient or
less reliable or because we seek to uncover new physical relations that
hold across an even wider range of climates.
The first step is to propose a physical transformation to imple-
ment. We can do this through knowledge of robust physical or statis-
tical relations that link and/or preserve distributions (e.g., state
equations, self-­similarities, conservation laws, and accurate empiri-
cal relations) as modeled in section SB2. These relations help derive
invariants [e.g., (94)] under a change in thermodynamic conditions.
Before taking the time to implement this transformation in the ML
workflow, we can verify that the PDFs of the transformed inputs/
outputs (approximately) match in the training and target climates.
Ideally, the joint PDFs of the transformed inputs/outputs would
match. In practice, because it is easier to transform one variable at a
time and the data are often insufficient in the target climate, we can
fall back on the necessary (but not sufficient) condition that the uni-
variate PDFs of the transformed inputs/outputs must match in the
training and target climates. Mathematically, this match can be quan-
tified using PDF distance metrics.
An additional challenge is that the original and transformed vari-
ables may have different units and range, meaning that any nonlinear
distance metric will complicate the PDF comparison. To address this,
we normalize the PDFs and their support variables X so that the
PDFs’ domains strictly lie within [0,1]. For a given variable, we use the
same normalization factors across climates
def X − mincl X
Xnorm = (3)
maxcl X − mincl X
def PDF
PDFnorm = 1
(4)
∫0
dXnorm × PDF(Xnorm )
where PDFnorm is the transformed PDF and Xnorm is its transformed
support; and maxcl and mincl, respectively, refer to the maximum and
minimum operators over the variables’ domains and across climates,
i.e., over the (−4 K), (+0 K), and (+4 K) simulations.
10 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 9. Proposed five-­step workflow to find climate-­invariant transformations.
The transformations help ML models generalize from a reference (ref.) climate to a
target one, using (top) a baseline MLR as an initial guide.
Once the PDFs of each variable are normalized, we may pick any
informative PDF distance metric to quantify how PDFs match
across climates. Here, we pick the commonly used Hellinger dis-
tance between two PDFs p and q, formally defined (95) as
� 2. T. Beucler, I. Ebert-­Uphoff, S. Rasp, M. Pritchard, P. Gentine, “Machine learning for clouds
1 and climate” in Clouds and their Climatic Impacts: Radiation, Circulation, and Precipitation
∫0 2
def dx √ √
(5)
ℋ(p, q) = [ p(x) − q(x)]2
This distance is symmetric (i.e., the arguments’ order does not af-
fect the outcome) and easy to interpret: ℋ(p, q) is bounded by 0
(when p = q) and 100% (when p is zero whenever q is positive and
vice versa). In section SD1, we show that our results using the Hellinger
distance (see table S1) are consistent with those using the Jensen-­
Shannon distance (see table S3) (96), a PDF distance metric giving
large weights to the PDFs’ tails that tend to be particularly problem-
atic for generalization purposes.
Once the univariate PDFs of the physically transformed variables
match across climates, the second step is to train two inexpensive or
“baseline” models on the reference climate to quickly check whether
the transformation improves an ML model’s generalization ability: (i)
a raw-­data model without the transformation and (ii) a climate-­
invariant model with the transformation. If the transformation does
not improve the baseline model’s generalization abilities [i.e., (ii) per-
forms worse than (i) in the target climate], then the transformation
may not be appropriate. Note that we trained MLR baselines to create
climate-­invariant NNs, but the ML model used to define the baseline
should be tailored to the desired final ML model.
If the transformation improves the inexpensive baseline model’s
performances, then the last step is to train the raw-­data and climate-­
invariant versions of the desired ML model (usually nonlinear) on
the reference climate. If the physical transformation improves the
desired ML model’s generalization abilities (i.e., the climate-­invariant
model beats the raw-­data model in the target climate using the same
performance metric calculated over a validation set), then we may
keep the transformation. This workflow may be repeated for the ML
model’s additional input/output variables until the emulated map-
ping is as climate-­invariant as possible.
Before applying this workflow to subgrid thermodynamics clo-
sures, we underline one of its key challenges: Because some transfor-
mations are much more impactful than others, it is often not possible
to develop each physical transformation independently. In our case,
the specific humidity inputs vary the most across climates, meaning
that transforming specific humidity affects the model’s generalization
abilities the most. As a result, initial experiments that independently
tested the effect of transforming temperature suggested a negative
impact of temperature transformation on generalization ability (not
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
shown). This initial result was later invalidated by experiments that
jointly transformed specific humidity and temperature. Following
this, we adopt a progressive input transformation approach, where the
most important inputs are transformed first: specific humidity, then
temperature, and lastly surface energy fluxes.
Supplementary Materials
This PDF file includes:
Supplementary Text
Figs. S1 to S15
Tables S1 to S3
References
REFERENCES AND NOTES
1. M. Reichstein, G. Camps-­Valls, B. Stevens, M. Jung, J. Denzler, N. Carvalhais, Prabhat, Deep
learning and process understanding for data-­driven Earth system science. Nature 566,
195–204 (2019).
(John Wiley & Sons, 2023), pp. 325–345.
3. C. Irrgang, N. Boers, M. Sonnewald, E. A. Barnes, C. Kadow, J. Staneva, J. Saynisch-­Wagner,
Towards neural Earth system modelling by integrating artificial intelligence in Earth
system science. Nat. Mach. Intell. 3, 667–674 (2021).
4. C. O. de Burgh-­Day, T. Leeuwenburg, Machine learning for numerical weather and
climate modelling: A review. Geosci. Model Dev. 16, 6433–6477 (2023).
5. M. J. Molina, T. A. O’Brien, G. Anderson, M. Ashfaq, K. E. Bennett, W. D. Collins, K. Dagon,
J. M. Restrepo, P. A. Ullrich, A review of recent and emerging machine learning applications
for climate variability and weather phenomena. Artif. Intell. Earth Syst. 2, 220086 (2023).
6. P. Ukkonen, R. Pincus, R. J. Hogan, K. P. Nielsen, E. Kaas, Accelerating radiation
computations for dynamical models with targeted machine learning and code
optimization. J. Adv. Model. Earth Syst. 12, e2020MS002226 (2020).
7. A. Belochitski, P. Binev, R. De Vore, M. Fox-­Rabinovitz, V. Krasnopolsky, P. Lamby, Tree
approximation of the long wave radiation parameterization in the NCAR CAM global
climate model. J. Comput. Appl. Math. 236, 447–460 (2011).
8. J. J. Gristey, F. Graham, I. B. Glenn, K. S. Schmidt, H. Chen, On the relationship between
shallow cumulus cloud field properties and surface solar irradiance. Geophys. Res. Lett.
47, e2020GL090152 (2020).
9. R. Lagerquist, D. Turner, I. Ebert-­Uphoff, J. Stewart, V. Hagerty, Using deep learning to emulate
and accelerate a radiative transfer model. J. Atmos. Oceanic Tech. 38, 1673–1696 (2021).
10. M. Chantry, S. Hatfield, P. Dueben, I. Polichtchouk, T. Palmer, Machine learning emulation
of gravity wave drag in numerical weather forecasting. J. Adv. Model. Earth Syst. 13,
e2021MS002477 (2021).
11. Z. I. Espinosa, A. Sheshadri, G. R. Cain, E. P. Gerber, K. J. Dalla Santa, Machine learning
gravity wave parameterization generalizes to capture the QBO and response to
increased CO2. Geophys. Res. Lett. 49, e2022GL098174 (2022).
12. D. Matsuoka, S. Watanabe, K. Sato, S. Kawazoe, W. Yu, S. Easterbrook, Application of deep
learning to estimate atmospheric gravity wave parameters in reanalysis data sets.
Geophys. Res. Lett. 47, e2020GL089436 (2020).
13. J. Yuval, P. A. O’Gorman, Neural-­network parameterization of subgrid momentum
transport in the atmosphere. J. Adv. Model. Earth Syst. 15, e2023MS003606 (2023).
14. H. Morrison, M. van Lier-­Walqui, M. R. Kumjian, O. P. Prat, A Bayesian approach for
statistical-­physical bulk parameterization of rain microphysics. Part I: Scheme
description. J. Atmos. Sci. 77, 1019–1041 (2019).
15. A. Seifert, S. Rasp, Potential and limitations of machine learning for modeling warm-­rain
cloud microphysical processes. J. Adv. Model. Earth Syst. 12, e2020MS002301 (2020).
16. A. Gettelman, D. J. Gagne, C.-­C. Chen, M. W. Christensen, Z. J. Lebo, H. Morrison,
G. Gantos, Machine learning the warm rain process. J. Adv. Model. Earth Syst. 13,
e2020MS002268 (2021).
17. B. François, S. Thao, M. Vrac, Adjusting spatial dependence of climate model outputs with
cycle-­consistent adversarial networks. Climate Dynam. 57, 3323–3353 (2021).
18. B. Pan, G. J. Anderson, A. Goncalves, D. D. Lucas, C. J. W. Bonfils, J. Lee, Y. Tian, H.-­Y. Ma,
Learning to correct climate projection biases. J. Adv. Model. Earth Syst. 13,
e2021MS002509 (2021).
19. V. Zantedeschi, F. Falasca, A. Douglas, R. Strange, M. J. Kusner, D. Watson-­Parris, Cumulo:
A dataset for learning cloud classes. arXiv:1911.04227 [quant-­ph] (2019).
20. S. Rasp, H. Schulz, S. Bony, B. Stevens, Combining crowdsourcing and deep learning to
explore the mesoscale organization of shallow convection. Bull. Am. Meteorol. Soc. 101,
E1980–E1995 (2020).
11 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
21. D. Watson-­Parris, S. A. Sutherland, M. W. Christensen, R. Eastman, P. Stier, A large-­scale
analysis of pockets of open cells and their radiative impact. Geophys. Res. Lett. 48,
e2020GL092213 (2021).
22. L. Denby, Discovering the importance of mesoscale cloud organization through
unsupervised classification. Geophys. Res. Lett. 47, e2019GL085190 (2020).
23. Y. Han, G. J. Zhang, X. Huang, Y. Wang, A moist physics parameterization based on deep
learning. J. Adv. Model. Earth Syst. 12, e2020MS002076 (2020).
24. N. D. Brenowitz, C. S. Bretherton, Prognostic validation of a neural network unified
physics parameterization. Geophys. Res. Lett. 45, 6289–6298 (2018).
25. V. M. Krasnopolsky, M. S. Fox-­Rabinovitz, A. A. Belochitski, Using ensemble of neural
networks to learn stochastic convection parameterizations for climate and numerical
weather prediction models from data simulated by a cloud resolving model. Adv. Artif.
Neural Syst. 2013, 485913 (2013).
26. L. Breiman, Random forests. Mach. Learn. 45, 5–32 (2001).
27. P. A. O’Gorman, J. G. Dwyer, Using machine learning to parameterize moist convection:
Potential for modeling of climate, climate change, and extreme events. J. Adv. Model.
Earth Syst. 10, 2548–2563 (2018).
28. A. Hernanz, J. A. García-­Valero, M. Domínguez, E. Rodríguez-­Camino, A critical view on
the suitability of machine learning techniques to downscale climate change projections:
Illustration for temperature with a toy experiment. Atmos. Sci. Lett. 23, e1087 (2022).
29. S. Rasp, M. S. Pritchard, P. Gentine, Deep learning to represent subgrid processes in
climate models. Proc. Natl. Acad. Sci. 115, 9684–9689 (2018).
30. T. Beucler, M. Pritchard, P. Gentine, S. Rasp, “Towards physically-­consistent, data-­driven
models of convection” in IGARSS 2020 -­2020 IEEE International Geoscience and Remote
Sensing Symposium (IEEE, 2020), pp. 3987–3990.
31. S. K. Clark, N. D. Brenowitz, B. Henn, A. Kwa, J. M. Gibbon, W. A. Perkins, O. Watt-­Meyer,
C. S. Bretherton, L. M. Harris, Correcting a 200 km resolution climate model in multiple
climates by machine learning from 25 km resolution simulations. J. Adv. Model. Earth Syst.
14, e2022MS003219 (2022).
32. A. Doury, S. Somot, S. Gadat, A. Ribes, L. Corre, Regional climate model emulator based
on deep learning: Concept and first evaluation of a novel hybrid downscaling approach.
Climate Dynam. 60, 1751–1779 (2023).
33. A. P. Guillaumin, L. Zanna, Stochastic-­deep learning parameterization of ocean
momentum forcing. J. Adv. Model. Earth Syst. 13, e2021MS002534 (2021).
34. M. J. Molina, D. J. Gagne, A. F. Prein, A benchmark to test generalization capabilities of
deep learning methods to classify severe convective storms in a changing climate. Earth
Space Sci. 8, e2020EA001490 (2021).
35. Y. LeCun, Y. Bengio, “Convolutional networks for images, speech, and time series” The
Handbook of Brain Theory and Neural Networks (MIT Press, 1995), p. 276.
36. S. Ioffe, C. Szegedy, “Batch normalization: Accelerating deep network training by
reducing internal covariate shift” in International Conference on Machine Learning (PMLR,
2015), pp. 448–456.
37. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: A simple
way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958
(2014).
38. X. Ying, An overview of overfitting and its solutions. J. Phys. Conf. Ser. 1168, 022022 (2019).
39. G. Wilson, D. J. Cook, A survey of unsupervised deep domain adaptation. ACM Trans.
Intell. Syst. Technol. 11, 1–46 (2020).
40. V. M. Patel, R. Gopalan, R. Li, R. Chellappa, Visual domain adaptation: A survey of recent
advances. IEEE Signal Process. Mag. 32, 53–69 (2015).
41. D. Tuia, C. Persello, L. Bruzzone, Domain adaptation for the classification of remote
sensing data: An overview of recent advances. IEEE Trans. Geosci. Remote Sens. 4, 41–57
(2016).
42. S. J. Pan, Q. Yang, A survey on transfer learning. IEEE Trans. Knowl. Data Eng. 22,
1345–1359 (2010).
43. W. M. Kouw, M. Loog, A review of domain adaptation without target labels. IEEE Trans.
Pattern Anal. Mach. Intell. 43, 766–785 (2021).
44. J. Ling, A. Kurzawski, J. Templeton, Reynolds averaged turbulence modelling using deep
neural networks with embedded invariance. J. Fluid Mech. 807, 155–166 (2016).
45. D. Randall, M. Khairoutdinov, A. Arakawa, W. Grabowski, Breaking the cloud
parameterization deadlock. Bull. Am. Meteorol. Soc. 84, 1547–1564 (2003).
46. S. Bony, B. Stevens, D. M. W. Frierson, C. Jakob, M. Kageyama, R. Pincus, T. G. Shepherd,
S. C. Sherwood, A. P. Siebesma, A. H. Sobel, M. Watanabe, M. J. Webb, Clouds, circulation
and climate sensitivity. Nat. Geosci. 8, 261–268 (2015).
47. T. Schneider, J. Teixeira, C. S. Bretherton, F. Brient, K. G. Pressel, C. Schär, A. P. Siebesma,
Climate goals and computing the future of clouds. Nat. Clim. Change 7, 3–5 (2017).
48. P. Gentine, V. Eyring, T. Beucler, “Deep learning for the parametrization of subgrid
processes in climate models” in Deep Learning for the Earth Sciences: A Comprehensive
Approach to Remote Sensing, Climate Science, and Geosciences (John Wiley & Sons, 2021),
pp. 307–314.
49. N. D. Brenowitz, T. Beucler, M. Pritchard, C. S. Bretherton, Interpreting and stabilizing
machine-­learning parametrizations of convection. J. Atmos. Sci. 77, 4357–4375 (2020).
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
50. J. Ott, M. Pritchard, N. Best, E. Linstead, M. Curcic, P. Baldi, A Fortran-­Keras deep learning
bridge for scientific computing. Sci. Program. 2020, 8888811 (2020).
51. P. Forster, T. Storelvmo, K. Armour, W. Collins, J.-­L. Dufresne, D. Frame, D. J. Lunt,
T. Mauritsen, M. D. Palmer, M. Watanabe, M. Wild, H. Zhang, “The Earth’s energy budget,
climate feedbacks, and climate sensitivity” in Climate Change 2021: The Physical Science
Basis. Contribution of Working Group I to the Sixth Assessment Report of the
Intergovernmental Panel on Climate Change, V. Masson-­Delmotte, P. Zhai, A. Pirani,
S. L. Connors, C. Péan, S. Berger, N. Caud, Y. Chen, L. Goldfarb, M. I. Gomis, M. Huang,
K. Leitzell, E. Lonnoy, J. B. R. Matthews, T. K. Maycock, T. Waterfield, O. Yelekçi, R. Yu,
B. Zhou, Eds. (Cambridge Univ. Press, 2021), pp. 923–1054.
52. G. Mooers, M. Pritchard, T. Beucler, J. Ott, G. Yacalis, P. Baldi, P. Gentine, Assessing the
potential of deep learning for emulating cloud superparameterization in climate models
with real-­geography boundary conditions. J. Adv. Model. Earth Syst. 13, e2020MS002385
(2021).
53. P. Gentine, M. Pritchard, S. Rasp, G. Reinaudi, G. Yacalis, Could machine learning break the
convection parameterization deadlock? Geophys. Res. Lett. 45, 5742–5751 (2018).
54. J. Yuval, P. A. O’Gorman, Stable machine-­learning parameterization of subgrid processes
for climate modeling at a range of resolutions. Nat. Commun. 11, 3295 (2020).
55. J. Yuval, P. A. O’Gorman, C. N. Hill, Use of neural networks for stable, accurate and
physically consistent parameterization of subgrid atmospheric processes with good
performance at reduced precision. Geophys. Res. Lett. 48, e2020GL091363 (2021).
56. M. Khairoutdinov, D. Randall, C. De Mott, Simulations of the atmospheric general
circulation using a cloud-­resolving model as a superparameterization of physical
processes. J. Atmos. Sci. 62, 2136–2154 (2005).
57. M. S. Pritchard, C. S. Bretherton, Causal evidence that rotational moisture advection is
critical to the superparameterized Madden-­Julian oscillation. J. Atmos. Sci. 71, 800–815
(2014).
58. J. A. Andersen, Z. Kuang, Moist static energy budget of MJO-­like disturbances in the
atmosphere of a zonally symmetric aquaplanet. J. Climate 25, 2782–2804 (2012).
59. M. F. Khairoutdinov, D. A. Randall, A cloud resolving model as a cloud parameterization in
the NCAR Community Climate System Model: Preliminary results. Geophys. Res. Lett. 28,
3617–3620 (2001).
60. T. Beucler, M. Pritchard, S. Rasp, J. Ott, P. Baldi, P. Gentine, Enforcing analytic constraints in
neural networks emulating physical systems. Phys. Rev. Lett. 126, 098302 (2021).
61. M. Wang, V. E. Larson, S. Ghan, M. Ovchinnikov, D. P. Schanen, H. Xiao, X. Liu, P. Rasch,
Z. Guo, A multiscale modeling framework model (superparameterized CAM5) with a
higher-­order turbulence closure: Model description and low-­cloud simulations. J. Adv.
Model. Earth Syst. 7, 484–509 (2015).
62. J. W. Hurrell, J. J. Hack, D. Shea, J. M. Caron, J. Rosinski, A new sea surface temperature
and sea ice boundary dataset for the Community Atmosphere Model. J. Climate 21,
5145–5153 (2008).
63. S. N. Tulich, A strategy for representing the effects of convective momentum transport in
multiscale models: Evaluation using a new superparameterized version of the Weather
Research and Forecast model (SP-­WRF). J. Adv. Model. Earth Syst. 7, 938–962 (2015).
64. N. D. Brenowitz, C. S. Bretherton, Spatially extended tests of a neural network
parametrization trained by coarse-­graining. J. Adv. Model. Earth Syst. 11, 2728–2744 (2019).
65. T. Bolton, L. Zanna, Applications of deep learning to ocean data inference and subgrid
parameterization. J. Adv. Model. Earth Syst. 11, 376–399 (2019).
66. S. T. Garner, D. M. W. Frierson, I. M. Held, O. Pauluis, G. K. Vallis, Resolving convection in a
global hypohydrostatic model. J. Atmos. Sci. 64, 2061–2075 (2007).
67. W. R. Boos, A. Fedorov, L. Muir, Convective self-­aggregation and tropical cyclogenesis
under the hypohydrostatic rescaling. J. Atmos. Sci. 73, 525–544 (2016).
68. P. A. O’Gorman, Z. Li, W. R. Boos, J. Yuval, Response of extreme precipitation to uniform
surface warming in quasi-­global aquaplanet simulations at high resolution. Philos. Trans.
A. Math. Phys. Eng. Sci. 379, 20190543 (2021).
69. R. B. Neale, B. J. Hoskins, A standard test for AGCMs including their physical
parametrizations: I: The proposal. Atmos. Sci. Lett. 1, 101–107 (2000).
70. A. Géron, Hands-­On Machine Learning With Scikit-­Learn, Keras, and TensorFlow (O’Reilly
Media Inc., 2022).
71. A. Karpatne, I. Ebert-­Uphoff, S. Ravela, H. A. Babaie, V. Kumar, Machine learning for the
geosciences: Challenges and opportunities. IEEE Trans. Knowl. Data Eng. 31, 1544–1554
(2019).
72. J. R. Holton, D. O. Staley, An introduction to dynamic meteorology. Am. J. Phys. 41,
752–754 (1973).
73. A. Siebesma Pier, Sandrine Bony, Christian Jakob, Bjorn Stevens, Eds., Clouds and Climate:
Climate Science’s Greatest Challenge (Cambridge Univ. Press, 2020).
74. S. Manabe, R. T. Wetherald, Thermal equilibrium of the atmosphere with a given
distribution of relative humidity. J. Atmos. Sci. 24, 241–259 (1967).
75. R. G. Brown, C. Zhang, Variability of midtropospheric moisture and its effect on cloud-­top
height distribution during TOGA COARE. J. Atmos. Sci. 54, 2760–2774 (1997).
76. C. S. Bretherton, M. E. Peters, L. E. Back, Relationships between water vapor path and
precipitation over the tropical oceans. J. Climate 17, 1517–1528 (2004).
12 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
77. C. E. Holloway, J. David Neelin, Temporal relations of column water vapor and tropical
precipitation. J. Atmos. Sci. 67, 1091–1105 (2010).
78. J. T. Seeley, N. Jeevanjee, D. M. Romps, FAT or FiTT: Are anvil clouds or the tropopause
temperature invariant? Geophys. Res. Lett. 46, 1842–1850 (2019).
79. D. L. Hartmann, K. Larson, An important constraint on tropical cloud‐climate feedback.
Geophys. Res. Lett. 29, 10.1029/2002GL015835 (2002).
80. Z. Kuang, D. L. Hartmann, Testing the fixed anvil temperature hypothesis in a
cloud-­resolving model. J. Climate 20, 2051–2057 (2007).
81. F. Ahmed, J. David Neelin, Reverse engineering the tropical precipitation-­buoyancy
relationship. J. Atmos. Sci. 75, 1587–1608 (2018).
82. F. Ahmed, Á. F. Adames, J. D. Neelin, Deep convective adjustment of temperature and
moisture. J. Atmos. Sci. 77, 2163–2186 (2020).
83. K. A. Schiro, F. Ahmed, S. E. Giangrande, J. D. Neelin, GoAmazon2014/5 campaign points
to deep-­inflow approach to deep convection across scales. Proc. Natl. Acad. Sci. U.S.A.
115, 4577–4582 (2018).
84. F. Ahmed, J. David Neelin, Protected convection as a metric of dry air influence on
precipitation. J. Climate 34, 3821–3838 (2021).
85. D. L. Hartmann, Global Physical Climatology, vol. 103 (Elsevier, 2015).
86. S. M. Lundberg, S.-­I. Lee, A unified approach to interpreting model predictions. Adv.
Neural Inf. Process. 30, 4768–4777 (2017).
87. L. S. Shapley, “A value for n-­person games” in Contributions to the Theory of Games
(Princeton Univ. Press, 1953), pp. 307–317.
88. M. J. Herman, Z. Kuang, Linear response functions of two convective parameterization
schemes. J. Adv. Model. Earth Syst. 5, 510–541 (2013).
89. T. Beucler, T. Cronin, K. Emanuel, A linear response framework for radiative-­convective
instability. J. Adv. Model. Earth Syst. 10, 1924–1951 (2018).
90. Z. Kuang, Linear stability of moist convecting atmospheres. Part I: From linear response
functions to a simple model and applications to convectively coupled waves. J. Atmos.
Sci. 75, 2889–2907 (2018).
91. P.-­J. Kindermans, S. Hooker, J. Adebayo, M. Alber, K. T. Schütt, S. Dähne, D. Erhan, B. Kim,
“The (un) reliability of saliency methods” in Explainable AI: Interpreting, Explaining and
Visualizing Deep Learning (Springer, 2019), pp. 267–280.
92. C. E. Holloway, J. D. Neelin, The convective cold top and quasi equilibrium. J. Atmos. Sci.
64, 1467–1487 (2007).
93. A. Zheng, A. Casari, Feature Engineering for Machine Learning: Principles and Techniques for
Data Scientists (O’Reilly Media Inc., 2018).
94. T. G. Shepherd, Symmetries, conservation laws, and Hamiltonian structure in geophysical
fluid dynamics. Adv. Geophys. 32, 287–338 (1990).
95. R. Beran, Minimum Hellinger distance estimates for parametric models. Ann. Stat. 5,
445–463 (1977).
96. D. M. Endres, J. E. Schindelin, A new metric for probability distributions. IEEE Trans. Inf.
Theory 49, 1858–1860 (2003).
97. S. Rasp, T. Beucler, G. Reinaudi, P. Gentine, tbeucler/CBRAIN-­CAM: Climate-­invariant
branch, second release, Zenodo (2023); https://doi.org/10.5281/zenodo.8140413.
98. T. Beucler, M. Pritchard, L. Peng, J. Yuval, Climate-­invariant machine learning, Zenodo
(2023); https://doi.org/10.5281/zenodo.8140536.
99. S. Rasp, raspstephan/CBRAIN-­CAM, Zenodo (2018); https://doi.org/10.5281/
zenodo.1402384.
100. H. Parishani, M. S. Pritchard, C. S. Bretherton, C. R. Terai, M. C. Wyant, M. Khairoutdinov,
B. Singh, Insensitivity of the cloud response to surface warming under radical changes to
boundary layer turbulence and cloud microphysics: Results from the ultraparameterized
CAM. J. Adv. Model. Earth Syst. 10, 3139–3158 (2018).
101. J. Yuval, yaniyuval/Neural_nework_parameterization: Associated code and data for use
of neural networks for stable, accurate and physically consistent parameterization of
subgrid atmospheric processes with good performance at reduced precision, Zenodo
(2021); https://doi.org/10.5281/zenodo.4118346.
102. M. F. Khairoutdinov, D. A. Randall, Cloud resolving modeling of the ARM summer 1997
IOP: Model formulation, results, uncertainties, and sensitivities. J. Atmos. Sci. 60, 607–625
(2003).
Beucler et al., Sci. Adv. 10, eadj7250 (2024) 7 February 2024
103. K. A. Emanuel, Atmospheric Convection (Oxford Univ. Press, 1994).
104. Á. F. Adames, S. W. Powell, F. Ahmed, V. C. Mayta, J. D. Neelin, Tropical precipitation
evolution in a buoyancy-­budget framework. J. Atmos. Sci. 78, 509–528 (2021).
105. D. Maraun, Bias correcting climate change simulations – A critical review. Curr. Clim.
Change Rep. 2, 211–220 (2016).
106. N. Jeevanjee, D. M. Romps, Mean precipitation change from a deepening troposphere.
Proc. Natl. Acad. Sci. 115, 11465–11470 (2018).
107. M. S. Singh, P. A. O’Gorman, Upward shift of the atmospheric general circulation under
global warming: Theory and simulations. J. Climate 25, 8259–8276 (2012).
108. D. Irving, A minimum standard for publishing computational results in the weather and
climate sciences. Bull. Am. Meteorol. Soc. 97, 1149–1158 (2016).
109. M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S. Corrado, A. Davis,
J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia,
R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mane, R. Monga, S. Moore, D. Murray,
C. Olah, M. Schuster, J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke,
V. Vasudevan, F. Viegas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, X. Zheng,
TensorFlow: Large-­scale machine learning on heterogeneous distributed systems.
arXiv:1603.04467 (2016).
110. J. Kiefer, J. Wolfowitz, Stochastic estimation of the maximum of a regression function.
Ann. Math. Stat. 23, 462–466 (1952).
111. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin,
N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. De Vito, M. Raison, A. Tejani,
S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala, Pytorch: An imperative style,
high-­performance deep learning library. arXiv:1912.01703 (2019).
112. L. N. Smith, “Cyclical learning rates for training neural networks” in 2017 IEEE Winter
Conference on Applications of Computer Vision (WACV) (IEEE, 2017), pp. 464–472.
113. F. Chollet, Keras: The python deep learning library, Astrophysics Source Code Library,
ascl-­1806 (2018); https://keras.io/.
Acknowledgments
Funding: T.B. and M.P. acknowledge funding from NSF grants OAC1835863 and AGS-­1734164.
P.G., M.P., S.Y., and J.L. acknowledge NSF funding from the LEAP Science and Technology center
grant (2019625-­STC). M.P. acknowledges the DOE’s Exascale Computing Project (17-­SC-­20-­SC),
Biological and Environmental Research (BER) program (DE-­SC0022331), and Advanced
Scientific Computing Research (ASCR) program (DE-­SC0022255). M.P. and S.Y. also
acknowledge the DOE EAGLES project (74358). J.Y. acknowledges support from the EAPS
Houghton-­Lorenz postdoctoral fellowship. P.G. thanks NSF grant AGS-­1734164 and USMILE
European Research Council grant. P.G., P.A.O., and J.Y. received M2LInES research funding by
the generosity of Eric and Wendy Schmidt by recommendation of the Schmidt Futures
program. F.A. and J.D.N. acknowledge NSF AGS-­1936810; F.A. acknowledges NSF AGS-­2225956.
We also thank the Extreme Science and Engineering Discovery Environment supported by NSF
grant number ACI-­1548562 (charge numbers TG-­ATM190002 and TG-­ATM170029), the Texas
Advanced Computing Center (TACC; ATM-­20009), and HPC support from Cheyenne
(doi:10.5065/D6RX99HX) provided by NCAR’s Computational and Information Systems
Laboratory (sponsored by the NSF) for computational resources. Contributions:
Conceptualization: T.B., M.P., S.R., and P.G. Methodology: All authors. Investigation: T.B., M.P., J.Y.,
A.G., L.P., J.L., and S.Y. Visualization: T.B., M.P., P.G., J.D.N., and J.Y. Supervision: T.B., M.P., and P.G.
Writing—original draft: T.B. Writing—review and editing: All authors. We thank W. Boos for
providing the SAM hypohydrostatic simulation output; J. Meyerson for helping us improve
figures; and A. Jenney, I. Ebert-­Uphoff, and the CBRAIN discussion group for help and advice
that greatly benefited the present manuscript. Competing interests: The authors declare that
they have no competing interests. Data and materials availability: All data needed to
evaluate the conclusions in the paper are present in the paper and/or the Supplementary
Materials. See section SA for details about code and data availability.
Submitted 13 July 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adj7250
13 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

OPTICS Copyright © 2024 The

Authors, some rights
Probing electron-­hole Coulomb correlations reserved; exclusive
licensee American
in the exciton landscape of a twisted Association for the
Advancement of
semiconductor heterostructure Science. No claim to
original U.S.
Government Works.
Jan Philipp Bange1, David Schmitt1, Wiebke Bennecke1, Giuseppe Meneghini2, Distributed under a
AbdulAziz AlMutairi3, Kenji Watanabe4, Takashi Taniguchi5, Daniel Steil1, Sabine Steil1, Creative Commons
R. Thomas Weitz1,6, G. S. Matthijs Jansen1, Stephan Hofmann3, Samuel Brem2, Ermin Malic2,7, Attribution License 4.0
Marcel Reutzel1*, Stefan Mathias1,6* (CC BY).

In two-­dimensional semiconductors, cooperative and correlated interactions determine the material’s excitonic prop-
erties and can even lead to the creation of correlated states of matter. Here, we study the fundamental two-­particle
correlated exciton state formed by the Coulomb interaction between single-­particle holes and electrons. We find that
the ultrafast transfer of an exciton’s hole across a type II band-­aligned semiconductor heterostructure leads to an un-
expected sub-­200-­femtosecond upshift of the single-­particle energy of the electron being photoemitted from the
two-­particle exciton state. While energy relaxation usually leads to an energetic downshift of the spectroscopic signa-
ture, we show that this upshift is a clear fingerprint of the correlated interaction of the electron and hole parts of the
exciton. In this way, time-­resolved photoelectron spectroscopy is straightforwardly established as a powerful method
to access electron-­hole correlations and cooperative behavior in quantum materials. Our work highlights this capabil-
ity and motivates the future study of optically inaccessible correlated excitonic and electronic states of matter.

INTRODUCTION
An exciton is a prime example of a quasiparticle that is built up by
electrons and holes bound together via Coulomb interaction. As in
the case of a hydrogen atom, the exciton’s properties are described
by its quantum number, its binding energy, and its Bohr radius (1).
For low-­dimensional materials, these key parameters can be sub-
stantially altered by cooperative interactions with surrounding qua-
siparticles (2, 3). To study such cooperative and emergent behavior,
artificial stacks of two-­dimensional transition metal dichalcogen-
ides (TMDs) have been shown to provide an exceptional playground
for manipulating exciton properties. Examples include the ultrafast
formation of interlayer excitons whose electron and hole compo-
nents are charge-­separated across the neighboring TMD layers (4–8),
the confinement of excitons in a moiré potential well (9–12), the
creation of correlated interlayer exciton insulators (13, 14) and ex-
citon crystals (15, 16), and even the stabilization of Bose-­Einstein
condensates (17).
It is therefore of fundamental importance to obtain insight
into the energy landscape and the ultrafast dynamics of the two-­
particle correlated exciton state (18, 19). In TMD semiconductors,
momentum-­indirect and spin-­forbidden excitons play a substantial
role but are mostly inaccessible (7, 20) using all-­optical experimental
techniques (21, 22). Recently, time-­and angle-­resolved photoelectron
spectroscopy (trARPES) experiments have been shown to be a pow-
erful technique to fill this gap and to simultaneously probe the
1
I. Physikalisches Institut, Georg-­August-­Universität Göttingen, Friedrich-­Hund-­Platz 1,
37077 Göttingen, Germany. 2Fachbereich Physik, Philipps-­Universität Marburg,
35032 Marburg, Germany. 3Department of Engineering, University of Cambridge,
Cambridge CB3 0FA, UK. 4Research Center for Functional Materials, National Insti-
tute for Materials Science, 1-­1 Namiki, Tsukuba 305-­0044, Japan. 5International
Center for Materials Nanoarchitectonics, National Institute for Materials Science,
1-­1 Namiki, Tsukuba 305-­0044, Japan. 6International Center for Advanced Studies
of Energy Conversion (ICASEC), University of Göttingen, Göttingen, Germany. 7De-
partment of Physics, Chalmers University of Technology, Gothenburg, Sweden.
*Corresponding author. Email: marcel.​reutzel@​phys.​uni-­goettingen.​de (M.R.); smathias@​
uni-­goettingen.​de (S.M.)
energy landscape and dynamics of optically bright and dark excitons
in monolayer (23–25) and twisted bilayer (8, 12, 26, 27) TMDs.
When using photoelectron spectroscopy, there is a fundamental
aspect that needs to be considered (Fig. 1A): In the photoemission
process, the Coulomb correlation between the electron and hole
components of the exciton is broken. This is because a single-­particle
photoelectron is collected with the detector and a single-­particle
hole remains in the material (28–31). In consequence, photoelec-
trons originating from excitons are detected at the exciton bind-
ing energy below the conduction band minimum (8, 23–25, 32) and
show a hole-like energy-­momentum dispersion (32, 33). In this
way, trARPES provides natural access to the electron contribution
of the exciton and can be used to quantify the charge transfer of the
exciton’s electron across a type II band-­aligned heterostructure
(Fig. 1B) (8, 27). However, to this day, only very limited energy-­
and momentum-­resolved spectroscopic information on the exciton’s
hole component is reported (12). Specifically, in contrast to all-­
optical spectroscopies (4–6, 18, 34–37), it has not been shown that
trARPES can be applied to monitor the charge-­transfer dynamics of
the exciton’s hole across the TMD interface (Fig. 1C).
Here, we demonstrate how the Coulomb interaction between the
electron-­and the hole components of the intra-­and interlayer exci-
tons facilitates the study of the ultrafast hole-­transfer mechanism in
a twisted WSe2/MoS2 heterostructure. We experimentally observe
an increase in the exciton’s photoelectron energy upon the hole-­
transfer process across the interface. This is unexpected at first be-
cause the electron remains rigid in the conduction band minimum
during this hole-­transfer process (Fig. 1C) and because any relax-
ation mechanism is typically expected to cause an overall decrease
in the measured electronic quasiparticle energies. However, when
taking the correlated nature of the electron-­hole pair into account,
despite an overall decrease in the quasiparticle energies, we show
that such an increase due to hole transfer must be expected for the
corresponding exciton’s photoelectron. Our work provides micro-
scopic insights into the ultrafast hole-­transfer mechanism and, more

Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024 1 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
generally, highlights the potential of time-­resolved momentum mi-
croscopy to probe optically inaccessible correlated excitonic and
electronic states of matter.
RESULTS
Energy landscape and photoemission fingerprints of bright
and dark excitons
We start the analysis of the hole-­transfer dynamics by first calculat-
ing the full energy landscape and formation dynamics of bright and
Fig. 1. Probing Coulomb-­correlated electron-­hole pairs and their femtosec-
ond dynamics using momentum microscopy. (A) Schematic illustration of the
photoemission process from excitons. Visible femtosecond light pulses (red) are
used to optically excite bright excitons that fully reside in the MoS2 monolayer. The
transfer of the hole component into the WSe2 monolayer leads to the formation of
charge-­separated interlayer excitons (black arrow). A time-­delayed extreme ultra-
violet laser pulse (blue) breaks the exciton; single-­particle electrons are detected in
the photoelectron analyzer and single-­particle holes remain in the WSe2 mono-
layer. (B and C) Single-­particle energy-­level alignment of the valence and conduc-
tion bands (v and c) of MoS 2 and WSe 2. K W-­K Mo excitons are formed due to
interlayer charge transfer of the exciton’s hole or electron, respectively, from intra-
layer KMo-­KMo or KW-­KW excitons. Note that in (C), the electron contribution to the
exciton remains rigid in the conduction band minimum of MoS2 during the hole-­
transfer process. In the abbreviation of the excitons, the capital letters and the sub-
scripts denote the valley (K, Σ, and Γ) and the layer (W and Mo) where the exciton’s
hole (first letter) and electron (second letter) are localized. It is not differentiated
between momentum-­direct and momentum-­indirect excitons (e.g., K W/Mo and
K'W/Mo or Σ and Σ') because those cannot be differentiated in the photoemission
experiment (see fig. S7).
Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024
dark excitons in the twisted WSe2/MoS2 heterostructure on a micro-
scopic footing (details in Supplementary Text). The optically excited
A1s excitons in the WSe2 and the MoS2 layer and their cascaded re-
laxation via layer-­hybridized excitons to the lowest energy interlayer
excitons are illustrated in Fig. 2A. If the heterostructure is excited
resonantly to the A1s-­exciton of WSe2 with 1.7 eV pulses, then only
intralayer KW-­KW A1s excitons are optically excited and decay in a
cascaded transition via layer hybridized KW-­Σ excitons to interlayer
KW-­KMo excitons, as we have discussed in detail in our earlier work
(8, 27) (i.e., KW-­KW → KW-­Σ → KW-­KMo; Fig. 2A, left-­hand side). In
the single-­particle picture, this cascaded transition can be associated
with the transfer of the exciton’s electron across the TMD interface
(Fig. 1B).
Complementary, if the hole-­transfer process across the WSe2/
MoS2 interface is considered (Fig. 1C), then the dynamics must be
initiated by an excitation of MoS2 A1s excitons with 1.9 eV light
pulses (KMo-­KMo excitons in Fig. 2A, right-­hand side). Exploiting
the density matrix formalism, we calculate the excitonic energy
landscape (details below), and track the exciton dynamics, finding
that the most efficient mechanism to form interlayer KW-­KMo exci-
tons occurs via layer hybridized Γ-­KMo excitons, where the exciton’s
electron resides in the KMo valley of MoS2 and the exciton’s hole can
be found in the layer-­hybridized valence bands at the Γ valley (38).
Hence, the hole-­transfer dominantly occurs via the KMo-­KMo → Γ-­
KMo → KW-­KMo exciton cascade.
To differentiate the spectral contributions of different excitons in
the experiment, we apply our setup for femtosecond momentum
microscopy (39, 40) that provides direct access to the photoemission
energy-­momentum fingerprint of excitons (Fig. 2, B to E). In Fig. 2E,
the momentum map of the intralayer KMo-­KMo exciton is shown af-
ter resonant optical excitation with 1.9-­eV pump pulses. Photoelec-
trons are detected at the in-­plane momenta of the KMo and K′Mo
valleys (0 ps). For better visibility, the Brillouin zone of MoS2 is over-
laid in dark red. Because 1.9-­eV pump photons also non-­resonantly
excite KW-­KW excitons in WSe2, the momentum map in Fig. 2C
shows photoemission yield at the KW and K′W valleys of WSe2 (or-
ange hexagon, 0 ps). Note that the Brillouin zone of WSe2 is rotated
by 9.8∘ ± 0.8∘ with respect to MoS2. Moreover, weak photoemission
yield from hybrid KW-­Σ excitons is detected at the Σ and Σ' valleys
(grey hexagon). At a pump-­probe delay of 10 ps (Fig. 2D), the major
part of the intralayer excitons has decayed either via the electron-­or
the hole-­transfer process, and spectral yield is dominated by the en-
ergetically most stable excitation, i.e., the interlayer KW-­KMo exci-
tons (fig. S4). For these interlayer excitons, the electron and the hole
contributions are now separated between both monolayers of the
heterostructure, and the exciton photoemission momentum finger-
print has to be described within the moiré mini-­Brillouin zones built
up by the κ valleys whose in-­plane momentum can be constructed
by the reciprocal lattice vectors of WSe2 and MoS2 (Fig. 2D, black
hexagon) (8, 26).
Hole-­and electron-­transfer dynamics
Having identified the exciton fingerprints in the photoemission ex-
periment, we can now proceed with the analysis of the hole-­transfer
dynamics. For this, fig. S4 provides an overview of the pump-­probe
delay-­dependent evolution of photoemission intensity from intra-
layer KMo-­KMo and KW-­KW excitons, the hybrid KW-­Σ exciton, and
the interlayer KW-­KMo exciton after optical excitation with 1.9 eV
(fluence: 140 μJ/cm2; optically excited exciton densities of 7 ×1011
2 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Energy landscape and energy-­momentum fingerprints of excitons in WSe2/MoS2. (A) Calculated low-­energy exciton landscape of intralayer, hybrid, and inter-
layer excitons. The electron-­and hole-­transfer processes can be initiated via excitation with 1.7-­and 1.9-­eV light pulses, respectively, and proceed via the KW-­KW → KW-­Σ→
KW-­KMo and KMo-­KMo → Γ-­KMo → KW-­KMo cascades. The solid and dashed arrows, respectively, indicate exciton-­phonon scattering events leading to inter-­and intravalley
thermalization of the exciton occupation. The effective mass of the exciton dispersion is extracted from many-­body calculations. The inset schematically shows the align-
ment of the WSe2 and MoS2 Brillouin zones and indicates the high-­symmetry points in the first Brillouin zone. (B) Energy-­and momentum-­resolved photoemission spec-
trum along the Γ-­Σ-­KW direction (inset) measured on the WSe2/MoS2 heterostructure after photoexcitation with 1.9-­eV light pulses at a delay of 10 ps. The WSe2 and MoS2
valence band maxima are labeled with EvW and EvMo, respectively. (C to E) Photoemission momentum fingerprints of the (C) intralayer KW-­KW exciton (0 ps), the (D) inter-
layer KW-­KMo exciton (10 ps), and the (E) intralayer KMo-­KMo exciton (0 ps) after photoexcitation with 1.9-­eV light pulses. The photoelectron energies of the momentum
maps are given in the figure with respect to the energy of the WSe2 valence band maximum and are indicated by colored arrowheads in (B). The energetic width of the
arrowheads indicates the energy range used for generating the momentum maps (C, D, and E). The Brillouin zones of WSe2, MoS2, and the moiré superlattice are overlaid
on the data by orange, dark red (dashed), and black hexagons, respectively.

and 3.5 ×1012 cm−2 in WSe2 and MoS2 (41), respectively). The for-
mation and thermalization dynamics of all accessible excitons indi-
cate that electron-­and hole-­transfer processes contribute to the
formation of interlayer KW-­KMo excitons, which, in consequence, we
have to distinguish. To do so, we directly compare the interlayer KW-­
KMo exciton rise time for 1.7-­and 1.9-­eV pumping. In Fig. 3A, the
black data points show the pump-­probe delay-­dependent buildup of
interlayer KW-­KMo exciton photoemission intensity that is formed
by electron-­and hole-­transfer processes (1.9-­eV pump photons).
For comparison, the green data points show the pump-­probe delay-­
dependent buildup of the interlayer KW-­KMo exciton intensity that is
formed only via the electron transfer process (1.7-­eV pump pho-
tons, fluence: 280 μJ/cm 2, exciton density: 5.4 ×10 12 cm −2). It
is directly obvious that there is a strong hierarchy of timescales
for the electron-­and hole-­transfer processes: When considering the
electron-­only transfer process (green symbols), the interlayer exci-
ton signal increases rapidly with pump-­probe delay and saturates
on the sub-­200-­fs timescale. A quantitative evaluation with rate
equation modeling yields a formation time of te − transfer = 40 ± 10 fs
(see Supplementary Text). In contrast, the joint buildup of interlayer
KW-­KMo excitons via electron-­and hole-­transfer processes after 1.9-­eV
excitation saturates on the 1-­ps timescale (black symbols). For fur-
ther analyzing this dataset, we assume that the 1.9-­eV pump pulses
excite A1s excitons in WSe2 and MoS2 in a 1:5 ratio, as estimated
from the optical absorption coefficient of both monolayers (41), and
take the already deduced electron-­transfer time te − transfer = 40 ± 10 fs
into account. From this fit, we extract th − transfer = 2.2 ± 1 ps, which
is more than an order of magnitude larger than the electron-­transfer
time te − transfer (see rate equation analysis based on fig. S3).
Hence, our experimental data imply that the interlayer hole-­
transfer mechanism across the WSe2/MoS2 heterointerface is sub-
stantially slower compared to the electron-­transfer mechanism. To
understand our findings on a microscopic footing, we exploit the
density matrix formalism to derive excitonic equations of motion
within the energy landscape of excitons shown in Fig. 2A and fig. S7
(see details in Supplementary Text) (38, 42). Here, we incorporate
exciton-­light and exciton-­phonon interaction and assume again that
the 1.9-­eV pump pulses excite A1s excitons in WSe2 and MoS2 in
a 1:5 ratio (41). We find an excellent qualitative agreement of the
microscopic model calculations (Fig. 3B) with the experimentally

Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024 3 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
quantified rise time (Fig. 3A) of interlayer KW-­KMo excitons: The
electron-­only transfer process saturates for delays <200 fs (green),
while the combined electron-­and hole-­transfer dynamics lead to an
increasing interlayer KW-­KMo exciton occupation for substantially
longer delays (black). Hence, in experiment and theory, we find that
the electron-­transfer dynamics is roughly one order of magnitude
faster than the hole-­transfer dynamics.
To understand this drastic difference in the rise time of interlayer
KW-­KMo exciton formation via the electron-­versus the hole-­transfer
process, we evaluate the calculated exciton dynamics in more detail
and make two major observations: First, it is important to realize
that the exciton energy difference between the optically excited in-
tralayer exciton and the interlayer exciton is roughly 200 meV larger
in the case of 1.9-­eV excitation, which initiates the hole-­transfer
process (see exciton energies in Fig. 2A and fig. S7). The dissipation
Fig. 3. Femtosecond-­to-­picosecond evolution of the hole-­and electron-­transfer
dynamics. (A) Direct comparison of the interlayer KW-­KMo exciton formation dy-
namics if the heterostructure is excited resonantly to the intralayer KW-­KW exciton
energy of WSe2 (1.7 eV, green circles) or the intralayer KMo-­KMo exciton of MoS2 (1.9 eV,
black circles). While the electron-­only transfer process (1.7 eV) leads to a saturation
of photoemission yield from interlayer KW-­KMo excitons on the <200-­fs timescale,
the combined electron-­and hole-­transfer dynamics (1.9 eV) leads to an increasing
photoemission yield up to 1 ps. The momentum-­filtered regions of interest (black
circles) used in the 1.7-­eV (green contour) and 1.9-­eV (black contour) measure-
ments are shown in the bottom panel. The κ valley that overlaps with the original
KMo valley is excluded in the analysis of the 1.9-­eV measurement. (B) Microscopic
model calculations of the interlayer KW-­KMo exciton formation dynamics. The green
curve describes the temporal evolution of the occupation of interlayer KW-­KMo ex-
citons after photoexcitation of intralayer KW-­KW excitons. For the black curve, the
interlayer KW-­KMo exciton formation dynamics is induced by the initial excitation of
intralayer KW-­KW and KMo-­KMo excitons. Note that the model calculations do not
include additional decay processes.
Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024
of this extra amount of energy via exciton-­phonon scattering events
with typical phonon frequencies of 0.03 eV (43) leads to overall
slower hole-­transfer dynamics (arrows in Fig. 2A) (42, 44). In addi-
tion, the first step of the exciton cascade leading to the formation of
either KW-­Σ or Γ-­KMo excitons in the electron-­and hole-­transfer
process, respectively, is markedly different. In the first Brillouin zone,
the Σ and Σ' valleys are each threefold degenerate, while there is
only one Γ valley (Fig. 2A, inset). Therefore, the density of final states
for the KW-­KW → KW-­Σ versus the KMo-­KMo → Γ-­KMo transition
is notably different (42, 43, 45–47). In consequence, hybrid KW-­Σ
excitons are more efficiently formed than hybrid Γ-­KMo excitons,
favoring faster interlayer exciton formation dynamics for the
electron-­transfer channel compared to the hole-­transfer channel.
Last, we want to point out two important deviations in the exci-
ton dynamics between experiment and theory. First, on the few
picosecond timescale, we find that the calculated occupation of in-
terlayer KW-­KMo excitons increases up to ≈4 ps and is composed of
a 1:5 ratio of interlayer excitons that are formed from A1s excitons
initially excited in the WSe2 and MoS2 layers (fig. S8). In contrast, in
the experiment, the respective photoemission intensity saturates at
roughly 1 ps and the 1:5 ratio cannot be identified (1.9-­eV excitation;
Fig. 3). This deviation between experiment and theory can be un-
derstood by the fact that radiative and defect-­assisted decay pro-
cesses of intralayer, hybrid, and interlayer excitons with lifetimes
ranging from 1 ps to tenths of picoseconds (8, 27, 34, 35) are not
included in the model calculations. Hence, the model calculations
overestimate the exciton occupation at large pump-­probe delays.
Second, we find that the experimental data for 1.7-­and 1.9-­eV
excitation rises faster than estimated from the model calculations
(sub-­200-­fs timescale in Fig. 3). This deviation could be related to
the fact that the model calculations do not consider exciton-­exciton
scattering events, which might already contribute to the dynamics
in the experiment (25, 48, 49). Although an in-­depth pump fluence-­
dependent analysis of these dynamics appears to be highly interest-
ing, it is beyond the scope of this manuscript, and, in the following,
we focus on the identification of a spectroscopic fingerprint of the
hole-­transfer process.
The spectroscopic signature of a correlated
hole-­transfer process
On the basis of this hierarchy of timescales between the electron-­
and the hole-­transfer process, it is possible to separate the interlayer
exciton formation dynamics: For delays >200 fs, the change in the
exciton photoemission yield from the interlayer KW-­KMo exciton is
mainly caused by hole-­transfer processes. Hence, the final ambition
of our work is the unambiguous discrimination of the photoemis-
sion spectral signature of intralayer KMo-­KMo and interlayer KW-­KMo
excitons, where, in both cases, the electron contribution to the exci-
ton is situated in the conduction band minimum of the MoS2 layer
(compare Fig. 1C).
In the most naive picture of photoemission, it might be expected
that trARPES only yields information on the exciton’s electron.
Hence, the experiment would not distinguish between photoelectrons
being emitted from the conduction band minimum of MoS2, irre-
spective of whether they result from the breakup of intralayer KMo-­
KMo or interlayer KW-­KMo excitons (Fig. 1C). However, it is known
that the spectral function in photoemission contains information
about many-­body interactions (50), and this is also the case for the
correlated electron-­hole pair. This leads to a very nonintuitive and
4 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

intriguing experimental observation. Figure 4A shows the pump-­
probe delay evolution of energy distribution curves (EDCs) filtered
for photoelectron yield at the κ valley, whose momentum coincides
with the KMo valley, i.e., the momentum region where photoelectron
yield from intralayer KMo-­KMo and interlayer KW-­KMo excitons is
expected (Fig. 4A, inset). Astonishingly, we find that the energy of
the photoelectrons shifts up as a function of pump-­probe delay from
E − EvW = 0.93 ± 0.03 eV at 15 fs to E − EvW = 1.10 ± 0.03 eV at 1 ps,
i.e., a shift of ΔEPES
h−transfer
= 0.17 ± 0.04 eV (Fig. 4B). At first glance,
this is an unexpected observation: In temporal overlap of the pump
and the probe laser pulses, the optical excitation deposits energy into
the system, and the system subsequently relaxes from its excited
state to energetically more favorable states via scattering processes.
In consequence, energy-­resolved pump-­probe photoemission spec-
troscopies of single-­particle charge carriers typically show that the
mean kinetic energy of the photoelectrons decreases with pump-­
probe delay (51). An increasing mean kinetic energy might indicate
higher-­order scattering processes such as Auger recombination (49, 52).
For Auger recombination, however, we would expect to observe a
decreasing mean kinetic energy on the few-­picosecond timescale as
the overall exciton density and thus the efficiency for Auger recom-
bination decreases. However, the long-­time evaluation of the mean
photoelectron energy clearly excludes this scenario (Fig. 4A). In ad-
dition, by evaluating the pump-­probe delay evolution of the energy
position of the MoS2 valence band maxima, we can exclude a photo-
induced renormalization of the band energies (53, 54) (fig. S5). We thus
search for the origin of the apparent increase of the mean kinetic
energy beyond the single-­particle picture, i.e., in the photoemission
from excitons whose occupation is dynamically transferring from
intralayer KMo-­KMo to interlayer KW-­KMo excitons.
So far, we have referenced the energies of all emitted single-­
particle photoelectrons to the valence band maximum of WSe2 (left
energy axis in Fig. 4B). However, especially for the intralayer KMo-­
KMo exciton that fully resides in the MoS2 layer, this is clearly not the
intrinsically relevant energy axis. We overcome this shortcoming by
using an energy scale that is more direct to photoemission from ex-
citons by relating the total energy before (E = E0 + Eexc + ℏω) and
after (E = E0 − Ehole + Eelec) the breakup of the correlated electron-­
hole pair (55). Here, Eexc is the energy necessary to resonantly excite
an exciton with a two-­particle binding energy Ebin (compare exciton
energy landscape in Fig. 2A); Ehole and Eelec denote the energy of the
single-­particle hole and electron state after the breakup of the exciton,
respectively; E0 is the ground state energy and ℏω is the photon en-
ergy. As energy needs to be conserved when the exciton is broken, the
energy of the detected single-­particle electron can be expressed as
Eelec = Ehole + Eexc + ℏ𝜔 (1)
Therefore, Eq. 1 fixes the energy of the single-­particle hole Ehole
remaining in the sample as the natural reference point of the photo-
electron energy axis for each probed exciton (at a given probe photon
energy ℏω). For the intralayer KMo-­KMo excitons and the interlayer
KW-­KMo excitons, respectively, the valence band maxima of MoS2
( EvMo) and WSe2 ( EvW ) set the energy scale [see Fig. 1 (B and C) and
band energies labeled in Fig. 2B)]. Following Eq. 1, we can directly
quantify the exciton energies of intralayer KMo-­KMo and interlayer KW-­KMo
excitons from the photoemission data to Eexc MoMo = 1.93 ± 0.08 eV and
WMo = 1.10 ± 0.03 eV
Eexc , respectively, which are in excellent agreement
with earlier results obtained with photoluminescence spectroscopy
[(Eexc,PL
MoMo
= 1.9 eV and Eexc,PL
WMo
= 1.1 eV; horizontal lines in Fig. 4B)]

Fig. 4. Coulomb correlation–induced excitonic energy fingerprints. (A) Pump-­probe delay evolution of the energy distribution curves (EDCs) filtered at the momen-
tum region of the KMo valley of MoS2 (region of interest indicated in the inset, 1.9-­eV excitation). At this high-­symmetry point, photoemission yield from intralayer KMo-­KMo
and interlayer KW-­KMo excitons is expected (see Fig. 1C). As intralayer KMo-­KMo excitons decay and form interlayer KW-­KMo excitons, the peak maxima of the photoelectron
energy shows an upshift by ΔEPESh−transfer
= 0.17 ± 0.04 eV (curved arrow). (B) Selected EDCs for pump-­probe delays of 15 fs (dark red) and 1 ps (black) illustrating an ener-
getic upshift of the exciton photoemission signal. The horizontal bars indicate expected photoelectron energies for the intralayer KMo-­KMo (dark red) and interlayer KW-­KMo
(black) excitons calculated with Eq. 1 and data from photoluminescence measurements (56, 57). The left and right energy axes in black and dark red show the correspond-
ing energy scales with respect to the valence band maximum of WSe2 and MoS2.

Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024 5 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(56, 57). In consequence, we can explain the experimentally observed
upshift of the photoelectron energy by ΔEPES h−transfer
= 0.17 ± 0.04 eV
with the energy difference between the single-­particle electron final
states Eelec of the interlayer KW-­KMo and the intralayer KMo-­KMo exci-
tons, i.e., with (EvW + Eexc
WMo + ℏω) − (E Mo + E MoMo + ℏω) ≈ 0.17 eV
v exc
(with EvW − EvMo = 1.00 ± 0.07 eV, see Fig. 2B). Hence, the energetic
upshift is a direct consequence of the breakup of the correlated
electron-­hole pair during the photoemission process.
Although the photoelectron energy increases during the hole-­
transfer process, we strongly emphasize that the overall energy of the
system relaxes by ΔEexc h−transfer = E WMo − E MoMo = − 0.83 ± 0.09 eV
exc exc
(see Fig. 2A). Consistently, if the same analysis is performed for the
electron-­only transfer process after photoexcitation with 1.7 eV pump
pulses, then we find a reduction of the overall exciton energy by
e−transfer = E WMo − E WW = − 0.46 ± 0.07 eV (fig. S6). In this case,
ΔEexc exc exc
where the exciton’s hole remains in the WSe2 VBM (Fig. 1B), the
reduction of the exciton energy directly translates to a reduction of the
single-­particle photoelectron energy (ΔEPESe−transfer
= − 0.46 ± 0.07 eV).
Therefore, as expected, interlayer charge transfer always leads to a
reduction of the exciton energy Eexc, which might, however, result in
an up-­or a downshift of the photoelectron energy in the photoemis-
sion spectrum.
DISCUSSION
We have shown that femtosecond momentum microscopy is a pow-
erful tool to study the correlated interaction between the exciton’s
electron and hole in twisted semiconductor heterostructures. Exem-
plarily, we show that the photoelectron of the correlated two-­particle
exciton contains direct information about the hole state. We use this
correlation in combination with microscopic and material-­specific
theory to directly follow an ultrafast interlayer hole-­transfer process
that would otherwise be elusive. Our work opens up means for the
future study of correlated states of matter in two-­dimensional quan-
tum materials.
MATERIALS AND METHODS
The time-­and angle-­resolved photoemission data are measured with
a time-­of-­flight momentum microscope (Surface Concept) (58, 59)
that is connected to a table-­top high harmonic generation beamline
driven by a 300-­W fiber laser system (AFS Jena) (40, 60). The overall
experimental setup and its application to exfoliated two-­dimensional
materials are described in (39) and (8), respectively.
In all experiments, the exciton dynamics are induced by resonant
optical excitation of the A1s-­excitons of WSe2 or MoS2. Therefore,
1.7-­and 1.9-­eV pump pulses with a duration of 50 fs are used (s-­
polarized), respectively. After a variable pump-­probe delay, photo-
emission is induced by 26.5-­eV light pulses (20 fs, p-­polarized).
For the characterization of the temporal resolution and the de-
termination of absolute time zero of the experiment, we have mea-
sured the pump-­probe delay-­dependent photoemission yield of
sidebands of the valence bands formed due to the laser-­assisted pho-
toelectric effect (40, 61). In fig. S1, a cross-­correlation of the pump
and probe laser pulse is shown, where both laser pulses are p-­
polarized (1.9-­eV pump pulses). The gray line is a Gaussian fit to the
data yielding a full width at half maximum of 60 ± 5 fs.
The 9.8∘ ± 0.8∘ twisted WSe2/MoS2 heterostructure is stamped
onto a 20-­to 30-­nm-­thick hBN (62) spacer layer and a p +-­doped
Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024
native oxide silicon waver. Before the momentum microscopy ex-
periments, the sample is annealed for 1 hour to 670 K. Details on the
sample fabrication and characterization (e.g., twist angle) are de-
scribed in (8).
Supplementary Materials
This PDF file includes:
Supplementary Text
Figs. S1 to S8
Table S1
References
REFERENCES AND NOTES
1. G. Wang, A. Chernikov, M. M. Glazov, T. F. Heinz, X. Marie, T. Amand, B. Urbaszek,
Colloquium: Excitons in atomically thin transition metal dichalcogenides. Rev. Mod. Phys.
90, 021001 (2018).
2. A. Chernikov, T. C. Berkelbach, H. M. Hill, A. Rigosi, Y. Li, O. B. Aslan, D. R. Reichman,
M. S. Hybertsen, T. F. Heinz, Exciton binding energy and nonhydrogenic rydberg series in
monolayer WS2. Phys. Rev. Lett. 113, 076802 (2014).
3. K. He, N. Kumar, L. Zhao, Z. Wang, K. F. Mak, H. Zhao, J. Shan, Tightly bound excitons in
monolayer WSe2. Phys. Rev. Lett. 113, 026803 (2014).
4. C.-­H. Lee, G.-­H. Lee, A. M. van der Zande, W. Chen, Y. Li, M. Han, X. Cui, G. Arefe,
C. Nuckolls, T. F. Heinz, J. Guo, J. Hone, P. Kim, Atomically thin p-­n junctions with
van der Waals heterointerfaces. Nat. Nanotechnol. 9, 676–681 (2014).
5. X. Hong, J. Kim, S.-­F. Shi, Y. Zhang, C. Jin, Y. Sun, S. Tongay, J. Wu, Y. Zhang, F. Wang,
Ultrafast charge transfer in atomically thin MoS2/WS2 heterostructures. Nat. Nanotechnol.
9, 682–686 (2014).
6. F. Ceballos, M. Z. Bellus, H.-­Y. Chiu, H. Zhao, Ultrafast charge separation and indirect
exciton formation in a MoS2-­MoSe2 van der waals heterostructure. ACS Nano 8,
12717–12724 (2014).
7. P. Merkl, F. Mooshammer, P. Steinleitner, A. Girnghuber, K. Q. Lin, P. Nagler, J. Holler,
C. Schüller, J. M. Lupton, T. Korn, S. Ovesen, S. Brem, E. Malic, R. Huber, Ultrafast transition
between exciton phases in van der Waals heterostructures. Nat. Mater. 18, 691–696
(2019).
8. D. Schmitt, J. P. Bange, W. Bennecke, A. AlMutairi, G. Meneghini, K. Watanabe, T. Taniguchi,
D. Steil, D. R. Luke, R. T. Weitz, S. Steil, G. S. M. Jansen, S. Brem, E. Malic, S. Hofmann,
M. Reutzel, S. Mathias, Formation of moiré interlayer excitons in space and time. Nature
608, 499–503 (2022).
9. K. L. Seyler, P. Rivera, H. Yu, N. P. Wilson, E. L. Ray, D. G. Mandrus, J. Yan, W. Yao, X. Xu,
Signatures of moiré-­trapped valley excitons in MoSe2/WSe2 heterobilayers. Nature 567,
66–70 (2019).
10. E. M. Alexeev, D. A. Ruiz-­Tijerina, M. Danovich, M. J. Hamer, D. J. Terry, P. K. Nayak, S. Ahn,
S. Pak, J. Lee, J. I. Sohn, M. R. Molas, M. Koperski, K. Watanabe, T. Taniguchi,
K. S. Novoselov, R. V. Gorbachev, H. S. Shin, V. I. Fal’ko, A. I. Tartakovskii, Resonantly
hybridized excitons in moiré superlattices in van der Waals heterostructures. Nature 567,
81–86 (2019).
11. K. Tran, G. Moody, F. Wu, X. Lu, J. Choi, K. Kim, A. Rai, D. A. Sanchez, J. Quan, A. Singh,
J. Embley, A. Zepeda, M. Campbell, T. Autry, T. Taniguchi, K. Watanabe, N. Lu,
S. K. Banerjee, K. L. Silverman, S. Kim, E. Tutuc, L. Yang, A. H. MacDonald, X. Li, Evidence for
moiré excitons in van der Waals heterostructures. Nature 567, 71–75 (2019).
12. O. Karni, E. Barré, V. Pareek, J. D. Georgaras, M. K. L. Man, C. Sahoo, D. R. Bacon, X. Zhu,
H. B. Ribeiro, A. L. O’Beirne, J. Hu, A. Al-­Mahboob, M. M. M. Abdelrasoul, N. S. Chan,
A. Karmakar, A. J. Winchester, B. Kim, K. Watanabe, T. Taniguchi, K. Barmak, J. Madéo,
F. H. da Jornada, T. F. Heinz, K. M. Dani, Structure of the moiré exciton captured by
imaging its electron and hole. Nature 603, 247–252 (2022).
13. L. Ma, P. X. Nguyen, Z. Wang, Y. Zeng, K. Watanabe, T. Taniguchi, A. H. MacDonald,
K. F. Mak, J. Shan, Strongly correlated excitonic insulator in atomic double layers. Nature
598, 585–589 (2021).
14. Z. Zhang, E. C. Regan, D. Wang, W. Zhao, S. Wang, M. Sayyad, K. Yumigeta, K. Watanabe,
T. Taniguchi, S. Tongay, M. Crommie, A. Zettl, M. P. Zaletel, F. Wang, Correlated interlayer
exciton insulator in heterostructures of monolayer WSe2 and moiré WS2/WSe2. Nat. Phys.
18, 1214–1220 (2022).
15. Y. Slobodkin, Y. Mazuz-­Harpaz, S. Refaely-­Abramson, S. Gazit, H. Steinberg, R. Rapaport,
Quantum phase transitions of trilayer excitons in atomically thin heterostructures.
Phys. Rev. Lett. 125, 255301 (2020).
16. L. Sigl, F. Sigger, F. Kronowetter, J. Kiemle, J. Klein, K. Watanabe, T. Taniguchi, J. J. Finley,
U. Wurstbauer, A. W. Holleitner, Signatures of a degenerate many-­body state of interlayer
excitons in a van der waals heterostack. Phys. Rev. Res. 2, 042044 (2020).
6 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
17. Z. Wang, D. A. Rhodes, K. Watanabe, T. Taniguchi, J. C. Hone, J. Shan, K. F. Mak, Evidence of
high-­temperature exciton condensation in two-­dimensional atomic double layers.
Nature 574, 76–80 (2019).
18. C. Jin, E. Y. Ma, O. Karni, E. C. Regan, F. Wang, T. F. Heinz, Ultrafast dynamics in van der
Waals heterostructures. Nat. Nanotechnol. 13, 994–1003 (2018).
19. R. Perea-­Causin, D. Erkensten, J. M. Fitzgerald, J. J. P. Thompson, R. Rosati, S. Brem, E. Malic,
Exciton optics, dynamics, and transport in atomically thin semiconductors. APL Mater. 10,
100701 (2022).
20. C. Poellmann, P. Steinleitner, U. Leierseder, P. Nagler, G. Plechinger, M. Porer,
R. Bratschitsch, C. Schüller, T. Korn, R. Huber, Resonant internal quantum transitions and
femtosecond radiative decay of excitons in monolayer WSe2. Nat. Mater. 14, 889–893
(2015).
21. E. Malic, M. Selig, M. Feierabend, S. Brem, D. Christiansen, F. Wendler, A. Knorr,
G. Berghäuser, Dark excitons in transition metal dichalcogenides. Phys. Rev. Mater. 2,
014002 (2018).
22. T. Mueller, E. Malic, Exciton physics and device application of two-­dimensional transition
metal dichalcogenide semiconductors. NPJ 2D Mater. Appl. 2, 29 (2018).
23. J. Madéo, M. K. L. Man, C. Sahoo, M. Campbell, V. Pareek, E. L. Wong, A. Al-­Mahboob,
N. S. Chan, A. Karmakar, B. M. K. Mariserla, X. Li, T. F. Heinz, T. Cao, K. M. Dani, Directly
visualizing the momentum-­forbidden dark excitons and their dynamics in atomically thin
semiconductors. Science 370, 1199–1204 (2020).
24. R. Wallauer, R. Perea-­Causin, L. Münster, S. Zajusch, S. Brem, J. Güdde, K. Tanimura,
K.-­Q. Lin, R. Huber, E. Malic, U. Höfer, Momentum-­resolved observation of exciton
formation dynamics in monolayer WS2. Nano Lett. 21, 5867–5873 (2021).
25. A. Kunin, S. Chernov, J. Bakalis, Z. Li, S. Cheng, Z. H. Withers, M. G. White,
G. Schönhense, X. Du, R. K. Kawakami, T. K. Allison, Momentum-­resolved exciton
coupling and valley polarization dynamics in monolayer WS2. Phys. Rev. Lett. 130,
046202 (2023).
26. D. Schmitt, J. P. Bange, W. Bennecke, G. Meneghini, A. AlMutairi, M. Merboldt, J. Pöhls,
K. Watanabe, T. Taniguchi, S. Steil, D. Steil, R. T. Weitz, S. Hofmann, S. Brem, G. S. Matthijs
Jansen, E. Malic, S. Mathias, M. Reutzel, Ultrafast nano-­imaging of dark excitons.
arXiv:2305.18908 (2023).
27. J. P. Bange, P. Werner, D. Schmitt, W. Bennecke, G. Meneghini, A. AlMutairi, M. Merboldt,
K. Watanabe, T. Taniguchi, S. Steil, D. Steil, R. T. Weitz, S. Hofmann, G. S. M. Jansen,
S. Brem, E. Malic, M. Reutzel, S. Mathias, Ultrafast dynamics of bright and dark excitons
in monolayer WSe2 and heterobilayer WSe2/MoS2. 2D Mater. 10, 035039 (2023).
28. E. Perfetto, D. Sangalli, A. Marini, G. Stefanucci, First-­principles approach to excitons in
time-­resolved and angle-­resolved photoemission spectra. Phys. Rev. B 94, 245303
(2016).
29. A. Rustagi, A. F. Kemper, Photoemission signature of excitons. Phys. Rev. B 97, 235310
(2018).
30. A. Steinhoff, M. Florian, M. Rösner, G. Schönhoff, T. O. Wehling, F. Jahnke, Exciton fission in
monolayer transition metal dichalcogenide semiconductors. Nat. Commun. 8, 1166
(2017).
31. D. Christiansen, M. Selig, E. Malic, R. Ernstorfer, A. Knorr, Theory of exciton dynamics in
time-­resolved ARPES: Intra-­and intervalley scattering in two-­dimensional
semiconductors. Phys. Rev. B 100, 205401 (2019).
32. S. Dong, M. Puppin, T. Pincelli, S. Beaulieu, D. Christiansen, H. Hübener,
C. W. Nicholson, R. P. Xian, M. Dendzik, Y. Deng, Y. W. Windsor, M. Selig, E. Malic,
A. Rubio, A. Knorr, M. Wolf, L. Rettig, R. Ernstorfer, Direct measurement of key exciton
properties: Energy, dynamics, and spatial distribution of the wave function. Nat. Sci. 1,
e10010 (2021).
33. M. K. L. Man, J. Madéo, C. Sahoo, K. Xie, M. Campbell, V. Pareek, A. Karmakar, E. L. Wong,
A. Al-­Mahboob, N. S. Chan, D. R. Bacon, X. Zhu, M. M. M. Abdelrasoul, X. Li, T. F. Heinz,
F. H. da Jornada, T. Cao, K. M. Dani, Experimental measurement of the intrinsic excitonic
wave function. Sci. Adv. 7, eabg0192 (2021).
34. H. Zhu, J. Wang, Z. Gong, Y. D. Kim, J. Hone, X. Y. Zhu, Interfacial charge transfer
circumventing momentum mismatch at two-­dimensional van der Waals heterojunctions.
Nano Lett. 17, 3591–3598 (2017).
35. J. E. Zimmermann, M. Axt, F. Mooshammer, P. Nagler, C. Schüller, T. Korn, U. Höfer,
G. Mette, Ultrafast charge-­transfer dynamics in twisted MoS2/WSe2 heterostructures. ACS
Nano 15, 14725–14731 (2021).
36. V. R. Policht, M. Russo, F. Liu, C. Trovatello, M. Maiuri, Y. Bai, X. Zhu, S. Dal Conte, G. Cerullo,
Dissecting interlayer hole and electron transfer in transition metal dichalcogenide
heterostructures via two-­dimensional electronic spectroscopy. Nano Lett. 21, 4738–4743
(2021).
37. Z. Wang, P. Altmann, C. Gadermaier, Y. Yang, W. Li, L. Ghirardini, C. Trovatello, M. Finazzi,
L. Duò, M. Celebrano, R. Long, D. Akinwande, O. V. Prezhdo, G. Cerullo, S. D. Conte,
Phonon-­mediated interlayer charge separation and recombination in a MoSe2/WSe2
heterostructure. Nano Lett. 21, 2165–2173 (2021).
38. G. Meneghini, M. Reutzel, S. Mathias, S. Brem, E. Malic, Hybrid exciton signatures in ARPES
spectra of van der waals materials. ACS Photonics 10, 3570–3575 (2023).
Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024
39. M. Keunecke, C. Möller, D. Schmitt, H. Nolte, G. S. M. Jansen, M. Reutzel, M. Gutberlet,
G. Halasi, D. Steil, S. Steil, S. Mathias, Time-­resolved momentum microscopy with a 1 MHz
high-­harmonic extreme ultraviolet beamline. Rev. Sci. Instrum. 91, 063905 (2020).
40. M. Keunecke, M. Reutzel, D. Schmitt, A. Osterkorn, T. A. Mishra, C. Möller, W. Bennecke,
G. S. M. Jansen, D. Steil, S. R. Manmana, S. Steil, S. Kehrein, S. Mathias, Electromagnetic
dressing of the electron energy spectrum of Au(111) at high momenta. Phys. Rev. B 102,
161403 (2020).
41. Y. Li, A. Chernikov, X. Zhang, A. Rigosi, H. M. Hill, A. M. van der Zande, D. A. Chenet,
E.-­M. Shih, J. Hone, T. F. Heinz, Measurement of the optical dielectric function of
monolayer transition-­metal dichalcogenides: MoS2, MoSe2, WS2, and WSe2. Phys. Rev. B
90, 205422 (2014).
42. G. Meneghini, S. Brem, E. Malic, Ultrafast phonon-­driven charge transfer in van der Waals
heterostructures. Nat. Sci. 2, e20220014 (2022).
43. Z. Jin, X. Li, J. T. Mullen, K. W. Kim, Intrinsic transport properties of electrons and holes in
monolayer transition-­metal dichalcogenides. Phys. Rev. B 90, 045422 (2014).
44. S. Ovesen, S. Brem, C. Linderälv, M. Kuisma, T. Korn, P. Erhart, M. Selig, E. Malic,
Interlayer exciton dynamics in van der Waals heterostructures. Commun. Phys. 2, 23
(2019).
45. M. Selig, G. Berghäuser, M. Richter, R. Bratschitsch, A. Knorr, E. Malic, Dark and bright
exciton formation, thermalization, and photoluminescence in monolayer transition metal
dichalcogenides. 2D Mater. 5, 035017 (2018).
46. A. Raja, M. Selig, G. Berghäuser, J. Yu, H. M. Hill, A. F. Rigosi, L. E. Brus, A. Knorr, T. F. Heinz,
E. Malic, A. Chernikov, Enhancement of exciton-­phonon scattering from monolayer to
bilayer WS2. Nano Lett. 18, 6135–6143 (2018).
47. J. Lindlau, M. Selig, A. Neumann, L. Colombier, J. Förste, V. Funk, M. Förg, J. Kim,
G. Berghäuser, T. Taniguchi, K. Watanabe, F. Wang, E. Malic, A. Högele, The role of
momentum-­dark excitons in the elementary optical response of bilayer WSe2. Nat. Commun.
9, 2586 (2018).
48. D. Sun, Y. Rao, G. A. Reider, G. Chen, Y. You, L. Brézin, A. R. Harutyunyan, T. F. Heinz,
Observation of rapid exciton-­exciton annihilation in monolayer molybdenum disulfide.
Nano Lett. 14, 5625–5629 (2014).
49. D. Erkensten, S. Brem, K. Wagner, R. Gillen, R. Perea-­Causín, J. D. Ziegler, T. Taniguchi,
K. Watanabe, J. Maultzsch, A. Chernikov, E. Malic, Dark exciton-­exciton annihilation in
monolayer WSe2. Phys. Rev. B 104, L241406 (2021).
50. A. Damascelli, Z. Hussain, Z.-­X. Shen, Angle-­resolved photoemission studies of the
cuprate superconductors. Rev. Mod. Phys. 75, 473–541 (2003).
51. M. Bauer, A. Marienfeld, M. Aeschlimann, Hot electron lifetimes in metals probed
by time-­resolved two-­photon photoemission. Prog. Surf. Sci. 90, 319–376 (2015).
52. S. Mathias, S. Eich, J. Urbancic, S. Michael, A. V. Carr, S. Emmerich, A. Stange,
T. Popmintchev, T. Rohwer, M. Wiesenmayer, A. Ruffing, S. Jakobs, S. Hellmann,
P. Matyba, C. Chen, L. Kipp, M. Bauer, H. C. Kapteyn, H. C. Schneider, K. Rossnagel,
M. M. Murnane, M. Aeschlimann, Self-­amplified photo-­induced gap quenching in a
correlated electron material. Nat. Commun. 7, 12902 (2016).
53. A. Chernikov, C. Ruppert, H. M. Hill, A. F. Rigosi, T. F. Heinz, Population inversion and
giant bandgap renormalization in atomically thin WS2 layers. Nat. Photonics 9, 466–470
(2015).
54. F. Liu, M. E. Ziffer, K. R. Hansen, J. Wang, X. Zhu, Direct determination of band-­gap
renormalization in the photoexcited monolayer MoS2. Phys. Rev. Lett. 122, 246803 (2019).
55. M. Weinelt, M. Kutschera, T. Fauster, M. Rohlfing, Dynamics of exciton formation at the
Si(100) c(4 x 2) surface. Phys. Rev. Lett. 92, 126801 (2004).
56. O. Karni, E. Barré, S. C. Lau, R. Gillen, E. Y. Ma, B. Kim, K. Watanabe, T. Taniguchi,
J. Maultzsch, K. Barmak, R. H. Page, T. F. Heinz, Infrared interlayer exciton emission in
MoS−2/WSe2 heterostructures. Phys. Rev. Lett. 123, 247402 (2019).
57. J. Kunstmann, F. Mooshammer, P. Nagler, A. Chaves, F. Stein, N. Paradiso, G. Plechinger,
C. Strunk, C. Schüller, G. Seifert, D. R. Reichman, T. Korn, Momentum-­space indirect
interlayer excitons in transition-­metal dichalcogenide van der Waals heterostructures.
Nat. Phys. 14, 801–805 (2018).
58. K. Medjanik, O. Fedchenko, S. Chernov, D. Kutnyakhov, M. Ellguth, A. Oelsner,
B. Schönhense, T. R. F. Peixoto, P. Lutz, C.-­H. Min, F. Reinert, S. Däster, Y. Acremann,
J. Viefhaus, W. Wurth, H. J. Elmers, G. Schönhense, Direct 3D mapping of the Fermi surface
and Fermi velocity. Nat. Mater. 16, 615–621 (2017).
59. B. Krömker, M. Escher, D. Funnemann, D. Hartung, H. Engelhard, J. Kirschner,
Development of a momentum microscope for time resolved band structure imaging.
Rev. Sci. Instrum. 79, 053702–053707 (2008).
60. M. Düvel, M. Merboldt, J. P. Bange, H. Strauch, M. Stellbrink, K. Pierz, H. W. Schumacher,
D. Momeni, D. Steil, G. S. M. Jansen, S. Steil, D. Novko, S. Mathias, M. Reutzel, Far-­from-­
equilibrium electron-­phonon interactions in optically excited graphene. Nano Lett. 22,
4897–4904 (2022).
61. G. Saathoff, L. Miaja-­Avila, M. Aeschlimann, M. M. Murnane, H. C. Kapteyn, Laser-­assisted
photoemission from surfaces. Phys. Rev. A 77, 022903 (2008).
62. T. Taniguchi, K. Watanabe, Synthesis of high-­purity boron nitride single crystals under
high pressure by using Ba-­BN solvent. J. Cryst. Growth 303, 525–529 (2007).
7 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
63. G. Schönhense, D. Kutnyakhov, F. Pressacco, M. Heber, N. Wind, S. Y. Agustsson,
S. Babenkov, D. Vasilyev, O. Fedchenko, S. Chernov, L. Rettig, B. Schönhense, L. Wenthaus,
G. Brenner, S. Dziarzhytski, S. Palutke, S. K. Mahatha, N. Schirmel, H. Redlin,
B. Manschwetus, I. Hartl, Y. Matveyev, A. Gloskovskii, C. Schlueter, V. Shokeen,
H. Duerr, T. K. Allison, M. Beye, K. Rossnagel, H. J. Elmers, K. Medjanik, Suppression
of the vacuum space-­charge effect in fs-­photoemission by a retarding electrostatic
front lens. Rev. Sci. Instrum. 92, 053703 (2021).
64. S. Brem, K.-­Q. Lin, R. Gillen, J. M. Bauer, J. Maultzsch, J. M. Lupton, E. Malic, Hybridized
intervalley moiré excitons and flat bands in twisted WSe2 bilayers. Nanoscale 12,
11088–11094 (2020).
65. J. Hagel, S. Brem, C. Linderälv, P. Erhart, E. Malic, Exciton landscape in van der Waals
heterostructures. Phys. Rev. Res. 3, 043217 (2021).
66. S. Brem, M. Selig, G. Berghaeuser, E. Malic, Exciton relaxation cascade in two-­dimensional
transition metal dichalcogenides. Sci. Rep. 8, 8238 (2018).
67. M. Kira, S. W. Koch, Many-­body correlations and excitonic effects in semiconductor
spectroscopy. Prog. Quantum Electron. 30, 155–296 (2006).
68. H. Haug, S. W. Koch, Quantum Theory of the Optical and Electronic Properties of
Semiconductors (World Scientific Publishing Company, 2009).
69. E. Malic, A. Knorr, Graphene and Carbon Nanotubes: Ultrafast Optics and Relaxation
Dynamics (John Wiley & Sons, 2013).
70. S. Brem, C. Linderälv, P. Erhart, E. Malic, Tunable phases of moiré excitons in van der Waals
heterostructures. Nano Lett. 20, 8534–8540 (2020).
Bange et al., Sci. Adv. 10, eadi1323 (2024) 7 February 2024
Acknowledgments
Funding: This work was funded by the Deutsche Forschungsgemeinschaft (DFG; German
Research Foundation) (432680300/SFB 1456, project B01; 217133147/SFB 1073, projects B07
and B10; and 223848855/SFB 1083, project B9). A.A. and S.H. acknowledge funding from
EPSRC (EP/T001038/1 and EP/P005152/1). A.A. acknowledges financial support from the Saudi
Arabian Ministry of Higher Education. E.M. acknowledges support from the European Union’s
Horizon 2020 research and innovation program under grant agreement no. 881603 (Graphene
Flagship). K.W. and T.T. acknowledge support from JSPS KAKENHI (grant numbers 19H05790,
20H00354, and 21H05233). Author contributions: D.St., S.S., R.T.W., G.S.M.J., S.H., S.B., E.M.,
M.R., and S.M. conceived the research. D.Sc., J.P.B., and W.B. carried out the time-­resolved
momentum microscopy experiments. J.P.B. and D.Sc. analyzed the data. G.M. performed the
microscopic model calculations. A.A. fabricated the heterostructure sample. All authors
discussed the results. M.R. and S.M. were responsible for the overall project direction and
wrote the manuscript with contributions from all coauthors. K.W. and T.T. synthesized the hBN
crystals. Competing interests: The authors declare that they have no competing interests.
Data and materials availability: All data needed to evaluate the conclusions of the paper are
present in the paper and/or the Supplementary Materials.
Submitted 5 April 2023
Accepted 10 January 2024
Published 7 February 2024
10.1126/sciadv.adi1323
8 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

OPTICS Copyright © 2024 The

Authors, some rights
Broadband and large-­aperture metasurface edge reserved; exclusive
licensee American
encoders for incoherent infrared radiation Association for the
Advancement of
Science. No claim to
Brandon T. Swartz1, Hanyu Zheng1, Gregory T. Forcherio2, Jason Valentine1* original U.S.
Government Works.
The prevalence of computer vision systems necessitates hardware-­based approaches to relieve the high compu- Distributed under a
tational demand of deep neural networks in resource-­limited applications. One solution would be to off-­load low-­ Creative Commons
level image feature extraction, such as edge detection, from the digital network to the analog imaging system. To Attribution
that end, this work demonstrates incoherent, broadband, low-­noise optical edge detection of real-­world scenes NonCommercial
by combining the wavefront shaping of a 24-­mm aperture metasurface with a refractive lens. An inverse design License 4.0 (CC BY-­NC).
approach is used to optimize the metasurface for Laplacian-­based edge detection across the 7.5-­to 13.5-­μm LWIR
imaging band, allowing for facile integration with uncooled microbolometer-­based LWIR imagers to encode edge
information. A polarization multiplexed approach leveraging a birefringent metasurface is also demonstrated as
a single-­aperture implementation. This work could be applied to improve computer vision capabilities of resource-­
constrained systems by leveraging optical preprocessing to alleviate the computational requirements for high-­
accuracy image segmentation and classification.

INTRODUCTION apertures, a single aperture with wavelength (23, 24), or polarization

Convolutional neural networks (CNNs) are widely used in computer
vision for image segmentation and classification, owing to their high
accuracy and versatility (1). Increasing the complexity of a CNN by
adding layers or channels generally improves performance but comes
at the expense of increased power consumption and calculation time
(2), as well as increases in system size and weight, putting practical
performance limits on resource-­constrained systems (3). Although
network architectures (2) and hardware accelerators (3) have gradu-
ally improved in recent years, the imaging systems in computer vision
still have a largely untapped potential for improving performance and
efficiency of the computational tasks. One avenue for improvement is
to off-­load low-­level feature encoding operations, such as edge detec-
tion, into an optical front-­end (4). Off-­loading these operations,
which are typically carried out by the first layers of a CNN (1, 5), into
an analog optical system could facilitate faster and more energy-­
efficient CNNs without sacrificing accuracy.
Optical edge detection is readily achievable under coherent illumi-
nation because interference in coherent systems allows for direct,
fully optical edge detection. A conventional approach has been to use
a 4f system to apply a frequency space filter to incoming light (6–9),
although recent approaches using metasurfaces (10–16) have achieved
similar filtering in a more compact form; photonic crystal slabs (17,
18), multilayer films (19, 20), and dielectric reflectors (21) have also
been proposed. Most real-­world imaging applications are passive, in-
volving incoherent illumination or thermal emission, for which it is
not possible to perform fully optical edge detection due to inherent
constraints on the incoherent point spread function (PSF). The inco-
herent PSF is an intensity distribution, thus is constrained to real,
positive values, whereas all PSFs capable of edge detection require
both positive and negative values (22). However, a bipolar PSF can
be synthesized with a hybrid optoelectronic system, which collects
two images using two different PSFs and then performs pixel-­by-­pixel
digital subtraction (6). This can be achieved with two separate
1 ventional refractive lens with a metasurface corrector (37). In this
Vanderbilt University, 2301 Vanderbilt Place, Nashville, TN 37235, USA. 2Naval Sur-
face Warfare Center, Crane, 300 Highway 361, Crane, IN 47522, USA.
*Corresponding author. Email: jason.​g.​valentine@​vanderbilt.​edu
(25) multiplexing.
While previous approaches have successfully demonstrated inco-
herent edge detection in simulated or laboratory conditions (23–25),
they have not fully addressed the challenges associated with edge de-
tection in real-­world scenes involving ambient illumination or ther-
mal emission. Incoherent light sources are inherently broadband;
thus, practical optic design must consider chromatic aberration. In
addition, the broadband PSFs of the optical system will determine not
only the quality and resolution of the edge images but also the signal-­
to-­noise ratio (SNR) because of noise accumulation during digital
subtraction (26, 27). A bandwidth filter or reduced aperture could be
used to simplify the optic design, but doing either would block valu-
able signal, limiting the sensitivity of the system and hindering its use
in low-­light settings. This is particularly true in the long-­wave infra-
red (LWIR) regime, which has become ubiquitous in autonomous
systems due to the ability to see in low-­visibility (28, 29) conditions
and to identify objects based on their thermal signature (30). Imple-
mentation of optical edge detection for LWIR thermal emission is es-
pecially difficult because thermal emission is not only incoherent and
broadband but also limited in signal and contrast (29). Therefore, to
achieve optimal performance, the optics should be designed to oper-
ate across the camera’s entire bandwidth and aperture while simulta-
neously considering the quality, resolution, and signal in the resulting
edge images.
Metasurfaces offer a compact platform for efficient control of
phase, amplitude, and polarization of light, which makes them a valu-
able tool for shaping PSFs for optical edge detection. The aperture
size, numerical aperture, and operational bandwidth of metalenses
are related, however, ultimately leading to narrowing of the opera-
tional bandwidth as the lens size is increased (31, 32). To address this
issue, various strategies such as multilayer (33), cascaded (34), aperi-
odic (35), or extended depth-­of-­focus (36) metasurfaces have been
explored to improve chromatic performance. Nevertheless, these ap-
proaches still face such challenges as limited focusing efficiency or
small device size. An alternative solution involves combining a con-
approach, the refractive lens provides naturally broadband focusing,
while the metasurface imparts a wavelength-­dependent phase

Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024 1 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
correction—with a substantially relaxed maximum dispersion re-
quirement—leading to large-­scale, broadband performance.
Here, we demonstrate an LWIR electro-­optical edge detection sys-
tem that works in real-­world conditions by combining a refractive
optic with a large-­aperture, inversely designed metasurface to create
positive and negative components of a bipolar Laplacian kernel. The
system is demonstrated first in a two-­aperture approach, followed by
extension into a single-­aperture architecture using polarization multi-
plexing. In both cases, combination with a refractive optic allows us to
overcome the typical bandwidth limitations associated with large ap-
erture metalenses. In addition, inverse design is used to optimize per-
formance for broadband LWIR emission while also preserving the
SNR in the synthesized edge images. The system is used for LWIR
electro-­optical edge detection of thermal emission from real-­world
scenes, which demonstrates the viability of this approach for imple-
mentation as an accelerated front-­end for practical computer vision
systems.
RESULTS
Principals of design
Designing an optic for edge detection can be approached by shaping
the effective PSF into a derivative kernel, such as the Laplacian of
Gaussian (LoG) kernel (38), as shown in Fig. 1. Any kernel used for
edge detection must have both positive and negative values. In con-
trast with coherent edge detection systems (18) that benefit from lin-
earity with respect to the electric fields of light, incoherent imaging
systems are linear with respect to light intensity, which constrains the
incoherent PSF to only positive values. Hence, it is not possible to op-
tically detect edges with a single PSF under incoherent illumination.
Bipolar edge detection kernels can be synthesized by an optoelectronic
system, however, by separating the desired kernel into its constituent
Fig. 1. Principal of bipolar optoelectronic edge detection from incoherent il-
lumination. The Laplacian (∇2) kernel is implemented as a LoG function, which can
be split into two components: a positive-­valued component approximated by the
PSF of a refractive lens and a negative-­valued component approximated by the
PSF of a spiral phase metasurface and a refractive lens.
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
positive and negative components. Figure 1 shows how the LoG ker-
nel can be decomposed into positive and negative component PSFs,
which can be formed by an incoherent imaging system; a typical re-
fractive lens can form the positive component, and the combination
of a specifically designed phase plate with a refractive lens can form an
annular ring as the negative component. For example, a spiral phase
plate produces an annular ring-­shaped PSF due to a phase singularity
at the center of the focal spot (39). Digital subtraction of images col-
lected with the two optics under the same acquisition conditions
yields an edge-­ encoded image corresponding to an approximate
LoG kernel.
Metasurface design
While the spiral phase plate illustrated in Fig. 1 produces a Laplac­
ian kernel, it is not optimized for operational bandwidth, edge reso-
lution, or SNR. To realize an imaging system where these factors are
considered, an inverse design process was used for optimization of
the negative component of the PSF. The positive component was
taken as the PSF from the 24-­mm F/1 refractive lens of an off-­the-­
shelf FLIR Boson 640, and the negative component is a superposi-
tion of the same refractive lens with an optimized metasurface. To
accomplish this, a broadband library of cylindrical silicon meta-­
atoms, sitting on a double side–polished silicon substrate, was gen-
erated using rigorous coupled-­wave analysis (RCWA). The phase
and transmittance of the library is plotted in Fig. 2 (A and B). Cylin-
drical geometry was chosen to facilitate fabrication using scalable
mask-­based photolithography. Pillar height and pitch were chosen
to be 6.5 and 5 μm, respectively, based on parameter sweeps. Al-
though this library has smooth phase coverage across bandwidth, it
does not allow for prescribed control of dispersion; thus, it is not
possible to completely avoid chromatic aberration using this library,
especially for the broad 7.5-­to 13.5-­μm LWIR imaging band and
24-­mm aperture studied here, making forward design approaches
difficult.
To optimize the broadband PSF within the constraints of the
meta-­atom library, the inverse design algorithm used gradient de-
scent to set the dimensions of the pillars in each unit cell for mini-
mal error in synthesized edge images, as shown in the schematic in
Fig. 2C. A metasurface was initialized with randomly radially
varying meta-­atom geometry across its area; initializing with radi-
ally varying parameters instead of fully random produced results
with more symmetric PSFs, without constraining the metasurface
to be fully axisymmetric. The incoherent PSF of the optic system in
both positive and negative configurations was calculated using the
angular spectrum method at evenly spaced wavelengths across the
design bandwidth. The positive and negative PSFs were scaled to
unity total intensity before evaluating their difference to calculate
the effective bipolar PSF. This effective PSF was convolved with a
Siemens star used for training. This simulated image was com-
pared with a target edge image defined by filtering the training
image with an exact LoG kernel; by comparing the simulated and
target images, as opposed to directly comparing the simulated and
target PSF, the solver finds designs that produce Laplacian edge
images without rigidly constraining the designed system PSF to
the LoG function. Gradient descent was performed with respect to
the loss function defined in Eq. 1
�
∑
2 ∣O∣ �
L= (Ii − Itarget )2 + α ∑ +β ∣ ∇D ∣2 (1)
∣ Ii ∣
2 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 2. Metasurface design. (A) Phase and (B) transmittance maps of polarization-­insensitive meta-­atom library. (C) Process diagram for inverse design method. (D and
E) Simulated broadband combined and component PSFs for the (D) inversely designed and (E) spiral phase metasurfaces.
The loss function includes three terms: a sum of squared errors
between the simulated and target edge images, an inverse of the rela-
tive signal strength in the edge image, and a penalty on the spatial
gradient of the meta-­atom width to smooth the metasurface phase
profile, which improves convergence speed and broadband perfor-
mance. Hyperparameters α and β set the relative weight of the terms.
Including a term that rewards high signal strength helps avoid un-
necessary overlap between component PSFs that leads to signal losses
during digital subtraction. When the two PSFs overlap, the resulting
signal in the difference image will cancel out, but the noise, which is
uncorrelated between images, will accumulate (26). For comparison,
a heuristic, forward design process was used to create a spiral phase
metasurface at a central operation wavelength of 10 μm.
The advantages of the inverse design process can be observed by
comparing the simulated, broadband effective PSFs for the inversely
designed metasurface and the spiral phase metasurface shown in
Fig. 2 (D and E, respectively). The broadband PSF for the spiral phase
metasurface is no longer symmetric because of dispersion, whereas
the PSF for the optimized metasurface has sufficient symmetry for
isotropic edge detection. The optimized metasurface is also better at
reducing overlap between component PSFs; the signal loss due to
component PSF overlap is 51% lower using the optimized design
compared to the spiral phase. In addition, the optimized bipolar PSF
has a region with near-­zero intensity separating the positive and nega-
tive regions. This PSF shape increases the system response to high
spatial frequencies while preserving the desired Laplacian filtering ef-
fect at low frequencies, resulting in more signal at sharp edges com-
pared to an exact LoG (see fig. S1). There is a trade-­off between the
resolution and the SNR of the designed devices. The SNR is improved
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
by increasing the separation between the positive and negative PSFs,
which means that better SNR solutions have larger effective focal
spots and therefore lower resolution. The desired balance between
SNR and resolution for a particular application can be achieved by
adjusting the weight α of the SNR loss term as well as the variance of
the gaussian used to generate the target edge image.
Experimental characterization
Both the inversely design and spiral phase metasurfaces were fabri-
cated (Fig. 3, A and B) and characterized with a combination of
laboratory and real-­world imaging. The broadband transmission ef-
ficiency of each metasurface was measured, and transmission losses
were compensated for during edge image synthesis. In practice, this
correction factor could be incorporated into the camera’s nonuni-
formity correction and therefore would not increase the number of
digital operations needed to form the edge image. The measured
transmission efficiency of the inversely designed metasurface was
38%, which includes losses from the metasurface reflection, back-
side reflection, and substrate absorption, which were estimated to be
43.3, 30.3, and 2.6%, respectively. The transmission efficiency of the
spiral phase metasurface was 44%. Maximizing the transmittance of
the metasurfaces was not the primary goal of this study, and it could
be increased by using a lower-­index substrate, adding an antireflec-
tive coating on the back side, or embedding the meta-­atoms in a
low-­index coating to decrease the impact of the impedance mis-
match between the substrate and air. The overall performance of
spatial filters, such as this, is better quantified by the maximum spa-
tial frequency for which they still retain the desired modulation
transfer function (MTF), and the magnitude of their MTF at that
3 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 3. Comparison of results between inverse design metasurface (top row) and spiral phase metasurface (bottom row). (A and B) Images of fabricated metasur-
faces. Inset: SEM images of region on metasurface. Scale bars, 10 μm. (C and D) Experimental edge images of blackbody spoke target at 315 K. Edge images of 0.104 cy/
mrad bar target at (E and F) 2 K and (G and H) 10 K temperature differential.
spatial frequency. The calculated effective MTF of the inversely de-
signed two-­aperture system exhibits the desired parabolic behavior
(24) up to a maximum at 0.272 cy/mrad, which could be considered
the spatial bandwidth for differentiation. The magnitude of the MTF
of our system at that frequency is 0.854 and can be compared to the
exact LoG that has a peak value of 1.0 (see fig. S1). The high MTF of
the inversely designed metasurface is achieved because of the mini-
mal overlap between the broadband component PSFs.
The experimental signal strength and resolution of the edge detec-
tion imaging system were characterized by imaging a blackbody Sie-
mens star with a temperature of 315 K. Figure 3 (A and B) shows
resulting edge images formed using the spiral phase and optimized
metasurfaces, respectively; this pair of images uses the same color
scale to demonstrate the increased signal strength in the optimized
metasurface. The peak edge signal using the inversely designed meta-
surface was on average 67% higher across all edge orientations and
spatial frequencies, which means that the optimized metasurface de-
livers a higher SNR and thus can be used to detect edges in lower
contrast scenes. The edge detection for the optimized metasurface is
also invariant to edge orientation, unlike the spiral phase, which de-
tects some edge orientations more strongly due to the asymmetry of
the spiral phase PSF under broadband illumination.
The limit on the minimum contrast needed for edge detection was
explored by imaging blackbody bar targets with low temperature dif-
ferentials, as shown in Fig. 3 (E to H). The inverse design metasurface
begins to detect edges for a temperature differential as low as 1 K (see
fig. S4). The strength and accuracy of the detected edges quickly im-
proves with temperature. At a 2-­K differential, some edges are visible
for both the spiral phase and optimized metasurface images. Howev-
er, the edges in the optimized metasurface image are brighter and
more reliably detected on each bar compared to the spiral phase meta-
surface image, which is missing edges in large regions on the image.
At a temperature difference of 10 K, both metasurfaces can reliably
detect edges. However, the spiral phase image still exhibits noticeable
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
noise. On the other hand, the optimized metasurface provides a
stronger edge signal, which effectively overpowers the noise.
To achieve low contrast edge detection, the inverse design priori-
tized SNR over resolution, which imposed a limit on the maximum
spatial frequency for identifying edges distinctly. The maximum spa-
tial frequency for distinct edge detection was determined from the
spoke target images in Fig. 3 and defined as the frequency where the
peak edge signal has at least 50% contrast with the trough between
edges. The maximum spatial frequency with at least 50% edge con-
trast is 0.076 cy/mrad for the spiral phase metasurface and 0.053 cy/
mrad for the optimized metasurface. This is a conservative criterion,
as evidenced by the 0.1 cy/mrad bar targets in Fig. 3 (G and H), which,
despite being above the cutoff, still have clearly visible edges, albeit
with some noticeable background signal.
The edge detector acts as a high-­pass filter; features much smaller
than the above limit are entirely above the cutoff frequency. Instead of
being outlined, the entire feature is highlighted in the edge image,
with enhanced contrast because the uniform background signal is re-
moved. The smallest feature that can be detected in this manner is
only limited by the resolution of the original imaging system before
modification of the metasurface. This is demonstrated in the real-­
word imagery in Fig. 4, which shows a conventional LWIR image and
its corresponding edge image of a distant tower with many small win-
dows and architectural features. The tower is very prominently out-
lined, while the small windows and features are highlighted. Figure 4E
is a blown-­up view of the edge image, in which you can see the indi-
vidual grilles on the windows and the patterns in the stonework, de-
spite the small size and low contrast of these features in the unaltered
LWIR image. If an application requires specific edge locations for
these small features, the zero crossings in the image can be found as
shown in Fig. 4 (C and F), as is typical for Laplacian-­based edge detec-
tion (38). In such applications, the optical system presented here re-
places the digital LoG convolution, decreasing computation time.
Edge images for a variety of other real-­world scenes are depicted in
4 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Demonstration of inverse design metasurface for edge encoding of real-­world scenes. (A) LWIR image of tower. (B) Experimental edge image corresponding
to (A). (C) Edge locations determined from zero-­crossing in (B) with brightness scaled by edge strength. (D to F) Zoomed-­in view of (A) to (C) for region marked in (A). Edge
image of (G) car near building entrance, (H) stadium seats, and (I) overlapping buildings.

Fig. 4 (H to J), which demonstrate the edge detector performance for
a wide span of feature sizes and contrasts; additional full-­size images
are available in the Supplementary Materials.
Single-­aperture design using
polarization-­sensitive metasurface
The ability for metasurfaces to multiplex according to polarization
potentiates the implementation of single-­aperture optoelectronic fea-
ture detection, in contrast with the two-­aperture approach described
in the previous sections, which would require careful alignment of
two temporally synchronous cameras to capture the positive and neg-
ative component images while accounting for parallax errors. The
single-­aperture approach uses a refractive lens for broadband focus-
ing as well as a Si (40, 41) refringent metasurface (40, 41) that inde-
pendently shapes the PSF for light with two orthogonal polarization
states into the positive and negative component PSFs. Using a polar-
imetry enabled camera, separate images would be simultaneously col-
lected for the two polarizations yielding the synthesized edge image
through digital subtraction.
To realize single-­aperture edge imaging, the inverse design of a
polarization-­sensitive metasurface was performed using a similar
method as elaborated earlier. The library of rectangular birefringent
meta-­atoms shown in Fig. 5 (A and B) was used. The design of the
polarization-­sensitive metasurface involved optimizing the x-­ and y-­
axis widths of each meta-­atom to independently control the phase
profiles for x-­ and y-­polarized light. Because this model had twice as
many meta-­atom parameters to optimize and required simultaneous
optimization of the PSFs for both polarization states, its memory re-
quirement was much larger. As a consequence, it was not possible to
simulate the PSF at fine wavelength samples across the entire band-
width in each iteration. To make this work with the available comput-
ing resources, in each iteration, the PSF was calculated for a small,
randomly selected batch of wavelengths, instead of evenly sampling
across the bandwidth. To improve the robustness of the solution,
the location of the point source used to calculate the PSF was also
dithered. Last, the adaptive loss weighting method by Heydari and
Mehmood (42) was implemented to retain fast convergence despite
the additional randomness due to mini-­batching.
The optimized phase profiles in each polarization state are shown
in Fig. 5 (C and D). The PSF for x-­polarized light is tightly focused, as
apparent the corresponding bar target image in Fig. 5E, while the PSF
for y-­polarized light is shaped into an annular ring, resulting in the
slightly blurred image in Fig. 5F. Digital subtraction of the images for
the two polarization states results in the edge image in Fig. 5H, which

Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024 5 of 8

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 5. Results for polarization multiplexing metasurface. (A) Phase and
(B) transmittance of polarization-­sensitive meta-­atom library at the center wave-
length (10.5 μm). Inset: SEM images of region on metasurface. Scale bar, 2 μm. (C and
D) Phase profiles of inversely designed metasurface at center wavelength for x-­ and
y-­polarized light; insets show PSFs for corresponding polarization. (E and F) Images
of 0.1 cy/mrad bar target at 350 K for x and y polarization. (G) Effective broadband,
bipolar PSF. (H) Edge image synthesized from subtraction of (F) and (G).
is consistent with the simulated effective bipolar PSF in Fig. 5G. The
increased noise is primarily due to ambient emissions reflected into
the camera from a polarizer, which would not be necessary when
implementing this approach with a polarimetric camera. The calcu-
lated effective MTF of the one aperture design, normalized for reflec-
tion losses, has a maximum of 1.0 at a frequency of 0.258 cy/mrad (see
fig. S1). This higher magnitude is attributed to the extra design
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
degrees of freedom afforded by the birefringent meta-­atom library, as
well as the improvements to the inverse design algorithm added for
the single-­aperture design (wavelength mini-­batching, point source
dithering, and adaptive loss weighting). This demonstration illustrates
that the presented approach can be applied to design an effective
single-­aperture optoelectronic edge encoder using polarization multi-
plexing. When used with a polarimetric camera, this approach facili-
tates live edge detection in a lightweight system with no need for
multiple aperture alignment.
DISCUSSION
In conclusion, we have presented an electro-­optical system for broad-
band, high-­SNR, incoherent edge detection with operation across the
7.5-­to 13.5-­μm LWIR imaging band. This performance was enabled
by a hybrid optical system comprising a refractive optic and a meta-
surface combined with multi-­objective inverse design. These advanc-
es enable imaging of real-­world scenes using off-­the-­shelf hardware,
and the approach was extended to a single-­aperture approach by us-
ing a birefringent metasurface in combination with polarization dis-
crimination at the imaging sensor. While the presented systems are
designed to apply a Laplacian filter, the approach can be applied to
arbitrary differentiation and convolution operations.
This proof-­of-­concept demonstration of high-­SNR, broadband
optical feature detection could be adapted to achieve practical op-
tical front-­end preprocessing for computer vision systems. Off-­
loading low-­level feature detection onto an imaging system has the
potential to improve the performance of common computer vision
systems by reducing both power consumption and latency. Typical
digital convolutional layers consist of several parallel convolution-
al channels, which could be achieved optically using a division of
aperture mini-­lens array (43), combined with a birefringent meta-
surface array used to shape the PSF of each mini-­lens. Whereas the
computation time for digital convolution scales with the size of the
convolution kernel, optical convolution speed is invariant to ker-
nel size and could enable substantial acceleration when used with
shallow networks enabled by large convolution kernels (44). This
could be particularly impactful for autonomous and remote sens-
ing systems that face size, weight, and power consumption con-
straints but must also perform on-­board processing due to latency
or communication bandwidth requirements.
MATERIALS AND METHODS
Simulation
The meta-­atom libraries were generated using the RCWA software
Reticolo Software for Grating Analysis (45). The inverse design algo-
rithm leveraged the Pytorch machine learning package, which en-
abled GPU computing and automatic gradient calculation. Forward
propagation steps were calculated using the angular spectrum meth-
od, using a memory-­efficient two-­dimensional fast Fourier transform
algorithm to enable simulation of larger devices. Inverse design calcu-
lations were performed on an Nvidia Quadro M6000 24 Gb GPU. The
image used for training was a 60-­spoke Siemens star target with a
resolution of 600 by 600 pixels (7.2 mm by 7.2 mm on the focal plane
array). The target edge images were formed by convolving the Sie-
mens star image with a LoG kernel. The variance of the LoG is a tun-
able hyperparameter that influences the resolution of the design; the
presented designs were generated using a variance of 0.82 mrad.
6 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Metasurface fabrication
The metasurfaces were composed of etched pillars on undoped, dou-
ble side–polished Si substrates with 280-­μm thickness. Negative pho-
tolithography masks containing the inversely designed geometric
features were ordered from Digidat (Pasadena, CA) and subsequently
used to pattern a layer of s1805 positive photoresist on the Si substrate.
The pattern was transferred to a 125-­nm Al2O3 hard mask by electron
beam deposition followed by liftoff using Microposit Remover 1165.
The hard mask pattern was transferred to the pillars by selective reac-
tive ion etching of the exposed Si to a depth corresponding to the
designed height of the pillars (6.5 μm for inversely designed metasur-
faces, 7.0 μm for the spiral phase metasurface). The thin remaining
layer of Al2O3 on the tops of the pillars was not removed, which may
have introduced a small absorptive loss (<10%) at the higher wave-
lengths but otherwise did not substantially alter the metasurface
phase because of the low thickness and refractive index of Al2O3.
Experimental setup
Laboratory and field imagery were collected on an FLIR Boson 640
camera with a 24-­mm F/1 lens with 18° horizontal field of view. Most
of the built-­in preprocessing for the Boson was disabled using the
FLIR Boson app v3.0, with the exception of the nonuniformity cor-
rection and defect pixel replacement. The Boson was aligned with a
cage system to facilitate quick and consistent insertion and removal of
the metasurface between complementary frames. The position of the
metasurface on the cage was controlled such that the metasurface
would be within 1 mm of the Boson lens mount during imaging, with
the metasurface facing toward the camera. Laboratory imaging of bar
targets was performed on a 60-­inch focal length optical collimator
bench (Santa Barbara Infrared Inc.; Santa Barbara, CA) with a black-­
body source projected through various target plates.
The transmission efficiency of each metasurface was measured by
imaging the blackbody source at temperatures from 300 and 370 K in
increments of 10 K with and without the metasurface installed; the
efficiency was calculated via regression as the ratio of the changes in
signal with and without the metasurface installed. These transmission
calculations included pixels in uniform regions within a 1.4° square
field of view at the center of the image.
Images and further details regarding the mounting apparatus used
for all imaging tests are included in fig. S2. The signal strength and
resolution of the designed systems was characterized in a Siemens star
imaging test, which is elaborated in fig. S3. The minimum tempera-
ture differential characterization was performed using a differential
black body, which is shown in fig. S4.
Supplementary Materials
This PDF file includes:
Figs. S1 to S12
References
REFERENCES AND NOTES
1. Y. Lecun, Y. Bengio, G. Hinton, Deep learning. Nature 521, 436–444 (2015).
2. K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, Z. Yang,
Y. Zhang, D. Tao, A survey on vision transformer. IEEE Trans. Pattern Anal. Mach. Intell. 45,
87–110 (2023).
3. D. Ghimire, D. Kil, S. Kim, A survey on efficient convolutional neural networks and
hardware acceleration. Electronics 11, 945 (2022).
4. M. K. Chen, X. Liu, Y. Sun, D. P. Tsai, Artificial intelligence in meta-­optics. Chem. Rev. 122,
15356–15413 (2022).
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
5. G. Habib, S. Qureshi, Optimization and acceleration of convolutional neural networks: A
survey. J. King Saud Univ. Comput. Inf. 34, 4244–4268 (2022).
6. W. T. Rhodes, Incoherent spatial filtering. Opt. Eng. 19, 193323 (1980).
7. J. Zhou, H. Qian, J. Zhao, M. Tang, Q. Wu, M. Lei, H. Luo, S. Wen, S. Chen, Z. Liu,
Two-­dimensional optical spatial differentiation and high-­contrast imaging. Natl. Sci. Rev.
8, nwaa176 (2020).
8. X. Qiu, F. Li, W. Zhang, Z. Zhu, L. Chen, Spiral phase contrast imaging in nonlinear optics:
Seeing phase objects using invisible illumination. Optica 5, 208–212 (2018).
9. S. Fürhapter, A. Jesacher, S. Bernet, M. Ritsch-­Marte, Spiral phase contrast imaging in
microscopy. Opt. Express 13, 689–694 (2005).
10. W. Fu, D. Zhao, Z. Li, S. Liu, C. Tian, K. Huang, Ultracompact meta-­imagers for arbitrary
all-­optical convolution. Light Sci. Appl. 11, 62 (2022).
11. H. Kwon, D. Sounas, A. Cordaro, A. Polman, A. Alù, Nonlocal metasurfaces for optical
signal processing. Phys. Rev. Lett. 121, 173004 (2018).
12. S. Abdollahramezani, O. Hemmatyar, A. Adibi, Meta-­optics for spatial optical analog
computing. Nanophotonics 9, 4075–4095 (2020).
13. L. Wan, D. Pan, T. Feng, W. Liu, A. A. Potapov, A review of dielectric optical metasurfaces
for spatial differentiation and edge detection. Front. Optoelectron. 14, 187–200 (2021).
14. A. Komar, R. A. Aoni, L. Xu, M. Rahmani, A. E. Miroshnichenko, D. N. Neshev, Edge
detection with mie-­resonant dielectric metasurfaces. ACS Photonics 8, 864–871 (2021).
15. J. Zhou, J. Zhao, Q. Wu, C. Chen, M. Lei, G. Chen, F. Tian, Z. Liu, Nonlinear computational
edge detection metalens. Adv. Funct. Mater. 32, 2204734 (2022).
16. Z. Wang, G. Hu, X. Wang, X. Ding, K. Zhang, H. Li, S. N. Burokur, Q. Wu, J. Liu, J. Tan,
C.-­W. Qiu, Single-­layer spatial analog meta-­processor for imaging processing. Nat.
Commun. 13, 2188 (2022).
17. C. Guo, M. Xiao, M. Minkov, Y. U. Shi, S. Fan, Photonic crystal slab Laplace operator for
image differentiation. Optica 5, 251–256 (2018).
18. Y. Zhou, H. Zheng, I. I. Kravchenko, J. Valentine, Flat optics for image differentiation. Nat.
Photonics 14, 316–323 (2020).
19. W. Xue, O. D. Miller, High-­NA optical edge detection via optimized multilayer films. J. Opt.
23, 125004 (2021).
20. C. Jin, Y. Yang, Transmissive nonlocal multilayer thin film optical filter for image
differentiation. Nanophotonics 10, 3519–3525 (2021).
21. T. Zhu, C. Guo, J. Huang, H. Wang, M. Orenstein, Z. Ruan, S. Fan, Topological optical
differentiator. Nat. Commun. 12, 680 (2021).
22. R. Maini, H. Aggarwal, Study and comparison of various image edge detection
techniques. Int. J. Image Process 3, 1–11 (2009).
23. H. Wang, C. Guo, Z. Zhao, S. Fan, Compact incoherent image differentiation with
nanophotonic structures. ACS Photonics 7, 338–343 (2020).
24. X. Zhang, B. Bai, H. Sun, G. Jin, J. Valentine, Incoherent optoelectronic differentiation
based on optimized multilayer films. Laser Photon Rev. 16, 2200038 (2022).
25. D. Hazineh, S. W. D. Lim, Q. Guo, F. Capasso, T. Zickler, Polarization multi-­image synthesis
with birefringent metasurfaces, in 2023 IEEE International Conference on Computational
Photography (IEEE, 2023).
26. A. W. Lohmann, W. T. Rhodes, Two-­pupil synthesis of optical transfer functions. Appl.
Optics 17, 1141–1151 (1978).
27. J. N. Mait, W. T. Rhodes, Pupil function design algorithm for bipolar incoherent spatial
filtering. Appl. Optics 28, 1474–1488 (1989).
28. A. Mukhtar, L. Xia, T. B. Tang, Vehicle detection techniques for collision avoidance systems:
A review. IEEE Trans. Intell. Transp. Syst. 16, 2318–2338 (2015).
29. X. Dai, X. Yuan, X. Wei, TIRNet: Object detection in thermal infrared images for
autonomous driving. Appl. Intell. 51, 1244–1261 (2021).
30. M. Teutsch, T. Mueller, M. Huber, J. Beyerer, Low resolution person detection with a
moving thermal infrared camera by hot spot classification, in Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern Recognition Workshops (IEEE,
2014), pp. 209–216.
31. A. Arbabi, A. Faraon, Advances in optical metalenses. Nat. Photonics 17, 16–25 (2023).
32. M. K. Chen, Y. Wu, L. Feng, Q. Fan, M. Lu, T. Xu, D. P. Tsai, Principles, functions, and
applications of optical meta-­lens. Adv. Opt. Mater. 9, (2021).
33. Y. Zhou, I. I. Kravchenko, H. Wang, J. R. Nolen, G. Gu, J. Valentine, Multilayer noninteracting
dielectric metasurfaces for multiwavelength metaoptics. Nano Lett. 18, 7529–7537 (2018).
34. A. S. Backer, Computational inverse design for cascaded systems of metasurface optics.
Opt. Express 27, 30308–30331 (2019).
35. H. Chung, O. D. Miller, High-­NA achromatic metalenses by inverse design. Opt. Express 28,
6945–6965 (2020).
36. E. Bayati, R. Pestourie, S. Colburn, Z. Lin, S. G. Johnson, A. Majumdar, Inverse designed
extended depth of focus meta-­optics for broadband imaging in the visible.
Nanophotonics 11, 2531–2540 (2022).
37. W. T. Chen, A. Y. Zhu, J. Sisler, Y. W. Huang, K. M. A. Yousef, E. Lee, C. W. Qiu, F. Capasso,
Broadband achromatic metasurface-­refractive optics. Nano Lett. 18, 7801–7808 (2018).
38. D. Marr, E. Hildreth, Theory of edge detection. Proc. R. Soc. Lond. B Biol. Sci. 207, 187–217
(1980).
7 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
39. E. Hecht, Optics (Pearson, ed. 5, 2016).
40. A. Arbabi, Y. Horie, M. Bagheri, A. Faraon, Dielectric metasurfaces for complete control of
phase and polarization with subwavelength spatial resolution and high transmission.
Nat. Nanotechnol. 10, 937–943 (2015).
41. X. Wang, H. Wang, J. Wang, X. Liu, H. Hao, Y. S. Tan, Y. Zhang, H. Zhang, X. Ding, W. Zhao,
Y. Wang, Z. Lu, J. Liu, J. K. W. Yang, J. Tan, H. Li, C.-­W. Qiu, G. Hu, X. Ding, Single-­shot
isotropic differential interference contrast microscopy. Nat. Commun. 14, 2063 (2023).
42. A. A. Heydari, A. Mehmood, SRVAE: Super resolution using variational autoencoders, in
Pattern Recognition and Tracking XXXI, M. S. Alam, Ed. (SPIE, 2020), p. 30.
43. J. L. Pezzaniti, D. B. Chenault, A division of aperture MWIR imaging polarimeter, in
Polarization Science and Remote Sensing II, J. A. Shaw, J. S. Tyo, Eds. (SPIE, 2005), vol. 5888,
p. 58880V.
44. X. Ding, X. Zhang, J. Han, G. Ding, Scaling up your kernels to 31×31: Revisiting large
kernel design in CNNs, in 2022 IEEE/CVF Conference on Computer Vision and Pattern
Recognition (CVPR) (IEEE, 2022), pp. 11953–11965.
45. J. P. Hugonin, P. Lalanne, RETICOLO software for grating analysis. arXiv:2101.00901
[physics.optics] (2021).
Swartz et al., Sci. Adv. 10, eadk0024 (2024) 7 February 2024
46. A. Trujillo-­Pino, K. Krissian, M. Alemán-­Flores, D. Santana-­Cedrés, Accurate subpixel edge
location based on partial area effect. Image Vis. Comput. 31, 72–90 (2013).
Acknowledgments
Funding: This work was funded by the Office of Naval Research under grant N00014-­21-­
1-­2468 (B.T.S., H.Z., and J.V.) and supported by the Naval Innovative Science and Engineering
(NISE) program (G.T.F.). The approved document control number is 0543-­688-­23. Author
contributions: Conceptualization: J.V. and G.T.F. Methodology: B.T.S. and H.Z. Software: B.T.S.
and H.Z. Investigation: B.T.S. and G.T.F. Writing—original draft: B.T.S. Writing—review and
editing: J.V. and G.T.F. Competing interests: The authors declare that they have no competing
interests. Data and materials availability: All data needed to evaluate the conclusions in the
paper are present in the paper and/or the Supplementary Materials.
Submitted 27 July 2023
Accepted 9 January 2024
Published 7 February 2024
10.1126/sciadv.adk0024
8 of 8
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

C O N D E N S E D M AT T E R P H Y S I C S Copyright © 2024 The

Authors, some rights
Fully gapped pairing state in spin-­triplet reserved; exclusive
licensee American
superconductor UTe2 Association for the
Advancement of
Science. No claim to
Shota Suetsugu1*, Masaki Shimomura1, Masashi Kamimura1, Tomoya Asaba1, Hiroto Asaeda1, original U.S.
Yuki Kosuge1, Yuki Sekino1, Shun Ikemori1, Yuichi Kasahara1, Yuhki Kohsaka1, Minhyea Lee2, Government Works.
Youichi Yanase1, Hironori Sakai3, Petr Opletal3, Yoshifumi Tokiwa3, Yoshinori Haga3, Distributed under a
Yuji Matsuda1* Creative Commons
Attribution License 4.0
The recently discovered superconductor UTe2 is a promising candidate for spin-­triplet superconductors, but the (CC BY).
symmetry of the superconducting order parameter remains highly controversial. Here, we determine the super-
conducting gap structure by the thermal conductivity of ultraclean UTe2 single crystals. We find that the a-­axis
thermal conductivity divided by temperature κ/T in zero-­temperature limit is vanishingly small for both magnetic
field H‖a and H‖c axes up to H/Hc2 ∼ 0.2, demonstrating the absence of nodes around the a axis contrary to the previous
belief. The present results, combined with the reduction of nuclear magnetic resonance Knight shift, indicate that
the superconducting order parameter belongs to the isotropic Au representation with a fully gapped pairing state,
analogous to the B phase of superfluid 3He. These findings reveal that UTe2 is likely to be a long-­sought three-­
dimensional strong topological superconductor, hosting helical Majorana surface states on any crystal plane.

INTRODUCTION was supported by scanning tunneling spectroscopy experiments (17).

Spin-­triplet pairing states have aroused tremendous interest because
of the emergence of Majorana quasiparticles (1) and their potential
application to fault-­tolerant quantum information processing (2, 3).
The most famous example is the superfluid 3He (4), and the quest for
its superconducting analog has been a long-­standing issue in condensed
matter physics. The promising candidate is the recently discovered
heavy fermion superconductor UTe2 (5). UTe2 exhibits highly unusual
superconducting properties, including extremely large upper critical
fields well exceeding the Pauli limit (5, 6), coexistence of super
conductivity and ferromagnetic order in high magnetic field (7, 8),
reentrant superconductivity that resembles to ferromagnetic super-
conductors URhGe and UCoGe (7, 9), and peculiar behavior of
nuclear magnetic resonance (NMR) Knight shift and absence of
coherence peak in the superconducting state (5, 10–12). All of these
notable properties are indicative of the spin-­triplet pairing state.
Moreover, this unconventional superconducting state occurs in the
paramagnetic state at ambient pressure. This is in contrast to the other
ferromagnetic superconductors; at ambient pressure, UGe2 shows a
ferromagnetic order (13), and, in URhGe and UCoGe, supercon-
ductivity coexists with ferromagnetic order (14, 15). Therefore, UTe2 is
a long-­sought material that allows us to examine superconducting
properties in more detail using various probes.
Of particular interest is the superconducting gap structure of UTe2,
which is of primary importance for understanding the peculiar super-
conducting state associated with the spin-­triplet pairing state. Despite
the intensive research efforts, however, the superconducting order
parameter has been highly controversial, and its determination is
still challenging. At an early stage, a chiral superconducting state
with a multicomponent order parameter has been suggested by the
double peak of the specific heat and polar Kerr effect (16), which
1
Department of Physics, Kyoto University, Kyoto 606-­8502, Japan. 2Department of
Physics, University of Colorado Boulder, Boulder, CO 80309, USA. 3Advanced Science
Research Center, Japan Atomic Energy Agency, Tokai, Ibaraki 319-­1195, Japan.
*Corresponding author. Email: suetsugu.​shota.​4n@​kyoto-­u.​ac.​jp (S.S.); matsuda@​
scphys.​kyoto-­u.​ac.​jp (Y.M.)
However, as the sample quality improves, a single peak is observed
in the specific heat (18–20) and does not split into two peaks even
under uniaxial strain (21). Moreover, recent Kerr effect experiments
on clean crystals report no evidence of broken time-­reversal symmetry
(22). Considering the orthorhombic crystal structure of UTe2 (Fig. 1A),
although the chiral state may appear at the surface, chiral supercon-
ductivity is unlikely to be realized in the bulk.
Given no solid evidence for the bulk chiral superconductivity in
UTe2, there are four possible gap symmetries for spin-­triplet super-
conductivity, {Au, B1u, B2u, B3u} (23). The Au symmetry is fully gapped
(Fig. 1B), corresponding to the B phase of superfluid 3He. On the
other hand, B1u, B2u, and B3u symmetries have point nodes at isolated
points on the Fermi surface along the c, b, and a axes, respectively
(Fig. 1C). Several measurements, such as thermal conductivity (24),
penetration depth (25, 26), and specific heat (27), have reported the
presence of the low-­energy quasiparticle excitations, suggesting the
presence of point nodes rather than line nodes. NMR measurements
show that, below the superconducting transition temperature Tc, the
Knight shift along the b and c axes decreases (11, 12). This excludes
the B1u and B2u symmetries. On the basis of these results, it has been
widely discussed that the B3u symmetry with point nodes along the
a axis accounts for the superconducting gap of UTe2. The point node(s)
around the a axis has also been suggested by thermal conductivity (24).
However, it should be stressed that, because the low-­energy qua-
siparticle excitations in spin-­triplet superconductors are extremely
sensitive to impurities (28), the determination of the gap symmetry
in the previous measurements may be hindered by impurity. More-
over, as mentioned previously, the gap structure near the surface may
be distinct from the bulk. Therefore, it is premature to conclude the
presence of point nodes along the a axis. Determining whether the
gap symmetry is Au or B3u is of crucial importance because the topo-
logical properties of these two symmetries are fundamentally different
(29). Thus, the situation calls for reexamining the gap structure using
a bulk probe on crystals with an extremely low impurity concentration.
Here, we determined the superconducting gap structure of ultra-
clean UTe2 crystals (30, 31) by thermal conductivity κ, which is a

Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024 1 of 6

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
A D 500
#2
15
ρ (microhm·cm)
ρ (microhm·cm)
400
10 RRR = 630
300
5
200 0 present results indicate ξa ≈ ξc. This suggests that the average Fermi
U
100
Te 0
B 0 100
T (K)
E 20
#1
15 H//c
µ0Hc2 (T)
Fermi surface
10
C 0
5 T (K)
a H//a
0 K2 mol. The Kadowaki-­Woods ratio, A/γ2 ≈ 2 × 10−5 microhm·cm
0 1
Fig. 1. Crystal structure, resistivity, and phase diagram of UTe2. (A) Crystal
structure of UTe2. The gray and green spheres represent U and Te atoms, respec-
tively. (B and C) Structure of the superconducting gap Δ for Au (B) and B3u (C) sym-
metries for spherical Fermi surface (blue sphere). The B3u state has point nodes (red
points) along the a axis. (D) Temperature dependence of resistivity ρ for sample #2.
The inset shows ρ as a function of T2 at low temperatures. The residual resistivity ρ0
is obtained by a fit to ρ(T) with ρ0 + AT2 (dashed line). (E) H-­T phase diagram deter-
mined by resistivity measurements of sample #1. The linear extrapolations to T = 0
yield the upper critical fields μ0Hc2 of ∼12 T for H‖a and ∼17 T for H‖c. The inset
shows an enlarged view near Tc.
bulk and directional probe of the superconducting gap structure
(32). In contrary to the previous reports, the present results demon-
strate the absence of any type of nodes at and around the a axis. This
indicates that the superconducting order parameter of UTe2 belongs
to the isotropic Au representation with a fully gapped pairing state,
analogous to the B phase of superfluid 3He.
RESULTS
The superconducting transition temperature Tc of both crystals #1 and
#2 is 2.1 K, which is higher than the previously reported value of typi-
cally ∼2.0 K (20). Figure 1D displays the temperature (T) dependence
of the resistivity ρ(T) along the a axis for sample #2. The normal state
resistivity is well fitted by ρ(T) = ρ0 + AT2 (inset of Fig. 1D). The qua-
dratic temperature dependence of ρ, a characteristic of Fermi liquid,
indicates the importance of the electron-­electron correlation. The
residual resistivity ratio (RRR ≡ ρ(300 K)/ρ0) of 260 for sample #1
and 630 for sample #2 is nearly 10 and 30 times larger than that
reported in crystal with RRR = 22 (24), respectively. Hereafter, the
samples with RRR = 22 (24), 260 (#1), and 630 (#2) will be referred
to as samples in moderately clean, very-­clean, and ultraclean regions,
respectively. It should be noted that clear quantum oscillations are
reported in the sample with RRR = 220 (30). These confirm the high
quality of the present crystals. By comparing these crystals with different
RRRs, we can obtain pivotal information about the quasiparticle exci-
tations that are intimately related to the superconducting gap function.
The upper critical fields determined by the resistive transition of
c
sample #1 for H‖a and H‖c, Hc2, and Hc2, respectively, are displayed
a
a c
in Fig. 1E. At zero temperature, Hc2 (0) and Hc2 (0) are approximately
12 and 17 T, respectively. The upper critical fields of sample #2 is
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
a c
close to those of #1. As shown in the inset, Hc2 is very close to Hc2 in
the vicinity of Tc. Because the initial slope at Tc is proportional to the
a,c
orbital limiting field, − (dHc2 ∕dT) ∣Tc ∝ 1 ∕ ξb ξc,a, where ξa, ξb, and
ξc are the coherence length along the a, b and c axes, respectively, the
0 10 20 30
2 2
velocity along the a axis is close to that along the c axis.
T (K ) The heat capacity C of sample #1 exhibits a very sharp transition
200 300 at Tc with no peak splitting (fig. S1). For comparison, we plot the
reported data for a very clean crystal with Tc = 2.1 K (30), whose
RRR = 220 is close to sample #1, and for a crystal with Tc = 1.7 K
(20). The temperature dependence of C/T for sample #1 is very close
µ0Hc2 (T)
2
to the data of the very clean crystal with a similar RRR that shows a
negligibly small residual C/T in the zero temperature limit. These
results further support the high quality of the present crystals. The
1.5 2.0
electronic heat capacity coefficient in the normal state γ is 120 mJ/
2 3
T (K)
(mol K/mJ)2, is close to that of typical correlated systems (33).
Figure 2 shows the T-­dependence of κ/T with applied thermal
current Q‖a for the very clean (#1) and ultraclean (#2) crystals, along
with κ/T for the moderately clean crystal (24). In the normal state,
both electron and phonon contribute to the thermal conductivity. To
evaluate the electron contribution in the normal state at very low
temperatures, we measure the normal state thermal conductivity
κN/T (open blue circles) by applying magnetic fields of μ0H = 12 T,
which exceeds Hc2 a
. The blue dashed line represents the electron contri-
bution yielded from the Wiedemann-­Franz (WF) law L0/ρN. Here,
L0 = 2.44 × 10−8 W·ohm·K−2 is the Lorenz number, and the normal
state resistivity ρN below Tc is calculated from the extrapolation above
Tc by assuming ρN(T) = ρ0 + AT2. As shown in Fig. 2, κN/T well
coincides with L0/ρN below 0.5 K, indicating that the WF law holds.
This demonstrates that, for the very clean and ultraclean crystals,
the electron contribution is dominant in the normal state at very low
temperatures. In zero field, κ/T exhibits a distinct kink upon entering
the superconducting state, increases steeply, and reaches a maximum
at ∼1.5 and ∼1.2 K for samples #1 and #2, respectively. As observed
in several strongly correlated systems (34–36), the enhancement of
κ/T below Tc is attributed to the rapid enhancement of the quasiparticle
mean free path, which is caused by the suppression of the electron-­
electron inelastic scattering rate due to the superconducting gap
#1
2.0
12 T(H//a)
1.5
κ/T (W/K m)
2
#2 #1
L 0 /ρN
1.0 0T #1
0T
(24)
0.5 0.05 T
L 0 /ρN (24)
0.0
0.0 1.0 2.0 3.0
T (K)
Fig. 2. Temperature dependence of thermal conductivity of UTe2. In zero field,
κ/T shows a kink at Tc and then increases up to ∼1.5 and ∼1.2 K for samples #1 and #2,
respectively. The blue dashed line represents the electronic contribution estimated
by L0/ρN for sample #1. At low temperatures, L0/ρN is close to the normal state value
of κ/T at 12 T for H‖a (blue open circles), indicating that κ is dominated by the
electronic contribution. For comparison, we plot κ/T at 0.05 T and L0/ρN for the
moderately clean crystal with RRR = 22 (24).
2 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

formation. The enhancement of κ/T of sample #2 is more substantial
than sample #1 because the mean free path is larger in samples with
larger RRR. We note that the enhancement of κ/T in the moderately
clean crystal is much smaller than in these crystals. Moreover, κ0N/T ≡
κN/T (T → 0) of sample #1 determined by the data taken at 12 T is
1.9 W/K2 m, one order of magnitude larger than 0.1 W/K2 m re-
ported for the moderately clean crystal (24). The difference of κ0N/T
also reflects the significantly enhanced mean free path in the very
clean and ultraclean crystals.
Figure 3A shows the T-­dependence of κ/T at low temperatures
for samples #1 and #2. The thermal conductivity in the supercon-
ducting state has quasiparticle and phonon contributions, κ = κqp +
κph. In the boundary-­limited scattering regime at low temperatures,
κph/T is proportional to ℓphT2, where ℓph is the phonon mean free
path. While ℓph is T-­independent for diffuse scattering limit, resulting
in κph/T ∝ T2, it is T−1-­dependent for specular reflection, resulting
in κph/T ∝ T. In real systems, κph/T ∝ Tp with 1 ≤ p ≤ 2. Because
κph/T becomes zero at T = 0, κ0/T ≡ κ/T (T → 0) contains only the
quasiparticle contribution. It is apparent the linear extrapolation of
κ/T to T = 0 yields a negative intercept. We find that κ/T for both
samples is best fitted by κ/T ∝ Tα with α = 1.48 below 0.3 K (inset of
Fig. 3A). These results indicate the absence of residual thermal con-
ductivity κ0/T ≈ 0. For unconventional superconductors with line
nodes in the energy gap, a finite residual term is expected because of
the existence of a residual normal fluid (37). Therefore, the present
results provide evidence for the absence of line nodes. This is further
supported by the estimation of the residual term in a line-­nodal super
conductor. For line nodes, the universal expression of κ0/T for unitary
L ℏ
scattering is given by μ λ0 2 πΔ , where μ0 is the permeability of vacuum
0 0
and λ is the penetration depth. Using 2Δ0 = 3.5kBTc and λ ≈ 1000
nm (24, 25), we obtain κ0/T ≈ 0.054 W/K2 m (gray dashed line).
Clearly, κ/T at the lowest temperature is less than the calculated
value for the line nodes.
We next examine the presence/absence of point nodes. It is
well-­known that the heat transport at low magnetic fields in nodal
superconductors is fundamentally different from that in full-­gap super-
conductors. In magnetic fields, the energy of quasiparticles with
momentum ℏk in a circulating supercurrent flow vs shifts as Ek →
Ek − ℏk · vs. As a result of this Doppler shift, the thermal conductivity
in nodal superconductors is dominated by the contribution from
delocalized √ quasiparticles, leading to an initial steep increase of
κ(H) ∕ T ∝ H for line nodes and κ(H)/T ∝ ∣H log H∣ for point
nodes (38). By contrast, in full-­gap type II superconductors in the
clean limit, all quasiparticles are bound to the vortex cores, and the
magnetic field has a negligible effect on the heat transport (39–41)
except in the vicinity of Hc2, where the vortex cores overlap.
Thermal conductivity is a directional probe sensitive to the quasi-
particles with momentum parallel to the thermal current (k · Q ≠ 0)
and perpendicular to the magnetic field (k × H ≠ 0) because H ⊥ vs.
To investigate the quasiparticle excitations around the a axis, we
measured κ along the a axis (Q‖a) for H‖c and H‖a (Fig. 3, B and C).
As seen in the insets, while the former geometry sensitively probes
the point node at the a axis, the latter geometry selectively probes
quasiparticle excitations from point nodes that are off a axis but have
momentum in the a-­axis direction. For the very clean crystal (#1),
κ/T collapses into a single curve and decreases almost linearly below
∼0.25 K for both H‖c and H‖a. For the ultraclean crystal (#2), κ/T
normalized by the normal state value κ0N/T for H‖c is nearly half of
that for the very clean crystal below ∼0.3 K (Fig. 3B). Simple linear
extrapolations to T = 0 give negative intercepts for all data, indicating
a very small residual κ/T at T = 0. To obtain the zero temperature
limit of κ/T quantitatively, the data are fitted using κ/T = κ0/T + ATα
with κ0/T ≥ 0 and 1 ≤ α ≤ 2. This yields the vanishingly small κ0/T at
low fields for both samples (figs. S2 to S4A).
Figure 4A shows κ0/T normalized by κ0N/T as a function of H/Hc2.
κ0/T of the very clean crystal (#1) is less than 1% of the normal-­state
value even at H ∼ 0.09 Hc2 c
for H‖c and at H ∼ 0.2 Hc2 a
for H‖a.
Moreover, as shown in the inset, (κ0/T)/(κ0N/T) of the ultraclean
crystal (#2) for H‖c is significantly suppressed from that of the very
clean crystal. For comparison, we plot data for two other U-­based su-
perconductors, UPt3 (42) and URu2Si2 (36), which are believed to
have both point and line nodes. In notable contrast to the present
results, κ0/T for UPt3 and URu2Si2 exhibits a steep increase at low
fields. As shown in Fig. 4A (see also fig. S4B), even at 0.13 K, κ/T of
c
the ultraclean crystal is only 2% of κ0N/T up to H ∼ 0.12 Hc2, still
much smaller than (κ0/T)/(κ0N/T) for UPt3 and URu2Si2. These results

A 0.6 B C 0.15
0.3 H//c 1T 0.4 T
vs 0.4 T 1.5 T
0.10
κ/T (W/K m)

0.2 2T Q//H//a
2

2T
(κ/T) /(κ0N/T)

(κ/T) /(κ0N/T)
/T (W/K m)

0.4 #1 1T 0.10
0.1 1.5 T vs
2

Q//a 1.1 T
#2 #1
0.0
0.00 0.10 0.05
0.2 2T 0.05
κ/T

1.48 1.48
T (K ) #1 1.5 T
#2
µ 0H = 0 T 0.5 T
0.0 0.00 0.00
0.0 0.1 0.2 0.3 0.0 0.1 0.2 0.3 0.0 0.1 0.2 0.3
T (K) T (K) T (K)

Fig. 3. Temperature dependence of thermal conductivity of UTe2 at low temperatures. (A) Thermal conductivity divided by temperature, κ/T, in zero field for the very
clean (#1) and ultraclean (#2) crystals. The data are extrapolated to T = 0 by κ/T = κ0/T + ATα with κ0/T ≥ 0 (black dashed and dotted lines), yielding κ0/T ≈ 0 and α = 1.48
for both samples. The inset displays κ/T and the fit as a function of T1.48. The gray dashed horizontal lines represent the universal constant for line nodes (see Results section).
(B and C) Temperature dependence of κ/T normalized by the normal state value in the zero temperature limit κN0/T for magnetic fields H parallel to the crystal c axis (B)
and a axis (C) with applied thermal current Q‖a. The normal state value κ0N/T of the very clean crystal for H‖a is determined by the data at 12 T (blue open circles in Fig. 2).
κ0N/T for other configurations is approximated by L0/ρ0. As illustrated in the inset, κ/T for H‖c (H‖a) selectively probes the quasiparticles in the red-­shaded area and is
sensitive to the nodes (blue circles) at (around) the a axis.

Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024 3 of 6

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
indicate that UTe2 is essentially different from other U-­based un-
conventional superconductors in that only a very few delocalized
quasiparticles are excited, even well inside the vortex state. In
Fig. 4B, we compare κ0/T of UTe2 with typical s-­wave superconductor
Nb (39) and nodal superconductor UPt3 (42) over a wide field range.
For Nb, we show the data measured in both ascending and descending
magnetic fields, although the difference is small. Here, we restrict
the argument of H-­dependence for H‖c to the low field regime,
where the normal-­state magnetoresistance is small and κ0N/T can be
approximated by L0/ρN (fig. S5). It is evident that H-­dependence of
UTe2 is very close to that of Nb. These results lead us to conclude
that quasiparticle excitations with the velocity component along the
a axis are negligibly small, ruling out the presence of point nodes at
and around the a axis. This indicates that κ in zero field is dominated
by the phonon contribution κph at low temperatures, consistent with
the power law behavior κ/T ∝ T1.48 (see inset of Fig. 3A).
The field dependence of κ/T at 0.22 K for H‖c and H‖a (Fig. 4C),
which are both nearly constant at low fields, further supports the
absence of point nodes around the a axis. It has been reported that
the thermal conductivity of the d-­wave superconductor CeCoIn5 shows
little H-­dependence at √
very low temperatures (35, 43). One possible
explanation is that the H dependence of the quasiparticle density
of states is canceled out by the reduction of the quasiparticle mean
√ structure of this material.
free path, which is proportional to intervortex distance ∝ 1 ∕ H
(44, 45). However, such a cancellation is not expected for the full
gap and point-­nodal superconductors. Another possibility of the
multigap effect has also been pointed out (43). However, it is
highly unlikely that the vanishingly small and nearly field inde-
pendent κ/T in ultraclean UTe2 is caused by the multiband effect.
DISCUSSION
Having established the absence of any type of nodes at and around
the a axis, we discuss the superconducting order parameter belong-
ing to the irreducible representation of the D2h point group in UTe2.
Our results rule out the B3u state with point nodes along the a axis.
NMR measurements reported a clear reduction of the Knight shift
for H||b and H||c (11, 12). These rule out the possibilities of the B1u
and B2u states. It should be noted that, because the Knight shift
measurements for H||a contain large error bars, it is premature to
A B
H//c
0.01
0.2 #1
UPt 3
(κ0/T) /(κ0N/T)
(κ 0/T) /(κ0N/T)
#2
0.00
0.0 0.1
URu2Si2
0.1
(24)
#1
0.13 K H//c
0.0 1%
0.0 #2 0.1 #1 0.2
H//c H//a
H/Hc2
Fig. 4. Field dependence of thermal conductivity of UTe2. (A and B) The residual thermal conductivity, κ0/T, normalized by the normal state value, κ0N/T, as a function
of H/Hc2. For both H‖c and H‖a, κ0/T is less than 1% of κ0N/T up to H/Hc2 = 0.09, which provides evidence for the absence of any types of nodes at or around the a axis.
Orange open squares depict (κ/T)/(κ0N/T) of sample #2 for H‖c at T = 0.13 K (see also fig. S4B). The inset shows an enlarged view in the low field region for H‖c. For comparison,
we plot data for other uranium superconductors UPt3 (42) and URu2Si2 (36) (A) and typical s-­wave superconductor Nb (39) and nodal superconductor UPt3 (42) (B). We also
plot the data for the moderately clean crystal with RRR =22 for Q‖H‖a (24). (C) Field dependence of κ/T at 0.22 K for H‖c and H‖a for sample #1.
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
conclude the B3u state pointed out in (12). On the basis of these results,
we conclude that the superconducting order parameter in UTe2 is
represented by the Au symmetry.
The recent penetration depth measurements have proposed a B3u +
iϵAu state with ϵ comparable to unity (26). This state has point nodes
close to the b and c axes, which may not be sensitively detected by
the a-­axis thermal conductivity. However, such a state can be safely
ruled out because the double superconducting transitions, required
for such a multicomponent state (46), have never been observed in
clean UTe2 crystals (18–20), even in the presence of uniaxial strain (21).
Here, we comment on the results of the specific heat. The quadratic
temperature dependence of C/T has been reported in the very clean
crystal with RRR = 220 (30), suggesting the presence of point nodes.
However, we point out that the heat capacity data of UTe2 need to be
scrutinized because they include contributions from both local and
itinerant excitations and may include local excitations that are not
directly related to the quasiparticles excited from the superconducting
condensate. Recent muon spin rotation experiments have reported
that uranium defects induce local magnetic clusters that give a finite
magnetic contribution to C/T (47). Given this situation in UTe2, ther-
mal conductivity, which is not influenced by such local excitations,
is a more direct bulk probe for determining the superconducting gap
In addition to the spin-­triplet full-­gap superconductivity, an
unexpected and unique feature of the superconducting state of UTe2
is that the quasiparticle excitations are extremely sensitive to disorder
due to impurities/defects. In Fig. 4A and fig. S6, we compare (κ0/T)/
(κ0N/T) for crystals with different RRR as a function of H/Hc2 and
RRR, respectively. The data of RRR = 22 are taken from (24). In
contrast to the present results, it has been pointed out that the field
dependence of κ/T of the moderately clean crystal bears a resemblance
to that of superconductors with point nodes (24). Moreover, as shown
in the inset of Fig. 4A, (κ0/T)/(κ0N/T) of the ultraclean crystal (#2) is
one order of magnitude smaller than that of the very clean crystal
(#1) at H/Hc2 ∼ 0.09, indicating that the quasiparticle excitations are
still strongly affected by the disorder even in the ultraclean region
where RRR is well above 260. It is well-­known that impurities have a
considerable effect on quasiparticle excitations in unconventional
superconductors, and this appears to be more pronounced in UTe2.
The extreme sensitivity of quasiparticle excitations to the disorder
C 1.5
1.0
#1
UTe2 (Q//H//a)
T = 0.22 K
(24)
1.0
κ /T (W/K m)
2
H//a
0.5 UPt 3 Nb
0.5
#1
H//a H//c
#1
H//c
0 0
0 #2 0.5 1.0 0 2 4 6 8 10
H//c
H/Hc2 µ0H (T)
4 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
should provide important clues to the mechanism of supercon-
ductivity in UTe2 that still remains elusive. A plausible origin of the
disorder is uranium defects whose concentration is very sensitive
to sample growth conditions (31). It has been suggested that the ura-
nium defects locally disrupt long-­range magnetic correlations (47–
49) that may have a large impact on the quasiparticle excitations.
We emphasize that the fully gapped Au symmetry is the realization
of spin-­triplet superconductivity analogous to the B-­phase of super-
fluid 3He. The full-­gap spin-­triplet superconductivity has also been
suggested in UBe13 (50, 51). However, in UBe13, the information of
the Fermi surface is almost absent, and notable deviation from the
Fermi liquid behavior is observed in the normal state (51). There-
fore, it is challenging to investigate the topological properties of UBe13.
In contrast, for UTe2 that exhibits the conventional Fermi liquid be-
havior in the normal state, the topological properties are characterized
by the shape of the Fermi surface (29).
It should be noted that UTe2 likely has the three-­dimensional (3D)
Fermi surface. Recent measurements of angle-­resolved photoemission
spectroscopy reveal the presence of 3D electronic structure with closed
Fermi surface along the c axis (52, 53). The 3D structure is further
supported by the almost isotropic normal state resistivity (54) and
the nearly isotropic coherence lengths along the a and c axes (Fig. 1E).
We note that, although several experiments have reported a larger
anisotropy of ξa and ξc by a factor of 2 to 3 (27, 55, 56), this does not
change the argument of the dimensionality of the Fermi surface.
Although very recent quantum oscillation measurements reported
the presence of 2D Fermi surfaces (30), this does not exclude the 3D
Fermi surface. This is because the topology of the Fermi surfaces
may change in strong magnetic fields due to a Lifshitz transition
(57). Moreover, the electronic specific heat coefficient γe ∼ 100 mJ/
K2 mol obtained from the quantum oscillations is still smaller than
γe∼ 120 mJ/K2 mol from the specific heat, implying the existence of
undiscovered Fermi surface. Given the 3D Fermi surface, the Au state
is characterized by a nontrivial 3D winding number (29). In this
case, UTe2 is a long-­sought 3D strong topological superconductor
with emergent Majorana surface states on any crystal plane. After
completion of this work, we became aware of NMR experiments on
the ultraclean crystal (58). The reduction of the a-­axis Knight shift is
consistent with the fully gapped Au symmetry.
MATERIALS AND METHODS
High-­quality single crystals of UTe2 were grown by a flux method as
described in (31). We used two crystals #1 and #2 from different
batches. Sample #1 was cleaved into two pieces. The data for sample
#1 are primarily from measurements on one piece. The other piece
of sample #1 was used for resistivity experiments to obtain the upper
critical fields and magnetoresistance. The size of the former piece is
2511 μm (length) by 147.5 μm (width) by 77.5 μm (thickness). The
size of the latter piece is 2230 μm (length) by 204.5 μm (width) by
75.3 μm (thickness). The size of sample #2 is 2467 μm (length) by
178.8 μm (width) by 104 μm (thickness). The resistivity was measured
by a standard four-­probe method by applying current along the
crystal a axis. Four gold wires were attached by spot welding. The
specific heat was measured by a long–relaxation time method as
described in (59, 60).
Thermal conductivity κ was measured by the standard steady-­
state method by applying temperature gradient Q along the crystal a
axis. To obtain good contacts with low contact resistance of ∼0.5 ohms,
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
four gold wires were attached by spot welding. The gold wires serve
heat links to a 10-­kilohm chip resistor as a heater, two field cali-
brated thermometers, and a silver plate. The silver plate was fixed
with silver paste to a copper block as a heat bath.
Supplementary Materials
This PDF file includes:
Figs. S1 to S6
REFERENCES AND NOTES
1. M. Sato, Y. Ando, Topological superconductors: A review. Rep. Prog. Phys. 80, 076501
(2017).
2. D. A. Ivanov, Non-­Abelian statistics of half-­quantum vortices in p-­wave superconductors.
Phys. Rev. Lett. 86, 268–271 (2001).
3. A. Y. Kitaev, Fault-­tolerant quantum computation by anyons. Ann. Phys. 303, 2–30 (2003).
4. A. J. Leggett, A theoretical description of the new phases of liquid 3He. Rev. Mod. Phys 47,
331–414 (1975).
5. S. Ran, C. Eckberg, Q.-­P. Ding, Y. Furukawa, T. Metz, S. R. Saha, I.-­L. Liu, M. Zic, H. Kim,
J. Paglione, N. P. Butch, Nearly ferromagnetic spin-­triplet superconductivity. Science 365,
684–687 (2019).
6. D. Aoki, A. Nakamura, F. Honda, D. Li, Y. Homma, Y. Shimizu, Y. J. Sato, G. Knebel,
J.-­P. Brison, A. Pourret, D. Braithwaite, G. Lapertot, Q. Niu, M. Vališka, H. Harima,
J. Flouquet, Unconventional superconductivity in heavy fermion UTe2. J. Phys. Soc. Jpn.
88, 043702 (2019).
7. S. Ran, I.-­L. Liu, Y. S. Eo, D. J. Campbell, P. M. Neves, W. T. Fuhrman, S. R. Saha, C. Eckberg,
H. Kim, D. Graf, F. Balakirev, J. Singleton, J. Paglione, N. P. Butch, Extreme magnetic
field-­boosted superconductivity. Nat. Phys. 15, 1250–1254 (2019).
8. W. Knafo, M. Nardone, M. Vališka, A. Zitouni, G. Lapertot, D. Aoki, G. Knebel, D. Braithwaite,
Comparison of two superconducting phases induced by a magnetic field in UTe2.
Commun. Phys. 4, 40 (2021).
9. G. Knebel, W. Knafo, A. Pourret, Q. Niu, M. Vališka, D. Braithwaite, G. Lapertot, M. Nardone,
A. Zitouni, S. Mishra, I. Sheikin, G. Seyfarth, J.-­P. Brison, D. Aoki, J. Flouquet, Field-­reentrant
superconductivity close to a metamagnetic transition in the heavy-­fermion
superconductor UTe2. J. Phys. Soc. Jpn. 88, 063707 (2019).
10. G. Nakamine, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai, S. Kambe, A. Nakamura,
Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Superconducting properties of heavy fermion
UTe2 revealed by 125Te-­nuclear magnetic resonance. J. Phys. Soc. Jpn 88, 113703 (2019).
11. G. Nakamine, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai, S. Kambe,
A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Anisotropic response of spin
susceptibility in the superconducting state of UTe2 probed with 125Te-­NMR
measurement. Phys. Rev. B. 103, L100503 (2021).
12. H. Fujibayashi, G. Nakamine, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai,
S. Kambe, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Superconducting
order parameter in UTe2 determined by Knight shift measurement. J. Phys. Soc. Jpn. 91,
043705 (2022).
13. S. Saxena, P. Agarwal, K. Ahilan, F. Grosche, R. Haselwimmer, M. Steiner, E. Pugh, I. Walker,
S. Julian, P. Monthoux, G. Lonzarich, A. Huxley, I. Sheikin, D. Braithwaite, J. Flouquet,
Superconductivity on the border of itinerant-­electron ferromagnetism in UGe2. Nature
406, 587–592 (2000).
14. D. Aoki, A. Huxley, E. Ressouche, D. Braithwaite, J. Flouquet, J.-­P. Brison, E. Lhotel,
C. Paulsen, Coexistence of superconductivity and ferromagnetism in URhGe. Nature 413,
613–616 (2001).
15. N. T. Huy, A. Gasparini, D. E. de Nijs, Y. Huang, J. C. P. Klaasse, T. Gortenmulder, A. de Visser,
A. Hamann, T. Görlach, H. V. Löhneysen, Superconductivity on the border of weak
itinerant ferromagnetism in UCoGe. Phys. Rev. Lett. 99, 067006 (2007).
16. I. M. Hayes, D. S. Wei, T. Metz, J. Zhang, Y. S. Eo, S. Ran, S. R. Saha, J. Collini, N. P. Butch,
D. F. Agterberg, A. Kapitulnik, J. Paglione, Multicomponent superconducting order
parameter in UTe2. Science 373, 797–801 (2021).
17. L. Jiao, S. Howard, S. Ran, Z. Wang, J. O. Rodriguez, M. Sigrist, Z. Wang, N. P. Butch,
V. Madhavan, Chiral superconductivity in heavy-­fermion metal UTe2. Nature 579,
523–527 (2020).
18. L. P. Cairns, C. R. Stevens, C. D. O’Neill, A. Huxley, Composition dependence of the
superconducting properties of UTe2. J. Phys. Condens. Matter 32, 415602 (2020).
19. S. M. Thomas, C. Stevens, F. B. Santos, S. S. Fender, E. D. Bauer, F. Ronning, J. D. Thompson,
A. Huxley, P. F. S. Rosa, Spatially inhomogeneous superconductivity in UTe2. Phys. Rev. B.
104, 224501 (2021).
20. P. F. Rosa, A. Weiland, S. S. Fender, B. L. Scott, F. Ronning, J. D. Thompson, E. D. Bauer,
S. M. Thomas, Single thermodynamic transition at 2 K in superconducting UTe2 single
crystals. Commun. Mater. 3, 33 (2022).
5 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
21. C. Girod, C. R. Stevens, A. Huxley, E. D. Bauer, F. B. Santos, J. D. Thompson, R. M. Fernandes,
J.-­X. Zhu, F. Ronning, P. F. S. Rosa, S. M. Thomas, Thermodynamic and electrical transport
properties of UTe2 under uniaxial stress. Phys. Rev. B. 106, L121101 (2022).
22. M. O. Ajeesh, M. Bordelon, C. Girod, S. Mishra, F. Ronning, E. D. Bauer, B. Maiorov,
J. D. Thompson, P. F. S. Rosa, S. M. Thomas, Fate of time-­reversal symmetry breaking in
UTe2. Phys. Rev. X. 13, 041019 (2023).
23. D. Aoki, J.-­P. Brison, J. Flouquet, K. Ishida, G. Knebel, Y. Tokunaga, Y. Yanase,
Unconventional superconductivity in UTe2. J. Phys. Condens. Matter 34, 243002 (2022).
24. T. Metz, S. Bae, S. Ran, I.-­L. Liu, Y. S. Eo, W. T. Fuhrman, D. F. Agterberg, S. M. Anlage,
N. P. Butch, J. Paglione, Point-­node gap structure of the spin-­triplet superconductor UTe2.
Phys. Rev. B. 100, 220504 (2019).
25. S. Bae, H. Kim, Y. S. Eo, S. Ran, I.-­L. Liu, W. T. Fuhrman, J. Paglione, N. P. Butch, S. M. Anlage,
Anomalous normal fluid response in a chiral superconductor UTe2. Nat. Commun. 12,
2644 (2021).
26. K. Ishihara, M. Roppongi, M. Kobayashi, K. Imamura, Y. Mizukami, H. Sakai, P. Opletal,
Y. Tokiwa, Y. Haga, K. Hashimoto, T. Shibauchi, Chiral superconductivity in UTe2 probed by
anisotropic low-­energy excitations. Nat. Commun. 14, 2966 (2023).
27. S. Kittaka, Y. Shimizu, T. Sakakibara, A. Nakamura, D. Li, Y. Homma, F. Honda, D. Aoki,
K. Machida, Orientation of point nodes and nonunitary triplet pairing tuned by the
easy-­axis magnetization in UTe2. Phys. Rev. Res. 2, 032014 (2020).
28. V. P. Mineev, Low temperature specific heat and thermal conductivity in superconducting
UTe2. J. Phys. Soc. Jpn. 91, 074601 (2022).
29. J. Ishizuka, S. Sumita, A. Daido, Y. Yanase, Insulator-­metal transition and topological
superconductivity in UTe2 from a first-­principles calculation. Phys. Rev. Lett. 123, 217001 (2019).
30. D. Aoki, H. Sakai, P. Opletal, Y. Tokiwa, J. Ishizuka, Y. Yanase, H. Harima, A. Nakamura, D. Li,
Y. Homma, Y. Shimizu, G. Knebel, J. Flouquet, Y. Haga, First observation of the de
Haas–van Alphen effect and Fermi surfaces in the unconventional superconductor UTe2.
J. Phys. Soc. Jpn. 91, 083704 (2022).
31. H. Sakai, P. Opletal, Y. Tokiwa, E. Yamamoto, Y. Tokunaga, S. Kambe, Y. Haga, Single crystal
growth of superconducting UTe2 by molten salt flux method. Phys. Rev. Mater. 6, 073401 (2022).
32. Y. Matsuda, K. Izawa, I. Vekhter, Nodal structure of unconventional superconductors
probed by angle resolved thermal transport measurements. J. Phys. Condens. Matter 18,
R705–R752 (2006).
33. K. Kadowaki, S. Woods, Universal relationship of the resistivity and specific heat in
heavy-­fermion compounds. Solid State Commun. 58, 507–509 (1986).
34. R. C. Yu, M. B. Salamon, J. P. Lu, W. C. Lee, Thermal conductivity of an untwinned
YBa2Cu3O7−δ single crystal and a new interpretation of the superconducting state
thermal transport. Phys. Rev. Lett. 69, 1431–1434 (1992).
35. K. Izawa, H. Yamaguchi, Y. Matsuda, H. Shishido, R. Settai, Y. Onuki, Angular position of
nodes in the superconducting gap of quasi-­2D heavy-­fermion superconductor CeCoIn5.
Phys. Rev. Lett. 87, 057002 (2001).
36. Y. Kasahara, T. Iwasawa, H. Shishido, T. Shibauchi, K. Behnia, Y. Haga, T. D. Matsuda,
Y. Onuki, M. Sigrist, Y. Matsuda, Exotic superconducting properties in the electron-­hole-­
compensated heavy-­fermion “semimetal” URu2Si2. Phys. Rev. Lett. 99, 116402 (2007).
37. L. Taillefer, B. Lussier, R. Gagnon, K. Behnia, H. Aubin, Universal heat conduction in
YBa2Cu3O69. Phys. Rev. Lett. 79, 483–486 (1997).
38. T. Yamashita, T. Takenaka, Y. Tokiwa, J. A. Wilcox, Y. Mizukami, D. Terazawa, Y. Kasahara,
S. Kittaka, T. Sakakibara, M. Konczykowski, S. Seiro, H. S. Jeevan, C. Geibel, C. Putzke,
T. Onishi, H. Ikeda, A. Carrington, T. Shibauchi, Y. Matsuda, Fully gapped
superconductivity with no sign change in the prototypical heavy-­fermion CeCu2Si2. Sci.
Adv. 3, e1601667 (2017).
39. J. Lowell, J. Sousa, Mixed-­state thermal conductivity of type II superconductors. J. Low
Temp. Phys. 3, 65–87 (1970).
40. S. Y. Zhou, H. Zhang, X. C. Hong, B. Y. Pan, X. Qiu, W. N. Dong, X. L. Li, S. Y. Li, Nodeless
superconductivity in Ca3lr4Sb13: Evidence from quasiparticle heat transport. Phys. Rev. B
86, 064504 (2012).
41. Z. Zhang, Y. Xu, C. Kuo, X. Hong, M. Wang, P. Cai, J. Dong, C. Lue, S. Li, Nodeless
superconducting gap in the caged-­type superconductors Y5Rh6Sn18 and Lu5Rh6Sn18.
Supercond. Sci. Technol. 28, 105008 (2015).
42. H. Suderow, J. Brison, A. Huxley, J. Flouquet, Thermal conductivity and gap structure of
the superconducting phases of UPt3. J. Low Temp. Phys. 108, 11–30 (1997).
43. G. Seyfarth, J. P. Brison, G. Knebel, D. Aoki, G. Lapertot, J. Flouquet, Multigap
superconductivity in the heavy-­fermion system CecoIn5. Phys. Rev. Lett. 101, 046401 (2008).
44. M. Franz, Mixed-­state quasiparticle transport in high-­​Tc cuprates. Phys. Rev. Lett. 82,
1760–1763 (1999).
45. Y. Kasahara, Y. Nakajima, K. Izawa, Y. Matsuda, K. Behnia, H. Shishido, R. Settai, Y. Onuki,
Anomalous quasiparticle transport in the superconducting state of CeCoIn5. Phys. Rev. B
72, 214515 (2005).
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
46. H. S. Røising, M. Geier, A. Kreisel, B. M. Andersen, Thermodynamic transitions and
topology of spin-­triplet superconductivity: Application to UTe2. arXiv:2311.06097 (2023).
47. S. Sundar, N. Azari, M. R. Goeks, S. Gheidi, M. Abedi, M. Yakovlev, S. R. Dunsiger,
J. M. Wilkinson, S. J. Blundell, T. E. Metz, I. M. Hayes, S. R. Saha, S. Lee, A. J. Woods,
R. Movshovich, S. M. Thomas, N. P. Butch, P. F. S. Rosa, J. Paglione, J. E. Sonier, Ubiquitous
spin freezing in the superconducting state of UTe2. Commun. Phys. 6, 24 (2023).
48. Y. Tokunaga, H. Sakai, S. Kambe, Y. Haga, Y. Tokiwa, P. Opletal, H. Fujibayashi, K. Kinjo,
S. Kitagawa, K. Ishida, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Slow
electronic dynamics in the paramagnetic state of UTe2. J. Physical Soc. Japan 91, 023707
(2022).
49. Y. Iguchi, H. Man, S. M. Thomas, F. Ronning, P. F. S. Rosa, K. A. Moler, Microscopic imaging
homogeneous and single phase superfluid density in UTe2. Phys. Rev. Lett. 130, 196003
(2023).
50. I. Fomin, J. Brison, Subdominant interactions and Hc2 in UBe13. J. Low Temp. Phys. 119,
627–643 (2000).
51. Y. Shimizu, D. Braithwaite, D. Aoki, B. Salce, J.-­P. Brison, Spin-­triplet p-­wave
superconductivity revealed under high pressure in UBe13. Phys. Rev. Lett. 122, 067001
(2019).
52. L. Miao, S. Liu, Y. Xu, E. C. Kotta, C.-­J. Kang, S. Ran, J. Paglione, G. Kotliar, N. P. Butch,
J. D. Denlinger, L. A. Wray, Low energy band structure and symmetries of UTe2 from
angle-­resolved photoemission spectroscopy. Phys. Rev. Lett. 124, 076401 (2020).
53. S.-­I. Fujimori, I. Kawasaki, Y. Takeda, H. Yamagami, A. Nakamura, Y. Homma, D. Aoki,
Electronic structure of UTe2 studied by photoelectron spectroscopy. J. Physical Soc. Japan
88, 103701 (2019).
54. Y. S. Eo, S. Liu, S. R. Saha, H. Kim, S. Ran, J. A. Horn, H. Hodovanets, J. Collini, T. Metz,
W. T. Fuhrman, A. H. Nevidomskyy, J. D. Denlinger, N. P. Butch, M. S. Fuhrer, L. A. Wray,
J. Paglione, c-­Axis transport in UTe2: Evidence of three-­dimensional conductivity
component. Phys. Rev. B 106, L060505 (2022).
55. A. Rosuel, C. Marcenat, G. Knebel, T. Klein, A. Pourret, N. Marquardt, Q. Niu, S. Rousseau,
A. Demuer, G. Seyfarth, G. Lapertot, D. Aoki, D. Braithwaite, J. Flouquet, J. P. Brison,
Field-­induced tuning of the pairing state in a superconductor. Phys. Rev. X 13, 011022 (2023).
56. K. Ishihara, M. Kobayashi, K. Imamura, M. Konczykowski, H. Sakai, P. Opletal, Y. Tokiwa,
Y. Haga, K. Hashimoto, T. Shibauchi, Anisotropic enhancement of lower critical field in
ultraclean crystals of spin-­triplet superconductor candidate UTe2. Phys. Rev. Res. 5,
L022002 (2023).
57. Q. Niu, G. Knebel, D. Braithwaite, D. Aoki, G. Lapertot, G. Seyfarth, J.-­P. Brison, J. Flouquet,
A. Pourret, Fermi-­surface instability in the heavy-­fermion superconductor UTe2. Phys. Rev.
Lett. 124, 086601 (2020).
58. H. Matsumura, H. Fujibayashi, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai,
S. Kambe, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Large reduction in
the a-­axis Knight shift on UTe2 with Tc = 2.1 K. J. Phys. Soc. Jpn. 92, 063701 (2023).
59. Y. Wang, B. Revaz, A. Erb, A. Junod, Direct observation and anisotropy of the contribution
of gap nodes in the low-­temperature specific heat of YBa2Cu3O7. Phys. Rev. B 63, 094508
(2001).
60. O. J. Taylor, A. Carrington, J. A. Schlueter, Specific-­heat measurements of the gap
structure of the organic superconductors κ−(ET)2Cu[N(CN)2]Br and κ−(ET)2Cu(NCS)2.
Phys. Rev. Lett. 99, 057001 (2007).
Acknowledgments: We thank K. Ishihara, T. Shibauchi, and Y. Tokunaga for insightful
discussions. Funding: This work is supported by Grants-­in-­Aid for Scientific Research
(KAKENHI) (nos. 18H05227, 18H03680, 18H01180, 21 K13881, 21 K18145, 22H01181,
22H04933, 23H00089, 23H01132, 23H04871, 23 K03332, 23 K13060, and 23 K17353) and on
Innovative Areas “Quantum Liquid Crystals” (no. 19H05824) from the Japan Society for the
Promotion of Science, JST CREST (JPMJCR19T5), and PRESTO (no. JPMJPR2252; T.A.). M.L.
acknowledges the support by Japan Society of for the Promotion of Science via the visiting
scholar fellowship. Author contributions: H.S., P.O., Y.T., and Y.H. synthesized the high-­quality
single crystals. S.S., M.S., M.K., T.A., and Y.S. performed thermal conductivity experiments. S.S.,
M.S., M.K., Y.Ko., H.A., S.I., and T.A. conducted resistivity experiments. Y.Ka. performed specific
heat experiments. All authors analyzed the data and discussed the results. S.S. and Y.M.
prepared the manuscript with inputs from all authors. Competing interests: The authors
declare that they have no competing interests. Data and materials availability: All data
needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials.
Submitted 18 August 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk3772
6 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

C O N D E N S E D M AT T E R P H Y S I C S Copyright © 2024 The

Authors, some rights
Fully gapped pairing state in spin-­triplet reserved; exclusive
licensee American
superconductor UTe2 Association for the
Advancement of
Science. No claim to
Shota Suetsugu1*, Masaki Shimomura1, Masashi Kamimura1, Tomoya Asaba1, Hiroto Asaeda1, original U.S.
Yuki Kosuge1, Yuki Sekino1, Shun Ikemori1, Yuichi Kasahara1, Yuhki Kohsaka1, Minhyea Lee2, Government Works.
Youichi Yanase1, Hironori Sakai3, Petr Opletal3, Yoshifumi Tokiwa3, Yoshinori Haga3, Distributed under a
Yuji Matsuda1* Creative Commons
Attribution License 4.0
The recently discovered superconductor UTe2 is a promising candidate for spin-­triplet superconductors, but the (CC BY).
symmetry of the superconducting order parameter remains highly controversial. Here, we determine the super-
conducting gap structure by the thermal conductivity of ultraclean UTe2 single crystals. We find that the a-­axis
thermal conductivity divided by temperature κ/T in zero-­temperature limit is vanishingly small for both magnetic
field H‖a and H‖c axes up to H/Hc2 ∼ 0.2, demonstrating the absence of nodes around the a axis contrary to the previous
belief. The present results, combined with the reduction of nuclear magnetic resonance Knight shift, indicate that
the superconducting order parameter belongs to the isotropic Au representation with a fully gapped pairing state,
analogous to the B phase of superfluid 3He. These findings reveal that UTe2 is likely to be a long-­sought three-­
dimensional strong topological superconductor, hosting helical Majorana surface states on any crystal plane.

INTRODUCTION was supported by scanning tunneling spectroscopy experiments (17).

Spin-­triplet pairing states have aroused tremendous interest because
of the emergence of Majorana quasiparticles (1) and their potential
application to fault-­tolerant quantum information processing (2, 3).
The most famous example is the superfluid 3He (4), and the quest for
its superconducting analog has been a long-­standing issue in condensed
matter physics. The promising candidate is the recently discovered
heavy fermion superconductor UTe2 (5). UTe2 exhibits highly unusual
superconducting properties, including extremely large upper critical
fields well exceeding the Pauli limit (5, 6), coexistence of super
conductivity and ferromagnetic order in high magnetic field (7, 8),
reentrant superconductivity that resembles to ferromagnetic super-
conductors URhGe and UCoGe (7, 9), and peculiar behavior of
nuclear magnetic resonance (NMR) Knight shift and absence of
coherence peak in the superconducting state (5, 10–12). All of these
notable properties are indicative of the spin-­triplet pairing state.
Moreover, this unconventional superconducting state occurs in the
paramagnetic state at ambient pressure. This is in contrast to the other
ferromagnetic superconductors; at ambient pressure, UGe2 shows a
ferromagnetic order (13), and, in URhGe and UCoGe, supercon-
ductivity coexists with ferromagnetic order (14, 15). Therefore, UTe2 is
a long-­sought material that allows us to examine superconducting
properties in more detail using various probes.
Of particular interest is the superconducting gap structure of UTe2,
which is of primary importance for understanding the peculiar super-
conducting state associated with the spin-­triplet pairing state. Despite
the intensive research efforts, however, the superconducting order
parameter has been highly controversial, and its determination is
still challenging. At an early stage, a chiral superconducting state
with a multicomponent order parameter has been suggested by the
double peak of the specific heat and polar Kerr effect (16), which
1
Department of Physics, Kyoto University, Kyoto 606-­8502, Japan. 2Department of
Physics, University of Colorado Boulder, Boulder, CO 80309, USA. 3Advanced Science
Research Center, Japan Atomic Energy Agency, Tokai, Ibaraki 319-­1195, Japan.
*Corresponding author. Email: suetsugu.​shota.​4n@​kyoto-­u.​ac.​jp (S.S.); matsuda@​
scphys.​kyoto-­u.​ac.​jp (Y.M.)
However, as the sample quality improves, a single peak is observed
in the specific heat (18–20) and does not split into two peaks even
under uniaxial strain (21). Moreover, recent Kerr effect experiments
on clean crystals report no evidence of broken time-­reversal symmetry
(22). Considering the orthorhombic crystal structure of UTe2 (Fig. 1A),
although the chiral state may appear at the surface, chiral supercon-
ductivity is unlikely to be realized in the bulk.
Given no solid evidence for the bulk chiral superconductivity in
UTe2, there are four possible gap symmetries for spin-­triplet super-
conductivity, {Au, B1u, B2u, B3u} (23). The Au symmetry is fully gapped
(Fig. 1B), corresponding to the B phase of superfluid 3He. On the
other hand, B1u, B2u, and B3u symmetries have point nodes at isolated
points on the Fermi surface along the c, b, and a axes, respectively
(Fig. 1C). Several measurements, such as thermal conductivity (24),
penetration depth (25, 26), and specific heat (27), have reported the
presence of the low-­energy quasiparticle excitations, suggesting the
presence of point nodes rather than line nodes. NMR measurements
show that, below the superconducting transition temperature Tc, the
Knight shift along the b and c axes decreases (11, 12). This excludes
the B1u and B2u symmetries. On the basis of these results, it has been
widely discussed that the B3u symmetry with point nodes along the
a axis accounts for the superconducting gap of UTe2. The point node(s)
around the a axis has also been suggested by thermal conductivity (24).
However, it should be stressed that, because the low-­energy qua-
siparticle excitations in spin-­triplet superconductors are extremely
sensitive to impurities (28), the determination of the gap symmetry
in the previous measurements may be hindered by impurity. More-
over, as mentioned previously, the gap structure near the surface may
be distinct from the bulk. Therefore, it is premature to conclude the
presence of point nodes along the a axis. Determining whether the
gap symmetry is Au or B3u is of crucial importance because the topo-
logical properties of these two symmetries are fundamentally different
(29). Thus, the situation calls for reexamining the gap structure using
a bulk probe on crystals with an extremely low impurity concentration.
Here, we determined the superconducting gap structure of ultra-
clean UTe2 crystals (30, 31) by thermal conductivity κ, which is a

Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024 1 of 6

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
A D 500
#2
15
ρ (microhm·cm)
ρ (microhm·cm)
400
10 RRR = 630
300
5
200 0 present results indicate ξa ≈ ξc. This suggests that the average Fermi
U
100
Te 0
B 0 100
T (K)
E 20
#1
15 H//c
µ0Hc2 (T)
Fermi surface
10
C 0
5 T (K)
a H//a
0 K2 mol. The Kadowaki-­Woods ratio, A/γ2 ≈ 2 × 10−5 microhm·cm
0 1
Fig. 1. Crystal structure, resistivity, and phase diagram of UTe2. (A) Crystal
structure of UTe2. The gray and green spheres represent U and Te atoms, respec-
tively. (B and C) Structure of the superconducting gap Δ for Au (B) and B3u (C) sym-
metries for spherical Fermi surface (blue sphere). The B3u state has point nodes (red
points) along the a axis. (D) Temperature dependence of resistivity ρ for sample #2.
The inset shows ρ as a function of T2 at low temperatures. The residual resistivity ρ0
is obtained by a fit to ρ(T) with ρ0 + AT2 (dashed line). (E) H-­T phase diagram deter-
mined by resistivity measurements of sample #1. The linear extrapolations to T = 0
yield the upper critical fields μ0Hc2 of ∼12 T for H‖a and ∼17 T for H‖c. The inset
shows an enlarged view near Tc.
bulk and directional probe of the superconducting gap structure
(32). In contrary to the previous reports, the present results demon-
strate the absence of any type of nodes at and around the a axis. This
indicates that the superconducting order parameter of UTe2 belongs
to the isotropic Au representation with a fully gapped pairing state,
analogous to the B phase of superfluid 3He.
RESULTS
The superconducting transition temperature Tc of both crystals #1 and
#2 is 2.1 K, which is higher than the previously reported value of typi-
cally ∼2.0 K (20). Figure 1D displays the temperature (T) dependence
of the resistivity ρ(T) along the a axis for sample #2. The normal state
resistivity is well fitted by ρ(T) = ρ0 + AT2 (inset of Fig. 1D). The qua-
dratic temperature dependence of ρ, a characteristic of Fermi liquid,
indicates the importance of the electron-­electron correlation. The
residual resistivity ratio (RRR ≡ ρ(300 K)/ρ0) of 260 for sample #1
and 630 for sample #2 is nearly 10 and 30 times larger than that
reported in crystal with RRR = 22 (24), respectively. Hereafter, the
samples with RRR = 22 (24), 260 (#1), and 630 (#2) will be referred
to as samples in moderately clean, very-­clean, and ultraclean regions,
respectively. It should be noted that clear quantum oscillations are
reported in the sample with RRR = 220 (30). These confirm the high
quality of the present crystals. By comparing these crystals with different
RRRs, we can obtain pivotal information about the quasiparticle exci-
tations that are intimately related to the superconducting gap function.
The upper critical fields determined by the resistive transition of
c
sample #1 for H‖a and H‖c, Hc2, and Hc2, respectively, are displayed
a
a c
in Fig. 1E. At zero temperature, Hc2 (0) and Hc2 (0) are approximately
12 and 17 T, respectively. The upper critical fields of sample #2 is
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
a c
close to those of #1. As shown in the inset, Hc2 is very close to Hc2 in
the vicinity of Tc. Because the initial slope at Tc is proportional to the
a,c
orbital limiting field, − (dHc2 ∕dT) ∣Tc ∝ 1 ∕ ξb ξc,a, where ξa, ξb, and
ξc are the coherence length along the a, b and c axes, respectively, the
0 10 20 30
2 2
velocity along the a axis is close to that along the c axis.
T (K ) The heat capacity C of sample #1 exhibits a very sharp transition
200 300 at Tc with no peak splitting (fig. S1). For comparison, we plot the
reported data for a very clean crystal with Tc = 2.1 K (30), whose
RRR = 220 is close to sample #1, and for a crystal with Tc = 1.7 K
(20). The temperature dependence of C/T for sample #1 is very close
µ0Hc2 (T)
2
to the data of the very clean crystal with a similar RRR that shows a
negligibly small residual C/T in the zero temperature limit. These
results further support the high quality of the present crystals. The
1.5 2.0
electronic heat capacity coefficient in the normal state γ is 120 mJ/
2 3
T (K)
(mol K/mJ)2, is close to that of typical correlated systems (33).
Figure 2 shows the T-­dependence of κ/T with applied thermal
current Q‖a for the very clean (#1) and ultraclean (#2) crystals, along
with κ/T for the moderately clean crystal (24). In the normal state,
both electron and phonon contribute to the thermal conductivity. To
evaluate the electron contribution in the normal state at very low
temperatures, we measure the normal state thermal conductivity
κN/T (open blue circles) by applying magnetic fields of μ0H = 12 T,
which exceeds Hc2 a
. The blue dashed line represents the electron contri-
bution yielded from the Wiedemann-­Franz (WF) law L0/ρN. Here,
L0 = 2.44 × 10−8 W·ohm·K−2 is the Lorenz number, and the normal
state resistivity ρN below Tc is calculated from the extrapolation above
Tc by assuming ρN(T) = ρ0 + AT2. As shown in Fig. 2, κN/T well
coincides with L0/ρN below 0.5 K, indicating that the WF law holds.
This demonstrates that, for the very clean and ultraclean crystals,
the electron contribution is dominant in the normal state at very low
temperatures. In zero field, κ/T exhibits a distinct kink upon entering
the superconducting state, increases steeply, and reaches a maximum
at ∼1.5 and ∼1.2 K for samples #1 and #2, respectively. As observed
in several strongly correlated systems (34–36), the enhancement of
κ/T below Tc is attributed to the rapid enhancement of the quasiparticle
mean free path, which is caused by the suppression of the electron-­
electron inelastic scattering rate due to the superconducting gap
#1
2.0
12 T(H//a)
1.5
κ/T (W/K m)
2
#2 #1
L 0 /ρN
1.0 0T #1
0T
(24)
0.5 0.05 T
L 0 /ρN (24)
0.0
0.0 1.0 2.0 3.0
T (K)
Fig. 2. Temperature dependence of thermal conductivity of UTe2. In zero field,
κ/T shows a kink at Tc and then increases up to ∼1.5 and ∼1.2 K for samples #1 and #2,
respectively. The blue dashed line represents the electronic contribution estimated
by L0/ρN for sample #1. At low temperatures, L0/ρN is close to the normal state value
of κ/T at 12 T for H‖a (blue open circles), indicating that κ is dominated by the
electronic contribution. For comparison, we plot κ/T at 0.05 T and L0/ρN for the
moderately clean crystal with RRR = 22 (24).
2 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

formation. The enhancement of κ/T of sample #2 is more substantial
than sample #1 because the mean free path is larger in samples with
larger RRR. We note that the enhancement of κ/T in the moderately
clean crystal is much smaller than in these crystals. Moreover, κ0N/T ≡
κN/T (T → 0) of sample #1 determined by the data taken at 12 T is
1.9 W/K2 m, one order of magnitude larger than 0.1 W/K2 m re-
ported for the moderately clean crystal (24). The difference of κ0N/T
also reflects the significantly enhanced mean free path in the very
clean and ultraclean crystals.
Figure 3A shows the T-­dependence of κ/T at low temperatures
for samples #1 and #2. The thermal conductivity in the supercon-
ducting state has quasiparticle and phonon contributions, κ = κqp +
κph. In the boundary-­limited scattering regime at low temperatures,
κph/T is proportional to ℓphT2, where ℓph is the phonon mean free
path. While ℓph is T-­independent for diffuse scattering limit, resulting
in κph/T ∝ T2, it is T−1-­dependent for specular reflection, resulting
in κph/T ∝ T. In real systems, κph/T ∝ Tp with 1 ≤ p ≤ 2. Because
κph/T becomes zero at T = 0, κ0/T ≡ κ/T (T → 0) contains only the
quasiparticle contribution. It is apparent the linear extrapolation of
κ/T to T = 0 yields a negative intercept. We find that κ/T for both
samples is best fitted by κ/T ∝ Tα with α = 1.48 below 0.3 K (inset of
Fig. 3A). These results indicate the absence of residual thermal con-
ductivity κ0/T ≈ 0. For unconventional superconductors with line
nodes in the energy gap, a finite residual term is expected because of
the existence of a residual normal fluid (37). Therefore, the present
results provide evidence for the absence of line nodes. This is further
supported by the estimation of the residual term in a line-­nodal super
conductor. For line nodes, the universal expression of κ0/T for unitary
L ℏ
scattering is given by μ λ0 2 πΔ , where μ0 is the permeability of vacuum
0 0
and λ is the penetration depth. Using 2Δ0 = 3.5kBTc and λ ≈ 1000
nm (24, 25), we obtain κ0/T ≈ 0.054 W/K2 m (gray dashed line).
Clearly, κ/T at the lowest temperature is less than the calculated
value for the line nodes.
We next examine the presence/absence of point nodes. It is
well-­known that the heat transport at low magnetic fields in nodal
superconductors is fundamentally different from that in full-­gap super-
conductors. In magnetic fields, the energy of quasiparticles with
momentum ℏk in a circulating supercurrent flow vs shifts as Ek →
Ek − ℏk · vs. As a result of this Doppler shift, the thermal conductivity
in nodal superconductors is dominated by the contribution from
delocalized √ quasiparticles, leading to an initial steep increase of
κ(H) ∕ T ∝ H for line nodes and κ(H)/T ∝ ∣H log H∣ for point
nodes (38). By contrast, in full-­gap type II superconductors in the
clean limit, all quasiparticles are bound to the vortex cores, and the
magnetic field has a negligible effect on the heat transport (39–41)
except in the vicinity of Hc2, where the vortex cores overlap.
Thermal conductivity is a directional probe sensitive to the quasi-
particles with momentum parallel to the thermal current (k · Q ≠ 0)
and perpendicular to the magnetic field (k × H ≠ 0) because H ⊥ vs.
To investigate the quasiparticle excitations around the a axis, we
measured κ along the a axis (Q‖a) for H‖c and H‖a (Fig. 3, B and C).
As seen in the insets, while the former geometry sensitively probes
the point node at the a axis, the latter geometry selectively probes
quasiparticle excitations from point nodes that are off a axis but have
momentum in the a-­axis direction. For the very clean crystal (#1),
κ/T collapses into a single curve and decreases almost linearly below
∼0.25 K for both H‖c and H‖a. For the ultraclean crystal (#2), κ/T
normalized by the normal state value κ0N/T for H‖c is nearly half of
that for the very clean crystal below ∼0.3 K (Fig. 3B). Simple linear
extrapolations to T = 0 give negative intercepts for all data, indicating
a very small residual κ/T at T = 0. To obtain the zero temperature
limit of κ/T quantitatively, the data are fitted using κ/T = κ0/T + ATα
with κ0/T ≥ 0 and 1 ≤ α ≤ 2. This yields the vanishingly small κ0/T at
low fields for both samples (figs. S2 to S4A).
Figure 4A shows κ0/T normalized by κ0N/T as a function of H/Hc2.
κ0/T of the very clean crystal (#1) is less than 1% of the normal-­state
value even at H ∼ 0.09 Hc2 c
for H‖c and at H ∼ 0.2 Hc2 a
for H‖a.
Moreover, as shown in the inset, (κ0/T)/(κ0N/T) of the ultraclean
crystal (#2) for H‖c is significantly suppressed from that of the very
clean crystal. For comparison, we plot data for two other U-­based su-
perconductors, UPt3 (42) and URu2Si2 (36), which are believed to
have both point and line nodes. In notable contrast to the present
results, κ0/T for UPt3 and URu2Si2 exhibits a steep increase at low
fields. As shown in Fig. 4A (see also fig. S4B), even at 0.13 K, κ/T of
c
the ultraclean crystal is only 2% of κ0N/T up to H ∼ 0.12 Hc2, still
much smaller than (κ0/T)/(κ0N/T) for UPt3 and URu2Si2. These results

A 0.6 B C 0.15
0.3 H//c 1T 0.4 T
vs 0.4 T 1.5 T
0.10
κ/T (W/K m)

0.2 2T Q//H//a
2

2T
(κ/T) /(κ0N/T)

(κ/T) /(κ0N/T)
/T (W/K m)

0.4 #1 1T 0.10
0.1 1.5 T vs
2

Q//a 1.1 T
#2 #1
0.0
0.00 0.10 0.05
0.2 2T 0.05
κ/T

1.48 1.48
T (K ) #1 1.5 T
#2
µ 0H = 0 T 0.5 T
0.0 0.00 0.00
0.0 0.1 0.2 0.3 0.0 0.1 0.2 0.3 0.0 0.1 0.2 0.3
T (K) T (K) T (K)

Fig. 3. Temperature dependence of thermal conductivity of UTe2 at low temperatures. (A) Thermal conductivity divided by temperature, κ/T, in zero field for the very
clean (#1) and ultraclean (#2) crystals. The data are extrapolated to T = 0 by κ/T = κ0/T + ATα with κ0/T ≥ 0 (black dashed and dotted lines), yielding κ0/T ≈ 0 and α = 1.48
for both samples. The inset displays κ/T and the fit as a function of T1.48. The gray dashed horizontal lines represent the universal constant for line nodes (see Results section).
(B and C) Temperature dependence of κ/T normalized by the normal state value in the zero temperature limit κN0/T for magnetic fields H parallel to the crystal c axis (B)
and a axis (C) with applied thermal current Q‖a. The normal state value κ0N/T of the very clean crystal for H‖a is determined by the data at 12 T (blue open circles in Fig. 2).
κ0N/T for other configurations is approximated by L0/ρ0. As illustrated in the inset, κ/T for H‖c (H‖a) selectively probes the quasiparticles in the red-­shaded area and is
sensitive to the nodes (blue circles) at (around) the a axis.

Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024 3 of 6

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
indicate that UTe2 is essentially different from other U-­based un-
conventional superconductors in that only a very few delocalized
quasiparticles are excited, even well inside the vortex state. In
Fig. 4B, we compare κ0/T of UTe2 with typical s-­wave superconductor
Nb (39) and nodal superconductor UPt3 (42) over a wide field range.
For Nb, we show the data measured in both ascending and descending
magnetic fields, although the difference is small. Here, we restrict
the argument of H-­dependence for H‖c to the low field regime,
where the normal-­state magnetoresistance is small and κ0N/T can be
approximated by L0/ρN (fig. S5). It is evident that H-­dependence of
UTe2 is very close to that of Nb. These results lead us to conclude
that quasiparticle excitations with the velocity component along the
a axis are negligibly small, ruling out the presence of point nodes at
and around the a axis. This indicates that κ in zero field is dominated
by the phonon contribution κph at low temperatures, consistent with
the power law behavior κ/T ∝ T1.48 (see inset of Fig. 3A).
The field dependence of κ/T at 0.22 K for H‖c and H‖a (Fig. 4C),
which are both nearly constant at low fields, further supports the
absence of point nodes around the a axis. It has been reported that
the thermal conductivity of the d-­wave superconductor CeCoIn5 shows
little H-­dependence at √
very low temperatures (35, 43). One possible
explanation is that the H dependence of the quasiparticle density
of states is canceled out by the reduction of the quasiparticle mean
√ structure of this material.
free path, which is proportional to intervortex distance ∝ 1 ∕ H
(44, 45). However, such a cancellation is not expected for the full
gap and point-­nodal superconductors. Another possibility of the
multigap effect has also been pointed out (43). However, it is
highly unlikely that the vanishingly small and nearly field inde-
pendent κ/T in ultraclean UTe2 is caused by the multiband effect.
DISCUSSION
Having established the absence of any type of nodes at and around
the a axis, we discuss the superconducting order parameter belong-
ing to the irreducible representation of the D2h point group in UTe2.
Our results rule out the B3u state with point nodes along the a axis.
NMR measurements reported a clear reduction of the Knight shift
for H||b and H||c (11, 12). These rule out the possibilities of the B1u
and B2u states. It should be noted that, because the Knight shift
measurements for H||a contain large error bars, it is premature to
A B
H//c
0.01
0.2 #1
UPt 3
(κ0/T) /(κ0N/T)
(κ 0/T) /(κ0N/T)
#2
0.00
0.0 0.1
URu2Si2
0.1
(24)
#1
0.13 K H//c
0.0 1%
0.0 #2 0.1 #1 0.2
H//c H//a
H/Hc2
Fig. 4. Field dependence of thermal conductivity of UTe2. (A and B) The residual thermal conductivity, κ0/T, normalized by the normal state value, κ0N/T, as a function
of H/Hc2. For both H‖c and H‖a, κ0/T is less than 1% of κ0N/T up to H/Hc2 = 0.09, which provides evidence for the absence of any types of nodes at or around the a axis.
Orange open squares depict (κ/T)/(κ0N/T) of sample #2 for H‖c at T = 0.13 K (see also fig. S4B). The inset shows an enlarged view in the low field region for H‖c. For comparison,
we plot data for other uranium superconductors UPt3 (42) and URu2Si2 (36) (A) and typical s-­wave superconductor Nb (39) and nodal superconductor UPt3 (42) (B). We also
plot the data for the moderately clean crystal with RRR =22 for Q‖H‖a (24). (C) Field dependence of κ/T at 0.22 K for H‖c and H‖a for sample #1.
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
conclude the B3u state pointed out in (12). On the basis of these results,
we conclude that the superconducting order parameter in UTe2 is
represented by the Au symmetry.
The recent penetration depth measurements have proposed a B3u +
iϵAu state with ϵ comparable to unity (26). This state has point nodes
close to the b and c axes, which may not be sensitively detected by
the a-­axis thermal conductivity. However, such a state can be safely
ruled out because the double superconducting transitions, required
for such a multicomponent state (46), have never been observed in
clean UTe2 crystals (18–20), even in the presence of uniaxial strain (21).
Here, we comment on the results of the specific heat. The quadratic
temperature dependence of C/T has been reported in the very clean
crystal with RRR = 220 (30), suggesting the presence of point nodes.
However, we point out that the heat capacity data of UTe2 need to be
scrutinized because they include contributions from both local and
itinerant excitations and may include local excitations that are not
directly related to the quasiparticles excited from the superconducting
condensate. Recent muon spin rotation experiments have reported
that uranium defects induce local magnetic clusters that give a finite
magnetic contribution to C/T (47). Given this situation in UTe2, ther-
mal conductivity, which is not influenced by such local excitations,
is a more direct bulk probe for determining the superconducting gap
In addition to the spin-­triplet full-­gap superconductivity, an
unexpected and unique feature of the superconducting state of UTe2
is that the quasiparticle excitations are extremely sensitive to disorder
due to impurities/defects. In Fig. 4A and fig. S6, we compare (κ0/T)/
(κ0N/T) for crystals with different RRR as a function of H/Hc2 and
RRR, respectively. The data of RRR = 22 are taken from (24). In
contrast to the present results, it has been pointed out that the field
dependence of κ/T of the moderately clean crystal bears a resemblance
to that of superconductors with point nodes (24). Moreover, as shown
in the inset of Fig. 4A, (κ0/T)/(κ0N/T) of the ultraclean crystal (#2) is
one order of magnitude smaller than that of the very clean crystal
(#1) at H/Hc2 ∼ 0.09, indicating that the quasiparticle excitations are
still strongly affected by the disorder even in the ultraclean region
where RRR is well above 260. It is well-­known that impurities have a
considerable effect on quasiparticle excitations in unconventional
superconductors, and this appears to be more pronounced in UTe2.
The extreme sensitivity of quasiparticle excitations to the disorder
C 1.5
1.0
#1
UTe2 (Q//H//a)
T = 0.22 K
(24)
1.0
κ /T (W/K m)
2
H//a
0.5 UPt 3 Nb
0.5
#1
H//a H//c
#1
H//c
0 0
0 #2 0.5 1.0 0 2 4 6 8 10
H//c
H/Hc2 µ0H (T)
4 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
should provide important clues to the mechanism of supercon-
ductivity in UTe2 that still remains elusive. A plausible origin of the
disorder is uranium defects whose concentration is very sensitive
to sample growth conditions (31). It has been suggested that the ura-
nium defects locally disrupt long-­range magnetic correlations (47–
49) that may have a large impact on the quasiparticle excitations.
We emphasize that the fully gapped Au symmetry is the realization
of spin-­triplet superconductivity analogous to the B-­phase of super-
fluid 3He. The full-­gap spin-­triplet superconductivity has also been
suggested in UBe13 (50, 51). However, in UBe13, the information of
the Fermi surface is almost absent, and notable deviation from the
Fermi liquid behavior is observed in the normal state (51). There-
fore, it is challenging to investigate the topological properties of UBe13.
In contrast, for UTe2 that exhibits the conventional Fermi liquid be-
havior in the normal state, the topological properties are characterized
by the shape of the Fermi surface (29).
It should be noted that UTe2 likely has the three-­dimensional (3D)
Fermi surface. Recent measurements of angle-­resolved photoemission
spectroscopy reveal the presence of 3D electronic structure with closed
Fermi surface along the c axis (52, 53). The 3D structure is further
supported by the almost isotropic normal state resistivity (54) and
the nearly isotropic coherence lengths along the a and c axes (Fig. 1E).
We note that, although several experiments have reported a larger
anisotropy of ξa and ξc by a factor of 2 to 3 (27, 55, 56), this does not
change the argument of the dimensionality of the Fermi surface.
Although very recent quantum oscillation measurements reported
the presence of 2D Fermi surfaces (30), this does not exclude the 3D
Fermi surface. This is because the topology of the Fermi surfaces
may change in strong magnetic fields due to a Lifshitz transition
(57). Moreover, the electronic specific heat coefficient γe ∼ 100 mJ/
K2 mol obtained from the quantum oscillations is still smaller than
γe∼ 120 mJ/K2 mol from the specific heat, implying the existence of
undiscovered Fermi surface. Given the 3D Fermi surface, the Au state
is characterized by a nontrivial 3D winding number (29). In this
case, UTe2 is a long-­sought 3D strong topological superconductor
with emergent Majorana surface states on any crystal plane. After
completion of this work, we became aware of NMR experiments on
the ultraclean crystal (58). The reduction of the a-­axis Knight shift is
consistent with the fully gapped Au symmetry.
MATERIALS AND METHODS
High-­quality single crystals of UTe2 were grown by a flux method as
described in (31). We used two crystals #1 and #2 from different
batches. Sample #1 was cleaved into two pieces. The data for sample
#1 are primarily from measurements on one piece. The other piece
of sample #1 was used for resistivity experiments to obtain the upper
critical fields and magnetoresistance. The size of the former piece is
2511 μm (length) by 147.5 μm (width) by 77.5 μm (thickness). The
size of the latter piece is 2230 μm (length) by 204.5 μm (width) by
75.3 μm (thickness). The size of sample #2 is 2467 μm (length) by
178.8 μm (width) by 104 μm (thickness). The resistivity was measured
by a standard four-­probe method by applying current along the
crystal a axis. Four gold wires were attached by spot welding. The
specific heat was measured by a long–relaxation time method as
described in (59, 60).
Thermal conductivity κ was measured by the standard steady-­
state method by applying temperature gradient Q along the crystal a
axis. To obtain good contacts with low contact resistance of ∼0.5 ohms,
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
four gold wires were attached by spot welding. The gold wires serve
heat links to a 10-­kilohm chip resistor as a heater, two field cali-
brated thermometers, and a silver plate. The silver plate was fixed
with silver paste to a copper block as a heat bath.
Supplementary Materials
This PDF file includes:
Figs. S1 to S6
REFERENCES AND NOTES
1. M. Sato, Y. Ando, Topological superconductors: A review. Rep. Prog. Phys. 80, 076501
(2017).
2. D. A. Ivanov, Non-­Abelian statistics of half-­quantum vortices in p-­wave superconductors.
Phys. Rev. Lett. 86, 268–271 (2001).
3. A. Y. Kitaev, Fault-­tolerant quantum computation by anyons. Ann. Phys. 303, 2–30 (2003).
4. A. J. Leggett, A theoretical description of the new phases of liquid 3He. Rev. Mod. Phys 47,
331–414 (1975).
5. S. Ran, C. Eckberg, Q.-­P. Ding, Y. Furukawa, T. Metz, S. R. Saha, I.-­L. Liu, M. Zic, H. Kim,
J. Paglione, N. P. Butch, Nearly ferromagnetic spin-­triplet superconductivity. Science 365,
684–687 (2019).
6. D. Aoki, A. Nakamura, F. Honda, D. Li, Y. Homma, Y. Shimizu, Y. J. Sato, G. Knebel,
J.-­P. Brison, A. Pourret, D. Braithwaite, G. Lapertot, Q. Niu, M. Vališka, H. Harima,
J. Flouquet, Unconventional superconductivity in heavy fermion UTe2. J. Phys. Soc. Jpn.
88, 043702 (2019).
7. S. Ran, I.-­L. Liu, Y. S. Eo, D. J. Campbell, P. M. Neves, W. T. Fuhrman, S. R. Saha, C. Eckberg,
H. Kim, D. Graf, F. Balakirev, J. Singleton, J. Paglione, N. P. Butch, Extreme magnetic
field-­boosted superconductivity. Nat. Phys. 15, 1250–1254 (2019).
8. W. Knafo, M. Nardone, M. Vališka, A. Zitouni, G. Lapertot, D. Aoki, G. Knebel, D. Braithwaite,
Comparison of two superconducting phases induced by a magnetic field in UTe2.
Commun. Phys. 4, 40 (2021).
9. G. Knebel, W. Knafo, A. Pourret, Q. Niu, M. Vališka, D. Braithwaite, G. Lapertot, M. Nardone,
A. Zitouni, S. Mishra, I. Sheikin, G. Seyfarth, J.-­P. Brison, D. Aoki, J. Flouquet, Field-­reentrant
superconductivity close to a metamagnetic transition in the heavy-­fermion
superconductor UTe2. J. Phys. Soc. Jpn. 88, 063707 (2019).
10. G. Nakamine, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai, S. Kambe, A. Nakamura,
Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Superconducting properties of heavy fermion
UTe2 revealed by 125Te-­nuclear magnetic resonance. J. Phys. Soc. Jpn 88, 113703 (2019).
11. G. Nakamine, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai, S. Kambe,
A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Anisotropic response of spin
susceptibility in the superconducting state of UTe2 probed with 125Te-­NMR
measurement. Phys. Rev. B. 103, L100503 (2021).
12. H. Fujibayashi, G. Nakamine, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai,
S. Kambe, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Superconducting
order parameter in UTe2 determined by Knight shift measurement. J. Phys. Soc. Jpn. 91,
043705 (2022).
13. S. Saxena, P. Agarwal, K. Ahilan, F. Grosche, R. Haselwimmer, M. Steiner, E. Pugh, I. Walker,
S. Julian, P. Monthoux, G. Lonzarich, A. Huxley, I. Sheikin, D. Braithwaite, J. Flouquet,
Superconductivity on the border of itinerant-­electron ferromagnetism in UGe2. Nature
406, 587–592 (2000).
14. D. Aoki, A. Huxley, E. Ressouche, D. Braithwaite, J. Flouquet, J.-­P. Brison, E. Lhotel,
C. Paulsen, Coexistence of superconductivity and ferromagnetism in URhGe. Nature 413,
613–616 (2001).
15. N. T. Huy, A. Gasparini, D. E. de Nijs, Y. Huang, J. C. P. Klaasse, T. Gortenmulder, A. de Visser,
A. Hamann, T. Görlach, H. V. Löhneysen, Superconductivity on the border of weak
itinerant ferromagnetism in UCoGe. Phys. Rev. Lett. 99, 067006 (2007).
16. I. M. Hayes, D. S. Wei, T. Metz, J. Zhang, Y. S. Eo, S. Ran, S. R. Saha, J. Collini, N. P. Butch,
D. F. Agterberg, A. Kapitulnik, J. Paglione, Multicomponent superconducting order
parameter in UTe2. Science 373, 797–801 (2021).
17. L. Jiao, S. Howard, S. Ran, Z. Wang, J. O. Rodriguez, M. Sigrist, Z. Wang, N. P. Butch,
V. Madhavan, Chiral superconductivity in heavy-­fermion metal UTe2. Nature 579,
523–527 (2020).
18. L. P. Cairns, C. R. Stevens, C. D. O’Neill, A. Huxley, Composition dependence of the
superconducting properties of UTe2. J. Phys. Condens. Matter 32, 415602 (2020).
19. S. M. Thomas, C. Stevens, F. B. Santos, S. S. Fender, E. D. Bauer, F. Ronning, J. D. Thompson,
A. Huxley, P. F. S. Rosa, Spatially inhomogeneous superconductivity in UTe2. Phys. Rev. B.
104, 224501 (2021).
20. P. F. Rosa, A. Weiland, S. S. Fender, B. L. Scott, F. Ronning, J. D. Thompson, E. D. Bauer,
S. M. Thomas, Single thermodynamic transition at 2 K in superconducting UTe2 single
crystals. Commun. Mater. 3, 33 (2022).
5 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
21. C. Girod, C. R. Stevens, A. Huxley, E. D. Bauer, F. B. Santos, J. D. Thompson, R. M. Fernandes,
J.-­X. Zhu, F. Ronning, P. F. S. Rosa, S. M. Thomas, Thermodynamic and electrical transport
properties of UTe2 under uniaxial stress. Phys. Rev. B. 106, L121101 (2022).
22. M. O. Ajeesh, M. Bordelon, C. Girod, S. Mishra, F. Ronning, E. D. Bauer, B. Maiorov,
J. D. Thompson, P. F. S. Rosa, S. M. Thomas, Fate of time-­reversal symmetry breaking in
UTe2. Phys. Rev. X. 13, 041019 (2023).
23. D. Aoki, J.-­P. Brison, J. Flouquet, K. Ishida, G. Knebel, Y. Tokunaga, Y. Yanase,
Unconventional superconductivity in UTe2. J. Phys. Condens. Matter 34, 243002 (2022).
24. T. Metz, S. Bae, S. Ran, I.-­L. Liu, Y. S. Eo, W. T. Fuhrman, D. F. Agterberg, S. M. Anlage,
N. P. Butch, J. Paglione, Point-­node gap structure of the spin-­triplet superconductor UTe2.
Phys. Rev. B. 100, 220504 (2019).
25. S. Bae, H. Kim, Y. S. Eo, S. Ran, I.-­L. Liu, W. T. Fuhrman, J. Paglione, N. P. Butch, S. M. Anlage,
Anomalous normal fluid response in a chiral superconductor UTe2. Nat. Commun. 12,
2644 (2021).
26. K. Ishihara, M. Roppongi, M. Kobayashi, K. Imamura, Y. Mizukami, H. Sakai, P. Opletal,
Y. Tokiwa, Y. Haga, K. Hashimoto, T. Shibauchi, Chiral superconductivity in UTe2 probed by
anisotropic low-­energy excitations. Nat. Commun. 14, 2966 (2023).
27. S. Kittaka, Y. Shimizu, T. Sakakibara, A. Nakamura, D. Li, Y. Homma, F. Honda, D. Aoki,
K. Machida, Orientation of point nodes and nonunitary triplet pairing tuned by the
easy-­axis magnetization in UTe2. Phys. Rev. Res. 2, 032014 (2020).
28. V. P. Mineev, Low temperature specific heat and thermal conductivity in superconducting
UTe2. J. Phys. Soc. Jpn. 91, 074601 (2022).
29. J. Ishizuka, S. Sumita, A. Daido, Y. Yanase, Insulator-­metal transition and topological
superconductivity in UTe2 from a first-­principles calculation. Phys. Rev. Lett. 123, 217001 (2019).
30. D. Aoki, H. Sakai, P. Opletal, Y. Tokiwa, J. Ishizuka, Y. Yanase, H. Harima, A. Nakamura, D. Li,
Y. Homma, Y. Shimizu, G. Knebel, J. Flouquet, Y. Haga, First observation of the de
Haas–van Alphen effect and Fermi surfaces in the unconventional superconductor UTe2.
J. Phys. Soc. Jpn. 91, 083704 (2022).
31. H. Sakai, P. Opletal, Y. Tokiwa, E. Yamamoto, Y. Tokunaga, S. Kambe, Y. Haga, Single crystal
growth of superconducting UTe2 by molten salt flux method. Phys. Rev. Mater. 6, 073401 (2022).
32. Y. Matsuda, K. Izawa, I. Vekhter, Nodal structure of unconventional superconductors
probed by angle resolved thermal transport measurements. J. Phys. Condens. Matter 18,
R705–R752 (2006).
33. K. Kadowaki, S. Woods, Universal relationship of the resistivity and specific heat in
heavy-­fermion compounds. Solid State Commun. 58, 507–509 (1986).
34. R. C. Yu, M. B. Salamon, J. P. Lu, W. C. Lee, Thermal conductivity of an untwinned
YBa2Cu3O7−δ single crystal and a new interpretation of the superconducting state
thermal transport. Phys. Rev. Lett. 69, 1431–1434 (1992).
35. K. Izawa, H. Yamaguchi, Y. Matsuda, H. Shishido, R. Settai, Y. Onuki, Angular position of
nodes in the superconducting gap of quasi-­2D heavy-­fermion superconductor CeCoIn5.
Phys. Rev. Lett. 87, 057002 (2001).
36. Y. Kasahara, T. Iwasawa, H. Shishido, T. Shibauchi, K. Behnia, Y. Haga, T. D. Matsuda,
Y. Onuki, M. Sigrist, Y. Matsuda, Exotic superconducting properties in the electron-­hole-­
compensated heavy-­fermion “semimetal” URu2Si2. Phys. Rev. Lett. 99, 116402 (2007).
37. L. Taillefer, B. Lussier, R. Gagnon, K. Behnia, H. Aubin, Universal heat conduction in
YBa2Cu3O69. Phys. Rev. Lett. 79, 483–486 (1997).
38. T. Yamashita, T. Takenaka, Y. Tokiwa, J. A. Wilcox, Y. Mizukami, D. Terazawa, Y. Kasahara,
S. Kittaka, T. Sakakibara, M. Konczykowski, S. Seiro, H. S. Jeevan, C. Geibel, C. Putzke,
T. Onishi, H. Ikeda, A. Carrington, T. Shibauchi, Y. Matsuda, Fully gapped
superconductivity with no sign change in the prototypical heavy-­fermion CeCu2Si2. Sci.
Adv. 3, e1601667 (2017).
39. J. Lowell, J. Sousa, Mixed-­state thermal conductivity of type II superconductors. J. Low
Temp. Phys. 3, 65–87 (1970).
40. S. Y. Zhou, H. Zhang, X. C. Hong, B. Y. Pan, X. Qiu, W. N. Dong, X. L. Li, S. Y. Li, Nodeless
superconductivity in Ca3lr4Sb13: Evidence from quasiparticle heat transport. Phys. Rev. B
86, 064504 (2012).
41. Z. Zhang, Y. Xu, C. Kuo, X. Hong, M. Wang, P. Cai, J. Dong, C. Lue, S. Li, Nodeless
superconducting gap in the caged-­type superconductors Y5Rh6Sn18 and Lu5Rh6Sn18.
Supercond. Sci. Technol. 28, 105008 (2015).
42. H. Suderow, J. Brison, A. Huxley, J. Flouquet, Thermal conductivity and gap structure of
the superconducting phases of UPt3. J. Low Temp. Phys. 108, 11–30 (1997).
43. G. Seyfarth, J. P. Brison, G. Knebel, D. Aoki, G. Lapertot, J. Flouquet, Multigap
superconductivity in the heavy-­fermion system CecoIn5. Phys. Rev. Lett. 101, 046401 (2008).
44. M. Franz, Mixed-­state quasiparticle transport in high-­​Tc cuprates. Phys. Rev. Lett. 82,
1760–1763 (1999).
45. Y. Kasahara, Y. Nakajima, K. Izawa, Y. Matsuda, K. Behnia, H. Shishido, R. Settai, Y. Onuki,
Anomalous quasiparticle transport in the superconducting state of CeCoIn5. Phys. Rev. B
72, 214515 (2005).
Suetsugu et al., Sci. Adv. 10, eadk3772 (2024) 7 February 2024
46. H. S. Røising, M. Geier, A. Kreisel, B. M. Andersen, Thermodynamic transitions and
topology of spin-­triplet superconductivity: Application to UTe2. arXiv:2311.06097 (2023).
47. S. Sundar, N. Azari, M. R. Goeks, S. Gheidi, M. Abedi, M. Yakovlev, S. R. Dunsiger,
J. M. Wilkinson, S. J. Blundell, T. E. Metz, I. M. Hayes, S. R. Saha, S. Lee, A. J. Woods,
R. Movshovich, S. M. Thomas, N. P. Butch, P. F. S. Rosa, J. Paglione, J. E. Sonier, Ubiquitous
spin freezing in the superconducting state of UTe2. Commun. Phys. 6, 24 (2023).
48. Y. Tokunaga, H. Sakai, S. Kambe, Y. Haga, Y. Tokiwa, P. Opletal, H. Fujibayashi, K. Kinjo,
S. Kitagawa, K. Ishida, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Slow
electronic dynamics in the paramagnetic state of UTe2. J. Physical Soc. Japan 91, 023707
(2022).
49. Y. Iguchi, H. Man, S. M. Thomas, F. Ronning, P. F. S. Rosa, K. A. Moler, Microscopic imaging
homogeneous and single phase superfluid density in UTe2. Phys. Rev. Lett. 130, 196003
(2023).
50. I. Fomin, J. Brison, Subdominant interactions and Hc2 in UBe13. J. Low Temp. Phys. 119,
627–643 (2000).
51. Y. Shimizu, D. Braithwaite, D. Aoki, B. Salce, J.-­P. Brison, Spin-­triplet p-­wave
superconductivity revealed under high pressure in UBe13. Phys. Rev. Lett. 122, 067001
(2019).
52. L. Miao, S. Liu, Y. Xu, E. C. Kotta, C.-­J. Kang, S. Ran, J. Paglione, G. Kotliar, N. P. Butch,
J. D. Denlinger, L. A. Wray, Low energy band structure and symmetries of UTe2 from
angle-­resolved photoemission spectroscopy. Phys. Rev. Lett. 124, 076401 (2020).
53. S.-­I. Fujimori, I. Kawasaki, Y. Takeda, H. Yamagami, A. Nakamura, Y. Homma, D. Aoki,
Electronic structure of UTe2 studied by photoelectron spectroscopy. J. Physical Soc. Japan
88, 103701 (2019).
54. Y. S. Eo, S. Liu, S. R. Saha, H. Kim, S. Ran, J. A. Horn, H. Hodovanets, J. Collini, T. Metz,
W. T. Fuhrman, A. H. Nevidomskyy, J. D. Denlinger, N. P. Butch, M. S. Fuhrer, L. A. Wray,
J. Paglione, c-­Axis transport in UTe2: Evidence of three-­dimensional conductivity
component. Phys. Rev. B 106, L060505 (2022).
55. A. Rosuel, C. Marcenat, G. Knebel, T. Klein, A. Pourret, N. Marquardt, Q. Niu, S. Rousseau,
A. Demuer, G. Seyfarth, G. Lapertot, D. Aoki, D. Braithwaite, J. Flouquet, J. P. Brison,
Field-­induced tuning of the pairing state in a superconductor. Phys. Rev. X 13, 011022 (2023).
56. K. Ishihara, M. Kobayashi, K. Imamura, M. Konczykowski, H. Sakai, P. Opletal, Y. Tokiwa,
Y. Haga, K. Hashimoto, T. Shibauchi, Anisotropic enhancement of lower critical field in
ultraclean crystals of spin-­triplet superconductor candidate UTe2. Phys. Rev. Res. 5,
L022002 (2023).
57. Q. Niu, G. Knebel, D. Braithwaite, D. Aoki, G. Lapertot, G. Seyfarth, J.-­P. Brison, J. Flouquet,
A. Pourret, Fermi-­surface instability in the heavy-­fermion superconductor UTe2. Phys. Rev.
Lett. 124, 086601 (2020).
58. H. Matsumura, H. Fujibayashi, K. Kinjo, S. Kitagawa, K. Ishida, Y. Tokunaga, H. Sakai,
S. Kambe, A. Nakamura, Y. Shimizu, Y. Homma, D. Li, F. Honda, D. Aoki, Large reduction in
the a-­axis Knight shift on UTe2 with Tc = 2.1 K. J. Phys. Soc. Jpn. 92, 063701 (2023).
59. Y. Wang, B. Revaz, A. Erb, A. Junod, Direct observation and anisotropy of the contribution
of gap nodes in the low-­temperature specific heat of YBa2Cu3O7. Phys. Rev. B 63, 094508
(2001).
60. O. J. Taylor, A. Carrington, J. A. Schlueter, Specific-­heat measurements of the gap
structure of the organic superconductors κ−(ET)2Cu[N(CN)2]Br and κ−(ET)2Cu(NCS)2.
Phys. Rev. Lett. 99, 057001 (2007).
Acknowledgments: We thank K. Ishihara, T. Shibauchi, and Y. Tokunaga for insightful
discussions. Funding: This work is supported by Grants-­in-­Aid for Scientific Research
(KAKENHI) (nos. 18H05227, 18H03680, 18H01180, 21 K13881, 21 K18145, 22H01181,
22H04933, 23H00089, 23H01132, 23H04871, 23 K03332, 23 K13060, and 23 K17353) and on
Innovative Areas “Quantum Liquid Crystals” (no. 19H05824) from the Japan Society for the
Promotion of Science, JST CREST (JPMJCR19T5), and PRESTO (no. JPMJPR2252; T.A.). M.L.
acknowledges the support by Japan Society of for the Promotion of Science via the visiting
scholar fellowship. Author contributions: H.S., P.O., Y.T., and Y.H. synthesized the high-­quality
single crystals. S.S., M.S., M.K., T.A., and Y.S. performed thermal conductivity experiments. S.S.,
M.S., M.K., Y.Ko., H.A., S.I., and T.A. conducted resistivity experiments. Y.Ka. performed specific
heat experiments. All authors analyzed the data and discussed the results. S.S. and Y.M.
prepared the manuscript with inputs from all authors. Competing interests: The authors
declare that they have no competing interests. Data and materials availability: All data
needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials.
Submitted 18 August 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk3772
6 of 6
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

OPTICS Copyright © 2024 The

Authors, some rights
Self-­assembled, disordered structural color from fruit reserved; exclusive
licensee American
wax bloom Association for the
Advancement of
Science. No claim to
Rox Middleton1,2*, Sverre Aarseth Tunstad1, Andre Knapp3, Sandra Winters1,4, Susan McCallum5, original U.S.
Heather Whitney1* Government Works.
Distributed under a
Many visually guided frugivores have eyes highly adapted for blue sensitivity, which makes it perhaps surprising Creative Commons
that blue pigmented fruits are not more common. However, some fruits are blue even though they do not contain Attribution
blue pigments. We investigate dark pigmented fruits with wax blooms, like blueberries, plums, and juniper cones, NonCommercial
and find that a structural color mechanism is responsible for their appearance. The chromatic blue-­ultraviolet License 4.0 (CC BY-­NC).
reflectance arises from the interaction of the randomly arranged nonspherical scatterers with light. We reproduce
the structural color in the laboratory by recrystallizing wax bloom, allowing it to self-­assemble to produce the blue
appearance. We demonstrate that blue fruits and structurally colored fruits are not constrained to those with blue
subcuticular structure or pigment. Further, convergent optical properties appear across a wide phylogenetic
range despite diverse morphologies. Epicuticular waxes are elements of the future bioengineering toolbox as
sustainable and biocompatible, self-­assembling, self-­cleaning, and self-­repairing optical biomaterials.

INTRODUCTION Blue-­reflective fruits might be less rare than currently documented,

Blueberries are observably blue; however, the pigments found in
blueberries are not. Recent work demonstrates explicitly that the
color variation of blueberries does not predominantly depend on
pigmentation (1). The anthocyanins, contained in high concentra-
tions in these fruits, generally have dark red scattering profiles. How
then do fruits like blueberries and sloes produce their blue color? In
many biological materials, nanostructures of nonabsorbing molecules
interact with light via wave interference to reflect bright, highly satu-
rated colors, creating “structural color” (2). Here, we analyze how
the nanostructure of epicuticular wax, which arises by self-­assembly
(3), produces the coloration of a diverse group of “bloom fruits.” We
find that the appearance is dominated by scattering from the random
assembly of nonspherical particles and, unlike the coherent inter-
ference found in other structurally colored fruits that rely on periodic
lengthscales within the material, is dominated by form factor contribu-
tion. We define bloom fruits as those with dark pigmentation and an
epicuticular wax crystal coating, referred to as “bloom.”
Having a high discriminatory sensitivity to blue light is extremely
common across vertebrates with color vision (4) and, in particular,
visually guided frugivores, making blue discriminable or “chromati-
cally salient” and therefore a valuable signaling color (5). However,
identification of blue fruits is rare; in a broad global-­scale study, blue
fruits constituted a group so small that they were classed as outliers
(6). This is in part explained by the relatively rare deployment of
blue biological pigments (7). Blue-­reflective pigments require absorp-
tion of lower energy photons by large, energetically expensive mole-
cules (8). Blue pigment-­derived color exists in plants, especially in
flowers (9), and in some fruits such as Dianella tasmanii (10). However,
these molecules are energetically costly, and structural color can offer
a less costly route to blue color. A handful of other known blue fruits
have been identified as structurally colored (see Fig. 1) (11–18).
1
University of Bristol, Bristol, UK. 2Technische Universität Dresden, Dresden, Germany.
3
Leibniz Polymer Institut, Dresden, Germany. 4University of Helsinki, Helsinki, Finland.
5
James Hutton Institute, Dundee, UK.
*Corresponding author. Email: heather.​whitney@​bristol.​ac.​uk (H.W.); r.​middleton@​
bristol.​ac.​uk (R.M.)
as bloom fruits have been overlooked in color surveys, perhaps because
their strong anthocyanin pigmentation develops with ripeness (19)
but, reflecting red/black, does not determine their blue ultraviolet
(UV) color appearance (1, 20).
Plant epicuticular waxes are self-­ordering and self-­repairing (21)
surface coatings that can determine properties of the plant surface
such as self-­cleaning, wettability (22), and resistance to insect and
microbial pathogens. Plant waxes share with biomaterials like cellu-
lose a very wide and multifaceted range in biological function (21).
An extensive body of work on epicuticular wax crystals exists. It is
largely focused on hydrophobicity properties but has identified epicu-
ticular wax crystals as developing external to the plant cuticle
through self-­assembly from crystallization (23, 24), which can be
reproduced in many (but so far not all) instances in the laboratory
(3), including complex coiled structures.
The optics of epicuticular waxes has been the focus of surprisingly
little analysis. An exception is a demonstration of thin-­film interference
producing a gold sheen on Tradescantia leaves (25). High UV re-
flectance in Picea pungens was reported in 1975 (26) along with
the reflectance spectrum and later the UV reflectance from other
leaves (27). Although the P. pungens mechanism was not discussed by
the authors of that paper, it has since been generally attributed to
Rayleigh scattering (28). Rayleigh, Tyndall, and Mie scatterings refer
to single-­particle scattering effects, with different degrees of precision
(29). While all nonmolecular, non-­iridescent color effects were once
generally attributed to this mechanism (30, 31), the imprecision of
its application in dense biomaterials [in comparison to dilute suspen-
sions and atmospheres, where it is well understood (32)] has meant
that many of these assumptions have been overturned (33). Despite
the general assumption, the color of bloom fruits was not previously
listed prominently in biomaterials attributed to Rayleigh scattering.
Mie scattering, the analytical result for single scattering from
spheres and infinite cylinders 4(34), has attracted particular interest,
in part because of its analytic solution but also because of its application
in producing vibrant spherical-­colloid colorant materials (35). Mie
resonance in spherical particles produces coloration corresponding
to particle diameter, allowing peaks across the optical spectrum.

Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024 1 of 9

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Structural color in wax bloom produces blue appearance on fruits across a wide phylogeny. (A) Undamaged highbush blueberries growing on the plant.
(B) Blueberry with (i) unmodified wax, (ii) mechanical wax removal, (iii) chloroform wax removal, (iv) surface application of (almost) index-­matching oil n = 1.518, (v)
surface application of water, (vi) outer skin peeled to reveal flesh, and (vii) underside of peeled skin. (C) Transmission optical microscopy of peeled blueberry skin from
internal edge showing red pigmentation in epidermal cells. Scale bar, 200 μm. (D) (i and ii) A selection of plum (Prunus domestica) fruits with different cell pigmentation,
(i) with wax intact and (ii) with wax removed. (E) Confirmed occurrence of bloom disseminules across phylogeny. Previously known structurally colored (multilayer)
fruits are shown in dark blue: Delarbrea michieana (11), Elaeocarpus angustifolia (12), Pollia species (13, 14), Margaritaria nobilis (15), Viburnum species (16, 17), and
Lantana strigocamara (18).

Theoretical analysis of Mie scattering has investigated the transition
between single-­particle scattering and assemblies with resonance or
multiple scattering, as dense thick assemblies transition to sparse
thin ones (36).
Study of Mie assemblies as disordered photonic glasses have shown
that the optical response can be broken down into a “form factor”
contribution from individual particles, and a “structure factor” contri-
bution from their spatial relationships to one another. Although
negligible in photonic crystals, a form factor contribution appears in
photonic glasses due to departure from strictly periodic scattering
centers. Analysis of this has shown that form factor, as a minority
contribution to overall reflectance, limits purity of non-­blue colors
(37). It is known that although multiple scattering is important, in
the low-­scattering limit of very thin samples, the spectral response
from individual particles can qualitatively predict spectral shape
(38). Analyzing the interaction of light with epicuticular wax expands
our understanding of the materials, advancing their potential for
application. Biomaterials have huge value in sustainable and biologi-
cally compatible bioinspired engineering applications. For example,
research into the optical and self-­assembly properties of cellulose
has led to an explosion in innovation in coatings, sensors, and colo-
rants (39). Epicuticular waxes show particular promise due to the
wide range of morphologies (40) the self-­assembled crystals take, as
well as the range of biological adaptation in environment and function.
RESULTS
Epicuticular waxes produce structural color across a wide
phylogenetic range
Highbush blueberry (Vaccinium corymbosum) fruit is used in Fig. 1
(A to C) to demonstrate the epicuticular wax bloom characteristics.
The fruit appears blue in trichromat human vision (Fig. 1, A and
Bi), but the color is removed by mechanical abrasion of the surface
(Fig. 1Bii), rinsing with chloroform (Fig. 1Biii), and application of
immersion oil (n = 1.5) (Fig. 1Biv), while wetting with water leaves
it unaffected (Fig. 1Bv). The pigment in the epidermal cells is very
dark (Fig. 1Bvi) but, when released by cell breakage as juice, appears
red over the whitish flesh (Fig. 1Bvii). Transmission optical micros-
copy under very high intensity light confirms the red color (Fig. 1C)
of the epidermal cell pigment, although its strong absorption makes
the macroscopic appearance of ripe dewaxed blueberry surface
black. The dissolved epicuticular wax is transparent in (chloroform)
solution over the visible range. Spectrophotometry shows a strong
absorption in UV below 300 nm for blueberry wax, absent in other

Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024 2 of 9

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

species (fig. S1), which accounts for a sub–300 nm reflectance drop-­
off observed only in blueberries. Color arising from structure is
nonabsorptive and, therefore, semitransparent. Wax bloom, in com-
mon with other structural color, therefore requires an underlying
absorptive background to be the sole source of coloration. Reflectance
from dewaxed surfaces and exposed pigmented skin is shown in
fig. S2. It also produces mixed colors in conjunction with non-­black
pigmentation, as seen in a range of plum cultivars in Fig. 1D. In
many such cases, wax bloom still contributes substantially to the
overall color. This is important both in fruits, in which strong color
mixing occurs, and in the transition of appearance of unripe fruits,
which generally develop from mixed green-­UV through red-­UV to
blue-­UV during maturation (fig. S3). We identified these traits of
bloom fruits in fruits (and in conifers, cones) from species across
seven phylogenetic orders, which are indicated in Fig. 1E.
For each species, we examined the morphology of the epicuticular
wax crystals by scanning electron microscopy (SEM), for which a
range of species are shown in Fig. 2. Epicuticular wax crystal mor-
phologies vary widely between genera according to chemical compo-
sition and are commonly categorized by shape (41, 42). Here, we
used four shape categories (“ring,” “rod,” “slab,” and “tube”) to capture
the characteristic dimensions plotted in fig. S4. Figure 2 (M to P)
shows surface optical reflectance spectra grouped by wax morphology.
Despite the major differences in the particle shape, the spectra share
a monotonic inverse relationship between reflectance intensity and
wavelength. The same or similar optical spectra are produced by
waxes from diverse families of wax morphologies. The phenomenon
shows apparent convergence in spectral reflectance across the phylo-
genetic range, despite morphological variation in the wax particles.
Epicuticular wax spectra are dominated by random
scattering from particles rather than periodic or
quasi-­periodic structure.
The comparison of bloom reflectance spectra with other spectral
signatures, as in Fig. 3A, shows a clear distinction between bloom
and other characteristic reflectance spectra. Other kinds of biological
pigmentary color, and structural color (both ordered and disordered),
show a peak in the visible wavelength range (corresponding in general
to the dominant visual appearance). Contrastingly, bloom spectra,
in common with the spectrum of the blue sky (produced by Rayleigh
scattering), have a monotonic inverse relationship between wave-
length and reflectance (29).
Spectral peaks from structural color arise from constructive inter-
ference of light waves with nanostructure arrays that have a charac-
teristic lengthscale corresponding to the reflected peak wavelength.
In photonic crystals, this characteristic lengthscale occurs on crystal
planes and produces angle-­dependent coloration (2). In photonic
glasses, directional disorder occurs between scattering loci, but a

Fig. 2. Wax morphologies differ strongly between species, but spectral signatures do not. (A to L) SEM images of wax bloom crystals; scale bar, 2 μm. Photo insets:
(A) Damson P. domestica. (B) Sloe, Prunus spinosa. (C) Victoria plum P. domestica, ssp. Intermedia. (D) Grape Vitis vinifera “Hamburg.” (E) Flowering currant, Ribes sanguineum
Pursh. (F) Bilberry Vaccinium myrtillus. (G) M. aquifolium. (H) Barberry, Berberis darwinii. (I) Juniperus occidentalis. (J) Juniperus virginiana. (K) Abies koreana. (L) Blueberry
V. corymbosum “Rocio.” (M to P) Average spectra from each species by morphology group. (M) Ring spectra. (N) Slab spectra. (O) Rod spectra. (P) Tube/blueberry spectra.

Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024 3 of 9

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Epicuticular wax shows no visual-­wavelength peak but is visually salient in both blue and UV channels for avian frugivores, indicating potential value in
chromatic contrast. (A) Normalized reflectances compare spectral shape. Structural color [highly ordered Pollia, quasi-­ordered Cotinga cotinga (28)] and pigment (phy-
cocyanin) compared to sloe bloom and irradiance measured from blue sky. Sunlight at ground level explains the blue-­sky cutoff below 300 nm. (B) (i to iv) All spectra
(solid lines) and 90% confidence intervals (shaded areas) are shown along with the single-­particle scattering model (thicker dotted/dashed lines) from FDTD scaled to the
peak intensity measured. In (iv), the blueberry model also includes the absorption below 300 nm measured in blueberry wax, as shown in fig. S1. The scattering profiles
from FDTD are scaled to match the intensity of the measured spectra, by slab: 3 × 1014; ring: 5 × 1014; rod: 3 × 1015; blueberry: 3 × 1014. Further variation of model parameters
is in fig. S6. (C) (i and ii) Projection of the bloom fruits with (i) wax intact and (ii) removed, into tetrachromat Turdus merula (Blackbird) visual space. (D) (i and ii) Projection
of the bloom fruits with (i) wax intact and (ii) removed, into trichromat human CIE color space. Outlines show the range in the other plot. (E) (i and ii) Color from calibrated
photos of five fruit surfaces are plotted in CIE space. (i) Bound of points from the photographed fruit body. (ii) Each fruit’s average. The range contains the palest blue
reflectance point, plotted from specular reflectance in (Di), and also shows the resultant full visible color range. (F) Chromatic and achromatic just noticeable differences
(JNDs) (62) show the discriminability of fruits with wax intact (bloom-­leaf, shown in blue) and wax removed (pigment-­leaf, shown in green) against green foliage. One JND
(discriminable in bright lighting) and three JND (discriminable across lighting conditions) thresholds are indicated with gray dashed lines.

Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024 4 of 9

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
characteristic inter-­locus distance is preserved, and thus, construc-
tive interference and a reflectance peak is also preserved but without
angle dependence (43). The analysis of disordered photonic glasses
(37) has shown that the disorder introduces additional scattering,
designated form factor. This depends on individual scatterers, rather
than their relationship to each other, which constitutes the structure
factor. Form factor corresponds to the scattering of light by individual
nonspherical particles of sub-­or near-­wavelength size. In dilute sys-
tems, single-­particle scattering has been identified since the late
19th century as “Rayleigh” scattering or as a “Tyndall” phenomenon
in which a monotonic relationship between wavelength and scattering
is observed, where shorter wavelengths are scattered more (29). As
the analog in dense biomaterials, single-­particle scattering as form
factor contributions in photonic glasses also has this bias, with high
relative reflectance in blue wavelengths therefore limiting the non-­
mixed color palette of photonic glasses to blue (37). Other colors of
photonic glasses may be produced, but where disorder makes it a
“glass” rather than a crystal, the color is usually mixed, appearing
light or nonsaturated rather than saturated (44).
The lack of a reflectance peak in bloom reflectance spectra is no-
table because it indicates a lack of constructive interference in the
material, in turn a sign of no characteristic lengthscale order (2). We
made Fourier transforms and structure measurements from the EM
measurements of bloom wax in cross section and normal to the surface.
These indeed showed no sign of short-­or long-­range order. Wax
crystals have different uniform stereotypical shapes, distributions,
sizes, and thicknesses depending on species (fig. S4), but none show
evidence of characteristic spacing between adjacent particles.
To understand the case in which coherent interference could never-
theless explain the result (i.e., to construct a plausible structure factor
model depending on packing of the regular particle dimensions), we
constructed a one-­dimensional (1D) model of multiple scattering from
very tight, disordered packing of ring particles. In this model, a peak
in the reflectance spectrum could occur at sufficiently low wavelength
that is not measured, and the measured spectral tail is interpreted as
a monotonic spectrum. With assumption of very close packing and
actual dimensions smaller than measured (i.e., assuming strong pres-
ence of artifacts in measurement), the closest fit model shows a loose
approximation to the observed spectra for ring particles (fig. S5),
although it has a flattened profile at low wavelengths, which is not
observed. Furthermore, such packing would only be feasible with some
(high symmetry) crystal shapes, despite experimentally similar mea-
sured optical responses between different bloom wax crystal morpho
logies. We used the same approach to confirm that the conformation of
thin-­film fringes predicted by the parameters measured do not provide
adequate explanation of the effect (also fig. S5).
Since we could identify no contribution from a periodic structure
factor of the highly disordered nanostructure, we modeled the spectra
as arising from form factor contributions, as modeled in previous low-­
scattering approximation (38). We used a Monte Carlo approach, aver-
aging over many single-­particle scattering events, modeled using
computational Maxwell equation solver [finite-­difference time-­domain
(FDTD)]. This approximation requires the assumption of small scatter-
ing cross section and thin layer thickness so that the impact of scattering
events on the EM field is negligible with respect to the incident wave.
For the low-­multiple scattering regime, the structure thickness
must be smaller than or similar to the transport length or distance
traveled by light through the material without scattering. This assump-
tion is suggested by previous work on a similar system in which
Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024
multiple-­scattering models break down for thin films with thickness
less than the transport length of the material (36). In the work
referenced, a dense Mie particle assembly of polystyrene spheres
between 94 and 138 nm in diameter is analyzed. The transport
length can be calculated analytically for spherical particles and was
found in that work to be ~8 μm. Reflectance from a film thickness of
6 μm was found to depart from multiple-­scattering assumptions.
Taking this transport length as a starting point, polystyrene has a
refractive index of around 1.6 (slightly larger than the refractive index
for wax assumed here) (45). The epicuticular wax shapes also have a
far reduced geometric cross section with respect to a filled sphere.
Both these features serve to reduce effective volume of interaction
between particle and light, increasing the transport length (34). Last,
the thickness of the material present on bloom surfaces is found to
be 2 to 4 μm, larger only in sparse rod structures. We therefore expect
the departure from a strictly multiple-­scattering model to be appro-
priate in epicuticular wax nanostructure, until a much greater thick-
ness is observed. Instead, as in (38), we used a scattering model in
this instance, which analyzed individual particles (whereas in the
previous work this was spheres, our particles are nonspherical). The
models are shown with the experimental data in Fig. 3B (i to iv).
Despite its limitations, the single-­particle model approximates the
measured spectra qualitatively and supports the dominant role for
form factor contribution in the scattering profile, analogously to the
successful model of spectral shape through spherical form factor
models (38) and the blue-­bias minor contribution identified in pho-
tonic glass arising from form factor (37). Although there is no spectral
peak in the visual wavelength range, the activation of visual channels
in UV and in blue, and its differential with longer wavelength re-
ceptors, is sufficient to produce chromatic salience both in a tetra-
chromatic avian (blackbird) visual model (Fig. 3C, i and ii) and in
human trichromatic vision, plotted in CIE space (Fig. 3D, i and ii).
The projection of the specular reflectance shows that the fruits are
all blue when waxy and dark/red when dewaxed. The color appears
desaturated when projected from spectral profile. For dewaxed
fruits, this is because they were very dark (low reflectance in all
wavelengths); however, the relatively low saturation for waxed fruits
is due to the lack of a peak in the spectrum. Nevertheless, under
natural lighting conditions, the fruits appear to human trichromat
vision relatively saturated. Color spaces appearing across fruit sur-
faces extracted from color-­calibrated photos and plotted into CIE
space in Fig. 3E (i and ii) show that the visual appearance actually
encompasses a range across the unevenly lit surface, with both more
tones and the unsaturated tones (in Fig. 3C, i and ii) measured by
specular optical-­axis reflectance.
Fruit coloration visibility is strongly affected by background color.
The contrast of the fruits with respect to green foliage in both lumi-
nance and chromaticity (for the specular spectra) is therefore plotted
in Fig. 3F, comparing fruits with wax bloom and after wax is removed
to reveal underlying pigment. Although waxy fruits have lower lumi-
nance discriminability (the luminance of a waxy fruit is closer to a
leaf than a dewaxed one), the chromatic discriminability is strongly
enhanced (the color contrast of bloom fruits with green leaves is
highly discriminable and much higher than dewaxed fruits).
Replication of structural color from fruit bloom
We extracted epicuticular wax from Mahonia aquifolium fruits
(Fig. 4A) by fast dipping in chloroform to extract only the outer
(epicuticular) layer of wax (46). We recrystallized the wax via
5 of 9
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

thermal evaporation and deposition (24) under vacuum onto non-­ The identification of visually salient blue-­UV structural color from
chromatic black card (Fig. 4B). The wax, having been transparent in wax bloom demonstrates structural color in a much wider range of
solution (fig. S1A) and white as a solute remaining after solvent fruits than just those with subcuticular structure. The phylogenetically
evaporation, self-­assembled after deposition to replicate a similar widespread occurrence indicates that it has evolved multiple times,
nanostructure (Fig. 4, C and D) and corresponding blue coloration likely many more than reported here. Its identification in part resolves
(Fig. 4F) close to the spectral measurement from naturally occurring the paradoxical situation of very few fruits seemingly using effective
M. aquifolium fruit surfaces (Fig. 4E). The production of blue color blue signaling colors, as structurally colored bloom overlaying dark
by nanostructure is therefore demonstrated. Increased deposition anthocyanin pigmentation to produce blue-­UV chromatic reflectance
time and weight showed a corresponding increase in the reflectance is widespread. The spectral similarity indicates potential convergence,
brightness of the deposited coating. despite diverse structures. This mirrors the convergence of appearance
observed in petal striations to produce a blue sheen in flowers (48).
We see therefore a second example of structural color in plants showing
DISCUSSION broad taxonomic stability, despite disparate structures. In both cases,
This work adds the widespread matte blue–UV chromaticity of bloom a high degree of disorder plays a key role in the color produced and
fruit to the known photonic effects that occur naturally in epicuticular robustness to differences in that structure within a bounded range.
wax. This includes the gold sheen on Tradescantia pallida leaves (25), The clear relevance in both cases to plant-­animal interaction through
whitish “glaucous” wax on succulent leaves (47), and highly UV-­ visually guided foraging is also notable.
reflective coating on P. pungens needles (26). The known range shows The use of structural color here overcomes the challenge that the
already that wax nanostructure has a versatile range of photonic effects. dense, nutritionally valuable anthocyanins make fruits dark and there-
The reproduction of the effect here [through already established fore potentially inconspicuous (49). As observed also in Viburnum
techniques (3)] is a promising indication that other wax optics tinus (16), once overlaid with bloom, darker anthocyanins enhance
might also be accessible through recrystallization, and through the chromaticity of the fruit by reducing color mixing. This is also
adapting morphology and orientation of the structures, as in other important in maturation (fig. S3): Bloom is present throughout fruit
photonic biomaterials. The prominence of the effect reported here development, displaying a bright UV signal, but only once the bright
indicates that there may be much more to understand about the pigmentary scattering of early ripening stages is removed is the final
range of optical phenomena in wax nanostructures. Further, the blue-­UV mixed color, due entirely to the bloom structure, visible.
analysis indicating the importance of form factor in the effect pro- Given the other functional benefits of bloom for plant health (40),
duces a lens on both the discriminate fine-­tuning of structures that there is also the possibility that its chromatic visibility could act as
has produced the convergent coloration in these instances and an honest signal, displaying an innate high-­quality trait. The pheno
notable flexibility in morphological groups in which it has arisen. menon is observed in unrelated cultivated and wild fruits, indicating

Fig. 4. Reproduction of structurally colored wax by in vitro recrystallization. (A and B) Photographs of (A) M. aquifolium fruits and (B) redeposited M. aquifolium wax
(left 1.75 mg, right 2.02 mg, both deposited over 1 hour at 120°C). (C and D) SEM images showing the crystal structure of epicuticular wax in (C) fruit surface and (D) rede-
posited sample. (E and F) Spectral measurements from (E) fruit surface and (F) redeposited samples; legends show initial wax weight, crucible temperature, and deposi-
tion duration.

Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024 6 of 9

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
convergent evolution in a signaling organ, but given the complexity
of fruit ecology, behavioral tests would be required to understand
whether the coloration enhances frugivore attraction.
Behavioral tests of the impact of bloom on bird attraction have
reported reflectance spectra of some species included here, although
the aim was not to understand its origin. Moreover, the studies focused
specifically on UV reflectance. From this work, it is clear that the
UV reflectance is visually salient to some birds, but in short-­range
“cafeteria experiments,” it is not clear whether UV itself enhances
attraction to birds (50, 51). The sloped monotonic spectral signature
allows the coloration to be salient in both UV and blue visual channels.
Crucially, it also features lower stimulation of longer-­wavelength visual
receptors, making it “colored” for both tetrachromats and trichromats,
rather than white. Although these colors could in some cases be
added by bright underlying pigments, in bloom fruits, the underlying
pigment is dark and absorptive.
The proposed mechanism extends the role of form factor contribu-
tion in structural color, from prominent as described in photonic
glasses (37), to primary in its role in these (2 to 6 μm) coatings
(fig. S7). Although thin, this thickness is of the same order as structural
color with coherent interference like multilayers (2), which justifies
the analysis of its structure factor. This mechanism is analogous to
resonant Mie scattering in spherical particles (52), the fabrication of
which is of contemporary interest but shows the prominent role of
scattering from nonspherical particles in dense materials.
The very simple optical model suggested here is successful, and
we propose that, analogously to dominance in disordered photonic
glass of single scattering from Mie spheres, it produces insight into
color production in these ultrathin materials. However, there is much
more to be understood about the materials and the models, which
should form the basis of future work, and a full theoretical investiga-
tion into the range of architectures observed to produce the effect.
The most important question is in the role of multiple scattering.
There must be a role for multiple scattering in these materials, as the
density of the scatterers implies interaction between them. More-
over, as nicely demonstrated in the Ewald sphere construction (53),
transport length decreases with decreasing wavelength, making
multiple scattering increasingly likely. The role of “imperfect shapes”
and highly disordered nearest-­neighbor relationships we suggest are
particularly of interest. Given the spectral profiles identified and
previous work on multiple scattering (54), we would expect contri-
bution from multiple scattering also to be entirely disordered. Un-
like in a photonic glass, where the disorder has a visible-­lengthscale
average coherence length, because of the highly irregular shapes, we
expect the disorder in this case to produce greater scattering with
shorter wavelengths far into the UV. Therefore, wax bloom is an ap-
parently unusual example in biological coloration beyond white-
ness, where truly random particulate scattering dominates, although
its pervasiveness indicates that there may be other unexamined ex-
amples of the effect in biological materials. The lack of periodic
structure factor contribution could suggest enhanced disorder of the
material, as in Cyphochilus beetles, where the disorder is hypothesized
to be enhanced to ensure that no coherent interference effects are
present (54).
The simple reproduction of such self-­assembled nonspherical
particles in wax crystals is a direct benefit of working with the bio-
material itself, and a potential that could also be extended to other
nanoparticle fabrication. Recrystallization of epicuticular waxes through
thermal deposition in vacuum is a well-­established technique that has
Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024
been used to create complex hierarchical architecture (24). Although
vacuum thermal evaporation does not replicate natural biological
crystallization conditions, it allows wax molecules to reassemble
without obstruction by additional forces. Understanding how this
can be achieved outside of vacuum conditions for different waxes, as
occurs on biological surfaces, should be the subject of further work.
We present here an instance of structural color from wax, which
indicates the potential for a wider range of photonic effects arising
from epicuticular wax (for example, in albedo-­enhancing white
coatings). Recrystallization of the structurally colored bloom directly
demonstrates the bio-­replicative engineering application. Both dis-
ordered photonic glasses (55) and single-­particle spherical Mie scatter-
ing materials (44) have been produced in artificial materials, and the
scalable production of these is the focus of ongoing research (56).
Biologically derived structural color offers a sustainable, biocompatible,
multifunctional approach to self-­assembled, self-­healing, and self-­
cleaning colorants and coatings. Epicuticular wax, in particular, has
multiple known functionalities (40), making its applications poten-
tially tunable to different optical/industrial applications. The straight-
forward self-­assembly (57) and range and tunability of wax nano
structure (42), which have allowed for the coloration both observed
and reproduced here, indicate a rich potential for application in
nonabsorptive photonic surfaces.
MATERIALS AND METHODS
Materials
We identified and collected samples from botanic (Edinburgh,
Dresden) gardens and private and community gardens in Bristol
(UK), public access land, and the James Hutton Institute polytunnels
or commercially. Shop-­bought samples were compared to samples
grown in known conditions, and no artifacts from commercial sur-
face treatments were observed. We removed wax from the surface of
the fruits using either gentle mechanical abrasion with paper tissue
or wiping with tissue soaked in chloroform. Mechanical rubbing
produced a smooth, specular reflecting surface, while chloroform
dissolution produced a matte surface. This difference is because
mechanical rubbing removes the relevant particulate epicuticular
wax while leaving intact underlying smooth wax that is fused with
intracuticular wax that permeates the cuticle. The surface produced
is a polished artifact of rubbing. Using chloroform strips the relevant
epicuticular wax back and may remove part of the outer layer of intra
cuticular wax as well. This is an important concern and is fully ex-
plored in previous literature (46). This, too, produces an “artificially
textured” surface but not a smooth polished one as with rubbing.
This affected the dewaxed spectra as shown in fig. S2.
Imaging
We took macroscale photographs using a Sony a300 DSLR and Sony
SAL30M28 macrolens and x-­rite ColorChecker for fruit image cali-
bration (processed using MATLAB Image processing toolbox). For
spectral measurement, we used a DH-­2000-­BAL OceanOptics lamp,
QR400-­7-­UV-­Vis double-­ended fiber, and an Avantes Flame UV-­Vis
Spectrometer. We calibrated it on an Avantes white diffuser standard.
We took reflectance measurements using a metal mount-­ block
holding the fiber normal to the surface (tangent to the fruit surface
in the case of round fruits) at 7-­mm distance. Spectra shown in
Fig. 2 are averages across samples, smoothed over λ = 7.5 nm. Im-
mersion oil was Zeiss Immersol 518. We prepared samples for SEM
7 of 9
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
by mounting them on carbon tape and coating them with 5-­nm gold
in a Quorum Q150RES sputter coater; then, we imaged them in
Zeiss Evo 15 ESEM under vacuum (standard settings: 10 kV, 50 pA).
Visual modeling
For visual modeling, we used pavo version 2 (58) in R version 4.1.2.
We used visual models (4) to convert spectral reflectance measure-
ments into quantal catches to analyze colors as perceived by trichro-
matic humans and tetrachromatic avian seed dispersers (represented
by the European blackbird Turdus merula). We implemented a D65
illuminant to model daylight conditions in the Von Kries model. We
averaged spectral reflectance measurements for each species and
smoothed them using a span of 8 nm. We modeled Blackbird vision
(using vismodel) based on five photoreceptors [peak sensitivities:
557, 504, 454, 373, and 557 nm; cone cutoff wavelengths: 570, 515,
414, 330, and 439 nm; oil droplet types: R, Y, C, T, and P; cone types:
long (LW), medium (MW), short (SW), and UV cones, which perceive
color, and double cones, which perceive luminance] (59, 60).
To determine the detectability of fruit against the background, we
calculated just noticeable differences (JNDs). One JND is the discrimi-
nation threshold between similar colors in ideal conditions, calcu-
lated using the receptor noise limited model (here, JNDs are calculated
pairwise between each fruit with wax both on and off and an average
blueberry leaf spectrum (a characteristic chlorophyll-­dominated
green leaf), using relevant cone ratios; blackbird: 1:1.78:2.21:1.96
(UV:S:M:L) and Weber fraction [blackbird: w = 0.1 (58)]. The averaged
blueberry leaf spectrum is used as an approximate model of back-
ground foliage for all fruits.
Optical modeling
We made optical models using Lumerical 2019b Finite Difference
IDE, a commercial Maxwell equation solver. We estimated the refrac-
tive index of the particles as 1.47 (61). This is a low-­precision assump-
tion, and future work should aim to improve precision on the
dispersive refractive index. Absorption of several waxes was measured
(see fig. S1A) and found to be extremely low, except for blueberry
waxes below 350 nm. They were therefore treated as negligible and
complex refractive indices were not used in the FDTD model, al-
though in the blueberry model they were applied afterward, as shown
in fig. S1B. We performed simulations with a source that averaged
over two orthogonally polarized pulsed plane wave sources (160 to
600 nm) and the far-­field scattering integrated over the 180° back-
scatter direction. No planes of symmetry were introduced, and simu-
lation area boundary conditions were perfectly matched layer (PML)
absorbing boundaries. We used individual 3D particles oriented
within the simulation volume with varied orientation of the target
particle with respect to the k-­vector according to orientations observed
in structural analysis and then averaged the full backscattered spectra.
Wax extraction and redeposition
We extracted plant wax dominated by the fraction of epicuticular
wax from fruits by fast (~2 s) dipping in chloroform (46). We filtered
the solution (Fisherbrand grade 100 filter paper) and evaporated it
to leave a dry white residue under a gentle stream of nitrogen. We
loaded the 2 mg of the wax into an aluminum crucible in an adapted
Balzers MED-­010 thermal evaporator and decreased the pressure to
5 × 10−5 mbar. The crucible was heated over 2 min and held at 120°C
for 30 min or 1 hour under vacuum. The 1 cm–by–2 cm flat substrate
was held face-­down 2.5 cm above the crucible. After evaporation, we
Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024
released the chamber quickly (over approximately 10 s) to room
temperature and pressure. On removal from the device, the coating
is immediately visible, indicating either crystallization in the device
under vacuum or relatively quickly during the fast repressurizing of
the device on opening.
Supplementary Materials
This PDF file includes:
Figs. S1 to S7
References
REFERENCES AND NOTES
1. Y. Yan, M. Dossett, S. D. Castellarin, Cuticular waxes affect fruit surface color in blueberries.
Plants People Planet 5, 736–751 (2023).
2. Kinoshita S., Structural Colors in the Realm of Nature (World Scientific Publishing Co.,
2008), pp. 076401–076431.
3. I. Meusel, C. Neinhuis, C. Markstädter, W. Barthlott, Chemical composition and recrystallization
of epicuticular waxes: Coiled rodlets and tubules. Plant Biol. 2, 462–470 (2000).
4. A. Kelber, M. Vorobyev, Animal colour vision—Behavioural tests and physiological
concepts. Biol. Rev. Camb. Philos. Soc. 78, 81–118 (2003).
5. D. Osorio, M. Vorobyev, Colour vision as an adaptation to frugivory in primates. Proc. Biol.
Sci. 263, 593–599 (1996).
6. K. E. Stournaras, E. Lo, K. Böhning-­Gaese, E. Cazetta, D. M. Dehling, M. Schleuning,
M. C. Stoddard, M. J. Donoghue, R. O. Prum, H. Martin Schaefer, How colorful are fruits? Limited
color diversity in fleshy fruits on local and global scales. New Phytol. 198, 617–629 (2013).
7. P. R. Denish, J.-­A. Fenger, R. Powers, G. T. Sigurdson, L. Grisanti, K. G. Guggenheim,
S. Laporte, J. Li, T. Kondo, A. Magistrato, M. P. Moloney, M. Riley, M. Rusishvili,
N. Ahmadiani, S. Baroni, O. Dangles, M. Giusti, T. M. Collins, J. Didzbalis, K. Yoshida,
J. B. Siegel, R. J. Robbins, Discovery of a natural cyan blue: A unique food-­sourced
anthocyanin could replace synthetic brilliant blue. Sci. Adv. 7, eabe7871 (2021).
8. R. M. Christie, Why is indigo blue? Biotech. Histochem. 82, 51–56 (2007).
9. M. Shiono, N. Matsugaki, K. Takeda, Structure of the blue cornflower pigment. Nature
436, 791 (2005).
10. S. J. Bloor, Deep blue anthocyanins from blue Dianella berries. Phytochemistry 58,
923–927 (2001).
11. D. W. Lee, G. T. Taylor, A. K. Irvine, Structural fruit coloration in Delarbrea michieana
(Araliaceae). Int. J. Plant Sci. 161, 297–300 (2000).
12. D. W. Lee, Ultrastructural basis and function of iridescent blue colour of fruits in
Elaeocarpus. Nature 349, 260–262 (1991).
13. S. Vignolini, P. J. Rudall, A. V. Rowland, A. Reed, E. Moyroud, R. B. Faden, J. J. Baumberg,
B. J. Glover, U. Steiner, Pointillist structural color in Pollia fruit. Proc. Natl. Acad. Sci. U.S.A.
109, 15712–15715 (2012).
14. R. Middleton, E. Moyroud, P. J. Rudall, C. J. Prychid, M. Conejero, B. J. Glover, S. Vignolini,
Using structural colour to track length scale of cell-­wall layers in developing Pollia
japonica fruits. New Phytol. 230, 2327–2336 (2021).
15. S. Vignolini, T. Gregory, M. Kolle, A. Lethbridge, E. Moyroud, U. Steiner, B. J. Glover,
P. Vukusic, P. J. Rudall, Structural colour from helicoidal cell-­wall architecture in fruits of
Margaritaria nobilis. J. R. Soc. Interface 13, 20160645 (2016).
16. R. Middleton, M. Sinnott-­Armstrong, Y. Ogawa, G. Jacucci, E. Moyroud, P. J. Rudall,
C. Prychid, M. Conejero, B. J. Glover, M. J. Donoghue, S. Vignolini, Viburnum tinus fruits use
lipids to produce metallic blue structural color. Curr. Biol. 30, 3804–3810.e2 (2020).
17. M. A. Sinnott-­Armstrong, R. Middleton, Y. Ogawa, G. Jacucci, E. Moyroud, B. J. Glover,
P. J. Rudall, S. Vignolini, M. J. Donoghue, Multiple origins of lipid-­based structural colors contribute
to a gradient of fruit colors in Viburnum (Adoxaceae). New Phytol. 237, 643–655 (2023).
18. M. A. Sinnott-­Armstrong, Y. Ogawa, G. T. van de Kerkhof, S. Vignolini, S. D. Smith,
Convergent evolution of disordered lipidic structural colour in the fruits of Lantana
strigocamara (syn. L. camara hybrid cultivar). New Phytol. 235, 898–906 (2022).
19. S. W. Chung, D. J. Yu, H. J. Lee, Changes in anthocyanidin and anthocyanin pigments in
highbush blueberry (Vaccinium corymbosum cv. Bluecrop) fruits during ripening. Hortic.
Environ. Biotechnol. 57, 424–430 (2016).
20. R. A. Moyer, K. E. Hummer, C. E. Finn, B. Frei, R. E. Wrolstad, Anthocyanins, phenolics, and
antioxidant capacity in diverse small fruits: Vaccinium, Rubus, and Ribes. J. Agric. Food
Chem. 50, 519–525 (2002).
21. H. Bargel, K. Koch, Z. Cerman, C. Neinhuis, Structure-­function relationships of the plant
cuticle and cuticular waxes—A smart material? Funct. Plant Biol. 33, 893–910 (2006).
22. W. Barthlott, C. Neinhuis, Purity of the sacred lotus, or escape from contamination in
biological surfaces. Planta 202, 1–8 (1997).
23. H. J. Ensikat, M. Boese, W. Mader, W. Barthlott, K. Koch, Crystallinity of plant epicuticular
waxes: Electron and x-­ray diffraction studies. Chem. Phys. Lipids 144, 45–59 (2006).
8 of 9
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
24. B. Bhushan, Y. C. Jung, A. Niemietz, K. Koch, Lotus-­like biomimetic hierarchical structures
developed by the self-­assembly of tubular plant waxes. Langmuir 25, 1659–1666
(2009).
25. G. T. Van De Kerkhof, L. Schertel, R. N. Poon, G. Jacucci, B. J. Glover, S. Vignolini, Disordered wax
platelets on Tradescantia pallida leaves create golden shine. Faraday Discuss. 223, 207–215
(2020).
26. J. B. Clark, G. R. Lister, Photosynthetic action spectra of trees. Plant Physiol. 55, 407–413
(1975).
27. R. H. Grant, G. M. Heisler, W. Gao, M. Jenks, Ultraviolet leaf reflectance of common urban
trees and the prediction of reflectance from leaf surface characteristics. Agric. For.
Meteorol. 120, 127–139 (2003).
28. T. C. Vogelmann, Plant tissue optics. Annu. Rev. Plant. Physiol. Plant. Mol. Biol. 44, 231–251 (1993).
29. M. Kerker, Blue skies and the Tyndall effect. J. Chem. Educ. 48, 389 (1971).
30. J. R. Byers, Tyndall blue and surface white of tent caterpillars, Malacosoma SPP. J. Insect
Physiol. 21, 401–415 (1975).
31. J. Huxley, The coloration of Papilio zalmoxis and P. antimachus, and the discovery of
Tyndall blue in butterflies. Proc. R. Soc. B. 193, 441–453 (1976).
32. G. S. He, H.-­Y. Qin, Q. Zheng, Rayleigh, Mie, and Tyndall scatterings of polystyrene microspheres
in water: Wavelength, size, and angle dependences. J. Appl. Phys. 105, 023110 (2009).
33. R. O. Prum, Blue integumentary structural colours in dragonflies (Odonata) are not
produced by incoherent Tyndall scattering. J. Exp. Biol. 207, 3999–4009 (2004).
34. K. Baek, Y. Kim, S. Mohd-­Noor, J. K. Hyun, Mie resonant structural colors. ACS Appl. Mater.
Interfaces 12, 5300–5318 (2020).
35. T. Okazaki, H. Sugimoto, T. Hinamoto, M. Fujii, Color toning of mie resonant silicon
nanoparticle color inks. ACS Appl. Mater. Interfaces 13, 13613–13619 (2021).
36. V. Hwang, A. B. Stephenson, S. Barkley, S. Brandt, M. Xiao, J. Aizenberg, V. N. Manoharan,
Designing angle-­independent structural colors using Monte Carlo simulations of
multiple scattering. Proc. Natl. Acad. Sci. U.S.A. 118, e2015551118 (2021).
37. G. Jacucci, S. Vignolini, L. Schertel, The limitations of extending nature’s color palette in
correlated, disordered systems. Proc. Natl. Acad. Sci. U.S.A. 117, 23345–23349 (2020).
38. V. Hwang, A. B. Stephenson, S. Magkiriadou, J.-­G. Park, V. N. Manoharan, Effects of
multiple scattering on angle-­independent structural color in disordered colloidal
materials. Phys. Rev. E 101, 012614 (2020).
39. R. M. Parker, T. H. Zhao, B. Frka-­Petesic, S. Vignolini, Cellulose photonic pigments. Nat.
Commun. 13, 3378 (2022).
40. K. Koch, W. Barthlott, Plant epicuticular waxes: Chemistry, form, self-­assembly and
function. Nat. Prod. Commun. 1, 1067–1072 (2006).
41. K. Koch, A. Dommisse, A. Niemietz, W. Barthlott, K. Wandelt, Nanostructure of epicuticular
plant waxes: Self-­assembly of wax tubules. Surf. Sci. 603, 1961–1968 (2009).
42. W. Barthlott, C. Neinhuis, D. Cutler, F. Ditsch, I. Meusel, I. Theisen, H. Wilhelmi, Classification and
terminology of plant epicuticular waxes. Bot. J. Linn. Soc. 126, 237–260 (1998).
43. H. Noh, S. F. Liew, V. Saranathan, S. G. J. J. Mochrie, R. O. Prum, E. R. Dufresne, H. Cao, How
noniridescent colors are generated by quasi-­ordered structures of bird feathers. Adv.
Mater. 22, 2871–2880 (2010).
44. G. Shang, M. Eich, A. Petrov, Photonic glass based structural color. APL Photonics 5, 060901
(2020).
45. E. Marx, G. W. Mulholland, Size and refractive index determination of single polystyrene
spheres. J. Res. Natl. Bur. Stand. 88, 321–338 (1983).
46. R. Jetter, S. Schäffer, M. Riederer, Leaf cuticular waxes are arranged in chemically and
mechanically distinct layers: Evidence from Prunus laurocerasus L. Plant Cell Environ. 23,
619–628 (2000).
47. T. W. Mulroy, Spectral properties of heavily glaucous and non-­glaucous leaves of a
succulent rosette-­plant. Oecologia 38, 349–357 (1979).
48. E. Moyroud, T. Wenzel, R. Middleton, P. J. Rudall, H. Banks, A. Reed, G. Mellers, P. Killoran,
M. M. Westwood, U. Steiner, S. Vignolini, B. J. Glover, Disorder in convergent floral
nanostructures enhances signalling to bees. Nature 550, 469–474 (2017).
49. E. Cazetta, H. M. Schaefer, M. Galetti, Why are fruits colorful? The relative importance of
achromatic and chromatic contrasts for detection by birds. Evol. Ecol. 23, 233–244 (2009).
50. M. F. Willson, C. J. Whelan, Ultraviolet reflectance of fruits of vertebrate-­dispersed plants.
Oikos 55, 341 (1989).
Middleton et al., Sci. Adv. 10, eadk4219 (2024) 7 February 2024
51. H. Siitari, J. Honkavaara, J. Viitala, Ultraviolet reflection of berries attracts foraging birds. A
laboratory study with redwings (Turdus iliacus) and bilberries (Vaccinium myrtillus). Proc.
Soc. B. Biol. Sci. 266, 2125–2129 (1999).
52. Y. Han, Z. Meng, Y. Wu, S. Zhang, S. Wu, Structural colored fabrics with brilliant colors, low
angle dependence, and high color fastness based on the Mie scattering of Cu2O spheres.
ACS Appl. Mater. Interfaces 13, 57796–57802 (2021).
53. L. Maiwald, S. Lang, D. Jalas, H. Renner, A. Y. Petrov, M. Eich, Ewald sphere construction for
structural colors. Opt. Express 26, 11352–11365 (2018).
54. B. D. Wilts, X. Sheng, M. Holler, A. Diaz, M. Guizar-­Sicairos, J. Raabe, R. Hoppe, S.-­H. Liu,
R. Langford, O. D. Onelli, D. Chen, S. Torquato, U. Steiner, C. G. Schroer, S. Vignolini,
A. Sepe, Evolutionary-­optimized photonic network structure in white beetle wing scales.
Adv. Mater. 30, 1702057 (2018).
55. J. D. Forster, H. Noh, S. F. Liew, V. Saranathan, C. F. Schreck, L. Yang, J.-­G. Park, R. O. Prum,
S. G. J. Mochrie, C. S. O’Hern, H. Cao, E. R. Dufresne, Biomimetic isotropic nanostructures
for structural coloration. Adv. Mater. 22, 2939–2944 (2010).
56. E. S. A. Goerlitzer, R. N. Klupp Taylor, N. Vogel, Bioinspired photonic pigments from
colloidal self-­assembly. Adv. Mater. 30, e1706654 (2018).
57. K. Koch, W. Barthlott, S. Koch, A. Hommes, K. Wandelt, W. Mamdouh, S. De-­Feyter,
P. Broekmann, Structural analysis of wheat wax (Triticum aestivum, c.v. “Naturastar” L.):
From the molecular level to three dimensional crystals. Planta 223, 258–270 (2006).
58. R. Maia, H. Gruson, J. A. Endler, T. E. White, pavo2: New tools for the spectral and spatial
analysis of colour inr. Methods Ecol. Evol. 10, 1097–1107 (2019).
59. N. S. Hart, J. C. Partridge, I. C. Cuthill, Visual pigments, oil droplets and cone photoreceptor
distribution in the European starling (Sturnus vulgaris). J. Exp. Biol. 201, 1433–1446 (1998).
60. N. S. Hart, M. Vorobyev, Modelling oil droplet absorption spectra and spectral sensitivities
of bird cone photoreceptors. J. Comp. Physiol. A Neuroethol. Sens. Neural Behav. Physiol.
191, 381–392 (2005).
61. S. Kumar, A. P. Bhondekar, P. Jain, S. Bagchi, A. Sharma, R. Kumar, S. Mishra, Artificial lipid
membrane: Surface modification and effect in taste sensing. IOP Conf. Ser. Mater. Sci. Eng.
360, 012039 (2018).
62. M. Vorobyev, D. Osorio, Receptor noise as a determinant of colour thresholds. Proc. Soc. B
Biol. Sci. 265, 351–358 (1998).
63. W. Chu, H. Gao, S. Cao, X. Fang, H. Chen, S. Xiao, Composition and morphology of
cuticular wax in blueberry (Vaccinium spp.) fruits. Food Chem. 219, 436–442 (2017).
Acknowledgments: We wish to acknowledge J. Mantell and the Wolfson Bioimaging Facility
for their support and assistance in this work. We thank I. Cuthill for extensive comments and
suggestions on a previous version of the manuscript, P. Heard for help with FIB for fig. S7,
N. Roberts for the use of the Ecology of Vision optical laboratory and further comments, and
D. Voigt for contribution to methodology discussion with A.K. We also thank M. Feingold and
the Royate Hill community plum orchard, M. Hughes and P. Brownless at the Royal Botanic
Garden Edinburgh, and B. Ditsch at the TUD Botanischer Garten Dresden for help collecting
and identifying samples. Funding: This work was supported by Engineering and Physical
Sciences Research Council EP/R513179/1 (R.M.), Bristol Centre for Agricultural Innovation
project grant 2019/20 (R.M.), Biotechnology and Biological Sciences Research Council BB/
S00873X/1 (S.A.T.), The Royal Society IES\R2\212116 (R.M.), and Biotechnology and Biological
Sciences Research Council Flexible Talent Mobility Award (2 and 3) (R.M., H.W., and S.M.).
Author contributions: Conceptualization: S.M., R.M., S.A.T., and H.W. Methodology: R.M.,
S.A.T., A.K., S.W., and H.W. Formal analysis: R.M., S.A.T., and S.W. Investigation: A.K., S.M., R.M.,
and S.A.T. Samples: R.M., S.A.T., and S.M. Writing—original draft: R.M. Writing—review and
editing: All authors. Funding acquisition: R.M., H.W., S.M., and S.A.T. Competing interests: The
authors declare that they have no competing interests. Data and materials availability: All
data needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials. Raw data and analysis scripts are available at 10.6084/m9.
figshare.24552778.
Submitted 22 August 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk4219
9 of 9
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

M AT E R I A L S S C I E N C E Copyright © 2024 The

Authors, some rights
Real-­time in vivo monitoring of intraocular pressure reserved; exclusive
licensee American
distribution in the anterior chamber and vitreous Association for the
Advancement of
chamber for diagnosis of glaucoma Science. No claim to
original U.S.
Government Works.
Hunkyu Seo1,2†, Yeon-­Mi Hong1,2†, Won Gi Chung1,2, Wonjung Park1,2, Jakyoung Lee1,2, Distributed under a
Hong Kyun Kim3, Suk Ho Byeon4, Dai Woo Kim3*, Jang-­Ung Park1,2,5,6* Creative Commons
Attribution
Glaucoma causes irreversible vision loss due to optic nerve damage and retinal cell degeneration. Since high NonCommercial
intraocular pressure (IOP) is a major risk factor for glaucoma development, accurate IOP measurement is crucial, License 4.0 (CC BY-­NC).
especially intravitreal IOP affecting the optical nerve and cells. However, conventional methods have limits in
selectively and directly detecting local retina pressure. Here, we present continuous measurements of local IOP
values in the anterior chamber and vitreous chamber of living animals using minimally invasive probes with
pressure-­sensitive transistors. After inducing glaucoma in animal models, we compared the local IOP distribution
between normal and glaucomatous eyes. We also compared IOP values detected in the cornea using tonometry
measurements. Our findings revealed that glaucoma induced higher IOP in the vitreous chamber than in the
anterior chamber, indicating that measuring IOP in the vitreous chamber is key to the glaucoma model. This prog-
ress offers future directions for diagnosis and treatment of glaucoma.

INTRODUCTION
Glaucoma is an optic neuropathy that results primarily from damage
to the axons of retinal ganglion cells (RGCs) as they exit the eye at the
optic nerve head (ONH). The laminar region of the ONH, known as
the lamina cribrosa (LC), is a major site of RGC vulnerability and the
location of RGC axonal injury (1). Intraocular pressure (IOP) induces
the deformation of the LC, and these changes promote damage to
axons and their cell bodies by various mechanisms, including block-
ing the axonal transport and reducing the diffusion of nutrients from
capillaries inside the laminar beam to the adjacent axons (2). Since
IOP is the only modifiable biomarker of glaucoma (3, 4), accurate
measurement of IOP is essential for the early diagnosis of glaucoma
and the implementation of appropriate treatment strategies.
Among the various methods for IOP measurements, including
smart contact lenses, Goldmann applanation tonometry is considered
to be the gold standard (5–13). These methods, such as tonometry
and smart contact lenses, indirectly calculate the values of IOP by
measuring the deformation of the cornea (14–17). Thus, occasional
misalignment of externally applied forces on the surface of the cornea,
as well as discrepancies in the sizes of contact lenses with ocular cur-
vature, may result in miscalculated IOP values. In addition, these
methods are limited in selectively measuring local IOP distribution in
the anterior chamber and the vitreous chamber. To better illustrate
the change in IOP, Fig. 1A shows the structure of an eye and the flow
of aqueous humor. The aqueous humor, which helps in maintaining
1
Department of Materials Science and Engineering, Yonsei University College of
Engineering, Seoul 03722, Republic of Korea. 2Center for Nanomedicine, Institute
for Basic Science (IBS), Yonsei University, Seoul 03722, Republic of Korea. 3Depart-
ment of Ophthalmology, Kyungpook National University School of Medicine, Daegu
41944, Republic of Korea. 4Department of Ophthalmology, Yonsei University College
of Medicine, Seoul 03722, Republic of Korea. 5Department of Neurosurgery, Yonsei
University College of Medicine, Seoul 03722, Republic of Korea. 6Graduate Program
of Nano Biomedical Engineering (NanoBME), Advanced Science Institute, Yonsei
University, Seoul 03722, Republic of Korea.
*Corresponding author. Email: jang-­ung@​yonsei.​ac.​kr (J.-­U.P.); proector97@​gmail.​
com (D.W.K)
†These authors contributed equally to this work.
the IOP, is generated in the ciliary body and it is drained through the
anterior chamber into the trabecular meshwork (3, 18). Disruptions
in this pathway can lead to an imbalance between the generation and
the release of the aqueous humor (4), resulting in its accumulation
and increased levels of IOP. Elevated IOP can cause damage to both
the retina and the optic nerve (19). However, note that the cornea is
located at the front of the eye, while the optic nerve is situated at the
back. In studies involving rabbits and pigs (20), tonometer-­measured
IOP is well correlated with the IOP measured in the anterior chamber,
but it tends to be underestimated (21). Moreover, some studies have
reported the difference in IOP between the anterior chamber and the
vitreous chamber, but these studies were focused on enucleated eyes
(22, 23). Therefore, there is a need to confirm these differences in vivo.
The biomechanical characteristics of the cornea and the sclera can
vary substantially among individuals. For instance, in conditions such
as pseudoexfoliative glaucoma, differences in scleral stiffness can af-
fect IOP notably (24). These variations in biomechanical properties
can affect the accuracy of the conventional methods in measuring the
actual IOP applied to the optic nerve. Therefore, for an accurate diag-
nosis of glaucoma, it is very important to obtain real-­time, in vivo lo-
cal distribution of the IOP measured in the anterior chamber and the
vitreous chamber, as well as the IOP values at the cornea.
A highly sensitive and minimally invasive device also may be re-
quired to monitor IOP using an implant. While many highly sensitive
pressure sensors are available, most of them are designed for monitor-
ing external physical signals that occur outside the body (25–30).
However, due to their bulky size, these sensors are not suitable for
minimally invasive implantation into the body (31). Previous at-
tempts have been made to insert pressure sensors into organs inside
the human body, such as the brain or liver (32–34). However, catego-
rizing these sensors as minimally invasive is misleading since they are
several millimeters in size, and when they are used for monitoring
physical signals in the human eye, potential damage to the eye and
measurement errors can occur.
To address these limitations, we have developed a minimally
invasive, probe-­type pressure-­sensitive transistor for selectively

Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024 1 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

monitoring the local distribution of IOP in specific areas of the
anterior chamber and the vitreous chamber. This field-­effect tran-
sistor (FET) can be used as a pressure sensor to detect IOP di-
rectly, and it is designed in the shape of a sharp needle to minimize
damage when it is inserted into the specific eye chambers of living
animals and to reduce its impact on IOP distribution. This pressure-­
sensitive FET exhibits high sensitivity and fast response time.
Moreover, its fine geometry minimizes insertion-­related damage
and ensures reliable IOP monitoring in vivo. Using this pressure
sensor, we conducted real-­time monitoring of the local IOP distri-
bution in the anterior chamber and the vitreous chamber of live
animals. Also, we induced glaucoma through cauterization and
compared the differences in the IOP distribution between normal
animals and animal models with glaucoma. These IOP values that
were monitored in these two chambers using our pressure sensors
also were compared with gold standard tonometry measurements
that detect IOP in corneal deformation. With the induction of
glaucoma, the vitreous chamber experienced relatively higher IOP
than the anterior chamber. This indicated that measuring IOP in
the vitreous chamber is key to the glaucoma model, and previous
methods of calculating IOP from corneal deformation may have
limitations in accurately measuring the IOP in the retina in the
glaucoma model. Our study involved monitoring the IOP of
normal animals and acute glaucomatous hyaluronic models,
and these results can provide valuable insights for accurately diag-
nosing glaucoma.

Fig. 1. Probe-­type pressure sensor for monitoring IOP distribution. (A) Schematic illustration of eye structure and IOP generation due to the flow of the aqueous
humor. (B) Schematic layouts of the probe-­type pressure sensor. (C) Photograph of the probe-­type pressure sensor. Scale bar, 5 mm. (D) Optical micrograph of the
pressure-­sensitive transistor. Source/drain electrodes were patterned on both sides of the isolated Si channel (left). The transistor was completed by assembling a dielec-
tric layer and gate electrode (right). Scale bars, 100 μm. (E) The transfer characteristics of the transistor (VD = 1 V). (F) The output characteristics of the transistor (VG = −10
to 50 V). (G) Schematic illustration of IOP sensing by a probe-­type pressure sensor in the vitreous chamber. PDMS, polydimethylsiloxane.

Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024 2 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
RESULTS
Fabrication of an eye-­implantable probe with a
pressure-­sensitive transistor
As illustrated in Fig. 1B, the eye-­implantable probe with a pressure-­
sensitive FET is composed of two panels of biocompatible polyimide.
A single-­crystalline silicon layer (thickness: 205 nm) was used as the
FET channel (channel length: 70 μm, width: 10 μm) and transferred
from a silicon-­on-­insulator (SOI) wafer onto the bottom polyimide
(PI) panel (thickness: 25 μm) using a polydimethylsiloxane (PDMS)
stamp. Before this transfer, an SU-­8 layer (thickness: 1.5 μm) was pre-
spun as an adhesive on the bottom PI panel. After transferring the Si
channel, source/drain electrodes (Cr/Au, 2 nm/300 nm) and inter-
connects were deposited using an e-­beam evaporator, and then they
were photolithographically patterned. Successively, an elastomeric
partition spacer (thickness: 50 μm) of PDMS to define a local air gap
was coated to form the air dielectric layer of this FET. Separately, the
gate (Cr/Au, 2 nm/300 nm) and interconnects were deposited and
patterned on the other PI panel (thickness: 25 μm). This top-­gate elec-
trode was brought into conformal contact with the PDMS layer and
fully covered the top of the air-­dielectric layer by laminating these two
PI panels. Sealing the entire sidewalls of the resulting device using a
biomedical-­grade elastomer (SILASTIC MDX4-­4210, Dow Corning)
completed the fabrication of the eye-­implantable probe with a
pressure-­sensitive FET. The fabrication process is described in detail
in Materials and Methods, and it also is illustrated schematically in
fig. S1. The height of the air gap was determined by the thickness of
the elastomeric partition spacer between the Si channel and the top
gate, and it was decreased by applying compressive pressure with in-
creasing capacitance of the metal-­air-­channel structure. This pressure-­
sensitive capacitance change enables the individual FET to act solely
as a single tactile pressure sensor.
Figure 1C shows a photograph of the resulting eye-­implantable
probe. This probe was 190 μm wide and 110 μm thick, and the
pressure-­sensitive FET was positioned at the tip. Figure 1D presents
an optical micrograph of this FET. In the figure, the left inset shows
the source/drain electrode with the channel, and the right inset shows
their assembly with the gate electrode. Figure 1 (E and F) plots the
transfer and output characteristics of this FET, respectively (without
applying pressure). The mobility of the device was calculated as
~520 cm2 V−1 S−1, and the on/off ratio and threshold voltage were
4.3 × 103 and 12.3 V, respectively. The calculations of these character-
istics are described in the Supplementary Materials. As plotted in
fig. S2, this air-­dielectric FET displayed negligible hysteresis due to the
clean interface between the channel and the gate, which was suitable
for rapid and reliable responses.
As illustrated in Fig. 1G, this long sharp probe with a pressure-­
sensitive FET was implanted directly into the anterior chamber or
vitreous chamber of an eye in vivo. The insertion of this probe was
performed after piercing a small hole in the eyeball using a 26-­gauge
needle. The outer diameter of this needle was only 474 μm, and the
cross-­sectional area of our probe (width: 190 μm, thickness: 110 μm)
was small enough to be inserted through this hole. After this probe
was inserted into the anterior chamber or the vitreous chamber (close
to the surface of the retina), the hole in the eyeball was sealed com-
pletely by a tissue adhesive (Vetbond, 3M). The tissue adhesive rap-
idly cured within seconds. Following the sealing process, we allowed
a stabilization period of 10 min to regulate the IOP before conducting
the experiment. Then, we compared the change in IOP using a to-
nometer (ICare Pro, Tonolab, ICare) before the insertion of this probe
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
and after stabilization to confirm that there was a negligible difference
in the IOP before proceeding with the subsequent experiment. The
small size of the hole and additional sealing using a tissue adhesive
effectively prevented any leakage of the aqueous humor. The tip of the
probe that was inserted was sufficiently small compared to the ocular
volume, and its position was stationary without its bending inside the
eye due to the isotropic intraocular forces caused by the aqueous hu-
mor inside the anterior chamber or vitreous chamber. Then, the
pressure-­sensitive transistor operated after connecting it to the source
meter and power supply for measuring the changes in the drain cur-
rent and in the IOP (drain voltage: 1 V, gate bias: 20 V). During the
measurement of IOP using this probe, the animals were anesthetized
by injection of ketamine hydrochloride and xylazine. For example, the
rats were anesthetized via intraperitoneal injection (ketamine hydro-
chloride: 75 mg kg−1, xylazine: 10 mg kg−1), and the rabbits were in-
jected intramuscularly with (ketamine hydrochloride: 50 mg kg−1,
xylazine: 5 mg kg−1). In addition, the rabbits were deeply anesthetized
by inhalation of 3% isoflurane for stable measurements.
Pressure-­sensing performances and in vitro
IOP measurements
Figure 2A illustrates the sensing mechanism of the pressure-­sensitive
FET. When the IOP caused by the vitreous humor in an eye presses
this FET, the thicknesses of the air gap and the PDMS partition spacer
decreased as the capacitance of the gate-­air-­Si channel structure and
the drain current (ID) increased. Because of the elastic property of
PDMS, this FET can act as a single pressure sensor solely with no in-
tegration of an additional component layer. To evaluate the pressure-­
sensing characteristics, a well-­defined load was applied in the pressure
range from 1.3 to 8 kPa using an experimental setup that consisted of
a motorized z-­axis stage (Mark-­10 ESM 303) and a force gauge
(Mark-­10 M7-­20). Figure 2B presents the real-­time measurement of
the relative change [ΔID/I0 (%)] in ID at VG = 20 V and VD = 1 V
under different magnitudes of pressure, where I0 is the current at zero
Pascal and ∆ID = I − I0, which denotes the variation of the ID during
the stepwise pressure loading. The pressure increased stepwise from
1.3 kPa (equivalent to 10 mmHg) to 8 kPa (60 mmHg), with 60 mmHg
being the maximum range that the human IOP typically exhibits. This
graph shows that the response to the applied pressure was distin-
guished clearly as a step-­like response. Figure 2C is a plot of the rela-
tive change in ID with respect to the applied pressure, and ∆ID/I0
increased linearly within the pressure range. The sensitivity of this
pressure sensor was obtained from the slope of this plot, and it was
expressed as [ID/I0 (%)]/∆P, where ∆P denotes the applied pressure.
The sensitivity of the pressure-­sensitive FET was determined to be
0.44% kPa−1 (equivalent to 0.059% mmHg−1) in the pressure range
below 60 mmHg. The signal-­ to-­
noise ratio (SNR) is defined as
SNR = μ/σ, with μ representing the average value of relative ID when
1-­mmHg pressure is applied, and with σ indicating the SD of the noise
levels when the pressure is released. The SNR measurement of our
sensor at 1 mmHg was 32 (μ = 1.5215 nA, σ = 0.0491 nA at 1 mmHg).
We calculated that the minimum detectable IOP, with an SNR of ~3,
is ~0.094 mmHg. This exceeds the resolution of commercially avail-
able tonometers, which typically offer a resolution of 0.1 mmHg, as
well as the previously reported IOP sensors (fig. S3) (35–37).
Figure 2D shows that this pressure-­sensitive FET operated stably;
it had a negligible change in signals (∆ID/I0) after 1000 cycles of
repetitive pressure loading at 8 kPa (60 mmHg). In addition, this pres-
sure sensor exhibited a fast response (~41 ms) and recovery time
3 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(~54 ms) with the applied pressure of 60 mmHg, as plotted in
Fig. 2E. For implantable electronic biomedical devices, the electrical
leakage (that can be caused by environmental factors or insufficient
insulation) is one of the biggest challenges. In particular, in the case of
an eye, the leakage current generated by implantable devices can
cause retinal damage by electrically stimulating optical neurons,
which can be fatal. Therefore, the leakage current of an implantable
electronic device should be less than 10 μA (38–40). To investigate
this leakage, this probe-­type device was immersed inside a balanced
Fig. 2. Sensing properties of the pressure sensor. (A) Schematic illustration of
the sensing mechanism. (B) Real-­time measurements of the relative change in
drain current (ΔID/I0) according to the various pressures (VG = 20 V, VD = 1 V). (C) The
relative change in ID with the applied pressures. S represents the sensitivity of the
pressure sensor. (D) Relative change in the ID of the pressure sensor that occurred
after 1000 cycles of repeated compression to 60 mmHg. (E) Response and recovery
time of the pressure sensor when a pressure of 60 mmHg is applied. (F) Compari-
son of sensing properties in a pressure sensor with one side fixed to the bottom
and a pressure sensor floating on fluid. (G) The relative change in ID with applied
water pressure as the pressure is increased stepwise. S represents the sensitivity of
the pressure sensor. (H) Correlation between the IOP measurements taken using
the pressure sensor and the manometer. PI, polyimide.
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
salt solution (BSS) for 7 days. BSS is the perfusate that is used during
the surgery on human eyes, and it is the solution most similar in com-
position to the aqueous humor and the vitreous humor. As shown in
fig. S4A, the pressure-­sensitive FET preserved its sensing perfor-
mance with negligible deviation in signals (∆ID/I0). Also, the leakage
current was measured to be less than 1 nA, which was small enough
for the leakage current of implantable ocular devices. Similarly, this
device also exhibited negligible changes in signals when it was im-
mersed in a phosphate-­buffered saline (PBS) solution for 7 days
(fig. S4B). In addition, there was negligible variation in the sensitivity
within the temperature range of 30° to 45°C (fig. S5). In Materials and
Methods, we conducted tests on human retinal pigment epithelial
cells to assess the cytotoxicity of the pressure-­sensing probe. Accord-
ing to fig. S6, the cell viability was 84.7 ± 5.8%, indicating no sub-
stantial difference compared to commercially available ophthalmic
implantable devices (41, 42). Therefore, we concluded that our IOP-­
sensing probe is not substantially cytotoxic to the human eyes.
After immersing this probe inside a water tank, the pressure-­
sensing performance of this device was compared when it was at-
tached to the bottom of the tank versus suspended stationarily in
water (at the same height) without its attachment. As exhibited in
Fig. 2F, this pressure sensor exhibited negligible signal differences
in these two cases. The vitreous chamber and the anterior chamber are
filled with aqueous humor and vitreous humor, which are liquids.
Therefore, the pressure-­sensing performance of our probe was com-
pared with the measurement using a commercial pressure sensor
(MS5803-­14BA, TE Connectivity) when these two different devices
were submerged in water. Our pressure-­sensitive FET detected the
same pressure values as the commercial device in the pressure range
below 60 mmHg, and Fig. 2G shows that the signal (∆ID/I0) of our
sensor increased linearly with the applied pressure even when im-
mersed in water without substantial difference in sensitivity (percent-
age difference between air and water cases: 14%). Also, we compared
the sensing performance of our probe with the measurement of a con-
ventional manometer. As shown in fig. S7, our probe and a manome-
ter were implanted together into a bovine eye to obtain their real-­time
IOP measurements. Figure 2H and fig. S8 indicate that these two dif-
ferent ways of measuring IOP, i.e., using either our pressure-­sensitive
FET or the manometer, had a good correlation [intraclass correlation
coefficient (ICC) = 0.997) with the coefficient of determination (r2)
of 0.99, which indicated good reproducibility for both IOP mea-
surements.
In vivo studies for monitoring IOP distribution in
live animals
Using live rabbit models, our probe-­type pressure sensors were used
to monitor the local IOP distribution in both the anterior chamber
and the vitreous chamber, and a commercially available tonometer
(ICare Pro, ICare) was used to detect IOP in the eye based on the de-
formation of the cornea. Figure 3A illustrates the experimental setup,
and Fig. 3B presents a photograph of the two pressure sensors that
were implanted. Before conducting further experimentation, we en-
sured that the implantation of these probes caused negligible IOP dif-
ferences, as measured by tonometry (Fig. 3C). For example, the
rabbit’s IOP initially remained at 9.43 ± 0.39 mmHg (before insertion
of the probe), and it was 9.49 ± 0.32 mmHg after the probe insertion.
Also, after this probe insertion, the IOP in the contralateral control
eye was recorded as 9.56 ± 0.66 mmHg, indicating a negligible differ-
ence. Subsequently, we started real-­time monitoring of localized IOP
4 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
distributions in the cornea, anterior chamber, and vitreous chamber.
A comprehensive experimental setup, including the measuring instru-
ments, is shown in fig. S9. Figure 3D and fig. S10 plot the IOP distri-
bution for the normal rabbit model. The IOP values within these three
areas were maintained similarly within the normal IOP range and no
drastic variations occurred. Also, Fig. 3E presents the mean values
and SDs of IOP in the three regions during the measurement peri-
od (cornea: 9.46 ± 1.24 mmHg, anterior chamber: 9.73 ± 1.47 mmHg,
vitreous chamber: 9.67 ± 1.44 mmHg), and substantial differences
were observed among the three areas. Comparing the IOP values
detected by the tonometer (IOPtono) at the cornea with those mea-
sured via our pressure sensors implanted in both chambers (IOPchamber),
Fig. 3F indicates a strong correlation (anterior chamber: r = 0.815
and ICC = 0.769, vitreous chamber: r = 0.754 and ICC = 0.729).
The differences between the IOP values obtained at the cornea (by the
tonometer) and those measured in the chambers are depicted in
Fig. 3G. The variations in IOP in these regions were less than 0.5
mmHg, and the distribution of IOP among the cornea, anterior
chamber, and vitreous chamber in a normal rabbit model indicates
the reliability of conventional tonometry measurements in the nor-
mal IOP range.
Hyaluronic acid is a naturally occurring polysaccharide that is
used as a viscoelastic tool in ophthalmological surgery, and it also is
included in artificial tears. It is used to increase IOP to prevent the
collapse of the anterior chamber and to facilitate the manipulation of
ocular tissues during surgery. As illustrated in Fig. 4A, we injected 0.2 ml
of hyaluronic acid into the anterior chamber in a live normal rabbit
model to increase the volume of aqueous humor in the anterior cham-
ber and to raise the IOP level. Figure 4B is a photograph of this injec-
tion of hyaluronic acid through a syringe (needle outer diameter:
0.474 mm), and the photograph shows that there was no leakage of
hyaluronic acid, ocular fluid, or blood. Figure 4C and fig. S11 present
the real-­time measurements of the local IOP distribution in the ante-
rior chamber and the vitreous chamber (using our pressure-­sensitive
probes) as well as at the cornea (using a tonometer), after the injection
of the hyaluronic acid. The injection of the hyaluronic acid induced an
acute increase in IOP (within 30 s) in both chambers. Subsequently,
the increased IOP decreased gradually and eventually returned to the
normal IOP range. Figure 4D compares the probe-­measured IOP in
the vitreous chamber and anterior chamber (IOPchamber) with the
tonometer-­measured IOP (IOPtono) at the cornea. Despite the injec-
tion of the hyaluronic acid, the IOP distributions between the cornea,
the anterior chamber, and the vitreous chamber were well aligned
within the entire IOP range of 7 to 60 mmHg. For example, Fig. 4 (E
and F) shows high concordance of IOP values in the anterior chamber
(IOPant) and vitreous chamber (IOPvit) in the normal IOP range (7 to
15 mmHg) and high IOP conditions (15 to 60 mmHg), respectively.
These in vivo results in a live rabbit model are consistent with previ-
ous in vitro studies using enucleated porcine eyes (17, 22, 43). How-
ever, as shown in Fig. 4G, our in vivo experiments showed that
the absolution difference in IOP between the vitreous chamber and the
anterior chamber, i.e.,│IOPvit − IOPant│, increased slightly in the
high IOP condition (from immediately after the injection of hyal-
uronic acid until the IOP saturation to the normal range). This abso-
lute difference was less than 3 mmHg in the normal IOP range, but it
was more than 3 mmHg in 36% of the points detected at high levels of
IOP (Fig. 4H). In addition, a small increase also was observed when
the relative disparity was considered, as shown in fig. S12A. Specifically,
the positions with a difference of 20% or more increased from 2.3%
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
for normal IOP to 13.2% for the high IOP range (fig. S12B). In view of
these results, IOP measurements using conventional tonometry are re-
liable in all ranges of IOP, but the transient difference (i.e., │IOPvit −
IOPant│) in high IOP ranges, due to temporary acute fluctuations in
IOP in the anterior chamber, may not be reflected accurately in con-
ventional tonometry because it detects IOP only through changes in
the cornea that reflect IOP in the anterior chamber, not in the vitreous
chamber.
In vivo studies for monitoring IOP distribution in normal and
glaucomatous models
The aqueous humor is secreted by the ciliary epithelium and it flows
through the trabecular meshwork into Schlemm’s canals, which are
located between the iris and the cornea. When the eye makes too
much aqueous humor or the drainage system does not work properly,
the accumulated fluid increases the pressure in both the anterior and
vitreous chambers, and it is commonly linked to elevated IOP. In glau-
coma models, the pressure in the vitreous chamber may differ from
the pressure in the anterior chamber due to various factors, such as
the presence of the lens and variations in fluid dynamics and drainage
pathways. Thus, measuring the pressure in the vitreous chamber is
crucial because it can provide valuable information about the condi-
tion of the retina, which helps in early diagnosis of problems, such as
glaucoma, and it enables effective management to prevent the loss of
vision. In addition, this measurement of both sides may be important
for understanding the pathophysiological progression of disease in
glaucoma. We utilized this pressure-­sensitive FET to monitor the dis-
tribution of IOP in the eyes of rats. Glaucoma was induced in rats by
cauterizing the two episcleral veins located in the upper part of the eye
(Fig. 5A and fig. S13). The induction of glaucoma was confirmed by
hematoxylin and eosin (H&E) staining of the retina. Figure 5B pres-
ents the degradation of the uppermost RGC layer in the glaucoma
model, compared to the normal model case. The RGC density in
glaucoma-­induced rats was only 60.2% compared to the normal
group (fig. S14). Then, we monitored the local IOP values in both the
vitreous and anterior chambers of these rats’ eyes continuously after
inserting pressure-­sensor probes inside these chambers. Simultane-
ously, for comparison, we also measured IOP in the cornea using a
tonometer (Fig. 5C). Also, Fig. 5D shows a photograph of these two
pressure-­sensor probes, each inserted into the anterior chamber and
the vitreous chamber, without any bleeding or substantial leakage of
ocular fluid. In addition, fig. S15 indicates that the change in the IOP
caused by the probe we inserted was negligible.
Real-­time IOP measurements were conducted for 30 min in three
eyes each from the normal model and the glaucoma model (Fig. 5E
and figs. S16 and S17). Snapshot IOP measurements using a commer-
cial tonometer (Tonolab, ICare) were performed at 150-­s intervals. As
shown in Fig. 5F, in the normal model, the tonometry measurement
showed an average IOP value of 10.33 ± 0.65 mmHg, pressure-­sensor
detection exhibited 10.62 ± 0.62 mmHg in the anterior chamber, and
10.87 ± 0.69 mmHg in the vitreous chamber. This normal model pre-
sented a similar IOP distribution in the cornea, the anterior chamber,
and the vitreous chamber. However, in the glaucoma model, the cor-
neal IOP (tonometry detection: 15.54 ± 0.81 mmHg on average) was
almost similar to the anterior chamber’s IOP (pressure-­sensor mea-
surement: 15.43 ± 0.76 mmHg), but the vitreous chamber’s IOP
(pressure-­sensor measurement: 19.51 ± 1.29 mmHg) was noticeably
higher. This glaucoma model exhibited substantially higher IOP in all
three regions (i.e., the cornea, the anterior chamber, and the vitreous
5 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. IOP distribution in a live rabbit model at normal IOP level. (A) Schematic illustration of an experimental setup for measuring the IOP distribution with a tonom-
eter and the probe-­type pressure sensor on a live rabbit model. (B) Photograph of the pressure sensors inserted into the eye of a live rabbit. Scale bar, 5 mm. (C) The aver-
age IOP of the rabbit’s eye before and after the insertion of the device. A tonometer was used to measure the IOP. (D) Real-­time measurements of the IOP distribution in
the vitreous chamber, anterior chamber, and tonometer. (E) Normal IOP is measured with a tonometer and a pressure sensor. The normal IOPs of the anterior chamber and
the vitreous chamber were measured using a probe-­type pressure sensor. (F) Local IOP distribution of the anterior chamber and the vitreous chamber (IOPchamer) com-
pared to the tonometer (IOPtono) in a normal IOP level. (G) The difference between the IOP (ΔIOP = IOPchamber − IOPtono) measured in the chambers of the rabbit’s eye and
the IOP measured using a tonometer.

Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024 6 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. IOP distribution in a live rabbit model at a high IOP level. (A) Schematic illustration of the hyaluronic injection method for inducing a high IOP level. (B) Photo-
graph of hyaluronic injection into the anterior chamber. Scale bar, 5 mm. (C) Real-­time measurements of IOP distribution with pressure sensors and tonometer. The green
part indicates where the high IOP level was induced by hyaluronic injection. (D) Local IOP distribution of the anterior chamber and the vitreous chamber (IOPchamber)
compared to the tonometer (IOPtono) in high IOP level. (E) Local IOP distribution between the anterior chamber (IOPant) and the vitreous chamber (IOPvit) at the normal IOP
level. (F) Local IOP distribution between the anterior chamber (IOPant) and the vitreous chamber (IOPvit) at a high IOP level. (G) IOP differences (IOPvit − IOPant) between
the anterior chamber and the vitreous chamber. (H) The distribution of IOP difference (│IOPvit − IOPant│) values in the normal IOP level and the high IOP level.

Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024 7 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. IOP distribution in normal and glaucoma rats. (A) Schematic illustration of glaucoma induction by cauterization of the episcleral veins. (B) Hematoxylin and eosin
staining of the normal model and glaucoma model. GCL, ganglion cell layer; INL, inner nuclear layer; ONL, outer nuclear layer; IS/OS, photoreceptor inner and outer segments;
RPE, retinal pigment epithelium. Scale bar, 100 μm. (C) Schematic illustration of an experimental setup for measuring the IOP distribution in rats. (D) Photograph of the probe-­
type pressure sensors inserted into the eye of live rats. Scale bar, 2 mm. (E) Real-­time measurements of the IOP distribution in the vitreous chamber, anterior chamber, and
tonometer with normal and glaucoma models. (F) Changes in IOP measured by tonometer and in the vitreous chamber and anterior chamber, caused by glaucoma. (G) Local
IOP distribution of the anterior chamber and the vitreous chamber compared to the IOP measured with the tonometer. (H) Local IOP distribution of the anterior chamber and
the vitreous chamber in a normal eye and glaucoma-­induced eye. (I) The IOP differences between IOP in the anterior chamber and vitreous chamber compared to tonometer-­
measured IOP in normal and glaucoma models. (J) The percentage of time at which the difference in IOP in the vitreous chamber compared to values in other regions (top
inset: tonometer; bottom: anterior chamber). (K) The Pearson’s correlation coefficient between the tonometer and the vitreous chamber (top inset) and the Pearson’s correla-
tion coefficient between the anterior chamber and the vitreous chamber (bottom) that measured in both normal and glaucoma models.

Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024 8 of 13

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
chamber) compared to the case of the normal model case, indicating
a successful increase in IOP through cauterization. Three mice were
tested for each model, and figs. S18 and S19 show the average IOP of
each subject for normal and glaucoma-­induced eyes, respectively. Fig-
ure 5G compares the IOP of each chamber to the corneal IOP for the
normal and glaucoma models. In particular, Fig. 5H plots all IOP val-
ues detected in the anterior chamber and vitreous chamber of each
subject (in the normal and glaucoma models). The red dotted line
indicates the value at which IOP in the vitreous chamber and the an-
terior chamber are identical. In normal eyes, most values were around
the dotted line (in both the anterior and vitreous chambers). Howev-
er, while IOP values in the anterior chamber of the glaucoma-­induced
eye were still around this red dotted line, all values in the vitreous
chamber of the glaucoma model were above this baseline. Figure 5H
indicates that, with the induction of glaucoma, the vitreous chamber
experienced relatively higher IOP than the anterior chamber experi-
enced. All other subjects presented the same trend, as shown in
fig. S20.
Furthermore, Fig. 5I shows the difference between the corneal IOP
and each chamber’s IOP in the normal and glaucoma models. Although
the normal model exhibited negligible difference in IOP, the glaucoma
model presented a substantial difference in IOP (3.84 ± 1.24 mmHg)
between the vitreous chamber and the cornea, indicating a higher
IOP in the vitreous chamber. This emphasizes that the development
of glaucoma resulted in substantially higher IOP in the vitreous
chamber than it did in the anterior chamber and cornea. Figure 5J
shows the temporal percentage during which the vitreous chamber’s
IOP was higher than the IOP values of the other two regions (i.e., the
cornea and the anterior chamber). In the case of normal eyes, the dif-
ference between the vitreous chamber’s IOP and the corneal IOP is
maintained below 10% for 71.8% of the total measured time. (Simi-
larly, the difference between the vitreous chamber’s IOP and the ante-
rior chamber’s IOP was maintained below 10% for 93.4% of the total
time). On the other hand, in the case of glaucomatous eyes, the period
in which the vitreous chamber’s IOP was more than 10% higher than
the corneal IOP was 97.5% of the total measurement time. (For the
anterior chamber of the glaucoma model, this period was 98.9%.) Fig-
ure 5K compares the Pearson’s correlation coefficient before and after
the induction of glaucoma induction, and all three subjects in the
glaucoma model had decreases in this coefficient compared to the
cases involving the normal model. Our results indicated that glauco-
ma alters the local distribution of IOP in the eye, with a marked in-
crease of the IOP in the vitreous chamber, where RGC damage occurs.
This is thought to be due to differences in the biomechanical proper-
ties between the anterior chamber and the vitreous chamber. In par-
ticular, the lens occupies a larger proportion of the total eye in the rat
than in the rabbits. In addition, glaucoma causes tissue remodeling of
the ocular components such as the cornea and sclera, leading to a
change in biomechanical properties. This change can maximize the
IOP difference between the anterior chamber and the vitreous cham-
ber. This means that it may be difficult to use conventional tonometry
for measurements in glaucoma models because it only measures IOP
through changes in the cornea which primarily reflects the IOP in the
anterior chamber, rather than in the vitreous chamber.
DISCUSSION
This in vivo study presents a minimally invasive, probe-­type pressure-­
sensitive FET for monitoring the local IOP distribution in the anterior
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
chamber and the vitreous chamber of living animals. To use this FET
as an IOP sensor, it was designed with a long, sharp, and needle shape
to facilitate its direct insertion into specific local areas of an eye. High
sensitivity, as well as the fast response of this air-­dielectric sensor
structure, enabled reliable and accurate monitoring of acute changes
in the IOP in animal models. Also, this implantation did not cause any
bleeding or leakage of ocular fluid. A comparison of IOP values mea-
sured with a tonometer before and after the implantation of our probe
confirmed that this insertion did not alter the IOP.
With the sensor that was fabricated, first, we confirmed the differ-
ence in IOP distribution in both normal and glaucoma models of liv-
ing animals. When comparing the IOP distribution between normal
and glaucomatous eyes in vivo models after the induction of glauco-
ma, both the vitreous chamber and the anterior chamber had distinct
differences in the IOP, demonstrating that the IOP can vary depend-
ing on its intraocular location. The normal model showed an align-
ment of the distribution of the IOP among the cornea, the anterior
chamber, and the vitreous chamber, but the glaucoma model present-
ed notably higher IOP in the vitreous chamber, where RGC damage
occurred. This means that it may be difficult for conventional tonom-
eters to measure IOP accurately in glaucoma models because they
measure IOP only through changes in the cornea, which reflects IOP
in the anterior chamber rather than in the vitreous chamber. Further-
more, when hyaluronic acid was injected into the anterior chamber in
a living normal model to increase the volume of aqueous humor in
this chamber, our continuous monitoring sensor observed transient
and acute IOP differences between the anterior chamber and the vit-
reous chamber in high IOP ranges. Given that the IOP spikes that are
caused by radical fluctuations can accelerate glaucoma (44), these de-
viations are essential factors to consider in the management of glau-
coma. However, these transient and temporary fluctuations in IOP
may not be reflected accurately by snapshot measurements of conven-
tional tonometry. This occurrence is due to the biomechanical charac-
teristics of the eye. The fluids that fill the anterior and vitreous
chambers have differences in viscosity and properties. These two ar-
eas, separated by the lens, display volume discrepancies. There are
also variations in the modulus of the cornea and sclera that surround
these areas. The IOP difference observed in these two compartments
is expected to be due to these factors. Moreover, conditions like glau-
coma can cause reorganization-­ induced changes in ocular tissue
characteristics, which can worsen the IOP imbalances.
Although the distributions of IOP measured using our probe-­type
pressure sensors indicate a potential approach for the analysis of glau-
coma, our experiments were limited to anesthetized animal models
due to the invasiveness of our device. Therefore, assessing IOP distri-
bution in non-­anesthetized subjects during everyday activities or
study factors that alter IOP (such as posture or exercise) would pro-
vide future strategies for managing glaucoma. However, our current
device form with wired interconnections requires the use of large,
external measuring equipment, and its invasive nature poses a risk of
contagion in everyday situations. Consequently, there is a need for
wireless operation of devices by integrating our sensor with wireless
IC chips and wireless antennas. Then, a secure fixation method for
such wireless devices can be achieved by inserting them into the su-
prachoroidal space.
The suprachoroidal space, which is situated between the sclera
and the choroid, provides an appealing opportunity for the place-
ment of a wireless device, which would eliminate the necessity for
direct intraocular penetration (45). This space displays potential for
9 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
multiple applications, including the delivery of drugs or serving as an
engraftment site for ocular prosthetics (46). Positioned right under
the sclera, the suprachoroidal space offers enough room for the im-
plantation of a device similar in size to an eyeball. The suprachoroidal
space, which is a void between membranes, impedes external devices
from penetrating the chamber, thereby reducing the risk of infection.
This method presents substantial benefits in terms of biocompatibility
and long-­term stability. Although various difficulties remain to be re-
solved, such as wireless functionality and miniaturization, the imple-
mentation of these techniques in the future is anticipated to allow the
assessment of IOP distribution in daily life.
The measured distribution of IOP suggests the need for direct IOP
measurements applied to the retina within the vitreous chamber in a
clinical setting. Specifically, the measurement of IOP in the vitreous
chamber is crucial for glaucoma patients because it can damage the
retina. The aforementioned strategy of inserting devices into the su-
prachoroidal space can be utilized in this regard. Currently, efforts are
being made to transplant nonelectronic devices, such as stents, into
the suprachoroidal space for the treatment of glaucoma (47, 48). With
advancements in wireless capabilities and miniaturization, it will be
possible in the future to insert devices into this space. This approach
has the potential to minimize damage to the vitreous humor and the
lens, thus preserving the natural function of vision without obstruct-
ing the field of view. The device also will help people deal with
concerns about portability and the inconsistent measurements that
commonly are associated with conventional tonometers. This allows
for the direct IOP assessment on the retina. Continuous monitoring
of the IOP permits the identification of increased pressure, which, in
turn, permits the use of drugs to avert injury to the RGCs and the
optic nerve. Ultimately, the adoption of patient-­centered interfaces for
measuring IOP enables the patients’ participation in prevention and
timely therapeutic interventions.
This FET-­based pressure sensor probe is expected to be used for
various other purposes. As an example, the real-­time measure-
ment of intracranial pressure (ICP) is another suitable application
scenario for our probe-­type pressure sensor. ICP can fluctuate due
to brain-­related conditions such as brain hemorrhage and tumors,
so, after brain surgery, continuous monitoring may be required to
confirm the recovery and detect potential side effects (49, 50). Fur-
thermore, ICP shares a similar range with IOP and also can indi-
cate locally different pressures (51). A FET-­based pressure probe is
highly suitable for monitoring ICP because it has minimal inva-
siveness and the capability for precise pressure measurement with-
in the specified pressure range. Real-­ time monitoring of ICP
allows for the confirmation of the progression of the disease and
the postsurgical recovery situation because it provides various
information about ICP and brain diseases. In this way, our probe
has notable potential and will provide advantages for biomedical
purposes.
MATERIALS AND METHODS
Fabrication process of a soft probe-­type pressure sensor
The pressure sensor is fabricated by dividing it into three parts, i.e., (i)
a top film with channel and source/drain electrodes, (ii) a bottom film
with gate electrodes, and (iii) a deformable dielectric layer. A PI film
(25 μm) is attached to the cleaned Si wafer with spin-­coated PDMS
(Sylgard 184, Dow Corning, base:agent = 10:1) at 3000 rpm for 30 s as
an adhesive layer. The PI substrate for the source/drain electrodes was
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
spin-­coated by a negative photoresist (SU-­8 2002, MicroChem) at
4000 rpm for 40 s for the transfer of the Si channel. After the SU-­8
coating was applied, it was cured by exposure to ultraviolet (UV) for
30 s. The Si channel for the transistor was acquired from the top Si
layer (205 nm) of the SOI wafer (205-­nm Si on 400-­nm buried oxide
layer, Soitec). The channel was patterned using a positive photoresist
(S1818, MicroChem) with photolithography on the SOI wafer. S1818
was spin-­coated at 4000 rpm for 40 s, and it was exposed to UV
(365 nm) for 13 s. The adhesion promoter (Surpass 4000, Dis-­Chem)
was used to prevent the failure of the channel in the reactive ion etch-
ing (RIE) process. The channel (length: 10 μm, width: 70 μm) was
etched by RIE using sulfur hexafluoride [SF6, 25 standard cubic centi-
meters per minute (SCCM); Ar, 55 SCCM/300 W/40 s]. After the
channel was etched, the residue of the photoresist was removed
completely with Piranha solution (sulfuric acid:hydrogen perox-
ide = 3:1) for 15 min. The buried oxide layer was etched anisotropi-
cally in a 49% hydrogen fluoride (HF) solution for 9 min and 20 s
to detach the Si channel from the SOI wafer. Then, the channel was
transferred to the PI substrates coated with the SU-­8 layer using
the PDMS stamp (base:agent = 4:1). The metal layers (Cr/Au,
2 nm/300 nm) were deposited for electrodes on the PI substrates by
e-­beam evaporation, and the source/drain electrodes were patterned
using S1818 (3000 rpm, 30 s) photolithographically. For patterning
electrodes, S1818 was exposed to UV for 15 s, and it was removed by
a developer (AZ 300 MIF, Merck) for 45 s. Then, the electrodes were
wet etched by Au etchant (Gold ETCH TFA, Transene) for 30 s and Cr
etchant (Cr Etch 905 N, Transene) for 10 s. First, the photoresist resi-
due was removed through UV exposure (5 min) and developer
(15 min); then, it was cleared completely by RIE (O2, 40 SCCM/200
mTorr/100 W/40 s). Except for the channel and the electrode pads for
interconnection, the SU-­8 layer (thickness: 1.5 μm) was spin-­coated at
4000 rpm for 40 s for the passivation of the electrodes. The PDMS film
(thickness: 50 μm) for the dielectric layer was pattered by using a laser
ablation method (CO2 laser, Epilog Laser, Inc). Then, the top film and
bottom film were patterned for a needle shape using RIE (SF6, 25
SCCM; Ar, 55 SCCM; O2, 90 SCCM/200 W/35 min). The copper film
(thickness: 3 μm), which was made in the shape of a needle, was used
as the mask to protect the films from the RIE, and it was deposited
and patterned by photolithography. Then, the top film, bottom film,
and dielectric layer were assembled using an optical microscope
(OM), and the side of the pressure sensor was sealed by a biomedical
grade elastomer (SILASTIC MDX4-­4210, Dow Corning).
Characterization of pressure-­sensitive transistor
The characteristics of the pressure-­sensitive transistor, including
the transfer and output characteristics, were measured through a
probe station with a parameter analyzer (Keithley 4200-­SCS). The
current between the source electrode and the drain electrode was
measured using a three-­terminal measurement system. Then, to
assess the performance of pressure sensing, an experimental setup
was configured using equipment that was capable of applying pre-
cise pressure. The pressure was applied up to 60 mmHg by using a
movable z-­axis stage (Mark-­10 ESM303) and a highly accurate
force gauge (Mark-­10 M7-­20). At this time, the change in the drain
current in the pressure sensor was measured through source me-
ters (2400 series, Keithley). One source meter was used to apply a
constant voltage to the gate electrode, and another source meter
was connected to the source electrode and drain electrode to mea-
sure the drain current.
10 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Cell culture and cell viability assay
ARPE-­19 cells (retinal pigment epithelial cell line, American Type
Culture Collection, Manassas, VA, USA) were cultured in Dulbecco’s
modified Eagle’s medium (DMEM)/F12 (Thermo Fisher Scientific,
Inc., Waltham, MA, USA) supplemented with GlutaMax, 10% heat-­
inactivated fetal bovine serum (FBS), and penicillin (50 U ml−1)/
streptomycin (50 μg ml−1) (both from Welgene Inc., Gyeongsangbuk-­
do, Korea). The culture media were replaced with fresh media
every 3 days and grown at 37°C in a humidified 5% CO2. For the
cell viability assay, the cells were plated at a density of 1 × 105 cells
ml−1 in 96-­well plates. The fabricated pressure sensor was steril-
ized with 70% ethanol for 3 min, rinsed in H2O for 3 min, and then
exposed to UV for 30 min. The sensor was incubated with DMEM/
F12 1% FBS supplemented GlutaMax for 24 hours, and the sensor-­
immersed medium (SIM) was collected and stored at −80°C until
it was used. Each well was treated with 100 μl of SIM or only media
as a control group for 24 hours, and viability was analyzed using
Quanti-­MAX WST-­8 Cell Viability Assay Kit according to the pro-
tocol provided by Biomax (Seoul, Korea). The optical density was
measured at 450 nm using Victor X5 2030 Multiable Plate Readers
(PerkinElmer).
Bovine eyeball model of ocular hypertension
In vitro tests were conducted by using a bovine eyeball. The IOP of
the bovine eyeball was controlled using a syringe pump by insert-
ing the PBS solution into the vitreous chamber. A commercial ma-
nometer (EM201B, UEi) was inserted into the vitreous chamber of
the eye to measure the IOP of the vitreous chamber in the bovine
eyeball. Also, the fabricated pressure sensor was inserted into the
vitreous chamber to measure the IOP. The change in the drain cur-
rent of the pressure sensor was characterized using source meters
(2400 series, Keithley). One source meter was connected to the
source/drain electrode to apply the drain voltage, and the other
source meter was connected to the gate electrode to apply the gate
voltage. The change in drain current due to the IOP was measured
in real time, and the IOP through the manometer was measured
every 15 s.
Animal models
All of the research in this paper complies with all relevant ethical
regulations. Animal procedures were conducted on the basis of the
guidelines of the Institutional Animal Care and Use Committee of
the Daegu-­Gyeongbuk Medical Innovation Foundation (DGMIF,
DGMIF-­21090302-­01). The Institutional Animal Care and Use
Committee of the Daegu-­Gyeongbuk Medical Innovation Founda-
tion was the ethics review committee. The experimental proce-
dures were carried out in rabbits (New Zealand white rabbit, male,
3.0 kg) and rats (Orient Bio, male, 150 g). The animals were main-
tained on a 12/12-­hour light/dark cycle in a temperature-­controlled
room (23° ± 2°C).
Rat models of glaucoma for IOP measurement
The chronic IOP elevation model (glaucoma model) was implanted
surgically in the left eyes of three male Sprague-­Dawley rats (weighing
>150 g). Before surgery, the rats were anesthetized with an intraperi-
toneal injection of ketamine hydrochloride (75 mg kg−1) and xylazine
(10 mg kg−1) before proceeding to any procedures. The two episcleral
(dorsal and temporal) veins of the left eye were caused by hand-­held
cautery (Change-­A-­Tip, Aaron, USA), while the right eye did not
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
receive surgical procedures and was used as a control. After cauteriza-
tion, the rats were returned to the cage. Three weeks after the chronic
glaucoma animal model, the IOP was measured with two probe-­type
pressure sensors.
Rabbit models for IOP measurement
Rabbits (New Zealand white male rabbits) that weighed 2.5 to
3.0 kg were intramuscularly preanesthetized with ketamine hydro-
chloride (50 mg kg−1) and xylazine 10% (5 mg kg−1) and anesthe-
tized deeply by inhalation of 3% isoflurane (HanaPharm Co. Ltd.,
Seoul Korea). Topical anesthesia (0.4% oxybuprocaine drops) was
applied to the eyes. The IOP of the live rabbits was measured at the
cornea by using the tonometer (ICare Pro, ICare). In addition, the
probe-­type pressure sensors were inserted into the anterior cham-
ber and vitreous chamber. In addition, hyaluronic acid was used to
induce high IOP. Hyaluronic acid is a viscous substance that is in-
cluded in artificial tears, and it is used for surgery or eye treatment
by injecting it into the eye. To increase IOP, 0.2 ml of hyaluronic
acid was used per session.
IOP measurement with the pressure sensor in the
animal model
The pressure sensor was inserted after a hole was made using a
26-­gauge needle. The outer diameter of the 26-­gauge needle was
0.474 mm, and it can make a hole sufficient for the pressure sensor to
enter. The pressure sensor worked by connecting to the source meter
and power supply. The source electrode and drain electrode were con-
nected to both ends of the source meter (2400 series, Keithley) to
measure the change in drain current. A drain voltage of 1 V was ap-
plied through the source meter. The gate electrode was connected to
the power supply (2260b series, Keithley), and a voltage of 20 V was
applied. The drain current was measured in real time through soft-
ware. The measured drain current was calculated through sensitivity
and converted into pressure.
Tissue processing and pathological damage assay
The rat’s retinas were enucleated, and the anterior segments were
removed. Retinas were fixed in Davidson’s fixative (Hartmann’s
fixative; 32% ethanol, 2.2% neutral buffered formalin, and 11%
glacial acetic acid in distilled water) (Merck, Darmstadt, Germany)
for 4 hours at room temperature and then in 4% paraformalde-
hyde overnight at 4°C. Eyecups were washed with PBS several
times and embedded in paraffin. Five-­micrometer sections were
cut and placed in a heat plate at 56°C for 30 min. Following de-
paraffinization and rehydration, sections were stained with an
H&E stain kit according to the protocol provided by Vector,
Brussels, Belgium). Then, they were dehydrated in alcohol and
cleared in xylene. The image of the section was captured by an
Olympus BX53M microscope in three random fields per section
and used to assess pathological damage. All photo adjustments
were carried out equally across the sections. The number of
RGCs was counted.
Statistical analysis
The data were expressed as means ± SD. Statistical analysis of the P
value was conducted using an open-­source MATLAB code. Signifi-
cant differences were assessed using the unpaired Student’s t test and
indicated as n.s. P > 0.05, *P > 0.01, **P > 0.001, ***P > 0.0001,
and ****P < 0.0001.
11 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Supplementary Materials
This PDF file includes:
Supplementary Text
Figs. S1 to S20
REFERENCES AND NOTES
1. J. C. Downs, C. A. Girkin, Lamina cribrosa in glaucoma. Curr. Opin. Ophthalmol. 28,
113–119 (2017).
2. J.-­A. Kim, S. H. Lee, D. H. Son, T.-­W. Kim, E. J. Lee, M. J. A. Girard, J. M. Mari, Morphologic
changes in the lamina cribrosa upon intraocular pressure lowering in patients with
normal tension glaucoma. Invest. Ophthalmol. Vis. Sci. 63, 23 (2022).
3. R. N. Weinreb, C. K. S. Leung, J. G. Crowston, F. A. Medeiros, D. S. Friedman, J. L.
Wiggs, K. R. Martin, Primary open-­angle glaucoma. Nat. Rev. Dis. Primers. 2, 16067
(2016).
4. J. Kountouras, C. Zavos, D. Chatzopoulos, Primary open-­angle glaucoma:
Pathophysiology and treatment. Lancet 364, 1311–1312 (2004).
5. Z. Nie, J. W. Kwak, M. Han, J. A. Rogers, Mechanically active materials and devices for
bio-­interfaced pressure sensors—A review. Adv. Mater. 2205609 (2023).
6. M. Kareem, T. Agbaje, F. Alam, H. Butt, A review of thin films used in smart contact lenses.
Adv. Funct. Mater. 2300363 (2023).
7. Z. Rao, F. Ershad, A. Almasri, L. Gonzalez, X. Wu, C. Yu, Soft electronics for the skin:
From health monitors to human–machine interfaces. Adv. Mater. Technol. 5, 2000233
(2020).
8. B. Tian, C. M. Lieber, Nanowired bioelectric interfaces. Chem. Rev. 119, 9136–9152
(2019).
9. K. W. Cho, S.-­H. Sunwoo, Y. J. Hong, J. H. Koo, J. H. Kim, S. Baik, T. Hyeon, D.-­H. Kim, Soft
bioelectronics based on nanomaterials. Chem. Rev. 122, 5068–5143 (2022).
10. A. Rodrigo-­Navarro, S. Sankaran, M. J. Dalby, A. del Campo, M. Salmeron-­Sanchez,
Engineered living biomaterials. Nat. Rev. Mater. 6, 1175–1190 (2021).
11. M. H. Kang, G. J. Lee, J. H. Yun, Y. M. Song, NFC-­based wearable optoelectronics working
with smartphone application for untact healthcare. Sensors 21, 878 (2021).
12. J. D. Stein, A. P. Khawaja, J. S. Weizer, Glaucoma in adults—Screening, diagnosis, and
management: A review. JAMA 325, 164–174 (2021).
13. J. Kim, M. Kim, M.-­S. Lee, K. Kim, S. Ji, Y.-­T. Kim, J. Park, K. Na, K.-­H. Bae, H. Kyun Kim, F. Bien,
C. Young Lee, J.-­U. Park, Wearable smart sensor systems integrated on soft contact lenses
for wireless ocular diagnostics. Nat. Commun. 8, 14997 (2017).
14. J. Kim, J. Kim, M. Ku, E. Cha, S. Ju, W. Y. Park, K. H. Kim, D. W. Kim, P.-­O. Berggren, J.-­U. Park,
Intraocular pressure monitoring following islet transplantation to the anterior chamber
of the eye. Nano Lett. 20, 1517–1525 (2020).
15. Z. Liu, G. Wang, C. Ye, H. Sun, W. Pei, C. Wei, W. Dai, Z. Dou, Q. Sun, C.-­T. Lin, Y. Wang,
H. Chen, G. Shen, An ultrasensitive contact lens sensor based on self-­assembly
graphene for continuous intraocular pressure monitoring. Adv. Funct. Mater. 31,
2010991 (2021).
16. J. Kim, J. Park, Y.-­G . Park, E. Cha, M. Ku, H. S. An, K.-­P. Lee, M.-­I. Huh, J. Kim, T.-­S. Kim,
D. W. Kim, H. K. Kim, J.-­U. Park, A soft and transparent contact lens for the wireless
quantitative monitoring of intraocular pressure. Nat. Biomed. Eng. 5, 772–782
(2021).
17. Y. Ye, Y. Ge, Q. Zhang, M. Yuan, Y. Cai, K. Li, Y. Li, R. Xie, C. Xu, D. Jiang, J. Qu, X. Liu, Y. Wang,
Smart contact lens with dual-­sensing platform for monitoring intraocular pressure and
matrix metalloproteinase-­9. Adv. Sci. 9, e2104738 (2022).
18. J. B. Jonas, T. Aung, R. R. Bourne, A. M. Bron, R. Ritch, S. Panda-­Jonas, Glaucoma. Lancet
390, 2183–2193 (2017).
19. R. N. Weinreb, T. Aung, F. A. Medeiros, The pathophysiology and treatment of glaucoma:
A review. JAMA 311, 1901–1911 (2014).
20. L. S. Abrams, S. Vitale, H. D. Jampel, Comparison of three tonometers for measuring
intraocular pressure in rabbits. Invest. Ophthalmol. Vis. Sci. 37, 940–944 (1996).
21. P. Hallberg, K. Santala, C. Lindén, O. A. Lindahl, A. Eklund, Comparison of Goldmann
applanation and applanation resonance tonometry in a biomicroscope-­based in vitro
porcine eye model. J. Med. Eng. Technol. 30, 345–352 (2006).
22. J. L. Hernández-­Verdejo, M. A. Teus, G. Bolivar, Simultaneous measurement of intraocular
pressure in the anterior chamber and the vitreous cavity. Acta Ophthalmol. 88,
e265–e268 (2010).
23. Y. J. Kim, K. S. Choi, Pressure difference between the anterior chamber and the vitreous
cavity in eyes with pupillary block. Curr. Eye Res. 40, 572–578 (2015).
24. N. Feltgen, D. Leifert, J. Funk, Correlation between central corneal thickness,
applanation tonometry, and direct intracameral IOP readings. Br. J. Ophthalmol. 85,
85–87 (2001).
25. J. Jang, H. Kim, S. Ji, H. J. Kim, M. S. Kang, T. S. Kim, J. Won, J.-­H. Lee, J. Cheon, K. Kang,
W. B. Im, J.-­U. Park, Mechanoluminescent, air-­dielectric MoS2 transistors as active-­matrix
pressure sensors for wide detection ranges from footsteps to cellular motions. Nano Lett.
20, 66–74 (2020).
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
26. T. Saha, T. Songkakul, C. T. Knisely, M. A. Yokus, M. A. Daniele, M. D. Dickey, A. Bozkurt,
O. D. Velev, Wireless wearable electrochemical sensing platform with zero-­power osmotic
sweat extraction for continuous lactate monitoring. ACS Sens. 7, 2037–2048 (2022).
27. J. Jang, S. Ji, G. K. Grandhi, H. B. Cho, W. B. Im, J.-­U. Park, Multimodal digital x-­ray scanners
with synchronous mapping of tactile pressure distributions using perovskites. Adv. Mater.
33, e2008539 (2021).
28. Y. Dai, S. Dai, N. Li, Y. Li, M. Moser, J. Strzalka, A. Prominski, Y. Liu, Q. Zhang, S. Li, H. Hu,
W. Liu, S. Chatterji, P. Cheng, B. Tian, I. McCulloch, J. Xu, S. Wang, Stretchable redox-­active
semiconducting polymers for high-­performance organic electrochemical transistors. Adv.
Mater. 34, e2201178 (2022).
29. Y.-­G. Park, G.-­Y. Lee, J. Jang, S. M. Yun, E. Kim, J.-­U. Park, Liquid metal-­based soft
electronics for wearable healthcare. Adv. Healthc. Mater. 10, e2002280 (2021).
30. Y.-­G. Park, I. Yun, W. G. Chung, W. Park, D. H. Lee, J.-­U. Park, High-­resolution 3D printing for
electronics. Adv. Sci. 9, e2104623 (2022).
31. B. B. Bhatt, L. Kumar, A. Kushwaha, D. Gupta, An ultra-­compressible piezoresistive strain
and pressure sensor based on RGO-­CNT-­Melamine foam composite for biomedical
sensing. Sens. Actuators A 331, 112875 (2021).
32. Y. Jeong, J. Park, J. Lee, K. Kim, I. Park, Ultrathin, biocompatible, and flexible pressure sensor
with a wide pressure range and its biomedical application. ACS Sens. 5, 481–489 (2020).
33. Q. Yang, S. Lee, Y. Xue, Y. Yan, T. Liu, S. Kang, Y. J. Lee, S. H. Lee, M. Seo, D. Lu, J. Koo,
M. R. MacEwan, R. T. Yin, W. Z. Ray, Y. Huang, J. A. Rogers, Materials, mechanics designs,
and bioresorbable multisensor platforms for pressure monitoring in the intracranial
space. Adv. Funct. Mater. 30, 1910718 (2020).
34. H. Ouyang, Z. Li, M. Gu, Y. Hu, L. Xu, D. Jiang, S. Cheng, Y. Zou, Y. Deng, B. Shi, W. Hua,
Y. Fan, Z. Li, Z. Wang, A bioresorbable dynamic pressure sensor for cardiovascular
postoperative care. Adv. Mater. 33, e2102302 (2021).
35. E. M. Hoffmann, F.-­H. Grus, N. Pfeiffer, Intraocular pressure and ocular pulse amplitude using
dynamic contour tonometry and contact lens tonometry. BMC Ophthalmol. 4, 4 (2004).
36. J. O. Lee, H. Park, J. Du, A. Balakrishna, O. Chen, D. Sretavan, H. Choo, A microscale optical
implant for continuous in vivo monitoring of intraocular pressure. Microsyst. Nanoeng. 3,
17057 (2017).
37. K. H. Kim, J. O. Lee, J. Du, D. Sretavan, H. Choo, Real-­time in vivo intraocular pressure
monitoring using an optomechanical implant and an artificial neural network. IEEE Sens.
J. 17, 7394–7404 (2017).
38. C. D. Swerdlow, W. H. Olson, M. E. O’Connor, D. M. Gallik, R. A. Malkin, M. Laks,
Cardiovascular collapse caused by electrocardiographically silent 60-­Hz intracardiac
leakage current. Circulation 99, 2559–2564 (1999).
39. M. M. Laks, R. Arzbaecher, J. J. Bailey, D. B. Geselowitz, A. S. Berson, Recommendations for
safe current limits for electrocardiographs. A statement for healthcare professionals from
the Committee on Electrocardiography, American Heart Association. Circulation 93,
837–839 (1996).
40. H.V. Pipberfer, R. C. Arzbaecher, A. S. Berson, S. A. Briller, D. B. Geselowitz, L. G. Horan,
P. Rautaharju,O. H. Schmitt,Amendment of recommendations for standardization of
specifications for instruments in electrocardiography and vectorcardiography concerning
safety and electrical shock hazards. IEEE Trans. Biomed. Eng. 20, 140–141 (1973).
41. P. Szurman, K. Gillmann, A.-­M. Seuthe, H. B. Dick, E. M. Hoffmann, A. Mermoud,
M. J. Mackert, R. N. Weinreb, H. L. Rao, K. Mansouri, EYEMATE-­SC Study Group, EYEMATE-­SC
Trial: Twelve-­month safety, performance, and accuracy of a suprachoroidal sensor for
telemetric measurement of intraocular pressure. Ophthalmology 130, 304–312 (2023).
42. P. Szurman, K. Mansouri, H. B. Dick, A. Mermoud, E. M. Hoffmann, M. Mackert,
R. N. Weinreb, H. L. Rao, A.-­M. Seuthe, EYEMATE-­SC Study Group, Safety and performance
of a suprachoroidal sensor for telemetric measurement of intraocular pressure in the
EYEMATE-­SC trial. Br. J. Ophthalmol. 107, 518–524 (2023).
43. K. Nagae, H. Sawamura, M. Aihara, Investigation of intraocular pressure of the anterior
chamber and vitreous cavity of porcine eyes via a novel method. Sci. Rep. 10, 20552
(2020).
44. C. W. McMonnies, Intraocular pressure spikes in keratectasia, axial myopia, and glaucoma.
Optom. Vis. Sci. 85, 1018–1026 (2008).
45. E. Moisseiev, A. Loewenstein, G. Yiu, The suprachoroidal space: From potential space to a
space with potential. Clin. Ophthalmol. 10, 173–178 (2016).
46. P. Emami-­Naeini, G. Yiu, Medical and surgical applications for the suprachoroidal space.
Int. Ophthalmol. Clin. 59, 195–207 (2019).
47. J. A. Kammer, K. M. Mundy, Suprachoroidal devices in glaucoma surgery. Middle East Afr.
J. Ophthalmol. 22, 45–52 (2015).
48. T. De Francesco, I. I. K. Ahmed, Surgical augmentation of the suprachoroidal space:
A novel material and implant. Clin. Ophthalmol. 17, 2483–2492 (2023).
49. J. P. Berdahl, M. P. Fautsch, S. S. Stinnett, R. R. Allingham, Intracranial pressure in primary
open angle glaucoma, normal tension glaucoma, and ocular hypertension: A case–
control study. Invest. Ophthalmol. Vis. Sci. 49, 5412–5418 (2008).
50. J. Shin, Z. Liu, W. Bai, Y. Liu, Y. Yan, Y. Xue, I. Kandela, M. Pezhouh, M. R. MacEwan, Y. Huang,
W. Z. Ray, W. Zhou, J. A. Rogers, Bioresorbable optical sensor systems for monitoring of
intracranial pressure and temperature. Sci. Adv. 5, eaaw1899 (2019).
12 of 13
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
51. K. B. Evensen, P. K. Eide, Measuring intracranial pressure by invasive, less invasive or
non-­invasive means: Limitations and avenues for improvement. Fluids Barriers CNS 17,
34 (2020).
Acknowledgments
Funding: This work was supported by the Ministry of Science and ICT (MSIT), the Ministry of
Trade, Industry and Energy (MOTIE), the Ministry of Health & Welfare, and the Ministry of Food
and Drug Safety of Korea through the National Research Foundation (2023R1A2C2006257);
Nano Material Technology Development Program (2021M3D1A2049914); ERC Program
(2022R1A5A6000846 and 2020R1A5A1019131); the Technology Innovation Program
(20013621, Center for Super Critical Material Industrial Technology); the Korea Medical Device
Development Fund grant (RMS 2022-­11-­1209 / KMDF RS-­2022-­00141392); and the Korea
Initiative for fostering University of Research and Innovation (KIURI) Program of the National
Research Foundation (NRF) (NRF-­2020M3H1A1077207). Also, we thank the financial support of
Seo et al., Sci. Adv. 10, eadk7805 (2024) 7 February 2024
the Institute for Basic Science (IBS-­R026-­D1). Author contributions: H.S. and Y.-­M.H. carried
out the experiments, analyzed the data, and wrote the manuscript. W.G.C., W.P., and J.L.
contributed to the fabrication of the device and the rabbit experiments. H.K.K. and S.H.B.
contributed to the rabbit experiments. D.W.K. contributed to the planning of the project and
the animal experiments. J.-­U.P. oversaw all phases of this research. All authors discussed and
commented on the manuscript. Competing interests: The authors declare that they have no
competing interests. Data and materials availability: All data needed to evaluate the
conclusions in the paper are present in the paper and/or the Supplementary Materials.
Submitted 10 September 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk7805
13 of 13
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

M AT E R I A L S S C I E N C E Copyright © 2024 The

Authors, some rights
ForceGen: End-­to-­end de novo protein generation reserved; exclusive
licensee American
based on nonlinear mechanical unfolding responses Association for the
Advancement of
using a language diffusion model Science. No claim to
original U.S.
Government Works.
Bo Ni1, David L. Kaplan2, Markus J. Buehler1,3* Distributed under a
Creative Commons
Through evolution, nature has presented a set of remarkable protein materials, including elastins, silks, keratins Attribution
and collagens with superior mechanical performances that play crucial roles in mechanobiology. However, going NonCommercial
beyond natural designs to discover proteins that meet specified mechanical properties remains challenging. Here, License 4.0 (CC BY-­NC).
we report a generative model that predicts protein designs to meet complex nonlinear mechanical property-­design
objectives. Our model leverages deep knowledge on protein sequences from a pretrained protein language model
and maps mechanical unfolding responses to create proteins. Via full-­atom molecular simulations for direct validation,
we demonstrate that the designed proteins are de novo, and fulfill the targeted mechanical properties, including
unfolding energy and mechanical strength, as well as the detailed unfolding force-­separation curves. Our model
offers rapid pathways to explore the enormous mechanobiological protein sequence space unconstrained by biological
synthesis, using mechanical features as the target to enable the discovery of protein materials with superior
mechanical properties.

INTRODUCTION
Proteins present an elegant yet complex and rich design platform.
The various functions and outstanding properties of proteins can be
attributed to the folded three-­dimensional (3D) structures, encoded
by the underlying one-­dimensional (1D) primary sequences consisting
of about 20 naturally occurring amino acids (1). Through evolution,
nature has demonstrated great success in “designing” proteins as a
set of critical building blocks that constitute fundamental functions
of all life, and specifically remarkable biomaterials, ranging from the
structural hierarchies in collagens, complex assemblies such as silk,
to tissue assemblies such as muscle and skin (2–5). In these various
tissues and systems, the detailed mechanical signature—often, their
response to mechanical pulling—is an essential feature for mechano-
biology (6–9).
At the same time, there remains vast design space of mechanically
optimized proteins yet unexplored by nature given the enormous
possibilities of protein sequences (10). Hence, inspired by nature, dis-
covering de novo proteins may unlock potentially unprecedented
properties and functions (3, 10–16). However, this enormous design
space and costs associated with experimental testing present great
challenges in finding effective tools to design de novo protein sequences
that meet a set of interested functions or properties (14–19).
In recent years, the development of deep learning approaches and
their applications to proteins have provided fast avenues for protein
study and design. For forward problems focused on structure identifi-
cation, deep learning–based tools such as AlphaFold2 (20) and
RoseTTAFold (21) represent a breakthrough in achieving competitive
accuracy with experimental methods in predicting 3D folded struc-
tures based on protein sequences at a much reduced cost. (22) Built
1
Laboratory for Atomistic and Molecular Mechanics (LAMM), Massachusetts Institute
of Technology, 77 Massachusetts Ave., Cambridge, MA 02139, USA. 2Department of
Biomedical Engineering, Tufts University, Medford, MA 02155, USA. 3Center for Compu-
tational Science and Engineering, Schwarzman College of Computing, Massachusetts
Institute of Technology, 77 Massachusetts Ave., Cambridge, MA 02139, USA.
*Corresponding author. Email: mbuehler@​mit.​edu
upon these approaches, other protein folding tools [e.g., Omegafold
(23), RGN2 (24), HelixFold-­single (25), and ESMFold (26)] have been
exploring the application of large language models. By removing
dependence on multiple sequence alignments (MSAs) as the input,
improvements in further reducing computational costs and achieving
better predictions for orphan and rapidly evolving proteins have
been demonstrated (23, 24, 26, 27).
End-­to-­end models based on deep learning that predict various
structural features [e.g., secondary structure type and content (28–
33), binding sites (34), and surfaces (35)] and properties [e.g., solu-
bility (16, 36, 37), melting temperature (38), natural vibrational
frequencies (39, 40), and strength (41)] for given sequences have
also been reported. At the sample time, the inverse design of de novo
proteins that meet desired structural or property features presents a
more challenging task. On one hand, facing the enormous sequence
space, search algorithms teamed with efficient deep learning–based
forward predictors (30, 42, 43) may still suffer from inefficient explora-
tion and the design accuracy and varieties of the discovered sequences
are not easily controlled. On the other hand, recently emergent genera-
tive models (44–49) provide a direct map from the desired character-
istics to potential designs and are becoming an emerging paradigm
for various materials research and design (50–55), including proteins.
For example, using an attention-­based diffusion model trained on
secondary structure data, de novo protein sequences can be generated
based on secondary structure design objectives (56). However, these
generative design models often focus on structural level design [such
as secondary structures (56) or detailed protein backbone shapes
(57–60)]. In contrast, development of generative models aimed at
end-­to-­end design from property of interest to protein sequence
remains rare (61).
Here, we focus on nanomechanical properties (62–65) of proteins.
Thanks to the advent of single-­molecule technology (66) [e.g., atomic
force microscopy (AFM) (67, 68), optical tweezers (69, 70), and
magnetic tweezers (71, 72)], the measurement of protein unfolding
under an applied mechanical force provides a unique molecular basis
for understanding protein deformation (elasticity/plasticity) and

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 1 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

fracture (64) and can play key roles in affecting some macroscopic
mechanical properties of protein-­based materials due to the inherent
structural hierarchy. For example, via experimental measurements
and theoretical analysis, it has been demonstrated that toughness of
synthetic protein hydrogels can be correlated to the mechanical un-
folding responses of the protein molecules and that mechanically
strong folded proteins can result in tough hydrogel designs (73).
Therefore, generating de novo proteins that meet desired mechanical
unfolding responses can represent a key molecular level design step
in protein-­based material designs. Compared with previous protein
design cases, this problem presents some unique challenges. First,
this is a property-­to-­sequence end-­to-­end design task bypassing the
structure level, which is expected to be more difficult than previous
structure-­to-­sequence design tasks (56, 60). Second, the available or
affordable data on mechanical unfolding responses of known proteins
(74) are rare when compared to those for protein structures (75) or
sequences (76). Besides mechanical properties, we expect that these
two challenges are also shared by many other property-­to-­sequence
design tasks in proteins.
To address this problem, here, we combine an attention-­based
(77) diffusion model (56) with a pretrained large language model
(26) for proteins to construct a generative deep learning model that
predicts amino acid sequences and 3D protein structures based on
mechanical unfolding responses as design objectives. In a singular
workflow (Fig. 1), we start with performing a large series of full-­atom
molecular dynamics (MD) to simulate the mechanical unfolding
process of Protein Data Bank (PDB) (75) proteins and recording the
force responses (Fig. 1A). Then, we construct a protein language dif-
fusion model (pLDM) by translating the protein sequences into a
word probability latent space using a pretrained protein language
model (pLM) and training a diffusion model to learn the map between
sequence representations and the force-­separation responses (Fig. 1B).
At deployment, the trained pLDM predicts sequence candidates based
on the given unfolding force conditions and the integrated folding
algorithm (i.e., OmageFold) (23) determines the 3D structures of
the resulting sequences. For validation, we compare the designed
sequences with known proteins to analyze novelty (Fig. 1C) and test
the designed proteins using MD to compare the mechanical properties
and unfolding responses with input conditions. To prepare the design
pipeline for further experimental validation, other properties key to
experimental synthesis and testing, such as solvent accessible surface
area (SASA) (78), solubility, or stability (36, 79), can be estimated using
available predicting tools (36, 78, 79) to further screen for preferred
protein candidates (Fig. 1D). Through well-­controlled comparisons,
we demonstrate that our pLDM outperforms the vanilla diffusion
model with or without an iterative design scheme. Built upon the
property-­to-­sequence generation capability of our model and the
broad potential of protein materials in achieving superior mechanical

Fig. 1. Workflow of developing the end-­to-­end protein generation model. (A) Curating a PDB protein dataset on their mechanical properties by unfolding protein
chains by force in MD simulations. (B) Overview of the conditioned protein language diffusion model (pLDM) developed here. (C) Analyzing the novelty of the generated
protein sequences via protein-­protein BLAST tests. (D) Validating the mechanical properties of the designed protein candidates using folding tools and mechanical un-
folding tests and predicting other properties (e.g., solubility or stability) for further screening of the desired protein candidates.

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 2 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

properties, as well as other interesting properties (80–83) [e.g., optical
(82, 83), electronic (80), energy storage (81), etc.], we expect that
our end-­to-­end design model can be useful in numerous biological
and engineering applications for the property-­targeted generative
design of various protein material systems.
RESULTS
Full-­atom modeling of unfolding proteins by force
Inspired by single-­molecule force spectroscopy (67), we simulate
the unfolding process of protein chains under mechanical force to
understand their mechanical properties at the molecular level. As
shown in Fig. 2A, we start with PDB proteins with experimentally
chain, where we assume the average length of each amino acid is 3.6 Å
(86) and Lcon = N × 3.6 Å. Further details on the MD simulations
can be found in Materials and Methods. Movies of the unfolding
trajectory of some selected PDB protein examples can be found in
the Supplementary Materials.
In Fig. 2B, we smooth the raw force response (the red curve) to
get rid of high-­frequency fluctuations and get the unfolding response,
Fp(Lac), of the protein chain (the blue curve), from which we can
identify the toughness and strength of the protein molecule using
the unfolding energy T and the maximal value of force Fmax defined
as the following.
Lcon

∫
measured 3D structures. Using full-­atom MD with the CHARMM T= Fp dLac (1)
force field (84) and a generalized Born implicit solvent model (85),
we first relax the protein molecule at body temperature (i.e., 310 K)
to reach equilibrium conformation. Then, we stretch the protein
chain of N amino acids along the direction connecting the two chain Lac ≤ Lcon
Fmax = max {Fp } (2)
ends (i.e., a and b) by fixing one end and steering the other with a
spring (i.e., the segment between b and c in Fig. 2A) of a force con- To curate a dataset based on naturally existing proteins, we use
stant k = 0.5 kcal/(mol Å2) at a constant velocity v = 0.1 Å/ps. The the Biomolecule Stretching Database (BSDB) (74) as guidance and
pulling force, Fp, is recorded every 0.2 ps until the distance between select 7026 PDB proteins that have no gaps in their experimentally
two pulling ends, Lac, reaches the contour length, Lcon, of the protein determined structures and consist of no more than 128 amino acids.

Fig. 2. Mechanical unfolding of proteins and mechanical properties dataset curation. (A) Full-­atom simulation of mechanically unfolding a PDB protein chain using
steered MD. (B) Collecting (red data) and smoothing (blue data) the pulling force history during the whole unfolding process and converting it into a vector representation
(green triangle dots). (C) Collecting pulling force curves during mechanically unfolding for a large member of PDB proteins. (D) The distributions of the unfolding energy
(left) and maximal pulling force (right) for the PDB protein training set developed here.

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 3 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

Then, we collect their structures directly from the PDB (75) and ap-
ply the protocol above to test their mechanical unfolding responses. (3)
F�⃑p = {Fp (Lac
i
): i = 0, 1, 2, … , N}
An overview of the distributions of the unfolding responses and me-
chanical properties are shown in Fig. 2 (C and D). Specifically, in where N is the sequence length of the protein chain and we sample
Fig. 2D, the unfolding energy or toughness shows a bimodal distri- the pulling force when the distance between the pulling ends reaches
bution while the strength presents a unimodal one; in Fig. 2C, one i =i×L
Lac con ∕ N . For Lac that is smaller than the value of the initial
i

can observe that there exist various unfolding responses among pro-
teins. For example, the maximal force may appear as the peak in the
middle of unfolding process or near the end when reaching the con-
tour length, which may indicate very different deformation mecha-
nisms. An in-­depth study of these mechanical properties and their
dependence on the structural features and sequences for a large
number of proteins is of great research interest and should be car-
ried out in the near future. Further insight can also be obtained from
numerous experimental studies using AFM-­based force generative
outcomes during unfolding states of proteins. Here, we focus on ap-
plying these freshly collected data to develop protein generation
models. To efficiently label the pulling force responses during the
full unfolding process, we introduce a pulling force vector F�⃑p (green
triangles in Fig. 2B) to represent the full response as the following
equilibrium conformation, we simply define the force values as zeros.
Such a vector representation can adjust to the protein sequence
length automatically; that is, longer/shorter proteins with potentially
more/fewer unfolding details have more/fewer sampling points evenly
distributed. Next, we develop DL models to generate protein sequences
that meet the given mechanically unfolding responses represented
in terms of the pulling force vector F�⃑p.
pLDM and inverse design for mechanical signatures
To solve the conditioned protein design task, here we develop a
pLDM by combining a pretrained pLM (26) and an attention-­based
diffusion model (56). Figure 3A depicts an overview of the model
developed in the present work. The pLM (on the right of Fig. 3A) is
pretrained on large amounts of protein sequences data (76) to form

Fig. 3. Overview of the pLDM. (A) Structure of the developed model, pLDM. It combines a protein language model (pLM) pretrained on large protein sequence data and
a trainable attention-­based diffusion model. We use the pretrained pLM to translate protein sequence representations between the tokenized sequence space and the
word probability latent space. The diffusion model, with a building block of a 1D U-­net, is trained to predict the noise added at each diffusion steps, thus gradually removing
them to generate meaningful sequence representations at deployment. (B) Depiction of the 1D U-­net architecture that translates an input Ii into an output Oi under a
condition set Ci. The model features 1D convolutional layers, as well as self-­/cross-­attention layers as shown on the right.

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 4 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

internal representations that better understand not only sequences
but also structures and properties of proteins (33). We leverage this
knowledge by applying pLMs to translate protein sequences from
the tokenized sequence space into the word probability latent space.
Then, we train a diffusion model developed in the previous work to
operate in this probability latent space. The diffusion model is built
upon a 1D U-­Net architecture with attention mechanisms (Fig. 3B).
At deployment, starting with the given condition (on the left of
Fig. 3A) and random signal seed, the diffusion model predicts and
removes the noise at each step and produces meaningful sequence
probability tensors, which are then translated back into protein
sequence using the fixed pLM. There are multiple choices for the
pretrained pLM and larger pLMs require higher computing resource
and cost (26, 33, 87). For computational efficiency, in the following,
we focus on the results that adopt a medium-­sized pretrained model
with 150 million parameters from the ESM-­2 series (we find that
this yields improved performances as is shown in the discussion in
the next subsection) (88).
Once the model has been trained, we demonstrate the performance
of the developed pLDM by testing it with various mechanical un-
folding responses, including those that come from naturally existing
proteins and those that are de novo. The generated sequences are
folded into 3D structures using OmegaFold (23) and then undergo
the same mechanical unfolding tests using full-­atom MD simulations.
With protein BLAST (89) test and comparing pulling force responses
with the input, we examine the novelty of the generated sequences
and the accuracy of the protein design.
For protein design with the mechanical unfolding responses that
correspond to naturally existing proteins, we test the model with the
pulling force records of PDB proteins in the test set, with which the
model has not been trained. Figure 4 shows some examples of
the designed proteins and their mechanical unfolding responses. In
terms of the design target, the conditioned pulling force paths (red
curves) in Fig. 4 (A to F) represent a variety of different patterns,
including simple ones that show that the pulling force nearly keeps
increasing during the unfolding process (Fig. 4D), the examples that

Fig. 4. Results for protein generation based on mechanical unfolding responses of naturally existing proteins. (A) to (F) show a variety of representative cases of
different unfolding force paths (red curves), including the one that nearly keeps increasing (D), the ones that show local peaks during an overall increasing trend (A, B, and
F), the one that meets an oscillating plateau and then increases (E), and the one that reaches a high peak in the early stage (C). The proteins generated by our model
demonstrate pulling force patterns (blue curves) that follow the trend of design objectives. Because of the complex and highly oscillating nature of pulling force response
during mechanical unfolding of proteins, we use R2 value and relative L2 error (listed in each panel) to measure the accuracy of the design in following the overall trend
and quantitative values. Corresponding to the various pulling force responses, the generated proteins show a variety of structures, including high α helix content (A, B,
and F), a mix of β sheets and random coils (C), a mix of an α helix segment and random coils (D) and a mix of β sheets and α helices (E).

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 5 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

show local peaks in early stages of unfolding during an overall in-
creasing trend of the pulling force (Fig. 4, A, B, and F), the ones that
reach an oscillating plateau and then increase (Fig. 4E), and those
that achieve high peak in the initial stage of unfolding (Fig. 4C).
Despite this complexity of protein unfolding scenarios, and the
oscillating nature of pulling force, the proteins generated by our
model demonstrate pulling force responses (blue curves) that in
general closely follow the design objectives. We use multiple metrics,
including R2 and relative L2 error (see Materials and Methods for
details), to measure the accuracy of design in meeting the overall
trend as well as quantitative values of the mechanical responses.
Corresponding to the various patterns of pulling force responses,
the generated proteins also show a variety of internal structures. For
example, a relatively simple combination of an α helix segment and
random coils without complex entanglement or close spatial packing
in Fig. 4D produces an unfolding process with consistently increasing
pulling forces with few oscillations. An entangled mix of α helices
and β sheets in Fig. 4E yields an unfolding pulling force path with a
plateau region of strong oscillations. Unfolding of the β sheets in
Fig. 4C can be related to the high local peak in pulling force history.
In Fig. 4 (A, B, and F), α helix segments of different lengths and
spatial arrangements can produce similar trends but different quanti-
tative pulling force records. More details on the mechanical unfolding
process of some of these cases, including cases A, C and E, can be
found in the MD trajectory movies in the Supplementary Materials.
On the novelty of these generated proteins, we apply basic local
alignment search tool (BLAST) analysis (89) to the predicted amino
acid sequences to access whether, and to what extent, they represent
de novo sequences or closely related forms of known proteins. Table 1
shows the results of the BLAST analysis for the various cases listed
in Fig. 4. We find that even though the input design targets are from
existing PDB proteins, many of the generated protein sequences
(cases shown in Fig. 4, A to D and F) do not match any sequences in
the database of known proteins with standard BLSAT analysis (i.e.,
returning “no significant similarity found” in protein BLAST test) and
are de novo ones. The model can also produce sequences (e.g., case
E in Fig. 4) that show some similarity to the existing proteins. However,
the most similar example found (i.e., 8CH0) is not included in the
training and testing set. While the model is only trained on a very
small portion of PDB proteins, with the pulling force corresponding
to existing PDB proteins as an input, we expect the possibility of the
model “rediscovering” sequences that show some similarities to the

Table 1. Results of the BLAST analysis and predicted solubility for the various generated proteins (from Fig. 4) based on existing mechanical unfolding
responses. Given the pulling force vectors of existing proteins as the design condition, the model still shows high probability in predicting sequences that show
little similarity to existing proteins as can be seen from the BLAST results (A to D and F). For other cases, sequences with some similarity to known proteins can
be predicted (E). The predicted solubility is scaled with the experimental dataset with a population average of 0.45. The listed solubility values are all larger than
0.45, thus predicting to have a higher solubility than the average soluble E. coli proteins in the experimental dataset (90).
Case Sequence BLAST result: the sequence producing the most signifi- Predicted solubility
cant alignment
Among PDB proteins Beyond PDB proteins

A MLIEGTQELIHQKLAKGKT- – No significant similarity 0.939

VLVQRYVAKGLQVDDNTEDL- found (NSSF)
LANAKNYLNPDQIERSIAYAQK-
IEEMEGDDMFKVALV
B MKKKVRMEQNEQKKQVY- – NSSF 0.876
QELNDKVENDEALAPKS-
VALYIAALKEKEEAGKIPHHF-
NLLERLKLTITSCRFFLLKIQN-
NDTKLQKRRKFIDETIQLAREI-
YEKQDNK
C MGKITPVVLAGGKQK- – NSSF 1.000
EDEETLDGGEILTKDG-
KTLKLISDAQVAVMN-
VKQVQEGTYEGSQVIEEDG-
VRGNYVSYVGK
D GSSGSSGRDVTQQTNKCCR- – NSSF 0.676
RCSRKPHCCIKAWRPRSSD-
LYYHEKHTHSGPSSG
E MNTPEHMTAVVQRYVAAL- 100% query cover, 73.6% – 0.774
NGGDLDGIVALFADDATVED- identical with 8CH0
PVGFQNVSGKAADANFYESP-
GFLDLVKALTGPVRAFGNEK-
FFAMIVFFEYEGTKTVVAGI-
DHIRFNGAGKVVSMRAYF-
DEKNIHASA
F MPWHHHGSSGLVQTG- – NSSF 0.606
MAATGLKDFIVEAYPKKPD-
DIIKVCRSEPSAGYWWCED-
VQNEVKQKCLSKKQRQVKAQ

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 6 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

known proteins. Further measures may be utilized to boost the novelty
of design for such cases, including multi-­shot design and selecting
the best one based on BLAST results. In current work, we focus on
understanding the performance of the current model.
Besides examining individual cases, we also show the distributions
of design accuracy and novelty for a larger number of testing cases.
Figure 5 demonstrates the results of 187 generated proteins based on
various pulling force conditions from the standalone test set. On the
mechanical unfolding responses, the R2 and relative L2 error between
the measured pulling force vectors of the generated proteins and the
input conditions among cases show unimodal distributions with
median values of 0.56 and 0.36 (Fig. 5, A and B), respectively. The
distributions indicate that for many of the cases, the designed proteins
follow the input conditions in terms of the trend and value reasonably
well during the whole mechanical unfolding process. However, as
demonstrated in individual cases (Fig. 4), it remains challenging for
designed proteins to precisely follow the input pulling force values at
each unfolding step. This can also be seen by comparing components
of pulling forces of all cases together in Fig. 5C. While conditioned
input values of pulling force components and the measured ones
based on the generated proteins in general share the same trend
(that is, the distribution centers at the y = x dashed line as the visual
guide in Fig. 5C), the finite width of the distribution cloud deviating
away from the ideal case indicates that for individual components,
there could be considerable mismatches.
The limited component-­wise accuracy demonstrates the difficulty
and challenge of designing proteins based on detailed mechanical
unfolding responses, even with the current model. At the same time,
the proteins generated by our model still show reasonable agreement
between the achieved and the conditioned mechanical properties, in-
cluding toughness (Fig. 5D) and strength (Fig. 5E). Strength defined
as the maximum of the pulling force shows an R2 value of 0.41
(Fig. 5E), slightly smaller than that of the pulling force components
(0.54 as listed in Fig. 5C). At the same time, an R2 value of 0.93,
much higher than that of pulling force components (Fig. 5C), was
observed for toughness (Fig. 5D), which is defined as the unfolding
energy over the whole unfolding process (Eq. 1). This difference in
R2 values indicates that when the entire unfolding process is considered,
the component-­wise error tends to cancel each other and the de-
signed proteins follow the input conditions in terms of toughness
more sensitively. On the novelty of the designed proteins, Fig. 5F
shows a bimodal distribution of the highest percent identity found
via protein BLAST analysis for all the generated sequences. The
highest peak (on the left in Fig. 5F) corresponds to the cases where
the generated proteins have little similarity to the existing/known
ones and are totally de novo. There also exists the other weaker peak

Fig. 5. Overall quality of generating proteins based on mechanical unfolding responses that correspond to naturally existing proteins in the test set. We test the
model with mechanical unfolding responses from 187 proteins in the standalone test set. On the pulling force response, (A) and (B) show the distributions of R2 (A) and
relative L2 error (B) for comparing the pulling force response of each designed protein with the input condition while (C) shows the comparison in terms of pulling force
components for all testing cases. On the overall mechanical properties, (D) and (E) compare the designed proteins with input conditions in terms of unfolding energy (i.e.,
toughness) and maximal pulling force (i.e., strength). On the novelty of the designed sequences, (F) shows the distribution of the highest percent identity found via BLAST
test.

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 7 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

on the right for cases in which the generated proteins are similar to
the known ones. The bimodal distribution echoes the result of indi-
vidual cases listed in Table 1, and the relative height of the two peaks
indicates that our model has a stronger tendency in generating de
novo sequence designs.
To further boost the novelty of designed sequences, different mea-
sures could be taken, including how the model is applied and how
the input conditions are constructed. Here, we discuss one possibility
of using de novo mechanical unfolding responses, i.e., pulling force
vectors that do not correspond to existing proteins, as the input.
There is still a lack of complete rules on how to identify or construct
physically achievable pulling force vectors for all possible amino acid
sequences. To explore the de novo mechanical unfolding responses,
here we start with existing pulling force vectors and mute them using
a mixing scheme. As shown in Fig. 6A, two pulling force vectors from
PDB proteins 2AAN and 1KN7 were chosen and show different
patterns during the unfolding process, one undergoing an oscillating
plateau in the early stages while the other continues to increase.
These differences in pulling force can be attributed to the different
internal protein structures and sequence length. 2AAN shows a
high β sheet content in a closely packed conformation. In compari-
son, 1KN7 has a mix of α helix segments and unstructured coils
with a more open spatial arrangement. We mix the pulling force
vectors of each of the proteins with respect to the same normalized
pulling end gap at different ratios (i.e., 1/3 to 2/3 for mix 1 and 2/3
to 1/3 for mix 2) and then use these as the de novo mechanical un-
folding response to generate proteins with our model. In addition,
in our model, the length of the generated sequence (i.e., N) can be
controlled through the length of the input pulling force vector (i.e.,
N + 1). Here, we intentionally choose N = 99, which is different
from that of both base proteins when constructing the de novo input
pulling force vectors. The results of the proteins designed with these
de novo pulling force vectors are shown in Fig. 6 (B and C). Similar
to previous designs listed in Fig. 4, the mechanical unfolding re-
sponses of the designs follow the input conditions, even though this
time they come from a mix of proteins of different structures and
mechanical properties. Moreover, the designed proteins adopt inter-
nal structures of a mix of α helices with different compactness,
which show little similarity to those of the base proteins. Another
set of examples is shown in Fig. 6 (D to F) with base proteins of dif-
ferent internal structures. Again, the designed proteins fulfill the
targeted de novo mechanical unfolding responses reasonably well.
Comparing the structures of the generated proteins including
the ones in Fig. 6, one can raise intriguing questions about the un-
derlying relation between protein structure and mechanical properties.
For example, as seen in in Fig. 6, none of the generated proteins have
β sheets even though some of the base proteins do. Instead, many of
the designed proteins are composed of α helices and β turns and
packed in a distinct conformation. As the unfolding force response
may be related to individual secondary structures unfolding, as well
as the interactions between secondary structures, one may examine
whether the model has learned to search for certain topological features,
like different compactness, to generate the desired force-­separation
patterns. To investigate this possibility, we calculate the SASA using
dr-­sasa (78) based on the relaxed protein structure. This helps us to
assess the compactness of the folded structures for both generated
and base proteins. The SASA and their values normalized with
respective to the sequence length are listed in table S1. Among the
base proteins in the mixing pair (Fig. 6, A and D), the one with a
higher unfolding force level (base protein 1 in Fig. 6A or base pro-
tein 3 in Fig. 6D) includes closely packed β sheet structures with a
normalized SASA smaller than that with a lower unfolding force
(base protein 2 in Fig. 6A or base protein 4 in Fig. 6D), and with
loosely packed internal structures. Among the mixtures (Fig. 6, B
and C, or Fig. 6, E and F), similar trends are observed: A higher
unfolding force level indicates a relatively smaller SASA and more

Fig. 6. Results for protein generation based on de novo mechanical unfolding responses. For de novo design input, in (A) and (D), we construct mechanical unfolding
responses by mixing existing ones at a different ratio and changing the targeted sequence length. Here, we intentionally choose pulling forces of different patterns, one
that monotonically increases (base 2 in A and base 4 in D) and the other that first reaches a plateau then increases (base 1 in A and base 3 in D) as the bases to mix. By
choosing the mixing ratio, the de novo inputs cover a transition between these two pulling force patterns. (B), (C), (E), and (F) show the results of the generated proteins.
On the pulling force, the generated proteins follow the design input closely in terms of both the overall pattern and quantitative values throughout the unfolding process.
On the structure, interestingly, the generated proteins show a mix of α helix and random coil segments, which is different from some of the base proteins of high β sheet
content. On the sequence length, the generated proteins have a length that is different from both base proteins.

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 8 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

closely packed internal structures (Fig. 6, B or E). However, between
the mixes and the base pairs of proteins, no clear relation in terms of
SASA or normalized SASA can be identified. The mix can have a
normalized SASA falling in between the base proteins or can be larger
than both bases. Therefore, systematic generation of more examples
and in-­depth study of the patterns and relationship of their structural
and mechanical properties would likely be required to gain additional
understanding of protein sequence-­structure-­property relationship
as well as the learned tendency of our model, which can be done in
future work, along with additional experimental validation.
With the obtained de novo pulling force responses as the input, it
is interesting to investigate the novelty of the generated proteins. As
shown in Table 2, all of the designs have no significant similarity among
known protein sequences according to protein BLAST analysis. The
case study shown here demonstrates that our generative model can
also be used as an effective tool to explore the unknown possibilities
in property space of mechanical unfolding responses by constructing
de novo input conditions. In return, it is more probable to discover
de novo designs of proteins.
Besides using the mixing scheme demonstrated above, other ways
can be used to create or discover de novo mechanical unfolding
responses as the design input. One possibility is to train a generative
model (e.g., a diffusion model) on the pulling force data curated here.
At deployment, mechanical unfolding responses can be generated
with or without input conditions. Because the possible pulling force
vectors should in principle be determined by the sequence-­structure-­
property relationship of proteins, in-­depth investigation on this pro-
cess deserves a separate study, where the generation model developed
here can serve as an effective tool in navigating the complex design
space for proteins.
To prepare the protein design pipeline for further experimental
testing, we can also integrate available predicting tools (36, 78, 79) to
estimate properties of generated proteins that are key to experimental
synthesis, manipulation, and test. For example, by applying the solu-
bility predictor based on protein sequences, we can screen generated
proteins for desired solubility. As listed in the last column of Tables 1
and 2, surprisingly, we find many of the generated proteins are predicted
to have a solubility higher (i.e., larger than 0.45) than the average
soluble Escherichia coli protein from the experimental solubility data-
set (90), which indicates encouraging potential for successful experi-
mental synthesis, use, and manipulation.
Benefits of using pLDM in small-­data generation tasks
Combining the above results of protein designs based on existing or
de novo mechanical unfolding responses, we have demonstrated that
the pLDM developed here can achieve reasonably good design
accuracy for both detailed unfolding pulling force history and overall
mechanical properties. At the same time, many of the generated se-
quences are totally de novo, having no significant similarity among
any existing or known proteins. This is surprising considering the

Table 2. Results of the BLAST analysis and predicted solubility for the various generated proteins (from Fig. 6) based on de novo mechanical unfolding
responses. Given the de novo pulling force vectors as the design condition, the model shows high probability in predicting de novo sequences that show little
similarity to existing proteins as can be seen from the BLAST results (for Mix 1 to 4). The predicted solubility is scaled with the experimental dataset with a
population average of 0.45. The listed solubility values are all larger than 0.45, thus predicting to have a higher solubility than the average soluble E. coli proteins
in the experimental dataset (90).
Case Sequence BLAST result: the sequence producing the most signifi- Predicted solubility
cant alignment
Among PDB proteins Beyond PDB proteins

For mix 1 MDDALLLKAMQQLLLA- – No significant similarity 0.742

PIRVKEDDPLVRRDAAIG- found (NSSF)
FAPDGVRVDFEYTAKVD-
LAKATLDEVGLKGANTTQP-
PRPIANKLPPPIVVLASKLLEI-
YKELKQL
For mix 2 GSSGSSGYGRQVTTR- – NSSF 0.814
SPRETTSLSFDIDREPMEFN-
QLKAQELMMPFNLKALDT-
GRFNRPLQFVEQAKGK-
MEKALLKKATDPVQALPK-
KDLSEGISPKMG
For mix 3 MMDTPKLMDELKDYAPQPAR- – NSSF 0.761
RALNLTNPRTAAVPEKTG-
DDVAPFFDHAAEKENLGF-
HHEVANDNWESEAKFLK-
LTKVPVSPQVIYNAAGLL-
FEAARKTP
For mix 4 GSSGSSGPSTYKNPG- – NSSF 0.657
DRFFTTSYFTDPELEAGQFE-
VRTKDKMLNGITLLQQKP-
CGKSCELFVDQNKKAVEEK-
KKKLMLTQMAAYYQQDDLT-
MASGPSSG

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 9 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

fact that (i) the design task is challenging, and (ii) the training data
used are relatively small compared with previous work (56). Specifi-
cally, the design task here is to map detailed pulling force responses
directly to protein sequences, bypassing predicting the 3D struc-
tural transition of protein chains during the unfolding process. On
the basis of MD simulations and experimental studies, it is clear that
the unfolding process of protein structures under mechanical forces
corresponds to complex uncoiling of the backbone and breaking of
hydrogen bonds within the 3D hierarchical structures of proteins.
Theoretically, we can expect that this design task could be more
challenging than designing protein sequences based on structural
conditions (e.g., secondary structure types). At the same time, for
the pLDM discussed above, which designs sequences with a length,
N, smaller than or equal to 126, it was trained with a dataset of 6863
proteins, mainly limited by the high computational cost to curate
the data on mechanical properties. The size of the dataset is small
compared with the PDB proteins of known structures (i.e., about
13,644 for single-­chain protein with N ≤ 126), and even further
dwarfed by the number of all possible protein sequences (i.e., more
than 20126).
Given such a challenging protein design problem with a limited
small set of labeled data, we discuss the strength of the current
pLDM by comparing it with the protein diffusion models (pDMs)
developed previously (56). The structure of the adjusted pDM is
summarized in fig. S1A. The main difference between pLDM (Fig. 2B)
and pDM is about the space where the diffusion/denoising processes
occur during the training/inferring. While the pDM operates in the
tokenized protein sequence space, pLDM performs in the word
probability latent space formed by pretraining the pLM on the available
protein sequence dataset (i.e., UniRef50 dataset for ESM-­2 pLM
adopted here), which is much larger than the datasets on protein
structures or mechanical properties. For various downstream pre-
diction tasks, the adoption of the pretrained pLM as a base model
has proven to be beneficial (33). For example, the ESM-­2 model adopted
here has been used for protein folding prediction in ESMFold (26,
87) and achieved comparable accuracy with the folding tool using
MSA as the input (e.g., AlphaFold2) but at a lower computational
cost. At the same time, for inverse generation, the integration of a
large language model and diffusion model has proven to be beneficial
in conditioned generation tasks for images (48, 91) and texts (92, 93)
[e.g., Diffusion-­LM can achieve fine-­grained controls on syntactic
structure for text generation (93)]. However, the applications of pre-
trained pLM in the generative tasks of protein design remain rare
(94) and the potential benefits are to be explored. Here, the mechanical
unfolding response of the conditioned protein design task and our
results provide a concrete example to study this aspect of the process.
To compare with pLDM, two alternative models (AMs) based on
pDMs were constructed. The first one, AM1, uses one pDM and de-
signs the protein sequence for a given pulling force vector in one
shot, similar to the pLDM developed above. The second AM, AM2,
consists of two pDMs as the protein designer and protein predictor
separately. The designer is the same as the first pDM model, AM1,
while the predictor is trained to predict pulling force vectors based
on the given protein sequences. At deployment, the protein designer
iteratively generates sequence candidates and the protein predictor
then evaluates them to pick the most accurate one among five attempts,
forming an on-­the-­fly iterative design scheme (fig. S1B). Both models
are trained on the same dataset discussed earlier and tested with the
same generation task based on existing mechanical unfolding re-
sponses from the test set. A summary of the performance of the two
pDM-­based models together with those of the pLDM is listed in
Table 3. We use various metrics, R2 and relative errors, to evaluate
the design accuracy of the models on the whole test set, considering
not only the overall mechanical properties (i.e., toughness and
strength) but also the detailed mechanical unfolding responses in
terms of the pulling force vectors. For most of the comparisons (9
out the 11 rows), pLDM achieves the best performance (e.g., the
highest R2 or the lowest error). This result demonstrates that, by

Table 3. Performance comparison between the current pLDM with one-­shot prediction (last column) and the protein diffusion model with one-­shot
prediction (third to the last column) as well as iterative predictions (second to the last column). Tested with the existing unfolding responses from the test
set, the pLDM shows an overall better performance in fulfilling the design target by achieving the best results (indicated with the underline, the minimum for
errors and the maximum for R2) in most rows considering mechanical properties as well as the detailed pulling force responses.
Performance on the test set AM1: pDM with AM2: iterative The current model:
one-­shot gener- prediction with pLDM with one-­
ation pDM-­based pro- shot prediction
tein designer and
predictor

Toughness (i.e., unfolding energy) R2 0.86 0.87 0.93

Mean 0.151 0.150 0.147
Relative L1 error
Median 0.121 0.123 0.102
Strength (i.e., Fmax) R2 0.09 0.17 0.41
Mean 0.188 0.164 0.188
Relative L1 error
Median 0.151 0.113 0.149
Pulling force vectors As vectors Mean 0.427 0.418 0.452
R2
Median 0.526 0.522 0.563
Mean 0.399 0.402 0.398
Relative L2 error
Median 0.377 0.382 0.362
As components R2 0.476 0.499 0.537

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 10 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

integrating the pretrained pLM, pLDM achieves enhancement in
producing more accurate designs even when being trained only on
a relatively small dataset. This advantage of pLDM with respect to
pDM can be helpful for other property-­to-­sequence design tasks for
proteins, given that usually data on protein mechanical or other
properties are more limited or costly to collect.
DISCUSSION
Generating de novo proteins based on their mechanical unfolding
responses presents unique challenges in property-­targeted protein
design. For rational design strategies, it is hard to grasp the complex
relationships between sequences, structures, and properties. For
data-­driven methods, the labeled data on the mechanical properties
are often costly to collect and limited in number, especially given the
enormous possibilities in protein sequence space.
Here, we have developed a pLDM as an effective tool to tackle these
challenges and generate de novo proteins that meet the mechanical
properties’ design objectives in an end-­to-­end manner. The pLDM
developed combines the pretrained pLM and diffusion model as key
components and leverages the strength of both. The pLM part is pre-
trained on the abundant protein sequence data and thus provides an
effective representation of protein sequences in its latent space. The
diffusion model part only operates in this latent space and learns the
map between detailed pulling force responses and the sequence rep-
resentation using only a relatively small set of data curated by per-
forming full-­atom simulations. By examining the unfolding details
of individual designs and the statistics of the mechanical properties
of many cases, we demonstrate that the proteins designed by our
model meet the targeted overall mechanical properties, including
toughness and strength, as well as the detailed unfolding force vectors
with reasonably good accuracy. Moreover, the sequences generated
are mostly de novo, sharing very limited similarity with existing/
known proteins. Given the mechanical unfolding responses from
known PDB proteins as the design input, our model still shows a
strong tendency in discovering de novo proteins as alternatives.
Constructing de novo unfolding responses as the input via a mixing
scheme further boosts the probability of generating de novo designs.
Finally, through controlled comparisons, we show that the pLDM
outperforms the vanilla pDM with or without an iterative design
scheme in achieving better design accuracy, thus clearly demonstrat-
ing the benefits of combining pretrained pLM and diffusion model
in the pLDM developed here. A short summary of these key aspects
about the pLDM is listed in Table 4.
As the initial steps of developing property-­to-­sequence generative
models for de novo protein design, here we adopt the force-­separation
curves collected from MD simulations, in the hope of achieving
consistency and relevance, avoiding bias from simplified models,
and curating sufficient data points for DL model training. A few
clarifications on the mechanical unfolding response data are included
in the following.
First, there exist differences as well as commonalities between
MD results and experimental measurements of the force-­separation
curves of protein unfolding that deserve more nuanced discussion.
The mechanical unfolding process of proteins often involves entropic
elasticity, a transition to energetic elasticity, and bond breaking (95).
Thus, the force response often shows strong rate dependence. Limited
by computational power, the MD simulations are performed at a
pulling speed several orders faster than that in the experimental
tests. Therefore, the corresponding unfolding mechanisms (e.g.,
sequential or simultaneous rupture of several hydrogen bonds) can
be different (96), and a direct comparison of the force records is often
challenging. It should be pointed out that our current model, trained
with all-­atom MD data, is not intended for designing proteins that
directly meet the given pulling force response at a different pulling
speed. Instead, we use the MD data under the fixed pulling speed as
a consistent representation of mechanical properties of protein.
While the absolute values of strength and toughness measured by
MD may change with the pulling speed, the relative rank of the
mechanical properties of the proteins often remain robust. At the
same time, there do exist methods and procedures to bridge MD
and experimental results (97, 98). For example, built upon the
steered MD trajectory calculated here, further MD simulations can
be performed to calculate the mean force potential during unfolding
using statistical sampling methods. Unfolding force distribution in
experimentally relevant regimes can be predicted based on the
mean force potential via transition-­state theory and Monte Carlo
simulations (97). Therefore, the design goal of our current model can
be connected to the response under experimental relevant pulling
speed and force level with extra calculations and sampling efforts.
Second, our MD data include fundamental atomistic details and
avoid bias from bottom-­up coarse-­grained (CG) or theoretical models
with specific assumptions. By tracking atomic motions during
unfolding, the MD results require little predefined assumptions like

Table 4. A short summary of the performance of protein language diffusion model developed in the present work and other models discussed.
Model name Tested input conditions Design accuracy Design novelty

The developed model: protein
language diffusion model using
one-­shot design
AM1: Protein diffusion model using
one-­shot design
AM2: Protein diffusion model using
multi-­shot iterative design
Mechanical unfolding responses
from naturally existing proteins
De novo mechanical unfolding
responses
Mechanical unfolding responses
from existing proteins
Mechanical unfolding responses
from existing proteins
Good agreement with the designed
pulling force responses as well as
the strength and toughness in trend
and values
Similar to the above
Slightly weaker than AM2
Weaker than the developed model
Tend to generate de novo ones, but
can also rediscover ones that show
some similarity to existing proteins
More probable to discover de novo
sequences
–
–

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 11 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
those in CG models (99) or theoretical worm-­like chain (WLC)
models (100). At the same time, the information collected from the
full-­atom MD simulation can be used to fit parameters in CG and
WLC models. The diffusion model was directly trained on these
force patterns and learned to pick relevant features via the attention
mechanisms embedded, avoiding any human intervention or pre-­
knowledge on the subject.
Third, similar models can be developed when sufficient data on
mechanical unfolding force under other testing protocols become
available. As an initial step to prove this framework, our current
model is trained and validated with the freshly curated MD data and
the consistent MD test protocol. At the same time, similar models
can be straightforwardly trained with other sufficient databases. In
particular, when a large number of force-­separation curves measured
from a standard experimental protocol become available, models
with similar architectures can be developed via direct training or
transfer learning with them. With such models, consistent validation
will require synthesis and test of the designed proteins in the wet lab.
This requires well-­planned in-­depth design for specific goals while
the current work is focused on model development and has provided
self-­consistent validation. We will save the experimental studies for
future study.
Finally, the specific choice of pulling force direction in our MD
protocol is clarified here. It has been demonstrated that the me-
chanical unfolding responses of proteins can be affected by the
detailed folding geometry and unfolding pathway (101). To effectively
record the force history that corresponds to the deformation and
uncoiling events of protein internal structures, we designed the test
protocol to apply the mechanical force along the direction that connects
the two ends of a protein chain after relaxation. When the mechanical
pull is applied in another direction, the monomer is likely to first
undergo a rotational motion to align along with the current pull
direction before meaningful unfolding events happen given that in
a quasi-­static loading process rotation usually requires a smaller
load than unfolding events like breaking hydrogen bonds. Once the
protein monomer aligns along with the pulling direction, for the
unfolding process that follows, we expect that a force record pattern
similar to ours could be collected and our protocol and results remain
relevant and robust. At the same time, some extreme cases could exist.
For protein monomers mainly with weak internal folding structures
(e.g., unstructured random coils), the load to rotate the monomer can
be comparable to or larger than that of deformation and unfolding.
We expect the detailed force pattern could be affected more strongly
by the choice of pull direction in those cases.
The freshly curated MD dataset and the pLDM developed here
offer a unique and powerful means to investigate the underlying
sequence-­structure-­property relationship and explore the enormous
protein sequence spaces with molecular mechanical properties as
guidance, to meet specific mechanobiology properties. With the
available dataset, one can conduct a detailed survey on the internal
structural features (e.g., secondary structures) of the proteins and
their correlation with the unfolding force pattern to see whether there
is any uneven distribution centered on certain secondary structure
patterns (e.g., α helix and β sheet) among PDB proteins and our training
set for certain unfolding force patterns. Applying our model, future
studies can start with a systematic study on the relationships between
sequence-­structure-­mechanical properties in proteins.
For example, as demonstrated in Fig. 6, one can construct de novo
mechanical unfolding responses by mixing those existing PDB proteins
Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024
at different ratios and our model can generate protein candidates
that meet those mechanical unfolding responses. With such generating
capability, one can systematically classify the patterns of various
unfolding responses of the existing PDB proteins as shown in Fig. 2C,
construct de novo force-­separation responses transferring between
those different patterns, and generate the corresponding proteins, thus
studying how the patterns of the unfolding force responses and their
transition affect the protein sequence mutations and internal struc-
ture variations. At the same time, as the designed de novo proteins in-
crease in number, their sequences and mechanical unfolding responses
can be used as growing data to gradually increase the protein dataset on
mechanical properties and our model can be further trained on this
growing set. With the more powerful model, one can further study
some challenging topics, such as designing proteins with optimal
mechanical properties or even their combination in various engi-
neering and biological applications.
While we have developed the pLDM that takes the mechanical
unfolding responses as the design conditions here, we expect that
similar pLDM frameworks can be generalized for other property-­
to-­sequence design tasks in proteins. The enhancement brought by
merging pretrained pLMs can be inspiring for other design tasks,
especially where only small datasets on the property of interest are
available or affordable at the beginning. At the same time, going
beyond only one type of condition as the design target, our pLDM
can also be generalized for design tasks under multiple objectives,
given the flexibility of the diffusion model in incorporating these
conditions (Fig. 3B). Combining the previous work using a pDM
(56), one example can be taking both secondary structure and un-
folding forces as the design target. Also, during the generation
process, techniques like inpainting through selective masking or
biasing certain amino acids (102) are straightforward to imple-
ment. Combining these under the pLDM framework, we envision
a comprehensive generative model that moves towards designing
proteins at all levels, including sequence, structure, and properties
in harmony.
MATERIALS AND METHODS
Protein mechanical unfolding simulations by MD
We use Nanoscale Molecular Dynamics (NAMD) to perform full-­atom
MD simulations. The interaction between protein atoms is described by
the CHARMM force field (84). We adopt a generalized Born implicit
solvent model (85) for the effect of solvent on proteins. Compared
with simulations with an explicit solvent model, our setup balances
the accuracy and the computational costs. We develop a parallel
workflow to simulate the mechanical unfolding process of about
7026 proteins of various sequence lengths.
Dataset
We curate the dataset based on the MD results. Key information for
each protein case includes PDB ID, protein sequence, sequence
length, pulling force vector, strength, and toughness. See Fig. 2 for
details on their distributions. We use 85% of the dataset for training
and keep 15% for testing.
Design of the neural network architectures and training
The pLDM developed here consists of a pretrained pLM and a diffusion
model. Only the latter is trainable. For the pretrained pLM, we use a
variant with 150 million parameters from the ESM-­2 series of models
12 of 15
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

(26, 88). During training, we first propagate a mini-­batch of B tokenized
sequences with a length smaller than or equal to N in terms of a B ×
1 × N tensor to the last hidden layer of the pretrained pLM to compute
the logits, then normalize them into a B × M × N tensor, where the
M components in the second dimension represent the probability of
that position being each of the M words in the pLM model. The dif-
fusion model only operates in this word probability space introduced
by the pLM. At deployment, the output of the diffusion model is
translated back to the tokenized sequences by sampling the word
with the highest probability.
Protein folding
We adopt OmegaFold (23) for rapid prediction of protein structures
from the sequences. OmegaFold offers a rapid alternative as it does
not require MSA, yet produces results of similar accuracy as Alpha-
Fold2 (20) and trRosetta (103) (and similar, related state of the
art methods).
Design accuracy evaluation
We use various metrics to compare the measured mechanical un-
folding responses and mechanical properties with the input design
conditions for individual designs as well as predictions for the
whole test set.
For vectors, including the unfolding pulling force vector for one
protein and toughness or strength for proteins in the test set, the R2
and relative L2 error are defined as the following
(xi − yi )2
∑
R2 [�x⃑, y�⃑] = 1 − ∑i
2 Nat. Mater. 14, 23–36 (2015).
i (xi − x)
√∑
2 13. F. Barthelat, Z. Yin, M. J. Buehler, Structure and mechanics of interfaces in biological
‖ x�⃑ − y�⃑ ‖ i (xi − yi )
L2rela [�x⃑, y�⃑] = =
‖ x�⃑ ‖
√∑
2 (5)
i (xi )
where x�⃑ is the ground truth or input vector and y�⃑ is the measured
one from the predictions, xi and yi are their components, and x is the
mean of the components xi.
For scalars, including the toughness and strength for one protein,
the relative L1 error is defined as the following
∣y − x∣
L1rela [x, y] = (6)
∣x∣
where x is the ground truth or input value and y is the measured
value based on the prediction.
BLAST analysis
The BLAST analysis (89) for the various cases is conducted using the
blastp (protein-­protein BLAST) algorithm and the nonredundant
protein sequences (nr) database.
Visualization
We use Visual Molecular Dynamics (104) for visualization of the
protein structures.
Software versions and hardware
We use Python 3.9.16, PyTorch 1.12.1 + cu13 (105) with CUDA (CUDA
version 12.0), and an NVIDIA Tesla V100 with 32 GB VRAM for
training and inference.
Supplementary Materials
This PDF file includes:
Fig. S1
Table S1
Legends for movies S1 to S8
Legends for data S1 to S4
Other Supplementary Material for this manuscript includes the following:
Movies S1 to S8
Data S1 to S4
REFERENCES AND NOTES
1. G. A. Petsko, D. Ringe, Protein Structure and Function (New Science Press, 2004).
2. D. López Barreiro, J. Yeo, A. Tarakanova, F. J. Martin-­Martinez, M. J. Buehler, Multiscale modeling
of silk and silk-­based biomaterials—A review. Macromol. Biosci. 19, e1800253 (2019).
3. G. Gronau, S. T. Krishnaji, M. E. Kinahan, T. Giesa, J. Y. Wong, D. L. Kaplan, M. J. Buehler, A review
of combined experimental and computational procedures for assessing biopolymer
structure–process–property relationships. Biomaterials 33, 8240–8255 (2012).
4. C. Vepari, D. L. Kaplan, Silk as a biomaterial. Prog. Polym. Sci. 32, 991–1007 (2007).
5. S. Ling, D. L. Kaplan, M. J. Buehler, Nanofibrils in nature and materials engineering. Nat.
Rev. Mater. 3, 18016 (2018).
6. K. A. Jansen, D. M. Donato, H. E. Balcioglu, T. Schmidt, E. H. J. Danen, G. H. Koenderink, A
guide to mechanobiology: Where biology and physics meet. Biochim. Biophys. Acta
1853, 3043–3052 (2015).
7. A. E. M. Beedle, S. Garcia-­Manyes, The role of single-­protein elasticity in mechanobiology.
Nat. Rev. Mater. 8, 10–24 (2023).
8. J. L. Balestrini, J. K. Skorinko, A. Hera, G. R. Gaudette, K. L. Billiar, Applying controlled
non-­uniform deformation for in vitro studies of cell mechanobiology. Biomech. Model.
Mechanobiol. 9, 329–344 (2010).
9. Z. Qin, M. J. Buehler, L. Kreplak, A multi-­scale approach to understand the
mechanobiology of intermediate filaments. J. Biomech. 43, 15–22 (2010).
10. P.-­S. Huang, S. E. Boyken, D. Baker, The coming of age of de novo protein design. Nature
537, 320–327 (2016).
(4) 11. U. G. K. Wegst, H. Bai, E. Saiz, A. P. Tomsia, R. O. Ritchie, Bioinspired structural materials.
12. G. X. Gu, M. Takaffoli, M. J. Buehler, Hierarchically enhanced impact resistance of
bioinspired composites. Adv. Mater. 29, 1700060 (2017).
materials. Nat. Rev. Mater. 1, 1–16 (2016).
14. W. Huang, A. Tarakanova, N. Dinjaski, Q. Wang, X. Xia, Y. Chen, J. Y. Wong, M. J. Buehler,
D. L. Kaplan, Design of multistimuli responsive hydrogels using integrated modeling and
genetically engineered silk–elastin-­like proteins. Adv. Funct. Mater. 26, 4113–4123 (2016).
15. S. T. Krishnaji, G. Bratzel, M. E. Kinahan, J. A. Kluge, C. Staii, J. Y. Wong, M. J. Buehler,
D. L. Kaplan, Sequence–structure–property relationships of recombinant spider silk
proteins: Integration of biopolymer design, processing, and modeling. Adv. Funct. Mater.
23, 241–253 (2013).
16. M. J. Buehler, Generative pretrained autoregressive transformer graph neural network
applied to the analysis and discovery of novel proteins. J. Appl. Phys. 134, 084902 (2023).
17. A. Paladino, F. Marchetti, S. Rinaldi, G. Colombo, Protein design: From computer models
to artificial intelligence. WIREs Comput. Mol. Sci. 7, e1318 (2017).
18. Z. Qin, L. Wu, H. Sun, S. Huo, T. Ma, E. Lim, P. Y. Chen, B. Marelli, M. J. Buehler, Artificial
intelligence method to design and fold alpha-­helical structural proteins from the
primary amino acid sequence. Extreme Mech. Lett. 36, 100652 (2020).
19. J. Wang, H. Cao, J. Z. H. Zhang, Y. Qi, Computational protein design with deep learning
neural networks. Sci. Rep. 8, 6349 (2018).
20. J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool,
R. Bates, A. Žídek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard,
A. Cowie, B. Romera-­Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman,
E. Clancy, M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein,
D. Silver, O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, D. Hassabis, Highly accurate
protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
21. M. Baek, F. DiMaio, I. Anishchenko, J. Dauparas, S. Ovchinnikov, G. R. Lee, J. Wang,
Q. Cong, L. N. Kinch, R. Dustin Schaeffer, C. Millán, H. Park, C. Adams, C. R. Glassman,
A. DeGiovanni, J. H. Pereira, A. V. Rodrigues, A. A. Van Dijk, A. C. Ebrecht, D. J. Opperman,
T. Sagmeister, C. Buhlheller, T. Pavkov-­Keller, M. K. Rathinaswamy, U. Dalwadi, C. K. Yip,
J. E. Burke, K. Christopher Garcia, N. V. Grishin, P. D. Adams, R. J. Read, D. Baker, Accurate
prediction of protein structures and interactions using a three-­track neural network.
Science 373, 871–876 (2021).
22. M. Varadi, S. Anyango, M. Deshpande, S. Nair, C. Natassia, G. Yordanova, D. Yuan, O. Stroe,
G. Wood, A. Laydon, A. Zídek, T. Green, K. Tunyasuvunakool, S. Petersen, J. Jumper,
E. Clancy, R. Green, A. Vora, M. Lutfi, M. Figurnov, A. Cowie, N. Hobbs, P. Kohli,

Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024 13 of 15

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
G. Kleywegt, E. Birney, D. Hassabis, S. Velankar, AlphaFold Protein Structure Database:
Massively expanding the structural coverage of protein-­sequence space with
high-­accuracy models. Nucleic Acids Res. 50, D439–D444 (2022).
23. R. Wu, F. Ding, R. Wang, R. Shen, X. Zhang, S. Luo, C. Su, Z. Wu, Q. Xie, B. Berger, J. Ma,
J. Peng, High-­resolution de novo structure prediction from primary sequence. bioRxiv
500999 [Preprint] (2022). https://doi.org/10.1101/2022.07.21.500999.
24. R. Chowdhury, N. Bouatta, S. Biswas, C. Rochereau, G. M. Church, P. K. Sorger,
M. Alquraishi, Single-­sequence protein structure prediction using language models from
deep learning. bioRxiv 454840 [Preprint] (2021). https://doi.
org/10.1101/2021.08.02.454840.
25. X. Fang, F. Wang, L. Liu, J. He, D. Lin, Y. Xiang, K. Zhu, X. Zhang, H. Wu, H. Li, L. Song, A
method for multiple-­sequence-­alignment-­free protein structure prediction using a
protein language model. Nat. Mach. Intell. 5, 1087–1096 (2023).
26. Z. Lin, H. Akin, R. Rao, B. Hie, Z. Zhu, W. Lu, N. Smetanin, R. Verkuil, O. Kabeli, Y. Shmueli,
A. dos Santos Costa, M. Fazel-­Zarandi, T. Sercu, S. Candido, A. Rives, Evolutionary-­scale
prediction of atomic-­level protein structure with a language model. Science 379,
1123–1130 (2023).
27. X. Fang, F. Wang, L. Liu, J. He, D. Lin, Y. Xiang, X. Zhang, H. Wu, H. Li, L. Song, HelixFold-­
Single: MSA-­free protein structure prediction by using protein language model as an
alternative. arXiv:2207.13921 [q-­bio.BM] (2022).
28. M. H. Høie, E. N. Kiehl, B. Petersen, M. Nielsen, O. Winther, H. Nielsen, J. Hallgren,
P. Marcatili, NetSurfP-­3.0: Accurate and fast prediction of protein structural features by
protein language models and deep learning. Nucleic Acids Res. 50, W510–W515 (2022).
29. B. Zhang, J. Li, Q. Lü, Prediction of 8-­state protein secondary structures by a novel deep
learning architecture. BMC Bioinformatics 19, 293 (2018).
30. C. H. Yu, W. Chen, Y. H. Chiang, K. Guo, Z. Martin Moldes, D. L. Kaplan, M. J. Buehler,
End-­to-­end deep learning model to predict and design secondary structure content of
structural proteins. ACS Biomater Sci. Eng. 8, 1156–1165 (2022).
31. G. Pollastri, A. McLysaght, Porter: A new, accurate server for protein secondary structure
prediction. Bioinformatics 21, 1719–1720 (2005).
32. C. Mirabello, G. Pollastri, Porter, PaleAle 4.0: High-­accuracy prediction of protein secondary
structure and relative solvent accessibility. Bioinformatics 29, 2056–2058 (2013).
33. A. Elnaggar, M. Heinzinger, C. Dallago, G. Rehawi, Y. Wang, L. Jones, T. Gibbs, T. Feher,
C. Angerer, M. Steinegger, D. Bhowmik, B. Rost, ProtTrans: Toward understanding the
language of life through self-­supervised learning. IEEE Trans. Pattern Anal. Mach. Intell. 44,
7112–7127 (2022).
34. J. Tubiana, D. Schneidman-­Duhovny, H. J. Wolfson, ScanNet: An interpretable geometric
deep learning model for structure-­based protein binding site prediction. Nat. Methods
19, 730–739 (2022).
35. F. Sverrisson, J. Feydy, B. E. Correia, M. M. Bronstein, Fast end-­to-­end learning on protein
surfaces. bioRxiv 424589 [Preprint] (2020). https://doi.org/10.1101/2020.12.28.424589.
36. V. Thumuluri, H. M. Martiny, J. J. Almagro Armenteros, J. Salomon, H. Nielsen,
A. R. Johansen, NetSolP: Predicting protein solubility in Escherichia coli using language
models. Bioinformatics 38, 941–946 (2022).
37. M. J. Buehler, MeLM, a generative pretrained language modeling framework that solves
forward and inverse mechanics problems. J. Mech. Phys. Solids 181, 105454 (2023).
38. E. Khare, C. Gonzalez-­Obeso, D. L. Kaplan, M. J. Buehler, CollagenTransformer: End-­to-­end
transformer model to predict thermal stability of collagen triple helices using an nlp
approach. ACS Biomater Sci. Eng. 8, 4301–4310 (2022).
39. Y. Hu, M. J. Buehler, End-­to-­end protein normal mode frequency predictions using
language and graph models and application to sonification. ACS Nano 16, 20656–20670
(2022).
40. K. Guo, M. J. Buehler, Rapid prediction of protein natural frequencies using graph neural
networks. Digit. Discov. 1, 277–285 (2022).
41. F. Y. C. Liu, B. Ni, M. J. Buehler, PRESTO: Rapid protein mechanical strength prediction with
an end-­to-­end deep learning model. Extreme Mech. Lett. 55, 101803 (2022).
42. A. J. Lew, M. J. Buehler, A deep learning augmented genetic algorithm approach to
polycrystalline 2D material fracture discovery and design. Appl. Phys. Rev. 8, 041414
(2021).
43. E. Khare, C. H. Yu, C. Gonzalez Obeso, M. Milazzo, D. L. Kaplan, M. J. Buehler, Discovering
design principles of collagen molecular stability using a genetic algorithm, deep
learning, and experimental validation. Proc. Natl. Acad. Sci. U.S.A. 119, e2209524119
(2022).
44. G. Dong, G. Liao, H. Liu, G. Kuang, A review of the autoencoder and its variants: A
comparative perspective from target recognition in synthetic-­aperture radar images.
IEEE Geosci. Remote Sens. Mag. 6, 44–68 (2018).
45. I. Goodfellow, J. Pouget-­Abadie, M. Mirza, B. Xu, D. Warde-­Farley, S. Ozair, A. Courville,
Y. Bengio, Generative adversarial networks. Commun ACM 63, 139–144 (2020).
46. J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models. Adv Neural Inf Process
Syst 33, 6840–6851 (2020).
47. G. Marcus, E. Davis, S. Aaronson, A very preliminary analysis of DALL-­E 2.
arXiv:2204.13807 [cs.CV] (2022).
Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024
48. C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, S. K. S. Ghasemipour, B. K. Ayan,
S. S. Mahdavi, R. G. Lopes, T. Salimans, J. Ho, D. J. Fleet, M. Norouzi, Photorealistic
text-­to-­image diffusion models with deep language understanding. arXiv:2205.11487
[cs.CV] (2022).
49. R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-­resolution image synthesis
with latent diffusion models. Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition 2022-­June, 10674–10685 (2021).
50. M. Z. Makoś, N. Verma, E. C. Larson, M. Freindorf, E. Kraka, Generative adversarial
networks for transition state geometry prediction. J. Chem. Phys. 155, 024116 (2021).
51. T. Lebese, B. Mellado, X. Ruan, The use of Generative Adversarial Networks to characterise
new physics in multi-­lepton final states at the LHC. arXiv:2105.14933 [hep-­ph] (2021).
52. Z. Yang, C. H. Yu, K. Guo, M. J. Buehler, End-­to-­end deep learning method to predict
complete strain and stress tensors for complex hierarchical composite microstructures. J.
Mech. Phys. Solids 154, 104506 (2021).
53. Z. Yang, C. H. Yu, M. J. Buehler, Deep learning model to predict complex stress and strain
fields in hierarchical composites. Sci. Adv. 7, eabd7416 (2021).
54. M. J. Buehler, FieldPerceiver: Domain agnostic transformer model to predict multiscale
physical fields and nonlinear material properties through neural ologs. Mater. Today 57,
9–25 (2022).
55. B. Ni, H. Gao, A deep learning approach to the inverse problem of modulus identification
in elasticity. MRS Bull. 46, 19–25 (2021).
56. B. Ni, D. L. Kaplan, M. J. Buehler, Generative design of de novo proteins based on
secondary-­structure constraints using an attention-­based diffusion model. Chem 9,
1828–1849 (2023).
57. Z. Lin, T. Sercu, Y. LeCun, A. Rives, Deep generative models create new and diverse
protein structures, in Machine Learning in Structural Biology Workshop at the 35th
Conference on Neural Information Processing Systems (MLSB, 2021).
58. N. Anand, T. Achim, Protein structure and sequence generation with equivariant
denoising diffusion probabilistic models. arXiv:2205.15019 [q-­bio.QM] (2022).
59. B. L. Trippe, J. Yim, D. Tischer, D. Baker, T. Broderick, R. Barzilay, T. Jaakkola, Diffusion
probabilistic modeling of protein backbones in 3D for the motif-­scaffolding problem.
arXiv:2206.04119 [q-­bio.BM] (2022).
60. J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern,
A. J. Borst, R. J. Ragotte, L. F. Milles, B. I. M. Wicky, N. Hanikel, S. J. Pellock, A. Courbet,
W. Sheffler, J. Wang, P. Venkatesh, I. Sappington, S. V. Torres, A. Lauko, V. De Bortoli,
E. Mathieu, S. Ovchinnikov, R. Barzilay, T. S. Jaakkola, F. DiMaio, M. Baek, D. Baker, De novo
design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023).
61. A. Madani, B. Krause, E. R. Greene, S. Subramanian, B. P. Mohr, J. M. Holton, J. L. Olmos,
C. Xiong, Z. Z. Sun, R. Socher, J. S. Fraser, N. Naik, Large language models generate functional
protein sequences across diverse families. Nat. Biotechnol. 41, 1099–1106 (2023).
62. M. Mora, A. Stannard, S. Garcia-­Manyes, The nanomechanics of individual proteins. Chem.
Soc. Rev. 49, 6816–6832 (2020).
63. J. Alegre-­Cebollada, Protein nanomechanics in biological context. Biophys. Rev. 13,
435–454 (2021).
64. M. J. Buehler, S. Keten, T. Ackbarow, Theoretical and computational hierarchical
nanomechanics of protein materials: Deformation and fracture. Prog. Mater. Sci. 53,
1101–1241 (2008).
65. H. Lu, B. Isralewitz, A. Krammer, V. Vogel, K. Schulten, Unfolding of titin immunoglobulin
domains by steered molecular dynamics simulation. Biophys. J. 75, 662–671 (1998).
66. K. C. Neuman, A. Nagy, Single-­molecule force spectroscopy: Optical tweezers, magnetic
tweezers and atomic force microscopy. Nat. Methods 5, 491–505 (2008).
67. E. M. Puchner, H. E. Gaub, Force and function: Probing proteins with AFM-­based force
spectroscopy. Curr. Opin. Struct. Biol. 19, 605–614 (2009).
68. M. L. Hughes, L. Dougan, The physics of pulling polyproteins: A review of single molecule
force spectroscopy using the AFM to study protein unfolding. Rep. Prog. Phys. 79, 076601
(2016).
69. X. Zhang, L. Ma, Y. Zhang, High-­resolution optical tweezers for single-­molecule
manipulation. Yale J. Biol. Med. 86, 367–383 (2013).
70. D. B. Ritchie, M. T. Woodside, Probing the structural dynamics of proteins and nucleic
acids with optical tweezers. Curr. Opin. Struct. Biol. 34, 43–51 (2015).
71. R. Tapia-­Rojo, E. C. Eckels, J. M. Fernández, Ephemeral states in protein folding under
force captured with a magnetic tweezers design. Proc. Natl. Acad. Sci. U.S.A. 116,
7873–7878 (2019).
72. X. Zhao, X. Zeng, C. Lu, J. Yan, Studying the mechanical responses of proteins using
magnetic tweezers. Nanotechnology 28, 414002 (2017).
73. J. Wu, P. Li, C. Dong, H. Jiang, B. Xue, X. Gao, M. Qin, W. Wang, B. Chen, Y. Cao, Rationally
designed synthetic protein hydrogels with predictable mechanical properties. Nat.
Commun. 9, 620 (2018).
74. M. Sikora, J. I. Sulkowska, B. S. Witkowski, M. Cieplak, BSDB: The biomolecule stretching
database. Nucleic Acids Res. 39, D443–D450 (2011).
75. H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov,
P. E. Bourne, The Protein Data Bank. Nucleic Acids Res. 28, 235–242 (2000).
14 of 15
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
76. UniProt Consortium, UniProt: The Universal Protein Knowledgebase in 2023. Nucleic
Acids Res. 51, D523–D531 (2023).
77. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser,
I. Polosukhin, Attention is all you need. Adv Neural Inf Process Syst 2017-­December,
5999–6009 (2017).
78. J. Ribeiro, C. Ríos-­Vera, F. Melo, A. Schüller, Calculation of accurate interatomic contact
surface areas for the quantitative analysis of non-­bonded molecular interactions.
Bioinformatics 35, 3499–3501 (2019).
79. M. Hebditch, M. A. Carballo-­Amador, S. Charonis, R. Curtis, J. Warwicker, Protein–Sol: A
web tool for predicting protein solubility from sequence. Bioinformatics 33, 3098–3100
(2017).
80. P. B. Dennis, E. L. Onderko, J. M. Slocik, L. J. Bird, D. A. Phillips, W. J. Crookes-­Goodson,
S. M. Glaven, Proteins for bioinspired optical and electronic materials. MRS Bull. 45,
1027–1033 (2020).
81. T. Wang, D. He, H. Yao, X. Guo, B. Sun, G. Wang, Development of proteins for high-­
performance energy storage devices: Opportunities, challenges, and strategies. Adv.
Energy Mater. 12, 2202568 (2022).
82. G. Qin, P. B. Dennis, Y. Zhang, X. Hu, J. E. Bressner, Z. Sun, W. J. Crookes-­Goodson,
R. R. Naik, F. G. Omenetto, D. L. Kaplan, Recombinant reflectin-­based optical materials.
J. Polym. Sci. B Polym. Phys. 51, 254–264 (2013).
83. J. Ren, Y. Wang, Y. Yao, Y. Wang, X. Fei, P. Qi, S. Lin, D. L. Kaplan, M. J. Buehler, S. Ling,
Biological material interfaces as inspiration for mechanical and optical material designs.
Chem. Rev. 119, 12279–12336 (2019).
84. K. Vanommeslaeghe, E. Hatcher, C. Acharya, S. Kundu, S. Zhong, J. Shim, E. Darian,
O. Guvench, P. Lopes, I. Vorobyov, A. D. Mackerell Jr., CHARMM general force field
(CGenFF): A force field for drug-­like molecules compatible with the CHARMM all-­atom
additive biological force fields. J. Comput. Chem. 31, 671–690 (2010).
85. A. V. Onufriev, D. A. Case, Generalized born implicit solvent models for biomolecules.
Annu. Rev. Biophys. 48, 275–296 (2019).
86. S. R. K. Ainavarapu, J. Brujić, H. H. Huang, A. P. Wiita, H. Lu, L. Li, K. A. Walther,
M. Carrion-­Vazquez, H. Li, J. M. Fernandez, Contour length and refolding rate of a small
protein controlled by engineered disulfide bonds. Biophys. J. 92, 225–233 (2007).
87. Z. Lin, H. Akin, R. Rao, B. Hie, Z. Zhu, W. Lu, A. Dos, S. Costa, M. Fazel-­Zarandi, T. Sercu,
S. Candido, A. Rives, M. Ai, Language models of protein sequences at the scale of
evolution enable accurate structure prediction. bioRxiv 500902 [Preprint] (2022). https://
doi.org/10.1101/2022.07.20.500902.
88. facebookresearch/esm: Evolutionary Scale Modeling (esm): Pretrained language models
for proteins. https://github.com/facebookresearch/esm.
89. S. F. Altschul, W. Gish, W. Miller, E. W. Myers, D. J. Lipman, Basic local alignment search
tool. J. Mol. Biol. 215, 403–410 (1990).
90. T. Niwa, B. W. Ying, K. Saito, W. Jin, S. Takada, T. Ueda, H. Taguchi, Bimodal protein
solubility distribution revealed by an aggregation analysis of the entire ensemble of
Escherichia coli proteins. Proc. Natl. Acad. Sci. U.S.A. 106, 4201–4206 (2009).
91. A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, M. Chen, Hierarchical text-­conditional image
generation with CLIP latents. arXiv:2204.06125 [cs.CV] (2022).
92. H. Zou, Z. M. Kim, D. Kang, A survey of diffusion models in natural language processing.
arXiv:2305.14671 [cs.CL] (2023).
93. X. Lisa Li, J. Thickstun, I. Gulrajani, P. Liang, T. B. Hashimoto, Diffusion-­LM improves
controllable text generation. [Preprint] (2022). https://github.com/XiangLi1999/
Diffusion-­LM.git.
94. Z. Gao, C. Tan, S. Z. Li, DiffSDS: A language diffusion model for protein backbone inpainting
under geometric conditions and constraints. arXiv:2301.09642 [q-­bio.QM] (2023).
Ni et al., Sci. Adv. 10, eadl4000 (2024) 7 February 2024
95. M. J. Buehler, S. Y. Wong, Entropic elasticity controls nanomechanics of single
tropocollagen molecules. Biophys. J. 93, 37–43 (2007).
96. T. Ackbarow, X. Chen, S. Keten, M. J. Buehler, Hierarchies, multiple energy barriers, and
robustness govern the fracture mechanics of α-­helical and β-­sheet protein domains.
Proc. Natl. Acad. Sci. U.S.A. 104, 16410–16415 (2007).
97. P.-­C. Li, D. E. Makarov, Theoretical studies of the mechanical unfolding of the muscle
protein titin: Bridging the time-­scale gap between simulation and experiment.
J. Chem. Phys. 119, 9260–9268 (2003).
98. P. Cao, G. Yoon, W. Tao, K. Eom, H. S. Park, The role of binding site on the mechanical
unfolding mechanism of ubiquitin. Sci. Rep. 5, 8757 (2015).
99. S. Kmiecik, D. Gront, M. Kolinski, L. Wieteska, A. E. Dawid, A. Kolinski, Coarse-­grained
protein models and their applications. Chem. Rev. 116, 7898–7936 (2016).
100. S. Keten, M. J. Buehler, Strength limit of entropic elasticity in beta-­sheet protein domains.
Phys. Rev. E Stat. Nonlin. Soft Matter Phys. 78, 061913 (2008).
101. R. A. Nome, J. M. Zhao, W. D. Hoff, N. F. Scherer, Axis-­dependent anisotropy in protein
unfolding from integrated nonequilibrium single-­molecule experiments, analysis, and
simulation. Proc. Natl. Acad. Sci. U.S.A. 104, 20799–20804 (2007).
102. J. Dauparas, I. Anishchenko, N. Bennett, H. Bai, R. J. Ragotte, L. F. Milles, B. I. M. Wicky,
A. Courbet, R. J. de Haas, N. Bethel, P. J. Y. Leung, T. F. Huddy, S. Pellock, D. Tischer, F. Chan,
B. Koepnick, H. Nguyen, A. Kang, B. Sankaran, A. K. Bera, N. P. King, D. Baker, Robust deep
learning–based protein sequence design using ProteinMPNN. Science 378, 49–56 (2022).
103. Z. Du, H. Su, W. Wang, L. Ye, H. Wei, Z. Peng, I. Anishchenko, D. Baker, J. Yang, The trRosetta
server for fast and accurate protein structure prediction. Nat. Protoc. 16, 5634–5651
(2021).
104. W. Humphrey, A. Dalke, K. Schulten, VMD: Visual molecular dynamics. J. Mol. Graph. 14,
33–38 (1996).
105. A. Paszke, S. Gross, F. Massa, A. Lerer, J. B. Google, G. Chanan, T. Killeen, Z. Lin,
N. Gimelshein, L. Antiga, A. Desmaison, A. K. Xamla, E. Yang, Z. Devito, M. R. Nabla,
A. Tejani, S. Chilamkurthy, Q. Ai, B. Steiner, L. F. Facebook, J. B. Facebook, S. Chintala,
PyTorch: An imperative style, high-­performance deep learning library. Adv Neural Inf
Process Syst 32, (2019).
Acknowledgments
Funding: We acknowledge support from USDA (2021-­69012-­35978), DOE-­SERDP (WP22-­
S1-­3475), ARO (79058LSCSB, W911NF-­22-­2-­0213, and W911NF2120130), and the MIT-­IBM
Watson AI Lab and MIT’s Generative AI Initiative. Additional support from NIH (U01EB014976
and R01AR077793) and ONR (N00014-­19-­1-­2375 and N00014-­20-­1-­2189) is acknowledged.
Author contributions: Conceptualization: M.J.B. and D.L.K. Investigation: B.N. and M.J.B.
Methodology: B.N., M.J.B., and D.L.K. Resources: M.J.B. and D.L.K. Funding acquisition: M.J.B.
and D.L.K. Data curation: B.N. and M.J.B. Validation: M.J.B. and B.N. Supervision: M.J.B. and D.L.K.
Formal analysis: B.N. and M.J.B. Software: B.N. and M.J.B. Project administration: M.J.B. and
D.L.K. Visualization: B.N., M.J.B., and D.L.K. Writing—original draft: B.N., M.J.B., and D.L.K.
Writing—review and editing: B.N., M.J.B., and D.L.K. Competing interests: The authors declare
that they have no competing interests. Data and materials availability: All data needed to
evaluate the conclusions in the paper are present in the paper and/or the Supplementary
Materials.
Submitted 16 October 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adl4000
15 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

H E A LT H A N D M E D I C I N E Copyright © 2024 The

Authors, some rights
Inhaled immunoantimicrobials for the treatment of reserved; exclusive
licensee American
chronic obstructive pulmonary disease Association for the
Advancement of
Science. No claim to
Junliang Zhu1†, Xiaohui Li2†, Yang Zhou1, Chenglong Ge1, Xudong Li1, Mengying Hou1, original U.S.
Yuansong Wei1, Yongbing Chen2, Kam W. Leong3, Lichen Yin1,2* Government Works.
Distributed under a
Effective therapeutic modalities and drug administration strategies for the treatment of chronic obstructive pulmo- Creative Commons
nary disease (COPD) exacerbations are lacking. Here, mucus and biofilm dual-­penetrating immunoantimicrobials Attribution
(IMAMs) are developed for bridging antibacterial therapy and pro-­resolving immunotherapy of COPD. IMAMs are NonCommercial
constructed from ceftazidime (CAZ)–encapsulated hollow mesoporous silica nanoparticles (HMSNs) gated with a License 4.0 (CC BY-­NC).
charge/conformation-­transformable polypeptide. The polypeptide adopts a negatively charged, random-­coiled
conformation, masking the pores of HMSNs to prevent antibiotic leakage and allowing the nebulized IMAMs to ef-
ficiently penetrate the bronchial mucus and biofilm. Inside the acidic biofilm, the polypeptide transforms into a
cationic and rigid α helix, enhancing biofilm retention and unmasking the pores to release CAZ. Meanwhile, the
polypeptide is conditionally activated to disrupt bacterial membranes and scavenge bacterial DNA, functioning as
an adjuvant of CAZ to eradicate lung-­colonizing bacteria and inhibiting Toll-­like receptor 9 activation to foster in-
flammation resolution. This immunoantibacterial strategy may shift the current paradigm of COPD management.

