BCG vaccination for Steps toward regulating Landfills emit methane

cattle pp. 1410 & 1433 indoor air quality p. 1418 persistently p. 1499

$15
29 MARCH 2024
science.org

A radical new
particle accelerator
concept emerges. Call it physicists’

MUON SHOT p. 1405

Deposited by scientists,
to further cancer research

Access a unique cell line collection

including models of:

Tumours and tumourigenicity

Drug resistance

Non-tumourigenic
immortalised cells

ut
Genes knocked-in/out

elled
Reporter-labell

Stem cells
cel

View the
collection:

Visit us at
Booth

#1314
14

Learn more about CancerTools.org, a non-profit,

cancer research tools biorepository ncertools.org
canc
 Every 8 minutes,
a CST antibody gets cited.
®

Take a minute to learn why.

Scientists have recently demonstrated the importance of elucidating spatial local-

ization and proximity of immune cells expressing immune checkpoints in the tumor
microenvironment. Shown here are tumor cells stained with PD-L1 XP® Rabbit mAb
#13684 in yellow, as well as T cells stained with PD-1 XP Rabbit mAb #86163 in green.

Reproducible results are critical for publication. That’s why CST antibodies
have earned more citations per product than those of any other antibody provider.
Eight of the 12 most-cited antibodies are from CST, including the #1-cited primary and
secondary antibodies. And over one third of the antibodies on CiteAb’s Top 100 list* are
ours—more than any other company’s. Why? Because we stringently validate our
antibodies — and guarantee them to work as expected in the specific application
they’ll be used in. Researchers around the world trust CST antibodies to give them
reliable, reproducible results. You can, too.

cellsignal.com
© 2023 Cell Signaling Technology, Inc. All rights reserved. | All other trademarks are the property of their respective owners. Visit cellsignal.com/trademarks for more information.
*CiteAb, "What were the 100 most popular antibodies of 2021?" Sept 15, 2022 23-BCH-03734
 CONTENTS
2 9 M A R C H 2 0 24
VO LU M E 3 8 3
ISSUE 6690

1410 & 1433

NEWS
IN BRIEF
1396 News at a glance
IN DEPTH
1398 Massive RNA sequencing
effort proposed
U.S. plan would harness the “RNome”
for medicine and more—but funding is
uncertain By E. Pennisi
1399 West Virginia OKs intelligent design
in schools
Governor signs bill allowing teachers to discuss
antievolutionary “theories” By J. Mervis
1400 Firms aim to capture carbon in
the oceans
Schemes to process seawater may be cheaper
than snagging carbon dioxide from the air
By R. F. Service
1401 NIH’s flat budget for 2024 will force
‘very difficult decisions’
PHOTO: MICHELE SPATARI/GETTY IMAGES
1403 Africa may lead rollout of long-
lasting HIV drug
U.S. agency plans to make inexpensive
prevention shots widely available for people
at risk By J. Cohen
1404 States in India survey
caste membership
Censuses, the first in a century, chart social
changes and shape affirmative action
programs By V. Chandrashekhar
FEATURES
1405 The dream machine
An accelerator known as a muon collider
could revolutionize particle physics—if it can
be built By A. Cho
INSIGHTS
PERSPECTIVES
1410 Vaccines to control tuberculosis
in cattle
1414 A mitotic stopwatch determines
cell fate
Surveillance of mitotic timing prevents
amplification of damaged cells
By A. P. Bertolin and V. Gottifredi
RESEARCH ARTICLE p. 1441
1416 A more biofriendly piezoelectric
material
A ferroelectric molecular crystal displays
characteristics required for implantation
By L. Wang and R-W Li
RESEARCH ARTICLE p. 1492
1417 Ingeborg Levin (1953—2024)
Pioneer in radiocarbon and atmospheric
research By F. Vogel and S. Hammer
POLICY FORUM
1418 Mandating indoor air quality
for public buildings
If some countries lead by example, standards
may increasingly become normalized
By L. Morawska et al.
BOOKS ET AL.

Belated Senate and House compromise
eliminates policy riders restricting certain
studies By J. Kaiser
1402 Transgenic marmosets mimic
Parkinson’s symptoms
New monkey models promise insight into early
stages of degenerative brain disease
By D. Normile
The age-old cattle disease has resisted
rigorous control, but the BCG vaccine may do
better By A. L. Michel
RESEARCH ARTICLE p. 1433
1412 Chemodiversity in freshwater health
Dissolved organic matter may offer a way to
track and restore the health of fresh waters
By A. J. Tanentzap and J. A. Fonvielle
1421 American climate migration
Increasingly inhospitable conditions will
change the nation’s demography, argues a
journalist By M. R. Aczel and M. E. Mor Barak
1422 Framing tough problems
An engineer urges readers to think more
holistically about complex challenges
By D. Riley

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1389

The NOMIS & Science Young Explorer Award recognizes and rewards
early-career M.D., Ph.D., or M.D./Ph.D. scientists that perform research at the
intersection of the social and life sciences. Essays written by these bold
researchers on their recent work are judged for clarity, scientific quality,
creativity, and demonstration of cross-disciplinary approaches to address
fundamental questions.

A cash prize of up to USD 15,000 will be awarded to essay winners, and their
engaging essays will be published in Science. Winners will also be invited to
share their work and forward-looking perspective with leading scientists in
their respective fields at an award ceremony.

Apply by May 15, 2024

 CONTENTS

LETTERS 1434 Coronavirus

1424 Chile’s Valparaíso hills on fire
By M. E. González et al.
Design of a SARS-CoV-2 papain-like protease
inhibitor with antiviral efficacy in a mouse
model B. Tan et al.
1506
1424 China’s plan to combat antimicrobial
resistance By X. Li et al. 1441 Cell cycle
Control of cell proliferation by memories of
1425 Past as prologue: The fern mandolin mitosis F. Meitinger et al.
By M. A. Handley PERSPECTIVE p. 1414

PODCAST
1448 Plant science
Biosynthesis of the allelopathic alkaloid

RESEARCH gramine in barley by a cryptic oxidative

rearrangement S. L. Dias et al.

1455 Hydrogels
IN BRIEF A hyperelastic hydrogel with an ultralarge
1428 From Science and other journals reversible biaxial strain L. Chen et al.

RESEARCH ARTICLES
1461 Machine learning
Mechanism for feature learning in 1492 Piezoelectrics
1431 Structural biology neural networks and backpropagation- Biodegradable ferroelectric molecular crystal
Molecular mechanism of dynein-dynactin free machine learning models with large piezoelectric response
complex assembly by LIS1 K. Singh et al. A. Radhakrishnan et al. H.-Y. Zhang et al.
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT:
PERSPECTIVE p. 1416
DOI.ORG/10.1126/SCIENCE.ADK8544
1467 Ultrafast optics
Ultrafast Kapitza-Dirac effect K. Lin et al. 1499 Methane emissions
1432 Zoonomia
Vocal learning–associated convergent Quantifying methane emissions from
1471 Neuroscience United States landfills D. H. Cusworth et al.
evolution in mammalian proteins and
Oxygen imaging of hypoxic pockets in the
regulatory elements M. E. Wirthlin et al.
mouse cerebral cortex F. R. M. Beinlich et al.
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT:
DOI.ORG/10.1126/SCIENCE.ABN3263 DEPARTMENTS
1478 Neuroscience 1395 Editorial
1433 Vaccination Selection of experience for memory by Strengthen the case for DEI
BCG vaccination reduces bovine tuberculosis hippocampal sharp wave ripples W. Yang et al. By S. Malcom
transmission, improving prospects for
elimination A. Fromsa et al. 1484 Metabolism 1506 Working Life
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT: Pyrimidines maintain mitochondrial pyruvate Shaken up by a shaker By E. L. Eberhardt
DOI.ORG/10.1126/SCIENCE.ADL3962 oxidation to support de novo lipogenesis
PERSPECTIVE p. 1410 U. Sahu et al.
ON THE COVER
By colliding weighty, unstable subatomic
particles called muons, physicists hope to
reach the high energies needed for new dis-
coveries more quickly and cheaply than with
a more-conventional proton collider. But the
unproven technology is challenging because
muons decay in a fraction of a second. In the
simulation shown here,
collision products streak
through a haze of extra-
neous particles from the
CREDITS: (LEFT TO RIGHT) CARBON MAPPER; ROBERT NEUBECKER

muons’ decay. See page

1405. Image: Lawrence
Lee and Charles Bell,
University of Tennessee,
Knoxville

1499 AAAS News & Notes ................................ 1426

Methane plumes from a landfill in Georgia, USA, detected remotely by satellite Science Careers .......................................1505

SCIENCE (ISSN 0036-8075) is published weekly on Friday, except last week in December, by the American Association for the Advancement of Science, 1200 New York Avenue, NW, Washington, DC 20005. Periodicals mail
postage (publication No. 484460) paid at Washington, DC, and additional mailing offices. Copyright © 2024 by the American Association for the Advancement of Science. The title SCIENCE is a registered trademark of the AAAS. Domestic
individual membership, including subscription (12 months): $165 ($74 allocated to subscription). Domestic institutional subscription (51 issues): $2627; Foreign postage extra: Air assist delivery: $107. First class, airmail, student, and
emeritus rates on request. Canadian rates with GST available upon request, GST #125488122. Publications Mail Agreement Number 1069624. Printed in the U.S.A.
Change of address: Allow 4 weeks, giving old and new addresses and 8-digit account number. Postmaster: Send change of address to AAAS, P.O. Box 96178, Washington, DC 20090–6178. Single-copy sales: $15 each plus shipping and
handling available from backissues.science.org; bulk rate on request. Authorization to reproduce material for internal or personal use under circumstances not falling within the fair use provisions of the Copyright Act can be obtained
through the Copyright Clearance Center (CCC), www.copyright.com. The identification code for Science is 0036-8075. Science is indexed in the Reader’s Guide to Periodical Literature and in several specialized indexes.

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1391

Sponsored Feature
Putting
community
health at the
center of city life
Social interactions connect and shape us. They also are helping
researchers at Morgan State University address the most pressing
health issues in urban settings.
Humans are social beings, and our interactions within our community
mold who we are. But not all communities have the same resources,
which has reverberations throughout many aspects of life and
contributes to health disparities.
To tackle those disparities, Morgan State University, a historically
Black institution in Baltimore, has two research centers working
against the health consequences that stem from unfair historical
policies—the Center for Urban Health Equity (CUHE) and the Research
Centers in Minority Institutions’ (RCMI) Center for Urban Health
Disparities Research and Innovation (referred to as RCMI@Morgan).
“A lot of cities across the country were shaped by policies such
as redlining,” says Lawrence Brown, a research scientist at CUHE,
referring to the practice of refusing mortgage loans to communities
of color. “Local policies, zoning laws, and ordinances were very much
responsible for helping segregate a lot of areas by race.”
The legacy of these events still haunts many American cities.
Some communities of color lack resources, such as grocery stores,
high-quality schools, functional infrastructure, and jobs that pay living
wages. Instead, the overabundance of factories, railways, and busy
roads diminishes air and water quality. According to the National
Institute of Allergy and Infectious Diseases, minority groups experience
higher rates of asthma, HIV/AIDS, tuberculosis, and COVID-19.
Susceptibility to these diseases is strongly influenced by a set of
community-wide risk factors called “social determinants of health.”
The Maryland-state-funded CUHE was established in 2021, a year
into the devastating COVID-19 pandemic that had a disproportional
PHOTO: PROVIDED BY MORGAN STATE UNIVERSITY
impact on African Americans. According to the National Institutes of
Health, approximately 98 out of every 100,000 African Americans have
died from COVID-19, a mortality rate that is a third higher than that for
Latinos, and more than double that for whites and Asians, bringing up a
larger discussion around urban health.
One contributing factor to health discrepancies is that underserved
communities are less likely to receive a prompt diagnosis, the first step
in getting appropriate care. For example, a commonplace healthcare
 Produced by the Science   

algorithm was found to underestimate how sick Black people were. that you get funding, and tomorrow you start researching,” she says.
This algorithmic bias meant Black people would be less likely to “Much of the work around equity and community engagement involves
receive necessary treatment than white people scored with that relationship building, so we have spent time doing that.”
same algorithm. In an NIH-funded collaboration with Morgan’s Center Crucially, active community engagement helps CUHE pressure-test
for Equitable Artificial Intelligence and Machine Learning Systems their research plans and get buy-in from residents. Brown explains that
(CEAMLS), RCMI@Morgan researchers are working to reduce the role Black communities are often skeptical of institutions that are likely to
that algorithmic bias plays in health disparities. misunderstand their struggles—especially older institutions that were
“Structural systemic complicit in segregation or
issues around education, slavery. But as a local and
wealth-gathering, cultural, historically Black university,
racial, and ethnic biases Morgan is uniquely
drive what we see,” says positioned to operate within
Kim Dobson Sydnor, the community as a trusted
director of CUHE. These partner. CUHE researchers
are the root of racial health engage with community
inequalities, she says, ambassadors who let
which don’t exist naturally. them know what residents
They’re the consequence need and what they are
of injustice. “The genetic concerned about, and the
differences between one Family League of Baltimore
person and another based is one group that helps
on the color of skin are just facilitate that interaction.
minuscule,” says Sydnor. The work of CUHE isn’t
“So, there’s no reason limited to health issues,
why that should exist at a Artwork calling to dismantle systems that hurt the “Black butterfly,” a term coined by Dr. Brown to refer to though. The center partners
Baltimore’s black neighborhoods.
population level.” with researchers from
across Morgan on a range of topics, from exploring tiny smart homes
Partnering with the community for seniors in urban settings to enhancing educational equity through
Although CUHE is locally focused, Baltimore’s issues are not unique. family engagement.
“Urban conditions unfortunately tend to replicate themselves,” says Community engagement is also central to RCMI@Morgan.
Sydnor. “The history may be a little different, the context may be a little Established in 2019 to tackle health issues that disproportionately
different, but the outcome tends to be the same.” affect minority communities, RCMI@Morgan is funded by the National
PHOTO: PROVIDED BY MORGAN STATE UNIVERSITY

Researchers at CUHE study many social determinants of health.
Brown uses historical data, for example, to analyze the influence
urban spaces have on health. Other projects examine the relationships
between variables such as food deserts, mental health in the public
school population, and maternal and child health.
“I think that at the end of the day, it is about being a partner in
improving community health and using research as a tool to do that,”
says Sydnor. She explains that the center is still very new. “It’s not
Institute of Minority Health and Health Disparities. Paul Tchounwou,
director of RCMI@Morgan, says the center is developing “a biomedical
research ecosystem to foster innovative biomedical, social, behavioral,
and clinical research, and to improve minority health, reducing or
eliminating health disparity in our urban communities.”
One core function of RCMI@Morgan is connecting local residents
with researchers to help them define scientific projects and disseminate
results to the community. MorganCARES is one such project, focusing on
 Sponsored Feature
smoking prevention and cessation. “We find that the approach of really
engaging the community was more effective than many other programs
set up in hospitals and clinics,” says Tchounwou.
Collaborating with facilities and data
To address health disparities, both CUHE and RCMI@Morgan
collaborate with other Maryland institutions, such as Johns Hopkins
University. RCMI@Morgan has ties to the University of Maryland (UM),
The four aims of RCMI@Morgan
This center’s goal is to address health
issues that disproportionately affect
minority communities.
1. Enhance Morgan’s health disparities research
infrastructure and capacity within the areas of
basic biomedical and behavioral/public health
research in infectious diseases including sexually
transmitted diseases, cancers and diabetes,
addiction research and abuse prevention, social
determinants of health, food security, and health
informatics.
2. Enhance high-quality research including
translational research on urban health and
health disparities through increased external
funding, publications, and scientific services to the
community.
3. Facilitate collaborations between basic biomedical
and social/behavioral faculty researchers and create
a collaborative and supportive environment for
faculty career development, especially for new and
WWW.MORGAN.EDU/RCMI/SPECIFIC-AIMS
early-career faculty.
4. Build sustainable partnerships with two local
research-intensive institutions, Johns Hopkins
University and the University of Maryland Baltimore,
as well as local government and community-based
organizations dealing with health disparities.
Produced by the Science   

and CUHE is also looking to partner with other HBCUs in the state.
So far, projects include investigating disparities in the rates of
cancer, COVID-19, and neurodegenerative diseases, and often bring
together teams of researchers with different strengths. “Many of these
projects have collaborations, not just from the biological aspect, but
the public health aspect, and also the engineering aspect in terms of
the technology being developed to address some of the key research
questions,” Tchounwou says.
RCMI@Morgan has a core facility with state-of-the-art equipment,
but if different instrumentation is needed, Morgan researchers have
access to facilities at UM as well. Some UM School of Medicine faculty
serve on the center’s advisory board. “We leverage research resources
to not duplicate effort and to make sure we create synergy that will
benefit all of our institutions,” says Tchounwou.
In addition to academic partnerships, CUHE has built relationships
with city and state entities to investigate mental health and
substance use issues in public schools and to analyze data to assess
whether health initiatives are actually working. Sydnor says that
government staff often ask for advice from CUHE researchers on
these programs.
The center also provides professional development opportunities
and training to junior researchers, such as postdoctoral fellows and
assistant professors. The goal is to learn how to shape pilot projects
and obtain grant funding for them from the National Institutes of Health
and the National Science Foundation.
It all comes down to rolling up sleeves and just doing the hard work—
together. “It is just astonishing how many people continue to work in
silos around this work,” says Sydnor. “We’re not going to be the problem
solvers on our own.” What’s needed is collaboration and education.
“This is a partnered enterprise,” she says. “It’s only going to be
successful to the extent that we can work with a lot of different folks.”
Sponsored by
Growing the Future, Leading the World
 EDITO RIAL
Strengthen the case for DEI
T
hree years ago, addressing racial justice in the
United States moved firmly into the mainstream.
Following the murder of George Floyd, the ongo-
ing struggle for social justice was again laid bare,
and pledges to improve diversity, equity, and in-
clusion (DEI) began sprouting everywhere. Now,
the pendulum is swinging back on these com-
mitments. A backlash against DEI initiatives is rising
across all sectors, especially at the state level. Last year’s
decision by the US Supreme Court to strike down the
consideration of students’ racial status in college admis-
sions has emboldened many who oppose any advance-
ment of DEI. Although there has been specific atten-
tion to higher education, other sectors have also been
attacked. The retreat includes recent
anti-DEI legislation that would affect
structures, programs, practices, and
curricula that aim to support success
for all, including persons who have
“We must all
been historically excluded from or
marginalized within science, technol-
move from
ogy, engineering, mathematics, and
medicine (STEMM). Before this back-
valuing what
lash worsens, DEI advocates, the sci-
entific community, universities, and
we measure
federal agencies need to collectively
call out the dangers of setting aside
to measuring
DEI and come up with robust ways to
demonstrate its value to society. what we value.”
DEI in STEMM is an element of
its excellence. It encompasses issues
beyond race and ethnicity, including those of gender,
disability, and other aspects of identity that have his-
torically been used to constrict opportunities. STEMM
should ideally benefit all of society. However, this will
not happen until the country creates a STEMM com-
munity as diverse as the population it should serve.
Perspectives and experiences matter in what one cre-
ates. Limiting opportunity to participate in scientific
and technological innovation means that science fails
to meet many of society’s needs. For example, one
study reports that women researchers in the United
States are more likely to make innovations that ben-
efit women as a whole but are less likely to participate
in commercial patenting. Their relative absence is a
loss for women and for the world economy. Critics im-
ply that DEI promotes mediocrity, whereas research
shows the exact opposite.
Turning a blind eye to these benefits, at least 38 states
in the US have recently introduced new anti-DEI leg-
PHOTO: CHRIS FLYNN
SCIENCE science.org
islation (11 are now law, another two await governors’
expected endorsement). The measures include prohibit-
ing public colleges from having DEI offices and staff,
banning mandatory DEI training, and forbidding the
use of diversity statements. Anti-DEI proponents mis-
leadingly claim that these practices break antidiscrimi-
nation laws and misuse public funds. However, teaching
the facts of our history and evidence of inequality is Shirley Malcom
central to academic freedom and every institution’s
is a senior advisor
mission. Evidence and truth become casualties.
and director of
These movements at the state level presage a threat
the STEM Equity
to federal DEI programs. STEMM college programs in
both public and private institutions receive funding Achievement (SEA)
from specific federal agencies, including the National Change initiative
Science Foundation, National Insti- at the American
tutes of Health, and NASA, to broaden Association for
participation, address health dispari- the Advancement
ties, and expand interest in STEMM.
Some of these programs have been in
of Science (AAAS,
the publisher
place since the Civil Rights era. So far,
their missions remain intact, but it
of Science),
Washington, DC,
is not clear for how long before aca-
demia and federal agencies follow the
USA. smalcom@
aaas.org
states and corporate America’s lead in
backing away from DEI.
To push back against the critics, it
is important to remember why DEI ef-
forts are so important and agree on
the best ways to gauge their success.
The success of STEMM is measured
not only by publications and head
counts of underrepresented groups in STEMM fields
but also by creating a culture of inclusion and respect.
It’s crucial not to lose sight of the kind of scientific com-
munity the federal government aims to support. In this
vein, DEI advocates, the scientific community, universi-
ties, and federal agencies must together consider where
processes can be modified within institutions to remove
barriers to participation by all. Importantly, measure-
ments on many fronts—from innovation output to
impact in the classroom—must be made to monitor
progress toward this goal. For example, if improving
teaching, adding relevant context, and supporting a
sense of belonging within introductory courses in-
crease retention of diverse STEMM students, then these
should be carefully tracked. Such efforts will provide
the data needed to answer critics and strengthen the
case for DEI. We must all move from valuing what we
measure to measuring what we value.
–Shirley Malcom
10.1126/science.adp4397
29 MARCH 2024 • VOL 383 ISSUE 6690 1395
 “ The destruction of known [virus samples] would no

NEWS ”
longer eliminate the threat of smallpox reemergence.
A U.S. National Academies of Sciences, Engineering, and Medicine report urging better
vaccines and treatments because synthetic biology could create the smallpox virus. Since the disease
was eradicated, virus samples have been guarded at two facilities in the U.S. and Russia.

downturn in domestic postdoc numbers

IN BRIEF Edited by Jeffrey Brainard underscores concerns that the academic
community is facing a postdoc shortage and
that early-career scientists are increasingly
favoring higher paid positions in industry.
Several groups have recommended how the
government could better support postdocs.

Superconductor misconduct found

| A physicist who
M AT E R I A L S S C I E N C E
claimed to have discovered room-
temperature superconductors engaged
in research misconduct, an investigative
committee concluded last week. Ranga
Dias of the University of Rochester gained
worldwide acclaim with a pair of papers
in 2020 and 2023 in Nature describing
two compounds that appeared to conduct
electricity without resistance near ambient
temperature. Both papers have since been
retracted, and concerns about the research
South Korean President Yoon Suk Yeol (center) and EU leaders announced a research funding deal. prompted the university to investigate.
Three internal inquiries didn’t find wrong-
FUNDING doing, but a review by outside experts did,
according to a university spokesperson. The
South Korea joins Horizon Europe university declined to release the investiga-
tion’s report. In an email to Science, Dias
called the report’s conclusions “baseless.”

S
outh Korea will participate in the €95.5 billion ($104 billion)
He remains on the faculty but is not teach-
Horizon Europe R&D program, the first East Asian country to ing, and all his students were removed
do so, the European Commission announced last week. South from his group last year.
Korean scientists will compete for grants on an equal footing
with their European counterparts; in return, South Korea will Edited pig organs transplanted
contribute an as-yet-undisclosed amount to the 7-year program, BIOMEDICINE | In a pair of firsts, research
which expires in 2027. The deal comes less than a year after New Zealand teams last week reported transplanting
became the first country from outside of the European region to join into humans a kidney and a liver from pigs
Horizon Europe, as the European Union seeks to internationalize the genetically modified to reduce the risk of
immune rejection. The kidney, transplanted
program. Canada’s official entry is also pending; Singapore and Japan
into a 62-year-old man at Massachusetts
are in early stages of discussions to join. Some EU scientists worry the General Hospital, came from a pig whose
program’s expansion could make it harder for researchers from smaller genome had received a whopping 69 edits
European countries to successfully compete for grants. by researchers using the CRISPR genome
editor. The edits included adding human
genes to aid circulation, STAT reported. The
other transplant, at Xijing Hospital of the
8% decline, from about 30,000 to 27,000, Air Force Medical University in China, was
Domestic U.S. postdocs drop is the largest year-to-year percentage-wise the first involving a pig liver. The organ,
PHOTO: KYODO VIA AP IMAGES

WO R K F O R C E | The number of U.S. citizens
and permanent residents working as post-
doctoral researchers in the United States,
especially in the biological and biomedical
sciences, fell sharply in 2022, the National
Science Foundation said last week. The
drop in the history of the agency’s Survey
of Graduate Students and Postdoctorates
in Science and Engineering, which has
collected data since 1980. Meanwhile, the
number of postdocs with temporary visas
increased from about 34,000 to 35,000. The
from a pig modified using other genetic
techniques, was placed in a person “clini-
cally dead” from brain damage, and was
removed after 10 days, Nature reported.
The surgeons saw no signs of immune
rejection, and the liver produced bile as

1396 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


hoped. The two results raised optimism
that the long-struggling field of xeno-
transplantation can help alleviate a short-
age of human organ donations. Kidneys
are in highest demand.
Asian nations urged to share data
C L I M AT E S C I E N C E | Seven countries in the
Hindu Kush Himalaya region must expand
scientific collaborations and data sharing
to address “enormous and growing” risks
posed by climate change to water supplies
from three major rivers supporting nearly
1 billion people, new reports say. All three
rivers—the Indus, the Ganges, and the
Brahmaputra—originate in the region’s icy
mountain ranges, where rapid warming is
accelerating the melting of some glaciers
and altering precipitation patterns. Those
changes, together with growing popula-
tions and increased demand for water,
have sobering implications for the nations
along the rivers—Afghanistan, Bangladesh,
Bhutan, China, India, Nepal, and Pakistan—
IMAGE: EHT COLLABORATION
say the authors of the reports, released
last week by the International Centre for
Integrated Mountain Development and the
Australian Water Partnership. Risks include
increased water shortages in some areas
and flooding in others. Although the seven
SCIENCE science.org
nations have agreements to collaboratively
manage the rivers, sharing of hydrological,
environmental, and socioeconomic data is
insufficient, the authors say. National secu-
rity concerns have led some governments to
keep some data private.
Trust in science rose despite Trump
P U B L I C O P I N I O N | U.S. residents’ faith
in scientists grew during the presidency
of Donald Trump despite his attacks on
individual scientists and federal research
agencies, a study says. But doubts grew
as well. Drawing on results from surveys
taken in late 2016 and late 2020, the
authors speculate that the COVID-19 pan-
demic drove people in the United States
to seek out reliable information about the
disease. The share of respondents holding
science in high or very high regard jumped
from 22% in late 2016 to 57% at the end of
2020, they report, whereas those express-
ing low or very low trust grew from 2%
to 13%. The size of the pool with neutral
views shrank dramatically, from 75% to
29%. The bimodal increases occurred
across the political spectrum. Conservative
Republicans, for example, added
20 percentage points on the low end of
a five-point scale of trust and 21 points
ASTRONOMY
Black hole pic hints
at Milky Way jet
S
trong magnetic field lines in
the first polarized-light image
of the supermassive black hole
at the center of the Milky Way
suggest it could be emitting a
high-energy jet of matter. The image,
produced by the Event Horizon
Telescope (EHT) and released this
week, depicts matter swirling around
the black hole, known as Sagittarius
A* (Sgr A*). The field lines are revealed
by gas spiraling around them, which
emits polarized light; a strong field is
thought to be key to forming a jet. The
image resembles one provided by the
EHT in 2021 of the black hole at the
center of the galaxy M87; that object,
1600 times more massive than Sgr
A*, emits a jet nearly 5000 light-years
long. Sgr A* is much less active. But
researchers now hope future upgrades
to the EHT—a worldwide array of radio
dishes that in 2019 released the first
image of a black hole—will reveal the
Milky Way, too, has a jet.
on the high end. The study appeared in
Science and Public Policy.
Bird virus appears in U.S. cows
I N F E CT I O U S D I S E A S E S | The flu virus
that has wreaked havoc on birds around
the world appears to have surfaced in U.S.
cows. Texas authorities said this week
they found the H5N1 strain, and the U.S.
Department of Agriculture said testing
showed “consistency” with the clade of
virus that has killed millions of wild birds
and forced culling of poultry. Kansas and
New Mexico also reported infected cows,
which showed limited symptoms. There
were no plans to cull them; pasteuriza-
tion of their milk kills the virus. Dead
birds were found on some of the affected
dairy farms. H5N1 has infected dozens of
other mammalian species but with limited
spread within them. Officials have said the
risk to humans remains low. Preliminary
studies have shown no signs that virus in
the cows has changed and become more
transmissible to humans, the U.S. National
Veterinary Services Laboratories says. But
scientists worry the unprecedented amount
of spread to mammals offers the virus
more opportunities to adapt and become
more transmissible and dangerous.
29 MARCH 2024 • VOL 383 ISSUE 6690 1397
NE WS
IN DEP TH
Among the 10 bases in this RNA molecule from SARS-CoV-2, two have modifications (teal) that likely affect the RNA’s function.
By Elizabeth Pennisi
I
n 2021, as a new kind of vaccine began
to protect millions from COVID-19,
many people heard about messenger
RNA (mRNA) for the first time since
high school biology. The molecule, which
transfers DNA’s code out of a cell’s nu-
cleus to guide protein production, is just one
of several types of RNA that profoundly affect
how cells function. Now, the U.S. National
Academies of Sciences, Engineering, and
Medicine (NASEM) is calling for a 15-year
effort to study and catalog the thousands of
RNA sequences made in cells and the ways
they can go wrong.
What researchers are calling an “RNome”
project could spur the development of new
disease treatments, vaccines, and agricultural
innovations. “There’s enormous power in
RNA,” says geneticist Fred Tyson, a scientific
program director at the National Institute
of Environmental Health Sciences (NIEHS),
which cosponsored the report. Understand-
ing RNA biology “can really alter our lives.”
But so far, no funding has been committed to
the effort, which will cost billions of dollars.
Far from being just a genetic messenger,
RNA serves diverse roles. Transfer RNA deliv-
ers amino acids to the ribosome—also partly
composed of RNA—that assembles proteins.
Other RNAs silence genes or affect their activ-
ity. How these RNAs behave reflects not just
their sequence, but also how they are modi-
fied before being exported from the cell’s nu-
cleus, often by the addition of chemical side
groups. If RNA is a house, “modifications are
windows and electrical outlets” that make it
function, says Stefanie Kaiser, an analytical
chemist at Goethe University Frankfurt.
“Most diseases are due to some dysregula-
tion of RNA,” including certain muscular dys-
trophies and cancers, says Vivian Cheung, an
1398 29 MARCH 2024 • VOL 383 ISSUE 6690
CELL BIOLOGY
Massive RNA sequencing
effort proposed
U.S. plan would harness the “RNome” for
medicine and more—but funding is uncertain
RNA biologist at the University of Michigan.
Host cell RNAs are also co-opted by HIV and
other viruses. “We know [RNA] screws up,”
says Peter Dedon, a biological engineer at the
Massachusetts Institute of Technology. “But
we don’t know how it operates very well.”
In a 2021 paper, Cheung and her colleagues
called for large-scale studies to sequence
RNA, including all modifications. After a
workshop on the topic, sponsored by NIEHS
and the National Human Genome Research
Institute (NHGRI), Cheung, with those in-
stitutes and the Rhode Island–based Warren
Alpert Foundation, asked NASEM to develop
a road map for cataloging all naturally occur-
ring RNAs and their modifications.
“This report is very much modeled on
the NASEM report that initiated the Human
Genome Project,” completed in 2003, says
Cheung, who wasn’t involved in the new re-
port’s drafting. But an RNome is more com-
plex than a genome. For one thing, frequent
modifications to RNA mean there will be no
fixed, reference sequence like the one pro-
duced for the human genome. For a given
RNA molecule, researchers will have to docu-
ment “not only the sequence, but also the type
and location of the modification[s],” adds
Chuan He, a chemical biologist at the Uni-
versity of Chicago who runs an RNA research
center for NHGRI.
For that reason, NASEM proposes not a
singular sequencing effort, but a project fo-
cused on “enabling technologies and infra-
structure,” says Carolyn Hutter, director of
genome sciences at NHGRI. It suggests that
in 15 years, the project should have mapped
modifications under different conditions in
model organisms such as worms and cata-
loged them in human diseases.
Though the report focuses on human
heath, it notes that understanding the role
of RNA modifications in plants could lead to
new ways to improve disease or drought re-
sistance and increase yield in crops.
NASEM suggests the National Institute of
Standards and Technology develop synthetic
modified RNA molecules that research-
ers can use as references to test sequencing
technologies and lab protocols. The National
Institutes of Health (NIH) should establish
a centralized database for RNAs, along with
rules about how to describe them.
For lack of a practical way to decipher
RNA directly, researchers currently have to
convert it back into DNA to sequence it, a
transformation that strips away any modifi-
cations. The report calls for improvements in
RNA sequencing and in chromatography and
mass spectrometry, which can identify the
chemical modifications.
Sequencing company Oxford Nanopore
Technologies has developed biochemical kits
and software for sequencing RNA directly,
and says that by 2025, its sequencers will also
reliably detect three of the most common
kinds of modifications to each base.
Dedon estimates that an RNome like
the report describes would require about
IMAGE: JANET IWASA AND RACHEL TORREZ; NATIONAL ACADEMY OF SCIENCES
$30 billion—10 times more than the Human
Genome Project. Cheung envisions a more
modest effort costing $300 million per year
for the next 5 years. But the funding is not
assured, given budget constraints facing U.S.
federal agencies, notes Eric Green, director
of NHGRI. Last year, a more modest RNA
sequencing proposal was passed over by a
trans-NIH funding initiative, he notes.
Cheung is still hopeful. In Rhode Island,
one senator, Jack Reed (D), has begun to
push for an RNome. It makes sense for poli-
ticians to back such a project, says David
Hirsch, chairman of the board at Warren Alp-
ert. “All you have to do is look at … the mRNA
vaccine,” he says. When it comes to ways to
harness RNA, “that’s the tip of the iceberg.” j
science.org SCIENCE

EDUCATION
West Virginia OKs intelligent design in schools
Governor signs bill allowing teachers to discuss antievolutionary “theories”
By Jeffrey Mervis
I
n 2005, then–U.S. District Court Judge
John Jones ruled that intelligent design
(ID)—the idea that life is too complex
to have evolved without nudging from
supernatural forces—cannot be taught
in public school biology courses because
it is not a scientific theory. This month, the
West Virginia legislature found a work-
around, approving a bill that doesn’t name
ID but will nevertheless allow public school
teachers there to discuss it in the classroom.
The bill, which the state’s governor, Jim
Justice (R), signed on 22 March, is the latest
example of what Nicholas Matzke, an evolu-
tion educator at the University of Auckland,
has called “legislation that avoids mention-
ing creationism in any of its varieties but
advances creationist antievolutionism.” It
comes nearly 2 decades after Jones, now
president of Dickinson College, told the
Dover, Pennsylvania, school board that its
policy on teaching ID violated the so-called
“establishment” clause of the U.S. Constitu-
tion, which bans the government from tak-
ing action favoring any religion.
Last year, the West Virginia Senate passed
a bill sponsored by State Senator Amy
Grady (R) that would have specifically al-
lowed teachers to talk about ID “as a theory
of how the universe and humanity came to
exist.” But the measure died in the House
of Delegates.
This year, Grady reintroduced the bill but
revised it to remove the words “intelligent
design.” Her colleagues approved this new
version in late January and the House fol-
lowed suit on 9 March.
The one-sentence legislation now declares
that “no local school board, school super-
intendent, or school principal shall prohibit
a public school classroom teacher from dis-
cussing and answering questions from stu-
dents about scientific theories of how the
PHOTO: WILL PRICE/WEST VIRGINIA LEGISLATURE
universe and/or life came to exist.”
“It seems innocuous,” says Glenn Branch
of the National Center for Science Education,
a nonprofit that tracks attacks on the teach-
ing of evolution and climate change. “But
Grady said during floor debate that it would
allow for discussion of intelligent design.”
Grady did so after being grilled in late
January on the bill’s language and intent by
State Senator Mike Woelfel. “Is there a defi-
nition of scientific theory in this bill?” asked
SCIENCE science.org
Woelfel, a lawyer who leads the contingent
of three Democrats in the 34-member state
Senate. “There is not,” replied Grady, a vet-
eran elementary school teacher who chairs
the Senate Education Committee. But then
she offered her own.
“The definition of a theory is that there is
some data that proves something to be true,”
Grady explained. “But it doesn’t have to be
proven entirely true.”
“Can ID be taught by a teacher?” Woelfel
then asked. Grady conferred with the Sen-
ate’s general counsel, then replied: “It would
not prevent them from teaching it.”
Her bill is part of the “third wave of anti-
evolutionism” legislation, as Matzke puts it,
that followed the 2005 Dover ruling. Three
southern states—Mississippi in 2006, Louisi-
ana in 2008, and Tennessee in 2012—passed
what Matzke calls “stealth creationism” bills
that avoid mentioning religion explicitly and
do not require teachers to talk about ID.
The Mississippi law also gives teach-
ers permission to discuss student que-
ries on the origins of life. But unlike the
West Virginia legislation, it doesn’t use the
phrase “scientific theories” in describing
such conversations.
During debate on the bill, Grady said she
introduced it to help teachers who fear re-
prisal if they mention ID when responding
to questions from students. “A lot of teach-
ers have told me that they’re not comfortable
bringing it up,” Grady said, without citing
any examples. “I say, just let them teach.”
Branch says West Virginia teachers can
already discuss scientific theories of how
West Virginia legislators overwhelmingly approved a bill intended to allow the teaching of intelligent design.
NE WS
life evolved. Ken Miller, an emeritus pro-
fessor of biology at Brown University who
testified as an expert witness for the plain-
tiffs in Kitzmiller v. Dover, believes the leg-
islation’s vagueness is dangerous as well
as unnecessary.
“Using the phrase ‘scientific theories’ with-
out defining it creates a lot of uncertainty
and opens the door to potential litigation,”
Miller says. “I see it as an attempt to confuse
people and provoke a controversy.” The 2005
decision split the community and ultimately
cost the Dover school board $1 million in le-
gal fees, he notes.
Even the Discovery Institute, which touts
itself as “the intellectual home of intelligent
design,” thinks the pending West Virginia
law could complicate its advocacy of ID even
though it is not mentioned by name. “When
ID gets brought into public schools through
curricular policies or legislation, this leads
to politicization of the scientific debate …
and witch hunts against ID-friendly scien-
tists,” Casey Luskin, an associate director
of its Center for Science and Culture, wrote
this month.
Woelfel predicts the West Virginia mea-
sure will be challenged in court, where he
thinks its intent to open the door to ID will
become clear.
“I had 12 years of Catholic school educa-
tion teaching me about the existence of a cre-
ator of heaven and Earth,” he says. “But after
my election [to the state Senate] I took an
oath to uphold the [U.S.] Constitution. And
it’s fairly obvious to me that this bill runs
afoul of the establishment clause.” j
29 MARCH 2024 • VOL 383 ISSUE 6690 1399
NE WS | I N D E P T H
CLIMATE CHANGE
Firms aim to capture carbon in the oceans
Schemes to process seawater may be cheaper than snagging carbon dioxide from the air
By Robert F. Service
E
very year, hundreds of container ships
slide into the Port of Los Angeles, the
busiest in the Western Hemisphere.
Belching carbon dioxide (CO2), they
deliver some $300 billion in goods to
trucks and railcars that add their own
pollution to our warming planet.
But one long gray barge docked at the
port is doing its part to combat climate
change. On the barge, which belongs to
Captura, a Los Angeles–based startup, is
a system of pipes, pumps, and containers
that ingests seawater and sucks out CO2,
which can be used to make plastics and fu-
els or buried. The decarbonated seawater
is returned to the ocean, where it absorbs
more CO2 from the atmosphere, in a small
strike against the inexorable rise of the
greenhouse gas.
After a yearlong experiment with
the barge, which is designed to capture
100 tons of CO2 per year, Captura is plan-
ning to open a 1000-ton-per-year facility
later this year in Norway that will bury the
captured CO2 in rock formations under the
North Sea. Equatic, another Los Angeles–
based startup, is launching an even larger
3650-ton-per-year ocean CO2 capture plant
this year in Singapore, and other compa-
nies are planning demos as well.
How to siphon carbon from the seas
Earth’s oceans naturally concentrate carbon dioxide (C02) from the atmosphere. To extract it, Captura, a California startup, uses renewable electricity to change the water
chemistry. The CO2 bubbles out and is then captured and stored beneath the seabed. The processed water is returned to the ocean, where it can absorb more CO2.
A small amount of seawater
is diverted to a unit that uses
electricity to split water into
an acid and base.
Filter
Seawater is pulled in
and flows continuously
o tinuously
ontinuously
he plant.
through the pla
plantt
Dissolved carbon
1400 29 MARCH 2024 • VOL 383 ISSUE 6690
Although these ocean capture ventures
deploy different chemical processes, most
are designed to use renewable electricity
to extract CO2 from seawater. “We’re kind
of doing the reverse of burning fossil fu-
els,” says Nicholas Ward, an oceanographer
at the Pacific Northwest National Labora-
tory who is working with Ebb Carbon, a
California-based startup. “We’re speeding
up the rate that the ocean can take in car-
bon dioxide.”
Proponents say capturing CO2 from the
ocean should be easier and cheaper than
a seemingly more direct approach: snag-
ging it directly from the air. Direct air
capture, which relies on fans to sweep air
past absorbent chemicals, currently costs
between $600 to $1000 per ton of CO2 re-
moved, largely because atmospheric CO2
is so dilute, making up less than 0.05% of
the air by volume. Earth’s oceans, in con-
trast, hold the gas at a concentration nearly
150 times higher, and absorb roughly 30%
of all CO2 emissions each year.
Because seawater concentrates CO2 nat-
urally, “it’s much more efficient to work
with the ocean,” says Ruben Brands, CEO
of SeaO2, a Dutch ocean carbon capture
company. Companies say they should ulti-
mately be able to capture CO2 at $100 per
ton, or less, meeting a U.S. Department of
Energy target for 2032.
Acid is added to
_ _
+ +
seawater, converting
dissolved carbon
into CO2 gas.
CO2 Membranes
Some startups are trying to harness
the ocean by adding alkaline rocks di-
rectly into the seas to promote the con-
version of CO2 to bicarbonate ions, which
helps the ocean retain the carbon (Science,
1 December 2023, p. 988). Others are
burying organic matter in the deep ocean
(Science, 5 January, p. 11). But ocean capture
is moving faster than other ocean CO2 re-
moval approaches, says Sifang Chen, man-
aging science and innovation advisor for
Carbon180, an environmental group that
tracks ocean carbon removal efforts.
That’s partly because the systems can
usually be bolted on to desalination plants,
wastewater treatment facilities, and other
large water-processing infrastructure. “The
technology is scalable,” says Erika La Plante,
a geochemist at the University of California,
Davis who also heads Equatic’s carbon re-
moval monitoring and verification efforts.
Captura’s process starts by piping sea-
water to a reactor that uses electricity to
split a few of the water molecules into
positively charged protons and negatively
charged hydroxyl ions. Those compounds
react with sodium and chloride ions in the
water to produce hydrochloric acid and
sodium hydroxide, a base. The acid reacts
with bicarbonate ions in the water, caus-
ing CO2 to bubble out into a storage tank.
The base is then added to the seawater,
CO2 Base is returned
to the seawater to
neutralize acidity.
Pump
GRAPHIC: N. BURGESS/SCIENCE
CO2 is extracted Seawater with restored
using membranes CO2 absorption capacity
and a vacuum pump. is returned to the ocean.
science.org SCIENCE
neutralizing the acid before the effluent is
discharged back into the ocean, ready to
absorb more CO2 (see graphic, p. 1400).
Oldham says the process doesn’t add any-
thing into the ocean environment that
wasn’t already there. “Our only output is
CO2 and decarbonized water.”
Equatic takes a different tack, although
its process also starts with an electricity-
fed reactor that converts seawater into
acids and bases. Alkaline rocks are added
to neutralize the acid, whereas the base
causes dissolved CO2 to form bicarbon-
ate, and nudges magnesium and calcium
ions to precipitate out as solid carbon-
ates, removing CO2. The carbonates can be
used as ingredients for cement and other
industrial products, and the reactor also
produces hydrogen gas, a valuable green
energy source, and free CO2. Last year,
Boeing inked a 5-year, $50 million deal
with Equatic to buy 62,000 tons of cap-
tured CO2 and 2100 tons of hydrogen to
make sustainable aviation fuel.
One challenge for all these approaches is
determining exactly how much CO2 the de-
carbonated water absorbs and at what rate,
figures that are critical for selling carbon
credits to companies looking to offset their
emissions. To avoid that uncertainty, Equatic
plans to pipe its CO2-stripped seawater to
the top of a 13-meter-high cooling tower and
drop it through the air, where it absorbs at-
mospheric CO2 in a way that can be precisely
measured before it is returned to the ocean.
Still, the absorption rates will vary
based on where the water is discharged,
how readily it mixes with surrounding wa-
ter, and whether natural processes push
it down to the ocean floor where it can’t
interact with the atmosphere, Chen notes.
“That makes site selection a really impor-
tant part of this process,” she says.
Ocean capture advocates are seeking more
government support. In the United States,
direct air capture plants earn a $180 tax
credit per ton of sequestered CO2. Ocean CO2
capture efforts currently don’t qualify. “A
similar tax incentive for water-based CO2 re-
moval is absolutely needed,” says Dan Deviri,
CEO of CarbonBlue, an Israel-based startup.
Even if the technology takes off, it will
have to scale up massively to make a dent
in offsetting global emissions. According
to the Intergovernmental Panel on Cli-
mate Change, by 2050 engineered carbon
removal efforts will need to remove some
5 billion tons of CO2 every year to limit the
global temperature increase to 1.5°C. So
far, the ocean capture companies are pull-
ing out only thousands of tons. Matthew
Eisaman, a physicist at Yale University and
chief scientist at Ebb Carbon, says, “We
have an enormous challenge ahead of us.” j
SCIENCE science.org
SCIENCE POLICY
NIH’s flat budget for 2024 will
force ‘very difficult decisions’
Belated Senate and House compromise eliminates policy
riders restricting certain studies
By Jocelyn Kaiser
C
ongress has given the National Insti-
tutes of Health (NIH) an essentially flat
budget of $47.1 billion, in a final 2024
spending bill that lawmakers approved
last week in time to avert a partial gov-
ernment shutdown. But researchers
dismayed by the bottom line were pleased
that several policy directives they opposed
have been stripped from the legislation.
Lawmakers on the Senate committee
that funds NIH say their bill provides a tiny,
$300 million bump, just one-third of the
$920 million increase requested by President
Joe Biden. The budget, part of a $1.2 trillion
package covering six federal agen-
cies that Biden quickly signed,
comes after years of generous “We are taking
increases for NIH. But because
the increase only partially makes
a step back.”
up for a mandated $678 million Mary Woolley,
drop in a pot of money known as Research!America
the 21st Century Cures Act fund,
which is separate from NIH’s base budget,
lobbyists note that the final bill actually im-
poses a $378 million cut on the agency.
The meager outcome was no surprise:
Once the president and Congress agreed to
tight spending caps in May 2023, the NIH
community started to prepare for little if any
new funding, says Jennifer Zeitzer, public af-
fairs director at the Federation of American
Societies for Experimental Biology (FASEB).
And with this fiscal year already half
over, Zeitzer notes, NIH Director Monica
Bertagnolli “and individual [NIH institutes]
will have to make some very difficult deci-
sions and on a very quick time frame.”
FASEB and other groups say they are
grateful Congress was able to settle on any
spending bill for NIH’s parent agency, the
Department of Health and Human Services
(HHS). The alternative would have been a
yearlong measure freezing NIH’s budget at
the 2023 level. Still, “We are taking a step
back,” Mary Woolley, president and CEO
of the biomedical research lobbying group
Research!America, said in a statement. “The
constraints these leaders contended with are
reflected in funding levels for these agencies
that lag inflation, need, and opportunity.”
The final budget agreement favors a few
research areas. The National Cancer Institute
would receive a $120 million raise for re-
search grants, Alzheimer’s disease research is
in line for a $100 million increase, and men-
tal health research receives $75 million more
than in 2023. But most of NIH’s 27 institutes
and centers will remain at roughly the 2023
funding levels. The budget for the Advanced
Research Projects Agency for Health created
2 years ago remains at $1.5 billion; Biden had
requested a $1 billion increase.
The bill does not include policy provisions
championed by the Republican majority in
the U.S. House of Representatives that would
have curbed diversity efforts at NIH and re-
stricted federal funding for fetal
tissue studies and research on
gender-affirming care. It drops
a proposal to bar HHS funding
for the EcoHealth Alliance, a
U.S. nonprofit that has four ac-
tive NIH grants and in the past
worked with virologists in Wu-
han, China, on studies some critics suggest
could have sparked the COVID-19 pandemic.
And it does not include a House-
proposed ban on HHS funding for so-called
gain-of-function (GOF) studies that could
make risky viruses more harmful to people.
Opponents of that language argued it could
have curtailed work on vaccines and low-
risk pathogens such as cold viruses.
The American Society for Microbiology,
which had raised concerns about the ban,
“is grateful that Congress chose not to cut off
specific areas of scientific research, instead
leaving it to those with the requisite exper-
tise to ensure that research is being con-
ducted safely,” says Allen Segal, the group’s
chief strategy and public affairs officer.
The society, however, has reserva-
tions about a request accompanying the
spending bill that HHS review all grants
screened for GOF work. “Every effort
should be made to protect the privacy of
researchers and ensure that this doesn’t
lead to unwarranted interference with the
scientific process,” Segal says. The White
House is drafting a new policy for what is
also known as “enhanced potential pan-
demic pathogen” research. j
29 MARCH 2024 • VOL 383 ISSUE 6690 1401

BIOMEDICINE
Transgenic marmosets mimic
Parkinson’s symptoms
New monkey models promise insight into early stages
of degenerative brain disease
By Dennis Normile
B
y the time a person shows symptoms
of Parkinson’s disease, neurons in a
part of their brain key to movement
have already quietly died. To learn
how this process unfolds, identify
warning signs, and test treatments, re-
searchers have long wanted an animal model
of the disease’s early stages. Now, they may
have one: a cohort of transgenic marmosets,
described at a conference on nonhuman pri-
mate models in Hong Kong last month.
The animals, which neuroscientist
Hideyuki Okano of Keio University and col-
leagues created using a mutated protein that
seems to drive Parkinson’s in some people,
closely mimic the disease’s onset and progres-
sion. And they have enabled Okano’s team to
identify what could be an early, predictive
sign of disease in brain imaging.
The model could be “transformative” for
Parkinson’s studies, says neurobiologist Peter
Strick of the University of Pittsburgh, who at-
tended the meeting, organized by the Hong
Kong University of Science and Technology,
Stanford University, and the University of
California San Francisco. “We desperately
need nonhuman primate models that re-
capitulate the natural onset and progression”
of conditions like Parkinson’s, he says.
Parkinson’s, which afflicts an estimated
8.5 million people, is thought to be trig-
gered by a combination of genetic and en-
vironmental factors, such as exposure to
toxic chemicals. It sets in as neurons that
1402 29 MARCH 2024 • VOL 383 ISSUE 6690
Marmosets mature faster than larger monkeys,
aiding studies of neurodegenerative disease.
produce the chemical messenger dopamine
in the substantia nigra, an area of the brain
that controls movement, die off. Early symp-
toms include tremors, muscle stiffness, and
hesitant motions. The disease can later affect
cognition and lead to dementia. Researchers
think one cause of neuronal death may be
abnormal versions of a protein called alpha-
synuclein that misfold and form toxic clumps
in the brain years before symptoms emerge.
Animal models of the disease exist, but
they don’t mimic the gradual die-off. Instead,
most use chemicals to destroy neurons in
rats or monkeys, mimicking advanced dis-
ease stages where damage to the brain has
already been done. To model earlier stages,
Okano’s team used a virus to deliver a gene
variant that produces a mutated alpha-
synuclein protein into marmoset embryos.
(Marmosets’ fast aging compared with other
monkeys allows for more efficient tracking
of progressive diseases.) The team implanted
the modified embryos into surrogate mothers
to produce transgenic animals.
Positron emission tomography scans of
the marmosets showed progressive loss of
dopamine-producing neurons. With age,
they began to show the tremors, decline
in physical activity, and hesitant move-
ments seen in human patients, and they
improved when given the Parkinson’s drug
levodopa. The animals even developed
rapid eye movement (REM) sleep behavior
disorder—the limb flailing and shouting
while dreaming that is an early warning
sign of Parkinson’s.
The work “is one giant step forward” to-
ward a nonhuman primate model that mim-
ics the entire degeneration process, says
neuroscientist Michele Basso, director of
the Washington National Primate Research
Center. (A Chinese research team reported
creating transgenic monkeys with an alpha-
synuclein mutation in 2014, but they did not
show the full range of Parkinson’s symptoms.)
Searching for early brain anomalies,
Okano’s team used functional magnetic reso-
nance imaging to measure changes in blood
flow indicating neural activity. The trans-
genic marmosets showed unusually high
activity in the pontine nucleus, another part
of the brain involved in motor control, even
before symptoms appeared. Separately, the
team found that humans with REM sleep
behavior disorder also have enhanced pon-
tine nucleus activity. This activity “could be a
preclinical biomarker of Parkinson’s disease,”
Okano concluded. Basso calls the finding “in-
teresting and novel.”
It’s not yet clear whether this potential bio-
marker might emerge earlier or prove more
useful than others under study in humans,
such as abnormal alpha-synuclein in skin or
spinal fluid samples, or other brain abnor-
malities detectable by imaging.
The monkeys also may not perfectly model
human disease. The virus that delivered the
mutant alpha-synuclein gene may have ran-
domly inserted multiple copies into the mon-
key chromosome, potentially producing the
protein in multiple organs. Those changes,
not seen in human disease, could skew stud-
ies testing the effects of potential therapeu-
tics. But given the current dearth of animal
models, Strick says researchers shouldn’t “let
perfect be the enemy of the good.” Okano de-
clined to comment because his group is pre-
paring a paper describing the research.
The Parkinson’s model joins other primate
models for neurological conditions. At the
same meeting, neuroscientist Guoping Feng
of the Massachusetts Institute of Technology
reported on his effort to produce animals ex-
hibiting a rare type of autism called Phelan-
McDermid syndrome by editing out a gene
called SHANK3 that is absent or mutated in
the disease. The animals developed social
and cognitive symptoms that improved after
gene therapy to reintroduce the gene. Illinois-
based Jaguar Gene Therapy will start a phase
1 trial of the approach in adults with Phelan-
McDermid syndrome later this year.
PHOTO: SAM OGDEN/SCIENCE SOURCE
Neuroscientist Mu-ming Poo of the Chinese
Academy of Sciences’s Institute of Neuro-
science predicts the next 5 years will bring
multiple monkey models with “clear impact
in preclinical tests of therapeutic approaches
for brain diseases.” The reports by Okano and
Feng, Strick adds, form “the leading edge of
what could be a revolution in the field.” j
science.org SCIENCE

GLOBAL HEALTH
Africa may lead rollout of long-lasting HIV drug
U.S. agency plans to make inexpensive prevention shots widely available for people at risk
By Jon Cohen
T
ools to fight HIV tend to come late to
sub-Saharan Africa, the region hardest
hit by the epidemic. After powerful,
lifesaving cocktails of HIV drugs came
to market in 1996, it took 7 years before
they began to reach large numbers of
people living with the virus there. When pills
to prevent, rather than treat, HIV infection
were introduced in 2012—a strategy known
as pre-exposure prophylaxis (PrEP)—Africa
was again slow to benefit.
But with the next revolution in HIV
prevention—an injectable, long-lasting ver-
sion of PrEP—Africans may actually soon
lead the pack. Not many people
in rich countries have started to
take this formulation, mainly
because of insurance hassles for
the expensive drug. But inject-
able PrEP is now on the cusp
of being widely introduced in
Africa, thanks to the President’s
Emergency Plan for AIDS Relief
(PEPFAR), a U.S. government
program, which has purchased
it at a steep discount.
“Over the next 2 years, we will
see more injectable PrEP use in
East and Southern Africa than
we’ll see in the U.S.,” predicts
Mitchell Warren, who heads
AVAC, an advocacy group for
HIV prevention. “That’s turning
history on its head.” This Kenyan clinic helped recruit people to test a new HIV-prevention drug.
PEPFAR had provided 24,000
doses of injectable PrEP in Zambia, Zimba-
bwe, and Malawi by 6 March and has plans
for an “aggressive scale-up,” says PEPFAR
head John Nkengasong. The drug has “the
potential to bend the curve on the annual
1.3 million new HIV infections globally,”
Nkengasong says, but the availability and
cost of injectable PrEP “are still a big con-
cern” and could limit its impact. (PEPFAR’s
own future funding has been in limbo, but
congressional negotiators last week agreed
to extend it through March 2025.)
PrEP first proved its worth as daily pills
made by Gilead, but they only work if people
take them consistently—which many find
difficult to do. The long-acting, injectable
PHOTO: SEARCH
version of PrEP, made by the pharmaceuti-
cal company ViiV Healthcare, contains the
antiviral cabotegravir (CAB-LA); a shot once
SCIENCE science.org
every 2 months suffices. Big international
trials that included participants in Africa
showed the drug to be more effective than
oral PrEP in men who have sex with men
and transgender women and, separately, in
women. The U.S. Food and Drug Administra-
tion approved CAB-LA in December 2021,
and the European Commission authorized it
in September 2023. Other long-acting inject-
ables that last 4 or even 6 months are now
being tested.
But in the United States, CAB-LA costs
more than $23,000 annually for the
2-month shots. Patients and health care
providers have had trouble getting their
health insurers to pay for injections, in part
because PrEP pills, now available from ge-
neric companies, cost as little as $300 per
year. As a result, only 11,000 people in the
U.S. had started to use CAB-LA by the end
of 2023, ViiV says.
ViiV was assailed by activists in 2022 for
not making inexpensive versions of CAB-LA
available in Africa, but the company has
stressed it is committed to selling CAB-LA
at “a nonprofit price” in low-income coun-
tries until a generic version is available.
“We seek to make our medicines widely
available to those who need it—regardless
of income or where they live, driven by pub-
lic health needs,” ViiV said in a statement
to Science. PEPFAR pays $30 per 2-month
dose, or $180 per year. There is also PrEP
in a vaginal ring version that works for
1 month and costs PEPFAR $13 each.
NE WS | I N D E P T H
A study in Uganda and Kenya, presented
at a U.S. HIV/AIDS conference this month,
suggested that given the choice, people at
risk of HIV chose CAB-LA over oral PrEP—
but simply having options increased PrEP
use overall. “There is much interest in inject-
able PrEP,” says Moses Kamya, an epidemio-
logist at Makerere University who presented
the findings. It’s “not just replacing oral
PrEP,” he adds. “It expands the pie.”
ViiV will have at least 1.2 million doses
of CAB-LA available for low- and middle-
income countries through 2025, and about
30% of those will go to PEPFAR. But so far,
not a single country in sub-Saharan Africa
has put in an order to purchase the drug with
its own money, says Linda-Gail
Bekker, who runs the Desmond
Tutu HIV Centre at the Uni-
versity of Cape Town. In South
Africa, which has the world’s
highest number of people living
with HIV and purchases more
oral PrEP than any country, an
investment analysis published
in December 2023 concluded
CAB-LA would not be cost ef-
fective, even at its discounted
price. “The modelers have fig-
ured that the bang is not worth
the buck,” Bekker says. She’s
heartened that at least PEPFAR
will soon start to offer CAB-LA
in South Africa.
For mass introduction, she
and others say, the price will
need to come down further. ViiV
has signed a voluntary licensing agreement
with the Medicines Patent Pool that makes it
possible to cut deals with companies that can
produce the drug more cheaply. Three gener-
ics manufacturers have licensed the drug,
but Warren estimates it will take at least
2 years before they can deliver. The generics
must first be shown to work as well as ViiV’s
drug, and they are unlikely to be as cheap as
pills, because injectable PrEP is more com-
plicated to manufacture.
Warren is optimistic that if injectable
PrEP is accessible in Africa, it will eventually
become popular, but he stresses that PrEP
pills and vaginal rings still have their place
as well. “There’s no such thing as ‘better’
PrEP,” Warren says. “Each one is different.
Now you can say: ‘Let’s talk about these dif-
ferent options.’” j
29 MARCH 2024 • VOL 383 ISSUE 6690 1403
NE WS | I N D E P T H
SOCIAL SCIENCE
States in India survey caste membership
Censuses, the first in a century, chart social changes and shape affirmative action programs
By Vaishnavi Chandrashekhar
T
he more than 1 billion Hindus living
in India belong to more than 4000
historical castes, a hereditary identity
that shapes social and economic status
and mobility. But it has been almost a
century since caste membership was
counted. India’s federal government, citing
logistical and political reasons, has been
reluctant to conduct another survey. Now,
states are taking things into their own hands.
What they are learning could help India bet-
ter address economic and health inequities
and allow social scientists to chart the
changes sweeping the country.
In October 2023, the northern state
of Bihar released findings from a year-
long survey that provided the first de-
tailed numbers on social composition
in decades. Other states are following
suit. Andhra Pradesh in January be-
gan its own such caste survey, sending
thousands of enumerators with lists
of previously known caste groups to
match with state residents. And last
month, the southern state of Telan-
gana passed a resolution to conduct a
“door-to-door” survey to gather com-
prehensive data on different caste
groups within its borders.
The trend reflects increasing de-
mand in the country for more detailed
data on caste, which is linked to po-
litical representation and affirmative survey, one of many efforts to update 100-year-old caste data in India.
action for historically disadvantaged
groups in public sector jobs and education,
including medicine, engineering, and science
colleges. But caste data are also politically
sensitive, says Sukhadeo Thorat, professor
emeritus at the Centre for the Study of Re-
gional Development at Jawaharlal Nehru
University (JNU). “Some people don’t want
caste disparities to be revealed,” he says, for
fear of revealing persistent discrimination.
India’s last caste-focused census, con-
ducted in 1931 by the British colonial
government, asked respondents about
their religion and caste. It produced de-
tailed numbers, including an overall
estimate of 4147 castes, but Indian na-
tionalists saw the exercise as a tactic to
stoke social divides. That concern kept
detailed caste documentation out of sub-
sequent censuses, including after In-
dia gained independence in 1947. In the
1404 29 MARCH 2024 • VOL 383 ISSUE 6690
1990s, an expansion of affirmative action
policies made caste data a political hot potato.
But many Indian sociologists and demo-
graphers say the time is ripe to update out-
dated data sets, as policymakers seek to bet-
ter target welfare and health programs and
state governments expand affirmative action.
“We are relying on almost 100-year-old data
for such an emotive and important topic,”
says Sonalde Desai, a demographer and pro-
fessor at the National Council of Applied Eco-
nomic Research.
Updating the century-old caste data will
be a challenge, experts say. There is the sheer
Three men from three different castes complete the Bihar caste
number of castes: more than 5000, apart
from innumerable subcastes, per some re-
cent estimates. Some people may also claim
multiple caste identities, Desai says, making
them harder to categorize.
In addition, any caste census must cap-
ture not just the size of a population, but its
socioeconomic condition. “Reality is inter-
sectional, and data has to bring that reality to
the forefront,” Thorat says.
Bihar’s census could become a model for
other states. The state trained and dispatched
270,000 enumerators and another 40,000
supervisors to gather and check the data of
130 million people, in two phases over almost
a year. Enumerators asked people to identify
as belonging to one of 214 castes listed in
state records. They were then prompted to
answer 16 questions on income, education
level, housing, and land ownership.
The Bihar government is already putting
the data to use. The census showed that tradi-
tionally privileged groups such as Brahmins
were overrepresented in government jobs
compared with their population, whereas
some historically disadvantaged castes such
as the Mushawars were underrepresented.
Disadvantaged castes comprised 84% of the
population, the survey found, and so the state
hiked the proportion of government jobs and
school and college places reserved for those
castes from 50% to 65%.
“The extent to which upper castes are
still better off is clear, and is greater than
expected,” says sociologist Satish
Deshpande, who taught at Delhi Uni-
versity until last year. The Bihar cen-
sus also revealed variations within the
disadvantaged castes, with one or two
of those communities consistently do-
ing better, he notes. “But those inter-
nal differences do not erase the larger
fact of caste inequality.”
It’s not just the most under-
privileged castes that stand to gain
from such surveys. The middle-tier but
powerful Maratha caste in the west-
ern state of Maharashtra demanded
affirmative action citing “social and
economic backwardness.” It led to a
hasty survey in January, in which enu-
merators were sent out at short notice
to cover a very large population. That
widely criticized survey found that
21% of the Maratha community lived
below the poverty line, allowing the
state government to reserve government jobs
for them ahead of elections.
Because caste varies by region, Surinder
Jodhka, professor of sociology at JNU, argues
PHOTO: KARISHMA MEHROTRA/THE WASHINGTON POST VIA GETTY IMAGES
surveys should be led by states. For now, a
federal effort seems unlikely. The Indian Na-
tional Congress, the major opposition party,
has called for a national caste survey, but the
ruling party continues to be opposed. Either
way, frequent surveys will reveal which com-
munities, or subcommunities, are doing well
and can be moved out of affirmative action
categories, and which ones still need those
benefits, he notes. “Data should be fluid and
dynamic, like the situation on the ground.”
At the same time, “Some people are con-
cerned that enumerating caste will solidify
identities,” Jodhka says. “And that may
happen, but the only way to get past that
is with data.” j
science.org SCIENCE
 FEATURES
Colliding muon beams
generate a muon and an
electron—possible signals of
a Higgs boson decay—amid
copious background particles
(gray) in a simulation.

THE DREAM MACHINE

An accelerator known as a muon collider could revolutionize
particle physics—if it can be built By Adrian Cho
BELL/UNIVERSITY OF TENNESSEE, KNOXVILLE
ILLUSTRATION: LAWRENCE LEE AND CHARLES

Y
oung people supposedly enjoy the
luxury of time, but perhaps not if
they’re particle physicists. For de-
cades, physicists have peered into
the universe’s inner workings by
smashing subatomic particles to-
gether at ever higher energies. But
the next highest energy collider
may not be built for 50 years. And
Tova Holmes, 34 and a particle physicist at
the University of Tennessee, Knoxville, wor-
ries her career could slip away before she
ever sees such a machine. “I will be defi-
nitely not still working, possibly not alive,”
Holmes says.
That’s one reason she and dozens of her
contemporaries are pushing to develop an
exotic new collider. The current highest
energy atom smasher, the 27-kilometer-
long Large Hadron Collider (LHC) at the
European particle physics laboratory,
CERN, fires protons into protons. In 2012
it discovered a long-sought particle called
the Higgs boson in the field’s greatest
triumph. CERN is now hoping to build a
much bigger, higher energy proton col-
lider, nearly 100 kilometers around—but

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1405

NE WS | F E AT U R E S
not until 2070 or 2080. So Holmes and oth-
ers want to explore an alternative: a col-
lider that would smash energetic muons,
heavier cousins of electrons, into equally
energetic antimuons.
A muon collider could be much smaller
and cheaper than a functionally equivalent
proton collider, advocates say. It could fit
on the 2750-hectare campus of the United
States’s dedicated particle physics lab,
Fermi National Accelerator Laboratory
(Fermilab), enabling the U.S. to reclaim
the lead in the continuing competition
for the highest energy collider. Most im-
portant, younger physicists say, it might
be built sooner than a more conventional
competitor, perhaps in as few as 25 years.
“If you want you can add 10 years to that,
that’s still a lot better than when I’m dead,”
Holmes says.
There’s just one catch: Nobody knows
whether a muon collider can actually be
built. That’s because unlike the proton
or the electron, the muon isn’t eternal,
but decays in just a fraction of a second.
“The challenge, if you want to capture it
in one word, is that the muon is unstable,”
says Sergo Jindariani, a particle physicist
at Fermilab. “So every stage of accelera-
tion has to be incredibly fast.” From gen-
erating the muons, to gathering them into
compact beams, to detecting the particles
produced in their collisions, the machine
presents novel technological challenges.
Nevertheless, some physicists, especially
in the U.S., are eager to tackle the task. In
December 2023, a federal advisory sub-
committee called the Particle Physics Proj-
ect Prioritization Panel (P5) laid out a road
map for the next decade of research in the
U.S. The P5 report calls for R&D on a muon
collider, stating, “This is our muon shot.”
But what exactly are physicists shooting
for and what obstacles must they overcome
to realize their dream machine?
AS ALBERT EINSTEIN OBSERVED, energy
equals mass. So, by smashing subatomic
particles together at high energies, physi-
cists can blast out new particles—fleeting,
massive entities not seen since the big
bang. By creating them and watching them
decay, they have pieced together a theory
of fundamental particles and forces called
the standard model.
The theory includes the four types of
particles in ordinary matter: the electron,
which whirls around the atomic nucleus;
the electron neutrino, which emerges in
a type of nuclear decay; and the up quark
and the down quark, which bind in trios to
form the protons and neutrons in atomic
nuclei. Two sets of similar but heavier
particles can pop into brief existence: the
1406 29 MARCH 2024 • VOL 383 ISSUE 6690
muon, muon neutrino, charm quark, and
strange quark; and the tau, tau neutrino,
top quark, and bottom quark.
The standard model describes how these
particles and their antiparticles interact
through three forces: electromagnetism, the
strong nuclear force that binds quarks, and
the weak nuclear force that enables quarks
and other particles to decay and change
type. (The theory ignores gravity.) The
forces emerge as the matter particles ex-
change other particles. The electromagnetic
force is conveyed by the photon, the strong
force by the gluon, and the weak force by
particles called the W boson and Z boson.
The Higgs boson completes the theory by
helping give the other particles mass.
Physicists worked out much of this with
colliders, typically circular accelerators
that send beams of electrically charged par-
ticles racing in opposite directions at near–
light-speed to smash them together. One
type fires electrons into their antimatter
“The challenge, if you want
to capture it in one word,
is that the muon is unstable.”
Sergo Jindariani,
Fermi National Accelerator Laboratory
counterparts, positrons, but these e+e-
colliders struggle to reach high energies.
That’s because charged particles radiate x-
rays as they circulate, and less massive par-
ticles radiate more than heavier ones. With
a mass just 0.05% as big as that of a proton,
electrons radiate strongly enough to limit
their energy. At some point, a circular e+e-
collider becomes like a leaky bucket, losing
energy as fast as it is pumped in.
To reach higher energies, physicists usu-
ally build bigger, more powerful hadron
colliders that smash protons into either
protons or antiprotons. But hadron col-
liders have limitations, too. A proton is a
bundle of quarks and gluons, so when two
protons strike each other, just a single con-
stituent in each is likely to collide directly,
converting less than 10% of the protons’
energy into new particles and reducing the
machine’s energy advantage. The remain-
ing fragments of the protons also generate
sprays of extraneous particles. In contrast,
an electron-positron collision consumes the
colliding particles’ full energy and produces
no extraneous sprays, so physicists often
use an e+e- collider to scrutinize the new
particles discovered at a hadron collider.
For example, in 1983 CERN physicists
used a proton-antiproton collider with an
energy of 540 gigaelectron volts (GeV) to
discover the W and Z bosons. CERN then
bored the 27-kilometer tunnel on the border
of Switzerland and France to build the Large
Electron-Positron collider, which enabled re-
searchers to study the W and Z in detail and
deduce that there are most likely just three
generations of matter particles. Then, in the
same tunnel, CERN built the LHC, designed
to smash protons at 14 teraelectron volts
(TeV). It found the Higgs.
That discovery neatly completed the
standard model but left physicists vexed.
The theory has obvious shortcomings—for
example, it doesn’t include dark matter,
the mysterious stuff thought to make up
85% of the universe’s matter. Researchers
hoped the LHC would cough up other new
particles that would lead to a deeper un-
derstanding. But so far, it has turned up
nothing more.
So CERN plans a similar progression
of bigger machines after the LHC shuts
down in 2041. For 15 billion Swiss francs—
roughly $17 billion—the lab envisions
digging a 91-kilometer-long tunnel and
installing an e+e- ring called the Future
Circular Collider-ee (FCC-ee) that would
run at 350 GeV, the likely limit for such a
machine. To be built by the mid-2040s, it
would examine the Higgs as the LHC can-
not. Thirty years later, CERN would replace
it with a 100-TeV hadron collider called the
FCC-hh that would seek new particles and
could cost $50 billion.
However, a muon collider might com-
bine the strengths of hadron and e+e-
colliders—and be faster and cheaper to
build. With a mass 207 times an electron’s,
a muon radiates far less energy as it goes
in circles, so a muon collider could reach
much higher energy than a circular e+e-
machine. At the same time, muons are fun-
damental particles that put all their energy
into a collision, enabling a muon collider
to compete with a hadron collider running
at 10 times the energy. So the muon col-
lider could be much smaller and, hence,
cheaper. A 10-TeV muon collider could be
had for $18 billion, advocates estimate.
THE CONCEPT OF A MUON COLLIDER dates
back decades, and in 2010 Fermilab
launched a program to develop it. How-
ever, 6 years later, the Department of
Energy, which funds most U.S. particle
physics research, stopped the program to
direct resources to another experiment: a
$3.3 billion project, currently under con-
struction, to shoot a beam of muon neutrinos
from Fermilab in Illinois to a gigantic sub-
terranean detector in South Dakota. The
aborted muon collider effort was more
popular with accelerator experts than par-
ticle physicists, says Diktys Stratakis, an
science.org SCIENCE
A smashing idea High-energy
rapid cycling
A muon collider would smash high-energy muons—heavier, unstable cousins of synchotron
electrons—into their antiparticles in two huge particle detectors. In its ability
to blast out massive new particles, it should rival a more conventional proton
collider running at an energy 10 times as high. It would also be smaller
and potentially much cheaper—if it can be built. To make a muon collider, Collider ring
physicists will have to generate muons, wrangle them into compact (~10-km circumference)
beams, and smash them together in the few milliseconds
before the particles decay. They’ll also have to cope
with radiation emanating from the muon beams.
m+
Graphic by Austin Fisher
m+
2
p+ m– 3
1 m–
p+ p –

Particle detector
Proton source Muon source Ionization cooling Low-energy rapid
channels cycling synchotron 1 km

1 Making muons 2 Bunching them into beams

Protons (p+) fired into a graphite target would generate negatively The muons would pass through a material such as liquid hydrogen and lose energy as
charged pions (p–), which would decay in flight to make negatively they ionize the atoms. The loss would make them swirl in a magnetic field in ever-tighter
charged muons (m–). The collisions would also yield positive pions (p+), spirals while RF cavities would accelerate them in one direction, forming a compact
which would decay into positively charged antimuons (m+). beam. Realizing such ionization cooling may be physicists’ biggest challenge.

