ll
Leading Edge
Editorial
Science Has a Racism Problem
We are the editors of a science journal, committed to publishing
and disseminating exciting work across the biological sciences.
We are 13 scientists. Not one of us is Black. Underrepresentation
of Black scientists goes beyond our team—to our authors, re-
viewers, and advisory board. And we are not alone. It is easy
to divert blame, to point out that the journal is a reflection of
the scientific establishment, to quote statistics. But it is this
epidemic of denial of the integral role that each and every mem-
ber of our society plays in supporting the status quo by failing to
actively fight it that has allowed overt and systemic racism to
flourish, crippling the lives and livelihoods of Black Americans,
including Black scientists.
Science has a racism problem.
Look to the history of human genetics, a field that has been
used repeatedly as scientific rationale for the definition of human
‘‘races’’ and to support inherent inequalities. Proponents of eu-
genics use the alleles we carry as reason to declare racial supe-
riority, as if expression of a lactase gene has bearing on one’s hu-
manity. Race is not genetic.
Look to the exploitation of Black research subjects. Acknowl-
edge the sheer volume of past and current scientific research
made possible by cells stolen decades ago from Henrietta
Lacks, a Black woman with cancer. Remember the Tuskeegee
syphilis study that intentionally withheld appropriate treatment
to hundreds of Black men. Think about the issues of consent,
of ownership, and of medical ethics and do not overlook the
shared role of race in these violations.
Look to the extreme disparity in the genetic and clinical data-
bases scientists have built, with the overwhelming majority of
data from white Americans of European descent and the result-
ing dearth of understanding of health and disease in Black indi-
viduals. Read statistics about morbidity and mortality disparities
in hospitals around the country, highlighted by the current
pandemic—ask why Black women are five times more likely
than white women to die during pregnancy, or why Black infants
are twice as likely to die as white babies born in the US. Black
health has never been the priority.
Science has a racism problem. And it is not limited to scientific
discoveries and their attendant usage. The scientific establish-
ment, scientific education, and the metrics used to define scien-
tific success have a racism problem as well.
Black Americans face a mountain of challenges built on cen-
turies of systemic structural racism and the United States’ his-
tory of slavery and racial oppression. Educational opportunities,
mentorship and representation, and our ingrained, often uncon-
scious attitudes all play a role. The gatekeeping system in
academia, industry, and scientific organizations was not de-
signed to correct for centuries of compounded disadvantage
and oppression. It is time for renovation.
We urge our community members who have the means to
enact change to do so. Hiring committees, educators, mentors,
admissions committees, classmates, researchers—what can
you do to raise up Black students and colleagues in your com-
munities and institutions? None of us individually can stem the
tide of racism or rebuild an unjust society, but every ac-
tion helps.
We are part of the problem, as are all of us who do not press for
change on a daily basis. It should not have taken the recent
deaths of George Floyd, Breonna Taylor, and Ahmaud Arbery
for us to speak and to act. We are asking ourselves what we
can do to be stronger allies, stronger anti-racists.
Cell stands with our Black readers, reviewers, authors, and
colleagues. We are committed to listening to and amplifying their
voices, to educating ourselves, and to finding ways that we can
help and do better. We alone cannot fix racism. But we have the
advantage of having a platform, so we will put in the work, we will
listen, and we will act.
As a start, we are committing to the following actions to high-
light and increase representation of Black scientists:
1. Representing – we will feature and amplify Black and other
underrepresented minority authors of Cell papers on so-
cial media. If you are a person of color and you wish to
be highlighted in this way, please tell us. Email the editor
of your paper with the subject line ‘‘Faces of Cell’’ at any
point in the publication process, and we will be honored
to post about your paper with your photo and/or your
Twitter handle and to re-tweet and amplify your own posts
and stories.
2. Educating – we are committed to featuring issues of
importance to the scientific community in our pages.
We pledge to purposefully highlight Black authors and
perspectives in the review and commentary content
that we commission and publish and to share these
with the greater scientific community. Has your depart-
ment or institute already made changes or launched
successful initiatives? Tell us, and we will try to find
ways to share those stories. Have new ideas? Let
us know.
3. Diversifying – we pledge to improve the diversity of our
advisory board and our reviewer pool, using our experi-
ence with gender equity initiatives to increase represen-
tation of non-white scientists, which is far too low. We
are actively investigating ways to improve diversity
through our outreach, recruiting, and hiring efforts, at
Cell and across Cell Press. If you are a Black scientist
with an interest in editorial careers, get in touch. We’re
eager to talk.
4. Listening – we are editors because we want to learn. If
there are ways that we can use our voice and our platform
to help the Black scientist community, we want to hear
them. Please email us if you have concrete ideas for
Cell 181, June 25, 2020 ª 2020 Published by Elsevier Inc. 1443
 ll
perspectives you want to see or creative ways that you
think we can help. We promise to hear them.
We and our colleagues across Cell Press hope to serve as one
small part of amplifying Black voices in STEM, and this is just the
Editorial
beginning. We are learning, and we will almost certainly make
mistakes along the way. But silence is not, and never should
have been, an option.
Science has a racism problem. Scientists are problem solvers.
Let’s get to it.

The Cell Editorial Team

https://doi.org/10.1016/j.cell.2020.06.009

1444 Cell 181, June 25, 2020

 ll
Leading Edge

Commentary
How Support of Early Career Researchers
Can Reset Science in the Post-COVID19 World
Erin M. Gibson,1,12,* F. Chris Bennett,2 Shawn M. Gillespie,3 Ali Deniz Güler,4 David H. Gutmann,5 Casey H. Halpern,6
Sarah C. Kucenas,4 Clete A. Kushida,1 Mackenzie Lemieux,7 Shane Liddelow,8 Shannon L. Macauley,9 Qingyun Li,10
Matthew A. Quinn,11 Laura Weiss Roberts,1 Naresha Saligrama,5 Kathryn R. Taylor,3 Humsa S. Venkatesh,3 Belgin Yalçın,3
and J. Bradley Zuchero6
1Department of Psychiatry and Behavioral Sciences, Stanford University School of Medicine, Palo Alto, CA 94305, USA
2Department of Psychiatry, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA 19104, USA
3Department of Neurology and Neurological Sciences, Stanford University School of Medicine, Palo Alto, CA 94305, USA
4Department of Biology, University of Virginia, Charlottesville, VA 22903, USA
5Department of Neurology, Washington University School of Medicine, St. Louis, MO 63110, USA
6Department of Neurosurgery, Stanford University School of Medicine, Palo Alto, CA 94305, USA
7The Salk Institute for Biological Studies, La Jolla, CA 92037, USA
8Neuroscience Institute, NYU School of Medicine, New York, NY 10016, USA
9Department of Internal Medicine, Wake Forest School of Medicine, Winston-Salem, NC 27101, USA
10Departments of Neuroscience and Genetics, Washington University School of Medicine, St. Louis, MO 63110, USA
11Department of Pathology, Section on Comparative Medicine, Wake Forest School of Medicine, Winston-Salem, NC 27517, USA
12Lead Contact

*Correspondence: egibson1@stanford.edu
https://doi.org/10.1016/j.cell.2020.05.045

The COVID19 crisis has magnified the issues plaguing academic science, but it has also provided the scien-
tific establishment with an unprecedented opportunity to reset. Shoring up the foundation of academic sci-
ence will require a concerted effort between funding agencies, universities, and the public to rethink how we
support scientists, with a special emphasis on early career researchers.

The novel coronavirus, SARS-CoV-2, has
placed science at the center of every con-
versation, amplifying the importance of
scientific research to economic stability,
healthcare infrastructure, and disaster
preparedness. In academic science, re-
covery from the immediate COVID19
crisis will require departments, univer-
sities, private foundations, federal
agencies, and the public to work together
collaboratively and comprehensively. The
goal of recovery should not be to return to
‘‘normal’’ but, rather, to reset. Here, we
argue that recovery provides us with the
opportunity to address three systemic is-
sues that plague the conduct of research
in the twenty-first century, with an
emphasis on supporting early career re-
searchers who are the most vulnerable.
The strategies needed to ensure stability
and success of early career scientists
post-COVID19 can be adapted to chip
away at the systemic issues affecting the
scientific establishment.
Excess Does Not Equal Excellence
Science has changed immensely over the
past 50 years. More has become better:
more experiments per paper, more pa-
pers per year, more expectations and re-
quirements for grants and tenure, more
opinions from reviewers. The scientific
community rewards quantity over quality.
Most scientists can easily name a seminal
paper; many were published long before
the 2000s, and many had, at most, a
handful of figures. Today, papers are
often published with a plethora of supple-
mental figures that will largely go unread
and underappreciated. The desire for
‘‘more’’ results in delays in publication,
the awarding of grants, and career
advancement for early career re-
searchers; it also stymies creativity and
encourages the proliferation of low-qual-
ity journals.
Diversification Leads to Discovery
This crisis is exacerbating the well-docu-
mented discrimination afflicting academic
science (Monroe et al., 2008). Women,
parents, and individuals who identify as
racial or ethnic minorities leave science,
technology, engineering, and math
(STEM) fields as early career researchers
at an excessively high rate in the best of
times and will undoubtedly suffer more
from the present lab closures. The re-
sponsibilities of family life disproportion-
ately impact women. A parent who is
trying to homeschool their children,
manage household duties, and work will
have left little time to further their own
scientific agenda. Faculty with family re-
sponsibilities—women specifically—
must be supported. The COVID19 crisis
will only highlight the rampant diversity is-
sues plaguing the scientific establish-
ment, many of which begin with the
loss of women and minorities during
early career stages and may lead to
further disenfranchisement of the disad-
vantaged (Malisch et al., 2020).
Rethink the Fundamentals of
Funding
The current model of academic science is
heavily reliant upon federal funding, even
though agencies such as the National In-
stitutes of Health (NIH) were not built to
sustain such expectations. The federal
government’s funding capacity has signif-
icantly diminished as the cost of science
has radically increased. The 2019 defense

Cell 181, June 25, 2020 ª 2020 Elsevier Inc. 1445

 ll
budget was $685 billion while the 2019
NIH budget was $39 billion. The COVID19
crisis has clearly amplified that the great-
est risk to American life is not war, but dis-
ease. Funding is needed at all levels; how-
ever, early career researchers should be
particularly supported as the consistent
trend of shifting funding away from
younger researchers has no end in sight
(Daniels, 2015).
Ensuring a Durable Future for
Academic Science Post-COVID19
Recovery from the immediate COVID19
crisis necessitates a multi-pronged
approach including fiscal and non-fiscal
strategies to help graduate students,
postdoctoral fellows, and early and later
career faculty. This pandemic has partic-
ularly impacted senior postdoctoral fel-
lows seeking academic faculty positions
and early career faculty seeking to estab-
lish themselves as independent investi-
gators. Special consideration for these
early career researchers is key to over-
coming the crisis and strengthening the
foundations of academic science. Our
action plan proposed below is not an
exhaustive list of all possible recommen-
dations for supporting scientists, nor is it
inclusive of every academic scientist’s
specific circumstance. Not all of our sug-
gestions are applicable at every univer-
sity or institution, as each will have its
own unique set of challenges. We
acknowledge that monetary support will
be limited due to the deteriorating eco-
nomic situation and drastic loss of reve-
nue from clinical operations for most
medical campuses. While the immediate
goal of the recommendations is to pro-
vide support for scientists from funding
agencies, universities, departments, and
the public following COVID19, this sup-
port also provides solutions to the three
major challenges. Solutions to these sys-
temic issues (i.e., Excess Does Not Equal
Excellence, Diversification Leads to Dis-
covery, or Funding Agencies) are inter-
woven across the structure of academic
science, allowing us to comprehensively
tackle these issues at all levels. Plans
for recovery from the COVID19
pandemic must ensure as much continu-
ity as possible in research while
improving upon existing infrastructures
in order to provide a more inclusive,
cohesive, and efficient future for the
1446 Cell 181, June 25, 2020
next generation of independent sci-
entists.
Funding Agencies
Grantsmanship
The resiliency of research is dependent
upon the support of funding agencies.
Like the broader scientific community,
funding agencies will need to adapt their
strategies and structure to fit the chang-
ing times. Simplification of grant applica-
tion processes, including fewer supple-
mental documentations and more
implementation of letter-of-intent formats
prior to full proposals, could increase effi-
ciency for both the funding agency and
researcher. Lab closures will undoubtedly
create a void in the preliminary data that
are necessary to obtain most awards.
Early career researchers who had less
time to acquire these data prior to lab
shutdowns will be the most affected.
Funding agencies could introduce pol-
icies and programs targeted at early in-
vestigators that require fewer preliminary
data (similar to the National Institute of
Mental Health [NIMH] Brain Research
through Advancing Innovative Neurotech-
nologies [BRAIN] Initiative R01 or the
DP2), reducing the excess in data
required for most grants. Grants submit-
ted by graduate students, postdoctoral
fellows, and early career faculty who do
not have sufficient preliminary data per
current standards should be given special
consideration. Currently, many of the new
funding opportunities by funding
agencies, such as the NIH, are geared to-
ward supplements to existing grants or
COVID-related research. As there will
likely be restrictions or reductions to new
funding opportunities in the coming years
due to fiscal shortages, faculty with exist-
ing grants might help early career faculty
by including them in their supplemental
applications. Including early career fac-
ulty will also foster collaboration and
resource sharing, both of which will be vi-
tal during this time (Excess Does Not
Equal Excellence and Rethink the Funda-
mentals of Funding).
Extension of Deadlines, Timelines,
and Funding
Numerous funding agencies have already
implemented deadline extensions, but
deadlines must be further extended for
the duration of lab disruptions. It is also
imperative that funding agencies extend
Commentary
early investigator status for grant applica-
tions and implement no-cost extensions
for currently held grants. Additional bridge
funding programs may be especially
important for faculty who are between
projects or aiming to switch areas of study
following the COVID19 crisis.
Universities
Extensions for Tenure: Faculty
Most universities have added one-year
extensions to the tenure tracks of early
career researchers, but sliding extensions
may better support the success of vulner-
able academics. Many early career inves-
tigators may request extensions during
lab closures, but they should also have
the ability to go up for tenure early if the
opportunity arises. Ensuring the promo-
tion and advancement of marginalized
groups such as women, who make up <
30% of STEM faculty, is even more imper-
ative post-COVID19. COVID19-initiated
resetting of expectations for the publish-
ing, teaching, mentorship, and service re-
quirements for tenure may not only help
minimize the excesses innate to the cur-
rent tenure structure, but also may help
foster environments that can acknowl-
edge implicit biases and keep marginal-
ized groups from disproportionately
leaving STEM fields. Tenure expectations
for the next generation of early career re-
searchers may need to account for
increased variability between faculty that
is exacerbated by the COVID19 crisis
and allow for more flexibility in the pro-
cess. This crisis has amplified how the
antiquated one-size-fits-all guidelines
only encourage the disenfranchisement
of women and racial or ethnic minorities
(Diversification Leads to Discovery and
Excess Does Not Equal Excellence).
Trainees
The current crisis will have a dramatic
trickle-down effect, and numerous hiring
freezes are already in place. Mechanisms
to allow postdoctoral fellows or graduate
students in their final year to continue in
their current positions should be enacted,
if necessary, and if labs or universities are
able to provide fiscal support. Current clo-
sures are also disrupting the ability of
many graduate students to complete their
rotations. Universities could extend the
timeline for rotations and potentially cover
graduate students’ stipends. Trainees,
particularly postdoctoral fellows, may

Commentary
have limited ability to extend their period
of training due to visa restrictions. Univer-
sities should coordinate with federal
agencies to pursue strategies aimed at
extending visa expiration timelines, allow-
ing trainees to complete work that was
delayed due to the COVID19 crisis. These
mechanisms are needed to assure that
we do not lose an entire generation of sci-
entists following the coronavirus crisis.
Curtailment of Applicable Hiring
Freezes
Many universities have implemented hiring
freezes for faculty and staff for the
remainder of the year or beyond. Univer-
sities should not limit the ability of early
career faculty to hire postdoctoral fellows
and staff, however. Restricting early career
faculty from hiring technical assistance and
lab managers will stymie their ability to
generate preliminary data, which will
consequently limit grant and paper sub-
missions and delay career advancement.
Even a short hiring freeze could have
devastating effects on the ability of early
career faculty labs to succeed. Allowing
early career faculty to continue hiring will
also help to ease the bottleneck of grad-
uate students looking for postdoctoral or
research scientist positions within the
next few years. Hiring freezes at any level
will disproportionately affect early career
individuals and oversaturate the market
with qualified candidates. Permitting
ongoing interviews for faculty positions,
even if the official hire date is postponed,
could alleviate stress on the postdoc popu-
lation and expedite the hiring process when
hiring freezes are lifted. The faculty search
process serves as a valuable feedback
mechanism for postdoctoral fellows that
sometimes has an impact on career path.
Halting all hiring and all faculty searches
may drive talented postdocs, especially
women and members of ethnic or racial mi-
norities, out of academia (Diversification
Leads to Discovery).
Institutional Funds and Startup
Packages
Although universities may curtail
spending from institutional funds, special
consideration should be given to new
and early career faculty. Early career fac-
ulty must retain access to their startup
packages during this time. Institutional
funds should be released for salary sup-
port for early career faculty and for all
staff, students, and trainees in their labs.
If startup funds are set to expire, the expi-
ration date should be extended. New fac-
ulty should be given the funds needed to
establish their labs once research activ-
ities resume (Rethink the Fundamentals
of Funding).
Supplementation
The economic toll caused by shelter-in-
place will undoubtedly be significant,
including the reduction in funding through
endowments and charitable giving. We
fully acknowledge that monetary supple-
mentation may be difficult for universities
following the COVID19 crisis. Any combi-
nation of fiscal supplementation with
other mechanisms of non-fiscal support
should be considered. Universities might
implement new or expanded fellowships
for postdocs and graduate students, add
to existing startup packages for faculty,
assist with the purchasing of equipment
or expand shared equipment funding, or
create subsidies or joint ventures with
federal programs similar to unemploy-
ment or re-deployment programs. Univer-
sities might supplement pay or provide
reimbursement for staff, postdoc, and
graduate student salaries during the dura-
tion of academic closures.
Supplementation: Per Diem Costs
Many universities have per diem policies
that differ based on funding source, with
reduced per diem costs associated with
federal grants. Early career faculty without
federal funding have per diem costs double
that of other labs. Universities could imple-
ment mechanisms to reduce or
supplement animal costs that will be
accrued during lab closures and when
labs reopen and expand their animal
colonies (Rethink the Fundamentals of
Funding).
Supplementation: Childcare
Initiatives
Onsite daycare facilities support postdoc-
toral fellows and faculty with young chil-
dren. These family care centers are critical
to narrowing the gap and slow the attrition
of women and parents in science. Univer-
sities could work with early childhood ed-
ucation programs to establish or expand
daycare and preschool programs,
providing free or subsidized childcare for
faculty and teaching opportunities for
early education majors. Universities might
also reach out to current or retired teach-
ers seeking supplemental income (Diver-
sification Leads to Discovery).
ll
Supplementation: Access to
Technology
Universities should encourage and enable
graduate students and postdocs to use
this time to learn new computational skills
in anticipation of reductions in ability to do
work at the bench. Many university-
offered computational courses were
over-committed during lab closures due
to a significant increase in enrollment re-
quests. Universities should make a
concerted effort to increase bandwidth
and capacity for computational courses.
Many free online resources are also avail-
able to supplement the acquisition of cod-
ing skills.
Departments: Administrative and
Teaching Load
Administrative and teaching expecta-
tions should be reevaluated during uni-
versity closures. Departments should re-
assess administrative and teaching
loads, especially for early career faculty
whose promotions are contingent upon
teaching requirements. This is especially
important, since female scientists gener-
ally have increased teaching loads and
more advisory expectations than male
scientists (Gibney, 2017), which could
disproportionately delay scientific recov-
ery of female scientists from COVID19
closures (Diversification Leads to Dis-
covery).
Mentorship
Mental Health
The COVID19 crisis and subsequent lab
closures will take an incredible toll on
mental health. Early career faculty who
have yet to establish themselves or their
research independently and postdocs
whose future job prospects are now
significantly limited will be especially
impacted by prolonged lab shutdowns.
Department chairs, division leaders,
and mentors should do their best to
check in with early career faculty and
postdocs during this time. Mentoring
will be key both during and after this
crisis. Establishing scheduled virtual
meetings during social distancing and
in-person meetings after labs are reop-
ened could help alleviate some mental
stress. University mental health re-
sources are also available for anyone
who needs support. As students gener-
ally contact female faculty about mental
health issues more frequently than male
Cell 181, June 25, 2020 1447
 ll
Commentary

and curtailment of patient enrollment in

clinical trials, they will also have the extra
physical and mental stressors of working
in the hospital during a crisis. Establish-
ment of protocols to aid clinician-scien-
tists is imperative to ensuring their impor-
tant contributions to science. Just as
senior faculty mentoring will be critical
for junior faculty and graduating postdocs
to successfully transition to a post-COVID
era in the basic sciences, this type of
mentorship protocol may be even more
critical for clinician-scientists, many of
whom do not have doctorates beyond
the medical degree.