INTRODUCTION
Chronic obstructive pulmonary disease (COPD) has become a ma-
jor global health problem with increasing prevalence (1, 2). Accord-
ing to World Health Organization statistics, COPD is responsible for
more than 3 million deaths annually (3). While various therapeutic
strategies have been developed to alleviate symptoms and prevent
complications, limited success has been achieved in slowing or halt-
ing disease progression (4). As a major cause of COPD, smoking
damages lung tissues and suppresses the immune response, leading
to aberrant and persistent inflammation (5). Meanwhile, smoking
triggers the massive proliferation of goblet cells, causing dense mu-
cus accumulation in the airway and thereby leading to restricted
breathing, airway obstruction, and alveolar structural changes (6).
Acute exacerbations, which are often triggered by bacterial infection,
lead to further inflammation and airway damage. Even worse, due to
the thickening of the mucus layer, bacteria tend to aggregate at the
bottom to form biofilms, which in turn promote bacterial prolifera-
tion (7). In addition, insufficient oxygen supply caused by restricted
breathing promotes anaerobic respiration of bacteria, leading to the
decrease of pH inside biofilms, which is beneficial for the reproduc-
tion of anaerobic bacteria. Therefore, to effectively treat COPD exac-
erbations, it is necessary to simultaneously clear the colonized
bacteria and alleviate pulmonary inflammation (8).
Antibiotics with broad-­spectrum antibacterial activity are the
most widely used pharmaceuticals for anti-­infection treatment in
patients with COPD (9). However, their efficacy is often suboptimal
because of insufficient entry into bacteria and the occurrence of
drug resistance (10–14). For instance, Pseudomonas aeruginosa,
a causative pathogen of exacerbations in patients with severe
COPD, has a low-­permeability outer membrane, which limits the
1
Institute of Functional Nano and Soft Materials (FUNSOM), Jiangsu Key Laboratory
for Carbon-­Based Functional Materials and Devices, Soochow University, Suzhou
215123, China. 2Department of Thoracic Surgery, The Second Affiliated Hospital of
Soochow University, Suzhou 215002, China. 3Department of Biomedical Engineer-
ing, Columbia University, New York, NY, USA.
*Corresponding author. Email: lcyin@​suda.​edu.​cn
†These authors contributed equally to this work.
penetration of antibiotics into the cells (15, 16). Antimicrobial pep-
tides (AMPs) and AMP mimics, which typically consist of cationic
and hydrophobic amino acid residues, serve as important alterna-
tives to antibiotics and can kill bacteria by directly destabilizing/
disrupting their membranes (17). In addition, AMPs and their
mimics can often serve as antibiotic adjuvants by enhancing the per-
meability of the bacterial cell membrane (18–21). Therefore, the
combined use of AMPs and antibiotics can greatly boost the bacte-
ricidal efficacy of antibiotics, lower the required clinical dosages of
antibiotics, and reduce the emergence of drug resistance (22, 23).
For anti-­infection management of COPD, nebulization is the
most effective and convenient administration route because it can
increase the local concentration of antimicrobials in the lung while
reducing systemic exposure to diminish potential side effects (24,
25). However, effective nebulization therapy is greatly challenged by
the multiple pathophysiological barriers (26–30). On the one hand,
COPD is characterized by hypersecretion of mucus in the upper
airway and bronchi, which is enriched in negatively charged mucin
glycoproteins (31–33). Thus, after nebulization, positively charged
antimicrobials tend to become trapped by the viscous mucus layer
lining above the bronchial epithelium owing to strong electrostatic
attraction with mucin (34), and they are subsequently cleared due to
rapid mucus turnover (35–37). On the other hand, bacteria coloniz-
ing the lungs form biofilms that are held together by an extracellular
polymeric substance (EPS) matrix consisting of bacteria-­secreted
extracellular DNA (eDNA), polysaccharides, and glycoproteins (38–
39). The EPS matrix can capture positively charged antimicrobials to
hamper their penetration into the biofilm and prevent their direct
contact with bacteria (40–43). Therefore, effective mucus and biofilm
penetration are critically imperative for inhaled antimicrobials.
In addition to bacterial infection, COPD is often accompanied
by persistent pulmonary inflammation, which is mainly induced by
bacteria-­secreted eDNA, a pathogen-­associated molecular pattern
(44) that causes excessive activation of Toll-­like receptor 9 (TLR9) in
multiple types of immune cells, resulting in an inflammatory micro-
environment characterized by excessive proinflammatory cytokine
production, reactive oxygen species (ROS) release, and immune cell

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 1 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
infiltration (45–47). Notably, the burst release of genomic DNA
(gDNA) from bacteria killed by antimicrobials may result in a cyto-
kine storm, thereby severely impairing lung function and leading to
poor prognosis (48). Therefore, antibacterial immunotherapy via
bacterial eDNA/gDNA scavenging in combination with antimicro-
bial therapy is highly desired for the management of COPD but re-
mains largely unexplored (49–53).
Here, mucus and biofilm dual-­penetrating immunoantimicrobi-
als (IMAMs) with conditional bacterial killing and eDNA/gDNA-­
scavenging capabilities are developed for the nebulized treatment
of COPD. The IMAMs are composed of ceftazidime (CAZ)–
encapsulated hollow mesoporous silica nanoparticles (HMSNs) gated
with an outer shell of charge/conformation-­transformable polypep-
tides. The polypeptides, which bear both positively charged quaternary
amine and negatively charged carboxyl in their side chains, adopt a
negative charge and random-­coiled conformation at neutral pH as a
consequence of electrostatic attractions among their side chains.
Thus, the flexible, random-­coiled polypeptides stack on the HMSN
surface to mask the pores and prevent CAZ leakage. Upon inhala-
tion in mice with COPD, the negatively charged IMAMs efficiently
penetrate the mucus and biofilm due to their electrostatic repulsion
with glycoproteins and polysaccharides. Inside the biofilm which is
mildly acidic, the polypeptides are conditionally activated upon
acid-­triggered removal of their carboxyl groups, transforming from
a negatively charged, flexible random coil with low membrane activ-
ity to a positively charged, rigid α helix with a potent capacity to
disrupt bacterial membranes. In the meantime, the helical polypep-
tides protrude away from the HMSN surface due to their rigid, rod-­
like nature, unmasking the pores to enable on-­demand CAZ release.
Consequently, the helical antimicrobial polypeptides synergize with
CAZ to eradicate the colonized bacteria. Moreover, the positively
charged polypeptides can effectively scavenge bacterial eDNA and
gDNA to inhibit the infiltration and activation of immune cells, pro-
moting inflammation resolution in the context of COPD (Fig. 1).
RESULTS
Preparation and characterization of IMAMs
HMSNs with surface-­immobilized amine groups were used to initi-
ate the ring-­ opening copolymerization of γ-­(6-­chlorohexyl)-­​L-­
glutamic acid–derived N-­ carboxyanhydride (CH NCA) and
Nε-­tert-­butyloxycarbonyl-­​L-­lysine N-­carboxyanhydride (Boc-­lys
NCA) at the molar ratio of 4:1, followed by the quaternization reac-
tion of chloro groups with N,N-­dimethyl-­1-­octanamine and depro-
tection of the Boc groups, yielding cationic polypeptide-­decorated
nanoparticles (referred to as MEK NPs). MEK NPs were then conju-
gated with cis-­ aconitic anhydride (CA) on the terminal amine
groups of the polypeptide side chains, ultimately generating acid-­
responsive MEKCA NPs (figs. S1 to S6). On the basis of thermogravi-
metric analysis, the average degree of polymerization of individual
polypeptide segments was determined to be 86 (fig. S7). Transmis-
sion electron microscopy (TEM) image revealed the hollow and
spherical morphology of MEKCA NPs (Fig. 2A). The polypeptide
segments in MEK NPs contained positively charged quaternary
amines and primary amines on the side chain terminals, and they
adopted a typical helical secondary structure, as evidenced by the
circular dichroism (CD) spectrum (Fig. 2B). In comparison, after
CA modification, the polypeptide segments in MEKCA NPs con-
tained both positively charged quaternary amine and negatively
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
charged carboxyl in their side chains, and they adopted a random-­
coiled conformation in aqueous solution, mainly due to strong elec-
trostatic attractions among their side chains (54, 55). Accordingly,
MEK NPs had a positive zeta potential of +45 mV, while MEKCA
NPs had a negative zeta potential of −30 mV (Fig. 2C). The α helical
polypeptide segments in MEK NPs tended to protrude away from
the HMSN surface due to their rigid, rod-­like nature, while the flex-
ible, random-­coiled polypeptide segments in MEKCA NPs allowed
flat stacking on the HMSN surface. As a result, the hydrodynamic
diameter of MEKCA NPs (123 nm) was smaller than that of MEK
NPs (145 nm; Fig. 2D and table S1). CAZ was then encapsulated
within the cavity of HMSNs, and the flexible, random-­coiled poly-
peptides could gate the pores and prevent premature leakage of
CAZ, generating MEKCA/CAZ NPs with a drug loading content
(DLC) of 16.4% and drug loading efficiency (DLE) of 35.6%.
Acid-­responsive transformation of IMAMs and drug release
The acid-­triggered structural transformation and drug release
profile of MEKCA NPs were then explored. MEKSA NPs, nonrespon-
sive analogs of MEKCA NPs, were similarly prepared using succinic
anhydride (SA) instead of CA (fig. S6). As shown by the CD spectra,
the secondary structure of the polypeptide segments in MEKCA
NPs but not in MEKSA NPs transformed from a random coil to a
helix after incubation in a mildly acidic environment (pH 6.5) main-
ly because the acid-­triggered removal of carboxyl groups and restora-
tion of lysine residues eliminated the electrostatic attractions among
the polypeptide side chains (Fig. 2E). Thus, the acid-­treated
MEKCA NPs transformed back to MEK NPs, conferring an increase
in particle size from 123 to 143 nm (Fig. 2F) and alteration of the zeta
potential from negative (−30 mV) to positive (+44 mV; Fig. 2G). As
a consequence of the acid-­triggered transition of the polypeptide
segments from a flexible random coil to a rigid helix, the pores of the
HMSNs were unmasked to facilitate drug release (Fig. 2H), with a
cumulative release rate of ~45% within the first 4 hours and ~83% in
the following 20 hours at pH 6.5 (Fig. 2I and fig. S8). In comparison,
under neutral conditions (pH 7.4), only 29% of the CAZ was re-
leased from MEKCA/CAZ NPs within 24 hours. Consistently, acid
treatment showed a negligible effect on CAZ release from the non-
responsive MEKSA/CAZ NPs, with a limited drug release rate of
25% after 24-­hour incubation (Fig. 2I).
Mucus penetration of MEKCA NPs in vitro and in vivo
Engineering negatively charged surfaces represents an effective
strategy to construct “mucus-­inert” nanocarriers and reduce their
entrapment in the mucus layer (56). Here, the penetration efficiency
of Cy5-­labeled MEKCA NPs (Cy5MEKCA NPs) in purulent sputum
from patients with cystic fibrosis (CF) was first evaluated. As shown
in Fig. 3A, the transport efficiency (TE) of the negatively charged
Cy5
MEKCA NPs was eightfold higher than that of the positively
charged Cy5MEK NPs. Subsequently, the multiple particle tracking
(MPT) assay was used to investigate the motility of Cy5MEKCA NPs
in CF sputum. As shown in Fig. 3 (B and C), the range of Brownian
motion of Cy5MEKCA NPs was substantially more extensive than
that of Cy5MEK NPs, with a mean square displacement (<MSD>)
value within 10 s that was higher than that of Cy5MEK NPs by ~3
orders of magnitude. Accordingly, the effective diffusivity (Deff) of
Cy5
MEKCA NPs was also significantly higher than that of Cy5MEK
NPs (Fig. 3D). For further validation, the diffusive behavior of
Cy5
MEKCA NPs in mucus was assessed by the photobleaching assay,
2 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Schematic illustration of IMAMs penetrating mucus and biofilm for the immunoantibacterial treatment of COPD. (A) A scheme illustrating the
construction of IMAMs (MEKCA/CAZ NPs). (B) A scheme illustrating IMAMs penetrating mucus and biofilm, followed by the charge/conformation transformation
in the acidic environment. MEKCA NPs synergize with CAZ to eliminate bacterial infection, scavenge eDNA and gDNA, avoid excessive activation of immune cells,
and ultimately achieve antimicrobial and anti-­inflammatory effects. TFA, trifluoroacetic acid; IL-­6, interleukin-­6; TNF-­α, tumor necrosis factor–α.

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 3 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Characterization and acid-­triggered transformation of IMAMs. (A) TEM image of MEKCA NPs (scale bar, 100 nm). CD spectra [(B), 1 mg/ml], zeta potential
(C), and particle size (D) of MEK NPs and MEKCA NPs (n = 3). (E) CD spectra of MEKCA NPs and MEKSA NPs (1 mg/ml) in deionized water at pH 6.5 and 7.4. Particle size (F) and
zeta potential (G) of MEKCA NPs and MEKSA NPs at pH 6.5 and 7.4 (n = 3). (H) Illustration of acid-­triggered transformation of MEKCA NPs. At pH 6.5, the random-­coiled,
negatively charged, flexible polypeptides on the HMSN surfaces transform into helical, positively charged, rigid polypeptides, thus unmasking the pores of HMSNs to re-
lease the encapsulated CAZ. (I) Cumulative release of CAZ from MEKCA/CAZ NPs and MEKSA/CAZ NPs in phosphate-­buffered saline (PBS) (pH 6.5 or 7.4, n = 3).

wherein the decrease in the fluorescence intensity in the photo-
bleached region negatively correlated with the diffusion efficiency of
NPs (57). Upon photobleaching of the yellow circled region for
150 s, Cy5MEKCA NPs showed only a slight fluorescence decrease
in comparison to the drastic fluorescence decrease (>60%) observed
for Cy5MEK NPs (Fig. 3, E and F). The above findings thus indicated
that the negatively charged MEKCA NPs had markedly higher trans-
lational motility in mucus than the positively charged MEK NPs
mainly because of the electrostatic repulsion between MEKCA NPs
and negatively charged mucin glycoproteins. To verify this hypothesis,
mucin adsorption on MEKCA NPs was assessed by measuring the
formation of NP/mucin aggregates (Fig. 3G). As expected, after in-
cubation with 0.3% mucin, Cy5MEK NP/mucin aggregates showed
25-­fold higher fluorescence intensity than Cy5MEKCA NP/mucin ag-
gregates, and similar results were also observed in the presence of
0.5% mucin.
The mucus penetration of MEKCA NPs after nebulization was
further explored via observation of their distribution in the lung tis-
sue by confocal laser scanning microscopy (CLSM). As illustrated in
Fig. 3H, the positively charged Cy5MEK NPs were mainly localized
on top of the goblet cells, indicating severe entrapment of Cy5MEK
NPs in the mucus layer lining above the epithelium. In drastic con-
trast, Cy5MEKCA NPs could effectively penetrate the mucus barrier
and were largely distributed into the deeper regions of the lung epi-
thelium. All the above findings collectively demonstrated the de-
sired capability of MEKCA NPs to penetrate mucin reticular fiber
meshes both in vitro and in vivo, which was mainly attributed to
their resistance to mucin adsorption. Moreover, MEKCA/CAZ NPs
had the desired stability during mucus penetration, as evidenced by
the unaltered size of NPs and low CAZ release level after 48-­hour
incubation in the mucus (fig. S9).
Pulmonary retention of MEKCA/CAZ NPs
The lung-­deposited fraction of the nebulized Cy5MEKCA/CAZ
NPs was first estimated by determining the fluorescence intensity
in the bronchoalveolar lavage fluid (BALF) at different time points
after nebulization. As shown in fig. S10, the lung-­deposited frac-
tion was calculated to be 33% at 1 hour after nebulization. The
pulmonary retention of Cy5MEKCA NPs was then explored via ex
vivo fluorescence imaging of excised lung tissues at different time

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 4 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Mucus penetration and lung retention of IMAMs. (A) Transport efficiency of Cy5MEK NPs and Cy5MEKCA NPs across CF sputum after 8-­hour incubation (n = 3).
(B) Representative trajectories of Cy5MEK NPs and Cy5MEKCA NPs in CF sputum during 20-­s movies. (C) Distribution of the logarithmic effective diffusivities [Log10(Deff )] of
individual NPs in CF sputum (n > 100 NPs per experiment). (D) The mean square displacement (<MSD>) values of individual NPs in CF sputum as a function of time (n = 3).
(E) Representative fluorescence images of Cy5MEK NPs and Cy5MEKCA NPs in CF sputum upon photobleaching in the circled area for different times (scale bar, 40 μm).
(F) Relative fluorescence intensity in the irradiated region (within the yellow circle) in (E) (n = 3). (G) Fluorescence intensity of Cy5MEK NPs and Cy5MEKCA NPs that formed
aggregates with mucin glycoproteins at different mucin concentrations (0.3 or 0.5%, w/v, n = 3). (H) Confocal laser scanning microscopy (CLSM) images of lung tissue
sections harvested from COPD mice 1 hour after nebulization of Cy5MEK NPs or Cy5MEKCA NPs (scale bar, 50 μm). Cell nuclei were stained with 4′,6-­diamidino-­2-­phenylindole
(DAPI). Ex vivo fluorescence imaging of excised lungs from COPD mice (I) and Cy5MEK level in excised lungs (J) at 0.5, 2, 4, 6, and 12 hours after nebulization of Cy5MEK NPs
or Cy5MEKCA NPs (n = 3). Ex vivo fluorescence imaging of excised lungs from COPD mice (K) and calcein level in excised lungs (L) at 0.5, 2, 4, 6, and 12 hours after nebuliza-
tion of MEKCA/calcein NPs or free calcein (n = 3). For (D) and (G), statistical analysis was calculated via Student's t test. ***P < 0.001.

points post-­nebulization. As illustrated in Fig. 3 (I and J), the
positively charged Cy5MEK NPs were rapidly cleared from the
lung, as evidenced by the drastic fluorescence attenuation at 2 hours
after administration mainly because of their entrapment by the
mucus layer followed by rapid mucociliary clearance. In com-
parison, Cy5MEKCA NPs, with robust mucus penetration capabil-
ity, could be retained in the lung for over 12 hours. To further
investigate the pulmonary retention of the antibiotic, calcein, a
fluorescent molecule with similar water solubility to CAZ, was
encapsulated within MEKCA NPs. Free calcein was quickly cleared
from the lung within 4 hours after nebulization, while appreciable
pulmonary accumulation could still be noted for MEKCA/calcein
NPs at 6 hours after nebulization (Fig. 3, K and L). These results
therefore demonstrated that MEKCA NPs could achieve pro-
longed pulmonary retention with the encapsulated antibiotic af-
ter nebulization, which should benefit their antimicrobial efficacy
against COPD.
Biofilm penetration and antibacterial efficiency of MEKCA/
CAZ NPs in vitro
To evaluate the biofilm permeation and retention capabilities of
MEKCA NPs, a P. aeruginosa biofilm was prepared according to the

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 5 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
previously reported method (58) and incubated with Cy5-­labeled
NPs before observation by CLSM. As shown in Fig. 4 (A and B) and
fig. S11, the negatively charged Cy5MEKSA NPs showed negligible
accumulation throughout the biofilm because they lacked sufficient
binding affinity with the EPS matrix to be retained in the biofilm
and were thus easily cleared. The red fluorescence of Cy5MEK NPs
markedly decreased with their penetration into deeper regions of
the biofilm, conferring a penetration depth of 1.6 μm. This occurred
mainly because the positively charged NPs had strong electrostatic
attractions with the EPS matrix and limited penetration into the
biofilms. In comparison, Cy5MEKCA NPs showed pronounced distri-
bution across the entire biofilm with a penetration depth >2.7 μm,
demonstrating their potent penetration capability which was en-
abled by the acid-­triggered charge reversal from negative to positive
inside the biofilm.
Then, the antimicrobial efficiency of the NPs was investigated.
MEKCA NPs efficiently killed Escherichia coli, P. aeruginosa, and
Staphylococcus aureus at pH 6.5, with minimum inhibitory concen-
trations (MICs) of 25 to 50 μg/ml (Fig. 4C). However, they showed
markedly higher MICs of 500 to 1000 μg/ml at pH 7.4 (table S2),
which indicated the conditional antimicrobial efficiency of MEKCA
NPs upon acid-­triggered transformation of the polypeptide seg-
ments from negatively charged random coils to positively charged α
helices. In support of this observation, the nontransformable MEKSA
NPs also exhibited weak antimicrobial capabilities at pH 6.5 (MIC
>1000 μg/ml). As a racemic analog of MEKCA NPs, DL-­MEKCA NPs
were similarly prepared from γ-­(6-­chlorohexyl)-­​DL-­glutamic acid–
derived N-­carboxyanhydride (DL-­ CH NCA) and Nε-­tert-­
butyloxycarbonyl-­​L-­lysine N-­carboxyanhydride (DL-­Boc-­lys NCA),
and the polypeptide segments in DL-­MEKCA NPs had the same
chemical structure as those in MEKCA NPs but could not transform
into α helices upon acid treatment. DL-­MEKCA NPs showed fourfold
higher MICs than MEKCA NPs against three tested bacterial strains
(Fig. 4C) mainly because the stiff, α helical structure featured potent
membrane-­disruptive activity (59, 60). Consistent with this result,
field emission scanning electron microscopy (FE-­SEM) observa-
tions revealed that the bacterial membrane was ruptured and mis-
shapen after treatment with MEKCA NPs at pH 6.5 (Fig. 4D),
indicating that the cationic, helical polypeptides immobilized on the
NPs could efficiently kill bacteria by damaging their cell membranes.
After encapsulation of the antibiotic CAZ, MEKCA/CAZ NPs ex-
hibited an even stronger antibacterial effect at pH 6.5, with notably
decreased MICs in comparison to those of free CAZ and MEKCA
NPs (Fig. 4C). The fractional inhibitory concentration index (FICI)
of CAZ and MEKCA NPs was lower than 0.5 in the three tested bac-
terial strains, which indicated a strong synergistic effect between the
antimicrobial polypeptide and antibiotic. The bactericidal kinetics
of MEKCA NPs, CAZ, and MEKCA/CAZ NPs against the three tested
bacterial strains again demonstrated the higher sterilization effi-
ciency of MEKCA/CAZ NPs than MEKCA NPs and CAZ (fig. S12).
Similar results were also obtained with the live/dead staining images
(Fig. 4E), wherein MEKCA/CAZ NPs outperformed free CAZ and
MEKCA NPs in inducing pronounced bacterial death (red fluores-
cence). Because of their robust biofilm penetration capability,
MEKCA NPs also markedly inhibited the formation and dispersal
of P. aeruginosa biofilms, outperforming both free CAZ and
MEKCA NPs (Fig. 4F and fig. S13). In addition, MEKCA/CAZ NPs
caused negligible cytotoxicity in human fetal lung fibroblast (HLF1)
cells and mouse macrophages (RAW 264.7) at pH 6.5 and 7.4 at
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
concentrations up to 12.5 μg/ml (fig. S14), indicating their desired
cytocompatibility.
MEKCA NP–mediated cfDNA scavenging and
anti-­inflammatory effects in vitro
In the context of COPD, cell-­free DNA (cfDNA) released from dy-
ing somatic cells and bacteria activates TLR9 in the endosomes of
immune cells such as macrophages and neutrophils, thus triggering
severe inflammation (45, 61). To confirm the relationship between
cfDNA and inflammation in COPD, sputum from 15 patients with
acute exacerbation COPD (AECOPD) and 15 healthy volunteers
were harvested. Patients with active AECOPD showed significantly
higher sputum levels of cfDNA, tumor necrosis factor–α (TNF-­α),
and nitric oxide synthase (iNOS) than healthy volunteers (figs. S15
and S16). In addition, the cfDNA level was strongly associated with
AECOPD severity, as evidenced by the strong correlation between
the cfDNA level and the TNF-­α or iNOS level (Fig. 5, A and B), sug-
gesting that cfDNA is an important target for COPD management.
Therefore, we first determined the cfDNA-­scavenging efficiency
of MEKCA NPs in human sputum from patients with AECOPD. As
illustrated in Fig. 5C, after incubation with MEKCA NPs at pH 6.5,
the cfDNA level in the sputum was decreased by 55%, mainly as a
result of the strong electrostatic interaction between the cationic
polypeptide and cfDNA. We further investigated the composition of
cfDNA. The ratio of bacterial DNA to total cfDNA remained con-
stant before and after MEKCA NP treatment, as evidenced by similar
expression of 16S ribosomal RNA (a bacterium-­specific DNA se-
quence) in the cfDNA bound by MEKCA NPs and the cfDNA in the
untreated sputum (fig. S17). However, there was a difference in
the composition of bacterial DNA before and after MEKCA NP
treatment. As illustrated in fig. S18, in the bacterial DNA bound by
MEKCA NPs, the proportion of DNA from gram-­negative bacteria
(Proteobacteria, Bacteroidetes, Fusobacteria, and Epsilonbacteraeota)
was increased, while that from gram-­positive bacteria (Firmicutes
and Actinobacteria) was decreased. This may have occurred because
the thick and dense cell walls of gram-­positive bacteria serve as a
barrier against bacterial DNA release, thus hampering effective
DNA binding by MEKCA NPs.
Bacterial DNA has abundant hypomethylated CpG fragments
and hence shows strong potency in activating TLR9 and inducing
inflammation (62). Hence, we further focused on evaluating the
bacterial DNA-­scavenging capabilities of MEKCA NPs. After bacte-
ria are killed by antimicrobials, bacterial gDNA is extensively re-
leased, serving as the main cause of the cytokine storm (63). Thus,
the ability of MEKCA NPs to scavenge gDNA, interrupt the TLR9
pathway, and inhibit proinflammatory cytokine secretion in im-
mune cells was first evaluated. As shown by agarose gel electropho-
resis, acid (pH 6.5)–treated MEKCA NPs, which had completely
transformed into positively charged MEK NPs, could completely
scavenge gDNA at a weight ratio ≥5:1 (Fig. 5D). Accordingly, un-
treated MEKCA NPs and MEKSA NPs lacking the acid-­responsive
charge reversal property showed negligible gDNA-­scavenging ef-
fects (figs. S19A and S20A). We further explored whether other bac-
terial damage-­ associated molecular patterns (DAMPs) such as
proteins and cell membrane fragments would influence the DNA-­
scavenging efficiency of MEKCA NPs. As illustrated in fig. S21,
MEKCA NPs could still completely scavenge bacterial DNA in
sonication-­generated, unpurified bacterial DAMPs at the NP/DNA
weight ratios ≥5:1, indicating that MEKCA NPs could efficiently
6 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. In vitro biofilm penetration and antibacterial activity of IMAMs. (A) CLSM images of P. aeruginosa biofilm after incubation with Cy5MEKCA NPs, Cy5MEKSA NPs, or
Cy5
MEK NPs for 2 hours (scale bar, 100 μm). P. aeruginosa biofilm was stained with SYTO 9. (B) Relative fluorescence intensity of biofilms at different depths after treatment
with various NPs. The top of the biofilm was set as 0 μm. (C) Minimum inhibitory concentration (MIC) of CAZ and various NPs against different bacterial strains and the
calculated fractional inhibitory concentration index (FICI) between CAZ and MEKCA NPs. (D) FE-­SEM images of P. aeruginosa, S. aureus, and E. coli after 1-­hour treatment
with MEKCA NPs at the MIC. (E) CLSM images of S. aureus, E. coli, and P. aeruginosa after 2-­hour treatment with CAZ, MEKCA NPs, or MEKCA/CAZ NPs at the MIC (scale bar, 15 μm).
Live and dead bacteria were stained with calcein acetoxymethyl ester and propidium iodide, respectively. (F) The biomass of S. aureus, E. coli, and P. aeruginosa biofilms
after 12-­hour treatment with PBS, CAZ, MEKCA NPs, or MEKCA/CAZ NPs at the MIC (n = 3). OD570, optical density at 570 nm. Statistical analysis was calculated via one-­way
ANOVA test. **P < 0.01 and ***P < 0.001.

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 7 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. In vitro cfDNA scavenging and anti-­inflammatory efficacy of IMAMs. Linear regression curves illustrating the correlation between the cfDNA level and the
TNF-­α (A) or iNOS (B) level in the sputum of patients with AECOPD. The green shading around the fitted regression line indicates the 95% confidence interval. (C) cfDNA
level in the sputum from patients with AECOPD after treatment with PBS or MEKCA NPs at pH 6.5 (n = 15). (D) gDNA binding by MEKCA NPs at various NP/DNA weight ratios,
as evaluated by agarose gel electrophoresis. (E) Relative TLR9 activation of HEK-­Blue mTLR9 cells after treatment with gDNA and MEKCA NPs (n = 3). The value obtained for
HEK-­Blue mTLR9 cells treated with gDNA and PBS served as 100%. (F) Flow cytometric diagrams of mouse neutrophils and calculated percentages of activated neutrophils
(Ly6GhighCD11bhigh) after incubation with gDNA and MEKCA NPs (n = 3). Secreted TNF-­α and IL-­6 levels of neutrophils (G) and macrophages [RAW 264.7, (H)] treated with
gDNA and MEKCA NPs (n = 3). (I) eDNA binding by MEKCA NPs at various NP/DNA weight ratios, as evaluated by agarose gel electrophoresis. (J) Relative TLR9 activation in
HEK-­Blue mTLR9 cells after treatment with eDNA and MEKCA NPs (n = 3). The value obtained for HEK-­Blue mTLR9 cells treated with eDNA and PBS served as 100%. (K) Se-
creted TNF-­α and IL-­6 levels of neutrophils treated with eDNA and MEKCA NPs (n = 3). In the above assessments, cells that were not subjected to gDNA/eDNA challenge or
NP treatment served as the control. For (C), statistical analysis was calculated via Student's t test. ***P < 0.001. For (E) to (H) and (J) to (K), statistical analysis was calculated
via one-­way ANOVA test. **P < 0.01 and ***P < 0.001.

scavenge DNA in the presence of other DAMPs. Then, HEK-­Blue
mTLR9 cells (mouse TLR9 reporter cells) were used to evaluate
TLR9 activation after coincubation with bacterial gDNA and MEKCA
NPs. In accordance with the gDNA-­scavenging results, MEKCA NPs
at pH 6.5 markedly decreased gDNA-­induced TLR9 activation by
51%, while MEKCA NPs at pH 7.4 and MEKSA NPs showed minimal
inhibitory effects against TLR9 activation (Fig. 5E and fig. S22A).
Inhibition of abnormal TLR9 signaling contributes to the inactiva-
tion of immune cells such as neutrophils, a major immune cell pop-
ulation participating in inflammation. Here, after coincubation with
MEKCA NPs at pH 6.5, the number of gDNA-­activated neutrophils
(Ly6GhighCD11bhigh) was decreased from 91 to 70% (Fig. 5F), and,
consistently, the levels of TNF-­α and interleukin-­6 (IL-­6) secreted by
neutrophils were decreased by 60 and 33%, respectively (Fig. 5G).
Similarly, in macrophages (RAW 264.7 cells), another important
immune cell population for bacterial elimination, acid-­ treated
MEKCA NPs also decreased the levels of secreted TNF-­α and IL-­6
upon gDNA challenge by 59 and 63%, respectively (Fig. 5H), indi-
cating that MEKCA NPs were able to inhibit the gDNA-­induced ac-
tivation of multiple types of immune cells.
In the context of COPD, live bacteria also constantly secrete
eDNA, which stimulates immune cells to cause persistent inflam-
mation (63). Therefore, we then explored the potency of MEKCA
NPs in scavenging eDNA and mitigating immune activation. Con-
sistent with the above findings related to gDNA-­scavenging, acid
(pH 6.5)–treated MEKCA NPs but not untreated MEKCA NPs or
nonresponsive MEKSA NPs effectively scavenged eDNA at a weight
ratio ≥5:1 (Fig. 5I and figs. S19B and S20B) and accordingly inhib-
ited the eDNA-­induced activation of TLR9 (Fig. 5J and fig. S22B).
As a consequence, eDNA-­induced secretion of TNF-­α and IL-­6
from neutrophils was decreased by 55 and 33%, respectively, after
treatment with MEKCA NPs at pH 6.5 (Fig. 5K). The above results

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 8 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
collectively substantiated that, in addition to the bactericidal activity,
MEKCA NPs were also capable of scavenging bacterial DNA to inhibit
inflammation.
Besides bacterial DNA, the cfDNA released from dying cells in
the body is one of the leading causes of chronic inflammation (52,
64). Therefore, we tested the ability of MEKCA NPs to inhibit the over-
activation of immune cells using macrophage-­generated cfDNA.
As illustrated in fig. S23, at the NP/cfDNA weight ratio of 5:1,
MEKCA NPs efficiently inhibited the secretion levels of TNF-­α and
IL-­6 in neutrophils by 54% and 61%, respectively, indicating that
MEKCA NPs could inhibit chronic inflammation in COPD through
scavenging of cfDNA.
To explore whether the captured DNA would be released again
from the NPs to elicit an immune-­stimulatory effect, eDNA release
from the MEKCA/eDNA complex was evaluated after incubation
with mucus enriched with negatively charged mucin glycoproteins.
As shown in fig. S24A, a minimal amount of eDNA was released
from MEKCA NPs after incubation in mucus for up to 24 hours, in-
dicating that the binding between eDNA and MEKCA NPs was
strong. Moreover, the mixture of the MEKCA/eDNA complex and
mucus collected at different time intervals did not cause appreciable
up-­regulation of TNF-­α and IL-­6 in neutrophils (fig. S24, B and C),
indicating that the captured eDNA would not be released again
from the MEKCA/eDNA complex to elicit an immune-­stimulatory
effect. Meanwhile, CAZ release from MEKCA/CAZ NPs that had
captured eDNA was only slightly slower than that in the absence of
eDNA but much faster than that of MEKSA/CAZ NPs (fig. S25), in-
dicating that the binding between eDNA and NPs would not notice-
ably hinder the release of CAZ.
MEKCA/CAZ NP–mediated antibacterial effects in vivo
A mouse COPD model was established by subjecting mice to smok-
ing treatment, intratracheal lipopolysaccharide (LPS) challenge,
and P. aeruginosa infection (29, 65), and the antibacterial activity of
nebulized NPs was explored in COPD mice (Fig. 6A). Preliminary
experiments were conducted to optimize the dose of MEKCA/CAZ
NPs, and the dose of 2.4 mg CAZ/kg (12.6 mg MEKCA NPs/kg) was
adopted because further increase in the NPs dose did not bring ad-
ditional benefit for the therapeutic effect (fig. S26). The properties
of NPs, including acid responsiveness, mucus permeability, DNA-­
scavenging capacity, antibacterial property, and drug release capac-
ity, remained unchanged after nebulization (fig. S27). The BALF
collected from phosphate-­ buffered saline (PBS)–treated COPD
mice formed a large number of colonies on Luria-­Bertani (LB) agar
plates, and CAZ treatment led to a marginal decrease in the num-
ber of bacterial colonies (Fig. 6B). MEKCA NPs showed some in-
hibitory effect against P. aeruginosa growth, but, in comparison, the
number of bacterial colonies dramatically decreased after MEKCA/
CAZ NP treatment. Quantitative analysis also confirmed the syner-
gistic antimicrobial effect of MEKCA NPs and CAZ, as evidenced by
a significant decrease in colony-­forming units (CFUs) (to ~2%) af-
ter treatment with MEKCA/CAZ NPs, with counts close to those of
uninfected mice.
MEKCA/CAZ NP–mediated antioxidative stress and
anti-­inflammatory effects in vivo
Consistent with the in vitro study, nebulized MEKCA/CAZ NPs
notably decreased the DNA concentration in BALF collected
from COPD mice, which benefited from MEKCA NP–mediated
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
DNA scavenging (Fig. 6C). Consequently, MEKCA/CAZ NPs sig-
nificantly decreased the concentrations of proinflammatory cy-
tokines, including TNF-­α, IL-­1β, IL-­6, and IL-­8, in the serum
and lung tissue of COPD mice, outperforming MEKCA NPs and
free CAZ (Fig. 6D and fig. S28). Among these cytokines, IL-­6, a
prototypic proinflammatory cytokine, can initiate the phosphor-
ylation of signal transducer and activator of transcription 3
(STAT3), an important second messenger of IL-­6 trans-­signaling,
to increase acute-­phase protein production and further amplify
the inflammatory cascade (66). Here, after nebulization of MEKCA/
CAZ NPs, the overexpression of STAT3 and p65 and the phos-
phorylation of inhibitor of nuclear factor κBα (IκBα) in COPD
mice were obviously inhibited (Fig. 6E), indicating effective dis-
ruption of JAK-­STAT signaling. Accordingly, the total protein
level in the BALF of COPD mice was reduced by 58%, mainly
owing to the reduction in proinflammatory cytokines and acute-­
phase proteins (fig. S29).
MEKCA/CAZ NP–mediated bacterial killing and alleviation of
inflammation also ameliorated the pathophysiological microenvi-
ronment in the lungs of COPD mice. In particular, oxidative stress
was relieved by MEKCA/CAZ NPs. The H2O2 level in the lung was
reduced by 64% (Fig. 6F); pulmonary levels of malondialdehyde
(MDA) and myeloperoxidase (MPO), two critical mediators in-
volved in the generation of ROS, almost recovered to normal values
(Fig. 6, G and H); and pulmonary levels of superoxide dismutase
(SOD) and nuclear factor erythroid 2–related factor 2 (Nrf2), fac-
tors responsible for scavenging free radicals and maintaining the
redox balance, respectively, were also obviously increased (Fig. 6, I
and J). Furthermore, the overproduction of matrix metallopeptidase
9 (MMP-­9), a tissue-­destroying enzyme, was reversed to nearly the
normal level after nebulization of MEKCA/CAZ NPs (Fig. 6K). These
findings therefore collectively confirmed the pronounced anti-­
inflammatory and antioxidant effects of nebulized IMAMs in the
management of COPD.
MEKCA/CAZ NP–mediated inflammation resolution in vivo
During bacterial infection, over-­recruitment, and overactivation
of immune cells lead to tissue edema and increased vascular per-
meability, resulting in the failure of inflammation resolution (67,
68). Here, total cell counts in BALF collected from COPD mice
showed a 66% reduction after MEKCA/CAZ NP treatment, indicat-
ing that MEKCA/CAZ NPs effectively inhibited immune cell infil-
tration (fig. S30). More specifically, after treatment with MEKCA/
CAZ NPs, the infiltration of neutrophils (Ly6GhighCD11bhigh) was
reduced by 92%, which was close to the level observed in healthy
mice (Fig. 7A). Similarly, the number of recruited macrophages
(CD45+CD11bhighCD11chigh) in lung tissues was decreased from
28 to 5%, comparable to the normal level in healthy mice (3%;
Fig. 7B). After MEKCA/CAZ NP treatment, the abundance of mac-
rophages with the M1 phenotype (F4/80+CD45+CD86+, a typical
proinflammatory phenotype) was significantly decreased (from
33 to 1%; fig. S31), and the ratio of M2-­ type macrophages
(CD45+F4/80+CD206+, a pro-­resolving phenotype) to M1-­type macro-
phages was increased from 0.2 to 19 (Fig. 7C), indicating that
MEKCA/CAZ NPs effectively polarized macrophages from the
proinflammatory M1 phenotype to the anti-­ inflammatory M2
phenotype. In addition, the numbers of other immune cells in lung
tissues, such as helper T cells (CD4+CD3+) and cytotoxic T cells
(CD8+CD3+), were restored to nearly normal levels (Fig. 7, D and E).
9 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. Antibacterial and anti-­inflammatory efficacy of IMAMs in COPD mice. (A) Timeline of the in vivo study. Mice were subjected to smoking treatment (four
cigarettes per day) from days 1 to 21, intratracheal administration of LPS (1 mg/ml in saline, 20 μl) on day 7, and intratracheal administration of P. aeruginosa (1 × 105
CFU/ml, 10 μl) on day 21 to induce COPD. COPD mice were nebulized with PBS, CAZ, MEKCA NPs, or MEKCA/CAZ NPs (2.4 mg CAZ/kg, 12.6 mg MEKCA/kg) on days 22, 23,
and 24. Healthy mice that were not subjected to NP treatment served as the normal control. Animals were euthanized on day 25 before the following assessments.
(B) Images and counts of P. aeruginosa colonies on LB agar plates inoculated with the collected BALF (n = 6). (C) Bacterial DNA concentration in the BALF (n = 4).
(D) Proinflammatory cytokine levels in lung tissues (n = 6). (E) Signal transducer and activator of transcription 3 (STAT3), p65, and p-­IκBα levels in lung tissues, as de-
termined by Western blotting. H2O2 (F), malondialdehyde (MDA) (G), myeloperoxidase (MPO) (H), superoxide dismutase (SOD) (I), nuclear factor erythroid 2–related
factor 2 (Nrf2) (J), and metallopeptidase 9 (MMP-­9) (K) levels in lung tissues (n = 6). For (B) to (D) and (F) to (K), statistical analysis was calculated via one-­way ANOVA
test. *P < 0.05, **P < 0.01, and ***P < 0.001.

All the above results verified that nebulized MEKCA/CAZ NPs
could inhibit immune cell infiltration and prevent immune cells
from differentiating/polarizing to proinflammatory phenotypes,
thus promoting effective inflammation resolution for the manage-
ment of COPD.
MEKCA/CAZ NP–mediated recovery of pulmonary function
Last, the pulmonary function of COPD mice was assessed by
measuring PaO2 and PaCO2 in the arterial blood as well as the
blood pH. As shown in Fig. 7 (F to H), the decreased PaO2,
increased PaCO2, and acidified pH observed in COPD mice
were restored to nearly normal levels after treatment with MEKCA/
CAZ NPs. MEKCA/CAZ NPs also substantially outperformed
CAZ and MEKCA NPs, again verifying their synergistic effect in
reducing pulmonary damage and promoting lung function recovery.
In support of this observation, hematoxylin and eosin (H&E)
staining of lung tissues further revealed that after MEKCA/CAZ
NP treatment, pathological symptoms, including alterations in
the alveolar structure, alveolar wall thinning, and interstitial
edema, were notably ameliorated (Fig. 7I).

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 10 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. IMAM-­mediated inflammation resolution and pulmonary function recovery in COPD mice. Flow cytometric diagrams and calculated percentages of neutro-
phils [Ly6GhighCD11bhigh, (A)] and recruited macrophages [CD45+CD11bhighCD11chigh, (B)] in the single-­cell suspensions derived from the lung tissues of COPD mice
(n = 6). (C) Flow cytometric diagrams of M2-­type macrophages (CD45+F4/80+CD206+) and M1-­type macrophages (CD45+F4/80+CD86+) and the calculated CD206+/
CD86+ ratio (n = 6). Flow cytometric diagrams and calculated percentages of CD4+ T cells [CD4+CD3+, (D)] and CD8+ T cells [CD8+CD3+, (E)] (n = 6). PaO2 (F), PaCO2
(G), and pH (H) of arterial blood (n = 6). (I) Hematoxylin and eosin–stained lung tissue sections (scale bar, 100 μm). For (A) to (H), statistical analysis was calculated via
one-­way ANOVA test. **P < 0.01 and ***P < 0.001. Mice received the same drug administration in Fig. 6A.

Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024 11 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Biocompatibility of MEKCA/CAZ NPs in vivo
The systemic toxicity of nebulized MEKCA/CAZ NPs (2.4 mg CAZ/
kg, 12.6 mg MEKCA NPs/kg) was examined in healthy mice after
three administrations with one day between each administration.
Compared to those of PBS-­treated mice, representative hematologi-
cal and biochemical parameters showed no abnormalities in MEKCA/
CAZ NP–treated mice (fig. S32). In addition, in H&E-­stained cross
sections of major organs (kidney, lung, liver, spleen, and heart), no
necrosis, inflammation, edema, or other pathological symptoms
were detected (fig. S33). These results therefore supported the de-
sired biocompatibility of MEKCA/CAZ NPs after nebulization.
DISCUSSION
In conclusion, we developed a robust IMAM based on polypeptide-­
gated HMSNs to achieve antibacterial and pro-­resolving immuno-
therapy for COPD upon administration via nebulization. The
polypeptide adopted an anionic, flexible, random coil conformation
at neutral pH and underwent an acid-­triggered transformation into a
cationic, rigid α helix inside the biofilm. This property has the follow-
ing major advantages for administration via nebulization and COPD
management. (i) It endows the IMAMs with robust mucus and bio-
film penetration, which are the two most critical processes for the
delivery of antimicrobials via nebulization. (ii) It allows on-­demand
CAZ release inside the biofilm, enhancing bioavailability and speci-
ficity. (iii) The acid-­triggered structural transformation led to condi-
tional activation of the polypeptide, which achieved a “one arrow,
three hawks” capability by serving as a membrane-­disruptive anti-
bacterial agent, antibiotic adjuvant, and bacterial DNA scavenger.
With these unique properties, the IMAMs demonstrated potent
therapeutic efficacy in mice with COPD exacerbations, effectively
eradicating lung-­colonizing bacteria and inhibiting TLR9 activation
to alleviate pulmonary inflammation. This study reports a previously
unreported approach to manipulating the secondary structure tran-
sition of synthetic polypeptides and exploits their flexibility/rigidity
for switchable caging. It also provides an effective strategy to over-
come the multiple physiological barriers encountered after delivery
of antimicrobials via nebulization. The proposed immunoantibacte-
rial approach renders an important addition to the clinical manage-
ment of COPD and thus holds profound translational potential.
Compared with the latest progress in clinical COPD treatment
and recently reported preclinical therapeutic strategies, our IMAMs
have significant advantages. Currently, antibiotics are widely used in
the antibacterial treatment of COPD, and a series of delivery strate-
gies have been developed to improve bactericidal efficiency and re-
duce the side effects (29, 69). However, bacterial resistance still
remains an unresolved challenge. By combining antibiotics with
membrane-­disrupting polypeptides, our IMAMs could not only re-
duce the dose of antibiotics but also avoid drug resistance. Moreover,
recent preclinical anti-­inflammatory strategies for COPD treatment
mainly focus on immunosuppression by means of cytokine-­directed
therapy or the use of chemokine receptor antagonists (70), while
these approaches unavoidably increase the risk of infection. Through
scavenging of bacterial DNA, the IMAMs could inhibit the hyperac-
tivation of immune cells without hurdling the normal antibacterial
immune response, thus favoring improved prognosis.
Despite the potency of IMAMs in sterilization and anti-­
inflammation, it should be recognized that certain limitations must
be overcome with regard to potential translation. For instance, CAZ
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
has low antibacterial efficacy against Gram-­positive bacteria and
may cause adverse side effects during long-­term use, and hence, it is
important to seek alternative antibiotics with higher efficiency and
security. Meanwhile, the repeated nebulization of silica-­based nano-
materials may raise concerns about pulmonary fibrosis, and thus,
the long-­term safety of IMAMs alongside the degradation profiles of
silica nanomaterials should be systematically studied. Specifically,
although it is widely reported that the pulmonary delivery of meso-
porous silicon would not cause severe toxicity or inflammation be-
cause the mesoporous silicon can be degraded by the mononuclear
phagocyte system into orthosilicic acid and lastly excreted by the
renal and hepatobiliary way (71), whether the polypeptide decora-
tion would influence the biosafety and clearance of mesoporous sili-
con still needs to be studied. In addition, a good manufacturing
practice should be established to manage the issue of batch-­to-­batch
variance during large-­scale production. Nevertheless, our study
clearly demonstrates that the vast therapeutic potential of IMAMs is
poised to shift the paradigm in COPD management.
MATERIALS AND METHODS
Preparation and characterization of CAZ-­loaded MEKCA NPs
(MEKCA/CAZ NPs)
MEKCA NPs (10 mg) were ultrasonically dispersed in Na2SO4 so-
lution (2 ml, 0.5 M), into which a solution of CAZ (5 mg) in DMSO
(0.3 ml) was added and stirred at room temperature for 12 hours in
the dark. After ultrafiltration [molecular weight cutoff (MWCO) =
3000 Da] and washing with PBS (pH 7.4) to remove Na2SO4 and the
un-­encapsulated CAZ, MEKCA/CAZ NPs were obtained. MEKCA/
calcein NPs were prepared from MEKCA NPs (10 mg) and calcein
(5 mg) using the same procedure.
To determine the CAZ content in the NPs, MEKCA/CAZ NPs
(2 mg) were incubated with hydrofluoric acid (HF) solution (5 M, 1 ml)
for 30 min to dissolve HMSNs. The concentration of CAZ was deter-
mined by monitoring the absorbance at 255 nm (OD255) by ultraviolet-­
visible spectrometry. DLC (wt %) and DLE (wt %) were calculated
according to the following formula: DLC (%) = amount of loaded
drug × 100%/amount of drug-­loaded NPs; DLE (%) = amount
of loaded drug × 100%/amount of drug in feed. Calcein-­loaded
MEKCA NPs (MEKCA/calcein NPs) were prepared using the same
method. The calcein content was determined by spectrofluorometry
(λex = 495 nm, λem = 515 nm).
In vitro drug release study
To explore the pH-­triggered CAZ release profiles, MEKCA/CAZ
NPs, MEKSA/CAZ NPs, or free CAZ (5 ml, containing 0.8 mg CAZ)
were placed in a dialysis bag (MWCO = 3500 Da), immersed in PBS
(pH 7.4 or 6.5, 45 ml), and incubated at 37°C with gentle shaking
(100 rpm). At desired time intervals, an aliquot of the release medi-
um (2 ml) was withdrawn and refreshed with an equal volume of
fresh medium. The drug release amount in the harvested aliquot was
determined by measuring the OD255.
Mucus penetration of NPs
Purulent sputum obtained from patients with CF was diluted with
PBS (pH 7.4) for 20-­folds and deposited in the apical chamber of the
Transwell insert (0.65 cm2, pore size of 0.4 μm, Corning, NY) which
was placed in the 24-­well plate. Cy5MEKCA NPs or Cy5MEK NPs
(50 μg/ml, 10 μl) were added to the apical chamber, and PBS (pH 7.4,
12 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
600 μl) was added to the basolateral chamber. After incubation at
37°C for 8 hours, an aliquot of the mixture (100 μl) was taken out
from the basolateral compartment. The concentration of NPs in the
aliquot was quantified by spectrofluorometry (λex = 649 nm,
λem = 670 nm). The TE of NPs was calculated according to the fol-
lowing equation: TE (%) = (F0 − F1) × 100%/F0, where F0 and F1
denote the amount of NPs that penetrated to the basolateral medi-
um in the absence or presence of purulent sputum, respectively.
MPT in mucus
The Brownian motion of NPs in the purulent sputum was investi-
gated by the MPT method (72). Briefly, Cy5MEKCA NPs or Cy5MEK
NPs (50 μg/ml, 10 μl) were mixed with diluted CF sputum (20-­fold
diluted with PBS, 200 μl) and transferred to an eight-­well slide
chamber, followed by 30-­min equilibrium at 37°C. Movies of NP
motion were captured at 3 fps using CLSM (Leica, TCS SP5, Germany).
Imaris software was used to analyze the Brownian movement, and
the track of individual NPs was expressed in the form of mean
square displacements (<MSD>) as a function of time scale (τ).
Moreover, the distributions of effective diffusivities (Deff) were
also counted.
Photobleaching/recovery study
Cy5
MEKCA NPs or Cy5MEK NPs (50 μg/ml, 10 μl) were mixed with
diluted CF sputum (20-­fold diluted with PBS, 200 μl), and then in-
cubated at 37°C for 30 min. The region of interest was irradiated
(649 nm, 10 s) at 100% laser power every 17.12 s using the laser
beam of CLSM, and fluorescence images were captured at specific
time intervals. The average fluorescence intensity within the irradi-
ated circular region was quantified using the ImageJ software (57).
Mucin aggregation study
Cy5
MEKCA NPs or Cy5MEK NPs (5 mg/ml, 100 μl) were mixed with
mucin solution (0.3% or 0.5% in PBS, pH 7.4, 2 ml), vortexed, and
incubated at 37°C for 30 min in a shaker. The mixture was centri-
fuged at 1500 rpm for 2 min and the precipitate was washed twice
with PBS. Then, the precipitate was incubated with NaOH (200 μl,
5 M) for 10 min, and the fluorescence intensity was measured
(λex = 649 nm, λem = 670 nm). The amount of the mucin/NP aggre-
gates was then calculated (73).
Establishment of mouse COPD model
The animal experimental protocols were reviewed and approved by
the Institutional Animal Care and Use Committee, Soochow Uni-
versity [SYXK(Su)2017-­0043]. Female BALB/c mice were daily
exposed to cigarette (Tai Shan cigarette, Jinan, China) smoke for
20 days (four cigarettes/day) and were intratracheally administered
with LPS solution (1 mg/ml in saline, 20 μl) on day 7. On day 21,
P. aeruginosa (1 × 105 CFU/ml, 10 μl) was intratracheally adminis-
tered to simulate acute exacerbation of COPD.
In vivo mucus penetration, pulmonary distribution, and
pulmonary retention
On day 22, COPD mice received nebulization of Cy5MEKCA NPs or
Cy5
MEK NPs (50 μg Cy5MEK/ml, 10 μl). Nebulization was conduct-
ed using a MicroSprayer Aerosolizer and FMJ-­250 High-­Pressure
Syringe (YSKD Biotechnology Co. Ltd., Beijing, China). At 1 hour
after administration, mice were euthanized and lung tissues were
collected, fixed in 4% paraformaldehyde, dehydrated in saturated
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
sucrose solution, embedded in OCT, cryo-­sectioned at 10 μm in
thickness, stained with 4′,6-­diamidino-­2-­phenylindole for 10 min,
and observed by CLSM.
In a parallel study, COPD mice received nebulization of free cal-
cein, MEKCA/calcein NPs, Cy5MEKCA NPs, or Cy5MEK NPs at 2.4 mg
calcein/kg or 12.6 mg Cy5MEK/kg. At predetermined time intervals,
mice were euthanized, and lungs were collected and imaged using
the in vivo imaging system (λex = 494 nm, λem = 515 nm for calcein;
λex = 649 nm, λem = 670 nm for Cy5). Lung tissues (100 mg) were
then homogenized with PBS (1 ml) and the concentrations of
Cy5
MEK and calcein were determined by spectrofluorometry. Re-
sults were presented as μg Cy5MEK or calcein per mg tissue.
Determination of MIC against E. coli, P. aeruginosa,
and S. aureus
E. coli, P. aeruginosa, and S. aureus were cultured in LB medium at
37°C. CAZ, MEKCA NPs, DL-­MEKCA NPs, or MEKSA/CAZ NPs
were serially diluted with LB medium (pH 6.5). Then, CAZ or NPs
(200 μl) at different concentrations were mixed with bacteria (2 μl,
1 × 108 CFU/ml) and added into 96-­well plates followed by incuba-
tion at pH 6.5 for 24 hours. In addition, MEKCA NPs were diluted
with LB medium (pH 7.4) and incubated with bacteria at pH 7.4 for
24 hours. Growth inhibition was determined by measuring the
OD600 of each well using a microplate reader and the MIC was con-
sidered as the lowest CAZ or NP concentration that corresponded
to OD ≤ 0.08.
Determination of the FICI
The antibacterial synergism between CAZ and MEKCA NPs was as-
sessed using the checkerboard method (19). Briefly, serial dilutions
of CAZ and MEKCA NPs in LB medium (pH 6.5) were prepared and
added to a 96-­well plate (100 μl per well). Bacterial inoculum was
then added (100 μl per well, 1 × 106 CFU/ml in LB medium), and
the plate was incubated at pH 6.5 for 24 hours. The bacterial concen-
tration was assessed by measuring the OD600. The FICI was calcu-
lated as follows: FICI = (MIC of CAZ in combination/MIC of CAZ
alone) + (MIC of MEKCA in combination/MIC of MEKCA alone).
FICI is interpreted as follows. For FICI ≤ 0.5, there is synergism
between the agents; for 0.5 < FICI ≤ 4, there is no synergism be-
tween the agents; for FICI > 4, there is antagonism between the
agents (19).
Determination of the bactericidal kinetics
Bacteria were grown, diluted, and aliquoted into 96-­well plates as de-
scribed for the MIC assay. Bacterial suspension (150 μl) was mixed
with LB medium (50 μl) containing MEKCA NPs and/or CAZ at vari-
ous final concentrations (0.07 × MIC of MEKCA NPs and 0.04 × MIC
of CAZ for S. aureus; 0.3 × MIC of MEKCA NPs and 0.01 × MIC of
CAZ for E. coli; and 0.09 × MIC of MEKCA NPs and 0.02 × MIC of
CAZ for P. aeruginosa). The plates were sealed and incubated at 37°C
with shaking at 200 rpm. After inoculation for 0, 2, 4, 6, 12, and
24 hours, growth inhibition was determined by measuring the OD600
of each well using a microplate reader.
Live/dead staining of bacteria
The bacterial suspensions (E. coli, P. aeruginosa, and S. aureus,
1 × 106 CFU/ml in LB medium) were incubated with CAZ, MEKCA
NPs, or MEKCA/CAZ NPs at their MIC at pH 6.5 for 2 hours. Bacte-
rial solution treated with PBS was used as the control. Then, the
13 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
bacterial suspensions were subjected to live/dead staining using the
bacterial viability kit containing calcein acetoxymethyl ester (calce-
in AM, 2 μM) and propidium iodide (PI, 1.5 μM) according to the
manufacturer’s protocol. Live bacteria were stained green with cal-
cein AM and dead bacteria were stained red with PI. After washing
with PBS three times, bacteria were observed by CLSM.
Penetration of NPs in biofilms
P. aeruginosa suspensions (1 × 108 CFU/ml, 100 μl) were cultured in
glass dishes with LB medium at 37°C for 2 days and the LB medium
was refreshed every 24 hours to obtain the P. aeruginosa biofilm.
P. aeruginosa biofilms were washed three times with PBS (pH 6.5)
and then incubated with Cy5MEKCA NPs, Cy5MEK NPs, or Cy5MEKSA
NPs (10 μl, 2 μg/ml) for 2 hours. Next, the biofilms were washed
with PBS and then stained with SYTO 9 (2 μM, 100 μl) for 15 min.
After being washed with PBS three times, the penetration of NPs in
biofilms was observed by CLSM. The resulting CLSM planes were
then rendered into three-­dimensional images in the ZEN software
(Zeiss LSM 810). The Cy5 fluorescence intensities at different planes
were analyzed using the ImageJ software. The penetration depth was
defined as the distance from the biofilm surface to the intra-­biofilm
site where the fluorescence intensity decreased by 85% compared to
the maximal fluorescent intensity at the biofilm surface.
Biofilm inhibition assay
Inhibition of biofilm formation was evaluated by crystal violet (CV)
staining (74). Briefly, P. aeruginosa, E. coli, and S. aureus suspensions
(200 μl, 1 × 108 CFU/ml) were incubated with CAZ, MEKCA NPs, or
MEKCA/CAZ NPs at their MIC in 96-­well plates at pH 6.5 for
24 hours. PBS was used as the control. The culture medium was re-
moved, and wells were washed with PBS three times, followed by the
addition of CV solution (200 μl per well, 1% in deionized water) to
stain the biofilms for 20 min. Next, the CV solution was removed,
and the wells were washed three times with PBS. Subsequently, 33%
acetic acid (200 μl) was added and shaken (100 rpm) for 15 min to
dissolve the stained biofilm. The optical density at OD570 (optical
density at 570 nm) was measured using a microplate reader.
Similarly, the biofilm dispersal efficiency of NPs was determined.
Briefly, the formed bacterial biofilm was incubated with CAZ, MEKCA
NPs, or MEKCA/CAZ NPs at their MIC in 96-­well plates at pH 6.5 for
12 hours. PBS was used as the control. The biofilm was then stained
with CV solution as described above followed by determination
of OD570.
Observation of bacteria morphology
E. coli, P. aeruginosa, and S. aureus (1 × 107 CFU/ml in LB medium)
were incubated with PBS or MEKCA NPs at their MIC at pH 6.5 for
1 hour. Bacteria were collected by centrifugation at 5000 rpm for
5 min, washed with PBS three times, and resuspended in PBS (1 ml).
The bacterial suspension (10 μl, 1 × 106 CFU/ml) was placed on
slides, fixed in 2.5% glutaraldehyde solution at 4°C for 2 hours, and
then dehydrated via sequential treatment with 30, 50, 70, 80, 90, and
100% ethanol (30 min for each concentration). After being sprayed
with gold, the samples were observed using FE-­SEM (Zeiss G500).
Patient sputum sample collection and analysis
Sputum samples for the clinical correlation analysis were obtained
from 15 patients with AECOPD and 15 healthy volunteers from the
Affiliated Nantong Hospital 3 of Nantong University. The collection
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
of samples was approved by the Ethics Committee at the Affiliated
Nantong Hospital 3 of Nantong University. All participants in this
study signed an informed consent form before sample collection and
were informed of the potential benefits and risks of participating.
The sputum was diluted 20-­fold with PBS (1 ml) and centrifuged
(1500 rpm, 5 min). The concentrations of TNF-­α and iNOS in the
supernatant were determined using enzyme-­linked immunosorbent
assay (ELISA) kits. In a parallel study, sputum was treated with di-
thiothreitol (DTT) (0.2 M) for 0.5 hours and centrifuged (1500 rpm,
5 min). The supernatant was collected and purified using the DNA
extraction and purification kit, and the cfDNA concentration was
determined using the dsDNA HS assay kit. The clinical correlation
analysis was performed by calculating the correlation between cfDNA
level and TNF-­α or iNOS levels using GraphPad Prism. In addition,
the composition of cfDNA was confirmed using the 16S RNA analysis
using primers (338F 5′-­ACTCCTACGGGAGGCAGCAG-­3′; 806R
5′-­GGACTACHVGGGTWTCTAAT-­3′) at Sinomics (Shanghai, China)
(51). The cycle threshold of each sample was recorded during poly-
merase chain reaction amplification.
Composition determination of cfDNA in sputum before and
after NP treatment
The sputum (20-­fold diluted with PBS) from patients with AECOPD
was treated with DTT (0.2 M) for 0.5 hours and centrifuged (1500
rpm, 5 min). The supernatant was collected and incubated with
MEKCA NPs (200 μg/ml) at pH 6.5 for 1 hour. After centrifuga-
tion (12,000 rpm, 15 min), the precipitate was collected and treated
with heparin (0.5 mg/ml, 100 μl) for 24 hours. After centrifuga
tion (12,000 rpm, 15 min), the supernatant was collected and puri-
fied using the DNA extraction and purification kit. The composi-
tion of the purified cfDNA was confirmed using the 16S RNA
analysis. The abundance of species-­level taxa of cfDNA was deter-
mined by sequencing at Sinomics with a MiSeq PE300 (Illumina,
USA). Sequences were aligned and taxonomically assigned with
the Silva database. Analyses were performed on the I-­Sanger Cloud
Platform.
Scavenging of bacterial DNA
P. aeruginosa suspension (1 ml, 1 × 108 CFU/ml) was centri-
fuged at 5000 rpm for 5 min. The precipitate was collected, lysed
with the radioimmunoprecipitation assay lysis buffer (200 μl),
and purified by the DNA extraction kit to obtain gDNA. In
addition, the supernatant was collected and purified by the
DNA extraction kit to obtain eDNA. In a parallel study, the pre-
cipitate was collected, disrupted with a homogenizer (Thermo
Fisher Scientific FS30D, 100 W, 1.5 min), and centrifuged at
5000 rpm for 5 min. The supernatant was collected as DAMP
DNA. The concentrations of gDNA, eDNA, and DAMPs DNA
were measured using NanoDrop 2000 (Thermo Fisher Scien-
tific, Massachusetts, USA).
The bacterial DNA-­scavenging capacity of MEKCA NPs was
determined via the ethidium bromide (EB) exclusion assay. Brief-
ly, MEKCA NPs were mixed with eDNA, gDNA, or DAMPs DNA
at the weight ratios of 10, 5, 1, 0.2, and 0.1 and incubated at pH 6.5
or 7.4 for 1 hour. As a control, MEKSA NPs were similarly incu-
bated with bacterial DNA at pH 6.5. The mixture was subjected to
electrophoresis in 1% agarose gel at 90 mV for 30 min, and the
DNA migration was visualized by the GE Amersham Imager 600
(Boston, USA).
14 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
In vitro anti-­inflammation efficiency
RAW 264.7 cells or neutrophils (extracted from mouse bone marrow by
using a neutrophil extraction kit) were seeded on 96-­well plates at
1 × 104 cells per well and cultured in a humidified atmosphere with 5%
CO2 for 24 hours. Cells were then incubated with bacterial DNA
(eDNA or gDNA, 100 pg/ml) in the presence of PBS or NPs (MEKCA
NPs or MEKSA NPs, 500 pg/ml) at pH 6.5 for 24 hours. Cells without
gDNA/eDNA challenge or NP treatment served as the control. Extra-
cellular TNF-­α and IL-­6 levels were then quantified using ELISA kits.
In a parallel study, neutrophils were seeded on 12-­well plates at
1 × 104 cells per well, cultured for 24 hours, and then incubated
with gDNA and NPs as described above. After incubation at 37°C
for 2 hours, cells were washed with PBS (1 ml × 3), centrifuged
(1000 rpm, 10 min), and resuspended in PBS (0.5 ml). The single-­cell
suspension was stained with phycoerythrin (PE)–anti-­ CD11b
(1:100) and fluorescein isothiocyanate (FITC)–anti-­Ly6G (1:100) in
cold 1% bovine serum albumin–containing PBS on ice for 20 min.
Cells were washed with PBS before being subjected to flow cytomet-
ric analysis (BD Accuri C6). Flow cytometry data were analyzed us-
ing FlowJo software (Version 10.8.1, BD Biosciences).
To determine the anti-­inflammation efficiency of NPs against
cfDNA from somatic cells, RAW 264.7 cells suspended in PBS
(1 × 106 cells/ml) were sonicated using a homogenizer (100 W, 1.5 min)
and centrifuged at 5000 rpm for 5 min. The supernatant was col-
lected and purified by the DNA extraction kit to obtain cfDNA. Then,
neutrophils were cultured and incubated with cfDNA (2 μg/ml) and
MEKCA NPs (10 μg/ml) as described above. Cells without cfDNA
challenge or NP treatment served as the control. Extracellular TNF-­α
and IL-­6 levels were then quantified using ELISA kits.
Inhibition of TLR activation
To study the inhibition of TLR9 activation, HEK-­Blue mTLR9 cells
were seeded on 96-­well plates at 1 × 104 cells per well and cultured in
Dulbecco’s minimum essential medium containing 10% fetal bovine
serum and 1% penicillin-­streptomycin at 37°C in a humidified atmo-
sphere with 5% CO2 for 24 hours. Cells were then incubated with
DNA (eDNA or gDNA, 100 pg/ml) in the presence of PBS or NPs
(MEKCA NPs or MEKSA NPs, 500 pg/ml) at pH 6.5 or 7.4. Cells with-
out gDNA/eDNA challenge or NP treatment served as the control.
After incubation at 37°C for 24 hours, the culture medium (20 μl) was
collected and incubated with QUANTI-­Blue medium (80 μl) at 37°C
for 30 min. Inhibition of TLR activation was determined by mea-
suring the OD650 using a microplate reader. Results were presented
as percentage absorbance of control cells that were treated with
DNA and PBS.
Blood gas analysis
On days 22, 23, and 24, COPD mice were nebulized with PBS, CAZ,
MEKCA NPs, or MEKCA/CAZ NPs at 2.4 mg CAZ/kg or 12.6 mg
MEKCA/kg once a day. Healthy mice without LPS/P. aeruginosa chal-
lenge or NP treatment served as the normal control. Blood samples
were obtained from the arterial carotids at 24 hours after the third
administration and were directly subjected to the measurement of
PaO2, PaCO2, and pH using the blood gas analyzer (Radiometer,
Shanghai, China).
BALF collection and analysis
COPD mice were nebulized with NPs as described above. At
24 hours after the third administration, mice were euthanized, and
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
BALF was collected and centrifuged (15,000 rpm, 4°C, 10 min). The
precipitate was collected and resuspended in PBS to allow for total
cell counting. The supernatant was collected, and the total protein
level was determined using the BCA kit. The remaining supernatant
was purified using the DNA extraction and purification kit. The
concentration of bacterial DNA in the purified BALF was deter-
mined using the dsDNA HS assay kit for Qubit.
In vivo antibacterial efficacy
In vivo antibacterial efficiency was assessed by calculating the num-
ber of bacteria in BALF. Briefly, the collected BALF was diluted with
PBS for 1000-­folds and then spread on the agar plate at 200 μl per
plate. After a 24-­hour culture at 37°C, the number of bacteria colo-
nies was counted using the colony counter.
In vivo anti-­inflammatory efficiency
At 24 hours after the third administration, mice were euthanized,
and peripheral blood was collected and centrifuged at 500g for
10 min at 4°C to extract the serum. The serum levels of pro-­
inflammatory cytokines (TNF-­α, IL-­1β, and IL-­6) were determined
using ELISA kits. In addition, lung tissues were collected and ho-
mogenized with the passive lysis buffer (1 ml) containing a cocktail
protease inhibitor. The STAT3, p65, and p-­IκBα protein levels in
lung tissues were determined by Western blot. The homogenate was
further centrifuged at 15,000 rpm for 10 min at 4°C, and the concen-
trations of pro-­inflammatory cytokines (TNF-­α, IL-­1β, IL-­8, and IL-­
6), Nrf2, MMP-­9, and oxidative stress markers (MPO, H2O2, SOD,
and MDA) in the supernatant were determined using ELISA kits or
commercial assay kits, respectively.
Immune cell infiltration and macrophage polarization in vivo
At 24 hours after the third administration, the lung tissues were har-
vested, washed with PBS three times, gently ground in PBS, treated
with trypsin (3 ml) at 37°C for 0.5 hours, and filtered with a nylon
mesh to obtain a single-­cell suspension. After being washed with
PBS (1 ml × 3), the single-­cell suspension was divided into four
equal fractions and stained with various antibodies before being
subjected to flow cytometric analysis. In the first fraction, cells were
stained with PE-­conjugated anti-­mouse CD11b antibody (1:100)
and FITC-­conjugated anti-­mouse Ly6G antibody (1:100), and neu-
trophils were identified as Ly6GhighCD11bhigh cells. In the second
fraction, cells were stained with FITC-­conjugated anti-­mouse CD45
antibody (1:100), PerCP-­Cy5.5–conjugated anti-­mouse F4/80 anti-
body (1:100), allophycocyanin (APC)–conjugated anti-­ mouse
CD86 antibody (1:100), and PE-­conjugated anti-­mouse CD206
antibody (1:100), and M1-­ t ype macrophages and M2-­ t ype
macrophages were identified as CD45+F4/80+CD86+ cells and
CD45+F4/80+CD206+ cells, respectively. In the third fraction, cells
were stained with APC-­ conjugated anti-­ mouse CD45 antibody
(1:100), FITC-­conjugated anti-­mouse CD11c antibody (1:100), and
PE-­conjugated anti-­mouse CD11b antibody (1:100), and tissue-­
recruited macrophages were identified as CD45+CD11bhighCD11chigh
cells. In the fourth fraction, cells were stained with PerCP-­Cy5.5–
conjugated anti-­mouse CD4 antibody (1:100), PE-­conjugated anti-­
mouse CD8 antibody (1:100), and FITC-­conjugated anti-­mouse
CD3 antibody (1:100), and CD8+ T cells and CD4+ T cells were
identified as CD8+CD3+ cells and CD4+CD3+ cells, respectively.
Flow cytometry data were analyzed using FlowJo software (Version
10.8.1, BD Biosciences).
15 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Supplementary Materials
This PDF file includes:
Supplementary Text
Figs. S1 to S33
Tables S1 and S2
References
REFERENCES AND NOTES
1. S. Suissa, S. Dell'Aniello, P. Ernst, Long-­term natural history of chronic obstructive
pulmonary disease: Severe exacerbations and mortality. Thorax 67, 957–963 (2012).
2. M. Divo, C. Cote, J. P. de Torres, C. Casanova, J. M. Marin, V. Pinto-­Plata, J. Zulueta,
C. Cabrera, J. Zagaceta, G. Hunninghake, B. Celli, Comorbidities and risk of mortality in
patients with chronic obstructive pulmonary disease. Am. J. Respir. Crit. Care Med. 186,
155–161 (2012).
3. J. Vestbo, S. S. Hurd, A. G. Agusti, P. W. Jones, C. Vogelmeier, A. Anzueto, P. J. Barnes,
L. M. Fabbri, F. J. Martinez, M. Nishimura, R. A. Stockley, D. D. Sin, R. Rodriguez-­Roisin,
Global strategy for the diagnosis, management, and prevention of chronic obstructive
pulmonary disease. Am. J. Respir. Crit. Care Med. 187, 347–365 (2013).
4. B. R. Celli, J. A. Wedzicha, Update on clinical aspects of chronic obstructive pulmonary
disease. N. Engl. J. Med. 381, 1257–1266 (2019).
5. M. Yoshida, S. Minagawa, J. Araya, T. Sakamoto, H. Hara, K. Tsubouchi, Y. Hosaka,
A. Ichikawa, N. Saito, T. Kadota, N. Sato, Y. Kurita, K. Kobayashi, S. Ito, H. Utsumi, H. Wakui,
T. Numata, Y. Kaneko, S. Mori, H. Asano, M. Yamashita, M. Odaka, T. Morikawa,
K. Nakayama, T. Iwamoto, H. Imai, K. Kuwano, Involvement of cigarette smoke-­induced
epithelial cell ferroptosis in COPD pathogenesis. Nat. Commun. 10, 3145 (2019).
6. V. Y. Lin, N. Kaza, S. E. Birket, H. Kim, L. J. Edwards, J. LaFontaine, L. B. Liu, M. Mazur,
S. A. Byzek, J. Hanes, G. J. Tearney, S. V. Raju, S. M. Rowe, Excess mucus viscosity and
airway dehydration impact COPD airway clearance. Eur. Respir. J. 55, 1900419 (2020).
7. L. Martinez-­Solano, M. D. Macia, A. Fajardo, A. Oliver, J. L. Martinez, Chronic Pseudomonas
aeruginosa infection in chronic obstructive pulmonary disease. Clin. Infect. Dis. 47,
1526–1533 (2008).
8. R. P. Dickson, F. J. Martinez, G. B. Huffnagle, The role of the microbiome in exacerbations
of chronic lung diseases. Lancet 384, 691–702 (2014).
9. L. Nici, M. J. Mammen, E. Charbek, P. E. Alexander, D. H. Au, C. M. Boyd, G. J. Criner,
G. C. Donaldson, M. Dreher, V. S. Fan, A. S. Gershon, M. L. K. Han, J. A. Krishnan,
F. J. Martinez, P. M. Meek, M. Morgan, M. I. Polkey, M. A. Puhan, M. Sadatsafavi, D. D. Sin,
G. R. Washko, J. A. Wedzicha, S. D. Aaron, Pharmacologic management of chronic
obstructive pulmonary disease. An official American thoracic society clinical practice
guideline. Am. J. Respir. Crit. Care Med. 201, e56–e69 (2020).
10. X. Ding, C. Yang, W. Moreira, P. Y. Yuan, B. Periaswamy, P. F. de Sessions, H. M. Zhao,
J. R. Tan, A. Lee, K. X. Ong, N. Park, Z. C. Liang, J. L. Hedrick, Y. Y. Yang, A macromolecule
reversing antibiotic resistance phenotype and repurposing drugs as potent antibiotics.
Adv. Sci. 7, 2001374 (2020).
11. D. S. Hui, M. Ip, T. Ling, S. C. Chang, C. H. Liao, C. G. Yoo, D. K. Kim, H. I. Yoon,
V. Udompanich, S. Mogmeud, R. Muttalif, A. M. Salleh, C. Roa, M. Mendoza,
C. Fajardo-­Ang, P. Soepandi, F. Isbaniah, E. Burhan, P. Sudarmono, H. Mangunnegoro,
H. H. Liu, A multicentre surveillance study on the characteristics, bacterial aetiologies and
in vitro antibiotic susceptibilities in patients with acute exacerbations of chronic
bronchitis. Respirology 16, 532–539 (2011).
12. J. M. V. Makabenta, A. Nabawy, C. H. Li, S. Schmidt-­Malan, R. Patel, V. M. Rotello,
Nanomaterial-­based therapeutics for antibiotic-­resistant bacterial infections. Nat. Rev.
Microbiol. 19, 23–36 (2021).
13. J. Y. Xie, M. Zhou, Y. X. Qian, Z. H. Cong, S. Chen, W. J. Zhang, W. N. Jiang, C. Z. Dai, N. Shao,
Z. M. Ji, J. C. Zou, X. M. Xiao, L. Q. Liu, M. Z. Chen, J. Li, R. H. Liu, Addressing MRSA infection
and antibacterial resistance with peptoid polymers. Nat. Commun. 12, 5898 (2021).
14. S. J. Lam, N. M. O'Brien-­Simpson, N. Pantarat, A. Sulistio, E. H. H. Wong, Y. Y. Chen,
J. C. Lenzo, J. A. Holden, A. Blencowe, E. C. Reynolds, G. G. Qiao, Combating multidrug-­
resistant Gram-­negative bacteria with structurally nanoengineered antimicrobial peptide
polymers. Nat. Microbiol. 1, 16162 (2016).
15. N. Chaudhary, B. Aggarwal, V. Saini, P. S. Yavvari, P. Sharma, A. Srivastava, A. Bajaj, Polyaspartate-­
derived synthetic antimicrobial polymer enhances the activity of rifampicin against
multidrug-­resistant Pseudomonas aeruginosa infections. Biomater. Sci. 10, 5158–5171 (2022).
16. R. S. Li, J. H. Liu, C. Wen, Y. R. Shi, J. Ling, Q. Cao, L. Wang, H. Shi, C. Z. Huang, N. Li,
Transformable nano-­antibiotics for mechanotherapy and immune activation against
drug-­resistant Gram-­negative bacteria. Sci. Adv. 9, eadg9601 (2023).
17. S. C. Yang, J. Y. Leong, Y. M. Wang, R. Sim, K. H. Tan, Y. H. Chua, N. Tan, A. L. Z. Lee, J. Tay,
Y. Y. Yang, Drug-­free neutrally charged polypeptide nanoparticles as anticancer agents.
J. Control. Release 345, 464–474 (2022).
18. S. W. Mu, Y. W. Zhu, Y. Wang, S. Qu, Y. C. Huang, L. Zheng, S. Duan, B. R. Yu, M. Qin, F. J. Xu,
Cationic polysaccharide conjugates as antibiotic adjuvants resensitize multidrug-­
resistant bacteria and prevent resistance. Adv. Mater. 34, e2204065 (2022).
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
19. Z. Y. Si, Z. Hou, Y. S. Vikhe, K. R. V. Thappeta, K. Marimuthu, P. P. De, O. T. Ng, P. Li, Y. B. Zhu,
K. Pethe, M. B. Chan-­Park, Antimicrobial effect of a novel chitosan derivative and its
synergistic effect with antibiotics. ACS Appl. Mater. Interfaces 13, 3237–3245 (2021).
20. C. D. Fjell, J. A. Hiss, R. E. W. Hancock, G. Schneider, Designing antimicrobial peptides:
Form follows function. Nat. Rev. Drug Discov. 11, 37–51 (2012).
21. I. I. Ponmalar, J. Swain, J. K. Basu, Escherichia coli response to subinhibitory concentrations
of colistin: Insights from a study of membrane dynamics and morphology. Biomater. Sci.
10, 2609–2617 (2022).
22. Y. Zhao, Q. Q. Guo, X. M. Dai, X. S. Wei, Y. J. Yu, X. L. Chen, C. X. Li, Z. Q. Cao, X. E. Zhang, A
biomimetic non-­antibiotic approach to eradicate drug-­resistant infections. Adv. Mater.
31, e1806024 (2019).
23. P. Salas-­Ambrosio, A. Tronnet, P. Verhaeghe, C. Bonduelle, Synthetic polypeptide
polymers as simplified analogues of antimicrobial peptides. Biomacromolecules 22,
57–75 (2021).
24. R. Wilson, S. Sethi, A. Anzueto, M. Miravitlles, Antibiotics for treatment and prevention of
exacerbations of chronic obstructive pulmonary disease. J. Infect. 67, 497–515 (2013).
25. Y. Ju, C. Cortez-­Jugo, J. Q. Chen, T. Y. Wang, A. J. Mitchell, E. Tsantikos,
N. Bertleff-­Zieschang, Y. W. Lin, J. Y. Song, Y. Z. Cheng, S. Mettu, M. A. Rahim, S. J. Pan,
G. W. Yun, M. L. Hibbs, L. Y. Yeo, C. E. Hagemeyer, F. Caruso, Engineering of nebulized
metal-­phenolic capsules for controlled pulmonary deposition. Adv. Sci. 7, 1902650
(2020).
26. J. Weers, Inhaled antimicrobial therapy—Barriers to effective treatment. Adv. Drug Deliv.
Rev. 85, 24–43 (2015).
27. M. Kesimer, A. A. Ford, A. Ceppe, G. Radicioni, R. Cao, C. W. Davis, C. M. Doerschuk,
N. E. Alexis, W. H. Anderson, A. G. Henderson, R. G. Barr, E. R. Bleecker, S. A. Christenson,
C. B. Cooper, M. K. Han, N. N. Hansel, A. T. Hastie, E. A. Hoffman, R. E. Kanner, F. Martinez,
R. Paine, P. G. Woodruff, W. K. O'Neal, R. C. Boucher, Airway mucin concentration as a
marker of chronic bronchitis. N. Engl. J. Med. 377, 911–922 (2017).
28. C. Menzel, A. Bernkop-­Schnurch, Enzyme decorated drug carriers: Targeted swords to
cleave and overcome the mucus barrier. Adv. Drug Deliv. Rev. 124, 164–174 (2018).
29. Z. B. Li, G. H. Luo, W. P. Hu, J. L. Hua, S. Y. Geng, P. K. Chu, J. Zhang, H. Y. Wang, X. F. Yu,
Mediated drug release from nanovehicles by black phosphorus quantum dots for
efficient therapy of chronic obstructive pulmonary disease. Angew. Chem. Int. Ed. 59,
20568–20576 (2020).
30. L. Wu, W. Shan, Z. R. Zhang, Y. Huang, Engineering nanomaterials to overcome the
mucosal barrier by modulating surface properties. Adv. Drug Deliv. Rev. 124, 150–163
(2018).
31. Z. H. Wang, M. Y. Chu, N. Yin, W. T. Huang, W. Liu, Z. Z. Zhang, J. J. Liu, J. J. Shi, Biological
chemotaxis-­guided self-­thermophoretic nanoplatform augments colorectal cancer
therapy through autonomous mucus penetration. Sci. Adv. 8, eabn3917 (2022).
32. T. T. Yang, A. H. Wang, D. Nie, W. W. Fan, X. H. Jiang, M. R. Yu, S. Y. Guo, C. L. Zhu, G. Wei,
Y. Gan, Ligand-­switchable nanoparticles resembling viral surface for sequential drug
delivery and improved oral insulin therapy. Nat. Commun. 13, 6649 (2022).
33. S. K. Lai, D. E. O'Hanlon, S. Harrold, S. T. Man, Y. Y. Wang, R. Cone, J. Hanes, Rapid transport
of large polymeric nanoparticles in fresh undiluted human mucus. Proc. Natl. Acad. Sci.
U.S.A. 104, 1482–1487 (2007).
34. W. C. Zhou, B. Li, R. T. Min, Z. Z. Zhang, G. T. Huang, Y. C. Chen, B. D. Shen, Q. Zheng,
P. F. Yue, Mucus-­penetrating dendritic mesoporous silica nanoparticle loading drug
nanocrystal clusters to enhance permeation and intestinal absorption. Biomater. Sci. 11,
1013–1030 (2023).
35. C. L. Ge, J. D. Yang, S. Z. Duan, Y. Liu, F. H. Meng, L. C. Yin, Fluorinated α-­helical
polypeptides synchronize mucus permeation and cell penetration toward highly efficient
pulmonary siRNA delivery against acute lung injury. Nano Lett. 20, 1738–1746 (2020).
36. L. M. Ensign, R. Cone, J. Hanes, Oral drug delivery with polymeric nanoparticles: The
gastrointestinal mucus barriers. Adv. Drug Deliv. Rev. 64, 557–570 (2012).
37. J. D. Yang, S. Z. Duan, H. Ye, C. L. Ge, C. X. Piao, Y. B. Chen, M. Y. Lee, L. C. Yin, Pro-­peptide-­
reinforced, mucus-­penetrating pulmonary siRNA delivery mitigates cytokine storm in
pneumonia. Adv. Funct. Mater. 31, 2008960 (2021).
38. J. W. Costerton, P. S. Stewart, E. P. Greenberg, Bacterial biofilms: A common cause of
persistent infections. Science 284, 1318–1322 (1999).
39. H. Koo, R. N. Allan, R. P. Howlin, P. Stoodley, L. Hall-­Stoodley, Targeting microbial biofilms:
Current and prospective therapeutic strategies. Nat. Rev. Microbiol. 15, 740–755 (2017).
40. S. Tian, L. Z. Su, Y. Liu, J. J. Cao, G. Yang, Y. J. Ren, F. Huang, J. F. Liu, Y. L. An,
H. C. van der Mei, H. J. Busscher, L. Q. Shi, Self-­targeting, zwitterionic micellar dispersants
enhance antibiotic killing of infectious biofilms-­An intravital imaging study in mice. Sci.
Adv. 6, eabb1112 (2020).
41. Y. Liu, H. J. Busscher, B. R. Zhao, Y. F. Li, Z. K. Zhang, H. C. van der Mei, Y. J. Ren, L. Q. Shi,
Surface-­adaptive, antimicrobially loaded, micellar nanocarriers with enhanced penetration
and killing efficiency in staphylococcal biofilms. ACS Nano 10, 4779–4789 (2016).
42. D. F. Hu, Y. Y. Deng, F. Jia, Q. Jin, J. Ji, Surface charge switchable supramolecular
nanocarriers for nitric oxide synergistic photodynamic eradication of biofilms. ACS Nano
14, 347–359 (2020).
16 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
43. Y. Y. Wang, A. Shukla, Bacteria-­responsive biopolymer-­coated nanoparticles for biofilm
penetration and eradication. Biomater. Sci. 10, 2831–2843 (2022).
44. J. A. Jurcisek, K. L. Brockman, L. A. Novotny, S. D. Goodman, L. O. Bakaletz, Nontypeable
Haemophilus influenzae releases DNA and DNABII proteins via a T4SS-­like complex and
ComE of the type IV pilus machinery. Proc. Natl. Acad. Sci. U.S.A. 114, E6632–E6641
(2017).
45. H. Hemmi, O. Takeuchi, T. Kawai, T. Kaisho, S. Sato, H. Sanjo, M. Matsumoto, K. Hoshino,
H. Wagner, K. Takeda, S. Akira, A Toll-­like receptor recognizes bacterial DNA. Nature 408,
740–745 (2000).
46. J. I. F. Bass, D. M. Russo, M. L. Gabelloni, J. R. Geffner, M. Giordano, M. Catalano,
A. Zorreguieta, A. S. Trevani, Extracellular DNA: A major proinflammatory component of
Pseudomonas aeruginosa biofilms. J. Immunol. 184, 6386–6395 (2010).
47. Z. Su, L. T. Kong, Y. Dai, J. Tang, J. W. Mei, Z. Z. Qian, Y. Y. Ma, Q. M. Li, S. H. Ju, J. X. Wang,
Bioresponsive nano-­antibacterials for H2S-­sensitized hyperthermia and
immunomodulation against refractory implant-­related infections. Sci. Adv. 8, eabn1701
(2022).
48. P. M. Gallo, G. J. Rapsinski, R. P. Wilson, G. O. Oppong, U. Sriram, M. Goulian, B. Buttaro,
R. Caricchio, S. Gallucci, C. Tukel, Amyloid-­DNA composites of bacterial biofilms stimulate
autoimmunity. Immunity 42, 1171–1184 (2015).
49. Z. X. Tu, Y. L. Zhong, H. Z. Hu, D. Shao, R. N. Haag, M. Schirner, J. Lee, B. Sullenger,
K. W. Leong, Design of therapeutic biomaterials to control inflammation. Nat. Rev. Mater.
7, 557–574 (2022).
50. H. Y. Huang, W. Y. Pan, Y. F. Wang, H. S. Kim, D. Shao, B. D. Huang, T. C. Ho, Y. H. Lao,
C. H. Quek, J. Y. Shi, Q. M. Chen, B. Shi, S. M. Zhang, L. Zhao, K. W. Leong, Nanoparticulate
cell-­free DNA scavenger for treating inflammatory bone loss in periodontitis. Nat.
Commun. 13, 5925 (2022).
51. C. X. Shi, J. Dawulieti, F. Y. Shi, C. Yang, Q. Qin, T. F. Shi, L. Z. Wang, H. Z. Hu, M. D. Sun,
L. Ren, F. Chen, Y. Zhao, F. Liu, M. Li, L. Mu, D. Liu, D. Shao, K. W. Leong, J. She, A
nanoparticulate dual scavenger for targeted therapy of inflammatory bowel disease. Sci.
Adv. 8, eabj2372 (2022).
52. E. K. Holl, K. L. Shumansky, L. B. Borst, A. D. Burnette, C. J. Sample, E. A. Ramsburg,
B. A. Sullenger, Scavenging nucleic acid debris to combat autoimmunity and infectious
disease. Proc. Natl. Acad. Sci. U.S.A. 113, 9728–9733 (2016).
53. H. Y. Liang, Y. Z. Yan, J. J. Wu, X. F. Ge, L. Wei, L. X. Liu, Y. M. Chen, Topical nanoparticles
interfering with the DNA-­LL37 complex to alleviate psoriatic inflammation in mice and
monkeys. Sci. Adv. 6, eabb5274 (2020).
54. C. L. Ge, J. L. Zhu, H. Ye, Y. S. Wei, Y. H. Lei, R. X. Zhou, Z. Y. Song, L. C. Yin, Rational
construction of protein-­mimetic nano-­switch systems based on secondary structure
transitions of synthetic polypeptides. J. Am. Chem. Soc. 145, 11206–11214 (2023).
55. M. H. Xiong, Z. Y. Han, Z. Y. Song, J. Yu, H. Z. Ying, L. C. Yin, J. J. Cheng, Bacteria-­assisted
activation of antimicrobial polypeptides by a random-­coil to helix transition. Angew.
Chem. Int. Ed. 56, 10826–10829 (2017).
56. X. Zhang, H. B. Cheng, W. Dong, M. X. Zhang, Q. Y. Liu, X. H. Wang, J. Guan, H. Y. Wu,
S. R. Mao, Design and intestinal mucus penetration mechanism of core-­shell
nanocomplex. J. Control. Release 272, 29–38 (2018).
57. X. Zhu, J. Wu, W. Shan, Z. Zhou, M. Liu, Y. Huang, Sub-­50 nm nanoparticles with
biomimetic surfaces to sequentially overcome the mucosal diffusion barrier and the
epithelial absorption barrier. Adv. Funct. Mater. 26, 2728–2738 (2016).
58. F. Gao, T. Y. Shao, Y. P. Yu, Y. J. Xiong, L. H. Yang, Surface-­bound reactive oxygen
species generating nanozymes for selective antibacterial action. Nat. Commun. 12,
745 (2021).
59. M. H. Xiong, M. W. Lee, R. A. Mansbach, Z. Y. Song, Y. Bao, R. M. Peek, C. Yao, L. F. Chen,
A. L. Ferguson, G. C. L. Wong, J. J. Cheng, Helical antimicrobial polypeptides with radial
amphiphilicity. Proc. Natl. Acad. Sci. U.S.A. 112, 13155–13160 (2015).
60. M. H. Xiong, Y. Bao, X. Xu, H. Wang, Z. Y. Han, Z. Y. Wang, Y. Q. Liu, S. Y. Huang, Z. Y. Song,
J. J. Chen, R. M. Peek, L. C. Yin, L. F. Chen, J. J. Cheng, Selective killing of Helicobacter pylori
with pH-­responsive helix-­coil conformation transitionable antimicrobial polypeptides.
Proc. Natl. Acad. Sci. U.S.A. 114, 12675–12680 (2017).
61. J. Lee, J. W. Sohn, Y. Zhang, K. W. Leong, D. Pisetsky, B. A. Sullenger, Nucleic acid-­binding
polymers as anti-­inflammatory agents. Proc. Natl. Acad. Sci. U.S.A. 108, 14055–14060
(2011).
62. M. Krieg, CpG motifs in bacterial DNA and their immune effects. Annu. Rev. Immunol. 20,
709–760 (2002).
63. R. Karki, T. D. Kanneganti, The ‘cytokine storm’: Molecular mechanisms and therapeutic
prospects. Trends Immunol. 42, 681–705 (2021).
Zhu et al., Sci. Adv. 10, eabd7904 (2024) 7 February 2024
64. H. Y. Liang, B. Peng, C. Dong, L. X. Liu, J. J. Mao, S. Wei, X. L. Wang, H. S. Xu, J. Shen,
H. Q. Mao, X. H. Gao, K. W. Leong, Y. M. Chen, Cationic nanoparticle as an inhibitor of
cell-­free DNA-­induced inflammation. Nat. Commun. 9, 4291 (2018).
65. G. G. Gunsel, T. M. Conlon, A. Jeridi, R. Kim, Z. Ertuz, N. J. Lang, M. Ansari, M. Novikova,
D. S. Jiang, M. Strunz, M. Gaianova, C. Hollauer, C. Gabriel, I. Angelidis, S. Doll, J. C. Pestoni,
S. L. Edelmann, M. S. Kohlhepp, A. Guillot, K. Bassler, H. P. V. Eeckhoutte, O. Kayalar,
N. Konyalilar, T. Kanashova, S. Rodius, C. Ballester-­Lopez, C. M. G. Robles, N. Smirnova,
M. Rehberg, C. Agarwal, I. Krikki, B. Piavaux, S. E. Verleden, B. Vanaudenaerde,
M. Konigshoff, G. Dittmar, K. R. Bracke, J. L. Schultze, H. Watz, O. Eickelberg, T. Stoeger,
G. Burgstaller, F. Tacke, V. Heissmeyer, Y. Rinkevich, H. Bayram, H. B. Schiller, M. Conrad,
R. Schneider, A. O. Yildirim, The arginine methyltransferase PRMT7 promotes
extravasation of monocytes resulting in tissue injury in COPD. Nat. Commun. 13, 1303
(2022).
66. D. Heise, A. D. Soria, S. Hansen, C. Dambietz, M. Akbarzadeh, A. F. Berg, G. H. Waetzig,
S. A. Jones, R. Dvorsky, M. R. Ahmadian, J. Scheller, J. M. Moll, Selective inhibition of IL-­6
trans-­signaling by a miniaturized, optimized chimeric soluble gp130 inhibits TH17 cell
expansion. Sci. Signal. 14, eabc3480 (2021).
67. G. G. Brusselle, G. F. Joos, K. R. Bracke, New insights into the immunology of chronic
obstructive pulmonary disease. Lancet 378, 1015–1026 (2011).
68. A. Zarbock, M. Allegretti, K. Ley, Therapeutic inhibition of CXCR2 by Reparixin attenuates
acute lung injury in mice. Br. J. Pharmacol. 155, 357–364 (2008).
69. S. Pramanik, S. Mohanto, R. Manne, R. R. Rajendran, A. Deepak, S. J. Edapully, T. Patil,
O. Katari, Nanoparticle-­based drug delivery system: The magic bullet for the treatment of
chronic pulmonary diseases. Mol. Pharm. 18, 3671–3718 (2021).
70. T. T. Hansel, P. J. Barnes, New drugs for exacerbations of chronic obstructive pulmonary
disease. Lancet 374, 744–755 (2009).
71. A. Garcia-­Fernandez, F. Sancenon, R. Martinez-­Manez, Mesoporous silica nanoparticles
for pulmonary drug delivery. Adv. Drug Deliv. Rev. 177, 113953 (2021).
72. P. Mastorakos, A. L. da Silva, J. Chisholm, E. Song, W. K. Choi, M. P. Boyle, M. M. Morales,
J. Hanes, J. S. Suk, Highly compacted biodegradable DNA nanoparticles capable of
overcoming the mucus barrier for inhaled lung gene therapy. Proc. Natl. Acad. Sci. U.S.A.
112, 8720–8725 (2015).
73. C. Bao, B. Liu, B. Li, J. J. Chai, L. W. Zhang, L. L. Jiao, D. Li, Z. Q. Yu, F. Z. Ren, X. H. Shi, Y. Li,
Enhanced transport of shape and rigidity-­tuned α-­lactalbumin nanotubes across
intestinal mucus and cellular barriers. Nano Lett. 20, 1352–1361 (2020).
74. L. Wang, L. W. Sun, X. Zhang, H. Y. Wang, L. J. Song, S. F. Luan, A self-­defense hierarchical
antibacterial surface with inherent antifouling and bacteria-­activated bactericidal
properties for infection resistance. Biomater. Sci. 10, 1968–1980 (2022).
75. M. H. Xiong, M. W. Lee, R. A. Mansbach, Z. Y. Song, Y. Bao, R. M. Peek Jr., C. Yao, L.-F. Chen,
A. L. Ferguson, G. C. L. Wong, J. J. Cheng . Helical antimicrobial polypeptides with radial
amphiphilicity. Proc. Natl. Acad. Sci. U.S.A. 112, 13155–13160 (2015).
76. Y. C. Xia, Z. Y. Song, Z. Z. Tan, T. R. Xue, S. Q. Wei, L. Y. Zhu, Y. F. Yang, H. L. Fu, Y. J. Jiang,
Y. Lin, Y. B. Lu, A. L. Ferguson, J. J. Cheng, Accelerated polymerization of N-­
carboxyanhydrides catalyzed by crown ether. Nat. Commun. 12, 732 (2021).
77. Z. Y. Song, H. L. Fu, J. Wang, J. S. Hui, T. R. Xue, L. A. Pacheco, H. Y. Yan, R. Baumgartner,
Z. Y. Wang, Y. C. Xia, X. F. Wang, L. C. Yin, C. Y. Chen, J. Rodríguez-­López, A. L. Ferguson,
Y. Lin, J. J. Cheng, Synthesis of polypeptides via bioinspired polymerization of in situ
purified N-­carboxyanhydrides. Proc. Natl. Acad. Sci. U. S. A. 116, 10658–10663 (2019).
Acknowledgments
Funding: This work was supported by the National Natural Science Foundation of China
(52325305, 82241008, and 52273144), Collaborative Innovation Center of Suzhou Nano
Science & Technology, 111 Project, Suzhou Key Laboratory of Nanotechnology and
Biomedicine, and Joint International Research Laboratory of Carbon-­Based Functional
Materials and Devices. Author contributions: Methodology: L.Y., K.W.L., J.Z., Xiaohui Li, Y.Z.,
and C.G. Experiment: J.Z., Xiaohui Li, M.H., and Y.W. Supervision: L.Y. Writing—original draft:
J.Z., Xudong Li, C.G., and L.Y. Writing—draft revision and proofreading: J.Z., Y.Z., Y.C., K.W.L., and
L.Y. Competing interests: The authors declare that they have no competing interests. Data
and materials availability: All data needed to evaluate the conclusions in the paper are
present in the paper and/or the Supplementary Materials.
Submitted 3 July 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.abd7904
17 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

BIOPHYSICS Copyright © 2024 The

Authors, some rights
Noncanonical interaction with microtubules via the reserved; exclusive
licensee American
N-­terminal nonmotor domain is critical for the Association for the
Advancement of
functions of a bidirectional kinesin Science. No claim to
original U.S.
Government Works.
Sudhir K. Singh1†‡, Nurit Siegler1†, Himanshu Pandey1†§, Neta Yanir1, Mary Popov1, Distributed under a
Alina Goldstein-­Levitin1, Mayan Sadan1, Garrett Debs2, Raz Zarivach3,4,5, Gabriel A. Frank3,4,5, Creative Commons
Itamar Kass4, Charles V. Sindelar2, Ran Zalk4*, Larisa Gheber1,4* Attribution
NonCommercial
Several kinesin-­5 motors (kinesin-­5s) exhibit bidirectional motility. The mechanism of such motility remains un- License 4.0 (CC BY-­NC).
known. Bidirectional kinesin-­5s share a long N-­terminal nonmotor domain (NTnmd), absent in exclusively plus-­
end–directed kinesins. Here, we combined in vivo, in vitro, and cryo–electron microscopy (cryo-­EM) studies to
examine the impact of NTnmd mutations on the motor functions of the bidirectional kinesin-­5, Cin8. We found
that NTnmd deletion mutants exhibited cell viability and spindle localization defects. Using cryo-­EM, we exam-
ined the structure of a microtubule (MT)-­bound motor domain of Cin8, containing part of its NTnmd. Modeling
and molecular dynamic simulations based on the cryo-­EM map suggested that the NTnmd of Cin8 interacts with
the C-­terminal tail of β-­tubulin. In vitro experiments on subtilisin-­treated MTs confirmed this notion. Last, we
showed that NTnmd mutants are defective in plus-­end–directed motility in single-­molecule and antiparallel MT
sliding assays. These findings demonstrate that the NTnmd, common to bidirectional kinesin-­5s, is critical for their
bidirectional motility and intracellular functions.

INTRODUCTION to plus-­end–directed motility under conditions of lower ionic strength

Kinesin-­5 motor proteins (kinesin-­5s) are essential for mitotic chro-
mosome segregation, facilitating mitotic spindle assembly, mainte-
nance, and elongation [reviewed in (1–4)]. These proteins are bipolar
kinesins, with two pairs of catalytic motor domains located on op-
posite sides of an elongated homotetrameric complex (5–7). By vir-
tue of this bipolar structure, kinesin-­5s cross-­link the antiparallel
microtubules (MTs) of the spindle, and by moving on the two cross-­
linked MTs, they provide the outwardly directed force essential for
spindle pole separation during spindle assembly and anaphase B
spindle elongation (1–4).
Within the mitotic spindle, the MTs are oriented with their dy-
namic plus-­ends facing the middle of the spindle, while the less dy-
namic minus-­ends are concentrated at the poles. It would seem that,
as a result of this MT architecture, the function of kinesin-­5s in sepa-
rating the spindle poles could be achieved only via plus-­end–directed
motility on the two MTs that they cross-­link (1–4). However, several
studies have indicated that the fungal kinesin-­5s, Cin8 and Kip1 from
Saccharomyces cerevisiae and Cut7 from Schizosaccharomyces pombe,
are bidirectional motors that can move in opposite directions on the
MTs, depending on the experimental conditions (8–11). Cin8 and
Kip1 were shown to be minus-­end–directed as single molecules mov-
ing on single MTs under high ionic strength conditions and to switch
1
1Department of Chemistry, Ben-Gurion University of the Negev, Beer-Sheva
8410501, Israel. 2Department of Molecular Biophysics and Biochemistry, Yale Univer-
sity, New Haven, CT 06510, USA. 3Department of Life Sciences, Ben-­Gurion Univer-
sity of the Negev, Beer-­Sheva 8410501, Israel. 4Ilse Katz Institute for Nanoscale
Science & Technology, Ben-­Gurion University of the Negev, Beer-­Sheva 8410501,
Israel. 5National Institute for Biotechnology in the Negev, Ben-­Gurion University of
the Negev, Beer-­Sheva 8410501, Israel.
*Corresponding author. Email: ranz@​bgu.​ac.​il (R.Z.); lgheber@​bgu.​ac.​il (L.G.)
†These authors contributed equally to this work.
‡Present address: Department of Microbiology, Institute of Medical Sciences, Banaras
Hindu University, Varanasi-­221 005, India.
§Present address: Department of Biomedical Engineering, Pennsylvania State
University, University Park, PA, USA.
and in a multi-­motor MT gliding assay (9–11). Cin8 was also shown
to be plus-­end–directed in an antiparallel MT sliding assay under
high ionic strength conditions (10–12). Deletion of the large insert in
loop 8 (L8) of Cin8 enhanced its tendency to move in the minus-­end
direction of the MTs under low ionic strength conditions (11, 13),
while deletion of the C-­terminal tail of Cin8 abolished the minus-­
end–directed preference of single-­molecule motility at high ionic
strength (14). Additional factors that have been shown to induce a
switch from minus-­to plus-­end directionality are motor and nonmo-
tor crowding of the MTs for Cut7 (15) and accumulation in clusters
for Cin8 (12). Nonetheless, despite the above experimental and the-
oretical studies showing that the bidirectional motility of fungal
kinesin-­5s is essential for spindle assembly (12, 16), the molecular
mechanism of bidirectional stepping remains unclear.
A clue to unraveling this mechanism may lie in a unique feature
shared by the bidirectional kinesin-­5s, namely, a long N-­terminal
nonmotor domain (NTnmd) of ~50 to 70 amino acids whose role in
regulating kinesin-­5 functions remains unknown at present. This long
extension is not present in exclusively plus-­end–directed kinesin-­5
and kinesin-­1 motors (1, 3), suggesting that it reflects an evolution-
ary adaptation of the bidirectional kinesin-­5s to their functions. To
reveal the function and conformation of the NTnmd, we started this
study by examining intracellular phenotypes of NTnmd deletion mu-
tants of Cin8. We found that the NTnmd is critical for the intracel-
lular functions of Cin8 since deletions in this region produced Cin8
mutants that were not functional in cells and exhibited spindle local-
ization defects. To investigate these defects, we conducted a cryo–
electron microscopy (cryo-­EM) study designed to determine the
structure of the MT-­bound kinesin-­5 Cin8 motor containing 37
amino acids of the NTnmd. We observed a weak cryo-­EM density
volume close to the MTs, corresponding to the NTnmd of Cin8.
Modeling and molecular dynamics simulations revealed that this re-
gion likely takes the form of an α helix that interacts with β-­tubulin

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 1 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

near its C-­terminal tail. In in vitro experiments on MTs treated with
subtilisin to remove C-­terminal tubulin tails (17, 18), we confirmed
that the NTnmd of Cin8 interacts with the MTs via their C-­terminal
tubulin tails. Thus, we provide here structural and functional evi-
dence of a noncanonical interaction between the NTnmd of a bidi-
rectional kinesin-­5 motor with MTs. In addition, we showed in a
single-­molecule motility assay that N-­terminal deletion mutants of
Cin8 exhibited a minus-­end–directed bias, failing to switch from
minus-­to plus-­end–directed motility. Last, we demonstrated that an
NTnmd deletion mutant that failed to support cell viability as the sole
source of Cin8 was severely defective in producing plus-­end–directed
antiparallel MT sliding. On the basis of our results, we propose that
the noncanonical interaction of bidirectional kinesin-­5s with MT C-­
terminal tails via the NTnmd is crucial for their bidirectional motility
and intracellular functions.
RESULTS
Viability of cells expressing NTnmd deletion mutants of Cin8
The NTnmd of bidirectional kinesin-­5s is considerably longer than
that of the exclusively plus-­end–directed kinesin-­5 and kinesin-­1
motors (Fig. 1A) (3). To examine the role of the sequences in the
NTnmd in regulating the function of Cin8, we generated N-­terminal
deletion variants and examined their function in S. cerevisiae cells.
In the preparation of the N-­terminal deletion mutants, the deletions
were based on the division of the N-­terminal extension into four seg-
ments, according to sequence similarity with other bidirectional, exclu-
sively plus-­end–directed kinesin motors, as follows (Fig. 1, A and B): (i)
the most upstream N-­terminal sequence between the two methionines
of Cin8, i.e., M1 and M39, which is equivalent to the N-­terminal exten-
sion of other bidirectional kinesin-­5s but shares low sequence similarity
with those motors and is rich in positively charged amino acids; (ii)

ScCin8
Bidirectional
ScKip1
kin5 SpCut7
Undetermined AnBimC
AgKin5
directionality kin5 KlKin5
DmKlp61F
Plus-end– XlEg5
directed kin5 HsEg5
DmKHC
Kinesin-1 MmKHC
HsKHC

Bidirectional ScCin8
ScKip1
kin5
SpCut7
Undetermined AnBimC
AgKin5
directionality kin5 KlKin5
DmKlp61F
Plus-end– XlEg5
directed kin5 HsEg5
DmKHC
Kinesin-1 MmKHC
HsKHC
Bidirectional Plus-end–directed kinesin-1 β1
kinesin-5 homology kinesin-5 homology CS

Fig. 1. Viability of cells expressing NTnmd deletion mutants of Cin8. (A) Multiple sequence alignment of the NTnmd of kinesin motors. Kinesin-­5 (bidirectional, of
undetermined directionality, and exclusively plus-­end–directed) and kinesin-­1 motors are indicated on the left. Structural elements are indicated on the bottom. The first
and last amino acids are indicated at the beginning and the end of each sequence. The organisms are as follows: Sc, Saccharomyces cerevisiae; Sp, Schizosaccharomyces
pombe; An, Aspergillus nidulans; Ag, Ashbya gossypii; Kl, Kluyveromyces lactis; Dm, Drosophila melanogaster; Xl, Xenopus laevis; Hs, Homo sapiens; Mm, Mus musculus. (B) Top:
NTnmd sequence of Cin8 (amino acids 1 to 79), numbers indicate amino acid position, letters a to e indicate sequences that were deleted in the different mutants (see
text). Bottom: Viability of cin8Δkip1Δ S. cerevisiae cells expressing wt Cin8 or variants of Cin8 carrying deletions in the NTnmd as the sole source of kinesin-­5 function. Cin8
variants are indicated on the left of each panel. Cells were plated in serial dilutions (indicated by blue triangles) at the temperatures indicated on top. “Vector,” cells contain-
ing a plasmid that does not contain Cin8.

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 2 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
M39-­N45, which shares weak similarity with parts of the N-­terminal
regions of bidirectional kinesin-­5s and fungal kinesin-­5s whose direc-
tionality is undetermined (Fig. 1A); (iii) T46-­P69, which is a segment
equivalent to N-­terminal extensions in bidirectional and plus-­end–
directed kinesin-­5s but exhibits low sequence homology with these
kinesin-­5s; and (iv) N70-­I75, which is the nearest sequence to the mo-
tor domain and shares similarity with sequences believed to form the
cover neck bundle (CNB) and to stabilize neck linker (NL) docking in
kinesin-­1 and kinesin-­5s (19–24). On the basis of this division, we gen-
erated Cin8 variants that carry separate deletions of these segments
(designated Cin8Δ1–38, Cin8Δ39–45, Cin8Δ46–69 and Cin8Δ70–75,
respectively) or deletions of amino acids (v) M39-­P69 (Cin8Δ39–69)
and (vi) M1-­P69 (Cin8Δ1–69).
We started by examining the ability of the N-­terminal deletion
variants to support cell viability as the sole source of kinesin-­5 func-
tion in S. cerevisiae cells deleted of the chromosomal copies of CIN8
and KIP1 (Fig. 1B) (25–27). We had previously demonstrated that
transcription of Cin8 starts from the first methionine, M1, in the
NTnmd and not from M39 (28), and here, we showed that deletion
of the first 38 amino acids of the N-­terminal extension produced a
variant that is functional in cells, similar to wild-­type (wt) Cin8
(Fig. 1C). The variant Cin8Δ46–69 also had little impact on the vi-
ability of cells in the absence of the overlapping Kip1. The internal
deletion of amino acid M39-­N45 (Cin8Δ39–45) reduced cell viabil-
ity at both examined temperatures, while internal deletion of the
sequence nearest to the motor domain, N70-­I75 (Cin8Δ70–75), al-
most completely abolished the viability of the yeast cells (Fig. 1B).
Similar abrogation of cell viability was also caused by deletion of the
first 69 amino acids of Cin8 (Cin8Δ1–69). Last, although separate
deletions of amino acids M39-­N45 and T46-­P69 did not cause a le-
thal phenotype, an internal deletion that combines these two seg-
ments (Cin8Δ39–69) failed to support cell viability (Fig. 1B). These
results indicate that the NTnmd is critical for the intracellular mi-
totic functions of Cin8 and deletion of the major part of the NTnmd
(Cin8Δ1–69) produces a Cin8 that is not functional in cells.
Intracellular localization of NTnmd deletion mutants of Cin8
To further understand the defects inflicted by mutations in NTnmd
(Fig. 1B), we next examined the intracellular localization of green
fluorescent protein (GFP)–tagged N-­terminal deletion variants in
CIN8-­deleted cells (with KIP1 intact) expressing the tdTomato-­tagged
spindle-­pole body component Spc42 (Fig. 2, A and B) (9). We divided
all budded cells into four spindle morphology categories: monopolar
spindles, before spindle formation, in which the two spindle pole
bodies have not been separated; short bipolar spindles, less than
1.8 μm, which indicate either S-­phase or metaphase cells; early ana-
phase spindles of 1.8 to 4 μm; and late anaphase spindles, longer than
4 μm. We found that in cells expressing Cin8Δ39–69 (i.e., the deletion
that causes a lethal phenotype in cin8Δkip1Δ cells) close to 30% of
cells with monopolar spindles did not produce a Cin8-­GFP signal. In
contrast, in cells expressing wt Cin8, only ~7% of cells did not pro-
duce a Cin8-­GFP signal (Fig. 2B). For the cells for which there was no
detectable GFP signal (“no Cin8” phenotype; Fig. 2B), we concluded
that Cin8 is degraded and that it is not cytoplasmatic. This finding
suggests that the deletion of amino acids 39 to 69 generated a Cin8
variant that is less stable in cells, and this reduced stability may par-
tially explain the reduced viability in cells expressing this variant as a
sole source of kinesin-­5. In contrast, Cin8 variants deleted amino ac-
ids 39 to 69 and 70 to 75 (Cin8Δ39–69 and Cin8Δ70–75) [which also
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
exhibited reduced viability in cin8Δkip1Δ cells (Fig. 1B)] exhibited a
mislocalization pattern in the nucleus and/or nuclear MTs, compared
to wt Cin8, which concentrated mainly at the poles (Fig. 2B and
fig. S1). These findings indicate that these variants were either defec-
tive in minus-­end–directed motility and/or exhibited reduced affinity
to MTs (12, 29). Consistently, cells expressing these two variants
(Cin8Δ39–69 and Cin8Δ70–75) exhibited a considerably lower per-
centage of short bipolar spindles compared to wt Cin8, indicating that
these variants were defective in bipolar spindle assembly (Fig. 2B).
Last, cells expressing the Cin8Δ1–69 variant (which was unable to
support viability in cin8Δkip1Δ cells) also exhibited an increased per-
centage of mislocalized motors compared to cells expressing wt Cin8,
but this difference did not reach statistical significance (Fig. 2B). To-
gether, our results indicate that sequences within the NTnmd of Cin8
regulate the intracellular localization of this motor, mainly before
spindle assembly, which, in turn, affects Cin8 functionality in cells.
Cryo-­EM density map of the MT-­bound Cin8 motor domain
To understand the reason behind the defective functionality and lo-
calization of NTnmd mutants, we next examined the structure of the
N-­terminal region by performing single-­particle cryo-­EM analysis
of the Cin8 motor domain bound to MTs in the presence of a non-­
hydrolyzable adenosine 5′-­triphosphate (ATP) analog, adenylyl imi-
dodiphosphate (AMPPNP), followed by fitting a model into the
cryo-­EM map. Since we found that deletion of the first 38 amino
acids does not interfere with Cin8 function (Fig. 1B), the Cin8 mo-
tor domain construct used for cryo-­EM reconstruction started from
the second methionine (M39) and contained 37 amino acids of the
NTnmd (Fig. 3, A and B, and fig. S2, A to D). Previously, it had been
suggested that the large 99–amino acid insert in L8 of Cin8 induces
noncanonical super-­stoichiometric binding of Cin8 to MTs (but was
too flexible to be resolved in a cryo-­EM study) (30). Therefore, to
eliminate the effects of noncanonical MT-­binding and Cin8 cluster-
ing, the large L8 of Cin8 was replaced by the equivalent 7 amino
acids of the Kip1 L8 (31); hereafter referred to as Cin8L8Kip1 (fig. S2,
A to D). A model of Cin8L8Kip1 was then built into the 6.1-­Å resolu-
tion cryo-­EM density map (fig. S2, E and F, and table S1). The local
resolution map generated by the CryoRes software (32) indicates
that the highest resolution of the obtained map is in the motor-­MT
interface and the less certain parts of the model are in the periphery
of the map (fig. S3A). We found that the three-­dimensional (3D)
structure of the Cin8L8Kip1 motor domain shares high similarity with
the 3D structures of kinesin motors from other subfamilies (33–35),
namely, a central β sheet with three α helices on the side facing the
MTs and three α helices on the opposite side (Fig. 3, A, B, and D, and
fig. S3, B and C).
Since the model of the Cin8 motor domain has not been reported
before, we compared different elements of the obtained structure to
those of the exclusively plus-­end–directed Homo sapiens kinesin-­5
Eg5 (crystal structure) (36) and of the bidirectional S. pombe kinesin-
­5 Cut7 (cryo-­EM structure) (Fig. 3D and figs. S3, B and C, and S4)
(37). The structure of the ATP-­binding pocket of the Cin8L8Kip1 mo-
tor domain is consistent with the AMPPNP-­bound state (Fig. 3 and
fig. S4) (38–40). The nucleotide binding and hydrolysis elements
[P-­loop, L9 (switch I), and L11 (switch II) (38–40)] in the 3D model
of the Cin8L8Kip1 motor domain are similar to those in Eg5 and Cut7
(Fig. 3D and figs. S3, B and C, and S4). This result is consistent with
the high sequence similarity between nucleotide binding and hydro-
lysis elements of Cin8, Eg5, and Cut7 (fig. S4C).
3 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Intracellular localization of GFP-­tagged NTnmd deletion mutants of Cin8. S. cerevisiae cells used for this analysis were deleted for the chromosomal copy of
CIN8 and expressed the tdTomato-­tagged spindle pole component Spc42. Cells were imaged in bright field (BF), as well as red (spindle pole body; SPB) and green (Cin8-­
GFP) channels. (A) Representative images of different spindle localization phenotypes of Cin8 variants. Scale bar, 2 μm. (B) Average percentage of budded cells, with dif-
ferent phenotypes in each spindle-­length category, indicated at the bottom of the panel. Cin8 N-­terminal variants are indicated on the right. A schematic representation
of each localization phenotype is shown at the bottom of the panel. *P < 0.05, **P < 0.01, and ***P < 0.001, calculated by analysis of variance (ANOVA), followed by the
Dunnett’s test (84).

L5 of the kinesin motor domain interrupts the sequence of α2 and
is located near the ATP-­binding P-­loop (35, 41). Some previous stud-
ies have shown that L5 of Eg5 regulates adenosine triphosphatase
(ATPase) activity (42–44), and others have suggested that this loop is
involved in a mechanism that gates the ATPase cycle (45–47). L5 has
thus been investigated as a target for small-­molecule allosteric inhibi-
tors, such as monastrol and (+)-­S-­trityl-­​l-­cysteine (STLC), which dis-
rupt the ATPase cycle, thereby arresting cells in mitosis (and as a
result are considered as potential anticancer agents) (44, 48–51). The
L5 of Eg5 is three amino acids longer than the L5s of Cin8 and Cut7
(Fig. 3C), which may explain a previous report that the L5-­binding
inhibitor STLC does not inhibit Cut7 (37). According to our structure,
similar to Cut7, the conformation of L5 in the presence of AMPPNP
is shifted toward the plus-­end of the MT compared to Eg5 (Fig. 3D).
Thus, we conclude that the Cin8 L5 assumes a conformation that is
different from that of the Eg5 L5 but similar to that of the Cut7 L5. In
keeping with this conclusion, we found, in a single-­molecule motility
assay, that in the presence of 150 mM KCl and 1 mM ATP, the move-
ment of Cin8 in the minus-­end direction of the MTs is not inhibited
by STLC (Fig. 3E). This finding indicates that, similar to the bidirec-
tional Cut7, Cin8 is not inhibited by an L5-­binding allosteric inhibitor.
Conformation of the NTnmd
To date, the structure of the NTnmd of bidirectional kinesin motors
has not been examined, most probably because the variants used in
structural studies lacked this region (37). The NTnmd was also absent

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 4 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Reconstruction of the MT-­bound motor domain structure of Cin8L8Kip1. (A) Structure of MT-­bound Cin8L8Kip1, reconstructed from the cryo-­EM density map (gray)
obtained in this study (EMD-­10625). The reconstruction shows the Cin8L8Kip1 motor domain (with the 99–amino acid insert in L8 replaced by seven amino acids of the L8 of
Kip1), in the presence of AMPPNP attached to a Taxol-­stabilized MT. (B) Zoom-­in view of the Cin8L8Kip1 motor domain. The color-­coding is as follows: MT-­bound Cin8L8Kip1
cryo-­EM density map, light gray; Cin8L8Kip1, dusky pink; N-­terminal region, blue; NL, red; α-­tubulin, light blue; β-­tubulin, light purple. The nucleotides are as follows: AMPPNP
bound to Cin8L8Kip1, GTP bound to α-­tubulin, and GDP bound to β-­tubulin. (C) Multiple sequence alignment of the structural elements surrounding L5, namely, P-­loop, α2a,
L5, and α2b. Structural elements are indicated at the bottom of the panel. (D) Side view of Cin8L8Kip1 [the structural model is reconstructed from the cryo-­EM density map
obtained in this study, colored as in panel (B)], with Cin8L8Kip1 L5 indicated in yellow, compared with L5 of Cut7 [Protein Data Bank (PDB): 6S8M, green] and L5 of Eg5 (PDB:
3HQD, light blue). (E) Effect of (+)-­S-­trityl-­​l-­cysteine (STLC) on single-­molecule, minus-­end–directed motility of Cin8. The bar graph shows the average velocity (±SEM)
calculated from a single-­molecule motility assay in the absence or presence of 5 or 50 μM STLC. All the experiments were conducted at a constant concentration of 2%
Me2SO. The numbers of Cin8 motors analyzed are shown in parentheses. *P < 0.05; significance was calculated by ANOVA, followed by Tukey’s test (85).

in the Cin8 variant whose density map was recently published by
Bell et al. (30). We observed that the most pronounced difference be-
tween the cryo-­EM density map of the Cin8 variant obtained in the
present study versus the maps of the Cin8 variant reported by Bell et al.
(30) and the maps of Cut7 (37) and of Eg5 (47) lies in a volume of weak
density (with Gaussian filter applied) that is absent from the maps in
the latter three studies (Fig. 4, A and B). This volume of weak density
emerges near the MT protofilament on the same side of the motor
domain where the N-­terminal and the NL are located (Fig. 4). Since
the variant used in the present study lacks the native large insert in the
L8 of Cin8, this volume of weak density cannot be attributed to this
loop. Since the Cin8 variant used here contains 37 amino acids of the
NTnmd, it is very likely that the additional density results from a se-
quence at the N terminus of Cin8. Thus, we provide direct evidence for
the conformation of the NTnmd and its possible interactions. On the
basis of our analysis, the NTnmd of Cin8 probably interacts with the
MT lattice between two adjacent protofilaments.
To obtain a more detailed insight into the possible structure and
interactions of the NTnmd, we performed homology modeling of
this region, followed by fitting to the obtained cryo-­EM density map
(Fig. 5). The NTnmd of Cin8 was modeled using the structure of
human kinesin-­13 Kif2A bound to bent αβ-­tubulin heterodimers
[Protein Data Bank (PDB) code: 6BBN (52)]. The 3D model was
then fitted to the cryo-­EM density, enabling refinement of the mod-
el. On the basis of this procedure, we were able to fit an additional 30
amino acids from the N terminus (N43 to E72) into the model
(Fig. 5A). Following the modeling, restraint-­free molecular dynam-
ics simulations in water were used to study the structural stability
and dynamics of the Cin8 3D model. The outcome of these simula-
tions indicated that the N terminus is highly flexible, which explains
why the density map for this region is weakly resolved. Although the
NTnmd is highly flexible, the simulation outcome indicates that this
region highly populates a conformation in which it interacts with
tubulin (fig. S5).

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 5 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
A B
Cin8L8Kip1
Cut7
Eg5
70Cin8
Fig. 4. Cryo-­EM density map of the NTnmd of Cin8L8Kip1. Cryo-­EM density map
of Cin8L8Kip1 bound to MTs obtained in this work (light gray), compared to the den-
sity maps of Cut7 (EMD-­3527, PDB: 6S8M, green) (37), Eg5 (EMD-­2077, PDB: 4AQV,
yellow) (47), and a Cin8 variant lacking the first 70 amino acids of the N-­terminal
region (EMD-­8408, with a model of Cin8 obtained in this study, 70Cin8, blue) (30),
all bound to MTs. (A) Cross-­sectional view (from minus-­to plus-­end of the MTs).
(B) A perpendicular-­to-­MT view; directionality of the MT is indicated on the left, at
a 90° rotation compared to (A). The 70Cin8 variant includes the 99–amino acid in-
sert in L8 of Cin8. Its density map was obtained in the presence of ADP-­AlFx which
corresponds to the ADP-­Pi nucleotide state. All density maps (except for the den-
sity map of Eg5, which has a 9.7-­Å resolution) were filtered to a lower resolution of
20 Å to prevent artifacts. The most pronounced difference between the four struc-
tures is a low-­intensity feature in the Cin8L8Kip1 cryo-­EM density map (mauve ar-
rows), which is absent from the cryo-­EM density maps of Cut7, Eg5, and 70Cin8.
This feature is proximal to the N terminus of the motor domain, making contact
with the tubulin surface and extending toward the Cin8L8Kip1 motor domain bound
to the adjacent MT protofilament. Models were fitted to the maps using the “fit in
map” feature of UCSF-­Chimera (75).
The cryo-­EM fitted model reveals that to fit this weak cryo-­EM
density map (Fig. 4), the NTnmd must undergo a turn that diverts it
away from the NL (Fig. 5, green arrow). The model of the N-­terminal
region also displays an α helix (amino acids Q48 to N60), whose
distal N-­terminal end lies in proximity to β-­tubulin (Fig. 5), close to
its C-­terminal tail (fig. S6). Nonetheless, the resolution of this extra
weak density was not sufficient to unambiguously define the confor-
mation or the exact placement of this predicted helix.
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
A B
Fig. 5. Modeling analysis reveals a unique orientation of the NTnmd and its
possible interaction with β-­tubulin. (A and B) Top (A) and back (B) views of the
molecular model based on the cryo-­EM structure and the weak intensity volume
obtained in this work. Color-­coding is as follows: Cin8L8Kip1, pink; NL, red; N70-­I75,
blue; T46-­P69, cyan; N43-­N45, magenta. The first and last Cin8 amino acids of this
model, N43 and F432, are indicated. Green arrows indicate the approximate posi-
tion of the turn that causes the NTnmd to deviate away from the NL.
In vitro motor activity of NTnmd Cin8 variants on
subtilisin-­treated MTs
The structural data presented here suggest that, upon MT binding,
Cin8 interacts with the C-­terminal tails of tubulin via its NTnmd
(Figs. 4 and 5 and fig. S6). To directly test this mechanism of interac-
tion, we examined the binding and motility of wt and selected NTnmd
mutants of Cin8 on MT treated with subtilisin, an agent that has been
shown to remove C-­terminal tubulin tails (Figs. 6 and 7) (17, 18). To
this end, we used a single-­molecule motility assay, in which we char-
acterized the activity of GFP-­tagged full-­length N-­terminal variants,
overexpressed and purified from S. cerevisiae cells, on fluorescently
labeled GMPCPP-­stabilized polarity-­marked MTs (12, 53). These
experiments were conducted in the presence of 1 mM ATP and
70 mM KCl. At higher salt concentrations, no binding of wt Cin8 to
subtilisin-­treated MTs was observed, indicating that C-­terminal
tubulin tails facilitated Cin8 binding to MTs in high-­salt solutions.
We observed MT binding and processive unidirectional motility of
wt Cin8 and its NTnmd deletion mutants (Fig. 6). Following treat-
ment of plus-­end–labeled MTs with subtilisin, very few MTs retained
their plus-­end labeling. Nonetheless, we were able to confirm proces-
sive minus-­end–directed motility for wt Cin8 and NTnmd mutants
on those MTs that retained their plus-­end labeling (Fig. 6A). To de-
termine the velocity of Cin8 variants on subtilisin-­treated MTs, we
performed mean square displacement (MSD) analysis of motility
trajectories (Fig. 6B) (9, 11, 54). MSD analysis was chosen because it
does not require polarity labeling of the MTs and produces similar
results to mean displacement (MD) analysis (11). We found that for
wt Cin8, velocity on subtilisin-­treated MTs was considerably higher
than on untreated MTs (Fig. 6B). This result further supports the no-
tion that C-­terminal tubulin tails contribute to MT binding of Cin8.
In addition, on untreated MTs, the velocity of the NTnmd deletion
mutants was faster than that of the wt Cin8 (Fig. 6B). As discussed
below (Fig. 8 and Table 1), this result indicates a weaker interaction
6 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
A A
B B
Fig. 6. Motility of wt Cin8 and NTnmd deletion mutants on subtilisin-­treated
MTs. (A) Time-­lapse recordings of Cin8-­GFP motor motility (green) on polarity-­
marked, subtilisin-­treated MTs (red). Bright red seeds indicate the plus-­end labeling
of the MTs. Cin8 variants are indicated on the top of the panel. Arrows point to the
Cin8-­GFP motors moving processively in the minus-­end direction, away from the
plus-­end red bright seed. The polarity of the MTs is indicated at the bottom of the panel.
Scale bar, 2 μm; time intervals: 10 s. (B) Representative kymographs and average
velocities (V) ± SD of wt Cin8 and NTnmd deletion mutants Cin8Δ1–38 and
Cin8Δ1–69, indicated on the left, in the presence of 70 mM KCl and 100 μM ATP. MTs
were either untreated (left) or treated with subtilisin (right), indicated on the top of
the panel. Since not all the MTs retained their polarity mark, average velocity (V)
was determined using mean square displacement analysis (9, 11). Velocities are in-
dicated on the right of each kymograph set. Numbers in parentheses indicate the
number of moving motors in each case.
with MTs of the NTnmd mutants compared to the wt Cin8. In con-
trast, on subtilisin-­treated MTs, velocities of the NTnmd mutants
were considerably lower, compared to the wt Cin8 (Fig. 6B, right),
indicating that the effect of the NTnmd on the motility of Cin8 is
strictly dependent on the presence of C-­terminal tubulin tails.
We next examined the binding of wt Cin8 and NTnmd deletion
to the MTs by determining the number of Cin8 motors bound to
subtilisin-­treated MTs compared to untreated MTs in the presence
of ATP (Fig. 7) (13, 29). First, we found that on untreated MTs, the
number of NTnmd mutants (Cin8Δ1–38 and Cin8Δ1–69) bound to
MTs was markedly reduced compared to the wt Cin8 (Fig. 7). This
result indicates that the NTnmd of Cin8 is directly involved in the
MT binding of Cin8. We found that treatment with subtilisin re-
duced the number of wt Cin8 motors bound to MTs (Fig. 7). This
result indicates that the C-­terminal tubulin tails are directly involved
in Cin8 binding to the MTs since, in their absence, Cin8 binding to
MTs is reduced. Treatment with subtilisin did not affect the binding
of the NTnmd deletion mutants to the MTs (Fig. 7), indicating that
when NTnmd is deleted, partially or completely, the C-­terminal tu-
bulin tails do not contribute to MT binding of Cin8. Thus, this result
strongly supports the notion that the NTnmd induces MT binding
via direct interaction with the C-­terminal tails.
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
Fig. 7. Binding of Cin8 NTnmd variants to subtilisin-­treated MTs. (A) Represen-
tative images of Cin8 (green) bound to surface-­attached MTs (red), in the presence
of 100 μM ATP. Cin8 variants are indicated at the top of the figure. MTs were either
untreated (top) or treated with subtilisin (bottom), as indicated on the left. (B) Aver-
age number of motors per micrometer of MT length (±SEM) for Cin8 bound to
surface-­attached MTs in the presence of ATP, in 500-­μm2 fields. Numbers of ana-
lyzed fields are indicated in parentheses. ***P < 0.001; ns, not significant. Signifi-
cance was evaluated using ANOVA, followed by Tukey’s test.
Directionality switch and antiparallel MT sliding of
NTnmd-­deleted mutants of Cin8
To further understand the relationship between the motor functions
of NTnmd mutants of Cin8 and their intracellular defects (Figs. 1B
and 2), we examined the motile properties of the variants in vitro.
Experiments were performed at a saturating ATP concentration and
at an intermediate ionic strength (110 mM KCl) since, under high
ionic strength conditions, some N-­terminal deletion mutants failed
to bind to MTs. We have previously demonstrated that one of the
factors that affect the directionality and velocity of Cin8 is its accu-
mulation in clusters on MTs (12, 54) and that MT-­bound motile clus-
ters of Cin8 can split and merge while remaining motile, indicating
that these clusters are active motors and not merely nonfunctional
aggregates (54). Thus, here, we examined the effects of mutations in
the N-­terminal region on the motility of single molecules and clus-
ters separately. For this purpose, we sorted single molecules from
clusters of Cin8, based on their fluorescence intensities, as previously
described (54); see also Materials and Methods. We found that under
the examined conditions, consistent with previous reports (11, 55),
single molecules and clusters of wt Cin8 moved in a bidirectional
manner, exhibiting motility trajectories in both plus-­and minus-­end
directions on the MTs (Fig. 8).
The motile behavior of the Cin8Δ70–75 variant with the internal
deletion of amino acids 70 to 75 [a sequence homologous to the N-­
terminal region of kinesin motors that is believed to form the CNB and
stabilize NL docking (19–24)] differed markedly from the behaviors of
the wt Cin8 and variants with other N-­terminal deletions (Fig. 8 and
7 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B

Fig. 8. In vitro single-­molecule motility of N-­terminal deletion Cin8 variants. (A) Representative kymographs of movements of Cin8 N-­terminal variants in the pres-
ence of 110 mM KCl. The polarity of the MTs is indicated at the bottom of each kymograph. Asterisks indicate Cin8 clustering at minus-­ends of the MTs. (B) Trajectories of
motility of Cin8 N-­terminal deletion variants. The trajectories of single molecules (left) and clusters (right) were used for the analysis of motile properties. Differentiation
between single motors and clusters was performed as previously described (13, 54). The number of trajectories and the percentage of movements toward the plus-­ends
of MTs are indicated at the bottom right of each graph. Cin8 variants for (A) and (B) are indicated on the left of the figure.

fig. S7). While the other Cin8 variants exhibited processive motility in
either the minus-­or plus-­end directions, the Cin8Δ70–75 mutant ex-
hibited bidirectional motility with frequent changes in directionality
within the same run (Fig. 8 and fig. S7). To determine whether this
behavior was ATP-­dependent, we examined the motility of Cin8Δ70–75
in the presence of adenosine 5′-­diphosphate (ADP) (fig. S7). In con-
trast to wt Cin8, which exhibited a minus-­end–directed bias in the
presence of ATP but not ADP, the displacement of Cin8Δ70–75 was
symmetrically bidirectional, with similar instantaneous displacement
distributions in the presence of either one of the nucleotides. These re-
sults indicate that the motility of Cin8Δ70–75 was ATP-­independent
and was mainly diffusion-­driven (fig. S7). In keeping with these find-
ings, it was previously demonstrated that, in the absence of N-­terminal
sequences, the NL of the human kinesin-­5 Eg5 was unable to dock onto
the motor domain (56). Thus, our results indicate that, similar to
kinesin-­5 Eg5, at least part of the amino acids 70 to 75 sequence of Cin8
may serve as a cover strand, enabling CNB formation during the
ATPase cycle, which is essential for NL docking, correct ATPase activ-
ity, and ATP-­dependent motility.
Examination of the motility of other N-­terminal variants revealed
that deletion of the first 38 amino acids (Cin8Δ1–38) decreased the
percentage of Cin8 clusters bound to MTs (Fig. 8B and Table 1),
which indicates that the first 38 amino acids of Cin8 contribute to the
accumulation in clusters of MT-­bound Cin8. In contrast, deletion of
the first 69 amino acids of the N terminus of Cin8 increased the per-
centage of clusters of MT-­bound motile Cin8 (Fig. 8 and Table 1).
The average intensities of the wt and Cin8Δ1–69 clusters [in arbi-
trary units (a.u.)] were not statistically different [254 ± 8 (SEM;
n = 22) versus 310 ± 22 (SEM; n = 87) for wt and Cin8Δ1–69, re-
spectively]. Although the mechanism by which the deletion of NTnmd
increases motor clustering is not clear, a possible explanation for this
finding is that motor clustering contributes to the stability of this mu-
tant (Cin8Δ1–69).
We found—for both single molecules and clusters of Cin8 variants—­
that all the deletions in the N-­terminal region, except for amino acids
70 to 75, increased the minus-­end–directed bias of Cin8 motors, there-
by increasing the percentage of trajectories in the minus-­end direction
and also increasing their velocity (Fig. 8 and Table 1). We have previ-
ously demonstrated that decreasing the affinity between Cin8 and MTs,
either by increasing the ionic strength or by introducing a negatively
charged phospho-­mimetic mutation at the motor-­MT interface, in-
creased the minus-­end–directed bias, thereby increasing the percent-
age and velocity of minus-­end–directed trajectories (11, 55). Together,
those results and the findings of the current study suggest that deletions
in the N-­terminal region decrease the affinity of Cin8 for the MTs, sup-
porting the notion that the sequences of the N-­terminal region of
Cin8 contribute to direct motor–MT interactions. We also found that
the deletion of these sequences of the N-­terminal region decreased the

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 8 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Table 1. Motile properties of the N-­terminal deletion Cin8 variants. MD analysis was used for the determination of velocities (V), durations of motile
episodes, percent of total runs longer than 60 s, and percentages of plus-­end–directed movements for Cin8 N-­terminal variants that exhibit directional motility.
Negative values of velocity represent motility in the minus-­end direction of the MTs; n, number of Cin8 motors analyzed.
Monomers Clusters

Cin8 V ± SD Dura- % of % of to- n V ± SD Duration % of % of to- n % of

variant (nm/s) tion of (+)-­end tal runs (nm/s) of motile (+)-­end tal runs clusters
motile move- >60 s episodes move- >60 s
epi- ments ± SEM ments
sodes ± (s)
SEM (s)
wt Cin8 −32 ± 0.4 38 ± 4 29 17 28 −18 ± 1 44 ± 5 41 23 22 44
Δ1–38 −65 ± 1 34 ± 4 13 23 40 −56 ± 1 29 ± 4 28 29 14 26
Δ1–69 −175 ± 2 11 ± 1 7 2 56 −107 ± 1 18 ± 1 9 24 96 63
Δ39–45 −281 ± 5 13 ± 1 0 0 46 −205 ± 1 12 ± 1 18 13 39 48

duration of motile episodes and the total duration of motile runs of
both single molecules and clusters of Cin8 (Fig. 8 and Table 1). This
result is also consistent with the decreased affinity between the mutant
motors and MTs. Last, we recently proposed a molecular model that
explains the directionality switch from minus-­to plus-­end–directed
motility. According to this model, any force that opposes the minus-­
end–directed motility of Cin8 (referred to as drag) induces a change to
the plus-­end direction (54). Thus, the NTnmd, which induces such a
switch (Fig. 8 and Table 1), exerts a drag force on the moving Cin8,
consistent with increased MT binding by Cin8 via its NTnmd.
Deletion of the first 38 amino acids caused the smallest effect on
the motility of Cin8, as is consistent with the finding that this variant
does not affect cell viability when expressed as the sole source of
kinesin-­5 function (Fig. 1B); this finding thus demonstrates a cor-
relation between effects on motor motility and cell viability. Consis-
tently, deletions of the first 69 amino acids and amino acids 39 to 45
produced the most pronounced effect on motor motility (Fig. 8 and
Table 1) and markedly affected cell viability (Fig. 1B).
As mentioned above, we have previously demonstrated that accu-
mulation in clusters of MT-­bound motors is one of the major factors
that induces a directionality switch from fast minus-­end–directed mo-
tility to slow plus-­end–directed motility of Cin8 (12, 54). In the current
study, deletions within the NTnmd markedly reduced the percentage
of plus-­end–directed trajectories of single molecules and clusters, com-
pared to wt Cin8 (Fig. 8B and Table 1). The deletion of almost the en-
tire NTnmd (Cin8Δ1–69) produced the most dramatic effect: In this
variant, 63% of the moving motors clustered, but only 9% of those clus-
ters moved in the plus-­end direction of the MTs (Fig. 8B and Table 1).
In comparison, 43% of the moving wt Cin8 motors clustered, and 41%
of those clusters moved in the plus-­end direction of the MTs (Fig. 8B
and Table 1). Thus, our results clearly indicated that deletions within
the NTnmd reduce the ability of Cin8 clusters to switch directionality
from minus-­to plus-­end–directed motility.
An additional condition under which Cin8 has been demonstrat-
ed to be plus-­end–directed is during sliding apart antiparallel MTs
in vitro (10–12). This plus-­end–directed motility is believed to enable
kinesin-­5s, Cin8 included, to separate the spindle poles apart when
performing their critical role in spindle assembly (12). Since deleting
the whole NTnmd of Cin8 (Cin8Δ1–69) prevented the switch from
minus-­to plus-­end–directed motility upon motor clustering in the
single-­molecule motility assay (Fig. 8 and Table 1), we examined how
this mutation affects the ability of Cin8 to produce plus-­end–directed
antiparallel MT sliding in vitro (Fig. 9). For this purpose, we labeled
the plus-­ends of the MT with a bright rhodamine seed (12) and po-
lymerized two sets of MTs, in the presence and absence of biotin-­
labeled tubulin. The biotinylated MTs were attached to the surface,
followed by the addition of biotin-­lacking MTs in the presence of
Cin8 (12, 29). Consistent with previous reports (10, 12, 29), we ob-
served that the majority of antiparallel MT sliding events (86%) me-
diated by the wt Cin8 were plus-­end–directed (Fig. 9 and movie S1),
as is to be expected from the physiological functions of Cin8 in spin-
dle assembly. In contrast, only a third of the MT sliding events pro-
duced by Cin8Δ1–69 were plus-­end–directed. For this mutant, we
found similar percentages of minus-­and plus-­end–directed antipar-
allel MT sliding events and bidirectional MT sliding (Fig. 9 and mov-
ies S2 to S4). The average velocity of plus-­end–directed sliding events
of the Cin8Δ1–69 mutant was significantly faster than that of the wt
Cin8 (Fig. 9), similar to the result obtained in the single-­molecule
motility assay in the minus-­end direction (Figs. 6 and 8 and Table 1).
These results indicate that deletion of the NTnmd severely impairs
the plus-­end–directed motility of Cin8 during the sliding of MTs.
Since minus-­end–directed and bidirectional MT sliding cannot con-
tribute to spindle pole body (SPB) separation during spindle assem-
bly, this sliding behavior is likely to be the reason for the severely
defective viability of S. cerevisiae cells expressing the Cin8Δ1–69
variant as the sole source of kinesin-­5 activity (Fig. 1B). This result is
also consistent with the notion that—in addition to directly increas-
ing the MT-­motor interactions—the NTnmd regulates the direction-
ality of Cin8: The deletion of the NTnmd greatly reduces the ability of
Cin8 to move in the plus-­end direction of the MTs under a variety of
experimental conditions.
Together, our structural and motility data indicate that sequenc-
es in the NTnmd of Cin8 interact with the MT lattice via the C-­
terminal tubulin tails. This interaction increases the affinity of Cin8
for MTs and affects its velocity, directionality, and ability to induce
plus-­end–directed antiparallel MT sliding, thus indicating the im-
portance of the NTnmd for the intracellular functions of Cin8.
DISCUSSION
The bidirectional kinesin-­5s share a large NTnmd of 50 to 70 amino
acids, which is not found in the exclusively plus-­ end–directed

Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024 9 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 9. Antiparallel MT sliding induced by wt Cin8 and Cin8Δ1–69. Representa-
tive time-­lapse recordings of different MT sliding events are presented, as indicated
at the top of the figure. For wt Cin8, plus-­end–directed antiparallel MT sliding is
shown (movie S1); for Cin8Δ1–69, plus-­end–directed (movie S2), minus-­end–directed
(movie S3), and bidirectional (movie S4) antiparallel MT sliding are shown. The time
interval between the frames is indicated at the top of each time-­lapse sequence;
total run time: 300 s. White arrowheads indicate the stationary plus-­end bright la-
bel of the surface-­attached MTs. Yellow arrowheads indicate the mobile plus-­end
bright label of the sliding MTs. Top: Schematic representations of the different slid-
ing events. Rhodamine-­labeled MTs are shown as light-­red tubes; plus-­end label-
ing is indicated by dark red coloring of the MTs; surface binding of the MTs via an
avidin-­biotin bond is indicated by “B”; Cin8 motors are shown as green vertical
dumbbell shapes; MT polarities are indicated by “+” and “–”; red and magenta ar-
rows indicate the direction of movement of the sliding MT; green arrows indicate
plus-­end directionality of the motors; and cyan arrows indicate minus-­end direc-
tionality of the motors. Numbers (light gray) and percentages (dark gray) of the
different categories of MT sliding for each variant are indicated at the bottom of the
figure. Average velocities V (±SEM) for the different MT sliding categories are indi-
cated in green at the bottom of the figure. Numbers in parentheses indicate the
number of events. *P < 0.05, compared to plus-­end–directed sliding of wt Cin8,
calculated by ANOVA followed by Tukey’s test.
kinesin-­5s. The role of this domain in the regulation of the motile
properties and intracellular functions of the bidirectional kinesin-­5s
has not yet been established. To determine the role of this domain in
the function of the bidirectional kinesin-­5 Cin8 motors, we com-
bined intracellular and in vitro experiments with N-­terminal dele-
tion mutants and a cryo-­EM study to examine the structure of the
motor domain of Cin8 bound to MTs in the presence of AMPPNP.
Reconstruction of the structure of the Cin8 motor domain re-
vealed that it exhibits the features of canonical binding to MTs, similar
to kinesin-­like motors, with the NL orienting in the plus-­end direc-
tion of the MT in the presence of AMPPNP. However, in contrast to
the previously published structures of MT-­bound kinesin-­5s (4, 15,
24, 30, 37, 42, 47, 57), the Cin8 used in this study contained an addi-
tional 37 amino acids in its N terminus, upstream of the motor do-
main, thus enabling us to examine the structure of this region for
MT-­bound Cin8. The presence of these 37 amino acids was manifest-
ed in additional cryo-­EM density located close to the MT lattice
(Figs. 4 and 5). Such a density element is absent from the MT-­bound
motor domains of other bidirectional and exclusively plus-­end–directed
kinesin-­5s that lack the NTnmd. This finding indicates that it is likely
that this N-­ terminal region of Cin8 interacts directly with MTs
(Fig. 5). This mechanism of interaction between the NTnmd of Cin8
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
and the C-­terminal tubulin tails was further supported by in vitro ex-
periments on MTs treated with subtilisin (Figs. 6 and 7), which was
previously demonstrated to cleave C-­terminal tubulin tails (17, 18).
Although the interaction of the NTnmd with MTs was previously sug-
gested for the bidirectional S. pombe kinesin-­5 Cut7 (15), our work
provides direct structural and functional evidence for such interac-
tion for a bidirectional kinesin-­5 motor. Moreover, we demonstrated
that the deletion of sequences in this region produced phenotypic de-
fects in intracellular functions, with the deletion of the first 69 amino
acids resulting in a severely defective Cin8 mutant that was unable to
support yeast cell viability as the sole source of kinesin-­5 function
(Fig. 1B). In addition, we demonstrated that the motile and MT bind-
ing properties of these deletion mutants were consistent with their
reduced affinity to MTs (Table 1). We also showed that the N-­terminal
deletion mutants exhibited an impaired ability to switch directionality
from minus-­to plus-­end–directed motility (Fig. 8), which is impor-
tant for their intracellular functions (see below). Last, we demonstrat-
ed that deletion of the entire NTnmd impaired the ability of that
variant to produce efficient plus-­end–directed antiparallel MT sliding
(Fig. 9), which was consistent with its inability to support cell viability
as the sole source of kinesin-­5 function (Fig. 1B). Together, these find-
ings provide direct structural evidence that the NTnmd of a bidirec-
tional kinesin-­5 motor facilitates noncanonical MT binding, which is
important for its motile properties and intracellular functions.
As mentioned above, the structural and functional data presented
here suggest that the NTnmd of Cin8 interacts with the C-­terminal
tail of β-­tubulin (Figs. 5 to 7 and fig. S6). A similar concept was re-
cently demonstrated for the positively charged insertion in loop 12
(K-­loop), which is unique to the kinesin-­3 subfamily: It was demon-
strated that the presence of the K-­loop increases the ATPase rate and
MT affinity in the kinesin-­3 homolog Kif1B (58). Moreover, it was
demonstrated that the K-­loop increases the initial attachment rate of
the kinesin-­3 KIF1 to MTs (59, 60) by inducing a weak MT interac-
tion in the presence of ADP (60). It was also suggested that an α helix
immediately adjacent to the K-­loop facilitates a specific conformation
of the K-­loop that is essential for its function (61). Last, it was demon-
strated that the motility of Kif1A is regulated by pauses induced by
interactions between the K-­loop and the polyglutamylated tubulin C-­
terminal tails (62). Thus, it is possible that, similar to the kinesin-­3
motors, the bidirectional kinesin-­5s interact with the tubulin C-­
terminal tail, via the NTnmd, to enable their unique motor functions.
It is, however, puzzling why noncanonical binding to the MTs is
required for bidirectional kinesin-­5 motor function. It has previously
been reported that processive plus-­end motility of kinesin-­1 motors
is achieved, in part, by a gating mechanism that prevents simultane-
ous strong binding of the two kinesin-­1 motor domain heads to MTs
(63). It has also been reported that for kinesin-­1 to take steps in the
minus-­end direction, the motor had to be mutated to enhance the
flexibility of the NL, thereby enabling the production of a two-­headed
MT-­bound state (64). On the basis of the findings of the current study,
we propose that the noncanonical binding to MTs via the NTnmd
enables the MT two-­headed bound state, which may be one of the in-
termediate states leading to bidirectional stepping. Two of our recently
published studies support this notion. In the first, we demonstrated
that the large insert in L8 of Cin8 can bind MTs, without canonical
MT binding on the motor domain of Cin8. We also demonstrated
that the deletion of this large insert is essential for bidirectional
motility on the single-­molecule level and in surface MT gliding
assays (13). In the second study, we demonstrated that the flexibility
10 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
of NL anchoring to the motor domain of Cin8 is essential for Cin8
functionality (29). In light of the knowledge that flexibility of the NL
is correlated with the two-­headed MT bound state of plus-­end–directed
kinesin-­motors (65–67), it is possible that the flexibility of the NL
docking of Cin8 may also enable the two-­headed bound state, en-
abling stepping in two directions. Thus, we propose that the nonca-
nonical MT binding of Cin8, via the conserved NTnmd and the large
insert in L8, together with the flexible NL docking provide the neces-
sary molecular conditions enabling bidirectional motility.
The three bidirectional kinesin-­5s reported to date are motors of
fungal cells (8–11). We have recently proposed a physiological role for
the bidirectional motility of kinesin-­5s in these cells, in which the
nuclear envelope does not break down during mitosis. In contrast to
higher eukaryote cells, the dynein motor does not participate in the
initial spindle pole separation during mitotic spindle assembly in fun-
gal cells (1, 3). We thus proposed that the minus-­end–directed motil-
ity of kinesin-­5s is required to locate these motors near the spindle
poles before spindle assembly. At this location, the motors cluster and
capture MTs from neighboring poles. Then, by switching from minus-
­to plus-­end–directed motility, clusters of Cin8 slide apart the cap-
tured MTs from neighboring poles (12), providing spindle pole
separation and spindle assembly. On the basis of this model, the abil-
ity to switch from minus-­to plus-­end–directed motility is critical for
the mitotic function of fungal kinesin-­5s. One of the marked differ-
ences between the motile properties of wt Cin8 and the NTnmd mu-
tants is the inability of the mutants to switch to plus-­end–directed
motility—as single molecules and as clusters (Fig. 8 and Table 1).
Last, deletion of the greater part of the NTnmd severely impairs the
ability of the mutant to produce plus-­end–directed antiparallel MT
sliding (Fig. 9), which in turn, causes defects in spindle pole separa-
tion and formation of the mitotic spindle (12). We thus conclude that
the defective intracellular functionality of the NTnmd mutants, espe-
cially that of Cin8Δ1–69 (Figs. 1B and 2), results from their reduced
affinity to MTs and their inability to switch to plus-­end–directed mo-
tility and to produce plus-­end–directed MT sliding (Figs. 8 and 9 and
Table 1). These defective motility properties are the direct outcome of
the absence of noncanonical MT binding of the NTnmd mutants.
Although members of the kinesin superfamily share homology in
elements that are involved in the ATPase cycle and canonical MT
binding, the different subfamilies have evolved unique structural ele-
ments that enable specific motile and mechanical properties adapted
to their intracellular functions. The data presented here indicates
that the NTnmd of Cin8 serves as such an element, enabling the non-
canonical MT binding that is essential for its motor and intracellular
functions. Since the NTnmd is a feature common to bidirectional
kinesin-­5s, it is likely that the NTnmd is essential for the ability to
step in two directions and the intracellular functions of these motors.
MATERIALS AND METHODS
Protein expression and purification
All the functional in vivo and in vitro assays were performed on full-­
length Cin8 variants that carried only deletions in the NTnmd. Since
the stalk (bipolar assembly domain) and tail of these variants remain
intact, they are predicted to be tetramers and to localize properly to
the nucleus (5, 28, 57). Only the construct used for the cryo-­EM study
lacked the large L8 of Cin8 (replaced by the short L8 of Kip1). This
L8-­deleted construct was used so as to eliminate super-­stoichiometric
MT binding [with ~4 motor domains bound to one tubulin dimer
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
(30)] and to prevent possible masking by the large L8 (30) of the den-
sity contributed by the NTnmd in the cryo-­EM map.
For cryo-­EM analysis, the sequence encoded for the Cin8L8Kip1
monomeric motor domain [in which the NTnmd started from M39
and in which the 99–amino acid insert in L8 of Cin8 (amino acids 255
to 353) was replaced with the corresponding 7–amino acid sequence
of Kip1 (amino acids 234 to 240 of Kip1) (fig. S2, A and B)] was cloned
into the pET26B+ vector, with the addition of a 6× His-­tag at the C
terminus of the construct. Expression of the 6× His-­tagged Cin8L8Kip1
construct was induced in 250 ml of BL21-­CodonPlus (DE3)-­RIPL cell
culture with 0.2 mM isopropyl-­β-­​d-­thiogalactopyranoside (IPTG).
The culture was incubated at 18°C for 6 hours, after which the cells
were harvested by centrifugation at 5000 rpm for 15 min at room tem-
perature and resuspended in lysis buffer [50 mM Hepes, 500 mM KCl,
10 mM MgCl2, Triton X-­100 0.01%, 0.5 mM tris(2-­carboxyethyl)phos-
phine (TCEP), and 10% glycerol (pH 7.5)]. Cells were disrupted with a
probe-­type ultrasonicator, followed by centrifugation at 13,500 rpm
for 30 min at 4°C. The 6× His-­tagged protein was purified from the
supernatant by Ni-­NTA affinity chromatography, followed by size-­
exclusion chromatography on a Superdex 200 HR 16/60 column on an
ÄKTA FPLC system (GE Healthcare) equilibrated with size exclusion
elution buffer [50 mM Hepes, 250 mM KCl, 1 mM MgCl2, Triton
X-­100 0.002%, 0.5 mM TCEP, and 6% glycerol (pH 7.5)]. The eluted
proteins were fractionated on SDS—polyacrylamide gel electrophore-
sis and stained with Coomassie Brilliant Blue dye (fig. S2C). For
single-­molecule motility experiments, full-­length GFP-­and 6× His-­
tagged Cin8 was overexpressed and purified from S. cerevisiae cells
using Ni-­NTA affinity and Superose 6 size exclusion columns, as pre-
viously described (12, 54).
Sample preparation for cryo-­EM
The MT polymerization reaction mixture contained final concentra-
tions of porcine tubulin (10 mg/ml; Cytoskeleton Inc.) and 1.5 mM
GTP in general tubulin buffer (GTB; Cytoskeleton Inc.). The mixture
was incubated for 50 min at 37°C, followed by the addition of Taxol
(Sigma-­Aldrich) to a final concentration of 100 μM, and another
20-­min incubation at 37°C. After the polymerization of the MTs, the
mixture was diluted ninefold with GTB supplemented with 10 μM
Taxol (GTB-­Tx), and centrifuged for 20 min at 15,000g. The superna-
tant was discarded, and the pellet was resuspended in 40 μl of GTB-­
Tx. A pre-­reaction mixture was then prepared so as to provide final
concentrations of 0.1 mM AMPPNP and 100 mM KCl in GTB-­Tx. The
Cin8L8Kip1 construct was added in a ratio of 4:1 to the MTs (tubulin
concentration of the MTs was ~0.13 mg/ml). The mixture was mixed
gently and incubated for 10 min. Cryo-­EM samples were prepared by
plunge-­freezing into liquid ethane. Thereafter, 3 μl of the mixture was
applied onto glow-­discharged Lacey F/C grids (01883-­F), blotted for
20 s, and then vitrified by rapidly plunging into liquid ethane using an
FEI Vitrobot operating at room temperature and 100% humidity. The
frozen samples were stored in liquid nitrogen until imaging.
Cryo-­EM data collection and structure determination
Samples were loaded under cryogenic conditions and imaged in low-­
dose mode on an FEI Tecnai F30 Polara microscope (FEI, Eindhoven)
operated at 300 kV. Datasets were collected using SerialEM (68) and an
in-­house prepared semiautomated data collection script (69). Images
were collected by a K2 Summit direct electron detector fitted behind
an energy filter (Gatan Quantum GIF) set to ±10 eV around the zero-­
loss peak. The calibrated pixel size in the sample plane was 1.1 Å. The
11 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
detector was operated in a dose-­fractionated counting mode at a dose
rate of ~8 ē/pixel per second. Each dose-­fractionated movie comprised
50 frames, with a total electron dose of 80 ē/Å2. Data were collected at
a defocus range of −1.0 to −2.0 μm (fig. S2, E and F).
Movies were motion-­corrected using the MotionCor2 program
(70) and contrast transfer function estimation was done with the ctf-
find4.1 program (71). All further processing was done on the motion-­
corrected micrographs. Particles were picked “start to end” manually
with the RELION program (72). MT image segments from selected
imaged filaments (approximately one per 8-­nm repeat) were initially
sorted on the basis of their symmetry type by using RELION 3D clas-
sification. Kinesin-­decorated 12, 13, and 14 protofilament MT models
synthetically generated from PDB models were used to seed the clas-
sification. Segments from the resulting 14 protofilament MT class were
then subjected to 3D structure refinement using RELION with “asym-
metric” helical repeats (14-­fold tubular symmetry was not applied).
Additional “protofilament refinement” steps followed the scheme
presented in (73). Refined MT segment alignment parameters were
then “smoothed” to remove outliers and to fill in the gaps along the fila-
ment trajectories (74). Smoothed alignments were then 14-­fold sym-
metrized to obtain coordinates and Euler angle particles corresponding
to individual protofilaments. A wedge-­shaped mask was then used to
remove a single protofilament from the refined, 14-­protofilament MT
map. The resulting “13-­protofilament” map was subtracted from the
initial particle stack to generate a stack of protofilament particles. Ad-
ditional rounds of RELION refinement were then performed using the
wedge mask, and the resulting protofilament coordinates and 3D vol-
ume were used for a second round of subtraction: 13 individually
aligned, projected protofilaments were subtracted from each segment
to generate a stack of improved protofilament particles. This stack was
then classified to determine the shift register of (i.e., to identify) the al-
ternating ⍺-­ and β-­tubulin subunits within each protofilament. Shifts
were then applied to protofilament coordinates to put all the protofila-
ments in the correct ⍺/β register, and lastly, the protofilament particles
were subjected to an additional round of 3D structure refinement with
RELION. The nominal resolution of the resulting map (0.143 “gold
standard”) was 6.1 Å. An initial Cin8L8Kip1 motor domain cryo-­EM
structure model, based on the crystal structure of Eg5 with AMPPNP
as a nucleotide (PDB: 3HQD), was generated using SWISS-­MODEL
and fitted into the density map by manual real-­space refinement using
Coot v0.9. Figures of the maps and the fitted models were prepared us-
ing UCSF-­Chimera (75). The cryo-­EM map has been deposited in the
EMDB (accession code EMD-­10625).
Homology modeling
To determine the conformation of the NTnmd, homology modeling
was performed. The outcome was then fitted to the experimental den-
sity map obtained by cryo-­EM. Initially, the 3D structure of the hu-
man kinesin-­like protein KIF2A [6BBN (52)] was used as a template
to model the 3D structure of the kinesin-­5 Cin8 (Cin8L8Kip1 motor
domain; amino acids 39 to 438). MODELLER v9.25 was used to build
1000 homology models of Cin8 (76), and the models were sorted ac-
cording to a discrete optimized protein energy (DOPE-­HR; using a
bin size of 0.0125 nm) score. Thereafter, the modeled parts were fitted
into the density seen in the cryo-­EM, as described above.
Molecular dynamics
Simulation systems were prepared using the VMD package version
1.93 (77). The simulated protein was solvated in a water box with a
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
minimum solvation shell of 15 Å thickness. Simulated systems were
then neutralized by adding counter ions at random positions. Before
initiating free molecular dynamics simulations, all systems were mini-
mized and equilibrated. Initially, each system was submitted to 20K
steps of conjugate gradient minimization in which the protein’s heavy
atoms were positionally restrained. This minimization was followed
by two steps of equilibration. First, simulation systems were heated to
300 K while simulated for 200 ps under NVT conditions [moles (N)
volume (V) and temperature (T) are kept constant] with the solute’s
heavy atoms restrained. Second, all systems were simulated using the
NPT ensemble for 100 ns, while restraints were gradually removed.
Last, each system was submitted to free NPT 200-­ns length molecular
dynamics simulations. Simulations were run in triplicate starting from
different random seeds.
During the simulations, a time step of 2 fs was used. In all simu-
lations, periodic boundary conditions with the PME algorithm for
electrostatics were used (78). All simulations were performed at 300 K
and 1 atm, using Langevin dynamics with a Langevin piston with a
damping coefficient of 1 ps−1, an oscillation period of 100 fs, and a
decay time of 50 fs; switching and cutoff distances were 1 and 1.2 nm,
respectively, and pair list distance was 1.4 nm.
All simulations were performed using the NAMD molecular dy-
namics simulation package version 2.13 (78) with CUDA extensions
for accelerated calculation of long-­range electrostatic potentials and
nonbonded forces on graphics processing units (79). All structures
were modeled with the CHARMM 36/CMAP force field (80), and
water was explicitly represented by the TIP3P model (81).
Sequence alignments
All sequence alignments were calculated by the MUSCLE algorithm
(82) and by the Jalview 2.11.1.0 program (83).
S. cerevisiae cell viability assay
Viability assays of cells expressing HA-­tagged Cin8 variants as the sole
source of kinesin-­5 were performed as previously described (13, 29).
Briefly, the S. cerevisiae strains used for the assay were deleted of their
chromosomal copies of CIN8 and KIP1 and contained an endogenic
recessive cycloheximide-­resistance gene (cin8Δkip1Δcyhr), along with
a plasmid (pMA1208) encoding for wt Cin8 and a wt dominant cyclo-
heximide sensitivity gene (CYH). Following transformation with a
plasmid encoding for a specific Cin8 variant, the pMA1208 plasmid
was shuffled out by growth on yeast extract, peptone, and dextrose
medium containing cycloheximide (7.5 μg/ml) at different tempera-
tures. On plates containing cycloheximide, the shuffle plasmid was
removed and only cells expressing functional Cin8 remained viable.
Live cell imaging
Live cells were imaged as previously described (25, 26) using a
S. cerevisiae strain deleted of the chromosomal copy of CIN8 and en-
dogenously expressing both the tdTomato-­tagged SPB component
Spc42 (strains listed in table S2), and 3GFP-­tagged Cin8 variants un-
der a Cin8 native promoter from a centromeric plasmid, listed in ta-
ble S3. S. cerevisiae cells were grown overnight and diluted for 2 hours
before imaging in a medium lacking tryptophan. Images of 15 Z-­stack
planes of live yeast cells placed on a low-­fluorescence agarose gel at
room temperature were acquired in three channels—bright field,
green fluorescence, and red fluorescence, with a 0.5-­μm separation be-
tween the planes. For each variant, three to four experiments were per-
formed in which at least 200 cells were examined for each experiment.
12 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
For spindle length distribution in budded cells, as shown in Fig. 2,
the distance between the two Spc42-­tdTomato SPB components was
measured on the projections of the Z-­stacks. Accordingly, for each
variant, budded cells were categorized into four categories: monopo-
lar cells, short bipolar cells with a short <1.8-­μm spindle, early ana-
phase cells with a spindle length of 1.8 to 4 μm, and late anaphase cells
with a long spindle of length >4 μm. In addition, the budded cells
were categorized into three phenotypes, according to Cin8 localiza-
tion (as shown in Fig. 2) as normal, mislocalized, and no Cin8 signal,
except in the anaphase stages, where Cin8 normally has a diffusive/
mislocalized localization. Cells with monopolar spindles were de-
fined as budded cells with a single Spc42-­tdTomato SPB signal in the
mother cell. The percentage of budded cells in each category was av-
eraged for three experiments for each variant. For each variant, three
to four experiments were performed in which at least 200 cells were
examined for each experiment. Statistical analysis was performed by
one-­way analysis of variance (ANOVA), followed by Dunnett’s analy-
sis (84), compared to wt Cin8 in each spindle length category; degrees
of freedom = 12, and F factors 2.90 (α = 0.05) and 3.81 (α = 0.01).
Single-­molecule motility assay
This assay was performed as previously described (12, 54), using
commercially available tubulin (Cytoskeleton Inc., USA). Briefly,
GMPCPP-­stabilized, fluorescently labeled, biotinylated MTs were
polymerized with tubulin (1 mg/ml), biotinylated tubulin (0.08 mg/
ml), and rhodamine-­labeled tubulin (0.08 mg/ml), in the presence
of 1 mM GMPCPP for 1 hour at 37°C in GTB [80 mM Pipes, 0.5 mM
EGTA, and 2 mM MgCl2 (pH 6.9)]. For plus-­end polarity labeling,
additional rhodamine-­labeled tubulin at a concentration of 0.1 mg/ml
was added to the polymerized MTs, and the mixture was further
incubated for 45 min at 37°C. After polymerization, the mixture
was diluted 10-­fold in GTB containing 20 μM Taxol and centrifuged
for 15 min at 17,000g. The supernatant was discarded, and the MT
pellet was resuspended in 50 μl of GTB with 20 μM Taxol and stored
at 28°C. To prepare subtilisin-­treated MTs (17, 18), the MTs were
stabilized after polymerization with 20 μM Taxol and then exposed
to subtilisin (400 μg/ml) for 30 min at 37°C; thereafter, proteolysis
was inactivated with 10 mM phenylmethylsulfonyl fluoride. The
workup for the resulting mixture was then continued in the same
way as that for the nontreated MTs. Coverslips, silanized with 0.1%
dimethyldichlorosilane in trichloroethylene and assembled into
flow cells, were coated with biotinylated bovine serum albumin
(Sigma-­Aldrich) and then with NeutrAvidin (Life Technologies).
Following the attachment of the biotinylated MTs to the coverslips
through biotin-­NeutrAvidin interaction, 20 μl of the GFP-­labeled
Cin8 N-­terminal variant in motility buffer [50 mM tris-­HCl, 30 mM
Pipes, 110 mM KCl, 2 mM MgCl2, 1 mM EGTA, casein (50 μg/ml),
1 mM dithiothreitol, 1 mM ATP, 10 mM glucose, glucose oxidase
(100 μg/ml), catalase (80 μg/ml), 10 mM phosphocreatine, and
creatine kinase (50 μg/ml) (pH 7.2)] was perfused into the flow cell.
For the Cin8 inhibition assay, STLC stock solution at 50 mM in
100% Me2SO was prepared and subsequently diluted in Me2SO be-
fore addition to the motility buffer so as to maintain a constant
Me2SO concentration of 2%. Motility without STLC but in the pres-
ence of 2% Me2SO was also investigated to ascertain any Me2SO
inhibition effect. For Cin8-­GFP motility and photobleaching ex-
periments, 90 images were acquired at 1 fps with an exposure time
of 800 ms. All data were acquired in at least three independent
experiments.
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
MT sliding assay
MT sliding assay was performed on piranha-­cleaned salinized cover-
slips as previously described (29). MTs were polymerized using com-
mercially available tubulin (Cytoskeleton Inc.) with bright seeds at
their plus-­ends, with and without biotinylated tubulin. Biotinylated
MTs were first adhered to the coverslips via biotin-­NeutrAvidin inter-
action, and then Cin8 was added in motility buffer containing 120 mM
KCl supplemented with 5 μM Taxol, followed by a ~4-­min incubation
of immobilized MTs with Cin8. Then, free MTs (polymerized without
biotinylated tubulin) were added to the motility buffer, as above, and
incubated for an additional ~4 min before imaging. Time-­lapse imag-
ing for 300 s with 1-­s exposure and 10-­s intervals was then acquired.
Motor concentration was in the range of 0.08 to 0.28 mg/ml, adjusted
such that sufficient MT sliding events could be observed without ex-
tensive MT bundling. The velocity of the moving MTs was calculated
by dividing the distance of displacement on the MT by the corre-
sponding time. The directionality of sliding was classified according
to the plus-­end labeling of the stationary and sliding MTs. Bidirec-
tional sliding was assigned when both plus-­and minus-­end–directed
sliding of free MTs along stationary MTs was observed in the same
300-­s time-­lapse sequence. Statistical analysis of sliding velocities
compared to antiparallel plus-­end–directed sliding of wt Cin8 was
calculated by ANOVA followed by Tukey’s test (85).
Imaging
Imaging was performed using a wide-­f ield illumination Micro-
Manager controlled Zeiss Axiovert 200M microscope equipped with
a cooled CCD Andor Neo sCMOS camera and a Plan-­Apochromat
DIC 100×/1.4NA objective (pixel size: 124 nm). Filter sets (Chro-
ma) #49002, ET-­eGFP (FITC/Cy2) [for GFP] and #49004: ET-­CY3/
TRITC [for rhodamine and tdTomato] were used for imaging Cin8-­
GFP, rhodamine-­MTs, and tdTomato-­tagged SPB, respectively.
Single-­molecule motility and MT binding analyses
MT polarity was assigned based on bright plus-­end labeling (12).
Image analysis and kymographs were conducted with ImageJ-­Fiji
software (86). The x-­​y coordinates of motile GFP-­labeled motors for
MD analysis from the obtained kymographs were determined with
the Kymobutler program (87). MD analysis was performed as previ-
ously described (13, 29, 54). The MD values for all possible time inter-
vals (Ƭ) were calculated by averaging each time interval displacement,
calculated for all motility recordings of Cin8-­GFP on the MTs. Aver-
age velocities (V) were derived from the linear MD fit, MD = V·Ƭ.
The sorting of Cin8 motors into single molecules and clusters was
performed as previously described (13, 29, 53, 54). Our examination
of the fluorescence intensity of fluorescent Cin8 molecules as a
function of time showed that the intensity was bleached in steps of
⁓50 a.u., with each step representing the photobleaching of 1 GFP.
Since Cin8 is a homotetrameric motor and one GFP is attached to
each subunit chain, the Cin8 tetramer (referred to as a single Cin8
molecule in this article) contains four GFPs. Hence, all Cin8 motors
having an intensity ≤200 a.u. were likely to be single tetrameric
Cin8 molecules. To minimize the effect of GFP photobleaching on
the determination of the Cin8 cluster size, all motility measurements
were performed only on those Cin8 motors that started moving
within the first 30 s of each measurement. In addition, the Cin8-­GFP
fluorescence intensity was measured only in the first frame of its ap-
pearance, thereby reducing the likelihood of estimating the cluster
size of a photobleached motor. Using this method, we assigned
13 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
intensity ranges of Cin8-­GFP fluorescence as ≤200 and >200 a.u. for
single Cin8 molecules and Cin8 clusters, respectively.
Binding to MTs by the different Cin8 variants was quantitatively
determined as previously described (13, 29). The total number of MT-­
bound motors was divided by the total MT length in fields of 500 μm2.
The average number of MT-­bound motors per MT length (±SEM)
was determined in a number of fields, as indicated in Fig. 7B.
Supplementary Materials
This PDF file includes:
Figs. S1 to S7
Tables S1 to S3
Legend for movies S1 to S4
Other Supplementary Material for this manuscript includes the following:
Movies S1 to S4
REFERENCES AND NOTES
1. H. Pandey, M. Popov, A. Goldstein-­Levitin, L. Gheber, Mechanisms by which kinesin-­5
motors perform their multiple intracellular functions. Int. J. Mol. Sci. 22, 6420 (2021).
2. B. J. Mann, P. Wadsworth, Kinesin-­5 regulation and function in mitosis. Trends Cell Biol. 29,
66–79 (2019).
3. S. K. Singh, H. Pandey, J. Al-­Bassam, L. Gheber, Bidirectional motility of kinesin-­5 motor
proteins: Structural determinants, cumulative functions and physiological roles. Cell. Mol.
Life Sci. 75, 1757–1771 (2018).
4. A. Goulet, C. Moores, New insights into the mechanism of force generation by kinesin-­5
molecular motors. Int. Rev. Cell Mol. Biol. 304, 419–466 (2013).
5. J. E. Scholey, S. Nithianantham, J. M. Scholey, J. Al-­Bassam, Structural basis for the
assembly of the mitotic motor kinesin-­5 into bipolar tetramers. eLife 3, e02217 (2014).
6. S. Acar, D. B. Carlson, M. S. Budamagunta, V. Yarov-­Yarovoy, J. J. Correia, M. R. Ninonuevo,
W. Jia, L. Tao, J. A. Leary, J. C. Voss, J. E. Evans, J. M. Scholey, The bipolar assembly domain
of the mitotic motor kinesin-­5. Nat. Commun. 4, 1343 (2013).
7. A. S. Kashina, R. J. Baskin, D. G. Cole, K. P. Wedaman, W. M. Saxton, J. M. Scholey, A bipolar
kinesin. Nature 379, 270–272 (1996).
8. M. Edamatsu, Bidirectional motility of the fission yeast kinesin-­5, Cut7. Biochem. Biophys.
Res. Commun. 446, 231–234 (2014).
9. V. Fridman, A. Gerson-­Gurwitz, O. Shapira, N. Movshovich, S. Lakamper, C. F. Schmidt,
L. Gheber, Kinesin-­5 Kip1 is a bi-­directional motor that stabilizes microtubules and tracks
their plus-­ends in vivo. J. Cell Sci. 126, 4147–4159 (2013).
10. J. Roostalu, C. Hentrich, P. Bieling, I. A. Telley, E. Schiebel, T. Surrey, Directional switching of
the kinesin Cin8 through motor coupling. Science 332, 94–99 (2011).
11. A. Gerson-­Gurwitz, C. Thiede, N. Movshovich, V. Fridman, M. Podolskaya, T. Danieli,
S. Lakamper, D. R. Klopfenstein, C. F. Schmidt, L. Gheber, Directionality of individual
kinesin-­5 Cin8 motors is modulated by loop 8, ionic strength and microtubule geometry.
EMBO J. 30, 4942–4954 (2011).
12. O. Shapira, A. Goldstein, J. Al-­Bassam, L. Gheber, A potential physiological role for
bi-­directional motility and motor clustering of mitotic kinesin-­5 Cin8 in yeast mitosis. J.
Cell Sci. 130, 725–734 (2017).
13. H. Pandey, S. K. Singh, M. Sadan, M. Popov, M. Singh, G. Davidov, S. Inagaki, J. Al-­Bassam,
R. Zarivach, S. S. Rosenfeld, L. Gheber, Flexible microtubule anchoring modulates the
bi-­directional motility of the kinesin-­5 Cin8. Cell. Mol. Life Sci. 78, 6051–6068 (2021).
14. A. Duselder, V. Fridman, C. Thiede, A. Wiesbaum, A. Goldstein, D. R. Klopfenstein,
O. Zaitseva, M. E. Janson, L. Gheber, C. F. Schmidt, Deletion of the tail domain of the
kinesin-­5 Cin8 affects its directionality. J. Biol. Chem. 19, 620799 (2015).
15. M. Britto, A. Goulet, S. Rizvi, O. von Loeffelholz, C. A. Moores, R. A. Cross,
Schizosaccharomyces pombe kinesin-­5 switches direction using a steric blocking
mechanism. Proc. Natl. Acad. Sci. U.S.A. 113, E7483–E7489 (2016).
16. R. Blackwell, C. Edelmaier, O. Sweezy-­Schindler, A. Lamson, Z. R. Gergely, E. O'Toole,
A. Crapo, L. E. Hough, J. R. McIntosh, M. A. Glaser, M. D. Betterton, Physical determinants
of bipolar mitotic spindle assembly and stability in fission yeast. Sci. Adv. 3, e1601603
(2017).
17. L. Serrano, J. de la Torre, R. B. Maccioni, J. Avila, Involvement of the carboxyl-­terminal
domain of tubulin in the regulation of its assembly. Proc. Natl. Acad. Sci. U.S.A. 81,
5989–5993 (1984).
18. L. Serrano, J. Avila, R. B. Maccioni, Controlled proteolysis of tubulin by subtilisin:
Localization of the site for MAP2 interaction. Biochemistry 23, 4675–4681 (1984).
19. W. Hwang, M. J. Lang, M. Karplus, Force generation in kinesin hinges on cover-­neck
bundle formation. Structure 16, 62–71 (2008).
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
20. A. S. Khalil, D. C. Appleyard, A. K. Labno, A. Georges, M. Karplus, A. M. Belcher, W. Hwang,
M. J. Lang, Kinesin's cover-­neck bundle folds forward to generate force. Proc. Natl. Acad.
Sci. 105, 19247–19252 (2008).
21. Y.-­Z. Geng, T. Li, Q. Ji, S. Yan, Simulation study of interactions between kinesin’s neck
linker and motor domain. Cell. Molec. Bioeng. 7, 99–105 (2014).
22. G. Yi-­Zhao, J. Qing, L. Shu-­Xia, Y. Shi-­Wei, Initial conformation of kinesin's neck linker. Chin.
Phys. B 23, 108701 (2014).
23. B. G. Budaitis, S. Jariwala, D. N. Reinemann, K. I. Schimert, G. Scarabelli, B. J. Grant, D. Sept,
M. J. Lang, K. J. Verhey, Neck linker docking is critical for Kinesin-­1 force generation in
cells but at a cost to motor speed and processivity. eLife 14, 44146 (2019).
24. A. Goulet, J. Major, Y. Jun, S. P. Gross, S. S. Rosenfeld, C. A. Moores, Comprehensive
structural model of the mechanochemical cycle of a mitotic motor highlights molecular
adaptations in the kinesin family. Proc. Natl. Acad. Sci. U.S.A. 111, 1837–1842 (2014).
25. A. Goldstein, D. Goldman, E. Valk, M. Loog, L. J. Holt, L. Gheber, Synthetic-­evolution
reveals narrow paths to regulation of the Saccharomyces cerevisiae mitotic kinesin-­5 Cin8.
Int. J. Biol. Sci. 15, 1125–1138 (2019).
26. A. Goldstein, N. Siegler, D. Goldman, H. Judah, E. Valk, M. Koivomagi, M. Loog, L. Gheber,
Three Cdk1 sites in the kinesin-­5 Cin8 catalytic domain coordinate motor localization and
activity during anaphase. Cell. Mol. Life Sci. 74, 3395–3412 (2017).
27. R. Avunie-­Masala, N. Movshovich, Y. Nissenkorn, A. Gerson-­Gurwitz, V. Fridman,
M. Koivomagi, M. Loog, M. A. Hoyt, A. Zaritsky, L. Gheber, Phospho-­regulation of kinesin-­5
during anaphase spindle elongation. J. Cell Sci. 124, 873–878 (2011).
28. E. R. Hildebrandt, L. Gheber, T. Kingsbury, M. A. Hoyt, Homotetrameric form of Cin8p, a
Saccharomyces cerevisiae kinesin-­5 motor, is essential for its in vivo function. J. Biol. Chem.
281, 26004–26013 (2006).
29. A. Goldstein-­Levitin, H. Pandey, K. Allhuzaeel, I. Kass, L. Gheber, Intracellular functions
and motile properties of bi-­directional kinesin-­5 Cin8 are regulated by neck linker
docking. eLife 10, e71036 (2021).
30. K. M. Bell, H. K. Cha, C. V. Sindelar, J. C. Cochran, The yeast kinesin-­5 Cin8 interacts with
the microtubule in a noncanonical manner. J. Biol. Chem. 12, 797662 (2017).
31. M. A. Hoyt, L. He, K. K. Loo, W. S. Saunders, Two Saccharomyces cerevisiae kinesin-­related
gene products required for mitotic spindle assembly. J. Cell Biol. 118, 109–120 (1992).
32. M. Dai, Z. Dong, K. Xu, Q. C. Zhang, CryoRes: Local resolution estimation of cryo-­EM
density maps by deep learning. J. Mol. Biol. 435, 168059 (2023).
33. J. Turner, R. Anderson, J. Guo, C. Beraud, R. Fletterick, R. Sakowicz, Crystal structure of the
mitotic spindle kinesin Eg5 reveals a novel conformation of the neck-­linker. J. Biol. Chem.
276, 25496–25502 (2001).
34. R. D. Vale, T. Funatsu, D. W. Pierce, L. Romberg, Y. Harada, T. Yanagida, Direct
observation of single kinesin molecules moving along microtubules. Nature 380,
451–453 (1996).
35. E. P. Sablin, F. J. Kull, R. Cooke, R. D. Vale, R. J. Fletterick, Crystal structure of the motor
domain of the kinesin-­related motor ncd. Nature 380, 555–559 (1996).
36. C. L. Parke, E. J. Wojcik, S. Kim, D. K. Worthylake, ATP hydrolysis in Eg5 kinesin involves a
catalytic two-­water mechanism. J. Biol. Chem. 285, 5859–5867 (2010).
37. O. von Loeffelholz, A. Pena, D. R. Drummond, R. Cross, C. A. Moores, Cryo-­EM structure
(4.5-­A) of yeast kinesin-­5-­microtubule complex reveals a distinct binding footprint and
mechanism of drug resistance. J. Molec. Biol. 431, 864–872 (2019).
38. Z. Shang, K. Zhou, C. Xu, R. Csencsits, J. C. Cochran, C. V. Sindelar, High-­resolution
structures of kinesin on microtubules provide a basis for nucleotide-­gated force-­
generation. eLife 3, e04686 (2014).
39. C. V. Sindelar, A seesaw model for intermolecular gating in the kinesin motor protein.
Biophys. Rev. 3, 85–100 (2011).
40. C. V. Sindelar, K. H. Downing, An atomic-­level mechanism for activation of the kinesin
molecular motors. Proc. Natl. Acad. Sci. U.S.A. 107, 4111–4116 (2010).
41. F. J. Kull, E. P. Sablin, R. Lau, R. J. Fletterick, R. D. Vale, Crystal structure of the kinesin motor
domain reveals a structural similarity to myosin. Nature 380, 550–555 (1996).
42. W. M. Behnke-­Parks, J. Vendome, B. Honig, Z. Maliga, C. Moores, S. S. Rosenfeld, Loop L5
acts as a conformational latch in the mitotic kinesin Eg5. J. Biol. Chem. 286, 5242–5253
(2011).
43. J. S. Waitzman, A. G. Larson, J. C. Cochran, N. Naber, R. Cooke, F. Jon Kull, E. Pate, S. E. Rice,
The loop 5 element structurally and kinetically coordinates dimers of the human
kinesin-­5, Eg5. Biophys. J. 101, 2760–2769 (2011).
44. A. G. Larson, N. Naber, R. Cooke, E. Pate, S. E. Rice, The conserved L5 loop establishes the
pre-­powerstroke conformation of the Kinesin-­5 motor, eg5. Biophys. J. 98, 2619–2627
(2010).
45. J. M. Muretta, Y. Jun, S. P. Gross, J. Major, D. D. Thomas, S. S. Rosenfeld, The structural
kinetics of switch-­1 and the neck linker explain the functions of kinesin-­1 and Eg5. Proc.
Natl. Acad. Sci. U.S.A. 112, E6606–E6613 (2015).
46. J. M. Muretta, W. M. Behnke-­Parks, J. Major, K. J. Petersen, A. Goulet, C. A. Moores,
D. D. Thomas, S. S. Rosenfeld, Loop L5 assumes three distinct orientations during the
ATPase cycle of the mitotic kinesin Eg5: A transient and time-­resolved fluorescence study.
J. Biol. Chem. 288, 34839–34849 (2013).
14 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
47. A. Goulet, W. M. Behnke-­Parks, C. V. Sindelar, J. Major, S. S. Rosenfeld, C. A. Moores, The
structural basis of force generation by the mitotic motor kinesin-­5. J. Biol. Chem. 287,
44654–44666 (2012).
48. H. Asraf, R. Avunie-­Masala, M. Hershfinkel, L. Gheber, Mitotic slippage and expression of
survivin are linked to differential sensitivity of human cancer cell-­lines to the Kinesin-­5
inhibitor monastrol. PLOS ONE 10, e0129255 (2015).
49. O. Rath, F. Kozielski, Kinesins and cancer. Nat. Rev. Cancer 12, 527–539 (2012).
50. Z. Maliga, T. J. Mitchison, Small-­molecule and mutational analysis of allosteric Eg5
inhibition by monastrol. BMC Chem. Biol. 6, 2 (2006).
51. I. Leizerman, R. Avunie-­Masala, M. Elkabets, A. Fich, L. Gheber, Differential effects of
monastrol in two human cell lines. Cell. Mol. Life Sci. 61, 2060–2070 (2004).
52. D. Trofimova, M. Paydar, A. Zara, L. Talje, B. H. Kwok, J. S. Allingham, Ternary complex of
Kif2A-­bound tandem tubulin heterodimers represents a kinesin-­13-­mediated
microtubule depolymerization reaction intermediate. Nat. Commun. 9, 2628 (2018).
53. H. Pandey, T. Zvagelsky, M. Popov, M. Sadan, N. Yanir, A. Goldstein-­Levitin, N. Siegler,
S. Hershfinkel, Y. Abraham, R. Abraham, L. A. Gheber, L. Gheber, Motility of single
molecules and clusters of bi-­directional kinesin-­5 Cin8 purified from S. cerevisiae cells. J.
Vis. Exp. 2, 63425 (2022).
54. H. Pandey, E. Reithmann, A. Goldstein-­Levitin, J. Al-­Bassam, E. Frey, L. Gheber,
Drag-­induced directionality switching of kinesin-­5 Cin8 revealed by cluster-­motility
analysis. Sci. Adv. 7, eabc1687 (2021).
55. O. Shapira, L. Gheber, Motile properties of the bi-­directional kinesin-­5 Cin8 are affected
by phosphorylation in its motor domain. Sci Rep-­Uk 6, 25597 (2016).
56. Y. C. Zhao, F. J. Kull, J. C. Cochran, Modulation of the kinesin ATPase cycle by neck linker
docking and microtubule binding. J. Biol. Chem. 285, 25213–25220 (2010).
57. T. Bodrug, E. M. Wilson-­Kubalek, S. Nithianantham, A. F. Thompson, A. Alfieri, I. Gaska,
J. Major, G. Debs, S. Inagaki, P. Gutierrez, L. Gheber, R. J. McKenney, C. V. Sindelar,
R. Milligan, J. Stumpff, S. S. Rosenfeld, S. T. Forth, J. Al-­Bassam, The kinesin-­5 tail domain
directly modulates the mechanochemical cycle of the motor domain for anti-­parallel
microtubule sliding. eLife 9, 51131 (2020).
58. M. Matsushita, K. Y. Fau-­Mitsui, H. M. Fau-­Kanazawa, H. Kanazawa, Altered motor activity
of alternative splice variants of the mammalian kinesin-­3 protein KIF1B. Traffic 10,
1647–1654 (2009).
59. B. G. Budaitis, S. Jariwala, L. Rao, Y. Yue, D. Sept, K. J. Verhey, A. Gennerich, Pathogenic
mutations in the kinesin-­3 motor KIF1A diminish force generation and movement
through allosteric mechanisms. J. Cell Biol. 220, e202004227 (2021).
60. V. Soppina, K. J. Verhey, The family-­specific K-­loop influences the microtubule on-­rate but
not the superprocessivity of kinesin-­3 motors. Mol. Biol. Cell 25, 2161–2170 (2014).
61. A. J. Lam, L. Rao, Y. Anazawa, K. Okada, K. Chiba, M. Dacy, S. Niwa, A. Gennerich,
D. W. Nowakowski, R. J. McKenney, A highly conserved 310 helix within the kinesin motor
domain is critical for kinesin function and human health. Sci. Adv. 7, eabf1002 (2021).
62. D. V. Lessard, O. J. Zinder, T. Hotta, K. J. Verhey, R. Ohi, C. L. Berger, Polyglutamylation of
tubulin's C-­terminal tail controls pausing and motility of kinesin-­3 family member KIF1A.
J. Biol. Chem. 294, 6353–6363 (2019).
63. S. M. Block, Kinesin motor mechanics: Binding, stepping, tracking, gating, and limping.
Biophys. J. 92, 2986–2995 (2007).
64. B. E. Clancy, W. M. Behnke-­Parks, J. O. Andreasson, S. S. Rosenfeld, S. M. Block, A universal
pathway for kinesin stepping. Nat. Struct. Mol. Biol. 18, 1020–1027 (2011).
65. S. Shastry, W. O. Hancock, Interhead tension determines processivity across diverse
N-­terminal kinesins. Proc. Natl. Acad. Sci. U.S.A. 108, 16253–16258 (2011).
66. S. Shastry, W. O. Hancock, Neck linker length determines the degree of processivity in
kinesin-­1 and kinesin-­2 motors. Curr. Biol. 20, 939–943 (2010).
67. A. Yildiz, M. Tomishige, A. Gennerich, R. D. Vale, Intramolecular strain coordinates kinesin
stepping behavior along microtubules. Cell 134, 1030–1041 (2008).
68. D. N. Mastronarde, Automated electron microscope tomography using robust prediction
of specimen movements. J. Struct. Biol. 152, 36–51 (2005).
69. G. Davidov, G. Abelya, R. Zalk, B. Izbicki, S. Shaibi, L. Spektor, D. Shagidov,
E. G. Meyron-­Holtz, R. Zarivach, G. A. Frank, Folding of an intrinsically disordered
iron-­binding peptide in response to sedimentation revealed by Cryo-­EM. J. Am. Chem.
Soc. 142, 19551–19557 (2020).
70. S. Q. Zheng, E. Palovcak, J. P. Armache, K. A. Verba, Y. Cheng, D. A. Agard, MotionCor2:
Anisotropic correction of beam-­induced motion for improved cryo-­electron microscopy.
Nat. Methods 14, 331–332 (2017).
71. A. Rohou, N. Grigorieff, CTFFIND4: Fast and accurate defocus estimation from electron
micrographs. J. Struct. Biol. 192, 216–221 (2015).
Singh et al., Sci. Adv. 10, eadi1367 (2024) 7 February 2024
72. S. He, S. H. W. Scheres, Helical reconstruction in RELION. J. Struct. Biol. 198, 163–176
(2017).
73. G. E. Debs, M. Cha, X. Liu, A. R. Huehn, C. V. Sindelar, Dynamic and asymmetric
fluctuations in the microtubule wall captured by high-­resolution cryoelectron
microscopy. Proc. Natl. Acad. Sci. U.S.A. 117, 16976–16984 (2020).
74. D. Liu, X. Liu, Z. Shang, C. V. Sindelar, Structural basis of cooperativity in kinesin revealed
by 3D reconstruction of a two-­head-­bound state on microtubules. eLife 6, 24490 (2017).
75. E. F. Pettersen, T. D. Goddard, C. C. Huang, G. S. Couch, D. M. Greenblatt, E. C. Meng,
T. E. Ferrin, UCSF Chimera—A visualization system for exploratory research and analysis.
J. Comput. Chem. 25, 1605–1612 (2004).
76. B. Webb, A. Sali, comparative protein structure modeling using MODELLER. Curr. Protoc.
Bioinformatics 54, 5.6.1–5.6.37 (2016).
77. W. Humphrey, A. Dalke, K. Schulten, VMD: Visual molecular dynamics. J. Mol. Graph. 14,
27–38 (1996).
78. U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, L. G. Pedersen, A smooth particle
mesh Ewald method. J. Chem. Phys. 103, 8577–8593 (1995).
79. J. C. Phillips, R. Braun, W. Wang, J. Gumbart, E. Tajkhorshid, E. Villa, C. Chipot, R. D. Skeel,
L. Kalé, K. Schulten, Scalable molecular dynamics with NAMD. J. Comput. Chem. 26,
1781–1802 (2005).
80. R. B. Best, X. Zhu, J. Shim, P. E. Lopes, J. Mittal, M. Feig, A. D. Mackerell Jr., Optimization of
the additive CHARMM all-­atom protein force field targeting improved sampling of the
backbone φ, ψ and side-­chain χ(1) and χ(2) dihedral angles. J. Chem. Theory. Comput. 8,
3257–3273 (2012).
81. W. L. Jorgensen, J. Chandrasekhar, J. D. Madura, R. W. Impey, M. L. Klein, Comparison of
simple potential functions for simulating liquid water. J. Chem. Phys. 79, 926–935 (1983).
82. R. C. Edgar, MUSCLE: A multiple sequence alignment method with reduced time and
space complexity. BMC Bioinformatics 5, 113 (2004).
83. A. M. Waterhouse, J. B. Procter, D. M. Martin, M. Clamp, G. J. Barton, Jalview Version 2—A
multiple sequence alignment editor and analysis workbench. Bioinformatics 25,
1189–1191 (2009).
84. C. W. Dunnett, A multiple comparison procedure for comparing several treatments with a
control. J. Am. Stat. Assoc. 50, 1096–1121 (1955).
85. J. W. Tukey, Comparing individual means in the analysis of variance. Biometrics 5, 99–114
(1949).
86. J. Schindelin, I. Arganda-­Carreras, E. Frise, V. Kaynig, M. Longair, T. Pietzsch, S. Preibisch,
C. Rueden, S. Saalfeld, B. Schmid, J. Y. Tinevez, D. J. White, V. Hartenstein, K. Eliceiri,
P. Tomancak, A. Cardona, Fiji: An open-­source platform for biological-­image analysis. Nat.
Methods 9, 676–682 (2012).
87. M. A. H. Jakobs, A. Dimitracopoulos, K. Franze, KymoButler, a deep learning software for
automated kymograph analysis. eLife 8, e42288 (2019).
Acknowledgments: We thank S. Hershfinkel from the L.G. laboratory for purifying the
full-­length Cin8 variants used in the in vitro experiments with subtilisin-­treated MTs (Figs. 6
and 7). Funding: This work was supported in part by the Israel Science Foundation (ISF) Grants
#286-­18 and #629-­23 and the U.S.-­Israel Binational Research Foundation (BSF) Grant #2019008,
awarded to L.G. C.V.S. and G.D. were supported by the National Institute of General Medical
Sciences of the NIH under award number R01GM110530. Author contributions:
Conceptualization: S.K.S., N.S., H.P., A.G.-­L., R.Zal., and L.G. Funding acquisition: C.V.S. and L.G.
Project administration: C.V.S. and L.G. Supervision: S.K.S., H.P., A.G.-­L., C.V.S., and L.G.
Investigation: S.K.S., N.S., H.P., N.Y., M.P., A.G.-­L., M.S., R.Zar., I.K., R.Zal., and L.G. Validation: S.K.S.,
H.P., N.S., N.Y., M.P., A.G.-­L., M.S., I.K., G.A.F., C.V.S., R.Zal., and L.G. Formal analysis: S.K.S., N.S., H.P.,
N.Y., M.P., A.G.-­L., M.S., I.K., C.V.S., R.Zal., and L.G. Visualization: S.K.S., H.P., N.S., N.Y., M.P., A.G.-­L.,
I.K., R.Zal., and L.G. Data curation: S.K.S., H.P., N.S., A.G.-­L., I.K., C.V.S., R.Zal., and L.G. Resources:
S.K.S., H.P., N.S., G.A.F., C.V.S., R.Zal., and L.G. Methodology: S.K.S., N.S., H.P., G.D., R.Zar., G.A.F.,
I.K., C.V.S., R.Zal., and L.G. Software: G.D., I.K., G.A.F., and C.V.S. Writing—original draft: S.K.S.,
N.S., H.P., N.Y., R.Zar., I.K., R.Zal., and L.G. Writing—review and editing: S.K.S., H.P., N.S., A.G.-­L.,
M.S., I.K., R.Zal., and L.G. Competing interests: The authors declare that they have no
competing interests. Data and materials availability: All data needed to evaluate the
conclusions in the paper are present in the paper and/or the Supplementary Materials. The
cryo-­EM map has been deposited in the EMDB (accession code: EMD-­10625).
Submitted 13 April 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adi1367
15 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

D E V E L O P M E N TA L B I O L O G Y Copyright © 2024 The

Authors, some rights
Rac1 promotes kidney collecting duct repair by reserved; exclusive
licensee American
mechanically coupling cell morphology to mitotic entry Association for the
Advancement of
Science. No claim to
Fabian Bock1,2,3*, Xinyu Dong1, Shensen Li1, Olga M. Viquez1, Eric Sha1, Matthew Tantengco1, original U.S.
Elizabeth M. Hennen4, Erin Plosa5, Alireza Ramezani6,7, Kyle L. Brown1, Young Mi Whang1, Government Works.
Andrew S. Terker1,3, Juan Pablo Arroyo1,2,3, David G. Harrison8, Agnes Fogo9, Distributed under a
Cord H. Brakebusch10, Ambra Pozzi1,2,3,11, Roy Zent1,2,3,12* Creative Commons
Attribution
Prolonged obstruction of the ureter, which leads to injury of the kidney collecting ducts, results in permanent NonCommercial
structural damage, while early reversal allows for repair. Cell structure is defined by the actin cytoskeleton, which License 4.0 (CC BY-­NC).
is dynamically organized by small Rho guanosine triphosphatases (GTPases). In this study, we identified the Rho
GTPase, Rac1, as a driver of postobstructive kidney collecting duct repair. After the relief of ureteric obstruction,
Rac1 promoted actin cytoskeletal reconstitution, which was required to maintain normal mitotic morphology al-
lowing for successful cell division. Mechanistically, Rac1 restricted excessive actomyosin activity that stabilized
the negative mitotic entry kinase Wee1. This mechanism ensured mechanical G2-­M checkpoint stability and pre-
vented premature mitotic entry. The repair defects following injury could be rescued by direct myosin inhibition.
Thus, Rac1-­dependent control of the actin cytoskeleton integrates with the cell cycle to mediate kidney tubular
repair by preventing dysmorphic cells from entering cell division.