Proton beam High-density target Magnetic RF cavity Liquid

coil hydrogen

1m

p+ High-density Hydrogen Ionized

target molecule hydrogen

p– p– m– m–

Liquid hydrogen Ejected

Vacuum p+ Neutrino electron

3 Sifting through the shards

A pair of massive particle detectors would look for new particles
produced in the muon collisions and instantly decaying into more Particle from collision
familiar ones. Each detector would comprise the usual subsystems,
but would also possess special shields to tamp down the radiation
emanating from the muon beams. Neutral particle

Inset at right Solenoid magnet

Collision point

Chambers to
track muons Muon beam

Tracking chamber
(measures momentum) Piercing the haze Tungsten nozzle
Within each detector, two cones of tungsten
would surround the beam pipe to screen out the
Radiation shield Positrons
electrons and positrons generated by decaying
muons. Those particles would spiral into the and electrons
Calorimeter tungsten to produce low-energy neutrons and
(measures energy) 2m photons that should be relatively easy to
distinguish from the desired quarry: particles
produced by high-energy muon collisions.
 An accelerating cavity for the now-defunct Muon Ionization Cooling Experiment, which partially demonstrated technologies needed to make a muon beam.
accelerator physicist at Fermilab who took
part in it. “We had a very nice product, but
we didn’t have any customers,” he says.
That has changed, Stratakis says. “Now
we have a lot of customers coming to us and
saying, ‘Hey, can you build a collider with
these parameters?’”
One reason is the realization that, tech-
nologically, building the FCC-hh will be
“a lot more difficult than people thought,”
says Hitoshi Murayama, a theorist at the
University of California (UC), Berkeley who
chaired the recent P5 team. For example,
the machine would need steering magnets
with fields of 16 tesla—33% higher than the
current state of the art and likely unobtain-
able for 20 years, Murayama says.
At the same time, physicists also feel
a new urgency to reach higher energies.
That’s because of all that the LHC has
not found—at least, so far. No particles of
dark matter, the unseen stuff whose grav-
ity holds galaxies together. No mini–black
holes. None of particles predicted by super-
symmetry, a concept that posits a more
massive “superpartner” exists for every par-
ticle in the standard model. “If many new
things had been discovered at the LHC, we
would really want to study them with high
precision and maybe not worry too much
about going to higher energy,” says Rachel
Yohay, an experimentalist at Florida State
University who works on an LHC experi-
ment. “Since we haven’t, there is a lot of in-
terest to explore higher energies.”
1408 29 MARCH 2024 • VOL 383 ISSUE 6690
Perhaps most important, some physicists
say a muon collider would be the best tool to
address the pressing questions raised by the
discovery of the Higgs. The particle is part of
a concept called the Higgs mechanism that
was added to the standard model in the 1970s
to explain an abstruse but important puzzle.
Mathematical symmetries within the stan-
dard model suggest the weak and electro-
magnetic forces are different aspects of a
single “electroweak” force. Yet the electro-
magnetic force can, in principle, reach
across the universe, whereas the weak force
doesn’t even reach across a nucleus.
Physicists had a standard explanation for
the disparity. The electromagnetic force is
long range because the photon has no mass,
and the weak force is short range because
the particles that convey it, the W and Z,
are massive. But there was a catch. In the
standard model, simply plugging in the W
and Z masses ruins the mathematical sym-
metry that produces the weak force in the
first place. So those masses have to emerge
in some less direct way.
Enter the Higgs mechanism. It assumes
the vacuum of empty space contains a
Higgs field, like an electric field that never
switches off. The otherwise massless W and
Z interact with it to acquire energy and,
hence, mass. The field consists of Higgs bo-
sons lurking “virtually” in the vacuum.
But why does the Higgs field persist? The
standard model assumes the field interacts
with itself in a particular way that ensures
its energy is lowest when, instead of van-
ishing, it has some nonzero strength. But
nobody knows whether that assumption is
right, says Nathaniel Craig, a theorist at UC
Santa Barbara. “We know the Higgs field
has this background value everywhere in
space, but we don’t really know why.”
To probe how the Higgs field interacts with
itself, scientists need collisions that produce
multiple Higgs bosons. The LHC already de-
tects the occasional Higgs pair and should
see more after it is upgraded in 2026–29 to
quintuple its collision rate. CERN’s planned
FCC-ee would see still more. But researchers
would need to spot at least a few events with
three or four Higgs particles, which would
require a higher energy machine. For that
task a 10-TeV muon collider might have an
edge over a 100-TeV proton collider.
That’s because colliding protons almost
always shred each other through a strong- PHOTO: REIDAR HAHN/FERMI NATIONAL ACCELERATOR LABORATORY
force interaction, which will rarely produce
a Higgs. Muons interact through just the
electromagnetic and weak forces, so their
collisions are more likely to produce the
prized events. A muon collider’s ability to
more clearly probe Higgs physics sets it
apart, says Donatella Lucchesi, an experi-
mentalist at the University of Padua and
Italy’s National Institute of Nuclear Physics.
“This is an opportunity that we should not
miss, it is super important.”
A muon collider’s sensitivity to weak
interactions could also help it stalk other
quarry, such as dark matter. Theorists have
science.org SCIENCE

long thought the stuff might consist of
heavy particles that interact only through
the weak force, but experimenters have
yet to detect such weakly interacting mas-
sive particles (WIMPs). A muon collider
could produce WIMPs too massive and too
weakly interacting to be seen in current
searches at the LHC or in sensitive detec-
tors underground, Craig says.
The machine might even put super-
symmetry to the acid test. The decades-old
concept could explain many things, such as
where WIMPs come from, but its main job
is to solve a more abstruse problem. Quan-
tum mechanics predicts that known kinds
of particles should fleetingly emerge from
the vacuum around the Higgs and interact
with it to make it hugely massive. That
doesn’t happen and physicists don’t know
why. Supersymmetry provides an answer.
Its “central prediction” is that the Higgs
has weakly interacting superpartners of
modest mass that counteract that effect,
Craig says, and a muon collider would be
ideally suited for hunting them.
BUT CAN ONE ACTUALLY BE BUILT? A muon
collider would consist of familiar accelera-
tor parts: so-called RF cavities resonating
with radio waves to accelerate the particles,
and magnets to steer and focus the beams.
But it would have to work incredibly quickly
because the muon is so short lived. Sitting
still, a muon decays into an electron, a neu-
trino, and an antineutrino in 2.2 micro-
seconds. High-energy muons traveling at
near–light-speed endure longer, for milli-
seconds, because of the time dilation pre-
dicted by Einstein’s theory of relativity. But
that’s still just a blink of an eye.
The clock would start when physicists
generate the muons. They would fire a pro-
ton beam into a target to produce particles
called charged pions, which, swirling in a
magnetic field, would decay into muons.
The same technique already generates neu-
trinos for other experiments. But a muon
collider would require a target that could
handle a proton beam with several mega-
watts of power.
The muons would emerge in a cloud centi-
meters wide. To herd them into a beam a
few micrometers across, they would pass
through a low-density material such as
lithium hydride or liquid hydrogen. Still
swirling in a horizontal magnetic field, the
muons would ionize atoms in the material,
dissipating energy and quelling their buzz-
ing about. An RF cavity would accelerate
the muons into the next cooling cell. That
could be tricky, because RF cavities typi-
cally don’t work well in magnetic fields.
Next, two or more circular accelerators
known as synchrotrons would boost the
SCIENCE science.org
beams of muons and antimuons to their
final energy. As the particles in a synchro-
tron gain energy, the fields generated by the
steering magnets need to ramp up in syn-
chrony to keep the particles on a circular
path of a fixed radius. At the LHC that pro-
cess takes 20 minutes. In contrast, to keep
replenishing its beams, a muon collider
would require synchrotrons that could cycle
a blinding 400 times per second.
Finally, the beams would pass into a
smaller accelerator called a storage ring,
which could be as little as 10 kilometers
long—petite compared with the LHC. Within
it, one bunch of muons and one bunch of
antimuons would circulate at fixed energy in
opposite directions, colliding in the hearts of
two detectors on opposite sides of the ring.
A smaller ring provides an obvious benefit,
says Stephen Gourlay, an accelerator physi-
cist at Fermilab. “You get more turns before
the muons disappear.”
The machine’s various components
would push the frontiers of magnet tech-
nology, Gourlay says. “This machine is a
magnet builder’s dream—or nightmare.”
For example, the magnets in the rapid cy-
cling synchrotrons must change fields by
several tesla almost instantly. That process
would not only exert enormous mechani-
cal stresses on the magnets, it would also
require safely shunting tens of gigawatts of
power in and out of them with exquisite
efficiency, Gourlay says.
Physicists will have to shield the ma-
chinery from the electrons and positrons
gushing from the beams as the muons
decay. Doing so is especially complicated
in the detectors, as physicists must take
care not to block the particles they want
to observe—those coming from the muon
collisions. To deal with the extraneous de-
cay particles, a cone of tungsten would sur-
round the beam pipe on either side of the
collision point in each detector. Electrons
and positrons striking the cone would
convert mostly to low energy photons and
neutrons, which are easier to distinguish
from the desired signals.
Decaying muons also radiate energetic
neutrinos, creating a novel radiation safety
challenge. Shooting horizontally from the
collider ring, these elusive particles would
zip through the earth and emerge dozens of
kilometers away. A few would strike atomic
nuclei in the soil and change back into mu-
ons, which could emerge from the ground
as potentially dangerous radiation. The
neutrinos can’t be blocked, but digging the
collider’s tunnel deeper would allow them
to spread more and lower the intensity of
muon radiation. Building the collider on
movable mounts so the orientations of its
sections could be gradually changed would
NE WS | F E AT U R E S
also help, Stratakis says, by limiting the ra-
diation dose at any one place.
Developers’ biggest challenge may be
proving ionization cooling works. “The
things that have never been done before
get the highest priority because there’s
the greatest chance of a surprise,” says
Mark Palmer, an accelerator physicist at
Brookhaven National Laboratory who
headed Fermilab’s previous muon collider
effort. In 2020, the International Muon
Ionization Cooling Experiment at the
United Kingdom’s Rutherford Appleton
Laboratory showed individual muons spi-
raled through a cooling cell as predicted.
But it did not show that the process could
actually cool a muon beam, Palmer says.
FACING SO MANY UNKNOWNS, U.S. physicists
are not clamoring to start planning such a
project now, as some press reports have sug-
gested. Instead, they simply want support
for basic R&D, says Patrick Meade, a theo-
rist at Stony Brook University. “In the U.S.,
the zeroth-order thing is getting permission
to do any research in this direction.”
Physicists envision 7 years of research
at an annual cost of a few million dollars
to determine what kind of demonstra-
tion project—perhaps a prototype cooling
channel—would best prove a muon collider
is feasible, Stratakis says. They would then
spend a decade and $100 million or more
building and running a demonstrator, with
the aim of deciding by 2040 whether to go
on to build a collider, he says.
Others are musing about muons, as well.
In 2021, CERN started the International
Muon Collider Collaboration (IMCC), which
U.S. researchers hope to join. As for who
might host a muon collider, “the goal is
to hold off on that decision until the mo-
ment when funding agencies put money
on the table,” says Daniel Schulte, a CERN
accelerator physicist and leader of the
200-member IMCC.
Will supporters get that far? “In Europe
the muon collider is something like a plan
B,” Schulte says. And CERN could soon opt
for the first step in plan A: the FCC-ee. In an
online press conference in February, CERN
Director-General Fabiola Gianotti said the
lab hoped to decide by 2028 whether to
build that machine, noting a muon collider
is “not on the same timescale.”
Nevertheless, the idea of a muon collider
continues to fascinate some physicists, es-
pecially younger researchers. Lucchesi says
she can more readily find graduate students
to work on the concept than to join ongoing
experiments at the LHC. What’s compelling
isn’t “something that you do for the fourth
or fifth time just to improve the errors,” she
says. “What is attractive is something new.” j
29 MARCH 2024 • VOL 383 ISSUE 6690 1409
 INSIGHTS
PERSPECTIVES
INFECTIOUS DISEASE
Vaccines to control tuberculosis in cattle
The age-old cattle disease has resisted rigorous control, but the BCG vaccine may do better
By Anita L. Michel
T
uberculosis (TB) has plagued humans
around the globe for thousands of
years and still causes 10 million new
cases every year (1). The second leading
cause of death from infectious disease
in humans, TB also afflicts animals,
both domestic and wild. Bovine TB in cattle
causes socioeconomic implications as a re-
sult of decreased production, the compulsory
slaughter of infected animals, and associated
social knock-on effects (2). The BCG (Bacillus
1410 29 MARCH 2024 • VOL 383 ISSUE 6690
Calmette-Guérin) vaccine, which has been in
use for 100 years, is currently the only vaccine
available to protect children against human
TB (caused by Mycobacterium tuberculosis);
there is no TB vaccine for livestock (caused by
the related bacterium Mycobacterium bovis).
On page 1433 of this issue, Fromsa et al. (3)
report field data from studies in Ethiopia that
support the use of the human BCG vaccine in
cattle, which could help eliminate bovine TB.
There are several closely related human-
and animal-adapted TB-causing bacteria that
can cross species barriers given favorable di-
rect or indirect interspecies contact (4). The
dire consequences of this zoonosis for human
health were felt throughout Europe during
the first half of the 20th century. In the UK,
PHOTO: MICHELE SPATARI/GETTY IMAGES
an estimated 50,000 new TB cases in hu-
mans were caused by M. bovis in the 1930s.
The burden of M. bovis infection in humans
mirrored the soaring prevalence in the cattle
population, which can be attributed to the
intensification of livestock production and
the uncontrolled spread of bovine TB. Com-
pulsory pasteurization of milk and bovine
TB control programs were introduced in the
science.org SCIENCE

late 19th century and mid-20th century, re-
spectively, and together curbed the epidemic
in both humans and cattle in high-income
countries.
By sharp contrast, bovine TB remains
largely uncontrolled in low- and middle-
income countries (LMICs), posing a risk to
vulnerable communities with limited access
to health and veterinary services. Addition-
ally, bovine TB has a detrimental effect on
the agricultural industry and the livelihoods
of farmers. Some countries impose compul-
sory quarantine for infected herds, excluding
them from trade opportunities, or even en-
force the slaughter of infected cattle without
financial compensation (5). The socioeco-
nomic consequences are crippling for LMICs,
Department of Veterinary Tropical Diseases, Faculty of
Veterinary Science, University of Pretoria, Onderstepoort,
South Africa. Email: anita.michel@up.ac.za
SCIENCE science.org
BCG vaccination could bring relief from bovine
tuberculosis in Ethiopia’s Amhara region and other
regions in low- and middle-income countries.
and such measures are socially and culturally
unacceptable, calling for alternatives (6).
TB has always been and remains the larg-
est infectious disease threat to captive wild
animals in zoological collections (7). In free-
ranging wild animals, TB may pose a con-
servation concern to endangered species (8),
and certain wildlife populations can become
permanently infected reservoirs that jeopar-
dize bovine TB control (9, 10). Under these
circumstances, wildlife reservoir vaccination
could help control bovine TB in cattle (10).
The BCG vaccine constitutes a laboratory-
adapted live M. bovis. It is safe and can confer
protection to children against human TB, so
why is the BCG vaccine not used to protect
cattle? The strongest argument was around
the interference of the BCG vaccine with the
official diagnostic test (the tuberculin skin
test) that is used in bovine TB surveillance,
international trade, and bovine TB control
programs; the vaccine renders this test unre-
liable. Additionally, numerous earlier studies
reported that BCG induced partial instead of
complete protection in cattle, which appeared
to be limited to 1 to 2 years, necessitating re-
vaccination with uncertain effects on protec-
tion. As a result, such arguments hindered
the acceptance of BCG vaccination of cattle,
and thus the test-and-slaughter strategy
(whereby all herds are repeatedly tested and
infected animals are slaughtered) was con-
sidered a faster and more cost-effective way
to eliminate bovine TB. Several countries,
such as Germany and the Netherlands, have
achieved eradication of bovine TB using this
strategy, but many others lack the resources
or the political commitment or their efforts
have been hampered by spillback of M. bovis
from an infected wildlife reservoir (11, 12).
The increasing costs of bovine TB control
(larger herd sizes require longer duration of
control measures, which harms production)
and an apparent failure to effectively lower
the disease prevalence have led to a renewed
interest in cattle BCG vaccination—especially
in the UK (13)—as an additional rather than
a stand-alone approach. The development of
a modified diagnostic test that differentiates
vaccinated from infected animals (called the
DIVA test) has made this a possibility (14).
Moreover, a duration of 52 weeks for BCG-
induced immunity was demonstrated (13),
which covers most of the life span of the
modern commercial dairy cow.
Fromsa et al. carried out a natural trans-
mission study in cattle in Ethiopia in which
vaccinated and unvaccinated susceptible
cattle were in contact with infected animals.
A two-stage study design allowed the authors
to quantify the direct and indirect BCG vac-
cine efficacy. Indirect vaccine efficacy refers
to the spectrum of protective effects of vac-
cines. Animals that are not fully protected
(immune) may have a partial protection that
leads to a reduced risk of them transmitting
the disease and thus confers a benefit toward
reducing disease prevalence. Indeed, their
results show a superior indirect vaccine effi-
cacy (74%) compared with direct protection
(58%), and the total BCG vaccine efficacy
was 89%. Using a transmission model tai-
lored for the dairy sector in Ethiopia, the
authors concluded that BCG vaccination of
cattle has the potential to limit the spread
of bovine TB and could reduce the preva-
lence of the disease in this setting within
10 years, improving the prospect of elimina-
tion over a longer time frame.
Together, these advances strongly advo-
cate for the adoption of BCG vaccination
into the control strategy for bovine TB.
For various reasons, the test-and-slaughter
approach introduced decades ago in high-
income countries did not achieve the re-
duction or elimination of disease that was
expected. This realization has allowed the
concerns of LMICs about this approach to
enter the agenda. However, several ques-
tions remain with regard to BCG vaccina-
tion, including the influence of disease
prevalence, cattle breed, farming system,
and revaccination on vaccine efficacy. From
a practical perspective, the costs and avail-
ability of the vaccine and the DIVA test
reagents could become limiting factors,
especially in LMICs. Future field research
will need to address the scientific knowl-
edge gaps and should ideally encompass
all promising candidate vaccines, including
inactivated M. bovis and subunit vaccines
(15), so that the needs of different settings
and species in high-income countries and
LMICs can be met. j
REFERENCES AND NOTES
1. World Health Organization, Global Tuberculosis
Report (2023); https://www.who.int/publications/i/
item/9789240083851.
2. F. Olea-Popelka et al., Lancet Infect. Dis. 17, e21 (2017).
3. A. Fromsa et al., Science 383, eadl3962 (2024).
4. M. Fellag, A. Loukil, M. Drancourt, New Microb. New
Infect. 41, 100712 (2021).
5. N. Valdivieso, P. Retamal, Ir. Vet. J. 76, 20 (2023).
6. A. Zumla et al., Lancet Infect. Dis. 20, 642 (2020).
7. A. L. Michel et al., Transbound. Emerg. Dis. 60, 46 (2013).
8. M. Azhar et al., J. Wildl. Biodivers. 7, 1 (2023).
9. B. M. Buddle et al., Front. Vet. Sci. 5, 10.3389/
fvets.2018.00259 (2018).
10. P. M. Boggiatto, C. R. Kanipe, E. J. Putz, S. C. Olsen, M. V.
Palmer, J. Immunol. 211, 1173 (2023).
11. M. L. Boschiroli, Ir. Vet. J. 76, 25 (2023).
12. D. J. O’Brien et al., Ir. Vet. J. 76, 16 (2023).
13. T. Holder et al., Vaccine 41, 7290 (2023).
14. B. Bayissa et al., Transbound. Emerg. Dis. 69, e1 (2022).
15. F. Blanco, J. Sabio y García, F. Bigi, CABI Rev. 10.1079/
cabireviews.2023.0010 (2023).
10.1126/science.ado4333
29 MARCH 2024 • VOL 383 ISSUE 6690 1411
INSIGHTS | P E R S P E C T I V E S
HYPOTHESIS
Chemodiversity in freshwater health
Dissolved organic matter may offer a way to track and restore the health of fresh waters
By Andrew J. Tanentzap1 and
Jérémy A. Fonvielle2
F
reshwater ecosystems, such as lakes
and rivers, have been deteriorating
in many places worldwide (1). Grow-
ing research suggests that many as-
pects of freshwater health—that is, the
maintenance of chemical, physical,
and biological integrity—may depend on
the composition of dissolved organic mat-
ter (DOM) (2). DOM consists of thousands
of distinct organic compounds that mainly
originate from different plant and animal re-
mains. The resulting variety of compounds
in DOM has been called chemodiversity
to mirror the term biodiversity. Whereas
past research tested how the environment
shapes the chemodiversity of DOM (3, 4), a
new research frontier is understanding how
healthy ecosystems may depend
on chemodiversity so as to moni-
tor and reverse declines in fresh-
water health.
Four main lines of evidence
indicate that different types of
DOM influence the function-
ing of ecosystems. DOM from
the land around fresh waters
can supply nutrients to subsi-
dize food web productivity (5).
The elements bound to DOM,
such as nitrogen and phos-
phorus, may result in undesir-
able effects, such as promoting
concentrations of harmful
bloom-causing algae species
that impair water quality. Ad-
ditionally, DOM can either re-
duce or enhance the toxicity of
contaminants. For example, the
highly toxic organometallic pol-
lutant methylmercury, which is
derived from natural and an-
thropogenic sources, can vary
in bioavailability and accumula-
tion through aquatic food webs
owing to the composition of
DOM. Thiols (sulfur-containing
analogs of alcohols) can bind to
methylmercury and reduce its
bioavailability, and the concen-
tration of thiols varies by many
fold in the environment (6).
Moreover, DOM can change the The health of fresh waters, as in this small river outside Kuujjuaq,
metabolism of microorganisms
1412 29 MARCH 2024 • VOL 383 ISSUE 6690
in ways that influence wider biogeochemi-
cal cycles. Heterotrophic microbes are the
main consumers of DOM and annually re-
spire 1.2 to 3.3 Pg carbon (1 Pg = 1015 g) of
terrestrially derived DOM as carbon diox-
ide to acquire energy [adenosine triphos-
phate (ATP)] from organic compounds
(2). However, compounds vary widely in
the energy they release during degrada-
tion. Some compounds, such as those with
greater ratios of oxygen to carbon atoms,
require more energy to degrade, and so re-
sult in more respiration (7). Lastly, some
types of DOM can modify the environment
by reducing heat penetration into deep
waters. Cooler deepwater temperatures
can offset impacts of climate warming that
reduce habitat for cold-water species and
promote temperature-dependent hetero-
trophic respiration and release of green-
Canada, can be tracked by measuring chemodiversity.
house gases to the atmosphere.
DOM is thought to affect the function-
ing of fresh waters through the diverse
characteristics of individual compounds,
such as proteins, lipids, and lignin (which
adds structure and strength to plants).
Compounds vary in characteristics that
include their elemental composition,
chemical structure, and biological and
photochemical reactivity. For example,
some compounds, such as many carbohy-
drates produced by algae, are “biolabile”
because they are more thermodynami-
cally favorable for microbial degradation
or meet microbial resource needs versus
more “biorefractory” compounds, such as
lignins produced by higher land plants (8).
Therefore, DOM that is dominated by more
biolabile compounds can be more quickly
transferred into microbial biomass or pro-
vide energy to break down bio-
refractory compounds, thereby
changing water clarity and
carbon fluxes from fresh water
(9). Any compounds chelated to
DOM, including contaminants,
may also be more quickly trans-
ferred to higher trophic levels in
the food web. By contrast, DOM
that is dominated by more bio-
refractory compounds may re-
sult in carbon accumulating in
waters at higher concentrations
(10). Because different sources
of DOM (e.g., runoff and algal
blooms) and modifiers of DOM
(e.g., temperature, photoirradia-
tion, and microorganisms) vary
from subdaily to seasonal time-
scales, chemodiversity may also
be very sensitive to change and
thus an ideal measure of ecosys-
tem health (see the figure).
Measuring chemodiversity
can help predict the role of
DOM in the environment. Tra-
ditional approaches to study
DOM, such as those based on
optical qualities of fresh wa-
ter, measure the bulk proper-
ties of DOM. These approaches
PHOTO: A. TANENTZAP
provide useful, albeit general,
information about the compo-
sition of DOM, but are unable
to resolve the origin and fate of
individual compounds. Such ap-
science.org SCIENCE
 proaches can therefore mask the influence
of distinct compounds if most of the other
compounds in a mixture are different. For
example, optical properties such as fluores-
cence and absorbance may identify a large
fraction of DOM as biolabile. However, the
biolabile fraction may be associated with a
few compounds that can undergo
only highly specialized biochemi-
cal transformations by microbes,
which optical properties cannot
detect. This DOM mixture would
have very different consequences
for freshwater health (e.g., water
clarity and food web productiv-
ity) compared with a mixture that
had the same fraction of biolabile
compounds, but which could un-
dergo more numerous biochemical
transformations.
High-resolution mass spec-
trometry is the state-of-the-art for
measuring chemodiversity. By re-
solving the molecular composition
of DOM, mass spectrometry can
identify the characteristics of in-
dividual compounds, such as their
thermodynamic traits and stoi-
chiometry. This information can
be used to summarize the variety
and abundance of ecological roles
(e.g., biochemical transformations,
toxicity, and biolability) in chemo-
diversity and can be derived from
measuring DOM rather than sepa-
rately measuring many indicators
of freshwater health, including
organic nutrients, pollutants, and
biological condition.
Measuring chemodiversity with
high-resolution mass spectrometry
also provides insight into the deter-
minants of freshwater health. Com-
mon measures of ecosystem state
(e.g., temperature, pH, and concen-
trations of nitrogen, phosphorus,
and carbon) indicate the condi-
tion of a water body but not why
a given state has arisen. Because
different DOM sources have dis-
tinct composition, identifying the
compounds associated with a given
state can allow them to be linked to
their sources. For example, a major
tributary of Lake Erie shared nearly
30% more organic phosphorus
compounds with cropland drain-
age than with wastewater effluent,
GRAPHIC: A. FISHER/SCIENCE
1Ecosystems and Global Change Group,
School of the Environment, Trent University,
Peterborough, Canada.2Departments of
Plant Sciences and Biochemistry, University
of Cambridge, Cambridge, UK.
Email: atanentzap@trentu.ca
SCIENCE science.org
suggesting that inorganic fertilizers were ized by phosphonic acids or linear alkyl
the main source of phosphorus pollution benzene sulfonates) or agricultural runoff
in this region (11). Although phosphorus (enriched in tannins and lignins) (11). This
concentrations can be detected with simple information can only be revealed by mea-
colorimetric reactions, it cannot be deter- suring chemodiversity with high-resolution
mined whether the phosphorus is derived mass spectrometry, thus offering a more in-
from detergents in wastewater (character- tegrated way to assess the state and drivers
of freshwater health.
Knowledge of chemodiversity
The influence of dissolved organic matter measured with high-resolution
on freshwater health mass spectrometry can add new
Dissolved organic matter (DOM) is a complex mixture of compounds from information to improve freshwater
diverse sources in the catchment area of fresh waters. DOM undergoes health. For example, trace metals
biochemical modifications that influence physical, chemical, and biological such as iron limit the growth of
processes such as nutrient cycling, carbon storage, light absorption, and algae, including many toxic bloom-
food web interactions, which together determine ecosystem function. The forming species. Trace metals are
chemical characteristics of DOM mixtures, called chemodiversity, can be generally more available for uptake
determined by their sources and how labile the compounds are, which by algae when bound to organic
together influence ecosystem function and health. compounds and form ligands.
Measurements of iron concentra-
Sources of DOM in fresh water
tions or the optical properties of
Aerosols and precipitation Industrial chemicals DOM cannot identify iron-binding
compounds. By contrast, measur-
ing chemodiversity can help de-
termine whether iron is available
for biological uptake and trigger
interventions to reduce iron inputs,
especially where these come from
Urban runoff Wetlands runoff.
Another example of the ben-
efit of molecular-scale resolution
comes from the discovery that a
toxic compound from an anthro-
pogenic source caused decades
of unexplained mass die-offs in a
Agricultural runoff Forests socioeconomically important fish
species (Oncorhynchus kisutch) in
the Northwest US (12). The toxicant
was a rubber antioxidant used in
tires that washed into waterways
from particles wearing off on road-
Microorganisms Wildfires ways. The source of this toxicant
could only be identified by compar-
ing the chemodiversity of DOM of
affected waterways with that of sur-
rounding urban stormwater runoff.
There are other important bene-
fits of resolving DOM at the molecu-
How chemodiversity might influence ecosystem function lar scale. Some organic compounds
Aquatic source DOM mixture Terrestrial source in DOM, such as polyphenols and
Glucose (C6H12O6) Tannic acid (C76H52O46) highly unsaturated hydrocarbons,
Glycine (C2H5NO2) Lignin (C81H92O28) may be more likely to react with
chlorine during drinking-water
treatment to form by-products
that are harmful to human health
Ecosystem function
(13). However, the molecular iden-
tities of most of these by-products
have not been described before.
Only measurements of chemodi-
versity can start to identify these
compounds, which is necessary to
monitor their presence, identify
Biolabile Biorefractory their sources, and tailor treatments
Low chemodiversity High chemodiversity Low chemodiversity to reduce their presence.
29 MARCH 2024 • VOL 383 ISSUE 6690 1413
INSIGHTS | P E R S P E C T I V E S
Monitoring of chemodiversity can be
adapted for different aspects of freshwater
health. For example, chemodiversity can be
measured at subdaily or daily timescales to
protect drinking-water quality, whereas
seasonal measurements may be adequate
to assess the state of food webs. By moni-
toring all compounds at once, it will also
be possible to realize the benefits of both
targeted approaches that track specific
markers or contaminants and untargeted
approaches that track the properties of
DOM as a whole. Although there may be fi-
nancial barriers to uptake, the per-sample
cost of measuring chemodiversity is com-
parable to that of DNA sequencing, which
is widely used for biodiversity monitoring.
Forecasting how the chemodiversity of
DOM will affect freshwater health under
global environmental change is an impor-
tant future challenge. Climate warming
has generally been associated with more
biolabile compounds in DOM because heat
directly promotes their production and de-
composition by microbes, and indirectly
promotes their production by changing
land cover, such as favoring more nitro-
gen-rich vegetation (8). In cold climates,
warming temperatures could also release
previously preserved, highly biolabile com-
pounds, such as carbohydrates, from thaw-
ing permafrost and melting glaciers (14).
Linking freshwater health to changes in
chemodiversity, such as from the release
of biolabile compounds, would improve
predictions of ecosystem impacts from
environmental change. Similarly, the con-
sequences of changes to precipitation re-
gimes for chemodiversity and freshwater
health remain poorly understood. Drought
can extend water residence times that sub-
ject DOM to greater microbial processing
and photochemical degradation (2, 14),
thereby leading to greater chemodiversity
(3). However, drought can simultaneously
reduce chemodiversity by suppressing run-
off of terrestrial-derived DOM. Wetter cli-
mates with more flooding events may also
have mixed outcomes by diluting DOM and
increasing input from terrestrial sources.
Chemodiversity and thus the health of
fresh waters may be further modified by
other aspects of anthropogenic change. Ex-
amples include disturbances (such as wild-
fire and insect defoliation), eutrophication,
acidification, land-use and land-cover
change (including deforestation and ur-
banization), mining, and organic pollution
(including synthetic polymers, per- and
polyfluoroalkyl substances, chlorinated
hydrocarbons, and improper wastewater
discharge) (14). There are also longer-term
questions about how heterotrophs might
adapt to use new mixtures of DOM gen-
1414 29 MARCH 2024 • VOL 383 ISSUE 6690
erated by environmental change. As mi-
crobes transform and release compounds
into the environment, selection pressures
on their metabolic pathways from changes
in chemodiversity may feed back to alter
the composition of DOM even further.
If chemodiversity is to be a useful mea-
sure of freshwater health, approaches
are ultimately required to monitor and
manipulate DOM at the scale of entire
catchments. Although chemodiversity
is measured at the molecular level, it re-
flects the contribution of different biogeo-
chemical sources that supply a waterbody
(8). Therefore, by characterizing different
sources with high-resolution mass spec-
trometry and predicting their downstream
mixing with hydrological and source-
tracking models, chemodiversity and its
implications for ecosystem health can be
mapped across large spatial and tempo-
ral scales (11). DOM can also be manipu-
lated at the catchment scale by cultivating
different mixtures of vegetation around
and within fresh water that contribute to
downstream chemodiversity (8). For exam-
ple, different plant lineages contain differ-
ent metabolites in their leaves (15), which
will leach into water from litterfall and
soils, with different outcomes for ecosys-
tem function and services. Ultimately, ef-
fective protection of fresh waters and their
surrounding catchments can only emerge
if the determinants of ecosystem health are
understood. j
REFERENCES AND NOTES
1. E. R. Jones et al., Nat. Water 1, 602 (2023).
2. L. A. Kaplan, R. M. Cory, Stream Ecosystems in a
Changing Environment, J. B. Jones, E. H. Stanley, Eds.
(Academic Press, 2016), chap. 6, pp. 241.
3. T. Dittmar et al., Nat. Rev. Earth Environ. 2, 570 (2021).
4. D. N. Kothawala, A. M. Kellerman, N. Catalán, L. J. Tranvik,
Trends Ecol. Evol. 36, 113 (2021).
5. A. J. Tanentzap et al., Sci. Adv. 3, e1601765 (2017).
6. E. Seelen et al., Nat. Commun. 14, 6728 (2023).
7. F. Guillemette, P. A. del Giorgio, Limnol. Oceanogr. 56,
734 (2011).
8. A. M. Kellerman, T. Dittmar, D. N. Kothawala, L. J. Tranvik,
Nat. Commun. 5, 3804 (2014).
9. A. J. Tanentzap et al., Proc. Natl. Acad. Sci. U.S.A. 116,
24689 (2019).
10. J. C. Stegen et al., Nat. Commun. 9, 585 (2018).
11. M. R. Brooker, K. Longnecker, E. B. Kujawinski, M. H.
Evert, P. J. Mouser, Environ. Sci. Technol. 52, 6771 (2018).
12. Z. Tian et al., Science 371, 185 (2021).
13. C. Postigo et al., J. Hazard. Mater. 401, 123681 (2021).
14. M. A. Xenopoulos et al., Biogeochemistry 154, 323
(2021).
15. E. Defossez et al., Proc. Natl. Acad. Sci. U.S.A. 118,
e2013344118 (2021).
ACKNOWLEDGMENTS
The authors are supported by the Canada Research Chairs
Program, Natural Sciences and Engineering Research Council
Discovery Grant (RGPIN-2023-03977), European Research
Council Starting Grant (804673), and UK Research and
Innovation Directed Grant (NE/W00495X/1) to A.J.T.
10.1126/science.adg8658
CELL SIGNALING
A mitotic
stopwatch
determines
cell fate
Surveillance of mitotic
timing prevents amplification
of damaged cells
By Agustina P. Bertolin1 and
Vanesa Gottifredi2
M
itosis is a challenging time for cells
owing to the potential for structural
and numerical chromosomal ab-
normalities, termed chromosomal
instability, which is a hallmark of
cancer (1). In M phase, chromo-
somal instability is prevented by the spindle
assembly checkpoint, which delays cell cycle
progression until chromosomes are properly
attached to spindle microtubules, ensur-
ing proper separation into daughter cells
(2). Recently, another M phase checkpoint
was described: If a threshold mitotic dura-
tion is surpassed, proliferation of daughter
cells is arrested in a manner that requires
the tumor suppressor p53, the cell cycle in-
hibitor p21, ubiquitin-specific protease 28
(USP28), and p53-binding protein 1 (53BP1)
(3–6). On page 1441 of this issue, Meitinger
et al. (7) clarify how this process works, with
extended mitosis leading to the assembly of
a USP28-53BP1-p53 complex—the “mitotic
stopwatch”—which mediates daughter cell
arrest and may prevent cancer development.
Using human untransformed cell lines,
Meitinger et al. show that increasing mi-
totic durations past a time threshold in the
mother cell leads to progressively higher
amounts of stopwatch complexes, resulting
in p53 stabilization and high expression of
p21 (a transcriptional target of p53) in the
subsequent G1 phase. Furthermore, the mi-
totic stopwatch complexes are stable enough
to accumulate over multiple generations,
enabling the arrest of cells that have expe-
rienced successive subthreshold extended
mitoses. The outcomes experienced in G1 be-
cause of prolonged mitoses (which vary be-
tween transient arrest, permanent arrest, or
cell death) depend on the cell type being ana-
lyzed. In the case of transient arrest, however,
the benefits of the “extra” time in G1 and the
science.org SCIENCE
 Mitotic stopwatch formation
During an unperturbed M phase, levels of p53 remain unchanged compared with those of the previous S-G2 and G1 phases. If mitosis is prolonged beyond a time
threshold, the levels of a ternary soluble complex (53BP1-USP28-p53), referred to as the “mitotic stopwatch,” increase. Although 53BP1 is primarily associated with
chromatin during interphase, its solubility peaks during mitosis. Soluble 53BP1 is then phosphorylated by PLK1, triggering its interaction with USP28 and p53; this leads
to the expression of its transcriptional target p21, which induces arrest in G1. Prolonged arrest in G1 may induce senescence or cell death, depending on the cell type.
Normal
mitosis
Ub
p53
S-G2
USP28
53BP1 Chromatin
Mother Extended
cell mitosis
53BP1, p53-binding protein 1; PLK1, Polo-like kinase 1;
Ub, ubiquitin; USP28, ubiquitin-specific protease 28.
mechanism to overcome the cause of the pro-
longed mitosis are unclear, especially because
this checkpoint senses long mitotic duration
independently from the cause of the delay.
Meitinger et al. also investigated the
mechanisms that regulate mitotic stopwatch
formation and found that 53BP1 dissoci-
ates from chromatin during M phase. Once
soluble, 53BP1 is phosphorylated at multiple
residues by Polo-like kinase 1 (PLK1). These
phosphorylations promote the interaction of
53BP1 with p53 through the BRCA1 C-termi-
nal (BRCT) domain and with USP28 through
the Tudor domain (see the figure). The phos-
phorylation of other proteins by PLK1 also
prompts a time-dependent loss of affinity
of 53BP1 for chromatin (8), indicating that
the main determinant for stopwatch forma-
tion is yet to be found. Moreover, it remains
to be clarified whether a threshold of PLK1
activity or of soluble 53BP1 levels needs to be
surpassed for complex formation or whether
there is a positive correlation between PLK1
kinase activity and mitotic length.
Meitinger et al. suggest that the interac-
tion between USP28 and p53 may promote
p53 accumulation, likely through USP28-
mediated deubiquitination of p53. Previous
studies have demonstrated that the catalytic
activity of USP28 is essential for the G1 ar-
rest induced by prolonged mitosis (3). Fur-
thermore, USP28 has been shown to directly
deubiquitinate p53 in vitro and, when overex-
pressed, can stabilize p53 in human cells (3).
One possibility is that USP28 could counter
GRAPHIC: N. BURGESS/SCIENCE
p53 ubiquitination that is dependent on the
E3 ubiquitin-protein ligase MDM2, an onco-
protein that serves as the main p53 ubiquitin
1The Francis Crick Institute, London, UK. 2Fundación Instituto
Leloir, Consejo Nacional de Investigaciones Científicas y
Técnicas (CONICET), Buenos Aires, Argentina.
Email: agostina.bertolin@crick.ac.uk; vgottifredi@leloir.org.ar
SCIENCE science.org
G1
Mitotic
stopwatch
PLK1
Stopwatch G1
accumulation
ligase. A preliminary report (9) shows that
the levels of MDM2 are down-regulated in a
time-dependent manner during M phase and
contribute to the steep increase of p21 in the
ensuing G1. This suggests that MDM2 acts as
a mitosis timer that controls the level of ac-
tive p53 in G1 and that it will be important to
understand whether and how these two “tim-
ers” interact during prolonged mitosis.
The relevance of the mitotic stopwatch for
cancer was also investigated. Mitotic poisons
have been successfully used in the treatment
of solid cancers for more than 30 years. They
interfere with spindle microtubule dynamics
and thereby induce mitotic delay and stop-
watch activation. Not surprisingly, Meitinger
et al. show that more than half of human
cancers are likely impaired in their ability to
form a stopwatch, which provides an addi-
tional advantage for cancer cell proliferation.
Thus, assessing the levels of stopwatch func-
tionality may serve as a reliable indicator of
tumor sensitivity to mitotic poisons.
The mitotic stopwatch and the spindle
assembly checkpoint are two independent
mitotic surveillance processes that act in
parallel (3), and it will be important to un-
derstand how they are coordinated. Further-
more, 53BP1 associates with chromatin in G1
in structures known as 53BP1 bodies that are
linked to the protection of damaged-DNA re-
gions that accumulated in the previous cell
cycle (10), whereas soluble 53BP1 is a prereq-
uisite for stopwatch formation. Therefore,
elucidating how chromatin-associated 53BP1
bodies and soluble 53BP1-containing mitotic
stopwatches could coexist, compete, or syner-
gize in G1 merits further research.
Another aspect that needs to be clarified
is the involvement of DNA damage response
(DDR) kinases in stopwatch formation. Sev-
eral reports associate PLK1 activation with
Daughter cell
Proliferation
G1 arrest and
cell death
p21 expression
DDR kinase activation (11, 12). In particular,
the DDR checkpoint kinase 1 (CHK1) activity
in G2 modulates PLK1 activity and induces a
p53-p21–dependent arrest in the subsequent
G1 phase (13), similar to the one described
by Meitinger et al. Furthermore, compelling
evidence suggests that DNA damage can af-
fect the G2–M phase duration (14). Because
minor perturbations in the mitotic length
can activate the stopwatch, the role of the
DDR in this process should be investigated.
Finally, given that stopwatch activation can
induce both permanent cell cycle arrest and
cell death, future research should investigate
its effects on stem cells, particularly in terms
of reducing their proliferation potential.
Although all cancers share certain char-
acteristics, they also achieve this outcome
through a range of genetic and epigenetic
variations. The study of Meitinger et al.
helps reinforce the idea that identifying
and understanding the vulnerabilities and
differences among individual cancers, such
as whether tumor cells possess an active
mitotic stopwatch, may be more effective
than adopting a one-size-fits-all approach
to cancer treatment. j
REFERENCES AND NOTES
1. S. F. Bakhoum, L. C. Cantley, Cell 174, 1347 (2018).
2. P. Lara-Gonzalez et al., Semin. Cell Dev. Biol. 117, 86
(2021).
3. C. S. Fong et al., eLife 5, e16270 (2016).
4. B. G. Lambrus et al., J. Cell Biol. 214, 143 (2016).
5. F. Meitinger et al., J. Cell Biol. 214, 155 (2016).
6. Y. Uetake, G. Sluder, Curr. Biol. 20, 1666 (2010).
7. F. Meitinger et al., Science 383, 1441 (2024).
8. M. Burigotto et al., EMBO Rep. 24, e57234 (2023).
9. L. J. Fulcher et al., bioRxiv 10.1101/2023.05.26.542398
(2024).
10. A. P. Bertolin et al., Cancers 12, 2764 (2020).
11. B. Lemmens et al., Mol. Cell 71, 117 (2018).
12. D. Dwivedi et al., Nat. Commun. 14, 6088 (2023).
13. V. Lebrec et al., Dev. Cell 57, 638 (2022).
14. Ö. Özer, I. D. Hickson, Open Biol. 8, 180018 (2018).
10.1126/science.ado5703
29 MARCH 2024 • VOL 383 ISSUE 6690 1415
INSIGHTS | P E R S P E C T I V E S
MATERIALS SCIENCE
A more biofriendly piezoelectric material
A ferroelectric molecular crystal displays characteristics required for implantation
By Linping Wang1,2 and Run-Wei Li1,2,3
D
evices implanted in the human body
for sensing conditions and deliv-
ering treatments must be biosafe,
biocompatible, and biodegradable.
These requirements make piezo-
electric materials attractive for
their design because they convert mechan-
ical force into electricity and vice versa (1).
However, traditional piezoelectrics such
as lead zirconate titanate are permanent,
rather than transient, in nature (2, 3).
Although biodegradable piezoelectrics—
such as those including amino acids, col-
lagen, and chitin constituents—have the
desired properties for implantation and
are immune to infection and inflam-
mation risks (4), they suffer from weak
piezoelectric performance (5). On page
1492 of this issue, Zhang et al. (6) re-
port a simple ferroelectric molecular
crystal that exhibits piezoelectricity
performance that is more than 13 times
greater than that of amino acid–based
piezoelectric materials (7). When com-
posited with polyvinyl alcohol (PVA),
the material demonstrates high flex-
ibility and biodegradation and biosafety
in physiological environments. The find-
ing should advance piezoelectric materials
further into transplant applications.
Inorganic piezoelectrics (piezoelectric
ionic crystals) are known for their large
piezoelectric response, but their brittle na-
ture and resistance to sound wave propa-
gation hinder their use in medical and
biological applications (8). Conversely,
piezoelectric polymers (piezoelectric
molecular semicrystals) offer flexibility
and low acoustical impedance but only a
modest piezoelectric performance. For
both materials, the absence of biodegrad-
ability limits their use as transient im-
plantable devices.
A substantial advance in piezoelectric
material design is ferroelectrochemis-
1CAS Key Laboratory of Magnetic Materials and Devices,
Ningbo Institute of Materials Technology and Engineering,
Chinese Academy of Sciences, Ningbo, China. 2Zhejiang
Province Key Laboratory of Magnetic Materials and
Application Technology, Ningbo Institute of Materials
Technology and Engineering, Chinese Academy of
Sciences, Ningbo, China. 3College of Materials Science and
Optoelectronic Technology, University of Chinese Academy
of Sciences, Beijing, China. Email: runweili@nimte.ac.cn
1416 29 MARCH 2024 • VOL 383 ISSUE 6690
try, which centers on the strategic use of
chemical design methodologies to tailor
specific molecular ferroelectrics (9). Fer-
roelectric molecular crystals composed
of organic single-component compounds
have good solubility, potential biocompat-
ibility, and biodegradability (10). However,
despite these promising attributes, the
piezoelectric performance of these molec-
ular crystals falls substantially short, with
low piezoelectric coefficients (d33) of ≤40
pC/N (11). A desirable coefficient would be
>100 pC/N.
Zhang et al. report that the piezoelec-
tric molecular crystal HOCH2(CF2)3CH2OH
[2,2,3,3,4,4-hexafluoropentane-1,5-diol
(HFPD)] exhibits a high piezoelectric re-
“The exceptional performance…
points to their suitability as
carriers for drug delivery, as well
as for…self-powered transient
energy-harvesting devices…”
sponse of ~138 pC/N under no poling con-
dition (that is, without applying a strong
electric field across piezoelectric materials
to align the electric dipoles in a specific
direction). It also displays an exceptional
voltage constant (g33) of ~2450 × 10−3
Vm/N under no poling condition, indicat-
ing a large voltage output when the ma-
terial is subjected to mechanical stress or
strain in the same direction. Achieving
such properties in a ferroelectric molecu-
lar crystal is a milestone in piezoelectric
material development.
The underlying mechanism of this bio-
degradable ferroelectric molecular crystal
centers on a two-dimensional (2D) hy-
drogen bond network that is facilitated
by O–H···O interactions of the terminal
hydroxyl in the HFPD molecule. This in-
teraction plays a pivotal role in solubility
across various solvents. As well, this struc-
tural feature contributes to the biodegrad-
ability of the crystal, a crucial aspect of
its potential biomedical application. Fur-
thermore, Zhang et al. attribute the ma-
terial’s exceptional piezoelectric response
to the anisotropy of Young’s modulus,
which comes from the ordered arrange-
ment of the 2D hydrogen bond network
and fluorine atoms along the chain in the
crystal structure.
Zhang et al. also made HFPD-PVA film,
a composite that could be used in biomedi-
cal device design. It exhibits an array of at-
tributes that are essential for biomedical
applications. Beyond high flexibility and
biocompatibility, as demonstrated by cul-
turing the film with multiple types of cells,
these films also had no discernible toxicity
to cultured cells and exhibited exceptional
biodegradation and biosafety in physi-
ological environments, which was con-
firmed through in vivo assessment in rats,
including implantation tests beneath the
skin. Moreover, the HFPD-PVA (2:1) com-
posited films boasted strong piezoelec-
tricity, with a d33 measuring 34.3 pC/N
—eclipsing other organic piezoelec-
tric composites formed by biofriendly
molecular crystals such as γ-glycine
(5.3 pC/N).
The exceptional performance of
HFPD-PVA films points to their suit-
ability as carriers for drug delivery,
as well as for the development of self-
powered transient energy-harvesting
devices for therapeutic interventions.
Furthermore, these films hold promise
for regenerative medicine as scaffolds
that support tissue repair and regenera-
tion. Further research can focus on ex-
ploring elastification techniques for this
advanced piezoelectric material because
elastic ferroelectrics offer promising po-
tential to expand implantable device ap-
plications (11). The findings of Zhang et al.
are poised to transform the field of biomed-
ical engineering. j
REFERENCES AND NOTES
1. S. Chen et al., Adv. Mater. 35, 2208256 (2023).
2. S. Selvarajan, A. Kim, S. H. Song, IEEE Access 8, 68219
(2020).
3. K. Kapat, Q. T. H. Shubhra, M. Zhou, S. Leeuwenburgh,
Adv. Funct. Mater. 30, 1909045 (2020).
4. T. Liu et al., Nano Today 52, 101945 (2023).
5. F. Yang et al., Science 373, 337 (2021).
6. H.-Y. Zhang et al., Science 383, 1492 (2024).
7. D. Kim et al., Adv. Mater. 32, 1906989 (2020).
8. S. Gupta, G. Haiat, C. Laporte, P. Belanger, IEEE Trans.
Ultrason. Ferroelectr. Freq. Control 68, 1653 (2021).
9. H.-Y. Liu, H.-Y. Zhang, X.-G. Chen, R.-G. Xiong, J. Am.
Chem. Soc. 142, 15205 (2020).
10. S. Guerin et al., Nat. Mater. 17, 180 (2018).
11. L. Gao et al., Science 381, 540 (2023).
10.1126/science.ado5706
science.org SCIENCE
 RETROSPECTIVE
Ingeborg Levin (1953—2024)
Pioneer in radiocarbon and atmospheric research
By Felix Vogel1 and Samuel Hammer2
I
ngeborg Levin, groundbreaking physi-
cist who helped quantify the carbon
cycle and greenhouse gas emissions,
died on 10 February. She was 70. Inge-
borg established high-quality long-term
measurements to identify the radiocar-
bon present in carbon dioxide and inform
estimates of carbon fluxes. In the late 1970s,
climate change and greenhouse gases were
far from a mainstream interest, but Ingeborg
immediately recognized their importance
and the crucial role science would play in
informing policy. She also realized early on
that the carbon-14 (14C) introduced into the
atmosphere by nuclear weapons testing pre-
sented a singular opportunity to understand
how carbon moves from the atmosphere
into other carbon reservoirs, such as the bio-
sphere, soils, and oceans.
Born on 6 June 1953 in Erlangen, Ger-
many, Ingeborg earned a diploma in physics
from Heidelberg University in 1978 at the
then newly funded Institute of Environmen-
tal Physics. She stayed to earn her doctorate
in physics in 1984 and later established the
Institute’s carbon cycle group, which she led
until her death.
The network of global stations that Inge-
borg established tracks the atmospheric ra-
diocarbon content of carbon dioxide (14CO2)
using the systems she developed, which
are now standard across Europe. Others
had taken such measurements before, but
she was the first to insist on careful, high-
quality, long-term measurements across the
globe. She visited many of the stations and
secured funding for them, when needed. In-
geborg’s foresight allowed scientists today to
demonstrate beyond any doubt how fossil
fuel emissions have altered the atmosphere.
Sustaining these observations for more than
40 years, in the face of a science system that
seldom supports long-term monitoring, was
the challenge and triumph of her career.
Ingeborg’s 14CO2 records have been crucial
to understanding the global carbon cycle.
She used 14CO2 measurements from the
rooftop of her lab in Heidelberg to separate
atmospheric CO2 from fossil fuel burning
PHOTO: SAMUEL HAMMER
1Environment and Climate Change Canada, Toronto, ON,
Canada. 2Integrated Carbon Observation System Central
Radiocarbon Laboratory, Institute of Environmental
Physics, Heidelberg University, Heidelberg, Germany. Email:
felix.vogel@ec.gc.ca; shammer@iup.uni-heidelberg.de
SCIENCE science.org
and natural sources, a method that is now
the gold standard when quantifying fossil
fuel CO2 in the atmosphere. With her global
cooperative network, from the German Neu-
mayer station in Antarctica to the Canadian
Alert station in the high Arctic, Ingeborg
and her collaborators demonstrated that
most fossil fuel emissions indeed occur in
the Northern Hemisphere and that Southern
Ocean carbon exchange plays a vital role in
the global carbon cycle. She also used the
data to test the size and importance of the
terrestrial land carbon pool and to quantify
the magnitude of radioisotopes produced by
atmospheric nuclear weapons testing. Today,
Ingeborg’s long-term records from the Ver-
munt and Jungfraujoch stations in the Swiss
Alps tracking this 14C “bomb spike” are a
crucial calibration tool for radiocarbon dat-
ing. The records’ diverse applications also
include ensuring that bioplastics are bio-
based, catching illegal international trade of
elephant ivory, and understanding adult hu-
man cell growth.
Beyond radiocarbon, Ingeborg developed
methods to quantify greenhouse gas emis-
sions using radon-222 observations, an ap-
proach that has evolved into a research
community of its own. She was one of the
strongest voices in the atmospheric green-
house gas community, advocating the use of
observational data to improve greenhouse
gas emissions reporting to inform climate
change policies. Her work contributed to a
better understanding of the effects of meth-
ane, nitrous oxide, and sulfur hexafluoride.
Since the early days of the international
greenhouse gas measurement techniques
community, Ingeborg served not only as a
leading expert but also as an appreciated
colleague and friend. Ingeborg was a voice of
reason—and careful, precise science—in the
sometimes heated discussions about maxi-
mizing the potential of greenhouse gas data.
Later, she was among the founding members
of the Integrated Carbon Observation Sys-
tem (ICOS) network, which has become the
backbone of greenhouse gas measurements
in Europe and a prototype for the World
Meteorological Organization’s Global Green-
house Gas Watch.
We were among Ingeborg’s PhD students
and immediately learned how passion-
ate she was about science and precision,
both when performing measurements and
when developing ideas. Discussions with
her could sometimes be fierce, and there-
fore intimidating, but Ingeborg was also
very supportive and, above all, sincere. She
was an active collaborator who had enjoyed
hands-on work since childhood, when she
would build furniture in her father’s work-
shop. She led by example; during her annual
maintenance visits to the Jungfraujoch sta-
tion, she cleaned the equipment and avoided
many hours of paperwork and shipping cost
by personally transporting the samples. We
were constantly inspired by her dedication.
We were first struck by Ingeborg’s global
importance at a CO2 expert meeting in 2009.
For a week, experts with famous names
from around the world had been engaged
in discussions and arguments. On the last
day, when new recommendations had to be
agreed upon, Ingeborg took the stage. Sud-
denly, all eyes were on her. She was well in-
formed on every topic, in command of the
room, and able to cut through the chatter and
push through the necessary compromises.
Ingeborg was a leading researcher at a
time when that often meant being the only
woman in the room, a reality that presented
ongoing challenges. In 2020, she was the
first female scientist to receive the European
Geosciences Union’s Alfred Wegener Medal.
The award’s description rightly recognizes
her as integral to the field of atmospheric
science, highlights her integrity and rigorous
work, and distinguishes her as a role model
to men and women alike.
Never interested in judging herself or oth-
ers by citation index, Ingeborg measured her
contributions by the laboratories she helped
establish and the crucial role she played in
launching ICOS and building a global com-
munity through her decades-long inter-
national collaborations. She will be dearly
missed by the atmospheric greenhouse gas
community and beyond. j
10.1126/science.ado8559
29 MARCH 2024 • VOL 383 ISSUE 6690 1417
INSIGHTS

P OLICY FORUM
PUBLIC HEALTH

Mandating indoor air quality for public buildings

If some countries lead by example, standards may increasingly become normalized

By Lidia Morawska, Joseph Allen, William Bahnfleth, Belinda Bennett, Philomena M. Bluyssen, Atze Boerstra, Giorgio Buonanno, Junji Cao,
Stephanie J. Dancer, Andres Floto, Francesco Franchimon, Trish Greenhalgh, Charles Haworth, Jaap Hogeling, Christina Isaxon, Jose L. Jimenez,
Amanda Kennedy, Prashant Kumar, Jarek Kurnitski, Yuguo Li, Marcel Loomans, Guy Marks, Linsey C. Marr, Livio Mazzarella, Arsen Krikor
Melikov, Shelly L. Miller, Donald K. Milton, Jason Monty, Peter V. Nielsen, Catherine Noakes, Jordan Peccia, Kimberly A. Prather, Xavier Querol,
Tunga Salthammer, Chandra Sekhar, Olli Seppänen, Shin-ichi Tanabe, Julian W. Tang, Raymond Tellier, Kwok Wai Tham, Pawel Wargocki, Aneta
Wierzbicka, Maosheng Yao

P
eople living in urban and industrial-
ized societies, which are expanding
globally, spend more than 90% of
their time in the indoor environment,
breathing indoor air (IA). Despite de-
cades of research and advocacy, most
countries do not have legislated indoor air
quality (IAQ) performance standards for pub-
lic spaces that address concentration levels of
IA pollutants (1). Few building codes address
operation, maintenance, and retrofitting,
and most do not focus on airborne disease
transmission. But the COVID-19 pandemic
has made all levels of society, from commu-
nity members to decision-makers, realize the
importance of IAQ for human health, well-
being, productivity, and learning. We propose
that IAQ standards be mandatory for public
spaces. Although enforcement of IAQ perfor-
mance standards in homes is not possible,
homes must be designed and equipped so
that they could meet the standards.
For the past two decades, scientists have
called for national IAQ standards and laws
to be established (2), but so far, little action
has been taken. The approach to IA contrasts
sharply with outdoor air, for which qual-
ity is regulated and monitored and compli-
ance with regulations is enforced. The World
Health Organization (WHO) Global Air
Quality Guidelines (AQG) published in 2021
provide recommendations for concentration
levels of six pollutants and their averaging
times (PM2.5, PM10, NO2, SO2, CO, and O3)
and apply to both outdoor air and IA (3).
In cases for which IAQ standard and
guideline values were established by national
or association working groups, the outcomes
were inconsistent; often the criteria for the
same parameter differed by orders of magni-
tude. The reasons cited for limited progress
include different criteria in the selection of
the critical study, in the starting point, and
The list of author affiliations is provided in the
supplementary materials. Email: l.morawska@qut.edu.au
in the derivation procedure; the complex po-
litical, social, and legislative situation regard-
ing IAQ; the lack of an open, systematic, and
harmonized approach (4); and that establish-
ing an IAQ standard is always the result of
a compromise between scientific knowledge
and political will (5). Because of the heterog-
enous landscape of approaches needed, such
barriers remain intact despite the consider-
able IAQ research and evidence base devel-
oped over the past decades.
CHALLENGES
Source contributions
IA pollution originates from sources indoors
(including humans) and outdoors and from
chemical reactions between pollutants in IA
(6). Compliance with IAQ standards (that
refer to the concentrations of indoor pollut-
ants) would require controlling indoor emis-
sion sources (such as combustion, building
products, and cleaning products) and mini-
mizing the entry of outdoor pollutants in-
doors (for example, by filtering or treating
outdoor air to remove particles and chemical
compounds and reducing penetration of pol-
lutants through the building envelope).
During respiration, humans emit (in ad-
dition to CO2) particles that contain viruses
and bacteria. Most respiratory infections
are acquired indoors, through inhalation of
virus-laden airborne particles (7). However,
there are no exposure-response relationships
for respiratory pathogen concentrations in
IA, nor are there technologies available to
routinely monitor such pathogens in build-
ings in real time. We cannot control human
respiratory emissions in the same way that
we control emissions from other sources.
Monitoring
We cannot use the well-established ap-
proach that is used to measure outdoor
air quality to monitor IAQ. We cannot rely
on a monitoring network (in only selected
indoor public spaces) because every space
is different and is used differently, and we
cannot use modeling to predict pollution
concentration in one space by using the con-
centrations measured in other spaces. Com-
pliance monitors are too costly and complex
to deploy in all indoor spaces to monitor for
all six pollutants included in the WHO AQG
(3). However, there are environmental pa-
rameters that can already be monitored in
each room of each building, such as temper-
ature and relative humidity. The feasibility
of monitoring IAQ parameters in buildings
depends on the size, cost, robustness, and
silent operation of the sensor or monitor;
calibration; and ease of interpreting data.
But routine, real-time monitoring of indoor
pathogens is currently infeasible. In the ab-
sence of information on the concentration
of pathogens in IA, the question is which
proxy parameter or pollutant should be the
basis for legislation that targets airborne in-
fection transmission.
Legislation
Legislation comprises the system of rules—
or statutes—created and enforced by the
government of a jurisdiction. Guidelines, on
the other hand, are less formal, not manda-
tory, and generally not enforceable unless
adopted in legislation. Standards, also gener-
ally unenforceable unless they are adopted in
legislation, are typically voluntary in nature
and can set out requirements with respect
to design, operation, and performance. They
may be adopted in legislation and thus made
enforceable by law.
In terms of formal international law, there
are global treaties on transboundary air pol-
lution, but to date, no international treaty
requires or encourages adoption of ambient
air quality standards (8). It is conceptually
difficult to legislate for air quality standards
in general, let alone IAQ, because air quality
legislation is typically focused on a result or
outcome, rather than on behavior (for exam-
ple, imposing limits on pollution sources) (8).

1418 22 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

Other challenges include the scope of what to
regulate, how monitoring and enforcement
activities are undertaken, and who has re-
sponsibility for them.
At a country level, IA legislation is ham-
pered by the tremendous variability across
jurisdictions and the particulars of each
country’s legal structure. “Air pollution”
is not defined in air quality legislation in a
substantial number of countries (8). This pre-
sents a challenge for the development of laws
on IAQ. However, the United Nations (UN)
Sustainable Development Goals provide an
opportunity for global progress on IAQ (9).
Industry priorities
Many regulations reflect compromise be-
tween the needs for human protection and
for industry opportunities, with the regula-
tory process involving balanced participa-
tion from groups with different priorities
to reach consensus. There has not yet been
sufficient coordinated support to implement
IAQ regulations. The industry most closely
related to IAQ is the heating, ventilation,
and air conditioning (HVAC) industry, which
in response to market demand has evolved
to focus primarily on thermal comfort and
energy efficiency; the market has not yet de-
manded large-scale supply of technologies
to improve IAQ . Regulation could rapidly
change this demand, which may or may not
benefit the HVAC industry and many other
building industries. There will always be
some industries that do not benefit and/or
will require strategic change owing to new
regulations, so they would prefer the status
quo. There are groups who will be forced
into capital costs by regulation change (such
as property owners and their associations)
that must be convinced of need and value.
Thus, in the pursuit of new IAQ regulation,
market forces may mean that industry sup-
port is not guaranteed.
The social and political dimension
Introducing standards is complex, not only
because scientific parameters may be con-
tested or technically difficult to achieve
but also because human stakeholders have
different values, goals, and power, and
standards may have cultural or political
implications. A particular standard may be
unfeasible in any given setting (for exam-
ple, because it is unaffordable or blocked
by powerful individuals or groups), so
compromises must be made. Organizations
that choose (or are required) to implement
standards must go through a complex and
sometimes costly process to identify, as-
similate, implement, and adapt them.
ADDRESSING THE CHALLENGES
The proposed approach is based on science,
SCIENCE science.org
Proposed parameter levels
Values may be adjusted to reflect local
circumstances and priorities.
AVERAGING
LEVEL TIME OR SETPOINT
PM2.5, µg/m3 15(i) 1-hour
CO2, ppm 800 threshold
(absolute
value)(ii)
350(delta)(iii) threshold
CO, mg/m3 100(iv) 15 minutes(iv)
35(iv) 1 hour(iv)
10(iv) 8 hours(iv)
Ventilation, 14(v) When the
liters/s space
per person is occupied
(i) 24-hour level from (3). (ii) When 100% of air delivered to the
space is outdoor air, assuming outdoor CO2 concentration is
450 ppm; based on classroom scenario (see SM). (iii) Delta is
the difference between the actual CO2 concentration and the
CO2 concentration in the supply air. (iv) 8-hour averaging time,
from (15). (v) Clean air supply rate in the breathing zone; see
(12). At 25°C and 1 atm for CO 1 ppb = 1.15 g/m3. Threshold is
the concentration level of CO2 that must not be exceeded.
technology, and specific solutions that have
existed for some time and can now serve as
a basis for addressing a complex interdisci-
plinary problem.
Pollutants recommended by WHO
Low-cost sensors are a viable technology to
measure some of the six pollutants included
in the WHO AQG; however, not all six can
be realistically monitored in buildings, nor
do they all need to be monitored. The two
most relevant candidates for routine regula-
tory IAQ monitoring are PM2.5 and CO, for
which low-cost advanced sensors have dem-
onstrated stability, durability, and robust-
ness. Particulate matter in IA originates from
indoor and outdoor sources, and exposure
to PM2.5 is among the 10 leading risks (10).
CO arising from various natural processes is
present in the atmosphere at very low con-
centrations, but it is incomplete combustion
(indoor and outdoor) that can raise concen-
trations to levels harmful to humans. Indoor
CO should be routinely measured in areas
where outdoor CO concentrations exceed
regulations and where indoor combustion
takes place. In several countries, CO monitors
are mandated in spaces where combustion
takes place to alert to life-threatening levels
of gas, but these monitors are typically not
sufficiently sensitive to lower concentrations.
Carbon dioxide
Currently CO2 concentration values are
not included in the WHO AQG. However,
regardless of the potential harm it causes,
CO2 can serve as a proxy for occupant-emit-
ted contaminants and pathogens and as a
means to assess the ventilation rate. CO2
sensors are readily available, inexpensive,
and robust and can be used in all interiors.
The advantage of using CO2 as a proxy is
that although both pathogens and CO2 are
emitted during human respiratory activi-
ties, it is much easier to link CO2 concentra-
tions to these activities than to model risk
from the emissions of pathogens.
Ventilation
Ventilation with clean air is a key control
strategy for contaminants generated indoors.
The efficacy of ventilation in reducing infec-
tion risk has been demonstrated in many
studies (11). The role of ventilation is to re-
move and dilute human respiratory effluents
and body odors and other indoor-generated
pollutants at a rate high enough relative to
their production so that they do not accu-
mulate in IA. IA is replaced (diluted) with
outdoor air (assumed to be clean) or clean
recirculated air. Outdoor air ventilation rates
are almost always set according to criteria
of hygiene and comfort (perceived air qual-
ity). Effective air distribution (ventilated
air reaching the entire occupied zone and
airflow not directed from one person to an-
other) is a practical candidate for a standard.
The measured ventilation rate can be used as
a proxy of IAQ.
Although technologies for measuring ven-
tilation already exist in most modern me-
chanically ventilated buildings, monitoring
the ventilation rate in terms of clean air de-
livered to the space without considering the
number of occupants or their activities is not
sufficient to ensure adequate IAQ. One way
to assess the quality of ventilation is to con-
currently measure the CO2 concentration: If
it rises above an accepted threshold relative
to the outside concentration or concentra-
tion in the recirculated air brought into the
room, the ventilation is inadequate.
Suggested numerical levels
Below, we provide justification for proposed
numerical levels and their averaging times
for the pollutants and the parameters dis-
cussed above (see the table). Actual levels
adopted by countries and jurisdictions will
differ, reflecting local circumstances and
competing priorities.
PM2.5 concentration. It is proposed that the
WHO AQG 24 hours, 15 µg/m3 level be consid-
ered as the basis for IAQ standards, but with
a 1-hour averaging time because 24 hours is
much longer than people typically spend in
public places or, for that matter, that public
spaces are occupied. This is a compromise
between the realistic occupancy of and expo-
sure in public spaces and the need for rigor
in the derivation of the health-based value.
22 MARCH 2024 • VOL 383 ISSUE 6690 1419
INSIGHTS | P O L I C Y F O RU M
Using the WHO AQG value for 24-hour ex-
posure for 1-hour exposure is a conservative
approach that considers each environment
as though it were the only one where people
spend all their time.
CO2 concentration. To decide on a level that
would adequately control the risk of infec-
tion in public spaces, a scenario of exposure
must be defined and then a risk assessment
model be applied. We propose a scenario of
a classroom with one infected student [see
supplementary materials (SM)]. A ventila-
tion rate of 14 liter/s per person, keeping
CO2 concentrations at or below the standard
level proposed in the table, would ensure
that the reproduction number Re < 1 even
for respiratory pathogens with high trans-
missibility, such as severe acute respiratory
syndrome coronavirus 2 (SARS-CoV-2) Delta
and Omicron variants and measles. The rec-
ommended level of 800 parts per million is
within an already relatively narrow range of
values of the CO2 levels recommended by dif-
ferent organizations and countries (see SM).
This approach takes outdoor concentration
as a baseline. However, not only are outdoor
concentrations continually increasing be-
cause of emissions to the atmosphere that
outweigh removal, which must be taken into
account in the formation of the standard,
there are also variations between locations,
and at individual locations there are diurnal
and annual variations. Therefore, jurisdic-
tions should consider local CO2 baseline lev-
els when setting levels.
In indoor environments where the sup-
plied ventilation air is a mixture of outdoor
air and recirculated air, the CO2 concentra-
tion can be high, but the risk of infection
may be low provided that the supplied ven-
tilation air is sufficient. This is because the
recirculated air is often filtered, and most of
the pathogens are removed before it reenters
the space; however, gaseous pollutants, such
as CO2, are not removed by this process. The
actual (absolute) CO2 concentration in the
space and the difference between the actual
CO2 concentration and the CO2 concentra-
tion in the air delivered to the space (out-
door air delivered with natural ventilation or
air delivered by mechanical ventilation sys-
tems) are assumed as a proxy for ventilation.
Ventilation rate. The recommended rate of 14
liters/s per person, based on (12), is higher
than the WHO-recommended minimum
ventilation rate for nonresidential settings of
10 liters/s per person (3), or the highest cate-
gory I ventilation rate defined in the existing
standard ISO 17772-1. However, it is in line
with ventilation rate recommended by (11),
based on an experimental exposure study of
a cohort of school children.
1420 22 MARCH 2024 • VOL 383 ISSUE 6690
Legislation
As noted in the UN-EP 2021 report, one ad-
vantage of an IAQ regulatory framework is
the ability to place obligations on owners of
indoor premises (8). This contrasts with am-
bient air quality, which generally relates to
“unowned” air for which allocating responsi-
bility can be more difficult (2). Premises that
operate under extant legal frameworks (such
as workplaces, schools, and hospitals) may be
more amenable to regulatory control through
these frameworks (2) to consider as part of
the development of laws for IAQ (table S2).
IMPLEMENTATION OF STANDARDS
For IAQ standards to have practical value,
they must be implementable; buildings must
be designed, constructed, maintained, oper-
ated, or retrofitted to meet the standards,
given the intended use, and must be used ac-
cordingly. This should be checked at delivery
and routinely throughout the building life.
Standards must establish specifications for
IAQ and be technically feasible, affordable
to construct and operate, and compatible
with other priorities and constraints such as
energy use. Several means are available for
achieving IAQ that meets these objectives.
The use of natural or hybrid ventilation
(natural ventilation supplemented by me-
chanical ventilation when necessary) when
feasible can greatly reduce space condition-
ing energy requirements and associated oper-
ating costs. Stratified air supply (distributing
air to create vertical stratification of tempera-
ture and contaminant concentrations) by us-
ing displacement ventilation or underfloor
air supply and personal ventilation (supply
of clean air directly to the breathing zone of
each occupant) can have a positive impact.
For required delivery of outdoor air, high-ef-
ficiency air-to-air energy recovery is essential
and required by many energy standards.
Additional measures in support of ventila-
tion, such as air cleaning and disinfection,
can greatly reduce the need to increase out-
door air supply, which carries a substantial
energy penalty. Filtration of recirculated air
is an effective way to reduce concentration
of, and exposure to, airborne particulate
matter, allergens, and pathogens. Other air
treatment technologies may help inactivate
infectious airborne particles. Work is ongoing
to develop consensus methods for determin-
ing the effectiveness of some of these tech-
nologies and safety measures.
The use of demand control (modulating
control levels in response to need and acti-
vation of higher levels of protection) can be
guided by public health data, for example,
during annual influenza seasons or when a
new pathogen emerges with the potential to
cause an epidemic. The recently published
ASHRAE Standard 241–2023 Control of In-
fectious Aerosols (13) incorporates most of
the noted measures and is intended to apply
during periods of elevated risk of airborne
disease transmission.
Actions to address IAQ will add cost in
the short term and may not be prioritized by
many countries because of pressures on bud-
gets. However, if some countries lead by ex-
ample, we anticipate that IAQ standards will
increasingly become normalized. Social and
economic benefits in terms of public health,
well-being, and productivity and perfor-
mance will likely far outweigh the investment
costs in achieving clean IA. Few countries
realize the enormity of public health costs,
but disability-adjusted life years (DALYs) at-
tributable to IA pollution accounted for an
estimated 14.1% of the total DALYs in China
for the period from 2000 to 2017, and corre-
sponding financial costs (not including the
costs of IA-borne infection transmission) ac-
counted for 3.45% of China’s gross domestic
product (14). By making IAQ standards the
reality, we will improve our health and well-
being, and also save money. j
REFERENCES AND NOTES
1. L. Morawska, W. Huang, in Handbook of Indoor Air Quality,
Y. Zhang, P. Hopke, C. Mandin, Eds. (Springer, 2022), pp.
1–20.
2. R. Corsi, EM Pittsburgh-Air and Waste Management
Association, pp. 10–15 (2000).
3. WHO, “WHO global air quality guidelines: Particulate
matter (PM2.5 and PM10), ozone, nitrogen dioxide, sulfur
dioxide and carbon monoxide” (WHO, 2021).
4. L. Morawska, Proceedings of the Healthy Buildings
Conference, Espoo, Finland, 6 to 10 August 2000 (Curran,
2022), vol. 2022.
5. T. Salthammer, Chemosphere 82, 1507 (2011).
6. C. J. Weschler, Indoor Air 21, 205 (2011).
7. T. Greenhalgh et al., Lancet 397, 1603 (2021).
8. UN Environmental Program (UN-EP), “Regulating air
quality: The first global assessment of air pollution
legislation,” Air Pollution Series (UN-EP, 2021).
9. UN General Assembly (UN-GA), The 2030 agenda for
sustainable development (UN-GA, 2015).
10. J. D. Stanaway et al., Lancet 392, 1923 (2018).
11. G. Buonanno, L. Ricolfi, L. Morawska, L. Stabile, Front. Pub.
Health 10.3389/fpubh.2022.1087087 (2022).
12. J. G. Allen et al., “Proposed non-infectious air delivery
rates (NADR) for reducing exposure to airborne
respiratory infectious diseases” (The Lancet COVID-19
Commission, 2022).
13. ASHRAE, ASHRAE standard 241-2023. Control of infec-
tious aerosols (ASHRAE, 2023).
14. N. Liu et al., Lancet Planet. Health 7, e900 (2023).
15. WHO, “Guidelines for indoor air quality, selected pollut-
ants” (WHO, 2010).
ACKNOWLEDGMENTS
This paper was supported by the Australian Research
Council (ARC) Industrial Transformation Training
Centre (IC220100012) and ARC Laureate Fellowship
(FL220100082). The Engineering and Physical Sciences
Research Council (EPSRC) funded the CO-TRACE project
(grant EP/W001411/1), UK Research and Innovation [EPSRC,
Natural Environment Research Council (NERC), Australian
Human Rights Commission (AHRC)] funded the RECLAIM
Network Plus project (grant EP/W034034/1), and NERC
funded the GreenCities project (grant NE/X002799/1).
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adl0677
10.1126/science.adl0677
science.org SCIENCE

B O OKS et al .
CLIMATE CHANGE
American climate migration
Increasingly inhospitable conditions will change
the nation’s demography, argues a journalist
By Miriam R. Aczel1,2 and Michàlle E. Mor Barak3
C
urrently, less than 1% of Earth is too
hot to support human life, but re-
searchers estimate that by 2070 nearly
20% of the planet’s surface will be
outside humanity’s comfort zone. The
“bubble of unlivability” could include
up to a third of the people on Earth, and
existing inequalities will likely increase con-
flict. In the United States, vulnerable popula-
tions will be prone to disproportionate risk.
On the Move, by journalist
Abrahm Lustgarten, is a poignant
and meticulously researched ex-
ploration of climate change and
both its imminent and long-term
effects on human migration in the
US. Through analysis, personal
narratives, and projected future
scenarios, Lustgarten unveils the
stark reality of a world on the
brink of massive demographic The Overheating Earth Black and Latinx—are in the
shifts driven by an increasingly
inhospitable climate.
Lustgarten begins with a per- Farrar, Straus and Giroux,
sonal account of the moment he
recognized the climate crisis as a
PHOTO: PHILIP PACHECO/GETTY IMAGES
reality that no region will escape. His usual
view of the San Francisco skyline was re-
The reviewers are at the 1California Institute for Energy and
Environment, University of California, Berkeley, CA, USA;
2United Nations University Institute for Water, Environment
and Health (UNU-INWEH), Hamilton, Ontario, Canada; and
3Suzanne Dworak-Peck School of Social Work and Marshall
School of Business, University of Southern California, Los
Angeles, CA, USA. Email: aczel@berkeley.edu
SCIENCE science.org
placed by “a sepia-toned, smoke-filled uni-
verse,” he writes. “Just twelve miles away as
the crow flies, behind the ridge of parched
and brittle redwoods I could see from my
window, the Point Reyes National Seashore
was burning. Tall gray towers of smoke bil-
lowed upward, raining down soot.” He then
details how climate-driven migrations are
not a future possibility but rather a current
event, with historical precedents and emerg-
ing patterns that signal a profound shift in
how and where people can live.
Lustgarten predicts that Amer-
icans will see an influx of migrants
from the south and will experience
considerable internal migration as
well. Increasingly frequent and
severe wildfires, extreme weather
patterns, and sea level rise will
make some areas uninhabitable.
“The poorest neighborhoods—
On the Move: many of them predominantly
and the Uprooting lowest-lying areas, and they will
of America
Abrahm Lustgarten
suffer first,” he observes.
The book begins with an anal-
2024. 336 pp. ysis of climate change pressures,
providing a glimpse into what
life might look like within several decades.
Here, Lustgarten highlights the risks many
Americans have unknowingly accepted by
living in vulnerable areas and how eco-
nomic policies have exacerbated these risks.
In subsequent chapters, Lustgarten
projects the demographic shifts that are
likely to occur within the United States.
These include coastal residents moving
Smoke from various wildfires turned the skies orange
in San Francisco in 2020.
inland because of rising sea levels and
hurricanes, as well as migration from the
drought-stricken Midwest to regions with
stable water sources.
Economic disparities will grow as wealth-
ier people relocate to climate-resilient ar-
eas, Lustgarten writes, while the urban poor
will face increased hazards. Urban–rural
migration patterns may reverse, and large
interstate migrations are expected from
states such as California, Texas, and Florida
to the Pacific Northwest and Northeast, po-
tentially reshaping political landscapes.
Lustgarten references a 2017 study esti-
mating that climate change could cost the US
about 1.2% of its gross domestic product per
1°C of global warming, rising to a 4% loss by
century’s end—a total of around $9.3 trillion
in lost productivity per decade (1). But this
figure, he notes, “conceals the researchers’
most important finding: far more devastat-
ing loss is in store for the hottest and driest
parts of the South.” “In the poorest places the
tally for climate damages could add up to
nearly 19 percent of their per capita income
each year,” he adds.
The book concludes on a note of cautious
optimism, emphasizing that the severity of
climate change impacts, including the scale
of migration, can still be mitigated through
major and immediate actions to reduce
emissions and adapt to the changes that are
already here. “Catastrophe is not inevita-
ble,” Lustgarten writes. “Its arrival or sever-
ity depends on how quickly the world can
slow or even halt emissions, and then on
the actions that humanity and our govern-
ment leaders take to prepare for the change
that will come anyway.”
On the Move is a compelling call to un-
derstand, prepare for, and act on the climate
crisis. It challenges readers to confront the
reality of what is coming and to consider the
role each of us plays in shaping the future.
Combining rigorous research and engaging
storytelling, Lustgarten provides a frame-
work for navigating the uncertain terrain of
a warming world.
Amid the growing body of literature on
climate change, Lustgarten’s book provides a
crucial examination of the impacts—realized
and projected—in our own backyards and
how these changes are remaking society.
Importantly, it is both a call to action and
a blueprint for how to weather the coming
storm, highlighting historical injustices and
charting an equitable path forward. j
REFERENCES AND NOTES
1. S. Hsiang et al., Science 356, 1362 (2017).
10.1126/science.adn9371
29 MARCH 2024 • VOL 383 ISSUE 6690 1421
INSIGHTS | B O O K S
ENGINEERING
Framing tough problems
An engineer urges readers to think more holistically
about complex challenges
By Donna Riley
S
cience and technology studies schol-
ars such as Gary Downey have chal-
lenged engineers to reject the “given:
find” routine of predetermined prob-
lem solving and to move toward ac-
tive involvement in defining better
problems (1). In Wicked Problems, Guru
Madhavan offers a primer on this skill, tak-
ing the reader through a series of thought
experiments in problem framing that are
intended to empower both engineers and
non-engineers as engaged civic actors.
Madhavan employs a taxonomy that will
be familiar to many readers,
classifying problems as “hard,”
“soft,” “messy,” or “wicked.”
These categorizations chal-
lenge us to be successively
more holistic and integra-
tive—or, to use a more precise
term, more convergent—in our
thinking, and they encourage
readers to stay focused, criti-
cal, and reflective in their con-
sideration of problem framing
rather than rushing to propose
solutions. I could not read this
book passively and found my-
self framing and reframing
the problems it referenced as
hard, soft, messy, wicked, and
back again. While Madhavan
does not explicitly invoke this
critical reading strategy, I can
imagine it being particularly
effective in college classrooms.
The book is a whirlwind journey through
a broad range of historical events—from
Boston’s Great Molasses Flood of 1919 to
9/11. At every turn, readers are challenged
to develop a wicked ethics imagination to
deepen their sociotechnical understanding
and dream up new realities. Comparisons to
Henry Petroski or Samuel Florman are apt
here. Like those authors, Madhavan’s edu-
cational breadth is evident throughout the
book, and he strives to offer an accessible
version of sociotechnical analysis.
Madhavan anchors the book around the
The reviewer is at the School of Engineering, University of
New Mexico, Albuquerque, NM 87131, USA.
Email: riley1@unm.edu
1422 29 MARCH 2024 • VOL 383 ISSUE 6690
Wicked Problems:
How to Engineer a
Better World
Guru Madhavan
Norton, 2024. 384 pp.
story of a lesser-known inventor, Edwin
Link, who helped shape modern aviation
through the development of systems-based
flight trainers. By integrating technological
and psychological aspects of pilot training
in a pioneering virtual modality, the train-
ers helped pilots visualize and experience
critical aspects of flight before they were in
the air. The analogy here is direct: We must
train in problem framing and sociotechni-
cal thinking before we are asked to build
the proverbial plane while flying it so that
we can learn to anticipate the “unforeseen”
consequences of the various technologies
we create.
Edwin Link’s (right) virtual flight trainers helped pilots anticipate problems safely.
My favorite chapter takes up the issue of
maintenance: In our obsession with innova-
tion and all things new, we sometimes for-
get the critical importance of maintenance,
and local communities often pay the price
of deferred maintenance and planned obso-
lescence. For the past several years, a group
of science and technology studies scholars
including Lee Vinsel and Andrew Russell
have been discussing this problem (2), and
Madhavan’s book brings this conversation
to a broader audience. With the recent suc-
cess of right-to-repair legislation and new
investments in infrastructure, it is timely
to incorporate maintenance into engineer-
ing education and into our broader under-
standing of technology in society.
Bringing the work of science and technol-
ogy studies scholars to a wider audience is
admirable indeed. At the same time, I could
not help but identify missed opportunities in
the text for more. For example, in discuss-
ing the organizational behavior problems
leading up to the Challenger and Columbia
disasters, he misses a reference to Diane
Vaughan’s essential work on the subject (3).
One key skillset in the service of defining
better problems is understanding and deftly
navigating the power relations inherent
in wicked sociotechnical problems, where
unclear boundaries, uncertainty about un-
certainty, and values conflicts predominate.
Although Madhavan demon-
strates an awareness of how
power relations have acted
through history, at times I
wished he would have walked
readers through a problem
from historical contexts to so-
ciotechnical implications. He
recognizes that there would
have been no Great Molasses
Flood without the sugar-rum-
slavery triangle, for example.
But if these historical anteced-
ents are foundational to fram-
ing the problem, how could
one apply anti-racist refram-
ings? There are a number of
missed opportunities in the
book to connect engineer-
ing and empire and to ask
how today’s systems could be
otherwise.
As we confront major chal-
lenges of this century in a changing climate
or the emergence of artificial intelligence
and increasingly automated systems, engi-
neers’ role in society will increasingly be
one of technology curator and interpreter.
Wicked Problems is a reminder that to be a
force for good in the world, we need—and
long have needed—the intellectual power
developed through the breadth of a liberal
education. j
PHOTO: BETTMANN ARCHIVE
REFERENCES AND NOTES
1. G. Downey, Chem. Eng. Res. Des. 83, 583 (2005).
2. The Maintainers; https://themaintainers.org/.
3. D. Vaughan, The Challenger Launch Decision (Univ. of
Chicago Press, 1996).
10.1126/science.ado0913
science.org SCIENCE
Pushing the Boundaries of Knowledge
As AAAS’s first multidisciplinary, open access journal, Science Advances publishes
research that reflects the selectivity of high impact, innovative research you expect
from the Science family of journals, published in an open access format to serve
a vast and growing global audience. Check out the latest findings or learn how to
submit your research: ScienceAdvances.org