Public Initiatives
Make Science a National Priority
The current crisis has brought the impor-
tance of science and research to the
forefront of public life. Not only is science
critical for public health decision-mak-
ing, but a sustained investment in
research better positions political
leaders to efficiently deploy testing and
therapeutic solutions. Capitalizing this
momentum is crucial to engaging the
public in science and science funding.
Providing additional funding sources
Figure 1. The COVID19 crisis has magnified the systemic issues plaguing academic focused on conveying science to the
research. These include the often stifling excess requirements in publication, tenure, and
greater public and stimulating interest in
grant processes; the reliance on funding from national agencies that is catered towards
senior level researchers; and the lack of diversity in academic research due to the attrition of science through educational outreach is
women and racial or ethnic minorities during early career stages. critical. Exploiting technology and social
media to bring science and research
directly to the public will be vital in the
faculty (Bennett, 1982), equal encour- Faculty Mentorship Programs post-COVID19 world. Such technology
agement of mentorship from all faculty Once labs are reopened, pairing early might include mechanisms to allow pri-
is essential to not overburdening women career faculty with a later career faculty vate citizens to directly invest in science
faculty during this time (Diversification mentor of an established lab could facili- and scientists (Else, 2019; Miller, 2019),
Leads to Discovery). tate more effective research programs including simplified website-based
Graduate Student Programs and allow for resource sharing. Later donation platforms or inclusion on elec-
Mentoring graduate students throughout career faculty could be incentivized to tion ballots. This is necessary for estab-
lab closures and after reopening should help early career researchers through re- lishing new funding sources for scien-
be strongly encouraged. Those con- ductions in teaching or administrative tists, potentially supplementing the
ducting experiments will be most affected loads, supplementations to animal care dearth of funding for early career re-
by lab closures, and this should be explic- costs, core facility usages, or other means searchers at federal funding agencies
itly acknowledged by faculty and mentors. of reimbursement and/or subsidizations. (Rethink the Fundamentals of Funding).
Universities must assure graduate stu- Investment of later career faculty in the Enhanced Scientific Transparency
dents that graduate programs will be stabi- success of early career faculty will help The COVID19 crisis has revealed a lack of
lized and that admittance will not be to ensure stability and success in the public understanding about how science
decreased. For many faculty, graduate younger generation of independent re- is funded, conducted, and reported. The
students are the major workforce of the searchers. current administration’s belief that the
lab. To ensure that faculty can successfully Clinician-Scientists NIH is ‘‘giving away $32 billion a year’’
build and sustain a lab, continued ability to Faculty who have clinical responsibilities should be cause for concern (DeYoung
attract graduate students is necessary. also necessitate special consideration et al., 2020). Much of the mistrust evident
This is especially important for new inves- during this time, especially if they are on between the scientific establishment and
tigators, as getting postdoctoral fellows the front lines. These individuals will not the general population is rooted in lack
can be more challenging for newer faculty. only lose productivity due to lab closures of transparency and community

1448 Cell 181, June 25, 2020


Commentary
involvement in science. Taking scientists
out of the ‘‘ivory tower’’ and increasing
accessibility through technology may
help to assuage the mistrust that hinders
our preparedness in times of crisis. Peo-
ple cannot support what they do not un-
derstand. Removing excess requirements
in publishing, grantsmanship, and tenure
expectations could have the added
benefit of creating more time for scientists
to interact in the public domain. Scientists
must work on building the trust that is
imperative to success as a community,
and early career scientists are primed to
help pave this new future (Excess Does
Not Equal Excellence).
Conclusions
Beyond the immediate challenges of re-
turning to laboratories and research ca-
reers, the COVID19 crisis has exposed
some of the underlying weaknesses and
problems that permeate the current sci-
entific enterprise (Figure 1). For example,
editors are asking reviewers to not
request more experiments unless abso-
lutely necessary to validate the core
claims of a manuscript during the review
process. Most are applauding this effort
to minimize excess and calling for its
continued implementation even after sci-
entists are able to get back to the bench.
All institutions, funding agencies, depart-
ments, and members of the scientific
community should speak openly and
honestly about the difficulties faced dur-
ing the current situation. Early career
researchers should be involved in the de-
cision-making processes, as they repre-
sent the future of science and academic
leadership. The COVID19 crisis has pro-
vided us with the unique opportunity to
reflect upon the present norms and enact
change through fiscal and non-fiscal
strategies. Our hope is that this
pandemic will allow us to chart a new
course for science, both academically
and socially, and to begin to address
the core challenges of research, with a
special focus on supporting the next
generation of independent scientists.
DECLARATION OF INTERESTS
Dr. Roberts serves as Editor-in-Chief of books for
the American Psychiatric Association Publishing
Division and as Editor-in-Chief of the journal Aca-
demic Medicine. Unrelated to this publication, Dr.
Roberts serves as an advisor for the Bucksbaum
Institute of the University of Chicago Pritzker
School of Medicine and owns the small business
Terra Nova Learning Systems.
REFERENCES
Bennett, S.K. (1982). Student perceptions of and
expectations for male and female instructors: Evi-
dence relating to the question of gender bias in
teaching evaluation. J. Educ. Psychol. 74,
170–179.
ll
Daniels, R.J. (2015). A generation at risk: young in-
vestigators and the future of the biomedical work-
force. Proc. Natl. Acad. Sci. USA 112, 313–318.
DeYoung, K., Sun, L.H., and Rauhala, E. (2020).
Americans at World Health Organization trans-
mitted real-time information about coronavirus to
Trump administration. The Washington Post, Avail-
able from. https://www.washingtonpost.com/
world/national-security/americans-at-world-health-
organization-transmitted-real-time-information-
about-coronavirus-to-trump-administration/2020/
04/19/951c77fa-818c-11ea-9040-68981f488eed_
story.html.
Else, H. (2019). Crowdfunding research flips sci-
ence’s traditional reward model. Nature, Available
from. https://www.nature.com/articles/d41586-
019-00104-1.
Gibney, E. (2017). Teaching load could put
female scientists at career disadvantage. Na-
ture, Available from. https://www.nature.com/
news/teaching-load-could-put-female-scientists-at-
career-disadvantage-1.21839.
Malisch, J.L., Harris, B.N., Sherrer, S.M., Lewis,
K.A., Shepherd, S.L., McCarthy, P.C., Spott, J.L.,
Karam, E.L., Moustaid-Moussa, N., McCrory Cal-
arco, et al. (2020). In the wake of COVID-19,
academia needs new solutions to ensure gender eq-
uity. Proceedings of the National Academy of Sci-
ence. https://doi.org/10.1073/pnas.2010636117.
Miller, Z. (2019). The best platforms for crowd-
funding science research. The Balance: Small
Business, Available from. https://www.
thebalancesmb.com/top-sites-for-crowdfunding-
scientific-research-985238.
Monroe, K., Ozyurt, S., Wrigley, T., and Alexander,
A. (2008). Gender Equality in Academia: Bad News
from the Trenches, and Some Possible Solutions.
Perspectives on Politics 6, 215–233.
Cell 181, June 25, 2020 1449
 ll
Leading Edge

Previews
Cap-Snatching Leads to Novel Viral Proteins
Alistair B. Russell1,*
1Division of Biology, University of California, San Diego, San Diego, CA, USA

*Correspondence: a5russell@ucsd.edu
https://doi.org/10.1016/j.cell.2020.05.044

Some negative-sense RNA viruses prime mRNA transcription using host 50 cap sequences, usurping host
translational machinery and evading antiviral surveillance. In this issue of Cell, Ho et al. identify an additional
consequence of this viral strategy: the acquisition of upstream start codons from host-derived sequences
and subsequent translation of novel viral products.

Canonical eukaryotic mRNAs possess a
50 methylguanosine cap and a 30 polyade-
nosine tail. These features, combined, re-
cruit cellular translational machinery and
mark these molecules as ‘‘belonging’’ in
the cytoplasmic compartment wherein
they are accessible to ribosomes. Influ-
enza A virus (IAV), lacking its own capping
machinery, produces capped mRNAs
through a process known as cap-snatch-
ing. The viral polymerase associates with
an actively transcribing RNA polymerase
II complex and cleaves the nascent host
mRNA, thereafter using the cleaved prod-
uct to prime viral mRNA production. This
process appears to roughly target caps
according to their relative abundance in
the nuclear compartment, resulting in
diverse cap sequences associated with
a given viral mRNA (Walker and Fo-
dor, 2019).
Previously, little to no consequence has
been ascribed to the composition of host-
derived sequences appended to the 50
end of viral mRNAs, and the heterogeneity
of such sequences is certainly consistent
with a lack of function in viral transcription
and translation. Ho et al. (2020) posited
that AUG codons snatched in such a
manner might permit the recruitment of ri-
bosomes upstream of the canonical IAV
start codons—producing either N-termi-
nal extensions to known viral proteins or
novel frameshifted peptides. The authors
found that AUG codons could be readily
identified in host-derived 50 sequences in
two different IAV strains in two different
cell types, indicating that there appeared
to be no specific mechanism precluding
their acquisition by viral transcriptional
machinery. Furthermore, Ho et al. (2020)
were able to detect initiating ribosomes
associated with upstream, host-derived
start codons, and, for a subset of these
proteins, they could detect expression
by mass spectrometry. Combined, these
represent relatively unequivocal evidence
that host-derived alternative start codon
acquisition can drive protein production
in IAV.
What is the consequence of translation
from host-derived start codons? Such
proteins could play a role in viral infection
via at least two, non-exclusive mecha-
nisms. (1) They could represent bone-
fide, functional, overprinted products
that serve significant roles in the viral life
cycle, and (2) they could provide targets
for adaptive immunity (Figure 1). With
respect to the first, the genomes of RNA
viruses are restricted in length and over-
printing, encoding multiple proteins from
the same overlapping sequence, is a
common mechanism by which additional
coding capacity is procured. Ribosomal
frameshifting, encoded alternative start
sites, and alternative splicing are, among
other strategies, utilized by IAV to
generate alternative protein products
(Chen et al., 2001; Dubois et al., 2014;
Jagger et al., 2012). Ho et al. (2020) found
broad conservation between IAV strains
of an overprinted protein that would be
generated from upstream host-derived
start codons, consistent with a functional
hypothesis. However, other evolutionary
constraints such as the sequence of the
‘‘primary’’ influenza protein, genome
packaging constraints, and viral promoter
sequences would also produce signa-
tures of conservation in these genomic re-
gions, a point also noted by the authors.
Furthering a functional argument, Ho
et al. (2020) found evidence that disrupt-
ing the expression of one of these over-
printed proteins or an N-terminal exten-
sion to a viral protein driven by an
upstream host-derived start codon
altered in vivo pathogenesis. Caution
should still be taken when trying to link
these virulence data to a true biological
function; infection outcome for the host
is ancillary to viral fitness, and, even in a
model that perfectly recapitulates the
host response, the route and dosage of
delivery can produce an incredible
breadth of outcomes that may have little
resemblance to the normal course of dis-
ease. Therefore, without a specific pro-
posed mechanism, these data do not yet
support inferring a function for these novel
proteins in IAV but also do not preclude
such a role.
In addition to potential novel functional
proteins, this discovery is incredibly
important in light of the second possibil-
ity—that these upstream start codons
may generate targets for immune surveil-
lance. Any protein, regardless of function
or lack thereof, once degraded can
generate potential peptides for display
via MHC-I and mark a cell for subsequent
destruction and recruitment of a more
robust immune response. One could
even posit that, unlike full-length, canoni-
cal viral proteins, out-of-frame, or even
N-terminal extended, proteins may lack
functional selection driving stability.
Without such selection, these proteins
may be rapidly targeted to the protea-
some, and thus preferentially presented
via MHC-I, and represent a heretofore un-
recognized target for antiviral immunity. It
has been increasingly recognized that
‘‘off-products’’ of viral replication, be
they aberrant peptides or even aberrant
genomes, are frequently the trigger for
an immune response (López, 2014; Wei
and Yewdell, 2019). This is perhaps an

1450 Cell 181, June 25, 2020 ª 2020 Elsevier Inc.


Previews
Figure 1. Biological Impacts of Cap-Snatching
IAV engages in cap-snatching, producing chimeric host-viral mRNAs with a 50 methylguanosine cap (host sequence, blue; viral, black). Previously identified
implications of this strategy (left) include global host transcriptional suppression (cap-stolen from actively transcribing RNA polymerase II, shown in blue),
recruitment of ribosomes to viral messages (ribosome, brown), and evasion of antiviral surveillance by RIG-I (red)-like receptors (Decroly et al., 2011). New
implications from this study identifying translation from host-derived start codons (left, N-terminal extension and frameshifted proteins, red) include potential
novel viral functions and additional targets for host surveillance (T cell engaging in MHC-I detection shown).
unavoidable feature of the rapid, expo-
nential growth of viral populations wherein
fidelity of replication may be at odds with
replicative speed (Fitzsimmons et al.,
2018). Ho et al. (2020) found predicted
MHC-I epitopes within an overprinted
peptide, that these epitopes varied be-
tween IAV strains consistent with poten-
tial immune pressure, and that an out-of-
frame canonical MHC-I peptide could be
produced and displayed—thus formally,
MHC-I targets can be generated via
acquisition of host-derived start codons.
A critical next step is establishing whether
peptides derived from native proteins
generated by upstream host-derived start
codons are displayed via MHC-I, and if
so, what role they may play in viral surveil-
lance and clearance. Such work becomes
even more important in light of recent at-
tempts to generate T cell responses
against IAV, which, although they may
not prevent infection, are still thought
have the potential to reduce morbidity
and mortality (La Gruta and Turner, 2014).
Cap-snatching is not unique to IAV. To
that end, Ho et al. (2020) explored 50 cap
sequences from influenza B virus and
lassa virus and found host-acquired
AUG codons in both viral species. Using
a plasmid-based system that recapitu-
lates the replicon of Heartland banyangvi-
rus, a Bunyavirus, they further confirmed
translation of a luciferase lacking a start
codon, consistent with acquisition of
host-derived start codons in frame with
their reporter. Therefore, the findings
with IAV have widespread relevance
across many negative-sense RNA vi-
ruses. What we make of this moving for-
ward will take a significant amount of
work, but such work will have been
made possible by this initial observation.
We must uncover whether these alterna-
tive products serve functional roles,
whether they represent significant con-
tributors to MHC recognition of viral infec-
tion, and, not touched upon in this study
but important nevertheless, whether se-
lection against particular N-terminal ex-
tensions or immunogenic peptides con-
strains the evolutionary trajectories of
viral species that rely on cap-snatching.
Ho et al. (2020) have provided us with
the critical step, identifying that such fea-
tures can impact viral biology, and it will
be exciting to explore the implications of
this study.
ACKNOWLEDGMENTS
Work from A.B.R. in this field is supported in part
by the Damon Runyon Cancer Research Founda-
tion (DFS-36-19) and NIAID (1K22AI141678).
REFERENCES
Chen, W., Calvo, P.A., Malide, D., Gibbs, J., Schu-
bert, U., Bacik, I., Basta, S., O’Neill, R., Schickli, J.,
Palese, P., et al. (2001). A novel influenza A virus
mitochondrial protein that induces cell death.
Nat. Med. 7, 1306–1312.
ll
Decroly, E., Ferron, F., Lescar, J., and Canard, B.
(2011). Conventional and unconventional mecha-
nisms for capping viral mRNA. Nat. Rev. Microbiol.
10, 51–65.
Dubois, J., Terrier, O., and Rosa-Calatrava, M.
(2014). Influenza viruses and mRNA splicing: doing
more with less. MBio 5, e00070–14.
Fitzsimmons, W.J., Woods, R.J., McCrone, J.T.,
Woodman, A., Arnold, J.J., Yennawar, M., Evans,
R., Cameron, C.E., and Lauring, A.S. (2018). A
speed-fidelity trade-off determines the mutation
rate and virulence of an RNA virus. PLoS Biol. 16,
e2006459.
Ho, J.S.Y., Angel, M., Ma, Y., Sloan, E., Wang, G.,
Martinez-Romero, C., Alenquer, M., Roudko, V.,
Chung, L., Zheng, S., et al. (2020). Hybrid
gene origination creates human-virus chimeric
proteins during infection. Cell 181, this issue,
1502–1517.
Jagger, B.W., Wise, H.M., Kash, J.C., Walters, K.-
A., Wills, N.M., Xiao, Y.-L., Dunfee, R.L., Schwartz-
man, L.M., Ozinsky, A., Bell, G.L., et al. (2012). An
overlapping protein-coding region in influenza A
virus segment 3 modulates the host response.
Science 337, 199–204.
La Gruta, N.L., and Turner, S.J. (2014). T cell medi-
ated immunity to influenza: mechanisms of viral
control. Trends Immunol. 35, 396–402.
López, C.B. (2014). Defective viral genomes: crit-
ical danger signals of viral infections. J. Virol. 88,
8720–8723.
Walker, A.P., and Fodor, E. (2019). Interplay be-
tween Influenza Virus and the Host RNA Polymer-
ase II Transcriptional Machinery. Trends Microbiol.
27, 398–407.
Wei, J., and Yewdell, J.W. (2019). Flu DRiPs in
MHC Class I Immunosurveillance. Virol. Sin. 34,
162–167.
Cell 181, June 25, 2020 1451
 ll
Previews

Role of Microbiota-Derived Bile

Acids in Enteric Infections
Casey M. Theriot1 and William A. Petri, Jr.2,*
1North Carolina State University, Raleigh, NC, USA
2Department of Medicine, University of Virginia, PO Box 801340, Charlottesville, VA 22908-1340, USA
*Correspondence: wap3g@virginia.edu
https://doi.org/10.1016/j.cell.2020.05.033

In this issue of Cell, Alavi et al. report that infection by Vibrio cholerae is blocked by gut microbiome-mediated
hydrolysis of bile acids. Cholera therefore joins amebic dysentery and Clostridioides difficile colitis as enteric
infections profoundly influenced by the microbiome’s impact on bile acid metabolism.