INTRODUCTION
Obstructive uropathy is severe kidney damage that results from func-
tional or structural restriction of urinary flow (1, 2). Obstructive
uropathy is common and accounts for nearly 10% of all acute and
chronic kidney disease cases in the general population (1). The build-
up of pressure within the urinary collecting system leads to mechani-
cally induced collecting duct (CD) dilatation with flattening, atrophy,
and apoptosis of CD epithelial cells (3, 4). Without timely decompres-
sion, irreversible progressive renal injury requiring dialysis often en-
sues (5, 6). Despite notable diagnostic and procedural advances in
recognizing and relieving urinary obstruction, the outcomes for pa-
tients after obstructive uropathy have not improved (7). This is due to
a gap in understanding how the CD repairs and a lack of targeted
molecular interventions.
Mechanisms of renal tubular epithelial repair have primarily been
studied in the proximal tubule exposed to ischemic or toxic injury (8).
These kidney epithelial cells are largely quiescent at baseline but rap-
idly reenter the cell cycle when injured and replace the cells that died
by a process of dedifferentiation (8). Arrest in the G2-­M phase of the
cell cycle is a well-­described maladaptive reaction of proximal tubule
injury leading to epithelial cell senescence that promotes local inflam-
mation and fibrosis (8). Similar to the proximal tubule, the CD cells
1
Division of Nephrology and Hypertension, Department of Medicine, Vanderbilt Uni-
versity Medical Center, Nashville, TN, USA. 2Department of Veterans Affairs Hospital,
Tennessee Valley Healthcare System, Nashville, TN, USA. 3Vanderbilt Center for Kidney
Disease, Vanderbilt University Medical Center, Nashville, TN, USA. 4Department of Bio-
medical Engineering, Vanderbilt University, Nashville, TN, USA. 5Division of Neonatol-
ogy, Department of Pediatrics, Vanderbilt University Medical Center, Nashville, TN,
USA. 6Interdisciplinary Center for Quantitative Modeling in Biology, University of Cali-
fornia, Riverside, CA, USA. 7Department of Physics and Astronomy, University of Cali-
fornia, Riverside, CA, USA. 8Division of Clinical Pharmacology, Department of Medicine,
Vanderbilt University Medical Center, Nashville, TN, USA. 9Department of Pathology,
Microbiology and Immunology, Vanderbilt University Medical Center, Nashville, TN,
USA. 10Biotech Research Center, University of Copenhagen, Copenhagen DK-­2200,
Denmark. 11Department of Physiology and Molecular Biophysics, Vanderbilt Univer-
sity School of Medicine, Nashville, TN, USA. 12Department of Cell and Developmental
Biology, Vanderbilt University School of Medicine, Nashville, TN, USA.
*Corresponding author. Email: fabian.​bock@​vumc.​org (F.B.); roy.​zent@​vumc.​org (R.Z.)
have notable self-­repair capacity mediated by reentry into the cell cy-
cle and increasing cell division following obstruction (3). However,
the mechanisms of CD repair are poorly defined and likely morpho-
logically and functionally distinct from the proximal tubule.
Kidney epithelial cell repair after injury requires dynamic mor-
phological change mediated by the actin cytoskeleton (9). There is
substantial actin dysregulation soon after initial tubular injury, and
loss of actin organization destroys architecture and hampers tubular
resorptive function (9, 10). Consistent with this, animal studies indi-
cate that therapeutic tubular actin cytoskeletal stabilization by en-
hancing actin branching protects the kidney in various experimental
injury models (11). Despite these findings, it is unclear what critical
cellular functions are required in kidney tubular epithelial repair,
whether dynamic actin reorganization plays a role and how this is
molecularly coordinated.
Rac1 is a small canonical Rho guanosine triphosphatase (GT-
Pase) that regulates actin cytoskeletal dynamics (12). Integrin–
extracellular matrix interactions activate Rac1, and activated
[guanosine 5′-­triphosphate (GTP)–bound] Rac1 polymerizes and
organizes the actin cytoskeleton via various downstream effectors
to promote lamellipodia formation and cellular migration (13).
Moreover, in epithelia, Rac1 promotes polarized actin organiza-
tion and cell-­cell junction stability and maintains apicobasal po-
larity (14–16). We and others showed that Rac1 is required for
normal postdevelopment CD structure and function (17, 18) as it
maintains actin cytoskeletal branching at cell-­cell junctions that
restricts excessive actomyosin activity and maintains cell shape
and polarity (18). In addition, there is strong evidence that Rac1
function in epithelial cells goes beyond actin organization (19).
Rac1 stimulates cell cycle progression by activating traditional
proliferative cell survival pathways (extracellular signal–regulated
kinase, p38–mitogen-­activated protein kinase, Akt, and c-­Jun N-­
terminal kinase) via downstream effectors that overlap with its ac-
tin cytoskeletal function (e.g., p21-­activated protein kinase 1),
raising the possibility that the actin organization and cell cycle
progression functions of Rac1 are intimately linked (20–23).

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 1 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
In this study, we analyzed the role of Rac1 in CD repair following
irreversible prolonged and short reversible obstructive kidney injury.
We show that CDs lacking Rac1 are unable to reconstitute their mor-
phology or proliferate normally after relief of obstruction, as this
small GTPase plays a key role in integrating actin cytoskeletal organi-
zation and cell cycle control. Mechanistically, we show that Rac1 regu-
lates the orderly actin-­dependent progression through G2-­M that
ensures reliable cell division.
RESULTS
Rac1 is required to maintain CD integrity during
prolonged obstruction
To determine whether Rac1 is essential in maintaining collecting sys-
tem epithelial structural integrity in a model with morphological
damage of the CD, we induced prolonged obstruction (>7 days) in
mice, which leads to substantial parenchymal loss with apoptosis pre-
vailing over regeneration (3). We deleted Rac1 in the CDs at embry-
onic day 19 (E19) by crossing Rac1flox/flox (Rac1f/f) mice with AQP2
(Aquaporin 2)–Cre mice. Successful deletion of Rac1 in the collecting
system of AQP2:Rac1f/f mice was confirmed by immunofluorescence
of medullary CDs and immunoblotting of papillary lysates (fig. S1).
AQP2:Rac1f/f mice did not have any histological abnormalities or
functional differences in the kidney at baseline (Fig. 1A and fig. S2).
We performed prolonged unilateral ureteral obstruction (UUO)
and analyzed the kidneys after 10 days of UUO. As expected, this in-
jury resulted in medullary tubular dilatation with epithelial flattening
in Rac1f/f mice, which was more pronounced in the kidneys of
AQP2:Rac1f/f mice. In keeping with this finding, the AQP2:Rac1f/f
kidneys displayed significantly increased fibrosis (Sirius Red) and
medullary apoptosis with decreased proliferation (Ki-­67, all active cell
cycle phases) (Fig. 1, A and B) compared to Rac1f/f kidneys indicating
more severe obstruction-­induced epithelial damage. Given the role of
Rac1 in regulating actin cytoskeletal organization, we performed
thick frozen sectioning of fresh tissue, followed by optical clearing
and high-­resolution three-­dimensional (3D) confocal imaging of
the actin cytoskeleton of CDs demarcated by AQP2. There was no
difference in epithelial actin structure at baseline between genotypes
(Fig. 1C), and after UUO, Rac1f/f CDs demonstrated flattening but a
largely preserved pattern of basolateral filamentous actin (F-­actin). By
contrast the AQP2:Rac1f/f CDs had a near-­complete loss of F-­actin,
which was more obvious in the 3D reconstruction of the tubular actin
cytoskeleton (Fig. 1, C and D, and movie S1). There was no general
decrease in F-­actin intensity surrounding the CDs, suggesting that the
F-­actin defect in the AQP2:Rac1f/f mice was CD specific (fig. S3).
Together, these results show that Rac1 plays a critical role in protect-
ing the kidney collecting system from injury following prolonged
obstruction.
Rac1 promotes CD repair, morphological reconstitution, and
F-­actin recovery
We next defined whether Rac1 played a role in CD repair following
injury using a well-­established model of reversible UUO (24–26). In
this model, microvascular clamps are used instead of sutures for ure-
teral ligation allowing for successful decompression and injury recov-
ery (Fig. 2A and fig. S4). Kidneys were reversibly obstructed for five
days, which is a time point after which damage is still recoverable (25).
There were no differences in tubular dilatation or fibrosis (assessed by
picrosirius red) at 5 days of injury; however, the tubular dilatation and
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
fibrosis did not regress in the reversed AQP2:Rac1f/f mice, suggesting
that they were unable to undergo tubular repair (Fig. 2, B to E). To
assess whether the CD cells could regain their epithelial integrity
following injury, we labeled AQP2+ CDs with the basolateral adher-
ens junction marker E-­cadherin. There was a loss of the typical baso-
lateral E-­cadherin pattern in both groups following obstruction,
which was reestablished in repairing Rac1f/f CDs; however, there
was persistently decreased E-­cadherin height and lateral density in
the AQP2:Rac1f/f CDs (Fig. 2, F and G). Medullary tubular dilatation,
inflammatory infiltrates, destruction of renal parenchymal architec-
ture, and fibrosis were still present 1 month after reversal of obstruc-
tion in the AQP2:Rac1f/f mice, while the Rac1f/f kidneys appeared to
be morphologically normal (fig. S5).
To support the structure of a cell, the actin cytoskeleton forms
F-­actin out of globular actin monomers (27). To examine the
actin cytoskeleton in greater detail, we performed optical clear-
ing of fresh thick inner medullary slices and performed 3D high-­
resolution confocal imaging of F-­actin. Both Rac1f/f and AQP2:Rac1f/f
CDs showed dysmorphic F-­actin and diminished apical F-­actin
density following UUO. This was reversible in the Rac1f/f CDs;
however, it persisted in AQP2:Rac1f/f CDs (fig. S6). 3D recon-
struction of the phalloidin-­labeled actin cytoskeleton, differen-
tial segmentation, and masking of the apical versus basolateral
F-­actin cytoskeleton revealed decreased density and increased
elongation of the tubular apical F-­actin meshwork in the flat-
tened AQP2:Rac1f/f CD epithelium during repair, whereas the F-­
actin structure returned to normal in control CDs (Fig. 2, H to J,
and movie S2). To verify that Rac1 plays a role for reconstitution
of the actin cytoskeleton after injury, we subjected a monolayer
of silicon-­rhodamine (SiR)-­actin–labeled isolated Rac1f/f and
Rac1−/−CD cells [generated as described in (18)] to two-­photon
laser ablation. Apical F-­actin formation in cells surrounding the
site of injury was clearly present in Rac1f/f but not Rac1−/− CD
cells (fig. S7 and movie S3). In summary, Rac1 is required to re-
constitute epithelial F-­actin architecture during the repair pro-
cess both in vivo and in vitro.
Rac1 is required for CD proliferation and maintains CD cells
in G2-­M during repair
In addition to morphological reconstitution, epithelial CD repair re-
quires proliferation (3). We therefore investigated whether this aspect
of CD repair in AQP2:Rac1f/f kidneys was abnormal. We observed
decreased CD cellularity in the dysmorphic AQP2:Rac1f/f kidneys,
when compared to the Rac1f/f kidneys, suggesting a proliferation de-
fect (fig. S8). Thus, we labeled full medullary slices with AQP2 and
Ki-­67 (which marks all active cell cycle phases) and performed 3D
high-­resolution multiplex imaging of optically cleared tissue slices to
determine Ki-­67–positive cells inside CDs (AQP2+). We performed
surface reconstruction, masking, and spot-­to-­surface thresholding
algorithms (28) to highlight proliferating CD cells (white dots). We
found that ureteral obstruction induces a proliferative response in
both groups, but AQP2:Rac1f/f CDs were unable to maintain the pro-
liferative state and showed markedly reduced Ki-­67 inside AQP2+
tubules during repair in comparison to controls (Fig. 3, A and B, and
fig. S9). In addition to the decreased proliferation, there was also in-
creased apoptosis in repairing AQP2:Rac1f/f CDs compared to con-
trols (Fig. 3, C and D). Thus, the repair phase of AQP2:Rac1f/f CDs is
characterized by both cell death and hypoproliferation compared to
Rac1f/f CDs.
2 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Rac1 is required to maintain epithelial and F-­actin integrity during prolonged obstruction. (A) Hematoxylin and eosin (H&E)–stained paraffin kidney sections
in the first and second rows [scale bars, 100 μm (top row) and 50 μm (insets); area for insets indicated by a black dashed box), Sirius Red staining in the third row (scale
bars, 50 μm), apoptosis [terminal deoxynucleotidyl transferase–mediated deoxyuridine triphosphate nick end labeling (TUNEL)] staining in the fourth row (nuclei are
magenta, and positive nuclei are green; scale bars, 50 μm), and Ki-­67 in the fifth row [3,3′-­Diaminobenzidine (DAB), brown; scale bars, 50 μm] of medullary regions of
control (Rac1f/f ) or AQP2:Rac1f/f kidneys at baseline or 10 days after obstruction (10-­day UUO, ureteral ligation with sutures). (B) Quantification of fibrosis [as percentage
of Sirius Red–positive area per high-­power field (HPF)], TUNEL (positive cells per HPF), and Ki-­67 (positive cells per HPF) from (A). bl, baseline. (C) Thick fresh frozen medul-
lary kidney slices stained for AQP2 (CDs) and F-­actin of baseline and obstructed (10-­day UUO) control and AQP2:Rac1f/f kidneys. First row depicts a cross section (scale bars,
10 μm), second row shows the F-­actin channel with the colors inverted (scale bars, 10 μm), and the third row shows 3D F-­actin reconstructions (scale bars, 5 μm). Red
arrows in the second row highlight basolateral F-­actin that is deficient in the injured mutant CDs. (D) Quantification of total tubular F-­actin content as outlined in the
methods. AU, arbitrary units. n = 5 mice per group with each dot in the scatter plots representing an individual sample. Bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 3 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Rac1 promotes CD repair, morphological reconstitution, and F-­actin recovery. (A) Schematic representation of the reversible UUO model outlining surgical clamp
placement and removal for collecting system decompression. (B) H&E paraffin kidney sections of Rac1f/f and AQP2:Rac1f/f mice at baseline, after 5 days of UUO and 1 week
after reversal of obstruction (“reversed”) in the top row (scale bars, 50 μm) with a black dashed box indicating insets shown below (scale bars, 20 μm). (C) Quantification of
medullary tubular diameter in (B) in normalized pixels (normalized to frame size that was equal between groups). (D) Sirius Red stained kidney sections (scale bars, 100 μm).
(E) Quantification of fibrosis in (D) as percent Sirius Red positive per HPF. (F) E-­cadherin immunostaining was performed on paraffin kidney sections of baseline, obstructed,
and reversed Rac1f/f and AQP2:Rac1f/f mice, and the CDs were marked by AQP2 (scale bars, 10 μm); sections were analyzed by confocal microscopy. (G) Quantification of
normalized lateral E-­cadherin intensity (left) and lateral E-­cadherin height in micrometers (right) from (F) using ImageJ. (H) 3D reconstructions of AQP2-­positive F-­actin–
labeled CDs with segmentation and masking of the basolateral (yellow) and apical (white) F-­actin using Imaris. Top rows display a top view of both masking channels, and
bottom rows show side views of apical F-­actin. n = 6 mice per group. Scale bars, 5 μm. (I) Quantification of the apical F-­actin major axis length in micrometers (details in Ma-
terials and Methods) with each dot in the scatter plots representing measurements. (J) Schematic depiction of the results showing abnormal longitudinal extension of the
apical F-­actin meshwork of repairing AQP2:Rac1f/f CDs. Bars are means ± SD. ns, not significant. *P < 0.05. Panels (A) and (J) were created with Biorender.com.

Rac1 is required for the G1-­S phase transition in proliferating can-
cer cells (20, 29), so we investigated whether there was decreased S
phase entry and DNA synthesis in Rac1-­null CDs. We pulsed mice
during repair with intraperitoneal 5-­bromo-­2′-­deoxyuridine (BrdU)
and examined the tissue histologically during the repair process.
Unexpectedly, there was no difference in the BrdU–to–Ki-­67 ratio
between groups during repair (fig. S10). Since G2-­M arrest has been
implicated in failed tubular repair in the proximal tubule, we predicted
that repairing AQP2:Rac1f/f CD cells would enrich in G2-­M. We
therefore performed a DNA content–based cell cycle assessment
in vivo in AQP2-­labeled CD cells (30). There were less repairing
AQP2:Rac1f/f CD cells in G2-­M compared to controls, suggesting that

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 4 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Rac1 promotes proliferation and normal mitotic progression during repair. (A) Optically cleared full medullary slices of baseline, obstructed (“UUO”), and
unobstructed reversed Rac1f/f and AQP2:Rac1f/f kidneys. Ki-­67 signal inside AQP2+ CDs was filtered and converted to white dots. Scale bars, 100 μm. (B) Quantification of
(A) showing Ki-­67–positive cells per 100-­μm-­long CD segment with each dot representing individual mice. n = 4 mice per group. (C) Apoptosis (TUNEL; white) labeling of
CDs (AQP2+; magenta) of paraffin kidney sections in reversed Rac1f/f and AQP2:Rac1f/f mice. (D) Quantification of TUNEL-­positive cells in CDs per medullary HPF with each
dot representing individual samples. n = 4 mice per group. (E) Representative flow cytometry plots of AQP2+ CD cells of reversed Rac1f/f and AQP2:Rac1f/f kidneys showing
side scatter (SSC-­A) against DNA [4′,6-­diamidino-­2-­phenylindole (DAPI)] with a G2-­M phase–specific cell cycle gate and corresponding subpopulation percentage shown
in the plot. (F) Quantification of G2-­M phase–specific cell populations as shown in the gating in (E) with four mice (dots) per group; bars are means ± SD. (G) Mitotic (pH3-­
positive; green) metaphase cells (condensed, aligned DNA) of repairing (reversed) CDs (AQP2; magenta) of Rac1f/f and AQP2:Rac1f/f mice were analyzed using 3D super-­
resolution confocal imaging. The far left column shows orthogonal Z-­slices (scale bars, 2.5 μm; yellow continuous line outlines the apical lumen) as indicated by the yellow
dashed line in the cross section in the middle column (scale bars, 5 μm). The far right column depicts insets as outlined by a red continuous box (scale bars, 2.5 μm).
(H) Relative distribution of metaphase abnormalities based on a morphological assessment as shown in (G). At least 10 mitoses were analyzed per group. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 5 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Rac1 was not required to prevent a G2-­M arrest but instead to main- results in aneuploidy (34–36). The aneuploidy rates were low and not
tain CD cells in G2-­M (Fig. 3, E and F). These data suggested that increased in repairing AQP2:Rac1f/f CD cells in vivo compared to
there was likely a defect in cell division [M phase (mitosis)] during Rac1f/f CDs (fig. S15). We next defined whether AQP2:Rac1f/f CD
G2-­M, which is a known actin cytoskeleton–dependent process. We cells were unable to undergo mitosis due to problems with metaphase
therefore evaluated cells undergoing mitosis [marked by phosphohis- cell rounding, which is an essential F-­actin–dependent process that
tone H3 (pH3)] and found that there was a defect in pH3/AQP2 creates the space and geometry for spindle formation and chromo-
double-­positive G2-­M phase cells in repairing AQP2:Rac1f/f CDs some separation in epithelial cells and is closely linked to normal cell
suggesting a problem with mitosis (fig. S11). cycle progression through G2-­M (37–40). To do this, we performed
deep 3D super-­resolution confocal imaging of optically cleared med-
Rac1 promotes normal mitotic progression ullary slices after reversal of ureteral obstruction, followed by 3D re-
To understand the mechanism for the M phase defect in repairing construction of the mitotic metaphase F-­actin surface. Rac1f/f CDs
AQP2:Rac1f/f CD cells, we performed high-­resolution 3D confocal rounded up their metaphase F-­actin surface and assumed a near-­
imaging of the mitotic substages in vivo during repair. During meta- spherical shape within the otherwise cuboidal epithelium, whereas
phase, Rac1f/f CD cells enlarged, their condensed chromosomes AQP2:Rac1f/f CDs displayed decreased circularity and increased flat-
(marked by pH3) aligned, and the dividing cells stayed within the tening (Figure 5, A to C). Next, we analyzed mitotic F-­actin rounding
plane of the epithelium (Fig. 3G). By contrast, in AQP2:Rac1f/f CD in vitro by performing 3D super-­resolution imaging of the F-­actin
cells, the condensed chromosomes appeared dysmorphic, and there cytoskeleton of mitotic metaphase CD cells in postwound closure
was an increased rate of mitotic extrusion and more luminal mitotic epithelial layers. The rounded metaphase morphology of Rac1−/− CD
cells (Fig. 3, G and H, and fig. S12). These data suggest that Rac1 is cells was severely altered with an increased axial ratio and decreased
required for normal metaphase morphology. circularity compared to controls (Fig. 5, D to F). A similar rounding
Mitosis is the shortest cell cycle phase and occurs on a scale of defect was observed when cells proliferated after mechanical stretch
minutes, which makes it difficult to study (31). We therefore (fig. S16). To exclude that the rounding defect was an artifact of our
performed 3D confocal live imaging of mitosis in proliferating 2D culture system, we grew CD cells as 3D spheroids in Matrigel, after
postscratch Rac1f/f and Rac1−/− CD cell monolayers grown on which we performed 3D high-­resolution imaging of mitotic F-­actin
Matrigel-­coated glass bottom dishes using SPY probes. In this model morphology. This also showed misplaced and dysmorphic mitotic
of stretch-­induced mitosis, cells initially stretch over the wound area metaphase F-­actin in Rac1−/− CD cells (fig. S17). To test whether the
and then undergo waves of cell division in the scratch-­adjacent areas F-­actin–dependent rounding defect was sufficient to cause mitotic in-
(32). When the chromosomes condensed in the interphase to pro- stability in metaphase, we treated postwound closure CD cell layers
metaphase, transition control cells rounded up, pushed against their with low-­ dose nocodazole, which activates the spindle assembly
neighbors, and successfully underwent chromosome separation (ana- checkpoint and arrests and maintains mitotic cells in the rounded
phase and telophase) and cytokinesis within the plane of the regener- metaphase stage without affecting F-­actin (39, 41, 42). We monitored
ating epithelial cell layer (Fig. 4, A and B). By contrast, Rac1−/− CD the mitotic state over time by super-­resolution confocal imaging of
cells briefly rounded up in prometaphase but then failed to maintain F-­actin and by immunoblotting for pH3. Compared to metaphase-­
their morphology resulting in mitotic delay or mechanical mitotic cell arrested controls, which remained stably rounded over time, the
extrusion (Fig. 4, A and B, and movie S4). As expected, epithelial F-­ Rac1−/− CD cells demonstrated a disrupted metaphase F-­actin struc-
actin organization was disrupted at the migrating wound edge mar- ture and increased cortical blebbing and mitotic catastrophes (mitotic
gins of Rac1−/− CD cell layers, consistent with the role of this small cell death) (Fig. 5, G and H). Furthermore, immunoblotting for the
GTPase in promoting F-­actin organization in repairing CD cells mitotic marker pH3 revealed that metaphase-­arrested Rac1−/− CD
(fig. S13, A and B). There were also less mitotic cells in the scratch-­ cells were unable to maintain the mitotic state as indicated by a de-
adjacent areas of repairing Rac1−/− CD cell layers (fig. S13, C and D). crease in pH3 that did not occur in controls (fig. S18). Thus, Rac1 is
To further verify this observation in another model where cells are required for epithelial mitotic metaphase stability by promoting F-­
stretched to mimic ureteral obstruction (33), we grew cells on coated actin–dependent mitotic rounding.
silicone membranes and exposed them to noncyclical uniaxial stretch
(10%, 3.5 hours) in a Flexcell bioreactor (Fig. 4C). Similar to the Rac1 promotes mitotic rounding by regulating actomyosin
scratch assay, poststretch Rac1−/− CD cell layers demonstrated in- Epithelial actin cytoskeletal networks generate force and contractility
creased F-­actin disruption (Fig. 4D), fewer cells in G2-­M (Fig. 4, E and by incorporating the motor protein myosin that determines the
F), and less mitosis (Fig. 4, G and H). These data suggest that during mechanical properties of an epithelium (43). We previously showed
repair from mechanical injury Rac1 not only maintains F-­actin orga- that Rac1-­depedent actin cytoskeletal organization is required to re-
nization but also promotes normal cell cycling through G2-­M by strict excessive actomyosin activity in CD cells by maintaining actin
maintaining mitotic progression. branching (18). We therefore determined whether the abnormality in
mitotic morphology was a mechanical defect due to actomyosin dys-
Rac1 regulates F-­actin–dependent mitotic rounding regulation. We confirmed that there is increased activated actomyosin
during repair in repairing AQP2:Rac1f/f CDs compared to controls by immunos-
To investigate the underlying cause for the mitotic defects, we initially taining for phospho–myosin light chain (pMLC) (Fig. 6, A and B). We
imaged the microtubular spindle apparatus in the different mitotic next investigated whether actomyosin was dysregulated around dys-
substages but found no gross differences between Rac1f/f and Rac1−/− morphic mitotic metaphase cells in regenerating Rac1−/− CD mono-
CD cells (fig. S14). We next investigated whether Rac1 deficiency af- layers in vitro. While Rac1f/f CD cells showed some enrichment of
fected the process of cytokinesis, which is the physical separation of pMLC in the rounded mitotic actin cortex, this was not observed
daughter cells at the end of mitosis. Failure of cytokinesis typically in mitotic Rac1−/− CD cells. Instead, strong, and diffuse actomyosin

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 6 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Rac1 promotes mitotic progression, F-­actin integrity, and a G2-­M cell cycle shift after CD cell injury in vitro. (A) Live confocal mitosis imaging of SPY650-­DNA
(blue)– and SPY555-­actin (white)–labeled CD cells in vitro. Orthogonal Z-­slices are also shown. Representative mitotic defects are shown for Rac1−/− CD cells. Scale bars,
10 μm. The mitotic cell was manually recolored in green. (B) Relative distribution of mitotic defects in vitro during live imaging cell division sequences with at least
10 mitoses analyzed per group. (C) Schematic presentation of the Flexcell bioreactor. Created with Biorender.com (D) Super-­resolution confocal images of apical cross
sections of phalloidin-­647 (F-­actin)–labeled Rac1f/f and Rac1−/− CD cell monolayers before stretch (baseline) and after stretch (“stretched”). Images are shown as color-­
inverted grayscale images and are representative of three repeat experiments (scale bars, 20 μm). The red dashed box indicates the region of the inset shown in the
rightmost column (scale bars, 5 μm). (E) Representative flow cytometry cell cycle histograms of propidium iodide (“DNA”)–labeled CD cells in vitro before stretch (baseline;
blue) and after stretch (stretched, green). Plots are normalized to mode (% max), and the indicated range highlights the G2-­M population. (F) Quantification of the relative
G2-­M population with bars as means ± SD and each dot representing individual samples from three independent repeat experiments. (G) Baseline and poststretch
(stretched) Rac1f/f and Rac1−/− CD cell monolayers were stained for mitotic cells (pH3; green) and nuclei (DNA; blue). Images are representative of three experiments. Scale
bars, 20 μm. (H) Quantification of mitotic cells (pH3-­positive) per HPF of Rac1f/f and Rac1−/− CD cells before stretch (baseline; blue) and after stretch (stretched; green).
Individual dots represent repeat experiments, and bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 7 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. Rac1 regulates normal mitotic metaphase rounding. (A) Cross sections of F-­actin–labeled (white) thick frozen sections. CDs (AQP2; magenta) and mitotic meta-
phase cells (pH3; green) of reversed/repairing mice (scale bars, 5 μm). Orthogonal Z-­slices of the mitotic cells are shown next to the cross section as indicated by the yellow
(“z”). The red dashed box indicates color-­inverted mitotic metaphase F-­actin outlined by a continuous cyan line (scale bars, 3 μm). The rightmost panel shows 3D recon-
structions of the metaphase F-­actin (scale bars, 3 μm). Images are representative of at least 10 mitotic figures per group pooled from at least three mice per group. (B and
C) Quantification of metaphase circularity (4π × area/perimeter2) and height (in micrometers) showing a minimum of 10 measurements per group. Bars are means ± SD.
(D) In vitro confocal imaging of F-­actin (white)– and DNA (blue)–labeled mitotic metaphase CD cells (scale bars, 5 μm). Orthogonal Z-­views are shown next to the cross
section with the Z-­level indicated by a yellow dashed line. 3D metaphase F-­actin surface reconstructions are shown (scale bars, 3 μm). (E and F) Quantification of meta-
phase circularity and height (in micrometers) of control and Rac1−/− CD cells showing a minimum of 10 measurements per group over at least three experiments. Bars are
means ± SD. (G) F-­actin (white)– and DNA (cyan)–labeled metaphase-­arrested [using nocodazole (100 ng/ml), condensed chromosomes] Rac1f/f and Rac1−/− CD cells
analyzed by confocal microscopy over time as indicated. Metaphase defects of Rac1−/− CD cells are shown including F-­actin disorganization, surface blebbing, and nucle-
ar fragmentation with micronuclei formation indicative of mitotic catastrophes (scale bars, 5 μm). Images are representative of at least two experiments. (H) Quantification
of mitotic catastrophes per HPF of nocodazole-­arrested control and Rac1−/− CD cells showing a minimum of 12 measurements per group. Bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 8 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. Rac1 promotes mitotic rounding by regulating actomyosin. (A) Representative images of activated actomyosin (pMLC) in reversed mice. CDs are labeled with DBA
(dolichus biflorus agglutinin; green). pMLC is shown with a fire color scheme applied. (B) Quantification of average pMLC intensity with dots representing four mice per group.
(C) pMLC (white)– and DNA (cyan)–labeled CD cell monolayers with a metaphase shown in the center. A pMLC fire color conversion is shown. Three repeat experiments. Scale
bars, 20 μm (A and C). (D) pMLC intensity in the perimitotic area with dots representing 10 individual measurements. (E) F-­actin (white)– and DNA (cyan)–labeled CD cells
grown on a Matrigel-­coated polydimethylsiloxane (PDMS) substrate with fluorescent nanobeads. Mitotic cell colored in green. Traction force maps are shown (in pascals).
Scale bars, 20 μm. (F) Quantification of perimitotic forces. (G) F-­actin (white)– and DNA (cyan)–labeled mitotic metaphase CD cells. Bleb, blebbistatin (5 μM). Scale bars, 10 μm.
Red box outlines Metaphase F-­actin shape shown on the right (scale bar, 5 μm). (H) Circularity quantification of mitotic metaphase F-­actin as shown in (G). A minimum of
10 measurements are shown per group. (I) Live confocal mitosis imaging of CD cells in vitro. Bleb, 5 μM. Scale bars, 10 μm. The mitotic cell was colored in green. (J) Relative
distribution of mitotic defects in vitro. At least 10 mitoses per group. In vitro experiments are representative of n = 3. (K) In vivo injection protocol of blebbistatin [5 days, 2 mg/kg
per day, intraperitoneally (i.p.)]. Created with Biorender.com. (L) Super-­resolution confocal imaging of mitotic CD cells during repair (scale bars, 10 μm). The yellow box outlines
the insets that are shown in the top row, which depict color-­inverted metaphase F-­actin (scale bars, 5 μm). (M) Quantification of mitotic metaphase circularity from (L) showing
n = 4 mice per group. Bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 9 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
activation was seen in the Rac1−/− CD cells that surround mitotic cells
(Fig. 6, C and D). To determine whether the dysregulation of acto-
myosin translates into increased mechanical tension around mitotic
cells, we performed traction force microscopy and found that Rac1−/−
CD cell layers displayed increased mechanical force in the area adja-
cent to mitotic prometaphase cells when compared to Rac1f/f CD
cells, suggesting that actomyosin dysregulation leads to a force imbal-
ance and increased tension around mitotic cells (Fig. 6, E and F).
These data suggest that Rac1 restricts excessive actomyosin force
generation, which is likely required to maintain normal mitotic mor-
phology. Direct myosin inhibition has been shown to decrease the
mechanical force generation by epithelial cells (44). Hence, we hy-
pothesized that direct myosin inhibition would reverse the Rac1−/−
phenotype of failed rounding and cell division. We therefore inhibited
the mechanical force on mitotic cells by treating Rac1−/− CD cell
layers in vitro with low doses of the direct myosin inhibitor blebbi-
statin. This treatment reversed the F-­actin mitotic rounding defect in
Rac1−/− CD cells, suggesting that Rac1 promotes normal mitotic F-­
actin rounding by modulating actomyosin (Fig. 6, G and H). Simi-
larly, blebbistatin reversed the F-­actin rounding defect in Rac1−/− CD
spheroids in a 3D matrix (fig. S19). We then assessed whether excess
actomyosin activity inhibition was sufficient to rescue the defect in
mitotic progression of dividing Rac1−/− CD cells by treating mono-
layers with low-­dose blebbistatin and performing live confocal mito-
sis imaging. Blebbistatin reversed the mitotic extrusion and metaphase
delay defects in dividing Rac1−/− CD cells (Fig. 6, I and J, and movie
S5). To assess whether this finding could be recapitulated in vivo, we
administered blebbistatin to mice upon clip removal and reversal of
obstruction (Fig. 6K). Five days after daily blebbistatin injection, we
performed in vivo super-­resolution confocal microscopy of the meta-
phase actin cytoskeleton of fresh frozen medullary slices and found
that myosin-­inhibited mitotic AQP2:Rac1f/f CD cells have a markedly
improved morphology with increased size and circularity, whereas
vehicle-­treated mitotic mutant CD cells remained flat and dysmor-
phic with decreased rounding (Fig. 6, L and M). Together, these
data indicate that Rac1 promotes actin-­dependent mitotic metaphase
rounding by controlling actomyosin, which is required to maintain
mitotic stability.
Direct actomyosin inhibition is sufficient to reverse the
repair defect of AQP2:Rac1f/f kidneys
Since blebbistatin rescued the mitotic rounding defects, we deter-
mined its effects on CD cell proliferation and repair following ob-
struction. To do this, we labeled full medullary slices collected after
the obstruction was reversed with AQP2 and Ki-­67. We then per-
formed confocal high-­resolution multiplex imaging and postimaging
surface reconstructions applying spot-­to-­surface thresholding algo-
rithms. We found that blebbistatin treatment rescued the proliferation
defect and decreased apoptosis of repairing AQP2:Rac1f/f kidneys
with similar proliferation rates to untreated controls (Fig. 7 A and B,
and figs. S20 and S21). We confirmed that blebbistatin does not in-
duce CD cell proliferation in uninjured control kidneys at baseline
(fig. S22). As blebbistatin administration during obstructive injury
has been shown to affect macrophage motility (45), we counted
immunolabeled total macrophages (CD68+) around untreated and
blebbistatin-­treated repairing AQP2:Rac1f/f CDs. There was no major
difference, suggesting that modulation of immune cell infiltration
is not a confounding mechanism whereby blebbistatin rescues the re-
pair defect in the absence of Rac1 (fig. S23). When we performed
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
DNA content–based cell cycle assessment by flow cytometry of dis-
sociated kidneys, we noted that actomyosin inhibition also rescued
the G2-­M cell cycle defect of repairing AQP2:Rac1f/f CDs (Fig. 7, C
and D). We further showed that blebbistatin ameliorated medullary
tubular dilatation and fibrosis to similar levels as untreated controls
(Fig. 7, E to G). Last, we performed thick frozen sectioning of fresh
tissue, followed by high-­resolution 3D confocal imaging and recon-
struction of the actin cytoskeleton to demonstrate full restoration of
apical F-­actin density and morphology of the apical actin cytoskele-
ton of actomyosin-­inhibited AQP2:Rac1f/f CDs when compared to
the vehicle-­treated AQP2:Rac1f/f CDs (Fig. 7, H to J, and movie S6).
These data indicate that Rac1-­dependent modulation of actomyosin is
required to promote normal proliferation and cell cycling through
G2-­M during repair, which are essential for epithelial CD repair.
Rac1 restricts myosin activation by inhibiting the RhoA
activator guanine nucleotide exchange factor–H1
Next, we sought to determine the mechanism whereby Rac1 sup-
pressed actomyosin activity. Classically, Rac1 regulates myosin activ-
ity via its mutual antagonism with the other canonical Rho-­GTPase
RhoA (46, 47). Although we previously showed that RhoA activity is
not changed in asynchronous (non–cell cycle–synchronized) Rac1−/−
CD cells (18), it is possible that interactions between Rac1 and RhoA
occur in a cell cycle phase–specific manner in kidney CD cells. To
investigate this, we first synchronized cells at the beginning of S phase
using a double thymidine block (enriched for cyclin E) or at the be-
ginning of G2 using the cyclin-­dependent kinase 1 (Cdk1) inhibitor
RO-­3306 (enriched for cyclin B1) (Fig. 8A) (48–50). We then mea-
sured RhoA activity using G-­LISA and found that Rac1−/− CD cells
have abnormally increased RhoA activity specifically at the beginning
of G2 (Fig. 8B). The antagonistic cross-­talk between Rac1 and RhoA is
mainly mediated by upstream regulatory guanine nucleotide ex-
change factors (GEFs) (47). The microtubule-­associated GEF-­H1
(also known as Arhgef2) is a RhoA-­activating GEF that mediates the
cell cycle phase–specific cross-­talk between microtubules and the
actin cytoskeleton, and its activity is regulated through a cycle of mi-
crotubule binding and release (51–55). This cycle is controlled by
Rac1-­dependent serine-­886 phosphorylation of GEF-­H1 via its effec-
tor p21-­Activated kinase 1 (Pak1) that promotes the microtubular
translocation and inhibition of GEF-­H1 (56). GEF-­H1 was recently
shown to be released upon interphase microtubular disassembly pro-
moting mitotic entry and controlling mitotic cell shape (57). On this
basis, we hypothesized that GEF-­H1 controls the antagonistic cell cy-
cle phase–specific Rac1-­RhoA interaction. Super-­resolution imaging
of microtubules in uniaxially stretched G2-­shifted CD cells revealed
increased free (not associated with microtubules) GEF-­H1 in Rac1−/−
CD cells indicating increased GEF-­H1 activity toward RhoA (Fig. 8, C
and D). Consistent with this, the inhibitory S886 phosphorylation of
GEF-­H1 was significantly decreased in Rac1−/− CD cells (Fig. 8, E and
F). GEF-­H1 knockdown in Rac1−/− CD cells abolished the abnormal-
ly elevated RhoA activity and myosin phosphorylation in G2 (Fig. 8, G
to J). Last, GEF-­H1 knockdown in Rac1−/− CD cells was sufficient to
rescue the cell stretch–triggered G2-­M phase cell cycle shift and re-
store mitotic rounding (Fig. 8, K to M). These data suggest that Rac1
suppresses abnormal actomyosin activation in the G2 cell cycle phase
through RhoA antagonism by promoting the phosphorylation, inhi-
bition, and microtubular association of the RhoA-­activating GEF-­H1.
As GEF-­H1 activates RhoA that regulates actomyosin via Rho-­
associated coiled-­coil containing protein kinases (ROCK1/2) (58), we
10 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Actomyosin inhibition reversed the repair defect of AQP2:Rac1−/− mice. (A) Optically cleared medullary slices of reversed (repairing) mice labeled with AQP2
(CDs) and Ki-­67. 3D reconstructions are shown. Ki-­67 signal inside AQP2+ CDs was filtered and converted to white dots. Scale bars, 100 μm. (B) Quantification of (A) show-
ing Ki-­67–positive cells per 100-­μm-­long CD segment with dots representing individual mice (n = 5). (C) Flow cytometry plots of AQP2+ CD cells of reversed kidneys
showing side scatter (SSC-­A) against DNA (DAPI) with a G2-­M phase–specific cell cycle gate. Subpopulation percentages are shown in the plot. (D) Quantification of G2-­M
CD cells with at least three mice (one dot) per group. (E) H&E paraffin kidney sections in the first row (scale bars, 20 μm) of medullary regions and Sirius Red staining in the
second row (scale bars, 50 μm). n = 5 mice per group. (F and G) Quantification of medullary tubular diameter as normalized pixels and fibrosis (as percentage of Sirius
Red–positive area per HPF) with dots representing individual mice. (H) F-­actin labeling (white) of CDs (AQP2; magenta) in thick fresh frozen medullary kidney slices in re-
versed kidneys analyzed by 3D super-­resolution confocal. First row depicts two cross sections (left original, right F-­actin channel; scale bars, 10 μm), and second row shows
oblique views of 3D reconstructions of the apical CD F-­actin (scale bars, 7 μm). (I and J) Quantification of apical F-­actin intensity and apical F-­actin major axis length; dots
represent individual mice. n = 5 mice per group. Bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 11 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 8. Rac1 restricts myosin activation by inhibiting the RhoA activator GEF-­H1. (A) Immunoblots of asynchronous (Asyn), G1-­S–synchronized, or S phase–synchronized
CD cells to confirm cell cycle phase–specific enrichment (cyclin E, S phase; cyclin B1, G2 phase). OD490, optical density at 490 nm. (B) Quantification of RhoA activity
(asynchronous, S, and G2) using G-­LISA and measuring absorbance at 490 nm. Individual dots represent repeat experiments (n = 3). (C) Super-­resolution confocal imaging of
microtubules (magenta) and GEF-­H1 (green) of CD cells after stretch. Scale bars, 10 μm (first row) and 2.5 μm (second row). The white dashed box indicates the region of
interest for the insets in the second row. (D) Line scan profile (white dashed line) normalized to the highest (100%) and lowest (0%) intensity. Representative of three
repeat experiments and 10 measurements per group. (E and F) Immunoblotting of G2-­synchronized cell lysates for phosphorylated (serine-­886, S886) and total GEF-­H1
in biological duplicates. Three repeat experiments were quantified in (F). Fold change values ± SD. (G) Immunoblotting of empty vector (EV)–transfected or GEF-­H1
knockdown short hairpin RNA (shRNA)–transfected CD cells. (H) Quantification of RhoA activity in G2 cell cycle fractions. Individual dots represent repeat experiments
(n = 3). (I and J) Immunoblotting of G2 cell cycle fractions of EV or GEF-­H1 knockdown transfected for total and pMLC (serine-­19). The phospho-­to-­total myosin light
chain (pMLC–to–t-­MLC) ratio is shown as fold change ± SD with individual dots representing repeat experiments (n = 3). (K) Quantification of G2-­M cell cycle fractions
(in percentage) of the indicated groups after induction of proliferation using stretch by flow cytometry using propidium iodide. Individual dots represent experiments
(n = 3). (L and M) F-­actin (white)– and DNA (blue)–labeled mitotic metaphase EV-­or GEF-­H1 knockdown–transfected Rac1−/− CD cells (scale bars, 5 μm). Mitotic rounding
(circularity) is quantified in (M) showing a minimum of 10 measurements per group. Bars are means ± SD. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 12 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
tested whether pharmacological inhibition of ROCK1/2 would rescue
the mitotic rounding and epithelial repair defect. Similar to direct
myosin inhibition with blebbistatin, ROCK inhibition with Y-­27632
rescued the defects in mitotic morphology and promoted mitotic
rounding in Rac1−/− CD cells (fig. S24). Next, we injected AQP2:Rac1f/f
mice with Y-­27632 upon reversal of ureteral obstruction [10 mg/kg,
intraperitoneally (i.p.), per day for 5 days]. Inhibition of ROCK in vivo
in repairing AQP2:Rac1f/f mice restored CD mitotic rounding, prolif-
eration, and tissue and actin cytoskeleton morphology, mimicking
blebbistatin treatment (figs. S24 and S25). These data support a mech-
anistic model whereby Rac1 promotes repair via suppression of the
GEF-­H1-­RhoA-­ROCK-­actomyosin signaling axis that is required to
maintain myosin-­dependent epithelial morphology and cell cycling.
Rac1-­dependent actomyosin regulates a mechanical G2-­M
checkpoint to prevent premature mitotic entry
Our findings suggested that Rac1-­dependent actomyosin regulation
via RhoA antagonism is required to promote the orderly cell cycle
progression through G2-­M, specifically mitosis. Unexpectedly, how-
ever, we found that dysmorphic Rac1-­deficient cells entered mitosis
and did not go into cell cycle arrest, suggesting an abnormal G2-­M
checkpoint override in cells that were clearly malformed. We there-
fore investigated whether failed cell cycle checkpoint activation con-
tributed to our Rac1-­deficient phenotype of mitotic dysmorphology,
rounding failure, extrusion, and impaired proliferation. We per-
formed cell cycle synchronization at the G1-­S border of control and
Rac1−/− CD cells in vitro using a double thymidine block and immu-
noblotted for cyclin B1 as a positive control as it rises steadily after
thymidine washout, peaks in G2, and then drops sharply in mitosis
(59). We found that while the cyclin B1 peak is preserved in Rac1−/−
CDs, it is notably shifted toward earlier time points (fig. S26). To test
whether this is associated with earlier mitosis, we immunoblotted for
the mitotic marker pH3 that confirmed abnormally early mitosis in
Rac1−/− CD cells (fig. S26). Next, to specifically study mitotic entry,
we synchronized cells in G2 using the reversible Cdk1 inhibitor RO-­
3306 (49, 60). Rac1f/f CD cells entered mitosis within 60 min of RO-­
3306 washout, whereas Rac1−/− CD cells rapidly entered mitosis
within ~15 min indicating a shortened G2 phase and premature mi-
totic entry (Fig. 9, A and B). G2-­M checkpoint override has recently
been linked to mitotic cell death in cancer cells (61–63). We thus in-
vestigated whether the premature mitotic entry is related to increased
mitotic instability and cell death. Rac1−/− CD cells demonstrated in-
creased cleaved caspase 3 activation upon release into G2 and mitotic
entry indicating increased mitosis-­associated cell death (Fig. 9, C and
D). The orderly progression through G2 and mitotic entry is con-
trolled by the G2-­M checkpoint kinase Wee1, which phosphorylates
and inhibits Cdk1 on tyrosine-­15 (Y15) to suppress early mitotic
events of cells unprepared for mitosis (64, 65). Wee1 inhibition leads
to loss of G2-­M checkpoint inhibition and early forced mitotic entry,
reminiscent of the Rac1−/− phenotype (61–63). To test whether Wee1
is regulated by Rac1, we synchronized cells in G2 using RO-­3306 and
obtained lysates immediately after washout (G2 fraction). We found
that Wee1 and Y15-­Cdk1 levels were reduced in Rac1−/− CD cells in
G2 (Fig. 9, E and F). Since Wee1 levels and function are regulated by
proteasomal degradation (64, 66, 67), we treated asynchronous con-
trol and Rac1−/− CD cells with the protein synthesis inhibitor cyclo-
heximide and followed Wee1 levels over time. While Wee1 levels
remained relatively stable over 5 to 7 hours in Rac1f/f CD cells, they
dropped sharply after 1 to 2 hours in Rac1−/− CD cells indicating that
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
Rac1 is required for Wee1 protein stability (Fig. 9, G and H). As Wee1
levels are modulated by actomyosin and degraded by excessive myo-
sin contractility (65), we tested whether Wee1 instability was revers-
ible with blebbistatin treatment of the Rac−/− CD cells. This was
indeed the case (Fig. 9, G and H), suggesting that the effect in the
Rac1-­deficient CD cells was mediated by excessive actomyosin-­
dependent forces. Furthermore, myosin inhibition in G2-­synchronized
Rac1−/− CD cells with blebbistatin or an ROCK inhibitor abrogated
premature mitotic entry, suggesting that the actomyosin dependent
stabilization of Wee1 downstream of Rac1 is required for normal mi-
totic entry (Fig. 9, I and J, and fig. S27).
Wee1 inhibition recapitulates Rac1 deficiency
To determine whether Rac1-­dependent Wee1 stabilization was suffi-
cient to promote normal mitotic entry, we treated G2-­synchronized
Rac1f/f CD cells with the Wee1-­specific inhibitor adavosertib (MK-­
1775). These cells behaved similar to the Rac1−/− CD cells showing
premature mitotic entry and mitosis-­associated cell death indicating
that Rac1-­dependent regulation of Wee1 is sufficient to control mi-
totic entry (Fig. 10, A to D). We also verified that Wee1 inhibition re-
sulted in abnormal actin–dependent mitotic rounding and mitotic
progression in postwound closure Rac1f/f CD cells or CD spheroids in
a 3D matrix in vitro (Fig. 10, E to H, and fig. S28 and movie S7). Thus,
Rac1-­dependent stabilization of Wee1 controls mitotic entry that is
required to maintain mitotic morphology and progression.
Rac1 levels are associated with preserved CD morphology in
human chronic kidney disease
Our mouse model indicates that Rac1 is essential for maintaining and
rebuilding the structural integrity of injured CDs. In humans, CD in-
tegrity is a major determinant of renal function, and many known
causes of chronic kidney disease are associated with notable CD mor-
phological abnormalities (68–70). Thus, we analyzed the protein level
of total Rac1 in the CDs of kidney biopsies from healthy controls or
individuals with chronic kidney disease with fibrosis. Rac1 is readily
detected in the CDs of healthy controls (Fig. 11A), and its expression
is preserved in CDs from patients with chronic kidney disease who
maintain their typical cuboidal to columnar morphology (Fig. 11, A
and B). By contrast, the levels of Rac1 are markedly decreased in the
dysmorphic CDs of individuals with chronic kidney disease, suggest-
ing that Rac1 expression plays a role in the preservation of CD mor-
phology in chronic kidney disease (Fig. 11C).
Mechanical CD cell injury results in Hace1-­mediated Rac1
ubiquitination and degradation
The finding that total Rac1 levels correlate with preserved morpholo-
gy in human chronic kidney disease (CKD) suggested that this is a
critical event in mediating irreversible chronic injury of CDs. In addi-
tion, proteasome-­mediated degradation of Rac1 was shown to con-
tribute to loss of epithelial morphology (71). We therefore sought to
determine the mechanism whereby total Rac1 is regulated. The Rac1-­
specific ubiquitin ligase was identified to be the HECT domain and
ankyrin repeat containing E3 ubiquitin protein ligase 1 (HACE1),
which polyubiquitinates (>50 kDa) Rac1 and efficiently targets Rac1
for degradation (72–74). Furthermore, Hace1 was recently identified
as a mechanosensitive ubiquitin ligase linking extracellular matrix
compliance to Rac1-­driven migration and cell cycling (75). When we
assessed the levels of Hace1 in our murine model, we found it to be
up-­regulated in irreversible but not in reversible ureteral obstruction.
13 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 9. Rac1 via actomyosin regulates the G2-­M checkpoint kinase Wee1 to prevent premature mitotic entry. (A and B) CD cells were G2-­synchronized using RO-­
3306, and lysates were collected at the indicated time points after RO-­3306 washout (“G2 release”). Lysates were immunoblotted for pH3 to monitor mitotic entry. Three
repeat experiments were quantified in (B). Fold change values ± SD. Arrows highlight the first pH3 peak of the respective groups indicating mitotic entry. (C and D) G2-­
synchronized CD cells were immunoblotted for cleaved caspase 3 (cl-­Casp3) after G2 release and quantified in (D) as fold change values ± SD (n = 3). Asterisk (*) denotes
between-­group significance at the corresponding time point. (E and F) CD cells were G2-­synchronized, and lysates were obtained immediately upon G2 washout (G2, T0)
and immunoblotted in biological duplicates. p-­Cdk1 Y15: phosphorylated Cdk1 tyrosine-­15. Three repeat experiments were quantified in (F). Fold change values ± SD. (G
and H) Asynchronous Rac1f/f (+DMSO) and Rac1−/− [+DMSO or blebbistatin (5 μM)] CD cells were treated with cycloheximide (CHX; 100 μg/ml), and lysates were obtained
at the indicated time points and immunoblotted for Wee1. Three repeat experiments are quantified in (H) as fold change ± SD. Asterisk (*) denotes significance between
Rac1−/− and Rac1f/f or blebbistatin-­treated Rac1−/− at the corresponding time point. (I and J) Rac1f/f and Rac1−/− CD cells were G2-­synchronized and treated with vehicle
(DMSO) or 5 μM blebbistatin upon G2 release. Lysates were collected at the indicated time points and immunoblotted for pH3 to monitor mitotic entry. Three independent
repeat experiments are quantified (for Rac1f/f, only the vehicle control is shown) in (J) as fold change values ± SD. Asterisk (*) denotes significance between Rac1−/− and
Rac1f/f or blebbistatin-­treated Rac1−/− at the corresponding time point. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 14 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 10. Wee1 inhibition phenocopies Rac1 deficiency in mitosis. (A to D) Rac1f/f and Rac1−/− CD cells were G2-­synchronized using RO-­3306 and treated with the Wee1-­
specific inhibitor MK-­1775 (1 μM) upon G2 release. Lysates were collected at the indicated time points and immunoblotted for pH3 to monitor mitotic entry or cleaved
caspase 3 to monitor cell death. Densitometry was used to quantify fold changes ± SD of three repeat experiments in (B) and (D). Arrows in (B) highlight the first pH3 peak
indicating mitotic entry. (E) F-­actin (white)– and DNA (blue)–labeled Rac1f/f and Rac1−/− (+MK-­1775; 1 μM) CD cell monolayers analyzed by confocal microscopy with a
mitotic metaphase cell shown in the center (scale bars, 10 μm). The top row column depicts metaphase F-­actin (scale bars, 5 μm) as outlined by a red continuous box in
the bottom row. Images are representative of at least three experiments. (F) Circularity quantification of mitotic metaphase F-­actin as shown in (E) with a minimum of
10 measurements shown per group. Bars are means ± SD. (G) Live confocal mitosis imaging of SPY650-­DNA (blue)– and SPY555-­actin (white)–labeled vehicle (DMSO)– or
MK-­1775 (1 μM)–treated Rac1f/f CD cells in vitro. The mitotic cell was manually segmented and colored in green. Scale bars, 10 μm. Sequences are representative of three
repeat experiments with at least three to four mitoses analyzed per experiment. (H) Relative distribution of mitotic defects in vitro during live imaging cell division se-
quences with at least 10 mitoses analyzed per group. *P < 0.05.

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 15 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 11. Rac1 is associated with preserved CD morphology in chronic kidney
disease. (A) Rac1 immunostaining (white) of human kidney biopsy specimens (par-
affin) with the CDs marked by AQP2 (green). Scale bars, 20 μm. Control (“healthy”)
biopsy displays a typical columnar to cuboidal CD morphology. Chronic kidney dis-
ease biopsy specimens were grouped according to CD morphology (preserved ver-
sus dysmorphic). Two or 4 representative stainings from different subjects (n = 3
healthy controls and 6 different chronic kidney disease biopsy specimens) are
shown per group. The rightmost panel shows an inset that is outlined by a dashed
yellow box in the original image. Scale bars, 10 μm. (B) Quantification of normalized
Rac1 staining intensity per cell with dots representing the intensity per sample with
35 cells per sample analyzed. Bars are means ± SD. *P < 0.05. (C) Schematic repre-
sentation of the proposed association between Rac1 and CD morphology. Created
with Biorender.com.
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
Consistent with this, Rac1 levels decrease in irreversible but not re-
versible injury (fig. S29). To assess the mechanistic relevance of this
finding, we used a well-­established in vitro cell compression model
where confluent cells are covered with a 1:1 mixture of low gelling
agarose and full medium creating a thin gel layer with a defined height
and weight that provides low but constant pressure deforming the cell
layer (fig, S30, A and B) (76–79). Long (4 hours)—but not short (1
hour)—cell compression reduced Rac1 and increased Hace1 levels
(fig. S30, C and D). We then performed Rac1 immunoprecipitation,
followed by Western blotting for ubiquitination at different time
points after cell compression. These studies indicate that long but not
short cell compression of CD cells in vitro lead to polyubiquitination
and degradation of Rac1 (fig. S30E). This effect was abolished by ge-
netic inhibition of Hace1 using lentivirus-­mediated transfection of
Hace1 short hairpin RNA (shRNA) into compressed control cells
(fig. S30, F to H). These data suggest a model whereby chronic me-
chanical epithelial injury results in Hace1-­mediated Rac1 degrada-
tion, which potentially hampers cellular adaption to a higher
mechanical pressure environment as found in chronic human kidney
disease with a profibrotic stiff extracellular matrix.
DISCUSSION
Relieving kidney collecting system obstruction before irreversible fi-
brosis results in tubule repair and regeneration. This requires complex
cellular processes that include reconstitution of cell morphology and
proliferation. In this study, we show that Rac1-­dependent regulation
of the actin cytoskeleton, a key component for normal cell morphol-
ogy, is required for the cell proliferation necessary for CD repair
following reversal of obstructive injury. We demonstrate that an ab-
normal actin cytoskeleton induced by deleting Rac1 in CDs results in
a dysmorphic epithelial cell that aberrantly enters mitosis and fails to
maintain mitotic rounding resulting in mitotic extrusion, apoptosis,
and impaired proliferation and repair. This is mediated by the inabil-
ity of the abnormal actin cytoskeleton to stabilize the G2-­M check-
point inhibitor Wee1 (fig. S31). This finding together with our
observation that decreased Rac1 expression is observed in dysmor-
phic CDs of patients with chronic kidney disease suggests that Rac1
mediates the morphological integrity of CD repair. Furthermore, it
highlights that preventing dysmorphic cells from aberrantly entering
mitosis is critical for epithelial repair.
There is little data on the role of Rac1 in kidney injury. Most stud-
ies were performed in the podocyte where Rac1 deletion induced no
discernible developmental phenotype (80). However, selective Rac1
deletion in podocytes was shown to be protective in a model of prot-
amine induced podocyte foot process effacement (80). Consistent
with this finding, abnormal activation of Rac1 expressed in podocytes
were also shown to induce functional alterations, maladaptive cyto-
skeletal remodeling, and focal segmental glomerulosclerosis (81–83).
These studies are in contrast with the finding that Rac1 deletion in-
creased podocyte loss and glomerulosclerosis in diabetic nephropa-
thy and Adriamycin models (84). Thus, in the podocyte, Rac1 exerts
different effects depending on the injury model. There is even less
data on the role of Rac1 in kidney tubule development and disease.
We and others showed that Rac1 is required for normal development
and function of the kidney collecting system (17, 18). In the context
of tubule injury, Rac1 contributes to oxalate-­induced nephropathy by
increasing reduced form of nicotinamide adenine dinucleotide phos-
phate–mediated cell injury and it enhances transforming growth
16 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
factor–β1–driven epithelial dysfunction in a UUO model (85). Last,
Rac1 expression is increased following acute ischemic injury of proxi-
mal tubules (86). We now show that decreased Rac1 expression cor-
relates with CD tubule damage and morphology in human individuals
with chronic kidney disease. Furthermore, we show that loss of Rac1
in the CD makes mice more susceptible to UUO injury and prevents
CD recovery in a reversible UUO model. Our observations are con-
sistent with the finding that deleting Rac1 in luminal epithelial cells in
the breast caused ductal apoptotic cell shedding and defective epithe-
lial regeneration during the reproductive cycle (87). Our data are also
consistent with the recent observation that exogenous administration
of constitutively active and cell-­permeable Rac1 peptides protects the
lung epithelium against pathogen-­induced injury by stabilizing the
actin cytoskeleton (88). Thus, like in the mammary gland or the lung,
Rac1 is critical in maintaining CD cytoskeletal structure in the setting
of injury and injury recovery. We substantially extend these studies by
defining how Rac1-­dependent actin organization is an essential re-
quirement for normal cell cycling and cell division control.
One of the unique aspects of the reversible UUO model we used in
this study is that it allows one to define how Rac1 mediates CD cell
proliferation following injury. The CD is an excellent model to under-
stand this process, as unlike many differentiated epithelial cell types
(e.g., podocytes) that die after severe injury, quiescent CD cells are
able reenter the cell cycle and proliferate (89). We found that Rac1
maintains CD cells in the G2-­M cell cycle phase during repair, sug-
gesting that Rac1 may play a role in “slowing” of cell cycle progression.
This was unexpected as failed repair of the injured proximal tubule is
caused by a G2-­M arrest and rapid and sustained proliferation of
proximal tubule cells is required for recovery (8, 90). These discrepant
data likely indicate cell type–specific control of the cell cycle and un-
derscore our incomplete understanding of the physiologic meaning of
cell cycle states. Our data extend prior observations from other model
systems that transient stalling or reversible prolongation of the cell
cycle in G2 is critical for successful repair of tissue injury (8, 91–94).
Cell cycle slowing after the acute kidney injury phase confers protec-
tion against chronic epithelial injury (92). The reason for as well as the
mechanism of this physiologic cell cycle and G2 prolongation has not
been defined. We found that Rac1 stabilizes Wee1 to prevent early and
rapid mitotic entry during G2. Wee1 in yeast and Xenopus constitutes
a “morphogenesis checkpoint” that puts the brakes on cell cycle pro-
gression allowing for the coordination of epithelial morphogenesis
and proper actin cytoskeletal organization with mitosis (95–98).
These data suggest that physiologic G2 prolongation is a key step in
epithelial repair. It is unclear whether this is a conserved element of
repair controlled by Rac1 in other organs or cell types.
We found that Rac1-­deficient CD cells are unable to undergo nor-
mal mitotic rounding and fail the progression through mitosis. This
result conflicts with other studies showing that Rac1 deficiency has no
effect on mitosis but, instead, its inactivation is required for successful
cytokinesis (99). In cancer cells, Rac1 activity is suppressed at the cell
equator during metaphase and anaphase to promote RhoA-­dependent
restriction of actomyosin contractility to a narrow zone for efficient
membrane ingression in early cytokinesis (99, 100). While we cannot
rule out that Rac1 regulation is required for efficient cytokinesis dur-
ing repair, we did not detect a difference in abnormal polyploidy or
aneuploidy (common outcomes of failed cytokinesis) in repairing CD
cells either expressing or lacking Rac1. Since we found that CD cells
lacking Rac1 are extruded or die in metaphase, it is also conceivable
that cells are eliminated before other mitotic defects become apparent.
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
We find that Rac1-­driven suppression of RhoA-­ROCK–dependent
myosin contractility is essential for early G2 cell cycling and epithelial
mitotic integrity. This finding extends prior experimental evidence
that constitutive RhoA activation in epithelial monolayers is sufficient
to drive mitotic rounding failure and mitosis-­associated cell extrusion
(101, 102). Well-­known molecular mechanisms exist where mechani-
cally unfit epithelial cells (loss of polarity and flattened morphology)
are eliminated from the epithelium by extrusion or apoptosis via le-
thal p53 induction (103). Tight spatial and temporal regulation of
Rac1-­RhoA antagonism is likely a critical part of this mechanism.
When we investigated the mechanism underpinning the prolifera-
tion and mitosis defect, we found that Rac1 is required to stabilize the
mitotic entry and G2 checkpoint kinase Wee1. Our findings extend
previous data showing that Madin-­Darby canine kidney cells have a
mechanical G2-­M checkpoint whereby increased E-­cadherin tension
at lower cell densities leads to myosin contractility-­dependent degra-
dation of Wee1 and rapid mitotic entry (65). Conversely, when cell
density is high, E-­cadherin forces are low, and Wee1 is stabilized and
G2-­prolonged (65). This endows the epithelial cell cycle with a mecha-
noresponsive mechanism that is critical in adapting the proliferative
response to external stress (e.g. during repair). The upstream regula-
tors of mechanoresponsive G2 and mitotic entry regulation in epithe-
lial cells remain undefined. We propose that Rac1—by controlling
actomyosin tension and cell-­cell junction integrity—could act as a
master regulator of Wee1-­dependent mechanical G2-­M checkpoint
regulation in polarized epithelial cells. A key question is how Wee1
protein levels are stabilized downstream of actomyosin contractility
and cell tension. The specific ubiquitin ligases that target Wee1 in a
cell cycle specific manner are known (67, 104, 105). It would be inter-
esting to define the mechanism whereby epithelial tension and myo-
sin contractility regulate Wee1 and its ubiquitination and proteasomal
degradation.
The repairing kidney epithelium is a highly confined space. How
morphologically challenging processes such as rapid rebuilding of the
actin cytoskeleton and cell division are spatially enabled is unknown.
We found here that Rac1-­controlled actin cytoskeletal organization is
an essential requirement for the repairing CD cells to maintain their
shape. Without Rac1 the proliferating epithelium is dysmorphic,
premitotic cells in G2 fail to go into the appropriate cell cycle pause
and are unable to exert pressure on their neighbors and fail to divide.
Targeting this mechanism through direct actomyosin inhibition
reversed the Rac1-­deficient phenotype. Furthermore, we found
that prolonged mechanical epithelial injury leads to proteasome-­
dependent Rac1 degradation. This extends recent experiments show-
ing that volumetric cell compression suppresses Rac1 and tilts the
Rac1/RhoA balance toward RhoA with increased myosin activation
resulting in maladaptive cytoskeletal remodeling (76). Either forced
Rac1 activation or myosin inhibition with an ROCK inhibitor was
sufficient to recover the compressed cellular phenotype (76). These
observations and our data highlight Rac1 as a central regulator of me-
chanical cell states in injury and repair. This raises the question
whether “mechanical” therapeutic approaches or mechanobiology-­
based treatment designs that target the actin cytoskeleton can en-
hance repair after mechanical kidney damage. This has been proposed
and tested in other organs. For example, lungs lacking the small Rho
GTPase Cdc42 develop pulmonary fibrosis that is caused by mechan-
ical cytoskeletal tension on alveolar epithelial cells (106). Implanta-
tion of a custom-­made silicone prosthesis alleviated mechanical
tension in these lungs, enhanced regeneration, and prevented fibrosis
17 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(106). In addition, mechanical therapies (e.g. based on shockwaves)
are already known to enhance recovery for other injured tissues such
as the heart, skin, and bone (107). Exploiting mechanoregulatory cues
such as extracellular matrix stiffness is an established and powerful
avenue to control stem cell differentiation or cancer cell behavior
(108). However, it should be noted that mechanical targeting ap-
proaches (e.g., myosin inhibition) are not cell specific. For example,
blebbistatin can affect macrophage motility in kidney injury models
that could contribute to its nephroprotective effects. (45). Neverthe-
less, our data suggest that mechanical targeting of actin cytoskeletal-­
based mechanisms may represent a viable therapeutic principle to
enhance postobstructive kidney regeneration, which is worth future
exploration.
In conclusion, we provide evidence that Rac1 in kidney CD or-
chestrates repair by promoting epithelial F-­actin organization and
orderly actin-­dependent mitotic entry and progression. This mecha-
nism reveals how epithelial morphology, cell cycle, and mitotic entry
are mechanically coupled.
MATERIALS AND METHODS
Mice
All experiments were approved by the Vanderbilt University Institu-
tional Animal Use and Care Committee and conducted in Associa-
tion for Assessment and Accreditation of Laboratory Animal Care
(AALAC)-­accredited facilities. All mice were backcrossed onto the
C57BL/6N background for multiple generations. Rac1f/f mice were
previously described (109) and were crossed with AQP2Cre mice (the
Jackson laboratory). Age-­and gender-­matched littermates (Rac1f/f
mice) were used as controls.
UUO or reversible ureteral obstruction (R-­UUO) was performed
as previously described (24–26) on 8-­to 12-­week-­old male Rac1f/f
or AQP2Cre:Rac1f/f mice (20 to 25 g). Mice were anesthetized with
ketamine-­xylazine and administered preoperative ketofen and iso-
tonic saline. The right ureter was exposed, and clamping was per-
formed on the ureter just distal to the renal pelvis with either sutures
(UUO) or an atraumatic microvascular clamp (R-­UUO, Fine Sci-
ence Tools, 3.5-­mm × 1-­mm, 5-­to 15-­g press, #00396-­01). For irre-
versible UUO, obstructed kidneys were harvested 10 days after
suture placement. For R-­UUO, the peritoneal cavity was reopened,
and successful ureteral obstruction was confirmed by assessing the
presence of hydronephrosis and paleness of the right kidney. The
ureteral clamp was removed using a clamp applier (Fine Science
Tools, #18040-­14), and mice were euthanized on days 5, 7, and 30
after clamp removal. Upon euthanasia, successful decompression
was confirmed by visual confirmation of the absence of hydrone-
phrosis and ureteral bulging (fig. S4). A subgroup of animals was
intraperitoneally injected with dimethyl sulfoxide (DMSO) (mock),
blebbistatin (2 mg/kg, daily; Sigma-­Aldrich, #B0560) or Y-­27632
(10 mg/kg, daily; Tocris Bioscience #1254), doses within a range
previously found to affect organ injury in various animal models
(45, 110–112). Plasma blood urea nitrogen (Bioassay Systems Quan-
tiChrom Urea Assay Kit, #DIUR-­100) was measured per the manu-
facturer’s instructions, and spot urine osmolality (in milliosmolal
per kilogram; freezing point depression osmometer) was measured
as previously described (18). For BrdU pulse labeling, in vivo mice
were injected with BrdU (100 mg/kg; Sigma-­Aldrich, #B5002) as a
single intraperitoneal injection for 3 to 4 hours before euthanasia.
Animal numbers are stated in each legend.
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
Histology
Kidneys were removed, coronally cut with a razor in 2-­to 3-­mm-­thick
slices, fixed in 4% paraformaldehyde (PFA) for 30 min, and embed-
ded in paraffin. Paraffin tissue sections (5 μm) were dewaxed and re-
hydrated by successive immersion for 5 min in 100% xylene and 100,
90, and 70% ethanol solutions and stained with hematoxylin and eo-
sin (H&E) for morphological evaluation by light microscopy. Colla-
gen accumulation in kidney sections was determined by staining for
Sirius Red and quantified by image analysis using ImageJ [National
Institutes of Health (NIH)] as described (https://imagej.nih.gov/ij/
docs/examples/stained-­ sections/index.html). Fibrillar collagen was
expressed as percentage area occupied by picrosirius red–positive
structures/microscopic field. A minimum of five nonoverlapping kid-
ney cortex areas were analyzed for each kidney. For 2D immunofluo-
rescence, paraffin kidney sections were subjected to heat-­induced
antigen retrieval in a citrate-­buffered solution (BioGenex, #HK086-­
9K). Sections were blocked, and antibodies were diluted in 3% bovine
serum albumin (BSA), 10% normal horse serum, and 0.01% Tween
20 in phosphate-­buffered saline (PBS). For mounting, ProLong Gold
Antifade was used (Thermo Fisher Scientific, #P10144). CDs were la-
beled with AQP2 (Cell Signaling Technology, #3487), preconjugated
[Alexa Fluor (AF)–488 or AF-­647] AQP2 antibodies (Santa Cruz
Biotechnology, #sc-­515770), or fluorescein isothiocyanate (FITC)–
conjugated dolichus biflorus agglutinin (DBA; Vector Labs, #-­1031).
Terminal deoxynucleotidyl transferase–mediated deoxyuridine tri-
phosphate nick end labeling (TUNEL) assay was performed on
paraffin sections as described by the manufacturer’s instructions
(Sigma-­Aldrich, #11684795910). Other antibodies include anti-­BrdU
(Cell Signaling Technology, #5292), pH3 (S10) antibody (Active
Motif, #39098), mouse anti–E-­cadherin (BD Biosciences, #610181),
pMLC (Cell Signaling Technology, #3671), and FITC preconjugated
anti-­Ki-­67 (Thermo Fisher Scientific, #11-­5698-­82). Secondary
antibodies included the following: AF-­488 anti-­rabbit (Thermo Fisher
Scientific, #A21206), AF-­647 anti-­rabbit (Thermo Fisher Scientif-
ic, #A21245), AF-­647 anti-­mouse (Thermo Fisher Scientific,
#A212236), and AF-­555 anti-­rat (Thermo Fisher Scientific, #78945).
4′,6-­Diamidino-­2-­phenylindole (DAPI) (Cell Signaling Technology,
#4083) was routinely added at a 1:500 dilution to secondary antibody
incubation to label DNA. For all histological quantifications, scatter
plot dots represent the average per mouse with at least 10 measure-
ments per mouse.
Determination of mitotic stages
Analysis of mitotic stages in vitro and in vivo was performed using
high-­resolution confocal microscopy images. In vivo prophase was
defined as a cell with nuclear positivity for pH3 and condensed
chromatin relative to interphase cells. Metaphase was defined as a
pH3-­positive cell with DNA aligned at the center of the cell. In
time-­lapse imaging in vitro, mitotic condensation of prophase
DNA and the increasing condensation during metaphase were
used to determine the transition from prophase to metaphase.
Anaphase was generally defined as a cell with an oblong mem-
brane containing a clear separation of DNA to two poles. Telo-
phase was defined as cells that have undergone membrane division
with a clear separation between DNA and membrane within two
daughter cells. For visualization of live imaging data, mitotic cells
were manually segmented using Fiji/ImageJ, and a green lookup
table (LUT) was applied to mitotic DNA. All groups were pro-
cessed identically.
18 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Optical clearing
Optical clearing was performed as previously described with
some modifications (28, 113, 114). In vivo transcardiac perfusion
with 10 to 15 ml of PBS infused over 5 min was used to wash out
blood. The kidney was removed and coronally sliced with a razor
at 2 to 3 mm in thickness, the cortex was removed, and slices were
incubated for 30 min in freshly prepared 4% PFA rocking at room
temperature, washed three times in PBS, and then incubated in
blocking buffer (2% BSA, 0.2% Tween 20, 0.2% Triton X-­100, and
0.02% sodium azide) for 4 to 5 hours at room temperature. Next,
sections were incubated in primary antibody (at least 1:50 dilu-
tion depending on the primary) in blocking buffer for 3 days at
room temperature on a rocking platform, washed, and then incu-
bated in secondary antibodies (+DAPI) for 3 days at room tem-
perature. Slices were then washed thoroughly three times in PBS
with 2 to 3 hours per washing step. Slices were dehydrated in a
methanol series (25, 50, 75, 100, and 100%), 5 min each step, and
transferred into BABB solution (benzyl alcohol and benzyl ben-
zoate at a 1:2 ratio) in an Eppendorf tube for 20 min. Once the
slices were translucent, they were placed into a glass-­bottom dish
(MatTek, #P35-­1.5-­14-­C) containing 100 μl of BABB solution for
confocal imaging.
Confocal microscopy and image processing
Images were collected with confocal microscopy at super-­resolution
using a Zeiss LSM 980 confocal microscope equipped with an in-
verted Axio Observer 7 and Airyscan 2 detector. The objective
used was a 63×/1.4 numerical aperture (NA) Plan Apochromat oil
or 10×/0.50 NA Plan Apochromat (for low-­powered scanning of
Ki-­67–labeled and optically cleared medullary slices). Airyscan
super-­resolution images were acquired under identical settings for
all groups and images. Acquisition and 2D Airyscan processing of
acquired images was done using ZEN Blue software (Carl Zeiss).
For optimal optical sectioning 3D reconstructions, auto Z bright-
ness corrections were applied that corrects for differences in
optical tissue density during deep imaging. To computationally
subtract bleed-­through between channels, spectral profiles of each
fluorescence channel were analyzed, and linear unmixing was per-
formed using ZEN Blue. Next, image stacks were exported into
Fiji/ImageJ for brightness/contrast adjustments and LUT assign-
ment. Image stacks were then exported into Imaris (Bitplane). For
AQP2/Ki-­67 double labeled medullary slices, AQP2+ tubules were
converted into surfaces using automatic segmentation and thresh-
olding. Next, Ki-­67 was converted into spots using the “create
spots” function and thresholding by quality. Last, the Imaris XTen-
sion (written in MATLAB), “spots close to surface” was activated
and run with a zero-­distance threshold to filter the spots inside
AQP2+ surfaces. For display, Ki-­67 was shown as spots, while for
AQP2, the original 3D reconstruction is shown. For tubular 3D
F-­actin reconstructions, in vivo image stacks were exported into
Imaris. A surface around subapical F-­actin was manually created
for the whole-­stack segmentation. Next, a mask was created on the
F-­actin channel (AF-­647), and the voxels outside the subapical ac-
tin surface were set to zero. This results in displaying apical F-­actin
only. To display basolateral F-­actin or total tubular F-­actin, an-
other surface was created around the whole tube, and subapical, as
well as actin outside the tube, was masked. For mitotic metaphase
F-­actin surface reconstruction, Imaris software was used as out-
lined above using surface creation and manual segmentation.
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
Thick frozen sectioning and F-­actin labeling in vivo
For actin cytoskeletal preservation in live tissue, kidneys were perfu-
sion fixed in 50:50 mix of 2% PFA and 2% glutaraldehyde in PBS, and
the renal papillae including the upper inner medulla were dissected.
For cryoprotection, tissue was then incubated in 30% sucrose for
48 hours, after which they were transferred into cryomolds for opti-
mal cutting temperature (OCT) embedding (Fisher Scientific, #23-­
730-­571). Thick cryosectioning was performed at 50 μm in thickness
using a Leica cryostat. Sections were dried, rehydrated, and incubated
with AF-­647–phalloidin (Thermo Fisher Scientific, #A22287; 1:50)
for 24 hours. For simple immersion, aqueous-­based optical clearing
sections were incubated in refractive index matching solutions (40 g
of Histodenz in 30 ml of 0.02 M phosphate buffer with 0.01% sodium
azide, brought to a pH of 7.5 with NaOH) for 2 to 3 hours or until
sections became translucent. 3D super-­resolution confocal imaging
and processing was performed as outlined above. For assessment of
total actin content, the mean global gray value intensity (of back-
ground normalized Z-­stacks for each dataset) of maximum intensity
projections (five Z-­slices each) was obtained using ImageJ. For apical
F-­actin intensity, a similar operation was performed, but the mean
gray value of regions of interest drawn around subapically segmented
F-­actin was used. To obtain the apical F-­actin major axis length, or-
thogonal projections of whole tubular Z-­stacks sliced across the me-
dian were generated. Longitudinal cell-­to-­cell junction distances were
then measured and calculated using ImageJ’s line tool. Annotated ex-
amples of histological quantifications are further illustrated in fig. S32.
Cell culture and cell cycle synchronization
Rac1−/− CD cells were generated and cultured as previously described
(18). To assess mitotic rounding in polarized monolayers, cells were
grown on Transwell inserts consisting of polyvinyl pyrrolidone–free
polycarbonate filters with 0.4-­μm pores (Corning, #3460) as previ-
ously described. Confluent cells were fixed after wounding (see be-
low) in 4% formaldehyde and processed for AF-­ 647–phalloidin,
activated actomyosin (pMLC S19, Cell Signaling Technology, #3671),
and DAPI labeling. Cell cycle synchronization at the G1-­S boundary
was performed using a double thymidine block as previously de-
scribed (48). Briefly, thymidine was added at a final concentration of
2 mM to CD cells at 50 to 60% confluency for 18 hours. Next, cells
were washed in PBS once and regular prewarmed medium [Dulbec-
co’s modified Eagle’s medium (DMEM) and 10% fetal bovine serum
(FBS)] was added back for 9 hours. After this, a second round of 2 mM
thymidine was added for 18 hours after which cells were washed in
PBS twice and incubated in regular medium for timed collections.
Cell cycle synchronization in G2 at the G2-­M boundary was per-
formed by treating CD cells with 2 μM Cdk1 inhibitor RO-­3306
(R&D Systems, #4181) for 20 hours, washed twice with PBS, and then
cultured in full regular medium. Other cell culture treatments included
nocodazole (Tocris Bioscience, #1228) at 100 ng/ml or the Wee1
inhibitor Adavosertib (MK-­1775, Tocris Bioscience, #7589) at 1 μM,
blebbistatin (Sigma-­Aldrich, #203391) at 5 μM, or the ROCK inhibi-
tor Y-­27632 at 1 μM. For cell treatments, all drugs were dissolved for
stock solutions and used according to the manufacturer’s recommen-
dations. Solvents for stock solutions were used as negative control
treatments in respective experiments. All in vitro experiments were
repeated three times unless otherwise indicated. For lentiviral trans-
fections of shRNA, the following vectors were ordered prepackaged in
lentiviruses (VectorBuilder). For GEF-­H1 (Arhgef2), pLV[shRNA]-­
EGFP:T2A:Puro-­U6>mArhgef2 (ID: VB900050-­0695dkh), and for
19 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Hace1, pLV[shRNA]-­EGFP:T2A:Puro-­U6>mHace1 (ID: VB900068-­
4913bgh). For the transduction of target cells, mouse CD cells were
cultured to 50% confluency in DMEM (10% FBS). The recommended
amount of virus per the manufacturer’s instructions was used and
added in virus-­containing Opti-­MEM with polybrene. Successful
transfection was monitored using the green fluorescent protein (GFP)
signal under the ZOE Fluorescent Cell Imager (Bio-­Rad) and con-
firmed by immunoblotting for the target protein.
Wounding assay
Scratch assays were performed as previously described (18). Briefly,
CD cells were grown to confluence on Transwell inserts consisting
of polyvinyl pyrrolidone–free polycarbonate filters with 0.4-­μm
pores (Corning, #3460). For wounding, confluent epithelial cell
sheets were scratched with a 10-­μl pipet tip. An area adjacent to the
scratch was monitored using the ZOE Fluorescent Cell Imager
(Bio-­Rad). As soon as the scratched area was close and the area was
fully covered with cells (15 hours for Rac1f/f and 24 hours for
Rac1−/− CD cells), the cell sheets were fixed, F-­actin was visualized
by phalloidin, and DNA with DAPI and mitotic cells were labeled
with anti-­pH3.
Cell stretch and compression
Cells were stretched as previously described (33, 115). Briefly, to
expose cells to uniaxial noncyclical mechanical stretch, CD cells
were grown to confluency on collagen-­coated flexible silicone six-­
well culture plates within a gasekt (Flexcell International Corpora-
tion, Burlington, NC, USA) in regular full DMEM. A controlled
vacuum is applied resulting in the membrane deformation across
the planar face of the postcreating uniform axial strain. See also
schematic representation in Fig. 4C. Static stretch was applied by
increasing the percent elongation (degree of stretch) from zero to
10 at ~0.3 to 0.5% per min during the 30 min (ramp up) and then
holding at 10% elongation for 3 hours. At the end of the protocol,
the percent elongation was decreased to zero over 5 to 10 min be-
fore removing plates from the vacuum manifold. Control plates
were not subjected to stretch. Cell compression was performed as
previously described (76–79). Briefly, low–gelling temperature aga-
rose (Sigma-­Aldrich, #A9045) was dissolved in PBS at 3% concen-
tration by heating in a microwave oven, followed by mixing the
solution with DMEM cell culture medium with 20% FBS in a 1:1 ratio
to form a 1.5% agarose gel solution containing full medium
(DMEM and 10% FBS), and poured into a round petri dish. Round
agarose cushions (diameter, ~20 mm) were cut, gently placed on
top of the cells, and weighted down with around 25 g of stainless-­
steel balls in which prior studies have shown to create ~5.8 mmHg
of constant pressure. For enrichment of ubiquitinylated Rac1 in
compressed CD cells, immunoprecipitation was performed as
previously described (116). Briefly, cells were harvested with
M-­PER after compression and gently sonicated, and debris was
removed by centrifugation (13,000 rpm for 10 min). Next, 10 μl
of immunoprecipitation-­validated Rac1 antibody (ProteinTech,
#24072-­1-­AP) was added to 150 μg of protein and incubated under
agitation at 4°C overnight. Then, 80 μl of 50% slurry of Protein A
agarose beads was added (Thermo Fisher Scientific/Pierce Protein
A Agarose, #20333) and incubated overnight at 4°C. Beads were
then precipitated, washed, and resuspended in 1× SDS-­loading
buffer for Western blotting of anti-­ubiquitin (Cell Signaling Tech-
nology, #58395 or #3936).
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
Live mitosis imaging
For live imaging of mitosis, control and Rac1−/− CD cells were grown
to confluence on glass-­bottom live imaging MatTek dishes (MatTek,
#P35G-­ 1.5-­
14-­C) precoated with growth factor–reduced Matrigel
(Corning, #356230). Wounding was performed as described above.
Upon wound closure and 1 hour before time-­lapse imaging, the
following live-­cell fluorogenic labeling probes were added according
to the manufacturer’s instructions (at 1× dilution of a 1000× stock)
in phenol red–free complete DMEM: SPY650-­DNA (Cytoskeleton,
#CY-­SC501) to label DNA and SPY555-­actin (Cytoskeleton, #CY-­
SC202) to label F-­actin. High-­resolution live confocal imaging was
carried out using a Zeiss LSM 980 microscope using a 63×/1.4 NA
Plan Apochromat oil objective and equipped with a 37°C humidified
incubator supplied with 5% CO2. Images were acquired at in the
Airyscan Multiplex SR-­8Y mode with subsequent deconvolution. In
general, one image every 10 min with 10 optical sections (every 1 μm)
per time point per frame was acquired. For visualization purposes, 4D
confocal stacks were subsequently exported into Fiji/ImageJ for man-
ual segmentation of mitotic chromosomes and green LUT reassign-
ment. A minimum of 10 mitotic events per condition were recorded.
Mitotic metaphase delay was defined as the inability to progress
through and successfully complete metaphase within 60 min. Treat-
ments (Wee1 inhibitor MK-­1775 at 1 μM or blebbistatin at 5 μM)
were added 30 min before time lapse imaging in phenol red–free
complete DMEM.
RhoA activity assay
RhoA GTPase activation assay was performed using the colorimetric
G-­LISA RhoA activation assay (Cytoskeleton, #BK124) according to
the manufacturer’s instructions and as previously described (117). Af-
ter the experimental treatments, cells were rinsed with ice-­cold PBS
and homogenized in ice-­cold lysis buffer. Protein was quantified and
sample concentration adjusted to 1 μg/μl using BCA Protein Assay
Kit (Thermo Fisher Scientific, #23225). Binding buffer was added, and
assays were performed using 20 μg of protein per well. Target binding
to the Rho GTP-­binding protein coated well was facilitated by 30 min
of incubation at 4°C with shaking, followed by washing. Antigen-­
presenting buffer was added for 2 min and removed, and samples were
incubated with the anti-­RhoA antibody (1:250) for 45 min at room
temperature. Samples were washed three times and incubated with
horseradish peroxidase–conjugated secondary antibody for 45 min
at room temperature. Horseradish peroxidase detection reagent was
added, the samples were incubated briefly, and absorbance was mea-
sured at 490 nm using a microplate reader (Molecular Devices).
3D spheroid culture
3D CD spheroids were generated as previously described (118). Brief-
ly, a CD cell suspension of 100,000 cells/100 μl in complete medium
was mixed with Matrigel (Corning, #356230) in a 1:1 ratio. 200 μl of
the cell:Matrigel mix were added to the media chambers of an of an
eight-­well glass slide (Millipore, #PEZGS0816). Upon solidification in
the incubator (at 37°C and 5% CO2) for 30 min, 200 μl of prewarmed
complete medium was added to the hardened cell:Matrigel mix. Cells
were cultured for 4 days, while monitoring spheroid formation using
a cell culture inverted light microscope. Medium was changed daily.
Treatments (Wee1 inhibitor MK-­1775 at 1 μM or blebbistatin at
2 μM) were added 16 hours before spheroid harvesting. For spheroid
harvesting, the gels were washed twice with calcium-­free PBS, the
Matrigel was dissolved, and the spheroids were fixed by incubating
20 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

the gels with fresh 4% PFA for 30 min at room temperature. Permea- 1 mM CaCl2 and 0.5 mM MgCl2 and incubated at 37°C for 1 hour.
bilization, blocking, immunocytochemistry, and immunofluores- Samples were then placed on ice and passed through a 70-­μm cell
cence were then performed per standard protocols. Spheroids were strainer (Thermo Fisher Scientific, #08-­771-­2). Cells were centrifuged
stained for F-­actin using AF-­647–conjugated phalloidin, α-­tubulin (5000g for 5 min at 4°C) and resuspended in 1.5 ml of 0.5% FBS/
(Cell Signaling Technology, #2144), and DNA (DAPI) and imaged us- PBS. Cell concentrations were then normalized to 1 million cells per
ing a Zeiss LSM 980 in super-­resolution multiplex mode, followed by ml across all samples using an automated cell counter (AE Bios, Cel-
mitotic F-­actin surface reconstructions using Imaris (Bitplane). lometer K2). Samples were incubated for 10 min at room temperature
with Fc blocking antibody, followed by incubation with anti-­AQP2
Traction force microscopy (Cell Signaling Technology, #3487) and anti-­pH3 (both 2 μl per 1 Mio
Traction force microscopy of CD cell layers was performed as briefly cells) and DAPI (1:200). Flow cytometry data were acquired with a
described with some modifications (119–121). For polydimethylsi- BD LSRFortessa analyzer at the Vanderbilt University Medical Center
loxane (PDMS) substrate preparation, the two-­part PDMS material Flow Cytometry Shared Resource. Data files were analyzed by flo-
Sylgard-­184 was used according to the manufacturer’s instructions. reada.io, and we performed cell cycle analysis by gating a plot of side
On the basis of previous characterizations, we chose an elastomer-­ or forward scatter versus DAPI staining. All runs included staining
to-­curer ratio of 80:1 that gives rise to a substrate stiffness of ~9 kPa. controls (secondary antibodies only) and compensation controls. Cell
One hundred microliters of the PDMS substrate mixture was applied cycle assessment in vitro was performed as previously described
to the coverslip of glass-­bottom imaging dishes (MatTek) using a cut (122). Briefly, CD cells were grown according to the required experi-
P200 tip and cured for 2 hours at 80°C. Next, the PDMS surface mental condition and detached using Accutase. Cell aggregates were
was silanized by adding 2 ml of 5% (v/v) 3-­aminopropyltriethoxysilane carefully broken up by pipetting up and down. The cells were then
(APTES) solution (in 100% ethanol) on top of the cured PDMS spun down at 1500 rpm for 5 min. The cell pellet was resuspended in
surface for 10 min. APTES was removed, and the PDMS surface was ice-­cold 70% ethanol while vortexing, and cells were fixed in ethanol
washed three times with 100% ethanol and dried at 80°C for 30 min. overnight. Cells were then spun down, ethanol was removed, and
Next, a 0.05% (v/v) fluorescent Nanobead solution (FluoSpheres the cell pellet was washed in PBS. Cells were then resuspended in
carboxylate-­modified microspheres, Thermo Fisher Scientific, propidium iodide staining solution [0.1% Triton X-­100 in PBS,
#F8811; diameter, 0.2 μm) was prepared in water and gently sonicated deoxyribonuclease-­free ribonuclease (100 μg/ml), and propidium io-
in a water bath sonicator. To functionalize the PDMS substrate with dide (50 μg/ml)] and subjected to flow cytometry.
nanobeads, the bead solution was incubated in the PDMS dish for
5 min at room temperature. After the bead solution was removed, the Immunoblotting
plate was washed with water three times and again dried at 80°C for Immunoblotting was performed as previously described (18). Briefly,
30 min. To inactivate the bead surface, the PDMS substrate was incu- cell lysates were prepared using M-­PER reagent (Thermo Fisher Sci-
bated in 100 mM tris solutions for 10 min at room temperature and entific; #78501). Lysates were resolved in 10 and 12% or gradient 4 to
again dried. PDMS dishes were then coated with Matrigel as outlined 20% SDS–polyacrylamide gel electrophoresis depending on protein
above and sterilized using 1% pluronic, after which CD cells were size and then transferred to nitrocellulose membranes. Membranes
seeded and grown to 100% confluence. Upon addition of live-­imaging were incubated with the primary antibody, followed by IRDye fluo-
dyes (SPY probes as outlined above), frames of cells were monitored rescent dyes secondary antibodies, and a LI-­COR Biosciences Odys-
using a Zeiss LSM 980 confocal and the live imaging setup as de- sey infrared imaging system was used for detection. Immunoreactive
scribed above to detect chromosome condensation upon mitotic en- bands were quantified by densitometry analysis using ImageJ or Im-
try. Stacks of five images with the fluorescent beads were obtained age Studio Lite. To quantify levels of protein phosphorylation, optical
during prometaphase. For the references image, fluorescent bead im- density of bands was normalized to total protein or β-­actin. For kid-
age stacks were obtained of the same frame after addition of 10% SDS ney tissue analysis, papillae were removed and mechanically lysed in
(reference image without cells). Bead image stacks with and without T-­PER reagent (Thermo Fisher Scientific, #78510) using a Polytron
cells were aligned using the slice alignment plug-­in (ImageJ), and gel homogenizer, centrifuged, and then subjected to immunoblotting.
deformation was calculated using the particle imaging velocimetry Primary antibodies used are pH3 (Cell Signaling Technology, #9701),
plug-­in (ImageJ). Forces were determined and force maps were gener- cleaved caspase 3 (Cell Signaling Technology, #9664), Wee1 (Novus
ated using the Fourier transform traction cytometry plug-­in (ImageJ). Biologicals, #NBP1-­33506), actin (Cell Signaling Technology, #4967),
cyclin B1 (Cell Signaling Technology, #4138), Rac1 (Millipore, #05-­
Flow cytometry 389), α-­tubulin (Cell Signaling Technology, #2144), Cdk1 (Cell Sig-
Flow cytometry and DNA content–based cell cycle assessment on naling Technology, #77055), and phospho–(Y15) Cdk1 (Cell Signaling
kidney tissue was performed as previously described with some mod- Technology, #4539). The secondary antibodies were IRDye 800CW
ifications (30). Briefly, kidneys were harvested and coronally sliced, anti-­rabbit (#926-­32213 and #926-­32211), IRDye 800CW anti-­mouse
and slices with an equal weight between groups were minced in fresh (#926-­32212 and #926-­32210), IRDye 680RD anti-­rabbit (#926-­
4% PFA on ice with 1:100 protease inhibitors. Tissue was then washed 68073 and #926-­68071), and IRDye 680 anti-­mouse (LI-­COR Biosci-
and pelleted by brief centrifugation, and additional mechanical tissue ences, #926-­68072 and # 926-­68070).
dissociation was performed using the gentleMACS Dissociator. The
tissue was then resuspended in a detergent solution (0.2% Triton X-­ Two-­photon laser-­induced injury in vitro
100) to permeabilize for 30 min at room temperature. After pelleting Induction of two-­ photon induced cellular microlesions was per-
and washing in PBS, the tissue was resuspended in dissociation formed as reported with some modifications (123). Briefly, control
solution [collagenase type 2, 1 mg/ml (Thermo Fisher Scientific, and Rac1−/− CD cells were grown to confluent monolayers on Tran-
#17101015) and dispase II, 1 mg/ml (Sigma-­Aldrich, #D4693)] in swell inserts (Corning, #3460). The Transwell filters containing the

Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024 21 of 24

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
confluent monolayers were excised using a razor, inverted, and
mounted onto the coverslip of a complete phenol-­free medium–
containing glass-­bottom dish (MatTek, #P35-­1.5-­14-­C). F-­actin was
labeled using SiR-­actin (Cytoskeleton, #CY-­SC001) or SPY555-­actin
(Cytoskeleton, #CY-­SC202) according to the manufacturer’s instruc-
tions. This was mounted on the stage of a Zeiss LSM 780 equipped
with a tunable-­range coherent chameleon Ti:sapphire multiphoton
laser. A 63×/1.40 NA Plan Apochromat oil Differential Interference
Contrast (DIC) objective was used for imaging. Experiments were
performed at 37°C in 25 mM Hepes. After collecting control images,
a small rectangular region (2 μm × 3 μm) of a single cell was scanned
at high Ti-­sapphire laser power (730 nm; 80% intensity) for 25 itera-
tions (duration, 8 to 10 s). F-­actin images of the same frame were sub-
sequently captured as indicated in the figure legend.
Human samples
Formalin-­fixed, paraffin-­embedded tissue sections were obtained
from deidentified tumor nephrectomy (renal cell cancer) specimens
through the Department of Pathology at Vanderbilt University Medi-
cal Center (VUMC). Control Images were obtained from the nontu-
mor healthy portion of the section. CKD biopsy specimens were
randomly selected and had to have at least CKD stage 2 and 10 to 20%
tubulointerstitial fibrosis. This use of human deidentified samples was
approved by the VUMC institutional review board.
Statistics
Data are shown as means ± SD or SEM. Unpaired two-­tailed t test was
used to evaluate statistically significant differences (*P < 0.05) be-
tween two groups. One-­way analysis of variance (ANOVA) was used
to test statistical significance among multiple groups. Post hoc com-
parisons of ANOVA were corrected with the method of Tukey. For
fold change calculations, the control number was normalized to 1 for
each comparison and does not have an error bar. Data distribution
was assumed to be normal, but this was not formally tested.
Supplementary Materials
This PDF file includes:
Figs. S1 to S32
Other Supplementary Material for this manuscript includes the following:
Movies S1 to S7
REFERENCES AND NOTES
1. J. Yaxley, W. Yaxley, Obstructive uropathy—Acute and chronic medical management.
World J Nephrol 12, 1–9 (2023).
2. B. A. Molitoris, Low-­flow acute kidney injury: The pathophysiology of prerenal azotemia,
abdominal compartment syndrome, and obstructive uropathy. Clin. J. Am. Soc. Nephrol.
17, 1039–1049 (2022).
3. M. S. Forbes, B. A. Thornhill, J. J. Minor, K. A. Gordon, C. I. Galarreta, R. L. Chevalier,
Fight-­or-­flight: Murine unilateral ureteral obstruction causes extensive proximal tubular
degeneration, collecting duct dilatation, and minimal fibrosis. Am. J. Physiol. Renal
Physiol. 303, F120–F129 (2012).
4. M. J. Hiatt, L. Ivanova, P. Trnka, M. Solomon, D. G. Matsell, Urinary tract obstruction in the
mouse: The kinetics of distal nephron injury. Lab. Invest. 93, 1012–1023 (2013).
5. Z. M. El-­Zoghby, J. C. Lieske, R. N. Foley, E. J. Bergstralh, X. Li, L. J. Melton III,
A. E. Krambeck, A. D. Rule, Urolithiasis and the risk of ESRD. Clin. J. Am. Soc. Nephrol. 7,
1409–1415 (2012).
6. J. Yang, B. G. Sun, H. J. Min, Y. B. Son, T. B. Kim, J. Lee, S. W. Oh, M. G. Kim, W. Y. Cho,
S. Y. Ahn, G. J. Ko, Y. J. Kwon, J. J. Cha, Y. S. Kang, D. R. Cha, S. K. Jo, Impact of acute kidney
injury on long-­term adverse outcomes in obstructive uropathy. Sci. Rep. 11, 23639 (2021).
7. P. Trnka, M. J. Hiatt, A. F. Tarantal, D. G. Matsell, Congenital urinary tract obstruction:
Defining markers of developmental kidney injury. Pediatr. Res. 72, 446–454 (2012).
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
8. K. Lee, G. L. Gusella, J. C. He, Epithelial proliferation and cell cycle dysregulation in kidney
injury and disease. Kidney Int. 100, 67–78 (2021).
9. M. Genesca, A. Sola, G. Hotter, Actin cytoskeleton derangement induces apoptosis in
renal ischemia/reperfusion. Apoptosis 11, 563–571 (2006).
10. P. R. Corridon, S. H. Karam, A. A. Khraibi, A. A. Khan, M. A. Alhashmi, Intravital imaging of
real-­time endogenous actin dysregulation in proximal and distal tubules at the onset of
severe ischemia-­reperfusion injury. Sci. Rep. 11, 8280 (2021).
11. K. Mukherjee, C. Gu, A. Collins, M. Mettlen, B. Samelko, M. M. Altintas, Y. R. Sudhini,
X. Wang, R. Bouley, D. Brown, B. P. Pedro, S. L. Bane, V. Gupta, P. T. Brinkkoetter,
H. Hagmann, J. Reiser, S. Sever, Simultaneous stabilization of actin cytoskeleton in
multiple nephron-­specific cells protects the kidney from diverse injury. Nat. Commun.
13, 2422 (2022).
12. E. E. Bosco, J. C. Mulloy, Y. Zheng, Rac1 GTPase: A “Rac” of all trades. Cell. Mol. Life Sci. 66,
370–374 (2009).
13. A. L. Berrier, R. Martinez, G. M. Bokoch, S. E. LaFlamme, The integrin beta tail is required
and sufficient to regulate adhesion signaling to Rac1. J. Cell Sci. 115, 4285–4291 (2002).
14. K. D. Liu, A. Datta, W. Yu, P. R. Brakeman, T. S. Jou, M. A. Matthay, K. E. Mostov, Rac1 is
required for reorientation of polarity and lumen formation through a PI 3-­kinase-­
dependent pathway. Am. J. Physiol. Renal Physiol. 293, F1633–F1640 (2007).
15. L. E. O’Brien, T. S. Jou, A. L. Pollack, Q. Zhang, S. H. Hansen, P. Yurchenco, K. E. Mostov,
Rac1 orientates epithelial apical polarity through effects on basolateral laminin
assembly. Nat. Cell Biol. 3, 831–838 (2001).
16. W. Yu, A. Datta, P. Leroy, L. E. O’Brien, G. Mak, T. S. Jou, K. S. Matlin, K. E. Mostov,
M. M. Zegers, Beta1-­integrin orients epithelial polarity via Rac1 and laminin. Mol. Biol. Cell
16, 433–445 (2005).
17. N. Ayuzawa, M. Nishimoto, W. Kawarazaki, S. Oba, T. Marumo, A. Aiba, T. Fujita, Rac1
deficiency impairs postnatal development of the renal papilla. Sci. Rep. 12, 20310 (2022).
18. F. Bock, B. C. Elias, X. Dong, D. V. Parekh, G. Mernaugh, O. M. Viquez, A. Hassan,
V. R. Amara, J. Liu, K. L. Brown, A. S. Terker, M. Chiusa, L. S. Gewin, A. B. Fogo,
C. H. Brakebusch, A. Pozzi, R. Zent, Rac1 promotes kidney collecting duct integrity by
limiting actomyosin activity. J. Cell Biol. 220, e202103080 (2021).
19. W. F. Liu, C. M. Nelson, D. M. Pirone, C. S. Chen, E-­cadherin engagement stimulates
proliferation via Rac1. J. Cell Biol. 173, 431–441 (2006).
20. N. Molinie, S. N. Rubtsova, A. Fokin, S. P. Visweshwaran, N. Rocques, A. Polesskaya,
A. Schnitzler, S. Vacher, E. V. Denisov, L. A. Tashireva, V. M. Perelmuter, N. V. Cherdyntseva,
I. Bieche, A. M. Gautreau, Cortical branched actin determines cell cycle progression. Cell
Res. 29, 432–445 (2019).
21. A. S. Mohan, K. M. Dean, T. Isogai, S. Y. Kasitinon, V. S. Murali, P. Roudot, A. Groisman,
D. K. Reed, E. S. Welf, S. J. Han, J. Noh, G. Danuser, Enhanced dendritic actin network
formation in extended lamellipodia drives proliferation in growth-­challenged Rac1(P29S)
melanoma cells. Dev. Cell 49, 444–460.e9 (2019).
22. Z. Wang, E. Pedersen, A. Basse, T. Lefever, K. Peyrollier, S. Kapoor, Q. Mei, R. Karlsson,
A. Chrostek-­Grashoff, C. Brakebusch, Rac1 is crucial for Ras-­dependent skin tumor
formation by controlling Pak1-­Mek-­Erk hyperactivation and hyperproliferation in vivo.
Oncogene 29, 3362–3373 (2010).
23. J. Liang, L. Oyang, S. Rao, Y. Han, X. Luo, P. Yi, J. Lin, L. Xia, J. Hu, S. Tan, L. Tang, Q. Pan,
Y. Tang, Y. Zhou, Q. Liao, Rac1, a potential target for tumor therapy. Front. Oncol. 11,
674426 (2021).
24. L. Li, D. Zepeda-­Orozco, V. Patel, P. Truong, C. M. Karner, T. J. Carroll, F. Lin, Aberrant planar
cell polarity induced by urinary tract obstruction. Am. J. Physiol. Renal Physiol. 297,
F1526–F1533 (2009).
25. A. Narvaez Barros, R. Guiteras, A. Sola, A. Manonelles, J. Morote, J. M. Cruzado, Reversal
unilateral ureteral obstruction: A mice experimental model. Nephron 142, 125–134 (2019).
26. A. L. Cochrane, M. M. Kett, C. S. Samuel, N. V. Campanale, W. P. Anderson, D. A. Hume,
M. H. Little, J. F. Bertram, S. D. Ricardo, Renal structural and functional repair in a mouse
model of reversal of ureteral obstruction. J. Am. Soc. Nephrol. 16, 3623–3630 (2005).
27. K. Rottner, J. Faix, S. Bogdan, S. Linder, E. Kerkhoff, Actin assembly mechanisms at a
glance. J. Cell Sci. 130, 3427–3435 (2017).
28. T. Saritas, V. G. Puelles, X. T. Su, J. A. McCormick, P. A. Welling, D. H. Ellison, Optical clearing in the
kidney reveals potassium-­mediated tubule remodeling. Cell Rep. 25, 2668–2675.e3 (2018).
29. L. Liu, H. Zhang, L. Shi, W. Zhang, J. Yuan, X. Chen, J. Liu, Y. Zhang, Z. Wang, Inhibition of
Rac1 activity induces G1/S phase arrest through the GSK3/cyclin D1 pathway in human
cancer cells. Oncol. Rep. 32, 1395–1400 (2014).
30. M. Manolopoulou, B. K. Matlock, S. Nlandu-­Khodo, A. J. Simmons, K. S. Lau,
M. Phillips-­Mignemi, A. Ivanova, C. E. Alford, D. K. Flaherty, L. S. Gewin, Novel kidney
dissociation protocol and image-­based flow cytometry facilitate improved analysis of
injured proximal tubules. Am. J. Physiol. Renal Physiol. 316, F847–F855 (2019).
31. S. Sivakumar, J. R. Daum, G. J. Gorbsky, Live-­cell fluorescence imaging for phenotypic
analysis of mitosis. Methods Mol. Biol. 1170, 549–562 (2014).
32. S. A. Gudipaty, J. Lindblom, P. D. Loftus, M. J. Redd, K. Edes, C. F. Davey, V. Krishnegowda,
J. Rosenblatt, Mechanical stretch triggers rapid epithelial cell division through Piezo1.
Nature 543, 118–121 (2017).
22 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
33. F. Cachat, B. Lange-­Sperandio, A. Y. Chang, S. C. Kiley, B. A. Thornhill, M. S. Forbes,
R. L. Chevalier, Ureteral obstruction in neonatal mice elicits segment-­specific tubular cell
responses leading to nephron loss. Kidney Int. 63, 564–575 (2003).
34. B. Lacroix, A. S. Maddox, Cytokinesis, ploidy and aneuploidy. J. Pathol. 226, 338–351 (2012).
35. J. Pampalona, C. Frias, A. Genesca, L. Tusell, Progressive telomere dysfunction causes
cytokinesis failure and leads to the accumulation of polyploid cells. PLOS Genet. 8,
e1002679 (2012).
36. G. Hognas, S. Tuomi, S. Veltel, E. Mattila, A. Murumagi, H. Edgren, O. Kallioniemi, J. Ivaska,
Cytokinesis failure due to derailed integrin traffic induces aneuploidy and oncogenic
transformation in vitro and in vivo. Oncogene 31, 3597–3606 (2012).
37. A. V. Taubenberger, B. Baum, H. K. Matthews, The mechanics of mitotic cell rounding.
Front. Cell Dev. Biol. 8, 687 (2020).
38. J. L. Monster, L. Donker, M. J. Vliem, Z. Win, H. K. Matthews, J. S. Cheah, S. Yamada,
J. de Rooij, B. Baum, M. Gloerich, An asymmetric junctional mechanoresponse coordinates
mitotic rounding with epithelial integrity. J. Cell Biol. 220, e202001042 (2021).
39. B. Sorce, C. Escobedo, Y. Toyoda, M. P. Stewart, C. J. Cattin, R. Newton, I. Banerjee,
A. Stettler, B. Roska, S. Eaton, A. A. Hyman, A. Hierlemann, D. J. Muller, Mitotic cells
contract actomyosin cortex and generate pressure to round against or escape epithelial
confinement. Nat. Commun. 6, 8872 (2015).
40. O. M. Lancaster, M. Le Berre, A. Dimitracopoulos, D. Bonazzi, E. Zlotek-­Zlotkiewicz,
R. Picone, T. Duke, M. Piel, B. Baum, Mitotic rounding alters cell geometry to ensure
efficient bipolar spindle formation. Dev. Cell 25, 270–283 (2013).
41. E. Hoijman, D. Rubbini, J. Colombelli, B. Alsina, Mitotic cell rounding and epithelial
thinning regulate lumen growth and shape. Nat. Commun. 6, 7355 (2015).
42. M. Novais-­Cruz, M. Alba Abad, W. F. van IJcken, N. Galjart, A. A. Jeyaprakash, H. Maiato,
C. Ferras, Mitotic progression, arrest, exit or death relies on centromere structural
integrity, rather than de novo transcription. eLife 7, e36898 (2018).
43. T. Svitkina, The actin cytoskeleton and actin-­based motility. Cold Spring Harb. Perspect.
Biol. 10, a018267 (2018).
44. C. G. Vasquez, A. C. Martin, Force transmission in epithelial tissues. Dev. Dyn. 245,
361–371 (2016).
45. J. Si, Y. Ge, S. Zhuang, R. Gong, Inhibiting nonmuscle myosin II impedes inflammatory
infiltration and ameliorates progressive renal disease. Lab. Invest. 90, 448–458 (2010).
46. L. K. Nguyen, B. N. Kholodenko, A. von Kriegsheim, Rac1 and RhoA: Networks, loops and
bistability. Small GTPases 9, 316–321 (2018).
47. J. L. MacKay, S. Kumar, Simultaneous and independent tuning of RhoA and Rac1 activity
with orthogonally inducible promoters. Integr. Biol. (Camb) 6, 885–894 (2014).
48. G. Chen, X. Deng, Cell synchronization by double thymidine block. Bio. Protoc. 8, e2994 (2018).
49. L. T. Vassilev, Cell cycle synchronization at the G2/M phase border by reversible inhibition
of CDK1. Cell Cycle 5, 2555–2556 (2006).
50. R. C. Wang, Z. Wang, Synchronization of cultured cells to G1, S, G2, and M phases by
double thymidine block. Methods Mol. Biol. 2579, 61–71 (2022).
51. P. Nalbant, Y. C. Chang, J. Birkenfeld, Z. F. Chang, G. M. Bokoch, Guanine nucleotide
exchange factor-­H1 regulates cell migration via localized activation of RhoA at the
leading edge. Mol. Biol. Cell 20, 4070–4082 (2009).
52. Y. C. Chang, P. Nalbant, J. Birkenfeld, Z. F. Chang, G. M. Bokoch, GEF-­H1 couples
nocodazole-­induced microtubule disassembly to cell contractility via RhoA. Mol. Biol. Cell
19, 2147–2153 (2008).
53. J. N. Heck, S. M. Ponik, M. G. Garcia-­Mendoza, C. A. Pehlke, D. R. Inman, K. W. Eliceiri,
P. J. Keely, Microtubules regulate GEF-­H1 in response to extracellular matrix stiffness. Mol.
Biol. Cell 23, 2583–2592 (2012).
54. M. Krendel, F. T. Zenke, G. M. Bokoch, Nucleotide exchange factor GEF-­H1 mediates
cross-­talk between microtubules and the actin cytoskeleton. Nat. Cell Biol. 4, 294–301
(2002).
55. J. Birkenfeld, P. Nalbant, B. P. Bohl, O. Pertz, K. M. Hahn, G. M. Bokoch, GEF-­H1 modulates
localized RhoA activation during cytokinesis under the control of mitotic kinases. Dev.
Cell 12, 699–712 (2007).
56. F. T. Zenke, M. Krendel, C. DerMardirossian, C. C. King, B. P. Bohl, G. M. Bokoch,
p21-­activated kinase 1 phosphorylates and regulates 14-­3-­3 binding to GEF-­H1, a
microtubule-­localized Rho exchange factor. J. Biol. Chem. 279, 18392–18400 (2004).
57. K. Leguay, B. Decelle, I. E. Elkholi, M. Bouvier, J. F. Cote, S. Carreno, Interphase microtubule
disassembly is a signaling cue that drives cell rounding at mitotic entry. J. Cell Biol. 221,
e202109065 (2022).
58. E. Joo, M. F. Olson, Regulation and functions of the RhoA regulatory guanine nucleotide
exchange factor GEF-­H1. Small GTPases 12, 358–371 (2021).
59. M. Barbiero, L. Cirillo, S. Veerapathiran, C. Coates, C. Ruffilli, J. Pines, Cell cycle-­dependent
binding between Cyclin B1 and Cdk1 revealed by time-­resolved fluorescence correlation
spectroscopy. Open Biol. 12, 220057 (2022).
60. N. P. Chen, J. Aretz, R. Fassler, CDK1-­cyclin-­B1-­induced kindlin degradation drives focal
adhesion disassembly at mitotic entry. Nat. Cell Biol. 24, 723–736 (2022).
61. X. Meng, L. L. Laidler, E. A. Kosmacek, S. Yang, Z. Xiong, D. Zhu, X. Wang, D. Dai, Y. Zhang,
X. Wang, P. Brachova, L. Albitar, D. Liu, F. Ianzini, M. A. Mackey, K. K. Leslie, Induction of
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
mitotic cell death by overriding G2/M checkpoint in endometrial cancer cells with
non-­functional p53. Gynecol. Oncol. 128, 461–469 (2013).
62. J. C. Martin, J. R. Sims, A. Gupta, A. V. Bakin, J. E. Ohm, WEE1 inhibition augments CDC7
(DDK) inhibitor-­induced cell death in Ewing sarcoma by forcing premature mitotic entry
and mitotic catastrophe. Cancer Res. Commun. 2, 471–482 (2022).
63. M. Aarts, R. Sharpe, I. Garcia-­Murillas, H. Gevensleben, M. S. Hurd, S. D. Shumway,
C. Toniatti, A. Ashworth, N. C. Turner, Forced mitotic entry of S-­phase cells as a
therapeutic strategy induced by inhibition of WEE1. Cancer Discov. 2, 524–539 (2012).
64. A. Ghelli, L. di Rorà, C. Cerchione, G. Martinelli, G. Simonetti, A WEE1 family business:
Regulation of mitosis, cancer progression, and therapeutic target. J. Hematol. Oncol. 13,
126 (2020).
65. L. Donker, R. Houtekamer, M. Vliem, F. Sipieter, H. Canever, M. Gomez-­Gonzalez,
M. Bosch-­Padros, W. J. Pannekoek, X. Trepat, N. Borghi, M. Gloerich, A mechanical G2
checkpoint controls epithelial cell division through E-­cadherin-­mediated regulation of
Wee1-­Cdk1. Cell Rep. 41, 111475 (2022).
66. M. Sasaki, T. Terabayashi, S. M. Weiss, I. Ferby, The tumor suppressor MIG6 controls mitotic
progression and the G2/M DNA damage checkpoint by stabilizing the WEE1 kinase. Cell
Rep. 24, 1278–1289 (2018).
67. C. Penas, V. Ramachandran, S. Simanski, C. Lee, F. Madoux, R. J. Rahaim, R. Chauhan,
O. Barnaby, S. Schurer, P. Hodder, J. Steen, W. R. Roush, N. G. Ayad, Casein kinase
1δ-­dependent Wee1 protein degradation. J. Biol. Chem. 289, 18893–18903 (2014).
68. H. Wu, R. Gonzalez Villalobos, X. Yao, D. Reilly, T. Chen, M. Rankin, E. Myshkin, M. D. Breyer,
B. D. Humphreys, Mapping the single-­cell transcriptomic response of murine diabetic
kidney disease to therapies. Cell Metab. 34, 1064–1078.e6 (2022).
69. J. Park, R. Shrestha, C. Qiu, A. Kondo, S. Huang, M. Werth, M. Li, J. Barasch, K. Susztak,
Single-­cell transcriptomics of the mouse kidney reveals potential cellular targets of
kidney disease. Science 360, 758–763 (2018).
70. A. B. Fogo, M. A. Lusco, N. K. Andeen, B. Najafian, C. E. Alpers, AJKD atlas of renal
pathology: Lithium nephrotoxicity. Am. J. Kidney Dis. 69, e1–e2 (2017).
71. E. A. Lynch, J. Stall, G. Schmidt, P. Chavrier, C. D’Souza-­Schorey, Proteasome-­mediated
degradation of Rac1-­GTP during epithelial cell scattering. Mol. Biol. Cell 17, 2236–2242 (2006).
72. A. Mettouchi, E. Lemichez, Ubiquitylation of active Rac1 by the E3 ubiquitin-­ligase
HACE1. Small GTPases 3, 102–106 (2012).
73. S. Singh, S. Machida, N. K. Tulsian, Y. K. Choong, J. Ng, S. Shankar, Y. Liu,
K. V. Chandiramani, J. Shi, J. Sivaraman, Structural basis for the enzymatic activity of the
HACE1 HECT-­type E3 ligase through N-­terminal helix dimerization. Adv. Sci. (Weinh) 10,
e2207672 (2023).
74. S. Torrino, O. Visvikis, A. Doye, L. Boyer, C. Stefani, P. Munro, J. Bertoglio, G. Gacon,
A. Mettouchi, E. Lemichez, The E3 ubiquitin-­ligase HACE1 catalyzes the ubiquitylation of
active Rac1. Dev. Cell 21, 959–965 (2011).
75. S. Petracchini, D. Hamaoui, A. Doye, A. Asnacios, F. Fage, E. Vitiello, M. Balland, S. Janel,
F. Lafont, M. Gupta, B. Ladoux, J. Gilleron, T. M. Maia, F. Impens, L. Gagnoux-­Palacios,
M. Daugaard, P. H. Sorensen, E. Lemichez, A. Mettouchi, Optineurin links Hace1-­
dependent Rac ubiquitylation to integrin-­mediated mechanotransduction to control
bacterial invasion and cell division. Nat. Commun. 13, 6059 (2022).
76. X. Gong, R. Nguyen, Z. Chen, Z. Wen, X. Zhang, M. Mak, Volumetric compression shifts
rho GTPase balance and induces mechanobiological cell state transition. bioRxiv
2023.10.08.561452 [Preprint] (2023). https://doi.org/10.1101/2023.10.08.561452.
77. C. F. Yeh, D. S. Juang, Y. W. Chen, D. Rodoplu, C. H. Hsu, A portable controllable
compressive stress device to monitor human breast cancer cell protrusions at single-­cell
resolution. Front. Bioeng. Biotechnol. 10, 852318 (2022).
78. G. Cheng, J. Tse, R. K. Jain, L. L. Munn, Micro-­environmental mechanical stress controls
tumor spheroid size and morphology by suppressing proliferation and inducing
apoptosis in cancer cells. PLOS ONE 4, e4632 (2009).
79. J. M. Tse, G. Cheng, J. A. Tyrrell, S. A. Wilcox-­Adelman, Y. Boucher, R. K. Jain, L. L. Munn,
Mechanical compression drives cancer cells toward invasive phenotype. Proc. Natl. Acad.
Sci. U.S.A. 109, 911–916 (2012).
80. S. M. Blattner, J. B. Hodgin, M. Nishio, S. A. Wylie, J. Saha, A. A. Soofi, C. Vining,
A. Randolph, N. Herbach, R. Wanke, K. B. Atkins, H. Gyung Kang, A. Henger, C. Brakebusch,
L. B. Holzman, M. Kretzler, Divergent functions of the Rho GTPases Rac1 and Cdc42 in
podocyte injury. Kidney Int. 84, 920–930 (2013).
81. X. Wan, M. S. Lee, W. Zhou, Dosage-­dependent role of Rac1 in podocyte injury. Am. J.
Physiol. Renal Physiol. 310, F777–F784 (2016).
82. R. Robins, C. Baldwin, L. Aoudjit, J. F. Cote, I. R. Gupta, T. Takano, Rac1 activation in
podocytes induces the spectrum of nephrotic syndrome. Kidney Int. 92, 349–364 (2017).
83. H. Yu, H. Suleiman, A. H. Kim, J. H. Miner, A. Dani, A. S. Shaw, S. Akilesh, Rac1 activation in
podocytes induces rapid foot process effacement and proteinuria. Mol. Cell. Biol. 33,
4755–4764 (2013).
84. R. Asao, T. Seki, M. Takagi, H. Yamada, F. Kodama, Y. Hosoe-­Nagai, E. Tanaka, J. A. O. Trejo,
K. Yamamoto-­Nonaka, Y. Sasaki, T. Hidaka, T. Ueno, M. Yanagita, Y. Suzuki, Y. Tomino,
K. Asanuma, Rac1 in podocytes promotes glomerular repair and limits the formation of
sclerosis. Sci. Rep. 8, 5061 (2018).
23 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
85. V. Thamilselvan, M. Menon, S. Thamilselvan, Selective Rac1 inhibition protects renal
tubular epithelial cells from oxalate-­induced NADPH oxidase-­mediated oxidative cell
injury. Urol. Res. 40, 415–423 (2012).
86. A. Caron, R. R. Desrosiers, R. Beliveau, Kidney ischemia-­reperfusion regulates expression
and distribution of tubulin subunits, β-­actin and rho GTPases in proximal tubules. Arch.
Biochem. Biophys. 431, 31–46 (2004).
87. N. Akhtar, W. Li, A. Mironov, C. H. Streuli, Rac1 controls both the secretory function of the
mammary gland and its remodeling for successive gestations. Dev. Cell 38, 522–535 (2016).
88. G. A. Gusarova, S. R. Das, M. N. Islam, K. Westphalen, G. Jin, I. O. Shmarakov, L. Li,
S. Bhattacharya, J. Bhattacharya, Actin fence therapy with exogenous V12Rac1 protects
against acute lung injury. JCI Insight 6, e135753 (2021).
89. L. Lasagni, E. Lazzeri, S. J. Shankland, H. J. Anders, P. Romagnani, Podocyte mitosis -­A
catastrophe. Curr. Mol. Med. 13, 13–23 (2013).
90. L. Yang, T. Y. Besschetnova, C. R. Brooks, J. V. Shah, J. V. Bonventre, Epithelial cell cycle
arrest in G2/M mediates kidney fibrosis after injury. Nat. Med. 16, 535–543 (2010).
91. D. P. DiRocco, J. Bisi, P. Roberts, J. Strum, K. K. Wong, N. Sharpless, B. D. Humphreys,
CDK4/6 inhibition induces epithelial cell cycle arrest and ameliorates acute kidney injury.
Am. J. Physiol. Renal Physiol. 306, F379–F388 (2014).
92. Y. Osaki, M. Manolopoulou, A. V. Ivanova, N. Vartanian, M. P. Mignemi, J. Kern, J. Chen,
H. Yang, A. B. Fogo, M. Zhang, C. Robinson-­Cohen, L. S. Gewin, Blocking cell cycle
progression through CDK4/6 protects against chronic kidney disease. JCI Insight 7,
e158754 (2022).
93. K. Kozyrska, G. Pilia, M. Vishwakarma, L. Wagstaff, M. Goschorska, S. Cirillo, S. Mohamad,
K. Gallacher, R. E. Carazo Salas, E. Piddini, p53 directs leader cell behavior, migration, and
clearance during epithelial repair. Science 375, eabl8876 (2022).
94. A. Cosolo, J. Jaiswal, G. Csordas, I. Grass, M. Uhlirova, A. K. Classen, JNK-­dependent cell
cycle stalling in G2 promotes survival and senescence-­like phenotypes in tissue stress.
eLife 8, e41036 (2019).
95. M. S. Murakami, S. A. Moody, I. O. Daar, D. K. Morrison, Morphogenesis during Xenopus
gastrulation requires Wee1-­mediated inhibition of cell proliferation. Development 131,
571–580 (2004).
96. A. Vassilopoulos, Y. Tominaga, H. S. Kim, T. Lahusen, B. Li, H. Yu, D. Gius, C. X. Deng, WEE1
murine deficiency induces hyper-­activation of APC/C and results in genomic instability
and carcinogenesis. Oncogene 34, 3023–3035 (2015).
97. J. N. McMillan, R. A. Sia, D. J. Lew, A morphogenesis checkpoint monitors the actin
cytoskeleton in yeast. J. Cell Biol. 142, 1487–1499 (1998).
98. D. R. Kellogg, Wee1-­dependent mechanisms required for coordination of cell growth and
cell division. J. Cell Sci. 116, 4883–4890 (2003).
99. M. Chircop, Rho GTPases as regulators of mitosis and cytokinesis in mammalian cells.
Small GTPases 5, e29770 (2014).
100. R. N. Bastos, X. Penate, M. Bates, D. Hammond, F. A. Barr, CYK4 inhibits Rac1-­dependent
PAK1 and ARHGEF7 effector pathways during cytokinesis. J. Cell Biol. 198, 865–880 (2012).
101. K. A. Anton, M. Kajita, R. Narumi, Y. Fujita, M. Tada, Src-­transformed cells hijack mitosis to
extrude from the epithelium. Nat. Commun. 9, 4695 (2018).
102. J. M. Vasiliev, T. Omelchenko, I. M. Gelfand, H. H. Feder, E. M. Bonder, Rho overexpression
leads to mitosis-­associated detachment of cells from epithelial sheets: A link to the
mechanism of cancer dissemination. Proc. Natl. Acad. Sci. U.S.A. 101, 12526–12530 (2004).
103. L. Wagstaff, M. Goschorska, K. Kozyrska, G. Duclos, I. Kucinski, A. Chessel,
L. Hampton-­O’Neil, C. R. Bradshaw, G. E. Allen, E. L. Rawlins, P. Silberzan, R. E. Carazo Salas,
E. Piddini, Mechanical cell competition kills cells via induction of lethal p53 levels. Nat.
Commun. 7, 11373 (2016).
104. M. C. Aublette, T. A. Harrison, E. J. Thorpe, M. S. Gadd, Selective Wee1 degradation by
PROTAC degraders recruiting VHL and CRBN E3 ubiquitin ligases. Bioorg. Med. Chem. Lett.
64, 128636 (2022).
105. F. Dang, L. Nie, W. Wei, Ubiquitin signaling in cell cycle control and tumorigenesis. Cell
Death Differ. 28, 427–438 (2021).
106. H. Wu, Y. Yu, H. Huang, Y. Hu, S. Fu, Z. Wang, M. Shi, X. Zhao, J. Yuan, J. Li, X. Yang, E. Bin,
D. Wei, H. Zhang, J. Zhang, C. Yang, T. Cai, H. Dai, J. Chen, N. Tang, Progressive pulmonary
fibrosis is caused by elevated mechanical tension on alveolar stem cells. Cell 180,
107–121.e17 (2020).
107. R. Tassinari, E. Olivi, C. Cavallini, V. Taglioli, C. Zannini, M. Marcuzzi, O. Fedchenko,
C. Ventura, Mechanobiology: A landscape for reinterpreting stem cell heterogeneity and
regenerative potential in diseased tissues. iScience 26, 105875 (2023).
108. S. M. Naqvi, L. M. McNamara, Stem cell mechanobiology and the role of biomaterials in
governing mechanotransduction and matrix production for tissue regeneration. Front.
Bioeng. Biotechnol. 8, 597661 (2020).
109. A. Chrostek, X. Wu, F. Quondamatteo, R. Hu, A. Sanecka, C. Niemann, L. Langbein,
I. Haase, C. Brakebusch, Rac1 is crucial for hair follicle integrity but is not essential for
maintenance of the epidermis. Mol. Cell. Biol. 26, 6957–6970 (2006).
110. B. Zhao, Z. Qi, Y. Li, C. Wang, W. Fu, Y. G. Chen, The non-­muscle-­myosin-­II heavy chain
Myh9 mediates colitis-­induced epithelium injury by restricting Lgr5+ stem cells. Nat.
Commun. 6, 7166 (2015).
Bock et al., Sci. Adv. 10, eadi7840 (2024) 7 February 2024
111. J. Zhang, Z. Pan, J. Zhou, L. Zhang, J. Tang, S. Gong, F. Li, B. Yu, Y. Zhang, J. Kou, The myosin
II inhibitor, blebbistatin, ameliorates pulmonary endothelial barrier dysfunction in acute
lung injury induced by LPS via NMMHC IIA/Wnt5a/β-­catenin pathway. Toxicol. Appl.
Pharmacol. 450, 116132 (2022).
112. Y. Zhang, L. Li, Y. Zhao, H. Han, Y. Hu, D. Liang, B. Yu, J. Kou, The myosin II inhibitor,
blebbistatin, ameliorates FeCl3-­induced arterial thrombosis via the GSK3β-­NF-­κB
pathway. Int. J. Biol. Sci. 13, 630–639 (2017).
113. D. J. Jafree, D. A. Long, P. J. Scambler, D. Moulding, Tissue clearing and deep imaging of
the kidney using confocal and two-­photon microscopy. Methods Mol. Biol. 2067,
103–126 (2020).
114. D. El-­Nachef, A. M. Martinson, X. Yang, C. E. Murry, W. R. MacLellan, High-­resolution 3D
fluorescent imaging of intact tissues. Int. J. Cardiol. Cardiovasc. Dis. 1, 1–14 (2021).
115. R. Loperena, J. P. Van Beusecum, H. A. Itani, N. Engel, F. Laroumanie, L. Xiao, F. Elijovich,
C. L. Laffer, J. S. Gnecco, J. Noonan, P. Maffia, B. Jasiewicz-­Honkisz,
M. Czesnikiewicz-­Guzik, T. Mikolajczyk, T. Sliwa, S. Dikalov, C. M. Weyand, T. J. Guzik,
D. G. Harrison, Hypertension and increased endothelial mechanical stretch promote
monocyte differentiation and activation: Roles of STAT3, interleukin 6 and hydrogen
peroxide. Cardiovasc. Res. 114, 1547–1563 (2018).
116. Y. S. Choo, Z. Zhang, Detection of protein ubiquitination. J. Vis. Exp., 1293 (2009).
117. A. Chronopoulos, S. D. Thorpe, E. Cortes, D. Lachowski, A. J. Rice, V. V. Mykuliak, T. Rog,
D. A. Lee, V. P. Hytonen, A. E. Del Rio Hernandez, Syndecan-­4 tunes cell mechanics by
activating the kindlin-­integrin-­RhoA pathway. Nat. Mater. 19, 669–678 (2020).
118. R. H. Giles, H. Ajzenberg, P. K. Jackson, 3D spheroid model of mIMCD3 cells for studying
ciliopathies and renal epithelial disorders. Nat. Protoc. 9, 2725–2731 (2014).
119. A. Zancla, P. Mozetic, M. Orsini, G. Forte, A. Rainer, A primer to traction force microscopy.
J. Biol. Chem. 298, 101867 (2022).
120. J. L. Martiel, A. Leal, L. Kurzawa, M. Balland, I. Wang, T. Vignaud, Q. Tseng, M. Thery,
Measurement of cell traction forces with ImageJ. Methods Cell Biol. 125, 269–287 (2015).
121. J. L. Teo, C. T. Lim, A. S. Yap, T. B. Saw, A biologist’s guide to traction force microscopy
using polydimethylsiloxane substrate for two-­dimensional cell cultures. STAR Protoc. 1,
100098 (2020).
122. L. C. Crowley, G. Chojnowski, N. J. Waterhouse, Measuring the DNA content of cells in
apoptosis and at different cell-­cycle stages by propidium iodide staining and flow
cytometry. Cold Spring Harb. Protoc. 2016, doi: 10.1101/pdb.prot087247 (2016).
123. E. Aihara, N. M. Medina-­Candelaria, H. Hanyu, A. L. Matthis, K. A. Engevik, C. B. Gurniak,
W. Witke, J. R. Turner, T. Zhang, M. H. Montrose, Cell injury triggers actin polymerization
initiating epithelial restitution. J. Cell Sci. 131, jcs216317 (2018).
Acknowledgments: We are grateful to the cell imaging shared resource (CISR), our flow
cytometry cores at Vanderbilt and the Nashville VA, and our pathology core (Translational
Pathology Shared Resource), and we would like to thank J. Schafer, C. Warren, B. Matlock, and
D. Flaherty for technical assistance. All scientific illustrations including the graphical abstract
were created with Biorender.com [Figs. 2 (A and J), 4C, 6K, and 11C; and figs. S5A, S10A, S30A,
and S31). Funding: These studies were supported by NIH grants K08 DK134879 (to F.B.),
DK069921 (to R.Z.), DK088327 (to R.Z.), DK127589 (to R.Z.), R01 HL163195 (to E.P.), K08
DK135931-­01 (to J.P.A.), DP5OD033412 (to A.S.T.), R01 DK119212 (to A.P.), P30 DK114809 (to
A.P.), and R01 DK056942 (to A.F.); VA Merit awards I01-­BX002196 (to R.Z.) and 1I01BX002025 (to
A.P.); American Society of Nephrology Kidney Cure Ben J. Lipps Fellowship (to F.B.); and a
Kidney Cure Pre-­Doctoral Fellowship (to X.D.), Vanderbilt Faculty Research Scholar Award (to
F.B.). Microscopy was performed using the Vanderbilt Cell Imaging Shared Resource
(supported by NIH grant P30-­CA068485 and the Department of Veteran Affairs), Flow
cytometry experiments were performed in the Nashville VA flow cytometry core (U.S.
Department of Veterans Affairs, Tennessee Valley Healthcare System, Nashville, TN) and the
VMC Flow Cytometry Shared Resource (supported by the Vanderbilt Ingram Cancer Center,
P30-­CA68485, and the Vanderbilt Digestive Disease Research Center, DK058404). Routine
histological processing was performed in our Translational Pathology Shared Resource that is
supported by NCI/NIH Cancer Center support grant P30-­CA068485. Author contributions:
Conceptualization: F.B., E.M.H., E.P., A.S.T., J.P.A., A.P., and R.Z. Data curation: F.B. Formal analysis:
F.B., X.D., E.S., and K.L.B. Funding acquisition: F.B, D.G.H., and R.Z. Investigation: F.B., X.D, E.S.,
M.T., E.M.H., J.P.A., D.G.H., and R.Z. Methodology: F.B., X.D., O.M.V., E.M.H., J.P.A., D.G.H., and R.Z.
Project administration: F.B., M.T., D.G.H., and R.Z. Resources: D.G.H., Y.M.W., E.M.H., C.H.B., R.Z.,
A.R., J.P.A., F.B., and A.F. Software: E.S., E.M.H., A.R., and K.L.B. Supervision: F.B., D.G.H., and R.Z.
Validation: F.B., X.D., S.L., E.P., D.G.H., and R.Z. Visualization: F.B., X.D., K.L.B., D.G.H., and R.Z.
Writing (original draft): F.B. and R.Z. Writing (review and editing): F.B., E.S., J.P.A., and R.Z.
Competing interests: The authors declare that they have no competing interests. Data and
materials availability: All data needed to evaluate the conclusions in the paper are present in
the paper and/or the Supplementary Materials.
Submitted 17 May 2023
Accepted 3 January 2024
Published 7 February 2024
10.1126/sciadv.adi7840
24 of 24
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

DISEASES AND DISORDERS Copyright © 2024 The

Authors, some rights
NU6300 covalently reacts with cysteine-­191 of reserved; exclusive
licensee American
gasdermin D to block its cleavage and palmitoylation Association for the
Advancement of
Science. No claim to
Xueqin Jiang1†, Xinlu Zhang1†, Xiaoying Cai1†, Na Li1†, Hongyu Zheng2, Minghai Tang1, original U.S.
Jiangli Zhu3, Kaiyue Su1, Ruijia Zhang1, Neng Ye1, Jing Peng1, Min Zhao4, Wenshuang Wu5*, Government Works.
Jianhong Yang1*, Haoyu Ye1* Distributed under a
Creative Commons
Gasdermin D (GSDMD) serves as a vital mediator of inflammasome-­driven pyroptosis. In our study, we have iden- Attribution
tified NU6300 as a specific GSDMD inhibitor that covalently interacts with cysteine-­191 of GSDMD, effectively NonCommercial
blocking its cleavage while not affecting earlier steps such as ASC oligomerization and caspase-­1 processing in License 4.0 (CC BY-­NC).
AIM2-­and NLRC4-­mediated inflammation. On the contrary, NU6300 robustly inhibits these earlier steps in NLRP3
inflammasome, confirming a unique feedback inhibition effect in the NLRP3-­GSDMD pathway upon GSDMD tar-
geting. Our study reveals a previously undefined mechanism of GSDMD inhibitors: NU6300 impairs the palmi-
toylation of both full-­length and N-­terminal GSDMD, impeding the membrane localization and oligomerization of
N-­terminal GSDMD. In vivo studies further demonstrate the efficacy of NU6300 in ameliorating dextran sodium
sulfate–induced colitis and improving survival in lipopolysaccharide-­induced sepsis. Overall, these findings high-
light the potential of NU6300 as a promising lead compound for the treatment of inflammatory diseases.

INTRODUCTION GSDMD acts as an effector of inflammasome signaling and is impli-

Gasdermin D (GSDMD) is an essential pyroptosis executioner and
a member of the gasdermin family, which includes GSDMA, GSDMB,
GSDMC, GSDMD, GSDME, and DFNB59 (1, 2). It consists of two
domains, the 31-­kDa N terminus (GSDMD-­N) and the 22-­kDa C
terminus (GSDMD-­C), separated by a linker region (3). The first
flexible loop of GSDMD-­C, located between GSDMD-­N and the
linker helix, stretches out and inserts into the GSDMD-­N pocket,
stabilizing the conformation of the full-­length protein (4). Upon
activation, inflammatory caspases cleave GSDMD, producing and
releasing GSDMD-­N and GSDMD-­C (3). Meanwhile, full-­length
GSDMD or N-­terminal GSDMD is palmitoylated at cysteine-­191
(C191) upon activation (5, 6). Then, palmitoylated GSDMD-­N in-
teracts with acidic phospholipids in the inner leaflet of the plasma
membrane, forming functional pores with an inner diameter of 10
to 15 nm through oligomerization (3, 7–9). Excessive pore forma-
tion leads to compromised plasma membrane integrity, cellular
swelling, secretion of interleukin-­1β (IL-­1β) and IL-­18, and, ulti-
mately, pyroptotic cell death (10).
The critical role of GSDMD as a pyroptosis executioner has made
it a prominent topic in the field of immunology, which has been
linked to immune defense and numerous diseases (11). Studies have
shown that GSDMD deletion or inhibition notably decreases pyrop-
tosis and remarkably protect mice from sepsis, which is a fatal
condition requiring urgent medical solutions (12–14). In addition,
1
Department of Biotherapy, Cancer Center and State Key Laboratory of Biotherapy,
West China Hospital, Sichuan University, Chengdu 610041, China. 2School of Phar-
macy, Chengdu Medical College, Chengdu 610500, China. 3Department of Urology,
Cancer Center and State Key Laboratory of Biotherapy, West China Hospital, Sichuan
University, Chengdu 610041, China. 4Laboratory of Metabolomics and Drug-­induced
Liver Injury, Department of Gastroenterology and Hepatology, Frontiers Science
Center for Disease-­related Molecular Network, West China Hospital, Sichuan Uni-
versity, Chengdu 610041, China. 5Division of Thyroid Surgery, Department of General
Surgery and Laboratory of Thyroid and Parathyroid Disease, Frontiers Science
Center for Disease-­related Molecular Network, West China Hospital, Sichuan Univer-
sity, Chengdu 610041, China.
*Corresponding author. Email: wenshuang_wu@​163.​com (W.W.); yjh1988@​scu.​
edu.​cn (J.Y.); haoyu_ye@​scu.​edu.​cn (H.Y.)
†These authors contributed equally to this work.
cated in several human diseases, including nonalcoholic steatohepa-
titis (15), cardiovascular disease (16), inflammatory bowel disease
(17), type II diabetes (17, 18), rheumatoid arthritis (19), cancer (20,
21), and Alzheimer’s disease (22). Consequently, on account of the
crucial function of GSDMD in pyroptosis, it has emerged as an ideal
and attracting target for therapeutic intervention of inflammasome-­
driven diseases.
Since GSDMD is a central effector protein in regulating inflam-
matory diseases (23, 24), there has been substantial interest in devel-
oping drugs that directly target GSDMD. Such drugs not only
provide critical insights for exploring and establishing previously
unknown mechanism for the study of pyroptosis but also offer a
promising avenue for clinical studies in various inflammatory dis-
eases. Up to now, three covalent small molecules have been reported
to directly target GSDMD: disulfiram (Dis) (14), necrosulfonamide
(NSA) (25), and dimethyl fumarate (DMF) (26, 27). These small
molecules react with the free thiol group at cysteine-­191/192
(C191/192) in GSDMD, thereby blocking pore formation and py-
roptosis (14, 25, 26).
In this study, we conducted a screening of small molecules to
identify inhibitors of pyroptotic cell death based on lactate dehydro-
genase (LDH) release. Intriguingly, we found that NU6300 was a
potent inhibitor of pyroptotic cell death, and it could reduce cyto-
kine release and propidium iodide (PI) intake in human monocytes
and murine macrophages. Mechanistically, NU6300 covalently
modified human C191 in GSDMD to block its cleavage but had no
effect on the earlier steps such as apoptosis-­associated speck-­like
protein containing a caspase recruitment domain (ASC) oligomer-
ization and caspase-­1 processing in AIM2 and NLRC4 inflamma-
somes, proving its high selectivity on GSDMD. On the contrary, we
observed robust inhibition of NU6300 on earlier steps of NLRP3
inflammasome, implying a feedback inhibition effect on NLRP3 in-
flammasome when GSDMD was targeted. We found that NU6300
impaired the palmitoylation of both full-­length and N-­terminal
GSDMD, resulting in the subsequent impediment of GSDMD-­N
membrane localization. Moreover, we demonstrated that NU6300

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 1 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
administration attenuated the release of proinflammatory cytokines
and provided protection in mice models of colitis and sepsis. Col-
lectively, our results identified NU6300 as a pyroptotic inhibitor that
directly targeted GSDMD and held great promise as a potential
intervention for inflammatory diseases.
RESULTS
Identification of NU6300 as a potent inhibitor of pyroptosis
Pyroptosis, an inflammatory form of programmed cell death trig-
gered by inflammatory stimuli, has emerged as a promising thera-
peutic strategy for inflammatory diseases (28). In our research, we
performed a screening of covalent compound library containing
565 compounds for inhibitors of pyroptotic cell death based on
LDH release on lipopolysaccharide (LPS) plus nigericin-­induced
pyroptosis in THP-­1 cells (Fig. 1A and data S1). Among the active
compounds, NU6300 (Fig. 1, B and C) exhibited a dose-­dependent
rescue effect on pyroptosis, with half maximal inhibitory concentra-
tion (IC50) values of 0.89 and 0.93 μM in THP-­1 cells and bone
marrow–derived macrophages (BMDMs), respectively (fig. S1A).
Furthermore, NU6300 blocked cell death driven by AIM2 or NLRC4
inflammasomes (Fig. 1, D and E). Apart from canonical inflamma-
somes, NU6300 also inhibited the noncanonical inflammasome-­
mediated pyroptosis in THP-­1 cells and BMDMs (Fig. 1F). NU6300
showed better inhibitory effects on pyroptosis in both THP-­1 cells
and BMDMs compared with known GSDMD inhibitors such as Dis
and NSA, as confirmed by LDH analysis (Fig. 1G and fig. S1B). In
addition, we verified that NU6300 displayed no effect on necroptosis
(Fig. 1H). The pyroptosis inhibition effect of NU6300 was further
validated by PI uptake assay (Fig. 1I and fig. S1C) and live cell imag-
ing of THP-­1 cells (fig. S1D). Meanwhile, transmission electron mi-
croscopy analysis revealed that NU6300 mitigated the characteristic
cytoplasmic swelling and plasma membrane rupture associated with
pyroptotic cell death (Fig. 1J). Collectively, these findings suggested
that NU6300 effectively inhibited pyroptosis triggered by canonical
or noncanonical inflammasomes in both human and mouse cells.
Previous studies indicated that NU6300 was a covalent cyclin-­
dependent kinase 2 (CDK2) inhibitor (29). Thus, we further tested
whether NU6300 inhibited pyroptosis through targeting CDK2. We
found that another CDK2 inhibitor, CVT-­313, did not successfully
inhibit pyroptotic cell death induced by nigericin in THP-­1 cells
(fig. S2A). We then used lentiviral short hairpin RNA (shRNA) to
silence CDK2 expression in THP-­1 cells (fig. S2, B and C) and
observed no pyroptosis inhibition effect in sh-­​CDK2 THP-­1 cells
after nigericin treatment (fig. S2D). These results indicated that the
rescue effect of NU6300 might be attributed to off-­target effects
rather than direct inhibition of CDK2 activity.
NU6300 directly interacts with GSDMD
To investigate the potential target of NU6300 in pyroptosis, we per-
formed the drug affinity responsive target stability (DARTS) assay
using protein lysates from THP-­1 cells. We observed a prominent
protected band at around 50 kDa on the Coomassie brilliant blue
staining gel in proteolyzed extracts of cells treated with NU6300,
indicating a potential interaction with a protein of interest (Fig. 2A).
Using shotgun proteomics based on peptide mass fingerprinting
and tandem mass spectrometry (MS/MS), we identified this band as
the pyroptosis-­related protein GSDMD (table S1). To validate this
interaction, we performed immunoblot on DARTS samples in
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
preferred targets (CAP1, NAPRT, TCP1, UMPS, and GSDMD) in
biological MS assay, confirming that NU6300 effectively protected
GSDMD from protease hydrolyzation (Fig. 2, B to G, and fig. S3A).
We observed a similar binding pattern with purified GSDMD
(Fig. 2H and fig. S3, B and C). In addition, cellular thermal shift
assay (CETSA) and thermal shift assay (TSA) demonstrated that
NU6300 markedly and dose-­dependently enhanced the thermal
stability of GSDMD protein in both THP-­1 cell lysates and purified
GSDMD protein accompanying increasing temperature (Fig. 2, I
and J). Biolayer interferometry (BLI) analysis revealed a specific
binding dissociation constant (KD) value of 29.1 μM for NU6300 to
GSDMD with Kon and Kdis rates of 1.51 × 103 (1/ms) and 4.39 × 10−2
(1/s) (Fig. 2K), respectively. Overall, these results confirmed that
GSDMD was the direct target of NU6300 in pyroptosis.
NU6300 covalently reacts with C191 of GSDMD
A previous report indicated that NU6300 covalently reacted with
Lys89 of CDK2 through its vinyl sulfone moiety (29). Given that vinyl
sulfone is an emerging electrophile capable of covalently reacting
with residues such as cysteine or lysine (30, 31), we hypothesized
that NU6300 could react covalently with GSDMD. We conducted
liquid chromatography MS/MS (LC-­MS/MS) to investigate the ef-
fect of NU6300 on GSDMD by incubating recombinant human
GSDMD protein with NU6300, revealing a modification of C191
(Fig. 3A), which was important for GSDMD oligomerization. In
addition, we detected C56 and C268 modification (fig. S4A). Fur-
thermore, a competitive binding assay demonstrated that preincu-
bation with N-­ acetylcysteine (NAC), a highly reactive cysteine
residue that can be inactivated by thiol-­reactive compounds (32),
notably reversed the inhibitory effects of NU6300 on nigericin-­
induced pyroptosis in THP-­1 cells (Fig. 3B), confirming that NU6300
targeted reactive cysteine. As these three cysteines are all located
in the N terminus of GSDMD (p30), which can trigger pyroptosis
alone, we expressed N-­terminal (p30) in human embryonic kidney
(HEK)–293T cells for further validation (fig. S4B). Cytotoxicity
analysis revealed that NU6300 concentration-­dependently amelio-
rated the pyroptotic cell death triggered by expression of the p30
fragment in HEK-­293T cells (fig. S4C). In addition, PI staining
demonstrated that treatment with NU6300 led to a notable reduc-
tion in cell death in p30-­transfected cells (fig. S4D). We conducted
further investigation by introducing cysteine-­to-­alanine mutations
in several conserved human cysteines (C38, C56, C191, and C268)
within the p30 fragment. Consistent with the previous report (25), it
was observed that C191A mutation on the p30 fragment signifi-
cantly reduced the cell death compared to the vehicle or full-­length
GSDMD, while C38A, C56A, and C268A mutations had no effect
on cytotoxicity (Fig. 3C). In HEK-­293T cells expressing p30, p30-­
C38A, p30-­C56A, p30-­C268A, and p30-­C38A/C56A/C268A frag-
ments, NU6300 reduced pyroptotic cell death, while it showed no
notable reduction in cell death in p30-­ C191A and p30-­ C38A/
C56A/191A/C268A fragments (Fig. 3, C and D). Furthermore, mi-
croscale thermophoresis (MST) assay further revealed that NU6300
had no direct binding with GSDMD C191A (fig. S4E). Collectively,
C191 is the important site for covalent modification, and the bind-
ing of NU6300 to C56 and C268 is nonspecific and does not affect
the function of GSDMD. To further verify the binding of NU6300 to
C191 and its pyroptosis inhibition, we synthesized compound NU2,
which lacked the vinyl group present in the structure of NU6300
(fig. S5, A and B). Compared to NU2, NU6300 and GSDMD inhibitor
2 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. NU6300 inhibits pyroptosis. (A) Experimental schematic of screening the active small molecules using LDH analysis. (B) THP-­1 cells were stimulated with LPS
(1 μg/ml) for 3 hours and then treated with different compounds (2 μM) for 40 min before they were induced with nigericin (10 μM) for 35 min, and the inhibition rate of
cell death was determined by LDH analysis with a cutoff value of 75% inhibition. (C) Chemical structure of NU6300. (D and E) THP-­1 cells and BMDMs were primed with
LPS and incubated with NU6300 before stimulation with poly(deoxyadenylic-­deoxythymidylic) acid [poly(dA:dT)] (500 ng/ml) (D) or flagellin (250 ng/ml) (E) for 6 hours,
followed by LDH analysis of cell death. (F) THP-­1 cells and BMDMs were treated with Pam3CSK4 (400 ng/ml) for 3 hours and incubated with NU6300 before stimulation
with cytosolic LPS (1.5 μg/ml) overnight, followed by LDH analysis of cell death. (G) THP-­1 cells and BMDMs were primed with LPS and incubated with NU6300 (2 μM), NSA
(10 μM), or Dis (20 μM) before nigericin induction, followed by LDH analysis of cell death. (H) HT-­29 cells were pretreated with or without NU6300 (2 μM) or NSA (10 μM)
for 1 hour before stimulation with TNFα (25 ng/ml) (T), 400 nM SMAC mimetic (S), and 20 μM z-­VAD-­fmk (Z) for 24 hours, and cell viability was analyzed by CCK8 assay.
(I) Kinetic analysis of PI uptake and cellular membrane permeability after treatment with NU6300 in THP-­1 and BMDMs. (J) Transmission electron microscope observation
of THP-­1 cells morphology. Red arrows indicate representative organelles. Scale bars, 1 μm. Graphs showed means ± SEM, n = 3. Statistics were analyzed by one-­way
analysis of variance (ANOVA). *P < 0.05, **P < 0.01, and ***P < 0.001. ns, not significant.

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 3 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. NU6300 directly interacts with GSDMD. (A) DARTS assay in LPS-­primed THP-­1 cells treated with dimethyl sulfoxide (DMSO) or NU6300 (1000, 100, and 10 μM). The
lysate samples were digested with pronase (1:500, pronase-­to-­protein mass ratio) and analyzed by SDS–polyacrylamide gel electrophoresis and Coomassie brilliant blue
staining. (B to E) Immunoblot assay to detect the CAP1 (B), NAPRT (C), TCP1 (D), and UMPS (E) proteins by DARTS analysis in LPS-­primed THP-­1 cells. (F to H) Immunoblot
assay to GSDMD protein in THP-­1 cells (F), BMDMs (G), and purified protein (H), and the relative density of GSDMD protein was obtained by normalization to the control
group. (I and J) CETSA in LPS-­primed THP-­1 cell lysates (I) or TSA (J) in purified GSDMD protein. The samples were incubated with NU6300 (10 and 20 μM) at various tem-
peratures, the relative density of GSDMD protein was normalized by the sample without heating, and ΔT1 and ΔT2 indicated the thermal shift of NU6300 (10 and 20 μM)
as compared to DMSO control. (K) The binding of GSDMD with NU6300 was evaluated by BLI analysis, and the equilibrium binding signal (Req) was plotted against concen-
tration of analyte. Data were presented as means ± SEM, n = 3. One-­way ANOVA or two-­way ANOVA was used. *P < 0.05, **P < 0.01, and ***P < 0.001. ns, not significant.

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 4 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. NU6300 directly binds to C191 of GSDMD. (A) MS/MS spectra of the corresponding human GSDMD peptide FSLPGATCLQGEGQGHLSQK modified on C191 after GSDMD
incubation with NU6300. (B) NU6300 (0.5, 1, and 2 μM) and NAC (500 μM) were preincubated for 1 hour before stimulation with nigericin (10 μM) from LPS-­primed THP-­1 cells for
LDH assay. (C) HEK-­293T cells were transfected with either GSDMD, p30, p30-­C38A, p30-­C56A, p30-­C191A, or p30-­C268A and treated with DMSO or NU6300 (2.5, 5, and 10 μM) be-
fore LDH analysis. (D) Inhibitory effect of NU6300 on pyroptosis was evaluated by transfection with GSDMD, p30, p30-­C191A, p30-­C38A/C56A/C268A, or p30-­C38A/C56A/C191A/
C268A in HEK-­293T cells. (E) MST analysis by incubation of GSDMD protein (0.2 μM) with NU6300 or NU2. (F) Cell cytotoxicity of THP-­1 cells was analyzed by stimulation with nigericin
and then incubated with NU6300 or NU2 (0.5, 1, 2, and 4 μM). Data were expressed as means ± SEM, n = 3. Comparisons were calculated by one-­way ANOVA. ***P < 0.001. ns, not
significant.

Dis obviously protected against GSDMD degradation in THP-­1 cells
and BMDMs (fig. S5C). As expected, MST assay further revealed
binding of NU6300 to GSDMD with a KD value of 36.12 μM, whereas
no direct binding was observed between GSDMD and NU2 (Fig. 3E);
in addition, incubation with NU2 revealed no inhibition of pyroptotic
cell death (Fig. 3F). These results strongly indicated that NU6300
directly bound to C191 on GSDMD.
NU6300 blocks the cleavage and palmitoylation of GSDMD
To identify whether NU6300 was selective for GSDMD in
inflammasome-­GSDMD signals, we investigated the effect of
NU6300 on earlier steps. As expected, NU6300 exhibited no or
only weak inhibition effect on ASC oligomerization in AIM2,
NLRC4, and noncanonical inflammasomes (Fig. 4, A and B, and
fig. S6) and no effect on caspase-­1 activation in AIM2 and NLRC4

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 5 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
inflammasomes (Fig. 4, C and D), while NU6300 obviously
blocked GSDMD cleavage in AIM2 and NLRC4 inflammasomes
(Fig. 4, E and F), proving the selectivity of NU6300 on GSDMD.
Recent data demonstrated that palmitoylation of C191 was a
requirement for GSDMD membrane translocation and pore forma-
tion (5, 6). We further investigated whether NU6300 acted on pal-
mitoylation of GSDMD. As exhibited in Fig. 5 (A to C), NU6300
substantially blocked palmitoylation level of full-­length GSDMD in
AIM2 and NLRC4 inflammasomes and also obviously inhibited
palmitoylation level of GSDMD-­N in HEK-­293T cells transfected
with GSDMD-­N. We then extended our investigation into the trans-
location of GSDMD-­N to the membrane in AIM2 and NLRC4
inflammasomes, revealing NU6300’s inhibition of GSDMD-­N
translocation in both inflammasomes (Fig. 5, D and E). However,
the conclusion remained elusive because NU6300 concurrently
blocked GSDMD cleavage and subsequently reduced N-­terminal
domain levels. To resolve this ambiguity, we used HEK-­293T cells
transfected with GSDMD-­N to bypass the impact of cleavage. The
results corroborated that NU6300 could prevent the GSDMD-­N
translocation to membrane (Fig. 5F). In addition, NU6300 promi-
nently interrupted the formation of p30 oligomers in HEK-­293T
cells transfected with GSDMD-­N (Fig. 5G). Together, NU6300
covalently modified C191 of GSDMD, blocked its cleavage and
palmitoylation, and consequently inhibited GSDMD-­N membrane
translocation and further oligomerization.
To investigate whether NU6300 was specific to GSDMD-­
mediated pyroptosis, we also detected its potential binding to oth-
er gasdermin family members using DARTS. The results showed
that NU6300 failed to protect GSDMA, GSDMB, GSDMC, and
DFNB59, except for GSDME (fig. S7, A to E). However, NU6300
did not inhibit GSDME cleavage and GSDME-­mediated cell death
in etoposide-­treated RAW264.7 cells (fig. S7, F to H), indicating
that NU6300 did not target GSDME in micromole concentration
in cells and the GSDME protection effect of NU6300 on DARTS
assay might be a nonspecific result. These results confirmed the
selectivity of NU6300 targeting GSDMD over other gasdermin
family members.
GSDMD inhibitors show feedback inhibition on
NLRP3 inflammasome
NU6300 exhibited a quite strong inhibition effect on earlier steps
of NLRP3 inflammasome, with a remarkable inhibition on ASC
oligomerization, caspase-­1 activation, GSDMD cleavage, membrane
translocation, and the release of IL-­1β in nigericin-­triggered NLRP3
inflammasome (Fig. 6, A to F). These results impelled us to investi-
gate whether NU6300 also targeted NLRP3 or caspase-­1. However,
DARTS assay revealed that NU6300 had no protective effect on
NLRP3 and caspase-­1, and in addition, no direct binding was ob-
served between NU6300 and NLRP3 protein by surface plasmon
resonance (SPR) analysis (fig. S8, A to C), indicating no binding of
NU6300 to NLRP3 or caspase-­1. We further examined the effects of
NU6300 and other two known GSDMD inhibitors (NSA and Dis)
on earlier steps of the NLRP3 inflammasome–GSDMD signals. All
GSDMD inhibitors exhibited noticeable inhibition effects on ASC
oligomerization and the cleavage of the pro–caspase-­1, pro–IL-­1β,
and GSDMD (Fig. 6, G and H). These results suggested that there
was a common feedback inhibition effect on NLRP3 inflamma-
some when GSDMD was inhibited, which was in accordance with
previous studies (33, 34).
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
NU6300 alleviates DSS-­induced colitis and LPS-­induced
sepsis in mice
Aberrant inflammatory responses have been implicated in various
pathological conditions, including inflammatory bowel disease and
sepsis (35, 36). Activation of GSDMD, a key executioner of pyroptosis,
has been observed during intestinal inflammation in a chemically
induced colitis model (37). To explore the therapeutic potential of
NU6300, we investigated its effect on dextran sulfate sodium (DSS)–
induced colitis mouse model. We challenged C57BL/6J mice with
3.25% DSS for 6 days, followed by normal drinking water and concur-
rent intraperitoneal administration of NU6300 at different doses (5, 10,
and 20 mg/kg) or NSA (20 mg/kg) for five consecutive days (Fig. 7A).
Notably, NU6300 exhibited a significant dose-­dependent mitigation of
body weight loss and ameliorated the severity of DSS-­induced colitis, as
indicated by disease activity index (DAI) scores reflecting body weight
loss, diarrhea, and evident rectal bleeding (Fig. 7, B and C). Further-
more, NU6300 administration markedly improved the shortening of
the colon length, a characteristic feature of DSS-­induced colitis (Fig. 7D
and fig. S9A). Meanwhile, histological examination confirmed the ben-
eficial effects of NU6300 on colonic inflammasome, as it mitigated
DSS-­induced epithelial and mucosal damage, crypt dilation, and goblet
cell depletion (Fig. 7E). NU6300 also demonstrated greater efficacy in
ameliorating colonic inflammation compared to NSA, a known inhibi-
tor of GSDMD-­medicated pyroptosis (38). Both NU6300 and NSA ef-
fectively reduced the expression of proinflammatory cytokines IL-­1β
and tumor necrosis factor–α (TNFα) by enzyme-­linked immunosor-
bent assay (ELISA) analysis in the colonic area (Fig. 7F) and decreased
the level of cleaved caspase-­1 and GSDMD protein, as determined by
immunoblot analysis (Fig. 7G and fig. S9B). NU6300 greatly improved
the severity of DSS-­induced colitis compared to NSA administration,
suggesting its better therapeutic benefits in ameliorating colitis.
Considering the promising protective effects of GSDMD inhibi-
tors in the LPS-­induced mouse model of sepsis (39, 40), we pro-
ceeded to assess the potential of NU6300 in this regard. We
administered NU6300 (5 and 10 mg/kg) or NSA (10 mg/kg) intra-
peritoneally to the BALB/c mice for 1 hour before LPS (8 mg/kg)
injection (Fig. 7H). Survival analysis showed that mice treated with
NU6300 and NSA were effectively protected against the lethal con-
sequences of LPS-­induced sepsis (Fig. 7I). After a 4-­hour LPS chal-
lenge, the NU6300-­treated groups exhibited a significant reduction
in the concentrations of IL-­1β and TNFα in spleen (Fig. 7, J and K).
In addition, spleen from NU6300-­treated groups exhibited a sig-
nificant reduction in spleen index compared to LPS-­stimulated
mice (Fig. 7L). To check whether the survival protection effect of
NU6300 was due to GSDMD inhibition, we compared the survival
curves of wild-­type and GSDMD−/− C57BL/6J mice. GSDMD−/−
mice were resistant to LPS (50 mg/kg)–induced death as expected,
and notably, NU6300 (10 mg/kg) protected wild-­type mice from
LPS-­induced death but did not significantly affect the survival of
GSDMD−/− mice (Fig. 7M). Moreover, the reduction extent of
splenic IL-­1β and TNFα by NU6300 in GSDMD−/− mice was not as
notable as that in wild-­type mice (Fig. 7, N and O), and NU6300
totally lost its pyroptosis inhibition activity in GSDMD−/− BMDMs
(Fig. 7P). All these results indicated that GSDMD was the main tar-
get attributed to the protection effect of NU6300 on sepsis model.
However, we noticed that, in GSDMD−/− mice, NU6300 also weakly
inhibited the expression of splenic TNFα (Fig. 7O), suggesting
that there might be other targets of NU6300 accounting for its anti-­
inflammatory activity.
6 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. NU6300 blocks the cleavage of GSDMD. (A and B) ASC oligomerization detection in THP-­1 cells or BMDMs by disuccinimidyl suberate cross-­linking assays. Cells were primed
with LPS (1 μg/ml), incubated with NU6300 (2 μM), and then transfected with poly(dA:dT) (500 ng/ml) (A) and flagellin (250 ng/ml) (B) for 6 hours. (C and D) The caspase-­1 activity
was measured by Caspase-­Glo 1 reagent after AIM2 inflammasome (C) or NLRC4 inflammasome (D) activation in THP-­1 cells and BMDMs. (E and F) GSDMD oligomerization and
cleavage of GSDMD in THP-­1 cells and BMDMs after AIM2 inflammasome (E) or NLRC4 inflammasome (F) activation. Data were expressed as means ± SEM, n = 3. Comparisons were
calculated by one-­way ANOVA. *P < 0.05, **P < 0.01, and ***P < 0.001. ns, not significant.

Safety and pharmacokinetic properties of NU6300
To evaluate the druggability of NU6300, we first explored potential
cardiotoxicity in HEK-­293 cells stably transfected with the human
ether-­a-­go-­go–related gene (hERG) potassium channel, and NU6300
blocked hERG activity with an IC50 of 5.94 μM (fig. S10A). In the
maximum tolerated dose (MTD) investigation, we administered
three different doses, namely, 80, 100, and 120 mg/kg. Notably, mice
in the 120 mg/kg group experienced fatalities, while the 100 mg/kg
group demonstrated a mere deceleration in weight gain, with no
incidences of rodent fatalities. Consequently, the MTD of NU6300
in mice was determined to be 100 mg/kg. NU6300 had little effect
on heart, spleen, lung, and kidney tissues, except for a weak toxicity

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 7 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. NU6300 blocks the palmitoylation and membrane translocation of GSDMD-­N. (A and B) THP-­1 cells were primed with LPS (1 μg/ml) for 3 hours, followed by NU6300
(2 μM), and then transfected with poly(dA:dT) (500 ng/ml) (A) and flagellin (250 ng/ml) (B) for 6 hours, and cell lysates were treated with or without HA and subjected to the acyl-­biotin
exchange (ABE)–palmitoylation assay. (C) ABE-­palmitoylation assay in HEK-­293T cells transfected with p30 and incubated with NU6300 (5 μM). (D and E) Expression of GSDMD-­NT
in cell membrane and cytoplasm after treatment with NU6300 and transfected with poly(dA:dT) (D) or flagellin (E). (F and G) Expression of GSDMD-­NT (p30) protein in cell membrane
and cytoplasm (F) and the GSDMD-­NT oligomerization (G) in HEK-­293T cells transfected with p30 and incubated with NU6300 (5 μM). Data were expressed as means ± SEM, n = 3.
Comparisons were calculated by one-­way ANOVA. *P < 0.05, **P < 0.01, and ***P < 0.001. ns, not significant.

to the liver, as evidenced by mouse body weight, organ morphology,
and organ coefficients (fig. S10, B to D). The 5-­day toxicity showed
that a slight deceleration in body weight gain and no obvious path-
ological injury was observed in heart, liver, spleen, lung, and kid-
ney by hematoxylin and eosin (H&E) staining (fig. S10, E and F),
implying that NU6300 (20 mg/kg) had no apparent toxicity to vital
organs in vivo. We determined the single-­dose pharmacokinetic
profile of NU6300 in C57BL/6J mice by administering NU6300
intravenously (20 mg/kg) and intraperitoneally (20 mg/kg). The
area under the curve was 137.20 μg hour−1 liter−1, with a Cmax of
190.09 μg/liter and a half-­life of 12.38 hours for intraperitoneal
administration (table S2).

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 8 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. GSDMD inhibitors show a feedback inhibition on NLRP3 inflammasome. (A) Oligomerization detection by disuccinimidyl suberate cross-­linking assays. THP-­
1 cells and BMDMs were primed with LPS (1 μg/ml), followed by NU6300 (2 μM) treatment, and stimulated with nigericin (10 μM), and then the ASC oligomerization was
detected. (B) LPS-­primed THP-­1 cells were incubated with NU6300 (2 μM) or z-­VAD-­fmk (20 μM) before stimulation with nigericin, and then the caspase-­1 activity was
measured by Caspase-­Glo 1 reagent. (C) The cleavage and oligomerization of GSDMD were analyzed in LPS plus nigericin-­activated THP-­1 cells and BMDMs. (D) Membrane
translocation assay in LPS and nigericin-­activated THP-­1 cells. (E) Confocal fluorescence microscope study for GSDMD membrane staining after incubation with NU6300
or z-­VAD-­fmk in LPS plus nigericin-­activated THP-­1 cells. Scale bars, 10 μm. (F) The release of IL-­1β was detected by ELISA analysis. (G and H) LPS-­primed THP-­1 cells were
incubated with NU6300 (2 μM), z-­VAD-­fmk (20 μM), NSA (10 μM), or Dis (20 μM) before stimulation with nigericin. The treated cells were analyzed for ASC oligomerization
(G) and pro–caspase-­1, GSDMD, and pro–IL-­1β cleavage, the maturation and release of caspase-­1 and IL-­1β (H) by immunoblot of culture supernatants (Sup) or whole-­cell
lysate (WCL). Graphs showed means ± SEM, n = 3. Statistics were analyzed by one-­way ANOVA. **P < 0.001 and ***P < 0.001.

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 9 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. NU6300 alleviates DSS-­induced colitis and LPS-­induced sepsis in mice. (A) Flowchart of colitis mouse model induced by DSS (n = 6). (B) Daily loss weight of
mice (n = 6). (C) The DAI was recorded (n = 6). (D) Colon length of mice (n = 6). (E) H&E staining to observe the pathological changes of colons. Scale bar, 200 μm. (F) The
IL-­1β and TNFα levels in colon tissues were determined by ELISA analysis (n = 6). (G) The protein expression of cleaved GSDMD in colons was detected by immunoblot
assay (n = 3). (H) Flowchart of sepsis mouse model induced by LPS. (I) The survival curves were analyzed by using the log-­rank (Mantel-­Cox) test (n = 10). (J to L) Levels of
IL-­1β (J) and TNFα (K) in spleen were measured by ELISA, and spleen index (L) was analyzed after administration with LPS for 4 hours (n = 6). (M) Survival curves in wild-­type
and GSDMD−/− C57BL/6J mice after administration with or without NU6300 (10 mg/kg) for 1 hour and then after challenge with LPS (50 mg/kg) (n = 10). (N and O) Mice
were pretreated with or without NU6300 (10 mg/kg) and then injected with LPS (50 mg/kg) for 4 hours, and the levels of IL-­1β (N) and TNFα (O) in spleen were measured
by ELISA (n = 6). (P) Cell cytotoxicity of BMDMs in wild-­type and GSDMD−/− mice was analyzed by incubation with NU6300 and stimulation with nigericin (n = 3). Data
were presented as means ± SEM, and one-­way ANOVA was performed. *P < 0.05, **P < 0.01, and ***P < 0.001. ns, not significant.