G O L D O P E N AC C E S S , D I G I TA L , A N D F R E E T O A L L R E A D E R S

LET TERS
Edited by Jennifer Sills
Chile’s Valparaíso
hills on fire
Although naturally occurring wildfires have
been infrequent in Chile (1), human-caused
ignitions, flammable plantations, and pro-
longed droughts now make the region one
of the most fire-prone areas in the world (2).
In February, a record-setting wildfire burned
through thousands of hectares of dry,
stressed vegetation in the Valparaíso region
(3), destroying thousands of homes and lead-
ing to 133 human fatalities (4). Chile must
take steps to mitigate the disastrous impacts
of such wildfires (5).
Central Chile has experienced an intense
and uninterrupted drought since 2010,
which has increased the size and severity
of human-caused wildfires (6–8). In the
Valparaíso region, large and catastrophic
fires have burned frequently at the interface
of wildland and urban areas during the past
few years (9, 10), and the 2024 Valparaíso
hills wildfire was the most damaging (4).
Starting inside the Lago Peñuelas National
Reserve, the allegedly intentional fire rapidly
propagated as a result of high temperatures,
low humidity, strong winds, and abundant
dry fuels (11). A rugged topography with
unregulated housing development, lack of
vegetation management, and fire-prone
pine and eucalyptus plantations created the
perfect ingredients for the spread of this
catastrophic fire (11).
Predictions indicate that fire weather con-
ditions in central Chile will remain favorable
to megafires (12). Mitigating the threat of
wildfire will require robust governance and
land-use planning. Chile must create less
1424 29 MARCH 2024 • VOL 383 ISSUE 6690
fire-prone landscapes by restoring and man-
aging native forest vegetation and removing
highly flammable forest plantations, espe-
cially those close to wildland–urban area
interfaces. Reducing fire-prone landscapes
should include prohibiting the conversion of
recently burned native forests into exotic for-
est plantations or new urban developments.
The government should also strengthen
fire prevention programs to reduce human-
caused ignitions. Property owners and the
Forest Service must share responsibility to
reduce flammable fuels and to build defen-
sive firebreaks at the interface of wildland
and urban areas. Multiple measures will be
required to increase resilience and safety
in the face of climate change and extreme
wildfire events.
Mauro E. González1,2,3*, Alexandra D. Syphard4,5,
A. Paige Fischer6, A. A. Muñoz3,7,8, Alejandro
Miranda3,9
1Instituto de Conservación, Biodiversidad
y Territorio, Universidad Austral de Chile,
5090000 Valdivia, Chile. 2Center for Fire and
Socioecosystem Resilience (FireSES), Universidad
Austral de Chile, 5090000 Valdivia, Chile. 3Center
for Climate and Resilience Research (CR2),
8370451 Santiago, Chile. 4Conservation Biology
Institute, Corvallis, OR 97331, USA. 5Department of
Geography, San Diego State University, San Diego,
CA 92110, USA. 6Western Forest and Fire Initiative,
School for Environment and Sustainability
University of Michigan, Ann Arbor, MI 48109, USA.
7Laboratorio de Dendrocronología y Estudios
Ambientales, Instituto de Geografía, Pontificia
Universidad Católica de Valparaíso, 2340025
Valparaíso, Chile. 8Centro de Acción Climática,
Pontificia Universidad Católica de Valparaíso,
2340025 Valparaíso, Chile. 9Laboratorio
de Ecología del Paisaje y Conservación,
Departamento de Ciencias Forestales, Universidad
de la Frontera, 4811230 Temuco, Chile.
*Corresponding author.
Email: maurogonzalez@uach.cl
REFERENCES AND NOTES
1. J. E. Keeley, W. J. Bond, R. A. Bradstock, J. G. Pausas, P.
W. Rundel, Fire in Mediterranean Ecosystems: Ecology,
The neighborhood El Olivar,
Viña del Mar, was badly burned
during Chile’s Valparaíso fire.
Evolution and Management (Cambridge Univ. Press,
2012).
2. Nat. Clim. Chang. 7, 755 (2017).
3.
4.
A. Miranda et al., Nat. Plants 9, 1810 (2023).
A. Correal, J. Bartlett, “‘We’re broken’: Wildfires on
Chile’s coast kill 112 and leave hundreds missing,” New
York Times, 4 February 2024.
5. A. Duane et al., Clim. Change 65, 165 (2021).
6. R. Garreaud et al., Hydrol. Earth Syst. Sci. 21, 6307
(2017).
7. M. E. González et al., Ecosphere 9, e02300 (2018).
8. D. M. J. S. Bowman et al., Ambio 48, 350 (2019).
9. A. Miranda et al., Earth Syst. Sci. Data 14, 3599 (2022).
10. P. Sarricolea et al., Sci. Total Environ. 706, 135894 (2020).
11. J. Kimutai et al., “Despite known coastal cooling trend,
risk of deadly wildfires in central Chile increasing with
changing land management in a warming climate,”
Scientific Report by The World Weather Attribution
(Imperial College London, 2024); https://spiral.
imperial.ac.uk/handle/10044/1/109375.
12. Intergovernmental Panel on Climate Change, “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 et al., Eds. (Cambridge Univ. Press, 2021).
10.1126/science.ado5411
China’s plan to combat
antimicrobial resistance
China, the leading global producer and
consumer of antibiotics, faces growing
antimicrobial resistance as a result of
inappropriate use of antibiotics in clinical,
agricultural, livestock, and aquacultural set-
tings (1, 2). China’s first national action plan
to contain antimicrobial resistance yielded
positive outcomes, but the prevalence of
clinical multidrug-resistant bacteria in China
remains high (3). The effects of global cli-
PHOTO: ARIEL A. MUÑOZ
mate change and pollution are exacerbating
the problem (4). China has now announced
a second national action plan (5), with the
goal of helping to prevent the global spread
of multidrug-resistant bacteria.
science.org SCIENCE

Antimicrobial resistance is now common
in China and across the world. In 2023, the
China Antimicrobial Surveillance Network
reported antimicrobial resistance rates (the
percentage of antibiotic-resistant strains
found in all strains identified within a given
period of time) above 50% for third-gener-
ation cephalosporin-resistant Escherichia
coli and 29.6% for methicillin-resistant
Staphylococcus aureus, as well as high
rates of meropenem-resistant Acinetobacter
baumannii (73.7%), Klebsiella pneumoniae
(26%), and Pseudomonas aeruginosa (17.4%)
(6). The prevalence of carbapenem-resistant
Gram-negative bacteria also increased, caus-
ing life-threatening infections, especially
when such bacteria exhibit resistance to
last-resort antibiotics, such as tigecycline
and colistin (7).
Climate change has led to conditions
that favor increased antimicrobial resis-
tance, including mounting temperatures
and humidity (8, 9). Fine particulate matter
(PM2.5), which contains diverse antibiotic-
resistant bacteria and antibiotic resistance
genes, is also associated with the occurrence
of antimicrobial resistance (10). An annual
increase of 10% in PM2.5 concentration is
projected to increase antimicrobial resis-
tance by 1.1% (10).
China’s new national action plan to
combat antimicrobial resistance focuses
on effective stewardship, rational use of
antimicrobials, and establishment of active
surveillance systems (5). Fully implement-
ing this plan is important, but additional
measures will be required, including the
implementation of the One Health frame-
work (11) and intensifying the research
and development of new antimicrobial
agents. China also needs to promote
rapid diagnostic techniques and establish
comprehensive surveillance systems that
provide whole-genome sequencing and
source-tracking for multidrug-resistant
bacteria (12). Most importantly, China
should take steps to optimize antimicrobial
prescription practices in hospitals and
to heighten public awareness about the
appropriate use of antibiotics.
Xinyang Li1, Yuye Wu1, Tian Jiang2, Bin Chen3, Rui
Feng4, Jun Zhang1, Xinyou Xie1, Zhi Ruan1*
1Department of Clinical Laboratory, Sir Run Run
Shaw Hospital, Zhejiang University School of
Medicine, Key Laboratory of Precision Medicine
in Diagnosis and Monitoring Research of Zhejiang
Province, Hangzhou, China. 2Department of
Clinical Laboratory, Affiliated Wenling Hospital,
Wenzhou Medical University, Wenling, China.
PHOTO: MARGARET HANDLEY
3Assisted Reproduction Unit, Department of
Obstetrics and Gynecology, Sir Run Run Shaw
Hospital, Zhejiang University School of Medicine,
Key Laboratory of Reproductive Dysfunction
Management of Zhejiang Province, Hangzhou,
China. 4College of Environmental and Resource
Sciences, Zhejiang University, Hangzhou, China.
*Corresponding author. Email: r_z@zju.edu.cn
SCIENCE science.org
INSIGHTS
REFERENCES AND NOTES http://www.chinets.com/Content/File/CHINET%20
1. A. Mann et al., Curr. Res. Microb. Sci. 2, 100030 (2021). 2023%E5%B9%B4%E7%BB%86%E8%8F%8C%E8
2. S. Chen, J. Zhang, Y. Wu, China CDC Wkly. 5, 492 (2023). %80%90%E8%8D%AF%E7%9B%91%E6%B5%8B%
E7%BB%93%E6%9E%9C.pdf [in Chinese].
3. L. Ding, F. Hu, J. Antimicrob. Chemother. 78, 558 (2023).
4. C. Wong, Nature 10.1038/d41586-023-04077-0 (2024). 7. H. Giamarellou, I. Karaiskos, Antibiotics 11, 1009 (2022).
5. National Health Commission, “China: Second 8. R. Pallares-Vega et al., Front. Microbiol. 12, 656250 (2021).
national action plan for containing antimicrobial 9. W. Li et al., Lancet Region. Health West. Pacific 30,
100628 (2023).
resistance 2022–2025” (World Health Organization,
2022); https://www.who.int/publications/m/item/ 10. Z. Zhou, Lancet Planet. Health 7, e649 (2023).
china-second-amr-national-action-plan-2022-2025. 11. Z. Ardakani et al., One Health 17, 100647 (2023).
6. China Antimicrobial Surveillance Network (CHINET), 12. Y. Feng et al., Nucleic Acids Res. 49, D644 (2021).
“CHINET 2023 bacterial resistance monitoring
results (January–December 2023)” (CHINET, 2023); 10.1126/science.ado5186
PAST AS PROLOGUE
The fern mandolin
My dad’s prized possession was a Gibson 1920s Lloyd Loar fern mandolin. Inlaid
with mother of pearl, the handmade instrument made a marvelous sound, perfect
for playing traditional bluegrass melodies. As kids, my brother and I would visit
the fern mandolin in our dad’s storage locker—the type with a big pull-up door. In
hindsight, the unique, delicate piece seemed out of place in the vast industrial stor-
age park landscape, but the locker was my dad’s only permanent address. He couch-
surfed, rented rooms, and sometimes lived out of his car. People called him a drifter.
But when he played the fern mandolin and sang, his captivating grin filled the space.
There was so much promise in that locker!
In his youth, my dad had played in a bluegrass band. He often sang songs that
reflected his rural background growing up in Texas and his experience moving to
California during the Dust Bowl of the 1930s. The folk song “Freight Train,” about the
lonesome life of riding on the cross-country
freight trains, was also one of his favorites.
When I was in my early 20s and just
starting my studies to become an epidemi-
ologist, my dad disappeared. My brother
and I searched for him at his most recent
address and at local flea markets, where he
liked to go to trade guitar and bicycle parts.
We never found him.
Now, I work with an interdisciplinary
team of investigators and clinicians, focus-
ing on understanding and preventing home-
lessness. We’ve explored when stays with
The fern mandolin that belonged to the family members do and don’t work out, and
author’s father recently surfaced at a rare we’ve worked with a cohort of unhoused
instrument auction. people over time to learn about how their
needs change, and how they care for each
other, as they age. The idea of helping people stay in housing, or improving their lives
as they revolve around the uncertainty of a storage locker existence, keeps me going on
my research-weary days.
I often remember how natural it seemed to spend time with my dad at his locker
and how unimportant it was to me that he didn’t always have an address. Belongings in
lockers, on the street, or seized during evictions reveal people’s complex inner lives, his-
tories, and talents. I want our society to appreciate, as I try to, the richness of the lives
of those who are unhoused or marginally housed.
We never knew what happened to my dad’s storage locker, but a few years ago, the
fern mandolin resurfaced at a rare instrument auction. My brother and I went to see
it. It was still beautiful.
Margaret A. Handley
University of California, San Francisco, San Francisco, CA, USA. Email: margaret.handley@ucsf.edu
10.1126/science.adn4676
Call for Submissions Past as Prologue is an occasional feature highlighting the role of family history in the life of
scientists. What role did your family background play in your decision to pursue science, your field, or your career?
Submit your story to www.submit2science.org.
29 MARCH 2024 • VOL 383 ISSUE 6690 1425
 AAAS NEWS & NOTES
Theresa Maldonado is newest AAAS president-elect
Morton Ann Gernsbacher and Juan S. Ramírez Lugo reelected to AAAS Board of Directors
By Andrea Korte
Members of the American Association for the Advancement of Sci-
ence have elected Theresa Maldonado, vice president for research and
innovation at the University of California, to serve as the organization’s
president-elect.
AAAS members also elected two members of its Board of Direc-
tors to new terms: Morton Ann Gernsbacher of the University of
Wisconsin–Madison and Juan S. Ramírez Lugo of the University of
Puerto Rico, Río Piedras.
“In selecting Theresa Maldonado, Morton Ann Gernsbacher, and
Juan Ramírez Lugo, the AAAS community should have great confi-
dence in its elected leadership for the next few years,” said Sudip S.
Parikh, chief executive officer of AAAS and executive publisher of the
Science family of journals. “Each of them brings a unique perspective
and many strengths to the Board, and I’m glad to
have them as part of our leadership team.”
Members cast their votes between Jan. 24 and
Feb. 5, and terms begin immediately. Maldonado
will serve a 1-year term as president-elect, followed
by 1 year as AAAS president and 1 year as immedi-
ate past-president. Willie E. May of Morgan State
University, elected in 2023, is now AAAS president,
and Keith Yamamoto of the University of California,
San Francisco, is now immediate past-president.
Gernsbacher and Ramírez Lugo will serve 4-year
terms as members of the Board.
Gilda Barabino, chair of the AAAS Board of Direc-
tors and president of Olin College of Engineering,
celebrated the news: “I am delighted to welcome
Dr. Maldonado to AAAS as president-elect. As a
champion for engineering and for a more inclusive
scientific enterprise, Theresa is an outstanding addition to the Board. I
am very much looking forward to working with her. I am also so pleased
that we get to keep the expertise and comradery of Morton and Juan
at our board table for the next 4 years. They have both made vitally
important contributions to our deliberations during their time on
the Board, and I look forward to their continued input and leadership.”
Maldonado, who joined the University of California Office of the
President in 2020 from the University of Texas at El Paso, leads large-
scale efforts across the UC system’s 10 campuses, three UC-managed
Department of Energy national laboratories, UC Agriculture & Natural
Resources, and UC Health. She describes herself as a “dot connector
and systems thinker” and sees AAAS serving several unique roles in a
broadened scientific enterprise: as an advancer of systems thinking,
as an aligner of the competing interests in the many communities—
including academia, industry, and government—that make up the
scientific enterprise; and as a trusted listener.
“The potential to authentically engage different people, diverse
thoughts, and systems thinking is tremendous; we just need the align-
ment and sense of urgency,” Maldonado said in her candidate state-
ment. “AAAS is at a pivotal moment as a professional society to enable
and guide the scientific enterprise, policy-makers, and an inclusive
global scientific citizenry to reimagine and influence future directions
for the benefit of society.”
1426 29 MARCH 2024 • VOL 383 ISSUE 6690
Gernsbacher, the Vilas Research Professor and Sir Frederic C.
Bartlett Professor of Psychology at the University of Wisconsin–
Madison, has served in many AAAS governance roles, most recently
as the chair of the AAAS Council while representing Section J:
Psychology. In her role as chair of the Council, she served simulta-
neously as a member of the Board. The Council is responsible for
electing Fellows and affiliates and adopting resolutions and state-
ments on matters affecting AAAS. She has also previously served
as a member-at-large and a member of the Electorate Nominating
Committee for the section and as a member of AAAS’s Governance
Modernization Working Group.
Gernsbacher, whose research focuses on the cognitive neurosci-
ence of human communication and attention, was elected as a Fellow
of AAAS in 1995.
Said Gernsbacher in her candidate statement, “I am most excited
to play a role in more fully and broadly engaging
AAAS members, which will in turn strengthen
AAAS’s contributions to the scientific enterprise.
I’m eager to involve more members in advocat-
ing for scientific evidence; science diplomacy
and education; and human rights, ethics, and law.
By enlivening and increasing our already strong
membership, I believe we will be better positioned
to accomplish our important goals.”
Ramírez Lugo, associate professor of biology
and associate dean for academic affairs in the Col-
lege of Natural Sciences at the University of Puerto
Rico, Río Piedras, was appointed to the AAAS
Board of Directors in 2022 to complete the unex-
pired term of Alondra Nelson, who stepped down
upon her appointment to the White House Office
of Science and Technology Policy. In addition to
his service as a Board member, Ramírez Lugo is a past president of
AAAS’s Caribbean Division and served as a delegate to the AAAS
Council from 2017 to 2020. Like Gernsbacher, he was a member of the
Governance Modernization Working Group.
“AAAS, serving as the nexus for scientists globally, occupies a
unique position as a catalyst for transformative action. Its potential
lies in cultivating a platform where interactions are not only fluid
but also frequent, fostering equal agency and accountability across
academia, industry, and government. This platform evolves into a
space where voices seamlessly traverse both vertically and horizon-
tally across governance structures and, significantly, among AAAS
members,” said Ramírez Lugo in his candidate statement.
Parikh also acknowledged the other candidates on the AAAS
ballot. “I would also like to thank Caroline Montojo and John Drazan
for agreeing to stand for election,” Parikh said. “They were outstand-
CREDIT: UNIVERSITY OF CALIFORNIA
ing candidates, and we are looking forward to working with them in
other capacities to pursue our shared mission of advancing science in
service to society.”
AAAS members also voted for the following positions within each
of AAAS’s 24 disciplinary Sections: Section Chair, Council Member,
Nominations/Leadership Development Chair, Early-Career Represen-
tative, and Non-Academic Representative. Full results are available
on AAAS.org.
science.org SCIENCE
CAL L FO R PAP E RS

Cyborg and Bionic Systems Beijing Institute

of Technology (BIT) and distributed by the American Association for the Advancement of Science (AAAS). The
journal publishes original, peer-reviewed articles based on fundamental, applied science, or their interaction. Cyborg
and Bionic Systems promotes the knowledge interchange and hybrid system codesign between living beings and

systems (CBS), mainly including robotics, biomedical engineering and neuro-engineering.

Submit your research to Cyborg and Bionic System today!

Learn more at spj.science.org/journal/cbsystems

The Science Partner Journal (SPJ) program was established by the American Association for the Advancement of Science (AAAS), the
Science family of journals. The SPJ program features high-quality, online-only, Open Access publications produced
in collaboration with international research institutions, foundations, funders and societies. Through these collaborations, AAAS furthers its

Visit us at spj.science.org

@SPJournals @SPJournals and @CBSystems_

A R T I C L E P R O C E S S I N G C H A R G E S WA I V E D U N T I L J U LY 2 0 2 3
 RESEARCH
IN S CIENCE JOURNAL S
Edited by Michael Funk

PLANT SCIENCE

Controlling gramine levels in barley

T
he defensive alkaloid gramine in barley and other grasses
is important for chemical defense against insects, but it
makes grasses less palatable to ruminants. One impor-
tant goal for breeders is to arrive at a balance where both
the protective and the palatability features of grasses are
maintained through targeted gramine engineering. Dias et al.
used pan-genome sequencing of barley to identify a cluster
of two genes for gramine biosynthesis and characterized an
enzyme that carries out a cryptic oxidative rearrangement of
tryptophan. The authors were able to modulate gramine levels
in yeast, model plants, and barley varieties, thus demonstrat-
ing the potential for integrating synthetic biology and genetic
engineering aimed at targeted optimization of gramine-linked
traits. —AWa Science p. 1448, 10.1126/science.adk6112

Unpalatable alkaloids in barley and other grasses are

important targets for plant engineering.

HYDROGELS
Biaxially super-
stretchable hydrogels
Hydrogels are typically com-
posed of cross-linked networks
of polymers that are highly
swollen with water. Depending
on the specific architecture, they
may be able to undergo a large
degree of stretching, although
usually only in one direction.
Chen et al. developed a polyelec-
from mechanical damage. —MSL
Science p. 1455, 10.1126/science.adh3632
ULTRAFAST OPTICS
Dynamics of
photoelectron emission
Diffraction of light from a matter
grating stems from the wave
property of electromagnetic
radiation. Likewise, the wave
properties of electrons prompted
effect should provide a valuable
tool for ultrafast electron spec-
troscopy studies. —ISO
Science p. 1467, 10.1126/science.adn1555
METHANE EMISSIONS
Gassy trash
Methane is the most important
trace greenhouse gas after
carbon dioxide, and anthropo-
genic emissions account for
more than half of the global total.
majority of sites. These results
underline the need for better
monitoring of landfill emissions to
help guide climate change mitiga-
tion policy. —HJS
Science p. 1499, 10.1126/science.adi7735
MACHINE LEARNING
How neural networks
learn features
How neural networks learn
PHOTO: LADEMANNMEDIA/ALAMY STOCK PHOTO

trolyte hydrogel that will form a
“pearl necklace” structure in the
relaxed state. The pearl regions,
which form between cross-links,
contain a reservoir of chain
segments that can uncoil when
the polymer is stretched. This
structure allows the hydrogels to
be reversibly biaxially stretched
to a large areal strain and to heal
the proposal from Kapitza and
Dirac in 1933 that electrons
should be diffracted from a strong
standing wave of light. Using an
intense pulsed standing wave, Lin
et al. added a temporal dimension
when probing the evolution of a
photoelectron wave packet. Such
phase sensitivity of the demon-
strated dynamical Kapitza-Dirac
Landfills containing solid waste
are potentially major sources of
methane, but their importance
has remained poorly constrained.
Cusworth et al. report data
gathered by airborne imaging
spectrometers from about 20%
of open US landfills to show
that considerable point source
emissions can be detected at a
problem-specific features (pat-
terns in data) and how features
emerge through training remain
major unsolved problems in
machine learning. Understanding
this mechanism provides the
opportunity to design networks
with improved reliability and the
model transparency needed for
various practical applications.

1428 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 Radhakrishnan et al. proposed
the deep neural feature ansatz,
which states that neural feature
learning occurs by up-weighting
the features that are most influ-
ential on model output, a process
that was formulated mathemati-
cally in terms of the average
gradient outer product and was
supported by numerical experi-
ments and theoretical results.
The presented mechanism
provides a backpropagation-free
approach for feature learning in
various machine learning models,
including those that previously
had no such capabilities. —YS
Science p. 1461, 10.1126/science.adi5639
NEUROSCIENCE
Measuring oxygen
in the brain
The brain must finely balance
delivery and demand of oxygen
to constantly maintain tissue
oxygenation. However, our under-
standing of the dynamics of brain
tissue oxygen tension during
HUMANOID ROBOTS
IN OTHER JOURNALS Edited by Caroline Ash
Robot face mirrors human and Jesse Smith
expressions
Humanoid robots are capable of
mimicking human expressions
by perceiving human emotions
and responding after the human
has finished their expression.
However, a delayed smile can
feel artificial and disingenuous
compared with a smile occur-
ring simultaneously with that of
a companion. Hu et al. trained
an anthropomorphic facial
robot named Emo to display
an anticipatory expression to
match its human companion.
Emo is equipped with 26 motors
and a flexible silicone skin to
provide precise control over its
facial expressions. The robot
was trained with a video dataset
of humans making expressions.
By observing subtle changes in
a human face, the robot could
predict an approaching smile
839 milliseconds before the PLANT SCIENCE
human smiled and adjust its face
to smile simultaneously. —MY
Sodium flux in plants

physiological conditions is still
limited. Beinlich et al. used a bio-
luminescent oxygen indicator to
examine oxygen partial pressure
in different parts of the mouse
brain at high spatial and temporal
resolution. They discovered that
transient and spatially restricted
periods of hypoxia occurred
spontaneously, a phenomenon
they named “hypoxic pockets.”
Systematically exploring the
response to various experimental
conditions indicated that physical
activity such as running reduces
the occurrence of hypoxic
regions. —PRS
Science p. 1471, 10.1126/science.adn1011
PHOTO: FELIX BEINLICH
Sci. Robot. (2024) 10.1126/scirobot.adi4724
MULTIPLE SCLEROSIS
Toward personalized
therapy
The autoimmune disease
multiple sclerosis (MS) is a
highly heterogeneous disease
with many different treatment
options. However, it is not clear
whether certain features of MS
are associated with distinct
immune signatures or if they
would benefit from particular
therapies. Gross et al. used
peripheral blood mononuclear
cells and serum collected from
two independent cohorts of
patients with MS to identify
three endophenotypes of the
disease. These peripheral blood
immune signatures distin-
guished patients with distinct
clinical disease trajectories and
were correlated with efficacy
of interferon-b treatment.
Peripheral blood analysis could
E
fficient management of sodium flux is vital to plants
to enable salinity tolerance and growth in saline soils.
Sodium flux in plants is controlled by loading and
unloading of sodium into and out of the xylem, a trans-
port tissue in vascular plants. Loading and unloading
is mediated by the plasma membrane sodium/hydrogen
antiporter SOS1 and by the channel-like HKT1;1 transporter
protein, respectively. However, how these opposing transport
systems are co-regulated is not well understood. Gámez-
Arjona et al. show that under salt stress, Arabidopsis SOS3
recruits SOS1 to the plasma membrane, where it is activated
by the SOS2/3 kinase complex. Concurrently, SOS3 interacts
with HKT1;1 and degrades it. Thus, SOS3 acts as a molecular
switch shifting the balance between sodium retention and
export into and out of the root. —AWa
Proc. Natl. Acad. Sci. U.S.A. (2024) 10.1073/pnas.2320657121
Xylem, pictured here by light microscopy, transports dissolved
salts from a plant’s roots to its other parts.
VASGUT-BRAIN AXIS is unknown. Espinosa-Oliva et al.
analyzed postmortem gut and
Irritable gut–brain brain samples from subjects
pathology diagnosed with inflamma-
Inflammatory bowel diseases are tory bowel disease and found
chronic intestinal disorders such accumulation of pathogenic
as ulcerative colitis and Crohn’s a-synuclein aggregates in the gut

Composite microscopy image of
a mouse brain expressing a
bioluminescent oxygen sensor
thus be used to guide personal-
ized treatment regimens for
patients with MS. —CM
Sci. Transl. Med. (2024)
10.1126/scitranslmed.ade8560
disease, which are characterized
by chronic systemic inflamma-
tion. However, the effect of these
diseases on the brain in relation
to neurodegenerative diseases
and the midbrain. Similarly, in a
rat model of gut inflammation, the
authors found that a-synuclein
accumulation and increased brain
inflammation were prevented by

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1429

RESEARCH | I N O T H E R J O U R NA L S

CONSERVATION

Fallow deer map human fortunes

uropean fallow deer were initially domesticated about the same time as other livestock and now occur globally, in some places becoming

E an invasive species. The presence of sustained wild populations complicates their study and conservation status. Baker et al. have
coupled historical records with mitochondrial sequences and biomolecular data from ancient and modern fallow deer. Their analysis
narrows down fallow deer origins to two clades found in glacial refugia in the Balkans and Anatolia. Humans transported fallow deer over
long distances, initially as part of the Greco-Roman worship of Artemis and Diana, and later during colonialism as status symbols. —CNS
Proc. Natl. Acad. Sci. U.S.A. (2024) 10.1073/pnas.2310051121

Fallow deer have a complicated genetic history and conservation status as a result of domestication and transport of the European species Dama dama.

cutting the vagus nerve. These
results suggest that chronic gut
inflammation might promote
pathogenic a-synuclein accumu-
lation in the brain. —MMa
Neuropathol. Appl. Neurobiol. (2024)
10.1111/nan.12962
RHIZOSPHERE
Amending soil microbiota
Plants struggle to grow in
dense clay soils that waterlog,
compact, and exclude aera-
tion. Therefore, farmers apply
nutrient-rich amendments to
improve soil structure in clayey
soil using a process called deep
banded amendment. How soil
microorganisms respond to
such amendments is poorly
understood. Vido et al. tested
various chemical and complex
growth expansion prompted by
the nutrient amendments, which
together helped to redistribute
microbiota and then drive further
nutrient cycling. —CA
Environ. Microbiol. (2024)
10.1111/1462-2920.16587
OPTICAL TWEEZING
Going large with optical
tweezer arrays
Optical tweezers are versatile
tools for trapping and manipulat-
ing a variety of quantum species.
For the development of practi-
cal quantum simulators, the
number of trapping sites in an
array should be on the order of
several thousand or more and
be individually addressable. Two
groups independently present
an acousto-optic deflector, spatial
light modulator, and digital mirror
deflector to form trap arrays
exceeding 10,000 sites (but not
yet loaded with atoms). Both of
these approaches are encour-
aging for the development of
large-scale quantum simulation
and processing architectures.
—ISO
Optica (2024) 10.1364/
OPTICA.513551, 10.1364/OPTICA.512155
ECONOMICS
Closing the gender gap
in pay
For similarly qualified candidates
in an online engineering job mar-
ket, a 1.4% gender gap in starting
salary was fully accounted for by
women initially asking for 2.9%
gender gaps in initial requested
salaries and starting salaries
for new hires without reducing
the number of offers made to
women. —BW
Q. J. Econ. (2024)
10.1093/qje/qjae004
THERMOELECTRICS
Controlled migration
Waste heat, especially at high
temperature, can be recovered
with thermoelectric devices.
However, developing stable
and efficient materials at high
temperature is difficult. Hu et
al. developed a more stable
high-temperature copper sel-
enide thermoelectric. The high
mobility of copper ions boosts
the thermoelectric properties by
decreasing the thermal conduc-
PHOTO: CHRISTIAN HÜTTER/ALAMY STOCK PHOTO

organic interventions in controlled
environment column experiments
and used 16S rRNA amplicon
sequencing to monitor changes in
microbial communities. Whereas
the amendments all changed
microbial communities, the
changes were more profound in
the presence of growing plants.
The authors suggest that the
synergism is caused by plant root
work that takes the trap arrays to
that level. Pause et al. adopted a
tiling technique in which multiple
microlens-generated tweezers
were interleaved to generate
engineered arrays containing
over 3000 trapping sites, which
they then loaded with over 400
rubidum atoms. Zhang et al.
introduce an optical modulation
method using a combination of
less pay than men did. Roussille
leveraged data from Hired.com
reflecting about 7600 hires,
6500 firms, and 114,000 job
seekers who had to fill an empty
field with their desired salary. In
mid-2018, the site began prefill-
ing this salary request with the
median initial salary offered by
firms to similar candidates over
the prior year. This eliminated
tivity and by acting as another
charge carrier. However, the high
mobility also decreases stability.
The authors controlled ion migra-
tion by adding a small amount of
silver hexafluoroantimonate, thus
improving stability while main-
taining the high thermoelectric
efficiency. —BG
Nat. Mater. (2024)
10.1038/s41563-024-01815-1

1430 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


◥ RESULTS: We show that JIP3’s short N-terminal
RESEARCH ARTICLE SUMMARY
STRUCTURAL BIOLOGY
Molecular mechanism of dynein-dynactin
complex assembly by LIS1
Kashish Singh†, Clinton K. Lau†, Giulia Manigrasso, José B. Gama, Reto Gassmann, Andrew P. Carter*
INTRODUCTION: Cytoplasmic dynein-1 (dynein) are linked to developmental and neurological
is a microtubule motor required for fundamen- disorders. LIS1 unlocks dynein’s autoinhibited
tal processes such as intracellular trafficking, “phi” state by binding to the motor domain.
organelle positioning, and the function of the However, it remains unclear whether this alone
mitotic spindle. Dynein activation requires the mediates its role in complex assembly.
assembly of a ~4-MDa complex with its co-
factor dynactin and a cargo-specific adaptor RATIONALE: In this study, we set out to un-
protein. In cells, a diverse set of these acti- derstand whether JIP3 activates dynein using
vating adaptors link dynein to its many cargos. in vitro motility assays. We then investigated
They typically contain long coiled coils and the molecular basis of JIP3-dependent activa-
specific motifs for binding dynein-dynactin. tion by visualizing these complexes, bound
However, some, such as the lysosomal adaptor to microtubules in the presence of the non-
JIP3 (c-Jun N-terminal kinase-interacting pro- hydrolyzable adenosine triphosphate analog
tein 3), contain notably shorter coiled coils adenylyl-imidodiphosphate (AMP-PNP), using
and appear to lack some of the required mo- cryo–electron microscopy. The inclusion of LIS1
tifs, raising the question of how they work. trapped the dynein-dynactin-JIP3 complex in its
Additionally, dynein-dynactin complex assem- assembly pathway, providing an explanation for
bly requires regulatory factors, such as LIS1 how it and dynactin’s extended p150Glued sub-
(lissencephaly-1) protein, mutations in which unit (p150) stimulate the process.
Dynein
LIS1 Dynein-dynactin-JIP3-LIS1 complex
(open)
on microtubules
Dynactin
Dynactin
p150
Dynein intermediate chain “Phi” Dynein-A
dynein Dynein-B
relieves dynactin autoinhibition
JIP3
JIP3
Flexible and labile p150
complex
DIC-N LIS1
LIS1 primes
dynein-dynactin
for adaptor binding
LIS1 releases
from first dynein
Robust complex Second
assembly dynein
Both motors
are active
Second dynein
binds to
microtubules
Microtubule
Model for LIS1-mediated dynein-dynactin-adaptor complex assembly. Dynein intermediate chain binds
to dynactin’s p150 subunit, relieving dynactin autoinhibition. This allows LIS1 (attached to dynein) to
bind p150, which stimulates complex assembly by priming dynein-dynactin for adaptor binding. LIS1 keeps
dynein-A detached from microtubules, whereas dynein-B, recruited to the complex, binds to the microtubule.
Finally, LIS1 disengages from dynein-A upon movement initiation.
SCIENCE science.org
RESEAR CH
coiled coil is sufficient to activate dynein-dynactin.
Although not absolutely required for activation,
JIP3 contains a Spindly motif. This dynactin-
interacting site was previously missed owing to
being ~200 amino acids away from the N-terminal
coiled coil. We find that JIP3, like many activat-
ing adaptors, contains an autoinhibitory mech-
anism. In JIP3’s case, this mechanism involves a
short internal helix that mimics the helices from
dynein’s light intermediate chain, which inter-
act with all characterized adaptors.
The structure of the dynein-dynactin-JIP3-
LIS1 complex contains two dyneins (A and B).
Dynein-B is bound to microtubules as seen
in a previous structure. In contrast, dynein-A
is bound to two LIS1 molecules and detached
from the microtubule. The presence of LIS1
leads to docking of dynactin’s p150 arm along
the length of the dynein. It also resolves p150’s
interaction with a critical helix at the N ter-
minus of dynein’s intermediate chain (DIC-N),
known to be needed for dynein activation.
Our structure and associated experiments pro-
vide insights into the dynein-dynactin-adaptor
(DDA) assembly pathway. We find that p150 is in
an open conformation. Comparison with a pre-
vious folded state suggests that DIC-N binds and
opens up the p150 first, allowing subsequent in-
teractions with LIS1 and dynein. We show that
LIS1 needs to be able to link p150 and the dynein-A
motor to stimulate DDA complex formation. We
propose that LIS1 works by tucking dynein un-
der the p150 arm, priming it for efficient adaptor
binding. The next step is dynein-B recruitment.
We propose that LIS1 is preferentially released
from this motor, as it lacks reinforcement from
p150. Therefore, dynein-B is the first motor to con-
tact the microtubule, leading to the intermediate
state that we have resolved. The likely next step
is the release of LIS1 from dynein-A, allowing
both motors to drive processive movement.
CONCLUSION: We show how the protein JIP3 ac-
tivates dynein, despite its unusual architecture,
providing insights into the minimal requirements
of dynein-dynactin cargo adaptors. We reveal in-
teractions between dynein, dynactin, and LIS1
that occur during the formation of active dynein
complexes. On the basis of our observations, we
propose a model where LIS1 cooperates with
p150 to orient dynein and dynactin, priming them
for productive adaptor binding. Our study un-
covers the intricate set of interactions occurring
during the formation of active dynein complexes
and how LIS1 helps stimulate them.
▪
The list of author affiliations is available in the full article online.
*Corresponding author. Email: cartera@mrc-lmb.cam.ac.uk
†These authors contributed equally to this work.
Cite this article as K. Singh et al., Science 383, eadk8544
(2024). DOI: 10.1126/science.adk8544
READ THE FULL ARTICLE AT
https://doi.org/10.1126/science.adk8544
29 MARCH 2024 • VOL 383 ISSUE 6690 1431
RES EARCH

◥ RESULTS: First, we studied convergent evolution

RESEARCH ARTICLE SUMMARY in protein-coding regions using the RERconverge
and HyPhy methods to find 200 significantly
ZOONOMIA associated genes. The genes that tend to be
under higher constraint in vocal learning mam-
Vocal learning–associated convergent evolution in mals are enriched for genes involved in human
autism. However, the vast majority of genes are
mammalian proteins and regulatory elements driven by signals from only one or two clades of
vocal learning mammals, suggesting that a large
Morgan E. Wirthlin†, Tobias A. Schmid†, Julie E. Elie†, Xiaomeng Zhang, Amanda Kowalczyk, component of the genetic basis for the trait may
Ruby Redlich, Varvara A. Shvareva, Ashley Rakuljic, Maria B. Ji, Ninad S. Bhat, Irene M. Kaplow, lie instead in the convergent evolution of regu-
Daniel E. Schäffer, Alyssa J. Lawler, Andrew Z. Wang, BaDoi N. Phan, Siddharth Annaldasula, latory elements. To explore that hypothesis, we
Ashley R. Brown, Tianyu Lu, Byung Kook Lim, Eiman Azim, Zoonomia Consortium, Nathan L. Clark, performed an anatomical and functional charac-
Wynn K. Meyer, Sergei L Kosakovsky Pond, Maria Chikina, terization of the Egyptian fruit bat motor cortex.
Michael M. Yartsev*‡, Andreas R. Pfenning*‡ We identified a subregion of the motor cortex
that is implicated in vocal production and direct-
ly projects to the motoneurons controlling the
INTRODUCTION: Vocal production learning (“vocal RATIONALE: Despite evidence for the conver- bat’s larynx. This allowed us to profile candidate
learning”), or the ability to modify vocalizations gent evolution of vocal learning at the behav- regulatory elements active in this vocalization-
according to the social environment, forms the ioral, anatomical, and gene expression levels associated subregion of the motor cortex by mea-
basis of human speech production. Among the in vertebrates, the genetic underpinnings suring open chromatin. These open chromatin
Boreoeutherian mammals, this trait has evolved of vocal learning and human speech in mam- regions and 222 mammalian genomes of the
independently in four different lineages: hu- mals are poorly understood. New machine Zoonomia Consortium served as input to the
mans, bats, cetaceans, and pinnipeds. In verte- learning approaches and the newly sequenced Tissue-Aware Conservation Inference Toolkit
brates, the evolution of vocal learning behavior mammalian genomes of the Zoonomia (TACIT) machine learning approach, which was
has been associated with the evolution of brain Consortium provide the foundation to rigo- applied to find 50 candidate regulatory elements
anatomical features, including cortical long- rously study this question. The repeated evo- whose predicted motor cortex open chromatin
range projection neurons (e.g., songbirds and lution of vocal learning across mammals measurements across mammals are highly cor-
humans). Moreover, neural circuits for the allows us to determine which parts of the related with the presence of vocal learning behav-
production of learned vocalization display con- genome are significantly associated with ior. Many of these open chromatin regions were
vergent evolution in patterns of gene expression. the behavior. near genes associated with autism, and they tended
to overlap with open chromatin specific to the
long-range projection neurons that have been
implicated in the evolution of vocal learning.

CONCLUSION: Although it is impossible to know

which parts of the genome evolved for human
speech production, we are able to use the repeated
evolution of a component of that behavior, vocal
learning, to find significantly associated genes
and noncoding regions. Our results demonstrate
that the presence of vocal learning behavior in a
given clade leads to weak selective pressure across
a broad range of genes and stronger selective
pressure across a smaller number of motor cortex
noncoding regions. These genes and noncoding
regions show an association with autism, which
suggests that there are shared regulatory net-
works for vocal and social behavior that tend to
adapt in similar ways when a lineage evolves
vocal learning behavior. More broadly, our re-
sults suggest that the evolutionary history of se-
lective pressures across a location in the genome
can provide insight into how that region might
influence human behavior.
(RERconverge and HyPhy) ▪
*Corresponding author. Email: myartsev@berkeley.edu
(M.M.Y.); apfenning@cmu.edu (A.R.P.)
†These authors contributed equally to this work.
‡These authors contributed equally to this work.
Finding vocal learning–associated regions of the mammalian genome. We compared the evolution of
Cite this article as M. Wirthlin et al., Science 383,
vocal learning behavior to the evolution of coding and noncoding elements of the genome, leveraging eabn3263 (2024). DOI: 10.1126/science.abn3263
anatomical, electrophysiological, and epigenomic experiments in the Egyptian fruit bat orofacial motor cortex
(ofM1). We show convergent evidence of the importance of long-range projection neurons and autism- READ THE FULL ARTICLE AT
associated gene networks. https://doi.org/10.1126/science.abn3263

1432 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESEAR CH

◥ This substantial indirect effect of BCG vaccina-

RESEARCH ARTICLE SUMMARY tion exceeded the observed direct protection
against infection (58%; 95% CrI: 34 to 73%),
VACCINATION and the combined effects translated to a total
vaccine efficacy of 89% (95% CrI: 74 to 96%).
BCG vaccination reduces bovine tuberculosis Vaccinated animals exhibited substantially
lower total visible lesion scores compared with
transmission, improving prospects for elimination unvaccinated controls, which is consistent with
the notion that BCG vaccination reduces dis-
Abebe Fromsa†, Katriina Willgert†, Sreenidhi Srinivasan†, Getnet Mekonnen, Wegene Bedada, ease severity and potentially infectiousness.
Balako Gumi, Matios Lakew, Biniam Tadesse, Berecha Bayissa, Asegedech Sirak, Musse Girma Abdela, A stochastic metapopulation transmission
Solomon Gebre, Tesfaye Chibssa, Maroudam Veerasami, H. Martin Vordermeier, Douwe Bakker, model, calibrated with data from Ethiopia,
Stefan Berg, Gobena Ameni, Nick Juleff, Mart C. M. de Jong, James Wood, suggests that routine calfhood BCG vaccination
Andrew Conlan*, Vivek Kapur* has the potential to prevent the predicted ex-
pansion of bTB in dairy herds and bring the
population average reproduction ratio below
INTRODUCTION: Bovine tuberculosis (bTB) poses is critical for the evaluation of BCG in cattle 1 within as few as 10 years, resulting in a sub-
a substantial global threat to animal health, because the primary effect is to reduce the stantial decrease in predicted bTB prevalence
food security, and human well-being. Although extent and rate of progression of lesions rather as compared with baseline scenarios without
proven effective in many high-income coun- than to provide sterilizing protection. Our study vaccination. The results highlight the critical
tries, the traditional test-and-slaughter approach addresses this gap by performing a natural importance of the combined direct and indirect
for bTB control is expensive and impractical transmission experiment with bTB in cattle, effects of BCG vaccination in enabling bTB
for socioeconomic reasons in many regions using a crossover design approach. This method elimination.
where the disease remains endemic. This has enabled a more realistic and robust assessment The findings suggest that BCG vaccination
necessitated a need for alternative bTB con- of BCG’s true impact on bTB transmission, represents an important tool for bTB control,
trol strategies, with Bacille Calmette-Guérin quantifying both the direct efficacy of BCG as particularly in resource-limited settings where
(BCG) vaccination presenting a promising op- well as its effect on reducing transmission. We traditional methods are impractical. The results
tion. However, the effectiveness of BCG in con- developed a mechanistic transmission model also suggest that achieving elimination through
trolling bTB by reducing onward transmission to explore the potential of using BCG vaccina- vaccination alone would require a long-term
remains unclear. This study investigated both tion in Ethiopia, where the transmission risk commitment, as the full benefits may take de-
the direct and previously unexplored indirect of bTB varies considerably between herds and cades to be realized. Our studies highlight a need
effects of BCG vaccination on bTB transmis- the relatively infrequent trading of animals is for further research on the duration of efficacy,
sion in cattle, thereby providing key missing projected to contribute to a gradual yet sub- including the potential for extending protec-
insights for control. stantial increase in prevalence. tion through revaccination, as well as the impact
on cross-species transmission.
RATIONALE: Traditional vaccine efficacy evalua- RESULTS: The natural transmission study showed
tions cannot measure the impact of vaccination a 74% reduction in bTB transmission [95% CONCLUSION: Our study demonstrates remark-
on reducing onward transmission from infected credible interval (CrI): 46 to 89%] in vacci- able and previously unrecognized indirect ef-
individuals. This mode of action of vaccination nated as compared with unvaccinated animals. fects of BCG vaccination on bTB transmission,
extending beyond its direct protective effect.
Seeder – Sentinel Scenario analyses with mechanistic models
Unvaccinated – Unvaccinated Vaccinated – Unvaccinated
for transmission in Ethiopia suggest that
Unvaccinated – Vaccinated Vaccinated – Vaccinated
implementation of BCG vaccination may en-
Naturally infected Naturally infected able effective bTB control and progress toward
seeders seeders 3 elimination. Moreover, these findings sug-
gest that BCG may provide an effective method
Transmission threshold
R=1 of control in resource-limited settings where
current test-and-slaughter approaches are
Posterior density

2 unfeasible. Lastly, the crossover trial design

incorporating natural transmission provides
a general framework for studying other vac-
1
cines and interventions aimed at reducing
Unvaccinated seeders Vaccinated seeders onward transmission of TB, with broad ap-
plicability to other infectious diseases of ani-
mals, including humans.
0
▪
0 2 4 6 The list of author affiliations is available in the full article online.
Unvaccinated Vaccinated Experimental R *Corresponding author. Email: vxk1@psu.edu (V.K.);
ajkc2@cam.ac.uk (A.J.K.C.)
†These authors contributed equally to this work.
Quantifying BCG vaccination’s total efficacy against bovine tuberculosis. Sentinel calves, both
Cite this article as: A. Fromsa et al., Science 383, eadl3962
BCG-vaccinated and unvaccinated, were exposed to seeder cattle to measure direct efficacy through (2024). DOI: 10.1126/science.adl3962
IGRA-conversion times over 12 months. Subsequently, these sentinels were used to determine BCG’s indirect
transmission-reducing effects, and the results helped inform development of models for evaluating TB READ THE FULL ARTICLE AT
elimination strategies. https://doi.org/10.1126/science.adl3962

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1433

RES EARCH

◥ Discovery of the binding site for ubiquitin

RESEARCH ARTICLES Val70 as a new drug binding site on PLpro
The noncovalent PLpro inhibitor XR8-24 has
CORONAVIRUS an IC50 of 0.56 mM in a fluorescence resonance
energy transfer (FRET) enzymatic assay and
Design of a SARS-CoV-2 papain-like protease an EC50 of 1.2 mM in an antiviral assay (24).
Compound 7 (Cp7) is a rationally designed
inhibitor with antiviral efficacy in a mouse model covalent PLpro inhibitor with a fumarate re-
active warhead that inhibits PLpro with an IC50
Bin Tan1†, Xiaoming Zhang2†, Ahmadullah Ansari3,4†, Prakash Jadhav1†, Haozhou Tan1, Kan Li1, of 0.094 mM and SARS-CoV-2 replication with
Ashima Chopra3,4‡, Alexandra Ford5, Xiang Chi2, Francesc Xavier Ruiz3,4*, Eddy Arnold3,4*, an EC50 of 1.1 mM (25). Inspired by these re-
Xufang Deng2,6*, Jun Wang1* sults, we subsequently designed a hybrid cova-
lent PLpro inhibitor Jun11313 by converting
The emergence of SARS-CoV-2 variants and drug-resistant mutants calls for additional oral antivirals. the naphthalene in Cp7 to 3-phenylthiophene
The SARS-CoV-2 papain-like protease (PLpro) is a promising but challenging drug target. We designed (Fig. 1A); Jun11313 was tested alone and was
and synthesized 85 noncovalent PLpro inhibitors that bind to a recently discovered ubiquitin binding found to potently inhibit PLpro with an IC50 of
site and the known BL2 groove pocket near the S4 subsite. Leads inhibited PLpro with the inhibitory 0.12 mM, and it was chosen as the initial hit for
constant Ki values from 13.2 to 88.2 nanomolar. The co-crystal structures of PLpro with eight leads the following inhibitor design.
revealed their interaction modes. The in vivo lead Jun12682 inhibited SARS-CoV-2 and its variants, including The lead covalent compound Jun11313 was
nirmatrelvir-resistant strains with EC50 from 0.44 to 2.02 micromolar. Oral treatment with Jun12682 co-crystallized with SARS-CoV-2 PLpro to vis-
improved survival and reduced lung viral loads and lesions in a SARS-CoV-2 infection mouse model, ualize the key interactions for binding. We
suggesting that PLpro inhibitors are promising oral SARS-CoV-2 antiviral candidates. determined the co-crystal structure at 2.85-Å
resolution (see Methods). As expected, the fu-
marate ester electrophile reacts to form a co-
he COVID-19 pandemic has created an sivir or nirmatrelvir have been identified from valent bond with the catalytic Cys111 (Fig. 1B

T urgent need for orally bioavailable anti-

virals. The FDA has approved three direct-
acting antivirals, including two viral
RNA-dependent RNA polymerase inhib-
itors, remdesivir and molnupiravir, and one
viral main protease (Mpro) inhibitor, nirma-
trelvir (1). Remdesivir is administered by in-
travenous injection and its use is limited to
hospitalized patients. The clinical efficacy of
remdesivir is also controversial (2). Molnu-
piravir is a mutagen and should not be used
by pregnant women (3). Nirmatrelvir is co-
administered with ritonavir, a CYP3A4 inhib-
itor, to improve the in vivo half-life (4). For this
reason, Paxlovid, a combination of nirmatrel-
vir and ritonavir, has drug-drug interaction
concerns. Ensiltrelvir is a second-generation
Mpro inhibitor approved in Japan (5). Although
it is highly efficacious in clinical trials (6), en-
sitrelvir is a potent CYP3A inhibitor that may
lead to severe adverse drug-drug interactions
with other medications (7, 8). Mutant SARS-
CoV-2 viruses with resistance against remde-
1
Department of Medicinal Chemistry, Ernest Mario School of
Pharmacy, Rutgers, the State University of New Jersey,
Piscataway, NJ 08854, USA. 2Department Physiological
Sciences, College of Veterinary Medicine, Oklahoma State
University, Stillwater, OK 74078, USA. 3Center for Advanced
Biotechnology and Medicine, Rutgers, the State University of
New Jersey, Piscataway, NJ 08854, USA. 4Department of
Chemistry and Chemical Biology, Rutgers, the State University
of New Jersey, Piscataway, NJ 08854, USA. 5Department of
Veterinary Pathobiology, College of Veterinary Medicine,
Oklahoma State University, Stillwater, OK 74078, USA.
6 binding site that accommodates the ubiqui-
Oklahoma Center for Respiratory and Infectious Diseases,
Oklahoma State University, Stillwater, OK 74078, USA.
*Corresponding author. Email: junwang@pharmacy.rutgers.edu (J.W.);
xufang.deng@okstate.edu (X.D.); arnold@cabm.rutgers.edu (E.A.);
xavier@cabm.rutgers.edu (F.X.R.)
†These authors contributed equally to this work.
‡Present address: Vanderbilt University
viral passage experiments in cell culture (9–11)
and drug-treated COVID-19 patients (12–15).
Therefore, additional antivirals with alterna-
tive mechanisms of action are urgently needed
to combat drug-resistant and emerging SARS-
CoV-2 variants.
The papain-like protease (PLpro) is one of
two viral cysteine proteases encoded by SARS-
CoV-2. PLpro cleaves the viral nonstructural
polyproteins (nsps) at the nsp1/2, nsp2/3, and
nsp3/4 junctions and is critical for viral repli-
cation (16). In addition, PLpro suppresses the
host immune response through cleaving ubiq-
uitin and interferon-stimulated gene 15 (ISG15)
modifications from host proteins (17–19). The
SARS-CoV-2 PLpro sequence is 82.9% identical
to SARS-CoV-1 PLpro. PLpro is highly conserved
among SARS-CoV-2 variants (20), rendering
it a high-profile antiviral drug target (19, 21).
Before our study, noncovalent and covalent
PLpro inhibitors were reported (20, 22). How-
ever, despite decades of medicinal chemistry
optimization and high-throughput screen-
ing, PLpro inhibitor development is still at the
preclinical stage and lags far behind Mpro
inhibitors (20). PLpro substrates contain a
consensus motif LXGG↓(N/K/X) (16). The S1
and S2 subsites of PLpro form a narrow tunnel
for binding two glycines (23). The absence of
S1 and S2 binding pockets presents a chal-
lenge in designing highly potent PLpro inhib-
itors. In this study, we describe the design of
potent PLpro inhibitors by exploiting a drug-
tin Val70 side chain (Val70Ub). We validate
PLpro as a drug target by demonstrating in vivo
antiviral efficacy of a designed PLpro inhibitor
Jun12682 with oral administration in SARS-
CoV-2–infected mice.
and fig. S1B).
Previous crystal structures of PLpro with in-
hibitors revealed a conformationally flexible
blocking loop 2 (BL2), remote from the active
site cysteine, which is stabilized by inhibitor or
substrate binding (20). Additionally, the co-
crystal structure with XR8-24 revealed a
neighboring site interacting with the phenyl-
thiophene group of the inhibitor, the BL2
groove (Fig. 1C) (24). Notably, the thienyl
group in Jun11313 is oriented toward the op-
posite side of the BL2 groove compared with
the pyrrolidine-substituted thienyl group in
XR8-24 (Fig. 1C). The phenylthiophene in
XR8-24 makes extensive contacts with the
BL2 groove (Fig. 1C). In comparison, the phenyl-
thienyl group of Jun11313 is oriented toward
Pro247, making van der Waals contacts with
both Pro248 and Pro247, and CH–p and S–p
interactions with Met208 (Fig. 1B). Although
the strength of these interactions is difficult
to estimate and has been understudied in
drug design (26, 27), here they may be the
culprit for Jun11313’s unexpected conforma-
tion (Fig. 1B). If the methyl pyrrolidine group
in XR8-24 were arranged as the phenylthienyl
group of Jun11313, it should orient in a very
solvent-exposed region to avoid a clash with
Met208 (fig. S1B).
Superimposition of PLpro structures com-
plexed with ubiquitin and Jun11313 revealed
that the thienyl group occupies the same hy-
drophobic site as Val70 from ubiquitin (Fig.
1D). Therefore, we designate this pocket as the
Val70Ub site. The comparison of the co-crystal
structures of PLpro with Jun11313 and with
ubiquitin suggests that this site is essential
for binding both the ubiquitin substrate and
this inhibitor. Similarly, the superimposition of

1434 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 1. X-ray crystal structure of the

covalent inhibitor Jun11313 with
SARS-CoV-2 PLpro and structure-based
design of biarylphenyl SARS-CoV-2
PLpro inhibitors. (A) Design of the hybrid
covalent inhibitor Jun11313 based on
XR8-24 and Cp7. (B) Atomic model of the
Jun11313 (green balls and sticks) binding
site in PLpro (light gray sticks represent
residues within a 5 Å distance of the
inhibitor), with hydrogen bonds indicated
with black dashed lines. The Jun11313
polder map (an unbiased omit map that
removes the bulk solvent contribution
near the inhibitor) (38) is displayed as a
gray mesh with 4s contours. (C) Super-
position of the PLpro-Jun11313 structure
onto the structure of the PLpro-XR8-24
complex (PDB 7LBS), with XR8-24 repre-
sented by yellow sticks and spheres, with
the relevant residues for binding of both
compounds indicated. (D) Superimposed
x-ray crystal structures of SARS-CoV-2
PLpro with Jun11313 (green) (PDB: 8UVM),
XR8-24 (yellow) (PDB: 7LBS), and
ubiquitin (orange) (PDB: 6XAA). (E) Generic
chemical structure of the designed
biarylphenyl PLpro inhibitors. Critical inter-
actions are highlighted. (F) Flow chart for
the lead optimization of PLpro inhibitors.
Jun12682 was selected as the in vivo
lead candidate.

PLpro structures complexed with ISG15 showed
that the thienyl group interacts with the analo-
gous Asn151-Leu152 site from ISG15 (fig. S1C).
Rational design of biarylphenyl PLpro inhibitors
The unexpected binding pose of Jun11313 in
PLpro led us to hypothesize that potent PLpro
inhibitors can be designed by simultaneously
engaging the BL2 groove pocket and the Val70Ub
hydrophobic site (Fig. 1E), which has not been
previously explored for PLpro inhibitor design
(20). We designed and synthesized a library of
85 biarylphenyl benzamide compounds (fig.
S2). Aryl substitutions were installed on the
3 and 5 positions of the phenylethylamine
ring to engage hydrophobic interactions with
residues in the BL2 groove and the ubiquitin
Val70Ub binding site (Fig. 1E). In addition,
diverse amines were installed on the meta-
position of the benzoic acid to engage elec-
trostatic interactions with Glu167 (24). The
central 1-phenylethyl benzamide core struc-
ture was kept intact to maintain the critical
hydrogen bonds and p-p interactions with
the BL2 loop. All synthesized compounds were
initially tested in the FRET-based enzymatic
assay and the cytotoxicity assay in Vero E6
cells (Fig. 1F). Promising lead compounds were
then tested in a secondary FlipGFP PLpro as-
say and SARS-CoV-2 antiviral assay. FlipGFP
PLpro is a cell-based assay that validates intra-
cellular PLpro target engagement (23). Next,
leads were profiled for in vitro microsomal
stability and in vivo oral pharmacokinetic (PK)
properties in mice.
A complete list of the designed biarylphenyl
PLpro inhibitors is shown in fig. S2, with rep-
resentative examples in Fig. 2. Among the 85
compounds tested in the FRET assay, 26 had
IC50 < 100 nM, 42 had IC50 between 100 and
200 nM, and 14 had IC50 between 200 and
400 nM. The control compound GRL0617 had
IC50 values of 1.92 mM. GRL0617 was originally
developed for SARS-CoV PLpro and was later
repurposed for SARS-CoV-2 PLpro (28). The
inhibitory constant Ki was determined for po-
tent compounds (Fig. 2). Jun11875 inhibited
PLpro with a Ki of 13.2 nM, a 104-fold improve-
ment over GRL0617 (Ki = 1374 nM). However,
thiophene-containing compounds were gen-
erally cytotoxic in Vero E6 cells (CC50 < 20 mM)
(Fig. 2B and fig. S2), possibly due to the forma-
tion of reactive metabolites (29). Next, we
examined several symmetric and asymmetric
five- and six-membered aromatic substitutions
at the 3- and 5-positions, respectively, to miti-
gate cellular cytotoxicity while maintaining
enzymatic inhibition (Fig. 2, C to F). Among
the list of heterocycle substitutions examined,
pyrazole was shown to confer potent PLpro in-
hibition and reduced cellular cytotoxicity. It
is noted that despite the potent PLpro enzy-
matic inhibition for Jun1247 (IC50 = 165.3 nM),
Jun121210 (IC50 = 81.9 nM), Jun121911 (IC50 =
73.2 nM), Jun12208 (IC50 = 91.6 nM), and
Jun12242 (IC50 = 90.5 nM), their cellular ac-
tivities were moderate to weak as shown by
the FlipGFP assay (EC50 > 10 mM) (fig. S2),
which might be due to poor membrane per-
meability. Jun12682 inhibited PLpro with a
Ki of 37.7 nM and displayed an EC50 of 1.1 mM
in the FlipGFP PLpro assay, more than 20-fold
improved from GRL0617 (EC50 = 22.4 mM).
Prioritized lead compounds were tested in
the SARS-CoV-2 antiviral assay and had EC50
from 0.23 to 1.15 mM (Fig. 2). Jun12682 had
EC50 values of 0.42 and 0.51 mM in the icSARS-
CoV-2-nLuc reporter virus and plaque assays,
respectively (fig. S3). To examine the potential
of PLpro inhibitors in inhibiting SARS-CoV-2
variants and drug-resistant mutants, we selected

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1435

RES EARCH | R E S E A R C H A R T I C L E S

A B C F
N F N
H2 N N N
NH2 S
O O
N
H S N
N H H
N O N
O N
S O O
N

GRL0617 Jun11875 Jun12129 Jun12303 Jun12162

FRET IC50 = 1918.0 150.0 nM IC50 = 66.2 2.0 nM IC50 = 90.9 4.6 nM IC50 = 121.5 4.0 nM IC50 = 98.3 7.0 nM
Ki = 1374 58 nM Ki = 13.2 2.0 nM Ki = 75.5 2.0 nM Ki = 88.2 6.0 nM Ki = 33.6 3.0 nM
CC50 = 107.3 10.9 µM (Vero) CC50 = 4.5 1.6 µM CC50 = 8.3 1.2 µM CC50 = 56.8 7.4 µM CC50 = 42.3 2.3 µM
FlipGFP EC50 = 22.4 2.3 µM PDB: 8UUY FlipGFP EC50 = 2.6 0.2 µM FlipGFP EC50 = 2.1 0.2 µM
SARS-CoV-2 Antiviral Antiviral EC50 = 0.32 µM Antiviral EC50 = 0.23 µM
EC50 > 20 µM T1/2 = 136 min T1/2 = 76.7 min
PDB: 8UUG PDB: 8UUU

D N N
N
O

N
CF3 N
O N N N N
N N N N S
H
N
N
N O

Jun11941 Jun12199 Jun12197 Jun12603 Jun12351 Jun12145

IC50 = 151.4 4.0 nM IC50 = 108.8 10.2 nM IC50 = 102.7 10.2 nM IC50 = 112.2 6.0 nM IC50 = 98.8 5.2 nM IC50 = 108.5 6.2 nM
Ki = 34.3 3.0 nM Ki = 47.6 3.0 nM Ki = 33.2 3.0 nM Ki = 39.8 4.0 nM Ki = 29.3 3.0 nM Ki = 35.2 2.0 nM
CC50 = 54.8 8.3 µM CC50 = 63.2 11.3 µM CC50 = 43.7 5.1 µM CC50 = 61.0 19.3 µM CC50 > 125 µM CC50 = 31.3 5.4 µM
FlipGFP EC50 = 1.8 0.2 µM FlipGFP EC50 = 0.8 0.1 µM FlipGFP EC50 = 0.6 0.1 µM FlipGFP EC50 = 2.4 0.4 µM FlipGFP EC50 = 2.0 0.4 µM FlipGFP EC50 = 1.3 0.3 µM
Antiviral EC50 = 0.31 µM Antiviral EC50 = 0.45 µM Antiviral EC50 = 0.35 µM Antiviral EC50 = 0.50 µM Antiviral EC50 = 0.59 µM Antiviral EC50 = 0.58 µM
T1/2 > 145 min T1/2 = 79.7 min T1/2 = 116.0 min T1/2 > 145 min T1/2 = 26.3 min
PDB: 8UUF PDB: 8UUH PDB: 8UUV PDB: 8UUW

E F
N N
N N
N N
O O F O N
N N
O N N
H H N
N
N N
N N
N O N O

Jun12682 Jun12763 Jun12395 Jun12713 Jun12602

IC50 = 106.8 5.0 nM IC50 = 164.6 23.0 nM IC50 = 121.9 8.0 nM IC50 = 86.7 6.0 nM IC50 = 116.7 9.2 nM
Ki = 37.7 3.0 nM Ki = 37.2 4.0 nM Ki = 47.1 4.0 nM Ki = 34.0 3.0 nM Ki = 46.9 3.0 nM
CC50 = 61.3 4.5 µM CC50 > 125 µM CC50 = 109.4 16.2 µM CC50 = 107.2 6.2 µM CC50 > 125 µM
FlipGFP EC50 = 1.1 0.1 µM FlipGFP EC50 = 2.0 0.3 µM FlipGFP EC50 = 2.9 0.3 µM FlipGFP EC50 = 2.7 0.6 µM FlipGFP EC50 = 14.7 2.5 µM
Antiviral EC50 = 0.42 µM Antiviral EC50 = 0.96 µM Antiviral EC50 = 1.15 µM Antiviral EC50 = 0.86 µM Antiviral EC50 = 0.80 µM
T1/2 = 82.4 min T1/2 = 88.3 min T1/2 > 145 min
PDB: 8UOB

Fig. 2. Representative biarylbenzamide series of SARS-CoV-2 PLpro
inhibitors. (A) Positive control GRL0617. (B) Thiophene-containing PLpro
inhibitors. (C to F) Pyrazole-containing PLpro inhibitors. The FRET IC50
and Ki values were determined in the enzymatic assay using the substrate
Dabcyl-FTLRGG/APTKV-E(Edans). Cytotoxicity CC50 was determined in Vero
E6 cells with 48-hour incubation. FlipGFP PLpro EC50 values were determined
in 293T cells transfected with PLpro and FlipGFP-expressing plasmids.
SARS-CoV-2 antiviral EC50 values were determined using the icSARS-CoV-2-
nLuc reporter virus in a Caco-2 line expressing hACE2 and hTMPRSS2
(Caco2-AT) with 2-mM p-gp inhibitor CP-100356. T1/2 refers to the half-life
of PLpro inhibitors in mouse liver microsomes. Data are presented as mean ±
SD of three technical replicates.

two PLpro inhibitors, Jun11941 and Jun12682,
to test against SARS-CoV-2 delta and omicron
variants and three recombinant SARS-CoV-2
viruses that are resistant to nirmatrelvir, rNsp5-
S144M, rNsp5-L50F/E166V, and rNsp5-L50F/
E166A/L167F (Fig. 3, A to C). S144M and
E166V are nirmatrelvir-resistant mutations
identified from enzymatic assays and viral pas-
sage experiments (9, 10). Jun11941 and Jun12682
showed consistent antiviral activity against
these viruses with EC50 fold increases less than
1.5 and 2.0, respectively, compared with the
wild type (WT). In comparison, the rNsp5-
S144M, rNsp5-L50F/E166V, and rNsp5-L50F/
E166A/L167F showed resistance to nirmatrelvir
with EC50 fold-increases of 12.5, 24.2, and 21.7,
respectively, compared with the WT.
Mechanism of action of Jun12682 in inhibiting
SARS-CoV-2 PLpro
PLpro is known to antagonize the host innate
immune response upon viral infection by hy-
drolyzing the isopeptide bond between ubiq-
uitin and ISG15 to lysine side chains of host
proteins (17, 18). To characterize whether
Jun12682 inhibits the deubiquitinating and
deISGylating activities of SARS-CoV-2 PLpro,
we performed the PLpro enzymatic assay using
Ub-AMC and ISG15-AMC substrates (23).
Jun12682 caused enzymatic inhibition with Ki
values of 63.5 and 38.5 nM in the ubiquitin-
AMC (7-amino-4-methylcoumarin) and ISG15-
AMC FRET assays (Fig. 3, D to G). The results
are consistent with inhibition of PLpro-mediated
cleavage of Ub-AMC and ISG15-AMC by GRL0617
and XR8-24 (23, 24). To profile the off-target
effect, Jun12682 was tested against the two
closest human structural homologs of PLpro,
USP7 and USP14 (24, 30). Jun12682 did not
inhibit USP7- and USP14-catalyzed Ub-AMC
hydrolysis at up to 40 mM (Fig. 3, H and I).
Jun12682 also showed dose-dependent sta-
bilization of SARS-CoV-2 PLpro and was more
potent than GRL0617 in the differential scan-
ning fluorimetry assay (Fig. 3J).
X-ray crystal structures of SARS-CoV-2 PLpro
with inhibitors
The x-ray co-crystal structures of SARS-CoV-2
PLpro were solved for eight biarylphenyl PLpro in-
hibitors, Jun11941, Jun12129, Jun12303, Jun12162,
Jun12199, Jun12197, Jun12145, and Jun12682

1436 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 3. Antiviral activity of PLpro inhibitors against SARS-CoV-2 variants
and mechanistic studies of Jun12682. Antiviral activity of nirmatrelvir
(A), Jun11941 (B), and Jun12682 (C) against SARS-CoV-2 WA1 strain (WT)
(USA-WA1/2020), Omicron strain (lineage B.1.1.529, BA.2; HCoV-19/USA/CO-
CDPHE-2102544747/2021), Delta strain (Lineage B.1.617.2; hCoV-19/USA/
MD-HP05647/2021), and nirmatrelvir-resistant strains rNsp5-S144M, rNsp5-
L50F/E166V, and rNsp5-L50F/E166A/L167F was determined using a cell viability
assay with Vero-AT cells. Data in (A), (B), and (C) are representative results
of at least two independent experiments and presented as mean ± SD of
three technical replicates. (D) IC50 curve of Jun12682 in inhibiting SARS-CoV-2
PLpro hydrolysis of Ub-AMC. (E) Ki curve of Jun12682 in inhibiting SARS-
CoV-2 PLpro hydrolysis of Ub-AMC. (F) IC50 curve of Jun12682 in inhibiting
SARS-CoV-2 PLpro hydrolysis of ISG15-AMC. (G) Ki curve of Jun12682 in
inhibiting SARS-CoV-2 PLpro hydrolysis of ISG15-AMC. Data in (D), (E), (F),
and (G) are presented as mean ± SD of three technical replicates. (H) Counter
screening of Jun12682 in inhibiting USP7 hydrolysis of Ub-AMC. (I) Counter
screening of Jun12682 in inhibiting USP14 hydrolysis of Ub-AMC. (J) Differential
scanning fluorimetry assay of Jun12682 and GRL0617 in stabilizing the
SARS-CoV-2 PLpro. Data in (H), (I), and (J) are presented as mean ± SD of
two technical replicates.