Vibrio cholerae causes watery diarrhea so
severe that it kills by dehydration within
hours. We are now experiencing the
7th pandemic of cholera, all 7 of which
likely originated in the Indian subconti-
nent, with current estimates of up to 3
million cases and 100,000 deaths annu-
ally. That cholera is water-borne was es-
tablished by the physician John Snow in
1854 by linking victims of the London
Broad Street cholera epidemic not to
bad air but to the Broad Street water
pump. V. cholerae exists in aquatic envi-
ronments on the surface and in the intes-
tine of copepods (a type of small crusta-
cean). This leads to sporadic outbreaks
near rivers in the Indian subcontinent,
amplified by human fecal-oral spread of
the bacteria during outbreaks causing
pandemics.
Diarrhea is caused by cholera toxin, an
enzyme that ADP-ribosylates the Gs pro-
tein that regulates adenylate cyclase,
leading to a cyclic AMP (cAMP)-mediated
chloride ion (Cl ) secretion. Cholera toxin
and the toxin coregulated pili (TCP) are
regulated by TcpP, a membrane bound
transcriptional activator. There is sub-
stantial person-to-person variation in the
severity of cholera, with one explanation
being personal differences in the micro-
biome regulating virulence gene expres-
sion by TcpP. However, we still lack a
complete understanding of the factors
that cause person-to-person variation on
severity of cholera.
In this issue of Cell, Ansel Hsiao and
colleagues demonstrate that hydrolysis
of the bile acid taurocholate to cholate
by the gut microbiome blocks TcpP acti-
vation and cholera colonization. This pa-
per adds to emerging literature on how
the microbiome impacts infectious dis-
eases via microbial metabolism of bile
acids in the gut.
The primary bile acids cholate (CA) and
chenodeoxycholate (CDCA) are made by
the liver, where they are conjugated with
a glycine or taurine before being secreted
into the duodenum. As they make their
way through the small intestine, 95% of
bile acids are absorbed in the terminal
ileum through the enterohepatic system,
a majority being conjugated bile acids
(Figure 1). Gut microbes that encode bac-
terial bile salt hydrolase bsh genes can de-
conjugate or cleave the glycine and taurine
from conjugated bile acids to yield decon-
jugated bile acids (e.g., taurocholate
[TCA]/taurine and CA). This is a critical
first step in microbial bile acid metabolism
that leads to all subsequent biotransfor-
mations. The deconjugated bile acids
that reach the large intestine are then
metabolized by members of the gut micro-
biota into secondary bile acids, including
deoxycholate (DCA). (Foley et al., 2019).
Here, Alavi et al. (2020) demonstrate
that the composition of the gut micro-
biome contributes to resistance to
cholera. By reconstituting germ-free
mice with defined communities of human
microbiome bacteria, they discovered
that the commensal bacterium Blautia
obeum mediates resistance. They show
that B. obeum mediates resistance in
mice through degrading TCA to CA and
that the abundance of the B. obeum
BSH enzyme correlated with resistance
in humans. In the absence of B. obeum-
dependent degradation, TCA induces
the TcpP virulence regulator to cause
disease. Cholera therefore joins a list of
enteropathogens whose ability to cause
disease is regulated by microbiome meta-
bolism of bile acids.
A second example is the parasite
Entamoeba histolytica, an ameba that in-
vades the intestine by eating the epithelial
lining in a process called trogocytosis.
Burgess et al. (2020) recently identified
the intestinal bacterium Clostridium scin-
dens as providing protection from amebic
colitis. Introduction of C. scindens into
the microbiome of a mouse altered the
bone marrow by inducing expansion
of granulocyte-monocyte progenitors
(GMPs). C. scindens-mediated protec-
tion from amebiasis could be transferred
with adoptive transfer of bone marrow
to a naive mouse and act via an
increased recruitment of polymorphonu-
clear neutrophils to the colon. Because
C. scindens can dehydroxylate CA to
DCA, Burgess et al. (2020) tested if this
mediated alteration of the marrow. In
fact, administration of DCA alone pro-
vided complete protection from amebi-
asis via GMP expansion, demonstrating
microbiome-to-bone-marrow communi-
cation via bile acids.
A final example relates to Clostridioides
difficile infection (CDI). C. difficile is a
Gram-positive spore forming bacillus
and the most common cause of hospi-
tal-acquired, antibiotic-associated diar-
rhea. Ingestion of the spore form of
C. difficile in an individual with a dysbiotic
microbiome (usually due to prior antibiotic
therapy) leads to infection of the large in-
testine by the vegetative stage of
C. difficile. Primary and secondary bile
acids have been shown to impact
C. difficile vegetative growth as well as
spore germination and toxin activity. For
example, the primary bile acid TCA

1452 Cell 181, June 25, 2020 ª 2020 Elsevier Inc.

 ll
Previews

Figure 1. Role of Microbially-Derived Bile Acids in Enteric Infections

Role of hepatically synthesized primary and microbially derived secondary bile acids (green color) in enteric infections. Primary bile acids cholate (CA) and
chenodeoxycholate (CDCA) are synthesized from cholesterol by hepatocytes. Primary bile acids can be further modified via conjugation with taurine or glycine
within the liver (T(G)CA, T(G)CDCA). Once synthesized, primary bile acids enter into bile. Bile is stored in the gallbladder until released in the duodenum following
ingestion of a meal. Once within the GI tract, gut microbial enzymes convert host-derived conjugated primary bile acids into secondary bile acids. For example,
the primary bile acid taurocholate (TCA) is metabolized by the bile salt hydrolase (bsh) of Blautia obeum into CA. CA in turn is metabolized by Clostridium scindens
7a-dehydoxylation via the bai operon into deoxycholate (DCA). The primary bile acid TCA is essential for V. cholerae TcpP virulence activation and for germination
of C. difficile spores. Conversion of TCA to CA by B. obeum BSH protects against cholera and C. difficile. Conversion by C. scindens bai operon of the secondary
bile acid CA to DCA increases granulocyte monocyte progenitors (GMPs) in the bone marrow to provide gut polymorphonuclear neutrophil (PMN) protection from
the parasite Entamoeba histolytica.

induces spore germination, and DCA in-
hibits C. difficile growth.
Like the amebic colitis example,
C. scindens is implicated as having a pro-
tective role in CDI. First, depletion of
C. scindens is associated with more se-
vere disease in humans and mice. More-
over, in the mouse model of CDI, reconsti-
tution of C. scindens was able to partially
restore colonization resistance against
CDI in mice, and colonization resistance
was associated with secondary bile acid
synthesis (Buffie et al., 2015). In patients
with recurrent CDI, high levels of conju-
gated primary bile acids and reduced sec-
ondary bile acids were observed in feces
when compared to healthy individuals.
Successful treatment of recurrent CDI
with fecal microbial transplant restored
the level of fecal secondary bile acids,
specifically, DCA and LCA (Weingarden
et al., 2015; Seekatz et al., 2018). Most
recently, Reed et al. (2020) have shown
that several commensal Clostridia encod-
ing the bai operon (which encodes en-
zymes that convert cholate into the sec-
ondary bile acid deoxycholate), but not
all, are able to inhibit C. difficile growth
due to the conversion of CA to DCA,
which should provide protection. To sum-
marize, there is strong in vitro and sup-
portive in vivo evidence that microbial
metabolism of bile acids has a direct
impact on C. difficile spore germination,
growth, and toxin production and activity.
Bile acids also regulate the immune
system, as demonstrated in amebic dys-
entery with DCA protecting by increasing
marrow GMPs. Additional examples of a
direct impact of bile acids on the immune
system include the secondary bile acid
LCA regulating Th17 responses by inter-
fering with RORyT transcriptional activity
(Hang et al., 2019) and, in the context of
colitis, bile acids interacting with macro-
phages to induce IL-10, which polarizes
T cells to a regulatory phenotype (Biagioli
et al., 2017).
In summary, Alavi et al. (2020) have
deepened our understanding of how
composition of the gut microbiome pro-
tects from cholera by the discovery of
the role of bacterially encoded bile salt hy-
drolases. As microbiome science moves
from description to mechanism, bile acids
synthesized by the host and further con-
verted by the microbiota are becoming
central players, both for their direct
impact on enteropathogens and for their
impact on the immune system.
ACKNOWLEDGMENTS
Work from the authors’ labs is supported by
National Institutes of Health grants R35
GM119438 (to C.M.T.), 2R37 AI026649-31,
and R01 AI043596-22 (to W.A.P.) and the Henske
and McGrath families.
REFERENCES
Alavi, S., Mitchell, J.D., Cho, J.Y., Liu, R., MacBeth,
J.C., and Hsiao, A. (2020). Interpersonal gut
microbiome variation drives susceptibility and
resistance to Vibrio cholerae. Cell 181, this issue,
1533–1546.
Biagioli, M., Carino, A., Cipriani, S., Francisci, D.,
Marchianò, S., Scarpelli, P., Sorcini, D., Zampella,
A., and Fiorucci, S. (2017). The bile acid receptor
GPBAR1 regulates the M1/M2 phenotype of intes-
tinal macrophages and activation of GPBAR1 res-
cues mice from murine colitis. J. Immunol. 199,
718–733.
Buffie, C.G., Bucci, V., Stein, R.R., McKenney,
P.T., Ling, L., Gobourne, A., No, D., Liu, H., Kinne-
brew, M., Viale, A., et al. (2015). Precision micro-
biome reconstitution restores bile acid mediated

Cell 181, June 25, 2020 1453

 ll
resistance to Clostridium difficile. Nature 517,
205–208.
Burgess, S.L., Leslie, J.L., Uddin, M.J., Oakland,
D.N., Gilchrist, C.A., Moreau, G.B., Watanabe, K.,
Saleh, M.M., Simpson, M., Thompson, B.A., et al.
(2020). Gut microbiome communication with
bone marrow regulates susceptibility to amebiasis.
J. Clin. Invest. https://doi.org/10.1172/JCI133605.
Foley, M.H., O’Flaherty, S., Barrangou, R., and
Theriot, C.M. (2019). Bile salt hydrolases: Gate-
keepers of bile acid metabolism and host-micro-
biome crosstalk in the gastrointestinal tract. PLoS
Pathog. 15, e1007581.
Mapping the Uncharted Territories of Human
Brain Malignancies
Daan Juri Kloosterman1 and Leila Akkari1,*
1Division of Tumor Biology and Immunology, Netherlands Cancer Institute, Oncode Institute, Amsterdam 1066CX, the Netherlands
*Correspondence: l.akkari@nki.nl
https://doi.org/10.1016/j.cell.2020.06.003
Despite its success in multiple tumor types, immunotherapy remains poorly efficacious in brain malignancies.
In this issue of Cell, Friebel et al. and Klemm et al. provide in-depth insights into the versatile nuances of
immune cells in primary and metastatic brain tumors, granting the field with a rich framework to explore novel
therapeutic avenues.
The development of therapeutic strate-
gies enlisting cells composing the tumor
microenvironment (TME) has revolution-
ized clinical approaches in recent years,
exemplified by T cell immunotherapy in
cancer treatment (Waldman et al., 2020).
Yet, in organs as unique as the brain
with its blood-brain barrier (BBB)-pro-
tected niche, the current understanding
of how tumors intrinsically emerging in
the central nervous system or originating
from extracranial sites sculpt the TME to
their advantage remains limited, thwarting
efficient immune cell harnessing in thera-
peutic intervention.
Glioblastoma (GBM) and brain metas-
tases (BrMs) bear some of the worst prog-
noses for cancer patients. Discoveries of
brain metastases drivers (Brastianos
et al., 2015), genomic alterations of GBM
subtypes (Brennan et al., 2013), and di-
versity at the single-cell level (Patel
et al., 2014; Tirosh et al., 2016) have
shed light into inter- and intra-tumoral
heterogeneity of brain lesions with a focus
1454 Cell 181, June 25, 2020 ª 2020 Elsevier Inc.
Hang, S., Paik, D., Yao, L., Kim, E., Trinath, J., Lu,
J., Ha, S., Nelson, B.N., Kelly, S.P., Wu, L., et al.
(2019). Bile acid metabolites control TH17 and
Treg cell differentiation. Nature 576, 143–148.
Reed, A.D., Nethery, M.A., Stewart, A., Barrangou,
R., and Theriot, C.M. (2020). Strain-dependent in-
hibition of Clostridioides difficile by commensal
Clostridia encoding the bile acid inducible (bai)
operon. J Bacteriol. https://doi.org/10.1128/JB.
00039-20.
Seekatz, A.M., Theriot, C.M., Rao, K., Chang, Y.M.,
Freeman, A.E., Kao, J.Y., and Young, V.B. (2018).
Restoration of short chain fatty acid and bile acid
on tumor cell features, while introducing
the notion that stromal composition can
be shaped according to different cancer
mutational statuses. Further evidence
elaborating on this concept have delin-
eated the particular phenotype of tumor-
associated macrophages (TAMs) in spe-
cific subsets of GBM (Wang et al., 2017),
with TAM enrichment generally associ-
ated with poor disease outcome (Kiel-
bassa et al., 2019). However, the extent
of myeloid cell diversity, including TAM
ontogeny and distinct education, has re-
mained largely unexplored in human brain
tumors.
In this issue of Cell, two studies (Friebel
et al., 2020; Klemm et al., 2020) per-
formed extensive and comprehensive an-
alyses of GBM and BrM immune land-
scapes from large cohorts of patients,
thus providing much needed information
on the phenotype and transcriptional pro-
grams acquired by components of the
TME, in a cell-type- and disease-specific
manner.
Previews
metabolism following fecal microbiota transplanta-
tion in patients with recurrent Clostridium difficile
infection. Anaerobe 53, 64–73.
Weingarden, A., González, A., Vázquez-Baeza, Y.,
Weiss, S., Humphry, G., Berg-Lyons, D., Knights,
D., Unno, T., Bobr, A., Kang, J., et al. (2015). Dy-
namic changes in short- and long-term bacterial
composition following fecal microbiota transplan-
tation for recurrent Clostridium difficile infection.
Microbiome 3, 10. https://doi.org/10.1186/
s40168-015-0070-0.
Through the use of multiparameter fluo-
rescence-activated cell sorting followed by
RNA sequencing (Klemm et al., 2020) and
CYTOF analyses (Friebel et al., 2020), the
authors assessed the abundance and het-
erogeneity of tissue-resident and peripher-
ally recruited leucocytes in BrMs and gli-
omas. Microglia (MG) dominated the TME
of IDHmut gliomas, considered to be less
aggressive brain lesions that displayed
limited infiltration of leucocytes. These re-
sults contrasted with the immunological
landscapes of IDHwt GBM and BrMs,
which were enriched by peripherally re-
cruited monocyte-derived macrophages
(MDMs). BrMs from multiple primary tumor
origins exhibited substantial infiltration of
T cells and neutrophils, with melanoma-
to-brain BrMs singled out in both studies,
harboring numerous CD4+ and CD8+
T cells subsets, and neutrophils heavily
infiltrating breast BrMs (Figure 1A).
The prominence of TAMs prompted the
authors to perform further analyses of MG
and MDM subset composition, spatial
 ll
Previews

Figure 1. Immune Landscape Composition and Features across Different Types of Brain Tumors
(A) Content of immune cells in the GBM and BrM TME. Tissue-resident microglia (MG) and monocyte-derived macrophages (MDMs) constitute the dominant cell
types composing the immune landscape of GBM in the depicted distinct abundances, according to the IDH mutational status of GBM. Recruited at the tumor site
from the peripheral circulation, leucocytes, including CD4+ T cells, CD8+ T cells, neutrophils, and monocytes, enter the brain parenchyma through the blood-brain
barrier and infiltrate brain tumors, a process enhanced in brain metastases, resulting in increased content of these recruited leucocytes in the BrM TME.
(B) Overview of the functional features acquired by each immune cell type in a disease-specific manner. Macrophages (MG and MDMs) are heterogenous across
tumor types and display the highest magnitude of plasticity, which ranges between ECM remodeling to immunomodulatory roles in IDHwt GBM and BrMs. Greatly
enriched in BrMs, CD4+ T cells and CD8+ T cells display features of hyporesponsiveness and exhaustion, whereas T cells in GBM are characterized by low
proliferative abilities and activation markers. Natural killer cells present an immature phenotype in IDHwt GBM in comparison to the cytotoxic features displayed in
IDHmut gliomas and BrMs.

Cell 181, June 25, 2020 1455

 ll
organization, and phenotype and tran-
scriptional programs in GBM and BrM in
order to establish their nexus role in regu-
lating the immune landscape in a disease-
specific manner. Using orthogonal ap-
proaches of RNA sequencing (Klemm
et al., 2020) and mass cytometry (Friebel
et al., 2020), both studies showed that
the plasticity of TAMs largely relied upon
the high magnitude of MDM phenotype
adaptation and transcriptional programs,
leading to limited commonalities between
TAM subsets across different diseases.
The prognostic value of myeloid cell
abundance was similarly relevant only
when examining MDM features in GBM.
Indeed, monocyte infiltration, which was
prominent in IDHmut GBM, only correlated
with a trend in increased GBM patient sur-
vival, whereas expression of the pan-
MDM marker CD163+ (Friebel et al.,
2020) or expression of the MDM-repre-
sentative gene set (Klemm et al., 2020)
both correlated with poor survival in low-
and high-grade GBM.
Strikingly, MDMs’ heterogeneity was
not a random feature across brain tumors
but distinct and dictated by each tumor
type. The origin of MDM diversity was
proposed to rely upon distinct differentia-
tion trajectories from their monocyte pro-
genitors, thus giving rise to multiple MDM
subsets carrying specific phenotypes
with different prognostic values. Single-
cell proteomic analysis was leveraged to
ascribe a positive survival outcome to
the CD163+MRC1+ MDM subset in low-
grade GBM (Friebel et al., 2020).
The findings that MG and MDMs shared
some, but limited, features within the
same disease setting suggest that the
type of tumor in addition to the affected
tissue heavily weighs onto the program-
ming of recruited immune cells. This is
further exemplified by the distinct reactive
phenotype acquired by MG in IDHmut
GBM compared to IDHwt tumors (Friebel
et al., 2020). Importantly, the transcrip-
tomic education acquired by either TAM
subsets did not reflect the defined M1-
like or M2-like macrophage polarization
phenotypes but included shades of either
activation states (Klemm et al., 2020).
A key distinctive feature between BrM
and GBM was revealed when assessing
the spatial organization and phenotype
of infiltrating leucocytes. TAMs and
T cells were closely interacting in the
1456 Cell 181, June 25, 2020
BrM TME but not in gliomas (Klemm
et al., 2020), with BrM CD4+ and CD8+
T cells exhibiting anergic and exhaustion
signatures, respectively. Flawed T cell
features can be explained by chronic acti-
vation through excessive antigen presen-
tation, including by TAMs, which exert
greater immunomodulatory functions in
BrM than in IDHwt GBM and express mul-
tiple co-inhibitory receptors (Figure 1B).
These major differences are in line with
the outcome of immune checkpoint
blockade, which showed promising effi-
cacy in melanoma BrMs, where T cells
are abundant and can be rewired, but little
to no response in T cell-excluded primary
brain tumors.
Primary cancer cells or their metastatic
counterparts face considerable chal-
lenges to strive within the unique brain
TME, which can only be overcome by
hijacking a selective niche, leading to a
paralleled evolution of stromal and im-
mune cells. The deep insights into the di-
versity of specialized immune landscapes
shaped by GBM and BrMs presented in
the studies discussed here highlight the
need for careful identification of the
ontogeny, cellular phenotype, and local
education of myeloid cell subsets, without
simply utilizing content to predict survival
outcomes or to devise therapeutic inter-
ventions. Holistic approaches using sin-
gle-cell RNA sequencing and spatial
determinants of the GBM and BrM TME
will be necessary to understand how
cellular plasticity gives rise to these
subtype-specific immunological niches.
Indeed, several questions remained to
be answered in light of these novel re-
sources: do brain tumors poise peripheral
immune cells for reprogramming, beyond
affecting their recruitment? How does the
genetic make-up of brain tumor cells
shape the education of each distinct im-
mune cell type? What underlies the dy-
namic acquisition of these phenotypes in
the course of tumor malignancy, and are
these altered by standard of care treat-
ment? The ability of TAMs to remodel
the extracellular matrix (ECM) landscape
in GBM and BrM additionally opens the
perspective to apply tumor-tissue archi-
tecture therapies to brain malignancies
and either disrupt the locally shaped niche
or facilitate drug penetration though the
BBB. A major challenge will then be to
design rational and optimized timing and
Previews
methodology of combination therapies
to account for the progressive changes
undergone by the TME (Rothschilds and
Wittrup, 2019). These may include neo-
adjuvant treatment approaches, which
have proven successful in recurrent
GBM (Cloughesy et al., 2019) and could
be guided by artificial-intelligence-based
radiogenomics to dynamically monitor
the brain TME (Rudie et al., 2019) and
stratify patients into tailored therapeutic
intervention.
The vivid immune cell heterogeneity
parallels their remarkably challenging tar-
geting to achieve clinical response in
brain tumors and emphasize the need to
further enrich our knowledge of the im-
mune contexture in a dynamic setting.
Collectively, these studies provide much
needed insights into the multifaceted nu-
ances of immune cells in primary and met-
astatic brain tumors and resources for the
brain tumor community to explore novel
therapeutic targets.
REFERENCES
Brastianos, P.K., Carter, S.L., Santagata, S., Cahill,
D.P., Taylor-Weiner, A., Jones, R.T., Van Allen,
E.M., Lawrence, M.S., Horowitz, P.M., Cibulskis,
K., et al. (2015). Genomic Characterization of Brain
Metastases Reveals Branched Evolution and
Potential Therapeutic Targets. Cancer Discov. 5,
1164–1177.
Brennan, C.W., Verhaak, R.G., McKenna, A., Cam-
pos, B., Noushmehr, H., Salama, S.R., Zheng, S.,
Chakravarty, D., Sanborn, J.Z., Berman, S.H.,
et al.; TCGA Research Network (2013). The so-
matic genomic landscape of glioblastoma. Cell
155, 462–477.
Cloughesy, T.F., Mochizuki, A.Y., Orpilla, J.R.,
Hugo, W., Lee, A.H., Davidson, T.B., Wang, A.C.,
Ellingson, B.M., Rytlewski, J.A., Sanders, C.M.,
et al. (2019). Neoadjuvant anti-PD-1 immuno-
therapy promotes a survival benefit with intratu-
moral and systemic immune responses in recur-
rent glioblastoma. Nat. Med. 25, 477–486.
Friebel, E., Kapolou, K., Unger, S., Núñez, N.G.,
Utz, S., Rushing, E.J., Regli, L., Weller, M., Greter,
M., Tugues, S., et al. (2020). Single-Cell Mapping of
Human Brain Cancer Reveals Tumor-Specific In-
struction of Tissue-Invading Leukocytes. Cell
181, this issue, 1626–1642.
Kielbassa, K., Vegna, S., Ramirez, C., and Akkari,
L. (2019). Understanding the Origin and Diversity
of Macrophages to Tailor Their Targeting in Solid
Cancers. Front. Immunol. 10, 2215.
Klemm, F., Maas, R.R., Bowman, R.L., Kornete,
M., Soukup, K., Nassiri, S., Brouland, J.P., Iacobu-
zio-Donahue, C.A., Brennan, C., Tabar, V., et al.
(2020). Interrogation of the microenvironmental
landscape in brain tumors reveals disease-specific

Previews
alterations of immune cells. Cell 181, this issue,
1643–1660.
Patel, A.P., Tirosh, I., Trombetta, J.J., Shalek, A.K.,
Gillespie, S.M., Wakimoto, H., Cahill, D.P., Nahed,
B.V., Curry, W.T., Martuza, R.L., et al. (2014). Sin-
gle-cell RNA-seq highlights intratumoral heteroge-
neity in primary glioblastoma. Science 344,
1396–1401.
Rothschilds, A.M., and Wittrup, K.D. (2019). What,
Why, Where, and When: Bringing Timing to Im-
muno-Oncology. Trends Immunol. 40, 12–21.
Rudie, J.D., Rauschecker, A.M., Bryan, R.N.,
Davatzikos, C., and Mohan, S. (2019). Emerging
Applications of Artificial Intelligence in Neuro-
Oncology. Radiology 290, 607–618.
Tirosh, I., Izar, B., Prakadan, S.M., Wadsworth,
M.H., 2nd, Treacy, D., Trombetta, J.J., Rotem, A.,
Rodman, C., Lian, C., Murphy, G., et al. (2016). Dis-
secting the multicellular ecosystem of metastatic
melanoma by single-cell RNA-seq. Science 352,
189–196.
ll
Waldman, A.D., Fritz, J.M., and Lenardo, M.J.
(2020). A guide to cancer immunotherapy: from
T cell basic science to clinical practice. Nat.
Rev. Immunol. https://doi.org/10.1038/s41577-
020-0306-5.
Wang, Q., Hu, B., Hu, X., Kim, H., Squatrito, M.,
Scarpace, L., deCarvalho, A.C., Lyu, S., Li, P., Li,
Y., et al. (2017). Tumor Evolution of Glioma-
Intrinsic Gene Expression Subtypes Associates
with Immunological Changes in the Microenviron-
ment. Cancer Cell 32, 42–56.6.
Cell 181, June 25, 2020 1457
 ll
Leading Edge

Perspective
Pandemic Preparedness: Developing Vaccines
and Therapeutic Antibodies For COVID-19
Gregory D. Sempowski,1,2,* Kevin O. Saunders,3 Priyamvada Acharya,3 Kevin J. Wiehe,1 and Barton F. Haynes1,4,*
1Department of Medicine, Duke Human Vaccine Institute, Duke University School of Medicine, Durham, NC 27710, USA
2Department of Pathology, Duke Human Vaccine Institute, Duke University School of Medicine, Durham, NC 27710, USA
3Department of Surgery, Duke Human Vaccine Institute, Duke University School of Medicine, Durham, NC 27710, USA
4Department of Immunology, Duke Human Vaccine Institute, Duke University School of Medicine, Durham, NC 27710, USA

*Correspondence: gregory.sempowski@duke.edu (G.D.S.), barton.haynes@duke.edu (B.F.H.)

https://doi.org/10.1016/j.cell.2020.05.041

The SARS-CoV-2 pandemic that causes COVID-19 respiratory syndrome has caused global public
health and economic crises, necessitating rapid development of vaccines and therapeutic counter-
measures. The world-wide response to the COVID-19 pandemic has been unprecedented with
government, academic, and private partnerships working together to rapidly develop vaccine
and antibody countermeasures. Many of the technologies being used are derived from prior gov-
ernment-academic partnerships for response to other emerging infections.