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 10 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Overall, the narrow safety window for cardiotoxicity and low Three reported covalent GSDMD inhibitors, DMF, Dis, and NSA,
exposure in blood indicated that the druggability of NU6300 was have different mechanisms of targeting GSDMD (14, 25, 26). Dis and
not satisfactory. Further structural optimization was required to NSA have been reported to inhibit GSDMD oligomerization but with
enhance its druggability. no effect on its cleavage (14, 25), while DMF has been reported to
inhibit the interaction between GSDMD and caspase-­1, thereby pre-
venting GSDMD cleavage (26). NU6300 inhibited GSDMD cleavage
DISCUSSION through a mechanism similar to DMF. Recent studies have also shown
Dysregulation of inflammation response is centrally incentive in the that reactive oxygen species–dependent palmitoylation of the C191
pathogenesis of various inflammatory diseases, such as sepsis, acute residue in GSDMD is an important modification for its membrane
intestinal inflammation, and cancers (35, 41, 42). Recently, a new localization and subsequent oligomerization (5, 6, 53, 54). Our re-
type of inflammatory programmed cell death, known as pyroptosis, search further demonstrated that NU6300 could remarkably inhibit
has been found. It is triggered by the activation of canonical and the palmitoylation of full-­length GSDMD and GSDMD-­N. In p30-­
noncanonical inflammasomes, notably via NLRP3 inflammasome transfected HEK-­239T cells, we observed that NU6300 could inhibit
activation (1, 43, 44). As GSDMD is a newly identified pyroptosis the palmitoylation and membrane localization. The palmitoylation-­
executioner downstream of inflammasome activation, increasing related mechanism has not been reported for existing GSDMD in-
interest is thus emerging in pursuing GSDMD as a novel therapeutic hibitors, which will have guiding implications for the discovery and
target for alleviating pyroptosis diseases. mechanistic studies of future GSDMD inhibitors. These studies
While drug discovery efforts have historically focused on re- suggested that NU6300 exhibited a dual-­mechanism approach, effec-
versibly binding molecules, recent success with covalent inhibi- tively inhibiting both the cleavage and palmitoylation processes of
tors has highlighted their potential as therapeutic agents (45, 46). GSDMD. Palmitoylation could occur on both the full-­length GSDMD
Hence, targeted covalent inhibitors have become integral parts of and GSDMD-­N, suggesting that the palmitoylation of GSDMD was
drug discovery approaches (46, 47). NU6300, a CDK2 inhibitor, independent of its cleavage (5, 6). Furthermore, a specific palmi-
with vinyl sulfone as the warhead, was identified as a potent inhibi- toylation inhibitor, 2BP, did not inhibit caspase-­1–mediated GSDMD
tor of pyroptotic cell death. It is the first covalent CDK2 inhibitor cleavage but inhibited LPS and nigericin-­induced cell death in a dose-­
with Lys89 as the binding site, a residue that lies just outside the dependent manner (5), indicating that the cleavage of GSDMD was
CDK2 adenosine triphosphate (ATP)–binding cleft (29). In addi- independent of its palmitoylation. This implied that the inhibition of
tion to its involvement in cell cycle regulation, CDK2 was reported GSDMD cleavage and palmitoylation by NU6300 were two indepen-
as a new target for mitigating neutrophil migration and relieving dent mechanisms that contribute to the overall inhibition of pyrop-
inflammatory ailments (48). However, we observed no pyroptosis tosis. Structurally, NU6300 provided a new skeleton for GSDMD
inhibition effect in both noncovalent CDK2 inhibitor CVT-­313– inhibitor that is worthy of further research. However, C191 is located
treated and CDK2 knockdown THP-­1 cells after nigericin induc- on the edge of GSDMD (55), and there is no protein cavity available
tion, indicating that NU6300 might act on other targets to regulate for small-­molecule binding and design. Therefore, it might be difficult
pyroptosis. to further design potential compounds based on the aforementioned
DARTS is a universally applicable target identification approach three small molecules and NU6300. The future research might involve
that analyzes direct drug binding to targets (49), which requires no the development of compounds targeting undefined functional sites
modification of the drug and is independent of any biological of GSDMD.
effects of the drug (50–52). In our research, we performed DARTS As NU6300 failed to impede the preliminary ASC oligomeriza-
to reveal the potential target of NU6300 account for the anti-­ tion and caspase-­1 processing provoked by AIM2 and NLRC4 in-
pyroptotic effects. Consequently, we identified that GSDMD was flammasome activation, these studies demonstrated that NU6300
the most relevant target. C38, C56, C191, and C268 are four con- displayed selectivity. However, NU6300, together with two known
served cysteines in the p30 fragment of GSDMD, and NU6300 GSDMD inhibitors (Dis and NSA), exhibited strong inhibition ef-
modified C56, C191, and C268, as indicated by MS data. We thus fects on NLRP3 inflammasome. These findings revealed a common
conducted mutation research on these four cysteines to uncover the feedback inhibition effect on NLRP3 inflammasome when GSDMD
potential binding site of NU6300 within p30 fragment. Both single-­ was inhibited. GSDMD is widely acknowledged as a typical down-
site and multi-­site mutations indicated that C191 was the only modi- stream pyroptosis executioner in the NLRP3 inflammasome–GSDMD
fication on GSDMD associated with NU6300’s inhibitory effect on pathway, while some studies implied that GSDMD might transcend
pyroptosis. In general, lysine-­and cysteine-­reactive small molecules its canonical role by stimulating NLRP3 activation (34). Our results
might covalently bind to other proteins. However, our research has provided solid evidences to support this notion. The underlying
confirmed that these additional modifications were not relevant to mechanism might be explained as follows: When GSDMD was
the GSDMD-­mediated inflammatory pathway discussed in this inhibited, the K+ efflux was blocked because of abolishment of
study. First, NU6300 suppressed the cleavage, palmitoylation, and GSDMD pore-­forming capability, which augmented intracellular
subsequent biochemical processes of GSDMD without affecting K+ concentration and consequently inhibited NLRP3 activation
ASC oligomerization and caspase-­1 activation in AIM2 and NLRC4 (56, 57). The feedback inhibition findings suggested that target-
inflammasomes. This indicated that these additional modifications ing GSDMD might be more effective on NLRP3 than other in-
of NU6300 did not influence the upstream steps. Second, in the mu- flammasomes.
tation experiment, we observed NU6300 deactivation upon GSDMD Despite being a covalently irreversible CDK2 inhibitor, no
C191A mutation, providing evidence that the additional modifica- studies have been reported on the efficacy of NU6300 in animal
tions of NU6300 did not affect the downstream steps of GSDMD in experiments (29). In this study, we evaluated the pharmacological
mediating pyroptosis. effect of NU6300 on mouse models of colitis and sepsis, given the

Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024 11 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
well-­known role of pyroptosis and IL-­1β release in acute inflamma-
tory bowel diseases and sepsis. As speculated, NU6300 ameliorated
the pathology, GSDMD cleavage, and release of inflammatory cyto-
kines associated with acute colitis induced by DSS. Meanwhile,
NU6300 reduced the cleavage of caspase-­1, indicating that NU6300
primarily acted by limiting feedforward NLRP3 activation and
reducing GSDMD pore formation in this model. However, there
might also be other mechanisms involved. For instance, NU6300
might directly target upstream signals such as NLRP3 or caspase-­1.
Nevertheless, our findings revealed no such direct interactions.
NU6300 also improved the survival of septic mice and suppressed
the release of splenic inflammatory mediators induced by LPS, while
it only slightly suppressed the expression of TNFα in GSDMD−/−
mice, indicating that GSDMD was the main target attributed to the
protection effect of NU6300 on sepsis model and other targets of
NU6300 accounting for its anti-­inflammatory activity might also be
involved.
In summary, our experiments demonstrated that NU6300, an
inhibitor of GSDMD, mitigated GSDMD-­mediated inflammation
both in vitro and in vivo. Mechanistically, NU6300 covalently
reacted with C191 of GSDMD to block its cleavage and palmi-
toylation (Fig. 8). Inhibition of GSDMD held great promise as a
therapeutic strategy for inflammatory pyroptosis disorders, and
NU6300 represented a promising lead compound for the treatment
of inflammatory diseases.
MATERIALS AND METHODS
Plasmids, reagents, and antibodies
Human GSDMD was cloned into a modified pCS2 vector with an
N-­terminal 3 × Flag tag for transient expression in HEK-­293T cells.
All cloning and mutagenesis were verified through DNA sequenc-
ing. GSDMD full length (1–484), C-­GSDMD (276–484), GSDMD-
­N/p30 (1–275), p30-­C38A, p30-­C56A, p30-­C191A, p30-­C268A, p30-­C38A/
C56A/C268A, and p30-­C38A/C56A/C191A/C268A point mutations
were generated by polymerase chain reaction (PCR)–mediated muta-
genesis of the cysteine to alanine using KOD Plus Neo DNA poly-
merase (Toyobo) and subcloning of the mutant cDNAs into the
vector using the MonClone Hi-­Fusion Cloning Mix V2 kit (Monad,
China, Shanghai). For recombinant expression in Escherichia coli,
the cDNAs were cloned into a modified pET28a vector with an
N-­terminal 6 × His-­SUMO tag. Truncation mutants of GSDMD
were constructed by the standard PCR cloning strategy and inserted
into the corresponding vectors. All plasmids were verified by DNA
sequencing.
Phorbol 12-­myristate 13-­acetate (PMA), LPS (serotypes O111:
B4 and O55: B5), poly(deoxyadenylic-­d eoxythymidylic) acid
[poly(dA:dT)], TNFα, and SMAC mimetic were from Sigma-­
Aldrich. The pan-­caspase inhibitor z-­VAD-­fmk and Dis were from
ApexBio. NSA was purchased from TargetMol. Nigericin was ob-
tained from InvivoGen. Flagellin was from Enzo Life Sciences.
Caspase-­Glo 1 inflammasome assay and CytoTox 96 nonradioac-
tive cytotoxicity assay were from Promega. Lipofectamine 2000 re-
agent was purchased from Invitrogen. The antibodies adopted in
experiments include NLRP3 antibody (Abcam), caspase-­1 anti-
body (p45, p20, AdipoGen), IL-­1β antibody (p31, p17, Affinity),
ASC antibody (AdipoGen), GSDMA (Huabio), GSDMB (Abcepta),
GSDMC (Huabio), GSDMD antibody (Huabio), GSDME antibody
(Abcam), DFNB59 antibody (ImmunoWay), NAPRT (Santa Cruz
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
Biotechnology), UMPS (Huabio), CAP1(Zen Bioscience), TCP1
(Abways), β-­actin antibody (Abways), DDDDK-­ Tag (Flag, Ab-
clonal), and glyceraldehyde-­ phosphate dehydrogenase (GAP-
3-­
DH) (Abclonal).
Mice
Animal experiments were performed on 8-­week-­old male C57BL/6J
mice (Animal Center of Beijing HFK Bioscience Co Ltd., Beijing,
China) and BALB/c mice (Speyford Biotechnology Co. Ltd., Beijing,
China). GSDMD knockout mice were purchased from Gemphar-
matech Co. Ltd. Mice were maintained in specific pathogen–free
(SPF)–grade facility and quarantined for a week before the formal
experiments. Animal experiments were conducted according to
protocols approved by the Animal Care and Use Committees of
State Key Laboratory of Biotherapy (Sichuan University).
Cell lines and stimulation
THP-­1 cells were cultured in RPMI 1640 (Gibco) supplemented
with 10% fetal bovine serum (Procell) and penicillin-­streptomycin
(100 U/ml; Gibco). BMDMs were obtained from male C57BL/6J
mice by culturing in Dulbecco’s modified Eagle’s medium (DMEM)
supplemented with 30% L929 cell culture supernatant for 6 days.
HEK-­293T cells and HT-­29 cells were grown in DMEM with 10%
fetal bovine serum and penicillin-­streptomycin (100 U/ml).
Differentiated THP-­1 cells were acquired by coculturing with
PMA (100 ng/ml) for 16 hours, then primed with LPS (1 μg/ml)
for 3 hours, and pretreated with NU6300 (2 μM) for 40 min be-
fore stimulation with nigericin (10 μM) for 35 min. For AIM2
and NLRC4 inflammasome activation, PMA differentiated cells
were primed with LPS and then transfected with poly(dA:dT)
(500 ng/ml) or flagellin (250 ng/ml) for 6 hours. For noncanoni-
cal inflammasome activation, cells were pretreated with Pam3C-
S4K (400 ng/ml) for 3 hours and then stimulated with cytosolic
LPS (1.5 μg/ml). Transient transfection of HEK-­293T cells was
performed using ExFect Transfection Reagent (Vazyme) accord-
ing to the manufacturer’s instructions. Necroptosis was induced
in HT-­29 cells by adding TNFα (25 ng/ml) together with 400 nM
SMAC mimetic and 20 μM z-­VAD-­f mk for 24 hours in the pres-
ence or absence of NU6300. Cell viability was determined by CCK8
analysis.
Assays of cell death
Culture supernatants in inflammasome-­activated cells were mea-
sured by LDH assay according to the manufacturer’s instruction.
For cytotoxicity in HEK-­293T cells, cells were transfected with the
plasmid with a complete medium change after 4 hours, followed by
NU6300 stimulation for 16 hours before performing LDH determi-
nation. To assess cell cytotoxicity and permeability, PI staining assay
was conducted. After NU6300 treatment, the cells were cultured in
buffered salt solution [25 mM Hepes, 5 mM glucose,120 mM NaCl,
4 mM KCl, 1.5 mM CaCl2, 1 mM magnesium chloride, and 0.1%
bovine serum albumin (BSA) at pH 7.4] containing PI (2.5 μg/ml).
Maximum fluorescence was measured by lysing each well with 0.1%
Triton X-­100. Then, the fluorescence at 535/617 nm (excitation/
emission) was continuously recorded for 130 min after nigericin
activation at 5-­min intervals using a Biotek Synergy plate reader,
and the images were captured with Nikon microscope. For cell death
in RAW264.7 macrophages for GSDME experiments, the cells were
plated with 2 × 105 cells/ml in six-­well plate, treated with NU6300
12 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 8. The mechanism of NU6300 modulating pyroptosis.
for 3 hours, and then stimulated with etoposide (100 μM) for 16 hours
for GSDME-­mediated cell death.
Transmission electron microscopy
LPS priming PMA-­differentiated THP-­1 cells were treated with
NU6300 (2 μM) for 40 min and then induced by nigericin (10 μM)
for 35 min. After incubation, the cells were washed with phosphate-­
buffered saline (PBS), collected, and then fixed with 0.5% glutar-
aldehyde at 4°C for 5 min, followed by centrifugation at 12,000 rpm
for 15 min. Next, the cell precipitate was fixed in 3% glutaraldehyde.
Ultrathin sections were imaged under a JEM-­1400PLUS transmis-
sion electron microscope (JEOL Ltd., Tokyo, Japan).
Measurement of cytokines
The concentrations of cytokine IL-­1β in culture supernatants from
THP-­1 cells, BMDMs, colon and spleen from mice, and TNFα in colon
and spleen from mice were quantified by ELISA kit (4A BIOTECH)
according to the manufacturer’s instruction.
Immunoblot
Medium supernatants from macrophages were precipitated by
fixing with ice-­cold trichloroacetic acid overnight and dissolved
in precooled acetone after centrifugation, followed by centrifuga-
tion to obtain supernatant samples. Whole-­cell lysates were pre-
pared by lysing cell pellets in radioimmunoprecipitation assay
(RIPA) lysis buffer with freshly added protease inhibitors and
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
clarified by centrifugation. All samples were separated by SDS–
polyacrylamide gel electrophoresis (SDS-­PAGE) and transferred
to polyvinylidene difluoride membranes. Following incubation in
the primary antibody, the membranes were incubated in horse-
radish peroxidase–conjugated secondary antibody (Abways) and
signals were captured using a gel imaging system (Shanghai Qin-
xiang, ChemiScope 3200, China).
DARTS and protein site modification
LPS-­priming PMA-­differentiated THP-­1 cells were harvested and
lysed with NP-­40 lysis buffer and centrifuged at 13,000 rpm for 15 min
at 4°C. The total protein content or purified GSDMD protein (0.2 μg/ml)
was incubated with or without NU6300 (1000, 100, and 10 μM) for
50 min and then digested with pronase (at the indicated protease to
protein ratios) for 30 min, followed by termination by addition of pro-
tease inhibitor cocktail for 10 min and 5 × SDS loading buffer to boil
for 10 min at 100°C. The samples were separated via SDS-­PAGE and
examined by immunoblot analysis. For site modification, we incubated
100 μl of purified human GSDMD (0.45 mg/ml) with NU6300 at a mo-
lar ratio of 1:20 (GSDMD protein:NU6300) at room temperature for
3 hours. The unbound small molecules were removed by transferring
the sample to a 10-­kDa Ultra-­0.5 ultrafiltration tube and rinsed with
washing buffer [20 mM tris-­HCl (pH 8.0), 300 mM NaCl, and 0.5 mM
Tris(2-­carboxyethyl)phosphine hydrochloride (TCEP)], centrifuged
at 10,000 rpm for 5 min, and repeated three times. The sample was col-
lected for SDS-­PAGE before Coomassie brilliant blue staining.
13 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
MS and sample preparation
For identification of proteins or protein site modification by MS,
Coomassie brilliant blue–stained gels were excised and cut into 2-­to
3-­mm-­size pieces and transferred into 1.5-­ml polypropylene tubes.
Cut gels were destained before incubation with 100% acetonitrile
(ACN) at room temperature. After lyophilizing the solvent, samples
were digested with sequencing-­grade trypsin (Promega, 5 ng/μl) for
20 hours and extracted in 60% ACN/0.1% trifluoroacetic acid, fol-
lowed by lyophilization. The samples were processed for LC-­MS/MS
analysis with mobile phase consisting of a 0.1% formic acid aqueous
solution (liquid A) and a 0.1% formic acid ACN solution (liquid B).
The column was balanced by 95% mobile phase A. Samples were
loaded onto a Thermo Fisher Scientific EASY trap column (100 μm
by 2 cm, 5 μm, 100 Å, C18) and separated by Thermo Fisher Scien-
tific analytical column (75 μm by 25 cm, 5 μm, 100 Å, C18). The
gradient elution program was as follows: 0 to 40 min, 5 to 28% B; 40
to 42 min, 28 to 90% B; 42 to 60 min, 90% B (v/v). The samples were
separated by capillary high-­performance liquid chromatography,
and then analyzed by MS using a Q-­Exactive mass spectrometer
(Thermo Finnigan, San Jose, CA, USA) in positive ion mode under
the following conditions: 375 to 1800 mass/charge ratio (m/z) pre-
cursor ion scan range with a mass resolution of 120,000 at m/z 200,
MS1 automatic gain control: 4 × 105, ion implantation time: 50 ms,
cycle time 3 s between master scans. The raw data for MS analysis
are raw files, and the software MaxQuant 2.0.1.0 is used for library
identification and quantitative analysis.
Protein expression and purification
For expression of full-­length human GSDMD, E. coli BL21 (DE3)
cells harboring the GSDMD plasmid (pET28a-­6 × His-­SUMO
vector) were grown in LB medium supplemented with kanamy-
cin (50 μg/ml) at 18°C overnight after induction with 0.4 mM
isopropyl-­β -­​d -­
t hiogalactopyranoside (IPTG) when OD 600
reached 0.8. The cells were harvested by centrifugation (4500 rpm,
15 min, 4°C), ultrasonicated in the buffer [20 mM tris-­HCl buf-
fer (pH 8.0), 300 mM NaCl, 5 mM imidazole, 20 mM MgCl2, 10 mM
KCl, 0.5 mM TCEP, 0.1 mM protease inhibitor], and disrupted
by high-­pressure micro-­fluidization on ice and centrifugation
(15,000 rpm, 30 min, 4°C) to isolate the supernatant containing
the target protein. The supernatant was then incubated with Ni–
nitrilotriacetic acid resin for 1 hour at 4°C and then washed with
lysis buffer, after which elution was performed in the lysis buffer
supplemented with 300 mM imidazole. The eluted protein was
subsequently purified using a desalting column and subjected to
SUMO tag overnight cleavage at 4°C by the addition of ULP1
protease. Protein was further subjected to a second round of pu-
rification. The purified GSDMD protein was further purified us-
ing a superdex 200 gel filtration column (GE Healthcare Life
Sciences) preequilibrated with 20 mM tris-­HCl buffer (pH 8.0),
300 mM NaCl, and 0.5 mM TCEP. The purified GSDMD was
concentrated to 10 mg/ml before freezing at −80°C.
CETSA and TSA
For CETSA analysis, the supernatants of LPS-­priming PMA-­
differentiated THP-­1 cells or BMDMs lysed in modified RIPA buffer
were harvested and incubated with dimethyl sulfoxide (DMSO) or
NU6300. The samples were then subjected to controlled heating,
with gradually increasing temperatures from 45° to 72°C for 6 min,
followed by cooling at room temperature for 3 min. The samples
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
were resuspended with 5 × SDS loading buffer and boiled for 10 min
before SDS-­PAGE for immunoblot analysis. In terms of TSA
analysis, the purified GSDMD protein was incubated with DMSO
or NU6300 and then heated with increasing temperatures from
48° to 72°C for 6 min for subsequent immunoblot analysis as
above described.
BLI assay
Biotinylated GSDMD was prepared, and the BLI analysis was
performed as previously described (58). Briefly, the GSDMD
protein was dissolved in PBS (100 μg/ml) and subjected to bio-
tinylate with EZ-­L ink NHS-­LC-­LC-­Biotin kit (Thermo Fisher
Scientific). Before each assay, streptavidin (SA) biosensors were
pre-­wetted in double-­distilled water to record the baseline. The
biotinylated human GSDMD protein was immobilized on SA-­
coated 96-­well plate surface, after which a 200-­μl solution con-
taining DMSO or various concentrations of NU6300 diluted by
PBS (containing 2.5% DMSO and 0.02% Tween 20) was added to
each well. The binding experiment was performed with repeated
cycles consisting of four major steps: initial loading, baseline,
association, and dissociation. Data collection and analysis were
performed by ForteBio Octet (Port Washington, NY, USA).
MST assay
Protein sample was swapped on the desalting column, and the flow-­
through was collected after the desalting column had been rinsed
with 300 μl of labeling buffer and centrifuged three times at 1500g.
Then, the protein was covalently attached by fluorescent label dye
RED-­NHS. NU6300 was diluted to the indicated concentration and
incubated with labeled GSDMD protein for 30 min in assay buffer
[20 mM tris (pH 8.0), 300 mM NaCl, 0.5 mM TCEP, and 0.5‰
Tween 20]. The samples were loaded into the Monolith NT.115 cap-
illaries for MST assay. The KD value was calculated using the mass
action equation via the MO Affinity Analysis V2.1.3 software.
Oligomerization assay
For oligomerization, the treated cells were lysed in NP-­40 lysis
buffer and incubated on a shaker for 30 min at 4°C. After cen-
trifugation at 6000g for 15 min, the cell pellets were cross-­linked
with fresh disuccinimidyl suberate (2 mM) for 30 min at 37°C
and pelleted by centrifugation. The pellets were denatured in
SDS sample buffer and subjected to SDS-­PAGE analysis.
Immunostaining
PMA-­differentiated THP-­1 cells were seeded on coverslips over-
night and primed with LPS, treated with DMSO or NU6300, and
then induced with nigericin. Activated cells were fixed with
precooled methanol and blocked with 5% BSA, followed by incu-
bation with GSDMD antibody and fluorochrome-­conjugated
anti-­rabbit immunoglobulin G (IgG). The nuclei were counter-
stained with 4′,6-­diamidino-­2-­phenylindole (DAPI), and GSD-
MD membrane location was observed by using a Leica confocal
microscope.
Membrane localization assay
The treated cells were collected, and then cell membrane and
cytoplasmic proteins were extracted by a cell membrane protein
and cytoplasmic protein extraction kit (Beyotime) in accor-
dance with the manufacturer’s instructions.
14 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Acyl-­biotin exchange assay
To detect the palmitoylation of GSDMD, treated cells were lysed with
RIPA lysis buffer containing protease inhibitor cocktail. One milli-
gram of protein was then incubated with 1% Triton X-­100 and
final 25 mM N-­ethylmaleimide for 30 min at 4°C with gentle full-­angle
rotation and then precipitated with chloroform-­methanol. The air-­
dried pellet was resuspended in buffer (4% SDS, 5 mM EDTA, and
50 mM tris-­HCl, pH 7.4) and divided into two aliquots. The two ali-
quots were incubated with 800 μl of PBS containing 0.2% Triton X-­100
and 4 mM HPDP-­Biotin with or without 0.7 M hydroxlamin (HA)
through rotation for 2 hours at 25°C and precipitated protein by repeat-
ing the chloroform-­methanol assay. To purify biotinylated protein, SA-­
agarose beads were added and incubated at 4°C for 2 hours. Afterward,
the beads were washed with PBS containing 150 mM NaCl and 0.1%
Triton X-­100 to remove any nonspecifically bound proteins. The bioti-
nylated proteins were eluted by 50 μl of buffer and boiled for SDS-­
PAGE analysis.
Caspase-­1 activity assay
The caspase-­1 activity in cultured cells was detected using Caspase-­
Glo 1 inflammasome assay kit according to the protocol. THP-­1 cells
were seeded in 96-­well plates with 2 × 105 cells/ml, differentiated
with PMA overnight, primed with LPS, treated with DMSO or
NU6300, and then induced with nigericin. After induction, the cul-
ture medium was removed, 100 μl of Caspase-­Glo 1 reagent or
YVAD-­CHO reagent was added, and the contents were gently mixed.
After incubation at room temperature for 1 hour, luminescence was
measured using a Biotek Synergy plate reader.
SPR analysis of recombinant protein
Recombinant human NLRP3 (residues 1 to 1036) was acquired from
CUSABIO CSB-­EP822275HU (A4), diluted to 50 μg/ml in sodium
acetate (pH 4.0), and then immobilized to the Series S Sensor Chip
CM5 (BR100530, Cytiva) through amine-­coupling chemistry accord-
ing to the manufacturer’s instructions. Flow cell 1 was immobilized
without protein to enable reference subtraction in PBST (PBS contain-
ing 0.5% Tween 20, 0.5% DMSO), and flow cell 2 was the detection
cell. Single-­cycle kinetics was performed for the kinetic experiments,
and the analyte compounds were prepared using a fivefold increased
concentration gradient in PBST. The equilibrium dissociation constant
(KD) was calculated by Biacore T200 (Cytiva, USA) evaluation software.
DSS-­induced colitis model in mice
The male C57BL/6J mice weighing 18 to 22 g were randomly divided
into six groups (n = 6), and the normal group was allowed to drink water
freely. The other groups (DSS, NU6300, and NSA treatment groups)
were given 3.25% DSS in drinking water (36,000 to 50,000 MW, Yeasen
Biotechnology Co. Ltd., Shanghai) for 6 days, followed by regular drink-
ing water for 5 days. Meanwhile, the treatment groups were intraperito-
neally injected with NU6300 (5, 10, and 20 mg/kg) or NSA (20 mg/kg)
as the positive group for five consecutive days. Daily weight loss was re-
corded, and DAI was monitored throughout the experiment based on
body weight, stool consistency, and hemoccult. The mice were sacrificed
on day 11, and the colons were collected for determination of colon
length, H&E staining analysis, cytokine analysis, and immunoblot assay.
LPS-­induced model of sepsis in mice
The BALB/c mice were induced by intraperitoneal injection of LPS
(8 mg/kg, O55:B5; Sigma-­Aldrich, catalog no. L2880) and administered
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
with NU6300 (5 or 10 mg/kg) or NSA (10 mg/kg). The wild-­type
and GSDMD−/− C57BL/6J mice were induced by intraperitoneal in-
jection of LPS (50 mg/kg) and administered with NU6300 (10 mg/kg).
For survival study, mice (n = 10 mice per group) were monitored
every 12 hours for a period of 96 hours after administration. To
measure cytokines (n = 6 mice per group), spleen was collected at
4 hours after LPS challenge, and the concentration of cytokine was
determined by ELISA analysis. For spleen index, spleen was excised
and spleen index was calculated according to the following equa-
tion: spleen index (g/g) = weight of spleen/body weight × 100%.
hERG test
The inhibitory effect of NU6300 on hERG currents was measured in
HEK-­293 cells stably transfected with hERG potassium channel
with patch clamp. The NU6300 was detected at concentrations of
0.3, 1, 3, 10, and 30 μM, and the positive control terfenadine was set
at 0.001, 0.01, 0.1, and 1 μM.
MTD of NU6300
The C57BL/6J mice were randomly allocated into three treatment
groups (80, 100, and 120 mg/kg) with an equal number of male and
female mice. After intraperitoneally administering the aforemen-
tioned doses, detailed observations were conducted every day in-
cluding physical appearance, general behavioral activities, body
weight, mortality, and any other abnormal symptoms. On the 14th day
after administration, organ coefficients analysis was conducted for
the MTD group.
In vivo toxicity study of NU6300
To assess the toxicity of NU6300 in vivo, healthy C57BL/6J mice
were given NU6300 (20 mg/kg) intraperitoneally, while an equal
volume of 0.9% saline was given to the control group. Mice were
sacrificed after 5 days of administration. The heart, spleen, liver,
kidneys, and lungs were collected for H&E staining.
Pharmacokinetic study
Male BALB/c mice were randomly divided into two groups (n = 3)
to receive either intravenous (20 mg/kg) or intraperitoneal (20 mg/kg)
administration of NU6300. Blood samples (20 μl each time) were
collected from the posterior orbital venous plexus at 5, 15, and 30 min
and 1, 2, 4, 6, 8, 10, and 24 hours after administration. Plasma samples
were obtained by centrifugation (3500 rpm, 4°C, 15 min), and then
the concentration of NU6300 in plasma was determined by LC-­MS/
MS analysis. Standard curve for NU6300 in plasma was generated
by adding various concentrations, including an internal standard, to
blank plasma. The Shimadzu LC-­30 AD system was used for separa-
tion, with a mobile phase of ACN/0.1% formic acid at a flow rate of
0.5 ml/min. All blood samples were centrifuged and quantified us-
ing the Shimadzu LC-­30 AD system coupled with AB SCIEX 5500
QTRAP mass spectrometer. Pharmacokinetic parameters were cal-
culated using DAS 2.0.
Statistics
Data were expressed as the SEM and performed at least three
times. All data summarization, visualization, and statistical
analyses were performed using GraphPad Prism v8.1.2 (Graph-
Pad Software, San Diego, CA). One-­way analysis of variance
(ANOVA) or two-­way ANOVA test was used to determine sta-
tistical significance, and differences with P < 0.05 were considered
15 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
significant. All significant values were indicated as follows: *P < 0.05,
**P < 0.01, and ***P < 0.001.
Supplementary Materials
This PDF file includes:
Supplementary Materials and Methods
Figs. S1 to S10
Tables S1 and S2
Legend for data S1
Other Supplementary Material for this manuscript includes the following:
Data S1
REFERENCES AND NOTES
1. J. J. Shi, Y. Zhao, K. Wang, X. Y. Shi, Y. Wang, H. W. Huang, Y. H. Zhuang, T. Cai, F. C. Wang,
F. Shao, Cleavage of GSDMD by inflammatory caspases determines pyroptotic cell death.
Nature 526, 660–665 (2015).
2. S. Feng, D. Fox, S. M. Man, Mechanisms of gasdermin family members in inflammasome
signaling and cell death. J. Mol. Biol. 430, 3068–3080 (2018).
3. J. J. Ding, K. Wang, W. Liu, Y. She, Q. Sun, J. J. Shi, H. Z. Sun, D. C. Wang, F. Shao,
Pore-­forming activity and structural autoinhibition of the gasdermin family. Nature 535,
111–116 (2016).
4. S. Y. Kuang, J. Zheng, H. Yang, S. H. Li, S. Y. Duan, Y. F. Shen, C. N. Ji, J. H. Gan, X. W. Xu,
J. X. Li, Structure insight of GSDMD reveals the basis of GSDMD autoinhibition in cell
pyroptosis. Proc. Natl. Acad. Sci. U.S.A. 114, 10642–10647 (2017).
5. A. Balasubramanian, L. Ghimire, A. Y. Hsu, H. Kambara, X. Liu, T. Hasegawa, R. Xu, M. Tahir,
H. Yu, J. Lieberman, H. R. Luo, Palmitoylation of gasdermin D directs its membrane
translocation and pore formation in pyroptosis. bioRxiv 2023.02.21.529402 [Preprint]
(2023). https://doi.org/10.1101/2023.02.21.529402.
6. G. Du, L. B. Healy, L. David, C. Walker, P. Fontana, Y. Dong, P. Devant, R. Puthenveetil,
S. B. Ficarro, A. Banerjee, J. C. Kagan, J. Lieberman, H. Wu, ROS-­dependent palmitoylation
is an obligate licensing modification for GSDMD pore formation. bioRxiv 531538
[Preprint] (2023). https://doi.org/10.1101/2023.03.07.531538.
7. L. Sborgi, S. Rühl, E. Mulvihill, J. Pipercevic, R. Heilig, H. Stahlberg, C. Farady, D. Müller,
P. Broz, S. Hiller, GSDMD membrane pore formation constitutes the mechanism of
pyroptotic cell death. EMBO J. 35, 1766–1778 (2016).
8. R. A. Aglietti, A. Estevez, A. Gupta, M. G. Ramirez, P. S. Liu, N. Kayagaki, C. Ciferri, V. M. Dixit,
E. C. Dueber, Gsdmd p30 elicited by caspase-­11 during pyroptosis forms pores in
membranes. Proc. Natl. Acad. Sci. U.S.A. 113, 7858–7863 (2016).
9. X. Liu, Z. B. Zhang, J. B. Ruan, Y. D. Pan, V. G. Magupalli, H. Wu, J. Lieberman,
Inflammasome-­activated gasdermin D causes pyroptosis by forming membrane pores.
Nature 535, 153–158 (2016).
10. J. Lieberman, H. Wu, J. C. Kagan, Gasdermin D activity in inflammation and host defense.
Sci. Immunol. 4, eaav1447 (2019).
11. C. L. Wang, J. B. Ruan, Mechanistic insights into gasdermin pore formation and regulation
in pyroptosis. J. Mol. Biol. 434, 16729 (2022).
12. N. Kayagaki, I. B. Stowe, B. L. Lee, K. O'Rourke, K. Anderson, S. Warming, T. Cuellar, B. Haley,
M. Roose-­Girma, Q. T. Phung, P. S. Liu, J. R. Lill, H. Li, J. Wu, S. Kummerfeld, J. Zhang,
W. P. Lee, S. J. Snipas, G. S. Salvesen, L. X. Morris, L. Fitzgerald, Y. Zhang, E. M. Bertram,
C. C. Goodnow, V. M. Dixit, Caspase-­11 cleaves gasdermin D for non-­canonical
inflammasome signalling. Nature 526, 666–671 (2015).
13. W. T. He, H. Q. Wan, L. C. Hu, P. D. Chen, X. Wang, Z. Huang, Z. H. Yang, C. Q. Zhong,
J. H. Han, Gasdermin D is an executor of pyroptosis and required for interleukin-­1β
secretion. Cell Res. 25, 1285–1298 (2015).
14. J. J. Hu, X. Liu, S. Y. Xia, Z. B. Zhang, Y. Zhang, J. X. Zhao, J. B. Ruan, X. M. Luo, X. W. Lou,
Y. Bai, J. H. Wang, L. R. Hollingsworth, V. G. Magupalli, L. Zhao, H. B. R. Luo, J. Kim,
J. Lieberman, H. Wu, FDA-­approved disulfiram inhibits pyroptosis by blocking gasdermin
D pore formation. Nat. Immunol. 21, 736–745 (2020).
15. B. Xu, M. Z. Jiang, Y. Chu, W. J. Wang, D. Chen, X. W. Li, Z. Zhang, D. Zhang, D. M. Fan,
Y. Z. Nie, F. Shao, K. C. Wu, J. Liang, Gasdermin D plays a key role as a pyroptosis
executor of non-­alcoholic steatohepatitis in humans and mice. J. Hepatol. 68, 773–782
(2018).
16. L. P. Du, X. Y. Wang, S. Y. Chen, X. G. Guo, The AIM2 inflammasome: A novel biomarker and
target in cardiovascular disease. Pharmacol. Res. 186, 106533 (2022).
17. K. Bulek, J. Zhao, Y. Liao, N. Rana, D. Corridoni, A. Antanaviciute, X. Chen, H. Wang,
W. Qian, W. A. Miller-­Little, S. Swaidani, F. Tang, B. B. Willard, K. McCrae, Z. Kang,
G. R. Dubyak, F. Cominelli, A. Simmons, T. T. Pizarro, X. Li, Epithelial-­derived gasdermin
D mediates nonlytic IL-­1β release during experimental colitis. J. Clin. Invest. 130,
4218–4234 (2020).
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
18. S. Lu, Y. R. Li, Z. J. Qian, T. S. Zhao, Z. W. Feng, X. G. Weng, L. L. Yu, Role of the
inflammasome in insulin resistance and type 2 diabetes mellitus. Front. Immunol. 14,
1052756 (2023).
19. Y. Ling, M. Xiao, Z. W. Huang, H. Xu, F. Q. Huang, N. N. Ren, C. M. Chen, D. M. Lu, X. M. Yao,
L. N. Xiao, W. K. Ma, Jinwujiangu capsule treats fibroblast-­like synoviocytes of rheumatoid
arthritis by inhibiting pyroptosis via the NLRP3/CAPSES/GSDMD pathway. Evid. Based
Complement. Alternat. Med. 2021, 4836992 (2021).
20. R. Yamagishi, F. Kamachi, M. Nakamura, S. Yamazaki, T. Kamiya, M. Takasugi, Y. Cheng,
Y. Nonaka, Y. Yukawa-­Muto, L. T. T. Thuy, Y. Harada, T. Arai, T. M. Loo, S. Yoshimoto, T. Ando,
M. Nakajima, H. Taguchi, T. Ishikawa, H. Akiba, S. Miyake, M. Kubo, Y. Iwakura, S. Fukuda,
W. Y . Chen, N. Kawada, A. Rudensky, S. Nakae, E. Hara, N. Ohtani, Gasdermin D-­mediated
release of IL-­33 from senescent hepatic stellate cells promotes obesity-­associated
hepatocellular carcinoma. Sci. Immunol. 3, eabl7209 (2022).
21. C. A. Traughber, G. M. Deshpande, K. Neupane, N. Bhandari, M. R. Khan, M. R. McMullen,
S. Swaidani, E. Opoku, S. Muppala, J. D. Smith, L. E. Nagy, K. Gulshan et al., iScience 26,
106076 (2023).
22. H. P. Shen, C. Y. Han, Y. Yang, L. Guo, Y. J. Sheng, J. Wang, W. Y. Li, L. P. Zhai, G. H. Wang,
Q. B. Guan, Pyroptosis executive protein GSDMD as a biomarker for diagnosis and
identification of Alzheimer's disease. Brain Behav. 11, e02063 (2021).
23. B. E. Burdette, A. N. Esparza, H. Zhu, S. Z. Wang, Gasdermin D in pyroptosis. Acta. Pharm.
Sin. B 11, 2768–2782 (2021).
24. X. Liu, S. Y. Xia, Z. B. Zhang, H. Wu, J. Lieberman, Channelling inflammation: Gasdermins in
physiology and disease. Nat. Rev. Drug Discov. 20, 384–405 (2021).
25. J. K. Rathkey, J. J. Zhao, Z. H. Liu, Y. H. Chen, J. Yang, H. C. Kondolf, B. L. Benson,
S. M. Chirieleison, A. Y. Huang, G. R. Dubyak, T. S. Xiao, X. X. Li, D. W. Abbott, Chemical
disruption of the pyroptotic pore-­forming protein gasdermin D inhibits inflammatory
cell death and sepsis. Sci. Immunol. 3, eaat2738 (2018).
26. F. Humphries, L. Shmuel-­Galia, N. Ketelut-­Carneiro, S. Li, B. W. Wang, V. V. Nemmara,
R. Wilson, Z. Z. Jiang, F. Khalighinejad, K. Muneeruddin, S. A. Shaffer, R. Dutta, C. Ionete,
S. Pesiridis, S. Yang, P. R. Thompson, K. A. Fitzgerald, Succination inactivates gasdermin D
and blocks pyroptosis. Science 369, 1633–1637 (2020).
27. M. D. Kornberg, P. Bhargava, P. M. Kim, V. Putluri, A. M. Snowman, N. Putluri, P. A. Calabresi,
S. H. Snyder, Dimethyl fumarate targets GAPDH and aerobic glycolysis to modulate
immunity. Science 360, 449–453 (2018).
28. S. M. Man, R. Karki, T. D. Kanneganti, Molecular mechanisms and functions of pyroptosis,
inflammatory caspases and inflammasomes in infectious diseases. Immunol. Rev. 277,
61–75 (2017).
29. E. Anscombe, E. Meschini, R. Mora-­Vidal, M. P. Martin, D. Staunton, M. Geitmann,
U. H. Danielson, W. A. Stanley, L. Z. Wang, T. Reuillon, B. T. Golding, C. Cano, D. R. Newell,
M. E. M. Noble, S. R. Wedge, J. A. Endicott, R. J. Griffin, Identification and characterization
of an irreversible inhibitor of CDK2. Chem. Biol. 22, 1159–1164 (2015).
30. J. C. Powers, J. L. Asgian, O. D. Ekici, K. E. James, Irreversible inhibitors of serine, cysteine,
and threonine proteases. Chem. Rev. 102, 4639–4750 (2002).
31. B. D. Mather, K. Viswanathan, K. M. Miller, T. E. Long, Michael addition reactions in
macromolecular design for emerging technologies. Prog. Polym. Sci. 31, 487–531 (2006).
32. T. Wolfram, M. Schwarz, M. Reuss, K. Lossow, M. Ost, S. Klaus, T. Schwerdtle, A. P. Kipp,
N-­Acetylcysteine as modulator of the essential trace elements copper and zinc.
Antioxidants 9, 1117 (2020).
33. S. Kesavardhana, R. K. S. Malireddi, T. D. Kanneganti, Caspases in cell death, inflammation,
and pyroptosis. Annu. Rev. Immunol. 38, 567–595 (2020).
34. K. S. G. de Sa, L. A. Amaral, T. S. Rodrigues, A. Y. Ishimoto, W. A. C. de Andrade,
L. de Almeida, F. Freitas-­Castro, S. S. Batah, S. C. Oliveira, M. T. Pastorello, A. T. Fabro,
D. S. Zamboni, Gasdermin-­D activation promotes NLRP3 activation and host resistance to
leishmania infection. Nat. Commun. 14, 1049 (2023).
35. R. Medzhitov, Origin and physiological roles of inflammation. Nature 454, 428–435
(2008).
36. F. Van Hauwermeiren, M. Lamkanfi, The NEK-­sus of the NLRP3 inflammasome. Nat.
Immunol. 17, 223–224 (2016).
37. C. M. Ma, D. X. Yang, B. W. Wang, C. Y. Wu, Y. Q. Wu, S. Li, X. Liu, K. Lassen, L. Dai, S. Yang,
Gasdermin D in macrophages restrains colitis by controlling cGAS-­mediated
inflammation. Sci. Adv. 6, eaaz6717 (2020).
38. W. C. Yang, K. X. Tao, Y. X. Wang, Y. Z. Huang, C. H. Duan, T. Wang, C. G. Li, P. Zhang, Y. P. Yin,
J. B. Gao, R. D. Li, Necrosulfonamide ameliorates intestinal inflammation via inhibiting
GSDMD-­medicated pyroptosis and MLKL-­mediated necroptosis. Biochem. Pharmacol.
206, 115338 (2022).
39. R. F. Shao, X. R. Lou, J. F. Xue, D. Y. Ning, G. B. Chen, L. H. Jiang, Review: The role of GSDMD
in sepsis. Inflamm. Res. 71, 1191–1202 (2022).
40. C. Q. Wu, W. Lu, Y. Zhang, G. Y. Zhang, X. Y. Shi, Y. Hisada, S. P. Grover, X. Y. Zhang, L. Li,
B. G. Xiang, J. M. Shi, X. A. Li, A. Daugherty, S. S. Smyth, D. Kirchhofer, T. Shiroishi, F. Shao,
N. Mackman, Y. N. Wei, Z. Y. Li, Inflammasome activation triggers blood clotting and host
death through pyroptosis. Immunity 50, 1401–1411.e4 (2019).
41. K. Schroder, J. Tschopp, The inflammasomes. Cell 140, 821–832 (2010).
16 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
42. K. V. Swanson, M. Deng, J. P. Y. Ting, The NLRP3 inflammasome: Molecular activation and
regulation to therapeutics. Nat. Rev. Immunol. 19, 477–489 (2019).
43. N. Kayagaki, S. Warming, M. Lamkanfi, L. Vande Walle, S. Louie, J. Dong, K. Newton, Y. Qu,
J. F. Liu, S. Heldens, J. Zhang, W. P. Lee, M. Roose-­Girma, V. M. Dixit, Non-­canonical
inflammasome activation targets caspase-­11. Nature 479, 117–121 (2011).
44. P. Broz, V. M. Dixit, Inflammasomes: Mechanism of assembly, regulation and signalling.
Nat. Rev. Immunol. 16, 407–420 (2016).
45. L. Boike, N. J. Henning, D. K. Nomura, Advances in covalent drug discovery. Nat. Rev. Drug
Discov. 21, 881–898 (2022).
46. S. Y. Chen, S. Lovell, S. M. Lee, M. Fellner, P. D. Mace, M. Bogyo, Identification of highly
selective covalent inhibitors by phage display. Nat. Biotechnol. 39, 490–498 (2021).
47. K. Bum-­Erdene, D. G. Liu, G. Gonzalez-­Gutierrez, M. K. Ghozayel, D. Xu, S. O. Meroueh,
Small-­molecule covalent bond formation at tyrosine creates a binding site and inhibits
activation of Ral GTPases. Proc. Natl. Acad. Sci. U.S.A. 117, 7131–7139 (2020).
48. A. Y. Hsu, D. C. Wang, S. Liu, J. Lu, R. Syahirah, D. A. Bennin, A. Huttenlocher, D. M. Umulis,
J. Wan, Q. Deng, Phenotypical microRNA screen reveals a noncanonical role of CDK2 in
regulating neutrophil migration. Proc. Natl. Acad. Sci. U.S.A. 116, 18561–18570 (2019).
49. B. Lomenick, R. Hao, N. Jonai, R. M. Chin, M. Aghajan, S. Warburton, J. N. Wang, R. P. Wu,
F. Gomez, J. A. Loo, J. A. Wohlschlegel, T. M. Vondriska, J. Pelletier, H. R. Herschman,
J. Clardy, C. F. Clarke, J. Huang, Target identification using drug affinity responsive target
stability (DARTS). Proc. Natl. Acad. Sci. U.S.A. 106, 21984–21989 (2009).
50. Y. F. Zhong, K. Lee, Y. Y. Deng, Y. M. Ma, Y. P. Chen, X. L. Li, C. G. Wei, S. M. Yang, T. M. Wang,
N. J. Wong, A. N. Muwonge, E. U. Azeloglu, W. J. Zhang, B. Das, J. C. J. He, R. J. Liu,
Arctigenin attenuates diabetic kidney disease through the activation of PP2A in
podocytes. Nat. Commun. 10, 45232 (2019).
51. W. Yang, B. Mu, J. You, C. Tian, H. Bin, Z. Xu, L. Zhang, R. Ma, M. Wu, G. Zhang, C. Huang,
L. Li, Z. Shao, L. Dai, L. Désaubry, S. Yang, Non-­classical ferroptosis inhibition by a small
molecule targeting PHB2. Nat. Commun. 13, 7473 (2022).
52. Z. Wang, Y. Sun, F. Lou, J. Bai, H. Zhou, X. Cai, L. Sun, Q. Yin, S. Tang, Y. Wu, L. Fan, Z. Xu,
H. Wang, X. Hu, H. Wang, Targeting the transcription factor HES1 by L-­menthol restores
protein phosphatase 6 in keratinocytes in models of psoriasis. Nat. Commun. 13, 7815 (2022).
53. P. Devant, E. Borsic, E. M. Ngwa, H. Xiao, E. T. Chouchani, J. R. Thiagarajah,
I. Hafner-­Bratkovic, C. L. Evavold, J. C. Kagan, Gasdermin D pore-­forming activity is
redox-­sensitive. Cell Rep. 42, 112008 (2023).
Jiang et al., Sci. Adv. 10, eadi9284 (2024) 7 February 2024
54. C. L. Evavold, I. Hafner-­Bratkovic, P. Devant, J. M. D'Andrea, E. M. Ngwa, E. Borsic,
J. G. Doench, M. W. LaFleur, A. H. Sharpe, J. R. Thiagarajah, J. C. Kagan, Control of
gasdermin D oligomerization and pyroptosis by the Ragulator-­Rag-­mTORC1 pathway.
Cell 184, 4495–4511.e19 (2021).
55. S. Y. Xia, Z. B. Zhang, V. G. Magupalli, J. L. Pablo, Y. Dong, S. M. Vora, L. F. Wang, T. M. Fu,
M. P. Jacobson, A. Greka, J. Lieberman, J. B. Ruan, H. Wu, Gasdermin D pore structure
reveals preferential release of mature interleukin-­1. Nature 593, 607–611 (2021).
56. Y. S. Xue, D. E. Tuipulotu, W. H. Tan, C. Kay, S. M. Man, Emerging activators and regulators
of inflammasomes and pyroptosis. Trends Immunol. 40, 1035–1052 (2019).
57. D. P. Zheng, T. Liwinski, E. Elinav, Inflammasome activation and regulation: Toward a
better understanding of complex mechanisms. Cell Discov. 6, 36 (2020).
58. G. T. Heller, F. A. Aprile, T. C. T. Michaels, R. Limbocker, M. Perni, F. S. Ruggeri, B. Mannini,
T. Lohr, M. Bonomi, C. Camilloni, A. De Simone, I. C. Felli, R. Pierattelli, T. P. J. Knowles,
C. M. Dobson, M. Vendruscolo, Small-­molecule sequestration of amyloid-­beta as a drug
discovery strategy for Alzheimer's disease. Sci. Adv. 6, eabb5924 (2020).
Acknowledgments: We thank E. Kong (Institute of Psychiatry and Neuroscience, Xinxiang
Medical University) for the palmitoylation research. Funding: This study was supported by the
National Natural Science Foundation of China (82174077 and 82272647), the Science and
Technology Department of Sichuan Province (2023NSFSC1772, 24QYCX0406, and
24NSFSC6854), the Foundation of Sichuan University (2018SCUH0067), and Science and
Technology Achievement Transformation Fund of West China Hospital, Sichuan University
(CGZH19010). Author contributions: Conceptualization: H.Y., J.Y., W.W., X.J., and N.L.
Methodology: X.J., X.Z., X.C., N.L., H.Z., M.T., J.Z., and K.S. Investigation: X.J., X.Z., X.C., R.Z., N.Y.,
J.P., and M.Z. Formal analysis: X.J., X.Z., X.C., and N.L. Visualization: X.J. and W.W. Supervision:
H.Y., J.Y., and W.W. Writing—original draft: X.J., X.Z., X.C., and N.L. Writing—review and editing:
X.J., H.Y., J.Y., and W.W. Competing interests: The authors declare that they have no
competing interests. Data and materials availability: All data needed to evaluate the
conclusions in the paper are present in the paper and/or the Supplementary Materials.
Submitted 26 May 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adi9284
17 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

PHYSIOLOGY Copyright © 2024 The

Authors, some rights
AMPK phosphorylation of FNIP1 (S220) controls reserved; exclusive
licensee American
mitochondrial function and muscle fuel utilization Association for the
Advancement of
during exercise Science. No claim to
original U.S.
Government Works.
Liwei Xiao1†, Yujing Yin1†, Zongchao Sun1†, Jing Liu1†, Yuhuan Jia1, Likun Yang1, Yan Mao1, Distributed under a
Shujun Peng2, Zhifu Xie3, Lei Fang4, Jingya Li3, Xiaoduo Xie2*, Zhenji Gan1* Creative Commons
Attribution
Exercise-­induced activation of adenosine monophosphate–activated protein kinase (AMPK) and substrate phos- NonCommercial
phorylation modulate the metabolic capacity of mitochondria in skeletal muscle. However, the key effector(s) of License 4.0 (CC BY-­NC).
AMPK and the regulatory mechanisms remain unclear. Here, we showed that AMPK phosphorylation of the fol-
liculin interacting protein 1 (FNIP1) serine-­220 (S220) controls mitochondrial function and muscle fuel utilization
during exercise. Loss of FNIP1 in skeletal muscle resulted in increased mitochondrial content and augmented
metabolic capacity, leading to enhanced exercise endurance in mice. Using skeletal muscle–specific nonphos-
phorylatable FNIP1 (S220A) and phosphomimic (S220D) transgenic mouse models as well as biochemical analysis
in primary skeletal muscle cells, we demonstrated that exercise-­induced FNIP1 (S220) phosphorylation by AMPK
in muscle regulates mitochondrial electron transfer chain complex assembly, fuel utilization, and exercise perfor-
mance without affecting mechanistic target of rapamycin complex 1–transcription factor EB signaling. Therefore,
FNIP1 is a multifunctional AMPK effector for mitochondrial adaptation to exercise, implicating a mechanism for
exercise tolerance in health and disease.

INTRODUCTION
Physical exercise, as the cornerstone of a healthy human lifestyle, trig-
gers cascades of physiological responses in the skeletal muscle, lead-
ing to enhanced energy expenditure and metabolic adaptations in all
vertebrates (1, 2). Mitochondria, often referred to as the “powerhouses
of the cell,” are the major cellular organelle that acts as an integrated
signal platform to control cellular metabolism in muscle fibers for en-
ergy generation and metabolic adaptation during exercise (3, 4). The
metabolic capacity and biogenesis of mitochondria are greatly pro-
moted within cells as physical activity increases. One of the key me-
diators involved in this process is the adenosine monophosphate
(AMP)–activated protein kinase (AMPK), an enzyme known as the
cellular energy sensor or mitochondria guardian (5, 6). During exer-
cise, the depletion of cellular energy, namely, adenosine triphosphate
(ATP), and the subsequent increase in the AMP/ATP ratio activate
AMPK, which phosphorylates effector proteins to promote glucose
uptake and fatty acid oxidation, achieving a condition of energy bal-
ance and cell homeostasis (5–7). In addition to its well-­known role in
mediating mitochondrial metabolic adaptation via transcription fac-
tor–mediated mitochondrial biogenesis, AMPK also modulates mi-
tochondrial activities through direct phosphorylation of multiple
mitochondrial proteins to control oxidative phosphorylation (8–10).
1
State Key Laboratory of Pharmaceutical Biotechnology and MOE Key Laboratory
of Model Animal for Disease Study, Model Animal Research Center, Division of
Spine Surgery, Department of Orthopedic Surgery, Nanjing Drum Tower Hospital,
The Affiliated Hospital of Nanjing University Medical School, Jiangsu Key Labora-
tory of Molecular Medicine, Chemistry and Biomedicine Innovation Center (ChemBIC),
Medical School of Nanjing University, Nanjing University, Nanjing, China. 2School
of Medicine, Shenzhen Campus of Sun Yat-­sen University, Sun Yat-­sen University,
Shenzhen, China. 3Shanghai Institute of Materia Medica, Chinese Academy of Sci-
ences, Shanghai, China. 4Jiangsu Key Laboratory of Molecular Medicine & Chemis-
try and Biomedicine Innovation Center, Medical School of Nanjing University,
Nanjing, China.
*Corresponding author. Email: ganzj@​nju.​edu.​cn (Z.G.); xiexd8@​mail.​sysu.​edu.​cn
(X.X.)
†These authors contributed equally to this work.
Drugs activate AMPK, act as exercise mimetics, and drive a mito-
chondrial oxidative program in the skeletal muscle (11–13). Exercise
training activates AMPK and improves many chronic disease condi-
tions through mitochondrial remodeling in the skeletal muscle (3, 8,
14, 15). Conversely, dysregulation of AMPK signaling and subse-
quent mitochondrial dysfunction in the skeletal muscle are frequent-
ly observed in the pathogenesis of human diseases, including obesity,
type 2 diabetes, and muscular dystrophy/atrophy (3, 4, 16, 17).
Accumulated evidence indicates that AMPK mediates exercise-­
induced mitochondrial modulation through phosphorylation of key
transcription factors such as transcription factor EB (TFEB) and per-
oxisome proliferation–activated receptor (PPAR)–γ coactivator 1α
(PGC1α), which subsequently induces gene expression for mito-
chondrial biogenesis and quality control through coordination with
estrogen-­related receptors (ERRs) or PPAR nuclear receptors (18–
21). In addition to de novo mitochondrial biogenesis, exercise is also
effective in promoting favorable remodeling of the muscle mito-
chondrial proteome to coordinate the requirements of efficient mito-
chondrial biogenesis, reorganization, and assembly (22–24). For
instance, mitochondrial electron transfer chain (ETC) respiratory
supercomplex assemblies increase in response to exercise in the skel-
etal muscle (22). However, it remains unclear whether there are oth-
er AMPK substrates that are directly involved in mitochondrial ETC
function, such as regulation of supercomplex assembly or regulation
of ETC complex activity. Folliculin (FLCN) interacting protein 1
(FNIP1) is an adaptor protein originally identified as a binding part-
ner for FLCN and AMPK (25). Previous studies have placed AMPK
downstream of FLCN/FNIP1, which functional as a guanosine tri-
phosphate–activating protein (GAP) for the RagC/D guanosine tri-
phosphatase (GTPase) to sense amino acids in mechanistic target of
rapamycin complex 1 (mTORC1) signaling (26–29). However, the
exact roles of FLCN/FNIP1 in sensing energy status in the AMPK
pathway or in sensing nutrients in mTOR signaling remain unclear.
Our previous studies indicated that Fnip1, as a target of the MyomiR
miR-­499, regulates mitochondrial function and muscle fiber type

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 1 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
switching in an AMPK/PGC1α-­dependent and -­independent man-
ner in myocytes and in skeletal muscle of mouse models (30–32).
Consistently, FLCN or FNIP1 knockout mouse models have been
shown to increase mitochondrial function and oxidative metabo-
lism in the skeletal muscle, and FNIP1 also affects metabolic pro-
cesses in other tissues (26, 33–35). Collectively, these studies suggest
a general role for FNIP1 in metabolic regulation. Nevertheless, how
FNIP1 mediates such a wide effect on metabolism and the exact
relationship between FNIP1 and AMPK or mTORC1 signaling remain
unclear. Deletion of FNIP1 led to both activation and inhibition of
AMPK phosphorylation (33, 34), and, similarly, both suppression
and activation of mTOR signaling by FNIP1 have also been reported
(29, 33, 34, 36). One of the common effectors of AMPK and
mTORC1 is TFEB, which mediates gene expression to control
lysosome biogenesis and autophagy in response to nutrient de-
privation and energetic stress (18, 37). AMPK and mTORC1 are
known to directly phosphorylate TFEB at different sites with op-
posing effects on TFEB nuclear translocation and transcription
(19, 37, 38). FNIP1 was also reported to be involved in this pro-
cess (18, 29, 39), and it seems that there is complex reciprocal
regulation among AMPK, FNIP1, and mTORC1 to control TFEB
translocation.
Using muscle-­specific gain-­and loss-­of-­function mouse models, we
investigated the physiological role of FNIP1 in mediating AMPK signaling
in the skeletal muscle, and it was revealed that FNIP1 was directly phos-
phorylated by AMPK at S220 in response to exercise training, which
appears to play an important role in the regulation of mitochondrial pro-
teomics and related metabolic regulation, according to genetic and bio-
chemical evidence. Notably, AMPK phosphorylation of FNIP1 (S220)
acts in parallel to AMPK-­FNIP1-­TFEB mitochondrial biogenesis signal-
ing to optimize mitochondrial fuel utilization, enhancing ATP produc-
tion and exercise capacity. Thus, FNIP1 is a multifunctional AMPK
effector for mitochondrial adaptation to exercise. These findings impli-
cate a target for exercise tolerance in the health and disease of
the skeletal muscle.
RESULTS
AMPK directly phosphorylates FNIP1 (S220) in vitro
and in vivo
AMPK and its effector phosphorylation are essential for mitochon-
drial regulation under energetic stress conditions, but the full reper-
toire of AMPK biological targets related to mitochondria in the skeletal
muscle in response to exercise remains unclear. Previous studies re-
ported that FNIP1, a key mitochondrial regulator, preferentially binds
to phosphorylated AMPK and regulates AMPK activity (25). Howev-
er, we observed a comparable FNIP1 binding capability with total
AMPK or phosphorylated AMPK upon treatment with the AMPK
activator A-­769662 (fig. S1A). We further tested whether FNIP1 is an
AMPK phosphorylation effector, as it contains a highly conserved op-
timal AMPK phosphorylation consensus (Fig. 1A). FNIP1 immuno-
precipitated from human embryonic kidney (HEK) 293T cell lysate
was phosphorylated at the S220 residue, as identified by phosphopep-
tide mass spectrometry (MS) (Fig. 1B). Specific antibody was gener-
ated against pS220 of FNIP1 [pFNIP1 (S220)], which recognizes
recombinant AMPK-­phosphorylated wild-­type (WT) FNIP1 but not
the nonphosphorylatable mutant FNIP1 (S220A) in vitro (Fig. 1C).
An increase in pFNIP1 (S220) signals was induced with FNIP1 (WT)
but not FNIP1 (S220A) in HEK293T cells treated with A-­769662
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
(Fig. 1D). These data validated the effectiveness of the pFNIP1 (S220)
antibody and the authentication of FNIP1 (S220) phosphorylation
by AMPK. Furthermore, the A-­769662–induced endogenous FNIP1
(S220) phosphorylation requires AMPK, as it was abolished in mouse
primary myotubes isolated from AMPKα1/α2 double knockout mice
(Fig. 1E) (31), and this phosphorylation could also be stimulated in
myocytes by various AMPK activators, such as 5-­Aminoimidazole-­4-­
carboxamide-­1-­beta-­4-­ribofuranoside (AICAR), glucose starvation,
and A-­769662 (Fig. 1F and fig. S1, B and C). S220 phosphorylation was
increased with AMPK phosphorylation stimulated by exercise train-
ing in muscle tissue from muscle-­specific WT FNIP1–overexpressing
(WT FNIP1 Tg) mice (Fig. 1G), implying that this phosphorylation is
functionally important for the AMPK-­ mediated physiological re-
sponse to exercise. We further tested whether the phosphorylation af-
fected the AMPK-­FNIP1 interaction; although S220-­phosphorylated
FNIP1 was structurally different from WT as predicted by AlphaFold2
and SWISS-­MODEL modeling (fig. S1, D and E), both FNIP1 (S220A)
and FNIP1 (S220D) exhibited no difference in AMPK binding com-
pared to FNIP1 (WT) (fig. S1F). FNIP1/FLCN complex functions as a
GAP for RagC and RagD GTPases and acts as an amino acid sensor
for amino acid–induced mTORC1 activation (28, 29); however, Fnip1
knockdown by small interfering RNA (siRNA) in HEK293T cells did
not affect amino acid–induced S6K phosphorylation, a canonical indi-
cator of mTORC1 activity (fig. S1G). Neither FNIP1 (S220A) nor
FNIP1 (S220D) rescue in the FNIP1-­deficient cells had any effect on
S6K phosphorylation induced either by amino acid or glucose (fig. S1,
G and H), suggesting that FNIP1 (S220) phosphorylation by AMPK
does not affect nutrient-­induced mTORC1 activity. Together, these
data strongly suggest that FNIP1 is a constitutive AMPK binding part-
ner and a bona fide phosphorylation substrate.
Skeletal muscle–specific deletion of FNIP1 leads to
enhanced exercise capacity
Despite FNIP1’s emerging role in the regulation of mitochondrial
and muscle fiber–type programming (31, 33), it remains unclear
whether and how FNIP1 contributes to the metabolic adaptation
and plasticity of skeletal muscle during physical exercise. To address
this, we crossed Fnip1-­floxed mice (Fnip1f/f, exon 6 flanked by two
loxP sites) with human skeletal actin Cre mice to create skeletal mus-
cle–specific FNIP1-­knockout (FNIP1 mKO) mice. FNIP1 protein
and mRNA expression was markedly reduced in muscle, but not in
other tissues, in FNIP1 mKO mice compared to WT littermates
(Fig. 2A and fig. S2, A and B), while the protein expression of FNIP2,
a close homolog of FNIP1, was increased in FNIP1 mKO muscle
relative to WT littermates (Fig. 2A and fig. S2C). To determine the
physiological impact of FNIP1 deficiency in muscle, we assessed the
acute running endurance performance of the FNIP1 mKO mice using
a run-­to-­exhaustion protocol on a motorized treadmill. Notably,
FNIP1 mKO mice could run for longer times and distances (~40%)
than their WT littermates (Fig. 2B). A forced maximal exercise ca-
pacity test revealed that FNIP1 mKO mice ran longer distances and
achieved higher maximal speeds than WT controls, despite a similar
energy substrate utilization indicated by the respiratory exchange
ratio (RER) during the course of exercise (Fig. 2, C to E). Consistent
with the enhanced exercise capacity, FNIP1 mKO mice consumed
more oxygen (as reflected by peak VO2) with higher maximal peak
oxygen consumption (VO2 max) than WT controls during the ex-
ercise period (Fig. 2, F and G); postexercise blood lactate levels,
an indicator of glycolysis, were decreased in FNIP1 mKO mice
2 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. AMPK phosphorylates FNIP1 (S220) in vitro and in vivo. (A) Clustal alignment of AMPK substrate consensus sequences and the conserved S220 phosphoryla-
tion site on FNIP1. (B) FNIP1 phosphorylation peptide identification by MS analysis. Flag-­FNIP1 was immunoprecipitated (IP), and the phosphopeptide containing
phosphorylated S220 was identified by liquid chromatography–MS/MS. m/z, mass/charge ratio. (C) AMPK in vitro kinase assays with FNIP1. Flag-­FNIP1 (WT) or (S220A)
IPs from compound C-­pretreated HEK293T cells were used as substrates for recombinant (Rec) AMPK; phosphorylation was detected by the FNIP1 (S220) antibody. CTL,
control. (D) FNIP1 (S220) phosphorylation in response to AMPK activation in HEK293T cells. HEK293T cells transfected with Flag-­FNIP1 (WT) or (S220A) were treated with
vehicle or 100 μM A-­769662 for 1 hour. (E and F) AMPK-­mediated FNIP1 (S220) phosphorylation in primary myotubes. (E) Top: Primary myotubes from WT or AMPKα1/
α2 double knockout (AKO) muscles were treated with either vehicle or 300 μM A-­769662 for 1 hour and immunoblotted with the indicated antibodies. Bottom: Quanti-
fication of FNIP1 (S220) phosphorylation by pFNIP1/FNIP1 signal ratios. AU, arbitrary units. (F) Primary myotubes were treated with AICAR (2 mM; for 1 hour) or glucose
starvation (GS) for 2 hours. Lysates were immunoblotted with the indicated antibodies. (G) Top: Representative immunoblots with white vastus lateralis (WV) muscle
total protein extracted from acutely exercised (Exe) WT FNIP1 Tg mice with the indicated antibodies. Bottom: Quantification of FNIP1 (S220) and AMPKα (T172) phos-
phorylation by pAMPK/AMPK and pFNIP1/FNIP1 signal ratios. n = 4 mice per group. Error bars are shown as the means ± SEM. Experiments (C) to (F) were repeated at
least three times. *P < 0.05; **P < 0.01; ***P < 0.001. The P value was determined by one-­way analysis of variance (ANOVA) coupled to Fisher’s least significant difference
(LSD) post hoc test.

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 3 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Deletion of muscle FNIP1 enhances exercise capacity in mice. (A) Representative immunoblotting analysis of FNIP1 and FNIP2 expression in tissue from indi-
cated mice. n = 4 to 5 mice per group. (B) Exercise endurance test in FNIP1 mKO mice. Left: A schematic depicting increments of speed over time. Middle and right: Bars
represent the running time and distance for 8-­week-­old male mice on a motorized treadmill. n = 11 mice per group. (C and D) Exercise capacity test in FNIP1 mKO mice.
(C) The schematic depicts increments of speed over time. (D) The RER in FNIP1 mKO mice. n = 7 mice per group. (E) Maximal speed and total running distance in FNIP1
mKO mice. n = 7 mice per group. (F) VO2 (oxygen consumption) during exercise. The gray-­shaded area under the VO2 line illustrates the difference in speed to exhaustion
in FNIP1 mKO mice compared to WT controls. n = 7 mice per group. (G) Peak VO2 during exercise in FNIP1 mKO mice. n = 6 to 7 mice per group. (H) Blood lactate levels in
indicated mice after a 25-­min exercise. n = 11 mice per group. (I to L) Blood glucose (I), TG (J), NEFA (K), and β-­hydroxybutyrate (β-­HB) (L) levels were determined in indi-
cated mice at rest (Sed) or after 150-­min exercise (Ex150). n = 4 to 5 mice per group. (M) Representative electron micrographs of soleus muscle showing mitochondria in
sections from indicated mice. n = 3 mice per group. Scale bars, 5.0 μm. Error bars are shown as the means ± SEM. *P < 0.05; **P < 0.01; ***P < 0.001. The P value was
determined by Student’s t test [(B), (D), (E), (G), and (H)] or one-­way ANOVA coupled to Fisher’s LSD post hoc test [(I) to (L)].

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 4 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(Fig. 2H); higher blood glucose, lower blood triglyceride (TG), and
nonesterified fatty acid (NEFA) levels were observed after 150 min
of endurance exercise in FNIP1 mKO mice (Fig. 2, I to K). Consis-
tently, blood ketone body β-­hydroxybutyrate levels were decreased
in FNIP1 mKO mice under both basal and postexercise conditions
(Fig. 2L), suggesting an enhancement of mitochondrial metabolism
and exercise capacity. Such an “athletic” phenotype was reflected by
the ultrastructure change of mitochondria in the muscle, as we ob-
served by electron microscopy. The mitochondrial area was promi-
nently increased in the FNIP1 mKO muscles compared to WT
control muscles. Moreover, the amount of intramyocellular lipid
droplets increased, indicating an increased reliance on lipids as en-
ergy substrates during running in FNIP1 mKO muscles (Fig. 2M).
Together, these results demonstrated a change in metabolic flexibil-
ity with enhanced mitochondrial activity and exercise endurance in
FNIP1 mKO mice.
Exercise-­induced FNIP1 (S220) phosphorylation controls
mitochondrial metabolism and exercise performance in mice
To pinpoint the potential function of FNIP1 (S220) phosphorylation
by AMPK in mitochondrial regulation in the skeletal muscle, muscle-­
specific transgenic (Tg) mice with FNIP1 (S220A) and FNIP1 (S220D)
expressed under the control of the muscle creatine kinase promoter
(hereafter indicated as S220A Tg and S220D Tg) were established
using the same strategy as that of WT FNIP1 Tg lines (Fig. 3A) (31).
These three Fnip1 Tg lines (WT FNIP1 Tg, S220A Tg, and S220D Tg)
exhibited similar Fnip1 mRNA levels in muscle tissues with 80-­
to 120-­fold changes compared to the nontransgenic (NTG) littermate
control mice (Fig. 3B). Consistently, immunoblotting experiments
confirmed that the Flag-­FNIP1 protein was overexpressed at equiva-
lent levels in the muscles of all three mouse lines (Fig. 3, C and D).
To examine the relevance of FNIP1 (S220) phosphorylation in exer-
cise physiology, we performed treadmill-­running tests comparing
WT FNIP1 Tg, S220A Tg, S220D Tg, and NTG littermates. S220A Tg
mice displayed a reduction in running time and distance, while there
were no differences for S220D Tg or WT FNIP1 Tg relative to NTG
controls (Fig. 3E). Consistent with reduced exercise tolerance, S220A
Tg mice, but not WT FNIP1 Tg mice, showed higher levels (~35%) of
blood lactate after exercise, indicating increased glycolysis (Fig. 3F).
To further evaluate muscle fuel utilization during exercise, NTG
or S220A Tg mice were also subjected to a forced maximal exercise
capacity test (VO2max test) consisting of increasing speed every 2 min
until exhaustion. S220A Tg mice had a higher RER than NTG con-
trols during exercise, and this was accompanied by reduced exercise
tolerance (Fig. 3, G and H). In addition, upon high-­intensity exercise
challenge, S220A Tg mice showed decreased maximal speed and dis-
tance compared to their NTG control littermates (Fig. 3I). This is con-
sistent with observations that S220A Tg mice consumed less oxygen
during the exercise period (as reflected by peak ΔVO2) than NTG
controls (Fig. 3J). Furthermore, S220A Tg mice showed lower blood
glucose levels than NTG controls under basal conditions and after
60 min of endurance exercise (Fig. 3K), while blood TG was higher
under basal conditions and after exercise in S220A Tg mice (Fig. 3L).
However, there were no changes in blood glucose and TG levels in
WT FNIP1 Tg mice relative to NTG controls under basal and postex-
ercise conditions (Fig. 3, K and L). There were no changes in fatty acid
levels in either WT FNIP1 Tg mice or S220A Tg mice relative to NTG
controls (Fig. 3M). Notably, blood ketone body β-­hydroxybutyrate
levels, mirroring the oxidation of blood fatty acids, also increased
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
in S220A Tg but not in WT FNIP1 Tg mice after exercise (Fig. 3N).
Muscle glycogen is well known as a key energy source for contraction
and is crucial for sustaining prolonged exercise. We found that glyco-
gen content in the muscle of FNIP1 mKO mice was higher than that
of WT mice, while it decreased in S220A Tg mice and was not changed
in S220D Tg or WT FNIP1 Tg mice compared to their control litter-
mates either by periodic acid–Schiff staining or glycogen content
measurement (fig. S3, A and B), indicating that FNIP1 and its S220
phosphorylation by AMPK might contribute to the regulation of
muscle glycogen metabolism. Together, these results demonstrate
that FNIP1 (S220) phosphorylation by AMPK links metabolic flex-
ibility to exercise performance through the promotion of mitochon-
drial function.
FNIP1 (S220) phosphorylation controls mitochondrial ETC
complex formation without affecting
mTORC1-­TFEB signaling
To delineate how FNIP1 (S220) phosphorylation regulates mito-
chondrial function by exercise-­induced AMPK activation in muscle
tissue, we performed transcriptome analysis by whole-­genome gene
expression profiling in the gastrocnemius (GC) muscle from FNIP1
mKO and WT control mice, S220A Tg, S220D Tg, WT FNIP1 Tg,
and NTG control mice, respectively. A total of 3550 differentially
expressed genes (DEGs) were detected in FNIP1 mKO muscle com-
pared to WT controls (Fig. 4A and fig. S3, C and D), including mito-
chondrial or lysosomal biogenesis signature genes related to PGC1α
and TFEB (Fig. 4A); however, no such DEGs were detected in S220A
Tg, S220D Tg, and WT FNIP1 Tg muscle relative to NTG controls
(Fig. 4A and fig. S4A). This result suggested that FNIP1 (S220) phos-
phorylation by AMPK did not affect TFEB-­related gene expression
in muscles as the whole FNIP1 knockout. A recent study reported
that mTORC1-­TFEB-­PGC1α signaling is important for AMPK-­
FNIP1–mediated mitochondrial biogenesis (18). To clarify if FNIP1
(S220) phosphorylation was involved in this process, we tested
AMPK and mTORC1-­TFEB signaling in S220A Tg muscle tissues.
The AMPK activity was comparable in S220A Tg and NTG muscle
either in sedentary or in exercised conditions (Fig. 4B); FNIP1
(S220A Tg) muscle exhibited normal mTORC1 activity toward ca-
nonical substrates S6K; nuclear translocation of TFEB, a noncanoni-
cal mTORC1 substrate (40), was also not changed between S220A
Tg and NTG muscle tissues by TFEB histochemical staining at both
baseline and exercise conditions (Fig. 4, C to H). Neither phospho-­
TFEB (S122) nor total TFEB band shift (indicative of TFEB overall
phosphorylation levels) were changed in both WT FNIP1 Tg and
S220A Tg muscle at baseline or exercise conditions compared to the
NTG control (Fig. 4, F to H). These data collectively suggest that
FNIP1 (S220) phosphorylation is likely not directly involved in the
regulation of mTORC1-­TFEB–mediated gene expression in the skel-
etal muscle. Nevertheless, the regulatory effects of FNIP1 (S220)
phosphorylation on mitochondria in the muscle were prominent, as
both pyruvate-­and succinate-­driven mitochondrial respiration rates
were markedly induced in the muscle of FNIP1 mKO mice com-
pared to WT controls. On the contrary, FNIP1 (S220A) suppressed
mitochondrial respiration rates, while there was no difference in
FNIP1 (S220D Tg) or WT FNIP1 Tg muscle (Fig. 5, A and B), indi-
cating that S220 phosphorylation is required for normal mitochon-
drial respiration rates.
Recent studies have revealed that AMPK can localize to mitochon-
dria in the skeletal muscle (41). To pinpoint the potential mechanism,
5 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. FNIP1 (S220) phosphorylation by AMPK regulates mitochondrial function and exercise performance in mice. (A) Schematic diagram depicting the Mck-­
driven WT Fnip1, S220A, and S220D transgenes. HGH, human growth hormone. (B) Quantitative reverse transcription polymerase chain reaction (RT-­PCR) analysis of
mRNA levels of Fnip1 transgenes from the gastrocnemius (GC) muscle of the indicated transgenic (Tg) mice. n = 5 to 12 mice per group. (C) Endogenous and Tg FNIP1
expression in muscle tissues. Representative immunoblotting analysis of protein extracts from the WV muscles of the indicated Tg mice with the indicated antibodies.
n = 3 mice per group. (D) Quantification of FNIP1 by signal ratios normalized to tubulin. (E) Exercise endurance test in Fnip1 Tg mice. Left: A schematic depicting incre-
ments of speed over time. Middle and right: Bars represent the mean running time and distance for the indicated Tg mice on a motorized treadmill. n = 6 to 13 mice per
group. (F) Blood lactate levels in WT FNIP1 Tg, S220A Tg, and NTG controls. n = 3 to 13 mice per group. (G to J) Exercise capacity test for Fnip1 Tg mice. n = 6 to 10 mice per
group. The schematic depicts the increments of speed over time (G). The RER during a graded exercise regimen (H), Maximal speed and total distance (I) as well as peak
VO2 and peak ΔVO2 during exercise (J). (K to N) Blood glucose (K), TG (L), NEFA (M), and β-­HB (N) levels at rest (Sed) or after 60 min of exercise (Ex60). n = 4 to 9 mice per
group. Error bars are shown as the means ± SEM. *P < 0.05; **P < 0.01; ***P < 0.001. NS, not significant. The P value was determined by Student’s t test [(B), (E), (F), and
(H) to (J)] or one-­way ANOVA coupled to Fisher’s LSD post hoc test [(D) and (K) to (N)].

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 6 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. FNIP1 (S220) phosphorylation regulates muscle mitochondrial function without affecting TFEB signaling. (A) Gene expression profiling of GC muscle in
mice. Heatmap analysis of DEGs in muscles of WT, FNIP1 mKO, and different transgenic (Tg) mice compared with NTG controls (n = 2 to 3 biological samples per group).
DEGs between different groups were analyzed by Gene Ontology (GO), and mRNA fold change is indicated by colored bar. (B) Left: Representative immunoblots of muscle
protein extracts from indicated mice following acute running exercise (Exe). Right: Quantification of AMPKα (T172) phosphorylation by pAMPK/AMPK signal ratios. n = 4
mice per group. (C) Left: Representative cross-­sectional images of plantaris muscle stained with anti-­TFEB antibody from indicated mice with or without exercise (Sed or
Exe). Scale bars, 50 μm. Right: Quantification of TFEB-­positive nuclear localization. n = 3 mice per group. (D and E) mTORC1 and AMPK signaling in Fnip1 Tg muscle. Im-
munoblotting analysis of WV muscle lysates from sedentary (Sed) and 60-­min exercised (Exe) mice of indicated genotype. n = 4 mice per group. (F) TFEB phosphorylation
analysis in Fnip1 Tg muscle. Immunoblotting analysis of muscle lysates from sedentary and 60-­min exercised mice of indicated genotype. To monitor the band shift of
TFEB, lysates of HEK293T cells with or without Torin1 treatment (CTL) were loaded as controls. (G to H) Quantification of pACC (S79), pAMPKα (T172), pS6K (T389), and
pTFEB (S122), phosphorylation by signal ratios normalized to NTG controls (=1.0) during sedentary or exercise conditions from (D) to (F). n = 4 to 8 mice per group. Error
bars are shown as the means ± SEM. NS, P > 0.05 was determined by one-­way ANOVA coupled to Fisher’s LSD post hoc test [(B) and (C)]. No statistical significance in (G)
and (H) calculated by Student's t test.

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 7 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
we checked the subcellular location of FNIP1. Immunofluorescence
costaining of Flag-­FNIP1 with the inner mitochondrial membrane
marker protein succinate dehydrogenase complex flavoprotein sub-
unit A (SDHA) revealed overlapping signals (fig. S4B). Moreover, we
isolated mitochondria from adult mouse skeletal muscle by sucrose
gradient fractionation and performed immunoblotting using anti-­
AMPK and anti-­FNIP1 antibodies. As shown in fig. S4C, we readily
detected both FNIP1 and AMPK in these enriched mitochondrial
fractions (fig. S4C). Notably, we did not detect FLCN in these enriched
mitochondrial fractions (fig. S4C), implicating that AMPK/FNIP1,
but not the GAP catalytic protein FLCN, is subcellularly localized or
associated with mitochondria in the skeletal muscle. We also took ad-
vantage of the skeletal muscle of FNIP1 mKO mice for fractionation,
and we observed a complete loss of FNIP1 but not AMPK in enriched
mitochondrial fractions (fig. S4D); again, we could not detect FLCN in
the mitochondrial fractions (fig. S4D). Using muscle-­specific FNIP1
(WT) and FNIP1 (S220A Tg) muscle tissues, we identified approxi-
mately 600 potential FNIP1 (WT)– or FNIP1 (S220A)–binding pro-
teins in exercising muscle tissue of the corresponding mice by the
coimmunoprecipitation (CoIP)–MS method (fig. S4E). Proteomic
analysis showed that both FNIP1 (WT) and FNIP1 (S220A) interacted
with major mitochondrial proteins, such as the respiratory complex
subunits (fig. S4F). Notably, FNIP1 (S220A) specifically interacted
with additional mitochondrial subunits that were not observed for
FNIP1 (WT), including proteins involved in reduced form of nicotin-
amide adenine dinucleotide (oxidized form) metabolism and sulfur
cluster binding (fig. S4G). These data suggested that nonphosphory-
lated FNIP1 (S220A) could enhance its interaction with mitochon-
dria. To further determine the effects of FNIP1 phosphorylation on
mitochondria, we checked the mitochondrial ETC (electron transport
chain) complex formation during exercise. Blue-­ native polyacryl-
amide gel electrophoresis (BN-­PAGE) of digitonin-­treated mitochon-
dria followed by immunoblotting with specific mitochondrial subunit
antibodies revealed that knockout of FNIP1 increased mitochondrial
ETC complex formation during exercise. In contrast, S220A sup-
pressed mitochondrial complex II (CII) and complex V (CV) forma-
tion, and there were no such effects in S220D Tg or WT FNIP1 Tg
muscle during exercise (Fig. 5, C and D). It has been reported that
mitochondrial chaperone genes regulate mitochondrial function by
modulating ETC complex assembly (42, 43). We also sought to deter-
mine whether mitochondrial chaperone gene expression is affected in
Fnip1 Tg muscles. However, we did not detect mRNA level changes for
most mitochondrial chaperones except a mild decrease in Clpp mRNA
in WT FNIP1 Tg and Dnaja3 mRNA in S220D Tg muscles compared
to NTG controls (fig. S5, A to C). We also found that Dnaja3 and
Foxred1 mRNA levels were mildly decreased in S220A Tg muscle
(fig. S5B). We found that Ppargc1α mRNA levels were mildly de-
creased in S220A Tg muscle (fig. S5B), whereas Ppargc1α gene expres-
sion was not changed in WT FNIP1 Tg or S220D Tg muscles (fig. S5,
A and C). Collectively, these results suggest that FNIP1 (S220) phos-
phorylation contributes to mitochondrial ETC complex formation
rather than affecting TFEB signaling in the skeletal muscle during
exercise.
DISCUSSION
AMPK is the metabolic fuel gauge that senses changes in the intra-
cellular AMP/ATP ratio in response to physical exercise. Activation
of skeletal muscle AMPK by exercise or exercise mimicking drugs
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
enhances mitochondrial metabolism and has become a powerful
therapeutic strategy for type 2 diabetes and associated metabolic
disorders. AMPK controls multiple aspects of mitochondrial func-
tion by phosphorylating effector proteins in diverse metabolic path-
ways involved in oxidative respiration and fatty acid oxidation. The
delineation of the full repertoire of AMPK targets and mechanisms
involved in the beneficial effects of exercise on muscle mitochon-
drial metabolism has implications for therapeutic approaches for
many human diseases, including metabolic disorders and muscular
dystrophy.
In this study, we identified that FNIP1, a hypothetical mitochon-
drial gatekeeper (44), was phosphorylated at the S220 site, contrib-
uting to AMPK regulation of mitochondrial function in response to
exercise in mouse models. We originally found that muscle-­specific
FNIP1 KO mice develop many features of an exercise-­trained ath-
letic phenotype, including increased endurance, enhanced mito-
chondrial capacity, and fat burning. While muscle physiological
exercise capacity is normal and mitochondrial function is un-
changed in muscle-­specific WT FNIP1 Tg and S220D Tg transgenic
mice, the lack of FNIP1 (S220) phosphorylation (S220A Tg) leads to
compromised mitochondrial function and an exercise intolerance
phenotype in mice. These results highlighted that single-­site phos-
phorylation at S220 of FNIP1 by exercise-­activated AMPK contrib-
utes to mitochondrial regulation and muscle exercise physiology.
The regulation of mitochondrial metabolism by AMPK in exer-
cising muscle occurs at multiple levels, including de novo mitochon-
drial biogenesis and mitochondrial proteomic remodeling (8, 17,
23). Previous studies have demonstrated that the TFEB and PGC1α
transcriptional regulatory circuit, including the nuclear receptors
PPARs and ERRs, are key transducers of exercise-­responsive mito-
chondrial gene expression for mitochondrial biogenesis and fuel
metabolism (12, 20, 45, 46). Evidence is also emerging that mito-
chondrial proteome remodeling induced by exercise in skeletal mus-
cle can affect mitochondrial function and muscle performance (22,
23). Our results suggested that exercise-­induced FNIP1 (S220) phos-
phorylation by AMPK participated in the regulation of mitochon-
drial function and related muscle fuel catabolism, possibly through
its compartmentalized subpool on mitochondria. The discovery of a
mitochondrial pool of FNIP1 and its importance for mitochondrial
respiration control underscores the complexity of energetic moni-
toring in the skeletal muscle. These data imply a mechanism for how
AMPK regulates mitochondrial function through its constitutive
binding partner FNIP1 in the exercising muscle. S220 is not con-
served in FNIP2, a close homolog of FNIP1, which suggests that
FNIP1 likely has a unique AMPK phosphorylation–dependent reg-
ulatory mechanism that FNIP2 lacks.
An interesting finding of our study is that FNIP1 (S220) phos-
phorylation by AMPK participates in the regulation of mitochondrial
function without affecting muscle TFEB signaling. Our results sug-
gested that FNIP1 (S220) phosphorylation acts independently of
TFEB signaling to regulate mitochondrial ETC complex assembly
and fuel metabolism. First, RNA sequencing (RNA-­seq) data revealed
no expression change of TFEB target genes in S220A Tg, S220D Tg,
or WT FNIP1 Tg muscle compared to NTG controls, despite promi-
nent lysosomal or mitochondrial genes being detected in FNIP1
mKO muscle compared to WT controls; second, we found that FNIP1
deficiency, as previously reported, did not affect mTORC1 activity to-
ward its canonical substrate S6K phosphorylation, and rescue of both
FNIP1 (S220A) and FNIP1 (S220D) phosphomutants exhibited no
8 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. FNIP1 (S220) phosphorylation regulates mitochondrial ETC complex formation and respiration capability in muscle. (A and B) Mitochondrial respiration
rate analysis in FNIP1 mKO and transgenic (Tg) muscle tissues. Mitochondrial respiration rates were determined from the extensor digital longus (EDL) muscle of the indi-
cated genotypes using pyruvate (A) or succinate (B) as substrates. Pyruvate/malate (Py/M)– or succinate/rotenone (Suc/Rot)–stimulated, adenosine diphosphate (ADP)–
dependent respiration, and oligomycin-­induced (Oligo) respiration are shown. n = 4 to 7 mice per group. (C and D) The effects of FNIP1 (S220) phosphorylation on
mitochondrial ETC complex assembly in the muscle. Mitochondrial proteins were extracted from the indicated mice after exercise and analyzed by BN-­PAGE as described
in Materials and Methods. Western blotting analysis was performed with anti-­SDHA (CII) (C) and anti-­ATP5A (CV) (D) antibodies. Equal total mitochondrial proteins were
loaded, and the band gray values were quantified using ImageJ. WT and FNIP1 mKO mice, exercise for 150 min; NTG and Fnip1 Tg mice, exercise for 60 min. n = 3 to 8 mice
per group. Error bars are shown as the means ± SEM. *P < 0.05; **P < 0.01; ***P < 0.001. The P value was determined by Student’s t test.

Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024 9 of 15

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
effects on S6K phosphorylation in cells or in muscle tissues; third,
nuclear translocation of TFEB, a noncanonical mTORC1 substrate,
was also not changed either in sedentary or in exercised muscle; nei-
ther phospho-­TFEB (S122) nor total TFEB band shift (indicative of
TFEB overall phosphorylation levels) were changed in both WT
FNIP1 Tg and S220A Tg muscle at baseline or exercise conditions
compared to the NTG control; fourth, there exists a FNIP1 subpool
on mitochondria, and FNIP1 (S220A) showed a differential interac-
tive network from FNIP1 (WT) for mitochondrial proteins. Mito-
chondrial ETC complex formation and respiratory function were
compromised in the S220A Tg muscle during exercise. Mitochon-
drial ETC formation in the S220A Tg muscle is likely not driven by
the mild changes in chaperones for ETC complexes, although a re-
duced muscle Ppargc1α may contributes to impaired mitochondrial
function in the S220A Tg muscle. Our results provide evidence that
phosphorylation of FNIP1 (S220) by AMPK controls exercise capac-
ity and regulates mitochondrial ETC function. The precise mecha-
nisms involved in the regulation of mitochondrial ETC formation via
FNIP1 (S220) need further dissection based on these discoveries.
While we cannot exclude the possibility that FNIP1 S220A acts indi-
rectly on mitochondrial ETC formation (via some indirect interac-
tion with key ETC formation regulatory proteins or even potentially
via indirect regulation at the transcriptional level), our findings sug-
gest that there exists an AMPK-­FNIP1 subpool on mitochondria that
may be involved in the regulation of mitochondrial ETC assembly,
possibly in an FLCN-­independent manner as FLCN, the catalytic
subunit of GAP activity toward RagC/D, was not present as a mito-
chondrial protein. Future studies will be necessary to further de-
lineate the mitochondrial-­localized AMPK-­FNIP1 mechanism in
regulating muscle physiology. It will also be of interest to investigate
whether phosphorylation of S220 regulates the protein stability
of FNIP1.
When this manuscript was being prepared, the Shaw group (18)
had recently published a study that identified five AMPK phosphory-
lation sites, including S220 in FNIP1. By delicate cellular mechanistic
analysis, they elegantly demonstrated that FNIP1 phosphorylation at
these sites by AMPK was coordinated with RagC and mTORC1 to
induce nuclear translocation of TFEB and subsequent PGC1α and
ERRα transcription, thus contributing to mitochondrial biogenesis
in response to mitochondrial poisoning. This and other previous
studies have highlighted the important role of FNIP1 in TFEB signal-
ing (18, 19, 28, 37, 39, 47). In our experimental system, no differences
in TFEB (S122) phosphorylation, total TFEB band shift, or transloca-
tion were observed in WT FNIP1 Tg or S220A Tg muscle tissue. Our
data here suggested that FNIP1 (S220) phosphorylation by AMPK is
not essential or not sufficient for the inhibition of TFEB phosphory-
lation by mTORC1. Nevertheless, we cannot exclude the possibility
that FNIP1 phosphorylation at additional site(s) by AMPK affects
TFEB phosphorylation. Future studies aimed at dissecting the essen-
tial phosphorylation site(s) for TFEB regulation will likely require
more stringent mouse models, including rescue of the FNIP1 mKO
muscle with single-­site mutants, such as FNIP1 (S220A) and FNIP1
(S220D), and with multiple-­site mutants, such as FNIP1 (SA4) or
FNIP1 (SA5), identified by Malik et al. (18).
Here, using genetic mouse models, we provided physiological evi-
dence that single-­site phosphorylation at S220 of FNIP1 by exercise-­
activated AMPK contributes to mitochondrial regulation through
mitochondrial proteomic remodeling and ETC complex assembly
in the skeletal muscle. Such a mechanism is in parallel to the
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
AMPK-­ FNIP1-­ TFEB mitochondrial biogenesis mediated by
multiple-­site FNIP1 phosphorylation. Nevertheless, it is also plau-
sible that both proteome or ETC complex remodeling and the bio-
genesis of mitochondria are coordinately regulated by AMPK
phosphorylation of FNIP1.
MATERIALS AND METHODS
Experimental design
The objective of this study was to investigate the key effector(s) of
AMPK and the regulation of muscle mitochondrial function and ex-
ercise performance. Using a combination of genetic mouse models
and biochemical approaches, our study highlights the critical role of
FNIP1 (S220) phosphorylation by AMPK in controlling mitochon-
drial function and muscle fuel utilization during exercise. These find-
ings implicate a target for exercise tolerance in the health and disease
of the skeletal muscle. The sample sizes are explicitly stated in the fig-
ure legends. No statistical methods were used to predetermine sample
sizes. Sample sizes were determined on the basis of previous experi-
ments using similar methodologies and are sufficient to account for
any biological/technical variability. For all experiments, samples/ani-
mals were randomly allocated to experimental groups and processed.
No data were excluded from the analyses. The investigators were not
blinded to allocation during experiments and outcome assessment
because the investigators needed to conduct genotyping polymerase
chain reactions (PCRs) at the age of 2 weeks for the mice. Blinding
was not relevant to the other experiments in cells because the investi-
gators needed to know what cell type they had to culture and process
the cells by themselves.
Animal studies
All animal studies were conducted in strict accordance with the insti-
tutional guidelines for the humane treatment of animals and were ap-
proved by the Institutional Animal Care and Use Committee (IACUC)
committees at the Model Animal Research Center of Nanjing Univer-
sity. The generation of Fnip1 floxed mice has been described previ-
ously in detail (48). Briefly, the Fnip1 targeting vector was generated
by inserting two loxP sequences into intron 5 and intron 6. To gener-
ate mice with a muscle-­specific disruption of the Fnip1 allele, Fnip1fl/fl
mice were crossed with mice expressing Cre recombinase under the
control of an human skeletal actin promoter (the Jackson laboratory,
stock no. 006139) to achieve muscle-­specific deletion of Fnip1 (FNIP1
mKO). Deletion of exon 6 in the FNIP1 mKO mice resulted in a read-
ing frameshift and premature termination codon in exon 7. The gen-
eration of FNIP1 transgenic (FNIP1 Tg) mice under the control of the
muscle creatine kinase promoter (a kind gift from E. N. Olson, Uni-
versity of Texas Southwestern) has been described elsewhere (31), and
the Fnip1 transgene was expressed efficiently in a skeletal muscle–spe-
cific manner with no overexpression of FNIP1 protein in other tissues,
such as the heart (31). Muscle-­specific FNIP1 S220A/S220D Tg mouse
models (FNIP1 S220A/S220D Tg) were established using the same
strategy as that of WT FNIP1 Tg lines. Briefly, cDNA encoding the
mouse Fnip1 mutant gene was separately cloned and inserted into the
EcoR V site downstream of the mouse Mck gene promoter. Site-­
directed mutagenesis was carried out using the QuikChange Kit
(Stratagene) according to the manufacturer’s protocol: the amino acid
S220 (S; 5′-­TCT-­3′), an FNIP1 phosphorylation site, was changed to
Ala (A; 5′-­GCT-­3′) or Asp (D; 5′-­GAT-­3′), which mimicked dephos-
phorylation or phosphorylation, respectively. The transgene was
10 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
linearized with Xho I and Sac II digestion and microinjected into
C57BL/6J embryos by the Tg mouse facility at the Model Animal
Research Center of Nanjing University. Tg mice were identified by
PCR amplification of a 534–base pair product using primers specific
for Fnip1 (5′-­TTTCCAACCTGCTTCATTCCACTCTTCA) and the
human growth hormone polyadenylate component of the muscle cre-
atine kinase (MCK) construct (5′-­AAGATTGTGCCACTGCA). Male
mice aged 8 to 12 weeks were used.
Recombinant AMPK purification and in vitro kinase assay
AMPKα1, AMPKβ1, AMPKγ1, and Calcium/calmodulin-­dependent
protein kinase kinase β (CaMKKβ) were subcloned and inserted
into the pET28b vector, and the bacterial expression and purifica-
tion of recombinant AMPKα1β1γ1 and CaMKKβ were performed.
Briefly, recombinant proteins were induced with 0.1 mM isopropyl-­
β-­​d-­thiogalactopyranoside at 22°C overnight in Escherichia coli
BL21(DE3) cells when the cell density reached 0.4 to 0.6 at 600 nm
(optical density at 600 nm). Cells were pelleted and resuspended in
lysis buffer containing 15% sucrose (w/v), 50 mM sodium phos-
phate (pH 7.5), 100 mM NaCl, 10 mM imidazole, and 1 mM β-­
mercaptoethanol with 1% Triton X-­ 100. After sonication and
centrifugation, the supernatant was collected and purified with
Ni–nitrilotriacetic acid resin. After washing with lysis buffer con-
taining 50 mM imidazole, the proteins were eluted with lysis buffer
containing 250 mM imidazole and dialyzed overnight with dialysis
buffer [50 mM tris-­HCl (pH 7.5), 1 mM β-­mercaptoethanol, and 1
mM EDTA] at 4°C to remove imidazole and then stored at −80°C
until use (49).
For the in vitro kinase assay, 200 ng of recombinant AMPKα1β1γ1
complex was preincubated with 20 ng of CaMKKβ to be fully acti-
vated in reaction buffer [20 mM tris-­HCl, 1 mM dithiothreitol, 5 mM
MgCl2, and 8 nM ATP (pH 7.5)] in a 30°C water bath for 2 hours.
Flag-­FNIP1 (WT and S220A) was purified from HEK293T cells
transfected with Flag-­FNIP1 (WT and S220A). Cells were pretreated
with the AMPK inhibitor compound C (10 μM) for 4 hours before
harvest, which inhibited AMPK-­dependent FNIP1 phosphorylation.
Cell lysates were collected 36 hours after transfection, and Flag-­
FNIP1 was purified with Flag antibody. Protein A/G beads and
200 ng of eluted substrate Flag-­FNIP1 (WT and S220A) protein
were incubated with 20 ng of activated AMPK in reaction buffer in
a 37°C water bath for 1 hour. After incubation, the reaction was
terminated by SDS loading buffer and heated to 95°C for 10 min,
followed by SDS-­PAGE and immunoblotting with specific an-
tibodies.
MS analysis
To identify phosphorylation sites of FNIP1, pcDNA5-­Flag-­FNIP1
was expressed in HEK293T cells, and the cells were treated with
A-­769662 before harvest. For Flag-­IP, cell lysates were incubated
with anti-­Flag M2 beads (Sigma-­Aldrich, A2220) at 4°C overnight.
IP of proteins and phosphorylated peptides/residues was performed
by Shanghai Applied Protein Technology Co. Ltd. To identify
FNIP1-­interacting proteins, muscle tissues from FNIP1 Tg or
FNIP1 S220A Tg mice were homogenized and then incubated
with anti-­Flag M2 beads (Sigma-­Aldrich, A2220) at 4°C over-
night. The samples were subjected to SDS-­PAGE, and stained with
Coomassie bright blue R-­250. For MS identification, the gel portions
containing the protein bands were excised, destained, dehydrated,
and subjected to trypsin digestion. The resulting peptides were
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
analyzed by a liquid chromatography–MS system (2D-­NanoLC/
TripleTOF5600) (50).
Exercise stress test
Mice were acclimated (run for 9 min at 10 m/min, followed by 1 min
at 20 m/min at a 10° incline) to the treadmill for two consecutive
days prior to the experimental protocol. Low-­intensity (endurance)
exercise studies were conducted. Briefly, fed mice were run for
10 min at 10 m/min, followed by a constant speed of 20 m/min at a
10° incline until exhaustion. Tail blood was taken after exercise and
measured for lactate (Lactate Scout, Senelab, Germany) according
to the manufacturer’s instructions (32).
RERs during exercise were determined using a high-­intensity ex-
ercise protocol. Briefly, mice were placed in an enclosed treadmill
attached to the Comprehensive Laboratory Animal Monitoring System
(Columbus Instruments) for 15 min at a 0° incline and 0 m/min.
The mice were then challenged with 2-­min intervals of increasing
speed at a 0° incline. The increasing speeds used in the protocol were
10, 14, 18, 22, 26, 28, 30, 32, 34, 36, 38, 40, 42, 44, 46, 48, 50, and 52
m/min. Measurements were collected before the exercise challenge,
throughout the challenge, and following failure.
Blood and tissue chemistry
Blood glucose levels were determined using a OneTouch UltraMini
glucose meter (OneTouch). Serum TG levels were determined using
a TG Kit (Wako, 290-­63701). Serum fatty acid levels were determined
using a NEFA kit (Wako, 294-­63601). Serum β-­hydroxybutyrate lev-
els were measured using the β-­hydroxybutyrate (Ketone Body) Colo-
rimetric Assay Kit (Cayman Chemical, 700190) according to the
manufacturer’s instructions.
Mitochondrial respiration studies
Mitochondrial respiration rates were measured in saponin-­permeablized
extensor digital longus muscle fibers with pyruvate or succinate as
substrates. Briefly, the muscle fibers were separated and transferred
to BIOPS buffer [7.23 mM K2EGTA, 2.77 mM CaK2EGTA, 20 mM
imidazole, 20 mM taurine, 50 mM potassium 2-­[N-­morpholino]-­
ethanesulfonic acid, 0.5 mM dithiothreitol, 6.56 mM MgCl2, 5.7 mM
ATP, and 14.3 mM phosphocreatine (pH 7.1)]. The muscle fiber
bundles were then permeabilized with saponin (50 μg/ml) in BIOPS
solution. Measurement of oxygen consumption in permeabilized
muscle fibers was performed in buffer Z [105 mM potassium
2-­[N-­morpholino]-­ethanesulfonic acid, 30 mM KCl, 10 mM KH2PO4,
5 mM MgCl2, bovine serum albumin (5 mg/ml), and 1 mM EGTA
(pH 7.4)] at 37°C and in the oxygen concentration range of 220 to
150 nmol of O2/ml in the respiration chambers of an Oxygraph 2K
(Oroboros Inc., Innsbruck, Austria). Following measurement of basal,
pyruvate (10 mM)/malate (5 mM) or succinate (5 mM)/rotenone (10
μM) respiration, maximal [adenosine diphosphate (ADP)–stimulated]
respiration was determined by exposing the mitochondria to 4 mM
ADP. Uncoupled respiration was evaluated following addition of
oligomycin (1 μg/ml). Respiration rates were determined and normalized
to tissue wet weight using Datlab 5 software (Oroboros Inc., Inns-
bruck, Austria), and the data were expressed as “picomole of O2 per
second per milligram wet weight” (31).
Mitochondrial isolation and BN-­PAGE
Mitochondria were isolated via sucrose gradient fractionation from
fresh mouse muscle as previously described (51). Briefly, minced
11 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
muscles from mice of different genotypes were homogenized with a glass
dounce in fractionation (FRAC) buffer [10 mM Hepes-­Na, 300 mM
sucrose, and 0.2 mM EDTA (pH 7.2)] and then centrifuged twice at
800g for 20 min at 4°C, and the resulting supernatant was termed
postnuclear lysate. The resulting supernatants were carefully trans-
ferred to a new tube and centrifuged at 8000g for 20 min at 4°C. The
pellet was resuspended in FRAC buffer and spun again until a solid
pellet formed at the bottom of the tube, designated the mitochondrial
fraction. The whole-­tissue lysates, postnuclear lysate, and mitochon-
drial fractions were resuspended in FRAC buffer and quantified by
bicinchoninic acid (BCA) assay using Pierce BCA Assay Kit Protocol
(Thermo Fisher Scientific).
BN-­PAGE analysis was performed. A total of 250 μg of mitochon-
dria isolated as described above were resuspended in solubilization buffer
[50 mM NaCl, 50 mM imidazole, 2 mM 6-­aminohexanoic, and 1 mM
EDTA (pH 7.0)]. Then, the mitochondria were incubated with 20%
digitonin on ice for 10 min. After centrifugation at 20,000g for 30 min at
4°C, the supernatants were collected. The samples were mixed with 50%
glycerol and 5% Coomassie G-­250 and subjected to 3.5 to 13% BN-­
PAGE for electrophoresis at 4°C. After the native gel electrophoresis was
conducted at 100 V for 30 min, cathode buffer B (50 mM tricine,
7.5 mM imidazole, and 0.02% Coomassie brilliant blue G-­250) was
changed to cathode buffer B/10 (50 mM tricine, 7.5 mM imidazole, and
0.002% Coomassie brilliant blue G-­250), and the running continued at
15 mA for ~3 hours. The gels were electroblotted on polyvinylidene
difluoride membranes for immunoblotting (51). Equal amounts of total
mitochondrial proteins were loaded, and the relative levels of protein
were quantified by band gray values calculated with ImageJ software.
Transmission electron microscopy
Mice were euthanized, and soleus muscles were dissected, cut into
small pieces, and fixed in 2.5% glutaraldehyde in 0.1 M phosphate buf-
fer overnight at 4°C. Then, the specimens were postfixed for 1.5 hours
with 1% osmium tetroxide in phosphate buffer, dehydrated through a
graded ethanol series, and embedded in Epon812. Ultrathin sections
(70 nm) were prepared, stained with uranyl acetate and lead citrate,
and examined by electron microscopy (JEOL-­1200EX) at JiNan WeiYa
Bio-­Technology Co. Ltd. (Jinan, China) in a double-­blind manner.
Histologic analysis
Muscle tissue was frozen in isopentane that had been cooled in liquid
nitrogen. Notably, the GC of each mouse was dissected as a whole.
Ten-­micrometer-­thick serial GC muscle cross sections were cut from
the knee cut side in a Leica CM1850 cryostat at −20°C and mounted
on positively charged glass slides. Transverse sections collected from
the widest part (mid-­belly) of the GC muscle were used for histo-
logical comparison to maintain consistency between different mice.
For TFEB immunohistochemistry (IHC), GC muscle cross sec-
tions were washed three times with phosphate-­buffered saline (PBS)
for 5 min each, fixed in ice-­cold 4% paraformaldehyde for 10 min,
washed with PBS for 5 min, and then permeabilized with ice-­cold
0.5% Triton X-­100–PBS for 10 min. IHC was performed using the
UltraSensitive SP (rabbit) IHC Kit (Maxim) according to the manu-
facturer’s instructions using rabbit TFEB antibodies (Bethyl Labora-
tories, A303-­673A; 1:200). A 3,3'-­Diaminobenzidine (DAB) staining
kit (Maxim, DAB-­0031) was used according to the manufacturer’s
instructions for development.
Periodic acid–Schiff staining was used to detect glycogen accu-
mulation. The muscle fibers were oxidized in 0.5% periodic acid for
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
5 min, rinsed three times in distilled water, and treated with Schiff ’s
reagent for 20 min. After extensive washing, the slides were counter-
stained with hematoxylin and lastly sealed with neutral resin.
Glycogen extraction and content measurement
For measurement of muscle glycogen, muscle tissue was acid hydro-
lyzed in 2 M HCl at 95°C for 2 hours, neutralized with an equal
volume of 2 M NaOH, and centrifuged at 12,000g for 10 min. The
liberated free-­glycosyl units of the supernatant were determined us-
ing the glucose hexokinase kit (Wako, 298-­65701) according to the
manufacturer’s instructions.
RNA analysis
Quantitative reverse transcription (RT)–PCR was performed as
described previously (48). Briefly, total RNA was extracted from
tissues using RNAiso Plus (Takara Bio). The purified RNA samples
were then reverse-­transcribed using the PrimeScript RT Re-
agent Kit with gDNA Eraser (Takara Bio). Real-­time quantita-
tive RT-­PCR was performed using the ABI Prism Step-­One
system with a Reagent Kit from Takara Bio. Specific oligonucle-
otide primers for target gene sequences are listed in table S1.
Arbitrary units of target mRNA were corrected to the expres-
sion of 36b4.
RNA-­seq studies
Transcriptomics analyses were performed using RNA-­seq as described
previously (52). Total RNA was isolated from the entire GC
muscle of FNIP1 mKO, WT FNIP1 Tg, FNIP1 S220A Tg, FNIP1
S220D Tg, and littermate control mice using RNAiso Plus (Takara
Bio). RNA-­s eq using an Illumina HiSeq 4000 was performed by
Beijing Novogene Bioinformatics Technology Co. Ltd. Two or
three independent samples per group were analyzed. Paired-­end
and 150-­nucleotide reads were obtained from the same sequencing
lane. The sequencing reads were then aligned to the UCSC
mm10 genome assembly using TopHat 2.0.14 with the default
parameters. Fragments per kb of exon per million mapped reads
were calculated using Cufflinks 2.2.1. The criteria for a regulated
gene were a fold change of >1.5 (either direction) and a significant
P value (<0.05) versus littermate controls. For pathway analysis,
the filtered datasets were uploaded into DAVID Bioinformatics
Resources 6.8 to review the biological pathways using the Func-
tional Categories database. Gene Ontology analysis was used to
interpret the data, and the regulated terms were ranked by P
value. The RNA-­s eq data have been deposited in the National
Genomics Data Center (NGDC) Genome Sequence Archive
(GSA) and are accessible through GSA Series accession numbers
CRA008211 (https://ngdc.cncb.ac.cn/search/?dbId=gsa&q=CRA0
08211), CRA011387 (https://ngdc.cncb.ac.cn/gsa/s/RdgAOIP6),
and CRA013258 (https://ngdc.cncb.ac.cn/gsa/s/SrO1O3I7).
Plasmid construction and site-­directed mutagenesis
pEGFP-­N1-­FNIP1 (Addgene, #49175) was purchased from Add-
gene. The FNIP1 coding sequence was then subcloned and inserted
into the pcDNA5 vector containing a Flag-­tag at the N terminus. Site
directed mutagenesis was performed using the QuikChange Kit
(Stratagene) according to the manufacturer’s protocol: the amino
acid S220, an FNIP1 phosphorylation site, was changed to Ala (A) or
Asp (D), which mimicked dephosphorylation or phosphorylation,
respectively. All constructs were confirmed by DNA sequencing.
12 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Primary cell culture
Primary muscle cells were isolated from the GC muscles of 4-­week-­old
male mice as previously described (31). Briefly, GC muscles from
both legs were removed. Minced tissue was digested in a collagenase/
dispase/CaCl2 solution for 1.5 hours at 37°C in a shaking bath.
Dulbecco’s modified Eagle’s medium (DMEM) supplemented with
10% fetal bovine serum (FBS) [Pre-plating medium (PPM)] was add-
ed, and samples were triturated gently before loading onto a Netwell
filter (70 μm; BD Biosciences). The cell suspension was pelleted at
1000 rpm for 5 min. Cells were then resuspended in PPM and plated
on an uncoated plate for differential plating. The cell suspension (not
adherent) was centrifuged for 5 min at 1000 rpm, and the pellet was
resuspended in growth medium (GM) [Ham’s F-­10 medium supple-
mented with 20% FBS and basic fibroblast growth factor (2.5 ng/ml)].
Cells were plated on collagen coated flasks for expansion. Cells were
fed daily with GM. For differentiation, cells were washed with PBS
and refed with 2% horse serum/DMEM differentiation medium and
refed daily. Cells were induced to differentiate for 3 days before vari-
ous experiments.
Cell transfection, RNA interference experiments, IP,
and immunofluorescence
Transient transfections in HEK293T cells were performed
using PEI Transfection Reagent (Polysciences) following the
manufacturer’s protocol. siRNAs (GenePharma) targeting
Fnip1 (siRNA pool: #1, 5′-­GCAGUUCACAGCAACCCAATT;
#2, 5′-­GGUGGCUACUGCUCAUCUUTT) were transfected into
cells at a final concentration of 50 nM using Lipofectamine 2000
transfection reagent (Invitrogen) according to the manufactur-
er’s instructions.
Whole lysates from HEK293T cells 48 hours after transfection
were used for CoIP studies. HEK293T cells were obtained from
American Type Culture Collection were cultured at 37°C and 5% CO2
in DMEM supplemented with 10% FBS, penicillin (1000 U/ml), and
streptomycin (100 g/ml). Cells were collected in lysis buffer [50 mM
tris (pH 7.4), 150 mM NaCl, 1 mM EDTA, 1% Triton X-­100, 1× cOm-
plete (Roche), 1 mM phenylmethylsulfonyl fluoride, 10 mM NaF, and
5 mM Na3VO4], and 1 μg of M2 anti-­Flag (MilliporeSigma) antibody
was incubated with extract and protein G–conjugated agarose beads.
The immunoprecipitated proteins were analyzed by immunoblotting.
For immunofluorescence, HeLa cells were transfected with the
indicated plasmids with Lipofectamine 2000 (Invitrogen). Thirty-­
six hours after transfection, the cells were washed with cold PBS and
fixed with 4% paraformaldehyde in PBS at room temperature for
15 min. Then, the cells were permeabilized with 0.1% Triton X-­100
and blocked with 10% goat serum at room temperature for 30 min,
followed by incubation with primary antibody diluted in PBS con-
taining 10% goat serum at 4°C overnight. The cells were then washed
three times with PBS and incubated with fluorescence dye–conju-
gated secondary antibodies in PBS containing 10% goat serum for
1 hour at room temperature. Images were taken by stimulated emis-
sion depletion superresolved microscopy (Abberior Instruments).
Antibodies and immunoblotting
Antibody directed against FLAG (F1804; 1:1000 dilution) was from
Sigma-­Aldrich; antibody directed against α-­tubulin (bs1699; 1:5000
dilution) was from Bioworld; antibodies directed against FLCN
(#3697S;1:1000 dilution), pTFEB (S122) (#86843; 1:1000 dilution),
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
TFEB (#37785S; 1:1000 dilution), pAMPKα (T172) (#2535; 1:1000
dilution), AMPKα (#5831; 1:1000 dilution), phospho-­ Acetyl-­
CoA
Carboxylase (pACC) (S79) (#11818; 1:1000 dilution), ACC (#3676;
1:1000 dilution), pS6K (T389) (#9234; 1:1000 dilution), and S6K
(#2708; 1:1000 dilution) were from Cell Signaling Technology; anti-
bodies directed against SDHA (14865-­1-­AP; 1:1000 dilution), ATP5A
(14676-­1-­AP; 1:1000 dilution), UQCRC2 (14742-­1-­AP; 1:1000 di-
lution), and COX4 (11242-­1-­AP; 1:1000 dilution) were from Protein-
Tech; antibody directed against Tim23 (#611222; 1:1000 dilution) was
from BD Biosciences; antibody directed against glyceraldehyde-­3-­
phosphate dehydrogenase (GAPDH) (#2133; 1:3000 dilution) was
from Signalway Antibody; antibody directed against FNIP2 (ab106611;
1:1000 dilution) was from Abcam, antibodies directed against FNIP1
(ab236547; 1:500 dilution) and pFNIP1 (S220) (1:200 dilution) were
developed in the laboratory of Zhenji Gan with the help with Abcam.
Specifically, after considering immunogenicity and uniqueness, C-­
QFCSPRRAF-­pS-­EQGP was chosen as the immunogen for the anti-
bodies against pS220, and cystine was used for conjugation to the
carrier protein. To generate a rabbit polyclonal antibody that recog-
nizes FNIP1 pS220, the antigen peptide was injected into rabbits, and
serum was collected and purified using an affinity column conjugated
with unmodified peptides to exclude antibodies recognizing unmodi-
fied proteins and then subjected to an affinity column conjugated with
the modified peptides to bind and purify the antibodies. The antibody
was then eluted and concentrated. The specificity of the pFNIP1
(S220) antibody was evaluated by immunoblotting in the presence of
blocking peptides. Western blotting studies were performed as previ-
ously described (31).
Statistical analysis
All mouse and cell studies were analyzed by Student’s t test when
two groups were compared. One-­way analysis of variance (ANOVA)
coupled to Fisher’s least significant difference (LSD) post hoc test
was used when more than two groups were compared. Data repre-
sent the means ± SEM, with a statistically significant difference de-
fined as a value of P < 0.05.
Supplementary Materials
This PDF file includes:
Figs. S1 to S5
Table S1
REFERENCES AND NOTES
1. M. Hargreaves, L. L. Spriet, Skeletal muscle energy metabolism during exercise. Nat.
Metab. 2, 817–828 (2020).
2. J. A. Hawley, M. Hargreaves, M. J. Joyner, J. R. Zierath, Integrative biology of exercise. Cell
159, 738–749 (2014).
3. Z. Gan, T. Fu, D. P. Kelly, R. B. Vega, Skeletal muscle mitochondrial remodeling in exercise
and diseases. Cell Res. 28, 969–980 (2018).
4. B. Egan, J. R. Zierath, Exercise metabolism and the molecular regulation of skeletal
muscle adaptation. Cell Metab. 17, 162–184 (2013).
5. S. Herzig, R. J. Shaw, AMPK: Guardian of metabolism and mitochondrial homeostasis. Nat.
Rev. Mol. Cell Biol. 19, 121–135 (2018).
6. D. G. Hardie, F. A. Ross, S. A. Hawley, AMPK: A nutrient and energy sensor that maintains
energy homeostasis. Nat. Rev. Mol. Cell Biol. 13, 251–262 (2012).
7. E. C. Cokorinos, J. Delmore, A. R. Reyes, B. Albuquerque, R. Kjøbsted, N. O. Jørgensen,
J. L. Tran, A. Jatkar, K. Cialdea, R. M. Esquejo, J. Meissen, M. F. Calabrese, J. Cordes,
R. Moccia, D. Tess, C. T. Salatto, T. M. Coskran, A. C. Opsahl, D. Flynn, M. Blatnik, W. Li,
E. Kindt, M. Foretz, B. Viollet, J. Ward, R. G. Kurumbail, A. S. Kalgutkar, J. F. P. Wojtaszewski,
K. O. Cameron, R. A. Miller, Activation of skeletal muscle AMPK promotes glucose disposal
13 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
and glucose lowering in non-­human primates and mice. Cell Metab. 25, 1147–1159.e10
(2017).
8. H. R. Spaulding, Z. Yan, AMPK and the adaptation to exercise. Annu. Rev. Physiol. 84,
209–227 (2022).
9. Z. Chen, C. Lei, C. Wang, N. Li, M. Srivastava, M. Tang, H. Zhang, J. M. Choi, S. Y. Jung, J. Qin,
J. Chen, Global phosphoproteomic analysis reveals ARMC10 as an AMPK substrate that
regulates mitochondrial dynamics. Nat. Commun. 10, 104 (2019).
10. E. Q. Toyama, S. Herzig, J. Courchet, T. L. Lewis Jr., O. C. Losón, K. Hellberg, N. P. Young,
H. Chen, F. Polleux, D. C. Chan, R. J. Shaw, Metabolism, AMP-­activated protein kinase
mediates mitochondrial fission in response to energy stress. Science 351, 275–281 (2016).
11. G. R. Steinberg, D. Carling, AMP-­activated protein kinase: The current landscape for drug
development. Nat. Rev. Drug Discov. 18, 527–551 (2019).
12. V. A. Narkar, M. Downes, R. T. Yu, E. Embler, Y. X. Wang, E. Banayo, M. M. Mihaylova,
M. C. Nelson, Y. Zou, H. Juguilon, H. Kang, R. J. Shaw, R. M. Evans, AMPK and PPARdelta
agonists are exercise mimetics. Cell 134, 405–415 (2008).
13. P. M. García-­Rovés, M. E. Osler, M. H. Holmström, J. R. Zierath, Gain-­of-­function R225Q
mutation in AMP-­activated protein kinase γ3 subunit increases mitochondrial biogenesis
in glycolytic skeletal muscle*. J. Biol. Chem. 283, 35724–35734 (2008).
14. J. C. Campos, L. H. Marchesi Bozi, B. Krum, L. R. Grassmann Bechara, N. D. Ferreira,
G. S. Arini, R. P. Albuquerque, A. Traa, T. Ogawa, A. M. van der Bliek, A. Beheshti,
E. T. Chouchani, J. M. Van Raamsdonk, T. K. Blackwell, J. C. B. Ferreira, Exercise preserves
physical fitness during aging through AMPK and mitochondrial dynamics. Proc. Natl.
Acad. Sci. U.S.A. 120, e2204750120 (2023).
15. M. K. Hesselink, V. Schrauwen-­Hinderling, P. Schrauwen, Skeletal muscle mitochondria as
a target to prevent or treat type 2 diabetes mellitus. Nat. Rev. Endocrinol. 12, 633–645
(2016).
16. H. M. O'Neill, S. J. Maarbjerg, J. D. Crane, J. Jeppesen, S. B. Jørgensen, J. D. Schertzer,
O. Shyroka, B. Kiens, B. J. van Denderen, M. A. Tarnopolsky, B. E. Kemp, E. A. Richter,
G. R. Steinberg, AMP-­activated protein kinase (AMPK) β1β2 muscle null mice reveal an
essential role for AMPK in maintaining mitochondrial content and glucose uptake during
exercise. Proc. Natl. Acad. Sci. U.S.A. 108, 16092–16097 (2011).
17. H. Zong, J. M. Ren, L. H. Young, M. Pypaert, J. Mu, M. J. Birnbaum, G. I. Shulman, AMP
kinase is required for mitochondrial biogenesis in skeletal muscle in response to chronic
energy deprivation. Proc. Natl. Acad. Sci. U.S.A. 99, 15983–15987 (2002).
18. N. Malik, B. I. Ferreira, P. E. Hollstein, S. D. Curtis, E. Trefts, S. Weiser Novak, J. Yu, R. Gilson,
K. Hellberg, L. Fang, A. Sheridan, N. Hah, G. S. Shadel, U. Manor, R. J. Shaw, Induction of
lysosomal and mitochondrial biogenesis by AMPK phosphorylation of FNIP1. Science
380, eabj5559 (2023).
19. M. Paquette, L. El-­Houjeiri, C. Z. L. P. Puustinen, P. Blanchette, H. Jeong, K. Dejgaard,
P. M. Siegel, A. Pause, AMPK-­dependent phosphorylation is required for transcriptional
activation of TFEB and TFE3. Autophagy 17, 3957–3975 (2021).
20. G. Mansueto, A. Armani, C. Viscomi, L. D'Orsi, R. De Cegli, E. V. Polishchuk, C. Lamperti,
I. Di Meo, V. Romanello, S. Marchet, P. K. Saha, H. Zong, B. Blaauw, F. Solagna, C. Tezze,
P. Grumati, P. Bonaldo, J. E. Pessin, M. Zeviani, M. Sandri, A. Ballabio, Transcription factor
EB controls metabolic flexibility during exercise. Cell Metab. 25, 182–196 (2017).
21. S. Jäger, C. Handschin, J. St-­Pierre, B. M. Spiegelman, AMP-­activated protein kinase
(AMPK) action in skeletal muscle via direct phosphorylation of PGC-­1alpha. Proc. Natl.
Acad. Sci. U.S.A. 104, 12017–12022 (2007).
22. C. Greggio, P. Jha, S. S. Kulkarni, S. Lagarrigue, N. T. Broskey, M. Boutant, X. Wang,
S. Conde Alonso, E. Ofori, J. Auwerx, C. Cantó, F. Amati, Enhanced respiratory chain
supercomplex formation in response to exercise in human skeletal muscle. Cell Metab.
25, 301–311 (2017).
23. N. J. Hoffman, B. L. Parker, R. Chaudhuri, K. H. Fisher-­Wellman, M. Kleinert, S. J. Humphrey,
P. Yang, M. Holliday, S. Trefely, D. J. Fazakerley, J. Stöckli, J. G. Burchfield, T. E. Jensen,
R. Jothi, B. Kiens, J. F. Wojtaszewski, E. A. Richter, D. E. James, Global phosphoproteomic
analysis of human skeletal muscle reveals a network of exercise-­regulated kinases and
AMPK substrates. Cell Metab. 22, 922–935 (2015).
24. B. N. Finck, D. P. Kelly, PGC-­1 coactivators: Inducible regulators of energy metabolism in
health and disease. J. Clin. Invest. 116, 615–622 (2006).
25. M. Baba, S. B. Hong, N. Sharma, M. B. Warren, M. L. Nickerson, A. Iwamatsu, D. Esposito,
W. K. Gillette, R. F. Hopkins III, J. L. Hartley, M. Furihata, S. Oishi, W. Zhen, T. R. Burke Jr.,
W. M. Linehan, L. S. Schmidt, B. Zbar, Folliculin encoded by the BHD gene interacts with a
binding protein, FNIP1, and AMPK, and is involved in AMPK and mTOR signaling. Proc.
Natl. Acad. Sci. U.S.A. 103, 15552–15557 (2006).
26. O. M. Siggs, A. Stockenhuber, M. Deobagkar-­Lele, K. R. Bull, T. L. Crockford, B. L. Kingston,
G. Crawford, C. Anzilotti, V. Steeples, S. Ghaffari, G. Czibik, M. Bellahcene, H. Watkins,
H. Ashrafian, B. Davies, A. Woods, D. Carling, A. Yavari, B. Beutler, R. J. Cornall, Mutation of
Fnip1 is associated with B-­cell deficiency, cardiomyopathy, and elevated AMPK activity.
Proc. Natl. Acad. Sci. U.S.A. 113, E3706–E3715 (2016).
27. M. Yan, M. C. Gingras, E. A. Dunlop, Y. Nouët, F. Dupuy, Z. Jalali, E. Possik, B. J. Coull,
D. Kharitidi, A. B. Dydensborg, B. Faubert, M. Kamps, S. Sabourin, R. S. Preston,
D. M. Davies, T. Roughead, L. Chotard, M. A. van Steensel, R. Jones, A. R. Tee, A. Pause, The
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
tumor suppressor folliculin regulates AMPK-­dependent metabolic transformation. J. Clin.
Invest. 124, 2640–2650 (2014).
28. Z. Y. Tsun, L. Bar-­Peled, L. Chantranupong, R. Zoncu, T. Wang, C. Kim, E. Spooner,
D. M. Sabatini, The folliculin tumor suppressor is a GAP for the RagC/D GTPases that
signal amino acid levels to mTORC1. Mol. Cell 52, 495–505 (2013).
29. C. S. Petit, A. Roczniak-­Ferguson, S. M. Ferguson, Recruitment of folliculin to lysosomes
supports the amino acid-­dependent activation of Rag GTPases. J. Cell Biol. 202,
1107–1122 (2013).
30. Z. Sun, L. Yang, A. Kiram, J. Yang, Z. Yang, L. Xiao, Y. Yin, J. Liu, Y. Mao, D. Zhou, H. Yu,
Z. Zhou, D. Xu, Y. Jia, C. Ding, Q. Guo, H. Wang, Y. Li, L. Wang, T. Fu, S. Hu, Z. Gan, FNIP1
abrogation promotes functional revascularization of ischemic skeletal muscle by driving
macrophage recruitment. Nat. Commun. 14, 7136 (2023).
31. L. Xiao, J. Liu, Z. Sun, Y. Yin, Y. Mao, D. Xu, L. Liu, Z. Xu, Q. Guo, C. Ding, W. Sun, L. Yang,
Z. Zhou, D. Zhou, T. Fu, W. Zhou, Y. Zhu, X. W. Chen, J. Z. Li, S. Chen, X. Xie, Z. Gan,
AMPK-­dependent and -­independent coordination of mitochondrial function and muscle
fiber type by FNIP1. PLOS Genet. 17, e1009488 (2021).
32. J. Liu, X. Liang, D. Zhou, L. Lai, L. Xiao, L. Liu, T. Fu, Y. Kong, Q. Zhou, R. B. Vega, M. S. Zhu,
D. P. Kelly, X. Gao, Z. Gan, Coupling of mitochondrial function and skeletal muscle fiber
type by a miR-­499/Fnip1/AMPK circuit. EMBO Mol. Med. 8, 1212–1228 (2016).
33. N. L. Reyes, G. B. Banks, M. Tsang, D. Margineantu, H. Gu, D. Djukovic, J. Chan, M. Torres,
H. D. Liggitt, S. D. Hirenallur, D. M. Hockenbery, D. Raftery, B. M. Iritani, Fnip1 regulates
skeletal muscle fiber type specification, fatigue resistance, and susceptibility to muscular
dystrophy. Proc. Natl. Acad. Sci. U.S.A. 112, 424–429 (2015).
34. H. Hasumi, M. Baba, Y. Hasumi, M. Lang, Y. Huang, H. F. Oh, M. Matsuo, M. J. Merino,
M. Yao, Y. Ito, M. Furuya, Y. Iribe, T. Kodama, E. Southon, L. Tessarollo, K. Nagashima,
D. C. Haines, W. M. Linehan, L. S. Schmidt, Folliculin-­interacting proteins Fnip1 and Fnip2
play critical roles in kidney tumor suppression in cooperation with Flcn. Proc. Natl. Acad.
Sci. U.S.A. 112, E1624–E1631 (2015).
35. H. Hasumi, M. Baba, Y. Hasumi, Y. Huang, H. Oh, R. M. Hughes, M. E. Klein, S. Takikita,
K. Nagashima, L. S. Schmidt, W. M. Linehan, Regulation of mitochondrial oxidative
metabolism by tumor suppressor FLCN. J. Natl. Cancer Inst. 104, 1750–1764 (2012).
36. J. A. Ramírez, T. Iwata, H. Park, M. Tsang, J. Kang, K. Cui, W. Kwong, R. G. James, M. Baba,
L. S. Schmidt, B. M. Iritani, Folliculin interacting protein 1 maintains metabolic
homeostasis during B cell development by modulating AMPK, mTORC1, and TFE3. J.
Immunol. 203, 2899–2908 (2019).
37. C. Settembre, R. Zoncu, D. L. Medina, F. Vetrini, S. Erdin, S. Erdin, T. Huynh, M. Ferron,
G. Karsenty, M. C. Vellard, V. Facchinetti, D. M. Sabatini, A. Ballabio, A lysosome-­to-­nucleus
signalling mechanism senses and regulates the lysosome via mTOR and TFEB. EMBO J. 31,
1095–1108 (2012).
38. J. A. Martina, Y. Chen, M. Gucek, R. Puertollano, MTORC1 functions as a transcriptional
regulator of autophagy by preventing nuclear transport of TFEB. Autophagy 8, 903–914
(2012).
39. R. E. Lawrence, S. A. Fromm, Y. Fu, A. L. Yokom, D. J. Kim, A. M. Thelen, L. N. Young, C. Y. Lim,
A. J. Samelson, J. H. Hurley, R. Zoncu, Structural mechanism of a Rag GTPase activation
checkpoint by the lysosomal folliculin complex. Science 366, 971–977 (2019).
40. G. Napolitano, C. Di Malta, A. Ballabio, Non-­canonical mTORC1 signaling at the lysosome.
Trends Cell Biol. 32, 920–931 (2022).
41. J. C. Drake, R. J. Wilson, R. C. Laker, Y. Guan, H. R. Spaulding, A. S. Nichenko, W. Shen,
H. Shang, M. V. Dorn, K. Huang, M. Zhang, A. B. Bandara, M. H. Brisendine, J. A. Kashatus,
P. R. Sharma, A. Young, J. Gautam, R. Cao, H. Wallrabe, P. A. Chang, M. Wong,
E. M. Desjardins, S. A. Hawley, G. J. Christ, D. F. Kashatus, C. L. Miller, M. J. Wolf,
A. Periasamy, G. R. Steinberg, D. G. Hardie, Z. Yan, Mitochondria-­localized AMPK responds
to local energetics and contributes to exercise and energetic stress-­induced mitophagy.
Proc. Natl. Acad. Sci. U.S.A. 118, e2025932118 (2021).
42. J. Song, J. M. Herrmann, T. Becker, Quality control of the mitochondrial proteome. Nat.
Rev. Mol. Cell Biol. 22, 54–70 (2021).
43. J. P. Castro, K. Wardelmann, T. Grune, A. Kleinridders, Mitochondrial chaperones in the
brain: Safeguarding brain health and metabolism? Front. Endocrinol. 9, 196 (2018).
44. A. G. Manford, F. Rodríguez-­Pérez, K. Y. Shih, Z. Shi, C. A. Berdan, M. Choe, D. V. Titov,
D. K. Nomura, M. Rape, A cellular mechanism to detect and alleviate reductive stress. Cell
183, 46–61.e21 (2020).
45. V. A. Narkar, W. Fan, M. Downes, R. T. Yu, J. W. Jonker, W. A. Alaynick, E. Banayo,
M. S. Karunasiri, S. Lorca, R. M. Evans, Exercise and PGC-­1α-­independent
synchronization of type I muscle metabolism and vasculature by ERRγ. Cell Metab. 13,
283–293 (2011).
46. Z. Gan, E. M. Burkart-­Hartman, D. H. Han, B. Finck, T. C. Leone, E. Y. Smith, J. E. Ayala,
J. Holloszy, D. P. Kelly, The nuclear receptor PPARβ/δ programs muscle glucose
metabolism in cooperation with AMPK and MEF2. Genes Dev. 25, 2619–2630 (2011).
47. R. M. Jansen, R. Peruzzo, S. A. Fromm, A. L. Yokom, R. Zoncu, J. H. Hurley, Structural basis
for FLCN RagC GAP activation in MiT-­TFE substrate-­selective mTORC1 regulation. Sci. Adv.
8, eadd2926 (2022).
14 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
48. Y. Yin, D. Xu, Y. Mao, L. Xiao, Z. Sun, J. Liu, D. Zhou, Z. Xu, L. Liu, T. Fu, C. Ding, Q. Guo,
W. Sun, Z. Zhou, L. Yang, Y. Jia, X. Chen, Z. Gan, FNIP1 regulates adipocyte browning and
systemic glucose homeostasis in mice by shaping intracellular calcium dynamics. J. Exp.
Med. 219, e20212491 (2022).
49. T. Pang, Z. S. Zhang, M. Gu, B. Y. Qiu, L. F. Yu, P. R. Cao, W. Shao, M. B. Su, J. Y. Li, F. J. Nan,
J. Li, Small molecule antagonizes autoinhibition and activates AMP-­activated protein
kinase in cells. J. Biol. Chem. 283, 16051–16060 (2008).
50. T. Fu, W. Sun, J. Xue, Z. Zhou, W. Wang, Q. Guo, X. Chen, D. Zhou, Z. Xu, L. Liu, L. Xiao,
Y. Mao, L. Yang, Y. Yin, X. N. Zhang, Q. Wan, B. Lu, Y. Chen, M. S. Zhu, P. E. Scherer, L. Fang,
H. L. Piao, M. Shao, Z. Gan, Proteolytic rewiring of mitochondria by LONP1 directs cell
identity switching of adipocytes. Nat. Cell Biol. 25, 848–864 (2023).
51. Z. Xu, T. Fu, Q. Guo, D. Zhou, W. Sun, Z. Zhou, X. Chen, J. Zhang, L. Liu, L. Xiao, Y. Yin, Y. Jia,
E. Pang, Y. Chen, X. Pan, L. Fang, M. S. Zhu, W. Fei, B. Lu, Z. Gan, Disuse-­associated loss of
the protease LONP1 in muscle impairs mitochondrial function and causes reduced
skeletal muscle mass and strength. Nat. Commun. 13, 894 (2022).
52. Q. Guo, Z. Xu, D. Zhou, T. Fu, W. Wang, W. Sun, L. Xiao, L. Liu, C. Ding, Y. Yin, Z. Zhou, Z. Sun,
Y. Zhu, W. Zhou, Y. Jia, J. Xue, Y. Chen, X. W. Chen, H. L. Piao, B. Lu, Z. Gan, Mitochondrial
proteostasis stress in muscle drives a long-­range protective response to alleviate dietary
obesity independently of ATF4. Sci. Adv. 8, eabo0340 (2022).
Acknowledgments
Funding: This work was supported by The Ministry of Science and Technology of China
grant 2018YFA0800700 and 2022YFA0806000 (to Z.G.); National Natural Science Foundation
of China grant nos. 91857105, 31922033, and 31871439 (to Z.G. and X.X.); Fundamental
Research Funds for the Central Universities grants 021414380511, 021414380529,
021414380533, and 021414380524 (to Z.G.); Natural Science Foundation of Guangdong
grant 2019A1515011105 (to X.X); project funded by China Postdoctoral Science Foundation
Xiao et al., Sci. Adv. 10, eadj2752 (2024) 7 February 2024
grant 2023M731633 (to Y.Y.); and Natural Science Foundation of Jiangsu Province grant
BK20230146 (to Y.Y.) Author contributions: Conceptualization: L.X., Y.Y., Z.S., J.Liu, X.X., and
Z.G. Methodology: L.X., Y.Y., Z.S., J.Liu, Y.J., L.Y., Y.M., S.P., Z.X., J.Li, X.X., and Z.G. Investigation:
L.X., Y.Y., Z.S., J.Liu, Y.J., L.Y., Y.M., S.P., and L.F. Resources: L.X., Y.Y., Z.S., J.Liu, Y.J., L.Y., Y.M., S.P., Z.X.,
L.F., J.Li, X.X., and Z.G. Data curation: J.Liu, X.X., and Z.G. Validation: L.X., Y.Y., Z.S., J.Liu, X.X., and
Z.G. Visualization: L.X., Y.Y., Z.S., J.Liu, X.X., and Z.G. Formal analysis: L.X., Y.Y., Z.S., J.Liu, Y.J., L.Y.,
Y.M., S.P., L.F., and Z.G. Software: J.Liu. Funding acquisition: Y.Y., X.X., and Z.G. Writing (original
draft): L.X., Y.Y., Z.S., J.Liu, X.X., and Z.G. Writing (review and editing): L.X., Y.Y., Z.S., J.Liu, X.X., and
Z.G. Project administration: X.X. and Z.G. Supervision: X.X. and Z.G. Competing interests: The
authors declare that they have no competing interests. Data and materials availability: All
data needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials. The MS proteomics data have been deposited to the
ProteomeXchange Consortium via the PRIDE partner repository with the dataset identifier
PXD046977. Submission details: project name, CoIP-­MS based method to identify FNIP1
binding proteins in exercising muscle; project accession, PXD046977 (www.ebi.ac.uk/pride/
archive/projects/PXD046977); the RNA-­seq data have been deposited in the NGDC GSA and
are accessible through GSA Series accession numbers CRA008211 (https://ngdc.cncb.ac.cn/
search/?dbId=gsa&q=CRA008211), CRA011387 (https://ngdc.cncb.ac.cn/gsa/s/RdgAOIP6),
and CRA013258 (https://ngdc.cncb.ac.cn/gsa/s/SrO1O3I7). The public muscle TFEB
overexpression–induced gene expression dataset is accessible through GEO Series accession
number GSE62975.
Submitted 23 June 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adj2752
15 of 15
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

BIOCHEMISTRY Copyright © 2024 The

Authors, some rights
Dimerization of a 5-­kDa domain defines the reserved; exclusive
licensee American
architecture of the 5-­MDa gammaproteobacterial Association for the
Advancement of
pyruvate dehydrogenase complex Science. No claim to
original U.S.
Government Works.
Sarah Meinhold†, Rafal Zdanowicz†, Christoph Giese, Rudi Glockshuber* Distributed under a
Creative Commons
The Escherichia coli pyruvate dehydrogenase complex (PDHc) is a ~5 MDa assembly of the catalytic subunits pyruvate Attribution
dehydrogenase (E1), dihydrolipoamide acetyltransferase (E2), and dihydrolipoamide dehydrogenase (E3). The NonCommercial
PDHc core is a cubic complex of eight E2 homotrimers. Homodimers of the peripheral subunits E1 and E3 associate License 4.0 (CC BY-­NC).
with the core by binding to the peripheral subunit binding domain (PSBD) of E2. Previous reports indicated that 12 E1 di-
mers and 6 E3 dimers bind to the 24-­meric E2 core. Using an assembly arrested E2 homotrimer (E23), we show that
two of the three PSBDs in the E23 dimerize, that each PSBD dimer cooperatively binds two E1 dimers, and that E3
dimers only bind to the unpaired PSBD in E23. This mechanism is preserved in wild-­type PDHc, with an E1 dimer:E2
monomer:E3 dimer stoichiometry of 16:24:8. The conserved PSBD dimer interface indicates that PSBD dimerization is
the previously unrecognized architectural determinant of gammaproteobacterial PDHc megacomplexes.

INTRODUCTION peripheral subunits E1 and E3 recognize the overlapping but non-

The pyruvate dehydrogenase complex (PDHc) is a huge multien-
zyme assembly that links glycolysis with the citric acid cycle in all
aerobic organisms. It catalyzes the transfer of an acetyl moiety from
pyruvate onto coenzyme A (CoA) to generate acetyl-­CoA, thereby
maintaining glucose homeostasis and mediating energy production
in the cell (1–4). Converting pyruvate, CoA, and nicotinamide ade-
nine dinucleotide (NAD+) into carbon dioxide, acetyl-­CoA, and
reduced form of NAD+ (NADH)/H+, the entire PDHc reaction cy-
cle comprises a series of five consecutive reactions (1, 5–8).
The overall reaction catalyzed by PDHcs is the same in pro-
karyotes and eukaryotes and achieved by three distinct catalytic
components. Across all domains of life, PDHcs contain the subunit
dihydrolipoamide acetyltransferase (E2) that forms the homo-­or
hetero-­oligomeric PDHc core, and the peripheral subunits pyruvate
dehydrogenase (E1) and dihydrolipoamide dehydrogenase (E3) that
associate with the core (5, 9–11). The E2 subunit is a multidomain
protein in which the individual domains are connected by flexible
linkers. Specifically, E2 contains one to three lipoyl domains (LDs)
in its N-­terminal segment, followed by a peripheral subunit binding
domain (PSBD) and a C-­terminal catalytic domain (CD) that is also
responsible for the assembly of the core (2, 11, 12). In Gram-­negative
bacteria, PDHc shows a cubic architecture with a core of 24 copies of
E2, arranged as eight E2 homotrimers (also termed E2 components)
occupying the vertices of the cube (13–15). A cubic core structure is
also preserved in the related α-­ketoglutarate and branched-­chain α-­
keto acid dehydrogenase complexes (16, 17). Eukaryotic PDHcs
have an icosahedral core structure composed of two different sub-
units, E2 and the E3-­binding protein (E3BP), in which E2 only in-
teracts with E1 and E3BP only binds E3 (18, 19). In bacterial PDHcs
with cubic core, multiple copies of E1 homodimers and E3 homodi-
mers (E1 and E3 components, respectively) bind to the PSBDs
in the flexible N-­t erminal part of the E2 subunits protruding
from the 24-­meric core to form an outer protein shell (Fig. 1A). The
ETH Zürich, Institute of Molecular Biology and Biophysics, Otto-­Stern-­Weg 5, 8093
Zürich, Switzerland.
*Corresponding author. Email: rudi@​mol.​biol.​ethz.​ch
†These authors contributed equally to this work.
identical binding sites in the small, 5-­to 7-­kDa PSBD of E2, al-
lowing for interactions with only E1 or E3 in a mutually exclusive
manner (20–24).
PDHc from Gram-­negative bacteria has been widely reported to
comprise a 24-­meric E2 core to which approximately 12 E1 dimers
and 6 E3 dimers are bound, resulting in a ~4.5-­MDa complex (12,
25–28). These data imply that only 18 of the 24 theoretically available
PSBD binding sites would be occupied. The role of the potentially
remaining six sites has so far not been clarified. For example, other
stoichiometries, sometimes largely deviating from the E1:E2:E3
(monomer) ratio of 1:1:0.5, were reported, with the only common
denominator that fully assembled PDHc contains twice as many
E1 than E3 subunits (29, 30). Up to now, the architectural principle
behind the stoichiometry of bacterial PDHcs, the specific arrange-
ment of the peripheral subunits relative to the core, and the reasons
underlying the functional advantages of a PDHc megacomplex
compared to smaller subcomplexes of E1, E2, and E3 that could con-
ceivably exhibit similar rates of catalysis (31, 32) are unknown. Be-
cause of potentially nonuniform subunit compositions and subunit
arrangements, as well as the intrinsic flexibility within PDHc subunits
required for efficient channeling of catalytic intermediates, no three-­
dimensional (3D) structure of a fully assembled PDHc (E2 core
saturated with the peripheral subunits E1 and E3) is available to date
at high resolution. Although atomic resolution models have been
solved for the individual components E1 (21, 33), E2 (13, 14, 34, 35),
and E3 (20, 23). Recent advances in cryo–electron microscopy and
cryo–electron tomography, however, suggested a model wherein the
peripheral subunits are flexibly tethered to the core and cluster
around the vertices of the E2 cube (14, 36, 37).
Here, we present a biochemical, biophysical, and structural anal-
ysis of PDHc from Escherichia coli. Our work focused on the gen-
eration and characterization of a minimal PDHc composed of a
homotrimeric E2 variant whose oligomerization beyond trimers is
arrested and which can be fully saturated with E1 dimers and/or E3
dimers. By combining analytical size exclusion chromatography
(SEC), analytical ultracentrifugation, optical spectroscopy, enzy-
matic assays, x-­ray crystallography, negative staining electron mi-
croscopy (NS-­EM) and cryogenic electron microscopy (cryo-­EM),

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 1 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Simplified model of E. coli PDHc, domain organization of the E2 constructs, and assessment of the oligomeric state of CD24 and CD3 at pH 7.4. (A) Cartoon
representation of E. coli PDHc. E1 and E3 homodimers form a peripheral protein shell around the cubic E2 core by associating with the PSBD of E2. Models used were as
follows: 24-­mer of E2 CDs [shades of blue, representative trimeric unit colored red; PDB: 1EAB (15)], E1 with E2 PSBD [green and blue; PDB: 4QOY (21)], E3 with E2 PSBD
[orange and blue; PDB: 1EBD (20)], and LDs (LD1 to LD3) of E2 [blue or red; PDB: 1QJO (83)]. Black lines correspond to the flexible linkers between the domains of each E2
subunit. Note that the cartoon does not provide specific information on the dynamics, exact subunit composition and interaction network within PDHc. (B) Domain or-
ganization of full-­length E. coli E2 (E224) and truncated E2 variants. Flexible linker regions are depicted in gray. (C) Analytical SEC on Superdex 200 Increase 5/150 GL (3 ml
volume) of CD24 (blue) and CD3 (red) recorded at 220 nm. The profiles are consistent with the homotrimeric state of CD3 within experimental error. Elution volumes of
molecular mass standard proteins (in kDa) are indicated. (D) SV-­AUC analysis of CD3 at pH 7.4 and 20°C. Normalized sedimentation coefficient distributions c (s) for the
indicated initial CD3 concentrations are shown. In the tested concentration range, a single CD3 species at 4.9 S was observed, confirming homogeneity of the preparation.

we show that this minimal ~700-­kDa complex (mini-­PDHc) har-
bors all structural principles determining the stoichiometry of wild-­
type PDHc (~5 MDa) and already has enzymatic activity comparable
to that of the wild-­type complex. In addition, we confirm that the E1
dimer:E2 monomer:E3 dimer ratio of 2:3:1 found for mini-­PDHc is
preserved in the wild-­type complex. Specifically, this subunit com-
position is the consequence of a previously unknown dimerization
reaction between PSBD domains within each E2 homotrimer: While
two E1 dimers associate with dimeric PSBD, E3 exclusively interacts
with the remaining unpaired PSBD. This balance ensures sufficient
E1 binding in the presence of excess E3 and sufficient E3 binding at
excess of E1. Last, on the basis of sequence conservation analysis, we
propose that this mechanism is conserved across Gammaproteobac-
teria, rendering the structural principles in E. coli PDHc applicable
to PDHcs in a multitude of other species, including medically rele-
vant human pathogens such as Pseudomonas aeruginosa, Salmonella
enterica, or Vibrio cholerae.
RESULTS
A two residue, C-­terminal truncation of E2 yields a
stable E2 trimer
Structural studies on the 24-­meric wild-­type E2 core (E224) from E. coli
had been hampered by its size and the flexibility of the N-­terminal
E2 domains. As only the CD of E2 contributes to the assembly of the
E224 core complex, we created a truncated construct of wild-­type E2
by deleting all LDs (LD1 to LD3) and PSBD of E2. We will refer to this
construct, spanning E2 residues 381 to 629, as CD24. To arrest oligo-
merization beyond trimers, we further deleted the two C-­terminal resi-
dues V628 and M629 that are involved in intertrimer interactions
(13, 14). This CD variant comprises residues 381 to 627 and will be re-
ferred to as CD3 (see Fig. 1B for nomenclature and domain organi-
zation of all E2 constructs used in this study). Analytical SEC revealed
a single peak corresponding to a considerably smaller hydrodynam-
ic radius of CD3 compared to CD24, which was consistent with the
homotrimeric state of CD3 within experimental error (Fig. 1C and
fig. S1). To assess homogeneity and to confirm the trimeric oligo-
merization state of CD3, we performed sedimentation velocity ana-
lytical ultracentrifugation (SV-­ AUC) experiments with different
initial CD3 concentrations. For all samples, a single species was ob-
served, with a sedimentation coefficient at 20°C and infinite dilution
in water (s020,w ) of 4.87 ± 0.01 S (Fig. 1D). Using sedimentation equi-
librium AUC (SE-­AUC), the molecular mass of the CD3 complex
was determined to 79.8 ± 0.4 kDa (fig. S2), confirming that CD3 was
indeed a homotrimer in solution (calculated molecular mass of CD3:
82.1 kDa). Far–ultraviolet (UV) circular dichroism spectra and co-
operative, temperature-­induced unfolding transitions indicated that
both CD3 and CD24 adopted well-­defined folds (fig. S3, A and B).
To confirm that the E2 fold was preserved in CD3 and to reveal
the structural basis of its arrested assembly to a 24-­mer, we crystal-
lized the protein and solved its structure at 1.97-­Å resolution with
molecular replacement using an E. coli E2 model [Protein Data
Bank (PDB): 4N72 (38)] (Fig. 2A). In addition, for a direct reference
to our CD3 structure, we determined the cryo-­EM structure of CD24
and reconstructed the map to 3.3-­Å global resolution with local
resolution estimates between 2.7 and 3.5 Å (Fig. 2B). As expected,
our analysis yielded a cubic assembly composed of eight E2 ho-
motrimers, each situated at the vertices of the cubic complex. The
two models of CD3 and CD24 resemble the previously published tri-
mer structure by Wang et al. (38) with an average root mean square

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 2 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Structures of the CD24 and CD3 complexes. (A) Overlay of CD3 (red) and a trimeric unit of CD24 (blue), demonstrating the two structures as highly similar. The
RMSDCα is provided. (B) Cryo-­EM map of the CD24 complex at 3.3-­Å resolution. The map is colored according to the local resolution estimates (color key provided; Å). The
interior part of the cube was reconstructed at about 2.7-­Å resolution, while the exterior region was refined to 3.0-­to 3.5-­Å resolution. The trimer-­trimer interface is high-
lighted with a gray box. (C) Interface between neighboring trimers in CD24 (top) and the corresponding view of CD3 (bottom). In the cubic CD24 complex, the two neigh-
boring trimers form a conserved E2 knob-­into-­hole interaction. Residues 624 to 629 (side chains shown) form the 310-­helix that binds in the hydrophobic pocket of a
neighboring trimer. This interaction is abolished in the trimeric CD3, where the C-­terminal L627 binds the hydrophobic groove of the same monomer (magnified view in
the bottom panel). (D) cryo-­EM and electron density maps for the C-­terminal segments of CD24 (top) and CD3 (bottom), respectively.

deviation for all alpha carbons (RMSDCα) per monomer of 1.01 and
0.62 Å, respectively. Both of our structures have a virtually identical
CD fold with an RMSDCα of 1.15 Å (Fig. 2A), indicating that the
removal of the last two amino acid residues neither impaired folding
of the CD nor its association to trimers but only inhibited 24-­mer
formation.
The biochemical and structural data showed that the small,
C-­terminal two-­residue truncation was sufficient to completely pre-
vent intertrimer contact formation and further oligomerization to
large, cubic complexes. Our CD24 model confirmed a conserved
“knob-­into-­hole” interaction between the neighboring E2 trimers
(Fig. 2, C and D, top panels), as first described by Mattevi et al. (13).
Here, residues 624 to 629 of one trimer form the short C-­terminal
310-­helix (knob) that fits into a hydrophobic pocket in a neighbor-
ing trimer (hole). Notably, the deletion of V628 and M629 in CD3
prevented further oligomerization via a different mechanism: It
breaks the C-­terminal 310-­helix, causing residues 624 to 627 to re-
main in an extended conformation and allowing L627 to dock intra-
molecularly into the hydrophobic groove that would be occupied by
V628 and M629 of a neighboring trimer in wild-­type E224 (Fig. 2, C
and D, bottom panels). This mechanism of arrested assembly differs
from that in a previously reported CD trimer, in which further as-
sembly was prevented sterically via a C-­terminal peptide exten-
sion (38).
The complex formed between E2, E1, and E3 has a
defined stoichiometry
To determine the binding stoichiometry of the peripheral sub-
units to trimeric E2, we used a C-­terminally truncated variant of
E2 (E23, residues 1 to 627) analogous to CD3 but still retaining the
LD domains LD1 to LD3 and the PSBD domain. We analyzed E23
for its ability to associate with E1 dimers and E3 dimers using ana-
lytical SEC. The same approach was used in parallel for the wild-­
type E224 core to directly compare the assembly of E23 and E224 with
the peripheral subunits. The results are summarized in Table 1
and Fig. 3.
When E224 (constant concentration) was titrated with E1 alone,
saturation was reached at 22.3 bound E1 dimers, which is very close
and within experimental error to a 1:1 ratio between E2 monomers
and E1 dimers. The same titration experiment with constant E23
concentration yielded 2.5 E1 dimers bound to E23 (Fig. 3A), corre-
sponding to a mixture of complexes with three (saturated) and two

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 3 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Table 1. Stoichiometry of E1 and E3 binding to E2, indicated as the fraction of occupied E2 monomers after saturation. n.a., not applicable.
E1 added E3 added Presaturation with E3 E1 added Presaturation with E1 E3 added

E23 0.85 ± 0.02 0.87 ± 0.05 0.66 ± 0.03 0.31 ± 0.02

E224 0.93 ± 0.05 0.75 ± 0.04 0.71 ± 0.04 0.35 ± 0.02
E224-­V364K 0.61 ± 0.04* 0.79 ± 0.05 n.a. 0.79 ± 0.05
E23-­V364K 0.86 ± 0.05 0.89 ± 0.03 n.a. 0.83 ± 0.06

*This value is underestimated due to partial dissociation of the complex during SEC (see fig. S9).

Fig. 3. Binding reactions between the homotrimer E23 and the peripheral subunits E1 and E3. The ratio of E1/E3 dimers bound per E2 monomer is plotted against
the respective ratio of E1 or E3 dimers added per E2 monomer (ratio of ligands per binding site). The ratio of E1/E3 dimers bound per E2 monomer was calculated from
the peak areas of complexed subunits in the respective chromatograms recorded at 280 nm [(A) and (C)] or 450 nm [flavin adenine dinucleotide (FAD) cofactor in E3]
[(B) and (D)] and normalized using the total peak area as a calibration curve. (A) Saturation of E23 with E1. (B) Saturation of E23 with E3. (C) Binding of E1 to the saturated
E3:E23 subcomplex displaces two E3 dimers. (D) Binding of E3 to the saturated E1:E23 subcomplex displaces one E1 dimer.

E1 dimers bound to E23. The results thus showed that all PSBD do-
mains in E23 and E224 can be saturated with E1 dimers.
Titration of E23 and E224 with E3 alone yielded 2.6 bound E3 dimers
for E23 and 17.9 bound E3 dimers for E224 (Fig. 3B). For E224, the
recorded stoichiometry is in good agreement with a previous report
by Reed et al. (25), where 22 E1 dimers and 19 E3 dimers were found
to bind to the E2 24-­mer.
Next, we investigated the competition between the peripheral
subunits E1 and E3 for binding to PSBD in E2. For this purpose, E23
or E224 were presaturated with either E1 or E3 and then titrated with
E3 or E1, respectively. The results, shown in Table 1 and Fig. 3, dem-
onstrated that each peripheral subunit could displace the other sub-
unit from the respective presaturated complex. Specifically, we
observed an average of 2.0 and 17.1 E1 dimers binding to the satu-
rated E3:E23 and E3:E224 subcomplexes, respectively. Reversely, an
average of 0.9 E3 dimers displaced E1 from the E23 trimer, while 8.3
E3 dimers associated with the E1:E224 complex. Figure 3 schemati-
cally summarizes the observed displacement reactions and their
stoichiometries for E23.
Together, our analysis of E224 in complex with E1 and E3 differs
from the previously published data on PDHc isolated from its na-
tive environment in that the reconstitution of the entire PDHc
from individual, purified subunits demonstrated that, indeed, all
24 PSBD domains of the E224 core can be saturated with the pe-
ripheral subunits: In contrast to the reported stoichiometry of
12 E1 dimers and 6 E3 dimers bound per E224 (25–28), we found
~16 E1 dimer and ~8 E3 dimer equivalents per E224 at saturating
concentrations of E1 and E3, corresponding to an E1 dimer:E2
monomer:E3 dimer ratio of 2:3:1. Notably, this PDHc subunit ratio
could be fully reproduced for the complex of E23 saturated with E1
and E3: Fig. 3 (C and D) shows that two E1 dimers and one E3 di-
mer associated with the E23 trimer, forming a complex that we will
refer to as mini-­PDHc. Consequently, the fully saturated PDHc
wild-­type complex can be considered a supramolecular structure
composed of eight identical mini-­PDHc subcomplexes located at
the vertices of the cube.
Next, we measured the overall enzymatic activity of mini-­PDHc
relative to wild-­type PDHc under nearly saturating concentrations
of the three substrates pyruvate, CoA, and NAD+ (fig. S3D). Mini-­
PDHc showed 60% of the wild-­type PDHc activity (fig. S3E). This
demonstrates that intersubunit exchange of catalytic intermediates
and facilitated substrate diffusion are already largely achieved with-
in the mini-­PDHc complex, making it a suitable target for further
structural investigation.

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 4 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Structural analysis of mini-­PDHcΔLD reveals PSBD-­mediated
linkage of two E1 dimers
For further structural characterization of mini-­PDHc with SV-­AUC
and EM, we constructed a simplified E2 variant in which the flexibly
attached, N-­terminal LDs were deleted (ΔLD). The resulting, ho-
motrimeric variant PSBD-­CD3 (residues 314 to 627; Fig. 1B) re-
tained the ability to bind E1 and E3 at the same molar ratios observed
for mini-­PDHc (fig. S4). The corresponding complex with two E1
dimers, the PSBD-­CD3 homotrimer, and one E3 dimer will be re-
ferred to as mini-­PDHcΔLD in the following. After purification by
SEC, mini-­PDHcΔLD sedimented as a single species with a sedimen-
tation coefficient (s20,w) of 16 S (Fig. 4A), demonstrating that mini-­
PDHcΔLD had a defined stoichiometry. The apparent molecular
mass calculated from the sedimentation data was ~530 kDa, consis-
tent with binding of two E1 dimers and one E3 dimer to one trimer
of PSBD-­CD3 (calculated mass: 604 kDa).
To gain more structural insight into the interaction between the
peripheral subunits and the E2 core, we analyzed mini-­PDHcΔLD
by EM. First, the complex was analyzed by NS-­EM to assess the or-
der and homogeneity of the particles. 2D class averages obtained
during the single-­particle analysis revealed two easily distinguish-
able subpopulations of particles. Under the applied NS-­EM condi-
tions, approximately 20% of all particles had two E1 dimers and one
E3 dimer bound (Fig. 4B), corresponding to the stoichiometry of
mini-­PDHcΔLD as determined via SEC and AUC. The remaining
80% of the particles consisted of two E1 dimers bound to PSBD-­
CD3 (Fig. 4C), from which E3 likely had dissociated during EM grid
preparation. We consider the loss of E3 an NS-­EM artifact, as this
dissociation reaction was never observed during SEC or AUC in so-
lution. The classes obtained for the two subpopulations of particles
are highly similar and follow an identical pattern: Two E1 dimers are
located close to each other, whereas E3 (if present) is more flexibly
bound on the side of the trimer of the E2 CDs and separated from
the E1 dimers.
In addition to the densities corresponding to E1, E2 CD, and E3,
a V-­shaped structure was observed between the E1 dimers and the
E2 CD trimer. This density appeared to orient the two E1 dimers in
a defined relative orientation and to prevent more random, unre-
strained E1 arrangements. The location of this density and the ob-
servation of apparent bridging between the E1 and the E2 subunits
suggested that the structure corresponded to two associated PSBD
domains within PSBD-­CD3 (Fig. 4, B and C).
Analysis of the low-­resolution maps of mini-­PDHcΔLD with
(Fig. 4D) and without (Fig. 4E) the E3 component showed that all
densities observed in 2D were resolved well enough to fit crystal
structures of the individual proteins. The estimated distances be-
tween E1 dimers or E3 dimers and the E2 CD trimer were approxi-
mately 50 to 60 Å. This is in agreement with the sum of folded PSBD
(~32 Å) (21) and the length of the flexible, 11-­residue linker con-
necting the PSBD with the CD (~38 Å for the fully extended linker,
E2 residues 374 to 384).
The PSBD density observed in our 2D class averages could also
be resolved in 3D demonstrating a clear connection between the
two bound E1 dimers. Despite the low resolution, the continuous
density observed here suggested an interaction mechanism in which
two of the three PSBDs of PSDB-­CD3 dimerize, the PSBDs in the
dimer preferentially bind E1 dimers, and the remaining unpaired
PSBD preferentially bind an E3 dimer.
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
Homodimeric PSBD binds two E1 dimers
To analyze a putative dimerization of PSBD, we recombinantly pro-
duced the isolated PSBD domain spanning E2 residues 314 to 379
and characterized the protein structurally via x-­ray crystallogra-
phy and EM.
The crystal structure of PSBD was solved to 1.64-­Å resolution. As
predicted from our EM analysis, PSBD indeed associated to a V-­
shaped homodimer (Fig. 5A and fig. S5) formed by two parallel α
helices (residues 331 to 339 and 357 to 372) from each monomer
that are connected with a 19-­residue long loop. Our refined model
closely resembles the available crystal structure of monomeric PSBD
from Geobacillus stearothermophilus, with RMSDCα of 0.74 Å (39).
While the N-­terminal region of PSBD (residues 327 to 334 and 350
to 354) had been shown to interact with peripheral subunits (20, 21),
our crystal structure showed that the more C-­terminal PSBD seg-
ment (residues 335 to 339 and 360 to 368) is responsible for dimer for-
mation. Hydrophobic residues line the interface between the PSBD
monomers, indicating that a network of hydrophobic interactions
holds them together, with a buried surface area of 872 and 844 Å2
per monomer according to PDBePISA (40) and COCOMAPS (41),
respectively (Fig. 5B). The amino acid residues identified as key con-
tributors to the hydrophobic interface include F339, V364, and I368
(Fig. 5C). Furthermore, SE-­AUC showed that purified PSBD is a
homodimer in solution with an experimentally determined mass of
14.0 ± 0.1 kDa (calculated mass of the dimer: 14.6 kDa, fig. S6).
Given that PSBD monomers were not detected by SV-­AUC even at
the lowest PSBD concentration used (8.9 μM PSBD monomer), we
can estimate an upper limit for the dissociation constant (KD) of the
isolated PSBD dimer of ~0.1 to 1.0 μM.
Next, we used NS-­EM to investigate whether isolated PSBD can
interact with E1 and if this interaction occurs between two E1 di-
mers, analogous to mini-­PDHcΔLD. The 2D class averages revealed
that the E1:PSBD complex shows a characteristic pattern of two
densities held close to each other (Fig. 5D). While the strong density
bridging the E1 subunits further confirmed the ability of dimeric
PSBD to associate with two E1 dimers, the complex appeared con-
formationally heterogeneous due to variable distances between the
E1 dimers.
To gain higher resolution, we attempted to solve the structure of
mini-­PDHcΔLD using cryo-­EM. Because of the flexible linker be-
tween PSBD and CD, chemical crosslinking with bis-­(succinimidyl)
suberate (BS3) was used to visualize the preferred conformational
states of the particles and to improve resolution. The obtained 2D class
averages showed that the E3 component could not be crosslinked to
PSBD and dissociated, possibly upon freezing (Fig. 5E). The two E1
dimers, flexibly connected with the PSBD dimer, adopted two main,
symmetrical conformations in the complex and were either very
close to each other, almost touching, or fully separated (Fig. 5E). The
2D alignments were driven primarily by PSBD-­linked E1 dimers, so
that the density for CD3 appeared blurry, indicating considerable
conformational freedom within the complex.
Focusing the alignments on a single E1 dimer to later extend onto
PSBDs not only allowed us to reconstruct dimeric E1 at ~4-­Å global
resolution but also revealed a high degree of heterogeneity in the
dataset (fig. S7). As expected, the estimated local resolution of our
map was nonuniform, where the E1 dimer was refined to 3.5 to 6 Å
while the density for PSBD was reconstructed to 6-­to 10-­Å resolution
(Fig. 5F). For model fitting, we used AlphaFold (42, 43) to predict a
5 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. SV-­AUC and NS-­EM analyses of mini-­PDHcΔLD. (A) Sedimentation coefficient distributions c (s) of mini-­PDHcΔLD. The initial concentrations of the complex
and the calculated sedimentation coefficient s20,w are indicated. Data were recorded at 20°C, pH 7.4. (B) Selection of 2D class averages generated for a subset of par-
ticles where E3 remained bound (mini-­PDHcΔLD). Protein components are labeled. (C) Selection of 2D class averages for a subset of particles from which E3 dissoci-
ated. Protein components are labeled. (D and E) NS-­EM density maps of mini-­PDHcΔLD (D) and mini-­PDHcΔLD from which E3 dissociated (E) colored in green (E1), red
(E2), and orange (E3) according to the identity of the complex subunit. Crystal structures of E. coli E1 with PSBD [PDB: 4QOY (21)], E. coli E3 [PDB: 4JDR (23)], and CD3
of E. coli E2 were fitted into the maps using UCSF ChimeraX (model map correlations of 0.88 to 0.92 at 15-­Å resolution) and are shown below the maps. Approximate
distances between the proteins are indicated.

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 6 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. Structure of E. coli PSBD homodimer and its binding to two dimers of E1. (A) Crystal structure of the dimeric PSBD of E. coli E2 at 1.64-­Å resolution. The two
monomers are colored red and pink, respectively. (B) Top view of the crystal structure shown in (A). Hydrophobic residues stabilizing a dimer interface are colored blue.
(C) Magnified view of the hydrophobic interface shown in (B). The residues with two or more hydrophobic contacts within <4-­Å distance are labeled. (D) NS-­EM 2D
class averages obtained for PSBD complexed with E1. Dimeric PSBD associates with two E1 dimers. Each density is identified for one class. (E) Cryo-­EM 2D class aver-
ages of the BS3-­crosslinked mini-­PDHcΔLD. (F) Model of the complex between one E1 dimer and the PSBD dimer, rigid-­body fitted into the cryo-­EM map after 3D vari-
ability analysis and local refinement. The nontransparent map is shown on the right and colored according to local resolution estimates. Models used were as follows:
PDB: 4QOY (E1 residues 56 to 886) and AlphaFold prediction of four E1 chains (residues 1 to 58) in complex with two PSBD chains (E2, residues 314 to 379). A downsized
copy of the AlphaFold model, colored according to per-­residue confidence score [predicted local distance difference test (pLDDT)], and the predicted aligned error
plot are provided. The low-­confidence regions correspond to disordered linkers between PSBD and CD of E2 and between the N-­terminal PSBD-­binding E1 segment
and the rest of the E1 dimer.

model of four E1 chains (E1 segments 1 to 58) in complex with two
PSBD chains (E2 residues 314 to 379). This model and the crystal
structure of the E1 dimer (PDB: 4QOY, residues 56 to 886) were
rigid-­body fitted into the map. While the E1 structure fitted well
into the EM density reconstructed at ~4-­Å resolution, the fit of the
AlphaFold model was more ambiguous within the less resolved re-
gion. Nevertheless, the E1:PSBD model presented here is in good
agreement with the observations made from 2D class averages and
explains how dimeric PSBD not only bridges between E1 and E2 but
also brings the two E1 dimers into close proximity.
PSBD dimerization is likely conserved in PDHcs
of Gammaproteobacteria
So far, PSBD of bacterial E2 subunits had only been described as a
monomeric domain (34, 44–46). We therefore investigated the de-
gree of sequence conservation among the prokaryotic PDHcs. On
the basis of AlphaFold predictions, we hypothesized that the PSBD
dimerization mechanism observed for E. coli E2 may be intrinsic to
gammaproteobacterial PDHcs. PSBD sequence analysis revealed a
high overall sequence similarity (more than ~60% relative to E. coli
PSBD) within this class of bacteria and near 100% conservation of
the residues forming the contacts with E1 or E3 (H327, P330, R333,
R337, and R354). Thus, the mode of interaction between E2 and the
peripheral subunits observed in E. coli PSBD is likely preserved
within Gammaproteobacteria (Fig. 6, A and B). The PSBD residues
V364 and K365 also show near full sequence conservation. They are
positioned in the middle of the second helix, distant from the bind-
ing interface with E1 or E3, indicating that they might play a role in
PSBD dimerization. Furthermore, the C-­terminal stretch of PSBD com-
prising residues 366 to 374 is presumed to elongate the helix and
further stabilize PSBD dimers, given that the hydrophobic nature of
amino acids pointing inward (A367, I368, and A371) is generally

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 7 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. PSBD sequence conservation and its interactions with the peripheral subunits E1 and E3. (A) Crystal structure of the E. coli PSBD colored according to the
degree of conservation among Gammaproteobacteria. The highly conserved residues, V364 and K365, which are not involved in binding E1 or E3, are labeled. (B) Se-
quence logo plot representing the per-­residue conservation within PSBD. The highly conserved V364 is marked with a star symbol to highlight the site of substitution with
lysine in the PSBDV364K variant. (C) Sedimentation coefficient distributions c (s) of PSBD (red) and PSBDV364K (blue) for 8.9 to 37 μM protein (monomer concentration).
Calculated sedimentation coefficients s20,w are indicated. Data were recorded at 20°C, pH 7.4. (D) Analytical SEC performed on the PSBD variants complexed with E1 (left)
or E3 (right). Elution volumes of molecular mass standard proteins (in kDa) are indicated. The monomeric PSBDV364K binds to a single dimer of either E1 or E3. Dimeric PSBD
associates with two dimers of E1, whereas E3 binds to monomeric PSBD. Dashed line denotes elution profile of an equimolar mixture of E1 dimers and PSBD dimers. All
E1 dimers accumulated in the peak corresponding to the complex of two E1 dimers bound to dimeric PSBD, showing that E1 binding is cooperative.

conserved. Another key residue of the hydrophobic PSBD dimer in-
terface is F339, which is either a phenylalanine or leucine in gam-
maproteobacterial PSBDs.
To confirm the role of V364 as a critical residue in the hydropho-
bic interface of the PSBD dimer, we substituted V364 with lysine
in the variant PSBDV364K. SV-­AUC analysis showed that PSBDV364K
sedimented as a uniform monomer with s20,w of 0.92 S and apparent
molecular mass of ~7.5 kDa (calculated mass 7.3 kDa) compared
to s20,w of 1.43 S and ~13.9 kDa determined for dimeric PSBD
(calculated mass 14.6 kDa) (Fig. 6C).
To test if the V364K substitution in PSBD affected the interac-
tion between PSBD and the peripheral subunits, we performed
analytical SEC runs for the complexes between E1 or E3 and either
the PSBD dimer or the PSBDV364K monomer (Fig. 6D). The samples
of E1 complexed with the two PSBD variants yielded distinct chro-
matograms. While only a small shift toward higher mass was ob-
served in the elution profile of the E1 dimer (~200 kDa) after the
addition of PSBDV364K (~7-­kDa mass increase due to binding of
PSBDV364K), the retention volume of the complex between the
PSBD dimer and E1 indicated formation of two E1 dimers bind-
ing per PSBD dimer (Fig. 6D, left). This is in full agreement with
the EM data described above, where particles of two E1 dimers
linked by dimeric PSBD were detected (Fig. 5D). In contrast, E3
showed small, essentially identical peak shifts toward higher mass
when incubated with the PSBD dimer or the PSBDV364K monomer
(Fig. 6D, right). This demonstrates that E3 dimers form stable
complexes exclusively with PSBD monomers.
Notably, analytical SEC also revealed that E1 binds cooperatively
to the PSBD dimer. Specifically, when E1 was mixed with an equi-
molar amount of PSBD dimers, all E1 dimers accumulated in the
saturated complex (two E1 dimers per PSBD dimer), while no com-
plexes comprising only a single E1 dimer could be detected. This
demonstrates that binding of the first E1 dimer to dimeric PSBD
strongly favors binding of the second E1 dimer (Fig. 6D, left).
For a further analysis of the PSBD:E3 interaction, complexes
between E3 dimers and either one or two PSBD chains were modeled
using AlphaFold. A structure of the E32:PSBD1 complex (single
PSBD chain) could be predicted with high confidence and in good

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 8 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
agreement with the crystal structure from G. stearothermophilus
(20) where PSBD was shown to bind as a monomer at the dimer
interface of E3 (fig. S8A). Although symmetric E3 homodimers have
two potential binding sites for monomeric PSBD, these cannot be
occupied simultaneously by two PSBD chains due to steric clashes
(fig. S8B). In contrast, the N-­terminal region of PSBD responsible
for E3 binding was predicted with considerably lower confidence for
E32:PSBD2 (two PSBD chains) (fig. S8, C and D), as the second
chain of the PSBD dimer caused steric hindrance preventing it from
interacting with the native binding site of the E3 dimer. In conclu-
sion, the AlphaFold models described above support our SEC and
EM data and point to monomeric PSBD as the single binding tar-
get for E3.
Dimerization of PSBD dictates PDHc stoichiometry
On the basis of the above results, we hypothesized that PSBD
dimerization within an E2 homotrimer is the critical determinant
for the subunit composition of E. coli PDHc and possibly many
other gammaproteobacterial PDHcs. To test this hypothesis, we
introduced the V364K substitution into E224 and E23 to disrupt the
PSBD dimerization interface and analyzed the resulting E224-­V364K
and E23-­V364K variants for complex formation with the peripheral
subunits using analytical SEC (Table 1, Fig. 7, and fig. S9). The
results demonstrated that E1 and E3 subunits exhibit different
preferences toward monomeric and dimeric PSBD. While E1 dimers
bind preferentially to dimeric PSBD, E3 displaces only E1 dimers
associated with monomeric PSBD.
As expected for a blocked PSBD dimerization in the V364K
variants of E2 and the inability of E3 to bind dimeric PSBD, we
observed no notable differences between wild-­type and mutant
E2 constructs in their maximum occupancy with E3 dimers (one
E3 dimer per E2 monomer; Table 1 and Fig. 7E). For binding of
E1, however, the occupancy of E2 monomers in E224-­V364K at a
2.5-­fold excess of E1 dimers dropped from 93% (E224 wild-­type)
to 61% (Fig. 7A and Table 1). The titration profile for E224-­V364K
also showed a clear curvature, indicative of weaker binding of
E1 dimers to monomeric PSBD in E224-­V364K (Fig. 7A) and pos-
sibly dissociation of E1 from the complex during SEC (fig. S9).
The preference of E1 dimers for dimeric compared to mono-
meric PSBD is best illustrated by the observation that E1 can no
longer displace E3 from PSBD monomers in E224-­V364K or E23-­
V364K. (Fig. 7, B to D, and Table 1). Conversely, E3 dimers essen-
tially displaced all E1 dimers from E224-­V364K or E23-­V364K (Fig. 7,
F and G, and Table 1), while two of three E1 dimers remained
bound in wild-­type E224 or E23 at an excess of E3 (Fig. 7H and
Table 1).
From the presented data, we conclude that two E1 dimers are
strongly bound to dimeric PSBD in E23 fully saturated with E1
dimers, and one E1 dimer is bound weakly to the remaining un-
paired PSBD. Only this weakly bound E1 dimer can then be dis-
placed by an E3 dimer, yielding a final E1 dimer: E2 monomer: E3
dimer ratio of 2:3:1 (Fig. 7H). This mechanism also guarantees that
the 2:3:1 ratio is maintained at oversaturation with E1 or E3 di-
mers. The preferred binding of two E1 dimers to a PSBD dimer
and the exclusive binding of E3 dimers to monomeric PSBD is also
consistent with the fact that an E1:E3 ratio of 2:1 was found in
most previous reports on the stoichiometry of E. coli PDHc in the
majority of which the face/edge model of PDHc was assumed (10,
12, 25, 27, 29).
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
DISCUSSION
Over the last decades, numerous reports provided important mech-
anistic and structural insights into the multienzyme assembly of
E. coli PDHc (14, 23, 33–37, 47). The biochemical, biophysical, and
structural analyses of wild-­type PDHc and its minimized variant
mini-­PDHc presented here allowed us to gain more detailed insight
into the structural basis of subunit stoichiometry and to identify the
PSBD of E2 as the main determinant of PDHc architecture.
To simplify the biological assembly and therefore facilitate data in-
terpretation, we followed an approach similar to that described for
PDHc from Thermoplasma acidophilum (31) and G. stearothermophilus
(48) and generated an E2 variant with arrested oligomerization be-
yond trimers. The mechanism of arrested assembly through deletion
of the C-­terminal two E2 residues was revealed by x-­ray crystallog-
raphy and cryo-­EM performed on the corresponding E2 CD constructs,
CD3 and CD24. The hydrophobic pocket, which is natively occupied by
the last two residues of a neighboring E2 trimer, was self-­complemented,
preventing formation of intertrimer contacts. Breakage of the C-­
terminal 310-­helix observed in our CD3 crystal structure results in
an extended C-­terminal stretch, which is functionally reminiscent
of the corresponding region in natively homotrimeric E2 found in
actinobacteria (32). There, a three-­residue insertion preceding the
C-­terminal 310-­helix shifts the helix orientation, causing it to occupy
the hydrophobic pocket in an intramolecular interaction. Our find-
ings, along with the reports mentioned above, highlight the critical
role of the E2 C-­terminal residues in mediating intertrimer contacts,
thereby determining the oligomeric state of the PDHc core complex.
Generation of the stable, trimeric E2 variant allowed us to as-
semble enzymatically active mini-­PDHc. We characterized the as-
sociation of peripheral subunits to monomeric and dimeric PSBD
and found differences that imply modulation of binding affinities of
E1 and E3 dependent on the oligomeric state of PSBD. While E1
dimers associate preferentially with dimeric PSBD, E3 dimers only
bind monomeric PSBD. Furthermore, given that E1 is only capable
of displacing E3 from the E3:E2 subcomplex carrying the wild-­type
(dimerization-­competent) PSBD, a synergistic effect between bind-
ing of peripheral subunits and PSBD dimerization may be present. It
is conceivable that PSBD dimerization ensures accommodation of
both E1 and E3, and thus PDHc activity, should the relative intracel-
lular concentrations of the peripheral subunits change.
In the 24-­meric E2 core, the dimerization of PSBDs belonging to
neighboring trimers is prevented by their spatial organization around
the cubic CD24 architecture. Within a single E2 trimer, the N termini
of the CDs converge at the threefold symmetry axis of the E2 trimer.
The distance between the corresponding threefold symmetry centers
of neighboring vertices is ~90 Å, rendering the formation of inter-
trimer PSBD dimers within the same PDHc complex impossible
(fig. S10A). Moreover, constraining the linkers to emerge from the
center of trimers increases the local PSBD concentrations (fig. S10B),
facilitating PSBD dimerization while leaving the third PSBD mono-
meric. This could also explain the crowding of flexibly tethered pe-
ripheral subunits around the E2 vertices previously reported by
Murphy and Jensen (37) and Škerlová et al. (14).
Our SEC experiments demonstrated that all 24 binding sites of
E224 can be occupied with a mixture of E1 and E3, whereas only 93
and 75% saturation can be reached in the presence of only E1 or E3,
respectively. Single-­particle cryo-­EM analyses on the saturated and
BS3-­crosslinked E1:E224 and E3:E224 subcomplexes suggest different
diameters for the peripheral protein shell formed by E1 alone or E3
9 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Effects of the V364K substitution in the PSBD domain on the stoichiometry of E1 and E3 binding. (A, B, E, and F) Titration of the E224 and E224-­V364K variants
with E1 and E3, analyzed by analytical SEC. The ratio of E1/E3 dimers bound per E2 monomer is plotted against the respective ratio of E1 or E3 dimers added per E2 mono-
mer (ratio of ligands per binding site). The ratio of E1/E3 dimers bound per E2 monomer was calculated from the peak areas of complexed subunits, determined from
chromatograms recorded at 280 nm [(A) and (B)] or 450 nm (FAD in E3) [(E) and (F)], and normalized using the total peak area as a calibration curve. The data obtained for
E224 is shown in light blue squares, and those for E224-­V364K is shown in dark blue triangles. (C, D, G, and H) Schemes summarizing the results of the titration experiments
in which preformed complexes between E2 trimers and either E1 or E3 were titrated with E3 or E1, respectively. The purple star indicates the V364K substitution within the
PSBD of E2. E1 dimers are depicted in green, E3 dimers are depicted in orange, and E2 trimers are depicted in blue.

alone (fig. S11). Specifically, the analysis of 2D class averages yielded
broad distance distributions with mean distances between periph-
eral subunits and the CD24 core of ~32 Å for E1 and ~15 Å for E3.
While E1 dimers appear to bind further away from the E2 core,
greatly expanding the overall volume of the particle, dimers of E3
tend to associate closer to the E2 surface. This discrepancy could be
attributed to the monomeric nature of E3-­bound PSBD and its cor-
responding linker, which is highly flexible with regions of local stiff-
ening due to its high proline content (49). In the context of E3-­bound
PSBD, this linker can adopt various conformational states, as op-
posed to the linkers of E1-­bound PSBD, whose conformational free-
dom might become restricted upon dimerization, causing them to
adopt a more extended state, thus increase the diameter of the pro-
tein shell formed by E1.
The discovery of intratrimeric PSBD dimers provides a struc-
tural basis for interpretation of PDHc stoichiometry and organiza-
tion of the peripheral subunits around the E2 core. In this study, we
could identify mini-­PDHc as an asymmetric building block of the
full complex and hence conclude that the subunit stoichiometry is
not driven by the octahedral symmetry of the core but by the oligo-
meric state of PSBD. Under saturating conditions, when all binding
sites are occupied with peripheral subunits, the asymmetry of the
trimer-­based mini-­PDHc breaks the symmetry in the periphery. In
this context, the three chains of an E2 trimer are likely no longer

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 10 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 8. Model for the organization of the peripheral subunits around the E2 core. (A) A mini-­PDHcΔLD composed of trimeric PSBD-­CD, two E1 dimers, and one E3
dimer was modeled based on NS-­EM data (Fig. 4) and distance restraints (fig. S11) and fitted to one of the trimers of the CD24 core. E1 dimers are shown in green, E3 dimer
is shown in orange, the 24-­mer of the E2 CD is shown in blue, with one trimer highlighted in red. The linkers connecting the CD and PSBD domains of the E2 trimer
were modeled in Coot using geometric restraints and distances extracted from NS-­EM and cryo-­EM. (B) One of the possible models of fully assembled PDHc gener-
ated by fitting mini-­PDHcΔLD to each of the E2 trimers in the CD24 core. All dimeric and monomeric PSBDs are fully occupied with E1 and E3 dimers, respectively,
avoiding steric clashes. LD1 to LD3 of E2 are omitted for clarity; however, the presented model allows for accommodation of all 72 LD domains.

equivalent in the catalytic cycle of PDHc. For example, dimerization
of two E2 chains via PSBD might favor the reductive acetylation of
the 6 LD domains from the E2 dimer and affect the previously
reported interchain acetyl transfer and disulfide exchange between
LDs (50–52). The transfer of reducing equivalents from LDs in the
E2 dimer to the LDs in unpaired E2 might be particularly important
because the three LDs from the unpaired E2 chain could be the pre-
ferred E3 substrates.
Overall, wild-­type PDHc can be considered a cubic assembly of
eight asymmetric mini-­PDHcs, which have a defined stoichiometry
but are structurally heterogeneous. This heterogeneity results not
only from the flexibility of the E2 linkers but also from the stochastic
dimerization of two of the three PSBDs within each of the eight E2
trimers. Although PDHc does not form a single, rigid structure,
there are certain rules by which the complex assembles: (i) Each ver-
tex of the E2 24-­mer has one dimeric and one monomeric PSBD; (ii)
one E3 dimer associates with monomeric PSBD, while two E1 di-
mers bind cooperatively (Fig. 6D, left) to dimeric PSBD, saturating
the vertex; (iii) the average distance from the core is larger for E1
compared to E3 (fig. S11). Figure 8A shows a model of 24-­meric E2
where one showcase vertex is saturated with E1 and E3 according to
these guidelines. Such an assembly can be extrapolated to all vertices
of the E2 cube, yielding a ~5.6-­MDa fully saturated complex, where
the exact positions of the peripheral subunits remain undefined
(Fig. 8B and movie S1). The organization of the peripheral subunits
around the E2 core in Fig. 8B is not static, as structural dynamics
and conformational heterogeneity are essential for the biological ac-
tivity of PDHc.
In conclusion, our findings provide insights into the architecture
of E. coli PDHc, a key complex in aerobic respiration, and present a
unique and previously unknown mechanism determining its compo-
sition. We confirmed that PDHc stoichiometry is not a result of simple
competitive binding of the peripheral subunits to the PSBDs of E2
but rather is determined by PSBD dimerization, which generates
distinct binding sites for E1 and E3. On the basis of sequence con-
servation analysis, we propose that this mechanistic principle applies
to PDHcs of all Gammaproteobacteria, thereby strengthening un-
derstanding of this critical metabolic enzyme in a wide range of bac-
terial species, including plant pathogens belonging to the Erwinia or
Xanthomonas genera, as well as human pathogens such as P. aeruginosa,
S. enterica, and V. cholerae.
MATERIALS AND METHODS
Plasmids
Genes encoding E1, E2, and E3 were amplified from an E. coli lysate
with polymerase chain reaction (PCR) using Q5 High-­Fidelity DNA
polymerase (New England Biolabs). The amplified genes were
cloned into the pET-­21a vector via the NdeI and SacI (E2 and E3)
and NheI and SacI (E1) restriction sites using T4 ligase (New England
Biolabs). The CD24 construct was derived from the E2 sequence by
PCR amplification of the respective fragment followed by ligation
into pET-­21a. All other modifications of the different E2 variants
were introduced using the QuikChange Lightning Site-­Directed
Mutagenesis Kit (Agilent Technologies).
Protein production and purification
Transformed E. coli BL21 (DE3) cells were grown with shaking at 37°C
in LB-­Miller medium (Becton Dickinson) with ampicillin (100 mg/liter)
to an optical density OD600 of 0.6 to 0.8. Media for production of the
E224 and E23 constructs were supplemented with 1 mM (±)-­α-­lipoic
acid (Sigma-­Aldrich), media for E3 production were supplemented
with 0.35 mM FAD (Sigma-­Aldrich), and media for production of
CD constructs were supplemented with additional NaCl (2 g/liter).

Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024 11 of 17

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Protein production was induced by the addition of 0.5 mM isopro-
pyl β-­​d-­1-­thiogalactopyranoside, and the cell suspensions were fur-
ther incubated with shaking for 4 hours at 37°C (PSBD constructs)
or 23 hours at 20°C (all other proteins). Frozen cell pellets were re-
suspended in lysis buffer [for E1: 20 mM Hepes-­NaOH (pH 7.4),
150 mM NaCl, 5 mM MgCl2, 2 mM thiamine pyrophosphate (TPP),
1 mM ethylenediaminetetraacetic acid (EDTA), and 1 mM dithioth-
reitol (DTT); for CD24: 20 mM Hepes-­NaOH (pH 7.4), 200 mM NaCl,
5 mM MgCl2, and 5 mM CaCl2; for all other proteins: 20 mM Hepes-­
NaOH (pH 7.4), 150 mM NaCl, 5 mM MgCl2, and 5 mM CaCl2].
Then, 1 mM phenylmethylsulfonyl fluoride, 50 μg/ml deoxyribonu-
clease I, and 1× cOmplete EDTA-­free protease inhibitor cocktail
(Roche) were added, and the cells were lysed by passing the suspen-
sion three times through a M110L microfluidizer (Microfluidics) at
11,000-­psi chamber pressure. The lysate was centrifuged at 4°C and
48,000g for 45 min, and the supernatant was collected for further
purification.
Briefly, the cleared lysates were subjected to precipitation with
ammonium sulfate, followed by hydrophobic interaction chroma-
tography, ion exchange chromatography (IEX), and SEC. The PSBD
constructs were purified by incubating lysates at 75°C for 10 min,
centrifuging, and subjecting the soluble fraction to IEX and SEC. For
further details on protein purification, please refer to table S1. All pro-
teins were flash-­frozen and stored at −80°C in GF buffer [20 mM
Hepes-­NaOH (pH 7.4), 140 mM KCl, 10 mM NaCl, and 1 mM MgCl2].
Determination of protein concentrations
The concentrations of the purified proteins were determined by
measuring absorbance at 280 nm using a Cary 300 spectrophotom-
eter (Agilent Technologies). Experimental molar extinction coeffi-
cients of the native proteins (ε280nm) were determined as described
by Gill and von Hippel (53) and were: ε280nm (E1) = 128,360 M−1
cm−1; ε280nm (E224) = 22,530 M−1 cm−1; ε280nm (E23) = ε280nm
(E23-­V364K) = 20,780 M−1 cm−1; and ε280nm (E224-­V364K) = 21,080 M−1
cm−1. E3 concentration was determined by measuring absorbance
of the FAD cofactor at 450 nm. An experimental FAD molar extinc-
tion coefficient at 450 nm of 11,170 M−1 cm−1 was corrected for its
decrease upon binding to native E3. For all other proteins, molar
extinction coefficients calculated by the Expasy ProtParam tool (54)
were used.
Analytical SEC
For analytical SEC, samples were prepared in GF buffer and loaded
onto a Superdex 200 Increase 5/150 GL column (Cytiva) connected
to a high-­performance liquid chromatography system (Agilent Tech-
nologies, 1100 Series). Chromatograms were recorded at 280 nm
(protein) and/or 450 nm (FAD cofactor bound to E3).
Analytical ultracentrifugation
Freshly thawed aliquots were used for CD3, PSBD-­CD3, and the
PSBD constructs. To assemble complexes between PSBD-­CD3 and
the peripheral subunits, the proteins were mixed in the following
molar ratio: 0.75:1:0.35 for mini-­PDHcΔLD (E1 dimer:PSBD-­CD3:E3
dimer) with PSBD-­CD3 at 35 μM monomeric concentration. After
incubation at room temperature for 5 min, the samples were loaded
on a Superdex 200 10/300 GL column (Cytiva) to remove unbound
subunits. All final samples were dissolved and diluted in GF buffer
[except for the CD3 construct in 20 mM Hepes-­NaOH (pH 7.4) and
150 mM NaCl] to give absorbance between 0.05 and 1.25 at 1.2-­cm
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
path length and a wavelength of 280 nm (samples containing PSBD-­
CD3), 230 nm (PSBD constructs), or both (CD3).
Sedimentation experiments were performed at 20°C in an Op-
tima XL-­A analytical ultracentrifuge (Beckman Coulter, USA) using
analytical cells containing double-­sector 12-­mm charcoal-­filled
Epon centerpieces and quartz windows. For SV-­AUC, samples with
different initial protein concentrations (400 μl) and reference buffer
(420 μl) were equilibrated at 20°C in the resting An-­50 Ti rotor for
at least 2 hours before acceleration to 40,000 rpm (CD3 and PSBD-­
CD3), 30,000 rpm (mini-­PDHcΔLD), or 50,000 rpm (PSBD and PSB-
DV364K). Radial scans of the cells were obtained by measuring
absorbance at 230 and/or 280 nm (depending on initial protein con-
centration), with a single acquisition in continuous mode and a step
size of 0.003 cm. Data were analyzed in SEDFIT according to a con-
tinuous sedimentation coefficient distribution model [c (s)] (55)
with frictional ratio, meniscus, and baseline as fitting parameters.
Calculated sedimentation coefficient distributions were normalized
using GUSSI (56).
For SE-­AUC experiments, samples containing the CD3 or PSBD
construct at different initial protein concentrations were succes-
sively centrifuged at 7400, 12,500, and 19,000 rpm (CD3) or 14,000,
24,000, and 36,000 rpm (PSBD) in an An-­50 Ti rotor. Sample
volumes were 140 and 200 μl for CD3 and PSBD, respectively. For
each rotor speed, radial scans were collected every 3 hours by mea-
suring absorbance at 280 nm (CD3) or 230 nm (PSBD), with a single
acquisition in step mode and a step size of 0.001 cm. Attainment of
equilibrium was assessed using the respective function in SEDFIT
(57). Datasets were globally fitted using SEDPHAT (58) and assum-
ing the presence of a single ideal species. Confidence limits for the
fitted molecular mass were determined using F statistics as imple-
mented in SEDPHAT (P = 0.95).
Values for buffer density, buffer viscosity, and protein partial specific
volumes were calculated using SEDNTERP (59): CD3 (0.7486 ml/g),
mini-­PDHcΔLD (0.7365 ml/g using four E1 chains, three PSBD-­CD3
chains, and two E3 chains), PSBD (0.7371 ml/g), and PSBDV364K
(0.7368 ml/g).
Circular dichroism experiments
Far-­UV circular dichroism spectra (205 to 255 nm) of CD3 and CD24
were recorded for 5 μM protein samples in circular dichroism buffer
[10 mM Mops-­NaOH (pH 7.4), 140 mM KCl, and 10 mM NaCl] using
a J-­715 spectropolarimeter (Jasco). For thermal unfolding, 5 μM CD3
or CD24 were prepared in circular dichroism buffer, and the melting
curves were recorded at 222 nm with a heating rate of 1°C/min between
25° and 95°C. To determine the apparent melting temperature Tm, the
data were fitted as described by Pace et al. (60) using the OriginPro
2018 software (OriginLab) assuming a two-­state model.
Crystallization, x-­ray data collection, and analysis
Crystallization plates for sitting drop vapor diffusion were set up at
21°C. The stock of 350 μM (9.5 mg/ml) CD3 in GF buffer was mixed
at a 1:1 ratio with precipitant solution and incubated at 4°C. Crystals
grew within 9 days in 100 mM tris-­HCl (pH 8.5), 200 mM MgCl2,
and 30% (w/v) polyethylene glycol (PEG) 4000; they were fished
from the drop and directly flash-­frozen. PSBD was concentrated to
3.25 mM (24 mg/ml) in GF buffer, and crystallization plates were set
up and incubated at 20°C with 1:1 mixing ratio with precipitant
solution. Crystals grew within a week in 100 mM sodium acetate
(pH 5.0), 2 mM ZnCl2, and 24% (w/v) PEG 6000. As cryoprotectant,
12 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
15% glycerol was added directly to the drop before the fished crys-
tals were flash-­frozen in liquid nitrogen.
X-­ray diffraction data were recorded at the X06DA beamline
(PXIII) at the Swiss Light Source (Paul Scherrer Institute, Villigen,
Switzerland). The data were indexed and scaled using XDS (61).
Phases were determined by molecular replacement using Phaser
(62) in the PHENIX software package (63). The structure of E. coli
E2 CD [PDB: 4N72 (38)] served as a search model for CD3; the
PSBD component of the E1:PSBD structure [PDB: 4QOY (21)] was
used as a search model for PSBD. The models were built in Coot (64)
and refined in PHENIX. Statistics are summarized in table S2, and
the atomic coordinates and structure factors are deposited in the
PDB under the accession codes 8OSY (CD3) and 8OQJ (PSBD). The
crystal structure of the PSBD dimer was analyzed for interface inter-
actions using PDBePISA (40) and COCOMAPS (41).
NS-­EM sample preparation and data collection
Mini-­PDHcΔLD was assembled by mixing 40 μM PSBD-­CD3 (mo-
nomeric concentration) with 30 μM E1 dimer and 15 μM E3 dimer
in GF buffer. After 5-­min incubation at room temperature, the sam-
ple was applied to a Superdex 200 Increase 5/150 GL column (Cytiva)
equilibrated in GF buffer. The collected sample was diluted to 55 nM
mini-­PDHcΔLD. E1:PSBD was prepared from 10 μM PSBD (mono-
meric concentration) and 12.5 μM E1 dimer. The samples were in-
cubated for 5 min at room temperature and injected onto Superdex
200 10/300 GL column (Cytiva) equilibrated in GF buffer. The sam-
ple was diluted to 35 nM E1:PSBD complexe.
NS-­EM grids were prepared as previously described (65). Briefly,
4 μl of the sample was applied to glow-­discharged (negative, 25 mA,
30 s; PELCO easiGlow, Ted Pella) Quantifoil copper EM grid (300 mesh)
coated with a carbon film. After a 1-­to 2-­min incubation, the sam-
ple was blotted away with a filter paper and the grid was washed in
two droplets of water. The grid was then stained for 1 to 2 min in two
droplets of 2% (w/v) aqueous uranyl acetate, blotted, and air-­dried.
NS-­EM grids were imaged on Morgagni 268 (100 kV) and Tecnai
F20 (200 kV, equipped with field emission gun) transmission elec-
tron microscopes (Thermo Fisher Scientific). For automated data
collection, the Tecnai F20 microscope equipped with the Falcon II
direct electron detector was used. Single micrographs were recorded
with the EPU software (Thermo Fisher Scientific) at a dose of ap-
proximately 50 e−/Å2 and a pixel size of 1.32 Å/pixel.
NS-­EM single-­particle analysis
All NS-­EM datasets were processed in a similar fashion. First, the
contrast transfer function (CTF) was estimated using the Patch CTF
Estimation routine in cryoSPARC v3.2 (66). After removal of poor-­
quality micrographs, a few hundred particles were manually selected,
extracted, and subjected to 2D classification. The selected 2D class
averages were then used as references for automated particle picking
from all the micrographs using the Template Picker. Particles were
extracted (200 pixel box size, binned to 2.31 Å/pixel for mini-­
PDHcΔLD; 100 pixel box size, binned to 3.96 Å/pixel for E1:PSBD)
and subjected to multiple rounds of 2D classification to remove
false-­positive picks before generating an ab initio model. The model
and selected particles were used in homogenous and/or nonuniform
3D refinements in cryoSPARC. In addition, the particles were also
imported (67) to Relion 3.1 (68, 69) where they were subjected to 2D
classification. The generated 2D class averages revealed complexes
with easily identifiable features.
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
Crystal structures of PDHc subunits were rigid-­body fitted into the
3D maps using UCSF ChimeraX (70), starting with an orientation
in which the C termini of the PSBD dimer pointed toward the
N termini of CD3. The correlation between the fitted models and
maps was determined using the fitmap command at resolution of
15 Å. The distances were measured between the N termini of the
E2 CDs and the N-­terminal E1 helices that bind PSBD. Because of
the higher orientational freedom of E3 and thus more ambiguous
fit to our density, the provided E2-­E3 distance was measured at the
closest point between the N terminus of the E2 CD and the sur-
face of E3.
Cryo-­EM sample preparation and data collection
Cryo-­EM datasets were recorded for two constructs: CD24 and mini-­
PDHcΔLD (E1:PSBD-­CD3:E3). CD24 was diluted to 66 nM complex
concentration. For mini-­PDHcΔLD, the complex was prepared by
mixing 12 μM PSBD-­CD3 (monomeric concentration), 8.1 μM E1
dimer, and 4.2 μM E3 dimer in GF buffer followed by a 5-­min incu-
bation at room temperature. The 1 μM mini-­PDHcΔLD was then chem-
ically fixed with an amine-­reactive crosslinker via incubation with
3 mM BS3 (Thermo Fisher Scientific) for 30 min at room tempera-
ture. The unreacted crosslinker was quenched by addition of 50 mM
tris-­HCl (pH 8.0) and incubation for 15 min at room tempera-
ture. Crosslinking and quality of the sample were assessed with SDS
polyacrylamide gel electrophoresis and NS-­EM. BS3-­crosslinked mini-­
PDHcΔLD complexes were then diluted to 125 nM before plunge-­
freezing on Quantifoil R2/2 holey carbon copper grids using a
Vitrobot Mark IV (Thermo Fisher Scientific) and the “floating carbon”
technique (71, 72). Briefly, a thin (1-­to 1.5-­nm) carbon film was floated
on sample for 1 min, recovered with the grid, and quickly mounted
in the Vitrobot set to 4°C and 100% humidity. A total of 5 μl of buffer
was then applied on the carbon side, blotted for 1 to 4 s, and plunge-­
frozen in liquid ethane.
The cryo-­EM data were recorded on two Titan Krios transmission
electron microscopes (300 kV; Thermo Fisher Scientific) equipped
with either K2 Summit (CD24) or K3 direct electron detectors (mini-­
PDHcΔLD), operating in counting mode, and using a slit width of
20 eV on a GIF-­Quantum energy filter (Gatan). EPU software was used
for the automated data collection (Thermo Fisher Scientific). The
CD24 dataset of ~3600 movies was recorded with a final pixel size of
0.84 Å/pixel and a dose of 80 e−/Å2. The data for mini-­PDHcΔLD
(~9300 movies) were collected with the same dose and a pixel size of
0.66 e−/Å2. For both datasets, the defocus was set between −1.4 and
−2.8 μm with 0.2-­μm increments.
Cryo-­EM single-­particle analysis
Early processing steps were identical for the cryo-­EM datasets of
CD24 and mini-­PDHcΔLD. First, the motion-­corrected and dose-­
weighted [MotionCor2 (73)] movies were imported into cryo-
SPARC v3.2 where the Patch CTF Estimation was used to estimate
the CTF, astigmatism, and relative ice thickness. After removal of
poor-­quality micrographs, a few hundred particles were manually
picked to generate 2D templates. Selected 2D class averages were
used for automated particle picking from 100 randomly chosen mi-
crographs using the Template Picker. After extraction and 2D clas-
sification, 2000 selected particles and 100 micrographs were used to
train the Topaz model (74, 75). Using the trained model, particles
were automatically picked from all micrographs, extracted, and sub-
jected to multiple rounds of 2D classification to remove false-­positive
13 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
picks. Dependent on the dataset, the next steps of processing were
adjusted according to the complex investigated.
The CD24 dataset of ~350,000 particles (256 pixel box size;
1.18 Å/pixel) was processed in cryoSPARC v3.2. Because of limited
particle orientations, octahedral symmetry was imposed starting
from an ab initio model and was kept throughout the processing.
The initial reference and particles were subjected to heterogeneous
refinement with two classes and default parameters. The final non-
uniform refinement was performed on approximately 280,000 se-
lected particles using dynamic masking and refinements of CTF
and per-­particle defocus. The map at 3.3-­Å resolution, filtered
based on local resolution estimates in cryoSPARC, was used for
model refinements. A monomer of E. coli E2 (PDB: 4N72) was
rigid-­body fitted to the filtered map in UCSF Chimera (76) and
real-­space refined in PHENIX with standard parameters using
protein secondary structure and side-­chain rotamer restraints. Af-
ter manual real-­space refinement in Coot to reduce clashscore and
rotamer outliers, an octahedral symmetry was applied to generate
the 24-­meric complex, which then underwent a final refinement
using noncrystallographic symmetry, protein secondary structure,
and side-­chain rotamer restraints. The model statistics were gener-
ated with MolProbity (77) using the comprehensive validation tool
in PHENIX.
The mini-­PDHcΔLD dataset was processed in cryoSPARC v3.2
and Relion 3.1. First, extracted particles (200 pixel box size; 1.98 Å/pixel)
were subjected to two rounds of 2D classification in cryoSPARC. Upon
selection of 2D class averages (~570,000 particles), particles were
imported into Relion where they underwent one more 2D classifica-
tion, resulting in a pool of approximately 330,000 particles. A single
initial model was created in Relion and used in 3D autorefinement
with default parameters. The roughly aligned particles were subjected
to 3D classification (one class, global angular searches, regulariza-
tion parameter T = 25) using the previous map from 3D refinement
as reference and a mask created in Relion, which covered a single E1
dimer and PSBD (E1-­PSBD mask). Upon completion, handedness
of the output map was flipped, the mask covering a single E1 dimer
plus PSBD was adjusted to the new map, and the 3D classification
with analogous parameters was performed. This procedure of mask
adjustment and 3D classification was repeated two more times for
best angular assignment. Next, the particles underwent 3D autore-
finement (E1-­PSBD mask; global angular searches), which recon-
structed the E1 density at ~4-­Å resolution. The particles were then
re-­extracted with 250 pixel box size and binned to 1.58 Å/pixel be-
fore being subjected to another 3D autorefinement with local angu-
lar searches and the E1-­PSBD mask. The particles underwent signal
subtraction using the mask on the aligned E1 dimer and PSBD, pre-
serving their density and removing the noisy signal of misaligned
CD3 and the second dimer of E1. The signal-­subtracted particles
(180 pixel box size, 1.58 Å/pixel) and their corresponding 3D recon-
struction were then imported back into cryoSPARC for further pro-
cessing. The nonuniform refinement, with the imported 3D volume
used as a reference and the E1-­PSBD mask created in cryoSPARC,
was performed. A mask covering only the PSBD density was then
generated from the volume-­erased (UCSF Chimera) map of the pre-
vious refinement and used in the 3D variability analysis set to cluster
mode and filtering resolution at 5 Å. A single cluster of about 61,000
particles was subjected to a nonuniform refinement followed by a
local refinement (10° and 5-­Å local searches; E1-­PSBD mask). Last,
filtering of the map based on local resolution estimates in
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
cryoSPARC was conducted. Structures of the dimeric E. coli E1
(PDB: 4QOY; residues 56 to 886) and the AlphaFold prediction (42,
43) of two PSBD chains (E2, residues 314 to 379) in complex with
four N-­terminal helices of E1 (residues 1 to 58) were rigid-­body fit-
ted into the cryo-­EM map using UCSF Chimera. Statistics are sum-
marized in table S3, and the atomic coordinates and EM maps are
deposited in the PDB and the Electron Microscopy Data Bank, re-
spectively, under the accession codes 8ORB and EMD-­17119 (CD24)
and EMD-­17126 (mini-­PDHc).
Measurement of enzymatic activity
For comparison of the enzymatic activities of wild-­type PDHc and
mini-­PDHc, complexes were assembled from 300 nM E224 or E23
(monomeric concentration) and saturating concentrations of E1 and
E3, according to our stoichiometry data. After 20-­min incubation at
room temperature, the apparent PDHc activity was measured spec-
troscopically according to the method described by Nemeria et al.
(78). Specifically, the reaction was started by 25-­fold rapid dilution
of the complex with reaction buffer [20 mM Hepes-­NaOH (pH 7.4),
140 mM KCl, 10 mM NaCl, 1 mM MgCl2, 5 mM pyruvate, 2.5 mM
NAD+, 0.2 mM TPP, 0.2 mM CoA, 2.6 mM DTT] using a plate reader
equipped with a dispenser module (Synergy 2, Agilent BioTek). Af-
ter 5-­s mixing time, the conversion of NAD+ to NADH was followed
for 40 s via the absorbance increase at 340 nm and 25°C [molar ex-
tinction coefficient of NADH at 340 nm: 6220 M−1 cm−1; (79)]. Ini-
tial slopes of the recorded curves were converted into initial velocity
of NADH production per E2 monomer (TON/[E2 monomer]) (see
fig. S3E) to normalize the activities of wild-­type PDHc and mini-­
PDHc. The final concentrations of PDHc subunits in all assays were:
E2 monomers (wild-­type or mini-­PDHc): 12 nM; E1 monomers:
16 nM; and E3 monomers: 8 nM. The enzymatic activity of wild-­type
PDHc or mini-­PDHc was measured for three independent complex
reconstitutions with two independent preparations each for E23 and
E224. Enzymatic parameters of PDHc were determined by varying
the concentration of one of the substrates, while the concentrations
of the other substrates were kept constant (5 mM for pyruvate, 0.2 mM
for CoA, and 2.5 mM for NAD+) (table S4). The data were analyzed
using Eq. 1
Pyr
vi = [E0 ] ⋅ kcat ⋅ ([Pyr]∕(KM + [Pyr]) ⋅
CoA
([CoA]∕(KM + [CoA]) ⋅ ([NAD+ ]∕(KM
NAD
+ [NAD+ ]) (1)
where vi is the recorded initial velocity; [E0] is the concentration of
fully assembled PDHc; [Pyr], [CoA], and [NAD+] are the concen-
trations of pyruvate, CoA, and NAD+, respectively; and KM , KM
Pyr
CoA
,
and KM are the respective KM values.
NAD
Determination of PDHc stoichiometry
The stoichiometry of complexes formed between E2 and the periph-
eral subunits was determined by addition of 0 to 12.5 μM E1 or E3
dimers to E2 at 5 μM monomeric concentration. Next, the stoichi-
ometry of the complex comprising all protein components (E1, E2,
and E3) was determined by incubating E2 variants (10 μM) for
5 min at room temperature with stoichiometric ratios of E1 or E3
found above and subsequent addition of the other peripheral subunit
(0 to 12.5 μM dimer) to the twofold diluted assembled subcomplexes.
To ensure full complex formation, the samples were incubated for
another 10 min at 25°C before carrying out analytical SEC as described
14 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
above. The peak areas of the assembled complex and free peripheral
subunits were determined using OriginPro 2018 or PeakFit (Systat
Software). The sum of the peak areas (total peak area) in each indi-
vidual SEC profile was plotted against the E1 dimer/E2 monomer or
E3 dimer/E2 monomer ratio and used as internal calibration curve
to determine the constant peak area of E2 (y-­axis intercept of the
calibration curve; the weak contribution of E2 to the total peak area
results from its low extinction coefficient at 280 nm). After subtrac-
tion of the contribution of E2 in the high–molecular mass peak, the
peak area of the peripheral subunit in complex with E2 was plotted
against the respective E1 dimer (E3 dimer)/E2 monomer ratio. The
intercept between the linear regression curves of the pre-­and post-­
saturation data points yielded the equivalence point. Using the cali-
bration curve for complexed E1, the peak area of bound E1 at the
equivalence point was converted to the [E1 dimer bound]/[E2 monomer]
ratio at the equivalence point. The workflow is illustrated in fig. S12.
The provided standard errors (SEs) were calculated using the SEs of
the linear regression curves.
Analysis of sequence conservation
The Basic Local Alignment Search Tool (BLAST; National Center
for Biotechnology Information) was utilized to search for 5000 or-
thologous gammaproteobacterial sequences using residues 327 to
376 of E. coli E2 as template. Sequences that were either incomplete
or not part of PDHc E2, despite high homology, were removed. The
resultant set of 2175 gammaproteobacterial PSBD sequences were
aligned with Clustal Omega (EMBL) and then converted into FASTA
format using the EMBOSS Seqret tool (EMBL). This sequence align-
ment together with the crystal structure of dimeric PSBD were up-
loaded to the ConSurf Server (80, 81) to visualize sequence conservation
in the context of a 3D model. The same alignment was also used for
the WebLogo server (82) to generate a sequence logo plot.
Supplementary Materials
This PDF file includes:
Figs. S1 to S12
Tables S1 to S4
Legend for movie S1
Other Supplementary Material for this manuscript includes the following:
Movie S1
REFERENCES AND NOTES
1. L. J. Reed, Multienzyme complexes. Acc. Chem. Res. 7, 40–46 (1974).
2. M. S. Patel, T. E. Roche, Molecular biology and biochemistry of pyruvate dehydrogenase
complexes. FASEB J. 4, 3224–3233 (1990).
3. M. S. Patel, T. E. Roche, R. A. Harris, Alpha-­keto acid dehydrogenase complexes (Molecular
and cell biology updates, Birkhäuser, Birkhäuser, 1996).
4. L. J. Reed, A trail of research from lipoic acid to α-­keto acid dehydrogenase complexes.
J. Biol. Chem. 276, 38329–38336 (2001).
5. M. Koike, L. J. Reed, W. R. Carroll, α-­Keto acid dehydrogenation Complexes. J. Biol. Chem.
238, 30–39 (1963).
6. L. J. Reed, R. M. Oliver, Structure-­Function Relationships in Pyruvate and α-­Ketoglutarate
Dehydrogenase Complexes in Structure and Function Relationships in Biochemical
Systems, F. Bossa, E. Chiancone, A. Finazzi-­Agrò, R. Strom, Eds. (Springer US, Boston, MA,
1982), pp. 231–241.
7. R. N. Perham, L. C. Packman, 2-­Oxo acid dehydrogenase multienzyme complexes:
Domains, dynamics, and Design. Ann. N. Y. Acad. Sci. 573, 1–20 (1989).
8. J. R. Guest, S. J. Angier, G. C. Russell, Structure, expression, and protein engineering of the
pyruvate dehydrogenase complex of Escherichia coli. Ann. N. Y. Acad. Sci. 573, 76–99
(1989).
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
9. A. de Kok, A. F. Hengeveld, A. Martin, A. H. Westphal, The pyruvate dehydrogenase
multi-­enzyme complex from Gram-­negative bacteria. Biochim. Biophys. Acta. 1385,
353–366 (1998).
10. D. L. Bates, M. J. Danson, G. Hale, E. A. Hooper, R. N. Perham, Self-­assembly and catalytic
activity of the pyruvate dehydrogenase multienzyme complex of Escherichia coli. Nature
268, 313–316 (1977).
11. M. S. Patel, N. S. Nemeria, W. Furey, F. Jordan, The pyruvate dehydrogenase complexes:
Structure-­based function and regulation. J. Biol. Chem. 289, 16615–16623 (2014).
12. L. J. Reed, M. L. Hackert, Structure-­function relationships in dihydrolipoamide
acyltransferases. J. Biol. Chem. 265, 8971–8974 (1990).
13. A. Mattevi, G. Obmolova, E. Schulze, K. H. Kalk, A. H. Westphal, A. de Kok, W. G. J. Hol,
Atomic structure of the cubic core of the pyruvate dehydrogenase multienzyme
complex. Science 255, 1544–1550 (1992).
14. J. Škerlová, J. Berndtsson, H. Nolte, M. Ott, P. Stenmark, Structure of the native pyruvate
dehydrogenase complex reveals the mechanism of substrate insertion. Nat. Commun.
12, 5277 (2021).
15. A. Mattevi, G. Obmolova, K. H. Kalk, A. Teplyakov, W. G. J. Hol, Crystallographic analysis of
substrate binding and catalysis in dihydrolipoyl transacetylase (E2p). Biochemistry 32,
3887–3901 (1993).
16. B. Nagy, M. Polak, O. Ozohanics, Z. Zambo, E. Szabo, A. Hubert, F. Jordan, J. Novaček,
V. Adam-­Vizi, A. Ambrus, Structure of the dihydrolipoamide succinyltransferase (E2)
component of the human alpha-­ketoglutarate dehydrogenase complex (hKGDHc)
revealed by cryo-­EM and cross-­linking mass spectrometry: Implications for the overall
hKGDHc structure. Biochim. Biophys. Acta Gen. Subj. 1865, 129889 (2021).
17. S. Liu, X. Xia, J. Zhen, Z. Li, Z. H. Zhou, Structures and comparison of endogenous
2-­oxoglutarate and pyruvate dehydrogenase complexes from bovine kidney. Cell Discov.
8, 126 (2022).
18. F. L. Kyrilis, D. A. Semchonok, I. Skalidis, C. Tüting, F. Hamdi, F. J. O’Reilly, J. Rappsilber,
P. L. Kastritis, Integrative structure of a 10-­megadalton eukaryotic pyruvate
dehydrogenase complex from native cell extracts. Cell Rep. 34, 108727 (2021).
19. B. O. Forsberg, S. Aibara, R. J. Howard, N. Mortezaei, E. Lindahl, Arrangement and
symmetry of the fungal E3BP-­containing core of the pyruvate dehydrogenase complex.
Nat. Commun. 11, 4667 (2020).
20. S. S. Mande, S. Sarfaty, M. D. Allen, R. N. Perham, W. G. J. Hol, Protein–protein interactions
in the pyruvate dehydrogenase multienzyme complex: Dihydrolipoamide
dehydrogenase complexed with the binding domain of dihydrolipoamide
acetyltransferase. Structure 4, 277–286 (1996).
21. P. Arjunan, J. Wang, N. S. Nemeria, S. Reynolds, I. Brown, K. Chandrasekhar, G. Calero,
F. Jordan, W. Furey, Novel binding motif and new flexibility revealed by structural
analyses of a pyruvate dehydrogenase-­dihydrolipoyl acetyltransferase subcomplex from
the Escherichia coli pyruvate dehydrogenase multienzyme complex. J. Biol. Chem. 289,
30161–30176 (2014).
22. R. A. W. Frank, J. V. Pratap, X. Y. Pei, R. N. Perham, B. F. Luisi, The molecular origins of
specificity in the assembly of a multienzyme complex. Structure 13, 1119–1130 (2005).
23. K. Chandrasekhar, J. Wang, P. Arjunan, M. Sax, Y. H. Park, N. S. Nemeria, S. Kumaran,
J. Song, F. Jordan, W. Furey, Insight to the interaction of the dihydrolipoamide
acetyltransferase (E2) core with the peripheral components in the Escherichia coli
pyruvate dehydrogenase complex via multifaceted structural approaches. J. Biol. Chem.
288, 15402–15417 (2013).
24. G. J. Domingo, H. J. Chauhan, I. A. D. Lessard, C. Fuller, R. N. Perham, Self-­assembly and
catalytic activity of the pyruvate dehydrogenase multienzyme complex from Bacillus
stearothermophilus. Eur. J. Biochem. 266, 1136–1146 (1999).
25. L. J. Reed, F. H. Pettit, M. H. Eley, L. Hamilton, J. H. Collins, R. M. Oliver, Reconstitution of
the Escherichia coli pyruvate dehydrogenase complex. Proc. Natl. Acad. Sci. U.S.A. 72,
3068–3072 (1975).
26. K. J. Angelides, S. K. Akiyama, G. G. Hammes, Subunit stoichiometry and molecular
weight of the pyruvate dehydrogenase multienzyme complex from Escherichia coli. Proc.
Natl. Acad. Sci. U.S.A. 76, 3279–3283 (1979).
27. C. A. CaJacob, P. A. Frey, J. F. Hainfeld, J. S. Wall, H. Yang, Escherichia coli pyruvate
dehydrogenase complex: Particle masses of the complex and component enzymes
measured by scanning transmission electron microscopy. Biochemistry 24, 2425–2431
(1985).
28. M. H. Eley, G. Namihira, L. Hamilton, P. Munk, L. J. Reed, α-­Keto acid dehydrogenase
complexes: XVIII. Subunit composition of the Escherichia coli pyruvate dehydrogenase
complex. Arch. Biochem. Biophys. 152, 655–669 (1972).
29. D. L. Bates, R. A. Harrison, R. N. Perham, The stoichiometry of polypeptide chains in the
pyruvate dehydrogenase multienzyme complex of E. coli determined by a simple novel
method. FEBS Lett. 60, 427–430 (1975).
30. R. N. Perham, E. A. Hooper, Polypeptide chain stoicheiometry in the self-­assembly of the
pyruvate dehydrogenase multienzyme complex of Escherichia coli. FEBS Lett. 73,
137–140 (1977).
15 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
31. N. L. Marrott, J. J. Marshall, D. I. Svergun, S. J. Crennell, D. W. Hough, J. M. van den Elsen,
M. J. Danson, Why are the 2-­oxoacid dehydrogenase complexes so large? Generation of
an active trimeric complex. Biochem. J. 463, 405–412 (2014).
32. E. M. Bruch, P. Vilela, L. Yang, A. Boyko, N. Lexa-­Sapart, B. Raynal, P. M. Alzari, M. Bellinzoni,
Actinobacteria challenge the paradigm: A unique protein architecture for a well-­known,
central metabolic complex. Proc. Natl. Acad. Sci. U.S.A. 118, e2112107118 (2021).
33. P. Arjunan, N. Nemeria, A. Brunskill, K. Chandrasekhar, M. Sax, Y. Yan, F. Jordan, J. R. Guest,
W. Furey, Structure of the pyruvate dehydrogenase multienzyme complex E1 component
from Escherichia coli at 1.85 Å resolution. Biochemistry 41, 5213–5221 (2002).
34. M. A. Robien, G. M. Clore, J. G. Omichinski, R. N. Perham, E. Appella, K. Sakaguchi,
A. M. Gronenborn, Three-­dimensional solution structure of the E3-­binding domain of the
dihydrolipoamide succinyltransferase core from the 2-­oxoglutarate dehydrogenase
multienzyme complex of Escherichia coli. Biochemistry 31, 3463–3471 (1992).
35. J. D. F. Green, E. D. Laue, R. N. Perham, S. T. Ali, J. R. Guest, Three-­dimensional structure of a
lipoyl domain from the dihydrolipoyl acetyltransferase component of the pyruvate
dehydrogenase multienzyme complex of Escherichia coli. J. Mol. Biol. 248, 328–343 (1995).
36. T. Wagenknecht, R. Grassucci, J. Berkowitz, C. Forneris, Configuration of interdomain
linkers in pyruvate dehydrogenase complex of Escherichia coli as determined by
cryoelectron microscopy. J. Struct. Biol. 109, 70–77 (1992).
37. G. E. Murphy, G. J. Jensen, Electron cryotomography of the E. coli pyruvate and
2-­oxoglutarate dehydrogenase complexes. Structure 13, 1765–1773 (2005).
38. J. Wang, N. S. Nemeria, K. Chandrasekhar, S. Kumaran, P. Arjunan, S. Reynolds, G. Calero,
R. Brukh, L. Kakalis, W. Furey, F. Jordan, Structure and function of the catalytic domain of
the dihydrolipoyl acetyltransferase component in Escherichia coli pyruvate
dehydrogenase complex. J. Biol. Chem. 289, 15215–15230 (2014).
39. R. A. W. Frank, C. M. Titman, J. V. Pratap, B. F. Luisi, R. N. Perham, A molecular switch and
proton wire synchronize the active sites in thiamine enzymes. Science 306, 872–876 (2004).
40. E. Krissinel, K. Henrick, Inference of macromolecular assemblies from crystalline state.
J. Mol. Biol. 372, 774–797 (2007).
41. A. Vangone, R. Spinelli, V. Scarano, L. Cavallo, R. Oliva, COCOMAPS: A web application to
analyze and visualize contacts at the interface of biomolecular complexes. Bioinformatics
27, 2915–2916 (2011).
42. J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool,
R. Bates, A. Žídek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie,
B. Romera-­Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy,
M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver,
O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, D. Hassabis, Highly accurate protein
structure prediction with AlphaFold. Nature 596, 583–589 (2021).
43. K. Tunyasuvunakool, J. Adler, Z. Wu, T. Green, M. Zielinski, A. Žídek, A. Bridgland, A. Cowie,
C. Meyer, A. Laydon, S. Velankar, G. J. Kleywegt, A. Bateman, R. Evans, A. Pritzel,
M. Figurnov, O. Ronneberger, R. Bates, S. A. A. Kohl, A. Potapenko, A. J. Ballard,
B. Romera-­Paredes, S. Nikolov, R. Jain, E. Clancy, D. Reiman, S. Petersen, A. W. Senior,
K. Kavukcuoglu, E. Birney, P. Kohli, J. Jumper, D. Hassabis, Highly accurate protein
structure prediction for the human proteome. Nature 596, 590–596 (2021).
44. Y. N. Kalia, S. M. Brocklehurst, D. S. Hipps, E. Appella, K. Sakaguchi, R. N. Perham, The
high-­resolution structure of the peripheral subunit-­binding domain of dihydrolipoamide
acetyltransferase from the pyruvate dehydrogenase multienzyme complex of Bacillus
stearothermophilus. J. Mol. Biol. 230, 323–341 (1993).
45. M. D. Allen, R. W. Broadhurst, R. G. Solomon, R. N. Perham, Interaction of the E2 and E3
components of the pyruvate dehydrogenase multienzyme complex of Bacillus
stearothermophilus. FEBS J. 272, 259–268 (2005).
46. N. Ferguson, T. D. Sharpe, P. J. Schartau, S. Sato, M. D. Allen, C. M. Johnson, T. J. Rutherford,
A. R. Fersht, Ultra-­fast barrier-­limited folding in the peripheral subunit-­binding domain
family. J. Mol. Biol. 353, 427–446 (2005).
47. J. S. Miles, J. R. Guest, S. E. Radford, R. N. Perham, Investigation of the mechanism of active
site coupling in the pyruvate dehydrogenase multienzyme complex of Escherichia coli by
protein engineering. J. Mol. Biol. 202, 97–106 (1988).
48. T. Peng, H. Lee, S. Lim, Isolating a trimer intermediate in the self-­assembly of E2 protein
cage. Biomacromolecules 13, 699–705 (2012).
49. J. D. Green, R. N. Perham, S. J. Ullrich, E. Appella, Conformational studies of the interdomain linker
peptides in the dihydrolipoyl acetyltransferase component of the pyruvate dehydrogenase
multienzyme complex of Escherichia coli. J. Biol. Chem. 267, 23484–23488 (1992).
50. J. Song, F. Jordan, Interchain acetyl transfer in the E2 component of bacterial pyruvate
dehydrogenase suggests a model with different roles for each chain in a trimer of the
homooligomeric component. Biochemistry 51, 2795–2803 (2012).
51. J. H. Collins, L. J. Reed, Acyl group and electron pair relay system: A network of interacting
lipoyl moieties in the pyruvate and alpha-­ketoglutarate dehydrogenase complexes from
Escherichia coli. Proc. Natl. Acad. Sci. U.S.A. 74, 4223–4227 (1977).
52. G. Hale, D. L. Bates, R. N. Perham, Subunit exchange in the pyruvate dehydrogenase
complex of Escherichia coli. FEBS Lett. 104, 343–346 (1979).
53. S. C. Gill, P. H. von Hippel, Calculation of protein extinction coefficients from amino acid
sequence data. Anal. Biochem. 182, 319–326 (1989).
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
54. E. Gasteiger, C. Hoogland, A. Gattiker, S. Duvaud, M. R. Wilkins, R. D. Appel, A. Bairoch,
Protein Identification and Analysis Tools on the Expasy Server, in The Proteomics Protocols
Handbook, J. M. Walker, Ed. (Humana Press, 2005), pp. 571–607.
55. P. Schuck, Size-­distribution analysis of macromolecules by sedimentation velocity
ultracentrifugation and Lamm equation modeling. Biophys. J. 78, 1606–1619 (2000).
56. C. A. Brautigam, Chapter Five -­Calculations and publication-­quality illustrations for
analytical ultracentrifugation data, in Methods Enzymol., J. L. Cole, Ed. (Academic Press,
2015), vol. 562, pp. 109–133.
57. P. Schuck, H. Zhao, C. A. Brautigam, R. Ghirlando, Basic Principles of Analytical
Ultracentrifugation (CRC Press, Boca Raton, FL, 2016), pp. 1–294.
58. P. Schuck, On the analysis of protein self-­association by sedimentation velocity analytical
ultracentrifugation. Anal. Biochem. 320, 104–124 (2003).
59. T. M. Laue, B. Shah, T. M. Ridgeway, S. L. Pelletier, Computer-­Aided Interpretation of
Sedimentation Data for Proteins, in Analytical Ultracentrifugation in Biochemistry and
Polymer Science, S. E. Harding, J. C. Horton, A. J. Rowe, Eds. (Royal Society of Chemistry,
Cambridge, U.K., 1992), pp. 90–125.
60. C. N. Pace, E. J. Hebert, K. L. Shaw, D. Schell, V. Both, D. Krajcikova, J. Sevcik, K. S. Wilson,
Z. Dauter, R. W. Hartley, G. R. Grimsley, Conformational stability and thermodynamics of
folding of ribonucleases Sa, Sa2 and Sa3. J. Mol. Biol. 279, 271–286 (1998).
61. W. Kabsch, XDS. Acta Crystallogr. D Biol. Crystallogr. 66, 125–132 (2010).
62. A. J. McCoy, R. W. Grosse-­Kunstleve, P. D. Adams, M. D. Winn, L. C. Storoni, R. J. Read,
Phaser crystallographic software. J. Appl. Crystallogr. 40, 658–674 (2007).
63. P. D. Adams, P. V. Afonine, G. Bunkóczi, V. B. Chen, I. W. Davis, N. Echols, J. J. Headd,
L.-­W. Hung, G. J. Kapral, R. W. Grosse-­Kunstleve, A. J. McCoy, N. W. Moriarty, R. Oeffner,
R. J. Read, D. C. Richardson, J. S. Richardson, T. C. Terwilliger, P. H. Zwart, PHENIX: A
comprehensive Python-­based system for macromolecular structure solution. Acta
Crystallogr. D Biol. Crystallogr. 66, 213–221 (2010).
64. P. Emsley, B. Lohkamp, W. G. Scott, K. Cowtan, Features and development of Coot. Acta
Crystallogr. D Biol. Crystallogr. 66, 486–501 (2010).
65. M. Ohi, Y. Li, Y. Cheng, T. Walz, Negative staining and image classification — powerful
tools in modern electron microscopy. Biol. Proced. Online 6, 23–34 (2004).
66. A. Punjani, J. L. Rubinstein, D. J. Fleet, M. A. Brubaker, cryoSPARC: Algorithms for
rapid unsupervised cryo-­EM structure determination. Nat. Methods 14, 290–296 (2017).
67. D. Asarnow, E. Palovcak, Y. Cheng, asarnow/pyem: UCSF pyem v0.5. (Zenodo, 2019);
10.5281/zenodo.3576630.
68. R. Fernandez-­Leiro, S. H. W. Scheres, A pipeline approach to single-­particle processing in
RELION. Acta Crystallogr. D Biol. Crystallogr. 73, 496–502 (2017).
69. J. Zivanov, T. Nakane, B. O. Forsberg, D. Kimanius, W. J. Hagen, E. Lindahl, S. H. W. Scheres,
New tools for automated high-­resolution cryo-­EM structure determination in RELION-­3.
eLife 7, e42166 (2018).
70. T. D. Goddard, C. C. Huang, E. C. Meng, E. F. Pettersen, G. S. Couch, J. H. Morris,
T. E. Ferrin, UCSF ChimeraX: Meeting modern challenges in visualization and analysis.
Protein Sci. 27, 14–25 (2018).
71. A. Scaiola, F. Mangia, S. Imseng, D. Boehringer, K. Berneiser, M. Shimobayashi, E. Stuttfeld,
M. N. Hall, N. Ban, T. Maier, The 3.2-­Å resolution structure of human mTORC2. Sci. Adv. 6,
eabc1251 (2020).
72. N. de Martín Garrido, W. Fu, K. Ramlaul, Z. Zhu, D. Miller, D. Boehringer, C. H. S. Aylett,
Direct transfer of electron microscopy samples to wetted carbon and graphene films via
a support floatation block. J. Struct. Biol. 213, 107677 (2021).
73. S. Q. Zheng, E. Palovcak, J.-­P. Armache, K. A. Verba, Y. Cheng, D. A. Agard, MotionCor2:
Anisotropic correction of beam-­induced motion for improved cryo-­electron microscopy.
Nat. Methods 14, 331–332 (2017).
74. T. Bepler, A. Morin, M. Rapp, J. Brasch, L. Shapiro, A. J. Noble, B. Berger, Positive-­unlabeled
convolutional neural networks for particle picking in cryo-­electron micrographs. Nat.
Methods 16, 1153–1160 (2019).
75. T. Bepler, K. Kelley, A. J. Noble, B. Berger, Topaz-­Denoise: General deep denoising models
for cryoEM and cryoET. Nat. Commun. 11, 5208 (2020).
76. E. F. Pettersen, T. D. Goddard, C. C. Huang, G. S. Couch, D. M. Greenblatt, E. C. Meng,
T. E. Ferrin, UCSF Chimera—A visualization system for exploratory research and analysis.
J. Comput. Chem. 25, 1605–1612 (2004).
77. V. B. Chen, W. B. Arendall III, J. J. Headd, D. A. Keedy, R. M. Immormino, G. J. Kapral,
L. W. Murray, J. S. Richardson, D. C. Richardson, MolProbity: all-­atom structure
validation for macromolecular crystallography. Acta Crystallogr. D Biol. Crystallogr. 66,
12–21 (2010).
78. N. Nemeria, A. Volkov, A. Brown, J. Yi, L. Zipper, J. R. Guest, F. Jordan, Systematic study of
the six cysteines of the E1 subunit of the pyruvate dehydrogenase multienzyme complex
from Escherichia coli: None is essential for activity. Biochemistry 37, 911–922 (1998).
79. B. L. Horecker, A. Kornberg, The extinction coefficients of the reduced band of pyridine
nucleotides. J. Biol. Chem. 175, 385–390 (1948).
80. M. Landau, I. Mayrose, Y. Rosenberg, F. Glaser, E. Martz, T. Pupko, N. Ben-­Tal, ConSurf 2005: The
projection of evolutionary conservation scores of residues on protein structures. Nucleic Acids
Res. 33, W299–W302 (2005).
16 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
81. H. Ashkenazy, S. Abadi, E. Martz, O. Chay, I. Mayrose, T. Pupko, N. Ben-­Tal, ConSurf 2016:
An improved methodology to estimate and visualize evolutionary conservation in
macromolecules. Nucleic Acids Res. 44, W344–W350 (2016).
82. G. E. Crooks, G. Hon, J.-­M. Chandonia, S. E. Brenner, WebLogo: A sequence logo generator.
Genome Res. 14, 1188–1190 (2004).
83. D. D. Jones, K. M. Stott, M. J. Howard, R. N. Perham, Restricted motion of the lipoyl-­lysine
swinging arm in the pyruvate dehydrogenase complex of Escherichia coli. Biochemistry
39, 8448–8459 (2000).
Acknowledgments: We thank P. Afanasyev for support with EM data collection and analysis,
M. Leibundgut for support with crystal structure determination, and J. Stanisich for revision of
this manuscript. We thank S. Chesnov (Functional Genomics Center Zurich) for recording mass
spectra. We also thank the Protein Crystallization Center at the University of Zurich for initial
crystallization screens. The x-­ray data were acquired at the Swiss Light Source (SLS) synchrotron
at the Paul Scherrer Institute (PSI). Cryo-­EM data were collected at the Scientific Center for
Optical and Electron Microscopy at the ETH Zurich (ScopeM). We thank the ScopeM and SLS
staff for technical support during data collection and P. Bachmann for IT support. Funding: This
work was supported by the Swiss National Science Foundation (SNSF) grant 310030_201234 to R.G.
Meinhold et al., Sci. Adv. 10, eadj6358 (2024) 7 February 2024
Author contributions: S.M., R.Z., C.G., and R.G. designed all experiments. S.M. and R.Z. cloned
expression vectors, set up crystallization plates. S.M. purified the proteins, performed analytical
SEC, circular dichroism, and enzymatic activity experiments. S.M. collected x-­ray diffraction
data, solved, and refined x-­ray structures. R.Z. collected, processed, and analyzed NS-­EM and
cryo-­EM data and performed PSBD sequence conservation analysis. C.G. conducted AUC
experiments and analyzed the data. S.M. and R.Z. prepared the figures. All authors contributed
to the writing of the manuscript. Competing interests: The authors declare that they have no
competing interests. Data and materials availability: All data needed to evaluate the
conclusions in the paper are present in the paper and/or the Supplementary Materials. The
structural data reported in this study are available in the Electron Microscopy Data Bank and
PDB under accession codes EMD-­17119 (CD24) and EMD-­17126 (E1:PSBD within mini-­PDHcΔLD)
and PDB ID 8ORB (CD24), 8OSY (CD3), and 8OQJ (PSBD).
Submitted 7 July 2023
Accepted 11 January 2024
Published 7 February 2024
10.1126/sciadv.adj6358
17 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

PLANT SCIENCES Copyright © 2024 The

Authors, some rights
Haplotype-­phased genome unveils the butylphthalide reserved; exclusive
licensee American
biosynthesis and homoploid hybrid origin of Association for the
Advancement of
Ligusticum chuanxiong Science. No claim to
original U.S.
Government Works.
Bao Nie1†, Xueqing Chen1†, Zhuangwei Hou1†, Miaoxian Guo1†, Cheng Li1, Wenkai Sun1, Distributed under a
Jiaojiao Ji1, Lanlan Zang1, Song Yang1, Pengxiang Fan2, Wenhao Zhang3,4, Hang Li5, Yuzhu Tan6, Creative Commons
Wei Li1, Li Wang1,7* Attribution
NonCommercial
Butylphthalide is one of the first-­line drugs for ischemic stroke therapy, while no biosynthetic enzyme for bu- License 4.0 (CC BY-­NC).
tylphthalide has been reported. Here, we present a haplotype-­resolved genome of Ligusticum chuanxiong, a long-­
cultivated and phthalide-­rich medicinal plant in Apiaceae. On the basis of comprehensive screening, four Fe(II)-­and
2-­
oxoglutarate–dependent dioxygenases and two CYPs were mined and further biochemically verified as
phthalide C-­4/C-­5 desaturases (P4,5Ds) that effectively promoted the forming of (S)-­3-­​n-­butylphthalide and bu-
tylidenephthalide. The substrate promiscuity and functional redundancy featured for P4,5Ds may contribute to
the high phthalide diversity in L. chuanxiong. Notably, comparative genomic evidence supported L. chuanxiong as
a homoploid hybrid with Ligusticum sinense as a potential parent. The two haplotypes demonstrated exceptional
structure variance and diverged around 3.42 million years ago. Our study is an icebreaker for the dissection of
phthalide biosynthetic pathway and reveals the hybrid origin of L. chuanxiong, which will facilitate the metabolic
engineering for (S)-­3-­​n-­butylphthalide production and breeding for L. chuanxiong.

INTRODUCTION TWIK-­related K+ channel 1 currents (12), preventing Alzheimer’s

dl-­3-­​n-­Butylphthalide (dl-­NBP) is one of the first-­line drugs for
ischemic stroke therapy, which is the commonest type of stroke, at-
tacking more than 7.63 million people and causing 3.29 million
deaths in 2019 (1). Distinct from the side effects and limitations of
recombinant tissue plasminogen activator (2, 3), dl-­NBP is promi-
nent for its good neuroprotection and safety in various animal mod-
els and humans (4–8) and was approved for marketing in China in
2004 (8). Moreover, the phase 2 clinical trial of NBP soft capsules in
patients with ischemic stroke was started in the United States in
2017 (9). To date, the annual sales of dl-­NBP have reached $1 billion
(https://doc.irasia.com/listco/hk/cspc/annual/2021/car2021.pdf).
dl-­NBP is a chemically synthetic racemic phthalide [an equal
mixture of d-­NBP and l-­NBP; the absolute configurations of d-­NBP
and l-­NBP are (R)-­NBP and (S)-­NBP, which are adopted hereafter].
Compared with (RS)-­NBP, the naturally distributed (S)-­NBP dem-
onstrated enhanced drug efficiency and reduced side effects. The ef-
fects of (S)-­NBP on intracellular cyclic guanosine monophosphate
and extracellular nitric oxide levels were antagonized by (R)-­NBP
(10), possibly resulting in a decrease in the drug efficiency of (RS)-­
NBP. Moreover, (S)-­NBP is more potent than (RS)-­NBP and (R)-­
NBP for antiplatelet (5), inhibiting apoptosis (11), hampering
1
Shenzhen Branch, Guangdong Laboratory of Lingnan Modern Agriculture, Key
Laboratory of Synthetic Biology, Ministry of Agriculture and Rural Affairs, Agricul-
tural Genomics Institute at Shenzhen, Chinese Academy of Agricultural Sciences,
Shenzhen 518120, China. 2College of Agriculture and Biotechnology, Zhejiang Uni-
versity, Hangzhou 310063, China. 3State Key Laboratory of Vegetation and Environ-
mental Change, Institute of Botany, Chinese Academy of Sciences, Beijing 100093,
China. 4College of Resources and Environment, University of Chinese Academy of
Sciences, Beijing 100049, China. 5School of Pharmaceutical Sciences, Sun Yat-­sen
University, Guangzhou 510006, China. 6State Key Laboratory of Southwestern
Chinese Medicine Resources, Pharmacy College, Chengdu University of Traditional
Chinese Medicine, Chengdu 611137, China. 7Key Laboratory for Quality Ensurance
and Sustainable Use of Dao-­di Herbs, National Resource Center for Chinese Materia
Medica, China Academy of Chinese Medical Sciences, Beijing 100700, China.
*Corresponding author. Email: wangli03@​caas.​cn
†These authors contributed equally to this work.
disease (13), indicating better neuroprotection potential. Thus, pure
(S)-­NBP is more promising for drug purposing than (RS)-­NBP (14).
However, (S)-­ NBP is less likely to be cost-­ efficiently produced
through chemical synthesis than (RS)-­NBP (15) and primarily relies
on extracting from Apiaceous plants (e.g., Ligusticum chuanxiong)
(16, 17). The low availability hampers the medicinal application of
(S)-­NBP. Metabolic engineering of biosynthetic pathways has been
proven to be effective and environment-­friendly in supplying plant
secondary metabolites (18), which may be a promising approach to
improve (S)-­NBP production at a relatively low cost.
However, little is known about phthalide biosynthesis. Previous
feeding experiments using 14C-­labeled acetic acid in Levisticum of-
ficinale revealed that six C─C units were incorporated into ligusti-
lide [4 in Fig. 1A, an analog of (S)-­NBP] without rearrangement,
which indicated a possible polyketide pathway for the phthalide
skeleton forming (fig. S1) (19, 20). However, other possible path-
ways for the phthalide scaffold formation could not be excluded
(21), such as cyclization of 12-­C fatty acid (fig. S1). Moreover, it
is very challenging to determine the intermediates in the polyketide
biosynthesis pathway without standard chemicals, although nu­
merous polyketide synthases have been identified (22). Thus, we
shift our attention to the downstream steps. (S)-­NBP belongs to
3-­substituted phthalides (type 1 in Fig. 1B), which is the most
broadly distributed type of phthalides in both plants and fungi
(Fig. 1B). It was primarily found in the Apiales species (Fig. 1B) (15,
17, 23, 24), particularly in L. chuanxiong. Apart from (S)-­NBP,
L. chuanxiong contains at least 94 other 3-­substituted phthalides
(fig. S2) (25), among which butylidenephthalide, senkyunolide A,
and ligustilide are the top three most abundant ones (2 to 4 in
Fig. 1A) (25, 26). They distinguish from (S)-­NBP only in C-­4/C-­5 or
C-­3/C-­8 olefinic bond (Fig. 1A), indicating that they are the likely
precursors or derivatives of (S)-­NBP. Although the content of (S)-­
NBP is usually low (lower than 0.001% of the dry weight), the con-
tent of its analogs is relatively abundant [the content of senkyunolide

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 1 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

A 11

9 10
8
4
3a 3
5
O2 O O O
6
7a 1
7
O O O O
(S)-NBP ( ) Butylidenephthalide ( ) Senkyunolide A ( ) Ligustilide ( )

B Type 1: Substituted Cn only on B (3- Type 4: Substituted Cn on neither A nor B

Type 2: Substituted Cn only on A (4/5/6/7- Type 5: Phthalide dimer / trimer
Type 3: Substituted Cn on both A and B
Lamiales
Boraginales
Solanales
Gentianales
4 Garryales
3a 3 Apiales
5 Dipsacales
A B O2 Asterales
Aquifoliales
6 7a 1 Ericales
7 Cornales
O Caryophyllales
Santalales
Rosales
Fagales
Fabales
Oxalidales
Brassicales
Eudicots Sapindales
Vitales
Buxales
Rauunculales
Poales
Zingiberales
Asparagales
Flowering Monocots
Liliales
plants Acorales
Piperales
Magnoliales
Amborellales
Polypodiales
Plant Ferns Cyatheales
Psilotales
Jungermanniales
Porellales
Liverworts Marchantiales
Treubiales
Pleosporales
Hysteriales
Botryosphaeriales
Venturiales
Eurotiales
Chaetothyriales
Ascomycota Hypocreales
Glomerellales
Micoascales
Diaporthales
Dikarya Ophiostomatales
Xylariales
Leotiomycetes
Polyporales
Basidiomycota Thelephorales
Russulales
Corticiales
Agaricales
Fungi
Blastocladiales
0 15 30 45 60
Species number

Fig. 1. Distribution and diversity of natural phthalides. (A) The four most abundant mono-­phthalides detected in L. chuanxiong. The differences among the four
phthalides are highlighted with blue (on the side chain) and red (within the ring) colors. (B) The phylogenetic distribution of natural phthalides, which were divided into
five types according to the substituted alkyl position and the polymerization level. Circles filled with colors indicate the presence of the corresponding natural phthalides,
and empty circles imply the absence. The horizontal bars denote the number of species in the corresponding order that have been reported to contain natural phthalides,
and the colors of the stacked bars designate the five types of natural phthalides. The distribution of phthalides was summarized from several reviews (15, 17, 23, 24). The
four target phthalides in this study all belong to type-­1 phthalides (3-­substituted phthalides).

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 2 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
A and ligustilide is usually dozens of times higher than that of (S)-­
NBP] (25, 26). Hence, the dissection of inter-­phthalide conversions
is crucial for enhancing the production of (S)-­NBP through pro-
moting the conversion process toward (S)-­NBP by overexpressing
validated enzymes in vivo. Thus, we focus on investigating the con-
version among (S)-­NBP and its three analogs, as it seems a feasible
starting point for unraveling the complete biosynthetic pathway.
L. chuanxiong is a medicinal plant in Apiaceae, the dried rhizome
of which is an important traditional Chinese medicine (TCM) used
to treat heart diseases. Despite L. chuanxiong has been cultivated for
thousands of years through asexual reproduction (vegetatively
propagated with expanded stem nodes) (Fig. 2A) (27), no wild pop-
ulation has been reported to date (28, 29). Thus, L. chuanxiong was
hypothesized to be a sterile cultivated variety (27–29). Several recent
phylogenetic studies revealed that L. chuanxiong belonged to Sino-
dielsia clade and several related species of L. chuanxiong were recog-
nized (30–32). Among these species, Ligusticum sinense was most
frequently proposed to be the wild progenitor of L. chuanxiong
based on limited evidence (27–30), which awaits further validation
based on genomic evidence. The high richness of phthalide and the
interesting biological background as a TCM make L. chuanxiong an
ideal system to investigate the (S)-­NBP biosynthesis pathway.
Here, we present the chromosome-­level and haplotype-­phased
genome of L. chuanxiong. By integrating multi-­omics data and bio-
chemical verification, we discovered that four Fe(II)-­and
2-­oxoglutarate–dependent dioxygenases (2OGDs) and two Cyto-
chrome P450s (CYPs) potentially promoted the conversions among
(S)-­NBP and its three analogs. In addition, we unexpectedly found
that L. chuanxiong was a homoploid hybrid based on comprehensive
genomic and evolutionary analyses. These findings shed light on the
knowledge gap of the phthalide biosynthetic pathway and the origin
of L. chuanxiong, which paves the way for the metabolic engineering
of (S)-­ NBP and provides insights into the future breeding of
L. chuanxiong varieties.
RESULTS
Genome assembly and annotation of L. chuanxiong
Fluorescence in situ hybridization (FISH) assay revealed 11 pairs of
homologous chromosomes in each metaphase root tip cell, which
confirmed that L. chuanxiong was a diploid species (28, 33), with
2n = 2X = 22 (Fig. 2B and fig. S3A). Further examinations detected
asymmetric telomere hybridization signals in at least three homolo-
gous chromosome pairs (fig. S3A), implying that structural varia-
tions (SVs) exist between haplotypes.
On the basis of 120-­Gb PacBio circular consensus sequencing
(CCS) long reads, the genome of L. chuanxiong was de novo as-
sembled to 7.01 Gb (tables S1 and S2), almost twice as the estimate
by flow cytometry (3.45 ± 0.38 Gb) (fig. S3B), echoing the high
genome heterozygosity (4.22%). Given that the phenomenon was
previously reported in other highly heterozygous genomes, e.g.,
Pogostemon cablin (3.69%) (34), we inferred that the assembled ge-
nome of L. chuanxiong contained two divergent haplotypes, which
was further validated by Hi-­C data. By Hi-­C technology, most con-
tigs (96.11% of the genome length) were anchored onto 22 chromo-
somes, and the predominant “1:1” Hi-­C cross-­link signals indicated
11 homoeologous chromosome pairs in L. chuanxiong (fig. S4). Us-
ing haplotype-­specific k-­mers based on SubPhaser (35), we further
phased the genome into two 11-­chromosome haplotypes, hereafter
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
referred to as “LcHapA” and “LcHapB” (fig. S5, A and B). Both hap-
lotype genomes exhibited high completeness from Benchmarking
Universal Single-­Copy Ortholog (BUSCO) (98.10% for LcHapA
and 98.10% for LcHapB) (table S2).
Multiple-­tissue RNA sequencing (RNA-­seq) data, ab initio pre-
diction, and homeolog protein evidence were combined for anno-
tation, totally identifying 101,862 high-­confidence protein-­coding
gene models (52,337 for LcHapA and 49,525 for LcHapB, respec-
tively; tables S2 and S3). Among them, 26,694 pairs were allelic
(syntenic and homoeologous) between LcHapA and LcHapB,
while the other half was haplotype specific. A total of 10,760,310
repetitive elements were annotated, accounting for 78.45% of the
genome, which is consistent with previous study (36). Long termi-
nal repeat (LTR) retrotransposons accounted for the largest pro-
portion (76.02%) of repetitive elements (table S4). For both
haplotypes of L. chuanxiong, expanded gene families mainly en-
riched in secondary metabolic pathways, while contracted gene
families related to immune response, inflorescence meristem
maintenance, and sexual reproduction (figs. S6 to S8). Consistent
with the previous studies (37), we only detected one whole-­genome
duplication (WGD) event shared by the subfamily Apioideae and
did not observe any recent specific WGD event for L. chuanxiong
(figs. S6 and S9).
Exceptional haplotypic structure variations in L. chuanxiong
The high heterozygosity, asymmetric FISH signals, and the high
proportion of haplotype-­specific genes implied great divergence be-
tween the two haplotypes of L. chuanxiong. Further syntenic analy-
sis revealed substantial SVs between LcHapA and LcHapB (Fig. 2C
and fig. S10). In general, 39 translocations and 42 inversions larger
than 1 Mb were detected, comprising 21.20% of the whole genome,
and about half of the SVs were larger than 5 Mb (table S5). On aver-
age, each chromosome contained 3.55 inversions and 3.82 translo-
cations larger than 1 Mb. SVs were most remarkable among Chr1,
Chr2, Chr 9, Chr 10, and Chr 11, where the largest inversion (~35 Mb)
and translocation (~37 Mb) occurred (Fig. 2C).
Such scale and number of haplotypic SVs were exceptional. We
compared the genomic percentage of SVs of L. chuanxiong with two
other diploid species with either moderate or high heterozygosity
(0.69% for Manihot esculenta and 1.28% for Malus pumila) (38, 39),
two alloploid species (Perilla frutescens and Miscanthus lutariori-
parius) (40, 41), and a diploid hybrid (Juglans regia × Juglans micro-
carpa) (42) (fig. S11). We found that the haplotypic SVs only
accounted for 0.13 to 0.16% of the genomes in the two diploid spe-
cies, which was 127-­to 170-­fold lower than that in L. chuanxiong. In
contrast, the inter-­subgenome SVs in hybrid or alloploid species,
representing interspecific SVs, made up 3.70 to 22.09% of the ge-
nome, which were more comparable with the genomic percentage of
haplotypic SVs in L. chuanxiong. These results jointly suggest the
SVs between LcHapA and LcHapB are more prevalent than expected
for the intraspecific level.
Hybrid origin of L. chuanxiong
The long history of asexual reproduction of L. chuanxiong and the
large genetic differentiation between its haplotypes inspired us to
speculate that L. chuanxiong might be derived from interspecific hy-
bridization instead of solely domesticating from its potential wild
progenitor, L. sinense (27–32). To test this hypothesis, we incorpo-
rated L. sinense into comparative genomic analyses. Given the
3 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

LG6B LG7B
LG5B 100 200 0

A
0

B LG
100
200

B
200
100

4 8B
300
0

LG
0
300 10
0 0
20 20
0
1 00
30
0
0
0

LG
3B
0
20

10

9B
LG

0
0

20
10
2 cm

0
0

0
B
200

LG1
LG2

100
100

0B
200
Fresh-cut rhizome

300
LG1B
200

LG11B
0
100

100
0

200
g fedc ba

2 00
LG11A

0
100

100

LG1A
0

200
300
LG1

300
4 cm

200

0
5 μm

100

A
0A

100
LG2
0

0 0
Stem nodes for

0 2
20
1 cm

0
LG

0
vegetative reproduction

10
10
0

3A
9

0
20
0
A
0

LG
30
0 30

LG
0
20
0

Rhizome 8A
0
10

4A
10
0 0
20

LG
0
300 300

LG7
200 0
100 100

A
200

LG5A
0
200 100 0 300

LG6A

C
Chr1 Chr2 Chr3 Chr4 Chr5 Chr6 Chr7 Chr8 Chr9 Chr10 Chr11
HapA

HapB

Linear region Inversion region Translocation region

LG1B LG1A LG2B LG2A LG9B LG9A LG10B LG10A LG11B LG11A
Fig. 2. Genomic features of L. chuanxiong. (A) Asexual plant of L. chuanxiong. The expanded stem nodes are used for vegetative reproduction. The dried swollen rhizomes are fa-
mous TCM “Chuanxiong.” (B) Genome circos plot and karyotype. (a) GC content, (b) TE density, (c) gene density, (d) to (f) expression spectrum of different tissues (leaf, stem, and rhi-
zome), and (g) karyotype. Chromosome number and telomeric repeats distribution (green) of each chromosome were detected by fluorescence in situ hybridization (FISH) assay.
Arrow heads in different colors distinguish different homologous chromosome pairs. Each major tick in the chromosome represents 100 Mbp, and each minor tick represents
50 Mbp. (C) SVs larger than 1 Mb between two haplotypes. Inversion and translocation are marked with yellow and green, respectively. Hi-­C heatmaps show obvious interactions (in
green circles) between Chr1 and Chr2, as well as among Chr9 to Chr11, which provides another piece of evidence for the detected SVs. Resolution of the Hi-­C heatmap is 2.5 Mb.

absence of the genome of L. sinense, we de novo assembled its tran-
scriptome and obtained 86,244 unigenes (table S6).
To evaluate the genetic distance among LcHapA, LcHapB, and
L. sinense, we conducted pairwise comparisons among their coding
sequences (CDSs) from three perspectives: SNPs (single-­nucleotide
polymorphisms), identity, and mapping coverage/quality. First, with
LcHapB as the reference, we detected fewer SNPs for L. sinense than
LcHapA (Fig. 3A). Second, we found that LcHapB and L. sinense
shared the highest mean identity (99.97%) of orthologous CDS,
while LcHapB and LcHapA shared the lowest (99.42%) (Fig. 3B).
Third, we mapped the RNA-­seq reads of L. sinense to each haplotype
and found overall higher mapping rates (88.84 versus 45.94%) and

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 4 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

A C
LcHapA vs. LcHapB Ls vs. LcHapB
1.25-3.22 Ma L. sinense
2000 Constraints
Chr1 Density 2.29-4.83 Ma LcHapB
0
Chr2 HapA
Angelica sinensis
Chr3
Coriandrum sativum
Chr4 Apium graveolens
Chr5 Daucus carota
Bupleurum chinense
Chr6
Centella asiatica
Chr7 Panax notoginseng
Chr8 Panax stipuleanatus
Aralia elata
Chr9
Lonicera japonica
Chr10 Cretaceous Paleogene Neogene

Chr11 130 104 78 52 26 0 Ma

B D 131
Ls vs. LcHapB A: LcHapA
Ls vs. LcHapA B: LcHapB
LcHapA vs. LcHapB 100
S: L. sinense
1.0 O: Outgroups
Counts
Density

50
0.5 35
28

0 B B A
95 96 97 98 99 100
S A S
A S B
Identity of orthologs (%) O O O
Fig. 3. Genetic divergence among haplotypes of L. chuanxiong (LcHapA/B) and L. sinense. (A) Comparison of SNP density within 100 kb nonoverlapped sliding win-
dows among LcHapA, LcHapB, and L. sinense (Ls). (B) The identity of orthologous CDS among LcHapA, LcHapB, and Ls. (C) The species tree and divergence time among
LcHapA, LcHapB, and Ls. The constraints for molecular clock calibration are marked with stars. Except that the stem node of C. sativum has a bootstrap support of 60, any
given node has a support of 100. (D) Number of gene trees for each identified phylogenetic relationship among LcHapA, LcHapB, and L. sinense.

mapping quality (12.69 versus 7.18) for LcHapB (fig. S12). Together,
the CDS of LcHapB was more identical with L. sinense than LcHapA,
indicating a closer genetic relationship between LcHapB and
L. sinense.
Alternatively, we constructed the phylogenetic trees with 194
single-­copy orthologs shared by LcHapA, LcHapB, L. sinense, and 10
other species from Apiales and Dipsacales (table S7). The species
tree strongly supported that LcHapB shared the most recent ances-
tor with L. sinense instead of LcHapA with bootstrap support value
of 100 (Fig. 3C). Specifically, 131 (67.53%) gene trees supported a
sister relationship between LcHapB and L. sinense, while only 38
(18.04%) presented a closer relationship between LcHapA and
LcHapB (Fig. 3D and fig. S13). Moreover, the molecular clock esti-
mated an earlier divergence age between LcHapA and LcHapB
(~3.42 Ma) than that between LcHapB and L. sinense (~2.12 Ma)
(Fig. 3C). Thus, the substantial genetic divergence observed be-
tween LcHapA and LcHapB might partially represent variation be-
tween parental species of the two haplotypes spanning 3.42 million

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 5 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

years (Ma). In summary, the collective evidence confirmed the dip-
loid hybrid nature of L. chuanxiong. LcHapB may be the descendant
of L. sinense or its relative species, while LcHapA originated from a
more distant species that need to be further investigated.
Balanced ASE in L. chuanxiong
Although some level of imbalance between the gene expression of pa-
rental alleles in the hybrid has been revealed by several studies (43,
44), our data showed a non-­bias pattern of the overall expression be-
tween LcHapA and LcHapB (Fig. 4A). To examine the allele-­specific
expression (ASE) of L. chuanxiong, we compared the expression level
of all 26,694 syntenic and homoeologous gene pairs between LcHapA
and LcHapB. We detected ASE in at least one tissue between 4381
(16.41%) allele pairs, while most alleles (> 70%) did not exhibit sig-
nificant ASE toward either haplotype in any tissue (Fig. 4B). Mean-
while, the ASE genes (ASEGs) toward each haplotype (A-­bias or
B-­bias genes) were roughly even (2224 versus 2157) (Fig. 4B and
fig. S14). These results indicate an overall balanced gene expression
between the two haplotypes of L. chuanxiong, with limited level of ASE.
We further investigated the distribution and functions of ASEGs.
Despite exceptional haplotypic SVs in L. chuanxiong, we found no
significant difference in the proportion of ASEGs between SV and
non-­SV regions in any tissue (Fisher’s exact test, P > 0.05; Fig. 4C
and tables S8 to S10), suggesting that SVs did not affect ASEG distri-
bution. Furthermore, we evaluated the evolutionary rate of ASEGs
and found that ASEGs showed significantly higher Ka (nonsynony-
mous substitution rate) and Ks (synonymous substitution rate) than
non-­ASEGs (Student’s t test, P < 0.05; Fig. 4D and fig. S15, A and B).
This observation indicates that the evolutionary rate of ASEGs is
higher. These fast-­evolving ASEGs were mainly enriched in adenos-
ine 5′-­triphosphate–binding Gene Ontology (GO) term (fig. S15, C
and D) for both A-­bias and B-­bias genes.
The inferred scheme of phthalide conversion and in silico
oxidase gene mining
On the basis of the high-­quality assembled and annotated genome,
we started to elucidate the conversion processes among (S)-­NBP (1),
butylidenephthalide (2), senkyunolide A (3), and ligustilide (4) in
L. chuanxiong. Given that little is known about the conversion direc-
tions and connectivity, we conducted an in vitro crude protein activ-
ity assay to explore the substrate-­product relationships. We initially
treated each of the four phthalides with the rhizome crude protein

A C LIN INV TRA

LcHapA LcHapB ns ns ns
ns ns ns
15 100

80
Percentage of ASEGs (%)
Log2 (FPKM + 1)

10
60

40
5

20

0
LIN INV TRA LIN INV TRA LIN INV TRA 0

Leaf Rhizome Stem Leaf Rhizome Stem

B A bias B bias Balanced Non-expression D ASEG Non-ASEG

80 ns
4
Percentage of genes (%)

60
3

40
Ka

20
1

0 0
Leaf Rhizome Stem Total LIN INV TRA

Fig. 4. ASE between the LcHapA and LcHapB of L. chuanxiong. (A) Comparison of gene expression between LcHapA and LcHapB in different SV and tissue types. It is
insignificant for all compared groups. (B) Percentage of genes among different allele-­specific expression (ASE) types across three tissues. Alleles were classified into four
types according to the expression imbalance between two haplotypes: bias toward LcHapA (A bias, in blue), bias toward LcHapB (B bias, in red), balanced expression be-
tween LcHapA and LcHapB (in light blue), and non-­expression (in gray). (C) Comparisons of the percentage of ASEGs among different SV types in three tissues. (D) Com-
parisons of Ka between ASEGs and non-­ASEGs in at least one tissue among different SV types and across the whole genome. LIN, linear region; INV, inversion; TRA,
translocation; Total, whole genome. *P < 0.05; **P < 0.01; ***P < 0.001; not significant (ns), P > 0.05.

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 6 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
extract (CPE) of L. chuanxiong. If any confident signal for the other
three phthalides was detected by gas chromatography–mass spec-
trometry (GC-­MS), then we would determine the conversion direc-
tion from the substrate phthalide toward the appeared one(s). As a
result, we found high-­confidence signals of (S)-­NBP and butylide-
nephthalide when senkyunolide A and ligustilide were used as sub-
strates (Fig. 5A). Although it is tempting to infer that these two
conversion processes are CPE-­catalyzed, it may instead be attribut-
able to the potential spontaneous reactions as described in previous
studies (45, 46, 47). Thus, we set substrates plus boiled CPE as the
negative control groups. In the boiled-­CPE group, (S)-­NBP and bu-
tylidenephthalide caused by spontaneous reaction were detected,
but they were significantly and marginally significantly lesser than
CPE groups in the conversion efficiency from ligustilide to butylide-
nephthalide (P = 0.001, one-­tail Wilcoxon rank sum exact test) and
from senkyunolide A to (S)-­NBP (P = 0.066, one-­tail Wilcoxon rank
sum exact test; Fig. 5B), respectively. These results confirmed that
the catalysis effect of CPE is not totally masked by spontaneous pro-
cess. These findings indicated these two conversions might contain
both spontaneous and enzyme-­catalyzed processes. Accordingly, we
next focused on unveiling the enzymes that could significantly pro-
mote these two conversion reactions (Fig. 5C).
As these two conversion steps are both desaturation processes,
we integrated multi-­omics data to mine the candidate oxidase genes.
Given that LcHapA and LcHapB contained around 50% haplotype-­
specific genes, we incorporated all genes from both haplotypes into
the base gene pool. On ground of the assumption that metabolite
accumulation is often positively correlated with the expression of
genes involved in metabolic biosynthesis pathway (48), we initially
selected the significantly up-­regulated [log2(fold change) > 1, P < 0.05,
false discovery rate (FDR) < 0.05] and highly expressed genes [frag-
ments per kilobase of exon model per million mapped fragments
(FPKM) > 20] in the rhizome, where the abundance of (S)-­NBP and
butylidenephthalide was peaked (Fig. 5D). In total, 2718 genes were
chosen (fig. S16). Subsequently, we screened out 2099 and 2174
genes, whose expression was significantly correlated with the con-
tent of (S)-­NBP and butylidenephthalide (Pearson correlation,
P < 0.05, r > 0.95), respectively. Given that genes of the DOXC family
in 2OGD superfamily and CYP superfamily typically catalyze oxi-
dation reactions in plant specialized metabolite (49, 50), we re-
trieved 14 2OGDs (DOXC family) and 45 CYPs in the 2099 as well
as the 2174 gene set (table S11). Although fatty acid desaturases
(FADs) are also typical enzymes for desaturation reactions (51),
none FADs are present in the 2099-­and 2174-­gene sets, and, thus,
they are not included for further verifications.
However, it is still experimentally laborious to characterize 59 genes.
We conducted phylogenetic analyses to further narrow down the
pool of candidate genes. For each DOXC clade or CYP family, we
only selected one to three highly expressed genes for functional
characterization (figs. S17 and S18). The final candidate set was lim-
ited to 10 2OGDs and 24 CYPs (Fig. 5E and table S11), among which
8 2OGDs and 11 CYPs were successfully cloned (Fig. 5E).
Four Lc2OGDs and two LcCYPs potentially promote the
phthalide C-­4/C-­5 desaturations
The enzymatic activities of 2OGDs were verified in Escherichia coli
system, while the functions of CYPs were characterized in Nicotia-
na benthamiana and yeast (Saccharomyces cerevisiae). The gene
functions were characterized by both feeding assays and in vitro
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
recombinant enzyme activity assays, as a solo method might be
limited in elucidating the enzymatic catalysis activity under the
noise generated by spontaneous processes. As a starting point, we
transiently expressed each cloned 2OGD or CYP and fed the E. coli
cells or N. benthamiana leaves with senkyunolide A and ligustilide
and then detected (S)-­NBP and butylidenephthalide. We found
that four 2OGDs (LCX6BG001106, Lc2OGD1; LCX6AG001172,
Lc2OGD2; LCX2BG002971, Lc2OGD3; and LCX7BG002917,
Lc2OGD4) and two CYPs (LCX5AG000851, LcCYP72A1132; and
LCX5BG004393s2, LcCYP716E110) could significantly promote
the conversion from senkyunolide A to (S)-­NBP, as well as ligusti-
lide to butylidenephthalide (Fig. 6 and figs. S19 and S20). As an-
ticipated, obvious signals of (S)-­NBP or butylidenephthalide caused
by spontaneous reactions were also detected in the empty vector
(EV) groups (GC-­MS peak plots in Fig. 6). However, signals in
Lc2OGD1 to Lc2OGD4 and LcCYPs added groups were approxi-
mately stronger by 1.63-­to 27.94-­fold than EV groups (bar plots in
Fig. 6). We found that fully identical oxidases were involved in the
two steps of conversion, which may result from the same desatura-
tion position (C-­4/C-­5 of phthalide) and similar structure of sen-
kyunolide A and ligustilide. Lc2OGD1 and Lc2OGD2 were
homoeologous alleles on Chr6, while Lc2OGD3 and Lc2OGD4
were located on Chr2B and Chr7B. The LcCYPs were distributed on
Chr5A and Chr5B. As a result, two phthalide C-­4/C-­5 desaturases
(P4,5Ds) were distributed in LcHapA and four in LcHapB.
To further validate the activity of these P4,5Ds, we carried out in
vitro enzymatic activity assay using purified recombinant enzymes
from E. coli and microsomal protein from yeast. Compared with the
EV groups, Lc2OGD2-­, Lc2OGD3-­, and both LcCYP-­added groups
revealed significantly higher (S)-­NBP production, and all Lc2OGDs
and LcCYPs showed significantly higher butylidenephthalide pro-
duction (boxplots in Fig. 6). Kinetics parameters of these P4,5Ds re-
vealed consistent catalysis efficiency with that observed in feeding
assays. In both assays, Lc2OGD2 was the most efficient 2OGD en-
zyme in producing (S)-­NBP (kcat/Km = 0.38 μM−1 min−1), Lc2OGD1
was the best performing 2OGD enzyme in promoting butylidene-
phthalide production (kcat/Km = 0.17 μM−1 min−1), and the two LcC-
YPs exhibited different preference in promoting different conversions
(Fig. 6 and fig. S21). The most reliable P4,5Ds were Lc2OGD2,
Lc2OGD3, and both LcCYPs that consistently promoted the forming
of (S)-­NBP/butylidenephthalide in both feeding assays and in vitro
enzymatic activity assays. Given the relatively weak enzymatic activ-
ity of these P4,5Ds (Km > 4.77 μM), it cannot be excluded that the
most efficient P4,5Ds in L. chuanxiong are not yet identified. Togeth-
er, four of the six P4,5Ds effectively mediated the two investigated
conversions and different enzymes demonstrated distinct preferences.
Redundancy and phylogenetically independent recruitment
of the P4,5Ds
As the desaturase function for phthalides has never been reported
in 2OGD and CYP superfamily, we further investigated the phy-
logenetic relationships between P4,5Ds and 96 characterized
2OGDs and 67 characterized CYPs (tables S12 to S14). Lc2OGD1
to Lc2OGD3 belonged to DOXC52 clade, while Lc2OGD4 was
located in DOXC30 clade [clade names following the previous
study (50)] (Fig. 5D). DOXC30 comprises the widely distributed
feruloyl–coenzyme A 6′-­hydroxylase involved in coumarin bio-
synthesis pathway (fig. S22 and table S12) (50). Nevertheless,
DOX52 targeted more diverse substrates, including dopamine,
7 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

Fig. 5. The hypothesis of the conversions among phthalides and mining for the candidate oxidase genes. (A) GC-­MS base peak plot of four phthalides after treating
each of them with CPE. The numbers 1 to 4 represent (S)-­NBP, butylidenephthalide, senkyunolide A, and ligustilide, respectively. (B) Comparison of the catalysis efficiency
between CPE and boiled CPE groups in promoting the conversion from 3 to 1 and from 4 to 2. All treatments were conducted for six times. One-­tail Wilcoxon rank sum
exact test was used for the significance test. ns, P > 0.05; **P < 0.01. (C) Hypothesis for phthalide conversions inferred from in vitro crude protein activity assay. The dashed
lines imply target conversions investigated in this study. (D) Content of phthalide 1 to 4 from the stem, petiole, leaf, and rhizome extracts of L. chuanxiong. Data are pre-
sented as the means ± SEM. (E) Candidate oxidases (2OGDs and CYPs) genes producing (S)-­NBP and butylidenephthalide. The heatmap details the expressional abundance
of each candidate gene across three replicates of four tissues. The gradient color of the scale bar represents relative expression. The gene names on the phylogenetic tree
marked with gray indicate genes not successfully cloned; black gene names imply inactive genes; red bold ones show active genes involved in phthalide conversions.

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 8 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

A D

B E

C F

Fig. 6. Lc2OGD1–4, LcCYP716E110, and LcCYP72A1132 mediate the phthalide conversions. (A to C) GC-­MS peak plots and the production of (S)-­NBP (1) [mass/
charge ratio (m/z) of 133] in the indicated combinations of senkyunolide A (3) and LcCYPs or Lc2OGDs. (D to F) GC-­MS peak plots and the production of butylidene-
phthalide (2) (m/z of 159) in the indicated combinations of ligustilide (4) and LcCYPs or Lc2OGDs. (A) and (D) The reactions investigated in this study. (B) and (E) The char-
acterization of LcCYPs in N. benthamiana and yeast through feeding assays (GC-­MS peak plots and bar plots) and in vitro enzymatic activity assays (boxplots), respectively.
(C) and (F) The characterization of Lc2OGDs in E. coli through feeding assays (GC-­MS peak plots and bar plots) and in vitro enzymatic activity assays (boxplots). All treat-
ments involving different genes and substrates were performed three to four times. The error bars in bar plots represent SEM. Empty vector (EV) was used as the negative
control. ns, P > 0.05; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. Weltch t test was used for statistical comparison.

4-­hydroxyphenylacetaldehyde (52), thebaine (53), and melatonin
(54). In addition, the two LcCYPs belong to two distant families,
CYP72 and CYP716, which are conventionally involved in terpe-
noid biosynthesis, especially triterpenoid tailoring (figs. S23 and
S24 and tables S13 and S14). Thus, P4,5Ds are likely to originate
from multiple pathways and distinct gene lineages.
In addition, the functional redundancy was noteworthy for
phthalide C-­4/C-­5 desaturations that potentially involved six genes
(Lc2OGD1 to Lc2OGD4, LcCYP72A1132, and LcCYP716E110).
This clue prompted us to investigate more potential P4,5Ds by reex-
amining the candidate CYPs for (S)-­NBP or butylidenephthalide
biosynthesis. Four more candidate CYP716Es were physically linked

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 9 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
to LcCYP716E110. Further examination for adjacent genomic re-
gions detected a densely distributed tandem duplication containing
19 copies in LcHapB (~662 Kb) and nine in LcHapA (~200Kb). Syn-
tenic analysis for different Apiaceous species showed five CYP716Es
in Bupleurum chinense, three in Apium graveolens, and two in Cori-
andrum sativum within this syntenic region, suggesting that CY-
P716E subfamily might be expanded in L. chuanxiong, especially in
LcHapB (figs. S25 and S26). These copies shared high–amino acid
sequence identity (57 to 96%) with LcCYP716E110, suggesting sim-
ilar catalytic functions, which await experimental verification in
the future.
DISCUSSION
Despite the wide application of NBP for ischemic stroke therapy (7,
14), no enzyme involved in its biosynthesis has been known to date.
The present study is an icebreaker for the dissection of phthalide
biosynthesis pathway, where four 2OGDs and two CYPs were
proved to potentially promote the conversions from senkyunolide A
to (S)-­NBP and ligustilide to butylidenephthalide in L. chuanxiong.
As senkyunolide A is the second abundant phthalide in L. chuanxiong
(Fig. 5D), ~47-­fold of the content of (S)-­NBP, up-­regulation of
P4,5Ds in vivo would be a promising way to enhance the content of
(S)-­NBP in L. chuanxiong. Our discovery will facilitate the future
molecular breeding of L. chuanxiong and metabolic engineering for
(S)-­NBP production.
An unusual feature of P4,5Ds is that the enzymatic reactions in-
volving spontaneous processes. Such reactions were also reported in
terpenoid, hyoscyamine biosynthesis, and S-­methyl thioester (55–
57). Because of the noise generated by spontaneous process, multiple
lines of evidence and proper negative control settings are necessary
to validate the function of enzymes. In this study, in vivo feeding as-
say and in vitro enzymatic assay brought out different revelation in
validating the function of P4,5Ds. The feeding assay appeared to
identify more P4,5Ds, which might be due to that the live cells in
feeding assay may provide a more suitable and stable enzymatic reac-
tion environment. However, the in vitro enzymatic assay using puri-
fied enzyme is particularly useful for eliminating false positive
enzymes. In sum, we advocate integrative methods in investigating
the role of enzymatic catalysis in such spontaneous reactions.
The substrate promiscuity and function redundancy were fea-
tured for P4,5Ds. Jointly considering the data from feeding assay
and in vitro enzymatic assay, at least four P4,5Ds could recognize
both senkyunolide A and ligustilide that differ in C-­3/C-­8 bond, in-
dicating that P4,5Ds are promiscuous to different phthalides and
possibly mediate other C-­4/C-­5 desaturating reactions, such as from
senkyunolide F to senkyunolide E (Fig. 7). Thus, the substrate pro-
miscuity may greatly contribute to the high phthalide diversity in
L. chuanxiong, as reported in other metabolic pathways (49, 58, 59).
Moreover, P4,5Ds are independently recruited from both 2OGD
and CYP lineages. To the best of our knowledge, the involvement of
both 2OGD and CYP in the same reaction is rarely reported (49, 50,
60). An exception is flavone synthase (FNS) that belongs to 2OGD
superfamily in Apiaceae but is CYP in other taxa (50). Noteworthy,
the reactions that FNS mediated are also desaturation, converting
the C─C in C-­3 of naringenin (flavanone) to C═C. This finding mo-
tivated us to speculated that more desaturation reaction in the form-
ing of natural product might involve both 2OGD and CYP. From
this perspective, our results unveil the probably underestimated
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
gene function redundancy in plant secondary metabolite biosyn-
thetic pathways and advocate a comprehensive survey of genes from
different families with similar functions.
In addition to the insights into (S)-­ NBP biosynthesis, the
haplotype-­phased genome also lifted the curtain of the origin of
L. chuanxiong. Notably, our comparative genomic analyses showed
that L. chuanxiong was a sterile hybrid from two different diploid
progenitors that diverged more than 3.42 Ma. One of the parents of
L. chuanxiong may be or is very related to L. sinense, explaining their
high morphological and genetic similarities. However, the other
parent is phylogenetically more distant and awaits further explora-
tion with a denser sampling of related Ligusticum species. Because
L. chuanxiong has been vegetatively cultivated for >1600 years in
China (27), rarely blooming and bearing mature seeds (28, 29), the
taxonomic treatment has been a puzzling problem. L. chuanxiong
has several potential synonyms, such as Ligusticum striatum (= Seli-
num striatum) and Ligusticum wallichii (28). In addition, Pu (29)
even treated it as a cultivated variety of L. sinense. Recent mole­
cular systematic studies recognized more related species with
L. chuanxiong, such as Ligusticum jeholense, Cnidium officinale,
Ligusticum nematophyllum and Ligusticum tenuissimum (30–32).
These species may be conspecific with or potential parents of
L. chuanxiong and thus should be included in future studies to re-
volve the nomenclatural and taxonomic issues of L. chuanxiong. Al-
though the sampling of related species is limited, our evidence
supports L. chuanxiong as a distinct hybrid species from L. sinense
and thus provides a more case for homoploid hybrid speciation
(HHS) that is increasingly recognized in recent years (61–63).
Among the HHS examples, the speciation of Ostrya intermedia via
alternatively inheriting parental isolating allelic genes is impressive
(62). However, L. chuanxiong revealed a different isolation mecha-
nism through asexual reproduction that acts as an unbridgeable
barrier for gene flow between L. chuanxiong and its parents. As
many diploid hybrid species propagate asexually, such as Oenothera
(Onagraceae) (64), Daphnia pulex (65), and Ranunculus (66), the
loss of sexual reproduction may be an alternative common HHS
mechanism. For L. chuanxiong, the sterility may result from several
reasons. The substantial chromosomal SVs have been frequently re-
ported to induce higher sterility levels because of the inappropriate
pairing and separation during the meiosis process (64, 67, 68).
Moreover, gene families related to inflorescence meristem mainte-
nance and sexual reproduction were significantly contracted in
LcHapA and LcHapB (fig. S7), which may limit the overall ability for
flowering of L. chuanxiong.
Although the two divergent haplotypes show substantial varia-
tion at both sequence and structure levels, they generally remain
balanced in gene expression and seem to both contribute to the
forming of (S)-­NBP. Although our discoveries of P4,5Ds represent a
small portion of the complete biosynthetic pathway of (S)-­NBP
(Fig. 7), the possibility that P4,5Ds are not dominant on solo-­
haplotype but on both haplotypes indicates that the higher content
of (S)-­NBP in L. chuanxiong than its potential parent Ligusticum si-
nensis (fig. S27) might benefit from the effect of heterosis on the
content of secondary metabolites (69). Thus, it highlights the sig-
nificance of further investigation of the likely parental species of
L. chuanxiong with a comprehensive sampling of species in Ligusti-
cum and design artificial interbreeding experiments to evaluate the
effect of heterosis on the content of important secondary metabo-
lites, such as (S)-­NBP, in the future study.
10 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e

O O O

Initial substrates CoA-S + 5× CoA-S OH

Acetyl-CoA Malonyl-CoA

Polyketide pathway or
fatty acid pathway

Phthalide tailoring and interconversion

C-3/C-8 desaturation C-3/C-8 desaturation

C-3/C-8 desaturation
O O O O
C-4/C-5 C-4/C-5 C-4/C-5
desaturation desaturation desaturation
O O O O
(S)-NBP Senkyunolide A Butylidenephthalide Ligustilide

C-4 hydroxylation C-4 hydroxylation C-9 hydroxylation C-9 hydroxylation

OH OH
OH OH

O O O O
C-4/C-5 C-4/C-5
O desaturation O O desaturation
O
(16) 4-hydroxy-3- Senkyunolide K Senkyunolide E Senkyunolide F
butylphthalide

Fig. 7. Proposed biosynthetic pathway for (S)-­NBP. The solid lines represent characterized steps in the present study, while the dotted lines represent potential but
unresolved steps. Different reaction types were distinguished with different colors. The upstream polyketide or fatty acid pathway is detailed in fig. S1. CoA, coenzyme A.

In conclusion, our study provided crucial genomic information
of L. chuanxiong identified several P4,5Ds and revealed the hybrid
origin of L. chuanxiong. These findings will greatly facilitate the fu-
ture metabolic engineering for (S)-­NBP production and point out
the direction for future breeding of L. chuanxiong varieties. How-
ever, several questions regarding the phthalide biosynthesis and ori-
gin of L. chuanxiong have yet to be answered. In addition to phthalide
C-­4/C-­5 desaturations, there exist various oxidatively tailoring pro-
cesses for phthalide (Fig. 7). The dissection of phthalide scaffold
forming is crucial for fully resolving the phthalide biosynthetic
pathway and for future metabolic engineering (fig. S1). In addition,
our study leaves a gap to fill, to investigate the most likely parents of
the hybrid, in our future research.
MATERIALS AND METHODS
Chemical reagents
All standard chemical reagents were purchased from commercial ven-
dors unless noted otherwise. The following phthalides were obtained
from commercial sources: (S)-­3-­​n-­butylphthalide [(S)-­NBP; Chengdu
Must Bio-­Technology Co. Ltd., Chengdu], butylidenephthalide
(Chengdu Must Bio-­Technology Co. Ltd., Chengdu), senkyunolide A
(Chengdu Herb Substance Co. Ltd., Chengdu), and ligustilide (Chengdu
Phytopurify Co. Ltd., Chengdu).
Plant materials
In November 2020, 50 seedlings of L. chuanxiong were collected
from the top-­geoherbalism region Pengzhou county (Sichuan, China),
and 10 seedlings of L. sinense from Lushi county (Henan, China)
were collected. These seedlings were subsequently transplanted to
the greenhouse of the Agricultural Genomics Institute at Shenzhen
(Chinese Academy of Agricultural Sciences, China). After growing
for 6 months, those healthy plants with similar biomass were se-
lected for genomic sequencing, RNA-­seq, and phthalide extraction.
The voucher specimens (vouchered as Nie1701 to Nie1703) of
L. chuanxiong have been deposited in SZG (Herbarium, Fairy Lake
Botanical Garden, Shenzhen, and Chinese Academy of Sciences).
The SZG barcodes are SZG00120205 to SZG00120207.
The 4-­to 5-­week seedlings of N. benthamiana were used for
Agrobacterium-­mediated gene transformation. The seedlings were
grown at laboratory temperature under growth lights with a
16-­hour/8-­hour light/dark cycle.

Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024 11 of 17

S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
Genomic DNA preparation, RNA extraction, and
cDNA preparation
High–molecular weight genomic DNA was extracted from 3 g of
fresh leaves of L. chuanxiong stored at −80°C using CTAB method
and purified with a QIAGEN Genomic kit (QIAGEN, Germany).
The DNA integrity and purity were monitored on 1% agarose gels
and by NanoDropTM One UVVis spectrophotometer (Thermo
Fisher Scientific, USA). DNA concentration was further measured
using the Qubit 4.0 Fluorometer (Invitrogen, USA).
Total RNA was extracted from each tissue (rhizome, stem, petiole,
and leaf) of L. chuanxiong using the RNeasy Plant Mini Kit (Omega,
USA). The first-­strand cDNAs were synthesized by reverse transcrip-
tion polymerase chain reaction (PCR) from the total RNA samples as
templates using the SuperScript III First-­Strand Synthesis System
with the oligo(dT) 20 primer (Thermos Fisher Scientific, USA).
Genome survey
The genome size of L. chuanxiong was estimated by flow cytometry
assay. The fresh leaves of three biological replicates were chopped in
LB01 lysis buffer [15 mM tris, 2 mM Na2 EDTA, 0.5 mM spermine
tetrahydrochloride, 80 mM KCl, 20 mM NaCl, 0.1% Triton X-­100,
and 15 mM β-­mercaptoethano (pH 7.0 to 8.0)] to release the nucleus.
The resulting nucleus suspension was filtered through a 40-­μm cell
strainer and then treated with ribonuclease A (20 μg ml−1) and
propidium iodide (PI; 20 μg ml−1). After a 30-­min incubation on ice
in the dark, the fluorescence intensity of PI-­stained nuclei was deter-
mined by flow cytometry (Beckman Coulter, CytoFLEX) with a
sample flow rate of 10 μl/min. We used fennel (Foeniculum vulgare)
(1C = 1.34 Gb) (70) as the internal reference. All analyses were per-
formed in triplicate for each sample.
The genomic heterozygosity of L. chuanxiong was surveyed by
k-­mer method based on 105-­Gb Illumina short-­read sequencing
data (~30×). Jellyfish (v2.2.10) (71) was performed to generate a 21-­
mer frequency distribution, which was then analyzed in Genome-
Scope 2.0 (72) to estimate heterozygosity.
Cytogenetic and FISH assay to determine karyotype
FISH and cytogenetic assays were conducted to determine the ploi-
dy level and chromosome number of L. chuanxiong. Two-­centimeter
root tips of seedlings were cut and treated with nitrous oxide on ice
for 2 hours. The chromosomes were then fixed for 10 min with 90%
glacial acetic acid. After washing twice with ddH2O, the samples
were chopped and incubated in a 25-­μl solution with a 3:1 mixture
of 2.0% cellulase and 1.0% pectinase for 1 hour at 37°C. The enzy-
matically hydrolyzed root tips were further mashed with tweezers
and rinsed thrice with 70% alcohol. The pretreated root-­tip cells
were suspended in 45% glacial acetic acid and dropped onto a clean
slide with a cover glass at 23°C. The slides containing cells in meta-
phase were selected using a light microscope.
The probes (TTTAGGG) for telomere hybridization were labeled
with fluorescein-­12–deoxyuridine triphosphate, a green fluorescein.
Labeled probe solution of 0.25 μl mixed with 8 μl of reaction buffer
[1× TE buffer and 2× SSC buffer (pH 7.0)] was added to each slide
containing cells in metaphase. The slides were incubated at 80°C for 5 min
and then transferred to an in situ hybridization box at 40°C for 12 hours.
After washing twice with ddH2O, the slides were stained with 8 μl
of 4′,6-­diamidino-­2-­phenylindole and covered with coverslips. The
cells in metaphase after hybridization were observed and imaged using
a high-­resolution fluorescence microscope (Olympus BX50, Japan).
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
Genome sequencing, assembly, and quality control
For the PacBio library construction, 15 μg of genomic DNA from
each leaf sample of L. chuanxiong was fragmented to ~15 kb using g-­
TUBEs (Covaris, USA). After removing short fragments and single-­
strand overhangs, the retained fragments were converted into the
proprietary SMRTbell library using the PacBio DNA Template Prepa-
ration Kit (Pacific Biosciences, CA, USA). Single-­molecule real-­time
sequencing was performed on a PacBio Sequel II sequencing platform.
For Hi-­C library construction, chromatin was first fixed in place
with formaldehyde in the nucleus and then extracted. The extracted
chromatin was digested with Dpn II. The 5′ overhangs of resulting
fragments were then filled in with biotinylated nucleotides, and free
blunt ends were ligated. After ligation, the DNA was purified from
protein and treated following the Illumina Next Generation manufac-
turer’s instructions. The libraries were subsequently sequenced on Il-
lumina Hiseq X, producing 296-­Gb 2 × 150–base pair paired-­end reads.
The raw data of PacBio subreads were filtered to HiFi reads by
PBccs (v6.4.0) (https://github.com/PacificBiosciences/ccs) and sub-
sequently assembled using Hifiasm (v0.16.0) (73). The initial as-
sembled contigs were anchored to chromosomes by 3D-­ DNA
pipeline (v201008) (74) and further manual adjustments were made
to produce a chromosome-­level genome. BUSCO (v5.4.3) (75) was
used for benchmarking the genome with the “embryophyte_odb10”
database.
Genome annotation
EDTA (v2.0.0) (76) was used to de novo identify, annotate, and clas-
sify the repetitive elements in the genome of L. chuanxiong. Before
protein-­coding gene annotation, the annotated repetitive elements
in the genome were soft-­masked using bedtools (v2.28.0) (77).
RNA-­seq raw reads of L. chuanxiong were filtered using fastx-­
toolkits (v0.0.14) (http://hannonlab.cshl.edu/fastx_toolkit/index.
html) and then assembled through Hisat2 (v2.2.1) (78) and Stringtie
(v2.2.0) (79). The raw assembly of transcripts was further validated
by PASA (v2.5.1) (80), which were then incorporated into the MAKER
(v3.01.03) (81) pipeline to automatedly identify protein-­ coding
genes. Last, the gene models identified by MAKER (v3.01.03) (81)
were updated by PASA (v2.5.1) (80). Functional annotations of the
protein-­coding genes were carried out by using BLASTP searches
against entries in both NCBI nonredundant protein (www.ncbi.nlm.
nih.gov/) and Swiss-­Prot (www.uniprot.org/) databases. The predic-
tion of conserved domains for the genes was performed by InterPro-
Scan (v5.11-­51.0) (82). The annotations of the GO terms (http://
geneontology.org/) and Kyoto Encyclopedia of Genes and Genomes
pathways [(KEGG) www.genome.jp/kegg/] for the genes were added
using eggNOG-­mapper (v2.1.10-­0) (83).
Haplotype phasing for the genome
SubPhaser (v1.2) (35) pipeline was used to phase the genome. Jelly-
fish (v2.2.10) (71) was first performed to generate a 17-­mer frequency
distribution. Then, a K-­means algorithm was used to cluster k-­mers
throughout the whole genome. The significance level of k-­mers clusters
was evaluated by the bootstrap method. Last, subgroups within
genome were identified by performing hierarchical clustering and
principal components analysis.
Detection of SVs and SNPs
Given the large genetic distance between the two haplotypes of
L. chuanxiong, it was challenging to accurately and directly align the
12 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
whole genomic sequence of LcHapA and LcHapB. Thus, we used the
syntenic gene block strategy to align and detect the SVs between
haplotypes. Jcvi (v1.1.18) (84) was used to detect collinear gene
blocks between LcHapA and LcHapB, and only those unique gene
block pairs (1:1) were remained for SV detection. Possible duplica-
tions caused by ancient WGD events or other fragment duplication
events were removed to avoid overestimating SVs.
To compare the sequence similarity among LcHapA, LcHapB,
and L. sinense, we performed SNP calling for LcHapA and L. sinense
using LcHapB genome as the reference based on RNA-­seq data.
First, we performed RNA-­seq for L. sinense. Second, we extracted
the LcHapA-­specific reads from the total reads initially mapped to
a combined reference of two haplotypes using the LcHapA genome.
Third, RNA-­seq reads of LcHapA and L. sinense were mapped to
LcHapB using Hisat2 (v2.2.1) (78). Subsequently, GATK (v4.2.4.0)
(85) best-­practice workflow was used to call SNPs with parameter
“-­-­s ample-­ploidy 1 -­-­intervals.” Last, the SNP number scanned
by 100-­kb nonoverlapped sliding windows was compared between
LcHapA and L. sinense.
De novo assembly of transcriptome of L. sinense
RNA-­seq raw reads of L. sinense were filtered using fastx-­toolkits
(v0.0.14) (http://hannonlab.cshl.edu/fastx_toolkit/index.html) and
then de novo assembled into transcripts using Trinity (v2.13.2)
(86). The longest transcript was chosen by Trinity in-­house scripts.
TransDecoder (v5.5.0) (https://github.com/TransDecoder/Trans-
Decoder) was used to predict CDS. The orthologous gene sets
among LcHapA, LcHapB, and L. sinense were obtained using the
bidirectional best matching (BBH) method performed through
BLAST (v2.13.0+) (87). Last, the CDS identity of each orthologous
gene pair among LcHapA, LcHapB and L. sinense was calculated
by MEGA11 (88).
Comparative genomic analyses
To clarify the phylogenetic relationship and divergence time
among LcHapA, LcHapB, L. sinense, and the other 10 species
(Fig. 3C and table S7), we constructed maximum likelihood (ML)
trees based on single-­copy orthologous genes obtained from RNA-­
seq data. In total, we identified 193 single-­copy orthologs using
OrthoFinder (v2.5.4) (89) with default settings and aligned each
of them by MUSCLE (v5.1.linux64) (90). These aligned orthologs
were concatenated by Phylosuite (v1.2.2) (91). ModelTest-­NG
(v0.1.7) (92) was used to detect the best-­fit amino acid substitution
model, based on which RAxML-­NG (v1.1.0) (93) was used to con-
struct the ML phylogeny with 1000 bootstrap analyses. MCMCTree
tool of PAML package (v4.9j) (94) was used to estimate the diver-
gence time, with three constraints for calibration (Fig. 3D and
table S15).
WGDI (v0.6.1) (95) was used to detect the WGD events in Apia-
les. The gene family expansion and contraction analyses were per-
formed using CAFE5 (v5.0) (96), where the gamma rate was set
from 1 to 7, and “k = 5” was selected for best fit. GO enrichments
were performed by ClusterProfiler (v4.6.0) (97) for those signifi-
cantly expanded or contracted gene families.
ASE analyses
To determine the ASEs, we comprehensively compared gene ex-
pression levels between LcHapA and LcHapB in different genomic
structure (linear, inversion, and translocation regions) and tissue
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
(rhizome, leaf, petiole, and stem) types. We initially identified al-
lele pairs in each syntenic gene block based on BBH methods by
BLAST (v2.13.0+) (87). Then, we quantified the expression level of
each allele based on RNA-­seq data. Clean RNA-­seq reads were
mapped to LcHapA and LcHapB through Hisat2 (v2.2.1) (78) and
assembled by Stringtie (v2.2.0) (79). The expression count matrix
for all allele pairs was constructed by a script prepDE.py down-
loaded from the stringtie website (http://ccb.jhu.edu/software/
stringtie/dl/prepDE.py). On the basis of this matrix, ASEGs
were identified using DESeq2 [P < 0.05, FDR < 0.05, log2(fold
change) > 1] (98). Ka and Ks of all allele pairs were calculated by
the Ks module of WGDI (v0.6.1) (95). Student’s t test was used to
compare Ka, Ks, or Ka/Ks between ASEGs and non-­ASEGs in each
genomic structure and tissue type with a significance level of
P < 0.05. Fisher’s exact test was used to compare the ASEGs fre-
quency between SV and non-­SV regions with a significance level
of P < 0.05.
Phthalide extraction and GC-­MS analysis
Approximately 100 mg of tissue samples were dissected and ho-
mogenized to a fine powder on a ball mill. Phthalides were extracted
using 5 to 10 volumes (w/v) of ethyl acetate at room temperature for
1 hour. The extracts were centrifuged twice (13,000 rpm for 20 min),
and supernatants were collected for GC-­MS analysis. The condi-
tions of GC-­ MS analyses for phthalide mainly followed other
studies (99).
GC-­MS analysis was carried out on an Agilent 7890B GC ma-
chine (Agilent Technologies, Waldbronn, USA) with an Agilent
7000C mass selective detector at 70 eV and a helium flow of 1.0 ml
min−1. The sample of 1 μl was injected at a split ratio 50:1 and ana-
lyzed on an Agilent HP-­5MS column. The column temperature was
set at 50°C for injection and then programmed at 5°C/min to
250°C. The spectrometers were operated in electron-­impact mode,
and the phthalides were analyzed with the selected reaction moni-
toring mode. The inlet and ionization source temperatures were set
at 280° and 300°C, respectively. The mass spectrums of (S)-­NBP,
butylidenephthalide, senkyunolide A, and ligustilide are presented
in fig. S28.
Crude protein activity assay
The crude protein activity assay mainly followed the protocol of
Vanholme et al. (100). Crude protein from ~200 mg of ground rhi-
zome tissue of L. chuanxiong was extracted by 1 ml of protein extrac-
tion buffer on ice for 1 hour (inverting the tubes every 5 to 10 min).
We then centrifuged the extract at 14,000 rpm for 10 min at 4°C and
transferred the supernatant to the other precooled tube. Protein
concentration of the CPE was quantified by Pierce 660-­nm Protein
Assay Reagent (Thermo Fisher Scientific). At last, we mixed 30 μg of
CPE with 40 μM of each target phthalide [(S)-­NBP, butylidene-
phthalide, senkyunolide A, and ligustilide] and subsequently de-
tected signals for the other three by GC-­MS.
Genome-­wide mining for P4,5Ds
We initially excluded those genes with an FPKM ≤20 and selected
those significantly up-­regulated genes in the rhizome [log2(fold
change) > 1, P < 0.05, FDR < 0.05] compared with leaf, stem, and
petiole using DESeq2 (96). We then calculated Pearson correlation
coefficients between contents of (S)-­NBP (also butylidenephthalide)
and the FPKM of genes in rhizome using the R package psych
13 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
(https://rdocumentation.org/packages/psych/versions/2.3.3). We selected
those genes with r > 0.95 and P < 0.05.
HMMER3 (v3.3.2) (101) was used to search 2OGDs and CYPs
with an e-­value of 1 × 10−6. HMMER profile PF00067 was used for
CYPs search, and PF14226 and PF03171 were used to search genes
in DOXC family (2OGD superfamily). The possible pseudogenes
(length of predicted CYPs < 300 amino acids and 2OGDs < 200
amino acids) were discarded. Gene structures of all candidate
2OGDs and CYPs were manually adjusted using IGV-­GSAman
(https://gitee.com/CJchen/IGV-­sRNA).
At last, we constructed ML phylogenies using RAxML-­ NG
(v1.1.0) (93) for candidate CYPs and 2OGDs. The protein alignment
was obtained using MUSCLE (v5.1.linux64) (90). The best-­fitting
amino acid substitution model was selected as LG+I+G4+F for
both CYPs and 2OGDs by ModelTest-­NG (v0.1.7) (92). For each
2OGD (DOXC family) clade or CYP family, we only selected one to
three highly expressed genes for cloning.
Cloning of candidate P4,5D genes
All biosynthetic genes tested in the present study were cloned from
the cDNA of L. chuanxiong. CDS for each gene was amplified via
PCR using the KOD One PCR master Mix (Toyobo). Oligonucle-
otide primers were designed to target the 5′ and 3′ homologous re-
gions for subsequent homologous recombination assembly into the
appropriate expression plasmid. Amplicon products of PCR were
analyzed by gel electrophoresis on a 1% (w/v) agarose gel, and the
positive bands were excised and purified using the Omega Gel DNA
Recovery Kit (Omega Research). A list of all primers used for CDS
cloning in this study is provided in table S16.
Functional characterization for candidate 2OGDs in
E. coil system
Candidate 2OGDs were cloned into the pMAL vector, a bacterial ex-
pression vector based on maltose-­binding protein tag, following the
manufacturer’s handbook. The purified plasmids were transformed into
Rosetta (DE3) strains of E. coli using heat shock transformation. Trans-
formed cells were selected overnight on Lysogeny broth (LB) plates with
ampicillin at 37°C. The positive colonies were screened by colony PCR
and further cultured in 200 ml of LB containing ampicillin at 200 rpm
and 37°C until the optical density at 600 nm (OD600) absorbance reached
0.6 to 0.8, and, then, 0.5 mM isopropyl-­ β-­​d-­thiogalactopyranoside
(MedChemExpress) was added to induce protein expression. For in vivo
enzyme activity characterization through feeding assay, the strains were
fed with 50 μM senkyunolide A or ligustilide and further cultured at
37°C for 4 hours under 200 rpm. At last, the product was extracted with
threefold volume of ethyl acetate. For in vitro enzyme activity character-
ization, the recombinant proteins were purified using the NEBExpress
Amylose Resin (New England Biolabs). Assay for the purified enzyme
activity was performed in a 200-­μl assay buffer [10 mM 2-­oxoglutarate,
10 mM dithiothreitol, 10 mM sodium ascorbate, 5 mM FeSO4, and 25 mM
tris-­HCl buffer (pH 7.5)] containing 30 μg of purified proteins and sen-
kyunolide A/ligustilide at concentrations ranging from 5 to 50 μM. The
reaction was carried out at 25°C for 2 hours and then extracted with
300 μl of ethyl acetate. The negative control of protein was prepared from
E. coli harboring the EV pMAL. GC-­MS analysis was performed to
detect and quantify the reaction products, as described above. Last,
GraphPad Prism software (v.10.0.2, GraphPad Software LLC) was used
to plot a Michaelis-­ Menten model curve and calculate the kinetic
parameters.
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
Functional characterization for candidate CYPs in tobacco
and yeast system
For in vivo enzyme activity characterization through feeding assay in
N. benthamiana, candidate CYPs were first cloned into the pSuper1300
vector and transformed into the GV3101 strains of Agrobacterium elec-
trocompetent cells. The GV3101strains were cultured overnight in LB
medium with kanamycin (50 μg/ml) at 28°C and 180 rpm. Then, 500 μl
of Agrobacterium suspension was transferred into 30 ml of fresh LB me-
dium containing rifampicin (25 μg/ml) and kanamycin (50 μg/ml) and
cultured overnight. Agrobacterium cells were collected by centrifugation
at 4000 rpm for 10 min and resuspended with induction medium
[100 μM acetosyringone, 10 mM MES, and 10 mM MgCl2 (pH 5.6)] for
1 to 2 hours at room temperature. The OD600 absorbance of each strain
resuspension was determined and diluted with induction medium to
~0.3. Each diluted strain was then infiltrated into the abaxial side of
three leaves from different 4-­to 5-­week N. benthamiana seedlings using
a 1-­ml needleless syringe. To test the enzymatic activity, senkyunolide A
or ligustilide (not present in N. benthamiana) solution (100 μM sub-
strate in deionized water) was co-­infiltrated into the Agrobacterium-­
infiltrated leaves after 3 days. The infiltrated leaf region with substrate
was marked and subsequently taken for phthalide extraction after 1 day.
Infiltrations for all treatments regarding different genes and substrates
were conducted in triplicate.
The in vitro enzymatic function characterization of candidate CYPs
was performed by heterologously expression in yeast (S. cerevisiae), mi-
crosomal protein extraction, and enzymatic activity assay. Candidate
CYPs were cloned into the pESC-­Ura vector and transformed into the
WAT11 yeast strain. WAT11 transformed with empty pESC-­Ura was
used as negative control. A single positive colony was initially grown in
10 ml of SD-­Ura liquid medium for about 48 hours at 28°C and 200 rpm.
The culture was then used to inoculate 250 ml of fresh SD-­Ura liquid
medium at 28°C and 200 rpm until the OD600 absorbance reached 1.0 to
1.2. Cells were then collected and washed twice with sterile water. Pro-
tein expression was then induced with yeast extract peptone dextrose
medium with 2% galactose for 16 hour at 28°C and 200 rpm. Then, the
culture was immediately used for microsomal protein isolation follow-
ing previous instructions (102). Enzyme reactions with CYP-­enriched
microsomal proteins were performed in tris-­HCl buffer (pH 7.5) con-
taining 5 μg of microsomal protein, 500 μM NADPH (reduced form of
nicotinamide adenine dinucleotide phosphate), 50 μM senkyunolide A
or ligustilide, and NADPH Regeneration System (Promega) in a total
reaction volume of 200 μl. Following the addition of all components, re-
actions were incubated at 25°C for 16 hours and then extracted with
300 μl of ethyl acetate. The reaction products were detected through
GC-­MS, as described above.
Supplementary Materials
This PDF file includes:
Figs. S1 to S28
Legends for tables S1 to S16
References
Other Supplementary Material for this manuscript includes the following:
Tables S1 to S16
REFERENCES AND NOTES
1. GBD 2019 Stroke Collaborators, Global, regional, and national burden of stroke and its
risk factors, 1990–2019: A systematic analysis for the global burden of disease study
2019. Lancet Neurol. 20, 795–820 (2021).
14 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
2. E. Berge, W. Whiteley, H. Audebert, G. M. De Marchis, A. C. Fonseca, C. Padiglioni,
N. Pérez de la Ossa, D. Strbian, G. Tsivgoulis, G. Turc, European Stroke Organisation (ESO)
guidelines on intravenous thrombolysis for acute ischaemic stroke. Eur. Stroke J. 6, I–LXII
(2021).
3. M. X. Dong, Q. C. Hu, P. Shen, J. X. Pan, Y. D. Wei, Y. Y. Liu, Y. F. Ren, Z. H. Liang, H. Y. Wang,
L. B. Zhao, P. Xie, Recombinant tissue plasminogen activator induces neurological side
effects independent on thrombolysis in mechanical animal models of focal cerebral
infarction: A systematic review and meta-­analysis. PLOS ONE 11, e0158848 (2016).
4. Y. P. Feng, D. Hu, L. Y. Zhang, Effect of dl-­butylphthalide (NBP) on mouse brain energy
metabolism in complete brain ischemia induced by decapitation. Acta Pharm. Sin. 30,
741–744 (1995).
5. Y. Peng, X. Zeng, Y. Feng, X. Wang, Antiplatelet and antithrombotic activity of
l-­3-­n-­butylphthalide in rats. J. Cardiovasc. Pharmacol. 43, 876–881 (2004).
6. C.-­L. Liu, S.-­J. Liao, J.-­S. Zeng, J.-­W. Lin, C.-­X. Li, L.-­C. Xie, X.-­G. Shi, R.-­X. Huang,
dl-­3n-­Butylphthalide prevents stroke via improvement of cerebral microvessels in RHRSP.
J. Neurol. Sci. 260, 106–113 (2007).
7. A. Wang, B. Jia, X. Zhang, X. Huo, J. Chen, L. Gui, Y. Cai, Z. Guo, Y. Han, Z. Peng, P. Jing,
Y. Chen, Y. Liu, Y. Yang, F. Wang, Z. Sun, T. Li, H. Sun, H. Yuan, H. Shao, L. Gao, P. Zhang,
F. Wang, X. Cao, W. Shi, C. Li, J. Yang, H. Zhang, F. Wang, J. Deng, Y. Liu, W. Deng, C. Song,
H. Chen, L. He, H. Zhao, X. Li, H. Yang, Z. Zhou, Y. Wang, Z. Miao, BAST Investigators,
Efficacy and safety of butylphthalide in patients with acute ischemic stroke: A
randomized clinical trial. JAMA Neurol. 80, 851–859 (2023).
8. X.-­X. Diao, K. Zhong, X.-­L. Li, D.-­F. Zhong, X.-­Y. Chen, Isomer-­selective distribution of
3-­​n-­butylphthalide (NBP) hydroxylated metabolites, 3-­hydroxy-­NBP and 10-­hydroxy-­NBP,
across the rat blood-­brain barrier. Acta Pharm. Sin. 36, 1520–1527 (2015).
9. L. Yang, H. Li, Y. Wu, H. Zhang, J. Du, Y. Chen, Efficacy of sequential N-­butylphthalide
therapy on psychiatric and behavioral functions in acute ischemic stroke. Medicine 100,
e27860 (2021).
10. C. Yan, Y. Feng, Effects of d-­3-­​n-­butylphthalide and l-­3-­​n-­butylphthalide on extracellular
NO level and intracellular cGMP level in primary cultured rat cortical neurons. Acta
Pharm. Sin. 33, 418–423 (1998).
11. Q. Chang, X.-­L. Wang, Effects of chiral 3-­​n-­butylphthalide on apoptosis induced by
transient focal cerebral ischemia in rats. Acta Pharm. Sin. 24, 796–804 (2003).
12. X. C. Ji, W. H. Zhao, D. X. Cao, Q. Q. Shi, X. L. Wang, Novel neuroprotectant chiral
3-­​n-­butylphthalide inhibits tandem-­pore-­domain potassium channel TREK-­1. Acta
Pharm. Sin. 32, 182–187 (2011).
13. Y. Peng, Y. Hu, S. Xu, P. Li, J. Li, L. Lu, H. Yang, N. Feng, L. Wang, X. Wang, L-­3-­n-­
butylphthalide reduces tau phosphorylation and improves cognitive deficits in AβPP/
PS1-­Alzheimer’s transgenic mice. J. Alzheimers Dis. 29, 379–391 (2012).
14. J. Marco-­Contelles, Y. Zhang, From seeds of Apium graveolens Linn. To a cerebral
ischemia medicine: The long journey of 3-­​n-­butylphthalide. J. Med. Chem. 63,
12485–12510 (2020).
15. A. León, M. Del-­Angel, J. Luisa Avila, G. Delgado, “Phthalide: Distribution in nature,
chemical reactivity, synthesis, and biological activity” in Progress in the Chemistry of
Organic Natural Products, A. D. Kinghorn, H. Falk, S. Gibbons, J. Kobayashi, Eds. (Springer,
2017), vol. 104, pp. 127–246.
16. Z. Chen, C. Zhang, F. Gao, Q. Fu, C. Fu, Y. He, J. Zhang, A systematic review on the rhizome
of Ligusticum chuanxiong Hort. (Chuanxiong). Food Chem. Toxicol. 119, 309–325 (2018).
17. J. J. Beck, S.-­C. Chou, The structural diversity of phthalides from the Apiaceae. J. Nat. Prod.
70, 891–900 (2007).
18. A. Cravens, J. Payne, C. D. Smolke, Synthetic biology strategies for microbial biosynthesis
of plant natural products. Nat. Commun. 10, 2142 (2019).
19. H. Mitsuhashi, M. Nomura, Studies on the constituents of Umbelliferae plants. XII.
Biogenesis of 3-­butylphthalide. Chem. Pharm. Bull. 14, 777–778 (1966).
20. M. Kobayashi, H. Mitsuhashi, Studies on the constituents of Umbelliferae plantx. XVII.
Structures of three new ligustilide derivatives from Ligusticum wallichii. Chem. Pharm.
Bull. 35, 4789–4792 (1987).
21. W. M. Feng, P. Liu, H. Yan, G. Yu, S. Zhang, S. Jiang, E. X. Shang, D. W. Qian, J. A. Duan,
Investigation of enzymes in the phthalide biosynthetic pathway in Angelica sinensis
using integrative metabolite profiles and transcriptome analysis. Front. Plant Sci. 13,
928760 (2022).
22. T. Naake, H. A. Maeda, S. Proost, T. Tohge, A. R. Fernie, Kingdom-­wide analysis of the
evolution of the plant type III polyketide synthase superfamily. Plant Physiol. 185,
857–875 (2021).
23. G. Lin, S. S. K. Chan, H. S. Chung, S. L. Li, “Chemistry and biological activity of natural
occurring phthalides” in Studies in Natural Products Chemistry, A. UrRahman, Ed. (Elsevier,
2005), vol. 32, pp. 611–669.
24. X. Wei, Y. Zeng, C. Sun, F. Meng, Y. Wang, Recent advances in natural phthalides:
Distribution, chemistry, and biological activities. Fitoterapia 160, 105223 (2022).
25. H. Yan, Y. Zhou, F. Tang, C. Wang, J. Wu, C. Hu, X. Xie, C. Peng, Y. Tan, A comprehensive
investigation on the chemical diversity and efficacy of different parts of Ligusticum
chuanxiong. Food Funct. 13, 1092–1107 (2022).
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
26. S. Huang, J. Chen, W. Li, S. Song, X. Li, H. Yu, F. Long, R. Chen, X. Bao, K. Chan, G. Lu,
Comparison of volatile compositions among four related Ligusticum chuanxiong herbs
by HS-­SPME-­GC-­MS. Processes 11, 196 (2023).
27. F. Shan, J. Hao, Herbal textual research on origin and development of Chuanxiong. China
Journal of Chinese Materia Medica 36, 2306–2310 (2011).
28. S. H. Qiu, Y. Q. Zeng, K. Y. Pan, Y. C. Tang, J. M. Xu, On the nomenclature of the Chinese
plant drug “Chuanxiong”. Acta Phytotaxonomica Sinica 17, 101–103 (1979).
29. F.-­T. Pu, A revision of the genus Ligusticum (Umbelliferae) in China. J. Syst. Evol. 29,
385–393 (1991).
30. C. Yuan, X. Sha, M. Xiong, W. Zhong, Y. Wei, M. Li, S. Tao, F. Mou, F. Peng, C. Zhang,
Uncovering dynamic evolution in the plastid genome of seven Ligusticum species
provides insights into species discrimination and phylogenetic implications. Sci. Rep. 11,
988 (2021).
31. J. Zhou, Y.-­Z. Gao, J. Wei, Z.-­W. Liu, S. R. Downie, Molecular phylogenetics of Ligusticum
(Apiaceae) based on nrDNA ITS sequences: Rampant polyphyly, placement of the
Chinese endemic species, and a much-­reduced circumscription of the genus. Int. J. Plant
Sci. 181, 306–323 (2020).
32. T. Ren, Z.-­X. Li, D.-­F. Xie, L.-­J. Gui, C. Peng, J. Wen, X.-­J. He, Plastomes of eight Ligusticum
species: Characterization, genome evolution, and phylogenetic relationships. BMC Plant
Biol. 20, 519 (2020).
33. Y. Wang, W. Xu, J. He, T. Tian, Making chromosomes of Rhizoma Chuanxiong and
analysing the karyotype. J. Chengdu Univ. 23, 11–13 (2004).
34. Y. Shen, W. Li, Y. Zeng, Z. Li, Y. Chen, J. Zhang, H. Zhao, L. Feng, D. Ma, X. Mo, P. Ouyang,
L. Huang, Z. Wang, Y. Jiao, H.-­B. Wang, Chromosome-­level and haplotype-­resolved
genome provides insight into the tetraploid hybrid origin of patchouli. Nat. Commun. 13,
3511 (2022).
35. K.-­H. Jia, Z.-­X. Wang, L. Wang, G.-­Y. Li, W. Zhang, X.-­L. Wang, F.-­J. Xu, S.-­Q. Jiao, S.-­S. Zhou,
H. Liu, Y. Ma, G. Bi, W. Zhao, Y. A. El-­Kassaby, I. Porth, G. Li, R.-­G. Zhang, J.-­F. Mao,
SubPhaser: A robust allopolyploid subgenome phasing method based on subgenome-­
specific k-­mers. New Phytol. 235, 801–809 (2022).
36. C.-­Q. Mao, X.-­F. Sha, J. Huang, S. Tao, F. Peng, Q. Li, C. Zhang, C. Yuan, Genome survey and
characteristic analysis of Ligusticum chuanxiong. Chin. Tradit. Herb. Drug 54, 907–914
(2023).
37. X. Han, C. Li, S. Sun, J. Ji, B. Nie, G. Maker, Y. Ren, L. Wang, The chromosome-­level genome
of female ginseng (Angelica sinensis) provides insights into molecular mechanisms and
evolution of coumarin biosynthesis. Plant J. 112, 1224–1237 (2022).
38. W. Hu, C. Ji, H. Shi, Z. Liang, Z. Ding, J. Ye, W. Ou, G. Zhou, W. Tie, Y. Yan, J. Yang, X. Yang,
Y. Wei, Z. Jin, J. Xie, M. Peng, W. Wang, A. Guo, B. Xu, J. Guo, S. Chen, L. Ma, M. Wang,
Z. Yang, X. Li, R. Li, S. Guo, X. Xiao, Z. Wan, F. An, J. Zhang, Q. Leng, Y. Li, R. Ming, K. Li,
Allele-­defined genome reveals biallelic differentiation during cassava evolution. Mol.
Plant 14, 851–854 (2021).
39. X. Sun, C. Jiao, H. Schwaninger, C. T. Chao, Y. Ma, N. Duan, A. Khan, S. Ban, K. Xu, L. Cheng,
G. Y. Zhong, Z. Fei, Phased diploid genome assemblies and pan-­genomes provide
insights into the genetic history of apple domestication. Nat. Genet. 52, 1423–1432
(2020).
40. Y. Zhang, Q. Shen, L. Leng, D. Zhang, S. Chen, Y. Shi, Z. Ning, S. Chen, Incipient
diploidization of the medicinal plant Perilla within 10,000 years. Nat. Commun. 12, 5508
(2021).
41. J. Miao, Q. Feng, Y. Li, Q. Zhao, C. Zhou, H. Lu, D. Fan, J. Yan, Y. Lu, Q. Tian, W. Li, Q. Weng,
L. Zhang, Y. Zhao, T. Huang, L. Li, X. Huang, T. Sang, B. Han, Chromosome-­scale assembly
and analysis of biomass crop Miscanthus lutarioriparius genome. Nat. Commun. 12, 2458
(2021).
42. T. Zhu, L. Wang, F. M. You, J. C. Rodriguez, K. R. Deal, L. Chen, J. Li, S. Chakraborty, B. Balan,
C. Z. Jiang, P. J. Brown, C. A. Leslie, M. K. Aradhya, A. M. Dandekar, P. E. McGuire,
D. Kluepfel, J. Dvorak, M. C. Luo, Sequencing a Juglans regia × J. microcarpa hybrid yields
high-­quality genome assemblies of parental species. Hortic. Res. 6, 55 (2019).
43. L. Shao, F. Xing, C. Xu, Q. Zhang, J. Che, X. Wang, J. Song, X. Li, J. Xiao, L.-­L. Chen,
Y. Ouyang, Q. Zhang, Patterns of genome-­wide allele-­specific expression in hybrid rice
and the implications on the genetic basis of heterosis. Proc. Natl. Acad. Sci. U.S.A. 116,
5653–5658 (2019).
44. P. Wang, M. Gu, X. Yu, S. Shao, J. Du, Y. Wang, F. Wang, S. Chen, Z. Liao, N. Ye, X. Zhang,
Allele-­specific expression and chromatin accessibility contribute to heterosis in tea
plants (Camellia sinensis). Plant J. 112, 1194–1211 (2022).
45. A. Zuo, L. Wang, H. Xiao, Study on degradation products of senkyunolide A and
senkyunolide I. Chin. Tradit. Herb. Drug 43, 2127–2131 (2012).
46. A. Schinkovitz, S. M. Pro, M. Main, S.-­N. Chen, B. U. Jaki, D. C. Lankin, G. F. Pauli, Dynamic
nature of the ligustilide complex. J. Nat. Prod. 71, 1604–1611 (2008).
47. K. Duric, Y. Liu, S.-­N. Chen, D. C. Lankin, D. Nikolic, J. B. McAlpine, J. B. Friesen, G. F. Pauli,
Studying mass balance and the stability of (Z)-­ligustilide from Angelica sinensis helps to
bridge a botanical instability–bioactivity chasm. J. Nat. Prod. 82, 2400–2408 (2019).
48. A. Rai, K. Saito, M. Yamazaki, Integrated omics analysis of specialized metabolism in
medicinal plants. Plant J. 90, 764–787 (2017).
15 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
49. C. C. Hansen, D. R. Nelson, B. L. Moller, D. Werck-­Reichhart, Plant cytochrome P450
plasticity and evolution. Mol. Plant 14, 1244–1265 (2021).
50. Y. Kawai, E. Ono, M. Mizutani, Evolution and diversity of the 2-­oxoglutarate-­dependent
dioxygenase superfamily in plants. Plant J. 78, 328–343 (2014).
51. J. E. Jeon, J.-­G. Kim, C. R. Fischer, N. Mehta, C. Dufour-­Schroif, K. Wemmer, M. B. Mudgett,
E. Sattely, A pathogen-­responsive gene cluster for highly modified fatty acids in tomato.
Cell 180, 176–187.e19 (2020).
52. H. Minami, E. Dubouzet, K. Iwasa, F. Sato, Functional analysis of norcoclaurine synthase in
Coptis japonica. J. Biol. Chem. 282, 6274–6282 (2007).
53. J. M. Hagel, P. J. Facchini, Dioxygenases catalyze the O-­demethylation steps of morphine
biosynthesis in opium poppy. Nat. Chem. Biol. 6, 273–275 (2010).
54. Y. Byeon, K. Back, Molecular cloning of melatonin 2-­hydroxylase responsible for
2-­hydroxymelatonin production in rice (Oryza sativa). J. Pineal Res. 58, 343–351 (2015).
55. B. R. Lichman, M. O. Kamileen, G. R. Titchiner, G. Saalbach, C. E. M. Stevenson,
D. M. Lawson, S. E. O’Connor, Uncoupled activation and cyclization in catmint reductive
terpenoid biosynthesis. Nat. Chem. Biol. 15, 71–79 (2019).
56. F. Zhang, F. Qiu, J. Zeng, Z. Xu, Y. Tang, T. Zhao, Y. Gou, F. Su, S. Wang, X. Sun, Z. Xue,
W. Wang, C. Yang, L. Zeng, X. Lan, M. Chen, J. Zhou, Z. Liao, Revealing evolution of
tropane alkaloid biosynthesis by analyzing two genomes in the Solanaceae family. Nat.
Commun. 14, 1446 (2023).
57. S. Helinck, H. E. Spinnler, S. Parayre, M. Dame-­Cahagne, P. Bonnarme, Enzymatic versus
spontaneous S-­methyl thioester synthesis in Geotrichum candidum. FEMS Microbiol. Lett.
193, 237–241 (2000).
58. E. O. Fukushima, H. Seki, K. Ohyama, E. Ono, N. Umemoto, M. Mizutani, K. Saito,
T. Muranaka, CYP716A subfamily members are multifunctional oxidases in triterpenoid
biosynthesis. Plant Cell Physiol. 52, 2050–2061 (2011).
59. Y. L. Xu, L. Li, K. Wu, A. J. Peeters, D. A. Gage, J. A. Zeevaart, The GA5 locus of Arabidopsis
thaliana encodes a multifunctional gibberellin 20-­oxidase: Molecular cloning and
functional expression. Proc. Natl. Acad. Sci. U.S.A. 92, 6640–6644 (1995).
60. S. C. Farrow, P. J. Facchini, Functional diversity of 2-­oxoglutarate/Fe(II)-­dependent
dioxygenases in plant metabolism. Front. Plant Sci. 5, 524 (2014).
61. Z. Wang, M. Kang, J. Li, Z. Zhang, Y. Wang, C. Chen, Y. Yang, J. Liu, Genomic evidence for
homoploid hybrid speciation between ancestors of two different genera. Nat. Commun.
13, 1987 (2022).
62. Z. Wang, Y. Jiang, H. Bi, Z. Lu, Y. Ma, X. Yang, N. Chen, B. Tian, B. Liu, X. Mao, T. Ma,
S. P. DiFazio, Q. Hu, R. J. Abbott, J. Liu, Hybrid speciation via inheritance of alternate
alleles of parental isolating genes. Mol. Plant 14, 208–222 (2021).
63. S. Lamichhaney, F. Han, M. T. Webster, L. Andersson, B. R. Grant, P. R. Grant, Rapid hybrid
speciation in Darwin's finches. Science 359, 224–228 (2018).
64. J. D. Hollister, S. Greiner, M. T. J. Johnson, S. I. Wright, Hybridization and a loss of sex
shape genome-­wide diversity and the origin of species in the evening primroses
(Oenothera, Onagraceae). New Phytol. 224, 1372–1380 (2019).
65. S. Xu, D. J. Innes, M. Lynch, M. E. Cristescu, The role of hybridization in the origin and
spread of asexuality in Daphnia. Mol. Ecol. 22, 4549–4561 (2013).
66. M. Pellino, D. Hojsgaard, T. Schmutzer, U. Scholz, E. Hörandl, H. Vogel, T. F. Sharbel,
Asexual genome evolution in the apomictic Ranunculus auricomus complex: Examining
the effects of hybridization and mutation accumulation. Mol. Ecol. 22, 5908–5921 (2013).
67. V. A. Lukhtanov, V. Dinca, M. Friberg, R. Vila, C. Wiklund, Incomplete sterility of
chromosomal hybrids: Implications for karyotype evolution and homoploid hybrid
speciation. Front. Genet. 11, 583827 (2020).
68. E. S. Tvedte, J. M. Logsdon Jr., A. A. Forbes, Sex loss in insects: Causes of asexuality and
consequences for genomes. Curr. Opin. Insect Sci. 31, 77–83 (2019).
69. P. Wang, M. Gu, S. Shao, X. Chen, B. Hou, N. Ye, X. Zhang, Changes in non-­volatile and
volatile metabolites associated with heterosis in tea plants (Camellia sinensis). J. Agric.
Food Chem. 70, 3067–3078 (2022).
70. F. Palumbo, G. Galla, N. Vitulo, G. Barcaccia, First draft genome sequencing of fennel
(Foeniculum vulgare Mill.): Identification of simple sequence repeats and their application
in marker-­assisted breeding. Mol. Breed. 38, 122 (2018).
71. G. Marcais, C. Kingsford, A fast, lock-­free approach for efficient parallel counting of
occurrences of k-­mers. Bioinformatics 27, 764–770 (2011).
72. T. R. Ranallo-­Benavidez, K. S. Jaron, M. C. Schatz, GenomeScope 2.0 and Smudgeplot for
reference-­free profiling of polyploid genomes. Nat. Commun. 11, 1432 (2020).
73. H. Cheng, E. D. Jarvis, O. Fedrigo, K. P. Koepfli, L. Urban, N. J. Gemmell, H. Li, Haplotype-­
resolved assembly of diploid genomes without parental data. Nat. Biotechnol. 40,
1332–1335 (2022).
74. O. Dudchenko, S. S. Batra, A. D. Omer, S. K. Nyquist, M. Hoeger, N. C. Durand,
M. S. Shamim, I. Machol, E. S. Lander, A. P. Aiden, E. L. Aiden, De novo assembly of the Aedes
aegypti genome using Hi-­C yields chromosome-­length scaffolds. Science 356, 92–95 (2017).
75. M. Manni, M. R. Berkeley, M. Seppey, F. A. Simao, E. M. Zdobnov, BUSCO update: Novel
and streamlined workflows along with broader and deeper phylogenetic coverage for
scoring of eukaryotic, prokaryotic, and viral genomes. Mol. Biol. Evol. 38, 4647–4654
(2021).
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
76. S. Ou, W. Su, Y. Liao, K. Chougule, J. R. A. Agda, A. J. Hellinga, C. S. B. Lugo, T. A. Elliott,
D. Ware, T. Peterson, N. Jiang, C. N. Hirsch, M. B. Hufford, Benchmarking transposable
element annotation methods for creation of a streamlined, comprehensive pipeline.
Genome Biol. 20, 275 (2019).
77. A. R. Quinlan, I. M. Hall, BEDTools: A flexible suite of utilities for comparing genomic
features. Bioinformatics 26, 841–842 (2010).
78. D. Kim, J. M. Paggi, C. Park, C. Bennett, S. L. Salzberg, Graph-­based genome alignment
and genotyping with HISAT2 and HISAT-­genotype. Nat. Biotechnol. 37, 907–915
(2019).
79. M. Pertea, G. M. Pertea, C. M. Antonescu, T. C. Chang, J. T. Mendell, S. L. Salzberg, StringTie
enables improved reconstruction of a transcriptome from RNA-­seq reads. Nat.
Biotechnol. 33, 290–295 (2015).
80. B. J. Haas, A. L. Delcher, S. M. Mount, J. R. Wortman, R. K. Smith Jr., L. I. Hannick, R. Maiti,
C. M. Ronning, D. B. Rusch, C. D. Town, S. L. Salzberg, O. White, Improving the Arabidopsis
genome annotation using maximal transcript alignment assemblies. Nucleic Acids Res.
31, 5654–5666 (2003).
81. M. S. Campbell, C. Holt, B. Moore, M. Yandell, Genome annotation and curation using
MAKER and MAKER-­P. Curr. Protoc. Bioinformatics 48, 4.11.1–4.11.39 (2014).
82. P. Jones, D. Binns, H.-­Y. Chang, M. Fraser, W. Z. Li, C. McAnulla, H. McWilliam, J. Maslen,
A. Mitchell, G. Nuka, S. Pesseat, A. F. Quinn, A. Sangrador-­Vegas, M. Scheremetjew,
S.-­Y. Yong, R. Lopez, S. Hunter, InterProScan 5: Genome-­scale protein function
classification. Bioinformatics 30, 1236–1240 (2014).
83. J. Huerta-­Cepas, K. Forslund, L. P. Coelho, D. Szklarczyk, L. J. Jensen, C. von Mering,
P. Bork, Fast genome-­wide functional annotation through orthology assignment by
eggNOG-­Mapper. Mol. Biol. Evol. 34, 2115–2122 (2017).
84. H. Tang, J. E. Bowers, X. Wang, R. Ming, M. Alam, A. H. Paterson, Synteny and collinearity
in plant genomes. Science 320, 486–488 (2008).
85. M. A. DePristo, E. Banks, R. Poplin, K. V. Garimella, J. R. Maguire, C. Hartl, A. A. Philippakis,
G. del Angel, M. A. Rivas, M. Hanna, A. McKenna, T. J. Fennell, A. M. Kernytsky,
A. Y. Sivachenko, K. Cibulskis, S. B. Gabriel, D. Altshuler, M. J. Daly, A framework for
variation discovery and genotyping using next-­generation DNA sequencing data. Nat.
Genet. 43, 491–498 (2011).
86. M. G. Grabherr, B. J. Haas, M. Yassour, J. Z. Levin, D. A. Thompson, I. Amit, X. Adiconis,
L. Fan, R. Raychowdhury, Q. Zeng, Z. Chen, E. Mauceli, N. Hacohen, A. Gnirke, N. Rhind,
F. di Palma, B. W. Birren, C. Nusbaum, K. Lindblad-­Toh, N. Friedman, A. Regev, Full-­length
transcriptome assembly from RNA-­Seq data without a reference genome. Nat.
Biotechnol. 29, 644–652 (2011).
87. C. Camacho, G. Coulouris, V. Avagyan, N. Ma, J. Papadopoulos, K. Bealer, T. L. Madden,
BLAST+: Architecture and applications. BMC Bioinform. 10, 421 (2009).
88. K. Tamura, G. Stecher, S. Kumar, MEGA11: Molecular evolutionary genetics analysis
version 11. Mol. Biol. Evol. 38, 3022–3027 (2021).
89. D. M. Emms, S. Kelly, OrthoFinder: Phylogenetic orthology inference for comparative
genomics. Genome Biol. 20, 238 (2019).
90. R. C. Edgar, MUSCLE: Multiple sequence alignment with high accuracy and high
throughput. Nucleic Acids Res. 32, 1792–1797 (2004).
91. D. Zhang, F. Gao, I. Jakovlić, H. Zou, J. Zhang, W. X. Li, G. T. Wang, PhyloSuite: An
integrated and scalable desktop platform for streamlined molecular sequence data
management and evolutionary phylogenetics studies. Mol. Ecol. Resour. 20, 348–355
(2020).
92. D. Darriba, D. Posada, A. M. Kozlov, A. Stamatakis, B. Morel, T. Flouri, ModelTest-­NG: A new
and scalable tool for the selection of DNA and protein evolutionary models. Mol. Biol.
Evol. 37, 291–294 (2020).
93. A. Stamatakis, RAxML version 8: A tool for phylogenetic analysis and post-­analysis of
large phylogenies. Bioinformatics 30, 1312–1313 (2014).
94. Z. Yang, PAML 4: Phylogenetic analysis by maximum likelihood. Mol. Biol. Evol. 24,
1586–1591 (2007).
95. P. Sun, B. Jiao, Y. Yang, L. Shan, T. Li, X. Li, Z. Xi, X. Wang, J. Liu, WGDI: A user-­friendly toolkit
for evolutionary analyses of whole-­genome duplications and ancestral karyotypes. Mol.
Plant 15, 1841–1851 (2022).
96. F. K. Mendes, D. Vanderpool, B. Fulton, M. W. Hahn, CAFE 5 models variation in
evolutionary rates among gene families. Bioinformatics 36, 5516–5518 (2021).
97. T. Wu, E. Hu, S. Xu, M. Chen, P. Guo, Z. Dai, T. Feng, L. Zhou, W. Tang, L. Zhan, X. Fu, S. Liu,
X. Bo, G. Yu, ClusterProfiler 4.0: A universal enrichment tool for interpreting omics data.
Innovation 2, 100141 (2021).
98. M. I. Love, W. Huber, S. Anders, Moderated estimation of fold change and dispersion for
RNA-­seq data with DESeq2. Genome Biol. 15, 550 (2014).
99. B. Yang, J. Chen, F. S.-­C. Lee, X. Wang, GC-­MS fingerprints for discrimination of Ligusticum
chuanxiong from Angelica. J. Sep. Sci. 31, 3231–3237 (2008).
100. B. Vanholme, I. Cesarino, G. Goeminne, H. Kim, F. Marroni, R. Van Acker, R. Vanholme,
K. Morreel, B. Ivens, S. Pinosio, M. Morgante, J. Ralph, C. Bastien, W. Boerjan, Breeding
with rare defective alleles (BRDA): A natural Populus nigra HCT mutant with modified
lignin as a case study. New Phytol. 198, 765–776 (2013).
16 of 17
S c i e n c e A d v an c e s | R e s e ar c h A r t i c l e
101. J. Mistry, R. D. Finn, S. R. Eddy, A. Bateman, M. Punta, Challenges in homology search:
HMMER3 and convergent evolution of coiled-­coil regions. Nucleic Acids Res. 41, e121
(2013).
102. J. Hu, S. Qiu, F. Wang, Q. Li, C. L. Xiang, P. Di, Z. Wu, R. Jiang, J. Li, Z. Zeng, J. Wang, X. Wang,
Y. Zhang, S. Fang, Y. Qiao, J. Ding, Y. Jiang, Z. Xu, J. Chen, W. Chen, Functional divergence
of CYP76AKs shapes the chemodiversity of abietane-­type diterpenoids in genus Salvia.
Nat. Commun. 14, 4696 (2023).
103. S. J. Gagne, J. M. Stout, E. Liu, Z. Boubakir, S. M. Clark, J. E. Page, Identification of olivetolic
acid cyclase from Cannabis sativa reveals a unique catalytic route to plant polyketides.
Proc. Natl. Acad. Sci. U.S.A. 109, 12811–12816 (2012).
104. F. Taura, M. Iijima, E. Yamanaka, H. Takahashi, H. Kenmoku, H. Saeki, S. Morimoto,
Y. Asakawa, F. Kurosaki, H. Morita, A novel class of plant type III polyketide synthase
involved in orsellinic acid biosynthesis from Rhododendron dauricum. Front. Plant Sci. 7,
1452 (2016).
105. P. Berman, L. A. de Haro, A. Jozwiak, S. Panda, Z. Pinkas, Y. Dong, J. Cveticanin, R. Barbole,
R. Livne, T. Scherf, E. Shimoni, S. Levin-­Zaidman, N. Dezorella, E. Petrovich-­Kopitman,
S. Meir, I. Rogachev, P. D. Sonawane, A. Aharoni, Parallel evolution of cannabinoid
biosynthesis. Nat. Plants 9, 817–831 (2023).
106. Y. Wang, H. Zhang, H. C. Ri, Z. An, X. Wang, J. N. Zhou, D. Zheng, H. Wu, P. Wang, J. Yang,
D. K. Liu, D. Zhang, W. C. Tsai, Z. Xue, Z. Xu, P. Zhang, Z. J. Liu, H. Shen, Y. Li, Deletion and
tandem duplications of biosynthetic genes drive the diversity of triterpenoids in Aralia
elata. Nat. Commun. 13, 2224 (2022).
Acknowledgments: We thank the support from Innovation Program of Chinese Academy of
Agricultural Sciences. We thank Z. Duan for providing seedlings of L. chuanxiong and Z. Guo for
providing seedlings of L. sinense. We thank J. Zhang for providing suggestions on experimental
details of cytogenetic assay. We thank Y. Peng for assistance in GC-­MS analysis. We thank Z. Bao
for help in the early stages of this project. We thank F. Wang, X. Qin, and D. Xie for help in the in
vitro enzyme activity assay. We also thank D. Nelson of P450 nomenclature committee for
Nie et al., Sci. Adv. 10, eadj6547 (2024) 7 February 2024
naming the CYPs. Funding: This work was supported by the National Key Research and
Development Program of China, grant 2023YFA0915800; National Natural Science Foundation
of China, grants 32300223, 32070242, and 82373837; Shenzhen Fundamental Research
Program, grant 20220817165436004; Shenzhen Science and Technology Program, grant
KQTD2016113010482651; Key Project at Central Government Level (The ability establishment
of sustainable use for valuable Chinese medicine resources), grant 2060302; Special Funds for
Science Technology Innovation and Industrial Development of Shenzhen Dapeng New
District, grants RC201901-­05 and PT201901-­19; China Postdoctoral Science Foundation, grant
2020 M672904; Basic and Applied Basic Research Fund of Guangdong, grant
2020A1515110912; Science, Technology, and Innovation Commission of Shenzhen
Municipality of China, grant ZDSYS20200811142605017; and Sichuan Science and Technology
Program, grant 2021JDJQ0040. Author contributions: Conceptualization: L.W., B.N., X.C., Z.H.,
and M.G. Plant material collection: B.N. Bioinformatics analyses: Z.H., B.N., and C.L. FISH
analysis: B.N. and Z.H. Cytogenetic assay: J.J. and L.Z. Gene cloning and characterization: X.C.,
W.S., and M.G. GC-­MS analyses: X.C., W.S., M.G., and S.Y. Visualization: B.N., X.C., Z.H., and M.G.
Supervision: L.W., W.L., and B.N. Writing—original draft: B.N., X.C., Z.H., and M.G. Writing—
review and editing: L.W., B.N., M.G., W.L., C.L., H.L., Y.T., W.Z., and P.F. Competing interests: The
authors declare that they have no competing interests. Data and materials availability: All
data needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials. The raw data of genome and transcriptome sequencing of L.
chuanxiong and L. sinense have been deposited to the Genome Sequence Archive at the
National Genomics Data Center (NGDC) under BioProject no. PRJCA017485. L. chuanxiong
genome files are available at figshare (https://figshare.com/projects/Ligusticum_chuanxiong_
genome/169136).
Submitted 9 July 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adj6547
17 of 17
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

STRUCTURAL BIOLOGY Copyright © 2024 The

Authors, some rights
Three-­step docking by WIPI2, ATG16L1, and ATG3 reserved; exclusive
licensee American
delivers LC3 to the phagophore Association for the
Advancement of

Shanlin Rao1,2, Marvin Skulsuppaisarn3,4, Lisa M. Strong2,5,6, Xuefeng Ren2,5,6,

Science. No claim to
original U.S.
Michael Lazarou2,3,4,7*, James H. Hurley2,5,6,8*, Gerhard Hummer1,2,9* Government Works.
Distributed under a
The covalent attachment of ubiquitin-­like LC3 proteins (microtubule-­associated proteins 1A/1B light chain 3) Creative Commons
prepares the autophagic membrane for cargo recruitment. We resolve key steps in LC3 lipidation by combining Attribution License 4.0
molecular dynamics simulations and experiments in vitro and in cellulo. We show how the E3-­like ligaseautophagy-­ (CC BY).
related 12 (ATG12)–ATG5-­ATG16L1 in complex with the E2-­like conjugase ATG3 docks LC3 onto the membrane in
three steps by (i) the phosphatidylinositol 3-­phosphate effector protein WD repeat domain phosphoinositide-­
interacting protein 2 (WIPI2), (ii) helix α2 of ATG16L1, and (iii) a membrane-­interacting surface of ATG3. Phospha-
tidylethanolamine (PE) lipids concentrate in a region around the thioester bond between ATG3 and LC3,
highlighting residues with a possible role in the catalytic transfer of LC3 to PE, including two conserved histi-
dines. In a near-­complete pathway from the initial membrane recruitment to the LC3 lipidation reaction, the
three-­step targeting of the ATG12–ATG5-­ATG16L1 machinery establishes a high level of regulatory control.

INTRODUCTION biology of protein ubiquitylation, and the related Neddylation,

Eukaryotic cells use autophagy for the wholesale degradation of bulk
cytosol and bulky substrates, including intracellular pathogens, pro-
tein aggregates, and mitochondria (1). Autophagy of the latter is
referred to as mitophagy (2, 3). Defects in mitophagy downstream of
the E3 ubiquitin (Ub) ligase Parkin and the Ub kinase PTEN-­induced
kinase 1 (PINK1) are implicated in familial Parkinson’s disease
(4). Autophagy is critical for cell homeostasis across a vast range of
physiological conditions, and its defects contribute to essentially all
the major late-­onset neurodegenerative diseases, cancer, and other
diseases (5). The covalent conjugation of the Ub-­like autophagy-­
related 8 (ATG8) proteins to the membrane lipid phosphatidyletha-
nolamine (PE) is a hallmark of autophagy (6). Atg8 is the sole and
founding member of this family in yeast, and it has six orthologs
in humans, LC3A/B/C, GABARAP, and GABARAPL1/2 (7). ATG8
family proteins bind to short motifs known in humans as LC3-­
interacting regions (LIRs). LIR motifs are found throughout the
machinery of autophagy, where their interactions facilitate cargo
sequestration in selective autophagy (8), autophagosome-­lysosome
fusion, and autophagic membrane breakdown (9, 10), and, indeed,
have some role in most steps in autophagy.
ATG8s are conjugated to membrane PE through a pathway
that has both analogies and differences with protein ubiquity-
lation. Ub and Ub-­like proteins are conjugated to proteins, usu-
ally via the ε-­amino group of Lys residues, by the sequential action
of E1, E2, and E3 enzymes (11, 12). The chemistry and structural
1
Department of Theoretical Biophysics, Max Planck Institute of Biophysics, Frankfurt am
Main, Germany. 2Aligning Science Across Parkinson’s (ASAP) Collaborative Research
Network, Chevy Chase, MD 20815, USA. 3Walter and Eliza Hall Institute of Medical
Research, Melbourne, Victoria, Australia. 4Department of Biochemistry and Molecu-
lar Biology, Biomedicine Discovery Institute, Monash University, Melbourne, Victoria,
Australia. 5Department of Molecular and Cell Biology, University of California, Berkeley,
Berkeley, CA 94720, USA. 6California Institute for Quantitative Biosciences, Universi-
ty of California, Berkeley, Berkeley, CA 94720, USA. 7Department of Medical Biology,
University of Melbourne, Melbourne, Victoria, Australia. 8Helen Wills Neuroscience
Institute, University of California, Berkeley, Berkeley, CA 94720, USA. 9Institute of Bio-
physics, Goethe University Frankfurt, 60438 Frankfurt am Main, Germany.
*Corresponding author. Email: lazarou.​m@​wehi.​edu.​au (M.L.); jimhurley@​berkeley.​
edu (J.H.H.); gerhard.​hummer@​biophys.​mpg.​de (G.H.)
SUMOylation, and similar pathways, have been elucidated in
great detail (11, 12). ATG8 conjugation begins with the action of
the E1-­like ATG7 and the E2-­like ATG3 enzymes (6). These enzymes
have the same overall fold and active-­site cysteine residue as their
cognate Ub E1 and E2 enzymes (13), as well as unique modifica-
tions that facilitate their mutual interaction (14) and the interac-
tion of ATG3 with membranes (15). Purified ATG3 can carry out
ATG8ylation on highly curved liposomes in vitro (15) in the
absence of its cognate E3, but in vivo (16) and in a giant unila-
mellar vesicle reconstitution system (17, 18), the downstream E3
complex components are essential.
The autophagic counterpart of the Ub E3 is the ATG12–ATG5-­
ATG16L1 complex (19), which is structurally and evolutionarily
unrelated to any of its functional equivalents in ubiquitylation.
The ATG12-­ATG5 unit is itself covalently bonded through an
ATG10-­dependent reaction (20). ATG12-­ATG5 binding allosteri-
cally activates ATG3 by increasing the exposure and reactivity of
its Cys264-­linked ATG8 thioester for transfer to PE (14, 21). The
ATG16L1 portion of the complex is responsible for delivery and
positioning on the membrane (19). ATG16L1 is itself delivered
to membranes by the β-­propeller protein WD repeat domain
phosphoinositide-­interacting protein 2 (WIPI2) (18, 22). WIPI2
(and other WIPIs) are recruited to membranes early in autophagy
induction by the lipid phosphatidylinositol 3-­phosphate [PI(3)P]
(23), which is generated by the class III PI 3-­kinase complex I
(PI3KC3-­C1) early in autophagy initiation (24).
The problem of how the chemistry and structural biology of a
protein ubiquitylation-­like system is adapted to act on a membrane
substrate has been one of the major open questions in the mechanis-
tic biochemistry of autophagy. A number of pieces of the puzzle
have come together in recent years. The structural basis of the as-
sembly of a fragment of ATG3 with the ATG12–ATG5-­ATG16L1
unit was worked out for the human proteins (21). ATG16L1 con-
tains an amphipathic helix α2, adjacent to its ATG5 binding site,
which is strongly sensitive to membrane curvature (25) and essen-
tial for promotion of LC3 lipidation in liposomes and in cells (26). It
is puzzling that ATG16L1 α2 is so important for catalysis, given that

Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024 1 of 12

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
WIPI2 is capable of recruiting ATG16L1 to membranes through
its WIPI2-­interacting region (W2IR) (18, 22). The structural basis
for human ATG16L1 recruitment by WIPI2 has also been worked
out (18, 27). The ATG12-­ATG5 and WIPI2-­binding regions of
ATG16L1 are separated by a coiled coil with >100 amino acids.
The resulting extended shape and its flexibility challenge experi-
mental structure determination of the full membrane-­bound WIPI2-­
ATG12–ATG5-­ATG16L1-­ATG3 system. Here, we approached the
problem beginning with large-­scale all-­atom simulations of the
WIPI2-­ATG12–ATG5-­ATG16L1-­ATG3 on lipid membrane. Pre-
dictions from the simulations were verified experimentally in vitro
and in cellulo. In this way, we connect structural and biochemical
information into a near-­complete view of the lipidation pathway.
RESULTS
Docking step 1: WIPI2 recruits ATG12–ATG5-­ATG16L1
loaded with ATG3-­LC3 to phagophore
As a key first step in targeting the lipidation machinery to the phag-
ophore membrane, we concentrated on the WIPI2-­mediated
membrane interaction of ATG12–ATG5-­ATG16L1. The central
homodimer-­forming coiled-­coil domain (residues 78 to 230) of the
human ATG16L1 protein is predicted (28) to form a continuous
stretch of α-­helical coiled coils spanning the major part of the
domain (~115 amino acids from the N-­terminal side), allowing
reconstruction of the dimeric ATG16L1 structure by fitting geomet-
ric parameters (29) based on Crick’s equations (fig. S1A) (30). The
resulting coiled-­coil structure is in excellent agreement with crystal
structures (31, 32) of the mouse ortholog in which an overlapping
region of the coiled-­c oil domain has been resolved (fig. S1A),
providing validation for our ATG16L1 model. Using AlphaFold (33,
34), we also obtained a structural model of the E2-­like ATG3 conju-
gase loaded with LC3 (fig. S1B). The predicted ATG3-­LC3 complex
adopts a conformation compatible both with binding to the E1-­like
ATG7 homodimer in the preceding step and with formation of a
Fig. 1. Structure and dynamics of the membrane-­recruited ATG12–ATG5-­ATG16L1-­WIPI2 complex loaded with ATG3-­LC3. (A) Ribbon and (semitransparent) sur-
face representation of the full LC3 lipidation machinery bound to a membrane mimicking the phagophore lipid composition, upon equilibration and atomistic MD simu-
lation. The ATG16L1 N-­terminal helix α2 and the W2IR are highlighted in gray. (B) Dynamics of the assembly, illustrated with a superimposition of conformations sampled
at 50-­ns intervals during the final 500 ns of one 1-­μs simulation trajectory. Flexibility of interdomain loops allows the ATG3-­LC3 conjugate (yellow/white) to explore the
region of space above the membrane to which the ATG12–ATG5-­ATG16L1 is anchored.
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
thioester bond between the catalytic Cys264 side chain of ATG3 and
the C-­terminal Gly120 of LC3 to yield the E2-­substrate conjugate
(fig. S1B). The core of the human ATG3 structure is architecturally
similar to the yeast and Arabidopsis Atg3 proteins, as has been previ-
ously reported (35), with an intrinsically disordered region (36)
forming a ~100-­residue loop that contains the ATG12-­binding se-
quence (21) (Fig. 1A) as well as a region predicted to participate in β
sheet formation in the presence of LC3 (fig. S1C). The AlphaFold-­
predicted intermolecular β sheet between ATG3 residues 95 to 110
and β2 of LC3 in our structural model is consistent with the pres-
ence of a noncanonical LIR motif in the flexible region of ATG3,
which was recently shown to be required for LC3 lipidation in cellulo
(37). Combined with crystallographic structures (21, 38) of ATG12-­
ATG5 in quaternary complex with a bound fragment of ATG3 and
the N-­terminal ATG5-­binding domain of ATG16L1, we present an
atomistic model of the full LC3 lipidation machinery consisting of
the E3-­like ATG12–ATG5-­ATG16L1 complex bound to the E2-­substrate
conjugate, ATG3-­LC3 (Fig. 1A).
To determine the configuration of the ATG12–ATG5-­ATG16L1
complex recruited to phagophore membranes by the PI(3)P effector
WIPI2, we first performed atomistic molecular dynamics (MD)
simulations of a WIPI2-­ATG16L1 cocrystal structure (18) in which
WIPI2 is bound to the W2IR of ATG16L1 (residues 207 to 230).
Initially placed at a minimum distance of ~2 nm above PI(3)P-­
containing membranes mimicking the endoplasmic reticulum (ER)
lipid composition (39), WIPI2 established spontaneous membrane
contacts in an expected orientation, with the two putative phos-
phoinositide binding sites (40–42) in its β-­propeller blades 5 and 6
interacting with PI(3)P (Fig. 1A) and the N-­terminal side of the
bound ATG16L1 segment oriented away from the membrane. Lipid
interactions were formed nearly exclusively in blades 5 and 6,
around the conserved FRRG motif and in the 6CD loop (fig. S2).
In extended MD simulations, we have previously demonstrated
the ability of the 6CD region to form a membrane-­inserting am-
phipathic helix that shows moderate curvature sensitivity (25). By
2 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
aligning our structural model of ATG12–ATG5-­ATG16L1 described
above to the membrane-­associated WIPI2-­ATG16L1 configuration
established during simulations, a first model of the membrane-­
recruited LC3 lipidation machinery was thus obtained.
Upon atomistic MD simulations of ATG12–ATG5-­ATG16L1
complexed with the ATG3-­LC3 conjugate and anchored via WIPI2
to membranes approximating the phagophore lipid composition,
all components maintained structural integrity across all five 1-­μs
simulation replicates (fig. S3). Flexing and tilting motions of the
dimeric coiled-­coil structure of ATG16L1 were accompanied by
considerable flexibility exhibited in the interdomain loop regions
of ATG16L1 and ATG3 (Fig. 1B), allowing the ATG3-­LC3 conju-
gate to explore favorable binding configurations near the mem-
brane (movie S1). However, membrane binding was observed only
in the case where ATG3-­LC3 was already at the membrane surface
upon initiation of the simulation. This finding indicates that the
upward tilt of the WIPI2-­attached coiled coil tends to keep the
ATG3-­LC3 conjugate above the membrane, even if direct interac-
tions of ATG3-­LC3 are possible in principle.
To reconcile the prevailing model of membrane recruitment
of ATG12–ATG5-­ATG16L1 by WIPI2 with the requirement for
the membrane-­interacting ATG16L1 N-­terminal helix α2 (26),
we hypothesized that upon initial recruitment through WIPI2,
direct membrane binding by helix α2 constitutes a crucial second
step in delivering ATG3-­LC3 nearer to the target membrane. We
focus here on the cis configuration in which the entire LC3
lipidation machinery becomes associated with the same patch of
membrane (43). Our molecular model does not rule out the alter-
native possibility whereby ATG12–ATG5-­ATG16L1 anchored at
omegasomal membranes would bridge an intermembrane distance
to facilitate LC3 conjugation to the nascent phagophore in trans
(22). However, it seems difficult to reconcile the trans model with
occupancy of the second WIPI2 site (27).
Docking step 2: Helix α2 of ATG16L1 pulls ATG3-­LC3
to membrane
As a possible second step in membrane targeting, we focused on
helix α2 of the ATG16L1 N-­terminal domain, which has been shown
to bind membranes (26) with a preference for positive membrane
curvature (25). For an ATG12–ATG5-­ATG16L1 complex attached
to the membrane via WIPI2, this mode of membrane interaction is
made possible by the flexibility of the ~30-­residue interdomain loop
between the N terminus of ATG16L1 and its coiled coil. We demon-
strated this ability of ATG16L1 to engage with the membrane simul-
taneously, at one end, through recruitment by WIPI2 and, at the
other end, via helix α2 by gently pulling the ATG16L1 helix α2
toward the membrane in steered MD simulations and then relaxing
the complex in extended MD simulations (Fig. 2A).
Building upon our previous MD simulations of ATG12–
ATG5-­ATG16L1 binding to curved membranes (25), with ATG16L1
interacting either at the membrane surface or with an embedded
hydrophobic face of the amphipathic helix α2, we obtained models
of membrane-­bound ATG12–ATG5-­ATG16L1 loaded with ATG3-­
LC3 (Fig. 2B). In atomistic MD simulations, the flexible ATG3 loop
then allowed ATG3-­LC3 to reach the membrane spontaneously
while maintaining its interactions with the α2-­anchored ATG12–
ATG5-­ATG16L1 complex (movie S2). In this configuration, we
found the N-­terminal helix of the ATG3 conjugase to embed into
the membrane (Fig. 2C and movie S2) without any biasing force.
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
Once inserted, the ATG3 amphipathic helix remained embedded
in the membrane through the course of the simulation, establishing
stable membrane contacts. This finding is consistent with the pre-
viously reported role of the ATG3 N terminus as an essential
membrane-­targeting element (44). Geometrically, the two ATG3-­
LC3 conjugates flexibly connected to the ATG12–ATG5-­ATG16L1
complex can simultaneously engage with the membrane for parallel
lipidation reactions.
Docking step 3: Catalytic domain of ATG3 forms stable
membrane interaction interface upon membrane insertion
of ATG3 N-­terminal helix
For a decisive third and final targeting step, we explored how the
catalytic domain established membrane contact. The ATG3 conju-
gase has been reported to show basal activity in vitro for catalyzing
LC3 conjugation in the absence of ATG12–ATG5-­ATG16L1 (45).
Having observed stable membrane association of the ATG3-­LC3
conjugate held near phagophore-­mimetic membranes by ATG12–
ATG5-­ATG16L1, we sought to further collect lipid contact data on
ATG3-­LC3 by initiating a set of longer (2-­μs) replicates of smaller
simulation systems containing the isolated conjugate placed directly
above membranes. In 8 of the 20 trajectories thus obtained, sponta-
neous membrane insertion of the ATG3 N-­terminal helix occurred
within the first 1 μs. Comparison of ATG3-­LC3 lipid contacts (after
insertion of the ATG3 helix) reveals a consistent membrane interac-
tion interface in the presence or absence of ATG12–ATG5-­ATG16L1
(Fig. 3A and fig. S4). Membrane insertion of the N-­terminal helix
is also accompanied by its increased ordering relative to the enzyme
body (fig. S5), consistent with a hypothesized role in controlling the
structural dynamics of the complex (46). We found that the ATG3
protein dominated the interactions of the conjugate with PE-­
containing membranes.
ATG3-­LC3 presents active site toward the membrane in
configuration conducive to lipidation reaction
With ATG3-­LC3 at the membrane, we explored the structural foun-
dation of the actual lipidation reaction. The folded core of ATG3
comprises a six-­stranded β sheet (strands β1 to β6) surrounded by α
helices (47). Among regions of ATG3-­LC3 that formed frequent
membrane interactions in our simulations were short sequences of
residues within intersecondary structure loops of the ATG3 core,
namely, (i) catalytic domain residues 208 to 211 and 242 to 243 of
the β3/β4 and β4/β5 loops, respectively; (ii) residues 262 to 265
encompassing the thioester-­forming Cys264 between β6 and the
succeeding α helix; and to a lesser extent (iii) residues 61 to 64 with-
in the β1/β2 loop. The catalytic site, which contains Cys264 of ATG3
covalently bonded to the C terminus of LC3, was situated centrally
on the membrane interaction interface identified above and exposed
toward membrane lipids (Fig. 3A). Furthermore, the ATG3-­LC3
conjugate formed distinct interactions with different types of lipids
present in the membrane, with PE localizing particularly near the
catalytic center (Fig. 3B). Our data thus suggest a preferred orienta-
tion of ATG3-­LC3 on the membrane that is compatible with catalyz-
ing LC3 conjugation to the phagophore.
Mutations at ATG3 membrane interaction face impair LC3
lipidation in vitro and in cellulo
The MD simulations identified a surface of ATG3 that was consis-
tently in contact with the membrane in the context of the larger
3 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Membrane-­bound ATG16L1 helix α2 brings ATG3-­LC3 into membrane contact. (A) Snapshot of the WIPI2-­recruited ATG12–ATG5-­ATG16L1 complex with the
N terminus of ATG16L1 also engaged in membrane interaction in an atomistic MD simulation. The ATG16L1 helix α2 had been steered gently to the membrane surface at
a speed of 0.5 nm ns−1, by application of harmonic restraints on the center-­of-­mass distance between the helix and the membrane. The complex was subsequently relaxed
in extended (1 μs) simulations. (B) Initial configuration of simulations representing a further stage of ATG3-­LC3 delivery by ATG12–ATG5-­ATG16L1, following membrane
recruitment via WIPI2. The hydrophobic face of helix α2 is embedded into the membrane at this stage. (C) Snapshot of ATG3-­LC3 delivered to the membrane while bound
to ATG12–ATG5-­ATG16L1, with the ATG3 N-­terminal helix (orange) spontaneously inserted between membrane lipids. Taken at t = 450 ns from a 1-­μs simulation replicate.

ATG12–ATG5-­ATG16L1-­ATG3–LC3B-­WIPI2 complex. Residues
in this patch include Lys62, Lys64, Lys208, Tyr209, Tyr210, Thr244, His262,
Cys264, Arg265, and His266 (Fig. 4A). The presence of Cys264 was ex-
pected, given this residue’s known role as the LC3 donor in the reac-
tion (6, 21, 48). We assayed LC3B conjugation activity in a small
unilamellar vesicle (SUV) system similar to that originally used
to demonstrate Atg8 conjugation activity of the Atg12–Atg5-­Atg16-­
Atg3 complex (48). Here, purified human proteins were used
(1.0 μM), WIPI2 (0.5 μM) was included in the protein mixture, and
10% PI(3)P was included in the SUVs (18). Activity was monitored
by the conversion of LC3B-­I to LC3B-­II. As expected, essentially
complete conversion was seen for wild type, while the mutation
C264A of the catalytic cysteine as a negative control completely
eliminated activity (Fig. 4, B and C). The mutation H262A also
completely abolished activity, suggesting a direct role in catalysis
beyond its membrane interactions alone. This is discussed further
below. Activity was nearly abolished in K208D and sharply reduced
in T244A, with small but significant reductions seen in K62D/K64D
and Y209A. Smaller apparent reductions were seen in Y210A,
R265A, and H266A. The observation that most of these mutations
had at least some effect on catalysis in the SUV system confirms
the predicted membrane interaction surface identified by the MD
simulations.
To determine whether the predicted membrane function had the
same function in living cells as in the reconstituted system, we gener-
ated an ATG3 knockout (KO) HeLa cell line. ATG3 KO was verified
by Western blotting (Fig. 5A). Starvation-­induced autophagic flux
was monitored with the HaloTag-­LC3B system based on the appear-
ance of a free HaloTag band (17). As expected, expression of the
wild-­type construct rescued autophagic flux in the KO cells, while no
flux was observed in the C264A rescue (Figs 5, A and B, and fig. S6).
The mutational effects on autophagic flux in the ATG3 KO cells mir-
ror the pattern seen in the SUV assays. Y210A, R265A, and H266A,
which have small (not statistically significant) reductions in activity
in SUVs (Fig. 4, B and C), manifest modest reductions in flux in cells
(Fig. 5, A and B). H262A and K208D show a nearly complete loss of
activity in both the SUV (Fig. 4, B and C) and flux assays (Fig. 5, A
and B). The effects of Y209A, T244A, and K62D/K64D are interme-
diate in both settings [Figs. 4 (B and C) and 5 (A and B)]. The rescue
of conversion of LC3B and GABARAPL1 was also monitored in cells

Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024 4 of 12

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. Membrane lipid contacts formed by the ATG3-­LC3 conjugate. (A) Ribbon representation of the ATG3-­LC3 conjugate structure before simulation, with the cata-
lytic site indicated with a red dashed circle around the thioester bond (left). Coloring ATG3-­LC3 residues by their mean frequency of membrane contacts (middle/right;
white to gray at increasing contact frequency), upon spontaneous insertion of the ATG3 N-­terminal helix in atomistic MD simulations, reveals a consistent membrane in-
teraction interface while bound to ATG12–ATG5-­ATG16L1 (final 500 ns of a 1-­μs trajectory) and in the absence thereof (final 1 μs of each of eight 2-­μs replicates). Top-­
ranked residues are highlighted as sticks: the only qualitatively discernible difference between membrane interaction data from the two independent simulation systems
is indicated with a red triangle. For clarity of comparison, lipid interaction data are projected onto the same view of the initial model before simulation. (B) Proportion of
membrane contacts formed by ATG3-­LC3 residues with different types of lipids present in the membrane, illustrated with PE, phosphatidylcholine (PC), and PI. Membrane
interaction data are averaged across the final 1 μs of eight 2-­μs replicates and normalized for each lipid type such that a value of 1.0 is assigned to the residue(s) showing
highest specificity for that lipid.

(fig. S6). Mutations that show a complete loss of activity in SUVs and
the flux assay were also negative for LC3B and GABARAPL1 conver-
sion. Mutations with intermediate defects in the SUV and flux assays
showed smaller defects in the ATG8 protein conversion, which is
attributed to differences in the stringency of the assays. The main
conclusion from the ATG3 KO experiments is that the membrane
interaction surface identified in the MD simulations accurately pre-
dicted loss of function in the biochemical and cellular assays.
Conserved His262 of HPC motif facilitates ATG3-­catalyzed
LC3 lipidation
Previous studies have shown that the transfer of LC3 from ATG3 to
lipid substrates is sensitive to pH and takes place more efficiently
under slightly basic conditions in vitro, most likely through an
effect on the ATG3 conjugase activity (49, 50). While the proton-
ation state of the target PE amine group is expected to show little
variation within the pH range of interest, we note the presence of
two histidine residues, His262 and His266, in close proximity to the
catalytic Cys264. Both histidines are fully conserved across ATG3
homologs and, with their characteristic pKa just below physiological
pH, serve as possible acidity sensors for the ATG3-­catalyzed reaction.
In atomistic MD simulations of the ATG3-­LC3 conjugate with
the His262 and His266 side chains both in their unprotonated state
(which is predicted to be the dominant species at pH ≥ 7), His266
remained oriented toward the protein interior with a minimum dis-
tance of ~1 nm to the nearest lipid (Fig. 6A). By contrast, frequent
lipid interactions formed by His262 are suggestive of a direct role in
the LC3 conjugation reaction. Whereas the nucleophile of the reac-
tion, the PE amine group, did spontaneously approach the backbone
carbonyl carbon of Gly120 (LC3) to be attacked (reaching a mini-
mum distance of ~0.4 nm), such interactions were infrequent.
Meanwhile, the unprotonated nitrogen of the His262 imidazole was
observed to interact with the positively charged primary amine of
PE headgroups within bonding distance (<0.2 nm) to the amine
proton (Fig. 6A). Furthermore, our simulations capture a configura-
tion in which the His262:PE interaction coincided with that between
PE and Gly120 (Fig. 6B).
His262 and Cys264 of human ATG3 form part of the HPC motif
that is conserved across orthologs of ATG3 as well as ATG10, an
E2-­like autophagic enzyme that catalyzes ATG12 conjugation to
ATG5 (6). Combined with previous (35, 51) and present evidence of
a critical role of His262 for ATG3 conjugase activity, our simulation

Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024 5 of 12

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Mutational analysis of ATG3 membrane interaction face. (A) Snapshot of ATG3-­LC3 upon spontaneous membrane association, taken at t = 1 μs of a 2-­μs simu-
lation replicate. Zoomed-­in views from the side (left) and bottom (right; with lipids omitted for clarity) are shown. Phosphatidylethanolamine lipids are highlighted in
green. (B) ATG3 in vitro LC3 lipidation. Do-­SUVs [70% PC:20% PE:5% PS:5% PI(3)P] were incubated with ATG3 wild type or mutant, ATG7, E3, WIPI2d, and LC3B. After 20 min,
samples were loaded onto a 4 to 15% SDS-­PAGE gel and stained with Coomassie blue. (C) Quantification of in vitro LC3 lipidation results, plotting the LC3B-­II percentage
in the total band intensities of LC3B-­I and LC3B-­II. P values were calculated using Student’s t test: not significant (NS), P ≥ 0.05; **, 0.001 < P < 0.01; ****P < 0.0001.

results are suggestive of a plausible reaction mechanism in which
the His262 imidazole ring would deprotonate the PE amine group for
nucleophilic attack on the Gly120 carbonyl of LC3 (Fig. 6C). As part
of such a proposed mechanism, the unique backbone conforma-
tional restraints conferred by the cyclic side chain of Pro263 in the
HPC motif would be crucial for orienting His262 and Cys264 side
chains in relative positions conducive to catalysis, explaining their
full conservation (fig. S7). The protein backbone conformation
conferred by Pro263 also holds the backbone amide of Cys264 within
bonding distance of the carbonyl oxygen of Gly120 (Fig. 6C), which
would stabilize the oxygen anion intermediate formed during the
reaction. Energetically favorable breakage of the thioester bond will
then yield the LC3-­PE conjugate, a stable amide product. Alterna-
tively, ATG3 has been reported to catalyze conjugation of ATG8
family proteins to phosphatidylserine (PS) lipids in the nonca-
nonical pathway of autophagy (52). In accordance with this, our
simulations of the ATG3-­LC3 conjugate also capture an analogous
membrane-­interacting configuration likely poised for reaction with
a PS molecule (fig. S8).
His266, the second of the two conserved histidine residues de-
scribed above, has been implicated in the pH-­dependent conjugase
activity of ATG3 in a recent study (35). To assess the effect of alter-
ing the protonation state of His266, we performed additional MD
simulations of the ATG3-­LC3 conjugate in which the His266 imidaz-
ole ring was doubly protonated. Notably, the extra proton destabi-
lized the local protein structure (Fig. 6D). A reorientation brought
the His266 side chain into direct membrane contact within the first
hundreds of nanoseconds in 7 out of 10 simulation replicates (Fig. 6,
E and F). These results are consistent with His266 fulfilling a pH-­
sensitive structural role, as previously proposed for its counterpart
in ATG3 orthologs (His236 in the yeast protein and His260 in Arabi-
dopsis) (50), and provide an explanation for the alternative confor-
mations in this region between available crystal structures obtained
at different pH values (47, 53).
DISCUSSION
Building upon an increasing collection of structural and biochemi-
cal data on the components and interactions that form the autopha-
gic LC3 lipidation machinery, we set out to complete the molecular
puzzle of how the E3-­like ATG12–ATG5-­ATG16L1 complex and
the E2-­like conjugase ATG3 deliver LC3 to phagophore membranes.
Results from atomistic MD simulations point toward a multistage
mechanism progressively localizing the ATG3-­LC3 conjugate near-
er to the target membrane and orienting the reactive center of LC3
conjugation toward lipid substrates. This process requires the sequen-
tial action of three previously identified membrane sensors within
the assembly: (i) WIPI2 as the PI(3)P effector protein that drives
membrane recruitment of ATG12–ATG5-­ATG16L1 (22, 54), (ii) the
curvature-­sensitive ATG16L1 helix α2 within the (ATG12–)ATG5-­
binding domain (25, 26), and (iii) the N-­terminal amphipathic helix
and membrane docking face of ATG3 (44).
As an emerging theme in cellular processes, with analogies to the
multistep process of docking in vesicle fusion (55), the stepwise

Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024 6 of 12

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 5. Mutations in ATG3 membrane interaction face impair function in cells.
(A) ATG3 KO stably expressing HaloTag-­LC3B with and without untagged ATG3 wild
type (WT) or mutants were starved in EBSS for 6 hours. Cells were pulse labeled
with 50 nM TMR-­conjugated Halo ligand before starvation. Cell lysates analyzed by
immunoblotting, showing one representative subset of data from triplicate experi-
ments (fig. S6). (B) Autophagy levels represented by percentage cleaved Halo were
obtained by calculating band intensities of free Halo (cleaved) compared to total
Halo (uncleaved plus cleaved). Significance was calculated by comparing KO
and mutants to WT. P values were calculated via one-­way ANOVA: **P < 0.01,
***P < 0.001, and ****P < 0.0001. Data shown are mean ± SD from three indepen-
dent experiments.
mechanism for the membrane targeting of LC3 provides additional
layers of regulatory potential to the autophagic pathway. On the pro-
tein side, phosphorylation and other posttranslational modifications
will affect the stability, accessibility, and affinity of the distinct interac-
tion elements. On the membrane side, variations in lipid composition
and phosphatidylinositol (PI) phosphorylation will modulate mem-
brane recruitment. The phagophore lipid composition in particular
modulates the recruitment of WIPI2 as anchor for ATG16L1 in dock-
ing step 1 as well as the membrane insertion of the ATG16L1 α2 helix
and the ATG3 N-­terminal helix in steps 2 and 3, respectively. Grow-
ing evidence points to a second WIPI2-­interacting site within the
ATG16L1 coiled-­coil domain (27), which would facilitate step 2 of
our model in a PI(3)P-­dependent manner. Occupancy of the second
site for WIPI2 has been proposed to facilitate LC3 lipidation following
the initial membrane recruitment of ATG12–ATG5-­ATG16L1, in line
with earlier observations of allosteric activation of the complex by
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
WIPI2 (43). Consistent with our structural model, a second WIPI2
molecule bound to the coiled-­coil region would pull the N-­terminal
side of ATG16L1 closer to the membrane surface, with the WIPI2
FRRG motif oriented toward the membrane (fig. S9), thereby facilitat-
ing the membrane insertion of the ATG16L1 α2 helix in step 2 of
our docking model. A three-­dimensional model of the complex of
ATG12–ATG5-­ATG16L1 loaded with ATG3-­LC3 anchored to the
bilayer by a second WIPI2 (fig. S9) favors lipidation in cis.
Through multi-­microsecond all-­atom MD simulations, collecting
a total of >50-­μs membrane docking trajectories, we have examined
the lipid-­interacting regions of the complete ATG3-­LC3 conjugate
for molecular determinants of the LC3 conjugation mechanism. The
spontaneous membrane insertion of the ATG3 N-­terminal helix in
our simulations is consistent with a role of this region in positioning
the protein onto the membrane for subsequent enzymatic activity
(15, 46, 56). Additional membrane-­interacting residues concentrate
around the Cys264 residue holding LC3. Mutations of these residues
impact lipidation both in vitro and in vivo, confirming the catalytic
relevance of the observed membrane interactions.
Simulations and experiments identify distinct roles for two fully
conserved histidine residues in the vicinity of the catalytic cysteine
of ATG3. We found neutral His266 to stabilize a catalytically compe-
tent structure of the active site, consistent with retention of full lipi-
dation activity by the H266A mutant. In contrast, protonation of
His266 disrupted the active site in our MD simulations, consistent
with a role of His266 as pH sensor (35). While uncharged His266
serves to stabilize the catalytic loop conformation, our data point to
active participation of His262 in the initiation of the LC3 lipidation
reaction. In particular, we found the unprotonated His262 imidazole
nitrogen to be positioned as proton acceptor from PE. Consistent
with a possible catalytic role, the H262A mutation abolished func-
tion. His262 is the starting residue in the highly conserved HPC mo-
tif (6) of ATG3, which is shared with the ATG10 conjugase family.
However, in ATG10, the counterpart of ATG3 His266 is a threonine,
which may reflect the distinct substrate specificity of the two en-
zymes (fig. S7).
The critical biological role of the ATG12–ATG5-­ATG16L1 com-
plex in mammalian autophagy (19), and before that, the role of the
corresponding Atg12–Atg5-­Atg16 complex in yeast (16), has long
been appreciated. Yet, the precise role of this complex in LC3 lipi-
dation has been challenging to define. The role of the extensive
structural elements linking the N-­terminal helix of ATG3 on the
one hand, and the established WIPI2-­dependent membrane dock-
ing site on the other, have proven difficult to characterize as the
membrane-­associated system is too large for nuclear magnetic res-
onance, yet too dynamic for x-­ray crystallography or single-­particle
cryo–electron microscopy. Under the “computational microscope”
of MD simulations, the role of the connecting elements in mediat-
ing a stepwise docking process has now been unveiled. As a core
element in the molecular machinery of selective autophagy, this far
more detailed insight into the membrane docking steps of LC3 will
undoubtedly facilitate the therapeutic targeting of autophagy in
Parkinson’s disease and other neurodegenerative diseases.
MATERIALS AND METHODS
Structural models of protein complexes
Atomistic models of the ATG3-­ATG12–ATG5-­ATG16L1 and
WIPI2d-­ATG16L1 complexes were based on crystal structures with
7 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. Two conserved histidine residues around ATG3 active site assume distinct roles. (A) Radial distribution function g(r) of the amine nitrogen atoms of PE lipids
as a function of their distance r from select protein atoms (arrow: direct contact). (B) Snapshot of ATG3-­LC3 interacting with membrane lipids during atomistic MD simula-
tion. The zoomed-­in view captures a PE lipid binding into the ATG3 active site, near the thioester bond under attack. The nearest PE amine proton is also within bonding
distance (<0.2 nm) of the unprotonated nitrogen atom of the His262 imidazole ring. (C) Possible mechanism for initiation of LC3 lipidation reaction, whereby the ATG3
His262 imidazole ring would facilitate a nucleophilic attack on the Gly120 carbonyl of LC3. The backbone amide of Cys264 is in position to stabilize the developing negative
charge on the Gly120 oxygen. Illustrated using the simulation snapshot of (B), showing only the “attacking” lipid for clarity. (D) Number of backbone hydrogen bonds in
ATG3 helix 265 to 280 in simulations of ATG3-­LC3 with ATG3 His266 in an uncharged or doubly protonated state. (E) Number of protein and membrane lipid contacts made
by the His266 imidazole ring in the uncharged and the doubly protonated state, respectively, in simulations of alternative models of the ATG3-­LC3 conjugate. (F) Snapshot
of membrane-­associated ATG3-­LC3 in which the side chain imidazole of His266 (indicated with a red dashed circle) was doubly protonated with a charge of +e. At t = 1 μs
of the 2-­μs simulation replicate shown, destabilization of the local protein structure has brought the His266 side chain into membrane contact.

Protein Data Bank (PDB) IDs 4NAW (21) and 7MU2 (18), respec-
tively. The ATG16L1 N-­terminal domain in the former complex was
replaced by a more complete structure [PDB ID: 4TQ0 (57)]. As intro-
duced previously (25), an alternative conformation of the same
ATG16L1 region was generated in PyMOL 2.3 (RRID:SCR_000305,
https://pymol.org/) (58) by rotation of helix α2 relative to helix α1 at
the Gln30/Ala31 hinge. A model for the dimeric central ATG16L1 do-
main was completed through (i) homology modeling of residues 141 to
225 using SWISS-­MODEL (RRID:SCR_018123, https://swissmodel.
expasy.org/) (59) based on crystal structures of the mouse protein
[PDB IDs: 6ZAY (32) and 6SUR (31)] and (ii) parameter fitting for
residues 78 to 193 with CCBuilder 2.0 (https://github.com/woolfson-­
group/ccbuilder2) (29) upon coiled-­coil prediction (28) by NPS@
(https://npsa-­prabi.ibcp.fr/) (60). Unstructured interdomain loops
were added using the DEMO server (https://zhanggroup.org/DEMO/)
(61) to yield an ATG16L1 dimer encompassing residues 1 to 247.
AlphaFold v2.2 (https://github.com/google-­deepmind/alphafold)
(34) was used to model ATG3-­LC3B in complex with the ATG7
homodimer. The Cys264 side chain of ATG3 was connected to the
LC3B C terminus by a thioester bond, parameterized using
CHARMM-­GUI (https://charmm-­gui.org/) (62, 63). The ATG5
Lys130 side chain was similarly connected to the ATG12 C termi-
nus, via an isopeptide bond. The ATG16L1 WD40 domain [dis-
pensable for canonical autophagy (64)] was excluded from the
model, as were the unstructured ATG12 residues 1 to 52 and WIPI2d
residues 1 to 11 and 362 to 425. Exposed N-­or C-­terminal groups
at the end(s) of each incomplete structure or truncated construct
were neutralized. Protonation states of amino acid side chains
were assigned according to pKa prediction by PROPKA 3 (https://
github.com/jensengroup/propka) (65). His183 and His255 at the puta-
tive PI(3)P binding sites of WIPI2d were protonated. Six models of
the ATG3-­LC3B conjugate were generated, with the imidazole of

Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024 8 of 12

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
ATG3 His262 uncharged (protonated at the δ-­ or ϵ-­nitrogen in alter-
native models) and that of His266 uncharged (protonated at δ-­ or ϵ-­
nitrogen in alternative models) or cationic (doubly protonated).
MD simulations
MD simulations were performed with GROMACS 2020 (RRID:SCR_014565,
http://gromacs.org/) (66) using the CHARMM36m force field (67).
Following the same protocol as previously described (25), all mem-
branes consisted of 60% dioleoylphosphatidylcholine (DOPC),
20% dioleoylphosphatidylethanolamine (DOPE), 5% dioleoylphos-
phatidylserine (DOPS), 10% 1-­palmitoyl-­2-­oleoyl-­*sn*-­glycero-­
3-­phosphoinositol (POPI), and 5% PI(3)P based on the ER lipid
composition (39) and were prepared initially in a coarse-­grained repre-
sentation using the insane method (https://github.com/Tsjerk/Insane)
(68). Curved membranes were constructed using LipidWrapper
(https://github.com/durrantlab/lipidwrapper) (69) by fitting the am-
plitude of the membrane buckle as a sine function of its x coordinate.
Each coarse-­grained membrane system was solvated with 150 mM of
aqueous NaCl, equilibrated for 200 ns and converted into an atomistic
representation using the CG2AT2 (https://github.com/owenvickery/
cg2at) (70) tool. Atomistic models of protein complexes were placed
above membranes after CG2AT2 conversion, followed by resolva-
tion and 10 ns of further equilibration. Simulation replicates were
independently prepared and equilibrated. During equilibration, har-
monic positional restraints with a force constant of 1000 kJ mol−1
were applied to nonhydrogen protein atoms or backbone beads. The
xy dimensions of buckled membrane systems were fixed in simula-
tions. System temperature and pressure were maintained at 310 K and
1 bar, respectively, using the velocity-­rescaling thermostat (71) and a
semi-­isotropic Parrinello-­Rahman barostat (72) during the produc-
tion phase. The integration time step was 2 fs. Long-­range electrostat-
ic interactions were treated using the smooth particle mesh Ewald
method (73, 74) with a real-­space cutoff of 1 nm, a Fourier spacing of
0.12 nm, and charge interpolation through fourth-­order B splines. The
LINCS linear constraint solver (LINCS) algorithm was used to con-
strain covalent bonds involving hydrogen atoms (75). Simulation tra-
jectories (table S1) were analyzed through the MDAnalysis 2.0 library
(https://mdanalysis.org/) (76, 77) in Python 3.6 (RRID:SCR_008394,
http://python.org/).
Protein expression and purification
ATG3 constructs used for in vitro lipidation assays were expressed in
Escherichia coli (BL21) DE3 star cells (Invitrogen, C601003). Cells were
grown in LB media at 37°C until an OD600 (optical density at 600 nm)
of 0.8 is reached. The culture was induced with 1 mM isopropyl-­β-­​d-­
thiogalactopyranoside and grown overnight at 18°C. Cells were pel-
leted and resuspended in 50 mM Hepes (pH 7.5), 300 mM NaCl, 2 mM
MgCl2, 10 mM imidazole, and 1 mM tris(2-­carboxyethyl)phosphine
(TCEP) supplemented with EDTA-­free protease inhibitors (Roche).
The cells were lysed via sonication, and lysate was clarified by centrifu-
gation (17,000 rpm for 1 hour at 4°C). The supernatant was then
applied to 1 ml of Ni–nitrilotriacetic acid resin. The resin was subse-
quently washed thoroughly with at least 100 column volumes (CV) of
lysis buffer, and the protein was eluted with lysis buffer supplemented
with 300 mM imidazole. The eluted proteins were concentrated and
loaded onto a Superdex 200 column (10/300 GL; GE Healthcare)
equilibrated with a buffer containing 25 mM Hepes (pH 7.5), 150 mM
NaCl, and 1 mM TCEP. Peak fractions corresponding to the pro-
tein were collected, pooled, snap-­frozen in liquid nitrogen, and stored
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
at −80°C. Purification of ATG12–ATG5-­ATG16L1 (dx.doi.org/10.17504/
protocols.io.­br6qm9dw), ATG7 (https://dx.doi.org/10.17504/protocols.­
io.­bsennbde), and LC3 (https://dx.doi.org/10.17504/protocols.io.­
j8nlkw82dl5r/v1) used for liposome lipidation assays was per­formed
as previously described (43). Purification of WIPI2d was performed
as previously described (https://dx.doi.org/10.17504/protocols.io.­
buxqnxmw) (18).
In vitro LC3 lipidation assays
A lipid mixture with a molar composition of 70% DOPC, 20%
DOPE, 5% DOPI(3)P, and 5% DOPS (Avanti Polar Lipids) was dried
under a nitrogen stream and put under vacuum overnight. Lipids were
resuspended at 1 mg/ml in the assay buffer [25 mM Hepes (pH 7.5),
135 mM NaCl, 2.7 mM KCl, and 1 mM TCEP], freeze-­thawed sev-
en times, and extruded 17 times through a 100 nM filter (Whatman).
Reactions were set up at room temperature in the assay buffer
to a final concentration of 1 μM of the indicated ATG3 construct,
1 μM ATG7, 1 μM E3, 500 nM WIPI2d, 5 μM LC3B, 0.5 mM adenos-
ine 5′-­triphosphate, 1 mM MgCl2, and liposomes (0.5 mg/ml). Fif-
teen microliters of reaction was quenched at 20 min with 4× lithium
dodecyl sulfate (LDS loading buffer, boiled at 60°C for 10 min, and
then loaded onto SDS–polyacrylamide gel electrophoresis (SDS-­
PAGE) gels. Protein bands were visualized with Coomassie blue.
Three biological replicates were performed. Protein band intensity of
LC3B-­I and LC3B-­II was analyzed by ImageJ (RRID:S­CR_003070,
https://imagej.nih.gov/ij/). Quantification of LC3B-­II formation was
plotted as percentage of total LC3B among the measured values for
each ATG3 protein in a bar graph. Averages and SDs were calculated.
The P values were calculated using an unpaired two-­tailed Student’s t
test. P values were considered as follows: not significant (NS), P ≥
0.05; *, 0.01 < P < 0.05; **, 0.001 < P < 0.01; ***, 0.0001 < P < 0.001;
and ****P < 0.0001. The LC3 lipidation assay protocol is available at
https://dx.doi.org/10.17504/protocols.io.e6nvwjxodlmk/v1.
Cloning and generation of stably expressing ATG3 wild type
and mutant HeLa cell lines
All HeLa cells (American Type Culture Collection, catalog no. CCL-­2,
RRID:CVCL_0030) used were cultured in Dulbecco’s modified
Eagle’s medium (DMEM) supplemented with 10%(v/v) fetal bovine
serum (Cell Sera), 10 mM Hepes, 1% (v/v) penicillin/streptomycin
antibiotic solution (Sigma-­Aldrich; P4333-­100ML), 1× GlutaMAX
(Gibco, 35050061), and 1× nonessential amino acids (Gibco,
11140050). All cells were stored under standard conditions in an
appropriate vessel in a humidified incubator at 37°C and a CO2 level
of 5%. Polymerase chain reaction products of ATG3 wild type and
ATG3 mutants were subcloned into linearized pMX-­IG backbone
using NEBuilder HiFi DNA Assembly Master Mix (New England
Biolabs, E2621L) containing an internal ribosomal entry site–yellow
fluorescent protein element for untagged expression of ATG3 wild
type and mutants. From this, the following ATG3 plasmids were
generated: pMX-­IG-­ATG3 (RRID:Addgene 212021), pMX-­IG-­
ATG3-­C264A (RRID:Addgene 212023), pMX-­IG-­ATG3-­Y209A
(RRID:Addgene 212024), pMX-­IG-­ATG3-­Y210A (RRID:Addgene
212025), pMX-­IG-­ATG3-­T244A (RRID:Addgene 212026), pMX-­
IG-­ATG3-­H262A (RRID:Addgene 212027), pMX-­IG-­ATG3-­R265A
(RRID:Addgene 212028), pMX-­IG-­ATG3-­H266A (RRID:Addgene
212029), pMX-­IG-­ATG3-­K208D (RRID:Addgene 212030), and
pMX-­IG-­ATG3-­K62D/K64D (RRID:Addgene 212031). All plas-
mids were verified by DNA sequencing. Stable cell lines were
9 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
generated using a retroviral system where pMRX-­IP-­HaloTag7-­
LC3 (RRID:Addgene 184899) and all pMX-­IG ATG3 (RRID:Addgene
212021) constructs alongside retroviral packaging plasmids vesicu-
lar stomatitis virus glycoprotein (RRID:Addgene 8454) and Gag-­pol
(RRID:Addgene 14887) were transfected into human embryonic
kidney 293T cells (RRID:CVCL_0063) using Lipofectamine LTX
(Invitrogen, 15338030) for 15 hours. The next day, transfection me-
dium was replaced with complete DMEM. After 24 hours, the retro-
viral supernatant was collected, filtered, and added to ATG3 KO
HeLa cells for 24 to 48 hours alongside polybrene (8 μg/ml; Sigma-­
Aldrich, H9268). After transduction, cells were allowed to recover
in full growth medium for 5 to 7 days before fluorescence sorting for
positive cells via fluorescence-­activated cell sorting. The protocol
used for generating stable cell lines using the retroviral system is
available at dx.doi.org/10.17504/protocols.io.81wgbyez1vpk/v1.
HaloTag-­LC3B starvation assay
Halo-­LC3B Assay was performed as previously described (78). Cells
were seeded at 400,000 cells per well in a six-­well plate 1 day before.
Cell were fed with 1 ml of complete DMEM for 1 hour followed by
incubation in complete DMEM containing 50 nM tetramethylrhoda-
mine (TMR)-­conjugated Halo ligand (Promega, GA1120) for 20 min.
Cells were then washed thrice with 1× PBS followed by incubation in
Earle's Balanced Salt Solution (EBSS) buffer (Thermo Fisher Scientific,
2410043) to induce autophagy by starvation for 6 hours. Afterward,
cells were washed with 1× PBS before harvesting with cell scrapers.
Cells were pelleted and lysed in 1× LDS sample buffer (Invitrogen,
NP0007) supplemented with 100 mM dithiothreitol (Sigma-­Aldrich,
10708984001). Samples were heated at 99°C with shaking for 10 min, and
protein concentration was measured using a nanodrop spectrophotom-
eter (Thermo Fisher Scientific). Twenty micrograms of protein per
sample was analyzed on 4 to 12% bis-­tris gels (Invitrogen, WG1402A)
according to the manufacturer’s instructions. Gels were electro-­
transferred to polyvinyl difluoride membranes (Immobilon) and im-
munoblotted using indicated antibodies. Quantification of percentage
autophagy was calculated by measuring the amount of cleaved Halo
against total Halo (cleaved plus uncleaved). Statistical analysis was
performed via one-­way analysis of variance (ANOVA) using the mul-
tiple comparisons function in GraphPrism 9 (RRID:SCR_002798,
http://graphpad.com/). A complete protocol is available at dx.doi.
org/10.17504/protocols.io.e6nvwdo9zlmk/v1.
Supplementary Materials
This PDF file includes:
Figs. S1 to S9
Table S1
Legends for movies S1 and S2
Other Supplementary Material for this manuscript includes the following:
Movies S1 and S2
REFERENCES AND NOTES
1. H. Morishita, N. Mizushima, Diverse cellular roles of autophagy. Annu. Rev. Cell Dev. Biol.
35, 453–475 (2019).
2. J. W. Harper, A. Ordureau, J.-­M. Heo, Building and decoding ubiquitin chains for
mitophagy. Nat. Rev. Mol. Cell Biol. 19, 93–108 (2018).
3. J. N. S. Vargas, M. Hamasaki, T. Kawabata, R. J. Youle, T. Yoshimori, The mechanisms and
roles of selective autophagy in mammals. Nat. Rev. Mol. Cell Biol. 24, 167–185 (2023).
4. A. M. Pickrell, R. J. Youle, The roles of PINK1, parkin, and mitochondrial fidelity in
Parkinson’s disease. Neuron 85, 257–273 (2015).
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
5. D. J. Klionsky, G. Petroni, R. K. Amaravadi, E. H. Baehrecke, A. Ballabio, P. Boya,
J. M. B.-­S. Pedro, K. Cadwell, F. Cecconi, A. M. K. Choi, M. E. Choi, C. T. Chu, P. Codogno,
M. I. Colombo, A. M. Cuervo, V. Deretic, I. Dikic, Z. Elazar, E.-­L. Eskelinen, G. M. Fimia,
D. A. Gewirtz, D. R. Green, M. Hansen, M. Jäättelä, T. Johansen, G. Juhász, V. Karantza,
C. Kraft, G. Kroemer, N. T. Ktistakis, S. Kumar, C. Lopez-­Otin, K. F. Macleod, F. Madeo,
J. Martinez, A. Meléndez, N. Mizushima, C. Münz, J. M. Penninger, R. M. Perera,
M. Piacentini, F. Reggiori, D. C. Rubinsztein, K. M. Ryan, J. Sadoshima, L. Santambrogio,
L. Scorrano, H.-­U. Simon, A. K. Simon, A. Simonsen, A. Stolz, N. Tavernarakis, S. A. Tooze,
T. Yoshimori, J. Yuan, Z. Yue, Q. Zhong, L. Galluzzi, F. Pietrocola, Autophagy in major
human diseases. EMBO J. 40, e108863 (2021).
6. Y. Ichimura, T. Kirisako, T. Takao, Y. Satomi, Y. Shimonishi, N. Ishihara, N. Mizushima,
I. Tanida, E. Kominami, M. Ohsumi, T. Noda, Y. Ohsumi, A ubiquitin-­like system mediates
protein lipidation. Nature 408, 488–492 (2000).
7. Y. Kabeya, N. Mizushima, T. Ueno, A. Yamamoto, T. Kirisako, T. Noda, E. Kominami,
Y. Ohsumi, T. Yoshimori, LC3, a mammalian homologue of yeast Apg8p, is localized in
autophagosome membranes after processing. EMBO J. 19, 5720–5728 (2000).
8. Å. B. Birgisdottir, T. Lamark, T. Johansen, The LIR motif—Crucial for selective autophagy.
J. Cell Sci. 126, 3237–3247 (2013).
9. K. Tsuboyama, I. Koyama-­Honda, Y. Sakamaki, M. Koike, H. Morishita, N. Mizushima, The
ATG conjugation systems are important for degradation of the inner autophagosomal
membrane. Science 354, 1036–1041 (2016).
10. T. N. Nguyen, B. S. Padman, J. Usher, V. Oorschot, G. Ramm, M. Lazarou, Atg8 family LC3/
GABARAP proteins are crucial for autophagosome-­lysosome fusion but not
autophagosome formation during PINK1/Parkin mitophagy and starvation. J. Cell Biol.
215, 857–874 (2016).
11. L. Cappadocia, C. D. Lima, Ubiquitin-­like protein conjugation: Structures, chemistry, and
mechanism. Chem. Rev. 118, 889–918 (2018).
12. K. Baek, D. C. Scott, B. A. Schulman, NEDD8 and ubiquitin ligation by cullin-­RING E3
ligases. Curr. Opin. Struct. Biol. 67, 101–109 (2021).
13. D. J. Klionsky, B. A. Schulman, Dynamic regulation of macroautophagy by distinctive
ubiquitin-­like proteins. Nat. Struct. Mol. Biol. 21, 336–345 (2014).
14. Y. Zheng, Y. Qiu, C. R. R. Grace, X. Liu, D. J. Klionsky, B. A. Schulman, A switch element in
the autophagy E2 Atg3 mediates allosteric regulation across the lipidation cascade. Nat.
Commun. 10, 3600 (2019).
15. S. Nath, J. Dancourt, V. Shteyn, G. Puente, W. M. Fong, S. Nag, J. Bewersdorf, A. Yamamoto,
B. Antonny, T. J. Melia, Lipidation of the LC3/GABARAP family of autophagy proteins relies
on a membrane-­curvature-­sensing domain in Atg3. Nat. Cell Biol. 16, 415–424 (2014).
16. A. Kuma, M. Hatano, M. Matsui, A. Yamamoto, H. Nakaya, T. Yoshimori, Y. Ohsumi,
T. Tokuhisa, N. Mizushima, The role of autophagy during the early neonatal starvation
period. Nature 432, 1032–1036 (2004).
17. C. Chang, X. Shi, L. E. Jensen, A. L. Yokom, D. Fracchiolla, S. Martens, J. H. Hurley,
Reconstitution of cargo-­induced LC3 lipidation in mammalian selective autophagy. Sci.
Adv. 7, eabg4922 (2021).
18. L. M. Strong, C. Chang, J. F. Riley, C. A. Boecker, T. G. Flower, C. Z. Buffalo, X. Ren,
A. K. H. Stavoe, E. L. F. Holzbaur, J. H. Hurley, Structural basis for membrane recruitment of
ATG16L1 by WIPI2 in autophagy. eLife 10, (2021).
19. N. Fujita, T. Itoh, H. Omori, M. Fukuda, T. Noda, T. Yoshimori, The Atg16L complex specifies
the site of LC3 lipidation for membrane biogenesis in autophagy. Mol. Biol. Cell 19,
2092–2100 (2008).
20. N. Mizushima, T. Noda, T. Yoshimori, Y. Tanaka, T. Ishii, M. D. George, D. J. Klionsky,
M. Ohsumi, Y. Ohsumi, A protein conjugation system essential for autophagy. Nature 395,
395–398 (1998).
21. Z. Metlagel, C. Otomo, G. Takaesu, T. Otomo, Structural basis of ATG3 recognition by the
autophagic ubiquitin-­like protein ATG12. Proc. Natl. Acad. Sci. U.S.A. 110, 18844–18849 (2013).
22. H. C. Dooley, M. Razi, H. E. J. Polson, S. E. Girardin, M. I. Wilson, S. A. Tooze, WIPI2 links LC3
conjugation with PI3P, autophagosome formation, and pathogen clearance by recruiting
Atg12–5-­16L1. Mol. Cell 55, 238–252 (2014).
23. T. Proikas-­Cezanne, Z. Takacs, P. Dönnes, O. Kohlbacher, WIPI proteins: Essential PtdIns3P
effectors at the nascent autophagosome. J. Cell Sci. 128, 207–217 (2015).
24. J. H. Hurley, L. N. Young, Mechanisms of autophagy initiation. Annu. Rev. Biochem. 86,
225–244 (2017).
25. L. E. Jensen, S. Rao, M. Schuschnig, A. K. Cada, S. Martens, G. Hummer, J. H. Hurley,
Membrane curvature sensing and stabilization by the autophagic LC3 lipidation
machinery. Sci. Adv. 8, eadd1436 (2022).
26. A. H. Lystad, S. R. Carlsson, L. R. de la Ballina, K. J. Kauffman, S. Nag, T. Yoshimori, T. J. Melia,
A. Simonsen, Distinct functions of ATG16L1 isoforms in membrane binding and LC3B
lipidation in autophagy-­related processes. Nat. Cell Biol. 21, 372–383 (2019).
27. X. Gong, Y. Wang, Y. Tang, Y. Wang, M. Zhang, M. Li, Y. Zhang, L. Pan, ATG16L1 adopts a
dual-­binding site mode to interact with WIPI2b in autophagy. Sci. Adv. 9, eadf0824
(2023).
28. A. Lupas, M. Van Dyke, J. Stock, Predicting coiled coils from protein sequences. Science
252, 1162–1164 (1991).
10 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
29. C. W. Wood, D. N. Woolfson, CCBuilder 2.0: Powerful and accessible coiled-­coil modeling.
Protein Sci. 27, 103–111 (2018).
30. F. H. C. Crick, The packing of α-­helices: Simple coiled-­coils. Acta Crystallogr. 6, 689–697
(1953).
31. J. Metje-­Sprink, J. Groffmann, P. Neumann, B. Barg-­Kues, R. Ficner, K. Kühnel, A. M. Schalk,
B. Binotti, Crystal structure of the Rab33B/Atg16L1 effector complex. Sci. Rep. 10, 12956
(2020).
32. S. Pantoom, G. Konstantinidis, S. Voss, H. Han, O. Hofnagel, Z. Li, Y.-­W. Wu, RAB33B recruits
the ATG16L1 complex to the phagophore via a noncanonical RAB binding protein.
Autophagy 17, 2290–2304 (2021).
33. J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool,
R. Bates, A. Žídek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie,
B. Romera-­Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy,
M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver,
O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, D. Hassabis, Highly accurate protein
structure prediction with AlphaFold. Nature 596, 583–589 (2021).
34. R. Evans, M. O’Neill, A. Pritzel, N. Antropova, A. Senior, T. Green, A. Žídek, R. Bates,
S. Blackwell, J. Yim, O. Ronneberger, S. Bodenstein, M. Zielinski, A. Bridgland,
A. Potapenko, A. Cowie, K. Tunyasuvunakool, R. Jain, E. Clancy, P. Kohli, J. Jumper,
D. Hassabis, Protein complex prediction with AlphaFold-­Multimer. bioRxiv 463034
[Preprint]. (2022). https://doi.org/10.1101/2021.10.04.463034.
35. Y. Ye, E. R. Tyndall, V. Bui, M. C. Bewley, G. Wang, X. Hong, Y. Shen, J. M. Flanagan,
H.-­G. Wang, F. Tian, Multifaceted membrane interactions of human Atg3 promote
LC3-­phosphatidylethanolamine conjugation during autophagy. Nat. Commun. 14, 5503
(2023).
36. H. Popelka, V. N. Uversky, D. J. Klionsky, Identification of Atg3 as an intrinsically disordered
polypeptide yields insights into the molecular dynamics of autophagy-­related proteins in
yeast. Autophagy 10, 1093–1104 (2014).
37. J. Farnung, M. Muhar, J. R. Liang, K. A. Tolmachova, R. M. Benoit, J. E. Corn, J. W. Bode,
Semisynthetic LC3 probes for autophagy pathways reveal a noncanonical LC3 interacting
region motif crucial for the enzymatic activity of human ATG3. ACS Cent. Sci. 9,
1025–1034 (2023).
38. C. Otomo, Z. Metlagel, G. Takaesu, T. Otomo, Structure of the human ATG12~ATG5
conjugate required for LC3 lipidation in autophagy. Nat. Struct. Mol. Biol. 20, 59–66
(2013).
39. L. Li, M. Tong, Y. Fu, F. Chen, S. Zhang, H. Chen, X. Ma, D. Li, X. Liu, Q. Zhong, Lipids and
membrane-­associated proteins in autophagy. Protein Cell 12, 520–544 (2021).
40. S. Baskaran, M. J. Ragusa, E. Boura, J. H. Hurley, Two-­site recognition of
phosphatidylinositol 3-­phosphate by PROPPINs in autophagy. Mol. Cell 47, 339–348
(2012).
41. R. Krick, R. A. Busse, A. Scacioc, M. Stephan, A. Janshoff, M. Thumm, K. Kühnel, Structural
and functional characterization of the two phosphoinositide binding sites of PROPPINs, a
β-­propeller protein family. Proc. Natl. Acad. Sci. U.S.A. 109, E2042–E2049 (2012).
42. R. Liang, J. Ren, Y. Zhang, W. Feng, Structural conservation of the two phosphoinositide-­
binding sites in WIPI proteins. J. Mol. Biol. 431, 1494–1505 (2019).
43. D. Fracchiolla, C. Chang, J. H. Hurley, S. Martens, A PI3K-­WIPI2 positive feedback loop
allosterically activates LC3 lipidation in autophagy. J. Cell Biol. 219, e201912098
(2020).
44. T. Hanada, Y. Satomi, T. Takao, Y. Ohsumi, The amino-­terminal region of Atg3 is essential
for association with phosphatidylethanolamine in Atg8 lipidation. FEBS Lett. 583,
1078–1083 (2009).
45. Y. Ichimura, Y. Imamura, K. Emoto, M. Umeda, T. Noda, Y. Ohsumi, In vivo and in vitro
reconstitution of Atg8 conjugation essential for autophagy. J. Biol. Chem. 279,
40584–40592 (2004).
46. T. Nishimura, G. Lazzeri, N. Mizushima, R. Covino, S. A. Tooze, Unique amphipathic α helix
drives membrane insertion and enzymatic activity of ATG3. Advances 9, eadh1281 (2023).
47. Y. Yamada, N. N. Suzuki, T. Hanada, Y. Ichimura, H. Kumeta, Y. Fujioka, Y. Ohsumi, F. Inagaki,
The crystal structure of Atg3, an autophagy-­related ubiquitin carrier protein (E2) enzyme
that mediates Atg8 lipidation. J. Biol. Chem. 282, 8036–8043 (2007).
48. T. Hanada, N. N. Noda, Y. Satomi, Y. Ichimura, Y. Fujioka, T. Takao, F. Inagaki, Y. Ohsumi, The
Atg12-­Atg5 conjugate has a novel E3-­like activity for protein lipidation in autophagy.
J. Biol. Chem. 282, 37298–37302 (2007).
49. K. Oh-­oka, H. Nakatogawa, Y. Ohsumi, Physiological pH and acidic phospholipids
contribute to substrate specificity in lipidation of Atg8. J. Biol. Chem. 283, 21847–21852
(2008).
50. M. Sakoh-­Nakatogawa, K. Matoba, E. Asai, H. Kirisako, J. Ishii, N. N. Noda, F. Inagaki,
H. Nakatogawa, Y. Ohsumi, Atg12-­Atg5 conjugate enhances E2 activity of Atg3 by
rearranging its catalytic site. Nat. Struct. Mol. Biol. 20, 433–439 (2013).
51. S. E. Kaiser, K. Mao, A. M. Taherbhoy, S. Yu, J. L. Olszewski, D. M. Duda, I. Kurinov, A. Deng,
T. D. Fenn, D. J. Klionsky, B. A. Schulman, Noncanonical E2 recruitment by the autophagy
E1 revealed by Atg7-­Atg3 and Atg7-­Atg10 structures. Nat. Struct. Mol. Biol. 19, 1242–1249
(2012).
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
52. J. Durgan, A. H. Lystad, K. Sloan, S. R. Carlsson, M. I. Wilson, E. Marcassa, R. Ulferts,
J. Webster, A. F. Lopez-­Clavijo, M. J. Wakelam, R. Beale, A. Simonsen, D. Oxley, O. Florey,
Non-­canonical autophagy drives alternative ATG8 conjugation to phosphatidylserine.
Mol. Cell 81, 2031–2040.e8 (2021).
53. M. Yamaguchi, K. Matoba, R. Sawada, Y. Fujioka, H. Nakatogawa, H. Yamamoto,
Y. Kobashigawa, H. Hoshida, R. Akada, Y. Ohsumi, N. N. Noda, F. Inagaki, Noncanonical
recognition and UBL loading of distinct E2s by autophagy-­essential Atg7. Nat. Struct. Mol.
Biol. 19, 1250–1256 (2012).
54. H. E. J. Polson, J. de Lartigue, D. J. Rigden, M. Reedijk, S. Urbé, M. J. Clague, S. A. Tooze,
Mammalian Atg18 (WIPI2) localizes to omegasome-­anchored phagophores and
positively regulates LC3 lipidation. Autophagy 6, 506–522 (2010).
55. A. E. Stanton, F. M. Hughson, The machinery of vesicle fusion. Curr. Opin. Cell Biol. 83,
102191 (2023).
56. Y. Ye, E. R. Tyndall, V. Bui, Z. Tang, Y. Shen, X. Jiang, J. M. Flanagan, H.-­G. Wang, F. Tian, An
N-­terminal conserved region in human Atg3 couples membrane curvature sensitivity to
conjugase activity during autophagy. Nat. Commun. 12, 374 (2021).
57. J. H. Kim, S. B. Hong, J. K. Lee, S. Han, K. H. Roh, K. E. Lee, Y. K. Kim, E. J. Choi, H. K. Song,
Insights into autophagosome maturation revealed by the structures of ATG5 with its
interacting partners. Autophagy 11, 75–87 (2015).
58. W. L. Delano, PyMOL: An open-­source molecular graphics tool. CCP4 Newsletter On Protein
Crystallography 40, 82–92 (2002).
59. A. Waterhouse, M. Bertoni, S. Bienert, G. Studer, G. Tauriello, R. Gumienny, F. T. Heer,
T. A. P. De Beer, C. Rempfer, L. Bordoli, R. Lepore, T. Schwede, SWISS-­MODEL: Homology
modelling of protein structures and complexes. Nucleic Acids Res. 46, W296–W303 (2018).
60. C. Combet, C. Blanchet, C. Geourjon, G. Deléage, NPS@: Network protein sequence
analysis. Trends Biochem. Sci. 25, 147–150 (2000).
61. X. Zhou, J. Hu, C. Zhang, G. Zhang, Y. Zhang, Assembling multidomain protein structures
through analogous global structural alignments. Proc. Natl. Acad. Sci. U.S.A. 116,
15930–15938 (2019).
62. S. Jo, T. Kim, V. G. Iyer, W. Im, CHARMM-­GUI: A web-­based graphical user interface for
CHARMM. J. Comput. Chem. 29, 1859–1865 (2008).
63. J. Lee, X. Cheng, J. M. Swails, M. S. Yeom, P. K. Eastman, J. A. Lemkul, S. Wei, J. Buckner,
J. C. Jeong, Y. Qi, S. Jo, V. S. Pande, D. A. Case, C. L. Brooks, A. D. MacKerell, J. B. Klauda,
W. Im, CHARMM-­GUI input generator for NAMD, GROMACS, AMBER, OpenMM, and
CHARMM/OpenMM simulations using the CHARMM36 additive force field. J. Chem.
Theory. Comput. 12, 405–413 (2016).
64. N. Fujita, T. Saitoh, S. Kageyama, S. Akira, T. Noda, T. Yoshimori, Differential involvement of
Atg16L1 in Crohn disease and canonical autophagy: Analysis of the organization of the
Atg16L1 complex in fibroblasts. J. Biol. Chem. 284, 32602–32609 (2009).
65. M. H. M. Olsson, C. R. Søndergaard, M. Rostkowski, J. H. Jensen, PROPKA3: Consistent
treatment of internal and surface residues in empirical pKa predictions. J. Chem. Theory
Comput. 7, 525–537 (2011).
66. M. J. Abraham, T. Murtola, R. Schulz, S. Páll, J. C. Smith, B. Hess, E. Lindahl, GROMACS: High
performance molecular simulations through multi-­level parallelism from laptops to
supercomputers. SoftwareX 1-­2, 19–25 (2015).
67. J. Huang, S. Rauscher, G. Nawrocki, T. Ran, M. Feig, B. L. De Groot, H. Grubmüller,
A. D. MacKerell, CHARMM36m: An improved force field for folded and intrinsically
disordered proteins. Nat. Methods 14, 71–73 (2017).
68. T. A. Wassenaar, H. I. Ingólfsson, R. A. Böckmann, D. P. Tieleman, S. J. Marrink,
Computational lipidomics with insane: A versatile tool for generating custom
membranes for molecular simulations. J. Chem. Theory Comput. 11, 2144–2155 (2015).
69. J. D. Durrant, R. E. Amaro, LipidWrapper: An algorithm for generating large-­scale
membrane models of arbitrary geometry. PLoS Comput. Biol. 10, e1003720 (2014).
70. O. N. Vickery, P. J. Stansfeld, CG2AT2: An enhanced fragment-­based approach for serial
multi-­scale molecular dynamics simulations. J. Chem. Theory Comput. 17, 6472–6482
(2021).
71. G. Bussi, D. Donadio, M. Parrinello, Canonical sampling through velocity rescaling.
J. Chem. Phys. 126, 014101 (2007).
72. M. Parrinello, A. Rahman, Polymorphic transitions in single crystals: A new molecular
dynamics method. J. Appl. Phys. 52, 7182–7190 (1981).
73. T. Darden, D. York, L. Pedersen, Particle mesh Ewald: AnN⋅log(N) method for Ewald sums
in large systems. J. Chem. Phys. 98, 10089–10092 (1993).
74. U. Essmann, L. Perera, M. L. Berkowitz, T. Darden, H. Lee, L. G. Pedersen, A smooth particle
mesh Ewald method. J. Chem. Phys. 103, 8577–8593 (1995).
75. B. Hess, H. Bekker, H. J. C. Berendsen, J. G. E. M. Fraaije, LINCS: A linear constraint solver for
molecular simulations. J. Comput. Chem. 18, 1463–1472 (1997).
76. N. Michaud-­Agrawal, E. J. Denning, T. B. Woolf, O. Beckstein, MDAnalysis: A toolkit for the
analysis of molecular dynamics simulations. J. Comput. Chem. 32, 2319–2327 (2011).
77. R. Gowers, M. Linke, J. Barnoud, T. Reddy, M. Melo, S. Seyler, J. Domański, D. Dotson,
S. Buchoux, I. Kenney, O. Beckstein, MDAnalysis: A Python package for the rapid analysis
of molecular dynamics simulations, in Proceedings of the 15th Python in Science
Conference, S. Benthall, S. Rostrup, Eds. (SciPy, 2016), pp. 98–105.
11 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
78. W. W.-­Y. Yim, H. Yamamoto, N. Mizushima, A pulse-­chasable reporter processing assay for
mammalian autophagic flux with HaloTag. eLife 11, e78923 (2022).
Acknowledgments: We thank members of the Aligning Science Across Parkinson’s Team
mito911 for advice and discussions. Funding: This work was supported by The Michael J. Fox
Foundation for Parkinson’s Research (MJFF) and Aligning Science Across Parkinson’s (ASAP)
initiative. MJFF administers the grant ASAP-­000350 on behalf of ASAP and itself (to J.H.H., M.L.,
and G.H.). This work was also supported by the Max Planck Society (to S.R. and G.H). Author
contributions: Conceptualization: L.M.S., M.L., J.H.H., and G.H. Methodology: S.R., M.S., L.M.S.,
X.R., M.L., and G.H. Investigation: S.R., M.S., L.M.S., and X.R. Formal analysis: S.R. and M.S.
Visualization: S.R., M.S., and L.M.S. Validation: S.R., M.S., L.M.S., X.R., and M.L. Data curation: M.S.
Resources: M.S., M.L., and G.H. Supervision: M.L., J.H.H., and G.H. Project administration: J.H.H.
and G.H. Funding acquisition: M.L., J.H.H., and G.H. Writing—original draft: S.R., L.M.S., J.H.H.,
Rao et al., Sci. Adv. 10, eadj8027 (2024) 7 February 2024
and G.H. Writing—review and editing: G.H. Competing interests: J.H.H. is a cofounder and
shareholder of Casma Therapeutics and receives research funding from Genentech and
Hoffmann-­La Roche. M.L. is a cofounder and member of the scientific advisory board of
Automera. All other authors declare that they have no competing interests. Data and
materials availability: Full data from MD simulations (10.5281/zenodo.8083723), in vitro LC3
lipidation assays (10.5281/zenodo.10091729), and HaloTag-­LC3B starvation assays (10.5281/
zenodo.10081243) have been uploaded at Zenodo. All data needed to evaluate the
conclusions in the paper are present in the paper and/or the Supplementary Materials.
Submitted 17 July 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adj8027
12 of 12
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

CANCER Copyright © 2024 The

Authors, some rights
Rational design of ICD-­inducing nanoparticles for reserved; exclusive
licensee American
cancer immunotherapy Association for the
Advancement of
Science. No claim to
Zhanzhan Zhang1,3†, Zheng Pan2,3†, Qiushi Li2,3, Qingqing Huang2,3, Linqi Shi2,3, Yang Liu1,2,4* original U.S.
Government Works.
Nanoparticle-­based cancer immunotherapy has shown promising therapeutic potential in clinical settings. How- Distributed under a
ever, current research mainly uses nanoparticles as delivery vehicles but overlooks their potential to directly mod- Creative Commons
ulate immune responses. Inspired by the endogenous endoplasmic reticulum (ER) stress caused by unfolded/ Attribution License 4.0
misfolded proteins, we present a rationally designed immunogenic cell death (ICD) inducer named NanoICD, (CC BY).
which is a nanoparticle engineered for ER targeting and retention. By carefully controlling surface composition
and properties, we have obtained NanoICD that can effectively accumulate in the ER, induce ER stress, and acti-
vate ICD-­associated immune responses. In addition, NanoICD is generally applicable to various proteins and
enzymes to further enhance the immunomodulatory capacity, exemplified by encapsulating catalase (CAT) to
obtain NanoICD/CAT, effectively alleviated immunosuppressive tumor microenvironment and induced robust an-
titumor immune responses in 4T1-­bearing mice. This work demonstrates engineered nanostructures’ potential to
autonomously regulate biological processes and provides insights into the development of advanced nanomedi-
cines for cancer treatment.