(resolution range of 2.5 to 3.1 Å; Fig. 4 and
table S1). The phenylthienyl group of Jun11313
binds to the Val70Ub site of SARS-CoV-2 PLpro
in the region where residues Val70Ub and Leu71Ub
at the end of a b strand in ubiquitin interact
with PLpro (an analogous situation happens
with residues Asn151 and Leu152 of ISG15;
fig. S1C) (Fig. 1B). This is a distinct property
of Jun11313 resulting from the unusual bind-
ing conformation of the phenylthienyl group,
which we exploited in the next compound
series (Fig. 4), exemplified by Jun12682. The
2.52-Å resolution co-crystal structure of PLpro
with Jun12682 reveals the “best of both worlds,”
with the N-ethyl and N-methyl pyrazole sub-
stituents in the phenyl moiety binding toward
the Met208/Pro247 direction (Val70Ub site)
and the Pro248/Tyr264/Tyr268 direction (BL2
groove), respectively (Fig. 4). The Met208
S-methyl moiety maintains CH–p and S–p
interactions with the N-ethyl pyrazole ring and
the Met208 methyl also makes van der Waals
contacts with the Pro247 ring. Moreover, the
phenyl group of Jun12682 has p–p interactions
with the side chain of Tyr264 (Fig. 4). Figure 4
also shows the unbiased electron density for
PLpro inhibitors Jun12145, Jun12303, Jun12199,
Jun12162, Jun12129, Jun12197, and Jun11941,
in a similar orientation as Jun12682. The disub-
stituted phenyl group has two alternative con-
formations in the Jun12145 and Jun12129
complex structures, which are related by a
180° rotation of the phenyl ring (fig. S4). In
concordance with the design strategy (Fig.
1E), the dimethylamino moieties of the flex-
ible side groups make electrostatic interactions
with the Glu167 carboxylate in all structures.
Despite the differences among compounds, the
interactions depicted in Fig. 4 between PLpro
and Jun12682 are largely conserved in all com-
plexes (fig. S4). The crystallographic data and
refinement statistics are listed in table S1.
In vitro and in vivo PK profiling identified
Jun12682 as an in vivo lead candidate
PLpro inhibitors with potent SARS-CoV-2 anti-
viral activity (EC50 ≤ 1 mM) and a high selec-
tivity index (SI > 50) were selected for in vitro
microsomal stability assay and in vivo oral PK
studies in mice. Most pyrazole-containing PLpro
inhibitors showed high stability in the mouse
microsomal stability assay (T1/2 > 60 min) (Fig. 2).
Next, 10 compounds were advanced to the in
vivo oral snap PK in C57BL/6J mice (Fig. 5, A
and B, and table S2). Compounds were dosed
in three male C57BL/6J mice per group at
50 mg/kg through oral gavage, blood samples

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1437

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 4. X-ray crystal structures

of SARS-CoV-2 PLpro with
biarylphenyl PLpro inhibitors.
Atomic model of the Jun12682
(orange balls and sticks) binding
site in SARS-CoV-2 PLpro (residues
within 5 Å of the inhibitor are
shown as light gray sticks), with
hydrogen bonds displayed in
black dashed lines, van der Waals
contacts as red dashed lines,
and p–p interactions as light
green dashed lines. Gallery of
the polder maps (38) of the inhib-
itors, displayed as a gray mesh
with contour levels between 2 and
4s. The PLpro inhibitory constant Ki
in the FRET enzymatic assay and
the SARS-CoV-2 antiviral activity
EC50 in Caco-2 cells were shown for
each compound.

were collected at 0.5, 1, 3, and 5 hours, and the
drug concentration was quantified by LC-MS/MS.
Jun12199 and Jun12682 showed the highest
in vivo plasma concentrations. Given the faster
absorption of Jun12682 compared with Jun12199,
Jun12682 was selected as an in vivo lead candi-
date. A 24-hour in vivo PK study was conducted
to determine the oral bioavailability of Jun12682
(Fig. 5C and table S3). Jun12682 had a rapid ab-
sorption after oral administration (p.o.) (50 mg/
kg) and reached the maximum plasma con-
centration at 1.7 hours (Tmax), with a peak
plasma concentration (Cmax) of 4537 ng/mL.
The clearance of Jun12682 was moderate, with
a half-life (t1/2) of 2.0 hours (table S3). The plas-
ma drug concentration was above the antiviral
EC90 value (1.59 mM) for 7 hours. The oral bio-
availability of Jun12682 was 72.8%.
Further profiling of the in vitro PK proper-
ties revealed that Jun12682 was highly stable
in the human microsomes [T1/2 = 131.9 min,
CLint(mic) = 10.5 mL/min/mg] and had high
selectivity for CYP1A2, 2C9, 2C19, 2D6, and 3A-M
(IC50 > 50.0 mM) (Fig. 5D). Jun12682 had a ki-
netic solubility of 180 mM and a thermodynamic
solubility greater than 5 mg/ml. The mouse
plasma protein binding was 85.4%. Overall,
Jun12682 showed favorable in vitro and in vivo
PK properties amenable for the in vivo anti-
viral efficacy study.
In vivo antiviral efficacy of Jun12682 in a
SARS-CoV-2 infection mouse model
To assess the in vivo antiviral efficacy of
Jun12682, we used a lethal SARS-CoV-2 mouse
model described in (31). This model involves
infecting young BALB/c mice (9 to 12 weeks
old) with a mouse-adapted SARS2-N501YMA30,
resulting in severe lung disease resembling
lung injuries in human patients. This model
has been widely used for evaluating SARS-
CoV-2 therapeutics and vaccine candidates
(32–35). To this end, young BALB/c mice were
intranasally infected with 5600 plaque forming
units (PFUs) of SARS2-N501YMA30 and orally
administered Jun12682 at 500 mg/kg twice
daily for 5 days (Fig. 5E). The weight loss plot
(Fig. 5F) illustrates that mice administered
with the vehicle experienced rapid body weight
loss exceeding 20%, leading to a 100% fatality
rate by 5 days post inoculation (DPI) (Fig. 5G).
By contrast, most mice treated with Jun12682
exhibited reduced weight loss, resulting in an
improved survival rate (0 versus 70%, P < 0.0001).
To further evaluate Jun12682’s in vivo efficacy,
two lower dosages (125 and 250 mg/kg) were
tested with reduced dosing times (from 5 days
down to 3 days) twice daily (Fig. 5H). Mice
treated with a dose of 250 mg/kg showed an
average of 10% maximum weight loss, provid-
ing evident protection from infection compared
with the vehicle- and 125 mg/kg–treated groups,
which exhibited over 20% weight loss (Fig. 5I).
Survival analyses demonstrated that Jun12682-
treated mice had statistically higher survival
rates compared with the vehicle group (0%
survival): 125 mg/kg (20%, P = 0.0428), and
250 mg/kg (100%, P < 0.0001) (Fig. 5J). The
3-day treatment of 250 mg/kg dose conferred
even better protection than the 5-day treatment
of 500 mg/kg dose, possibly due to reduced
drug toxicity. Lung viral load analyses revealed
that at 2 DPI, the vehicle-treated mice had
robust infections in the lungs (mean lung titer
of log10 9.17 ± 0.124 PFU/ml), whereas the
250 mg/kg Jun12682-treated mice had statis-
tically lower lung viral titers (mean lung titers
of log10 8.87 ± 0.194 PFU/ml, P = 0.0199) (Fig.
5K). The antiviral effect of the 250 mg/kg treat-
ment was more evident at 4 DPI with over a log
viral titer reduction compared with the vehicle
treatment (mean lung titers of log10 7.05 ± 0.401
and log10 5.73 ± 0.528 PFU/ml for the vehicle
and 250 mg/kg groups, respectively, P = 0.0156)
(Fig. 5K), corroborating the weight loss and
survival data (Fig. 5, I and J).
Quantitative PCR analysis of the RNA sam-
ples extracted from 2-DPI mice lungs showed
that the 250 mg/kg treatment reduced the
viral nucleocapsid (N) gene level (Fig. 5L) and the
expression of multiple inflammatory cytokines,

1438 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

A10000 B C 10000 D In vitro PK Jun12682

10000 T1/2 = 82.4 min
Mouse liver
CLint(mic) = 16.8 µL/min/mg
Concentrations (ng/ml)

Concentrations (ng/ml)

Concentrations (ng/ml)
microsomal stability
CLint(liver) = 66.6 mL/min/kg
1000 EC90
T1/2 = 131.9 min
Human liver
1000 microsomal stability CLint(mic) = 10.5 µL/min/mg
EC50
P.O. 50mg/kg Jun12682 CLint(liver) = 9.5 mL/min/kg
P.O. 50mg/kg
1000 Jun12682 100 P.O. 50mg/kg CYP1A2 (IC50 > 50.0 µM)
Jun12199 Jun12682
Jun12763 CYP2C9 (IC50 > 50.0 µM)
Jun12197 I.V. 10mg/kg
Jun12395 CYP inhibition CYP2C19 (IC50 > 50.0 µM)
Jun12713 100
Jun12602
Jun12603 10 CYP2D6 (IC50 > 50.0 µM)
Jun12351
Jun11941 CYP3A-M (IC50 > 50.0 µM)
Kinetic solubility 180 µM (90.0 µg/ml) PB (pH 7.4)
Thermodynamic
100 10 1 solubility >5mg/ml PB (pH 7.4)
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0 5 10 15 20 25 Mouse plasma
protein binding 85.4%
Time (h) Time (h) Time (h)

E IN inoculation BID for 5 days

Jun12682 dosing H IN inoculation BID for 3 days Jun12682 dosing
SARS2-N501YMA30 8h 16h 8h 16h 8h 16h 8h 16h 8h SARS2-N501YMA30 8h 16h 8h 16h 8h Virus inoculation
Virus inoculation
Mice necropsy

0 1 2 3 4 5 14 0 1 2 3 4 5 10
BALB/c mice Day post-infection BALB/c mice Day post-infection

F 500mg/kg, BID for 5 days G 500mg/kg, BID for 5 days I BID for 3 days J BID for 3 days

Body weight change (%)

Body weight change (%)

100 100
100
100

Survival Rate (%)

Survival Rate (%)

80 Vehicle
90 125mg p = 0.0428
90 60
Vehicle 50 Vehicle 250mg p < 0.0001
Jun12682 Vehicle 40
80 Jun12682 80
p < 0.0001 125mg 20
250mg
70 0 70 0
0 5 10 15 0 5 10 15 0 5 10 0 3 6 9 12
Day post-infection Day post-infection Day post-infection Day post-infection

K Vehicle L Vehicle M ** Vehicle O 5

Vehicle Q 4
10 2×10 -3 8×10 -4 *
Viral titer (Log10 PFU/mL)

Jun12682 **
Relative mRNA/18S rRNA

Relative mRNA/18S rRNA

Jun12682 Jun12682 Jun12682

Histopathological Score

Antigen Stainng Score

**** 4
9 3
1.5×10 -3 6×10 -4 *
3
8 ns
* 2
1×10 -3 4×10 -4
7 2
* 1
5×10 -4 2×10 -4
6 ** 1

5 0 0 0 0
2 4
se

es
M
ne

-6
0

Ve ck

n1 le
82
H
-

-1

L1

ea

at
N

IL

Ju hic
ge

Day post-infection
a/

26
IL
IF

r
XC

M
s

ilt
em
N

di

f
C

in
Ed

l
ia

ar
tit

ul
rs

sc
te

a
In

r iv
Pe

N Vehicle Jun12682 P Mock Vehicle Jun12682

10X

10X
20X

20X

Fig. 5. In vitro and in vivo PK profiling of PLpro inhibitors and in vivo following oral administration of 50 mg/kg of compound in 0.5% methylcellulose
antiviral efficacy of Jun12682. (A) Plasma drug concentration of Jun12199, and 2% Tween 80 in water. (C) Plasma drug concentration of Jun12682 in
Jun12197, Jun12713, Jun12603, and Jun11941 in C57BL/6J mice (6 to C57BL/6J mice (6 to 8 weeks old) following oral administration of 50 mg/kg and
8 weeks old) following oral administration of 50 mg/kg of compound in 0.5% i.v. injection of 10 mg/kg (n = 3 per group). (D) In vitro PK parameters of
methylcellulose and 2% Tween 80 in water (n = 3 per group). p.o., per os (oral Jun12682. (E) Experimental design for the 5-day treatment experiment. Young
administration). (B) Plasma drug concentration of Jun12682, Jun12763, BALB/c mice were intranasally (IN) inoculated with 5600 PFU of SARS2-
Jun12395, Jun12602, and Jun12351 in C57BL/6J mice (6 to 8 weeks old) N501YMA30 and subsequently orally administered 500 mg/kg Jun12682 or

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1439

RES EARCH | R E S E A R C H A R T I C L E S

vehicle twice a day (BID) for 5 days (BID_5). (F) Body weight loss and
(G) survival rate of the BID_5 mice experiment. (H) Experimental design for the
3-day treatment experiment. Young BALB/c mice were intranasally (IN)
inoculated with 5600 PFU of SARS2-N501YMA30 and subsequently orally
administered 125, 250 mg/kg Jun12682, or vehicle BID for 3 days (BID_3).
(I) Body weight loss and (J) survival rate of the BID_3 mice experiment. Data in
(F), G), (I), and (J) are pooled from two independent experiments (n = 10 per
group) and are shown as mean ± sem. The P values in (G) and (J) were
determined using a log-rank (Mantel–Cox) test. (K) Viral titers in lungs collected at
2 and 4 DPI from vehicle- or 250 mg/kg Jun12682–treated mice (n = 5 per group).
Data are mean ± sem and analyzed with unpaired t test with Welch’s correction.
*, P < 0.05. (L and M) Quantitative PCR analysis of viral nucleocapsid gene (L) and
cellular cytokines (M) in lungs collected at 2 DPI from vehicle- or 250 mg/kg
Jun12682–treated mice (n = 5 per group). (N to Q) Lungs collected at 4 DPI from
vehicle- or 250 mg/kg Jun12682–treated mice (n = 5 per group) were stained with
haematoxylin and eosin (H&E) (N) or immunostained for SARS-CoV-2 nucleocapsid
(P), and the pathological lesions and immunostaining were quantified (O) and (Q),
respectively. (N) H&E stained lungs from vehicle-treated infected mice exhibited
airway edema (asterisks), hyaline membranes (HM, arrowheads), and interstitial
thickness (number sign). Scale bars, 100 mm (top) and 50 mm (bottom). (O)
Summary scores of lung lesions (n = 5 per group). (P) Lungs from vehicle- or
Jun12682-treated mice (n = 5 per group) were immunostained to detect SARS-CoV-2
nucleocapsid protein (brown color staining). Scale bars, 100 mm (top) and 50 mm
(bottom). (Q) Summary scores of nucleocapsid immunostaining of lungs. Data in (L),
(M), (O), and (Q) are mean ± sem and analyzed with unpaired t test with Welch’s
correction. ns, not significant; *, P < 0.05; **, P <0.01; ****, P < 0.0001.

including IFN-b, IL-1b, IL-6, and CXCL10 (36)
(Fig. 5M). Histopathological analysis revealed
that lungs from the vehicle-treated, SARS2-
N501YMA30–infected mice at 4 DPI exhibited
multifocal pulmonary lesions, including lym-
phocytic perivascular cuffing, pulmonary edema,
hyaline membrane formation, and interstitial
thickening and inflammation compared with
Jun12682-treated mice (Fig. 5, N and O, and
fig. S5). Immunohistochemical analysis using
a monoclonal antibody to detect SARS-CoV-2
N protein in the lungs demonstrated strong
and expansive antigen staining in lungs from
vehicle-treated, infected mice, whereas Jun12682
treatment considerably decreased viral anti-
gen staining levels with a few sporadic positive
cells (Fig. 5, P and Q, and fig. S6), consistent
with the lung viral titer results (Fig. 5K). Mock-
infected lungs were negative for nucleocapsid
staining. Overall, the reduced viral replication
in the lungs and the expression of inflamma-
tory cytokines (Fig. 5, K to M) corroborate with
the reduced lung inflammation and N protein
staining at 4 DPI (Fig. 5, N to Q).
Conclusions
SARS-CoV-2 PLpro is a promising target, but
the shallow S1 and S2 subsites near the active
site have hindered the discovery of potent in-
hibitors. We have used a structure-based drug
design approach that resulted in the discovery
of the lead compound Jun12682, which ex-
ploits a previously unknown site Val70Ub, as
well as the known BL2 groove. We showed
that Jun12682 efficiently inhibits not only the
protease but the deubiquitinase and deISGylase
activity of PLpro, as blocking access to the Val70Ub
site impedes the binding of ubiquitin and ISG15.
Oral administration of Jun12682 efficiently
inhibited SARS-CoV-2 replication and miti-
gated SARS-CoV-2–induced lung lesions in vivo.
Jun12682 displayed consistent potency against
SARS-CoV-2 delta and omicron variants, and
nirmatrelvir-resistant strains, suggesting that
PLpro inhibitors can be used to combat cur-
rent and future SARS-CoV-2 variants and
nirmatrelvir-resistant mutants. We anticipate
that PLpro inhibitors could be used alone or in
combination with existing RdRp and Mpro in-
hibitors to achieve a synergistic effect. Combi-
nation therapy is an established strategy for
suppressing resistance evolution and reducing
side effects (37) and can be applied to address
potential drug-resistance mutations that might
emerge at the Val70Ub, BL2 groove, or alloste-
ric sites. Candidate molecules available for
combination therapy with PLpro inhibitors
include the oral Mpro inhibitors nirmatrelvir
and ensitrelvir, and oral RdRp inhibitors mol-
nupiravir and remdesivir analogs (1). Overall,
Jun12682 represents a promising candidate
for further development as an oral SARS-CoV-2
antiviral.
RE FERENCES AND NOTES
1. G. Li, R. Hilgenfeld, R. Whitley, E. De Clercq, Nat. Rev. Drug
Discov. 22, 449–475 (2023).
2. W. H. O. S. T. Consortium; WHO Solidarity Trial Consortium,
Lancet 399, 1941–1953 (2022).
3. L. D. Saravolatz, S. Depcinski, M. Sharma, Clin. Infect. Dis. 76,
165–171 (2023).
4. D. R. Owen et al., Science 374, 1586–1593 (2021).
5. Y. Unoh et al., J. Med. Chem. 65, 6499–6512 (2022).
6. H. Mukae et al., Clin. Infect. Dis. 76, 1403–1411 (2023).
7. R. Shimizu et al., Clin. Drug Investig. 43, 335–346 (2023).
8. R. Shimizu et al., Antimicrob. Agents Chemother. 66, e0063222
(2022).
9. S. Iketani et al., Nature 613, 558–564 (2023).
10. Y. Hu et al., ACS Cent. Sci. 9, 1658–1669 (2023).
11. L. J. Stevens et al., Sci. Transl. Med. 14, eabo0718 (2022).
12. S. Gandhi et al., Nat. Commun. 13, 1547 (2022).
13. N. S. Zuckerman, E. Bucris, D. Keidar-Friedman, M. Amsalem,
T. Brosh-Nissimov, Clin. Infect. Dis. 78, ciad494 (2023).
14. C. P. Sjaarda et al., JAMA Netw. Open 6, e2324963 (2023).
15. Y. Hirotsu et al., Med 4, 813–824.e4 (2023).
16. W. Rut et al., Sci. Adv. 6, eabd4596 (2020).
17. D. Shin et al., Nature 587, 657–662 (2020).
18. P. M. Wydorski et al., Nat. Commun. 14, 2366 (2023).
19. Y. M. Báez-Santos, S. E. St John, A. D. Mesecar, Antiviral Res.
115, 21–38 (2015).
20. H. Tan, Y. Hu, P. Jadhav, B. Tan, J. Wang, J. Med. Chem. 65,
7561–7580 (2022).
21. J. Lei, Y. Kusov, R. Hilgenfeld, Antiviral Res. 149, 58–74
(2018).
22. A. K. Ghosh, J. L. Mishevich, A. Mesecar, H. Mitsuya,
ChemMedChem 17, e202200440 (2022).
23. C. Ma et al., ACS Cent. Sci. 7, 1245–1260 (2021).
24. Z. Shen et al., J. Med. Chem. 65, 2940–2955 (2022).
25. B. C. Sanders et al., Nat. Commun. 14, 1733 (2023).
26. G. Platzer et al., Angew. Chem. Int. Ed. 59, 14861–14868
(2020).
27. B. R. Beno, K. S. Yeung, M. D. Bartberger, L. D. Pennington,
N. A. Meanwell, J. Med. Chem. 58, 4383–4438 (2015).
28. K. Ratia et al., Proc. Natl. Acad. Sci. U.S.A. 105, 16119–16124
(2008).
29. D. Gramec, L. Peterlin Mašič, M. Sollner Dolenc, Chem. Res. Toxicol.
27, 1344–1358 (2014).
30. T. Klemm et al., EMBO J. 39, e106275 (2020).
31. L. R. Wong et al., Nature 605, 146–151 (2022).
32. S. A. El-Kafrawy et al., PLOS Pathog. 18, e1010782 (2022).
33. L. Zhang et al., EMBO Mol. Med. 15, e17376 (2023).
34. J. Shi et al., Transl. Res. 248, 11–21 (2022).
35. V. Fumagalli et al., EMBO Mol. Med. 15, e17580 (2023).
36. H. Nobori et al., J. Antimicrob. Chemother. 77, 2984–2991
(2022).
37. Z. A. Shyr, Y. S. Cheng, D. C. Lo, W. Zheng, Drug Discov. Today
26, 2367–2376 (2021).
38. D. Liebschner et al., Acta Crystallogr. D. 73, 148–157 (2017).
AC KNOWLED GME NTS
We thank S. Cowan and C. Miller at the Immunopathology Core of
the Oklahoma Center for Respiratory and Infectious Diseases for
their technical assistance. We are grateful to K. Das, S. Burley,
and PDB staff members for helpful discussions. Funding: This
work was funded by the following: National Institutes of Health
grant U19AI171110 (to J.W. and E.A.) and National Institutes of
Health grant R01AI158775 (to J.W. and X.D.) Author contributions:
J.W., X.D., and E.A. conceived and supervised the research and
designed the experiments. J.W., B.T., and P.J. designed the inhibitors.
B.T. and P.J. performed chemical syntheses, separation, purification,
and structural characterizations. A.A., A.C., and F.X.R. performed
gene expression, protein purification, crystallization, and diffraction
data collection. A.A., A.C., and F.X.R., and E.A. determined and
analyzed the crystal structures. B.T., H.T., and K.L. performed
enzymatic inhibition assays, DSF assays, and cellular cytotoxicity
assays. X.Z. performed in vitro cellular antiviral assays and
in vivo antiviral studies. A.F. performed the histopathology and
immunohistochemistry (IHC) assessment. X.C. performed the
mouse tissue analysis and generated the recombinant SARS-CoV-2
viruses. J.W., B.T., X.Z., A.A., A.F., X.C., F.X.R., X.D., and E.A.
analyzed and discussed the data with the assistance of P.J.,
H.T., K.L., and A.C.. and J.W., A.A., F.X.R., E.A., and X.D. wrote the
manuscript with the assistance of B.T., X.Z., P.J., H.T., K.L., A.C.,
and A.F. Competing interests: Rutgers, the State University of
New Jersey, has applied for PCT patents 63/447,269 and 63/
594,200 that cover the PLpro inhibitors reported in this manuscript
and related compounds. The inventors include B.T., A.A., P.J.,
A.C., F.X.R., E.A., and J.W. Data and materials availability: The
PDB accession numbers for the coordinates and structure
factors of SARS-CoV-2 PLpro in complex with PLpro inhibitors are
8UUW (Jun12145), 8UUY (Jun12129), 8UUV (Jun12197), 8UUU
(Jun12162), 8UUH (Jun12199), 8UUG (Jun12303), 8UUF
(Jun11941), 8UOB (Jun12682), and 8UVM (Jun11313). All
other data are available in the main text or the supplementary
materials. License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.sciencemag.org/about/science-licenses-
journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adm9724
Materials and Methods
Figs. S1 to S6
Tables S1 to S3
Chemical Synthesis and Spectra
MDAR Reproducibility Checklist
References (39–50)
Submitted 30 November 2023; accepted 22 February 2024
10.1126/science.adm9724

1440 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

CELL CYCLE G1 persisted, and no subsequent M phase was

observed (Fig. 1C). Using the decrease in p21
Control of cell proliferation by memories of mitosis signal as a marker for the G1-S transition re-
vealed that the G1 duration, like p21-mNG ex-
Franz Meitinger1,2,3,4*, Hazrat Belal4, Robert L. Davis5, Mallory B. Martinez5, Andrew K. Shiau1,5, pression, progressively increased for subthreshold
Karen Oegema1,2,3*, Arshad Desai1,2,3* mitotic durations between 30 and 150 min (Fig.
1D). A similar increase in G1 duration, but not in
Mitotic duration is tightly constrained, and extended mitosis is characteristic of problematic cells prone to S+G2 duration, was observed using a cell cycle
chromosome missegregation and genomic instability. We show here that mitotic extension leads to the formation phase sensor in a cell line without tagged p21
of p53-binding protein 1 (53BP1)–ubiquitin-specific protease 28 (USP28)–p53 protein complexes that are (fig. S1, E and F). These results suggest that
transmitted to, and stably retained by, daughter cells. Complexes assembled through a Polo-like kinase an analog memory of mitotic extension is trans-
1–dependent mechanism during extended mitosis and elicited a p53 response in G1 that prevented the mitted to daughter cells, which leads to p21
proliferation of the progeny of cells that experienced an approximately threefold extended mitosis or successive expression in G1 that reflects the extent to
less extended mitoses. The ability to monitor mitotic extension was lost in p53-mutant cancers and some which the mother’s mitosis was prolonged. The
p53–wild-type (p53-WT) cancers, consistent with classification of TP53BP1 and USP28 as tumor suppressors. fact that the progressive increase in p21 abun-
Cancers retaining the ability to monitor mitotic extension exhibited sensitivity to antimitotic agents. dance was converted into a sharp threshold in
mother cell mitotic time for G1 arrest is consis-
tent with a stoichiometric inhibition of G1 Cdk
itosis is a complex event that occurs nontransformed cell lines (fig. S1B) by treating complexes by p21 exhibiting ultrasensitivity

M
within a tightly constrained time frame
(1). Prolonged mitosis is a sign of prob-
lems that can lead to chromosome mis-
segregation, a trigger event for genomic
instability (2, 3). Identifying cells that have
experienced extended mitosis is therefore an
effective means to identify potentially problem-
atic cells in a proliferative population. In hu-
man nontransformed immortalized retinal
pigment epithelial (RPE1) cells, when a thresh-
old mitotic duration is surpassed, the resulting
daughter cells arrest in G1 in a p53-dependent
manner. Arrest of daughter cells is independent
of the cause of the mitotic extension (4). In
addition to p53 and its transcriptional target,
the cell cycle kinase inhibitor p21, two addition-
al factors, the scaffold protein 53BP1 and the
deubiquitinase USP28, are required for daughter
cell arrest after extended mitosis (5–7) (Fig. 1,
A and B, and fig. S1, A and B). When spindle
assembly is prolonged after centriole removal
in the early mouse embryo or developing ner-
vous system, the observed cell death, embryonic
arrest, and microcephaly phenotypes are ame-
liorated by deletion of the TP53 or USP28
genes, indicating that this pathway is opera-
tional during development (8–13). However,
how mitotic extension is monitored and the
physiological significance of mitotic extension–
regulated control of cell proliferation remain
unknown.
An analog memory of mitotic extension is
transmitted to daughter cells
We monitored the response to mitotic exten-
sion in RPE1 cells (Fig. 1B) and two additional
1
Department of Cell and Developmental Biology, School of
Biological Sciences, University of California San Diego,
La Jolla, CA 92093, USA. 2Department of Cellular and
Molecular Medicine, University of California San Diego,
La Jolla, CA 92093, USA. 3Ludwig Institute for Cancer Research,
La Jolla, CA 92093, USA. 4Okinawa Institute of Science and
Technology Graduate University, Okinawa 904-0495, Japan.
5
Small Molecule Discovery Program, Ludwig Institute for Cancer
Research, La Jolla, CA 92093, USA.
*Corresponding author. Email: franz.meitinger@oist.jp (F.M.);
koegema@ucsd.edu (K.O.); abdesai@ucsd.edu (A.D.)
asynchronously growing cells with the spindle
assembly inhibitor monastrol for 6 hours. Be-
cause cells enter mitosis at various times, this
treatment generates cells with mitotic durations
between ~30 and ~400 min. After inhibitor
washout, cells completed mitosis, and the
resulting daughter cells were followed for
48 hours to determine whether they arrested,
underwent apoptosis, or continued to prolifer-
ate (Fig. 1B and fig. S1B). All three tested cell
lines were sensitive to extended mitosis that
exceeded a threshold of 90 to 110 min (normal
duration = 30 to 40 min; Fig. 1B and fig. S1B).
To investigate the observed sharp transition
in daughter cell fate as a function of mother
cell mitotic duration, we in situ–tagged the p53
effector p21, which is necessary for daugh-
ter cell G1 arrest (Fig. 1B), with mNeonGreen
(mNG). In the p21-mNG line, the threshold
mother cell mitotic duration for daughter cell
arrest was ~50% longer than in the parental
line (fig. S1, C and D), possibly due to the flu-
orescent protein tag. We used monastrol to
prolong mitosis and correlated mother cell
mitotic duration with p21-mNG expression in
daughter cells to determine whether p21 only
accumulated in daughters of mother cells
with mitotic duration that surpassed the arrest
threshold or if increasing mitotic durations
were encoded in progressively higher amounts
of p21 in G1 that trigger arrest when a thresh-
old is surpassed. After mitoses with a normal
duration of ~30 min, p21 expression was low
in most cells (Fig. 1, C and D). Extending mi-
totic duration to subthreshold times between
60 and 150 min led to progressive increases in
p21 abundance (Fig. 1, C and D, and fig. S1C).
p21-mNG signal was only observed after transit
into G1. For mitotic extensions below the 150-min
threshold, p21-mNG signal was detected in G1
and then abruptly decreased, possibly because
of ubiquitin-dependent degradation in early S
phase (14). p21-mNG was then detected again
in G2 before the next M phase (Fig. 1C). Above
the 150-min threshold, high p21-mNG signal in
(15, 16). Because the mechanism monitoring
mitotic duration acts in an analog manner and
is proportional to time, we refer to it as the
mitotic stopwatch.
Memory of mitotic extension is integrated
across cell cycles to control proliferation
The results above indicate that a memory of
mitotic duration is transmitted to immediately
produced daughter cells. We therefore inves-
tigated whether a memory of mitotic extension
can also be transmitted in a multigenerational
manner across successive cell cycles. For this
purpose, we developed a means to extend aver-
age mitotic duration from ~30 min to ~60 min,
well below the ~100-min threshold for immedi-
ate daughter cell G1 arrest (Fig. 1E). We achieved
this through inducible inactivation of the gene
encoding the mitotic checkpoint complex dis-
assembly factor Comet, which accelerates but
is not essential for anaphase onset [iCometD;
Fig. 1E and fig. S2, A and B; (17, 18)]. Although
it delays anaphase onset, Comet deletion does
not perturb spindle assembly or chromosome
alignment and segregation (Fig. 1E and fig.
S2C). Despite mother cells transiting mitosis in
comparable time intervals and without visible
segregation errors, the frequency of daughter
cell arrest after subthreshold mitotic durations
of 60 to 90 min increased progressively between
days 2 and 4 after induction of Comet deletion,
as cells experienced successive subthreshold
extended mitoses (Fig. 1, E and F, and fig. S2B).
Daughter cell arrest resulting from iCometD
was greatly reduced in cell lines lacking USP28
or p53 (Fig. 1F and fig. S2D). Thus, a mitotic
stopwatch-dependent memory of subthreshold
mitotic extension is transmitted and integrated
across sequential cell cycles to control cell pro-
liferation. An orthogonal perturbation by chemical
inhibition of PLK4, which leads to a subthresh-
old mitotic extension by delaying spindle as-
sembly (19), coupled with imaging of p21-mNG,
provided additional support for this conclusion
(fig. S3, A to G).

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1441

RES EARCH | R E S E A R C H A R T I C L E S
Fig. 1. Control of G1 progression and cell proliferation by memories of
mitosis. (A) Schematic highlighting that prolonging mitosis beyond a threshold
about three times longer than normal mitosis leads to G1 arrest of daughter
cells that requires p53, p21, 53BP1, and USP28. (B) Plots of the fate of daughter
cells as a function of mother cell mitotic duration for WT and CDKN1AD
(p21 knockout) RPE1 cell lines. Mother cells with mitotic durations between
30 and 400 min were generated by treating an asynchronous cell population with
a reversible inhibitor of spindle assembly for 6 hours. Mother cells were imaged
in the presence of the inhibitor to measure mitotic duration. After inhibitor
washout, mother cells completed mitosis and the resulting daughter cells were
imaged for 48 h to determine whether they divided again (gray), arrested (red),
or underwent apoptosis (black; not observed for RPE1 cells). Each bar represents
a single daughter cell, with bar height representing the mitotic duration of its
mother and color representing its fate. Percentage of arrested daughter cells
produced by mothers that spent ≥100 min in mitosis is noted above the black
lines. See also fig. S1, A and B. (C) Stills from time-lapse movies monitoring in
situ–tagged p21-mNG and H2B-RFP. Monastrol treatment and washout was used
1442 29 MARCH 2024 • VOL 383 ISSUE 6690
to prolong mitosis to different extents, and p21-mNG expression was
monitored in daughter cells. Shown are mother cells of different mitotic durations
(left) and one of their daughters (right). (D) Left: box-and-whiskers plot of
peak daughter cell p21-mNG expression in G1 as a function of mother cell mitotic
duration. Right: G1 duration of daughter cells produced by mother cells of the
indicated mitotic durations. Mean and 95% confidence intervval (CI) are
indicated. P values in (D) are from t tests (**P < 0.01; ***P < 0.001;
****P < 0.0001). See also fig. S1, C to F. (E) Panels from time-lapse movies of
representative control and iCometD cells expressing H2B-RFP. NEBD, nuclear
envelope breakdown. (F) Left: mitotic duration at different days after induction
of Comet knockout, which extends mitosis by slowing the disassembly of
mitotic checkpoint complexes; mean and SD are indicated. Right: frequency of
daughter cells that arrest produced by mothers with 60- to 90-min mitotic
duration on different days after induction of Comet knockout. Comet knockout
was also induced and analyzed in USP28D (shown here) and TP53sh
(fig. S2D) backgrounds. See also figs. S2 and S3, A to G. Scale bars in (C)
and (E), 5 mm.
science.org SCIENCE

Fig. 2. Mitosis-specific formation of stopwatch complexes transmits
memory of extended mitotic duration to daughter cells. (A) Schematic
highlighting the requirement for 53BP1 and USP28 for mitotic stopwatch
function. (B) Immunoblots monitoring solubility of 53BP1 and USP28 in
asynchronous and mitotic cell extracts. WCE, whole-cell extract. GAPDH and
histone H3 are soluble and chromatin-bound insoluble controls, respectively. See
also fig. S4A. (C to E) Analysis of 53BP1 immunoprecipitates from: asynchronous
cells or cells treated to arrest at indicated cell cycle stages (C), asynchronous
cells and cells treated to induce DNA damage or prolong mitosis (D), and cells of
the indicated genotypes treated to prolong mitosis (E). Treatment details are
indicated in fig. S4B. Inputs are soluble supernatants. IP, immunoprecipitate;
No Ab., beads-only control. a-tubulin and GAPDH served as loading controls. See
also fig. S4, C and D. (F) Top: schematic showing the location of the G1560K
point mutation in 53BP1’s Tudor domain that disrupts its interaction with USP28
(24) and was introduced by base editing of the endogenous TP53BP1 locus
(see fig. S4E). Bottom: analysis of 53BP1 immunoprecipitates from cells treated
to prolong mitosis as described in fig. S4B. (G) Functional analysis of the mitotic
SCIENCE science.org
RESE ARCH | R E S E A R C H A R T I C L E S
stopwatch for the indicated engineered mutant lines. The control graph is the
same as in Fig. 1B. (H) Analysis of 53BP1 immunoprecipitates from cells treated
to prolong mitosis or after release from synchronization at the G2-M boundary
using a CDK1 inhibitor into an unperturbed mitosis. Treatment details are
outlined in fig. 5A. Anti-pS10 H3, which monitors mitosis-specific phosphorylation
of histone H3 on Ser10, was used as a marker for mitosis and GAPDH served as a
loading control. (I) Analysis of 53BP1 immunoprecipitates from synchronized
cells held in mitosis for ~2 or ~8 h using the protocol schematized on the top.
GAPDH served as a loading control. Lanes shown are from a single exposure of
the same immunoblot. See also fig. S5B. (J) Analysis of the stability of stopwatch
complexes after their formation in prolonged mitosis. CDKN1AD cells were
used to avoid the G1 arrest observed after release from extended mitosis
(Fig. 1B) and were treated as indicated in the schematic above the blot. See also
fig. S5D. Twelve and 24 h after release from mitosis represents late G1/S and
G2 phases of the daughter cells’ cell cycle respectively. GAPDH served as a
loading control. Progression of daughter cells into the next mitosis was
confirmed by live imaging (see fig. S5E).
29 MARCH 2024 • VOL 383 ISSUE 6690 1443
RES EARCH | R E S E A R C H A R T I C L E S

Fig. 3. PLK1 kinase activity is central to the formation of stopwatch
complexes that are transmitted to daughter cells. (A) Top left: experimental
approach and list of inhibited mitotic and DNA damage kinases. Plots show
results of functional analysis of the mitotic stopwatch for the indicated
conditions. See also figs. S6 and S7, A to D. (B) Analysis of 53BP1
immunoprecipitates from cells treated to prolong mitosis with and without PLK1
inhibition. For treatment details, see fig. S7E. (C) Analysis of PLK1 activity–
dependent mitotic phosphorylation sites in the 53BP1 BRCT region. Schematic
above indicates the three sites mutated to nonphosphorylatable alanine.
Immunoblot below shows analysis of 53BP1 immunoprecipitates from cells
treated to prolong mitosis, comparing WT RPE1 and TP53BP1D RPE1 cells after
introduction of transgenic WT and phosphosite mutant forms of 53BP1. See also
fig. S8, A to H. Lanes shown are from a single exposure of the same immunoblot;
the full blot is shown in fig. S8D. (D) Analysis of p53 levels by immunostaining
after 4-day treatment with PLK4i, which delays spindle assembly (fig. S3A).
Transgene expression is heterogeneous; therefore, cells with comparable
transgene expression are shown. (E) Quantification of p53 fluorescence signal in
nuclei comparing dimethylsulfoxide (DMSO) and PLK4i treatment for the
indicated conditions. Because of the heterogeneity of transgene expression in
TP53BP1D cells, 53BP1 signal intensity was used to first select cells with
comparable expression (fig. S8H) and p53 signal was then quantified. The 10th to
the 90th percentile of measured values normalized to the average value in
DMSO-treated TP53BP1D+WT condition are plotted. P value is from a t test
(****P < 0.0001). (F) Schematic summary of mitotic extension being encoded
by PLK1 activity–dependent formation of stopwatch complexes that are stably
inherited by daughter cells in which, depending on their abundance, either
trigger immediate proliferation arrest or impart a memory of prolonged mitosis
into the subsequent cell cycle.

Mitotic stopwatch complexes transmit
memory of mitotic extension
To understand the molecular basis for the trans-
mission of an analog memory of extended
mitosis to daughter cells, we conducted bio-
chemical analysis of 53BP1 and USP28 in ex-
tracts prepared from mitotically arrested and
asynchronously cycling (primarily interphase)
cells. Although USP28 was soluble in both, 53BP1
was soluble in mitotic extracts but largely in-
soluble in interphase extracts (Fig. 2, A and B,
and fig. S4, A and B); low solubility of 53BP1
in interphase is consistent with its chromatin
association (20, 21). Immunoprecipitation
of 53BP1 from cells held in different cell cycle
states revealed co-immunoprecipitation of 53BP1,
USP28, and p53 specifically from mitotic cells
(Fig. 2C and fig. S4, B to D). Neither an inter-
action between 53BP1 and USP28 nor increased
53BP1 solubility was detected in cells with DNA
damage despite the increased abundance of p53,
indicating that complex formation in mitosis
is not a secondary consequence of DNA damage
(Fig. 2D and fig. S4, A and B). The cell cycle arrest
observed when mitosis is extended by centriole
depletion also does not exhibit any of the mo-
lecular signatures of the DNA damage–associated

1444 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 4. The mitotic stopwatch is compromised in cancers and influences the efficacy of
antimitotic agents. (A) Left: schematic of diverse tissue-of-origin cancer-derived cell lines annotated
as expressing p53-WT (blue) or p53-mutant (orange). For experimental confirmation of p53 status, see
fig. S10A. Right: representative plots analyzing mitotic stopwatch function in p53-WT cancer-derived cell lines.
See also fig. S10, B to E. (B) Plot of daughter cell arrest and/or death for mother cell mitotic durations above
the indicated thresholds in the p53-WT and p53-mutant cancer cell lines. (C) Top: USP28 schematic showing
mutations identified in p53-WT cancer lines that lack stopwatch function. Bottom: USP28 immunoblot for the
indicated cell lines. a-tubulin was used as a loading control. See also figs. S10, F and G, and S11. (D) Left:
mean relative proliferation in PLK4i versus DMSO plotted for successive 4-day intervals in a passaging assay
for 15 p53-WT and 3 p53-mutant cancer lines. See also fig. S12. p53-WT lines are grouped based on the
functionality of their mitotic stopwatch. Right: mean relative proliferation in PLK4i of CHP134 neuroblastoma
cells and derived isogenic clonal lines with mutations and/or knockdown of stopwatch complex components.
Error bars are the SD (n = 3). See also fig. S13, A to C. (E) Mean relative proliferation of parental CHP134
cells and derived lines knocked down or mutated for stopwatch components in 2 nM Taxol (left) or 40 nM
CENPEi (right); n = 9 for controls and n = 3 for other conditions. Error bars are the SD. P values are from
pairwise t tests comparing derived cell lines with the parental line (****P < 0.0001). See also fig. S13D.

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1445

RES EARCH | R E S E A R C H A R T I C L E S
p53 response (19). Immunoprecipitation of 53BP1
from mitotically arrested cells lacking USP28
or p53 revealed that 53BP1 independently in-
teracts with the other two complex components
(Fig. 2E), which is consistent with the C-terminal
tandem BRCT domain of 53BP1 interfacing with
p53 (22, 23) and a point mutation (G1560K) in
the Tudor domain of 53BP1 disrupting interac-
tion with USP28 (24). Introducing the G1560K
mutation by base editing selectively disrupted
the interaction of 53BP1 with USP28 in mito-
tically arrested cells (Fig. 2F and fig. S4E) and
inhibited stopwatch function to a similar extent
as did TP53BP1 knockout (Fig. 2G). Thus, stop-
watch complex formation is required to limit
the proliferation of daughter cells in response
to mitotic extension.
Although the stopwatch mechanism can de-
tect relatively small mitotic extensions, it is in-
sensitive to normal mitotic duration. Consistent
with this observation, the formation of com-
plexes containing 53BP1, USP28, and p53 was
observed in cells held in mitosis, but not in
cells released from synchronization at the G2/M
boundary into an unperturbed mitosis (Fig. 2H
and fig. S5A). Comparing stopwatch complex
formation after release from G1 arrest into the
microtubule polymerization inhibitor nocoda-
zole revealed that stopwatch complexes could
be detected in cells held for 2 hours in mitosis
and were present in increased amounts in cells
held in mitosis for 8 hours (Fig. 2I and fig. S5B).
A progressive increase in stopwatch complex
abundance was also observed after treatment
of asynchronous cells with nocodazole for 4 to
16 hours (fig. S5C).
The mitotic stopwatch mechanism monitors
mildly extended mitoses and integrates these
extensions over subsequent cell cycles to trigger
proliferation arrest (Fig. 1, E and F). This prop-
erty requires transmission of a molecular mem-
ory of extended mitotic duration across cell
cycles. Such a memory is unlikely to reside in
p21 protein or mRNA, because p21 protein is
degraded during S-phase entry (Fig. 1C) and
p21 mRNA lifetime is relatively short (25, 26).
We thus tested whether, once formed during
extended mitosis, stopwatch complexes were
stable enough to transmit the memory of ex-
tended mitosis across cell cycles. To assess the
stability of stopwatch complexes as cells pro-
gressed through G1, S, and G2, we used cells
lacking p21 because they do not arrest in G1
after release from extended mitosis (Fig. 1B).
Cells lacking p21 were held in mitosis and then
released to allow daughter cells to progress
through the cell cycle and eventually enter
the subsequent mitosis (fig. S5, D and E). This
analysis revealed that amounts of USP28 co-
immunoprecipitated with 53BP1 were compa-
rable during extended mitosis and when cells
were in G1/S or G2, 12 or 24 hours after release
from extended mitosis respectively (Fig. 2J).
More p53 was detected in the 53BP1 immuno-
1446 29 MARCH 2024 • VOL 383 ISSUE 6690
precipitates, and p53 abundance was higher
overall at the postrelease time points, indicat-
ing that 53BP1-USP28 complexes are not stoi-
chiometrically bound to p53 during prolonged
mitosis. By comparison, cells held in G1 or G2
without a prior extended mitosis did not con-
tain stopwatch complexes (Fig. 2C). Collect-
ively, these data indicate that mitotic stopwatch
complexes are only formed if mitosis is ex-
tended beyond its normal duration, increase
progressively in abundance during extended
mitosis, are inherited by daughter cells, and
are sufficiently stable to transmit the memory
of extended mitosis.
PLK1 kinase activity is required for mitotic
stopwatch complex formation and function
To address how mitotic stopwatch complexes
form during extended mitosis, we performed a
focused screen with chemical inhibitors target-
ing protein kinases that act in mitosis or after
DNA damage (Fig. 3A). CDK1, MPS1, and Aurora
B were not tested because they are essential to
maintaining an extended mitotic state in re-
sponse to defective spindle assembly. Kinase in-
hibitors were added at the same time as the
reversible spindle assembly inhibitor monastrol
used to extend mitosis (Fig. 3A). This screen
identified the mitotic kinase PLK1 as being
essential for the mitotic stopwatch (Fig. 3A).
By contrast, the kinases that acted after DNA
damage did not contribute to stopwatch func-
tion (Fig. 3A and fig. S6, A to C). Inhibition of
Aurora A had only a minor effect on stopwatch
function (fig. S7A), and BET bromodomain in-
hibition, which is a secondary consequence of
commonly used PLK1 inhibitors including BI2536
(27), had no effect (fig. S7B). A chemically dis-
tinct and PLK1-specific inhibitor also suppressed
stopwatch function (fig. S7C), and introduction
of a mutation conferring partial PLK1 inhibitor
resistance restored stopwatch function in the
presence of the inhibitor, confirming the spe-
cific requirement for PLK1 activity (fig. S7D).
53BP1 becomes soluble during mitosis (Fig. 2,
C and H) and was equally soluble in cells held
in mitosis with nocodazole, nocodazole and PLK1i,
or PLK1i only, indicating that 53BP1 release
from chromatin during mitotic entry does not
require PLK1 activity. Instead, PLK1 activity
appears to be required for the association of
USP28 and p53 with 53BP1 to form stopwatch
complexes when mitosis is extended (Fig. 3B
and fig. S7E). After release into G1, cells in which
PLK1 was inhibited during extended mitosis
behaved like cells experiencing normal-length
mitosis, with 53BP1 becoming insoluble upon
exit from mitosis. By contrast, 53BP1 remained
soluble if PLK1 was not inhibited during the ex-
tended mitosis and stopwatch complexes were
formed (fig. S7F). This analysis also confirmed
that p21 accumulation is only observed after
the transition to G1 (fig. S7F), as observed with
in situ–tagged p21 (Fig. 1C).
To determine whether PLK1 stimulates com-
plex formation by phosphorylating stopwatch
complex subunits, we focused on 53BP1 be-
cause it has PLK1 activity–dependent phos-
phorylation sites and putative PLK1-docking
sites, one of which has been validated (28, 29).
Because USP28 and p53 interface with the Tudor
and BRCT repeat regions of 53BP1, respectively
(23, 24, 30), we mutated 23 previously identi-
fied PLK1 activity–dependent phosphorylation
sites in 53BP1 that span this region (28), along
with 12 additional high-scoring candidate PLK1
sites predicted by GPS 5.0 (31). The 35 phosphor-
ylation sites targeted for mutation were
grouped into seven clusters, and we mutated
the validated PLK1-docking site in the 53BP1
N terminus (fig. S8, A and B).
Transgenes mutating the grouped sites to
alanine were introduced into TP53BP1D cells,
and stopwatch complex formation was analyzed
after prolonged mitosis (fig. S8, C to E). Dis-
rupting three phosphosites in 53BP1 that were
sensitive to PLK1 inhibition in vivo and located
in a disordered loop of the BRCT1 repeat that
interfaces with p53 (23) reduced interaction
with p53 in cells experiencing extended mitosis
(Fig. 3C and fig. S8, C to G). Mutation of these
sites also compromised stopwatch function, as
monitored by using PLK4 inhibition to extend
mitosis and measuring p53 levels by immuno-
staining (Fig. 3, D and E, and fig. S8H). A second
mutant targeting a cluster of sites between the
Tudor and BRCT domains reduced interaction
with both USP28 and p53 and compromised
stopwatch function (fig. S8, B and D); however,
this mutant also impaired nuclear localization
of 53BP1 (fig. S8E). These results indicate that
PLK1 activity directly contributes to stopwatch
complex formation during prolonged mitosis.
Collectively, the results above indicate that
mitotic entry releases 53BP1 from chromatin
through a PLK1 activity–independent mecha-
nism. Subsequently, if mitosis is prolonged, the
degree of mitotic extension is encoded in the
formation of progressively greater numbers of
stopwatch complexes (Fig. 3F). Stopwatch com-
plex formation involves the PLK1-regulated
interaction of 53BP1’s Tudor and BRCT domains
with USP28 and p53, respectively. For the p53
interaction, a set of mitotic PLK1 target sites
in 53BP1 appear to have an important role.
Incorporation of 53BP1 into stopwatch com-
plexes prevents it from reassociating with chro-
matin when daughter cells exit mitosis (Fig. 2J
and fig. S7F); instead, soluble stopwatch com-
plexes constitute a stable mark of mitotic ex-
tension that is transmitted to daughter cells.
In G1, stopwatch complexes stabilize p53, like-
ly through USP28-mediated deubiquitination,
leading to synthesis of an amount of p21 that
reflects the magnitude of the mitotic extension
(Fig. 3F). If p21 concentration is insufficient to
halt daughter cells in G1, stopwatch complexes
persist, transmitting a memory of extended
science.org SCIENCE

mitosis that enables the detection of cells that
experience sequential moderately prolonged
mitoses (Fig. 1, E and F, and figs. S2 and S3).
Frequent inactivation of the mitotic stopwatch
in p53-WT cancers
Difficulty in executing events such as spindle
assembly or chromosome-spindle attachment
extends mitosis through the spindle checkpoint
and is associated with elevated rates of chro-
mosome missegregation (32), which in turn lead
to aneuploidy and genomic instability. The mito-
tic stopwatch monitors mitotic duration and
halts the proliferation of cells that experience
either one highly delayed mitosis or successive
moderately delayed mitoses, making it well suited
to act as a fidelity filter that suppresses the
proliferation of potentially dangerous cells in
a population. Consistent with this notion, the
genes encoding the three stopwatch complex
subunits are tumor suppressor genes (33) (fig.
S9). Classification of USP28 as a tumor sup-
pressor gene is particularly notable because,
unlike TP53BP1 and TP53, it is not an integral
component of the DNA damage response (34).
Because stopwatch function relies on p53 (4),
the stopwatch is predicted to be inactive in the
~50% of human cancers that harbor p53 muta-
tions. This expectation was confirmed in three
p53-mutant cancer-derived cell lines (Fig. 4,
A and B, and fig. S10, A and B). A more in-
teresting question related to the physiological
function of the stopwatch is what its status
is in the ~50% of cancers that express p53-
WT (35, 36). To determine the status of the
stopwatch in p53-WT cancers, we surveyed
15 cell lines from pediatric and adult cancers
that were annotated as expressing p53-WT
(Fig. 4A). For all tested cell lines, expression of
functional p53 was confirmed by monitoring
cell proliferation after treatment with an in-
hibitor that stabilizes p53 by preventing its
ubiquitination by MDM2 (fig. S10A). Analysis of
stopwatch function partitioned the 15 p53-WT
cell lines into three groups: (i) those with a
functional stopwatch (n = 5), (ii) those with a
partially compromised stopwatch (n = 2), and
(iii) those with an inactive stopwatch (n = 8)
(Fig. 4, A and B, and fig. S10, C to E). In two
neuroblastoma-derived p53-WT cancer cell
lines, stopwatch activation led to cell death
instead of arrest (Fig. 4A, and fig. S10C), possibly
because of higher expression or activity of the
cell death machinery. Of the five cell lines with
a functional stopwatch, three appeared to have
a higher temporal threshold for stopwatch ac-
tivation (Fig. 4, A and B, and fig. S10C).
Consistent with bioinformatic analysis indi-
cating that USP28 and TP53BP1 exhibit a high
frequency of deleterious mutations in cancer
(33) (fig. S9), mutations in USP28 or TP53BP1
or both were present in four of the eight cell
lines lacking stopwatch function (Fig. 4C and
fig. S10, F and G). Creating mutations that re-
SCIENCE science.org
sembled cancer-associated mutations in the
nontransformed RPE1 cell line revealed that
introducing frameshift mutations in both alleles
of either USP28 or TP53BP1 eliminated protein
expression and abrogated stopwatch function
(fig. S11, A to E), whereas heterozygous frame-
shift mutations reduced protein expression by
~50% and compromised stopwatch function
(fig. S11, A to E). Three other p53-WT cancer
cell lines lacking stopwatch function had genetic
alterations that dampen p53 signaling (fig. S11,
F to H). In particular, mutations that truncate
and hyperactivate the p53-antagonizing phos-
phatase WIP1 (37) were associated with loss of
stopwatch function (Fig. 4B and fig. S11F), and
treatment with a WIP1 inhibitor (38) partially
restored stopwatch activity (fig. S11G). Col-
lectively, these results indicate that both
p53-mutant and a substantial proportion of
p53-WT human cancers have compromised
mitotic stopwatch function, which may en-
able them to tolerate problematic mitoses that
are both a cause and a consequence of ge-
nomic instability.
Mitotic stopwatch status predicts sensitivity
to antimitotic agents
Because the stopwatch monitors mitotic dura-
tion to control proliferation, its functional status
has the potential to influence the efficacy of
agents that prolong mitosis and are already
being used or are in development as cancer
therapeutics. To prolong mitosis to a moderate
extent, we used the PLK4 inhibitor centrinone
(19), which delays spindle assembly by causing
loss of centrosomes (fig. S12, A to D). Moni-
toring the proliferation of the three p53-mutant
and 15 p53-WT cancer cell lines revealed that
p53-mutant lines continued to proliferate in
PLK4i, albeit at reduced rates due to mitotic
challenges caused by centrosome loss [Fig. 4D
and fig. S12E (19)]. For the p53-WT cancer lines,
proliferation in PLK4i was broadly inversely
correlated with stopwatch status. p53-WT lines
with a functional mitotic stopwatch exhibited
rapid and penetrant cessation of proliferation
in PLK4i (Fig. 4D and fig. S12E); by contrast,
those with a compromised or absent stopwatch
exhibited improved proliferation in PLK4i (Fig.
4D). To extend this correlational analysis to an
isogenic context, CHP134 neuroblastoma cells
that have an intact stopwatch and are highly
sensitive to PLK4i were engineered to delete or
mutate the three stopwatch complex sub-
units. Inactivation of stopwatch components
improved proliferation in PLK4i relative to
that of parental CHP134 cells (Fig. 4D and
fig. S13, A to C), indicating that a functional
stopwatch enhances the response to PLK4i
treatment. Although PLK4 inhibitors have
the potential to target specific cancer types
(39, 40), they are just beginning to be eval-
uated clinically. Thus, we assessed whether
stopwatch status affected sensitivity to the
RESE ARCH | R E S E A R C H A R T I C L E S
widely used mitosis-targeting chemotherapy
agent taxol and to a clinically tested inhibitor
of the mitotic kinesin CENPE (41–45). Low-dose
taxol and CENPEi treatments extend mitosis
by mildly perturbing microtubule dynamics
and chromosome congression, respectively.
In both CHP134 neuroblastoma cells and
in a nontransformed model cell line, muta-
tion of stopwatch complex subunits improved
proliferation relative to the parental lines after
low-dose taxol or CENPEi treatment (Fig. 4E
and fig. S13D). Thus, stopwatch status may
influence the efficacy of therapeutic agents
currently in use or being developed to target
mitotic processes and could serve as a potential
biomarker for their use in cancer treatment.
Conclusions
Defects in chromosome segregation during mi-
tosis, which not only generate cells with incorrect
numbers of chromosomes but also precipitate
drastic chromosome rearrangements, are inter-
mediates during the generation of nearly all
cancers (46, 47). Given the hundreds of billions
of cells that divide in an adult human every day,
monitoring these mitoses to filter out poten-
tially problematic cells poses a major challenge.
This work shows that the degree of mitotic
extension is encoded through the PLK1 kinase–
dependent assembly of mitotic stopwatch com-
plexes that are transmitted to daughter cells.
The ability of this mechanism to detect subtle
repeated extensions of mitosis functions as a
fidelity filter in a proliferative cell population
and may explain its frequent inactivation in
cancers, as well as the classification of stopwatch
complex components as tumor suppressors.
Compromised stopwatch function is likely im-
portant for the tolerance of problematic mito-
ses that are contributors to and a consequence
of the aneuploidy and genomic instability that
is characteristic of cancers.
REFERENCES AND NOTES
1. A. R. Araujo, L. Gelens, R. S. Sheriff, S. D. Santos, Mol. Cell 64,
362–375 (2016).
2. N. T. Umbreit et al., Science 368, eaba0712 (2020).
3. O. Shoshani et al., Nature 591, 137–141 (2021).
4. Y. Uetake, G. Sluder, Curr. Biol. 20, 1666–1671 (2010).
5. F. Meitinger et al., J. Cell Biol. 214, 155–166 (2016).
6. B. G. Lambrus et al., J. Cell Biol. 214, 143–153 (2016).
7. C. S. Fong et al., eLife 5, e16270 (2016).
8. H. Bazzi, K. V. Anderson, Proc. Natl. Acad. Sci. U.S.A. 111,
E1491–E1500 (2014).
9. C. Xiao et al., EMBO Rep. 22, e51127 (2021).
10. R. Insolera, H. Bazzi, W. Shao, K. V. Anderson, S. H. Shi,
Nat. Neurosci. 17, 1528–1535 (2014).
11. T. P. Phan, A. J. Holland, Genes Dev. 35, 1551–1578
(2021).
12. T. P. Phan et al., EMBO J. 40, e106118 (2021).
13. M. Marjanović et al., Nat. Commun. 6, 7676 (2015).
14. A. R. Barr et al., Nat. Commun. 8, 14728 (2017).
15. J. E. Ferrell Jr., S. H. Ha, Trends Biochem. Sci. 39, 556–569
(2014).
16. H. W. Yang, M. Chung, T. Kudo, T. Meyer, Nature 549,
404–408 (2017).
17. K. D. Corbett, Prog. Mol. Subcell. Biol. 56, 429–455
(2017).
18. D. H. Kim et al., Nat. Commun. 9, 4354 (2018).
19. Y. L. Wong et al., Science 348, 1155–1160 (2015).
29 MARCH 2024 • VOL 383 ISSUE 6690 1447
RES EARCH | R E S E A R C H A R T I C L E S

20. A. Fernandez-Vidal, J. Vignard, G. Mirey, Int. J. Mol. Sci. 18, PLANT SCIENCE
2611 (2017).
21. M. Zimmermann, T. de Lange, Trends Cell Biol. 24, 108–117
(2014). Biosynthesis of the allelopathic alkaloid gramine in
22. D. J. Derbyshire et al., EMBO J. 21, 3863–3872
(2002).
23. W. S. Joo et al., Genes Dev. 16, 583–593 (2002).
barley by a cryptic oxidative rearrangement
24. R. Cuella-Martin et al., Cell 184, 1081–1097.e19 (2021).
25. J. R. Porter, B. E. Fisher, E. Batchelor, Cell Syst. 2, 272–282 Sara Leite Dias1†, Ling Chuang2,3†‡, Shenyu Liu2,3†, Benedikt Seligmann2, Fabian L. Brendel1,
(2016). Benjamin G. Chavez1, Robert E. Hoffie4, Iris Hoffie4, Jochen Kumlehn4, Arne Bültemeier2,3,
26. A. Scoumanne, S. J. Cho, J. Zhang, X. Chen, Nucleic Acids Res.
39, 213–224 (2011). Johanna Wolf2, Marco Herde5, Claus-Peter Witte5, John C. D’Auria1*, Jakob Franke2,3*
27. P. Ciceri et al., Nat. Chem. Biol. 10, 305–312 (2014).
28. A. N. Kettenbach et al., Sci. Signal. 4, rs5 (2011). The defensive alkaloid gramine not only protects barley and other grasses from insects but also
29. M. A. van Vugt et al., PLOS Biol. 8, e1000287 (2010).
negatively affects their palatability to ruminants. The key gene for gramine formation has remained
30. R. Cuella-Martin et al., Mol. Cell 64, 51–64 (2016).
31. C. Wang et al., Genomics Proteomics Bioinformatics 18, 72–80 elusive, hampering breeding initiatives. In this work, we report that a gene encoding cytochrome
(2020). P450 monooxygenase CYP76M57, which we name AMI synthase (AMIS), enables the production of
32. D. Cimini et al., J. Cell Biol. 153, 517–528 (2001). gramine in Nicotiana benthamiana, Arabidopsis thaliana, and Saccharomyces cerevisiae. We reconstituted
33. T. Davoli et al., Cell 155, 948–962 (2013).
34. P. A. Knobel et al., Mol. Cell. Biol. 34, 2062–2074 gramine production in the gramine-free barley (Hordeum vulgare) variety Golden Promise and
(2014). eliminated it from cultivar Tafeno by Cas-mediated gene editing. In vitro experiments unraveled that
35. K. H. Khoo, C. S. Verma, D. P. Lane, Nat. Rev. Drug Discov. 13, an unexpected cryptic oxidative rearrangement underlies this noncanonical conversion of an amino acid
217–236 (2014).
36. M. Olivier, M. Hollstein, P. Hainaut, Cold Spring Harb.
to a chain-shortened biogenic amine. The discovery of the genetic basis of gramine formation now
Perspect. Biol. 2, a001008 (2010). permits tailor-made optimization of gramine-linked traits in barley by plant breeding.
37. P. Kleiblova et al., J. Cell Biol. 201, 511–521 (2013).
38. A. G. Gilmartin et al., Nat. Chem. Biol. 10, 181–187
(2014).

39. F. Meitinger et al., Nature 585, 440–446 (2020).
40. Z. Y. Yeow et al., Nature 585, 447–452 (2020).
41. X. Qian et al., ACS Med. Chem. Lett. 1, 30–34
(2010).
42. J. Tischer, F. Gergely, J. Cell Biol. 218, 10–11 (2019).
43. C. Dumontet, M. A. Jordan, Nat. Rev. Drug Discov. 9, 790–803
(2010).
44. E. Bernabeu, M. Cagel, E. Lagomarsino, M. Moretton,
D. A. Chiappetta, Int. J. Pharm. 526, 474–495 (2017).
45. L. S. Penna, J. A. P. Henriques, D. Bonatto, Pharmacol. Ther.
173, 67–82 (2017).
46. L. Garribba, S. Santaguida, Front. Cell Dev. Biol. 10, 838928
(2022).
47. R. Li, J. Zhu, Nat. Rev. Mol. Cell Biol. 23, 250–265
(2022).
ACKN OW LEDG MEN TS
We thank M. Ohta and A. Schlientz for feedback on the manuscript,
R. Green for help with the model figure, J. Anzola and D. Jenkins
for technical assistance, M. Kahraman for compound sourcing,
and members of the Oegema and Desai labs for discussion.
Funding: This work was supported by the National Institutes of
Health (grant GM074207 to K.O. and grant GM074215 to A.D.);
the German Research Foundation (grant ME 4713/1-1 to F.M.);
the Ludwig Institute for Cancer Research (salary and other
support for K.O., A.K.S., and A.D.); the Okinawa Institute of
Science and Technology (F.M.); and the Japan Society for the
Promotion of Science (KAKENHI grant 23K05773 to F.M.).
Author contributions: Conceptualization: F.M., A.S., A.D.,
K.O.; Funding: F.M., K.O., A.D.; Investigation: F.M., H.B., R.D.,
A.S., K.O., A.D.; Methodology: F.M., H.B., R.D., M.M.; Resources:
F.M., A.S., A.D., K.O.; Writing – original draft: F.M., K.O., A.D.;
Writing – review & editing: F.M., A.S., R.D., K.O., A.D. Competing
interests: The authors declare no competing interests. Data
and materials availability: All data and materials are available
in the main text or the supplementary materials or upon
request to the corresponding authors. License information:
Copyright © 2024 the authors, some rights reserved; exclusive
licensee American Association for the Advancement of Science.
No claim to original US government works. https://www.science.
org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.add9528
Materials and Methods
Figs. S1 to S13
Tables S1 and S2
References (48–66)
MDAR Reproducibility Checklist
Submitted 14 July 2022; resubmitted 21 May 2023
Accepted 4 February 2024
10.1126/science.add9528
C
ereal crops show substantial vulnerabil-
ity to biological stress because of genetic
uniformity, intensive farming practices,
and a lack of physical barriers such as
woody tissues. Chemical defenses are
therefore essential to endow grasses with pro-
tection from insects, grazing animals, and
pathogens. In grasses of the family Poaceae
(Gramineae), an important defensive molecule
is the allelopathic indole alkaloid gramine (1)
(Fig. 1) (1, 2). Despite its structural simplicity,
gramine has a broad spectrum of activities
against aphids, fungi, viruses, and other plants
(3–5). It is found not only in wild grasses such
as Phalaris spp. (6) but also in barley (Hordeum
vulgare), the world’s fourth most-widely cul-
tivated cereal crop, according to FAOSTAT
Food and Agriculture Data from the Food
and Agriculture Organization of the United
Nations. Although gramine confers benefi-
cial protective properties, it can also negatively
affect its usability as fodder because of its
ruminant antifeedant properties (1, 5, 6). Hence,
manipulating gramine levels in Poaceae through
plant breeding could allow fine-tuning of the
positive and negative traits associated with
1
Department of Molecular Genetics, Leibniz Institute of Plant
Genetics and Crop Plant Research (IPK), Corrensstr. 3,
06466 Seeland OT Gatersleben, Germany. 2Institute of
Botany, Leibniz University Hannover, Herrenhäuser Straße 2,
30419 Hannover, Germany. 3Centre of Biomolecular Drug
Research, Leibniz University Hannover, Schneiderberg 38,
30167 Hannover, Germany. 4Department of Physiology and
Cell Biology, Leibniz Institute of Plant Genetics and Crop
Plant Research (IPK), Corrensstr. 3, 06466 Seeland OT
Gatersleben, Germany. 5Department of Molecular Nutrition
and Biochemistry of Plants, Leibniz University Hannover,
Herrenhäuser Straße 2, 30419 Hannover, Germany.
*Corresponding author. Email: dauria@ipk-gatersleben.de (J.C.D.);
jakob.franke@botanik.uni-hannover.de (J.F.)
†These authors contributed equally to this work.
‡Present address: Department of Natural Product Biosynthesis, Max
Planck Institute for Chemical Ecology, Hans-Knöll-Straße 8, 07745
Jena, Germany.
this alkaloid. Although gramine has been known
for more than 60 years, elucidation of the
complete genetic basis of gramine biosynthe-
sis has not been successful so far, hampering
modern breeding efforts (6).
Early studies using radiolabeled precursors
determined that gramine is derived from the
amino acid tryptophan (2), with aminomethy-
lindole (AMI) (3) being the key intermediate
(Fig. 1) (7, 8). AMI is methylated twice by a
previously described N-methyltransferase
(NMT) to yield gramine via the intermediate
N-methylaminomethylindole (MAMI) (4) (9).
Labeling studies suggested that AMI retains
the nitrogen of the a-amino group as well
as C-3 of tryptophan, whereas C-1 and C-2 are
lost (8, 10, 11). To date, neither the enzymatic
basis nor the mechanism for this highly un-
usual rearrangement are known. It has been
proposed that an enzyme that uses the co-
factor pyridoxal phosphate (PLP) might carry
out this reaction (fig. S1) (11, 12), but no such
enzyme has yet been identified. In this work,
we show that this elusive transformation is
performed not by a PLP-dependent enzyme
but by a cytochrome P450 monooxygenase,
CYP76M57, which we name AMI synthase
(AMIS). Using a combination of bioinformatics
and heterologous expression, we show that the
genes encoding CYP76M57 and NMT form a
gene cluster in barley that correlates with
the presence of gramine and are sufficient to
produce gramine in Nicotiana benthamiana,
Arabidopsis thaliana, baker’s yeast (Saccharomyces
cerevisiae), and a gramine-free barley cultivar.
Knockout of AMIS in a gramine-producing
barley cultivar prevented the production of
gramine (1). In vitro experiments with yeast
microsomes demonstrate how a cryptic oxi-
dative rearrangement converts tryptophan (1)
into AMI (3) via an iminium intermediate. Our
results reveal the genetic and enzymatic basis

1448 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


O
1 OH
2 NH2
3 NH2
Unknown enzyme
N N
H • Loss of C-1 and C-2
Tryptophan • Cleavage of C–C bond Aminomethylindole
(2) • Formation of C–N bond (AMI) (3)
Previous hypothesis [Wenkert (12) / O'Donovan (11)]:
PLP-dependent
Tryptophan enzyme
(2) Pyridoxal phosphate
(PLP)
This work:
CYP76M57
(AMI Synthase)
Tryptophan
(2) Heme, O2, NADPH
Fig. 1. Gramine (1) is a defensive alkaloid from barley (H. vulgare) and other Poaceae members. Gramine (1) conveys desirable as well as undesirable traits for
breeding. It is derived from tryptophan (2) and produced by an uncharacterized enzyme. SAM, S-adenosyl methionine. [Part of the figure was created with BioRender.com]
of an unusual biosynthetic chain-shortening
process of an amino acid that enables further
biotechnological applications of this alkaloid,
which is crucial for defense in cereals.
Results
Discovery of a clustered cytochrome P450
monooxygenase gene
To initiate our search for the missing gene(s)
involved in gramine biosynthesis, we analyzed
the recently released pan-genome of barley
that covers chromosome-scale assemblies of
20 varieties (13). Based on the previous hypoth-
esis proposing a dependence on the cofactor
PLP (fig. S1), our original intention was to select
gene candidates that encode PLP-dependent
enzymes. We noted a previous study that sug-
gested the possibility of gene clustering in
gramine biosynthesis (14). With this scenario
in mind, we analyzed the genomic context
of the previously reported NMT gene (9).
Only one gene encoding a protein typical
for specialized metabolism was found in the
vicinity of NMT, namely a cytochrome P450
monooxygenase gene (CYP) at a distance of
681 kb. The cluster region also includes short
sequences annotated as “transposon protein”
and “nonspecific serine/threonine protein
kinase” (Fig. 2A). By contrast, the closest gene
encoding a PLP-dependent enzyme is located
almost 5 Mb away.
Among the 20 barley lines included in the
pan-genome v1 (13), six do not contain gram-
ine (1) or contain only traces, whereas the
SCIENCE science.org
HvNMT
H SAM
Methylaminomethylindole
(MAMI) (4)
AMI (3)
Barley
AMI (3)
( Hordeum vulgare )
others produce gramine (1) in the range of
700 to 6600 pmol/mg fresh weight (FW) in the
leaves (15) (Fig. 2B). Genome analyses focused
on CYP and NMT revealed that the presence of
both genes is largely consistent with the pro-
duction of gramine. All gramine-producing
varieties contain both CYP and NMT, whereas
Golden Promise and Morex produce neither
gramine (1) nor AMI (3) and contain neither
CYP nor NMT. Apparent exceptions are repre-
sented by the four accessions Du-Li Huang
(ZDM_01467), Igri, Hockett, and Barke. Du-Li
Huang (ZDM_01467) contains a CYP with a
single nonconserved amino acid change [Ser211→
Trp (S211W)]. Lines B1K-04-12 and HOR3081
contain other single nonconserved amino acid
changes, namely Met300→Ile (M300I) and
Met459→Ile (M459I), respectively (table S1);
however, this does not seem to affect gramine
production. For Igri, Hockett, and Barke, the
annotated coding sequences of CYP are 100%
identical to CYP from HOR10350, yet none of
these varieties produce gramine. Further analysis
on the genomic level revealed that the CYP al-
leles in Barke, Hockett, and Igri are all annotated
to include the insertion of an intron, whereas
all other accessions analyzed consist of a single
exon containing the full open reading frame
(fig. S2). According to the genome annotation
(13), the corresponding introns include a long-
terminal repeat retrotransposon (fig. S2). The
presence of a transposon inside the CYP gene
may contribute to altered expression levels be-
cause regions with these elements are often si-
RESE ARCH | R E S E A R C H A R T I C L E S
NH N
HvNMT
N N
H SAM H
Gramine
(1)
+
Desirable factor
Adverse e ects on aphids
and competing plants
Anti-quality factor
Adverse e ects on
grazing ruminants
lenced (16). Taken together, this bioinformatics
analysis indicated a correlation between gram-
ine (1) production and CYP, suggesting that it
might play a key role in gramine biosynthesis.
CYP76M57 is required for gramine biosynthesis
To test if this cytochrome P450 monooxygenase
with the systematic name CYP76M57 is in-
volved in gramine biosynthesis, we transiently
expressed CYP76M57, NMT, and a combina-
tion of both in N. benthamiana (17). The ex-
pression of CYP76M57 led to the production of
a new compound with a retention time, mass,
and ultraviolet (UV) spectrum consistent with
AMI (3) (Fig. 3A). Coexpression of CYP76M57
and NMT resulted in a compound consistent
with gramine (1) based on a comparison of re-
tention time, mass, and UV spectra to a refer-
ence compound (Fig. 3A). To further corroborate
the identity of heterologously produced metab-
olites, we isolated AMI (3) (3.2 mg/g dry weight)
and gramine (1) (6.6 mg/g dry weight) and
confirmed their identity by nuclear magnetic
resonance (NMR) spectroscopy, a technique
used for the structural characterization of
organic compounds, in comparison to refer-
ence compounds (figs. S3 to S6 and tables S2
and S3). This suggested that CYP76M57 con-
verts tryptophan (2) into AMI (3) and conse-
quently completes the biosynthetic pathway of
gramine (1). We therefore renamed CYP76M57
to AMI synthase (AMIS).
Our next goal was to investigate whether
AMIS and NMT enable the reconstitution of
29 MARCH 2024 • VOL 383 ISSUE 6690 1449
RES EARCH | R E S E A R C H A R T I C L E S

gramine biosynthesis in other organisms. We
chose the model plant A. thaliana because of
the ease of the genetic transformation system
that is available. After stable transformation,
selected T1 plants constitutively expressing the
genes AMIS, NMT, or a combination of both
were tested for the presence of AMI (3) and
gramine (1) by reversed-phase ultraperfor-
mance liquid chromatography fluorescence de-
tection (RP-UPLC-FLD) (15). More than 88%
(N = 82) of the herbicide-resistant progeny of
all plants transformed with the AMIS construct
(N = 92) exhibited AMI (3) concentrations
above the limit of quantification. Likewise,
46% (N = 47) of herbicide-resistant plants
transformed with the construct containing
both transgenes (N = 102) produced gramine
(1) in levels above the limit of quantification. If
limits of detection are considered, 93% (N =
86) of the plants transformed with the AMIS
construct produced AMI (3), and 60% (N = 61)
of the plants transformed with the construct
containing both transgenes produced gramine
(1). NMT-transformed plants (N = 43), as well
as control plants (N = 16), did not contain
detectable levels of AMI (3) or gramine (1)
(Fig. 3B).
To show that AMIS and NMT also enable
gramine production in a nonplant organism,
we integrated both genes either individually or
in combination into the genome of S. cerevisiae
(Fig. 3C). The strains further contained genes
encoding Catharanthus roseus cytochrome
P450 reductase (CPR) and cytochrome b5 (CYB5)
(18) to support cytochrome P450 monooxy-
genase activity. Yeast strains expressing AMIS
produced 11.7 mg/liter of AMI (3). When both
AMIS and NMT were expressed, 3.2 mg/liter
of gramine (1) were produced, in addition to
3.8 mg/liter of AMI (3). Another new peak
was observed in this strain, which occurred
only when both AMIS and NMT were expressed.
Based on a synthetic reference compound, we
identified this compound as the previously re-
ported monomethylated intermediate MAMI
(4) (9), produced at a titer of 11.0 mg/liter. The
total alkaloid titer of 18.0 mg/liter in the AMIS
+NMT strain was slightly higher than the AMI
(3) levels in the AMIS strain. With this yeast
system, we also tested the effect of the AMIS
mutant S211W that we noted in the barley cul-
tivar Du-Li Huang (ZDM_01467) (Fig. 2B). In-
deed, this mutation resulted in a complete loss
of AMIS activity in yeast (fig. S7), which ex-
plains why we could not detect AMI (3) or
gramine (1) in this cultivar despite the presence
of the AMIS and NMT genes (Fig. 2B).
To corroborate the native role of AMIS, we
tested it in its host plant barley (H. vulgare).
For this, we introduced the gramine biosynthetic
genes into the gramine-free variety Golden
Promise and produced AMIS knockouts of the
gramine-producing variety Tafeno. To provide
the necessary genetic complementation for
Golden Promise plants to produce AMI (3) and
gramine (1), a construct for the integration of

Fig. 2. The formation of gramine (1) and AMI (3) correlates with a distantly clustered cytochrome P450 monooxygenase gene in barley. (A) Loci of CYP
(CYP76M57, AMIS) and NMT on chromosome 1H of H. vulgare cv. HOR10350. Arrow segments in zoomed-in regions indicate coding regions of exons. (B) Correlation of AMI
(3) and gramine (1) levels with the CYP and NMT gene content across the 20 accessions of the barley pan-genome v.1.0. The published genetic distance between accessions
based on long-terminal repeat retroelements (13) is shown. Accessions with apparent mismatches (underlined) are discussed in the main text. AA, amino acid.

1450 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 3. AMIS (CYP76M57) and NMT are sufficient for gramine (1) bio-
synthesis in plants and baker’s yeast. (A) Transient expression of AMIS and
NMT in N. benthamiana leaves. DAD, diode array detector; TIC, total ion
chromatogram. (B) Metabolite levels in T1 A. thaliana leaves stably expressing
AMIS and NMT. Data are shown as violin plots for control (N = 16), NMT (N = 43),
AMIS (N = 92), and AMIS+NMT (N = 102) plants. (C) Metabolite levels after
genomic expression of AMIS and NMT in baker’s yeast (S. cerevisiae). The bar
plot shows means ±SD and data points of three biological replicates. The diamond
symbol indicates the total alkaloid level [AMI (3) + MAMI (4) + gramine (1)].
(D) Metabolite levels of T0 plants of H. vulgare cv. Golden Promise, a cultivar that does
not produce gramine, transformed with AMIS and NMT. Data are shown as violin
plots for control (N = 8), NMT (N = 35), AMIS (N = 39) and AMIS+NMT (N = 10) Golden
Promise plants. (E) Metabolite levels of three independent transgenic H. vulgare cv.
Tafeno plants carrying a knockout construct for AMIS. Tafeno is a gramine-producing
variety. Data are shown as violin plots for control (N = 5) and knockout (N = 3)
Tafeno plants. In (B), (D), and (E), boxes and whiskers represent interquartile ranges
and minimum and maximum values, respectively, combined with kernel density
plots. [Part of the figure was created with BioRender.com]

HvAMIS (“HvAMIS p6i-d35S”) and one for proach to transfer both constitutive expression and carried out a chemical analysis to inves-
the integration of HvNMT (“HvNMT p6i-d35S”) cassettes into Golden Promise. To confirm tigate the effect of the transgenes (Fig. 3D).
were generated and used together in an transformation events during transgenesis, we Thirty-seven out of the 39 transformants that
Agrobacterium-mediated cotransformation ap- performed polymerase chain reaction (PCR) exclusively carried the AMIS gene produced

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1451

RES EARCH | R E S E A R C H A R T I C L E S

A C
EV O O
O
OH NH2
+ Buffer AMIS OH OH
NH NH O
NH2
Oxidative Hydrolysis
+ D-Trp N OH
1,2 -rearrangement H
O
N N N
H H H
+ L-Trp L-Trp (2) Iminium intermediate 5 AMI (3) Glyoxylic acid (6)

AMIS D F O O
O AMIS
OH
D OH OH
+ Buffer D 2 as substrate NH NH
D
3

AMI (3) Isotopologue Intensity (%)

NH2 NaBH4
N 30 oC N
N
H H H
+ D-Trp 2-d3 Iminium intermediate 5 Trapped intermediate 8
3
AMIS
2-d3 as substrate EV + L-Trp + NaBH4
3 132.1 2
D
+ L-Trp D NH2

AMIS + Buffer
N

TIC (SIM 226.8-229.8, 130.0-133.0)

H
3-d2

AMIS + L-Trp
B E 2
O
***
O
OH AMIS + L-Trp + Microsome
removal
+ NaBH4
6
2

NH2
N AMIS + L-Trp + NaBH4
H
n.s. n.s. Phenylhydrazine 8 2
37 oC,
AMIS 15 min O
EV AMIS Synthetic trapped intermediate
+
+
+ rp
+ Trp

8
Bu
D er
L-
L- (w

OH
-T

Tr /o

+
+ er
+
+
ff

Bu
L-
Bu
L- r

N
Tr
Tr

N
f
ffe

p
p

H
N

GA-phenyl-
AD
PH

hydrazone (7)
)

Fig. 4. AMIS catalyzes a cryptic oxidative rearrangement of L-tryptophan
(2) to iminium intermediate 5 that releases AMI (3) and glyoxylic acid (6).
(A) Representative chromatograms [extracted ion chromatogram (EIC) m/z 130]
of yeast microsome assays with AMIS that confirm L-tryptophan (2) as its
substrate. EV, empty vector. (B) Activity of AMIS depends on L-tryptophan (2) as
the substrate and NADPH. (C) Proposed mechanism for the AMIS reaction.
(D) AMIS reaction with L-tryptophan-2,3,3-d3 (2-d3) demonstrates that no loss
of deuterium label and therefore no oxidation occurs at C-3. M130 represents
the in-source fragment [M−NH3+H]+ of AMI (3). (E) Glyoxylic acid (6) (GA) is
produced as a by-product from L-tryptophan (2) by AMIS. Detection was
carried out after conversion to GA-phenylhydrazone (7) by high-performance
liquid chromatography at 324 nm. (F) Trapping of iminium intermediate 5
with NaBH4 from AMIS enzyme reactions with L-tryptophan (2), which results
in the formation of trapped intermediate 8. The data shown are total ion
chromatograms (TICs) from single-ion monitoring (SIM) mode combining
m/z 227 to 230 and 130 to 133 for maximum sensitivity. All bar plots show
means ±SD and data points of three replicates. *p < 0.05, ***p < 0.001, and
n.s. is not significant by Student’s t test.