Introduction cholesterolemia, osteoporosis, cancer, and infectious diseases

The novel coronavirus outbreak of SARS-CoV-2, as of May 14,
2020, has resulted in more than 4,300,000 individuals infected
and over 290,000 deaths worldwide (Dong et al., 2020). The cur-
rent pandemic, as well as the potential of future pandemics
based on estimates of undiscovered zoonotic infections (Carroll
et al., 2018), has brought to the forefront urgency and necessity
for rapid development of pandemic countermeasures. Two
countermeasures with promise for controlling the current
SARS-CoV-2 pandemic are recombinant neutralizing antibodies
(Ju et al., 2020; Walker and Burton, 2018) and vaccines (Graham,
2020; Graham et al., 2018; Graham and Sullivan, 2018) directed
against the virus that causes COVID-19, SARS-CoV-2. In partic-
ular, over the past 15 years, the NIAID Center for HIV/AIDS Vac-
cine Immunology (CHAVI) program (Burton et al., 2012; Haynes
et al., 2016), the NIH Vaccine Research Center (Kwong and Mas-
cola, 2012) as well as others, and, for the past three years, the
DARPA Pandemic Prevention Program (P3) program (Cable
et al., 2020; DARPA, 2017; Kose et al., 2019) have worked to
define the platforms and enable technology for HIV vaccine
development and rapid response to viral pandemics. Although
an HIV vaccine has not yet been developed, much of the technol-
ogy the HIV vaccine field has developed is now being used to
fight the COVID-19 pandemic. From the HIV field and the DARPA
preparedness programs have come teams and technologies that
are now responding to the COVID-19 epidemic to both isolate
SARS-CoV-2 neutralizing antibodies and develop SARS-CoV-2
vaccine candidates. Here we comment on some of the strategies
that are being used to develop antibody and vaccine counter-
measures for SARS CoV-2 (Figure 1).
Neutralizing antibodies. Antibodies isolated from a single B
cell are called monoclonal antibodies (mAbs) and have become
an effective new biologic class in our pharmacopeia with a
wide-range of FDA-approved mAbs for indications such as
arthritis and other inflammatory diseases, heart disease, hyper-
(Shepard et al., 2017). Recombinant human or humanized
monoclonal antibodies are proving to be safe, effective, and
highly specific in their ability to target a pathway, process, or
invading pathogen. More than 70 recombinant monoclonal an-
tibodies have now been approved by the FDA for use in the
treatment of infectious, autoimmune and inflammatory, malig-
nant, or cardiovascular diseases (Carter and Lazar, 2018; She-
pard et al., 2017). Specifically, recombinant neutralizing anti-
bodies for infectious diseases, such as for protection from
anthrax toxin and for the prevention of respiratory syncytial vi-
rus infection (Empey et al., 2010; Shepard et al., 2017), have
been approved by the FDA. Neutralizing antibodies are
currently in development for prevention and/or treatment of
HIV (Caskey et al., 2019; Gaudinski et al., 2019) and pending
approval for Ebola (Saphire et al., 2018).
Thus, recombinant neutralizing antibodies isolated from those
infected with SARS-CoV-2 are the most rapid and readily manu-
facturable immune intervention for passive administration that
might be developed to either prevent or treat COVID-19 disease
(Andreano et al., 2020; Brouwer et al., 2020; Ju et al., 2020;
Rogers et al., 2020; Seydoux et al., 2020). SARS-CoV-2 antibody
countermeasures will benefit from the last 20 years of antibody
optimization research that has discovered point mutations in
the Fc portion of antibodies that finetune antibody function and
circulation half-life (Saunders, 2019). Such mutations have
been described for the Fc region of IgG that have prolonged anti-
body half-life for up to 6–7 weeks (Gaudinski et al., 2019; Robbie
et al., 2013; Yu et al., 2016). Additionally, mutations are known
that can increase antibody-dependent infected cell killing and
antibody-dependent complement activation (Idusogie et al.,
2001; Richards et al., 2008). Given the ability of certain anti-
bodies to facilitate SARS-CoV-1 virus entry via engagement of
Fc receptors on host cells (Jaume et al., 2011), the introduction
of mutations that inhibit Fc binding to Fc receptors could also

1458 Cell 181, June 25, 2020 ª 2020 Elsevier Inc.


Perspective
Figure 1. Schema of Iterative and Synergistic Approaches Being
Used to Simultaneously Develop Both Vaccines and Antibody
Countermeasures for SARS-CoV-2/COVID-19
be important for successful development of SARS-CoV-2
neutralizing antibody treatments.
Neutralizing antibodies to the spike protein receptor binding
domain (RBD) protect mice from MERS, SARS-CoV-1, and
SARS-CoV-2 infection (Quinlan et al., 2020; Wang et al., 2018;
Zhou et al., 2018). Thus, neutralizing antibodies are under devel-
opment as proteins or gene-delivered formulations to prevent or
treat SARS-CoV-2 infection. One example of technology now
brought to bear on SARS-CoV-2 countermeasure work is the
strategy developed to isolate and screen for HIV neutralizing an-
tibodies without antibody gene cloning (Liao et al., 2009), and
ll
this technology was recently used in China to isolate the first
neutralizing antibodies to SARS-CoV-2 (Ju et al., 2020). As
seen with SARS-CoV-1 and HIV, viruses resistant to a particular
antibody can emerge among the circulating virus variants (Cas-
key et al., 2017; ter Meulen et al., 2006). In these cases, combi-
nations of two of more neutralizing antibodies can broaden the
efficacy of neutralizing antibody prophylaxis or treatment (Men-
doza et al., 2018; ter Meulen et al., 2006). Thus, large numbers of
effective neutralizing antibodies are needed so that antibody
cocktails can be formulated to provide optimal coverage of
circulating SARS-CoV-2 isolates.
The Defense Advanced Research Projects Agency (DARPA)
Pandemic Prevention Program (P3) within the Department of De-
fense has played a key role in funding and driving public/private
development of end-to-end platforms to rapidly develop anti-
body countermeasures. The DARPA P3 program aims to identify
and propagate viral pathogens, isolate human neutralizing anti-
body sequences, develop effective approaches to deliver
neutralizing antibodies as nucleic acids, and develop good
manufacturing practices for a final drug product to submit to
the Agency for Phase I approval within 60 days of receiving an
outbreak blood specimen (Cable et al., 2020; DARPA, 2017).
To meet this challenge, the Duke DARPA P3 program devel-
oped a ‘‘thaw and infect’’ permissive cell line array to rapidly
grow Risk Group 2 and 3 viruses, such as SARS-CoV-2. This
team also developed platform technologies for real-time virus
growth monitoring, fluorescently labeling viruses, antibody
focus/plaque-reduction neutralization assays, and whole virus
binding assays. By using innovative and time-tested approaches
from the HIV field (Klein et al., 2013; Liao et al., 2009), the
Duke P3 group developed a rapid, virus-independent, high-
throughput, semi-automated Ab sequence isolation and
screening pipeline in high-containment (Figure 2). This pipeline
has recently isolated a panel of H3N2 influenza neutralizing anti-
bodies that provide protection in mouse challenge models and is
now fully engaged in isolating a cocktail of neutralizing SARS-
CoV-2 human monoclonal antibodies. The human angiotensin-
converting enzyme 2 (ACE2) transgenic mouse (Bao et al.,
2020) will be incorporated into the pipeline as well as rhesus
and cynomolgus macaque models (Lu et al., 2020; Rockx
et al., 2020) as SARS-CoV-2 acquisition models to rapidly eval-
uate the ability of antibodies to protect from SARS-CoV-2 chal-
lenge and the possibility of antibody-induced enhancement of
disease (Peeples, 2020). Rapid movement of potent viral neutral-
izing antibodies to testing in the clinic requires an alternative to
typical recombinant antibody protein production. DARPA P3
performer teams and others are therefore developing gene-
delivered approaches (Figure 1) for rapid cGMP-scalable
manufacturing. One approach being used is modified RNA
encapsulated in lipid nanoparticles (LNPs) due to its proven util-
ity and potential for rapid manufacturing (Kose et al., 2019; Pardi
et al., 2018). In pre-clinical models, protective concentrations of
recombinant neutralizing antibodies have been successfully ex-
pressed from intravenously and intramuscularly delivered LNP-
encapsulated mRNA, supporting this approach for use in
response to a rapidly spreading pandemic (Kose et al., 2019;
Pardi et al., 2017b). Moreover, Crowe at al. (Vanderbilt P3
Performer Site), in collaboration with Moderna, has successfully
Cell 181, June 25, 2020 1459
 ll
Figure 2. Accelerated Platform Technology for Rapid B Cell Screening and Isolation of Pathogen Neutralizing Antibodies Being Used to
Isolate SARS-CoV-2 Antibodies
isolated and delivered a Chikungunya neutralizing antibody by
using mRNA in LNPs in a Phase 1 human study (Kose et al.,
2019). In addition to mRNA-LNP, DNA or viral vector approaches
are also being rapidly developed for pandemic prevention anti-
body delivery (Balazs et al., 2011; Muthumani et al., 2013).
SARS-CoV-2 vaccine development. Vaccines are the time-
honored method for establishing long-lived immune memory
for controlling infectious diseases, and technologies have been
developed such that vaccines can now be developed faster
than in previous times (Figure 1) (Graham, 2020; Graham et al.,
2018; Graham and Sullivan, 2018). Over 100 companies or aca-
demic institutions are working on COVID-19 vaccines with stra-
tegies that include recombinant vectors, mRNA in lipid nanopar-
ticles, DNA, inactivated virus, live attenuated virus, virus-like
particles, and protein subunits (Thanh Le et al., 2020; WHO,
2020b). Three vaccine candidates have already advanced to
Phase II testing that include an mRNA vaccine encoding the viral
spike protein from Moderna, an Adeno-type 5 vector vaccine ex-
pressing the S protein from CanSino Biologicals, and a chim-
panzee adenovirus encoding the spike protein from the Jenner
Institute in Oxford, UK. Five other vaccine candidates are also
now in phase I trials including other mRNA/LNP or DNA vaccines
as well as three forms of whole inactivated vaccines
(WHO, 2020b).
As seen from this rapid movement of SARS-CoV-2 vaccine
candidates into human trials, the time it takes to develop vac-
cines for emerging pathogens is decreasing from that in the
past for traditional childhood vaccines. Recently, a DNA vaccine
for the original SARS (SARS-CoV-1) was developed in
20 months, a vaccine for H5 influenza A/Indonesia/2006 in
11 months, a vaccine for H1 influenza A/California/2009
in 4 months, and a Zika virus vaccine in 3.5 months (Graham
et al., 2018). These successes have been brought about by inno-
vative technology and approaches that have allowed for rapid
identification and sequencing of new viral pathogens and new
1460 Cell 181, June 25, 2020
Perspective
technologies for vaccine delivery (Graham and Sullivan, 2018).
For the past 15 years, the HIV vaccine field has pioneered devel-
opment and use of recombinant antibody technology, advanced
computational methods, novel animal models, and new vaccine
delivery approaches to accelerate HIV vaccine immunogen
design and development (Burton et al., 2012; Caskey et al.,
2019; Haynes et al., 2019; Haynes et al., 2016; Klein et al.,
2013; Kwong and Mascola, 2018; Liao et al., 2009). This work
has deciphered the roadblocks for this most difficult-to-develop
HIV vaccine (Haynes et al., 2019). It is expected that the timeline
for a SARS-CoV-2 vaccine will be much faster and much easier
than for HIV-1. Investigators have worked to integrate these iter-
ative approaches for vaccine and antibody countermeasure
development and applied them to a rapid response to the
COVID-19 disease epidemic caused by SARS-CoV-2 (Figure 1).
The COVID-19 pandemic caused by SARS-CoV-2 was first
widely recognized in December 2019, and the first virus
sequence published online in January 2020. By March 16,
2020, the first mRNA/LNP vaccine trial developed by the VRC
in collaboration with Moderna had begun (NIH, 2020). A rapid
SARS-CoV2 vaccine development approach involves the inte-
gration of computational and structural-based immunogen
design strategies; production of immunogens as inactivated vi-
rus; DNA, mRNA, vectored or protein subunits; and immuno-
genic profiling in animal models prior to vaccine manufacturing
and testing in clinical trials (Figure 1). Computational biology
techniques have facilitated the rapid analysis of antibody and vi-
rus sequences for influenza, HIV, and now SARS-CoV-2 to
enable vaccine development (GISAID, 2020; Los Alamos Na-
tional Laboratory, 2020a; Saunders et al., 2019; Wiehe et al.,
2018). Monitoring of HIV evolution by using the Los Alamos
HIV Sequence Database (Los Alamos National Laboratory,
2020a) has been critical for HIV vaccine immunogen designs.
The SARS-CoV-2 virus is evolving, albeit at a slower rate than
HIV, and virus evolution is a concern for successful COVID-19