INTRODUCTION immunity, thereby transforming a “cold” tumor into an immuno-

Cancer immunotherapy has changed the paradigm of cancer treat-
ment by mobilizing the host immune system to recognize and
destroy cancer cells (1, 2). Recently, nanoparticles have been in-
creasingly used in cancer immunotherapy due to their abilities to
optimize biodistribution, improve circulation stability, and reduce
the side effects of therapeutic agents (1, 3). To date, most current
research on nano-­immunotherapies uses nanoparticles as delivery
vehicles. The design of nanoparticles usually focuses on prolonging
circulation time and improving delivery efficiency but largely ig-
nores the potential of nanoparticles to directly regulate the immune
system (4–6). Recently, some serendipitous discoveries have sug-
gested that certain solid nanoparticles or lipid nanoparticles with
specific chemical structures, composition, or formulations are di-
rectly involved in immune regulation (7–12). However, rationally
designed nanoparticles with immunomodulatory capacity have not
been reported so far. Nanoparticles are a special class of materials
with unique particle sizes and highly engineerable surfaces (13).
These properties allow tailor-­made nanoparticles to target specific
tissues, cells, and even organelles (14–16). Moreover, the binding
affinity between the nanoparticle and its target can be finely con-
trolled by tuning nano-­multivalent interactions, thus offering
nanoparticles great potential to modulate immune systems (17, 18).
Currently, most cancer immunotherapies are designed on the ba-
sis of the cancer immune cycle, which is a self-­propagating process
that elicits effective antitumor immune responses (19). However,
low immune cell infiltration and immunosuppressive networks in
tumor microenvironment (TME) render tumors less immunogenic
and severely suppress host immune responses (20). Recent studies
have shown that neoplastic cells undergoing immunogenic cell
death (ICD) exert a vaccine-­like function to generate antitumor
1
College of Chemistry, Key Laboratory of Functional Polymer Materials (Ministry of
Education), Nankai University, Tianjin 300071, China. 2State Key Laboratory of Medicinal
Chemical Biology, Nankai University, Tianjin 300071, China. 3School of Medical
Imaging, Tianjin Medical University, Tianjin 300203, China. 4Frontiers Science Center
for New Organic Matter Nankai University, Tianjin 300071, China.
*Corresponding author. Email: yliu@​nankai.​edu.​cn
†These authors contributed equally to this work.
genic “hot” tumor (21–23). ICD is considered a stress-­induced pro-
cess in which endoplasmic reticulum (ER) stress is required for the
release of tumor-­associated antigens (TAAs) and damage-­associated
molecular patterns (DAMPs) from tumor cells to provide antigenic-
ity and adjuvanticity, respectively (24–26). Recently, several clinical
trials have shown that pretreatment with ICD inducers notably im-
proves the response rate and survival of immunotherapies based on
checkpoint blockade (27). To date, many ICD induction strategies
have been developed, including treatment with chemical drugs,
such as anthracyclines, chemical protein phosphatase 1/GADD34
inhibitors, etc., and physical induction strategies, such as photody-
namic therapy and radiotherapy (24, 28, 29). Recently, nanoparticle-­
based ICD inducers with high ICD-­inducing efficacy and low side
effects have emerged (26). However, most nanoparticle-­based ICD
inducers are chemical inducers or photosensitizers delivered by
nano-­sized drug carriers. In these strategies, the ICD is induced by
the payload, while the nanomaterial has no immunomodulatory
function.
Given the relationship between ER stress and ICD induction,
disturbing ER homeostasis is essential for successful ICD induction
(24, 30). In normal cells, endogenous signals such as the accumula-
tion of unfolded or misfolded proteins in the ER lumen can disturb
ER proteostasis and induce ER stress (31, 32). However, the ER sub-
tly relieves this stress by retro-­translocating these proteins to the
cytoplasm and activating proteasomes for degradation (31–33).
Mimicking this accumulation process but avoiding the subsequent
stress relief should be an effective strategy to induce ER stress.
Therefore, we hypothesize that synthetic nanoparticles with the abil-
ity to target and retain in ER can effectively induce ER stress and
activate ICD-­associated immune responses. As a proof of concept,
we present here a rationally designed ICD inducer based on
nanoparticles engineered for ER targeting and ER retention (denot-
ed as NanoICD, Fig. 1). NanoICD was synthesized by a protein na-
noencapsulation method (15, 34, 35), which involved in
situ–copolymerized neutral monomers [acrylamide (AAm)], posi-
tively charged monomers {N-­(3-­aminopropyl)-­methacrylamine

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 1 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 1. Schematic representation of the synthesis and the mechanism of NanoICD to induce ICD. (A) Schematic illustration of the synthesis of NanoICD; (B) the
mechanism by which NanoICD induces ER stress; (C) encapsulated CAT and functional polymer surface of NanoICD/CAT synergistically regulate immune system for cancer
treatment. NanoICD with strong surface interaction with ER hinders the retro-­translocation process and effectively induces ER stress.
[APm]} (36), ER-­targeting ligand {N-­(2-­((4-­methylphenyl)sulfon-
amido)ethyl)acrylamide [ETL]} (37), and the cross-­linkers [N,N′-­
methylenebisacrylamide (BIS)] on the surface of each protein
molecule (Fig. 1A). The introduction of APm provides a positively
charged surface for NanoICD, which ensures the effective uptake of
NanoICD by cancer cells and the subsequent escape from lyso-
somes. Meanwhile, the integration of multiple ETLs allows Na-
noICD to target and tightly bind to the ER after the lysosomal
escape. By controlling monomer compositions and formulation
methodologies, NanoICD efficiently accumulates in the ER, induces
ER stress, and eventually activates ICD-­associated immune respons-
es (Fig. 1B). In this work, immunologically inert bovine serum albu-
min (BSA) was first used for the nanoencapsulation (NanoICD/
BSA) to better understand the immunomodulatory function of the
polymer surface. Furthermore, the BSA can be replaced with other
proteins and enzymes to provide additional functions to cooperate
with the ICD-­inducing function of the polymer surface. As a dem-
onstration, catalase (CAT) was used to perform the nanoencapsula-
tion to synthesize NanoICD/CAT (Fig. 1C). In addition to the
induction of ICD of tumor cells, NanoICD/CAT effectively allevi-
ated hypoxic and inflammatory TME by decomposing hydrogen
peroxide (H2O2) to oxygen (O2), resulting in the relief of tumor im-
munosuppression and further enhancing the antitumor immune
responses (38–40). Together, these unique features make NanoICD
an innovative strategy for effective cancer immunotherapy. This
work demonstrates that a nanoparticle with a rationally designed
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
surface can directly modulate the immune system without the aid of
conventional immunomodulating drugs, providing new insights
into the design of advanced nanomedicines for cancer immu-
notherapy.
RESULTS
The relationship between the amount of ETL on NanoICD
and its ability to induce ICD
To achieve ER accumulation, nanoparticles must first enter the cyto-
plasm and then bind to the ER. For NanoICD, cell internalization
and ER targeting ability are achieved by introducing APm and ETL
during the polymerization to provide positive surface charge and
ER-­targeting groups, respectively. In particular, the amount of ETL
determines the mass of NanoICD bound to ER and ultimately af-
fects the efficiency of ICD induction. To investigate the relationship
between the amount of ETL on NanoICD and its efficiency in in-
ducing ICD, we synthesized a series of NanoICD/BSA (with similar
zeta potential, ~3 mV; details in table S2) containing different
amounts of ETL (denoted as NanoICD/BSA-­n, where n represents
the amount of ETL on each NanoICD). Since the pre-­apoptotic cal-
reticulin (CRT) exposure on the cell surface is the most relevant
DAMP for ICD (5, 30), the level of surface CRT exposure was used
as the standard criterion to evaluate the ability of these NanoICD/
BSA to induce ICD. In this study, all these NanoICD/BSA were in-
cubated with mouse melanoma cells (B16F10) for 12 hours.
2 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Phosphate-­ buffered saline (PBS)– and paclitaxel (PTX)–treated
cells were used as negative and positive controls, respectively. For
better demonstration, a free form of ETL and a nanoparticle similar
to NanoICD/BSA but without ETL (denoted as nBSA) were used as
comparative groups. After incubation, cells were collected and
stained with ATTO-­488-­CRT and propidium iodide (PI) for flow
cytometric analysis. As shown in Fig. 2A, negligible pre-­apoptotic
CRT exposure (gated on PI−CRT+) was observed from cells treated
with nBSA (1.55%), NanoICD/BSA-­ 10 (1.75%), and NanoICD/
BSA-­20 (2.91%) compared to the negative control (PBS-­treated
cells, 1.15%). In contrast, the cells treated with NanoICD/BSA-­30
(24%) and NanoICD/BSA-­40 (38.3%) exhibit a significant amount
of PI−CRT+ cells. This result indicates that high surface density of
ETL is required for NanoICD/BSA to effectively induce ER stress.
Moreover, the free form of ETL did not induce any CRT exposure
(0.36%), even at a much higher dosage. These results suggest that
strong multivalent binding to ER is essential for NanoICD/BSA to
effectively trigger ER stress and induce ICD.
Next, we investigated the mechanism underlying the ICD-­
inducing activity of NanoICD/BSA. As shown in figs. S6 and S7, all
these NanoICD/BSA-­​n were efficiently internalized by B16F10 cells
and escaped from endosomes within 2 hours. This result suggests
that the different ICD-­inducing performance of NanoICD/BSA-­​n
was not affected by their cell internalization ability but may be
caused by their different ER-­targeting and retention ability. For the
verification, the intracellular distribution of NanoICD/BSA-­​n was
observed by a confocal laser scanning microscope (CLSM) at differ-
ent time points. As shown in Fig. 2B, the nBSA and NanoICD/BSA-­
10 were mainly distributed in the cytoplasm, and very few of them
were observed in the ER after 8 hours of incubation. NanoICD/BSA-­
20–treated cells exhibited obvious colocalization of NanoICD/BSA-­
20 with ER at 4 hours, but the colocalization significantly decreased,
and the NanoICD/BSA-­20 were transported back to the cytoplasm
at 6 and 8 hours. Despite the successful targeting to the ER, Na-
noICD/BSA-­20 failed to survive from the retro-­translocation effect
of ER. Considering the ineffectiveness of NanoICD/BSA-­20 in in-
ducing pre-­apoptotic CRT exposure of cells, this result suggests that
long-­term retention and accumulation of the nanoparticles in ER
may be essential for the induction of ICD. This is confirmed by the
results from the cells treated with NanoICD/BSA-­30 and NanoICD/
BSA-­40, where the cells showed significant pre-­apoptotic CRT ex-
posure (Fig. 2A), and the nanoparticles were mainly distributed in
the ER, and their ER accumulation increased along with the incuba-
tion time. These results were further confirmed by quantitative anal-
ysis of the colocalization of the NanoICD/BSA-­​n with ER by using
Manders’ coefficients M2 (fraction of NanoICD/BSA overlapping
ER) (41). As shown in Fig. 2C, M2 of cells treated with nBSA and
NanoICD/BSA-­ 10 were lower than 0.6 at all four time points,

Fig. 2. The relationship between the amount of ETL/APm and the ICD-­inducing activity of NanoICD/BSA. (A) Pre-­apoptotic CRT exposure on cell (B16F10) surface
after the treatment with PBS, ETL, PTX, nBSA, and NanoICD/BSA-­​n (n = 10, 20, 30, and 40). (B) CLSM images showing the ER targeting and retention capabilities of nBSA
and NanoICD/BSA-­​n (n = 10, 20, 30, 40) in B16F10 cells. (C) Quantitative analysis of the colocalization of nBSA and NanoICD/BSA-­​n with ER. Manders’ coefficient M2 indi-
cates the fraction of NanoICD overlapping ER. (D) QCM-­D experiment showing the binding affinities between nBSA, NanoICD/BSA-­​n (n = 20, 40), and ER. (E) Pre-­apoptotic
CRT exposure on the cell (B16F10) surface after treatment with (Z)-­NanoICD/BSA-­40 (Z = −2, 0, 1.5, 3, and 6 mV). (F) Apoptosis of B16F10 cells after treatment with (Z)-­
NanoICD/BSA-­40 (Z = −2, 0, 1.5, 3, and 6 mV). Scale bar, 10 μm.

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 3 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
suggesting that low surface density of ETL (<10 per nanoparticle)
is insufficient for NanoICD to achieve effective ER targeting.
NanoICD/BSA-­20–treated cells exhibited a much higher M2 (0.82) at
4 hours but gradually decreased to 0.49 after 8 hours of incubation,
suggesting that moderate density of ETL (~20 per nanoparticle) en-
hanced the interaction of NanoICD with ER, but the interaction was
not strong enough to overcome the retro-­translocation to achieve
long-­term retention and accumulation in ER. Of note, the M2 of
cells treated with NanoICD/BSA-­30 (0.86 at 6 hours, 0.90 at 8 hours)
and NanoICD/BSA-­40 (0.86 at 6 hours, 0.95 at 8 hours) was signifi-
cantly higher than other treatments and increased along with the
incubation time, suggesting that a high density of ETL (>30 per
nanoparticle) is required for NanoICD/BSA to effectively interact
with ER, overcome the retro-­translocation to be retained in ER, and
eventually induce ICD of the cell.
To verify this hypothesis, we then determined the mass of Na-
noICD/BSA bound to ER using a quartz crystal microbalance with
dissipation monitoring (QCM-­ D) (42). Three NanoICD/BSA-­​ n To further investigate the ICD-­inducing activity of NanoICD/
with different amounts of ETL, including 0 (nBSA), 20 (NanoICD/
BSA-­20), and 40 (NanoICD/BSA-­40), were used in this assay. Brief-
ly, nBSA and NanoICD/BSA (100 μg/ml) were first conjugated to
the Au sensor chip via Traut’s reagent and washed with PBS buffers
before exposure to freshly extracted ER (details in the Supplemen-
tary Materials). After the binding reached equilibrium, the flow
phase was replaced with PBS to simulate the retro-­translocation
process of the ER. The frequency shifts of QCM-­D were continuously
monitored, and the results are summarized in Fig. 2D. According to
the results, NanoICD/BSA-­40 (ΔF = 149.85 Hz) exhibited a signifi-
cantly higher maximum frequency shift than that of NanoICD/BSA-­
20 (ΔF = 110.14 Hz) and nBSA (ΔF = 91.68 Hz), indicating that the
high density of ETL confers a strong binding between NanoICD/
BSA-­40 and the ER. When rinsed with PBS, only a slight change in
frequency shift was observed in NanoICD/BSA-­ 40 groups
(ΔF = 130.8 Hz, ΔΔF = 19.05 Hz) compared to NanoICD/BSA-­20
(ΔF = 68.2 Hz, ΔΔF = 41.94 Hz) and nBSA groups (ΔF = 7 Hz,
ΔΔF = 84.67 Hz), indicating that the strong binding of NanoICD/
BSA-­40 offers the potential to overcome the retro-­translocation to
retain in the ER.
The relationship between the amount of APm on NanoICD
and its ability to induce ICD
Zeta potential is another factor that affects the ICD-­inducing ability
of NanoICD. By increasing the amount of APm during the polymer-
ization or neutralizing APm with succinic anhydride (SA) (20), we
obtained a series of NanoICD/BSA with similar ETL amount (~40)
but different zeta potentials [ranging from −2 to +6 mV, denoted as
(Z)-­NanoICD/BSA, where Z represents the zeta potential of each
NanoICD/BSA; details in table S1]. The ICD-­inducing capabilities
of these (Z)-­NanoICD were evaluated by flow cytometric analysis.
As shown in Fig. 2E, a significantly reduced population of PI−CRT+
cells were detected in the cells treated with (+1.5)-­NanoICD/BSA
(14.2%) compared to (+3)-­NanoICD/BSA (38.3%). In addition, the
amount of PI−CRT+ cells were almost undetectable when the cells
were treated with (0)-­NanoICD/BSA (2.33%) and (−2)-­NanoICD/
BSA (1.41%). The significantly reduced ICD-­inducing capacity of
NanoICD/BSA may be attributed to its reduced cellular uptake effi-
ciency as the decrease in zeta potential (fig. S6), indicating that ef-
fective cellular uptake is a prerequisite for NanoICD/BSA to induce
ICD. However, although (+6)-­NanoICD/BSA showed the highest
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
cellular uptake, only a small amount of PI−CRT+ cells were detected
from the cells treated with (+6)-­NanoICD/BSA (5.55%). In addi-
tion, (+6)-­NanoICD/BSA exhibited much higher cytotoxicity
than (+3)-­NanoICD/BSA (fig. S8). These results imply that
(+6)-­NanoICD/BSA might behave differently from (+3)-­NanoICD/
BSA after cellular uptake. To investigate, we next performed an an-
nexin V (AV)–PI–based apoptosis assay on the NanoICD/BSA-­
treated cells. As shown in Fig. 2F, the amount of late apoptotic cells
(AV+PI+) was significantly increased from 30.7% of (+3)-­NanoICD/
BSA to 66.1% of (+6)-­NanoICD/BSA, indicating that (+6)-­NanoICD/
BSA induced cells to undergo apoptosis rather than ICD. Therefore,
an appropriate zeta potential (~3 mV) to allow efficient cellular up-
take and endosomal escape while avoiding excessive cytotoxicity is
critical for NanoICD/BSA to effectively activate ICD-­associated im-
mune response. In the following, NanoICD/BSA will be explicitly
referred to as NanoICD/BSA with surface ETL groups of ~40 and a
zeta potential of ~3 mV, unless otherwise noted.
BSA, CLSM-­based analysis was performed to directly observe the
pre-­apoptotic exposure of CRT on the cell surface after treatment.
PBS, PTX, ETL, and nBSA were used as comparison groups for bet-
ter demonstration. As shown in Fig. 3A, both PTX-­and NanoICD/
BSA-­treated cells showed significant surface CRT exposure. In con-
trast, negligible CRT exposure was observed in cells treated with
PBS, ETL, and nBSA, which is consistent with the results of the flow
cytometric analysis (Fig. 2A). In addition to surface CRT exposure,
two other important indicators of ICD, including the secretion of
adenosine triphosphate (ATP) and the post-­apoptotic exodus of
high mobility group box 1 (HMGB-­1), were also evaluated (24). As
shown in Fig. 3B, significantly higher levels of ATP were detected in
cellular supernatants after treatment with NanoICD/BSA and PTX
than those of PBS, ETL, and nBSA. CLSM-­and enzyme-­linked im-
munosorbent assay (ELISA)–based analyses were then performed
to investigate the intracellular and extracellular levels of HMGB-­1
after different treatments. As shown in Fig. 3 (C and D), the decrease
in intracellular HMGB-­1 levels and increase in extracellular HMGB-
­1 levels were observed from the cells treated with PTX and Na-
noICD/BSA, indicating the exodus of HMGB-­1 after treatment.
Together, the observation of these three important indicators of ICD
suggests the great potential of NanoICD/BSA as an effective ICD
inducer.
The potential mechanism of NanoICD/BSA to induce ICD
Next, we investigated the mechanism behind this ICD-­inducing
ability of NanoICD/BSA. Since NanoICD/BSA that can effectively
trigger ICD-­associated immune responses showed efficient ER tar-
geting and retention, we hypothesized that the long-­term retention
of NanoICD/BSA in the ER might induce ER stress and the phos-
phorylation of eukaryotic translation initiation factor 2α (EIF2α).
For verification, the phosphorylation of EIF2α (pEIF2α) in cells was
evaluated by flow cytometry–based analysis after different treat-
ments. As shown in Fig. 3E, negligible changes in pEIF2α levels were
observed from the cells treated with ETL (0.28%) and nBSA (0.18%).
In contrast, pEIF2α levels were significantly up-­regulated in PTX
(7.79%)– and NanoICD/BSA-­treated cells (28.8%), suggesting that
NanoICD effectively induced ER stress. Similar results were ob-
served from the Western blot–based analysis (Fig. 3F), where
NanoICD/BSA-­treated cells showed the highest levels of pEIF2α. Last,
these samples were subjected to correlation analysis between pEIF2α
4 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. The potential mechanism of NanoICD/BSA to induce ICD of tumor cells. (A) CLSM images show the exposure of CRT on B16F10 surface after treatment with
PBS, PTX, ETL, nBSA, and NanoICD/BSA. (B) ATP levels in cellular supernatants after different treatments. (C) The exodus of HMGB-­1 from cell nuclei after different treat-
ments. (D) Extracellular concentrations of HMGB-­1 after treating the cells with PBS, PTX, ETL, nBSA, and NanoICD/BSA. (E) pEIF2α levels in B16F10 cells after treatment with
PBS, PTX, ETL, nBSA, and NanoICD/BSA. (F) Western blot analysis of the expression of β-­actin, EIF2α, and pEIF2α in B16F10 cells after different treatments. (G) The correla-
tions between pEIF2α levels and pro-­apoptotic CRT on cell surface after the treatment. (H) Heatmap and hierarchical clustering analysis of DEGs of B16F10 after different
treatments. (I) Top 30 enriched pathways in GO analysis in NanoICD/BSA-­treated B16F10 cells. (J) GSEA shows the up-­regulation of protein processing in ER, response to
unfolded protein, the ER UPR, and antigen processing and presentation pathways in NanoICD/BSA-­treated B16F10 cells. Data are presented as means ± SD from three
independent experiments (biological replicates, n = 3). Scale bars, 20 μm.

levels and surface CRT exposure, which revealed a strong positive
linear correlation (Pearson’s r = 0.9986, R2 = 0.9973) (Fig. 3G).
These results suggest that NanoICD/BSA induces ICD-­associated
immunogenicity by promoting ER stress and pEIF2α.
In addition, we also performed whole RNA sequencing to com-
pare gene expression levels in B16F10 cells after NanoICD/BSA,
nBSA, and PBS treatment for 24 hours. As shown in figs. S9 to S11,
principal components analysis revealed that differential gene ex-
pressions were closely correlated with the different treatments, indi-
cating that nanoparticles with different surface properties could
modulate the gene expression of B16F10. Further analysis of dif-
ferentially expressed genes (DEGs) revealed that 450 genes were
significantly up-­regulated, and 484 genes were significantly
down-­regulated in NanoICD/BSA-­treated cells compared to those
treated with nBSA (Fig. 3H and fig. S12). Gene Ontology (GO),
Kyoto Encyclopedia of Genes and Genomes (KEGG), and Gene Set
Enrichment Analyses (GSEA) were then performed to investigate
the biological functions of these DEGs. As expected, the ER unfold-
ed protein response (ER UPR), immune system process, response to
ER stress, and response to unfolded protein pathways were among
the top up-­regulated GO terms in NanoICD-­treated cells (Fig. 3, I
and J), and protein processing in ER and antigen processing and
presentation pathways were among the top up-­regulated KEGG
terms (fig. S13 and Fig. 3J). These up-­regulated gene terms sug-
gest that NanoICD/BSA acts as a biomimetic unfolded/misfolded
protein that induces ER stress and ER UPR, promotes antigen

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 5 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

processing and presentation, and ultimately activates immune sys- NanoICD/BSA (24.1%) resulted in a significant increase in the pre-­
tem processes. apoptotic surface CRT exposure on the cell surface compared to
PBS (2.11%), ETL (2.52%), nBSA (1.44%), and PTX (11.6%) treat-
ICD-­associated immune responses induced by NanoICD/BSA ments, suggesting efficient activation of ICD. Similar trends were
During the development of tumors, tumor cells transmit a “don’t eat observed from the evaluation of serum HMGB-­1 levels (fig. S17),
me” signal through the up-­regulation of CD47 and its binding to where PTX-­and NanoICD/BSA-­treated mice exhibited significantly
signal regulatory protein α (SIRPα) on the surface of antigen-­ higher HMGB1 levels in their serum compared to those treated with
presenting cells (APCs) via the CD47-­SIRPα pathway. This interac- PBS, ETL, and nBSA, which is consistent with the results obtained
tion effectively hinders the immune clearance of cancer cells. For a from in vitro HMGB-­1 analysis (Fig. 3D). Further flow cytometric
typical ICD, the surface-­exposed CRT acts as an “eat me” signal, analysis of mature DCs and effector memory cells confirmed this
mediated by the CRT–low density lipoprotein receptor (LPR1, CRT-­ result, with NanoICD/BSA-­treated mice showing significantly higher
LPR1) pathway, to promote the recognition and phagocytosis of levels of mature DCs (Fig. 4D and fig. S18; CD80hiCD86hi cells,
TAAs by APCs (43). For demonstration, B16F10 (prestained with 32.7% for NanoICD/BSA and 28.9% for PTX) and effector memory
DiD) were first incubated with PBS, PTX, ETL, nBSA, or NanoICD/ cells (Fig. 4E and fig. S19; CD44hiCD62L−, 19.2% for NanoICD/BSA
BSA and then cocultured with bone marrow–derived dendritic cells and 19.5% for PTX) than those receiving other treatments, indicat-
(BMDCs, prestained with DiO) for flow cytometric analysis. As ing the activation of ICD-­associated antitumor immune responses.
shown in fig. S14, NanoICD/BSA treatment (45.1%) significantly Next, we analyzed the tumor-­infiltrating lymphocyte (TIL) pop-
enhanced the phagocytosis of B16F10 by BMDCs compared to ulation in tumors to further investigate the immune activation
those treated with PBS (5.45%), ETL (8.33%), nBSA (9.43%), and achieved by NanoICD/BSA. As shown in Fig. 4F and fig. S20, the
even PTX (13.9%), indicating that NanoICD/BSA effectively pro- highest level of CD8+ TIL (CD45+CD3+ cells) was observed in tu-
moted BMDC-­ mediated phagocytosis of tumor cells. After the mors from mice treated with NanoICD/BSA (26.6%) compared to
phagocytosis, the cross-­presentation of TAAs was further studied. PBS (8.21%), ETL (12.9%), nBSA (15.4%), and PTX (21.6%). Simi-
This was achieved by using ovalbumin (OVA)–expressing B16F10 lar results were observed when analyzing the proliferation and ac-
cells (B16F10-­OVA) to perform a similar study as above. As shown tivity of TILs, in which NanoICD/BSA-­treated mice exhibited a
in fig. S15, NanoICD/BSA-­treated BMDCs exhibited the highest significantly higher level of Ki67hiCD3+ T cells (Fig. 4G and fig. S21;
levels of SIINFEKL-­H2kb complex (22.9%) compared to other treat- CD45+ cells, 20.3% for NanoICD/BSA and 10.5% for PTX) and
ments, suggesting that NanoICD/BSA effectively promoted the GzmBhiCD8+ T cells (Fig. 4H and fig. S22; CD45+CD3+ cells, 34.8%
cross-­presentation of OVA peptide on the major histocompatibility for NanoICD/BSA and 24.1% for PTX) than other treatments, indi-
protein I (MHC I) complex. In addition, the exodus of HMGB-­1 and cating the ability of NanoICD to maintain cytotoxic T lymphocyte
the secretion of ATP during ICD also serve as adjuvants to promote (CTL) proliferation/activity and elicit T cell–mediated antitumor
APC maturation (26). To this end, the proportion of mature BM- immunity. Next, a standard vaccination assay was performed to de-
DCs (CD11c+CD80+CD86+) was investigated by coculturing BM- termine whether NanoICD/BSA is sufficient to drive bona fide ICD
DCs with the B16F10 cells treated with NanoICD/BSA, nBSA, ETL, and elicit protective cognate anticancer immunity against tumor re-
PTX, and PBS. As shown in fig. S16, significantly enhanced BMDC currence and metastasis (28, 30, 44). B16F10 cells were first exposed
maturation was observed from cells treated with NanoICD/BSA to ETL, nBSA, and NanoICD/BSA and then washed and resuspend-
(43.5%) compared to those treated with PBS (16.2%), PTX (23.3%), ed in PBS to remove treatment reagents. Cells succumbing to PTX
ETL (19.1%), and nBSA (20.9%), indicating the high efficiency of were used as the positive control. The pretreated cells and PBS (no
NanoICD/BSA in promoting DC maturation. Quantitative analysis vaccination control) were then injected subcutaneously into the left
further confirmed these results, where the highest levels of phagocy- flank of C57BL/6 mice. One week later, the mice were rechallenged
tosis, antigen presentation, and maturation of BMDCs were ob- by injecting live B16F10 into the contralateral flank (Fig. 4I) or vein
served from cells treated with NanoICD/BSA. All these results (Fig. 4J). The tumor-­free survival of the mice was continuously
suggest the great potential of NanoICD/BSA as an effective monitored for 60 days after the challenge. As summarized in Fig. 4I,
ICD inducer. inoculation of NanoICD/BSA-­and PTX-­treated cells conferred im-
To evaluate the potential of NanoICD/BSA to induce ICD-­ munological protection against tumor recurrence in 40 and 20% of
associated immune responses and promote antitumor immunity the mice. In addition, the inoculation of NanoICD/BSA-­and PTX-­
in vivo, the antitumor efficacy of NanoICD/BSA was studied in treated cells also prevented lung metastasis of B16F10 cells (Fig. 4J).
B16F10-­bearing C57BL/6 mice. Seven days after inoculation with Collectively, these results indicate that NanoICD/BSA can effective-
B16F10 cells, tumor-­bearing mice were randomly divided into five ly drive bona fide ICD and induce protective cognate anticancer
groups (n = 6) and then intratumorally injected with PBS, PTX, immunity.
ETL, nBSA, or NanoICD/BSA. The tumor volumes were continu-
ously monitored for 22 days (as illustrated in Fig. 4A; details in the Enhancement of antitumor immune responses
Supplementary Materials). As shown in Fig. 4B, treatment with PTX by NanoICD/CAT
and NanoICD/BSA significantly inhibited the growth of B16F10 tu- Despite the effective generation of ER stress and the activation of
mors. This significantly enhanced the idea that the antitumor effect ICD by NanoICD/BSA, the therapeutic efficacy may still be far from
should be attributed to the activation of ICD-­associated antitumor expected because of the immunosuppressive environment of the tu-
immunity. For verification, typical ICD-­related indicators, including mor (1, 45). Inflammation and hypoxia are the two most studied
the pre-­apoptotic surface CRT exposure on tumor cells, mature DCs indicators of immunosuppression, which impair antitumor immune
in tumor-­draining lymph nodes, and effector memory cells in responses by reducing the migratory, cytolytic, and survival of effec-
the spleen, were analyzed. As shown in Fig. 4C, treatment with tor T cells and promoting the infiltration of immunosuppressive

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 6 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. NanoICD activates ICD-­associated antitumor immunity in vivo. (A) Schematic illustration of the experimental design. (B) Average tumor growth kinetics in
different groups (PBS, PTX, ETL, nBSA, and NanoICD/BSA). (C) Pre-­apoptotic CRT exposure on tumors from mice treated with PBS, ETL, PTX, nBSA, and NanoICD/BSA.
(D and E) The population of CD80+CD86+ DC cells (gated on CD11c+) in tumor-­draining lymph nodes (D) and CD44+CD62L− effector memory cells (gated on CD3+CD8+)
in spleen (E) from mice treated with PBS, PTX, ETL, nBSA, and NanoICD/BSA. (F to H) The population of TILs (F), Ki67+CD3+ (gating on CD45+) (G), and GzmB+CD8+ (gating
on CD45+CD3+) (H) in tumors from mice treated with PBS, PTX, ETL, nBSA, and NanoICD/BSA. (I) Schematic illustration of the establishment of the vaccination model
to evaluate the ability of NanoICD/BSA to against tumor recurrence (left), and the tumor free survival (TFS) of C57BL/6 mice after inoculation of pre-treated B16F10 cells
(PBS, PTX, ETL, nBSA, or NanoICD/BSA for 24 hours), followed by the inoculation of live B16F10 cells at the contralateral flank, n = 10 (right). (J) Schematic illustration of the
establishment of the vaccination model to evaluate the ability of NanoICD/BSA to against tumor lung metastases (left), and the lung metastases in mice after the indi-
cated treatments (right). Data are presented as means ± SD from six independent experiments for (B) (biological replicates, n = 6) and three independent experiments for
(D) to (H) (biological replicates, n = 3). Significant levels are *P < 0.05, **P < 0.01, and ****P < 0.0001.

cells and the accumulation of proinflammatory mediators (46).
Therefore, simultaneous immune activation and inflammation/hy-
poxia alleviation may be a promising strategy to enhance the antitu-
mor efficacy. Different from traditional ICD inducers, the highly
customizable structure of NanoICD provides the potential to inte-
grate immunomodulating functions by replacing the BSA with oth-
er functional proteins and enzymes during the synthesis. In addition,
coating NanoICD with a TME-­responsive polymer allows effective
regulation of its biodistribution after systemic administration. CAT,
which catalyzes the decomposition of H2O2 into O2, emerges as an
ideal candidate for alleviating immunosuppressive inflammation
and the hypoxic TME. For demonstration, CAT was used to perform
the nanoencapsulation to synthesize NanoICD/CAT. Subsequently,
NanoICD/CAT was coated with a poly(ethylene glycol) (PEG)–
based pH-­responsive polymer (PCA) to obtain NanoICD/CAT-­
PCA (Fig. 5A) (14, 16). The PEG segments of the coating effectively
protect NanoICD/CAT-­PCA from clearance by the mononuclear
phagocytosis system (MPS) and prevents undesired ICD activation.
Upon reaching acidic TME, the polymer coating was dissociated,
allowing the release and cellular uptake of NanoICD/CAT. Last, the
surface interaction between NanoICD/CAT and ER leads to the ac-
tivation of ICD-­associated immune responses, while the integrated
CAT efficiently alleviates tumor immunosuppression by decompos-
ing H2O2 to O2, resulting in the activation and enhancement of anti-­
tumor immune responses.
To construct NanoICD/CAT-­PCA, NanoICD/CAT was prepared
using a similar nanoencapsulation strategy by replacing BSA with
CAT (details in the Supplementary Materials). After the prepa-
ration, we first evaluated the mass of NanoICD/CAT bound
to ER using QCM-­D. Nanoparticles with structures similar to
NanoICD/CAT but without ETL (referred to as nCAT) were synthe-
sized as a comparison group. As shown in Fig. 5B, NanoICD/
CAT (ΔF =198.95 Hz) exhibited a significantly higher maximum
frequency shift than nCAT (ΔF = 124.05 Hz). When rinsed with
PBS, only a slight change in frequency shift was observed in NanoICD/
CAT (ΔF = 146.12 Hz, ΔΔF = 52.83 Hz) compared to nCAT

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 7 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. Schematic and the characterization of NanoICD/CAT-­PCA. (A) Schematic illustration of NanoICD/CAT-­PCA for synergistic immune activation. (B) QCM-­D ex-
periment showing the binding affinities between nCAT, NanoICD/CAT, and ER. (C) H2O2 degradation after treatment with CAT, nCAT, and NanoICD/CAT. (D) O2 generation
after treatment with CAT, nCAT, and NanoICD/CAT. (E) Size distribution and TEM image of NanoICD/CAT-­PCA. Scale bar, 100 nm. (F) Fluorescence spectra of NanoICD/CAT
and NanoICD/CAT-­PCA after incubation under different conditions. (G) Cellular uptake efficacy of NanoICD/CAT and NanoICD/CAT-­PCA under different conditions.
(H) Pre-­apoptotic CRT exposure on cell surface after treatment with PBS, NanoICD/CAT, NanoICD/CAT-­PCA (pH 7.4), and NanoICD/CAT-­PCA (pH 6.5). (I) The exodus of HMGB-­1
from cell nuclei after different treatments. (J) Ex vivo images of tumor and major organs of mice treated with NanoICD/CAT and NanoICD/CAT-­PCA at 24 hours after injec-
tion. The CAT of NanoICD/CAT and NanoICD/CAT-­PCA was prelabeled with Cy5. Scale bars, 25 μm. a.u., arbitrary units.

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 8 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(ΔF = 1.79 Hz, ΔΔF = 122.26 Hz), suggesting the strong binding of
NanoICD/CAT to ER. Next, the catalytic activity of NanoICD/CAT
was evaluated. As shown in fig. S23, 88.3% of the original CAT ac-
tivities were retained after the nanoencapsulation. The catalytic ac-
tivity enables NanoICD/CAT to decompose H2O2 (Fig. 5C) and
generate O2 effectively (Fig. 5D), thus providing the potential to al-
leviate tumor hypoxia and remodel immunosuppressive TME to
enhance cancer immunotherapy.
To avoid MPS clearance and undesired immune activation,
NanoICD/CAT was coated with a PEG-­based pH-­responsive polymer,
PCA (mPEG113-­PLys120/CA, PCA), to obtain NanoICD/CAT-­PCA
(detailed preparation in the Supplementary Materials). Transmis-
sion electron microscopy (TEM) and dynamic light scattering indi-
cated the successful formation of spherical nanoparticles with a
diameter of 44.25 nm and a zeta potential of −9.44 mV (Fig. 5E and
fig. S24). Förster resonance energy transfer (FRET)–based analysis
shows that a significant FRET signal was observed from NanoICD/
CAT-­PCA (Fig. 5F), indicating the successful coating of the polymer
on NanoICD/CAT. The polymer coating was readily detached from
NanoICD/CAT-­PCA upon reaching acidic TME, which was con-
firmed by the decrease in FRET signal when NanoICD/CAT-­PCA
was incubated at pH 6.5. In addition, the cellular uptake efficiency of
NanoICD/CAT-­PCA was investigated under different conditions.
As shown in the confocal images (Fig. 5G), negligible fluorescence
signal was detected from cells treated with NanoICD/CAT-­PCA
(pH 7.4). In contrast, significantly higher fluorescence signals were
observed in cells treated with NanoICD/CAT and NanoICD/
CAT-­PCA (pH 6.5). This result was then confirmed by the flow cy-
tometric analysis (fig. S25). The significantly different cellular up-
take behaviors of NanoICD/CAT-­PCA under different conditions
(pH 7.4 and 6.5) may provide an opportunity to precisely control the
activation of ICD-­associated antitumor immunity. For demonstration,
the ICD-­inducing ability of NanoICD/CAT-­PCA was then investi-
gated under different conditions. Consistent with its cellular uptake
behavior, NanoICD/CAT-­PCA effectively induced CRT exposure
(Fig. 5H and fig. S26), HMGB-­1 release (Fig. 5I), and ATP secretion
(fig. S27) only under acidic conditions (pH 6.5), but not under phys-
iological conditions (pH 7.4). In addition, the polymer coating sig-
nificantly improved the biodistribution of NanoICD/CAT, with
significantly enhanced fluorescence signals observed in tumor tis-
sues of mice treated with NanoICD/CAT-­PCA compared to Na-
noICD/CAT (Fig. 5J). All these results indicate that NanoICD/
CAT-­PCA can be enriched and activated at the tumor site after sys-
temic administration, which is important for enhancing the efficien-
cy of immunotherapy and reducing side effects.
NanoICD/CAT-­PCA for enhanced cancer immunotherapy
Next, we evaluated the antitumor efficacy of NanoICD/CAT-­PCA
by simultaneously inducing ICD and remodeling the TME in vivo.
To demonstrate the versatility and effectiveness of the NanoICD
strategy in cancer immunotherapy, 4T1 breast tumor, which is
another tumor model with lower immunogenicity, was used for
in vivo studies. For better demonstration, PCA-­coated nCAT
(nCAT-­PCA), which can only remodel immunosuppressive TME,
and PCA-­coated NanoICD/BSA (NanoICD/BSA-­PCA), which can
only induce ICD, were used as the comparison groups to perform the
same test. As shown in Fig. 6A and fig. S28, NanoICD/CAT-­PCA
exhibited the most effective tumor suppression in mice compared
to other treatments. As a result, NanoICD/CAT-­PCA significantly
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
improved the survival of 4T1-­bearing BALB/c mice, with 50% of
tumor-­bearing mice alive at the observed 45 days (Fig. 6B). In addi-
tion, negligible body weight loss was observed during the treatment
(Fig. 6C). Histopathological analysis of major organs, including heart,
liver, spleen, lung, and kidney, showed no obvious abnormalities
or damage after the treatment, indicating the safety of nCAT-­PCA
(fig. S29).
To investigate the mechanism underlying the enhanced anti-­tumor
effect, the tumor tissues were harvested for flow cytometric analysis. As
shown in Fig. 6 (D and E), a significant increase in the infiltration of cy-
totoxic CD8+ T cells (CTLs, CD45+CD3+) was observed in the tumor
tissues of NanoICD/CAT-­PCA (45.7%) treatment compared to those of
PBS (6.28%), nCAT-­PCA (15.6%), and NanoICD/BSA-­PCA (15.9%).
CLSM observation of the tumor tissues confirmed these results (Fig. 6F),
with the highest fluorescent signal of CD8+ CTLs observed in tumor
sections from mice treated with NanoICD/CAT-­PCA. Furthermore, sig-
nificantly improved proliferation (Ki67hiCD8+ cells; Fig. 6G and fig. S30)
and activation (granzyme B+ CD8+; Fig. 6H and fig. S31) of CTLs were
also observed in the tumor of NanoICD/CAT-­PCA–treated mice, indi-
cating the effective activation of T cell–based antitumor immunity. In
addition, the intratumoral proportions of immunosuppressive cells were
also analyzed by flow cytometry. As shown in the results, nCAT-­PCA
and NanoICD/CAT-­PCA treatments resulted in a significant reduction
of M2-­like tumor-­associated macrophages (M2-­TAMs) (Fig. 6I and
fig. S32), regulatory T cells (Tregs) (Fig. 6J and fig. S33), and myeloid-­
derived suppressor cells (MDSCs) (Fig. 6K and fig. S34), indicating the
effective alleviation of immunosuppression by the catalytic activity of
CAT. Further analysis of cytokine (interleukin-­10, IL-­10; fig. S35A) and
immunosuppressive cytokines (transforming growth factor–β, TGF-­β;
fig. S35B) confirmed these results, with nCAT-­PCA and NanoICD/
CAT-­PCA treatment showing a significant down-­regulation of IL-­10
and TGF-­β.
The significantly improved antitumor immunity exhibited by
NanoICD/CAT-­PCA can be attributed to the synergy between ICD
activation and the remodeling of the immunosuppressive TME. To this
end, the surface CRT exposure in tumor cells was first analyzed using flow
cytometry (fig. S36). As shown in Fig. 7A, significantly higher PI−CRT+
cell population was observed from the tumor sections of mice treated
with NanoICD/BSA-­PCA and NanoICD/CAT-­PCA, which is consist­
ent with the results obtained from immunofluorescence imaging. Fur-
thermore, elevated levels of mature DCs (Fig. 7B and fig. S37) and
effector memory T cells (Fig. 7C and fig. S38) were identified in tumors
of mice treated with NanoICD/BSA-­PCA and NanoICD/CAT-­PCA, in-
dicating the activation of ICD-­associated antitumor immune responses.
Next, we assess the ability of NanoICD/CAT-­PCA to mitigate the in-
flammation and the hypoxia of TME. The H2O2 concentrations in tu-
mors were measured through the utilization of an OxiVisionGreen
probe. We found that both NanoICD/CAT-­PCA and nCAT-­PCA treat-
ments significantly reduced fluorescence signals (Fig. 7D) of OxiVision-
Green, indicating a decrease in H2O2 concentration within tumors. This
suggests that inflammation within the TME was effectively alleviated. In
addition, the decomposition of H2O2 led to the generation of O2 and
consequent alleviation of hypoxic conditions of TME, which was con-
firmed by the down-­regulation of hypoxia-­inducible factor–1α (HIF-­1α,
Fig. 7E). As a result of these properties, NanoICD/CAT-­PCA efficiently
elicited protective cognate antitumor immunity against tumor rechal-
lenge (Fig. 7F) and tumor lung metastasis (Fig. 7G). These findings high-
light the great potential of NanoICD as a platform technology for
enhanced cancer immunotherapy.
9 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. In vivo antitumor efficacy of NanoICD/CAT-­PCA. (A) Average tumor growth kinetics in different groups (PBS, nCAT-­PCA, NanoICD/BSA-­PCA, and NanoICD-­PCA),
n = 6. (B) The survival of tumor-­bearing mice after different treatments as indicated. (C) The variations of mouse weight during treatment, n = 6. (D and E) The population
of CD4+ T cells and CD8+ T cells (gated on CD45+CD3+) within tumors of mice after receiving different treatments (D) and quantitative results (E). (F) Immunofluorescence
images showing the infiltration of CD8+ T cells and CD4+ T cells within tumors after different treatments. (G to K) Quantitative analysis of Ki67hiCD8+ T cells (G), granzyme
B+ CD8+ T cells (H), M2-­like TAMs (I), Treg cells (J), and MDSCs (K) in tumors from mice treated with PBS, nCAT-­PCA, NanoICD/BSA-­PCA, and NanoICD/CAT-­PCA. Data are
presented as means ± SD of six independent experiments for (A) to (C) (biological replicates, n = 6) and three independent experiments for (E) and (G) to (K) (biological
replicates, n = 3). Significant levels are ***P < 0.001 and ****P < 0.0001. Scale bars, 100 μm.

DISCUSSION effects, there is a lack of rationally designed nanoparticles with

Nanoparticle-­based cancer immunotherapy has shown promising
therapeutic potential in clinical settings. However, current re-
search mainly focuses on the use of nanoparticles as delivery ve-
hicles but often overlooks their potential to directly modulate
immune responses. Although some serendipitous discoveries sug-
gest that certain nanoparticles may have immunoregulatory
immunomodulatory capacity. To this end, we have presented a
novel synthetic nanoparticle, NanoICD, capable of inducing
ICD. NanoICD has a rationally designed surface containing two
types of functional groups that determine its surface charge and
ER-­targeting potential. By controlling monomer composition and
formulation methods, NanoICD achieves efficient accumulation

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 10 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. NanoICD/CAT-­PCA activates ICD and remodels immunosuppressive TME in vivo. (A) Pre-­apoptotic CRT exposure on tumors from mice treated with PBS, nCAT-­
PCA, NanoICD/BSA-­PCA, and NanoICD/CAT-­PCA. (B) The population of CD80+CD86+ DC cells (gated on CD11c+) in tumor-­draining lymph nodes of mice after receiving
different treatments. (C) The population of CD44+CD62L− effector memory cells (gated on CD3+CD8+) in spleen of mice after receiving different treatments. (D) Flow cy-
tometric analysis showing H2O2 concentrations in tumor tissues after different treatments. (E) Immunofluorescence images showing HIF-­1α concentrations in tumor tis-
sues after different treatments. (F) The TFS of 4T1-­bearing BALB/c mice after different treatments as indicated. (G) The lung metastases in mice after the indicated
treatments. Scale bar, 200 μm. Data are presented as means ± SD from three independent experiments for (B) to (D) (biological replicates, n = 3) and six independent
experiments for (F) (biological replicates, n = 6). Significant levels are ***P < 0.001 and ****P < 0.0001. Scale bar, 200 μm.

in the ER, induces ER stress, and ultimately activates ICD-­
associated immune responses (Figs. 2 and 3). Animal studies based
on NanoICD/BSA confirmed the immunomodulatory function of
the polymer surface (Fig. 4). Furthermore, the synthesis method of
NanoICD is generally applicable to various proteins and enzymes,
so that the biological functions of the encapsulated protein or en-
zyme can synergize with the ICD-­inducing function of the poly-
mer surface to achieve enhanced antitumor performance (Fig. 5).
As demonstrated by encapsulating CAT, NanoICD/CAT-­PCA ef-
fectively alleviated the immunosuppressive TME and induced ro-
bust antitumor immune responses in 4T1 tumor–bearing mice
(Fig. 6), significantly suppressing tumor recurrence and lung me-
tastasis, and improving the survival rate of the mice (Fig. 7). These
unique features make NanoICD a novel strategy for effective can-
cer immunotherapy.
Moreover, the research presented in this study illustrates how
a specifically engineered nanoparticle is capable of modulating
the immune system independent of conventional immunomodu-
latory drugs. The study on the structure-­activity relationship con-
firmed the critical role of surface properties in the ICD-­inducing
ability of NanoICD. As the design of NanoICD was inspired by
the endogenous ER stress generated by unfolding/misfolding pro-
teins, this work demonstrates the potential of artificial nanostruc-
ture with a tailor-­made surface to regulate biological processes
autonomously. Consequently, this study provides a novel perspec-
tive for the development of advanced nanomedicines for cancer
immunotherapy.
Despite the significant efficacy of NanoICD in activating ICD-­
associated antitumor immune response, there remain certain is-
sues that necessitate our attention and resolution in future studies.
(i) The impact of macrophage phagocytosis on the overall ICD
process. Macrophage phagocytosis of NanoICD diminishes its
bioavailability in tumor cells and may lead to unpredictable altera-
tions in these macrophages. Thus, improving the targeting and

Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024 11 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
selectivity of NanoICD for tumor cells is imperative in our ongo-
ing research. (ii) The role of don’t-­eat-­me signals. While NanoICD
effectively induce the exposure of CRT on tumor surface and
prompt the endocytosis of tumors by phagocytic cells, its efficacy
may still be limited by don’t-­eat-­me signals, such as CD47. There-
fore, the combination of NanoICD with inhibitors targeting the
don’t-­eat-­
me signals, such as anti-­ CD47, is expected to yield
broader clinical benefits. (iii) The variability of human cell lines
compared to B16F10 and 4T1. Although we have demonstrated in
human cancer cell lines that NanoICD induces the release of
DAMPs, including CRT, HMGB-­1, and ATP (fig. S39), it is im-
perative to conduct further investigations, particularly in patient-­
derived xenografts and humanized mouse models, to corroborate
and validate this observation.
MATERIALS AND METHODS
Materials
All commercially available reagents and solvents were used as re-
ceived without further purification unless otherwise noted. BSA and
CAT were purchased from Mreda. AAm and N-­acryloxysuccinimide
(NAS) were purchased from Aladdin. APm and N-­tosylethylenediamine
were obtained from Energy Chemical. Acryloyl chloride, ammoni-
um persulfate (APS), BIS, and N,N,N′,N′-­tetramethylethylenediamine
(TEMED) were purchased from Bidepharm. Antibodies for CLSM
observation and flow cytometry were purchased from BioLegend
unless otherwise noted: phycoerythrin (PE)–conjugated anti-­mouse
CD11b (dilution 1:200, catalog no. 101207), fluorescein isothiocya-
nate (FITC)–conjugated anti-­mouse CD80 (dilution 1:200, catalog
no. 104705), allophycocyanin (APC)–conjugated anti-­mouse CD86
(dilution 1:200, catalog no. 105011), APC-­conjugated anti-­mouse
CD11c (dilution 1:200, catalog no. 117309), FITC-­conjugated anti-­
mouse SIINFEKL-­ MHC I (dilution 1:200, catalog no. 116505),
APC-­conjugated anti-­mouse CD3 (dilution 1:200, catalog no.
100235), PE-­conjugated anti-­mouse CD8a (dilution 1:200, catalog
no. 110707), FITC-­conjugated anti-­mouse CD44 (dilution 1:200,
catalog no. 156007), APC/Cy7-­conjugated anti-­mouse CD62L
(dilution 1:200, catalog no. 104427), APC/Cy7-­conjugated anti-­
mouse CD45 (dilution 1:200, catalog no. 103115), FITC-­conjugated
anti-­mouse Ki-­6 7 (dilution 1:200, catalog no. 652409), FITC-­
conjugated anti-­mouse granzyme B (dilution 1:200, catalog no.
515403), APC-­conjugated anti-­mouse F4/80 (dilution 1:200, catalog
no. 123115), PerCP/Cy5.5-­conjugated anti-­mouse CD206 (dilution
1:200, catalog no. 141715), APC-­conjugated anti-­mouse CD25
(dilution 1:200, catalog no. 101910), PE-­conjugated anti-­mouse
FOXP3 (dilution 1:200, catalog no. 118903), FITC-­conjugated anti-­
mouse Gr-­1 (dilution 1:200, catalog no. 108405), and FITC-­
conjugated anti-­mouse CD4 (dilution 1:200, catalog no. 100405).
ATTO 488–conjugated anti-­human CRT (dilution 1:200, catalog
no: SPC-­122) was obtained from Stressmarq. ELISA kits for HMGB-
­1 (E-­EL-­M0676c) and IL-­10 (E-­EL-­M0046c) were purchased from
Elabscience Biotechnology Co. Ltd., and TGF-­β was purchased from
Tianjin Anorikang Biotechnology Co. Ltd. Antibodies for Western
blot were purchased from Abcam. AV-­FITC apoptosis detection kit,
ATP assay kit, paraformaldehyde, ER-­Tracker Red, and Triton X-­
100 were purchased from Beijing Solarbio Science & Technology
Co. Ltd. Lysosome-­Tracker Red was obtained from Beyotime.
OxiVision Green hydrogen peroxide sensor was purchased from
AAT Bioquest.
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
Instruments
Nuclear magnetic resonance data were recorded on a Bruker
AV400 spectrometer. Ultraviolet-­visible spectra were recorded
with a Nanodrop Onec (Thermo Fisher Scientific, USA). Size
distribution and zeta potential were performed on a Brookhaven
ZETAPALS/BI-­2 00SM (Brookhaven Instrument, USA). TEM
measurements were performed on a Talos F200C electron mi-
croscope at an acceleration voltage of 120 kV. The mass spec-
trum of ETL was performed on a Fouier transform ion cyclotron
resonance mass spectrometer (Varian 7.0T FTMS). Flow cytom-
etry analysis was performed on a BD LSR Fortessa flow cytom-
etry. CLSM images were captured on a FluoView Confocal Laser
Scanning Microscopes-­FV1000. Living imaging was performed
with an IVIS Lumina imaging system (Caliper Life Sciences,
USA). Tumor tissues, spleen, and tumor-­draining lymph nodes
were homogenized with the GentleMACs Dissociator (FJ200-­
SH). The binding affinities between nanoparticles and ER were
analyzed with QCM-­D, equipped with a Q-­S ence E4 system (Q-­
Sence, Västra Frölunda, Sweden).
Syntheses of ETL
To a solution of N-­tosylethylenediamine (1.35 g, 6.3 mmol) in
CH2Cl2 (20 ml), acryloyl chloride (0.855 g, 9.45 mmol) and tri-
ethylamine (0.96 g, 9.45 mmol) in CH2Cl2 (10 ml) were added
dropwise to the solution at 0°C and stirred at room temperature
for another 2 hours. The reaction mixtures were then extracted
twice with water and saturated sodium chloride to remove unre-
acted acryloyl chloride and triethylamine. The prepared ETL
was then purified by column chromatography (PE:EA = 1:1)
(yield, 1.61 g, 95%).
Preparation of NanoICD
NanoICD was synthesized by a polymerization method that forms a
thin layer of cross-­linked polymer around each protein molecule.
Before polymerization, proteins (BSA or CAT) were reacted with
NAS to attach polymerization sites onto the surface, which was
achieved by adding NAS [10% in dimethyl sulfoxide (DMSO) w/v]
to the protein solution (5 mg/ml, pH 8.5, 50 mM sodium carbonate
buffer) at a molar ratio of 20:1 (NAS/protein) and reacting for 1 hour
at 4°C. The solution was then thoroughly dialyzed against pH 8.5
sodium carbonate buffer (50 mM) using a dialysis bag [molecular
weight cutoff (MWCO) = 10 kDa]. Monomers and cross-­linkers,
which were first prepared as stock solutions [AAm and APm: 20%
(w/v) in deionized water, ETL, and BIS: 10% (w/v) in DMSO], were
then added to the protein solution. The final protein concentration
was adjusted to 1 mg/ml by dilution with sodium carbonate buffer
(50 mM, pH 8.5). The polymerization was initiated by the addition
of TEMED/APS (2:1, w/w) and kept at 4°C for 1 hour. After the
polymerization, the unreacted monomers and cross-­linkers were
removed by centrifugal filtration (MWCO = 10 kDa). To prepare
NanoICD/BSA with different zeta potentials, the SA was used to
neutralize the polymerized APm in NanoICD/BSA to varying de-
grees. In detail, SA was first prepared as a stock solution (10 mg/kg
in DMSO), then added to the NanoICD/BSA solution (5 ml, 1 mg/ml,
pH 8.5, 50 mM sodium carbonate buffer) at different molar ratios
(SA/BSA, 25:1, 50:1, and 100:1), and reacted for 1 hour at
4°C. The unreacted SA was then removed by centrifugal filtration
(MWCO = 10 kDa). The final protein concentration was diluted to
1 mg/ml with sodium carbonate buffer and stored at 4°C for further
12 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
use. The molar ratios for the preparation of different NanoICD are
provided in tables S1 and S2.
Preparation of NanoICD/CAT-­PCA
To prepare NanoICD/CAT-­PCA, a PEG-­based pH-­responsive poly-
mer, PCA, was first synthesized according to our previously reported
methods (47, 48). Next, PCA was dissolved in sodium carbonate
buffer (pH 8.5, 50 mM) to make a 10 mg/ml stock solution and
mixed with NanoICD/CAT at a 10:1 molar ratio (PCA/BSA) to
afford NanoICD/CAT-­PCA.
Investigation of the mass of NanoICD bound to ER
The mass of NanoICD bound to ER was measured using a QCM-
­D. Au sensor chips were first modified with nBSA, NanoICD/BSA-­
20, NanoICD/BSA-­40, nCAT, and NanoICD/CAT (1.5 μM) and
then placed in the standard flow module (detailed in the Supple-
mentary Materials). Each sensor chip was washed with PBS buffer
for 1 hour at a flow rate of 10 μl/min and then equilibrated at a flow
rate of 2 μl/min until the baseline was stable. Then, freshly extracted
ER was added to the flow buffer and allowed to run at the flow rate
of 2 μl/min for 30 min. After reaching equilibrium, the flow phase
was replaced with PBS (flow rate of 2 μl/min). All the QCM experi-
ments were performed at 37°C.
Vaccination assays
All animal studies were performed in accordance with the Regu-
lations for the Administration of Affairs Concerning Experimen-
tal Animals (Tianjin, revised in June 2004) and were approved by
the Animal Ethics Committee of Nankai University (approval
number, 2023-­SYDWLL-­000235). The ability of NanoICD/BSA to
prevent tumor recurrence was evaluated using a vaccine assay.
Briefly, 1 × 106 B16F10 cells were first incubated with ETL (100 μM),
PTX (15 μM), nBSA (1.5 μM), and NanoICD/BSA (1.5 μM) for
24 hours and then washed and resuspended in PBS. Subsequently,
the treated cells and PBS (no vaccination control) were inoculated
subcutaneously into the lower flank of 6-­week-­old female C57BL/6
mice (vaccination). One week later, the mice were subcutaneously
injected with 1 × 105 untreated B16F10 cells into the contralateral
flank (rechallenge). Tumor incidence was routinely monitored up
to 60 days after the rechallenge.
The ability of NanoICD/CAT-­PCA to prevent tumor recurrence
was evaluated with a similar vaccine assay using a different type of
cancer cells. Briefly, 1 × 106 4T1 cells were inoculated subcutane-
ously into the left mammary fat pad of 6-­week-­old female BALB/c
mice. One week later, 200 μl of PBS, nCAT-­PCA (4 μM, 10 mg/kg),
NanoICD/BSA-­PCA (4 μM, 2.66 mg/kg), and NanoICD/CAT-­PCA
(4 μM, 10 mg/kg) were injected intravenously every 2 days for a
total of three doses. Subsequently, 1 × 105 4T1 cells were inoculated
subcutaneously into the contralateral flank (rechallenge). Tumor
incidence was routinely monitored up to 45 days after the
rechallenge.
Statistical analysis
Data are presented as means ± SD from at least three independent
experiments (n ≥ 3), and the significance levels are *P < 0.05,
**P < 0.01, ***P < 0.001, and ****P < 0.0001, analyzed by one-­
way or two-­way analysis of variance (ANOVA) (when more than
two groups were compared) with a Dunnett’s test. P < 0.05 or less
was considered significant.
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
Supplementary Materials
This PDF file includes:
Supplementary Text
Figs. S1 to S39
Tables S1 to S4
REFERENCES AND NOTES
1. R. S. Riley, C. H. June, R. Langer, M. J. Mitchell, Delivery technologies for cancer
immunotherapy. Nat. Rev. Drug Discov. 18, 175–196 (2019).
2. Y. Zhao, Z. Zhang, Z. Pan, Y. Liu, Advanced bioactive nanomaterials for biomedical
applications. Exp. Dermatol. 1, 20210089 (2021).
3. S. Musetti, L. Huang, Nanoparticle-­mediated remodeling of the tumor microenvironment
to enhance immunotherapy. ACS Nano 12, 11740–11755 (2018).
4. J. Xue, Z. Zhao, L. Zhang, L. Xue, S. Shen, Y. Wen, Z. Wei, L. Wang, L. Kong, H. Sun, Q. Ping,
R. Mo, C. Zhang, Neutrophil-­mediated anticancer drug delivery for suppression of
postoperative malignant glioma recurrence. Nat. Nanotechnol. 12, 692–700 (2017).
5. Z. Zhang, Y.-­X. Yue, L. Xu, Y. Wang, W.-­C. Geng, J.-­J. Li, X.-­L. Kong, X. Zhao, Y. Zheng, Y. Zhao,
L. Shi, D.-­S. Guo, Y. Liu, Macrocyclic-­amphiphile-­based self-­assembled nanoparticles for
ratiometric delivery of therapeutic combinations to tumors. Adv. Mater. 33, e2007719
(2021).
6. T.-­X. Zhang, Z. Zhang, Y.-­X. Yue, X.-­Y. Hu, F. Huang, L. Shi, Y. Liu, D.-­S. Guo, A general
hypoxia-­responsive molecular container for tumor-­targeted therapy. Adv. Mater. 32,
e1908435 (2020).
7. M. Luo, H. Wang, Z. Wang, H. Cai, Z. Lu, Y. Li, M. Du, G. Huang, C. Wang, X. Chen,
M. R. Porembka, J. Lea, A. E. Frankel, Y. X. Fu, Z. J. Chen, J. Gao, A STING-­activating
nanovaccine for cancer immunotherapy. Nat. Nanotechnol. 12, 648–654 (2017).
8. X. Wang, J. Wilhelm, W. Li, S. Li, Z. Wang, G. Huang, J. Wang, H. Tang, S. Khorsandi, Z. Sun,
B. Evers, J. Gao, Polycarbonate-­based ultra-­pH sensitive nanoparticles improve
therapeutic window. Nat. Commun. 11, 5828 (2020).
9. J. Yan, G. Wang, L. Xie, H. Tian, J. Li, B. Li, W. Sang, W. Li, Z. Zhang, Y. Dai, Engineering
radiosensitizer-­based metal-­phenolic networks potentiate STING pathway activation for
advanced radiotherapy. Adv. Mater. 34, e2105783 (2022).
10. X. Sun, Y. Zhang, J. Li, K. S. Park, K. Han, X. Zhou, Y. Xu, J. Nam, J. Xu, X. Shi, L. Wei, Y. L. Lei,
J. J. Moon, Amplifying STING activation by cyclic dinucleotide-­manganese particles for
local and systemic cancer metalloimmunotherapy. Nat. Nanotechnol. 16, 1260–1270
(2021).
11. Y. Lee, M. Jeong, J. Park, H. Jung, H. Lee, Immunogenicity of lipid nanoparticles and its
impact on the efficacy of mRNA vaccines and therapeutics. Exp. Mol. Med. 55, 2085–2096
(2023).
12. B. Li, A. Y. Jiang, I. Raji, C. Atyeo, T. M. Raimondo, A. G. R. Gordon, L. H. Rhym, T. Samad,
C. MacIsaac, J. Witten, H. Mughal, T. M. Chicz, Y. Xu, R. P. McNamara, S. Bhatia, G. Alter,
R. Langer, D. G. Anderson, Enhancing the immunogenicity of lipid-­nanoparticle mRNA
vaccines by adjuvanting the ionizable lipid and the mRNA. Nat. Biomed. Eng., (2023).
13. T. Sun, Y. S. Zhang, B. Pang, D. C. Hyun, M. Yang, Y. Xia, Engineered nanoparticles for drug
delivery in cancer therapy. Angew. Chem. Int. Ed. Engl. 53, 12320–12364 (2014).
14. Z. Zhang, Q. Wang, Q. Liu, Y. Zheng, C. Zheng, K. Yi, Y. Zhao, Y. Gu, Y. Wang, C. Wang,
X. Zhao, L. Shi, C. Kang, Y. Liu, Dual-­locking nanoparticles disrupt the PD-­1/PD-­L1
pathway for efficient cancer immunotherapy. Adv. Mater. 31, e1905751 (2019).
15. C. Zheng, Q. Wang, Y. Wang, X. Zhao, K. Gao, Q. Liu, Y. Zhao, Z. Zhang, Y. Zheng, J. Cao,
H. Chen, L. Shi, C. Kang, Y. Liu, Y. Lu, In situ modification of the tumor cell surface with
immunomodulating nanoparticles for effective suppression of tumor growth in mice.
Adv. Mater. 31, e1902542 (2019).
16. Z. Kang, C. Wang, Z. Zhang, Q. Liu, Y. Zheng, Y. Zhao, Z. Pan, Q. Li, L. Shi, Y. Liu, Spatial
distribution control of antimicrobial peptides through a novel polymeric carrier for safe
and efficient cancer treatment. Adv. Mater. 34, e2201945 (2022).
17. M. Rutter, N. Milosevic, A. David, Say no to drugs: Bioactive macromolecular therapeutics
without conventional drugs. J. Control. Release 330, 1191–1207 (2021).
18. J. Yang, L. Li, J. Kopecek, Biorecognition: A key to drug-­free macromolecular therapeutics.
Biomaterials 190-­191, 11–23 (2019).
19. D. S. Chen, I. Mellman, Oncology meets immunology: The cancer-­immunity cycle.
Immunity 39, 1–10 (2013).
20. Y. Zheng, Z. Zhang, Q. Liu, Y. Zhao, C. Zheng, J. Hao, K. Yi, Y. Wang, C. Wang, X. Zhao, L. Shi,
C. Kang, Y. Liu, Multifunctional nanomodulators regulate multiple pathways to enhance
antitumor immunity. ACS Appl. Bio. Mater. 3, 4635–4642 (2020).
21. M. Binnewies, E. W. Roberts, K. Kersten, V. Chan, D. F. Fearon, M. Merad, L. M. Coussens,
D. I. Gabrilovich, S. Ostrand-­Rosenberg, C. C. Hedrick, R. H. Vonderheide, M. J. Pittet,
R. K. Jain, W. Zou, T. K. Howcroft, E. C. Woodhouse, R. A. Weinberg, M. F. Krummel,
Understanding the tumor immune microenvironment (TIME) for effective therapy.
Nat. Med. 24, 541–550 (2018).
22. D. S. Chen, I. Mellman, Elements of cancer immunity and the cancer-­immune set point.
Nature 541, 321–330 (2017).
13 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
23. D. Bruni, H. K. Angell, J. Galon, The immune contexture and Immunoscore in cancer
prognosis and therapeutic efficacy. Nat. Rev. Cancer 20, 662–680 (2020).
24. W. Li, J. Yang, L. Luo, M. Jiang, B. Qin, H. Yin, C. Zhu, X. Yuan, J. Zhang, Z. Luo, Y. Du, Q. Li,
Y. Lou, Y. Qiu, J. You, Targeting photodynamic and photothermal therapy to the endoplasmic
reticulum enhances immunogenic cancer cell death. Nat. Commun. 10, 3349 (2019).
25. B. Feng, F. Zhou, B. Hou, D. Wang, T. Wang, Y. Fu, Y. Ma, H. Yu, Y. Li, Binary cooperative
prodrug nanoparticles improve immunotherapy by synergistically modulating immune
tumor microenvironment. Adv. Mater. 30, e1803001 (2018).
26. X. Duan, C. Chan, W. Lin, Nanoparticle-­mediated immunogenic cell death enables and
potentiates cancer immunotherapy. Angew. Chem. Int. Ed. Engl. 58, 670–680 (2019).
27. O. Kepp, L. Zitvogel, G. Kroemer, Clinical evidence that immunogenic cell death sensitizes
to PD-­1/PD-­L1 blockade. Onco. Targets. Ther. 8, e1637188 (2019).
28. M. Obeid, A. Tesniere, F. Ghiringhelli, G. M. Fimia, L. Apetoh, J. L. Perfettini, M. Castedo,
G. Mignot, T. Panaretakis, N. Casares, D. Metivier, N. Larochette, P. van Endert, F. Ciccosanti,
M. Piacentini, L. Zitvogel, G. Kroemer, Calreticulin exposure dictates the immunogenicity
of cancer cell death. Nat. Med. 13, 54–61 (2007).
29. M. Zhang, X. Jin, M. Gao, Y. Zhang, B. Z. Tang, A self-­reporting fluorescent salicylaldehyde-­
chlorambucil conjugate as a type-­II ICD inducer for cancer vaccines. Adv. Mater. 34,
e2205701 (2022).
30. Z. Zhang, Y. X. Yue, Q. Li, Y. Wang, X. Wu, X. Li, H. B. Li, L. Shi, D. S. Guo, Y. Liu, Design of
calixarene-­based ICD inducer for efficient cancer immunotherapy. Adv. Funct. Mater. 33,
2213967 (2023).
31. S. E. Dodgson, S. Kim, M. Costanzo, A. Baryshnikova, D. L. Morse, C. A. Kaiser, C. Boone,
A. Amon, Chromosome-­specific and global effects of aneuploidy in Saccharomyces
cerevisiae. Genetics 202, 1395–1409 (2016).
32. B. Tsai, Y. Ye, T. A. Rapoport, Retro-­translocation of proteins from the endoplasmic
reticulum into the cytosol. Nat. Rev. Mol. Cell Biol. 3, 246–255 (2002).
33. Y. Ye, Y. Shibata, C. Yun, D. Ron, T. A. Rapoport, A membrane protein complex
mediates retro-­translocation from the ER lumen into the cytosol. Nature 429,
841–847 (2004).
34. Y. Liu, J. Du, M. Yan, M. Y. Lau, J. Hu, H. Han, O. O. Yang, S. Liang, W. Wei, H. Wang, J. Li,
X. Zhu, L. Shi, W. Chen, C. Ji, Y. Lu, Biomimetic enzyme nanocomplexes and their use as
antidotes and preventive measures for alcohol intoxication. Nat. Nanotechnol. 8, 187–192
(2013).
35. Z. Zhang, Y. Gu, Q. Liu, C. Zheng, L. Xu, Y. An, X. Jin, Y. Liu, L. Shi, Spatial confined
synergistic enzymes with enhanced uricolytic performance and reduced toxicity for
effective gout treatment. Small 14, e1801865 (2018).
36. M. Yan, J. J. Du, Z. Gu, M. Liang, Y. F. Hu, W. J. Zhang, S. Priceman, L. L. Wu, Z. H. Zhou,
Z. Liu, T. Segura, Y. Tang, Y. F. Lu, A novel intracellular protein delivery platform based on
single-­protein nanocapsules. Nat. Nanotechnol. 5, 48–53 (2010).
37. Y. Liu, H.-­R. Jia, X. Han, F.-­G. Wu, Endoplasmic reticulum-­targeting nanomedicines for
cancer therapy. Smart Mater. Med. 2, 334–349 (2021).
38. S. Z. F. Phua, G. Yang, W. Q. Lim, A. Verma, H. Chen, T. Thanabalu, Y. Zhao, Catalase-­
integrated hyaluronic acid as nanocarriers for enhanced photodynamic therapy in solid
tumor. ACS Nano 13, 4742–4751 (2019).
Zhang et al., Sci. Adv. 10, eadk0716 (2024) 7 February 2024
39. Q. You, K. Zhang, J. Liu, C. Liu, H. Wang, M. Wang, S. Ye, H. Gao, L. Lv, C. Wang, L. Zhu,
Y. Yang, Persistent regulation of tumor hypoxia microenvironment via a bioinspired
Pt-­based oxygen nanogenerator for multimodal imaging-­guided synergistic
phototherapy. Adv. Sci. 7, 1903341 (2020).
40. H. Li, M. Wang, B. Huang, S. W. Zhu, J. J. Zhou, D. R. Chen, R. Cui, M. Zhang, Z. J. Sun,
Theranostic near-­infrared-­IIb emitting nanoprobes for promoting immunogenic
radiotherapy and abscopal effects against cancer metastasis. Nat. Commun. 12, 7149 (2021).
41. J. Xu, Y. Liu, Y. Li, H. Wang, S. Stewart, K. Van der Jeught, P. Agarwal, Y. Zhang, S. Liu,
G. Zhao, J. Wan, X. Lu, X. He, Precise targeting of POLR2A as a therapeutic strategy for
human triple negative breast cancer. Nat. Nanotechnol. 14, 388–397 (2019).
42. Y. Gu, Y. Zhao, Z. Zhang, J. Hao, Y. Zheng, Q. Liu, Y. Liu, L. Shi, An antibody-­like polymeric
nanoparticle removes intratumoral Galectin-­1 to enhance antitumor T-­cell responses in
cancer immunotherapy. ACS Appl. Mater. Interfaces 13, 22159–22168 (2021).
43. Y. Zhao, B. Han, J. Hao, Y. Zheng, J. Chai, Z. Zhang, Y. Liu, L. Shi, Bi-­specific macrophage
nano-­engager for cancer immunotherapy. Nano Today 41, 101313 (2021).
44. K. Tatsuno, T. Yamazaki, D. Hanlon, P. Han, E. Robinson, O. Sobolev, A. Yurter,
F. Rivera-­Molina, N. Arshad, R. L. Edelson, L. Galluzzi, Extracorporeal photochemotherapy
induces bona fide immunogenic cell death. Cell Death Dis. 10, 578 (2019).
45. Q. Chen, C. Wang, X. Zhang, G. Chen, Q. Hu, H. Li, J. Wang, D. Wen, Y. Zhang, Y. Lu, G. Yang,
C. Jiang, J. Wang, G. Dotti, Z. Gu, In situ sprayed bioresponsive immunotherapeutic gel for
post-­surgical cancer treatment. Nat. Nanotechnol. 14, 89–97 (2018).
46. C. T. Taylor, S. P. Colgan, Regulation of immunity and inflammation by hypoxia in
immunological niches. Nat. Rev. Immunol. 17, 774–785 (2017).
47. Q. Liu, K. Zhao, C. Wang, Z. Zhang, C. Zheng, Y. Zhao, Y. Zheng, C. Liu, Y. An, L. Shi, C. Kang,
Y. Liu, Multistage delivery nanoparticle facilitates efficient CRISPR/dCas9 activation and
tumor growth suppression in vivo. Adv. Sci. 6, 1801423 (2019).
48. L. Teng, Z.-­W. Shao, Y.-­S. He, J.-­Y. Lu, D.-­R. Zou, C.-­L. Feng, C.-­M. Dong, A glycosylated and
catechol-­crosslinked ε-­polylysine hydrogel: Simple preparation and excellent wound
hemostasis and healing properties. Chin. J. Polym. Sci. 40, 1110–1119 (2022).
Acknowledgments
Funding: This work was supported by the National Natural Science Foundation of China
(NSFC, grant nos. 22077073, 51933006, and 52203172) and the Fundamental Research Funds
for the Central Universities (Nankai University, grant no. 63206015). Author contributions:
Conceptualization: Y.L., Z.Z., and L.S. Methodology: Y.L., Z.Z., Z.P., Q.L., and Q.H. Investigation:
Z.Z., Z.P., Q.L., and Q.H. Visualization: Y.L. and Z.Z. Funding acquisition: Y.L., L.S., and Z.Z.
Supervision: Y.L. and L.S. Writing—original draft: Z.Z. Writing—review and editing: Y.L.
Competing interests: The authors declare that they have no competing interests. Data and
materials availability: All data needed to evaluate the conclusions in the paper are present in
the paper and/or the Supplementary Materials.
Submitted 1 August 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk0716
14 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

CANCER Copyright © 2024 The

Authors, some rights
Neoadjuvant radioimmunotherapy in pancreatic cancer reserved; exclusive
licensee American
enhances effector T cell infiltration and shortens their Association for the
Advancement of
distances to tumor cells Science. No claim to
original U.S.
Government Works.
Junke Wang1,2,3,4,5†, Jessica Gai1,2,3,4,6†, Tengyi Zhang1,2,3,4, Nan Niu1,2,3,4, Hanfei Qi1,2,7,8, Distributed under a
Dwayne L. Thomas II1,2,3,4,9, Keyu Li1,2,3,4, Tao Xia1,2,3,4, Christina Rodriguez1,2,4,6, Creative Commons
Rose Parkinson1,2,4,6,8, Jennifer Durham1,2,4,6,8, Thomas McPhaul1,2,3,9, Amol K. Narang1,2,3,4,10, Attribution
Robert A. Anders2,3,6,11, Arsen Osipov1,2,3,4,6,12, Hao Wang1,2,3,6,7,8, Jin He1,2,3,6,9, NonCommercial
Daniel A. Laheru1,2,3,4,6,8, Joseph M. Herman1,2,6,10,13, Valerie Lee1,2, Elizabeth M. Jaffee1,2,3,4,6,8, License 4.0 (CC BY-­NC).

Elizabeth D. Thompson1,2,3,4,11*, Qingfeng Zhu1,2,3,4,8*, Lei Zheng1,2,3,4,6,8,9*

Radiotherapy is hypothesized to have an immune-­modulating effect on the tumor microenvironment (TME) of

pancreatic ductal adenocarcinoma (PDAC) to sensitize it to anti–PD-­1 antibody (a–PD-­1) treatment. We collected
paired pre-­and posttreatment specimens from a clinical trial evaluating combination treatment with GVAX
vaccine, a–PD-­1, and stereotactic body radiation (SBRT) following chemotherapy for locally advanced PDACs
(LAPC). With resected PDACs following different neoadjuvant therapies as comparisons, effector cells in PDACs
were found to skew toward a more exhausted status in LAPCs following chemotherapy. The combination of
GVAX/a–PD-­1/SBRT drives TME to favor antitumor immune response including increased densities of GZMB+CD8+
T cells, TH1, and TH17, which are associated with longer survival, however increases immunosuppressive M2-­like
tumor-­associated macrophages (TAMs). Adding SBRT to GVAX/a–PD-­1 shortens the distances from PD-­1+CD8+
T cells to tumor cells and to PD-­L1+ myeloid cells, which portends prolonged survival. These findings have guided
the design of next radioimmunotherapy studies by targeting M2-­like TAM in PDACs.

INTRODUCTION
Pancreatic ductal adenocarcinoma (PDAC) has a dismal prognosis,
with an overall 5-­year survival rate of only 12% in the USA (1).
Approximately 30% of newly diagnosed PDAC presents as locally
advanced pancreatic cancer (LAPC) (2), which is considered surgi-
cally unresectable due to local involvement of adjacent vessels. A
common treatment strategy is induction chemotherapy followed by
chemoradiation or stereotactic body radiation (SBRT) (3). Histori-
cally, chemotherapy followed by SBRT resulted in downstaging and
successful R0 or R1 resection in 20% of LAPCs (4). We reported that
SBRT following induction chemotherapy resulted in higher rates of
complete pathologic responses (5) and attenuated lymphocyte
suppression (6). However, additional interventions are needed to
1
Department of Oncology, Johns Hopkins University School of Medicine, Baltimore,
MD 21287, USA. 2The Sidney Kimmel Comprehensive Cancer Center, Johns Hopkins
University School of Medicine, Baltimore, MD 21287, USA. 3The Pancreatic Cancer
Precision Medicine Center of Excellence Program, Johns Hopkins University School
of Medicine, Baltimore, MD 21287, USA. 4The Bloomberg Kimmel Institute for Cancer
Immunotherapy, Johns Hopkins University School of Medicine, Baltimore, MD
21287, USA. 5Division of Biliary Tract Surgery, Department of General Surgery, West
China Hospital, Sichuan University, Chengdu, Sichuan 610041, China. 6The Skip
Viragh Pancreatic Cancer Center, Johns Hopkins University School of Medicine,
Baltimore, MD 21287, USA. 7Quantitative Sciences Division, Department of Oncology,
Johns Hopkins University School of Medicine, Baltimore, MD 21287, USA. 8The Can-
cer Convergence Institute at Johns Hopkins, Johns Hopkins University School of
Medicine, Baltimore, MD 21287, USA. 9Department of Surgery, Johns Hopkins Uni-
versity School of Medicine, Baltimore, MD 21287, USA. 10Department of Radiation
Oncology, Johns Hopkins University School of Medicine, Baltimore, MD 21287,
USA. 11Department of Pathology, Johns Hopkins University School of Medicine,
Baltimore, MD 21287, USA. 12Cedars Sinai Medical Center, Los Angeles, CA, 90048,
USA. 13Northwell Health System, New Hyde Park, NY, 11042, USA.
*Corresponding author. Email: lzheng6@​jhmi.​edu (L.Z.); qzhu6@​jhmi.​edu (Q.Z.);
ethomp36@​jhmi.​edu (E.D.T.)
†These authors contributed equally to this work.
increase this rate because surgical resection is often required for
durable local tumor control.
Immunotherapy is a promising treatment alternative or adjunct
to standard of care for some aggressive cancers. Unfortunately, PDACs
respond poorly to immunotherapy (7). While tumor-­infiltrating
effector T cells (Teffs) portend a survival benefit in immunotherapy-­
responsive cancers, their presence is rare in the PDAC tumor micro-
environment (TME) (8). Immunotherapy has focused on manipulating
tumor antigen-­specific Teffs either by potentiating them through
vaccine therapy or T cell therapy or by unleashing their function
through immune checkpoint inhibitor (ICI) therapies such as anti–
programmed cell death protein 1 (PD-­1)/programmed cell death
ligand 1 (PD-­L1) antibodies. However, the major immune “defect”
in PDAC is its immunosuppressive TME, which impedes T cell
infiltration with immunotherapy (9).
Radiotherapy (RT) has been reported as one conventional treat-
ment modality for local or metastatic PDAC, although the clinical
benefit has not been clearly established (10–12). Emerging evidence
indicates that RT can prime the TME with Teffs by causing immuno-
genic cell death and then activating innate responses, which serves
as an “in situ vaccination” (13, 14). RT has the potential to improve
the efficacy of immunotherapies such as ICIs, which have achieved
clinical success for many cancers, but has not yet been successful for
PDACs. Preclinical and clinical studies suggest that the combination
of RT and ICIs exhibits a synergistic effect and enhanced antitumor
activity in varied tumors such as non–small cell lung cancer and
melanoma (15). In the KPC and Panc02 murine PDAC models, RT
combined with PD-­L1 blockade significantly improved tumor
response and prevented development of liver metastases (16). Al-
though we found that this combination offered local tumor control, it
failed to induce Teff infiltration into the tumors (17). Moreover, in

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 1 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
both subcutaneous and orthotopic mouse PDAC models, SBRT
combined with dual cytotoxic T lymphocyte-­associated antigen 4
(CTLA-­4) and PD-­1 blockades led to a modest growth delay in
tumors, which was associated with significant increases in CD8+
and CD4+ T cell infiltration (18). These findings provide a rationale
for the testing of RT combined with ICIs in patients with PDAC. In
a phase 1 clinical trial, SBRT combined with durvalumab or dur-
valumab plus tremelimumab demonstrated an acceptable safety
profile and a modest treatment benefit in patients with metastatic
PDAC (19). In a single-­arm phase 2 clinical trial, RT combined with
ipilimumab and nivolumab led to a short duration of disease control
and clinical benefit in 25 patients with metastatic, microsatellite
stable PDAC (20). In another randomized phase 2 study, SBRT com-
bined with nivolumab plus ipilimumab, compared to SBRT combined
with nivolumab, demonstrated a statistically significant improvement
in progression-­free survival and radiographic responses in both
SBRT-­targeted and non-­targeted lesions in patients with chemo
therapy-­refractory, metastatic PDAC (21). Nevertheless, the optimal
immune-­based mechanistic role of RT combined with immunotherapy
in treating PDAC remained unknown.
Our prior study showed that a granulocyte-­macrophage colony-­
stimulating factor-­secreting allogeneic PDAC cell vaccine (GVAX)
as a neoadjuvant therapy can induce the formation of tertiary lym-
phoid aggregates (LAs) and the up-­regulation of PD-­L1 in the LAs,
suggesting that GVAX may also prime PDAC to respond to ICIs
(22). Immunohistochemical analysis showed these LAs, with orga-
nized T cell and B cell zones with evidence of lymphoid neogenesis,
to be regulatory structures of adaptive immunity (22). More recently,
our platform trial (23, 24) testing either GVAX or GVAX in combi-
nation with anti–PD-­1 antibody (a–PD-­1) as neoadjuvant therapy
for resectable PDACs demonstrated that the combination of GVAX
and a–PD-­1 induced the infiltration of CD8+ T and CD4+ T cells into
tumors but reduced the amount of PD-­1+CD8+ and PD-­1+CD4+
T cells. Therefore, we and others have shown that it is possible to
turn the immunologic desert in PDACs into immune-­responsive
tumors; however, the response rates remain low in clinical trials in-
vestigating these regimens (25). To test the synergy between vaccine
therapy and RT in sensitizing PDACs for a–PD-­1 therapy, we
have completed a single-­center, single-­arm, phase 2 clinical trial
(NCT02648282) of combination GVAX, a–PD-­1 pembrolizumab,
and SBRT for patients with LAPC (26). Few studies have compre-
hensively investigated the immune-­modulating roles of RT in com-
bination with immunotherapy in human PDACs. Here, by performing
multiplex immunohistochemistry (IHC) on the prospectively col-
lected specimens from this phase 2 clinical trial, we conducted a
comprehensive assessment of the changes of TME in PDAC following
the combination treatment regimens of GVAX, a–PD-­1, and SBRT.
RESULTS
Combination treatment with pembrolizumab, vaccine
therapy, and SBRT induces CD4+ and CD8+ T lymphocyte
infiltration into posttreatment PDAC tumors
To investigate changes in tumor-­infiltrating immune cells in PDACs
after neoadjuvant immunotherapy with a GVAX vaccine therapy, an
a–PD-­1 pembrolizumab, and SBRT, we used a sequential staining
and stripping multiplex IHC technique (27). De-­identified speci-
mens were collected from 54 patients with evaluable LAPC who
were enrolled in a clinical trial (NCT02648282) (26) to receive
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
GVAX, pembrolizumab, and SBRT following induction chemotherapy
with either the FOLFIRINOX or the gemcitabine/abraxane combi-
nation chemotherapy regimens from July 2016 to January 2021
(fig. S1A). A pre-­GVAX/a–PD-­1/SBRT treatment core biopsy speci-
men was obtained from every participant. We obtained archived
surgically resected tumor samples from surgical candidates post-­
GVAX/a–PD-­1/SBRT treatment. For patients with unresectable
tumors, a posttreatment core biopsy was performed if feasible.
Among 54 evaluable patients’ specimens (table S1), those from
patients who have been followed on the clinical trial for at least
2 years between the start of the GVAX/a–PD-­1/SBRT treatment and
1 June 2022, including those who had died within 2 years, were
included in this study.
Slides were sectioned from formalin-­fixed paraffin-­embedded
(FFPE) blocks and stained by multiplex IHC with a panel of myeloid
and lymphoid cell markers (tables S2 and S3). The regions of interest
(ROIs) were selected within the biopsy tumor slides to cover a min-
imum of three different tumor cores from the same biopsy (fig. S1B).
A minimum of three ROIs in the tumor areas that each contained at
least one vaccine-­induced intratumoral tertiary LA were also selected
from each posttreatment surgically resected tumor and divided into
LAs and non-­LA tumor areas (fig. S1C). For posttreatment core
biopsy specimens, only tumor areas were selected. Pseudo-­colors
were assigned to each marker. ROIs were quantified as described
previously (27). As the signals from each staining were captured
separately, they would not interfere with each other. A combination
of positive and negative markers was used to identify each immune
cell subtype (Fig. 1A and table S4); thus, nonspecific staining of in-
dividual markers is unlikely to cause the misidentification of im-
mune cell subtypes (27). The density of each immune cell subtype
was calculated as the percentage of total cells identified by nuclear
staining within each ROIs. In comparison to resected PDACs from
patients treated by the standard of care chemotherapy and SBRT
(designated SBRT cohort, table S5), those who had upfront surgery
without neoadjuvant therapy (designated No-­NAT cohort, table S6),
and the published cohort from a neoadjuvant platform clinical trial
(NCT02451982) where patients received one cycle of GVAX and
a–PD-­1 nivolumab 2 weeks before the surgical resection (designated
GVAX/a–PD-­1 cohort) (23), the immune cell composition in the resected
PDACs from patients who received GVAX/a–PD-­1/SBRT appears to
have several key differences (Fig. 1B). Therefore, in the below studies,
we used specimens from these cohorts as comparisons (dataset S1).
To assess the effect of the GVAX/a–PD-­1/SBRT treatment on
patient’s PDAC immune-­microenvironment, we examined the
immune cell population in patient’s pretreatment biopsy samples
and the non-­LA tumor areas in posttreatment tumor specimens
from all available patients (Fig. 1C). We found that the combination
neoadjuvant therapy induced a variety of immune cell infiltration,
including a significant increase in the total CD4+ and CD8+ T cell
infiltration (Fig. 1C). We also found a significant increase in M1-­like
tumor-­associated macrophage (TAM) infiltration in posttreatment
tumors (Fig. 1C). There was no significant difference in the levels of
B cells, M2-­like TAM, and tumor-­associated neutrophil (TAN) infil-
tration between pre-­and posttreatment tumor samples (Fig. 1C and
fig. S2). These results suggested that the TME has changed toward
favoring an antitumor immune response following GVAX/a–PD-­1/
SBRT treatment.
Next, we compared immune cell populations between patients
who underwent PDAC resection and those with non-­resectable
2 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 1. Multiplex IHC analysis of PDACs with different types of neoadjuvant therapies. (A) Overlaid images of representative immune cell markers assigned with
pseudo-­colors in the resected PDACs following GVAX/a–PD-­1/SBRT treatment. Scale bars, 50 μm. (B) Immune cell composition in the resected PDAC TME following different
types of neoadjuvant therapies (n = 21 for “No-­NAT,” n = 17 for “SBRT,” n = 10 for “GVAX/a–PD-­1,” and n = 24 for “GVAX/a–PD-­1/SBRT”). (C) Summary of the density of all
immune cell subtypes examined as indicated in the pretreatment (n = 54) versus posttreatment (n = 45) tumor areas in PDACs with GVAX/a–PD-­1/SBRT treatment. (D and
E) Summary of the density of all immune cell subtypes analyzed as indicated in the pretreatment [(D); n = 24 for “resected” and n = 30 for “non-­resected”] and posttreatment
[(E); n = 24 for resected and n = 21 for non-­resected] tumor areas from the resected versus non-­resected PDACs with GVAX/a–PD-­1/SBRT treatment. Data were shown as
the means ± SEM or percentage and compared by unpaired t test; *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001; all others, not significant.

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 3 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

PDACs following GVAX/a–PD-­1/SBRT treatment (Fig. 1, D and E).
It should be noted that LAs were not consistently observed in post-
treatment biopsy specimens due to small tissue areas. Thus, tumor
areas were compared between posttreatment resected PDAC tumors
and non-­resected PDAC biopsy specimens. With the exception of TANs,
immune cell–subtype compositions between these two patient groups
were comparable before receiving GVAX/a–PD-­1/SBRT treatment
(Fig. 1D). In posttreatment tumor areas, there were significantly in-
creased infiltrations of immunosuppressive immune cell subtypes in
non-­resected PDACs in comparison to resected PDACs, including
T helper 2 (TH2), regulatory T cells (Tregs), and TANs, suggesting
that maintaining a less-­immunosuppressive TME as part of an im-
mune response to combination GVAX/a–PD-­ 1/SBRT treatment
may contribute to surgical resectability following neoadjuvant treat-
ment (Fig. 1E). Nevertheless, a higher density of TANs after induc-
tion chemotherapy and before GVAX/a–PD-­1/SBRT may contribute
to an inability to surgical resection following neoadjuvant therapy.
An increase in high-­quality CD8+ T cell population occurs
with combination GVAX + a–PD-­1 + SBRT treatment and
correlates with better survival in patients with
resectable PDAC
To understand the response and resistance mechanisms elicited by
our combination therapy in patients with PDAC, we examined
potential correlations between each subtype of T cells, myeloid cells,
and overall survival (OS). OS > 2 years was considered a “favorable”
survival for patients with PDAC following surgical resection based
on previously published correlative studies (22, 27). We did not
observe a correlation between OS and the infiltration of CD8+ T cells
and the granzyme B+ (GZMB+) CD8+ T cell subtype in tumor areas
before neoadjuvant immunotherapy in patients with either resectable
or non-­resectable PDAC (Fig. 2A). A positive correlation between
the infiltration of GZMB+CD8+ T cells, but not that of general
CD8+ T cells, in tumor areas and OS was found in patients with re-
sectable PDAC after treatment (Fig. 2B). However, this correlation
was not observed in patients with non-­resectable PDACs (Fig. 2B).
Furthermore, we observed a significant increase in the infiltration of
CD8+ T cells and GZMB+CD8+ T cells in tumor areas from pre-
treatment tumors to posttreatment tumors in resected PDACs as-
sociated with OS > 2 years, but not in unresected PDACs (Fig. 2C).
A significant increase in CD8+ T cells was also observed in unresected
PDACs associated with OS < 2 years (Fig. 2C), suggesting that an
increased infiltration of general CD8+ T cells is not prognostic.
Again, a significantly increased GZMB+CD8+ T cells, but not CD8+
T cells in general, in posttreatment LAs was associated with OS >
2 years in resected PDACs (Fig. 2D), suggesting that there is a specific
association of Teff cytotoxic function with survival. Overall, it ap-
pears that increased CD8+ T cells in the posttreatment PDACs may
simply be a response to the GVAX/a–PD-­1/SBRT treatment. How-
ever, increased GZMB+CD8+ T cells in the posttreatment PDACs
may not only signify an immune response to this treatment but also
represent improved clinical outcomes.

Fig. 2. Multiplex IHC analysis of CD8+ and GZMB+CD8+ T cells and their correlation with OS in PDACs with GVAX/a–PD-­1/SBRT treatment. (A and B) Correlative
analysis of OS with the densities of CD8+ and GZMB+CD8+ T cells, as indicated, in pretreatment [(A); n = 9 for “resected OS < 2y,” n = 15 for “resected OS > 2y,” n = 20 for
“non-­resected OS < 2y,” and n = 7 for “non-­resected OS > 2y”] and posttreatment [(B); n = 9 for resected OS < 2y, n = 15 for resected OS > 2y, n = 11 for non-­resected OS < 2y,
and n = 7 for non-­resected OS > 2y] tumor areas. (C) Correlative analysis of OS with the changes of the densities of CD8+ and GZMB+CD8+ T cells, as indicated, between
matched pre- and posttreatment tumor areas (n = 9 for resected OS < 2y, n = 15 for resected OS > 2y, n = 11 for non-­resected OS < 2y, and n = 7 for non-­resected OS >
2y). (D) Correlative analysis of OS with the densities of CD8+ and GZMB+CD8+ T cells, as indicated, in the posttreatment LAs in the resected PDACs (n = 9 for “OS < 2y” and
n = 15 for “OS > 2y”). 2y, 2 years. Data were shown as the means ± SD; comparisons by unpaired t test for (A), (B), and (D) and paired t test for (C); *P < 0.05 and **P < 0.01;
NS, not significant.

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 4 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

An increase in TH1 and TH17 density correlates with
improved OS in patients treated with the
combination GVAX/a–PD-­1/SBRT
We next evaluated the infiltration of CD4+ T cell subtypes in the
pre-­and posttreatment tissues and their correlation with OS (Fig. 3,
A to F, and fig. S3, A to R). By comparing paired pre-­and posttreatment
tumor areas, we found that the GVAX/a–PD-­1/SBRT treatment in-
duced a significant increase in the general CD4+ T cells, the TH1 and
TH17 subtypes, and the TH1/TH2 ratio in resected tumors with OS >
2 years, but not those with OS < 2 years (Fig. 3, A, B, D, and F). In
contrast, GVAX/a–PD-­1/SBRT treatment did not induce a signifi-
cant increase in the general CD4+ T cells, TH17, and the TH1/TH2
ratio in unresected tumors with OS > 2 years (Fig. 3, A, D, and F).
However, the densities of essentially all CD4+ T cell subtypes in
pretreatment tumors did not correlate with OS (fig. S3, A to F), sug-
gesting that treatment-­induced changes in CD4+ T cell subtypes
may have changed the clinical outcome. In contrast, there were
associations between higher TH2 density and OS < 2 years (fig. S3I),
between higher TH17 density and OS > 2 years (fig. S3L) in post-
treatment non-­LA tumor areas in resected tumors, and between higher
TH1 density and OS > 2 years in posttreatment LAs in resected
tumors (fig. S3N) and in posttreatment biopsies of unresected tumors
(fig. S3H), further supporting the same notion that the GVAX/a–
PD-­1/SBRT treatment may have changed the clinical outcomes.
These results are consistent with our previous findings showing that
TH17 signaling was enhanced in PDACs and associated with
improved survival following GVAX/a–PD-­1 treatment and that an
enhanced TH1 infiltration and TH1/TH2 ratio were also associated
with improved OS (23). In this current study, we also found a sig-
nificant increase in the infiltration of posttreatment Tregs, which are
known to play an immunosuppressive role according to previously
published studies (28, 29), in unresected PDACs associated with OS
> 2 years (Fig. 3E).
The EOMES-­mediated T cell exhaustion pathway predicts
clinical outcome of non-­resectable PDACs following
GVAX/a–PD-­1/SBRT treatment
We did not observe any significant difference in exhausted T cell
infiltration PD-­1+CD4+, PD-­1+CD8+, and EOMES+CD8+ T cells
between pre-­and posttreatment tumor areas, in tumor areas or LAs
between resected tumors and unresected tumors, or in tumor areas
or LAs between tumors associated with OS > 2 years and OS < 2 years

Fig. 3. Correlative analysis of OS with the changes of the densities of CD4+ T cell subtypes and myeloid cell subtypes between matched pre- and posttreatment
tumor areas in PDACs with GVAX/a–PD-­1/SBRT treatment. (A to F) Correlative analysis of OS with the changes of the densities of CD4+ T cell subtypes, as indicated,
between matched pre- and posttreatment tumor areas. General CD4+ T cells (A), TH1 (B), TH2 (C), TH1/TH2 ratio (D), Treg (E), and TH17 (F). (G to J) Correlative analysis of OS
with the changes of the densities of myeloid cell subtypes, as indicated, between matched pre- and posttreatment tumor areas. M1-­like TAM (G), M2-­like TAM (H), M1/M2-­like
TAM ratio (I), and TAN (J). Sample numbers [(A) to (J)]: n = 9 for resected OS < 2y, n = 15 for resected OS > 2y, n = 11 for non-­resected OS < 2y, and n = 7 for non-­resected
OS > 2y. All data were compared by paired t test; *P < 0.05, **P < 0.01, and ***P < 0.001.

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 5 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(fig. S4, A to L). The only exception is in the EOMES+CD8+ T cell
infiltration that exhibited a statistically significant elevation in post-
treatment tumor areas compared to that in pretreatment tumor ar-
eas in matched unresected PDACs associated with OS < 2 years
(fig. S4K). We further examined different subgroups of exhausted
CD8+ T cells and found no obvious difference in the composition
of PD-­1+EOMES+, PD-­1−EOMES+, PD-­1+EOMES−, and PD-­
1−EOMES− CD8+ T cells in pretreatment biopsy tumor areas be-
tween resected and unresected PDACs (fig. S4M). We found a
modest, but statistically significant, increase of the infiltration of
PD-­1+EOMES+ CD8+ T cells (2.20 versus 1.21% of total CD8+ T cells)
and a modest but statistically significant decrease in infiltration of
PD-­1−EOMES+ CD8+ T cells (6.15 versus 9.11% of total CD8+ T cells)
in unresected PDACs compared to resected PDACs (fig. S4N).
These results suggested that the EOMES-­mediated T cell exhaustion
pathway likely affects the clinical outcome of non-­resectable PDACs.
Nevertheless, this pathway does not seem to be a major T cell
exhaustion mechanism that accounts for the sensitivity and resistance
to the GVAX/a–PD-­1/SBRT treatment.
Changes in M2-­like TAMs following GVAX/a–PD-­1/SBRT
treatment correlates with patient outcomes
We also examined changes in myeloid cell populations following
neoadjuvant therapy (Fig. 3, G to J, and fig. S5, A to L), because they
are known suppressive regulators of T cell responses in the PDAC
TME (30). The densities of myeloid cell subpopulations including
TAMs and TANs in pretreatment tumor areas were mostly not
associated with OS in either resectable or non-­resectable patients
following GVAX/a–PD-­1/SBRT treatment (fig. S5, A to D). An
increase in M1-­like TAMs in posttreatment tumor areas compared
to that in pretreatment tumor areas was noted in both resected
PDACs associated with OS < 2 years and those with OS > 2 years
(Fig. 3G), suggesting that this increase is a response to the treatment
that does not affect clinical outcomes. However, an increase in the
M1/M2-­like TAM ratio, likely due to an increase of M1-­like TAMs,
was significantly associated with OS > 2 years in resected PDACs,
whereas a decrease in the M1/M2-­like TAM ratio, likely due to an
increase of M2-­like TAMs, was significantly associated with OS <
2 years in unresected PDACs (Fig. 3I). In addition, an increase in
the M2-­like TAMs was significantly associated with OS < 2 years
in unresected PDACs (Fig. 3H). Consistently, M2-­like TAMs in
posttreatment tumor areas in unresectable PDACs is the only my-
eloid cell subpopulation that negatively correlated with OS in the
posttreatment non-­LA tumor areas or LAs (fig. S5F). These results
suggest that a treatment-­induced increase in the M1/M2-­like TAM
ratio or M2-­like TAM density is not only an immune response to
this treatment but also potentially changes clinical outcomes. In
contrast, we did not observe any significant association between
TANs and OS in posttreatment tumor areas or LAs (fig. S5, H and L).
Similarly, the densities of PD-­L1+ myeloid cell subpopulations
including PD-­L1+ TAMs and PD-­L1+ TANs in pretreatment tumors
or posttreatment tumors did not correlate with OS in either resectable
or non-­resectable patients (fig. S5, M to U). When we compared
PD-­L1+ TAMs and PD-­L1+ TANs between pre-­and posttreatment
tumor areas from the same patients (fig. S5, V to X), we found that a
decrease in PD-­L1+ M2-­like TAMs was statistically significant in
resected PDACs associated with OS > 2 years (fig. S5W), suggest-
ing that the PD-­L1+ subtype may account for the impact of the
treatment-­induced increase in M2-­like TAMs on clinical outcomes.
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
Furthermore, there was a statistically significant decrease in PD-­L1+
TANs in unresected PDACs associated with OS < 2 years (fig. S5X),
suggesting that the role of TANs following GVAX/a–PD-­1/SBRT
treatment remains elusive and that TAN may be inferior to the PD-­
L1+ M2-­like TAM as the additional target in PDAC following induc-
tion chemotherapy and the GVAX/a–PD-­ 1/SBRT treatment in
contrast to its role in resectable PDAC described previously (23).
Untreated PDACs and chemotherapy/RT-­treated PDACs are
characterized by distinct TMEs
Our previous studies have demonstrated that GVAX/a–PD-­1 neo-
adjuvant therapy induces CD8+ T cell infiltration into the PDAC TME,
but not the high-­quality Teff subpopulation such as GZMB+CD8+
T cells (23). To assess the specific immune-­modulating effect of
chemotherapy and SBRT when given with the GVAX/a–PD-­1 regimen,
we compared the immune cell population between PDACs resected
following chemotherapy and the GVAX/a–PD-­1/SBRT treatment
from the current study and those from the prior study following
GVAX/a–PD-­1 treatment (23). We first compared the pretreatment
biopsy specimens before the GVAX/a–PD-­1/SBRT to those before
the GVAX/a–PD-­1 treatment (Fig. 4A). Of note, the pretreatment
biopsy specimens in the clinical trial of LAPCs were collected after
induction chemotherapy, thus demonstrating the effect of chemo-
therapy on the TME, whereas the pretreatment biopsy specimens in
the prior clinical trial of resectable PDAC were collected before
starting any treatment. We observed a significantly higher density of
infiltrated lymphoid cells in the pre-­radioimmunotherapy biopsies
from patients following induction chemotherapy, including CD8+
T cells and its exhausted subtype EOMES+CD8+ T cells, CD4+ T cells
and the TH1 subtype, PD-­1+ T cells, and DC-­SIGN+ immature den-
dritic cells (DCs), as well as a significantly lower density of natural
killer (NK) cells and TANs, comparing to those who had not started
any treatment (Fig. 4A). There were no significant differences be-
tween the levels of GZMB+CD8+ T cells, TH2 cells, TH17 cells, Tregs,
B cells, mast cells, TAMs, and DC-­LAMP+ mature DCs (Fig. 4A).
These results suggest that chemotherapy or a more advanced stage
of PDAC skews the effector cells in the TME toward a more ex-
hausted status.
SBRT enhances the GVAX/a–PD-­1–mediated increase of
GZMB+CD8+ Teffs
In contrast to the pre-­immunotherapy specimens, the post-­immunotherapy
resected tumor specimens exhibited nearly no significant differences
in immune cell subtypes between the non-­LA tumor areas following
either the GVAX/a–PD-­1/SBRT treatment or the GVAX/a–PD-­1 treat-
ment, including CD8+ T cells and its exhausted subtype EOMES+CD8+
T cells, PD-­1+ T cells, or NK cells (Fig. 4B). The only exception was
that tumors collected after GVAX/a–PD-­1/SBRT treatment had a
significantly lower TH2 infiltration than the tumors collected after
GVAX/a–PD-­1 treatment (Fig. 4B).
We also compared the immune cells compositions within the
GVAX-­induced LAs between these two cohorts (Fig. 4C). We ob-
served a significantly higher density of GZMB+CD8+ T cells and a
significantly lower density of Tregs in the posttreatment LAs in tumors
collected after the GVAX/a–PD-­1/SBRT treatment in comparison to
those collected after GVAX/a–PD-­1 treatment (Fig. 4C). This result
suggests that either chemotherapy, more likely SBRT, or both en-
hance the ability of the GVAX/a–PD-­1 treatment to increase high-­
quality effector CD8+ T cells and reduce Tregs in the setting of
6 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Summary of the density of all immune cell subtypes analyzed in resected PDACs with GVAX/a–PD-­1 treatment versus GVAX/a–PD-­1/SBRT treatment.
(A) Summary of the density of all immune cell subtypes analyzed as indicated in the pretreatment tumor areas in PDACs with GVAX/a–PD-­1 treatment (n = 10) versus
GVAX/a–PD-­1/SBRT treatment (n = 24). (B and C) Summary of the density of all immune cell subtypes analyzed as indicated in the posttreatment non-­LA tumor areas (B)
and LAs (C) in PDACs with GVAX/a–PD-­1 treatment (n = 10) versus GVAX/a–PD-­1/SBRT treatment (n = 24). Data were shown as the means ± SEM and compared by
unpaired t test; *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001; all others, not significant.

anti–PD-­1 treatment that is known to increase Tregs. It is likely based
on these results that SBRT mediates the immune modulating activ-
ity of the GVAX/a–PD-­1 therapy specifically in LAs as previously
shown (22, 23).
Although we did not observe an increase in GZMB+CD8+ T cells
in tumor areas following GVAX/a–PD-­1/SBRT treatment compared
to GVAX/a–PD-­1 treatment, we suspected that there may be a more
significant treatment-­induced change in GZMB+CD8+ T cells in the
tumors following GVAX/a–PD-­1/SBRT treatment. We therefore calcu-
lated the fold changes in CD8+ T cells and GZMB+CD8+ T cells from
pretreatment biopsy tumors to posttreatment resected tumors in
non-­LA tumor areas and compared them between the two cohorts
(fig. S6). Tumors following GVAX/a–PD-­1/SBRT treatment had a
significantly lower fold change in CD8+ T cells in the tumor areas
compared to tumors following GVAX/a–PD-­1 treatment (fig. S6A).
In contrast, tumors following GVAX/a–PD-­1/SBRT treatment did
not show a significantly higher fold change in GZMB+CD8+T cells in
the tumor areas compared to PDACs following GVAX/a–PD-­1 treat-
ment (fig. S6A). However, tumors following GVAX/a–PD-­1 /
S B RT treatment had a significantly lower fold change in CD8 +
T cells in the tumor areas in patients with OS < 2 years but not in
those with OS > 2 years in comparison with PDACs following GVAX/a–
PD-­1 treatment (fig. S6B). Nevertheless, compared to PDACs fol-
lowing GVAX/a–PD-­1 treatment, tumors following GVAX/a–PD-­1 /
SBRT treatment had a significantly higher fold change in GZMB+CD8+
T cells in the tumor areas in patients with OS > 2 years but not
in those with OS < 2 years (fig. S6C). These results suggested that,
although adding SBRT to GVAX/a–PD-­1 treatment leads to a de-
crease in the general population of CD8+ T cells in the tumors, it
leads to an increase in the high-­quality subpopulation of CD8+
T cells, such as the GZMB+CD8+ T cells, and subsequently results
in an improved survival.
Next, we assessed the role of GVAX/a–PD-­1 treatment in the
GVAX/a–PD-­1/SBRT combination treatment. We obtained a cohort
of PDACs resected after neoadjuvant gemcitabine or fluorouracil-­
based chemotherapy followed by SBRT (designated SBRT cohort)
for comparison (Fig. 5 and table S5). Within the tumors from this
SBRT cohort, we did not anticipate the observation of LAs often in-
duced by vaccine therapy. Thus, we compared only the tumor areas
from tumors in the SBRT cohort, and the non-­LA tumor areas fol-
lowing GVAX/a–PD-­1/SBRT treatment showed a significantly lower
density of CD8+ T cells, a significantly higher density of TH1, and a
significantly lower density of TH2 and Tregs. Tumors following
GVAX/a–PD-­1/SBRT treatment also showed a significantly lower
density of PD-­1+CD8+ T cells in the tumor areas. Nevertheless,
tumors following GVAX/a–PD-­1/SBRT treatment exhibited an
increased trend in GZMB+CD8+ T cells in the tumor areas, although
not statistically significant, likely due to the small sample size in the

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 7 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

spatial distances between several key tumor immune cells and tumors
cells in the different resected PDAC cohorts following different
types of neoadjuvant therapies. The ROIs selected from the above-­
described images were reprocessed in Halo software (Fig. 6A), and
the spatial distances between epithelial tumor cells marked by epi-
thelial cellular adhesion molecule (EpCAM), CD8+ T cells marked
by CD8, and myeloid cells marked by CSF-­1R were measured, re-
spectively (Fig. 6B and dataset S2). Distances from CD8+ T cell sub-
types (PD-­1+ and PD-­1−) and those from CSF-­1R+ myeloid cell
subtypes (PD-­L1+ and PD-­L1−; and CD68+CD163− M1-­like and
CD68+CD163+ M2-­like macrophages) were also measured.
As previously described (23, 34), we used 100 μm as the cutoff to
divide PDACs into those with the shorter average distances (<100 μm)
and those with the longer average distances (>100 μm) between two
Fig. 5. Summary of the density of all immune cell subtypes analyzed as indi-
cated in the posttreatment tumor areas in resected PDACs with SBRT treat-
types of cells. Here, we did the analysis by using different cutoffs to
ment (n = 17) versus GVAX/a–PD-­1/SBRT treatment (n = 24). Data were shown divide PDACs into two subgroups and found that different cutoffs
as the means ± SEM and compared by unpaired t test; *P < 0.05, **P < 0.01, and yielded consistent results and that the 100-­μm cutoff remains the
****P < 0.0001; all others, not significant. most optimal one (fig. S7). A significantly higher percentage of
tumors with shorter distances between PD-­1+CD8+ T cells and
tumor cells was observed in PDACs following GVAX/a–PD-­1/SBRT
SBRT cohort (P = 0.0595). Together, these results suggest that it treatment (13/17, 76.47%) than in those without neoadjuvant thera-
is not SBRT alone that modulates Teff infiltration but rather the py (2/19, 10.53%), with neoadjuvant SBRT therapy (4/15, 26.67%)
addition of SBRT to the GVAX/a–PD-­1 combo. or with GVAX/a–PD-­1 treatment (3/10, 30.00%), respectively (Fig. 6C).
In contrast, a significantly higher percentage of tumors with shorter
SBRT enhances GVAX/a–PD-­1 treatment–mediated distances between PD-­1−CD8+ T cells and tumor cells was observed
recruitment of immunosuppressive M2-­like TAMs in PDACs following GVAX/a–PD-­1 treatment (10/10, 100.00%) and
It has been described that the infiltration of immunosuppressive GVAX/a–PD-­1/SBRT treatment (16/17, 94.12%) than in those in
cells such as Tregs and M2-­like TAM into the TME are associated the SBRT cohort (9/15, 60.00%) (Fig. 6D).
with poor prognosis and aggressiveness of PDAC (31–33). Although The percentage of tumors with shorter distances between PD-­
we observed a decreased density of Tregs in LAs from PDAC tumors 1 +CD8 + T cells and PD-­L 1 +CSF-­1 R + myeloid cells in tumors
following GVAX/a–PD-­1/SBRT treatment compared to those fol- following GVAX/a–PD-­1/SBRT treatment (16/17, 94.12%) was higher
lowing GVAX/a–PD-­1 treatment, we found an increased M2-­like than that in tumors following GVAX/a–PD-­1 treatment (7/10, 70.00%);
TAM infiltration in LAs following GVAX/a–PD-­1/SBRT treatment however, the difference did not reach statistical significance (Fig. 6E).
(Fig. 4C). Therefore, our results suggest that adding SBRT to the GVAX/a– Nevertheless, a significantly higher percentage of tumors with shorter
PD-­1 treatment has led to increased M2-­like TAM infiltration. distances between PD-­1 +CD8 + T cells and PD-­L 1 +CSF-­1 R +
We did not observe differences in myeloid cell populations myeloid cells was observed in the cohort following GVAX/a–PD-­1/
including TANs, M1-­like TAMs, DC-­LAMP+ mature DCs, DC-­ SBRT treatment when compared to other cohorts (Fig. 6E). Furthermore,
SIGN+ immature DCs, and mast cells in tumors between the SBRT there were no statistically significant differences in the percentages
cohort and the cohort that received GVAX/a–PD-­1/SBRT treatment of tumors with shorter distances between CSF-­1R+ myeloid cells re-
(Fig. 5). However, we observed a significantly lower density of M2-­ gardless of PD-­L1 status and tumor cells among different cohorts
like TAMs in the cohort that received GVAX/a–PD-­1/SBRT treat- (fig. S8A). However, a significantly higher percentage of tumors
ment compared to the SBRT cohort (Fig. 5), further suggesting that with shorter distances between PD-­L1−CSF-­1R+ myeloid cells and
the addition of SBRT in the GVAX/a–PD-­1/SBRT treatment combi- tumor cells was noted in the cohort following GVAX/a–PD-­1/SBRT
nation caused an elevation in M2-­like TAM infiltration. treatment (17/17, 100.00%) than in the SBRT cohort (11/15, 73.33%)
(fig. S8A).
Adding SBRT to GVAX/a–PD-­1 treatment shortens We also compared the distances as a continuous variable be-
the distances between PD-­1+CD8+ T cells and tumor cells tween PD-­1+CD8+ or PD-­1−CD8+ T cells and tumor cells and dis-
and distances between PD-­1+CD8+ T cells and tances between PD-­1+CD8+ T cells and PD-­L1+CSF-­1R+ myeloid
+ +
PD-­L1 CSF-­1R myeloid cells cells in different cohorts (fig. S8B). The distances between PD-­1+CD8+
+
Previously, we found that CD8 Teffs mainly reside in LAs but rarely T cells and tumor cells and the distances between PD-­1+CD8+ T cells
in the vicinity of tumor cells following GVAX/a–PD-­1 treatment. and PD-­L1+CSF-­1R+ myeloid cells were significantly shorter in the
However, the higher density of GZMB+CD8+ T cells in LAs corre- GVAX/a–PD-­1/SBRT treatment cohort compared to any other
lated with shorter distance between CD8+ T cells and tumor cells cohort (fig. S8B). The distances between PD-­1−CD8+ T cells and
(23). Although our results showed that tumors following GVAX/ tumor cells were also significantly shorter in the GVAX/a–PD-­1/
a–PD-­1/SBRT treatment had a significantly higher fold change in SBRT treatment cohort compared to the SBRT or GVAX/a–PD-­1
GZMB+CD8+ T cells in the tumor areas in patients with OS > 2 years treatment cohort (fig. S8B). Comparing the GVAX/a–PD-­1/SBRT
compared to PDACs following GVAX/a–PD-­1 treatment (fig. S6C), treatment cohort and the No-­NAT cohort, the distances between
we wondered whether CD8+ Teffs were getting closer to tumor cells PD-­1−CD8+ T cells or general CD8+ T cells and tumor cells had
following GVAX/a–PD-­1/SBRT treatment. Thus, we measured the no significant difference (fig. S8B). It is possible that the distance

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 8 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. Spatial distance measurement between CD8+ T cells, CSF-­1R+ myeloid cells, and tumor cells in PDACs following different types of neoadjuvant therapies.
(A) A representative region from six-­marker multiplex IHC images integrated by Halo software. Scale bar, 100 μm. (B) Distance measurement schema of a representative
region. Positive signals (exemplified by the EpCAM staining signals) and the nearest neighbor cells (exemplified by CD8+ cells) are connected by gray lines whose lengths
are measured as the distances. (C to E) Comparison of percentages of PDACs with shorter (<100 μm) and longer (>100 μm) average distances between PD-­1+ or PD-­1−
CD8+ T cells and tumor cells [(C) and (D)] and between PD-­1+CD8+ T cells and PD-­L1+CSF-­1R+ myeloid cells (E) in PDACs following different types of neoadjuvant therapies
(n = 19 for No-­NAT, n = 15 for SBRT, n = 10 for GVAX/a–PD-­1, and n = 17 for GVAX/a–PD-­1/SBRT). Data were shown as the percentage and compared by chi-­square test;
*P < 0.05, **P < 0.01, and ****P < 0.0001.

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 9 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
between PD-­1−CD8+ T cells and tumor cells was shorter due to
the abundance of tumor cells in untreated tumors. Together, these
results showed that adding SBRT to GVAX/a–PD-­1 treatment short-
ened the distances between CD8+ T cells, particularly PD-­1+CD8+
T cells and tumor cells, and also the distance between PD-­1+CD8+
T cells and PD-­L1+CSF-­1R+ myeloid cells in PDACs. Thus, these
results suggest that adding SBRT to GVAX/a–PD-­1 treatment brings
more Teffs into the tumor vicinity and may also potentiate the effect
of anti–PD-­1/PD-­L1 blockade antibody by bringing PD-­1+CD8+
T cells and PD-­L1+CSF-­1R+ myeloid cells closer.
We also did the correlative analysis between the density of CD8+
T cells, PD-­1+CD8+ T cells, PD-­1−CD8+ T cells, or GZMB+CD8+
T cells and their distance to other cell types of interest (fig. S9, A to
H). The densities of CD8+ T cells and its subtypes including PD-­1+CD8+
T cells do not appear to have an impact on the distance between
PD-­1+CD8+ T cells and tumor cells or between other cell types of
interest and tumor cells (fig. S9, A to D). The higher density of
PD-­1+CD8+ T cells is in a strong trend associated with both the lon-
ger, not shorter, distance between PD-­1+CD8+ T cells and M1-­like
TAM and the longer distance between PD-­1+CD8+ T cells and
M2-­like TAM (fig. S9F). Although it is unknown how higher
density of PD-­1+CD8+ T cells could make their distance to TAMs
longer, these results further suggests that an increased abundance of
CD8+ T cells or a CD8+ T cell subtype is likely independent from
their distance to other cell types.
Bringing the PD-­1+CD8+ T cells closer to tumor cells and
closer to PD-­L1+ myeloid cells by adding SBRT to the
GVAX/a–PD-­1 treatment correlated with better OS
Next, we correlated the distances between CD8+ T cells, myeloid
cells, and tumor cells with OS (Fig. 7). The shorter distances be-
tween PD-­1+CD8+ T cells and tumor cells correlated with longer OS
in PDACs following GVAX/a–PD-­1/SBRT treatment; this correla-
tion is nearly statistically significant (P = 0.0889) (Fig. 7A). Such a
correlation was not present in other cohorts.
The distances between PD-­1−CD8+ T cells and tumor cells did
not correlate with OS in any of these cohorts (Fig. 7B). The dis-
tances between general CD8+ T cells and tumors similarly did not
correlate with OS in any of the cohorts (Fig. 7C). The shorter dis-
tances between PD-­1+CD8+ T cells and PD-­L1+CSF-­1R+ myeloid
cells correlated with longer OS only in the GVAX/a–PD-­1/SBRT
treatment cohort (Fig. 7D). These results suggest that adding
SBRT to the GVAX/a–PD-­1 treatment enhances activated T cell and
myeloid cell proximity that may also lead to prolonged survival
in PDACs.
Adding SBRT to GVAX/a–PD-­1 treatment elongated the
spatial distances between M2-­like TAMs and tumor cells
Since discovering that the shorter distances between PD-­1+CD8+
T cells and PD-­L1+CSF-­1R+ myeloid cells correlated with longer OS
only in the GVAX/a–PD-­1/SBRT treatment cohort, we hypothe-
sized that PD-­L1+CSF-­1R+ myeloid cells consist of TAMs. We found
that the distances between PD-­1+CD8+ T cells and M1-­like TAMs
or M2-­like TAMs, respectively, were near significantly or significantly
elongated in the GVAX/a–PD-­1/SBRT treatment cohort compared
to the GVAX/a–PD-­1 cohort (fig. S10A). Similarly, the distances
between tumor cells and M1-­like TAMs or M2-­like TAMs, respec-
tively, were significantly prolonged in the GVAX/a–PD-­1/SBRT
treatment cohort compared to the GVAX/a–PD-­1 cohort (fig. S10A).
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
Therefore, these TAMs must not fully overlap with PD-­L1+CSF-­1R+
myeloid cells. It is possible that the distances between PD-­L1− TAMs
and CD8+ T cells were modulated differently by the GVAX/a–PD-­1/
SBRT treatment. Supporting this notion, the distances between
PD-­1+CD8+ T cells and PD-­L1−CSF-­1R+ myeloid cells were not
significantly different between the GVAX/a–PD-­1/SBRT treatment
cohort and the GVAX/a–PD-­1 cohort (fig. S10A). Because of the
technical limitations in the number of markers to be included in the
distance analysis, we were not able to assess the distances between
PD-­1+CD8+ T cells and PD-­L1+ or PD-­L1− TAMs at present,
which should be measured in the future when there is appropriate
technology.
Nevertheless, we found that the distances between tumor cells
and M2-­like TAMs are potentially prognostic as they were notably
elongated in the patients with OS > 2 years compared to those
with OS < 2 years in the GVAX/a–PD-­1/SBRT treatment cohort
(P = 0.1049) but not in the GVAX/a–PD-­1 cohort (fig. S10B). This
result suggests that adding SBRT to GVAX/a–PD-­1 treatment may
offer survival benefit by elongating the distances between tumor
cells and pro-­cancerous macrophages.
DISCUSSION
To our knowledge, this is the most comprehensive studies to analyze
the major immune subtypes within the TME of LAPCs following
neoadjuvant immunotherapy and to compare the effect of different
types of neoadjuvant therapies on the TME of PDACs. This is the
largest study of PDACs with pre-­and post-­immunotherapy tumor
specimens and also with spatial analysis of immune cells by multiplex
IHC. The results from this study may potentially guide the design of
combination treatment strategies and development of previously
unidentified immunotherapy strategies for patients with PDAC,
particularly those at the locally advanced stage, as well as inform the
development of immunotherapeutics in combination with RT.
Both receiving chemotherapy and being diagnosed with a locally
advanced stage skew the effector cells in the TME of PDACs toward
a more exhausted state. However, SBRT specifically enhances Teff
population induced by a–PD-­1 combination immunotherapy that is
not observed when given with standard of care chemotherapy. This
response favored improved OS, suggesting an immunomodulatory
mechanism for RT in sensitizing PDAC tumors to anti–PD-­1 therapy.
Our results also suggested that such favorable antitumor immune
responses are not merely a result of SBRT but a combination of
GVAX/a–PD-­1/SBRT. These findings may explain the improved
survival observed in patients who received GVAX/a–PD-­1/SBRT as
the neoadjuvant therapy (26) and GVAX/a–PD-­1 as adjuvant therapy
(24) compared with that in patients who received only standard of
care SBRT before surgical resection (4).
The resectability rate in this study was 44% following the neo-
adjuvant therapy of GVAX/a–PD-­1/SBRT (26) compared to 20%
following SBRT in the historical control group (4), after both re-
ceived similar standard of care chemotherapies. A randomized con-
trolled study is needed to prove the significance of this approach for
improving the treatment of patients with PDAC. However, this
study suggests a potential underlying mechanism for improved
resectability, the reduction in suppressive T cell and TAN populations,
and the enhanced proximity of Teffs to tumor cells in the PDAC
TME. More specific targeting of these immunosuppressive immune
cell subtypes may further improve resectability.
10 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Correlative analysis of OS with the average distances between CD8+ T cells and tumor cells and between PD-­1+CD8+ T cells and PD-­L1+CSF-­1R+ myeloid
cells in PDACs following different types of neoadjuvant therapies. (A) Between PD-­1+CD8+ T cells and tumor cells; (B) between PD-­1−CD8+ T cells and tumor cells;
(C) between general CD8+ T cells and tumor cells; (D) between PD-­1+CD8+ T cells and PD-­L1+CSF-­1R+ myeloid cells. Sample numbers: n = 14 for “No-­NAT OS < 2y” and
n = 5 for “No-­NAT OS > 2y”; n = 8 for “SBRT OS < 2y” and n = 7 for “SBRT OS > 2y”; n = 3 for “GVAX/a–PD-­1 OS < 2y” and n = 7 for “GVAX/a–PD-­1 OS > 2y”; n = 9 for “GVAX/a–PD-­
1/SBRT OS < 2y” and n = 8 for “GVAX/a–PD-­1/SBRT OS > 2y.” Data were shown as the means ± SEM, and correlative analysis of OS with the spatial distances between two
groups was performed by unpaired t test; *P < 0.05, **P < 0.01, and ***P < 0.001.

Our previous study showed that a higher density of TANs in
pretreatment biopsies is associated with poorer prognosis following
neoadjuvant and adjuvant therapy of GVAX/a–PD-­1 for resectable
PDACs (23). Here, we did not observe a survival correlation with
TANs density either in pretreatment or in posttreatment tumors fol-
lowing GVAX/a–PD-­1/SBRT treatment within either the resected
PDACs or unresected PDACs. Nevertheless, the role of TANs in
RT remains interesting to be explored, as this study showed that
the density of TANs was significantly increased in the LAs in the
resected tumors following GVAX/a–PD-­1/SBRT treatment com-
pared to GVAX/a–PD-­1 treatment. It is possible that TANs would
only influence the resectability but not the long-­term outcome. In

Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024 11 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
contrast, further targeting M2-­like TAM is strongly supported by
the finding showing that RT-­induced M2-­like TAMs is associated
with poorer prognosis following GVAX/a–PD-­1/SBRT treatment.
We previously conducted a preclinical study targeting TAMs with a
CCR2/CCR5 dual antagonist in combination with radiation (35)
and have recently completed enrollment into a clinical trial of
CCR2/CCR5 dual antagonist in combination with nivolumab with
or without GVAX following induction chemotherapy and SBRT for
LAPCs (NCT03767582). The hypotheses raised by this study will be
tested once the outcome results of the second radioimmunotherapy
clinical trial become available.
There are several limitations in this study. First, some conclu-
sions drawn from this study are based on the comparisons between
tumors from different cohorts of patients obtained in different clinical
trials. Although such comparisons would be considered a strength
of this study, confounding factors such as different stages and neo-
adjuvant chemotherapies likely exist and vary among these patient
cohorts (table S7). Therefore, some conclusions must be further
validated by correlative studies using specimens from a randomized
clinical trial. Nevertheless, comparisons of specimens between dif-
ferent cohorts of patients have provided precious opportunities for
understanding the immunomodulating role of RT in combination
with immunotherapy in PDACs. Second, the analysis of the TME is
limited to multiplex IHC. Because of the treatment effect, tumor-­
infiltrating immune cells were only harvested in a small per-
centage of resected tumors; thus, a higher resolution study by
single-­cell RNA sequencing was not feasible. Bulk RNA sequenc-
ing was performed but could not distinguish gene expression
patterns between tumor and tumor-­infiltrating immune cells. Third,
although this study was conducted using the most sophisticated
cell-­to-­c ell distance measurements to date, this technique is
still limited by the number of overlaid markers and minimal densi-
ties of cells in the ROIs. Because of these limitations, we were
not able to assess the distances related to PD-­ L1+ TAMs and
+ +
GZMB CD8 T cells. This technique would also require further
improvement to feasibly apply this distance measurement method
to biopsy specimens.
Although our study supports an antitumor immunomodulating
role of RT, it also suggested that adding SBRT to the GVAX/a–PD-­1
treatment leads to an increase in M2-­like TAM infiltration, suggest-
ing that targeting M2-­like TAMs may further enhance antitumor
immune response. Supporting this notion, this study finds that add-
ing SBRT to GVAX/a–PD-­1 treatment elongates the distance be-
tween M2-­like TAMs and tumor cells in PDACs with longer survival.
We may learn in the near future whether targeting TAMs would further
enhance the antitumor immune response of a radio-­immunotherapy
in PDAC by analyzing the specimens from the above-­described
clinical trial of CCR2/CCR5 dual antagonist in combination with
nivolumab following chemotherapy and SBRT.
MATERIALS AND METHODS
Patients and specimens
From July 2016 to January 2021, 58 patients with LAPC at the Johns
Hopkins Hospital were enrolled in a clinical trial under a Johns
Hopkins Medical Institution Institutional Review Board approved
protocol (26). Written informed consent was obtained from all
patients. Of this group, 54 patients completed two cycles of cyclo-
phosphamide (Cy)/GVAX/a–PD-­1/SBRT therapy and were evaluable
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
for immune response (table S1). In previous studies, low-­dose Cy
before GVAX has proved to deplete immunosuppressive Tregs (22,
36). Twenty-­four patients had tumors resected after combination
therapy, and the remaining 30 patients were unresected. For the 54
enrolled patients, FFPE tissue blocks were obtained, including 54
pretreatment tumor specimens by endoscopic ultrasound-­guided
fine needle core biopsies and 45 posttreatment tumor specimens
composed of 24 surgically resected and 21 biopsied specimens. OS
was calculated from the first day of the first cycle of immunotherapy,
which would be 3 weeks before the first day of a 5-­day course of
SBRT at 6.6 gray (Gy) per day, to the date of death or to the last
follow-­up date on 1 June 2022. To conduct the survival correlation
analysis, patients whose OS is greater than 2 years or those who died
within 2 years were included. Patients from whom we obtained both
pre-­and posttreatment specimens were included in a matched pre-­
and posttreatment comparison.
Seventeen surgically resected PDACs following neoadjuvant
therapy of gemcitabine or fluorouracil-­based chemotherapy and
SBRT at 6.6 Gy per day for 5 days from March 2011 to October 2015
(designated SBRT cohort, table S5) and 21 surgically resected
PDACs without neoadjuvant therapy from February 1998 to May
2003 (designated No-­NAT cohort, table S6) were also included in
the study for comparison as historical control cohorts. For the SBRT
cohort, OS was calculated from the first day of SBRT treatment to the
date of death or to the last follow-­up date on 21 May 2020. For the No-­NAT
cohort, OS was calculated from the date of diagnosis to the date of
death or to the last follow-­up date on 16 October 2006. Such an old
cohort was chosen because neoadjuvant therapy was given to the
majority of localized PDACs more recently. FFPE tissue sections
were also obtained for these comparison groups. Furthermore, we
obtained our previously published immunological data without
modification from the 10 surgically resected PDACs following neoad-
juvant therapy of Cy/GVAX/a–PD-­1 (designated GVAX/a–PD-­1
cohort) (23), which were enrolled in another clinical trial
(NCT02451982).
Sequential IHC and image acquisition
Previous studies have described the sequential staining-­stripping
multiplex IHC protocol (27). In brief, deparaffinized 5-­μm FFPE tissue
sections were stained by hematoxylin (Dako, S3301), followed by
whole-­slide bright-­field scanning using NanoZoomer (Hamamatsu).
After subsequent endogenous peroxidase blocking and microwave
heat-­induced antigen retrieval with citrate buffer (pH 6.0; BioGenex,
HK080-­9K), sequential IHC iterations composed of staining, digital
image acquisition, and chromogen stripping were performed as pre-
viously described (27). Information on the optimized concentra-
tions and incubation times of primary antibodies, the incubation
time of horseradish peroxidase-­polymer antibody (Nichirei Biosciences
Inc.), and the aminoethyl carbazole (Vector Laboratories, SK-­4200)
reaction times for chromogenic detection were summarized in tables S2
and S3. Negative control images were chosen from those after the
last cycle of staining and chromogen stripping without primary
antibody staining. In this study, two separate staining panels were
designed, and each panel included 15 markers primarily defining
either myeloid or lymphoid cells (tables S2 and S3). Every immune
marker examined represented a specific aspect of the immune re-
sponse of interest, which was determined on the basis of the previ-
ously published study (27); thus, multiplicity was not controlled in
the analysis.
12 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Image processing and analysis
The digitized image processing and analyzing workflow included
sequential steps of image co-­registration, visualization, and quanti-
tative image analysis. First, digitized images scanned by NanoZoomer
were co-­registered via the specific CellProfiler v.2.2.1 pipeline designed
as previously described (27), shown in fig. S11. Tumor areas for
subsequent analyses were circled by pathologists. A minimum of
three rectangular ROIs (~3000 × 3000 pixels per ROI) in the vicinity
of tumor epithelia identified by the EpCAM staining, which are
known to be sufficiently representative of the whole-­tumor area as
previously described (27), were chosen (fig. S12). Each ROI from
the posttreatment surgically resected tumor contains at least one
vaccine-­induced intratumoral tertiary LA and would be divided
into LAs and non-­LA tumor areas for a secondary analysis (fig. S1C).
According to our prior studies (23), quantifying any three ROIs
would yield similar results, suggesting that three ROIs are represen-
tative for the whole-­tumor area. For pretreatment, posttreatment
biopsy specimens or resected tumors from the No-­NAT and SBRT
cohorts, ROIs were selected in the vicinity of tumor epithelia and
also contain CD45+ or CD3+ cells as previously described (27). Ac-
cording to our prior studies (23), quantifying any three ROIs that
cover a minimum of three biopsy cores yielded consistent im-
mune cell density results. Tumor epithelia, large intratumoral blood
vessels, and areas with tissue detachment were maximally excluded.
As shown in fig. S13, visualization was performed by converting co-­
registered images into individually pseudo-­colored single-­marker
staining images via ImageJ software v1.48 (National Institutes of
Health) (37) and Aperio ImageScope software v.12.3.2.8013 (Leica Bio-
systems). In the quantitative image analysis step, single-­cell segmen-
tation was obtained via the specific CellProfiler v.2.2.1 pipeline as
previously described (27), followed by image cytometry analysis and
quantification using the FCS Express 7 Image Cytometry software (De
Novo Software) (fig. S14). Lymphoid and myeloid cell subtypes were
defined by multiple markers as listed in table S4.
Spatial relationship assessment
All image analysis steps of the spatial relationship assessment were
performed via Halo Image Analysis Platform software v3.4.2986
(Indica Labs). The steps were as follows: (i) six different sequentially
stained images were deconvolved, registered, and fused into a pseu-
domultiplex immunofluorescent image; (ii) a pathologist annotated
the entire tumor region and excluded the necrotic region for the
next analysis; (iii) positive staining signals were identified by the
threshold adjustment of the signal intensity of each marker via
the HighPlex FL module; and (iv) distance and proximity mea-
surements were performed on the basis of the positive signal
detected with coordinates and by the spatial module included in
the software.
Quantification and statistical analysis
The density of each immune cell subtype was calculated by its per-
centage among all cells in ROIs. Student’s t test or chi-­square test
was used for comparison. Paired t test was used for the comparison
between pre-­and posttreatment biospecimens from same patients,
while unpaired t test was used for the comparison between two in-
dependent groups. Multiple testing correction was not used because
the analysis was either hypothesis testing or explorative in nature.
chi-­square test was used for the comparison between percentages.
All statistical analyses and figures were completed using GraphPad
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
Prism software v9.3.1 (GraphPad Software). A two-­sided P value of
<0.05 was considered statistically significant.
Supplementary Materials
This PDF file includes:
Figs. S1 to S14
Tables S1 to S7
Legends for datasets S1 and S2
Other Supplementary Material for this manuscript includes the following:
Datasets S1 and S2
REFERENCES AND NOTES
1. R. L. Siegel, K. D. Miller, N. S. Wagle, A. Jemal, Cancer statistics, 2023. CA Cancer J. Clin. 73,
17–48 (2023).
2. M. Hidalgo, Pancreatic cancer. N. Engl. J. Med. 362, 1605–1617 (2010).
3. J. Zhong, K. Patel, J. Switchenko, R. J. Cassidy, W. A. Hall, T. Gillespie, P. R. Patel, D. Kooby,
J. Landry, Outcomes for patients with locally advanced pancreatic adenocarcinoma
treated with stereotactic body radiation therapy versus conventionally fractionated
radiation. Cancer 123, 3486–3493 (2017).
4. G. Gemenetzis, V. P. Groot, A. B. Blair, D. A. Laheru, L. Zheng, A. K. Narang, E. K. Fishman,
R. H. Hruban, J. Yu, R. A. Burkhart, J. L. Cameron, M. J. Weiss, C. L. Wolfgang, J. He, Survival
in locally advanced pancreatic cancer after neoadjuvant therapy and surgical resection.
Ann. Surg. 270, 340–347 (2019).
5. J. He, A. B. Blair, V. P. Groot, A. A. Javed, R. A. Burkhart, G. Gemenetzis, R. H. Hruban,
K. M. Waters, J. Poling, L. Zheng, D. Laheru, J. M. Herman, M. A. Makary, M. J. Weiss,
J. L. Cameron, C. L. Wolfgang, Is a pathological complete response following neoadjuvant
chemoradiation associated with prolonged survival in patients with pancreatic cancer?
Ann. Surg. 268, 1–8 (2018).
6. A. T. Wild, J. M. Herman, A. S. Dholakia, S. Moningi, Y. Lu, L. M. Rosati, A. Hacker-­Prietz,
R. K. Assadi, A. M. Saeed, T. M. Pawlik, E. M. Jaffee, D. A. Laheru, P. T. Tran, M. J. Weiss,
C. L. Wolfgang, E. Ford, S. A. Grossman, X. Ye, S. G. Ellsworth, Lymphocyte-­sparing effect of
stereotactic body radiation therapy in patients with unresectable pancreatic cancer.
Int. J. Radiat. Oncol. Biol. Phys. 94, 571–579 (2016).
7. K. Foley, V. Kim, E. Jaffee, L. Zheng, Current progress in immunotherapy for pancreatic
cancer. Cancer Lett. 381, 244–251 (2016).
8. J. Kleponis, R. Skelton, L. Zheng, Fueling the engine and releasing the break:
Combinational therapy of cancer vaccines and immune checkpoint inhibitors.
Cancer Biol. Med. 12, 201–208 (2015).
9. A. A. Wu, V. Drake, H.-­S. Huang, S. Chiu, L. Zheng, Reprogramming the tumor
microenvironment: Tumor-­induced immunosuppressive factors paralyze T cells.
Onco. Targets. Ther. 4, e1016700 (2015).
10. A. Sultana, C. Tudur Smith, D. Cunningham, N. Starling, D. Tait, J. P. Neoptolemos,
P. Ghaneh, Systematic review, including meta-­analyses, on the management of locally
advanced pancreatic cancer using radiation/combined modality therapy. Br. J. Cancer 96,
1183–1190 (2007).
11. K. A. Goodman, C. Hajj, Role of radiation therapy in the management of pancreatic
cancer. J. Surg. Oncol. 107, 86–96 (2013).
12. X. Zhu, F. Li, X. Ju, Y. Shen, Y. Cao, F. Cao, F. Fang, S. Qing, Z. Jia, H. Zhang, Prediction of
overall survival after re-­irradiation with stereotactic body radiation therapy for pancreatic
cancer with a novel prognostic model (the SCAD score). Radiother. Oncol. 129, 313–318
(2018).
13. S. C. Formenti, S. Demaria, Radiation therapy to convert the tumor into an in situ vaccine.
Int. J. Radiat. Oncol. Biol. Phys. 84, 879–880 (2012).
14. S. Yasmin-­Karim, P. T. Bruck, M. Moreau, S. Kunjachan, G. Z. Chen, R. Kumar, S. Grabow,
S. K. Dougan, W. Ngwa, Radiation and local anti-­CD40 Generate an effective in situ
vaccine in preclinical models of pancreatic cancer. Front. Immunol. 9, 2030 (2018).
15. M. L. Aliru, J. E. Schoenhals, B. P. Venkatesulu, C. C. Anderson, H. B. Barsoumian,
A. I. Younes, L. S. K Mahadevan, M. Soeung, K. E. Aziz, J. W. Welsh, S. Krishnan, Radiation
therapy and immunotherapy: What is the optimal timing or sequencing? Immunotherapy
10, 299–316 (2018).
16. A. Azad, S. Yin Lim, Z. D'Costa, K. Jones, A. Diana, O. J. Sansom, P. Kruger, S. Liu,
W. G. McKenna, O. Dushek, R. J. Muschel, E. Fokas, PD-­L1 blockade enhances response of
pancreatic ductal adenocarcinoma to radiotherapy. EMBO Mol. Med. 9, 167–180 (2017).
17. K. Fujiwara, M. T. Saung, H. Jing, B. Herbst, M. K. Zarecki, S. Muth, A. Wu, E. Bigelow,
L. Chen, K. Li, N. Jurcak, A. B. Blair, D. Ding, M. Wichroski, J. Blum, N. Cheadle, J. Koenitzer,
L. Zheng, Interrogating the immune-­modulating roles of radiation therapy for a rational
combination with immune-­checkpoint inhibitors in treating pancreatic cancer. J.
Immunother. Cancer 8, e000351 (2020).
13 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
18. C. T. Stump, K. Roehle, N. Manjarrez Orduno, S. K. Dougan, Radiation combines with
immune checkpoint blockade to enhance T cell priming in a murine model of poorly
immunogenic pancreatic cancer. Open Biol. 11, 210245 (2021).
19. C. Xie, A. G. Duffy, G. Brar, S. Fioravanti, D. Mabry-­Hrones, M. Walker, C. M. Bonilla,
B. J. Wood, D. E. Citrin, E. M. Gil Ramirez, F. E. Escorcia, B. Redd, J. M. Hernandez, J. L. Davis,
B. Gasmi, D. Kleiner, S. M. Steinberg, J. C. Jones, T. F. Greten, Immune checkpoint blockade
in combination with stereotactic body radiotherapy in patients with metastatic
pancreatic ductal adenocarcinoma. Clin. Cancer Res. 26, 2318–2326 (2020).
20. A. R. Parikh, A. Szabolcs, J. N. Allen, J. W. Clark, J. Y. Wo, M. Raabe, H. Thel, D. Hoyos,
A. Mehta, S. Arshad, D. J. Lieb, L. C. Drapek, L. S. Blaszkowsky, B. J. Giantonio, C. D. Weekes,
A. X. Zhu, L. Goyal, R. D. Nipp, J. S. Dubois, E. E. van Seventer, B. E. Foreman, L. E. Matlack,
L. Ly, J. A. Meurer, N. Hacohen, D. P. Ryan, B. Y. Yeap, R. B. Corcoran, B. D. Greenbaum,
D. T. Ting, T. S. Hong, Radiation therapy enhances immunotherapy response in
microsatellite stable colorectal and pancreatic adenocarcinoma in a phase II trial. Nat.
Cancer 2, 1124–1135 (2021).
21. I. M. Chen, J. S. Johansen, S. Theile, J. X. Hjaltelin, S. I. Novitski, S. Brunak, J. P. Hasselby,
G. L. Willemoe, T. Lorentzen, K. Madsen, B. V. Jensen, E. E. Wilken, P. Geertsen, C. Behrens,
C. Nolsoe, K. L. Hermann, I. M. Svane, D. Nielsen, Randomized phase II study of nivolumab
with or without ipilimumab combined with stereotactic body radiotherapy for refractory
metastatic pancreatic cancer (CheckPAC). J. Clin. Oncol. 40, 3180–3189 (2022).
22. E. R. Lutz, A. A. Wu, E. Bigelow, R. Sharma, G. Mo, K. Soares, S. Solt, A. Dorman, A. Wamwea,
A. Yager, D. Laheru, C. L. Wolfgang, J. Wang, R. H. Hruban, R. A. Anders, E. M. Jaffee,
L. Zheng, Immunotherapy converts nonimmunogenic pancreatic tumors into
immunogenic foci of immune regulation. Cancer Immunol. Res. 2, 616–631 (2014).
23. K. Li, J. A. Tandurella, J. Gai, Q. Zhu, S. J. Lim, D. L. Thomas II, T. Xia, G. Mo, J. T. Mitchell,
J. Montagne, M. Lyman, L. V. Danilova, J. W. Zimmerman, B. Kinny-­Köster, T. Zhang,
L. Chen, A. B. Blair, T. Heumann, R. Parkinson, J. N. Durham, A. K. Narang, R. A. Anders,
C. L. Wolfgang, D. A. Laheru, J. He, A. Osipov, E. D. Thompson, H. Wang, E. J. Fertig,
E. M. Jaffee, L. Zheng, Multi-­omic analyses of changes in the tumor microenvironment of
pancreatic adenocarcinoma following neoadjuvant treatment with anti-­PD-­1 therapy.
Cancer Cell 40, 1374–1391.e7 (2022).
24. T. Heumann, C. Judkins, K. Li, S. J. Lim, J. Hoare, R. Parkinson, H. Cao, T. Zhang, J. Gai,
B. Celiker, Q. Zhu, T. McPhaul, J. Durham, K. Purtell, R. Klein, D. Laheru, A. de Jesus-­Acosta,
D. T. le, A. Narang, R. Anders, R. Burkhart, W. Burns, K. Soares, C. Wolfgang, E. Thompson,
E. Jaffee, H. Wang, J. He, L. Zheng, A platform trial of neoadjuvant and adjuvant antitumor
vaccination alone or in combination with PD-­1 antagonist and CD137 agonist antibodies
in patients with resectable pancreatic adenocarcinoma. Nat. Commun. 14, 3650 (2023).
25. T. Tsujikawa, T. Crocenzi, J. N. Durham, E. A. Sugar, A. A. Wu, B. Onners, J. M. Nauroth,
R. A. Anders, E. J. Fertig, D. A. Laheru, K. Reiss, R. H. Vonderheide, A. H. Ko, M. A. Tempero,
G. A. Fisher, M. Considine, L. Danilova, D. G. Brockstedt, L. M. Coussens, E. M. Jaffee, D. T. le,
Evaluation of cyclophosphamide/GVAX pancreas followed by Listeria-­Mesothelin
(CRS-­207) with or without nivolumab in patients with pancreatic cancer. Clin. Cancer Res.
26, 3578–3588 (2020).
26. V. Lee, D. Ding, C. Rodriguez, B. Onners, A. Narang, J. Meyer, J. M. Herman,
A. Hacker-­Prietz, R. A. Burkhart, W. Burns, J. He, A. de Jesus-­Acosta, R. B. Klein, D. Laheru,
D. T. le, A. Ryan, E. Sugar, L. Zheng, A phase 2 study of cyclophosphamide (CY), GVAX,
pembrolizumab (Pembro), and stereotactic body radiation (SBRT) in patients (pts) with
locally advanced pancreas cancer (LAPC). J. Clin. Oncol. 39, 4134–4134 (2021).
27. T. Tsujikawa, S. Kumar, R. N. Borkar, V. Azimi, G. Thibault, Y. H. Chang, A. Balter,
R. Kawashima, G. Choe, D. Sauer, E. el Rassi, D. R. Clayburgh, M. F. Kulesz-­Martin, E. R. Lutz,
L. Zheng, E. M. Jaffee, P. Leyshock, A. A. Margolin, M. Mori, J. W. Gray, P. W. Flint,
L. M. Coussens, Quantitative multiplex immunohistochemistry reveals myeloid-­inflamed
tumor-­immune complexity associated with poor prognosis. Cell Rep. 19, 203–217 (2017).
28. X. Wang, M. Lang, T. Zhao, X. Feng, C. Zheng, C. Huang, J. Hao, J. Dong, L. Luo, X. Li, C. Lan,
W. Yu, M. Yu, S. Yang, H. Ren, Cancer-­FOXP3 directly activated CCL5 to recruit FOXP3+Treg
cells in pancreatic ductal adenocarcinoma. Oncogene 36, 3048–3058 (2017).
29. D. Mougiakakos, A. Choudhury, A. Lladser, R. Kiessling, C. C. Johansson, Regulatory T cells
in cancer. Adv. Cancer Res. 107, (2010).
30. S. Pushalkar, M. Hundeyin, D. Daley, C. P. Zambirinis, E. Kurz, A. Mishra, N. Mohan, B. Aykut,
M. Usyk, L. E. Torres, G. Werba, K. Zhang, Y. Guo, Q. Li, N. Akkad, S. Lall, B. Wadowski,
J. Gutierrez, J. A. Kochen Rossi, J. W. Herzog, B. Diskin, A. Torres-­Hernandez, J. Leinwand,
W. Wang, P. S. Taunk, S. Savadkar, M. Janal, A. Saxena, X. Li, D. Cohen, R. B. Sartor,
Wang et al., Sci. Adv. 10, eadk1827 (2024) 7 February 2024
D. Saxena, G. Miller, The pancreatic cancer microbiome promotes oncogenesis by
induction of innate and adaptive immune suppression. Cancer Discov. 8, 403–416 (2018).
31. Y. Tang, X. Xu, S. Guo, C. Zhang, Y. Tang, Y. Tian, B. Ni, B. Lu, H. Wang, An increased
abundance of tumor-­infiltrating regulatory T cells is correlated with the progression and
prognosis of pancreatic ductal adenocarcinoma. PLOS One 9, e91551 (2014).
32. N. Hiraoka, K. Onozato, T. Kosuge, S. Hirohashi, Prevalence of FOXP3+ regulatory T cells
increases during the progression of pancreatic ductal adenocarcinoma and its
premalignant lesions. Clin. Cancer Res. 12, 5423–5434 (2006).
33. S. Yang, Q. Liu, Q. Liao, Tumor-­associated macrophages in pancreatic ductal
adenocarcinoma: Origin, polarization, function, and reprogramming. Front. Cell Dev. Biol.
8, 607209 (2020).
34. C. Yan, H. Huang, Z. Zheng, X. Ma, G. Zhao, T. Zhang, X. Chen, F. Cao, H. Wei, J. Dong,
P. Tang, H. Jiang, M. Wang, P. Wang, Q. Pang, W. Zhang, Spatial distribution of tumor-­
infiltrating T cells indicated immune response status under chemoradiotherapy plus
PD-­1 blockade in esophageal cancer. Front. Immunol. 14, 1138054 (2023).
35. J. Wang, M. T. Saung, K. Li, J. Fu, K. Fujiwara, N. Niu, S. Muth, J. Wang, Y. Xu, N. Rozich,
H. Zlomke, S. Chen, B. Espinoza, M. K. Henderson, V. Funes, B. Herbst, D. Ding,
C. Twyman-­Saint Victor, Q. Zhao, A. Narang, J. He, L. Zheng, CCR2/CCR5 inhibitor permits
the radiation-­induced effector T cell infiltration in pancreatic adenocarcinoma. J. Exp.
Med. 219, (2022).
36. D. Laheru, E. Lutz, J. Burke, B. Biedrzycki, S. Solt, B. Onners, I. Tartakovsky, J. Nemunaitis,
D. le, E. Sugar, K. Hege, E. Jaffee, Allogeneic granulocyte macrophage colony-­stimulating
factor-­secreting tumor immunotherapy alone or in sequence with cyclophosphamide for
metastatic pancreatic cancer: A pilot study of safety, feasibility, and immune activation.
Clin. Cancer Res. 14, 1455–1463 (2008).
37. C. A. Schneider, W. S. Rasband, K. W. Eliceiri, NIH Image to ImageJ: 25 years of image
analysis. Nat. Methods 9, 671–675 (2012).
Acknowledgments
Funding: This study was supported by NIH grants R01 CA169702 (L.Z.), R01 CA197296 (L.Z. and
E.M.J.), and P50 CA062924 (E.M.J. and L.Z.); a MISP grant from Merck Sharp & Dohme LLC,
a subsidiary of Merck & Co. Inc., Rahway, NJ, USA (L.Z.); a Gateway Foundation grant (L.Z.); and
the NCI Cancer Center Support Grant P30 CA006973. Author contributions: Conceptualization:
A.K.N., R.A.A., J.H., D.A.L., J.M.H., V.L., E.M.J., and L.Z. Methodology: N.N., K.L., A.K.N., A.O., D.A.L.,
V.L., E.M.J., E.D.T., Q.Z., and L.Z. Investigation: J.W., J.G., T.Z., N.N., D.L.T., T.X., C.R., J.D., T.M., R.A.A.,
A.O., J.H., V.L., and E.D.T. Visualization: J.W., J.G., R.A.A., A.O., Q.Z., and L.Z. Supervision: D.A.L.,
J.M.H., E.M.J., and L.Z. Validation: J.W., R.P., J.D., A.O., and E.M.J. Formal analysis: J.W., J.G., H.Q.,
T.M., H.W., and Q.Z. Data curation: J.W., C.R., R.P., A.O., and H.W. Software: T.M. and Q.Z.
Writing—original draft: J.W., J.M.H., and L.Z. Writing—review and editing: J.W., J.G., D.L.T., C.R.,
R.P., R.A.A., A.O., H.W., J.M.H., E.M.J., E.D.T., and L.Z. Funding acquisition: E.M.J. and L.Z. Project
administration: C.R., R.P., J.D., H.W., D.A.L., and L.Z. Resources: J.W., D.L.T., A.O., J.H., E.M.J., E.D.T.,
and L.Z. Competing interests: L.Z. receives grant support from Bristol-­Meyer Squibb, Merck,
AstraZeneca, iTeos, Amgen, NovaRock, Inxmed, Halozyme, and Abmeta. L.Z. is a paid
consultant/Advisory Board member at Biosion, Alphamab, NovaRock, Ambrx, Akrevia/Xilio,
QED, Natera, Novagenesis, Snow Lake Capitals, BioArdis, Tempus, Amberstone, Pfizer, Tavotek,
and Mingruizhiyao. L.Z. holds shares at Alphamab, Amberstone, and Mingruizhiyao. E.M.J.
receives grant support from Lustgarten Foundation, Bristol-­Meyer Squibb, Genentech, and
AstroZeneca; is a paid consultant for NextCure, Genocea, DragonFly, Stimit, CSTONE, Achilles,
and Candel; is on the advisory board of Parker Institute and Break Through Cancer; is a founder
of Abmeta Biotech; and is the Chief Medical Advisor for the Lustgarten Foundation. J.M.H.
consults and holds equity from Histosonics, consults for BTG, and receives research support to
the Northwell Health System from Canopy Cancer Collective/1440 Foundation. R.A.A. serves on
the advisory board of RAPT Therapeutics, Astra Zeneca, Merck SD, and Bristol-­Myers Squib and
receives research funding from RAPT Therapeutic, Bristol-­Myers Squib, and the National
Institute of Health. There is no relevant conflict of interest disclosed by all other authors.
Data and materials availability: All data needed to evaluate the conclusions in the paper are
present in the paper and/or the Supplementary Materials.
Submitted 7 August 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adk1827
14 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

CANCER Copyright © 2024 The

Authors, some rights
Atypical inflammatory kinase IKBKE phosphorylates reserved; exclusive
licensee American
and inactivates FoxA1 to promote liver tumorigenesis Association for the
Advancement of
Science. No claim to
Bing Gao1,2†, Xueji Wu2†, Lang Bu2†, Qiwei Jiang2†, Lei Wang2, Haining Liu1,2, Xiaomei Zhang2, original U.S.
Yuanzhong Wu3, Xiaoxing Li1,2, Jingting Li2, Ying Liang4*, Lixia Xu2,5*, Wei Xie2*, Jianping Guo2* Government Works.
Distributed under a
Physiologically, FoxA1 plays a key role in liver differentiation and development, and pathologically exhibits an Creative Commons
oncogenic role in prostate and breast cancers. However, its role and upstream regulation in liver tumorigenesis Attribution
remain unclear. Here, we demonstrate that FoxA1 acts as a tumor suppressor in liver cancer. Using a CRISPR-­based NonCommercial
kinome screening approach, noncanonical inflammatory kinase IKBKE has been identified to inhibit FoxA1 tran- License 4.0 (CC BY-­NC).
scriptional activity. Notably, IKBKE directly binds to and phosphorylates FoxA1 to reduce its complex formation
and DNA interaction, leading to elevated hepatocellular malignancies. Nonphosphorylated mimic Foxa1 knock-­in
mice markedly delay liver tumorigenesis in hydrodynamic transfection murine models, while phospho-­mimic
Foxa1 knock-­in phenocopy Foxa1 knockout mice to exhibit developmental defects and liver inflammation. Nota-
bly, Ikbke knockout delays diethylnitrosamine (DEN)-­induced mouse liver tumor development. Together, our find-
ings not only reveal FoxA1 as a bona fide substrate and negative nuclear effector of IKBKE in hepatocellular
carcinioma (HCC) but also provide a promising strategy to target IKBEK for HCC therapy.

INTRODUCTION
Liver cancer is the fourth leading cause of cancer death in both sexes
(8.2% of all cancer deaths) (1). It is worth noting that hepatocellular
carcinoma (HCC), the most common form of liver cancer, is usually
caused by excessive alcohol consumption, viral hepatitis [hepatitis B
virus (HBV) and HCV)], or nonalcoholic fatty liver disease with
inflammation that promotes liver fibrosis and tumorigenesis (2, 3).
Therefore, HCC is considered an inflammatory disease. Thus, in a
murine model of carcinogen-­induced HCC, inhibition of signal
transducer and activator of transcription 3 (STAT3) in hepatocytes
successfully blocked HCC development (4). In addition, proinflam-
matory cytokines such as interleukin-­6 (IL-­6) and tumor necrosis fac-
tor (TNF), which activate STAT3 and nuclear factor κB (NF-­κB)
signaling, respectively, play major roles in initiation and progression
of HCC (5, 6). Notably, inhibition of NF-­κB increased hepatocyte pro-
liferation in diethylnitrosamine (DEN)–induced and HBV surface
antigen–driven HCC murine models (7, 8). However, there are cur-
rently no effective strategies to combat HCC by targeting the NF-­
κB pathway.
The inhibitor of NF-­κB kinase subunit epsilon (IKBKE) belongs
to noncanonical IκB kinases. IKBKE and TNF receptor-­associated
factor (TANK)–binding kinase 1 (TBK1) have been reported to be
involved in energy storage and inflammatory programs in the liver
and adipose tissues, where abnormality of IKBKE contributes to obe-
sity (9). In addition to regulating NF-­κB (10) and STAT1–interferon
regulatory factor 3/7 (IRF3/7)–mediated inflammatory signaling
pathways (11), the function of IKBKE in tumorigenesis has also been
1
Center of Hepato-­Pancreate-­Biliary Surgery, the First Affiliated Hospital, Sun Yat-­sen
University, Guangzhou, Guangdong, 510275, China. 2Institute of Precision Medicine,
the First Affiliated Hospital, Sun Yat-­ sen University, Guangzhou, Guangdong
510275, China. 3State Key Laboratory of Oncology in South China, Collaborative
Innovation Center for Cancer Medicine, Sun Yat-­sen University Cancer Center,
Guangzhou, Guangdong 510060, China. 4Department of Nephrology, Guangzhou
Eighth People′s Hospital, Guangzhou Medical University, Guangdong 510060,
China. 5Department of Oncology, the First Affiliated Hospital, Sun Yat-­sen Univer-
sity, Guangzhou, Guangdong 510275, China.
*Corresponding author. Email: gz8hliangying@​yeah.​net (Y.L.); xulixia@​mail.​sysu.​
edu.​cn (L.X.); xiew56@​mail.​sysu.​edu.​cn (W.X.); guojp6@​mail.​sysu.​edu.​cn (J.G.)
†These authors contributed equally to this work.
decoded in many other ways. For example, IKBKE integrates with
STAT3 activation to compose a cytokine signaling network in a subset
of immune-­activated triple-­negative breast cancer (12), and IKBEK
also promotes the phosphorylation of Yes1-­associated transcriptional
regulator (YAP) to alleviate the effect of YAP on cellular antiviral
response (13), which occurs in approximately 30% patients with
HCC (14). As an inducible kinase, IKBKE could be induced by vari-
ous cytokines or chemokines, including phorbol 12-­myristate
13-­acetate (PMA) and lipopolysaccharide (LPS) (15, 16). Conceiv-
ably, inflammation-­related HCC would bear a high level of IKBKE;
however, the function of IKBKE in HCC is still under exploration.
Hepatocyte nuclear factor 3-­alpha (FoxA1), initially identified as a
liver-­specific transcription factor involved in hepatocyte differentia-
tion and development (17), has recently attracted more attention due
to its oncogenic role in tumorigenesis, especially in breast and pros-
tate cancer settings. FoxA1 can assist estrogen receptor (ER) or andro-
gen receptor (AR) to regulate downstream gene expression (18–20)
and has been considered as an effective target for prostate cancer
therapy due to its gain-­of-­function mutations in boosting prostate tu-
morigenesis (21). Sexual dimorphism of HCC that estrogens prevent
and androgens promote liver cancer, which is largely dependent on
the presence of Foxa1 and Foxa2 in DEN-­induced HCC murine mod-
el (22). Recently, FoxA1 has been suggested to play potential tumor
suppressor role in certain studies (23). Therefore, urgent efforts are
desired to identify FoxA1 upstream regulators and consequently de-
velop effective therapeutic interventions for HCC.
In conclusion, here, we initially reveal the tumor suppressor role
of FoxA1 in HCC and further integrate CRISPR-­based screening
approach, biochemical studies, and genetic mouse models to demon-
strate that accumulated IKBKE-­mediated phosphorylation and inac-
tivation of FoxA1 contribute to liver tumorigenesis, which would be
efficiently alleviated by IKBKE inhibitors.
RESULTS
FoxA1 plays a tumor suppressor role in HCC
Recently, the critical role and clinical relevance of FoxA1 as an onco-
gene in prostate and breast cancers has been established (18, 21);

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 1 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

however, its potential role in HCC has just begun to be explored (24– and K, and fig. S2, H and I), including cell cycle–related genes such as
27). Here, we observed that depletion of FoxA1 dramatically enhanced BUB1 mitotic checkpoint serine/threonine kinase (BUB1), cyclin D1
hepatic cancer cell colony formation, anchorage growth (Fig. 1, A and (CCND1), and the oncogene mesenchymal epithelial transition
B, and fig. S1, A to C), and tumor growth in mice (Fig. 1, C and D, and (MET) (Fig. 1, L and M). These findings collectively suggest that IK-
fig. S1, D to F). Depletion of FoxA1 significantly increased HCC or- BKE inhibits FoxA1 transcriptional activation (Fig. 1N).
ganoid formation (Fig. 1, E and F). Consistent with these findings,
re-­introduction of FoxA1 in FoxA1-­depleted HCC cells effectively IKBKE exerts oncogenic function in HCC by repressing FoxA1
rescued FoxA1 tumor suppressor effect in impairing cell colony for- Because of the functional negative regulation of FoxA1 by IKBKE in
mation (fig. S1, G and H). These data collectively implicate a potential HCC (Fig. 1, I to N), we sought to investigate the potential pathologi-
tumor suppressor role of FoxA1 in HCC. cal relevance of IKBKE in HCC. Bioinformatic analyses revealed that
However, compared to the frequent genetic alterations of FoxA1 in IKBKE exhibited approximately 4 to 8% amplification in the HCC set-
the prostate cancer setting, where FoxA1 acts as an oncogene with tings, and its expression was strongly associated with poor survival in
gain-­of-­function mutations (more than 4%) (21), alterations of FoxA1 both alcohol-­and hepatitis virus–associated HCC (mainly in the male
are rarely observed in patients with HCC. In contrast, bioinformatic patient cohort) (fig. S3, A to C). Furthermore, immunoblot analysis
analysis from The Cancer Genome Atlas dataset revealed that high implied that IKBKE was highly expressed in HCC tissues compared
mRNA levels of FoxA1 indicated a poor prognosis in patients with with adjacent normal tissues (fig. S1K). Immunohistochemistry
HCC, whereas high expression of FoxA2 or FoxA3 showed better out- (IHC) staining further revealed that IKBKE expression correlated
comes (fig. S1I). To this end, we sought to detect FoxA1 protein levels with the early stage of HCC to predict poor patient outcomes (fig. S3,
in HCC tissues. Notably increased FoxA1 protein was observed in D to F), suggesting that IKBKE should be an early event for
HCC tissues compared with paired adjacent normal tissues (fig. S1, J HCC. Similar to the oncogenic role of IKBKE in breast and other can-
and K), suggesting that the mRNA and protein levels of FoxA1 did cers (33), depletion of IKBKE significantly reduced HCC cell growth,
not well associated with its tumor suppressor function, suggesting colony formation (Fig. 2, A and B, and fig. S4A), tumor growth (Fig. 2,
that posttranslational modifications (PTMs) may play a key role in C to E, and fig. S4, B and C), and HCC organoid formation (Fig. 2, F
regulating the tumor suppressor of FoxA1. and G), and impaired NF-­κB and phosphatidylinositol 3-­kinase
(PI3K)–AKT pathways (fig. S4C). To explore the potential role of IK-
IKBKE represses FoxA1 transcriptional activation BKE in vivo, the chemical carcinogen DEN was used to generate a
Although several PTMs of FoxA1, including acetylation, methyla- murine model of hepatocarcinogenesis. The results showed that Ikbke
tion, and ubiquitination, have been reported in prostate cancer knockout (Ikbke−/−) mice were resistant to DEN-­induced liver in-
(28–30), the phosphorylation modification has not been identified flammation (with the marker F4/80), proliferation (with the marker
to affect FoxA1 function. To this end, we established a CRISPR-­ Ki67), and tumor formation (with the liver cancer marker CK19 and
Cas9–based kinome screen method (Fig. 1G). For that, a reporter normal liver tissue marker HNF4α) compared with counterpart mice
construct containing FoxA1-­response elements (FRE) was de- (Fig. 2, H and I, and fig. S4D), suggesting a protective effect of IKBKE
signed, which also exhibited unexpected cross-­reactivity to fork- on hepatocytes. Furthermore, knockdown of FoxA1 partially rescued
head BOX O3 (FoxO3a), but not forkhead BOX D4 (FoxD4) and IKBKE depletion-­induced cell growth inhibition (Fig. 2, J and K, and
forkhead BOX P1 (FoxP1) (fig. S2A). As a result, ectopic expres- fig. S4, E and F), accompanied with increased IKBKE substrates such
sion of both FoxA1 and FoxO3a could notably increase the expres- as pS473-­AKT and phosphor–NF-­κB inhibitor alpha (p-­Iκ-­Bα), sug-
sion of green fluorescent protein (GFP) reporter gene (fig. S2B). gesting that FoxA1 likely defects the oncogenic role of IKBKE in HCC.
Then, the reporter construct cells were infected with CRISPR ki-
nase pools (Fig. 1G), and the resulting cells were subjected to the IKBKE directly phosphorylates FoxA1
kinome screen analysis by flow cytometry–based sorting and deep To investigate the underlying mechanism of IKBKE in impairing
sequencing methods (31) to identify kinases that regulate FoxA1 FoxA1 transcriptional activity, we observed the stable interaction of
transcriptional activity (Fig. 1H and fig. S2, C and D). IKBKE and FoxA1 at both exogenous and endogenous levels (Fig. 3A
Among the screening results, we focused on potential kinases that and fig. S4, G and H). Furthermore, the Forkhead (FH) domain of
negatively regulate FoxA1 transcriptional activity, which may be fur- FoxA1, the major part of FoxA1 binding to DNA, was narrowed to
ther used as therapeutic targets. As a result, polo like kinase 3 (PLK3), specifically interact with IKBKE (Fig. 3B). While among FoxA family
AKT serine/threonine kinase 1 (AKT1), SIK family kinase 3 (SIK3), proteins, FoxA2 displayed a relatively lower binding affinity to IKBKE
NIMA-­related kinase 10 (NEK10), IKBKE, 3-­phosphoinositide– (fig. S4H). To disclose the potential role of IKBKE in regulating
dependent protein kinase 1 (PDK1), and mitogen-­activated protein FoxA1, we hypothesized that IKBKE could directly phosphorylate
kinase 6 (MAPK6) are ranked in the top of repressing FoxA1 tran- FoxA1 to inhibit its activity. To this end, a band shift of FoxA1, a com-
scriptional activity (Fig. 1H). Furthermore, ectopic expression of mon indicator of protein phosphorylation, was observed upon en-
these kinases was performed, among them, IKBKE, to a lesser extent forcing IKBKE expression, which could be reduced by addition of
of AKT1, was shown to readily reduce FoxA1 reporter activity (fig. S2E). lambda phosphatase (λ-­phosphatase) (fig. S5A). Similar results were
To validate the function of IKBKE in regulating FoxA1, ectopic ex- observed by using a phospho-­tag approach, where Myr-­IKBKE en-
pression of wild type (WT) or constitutively active IKBKE (Myr-­ hanced FoxA1 phosphorylation (fig. S5B). Furthermore, in vitro ki-
IKBKE) (32), but not TBK1 or kinase dead (DN) IKBKE-­K23M, nase assays revealed that IKBKE directly phosphorylated FoxA1
significantly decreased FoxAs (FoxA1-­3) transcription activity (Fig. 1I (fig. S5C).
and fig. S2, F and G). Furthermore, through transcriptomic analyses, To determine the phosphorylation residues of FoxA1 by IKBKE,
IKBKE depletion-­induced alterations of gene expression markedly we performed mass spectrometry analysis and observed that the S174
overlapped with those derived from FoxA1 knockdown cells (Fig. 1, J and S177 sites, located in the FH domain of FoxA1, were possibly

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 2 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B C

D E F

H I J K

L M N

Fig. 1. IKBKE negatively regulates FoxA1 activity and tumor suppressor function. (A) Immunoblotting (IB) analysis of whole-­cell lysate (WCL) derived from Huh7 and
Jhh7 cells infected with short hair RNAs (shRNAs) targeting FoxA1. GAPDH, glyceraldehyde-­3-­phosphate dehydrogenase. (B) Huh7 cells generated in (A) were subjected
to colony formation and soft agar assays. The relative colony numbers were quantified. Means ± SD, n = 3, Student’s t test. *P < 0.05 and **P < 0.01. (C and D) Huh7 cells
generated in (A) were subjected to xenograft mouse assays. Tumor sizes were monitored (C), and dissected tumors were weighed and calculated (D). Means ± SD, n = 6,
two-­way analysis of variance (ANOVA) test for (C) and Student’s t test for (D). *P < 0.05. (E) Micrograph shows representative shFoxA1 HCC organoids. Scale bars, 100 μm.
(F) Box-­whisker plot show relative size (diameter) of HCC tumoroids (E). Means ± SD, n = 10, Student’s t test. **P < 0.01. (G) Workflow of kinome screening using CRISPR-­
Cas9 system. (H) Scatter plot shows the top candidate single guide RNA (sgRNA) genes of high-­throughput sequencing for (G). (I) FoxA1 luciferase reporter assays were
performed, and the results normalized to level of EV (empty vector) activity. Means ± SD, n = 3, Student’s t test. **P < 0.01. IB analysis of WCL derived from 293T cells
transfected with indicated constructs (top). (J) Venn diagram indicates the numbers of differentially expressed genes (DEGs) between shFoxA1 and shIKBKE in Huh7 cells.
(K) Kyoto Encyclopedia of Genes and Genomes (KEGG) Pathway enrichment results of top enriched pathways in crossed genes (J). (L and M) Quantitative polymerase
chain reaction (qPCR) validation of crossed gene from (J) in Huh7 cell lines. (N) Schematic model indicates that IKBKE inactivates FoxA1 transcriptional activity.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 3 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B C D

E F G

H I J

Fig. 2. IKBKE promotes liver tumorigenesis. (A) iB analysis of Wcl derived from huh7 cells infected with shRnAs targeting IKBKE. (B) cells generated in (A) were sub-
jected to colony formation and soft agar assays. the relative colony numbers were quantified. Means ± Sd, n = 3, Student’s t test. *P < 0.05 and **P < 0.01. (C to E) cells
generated in (A) were subjected to xenograft mouse assays. tumor sizes were monitored (c), and dissected tumors (d) were weighed and calculated (e). Means ± Sd, n = 6,
two-way ANOVA test for (c) and Student’s t test for (d). *P < 0.05 and **P < 0.01. (F) Micrograph shows representative shiKBKe hcc organoids. Scale bars, 100 μm.
(G) Box-whisker plot show relative size (diameter) of hcc organoids (F). Means ± Sd, n = 10, Student’s t test. **P < 0.01. (H) Macroscopic images represent the hematoxylin
and eosin (h&e), AFP, Ki67, cK19, and hnF4a staining of Wt and Ikbke−/− mouse livers treated with or without den, and the immunofluorescence (iF) staining of F4/80 was
also represented. Scale bars, 100 μm. (I) liver/body ratio was quantified (means ± Sd, n = 6 or 7, Student’s t test) and the F4/80 positive cells were quantified (means ± Sd,
n = 8, Student’s t test). **P < 0.01 and ****P < 0.0001; # means no significance. (J and K) huh7 cells were infected with shRnAs against FoxA1 and selected with puromycin
(1 μg/ml) for 5 days. After that, the resulting huh7 cells were infected with IKBKE small interfering RnA and subjected to iB analysis (J) and colony formation assays [(K)
bottom], and the relative colony numbers were quantified (K, top). Means ± Sd, n = 3, Student’s t test. *P < 0.05 and **P < 0.01.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 4 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B C D

E F G H

I J K L

M N O

Fig. 3. IKBKE directly phosphorylates FoxA1 to repress its transcriptional activity. (A) IB analysis of WCL and immunoprecipitation (IP) products derived from Huh7
cells. Immunoglobulin G (IgG) was used as negative control. (B) Functional domain of full-­length hFoxA1 and its different truncated constructions are represented and
subjected for IB analysis. (C) A schematic presentation of the evolutionarily conserved putative IKBKE phosphor-­motif in hFoxA1. (D) Phospho-­tag SDS–polyacrylamide
gel electrophoresis (SDS-­PAGE) IB analysis of WCL derived from 293T cells transfected with indicated constructs. (E to G) IB analysis of WCL and IP products derived from
Huh7 cells infected with indicated shRNAs (E) or from Ikbke−/− and counterpart mice (F), as well as patients with HCC (G). Scr, sh-­scramble RNA. (H) C57/BL6 male mice were
gavage-­administrated by DEN (100 mg/kg) for 48 hours with equal volume of 0.9% saline as control. The mouse livers were harvested for IB analysis. (I) LPS (10 mg/kg) was
intraperitoneally injected into C57/BL6 male mice for 6 hours with equal volume of PBS as the control. The mouse livers were harvested for IB analysis. (J) Nonalcoholic
steatohepatitis (NASH) mouse model was established by feeding C57/BL6 male mice with high fat methionine-­choline deficient diet (HFMCD) (A06071301, Research Di-
ets) for 6 weeks. The mouse livers were harvested for IB analysis. (K and M) Luciferase reporter assays were performed in 293T cells transfected indicated constructs (as IB
analysis of top), and the reporter results were normalized to the level of EV activity. Means ± SD, n = 3, Student’s t test. **P < 0.01 and ***P < 0.001. (L) Electrophoretic
mobility shift assay (EMSA) was performed with cell lysates from 293T cells transfected with indicated constructs. EV, as a negative control. (N) Volcano plot represents the
DEGs in Huh7 shFoxA1 cells stabilized with WT-­and S177D-­FoxA1. (O) KEGG pathway enrichment results of top enriched pathways were obtained from Huh7 shFoxA1
cells stabilized with WT-­and S177D-­FoxA1. mTOR, mammalian target of rapamycin; ECM, extracellular matrix.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 5 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
phosphorylated (Fig. 3C and fig. S5, D and E). Notably, both S174 and
S177 are evolutionally conserved in the FoxA subfamily and different
species, as well as in some members of the Fox family (Fig. 3C). S174
and S177 also share a conserved IKBKE phosphorylation motif de-
fined in previously established substrates such as cylindromatosis
(CYLD) (Fig. 3C) (16, 34). To validate these phosphorylation events,
we performed phospho-­tag assays and observed that both S174A and
S177A mutant markedly reduced IKBKE-­mediated band shifts in
FoxA1 (Fig. 3D and fig. S5F). Subsequently, we developed specific an-
tibodies against phosphorylated FoxA1-­S174 and S177, respectively
(fig. S5G), and observed that WT or Myr-­IKBKE but not DN-­IKBKE
promoted FoxA1 phosphorylation both under physiological condi-
tions (fig. S5, H to K) and in vitro kinase assay (fig. S5L). However,
FoxA2 displayed a relatively low phosphorylation (fig. S5M), possibly
due to its lower binding affinity to IKBKE (fig. S4H). IKBKE-­mediated
phosphorylation of FoxA1 was detected at endogenous levels both in
HCC cells, human HCC specimens, and murine liver tissues treated
by DEN, LPS, or high-­fat diet (Fig. 3, E to J), suggesting that FoxA1 is
a bona fide substrate of IKBKE.
IKBKE represses FoxA1 transcriptional activity in a
phosphorylation-­dependent manner
To investigate the effect of IKBKE-­mediated phosphorylation on
FoxA1 transcriptional activity, the nonphosphorylated mimetic
(S174A and S177A) and phosphorylated mimetic (S174D and S177D)
FoxA1 mutants were generated, in contrast to ectopic expression of
IKBKE or S177D-­FoxA1, S177A mutant FoxA1 resulted in enhanced
FoxA1 binding to the corresponding DNA elements and elevated its
transcriptional activity (Fig. 3, K and L, and fig. S6A). Notably, S177A
mutation-­induced FoxA1 transcriptional activity could only be lesser
reduced compared with WT-­FoxA1 upon ectopic expression of IK-
BKE (Fig. 3M).
Next, compared with WT, the S177D mutant transcriptionally al-
tered the downstream target genes of FoxA1 (Fig. 3, N and O, and
fig. S6, B and C). Unexpectedly, both S174A and S174D exhibited
similar phenotypes to impair FoxA1 binding to DNA elements and
inhibit FoxA1 transcriptional activity (Fig. 3, K and L, and fig. S6A).
To point out this distinction, we analyzed the localization of S174 and
S177 sites in FoxA1, and observed that S174, but not S177, resides at a
critical position for FoxA1 binding to DNA (fig. S6D). Therefore, re-
placement of S174 with either Ala (S174A) or Asp (S174D) might al-
ter the conformation of FoxA1, thereby interfering its DNA binding
properties. Therefore, although S174 also undergoes IKBKE-­mediated
phosphorylation (Fig. 3C), we could not use S174A or S174D muta-
tion to mimic (non)-­phosphorylated S174, as usually done in the
phosphorylation study. Therefore, in this study, we mainly focused on
IKBKE-­mediated phosphorylation of FoxA1 at S177 to inhibit its
function.
IKBKE dissembles FoxA1 complex formation
Consistent with the previous finding that FoxA1 formed a dimer and
integrated with cofactors into a large complex to carry out its tran-
scriptional function (35), we observed that FoxA1 formed homodi-
mer with its FH domain (fig. S6E). Ectopic expression of IKBKE or
S177D-­FoxA1 mutant, but not DN-­IKBKE or S177A-­FoxA1 mutant,
markedly blocked FoxA1 dimerization (fig. S6, F and G). Further-
more, both IKBKE expression and the S177D mutant reduced FoxA1
complex formation, whereas the S177A mutant enhanced its complex
assembly (Fig. 4, A and B). Specifically, as a physiological partner,
Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024
GATA binging protein 4 (GATA4) interacts with FoxA1 and plays a
key role in its tumor suppressor function in HCC (24). Here, we ob-
served that forced expression of IKBKE and the S177D mutant
blocked, whereas overexpression of the S177A mutant enhanced
GATA4 binding to FoxA1 (Fig. 4, C to E, and fig. S6H). These findings
suggest that IKBKE phosphorylates FoxA1 to reduce its transcrip-
tional complex formation (Fig. 4F).
Biologically, S177D enhanced colony formation in cells (Fig. 4, G
and H, and fig. S7, A and B) and HCC organoid models compared to
WT and S177A-­FoxA1 (Fig. 4I). To study the role of phosphorylated
S177 in vivo, S177A and S177D-­Foxa1 knock-­in (designated Foxa1S177A
and Foxa1S177D) and knockout (designated Foxa1−/−) mice were gen
erated (fig. S7, C to E). Among them, both Foxa1S177D and Foxa1−/− dis-
played potential birth defects compared with intact and corresponding
heterogeneous mice, while Foxa1S177A had no obvious effect on mouse
birth and growth (fig. S7F). To explored whether the nonphospho-
mimic FoxA1 (S177A) acts as a tumor suppressor, a hydrodynamic
HCC mouse model was used, and the results showed that Foxa1S177A
greatly reduced Myr-­Akt1;N-­Ras–induced HCC (Fig. 5A), coupled
with reducing fibroblast formation (indicated by Masson), hepatotox-
icity [alpha fetoprotein positive (AFP+)], and inflammation (F4/80+)
(Fig. 5, B and C). This result indicates that non-­IKBKE phosphory-
lated FoxA1 (S177A) possibly enhances hepatic tissue differentiation
to antagonize tumorigenesis, consistent with our hypothesis that
S177A enhances FoxA1 binding to GATA4 to promote hepatic dif-
ferentiation, as previously reported (24).
However, because of the developmental defects of Foxa1S177D mice
and technique limitation in generating liver-­conditional Foxa1S177D
knock-­in mice, it was not possible to assess whether Foxa1S177D mice
are prone to liver tumorigenesis. Alternatively, we harvested 5-­to
12-­day-­old Foxa1S177D and counterpart WT (Foxa1WT) mice and
observed that Foxa1S177D mice phenocopied Foxa1−/− mice exhibiting
marked growth retardation of body and organs (fig. S7, G and H).
Consistent with the previous finding that Gata4 deficiency leads to
hypoplastic liver in mice and thus affects liver development (24),
Foxa1S177D mice also largely reduced mouse liver development and
suppressed it by enhancing liver injury (AFP+) and inflammation
(F4/80+) (Fig. 5, D to F), accompanied by impaired hepatocyte cycle
and inflammation (Fig. 5, G and H, and fig. S7, I to L). Together,
these findings demonstrate that IKBKE inhibits FoxA1 transcrip-
tional activity and tumor suppression in a phosphorylation-­
dependent manner.
IKBKE inhibitor alleviates liver tumorigenesis
Although only 4 to 8% of IKBKE amplifications were identified in
HCC by IHC staining, we identified IKBKE high expression in at
least 60% patients with HCC (figs. S1K and S3, D and E). Because
HCC is often present under inflammatory conditions, we also ob-
served that DEN, a typical carcinogen that induces HCC, could en-
hance the expression of Ikbke in mouse liver tissues (fig. S7M), which
was consistent with the above finding that Ikbke−/− reduced DEN-­
induced liver injury and inflammation compared with intact mice
(Fig. 2, H and I, and fig. S4D). Because IKBKE has been identified as
highly expressed and plays an important pathological role in HCC,
we tended to verify the efficacy of IKBKE inhibitors as potential ther-
apies for HCC. To this end, the IKBKE inhibitor compound 1
(COMPD1) was used (36), which substantially inhibited the IKBKE-­
induced band shift and phosphorylation of FoxA1 (fig. S8, A to C).
Furthermore, this inhibitor also markedly rescued the decreased
6 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B

C D E F

G H

Fig. 4. IKBKE phosphorylates FoxA1 to disassemble its complex formation. (A and B) IB analysis of gel filtration products derived from 293T cells transfected with
indicted constructions (A). MW, molecular weight. The quantitative trendline results of protein expression levels were quantified by Image J software (B). (C and E) IB
analysis of WCL and GST–pull-­down products derived from 293T cells transfected with indicated constructs. (D) IB analysis of WCL and IP products derived from Huh7 cells
infected with indicated shRNAs. Scr, sh-­scramble RNA, as a negative control. The GATA4 co-­IP levels were quantified. Means ± SD, n = 3, Student’s t test. *P < 0.05.
(F) Schematic model indicates that IKBKE could directly phosphorylate FoxA1 to repress its binding with GATA4. (G and H) Huh7 cells were infected with shRNAs against
FoxA1 and selected with puromycin for 5 days, and the resulting cells were infected with virus encoding different mutants of FoxA1 and selected with hygromycin for
5 days. Then, the cells were subjected to IB analysis (G), colony formation, and soft agar assays (H). The relative colony numbers were quantified. Means ± SD, n = 3, Student’s
t test. *P < 0.05 and **P < 0.01. (I) FoxA1-­deleted (shFoxA1) HCC organoids were infected with virus encoding different mutants of FoxA1 and selected with hygromycin
for 5 days. Micrograph shows the resulting HCC organoids. Scale bars, 100 μm (I). Box-­whisker plot show relative size (diameter) of HCC tumoroids. Means ± SD, n = 10,
Student’s t test. *P < 0.05 and **P < 0.01.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 7 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B C

D E F

G H

Fig. 5. Phosphorylation of FoxA1 represses its tumor suppressor roles. (A to C) Macroscopic liver images of WT and S177A-­KI mice with indicated treatment. Control
group mice were tail vein injected with empty vector in 2 ml of saline solution, while Myr-­mAKT1;N-­Ras-­V12;transposase constructs were injected in test groups (A). Also,
the surface tumors were calculated (A). Means ± SD, Student’s t test. *P < 0.05. Macroscopic liver images represented the H&E and Masson staining of WT and Foxa1S177A/
S177A
mice (scale bar, 2 mm) and the AFP and F4/80 IHC staining results of WT and Foxa1S177A/S177A mice (scale bars, 2 mm and 100 μm) (B). The F4/80-­and AFP-­positive cells
were quantified (C). Means ± SD, n = 8 for F4/80 and n = 3 for AFP, Student’s t test. *P < 0.05 and ***P < 0.001. (D to F) Macroscopic liver images represent the H&E stain-
ing of Foxa1+/+(WT), Foxa1S177D/+(Het), Foxa1S177D (Homo), Foxa1+/− and Foxa1−/− mice, and the AFP and F4/80 IHC staining (D and E). Scale bars, 100 μm. The F4/80-­and
AFP-­positive cells were quantified (F). Means ± SD, n = 8 for F4/80 and n = 3 for AFP, Student’s t test. *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001. (G) Gene Set
Enrichment Analysis of DGEs in (fig. S7I). The genes sorted along the horizontal axis corresponded to the running enrichment score at vertical axis. In the vertical axis, a
positive value indicates enrichment at the top, the left genes of the peak genes are core genes, and the negative values indicate the opposite. In the horizontal axis, the
red and blue heatmap means the genes expression abundance arrangement. The dark red means a larger log fold change (logFC), and dark blue means a smaller logFC.
(H) qPCR validation of crossed gene from (Fig. 1J) in Ikbke and Foxa1 knockout mouse liver tissues, respectively. Means ± SD, n = 3, Student’s t test. *P < 0.05, **P < 0.01,
and ***P < 0.001. NES, normalized enrichment score.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 8 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

FoxA1 transcriptional activity by IKBKE (Fig. 6A). In addition, FoxA3 also undergo IKBKE-­mediated phosphorylation, but with
COMPD1 not only restored the interaction of FoxA1 with its partner different trends. Due to the redundant and identical roles and tissue-­
GATA4 (Fig. 6, B and C), but also rescued its protein complex forma- context distribution of FoxA1-­3, whether IKBKE can regulate cellu-
tion (fig. S8D). Functionally, COMPD1 potently suppressed HCC lar biology and pathological functions by phosphorylating FoxA2/3,
growth both in cells and in vivo (Fig. 6, D to G, and fig. S8, E to H), especially in liver tumorigenesis, deserves further study. Accumulat-
while enhancing apoptosis (Fig. 6, F and G). COMPD1 decreased ing substrates of IKBKE have also recently been identified to affect
intact but with a lesser extent of FoxA1-­depleted, liver cancer cell inflammation and immune pathways. Although we have preliminar-
growth (Fig. 6, D to G, and fig. S8F), suggesting that IKBKE partially ily elucidated FoxA1 as the major downstream effector of IKBKE in
regulates HCC by inactivating FoxA1. To further evaluate the poten- HCC, more in vivo studies are needed by using Foxa1S177A knock-­in
tial antitumor role of COMPD1 in liver tumorigenesis in vivo, a C-­ mice, including DEN-­induced HCC models and crossing with C-­
Myc–driven HCC mouse model was used (37). After COMPD1 Myc–driven HCC mouse models.
administration (Fig. 6H), we observed a significant attenuation of As an inducible inflammation kinase, IKBKE is thought to be in-
C-­Myc–induced HCC with reduced FoxA1 phosphorylation com- duced by diverse physiological or pathological stimuli, including
pared with the vehicle-­treated group (Fig. 6, I to J), along with de- PMA, TNFα, and LPS. These cytokines or chemokines have been
creased liver injury (AFP+), inflammation (F4/80+), proliferation shown to play pivotal roles in liver tumorigenesis (40). Because HCC
(Ki67+), and increased apoptosis (cleaved Caspase 3) (Fig. 6, K and is mainly induced by hepatis virus and metabolic mediators, we hy-
L, and fig. S8I). These observations together suggest that targeting pothesize that IKBKE might be induced/activated under these differ-
IKBKE will effectively inhibit liver tumorigenesis, in part, by un- ent conditions by inhibiting the tumor suppressor FoxA1, thereby
leashing FoxA1 tumor suppressor function (Fig. 7). promoting liver tumorigenesis. We observe that DEN challenge can
also enhance IKBKE expression in mice, accompanied with FoxA1
phosphorylation. LPS administration and high-­fat diet feeding could
DISCUSSION also promote NF-­κB activation and IKBKE-­mediated FoxA1 phos-
Frequent mutations of FoxA1 are strongly associated with breast and phorylation (Fig. 3, I and J), suggesting the potential function of
prostate tumorigenesis (20, 21, 38); however, genetic alterations of IKBKE-­FoxA1 axis in promoting HCC. However, although we vali-
FoxA1 rarely occur in other tumors including liver cancer, suggesting dated that DEN-­induced HCC is IKBKE dependent, whether LPS-­
that FoxA1 may play distinct roles in these cancer settings. Initially, and high-­fat–induced liver damage and HCC are IKBKE activity
FoxA1 has been identified in rat liver tissues and plays an important dependent that, particularly via FoxA1 phosphorylation, are worth
role in hepatocyte differentiation and maturation (35, 39). For that, a further investigations. Meanwhile, although viral infections have
liver conditional Foxa1/2 double knockout mouse models were gen- been disclosed to activate IKBKE (41), because HBV/HCV could not
erated and displayed a sex-­dependent distinct function in DEN-­ infect mouse hepatocyte, we only postulate that this virus-­induced
induced HCC (22). Mutations in Gata4, a binding cofactor of FoxA1, IKBKE activation may result in FoxA1 inactivation and liver tumor-
could enhance liver tumorigenesis, and the FoxA1-­lysine demethyl- igenesis. Together, our findings suggest that under conditions of
ase 6A (KDM6A)–Rho GDP dissociation inhibitor beta (ARHGDIB) IKBKE amplification/activation genetically or stimulated with in-
axis can also suppress the malignancy of bladder cancer (23, 24); thus, flammation/carcinogens, FoxA1 undergoes IKBKE-­mediated phos-
we hypothesized that FoxA1 might play a tumor suppressor role in phorylation, which abolishes its transcriptional activity and tumor
HCC. Here, we disclosed FoxA1 as a tumor suppressor in HCC, al- suppressor function, thus leading to HCC. We also highlight the strat-
though high FoxA1 expression was observed in HCC tissues and as- egy to combat HCC with IKBKE-­specific inhibitors.
sociated with poor survival that are consistent with previous studies
(26, 27). The different findings of FoxA1 in prostate/breast cancer ver-
sus liver cancer may be due to the fact that FoxA1 regulates distinct MATERIALS AND METHODS
downstream genes by cooperating with different cofactors. Thus, al- Cell culture, transfection, fractionations, and viral infection
though genes in the transcriptional regulation of FoxA1 between ER+ Huh7, Jhh7, PLC/PRF5, HepG2, SK-­HEP-­1, SNU739, and 293T
breast cancer and AR+ prostate cancer overlapped, the differences in cells were obtained from American Type Culture Collection and
the overlap between FoxA1 as a tumor suppressor and oncogene need maintained in Dulbecco’s modified Eagle’s medium (DMEM); Li7,
to be further explored. SNU761, and SNU387 cells were maintained in RPMI 1640 medium
IKBKE has been recognized as an oncogene in a variety of tu- containing 10% fetal bovine serum (Gibco), penicillin (100 U/ml),
mors, including breast and lung cancers. Besides activating the NF-­ and streptomycin (100 μg/ml; Gibco). Cells transfected with Lipo-
κB pathway, IKBKE could also directly phosphorylate AKT and fectamine (Invitrogen) or Polyethylenimine [(PEI) Polysciences]
IRF3, thus, we commonly used phosphorylation of Iκ-­Bα with p-­ were performed according to the manufacturer’s instructions. Cell
IRF3 or p-­AKT to reflect the activation of IKBKE. Here, we find that fractionations were performed with the Cell Fractionation Kit
IKBKE inhibits FoxA1 transcriptional activity in a phosphorylation-­ (CST9038). IKBKE inhibitors COMPD1 (C1) was obtained from
dependent manner and exerts oncogenic roles in HCC, possibly Selleck (S8922).
through unleashing FoxA1-­mediated negative regulation of cell cy- For generating knockout/knockdown or reconstituted express-
cle–associated genes (such as BUB, CCND1, and MET). At the same ing cells, human embryonic kidney–293 T cells were used for pack-
time, a question is raised, how does IKBKE regulate FoxA1 to per- aging of lentiviral viruses. After transfection 48 and 72 hours,
form oncogenic functions once FoxA1 acts as an oncogene in breast/ supernatant was harvested and filtered through a 0.45-­μm syringe
prostate cancer? We postulate that IKBKE may regulate FoxA1 tran- filter and used to infect cells in the presence of polybrene (4 μg/ml).
scriptional profiles by selectively altering its interaction with cofac- Infected cells were selected using hygromycin B (200 μg/ml) or pu-
tors. Because of bearing similar phosphorylation motifs, FoxA2 and romycin (1 μg/ml) for 5 days.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 9 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

A B C D

F G

H I J

K L

Fig. 6. IKBKE inhibitor represses tumor growth partially by recovering FoxA1 activity. (A) Reporter assays were performed in 293T cells transfected indicated con-
structs treated with/without COMPD1 (5 μM for 4 hours), and the results were normalized to EV activity. Means ± SD, n = 3, Student’s t test. *P < 0.05 and **P < 0.01. (B
and C) IB analysis of WCL and GST–pull-­down products from 293T cells transfected indicated constructs and treated with/without COMPD1 (5 μM for 4 hours). HA-­FoxA1
co-­IP levels were quantified. Means ± SD, n = 3, Student’s t test. *P < 0.05 and **P < 0.01. (D to G) Huh7 cells were infected with shRNAs against FoxA1, selected with
puromycin for 5 days, and then subjected to xenograft assays treated with/without COMPD1 [intragastrical administration (IG) with 30 mg/kg per 2 days]. Tumors were
dissected (D) and weighed (E). Data were shown as means ± SD, n = 6, Student’s t test. *P < 0.05. Ki67 staining and cleaved Caspase 3 staining were quantified [(F) and (G)]
(means ± SD, n = 6, Student’s t test). *P < 0.05 and ***P < 0.001. (H) Illustration of the flow for treating Cebpb-­Tta-­TetO-­Myc mice with IKBKE inhibitor. Briefly, male mice
were terminated with doxycycline water (100 μg/ml) at age 4 weeks and then administrated with COMPD1 (I.G. 30 mg/kg) or vehicle every 2 days for 6 weeks. (I to L)
Macroscopic liver images of C-­Myc mice with/without COMPD1 treatment. The surface tumors were calculated (I). Means ± SD, n = 4, Student’s t test. *P < 0.05. The mouse
livers were harvested for IB analysis (J). The H&E, AFP, Ki67, F4/80, and cleaved Caspase 3 staining were represented (K). The F4/80 and cleaved Caspase3–positive cells
were quantified (L). Means ± SD, n = 6, Student’s t test. **P < 0.01 and ***P < 0.001. Scale bars, 2 mm for H&E and IHC and 100 μm for IF.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 10 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Illustration of the IKBKE-­FoxA1 axis in liver tumorigenesis. Schematic model indicates that under physiological conditions, FoxA1 forms complex with GATA4
to facilitate downstream differentiation-­associated transcriptional regulation in repressing tumorigenesis; on the other hand, elevated IKBKE induced by virus infection or
DEN administration could phosphorylate FoxA1 to reduce its complex formation, leading to inactivation of FoxA1 and promotion of tumorigenesis, highlighting the po-
tential role of IKBKE inhibitor for liver cancer therapies. The figure was generated with Biorender.com.