AMI (3) in amounts that averaged around
2204.05 pmol/mg FW. We also generated
10 plants that successfully integrated both
AMIS and NMT and produced, on average,
40.22 pmol/mg FW of AMI (3) and 1028.77 pmol/
mg FW of gramine (1).
To generate knockout plants, guide RNA
(gRNA) targets based on the mRNA coding
regions of AMIS were designed and used for
targeted Cas9-mediated mutagenesis in the
gramine-producing cultivar Tafeno. A vector whereas two (brh104AE2 and brh104E3, re-
with three specific gRNAs and a Cas9 coding spectively) were unmutated or carried hetero-
sequence was built using the CasCADE sys- zygous mutations and were therefore excluded
tem (19), and two Agrobacterium-mediated from the chemical analysis reported in Fig. 3E.
transformation experiments (brh104 and brh104A) Homozygous deletions consisted of a 47–base
were conducted. In total, seven plants were pair (bp) deletion between target motifs 3 and
regenerated. Genotyping revealed that five of 2 or a 34-bp deletion between target motifs
them (brh104E1a to brh104E1c, brh104E2, and 1 and 2 (see fig. S8). The 47-bp deletion creates
brh104AE1) were homozygous mutants ascrib- a shift of the reading frame, resulting in a pre-
able to three independent mutation events, mature stop codon and a truncated protein of

1452 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


44 amino acid residues. In addition to the 34-bp
deletion, plant brh104AE1 also carried a biallelic
1-bp insertion at target motif 3, resulting in the
alteration and deletion of a total of 16 amino
acid residues (fig. S8). All transgenic plants
showed a positive signal in the hpt-specific PCR
results. AMI (3) and gramine (1) concentra-
tions in the plants that resulted from the three
independent transformation events using the
knockout construct were below limits of quan-
tification, whereas the wild-type control plants
(N = 5) produced an average amount of
2109.13 pmol/mg FW of gramine (1) (Fig. 3E).
Altogether, the results of targeted knockout of
AMIS in barley cv. Tafeno and the expression
of AMIS and NMT in plants and yeast demon-
strate that the two enzymes AMIS and NMT,
possibly supported by endogenous cytochrome
P450 reductases as redox partners, are suffi-
cient to produce gramine (1).
AMIS catalyzes a rearrangement to an
iminium intermediate
The process of forming AMI (3) from trypto-
phan (2) is mechanistically unusual because
no apparent oxidation takes place that could
explain the necessity of an oxidative enzyme.
We therefore isolated microsomes from yeast
expressing AMIS, CPR, and CYB5 to study the
mechanism of AMIS. Incubation of AMIS with
L-tryptophan (2) resulted in the formation of
AMI (3), whereas no AMI (3) was found in
empty-vector controls (Fig. 4A). In the absence
of L-tryptophan (2), small amounts of AMI (3)
could be detected for microsomes containing
AMIS. We propose that this background AMI
(3) was either copurified during microsome
isolation or formed by activity of AMIS with
copurified L-tryptophan (2) before the start of
the assay. Assays containing D-tryptophan (99%
optical purity) showed no increase in AMI (3)
formation compared with those performed
using controls without any substrate (Fig. 4B).
We next wanted to confirm that AMIS is truly
a cytochrome P450 monooxygenase. When
NADPH (reduced form of nicotinamide ade-
nine dinucleotide phosphate), the common
electron donor for cytochrome P450 mono-
oxygenases, was omitted, the activity was
indistinguishable from that of the negative
control (Fig. 4B). This dependency on electron
transfer was also supported in yeast; yeast
strains that lacked genes for the redox part-
ners C. roseus CPR and CYB5 only produced
44% of AMI (3) compared with strains that
included these genes (fig. S9A). We also per-
formed microsome assays in oxygen-depleted
conditions, which was achieved by flushing
the assay mixtures with nitrogen. Under these
conditions, AMI (3) production dropped to
21% compared with the positive control but
could be restored to 98% by flushing the assay
with pressurized air (fig. S9B). Finally, we pro-
duced a truncated and tagged soluble version
SCIENCE science.org
of AMIS (tAMIS) in N. benthamiana that
lacked amino acids 2 to 23 to eliminate the
N-terminal membrane anchor typical of cyto-
chrome P450 monooxygenases and that con-
tained a hemagglutinin (HA) and StrepII
tag for affinity purification (fig. S10, A and
B). Purified tAMIS contained approximate-
ly equimolar amounts of heme (fig. S10C).
Taken together, our data confirm that AMIS
is, despite its unusual net reaction, a canonical
cytochrome P450 monooxygenase that con-
tains heme and depends on NADPH, redox
partners, and oxygen.
According to our data, AMIS catalyzes an
oxidative reaction. We therefore hypothesized
that the nitrogen of tryptophan is oxidized to
trigger a 1,2-rearrangement of the bond be-
tween C-3 and C-2 toward the nitrogen, lead-
ing to iminium intermediate 5 (Fig. 4C). This
proposal is in line with previous isotope-labeling
experiments in plants that show that the
nitrogen and C-3 of tryptophan (2) are re-
tained in AMI (3), whereas C-1 and C-2 get
eliminated in the form of an unknown mo-
lecular species (10). The oxidation could occur
by multiple possible mechanisms (fig. S11). Hy-
drolysis of iminium intermediate 5 would then
release AMI (3) as well as glyoxylic acid (6).
To verify this proposed mechanism, we per-
formed enzyme assays with isotopically labeled
tryptophan (2). Because AMI (3) shows com-
plete in-source fragmentation under loss of
NH3 (10), we analyzed the isotope composition
of this fragment at a mass/charge ratio (m/z)
of 130. Assays carried out with tryptophan-
2,3,3-d3 (2-d3) showed a clear +2 mass shift
for the fragment of AMI (3) but no increased
+1 signal (Fig. 4D). This is consistent with
our hypothesis and supports the previous mod-
el (11) that no oxidation occurs at C-3.
We next wanted to confirm that glyoxylic
acid (6) is the by-product of AMIS. We used
derivatization with phenylhydrazine to phenyl-
hydrazone adduct 7, which is more readily
detected (fig. S12) (20). Assay analysis was
again complicated by a persistent background
of glyoxylic acid (6) that was copurified during
microsome preparation or produced before
the assay started. Nonetheless, we detected a
statistically significant increase in glyoxylic
acid (6) levels in the presence of AMIS and
tryptophan (2) compared with controls (Fig. 4E).
We also independently quantified glyoxylic
acid (6) levels in N. benthamiana after tran-
sient expression; overexpression of AMIS re-
sulted in a significant increase in glyoxylic acid
(6) levels in planta (fig. S13).
If AMIS oxidizes the amino group of trypto-
phan (2), N-hydroxy-tryptophan (9) could be a
possible intermediate. Alternatively, N-hydroxy-
tryptophan (9) might also occur as a by-product
of AMIS upon premature release of an oxi-
dized but not yet rearranged intermediate. To
test these possibilities, we synthesized 9. Al-
RESE ARCH | R E S E A R C H A R T I C L E S
though N-hydroxy-tryptophan (9) was not stable
in solution, we did not observe spontaneous con-
version of 9 to AMI (3). N-hydroxy-tryptophan
(9) could also not be detected as a by-product
from enzyme reactions (fig. S14). Likewise, en-
zyme assays using N-hydroxy-tryptophan (9)
as a substrate did not lead to increased for-
mation of AMI (3) compared with those using
controls without substrate (fig. S14).
To corroborate that AMIS directly catalyzes
an oxidative rearrangement, we sought to trap
the proposed iminium intermediate with the
reducing agent NaBH4, which would lead to
secondary amine 8 (Fig. 4F). Assays carried
out in the presence of NaBH4 showed a new
peak with a retention time and tandem mass
spectrometry (MS/MS) spectrum consistent
with a synthetic reference of 8 (Fig. 4F and
fig. S15). This trapped intermediate 8 could
not be detected when either NaBH4, AMIS, or
tryptophan (2) were omitted. Also, it was not
detected when NaBH4 was added after termi-
nation of the reaction by removal of microsomal
proteins, supporting that iminium intermediate
5 is a true enzymatic intermediate (Fig. 4F).
When the AMIS reaction was carried out with
L-tryptophan-2,3,3-d3 (2-d3) as a substrate, a
+3 mass shift was observed for trapped inter-
mediate 8, providing further evidence that
AMIS does not carry out an oxidation at C-2
or C-3 (fig. S16). Assays with L-tryptophan-
15
N2 (2-15N2) suggested that both nitrogen
atoms are not separated during the reaction
catalyzed by AMIS, which supports an intra-
molecular reaction (fig. S17). Taken together,
our experiments indicate that iminium inter-
mediate 5 is a true intermediate from the AMIS-
catalyzed cryptic oxidative rearrangement of
L-tryptophan (2).
Discussion
Although the biological relevance of gramine
as a defense alkaloid in barley and other grasses
is well recognized (1, 6), the genetic and enzy-
matic basis of its formation has remained
elusive. This in turn has hindered modern
molecular breeding programs that focus on
sustainable pest management and sustain-
able agriculture. Labeling experiments led to
the hypothesis that an enzyme dependent on
the cofactor PLP is responsible for the forma-
tion of AMI (3) from tryptophan (2) (fig. S1)
(11, 12). This would have also been in line with
the canonical model of alkaloid biosynthesis,
which starts with the conversion of an amino
acid into a biogenic amine by a PLP-dependent
decarboxylase (21). In this work, we reveal that,
unexpectedly, a cryptic oxidative rearrange-
ment plays a key role in gramine biosynthesis.
A single cytochrome P450 monooxygenase,
CYP76M57, which we name AMIS, converts
tryptophan (2) into AMI (3) to complete the
gramine biosynthetic pathway in barley. The
two genes required for gramine biosynthesis
29 MARCH 2024 • VOL 383 ISSUE 6690 1453
RES EARCH | R E S E A R C H A R T I C L E S
form a gene cluster, which confirms the earlier
hypothesis that only two colocalized genes
are required for gramine biosynthesis (14).
The improved availability and quality of plant
genome data has accelerated the discovery of
biosynthetic gene clusters in plants (22, 23).
The large intergenic distance of 681 kb be-
tween AMIS and NMT is unusual, though. It is
possible that this serves a special regulatory
function (24, 25). Future plant genome–mining
approaches should therefore consider the pos-
sibility of distant gene clustering (22).
AMIS belongs to the CYP76 family of cyto-
chrome P450 monooxygenases. More than 30
CYP76s from specialized plant metabolic path-
ways have been characterized (fig. S18), mostly
from terpenoid biosynthesis (26). This includes
CYP76s from the AMIS-containing subfamily
CYP76M as well as from other subfamilies.
Notably, CYP76M7, CYP76M8, and CYP76M14
are involved in the biosynthesis of diterpe-
noids such as momilactone and other phytoa-
lexins in rice (27). Only two characterized
CYP76s accept amino acid–derived substrates—
CYP76AD6 and CYP76AD1 in betalain bio-
synthesis, which oxidize tyrosine and the re-
lated amino acid L-dopa (28). In contrast to
AMIS, these enzymes act on the aromatic part
of their amino acid substrate. The discovery of
AMIS highlights that the substrate and reac-
tion scope of the CYP76 family is broader than
previously described and emphasizes the over-
all plasticity that is well known for cytochrome
P450 monooxygenases in general (29, 30).
The direct rearrangement of an unmodified
amino acid as catalyzed by AMIS is atypical. In
the biosynthesis of glucosinolate-derived phy-
toalexins, superficially similar transformations
occur en route to brassinin and related phy-
toalexins (31, 32). However, these are based on
Lossen-type rearrangements of thiohydroximate-
O-sulfonates to isothiocyanates (33) and re-
quire multiple enzymatic steps to achieve the
required reactivity. Enzymatic modifications
of amino acids also occur in pathways of nat-
ural products that contain nonproteinogenic
a- or b-amino acids (34, 35). However, none of
these enzymes performs a rearrangement like
AMIS. Conceptually, the AMIS reaction resem-
bles the Stieglitz rearrangement, which re-
quires prior activation of the nitrogen and
typically involves aromatic substituents (36).
An oxidative variant of the Stieglitz rearrange-
ment for primary amines has been reported in
chemical synthesis (37). AMIS now provides a
direct enzymatic connection of an amino acid
to a chain-shortened biogenic amine; this reac-
tivity enabled by an unusual cytochrome P450
monooxygenase–mediated cryptic oxidation lies
outside of the scope of PLP-dependent amino
acid decarboxylases, the canonical entry point
into alkaloid biosynthesis (21). As such, our dis-
covery will facilitate new avenues in synthetic
biology to generate non-natural amines and
1454 29 MARCH 2024 • VOL 383 ISSUE 6690
alkaloids that might possess desirable biologic-
al activities.
With the description of the structural genes
for gramine biosynthesis complete, it will now
be possible to undertake targeted breeding
efforts aimed at modulating gramine levels
or to introduce this metabolite into new host
plants by incorporating its biosynthetic path-
way (38). These approaches will be tailored
to the myriad biological interactions in which
gramine participates. Compared with many
other biosynthetic pathways for metabolites
of relevance to crop protection, which often
require 10 or more different genes (39), gram-
ine production is achieved by only two genes.
Hence, reaching the ability to genetically engi-
neer organisms to produce gramine seems an
achievable goal.
RE FERENCES AND NOTES
1. M. Vicari, D. R. Bazely, Trends Ecol. Evol. 8, 137–141
(1993).
2. J. V. Lovett, A. H. C. Hoult, in Allelopathy, vol. 582 of ACS
Symposium Series, Inderjit, K. M. M. Dakshani, F. A. Einhelllig,
Eds. (American Chemical Society, 1994), pp. 170–183.
3. L. J. Corcuera, Phytochemistry 23, 539–541 (1984).
4. A. Lu et al., J. Agric. Food Chem. 67, 2148–2156
(2019).
5. K. S. Brown Jr., J. R. Trigo, in vol. 47 of The Alkaloids:
Chemistry and Pharmacology, G. A. Cordell, Ed. (Academic
Press, 1995), pp. 227–354.
6. A. D. Hanson, P. L. Traynor, K. M. Ditz, D. A. Reicosky, Crop Sci.
21, 726–730 (1981).
7. E. Leete, L. Marion, Can. J. Chem. 31, 1195–1202
(1953).
8. B. G. Gower, E. Leete, J. Am. Chem. Soc. 85, 3683–3685
(1963).
9. K. A. E. Larsson, I. Zetterlund, G. Delp, L. M. V. Jonsson,
Phytochemistry 67, 2002–2008 (2006).
10. E. Ishikawa, S. Kanai, M. Sue, Biochem. Biophys. Rep. 34,
101439 (2023).
11. D. O’Donovan, E. Leete, J. Am. Chem. Soc. 85, 461–463
(1963).
12. E. Wenkert, J. Am. Chem. Soc. 84, 98–102 (1962).
13. M. Jayakodi et al., Nature 588, 284–289 (2020).
14. T. J. Leland, R. Grumet, A. D. Hanson, Plant Sci. 42, 77–82
(1985).
15. S. Leite Dias et al., Plants 12, 1930 (2023).
16. B. Liu, M. Zhao, Curr. Opin. Plant Biol. 75, 102428 (2023).
17. L. Chuang, J. Franke, in Engineering Natural Product
Biosynthesis: Methods and Protocols, vol. 2489 of Methods
in Molecular Biology, E. Skellam, Ed. (Springer, 2022),
pp. 395–420.
18. S. Brown, M. Clastre, V. Courdavault, S. E. O’Connor, Proc. Natl.
Acad. Sci. U.S.A. 112, 3205–3210 (2015).
19. I. V. O. Hoffie, thesis, Institutionelles Repositorium der Leibniz
Universität Hannover (2022).
20. M. Lange, M. Mályusz, J. Chromatogr. B Biomed. Appl. 662,
97–102 (1994).
21. B. R. Lichman, Nat. Prod. Rep. 38, 103–129 (2021).
22. G. Polturak, Z. Liu, A. Osbourn, Curr. Opin. Green Sustain. Chem.
33, 100568 (2022).
23. S. J. Smit, B. R. Lichman, Nat. Prod. Rep. 39, 1465–1482
(2022).
24. J. Colinas, S. C. Schmidler, G. Bohrer, B. Iordanov, P. N. Benfey,
PLOS ONE 3, e3670 (2008).
25. R. M. Clark, T. N. Wagler, P. Quijada, J. Doebley, Nat. Genet. 38,
594–597 (2006).
26. U. Bathe, A. Tissier, Phytochemistry 161, 149–162 (2019).
27. N. Kitaoka et al., Plant Cell 33, 290–305 (2021).
28. G. Polturak et al., New Phytol. 210, 269–283 (2016).
29. C. C. Hansen, D. R. Nelson, B. L. Møller, D. Werck-Reichhart,
Mol. Plant 14, 1244–1265 (2021).
30. K. Malhotra, J. Franke, Beilstein J. Org. Chem. 18, 1289–1310
(2022).
31. M. S. C. Pedras, E. E. Yaya, Org. Biomol. Chem. 11, 1149–1166
(2013).
32. A. P. Klein, E. S. Sattely, Proc. Natl. Acad. Sci. U.S.A. 114,
1910–1915 (2017).
33. F. S. Hanschen, E. Lamy, M. Schreiner, S. Rohn, Angew. Chem.
Int. Ed. 53, 11430–11450 (2014).
34. F. Kudo, A. Miyanaga, T. Eguchi, Nat. Prod. Rep. 31, 1056–1073
(2014).
35. J. B. Hedges, K. S. Ryan, Chem. Rev. 120, 3161–3209
(2020).
36. Z. Wang, in Comprehensive Organic Name Reactions and
Reagents, Z. Wang, Ed. (Wiley, 2010), pp. 2673–2676.
37. W. Yamakoshi, M. Arisawa, K. Murai, Org. Lett. 21, 3023–3027
(2019).
38. J. Jirschitzka, D. J. Mattern, J. Gershenzon, J. C. D’Auria,
Curr. Opin. Biotechnol. 24, 320–328 (2013).
39. J. Zhang et al., Nature 609, 341–347 (2022).
AC KNOWLED GME NTS
We thank D. Nelson (Department of Molecular Science, University
of Tennessee, Memphis) and the P450 nomenclature committee
for naming CYP76M57. The group of J.F. thanks S. Krause,
G. Birkenbach, K. Körner, Y. Leye, and M. Fent for excellent
technical and horticultural support and M. Niehaus and L. Fischer
for helpful discussions. J.F. thanks C. Hertweck (Leibniz Institute
for Natural Product Research and Infection Biology, HKI) and
S. O’Connor (Max Planck Institute for Chemical Ecology) for helpful
discussions. The group of J.C.D. thanks E. Brueckner and
M. Gerres as integral members of the technical staff of the
Leibniz Institute of Plant Genetics and Crop Plant Research (IPK)
as well as the IPK gardener team. J.K. and
R.E.H. thank S. Sommerfeld and A. Knospe for excellent technical
assistance with barley transformation. The EasyClone-MarkerFree
Vector Set was a gift from I. Borodina (Addgene kit no.
1000000098). Parts of figs. S7 and S9 were created with
BioRender.com. Funding: J.F. acknowledges financial support from
the SMART BIOTECS alliance between the Technische Universität
Braunschweig and the Leibniz Universität Hannover, which is
supported by the Ministry of Science and Culture (MWK) of Lower
Saxony. Moreover, we thank the International Max Planck Research
School for supporting S.L.D. and the Leibniz Research Alliance
“Bioactive Compounds and Biotechnology” for the “GraB-ME”
seed-money grant to J.C.D. This work was supported by the
Deutsche Forschungsgemeinschaft (DFG) INST 187/741-1 FUGG
to C.-P.W. Author contributions: S.L.D., L.C., J.C.D., and
J.F. conceived the project, designed the experiments, analyzed
the data, and wrote the manuscript. S.L.D. generated the vectors
for A. thaliana transformation, carried out the transformation,
optimized the RP-UPLC-FLD method for metabolite measurement,
and performed the chemical analysis of the Arabidopsis, Golden
promise, and Tafeno plants. L.C. and J.W. performed transient
expression in N. benthamiana. L.C. and B.S. designed and
conducted enzyme assays and performed yeast microsome
purifications. S.L. designed isotope-labeling experiments,
synthesized and purified compounds, optimized analytical
conditions, and performed NMR analysis. B.S. designed and
performed the yeast metabolic engineering experiments. F.L.B.
analyzed genome data. F.L.B., B.G.C., R.E.H., and I.H. generated
vectors for gramine overexpression and knockout in barley.
R.E.H. carried out barley transformations. J.K. designed and
supervised barley transformation experiments. A.B., M.H., and
C.-P.W. designed biochemical experiments. A.B. carried out
biochemical experiments and heme quantification. M.H. and
C.-P.W. assisted with MS/MS and high-resolution MS
measurements. Competing interests: The authors declare
that they have no competing interests. Data and materials
availability: All data are available in the main text or the
supplementary materials. The coding sequence of AMIS has
been deposited in GenBank under accession number OR461264.
License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adk6112
Materials and Methods
Figs. S1 to S60
Tables S1 to S10
References (40–60)
MDAR Reproducibility Checklist
Submitted 7 September 2023; accepted 26 February 2024
10.1126/science.adk6112
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

HYDROGELS alkyl-water contacts and maximize favorable

polyions-water contacts (Fig. 1B) (14–18). We
A hyperelastic hydrogel with an ultralarge reversible synthesized AETC-25 hydrogels using poly[2-
(acryloyloxy)ethyl]trimethylammonium chloride
biaxial strain (PAETC) as building blocks (movie S1). The
heterogeneously hydrated polycations under
Lili Chen, Zhekai Jin, Wenwen Feng, Lin Sun, Hao Xu, Chao Wang* water-scarce conditions were confirmed using
density functional theory (DFT) calculations
Hyperelastic materials exhibit a nonlinear elastic response to large strains, whereas hydrogels typically (fig. S1 and supplementary text). To verify the
possess a low elastic range due to the nonuniform cross-linking and limited chain segments among PNC structure, we conducted atomic force mi-
cross-links. We developed a hyperelastic hydrogel that possesses a broader elastic range by introducing croscope (AFM) tests to observe the conforma-
a reversible pearl-necklace structure, in which beads are connected by strings. The subnanometric tion of a single PAETC chain deposited on a
beads can efficiently unfold and refold under cyclic mechanical strains; thus, the hydrogel can rapidly mica sheet after the evaporation of water. As
recover after being stretched to an areal strain of more than 10,000%. Additionally, the hydrogel can depicted in Fig. 1C, the polycations existed in
quickly heal from minor mechanical damages such as needle punctures and cuts. These advancements a PNC structure rather than the coiled con-
make our ionic hydrogels ideal for multifunctional pneumatic gripper materials; they simultaneously formations that were observed in neutral poly-
offer an ultralarge gripping range, self-sensing capabilities, and fast healing abilities. mer chains under water-scarce conditions
(figs. S2 and S3). The AFM height image at
the z dimension revealed bead sizes less than
ydrogels consist of cross-linked polymer of flexible and weakly hydrophobic polyelec- 2 nm. Cryogenic transmission electron mi-

H networks swelled in water molecules.

The elasticity of hydrogels stems from
the reversible conformation changes of
polymer chains among cross-links. The
maximum elastic deformation (lmax) of poly-
mer chains that contain N links of equal length
can be described as lmax ~ N1/2 (supplementary
text), which represents the ratio of the mean
square end-to-end distance of a fully extended
chain to that of a freely jointed chain (1). Owing
to the nonuniform cross-linking and the lim-
ited value of N (number of chain segments
among cross-links) within a swelling hydrogel
network, the lmax for traditional hydrogels
was limited to less than 800% (2–8). Ito et al.
used an unfixed slide-ring cross-linking strat-
egy to effectively increase the value of N during
the stretching process (9). By incorporating
slidable rings on the chains, the material was
capable of additional stretch by drawing from
transverse directions, resulting in a lmax of
1100%. However, lmax still follows an ex-
ponent of 1/2 with respect to N in existing
hydrogels. Furthermore, biaxial deformation
remains challenging because the chains ex-
tend along two directions; covalent cross-links
or unfixed slide-ring cross-links greatly limit
the stretchability because of biaxial deforma-
tion locking (1). The maximum linear strain
under equal biaxial tension is at least 30%
lower than that observed under uniaxial ten-
sion (10). Achieving superb biaxial reversibil-
ity faces even greater challenges, the reversible
biaxial areal strain in existing hydrogels is, at
present, limited to less than 2500% (11–13).
We achieved an ultralarge reversible biaxial
strain by incorporating pearl-necklace chains
(PNCs) into polymer networks. This was ac-
complished through heterogeneous hydration
Key Lab of Organic Optoelectronics and Molecular
Engineering, Department of Chemistry, Tsinghua University,
Beijing, China.
*Corresponding author. Email: chaowangthu@tsinghua.edu.cn
trolytes in water-scarce conditions. In contrast
to the freely jointed chain model commonly
used in hydrogels and rubbers, PNCs exhibit
a higher degree of shrinkage and allow for
the release of additional chain segments by
unfolding numerous subnanometric beads.
When the applied force is released, these
extended chains promptly refold back into
the initial PNC structure, which results in
hyperelasticity.
Hydrogels with pearl-necklace structures
The maximum deformation of a single chain is
determined by the ratio of the mean square
end-to-end distance of a fully extended chain
to that of its initial conformation. The maxi-
mum deformations of freely jointed chains
(FJCs) and coiled chains (CCs) approximately
follow the relationships lFJC ~ Ν 1/2 and lCC ~
Ν 2/3, respectively (supplementary text) (14, 15).
PNCs are chain conformations that are com-
posed of small, charged beads connected by
long and narrow strings, which were first
proposed in dilute polyelectrolyte solutions
30 years ago (supplementary text) (14–18). The
maximum deformation of PNCs falls between
that of CCs and FJCs, with Ν 1/2 < lPNC < Ν 2/3.
The monomer fraction in the beads is approx-
imately constant and close to 0.8 (14), which
indicates that lPNC tends to be closer to Ν 2/3
(Fig. 1A). Compared with typical hydrogels
that follow lFJC ~ Ν 1/2, hydrogels that contain
PNCs offer a higher potential for larger elastic
deformations. We selected polyelectrolytes with
strongly hydrophilic ionic groups and weakly
hydrophobic flexible alkyl chains, irrespective of
polycations or polyanions. When there is an in-
sufficient amount of water to fully hydrate poly-
mer chains, hydrophobic interactions between
alkyl chains compete against the electrostatic
interactions between charged groups. Conse-
quently, flexible polymer chains collapse into
poorly hydrated and soft beads connected by
highly hydrated strings to minimize unfavorable
croscopy (cryo-TEM) tests were performed on
our ultrathin AETC-25 hydrogel, the results
of which indicated a dense packing of PNC
structures, with most bead sizes less than 2 nm
(Fig. 1D and fig. S4). In comparison, no beads
were observed in traditional acrylamide hydro-
gels (fig. S5). Two-dimensional small-angle x-ray
scattering (2D SAXS) further confirmed the
PNC structure; the results showed a circular
scattering ring that demonstrates an isotropic
network structure (Fig. 1E) (19). Azimuthally
integrated 1D SAXS profile from 3 to 5.5 nm−1
(Fig. 1F) exhibited a wide peak, corresponding
to distances between neighboring beads of
1.14 to 2.09 nm, which indicates the presence
of subnanometric beads in large quantities.
Highly reversible pearl-necklace structure
The beads within the PNC structure can under-
go sequential unwinding when exposed to
external strains (20). Through in situ SAXS,
we witnessed the highly reversible unfolding
and refolding of PNC structures. Figure 2
presents the in situ SAXS images and profiles
of AETC-25 hydrogels at various strains. In
the undeformed state, only a circular scattering
ring in the 2D SAXS image and a broad peak in
the 1D SAXS profile corresponding to the initial
distances between neighboring subnanomet-
ric beads (1.37 nm) were observed. As the hy-
drogel was stretched, new scattering rings
and new peaks emerged in the low q region
(shown by the arrow position), and the ratio
of these new peaks gradually increased. This
indicates that the distances between beads
expanded because of the stretching of more
polymer strands out of the beads. The infor-
mation gathered from the SAXS peaks aligned
well with the theoretical values corresponding
to the applied strains. The maximum distance
between neighboring beads is calculated as
2p/qn, where qn represents the position of the
newest peak in the low q region. The maxi-
mum strain between neighboring beads is then

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1455

RES EARCH | R E S E A R C H A R T I C L E S

A B

Polycations or Polyanions
N 1 3a

Pearl-necklace chain Coiled chain

Heterogeneous hydration

N1 2 max N2 3 max Na / N 1 3 a N2 3

Na
Fully extended chain

max Na / N 1 2 a N1 2 Bead

String

N 1 2a
Freely jointed chain Pearl-necklace networks
C
3.0 nm

3.0 nm

1.5

-1.5

-3.0
0
0 100 200 300 400 500 nm

D E F
Intensity

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
100 150 200 250 300 350 400 450
-1
Intensity q (nm )

Fig. 1. Hydrogels with pearl-necklace structures. (A) Models of PNC, increasing in magnification from left to right. The dashed circle indicates
CC, and FJC structures. Here, a is the diameter of one globule, and N is the magnified region. (D) Cryo-TEM images of the AETC-25 hydrogels.
the number of globules in each chain model. (B) Hydrogel networks (E) 2D SAXS image of the AETC-25 hydrogels. (F) 1D SAXS profile of the
containing PNCs. (C) AFM images of PAETC deposited on a mica sheet at AETC-25 hydrogels.

defined as the maximum distance divided by
the initial distance minus one (table S1). At a
low applied strain (200%), the initially ran-
domly oriented PNCs gradually aligned along
the applied force, with only a few polymer
strands pulled out of the beads. Thus, the dis-
tances between beads changed little, causing
the calculated strain (42%) to deviate from the
applied strain (200%). With a larger stretch, a
substantial number of polymer strands were
drawn out of the beads, resulting in less
deviation from the applied strain. When our
AETC-25 hydrogel was released from a uni-
axial strain of 1500%, all the new scattering
rings and new peaks disappeared, and the

1456 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

400% 800% 1000% 1200% Intensity

400

200

100

Initial 1000%
60
200% 1200%
400% 1500%
40
200% 800% Return to initial 1500%
Intensity

20

10

0.2 0.4 0.6 0.8 1 2 3 4 5

-1
q (nm )

Stretched to
Initial 1500%
Return to initial

Fig. 2. Highly reversible pearl-necklace structure characterized by in situ
SAXS. New circular scattering rings in 2D SAXS images and new peaks in
1D SAXS profiles appeared with an increase in the applied strain, which showed
larger mean distances between beads due to the unfolding of PNC structures
(clockwise from bottom left to bottom right). These signals disappeared when
the applied strain of 1500% was released, which confirms the high reversibility
of the PNC structure over a large elastic range. The white double arrows
indicate the stretching direction. The center graph shows 1D SAXS profiles of
AETC-25 hydrogels at different applied strains. The photos at the bottom show
the AETC-25 hydrogels in situ stretched to 1500%. [Photos by the authors]

SAXS data of the recovered hydrogel closely
resembled that of the undeformed hydrogel.
This confirmed the excellent reversibility of the
PNC structure even at a cyclic strain of 1500%
(fig. S6). These findings suggest that under
large uniaxial or biaxial strains, apart from the
straightening of polymer chains, subnanometric
beads can unfold to elongate the polymer
strands between rigid interchain entanglements
(2, 21, 22). Upon release, facilitated by the lub-
rication of water molecules, the extended chains
quickly refold to their initial PNC structure,
owing to the low-energy state of the PNC
structure and the driving force of entropy.
Hyperelasticity
Our AETC-25 hydrogel demonstrates excep-
tional properties of high uniaxial stretchability
(3000 ± 350%) and ultralarge biaxial elasticity
without chemical cross-linkers (Fig. 3A and
fig. S7). These characteristics are attributed to
the reversible PNC structures lubricated by
water molecules and appropriate interchain
physical entanglements (figs. S8 and S9). Nu-
clear magnetic resonance (NMR), static light
scattering (SLS), and differential scanning calo-
rimetry (DSC) confirmed the chemical struc-
ture and properties of the AETC-25 hydrogel.
The monomer conversion was about 92%
through NMR (fig. S10), the weight average
molecular weight was estimated to be 3.895 ×
106 g mol−1 by SLS (fig. S11), and the glass
transition temperature (Tg) was −34.2°C from
DSC (fig. S12). Rheological tests demonstrated
that the AETC-25 hydrogel exhibits superb
elasticity with a broad frequency range from 0.1
to 100 rad s−1 in the rubbery plateau zone (fig.
S13). The hyperelastic hydrogels showed a weak
stretch rate dependence and fully recovered
from a strain of 1500% to the initial state
within seconds, displaying excellent reversi-
bility (Fig. 3, B to D, and movie S2). The
reversibility (U2/U1) is defined as the ratio of
the area covered by the second cycle loading
curve (U2) to that covered by the first cycle
(U1) on the nominal stress-strain curves. In
comparison to commercial rubbers, which ex-
hibit a U2/U1 less than 80% in 1 min at cyclic
strains below 800%, our AETC-25 hydrogels
surpassed expectations. At a uniaxial strain of
1500%, they achieved a U2/U1 value of 98.9%
and maintained reversibility above 98% in
subsequent cycles, outperforming most soft
hydrogels and rubbers (Fig. 3E and figs. S14 and
S15) (5, 9, 23–25). Fifty cyclic curves of AETC-25
hydrogels at a cyclic strain of 500% showed
substantial overlap, which indicates high consist-
ency and repeatability (fig. S16). The compres-
sibility of the AETC-25 hydrogels was also
impressive, with a weak compression rate de-
pendence (Fig. 3F) and superb cyclic perform-
ance. Even after 1000 cycles of 67% compressions
with a relaxation time of 1 s, the compression

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1457

RES EARCH | R E S E A R C H A R T I C L E S

A B C 3.0
12 14
20 wt% 100 mm min-1 1st
25 wt% 12 2.5
200 mm min-1 2nd
30 wt%
300 mm min-1 3rd
(MPa) 9

(MPa)

(MPa)
35 wt% 10 2.0
400 mm min-1 4th
8 500 mm min-1
6 1.5
true

true

true
6
1.0
3 4
0.5
2
0 0
0
0 1000 2000 3000 4000 5000 0 500 1000 1500 2000 2500 3000 0 400 800 1200 1600
D Strain (%) Strain (%) Strain (%)

Before Stretch Fully recover

30 cm 30 cm

4.8 m (Strain = 1500%)

E F G
Compression stress (MPa)

Compression stress (MPa)

Cyclic tensile strain (%)

This work 0.30 0.30

1500 100 mm min -1 1st 600th
0.25 200 mm min-1 0.25 100th 700th
1200 300 mm min-1 200th 800th
0.20
SR-0.3 (9) 400 mm min-1 0.20 300th 900th
900 500 mm min -1 400th 1000th
Ecoflex 00-30
0 0.15 0.15
Vulcanized NR
500th
600 Hybrid (5) Silicon
con
0.10
(LD-60)
rubberr
rub 0.10
PAM-CS-A (23) ((RTV-2)
(RT
TV
300 0.05 0.05
MM-0.2-6 (25) PVA-PAAm (24)
Tetra-0.33 (9)
0 0 0
0 20 40 60 80 100 0 20 40 60 0 20 40 60
U2/U1 (%) Compression strain (%) Compression strain (%)
H

Biaxial areal strain

S0 ~ 100 S0
Recover in
seconds

I J Linear strain (%)

1377
~ 1000 V0
V0 Inflate Deflate

Recover in seconds

Fig. 3. Hyperelasticity. All of the mechanical curves are based on data U2/U1 for different soft hydrogels and rubbers. The relaxation time was
from more than three independent trials. (A) True stress-strain curves of no more than 1 min. (F) Compression tests of the AETC-25 hydrogels at
the AETC hydrogels with different water content. strue, true stress. (B) True different loading rates. (G) Cyclic compression stress-strain curves of
stress-strain curves of the AETC-25 hydrogels at different loading rates. AETC-25 hydrogels at 67% compression strain. (H) Photo series of an
(C) Cyclic true stress-strain curves of AETC-25 hydrogels at a strain of AETC-25 cylindrical hydrogel (diameter of 2 cm) after stretching to
1500% for four cycles. The four curves almost fully overlap. (D) A photo an areal strain of 10,000% and recovery to its initial state (S0) in seconds.
of AETC-25 hydrogels stretched to 1500% and recovering in seconds. The (I) On the left is a photo series of a cuboid pneumatic device with a
hydrogel section on the left shows the hydrogel before stretching. The cavity of 0.54 ml after inflation with 550 ml of air, deflation, and recovery
hydrogel section on the right shows the hydrogel after stretching and to its initial state in seconds. V0, initial volume. (J) Finite element simulations
recovery to the initial state. (E) Ashby diagrams of cyclic tensile strain versus of the balloon-like cuboid pneumatic device. [Photos by the authors]

1458 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

A Healed wounds
~ 100 V0 Healing
V0 Needle in seconds
Inflate puncture Reinflate Zoom in

B C
12
Initial V1
10 2 weeks Bounce back
1 month
(MPa)

8 1 month at -20°C Inflate

6 The hydrogel balloon bearing
the weight of 50 kg
true

4 ~ 100 V1
2

0
0 500 1000 1500 2000 2500 3000 3500
Strain (%)

D 200 E 10000 F 1600

10 ml
Initial 20 ml 1400
1-s healing 8000 40 ml
150 60 ml 1200

R/R (%)
Z" (ohm)

80 ml
R (ohm)

6000 1000
100 ml
100 150 ml 800
200 ml
4000 600
50 400
2000
200
0 0 0
0 50 100 150 200 0 50 100 150 200 250 0 100 200 300 400 500

Z' (ohm) Time (s) V (ml)

G H I 1200
40 ml

900

R/R (%)
600

1-s healing Inflate Deflate 300

Grab Release
Pierce through Grab a small strawberry
0
0 5 10 15 20 25 30 35
J K Time (s)
1200
200 ml

900
R/R (%)

600
Release

Inflate Deflate
300

Grab a large ball Grab

0
0 5 10 15 20 25 30 35
Time (s)

Fig. 4. Hyperelastic, minor damage–insensitive, and fast-healing pneu-
matic grippers. (A) Photo series of a cuboid pneumatic device that is pierced
from the top, heals, and inflates again. (B) Stable mechanical properties of
AETC-25 hydrogels after being placed in air (RT = 25°C, RH = 40%) or in
the refrigerator at −20°C for different periods of time. The mechanical curves are
based on data from more than three independent trials. (C) Photo series of
a cuboid pneumatic device with a cavity of 27 ml (V1) after inflation, application
of a 50-kg weight. (D) Ionic conductivity of the AETC-25 hydrogels. Z′, the
real part of the impedance; Z″, the imaginary part of the impedance.
(E) Resistance (R)–time curves for various volumes of inflation and deflation of
grippers. (F) Resistance–volume (V) scatter diagram for grippers (data are
mean ±SD, N = 3). DR/R, change in resistance. (G) Size comparison of a
strawberry and the model of Saturn. (H) Photo series of pierced grippers
grabbing the strawberry when inflated with 40 ml of air. (I) Resistance-time
curves for grippers grabbing the strawberry. (J) Photo series of pierced
grippers grabbing the model of Saturn when inflated with 200 ml of air.
(K) Resistance-time curves for grippers grabbing the model of Saturn.
[Photos by the authors]

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1459

RES EARCH | R E S E A R C H A R T I C L E S
curves nearly overlapped (Fig. 3G). Furthermore,
the hydrogels exhibited exceptional resistance
to large deformations without breakage or
water extrusion (figs. S17 and S18). They could
withstand the weight of 4 metric tons and
rapidly recover to their initial states (movie
S3). Moreover, our AETC-25 hydrogels dem-
onstrated ultralarge biaxial elasticity. The
cylindrical hydrogel with a diameter of 2 cm
was biaxially stretched to an areal strain of
10,000% and quickly recovered to its initial
area, almost without any residual strain (Fig.
3H and movie S4). A cuboid pneumatic device
with a 0.54-ml cavity made of AETC-25 hydro-
gels expanded to about 1000 times its original
volume when inflated with air and then ra-
pidly recovered upon deflation (Fig. 3I, fig.
S19, and movie S5). In comparison, other mate-
rials like Ecoflex 00-30 and acrylamide hy-
drogels exhibited poor biaxial elasticity, with
a volume expansion no more than 200 times
their original volume (fig. S20 and movie S6).
We performed finite element simulations
using a polynomial isotropic hyperelastic mod-
el to describe the strain distribution of the
balloon-like cuboid pneumatic device, reveal-
ing a maximum linear strain of 1377%, which
is equivalent to a maximum areal strain of
more than 10,000% (Fig. 3J).
Hyperelastic, minor damage–insensitive, and
self-sensing pneumatic grippers
Our AETC-25 hydrogels possess desirable prop-
erties, including insensitivity to minor damage
and self-sensing capabilities, which make them
highly suitable for applications in soft electron-
ics and soft robotics, such as pneumatic grip-
pers. The AETC-25 hydrogels exhibited fast
healing abilities due to the electrostatic inter-
actions and water molecules with high mobility
(fig. S21 and table S2). When the hydrogels
were cut into halves, brought back into contact,
and pressed for 1 min, they could be stretched
to 200% and even up to 600% after 5 min of
healing (fig. S22). The AETC-25 hydrogels
showed even faster healing capabilities for
minor damage (fig. S23 and movie S7). When
the hydrogel balloon was penetrated at its
thinnest top area, it initially shrunk back to
its original volume. However, after a few sec-
onds of healing, the cuboid pneumatic device
was reinflated without any leakage, which dem-
onstrates the effective healing capabilities of
AETC-25 hydrogels (Fig. 4A and movie S8).
Moreover, when we used a sharp knife to
penetrate the cuboid pneumatic device, despite
this puncture, it still endured a volume expan-
sion of nearly 1000 times its initial volume
without any failure or leakage (fig. S24). Fast
self-healing plays a critical role in pneumatic
grippers because the potential gas leakage re-
sulting from punctures presents an important
limitation in practical applications. Maintain-
ing the integrity and functionality of the grip-
1460 29 MARCH 2024 • VOL 383 ISSUE 6690
pers even when punctured is essential for their
successful operation. Furthermore, our AETC-
25 hydrogels demonstrated high stability under
ambient conditions (table S3). Unlike tradi-
tional hydrogels that dried out when exposed
to air, our hydrogels maintained their high
stretchability and hyperelasticity even after
1 month of storage at a room temperature (RT)
of 25°C and a relative humidity (RH) of 40%,
benefitting from the balance between water
evaporation and water absorption (Fig. 4B and
figs. S25 to S27). Moreover, the hydrogel bal-
loon exhibited substantial impact resistance
(movie S9). The hydrogel balloon at 10,000%
volume expansion, with a minimal thickness
of approximately 50 mm, could even sustain
the impact of a human weight of 50 kg and
immediately recover once the weight was re-
moved (Fig. 4C, figs. S28 to S30, and movie
S10). Our hydrogels exhibited a high ionic
conductivity of 4.9 mS cm−1. Even when they
were cut into halves and subjected to a mere
1-s healing process, the hydrogels retained
their ionic conductivity (Fig. 4D). By combin-
ing hyperelasticity, fast healing properties, and
ionic conductivity, we created multifunctional
pneumatic grippers (fig. S31 and supplemen-
tary text). These grippers demonstrated a long-
term service life, mechanical stability, and
electrical resistance stability when inflated
with either 10 or 500 ml of air (RT = 25°C, RH =
40%) (Fig. 4, E and F; fig. S32; and movie S11).
They could handle objects with different cur-
vatures and volumes, such as delicate straw-
berries, as well as much larger objects, such
as a model of Saturn, despite the more than
100 times volume difference between these
objects (Fig. 4G). The grippers could even heal
from needle penetrations within 1 s and imme-
diately grab the objects, with resistance-time
curves accurately recording the grab-release
process (Fig. 4, H to K). The grippers could fur-
ther expand to three to four times their initial
length with continuous inflation (fig. S33). Our
multifunctional grippers offer advantages in
intrinsic self-sensing of volume changes, insen-
sitivity to minor damage, and ultralarge gripp-
ing range, thereby surpassing the capabilities of
previous soft pneumatic grippers (supplemen-
tary text) (26–30).
Conclusions
We achieved a hyperelastic hydrogel with an
ultralarge reversible biaxial strain by incor-
porating reversible PNCs into the polymer net-
works. Our hydrogel demonstrates a broader
elastic range compared with that of traditional
hydrogels, owing to the efficient unfolding and
refolding of subnanometric beads under cyclic
mechanical strains, and exhibits notable prop-
erties, such as rapid recovery after extreme
biaxial stretching (a reversible areal strain of
10,000%) and quick healing from minor mech-
anical damage. It also holds potential for use in
pneumatic materials with self-sensing capa-
bilities. These advancements hold great prom-
ise for next-generation soft pneumatic robots
that require hyperelasticity, fast healing, and
self-sensing.
REFERENCES AND NOTES
1. E. M. Arruda, M. C. Boyce, J. Mech. Phys. Solids 41, 389–412
(1993).
2. J. Kim, G. Zhang, M. Shi, Z. Suo, Science 374, 212–216
(2021).
3. X. Meng et al., Adv. Mater. 34, 2108243 (2022).
4. H. Lei et al., Nat. Commun. 11, 4032 (2020).
5. J.-Y. Sun et al., Nature 489, 133–136 (2012).
6. J. P. Gong, Y. Katsuyama, T. Kurokawa, Y. Osada, Adv. Mater.
15, 1155–1158 (2003).
7. M. Hua et al., Nature 590, 594–599 (2021).
8. L. Fu et al., Nature 618, 740–747 (2023).
9. C. Liu et al., Science 372, 1078–1081 (2021).
10. R. A. Dickie, T. L. Smith, Trans. Soc. Rheol. 15, 91–110
(1971).
11. D. Wirthl et al., Sci. Adv. 3, e1700053 (2017).
12. M. Baumgartner et al., Nat. Mater. 19, 1102–1109 (2020).
13. X. Liang et al., Adv. Mater. 33, 2102011 (2021).
14. P. Chodanowski, S. Stoll, J. Chem. Phys. 111, 6069–6081
(1999).
15. A. V. Dobrynin, M. Rubinstein, Prog. Polym. Sci. 30, 1049–1118
(2005).
16. S. B. Mahmoud, W. Essafi, A. Brûlet, F. Boué, Macromolecules
51, 9259–9275 (2018).
17. Y. Kantor, M. Kardar, Europhys. Lett. 27, 643–648 (1994).
18. Y. Kantor, M. Kardar, Phys. Rev. E 51, 1299–1312 (1995).
19. X. Li et al., Sci. Adv. 7, eabe8210 (2021).
20. M. N. Tamashiro, H. Schiessel, Macromolecules 33, 5263–5272
(2000).
21. L. Chen et al., Adv. Funct. Mater. 32, 2107538 (2021).
22. M. Li et al., Nat. Commun. 13, 2279 (2022).
23. Y. Yang, X. Wang, F. Yang, H. Shen, D. Wu, Adv. Mater. 28,
7178–7184 (2016).
24. J. Li, Z. Suo, J. J. Vlassak, J. Mater. Chem. B 2, 6708–6713
(2014).
25. Y. J. Wang et al., Chem. Mater. 31, 1430–1440 (2019).
26. T. J. Jones, E. Jambon-Puillet, J. Marthelot, P.-T. Brun, Nature
599, 229–233 (2021).
27. S. Terryn, J. Brancart, D. Lefeber, G. Van Assche,
B. Vanderborght, Sci. Robot. 2, eaan4268 (2017).
28. J. Shintake, V. Cacucciolo, D. Floreano, H. Shea, Adv. Mater. 30,
1707035 (2018).
29. Z. Shen et al., Adv. Mater. 34, 2203650 (2022).
30. Q. Ge et al., Sci. Adv. 7, eaba4261 (2021).
AC KNOWLED GME NTS
We thank the following individuals and organizations for
their valuable contributions and assistance throughout this
research: X. Miao from Shanghai Advanced Research Institute,
Chinese Academy of Sciences, Shanghai, China, for providing
technical support; Beam line BL16B1 at SSRF (Shanghai
Synchrotron Radiation Facility, China) for granting us beam
time; Y. Li and F. Yang from Tsinghua University Cryo-TEM
Facility of China National Centre for Protein Sciences
(Beijing) for their assistance with cryo-TEM measurements;
S. Yue from Tsinghua University Analysis Center for providing
support with AFM characterization; X. Tan, C. He, and
C. Chen from Tsinghua University for their assistance with
data collection; H. Ding from Suzhou Niumag Analytical
Instrument Corporation for helping with low-field NMR
measurements; R. Zong from Tsinghua University for assisting
with cryo-TEM analysis; H. Xiao from Tsinghua University for
his guidance in computational chemistry; S. Liu, B. Qin, W. Lai,
W. Xu, and X. Zhang from Tsinghua University for their
assistance with DSC tests; X. He and L. Tao from Tsinghua
University for their support with dynamic rheometer
measurements; the Tsinghua Xuetang Talents Program for
providing resources for computational chemistry; Z. Xu,
B. Xue, and Y. Cao from Nanjing University for their assistance
with mechanical characterization; Z. Cheng from the University
of Chicago for his involvement in modifications; Y. Dai, S. Zhang,
J. Si, and H. Xu from Tsinghua University for their assistance
with chemical analysis; and S. Ma from Hebei Aozhe New
Material Technology Company for providing industrial rubbers.
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

Funding: This research was supported by the National supplementary materials. License information: Copyright © Figs. S1 to S33
Natural Science Foundation of China (21890731 and 22075164). 2024 the authors, some rights reserved; exclusive licensee Tables S1 to S3
Author contributions: L.C. and C.W. conceived and designed American Association for the Advancement of Science. No claim References (31–33)
the experiments. L.C. carried out the experiments and to original US government works. https://www.science.org/ Movies S1 to S11
collected the overall data. L.C. and C.W. analyzed the data and about/science-licenses-journal-article-reuse
co-wrote the paper. All authors discussed the results and
commented on the manuscript. Competing interests: The SUPPLEMENTARY MATERIALS Submitted 27 February 2023; resubmitted 9 August 2023
authors declare no competing interests. Data and materials science.org/doi/10.1126/science.adh3632 Accepted 14 February 2024
availability: All data are available in the main text or the Materials and Methods 10.1126/science.adh3632

MACHINE LEARNING proportional to the average gradient outer

product (AGOP) with respect to the input to
Mechanism for feature learning in neural networks this layer. We presented extensive empirical
evidence that our ansatz holds for state-of-the-
and backpropagation-free machine learning models art models such as vision transformers (15),
deep convolutional networks [e.g., ResNet (2),
Adityanarayanan Radhakrishnan1,2†, Daniel Beaglehole3†, Parthe Pandit4,5, Mikhail Belkin5,3* VGG (16), AlexNet (17)], and transformer-based
language models from the GPT2 family (3). We
Understanding how neural networks learn features, or relevant patterns in data, for prediction is demonstrated that the AGOP of GPT2-based
necessary for their reliable use in technological and scientific applications. In this work, we presented a transformers recovered thematically related
unifying mathematical mechanism, known as average gradient outer product (AGOP), that characterized groups of tokens and that the AGOP of con-
feature learning in neural networks. We provided empirical evidence that AGOP captured features volutional networks recovered features related
learned by various neural network architectures, including transformer-based language models, to edge detection.
convolutional networks, multilayer perceptrons, and recurrent neural networks. Moreover, we By definition, AGOP is a backpropagation-
demonstrated that AGOP, which is backpropagation-free, enabled feature learning in machine learning free, data-dependent mathematical operator
models, such as kernel machines, that a priori could not identify task-specific features. Overall, we that can be applied to any predictive model.
established a fundamental mechanism that captured feature learning in neural networks and Thus, we used AGOP to enable feature learn-
enabled feature learning in general machine learning models. ing in general machine learning models through
an algorithm we called Recursive Feature Ma-
chine (RFM). As a model system, we used AGOP
eural networks have been central to major cific learning tasks, e.g., learning single-index

N
breakthroughs in biology (1), computer
vision (2), and language understanding
and generation (3). Yet, a major challenge
for their deployment in scientific or
technological applications is the limited under-
standing of how these models learn and use
features, or relevant patterns in data, for pre-
diction. Without understanding neural feature
learning, it is difficult to establish whether neu-
ral networks produce reliable, accurate, and
appropriate responses. Given the rapidly expand-
ing reach of neural network applications, it is
crucial to characterize the mechanism of neural
feature learning.
The ability of neural networks to learn fea-
tures from data is thought to be a central con-
tributor to their improved effectiveness over
classical machine learning models (4, 5). Despite
active research effort into neural feature learn-
ing, a unified mechanism that captures features
learned across neural architectures had not
been identified by prior work. Most prior work
on neural feature learning analyzed feature
learning in fully connected networks on spe-
1
Harvard School of Engineering and Applied Sciences,
Cambridge, MA 02138, USA. 2Broad Institute of MIT and
Harvard, Cambridge, MA 02142, USA. 3Computer Science
and Engineering, UC San Diego, La Jolla, CA 92093, USA.
4
Center for Machine Intelligence and Data Science, IIT
Bombay, Mumbai 400076, India. 5Halıcıoğlu Data Science
Institute, UC San Diego, La Jolla, CA 92093, USA.
*Corresponding author. Email: mbelkin@ucsd.edu
†These authors contributed equally to this work.
functions (6, 7), multi-index functions (8, 9), or
learning parity functions (4, 10). Recent work
(11) analyzed feature learning in deep, non-
linear fully connected networks that fit data
and minimized representation cost (given
by the sum of squares of network weights).
Works (5, 12–14) empirically identified spe-
cific classification and regression tasks where
fully connected and convolutional neural net-
works outperformed non–feature learning,
kernel methods. The work (12) proved that
two-layer convolutional networks outperformed
non–feature learning, convolutional kernel
methods on tasks where relevant features
for prediction are embedded in noisy back-
ground. The goal of our work was to identify
a single, unifying mechanism that captured
features learned across neural architectures,
including effective models used in practice.
Such a unifying mechanism would open new
avenues for advancing interpretability of neu-
ral networks and enabling feature learning in
machine learning models that cannot na-
tively learn task-relevant features, such as ker-
nel machines.
In this work, we presented a unifying mech-
anism for feature learning that captured features
learned across neural network architectures and
enabled feature learning with general machine
learning models. An overview of our results is
presented in Fig. 1. We posited the neural fea-
ture ansatz (NFA), which stated that features
extracted by a given neural network layer are
to enable feature learning with kernel ma-
chines (18), a classical machine learning mod-
el that has received renewed interest (19–21).
We showed that AGOP of standard kernel
machines and convolutional kernel machines
(20) recovered relevant, task-specific features.
For standard kernel machines operating on
vectorized data, these features were transfor-
mations of input features that were useful
for prediction. For convolutional kernel ma-
chines trained on standard image classifica-
tion tasks, we observed that AGOP with respect
to image patches recovered edge detectors,
a phenomenon that had previously been
observed for convolutional networks (22).
Moreover, we showed that reintroducing fea-
tures captured by AGOP of kernel machines
improved performance of these models, there-
by confirming that AGOP captured fea-
tures inherent in data that were relevant for
prediction.
Overall, we concluded that AGOP is a unify-
ing mechanism for feature learning that can
be used to characterize features learned by
neural network architectures and enable feature
learning in general machine learning models.
Results
Neural networks are nonlinear models that im-
′
plement functions of the form f : ℝdt →ℝct
where d denotes the dimensionality of input
features, c denotes the dimensionality of labels,
and t; t 0 respectively refer to input and label
sequence length for sequence models such as

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1461

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 1. Average gradient outer product (AGOP) captured features in neural
network architectures and enabled feature learning in general machine
learning models. (A) Eigenvectors of neural feature matrices (NFMs) and AGOPs
for five-layer ReLU MLPs trained on CelebA prediction tasks. (B) Eigenvectors of
NFMs and AGOPs for CNNs [VGG19 (16)] pretrained on ImageNet (27) captured
edge detectors. (C) Eigenvectors of NFMs and AGOPs in GPT2-architecture language
models trained on TinyStories (26) identified thematically related token groups. The
example shows how eigenvectors highlighted past-tense verbs in text (further
examples are given in Fig. 3 and fig. S10). (D) AGOP eigenvectors of kernel machines
trained on CelebA tasks recovered task-relevant features. AGOP eigenvectors of
convolutional kernel machines on SVHN (42) recovered edge detectors that could be
used for edge detection on images of different resolution and domain.

transformers or recurrent networks. These mod-
els are recursively defined as
f ðxÞ ¼ hLþ1 ðxÞ; hlþ1 ðxÞ
¼ fl ðhl ðxÞÞ for l∈f1; …; Lg ð1Þ
where h1 ðxÞ ¼ x , hl ðxÞ∈ℝdl tl are the repre-
sentations at the lth layer of the network on
the input x, L ∈ ℤþ denotes the depth of the
network, and fl : ℝdl tl →ℝdlþ1 tlþ1 is param-
ðlÞ
eterized with weight matrices fW i gki¼1 as a
multivariate function of the form
fl ðhl ðxÞÞ :
¼ ~f l W1ðlÞ uð11lÞ ; W1ðlÞ uð12lÞ ; …WkðlÞ uðklðÞm
 
ðl Þ ðl Þ
1Þ ; Wk ukm
ð2Þ
n o
ðlÞ
where uij denote subsets of entries of hl ðxÞ
(e.g., patches of images in convolutional net-
works or token embeddings in transformers).
By adjusting values of k; m, and ~fl , Eq. 2 can be
used to recover standard architectures such as
fully connected networks [also known as mul-
tilayer perceptrons (MLPs)], convolutional net-
works (CNNs), recurrent neural networks (RNNs),
and transformers. Examples are presented in
materials and methods.
To characterize feature learning in neural net-
works, we used Eq. 2 to understand how weight
matrices are involved in transforming hidden
units. Equation 2 implies that the features
extracted by the weights in layer l of a neural
network are characterized by right singular vec-
ðlÞ
tors and singular values of fWi gki¼1 , because
these terms rotate, reflect, and rescale compo-
nents of the input into this layer, hl ðxÞ. These
singular vectors and values can be recovered
from eigenvectors and eigenvalues of the matrices
ðlÞ ðlÞ
Wi T Wi , which we refer to as neural feature
matrices (NFMs). Suppose the neural
network
n
is trained on a dataset xðpÞ ; yðpÞ p¼1 . To
understand neural feature learning, we aimed
to characterize the structure of NFMs at the
end of training.
Our main finding was the NFA, which con-
nected NFMs at any intermediate layer of a
trained network, ^f , to its AGOP as follows:
ðlÞ ðlÞ
Wi T Wi º
n m
!!a
1 T
X X   
∇uðlÞ ^f xðpÞ ∇uðlÞ ^f xðpÞ
n p¼1 j¼1 ij ij
ð3Þ
where º denotes elementwise proportion-
ality, the power a > 0 applies to the matrix,
and ∇ ðlÞ ^f xðpÞ denotes the gradient of ^


uij
f with
ðlÞ
respect to the input uij of ~ fl . For networks
with multivariate outputs, we used the Jacobian
of the resulting model ^f , which is the natural
extension of the gradient. Note thata affects only
the magnitude of the AGOP eigenvalues and
does not affect the order of eigenvalues nor
the eigenvectors. The eigenvectors capture the
task-relevant directions used for identifying
features, which was the focus of our work. In
supplementary text A, we proved that the
ansatz with a ¼ 21 holds with equality when
training two-layer linear networks under
balanced initialization. Motivated by this theo-
retical insight, we chose a ¼ 21 for all empirical
results involving neural networks. Before pre-
senting empirical evidence for the NFA, we
provide intuition for the NFA and discuss
implications.
Intuition for the ansatz and implications
To build effective predictors on high-dimensional
data, it is important to identify features that
are relevant for prediction. A natural approach
to identifying important features is to dis-
tinguish between directions along which
the predictor varies most and least. When
applied to a predictor, AGOP mathematically
identifies task-relevant directions, with top
eigenvectors corresponding to the most re-
levant directions and bottom eigenvectors
corresponding to least relevant directions.
Thus, our ansatz explicates an internal mech-
anism in neural networks that identifies
task-relevant directions. Although such a
mechanism is absent in standard machine
learning models, including kernel machines,
our result implies that AGOP can be used to
enable task-relevant feature learning in these
standard models operating on general data
modalities.
Empirical evidence for the neural
feature ansatz
In Fig. 2, we presented empirical evidence
for the NFA across a wide variety of neural
network architectures, including those com-
monly used in practice. We stratified our
empirical evidence on the basis of architecture
type: transformers (15, 23), CNNs (2, 16, 17),
MLPs, and RNNs (24). In Fig. 2 and fig. S4,

1462 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 2. NFMs were highly correlated (mean Pearson correlation >0.8) with
AGOP in trained neural networks. Mean correlations were higher than those
for AGOP and NFM in untrained models and between NFM before and after
training (see materials and methods and fig. S1). Tables present mean Pearson
correlation across all layers (error bars indicate 1 SD). Box-and-whisker plots
show how correlations vary across layers (and heads in transformers). (A) Correlation
between Query, Key, Value NFMs, and AGOPs for vision transformers (Vit) pretrained
on ImageNet. In the box-and-whisker plot, yellow, blue, gray represent Query, Key,
Value NFM, and AGOP correlations, respectively. (B) Correlation between
Query, Key, Value NFMs, and AGOPs for GPT2-architecture language models
trained on Shakespeare text from (39) and the TinyStories dataset (26). Box-
and-whisker plot shows correlations for the TinyStories language model with
vocabulary size of 3200 [same color scheme as in (A)]. (C) Correlation between
NFMs and AGOPs in CNNs pretrained on ImageNet. (D) Correlation between
NFMs and AGOPs across 121 five–hidden layer ReLU MLPs trained on tabular
data tasks from (30).

we showed for each architecture that mean
Pearson correlation between AGOP and NFM
across all layers in trained networks was high
(i.e., > 0:8 in all cases). In materials and meth-
ods and figs. S1 to S3, we showed that this
correlation was considerably higher than (i)
correlation between AGOP and NFM in un-
trained networks and (ii) correlation between
the NFM in trained models and NFM in un-
trained, randomly initialized models. See ma-
terials and methods for details regarding all
architectures considered.
Given the computational simplicity of training
MLPs over transformers, CNNs, and RNNs,
in figs. S5 to S7, we analyzed how activation
function choice, width (defined as number of
hidden units per layer), and initialization scheme
affected correlation between NFM after training
and AGOP in MLPs. In this setting, we found
that the correlation between NFMs and AGOP
was higher in situations where previous work
identified benefits of feature learning, e.g.,
training low-width networks with default ini-
tialization (14) or training high-width networks
under small initialization (5). Lastly, in supple-
mentary text B and fig. S8, we showed that
AGOP accurately captured spurious features
learned by MLPs.
Characterization of features identified
by AGOP
Thus far, we established that NFMs after
training were correlated with AGOP across
various neural network architectures. We
now demonstrate that AGOPs of trained net-
works capture features in data that are relevant
for prediction. In the case of GPT2-architecture
language models, AGOPs identified related
groups of tokens and in the case of convolutional
networks, AGOP identified edge detectors.
AGOP recovered groups of related
tokens from GPT2-architecture
language models
Recent work (25) demonstrated that project-
ing the embedding of tokens onto the right
singular vectors of value matrices in GPT2
led to identification of groups of related to-
kens. The NFA stated that eigenvectors of
AGOP were equal (up to sign) to right sin-
gular vectors of weight matrices in neural
networks. Thus, the NFA implied that eigen-
vectors of AGOP were an alternative mech-
anism for capturing such related tokens. To
this end, we projected token embeddings from
a GPT2-architecture transformer trained on
TinyStories (26) onto the eigenvectors of value-
AGOPs and extracted top-weighted tokens
(see materials and methods). Our analysis
indicated that eigenvectors of value-AGOPs
recovered related token groups. In Fig. 3A, we
presented the top five eigenvectors of value-
AGOPs resulting in the most thematically re-
lated token groups according to ratings by
GPT4. See materials and methods and fig.
S9A for details on ranking methodology and
comparison with random projections of to-
ken embeddings. We further demonstrated
how value-AGOP eigenvectors could be used
to highlight related tokens in groups of text.
In particular, in Fig. 3B, we presented sam-
ples of context length 256 from TinyStories

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1463

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 3. AGOP of trained networks identified relevant features for
prediction. (A) Related token groups identified by projecting embedding
matrices onto eigenvectors of value-AGOPs of a GPT2-architecture language
model trained on TinyStories. Presented token groups were the 5 out of
124 with highest thematic coherence rating according to GPT4. Scores of all
124 token groups along with comparison to random projections of tokens
are presented in fig. S9. (B) Projecting groups of tokens onto eigenvectors in
(A) highlighted (in green) thematically related tokens. Examples are
presented for projection onto eigenvectors identifying (i) food and cooking
terms and (ii) name and animal terms. Additional examples are presented
in fig. S10. (C) Patch-AGOPs of VGG19 identified edges in ImageNet images
and progressively highlighted regions of images used for prediction.
(D) Eigenvectors of first layer patch-AGOP of VGG19 contained horizontal
and vertical edge detectors.

colored by their inner product with value-
AGOP eigenvectors corresponding to (i) food
and cooking terms and (ii) names and ani-
mals. In fig. S9B, we presented additional
samples of text ranked by their average
token correlation with value-AGOP eigenvec-
tors. Overall, these results showed that AGOP
captured text features learned by language
models.
AGOP recovered edge detectors
from CNNs
Previous work (22) presented evidence that
early layers of trained CNNs captured features
akin to edge detectors. Given that the covariances
of these filters were correlated with AGOP over
image patches (patch-AGOP), eigenvectors of
AGOP should capture similar features. In Fig.
3C, we visualized the features of ImageNet (27)
images captured by AGOP of VGG19 layers
(see materials and methods for implementa-
tion details of our visualization technique). In
particular, we observed that the patch-AGOP
computed with respect to the input of the first
convolutional layer directly acted as an edge
detector. Moreover, we observed that patch-
AGOP computed with respect to input of deeper
layers of VGG19 amplified specific regions of
images for prediction. For example, the first
and last rows of Fig. 3C showed that the eyes
of the dogs were used as predominant features
for classification, and row two showed that
pixels corresponding to the ladybug are pre-
dominantly used for classification. In Fig. 3D,
we demonstrated that eigenvectors of patch-
AGOP computed for the first layer of VGG19
contained horizontal and vertical edge detectors.
Furthermore, our visualization was markedly
different from previous approaches using class-
activation mappings (28). These prior works
attempted to identify features in individual
images, and we identified a single operator per
layer that was used to select features across
all images. Overall, these results showed that
AGOP captures relevant image features used
by CNNs for classification.
AGOP enabled feature learning with general
machine learning models
A powerful outcome of the ansatz was that it
decoupled feature extraction from choice of
machine learning model. Because AGOP is
an operator that can be applied independently
of the model choice, it serves as a universal
mechanism for enabling feature learning in
machine learning models, including those that
a priori could not learn task-specific features.
We demonstrated the effectiveness of AGOP
as a feature learning mechanism by using it
to enable feature learning in kernel machines,
a well-studied class of machine learning models
that does not adaptively extract task-specific
features (see supplementary text C for back-
ground reviewing kernel machines).
AGOP of kernel machines operating on
tabular data extracted relevant features
for prediction
We analyzed features identified by AGOP of kernel
machines operating on tabular (vectorized) data.
We trained kernel machines using the Laplace

kernel of the form K ðx; z Þ ¼ exp gkx z k2
and then computed the AGOP of the trained
kernel machine. In Fig. 4A, we visualized the
top eigenvector of the AGOP for such kernel
machines trained on prediction tasks from
CelebA (29), using vectorized images. Even
though kernel machines rely on a fixed kernel
function regardless of the task, we observed
that AGOP of trained kernel machines identi-
fied task-specific features. Moreover, in fig. S11,
we observed that these AGOP eigenvectors are
similar (cosine similarity of 0:99) to the eigen-
vectors of first-layer NFMs in MLPs trained on
the same tasks, even though these are different
classes of models.
Reintroducing features identified by AGOP of
kernel machines improved model performance:
Recursive feature machines
The fact that AGOP identifies task-specific fea-
tures provides an opportunity to improve kernel
machines. To this end, letting M denote the

1464 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 4. AGOP of kernel machines identified features that could be used to
improve model performance. All models in (B), (C), and (I) used the same
training, validation, and test splits with the same hyperparameters (see materials
and methods). (A) The top eigenvector of AGOP of Laplace kernel machines
trained on tasks from CelebA identified task-specific features. (B and
C) Reintroducing features from AGOP of Laplace kernel machines improved test
accuracy across classification and regression tasks, including the 121 classification
tasks from (30). (D to F) Patch-AGOP of Convolutional Neural Tangent Kernel
(CNTK) trained to classify stars in noisy background identified features that
amplified signal and reduced noise. (G) Eigenvectors of patch-AGOP of CNTKs
trained on SVHN identified edge detectors. (H) Eigenvectors from (G) were edge
detectors that identified edges in images of different resolution from ImageNet.
(I) Reintroducing patch-AGOP into CNTK improved performance in image
classification tasks.

AGOP of a trained kernel machine, we rein-
corporated such features into the kernel func-
1 1

tion by consideringK M 2 x; M 2 z . Moreover, we
iterated between solving kernel regression,
computing AGOP of the trained kernel machine,
and reintroducing AGOP features. We refer to
this iterative approach as a Recursive Feature
Machine (RFM). In Fig. 4, B and C, we demon-
strated that reintroduction of AGOP features
drastically improved performance of kernel
machines [8% increase in classification accu-
racy on Street-View-House-Numbers (SVHN)
and a double in the test R2 on synthetic re-
gression tasks from (8, 14)]. In Fig. 4C, we
provided a comprehensive analysis demon-
strating the benefit of reintroducing AGOP fea-
tures into kernel machines across 121 tabular
datasets from (30). AGOP features improved
overall average accuracy of kernel machines
across the 121 tasks with up to a 31% increase
in accuracy over kernel regression with the
Laplace kernel. Moreover, we observed that
reintroducing features led to better performance
primarily on larger datasets. Indeed, the only
points in Fig. 4C for which feature learning
was worse was for smaller datasets, e.g., those
with fewer than 1000 samples. In figs. S12 and
S13 and tables S1 to S4, we showed perform-
ance of RFM in relation to 181 other models on
the tabular data tasks from (30) and against all
other models and tasks from (31). Moreover,
we found that RFM required only a fraction of
the computational cost of training MLPs (fig.
S12). Computational complexity of computing
AGOP is presented in materials and methods.
All evaluation metrics are described in mate-
rials and methods.
AGOP of convolutional kernel machines
extracted features on image patches
Analogously to the case of using AGOP to
enable feature learning in standard kernel
machines, we used patch-AGOP to enable fea-
ture learning in convolutional kernel machines.
These convolutional kernels, such as the Con-
volutional Neural Tangent Kernel (CNTK) (20),
can be viewed as functions on image patches.
We thus trained convolutional kernel machines
and then computed patch-AGOP to identify rel-
evant features.
In Fig. 4, D to H, we presented examples
of features learned by convolutional ker-
nels on synthetic and real-world image data-
sets. We began with the illustrative synthetic
task of classifying whether a noisy image con-
tains a star pattern (see Fig. 4D for dataset
samples). Prior work (12) showed that CNNs
were able to far outperform kernel machines
in such synthetic settings and claimed that
CNNs were able to amplify signal and reduce
noise in such images in contrast to kernel
machines. After training CNTK on 50 samples
from this synthetic dataset, we computed the

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1465

RES EARCH | R E S E A R C H A R T I C L E S
patch-AGOP, M. We observed that eigenvectors
of M corresponded to patch transformations
that highlight yellow pixels, indicating that
the AGOP captured features relevant for this
specific prediction problem (Fig. 4E). In Fig.
1 6. A. Bietti, J. Bruna, C. Sanford, M. J. Song, in Advances
4F, we multiplied image patches by M 2 and
saw that such features amplified signal corre-
sponding to the star pattern and reduced noise.
In Fig. 4G, we visualized eigenvectors of AGOP
for convolutional kernel machines trained on
SVHN using K ðx; z Þ corresponding to CNTK
with a patch size of 3 × 3. We observed that
AGOP eigenvectors of the trained convolutional
kernel machine corresponded to edge detectors.
Indeed, Fig. 4H demonstrated that patch-AGOPs
of these models trained on SVHN automatically
identified edges in images of different resolution
and domains such as ImageNet. Lastly, in Fig. 4I,
we demonstrated that reintroducing the patch-
AGOP into convolutional kernel machines led
to improvements in performance across various
synthetic and real-world image classification
tasks. Consistent improvement across various
tasks confirmed that patch-AGOP of convolu-
tional kernels identifies relevant features for
image classification.
Remarks
Our results can be extended by recursively
applying AGOP to capture features on non-
linearly transformed versions of input data
(see supplementary text D). We found that
such deep, nonlinear feature learning through
AGOP, a process we call Deep RFM, improved
performance of convolutional kernel ma-
chines, closing the gap between kernel ma-
chines and neural networks (figs. S14 and
S15). We additionally showed that AGOP could
be effectively used in conjunction with alter-
native representations of data to substantially
increase model performance. In supplemen-
tary text E and fig. S16, we used AGOP to im-
prove performance of kernel machine image
classifiers trained on scattering transformed
images (32). Lastly, as AGOP captured task-
specific features on representations of data,
it could be additionally used to perform trans-
fer learning between tasks that required sim-
ilar subsets of features (see supplementary text
F and fig. S17).
Discussion
In this work, we presented a single unifying
mechanism for capturing feature learning in
neural architectures and enabling feature
learning in general machine learning models.
We posited the NFA, which connected neural
feature learning to a mathematical operator
known as AGOP. We presented empirical evi-
dence that AGOP accurately captured features
learned in neural architectures, including state-
of-the-art models used for vision tasks (e.g.,
vision transformers, ResNets, VGG, AlexNet)
and for language tasks (e.g., transformer based
1466 29 MARCH 2024 • VOL 383 ISSUE 6690
language models of the GPT2 family). In par-
ticular, we showed that AGOP of language
models captured groups of related tokens and
AGOP of CNNs captured edge detectors. More
broadly, we demonstrated that AGOP could
be used to enable feature learning in general
machine learning models such as kernel ma-
chines, which a priori could not learn task-
specific features. These task-specific features
could be reintroduced into kernel machines
to improve their performance. We conclude
with a discussion of implications and future
directions.
Advancing transparency in deep
neural networks
Given the broad reach of deep learning to-
day, it is crucial to understand how these
models learn and use features to make pre-
dictions. Precise characterization of neural
feature learning is a key step toward building
deep learning systems that are demonstrab-
ly safe and aligned with human preferences
(33). Our results advanced the transparency
of neural networks by providing a single
mechanism that could be used to charac-
terize and visualize features extracted across
various neural architectures operating on tab-
ular, image, and text data. Thus, we envision
that our results will be a key step toward build-
ing transparent and interpretable machine
learning models.
New avenues for building effective machine
learning models
Identification of a small set of task-relevant
features from a much larger set of features
is a key step for building effective, interpret-
able predictive models. Approaches for task-
relevant feature selection have been developed
and studied in classical machine learning
literature, including works for feature selec-
tion in linear models (34) and nonlinear
models in specific tabular data settings (e.g.,
multi-index models) (35–37). Prior to our work,
it was unclear how to disentangle neural fea-
ture learning from predictor construction
because both occur simultaneously when
training deep networks using backpropaga-
tion. Thus, our work presented an important
foundational advancement in framing neural
feature learning from the classical machine
learning perspective. We believe our findings
offer an opportunity to improve current neural
architectures by making them more compu-
tationally efficient, theoretically grounded,
and interpretable.
RE FERENCES AND NOTES
1. K. Tunyasuvunakool et al., Nature 596, 590–596 (2021).
2. K. He, X. Zhang, S. Ren, J. Sun, “Proceedings of the 2016 IEEE
conference on computer vision and pattern recognition (2016),
pp. 770–778.
3. A. Radford et al., Language models are unsupervised multitask
learners. OpenAI blog 1 (8), 9 (2019).
4. Z. Shi, J. Wei, Y. Lian, “Proceedings of the 2022 International
Conference on Learning Representations (2022).
5. G. Yang, E. J. Hu, “Tensor Programs IV: Feature learning
in infinite-width neural networks” in Proceedings of the 2021
International Conference on Machine Learning
(PMLR 2021), pp. 11727-11737.
in Neural Information Processing Systems (Curran Associates,
2022), pp. 9768–9783.
7. J. Ba et al., “High-dimensional asymptotics of feature learning:
How one gradient step improves the representation” in
Advances in Neural Information Processing Systems (Curran
Associates, 2022), pp. 37932–37946.
8. A. Damian, J. Lee, M. Soltanolkotabi, “Conference on Learning
Theory (PMLR, 2022), pp. 5413–5452.
9. E. Abbe, E. Boix-Adsera, T. Misiakiewicz, “Conference on
Learning Theory (PMLR, 2022), pp. 4782–4887.
10. A. Daniely, E. Malach, in Advances in Neural Information
Processing Systems (Curran Associates, 2020),
pp. 20356–20365.
11. A. Jacot, in Advances in Neural Information Processing
Systems (Curran Associates, 2023).
12. S. Karp, E. Winston, Y. Li, A. Singh, in Advances in Neural
Information Processing Systems (Curran Associates, 2021),
pp. 24883–24897.
13. P. M. Long, Properties of the after kernel. arXiv:2105.10585
[cs.LG] (2021).
14. N. Vyas, Y. Bansal, P. Nakkiran, Limitations of the ntk for
understanding generalization in deep learning. arXiv:2206.
10012[cs.LG] (2022).
15. A. Kolesnikov et al., “Proceedings of the 2021 International
Conference on Learning Representations (2021).
16. K. Simonyan, A. Zisserman, “Proceedings of the 2015
International Conference on Learning Representations
(2015).
17. A. Krizhevsky, I. Sutskever, G. E. Hinton, in Advances in Neural
Information Processing Systems (Curran Associates, 2012),
pp. 1097–1105.
18. B. Schölkopf, A. J. Smola, Learning with Kernels: Support
Vector Machines, Regularization, Optimization, and Beyond
(MIT Press, 2002).
19. A. Jacot, F. Gabriel, C. Hongler, in Advances in Neural
Information Processing Systems (Curran Associates, 2018),
pp. 8571–8580.
20. S. Arora et al., “On exact computation with an infinitely wide
neural net” in Advances in Neural Information Processing
Systems (Curran Associates, 2019), pp. 8141–8150.
21. R. Novak et al., “Proceedings of the 2020 International
Conference on Learning Representations (2020).
22. M. D. Zeiler, R. Fergus, “Proceedings of the 2014
European Conference on Computer Vision (Springer, 2014),
pp. 818–833.
23. A. Vaswani et al., “Attention is all you need” in Advances in
Neural Information
Processing Systems (Curran Associates, 2017),
pp. 5998–6008.
24. K. Cho et al., “Proceedings of the 2014 Conference on
Empirical Methods in Natural Language Processing (Association
for Computational Linguistics, 2014), pp. 1724–1734.
25. G. Dar, M. Geva, A. Gupta, J. Berant, Analyzing
transformers in embedding space. arXiv:2209.02535[cs.CL]
(2022).
26. R. Eldan, Y. Li, Tinystories: How small can language models be
and still speak coherent english? arXiv:2305.07759[cs.CL]
(2023).
27. O. Russakovsky et al., Int. J. Comput. Vis. 115, 211–252
(2015).
28. R. R. Selvaraju et al., “Proceedings of the 2017
International Conference on Computer Vision (IEEE, 2017),
pp. 618–626.
29. Z. Liu, P. Luo, X. Wang, X. Tang, “Proceedings of the 2015
International Conference on Computer Vision (IEEE, 2015),
pp. 3730-3738.
30. M. Fernández-Delgado, E. Cernadas, S. Barro, D. Amorim,
J. Mach. Learn. Res. 15, 3133–3181 (2014).
31. L. Grinsztajn, E. Oyallon, G. Varoquaux, in Advances in Neural
Information Processing Systems Datasets and Benchmarks
(2022), pp. 1–48.
32. J. Bruna, S. Mallat, IEEE Trans. Pattern Anal. Mach. Intell. 35,
1872–1886 (2013).
33. L. Ouyang et al., “Training language models to follow
instructions with human feedback” in Advances in Neural
Information Processing Systems (Curran Associates, 2022),
pp. 27730–27744.
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

34. R. Tibshirani, J. R. Stat. Soc. B 58, 267–288 (1996).
35. S. Trivedi, J. Wang, S. Kpotufe, G. Shakhnarovich, in
Uncertainty in Artificial Intelligence (Morgan Kaufmann, 2014),
pp. 819–828.
36. W. Härdle, T. M. Stoker, J. Am. Stat. Assoc. 84, 986–995
(1989).
37. S. Mukherjee, D.-X. Zhou, J. Shawe-Taylor, J. Mach. Learn. Res.
7, 519–549 (2006).
38. T. Wolf et al., “Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing: System Demonstrations
(Association for Computational Linguistics, 2020), pp. 38–45.
39. A. Karpathy, nanogpt (2022); https://github.com/karpathy/
nanoGPT.
40. S. Arora et al., “Proceedings of the 2020 International
Conference on Learning Representations (2020).
41. A. Radhakrishnan, D. Beaglehole, P. Pandit, M. Belkin, version
1, agop_feature_learning (2024); https://doi.org/10.5281/
zenodo.10676950.
42. Y. Netzer et al., “Reading digits in natural images with
unsupervised feature learning” in Advances in Neural
Information Processing Systems Workshop on Deep Learning
and Unsupervised Feature Learning (2011).
ACKN OW LEDG MEN TS
A.R. thanks A. Philippakis for feedback on the paper, C. Uhler for
support on this work, and E. Lander for discussions on
transformers. We thank D. Hsu, J. Tropp, and J. Lee for valuable
literature references. Funding: A.R. is currently supported by the
George F. Carrier postdoctoral fellowship in applied mathematics
at Harvard School of Engineering and Applied Sciences (SEAS) and
was supported by the Eric and Wendy Schmidt Center at the Broad
Institute of MIT and Harvard at the time of submission. P.P. is
currently supported by a Thakur Family Chair at IIT Bombay and
was supported by a Simons-HDSI postdoctoral fellowship at UCSD
at the time of submission. We acknowledge support from the
National Science Foundation (NSF) and the Simons Foundation for
the Collaboration on the Theoretical Foundations of Deep Learning
through awards DMS-2031883 and 814639 as well as the TILOS
institute (NSF CCF-2112665). This work used the programs XSEDE
(Extreme science and engineering discovery environment), which is
supported by NSF grant nos. ACI-1548562, and ACCESS
(Advanced cyberinfrastructure coordination ecosystem: services
and support), which is supported by NSF grant nos. 2138259,
2138286, 2138307, 2137603, and 2138296. Specifically, we used
the resources from SDSC Expanse GPU compute nodes, and NCSA
Delta system, via allocations TG-CIS220009. Author
contributions: A.R., D.B., P.P., and M.B. designed and performed
the research, analyzed the data, and wrote the paper. Competing
interests: None declared. Data and materials availability: All
data needed to evaluate the conclusions in the paper are present in
the paper or the supplementary materials. All image datasets
considered in this work, i.e., CelebA, SVHN, CIFAR10, CIFAR100,
GTSRB, MNIST, STL-10, PCAM, Caltech101, DTD, QMNIST, and
EMNIST, are publicly available for download via PyTorch.
ImageNet32 and ImageNet are publicly available for download
through the link to the ImageNet portal provided in (27).
TinyStories is publicly available for download through HuggingFace
(38). Shakespeare data are provided directly from (39). Tabular
data from (30) is publicly available through the link to code
provided in (40). Tabular data from (31) are publicly available
through the link to code provided in their paper. All code for (i)
computing correlation between AGOP and NFM for various
architectures (transformers, CNNs, RNNs, MLPs); (ii) visualizing
AGOP features in CNNs and transformer-based language models
of the GPT2-family; and (iii) training RFMs on tabular and image
data are available at (41). License information: Copyright © 2024
the authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original
US government works. https://www.sciencemag.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adi5639
Materials and Methods
Supplementary Text
Figs. S1 to S17
Tables S1 to S4
References (43–72)
Submitted 4 May 2023; resubmitted 16 January 2024
Accepted 22 February 2024
Published online 7 March 2024
10.1126/science.adi5639

ULTRAFAST OPTICS in electron microscopy (10–16). Despite the ex-

isting body of theoretical work (17–19), the
Ultrafast Kapitza-Dirac effect conventional Kapitza-Dirac effect on electrons
was not observed until 2001 (20) because of
Kang Lin1,2*, Sebastian Eckart2, Hao Liang3, Alexander Hartung2, Sina Jacob2, the experimental challenges mainly related to
Qinying Ji2,4, Lothar Ph. H. Schmidt2, Markus S. Schöffler2, Till Jahnke5, the overall weakness of the effect. In that ex-
Maksim Kunitski2, Reinhard Dörner2* periment, a narrowly collimated electron beam
produced by an electron gun was used. The
Similar to the optical diffraction of light passing through a material grating, the Kapitza-Dirac effect electrons traveled through a standing light wave,
occurs when an electron is diffracted by a standing light wave. In its original description, the which was formed by two counterpropagating
effect is time independent. Here, we extended the Kapitza-Dirac effect to the time domain. By laser beams. In the final electron momentum
tracking the spatiotemporal evolution of a pulsed electron wave packet diffracted by a 60-femtosecond spectrum, discrete diffraction peaks spaced by
(where one femtosecond = 10−15 seconds) standing wave pulse in a pump-probe scheme, we observed two-photon momenta, 2ħkg, were observed
time-dependent diffraction patterns. The fringe spacing in the observed pattern differs from that (20, 21) (Fig. 1A).
generated by the conventional Kapitza-Dirac effect. By exploiting this time-resolved diffraction scheme, The advent of the ultrafast laser technology
we can access the time evolution of the phase properties of a free electron and potentially image provides the opportunity to address the dy-
ionic potentials and electronic decoherences. namics of the Kapitza-Dirac photon scatter-
ing in an ultrafast pump-probe experiment
and explore the time dependence of the elec-

S
hortly after the wave nature of electrons
was demonstrated by diffracting an elec-
tron beam at a periodic crystal structure
(1), Kapitza and Dirac suggested that an
electron beam can also be diffracted by
an immaterial standing light wave (2). Today,
this is referred to as the Kapitza-Dirac effect
(3). It was shown that this diffraction occurs due
to the stimulated Compton back-scattering of
1
School of Physics, Zhejiang Key Laboratory of Micro-Nano
Quantum Chips and Quantum Control, Zhejiang University,
Hangzhou 310058, China. 2Institut für Kernphysik,
Goethe-Universität, 60438 Frankfurt am Main, Germany.
3
Max Planck Institute for the Physics of Complex
Systems, 01187 Dresden, Germany. 4Center for Ultrafast
Science and Technology, School of Chemistry and
Chemical Engineering, Shanghai Jiao Tong University,
Shanghai 200240, China. 5Max-Planck-Institut für
Kernphysik, 69117 Heidelberg, Germany.
*Corresponding author. Email: klin@zju.edu.cn (K.L.); doerner@
atom.uni-frankfurt.de (R.D.)
photons in the standing wave. During this
process, n times twice the photon momentum
ħkg is imparted onto the electron, causing the
deflection of the electron beam to discrete an-
gles. Unlike this particle perspective of photon
scattering, the standing wave can be thought
of as a spatially periodic ponderomotive po-
tential that leads to diffraction of the electron
wave packet (3). In matter optics, the Kapitza-
Dirac effect is used for the manipulation of
matter waves. For instance, a standing light
wave can serve as a coherent beam splitter for
particle bunches, including electrons, ions,
atoms, and molecules (3–6). Such a matter
splitter is an essential part of particle inter-
ferometers (7), which have many applications
ranging from extremely sensitive studies of
gravitational effects (8) to the generation of
spin-polarized electrons (9). Further, laser fields
are routinely used for shaping electron bunches
tron diffraction. In addition, strong laser fields
can be used for the production of electrons
through ionizing neutral atoms or molecules,
which are then subject to the Kapitza-Dirac
effect in the experiment. Electrons released
by such laser fields from atoms or molecules
have specific properties that are different from
those produced by an electron gun. The former
can be considered as a coherent electron wave
packet with a broad distribution in momen-
tum space, and the latter can be approximated
as a plane wave with a momentum-space dis-
tribution close to a delta function. The early
attempts (22) used two counterpropagating,
100-ps (where 1 ps = 10−12 s) laser pulses to
form a standing light wave. The pulsed stand-
ing wave ionized individual atoms in the gas
phase, setting the electrons free and subse-
quently diffracting them. As a result of a very
high photon-scattering rate in the intense

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1467

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 1. Momentum space perspective of the

conventional and the ultrafast Kapitza-Dirac effects.
(A) The conventional Kapitza-Dirac effect describes the
diffraction of a plane electron wave that can be
represented as a delta function in momentum space.
Thus, the width of the momentum distribution is much
narrower than two-photon momenta 2kg. In a standing
light wave, stimulated Compton backscattering of photons
shifts this narrow momentum distribution by ±2nkg,
where n is an integer. The resulting regularly spaced peak
structure in momentum space is time independent. (B) By
contrast, for the ultrafast Kapitza-Dirac effect, the
stimulated Compton backscattering transfers the
momentum to an electron wave packet that has a very
broad momentum distribution. The Compton
backscattering occurs during a short time interval
(femtoseconds) and at a variable time delay after the
birth of the wave packet. The interference between the original (blue Gaussian) and the momentum-shifted (red and green Gaussians) wave packets leads to time-
dependent fringes in momentum space.