Perspective
vaccine development. The SARS-CoV-2 spike protein RBD is the
prime target for vaccine-induced neutralizing antibodies
although other spike protein neutralizing epitopes are of interest.
However, comparison of the SARS-CoV2 RBD with that of
SARS-COV-1 reveals only partial homology although both retain
the ability to bind to the ACE2 as a receptor (Wrapp et al., 2020).
The GISAID database is proving to be helpful information for
monitoring the viral evolution of SARS-CoV-2 (GISAID, 2020),
and the Los Alamos National Laboratory is developing a website
with tools for analysis of global SARS-CoV-2 spike protein se-
quences (Korber et al., 2020; Los Alamos National Labora-
tory, 2020b).
Structural determination of the primary targets of neutralizing
antibodies, for example, hemagglutinin in influenza, Env in HIV,
and now the spike protein in SARS-CoV-2, has provided valu-
able atomic-level insight for vaccine design strategies. In partic-
ular, cryo-electron microscopy has enabled the rapid solution of
the structure of the SARS-CoV-2 spike protein (Walls et al., 2020;
Wrapp et al., 2020). Structural biology analysis of pathogens
combines structural, computational, biophysical, and biochem-
ical methods to understand interactions of pathogens with the
immune system (Henderson et al., 2020; LaBranche et al.,
2019; Murin et al., 2019; Saunders et al., 2019). Early this year,
structural biologists pivoted to apply technology developed for
HIV-1 envelope or respiratory syncytial virus (RSV) structural
biology to fast-track structure-based vaccine design for
COVID-19 (Lan et al., 2020; Walls et al., 2020; Wrapp et al.,
2020; Yuan et al., 2020). Currently, established pipelines for
high-resolution cryo-EM structural determination of the SARS-
CoV-2 spike are integrated with the computational teams, thus
providing atomic-level feedback to COVID-19 vaccine designs.
The past eight years of HIV-1 antibody discovery has provided
templates for HIV-1 vaccine design aiming to elicit broadly reac-
tive neutralizing antibodies (Kwong and Mascola, 2018; Sok and
Burton, 2018). From the study of the ontogeny of HIV neutralizing
antibodies it has become clear that an effective vaccine will likely
require multiple immunogens administered in a specific order to
facilitate proper antibody development to multiple neutralizing
targets on HIV (Haynes et al., 2019). Hopefully, the development
of SARS-CoV-2 neutralizing antibodies will require a much
simpler vaccination regimen like the Zika vaccine, where one im-
munization with one immunogen was sufficient to elicit protec-
tive neutralizing antibodies (Pardi et al., 2017a). Such a vaccine
would be amenable to rapid development, large-scale
manufacturing, and global administration.
What will follow rapidly now for prevention of COVID-19 will be
a number of mRNA/LNP (e.g., from Moderna/NIAID, BioNTech/
Fosum, Pharma/Pfizer) or DNA (e.g., Inovio) vaccines as well
as attenuated viruses, proteins, nanoparticles, and viral vectors
containing SARS-CoV-2 viral genes as vaccine candidates mov-
ing through safety and immunogenicity trials, and a smaller sub-
set of vaccine candidates will be tested in Phase III or efficacy tri-
als to continue to determine if they are safe, as well to determine
their efficacy. In parallel now with Phase I and II trials, it is impor-
tant to develop capacity for large-scale vaccine production, in
the event of a successful efficacy trial (Corey et al., 2020; Gra-
ham, 2020). It is possible that genetic immunization strategies
such as DNA or mRNA in LNPs can be manufactured more
ll
rapidly than proteins or viral vectors and can be more cost
effective.
In addition to efficacy being the primary goal of SARS-CoV-2
vaccine development, safety is also a major concern (Peeples,
2020). Immunization with a SARS-CoV-1 vaccine has induced
vaccine-associated immunopathology in the lung (Bolles et al.,
2011; Liu et al., 2019; Tseng et al., 2012). Both CD4 and CD8
T cell responses have been suggested to also be protective for
SARS-CoV-1 (Zhao et al., 2016). Thus, careful animal preclinical
studies as well as intense monitoring of human clinical trials will
be of critical importance to developing safe and effective anti-
COVID-19 antibody and vaccine countermeasures.
Finally, collaboration and coordination will be essential to
ending the pandemic. Globally, one example of private support
for COVID-19 research is the Coalition for Epidemic Prepared-
ness Innovations (CEPI). CEPI is raising funds for COVID-19 vac-
cine development and as well is funding vaccine development
projects (Gouglas et al., 2019). The Bill & Melinda Gates Founda-
tion has made a $250 million commitment to fight COVID-19 and
has established a Coronavirus Immunotherapy Consortium, or
CoVIC, to foster sharing and comparison of SARS-CoV-2 anti-
bodies to speed therapeutic antibody development (Bill & Me-
linda Gates Foundation, 2020). The recent formation of an NIH-
organized public-private partnership, termed Accelerating
COVID-19 Therapeutic Interventions and Vaccines (ACTIV)
(Corey et al., 2020; Kaiser, 2020), is necessary and will facilitate
a coordinated COVID-19 pandemic response. Globally, the
World Health Organization is playing a critical multinational coor-
dination and informational role (WHO, 2020a).
Summary
Government-funded and private initiatives have synergized to
provide countermeasure platforms to rapidly respond to the
SARS-CoV-2 pandemic. Continued cooperation among public
and private institutions coupled with speed of development of
antibody countermeasures and vaccines, with rapid evaluation
of their safety and efficacy, and early planning for scale-up and
manufacture will be critical for expeditious control of the global
COVID-19 pandemic.
ACKNOWLEDGMENTS
Funded by NIH grants AI142596 (B.F.H.), AI145687 and AI150415 (P.A.),
AI058607 (G.D.S.); Department of Defense HR0011-17-2-0069 (G.D.S.); and
the Translating Duke Health Initiative (P.A.). All authors wrote and edited the
manuscript. We thank Megan Averill for editorial assistance and David East-
erhoff and QiFeng Han for their contributions to the DARPA P3 program.
DECLARATION OF INTERESTS
B.F.H. and G.D.S. have a patent submitted for the emerging infections pre-
paredness platform and influenza antibodies (No. 62_902705).
REFERENCES
Andreano, E., Nicastri, E., Paciello, I., Pileri, P., Manganaro, N., Piccini, G.,
Manenti, A., Pantano, E., Kabanova, A., Troisi, M., et al. (2020). Identification
of neutralizing human monoclonal antibodies from Italian Covid-19 convales-
cent patients. bioRxiv. https://doi.org/10.1101/2020.05.05.078154.
Cell 181, June 25, 2020 1461
 ll
Balazs, A.B., Chen, J., Hong, C.M., Rao, D.S., Yang, L., and Baltimore, D.
(2011). Antibody-based protection against HIV infection by vectored immuno-
prophylaxis. Nature 481, 81–84.
Bao, L., Deng, W., Huang, B., Gao, H., Liu, J., Ren, L., Wei, Q., Yu, P., Xu, Y.,
Qi, F., et al. (2020). The pathogenicity of SARS-CoV-2 in hACE2 transgenic
mice. Nature. https://doi.org/10.1038/s41586-020-2312-y.
Bill & Melinda Gates Foundation (2020). Press Release: COVID-19 Therapeu-
tics Accelerator Awards $20 Million in Initial Grants to Fund Clinical Trials.
Bolles, M., Deming, D., Long, K., Agnihothram, S., Whitmore, A., Ferris, M.,
Funkhouser, W., Gralinski, L., Totura, A., Heise, M., and Baric, R.S. (2011). A
double-inactivated severe acute respiratory syndrome coronavirus vaccine
provides incomplete protection in mice and induces increased eosinophilic
proinflammatory pulmonary response upon challenge. J. Virol. 85,
12201–12215.
Brouwer, P., Caniels, T., van Straten, K., Snitselaar, J., Aldon, Y., Bangaru, S.,
Torres, J., Okba, N., Claireaux, M., Kerster, G., et al. (2020). Potent neutralizing
antibodies from COVID-19 patients define multiple targets of vulnerability. bio-
Rxiv. https://doi.org/10.1101/2020.05.12.088716.
Burton, D.R., Ahmed, R., Barouch, D.H., Butera, S.T., Crotty, S., Godzik, A.,
Kaufmann, D.E., McElrath, M.J., Nussenzweig, M.C., Pulendran, B., et al.
(2012). A Blueprint for HIV Vaccine Discovery. Cell Host Microbe 12, 396–407.
Cable, J., Srikantiah, P., Crowe, J.E., Jr., Pulendran, B., Hill, A., Ginsberg, A.,
Koff, W., Mathew, A., Ng, T., Jansen, K., et al. (2020). Vaccine innovations for
emerging infectious diseases-a symposium report. Ann. N Y Acad. Sci. 1462,
14–26.
Carroll, D., Daszak, P., Wolfe, N.D., Gao, G.F., Morel, C.M., Morzaria, S., Pa-
blos-Méndez, A., Tomori, O., and Mazet, J.A.K. (2018). The Global Virome
Project. Science 359, 872–874.
Carter, P.J., and Lazar, G.A. (2018). Next generation antibody drugs: pursuit of
the ‘high-hanging fruit’. Nat. Rev. Drug Discov. 17, 197–223.
Caskey, M., Schoofs, T., Gruell, H., Settler, A., Karagounis, T., Kreider, E.F.,
Murrell, B., Pfeifer, N., Nogueira, L., Oliveira, T.Y., et al. (2017). Antibody 10-
1074 suppresses viremia in HIV-1-infected individuals. Nat. Med. 23, 185–191.
Caskey, M., Klein, F., and Nussenzweig, M.C. (2019). Broadly neutralizing anti-
HIV-1 monoclonal antibodies in the clinic. Nat. Med. 25, 547–553.
Corey, B.L., Mascola, J.R., Fauci, A.S., and Collins, F.S. (2020). A strategic
approach to COVID-19 vaccine R&D. Science, eabc5312.
DARPA (2017). Pandemic Prevention Platform (P3). https://www.darpa.mil/
program/pandemic-prevention-platform.
Dong, E., Du, H., and Gardner, L. (2020). An interactive web-based dashboard
to track COVID-19 in real time. Lancet Infect. Dis. 20, 533–534.
Empey, K.M., Peebles, R.S., Jr., and Kolls, J.K. (2010). Pharmacologic ad-
vances in the treatment and prevention of respiratory syncytial virus. Clin.
Infect. Dis. 50, 1258–1267.
Gaudinski, M.R., Houser, K.V., Doria-Rose, N.A., Chen, G.L., Rothwell, R.S.S.,
Berkowitz, N., Costner, P., Holman, L.A., Gordon, I.J., Hendel, C.S., et al.; VRC
605 study team (2019). Safety and pharmacokinetics of broadly neutralising
human monoclonal antibody VRC07-523LS in healthy adults: a phase 1
dose-escalation clinical trial. Lancet HIV 6, e667–e679.
GISAID (2020). Next hCoV-19 App (Germany: Munich). https://www.gisaid.
org/epiflu-applications/next-hcov-19-app/.
Gouglas, D., Christodoulou, M., Plotkin, S.A., and Hatchett, R. (2019). CEPI:
Driving Progress Toward Epidemic Preparedness and Response. Epidemiol.
Rev. 41, 28–33.
Graham, B.S. (2020). Rapid COVID-19 vaccine development. Science,
eabb8923.
Graham, B.S., and Sullivan, N.J. (2018). Emerging viral diseases from a vacci-
nology perspective: preparing for the next pandemic. Nat. Immunol. 19, 20–28.
Graham, B.S., Mascola, J.R., and Fauci, A.S. (2018). Novel Vaccine Technol-
ogies: Essential Components of an Adequate Response to Emerging Viral Dis-
eases. JAMA 319, 1431–1432.
1462 Cell 181, June 25, 2020
Perspective
Haynes, B.F., Shaw, G.M., Korber, B., Kelsoe, G., Sodroski, J., Hahn, B.H.,
Borrow, P., and McMichael, A.J. (2016). HIV-Host Interactions: Implications
for Vaccine Design. Cell Host Microbe 19, 292–303.
Haynes, B.F., Burton, D.R., and Mascola, J.R. (2019). Multiple roles for HIV
broadly neutralizing antibodies. Sci. Transl. Med. 11, eaaz2686.
Henderson, R., Lu, M., Zhou, Y., Mu, Z., Parks, R., Han, Q., Hsu, A.L., Carter,
E., Blanchard, S.C., Edwards, R.J., et al. (2020). Disruption of the HIV-1 Enve-
lope allosteric network blocks CD4-induced rearrangements. Nat. Commun.
11, 520.
Idusogie, E.E., Wong, P.Y., Presta, L.G., Gazzano-Santoro, H., Totpal, K.,
Ultsch, M., and Mulkerrin, M.G. (2001). Engineered antibodies with increased
activity to recruit complement. J. Immunol. 166, 2571–2575.
Jaume, M., Yip, M.S., Cheung, C.Y., Leung, H.L., Li, P.H., Kien, F., Dutry, I.,
Callendret, B., Escriou, N., Altmeyer, R., et al. (2011). Anti-severe acute respi-
ratory syndrome coronavirus spike antibodies trigger infection of human im-
mune cells via a pH- and cysteine protease-independent FcgR pathway.
J. Virol. 85, 10582–10597.
Ju, B., Zhang, Q., Ge, X., Wang, R., Yu, J., Shan, S., Zhou, B., Song, S., Tang,
X., Yu, J., et al. (2020). Potent human neutralizing antibodies elicited by SARS-
CoV-2 infection. bioRxiv. https://doi.org/10.1101/2020.03.21.990770.
Kaiser, J. (2020). NIH organizes hunt for drugs. Science 368, 351, 351.
Klein, F., Mouquet, H., Dosenovic, P., Scheid, J.F., Scharf, L., and Nussenz-
weig, M.C. (2013). Antibodies in HIV-1 vaccine development and therapy. Sci-
ence 341, 1199–1204.
Korber, B., Fischer, W., Gnanakaran, S., Yoon, H., Theiler, J., Abfalterer, W.,
Foley, B., Giorgi, E., Bhattacharya, T., Parker, M., et al. (2020). Spike mutation
pipeline reveals the emergence of a more transmissible form of SARS-CoV-2.
bioRxiv. https://doi.org/10.1101/2020.04.29.069054.
Kose, N., Fox, J.M., Sapparapu, G., Bombardi, R., Tennekoon, R.N., de Silva,
A.D., Elbashir, S.M., Theisen, M.A., Humphris-Narayanan, E., Ciaramella, G.,
et al. (2019). A lipid-encapsulated mRNA encoding a potently neutralizing hu-
man monoclonal antibody protects against chikungunya infection. Sci. Immu-
nol. 4, eaaw6647.
Kwong, P.D., and Mascola, J.R. (2012). Human antibodies that neutralize HIV-
1: identification, structures, and B cell ontogenies. Immunity 37, 412–425.
Kwong, P.D., and Mascola, J.R. (2018). HIV-1 Vaccines Based on Antibody
Identification, B Cell Ontogeny, and Epitope Structure. Immunity 48, 855–871.
LaBranche, C.C., Henderson, R., Hsu, A., Behrens, S., Chen, X., Zhou, T.,
Wiehe, K., Saunders, K.O., Alam, S.M., Bonsignori, M., et al. (2019). Neutrali-
zation-guided design of HIV-1 envelope trimers with high affinity for the unmu-
tated common ancestor of CH235 lineage CD4bs broadly neutralizing anti-
bodies. PLoS Pathog. 15, e1008026.
Lan, J., Ge, J., Yu, J., Shan, S., Zhou, H., Fan, S., Zhang, Q., Shi, X., Wang, Q.,
Zhang, L., and Wang, X. (2020). Structure of the SARS-CoV-2 spike receptor-
binding domain bound to the ACE2 receptor. Nature 581, 215–220.
Liao, H.X., Levesque, M.C., Nagel, A., Dixon, A., Zhang, R., Walter, E., Parks,
R., Whitesides, J., Marshall, D.J., Hwang, K.K., et al. (2009). High-throughput
isolation of immunoglobulin genes from single human B cells and expression
as monoclonal antibodies. J. Virol. Methods 158, 171–179.
Liu, L., Wei, Q., Lin, Q., Fang, J., Wang, H., Kwok, H., Tang, H., Nishiura, K.,
Peng, J., Tan, Z., et al. (2019). Anti-spike IgG causes severe acute lung injury
by skewing macrophage responses during acute SARS-CoV infection. JCI
Insight 4, e123158.
Los Alamos National Laboratory (2020a). Los Alamos HIV Sequence Database
(US: Los Alamos, NM). http://www.hiv.lanl.gov/.
Los Alamos National Laboratory (2020b). SARS-CoV-2 Sequence Analysis
Pipeline. https://cov.lanl.gov/.
Lu, S., Zhao, Y., Yu, W., Yang, Y., Gao, J., Wang, J., Kuang, D., Yang, M.,
Yang, J., Ma, C., et al. (2020). Comparison of SARS-CoV-2 infections among
3 species of non-human primates. bioRxiv. https://doi.org/10.1101/2020.04.
08.031807.
Mendoza, P., Gruell, H., Nogueira, L., Pai, J.A., Butler, A.L., Millard, K., Leh-
mann, C., Suárez, I., Oliveira, T.Y., Lorenzi, J.C.C., et al. (2018). Combination

Perspective
therapy with anti-HIV-1 antibodies maintains viral suppression. Nature 561,
479–484.
Murin, C.D., Wilson, I.A., and Ward, A.B. (2019). Antibody responses to viral in-
fections: a structural perspective across three different enveloped viruses.
Nat. Microbiol. 4, 734–747.
Muthumani, K., Flingai, S., Wise, M., Tingey, C., Ugen, K.E., and Weiner, D.B.
(2013). Optimized and enhanced DNA plasmid vector based in vivo construc-
tion of a neutralizing anti-HIV-1 envelope glycoprotein Fab. Hum. Vaccin. Im-
munother. 9, 2253–2262.
NIH (2020). Press Release: NIH clinical trial of investigational vaccine for
COVID-19 begins.
Pardi, N., Hogan, M.J., Pelc, R.S., Muramatsu, H., Andersen, H., DeMaso,
C.R., Dowd, K.A., Sutherland, L.L., Scearce, R.M., Parks, R., et al. (2017a).
Zika virus protection by a single low-dose nucleoside-modified mRNA vacci-
nation. Nature 543, 248–251.
Pardi, N., Secreto, A.J., Shan, X., Debonera, F., Glover, J., Yi, Y., Muramatsu,
H., Ni, H., Mui, B.L., Tam, Y.K., et al. (2017b). Administration of nucleoside-
modified mRNA encoding broadly neutralizing antibody protects humanized
mice from HIV-1 challenge. Nat. Commun. 8, 14630.
Pardi, N., Hogan, M.J., Porter, F.W., and Weissman, D. (2018). mRNA vaccines
- a new era in vaccinology. Nat. Rev. Drug Discov. 17, 261–279.
Peeples, L. (2020). News Feature: Avoiding pitfalls in the pursuit of a COVID-19
vaccine. Proc. Natl. Acad. Sci. USA 117, 8218–8221.
Quinlan, B.D., Mou, H., Zhang, L., Guo, Y., He, W., Ojha, A., Parcells, M.S.,
Luo, G., Li, W., Zhong, G., et al. (2020). The SARS-CoV-2 receptor-binding
domain elicits a potent neutralizing response without antibody-dependent
enhancement. bioRxiv. https://doi.org/10.1101/2020.04.10.036418.
Richards, J.O., Karki, S., Lazar, G.A., Chen, H., Dang, W., and Desjarlais, J.R.
(2008). Optimization of antibody binding to FcgammaRIIa enhances macro-
phage phagocytosis of tumor cells. Mol. Cancer Ther. 7, 2517–2527.
Robbie, G.J., Criste, R., Dall’acqua, W.F., Jensen, K., Patel, N.K., Losonsky,
G.A., and Griffin, M.P. (2013). A novel investigational Fc-modified humanized
monoclonal antibody, motavizumab-YTE, has an extended half-life in healthy
adults. Antimicrob. Agents Chemother. 57, 6147–6153.
Rockx, B., Kuiken, T., Herfst, S., Bestebroer, T., Lamers, M.M., Oude Munnink,
B.B., de Meulder, D., van Amerongen, G., van den Brand, J., Okba, N.M.A.,
et al. (2020). Comparative pathogenesis of COVID-19, MERS, and SARS in a
nonhuman primate model. Science, eabb7314.
Rogers, T.F., Zhao, F., Huang, D., Beutler, N., Abbott, R.K., Callaghan, S., Gar-
cia, E., He, W.-t., Hurtado, J., Limbo, O., et al. (2020). Rapid isolation of potent
SARS-CoV-2 neutralizing antibodies and protection in a small animal model.
bioRxiv. https://doi.org/10.1101/2020.05.11.088674.
Saphire, E.O., Schendel, S.L., Fusco, M.L., Gangavarapu, K., Gunn, B.M.,
Wec, A.Z., Halfmann, P.J., Brannan, J.M., Herbert, A.S., Qiu, X., et al.; Viral
Hemorrhagic Fever Immunotherapeutic Consortium (2018). Systematic Anal-
ysis of Monoclonal Antibodies against Ebola Virus GP Defines Features that
Contribute to Protection. Cell 174, 938–952.e13, e913.
Saunders, K.O. (2019). Conceptual Approaches to Modulating Antibody
Effector Functions and Circulation Half-Life. Front. Immunol. 10, 1296.
Saunders, K.O., Wiehe, K., Tian, M., Acharya, P., Bradley, T., Alam, S.M., Go,
E.P., Scearce, R., Sutherland, L., Henderson, R., et al. (2019). Targeted selec-
tion of HIV-specific antibody mutations by engineering B cell maturation. Sci-
ence 366, eaay7199.
Seydoux, E., Homad, L.J., MacCamy, A.J., Parks, K.R., Hurlburt, N.K., Jenne-
wein, M.F., Akins, N.R., Stuart, A.B., Wan, Y.-H., Feng, J., et al. (2020). Char-
ll
acterization of neutralizing antibodies from a SARS-CoV-2 infected individual.
bioRxiv. https://doi.org/10.1101/2020.05.12.091298.
Shepard, H.M., Phillips, G.L., D Thanos, C., and Feldmann, M. (2017). Devel-
opments in therapy with monoclonal antibodies and related proteins. Clin.
Med. (Lond.) 17, 220–232.
Sok, D., and Burton, D.R. (2018). Recent progress in broadly neutralizing anti-
bodies to HIV. Nat. Immunol. 19, 1179–1188.
ter Meulen, J., van den Brink, E.N., Poon, L.L., Marissen, W.E., Leung, C.S.,
Cox, F., Cheung, C.Y., Bakker, A.Q., Bogaards, J.A., van Deventer, E., et al.
(2006). Human monoclonal antibody combination against SARS coronavirus:
synergy and coverage of escape mutants. PLoS Med. 3, e237.
Thanh Le, T., Andreadakis, Z., Kumar, A., Gómez Román, R., Tollefsen, S., Sa-
ville, M., and Mayhew, S. (2020). The COVID-19 vaccine development land-
scape. Nat. Rev. Drug Discov. 19, 305–306.
Tseng, C.T., Sbrana, E., Iwata-Yoshikawa, N., Newman, P.C., Garron, T., At-
mar, R.L., Peters, C.J., and Couch, R.B. (2012). Immunization with SARS co-
ronavirus vaccines leads to pulmonary immunopathology on challenge with
the SARS virus. PLoS ONE 7, e35421.
Walker, L.M., and Burton, D.R. (2018). Passive immunotherapy of viral infec-
tions: ‘super-antibodies’ enter the fray. Nat. Rev. Immunol. 18, 297–308.
Walls, A.C., Park, Y.J., Tortorici, M.A., Wall, A., McGuire, A.T., and Veesler, D.
(2020). Structure, Function, and Antigenicity of the SARS-CoV-2 Spike Glyco-
protein. Cell 181, 281–292.e6, e286.
Wang, L., Shi, W., Chappell, J.D., Joyce, M.G., Zhang, Y., Kanekiyo, M.,
Becker, M.M., van Doremalen, N., Fischer, R., Wang, N., et al. (2018). Impor-
tance of Neutralizing Monoclonal Antibodies Targeting Multiple Antigenic Sites
on the Middle East Respiratory Syndrome Coronavirus Spike Glycoprotein To
Avoid Neutralization Escape. J. Virol. 92, e02002–e02017.
WHO (2020a). Coronavirus disease (COVID-2019) technical guidance, World
Health Organization. https://www.who.int/emergencies/diseases/novel-
coronavirus-2019/technical-guidance.
WHO (2020b). Draft landscape of COVID-19 candidate vaccines – 20 March
2020. https://www.who.int/blueprint/priority-diseases/key-action/novel-
coronavirus-landscape-ncov.pdf?ua=1.
Wiehe, K., Bradley, T., Meyerhoff, R.R., Hart, C., Williams, W.B., Easterhoff, D.,
Faison, W.J., Kepler, T.B., Saunders, K.O., Alam, S.M., et al. (2018). Functional
Relevance of Improbable Antibody Mutations for HIV Broadly Neutralizing
Antibody Development. Cell Host Microbe 23, 759–765.e6, e756.
Wrapp, D., Wang, N., Corbett, K.S., Goldsmith, J.A., Hsieh, C.L., Abiona, O.,
Graham, B.S., and McLellan, J.S. (2020). Cryo-EM structure of the 2019-
nCoV spike in the prefusion conformation. Science 367, 1260–1263.
Yu, X.Q., Robbie, G.J., Wu, Y., Esser, M.T., Jensen, K., Schwartz, H.I., Bell-
amy, T., Hernandez-Illas, M., and Jafri, H.S. (2016). Safety, Tolerability, and
Pharmacokinetics of MEDI4893, an Investigational, Extended-Half-Life, Anti-
Staphylococcus aureus Alpha-Toxin Human Monoclonal Antibody, in Healthy
Adults. Antimicrob. Agents Chemother. 61, 61.
Yuan, M., Wu, N.C., Zhu, X., Lee, C.D., So, R.T.Y., Lv, H., Mok, C.K.P., and Wil-
son, I.A. (2020). A highly conserved cryptic epitope in the receptor binding do-
mains of SARS-CoV-2 and SARS-CoV. Science 368, 630–633.
Zhao, J., Zhao, J., Mangalam, A.K., Channappanavar, R., Fett, C., Meyerholz,
D.K., Agnihothram, S., Baric, R.S., David, C.S., and Perlman, S. (2016). Airway
Memory CD4(+) T Cells Mediate Protective Immunity against Emerging Respi-
ratory Coronaviruses. Immunity 44, 1379–1391.
Zhou, Y., Jiang, S., and Du, L. (2018). Prospects for a MERS-CoV spike vac-
cine. Expert Rev. Vaccines 17, 677–686.
Cell 181, June 25, 2020 1463
 ll
Leading Edge

Perspective
Molecular Transducers of Physical Activity
Consortium (MoTrPAC): Mapping the Dynamic
Responses to Exercise
James A. Sanford,1,12 Christopher D. Nogiec,2,12 Malene E. Lindholm,3,12 Joshua N. Adkins,1 David Amar,3
Surendra Dasari,4 Jonelle K. Drugan,5 Facundo M. Fernández,6 Shlomit Radom-Aizik,7 Simon Schenk,8
Michael P. Snyder,3 Russell P. Tracy,9 Patrick Vanderboom,4 Scott Trappe,10,11,12,* Martin J. Walsh,2,11,12,* and the
Molecular Transducers of Physical Activity Consortium
1Pacific Northwest National Laboratory, Richland, WA 99354, USA
2Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
3Stanford University, Stanford, CA 94305, USA
4Mayo Clinic, Rochester, MN 55901, USA
5National Institutes of Health, Bethesda, MD 20892, USA
6Georgia Institute of Technology, Atlanta, GA 30322, USA
7University of California, Irvine, Irvine, CA 92617, USA
8University of California, San Diego, La Jolla, CA 92093, USA
9University of Vermont, Burlington, VT 05405, USA
10Ball State University, Muncie, IN 47306, USA
11Senior author
12These authors contributed equally

*Correspondence: strappe@bsu.edu (S.T.), martin.walsh@mssm.edu (M.J.W.)

https://doi.org/10.1016/j.cell.2020.06.004

Exercise provides a robust physiological stimulus that evokes cross-talk among multiple tissues that when
repeated regularly (i.e., training) improves physiological capacity, benefits numerous organ systems, and de-
creases the risk for premature mortality. However, a gap remains in identifying the detailed molecular signals
induced by exercise that benefits health and prevents disease. The Molecular Transducers of Physical Activ-
ity Consortium (MoTrPAC) was established to address this gap and generate a molecular map of exercise.
Preclinical and clinical studies will examine the systemic effects of endurance and resistance exercise across
a range of ages and fitness levels by molecular probing of multiple tissues before and after acute and chronic
exercise. From this multi-omic and bioinformatic analysis, a molecular map of exercise will be established.
Altogether, MoTrPAC will provide a public database that is expected to enhance our understanding of the
health benefits of exercise and to provide insight into how physical activity mitigates disease.