Antibodies from Santa Cruz Biotechnology. Polyclonal anti-­ Flag antibody

All antibodies were diluted in Tris Buffered Saline with Tween-­20
(TBST) buffer with 5% nonfat milk for Western blot. Anti-­IKBKE
antibody (1:3000; 3416 s), anti-­FoxA1 antibody (1:3000; 53528 s),
anti–TNF receptor–associated factor 2 antibody (1:3000; 4724 s),
anti-­ubiquitin antibody (1:2000; 3933 s), anti–p–NF-­κB antibody
(1:3000; 3033 s), anti–C-­MYC antibody (1:3000; 18,583 s), anti–extra-
cellular signal–regulated kinase 1/2 (ERK1/2) antibody (1:3000; 4695
s), anti-­ERK1/2-­p antibody (1:3000; 4370 s), anti–Iκ-­Bα antibody
(1:3000; 4812 s), anti–p–Iκ-­Bα antibody (1:3000; 2859 s), anti-­IRF3
antibody (1:3000; 4302 s), anti–p–IRF-­3 antibody (1:3000; 37829 s),
anti–cleaved Caspase 3 antibody (1:50; 9664 s), anti-­AKT1 antibody
(1:3000; 75692 s), anti-­pSer473-­AKT antibody (1:3000; 4060), anti–
Myc-­tag antibody (1:3000; 2276 s), and anti–glutathione S-­transferase
(GST)–tag antibody (1:3000; 2625 s) were obtained from Cell Signal-
ing Technology (CST). Anti-­ p65 antibody (1:2000; 10745-­ 1-­
AP),
anti–GST-­ tag antibody (1:2000; 66001-­ 2-­
lgl), anti-­
AFP antibody
(1:100 14550-­1-­AP), F4/80 antibody (1:1000; 28463-­1-­AP) were ob-
tained from Proteintech. Anti–thiophosphate ester antibody (1:3000;
ab92570) were obtained from Abcam. Polyclonal anti-­FoxA1 anti-
body (1:100; EA32402) was obtained from ELGBIO. Anti-­Ki67 anti-
body (1:100; GB111499) was obtained from Servicebio. Anti-­IKKE
antibody (1:200;VMA00517) was obtained from Bio-­Rad. Polyclonal
anti–hemagglutinin (HA) antibody (1:1000; sc-­805) was obtained
(1:1000; F-­2425), monoclonal anti-­Flag antibody (1:3000; F-­3165,
clone M2), anti-­tubulin antibody (1:3000; T-­5168), anti-­vinculin
antibody (1:3000; V4505), anti-­ HA agarose beads (A-­ 2095),
peroxidase-­conjugated anti-­mouse secondary antibody (A-­4416), and
peroxidase-­conjugated anti-­rabbit secondary antibody (A-­4914) were
obtained from Sigma-­Aldrich. Mouse anti-­rabbit immunoglobulin G
LCS (1:10,000; A25022) was obtained from Abbkine. Glutathione
Sepharose 4B (17-­0756-­05) was obtained from GE Healthcare, Mono-
clonal anti-­HA antibody (1:3000; 901503), anti–glyceraldehyde-­3-­
phosphate dehydrogenase (1:3000; 649201) antibody was obtained
from BioLegend.
The polyclonal phosphorylation antibodies against pS174 and
pS177-­FoxA1 (1:1000) generated by Abclonal Technology were de-
rived from rabbit. The antigen sequence used for immunization was
FoxA1 amino acids 170 to 177 (KPPYSYIS) for pS174 and amino
acids 174 to 180 (C-­SYISLIT) for pS177 (S stands for phosphorylated
residue in this synthetic peptide). The antibodies were affinity puri-
fied using the antigen peptide column and counter-­selected with un-
modified antigen.
Plasmid construction
To create pCMV-­GST-­FoxA1-­FL (GST-­FoxA1) and pCDNA3-­HA-­
FoxA1 (HA-­FoxA1), rat and human FoxA1 cDNA was cloned into the

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 11 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Bam H1 and SaI 1 sites of pCMV-­GST-­vector and pCDNA3-­HA-­ Purification of GST-­tagged proteins from bacteria
vector, respectively. CMV-­GST-­FoxA1-­N, CMV-­GST-­FoxA1-­FH, Recombinant GST-­conjugated FoxA1 and mutants were generated by
CMV-­GST-­FoxA1-­M, and CMV-­GST-­FoxA1-­C were cloned into the transforming the BL21 (DE3) Escherichia coli strain with pGEX-­
mammalian expression GST-­fusion vectors. HA-­FoxA1-­S177A and FoxA1-­20aa, pGEX-­FoxA1-­S177A-­20aa, pGEX-­FoxA1-­S174A-­20aa,
HA-­FoxA1-­S177D constructs were generated by site-­directed muta- respectively. The cultured bacteria were grown with vigorous shaking
genesis using overlap extension polymerase chain reaction (PCR). All at 37°C to an outer diameter of 0.8 and then for 12 to 16 hours at 16°C
primers and short hair RNA (shRNA)/single guide RNA (sgRNA) by adding 0.1 mM isopropyl-­β-­​d-­thiogalactopyranoside. Recombi-
sequences were listed in table S1. pLenti-­HA-­Hygro-­FoxA1, pLenti-­ nant proteins were purified from harvested pellets and resuspended
HA-­Hygro-­FoxA1-­S177A, and pLenti-­HA-­Hygro-­FoxA1-­S177D were in 10 ml of EBC buffer for sonication. The supernatant was incubated
cloned into the mammalian expression pLenti-­HA-­Hygro vectors. with 50 μl of 50% glutathione sepharose slurry (Pierce) for 3 hours
Various FoxA1, IKBKE mutants were generated by site-­directed mu- at 4°C, and the glutathione beads were washed three times with
tagenesis using overlap extension PCR. All mutants were verified by phosphate-­buffered saline (PBS) buffer and eluted by elution buffer.
DNA sequencing. Recovery and yield of the desired proteins (or complexes) was con-
firmed by analyzing 10 μl of beads by Coomassie blue staining and
Human primary HCC cancer organoid culture and HCC quantified with bovine serum albumin (BSA) standards.
organoid generation
Human primary HCC organoids (L231T) were generated and cul- Peptide synthesis and dot immunoblot assays
tured in three-­dimensional Matrigel using DMEM/F12 medium FoxA1-­WT, FoxA1-­pS177 and FoxA1-­pS174 peptides used for dot
with 1% GlutaMAX, 10 mM Hepes, 1:50 B27 supplement, 1:100 N2 blot assays were synthesized by Abclonal. The sequences were listed
supplement, 1.25 mM N-­acetyl-­​l-­cysteine, 10% (v/v) Rspo-­1–con- as: Foxa1-­W T: Bio-­KRSYPHAKPPYSYISLITMAIQQA, Foxa1-­
ditioned medium, 30% (v/v) Wnt3a-­conditioned medium, 10 μM pS177: Bio-­KRSYPHAKPPYSYI(pS)LITMAIQQA, and Foxa1-­
nicotinamide, 10 nM recombinant human (Leu15)–gastrin I, recombi- pS174: Bio-­KRSYPHAKPPY(pS)YISLITMAIQQA.
nant human EGF (50 ng/ml), recombinant human FGF10 (100 ng/ml), Peptides were diluted into 2 mg/ml for further biochemical as-
recombinant human hepatocyte growth factor [(HGF) 25 ng/ml], 10 μM says. For dot blot assays, peptides were diluted with PBS and spotted
forskolin, 5 μM A8301, Noggin (25 ng/ml) ,10 μM Y27632, and 1% onto nitrocellulose membrane with the amount of 0.01, 0.03, 0.1,
penicillin/streptomycin. To knockdown FoxA1 and IKBKE, organ- 0.3, and 1 μg. The membrane was dried and blocked by soaking in
oids were digested into single-­cell suspension and infected with len- TBST buffer with 5% nonfat milk for immunoblot analysis.
tiviral shRNAs for 6 hours at 37°C and harvested at centrifugation
(500g) followed by resuspension in Matrigel. Also, packaging of DNA synthesis and pulldown assays
cDNA expressing viruses were infected for sh-­FoxA1 organoids. Before harvest the whole cell lysates, the synthesized Bio-­5xFRHE
Constructed organoids were maintained in the presence of hygro- DNA fragments were incubated with streptomycin beads 1 hour at
mycin (200 μg/ml) or puromycin (1 μg/ml) depending on the vi- room temperature (RT). After 48-­hour transfection, the transfected
ral vectors. cells were lysed in EBC buffer [50 mM tris (pH 7.5), 120 mM NaCl,
and 0.5% NP-­40] supplemented with protease inhibitors (cOmplete
Immunoprecipitation and immunoblotting analyses Protease Inhibitor Cocktail, Roche) and phosphatase inhibitors
To immunoprecipitate ectopically expressed HA-­tagged and GST-­ (PhosSTOP, Roche). The whole-­cell lysates were added into the in-
tagged proteins, transfected cells were lysed 48 hours after transfection cubated streptomycin beads mixture and incubated at 4°C shaker
in EBC buffer [50 mM tris (pH 7.5), 120 mM NaCl, and 0.5% NP-­40] for 2 to 4 hours. After three washes with NETN buffer, the proteins
supplemented with protease inhibitors (cOmplete Mini, Roche) and samples were resolved by SDS-­PAGE.
phosphatase inhibitors (PhosSTOP, Roche). The whole-­cell lysates
were immunoprecipitated with distinct agarose beads for HA and GST In vitro kinase assay
at 4°C for 2 to 4 hours. After three washes with NaCl–EDTA–Tris–NP-­ For kinase assay using GST-­tagged proteins from cell lysate as sub-
40 (NETN) buffer [20 mM tris (pH 8.0), 150 mM NaCl, 1 mM EDTA, strate, the GST-­FoxA1, GST-­FoxA1-­S177A and GST-­FoxA1-­S174A
and 0.5% NP-­40], the bound proteins samples were resolved by SDS– proteins were pulled down by 30 μl Glutathione-­sepharose slurry
polyacrylamide gel electrophoresis (SDS-­PAGE). from 293 T cells. The obtained beads were respectively mixed with
10× kinase buffer (CST), 200 μM adenosine 5′-­triphosphate (ATP),
Kinome screening and 100 ng of adenosine 5′-­thiotriphosphate (Abcam) and the
A lentiviral reporter vector pLenti-­hygro-­CMV-­FoxA1-­luci-­EGFP ddH2O and the 70 ng of re-­IKBKE (Sigma-­Aldrich) were added to
was used to integrate into 293T cells, and stable clones were obtained reach a 30 μl of kinase reaction system. Kinase reactions were first
after selected with hygromycin. Next, the clones were infected with carried out at 30°C for 30 min and stopped by 1 μl of 3 mM EDTA
CRISPR-­Cas9 kinase lentivirus library then selected for 7 days with and next incubated at RT for 2 hours by adding 1.5 μl of 50 mM p-­
the treatment of puromycin to establish the reporter system. The re- Nitrobenzyl mesylate (PNBM) and lastly stopped by 10 μl of 5×
porter system cells were subjected to flow cytometry sorting to collect SDS buffer. After reaction, the targeted proteins were resolved by
the GFP intensity high and low cells, which further be the preparation SDS-­PAGE. The phosphorylated FoxA1 was detected by specific
of genomic DNA extraction. Also, the CRISPR library sgRNAs were antibodies.
amplified by PCR and aligned to the collected DNA using high-­ For kinase assay using purified recombinant GST-­conjugated pro-
throughput sequencing technology (Novogene). After alignment, the teins as substrate, the GST-­FoxA1-­20aa, the GST-­FoxA1-­S177A-­20aa,
differentially expressed sgRNAs were calculated and the correspond- and the GST-­FoxA1-­S174A-­20aa were added into a 30 μl of kinase
ing genes are ranked. reaction system. After reaction, the targeted proteins were resolved by

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 12 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

SDS-­PAGE, stained by Coomassie blue and detected by specific an-
tibodies.
Immunofluorescence staining
Cells were transfected with indicated plasmids and grown on glass
coverslips, fixed with 4% paraformaldehyde at RT for 15 min, washed
three times with PBS, and then permeabilized with 0.05% Triton X-­
100 at RT for 10 min. Following washed three times with PBS, the
coverslips were blocked with 5% BSA for 30 min and then incubated
overnight with indicated antibodies at 4°C. Following washing three
times with PBST, the coverslips were incubated with secondary anti-
bodies, anti-­mouse Alexa Fluor 643 (CST), for 1 hour at room tem-
perature in dark. Following washing three times with PBST, the
coverslips were stained with TRITC phalloidin at RT for 30 min and
with 4′,6-­diamidino-­2-­phenylindole in the dark and mounted using
glycerin. The mounted coverslips were subject to detection with Zeiss
LSM880 with Airyscan microscopy.
reporter assay system (Promega) according to the manufacturer’s in-
structions. All results are means and standard deviating from experi-
ments performed in biological triplicates (n = 3).
Electrophoretic mobility shift assay
Biotin end-­labeled probe was prepared by SBS Genetech (China).
Electrophoretic mobility shift assay (EMSA) was performed using the
LightShift Chemiluminescent EMSA Kit (Thermo Fisher Scientific,
20148) according to the manufacturer’s instructions. Briefly, DNA
binding reactions were performed in 20 μl of system containing
biotin-­labeled oligonucleotides and nuclear extracts. Additional un-
labeled oligonucleotides were added for competition. Reaction prod-
ucts were then separated by nondenaturing polyacrylamide gel
electrophoresis. Next, the protein-­DNA complexes were transferred
onto a positively charged nylon membrane (GE), and the nylon mem-
brane was cross-­linking with excitation light (120 mJ/cm2) for 1.5 min
and lastly detected by chemiluminescence.

Liver tissues and IHC staining Gel filtration chromatography analysis

The HCC tissue microarray (TMA) was obtained from the First Af- 293T cells were separately transfected with indicated plasmids, then
filiated Hospital of Sun Yat-­sen University. The study was approved by washed with PBS, and lysed in 1 ml of EBC buffer. Gel filtration mark-
the Cantonal Ethics Committee of Sun Yat-­sen University, and in- er kit (Sigma-­Aldrich) was used to determine the protein molecular
formed consent was obtained from each patient. We also combined weight corresponding to the number of tubes for further analysis. A
another HCC tumor tissue microarrays (Biomax, BC03117a) to ana- total of 1 ml of lysate was loaded onto a 200 increase 10/300 GL col-
lyze the clinical relevance of IKBKE in HCC. IHC was performed as umn after filtered through a 0.22-­μm syringe filter. Chromatography
previous reported including heat-­induced antigen retrieval proce- was performed on the AKTA avant 150 with EBC buffer. One column
dures (42). Incubation with antibodies against IKBKE (1:200), Ki67 volume of eluates was fractionated with 500 μl in each fraction at the
antibody (1:100), AFP antibody (1:100), cleaved Caspase 3 antibody elution speed of 0.8 ml/min. A total of 40-­μl aliquots of each fraction
(1:50), and F4/80 antibody (1:1000) were carried out at 4°C overnight. were loaded onto SDS-­PAGE and detected with indicated antibodies.
All stainings were blindly evaluated by two pathologists by quantita-
tive imaging methods; the percentage of immunostaining and the Real-­time reverse transcriptional PCR and RNA
staining intensity were recorded. An H score was calculated using the sequencing analyses
following formula: H score = Σ (PI × I) = (percentage of cells of weak Total RNA was extracted using a phenol-­chloroform method with
intensity × 1) + (percentage of cells of moderate intensity × 2) + (per-TRIzol regent (Takara, 9109), and cDNA was reverse-­transcribed us-
centage of cells of strong intensity × 3). PI indicates the percentage of
ing the PrimeScript reverse transcription kit (Takara 6210A) accord-
positive cells versus all cells, and I represent the staining intensity. The
ing to the manufacturer’s instructions. Real-­time PCR was performed
scores over 6 were identified as “high,” and the scores less than 6 wereusing the SYBR Green Premix Pro Taq HS qPCR Kit II (Rox Plus)
identified as “low.” In addition, the scores between 6 and 9 were iden- (AG11719). The primers used are listed in table S1.
tified as “middle.” High-­quality total RNA was used for the preparation of se-
quencing libraries using the oligo-­dT method and the BGISEQ-­500
Histopathologic analysis platform (BGI-­Shenzhen, China). The sequencing data were fil-
Hematoxylin and eosin staining was performed on paraffin-­ tered with SOAPnuke (v1.5.2) by removing reads containing se-
embedded tissues according to manuscript (Solarbio, catalog no. quencing adapter; removing reads whose low-­quality base ratio
G1120). Liver fibrosis was assessed via Masson staining (Solarbio, (base quality less than or equal to 5) is more than 20%; removing
catalog no. G1340). Histopathological images were captured under reads whose unknown base (“N” base) ratio is more than 5aaaaa5,
fully automatic digital pathology slide scanner (kfbio) and analyzed afterward, clean reads were obtained and stored in FASTQ format.
by QuPath software. Images were quantified using ImageJ and Graph- The clean reads were mapped to the reference genome using the
Pad Prism version 6.0 software. reference coding gene set, and then expression level of gene was
calculated by RSEM (v1.1.12). The heatmap was drawn by pheat-
Dual-­luciferase reporter assay map (v1.0.8) according to the gene expression in different samples.
Oligonucleotide fragment containing three tandem fork-­head do- Essentially, differential expression analysis was performed using
main (FHKD) consensus (canonical or noncanonical) motifs (21) the DESeq2 (v1.4.5) with Q value ≤ 0.05 to take insight to the
with minor promoters were cloned into pGL3-­basic vectors (Prome- change of phenotype, Gene Ontology (www.geneontology.org/)
ga). The pGL3-­3xFoxA1-­Luc was transiently transfected using Lipo- and Kyoto Encyclopedia of Genes and Genomes KEGG (www.
fectamine with PLUS reagents into 293T cells along with Renilla kegg.jp/) enrichment analysis of annotated different expressed gene
(pRL-­SV40-­Renilla, Promega) as internal control. Response rations was performed by Phyper (http://en.wikipedia.org/wiki/Hypergeo-
are expressed relative to signal obtained from the empty vector con- metric_distribution) based on hypergeometric test. The significant
trol wells transfected with pCDN3-­HA vector. After transfection for levels of terms and pathways were corrected by Q value with a rig-
40 hours, the cells were harvested and analyzed with a dual-­luciferase orous threshold (Q value ≤ 0.05) by Bonferroni.

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 13 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Mass spectrometry analysis Animal ethics statement

Gel bands of proteins with catalase activity were excised for in gel All animal procedures were authorized by the Institutional Animal
digestion, and proteins were identified by mass spectrometry as we Care and Use Committee (IACUC) of Sun Yat-­sen University (SYSU)
previously described (43). Briefly, proteins were disulfide reduced (approval number: SYSU-­IACUC-­2021-­000527).
with 5 mM dithiothreitol and alkylated with 11 mM iodoacetamide.
In-­gel digestion was performed using sequencing grade-­modified Hydrodynamic gene delivery
trypsin in 50 mM ammonium bicarbonate at 37°C overnight. The We use sleeping beauty (SB) transposon system to induce hepato-
peptides were extracted twice with 1% trifluoroacetic acid in 50% carcinogenesis mouse model (46). Four-­week-­old male mice were
acetonitrile aqueous solution for 1 hour. The peptide extracts were hydrodynamic delivered with 20 μg of Myr-­Akt1 and 20 μg of N-­
then centrifuged in a SpeedVac to reduce the volume. The peptides RasV12 combined with 2.75 μg of SB transposase in 2 ml of saline
were resuspended in 20 μl of 0.1% trifluoroacetic acid, followed by via tail vein injections within 10 s. Control mice were injected with
centrifugation at 20,000g at 4°C for 15 min to remove any particu- 20 μg of N-­RasV12 and 2.75 μg of SB transposase. Livers were har-
late impurities. vested and used for further analysis after 4-­to 6-­week hydrody-
For liquid chromatography–tandem mass spectrometry (LC-­ namic transfection.
MS/MS) analysis, peptides were separated by a 40-­min gradient
elution at a flow rate of 0.300 μl/min with a Thermo-­Dionex Ul- DEN-­induced model of liver cancer
timate 3000 HPLC system, which was directly interfaced with a B6.Cg (Ikbke−/−)tm1Tman mice were obtained from the Jackson Labora-
Thermo Orbitrap Fusion Lumos mass spectrometer. The analyti- tory. The offspring were genotyped by PCR of genomic DNA derived
cal column was a homemade fused silica capillary column (75 μm from tail clippings, and the homozygous mouse were used for the fur-
in inner diameter, 150 mm in length; Upchurch, Oak Harbor, ther experiments. These mice were housed in a specific pathogen–free
WA) packed with C-­18 resin (300 A, 5 μm; Varian, Lexington, environment. Two-­week-­old male Ikbke−/− mice were treated with
MA). Mobile phase A was consisted of 0.1% formic acid, and mo- DEN (25 mg/kg in 0.9% saline; Sigma-­Aldrich) by intraperitoneal in-
bile phase B was consisted of 100% acetonitrile and 0.1% formic jection to initiate liver tumor formation. Control mice were intraperi-
acid. An LTQ-­Orbitrap mass spectrometer was operated in the toneally injected with 0.9% saline at the same age. Mice were euthanized
data-­dependent acquisition mode using Byonic software, and at the indicated ages, and liver tissues were collected and analyzed.
there is a single full-­scan mass spectrum in the Orbitrap (300 to
1500 mass/charge ratio, 120,000 resolution) followed by 3-­s data-­ Xenografted mouse study
dependent MS/MS scans in an ion-­routing multipole at 40% Mouse xenograft assays were performed as described previously
normalized high energy collision dissociation (HCD). MS/MS (42). Briefly, 1 × 106 Huh7 cells were injected into the flank of six fe-
spectra from each LC-­MS/MS run were searched against the hu- male mice (SYSU, 4 to 5 weeks of age). COMPD1 is administered by
man database using Proteome Discoverer (version 1.4) searching gavage every 2 days at a dose of 30 mg/kg. Tumor size was measured
algorithm. The search criteria were as follows: Full tryptic speci- every 3 days with a caliper, and tumor volume was determined with
ficity was required; two missed cleavages were allowed; carbami- the formals: L × W2 × 0.5, where L is the longest diameter and W is
domethylation was set as fixed modification; phosphorylation the shortest diameter. Nude mice (4 weeks old, male) were acquired
and oxidation (M) were set as variable modifications; precursor from Guangdong Yaokang Biotechnology company.
ion mass tolerance was 20 parts per million for all MS acquired in
the Orbitrap mass analyzer; and fragment ion mass tolerance was Generation and validation of Foxa1S177A and Foxa1S177D
0.8 Da for all MS2 spectra acquired in the LTQ. High confidence knock-­in mice
score filter (false discovery rate < 1%) was used to select the “hit” The Foxa1S177A and Foxa1S177D knock in mice were generated using
peptides, and their corresponding MS/MS spectra were manually CRISPR-­Cas9 technology by the Cyagen Biosciences. To generate
inspected. Foxa1 knock-­in mice, the gRNAs against Foxa1, the donor oligo con-
taining S177A (TCG to GCC) or S177D (TCG to GAT) mutation, si-
Colony formation and soft agar assays lent mutations (TCC to AGT, CTC to CTG), and Cas9 were co-­injected
Following procedures previously described in (44), we seeded cells into fertilized mouse eggs to generate targeted knock-­in offspring. F0
into six-­well plates (300 or 600 cells per well) and left for 12 to 20 days founder animal was identified by PCR, followed by sequencing analy-
until formation of visible colonies. Colonies were washed with PBS sis or Bcl l restriction analysis. For Foxa1−/− mice, using CRISPR-­
two times and fixed with 10% acetic acid/10% methanol for 1 hour Cas9 technology to delete 2 base pair in mouse genome, and the
and then stained with 0.4% crystal violet in 20% ethanol for offspring were identified by PCR followed by sequencing analysis.
2 hours. After staining, the plates were washed and air-­dried, and These mice were housed in a specific pathogen–free environment. All
colony numbers were counted and quantified. The anchorage-­ experimental procedures were approved by the IACUC (RN150D) at
independent cell growth assays were performed as described pre- SYSU. The research projects that are approved by the IACUC are op-
viously (45). Briefly, the assays were preformed using six-­well erated according the applicable institutional regulations. The institute
plates where the solid medium consists of two layers. The bottom is committed to the highest ethical standards of care for animals used
layer contains 0.8% noble agar and the top layer contains 0.4% agar for the purpose of continued progress in the field of human can-
suspended with 1 × 104 or 3 × 104 cells. A total of 500 μl of com- cer research.
plete DMEM medium was added every 7 days to keep the top layer
moisture, and 4 weeks later, the cells were stained with iodonitro- Spontaneous liver cancer model with COMPD1 treatment
tetrazolium chloride (1 mg/ml) (Sigma-­Aldrich, I10406) for colo- The B6.Cg-­Tg (Cebpb-­Tta) 5Bjd/J mice and FVB/N-­Tg (tetO-­MYC)
ny visualization and counting. 36aBop/J mice were obtained from the Jackson Laboratory. The

Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024 14 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
offspring were genotyped by PCR of genomic DNA derived from tail
clippings, and the right genotyped mice (Cebpb-­Tta-­TetO-­​MYC mod-
el) were used to establish spontaneous liver cancer model. These mice
were housed in a specific pathogen–free environment. Three-­to
4-­week-­old Cebpb-­Tta-­TetO-­MYC male mice were gavage-­treated with
COMPD1 [30 mg/kg per 2 days in 0.5% O-­carboxymethylcellulose
(CMC)–Na]. Control mice were gavage-­treated with 0.5% CMC-­Na
at the same age. Mice were euthanized at the indicated ages, and liver
tissues were collected and analyzed.
Quantification and statistical analyses
GraphPad Prism version 6.0 was used for statistical analyses. The dif-
ferences between experimental groups were tested by Student’s t test
or one-­way analysis of variance (ANOVA) test. Two-­way ANOVA
was also applied in the experiments involved with treatment and time
course. The difference of IKBKE expression in liver tissues was ana-
lyzed with chi-­squared (χ2) test. For mice, tumor formation analysis
was carried out by Kaplan-­Meier curve analysis (log-­rank test at
<0.05). For in vivo tumor experiments and IHC staining, data collec-
tion and analysis such as tumor measuring were performed by re-
searchers who were blinded to group information. No blinding was
involved in other experiments such as Western blot. The threshold for
statistical significance was set to P < 0.05. Number of n and addi-
tional statistical tests can be found in the figure legends.
Supplementary Materials
This PDF file includes:
Figs. S1 to S8
REFERENCES AND NOTES
1. F. Bray, J. Ferlay, I. Soerjomataram, R. L. Siegel, L. A. Torre, A. Jemal, Global cancer statistics
2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185
countries. CA Cancer J. Clin. 68, 394–424 (2018).
2. J. M. Llovet, R. K. Kelley, A. Villanueva, A. G. Singal, E. Pikarsky, S. Roayaie, R. Lencioni,
K. Koike, J. Zucman-­Rossi, R. S. Finn, Hepatocellular carcinoma. Nat. Rev. Dis. Primers 7, 6
(2021).
3. D. Q. Huang, H. B. El-­Serag, R. Loomba, Global epidemiology of NAFLD-­related HCC:
Trends, predictions, risk factors and prevention. Nat. Rev. Gastroenterol. Hepatol. 18,
223–238 (2021).
4. Q. Zhou, W. Tian, Z. Jiang, T. Huang, C. Ge, T. Liu, F. Zhao, T. Chen, Y. Cui, H. Li, M. Yao, J. Li,
H. Tian, A positive feedback loop of AKR1C3-­mediated activation of NF-­κB and STAT3
facilitates proliferation and metastasis in hepatocellular carcinoma. Cancer Res. 81,
1361–1374 (2021).
5. G. He, M. Karin, NF-­κB and STAT3 -­key players in liver inflammation and cancer. Cell Res.
21, 159–168 (2011).
6. W. E. Naugler, T. Sakurai, S. Kim, S. Maeda, K. Kim, A. M. Elsharkawy, M. Karin, Gender
disparity in liver cancer due to sex differences in MyD88-­dependent IL-­6 production.
Science 317, 121–124 (2007).
7. G. He, G.-­Y. Yu, V. Temkin, H. Ogata, C. Kuntzen, T. Sakurai, W. Sieghart,
M. Peck-­Radosavljevic, H. L. Leffert, M. Karin, Hepatocyte IKKbeta/NF-­kappaB inhibits
tumor promotion and progression by preventing oxidative stress-­driven STAT3
activation. Cancer Cell 17, 286–297 (2010).
8. Y. Sunami, M. Ringelhan, E. Kokai, M. Lu, T. O’Connor, A. Lorentzen, A. Weber,
A.-­K. Rodewald, B. Müllhaupt, L. Terracciano, S. Gul, S. Wissel, F. Leithäuser, D. Krappmann,
P. Riedl, D. Hartmann, R. Schirmbeck, P. Strnad, N. Hüser, J. Kleeff, H. Friess, R. M. Schmid,
F. Geisler, T. Wirth, M. Heikenwalder, Canonical NF-­κB signaling in hepatocytes acts as a
tumor-­suppressor in hepatitis B virus surface antigen-­driven hepatocellular carcinoma by
controlling the unfolded protein response. Hepatology 63, 1592–1607 (2016).
9. Q. He, X. Xia, K. Yao, J. Zeng, W. Wang, Q. Wu, R. Tang, X. Zou, Amlexanox reversed
non-­alcoholic fatty liver disease through IKKε inhibition of hepatic stellate cell. Life Sci.
239, 117010 (2019).
10. H. Guan, H. Zhang, J. Cai, J. Wu, J. Yuan, J. Li, Z. Huang, M. Li, IKBKE is over-­expressed in
glioma and contributes to resistance of glioma cells to apoptosis via activating NF-­κB. J.
Pathol. 223, 436–445 (2011).
Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024
11. K. A. Fitzgerald, S. M. McWhirter, K. L. Faia, D. C. Rowe, E. Latz, D. T. Golenbock, A. J. Coyle,
S.-­M. Liao, T. Maniatis, IKKepsilon and TBK1 are essential components of the IRF3
signaling pathway. Nat. Immunol. 4, 491–496 (2003).
12. T. U. Barbie, G. Alexe, A. R. Aref, S. Li, Z. Zhu, X. Zhang, Y. Imamura, T. C. Thai, Y. Huang,
M. Bowden, J. Herndon, T. J. Cohoon, T. Fleming, P. Tamayo, J. P. Mesirov, S. Ogino,
K.-­K. Wong, M. J. Ellis, W. C. Hahn, D. A. Barbie, W. E. Gillanders, Targeting an IKBKE
cytokine network impairs triple-­negative breast cancer growth. J. Clin. Invest. 124,
5411–5423 (2014).
13. S. Wang, F. Xie, F. Chu, Z. Zhang, B. Yang, T. Dai, L. Gao, L. Wang, L. Ling, J. Jia, H. van Dam,
J. Jin, L. Zhang, F. Zhou, YAP antagonizes innate antiviral immunity and is targeted for
lysosomal degradation through IKKɛ-­mediated phosphorylation. Nat. Immunol. 18,
733–743 (2017).
14. D. Zhou, C. Conrad, F. Xia, J.-­S. Park, B. Payer, Y. Yin, G. Y. Lauwers, W. Thasler, J. T. Lee,
J. Avruch, N. Bardeesy, Mst1 and Mst2 maintain hepatocyte quiescence and suppress
hepatocellular carcinoma development through inactivation of the Yap1 oncogene.
Cancer Cell 16, 425–438 (2009).
15. K. Brennan, C. Lyons, P. Fernandes, S. Doyle, A. Houston, E. Brint, Engagement of Fas
differentially regulates the production of LPS-­induced proinflammatory cytokines and
type I interferons. FEBS J. 286, 523–535 (2019).
16. R. T. Peters, S. M. Liao, T. Maniatis, IKKepsilon is part of a novel PMA-­inducible IkappaB
kinase complex. Mol. Cell 5, 513–522 (2000).
17. K. H. Kaestner, J. Katz, Y. Liu, D. J. Drucker, G. Schütz, Inactivation of the winged helix
transcription factor HNF3α affects glucose homeostasis and islet glucagon gene
expression in vivo. Genes Dev. 13, 495–504 (1999).
18. A. Arruabarrena-­Aristorena, J. L. V. Maag, S. Kittane, Y. Cai, W. R. Karthaus, E. Ladewig,
J. Park, S. Kannan, L. Ferrando, E. Cocco, S. Y. Ho, D. S. Tan, M. Sallaku, F. Wu, B. Acevedo,
P. Selenica, D. S. Ross, M. Witkin, C. L. Sawyers, J. S. Reis-­Filho, C. S. Verma, R. Jauch,
R. Koche, J. Baselga, P. Razavi, E. Toska, M. Scaltriti, FOXA1 mutations reveal distinct
chromatin profiles and influence therapeutic response in breast cancer. Cancer Cell 38,
534–550.e9 (2020).
19. S. Gao, S. Chen, D. Han, D. Barrett, W. Han, M. Ahmed, S. Patalano, J. A. Macoska, H. H. He,
C. Cai, Forkhead domain mutations in FOXA1 drive prostate cancer progression. Cell Res.
29, 770–772 (2019).
20. S. E. Glont, I. Chernukhin, J. S. Carroll, Comprehensive genomic analysis reveals that the
pioneering function of FOXA1 is independent of hormonal signaling. Cell Rep. 26,
2558–2565.e3 (2019).
21. E. J. Adams, W. R. Karthaus, E. Hoover, D. Liu, A. Gruet, Z. Zhang, H. Cho, R. D. Loreto,
S. Chhangawala, Y. Liu, P. A. Watson, E. Davicioni, A. Sboner, C. E. Barbieri, R. Bose,
C. S. Leslie, C. L. Sawyers, FOXA1 mutations alter pioneering activity, differentiation and
prostate cancer phenotypes. Nature 571, 408–412 (2019).
22. Z. Li, G. Tuteja, J. Schug, K. H. Kaestner, Foxa1 and Foxa2 are essential for sexual
dimorphism in liver cancer. Cell 148, 72–83 (2012).
23. L. Liu, J. Cui, Y. Zhao, X. Liu, L. Chen, Y. Xia, Y. Wang, S. Chen, S. Sun, B. Shi, Y. Zou, KDM6A-­
ARHGDIB axis blocks metastasis of bladder cancer by inhibiting Rac1. Mol. Cancer 20, 77 (2021).
24. F. O. Enane, W. H. Shuen, X. Gu, E. Quteba, B. Przychodzen, H. Makishima, J. Bodo, J. Ng,
C. L. Chee, R. Ba, L. S. Koh, J. Lim, R. Cheong, M. Teo, Z. Hu, K. P. Ng, J. Maciejewski,
T. Radivoyevitch, A. Chung, L. L. Ooi, Y. M. Tan, P.-­C. Cheow, P. Chow, C. Y. Chan, K. H. Lim,
L. Yerian, E. Hsi, H. C. Toh, Y. Saunthararajah, GATA4 loss of function in liver cancer
impedes precursor to hepatocyte transition. J. Clin. Invest. 127, 3527–3542 (2017).
25. S. He, J. Zhang, W. Zhang, F. Chen, R. Luo, FOXA1 inhibits hepatocellular carcinoma
progression by suppressing PIK3R1 expression in male patients. J. Exp. Clin. Cancer Res.
36, 175 (2017).
26. Y. Wang, L. Yang, T. Chen, X. Liu, Y. Guo, Q. Zhu, X. Tong, W. Yang, Q. Xu, D. Huang, K. Tu, A
novel lncRNA MCM3AP-­AS1 promotes the growth of hepatocellular carcinoma by
targeting miR-­194-­5p/FOXA1 axis. Mol. Cancer 18, 28 (2019).
27. Z. Yuan, M. Ye, J. Qie, T. Ye, FOXA1 promotes cell proliferation and suppresses apoptosis in
HCC by directly regulating miR-­212-­3p/FOXA1/AGR2 signaling pathway. Onco. Targets.
Ther. 13, 5231–5240 (2020).
28. S. H. Park, K.-­W. Fong, J. Kim, F. Wang, X. Lu, Y. Lee, L. T. Brea, K. Wadosky, C. Guo,
S. A. Abdulkadir, J. D. Crispino, D. Fang, P. Ntziachristos, X. Liu, X. Li, Y. Wan, D. W. Goodrich,
J. C. Zhao, J. Yu, Posttranslational regulation of FOXA1 by Polycomb and BUB3/USP7
deubiquitin complex in prostate cancer. Sci. Adv. 7, eabe2261 (2021).
29. S. Gao, S. Chen, D. Han, Z. Wang, M. Li, W. Han, A. Besschetnova, M. Liu, F. Zhou, D. Barrett,
M. P. Luong, J. Owiredu, Y. Liang, M. Ahmed, J. Petricca, S. Patalano, J. A. Macoska, E. Corey,
S. Chen, S. P. Balk, H. H. He, C. Cai, Chromatin binding of FOXA1 is promoted by
LSD1-­mediated demethylation in prostate cancer. Nat. Genet. 52, 1011–1017 (2020).
30. L. Wang, M. Xu, C.-­Y. Kao, S. Y. Tsai, M.-­J. Tsai, Small molecule JQ1 promotes prostate
cancer invasion via BET-­independent inactivation of FOXA1. J. Clin. Invest. 130,
1782–1792 (2020).
31. Y. Wu, L. Zhou, Z. Wang, X. Wang, R. Zhang, L. Zheng, T. Kang, Systematic screening for
potential therapeutic targets in osteosarcoma through a kinome-­wide CRISPR-­Cas9
library. Cancer Biol. Med. 17, 782–794 (2020).
15 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
32. W. Xie, Q. Jiang, X. Wu, L. Wang, B. Gao, Z. Sun, X. Zhang, L. Bu, Y. Lin, Q. Huang, J. Li, J. Guo,
IKBKE phosphorylates and stabilizes Snail to promote breast cancer invasion and
metastasis. Cell Death Differ. 29, 1528–1540 (2022).
33. A. Bainbridge, S. Walker, J. Smith, K. Patterson, A. Dutt, Y. M. Ng, H. D. Thomas, L. Wilson,
B. M. Cullough, D. Jones, A. Maan, P. Banks, S. R. McCracken, L. Gaughan, C. N. Robson,
K. Coffey, IKBKE activity enhances AR levels in advanced prostate cancer via modulation
of the Hippo pathway. Nucleic Acids Res. 48, 5366–5382 (2020).
34. W. Reiley, M. Zhang, X. Wu, E. Granger, S.-­C. Sun, Regulation of the deubiquitinating
enzyme CYLD by IkappaB kinase gamma-­dependent phosphorylation. Mol. Cell. Biol. 25,
3886–3895 (2005).
35. G. M. Bernardo, R. A. Keri, FOXA1: A transcription factor with parallel functions in
development and cancer. Biosci. Rep. 32, 113–130 (2012).
36. R. W. Jenkins, A. R. Aref, P. H. Lizotte, E. Ivanova, S. Stinson, C. W. Zhou, M. Bowden,
J. Deng, H. Liu, D. Miao, M. X. He, W. Walker, G. Zhang, T. Tian, C. Cheng, Z. Wei,
S. Palakurthi, M. Bittinger, H. Vitzthum, J. W. Kim, A. Merlino, M. Quinn, C. Venkataramani,
J. A. Kaplan, A. Portell, P. C. Gokhale, B. Phillips, A. Smart, A. Rotem, R. E. Jones, L. Keogh,
M. Anguiano, L. Stapleton, Z. Jia, M. Barzily-­Rokni, I. Cañadas, T. C. Thai, M. R. Hammond,
R. Vlahos, E. S. Wang, H. Zhang, S. Li, G. J. Hanna, W. Huang, M. P. Hoang, A. Piris,
J.-­P. Eliane, A. O. Stemmer-­Rachamimov, L. Cameron, M.-­J. Su, P. Shah, B. Izar, M. Thakuria,
N. R. Le Boeuf, G. Rabinowits, V. Gunda, S. Parangi, J. M. Cleary, B. C. Miller, S. Kitajima,
R. Thummalapalli, B. Miao, T. U. Barbie, V. Sivathanu, J. Wong, W. G. Richards, R. Bueno,
C. H. Yoon, J. Miret, M. Herlyn, L. A. Garraway, E. M. Van Allen, G. J. Freeman,
P. T. Kirschmeier, J. H. Lorch, P. A. Ott, F. S. Hodi, K. T. Flaherty, R. D. Kamm, G. M. Boland,
K.-­K. Wong, D. Dornan, C. P. Paweletz, D. A. Barbie, Ex vivo profiling of PD-­1 blockade
using organotypic tumor spheroids. Cancer Discov. 8, 196–215 (2018).
37. M. Jechlinger, K. Podsypanina, H. Varmus, Regulation of transgenes in three-­
dimensional cultures of primary mouse mammary cells demonstrates oncogene
dependence and identifies cells that survive deinduction. Genes Dev. 23, 1677–1688
(2009).
38. S. Wang, S. Singh, M. Katika, S. Lopez-­Aviles, A. Hurtado, High throughput chemical
screening reveals multiple regulatory proteins on FOXA1 in breast cancer cell lines. Int. J.
Mol. Sci. 19, 4123 (2018).
39. E. Lai, V. R. Prezioso, E. Smith, O. Litvin, R. H. Costa, J. E. Darnell Jr., HNF-­3A, a hepatocyte-­
enriched transcription factor of novel structure is regulated transcriptionally. Genes Dev.
4, 1427–1436 (1990).
40. J. Prieto, Inflammation, HCC and sex: IL-­6 in the centre of the triangle. J. Hepatol. 48,
380–381 (2008).
Gao et al., Sci. Adv. 10, eadk2285 (2024) 7 February 2024
41. K. Ohata, T. Ichikawa, K. Nakao, M. Shigeno, D. Nishimura, H. Ishikawa, K. Hamasaki,
K. Eguchi, Interferon alpha inhibits the nuclear factor kappa B activation triggered by X
gene product of hepatitis B virus in human hepatoma cells. FEBS Lett. 553, 304–308 (2003).
42. Q. Jiang, N. Zheng, L. Bu, X. Zhang, X. Zhang, Y. Wu, Y. Su, L. Wang, X. Zhang, S. Ren, X. Dai,
D. Wu, W. Xie, W. Wei, Y. Zhu, J. Guo, SPOP-­mediated ubiquitination and degradation of
PDK1 suppresses AKT kinase activity and oncogenic functions. Mol. Cancer 20, 100 (2021).
43. Q. Jiang, X. Zhang, X. Dai, S. Han, X. Wu, L. Wang, W. Wei, N. Zhang, W. Xie, J. Guo,
S6K1-­mediated phosphorylation of PDK1 impairs AKT kinase activity and oncogenic
functions. Nat. Commun. 13, 1548 (2022).
44. X. Zhang, Q. Jiang, Y. Su, L. Bu, Z. Sun, X. Wu, B. Gao, L. Wang, Y. Lin, W. Xie, J. Guo, AMPK
phosphorylates and stabilises copper transporter 1 to synergise metformin and copper
chelator for breast cancer therapy. Br. J. Cancer 128, 1452–1465 (2023).
45. J. Guo, X. Dai, B. Laurent, N. Zheng, W. Gan, J. Zhang, A. Guo, M. Yuan, P. Liu, J. M. Asara,
A. Toker, Y. Shi, P. P. Pandolfi, W. Wei, AKT methylation by SETDB1 promotes AKT kinase
activity and oncogenic functions. Nat. Cell Biol. 21, 226–237 (2019).
46. Z. Chen, W. Zhu, S. Zhu, K. Sun, J. Liao, H. Liu, Z. Dai, H. Han, X. Ren, Q. Yang, S. Zheng,
B. Peng, S. Peng, M. Kuang, S. Lin, METTL1 promotes hepatocarcinogenesis via m7G tRNA
modification-­dependent translation control. Clin. Transl. Med. 11, e661 (2021).
Acknowledgments: We thank members of the J.G. laboratory for critical reading and kind
suggestions of the manuscript. We thank X. Tian and H. Deng in Center of Protein Analysis
Technology, Tsinghua University for MS analysis. Funding: This work was supported by the
National Key Research and Development Program of China 2023YFC3402100 (to J.G.) and
National Natural Science Foundation of China 32070767 (to J.G.). Author contributions:
Conceptualization: J.G. Methodology: B.G., X.W., L.B., Q.J., X.L., and J.L. Investigation: B.G., X.W.,
L.B., and Q.J. Visualization: B.G., X.W., L.W., H.L., X.Z., and Y.W. Material support: Y.L., L.X., X.L., and
Y.W. Supervision: J.G., W.X., Y.L., and L.X. Writing—original draft: J.G., B.G., and W.X. Writing—
review and editing: J.G., W.X., B.G., Y.L., and L.X. Competing interests: The authors declare that
they have no competing interests. Data and materials availability: RNA sequencing data
have been deposited in NCBI Sequence Read Archive (SRA) with accession number
PRJNA1062288. All data needed to evaluate the conclusions in the paper are present in the
paper and/or the Supplementary Materials.
Submitted 9 August 2023
Accepted 8 January 2024
Published 7 February 2024
10.1126/sciadv.adk2285
16 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

BIOCHEMISTRY Copyright © 2024 The

Authors, some rights
MutSβ-­MutLβ-­FANCJ axis mediates the restart of DNA reserved; exclusive
licensee American
replication after fork stalling at Association for the
Advancement of
cotranscriptional G4/R-loops Science. No claim to
original U.S.
Government Works.
Esin Isik1‡, Kaustubh Shukla2†, Michaela Pospisilova3,4†, Christiane König1, Martin Andrs1, Distributed under a
Satyajeet Rao1§, Vinicio Rosano1, Jana Dobrovolna2, Lumir Krejci3,4, Pavel Janscak1,2* Creative Commons
Attribution
Transcription-­replication conflicts (TRCs) induce formation of cotranscriptional RNA:DNA hybrids (R-loops) stabilized NonCommercial
by G-­quadruplexes (G4s) on the displaced DNA strand, which can cause fork stalling. Although it is known that these License 4.0 (CC BY-­NC).
stalled forks can resume DNA synthesis in a process initiated by MUS81 endonuclease, how TRC-­associated G4/R-
loops are removed to allow fork passage remains unclear. Here, we identify the mismatch repair protein MutSβ, an
MLH1-­PMS1 heterodimer termed MutLβ, and the G4-­resolving helicase FANCJ as factors that are required for
MUS81-­initiated restart of DNA replication at TRC sites in human cells. This DNA repair process depends on the
G4-­binding activity of MutSβ, the helicase activity of FANCJ, and the binding of FANCJ to MLH1. Furthermore, we
show that MutSβ, MutLβ, and MLH1-­FANCJ interaction mediate FANCJ recruitment to G4s. These data suggest that
MutSβ, MutLβ, and FANCJ act in conjunction to eliminate G4/R-loops at TRC sites, allowing replication restart.

INTRODUCTION or in vitro transcription of plasmid regions containing a G-­rich

Genome replication is frequently challenged by various obstacles
on the DNA template, which can halt replication fork progression.
This phenomenon, known as DNA replication stress, can cause
genomic instability, a hallmark of cancer (1). A potent source of
DNA replication stress is active transcription if the transcription and
replication machineries converge on the same DNA template (2).
These transcription-­replication conflicts (TRCs) are markedly in-
creased upon overexpression of oncogenes such as CCNE1 and MYC
that induce firing of DNA replication origins within highly tran-
scribed genes in early S phase (3). Blockage of replication fork pro-
gression by head-­on transcription results from the formation of an
aberrant structure termed R-loop (4–6). In an R-loop, the nascent
RNA transcript pairs with the template DNA strand behind the tran-
scription complex, while the nontemplate strand loops out (7, 8). R-
loops impede DNA replication by inducing a fork remodeling process
known as replication fork reversal (6), which involves replisome dis-
assembly and the pairing of nascent DNA strands to form a DNA
duplex protected by BRCA2-­stabilized RAD51 filament (9). R-loops
tend to form in regions where the nontemplate strand is rich in gua-
nines (Gs) with the potential to fold into G-­quadruplex (G4) struc-
tures (10). These G4s may contribute to the formation of RNA:DNA
hybrids during transcription, possibly by stabilizing the single-­
stranded DNA (ssDNA) tract within the resulting R-loop (8, 11).
Intriguingly, these G4/R-loop structures, also termed G-loops, have
been directly observed by electron microscopy following intracellular
1
Institute of Molecular Cancer Research, University of Zurich, Winterthurerstrasse
190, 8057 Zurich, Switzerland. 2Institute of Molecular Genetics of the Czech Academy of
Sciences, Videnska 1083, 143 00 Prague, Czech Republic. 3Department of Biology
and National Centre for Biomolecular Research, Masaryk University, Kamenice 5/A7,
Brno 62500, Czech Republic. 4International Clinical Research Center, St Anne's
University Hospital, Pekarska 53, Brno 656 91, Czech Republic.
*Corresponding author. Email: pjanscak@​imcr.​uzh.​ch
†These authors contributed equally to this work.
‡Present address: Department of Oncologic Pathology, Dana-­Farber Cancer Institute,
450 Brookline Avenue, Mayer Building 641, Boston, MA 02215, USA.
§Present address: Swiss Institute for Experimental Cancer Research (ISREC), School
of Life Sciences, Ecole Polytechnique Fédérale de Lausanne (EPFL), 1015 Lausanne,
Switzerland.
sequence in the nontemplate strand (12).
Reversed replication forks at sites of TRCs can resume DNA syn-
thesis in a multistep process involving fork cleavage–religation
cycles mediated by MUS81-­E ME1 endonuclease and the DNA
ligase IV/XRCC4 complex coupled to transcription restart depen-
dent on the RNA polymerase II (RNAPII) elongation factor ELL (6).
A prerequisite for this sequential restart of transcription and repli-
cation at TRC sites is not only the removal of the R-loop but also
the removal of G4 structures on the nontemplate strand, which, if
persistent, would impede the progression of the reactivated repli-
some (13). While several human DNA helicases, including FANCJ,
BLM, WRN, PIF1, and RTEL1, are known to unwind G4 structures
in vitro (14), the specific helicase that removes G4 structures at
sites of R-loop–mediated TRCs to allow replication fork passage
is not known.
We have recently shown that the human mismatch repair (MMR)
factor MutSβ, a heterodimer of MSH2 and MSH3 proteins, binds to
G4 structures in vitro and regulates telomeric G4/R-loops in Alterna-
tive Lengthening of Telomeres (ALT) cancer cells to prevent telomere
fragility and excision into extrachromosomal C-­circles (15). The lack
of MutSβ caused accumulation of RNA:DNA hybrids and G4s also at
nontelomeric sites, suggesting a genome-­wide role for this protein in
G4/R-loop removal (15). In post-­replicative MMR, MutSβ is respon-
sible for the recognition of insertion/deletion loops generated by
polymerase slippage, whereas base-­base mismatches generated by
nucleotide misincorporation are recognized by MutSα, a heterodimer
of MSH2 and MSH6 proteins (16, 17). After mismatch recognition by
MutSα or MutSβ, an MLH1-­PMS2 heterodimer, known as MutLα, is
recruited to trigger downstream steps of the canonical MMR pathway
(18). MutLα acts as a latent endonuclease that is activated upon bind-
ing to the MutSα/MutSβ mismatch complex, generating nicks in the
discontinuous DNA strand in a manner dependent on the DNA-­
loaded form of proliferating cell nuclear antigen (PCNA) (19). These
strand breaks serve as an entry site for EXO1, which removes the mis-
match in a 5′-­to-­3′ exonucleolytic reaction controlled by replication
protein A (RPA), allowing nascent strand resynthesis (18, 19). In
addition to MutLα, MLH1 forms two other heterodimers in

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 1 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

mammalian cells: MutLβ (MLH1-­PMS1) and MutLγ (MLH1-­MLH3) lack of MutSα did not markedly increase RNH1(D210N)-­GFP foci
(20, 21). While MutLγ is known to have an essential function in mei- either in EdU+ or in EdU− cells (fig. S1C).
otic recombination and triplet repeat expansion (22–24), the function R-loop formation in S phase is promoted by head-­on TRCs (4).
of MutLβ remains elusive. MutLβ (PMS1) lacks the motif required for Using proximity ligation assay (PLA) followed by quantitative image-­
endonuclease activity (19), suggesting a role other than that in mis- based cytometry analysis, we found that the depletion of MSH3 but
match correction. not MSH6 increased colocalization between elongating form of
Here, we provide several lines of evidence, suggesting that the RNAPII (RNAPIIS2P, where S2P stands for phosphorylation of
restart of replication forks stalled by cotranscriptional G4/R-loops serine-­2 in the C-­terminal repeat domain of RNAPII) and the replisome
requires a coordinated action of MutSβ, MutLβ, and FANCJ. Our component PCNA predominantly in EdU+ nuclei of U2OS cells
data suggest a model wherein MutSβ and MutLβ mediate FANCJ (Fig. 1, A to D, and fig. S1D). This was prevented if cells were pre-
recruitment to G4s within R-loops at TRC sites for G4 unwinding. treated with the RNAPII transcription initiation inhibitor triptolide
The data also suggest that G4 unwinding is required for the removal (TRP) (Fig. 1, A to D). MSH3 depletion also increased S phase–
of R-loops. specific colocalization of elongating RNAPII with FANCD2
(fig. S1E), which is recruited to stalled replication forks (27). Together,
these data suggest that a lack of MutSβ causes persistent TRCs.
RESULTS Having determined that MutSβ deficiency induces S phase–
MutSβ deficiency induces R-loop–dependent specific accumulation of R-loops, we sought to investigate whether
replication stress this condition has an impact on the progression of DNA replica-
U2OS cells lacking MutSβ were shown to accumulate G4 structures tion forks. To this end, replication tracts in cells transfected with
and R-loops not only at telomeres but also at nontelomeric loci (15), MSH3, MSH6, and control siRNA (siLuc), respectively, were se-
suggesting a genome-­wide role for MutSβ in the regulation of these quentially pulse-­labeled with halogenated thymidine (T) analogs
aberrant structures. To explore this further, we made use of a U2OS 5-­chloro-­2′-­deoxyuridine (CldU) and 5-­iodo-­2′-­deoxyuridine (IdU)
T-­REx cell line inducibly expressing a green fluorescent protein for 30 min each, and the stretches of nascent DNA were then visual-
(GFP)–tagged mutant of ribonuclease (RNase) H1, RNH1(D210N)-­ ized on DNA fiber spreads by indirect immunofluorescence (Fig. 2A).
GFP, that binds but does not cleave RNA:DNA hybrids and thus en- Using this so-­called DNA fiber assay, we found that depletion of
ables visualization of the sites of R-loop formation as fluorescent MSH3 but not MSH6 impaired replication fork progression in U2OS
nuclear foci (25, 26). Cells were transfected with MSH3 small interfer- cells as revealed by shorter replication tracts compared to mock-­
ing RNA (siRNA) to selectively eliminate MutSβ, and the expression depleted cells (Fig. 2, A and B). We also analyzed the progression of
of RNH1(D210N)-­GFP was induced by adding doxycycline 24 hours sister replication forks, which, under unperturbed conditions or
before the analysis of individual cells by fluorescent microscopy upon a global replication slowing, would progress with similar rates,
(fig. S1A). Pulse labeling of nascent DNA with 5-­ethynyl-­deoxyuridine resulting in a symmetric pattern of replication tracts. However, in
(EdU) was also carried out to identify S phase cells. We observed that case of fork stalling by a barrier encountered on only one side of the
MSH3 depletion increased the number of RNH1(D210N)-­GFP foci replication origin, an asymmetric pattern of sister replication tracts
in EdU+ cells but not in EdU− cells, suggesting that MutSβ regulates would be observed, a phenotype termed sister fork asymmetry
R-loops formed specifically during S phase (fig. S1, B and C). We also (Fig. 2C). Intriguingly, we could observe a substantially higher asym-
analyzed R-loop levels in cells depleted of MSH6, which eliminates metry of sister replication tracts in MSH3-­deficient cells as compared
the MSH2-­MSH6 heterodimer MutSα (fig. S1A). We found that the to mock-­depleted cells or cells depleted of MSH6 (Fig. 2D). Similar

Fig. 1. MutSβ deficiency causes persistent TRCs. (A) Workflow of cell treatments for PLA to measure colocalization of elongating RNA polymerase II (RNAPIIS2P) and
PCNA in cell nuclei. U2OS cells transfected with appropriate siRNA were pulse-­labeled with EdU (10 μM) for 15 min. Where indicated, TRP (1 μM) was added 2 hours before
cell harvest. (B) Western blot analysis of extracts of U2OS cells transfected with indicated siRNAs. (C) Representative images of RNAPIIS2P/PCNA PLA foci in EdU-­positive
nuclei of U2OS cells transfected with indicated siRNAs and incubated without [dimethyl sulfoxide (DMSO)] or with TRP as in (A). Scale bar, 10 μm. (D) Quantification of PLA
foci in the images represented in (C). At least 741 nuclei were analyzed for each condition. Representative plot from two independent experiments yielding similar results
is shown. Images from PLA with PCNA or RNAPIIS2P antibody only (negative control, in green) were also analyzed. Red lines represent median values. ****P < 0.0001; ns,
not significant (Kruskal-­Wallis test followed by Dunn’s multiple comparisons test).

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 2 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. MutSβ deficiency induces R-­loop-­dependent replication stress. (A to F) MSH3 depletion induces replication fork stalling and micronucleation in U2OS cells.
(A) Top: Workflow of DNA fiber assays. Bottom: Representative images of replication tracts on DNA fibers of cells transfected with indicated siRNAs. (B) Quantification of
the lengths of replication tracts (CldU + IdU) in the images represented in (A) (n ≥ 311). (C) Representative images of symmetric and asymmetric replication tracts of sister
forks observed on DNA fiber spreads in (A). (D) Plot of the values of IdU tract length ratio of sister forks (sister fork ratio) obtained for indicated conditions (n ≥ 120). (E) Top:
Workflow of cytokinesis-­block micronucleus assay. Bottom: Representative images of binucleated cells with or without micronucleus (red arrow). (F) Quantification of the
percentage of micronucleus-­positive binucleated cells for indicated conditions. Data are means ± SEM (n = 3). (G to K) RNase H1 overexpression rescues replication fork
stalling and micronucleation in MSH3-­depleted cells. (G) Western blot analysis of extracts of U2OS T-­REx [RNH1-­GFP] cells transfected with indicated siRNAs and treated
with (+) or without (−) doxycycline (Dox; 1 ng/ml) to induce expression of GFP-­tagged RNase H1 (RNH1-­GFP). (H) Workflow of DNA fiber assays. (I) Quantification of the
lengths of DNA replication tracts for indicated conditions (n ≥ 402). (J) Plot of the values of sister fork ratio obtained for indicated conditions (n ≥ 211). (K) Top: Workflow
of cytokinesis-­block micronucleus assay. Bottom: Quantification of the percentage of micronucleus-­positive binucleated cells for indicated conditions. Data are
means ± SEM (n = 3). Kruskal-­Wallis test followed by Dunn’s multiple comparisons test was used in (B), (D), (I), and (J). ****P < 0.0001. One-­way ANOVA with Tukey’s mul-
tiple comparisons correction was used in (F) and (K). ***P < 0.001 and *P < 0.05. All DNA fiber experiments were performed three times. Red lines represent median val-
ues. Scale bars, 10 μm. ns, not significant.

results were obtained with two other cell lines, RPE-­1 and HeLa Kyoto
(fig. S1, F to I). These data suggest that MutSβ deficiency increases
the frequency of replication fork stalling events. In line with this con-
clusion, we observed that MSH3 depletion markedly induced the
formation of micronuclei in U2OS cells (Fig. 2, E and F), a chromo-
some instability phenotype resulting from chromosome segregation
defects in mitosis caused by regions of under-­replicated DNA (28,
29). On the other hand, MSH6 depletion increased micronucleation
frequency only slightly compared to MSH3-­depleted cells (Fig. 2F).
To investigate whether R-loops contribute to replication stress
observed in cells lacking MutSβ, we used a stable U2OS T-­REx cell
line with doxycycline-­regulated expression of the wild-­type form of
GFP-­tagged RNase H1 (6). We found that RNase H1 overexpression,
which eliminates R-loops (6), prevented fork slowing and sister fork

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 3 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

asymmetry as well as micronucleation in MSH3-­depleted cells
(Fig. 2). These results suggest that fork stalling and chromosomal
instability induced by MutSβ deficiency are caused by R-loops.
MutSβ is required for replication restart at R-loop–stalled
replication forks
Replication fork reversal induced by G4/R-loop structures is coun-
teracted by poly(adenosine 5′-­diphosphate–ribose) polymerase 1
(PARP1)–regulated replication restart via the MUS81-­LIG4-­ELL axis
(6). To determine how MutSβ averts R-loop–mediated replication
stress, we first tested whether replication fork slowing in MSH3-­
depleted cells can be rescued by PARP inhibition, which promotes
replication restart at G4/R-loop–stalled forks (6, 30, 31). By DNA fi-
ber assay, we found that PARP inhibition with olaparib failed to re-
store normal fork progression in MSH3-­depleted U2OS cells (fig. S2,
A and B). In contrast, olaparib almost completely prevented replica-
tion fork stalling in cells depleted of RTEL1 helicase (fig. S2, A and B),
known to suppress the accumulation of G4/R-loops (32). These data
suggest that MutSβ might mediate the elimination G4/R-loops at
sites of head-­on TRCs to facilitate replication restart via the MUS81-­
LIG4-­ELL axis. To explore this hypothesis, we used DNA fiber assay
to analyze the effect of MSH3 depletion on replication fork dynamics
upon treatment of cells with the G4-­stabilizing ligand pyridostatin
(PDS) or the DNA topoisomerase I inhibitor camptothecin (CPT),
which promote R-loop formation and hence induce G4/R-loop–
mediated fork stalling (6, 11, 32). As previously reported (6, 30, 31),
we observed that the treatment of mock-­depleted U2OS cells with
PDS or CPT during the second pulse labeling with IdU resulted in
replication tract shortening (IdU:CldU ratio < 1), which was largely
prevented by addition of olaparib 2 hours before DNA fiber labeling
(Fig. 3, A and B). However, PARP inhibition could not rescue
PDS-­or CPT-­induced replication fork slowing in MSH3-­depleted
cells (Fig. 3B and fig. S2, C and D), mirroring the scenario in cells with
defects in the MUS81-­LIG4-­ELL axis involved in restarting R-loop–
stalled forks (6). In contrast, the depletion of MSH6 or RTEL1 had
minimal impact on the rescue of PDS/CPT-­induced fork slowing by
PARP inhibition (Fig. 3B and fig. S2D). These results were reproduced
in noncancerous RPE-­1 cells (fig. S2E). Moreover, MSH3 or MSH6
depletion did not alter IdU:CldU ratio upon PARP inhibition alone
in U2OS cells (Fig. 3B). Thus, we conclude that MutSβ but not MutSα
plays a major role in the restart of DNA replication at sites of R-loop–
mediated TRCs, possibly by mediating G4/R-loop removal. Consis-
tent with the proposal that MutSβ promotes replication restart via the
MUS81-­LIG4-­ELL axis, we found that MSH3 depletion did not fur-
ther exacerbate replication fork stalling in MUS81 knockout HeLa
Kyoto cells (fig. S2F to H) (6).
MutSβ mediates replication restart at R-loop–stalled
replication forks through its G4-­binding activity
We have recently shown that MutSβ specifically binds to G4 DNA
structures in vitro (15). To explore the possible link between the
G4-­binding activity of MutSβ and its function in processing G4/R-
loops, we sought to generate a mutant of MutSβ that is defective
in G4 binding. Previous studies have shown that Y245S/K246E sub-
stitutions in the mismatch-­binding domain (MBD) of MSH3 abolish
the interaction of MutSβ with insertion/deletion loops and hairpin
loops in vitro (33, 34). We therefore investigated whether these muta-
tions also impair the binding of MutSβ to G4 DNA structures. To this
end, we used a 5′-­biotinylated 66-­nucleotide oligomer DNA oligonu-
cleotide composed of a stretch of 45 dTs followed by the d(GGGT)4
sequence, which adopts a thermally stable parallel G4 topology in the
presence of K+ ions (35). By pull-­down assay with streptavidin-­coated
magnetic beads, we compared the binding of wild-­type and Y245S/
K246E variants of MutSβ to this G4-­forming oligonucleotide after G4
folding in a KCl-­based buffer. The binding reactions were also supple-
mented with the ssDNA-­binding protein RPA to prevent nonspecific
binding of MutSβ to the stretch of dTs. We found that wild-­type
MutSβ bound effectively to the G4-­forming oligonucleotide both in
the absence and presence of bound RPA, suggesting a specific binding
(Fig. 4A). In contrast, the Y245S/K246E variant of MutSβ exhibited
markedly reduced binding to this DNA structure when compared
to the wild-­type protein (Fig. 4A). To substantiate these findings,
G4-­binding activities of the wild-­type and mutant forms of MutSβ
were compared by electrophoretic mobility shift assay (EMSA)

Fig. 3. MutSβ is required for replication restart at R-loop–stalled replication forks. (A) Top: Workflow of DNA fiber assays with U2OS cells. The PARP inhibitor olaparib
[10 μM; PARP inhibition (PARPi)/Pi] was added 2 hours before replication tract labeling and was also present during the labeling. PDS (10 μM) or CPT (100 nM) were added
together with IdU. Bottom: Representative images of DNA replication tracts of U2OS cells transfected with control siRNA (siLuc) and treated as indicated. Scale bar, 10 μm.
(B) Effect of depletion of MSH3 or MSH6 on replication fork progression in U2OS cells treated as indicated. Scatter plot of the IdU/CldU tract length ratio is shown. At least
328 replication tracts were scored in three independent experiments for each condition. Red lines represent median values. ****P < 0.0001 (Kruskal-­Wallis test followed
by Dunn’s multiple comparisons test). ns, not significant.

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 4 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. MBD of MutSβ mediates G4 DNA binding. (A) Binding of wild-­type (WT) and Y245S/K246E mutant forms of MutSβ (MutSβWT and MutSβY245S/K246E) to biotinylated
G4 DNA substrate b66-­G4 in the absence or presence of RPA. The G4 oligonucleotide (20 nM) was preincubated with RPA (40 nM) before addition of MutSβ (50 nM). After
incubation for 45 min at 4°C, DNA substrate was captured with streptavidin beads and bound proteins were detected by Western blotting. (B) Binding of WT and Y245S/
K246E forms of MutSβ to CEB1 G4 detected by EMSA. Top: 30 nM Cy3-­conjugated CEB1 G4 was incubated with indicated concentrations of MutSβ variants at 37°C for
10 min. DNA-­protein complexes were separated from the free oligonucleotide by agarose gel electrophoresis. Bottom: Quantification of gels represented in the top
panel. Data are means ± SD (n = 4).

using a cyanine 3 (Cy3)–conjugated oligonucleotide, CEB1, known to
fold into a parallel G4 structure, as confirmed by circular dichro-
ism measurements (fig. S3A). In a protein titration experiment, we
found that the Y245S/K246E MutSβ variant showed notably lower
binding to the CEB1 G4 structure compared to the wild-­type protein
(Fig. 4B). Together, these data suggest that MutSβ binds to G4 DNA
via its MBD.
To determine whether the binding of MutSβ to G4 is required for
the restart of R-loop–stalled replication forks, we established U2OS
T-­REx cell lines expressing either wild-­type or Y245S/K246E mutant
form of a GFP-­MSH3 chimera from a doxycycline-­regulated cyto-
megalovirus (CMV) promoter (Fig. 5A). Silent mutations were intro-
duced into both chimeras to confer siRNA resistance. We found that
ectopic expression of wild-­type MSH3 but not the Y245S/K246E vari-
ant could rescue replication fork slowing and sister fork asymmetry in
cells depleted of endogenous MSH3 (Fig. 5, B to D). Notably, these
findings could be reproduced in different cell clones (fig. S3, B to D).
In addition, ectopic expression of the Y245S/K246E variant of MSH3
did not suppress the formation of micronuclei induced by depletion
of endogenous MSH3, as seen with ectopic expression of wild-­type
MSH3 (Fig. 5E). In cells expressing the Y245S/K246E variant of
MSH3 instead of endogenous MSH3, replication fork slowing in-
duced by PDS or CPT could not be rescued by PARP inhibition, in
contrast to cells ectopically expressing wild-­type MSH3 (Fig. 5F and
fig. S3E). These data suggest that the G4-­binding activity of MutSβ is
required for the restart of R-loop–stalled replication forks via the
MUS81-­LIG4-­ELL axis.
Replication restart at R-loop–stalled forks
depends on MutLβ
In canonical MMR, mismatch recognition by MutSα or MutSβ is fol-
lowed by the recruitment of MutLα (MLH1-­PMS2 heterodimer),
which triggers downstream steps of the repair process (18). MutLβ
(MLH1-­PMS1 heterodimer) is not required for MMR (20); however,
a recent study showed that PMS1 depletion increases RNA:DNA
hybrid levels in HeLa cells (36). Given these findings, we sought to
investigate the possible roles of MutLα and MutLβ as factors that co-
operate with MutSβ to promote the reactivation of R-loop–stalled
forks via the MUS81-­LIG4-­ELL axis. We initially tested the ef-
fect of siRNA-­mediated depletion of PMS1 or PMS2 on replication
fork progression in U2OS cells. By DNA fiber assay, we found that
depletion of either protein impaired replication fork progression as
revealed by shorter replication tracts compared to mock-­depleted
cells (fig. S4, A and B). Notably, we observed a marked asymmetry of
sister replication tracts in PMS1-­depleted cells but not in PMS2-­
depleted cells (fig. S4C), suggesting that the lack of MutLβ but not
MutLα induces replication fork stalling. Overexpression of RNase H1
completely rescued replication fork slowing and sister fork asymme-
try in PMS1-­depleted cells, while it had no impact on the reduced rate
of replication fork progression in PMS2-­depleted cells (Fig. 6, A to C).
In addition, we found that depletion of PMS1 but not PMS2 increased
the frequency of micronucleation, which could be also prevented by
RNase H1 overexpression (Fig. 6D and fig. S4D). Together, these re-
sults suggest that MutLβ deficiency induces R-loop–dependent repli-
cation stress, whereas MutLα deficiency globally reduces replication
fork velocity presumably due to defective MMR. In line with this
conclusion, we observed a higher frequency of S phase–specific
RNH1(D210N)-­GFP foci and increased colocalization between elon-
gating RNAPII and PCNA in PMS1-­ deficient cells compared to
mock-­depleted cells or cells depleted of PMS2 (Fig. 6, E and F). Ulti-
mately, we tested the effect of PMS1 and PMS2 depletion on the res-
cue of PDS-­or CPT-­ induced replication fork slowing by PARP
inhibition in U2OS cells. We found that in cells depleted of PMS1,
PARP inhibition did not restore the normal rate of fork progression
upon treatment with PDS or CPT, whereas it did in cells depleted of
PMS2 (Fig. 6G). These results suggest that the restart of R-loop–
stalled forks via the MUS81-­LIG4-­ELL axis requires MutLβ.
Replication restart at R-loop–stalled forks requires the
helicase activity of FANCJ
Several DNA helicases have been shown to unwind G4 DNA
structures in vitro (14). One of these helicases, FANCJ, has been

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 5 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. MutSβ mediates replication restart at R-loop–stalled replication forks through its G4-­binding activity. (A to E) Y245S/K246E substitutions in MSH3 induce
replication fork stalling and micronucleation. (A) Western blot analysis of expression of WT or Y245S/K246E variants of GFP-­tagged MSH3 in U2OS T-­REx cells. Cells were
transfected with indicated siRNAs for 72 hours and treated with doxycycline (0.4 ng/ml) in the past 24 hours or left untreated to induce transgene expression. (B) Experi-
mental workflow of DNA fiber assays with cells in (A). (C) Quantification of replication tract lengths (CldU + IdU) for indicated conditions from three independent experi-
ments (n ≥ 325). Horizontal lines represent median values. (D) Plot of the values of IdU tract length ratio of sister forks (sister fork ratio) obtained for indicated conditions
(n ≥ 133). Horizontal lines represent median values. (E) Quantification of the frequency of micronuclei for indicated conditions. Data are means ± SEM (n = 3). (F) Y245S/
K246E substitutions in MSH3 abolish the rescue of PDS-­induced fork slowing by PARPi. Top: Workflow of DNA fiber assays with cells in (A). PARPi treatment was carried out
as in Fig. 3A. PDS (10 μM) was added together with IdU. Bottom: Scatter plot of the values of IdU/CldU tract length ratio obtained for indicated conditions in two indepen-
dent experiments (n ≥ 251). Horizontal lines represent median values. Kruskal-­Wallis test followed by Dunn’s multiple comparisons test was used in (C), (D), and (F).
****P < 0.0001. One-­way ANOVA with Tukey’s multiple comparisons correction was used in (E). ***P < 0.001, **P < 0.01, and *P < 0.05. ns, not significant.

identified as an interactor of MLH1 and PMS1 in human cells
(37, 38). In addition, using an affinity pull-­down assay with pu-
rified recombinant proteins, we found that FANCJ was bound to
MutLβ (fig. S5). Therefore, we sought to investigate whether
FANCJ is essential for the reactivation of R-loop–stalled forks
via the MUS81-­LIG4-­ELL pathway. To this end, we first evalu-
ated the effect of FANCJ deficiency on the progression of repli-
cation forks in unperturbed cells. We found that siRNA-­mediated
depletion of FANCJ reduced the rate of replication fork progres-
sion in U2OS cells, which was accompanied by a higher fre-
quency of replication fork stalling events as revealed by sister
fork asymmetry (fig. S6, A to C). A reduced replication speed
and asymmetric progression of sister replication forks could be
also observed in FANCJ knockout HeLa FlpIn T-­REx (HeLa
FIT) cells (fig. S6, D to G), generated by CRISPR-­C as9 technol-
ogy (39). The replication fork slowing and sister fork asymmetry
phenotypes induced by FANCJ depletion could be completely
rescued by RNase H1 overexpression (Fig. 7, A to D), suggesting
that, similar to MutSβ and MutLβ, FANCJ suppresses R-loop–
mediated replication stress. Consistently, FANCJ-­depleted cells
exhibited an increased incidence of unresolved TRCs as reflect-
ed by a higher frequency of S phase–specific RNH1(D210N)-­
GFP foci and PLA foci between elongating RNAPII and PCNA
compared to mock-­depleted cells (Fig. 7, E and F). Moreover, we
found that depletion of MSH3, PMS1, or MUS81 did not further
exacerbate replication fork slowing and sister fork asymmetry in
FANCJ knockout HeLa FIT cells, suggesting that all these pro-
teins act in the same pathway for resolution of R-loop–stalled
replication forks (fig. S6, H to J). Last, we tested the effect of
FANCJ deficiency on the progression of replication forks in cells
treated with PDS or CPT. We found that the unrestrained fork
progression conferred by PARP inhibition in cells treated with
these G4/R-loop–­inducing drugs was abrogated by FANCJ defi-
ciency (Fig. 7G and fig. S6, K and L). Together, these results

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 6 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. Replication restart at R-loop–stalled forks depends on MutLβ. (A to D) PMS1 depletion induces R-loop–dependent replication stress. (A) Top: Workflow of
DNA fiber assays with U2OS T-­REx [RNH1-­GFP] cells. Where required, overexpression of RNH1-­GFP was induced with doxycycline (1 ng/ml). Bottom: Western blot analysis
of extracts of U2OS T-­REx [RNH1-­GFP] cells transfected with indicated siRNAs. (B) Quantification of the lengths of DNA replication tracts (CldU + IdU) for indicated condi-
tions in three independent experiments (n ≥ 456). (C) Plot of the values of IdU tract length ratio of sister forks (sister fork ratio) obtained for indicated conditions (n ≥ 92).
(D) RNH1-­GFP overexpression suppresses micronucleation induced by PMS1 depletion in U2OS T-­REx [RNH1-­GFP] cells. Data are means ± SEM (n = 3). (E) PMS1 depletion
induces formation of nuclear foci of catalytically inactive form of RNase H1 [RNH1(D210N)-­GFP]. Percentage of EdU+ cells with >10 RNH1(D210N)-­GFP foci is plotted. Data
are means ± SEM (n = 3). At least 300 cells per condition were analyzed. (F) PMS1 depletion induces colocalization of elongating RNA polymerase II (RNAPIIS2P) and PCNA
in nuclei of U2OS cells as determined by PLA. Workflow of cell treatments was as in Fig. 1A. At least 178 nuclei were analyzed for each condition to quantify RNAPIIS2P/
PCNA PLA foci. A representative plot from two independent experiments is shown. (G) PMS1 depletion abolishes the rescue of PDS/CPT-­induced fork slowing by PARPi in
U2OS cells. Left: Workflow of DNA fiber assays as in Fig. 3A. Right: Scatter plot of the values of IdU/CldU tract length ratio obtained for indicated conditions from two inde-
pendent experiments (n ≥ 312). Kruskal-­Wallis test followed by Dunn’s multiple comparisons test was used in (B), (C), (F), and (G). ****P < 0.0001. One-­way ANOVA with
Tukey’s multiple comparisons correction was used in (D) and (E). ****P < 0.0001, ***P < 0.001, and *P < 0.05. Red lines represent median values. ns, not significant.

suggest that FANCJ is needed for the restart of G4/R-loop–stalled
forks via the MUS81-­LIG4-­ELL axis.
A K52R substitution in the ATP-­binding pocket of FANCJ
inactivates its ATPase/helicase function. This mutant was shown
to be inefficient in G4 DNA unwinding both in vitro and in vivo
(39, 40). To explore whether FANCJ exerts its function in coun-
teracting G4/R-loop–dependent replication stress through its
helicase activity, we stably transfected FANCJ knockout HeLa
FIT cells with vectors expressing either wild-­type FANCJ or its
K52R mutant. We found that ectopic expression of wild-­type
FANCJ rescued the replication fork stalling phenotypes in both
unchallenged and olaparib/PDS-­t reated FANCJ knockout cells,
whereas the expression of the K52R mutant of FANCJ did not
(Fig. 8, A to C). These data suggest that the helicase activity
of FANCJ is required for its function in the restart of R-loop–
stalled replication forks.

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 7 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Replication restart at R-loop–stalled forks depends on FANCJ. (A to D) FANCJ depletion induces R-loop–dependent replication stress. (A) Workflow of DNA fiber assays
with U2OS T-­REx [RNH1-­GFP] cells. Where required, overexpression of RNH1-­GFP was induced with doxycycline (1 ng/ml) for 24 hours. (B) Western blot analysis of extracts of U2OS
T-­REx [RNH1-­GFP] cells transfected with indicated siRNAs. (C) Quantification of the lengths of DNA replication tracts (CldU + IdU) for indicated conditions from three independent
experiments (n ≥ 402). (D) Plot of the values of IdU tract length ratio of sister forks (sister fork ratio) obtained for indicated conditions (n ≥ 96). (E) FANCJ depletion induces accu-
mulation of nuclear foci of catalytically inactive form of RNase H1 [RNH1(D210N)-­GFP]. Cells were cultured, treated, and analyzed as depicted in fig. S1 (A to C). Percentage of EdU+
cells with >10 RNH1(D210N)-­GFP foci is plotted. Data are means ± SEM (n = 3). At least 300 cells were analyzed in each experiment for each condition. (F) FANCJ depletion in-
duces colocalization of elongating RNA polymerase II (RNAPIIS2P) and PCNA in the nuclei of U2OS cells lacking FANCJ as determined by PLA. At least 303 nuclei were analyzed for
each condition to quantify RNAPIIS2P/PCNA foci. A representative plot of two independent experiments yielding similar results is shown. (G) FANCJ depletion abolishes the rescue
of PDS/CPT-­induced fork slowing by PARPi in U2OS cells. Top: Workflow of DNA fiber assays. Conditions for olaparib (PARPi) and PDS/CPT treatments were the same as in
Fig. 3A. Bottom: Plot of the values of IdU/CldU tract length ratio obtained for indicated conditions in three independent experiments (n ≥ 291). Kruskal-­Wallis test followed by
Dunn’s multiple comparisons test was used in (C), (D), (F), and (G). ****P < 0.0001. Paired t test was used in (E). **P < 0.01. Red lines represent median values. ns, not significant.

Reactivation of R-loop–stalled replication forks requires
MLH1-­FANCJ interaction
The findings presented thus far suggest a model wherein MutSβ and
MutLβ recruit FANCJ to unwind G4 structures within R-loops at
TRC sites, which is needed for replication restart via the MUS81-­
LIG4-­ELL axis. To monitor the recruitment of FANCJ to the sites of
G4s in cells, we sought to perform PLA using antibodies against
FANCJ and G4 structures (BG4). Cells were pulsed with EdU to mark
the S phase population and then subjected to PLA. We observed S
phase–specific PLA foci in nuclei of U2OS cells under unperturbed
conditions (Fig. 9, A and B, and fig. S7A). This phenotype was sup-
pressed by a 2-­hour pretreatment of cells with TRP, suggesting a de-
pendence on transcription (Fig. 9B and fig. S7A). We then tested
whether these PLA foci, indicating FANCJ/G4 colocalization, depend
on MutSβ and MutLβ. We found that siRNA-­mediated depletion of
MSH3 and PMS1, respectively, lowered the number of PLA foci in
U2OS cells without affecting FANCJ protein levels (Fig. 9, C and D).
These results suggest that both MutSβ and MutLβ are required for the
recruitment of FANCJ to G4s in vivo.
Previous studies identified a FANCJ variant containing K141/142A
substitutions that render FANCJ defective in its interaction with
MLH1 and abolish FANCJ recruitment to sites of DNA crosslinks
without affecting its helicase activity (38, 41, 42). Intriguingly, we
found that FANCJ knockout HeLa FIT cells expressing the K141/142A
mutant of FANCJ had a defect in the recruitment of FANCJ to G4s, as
revealed by PLA (Fig. 9, E and F). This observation supports the

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 8 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 8. Replication restart at R-loop–stalled forks requires the helicase activity of FANCJ. (A) K52R substitution in helicase motif I of FANCJ induces sister fork asym-
metry. Top: Workflow of DNA fiber assays with HeLa FlpIn T-­REx (FANCJ+/+) or CRISPR-­Cas9–generated FANCJ knockout HeLa FIT cells (FANCJ−/−). Where indicated, the
expression of WT or K52R (helicase-­dead) variants of FANCJ was induced with 1 μg/ml doxycycline (Dox) for 24 hours in FANCJ−/− cells. Bottom: Plot of the values of IdU
tract length ratio of sister forks (sister fork ratio) for indicated conditions. At least 100 tracts were scored for each condition in two independent experiments. (B) K52R
substitution in helicase motif I of FANCJ abolishes the rescue of PDS-­induced fork slowing by PARPi. Top: Workflow of DNA fiber assays with cells as in (A). Conditions for
olaparib (PARPi) and PDS treatments were the same as in Fig. 3A. Bottom: Plot of the values of IdU/CldU tract length ratio obtained for indicated conditions. At least 300
replication tracts were measured for each condition in two independent experiments. (C) Western blot analysis of extracts of cells in (A). Kruskal-­Wallis test followed by
Dunn’s multiple comparisons test was used in (A) and (B). ****P < 0.0001. Red lines represent median values. ns, not significant.

notion that the physical interaction between FANCJ and MLH1 is re-
quired for efficient recruitment of FANCJ to G4 sites. In addition,
complementation with the K141/142A mutant of FANCJ failed to
rescue replication fork slowing and sister fork asymmetry in FANCJ
knockout cells (Fig. 9G and fig. S7B). Moreover, in these cells, PARP
inhibition did not restore the normal rate of replication fork progres-
sion upon PDS treatment, as it did in FANCJ knockout cells comple-
mented with wild-­type FANCJ (Fig. 9H). These data suggest that
MLH1-­FANCJ interaction is required for replication restart at
G4/R-loop–stalled replication forks.
DISCUSSION
Cotranscriptional R-loops, a major source of DNA replication stress,
are formed preferentially in actively transcribed regions where the
nontranscribed strand contains runs of guanines that can fold into G4
structures. Specific G4 ligands that stabilize G4s increase R-loop levels in
these G-­rich regions and induce R-loop–dependent stalling of DNA
replication forks (6, 11, 32). This suggests that the formation of G4s in
the displaced nontranscribed strand stabilizes R-loops and promotes
their extension. Our earlier studies have shown that replication forks
stalled by G4/R-loops can resume DNA synthesis by a process involv-
ing fork cleavage-­religation cycles mediated by MUS81-­EME1 endo-
nuclease and the LIG4-­XRCC4 complex, which presumably relieve
the torsional stress in the DNA template generated by transcription-­
replication encounters, allowing for transcription and replication
restart (6). Here, we identify three additional factors that act in this
DNA repair process: the G4-­resolving helicase FANCJ; the MutSβ
heterodimer (MSH2-­MSH3), which specifically binds to G4s; and the
MutLβ heterodimer (MLH1-­PMS1), which physically interacts with
FANCJ. We show that replication restart at G4/R-loops depends on
the helicase activity of FANCJ, the G4-­binding activity of MutSβ, and
the binding of FANCJ to MLH1. Moreover, we present evidence sug-
gesting that MutSβ and MutLβ mediate the recruitment of FANCJ to
the sites of G4s in cell nuclei. On the basis of these findings, we pro-
pose a model wherein MutSβ, MutLβ, and FANCJ act in conjunction
to eliminate G4 structures within R-loops at TRC sites, thereby facili-
tating replication restart via the MUS81-­LIG4-­ELL axis (Fig. 10). It is
possible that the unwinding of G4s is essential for the helicase-­
mediated removal of R-loops. In line with this assumption, we found
that depletion of MSH3, PMS1, and FANCJ, respectively, induced
R-loop accumulation in S phase cells (Figs. 6E and 7E and fig. S1C).
Notably, we have recently identified the DEAD-­box helicase DDX17
as a factor that might be involved in R-loop unwinding at sites of
R-loop–mediated TRCs to promote replication restart via the MUS81-­
LIG4-­ELL axis (43). Consistently, DDX17 has been found to efficient-
ly unwind synthetic R-loop structures in vitro (43). To initiate strand
separation, DEAD-­box helicases are loaded directly onto the duplex
region of an RNA, aided by a proximal single-­stranded nucleic acid
region that does not have to be covalently linked to the helix (44).
Thus, in the context of an R-loop structure, DDX17 loading on the
RNA:DNA hybrid could be mediated through its binding to the dis-
placed nontranscribed strand (Fig. 10). In this scenario, the presence
of G4s in the single-­stranded region of the R-loop could potentially
block DDX17 loading, thereby preventing R-loop unwinding (Fig. 10).
The K141/K142A substitutions in FANCJ were also shown to
abolish the binding of FANCJ to G4 in vitro (45). However, the
K141/K142A mutant of FANCJ displayed G4-­unwiding activity if

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 9 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 9. Reactivation of R-loop–stalled replication forks requires MLH1-­FANCJ interaction. (A to D) Depletion of MSH3 or PMS1 impairs FANCJ recruitment to G4s in
U2OS cells. (A) Representative images of FANCJ/BG4 PLA foci in EdU-­positive nuclei. Where indicated, cells were treated with TRP (1 μM) for 2 hours. EdU was present the
last 15 min. Scale bar, 10 μm. (B) Quantification of FANCJ/BG4 PLA foci in images represented in (A). (C) Quantification of FANCJ/BG4 PLA foci in U2OS cells lacking MSH3
or PMS1. A representative plot from two independent experiments is shown (n ≥ 1364). Images from PLA with BG4 or FANCJ antibody only were also analyzed. (D) Western
blot analysis of extracts of cells in (C). (E and F) K141/142A substitutions in FANCJ abolish its recruitment to G4s in unchallenged cells. (E) Western blot analysis of extracts
of FANCJ+/+ or FANCJ−/− HeLa FIT cells. Where indicated, the expression of WT or K141/142A variants of FANCJ was induced with doxycycline (1 μg/ml) in FANCJ−/− cells.
(F) Quantification of FANCJ/BG4 PLA foci in cells in (E). A representative plot from two independent experiments is shown (n ≥ 420). (G) K141/142A substitutions in FANCJ
induce replication fork stalling in unchallenged cells. Top: Experimental workflow of DNA fiber assays with cells in (E). Bottom: Plot of the values of sister fork ratio obtained
for indicated conditions in two independent experiments (n ≥ 202). (H) K141/142A substitutions in FANCJ abolish the rescue of PDS-­induced fork slowing by PARPi. Top:
Workflow of DNA fiber assays with cells in (E). Bottom: Plot of the values of IdU/CldU tract length ratio obtained for indicated conditions in two independent experiments
(n ≥ 189). Kruskal-­Wallis test followed by Dunn’s multiple comparisons test was used in (B), (C), and (F) to (H). ****P < 0.0001. Red lines represent median values.
ns, not significant.

the DNA substrate contained 5′-­ssDNA overhang for FANCJ load-
ing (45). Moreover, we have found that the recruitment of FANCJ to
G4 sites is mediated by MutSβ and MutLβ. Thus, it is unlikely that
the observed defect of FANCJ K141/K142A-­expressing cells in re-
starting G4/R-loop–stalled forks is primarily caused by the inability
of the FANCJ mutant to recognize G4.
Our data implicating FANCJ in G4 unwinding during the restart of
R-loop–stalled forks are consistent with the previous observation of a
FANCJ requirement for unrestrained DNA synthesis in HLTF-­deficient
cells exposed to hydroxyurea (HU) (46), since HU-­induced replication
fork stalling is known to be caused by R-loops (47). Moreover, abroga-
tion of HU-­induced fork reversal by HLTF depletion has been shown to
trigger replication restart via the MUS81-­LIG4-­ELL axis (47). However,
a recent study using high-­resolution microscopy has shown that FANCJ,
in an interplay with RPA, regulates spontaneously formed G4s on newly
unwound DNA at replication forks, immediately behind the MCM heli-
case and before nascent DNA synthesis (48). The increased frequency of
these G4/replisome structures in FANCJ-­depleted cells was associated
with a reduction in EdU incorporation both at nuclear and individual
replisome level, particularly upon PDS treatment (48). However, it is not
clear to what extent these G4s contribute to replication fork stalling ob-
served in FANCJ-­deficient cells by DNA fiber assay, as this phenotype
was largely rescued by overexpression of RNase H1, suggesting a depen-
dence on G4/R-loop structures (Fig. 7D).
Cells defective in MLH1-­FANCJ interaction or FANCJ helicase
activity show hypersensitivity to DNA-­crosslinking agents such as
mitomycin C (MMC) (38). However, the underlying molecular
mechanism is not clear. In addition to DNA interstrand crosslinks,

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 10 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 10. Model for the elimination of G4/R-loops at TRC sites and the consequences of its failure. MutSβ binds to G4s in the non-­transcribed strand through its MBD
domain and recruits MutLβ-­FANCJ complex for G4 unwinding. This facilitates the loading of the DDX17 helicase on the ssDNA loop and the subsequent unwinding of the
R-loop, leading to the sequential restart of transcription and replication via the MUS81-­LIG4-­ELL pathway. Failure of G4 unwinding impairs DDX17 loading, leading to
persistent fork stalling. Sβ, MutSβ; Lβ, MutLβ; FJ, FANCJ; CMG, CDC45-­MCM2-­7-­GINS helicase.
MMC was also found to induce accumulation of R-loop–dependent
DNA damage in human cells (49). Moreover, replication fork stall-
ing induced by MMC can be rescued by inhibition of fork reversal
(50), a phenomenon observed for R-loop–stalled forks (6). Thus, it
will be interesting to explore whether MMC sensitivity of FANCJ-­
deficient cells stems from a defect in restarting R-loop–stalled forks
via the MUS81-­LIG4-­ELL axis.
MATERIALS AND METHODS
Plasmid constructions
The E. coli strain DH10B was used for all plasmid constructions
(Thermo Fisher Scientific). The vectors for inducible expression of
wild-­type and mutant (Y245S/K246E) variants of human MSH3
were constructed using the plasmid pEGFP-­c1-­hMSH3 that ex-
presses MSH3 as an N-­terminal fusion with GFP (34). The Afe I/
Kpn I fragment of pEGFP-­c1-­hMSH3 including the GFP-­MSH3
chimera was introduced into the multiple cloning site (Pme I/Kpn I)
of the plasmid pAIO to generate a regulatable transcription unit
with a CMV enhancer/promoter followed by two tet operators (51).
Using the QuickChange II Site-­Directed Mutagenesis Kit (Agilent
Technologies), a set of silent mutations was introduced into the
MSH3 open reading frame (ORF) to confer resistance to siMSH3#1
in the resulting transcript. Ultimately, this pAIO-­GFP-­MSH3
construct was subjected to another round of site-­directed mutagen-
esis to introduce Y245S/K246E mutations into the MSH3 gene
(pAIO-­GFP-­MSH3-­Y245S/K246E). Using site-­directed mutagene-
sis, K141/142A variant of FANCJ was generated in the FANCJ-­
pDONR221 vector (39). To construct the transfer vector for generation
of a bacmid encoding for His-­tagged PMS1, the PMS1 ORF was
amplified by polymerase chain reaction (PCR) using the primers 5′
-­TATATGGATCCATGCATCATCATCATCACCACGGTTCTGGT
ATGAAACAATTGCCTGCGG-­3′ and 5′-­ATATATAAGCTTTCA
TGTAGTTTCTGGAAGATAGGT-­3′, which introduce Bam HI and
Hind III sites, respectively, and the forward primer also introduces
an N-­terminal 6xHis-­tag to PMS1. As a template for PCR, the plas-
mid pFastBac1-­hPMS1 was used (20). The resulting PCR product
Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024
was cut with Bam HI and Hind III and cloned in pFastBac1 (pFastBac1-­
6xHis-­hPMS1).
Generation of stable cell lines
U2OS T-­REx cells were transfected with pAIO-­GFP-­MSH3 or pAIO-­
GFP-­MSH3-­Y245S/K246E constructs using TransIT-­X2 reagent
(Mirus). Transfected cells were selected in the presence of puromycin
(1 μg/ml; InvivoGen). Clones were isolated after 10 to 14 days of
growth in the selection medium. The expression of the desired
proteins upon doxycycline induction was tested by Western blotting
with anti-­GFP and anti-­MSH3 antibodies. CRISPR-­Cas9–generated
FANCJ HeLa FlpIn T-­REx (FANCJ−/− HeLa FIT) knockout cells (39)
were cotransfected with pDONR221-­based vector expressing either
wild-­type, K52R, or K141/142A variant of FANCJ, as well as an Flp-­
In-­compatible expression vector plasmid (pOG44) encoding for Flp
recombinase, using TransIT-­X2 reagent (Mirus). Transfected
FANCJ−/− HeLa FIT cells were selected in the presence of blasticidin
(15 μg/ml; InvivoGen) and hygromycin (150 μg/ml; InvivoGen). The
expression of these constructs allowing complementation of FANCJ-­
deficient cells with FANCJ variants upon doxycycline induction was
tested by Western blotting with anti-­FANCJ antibody.
Cell culture
U2OS (HTB-­96), HeLa Kyoto (CVCL_1922), and RPE-­1 cells (CRL-­
4000) were cultured in Dulbecco’s modified Eagle’s medium (DMEM;
Thermo Fisher Scientific) supplemented with 5% fetal calf serum
(FCS; Thermo Fisher Scientific) and streptomycin/penicillin (100 U/ml)
at 37°C in a humidified incubator containing 5% CO2. U2OS T-­
REx–derived cell lines were grown in DMEM supplemented with
Tet-­free approved 5% FCS, streptomycin/penicillin (100 U/ml), puro-
mycin (1 μg/ml), and hygromycin B (50 μg/ml; InvivoGen) at 37°C in
a humidified incubator containing 5% CO2. U2OS T-­REx cell lines
carrying RNH1-­GFP and RNH1(D210N)-­GFP transgenes, respec-
tively, were described previously (25). Expression of RNase H1 vari-
ants was induced by addition of doxycycline (catalog no. 631311,
Takara Bio) to a concentration of 1 ng/ml for 24 hours. Ectopic ex-
pression of MSH3 variants was tuned to be comparable with the level
11 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

of endogenous MSH3 by titration of doxycycline concentration. Typi- (E-­3) mouse monoclonal (1:1000; sc-­515302, Santa Cruz Biotechnol-
cally, doxycycline was added to a concentration of 0.4 ng/ml for ogy), PMS2 (clone A16-­4) mouse monoclonal (1:1000; 556415, BD
24 hours. HeLa FIT (52) and FANCJ−/− HeLa FIT cell lines were Pharmingen), BRIP1/FANCJ rabbit polyclonal (1:1000; NBP1-­31883,
grown in DMEM supplemented with 5% FCS at 37°C in a humidified Novus Biologicals), TFIIH p89 (S-­19) rabbit polyclonal (1:1000; sc-­
incubator at a 5% CO2-­containing atmosphere. Stably transfected 293, Santa Cruz Biotechnology), glyceraldehyde-­3-­phosphate dehy-
FANCJ−/− HeLa FIT cells were cultured in DMEM supplemented drogenase (0411) mouse monoclonal (1:1000; sc-­47724, Santa Cruz
with 5% FCS, blasticidin (15 μg/ml; InvivoGen), and hygromycin Biotechnology), β tubulin (clone TUB 2.1) mouse monoclonal
(150 μg/ml; InvivoGen). Expression of constructs in FANCJ−/− HeLa (1:1000; T-­4026, Sigma-­Aldrich), and γ-­tubulin rabbit polyclonal
FIT was induced by doxycycline (1 μg/ml) for 24 hours. The following (1:1000; T-­3559, Sigma-­Aldrich). The following secondary antibodies
drugs were used at the indicated final concentrations, unless stated used for Western blotting are as follows: goat anti-­rabbit IgG-­HRP
otherwise: PDS pentahydrochloride (10 μM; catalog no. 4763, Tocris (1:5000; A5050, Sigma-­Aldrich) and goat anti-­mouse IgG HRP
Biosciences), CPT (100 nM; catalog no. C9911, Sigma-­ Aldrich), (1:10,000; A4416, Sigma-­Aldrich).
Olaparib (10 μM; catalog no. S1060, Selleckchem), cytochalasin B
(2 μg/ml; catalog no. C6762, Sigma-­Aldrich), and TRP (1 μM; catalog Detection of R-loops by monitoring RNH1(D210N)-­GFP foci
no. T3652, Sigma-­Aldrich). U2OS T-­REx [RNH1(D210N)-­GFP] cells were grown on autoclaved
glass coverslips. RNH1(D210N)-­GFP expression was induced by ad-
siRNA transfections dition of doxycycline to a concentration of 1 ng/ml for 24 hours. To
Transfections of siRNAs (a final concentration of 40 nM) were done at mark the S phase population, cells were pulse-­labeled with EdU (10 μM)
30 to 40% confluency using Lipofectamine RNAiMAX (Invitrogen) for 15 min. After a brief wash with ice-­cold 1× PBS, cells were pre-­
according to the manufacturer’s instructions. Twenty-­four hours after extracted using ice-­cold pre-­extraction solution [25 mM Hepes-­
siRNA transfection, the medium was exchanged with fresh medium. NaOH (pH 7.7), 50 mM NaCl, 1 mM EDTA, 3 mM MgCl2, 300 mM
The sequences of the sense strand of siRNA duplexes are listed in sucrose, and 0.5% (v/v) Triton X-­100] on ice for 10 min and then
table S1. For ectopic expression of MSH3 variants in stable U2OS T-­ briefly washed with pre-­extraction solution that does not contain
REx cell lines, endogenous MSH3 was depleted by transfection of Triton X-­100. The cells were fixed with 4% (v/v) formaldehyde for
siMSH3#1 for a total time of 72 hours. Forty-­eight hours after siRNA 15 min at RT in the dark, and after several washes with 1× PBS, cells
transfection, doxycycline (0.4 ng/ml) was added to induce expression were permeabilized with 0.5% (v/v) Triton X-­100/1× PBS for 5 min
of MSH3 variants for 24 hours. In the figures where siMSH3 and at RT. Following three washes with 1× PBS, cells were blocked in
siPMS1 are indicated, data for siMSH3#1 and siPMS1#1, respectively, 3% bovine serum albumin (BSA)/1× PBS for 30 min at RT. To
are shown. visualize EdU incorporation, Click-­iT Plus EdU Alexa Fluor 594
Imaging reaction was performed according to the manufacturer’s
Preparation of cell extracts and Western blot analysis protocol (Thermo Fisher Scientific). After washing with 3%
Cells that were trypsinized and washed with 1× phosphate-­buffered BSA/1× PBS and then 1× PBS, the cells were counterstained with
saline (PBS) were resuspended in lysis buffer [50 mM tris-­HCl buffer 4′,6-­diamidino-­2-­phenylindole (DAPI) (1 μg/ml) in distilled water
(pH 7.5), 120 mM NaCl, 20 mM NaF, 1 mM EDTA, 6 mM EGTA, 15 mM and mounted using Fluoromount-­G (Invitrogen). All images were
Na-­pyrophosphate, and 0.5% (v/v) NP-­40] supplemented with prote- captured with a fluorescent microscope at ×63 magnification (Leica
ase inhibitor cocktail (cOmplete, EDTA free; Sigma-­Aldrich) and microscope, model DM6B coupled to the DMC 2900 digital camera).
phosphatase inhibitor cocktail (PhosSTOP, Roche). After sonication At least 300 nuclei were analyzed using ImageJ. DAPI signal was used
for 5 min with a diagenode sonicator, cell lysate was centrifuged at for the identification of the nuclei, and EdU signal was used for the
11,000 rpm for 30 min at 4°C to separate cellular debris from the sol- determination of S phase cells. The percentage of EdU-­positive cells
uble cell fraction. Bradford assay was carried out for the determi- with more than 10 RNH1(D210N)-­GFP foci was calculated.
nation of protein concentration. A total of 30 to 50 μg of total
protein from whole-­cell extracts were loaded onto 8 to 10% SDS–­ In situ PLA
polyacrylamide gel electrophoresis (SDS-­PAGE) gels. Proteins sepa- Cells were grown on autoclaved glass coverslips. To mark the S phase
rated by SDS-­PAGE were blotted onto a Hybond-­P polyvinylidene population, cells were pulse-­labeled with EdU (10 μM) for 15 min.
difluoride membrane (GE Healthcare) in a semi-­dry transfer appara- Cells were briefly washed with ice-­cold 1× PBS and incubated with
tus at 56 mA for 2 hours. After Ponceau staining, the membranes were ice-­cold 0.5% (v/v) Triton X-­100/1× PBS supplemented with protease
blocked in 2% ECL Prime blocking reagent (GE Healthcare) in TBS-­T inhibitor cocktail (cOmplete, EDTA free; Sigma-­Aldrich) for 10 min
[20 mM tris-­HCl (pH 7.4), 150 mM NaCl, and 0.1% (v/v) Tween 20] on ice for pre-­extraction. After washing with 1× PBS three times, cells
for 30 min at room temperature (RT). Then, the membranes were were fixed with 4% (v/v) formaldehyde for 15 min at RT in the dark.
incubated in primary antibodies diluted in 2% ECL blocking solution Following three washes with 1× PBS, coverslips were incubated with
at 4°C overnight. After three washes in TBS-­T, the membranes were prechilled 100% (v/v) methanol for 20 min at −20°C. Following three
incubated in the corresponding horseradish peroxidase (HRP)– washing steps with 1× PBS, cells were permeabilized with 0.2% (v/v)
coupled secondary antibody for 1 hour at RT. Then, the membranes Triton X-­100/1× PBS for 10 min at RT. Cells were washed with 1×
were washed three times with TBS-­T, and protein bands were detected PBS and incubated with blocking solution (5% BSA/1× PBS) for 30 min
using ECL Western blotting substrate (Pierce). The following primary at RT. Primary antibody (diluted in blocking solution) incubation was
antibodies were used for immunoblotting: GFP rabbit polyclonal performed overnight at 4°C. After several washes of coverslips with
(1:2000; ab290, Abcam), MSH3 (H300) rabbit polyclonal (1:1000; sc-­ 1× PBS, PLA was carried out according to the manufacturer’s proto-
11441, Santa Cruz Biotechnology), MSH6 (clone 44/MSH6) mouse col (Sigma-­Aldrich). Briefly, coverslips were incubated with anti-­
monoclonal (1:1000; 610918, BD Transduction Laboratories), PMS1 Mouse MINUS and anti-­Rabbit PLUS PLA probes (Sigma-­Aldrich)

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 12 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
for 1 hour at 37°C. Coverslips were then washed twice in Wash Buffer
A [0.01 M tris (pH 7.4), 0.15 M NaCl, and 0.05% (v/v) Tween 20] for
5 min by gentle shaking. The ligation step where minus and plus
probes were ligated was performed for 30 min at 37°C. After two
washes in wash buffer A for 5 min by gentle shaking, incubation with
“Duolink In Situ Detection Reagents Green” (Sigma-­ Aldrich) or
“Duolink In Situ Detection Reagents Red” (Sigma-­Aldrich) was car-
ried out for the amplification step for 100 min. Then, coverslips were
washed twice with 1× wash buffer B [0.2 M tris-­HCl (pH 7.5) and 0.1 M
NaCl] for 10 min and once with 0.01× wash buffer B for 5 min by
gentle shaking at RT. After a brief incubation with blocking solution,
the Click-­iT EdU reaction was performed according to the manufac-
turer’s protocol for the visualization of EdU incorporation (Thermo
Fisher Scientific). After washing with blocking buffer and then 1×
PBS, the cells were counterstained with DAPI (1 μg/ml) in distilled
water and mounted using Fluoromount-­G (Invitrogen). The repre-
sentative images were acquired with a Leica DM6B fluorescent micro-
scope at ×63 magnification. For determination of the number of PLA
foci, images were captured automatically on an IX83 microscope
(Olympus) equipped with ScanR imaging platform using 40×/0.9 nu-
merical aperture objective. Number of PLA foci per nucleus from at
least 1000 nuclei marked by DAPI signal was determined using ScanR
analysis software. EdU signal was used for the determination of S
phase cells. The primary antibodies used in PLA are: PCNA rabbit
polyclonal (1:500; catalog no. ab18197, Abcam), FANCD2 rabbit
polyclonal (1:500; NB100-­ 182, Novus Biologicals), RNAPII (H5)
mouse monoclonal (1:1000; 920204, BioLegend), BRIP1/FANCJ rab-
bit polyclonal (1:500; NBP1-­31883, Novus Biologicals), DNA/RNA
G-­quadruplex [BG4] recombinant mouse monoclonal (1:1000;
Ab00174-­1.1, Absolute Antibody).
DNA fiber assay
For measurements of replication fork dynamics under different con-
ditions, well-­established DNA fiber spreading assay was performed as
previously described (6). Briefly, cells were sequentially pulse-­labeled
with two thymidine analogs: 30 μM CldU (Sigma-­Aldrich) for 30 min
and, after three washes with 1× PBS, 250 μM IdU (Sigma-­Aldrich) for
30 min. After labeling, cells were washed three times with ice-­cold 1×
PBS to cease nucleotide incorporation and trypsinized. Cells were
then resuspended in 1× PBS to a concentration of 2.5 × 105 cells/ml.
The labeled cells were diluted with unlabeled cells in 1:1 ratio, and 2.5 μl
of cell suspension was spotted on the glass slide and air-­dried for 1 min.
The cells were then lysed on the glass slide with 7.5 μl of lysis buffer
[200 mM tris-­HCl (pH 7.5), 50 mM EDTA, and 0.5% (w/v) SDS]. The
slides were then incubated in horizontal position for exactly 9 min at
RT without moving and then tilted at 15° to 45° to allow the DNA fi-
bers to slowly stretch along the slide in 1 to 3 min. The slides were
air-­dried at RT and fixed with methanol/acetic acid (3:1) at 4°C over-
night. DNA fibers on glass slides were denatured with 2.5 M HCl for
1 hour at RT and then washed four times with 1× PBS and blocked
with blocking solution [2% BSA/1× PBS/0.1% (v/v) Tween 20] for
10 min at RT. Subsequently, DNA fibers were incubated with primary
antibodies diluted in blocking solution for 2.5 hours at RT. Rat mono-
clonal anti–5-­bromo-­2′-­deoxyuridine (BrdU) antibody (1:500; ab6326,
Abcam) was used to detect CldU and mouse monoclonal anti-­BrdU
antibody (1:100; 347580, BD Biosciences) was used to detect
IdU. Following five washes with PBS-­T [1× PBS/ 0.2% (v/v) Tween
20], slides were incubated for 1 hour in the dark at RT with the sec-
ondary antibodies diluted in the blocking solution. Donkey anti-­rat
Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024
Cy3 (1:150; 712-­166-­153, Jackson ImmunoResearch) was used against
anti-­CldU antibody and goat anti-­mouse Alexa 488 (1:300; A110334,
Thermo Fisher Scientific) was used against anti-­IdU antibody. The
secondary antibody incubation was followed by five washes with PBS-
­T, and then the slides were air-­dried before the mounting of coverslips
with Antifade Gold (Invitrogen) mounting medium. Images of la-
beled replication tracts were acquired with a Leica DM6B fluorescent
microscope (63×/1.40 oil immersion), and their lengths were mea-
sured by using segmented line tool of ImageJ.
Cytokinesis-­block micronucleus assay
Cells were cultured on autoclaved coverslips and the culture medium
was supplemented with cytochalasin B (2 μg/ml; micro-­filament as-
sembly inhibitor; C6762, Sigma Aldrich) for 16 hours before the cell
harvest to block cytokinesis. The cells were fixed with 4% (v/v) form-
aldehyde for 15 min at RT in the dark, and after several washes with
1× PBS, cells were counterstained with DAPI (1 μg/ml) in distilled
water and mounted using Fluoromount-­G (Invitrogen). Images were
captured with a Leica DM6B fluorescent microscope at ×63 magnifi-
cation, and the percentage of binucleated cells with micronuclei was
determined. At least 250 binucleated cells were analyzed in each
experiment.
Protein expression and purification
RPA was produced in bacteria and purified as previously described
(53). MutSβ, MutLβ, and FANCJ proteins were produced in Spodoptera
frugiperda (Sf9) insect cells using the Bac-­to-­Bac Baculovirus Expres-
sion System (catalog no. 10359016, Thermo Fisher Scientific) as pre-
viously described (20, 34, 39). Purification of MutSβ (wild-­type and
Y245S/K246E mutant) carrying a hexahistidine tag on MSH3 was
described previously (34). For purification of MutLβ, a 400-­ml culture
of Sf9 cells at 1 × 106 cells/ml was infected with freshly amplified recom-
binant baculoviruses encoding for hMLH1 and His-­hPMS1, respec-
tively. Seventy-­two hours after infection, cells were harvested by
centrifugation at 480g for 20 min at 4°C. The cell pellet was resus-
pended in 50 ml of lysis buffer A [25 mM Hepes-­NaOH (pH 7.6),
300 mM NaCl, 10% (v/v) glycerol, and 2 mM 2-­Mercaptoethanol] sup-
plemented with 10 mM imidazole, 0.1 mM phenylmethylsulfonyl flu-
oride (PMSF), and 1× protease inhibitor cocktail (Roche cOmplete,
EDTA free). Cells were disrupted by sonication on ice (four pulses of
30 s, 60% power, 70% cycle), and the lysate was centrifuged for 15 min
at 3200g, 4°C, followed by ultracentrifugation at 100,000g for 1 hour.
The supernatant was loaded onto a 5-­ml HisTrap FF column (GE
HealthCare) equilibrated with buffer A/10 mM imidazole. After bind-
ing, the column was washed with 75 ml of buffer A/20 mM imidazole
followed by 75 ml of buffer A/40 mM imidazole. MutLβ protein was
eluted with buffer A/250 mM imidazole. Peak fractions with MutLβ,
determined by SDS-­PAGE, were pooled, diluted 1.2× with buffer B
[25 mM Hepes-­NaOH (pH 7.6), 0.1 mM EDTA, 10% (v/v) glycerol,
and 1 mM dithiothreitol (DTT)] to a final concentration of NaCl of
250 mM. The sample was loaded onto a 5-­ml HiTrap Heparin column
(GE HealthCare) equilibrated with buffer B/250 mM NaCl. The col-
umn was washed with 100 ml of buffer B/250 mM NaCl. Bound pro-
teins were eluted with a 50-­ml linear gradient of NaCl from 250 to
800 mM. Peak fractions were pooled, diluted 3× with buffer B, and
loaded onto a 1-­ml Mono Q 5/50 GL column (GE HealthCare)
equilibrated with buffer B/150 mM NaCl. The column was washed
with 15 ml of buffer B/150 mM NaCl. Bound proteins were eluted
with a 20-­ml linear gradient of 150 to 550 mM NaCl in buffer B. Peak
13 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
fractions were pooled; dialyzed against 25 mM Hepes-­NaOH (pH 7.5),
20% (v/v) glycerol, 0.1 mM EDTA 110 mM NaCl, and 0.1 mM PMSF;
aliquoted; and stored at −80°C.
FANCJ was affinity purified via its N-­terminal Flag-­tag using
anti-­FLAG M2 Magnetic beads (Sigma-­Aldrich). Approximately
5 × 108 cells were infected with freshly amplified recombinant baculo-
viruses encoding for N-­terminally Flag-­tagged FANCJ. Forty-­
eight hours after infection, the cells were harvested by centrifugation at
480g for 20 min at 4°C. The pellet was resuspended in 4× pellet vol-
ume (cca. 20 ml) of lysis/binding buffer C [10 mM tris-­HCl (pH 7.4),
130 mM NaCl, 1.0% (v/v) Triton X-­100, 10 mM NaF, 10 mM
NaH2PO4, 10 mM Na4P2O7, and 10% (v/v) glycerol] supplemented
with 0.1 mM PMSF and 1× protease inhibitor cocktail (Roche cOm-
plete, EDTA free). Cells were lysed for 60 min at 4°C on a rotary
shaker. The lysate was centrifuged at 4°C for 15 min at 3200g, fol-
lowed by 45 min at 20,000g. The supernatant was mixed with 640 ml
of slurry anti-­Flag M2 magnetic beads equilibrated with buffer C
followed by incubation for 2 hours at 4°C on a rotary shaker. After
binding, the beads were separated from the supernatant with un-
bound proteins by using a magnetic rack. The resin was washed three
times with 30 ml of buffer D [50 mM tris-­HCl (pH 7.4), 0.5% (v/v)
NP-­40, and 10% (v/v) glycerol] supplemented with 500 mM NaCl
and 0.1 mM PMSF, followed by one wash in buffer D supplemented
with 150 mM NaCl and 0.1 mM PMSF. After quantitative removal of
all wash buffer, the beads were eluted three times, each with 1 ml of
elution buffer E [25 mM tris-­HCl (pH7.4), 100 mM NaCl, 0.1% (v/v/)
Tween 20, 10% (v/v) glycerol, 0.1 mM PMSF, 2 mM 2-­mercaptoethanol,
and 3xFlag-­peptide (150 μg/ml)]. Concentration of purified proteins
was determined by densitometric analysis of SDS-­PAGE gels cali-
brated by BSA.
His pull-­down assay
Purified MutLβ protein carrying a 6xHis tag on the PMS1 subunit
(1 μg) was incubated with purified Flag-­tagged FANCJ (0.5 μg) in
a buffer (500 μl) containing 50 mM tris-­HCl (pH 7.5), 0.1% (v/v)
NP-­40, 10% (v/v) glycerol, 150 mM NaCl, 1 mM 2-­mercaptoethanol,
and 10 mM imidazole for 2 hours at 4°C. Subsequently, Ni-­NTA
beads (20 μl; QIAGEN) were added and incubation was continued
for 2 hours at 4°C. Beads were washed three times with the above
buffer. The bound proteins were eluted with 200 mM imidazole
(50 μl) and analyzed by Western blotting using anti-­MLH1 and
anti-­FANCJ antibodies.
Biotin pull-­down assay
To form a G4 structure substrate, 5′-­biotin–labeled b66-­G4 oligo-
nucleotide (5′-­Biotin-­TTTTTTTTTTTTTTTTTTTTTTTTTTTT
TTTTTTTTTTTTTTTTTGGGTGGGTGGGTGGGTTTTTT-­3′)
was diluted to a concentration of 1 μM in NEBuffer 4 (1×) and then
heat-­denatured at 95°C for 5 min followed by 20-­min incubation at
85°C and slow cooling to RT. Binding reactions (50 μl) were per-
formed in a buffer containing 50 mM tris-­HCl (pH 7.5), 120 mM
KCl, and 0.1% (v/v) Triton X-­100. Initially, b66-­G4 at a concen-
tration of 20 nM was incubated with streptavidin-­coupled M280
Dynabeads (10 μl; Life Technologies) for 15 min at RT. Where indi-
cated, RPA (40 nM) was added to the mixture and incubation was
carried out in a rotary shaker for 15 min at 4°C to coat the single-­
stranded part of b66-­G4. Mixtures were then supplemented with
50 nM wild-­type or Y245S/K246E mutant MutSβ, and incubation was
continued for 45 min at 4°C. Last, the beads were washed three
Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024
times with the binding buffer, and bound proteins were analyzed
by Western blotting.
Electrophoretic mobility shift assay
Cy3-­labeled CEB1 G4 oligonucleotide (5′-Cy3-AGGGGGGAGG­
GAGGGTGG-­3′) was folded into G4 structure in 100 mM KCl by
heating at 95°C for 5 min, followed by slow cooling to RT. The folded
oligonucleotide was incubated at a concentration of 30 nM with in-
creasing concentrations of MutSβ (wild type or Y245S/K246E mu-
tant) in 20 mM tris-­HCl (pH 7.5), 5 mM MgCl2, 1 mM DTT, 75 mM
NaCl, and BSA (0.05 mg/ml) at 37°C for 10 min. The assembled
complexes were separated from free oligonucleotide on 0.8% aga-
rose gel in 1× TBE (Tris-­Borate-­EDTA) buffer supplemented with
10 mM KCl at 70 V for 35 min at 4°C. Gels were scanned on an
Amersham Typhoon Biomolecular Imager scanner (Cytiva) and
quantified with Multi Gauge version 3.2 (Fujifilm).
Circular dichroism spectroscopy
Circular dichroism spectra of 7.5 μM CEB1 G4 oligonucleotide in
10.5 mM KCl were collected from 200 to 330 nm with a spectral
bandwidth of 1 nm using a Chirascan CD Spectrometer. The spec-
trum shown is an average of three technical replicates.
Statistical analysis
Statistical analysis was performed with GraphPad Prism 9 soft-
ware using either paired Student’s t test, repeated-­measures one-­
way analysis of variance (ANOVA) with Tukey’s multiple comparisons
correction, two-­tailed nonparametric Kruskal-­Wallis test followed
by Dunn’s multiple comparisons test, or two-­t ailed nonparamet-
ric Mann-­Whitney test, where appropriate. Details of statistics
in each experiment can be found in the corresponding fig-
ure legends.
Supplementary Materials
This PDF file includes:
Figs. S1 to S7
Table S1
REFERENCES AND NOTES
1. M. K. Zeman, K. A. Cimprich, Causes and consequences of replication stress. Nat. Cell Biol.
16, 2–9 (2014).
2. T. Garcia-­Muse, A. Aguilera, Transcription-­replication conflicts: How they occur and how
they are resolved. Nat. Rev. Mol. Cell Biol. 17, 553–563 (2016).
3. M. Macheret, T. D. Halazonetis, Intragenic origins due to short G1 phases underlie
oncogene-­induced DNA replication stress. Nature 555, 112–116 (2018).
4. S. Hamperl, M. J. Bocek, J. C. Saldivar, T. Swigut, K. A. Cimprich, Transcription-­replication
conflict orientation modulates R-loop levels and activates distinct DNA damage
responses. Cell 170, 774–786.e19 (2017).
5. K. S. Lang, A. N. Hall, C. N. Merrikh, M. Ragheb, H. Tabakh, A. J. Pollock, J. J.
Woodward, J. E. Dreifus, H. Merrikh, Replication-­transcription conflicts generate
R-­Loops that orchestrate bacterial stress survival and pathogenesis. Cell 170,
787–799.e18 (2017).
6. N. Chappidi, Z. Nascakova, B. Boleslavska, R. Zellweger, E. Isik, M. Andrs, S. Menon,
J. Dobrovolna, C. Balbo Pogliano, J. Matos, A. Porro, M. Lopes, P. Janscak, Fork
cleavage-­religation cycle and active transcription mediate replication restart after fork
stalling at co-­transcriptional R-­loops. Mol. Cell 77, 528–541.e8 (2020).
7. J. R. Brickner, J. L. Garzon, K. A. Cimprich, Walking a tightrope: The complex balancing act
of R-­loops in genome stability. Mol. Cell 82, 2267–2297 (2022).
8. S. Hamperl, K. A. Cimprich, The contribution of co-­transcriptional RNA: DNA hybrid
structures to DNA damage and genome instability. DNA Repair 19, 84–94 (2014).
9. M. Berti, D. Cortez, M. Lopes, The plasticity of DNA replication forks in response to
clinically relevant genotoxic stress. Nat. Rev. Mol. Cell Biol. 21, 633–651 (2020).
14 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
10. D. Castillo-­Guzman, F. Chédin, Defining R-­loop classes and their contributions to genome
instability. DNA Repair 106, 103182 (2021).
11. A. De Magis, S. G. Manzo, M. Russo, J. Marinello, R. Morigi, O. Sordet, G. Capranico,
DNA damage and genome instability by G-­quadruplex ligands are mediated by R
loops in human cancer cells. Proc. Natl. Acad. Sci. U.S.A. 116, 816–825 (2019).
12. M. L. Duquette, P. Handa, J. A. Vincent, A. F. Taylor, N. Maizels, Intracellular transcription of
G-­rich DNAs induces formation of G-­loops, novel structures containing G4 DNA. Genes
Dev. 18, 1618–1629 (2004).
13. P. Castillo Bosch, S. Segura-­Bayona, W. Koole, J. T. Heteren, J. M. Dewar, M. Tijsterman,
P. Knipscheer, FANCJ promotes DNA synthesis through G-­quadruplex structures. EMBO J.
33, 2521–2533 (2014).
14. O. Mendoza, A. Bourdoncle, J. B. Boulé, R. M. Brosh Jr., J. L. Mergny, G-­quadruplexes
and helicases. Nucleic Acids Res. 44, 1989–2006 (2016).
15. D. Sakellariou, S. T. Bak, E. Isik, S. I. Barroso, A. Porro, A. Aguilera, J. Bartek, P. Janscak,
J. Peña-­Diaz, MutSβ regulates G4-­associated telomeric R-­loops to maintain telomere
integrity in ALT cancer cells. Cell Rep. 39, 110602 (2022).
16. J. Jiricny, The multifaceted mismatch-­repair system. Nat. Rev. Mol. Cell Biol. 7, 335–346
(2006).
17. Y. Habraken, P. Sung, L. Prakash, S. Prakash, Binding of insertion/deletion DNA
mismatches by the heterodimer of yeast mismatch repair proteins MSH2 and MSH3.
Curr. Biol. 6, 1185–1187 (1996).
18. P. Modrich, Mechanisms in eukaryotic mismatch repair. J. Biol. Chem. 281,
30305–30309 (2006).
19. F. A. Kadyrov, L. Dzantiev, N. Constantin, P. Modrich, Endonucleolytic function of
mutlα in human mismatch repair. Cell 126, 297–308 (2006).
20. M. Räschle, G. Marra, M. Nyström-­Lahti, P. Schär, J. Jiricny, Identification of hMutLβ,
a heterodimer of hMLH1 and hPMS1. J. Biol. Chem. 274, 32368–32375 (1999).
21. E. Cannavo, G. Marra, J. Sabates-­Bellver, M. Menigatti, S. M. Lipkin, F. Fischer, P. Cejka,
J. Jiricny, Expression of the MutL homologue hMLH3 in human cells and its role in
DNA mismatch repair. Cancer Res. 65, 10759–10766 (2005).
22. S. M. Lipkin, P. B. Moens, V. Wang, M. Lenzi, D. Shanmugarajah, A. Gilgeous,
J. Thomas, J. Cheng, J. W. Touchman, E. D. Green, P. Schwartzberg, F. S. Collins, P. E.
Cohen, Meiotic arrest and aneuploidy in MLH3-­deficient mice. Nat. Genet. 31,
385–390 (2002).
23. R. M. Pinto, E. Dragileva, A. Kirby, A. Lloret, E. Lopez, J. St. Claire, G. B. Panigrahi,
C. Hou, K. Holloway, T. Gillis, J. R. Guide, P. E. Cohen, G. M. Li, C. E. Pearson, M. J. Daly,
V. C. Wheeler, Mismatch repair genes Mlh1 and Mlh3 modify CAG instability in
huntington’s disease mice: Genome-­wide and candidate approaches. PLOS Genet. 9,
e1003930 (2013).
24. X. Zhao, Y. Zhang, K. Wilkins, W. Edelmann, K. Usdin, MutLγ promotes repeat
expansion in a Fragile X mouse model while EXO1 is protective. PLOS Genet. 14,
e1007719 (2018).
25. F. Teloni, J. Michelena, A. Lezaja, S. Kilic, C. Ambrosi, S. Menon, J. Dobrovolna,
R. Imhof, P. Janscak, T. Baubec, M. Altmeyer, Efficient pre-­mRNA cleavage prevents
replication-­stress-­associated genome instability. Mol. Cell 73, 670–683.e12 (2019).
26. M. Bauer, Z. Nascakova, A. I. Mihai, P. F. Cheng, M. P. Levesque, S. Lampart, R. Hurwitz,
L. Pfannkuch, J. Dobrovolna, M. Jacobs, S. Bartfeld, A. Dohlman, X. Shen, A. A. Gall,
N. R. Salama, A. Töpfer, A. Weber, T. F. Meyer, P. Janscak, A. Müller, The ALPK1/TIFA/NF-­κB axis
links a bacterial carcinogen to R-­loop-­induced replication stress. Nat. Commun. 11, 5117 (2020).
27. G. Lossaint, M. Larroque, C. Ribeyre, N. Bec, C. Larroque, C. Decaillet, K. Gari,
A. Constantinou, FANCD2 binds MCM proteins and controls replisome function
upon activation of s phase checkpoint signaling. Mol. Cell 51, 678–690 (2013).
28. S. Ying, S. Minocherhomji, K. L. Chan, T. Palmai-­Pallag, W. K. Chu, T. Wass, H. W.
Mankouri, Y. Liu, I. D. Hickson, MUS81 promotes common fragile site expression.
Nat. Cell Biol. 15, 1001–1007 (2013).
29. V. Naim, T. Wilhelm, M. Debatisse, F. Rosselli, ERCC1 and MUS81-­EME1 promote sister
chromatid separation by processing late replication intermediates at common
fragile sites during mitosis. Nat. Cell Biol. 15, 1008–1015 (2013).
30. M. Berti, A. R. Chaudhuri, S. Thangavel, S. Gomathinayagam, S. Kenig, M. Vujanovic,
F. Odreman, T. Glatter, S. Graziano, R. Mendoza-­M aldonado, F. Marino, B. Lucic,
V. Biasin, M. Gstaiger, R. Aebersold, J. M. Sidorova, R. J. Monnat, M. Lopes, A.
Vindigni, Human RECQ1 promotes restart of replication forks reversed by DNA
topoisomerase I inhibition. Nat. Struct. Mol. Biol. 20, 347–354 (2013).
31. A. Ray Chaudhuri, Y. Hashimoto, R. Herrador, K. J. Neelsen, D. Fachinetti, R. Bermejo,
A. Cocito, V. Costanzo, M. Lopes, Topoisomerase I poisoning results in PARP-­
mediated replication fork reversal. Nat. Struct. Mol. Biol. 19, 417–423 (2012).
32. P. Kotsantis, S. Segura-­Bayona, P. Margalef, P. Marzec, P. Ruis, G. Hewitt, R. Bellelli,
H. Patel, R. Goldstone, A. R. Poetsch, S. J. Boulton, RTEL1 regulates G4/R-­loops to
avert replication-­transcription collisions. Cell Rep. 33, 108546 (2020).
33. S. Gupta, M. Gellert, W. Yang, Mechanism of mismatch recognition revealed by
human MutSβ bound to unpaired DNA loops. Nat. Struct. Mol. Biol. 19, 72–78
(2012).
Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024
34. K. Burdova, B. Mihaljevic, A. Sturzenegger, N. Chappidi, P. Janscak, The mismatch-­binding
factor MutSβ can mediate ATR activation in response to DNA double-­strand breaks. Mol.
Cell 59, 603–614 (2015).
35. J. Johnson, R. Okyere, A. Joseph, K. Musier-­Forsyth, B. Kankia, Quadruplex formation
as a molecular switch to turn on intrinsically fluorescent nucleotide analogs. Nucleic
Acids Res. 41, 220–228 (2013).
36. S. Barroso, E. Herrera-­M oyano, S. Muñoz, M. García-­Rubio, B. Gómez-­G onzález,
A. Aguilera, The DNA damage response acts as a safeguard against harmful DNA–
RNA hybrids of different origins. EMBO Rep. 20, 47250 (2019).
37. E. Cannavo, B. Gerrits, G. Marra, R. Schlapbach, J. Jiricny, Characterization of the
interactome of the human MutL homologues MLH1, PMS1, and PMS2. J. Biol. Chem. 282,
2976–2986 (2007).
38. M. Peng, R. Litman, J. Xie, S. Sharma, R. M. Brosh Jr., S. B. Cantor, The FANCJ/MutLα
interaction is required for correction of the cross-­link response in FA-­J cells. EMBO J. 26,
3238–3249 (2007).
39. D. C. Odermatt, W. T. C. Lee, S. Wild, S. K. Jozwiakowski, E. Rothenberg, K. Gari,
Cancer-­associated mutations in the ironsulfur domain of FANCJ affect G-­quadruplex
metabolism. PLOS Genet. 16, 1008740 (2020).
40. Y. Wu, K. Shin-­ya, R. M. Brosh Jr., FANCJ Helicase defective in fanconia anemia and Breast
Cancer unwinds G-­Quadruplex DNA to defend genomic stability. Mol. Cell. Biol. 28,
4116–4128 (2008).
41. A. N. Suhasini, J. A. Sommers, P. A. Muniandy, Y. Coulombe, S. B. Cantor, J.-­Y. Masson,
M. M. Seidman, R. M. Brosh Jr., Fanconi anemia group J helicase and MRE11
nuclease interact to facilitate the DNA damage response. Mol. Cell. Biol. 33,
2212–2227 (2013).
42. S. Guillemette, A. Branagan, M. Peng, A. Dhruva, O. D. Scharer, S. B. Cantor, FANCJ
localization by mismatch repair is vital to maintain genomic integrity after UV irradiation.
Cancer Res. 74, 932–944 (2014).
43. B. Boleslavska, A. Oravetzova, K. Shukla, Z. Nascakova, O. N. Ibini, Z. Hasanova, M. Andrs,
R. Kanagaraj, J. Dobrovolna, P. Janscak, DDX17 helicase promotes resolution of
R-­loop-­mediated transcription-­replication conflicts in human cells. Nucleic Acids Res. 50,
12274–12290 (2022).
44. P. Linder, E. Jankowsky, From unwinding to clamping -­the DEAD box RNA helicase family.
Nat. Rev. Mol. Cell Biol. 12, 505–516 (2011).
45. C. G. Wu, M. Spies, G-­quadruplex recognition and remodeling by the FANCJ helicase.
Nucleic Acids Res. 44, 8742–8753 (2016).
46. M. Peng, K. Cong, N. J. Panzarino, S. Nayak, J. Calvo, B. Deng, L. J. Zhu, M. Morocz,
L. Hegedus, L. Haracska, S. B. Cantor, Opposing roles of FANCJ and HLTF protect forks and
restrain replication during stress. Cell Rep. 24, 3251–3261 (2018).
47. M. Andrs, H. Stoy, B. Boleslavska, N. Chappidi, R. Kanagaraj, Z. Nascakova, S. Menon,
S. Rao, A. Oravetzova, J. Dobrovolna, K. Surendranath, M. Lopes, P. Janscak, Excessive
reactive oxygen species induce transcription-­dependent replication stress. Nat. Commun.
14, 1791 (2023).
48. W. T. C. Lee, Y. Yin, M. J. Morten, P. Tonzi, P. P. Gwo, D. C. Odermatt, M. Modesti, S. B. Cantor,
K. Gari, T. T. Huang, E. Rothenberg, Single-­molecule imaging reveals replication fork
coupled formation of G-­quadruplex structures hinders local replication stress signaling.
Nat. Commun. 12, 2525 (2021).
49. M. L. García-­Rubio, C. Pérez-­Calero, S. I. Barroso, E. Tumini, E. Herrera-­Moyano, I. V. Rosado,
A. Aguilera, The Fanconi anemia pathway protects genome integrity from R-­loops. PLOS
Genet. 11, e1005674 (2015).
50. M. Vujanovic, J. Krietsch, M. C. Raso, N. Terraneo, R. Zellweger, J. A. Schmid,
A. Taglialatela, J. W. Huang, C. L. Holland, K. Zwicky, R. Herrador, H. Jacobs, D. Cortez,
A. Ciccia, L. Penengo, M. Lopes, Replication Fork slowing and reversal upon DNA
damage require PCNA polyubiquitination and ZRANB3 DNA translocase activity. Mol.
Cell 67, 882–890.e5 (2017).
51. M. M. Ghodgaonkar, P. Kehl, I. Ventura, L. Hu, M. Bignami, J. Jiricny, Phenotypic
characterization of missense polymerase-­δ mutations using an inducible protein-­
replacement system. Nat. Commun. 5, 4990 (2014).
52. A. Tighe, O. Staples, S. Taylor, Mps1 kinase activity restrains anaphase during an
unperturbed mitosis and targets Mad2 to kinetochores. J. Cell Biol. 181, 893–901
(2008).
53. L. A. Henricksen, C. B. Umbricht, M. S. Wold, Recombinant replication protein A:
Expression, complex formation, and functional characterization. J. Biol. Chem. 269,
11121–11132 (1994).
Acknowledgments: We thank S. Cantor for FANCJ constructs used in our preliminary
experiments, K. Gari for FANCJ−/− HeLa FIT cells and pDONR221-­based vectors expressing
either wild-­type or K52R variant of FANCJ, and J. Jiricny for stimulating discussions and
comments on the manuscript. Funding: This work was supported by grants from the Swiss
National Science Foundation (310030_184716 and 310030_214846), the Czech Science
Foundation (22-­08294S), the Cancer League of the Canton Zurich, and the Foundation for
Research in Science and the Humanities at the University of Zurich to P.J. L.K. was supported by
15 of 16
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

the Czech Science Foundation (21-­22593X, with J.D. as co-­principal investigator), Wellcome and/or the Supplementary Materials. Requests for reagents may be directed to and will be
Trust Collaborative Grant (206292/E/17/Z), and the European Union’s Horizon 2020 research fulfilled by P.J. (pjanscak@​imcr.​uzh.​ch).
and innovation programme under grant agreement no. 812829. Author contributions: P.J.
conceived the project. E.I., K.S., M.P., L.K., and P.J. designed experiments. E.I., K.S., M.P., C.K., S.R.,
M.A., J.D., and V.R. performed experiments and analyzed the data. L.K. and P.J. supervised the Submitted 11 August 2023
research. E.I. and P.J. wrote the manuscript with inputs from L.K., K.S., and M.P. Competing Accepted 8 January 2024
interests: The authors declare that they have no competing interests. Data and materials Published 7 February 2024
availability: All data needed to evaluate the conclusions in the paper are present in the paper 10.1126/sciadv.adk2685

Isik et al., Sci. Adv. 10, eadk2685 (2024) 7 February 2024 16 of 16

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

R E G E N E R AT I O N Copyright © 2024 The

Authors, some rights
Aligned cryogel fibers incorporated 3D printed scaffold reserved; exclusive
licensee American
effectively facilitates bone regeneration by enhancing Association for the
Advancement of
cell recruitment and function Science. No claim to
original U.S.
Government Works.
Yuxuan Wei1,2†, Hao Pan2,3†, Jianqiu Yang2,4, Canjun Zeng1*, Wenbing Wan4*, Shixuan Chen2* Distributed under a
Creative Commons
Reconstructing extensive cranial defects represents a persistent clinical challenge. Here, we reported a hybrid Attribution
three-­dimensional (3D) printed scaffold with modification of QK peptide and KP peptide for effectively promoting NonCommercial
endogenous cranial bone regeneration. The hybrid 3D printed scaffold consists of vertically aligned cryogel fibers License 4.0 (CC BY-­NC).
that guide and promote cell penetration into the defect area in the early stages of bone repair. Then, the conjugated
QK peptide and KP peptide further regulate the function of the recruited cells to promote vascularization and
osteogenic differentiation in the defect area. The regenerated bone volume and surface coverage of the dual
peptide-­modified hybrid scaffold were significantly higher than the positive control group. In addition, the dual
peptide-­modified hybrid scaffold demonstrated sustained enhancement of bone regeneration and avoidance of
bone resorption compared to the collagen sponge group. We expect that the design of dual peptide-­modified
hybrid scaffold will provide a promising strategy for bone regeneration.

INTRODUCTION
Now, the most commonly used clinical solutions for large cranial
bone defects include the use of autologous vascularized bone graft
(1, 2), polyetheretherketone (PEEK) cranial implant (3), and titanium/
titanium alloy mesh (4, 5). However, these modalities have well-­
established limitations after years of clinical use. The autologous
vascularized bone graft remains the clinical gold standard for recon-
structing large cranial defects. However, several drawbacks exist.
For instance, the mismatch between the graft shape and cranial defect
contours can lead to poor aesthetic outcomes. Autologous grafts
undergo substantial resorption over time, resulting in contour irregu-
larities and the need for secondary renovation (1, 2). In addition, the
donor site also suffers excruciating pain and infection risks. Although
synthetic PEEK implants and titanium meshes can avoid donor site
morbidity, they carry a high complication rate. Exposure and infection
frequently necessitate removal (6), and their rigidity causes discom-
fort (7). Moreover, the limited osseointegration caused by nondeg-
radation and dense structure can cause implant failure (8). Therefore,
there is an unmet need for resorbable cranial bone grafts that can
preclude donor site morbidity and disadvantages associated with al-
lografts and existing cranioplasty materials.
Collagen sponge is a commonly used biodegradable bone repair
material, and it has been widely used as osteoconductive bone graft
substitutes to promote bone regeneration in clinics (9, 10). The
highly porous structure allows cellular infiltration, vascularization,
and bone regeneration (11). However, collagen sponges lack inherent
osteoinductivity (9, 12). Moreover, the degradation of collagen sponge is
1
Department of Foot and Ankle Surgery, Center for Orthopaedic Surgery, the Third
Affiliated Hospital of Southern Medical University. Guangzhou, Guangdong 510630,
China. 2Zhejiang Engineering Research Center for Tissue Repair Materials, Wenzhou
Institute, University of Chinese Academy of Sciences, Wenzhou, Zhejiang 325000,
China. 3Department of Orthopaedic Surgery, the First Affiliated Hospital of Wenzhou
Medical University, Wenzhou, Zhejiang 325000, China. 4Department of Orthopaedic
Surgery, the Second Affiliated Hospital, Jiangxi Medical College, Nanchang University,
Nanchang, Jiangxi, 330006 China.
*Corresponding author. Email: zengcanjun@​163.​com (C.Z.); ndefy15220@​ncu.​edu.​
cn (W.W.); chensx@​ucas.​ac.​cn (S.C.)
†These authors contributed equally to this work.
before sufficient new bone formation, resulting in new bone formation
that cannot be continuously induced, and the formed osteoid tissue
may be resorbed. Ongoing challenges of biodegradable bone repair
material include balancing degradation with bone ingrowth (13, 14),
improving mechanical properties (15, 16), and incorporating bioac-
tive factors to enhance overall osteogenic performance (17). In our
previous study, we developed a class of cell-­and therapeutic agent–
free scaffolds made of polycaprolactone electrospun nanofibers with
predesigned three-­dimensional (3D) aligned nanotopography (18)
based on fully considering the advantages and disadvantages of collagen
sponge bone repair materials, which have higher porosity, longer degra-
dation time, and can enable rapid cell recruitment either from the sur-
rounding to the center or from the bottom to the top, resulting in rapid
bone regeneration (19). It inspires us with the importance of cell recruit-
ment in the early stages of bone repair.
Among various additive manufacturing processing technologies,
3D printing is the most widely used technology for preparing bone
repair materials. However, the 3D printed scaffolds (PSs) have low
cell seeding efficiency and cannot induce cell migration. Here, in the
present study, we hypothesized that introducing aligned fibers within
the 3D PS may further accelerate cell infiltration into the defect area
(18, 20), thereby boosting tissue regeneration. Here, one purpose of
this work is to study how to introduce aligned fibers inside the 3D
PS (Fig. 1A). Another issue that needs to be solved is how to regulate
the function of these recruited cells, such as enhancing cells’ osteo-
inductive ability (Fig. 1B). To regulate the cell behaviors of the
recruited cells, the QK peptide and KP peptide were used to modify
the aligned fibers and 3D PS, respectively. The QK peptide is an engi-
neered vascular endothelial growth factor (VEGF)–mimicking peptide,
which can enhance angiogenesis to provide more nutrients for these
recruited cells during bone healing (21, 22). The KP peptide is a
bone morphogenetic protein 2 (BMP-­2) mimicking peptide, which
can significantly enhance new bone formation (22–24). Therefore,
the introduction of aligned fibers within the 3D PS and the cell recruit-
ment ability of aligned fiber-­modified 3D PS were explored. In addition,
the pro-­angiogenesis and pro-­ossification ability of the QK peptide
and KP peptide conjugated hybrid scaffold were also investigated in
this study.

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 1 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
Fig. 1. ACFs incorporated 3D PS effectively facilitates bone regeneration by enhancing cell recruitment and function. (A) Fabrication procedure of ACFs incorporated
into a 3D PS and cell recruitment methods of PS/ACFs in the cranial bone defect. (B) PS/ACFs functionalized with KP and QK peptides facilitate critical-­sized cranial bone
defect repair in rats. Red arrows indicate the cell recruitment, blue arrows show the orientation of the newly formed collagenous fibers, and orange arrows indicate the
osteogenic differentiation. BMSC, bone marrow mesenchymal stromal cell.
RESULTS
Fabrication and characterization of ACFs incorporated 3D PS
The aligned cryogel fibers (ACFs) incorporated a 3D PS, which was
prepared by 3D bioprinting in combination with unidirectional freeze
casting. Briefly, a 3D lattice scaffold was printed using 5 weight %
(wt %) gelatin-­methacrylamide (Gel-­MA). The lyophilized PS was sub-
merged into a solution containing 0.18 wt % Gel-­MA and 0.02 wt %
gelatin for cryogel fiber formation. Then, the soaked PS was trans-
ferred to a − 80°C precooling stainless steel plate for unidirectional
freezing and chemical crosslinking over 24 hours (Fig. 2A). The PS
exhibited a colorless, transparent lattice structure, while the PS/
ACFs showed a similar morphology with some air bubbles within
the lattice holes. After lyophilization, the holes of PS/ACFs were
filled with white fibers in contrast to the PS scaffold (Fig. 2B). The
presence of air bubbles in PS/ACFs before freeze drying corresponded
to the formation of ACFs within the pores. The preparation of random
cryogel fibers (RCFs) incorporating a 3D PS was similar to PS/ACFs
but without unidirectional freeze-­casting (fig. S1A).
Scanning electron microscopy (SEM) images further found that
the holes of PS/ACFs were filled with fibers (top view), and these
fibers exhibited a vertically aligned structure (cross section) when
compared with unmodified PS (Fig. 2C). The PS/RCFs were filled
almost entirely with horizontal fibers and lamellas (fig. S1B). The
false color images (Fig. 2D) and the distribution of directionality
analysis (Fig. 2E) demonstrated that the formed ACFs had highly
coherent orientation, and the median fibers’ angles were nearly 90°
(Fig. 2E). The formed RCFs were random and almost perpendicular
to the wall of the holes (fig. S1, C and D). In addition, the diameter
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
of ACFs ranged from 1 to 5 μm, with a mean of 2.46 ± 0.86 μm
(Fig. 2F).
Moreover, the Young’s modulus of wet PS/ACFs was higher than
that of PS (fig. S2, A and B). The maximum strain of wet PS/ACFs
was slightly lower than that of PS, and there was no difference in the
maximum stress between the two groups (fig. S2, C and D). Both
lyophilized PS and PS/ACFs could reach swelling equilibrium with-
in 1 hour in phosphate-­buffered saline (PBS) buffer at 37°C (fig. S3A).
In vitro enzymatic degradation experiment showed that both PS and
PS/ACFs could be rapidly and completely degraded (within 3 hours) in
0.05% collagenase solution (fig. S3B).
In vitro cell seeding efficiency and in vivo cell recruitment
ability of PS/ACFs
As shown in Fig. 3A (i), human umbilical cord endothelial cells
(HUVECs) suspension was dropped to the top surface of the rehydrated
scaffold and incubated for 2 hours, and then the HUVECs seeded
PS, PS/RCFs, and PS/ACFs were transferred to fresh medium to cul-
ture another 24 hours. The 3D reconstructed confocal images showed
that HUVECs only adhered to the surface of the PS (depth less than
200 μm), and there were no cells in the hole area. In contrast, HUVECs
could adhere to both PS and ACFs/RCFs areas of PS/RCFs and PS/
ACFs. In addition, the cells were not only distributed on the surface
but also migrated downward to the bottom of the PS/RCFs (about
300 μm in depth) and the PS/ACFs (about 600 μm in depth)
(Fig. 3B). SEM images further visualized that HUVECs preferred to
attach to the RCFs/ACFs area rather than 3D printed Gel-­MA lines
(Fig. 3C). The Cell Counting Kit-­8 (CCK-­8) assay indicated that the
2 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 2. Morphological features of 3D printed scaffold embedded without/with ACFs. (A) Schematic illustrating the preparation procedures of ACFs incorporated 3D
PS by 3D printing in combination with freeze-­casting. (B) Gross view of PS and PS/ACFs before and after lyophilization. (C) SEM images showing the top view and cross
section of the PS and PS/ACFs. (D) False color images of ACFs indicate the angle mapping of fiber orientations. (E) Angle distribution of formed cryogel fibers. (F) Distribution
of cryogel fibers’ diameter. Oct, optimal cutting temperature compound; UV, ultraviolet.

PS/ACFs and PS/RCFs had twice as attached cells as PS after 24 hours
of culture (Fig. 3D).
To further explore the in vivo cell recruitment ability of PS/ACFs,
it was implanted subcutaneously in rats (Fig. 3A, ii). Hematoxylin
and eosin (H&E) images revealed that more infiltrated cells were
found in the gap of printed Gel-­MA lines of PS/ACFs compared with
PS after 1 and 2 weeks of implantation. The density of newly formed
tissue in the ACFs area of PS/ACFs is significantly denser than the
empty area of PS, and no fibrous capsules formed around the PS and
PS/ACFs (Fig. 3, E to G). In addition, we also explored the extracellular
matrix (ECM) morphology of newly formed tissue by trichrome
staining. Trichrome staining images found that the orientation of
ECM fibers of the PS group was randomly arranged. In contrast, it
exhibited an orderly arrangement in the PS/ACFs group (Fig. 3H),

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 3 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 3. The in vitro cell seeding efficiency and in vivo cell recruitment ability of the ACFs incorporated 3D PS. (A) (i) Schematic illustrating the in vitro cell seeding
and adhesion on the PS, PS/RCFs, and PS/ACFs. (ii) Schematic illustrating the in vivo cell recruitment and tissue ingrowth induced by PS/ACFs by a subcutaneous implan-
tation model in rats. (B and C) Confocal microscope 3D images and SEM images showing HUVECs adhesion and distribution on the PS, PS/RCFs, and PS/ACFs after 24 hours
of cell seeding. (D) Quantification of attached HUVECs on the PS, PS/RCFs, and PS/ACFs after 24 hours of culture. (E to G) H&E staining and semi-­quantitative analysis ex-
hibit the cell infiltration and tissue ingrowth of PS and PS/ACFs after 1 and 2 weeks of subcutaneous implantation. (H) Trichrome staining discovers the newly formed
collagenous fiber orientation of the ingrowth soft tissue in the PS and PS/ACFs groups after 2 weeks of operation. The false color images indicate the orientation of col-
lagenous fiber. (I) Angle distribution of the newly formed collagenous fibers after 2 weeks of operation. ***P < 0.001 and ****P < 0.0001. ns, not significant.

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 4 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
which consisted of the direction of ACFs. The false color images and
angle quantitative analysis further verified that the orientation of
collagen fibers in the PS/ACFs group was highly coherent compared
with those in the PS group (Fig. 3, H and I). Furthermore, we found
that the angiogenesis in the PS/ACFs group was better than that in
the PS group after 1 and 2 weeks of subcutaneous implantation (fig. S4).
Pro-­osteogenic properties of the KP
peptide–functionalized PS
To accelerate bone repair, we used KP peptide to modify the 3D PS
to promote the differentiation of recruited bone marrow mesenchymal
stromal cells (BMSCs) into osteoblasts. To verify the pro-­osteogenic
differentiation potential of the KP peptide–conjugated PS (PS-­KP),
three different concentrations of KP peptide (250, 500, and 750 μg/ml)–
modified PS scaffold were prepared (fig. S5), and the in vitro and
in vivo pro-­osteogenic ability of different PS-­KP were explored.
Alkaline phosphatase (ALP) staining result showed that the rat BMSCs
(rBMSCs) treated with PS-­KP250, PS-­KP500, and PS-­KP750 groups
had stronger ALP expression than that in the PS-­treated group after
7 days of coculture (fig. S6, A and B). The relative expression of ALP,
osteopontin (OPN), and osteocalcin (OCN) in the PS-­KP–treated
rBMSCs was increased compared with PS-­treated rBMSCs (fig. S6,
C to E). The in vitro results suggested that PS-­KP could promote
osteogenic differentiation of rBMSCs.
We also explored the pro-­osteogenic effects of PS-­KP using a
subcutaneous ectopic osteogenesis model in nude mice. The suspension
of rBMSCs was dropped onto the surface of the scaffolds and cocultured
for 24 hours, and then subcutaneous implantation of rBMSCs loaded
PS-­KP in nude mice (Fig. 4A). The micro-­computed tomography
(Micro-­CT) 3D reconstruction images of the isolated scaffold tissue
showed that there was apparent new bone formation in the PS-­KP500
group after 4 weeks of surgery, while there was almost no new bone
formation in the PS, PS-­KP250, and PS-­KP750 groups (Fig. 4B). The
axial, sagittal, and coronal images showed the distribution and morpho
logy of regenerated bone. The regenerated bone in the PS-­KP500 group
showed a lattice shape consistent with PS morphology (Fig. 4C). The
quantitative analysis of the regenerated bone volume (BV/TV, %),
bone surface (BS/TS, %), trabecular number (Tb.N, 1/100 mm), and
thickness (Tb.Th, μm) of the PS-­KP500 group was definitely in-
creased compared to the PS, PS-­KP250, and PS-­KP750 groups
(Fig. 4, D to G). Trichrome (Fig. 4H) and H&E staining (fig. S7)
showed that the PS-­KP500 group had the most osteoid matrix dis-
tributed in clumps. The PS-­KP750 and PS-­KP250 groups had a
small number of osteoid matrices. Almost no osteoid matrix was
detected in the PS group. Immunohistochemical staining showed
the expression of OPN in the PS-­KP500 and PS-­KP750 groups was
higher than in the PS and PS-­KP250 groups. The expression of OPN
in the PS-­KP500 group was also higher than in the PS-­KP750 group
(Fig. 4, I and J). Therefore, we determined KP peptide (500 μg/ml)
as an effective concentration for modifying 3D printed Gel-­ MA
hydrogel.
Pro-­angiogenic properties of QK
peptide–functionalized PS/ACFs
As shown in Fig. 5A and fig. S8A, the QK peptide with different
concentrations (100, 300, and 500 μg/ml) was conjugated to the
ACFs, and the subcutaneous implantation of different PS/ACFs-­QK
was used to verify the pro-­angiogenic properties of PS/ACFs-­QK. The
SEM images, false color images, and angle distribution analysis
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
showed the vertical alignment of the cryogel fibers in the ACFs-­
QK100, ACFs-­QK300, and ACFs-­QK500 groups (fig. S8, B to D).
The subcutaneous implanted scaffold with surrounding tissue was
isolated 2 weeks after surgery. Trichrome staining results indicated
that the newly formed blood vessels were gradually increased with
the increased concentration of QK peptide from 0 to 500 μg/ml
(Fig. 5, B and C). Immunohistochemical staining and semi-­quantitative
analysis further demonstrated that the area of newly formed blood
vessels in the PS/ACFs-­QK300 and PS/ACFs-­QK500 groups was
increased compared to the PS and PS/ACFs-­QK100 groups. The
area of newly formed blood vessels in the PS/ACFs-­QK500 group
was higher than the PS/ACFs-­QK300 group (Fig. 5, D and E).
Therefore, we determined QK peptide (500 μg/ml) as an effective
concentration for modifying the ACFs.
The bone pro-­regenerative ability of KP peptide– and QK
peptide–functionalized PS/ACFs in rat critical-­sized cranial
bone defect model
To verify the potential bone pro-­regenerative ability of KP peptide–
and QK peptide–functionalized PS/ACFs, we implanted the PS-­
KP500/ACFs-­QK500 into the critical-­sized cranial bone defects
(8 mm in diameter) in rats, the commercial collagen sponge was used
as a positive control (Fig. 6A and fig. S9). The x-­ray images (Fig. 6B)
and reconstructed Micro-­CT images (Fig. 6, C and D) showed that
the PS-­KP500/ACFs-­QK500 and collagen sponge groups showed
significantly more new bone formation than the control group after
4 and 8 weeks of surgery. The quantitative analysis further revealed
no difference in the regenerated bone volume (BV/TV, %) and surface
coverage (%) between the PS-­KP500/ACFs-­QK500 and collagen
sponge groups after 4 weeks of treatment. While the regenerated
bone volume (BV/TV, %) and surface coverage (%) of the KP500/
ACFs-­QK500 group were higher than the collagen group after 8 weeks
of treatment (Fig. 6, E and F).
We further analyzed the microstructure of the regenerated bone
of each group. The coronal section of the 3D reconstructed Micro-­CT
images of all groups showed that the new bone was thicker at 8 weeks
than at 4 weeks. The regenerated bone volume and distribution in the
PS-­KP500/ACFs-­QK500 group achieved the expected healing effect
after 8 weeks of surgery (fig. S10A). The trabecular pattern factor
(Tb.Pf) is an index used to describe the degree of concavity of the
trabecular surface. A decrease of Tb.Pf indicated that the trabecula
changes from rod bone to plate bone, and the negative value of Tb.Pf
showed that the measured bone tissue is a cortical bone. The Tb.Pf
values of the control and PS-­KP500/ACFs-­QK500 groups did not
change over time, while the Tb.Pf value of the collagen group
changed from negative to positive (fig. S10B). The spatial morpho-
logical and structural indexes of bone trabeculae [including trabecular
separation (Tb.Sp), trabecular number (Tb.N), and trabecular thick-
ness (Tb.Th)] can indicate the changes of bone microstructure.
The collagen sponge and PS-­KP500/ACFs-­QK500 groups both
had lower Tb.Sp than control group at 4 weeks, and only PS-­KP500/
ACFs-­QK500 group had lower Tb.Sp than the control group at 8 weeks
(fig. S10C). Tb.N in the collagen sponge and PS-­KP500/ACFs-­
QK500 groups were higher than in the control group at 4 weeks after
surgery, and only PS-­KP500/ACFs-­QK500 group had higher Tb.N
than that in the control group at 8 weeks after surgery (fig. S10D).
After 4 weeks of surgery, Tb.Th in the PS-­KP500/ACFs-­QK500
group was higher than that of the collagen sponge and control groups,
with no difference among them at 8 weeks (fig. S10E). The several
5 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 4. Pro-­osteogenic properties of the KP peptide-­functionalized PS. (A) Schematic illustrating the preparation procedures and verification of pro-­osteogenic property of
the KP peptide-­functionalized PS. The ectopic osteogenesis model was used by subcutaneous implantation of rBMSCs-­seeding scaffolds in nude mice (n = 5). (B) Micro-­CT
3D reconstruction images exhibited ectopic new bone formation in the PS, PS-­KP250 (250 μg/ml), PS-­KP500 (500 μg/ml), and PS-­KP750 (750 μg/ml) group after 4 weeks
of operation. (C) Micro-­CT axial, sagittal, and coronal views of scaffold area after 4 weeks of operation. (D to G) Regenerated bone volume (BV/TV, %), bone surface (BS/TS, %),
Tb.N (1/100 mm), and Tb.Th (μm) after 4 weeks of operation. (H) Trichrome staining of the decalcified scaffold area and surrounding tissue of PS, PS-­KP250, PS-­KP500, and
PS-­KP750 groups after 4 weeks of operation. (I and J) OPN immunohistochemical staining of the decalcified scaffold area of PS, PS-­KP250, PS-­KP500, and PS-­KP750 groups
after 4 weeks of operation and semi-­quantitative analysis. *P < 0.05, **P < 0.01, ***P < 0.001, and ****P < 0.0001. IOD, integral optical density.

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 6 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 5. Pro-­angiogenic properties of QK peptide-­functionalized PS/ACFs. (A) Schematic illustrating the preparation procedures and verification of pro-­angiogenic
property of the QK peptide-­functionalized PS/ACFs. The subcutaneous implantation of QK peptide-­functionalized PS/ACFs in rats was used to evaluate its pro-­angiogenic
properties (n = 4). (B) Trichrome staining of the scaffold area and surrounding tissues of PS/ACFs, PS/ACFs-­QK100 (100 μg/ml), PS/ACFs-­QK300 (300 μg/ml), and PS/ACFs-­QK500
(500 μg/ml) groups after 2 weeks of operation. (C) Quantification of the number of newly formed blood vessels in the PS/ACFs, PS/ACFs-­QK100, PS/ACFs-­QK300, and
PS/ACFs-­QK500 groups after 2 weeks of operation. (D) CD31 immunohistochemical staining of the scaffold area and surrounding tissues in the PS/ACFs, PS/ACFs-­QK100,
PS/ACFs-­QK300, and PS/ACFs-­QK500 groups after 2 weeks of operation. (E) Quantification of the area of newly formed blood vessels in the PS/ACFs, PS/ACFs-­QK100, PS/
ACFs-­QK300, and PS/ACFs-­QK500 group after 2 weeks of operation. *P < 0.05 and ***P < 0.001.

indexes suggest that in the early stages of bone regeneration, PS-­K P500/
ACFs-­Q K500 could promote the formation of bone trabeculae
more quickly and maintain and promote the morphological and struc-
tural remodeling of trabecular bone favoring dense rod trabeculae.
The histological observations revealed that a more organic matrix
of regenerated bone was detected in the PS-­KP500/ACFs-­QK500
and collagen sponge groups after 4 weeks of treatment, which was
consistent with the reconstructed Micro-­CT images (Fig. 7, A and
B). We found that the newly formed collagen fibers in the defect area
of the control and collagen sponge groups were all parallel to the
skull surface (direction of the arrow). In the PS-­KP500/ACFs-­QK500
group, a part of vertically aligned collagen fibers was detected, which
was introduced by ACFs. Some decalcified organic bone matrix formed
within the vertically aligned collagen fibers (direction of the arrow)
(Fig. 7B). We further analyzed the arrangement direction of colla-
gen fibers located at the edge and middle of the wound. The collagen
tissues in the control and collagen sponge groups were horizon-
tally arranged, while the collagen fibers in the PS-­KP500/ACFs-­
QK500 group were arranged vertically (Fig. 7, C to F). After 8 weeks
of treatment, H&E and trichrome staining showed more newly
formed organic bone matrix in the defect area 8 weeks after surgery
(Fig. 8, A and B). Immunohistochemical staining (Fig. 8C) and
semi-­quantitative analysis showed that the relative expression of
CD31, osteocalcin (OCN), osteopontin (OPN), and runt-­related
transcription factor 2 (RUNX2) in the PS-­KP500/ACFs-­QK500
group was the highest among the three groups (Fig. 8, D to G).

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 7 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 6. The bone pro-­regenerative ability of KP peptide-­and QK peptide-­functionalized PS/ACFs in rat critical cranial bone defect model. (A) Schematic illustrating
the implantation of experimental scaffold (PS-­KP500/ACFs-­QK500) and collagen sponge (positive control) to a critical-­sized cranial bone defect (φ 8 mm) in rats (n = 5).
(B) X-­ray raw images of the control, collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 4 and 8 weeks of operation. (C and D) Micro-­CT 3D reconstruction images
and coronal views of the control, collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 4 and 8 weeks of operation. (E and F) Regenerated bone volume (BV/TV, %)
and surface coverage (%) of the control, collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 4 and 8 weeks of operation. *P < 0.05, **P < 0.01, and ***P < 0.001.

DISCUSSION lamellar structure with directional arrangement of channels. On the

In this work, our strategy for repairing large-­area skull defects is to
first recruit many cells involved in bone regeneration to the defect
area through the aligned cyrogel fibers. Then, the conjugated QK
peptide promotes blood vessel formation to support nutrients to
help the recruited cells perform their functions. Meanwhile, the
conjugated KP peptide promotes osteogenic differentiation and ac-
celerates the formation of new bone. The ACFs incorporated 3D
printed scaffold effectively facilitates bone regeneration by enhancing
cell recruitment and function.
In this work, the most innovative point of the designed scaffold is
the introduction of vertically ACFs into the 3D printed frame. Direc-
tional freeze casting is a typical method for preparing scaffolds with
oriented structures (25, 26). However, the obtained scaffold is a
basis of the directional freeze casting technology, we successfully
transformed the lamellar structure to ACFs by controlling the solution
concentration and freezing polymerization temperature. Low con-
centrations of GelMA solutions (around 0.2%) could form ACFs
after freeze polymerization and lyophilization. In addition, the di-
mension, density, and orientation of cryogel fibers can affect local
cell recruitment. For example, the ACFs can promote cell migration
compared to the RCFs (Fig. 2B). In addition, we also tried to control
the density of ACFs, and we found that temperature is the key point.
The fibrous ACF scaffold formed under −20°C exhibited significantly
low density in comparison to the scaffold formed under −190°C. Low-­
density fiber scaffolds are more conducive to cell migration. More-
over, the dimension of the formed cryogel fibers is in micron size

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 8 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 7. Histological observations of bone regeneration after 4 weeks of treatment. (A and B) H&E and trichrome staining of the decalcified cranial bone of the control, collagen
sponge, and PS-­KP500/ACFs-­QK500 groups after 4 weeks of operation (n = 5). (C and E) Trichrome staining of newly formed soft tissues in the edge and center area of the control,
collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 4 weeks of operation. The false color images indicate the orientation of the regenerated soft tissue. (D and F) Angle distribution
of the regenerated collagenous fibers in the edge and center of the regenerated soft tissue after 4 weeks of operation. NB, new bone; C, collagen sponge; PS, printed scaffold.

(1 to 4 μm). We are still trying to get aligned cryogel nanofibers. The
nanoscale cryogel fibers can further promote cell adhesion, migration,
and proliferation because of the regulatory role of nanotopography cues.
Our previous study demonstrated that the 3D aligned nanofiber
scaffolds could enhance cell penetration and subsequently accelerate
granulation tissue formation (18, 27). In the present research,
incorporating ACF into a 3D PS enhanced granulation tissue formation
in the defect area. The mechanism of repairing large-­area bone losses
is endochondral ossification. The granulation tissues first form in
the defect area, and then cartilage forms within the granulation tissue
and gradually transforms into bone (28, 29). Our study found several
dissociative new bones in the vertically ACF area of the hybrid

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 9 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e

Fig. 8. Histological observations of bone regeneration after 8 weeks of treatment. (A and B) H&E and trichrome staining of the decalcified cranial bone of the control,
collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 8 weeks of operation (n = 5). (C) CD31, OCN, OPN, and RUNX2 immunohistochemical staining of the regener-
ated bone tissue of the control, collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 8 weeks of operation. (D to G) Quantification of integral OD of expressed CD31,
OCN, OPN, and RUNX2 in the control, collagen sponge, and PS-­KP500/ACFs-­QK500 groups after 8 weeks of operation. *P < 0.05, **P < 0.01, and ****P < 0.0001.

scaffold group. This may be one of the reasons that our materials can
accelerate bone regeneration.
We noticed that the regenerated bone volume in the collagen
sponge group at 8 weeks after surgery was not significantly higher
than that at 4 weeks after surgery. However, the immunohisto­
chemistry results related to ossification were better than those in the
control group. We speculated that bone resorption in the collagen
sponge group increased at week 8. We further performed tartrate-­
resistant acid phosphatase (TRAP) staining of the new bone area to
test the hypothesis on week 8. TRAP staining images showed many
osteoclasts existed around and inside the new bone of the collagen
sponge group, while those in the control and PS-­KP500/ACFs-­QK500

Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024 10 of 14

S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
groups were mainly inside the new bone (fig. S11). Our findings
were consistent with many reported studies (30, 31). Our study
solved bone resorption due to the prolonged in vivo degradation
time and enhanced ossification induced by KP peptide (32, 33).
These effects allow new bone to continuously form and gradually
mature. Also, this is the main reason our bone repair materials im-
prove the quality of bone regeneration.
In summary, we reported a hybrid 3D PS with modification of
QK peptide and KP peptide for effectively promoting endogenous
cranial bone regeneration. The hybrid 3D PS consists of vertically
ACFs that guide and promote cell penetration into the defect area in
the early stages of bone repair. Then, the conjugated QK peptide and
KP peptide further regulate the function of the recruited cells to
promote vascularization and osteogenic differentiation in the defect
area. The regenerated bone volume and surface coverage of the dual
peptide-­modified hybrid scaffold were significantly higher than the
positive control group (collagen sponge) after 8 weeks of implantation.
In addition, the dual peptide-­modified hybrid scaffold demonstrated
sustained enhancement of bone regeneration compared to the collagen
sponge group from 4 to 8 weeks. We expect that the strategy of
introducing ACFs and incorporating functional peptides will inspire
the design of next-­generation biomaterials for effective endogenous
bone regeneration with the desired quality.
MATERIALS AND METHODS
Materials
Type A porcine skin gelatin (catalog number: V900863), methacrylic
anhydride (catalog number: 276685), ammonium persulfate (APS)
(catalog number: 248614), tetramethylethylenediamine (TEMED)
(catalog number: T9281), and 2-­hydroxy-­4′-­(2-­hydroxyethoxy)-2-­
methylpropiophenone (photoinitiator 2959) (catalog number:
410896) were purchased from Sigma-­Aldrich (Shanghai, China).
Dulbecco’s modified Eagle’s medium (DMEM), high glucose (catalog
number: 11965092), minimum essential medium (α-­MEM) (catalog
number: 12571063), fetal bovine serum (FBS) (catalog number:
10099141C), and penicillin-­streptomycin (P-­S) (catalog number:
15140148) were obtained from Gibco (Shanghai, China). CD31
polyclonal antibody (catalog number: 28083-­1-­AP) and osteocalcin
rabbit polyclonal antibody (catalog number: 23418-­1-­AP) were
purchased from Proteintech (Wuhan, China). Anti-­osteopontin anti-
body (catalog number: ab63856) and anti-­RUNX2 antibody (catalog
number: ab236639) were acquired from Abcam (Shanghai, China).
TRAP kit (catalog number: G1492) was purchased from Beijing Solar-
bio Science & Technology Co. Ltd. (Beijing, China). The commercial
collagen sponge was purchased from Beijing Pashionbio Co. Ltd.
(Beijing, China). The BMP-­2 knuckle epitope-­derived peptide (KP)
[N to C sequence: Alloc-­KIPK(Ac) ASSVPTELSAISTLYL] and VEGF-­
derived peptide (QK) [N to C sequence: Alloc-­KLTWQELYQLK(Ac)
YK(Ac)GI] were commercially synthesized and identified by Nanjing
TGpeptide Biotechnology Co. Ltd. (Nanjing, China).
Preparation of highly substituted Gel-­MA
Highly substituted Gel-­MA was prepared according to the previous
literature. Briefly, 10 g of type A porcine skin gelatin was added to
100 ml of carbonate bicarbonate buffer (0.25 M) to be dissolved at
55°C under continuous stirring. Then, 2 ml of methacrylic anhydride
was added into the gelatin solution using a micropump. After that,
the reaction solution continued to react for 1 hour at 55°C while pH
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
was kept at 9.4. The final pH of the reaction solution was adjusted to
7.4 to terminate the reaction. The final solution was dialyzed in
deionized water (changed three to four times a day) at 37°C for 4 days
and lyophilized to obtain Gel-­MA. The product was stored at −20°C.
Preparation of 3D PSs embedded with ACFs
First, 0.025 g of photoinitiator 2959 was dissolved in 5 ml of deionized
water at 50°C away from light. Then, 0.25 g of Gel-­MA was added
into the above solution at 50°C under continuous stirring away from
light to attain 5 wt % Gel-­MA solution. The hydrogel scaffold was
produced by a 3D bioprinter (BioMaker4, SunP Biotech, Beijing,
China) and crosslinked and cured under ultraviolet (UV) light
(OmniCure S2000, Excelitas Technologies Corp., Canada). The printing
parameters were set as the following: printer’s needle, 25 gauge;
model height, 1.0 or 2.0 mm; layer height, 0.2 mm; line distance, 1.2 mm;
the lines of the adjacent two layers crossed (90°); printing speed,
6 mm/s; and extrusion speed, 1 mm3/s. The cured hydrogel was cut
into cylinders with a diameter of 8 mm, lyophilized, and stored at
−20°C. The preparation method of the KP peptide–modified Gel-­MA
hydrogel scaffold was the same as above. Bio-­ink containing different
concentrations (250, 500, and 750 μg/ml) of KP peptide was prepared
with 5 wt % Gel-­MA solution mixed with KP peptide at 37°C under
continuous stirring away from light.
The aligned cryogel fiber (ACF) was embedded in the lattices of
a 3D PS by unidirectional freeze casting technique. A 90 mg of Gel-­MA
and 10 mg of gelatin were dissolved in 50 ml of deionized water at
50°C under continuous stirring away from light to obtain a presolution
containing 0.18 wt % Gel-­MA and 0.02 wt % gelatin. Then, 4 ml of
presolution was mixed with 0.2 ml of APS (10 wt %) and 5.2 μl of
TEMED in an ice bath. The lyophilized 3D PS was immersed in the
above mixture for 1 min in an ice bath and then quickly transferred
to the stainless steel plate precooled at −80°C while ensuring that
the scaffold’s bottom fell to the plate simultaneously. The scaffold
continued to be placed at −80°C for 24 hours for chemical cross-
linking of Gel-­MA. The product was lyophilized and stored at
−20°C. In subsequent experiments, gelatin helps ACFs maintain its
vertical fiber structure in PBS at low temperatures. The preparation
method of QK peptide-­modified ACF was the same as above. QK pep-
tide was dissolved in the presolution in an ice bath under continuous
stirring away from light to attain different concentrations (100, 300,
and 300 μg/ml) of QK peptide. The scaffold continued to be placed
at −80°C for 24 hours for chemical crosslinking of Gel-­MA and QK
peptide.
Characterization of biomaterials
Gross photographs of the biomaterials were taken with a stereoscopic
microscope (Olympus SZ61). The microstructure was observed by
SEM (Hitachi SU8010, Japan). The directionality of aligned fibers
was analyzed using the “Directionality” and “OrientationJ” plugins
in Fiji software. The aligned fibers’ diameter distribution was obtained
by measuring 200 fiber segments using ImageJ software.
Compressive mechanical properties of the wet PS and PS/ACFs
after crosslinking were obtained by an electronic universal testing
machine (Instron 5944, USA). In compression parameter settings,
the model was cylindrical (diameter, 8 mm; height, 2 mm) and the
moving speed was 1 mm/min.
In the swelling test, the freeze-­dried scaffold was weighed (M0).
It was immersed in 37°C PBS buffer and weighed (M1) at a predeter-
mined time. The swelling ratio was calculated as M1/M0.
11 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
In vitro enzymolysis test, the freeze-­dried scaffold was weighed
after rehydration in PBS buffer at 37°C for 1 hour (M0). Then, the
rehydrated scaffold was immersed in 0.05% collagenase I (Gibco)
solution at 37°C (enzyme activity about 100 U/ml) and weighed
(M1) at a predetermined time point. The remaining mass ratio was
calculated as M1/M0.
In vitro biological properties of biomaterials
Cell seeding and adhesion on biomaterials
3D PS embedded with or without ACF (diameter, 8 mm; height, 1 mm)
was freeze-­dried and sterilized using UV light. HUVECs with red
fluorescent protein (RFP) fluorescence were cultured in high-­glucose
DMEM medium supplemented with 10% FBS and 1% P-­S at 37°C
and 5% CO2. PS/ACFs were immersed in PBS at 4°C for 4 hours
before biological experiments to remove the residual APS, TEMED,
and their by-­products. Scaffolds were put in a 48-­well plate with one
scaffold per well, and a complete medium was added. The scaffolds
were rehydrated at 37°C for 1 hour for swelling balance, and excess
liquid was discarded. HUVECs-­RFP suspension (50 μl; 2 × 106 cells/ml)
was dropped to the top surface of the scaffold and cultured at 37°C
and 5% CO2. After 2 hours, the fresh complete medium was supple-
mented, and the cells continued to be cultured for 24 hours. Then,
the cell-­loaded scaffolds were transferred to a new 48-­well plate for
the subsequent experiments.
The cell-­loaded semi-­quantitative analysis
The culture medium of the new 48-­well plate was replaced with
CCK-­8 solution according to the manufacturer’s protocol (Beyotime
Biotechnology) and incubated for 1 hour away from light. Absorbance
[optical density (OD) value)] at 450 nm was determined with a Micro-
plate Spectrophotometer (Epoch2, Bio-­Tec Instruments, USA).
CLSM and SEM observation of cell adhesion and distribution
on biomaterials
For confocal laser scanning microscope (CLSM) observation, the cell-­
loaded scaffold in the new 48-­well plate was fixed with 4% paraformal-
dehyde for 15 min and washed three times with PBS buffer. The cell
adhesion and distribution on the biomaterials were observed using
CLSM (Nikon A1, Japan) with the z-­stack range from 0 to 600 μm
and an interval of 10 μm. For SEM observation, the cell-­loaded scaffolds
fixed with 4% paraformaldehyde were dehydrated by gradient alcohol
(70 to 100%, 15 min for each class) and vacuum-­dried at 37°C for
24 hours. The scaffolds were sputtered with platinum (Leica EM
ACE600, Germany) and observed under SEM (Hitachi SU8010, Japan).
Subcutaneous implantation in rats
The Sprague-­Dawley rats [specific pathogen–free (SPF) grade, male,
10 weeks, 300 g] were purchased from the Experimental Animal Center
of Zhejiang Province. This animal experiment protocol was approved by
the Animal Research and Ethics Committee of Wenzhou Institute of
the University of Chinese Academy of Sciences (approval number:
WIUCAS22072901). The experimental groups used are 3D PS, 3D
PS embedded with ACF (PS/ACFs), 3D PS embedded with QK peptide–
modified (100, 300, and 500 μg/ml) ACF (PS/ACFs-­QK100, PS/
ACFs-­QK300, and PS/ ACFs-­QK500). There were eight samples
(four rats) in each group, and four samples in two rats were collected
1 and 2 weeks after surgery, respectively. The operation procedure
was as follows: The rat was anesthetized with 3% isoflurane in oxygen
for about 2 min, and then the back hair of the rat was shaved, disinfected
with iodophor twice, and cleaned with 75% ethanol gauze once. The skin
and subcutaneous tissue were cut to make a subcutaneous cavity, and
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
the scaffold was sent into the subcutaneous cavity, with one scaffold
in each cavity. The incision was closed and disinfected. The rat was
placed on a heating pad during the surgery to maintain its body
temperature. Euthanasia by an overdose of anesthesia was performed
on the rats after 1 and 2 weeks of surgery. The obtained subcutaneous
tissue samples were washed with PBS buffer and soaked in 4% parafor-
maldehyde for 1 day, then underwent alcohol gradient dehydration
and paraffin embedding, and then cut into 5-­μm-­thick slices for histo-
logical staining.
Ectopic osteogenesis in nude mice
Nude mice (SPF grade, male, 6 to 8 weeks, 18 to 25 g) were purchased
from the Experimental Animal Center of Zhejiang Province. This animal
experiment protocol has been approved by the Animal Research
and Ethics Committee of Wenzhou Institute of the University of Chinese
Academy of Sciences (approval number: WIUCAS22122001). Ex-
perimental groups used are 3D PSs and 3D PSs modified with KP
peptide (250, 500, and 750 μg/ml) (PS-­KP250, PS-­KP500, and PS-­
KP750). Six samples in three mice were collected in each group at
4 weeks after surgery. The procedures for cell seeding on scaffolds
were as follows: First, the freeze-­dried and UV-­sterilized scaffolds
were immersed in a complete medium (α-­MEM, 10% FBS, and 1%
P-­S) at 37°C for 1 hour for rehydration and then transferred to a 48-­
well plate. P3 rBMSC suspension (50 μl; 4 × 107 cells/ml) was
dropped to the top surface of each scaffold and cultured at 37°C and
5% CO2. After 2 hours, the fresh complete medium was supplemented,
and the cells continued to be cultured for 24 hours. Then, the cell-­
loaded scaffolds were for the subsequent subcutaneous implantation.
The procedures were as follows: The nude mouse was anesthetized
with 2% isoflurane in oxygen for about 2 min, the back skin was
disinfected with iodophor twice, and the back skin was cleaned with
ethanol gauze once. The skin and subcutaneous tissue were cut to
make a subcutaneous cavity, and the cell-­loaded scaffold was sent
into the subcutaneous cavity, one scaffold in each cavity. The incision
was closed and disinfected. The mouse was placed on a heating pad
during the surgery to maintain its body temperature. Euthanasia by
an overdose of anesthesia was performed on the mice after 4 weeks
of surgery. The subcutaneous tissue samples were washed with PBS
buffer, soaked in 4% paraformaldehyde for 3 days, and then transferred
to 70% ethanol, followed by a Micro-­CT scan, decalcification (10%
EDTA, 2 weeks, room temperature), and histological evaluation.
Critical-­sized rat cranial bone defect model
The Sprague-­Dawley rats (SPF grade, male, 12 weeks, 350 g) were
purchased from the Experimental Animal Center of Zhejiang Province.
This animal experiment protocol was approved by the Animal Research and
Ethics Committee of Wenzhou Institute of the University of Chinese
Academy of Sciences (Approval number: WIUCAS22121603). Experi-
mental groups used are control group (no scaffold), collagen sponge
group (positive control), and KP peptide–modified 3D PS embedded
with QK peptide–modified ACF (PS-­KP500/ACFs-­QK500) group.
Each group had 10 samples (10 rats), and 5 samples in 5 rats were
collected 4 and 8 weeks after surgery, respectively. The operation
procedures were as follows: The rat was anesthetized with 3% isoflurane
in oxygen for about 2 min, and then the rat calvarial hair was shaved,
disinfected with iodophor twice, and cleaned with 75% ethanol
gauze once. Sagittal incision was made at the top of the skull of the
rat. The skin, subcutaneous tissue, and periosteum were sequentially
incised to expose the calvaria. A critical-­sized cranial bone defect
12 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
(8 mm) was created by using a trephine bur mounting on a dentist
drill. The normal saline solution was continuously added to reduce
heat damage. The wound was cleaned with a normal saline solution,
and different treatments were applied to the bone defects. The perios-
teum, subcutaneous tissue, and skin were sutured, and the incision
was disinfected with iodophor. The rat was placed on a heating pad
during the surgery to maintain its body temperature. Euthanasia by
an overdose of anesthesia was performed on the rats after 4 and 8 weeks
of surgery. The rat cranium samples were isolated, washed with PBS
buffer, soaked in 4% paraformaldehyde for 3 days, and then trans-
ferred to 70% ethanol, followed by a Micro-­CT scan, decalcification
(10% EDTA, 4 weeks, room temperature), and histological evaluation.
Histological observation
After alcohol gradient dehydration and paraffin embedding, the tissue
samples were cut into 5-­μm-­thick slices for H&E staining, trichrome
staining, and immunohistochemical staining. Trichrome staining was
carried out according to the standard protocol of the manufacturer
(catalog number: MST-­8003, Fuzhou Maixin Biotechnology Develop-
ment Co. Ltd., China). During immunohistochemical staining, the
slices were deparaffinized and rehydrated, followed by antigen
retrieval in heated citrate buffer for 15 min (citrate buffer solution,
pH 6.0 at 90°C). Nonspecific antibody binding was prevented with
10% goat serum albumin solution. The slices were incubated overnight
at 4°C after adding the primary antibody. Then, the corresponding
secondary antibodies were added and incubated at room temperature
for 1 hour. Tissue color development using the DAB Horseradish
Peroxidase Color Development Kit and observation were carried
out according to the standard protocol of the manufacturer (catalog
number: ZLI-­9018, Beijing Zhongshan Jinqiao Biotechnology Co.
Ltd., China). At least five randomly selected fields were examined
for each group at each time point and used to assess the average
positive intensity. The positive expression of OPN, OCN, RUNX2,
and CD31 in the new bone was determined using Fiji software and
was presented as the integral OD.
Micro-­CT scanning and analysis
The ectopic osteogenesis and rat cranium samples were scanned using
high-­resolution Micro-­CT (SKYSCAN 1275, Bruker). Scanning para
meters included voltage (50 kV), current (60 μA), and image pixel
size (9.0 μm). CTAn software (Bruker) was used to analyze bone
volume, bone area, Tb.Th, Tb.N, and Tb.Pf. The Micro-­CT 3D images of
the isolated samples were reconstructed by CTvol (Bruker). The surface
coverage of regenerated bone was calculated using ImageJ software.
KP peptide–modified 3D PS promotes osteogenic
differentiation of rBMSC in vitro
ALP staining
P3 rBMSCs were seeded on a 48-­well plate with a cell density of
1 × 104 cells per well and cultured in α-­MEM containing 10% FBS
and 1% P-­S at 37°C and 5% CO2 condition. The medium was replaced
by an osteogenic induction medium [α-­MEM containing 10% FBS,
1% P-­S, ascorbic acid (50 μg/ml), 10 mM sodium β-­glycerophosphate,
and 10 nm dexamethasone] 24 hours later. A scaffold was added to
each well. The scaffolds were 3D PS modified with KP peptide at
different concentrations (250, 500, and 750 μg/ml) (PS-­KP250, PS-­
KP500, and PS-­KP750) after lyophilization and UV sterilization.
They were immersed in an osteogenic induction medium at 37°C for
1 hour for rehydration. The osteogenic induction medium was changed
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
every 2 to 3 days. The cells were cultured for 7 days. ALP staining
was performed according to the instructions (BCIP/NBT Alkaline
Phosphatase Color Development Kit, Beyotime Biotechnology, China).
Real-­time quantitative polymerase chain reaction
Total RNA of rBMSCs cocultured with 3D PSs modified with dif-
ferent concentrations of KP peptide in osteogenic induction me-
dium for 7 days was extracted, according to the standard protocol
of the manufacturer (Nanjing Vazyme Biotech Co.,Ltd.). Then, the
RNA was reverse-­transcribed into cDNA using SYBR PremexEx-
Taq (Takara, China). The mRNA levels of ALP, OPN, and OCN in
rBMSCs were detected by real-­time quantitative polymerase chain
reaction. Relative gene expression was normalized to housekeep-
ing gene glyceraldehyde-­3-­phosphate dehydrogenase (GAPDH).
The primers used in this study are as follows: ALP (forward primer:
CAACCTGACTGACCCTTCCC; reverse primer: CGATGGCCT­
CATCCATCTCC), OPN (forward primer: TGAGTTTGGCAGCT-
CAGAGG; reverse primer: TCGTCGTCGTCATCATCGTC), OCN
(forward primer: CAAGCAGGAGGGCAGTAAGG; reverse primer:
GAAGCCAATGTGGTCCGCTA), and GAPDH (forward primer:
GCGAGATCCCGCTAACATCA; reverse primer: CTCGTGGTTC­
ACACCCATCA).
Statistical analysis
All data were presented as means ± SD. Statistical analysis was
performed using GraphPad Prism 7.0 software. The difference
between the two groups of independent samples was evaluated
using a t test. Differences among groups were assessed by one-­way
analysis of variance (ANOVA) followed by Tukey’s multiple com-
parison test. The values of P < 0.05 were considered statistically
significant. The values of P < 0.01 were considered statistically
very significant.
Supplementary Materials
This PDF file includes:
Figs. S1 to S11
REFERENCES AND NOTES
1. C. E. Gillman, A. C. Jayasuriya, FDA-­approved bone grafts and bone graft substitute
devices in bone regeneration. Mater. Sci. Eng. C Mater. Biol. Appl. 130, 112466 (2021).
2. M. E. Elsalanty, D. G. Genecov, Bone grafts in craniofacial surgery. Craniomaxillofac.
Trauma Reconstr. 2, 125–134 (2009).
3. S. Berretta, K. Evans, O. Ghita, Additive manufacture of PEEK cranial implants:
Manufacturing considerations versus accuracy and mechanical performance. Mater. Des.
139, 141–152 (2018).
4. T. Takizawa, N. Nakayama, H. Haniu, K. Aoki, M. Okamoto, H. Nomura, M. Tanaka,
A. Sobajima, K. Yoshida, T. Kamanaka, K. Ajima, A. Oishi, C. Kuroda, H. Ishida, S. Okano,
S. Kobayashi, H. Kato, N. Saito, Titanium fiber plates for bone tissue repair. Adv. Mater. 30,
1703608 (2018).
5. I. Yadroitsava, A. Du Plessis, I. Yadroitsev, Bone regeneration on implants of titanium
alloys produced by laser powder bed fusion: A review, in Titanium for Consumer
Applications (Elsevier, 2019), chap. 12, pp. 197–233.
6. T. Maqbool, A. Binhammer, P. Binhammer, O. M. Antonyshyn, Risk factors for titanium
mesh implant exposure following cranioplasty. J. Craniofac. Surg. 29, 1181–1186 (2018).
7. J. A. Smith, S. Petersmann, F. Arbeiter, U. Schäfer, Optimization and manufacture of
polyetheretherketone patient specific cranial implants by material extrusion—A clinical
perspective. J. Mech. Behav. Biomed. Mater. 144, 105965 (2023).
8. P. Chen, J. Su, H. Wang, L. Yang, M. Li, Z. Li, J. Liu, S. Wen, Y. Zhou, C. Yan, Y. Shi, Aging
mechanism of polyetheretherketone powder during layer-­wise infrared radiation of
high-­temperature laser powder bed fusion. Mater. Des. 213, 110348 (2022).
9. D. Zhang, X. Wu, J. Chen, K. Lin, The development of collagen based composite scaffolds
for bone regeneration. Bioact. Mater. 3, 129–138 (2017).
10. D. Qin, N. Wang, X.-­G. You, A.-­D. Zhang, X.-­G. Chen, Y. Liu, Collagen-­based biocomposites
inspired by bone hierarchical structures for advanced bone regeneration: Ongoing
research and perspectives. Biomater. Sci. 10, 318–353 (2022).
13 of 14
S c i e n c e A d v a n c e s | R e s e ar c h A r t i c l e
11. S. J. Hollister, Porous scaffold design for tissue engineering. Nat. Mater. 4, 518–524 (2005).
12. E. García-­Gareta, M. J. Coathup, G. W. Blunn, Osteoinduction of bone grafting materials
for bone repair and regeneration. Bone 81, 112–121 (2015).
13. J. Henkel, M. A. Woodruff, D. R. Epari, R. Steck, V. Glatt, I. C. Dickinson, P. F. M. Choong,
M. A. Schuetz, D. W. Hutmacher, Bone regeneration based on tissue engineering
conceptions—a 21st century perspective. Bone Res. 1, 216–248 (2013).
14. G. N. Duda, S. Geissler, S. Checa, S. Tsitsilonis, A. Petersen, K. Schmidt-­Bleek, The decisive
early phase of bone regeneration. Nat. Rev. Rheumatol. 19, 78–95 (2023).
15. G. L. Koons, M. Diba, A. G. Mikos, Materials design for bone-­tissue engineering. Nat. Rev.
Mater. 5, 584–603 (2020).
16. A. K. Gaharwar, I. Singh, A. Khademhosseini, Engineered biomaterials for in situ tissue
regeneration. Nat. Rev. Mater. 5, 686–705 (2020).
17. E. M. Bueno, J. Glowacki, Cell-­free and cell-­based approaches for bone regeneration. Nat.
Rev. Rheumatol. 5, 685–697 (2009).
18. S. Chen, H. Wang, A. McCarthy, Z. Yan, H. J. Kim, M. A. Carlson, Y. Xia, J. Xie, Three-­
dimensional objects consisting of hierarchically assembled nanofibers with controlled
alignments for regenerative medicine. Nano Lett. 19, 2059–2065 (2019).
19. S. Chen, H. Wang, V. L. Mainardi, G. Talò, A. McCarthy, J. V. John, M. J. Teusink, L. Hong,
J. Xie, Biomaterials with structural hierarchy and controlled 3D nanotopography guide
endogenous bone regeneration. Sci. Adv. 7, eabg3089 (2021).
20. S. Chen, A. McCarthy, J. V. John, Y. Su, J. Xie, Converting 2D nanofiber membranes to 3D
hierarchical assemblies with structural and compositional gradients regulates cell
behavior. Adv. Mater. 32, e2003754 (2020).
21. L. D. D'Andrea, G. Iaccarino, R. Fattorusso, D. Sorriento, C. Carannante, D. Capasso,
B. Trimarco, C. Pedone, Targeting angiogenesis: Structural characterization and biological
properties of a de novo engineered VEGF mimicking peptide. Proc. Natl. Acad. Sci. U.S.A.
102, 14215–14220 (2005).
22. R. Li, C. Zhou, J. Chen, H. Luo, R. Li, D. Chen, X. Zou, W. Wang, Synergistic osteogenic and
angiogenic effects of KP and QK peptides incorporated with an injectable and
self-­healing hydrogel for efficient bone regeneration. Bioact. Mater. 18, 267–283
(2022).
23. S. K. Boda, Y. Almoshari, H. Wang, X. Wang, R. A. Reinhardt, B. Duan, D. Wang, J. Xie,
Mineralized nanofiber segments coupled with calcium-­binding BMP-­2 peptides for
alveolar bone regeneration. Acta Biomater. 85, 282–293 (2019).
24. L. Weng, S. K. Boda, H. Wang, M. J. Teusink, F. D. Shuler, J. Xie, Novel 3D hybrid nanofiber
aerogels coupled with BMP-­2 peptides for cranial bone regeneration. Adv. Healthc. Mater.
7, e1701415 (2018).
25. H. Zhang, I. Hussain, M. Brust, M. F. Butler, S. P. Rannard, A. I. Cooper, Aligned two-­and
three-­dimensional structures by directional freezing of polymers and nanoparticles. Nat.
Mater. 4, 787–793 (2005).
Wei et al., Sci. Adv. 10, eadk6722 (2024) 7 February 2024
26. P. Zhang, J. Li, L. Lv, Y. Zhao, L. Qu, Vertically aligned graphene sheets membrane for
highly efficient solar thermal generation of clean water. ACS Nano 11, 5087–5093 (2017).
27. J. Jiang, S. Chen, H. Wang, M. A. Carlson, A. F. Gombart, J. Xie, CO2-­expanded nanofiber
scaffolds maintain activity of encapsulated bioactive materials and promote cellular
infiltration and positive host response. Acta Biomater. 68, 237–248 (2018).
28. E. M. Thompson, A. Matsiko, E. Farrell, D. J. Kelly, F. J. O'Brien, Recapitulating
endochondral ossification: A promising route to in vivo bone regeneration. J. Tissue
Eng. Regen. Med. 9, 889–902 (2015).
29. Y. Liu, Z. Yang, L. Wang, L. Sun, B. Y. S. Kim, W. Jiang, Y. Yuan, C. Liu, Spatiotemporal
immunomodulation using biomimetic scaffold promotes endochondral ossification-­
mediated bone healing. Adv. Sci. 8, e2100143 (2021).
30. X. Ren, Q. Zhou, D. Foulad, A. S. Tiffany, M. J. Dewey, D. Bischoff, T. A. Miller, R. R. Reid,
T.-­C. He, D. T. Yamaguchi, B. A. C. Harley, J. C. Lee, Osteoprotegerin reduces osteoclast
resorption activity without affecting osteogenesis on nanoparticulate mineralized
collagen scaffolds. Sci. Adv. 5, eaaw4991 (2019).
31. H. K. Lin, Y. H. Pan, E. Salamanca, Y. T. Lin, W. J. Chang, Prevention of bone resorption by
HA/β-­TCP+ collagen composite after tooth extraction: A case series. Int. J. Environ. Res.
Public Health 16, 4616 (2019).
32. N. Shao, J. Guo, Y. Guan, H. Zhang, X. Li, X. Chen, D. Zhou, Y. Huang, Development of
organic/inorganic compatible and sustainably bioactive composites for effective bone
regeneration. Biomacromolecules 19, 3637–3648 (2018).
33. E. Jabbari, Osteogenic peptides in bone regeneration. Curr. Pharm. Des. 19, 3391–3402 (2013).
Acknowledgments
Funding: This work was financially supported by the National Natural Science Foundation of
China (grant no. 82102334 to S.C.), the Key Foundation of Zhejiang Provincial Natural Science
Foundation (grant no. LZ22C100001 to S.C.), and Key Area Research and Development
Program of Guangdong Province (grant no. 2020B090924004 to C.Z.). Author contributions:
S.C. and Y.W. conceived the study idea. S.C., W.W., C.Z., Y.W., and H.P. designed the experiments.
Y.W., H.P., and J.Y. performed the experiments. S.C., Y.W., H.P., and C.Z. performed the data
analysis and interpretation. All authors contributed to writing the manuscript, discussing the
results and implications, and editing the manuscript at all stages. Competing interests: The
authors declare that they have no competing interests. Data and materials availability: All
data needed to evaluate the conclusions in the paper are present in the paper and/or the
Supplementary Materials.
Submitted 15 October 2023
Accepted 5 January 2024
Published 7 February 2024
10.1126/sciadv.adk6722
14 of 14