Fig. 2. Observation of the ultrafast Kapitza-Dirac effect. (A to D) Experimentally
measured momentum distributions of the electron that is released upon
strong-field ionization of xenon atoms in a pulsed standing light wave (800 nm,
60 fs, 1.0 × 1014 W/cm2). The ionizing pulse is followed by a weak standing
light wave pulse (0.4 × 1014 W/cm2) that acts as a probe pulse. The probe
pulse is attenuated such that it cannot lead to the liberation of bound electrons
and thus only serves as a diffraction grating. Vertical axis shows the electron
momentum along the polarization direction (pz). Horizontal axis shows the
electron momentum along the light-propagation direction (px). Panels (A) to
(D) show the data taken for different time delays of –1, 1, 2, and 3 ps,
respectively. (E to H) 1D momentum distribution along the light-propagation
axis by selecting the range of pz between (0.45, 0.55) atomic units [shaded
areas in (A) to (D)]. The dashed curve in (E) is obtained by fitting a Gaussian
distribution. The solid curves in (F) to (H) show results from numerically
propagating a 1D electron wave packet followed by the interaction with a
standing wave at different time delays.

1468 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

fringe spacing was decreased continuously by

increasing the time delay between the pump
and probe pulses. This observation is different
from that of the conventional Kapitza-Dirac
A effect, where the diffraction pattern shows a
constant spacing of two-photon momenta. The
corresponding results depict the measured
2D electron momentum distribution along the
laser polarization (z axis) and propagation di-
rections (x axis) for various time delays (Fig. 2).
For the sake of visibility, a gating of |py| > 0.1 a.u.
(atomic units are used unless stated other-
wise) was applied to the dataset to remove the
B Coulomb-focusing-induced distortions around
the center of the 3D momentum distribution
(29, 30). The y axis points along the direction
of the gas jet. As a reference, Fig. 2A shows the
momentum distribution recorded at negative
time delay of –1 ps. In this case, the probe
pulse arrives before the pump pulse such that
the electron wave packet, which is released by
the pump pulse, does not experience diffrac-
C tion by the standing wave. The 2D momentum
distribution extends along the polarization di-
rection (pz) of the pump pulse resulting from
the acceleration of electrons by the pump pulse’s
electric field. Along the light-propagation di-
rection (px), the electron momentum distribu-
tion is much narrower because it is determined
solely by the initial velocity distribution of the
electrons upon tunneling. The distribution is
similar to the extensively studied momentum
distribution of electrons released by tunnel
ionization using a single traveling femtosec-
Fig. 3. Diffraction pattern as a function of time delay. (A and B) Measured (A) and simulated
ond laser pulse (31). The Kapitza-Dirac effect
(B) electron momentum distribution along px for light pulses as in Fig. 2 as a function of the time
manifests itself in the periodic vertical fringes,
delay between ionization (by the pump pulse) and probing by the non-ionizing standing light wave
i.e., an intensity modulation along the light-
(probe pulse). The data are integrated over |py| >0.1 a.u. (C) Fitted diffraction spacing from (A)
propagation direction (horizontal axes in Fig. 2).
as a function of the time delay. The solid curve is obtained from simulations in full analogy to
At a negative pump-probe delay (Fig. 2A), we
the experiment.
found along this direction a close to Gaussian
distribution centered at zero momentum with a
full width at half-maximum (FWHM) of ~0.25 a.u.
At positive time delays (Fig. 2, B to D), the electron
light field, the measured electron angular emis- light wave (pump pulse). After a variable time momentum distribution turned into fringe
sion distribution consisted of two peaks with delay, a weaker, non-ionizing, femtosecond stand- patterns in the light-propagation direction,
a spacing corresponding to thousands of pho- ing light wave (probe pulse) was applied to and the spacing between diffraction fringes
ton momenta. This is also referred to as “rain- diffract the emitted electron wave packet. The decreased with increasing time delay (Fig. 2,
bow” scattering (19) or channeling (23). No pump and the probe pulses were generated by F to H). Specifically, the spacing between the
peaks spaced by two-photon momenta were a commercial Ti:Sapphire laser system (Coherent diffraction peaks at 1, 2, and 3 ps was ~0.18,
observed [see (24) for theoretical modeling Legend Elite Duo, 25 fs, 800 nm, 10 kHz) by 0.09, and 0.06 a.u., respectively. To find a
predicting such interference fringes]. In a focusing two counterpropagating laser pulses quantitative relation between the diffraction
later experiment, a 55-fs standing wave was into the same target volume (26), where they spacing and the time delay, we performed a
used to scatter neutral atoms (25). However, were finally stretched to ~60 fs. Using a stand- continuous scan of the pump-probe delay from
neither of these experiments observed inter- ing wave as the pump locks the birth posi- 2 to 10 ps. Figure 3A clearly shows that the
ference fringes, nor did they have access to tion of the electron wave packet relative to spacing between the diffraction fringes de-
the temporal properties of the diffraction the field maxima of the probe standing wave creased continuously with increasing time de-
process. (see the supplementary materials). The three- lay. We found that the spacing between the
dimensional (3D) momenta of the electrons diffraction peaks equals one half the wave-
Fully time-resolved experiment on the were measured using a COLd Target Recoil length of the standing light wave lSW times
Kapitza-Dirac effect Ion Momentum Spectroscopy (COLTRIMS) the electron mass me (with me = 1 in atomic
The concept of our experiment is illustrated in reaction microscope (27, 28). units) divided by the time delay, i.e.
Fig. 1B. An electron wave packet was released With this experimental arrangement, we
from a xenon atom upon strong-field ioniza- observed time-dependent interference fringes lSW
Dpx ¼ me ð1Þ
tion using a highly intense, pulsed standing in the electron momentum distribution. The 2t

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1469

RES EARCH | R E S E A R C H A R T I C L E S
This result was unexpected because the conven-
tional Kapitza-Dirac effect is time independent.
Hereafter, we refer to this newly observed phe-
nomenon as the “ultrafast Kapitza-Dirac effect.”
Numerical modeling
To understand the underlying physics of the
ultrafast Kapitza-Dirac effect, we numerically
propagated a 1D electron wave packet, fol-
lowed by interaction with the pulsed standing
wave (28). We started with an initial electron
wave packet defined by the distribution ob-
served for the negative time delay (Fig. 2E)
and with a width of ~0.25 a.u. in momentum
space. This is a very broad distribution com-
pared with the transverse momentum dis-
tribution of a collimated electron beam (which
is often even approximated by a delta function
in momentum space). Specifically, in (20), the
electron beam had a momentum distribution
with a width of 2.2 × 10−4 a.u. along the light-
propagation direction. Subsequently, the initial
electron wave packet evolved under field-
free conditions. At different time delays, we
applied an ultrashort spatially periodic poten-
tial (which mimics the ponderomotive poten-
tial of the probe pulse) to diffract the electron
wave packet. The simulated momentum dis-
tributions show fringes (Figs. 2 and 3) that
agree very well with the experimental obser-
vation and its time-domain behavior.
The ultrafast Kapitza-Dirac effect can be con-
ceptionally understood as an interference be-
tween the initial electron wave packet and a
replica of the electron wave packet that has
been shifted in momentum space by 2kg re-
sulting from stimulated Compton backscatter-
ing of the photons of the probe pulse. Although
the initial electron wave packet evolves under
field-free conditions [thus, accompanied with a
phase evolution of ϕe ðt Þ ¼ p2x t=2], the altered
electrons experience the aforementioned 2kg
recoil along the light-propagation direction when
interacting with the standing wave. Accordingly,
this interaction leads to a momentum-shifted
electron wave packet with
an accompanying
2
phase of ϕ′e ðt Þ ¼ px T 2kg t=2 þ ϕSW, where
ϕSW encodes the relative phase of the standing
wave with respect to the initial electron wave
packet (28). The tiny change of energy of the
electron is compensated by an equivalent
change of the photon energy upon scattering,
which is within the bandwidth of the short
pulse. In our case, ϕSW ¼ 2p . The initial and
shifted wave packets interfere constructively
when the phase difference is Dϕe ¼ ϕ′e ϕe ¼
2np. Thus, the peaks in the final electron mo-
mentum distribution are
np

expected to be located
1 lSW
at ðpx Þn ≈ tkg
¼ n þ 2 2t , resulting in a mod-
ulation spacing of Dpx ¼ l2t SW
. 4. P. L. Gould, G. A. Ruff, D. E. Pritchard, Phys. Rev. Lett. 56,
Concluding remarks
The ultrafast Kapitza-Dirac effect differs fun-
damentally from the conventional Kapitza-Dirac
1470 29 MARCH 2024 • VOL 383 ISSUE 6690
effect. In momentum space, the latter gen-
erates extremely narrow (delta-function-like)
side bands in momentum space. Because these
sidebands and the original momentum dis-
tribution of the electron do not overlap, an
interference of these contributions is not pos-
sible. In the case of the ultrafast Kapitza-
Dirac effect, however, the momentum shift of
2kg is much smaller than the width of the
initial electron wave packet, which enables
an interference of the initial and scattered
contribution (Figs. 2 and 3). The observed
interference fringes not only provide insight
into the temporal dynamics of the studied
light-matter interaction, but we also envision
that the ultrafast Kapitza-Dirac effect can be
used as a powerful interferometer to measure
phase gradients of free electrons in general,
an observable that is otherwise inaccessible
using state-of-the-art electron spectroscopy.
Techniques such as velocity map imaging, dis-
persive electron analyzers, and COLTRIMS
reaction microscopes routinely access the am-
plitude of electron wave packets. Approaches
to learning about the phase of photoelectrons,
such as RABBITT (Reconstruction of Attosec-
ond Beating by Interference of Two-photon
Transitions) (32, 33) or HASE (Holographic
Angular Streaking of Electrons) (34, 35), rely
on intercepting the photoelectron’s creation
mechanism. By contrast, the ultrafast Kapitza-
Dirac effect acts as an electron interferometer
that reads out phase information long after
the ionization event, leaving the ionization
process undistorted. The interference fringes
shown in Figs. 2 and 3 for positive time delays
are therefore the result of probing a freely
propagating electron and accessing its quan-
tum-phase properties. Conversely, the con-
trast of the diffraction fringes can serve as
a probe for the coherence of the electron
wave packet. Thus, the ultrafast Kapitza-
Dirac effect could be used to trace the time
dependence of decoherence caused by in-
teraction with an environment such as a
large molecule or a neighboring metal clus-
ter surface (36, 37). Possible applications of
such pulsed Kapitza-Dirac interferometers are
investigating (i) phase entanglement between
several electrons or between electrons and
ions, (ii) electrons released from surfaces, or
(iii) chiral electron wave packets, tracking the
origin of their chirality as they emerge from
chiral molecules.
RE FERENCES AND NOTES
1. C. Davisson, L. H. Germer, Nature 119, 558–560 (1927).
2. P. L. Kapitza, P. A. M. Dirac, Math. Proc. Camb. Philos. Soc. 29,
297–300 (1933).
3. H. Batelaan, Rev. Mod. Phys. 79, 929–941 (2007).
827–830 (1986).
5. O. Nairz, B. Brezger, M. Arndt, A. Zeilinger, Phys. Rev. Lett. 87,
160401 (2001).
6. B. A. Stickler, M. Diekmann, R. Berger, D. Wang, Phys. Rev. X
11, 031056 (2021).
7. D. M. Giltner, R. W. McGowan, S. A. Lee, Phys. Rev. Lett. 75,
2638–2641 (1995).
8. X. Wu et al., Sci. Adv. 5, eaax0800 (2019).
9. M. M. Dellweg, C. Müller, Phys. Rev. Lett. 118, 070403
(2017).
10. A. H. Zewail, Science 328, 187–193 (2010).
11. H. Müller et al., New J. Phys. 12, 073011 (2010).
12. M. Kozák, N. Schönenberger, P. Hommelhoff, Phys. Rev. Lett.
120, 103203 (2018).
13. S. A. Koppell et al., Ultramicroscopy 207, 112834
(2019).
14. O. Schwartz et al., Nat. Methods 16, 1016–1020
(2019).
15. K. Wang et al., Nature 582, 50–54 (2020).
16. M. C. Chirita Mihaila et al., Phys. Rev. X 12, 031043
(2022).
17. R. Gush, H. P. Gush, Phys. Rev. D Part. Fields 3, 1712–1721
(1971).
18. M. V. Fedorov, Opt. Commun. 12, 205–209 (1974).
19. H. Batelaan, Contemp. Phys. 41, 369–381 (2000).
20. D. L. Freimund, K. Aflatooni, H. Batelaan, Nature 413, 142–143
(2001).
21. D. L. Freimund, H. Batelaan, Phys. Rev. Lett. 89, 283602
(2002).
22. P. H. Bucksbaum, D. W. Schumacher, M. Bashkansky, Phys.
Rev. Lett. 61, 1182–1185 (1988).
23. C. Salomon, J. Dalibard, A. Aspect, H. Metcalf,
C. Cohen-Tannoudji, Phys. Rev. Lett. 59, 1659–1662
(1987).
24. X. Li et al., Phys. Rev. Lett. 92, 233603 (2004).
25. S. Eilzer, H. Zimmermann, U. Eichmann, Phys. Rev. Lett. 112,
113001 (2014).
26. A. Hartung et al., Nat. Phys. 15, 1222–1226
(2019).
27. R. Dörner et al., Phys. Rep. 330, 95–192 (2000).
28. See the supplementary materials.
29. T. Brabec, M. Y. Ivanov, P. B. Corkum, Phys. Rev. A 54,
R2551–R2554 (1996).
30. A. Ludwig et al., Phys. Rev. Lett. 113, 243001 (2014).
31. S. V. Popruzhenko, J. Phys. At. Mol. Opt. Phys. 47, 204001
(2014).
32. P. M. Paul et al., Science 292, 1689–1692 (2001).
33. H. G. Muller, Appl. Phys. B 74, s17–s21 (2002).
34. S. Eckart, Phys. Rev. Res. 2, 033248 (2020).
35. D. Trabert et al., Nat. Commun. 12, 1697 (2021).
36. P. Sonnentag, F. Hasselbach, Phys. Rev. Lett. 98, 200402
(2007).
37. N. Kerker, R. Röpke, L. M. Steinert, A. Pooch, A. Stibor, New J.
Phys. 22, 063039 (2020).
38. Data for: K. Lin et al., Ultrafast Kapitza-Dirac effect,
Zenodo (2024); https://doi/10.5281/
zenodo.10617990.
AC KNOWLED GME NTS
We thank C. Lu and J. Jiang for figure plotting and
P. He and F. He for fruitful discussions. Funding: This
work was supported by the DFG (German Research
Foundation). S.E. was supported by the DFG through
Priority Programme SPP 1840 QUTIF. Author
contributions: K.L., S.E., A.H., S.J., Q.J., L.Ph.H.S.,
M.S.S., T.J., M.K., and R.D. performed the experiments.
K.L., S.E., M.K., and R.D. analyzed the data. M.K. and
H.L. performed the simulations. All authors contributed
to writing the manuscript. Competing interests: The
authors declare no competing interests. Data and
materials availability: All data are available in the
manuscript or the supplementary materials and have
been deposited at Zenodo (38). License information:
Copyright © 2024 the authors, some rights reserved;
exclusive licensee American Association for the
Advancement of Science. No claim to original US
government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adn1555
Materials and Methods
Supplementary Text
Figs. S1 to S3
Submitted 27 November 2023; accepted 20 February 2024
10.1126/science.adn1555
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

NEUROSCIENCE centrations (fig. S3D). DBLI is correlated with

DPO2 in a linear manner (Fig. 1F), although
Oxygen imaging of hypoxic pockets in the mouse availability of free O2 might limit the sensitiv-
ity at low O2 supply. GeNL BLI is thus an ac-
cerebral cortex curate measure of relative PO2.
Increased metabolism leads to increased
Felix R. M. Beinlich1*†, Antonios Asiminas1†, Verena Untiet1, Zuzanna Bojarowska1, Virginia Plá1, carbon dioxide (CO2) concentrations as a by-
Björn Sigurdsson1, Vincenzo Timmel2, Lukas Gehrig2, Michael H. Graber2, product of cellular respiration. CO2 is a known
Hajime Hirase1,3, Maiken Nedergaard1,3* vasodilator and alters pH (9, 10). We thus in-
vestigated the pH change in brain tissue upon
Consciousness is lost within seconds upon cessation of cerebral blood flow. The brain cannot store oxygen, and O2 calibration to exclude the possibility of
interruption of oxidative phosphorylation is fatal within minutes. Yet only rudimentary knowledge exists pH alterations causing the effect on BLI de-
regarding cortical partial oxygen tension (PO2) dynamics under physiological conditions. Here we introduce scribed above. Mice expressing the genetically
Green enhanced Nano-lantern (GeNL), a genetically encoded bioluminescent oxygen indicator for PO2 imaging. encoded fluorescent pH sensor pHuji (11) in
In awake behaving mice, we uncover the existence of spontaneous, spatially defined “hypoxic pockets” and the extracellular space were implanted with
demonstrate their linkage to the abrogation of local capillary flow. Exercise reduced the burden of hypoxic a fiber photometry probe and exposed to the
pockets by 52% compared with rest. The study provides insight into cortical oxygen dynamics in awake same protocol for O2 calibration under KX
behaving animals and concurrently establishes a tool to delineate the importance of oxygen tension in anesthesia. Manipulating O2 did not alter
physiological processes and neurological diseases. pH during O2 calibration (fig. S4).
Activation of the whisker barrel cortex by con-
tralateral whisker stimulation leads to an in-
he human brain uses ∼20% of total body tion of GeNL with furimazine depends on O2 crease in local blood flow and an increase in PO2.

T oxygen consumption at rest (1–3). Deliv-

ery and demand of oxygen (O2) are so
finely balanced that maintaining tissue
oxygenation may be the most critical of
all brain functions. Yet our understanding of
the dynamics of brain tissue oxygen tension
(Fig. 1A), and the intensity of the biolumine-
scence signal is linearly correlated to the avail-
ability of O2 when O2 is the rate-limiting factor
in the enzymatic reaction (8).

Green enhanced Nano-lantern can detect

(PO2) under physiological conditions remains
limited, largely because of the lack of spatially
precise measurement techniques for PO2 imag-
ing. Currently, tissue PO2 can be measured by
phosphorescence and by Clark-type electrodes
(4, 5). Neither approach provides sufficiently
high spatiotemporal sensitivity to detect phys-
iological changes in cortical PO2.
We developed a methodology to measure rel-
ative changes in PO2 using the oxygen-dependent
reaction of a luminescent substrate guided by
its enzyme expressed in astrocytes. This method
has a superior signal-to-noise ratio due to its
bioluminescence origin and a spatiotemporal
resolution that can visualize the dynamics of
cortical PO2 in awake behaving mice. Green
enhanced Nano-lantern (GeNL) is a lumines-
cent fusion protein consisting of the lucifer-
ase NanoLuc (6) and the fluorescent protein
mNeongreen (7). During the enzymatic con-
version of its luminescent substrate furima-
zine to furimamide, energy is emitted in the
form of light (6). The fluorescent fusion pro-
tein acts as a fluorescence amplifier, increasing
the quantum yield through Förster resonance
energy transfer (FRET). The enzymatic reac-
1
Division of Glial Disease and Therapeutics, Center for
Translational Neuromedicine, Faculty of Health and Medical
Sciences, University of Copenhagen, 2200 Copenhagen,
Denmark. 2School of Engineering, FHNW University of
Applied Sciences and Arts Northwestern Switzerland, 5210
Windisch, Switzerland. 3Division of Glial Disease and
Therapeutics, Center for Translational Neuromedicine,
University of Rochester Medical Center, Rochester, NY 14642, USA.
*Corresponding author. Email: felix.beinlich@sund.ku.dk (F.R.M.B.);
nedergaard@urmc.rochester.edu (M.N.)
†These authors contributed equally to this work.
oxygen in mouse cortex
To assess whether the oxygen dependency can
be used in vivo to visualize spontaneous PO2
dynamics in the brain, we expressed GeNL in
cortical astrocytes of wild-type mice (fig. S1A)
and measured bioluminescence intensity (BLI)
after topical administration of the substrate
furimazine through a cranial window (Fig. 1, A
and B, and fig. S2). BLI followed cerebral PO2
when O2 concentration of the breathing air
was changed stepwise from 10% to 40% under
ketamine-xylazine (KX) anesthesia. The baseline
level of air O2 was kept at 20% and then changed
for 1-min periods followed by a 1-min recovery
phase at baseline levels (Fig. 1C). Changing the
O2 concentration from 20% to 40% increased
BLI by ~200%, whereas a reduction of O2 con-
centration in the inhaled air to 10% decreased
BLI by ~50% from baseline (Fig. 1, D and E;
f ig. S3A; and movie S1). To record the absolute
PO2 in the BLI imaging field, we placed an O2-
sensitive Clark-type microelectrode in the field
of view during the calibration protocol (Fig. 1B).
Cerebral PO2 increased by 20 mmHg at 30% O2
and by 50 mmHg at 40% O2 and reduced by
10 mmHg from baseline at 10% O2 (Fig. 1, D and
E, and fig. S3A). The reaction time and slope
of BLI occurred in parallel with the electrode
recordings, indicating similar reaction times
with no time lag between change in O2 and
change in BLI (fig. S3, B and C). BLI peaked
faster than electrode recordings upon increase
of O2 concentration but declined at a similar
rate when O2 was lowered, perhaps indicat-
ing substrate availability limitations under
conditions with unphysiologically high O2 con-
We next imaged the whisker barrel cortex in
awake behaving mice while stimulating the
whiskers by air puffs to test whether BLI can
detect the sensory-induced cortical PO2 change.
Mice were exposed to a series of 10 brief whisker
stimulations (5 Hz, 50-ms duration for 10 s,
50-s interstimulation interval) (Fig. 1G) (12).
During whisker stimulation, BLI increased in
the field of view (Fig. 1H and movie S2), closely
following the individual series of whisker stimu-
lation trials (Fig. 1I). Moreover, KX-anesthetized
mice with simultaneous O2 microelectrode
recording showed similar consistent corre-
lation between BLI and DPO2 (fig. S5). Notably,
the amplitude and the timing of BLI response
differed between awake and KX-anesthetized
mice (fig. S6), supporting the notion that an-
esthesia dampens tissue PO2 and functional
hyperemia dynamics (13, 14).
Oxygen dynamics in the cerebral cortex
Continuous imaging of BLI showed that PO2
under resting conditions was highly dynamic,
exhibiting local transient dips in PO2 (Fig. 2, A
to C, and movie S3). These local hypoxic events
were spatially constricted, lasting several sec-
onds up to minutes, and typically showed a
sharp on- and offset in relative tissue PO2 (Fig. 2,
C to E). Given their negative relative amplitude
in PO2 and the characteristic sharply defined
border, we called these “hypoxic pockets.”
Hypoxic pockets were identified on the basis
of their negative amplitude [PO2(DB/B)], typical
sharp on- and offset, clear edges, and long
duration (fig. S7A and movie S4). Hypoxic pockets
were observed throughout 20-min-long record-
ings (Fig. 2, F to H). In KX-anesthetized mice,
200 ± 22 hypoxic pockets (mean ± SEM) were
detected during this time (Fig. 2I). Within a
single frame recorded at 1 Hz, 8.12 ± 0.04 hy-
poxic pockets were observed per square milli-
meter, covering 2.43 ± 0.02% of the field of view

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 147 1

RES EARCH | R E S E A R C H A R T I C L E S

A Bioluminescence oxygen imaging B GeNL Bioluminescence (BLI) 3595

a.u.

O2
electrode
266

O2/N2
GeNL + BLI

Furimazine + NLuc Furimamide + hv

O2 CO2

O2
electrode
C 10,789

a.u.

20% O2 40% 20% 30% 20% 10% 20% 107

D O
400 2
40%
30% 100
E 300 **** 80 **** F 300
20% 20% 20% 20% r2= 0.82
10% *** ****
300 75 200
**** 60 **** 200
PO2(mmHg)
PO2(mmHg)

PO2( B/B)
PO2( B/B)
PO2( B/B)

40
200 50 100 100
20
100 25 0 0
0

0 0 -100 -20 -100

0 60 120 180 240 300 360 420 10% 30% 40% 10% 30% 40% -20 0 20 40 60
Time (s) O2 O2 PO2(mmHg)
G Whisker puff H Air Puff 2512 I Air puff
Awake

900
Air puff
PO2(BLI)

a.u.
800
5 Hz, 50 ms

50 s 10 s 50 s
Virus 700
injection Craniotomy Imaging
0 200 400 600
3 weeks 5-6 hours 60 s 70 s 80 s 411 Time (s)

Fig. 1. Bioluminescence intensity of GeNL reports cerebral partial oxygen
pressure. (A) (Top) Scheme of the experimental setup. KX-anesthetized mice
expressing GeNL under glial fibrillary acidic protein promoter were placed under
a macroscope and exposed to different O2 concentrations. An O2-sensitive
microelectrode was inserted into the cortex through an acute craniotomy and
artificial cerebrospinal fluid supplemented with furimazine (0.25 mM).
(Bottom) Chemical reaction of furimazine to furimamide catalyzed by GeNL
under presence of O2 leading to light output in the form of bioluminescence.
(B) Overlay of BLI and mNeonGreen fluorescence excited at 490 nm. Dashed line
outlines the craniotomy; solid line outlines the oxygen-sensitive microelectrode.
(C) BLI changes with varying concentrations of O2 in the inhaled air. Single
frames (1-s exposure time) at indicated O2 concentrations throughout the
calibration protocol. (D) Mean trace of the BLI and PO2 as measured with the
O2 electrode at varying O2 concentration. Shaded area indicates standard error.
(E) BLI intensity [PO2(DB/B)] changes 162% (± 19.13, SEM) when O2 concentration
in the tidal air is doubled. A 10% increase in O2 induces a 66% (± 6.70) increase.
Under hypoxia, BLI is reduced by 29% (± 2.91). Accordingly, PO2 increases by
52.4 mmHg (± 2.4) when O2 concentration is doubled. A 10% increase in O2 increases
PO2 by 18.6 mmHg (±1.9), a 10% decrease reduces PO2 by 9.7 mmHg (±1.6).
(Left) One-way repeated measures analysis of variance (ANOVA) F1.272,10.18 = 82.66,
P < 0.0001 main effect of group. Tukey post hoc: 10% versus 30%, P < 0.0001; 10%
versus 40%, P < 0.000; 30% versus 40%, P = 0.0006. (Right) One-way repeated
measures ANOVA F1.523,12.18 = 219.7, P < 0.0001 main effect of group. Tukey post hoc:
10% versus 30%, P < 0.0001; 10% versus 40%, P < 0.000; 30% versus 40%,
P < 0.0001. (F) Dependency of change in BLI and corresponding change in PO2
recorded with the O2 electrode in the same region upon changes in O2 concentration
of the inhalation air. Data points are fitted with a linear regression with a coefficient of
determination, R2, of 0.82. Dots sharing the same color indicate data points from
the same mouse. (G) Cerebral PO2 was measured using BLI in somatosensory cortex
of awake stationary head-fixed mice subjected to air-puff whisker stimulation.
(H) Images from single frames at indicated time points before, during, and after a
whisker puff. (I) Trace of DPO2 upon whisker stimulation showing elevated PO2 during
each of the 10 repetitions. Means ± SEM are shown. ***P < 0.001, ****P < 0.0001.
[(D) to (F)] N = 9 trials from six mice. a.u., arbitrary units. Scale bars, 100 mm.

1472 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

(Fig. 2, J and K). Hypoxic pockets covered an To quantify whether PO2 during hypoxic signal interference by hemoglobin as the under-
area of 1823 ± 27 mm2 while lasting 48.2 ± 1.0 s pockets indeed reaches the hypoxic threshold lying cause of the hypoxic pockets, we recorded
(Fig. 2, L and M). PO2(DB/B) decreased by 27.2 ± in the cortex (≤18 mmHg) (15), we compared the mNeonGreen fluorescence of GeNL instead
4.5% from baseline before each hypoxic pocket the PO2(DB/B) decrease in hypoxic pockets with (fig. S9A). Hemoglobin absorption should also
(Fig. 2N). Hypoxic pockets had an average diam- those during hypoxia (10% inhaled O2), where affect the mNeonGreen fluorescence, which
eter of 45.29 ± 0.31 mm and an almost circular PO2 reached 11 mmHg (Fig. 1D). During hypoxia, would be subject to hemoglobin absorption, but
shape (Fig. 2, O and P). Hypoxic pockets often PO2(DB/B) was reduced by 29.2 ± 2.9%, which is does not depend on O2. However, no events with
occurred repeatedly in the same area. Within a similar to the decrease observed in hypoxic similar spatiotemporal characteristics as hypoxic
20-min recording, on average 73.2 ± 5.1 of those pockets (fig. S8A). Using the electrode and pockets were observed in the mNeonGreen
regions of interest (ROIs) were detected with BLI correlation, we determined that a 20.6% fluorescence traces (fig. S9, B to E).
on average 2.8 ± 0.1 hypoxic pockets (Fig. 2, Q DB/B reduction corresponds to the hypoxic O2 tension in and around venules is lower
and R). Each minute, 0.15 ± 0.01 hypoxic pockets threshold (fig. S8B). than in and around arterioles (17). We thus mea-
occurred within a given ROI, meaning that Spectral absorption by hemoglobin is a com- sured the distance of hypoxic pockets from
every ~7 min, a hypoxic pocket occurred at the mon cause of artifacts in imaging fluorescent arterioles and venules. Hypoxic pockets were
same place (Fig. 2S). biosensors in vivo (16). To exclude the possible closer to venules than to arterioles, with an

A Hypoxic pockets B Average projection C PO2 (BLI) D E Pocket vicinity

KX Pocket
400 s 420 s 440 s
7 1 8000
5 2
3

P O2 (BLI)
6000
1 3 4
6 5
4000
2 6
2 (z score, a.u.)

20 min, 1 Hz KX 7
Virus 8 4 8 2000
injection Craniotomy Imaging 360 420 480
3 weeks KX Time (s)
60 s

F Hypoxic pockets G I J K L M
20 400 15 5 15000 150
Pockets/mm²/s

4
mouse/mm²

Area/s (%)
300
Pockets/mm²

Seconds
15 10 10000 100
3

µm²
10 200
2
5 5000 50
5 100 1
0 0 0 0 0
0 Hypoxic Number Area Size Duration
0 400 800 1200
Time (s) Pockets
1200

1000
H 8
N O P Q R S
800 0 150 1.5 150 10 0.5
Time (s)

Pockets/ROI/min
6 -20 8 0.4
Pockets/ROI
Area (%)

600
ROI/mouse
B/B (%)

100 1.0 100

4 -40 6 0.3
a.u.
µm

400
-60 4 0.2
2 50 0.5 50
200
-80 2 0.1
0 0 -100 0 0.0 0 0 0.0
0.9
0.9
0 400 800 1200 PO 2 Diameter Circularity No. of KX KX
0.45 0.45 Time (s) ROIs
0 0
mm

Fig. 2. Characterization of hypoxic pockets in cortex of anesthetized mice. by hypoxic pockets for each frame. Shading indicates SEM. (I) Average number
(A) Cerebral PO2 was imaged in somatosensory cortex of KX-anesthetized mice of detected hypoxic pockets per square millimeter per mouse in a recording
over 20 min using BLI. (B) Average projection of BLI of the mouse cortex session. (J) Frequency distribution of number of hypoxic pockets per square
expressing GeNL in astrocytes after furimazine administration from a 20-min millimeter per second. (K) Frequency distribution of area covered by hypoxic
recording. (C) Time traces of the z-scored bioluminescent intensity of tissue PO2. pockets per second. (L) Frequency distribution of size of hypoxic pockets.
Numbers indicate corresponding ROI from manually drawn circles in (B). (M) Frequency distribution of duration of hypoxic pockets. (N) Frequency
(D) Images recorded at 400, 420, and 440 s from (B). Insets show magnification distribution of amplitude changes relative to baseline PO2 of hypoxic pockets.
of the ROIs indicated by arrowheads [2 from (C)]. Inner circle defines area of (O) Frequency distribution of diameter of hypoxic pockets. (P) Frequency
pocket. The outer circle defines pocket vicinity. (E) Time trace of cerebral PO2 distribution of circularity of hypoxic pockets. (Q) Average number of active
from the ROI indicated by cyan (inner) and black (outer) circle in (D). regions with reoccurring hypoxic pockets (ROIs) per mouse in a single recording
(F) (Top) Average distribution of hypoxic pockets from (B). (Bottom) Hypoxic session. (R) Frequency distribution of the number of hypoxic pockets for each
pockets during KX anesthesia in the form of ROIs over time displayed in an detected active region. (S) Frequency distribution of the number of hypoxic
x-y-t 3D rendering. Cyan regions denote signal. (G) Line plot showing average pockets for each detected active region per minute. N = 9 mice (eight recorded
number of hypoxic pockets per square millimeter per second over time. Shading for 20 min, one for 10 min). Violin plots show median and quartiles. Scale
indicates SEM. (H) Line plot showing average area of the field of view covered bars, 100 mm.

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1473

RES EARCH | R E S E A R C H A R T I C L E S

A B C D 140 125 G 0 *

PO2( B/B)

PO2( B/B)

PO2( B/B)
Hypercapnia 120
KX 100 -20
**** ****

N
Norm
100

orm
1000 ****
80 75 -40
0 300 600 900 Norm Hyper Norm Norm Hyper Norm
800 Time (s)
**** H
E 15 10
**** ****
150

Pockets/mm²/s
Pockets/mm²

Duration (s)
600 10 100

Time (s)
5

Hyper
CO2 5 50
400 0 0 0
10 % CO2 0 300 600 900 Norm Hyper Norm Norm Hyper Norm
3
Time (s)
2.0
I 10000
200 F

Size (µm²)
5 min 5 min 5 min

Area (%)
Area (%)
2
Virus N
Norm 1.0 5000
0
orm
injection Craniotomy Imaging 0.8 1
0.8
3 weeks KX 0 0.0 0
0 0 0 200 400 600 800 Norm Hyper Norm Norm Hyper Norm
Time (s)
mm
J K L M 300

Time (s)
Capillary stalling BLI Composite
KX Microsphere CAM 1
CAM 2 0
LED 0.8
0.4 0.8
CAM 1 0.4
LED 0 0
mm
300

Time (s)
Microspheres
TL
CAM 2 Fx
Virus
Microsphere i.v. M 0
injection Craniotomy Imaging 0.8
0.8
TL Fm
D2 D1 0.4 0.4
3 weeks KX 4x
0 0
mm
N 15 15 O 10 8 **** P 0 Q 30000 **** R 150
Pockets/mm²/s

****
Pockets/mm²

Size (µm²)

Duration (s)
6 -20
Area (%)

PO2( B/B)
Area (%)

10 10 20000 100
6 -40
4
5 4 -60
5 10000 50
2 2 -80
0 0 0 0 -100 0 0
0 100 200 300 Ctrl Stalling 0 100 200 300 Ctrl Stalling Ctrl Stalling Ctrl Stalling Ctrl Stalling
Time (s) Time (s)

S 230 s 245 s T 1400 800 U V W X

20 1.0 200 200
Microsphere(a.u.)

Fisher transformation (z)

0 0.5
1200 100 150
PO2 (BLI)

Offset (s)

Seconds
0.0
-20
B/B

750
-0.5 0 100
1000 -40
-1.0
-100 50
-60 -1.5

800 700 -80 -2.0 -200 0

200 250 300 PO 2 Correlation Offset Duration
Time (s)

Fig. 3. Effects of vasodilation and capillary stalling on hypoxic pockets. during, and after CO2 increase. (I) Size of hypoxic pockets before, during, and after
(A) Cerebral PO2 was measured in KX-anesthetized mice exposed to 10% CO2 in CO2 increase. (J) Cerebral PO2 was measured with BLI in KX-anesthetized mice. Capillary
the inhaled air for 10 min after acute craniotomy. (B) Averages of the location stalling was induced by intravascular injection of 4-mm microspheres before onset
of hypoxic pockets, before, during, and after hypercapnia. Norm, normocapnia; hyper, of imaging session. (K) The hybrid BLI-fluorescence microscope setup used to image
hypercapnia. (C) Hypoxic pockets during transition to and from hypercapnia, BLI and microsphere fluorescence simultaneously. (L) (Left) Average intensity
respectively, in the form of ROIs over time displayed in an x-y-t 3D rendering. Turquoise projection of BLI (top) and microsphere fluorescence (bottom). (Right) Overlay of
regions denote signal at normoxia, lilac regions denote signal during hypercapnia. BLI and microsphere fluorescence. (M) Hypoxic pockets in the form of ROIs
(D) (Left) Average trace of cerebral PO2 during the experiment. Lilac shading indicates over time displayed in an x-y-t 3D rendering. Turquoise regions denote signal
the period of increased CO2, turquoise shading indicates periods of normal CO2 levels. under control, and lilac regions denote signal after injection of microspheres.
(Right) Frequency distribution of PO2 per second in a 5-min window before, during, and (N) (Left) Average number of hypoxic pockets from control mice and after
after increased CO2. (E) (Left) Average trace of the number of hypoxic pockets per injection of microspheres. Shading indicates ± SEM. (Right) Frequency distribution
square millimeter. Lilac shading indicates period of increased CO2, turquoise shading of number of hypoxic pockets per second in the cortex of control and microsphere-
indicates periods of normal CO2 levels. (Right) Frequency distribution of number of injected mice. (O) (Left) Average area covered by hypoxic pockets in control
pockets per square millimeter per second before, during, and after increased CO2. mice and mice injected with microspheres. Shading indicates ± SEM. (Right)
(F) (Left) Average trace of the area covered by hypoxic pockets. Lilac shading indicates Frequency distribution of area covered by hypoxic pockets per second in the cortex
period of increase CO2, turquoise shading indicates periods of normal CO2 levels. of control and microsphere-injected mice. (P) Amplitude of hypoxic pockets.
(Right) Frequency distribution of area covered by hypoxic pockets per second in a 5-min (Q) Size of hypoxic pockets. (R) Duration of hypoxic pockets. (S) Images recorded
window before, during, and after increased CO2. (G) Amplitude of hypoxic pockets at 230 and 245 s from (L). Arrows indicate ROI. (T) Time trace of BLI (pink)
before, during, and after CO2 increase. (H) Duration of hypoxic pockets before, and microsphere fluorescence (blue) from the ROI indicated by arrow in (S).

1474 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


Dotted lines indicate time points of images in (S). Gray shading indicates
time window where microsphere was visible. (U) Amplitude of BLI decrease
during events where microspheres were stalled. (V) Fisher z-transformation
of correlation coefficient of BLI and microsphere fluorescence. (W) Offset
between BLI decrease normalized to microsphere onset. (X) Duration of
average distance of 28.1 ± 0.7 and 48.0 ± 1.2 mm,
respectively (fig. S10).
Tissue oxygenation is closely linked to the
availability of O2 and therefore to capillary
circulation of red blood cells. We thus hypoth-
esized that hypoxic pockets result from hemo-
dynamic changes in the microcirculation. We
used intrinsic optical spectroscopy imaging
(IOSI) (12, 16) to monitor hemoglobin dynamics
in the brain of resting awake mice (fig. S11A).
Hemoglobin dynamics were recorded at the
isosbestic point for total hemoglobin concen-
tration ([HbT]) (fig. S11B). The analysis iden-
tified areas of low [HbT] that shared their
characteristic onset and offset dynamics with
hypoxic pockets measured with BLI (fig. S9, C
to F). On average 0.9 ± 0.01 events of low [HbT]
were detected per square millimeter per second
in IOSI measurements, which lasted 16.5 ± 0.4 s
covering 7344 ± 402 mm2 (fig. S11, G to J).
During these events, [HbT] decreased by 1.9 ±
0.1 mmHg (fig. S11K). Whereas bioluminescence
allows only the measurement of tissue hypoxia,
IOSI reflects blood volume and thus transient
localized decreases of hemoglobin concentration
with similar temporal and spatial properties as
shown for hypoxic pockets.
Effects of vasodilation and capillary stalling
on hypoxic pockets
Brain activity is accompanied by transient in-
creases in blood flow due to vasodilation, a
phenomenon called neurovascular coupling or
functional hyperemia (18). We asked whether
hyperemia suppresses the number of hypoxic
pockets and induces vasodilation by increasing
CO2 in air (hypercapnia) (Fig. 3A) (9). The ele-
vation of CO2 in the inhaled air increased tis-
sue PO2 reversibly to 118.4 ± 0.3% concurrently
with a sharp decrease in the number of hypoxic
pockets per square millimeter per second in a
reversible manner, from 3.9 ± 0.1 to 2.3 ± 0.1
and back to 4.7 ± 0.1, a reduction of 41% when
CO2 was lowered (Fig. 3, B to E). Area covered
by hypoxic pockets changed by 0.54%, from
1.01 ± 0.03% to 0.47 ± 0.02%, a reduction of 53%
(Fig. 3F). PO2(DB/B) decrease within hypoxic
pockets during hypercapnia was reduced by 4%
and stayed low after hypercapnia during the
time of recording (Fig. 3G). Furthermore, the
duration of hypoxic pockets was reduced by
17 s, and their spatial expansion was reduced
by 65% (Fig. 3, H and I). After resolution of
hypercapnia, hypoxic pocket duration and size
recovered (Fig. 3, H and I). This observation is
consistent with the finding that hypercapnia
SCIENCE science.org
microsphere stalling. N = 6 mice, hypercapnia. N = 9 mice, control. N = 5 mice,
microsphere injection. Means ± SEM are shown. Violin plots show median
and quartiles. Bars indicating Tukey’s post hoc tests between groups
at their edges: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Scale
bars, 100 mm.
decreases the portion of poorly perfused capil-
laries in the rat brain cortex (19). To assess
whether an increased acidification during hyper-
capnia is involved (20), we measured extra-
cellular pH in the brain parenchyma during
10% CO2 as described above. A 5-min period of
10% CO2 decreased pHuji intensity by 6.4% in
a reversible manner (fig. S12). Typically, fluores-
cent proteins are quenched in acidic environ-
ments, yet we observed an increase in BLI
during hypercapnia, excluding the possibility
that the observation was an artificial effect of
pH on the sensor’s intensity.
Isoflurane is another potent vasodilator (21, 22)
that increases blood volume but not tissue pH
(23). We compared mice before and during iso-
flurane anesthesia and observed a decrease in
the number of hypoxic pockets similar to that
seen for CO2 (fig. S13, A and B). During iso-
flurane anesthesia, the number of hypoxic pockets
per second was effectively decreased from 6.2 ±
0.01 to 1.9 ± 0.03, a reduction of 69% (fig. S13C).
The area covered by hypoxic pockets was re-
duced by 0.83%, a decrease of 13% (fig. S13D).
The PO2(DB/B) decrease within hypoxic pockets
was 21% less in isoflurane-anesthetized mice
than in awake mice, and duration decreased by
15 s (fig. S13, E and F). The size of hypoxic
pockets was reduced by 13% (fig. S13G).
A direct consequence of hypercapnia and
isoflurane anesthesia–induced vasodilation is
an increase in tissue PO2 (13, 24). Increased tis-
sue PO2 might compensate for low-oxygenated
areas by increasing O2 diffusion and thus re-
ducing prevalence of hypoxic pockets (25). To
uncouple vasodilation and tissue oxygenation,
we increased the O2 in the inhalation air from
20% to 30%, inducing hyperoxia (fig. S14, A to C).
This led to a 55% increase in DPO2 during
hyperoxia, which is substantially higher than
the increase measured under hypercapnia (fig.
S14D). During hyperoxia, the number of hypoxic
pockets per second dropped from 3.5 ± 0.0.04
to 3.1 ± 0.03, a decrease of 11% (fig. S14E), which
is far less compared with the isoflurane and
hypercapnia conditions, where hypoxic pockets
were reduced by 69 and 41%, respectively.
During hyperoxia, PO2(DB/B) in hypoxic pockets
decreased 15.9% less than during normoxia (fig.
S14F), which might be explained by the accom-
panying increase in PO2 leading to an overall
increased tissue PO2, reducing the amplitude of
the hypoxic pockets. Furthermore, the duration
of hypoxic pockets was reduced by 13 s (41.5 ±
1 versus 28.5 ± 1 s), and their size increased
slightly by 13% (fig. S14, G to H). These obser-
RESE ARCH | R E S E A R C H A R T I C L E S
vations show that vasodilation more potently
controls the hypoxic pockets than does blood
oxygenation.
Local cerebral microcirculation is mainly reg-
ulated by changes of vascular resistance (26).
Reversible adhesion of circulating leukocytes
can effectively halt capillary blood flow, a phenom-
enon called capillary stalling (27). To directly
test whether capillary stalling elicits hypoxic
pockets, microspheres (diameter: 4 mm) were
delivered intravascularly (Fig. 3J) to occlude
capillaries (28, 29). We recorded O2 dynamics
and microspheres in the same field of view
simultaneously at 1 Hz using a hybrid BLI-
fluorescence microscope that uses the readout
time of the BLI camera to record microsphere
fluorescence with a second camera (Fig. 3, K to
M). After injection of microspheres, the num-
ber of hypoxic pockets decreased from 8.2 ±
0.07 to 5.9 ± 0.06 (Fig. 3N). In contrast, the
total area covered by hypoxic pockets per second
increased from 2.4% to 5.9%, an increase of
146% (Fig. 3O). Whereas PO2(DB/B) only mar-
ginally changed (Fig. 3P), the duration of
microsphere-induced hypoxic pockets was 11 s
shorter (Fig. 3R). Microsphere-induced hypoxic
pockets were 445% larger than in the control
group (Fig. 3Q). Taken together, upon micro-
sphere injection, the area of hypoxic tissue was
increased, as reflected by an increase in size
but not in the number of pockets. A plausible
explanation for this observation is that upon
microsphere injection, hypoxic pockets near
each other may have fused together, thus
decreasing the number of individually detected
events. When interrogating the dynamics of
individual spheres and local tissue O2 dynamics,
it was clear that influx of a microsphere quickly
triggered a local decrease in PO2 with a similar
temporal dynamic as the spontaneously occurring
hypoxic pockets (Fig. 3, S to X).
Hypoxic pockets are reduced by wakefulness
and further suppressed by locomotion
To detect O2 dynamics in awake animals and
assess the effect of locomotion, mice were trained
to tolerate head restraint. One group of awake
mice was recorded on a stationary platform,
whereas a second group of mice was placed on
a polystyrene ball allowing head-fixed locomo-
tion (Fig. 4A). The number of hypoxic pockets
in behaving mice was reduced by 17% com-
pared to KX-anesthetized mice (Fig. 4, B, D, and
F). Yet the pockets increased in surface area by
41% (Fig. 4J). Hypoxic pockets lasted on av-
erage 8 s less in awake mice, whereas the
29 MARCH 2024 • VOL 383 ISSUE 6690 1475
RES EARCH | R E S E A R C H A R T I C L E S

A B C D E
KX 1200
20
5

Pockets/mm²

Area (%)
800 15 4

Time (s)
10 3
400 2
5 1
Virus 0 0
injection Craniotomy Imaging 0 0 300 600 900 1200
1.2 0 300 600 900 1200
1.2 Time (s)
3 weeks KX
0 0 Time (s)
Wake 1200 20 5

Pockets/mm²
4

Area (%)
15

Time (s)
800 3
10
2
400 5 1
Virus 0 0
injection Craniotomy Imaging 0 0 300 600 900 1200 0 300 600 900 1200
1.2 1.2 Time (s) Time (s)
3 weeks 5-6 hours
0 0
Mobile 1200
20 5

Pockets/mm²
4

Area (%)
15
800
Time (s)

3
10
2
400 5 1
Virus 0 0
injection Craniotomy Imaging 0 0 300 600 900 1200 0 300 600 900 1200
1.2 1.2 Time (s) Time (s)
3 weeks 24 hours mm 0 0

F **** G H I *
15 **** 200
* 0
Pockets/mm²/s

5
****
Duration (s)
****
Area/s (%)

150

PO2( B/B)
10 **** 4 -50
3 **** 100
5 2 -100
1 50
0 0 0 -150
KX Wake Mobile KX Wake Move KX Wake Mobile KX Wake Mobile
J K L M
*

Hypoxic Burden (Ξ)

No. pockets same

15000 0.6 1×10 6

time and space
Area (µm²)

10000 0.4
Area 5×10 5

Time
Amplitude

5000 0.2
e
Tim
0 0.0 0
KX Wake Mobile KX Wake Mobile KX Wake Mobile

Fig. 4. Increased arousal level suppresses tissue hypoxia. (A) Cerebral PO2
was either measured in KX-anesthetized mice, in awake head-fixed mice
during quiet wakefulness in a MAG-1 mouse holder, or in mobile mice voluntarily
running on a Styrofoam sphere. (B) Distribution of hypoxic pockets in a single
recording session lasting 20 min. (C) Hypoxic pockets identified in (B) in the
form of ROIs over time displayed in an x-y-t 3D rendering. Turquoise regions
denote signal. (D) Average number of hypoxic pockets detected for each frame in
mice during KX anesthesia, wakefulness, and mobile wakefulness, respectively.
(E) Average area that hypoxic pockets cover in the field of view in percentage
detected for each frame in mice during KX anesthesia, wakefulness, and mobile
wakefulness, respectively. (F) Frequency distribution of number of hypoxic
pockets. (G) Frequency distribution of area covered by hypoxic pockets in the
field of view per second. (H) Frequency distribution of duration of hypoxic
pockets. (I) Frequency distribution of amplitude of hypoxic pockets. (J) Frequency
distribution of size of hypoxic pockets. (K) Frequency distribution of number of
hypoxic pockets sharing the same region per second. (L) Schematic illustrating the
parameters used to calculate the hypoxic burden for each hypoxic pocket (left; x)
and for an entire recording/mouse (right; X). (M) Hypoxic burden during KX
anesthesia, wakefulness, and mobile wakefulness. N = 9 mice, KX. N = 11 mice,
awake. N = 10 mice, mobile. Means ± SEM are shown. Violin plots show median and
quartiles. Bars indicating Tukey’s post hoc tests between groups at their edges:
*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Scale bars, 100 mm.

amplitude was unaffected (Fig. 4, H and I). dynamics were recorded (Fig. 4, A to C). Com- mice (Fig. 4J), whereas the number of ROIs of
Taken together, the PO2 maintains the same pared with immobilized awake mice, the num- hypoxic pockets did not differ between KX-
characteristics but is more dynamic in the ber of hypoxic pockets was reduced by 35%, anesthetized, mobile, and immobile mice (Fig.
awake state than in the anesthetized state, from 6.7 ± 0.05 to 4.3 ± 0.03, and the per- 4K). It is unlikely that the decrease of hypoxic
possibly reflecting the higher level of neuro- centage of area covered by hypoxic pockets at pockets during active locomotion is due to a
nal activity and thereby blood flow in wake- any given time was reduced by 33% (Fig. 4, D general increase in blood flow, as their highly
fulness compared with anesthesia (30, 31). to G). On average, hypoxic pockets lasted 7 s structured spatial characterization is preserved,
We also sought to determine whether the less in mice that could freely run (Fig. 4H). which is in contrast to the expected general linear
number of hypoxic pockets correlates with Further, the amplitude of hypoxic pockets increase in blood flow. We summarized the bur-
mice actively running, where increased blood was dampened by 10.4% compared with im- den of hypoxic pockets on the brain as a measure
flow and tissue PO2 is expected (32). Twenty- mobilized mice (Fig. 4I). The spatial coverage of the effect of area, duration, and amplitude
four hours after surgery, mice were placed on of the hypoxic pockets was further reduced by (Fig. 4L). The analysis showed that the hypoxic
an air-supported polystyrene ball (33), and PO2 26% in running mice compared with immobile burden is reduced by 52% in running mice

1476 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


compared with quiet awake mice (Fig. 4M).
Earlier studies have shown that sensory stimu-
lation can reduce prevalence of capillary stalls
(25). To test whether a similar paradigm sup-
presses the occurrence of hypoxic pockets, a
series of 10 whisker stimulations were deliv-
ered to anesthetized mice (fig. S15A). The stimu-
lation resulted in a 35% decrease in the number
of hypoxic pockets (fig. S15, B to H). Thus, func-
tional hyperemia suppressed the occurrence of
hypoxic pockets, whether initiated by sensory
stimulation in anesthetized mice or by active
running in awake behaving mice (fig. S15, I to M).
Discussion
This study shows that relative changes in PO2
can be monitored continuously in wide cortical
regions of awake behaving mice by a genetically
encoded bioluminescent oxygen indicator. By
monitoring the bioluminescence signal of GeNL
expressed in astrocytes, we found that corti-
cal PO2 constantly fluctuates during physiolog-
ical conditions, giving rise to spatially and
temporally defined hypoxic pockets. Manipu-
lations that either increased or blocked capillary
flow showed that local interruption of the mi-
crocirculation is responsible for the occurrence
of hypoxic pockets. This conclusion was sup-
ported by imaging (i.e., IOSI) of cortical hemo-
globin absorption. Monitoring of the local
hemoglobin concentration identified transient
local reduction in hemoglobin, which shared
characteristic onset and offset dynamics with
hypoxic pockets, thus providing an alternative
approach to validate the occurrence of spatially
restricted PO2 fluctuations during physiological
conditions.
Two-photon imaging of cortical capillaries
has documented that the velocity of red blood
cells can vary by more than a factor of 10 (34–36).
A characteristic trait of the mammalian cortex
is the unbalanced prevalence of arterioles and
venules. A modeling study predicted that low-
flow regions in the capillary bed are inevitable
owing to the many sources and sinks of blood
flow and that they tend to form around which-
ever vessel is more numerous (37). In contrast
to humans, where the ratio is 2.5:1 (arterioles:
venules), in mice, venules outnumber arterioles
by a factor of 2.6 (38), which is expected to
increase the low-flow regions preferentially
around venules. Indeed, our analysis supported
this notion by demonstrating that hypoxic pock-
ets tend to appear closer to venules than to
arterioles in mouse cortex.
We expressed GeNL under an astrocytic pro-
moter to take advantage of the fact that the fine
astrocytic processes infiltrate all parts of the
neuropil (39, 40). However, we do not expect
that the astrocyte-selective expression of GeNL
had substantial impact on the observations.
Oxygen diffuses freely across the neuropil with
no restrictions imposed by either plasma mem-
branes or by specific cell types (41, 42).
SCIENCE science.org
Why have hypoxic pockets not previously
been detected? Recording PO2 concentration
with a temporal resolution of 225 to 400 s per
frame found hypoxic micropockets in the cortex
of awake old mice using the phosphorescence
probe PtP-C343 (43, 44). However, the more
frequent hypoxic events reported lasted only
~50 s and could not be detected when imaging
with a slow, minute-lasting temporal resolu-
tion. Recently, 1/f-like fluctuations in PO2 have
been observed using Clark electrodes and
linked to red blood cell spacing heterogeneity,
in line with our observations (45). However,
PO2 declines with age (43, 46), and it would
be of interest to use the PO2 bioluminescence
imaging introduced here to assess whether
aging is linked to progressive increase in the
duration and/or spatial expansion of hypoxic
pockets. Another open question is whether the
transient hypoxic pockets contribute to the
noise observed in the functional magnetic res-
onance imaging (fMRI) blood oxygen level–
dependent (BOLD) signal. The hypoxic pockets
are below the spatial resolution of fMRI but are
likely contributing to the considerable noise
during rest (47). It is in this regard that our
analysis showed that the burden of the hypoxic
pockets decreased when neuronal activity was
increased, suggesting that the increase in BOLD
signal may in part reflect a drop in occurrence
of hypoxic pockets. This conclusion is sup-
ported by the finding that increase in capil-
lary blood flow reduces the relative portion of
capillary stalls (19).
Monitoring PO2 with BLI is limited to de-
tection of short-term fluctuations, excelling
at measuring relative changes over periods of
minutes rather than providing accurate baseline
PO2 comparisons over extended durations or
across groups. It is important to note that al-
though BLI is effective for assessing relative
PO2, it does not offer absolute quantification.
Capillary blood flow is essential for supply-
ing the brain with O2 and glucose needed to
support the high metabolic demand associated
with normal brain function (48). Numerous
studies have shown a link between reduced
cerebral blood flow and cognitive decline (49–52),
including changes in microvasculature struc-
ture and flow (29, 43). The existence of non-
perfused capillaries were discovered decades
ago (53). More recently, transient disruptions
of flow, caused by neutrophil adhesion, at the
single capillary level were identified as a po-
tential mechanism that contributes to cerebral
blood flow changes driving neurological defi-
cits (54, 55). Increased capillary stalling has been
observed in models of Alzheimer’s disease (54),
raising questions about the long-term impact
of capillary stalling and its potential role in
long-term neuronal viability. Hypoxia-induced
increase in expression of hypoxia inducible
factor 1a (HIF1a) impairs plasticity by disrupt-
ing synaptic physiology and spatial memory
RESE ARCH | R E S E A R C H A R T I C L E S
(56). Our study predicts that physical inac-
tivity has direct effects on tissue PO2 by favor-
ing capillary occlusions and increasing the
number of hypoxic pockets (fig. S16). Con-
versely, simply increasing sensory input or
locomotion rapidly suppress the occurrence of
hypoxic pockets, perhaps explaining the linkage
between sedentary lifestyle and an increased
risk of dementia (57, 58).
REFERENCES AND NOTES
1. S. S. Kety, in Metabolism of the Nervous System, D. Richter, Ed.
(Pergamon, 1957), pp. 221–237.
2. D. D. Clarke, L. Sokoloff, in Basic Neurochemistry: Molecular,
Cellular and Medical Aspects, G. J. Siegel, B. W. Agranoff,
R. W. Albers, S. K. Fisher, M. D. Uhler, Eds. (Lippincott-Raven,
ed. 6, 1999), pp. 637–689.
3. D. F. Rolfe, G. C. Brown, Physiol. Rev. 77, 731–758 (1997).
4. T. Takano et al., Nat. Neurosci. 10, 754–762 (2007).
5. B. Li et al., eLife 8, e42299 (2019).
6. M. P. Hall et al., ACS Chem. Biol. 7, 1848–1857 (2012).
7. K. Suzuki et al., Nat. Commun. 7, 13718 (2016).
8. D. Lambrechts et al., PLOS ONE 9, e97572 (2014).
9. R. A. Pollock, R. T. Jackson, A. A. Clairmont, W. L. Nicholson,
Arch. Otolaryngol. 100, 309–313 (1974).
10. J. Lee, T. Taira, P. Pihlaja, B. R. Ransom, K. Kaila, Brain Res.
706, 210–216 (1996).
11. Y. Shen, M. Rosendale, R. E. Campbell, D. Perrais, J. Cell Biol.
207, 419–432 (2014).
12. S. Holstein-Rønsbo et al., Nat. Neurosci. 26, 1042–1053 (2023).
13. D. G. Lyons, A. Parpaleix, M. Roche, S. Charpak, eLife 5,
e12024 (2016).
14. A.-K. Aydin, C. Verdier, E. Chaigneau, S. Charpak, Proc. Natl.
Acad. Sci. U.S.A. 119, e2200205119 (2022).
15. K. Xu, D. A. Boas, S. Sakadžić, J. C. LaManna, Adv. Exp.
Med. Biol. 977, 149–153 (2017).
16. Y. Ma et al., Proc. Natl. Acad. Sci. U.S.A. 113, E8463–E8471 (2016).
17. X. Zhu et al., Light Sci. Appl. 11, 138 (2022).
18. A. R. Nippert, K. R. Biesecker, E. A. Newman, Neuroscientist 24,
73–83 (2018).
19. A. Villringer, A. Them, U. Lindauer, K. Einhäupl, U. Dirnagl,
Circ. Res. 75, 55–62 (1994).
20. D. A. Brodie, D. M. Woodbury, Am. J. Physiol. 192, 91–94
(1958).
21. R. F. Cucchiara, R. A. Theye, J. D. Michenfelder, Anesthesiology
40, 571–574 (1974).
22. S. Gelman, K. C. Fowler, L. R. Smith, Anesth. Analg. 63,
557–565 (1984).
23. L. A. Low, L. C. Bauer, B. A. Klaunberg, PLOS ONE 11, e0154936
(2016).
24. D. Aksenov, J. E. Eassa, J. Lakhoo, A. Wyrwicz,
R. A. Linsenmeier, Neurosci. Lett. 524, 116–118 (2012).
25. Ş. E. Erdener et al., J. Cereb. Blood Flow Metab. 39, 886–900
(2019).
26. C. Iadecola, Neuron 96, 17–42 (2017).
27. O. Bracko, J. C. Cruz Hernández, L. Park, N. Nishimura,
C. B. Schaffer, J. Cereb. Blood Flow Metab. 41, 1501–1516 (2021).
28. P. Reeson, B. Schager, M. Van Sprengel, C. E. Brown,
Front. Cell. Neurosci. 16, 876746 (2022).
29. P. Reeson, K. Choi, C. E. Brown, eLife 7, e33670 (2018).
30. D. S. Greenberg, A. R. Houweling, J. N. D. Kerr, Nat. Neurosci.
11, 749–751 (2008).
31. K. Masamoto, T. Obata, I. Kanno, Adv. Exp. Med. Biol. 662,
57–61 (2010).
32. Q. Zhang et al., Nat. Commun. 10, 5515 (2019).
33. V. Untiet et al., Nat. Commun. 14, 1871 (2023).
34. E. Chaigneau, M. Oheim, E. Audinat, S. Charpak, Proc. Natl.
Acad. Sci. U.S.A. 100, 13081–13086 (2003).
35. H. S. Wei et al., Neuron 91, 851–862 (2016).
36. J. Grutzendler, M. Nedergaard, Trends Neurosci. 42, 528–536
(2019).
37. Y. Qi, M. Roper, Proc. Natl. Acad. Sci. U.S.A. 118, e2021840118
(2021).
38. P. Blinder et al., Nat. Neurosci. 16, 889–897 (2013).
39. K. Ogata, T. Kosaka, Neuroscience 113, 221–233 (2002).
40. E. A. Bushong, M. E. Martone, Y. Z. Jones, M. H. Ellisman,
J. Neurosci. 22, 183–192 (2002).
41. K. A. Kasischke et al., J. Cereb. Blood Flow Metab. 31, 68–81
(2011).
29 MARCH 2024 • VOL 383 ISSUE 6690 1477
RES EARCH | R E S E A R C H A R T I C L E S

42. A. Krogh, J. Physiol. 52, 409–415 (1919). NEUROSCIENCE

43. M. Moeini et al., Sci. Rep. 8, 8219 (2018).
44.
45.
X. Lu et al., Neurobiol. Aging 88, 11–23 (2020).
Q. Zhang, K. W. Gheres, P. J. Drew, PLOS Biol. 19, e3001298 (2021). Selection of experience for memory by hippocampal
46. C. Mathiesen, A. Brazhe, K. Thomsen, M. Lauritzen, J. Cereb.

47.
Blood Flow Metab. 33, 161–169 (2013).
S.-G. Kim, S. Ogawa, J. Cereb. Blood Flow Metab. 32,
sharp wave ripples
1188–1206 (2012).
48. C. Iadecola, Neuron 80, 844–866 (2013). Wannan Yang1,3, Chen Sun2, Roman Huszár1,3, Thomas Hainmueller1,4, Kirill Kiselev3, György Buzsáki1,3*
49. Y. Iturria-Medina et al., Nat. Commun. 7, 11934 (2016).
50. J. C. de la Torre, Stroke 33, 1152–1162 (2002).
51. R. N. Kalaria, Neurobiol. Aging 21, 321–330 (2000). Experiences need to be tagged during learning for further consolidation. However, neurophysiological
52. E. Farkas, P. G. Luiten, Prog. Neurobiol. 64, 575–611 (2001). mechanisms that select experiences for lasting memory are not known. By combining large-scale neural
53. W. Kuschinsky, O. B. Paulson, Cerebrovasc. Brain Metab. Rev. recordings in mice with dimensionality reduction techniques, we observed that successive maze
4, 261–286 (1992).
54. J. C. Cruz Hernández et al., Nat. Neurosci. 22, 413–420 (2019).
traversals were tracked by continuously drifting populations of neurons, providing neuronal signatures
55. Ş. E. Erdener et al., J. Cereb. Blood Flow Metab. 41, 236–252 (2021). of both places visited and events encountered. When the brain state changed during reward
56. A. Arias-Cavieres et al., eNeuro 7, ENEURO.0024-20.2020 consumption, sharp wave ripples (SPW-Rs) occurred on some trials, and their specific spike content
(2020).
decoded the trial blocks that surrounded them. During postexperience sleep, SPW-Rs continued to
57. S. Yan et al., Transl. Psychiatry 10, 112 (2020).
58. D. A. Raichlen et al., JAMA 330, 934–940 (2023). replay those trial blocks that were reactivated most frequently during waking SPW-Rs. Replay content of
59. F. R. M. Beinlich et al., Oxygen Imaging of Hypoxic Pockets in awake SPW-Rs may thus provide a neurophysiological tagging mechanism to select aspects of
the Mouse Cerebral Cortex [Data set], version 0.240215.0831, experience that are preserved and consolidated for future use.
DANDI archive (2024); https://doi.org/10.48324/
dandi.000891/0.240215.0831.
60. A. Asiminas, Temp2Spec/OxygenDynamics_Public:

Science_2024, version Science_2024, Zenodo (2024);
https://doi.org/10.5281/zenodo.10629901.
ACKN OW LEDG MEN TS
We thank T. Nagai and S. Inagaki (Department of Biomolecular
Science and Engineering, the Institute of Scientific and Industrial
Research, Osaka University, Japan) for providing GeNL plasmid
and helpful technical discussion. We also thank D. Xue for expert
assistance with the illustrations, Y. Zhao for assistance with the
data repository, and S. Goldman for comments on the manuscript.
Funding: This work was supported by the Dr. Miriam and Sheldon
G. Adelson Medical Research Foundation (M.N.); National Institutes
of Health grant R01AT012312 (M.N.); NINDS R01AT011439 (M.N.);
U19 NS128613 (M.N.); the Simons Foundation (M.N.); Novo Nordisk
Foundation NNF19OC0058058 (H.H.); Independent Research
Fund Denmark 0134-00107B (H.H.); Novo Nordisk Foundation
NNF20OC0066419 (M.N.); the Lundbeck Foundation R386-2021-
165 (M.N.); JPND/HBCI 1098-00030B (M.N.); JPND/Good Vibes
2092-00006B (M.N.); DOD W911NF2110006 (M.N.); the Lundbeck
Foundation R287-2018-2046 (V.U.); Independent Research Fund
Denmark 0134-00107B (H.H.); Independent Research Fund
Denmark 3101-00282B (M.N.); US Army Research Office grants
MURI W911NF1910280 (M.N.); Marie Skłodowska-Curie Fellowship
ANCoDy 101064009 (A.A.); and an ONO Rising Star Fellowship
(A.A.). Author contributions: Conceptualization: F.R.M.B., H.H.,
and M.N. Methodology: F.R.M.B., A.A., V.U., Z.B., V.P., V.T.,
L.G., B.S., M.H.G., H.H., and M.N. Software: A.A., F.R.M.B., B.S.,
V.T., L.G., and M.H.G. Formal analysis: F.R.M.B., A.A., V.P.,
Z.B., V.T., L.G., M.H.G., and B.S. Investigation: F.R.M.B. and Z.B.
Visualization: F.R.M.B. and A.A. Funding acquisition: H.H., M.N.,
V.U., and A.A. Project administration: F.R.M.B. and M.N. Supervision:
F.R.M.B., M.H.G., H.H., and M.N. Writing – original draft: F.R.M.B.,
A.A., H.H., and M.N. Writing – review & editing: F.R.M.B., A.A., H.H.,
and M.N. Competing interests: The authors declare that they have
no competing interests. Data and materials availability: The
microscope recordings are available at the DANDI archive (59). All
data are available in the manuscript and the supplementary
materials as data S1. Source code is deposited in Zenodo (60).
License information: Copyright © 2024 the authors, some rights
reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adn1011
Materials and Methods
Figs. S1 to S16
Tables S1 and S2
References (61–68)
MDAR Reproducibility Checklist
Movies S1 to S4
Data S1
Submitted 29 November 2023; accepted 23 February 2024
10.1126/science.adn1011
S
ome episodic events experienced during
the day are further consolidated during
sleep, whereas others are discarded (1).
Remembering events depends on pro-
cesses that occur both immediately after
encoding and thereafter during sleep (2–4).
However, no unified solutions have yet emerged
for brain mechanisms that assign which ex-
periences are selected for storage and which
are eliminated. A molecular candidate for syn-
aptic tagging has been proposed (5) as an eli-
gibility trace (6), but no neurophysiological
mechanism has been identified for online se-
lection of experience. Awake sharp wave ripples
(SPW-Rs) in the hippocampal system (3, 7–8)
are a potential candidate for selecting partic-
ular aspects of experience for future use (9). In
support of this hypothesis, salient features of
the experience, such as novelty and reward mag-
nitude, facilitate replay of multiple aspects
of waking experience and enhance memory
(2, 9, 10). To hold confounding external bias-
ing factors (such as reward or novelty) constant
and identify whether SPW-Rs assign credit to
particular events, we examined which events
within a session were selected by waking SPW-
Rs and continued to be replayed repeatedly
during sleep SPW-Rs. Recordings taken from
many hundreds of neurons simultaneously
in the dorsal CA1 region of the hippocampus
(n = 4469 cells from six mice) with dual-side
silicon probes (Fig. 1A) (11) and the application
of advanced analysis tools allowed us to visual-
ize and decode perpetually changing spike
contents of successive trials during experience
and relate them to spike sequences during
both awake and sleep SPW-Rs.
1
Neuroscience Institute and Department of Neurology, NYU
Grossman School of Medicine, New York University, New
York City, NY, USA. 2Mila - Quebec AI Institute, Montréal,
Quebec, Canada. 3Center for Neural Science, New York
University, New York City, NY, USA. 4Department of
Psychiatry, New York University Langone Medical Center,
New York City, NY, USA.
*Corresponding author. Email: gyorgy.buzsaki@nyulangone.org
To extract the sequential structure embed-
ded in the spike data, we used sequence non-
negative matrix factorization (seqNMF) (12).
seqNMF identified robust patterns that matched
the behavioral events of mice in the figure-
eight maze (Fig. 1, B to D, and fig. S1). To fur-
ther examine the differences in the sequences
between different events, we used uniform
manifold approximation and projection (UMAP)
to embed the same high-dimensional data in
a low-dimensional space (13). As expected,
qualitative UMAP visualization (Fig. 1E) re-
vealed that population activity of the hippo-
campus corresponded to a latent space that
topologically resembled the physical environ-
ment (14, 15). Notably, a prominent progres-
sion of states that corresponded to trial
sequences was observed after color-labeling
the manifold with trial block numbers (Fig. 1F)
(16). This observation was consistent across
different maze types and rodent species (fig.
S2) (17).
Within-session temporal evolution of neuronal
population activity
To quantify the trial sequence information
(18) present in the state space, we tested if trial
block membership can be accurately decoded
from the population activity (13, 19). Succes-
sive five trials composed a trial block, and a
specific label was assigned to each trial block
(Fig. 1, G and H). Decoding was performed
using a k-nearest neighbor (kNN) decoder,
followed by 10-fold cross-validation in the
original high-dimensional space (fig. S7B)
(materials and methods). Trial block member-
ship could be accurately decoded from the
original high-dimensional space (Fig. 1I) as
well as from the low-dimensional UMAP em-
bedding (fig. S3, C and E). These results were
further confirmed by using principal compo-
nent analysis (PCA) for dimensionality reduc-
tion or by using Bayesian decoding (Fig. 1J and
fig. S3), although these classical methods did

1478 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

A C R L D E E
1

Position
1 2 3 32

0
422

B L R

Cell no.
1
Position

0 0
0 50 0 250 500 750
Time (s) Time (s)

E Maze trajectory Neural manifold G Neural manifold I

1

Actual trial no.