INTRODUCTION
Exercise perturbs multiple systems from the whole body to the
molecular level in an integrated manner (Hawley et al., 2014).
However, in-depth fundamental knowledge into the molecular
and cellular mechanisms that are responsible for physical activ-
ity’s benefits on multiple organ systems and the diseases and
disorders that derive from inactivity is incomplete (Booth et al.,
2017; Neufer et al., 2015). A better understanding of these bio-
logical processes and pathways would allow for the develop-
ment of targeted exercise interventions and prescriptions and
provide a foundation for developing exercise-mimetic pharma-
cologic interventions.
The Molecular Transducers of Physical Activity Consortium
(MoTrPAC) was established to elucidate how exercise improves
health and ameliorates diseases by building a map of the molec-
ular responses to acute and chronic exercise. In 2014, a portfolio
analysis of National Institutes of Health (NIH) grants revealed that
most research regarding physical activity involved disease pre-
vention or treatment. In fact, almost all the grants which em-
ployed an exercise intervention only addressed health outcomes
and adherence issues. The MoTrPAC initiative provides a much
needed comprehensive program to understand the interplay be-
tween these biological systems with the goal of improving the
design of physical activity interventions. In addition, there is a
potential to identify molecular targets that can be manipulated
to mimic the effects of exercise in persons unable to do so for
a variety of reasons, such as physical disability, coma, or pa-
ralysis.
To address the gaps in knowledge about how exercise en-
hances health and ameliorates disease, multiple agencies at
the NIH—including the National Institute of Arthritis and Muscu-
loskeletal and Skin Diseases (NIAMS), the National Institute of
Diabetes and Digestive and Kidney Diseases (NIDDK), the Na-
tional Institute on Aging (NIA), and other institutes and centers
who participated in the trans-NIH Exercise Interest Group—pro-
posed the Common Fund program supporting MoTrPAC. To
create a substantive complex map of molecular transducers in

1464 Cell 181, June 25, 2020 ª 2020 Elsevier Inc.


Perspective
diverse populations across the lifespan, MoTrPAC established a
multi-site collaboration across the United States encompassing
various scientific disciplines: preclinical animal study sites and
human clinical exercise sites to perform the exercise testing, ex-
ercise interventions, and biospecimen collection; a consortium
coordinating center to manage sample collection, distribution
of samples, and consortium logistics; chemical analysis sites
to perform ‘omics analysis from the samples collected; and a
bioinformatics center to collaboratively facilitate data quality
control, bioinformatics analysis, and dissemination to make the
data and related resources available to the broad scientific com-
munity (Figure 1). The animal studies will enable analysis of the
effects of exercise on many different tissues that are not readily
obtainable in humans, thereby enabling a broad view of the sys-
temic effects of exercise. The collection of human specimens
(blood, muscle, and adipose) will permit the analysis of these
critical systems, which are central to the energetics of exercise
and appear to interact in a coordinated manner to improve over-
all metabolic health (Pedersen and Febbraio, 2012; Romijn et al.,
1993; Stanford and Goodyear, 2018). In addition to providing in-
formation concerning the effects of exercise at different physio-
logical and molecular levels, the large scope of this study (hu-
mans: N = 2,600; rats: N = 820) will create a complex,
integrative dataset that will be available to the scientific commu-
nity. This dataset and some associated biospecimens can be
leveraged by other groups by proposing ancillary studies to
MoTrPAC.
ll
Figure 1. General Overview of MoTrPAC
Preclinical Animal Study Sites (PASSs) (rats) and
Human Clinical Exercise Sites will collect bio-
specimen samples after acute and chronic exercise.
The biospecimen samples will be sent to a central
biorepository where they will be logged, processed,
and distributed to various Chemical Analysis Sites
(CASs) for ‘omics analysis. A portion of the bio-
specimen samples will be kept at the biorepository
for future ancillary study opportunities by the sci-
entific community. Data generated from the CASs
will be transferred to the Bioinformatics Center (BIC)
for a multi-omics, multi-species, multi-tissue, and
multi-time point integration of the data with the goal
of generating a molecular map of exercise. The data
will be available to the scientific community via the
MoTrPAC Data Hub: https://motrpac-data.org.
Figure design by Jill K. Gregory, Mount Sinai.
Preclinical Animal Study Sites
A primary goal of the Preclinical Animal
Study Sites (PASSs) investigations is to
enable the analysis of the systemic ef-
fects of exercise across many different
organs and blood, most of which cannot
be collected in humans. The first phase
of the PASS studies is being conducted
across three separate sites, with each
collecting numerous biospecimens after
acute (i.e., single bout) treadmill exercise
or chronic treadmill training of male and
female F344 rats at 6 and 18 months of
age (Figure 2). For both acute and chronic
exercise studies, the same biospecimens are being collected
from non-exercised, control animals; for the acute studies this in-
cludes groups accounting for potential effects of time of day and
time of feeding. This rat strain was selected for MoTrPAC
because there is a large body of previous exercise research uti-
lizing this strain, and rats provide larger amounts of tissue in
contrast to mice. Larger tissue samples allow for multiple assays
to be performed on the same sample, which supports bioinfor-
matics analysis for the development of the molecular map of ex-
ercise. Also, larger tissues will provide additional material for
ancillary studies. To control variation across the three PASSs,
the rats are being provided from the same NIA colony, and the
housing and feeding conditions are standardized across the
sites. Furthermore, because rats are most active during the
‘‘dark’’ light phase, all rats are first being acclimated to a reverse
light cycle for a minimum of 10 days, and with all exercise bouts
conducted during the dark cycle.
The PASS study design for the acute response to exercise en-
tails a single 30-min bout of treadmill running (intensity: 80%–
90% VO2max; incline: 5 ; speed: 6 months: male – 16.8 m/min,
female – 18.0 m/min; 18 months: male – 12.0 m/min, female –
13.8 m/min), with tissue collections occurring immediately
post-exercise and at six additional times up to 48 h after the ex-
ercise bout. This sampling series is weighted toward early time
points (0, 0.5, 1, and 4 h post-exercise) to capture the temporal
dynamics of the molecular responses, but it also includes
later time points (7, 24, and 48 h post-exercise) to capture
Cell 181, June 25, 2020 1465
 ll
Perspective

Figure 2. A General Schematic of the Preclinical and Clinical Studies

(A) Biospecimen collection for the rat studies (N = 820, including non-exercised controls) is ongoing and will be completed in Fall 2020. For both the acute (i.e.,
single bout; n = 526) and chronic phases (i.e., training; n = 294), 6- and 18-month-old male and female rats are being studied, and both exercise phases include a
cohort of non-exercised controls.
(B) 2,600 healthy individuals will be evaluated physiologically, and biospecimens will be collected to accommodate molecular probing of various tissues before
and after acute and chronic endurance or resistance exercise. For the adult cohort, the control participants will rest quietly (i.e., no exercise) during the acute bout
with biospecimen collections.
*Assessments include blood profile, body composition, heart health, VO2max and strength.
Figure design by Jill K. Gregory, Mount Sinai.

long-duration primary responses as well as secondary molecular
events (detailed protocols are available at https://motrpac.org/
protocols.cfm). To study the biological events that occur during
the early, intermediate, and later stages of endurance training,
the PASS study design for the chronic response to exercise,
which has been completed, entailed up to 8 weeks of treadmill
training (5 days per week at 70% VO2max), with tissues
collected 48 h after 1, 2, 4, and 8 weeks of training; incline, dura-
tion, and speed of exercise progressively increased on a daily to
weekly basis during the initial 6 weeks of training. Sex differ-
ences in both the acute exercise response and the training
response are being investigated along with other study aims.
The most powerful aspect of the PASS design is the breadth of
tissues collected. In addition to being studied in the context of
MoTrPAC, these will serve as a data resource for generating hy-
potheses for future studies. For both the acute exercise and
chronic exercise training studies (including the non-exercise
controls), as many as 27 biospecimens per rat are being
collected for potential analysis. In addition to biospecimen
collection, other phenotypic outcomes are being collected,
including blood lactate concentration, maximal oxygen con-
sumption (VO2max), and body composition. At Chemical Anal-
ysis Sites, initial biospecimens of focus will include those that
overlap with the human studies (i.e., plasma, skeletal muscle,
white adipose), as well as liver, heart, kidney, lungs, brain, and
brown adipose. It is expected that nucleic acid, proteomic, and
targeted metabolomic assays will be performed only on tissues
where the amount is non-limiting, whereas transcriptomics,
non-targeted metabolomics, and non-targeted lipidomics will
be performed on all tissues. Together, these assays are ex-
pected to provide molecular and physiological insights about
the effect of exercise on many different organs. Ultimately,
MoTrPAC should begin to explain how molecular transducers
function across an entire mammal (Pedersen and Feb-
braio, 2012).
After preliminary characterization of the changes that occur in
the initial set of analyses, a second phase of the PASS will
include mechanistic studies of exercise-induced molecules
that transduce stress resistance and circulating factors that
might be implicated in the health benefits of exercise. Additional
studies will focus on the adaptation to chronic resistance exer-
cise and the impact of age and sex on these responses, as
well as other studies that have yet to be determined.
Human Clinical Exercise Sites
The human component of MoTrPAC is an in-depth study of the
effects of two different forms of exercise (endurance and resis-
tance training) across multiple individuals of different ages

1466 Cell 181, June 25, 2020


Perspective
(including children) and sexes, as well as sedentary and highly
active individuals. This large cohort will be used to study the
response to exercise at the whole body and cellular levels and
attempt to identify the molecular underpinnings that might be
responsible for the adaptive process and variation among indi-
viduals. Several traditional methods from the field of exercise
physiology will be combined with novel biospecimen sampling
and high-throughput molecular analytical approaches that will
likely yield important insights into the effects of exercise on
health. The human study has many unique aspects that are high-
lighted below and will be conducted as a randomized controlled
trial (RCT) with an intent-to-treat design.
Participants
The goal is to recruit 270 children and adolescents (10–17 years
of age) who are low-active in endurance-type exercise and 1,980
healthy sedentary adults (age 18 years or greater) who will be
medically screened and randomly assigned to endurance
training (170 youth, 840 adults), resistance training (840 adults),
or non-exercise control (50 youth, 300 adults) (Figure 2). An addi-
tional group of highly active endurance- (50 youth, 150 adults)
and resistance- (150 adults) trained individuals will serve as com-
parators and will not participate in the MoTrPAC exercise training
programs. The recruitment and enrollment approach will be sex
balanced and provide participants across a wide range of ages
(10–17, 18–39, 40–59 and R60 year age groups) and of different
races.
Exercise Training Program: Adults
The sedentary adult participants randomized to endurance or
resistance exercise training will perform 12 weeks of supervised
exercise, 3 days per week, with progression in both volume and
intensity. Each endurance training session will be 1 h in dura-
tion and be evenly split between cycling and treadmill (walking/
running) exercise with intensity set to 60%–80% of heart rate
reserve and monitored in real time during each session. Each
resistance training session will target the whole body and consist
of eight total exercises (five upper body: chest press, military
press, seated row, triceps extension, biceps curl; three lower
body: leg press, leg curl, knee extension) at a prescribed plan
of 3 sets of 8–12 repetitions at an intensity of 60%–80% of
maximum for each exercise. These exercise protocols are well
known to improve clinically relevant parameters (i.e., VO2max
and muscular strength and hypertrophy) via alterations in meta-
bolic, biochemical, and molecular signatures (Coggan et al.,
1990; Gollnick et al., 1973; Raue et al., 2012; Rönn et al., 2014;
Timmons et al., 2010).
Acute Exercise Bout and Biospecimens Collections:
Adults
A unique feature of the MoTrPAC adult protocol will be the inte-
gration of strategic biospecimen collections (blood, muscle, and
adipose) before, during, and after standardized bouts of acute
exercise. Participants will perform a 40 to 45-min bout of exer-
cise (exercise-mode specific; rest for the non-exercise controls)
with biospecimens collected before and after 12 weeks of
training. The highly active group will perform the exercise-
mode specific bout only once. Compared to resting homeosta-
sis, these types of exercise challenges are expected to dramat-
ically increase metabolic rate 5- to 10-fold (Coggan et al., 1990;
Farinatti and Castinheiras Neto, 2011; Mulla et al., 2000; Romijn
ll
et al., 1993), bioenergetic flux >10-fold (Kjaer et al., 1991; Romijn
et al., 1993; Steensberg et al., 2000), and large dynamic range in
gene expression from small to >100-fold changes (Louis et al.,
2007; Radom-Aizik et al., 2013, 2014) and likely enhance
cross-talk among many organs (Pedersen and Febbraio, 2012).
Standardized conditions that control for physical activity,
time of day, and dietary intake will be implemented prior to
the acute exercise bout. On the day of an acute exercise bout
with biospecimen collections, volunteers will arrive at a Human
Clinical Center in the morning after an overnight fast and rest
comfortably for 0.5 h prior to obtaining baseline blood (antecu-
bital vein), skeletal muscle (vastus lateralis), and adipose (peri-
umbilical region) samples. Participants will then perform the
standardized acute exercise bout (or rest for the non-exercise
control group) with additional biospecimen samples (blood,
muscle, adipose) obtained 0.5 h (early), 4 h (middle), and
24 h (late) after exercise. These time points were chosen to
capture the dynamic changes in the response to exercise as
metabolic, post-translational, and epigenetic modifications
can occur quite rapidly (Barrès et al., 2012; Bolster et al.,
2003; Hoffman et al., 2015; Romijn et al., 1993), whereas
mRNA induction generally peaks a few hours after exercise
(Louis et al., 2007; Yang et al., 2005), and increases in protein
synthesis rates are detectable in the hours and days following
exercise (Phillips et al., 1997) (Figure 3). Additional blood sam-
ples will be collected during the endurance exercise bout (20-
and 40-min time points) and shortly after (10 min) both endur-
ance and resistance exercise bouts. All participants will have
pre-exercise biospecimen collections, but to reduce partici-
pant burden in the post-exercise phase, sedentary participants
will undergo skeletal muscle and adipose biopsies at one of
three time points (early, middle, late). The highly active partici-
pants will have muscle biopsies and blood at all time points,
whereas adipose biopsies will be collected at the pre and mid-
dle time points.
Acute Exercise Bout and Biospecimens Collections:
Pediatrics
Children and adolescents undergo critical periods of growth and
development, which are distinct from adult physiology. Pediatric
studies must also comply with additional ethical considerations
(Radom-Aizik and Cooper, 2016). Consequently, although the
pediatric arm of the study will mimic the adult protocol as closely
as possible, there are a few notable exceptions: (1) children who
are low-active in endurance-type exercise will be recruited
(versus sedentary adults) to account for the fact that children
are naturally more active than adults and also participate in
mandatory physical education classes; (2) no tissue biopsies
(only blood will be collected); (3) an acute bout of endurance ex-
ercise with blood collection will be performed in both the training
intervention and no-exercise control groups; (4) blood samples
will be collected in all participants before, 20 and 40 min during
exercise, and 10 min, 0.5 h, and 3.5 h into recovery; and (5) for
a subgroup of 170 low-active endurance exercise children and
adolescents who will be randomized to receive 12-week endur-
ance training (N = 120) or continue their standard practice (N =
50) (Figure 2), the endurance exercise intervention will be modi-
fied to provide an exercise intervention that is appropriate to the
pediatric participants’ age group. For middle and high school
Cell 181, June 25, 2020 1467
 ll
students the modes of endurance exercise will include cycling
and treadmill and also an option for elliptical and rowing ma-
chines. Elementary school students will be trained in a form of
circuit training (e.g., endurance activity stations: cycle ergom-
eter, steppers, individual jump rope, pacer, and sliders) to
keep the younger participants engaged.
1468 Cell 181, June 25, 2020
Perspective
Figure 3. Overview of Biospecimen Collec-
tions and Integrated Analysis Plan
Preclinical studies: For the acute exercise arm,
biospecimens (N = 25 per animal) are being
collected 0 (i.e., immediately post), 0.5, 1, 4, 7, 24,
or 48 h following exercise. For the chronic exercise
training arm, which has been completed, bio-
specimens (N = 28) were collected 48 h after rats
completed 1, 2, 4, or 8 weeks of treadmill training.
For both study arms, the same biospecimens are
being collected from an unexercised group of rats.
Clinical Studies: Human biospecimens (blood,
muscle, adipose; blood only for pediatrics) will be
obtained before and at 0.5, 4, and 24 h following
endurance or resistance exercise. The sedentary
participants will perform the acute exercise bout
with biospecimen collections twice (before and
after 12 weeks of training) whereas the highly
active participants will perform this one time (see
Human Clinical Sites description in the text for
more detail). The ‘omics data generated from
multiple different genomics, proteomics, and me-
tabolomics assays will be processed in assay-
specific pipelines. Omic-specific analyses fol-
lowed by state-of-the-art integrative methods will
be applied for a multi-omic analysis of the multiple
time points and tissues collected in MoTrPAC with
the goal of creating a map of the molecular
transducers of exercise. All data (pipelines, raw
and processed data, results, and integrative ana-
lyses) will be made publicly available through the
MoTrPAC Data Hub: https://motrpac-data.org.
Figure design by Jill K. Gregory, Mount Sinai.
Phenotype Measures: Pediatrics
and Adults
To complement the molecular map,
selected phenotypic measures will be ob-
tained. Participants will be assessed
before and after the 12-week intervention
period, whereas the highly active adult
comparator group members will be
assessed once. These measurements
include maximal oxygen consumption
on a cycle ergometer (VO2max), grip
strength, maximal isometric knee exten-
sion strength, body composition (DXA,
dual-energy X-ray absorptiometry), clin-
ical blood profiles, heart rate profiles dur-
ing the acute endurance and resistance
exercise bouts, substrate utilization (car-
bohydrate and fat) during the acute
endurance exercise bout at 65%
VO2max (adult endurance participants
only), and upper and lower body strength
(one-repetition maximum; adult resis-
tance participants only). Most of the adult
participants also will provide information on self-reported health
outcomes using PROMIS measures (www.nihpromis.org) that
will provide opportunities to investigate the effect of exercise
on mood, anxiety, and depression. In a subgroup of adult partic-
ipants, skeletal muscle and adipose histology (cell type, size,
capillarization) and single cell analysis across various ‘omics

Perspective
platforms will be conducted. Investigators are exploring the po-
tential for collecting and analyzing microbiome samples from a
subgroup of adult participants.
Implementation
To optimize this complex protocol, the adult component of the
study will be implemented in two phases. The first phase will
involve 150 adult participants and will require 4–6 months,
enabling assessment of participant and clinical burden and
feasibility, as well as allowing for refinement of the MoTrPAC pro-
tocol. In phase two, the remainder of the project with a target of
over 2,000 participants will be implemented.
Consortium Coordinating Center
The Consortium Coordinating Center (CCC) is composed of four
parts: an Administrative Coordinating Center (ACC), a Data Man-
agement and Quality Control (DMAQC) Core, an Exercise Inter-
vention Core (EIC), and a central Biorepository. The role of the
ACC is to enable the organization and governance of MoTrPAC
by facilitating key processes such as meeting logistics, IRB sub-
mission, and preparation of Manuals of Operations.
The Biorepository, working with the preclinical and clinical
sites, the DMAQC, and the Chemical Analysis Sites, oversees
sample collection, shipping, archiving, and distribution of human
and animal samples. This includes ensuring that homogeneous
cryo-pulverization of tissue samples occurs prior to distribution
of aliquots to the various Chemical Analysis Sites. Uniform sam-
ple processing is important to ensure that diverse data types can
be directly compared. Each tissue sample will also be stored for
future use by MoTrPAC and non-MoTrPAC investigators. Sam-
ples include serum, EDTA plasma, PAXgene-protected whole
blood and peripheral blood mononuclear cells, and vastus later-
alis skeletal muscle and subcutaneous abdominal adipose tissue
from humans and >20 different tissues from the preclinical ani-
mals. Each sample will be analyzed by the Chemical Analysis
Sites, and additional material will be archived for future use.
The Biorepository inventory system interacts with the DMAQC
to enable sample tracking, quality control, and other process
support systems.
Chemical Analysis Sites
To understand the exercise response in detail, an in-depth
analysis of molecular and ‘omic assays will be performed using
state-of-the-art laboratory techniques. Technologies include ge-
nomics, transcriptomics, DNA methylomics, targeted and untar-
geted proteomics, and targeted and untargeted metabolomics.
Genomic, Transcriptomic, and Regulatory Analyses
Evidence has shown, through more than 150 small cohort
studies (typically with under 50 participants analyzed) (Bouchard
et al., 2011; Pacheco et al., 2018), that exercise is accompanied
with massive changes at both the transcriptional and epige-
nomic levels in muscle, adipose, and most other tissue systems
(Lindholm et al., 2014; Ling and Rönn, 2014; Rönn and Ling,
2013) with the poorly understood influence of the underlying hu-
man genetic/environmental variation that exists between and
within populations (Leon ska-Duniec et al., 2016). Therefore,
recent scientific studies have been conducted generating data
reflecting some of the underlying genetic and epigenetic basis
for responses to exercise, physical activity, and training linking
ll
specific targets to benefits (Carter et al., 2017). Although several
studies (for review see Loos et al., 2015; Warburton et al., 2006)
have provided a rich source of information to develop a founda-
tion for larger and more comprehensive genomic, epigenomic,
and transcriptomic (GET) analyses, sufficiently powered studies
with the complementary detailed study design to make accurate
predictions as machine-learned models remain underdevel-
oped. Moreover, the information needed to understand the role
genetic variation plays in the response of individuals to acute
and chronic exercise remains limited. MoTrPAC, although pre-
dicted to be statistically underpowered for a genome-wide asso-
ciation study (GWAS), should be able to be statistically organized
and prioritized (Cantor et al., 2010) so that there is potential
benefit from the orthogonal measurements assessed through
other ‘omes, leading to improved mechanistic insight. Such in-
formation can begin as a knowledge base for enabling better
treatment considerations for a variety of diseases (whether acute
or chronic) through recognizing potential genetic and epigenetic
differences in responses to exercise and training. This could be
accomplished through identifying novel gene/genetic network
involvement, their corresponding changes in RNA transcripts
and how such genes are regulated at the epigenetic level from
adult and adolescent and between athletic and sedentary indi-
viduals, and associated sex differences in response to acute
and chronic exercise.
The goal of the GET assays are to map and measure changes
in the (1) RNA transcriptome and transcript isoforms including
small and micro RNA using RNA sequencing, (2) DNA methyl-
ation and chromatin accessibility from rat and human tissues us-
ing reduced representation bisulfite (RRBS) for rat or methyl CpG
hybrid capture for human specimens and ATAC-seq (assay for
transposase-accessible chromatin with sequencing), respec-
tively, and (3) genomic sequence and structure of all human par-
ticipants. The assays are expected to provide insights into
changes in biological processes as well as gene regulatory net-
works that occur in response to acute and chronic exercise. The
GET assay component of MoTrPAC will involve comprehensive
analyses of extensively curated rat and human MoTrPAC sam-
ples with an exercise intervention, contribute these data to public
databases, help identify candidate molecular transducers, eluci-
date new mechanisms that might explain the human response to
exercise, and cooperate with the Bioinformatics Center (BIC) to
develop predictive models of the individual response to physical
activity.
Proteomics
Proteins are important drivers of cellular structure, function, and
signal mediation (Cox and Mann, 2011); thus, uncovering the
pathways through which physical activity influences health re-
quires analysis of the proteome and the critical signaling-associ-
ated post-translational modifications of the proteome in various
tissues. To date, a number of proteomic studies have shown
important changes influenced by exercise (Burniston, 2008; Hoff-
man et al., 2015; Magherini et al., 2012; Sollanek et al., 2017).
The majority of this work has focused on skeletal muscle, which
is the tissue that actively performs the motions involved in exer-
cise, and blood and plasma, which circulate signals systemically
through the body and may be responsible for facilitating cross-
talk between organ systems. Furthermore, this research is often
Cell 181, June 25, 2020 1469
 ll
performed in the context of diabetes because of the role of muscle
and the interplay of exercise with insulin-resistance (Kleinert et al.,
2018). Although these studies have largely been constrained to
experimental models of exercise in animals or very small cohorts
of human subjects, the results are tantalizing and have identified
several proteins and signaling molecules that potentially play a
key role in the response to exercise. The large-scale and well-
controlled preclinical and clinical protocols adopted for MoTrPAC
will allow for expansion of this knowledge by providing a deeper
interrogation of the proteomic response to acute and chronic ex-
ercise in numerous tissues from individuals across a range of
fitness levels.
Importantly, proteomic analyses should be inclusive of not
only protein expression but also the state of protein post-trans-
lational modifications, such as phosphorylation or acetylation,
because these chemical moieties can act as rapid integrators
by dictating protein localization and enzymatic activity (Brandes
et al., 2009; Choudhary et al., 2014; Emmerich et al., 2011; Hunt-
er, 1995). Primarily, untargeted mass spectrometry methods and
targeted aptamer-based detection techniques will be employed
to probe changes in protein abundance and modifications
induced by exercise. Given that distinct tissues present techno-
logical challenges to discovery-based proteomic analysis (e.g.,
dynamic range in skeletal muscle), state-of-the-art instrumenta-
tion and protocols, including tandem mass tag labeling and frac-
tionation (Mertins et al., 2018), will be employed. Indeed, pilot
discovery-based proteomics efforts with muscle and other
tissues have yielded robust datasets with levels of protein
coverage exceeding previous studies, presenting a wealth of
opportunities to elucidate proteomic response to exercise and
integrate these findings with data obtained from GET and metab-
olomic studies of the same tissues.
Metabolomics
Complementing genomics, transcriptomics, epigenomics, and
proteomic studies, MoTrPAC will also carry out a highly compre-
hensive mapping of exercise-associated alterations in the me-
tabolome of both rats and humans. The metabolome is the total
collection of biologically active small molecules in a given organ-
ism (Nicholson and Wilson, 2003). This includes endogenous
molecules that are biosynthesized by metabolic networks in pri-
mary metabolism, molecules derived from diet or environmental
exposures (the exposome; Wild, 2005), and molecules derived
from the biosynthetic interactions with the microbiome. Metabo-
lomics can either be ‘‘targeted’’ to a set of known compounds
(e.g., certain acylcarnitines) or ‘‘non-targeted,’’ which attempts
to detect and relatively quantify as many metabolites as possible
(Dettmer et al., 2007). In the context of acute and chronic exer-
cise, metabolomics can provide sensitive and dynamic pheno-
typic patterns that closely reflect cellular and molecular changes
and will likely improve our understanding of the effects of exer-
cise beyond the individual pathway level (Heaney et al., 2017).
A number of studies have documented profound metabolomic
alterations associated with exercise (Fukai et al., 2016; Heaney
et al., 2017; Lewis et al., 2010; Xiao et al., 2016), but these typi-
cally involve smaller cohorts (n < 100), are limited to only one (or a
handful of) metabolomics assays, or focus primarily on alter-
ations in energy production pathways. The vast chemical diver-
sity of the metabolome, which includes lipids, sugars, amino
1470 Cell 181, June 25, 2020
Perspective
acids, and myriad other molecule types, and its wide dynamic
range (sub-nM to mM) implies that no single chemical assay
can adequately profile all metabolites in one experiment (Smilde
et al., 2005). To this end, MoTrPAC will employ a combination of
non-targeted and targeted approaches for mapping the broader
effects of exercise on both the metabolome and lipidome. These
will range from triple-quadrupole-based liquid chromatography-
mass spectrometry (LC-MS) using stable isotope-labeled inter-
nal standards for absolute quantification to high resolution MS
and tandem MS using reversed phase and hydrophilic interac-
tion LC for mapping relative changes in both known and
unknown molecular transducers. Targeted and non-targeted
LC-MS assays that focus on the non-polar fraction of the metab-
olome (the lipidome) will also be leveraged to map exercise ef-
fects on lipid metabolism and oxidation (Nieman et al., 2013,
2014). It is expected that these studies will provide insights
into energy metabolism and signaling molecules involved in the
response to exercise.
Exosomes
Exercise is a potent stimulus that has broad ranging systemic ef-
fects that are indubitable (Egan and Zierath, 2013). One prevail-
ing hypothesis is that circulating extracellular vesicles termed
exosomes play an important role in carrying training-induced
protein, mRNA, and microRNA (miRNA) cargo between organs
as a means of integrating responses to exercise (Safdar and Tar-
nopolsky, 2018) (Figure 4). Many techniques have been
developed for isolating exosomes from plasma to analyze their
molecular cargo (Barrachina et al., 2019). Importantly, these
techniques have been used to demonstrate that acute exercise
increases the abundance of a wide variety of exosome-associ-
ated proteins related to metabolic and immune regulation
(Whitham et al., 2018). Exosome isolation and analysis of
MoTrPAC specimens will further investigate these effects by
describing how exosome content is modulated in response to
endurance and resistance exercise. The identification of protein
and RNA signatures associated with exercise will shed light on
exosome-mediated inter-organ cross-talk and provide a frame-
work for studies to characterize the systemic response to phys-
ical activity.
Bioinformatics Center
The immediate goals of MoTrPAC will be vested in the ‘omic plat-
forms used and the data being generated, the quality of this data,
both meta and experimental, and how it will be utilized to map
the molecular transducers involving the responses to acute
and chronic exercise. Data from each assay will be collected at
the BIC and analyzed using consistent bioinformatic and analytic
pipelines, whenever possible. This will improve reproducibility,
interpretability, and ease in data harmonization across sites.
Assay-specific quantitative data will undergo quality control
assessment and be normalized to reduce undesirable sample-
to-sample variation, minimize batch effects, and deal with ana-
lyte heteroskedasticity typically observed in molecular abun-
dance datasets. Relative levels of analytes will be determined
and the changes in molecules and pathways in response to ex-
ercise deduced.
Investigators across the consortium will conduct a series of
integrative analyses with the end goal of creating a map of

Perspective
Figure 4. Role of Exosomes in Integrating the Exercise Response
across Organ Systems
Exosomes are small extracellular vesicles (EVs) that are packaged with func-
tional molecular cargo and released from most cell types. Exosomes released
in response to exercise can be transported in the blood to various other tissues
(adipose, brain, liver, etc.) where their contents can be released and have
biological impact. Protein and miRNA cargo of circulating EVs from MoTrPAC
samples will be analyzed to provide additional insight into this emerging bio-
logical phenomenon.
Figure design by Jill K. Gregory, Mount Sinai.
molecular transducers of physical activity and understanding the
dynamic biological changes that occur in response to acute and
chronic exercise (Figure 3). Unbiased statistical analysis will be
performed to understand the variability in the data, thereby
ll
revealing significant differential responses possibly interacting
with other clinical variables (e.g., sex, age, and VO2max). Sys-
tems biology methods and network algorithms will be used to
correlate the different ‘omics and provide interpretable modules,
focusing on regulatory processes. This will include basic inter-
pretation of differential abundance patterns via network and
pathway enrichment analyses. Additional analysis will be per-
formed to understand the correlation structure between ‘omics,
detecting latent subject classes with different temporal re-
sponses to exercise, and network propagation techniques to
highlight tissue-specific response (Amar and Shamir, 2014;
Amar et al., 2015; Cowen et al., 2017; Gallant et al., 2013; Hofree
et al., 2013; Jo et al., 2016; Schulz et al., 2012). In addition to
identifying differential molecular responses and their associa-
tions with physiological changes (e.g., molecules and pathways
that are associated with VO2max), MoTrPAC will attempt to build
predictive models of the effects of exercise on change in these
parameters. Plans are in place to study the effects of age, sex,
race, and exercise type on these changes. To enhance interpret-
ability, the results from all analyses above will be summarized us-
ing multiple visualization techniques.
Data and Resources Dissemination
As an NIH Common Fund Program, MoTrPAC aims to provide a
foundation for further research to be used by the broad biomedical
research community. The goal is to share all MoTrPAC data that
does not compromise personal health information (PHI) for gen-
eral research use. All of the produced raw and processed data,
analysis pipelines, and results will be made rapidly available to
the scientific community through the MoTrPAC Data Hub
(https://motrpac-data.org) and, whenever suitable, in the appro-
priate public repositories (e.g., Metabolomics Workbench for me-
tabolomics data and Database of Genotypes and Phenotypes for
sequencing data). We plan to follow the FAIR (findable, acces-
sible, interoperable, and reusable) data principles (Wilkinson
et al., 2016) to ensure that the data are widely available to and us-
able by the scientific community. In addition, we will be collabo-
rating with the Common Fund Data Ecosystem (https://
commonfund.nih.gov/dataecosystem) team on tools, techniques,
and external user training to maximize the utility of the MoTrPAC
data to the scientific community. The MoTrPAC Data Hub itself
is a cloud-native application which uses a service-oriented archi-
tecture. It is being hosted on the Google Cloud Platform. In addi-
tion to providing direct access to data, there will be tools for cus-
tomizable web-based data visualizations provided on the site.
Scientific Opportunities/Ancillary Studies
Although many activities are occurring within MoTrPAC, there
will be opportunities to expand the breadth and depth of
MoTrPAC. These opportunities include (1) additional interroga-
tion of clinical metadata derived from animal and human studies
and through the ‘omics applications and the interpretation of the
data beyond what has been proposed and approved through the
main MoTrPAC study, (2) addition or expansion of data collection
to the current parent protocol, and (3) accession of the bio-
repository for remaining biospecimen samples obtained from
the animal and human studies to perform complementary ana-
lyses on these tissues. The MoTrPAC Ancillary Study policy
Cell 181, June 25, 2020 1471
 ll
and proposal template are available at https://motrpac.org/
ancillarystudyguidelines.cfm.
Study Challenges
There are many challenges associated with a large multicenter
project generating a wide variety of data types. Standardization
procedures, operation manuals, and quality control steps are in
place to reduce the variation in exercise performance and eval-
uation and sample collections of the animal and human samples.
Participant and clinic burden are a concern given the large scope
and complexity of MoTrPAC; the consortium has made strategic
choices regarding the protocol design and biospecimens sam-
pling to help mitigate these challenges. Similarly, standardization
and quality control steps are in place for the data generated us-
ing multiple ‘omics platforms even for similar data types (e.g.,
metabolomics and proteomics). An initial implementation phase
(described above) was also put in place for the human studies to
further evaluate the adult protocol during the early stages of
recruitment, data collection, and analysis to identify any unfore-
seen issues.
Integrating heterogeneous data types across MoTrPAC will
require sophisticated tracking, data normalization, and analytic
approaches. MoTrPAC will generate large amounts of data, and
many measurements may not meet statistical significance at the
individual molecule level but may do so at the pathway level.
State-of-the art analytic tools are in place to help manage the
depth and breadth of data analysis, and these tools will continue
to evolve as MoTrPAC progresses. Finally, the data are complex;
ensuring that they are accessible and understandable to the broad
scientific community in a timely fashion is essential for this project
to be successful. Protection of participant clinical data (PHI, geno-
mics data) will be given the highest priority to protect each individ-
ual’s identity. Incorporation of useful visualization tools as well as
active engagement with the broader scientific community will be
equally important to fully capitalize on the MoTrPAC project.
Summary
When complete, MoTrPAC will deliver a map of the biological
molecules and pathways underlying the systemic effects of
acute and chronic exercise. The data, which will ultimately be
made freely available to the scientific community, will provide
unprecedented opportunities to begin to understand the path-
ways by which physical activity influences health. In the future,
it is expected that the knowledge gained will allow researchers
and health professionals to develop personalized exercise rec-
ommendations and provide insights into molecular targets that
could be manipulated to mimic some of the effects of exercise
in persons unable to do so.
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.06.004.
ACKNOWLEDGMENTS
The authors would like to gratefully acknowledge Jill K. Gregory, CMI, FAMI
(Certified Medical Illustrator) of the Icahn School of Medicine at Mount Sinai
for working with the writing group to generate the figures enclosed within
1472 Cell 181, June 25, 2020
Perspective
this article. Additionally, the authors would like to gratefully acknowledge the
expert administrative functions by Heather Kiesel of the University of
Florida for help organizing efforts by the MoTrPAC Writing Group to better
enable the completion of this article. The MoTrPAC Study is supported by
NIH grants U24OD026629 (Bioinformatics Center), U24DK112349,
U24DK112342, U24DK112340, U24DK112341, U24DK112326,
U24DK112331, U24DK112348 (Chemical Analysis Sites), U01AR071133,
U01AR071130, U01AR071124-01, U01AR071128, U01AR071150,
U01AR071160, U01AR071158 (Clinical Centers), U24AR071113 (Consortium
Coordinating Center), U01AG055133, U01AG055137, and U01AG055135
(PASS/Animal Sites). The views expressed are those of the authors and do
not necessarily reflect those of the NIH or the Department of Health and Hu-
man Services of the United States.
REFERENCES
Amar, D., and Shamir, R. (2014). Constructing module maps for integrated
analysis of heterogeneous biological networks. Nucleic Acids Res. 42,
4208–4219.
Amar, D., Yekutieli, D., Maron-Katz, A., Hendler, T., and Shamir, R. (2015). A
hierarchical Bayesian model for flexible module discovery in three-way time-
series data. Bioinformatics 31, i17–i26.
Barrachina, M.N., Calderón-Cruz, B., Fernandez-Rocca, L., and Garcı́a, Á.
(2019). Application of Extracellular Vesicles Proteomics to Cardiovascular Dis-
ease: Guidelines, Data Analysis, and Future Perspectives. Proteomics 19,
e1800247.
Barrès, R., Yan, J., Egan, B., Treebak, J.T., Rasmussen, M., Fritz, T., Caidahl,
K., Krook, A., O’Gorman, D.J., and Zierath, J.R. (2012). Acute exercise re-
models promoter methylation in human skeletal muscle. Cell Metab. 15,
405–411.
Bolster, D.R., Kubica, N., Crozier, S.J., Williamson, D.L., Farrell, P.A., Kimball,
S.R., and Jefferson, L.S. (2003). Immediate response of mammalian target of
rapamycin (mTOR)-mediated signalling following acute resistance exercise in
rat skeletal muscle. J. Physiol. 553, 213–220.
Booth, F.W., Roberts, C.K., Thyfault, J.P., Ruegsegger, G.N., and Toede-
busch, R.G. (2017). Role of Inactivity in Chronic Diseases: Evolutionary Insight
and Pathophysiological Mechanisms. Physiol. Rev. 97, 1351–1402.
Bouchard, C., Rankinen, T., and Timmons, J.A. (2011). Genomics and genetics
in the biology of adaptation to exercise. Compr. Physiol. 1, 1603–1648.
Brandes, N., Schmitt, S., and Jakob, U. (2009). Thiol-based redox switches in
eukaryotic proteins. Antioxid. Redox Signal. 11, 997–1014.
Burniston, J.G. (2008). Changes in the rat skeletal muscle proteome induced
by moderate-intensity endurance exercise. Biochim. Biophys. Acta 1784,
1077–1086.
Cantor, R.M., Lange, K., and Sinsheimer, J.S. (2010). Prioritizing GWAS re-
sults: A review of statistical methods and recommendations for their applica-
tion. Am. J. Hum. Genet. 86, 6–22.
Carter, A.C., Chang, H.Y., Church, G., Dombkowski, A., Ecker, J.R., Gil, E.,
Giresi, P.G., Greely, H., Greenleaf, W.J., Hacohen, N., et al. (2017). Challenges
and recommendations for epigenomics in precision health. Nat. Biotechnol.
35, 1128–1132.
Choudhary, C., Weinert, B.T., Nishida, Y., Verdin, E., and Mann, M. (2014). The
growing landscape of lysine acetylation links metabolism and cell signalling.
Nat. Rev. Mol. Cell Biol. 15, 536–550.
Coggan, A.R., Kohrt, W.M., Spina, R.J., Bier, D.M., and Holloszy, J.O. (1990).
Endurance training decreases plasma glucose turnover and oxidation during
moderate-intensity exercise in men. J. Appl. Physiol. 68, 990–996.
Cowen, L., Ideker, T., Raphael, B.J., and Sharan, R. (2017). Network propaga-
tion: a universal amplifier of genetic associations. Nat. Rev. Genet. 18,
551–562.
Cox, J., and Mann, M. (2011). Quantitative, high-resolution proteomics for
data-driven systems biology. Annu. Rev. Biochem. 80, 273–299.