1-5
1

Probability
Position
65-70 0
0 1-5 65-70
Predicted trial no.
(five-trial block)

F H J 2
65-70

Decoding error
65-70
Trial no. 1
70 1

AP

ye A
an

D
S
1-5 1-5

O
Ba C

EU
si
M

Fig. 1. Trial block identity can be decoded from the distinct temporal
evolution of neuronal population activity. (A) Illustration of two (out of four)
shanks of the dual-sided probe. (B) The figure-eight maze task, where mice
alternated between left and right arms for water reward (blue droplets). The
mouse’s trajectory along maze corridors is color coded by its linearized position.
(C) Trials 1 (right arm traversal) and 2 (left arm traversal) during the figure-eight
maze task. (Top) The linearized position of the mouse. Red, right traversal;
green, left traversal. (Bottom) A raster plot of 422 pyramidal cells that were
simultaneously recorded from the right dorsal CA1 region of a mouse’s
hippocampus, sorted by seqNMF, and color coded according to the linearized
position. This follows the same color scheme as in (B). (D) Trials 3 to 31
[of 70 trials in total (fig. S1)] of an example session. E, error trials. (E) (Left)
Running trajectory of a mouse in the figure-eight maze. (Right) UMAP embedding
of population activity. Each point corresponds to the low-dimensional
C
U
representation of one-binned spiking data. Both are colored according to the
mouse’s position, as in (B). (F) The same session as (E), but the running
trajectory and neural manifold are colored by trial block number [see color
key in (H)]. (G) Neural manifold of the same session as in (E) and (F) made by
using semisupervised UMAP trained on blocks of five trials. The manifold is
colored by linearized position. (H) Same as (G), but the neural manifold is
colored by trial block number. (I) Confusion matrix of trial block decoding from
the original high-dimensional space [same session as (A) to (H)]. (J) Trial
block identity decoding errors from across all sessions using UMAP, PCA,
Bayesian decoding, decoding from original high-dimensional space with Euclidean
distance (EUD), and cosine similarity (COS) as distance metrics. Each dot in
the violin plot indicates one session, pooled across n = 26 sessions from
six animals. Decoding error was measured in units of trial block in which five
trials were binned to one trial block.

not generate intuitive visualization for the
initial hypothesis generation process. Nota-
bly, we confirmed that UMAP embedding in
the low-dimensional space preserved the trial
block membership of the data compared with
that in the original high-dimensional space,
yielding consistent trial-decoding results be-
tween the low- and high-dimensional spaces
(fig. S4). Nevertheless, the main summary
statistics (across the entire dataset) through-
out the paper were performed in the original
high-dimensional space.
Intertrial variation of population activity
is not random
To probe the contribution of the individual
neurons to the aforementioned neuronal pop-
ulation features, we first plotted the tuning
curves of individual neurons across trials. We

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1479

RES EARCH | R E S E A R C H A R T I C L E S

A Example cell B Example cell C D E F

1-5
1
Left arm

Actual trial no.

65-70
1
Firing rate (Hz)

early

Probability
65-70 2158 65-70

Actual data

Trial no.
Position
Trial no.

Time (s)
20

10 1-5 0 1-5
late 1-5

65-70
0
0
0 0 100 200
1-5 65-70
0 100 200

1-5
1

Actual trial no.

Right arm 2158
Simulated data

Probability
Firing rate (Hz)

1
Time (s)

65-70

Position

Trial no.
20
late
10 0
0 1-5 1-5

65-70
0 100 200
0 65-70 0
0 100 200 1-5 65-70
Position (cm) Predicted trial no.
Position (cm)
(five-trial block)

G H I N.S. J
Place cells Nonplace cells N.S.
3
***

Decoding error
N.S.

Decoding error
N.S.
8 2
Decoding error

late 2
65-70 N.S.
6 Shuffle
Trial no.

4 1 1
early
2 1-5
early middle 0
0
0

0
0
0
0
0
0
0
0
50
10
15
20
25
30
35
40
45
late All C C
Data Simulation PC M-P NPC -NP
S SM No. of cells

Fig. 2. Contribution of single cell and subpopulation of cells to the trial block
identity coding. (A) Trial-by-trial firing rate of an example neuron during left-arm
(top) and right-arm trials (bottom) in the figure-eight maze. Color indicates the trial
block number. (B) Raster plot of an example neuron (top) and a simulated
neuron (bottom), generated from a Poisson process model based on the across-trial
mean firing rate of the real neuron (materials and methods). (C) (Top) UMAP
manifold generated from the population activity of all the neurons in one example
session. (Bottom) UMAP embedding from the simulated population activity. Both
are colored by linearized position. (D) Same as (C) but colored by the trial block
number [see color key in (E)]. (E) Neural manifold of the same session as (C)
and (D) made by using the semisupervised UMAP trained on blocks of five trials.
(F) Confusion matrix of trial block decoding results for the real (top) and simulated
population (bottom). For decoding results from other decoding methods, see fig. S7, E to
G. (G) Trial block decoding error of real (purple) and simulated (green) data across all
sessions (decoded from the original high-dimensional space) (***P < 10−10, unpaired
t test; n = 26 sessions from six animals). The dashed line indicates the chance level
from trial-shuffled data (fig. S8). (H) Neural manifold generated from place cells (left)
and nonplace cells (right) from the same session as (A) to (F) by using semisupervised
UMAP trained on blocks of five trials. (I) Trial block decoding error of all cells (red),
place cells (PC), and its size-matched control from all cells (SM-PC, purple), as well as
nonplace cells (NPC) and their size-matched control (SM_NPC, blue). Decoding
error was measured in units of trial block in which successive five trials were binned to
one trial block. (N.S., not significant; unpaired t test; n = 26 sessions from six animals.)
(J) Decoding error when downsampling the cells in the example session to
subsamples of varying cell numbers (from 50 to 450 cells). Error bar indicates the
standard error of the mean (SE) across 1000 subsamples. Decoding error was
measured in units of trial block in which successive five trials were binned to one
trial block). The red line shows the fitted exponential curve.

found diverse patterns of intertrial variability: fold of the simulated data also reflected the that only the real data were decoded correctly
within-place field rate remapping (fig. S6, A topology of the maze (Fig. 2C). By contrast, to the preceding or succeeding trials, occupy-
and B) (20), place fields emerging on later trial block identity information across trials ing the space immediately next to the diagonal
trials (bottom, Fig. 2A) (21), and shifting place vanished (Fig. 2, D and E). As an alternative of the confusion matrix (1.20 and 7.84 mean
fields across trials (top, Fig. 2A, and fig. S6C) to the standard 10-fold cross-validation meth- error for real and simulated data for one ex-
(22–24). In principle, decodability of trial block od (Fig. 1I and fig. S7, B and D), we imple- ample session) (Fig. 2F and fig. S7, C and E to
membership could be explained by either sto- mented a targeted validation method (a “leave G). The trial decoding accuracy of the real data
chastic or structured variation across trials. one trial out” procedure) that was expected to was significantly higher than that of the sim-
To identify the source of variability, we built a yield high decoding accuracy only if the state ulated or trial-shuffled data (P < 10−10; 1.51,
model that generated stochastic fluctuation space of the data changed in a structured way, 4.07, and 5.34 mean error for real, simulated,
in firing rate across trials. Each simulated evolving along one axis according to the se- and shuffled data, respectively) (Fig. 2G and
cell’s spiking activity was generated through quence of trial events. Because the training set fig. S8). These findings indicate that inter-
a Poisson process based on the tuning curve did not contain any data that shared the trial trial variation of population activity (drift) can-
of real neurons (Fig. 2B) (materials and meth- block membership with the test set, the test not be captured by random fluctuations at the
ods). We passed the simulated spike trains data could decode only to the next closest trial single neuron level. By using four different de-
through the same dimensionality-reduction block (not the same trial block) in the state coding methods, we confirmed that the system-
pipeline as we did for the real data. The mani- space (fig. S7, A and C). Indeed, we observed atic trial-dependent variation of the neuronal

1480 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

A B C D
Distance to manifold
Noise cloud
Position

1
off-manifold SPW-Rs

Trajectory length
422

E ***

Proportion (%)
Maze manifold 60
Cell no.

40

0 1.147 20
Time (s)
0
00 30.1 Maze SPW-R Shuffle
Time (s) on-manifold SPW-Rs events events

F G H
Position manifold Maze trajectory Trial manifold Maze trajectory 0.8

Proportion
Position

65-70

Trial no.
3

1 0
7
6 2 3
5 3 7 1-5 0
4 1 2 4 567 6 0 1 2 3 4 5 6
54
2 (Actual trial block no. –
1
decoded trial block no.)

Fig. 3. Maze SPW-Rs replay of trial block identity. (A) Spiking activity of an
example trial in the figure-eight maze. (Top) Linearized position of the mouse.
(Bottom) Raster plot of neuronal spiking activity sorted by seqNMF. Purple stars,
SPW-Rs. (B) Zoomed-in display of a replay event. (Top) Local field potential (LFP)
filtered in the ripple band. (Bottom) Raster plot of neurons belonging to a sequence
factor with significant reactivation strength (materials and methods). Neurons were
sorted in the same order as in (A). (C) All waking SPW-R events (red) in this
example session were embedded with the neural manifold data during navigation
(light gray), and the noise cloud consisted of negative samples (dark gray). Two
clusters of SPW-R replay events were distinguished: off- and on-manifold events
(materials and methods) (figs. S12 and S13). (D) SPW-R replays were classified as
significant if they were (1) close to the maze manifold and (2) their trajectory length
along the manifold was short in comparison to shuffled data. (E) Percentage of
significant replays in the maze (number of significant maze replay events/total
number of maze SPW-R events) compared with shuffled data. (***P < 10−4,
unpaired t test; n = 26 sessions from six animals). (F) UMAP embedding and the
decoded position for the same event as in (B). (Left) The neural manifold during
maze running (“position manifold,” colored by position). (Right) The position content
of each replay time bin was decoded to a position bin along the maze trajectory
according to the position label of its nearest neighbor on the manifold. The black
triangle represents the physical location of the mouse when the replay event took
place, and the pink-purple dots represent the neural embedding of seven successive
time bins of a SPW-R replay event (each time bin was 20 ms). (G) (Left) The
same event was embedded with the “trial manifold” and was colored by trial
block number. (Right) Trial content of each replay time bin was decoded as the
trial block label of its nearest neighbor on the manifold (each trial block
corresponds to five trials). (H) Distribution of differences between the actual trial
block of SPW-Rs events and their decoded trial block identity across all sessions.

population activity could not be explained by
electrode drift (figs. S5, S6, and S9) (25).
Both place cells (PC) and nonplace cells
(NPC) contributed to trial block membership
decoding (Fig. 2H) [mean decoding errors were
0.1, 0.42, 0.49, 0.53, and 0.59 for all cells, PCs,
NPCs, and the size-matched controls of PCs
and NPCs, respectively (Fig. 2I)]. The neural
manifolds of different trials were aligned better
with place cells than with nonplace cells. De-
coding accuracy deteriorated rapidly when
downsampled to <100 neurons in a session, in-
dicating that trial block identity is encoded in
the population (Fig. 2J and figs. S10 and S11).
Waking SPW-Rs replay current experience
Place and trial sequence encoding is associated
with theta frequency (6 to 9 Hz) oscillations
(“theta state”) (26) as animals actively navigate
in an environment. When animals stop con-
suming the water reward, theta activity gives
way to synchronous population events, SPW-
Rs (Fig. 3, A and B). Because spike sequences
within SPW-Rs are known to replay place field
sequences (27, 28), we asked whether trial
block identity could also be decoded in replay
events. A candidate SPW-R event was classi-
fied as a significant replay if both its distance
to manifold and trajectory length were sig-
nificantly shorter than those of shuffled dis-
tributions (Fig. 3, C and D, and figs. S12 to
S14) (materials and methods). The majority
of SPW-Rs occurred in the reward area (fig.
S15A), and about 33% of the awake SPW-Rs
were significant replays (Fig. 3E). From the
original high-dimensional space, we decoded
not only the spatial trajectory but also the trial
block identity of the significant replay events.
We validated the decoding results using four
different methods, including decoding from
the original high-dimensional space with dif-
ferent distance metrics and decoding from
the low-dimensional space with PCA and
UMAP. Different methods yielded consistent
decoding results (Fig. 3, F to G, and figs. S16
and S18, A to D). The within-event coherence
of trial replay was high (different time bins
within the same events coherently decoded to
the same trial block) but decreased rapidly
as we downsampled the number of cells in-
cluded for decoding (fig. S18, E to G). Next,
we examined whether the SPW-Rs replayed the
past, future, or current trial block. The spike
content of SPW-Rs decoded most reliably to

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1481

RES EARCH | R E S E A R C H A R T I C L E S

A B D Z
Z
Z
F
1
***
Position

*** ***
0

Proportion (%)
0 1. 1 3 8 0 1.1 0 2
318 Time (s) Time (s) 20

C E 25-30
Cell no.

Trial no.
1
65432 54321

0
ee
p ep le
-sl sle uff
st- Sh
1-5 Pre Po
0
0 90.1 0
Time (s)

G H I J
Proportion

R = 0.86 K
0.2 P= 3.1*10-34 1 1 *
Post sleep trial proportion

Maze

Proportion
1

Beta coefficient
0.5
0 ***

Proportion
1- 5 45-50
Proportion

0.5 0
0.5
0.2
0.5

0 Post-sleep

Proportion
1- 5 45-50 0 0.5
Proportion

re t

fle
Sh lay
Pr c a p lay
ee co r
p un
sl le e
0 e- yc ow ep le

uf
0.4

p
uff
et et rep

0 0.5 1 t- sle
0 os Sh
Th Th ze

0.2 Maze trial block distribution Left Right Sp

a

V
M

0 Arm ze
1-5 4 5- 5 0 ma
Trial no.
Maze replay Post-sleep replay Pre-sleep replay

Fig. 4. Replay of trial block identity and maze segments during sleep can unpaired t test; n = 26 sessions from six animals). (G) Distribution of the trial blocks
be predicted from waking SPW-Rs in the maze. (A) Spiking activity in the decoded from the population spike content of SPW-Rs in an example session. (Top)
figure-eight maze. (Top) Linearized position of the mouse. (Bottom) Raster Maze and postsleep replay trial block distribution pattern. (Bottom) Maze and
plot of neuronal spikes, sorted by seqNMF. Purple stars represent SPW-Rs. presleep replay trial block distribution pattern. (H) Correlation between trial block
(B) Zoomed-in display of an awake replay event in the maze. The raster plot distributions across trial blocks during maze and postsleep SPW-Rs replay (decoded
contains neurons belonging to the sequence factor with significant reactivation from the original high-dimensional space; Pearson correlation coefficient, R = 0.86;
strength (materials and methods). Neurons were sorted in the same order P < 3.1 × 10−34; n = 16 sessions from five animals) (results from all four different
as in (A). (C) Decoded position and trial block identity of successive 20-ms bins decoding methods, see fig. S18 ). (I) The predictive relationship between trial block
(one to six bins) of the same SPW-R replay event in (B). The black triangle distribution patterns of postsleep SPW-Rs and other candidate variables, including
represents the physical location of the mouse when the replay event took place. the trial block distribution patterns of theta cycle, theta power, presleep, and trial-
(D) Raster plot of a SPW-R replay event during sleep in the home cage. shuffled data (***P < 10−23 for awake replay; ***P < 10−3 for theta cycle number;
(E) Decoded position and trial block identity of successive 20-ms bins of the the relative predictive power of a given metric was considered nonsignificant when
same SPW-R replay event in panel (D). (A), (B), and (D) were colored according it overlapped with zero; n = 16 sessions from five animals) (for results from all four
to the linearized position of the mouse; (C) and (E) were colored according to different decoding methods, see fig. S18). (J) Proportion of maze segment replays (left
trial block number. (F) Percentage of significant replays during pre- and arm and right arm) during awake and postexperience sleep SPW-Rs in an example
postexperience sleep (presleep and postsleep, respectively) in the home cage session. (K) Proportion of sessions with same rank order between maze segments
compared with shuffled data. (Presleep versus postsleep, ***P < 10−5; postsleep during awake and post-maze sleep SPW-R replays, compared with shuffle data.
versus shuffled data, ***P < 10−8; presleep versus shuffled data, ***P < 10−8; (*P < 0.05, chi-square test; n = 13 sessions from five animals).

the present trial block (Fig. 3H) (4, 28–30). during postexperience sleep were highly cor- maze (P < 10−26, P = 0.06, P < 10−4, P = 0.89,
The strongest predictor of the trial block dis- related with that during maze replay, but not and P = 0.87 for in-maze replay, theta power,
tribution pattern for maze replay was the im- with preexperience sleep replay (R = 0.86, P < theta cycle count, presleep replay, and trial-
mobility time (fig. S19B). 10−36 for correlation between postexperience shuffled data, respectively) (Fig. 4I and fig.
sleep replay and maze replay) (Fig. 4, F to H, S19C) (32).
Postexperience sleep SPW-Rs replay events and figs. S17 and S19A). We compared which We examined whether replays of the left
selected by waking SPW-Rs factor best explained the trial block distribu- versus right arms during waking and post-
To examine the relationship between awake tion pattern during postexperience sleep by sleep SPW-Rs were correlated, exploiting the
SPW-Rs and those during postexperience sleep, using a mixed-effect linear regression anal- natural variability in the replay distribution
we compared the population activity of awake ysis. All decoding methods consistently showed patterns across different arms in different ses-
SPW-Rs in the maze with that during postex- that the strongest predictor for the trial block sions (Fig. 4J and fig. S15B). The correlation of
perience sleep in the home cage (Fig. 4, A to E). distribution pattern during postexperience maze-arm replays between wake and sleep
The distribution patterns of trial block identity sleep was that of the SPW-R replay in the SPW-Rs was significantly higher than in the

1482 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


shuffled data (Fig. 4K). The difference in replay
proportion across left and right arms could not
be explained by the difference in the decod-
ability (measured by spatial information) or the
difference in coverage (measured by number of
visits) between the two arms (fig. S20).
Waking SPW-Rs weigh and select the
experience, and sleep SPW-Rs consolidate it
Exploration of the environment is a regular al-
ternation between ambulation and rest phases,
enabling brain state changes (33–35), and re-
playing aspects of experience during SPW-Rs
(27, 28, 30). We hypothesize that waking SPW-Rs
represent a natural credit assignment (tag-
ging) mechanism of experiences (3, 7–9). They
tag selected neuronal patterns, possibly by
comparing them to previous experience and
relevance to the animal, and the tagged pat-
terns are reactivated numerous times during
SPW-Rs of postexperience sleep to consoli-
date the selected experience and combine it
with the existing knowledge base of the brain
(4, 29, 36, 37). Waking SPW-Rs may trigger
molecular mechanisms to induce long-term
synaptic changes (5, 38). At the functional
level, they may create a brain state–dependent
attractor. Consequently, when the hippocam-
pal network returns to a similar brain state,
such as non–rapid eye movement (NREM) sleep,
it continues generating patterns set forth by
the waking SPW-R attractor.
Previous observations support this scenario.
Cues relevant to future behavioral success can
bias the neuronal trajectories of SPW-Rs (39–41).
Perturbation of SPW-Rs prevents place field
and experience stabilization (7, 8, 42–48). In
humans, items for which electrophysiological
brain patterns are not reactivated during non-
REM sleep are forgotten. Replay can occur after
a single experience and the number of SPW-R
after learning predicts subsequent memory
performance (3, 31, 46, 49–51).
Salient features of the experience, such as
novelty (9, 52, 53), repetition (9), and reward
(10), increase the incidence of SPW-Rs and re-
plays (54, 55). In our experiments, we could
exclude differences in external factors that
could shape sleep replay patterns by demon-
strating distinctly changing neuronal popula-
tion activity patterns within the same maze
and session. Selective tagging of trial events by
waking SPW-Rs could occur because the per-
petually evolving population activity patterns
during the theta state serve as a temporal scaf-
fold to organize experiences (56) and differen-
tiate successive events. The sequence of theta
state–dependent turnover of neuronal assem-
blies, their selection by waking SPW-Rs, and the
numerous repetitions during sleep SPW-Rs may
represent a hippocampal network mechanism
of selective episodic memory consolidation.
Our observations suggest that rewards pro-
vide an affordance for shifting brain states
SCIENCE science.org
and that the waking SPW-R serves as a neuro-
physiological mechanism for memory selection
(49, 56). These findings establish a neurophys-
iological framework for multiple domains of
systems neuroscience, including credit assign-
ment (28, 37), representational drift (57), “event
remapping” (18), time coding (16, 58–60), and
memory editing (61). Our work also links a
shared principle of memory processing impor-
tant for both biological and artificial learning.
In particular, it relates to importance sampling
(62, 63) in machine learning, which enables
faster acquisition (64) and more robust gener-
alization (65).
RE FERENCES AND NOTES
1. M. Inostroza, J. Born, Annu. Rev. Neurosci. 36, 79–102 (2013).
2. A. Tambini, N. Ketz, L. Davachi, Neuron 65, 280–290 (2010).
3. B. P. Staresina, A. Alink, N. Kriegeskorte, R. N. Henson,
Proc. Natl. Acad. Sci. U.S.A. 110, 21159–21164 (2013).
4. J. L. McClelland, B. L. McNaughton, R. C. O’Reilly, Psychol. Rev.
102, 419–457 (1995).
5. U. Frey, R. G. M. Morris, Nature 385, 533–536 (1997).
6. R. S. Sutton, A. G. Barto, Reinforcement Learning: An
Introduction (MIT Press, ed. 2, 2018).
7. S. P. Jadhav, C. Kemere, P. W. German, L. M. Frank, Science
336, 1454–1458 (2012).
8. A. Fernández-Ruiz et al., Science 364, 1082–1086 (2019).
9. M. Huelin Gorriz, M. Takigawa, D. Bendor, Nat. Commun. 14,
8157 (2023).
10. H. Igata, Y. Ikegaya, T. Sasaki, Proc. Natl. Acad. Sci. U.S.A. 118,
e2011266118 (2021).
11. R. Huszár, Y. Zhang, H. Blockus, G. Buzsáki, Nat. Neurosci. 25,
1201–1212 (2022).
12. E. L. Mackevicius et al., eLife 8, e38471 (2019).
13. L. Mcinnes, J. Healy, J. Melville, UMAP: Uniform Manifold
Approximation and Projection for Dimension Reduction.
arXiv:1802.03426 [stat.ML] (2020).
14. W. Tang, J. D. Shin, S. P. Jadhav, Cell Rep. 42, 112246 (2023).
15. W. Guo, J. J. Zhang, J. P. Newman, M. A. Wilson, bioRxiv
2020.02.27.967794 [Preprint] (2020); https://doi.org/
10.1101/2020.02.27.967794.
16. A. Tsao et al., Nature 561, 57–62 (2018).
18. C. Sun, W. Yang, J. Martin, S. Tonegawa, Nat. Neurosci. 23,
651–663 (2020).
19. S. Schneider, J. H. Lee, M. W. Mathis, Nature 617, 360–368 (2023).
20. A. A. Fenton, R. U. Muller, Proc. Natl. Acad. Sci. U.S.A. 95,
3182–3187 (1998).
21. J. D. Monaco, G. Rao, E. D. Roth, J. J. Knierim, Nat. Neurosci.
17, 725–731 (2014).
22. M. R. Mehta, M. C. Quirk, M. A. Wilson, Neuron 25, 707–715 (2000).
23. I. Lee, G. Rao, J. J. Knierim, Neuron 42, 803–815 (2004).
24. I. Lee, A. L. Griffin, E. A. Zilli, H. Eichenbaum, M. E. Hasselmo,
Neuron 51, 639–650 (2006).
26. C. H. Vanderwolf, Electroencephalogr. Clin. Neurophysiol. 26,
407–418 (1969).
27. K. Diba, G. Buzsáki, Nat. Neurosci. 10, 1241–1242 (2007).
28. D. J. Foster, M. A. Wilson, Nature 440, 680–683 (2006).
29. G. Buzsáki, Neuroscience 31, 551–570 (1989).
30. J. Widloski, D. J. Foster, Neuron 110, 1547–1558.e8 (2022).
31. J. O’Neill, T. Senior, J. Csicsvari, Neuron 49, 143–155 (2006).
32. C. Drieu, R. Todorova, M. Zugaro, Science 362, 675–679 (2018).
33. A. V. Samsonovich, G. A. Ascoli, Learn. Mem. 12, 193–208 (2005).
34. D. Drai, N. Kafkafi, Y. Benjamini, G. Elmer, I. Golani, Behav.
Brain Res. 125, 133–140 (2001).
35. D. L. Kramer, R. L. McLaughlin, Am. Zool. 41, 137–153 (2001).
36. G. Buzsáki, Hippocampus 25, 1073–1188 (2015).
37. H. R. Joo, L. M. Frank, Nat. Rev. Neurosci. 19, 744–757 (2018).
38. S. A. Josselyn, S. Tonegawa, Science 367, eaaw4325 (2020).
39. D. Bendor, M. A. Wilson, Nat. Neurosci. 15, 1439–1444 (2012).
40. B. Rasch, C. Büchel, S. Gais, J. Born, Science 315, 1426–1429 (2007).
41. J. D. Rudoy, J. L. Voss, C. E. Westerberg, K. A. Paller, Science
326, 1079 (2009).
42. L. Roux, B. Hu, R. Eichler, E. Stark, G. Buzsáki, Nat. Neurosci.
20, 845–853 (2017).
43. I. Gridchyn, P. Schoenenberger, J. O’Neill, J. Csicsvari, Neuron
106, 291–300.e6 (2020).
44. G. Girardeau, K. Benchenane, S. I. Wiener, G. Buzsáki,
M. B. Zugaro, Nat. Neurosci. 12, 1222–1223 (2009).
RESE ARCH | R E S E A R C H A R T I C L E S
45. Y. Zhang et al., Proc. Natl. Acad. Sci. U.S.A. 118, e2016432118
(2021).
46. A. C. Singer, M. F. Carr, M. P. Karlsson, L. M. Frank, Neuron 77,
1163–1173 (2013).
47. A. D. Grosmark, F. T. Sparks, M. J. Davis, A. Losonczy,
Nat. Neurosci. 24, 1574–1585 (2021).
48. T. Schreiner, C. F. Doeller, O. Jensen, B. Rasch, T. Staudigl,
Cell Rep. 25, 296–301 (2018).
49. A. Berners-Lee et al., Neuron 110, 1829–1842.e5 (2022).
50. W. Ramadan, O. Eschenko, S. J. Sara, PLOS ONE 4, e6697 (2009).
51. N. Axmacher, C. E. Elger, J. Fell, Brain 131, 1806–1817 (2008).
52. B. Giri, H. Miyawaki, K. Mizuseki, S. Cheng, K. Diba, J. Neurosci.
39, 866–875 (2019).
53. S. Cheng, L. M. Frank, Neuron 57, 303–313 (2008).
54. M. F. Carr, S. P. Jadhav, L. M. Frank, Nat. Neurosci. 14, 147–153
(2011).
55. A. S. Gupta, M. A. A. van der Meer, D. S. Touretzky,
A. D. Redish, Neuron 65, 695–705 (2010).
56. A. Rubin, N. Geva, L. Sheintuch, Y. Ziv, eLife 4, e12247 (2015).
57. G. Buzsáki et al., in Temporal Coding of the Brain (Springer,
1994), pp. 145–172.
58. B. Shahbaba et al., Nat. Commun. 13, 787 (2022).
59. M. E. Montchal, Z. M. Reagh, M. A. Yassa, Nat. Neurosci. 22,
284–288 (2019).
60. M. W. Howard, M. J. Kahana, J. Math. Psychol. 46, 269–299 (2002).
61. P. Dayan, L. F. Abbott, Theoretical Neuroscience:
Computational and Mathematical Modeling of Neural Systems
(The MIT Press, 2001).
62. G. Alain, A. Lamb, C. Sankar, A. Courville, Y. Bengio, Variance
Reduction in SGD by Distributed Importance Sampling. arXiv:1511.
06481 [stat.ML] (2013).
63. T. Schaul, J. Quan, I. Antonoglou, D. Silver, Prioritized
Experience Replay (, 2016).
64. P. Zhao, T. Zhang, Proc. Mach. Learn. Res. 37, 1–9 (2015).
65. C. Sun et al., Contrastive introspection to identify critical steps in
reinforcement learning. arXiv:2210.05845 [cs.LG] (2022).
66. D. Levenstein et al., winnieyangwannan/buzcode, version 1,
Zenodo (2024); https://doi.org/10.5281/zenodo.10685428.
67. W. Yang, winnieyangwannan/Selection-of-experience-for-
memory-by-hippocampal-sharp-wave-ripples, Zenodo (2024);
https://doi.org/10.5281/zenodo.10685490.
AC KNOWLED GME NTS
We thank K. Tewatia and J. Paraiso for helping to solve technical
problems with data visualization and validate ripple detection
results. We thank A. Grosmark for sharing his data. We are grateful
for the insightful discussions with A. Williams and E. Hermansen
regarding UMAP and dimensionality reduction methods in
general. We thank E. Chinigo for her kind help with the centroid
displacement analysis. We thank I. Zutshi, M. Valero, M. Vöröslakos,
K. Mcclain, R. Swanson, Z. Zheng, C. Kemere, A. Kirsanov, and all
members of the Buzsaki lab for their insightful discussions and
feedback. We greatly appreciate D. Tingley, J. Carpenter, and
K. Hardcastle for their constructive feedback on the manuscript.
Funding: G.B. received funding from the National Institutes
of Health grants R01MH122391 and U19NS107616. Author
contributions: Conceptualization: W.Y. and G.B.; Methodology:
W.Y., C.S., and R.H.; Investigation: W.Y., R.H., and T.H.;
Visualization: W.Y., C.S., and K.K.; Funding acquisition: G.B.;
Project administration: G.B.; Supervision: G.B.; Writing – original
draft: W.Y., C.S., and G.B. Competing interests: The authors
declare that they have no competing interests. Data and
materials availability: All data are available in the main text and
supplementary materials or are deposited at Zenodo (66, 67).
Notated neurophysiological data are available at https://
dandiarchive.org/dandiset/000552/0.230630.2304, https://
buzsakilab.nyumc.org/datasets/GrosmarkAD/Achilles/Achilles_
10252013/, and https://buzsakilab.nyumc.org/datasets/
HainmuellerT/TH11_210605/. License information: Copyright ©
2024 the authors, some rights reserved; exclusive licensee
American Association for the Advancement of Science. No claim to
original US government works. https://www.science.org/about/
science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adk8261
Materials and Methods
Figs. S1 to S20
References (11–13, 17, 25, 61)
MDAR Reproducibility Checklist
Submitted 12 September 2023; accepted 23 February 2024
10.1126/science.adk8261
29 MARCH 2024 • VOL 383 ISSUE 6690 1483
RES EARCH | R E S E A R C H A R T I C L E S

METABOLISM rimidines, resulting in decreased abundance of

citrate and increased amounts of pyruvate (Fig.
Pyrimidines maintain mitochondrial pyruvate 1E and fig. S1I). Subjecting cells to 16 hours of
pyrimidine depletion did not exert a widespread
oxidation to support de novo lipogenesis impact on other metabolic pathways (fig. S2A
and table S2). Furthermore, this period of py-
Umakant Sahu1,2†, Elodie Villa1,2,†, Colleen R. Reczek2,3, Zibo Zhao1,2, Brendan P. O’Hara1,2, rimidine deprivation did not trigger any changes
Michael D. Torno1,2, Rohan Mishra4, William D. Shannon4, John M. Asara5, Peng Gao6, in the abundance of pyrimidine enzymes or
Ali Shilatifard1,2, Navdeep S. Chandel1,2,3, Issam Ben-Sahra1,2* in the activity of the mechanistic target of
rapamycin complex 1 (mTORC1), a critical regu-
Cellular purines, particularly adenosine 5′-triphosphate (ATP), fuel many metabolic reactions, but less lator of cellular metabolism (21, 22) (fig. S2,
is known about the direct effects of pyrimidines on cellular metabolism. We found that pyrimidines, B and C).
but not purines, maintain pyruvate oxidation and the tricarboxylic citric acid (TCA) cycle by regulating To assess whether changes in abundance of
pyruvate dehydrogenase (PDH) activity. PDH activity requires sufficient substrates and cofactors, pyruvate and citrate were associated with al-
including thiamine pyrophosphate (TPP). Depletion of cellular pyrimidines decreased TPP synthesis, a terations in glycolysis and the TCA cycle, we
reaction carried out by TPP kinase 1 (TPK1), which reportedly uses ATP to phosphorylate thiamine used stable isotope tracing of [13C6]-glucose or
(vitamin B1). We found that uridine 5′-triphosphate (UTP) acts as the preferred substrate for TPK1, [13C3]-pyruvate in control cells or cells deprived
enabling cellular TPP synthesis, PDH activity, TCA-cycle activity, lipogenesis, and adipocyte differentiation. of pyrimidines (Fig. 1F and fig. S3A). The citrate
Thus, UTP is required for vitamin B1 utilization to maintain pyruvate oxidation and lipogenesis. (M+2)/pyruvate (M+3) enrichment ratio de-
creased upon pyrimidine depletion, whereas
glucose entry remained unchanged (Fig. 1, G
yrimidine nucleotides are necessary for also supports glycosylation processes (13, 14), and H, and fig. S3, B and C). Pyrimidine de-

P nucleic acid synthesis, glycosylation, phos-

pholipid synthesis, cellular biomass, and
homeostasis (1–4). Pyrimidine produc-
tion occurs through two metabolic path-
ways: the de novo pyrimidine synthesis pathway
and the salvage pathway. The former results
from the assembly of the pyrimidine ring from
amino acids and other small molecules, where-
as the latter requires recycling of pyrimidine
nucleosides, such as uridine, from the sur-
rounding environment. These pathways are
distinct and separate processes (1, 5). Three
metabolic enzymes produce uridine mono-
phosphate (UMP) from de novo pyrimidine
synthesis.
The rate-limiting step of the de novo pyrim-
idine pathway is dihydroorotate synthesis from
glutamine, aspartate, bicarbonate, and adeno-
sine 5′-triphosphate (ATP) by the cytosolic tri-
catalytic enzyme CAD (carbamoyl-phosphate
synthetase 2, aspartate transcarbamoylase, and
dihydroorotase) (6, 7). Dihydroorotate dehydro-
genase (DHODH) oxidizes dihydroorotate in
the mitochondria to produce orotate, which
UMP synthase (UMPS) converts to UMP in
the cytosol (8–10). Cytidine triphosphate (CTP)
synthase (CTPS) converts UTP to CTP after
UMP phosphorylation.
UTP and CTP serve as building blocks for
nucleic acid synthesis (11, 12) (Fig. 1A). But UTP
1
Department of Biochemistry and Molecular Genetics,
Feinberg School of Medicine, Northwestern University,
Chicago, IL 60611, USA. 2Robert H. Lurie Comprehensive
Cancer Center, Northwestern University, Chicago, IL 60611,
USA. 3Department of Medicine, Northwestern University
Feinberg School of Medicine, Chicago, IL 60611, USA.
4 zyme phosphoribosylglycinamide formyltrans-
BioRankings, St. Louis, MO 63108, USA. 5Mass Spectrometry
Core, Beth Israel Deaconess Medical Center, Department of
Medicine, Harvard Medical School, Boston, MA 02115, USA.
6
Metabolomics Core Facility, Robert H. Lurie Comprehensive
Cancer Center, Northwestern University, Chicago, IL 60611, USA.
*Corresponding author. Email: issam.ben-sahra@northwestern.edu
†These authors contributed equally to this work.
and CTP functions in phospholipid synthesis
(15, 16) (Fig. 1A). Although purines, such as
adenylates (ADP, ATP), participate in energy
metabolism, including glycolysis and the tri-
carboxylic citric acid (TCA) cycle (17, 18), direct
effects of pyrimidines on central carbon me-
tabolism activity have not been described.
Requirement of cellular pyrimidines
for pyruvate catabolism
To understand how pyrimidine nucleotides
affect cellular metabolism, we performed tar-
geted metabolite profiling of human cervical
cancer cells (HeLa) treated with either sol-
vent or the DHODH inhibitor brequinar (BRQ)
at several time points (table S1). BRQ treatment
rapidly decreased abundance of pyrimidine
intermediates. Within 4 hours, citrate concen-
tration decreased, whereas pyruvate concentra-
tions increased (Fig. 1, B and C). By contrast, the
DHFR inhibitor methotrexate (MTX), which
affects one-carbon metabolism and purine and
thymidylate production, had no effect on steady-
state pyruvate and citrate levels (Fig. 1D). To
study pyrimidine depletion independent of
mitochondrial DHODH (19), we used CRISPR-
Cas9 to generate cells lacking CAD (DCAD) or
UMPS (DUMPS) human embryonic kidney
293E (HEK293E) and HeLa cell lines, respec-
tively (fig. S1, A and B), which showed a de-
crease in de novo pyrimidine synthesis, as
indicated by reduced 14C incorporation from
aspartate into RNA (fig. S1C). Uridine depletion
in DCAD or DUMPS cells reduced cell prolifer-
ation and resulted in uridine auxotrophy (fig.
S1, D and E). Depleting the de novo purine en-
ferase (GART) (DGART) (20) in HEK293E cells
inhibited de novo purine synthesis and in-
duced purine auxotrophy (fig. S1, F to H). Sim-
ilarly to DHODH inhibition, lack of uridine in
DUMPS HEK293E cells depleted cellular py-
pletion also decreased abundance of other
TCA-cycle metabolites such as a-ketoglutarate
(M+2), succinate (M+2), and fumarate (M+2)
in DCAD and DUMPS HEK293E cells (fig. S3C),
without noticeable changes in amounts of gly-
colytic intermediates (fig. S3, D through F). By
contrast, purine deprivation in DGART cells
decreased the abundance of glycolytic inter-
mediates (fig. S3G). Inhibition of de novo
pyrimidine synthesis with BRQ and the UMPS
inhibitor pyrazofurin (PRZ)—or genetic dele-
tion of CAD—consistently decreased pyruvate
catabolism and TCA-cycle activity in various
human cancer cell lines and mouse embryonic
stem (mES) cells (Fig. 1, I and J, and fig. S4, A
to E). However, depletion of purines did not
affect this process (fig. S4, F and G). To deter-
mine if the regulation of pyruvate catabolism
by pyrimidines is conserved across species, we
used Saccharomyces cerevisiae (yeast) and
Mus musculus (mouse) as model systems. Ura3-
deficient (ura3) S. cerevisiae cells, which are
uracil auxotrophs (fig. S4H), exhibited decreased
pyruvate catabolism upon acute (15-min) uracil
depletion (Fig. 1K and fig. S4, I and J). Mice
treated with a DHODH inhibitor also showed
lower liver pyruvate catabolism than did vehicle-
treated mice (Fig. 1, L and M). These findings
demonstrate that pyrimidine abundance can
influence pyruvate catabolism from unicellu-
lar yeast to mouse and human cells.
Cellular pyrimidines support pyruvate
dehydrogenase activity
The mitochondrial pyruvate dehydrogenase
(PDH) complex decarboxylates pyruvate into
acetyl–coenzyme A (acetyl-CoA), which enters
the TCA cycle. We assessed PDH activity by
measuring the release of [14C]-CO2 resulting
from the oxidation of [1-14C]-pyruvate (Fig. 2A).
Depletion of uridine in DUMPS or DCAD cells
decreased cellular PDH activity, which was

1484 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


Fig. 1. Depletion of cellular pyrimidines inhibits pyruvate catabolism.
(A) Illustration and metabolic role of the de novo pyrimidine synthesis pathway.
The DHODH inhibitor brequinar (BRQ) and UMPS inhibitor pyrazofurin (PRZ) are
indicated. (B) Steady-state metabolite profiles measured by liquid chromatography–
tandem mass spectrometry (LC-MS/MS) of HeLa cells treated with BRQ (1 mM).
(C) Glucose fate through glycolysis and TCA cycle. (D) Steady-state levels of
pyruvate, citrate, N-carbamoyl-L-aspartate, UMP, and ATP measured by LC-MS/MS
in HeLa cells treated with vehicle (DMSO, dimethyl sulfoxide), BRQ (1 mM), or
methotrexate (MTX, 4 mM) for 16 hours. (E) Same as shown in (D), but in DUMPS
HEK293E cells reconstituted or not with UMPS cDNA. (F) Carbon flow schematic
from [13C6]-glucose into the TCA cycle. (G) Fractional enrichment (%) of 13C-labeled
glucose (M+6) from [13C6]-glucose labeling in DUMPS HEK293E cells exposed to
uridine withdrawal time course. (H) Fractional enrichment (%) of citrate (M+2) from
SCIENCE science.org
RESE ARCH | R E S E A R C H A R T I C L E S
[13C6]-glucose labeling in DUMPS HEK293E cells reconstituted or not with UMPS
cDNA (I) Pyruvate catabolism from HEK293E cells labeled with [13C3]-pyruvate and
treated with vehicle (DMSO) or PRZ (1 mM) for 16 hours. (J) Pyruvate catabolism
and fractional enrichment (%) of 13C-labeled metabolites derived from [13C3]-
pyruvate in mouse embryonic stem cells treated as in (I). (K) Pyruvate catabolism
measured in S. cerevisiae ura3D cells labeled with 13C3-pyruvate for 30 min.
(L) Experimental design of the [13C6]-glucose isotope tracing in mice. (M) Measurement
of pyruvate catabolism in livers of mice treated with vehicle (PBS 70%, PEG-400
30%) or BRQ (20 mg/kg). Data in (M) are mean ± SD. n = 5 mice. For all the
other panels, data are mean ± SD. n = 3 independent replicates. *P < 0.05 for
multiple comparisons calculated using one-way analysis of variance (ANOVA) with
Tukey’s honest significant difference (HSD) test [(D), (E), (G), and (H)] or using a
two-tailed Student’s t test for pairwise comparisons [(I), (J), (K), and (M)].
29 MARCH 2024 • VOL 383 ISSUE 6690 1485
RES EARCH | R E S E A R C H A R T I C L E S
rescued by adding back uridine or reexpress-
ing UMPS in DUMPS cells (Fig. 2, B and C,
and fig. S5A). The pyrimidine-dependent de-
crease in PDH activity appeared not to result
from changes in abundance or protein products
of the genes that function in pyruvate me-
tabolism or glycolysis (Fig. 2D and fig. S5B).
Uridine-starved DCAD HEK293E cells showed
decreased acetyl-CoA (M+2) derived from py-
ruvate (M+3) (Fig. 2E). Similarly, acetyl-CoA
production was diminished in HeLa cells treated
with BRQ (fig. S5C) and in DCAD A375, DCAD
A549, and DCAD CAL-51 cancer cells depleted
of uridine (fig. S5D).
Because PDH and the TCA cycle generate
the reducing equivalent nicotinamide adenine
dinucleotide (NADH) in the mitochondria, we
investigated whether pyrimidine depletion al-
tered abundance of mitochondrial NAD(H).
DHODH inhibition decreased amounts of mito-
chondrial NADH but not those of NAD+ in HeLa
cells (Fig. 2, F and G). Similarly, genetic dele-
tion of UMPS or CAD decreased mitochon-
drial NADH, but not NAD+, in the absence of
exogenous uridine (Fig. 2, H and I, and fig.
S5E). In various DCAD cancer cell lines, elimi-
nation of uridine lowered the mitochondrial
NADH/NAD+ ratio (Fig. 2J and fig. S5, F to H).
Because NADH serves as an essential cofac-
tor for mitochondrial complex I activity in
the electron transport chain (3), we sought
to determine whether pyrimidine depletion
inhibited complex I activity. Pyrimidine de-
pletion caused by DHODH inhibition in wild-
type (WT) cells or uridine withdrawal in DUMPS
cells caused a significant decrease in mitochon-
drial complex I activity in live HEK293E cells,
which was recovered by UMPS cDNA expres-
sion (Fig. 2, K and L). Thus, pyruvate oxidation,
PDH activity, mitochondrial NADH mainte-
nance, and complex I activity all appear to re-
quire cellular pyrimidines.
Maintenance of thiamine pyrophosphate
synthesis in cells by pyrimidine nucleotides
To determine how pyrimidine depletion affects
PDH activity, we examined PDH phosphoryl-
ation by pyruvate dehydrogenase kinases (PDKs),
which reduces PDH activity (23, 24). However,
pyrimidines appeared to regulate PDH inde-
pendently of PDK (fig. S6A). We explored the
impact of pyrimidine depletion on CoA avail-
ability because CTP functions in CoA synthesis
(25, 26), but pyrimidine depletion did not af-
fect abundance of CoA (fig. S6B). Moreover,
pyrimidine depletion did not affect amounts
of newly synthesized CoA (M+4), as deter-
mined by stable isotope tracing with [13C3-15N]-
pantothenate (M+4) (fig. S6, C and D).
We conducted untargeted metabolite profil-
ing in DCAD HeLa cells grown with or without
uridine for 4 or 16 hours (Fig. 3A and tables S3
and S4). Changes in pyrimidine abundance
enriched the abundance of specific metabolites
1486 29 MARCH 2024 • VOL 383 ISSUE 6690
from various metabolic pathways, including
pyruvate, thiamine (vitamin B1), methionine,
tyrosine, and glucose metabolism (Fig. 3B).
Thiamine is a metabolite precursor required to
maintain pyruvate oxidation (27, 28). Thiamine
is phosphorylated by thiamine pyrophosphate
kinase 1 (TPK1) to produce thiamine pyrophos-
phate (TPP) (29), which functions in pyruvate
decarboxylation through PDH E1 activity (30).
TPP is also required for transketolase (TKT),
oxoglutarate dehydrogenase (OGDH), and
branched-chain alpha-keto acid dehydrogenase
(BCKDH) reactions (31) (Fig. 3C). Chemical or
genetic inhibition of pyrimidine synthesis sig-
nificantly decreased amounts of TPP in HeLa,
HEK293E, and mES cells (Fig. 3, D to F, and
fig. S6E). Genetic depletion of TPK1 decreased
abundance of TPP, pyruvate catabolism, and
amounts of the product of TKT sedoheptulose
7-phosphate (fig. S6, F to I). Thus, it appeared
that TPK1 activity might be regulated by py-
rimidine concentrations. To test this, we mon-
itored cellular TPP synthesis through stable
isotope tracing using [13C4]-thiamine labeling
(Fig. 3G). Depletion of uridine in DUMPS cells
decreased amounts of TPP (M+4) (Fig. 3H and
fig. S6J) without affecting TPK1 protein abun-
dance (fig. S6K). Additionally, BRQ suppressed
TPP synthesis, but MTX did not (Fig. 3I), indi-
cating that uridine or cytidine nucleotides gov-
ern TPP synthesis. Mice treated with BRQ or
PRZ also showed decreased abundance of TPP
in liver and a decrease in UTP without any
change in abundance of ATP (Fig. 3J), indicat-
ing that cellular pyrimidines have a specific
role in mouse-liver TPP synthesis.
To assess the contribution of ATP to TPP
synthesis, we compared TPK1 activity in DUMPS
and DGART cells cultured in the presence or
absence of uridine or inosine, respectively. Ino-
sine depletion reduced cellular abundance
of ATP but not TPP synthesis in DGART cells
(Fig. 3K), whereas uridine deprivation decreased
abundance of TPP (M+4) in DUMPS cells.
These findings indicated that pyrimidines, not
purines, limit cellular TPP synthesis (Fig. 3L).
Pyrimidine depletion also affected the activ-
ity of other TPP-utilizing enzymes, such as TKT,
OGDH, and BCKDH, as detected by metabolic
reaction–specific isotope tracing methods. TKT
activity was attenuated by pyrimidine depletion,
as reflected by the decrease in sedoheptulose
7-phosphate (M+7) derived from glucose (M+6)
(fig. S7, A and B). Depletion of pyrimidines also
reduced OGDH activity, reflected by the decrease
in succinyl-CoA (M+2) from the [13C2]-acetate
tracer, which incorporates into succinyl-CoA
independently of acetyl-CoA derived from PDH
activity (fig. S7, C and D). Pyrimidine depletion
similarly decreased BCKDH activity as moni-
tored by 13C2-leucine isotope tracing into acetyl-
CoA (M+1) (fig. S7, E and F). Given that PDH
has a low affinity for TPP compared with that of
TKT or OGDH (32, 33), PDH activity is likely
more sensitive to acute TPP depletion than
are other TPP-dependent enzymes. These find-
ings demonstrate that pyrimidines can sup-
port cellular TPP production and central carbon
metabolism.
Requirement of uridine triphosphate for TPK1
activity and pyruvate catabolism.
Textbooks have stated that TPK1 uses ATP
to phosphorylate thiamine and produce TPP
(34, 35). However, TPK1 might also use py-
rimidine nucleotides directly instead of ATP
to promote thiamine phosphorylation and TPP
synthesis. In vitro enzymatic studies showed
that TPK1 can use various nucleotides as phos-
phodonors, including UTP, without any clear
specificity (36, 37). We therefore explored
whether TPK1 uses UTP in human cells and
sustains pyruvate oxidation.
We tested whether exogenous nucleotides
could replenish abundance of TPP in digitonin-
permeabilized cells that had been depleted of
pyrimidines (38) (Fig. 4A). The addition of
UTP, but not CTP, ATP, or GTP, effectively re-
stored cellular TPP levels and promoted pyr-
uvate catabolism (Fig. 4, B and C, and fig. S8, A
to D). Additionally, exogenous UTP restored
TCA-cycle metabolism in pyrimidine-depleted
cells (Fig. 4D), indicating that UTP is required
for TPP synthesis and pyruvate oxidation. A
decline in pyruvate oxidation occurred when
cellular UTP concentrations dropped by ~50%
(fig. S8E). Moreover, direct addition of TPP to
permeabilized DUMPS cells partially restored
pyruvate catabolism, indicating that TPP de-
pletion resulting from depletion of cellular py-
rimidines can contribute to regulation of pyruvate
oxidation (Fig. 4E and fig. S8F). Moreover, the
depletion of CTP in cells genetically depleted
of CTPS1 and 2 did not affect pyruvate catab-
olism, further supporting the notion that UTP,
but not CTP, supports pyruvate catabolism in
human cells (fig. S8G).
To determine the phosphodonor preference
of TPK1 for thiamine phosphorylation, we in-
cubated purified human TPK1 with thiamine,
UTP, ATP, or CTP. Although TPK1 catalyzed
TPP synthesis from UTP, ATP, or CTP, UTP
was the preferred substrate for TPK1 in pro-
ducing TPP (fig. S9, A to C). TPK1 showed a
~10-fold higher affinity for UTP than ATP
(Km-UTP = 0.33 mM; Km-ATP = 5 mM) and dis-
played higher catalytic efficiency (kcat/Km) when
incubated with UTP instead of ATP (Fig. 4F
and fig. S9D) (kcat, catalytic rate constant; Km,
Michaelis constant). We conducted an equilib-
rium binding assay with radioactive UTP and
purified human TPK1, in which unlabeled UTP
completely displaced the bound radiolabeled
UTP from human TPK1 (Fig. 4G), demonstrat-
ing specific and direct binding. Through virtual
screening for nucleotide binding sites on the
TPK1 structure, we identified highly conserved
residues within two helices that potentially
science.org SCIENCE

Fig. 2. Pyrimidines are required to support PDH and mitochondrial
complex I activities. (A) Schematic illustrating the decarboxylation of
pyruvate step mediated by PDH. (B and C) 14CO2 production from [1-14C]-
pyruvate, reflecting PDH activity in DUMPS HEK293E cells cultured in the
presence or absence of uridine (200 mM) (B), reconstituted or not with
a UMPS cDNA construct (C). (D) Immunoblots of wild-type (WT) and DCAD
HEK293E cells cultured in the presence or absence of uridine (200 mM) for
the indicated times. (E) Normalized peak areas of 13C-labeled acetyl-CoA
derived from [13C3]-pyruvate labeling in DCAD HEK293E cells reconstituted
or not with CAD cDNA, cultured in the presence or absence of uridine
(200 mM, 16 hours). (F) Measurement of NADH/NAD+ ratio from whole-cell
extracts of HeLa cells treated with vehicle (DMSO) or BRQ (1 mM, 16 hours).
(G) NAD+, NADH levels, and ratio measured from isolated mitochondria of HeLa
cells treated with vehicle (DMSO) or BRQ (1 mM, 16 hours). (H and I) Same as
SCIENCE science.org
RESE ARCH | R E S E A R C H A R T I C L E S
shown in (F) and (G) but in DUMPS HEK293E cells reconstituted or not
with UMPS cDNA, in the presence or absence of uridine (200 mM, 16 hours).
(J) NADH/NAD+ ratios of isolated mitochondria from A375, A549, and
CAL-51 cells deleted for CAD (DCAD), in the presence or absence of uridine
(200 mM, 16 hours). (K) Measurement of respiratory complex I activity
and quantification of state III respiration in HeLa WT cells treated with
vehicle (DMSO) or BRQ (1 mM, 16 hours). (L) Same as shown in (K) but in
DUMPS HEK293E cells reconstituted or not with UMPS cDNA, in the presence
or absence of uridine (200 mM, 16 hours). Graphs representative of three
replicates shown in (K) and (L). For all other panels, data are mean ± SD.
n = 3 independent replicates.*P < 0.05 for multiple comparisons calculated
using one-way ANOVA with Tukey’s HSD test [(C), (E), (H), (I), and (L)]
or using a two-tailed Student’s t test for pairwise comparisons [(B), (F), (G),
(J), and (K)].
29 MARCH 2024 • VOL 383 ISSUE 6690 1487
RES EARCH | R E S E A R C H A R T I C L E S
Fig. 3. Cellular pyrimidine levels are required for thiamine pyrophosphate
synthesis. (A) Experimental design of the untargeted metabolomics experiment in
DCAD HEK293E cells. (B) Pathway impact analysis of untargeted metabolomics.
(C) Metabolic role of thiamine in cells. (D and E) Steady-state levels of thiamine
pyrophosphate (TPP) measured by LC-MS in HeLa (D) and HEK293E (E) cells treated
with vehicle (DMSO), BRQ (1 mM), or PRZ (1 mM). (F) Steady-state levels of thiamine,
TPP, and UDP in DUMPS HeLa cells reconstituted or not with UMPS cDNA, in the
presence or absence of uridine (200 mM, 16 hours). (G) Cellular TPP kinase 1 (TPK1)
activity measured with stable isotope tracing through measurement of carbon flow
from 13C4-thiamine into TPP. (H) Fractional enrichment (%) of 13C-labeled metabolites
derived from [13C4]-thiamine in DUMPS HeLa cells reconstituted or not with UMPS
cDNA in the presence or absence of uridine (200 mM, 16 hours). (I) Fractional
1488 29 MARCH 2024 • VOL 383 ISSUE 6690
enrichment (%) of 13C-labeled metabolites derived from [13C4]-thiamine in HeLa cells
treated with vehicle (DMSO), BRQ (1 mM), or methotrexate (MTX, 4 mM) for 16 hours.
(J) Experimental design and steady-state levels of TPP, UTP, and ATP from the
liver of mice treated with either vehicle (PEG-400 30%, PBS 70%), BRQ (20 mg/kg), or
PRZ (10 mg/kg) over 30 days. (K and L) Fractional enrichment (%) of 13C-labeled
metabolites derived from [13C4]-thiamine in DGART or DUMPS HeLa cells, cultured
in the presence or absence of uridine or inosine (200 mM, 6 hours). Normalized
peak areas of ATP and UTP are shown. Data in (J) are mean ± SD. n = 6 mice. For all
the other panels, data are mean ± SD. n = 3 independent replicates. *P < 0.05 for
multiple comparisons calculated using one-way ANOVA with Tukey’s HSD test
[(D), (F), (H), (I), and (J)] or by a two-tailed Student’s t test for pairwise
comparisons [(E), (K), and (L)].
science.org SCIENCE

Fig. 4. UTP is a substrate of TPK1 and supports pyruvate oxidation.
(A) Workflow of cellular permeabilization coupled with LC-MS/MS. (B) Steady-state
levels of TPP from DUMPS HEK293E cells cultured in the presence or absence of
uridine (200 mM), digitonin (10 mM), and supplemented with either vehicle (water) or
the indicated nucleotides (200 mM) for 8 hours prior to metabolite extraction.
(C) Citrate (M+2)/pyruvate (M+3) enrichment ratio from DUMPS HEK293E cells
treated as in (B) but labeled with [13C3]-pyruvate for 3 hours. (D) Fractional
enrichment (%) of 13C-labeled metabolites derived from [13C3]-pyruvate labeling in
DUMPS HEK293E cells measured and treated as in (C). Normalized peak areas of UTP
are shown. (E) Citrate (M+2)/pyruvate (M+3) ratio from DUMPS HEK293E cells
treated as in (B) and (C) but supplemented with indicated concentration of TPP and
labeled with [13C6]-glucose for the final 4 hours. (F) Kinetic assays with human TPK1
purified from E. coli incubated with increasing concentrations of UTP or ATP over time.
The product TPP was detected by LC-MS/MS. (G) Binding of radiolabeled UTP to TPK1.
SCIENCE science.org
RESE ARCH | R E S E A R C H A R T I C L E S
Cold UTP was added where indicated. (H) Virtual ligand screening and docking
simulation for UTP on human TPK1 identified putative UTP binding pocket. (I) SDS-
polyacrylamide gel electrophoresis (PAGE), followed by Coomassie blue staining,
was used to analyze human TPK1 protein preparation from E. coli. TPP levels
measured by LC-MS from purified human WT or mutants TPK1 incubated with
thiamine (1 mM), UTP (1 mM), or ATP (1 mM). (J) Immunoblots and TPP levels
from sgTPK1 HEK293E cells reconstituted with empty vector (EV), WT, D100A/F101A
(DA/FA), D100A, or F101A human TPK1. *P < 0.05 calculated using a one-way
ANOVA with Tukey’s HSD test for multiple comparisons [(B) to (E), (I), and (J)]
or a two-tailed Student’s t test for pairwise comparisons (G). Single-letter
abbreviations for the amino acid residues are as follows: A, Ala; C, Cys; D, Asp;
F, Phe; K, Lys; L, Leu; S, Ser; T, Thr; V, Val. In the mutants, other amino acids
were substituted at certain locations; for example, D100A indicates that aspartic
acid at position 100 was replaced with alanine.
29 MARCH 2024 • VOL 383 ISSUE 6690 1489
RES EARCH | R E S E A R C H A R T I C L E S
Fig. 5. Maintenance of de novo lipogenesis and adipocyte differentiation
by UTP. (A) Schematic illustrating the central role of citrate downstream of PDH
for the maintenance of de novo lipogenesis. (B) Relative incorporation of 14C
from [U-14C]-glucose into lipids in HEK293E cells treated with vehicle (DMSO),
BRQ (1 mM, 16 hours), or GSK 2194069 (FASNi, 20 mM, 16 hours). (C) 14C
incorporation from [2-14C]-pyruvate into lipids in 3T3-L1 fibroblasts treated with
vehicle (DMSO), BRQ (1 mM, 16 hours), CPI-613 (PDHi, 100 mM, 16 hours), or
GSK 2194069 (FASNi, 20 mM, 16 hours). (D) 14C incorporation from [2-14C]-
pyruvate into lipids in DUMPS HEK293E cells cultured in the presence or absence
of uridine (200 mM, 16 hours), stably reconstituted or not with UMPS cDNA.
1490 29 MARCH 2024 • VOL 383 ISSUE 6690
(E) 14C incorporation from [1-14C]-acetate into lipids in DUMPS HEK293E cells
cultured and treated as in (D). For a positive control, cells were treated with GSK
2194069 (FASNi, 20 mM, 3 hours). (F) DUMPS HEK293E cells reconstituted or not
with UMPS cDNA, in the presence or absence of uridine (200 mM, 16 hours) and
labeled with radioactive [1-14C]-palmitate. Measurement of 14C-carbon dioxide reflects
fatty acid oxidation. (G) 14C incorporation from [2-14C]-pyruvate into lipids in
DUMPS, DPDHA1, or DUMPS/DPDHA1 HEK293E cells treated as in (D). (H) 14C
incorporation from [2-14C]-pyruvate into lipids in DUMPS HEK293E cells cultured in
the presence or absence of uridine (200 mM), permeabilized with digitonin (10 mM),
and treated with vehicle (water) or UTP (1 mM) for 8 hours and labeled with
science.org SCIENCE

[2-14C]-pyruvate for the last 2 hours. (I) Schematic illustrating the adipocyte
differentiation of primary preadipocytes extracted from mouse fat pads. (J) Adipocyte
differentiation of mouse primary white preadipocytes treated with vehicle (DMSO) or
BRQ (200 nM) for 4 days in the presence or absence of uridine (200 mM) or acetate (1 mM).
Oil red O staining and quantification are shown. (K) Adipocyte differentiation of sgUMPS
form a UTP-binding pocket (Fig. 4H and fig.
S9, E and F). We used site-directed muta-
genesis of these conserved residues of TPK1
to generate various TPK1 mutants (fig. S9G).
The double mutants D100A and F101A had
reduced TPK1 activity when incubated with
UTP or ATP (Fig. 4I and fig. S9H) and showed
diminished UTP or ATP binding to TPK1 (fig.
S9I). The D100 residue was required for both
UTP- and ATP-mediated catalysis (Fig. 4I),
whereas the F101A single mutation impaired
only TPK1’s capacity to synthesize TPP from
UTP (Fig. 4I). To test whether the UTP-binding
sites of TPK1 promote cellular TPP synthesis
and pyruvate catabolism, we established cells
lacking TPK1 (sgTPK1) (fig. S10A) and recon-
stituted them with empty vector or with vec-
tors encoding WT TPK1 or mutant variants of
TPK1 (D100A and F101A) (Fig. 4J). Expression
of WT TPK1 in sgTPK1 cells restored TPP syn-
thesis, whereas cells expressing D100A or
F101A TPK1 made smaller amounts of TPP
(Fig. 4J). Moreover, cells expressing D100A
TPK1 displayed decreased pyruvate catabolism
compared with that of sgTPK1 cells reconsti-
tuted with WT TPK1 (fig. S10, B to D). These
findings reveal a specific role for UTP in con-
trolling TPP synthesis and pyruvate catabo-
lism in human cells.
Pyrimidines control de novo lipogenesis and
adipocyte differentiation
Our work indicates that UTP concentrations
influence TPP synthesis and citrate synthesis
downstream of PDH. Citrate can exit mitochon-
dria and contribute to the production of fatty
acids in the cytosol as acetyl-CoA, in addition
to being used by the TCA cycle (Fig. 5A). To
evaluate the effects of pyrimidine depletion on
de novo lipogenesis, we quantified the incor-
poration of carbons from [14C]-glucose, [14C]-
pyruvate, or [14C]-acetate into lipids in HEK293E,
HeLa cells, and mouse fibroblasts. As a con-
trol, we treated cells with a selective fatty acid
synthase inhibitor (FASNi, GSK 2194069) or a
PDH inhibitor (PDHi, CPI-613). FASNi signif-
icantly reduced lipogenesis (Fig. 5, B and C,
and fig. S11A). Pyrimidine depletion with BRQ
in HEK293E cells or mouse 3T3-L1 fibroblasts
diminished de novo lipogenesis compared with
that of control cells (Fig. 5, B and C, and fig.
S11A). These effects were similar to those in
cells either expressing TPK1 D100A or inhib-
ited for PDH and TPK1 (Fig. 5C and fig. S11, B
to D). Pyrimidine depletion decreased pyru-
vate catabolism and TPP concentrations in
3T3-L1 fibroblasts (fig. S11E). Moreover, uri-
SCIENCE science.org
3T3-L1 fibroblasts cultured in the presence of high (200 mM) or low (50 nM)
concentration of uridine. Oil red O staining and quantification are shown. (L) Regulation
of vitamin B1 metabolism and pyruvate oxidation by UTP levels. Data are mean ± SD.
n = 3 independent replicates. *P < 0.05 for multiple comparisons calculated
using a one-way ANOVA with Tukey’s HSD test [(B) to (H), and (K)].
dine depletion in DUMPS cells also decreased
de novo lipogenesis in cells labeled with [2-14C]-
pyruvate (Fig. 5D) but not in cells labeled with
[1-14C]-acetate (Fig. 5E), indicating that pyrim-
idine regulation of lipogenesis may occur through
control of PDH. Consistent with a decrease in
fatty acid synthesis, measurement of the fatty
acid b-oxidation through [1-14C]-palmitate la-
beling revealed that DUMPS cells depleted of
uridine had increased mitochondrial b-oxidation
(Fig. 5F).
To confirm the role of PDH in the effects of
pyrimidine concentrations on lipogenesis, we
generated HEK293E cells lacking PDHA1 as
well as UMPS and PDHA1 (DPDHA1-DUMPS)
through modification with CRISPR-Cas9 (fig.
S11F). Loss of PDHA1 and depletion of uridine in
DUMPS cells decreased lipogenesis (Fig. 5G),
and depletion of uridine in DUMPS-DPDHA1 cells
did not further decrease pyruvate-dependent
lipogenesis, suggesting that pyrimidine concen-
trations predominantly influence lipogenesis by
regulating PDH (Fig. 5G). Because UTP is nec-
essary for TPP synthesis and pyruvate oxidation,
we tested whether supplementing permeabi-
lized cells with UTP could prevent pyrimidine
depletion from inhibiting lipogenesis. UTP
supplementation effectively restored lipoge-
nesis in DUMPS and DCAD cells depleted of
uridine (Fig. 5H and fig. S11G). Given the mod-
ulation of lipogenesis in response to changes
in intracellular pyrimidine concentrations, we
tested further whether pyrimidines might in-
fluence adipogenesis. We examined the ability
of mouse primary preadipocytes to undergo
white-adipocyte differentiation in the presence
or absence of a low dose of BRQ (Fig. 5I). Low-
dose BRQ reduced adipogenesis of primary
adipocytes (Fig. 5J). Similarly, DHODH inhi-
bition decreased adipogenesis of 3T3-L1 fibro-
blasts without any discernible impact on their
mitotic clonal expansion capability (fig. S11, H
to K). Exogenous uridine or acetate partially
rescued the antidifferentiation effects induced
by pyrimidine depletion (Fig. 5J and fig. S11I).
Furthermore, 3T3-L1 fibroblasts lacking Umps
(sgUmps) (fig. S12A) also exhibited decreased
adipocyte differentiation capacity when cultured
with a low concentration of uridine (50 nM),
and this phenotype was rescued by increasing
uridine concentration in the medium or by
supplementation with acetate (Fig. 5K and fig.
S12B). 3T3-L1 cells deficient in TPK1 (sgTpk1)
or PDH (sgPdha1) also displayed decreased
adipocyte differentiation (fig. S12, C and D).
Overall, we demonstrated that the pyrimi-
dine nucleotide UTP supports cellular TPP
RESE ARCH | R E S E A R C H A R T I C L E S
synthesis and mitochondrial pyruvate oxida-
tion, enabling citrate utilization for de novo
lipogenesis and adipocyte differentiation (Fig.
5L and fig. S13).
Discussion
Control of pyruvate oxidation has been pri-
marily attributed to PDK-dependent phosphor-
ylation of PDH (27, 39). However, our study
reveals a role of pyrimidines in regulating py-
ruvate catabolism across different species. We
found that pyrimidines, not purines, stimu-
lated pyruvate catabolism by controlling TPK1
activity, which promotes TPP synthesis, an es-
sential cofactor for PDH. We discovered that
TPK1 effectively utilizes UTP, rather than ATP,
for cellular TPP generation. Purines, on the other
hand, do not limit TPP synthesis or pyruvate ca-
tabolism. As a result, UTP emerges as a poten-
tially important phosphodonor for vitamin
B1 metabolism. Thiamine utilization has been
linked to cancer cell growth and treatment,
underscoring the importance of understand-
ing its regulation in human cells (40, 41). The
pyrimidine-dependent control of TPP availa-
bility and PDH activity drives the fate of in-
termediates downstream of pyruvate, such as
citrate, which can either be catabolized into
the TCA cycle, recycled into the TCA cycle, or
directly used for lipid synthesis (42, 43). Al-
though UTP is required for pyruvate oxidation
and TPP synthesis, the decrease in amounts of
TPP in cells depleted of UTP is partially respon-
sible for the inhibition of pyruvate oxidation.
Thus, additional UTP-dependent mechanisms
that contribute to PDH regulation might exist.
Depletion of cellular pyrimidines led to de-
creases in both PDH activities and TCA-cycle
activities, which, in turn, resulted in decreased
mitochondrial NADH generation and impaired
complex I activity and respiration. This find-
ing unveils the possible role of pyrimidines in
controlling adipocyte differentiation through
the regulation of lipogenesis. We conclude that
changes in UTP concentrations may have pre-
viously unrecognized influences on thiamine
utilization, pyruvate catabolism, and cell fate.
REFERENCES AND NOTES
1. A. N. Lane, T. W. Fan, Nucleic Acids Res. 43, 2466–2485
(2015).
2. J. Zhu, C. B. Thompson, Nat. Rev. Mol. Cell Biol. 20, 436–450
(2019).
3. I. Martínez-Reyes, N. S. Chandel, Nat. Commun. 11, 102 (2020).
4. M. G. Vander Heiden et al., Cold Spring Harb. Symp.
Quant. Biol. 76, 325–334 (2011).
5. E. Villa, E. S. Ali, U. Sahu, I. Ben-Sahra, Cancers 11, 688 (2019).
6. G. Li, D. Li, T. Wang, S. He, Int. J. Mol. Sci. 22, 10253
(2021).
29 MARCH 2024 • VOL 383 ISSUE 6690 1491
RES EARCH | R E S E A R C H A R T I C L E S
7. F. Del Caño-Ochoa, M. Moreno-Morcillo, S. Ramón-Maiques,
Subcell. Biochem. 93, 505–538 (2019).
8. V. Aleman, P. Handler, J. Biol. Chem. 242, 4087–4096 (1967).
9. K. Shostak, M. E. Jones, Biochemistry 31, 12155–12161 (1992).
10. M. E. Jones, Annu. Rev. Biochem. 49, 253–279 (1980).
11. M. Goto, R. Omi, N. Nakagawa, I. Miyahara, K. Hirotsu,
Structure 12, 1413–1423 (2004).
12. H. L. Schultheisz, B. R. Szymczyna, L. G. Scott, J. R. Williamson,
J. Am. Chem. Soc. 133, 297–304 (2011).
13. T. Mio, T. Yabe, M. Arisawa, H. Yamada-Okabe, J. Biol. Chem.
273, 14392–14397 (1998).
14. A. L. Wolfe et al., Proc. Natl. Acad. Sci. U.S.A. 118, e2103592118
(2021).
15. E. P. Kennedy, S. B. Weiss, J. Biol. Chem. 222, 193–214
(1956).
16. D. B. Ostrander, D. J. O’Brien, J. A. Gorman, G. M. Carman,
J. Biol. Chem. 273, 18992–19001 (1998).
17. G. Weber, M. A. Lea, H. J. Hird Convery, N. B. Stamm,
Adv. Enzyme Regul. 5, 257–298 (1967).
18. A. I. Jonckheere, J. A. Smeitink, R. J. Rodenburg, J. Inherit.
Metab. Dis. 35, 211–225 (2012).
19. M. Bajzikova et al., Cell Metab. 29, 399–416.e10 (2019).
20. M. H. Soflaee et al., Nat. Commun. 13, 2698 (2022).
21. R. A. Saxton, D. M. Sabatini, Cell 168, 960–976 (2017).
22. A. J. Valvezan, B. D. Manning, Nat. Metab. 1, 321–333 (2019).
23. M. J. Holness, M. C. Sugden, Biochem. Soc. Trans. 31,
1143–1151 (2003).
24. S. Zhang, M. W. Hulver, R. P. McMillan, M. A. Cline, E. R. Gilbert,
Nutr. Metab. 11, 10 (2014).
25. T. Kupke, J. Biol. Chem. 277, 36137–36145 (2002).
26. R. Leonardi, S. Jackowski, Ecosal Plus 2, 10.1128/
ecosalplus.3.6.3.4 (2007).
27. L. R. Gray, S. C. Tompkins, E. B. Taylor, Cell. Mol. Life Sci. 71,
2577–2604 (2014).
28. D. Lonsdale, Evid. Based Complement. Alternat. Med. 3, 49–59
(2006).
29. L. de Jong, Y. Meng, J. Dent, S. Hekimi, Genetics 168, 845–854
(2004).
30. J. R. Butler, R. H. Pettit, P. F. Davis, L. J. Reed, Biochem.
Biophys. Res. Commun. 74, 1667–1674 (1977).
31. J. A. Mayr et al., Am. J. Hum. Genet. 89, 806–812 (2011).
32. V. I. Bunik, A. Tylicki, N. V. Lukashev, FEBS J. 280, 6412–6442
(2013).
33. J. R. Tate, P. F. Nixon, Anal. Biochem. 160, 78–87 (1987).
34. M. Sambon et al., Biochim. Biophys. Acta, Gen. Subj. 1866,
130071 (2022).
35. L. J. Baker, J. A. Dorocke, R. A. Harris, D. E. Timm, Structure 9,
539–546 (2001).
36. A. I. Voskoboyev, Y. M. Ostrovsky, Ann. N.Y. Acad. Sci. 378,
161–176 (1982).
37. M. Onozuka, K. Nosaka, J. Nutr. Sci. Vitaminol. 49, 156–162
(2003).
38. Y. Nonnenmacher et al., Metab. Eng. 43, 147–155 (2017).
39. A. R. Grassian, C. M. Metallo, J. L. Coloff, G. Stephanopoulos,
J. S. Brugge, Genes Dev. 25, 1716–1733 (2011).
40. J. A. Zastre, R. L. Sweet, B. S. Hanberry, S. Ye, Cancer Metab.
1, 16 (2013).
41. R. Guarecuco et al., Sci. Adv. 6, eabc7120 (2020).
42. P. K. Arnold et al., Nature 603, 477–481 (2022).
43. K. E. Wellen et al., Science 324, 1076–1080 (2009).
44. U. Sahu et al., Pyrimidines maintain mitochondrial pyruvate
oxidation to support de novo lipogenesis. Dryad (2024);
https://doi.org/10.5061/dryad.rxwdbrvfq.
ACKN OW LEDG MEN TS
We thank the remaining members of the Ben-Sahra laboratory for
their valuable contributions and discussions on the data presented
in this manuscript. Funding: This work was supported by grants
from the NIH (R01GM135587 and R01GM143334 to I.B.-S.,
5R35CA197532 and 5P01AG049665 to N.S.C., and R35CA197569
to A.S.); the LAM Foundation Established Investigator Award
(LAM0151E01-22 to I.B.-S.); and the American Cancer
Society Award (DBG-23-1039959-01-TBE to I.B.-S.). Author
contributions: U.S. and E.V. performed all experiments and
prepared the manuscript. P.G. and J.M.A. performed the LC-MS
analysis. C.R.R. and N.S.C. provided expertise for metabolic
experiments on animals. Z.Z. and A.S. provided the mouse
embryonic stem cells. B.P.O. and M.D.T. provided technical
assistance. R.M. and B.S. performed the computational analysis of
the untargeted metabolomics data. I.B.-S. supervised the project,
reviewed all experimental data, and prepared the manuscript.
All authors discussed the results and commented on the
manuscript. Competing interests: The authors declare that they
1492 29 MARCH 2024 • VOL 383 ISSUE 6690
have no competing interests. Data and materials availability: All SUPPLEMENTARY MATERIALS
constructs and cell lines established in this study are available science.org/doi/10.1126/science.adh2771
upon request. All data are available in the main text or the Materials and Methods
supplementary materials. Raw data are archived on Dryad (44) Supplementary Text
and Mendeley Data (https://data.mendeley.com/preview/ Figs. S1 to S13
sg634gy9v2). License information: Copyright © 2024 the References (45–51)
authors, some rights reserved; exclusive licensee American
Association for the Advancement of Science. No claim to original Submitted 21 February 2023; resubmitted 14 August 2023
US government works. https://www.science.org/about/science- Accepted 20 February 2024
licenses-journal-article-reuse 10.1126/science.adh2771
PIEZOELECTRICS
Biodegradable ferroelectric molecular crystal with
large piezoelectric response
Han-Yue Zhang1*†, Yuan-Yuan Tang2†, Zhu-Xiao Gu3†, Peng Wang1,3†, Xiao-Gang Chen2,
Hui-Peng Lv2, Peng-Fei Li2, Qing Jiang3, Ning Gu4*, Shenqiang Ren5*, Ren-Gen Xiong1,2*
Transient implantable piezoelectric materials are desirable for biosensing, drug delivery, tissue
regeneration, and antimicrobial and tumor therapy. For use in the human body, they must show
flexibility, biocompatibility, and biodegradability. These requirements are challenging for conventional
inorganic piezoelectric oxides and piezoelectric polymers. We discovered high piezoelectricity in a
molecular crystal HOCH2(CF2)3CH2OH [2,2,3,3,4,4-hexafluoropentane-1,5-diol (HFPD)] with a large
piezoelectric coefficient d33 of ~138 picocoulombs per newton and piezoelectric voltage constant
g33 of ~2450 × 10−3 volt-meters per newton under no poling conditions, which also exhibits good
biocompatibility toward biological cells and desirable biodegradation and biosafety in physiological
environments. HFPD can be composite with polyvinyl alcohol to form flexible piezoelectric films with a
d33 of 34.3 picocoulombs per newton. Our material demonstrates the ability for molecular crystals to
have attractive piezoelectric properties and should be of interest for applications in transient
implantable electromechanical devices.
he piezoelectric effect is a phenomenon polarization as well as their being usable in
T
in which certain dielectrics generate elec- the polycrystalline ceramic form (11–13). In
tronic polarization through an applied comparison, piezoelectric molecular crystals
stress, or conversely (1–3). The effect en- generally have weak piezoelectric performance
ables direct conversion between electrical caused by relatively weak intermolecular forces,
and mechanical energy, which allows materials with piezoelectric coefficient d33 < 40 pC/N
to be widely used in actuators, transducers, (table S1) (5). Ferroelectric polymers, such as
sensors, and energy harvesters (4–7). The pie- polyvinylidene fluoride (PVDF), belong to
zoelectric effect can exist in ionic, atomic, and piezoelectric molecular semicrystals, which
molecular crystals (8–10). Among them, piezo- have been widely used in flexible electronic de-
electric ionic crystals, represented by ferroelectric vices on the basis of their high pliability, light
oxides barium titanate (BTO) and lead zirconate weight, and low toxicity (14–19).
titanate (PZT), have a wide range of industrial Transient implantable piezoelectric materials
and commercial applications because of their are attracting interest for biosensing, drug de-
excellent electromechanical conversion, high livery, tissue regeneration, and antimicrobial
Curie temperature, and strong spontaneous and tumor therapy because nondegradable scaf-
folds and implanted devices would increase the
risk of infection and inflammation (20–22).
1
Jiangsu Key Laboratory for Biomaterials and Devices,
School of Biological Science and Medical Engineering,
For example, Liu et al. reported a biodegrad-
Southeast University, Nanjing 210009, P. R. China. 2Ordered able poly(L -lactic acid) piezoelectric nano-
Matter Science Research Center, Nanchang University, fiber scaffold that can be used as a battery-less
Nanchang 330031, P. R. China. 3Division of Sports Medicine
and Adult Reconstructive Surgery, Department of Orthopedic
electrical stimulator to promote chondro-
Surgery, Nanjing Drum Tower Hospital, Affiliated Hospital of genesis and cartilage regeneration (21). Piezo-
Medical School, Nanjing University, Nanjing 210008, Jiangsu, electrics intended for human body application
P. R. China. 4Medical School, Nanjing University, Nanjing should be biocompatible, biodegradable, and
210093, Jiangsu, P. R. China. 5Department of Materials
Science and Engineering, University of Maryland, College flexible (23–26). However, many of these re-
Park, MD 20742, USA. quirements cannot be met with conventional
*Corresponding author. Email: zhanghanyue@seu.edu.cn (H.-Y.Z.); inorganic ferroelectric oxides and ferroelectric
guning@nju.edu.cn (N.G.); sren@umd.edu (S.R.); xiongrg@seu.edu.
cn (R.-G.X.) polymers, especially those pertaining to biode-
†These authors contributed equally to this work. gradability. Piezoelectric molecular crystals,
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 1. Crystal structures and phase transitions of HFPD. (A) Digital
image of HFPD crystals. (B) The transition of 2D monolayer formed through
O−H···O interactions of terminal hydroxyl in molecules between two
phases triggered by thermal or stress fields. The yellow dashed lines
indicate the O−H···O hydrogen bonding interactions. (C) Packing view of
2D hydrogen bond layers. (D) The results of DSC and dielectric (along
the b axis) measurements, revealing a structural phase transition at
around 282 K. (E) (Top) Temperature-dependent and (bottom)
pressure-dependent Raman spectra show thermal- and pressure-induced
phase transitions.

which are basically composed of organic single-
component compounds, may be a promising
solution to these concerns because of their
decent solubility, potential biocompatibility,
and biodegradability (27). For example, bio-
molecular crystals can provide many poten-
tially beneficial properties of biomaterials such
as flexibility, biocompatibility, reliability, bio-
degradability, and environmental sustain-
ability (28–31). Yang et al. successfully created
wafer-scale heterostructured piezoelectric bio-
material thin films based on g-glycine crystals
(22). However, those piezoelectric biomaterials
have very weak piezoelectric performance, with
piezoelectric coefficients of d33 ≤ 10 pC/N,
such as lysozyme (6.5 pC/N) (29), DL-alanine
(4.8 pC/N) (30), collagen (0.89 pC/N) (31), and
g-glycine (10.4 pC/N) (22, 28), all of which are
much smaller than that of PVDF (~28 pC/N)
(table S1) (32).
We report a piezoelectric molecular crystal,
HOCH2(CF2)3CH2OH [2,2,3,3,4,4-hexafluo-
ropentane-1,5-diol (HFPD)], that shows strong
piezoelectric response, with a large d33 reaching
~138 pC/N and a piezoelectric voltage constant
g33 of ~2450 × 10−3 Vm/N under no poling
conditions. Discovery of a molecular crystal
with a d33 of more than 100 pC/N is a notable
development since the discovery in 1880 of
the piezoelectric effect (table S1) (5, 22, 33).
The HFPD–polyvinyl alcohol (PVA) films not
only exhibit excellent flexibility, high biocom-
patibility, and no obvious toxicity to biological
cells but also show great biodegradation and
biosafety in physiological environments. Our
HFPD-PVA (2:1) films demonstrate excellent
piezoelectricity, having a d33 of 34.3 pC/N, which
is much larger than that of other organic pie-
zoelectric composites formed by biofriendly
molecular crystals such as g-glycine (5.3 pC/N)
(22). A two-dimensional (2D) hydrogen bond
network through O–H···O interactions of ter-
minal hydroxyl in molecules induces decent
solubility in various solvents and results in the
biodegradability of HFPD crystal. The anisotropy
of Young’s modulus comes from the ordered
arrangement of a 2D hydrogen bond network
and fluorine atoms arranged on the chain in
the crystal structure, which might be an im-
portant reason for the large piezoelectric co-
efficient of HFPD crystal.
Crystal structures and phase
transition behaviors
We recrystallized transparent rod-like single
crystals of HFPD with the size of 1 by 1 by 5 mm
using the solvent diffusion method at room
temperature (Fig. 1A) (34). These crystals showed
good environmental stability (fig. S1) and ther-
mal stability up to 380 K (fig. S2). Single-
crystal x-ray diffraction (XRD) determined
that HFPD adopts polar crystal symmetry with
P1 space group at room temperature (table S2).
HFPD molecules connected with the adjacent
ones (Fig. 1B) to form a 2D hydrogen bond
layer through O−H···O hydrogen bonding inter-
actions (Fig. 1C and table S3). The asymmetric
distribution of the C−F bonds generates a non-
zero molecular dipole moment, which is super-
imposed on each other by the 2D hydrogen
bond layered stacking structure, resulting in the
generation of spontaneous polarization (fig. S3).
When the crystal was cooled below 259 K, we
observed slight changes in cell parameters
that revealed the occurrence of isomorphic
phase transitions (Fig. 1B and table S2).
Differential scanning calorimetry (DSC) curves
and temperature-dependent dielectric measure-
ments confirmed that HFPD shows a thermal
anomaly at 282 K upon heating (Fig. 1D).
Temperature-dependent XRD (fig. S4) and
Raman spectrum measurements (Fig. 1E) fur-
ther verified this thermodynamic phase tran-
sition. We also found that this structural phase
transition can be triggered under pressure.
To confirm this, we performed in situ pressure-
dependent Raman spectrum measurements using
diamond anvil cells. The pressure-dependent
Raman spectra exhibited a peak transforma-
tion around 2970 cm–1 (C–H stretching) when
applied with pressure up to 0.36 GPa, whereas
a strong narrow peak separated into three
successive peaks at 2958, 2972, and 2983 cm–1

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1493

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 2. Ferroelectric properties of HFPD. (A to C) P-E hysteresis loops pattern, which was obtained by successively applying a tip voltage of
for HFPD single crystals along (A) [100], (B) [010], and (C) [001] crystal +90 V and –80 V to a region with an initial single-domain state. (F) PFM
orientations, respectively. (D) PFM phase image captured on HFPD switching hysteresis loops for amplitude and phase signals on HFPD
crystalline thin film. (E) PFM phase image showing a box-in-box domain crystalline thin film.

(Fig. 1E). The triple-peak character returned to
initial single-peak form after the pressure was
released from 0.63 GPa, indicating the full re-
versibility of the pressure-induced phase tran-
sition. The Raman vibrations recorded at the
low-temperature phase correspond to those of
the high pressure–induced new phase, indicat-
ing that these two phases should be the same.
This was also confirmed with single-crystal
XRD of HFPD crystal under a high pressure of
0.6 GPa (table S4).
Ferroelectric properties
HFPD crystals have the desired ferroelectricity
with switchable spontaneous polarization be-
cause their polarization directions are roughly
perpendicular to the long axis of the molecules,
where the rotational reorientation of molecules
along the long axis can relatively easily achieve
polarization reversal. We implemented
polarization−electric field (P-E) hysteresis loop
tests for the HFPD to confirm their ferroelec-
tricity, using the double-wave method. The HFPD
crystals exhibit well-shaped P-E hysteresis loops
along different crystal orientations (Fig. 2, A to C).
The HFPD crystals hold moderate saturated
polarization (Ps) of 0.73, 4.88, and 6.34 mC/cm2
along the a, b, and c axes, respectively, indi-
cating obvious crystal anisotropy, which is
also reflected in the different coercive field
values. The coercive field is substantially lower
than that of PVDF, with a typical value of
hundreds of kilovolts per centimeter (18). We
measured the breakdown electric field of HFPD
to be 726.74 kV/mm, which is higher than that
of PVDF (346.57 kV/mm) (fig. S5 and table S5).
We further investigated the ferroelectric
behaviors and piezoresponse of HFPD on the
nanoscale by using piezoresponse force micros-
copy (PFM) characterization, which is a powerful
technique with which to characterize ferro-
electrics by measuring the dynamic electro-
mechanical response (35–38). The PFM phase
image clearly shows a polydomain structure of
the HFPD crystalline thin film, indicating the
existence of spontaneous polarization in dif-
ferent directions (Fig. 2D and fig. S6). To con-
firm the ferroelectricity of HFPD, we obtained
the box-in-box domain pattern (Fig. 2E and
fig. S7) by successively applying a tip voltage
of +90 V and –80 V to a region with an initial
single-domain state. This reversible domain-
switching behavior further verifies the ferro-
electricity of HFPD. Moreover, we performed
the PFM switching spectroscopy measurement
on the crystalline thin-film surface. The butter-
fly amplitude loops and 180° phase shifts when
the amplitude signal was at a minimum, demon-
strating robust ferroelectric switching (Fig. 2F).
Piezoelectric performance
We used the PFM measurements on the HFPD
crystalline thin films to evaluate their piezo-
electric performance. The XRD pattern indi-
cated that the orientation of HFPD crystal in
its crystalline thin film is mainly at the (012)
and (010) planes (fig. S8). We show in Fig. 3A
the relationship between piezoelectric amplitude
and frequency of the HFPD and PVDF when
the excitation frequency sweeps around the con-
tact resonance frequency. We fit the curves with
the damped harmonic oscillator model, and
HFPD exhibits a much larger piezoresponse
peak (34). We can obtain the intrinsic inverse
piezoelectric deformation of the sample by ac-
quiring the inverse optical lever sensitivity of
the PFM cantilever and correcting the reso-
nance magnification using Q factor. We repeated
such measurements over six randomly chosen
points on the HFPD crystalline thin film and
PVDF film under different ac excitation volt-
ages, respectively. It shows that the electro-
mechanical responses from HFPD and PVDF
adapt linear relationships with respect to their
piezoelectricity (Fig. 3B). We can estimate the
piezoelectric coefficient from the slope, which
we measured to be 119.3 pm/V for HFPD crys-
talline thin films and 18.8 pm/V for PVDF film
(21 pC/N measured with the quasi-static method).
We investigated the piezoelectric response
of HFPD crystals with the Berlincourt method,
based on the direct piezoelectric effect. We
measured the longitudinal piezoelectric co-
efficients d33 of HFPD crystals to be ~90 pC/N
on the (010) plane, ~70 pC/N on the (012) plane,
and a maximum of ~138 pC/N on the ð021 Þ
plane (Fig. 3C and fig. S9). Such promising d33
values of HFPD crystals are considerably large
among single-component organic piezoelectrics

1494 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 3. Piezoelectric properties of HFPD. (A) Piezoresponse of HFPD and PVDF
films versus excitation frequency measured with resonance-enhanced PFM and
analyzed by the damped harmonic oscillator model. (B) The derived amplitude-
excitation curve for HFPD and PVDF films by PFM tip. (C) Piezoelectric coefficient
of HFPD compared with some classical piezoelectric ionic, atomic, and molecular
crystals, including PZT (42), BTO (43), (TMCM)CdCl3 [trimethylchloromethyl
ammonium trichlorocadmium(II)] (5), triglycine sulfate (TGS) (44), g-glycine (45),
AlN (46), lysozyme (29), PVDF (32), poly(vinylidene fluoride-co-trifluoroethylene)
(PVDF-TrFE) (47), croconic acid (5), and more (the detailed data are included in
table S1). Measured d33 values are listed at the end of each bar. (D to F) 3D plots of
elastic modulus of HFPD crystals. (D) Directional Young’s modulus. (E) Spatial
dependence of shear modulus MAX. (F) Spatial dependence of shear modulus MIN.

and are competitive with the piezoelectric poly-
mer PVDF (34). The piezoelectric voltage co-
efficient, as a key indicator of performance,
directly represents a material figure of merit
for piezoelectric sensors. We estimated the
g33 of HFPD to be about 2450 × 10−3 Vm/N.
This value is nine times higher than that of
PVDF (122.4 × 10−3 to 263.5 × 10−3 Vm/N) and
seven times that of poly(vinylidene fluoride-
ran-trifluoroethylene) [P(VDF-TrFE)] (225.9 ×
10−3 to 357.7 × 10−3 Vm/N) (table S6) (32). Also,
we calculated the electrostriction coefficient
Q22 of HFPD crystal to be about 15.28 m4/C 2,
which is larger than that of PVDF (1.3 m4/C 2) (34).
To investigate the mechanical properties of
our crystals, we calculated the elastic stiffness
matrix cij (34). The small elastic stiffness con-
stant suggests that HFPD crystal is an inherently
mechanically soft material. To understand the
elastic properties more intuitively, we plotted
the directional Young’s modulus, which repre-
sents the uniaxial stiffness in different direc-
tions (Fig. 3D). The directional Young’s modulus
has a very large anisotropy, and the two maxima
are located in an irregular plane that is almost
elliptical. The anisotropy of Young’s modulus
comes from the ordered arrangement of 2D
hydrogen bond network and F atoms arranged
on the chain in the crystal structure. In addi-
tion, the spatial dependence of shear modulus
MIN surfaces with six maxima are regularly
nested within the MAX surfaces to form a shell
structure of shear modulus (Fig. 3, E and F).

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1495

RES EARCH | R E S E A R C H A R T I C L E S

Fig. 4. Characterization of HFPD-PVA films. (A) (Top) Schematic diagram
of the film fabrication by solution evaporation and digital images of the (bottoom
left) as-grown and (bottom right) curled film showing good flexibility. (B) (Left)
SEM image of the as-prepared HFPD-PVA (2:1) film and (Right) corresponding
elemental mapping for F element. (C) Stress-strain curves of as-prepared
films with various HFPD-to-PVA ratios. (D) (Left) XRD patterns simulated from
the single crystal of HFPD and measured from the as-prepared HFPD-PVA (2:1) and
pure PVA film. (Right) Schematic diagram of molecular arrangement in (012) plane.
(E) Young’s modulus, elongation at break and tensile strength of as-prepared
films with various HFPD-to-PVA ratios. (F) The piezoelectric coefficient d33 and
piezoelectric voltage coefficient g33 obtained from the as-prepared films with different
HFPD-to-PVA ratios. (G) (Left) The piezoelectric voltage output of a packaged HFPD-
PVA film with a 2:1 ratio before and after being immersed in PBS. (Right) Digital
images of as-prepared film encapsulated with PLA before and after degradation.

Such an exotic mechanical property might be
an important reason for the large piezoelectric
coefficient of the HFPD crystal. The pressure-
induced structural phase transition may be
another reason. It can be found that under a
certain pressure, some of the HFPD molecules
in the lattice are likely to rotate 45°, resulting
in a certain increase in d33 (figs. S10 and S11
and movie S1) (34). To compare with the experi-
mentally used piezoelectric strain constants
(commonly denoted as dij), the piezoelectric
stress coefficients eij and elastic compliances sij
should be calculated because dij = eij * sij, in
which the elastic compliance sij is obtained by
inverting the elastic stiffness matrix cij directly.
Combining the values of eij and cij matrix, we
can directly get the piezoelectric strain con-
stants dij (table S7)
2 3
37:6 101:8 64:2 109:1 455:3 266:7
4 13:7 46:4 29:6 50:4 166:2 93:0 5
8:0 26:1 21:7 17:3 94:7 52:6
According to the calculation results of this
matrix, the maximum allowable longitudinal
piezoelectric coefficient is up to 248 pC/N,
which means that the large value of d33 may
be experimentally measured, providing compet-
itive candidates for piezoelectric sensors and
fine-tuned microelectromechanical systems. Fur-
thermore, we used the Berry phase method
to estimate the ferroelectric polarization of
the crystal, from which the polarization with
–0.93, 4.22, and 6.22 mC/cm2 along the a, b,
and c axes can be extracted, in good accord-
ance with the symmetry requirement of space
group P1. We estimated by means of vector sum-
mation the total polarization to be 7.58 mC/cm2,
which is close to the measured value.
Flexible piezoelectric films
For piezoelectric devices compatible with the
human body such as artificial muscle, electronic
skin, and wearable devices, their flexibility is
of great concern. Current flexible piezoelectrics
mainly include PVDF-based polymers, inorganic-
polymer composites, and bio-piezoelectric ma-
terials (18, 22, 27, 39, 40). The HFPD crystals
present an elastic modulus (E) of 6.14, 8.85, and
4.30 GPa and hardness (H) of 437.4, 365.1, and
507.3 MPa for (100), (010), and (001) planes, re-
spectively (fig. S12) (34), in the same magni-
tude as PVDF (E = 3.28 GPa, H = 240.3 MPa)
(41), indicating the great application poten-
tial in flexible and body-compatible electro-
mechanical devices. Furthermore, the acoustic
impedance (z0) for HFPD [4.08 × 106 kg/(m2 s)]
is lower than that of P(VDF-TrFE) [4.51 × 106 kg/
(m2 s)], which is comparable with that of PVDF
[3.92 × 106 kg/(m2 s)], which matches that of the
body reasonably well, making it a promising
candidate for implantable ultrasound applica-
tions (table S8) (32).
Crystals are inherently rigid and brittle, which
makes them challenging to be used in flexible
systems. We thus prepared the HFPD-PVA films
by means of solution evaporation from their
corresponding mixture solution at 40°C (Fig.
4A) (34). The films exhibited excellent flexibil-
ity, biocompatibility, and piezoelectricity (Fig.
4A). The as-prepared films with the HFPD-to-
PVA ratio above 2:1 were relatively uniform
(fig. S13). The top-view (Fig. 4B and fig. S14)
and cross-sectional scanning electron micro-
scope (SEM) (fig. S14) images show the dis-
tribution of HFPD crystal in the HFPD-PVA
films. Among them, the corresponding elemen-
tal mapping image demonstrated that F elements
(only from HFPD) were uniformly distributed
in the HFPD-PVA (2:1) film (Fig. 4B and fig. S14).
We measured stress-strain curves to understand
the mechanical properties of as-prepared films
with various HFPD-to-PVA ratios (Fig. 4C and
table S9). The maximum tensile strains of these
films decreased gradually with the increased

1496 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

 RESE ARCH | R E S E A R C H A R T I C L E S

Fig. 5. Biocompatibility and biodegradability of HFPD. (A) Viability of
rBMSCs after incubation with various concentrations of HFPD. (B) Representative
images of live/dead staining (green and red labels indicate living and dead cells,
respectively) and phalloidine/DAPI staining (green and blue labeled cells indicate
cytoskeleton and nucleus of cells, respectively) of rBMSCs after incubation with
various concentrations of HFPD. (C) Schematic diagram of the process of in vivo
biocompatibility and biodegradability assessment through tail intravenous
injection of HFPD. (D) Time-dependent changes in fluorine levels in the blood of
rats after intravenous injection of HFPD (80 mg/kg) (n = 6 rats per group).
(E) Distribution of fluorine levels in major organs of rats, including heart, liver,
spleen, lung, and kidney of rats at 24 and 48 hours after injection. The fluorine
amounts were determined by using ion chromatography after tissue
homogenization using hydrogen peroxide (n = 6 rats per group). (F) Results of
serum biochemical tests conducted on treated rats (n = 6 rats per group).
(G) Hematoxylin and eosin (H&E) staining images of major organs collected
from control and rats administered with HFPD at 24 and 48 hours after
injection. Control rats received PBS injections. (H) Schematic diagram of the
piezoelectric performance measurement of PLA-packaged HFPD-PVA device in
SD rat. (I) Representative voltage generation of the implanted PLA-packaged
HFPD-PVA device in the rat knee area.

amounts of incorporated HFPD, indicating
the degradation of their stretchability (Fig.
4E). Neat PVA possessed a tensile strength of
14.5 ± 4.43 MPa, Young’s modulus of 223.18 ±
53.88 MPa, and elongation at break of 63.92 ±
18.37%. For films with a HFPD-to-PVA ratio
less than 1.5:1, the crystalline samples were
not distributed homogeneously on the films
(fig. S13), which made their mechanical proper-
ties fluctuate within a certain range. When the
HFPD-to-PVA ratio was greater than 1.5, we
observed a substantial increase in Young’s mod-
ulus. Hence, the films within the HFPD-to-PVA
ratio of 1.5 to 2.5 demonstrate the most balanced
stretchability and Young’s modulus (Fig. 4E).
The combination of PVA and HFPD piezoelec-
tric crystals largely maintains the good me-
chanical flexibility of polymers but sacrifices
the strength to some degree.
XRD spectra obtained from the as-prepared
HFPD-PVA (2:1) films showed one prominent
diffraction at 19.1°, which corresponds to the
(012) plane of the HFPD single crystal (Fig. 4D).
From the perspective of crystallography,
the (012) plane of HFPD crystal is the 2D hydro-
gen bond layer, where HFPD molecules are
connected with the adjacent ones through
O–H···O hydrogen bonding interactions (Fig.
4D). Thus, it could be rationally suggested
that the continuous crystallization of HFPD
with orientation in the as-prepared films
could mainly account for the induction of
piezoelectricity. We examined the piezoelec-
tric performances of films made from different
HFPD-to-PVA ratios by using the Berlincourt
method (Fig. 4F). The highest piezoelectric
coefficient d33 (~34.3 pC/N) was observed in
the film with a 2:1 ratio. This value is com-
parable with that of the typical ferroelectric
polymer PVDF and far beyond those of most
reported bio-organic films such as DL-alanine
(4.8 pC/N) (30), collagen (0.89 pC/N) (31), and
g-glycine (10.4 pC/N) (22). On the HFPD-
PVA film with the ratio of 2:1, the measured
d33 values were distributed from 32 to 37 pC/
N (table S10), confirming the good uniformity
of this piezoelectric film. Because of their low
permittivity (table S11), the piezoelectric volt-
age coefficient g33 value of HFPD-PVA film with
the ratio of 2:1 was calculated to be 1587.6 ×
10−3 Vm/N. This value is far beyond that of
PZT ceramics (20 × 10−3 to 40 × 10−3 Vm/N)
and an order of magnitude larger than that of
PVDF. Such a high g33 value indicates a large
voltage output of the composite film, which
is critical to the performance of piezoelectric
sensors. Films with other HFPD-to-PVA ratios
also exhibited appreciable g33 values (Fig.
4F and table S11).
We further investigated the piezoelectric per-
formance of the HFPD-PVA films by measuring
the generated output voltages after repetitively
striking them with a certain external mecha-
nical stress. An impulse force of 40 N was re-
peatedly applied to the film surface at a frequency
of 4 Hz. The as-prepared HFPD-PVA (2:1) films
exhibited a stable output peak-to-peak voltage
(Vpp) of around 18 V (Fig. 4G). We measured
the force- and frequency-dependent piezo-
electric output of the HFPD-PVA (2:1) films.
With the force and frequency increasing, the
generated output voltages gradually increased
(fig. S15). Because both PVA and HFPD are
water soluble, the HFPD-PVA films dissolved
into an aqueous solution after about 1 hour

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1497

RES EARCH | R E S E A R C H A R T I C L E S
(fig. S16). We thus used biodegradable poly-
lactic acid (PLA) encapsulation to construct a
controllable transient electromechanical device,
in which the duration time for the device to work
depends on the encapsulation properties. After
being immersed in the phosphate buffer solu-
tion (PBS) for several days, the packaged com-
posite film was degraded, and the electrical
signal Vpp decreased to zero (Fig. 4G).
Biocompatibility and biodegradability
We confirmed the in vitro biocompatibility of
this piezoelectric crystal by culturing it with
multiple types of cells, such as rat bone marrow–
derived mesenchymal stem cells (rBMSCs),
mouse calvarial pre-osteoblast cell line MC3T3-
E1 cells, and rat neural stem cells (rNSCs). We
conducted cell viability tests in culture medium
supplemented with 1, 5, and 10 mg/ml of HFPD
using the Cell Counting Kit-8 (CCK-8) assay. As
illustrated in Fig. 5A, HFPD demonstrated ex-
cellent biocompatibility toward rBMSCs, with
cell viability exceeding 90% even at a concen-
tration of up to 10 mg/ml. Furthermore, live/
dead and phalloidine/DAPI (4′,6-diamidino-2-
phenylindole) staining were performed to eval-
uate the potential cytotoxicity of HFPD toward
rBMSCs. As depicted in Fig. 5B, nearly all the
cells treated with various concentrations of
HFPD exhibited green fluorescence and simi-
lar spreading behavior accompanied by extended
filopodia compared with that of the control group.
Similar trends could be observed in MC3T3-E1
cells treated with various concentrations of
HFPD (fig. S17). Furthermore, our organic fluo-
ride piezoelectric HFPD showed no neurotoxin,
as evidenced by the high cell viability (fig. S18)
and no impact on the differentiation ability of
rNSCs (fig. S19). Taken together, these results
undoubtedly indicated that HFPD was of great
biocompatibility.
We conducted in vivo biodegradability assess-
ment through tail intravenous injection of HFPD
(80 mg/kg) into Sprague-Dawley (SD) rats
for 24 and 48 hours, as schematically depicted
in Fig. 5C. We measured the time-dependent
examination of the fluorine amounts in blood
(Fig. 5D); major organs including the heart,
liver, spleen, lung, and kidney (Fig. 5E); as well
as excreta, including urine and feces (fig. S20)
using ion chromatography after tissue homog-
enization, and the results further substanti-
ated that HFPD was biodegradable. Orbitrap
high-resolution mass spectrometry of urine in
SD rats intravenously injected with HFPD solu-
tion (80 mg/kg) confirmed that HFPD mole-
cules could be removed from the body (fig. S21).
Additionally, the analysis of serum parameters—
including alanine aminotransferase (ALT),
albumin (ALB), cholesterol (CHO), creatine
kinase-MB (CK-MB), and urea—all fell within
the normal range (Fig. 5F). Histological exa-
minations of major organs revealed no notice-
able structural or cellular alterations (Fig. 5G).
1498 29 MARCH 2024 • VOL 383 ISSUE 6690
Furthermore, the biosafety of HFPD was
validated by implanting a HFPD-PVA film
(10 by 10 mm) under the skin of SD rats. As
expected, all the serum parameters were within
the normal range (fig. S22). Neither major or-
gans (fig. S23) nor the skin tissue directly above
the HFPD-PVA film (fig. S24) exhibited discer-
nible inflammatory cell infiltration or patho-
logical features. To further extend the potential
biomedical application of HFPD-PVA film, we
constructed the PLA-packaged HFPD-PVA (2:1)
device and tested its piezoelectric performance
in vivo in SD rats. We implanted devices under
the skin in the posterior limb for substantial
mechanical-electrical transduction detection
(Fig. 5H). When we briefly pulled and released
the rat’s right leg at a certain beat, the em-
bedded device that was attached to the knee
area produced the corresponding electrical
signal Vpp of ~300 mV (Fig. 5I). All of these
results demonstrated that HFPD had great
biodegradation and biosafety, and these supe-
rior performances suggest that HFPD-PVA films
possessed great application potential as carriers
for drug delivery, self-powered transient energy-
harvesting devices for therapeutics, and bio-
degradable high-performance piezoelectric
scaffolds for regenerative medicine.
Summary
We have presented a simple ferroelectric mol-
ecular crystal HFPD with piezoelectricity better
than PVDF, and the obtained piezoelectricity
does not require the pretreatments of stretch-
ing and poling. An intriguing 2D hydrogen
bond network in HFPD shows distinct advan-
tages for film-forming fabrication and pollution-
free recycling. The HFPD-PVA films not only
exhibit excellent flexibility, high biocompatibility,
and no obvious toxicity to biological cells, they
also show great biodegradation and biosafety
in physiological environments. Our finding
provides a fundamental template and potential
inspiration for the development of transient
implantable electromechanical devices.
RE FERENCES AND NOTES
1. Y. Xu, Ferroelectric Materials and Their Applications (Elsevier,
2013).
2. F. Li et al., Science 364, 264–268 (2019).
3. S. H. Baek et al., Science 334, 958–961 (2011).
4. J. F. Scott, Science 315, 954–959 (2007).
5. Y.-M. You et al., Science 357, 306–309 (2017).
6. J. Harada et al., J. Am. Chem. Soc. 141, 9349–9357
(2019).
7. E. Lage et al., Nat. Mater. 11, 523–529 (2012).
8. M. E. Lines, A. M. Glass, Principles and Applications of
Ferroelectrics and Related Materials (Oxford Univ. Press, 2001).
9. J. Harada et al., J. Am. Chem. Soc. 140, 346–354
(2018).
10. W.-Q. Liao et al., Science 363, 1206–1210 (2019).
11. J. Y. Li, R. C. Rogan, E. Ustündag, K. Bhattacharya, Nat. Mater.
4, 776–781 (2005).
12. G. Huangfu et al., Science 378, 1125–1130 (2022).
13. C. Qiu et al., Nature 577, 350–354 (2020).
14. X. Qian, X. Chen, L. Zhu, Q. M. Zhang, Science 380, eadg0902
(2023).
15. M. Guo et al., Science 371, 1050–1056 (2021).
16. L. Gao et al., Science 381, 540–544 (2023).
17. H.-Y. Zhang, R.-G. Xiong, Science 381, 484–485
(2023).
18. Y. Liu et al., Nature 562, 96–100 (2018).
19. X. Chen et al., Science 375, 1418–1422 (2022).
20. M. T. Chorsi et al., Sci. Adv. 9, eadg6075 (2023).
21. Y. Liu et al., Sci. Transl. Med. 14, eabi7282 (2022).
22. F. Yang et al., Science 373, 337–342 (2021).
23. T. Someya, Z. Bao, G. G. Malliaras, Nature 540, 379–385
(2016).
24. C. Li et al., Nat. Rev. Mater. 5, 61–81 (2020).
25. Y. Jiang et al., Science 375, 1411–1417 (2022).
26. J. Xu et al., Science 355, 59–64 (2017).
27. S. Guerin et al., Nat. Mater. 17, 180–186 (2018).
28. A. Heredia et al., Adv. Funct. Mater. 22, 2996–3003
(2012).
29. A. Stapleton et al., Appl. Phys. Lett. 111, 142902 (2017).
30. D. Kim et al., Adv. Mater. 32, e1906989 (2020).
31. D. Denning et al., ACS Biomater. Sci. Eng. 3, 929–935
(2017).
32. F. S. Foster, K. A. Harasiewicz, M. D. Sherar, IEEE Trans.
Ultrason. Ferroelectr. Freq. Control 47, 1363–1371 (2000).
33. J. Curie, P. Curie, C. R. Acad. Sci. Paris 91, 294–297 (1880).
34. Materials and methods and discussions are available as
supplementary materials.
35. A. Gruverman, M. Alexe, D. Meier, Nat. Commun. 10, 1661
(2019).
36. D. Lee et al., Science 349, 1314–1317 (2015).
37. S. Kang et al., Science 376, 731–738 (2022).
38. W.-J. Xu et al., J. Am. Chem. Soc. 142, 16990–16998 (2020).
39. T. Tang et al., Natl. Sci. Rev. 10, nwad177 (2023).
40. S. Bhunia et al., Science 373, 321–327 (2021).
41. Z.-X. Zhang et al., Angew. Chem. Int. Ed. 61, e202210809
(2022).
42. S.-E. Park, T. R. Shrout, J. Appl. Phys. 82, 1804–1811 (1997).
43. D. A. Berlincourt, D. R. Curran, H. Jaffe, “Piezoelectric and
piezomagnetic materials and their function in transducers” in
Physical Acoustics: Principles and Methods, vol. 1, W. P. Mason,
Ed. (Academic Press, 1964), pp. 169–270.
44. C. Rai, B. Narayana Moolya, S. M. Dharmaprakash, Physica B
406, 1–7 (2011).
45. Q. Yu et al., Nano Energy 121, 109196 (2024).
46. A. Zoroddu, F. Bernardini, P. Ruggerone, V. Fiorentini,
Phys. Rev. B Condens. Matter 64, 045208 (2001).
47. J. L. Lutkenhaus, K. McEnnis, A. Serghei, T. P. Russell,
Macromolecules 43, 3844–3850 (2010).
AC KNOWLED GME NTS
We thank Y.-G. Wang (Peking University) for help with the test and
analysis of high-pressure single-crystal XRD experiments. Funding:
This work was supported by the National Natural Science
Foundation of China (21991142, 12304005, 32301264, and
22375082) and the Ten Science and Technology Problem of
Southeast University. Author contributions: H.-Y.Z. carried out
general characterizations, prepared the composite piezoelectric
films, conceived the study, and wrote the manuscript. X.-G.C.
performed crystallographic analysis. H.-P.L. carried out
ferroelectric and piezoelectric characterization. Z.X.G., P.W., and
Q.J. carried out the biocompatibility and biodegradability
assessments and biological implantable application. P.-F.L.
performed the computational studies. Y.-Y.T. carried out the PFM
study. N.G. and S.R. analyzed the data and wrote the manuscript.
R.-G.X. conceived the study and wrote the manuscript, with
input from other authors. Competing interests: The authors
declare no competing interests. Data and materials availability:
All data are available in the manuscript or the supplementary
materials. License information: Copyright © 2024 the authors,
some rights reserved; exclusive licensee American Association for
the Advancement of Science. No claim to original US government
works. https://www.science.org/about/science-licenses-journal-
article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adj1946
Materials and Methods
Supplementary Text
Figs. S1 to S24
Tables S1 to S11
References (48–159)
Movie S1
Submitted 12 June 2023; resubmitted 16 October 2023
Accepted 7 February 2024
10.1126/science.adj1946
science.org SCIENCE
 RESE ARCH | R E S E A R C H A R T I C L E S

METHANE EMISSIONS ing (SEM) walking surveys. These generally

involve a person equipped with a low-sensitivity
Quantifying methane emissions CH4 detector (e.g., a flame ionization detector)
walking along a serpentine path across por-
from United States landfills tions of the landfill and logging the coordinates
of any exceptionally high detected surface
Daniel H. Cusworth1,2*, Riley M. Duren1,2,3, Alana K. Ayasse1, Ralph Jiorle1, Katherine Howell1, concentrations >500 ppm [40 CFR 63.1958(d);
Andrew Aubrey1, Robert O. Green3, Michael L. Eastwood3, John W. Chapman3, Andrew K. Thorpe3, 40 CFR 63.1960(c) and (d)]. Walking surveys are
Joseph Heckler4, Gregory P. Asner4, Mackenzie L. Smith5, Eben Thoma6, Max J. Krause6, complicated by the fact that many locations on
Daniel Heins6, Susan Thorneloe6 an active landfill are unsafe to measure (e.g.,
walking along face areas where new trash is
Methane emissions from solid waste may represent a substantial fraction of the global anthropogenic deposited or steep side slopes). SEM surveys
budget, but few comprehensive studies exist to assess inventory assumptions. We quantified emissions are federally required only four times per year
at hundreds of large landfills across 18 states in the United States between 2016 and 2022 using for most landfills, which limits their ability to
airborne imaging spectrometers. Spanning 20% of open United States landfills, this represents the most capture any dynamics in emissions. SEM sur-
systematic measurement-based study of methane point sources of the waste sector. We detected significant vey accuracy is also highly dependent on the
point source emissions at a majority (52%) of these sites, many with emissions persisting over multiple human operator and exact choice of measure-
revisits (weeks to years). We compared these against independent contemporaneous in situ airborne ment locations, with the result that high-emission
observations at 15 landfills and established good agreement. Our findings indicate a need for long-term, locations potentially can be missed entirely.
synoptic-scale monitoring of landfill emissions in the context of climate change mitigation policy. Additionally, SEM measurements do not ex-
plicitly represent an emission rate, but instead
are designed to flag CH4 concentration “hot
andfill methane (CH4) emissions are esti- with buried organic waste will generate CH4 spots” that may indicate a potential regulatory

L
mated to make up nearly 20% of global
anthropogenic CH4 emissions (years 2000
to 2017) (1) and 17% of US anthropogenic
CH4 emissions (years 1990 to 2020) (2).
However, these estimates are almost entirely
driven by bottom-up process models that have
not been comprehensively validated by direct
measurement across a broad population of
global landfills and dumpsites. Landfill gas
emission models generally rely on waste ton-
nage, decay parameterizations, and some es-
timate of gas capture, if applicable (75% is
commonly used in US), to estimate annual
CH4 emissions (3, 4) These parameters are
difficult to generalize because they rely on fac-
tors intrinsic to a particular landfill, regional
waste stream (e.g., proportion of organic mate-
rial), operator practices, and jurisdictional over-
sight. To the best of our knowledge, direct
measurements of CH4 emissions at landfills
to date using surface or aircraft instruments
have largely been limited to a small number
of facilities due primarily to cost, which has
resulted in incomplete spatial and temporal
sampling. Given the diversity of operational and
environmental factors driving landfill emis-
sions, these observational limitations lead to
continued uncertainty in this sector’s contri-
bution to regional, national, and global CH4
emission inventories, which can complicate
assessing the efficacy of emission mitigation
efforts.
CH4 emissions at solid waste sites result
from several processes. Nearly every location
1 from hours to months. For safety reasons,
Carbon Mapper, Pasadena, CA, USA. 2Arizona Institutes for
Resilience, University of Arizona, Tucson, AZ, USA. 3NASA
Jet Propulsion Laboratory, Pasadena, CA, USA. 4Center for Global
Discovery and Conservation Science, Arizona State University,
Tempe, AZ, USA. 5Scientific Aviation, Boulder, CO, USA. 6US
Environmental Protection Agency, Washington, DC, USA.
*Corresponding author. Email: dan@carbonmapper.org
gas at some time scale (5, 6). A fraction of that
generated gas may escape to the atmosphere
through transport or diffusion through soil
layers by taking the path of least resistance
(7). Therefore, changes in barometric pressure
have been shown to influence emission varia-
bility (8–10). However, current datasets are
insufficient to represent pressure-emission re-
lationships at typical landfills with variable
topography, landfill design and operation,
waste composition and quantity, gas capture
and collection, water management, and daily
working face design and operation (10). Fugi-
tive emissions of landfill gas can be caused
by undersized air pollution control equip-
ment, cracks in cover due to drought, side
slope erosion, and how the working face is
operated. Emissions can also result from ex-
treme precipitation events, because landfill
gas wells can be disconnected from landfill
gas header pipes because of the high level of
liquids. These emissions may manifest as a
distributed diffuse “area source” over a wide
area of the waste site or as a “point source”
localized to a certain region or hot spot of the
site. Waste sites may also contain multiple
point or area sources (or both) at any given
time. Area source emissions are constrained
by the CH4 generation potential at a landfill,
but point sources are more likely related to
the dynamic operational nature of a landfill.
For example, planned maintenance or con-
struction at a landfill or equipment failures
can result in highly concentrated CH4 point
sources that can persist for periods ranging
operators may also “underpull” or apply less
vacuum to a gas collection system to avoid
excess oxygen from entering the soil.
In the US, CH4 is most frequently measured
at landfills through surface emission monitor-
exceedance. Actual quantification of landfill
emission fluxes requires the concurrent obser-
vation of CH4 concentration fields and surface
wind speed, which often requires the use of
sophisticated atmospheric transport modeling.
Other studies have leveraged various ground-
and aerial-based technologies to measure land-
fill gas emissions using technologies such as
eddy covariance, radial plume mapping, tracer
correlation, and flux chambers, among others
(11, 12). However, because of the complexity in-
volved in operating these measurement sys-
tems, these studies are often limited to a small
sample of landfills, making extrapolation to
larger waste sector dynamics difficult.
Although strategies to compare, design, and
scale emission quantification technologies tail-
ored for landfill CH4 quantification continue
to be developed, there is an immediate need to
make a baseline observational assessment of
CH4 emissions across a large swath of waste
sites. Remote sensing offers an efficient meth-
od for surveying widely dispersed waste sites
without costly and time-consuming efforts to
gain access to facilities with surface-based ob-
servations. In 2016 and 2017, airborne imaging
spectroscopy was used to observe >400 active
and closed landfills and waste diversion sites
in California as part of the California Methane
Survey (13). This observational approach is
sensitive to high-emission CH4 point sources
(typically >10 kg h−1 for typical wind speeds
and surface albedo), produces high spatial re-
solution plume maps of emission hot spots,
and can quantify emissions of those hot spots
under adequate observing conditions. Here,
“plume” refers to a region of contiguous pixels
of elevated CH4 concentrations that is ob-
served by an imaging spectrometer and attrib-
utable to landfill gas emissions. These plume
maps have in turn been used to guide operators

SCIENCE science.org 29 MARCH 2024 • VOL 383 ISSUE 6690 1499

RES EARCH | R E S E A R C H A R T I C L E S

in locating emission sources at landfills and

have prompted mitigation (14).
Patterns of high-emission point sources at
landfills revealed by the California Methane
Survey suggest that persistent super-emitter
activity could be prevalent more broadly across
the solid waste management sector in the US.
To test this hypothesis, we generated an ob-
servationally based CH4 dataset spanning a
diversity of US climate zones and jurisdictions,
including repeat observations over multiple
seasons and, in some cases, years. The sites
surveyed in this study represent the largest
airborne or ground-based survey of US land-
fills to date, totaling 250 US Environmental
Protection Agency (EPA) Greenhouse Gas Re-
porting Program (GHGRP) landfills in 18 US
states surveyed between 2018 and 2022.
We analyzed this statistically robust data-
set to assess at what rate point source emis-
sions are prevalent at large (i.e., GHGRP
reporting) managed landfills, how long they
persist, and whether the magnitude of quan-
tified emission rates is consistent with re-
ported values. This dataset is key for future
comparison with other non-US jurisdictions,
especially for regions that lack waste man- Fig. 1. Landfills flown between 2016 and 2022 using the AVIRIS-NG or the GAO. (A) Spatial extent of
agement but are looking to incorporate more US surveys. Blue dots represent all landfills with flyovers, and red dots represent landfills where point sources were
recommended practices for emission miti- detected on at least one overpass. (B) Total number of large (>20,000 MtCO2e reported to GHGRP) landfills
gation and public health improvement. In surveyed by state (light blue) and the number of landfills where we detected point sources in at least one overpass
this study, significant point source emis- (dark blue). (C) Average and SD of detection frequency, also known as persistence (number of detections/number
sions were detected at a majority of landfills, of overpasses), across all surveyed landfills as a function of the number of overpasses. Dashed lines represent
many with emissions persisting over multiple the average persistence for facilities flown at least three times on three different days.
revisits spanning several months and, in some
cases, over multiple years. These results show
the need for sustained measurements at land-
fills to provide operator guidance and to bet-
ter constrain emission variability.

Results
Landfills surveyed between 2018 and 2022
across the US, as well as landfills surveyed
between 2016 and 2017 during the California
Methane Survey (6), are summarized in Fig.
1A. These surveys deployed either the Next-
Generation Airborne Visible/Infrared Imag-
ing Spectrometer (AVIRIS-NG) operated by the
NASA Jet Propulsion Laboratory or the equiv-
alent imaging spectrometer onboard the Glo-
bal Airborne Observatory (GAO) operated by
Arizona State University. Both spectrometers
measure solar backscattered radiance from
380 to 2500 nm with 5-nm spectral sampl-
ing, enabling the estimation of atmospheric
column CH4 concentrations using a retrieval
algorithm. We used the columnwise matched Fig. 2. Example of multiple persistent point sources at a landfill. (A) At least two plumes can be seen
filter retrieval tuned to the CH4-absorbing wave- detected from a single overflight in May 2021. (B) The plume emanating from the blue circled region
lengths between 2200 and 2400 nm (15), con- potentially corresponds to a vent or unlit flare. (C to E) Emission from this vent persisted across all other
sistent with the California Methane Survey. overflights between May 2021 and June 2022. Visible basemaps are provided by Google Earth.
Typical flight altitudes ranged between 3 and
5 km above ground level, resulting in CH4 plume (average 5.5) to provide a basic constraint on in the materials and methods section of the
concentration maps with 3- to 5-m spatial res- emission persistence (number of detections/ supplementary materials.
olution. Surveys were designed to require a number of overpasses) and variability. Infor- The California Methane Survey performed
minimum of three overpasses on different days mation on emission quantification can be found a landscape assessment of 436 facilities across

1500 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE


Fig. 3. Comparison of emission rates at landfills derived using the SA mass-balance approach and
the GAO imaging spectrometer. (A) Fifteen landfill overpasses where comparison between emissions
was possible. The plus symbol indicates simultaneous data acquisition. The asterisk indicates asynchronous
but same-day observations within 2 hours of one another. Error bars represent 1 SD uncertainties on
emission rates. (B and C) CH4 observations at LF06 and LF12, respectively, from (A). The white arrow
indicates wind direction.
the waste sector in California, including active
landfills, closed landfills, dry digestion facili-
ties, and composting facilities (13). The study
only detected large point source emissions at
32 of those facilities, although emissions from
just those facilities made up a disproportion-
ate sector contribution compared with all
aggregated sector emissions quantified in
the survey (oil and gas, livestock, energy, and
wastewater treatment sectors). Of those 32
facilities, 21 (66%) were open landfills that re-
ported CH4 emissions of at least 50,000 metric
tons of carbon dioxide equivalents (MtCO2e) to
the GHGRP and 12 (38%) reported at least
100,000 MtCO2e. Given the observed preva-
lence of high point-source emission rates at
large open landfills in California [i.e., land-
fills that accept large quantities of waste on
an annual basis (4)] and in an effort to expand
coverage across diverse climatic zones and
jurisdictions, this study focused on similarly
classified facilities based on emissions reported
SCIENCE science.org
to the GHGRP, generally >50,000 MtCO2e.
Although they make up only a fraction of each
state’s waste facilities, together these facilities
represent on average 36% (range 3.8 to 81%) of
each state’s anticipated landfill emissions ac-
cording to the GHGRP. The 18 states in this
survey made up 67% of the US municipal
landfill emissions in the GHGRP (2019 report-
ing year). To our knowledge, this study repre-
sents the most systematic measurement-based
study to date of CH4 point sources from the
high-emission solid waste sector, spanning 20%
of ~1200 reported open landfills in the US.
We detected plumes at 52% of the landfills
that we surveyed (Fig. 1B), far exceeding the
point-source detection rate in other CH4 emis-
sion sectors. For example, airborne surveys in
California and the Permian Basin showed that
~0.2% and 1% of infrastructure had detectable
plumes, respectively (13, 16). However, landfills
are complex facilities with anticipated con-
tinuous emissions and are fundamentally dif-
RESE ARCH | R E S E A R C H A R T I C L E S
ferent from other anthropogenic emission
sectors. For example, in the oil and gas sector,
point sources detected by imaging spectrom-
eters are usually clearly associated with inter-
mittent but expected operations (maintenance,
venting, and flaring) or fugitive emissions
(leaks). The higher detection rate of point
sources at large landfills confounds any clear
separation between operational and anoma-
lous emission behavior because some contin-
uous CH4 emissions are always to be expected
at landfills. To underscore this point, Fig. 1C
shows the persistence (i.e., CH4 detection fre-
quency, where persistence equals the number
of detections divided by number of overflights)
of CH4 at surveyed landfills compared with the
persistence of oil and gas infrastructure in the
Permian Basin (16). As evidenced by the US
Environmental Protection Agency’s recently
finalized Super Emitter Program for oil and
gas production and proposed changes to the
Greenhouse Gas Reporting Program, high-
emission point sources (>100 kg h−1), regard-
less of persistence, are important to identify
given their contribution to net emissions. Ad-
ditionally, any such sources that persist over
time may indicate anomalous behavior (e.g.,
leaks, malfunction) that warrant expedited
repair. For the Permian, the average persist-
ence for facilities with at least three overflights
is 0.26, whereas for landfills, the persistence is
a higher 0.60.
Related to persistence, we calculated the
timescale or duration of point source activity
for each landfill during its period of observa-
tion (17). This metric is calculated as the length
of time that point sources were detected at a
landfill divided by the length of time the land-
fill was observed. We found a bimodal distribu-
tion across landfills (see the supplementary
materials, section S2), meaning that there ex-
ists a population of landfills where point source
activity only was observed for a short period of
time and another distinct population of land-
fills where point source activity apparently
persisted across nearly the entire observing
record. This long-duration population repre-
sents >60% of all landfills and 87% of all
quantified emissions. These results highlight
the distinct nature of point sources at landfills
compared with other sectors. Within the oil
and gas sector, point source duration time
scales are also bimodally distributed, but long-
duration sources make up a smaller fraction of
all point sources (17), whereas the majority of
landfills with point source detections show
more persistent and long-lasting emission
activity, highlighting the fundamentally dif-
ferent activities, equipment, and dynamics of
these sectors. In particular, mitigating persis-
tent landfill sources potentially poses a greater
climate benefit because they make up an out-
sized contribution of total emissions from that
sector and are more readily attributable and
29 MARCH 2024 • VOL 383 ISSUE 6690 1501
RES EARCH | R E S E A R C H A R T I C L E S
verifiable with on-site leak detection and re-
pair protocols.
Point source CH4 emissions at landfills may
result from complex operational dynamics, in-
cluding the constant movement of the active
or working face, maintenance of the gas cap-
ture system, delays between waste burial and
gas collection installation, construction of new
waste cells, etc. There may also be operational
inefficiencies or exceptional circumstances that
lead to emissions (e.g., poor maintenance of
cover material, insufficient vacuum applied to
wells, flooded wells, droughts creating cracks
in cover). This dynamic environment where
multiple factors could lead to point source
emissions may explain the higher detection
rate and persistence of CH4 point sources.
High-resolution plume maps can aid in un-
covering information about processes that lead
to point sources. However, ground informa-
tion from landfill inspections is also vital to
connecting observations to processes, because
causes of emissions may be due to subsurface
processes or small surface features that are
difficult to discern even with high-resolution
aerial imagery. During these surveys, data
were shared with several operators and feed-
back solicited regarding potential causes of
the detected point sources. Although we only
received limited responses, many plumes were
confirmed to be near the active working face,
near compromised wellheads, or in areas
where wells were being drilled. Sustained
efforts are needed to connect detections and
quantification to specific practices so that a
better database of emission factors can be de-
veloped to help improve management prac-
tices and understand the causes of landfill
CH4 emissions.
We observed many plumes where source
attribution was much clearer when compared
with high-resolution visible imagery. These
cases usually corresponded to easily distin-
guishable gas capture infrastructure. For exam-
ple, we surveyed a landfill in the southern US
in May 2021, October 2021, May 2022, and
June 2022 (Fig. 2, A to E). We observed ex-
ceptionally large (2000 to 6000 kg CH4 h−1)
plumes emanating from multiple points across
the face of the landfill at every airborne over-
pass. However, to the east of these massive
plumes is a smaller, though still significant,
plume emanating from gas capture infrastruc-
ture. Figure 2B shows a close-up of this facility.
The origin of the plume appears to be from
an unlit flare or vent stack, and plumes are
detected at every overpass between 2021 and
2022. Although the massive landfill plumes
to the west are larger in magnitude, the per-
sistent emissions from the unlit flare are still
concerning because these are emissions that
are not expected with a functioning gas cap-
ture and control system. At a minimum, these
excess emissions are normally expected to be
1502 29 MARCH 2024 • VOL 383 ISSUE 6690
flared instead of vented. To detemine whether trometers used in this study, which only de-
any flaring occurred at this site in between our tect strong point sources, SA measures the net
overflights, we queried satellite fire detections emission flux from a landfill, including the
using MODIS and VIIRS day and nighttime sum of diffuse area source fluxes and point
overpasses (18). The satellites did not discover source fluxes. Therefore, in comparing imag-
any thermal signatures indicative of flaring in ing spectrometer-derived emissions with SA,
the vicinity of this site. When we take the we would expect the former to produce lower
average emission rate from all overpasses of emission estimates than SA if at the time of
this flare stack (1470 ± 720 kg CH4 h−1) and overpass there are significant contributions
integrate across the ~12 months of observa- from area sources. However, if the net landfill
tions, the total emissions are 12,900 metric flux at the time of overpass was dominated by
tons CH4 or 322,000 MtCO2e. For reference, strong point source emissions, then we would
GHGRP reports CH4 emissions of 2,920,000 expect SA and imaging spectrometer–derived
MtCO2 e for this state’s total landfill sector, emissions to be comparable. SA generally re-
so this single large point source is equivalent quires 30 to 40 min of spiral observations to
to 11% of that portion of the state’s inventory. quantify emissions from a facility the size of a
Therefore, as we continue to use observations landfill. This enables as many as three to six
to uncover process-level complexities at land- overpasses with the airborne imaging spec-
fills, there is a population of CH4 mitigation trometer. This approach was initially dem-
candidates in which timely repairs could have onstrated with landfills in California in 2017,
a significant impact. but only a small number involved simultaneous
Emission estimates derived from imaging overflights, and intercomparison of measure-
spectrometers using the Integrated Mass En- ments separated by days to months was af-
hancement method have been evaluated in fected by source variability (13). In this study,
multiple controlled-release experiments and most of the intercomparison flights were con-
independent measurements (19–21). However, ducted simultaneously, as well as a few flights
given the aforementioned complexities with that occurred on the same day but were sepa-
landfills, including topography, meteorology, rated by up to 2 hours.
and multiple plume origin locations, we per- Figure 3 shows a comparison between SA-
formed extensive intercomparison with con- and GAO-derived emission rates that passed
temporaneous airborne surveys using Scientific quality control protocols at 15 landfill over-
Aviation’s (SA’s) mass balance approach (20). passes in several midwestern and southern
This measurement technique uses low-altitude US states (landfill names redacted). GAO emis-
aircraft equipped with cavity-ring down spec- sions represent the average of all imaging
trometers and a wind measurement system spectrometer observations acquired during
to conduct spiral surveys around a facility at an SA observation window. The results are
various altitudes (generally 500 to 1500 m). generally consistent (R2 = 0.69; fig. S7). Figure
The emission rate is calculated by applying 3, B and C, shows a visual example of a landfill
Gauss’s theorem to observed concentrations in which the GAO-derived emission rate was
and wind speeds. Unlike the imaging spec- smaller than the SA-derived rate (Landfill 06)
Fig. 4. Comparison of aerial emission rates with EPA GHGRP for landfills where point sources were
detected at least once. (A) Mean CH4 emission rates across all aerial overpasses compared with average
GHGRP emissions. The gray line represents the one-to-one line. The size of the dots corresponds to number
of overflights. Two red-colored dots in (A) correspond to landfills with 5+ years of observations in which
the observed emission trends are significant (P < 0.05). (B) Trends for these landfills. Black squares
represent GHGRP reporting for that year. Error bars represent 1 SD uncertainties on emission rates.
science.org SCIENCE

and an example in which GAO and SA showed
comparable emission rates (Landfill 12). In
both cases, the CH4 plumes observed by GAO
correspond closely with observed downwind
high CH4 concentrations observed by SA. The
generally good agreement between GAO and
SA builds confidence in the broader applica-
tion of the remote sensing method to the larger
population of landfills. Because remote obser-
vations are increasingly being considered as
contributors to routine CH4 emission monitor-
ing over large areas, confidence in emission
quantification is essential for their adoption.
Figure 4A shows a comparison between the
imaging spectrometer–quantified average emis-
sion rates and those emissions reported to the
GHGRP for those facilities. In the US, landfills re-
port emissions according to the Code of Federal
Regulations (40 CFR Part 98 Subpart HH)
either by modeling generated emission from
reported annual waste disposed (HH1/HH6)
or, for landfills with gas capture and collection
systems, from back-calculations based on re-
ported annual gas captured and assumed col-
lection efficiency (HH7/HH8). For landfills
where multiple years of observations are avail-
able through Carbon Mapper flights, we took
the average GHGRP across those years. Poor
correlation exists between aerial emission rates
and GHGRP (R2 = 0.07), which could be ex-
pected under sparse sampling. However, even
for landfills where we surveyed 10+ times
(20 landfills total), we still found little agree-
ment between emission estimates (R2 = 0.02).
This discrepancy appears equally in both di-
rections; there is a population of landfills (47%
of all sites) with aerial emissions that are
higher than GHGRP and a population (53%)
in which they are lower or did not show evi-
dence of any point source emissions. On aver-
age, aerial emission rates were a factor 2.7 higher
than GHGRP for all landfills and a factor 1.4
higher for landfills with 10+ unique over-
passes. Consistent with this study, independent
assessments of US emission inventories have
indicated a needed 1.25 to 1.5 scaling of waste
emissions to reconcile inventories with in situ
ground-based measurements and coarse-
resolution satellite observations (22, 23). Fur-
thermore, in some cases, coarse-resolution
satellite instruments (e.g., TROPOMI) can
quantify annualized emissions from individ-
ual landfills that are isolated from other emis-
sion sources (23). These annualized satellite
observations show better correlation with
our airborne datasets (15 landfills total) than
with GHGRP (for details, see the supple-
mentary materials, section S2). We also found
no significant aggregate bias in airborne re-
sults from seasonal and/or diurnal barometric
pressure variability (for details, see the sup-
plementary materials, section S3). Therefore,
our airborne data, although not continuous,
are still in aggregate likely indicative of gen-
SCIENCE science.org
eral trends of discrepancy with national
inventories.
Figure 4B shows yearly averaged aerial emis-
sion estimates from two landfills with 5+ years
of aerial sampling in which trends in emission
rates are significant (P < 0.05) based on an
ordinary least-squares fit to the data. For both
of these landfills, EPA GHGRP indicates an
insignificant trend in emission rates. In the
case of Landfill 17 (LF 17), airborne observa-
tions also suggest at least a factor of 2.5 under-
estimate compared with GHGRP. There could
be some significant operational issues at that
landfill leading to larger than predicted emis-
sions. Figure 4B therefore shows an example
of how one could use top-down information as
a check against reporting or enforcement pro-
tocols. When a significant number of atmo-
spheric observations sustained over many
months to years show persistent discrepancies
and diverging trends with bottom-up process-
based emission estimates, it highlights areas
for attention and action. Additionally, when
coupled with nimble application of emerging
onsite emission assessment approaches, a single
CH4 plume image with sufficient clarity could
trigger an expedited response to guide follow-up
root-cause analysis, improve general practices,
and potentially reduce emissions.
Landfill emissions are composed of some
fraction of spatially dispersed area and local-
ized point sources, but a typical ratio of area to
point source emissions for a given landfill in
most cases remains unknown and may vary
with site operations and environmental con-
ditions. These conditions may affect emission
pathways and gas collection system efficacy in
complex ways. The comparison of GAO and SA
suggests that for landfills with detectable
point sources, these emissions may make up
an outsized contribution against the total CH4
contribution. Airborne and satellite remote
sensing observations can provide some ini-
tial indications regarding the distribution of
physical emission types for landfills through
relatively frequent, wide-area monitoring of
high-emission point sources. In a tiered ob-
serving strategy, the remote sensing data can
be combined with net facility emission esti-
mates from mass balance aircraft (Fig. 3) and
more widespread deployment of continuous
surface monitoring to quantify the contribu-
tion of point sources relative to the net land-
fill emission flux and to better understand
variability.
Discussion
There are at least two use cases for plume-
scale remote sensing of landfill CH4 point
sources. The first is quick detection and pre-
cise geolocation of emission hot spots at a land-
fill. After communicating with some facility
operators, we attributed a few of our detected
plumes to specific operations (e.g., working
RESE ARCH | R E S E A R C H A R T I C L E S
face, well drilling, construction). More effort
is needed to connect the detected emission
hot spots to operations to better understand
the carbon impact of certain management prac-
tices and to help guide operators to areas on
landfills where remediation may be needed.
The EPA Inspector General issued a report in
2020 finding that EPA needs to improve over-
sight of how states implement air emissions
regulations for municipal solid-waste landfills
(14, 24). Much more attention toward Clean
Air Act Compliance is anticipated because of
the ongoing landfill inspections by both fede-
ral and state governments and the interest in
CH4 reduction.
The second use case is quantifying emission
rates to support evaluation of emission factors
used in reporting programs and inventories.
At this stage, we find a large discrepancy and
generally poor correlation between EPA GHGRP
bottom-up emission estimates and what we
observed from airborne platforms. This dis-
crepancy may be partially explained by sam-
pling, but there could also be systematic issues
with the models that underpin reporting pro-
grams. Ultimately, informed comparison of
emission rates derived from atmospheric mea-
surements with bottom-up calculations requires
an improved understanding of the site pro-
cesses that these airborne platforms detect.
However, regardless of root cause, the de-
tection rate and persistence of point source
emissions at landfills and the large magni-
tude of aerial emission rates found in this
study point to potential gaps in landfill mod-
els and/or calculation of emissions reported
to the GHGRP.
Reconciling top-down and bottom-up esti-
mates requires improved accounting for po-
tential point source emissions in inventories
in the context of current regulatory structures.
Unplanned emissions such as unlit flares or
large leaks from gas collection fields (partic-
ularly ones detectable by SEM) may not be
reflected in inventory estimates. Planned main-
tenance activities of limited temporal extent
could be represented at some level. Working
face emission potential is largely unknown. In
the future, in situ measurements and new site
metadata, such as changes in time-resolved gas
collection and maintenance event temporal
tracking, would complement the top-down data,
improving inventories and reducing unneces-
sary emissions. Additional multisensor field
studies are needed to conclusively determine
whether there are systematic biases in landfill
emission models used in CH4 inventories.
Airborne remote sensing platforms have
proven extremely valuable for initial surveys
and baseline assessments of CH4 emissions
across multiple sectors. However, the sus-
tained sampling of landfills recommended
by this study requires systems in which rou-
tine observation is logistically more feasible.
29 MARCH 2024 • VOL 383 ISSUE 6690 1503
RES EARCH | R E S E A R C H A R T I C L E S
Especially outside of the US, where many waste
sites in developing countries lack any form
of management or monitoring, emissions may
be disproportionately large compared with
other sectors and bottom-up models. As coun-
tries move toward incorporating best manage-
ment practices for waste, using atmospheric
measurements that can be deployed at scale
to verify emission reductions will be critical.
Satellites could provide a solution to sam-
pling when they are configured or capable of
scanning large areas frequently and have suf-
ficient sensitivity to point source emissions
(25). Preliminary studies with the TROPOMI,
GHGSat, and EMIT satellite instruments have
identified large point sources at a small sub-
set of global waste sites, many of which lack
management practices geared toward reducing
CH4 emissions (26, 27). The Carbon Mapper
Coalition plans to launch two Planet Tanager
satellites in 2024, which are optimized for CH4
and CO2 point source monitoring from space
and build on advances from NASA’s EMIT mis-
sion (28). This system will provide wide-area
coverage and frequent sampling to quantify
CH4 emissions from a large population of man-
aged and unmanaged waste sites around the
world. Satellites offer the ability to monitor
CH4 emissions from landfills and unmanaged
dumps across regions that are largely inac-
cessible because of workforce and resource
limitations. Satellites also offer more complete
coverage than aircraft in many regions given
high costs, logistics, and airspace restrictions.
Although not a complete solution to waste emis-
sion quantification, the ability to quantify and
precisely geolocate point source emissions rou-
tinely at global scale with a combination of these
remote sensing platforms represents an impor-
tant contribution to this sector. This informa-
tion, combined with multiscale information
from a tiered observing system, can be effec-
tive in accelerating mitigation if efforts to con-
nect observations with operators and regulators
are sustained.
RE FE RENCES AND N OT ES
1. M. Saunois et al., Earth Syst. Sci. Data 12, 1561–1623 (2020).
2. US Environmental Protection Agency, “Inventory of US
greenhouse gas emissions and sinks: 1990-2020” (EPA, 2024);
1504 29 MARCH 2024 • VOL 383 ISSUE 6690
https://www.epa.gov/ghgemissions/inventory-us-greenhouse-
gas-emissions-and-sinks.
3. US Environmental Protection Agency, “Background information
document for updating AP42 section 2.4 for estimating
emissions from municipal solid waste landfills (EPA, EPA/600/
R-08-116, 2008); https://cfpub.epa.gov/si/si_public_record_
report.cfm?Lab=NRMRL&dirEntryId=198363.
4. M. A. Barlaz, J. P. Chanton, R. B. Green, J. Air Waste Manag. Assoc.
59, 1399–1404 (2009).
5. M. A. Barlaz, Global Biogeochem. Cycles 12, 373–380
(1998).
6. M. A. Barlaz, W. E. Eleazer, W. S. Odle, X. Qian, Y. S. Wang,
“Biodegradative analysis of municipal solid waste in
laboratory-scale landfills” (EPA, 1997); https://
19january2017snapshot.epa.gov/www3/epawaste/conserve/
tools/warm/pdfs/Barlaz_Biodegradative_analyis_in_
laboratory_scale_landfills.pdf.
7. L. Theodore, P. S. Farber, in Perry’s Chemical Engineers’
Handbook, D. Green, M. Z. Southard, Eds. (McGraw Hill, 9th
ed., 2019); pp. 22-69–22-93.
8. K. Kissas, A. Ibrom, P. Kjeldsen, C. Scheutz, Waste Manag. 138,
234–242 (2022).
9. L. Xu, X. Lin, J. Amen, K. Welding, D. McDermitt, Global
Biogeochem. Cycles 28, 679–695 (2014).
10. P. M. Czepiel et al., Waste Manag. 23, 593–598 (2003).
11. J. Mønster, P. Kjeldsen, C. Scheutz, Waste Manag. 87, 835–859
(2019).
12. B. W. Mosher et al., Environ. Sci. Technol. 33, 2088–2094
(1999).
13. R. M. Duren et al., Nature 575, 180–184 (2019).
14. D. H. Cusworth et al., Environ. Res. Lett. 15, 054012
(2020).
15. D. R. Thompson et al., Geophys. Res. Lett. 43, 6571–6578
(2016).
16. D. H. Cusworth et al., Environ. Sci. Technol. Lett. 8, 567–573
(2021).
17. D. H. Cusworth et al., Proc. Natl. Acad. Sci. USA. 119,
e2202338119 (2022).
18. National Aeronautics and Space Administration, “Fire
Information for Resource Management System (FIRMS)”
(NASA, 2024); https://firms.modaps.eosdis.nasa.gov/.
19. S. H. El Abbadi et al., Comprehensive evaluation of aircraft-
based methane sensing for greenhouse gas mitigation.
EarthArXiv (2023); https://doi.org/10.31223/X51D4C.
20. A. K. Thorpe et al., Remote Sens. Environ. 266, 112681
(2021).
21. S. Conley et al., Atmos. Meas. Tech. 10, 3345–3358
(2017).
22. X. Lu et al., Atmos. Chem. Phys. 21, 4637–4657 (2021).
23. H. Nesser et al., EGUsphere 2023, 1–36 (2023).
24. US Environmental Protection Agency, “EPA needs to
improve oversight of how states implement air emissions
regulations for municipal solid waste landfills” (EPA,
2020); https://www.epa.gov/office-inspector-general/report-
epa-needs-improve-oversight-how-states-implement-air-
emissions.
25. D. J. Jacob et al., Atmos. Chem. Phys. 22, 9617–9646
(2022).
26. J. D. Maasakkers et al., Sci. Adv. 8, eabn9683 (2022).
27. A. K. Thorpe et al., Sci. Adv. 9, eadh2391 (2023).
28. R. O. Green et al., “The Earth surface mineral dust source
investigation: An Earth science imaging spectroscopy
mission” in 2020 IEEE Aerospace Conference, Big Sky, MT,
USA, 2020 (IEEE, 2020); https://doi.org/10.1109/
AERO47225.2020.9172731.
29. Carbon Mapper Public Data Portal; https://data.
carbonmapper.org.
30. Data for: D. H. Cusworth et al., Quantifying methane emissions
from United States landfills, Zenodo (2024); https://doi.org/
10.5281/zenodo.10642570.
AC KNOWLED GME NTS
The Carbon Mapper team acknowledges the generous
support of its philanthropic donors. Portions of this work were
performed at the Jet Propulsion Laboratory, California
Institute of Technology, under a contract with the National
Aeronautics and Space Administration (80NM0018D0004).
We acknowledge SA pilots A. Healy and R. Glaser for their help
in acquiring airborne mass balance data. We thank V. Mei
and M. Ruffatto and others from EPA region 5 and other regions
who aided in selecting landfill measurement sites. We thank
E. Tseng, E. Ayandele, and T. Frankiewicz for discussions on driving
factors for landfill emissions. The Global Airborne Observatory (GAO)
is managed by the Center for Global Discovery and Conservation
Science at Arizona State University. The GAO is made possible
by support from private foundations, visionary individuals, and Arizona
State University. Measurements conducted by Carbon Mapper
were not performed or funded by the EPA. Any mention of
trade names, manufacturers, or products does not imply an
endorsement by the US Government or the US EPA. EPA and its
employees do not endorse any commercial products, services,
or enterprises. Measurements conducted by SA were conducted
under EPA contract 68HERC20D0018, task order 68HERC21R0193 to
Jacobs Technical, Inc., and have been reviewed in accordance with
EPA policy and approved for publication. Funding: Funding for flight
operations and/or data analysis referenced herein was supported by
the Grantham Foundation for the Protection of the Environment,
Bloomberg Philanthropies, High Tide Foundation, the California Air
Resources Board (CARB), the EPA, the University of Arizona, and
NASA’s Carbon Monitoring System program. Author contributions:
Conceptualization: D.H.C., R.M.D., S.T., E.T.; Data curation: D.H.C.,
A.K.A., R.J., K.H.; Funding acquisition: R.M.D., S.T.; Investigation:
A.A., A.K.T., R.O.G., M.L.E., J.W.C., J.H., G.P.A., M.L.S.; Methodology:
D.H.C., R.M.D., A.K.A., M.L.S.; Supervision: D.H.C., R.M.D, A.A.,
S.T.; Visualization: D.H.C.; Writing – original draft: D.H.C.; Writing –
review and editing: D.H.C., R.M.D., A.K.A., A.K.T., E.T., M.J.K.,
D.H., S.T., G.P.A. Competing interests: The authors declare no
competing interests. Data and materials availability: All data are
available in the main text or the supplementary materials. Plume
datasets are also available for visualization and download publicly
(29, 30). License information: Copyright © 2024 the authors, some
rights reserved; exclusive licensee American Association for the
Advancement of Science. No claim to original US government works.
https://www.science.org/about/science-licenses-journal-article-reuse
SUPPLEMENTARY MATERIALS
science.org/doi/10.1126/science.adi7735
Materials and Methods
Supplementary Text
Figs. S1 to S8
References (31, 32)
Carbon Mapper Emission Data
SA Data
Submitted 17 May 2023; accepted 16 February 2024
10.1126/science.adi7735
science.org SCIENCE
myIDP:

ScienceCareers.org
A career plan customized
for you, by you.

For your career in science, there’s only one

Features in myIDP include:

 Exercises to help you examine your skills,
interests, and values.
 A list of 20 scientific career paths with a
prediction of which ones best fit your skills
and interests.
 A tool for setting strategic goals for the
coming year, with optional reminders to
keep you on track.
 Articles and resources to guide you through
the process.
 Options to save materials online and print
them for further review and discussion.
 Ability to select which portion of your IDP
you wish to share with advisors, mentors,
or others.
 A certificate of completion for users that
finish myIDP.

Visit the website and start

planning today!
myIDP.sciencecareers.org

In partnership with:
 WORKING LIFE
By Emily L. Eberhardt

Shaken up by a shaker

C
rash! Hrmpf! Screech! At 200 revolutions per minute, 4 liters of bacterial culture came loose
from the shaker’s adhesive pads. Facing the pungent dripping liquid and razor-sharp glass
shards, I burst into tears. I had reached my own breaking point. Three months earlier, I
passed my Ph.D. candidacy exam—only for the lab to be shut down the following week as
the COVID-19 pandemic set in. When the university began to reopen, I wanted to be the first
person back in the lab. I needed to make progress. I needed to be productive. But now, I felt
as shattered as my experiment. I didn’t realize then that this moment would force me to confront a
pattern of unhealthy behavior and make myself whole.

I’ve always striven for perfection. aspects of my life, setbacks at work

In elementary school, a big, red A+
at the top of an assignment was my
definition of success. After experienc-
ing a severe illness in eighth grade, I
confidently told my teachers I was go-
ing to grow up to be a scientist and
cure diseases. As an undergraduate, I
spent my Friday nights completing as-
signments early and studying. When I
advanced to graduate school, I had a
perfect 4.0 GPA.
But as a newly minted Ph.D. candi-
date, I struggled to fully understand
the intricacies of complex protocols,
troubleshoot experiments, and man-
age my time. Each day brought more
“I judged myself harshly for
failure, more unknowns, and more
shame. I felt I couldn’t do anything
needing help, but I managed to
right, and early pandemic life only push these feelings aside.”
made matters worse. As a natural
night owl, I would come in midafternoon and work until
the next lab members arrived in the morning. When I hit
snags in my experiments, there were few—if any—colleagues
around to troubleshoot, or even just commiserate.
The incubator incident, and the breakdown that followed,
forced me to face my unhealthy habits. My partner encour-
aged me to seek emergency mental health resources. I was
hospitalized and had to set aside my phone—and my to-do
lists, calendars, and emails—and truly take time off. I judged
myself harshly for needing help, but I managed to push these
feelings aside.
Once I was out of the hospital, I sought therapy and other
resources to help me take care of myself and set clearer bound-
didn’t feel as catastrophic, and I made
meaningful progress in the lab.
Still, every once in a while, I have
fallen right back into my old hab-
its. A few nights before giving a
programwide presentation, I stayed
up all night preparing talking points
and slides. In a sleep-deprived daze,
I hopped on an electric scooter to
head home, hit a bump, and crashed.
From the hard concrete, I stared
up at the sky and promised myself
another reset.
After getting treated for a concus-
sion, I called my therapist. We set
realistic time points for completing a
paper I was working on, instituted a
“minimal work after coming home”
rule, and even added a few vacations
to break up heavy work periods. These
strategies reduced my stress enough for me to complete my
thesis research, advance in my ballet class, and take each
day’s challenges as they came.
I’ve come to accept that mental illness and burnout
won’t just go away, and I’ll have to manage them for the
rest of my life and career. I attend regular therapy appoint-
ments and take medication. And now that I’ve recognized
my pattern—a build-up of stress and anxiety, followed by
an accident, mistake, or error in judgment and a resulting
breakdown—I’m better equipped to stop it from repeating.
I try to incorporate mindfulness techniques, asking myself
questions such as, “How do I feel right now?” or “Do I need
to step away for a minute?” When I notice I’m becoming
ILLUSTRATION: ROBERT NEUBECKER

aries. With the help of my mental health care team, friends,
and partner, I scheduled regular meals, prioritized sleep, and
resumed a beloved former hobby, ballet. However, I was still
hard on myself. I often fixated on how I could have seen the
signs of burnout earlier and done things differently, or worried
about what others would think. But as I started to enjoy more
overwhelmed, I force myself to take breaks. And if all else
fails, I head home, talk it out with my partner, and go to bed
knowing I’ll be able to try again the next day. j
Emily L. Eberhardt is a postdoc at the University of Michigan. Send your
career story to SciCareerEditor@aaas.org.

1506 29 MARCH 2024 • VOL 383 ISSUE 6690 science.org SCIENCE

Celebrating 50 years
of Passion for Science.
We all share a passion for science. Curiosity inspires
us. The research community’s quest to find answers
to pressing global issues, such as improving human
health, addressing food shortages, and combating
climate change, is what drives us.
NEB was created by scientists for scientists, and we
prioritize the advancement of science, stewardship of
the environment, and giving back to the world around
us. Since our establishment in 1974, we have remained
committed to developing high quality, innovative
products that not only empower your research but also
our own. Our profits have always funded an extensive
research program, which we believe is critical for
staying connected to our customers and helping to
drive scientific breakthroughs.
As we reflect on the last 50 years and look toward the
future, we are excited to support your research and
help you address these complex challenges.

We hope that by working together, we can shape

the science of tomorrow.

Thank you for your continued trust and support. Find out
how you can help us celebrate 50 years by visiting
www.neb.com/NEBturns50

Products and content are covered by one or more patents, trademarks and/or copyrights owned or controlled by New England Biolabs, Inc (NEB). The use of trademark symbols does not necessarily indicate that the name is
trademarked in the country where it is being read; it indicates where the content was originally developed. See www.neb.com/trademarks. The use of these products may require you to obtain additional third-party intellectual
property rights for certain applications. For more information, please email busdev@neb.com.
© Copyright 2024, New England Biolabs, Inc.; all rights reserved.
Pioneering Excellence in Immunotherapy Research
ProMab Biotechnologies’ offers a one-stop-shop for contract research
services to enhance antibody, mRNA-LNP, and cell therapy research.
As the world’s first CAR-T CRO and a long-established antibody platform for over 20 years,
ProMab can leverage diverse innovation pathways to fuel your next therapeutic research adventure.

mRNA-Lipid Nanoparticles CAR-T/NK/Macrophage Monoclonal & Bispecific Antibodies

ProMab Biotechnologies’ combines advanced synergies under one umbrella:

Completely validated mRNA-LNP platform supporting formulation and manufacturing
Chimeric antigen receptor design and development for a plethora of immune cell types
Antibody production and engineering offerings for research and therapeutic applications

Imagine the possibilities partnering with our end-to-end services.

For more details, scan theDiscover
QR codemore
or visit www.promab.com
| www.ProMab.com
Discover more | www.promab.com

2600 Hilltop Dr, Building B, Richmond, CA 94806

1.866.339.0871 | info@promab.com