Perspective
Dettmer, K., Aronov, P.A., and Hammock, B.D. (2007). Mass spectrometry-
based metabolomics. Mass Spectrom. Rev. 26, 51–78.
Egan, B., and Zierath, J.R. (2013). Exercise metabolism and the molecular
regulation of skeletal muscle adaptation. Cell Metab. 17, 162–184.
Emmerich, C.H., Schmukle, A.C., and Walczak, H. (2011). The emerging role of
linear ubiquitination in cell signaling. Sci. Signal. 4, re5.
Farinatti, P.T.V., and Castinheiras Neto, A.G. (2011). The effect of between-set
rest intervals on the oxygen uptake during and after resistance exercise ses-
sions performed with large- and small-muscle mass. J. Strength Cond. Res.
25, 3181–3190.
Fukai, K., Harada, S., Iida, M., Kurihara, A., Takeuchi, A., Kuwabara, K., Su-
giyama, D., Okamura, T., Akiyama, M., Nishiwaki, Y., et al. (2016). Metabolic
Profiling of Total Physical Activity and Sedentary Behavior in Community-
Dwelling Men. PLoS ONE 11, e0164877.
Gallant, A., Leiserson, M.D.M., Kachalov, M., Cowen, L.J., and Hescott, B.J.
(2013). Genecentric: a package to uncover graph-theoretic structure in high-
throughput epistasis data. BMC Bioinformatics 14, 23.
Gollnick, P.D., Armstrong, R.B., Saltin, B., Saubert, C.W., 4th, Sembrowich,
W.L., and Shepherd, R.E. (1973). Effect of training on enzyme activity and fiber
composition of human skeletal muscle. J. Appl. Physiol. 34, 107–111.
Hawley, J.A., Hargreaves, M., Joyner, M.J., and Zierath, J.R. (2014). Integra-
tive biology of exercise. Cell 159, 738–749.
Heaney, L.M., Deighton, K., and Suzuki, T. (2017). Non-targeted metabolomics
in sport and exercise science. J. Sports Sci. 37, 959–967.
Hoffman, N.J., Parker, B.L., Chaudhuri, R., Fisher-Wellman, K.H., Kleinert, M.,
Humphrey, S.J., Yang, P., Holliday, M., Trefely, S., Fazakerley, D.J., et al.
(2015). Global Phosphoproteomic Analysis of Human Skeletal Muscle Reveals
a Network of Exercise-Regulated Kinases and AMPK Substrates. Cell Metab.
22, 922–935.
Hofree, M., Shen, J.P., Carter, H., Gross, A., and Ideker, T. (2013). Network-
based stratification of tumor mutations. Nat. Methods 10, 1108–1115.
Hunter, T. (1995). Protein kinases and phosphatases: the yin and yang of pro-
tein phosphorylation and signaling. Cell 80, 225–236.
Jo, K., Jung, I., Moon, J.H., and Kim, S. (2016). Influence maximization in time
bounded network identifies transcription factors regulating perturbed path-
ways. Bioinformatics 32, i128–i136.
Kjaer, M., Kiens, B., Hargreaves, M., and Richter, E.A. (1991). Influence of
active muscle mass on glucose homeostasis during exercise in humans.
J. Appl. Physiol. 71, 552–557.
Kleinert, M., Parker, B.L., Jensen, T.E., Raun, S.H., Pham, P., Han, X., James,
D.E., Richter, E.A., and Sylow, L. (2018). Quantitative proteomic characteriza-
tion of cellular pathways associated with altered insulin sensitivity in skeletal
muscle following high-fat diet feeding and exercise training. Sci. Rep.
8, 10723.
Leon ska-Duniec, A., Ahmetov, I.I., and Zmijewski, P. (2016). Genetic variants
influencing effectiveness of exercise training programmes in obesity - an over-
view of human studies. Biol. Sport 33, 207–214.
Lewis, G.D., Farrell, L., Wood, M.J., Martinovic, M., Arany, Z., Rowe, G.C.,
Souza, A., Cheng, S., McCabe, E.L., Yang, E., et al. (2010). Metabolic signa-
tures of exercise in human plasma. Sci. Transl. Med. 2, 33ra37.
Lindholm, M.E., Marabita, F., Gomez-Cabrero, D., Rundqvist, H., Ekström,
T.J., Tegnér, J., and Sundberg, C.J. (2014). An integrative analysis reveals co-
ordinated reprogramming of the epigenome and the transcriptome in human
skeletal muscle after training. Epigenetics 9, 1557–1569.
Ling, C., and Rönn, T. (2014). Epigenetic adaptation to regular exercise in hu-
mans. Drug Discov. Today 19, 1015–1018.
Loos, R.J.F., Hagberg, J.M., Pérusse, L., Roth, S.M., Sarzynski, M.A., Wolf-
arth, B., Rankinen, T., and Bouchard, C. (2015). Advances in exercise, fitness,
and performance genomics in 2014. Med. Sci. Sports Exerc. 47, 1105–1112.
Louis, E., Raue, U., Yang, Y., Jemiolo, B., and Trappe, S. (2007). Time course
of proteolytic, cytokine, and myostatin gene expression after acute exercise in
human skeletal muscle. J. Appl. Physiol. 103, 1744–1751.
ll
Magherini, F., Abruzzo, P.M., Puglia, M., Bini, L., Gamberi, T., Esposito, F.,
Veicsteinas, A., Marini, M., Fiorillo, C., Gulisano, M., and Modesti, A. (2012).
Proteomic analysis and protein carbonylation profile in trained and untrained
rat muscles. J. Proteomics 75, 978–992.
Mertins, P., Tang, L.C., Krug, K., Clark, D.J., Gritsenko, M.A., Chen, L.,
Clauser, K.R., Clauss, T.R., Shah, P., Gillette, M.A., et al. (2018). Reproducible
workflow for multiplexed deep-scale proteome and phosphoproteome anal-
ysis of tumor tissues by liquid chromatography-mass spectrometry. Nat. Pro-
toc. 13, 1632–1661.
Mulla, N.A., Simonsen, L., and Bülow, J. (2000). Post-exercise adipose tissue
and skeletal muscle lipid metabolism in humans: the effects of exercise inten-
sity. J. Physiol. 524, 919–928.
Neufer, P.D., Bamman, M.M., Muoio, D.M., Bouchard, C., Cooper, D.M.,
Goodpaster, B.H., Booth, F.W., Kohrt, W.M., Gerszten, R.E., Mattson, M.P.,
et al. (2015). Understanding the Cellular and Molecular Mechanisms of Phys-
ical Activity-Induced Health Benefits. Cell Metab. 22, 4–11.
Nicholson, J.K., and Wilson, I.D. (2003). Opinion: understanding ‘global’ sys-
tems biology: metabonomics and the continuum of metabolism. Nat. Rev.
Drug Discov. 2, 668–676.
Nieman, D.C., Gillitt, N.D., Knab, A.M., Shanely, R.A., Pappan, K.L., Jin, F., and
Lila, M.A. (2013). Influence of a polyphenol-enriched protein powder on exer-
cise-induced inflammation and oxidative stress in athletes: a randomized trial
using a metabolomics approach. PLoS ONE 8, e72215.
Nieman, D.C., Shanely, R.A., Luo, B., Meaney, M.P., Dew, D.A., and Pappan,
K.L. (2014). Metabolomics approach to assessing plasma 13- and 9-hydroxy-
octadecadienoic acid and linoleic acid metabolite responses to 75-km cycling.
Am. J. Physiol. Regul. Integr. Comp. Physiol. 307, R68–R74.
Pacheco, C., Felipe, S.M.D.S., Soares, M.M.D.C., Alves, J.O., Soares, P.M.,
Leal-Cardoso, J.H., Loureiro, A.C.C., Ferraz, A.S.M., de Carvalho, D.P., and
Ceccatto, V.M. (2018). A compendium of physical exercise-related human
genes: an ’omic scale analysis. Biol. Sport 35, 3–11.
Pedersen, B.K., and Febbraio, M.A. (2012). Muscles, exercise and obesity:
skeletal muscle as a secretory organ. Nat. Rev. Endocrinol. 8, 457–465.
Phillips, S.M., Tipton, K.D., Aarsland, A., Wolf, S.E., and Wolfe, R.R. (1997).
Mixed muscle protein synthesis and breakdown after resistance exercise in
humans. Am. J. Physiol. 273, E99–E107.
Radom-Aizik, S., and Cooper, D.M. (2016). Bridging the Gaps: the Promise of
Omics Studies in Pediatric Exercise Research. Pediatr. Exerc. Sci. 28,
194–201.
Radom-Aizik, S., Zaldivar, F., Haddad, F., and Cooper, D.M. (2013). Impact of
brief exercise on peripheral blood NK cell gene and microRNA expression in
young adults. J. Appl. Physiol. 114, 628–636.
Radom-Aizik, S., Zaldivar, F.P., Jr., Haddad, F., and Cooper, D.M. (2014).
Impact of brief exercise on circulating monocyte gene and microRNA expres-
sion: implications for atherosclerotic vascular disease. Brain Behav. Immun.
39, 121–129.
Raue, U., Trappe, T.A., Estrem, S.T., Qian, H.-R., Helvering, L.M., Smith, R.C.,
and Trappe, S. (2012). Transcriptome signature of resistance exercise adapta-
tions: mixed muscle and fiber type specific profiles in young and old adults.
J. Appl. Physiol. 112, 1625–1636.
Romijn, J.A., Coyle, E.F., Sidossis, L.S., Gastaldelli, A., Horowitz, J.F., Endert,
E., and Wolfe, R.R. (1993). Regulation of endogenous fat and carbohydrate
metabolism in relation to exercise intensity and duration. Am. J. Physiol.
265, E380–E391.
Rönn, T., and Ling, C. (2013). Effect of exercise on DNA methylation and meta-
bolism in human adipose tissue and skeletal muscle. Epigenomics 5, 603–605.
Rönn, T., Volkov, P., Tornberg, A., Elgzyri, T., Hansson, O., Eriksson, K.-F.,
Groop, L., and Ling, C. (2014). Extensive changes in the transcriptional profile
of human adipose tissue including genes involved in oxidative phosphorylation
after a 6-month exercise intervention. Acta Physiol. (Oxf.) 211, 188–200.
Safdar, A., and Tarnopolsky, M.A. (2018). Exosomes as Mediators of the Sys-
temic Adaptations to Endurance Exercise. Cold Spring Harb. Perspect. Med.
8, a029827.
Cell 181, June 25, 2020 1473
 ll
Schulz, M.H., Devanny, W.E., Gitter, A., Zhong, S., Ernst, J., and Bar-Joseph,
Z. (2012). DREM 2.0: Improved reconstruction of dynamic regulatory networks
from time-series expression data. BMC Syst. Biol. 6, 104.
Smilde, A.K., van der Werf, M.J., Bijlsma, S., van der Werff-van der Vat, B.J.C.,
and Jellema, R.H. (2005). Fusion of mass spectrometry-based metabolomics
data. Anal. Chem. 77, 6729–6736.
Sollanek, K.J., Burniston, J.G., Kavazis, A.N., Morton, A.B., Wiggs, M.P., Ahn,
B., Smuder, A.J., and Powers, S.K. (2017). Global Proteome Changes in the
Rat Diaphragm Induced by Endurance Exercise Training. PLoS ONE 12,
e0171007.
Stanford, K.I., and Goodyear, L.J. (2018). Muscle-Adipose Tissue Cross Talk.
Cold Spring Harb. Perspect. Med. 8, a029801.
Steensberg, A., van Hall, G., Osada, T., Sacchetti, M., Saltin, B., and Klarlund
Pedersen, B. (2000). Production of interleukin-6 in contracting human skeletal
muscles can account for the exercise-induced increase in plasma interleukin-
6. J. Physiol. 529, 237–242.
Timmons, J.A., Knudsen, S., Rankinen, T., Koch, L.G., Sarzynski, M., Jensen,
T., Keller, P., Scheele, C., Vollaard, N.B.J., Nielsen, S., et al. (2010). Using mo-
lecular classification to predict gains in maximal aerobic capacity following
endurance exercise training in humans. J. Appl. Physiol. 108, 1487–1496.
1474 Cell 181, June 25, 2020
Perspective
Warburton, D.E.R., Nicol, C.W., and Bredin, S.S.D. (2006). Health benefits of
physical activity: the evidence. CMAJ 174, 801–809.
Whitham, M., Parker, B.L., Friedrichsen, M., Hingst, J.R., Hjorth, M., Hughes,
W.E., Egan, C.L., Cron, L., Watt, K.I., Kuchel, R.P., et al. (2018). Extracellular
Vesicles Provide a Means for Tissue Crosstalk during Exercise. Cell Metab.
27, 237–251.e4.
Wild, C.P. (2005). Complementing the genome with an ‘‘exposome’’: the
outstanding challenge of environmental exposure measurement in molecular
epidemiology. Cancer Epidemiol. Biomarkers Prev. 14, 1847–1850.
Wilkinson, M.D., Dumontier, M., Aalbersberg, I.J.J., Appleton, G., Axton, M.,
Baak, A., Blomberg, N., Boiten, J.-W., da Silva Santos, L.B., Bourne, P.E.,
et al. (2016). The FAIR Guiding Principles for scientific data management
and stewardship. Sci. Data 3, 160018.
Xiao, Q., Moore, S.C., Keadle, S.K., Xiang, Y.-B., Zheng, W., Peters, T.M.,
Leitzmann, M.F., Ji, B.-T., Sampson, J.N., Shu, X.-O., and Matthews, C.E.
(2016). Objectively measured physical activity and plasma metabolomics in
the Shanghai Physical Activity Study. Int. J. Epidemiol. 45, 1433–1444.
Yang, Y., Creer, A., Jemiolo, B., and Trappe, S. (2005). Time course of
myogenic and metabolic gene expression in response to acute exercise in hu-
man skeletal muscle. J. Appl. Physiol. 98, 1745–1752.
 Article

Host-Viral Infection Maps Reveal Signatures of

Severe COVID-19 Patients
Graphical Abstract Authors
Pierre Bost, Amir Giladi, Yang Liu, ...,
Benno Schwikowski, Zheng Zhang,
Ido Amit

Correspondence
benno@pasteur.fr (B.S.),
ido.amit@weizmann.ac.il (I.A.),
zhangzheng1975@aliyun.com (Z.Z.)

In Brief
A computational framework that allows
for the identification and characterization
of virus-infected cells as well as
bystander cell responses reveals how
SARS-CoV-2 alters the immune
responses of patients.

361667513

Highlights
d Viral-Track: a computational framework to analyze host-viral
infection maps

d Viral-Track sorts infected from bystander cells and reveals

virus-induced expression

d SARS-CoV-2 infects epithelial cells and alters immune

landscape in severe patients

d Co-infection of SARS-Cov-2 and hMPV affects monocytes

and dampens interferon response

Bost et al., 2020, Cell 181, 1475–1488

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.006 ll
 ll

Article
Host-Viral Infection Maps Reveal Signatures
of Severe COVID-19 Patients
Pierre Bost,1,2,3,6 Amir Giladi,1,6 Yang Liu,4,6 Yanis Bendjelal,2 Gang Xu,4 Eyal David,1 Ronnie Blecher-Gonen,1
Merav Cohen,1 Chiara Medaglia,1 Hanjie Li,1 Aleksandra Deczkowska,1 Shuye Zhang,5 Benno Schwikowski,2,*
Zheng Zhang,4,* and Ido Amit1,7,*
1Department of Immunology, Weizmann Institute of Science, Rehovot, Israel
2Systems Biology Group, Department of Computational Biology and USR 3756, Institut Pasteur and CNRS, Paris 75015, France
3Sorbonne Universite, Complexite du vivant, Paris 75005, France
4Institute for Hepatology, National Clinical Research Center for Infectious Disease, Shenzhen Third People’s Hospital, School of Medicine,

Southern University of Science and Technology, Shenzhen 518112, Guangdong Province, China
5Shanghai Public Health Clinical Center and Institute of Biomedical Sciences, Fudan University, Shanghai 201508, China
6These authors contributed equally
7Lead Contact

*Correspondence: benno@pasteur.fr (B.S.), zhangzheng1975@aliyun.com (Z.Z.), ido.amit@weizmann.ac.il (I.A.)

https://doi.org/10.1016/j.cell.2020.05.006

SUMMARY

Viruses are a constant threat to global health as highlighted by the current COVID-19 pandemic. Currently,
lack of data underlying how the human host interacts with viruses, including the SARS-CoV-2 virus, limits
effective therapeutic intervention. We introduce Viral-Track, a computational method that globally scans un-
mapped single-cell RNA sequencing (scRNA-seq) data for the presence of viral RNA, enabling transcriptional
cell sorting of infected versus bystander cells. We demonstrate the sensitivity and specificity of Viral-Track to
systematically detect viruses from multiple models of infection, including hepatitis B virus, in an unsuper-
vised manner. Applying Viral-Track to bronchoalveloar-lavage samples from severe and mild COVID-19 pa-
tients reveals a dramatic impact of the virus on the immune system of severe patients compared to mild
cases. Viral-Track detects an unexpected co-infection of the human metapneumovirus, present mainly in
monocytes perturbed in type-I interferon (IFN)-signaling. Viral-Track provides a robust technology for dis-
secting the mechanisms of viral-infection and pathology.

INTRODUCTION
The development of efficient vaccines against viral pathogens is
considered one of the biggest achievements of modern
medicine and has significantly contributed to the increase in life ex-
pectancy worldwide. However, no vaccines exist for many life-
threatening viruses such as HIV (Burton, 2019), Zika virus (Pierson
and Diamond, 2018), or hepatitis C virus (HCV) (Bailey et al., 2019).
Additionally, efficient broad-spectrum antiviral drugs are still
missing, making infectious diseases a significant challenge for
modern health systems. Viruses can also trigger or fuel non-infec-
tious diseases such as cancer (Young and Rickinson, 2004) and
are suspected to contribute to various other chronic diseases
such as Alzheimer disease (Itzhaki, 2018) and various auto-im-
mune disorders (Münz et al., 2009). The recent emergence of high-
ly pathogenic viruses such as the Ebola virus and the emerging
SARS-CoV-2 pandemic recalls the constant threat that viruses
represent to global health. So far, the SARS-CoV-2 pandemic
has caused a global financial and social catastrophe and is ex-
pected to make a significant long-lasting impact on human health
(Zhu et al., 2020). Despite intensive research efforts, little is known
thus far regarding the interaction of the SARS-CoV-2 virus with the
human host and, as a consequence, no efficient treatment has
been designed so far (Chen et al., 2020). Moreover, only few ther-
apeutic targets have been identified, highlighting the urgency to
develop additional strategies to dissect the virus-host interactions.
Single-cell RNA sequencing (scRNA-seq) is an emerging tech-
nology that has been extensively used to study several complex
diseases, including cancer (Li et al., 2019), neurodegeneration
(Keren-Shaul et al., 2017), and auto-immune (Zhang et al., 2019)
and metabolic diseases (Jaitin et al., 2019), providing new insights
and revealing new therapeutic targets and strategies (Yofe et al.,
2020). In the context of infectious diseases, scRNA-seq studies
identified the underlying cells and pathways interacting with
various pathogens (Drayman et al., 2019; Shnayder et al., 2018;
Steuerman et al., 2018; Zanini et al., 2018). During the immune
response to a pathogen, a limited number of antigen-positive or in-
fected cells initiate and modulate the host immune response
(Blecher-Gonen et al., 2019), while most of the tissue response is
propagated through cytokines, such as type I interferon (IFN)
signaling, to bystander, uninfected cells. It is therefore essential
to develop new analytical tools to identify the rare infected cells

Cell 181, 1475–1488, June 25, 2020 ª 2020 Elsevier Inc. 1475
 ll
in order to better understand complex host-virus interactions un-
derlying these pathologies. Multiple experimental tools have
been developed over the years to track virus-infected cells in vivo,
characterize the cellular state of the infected cells, and differentiate
them from their bystander neighbors. These include fluorescently
labeled pathogens or pathogens expressing fluorescent proteins
(De Baets et al., 2015; Blecher-Gonen et al., 2019), as well as
reporter mice (Lienenklaus et al., 2009). However, in the case of hu-
man clinical samples, these tools are limited, making the pathogen-
infected cells and viral reservoir cell types hard to detect.
Viruses exploit their host cells to first express viral genes, opti-
mize the cellular environment, and then fully activate the viral repli-
cation program. Because scRNA-seq technologies rely on polya-
denylated RNA isolation and amplification, current scRNA-seq
methods can, in theory, detect these viral RNA programs and
therefore enable accurate identification of the bona fide infected
cells and their unique properties at single-cell resolution. While
such an approach has already been used to study both in vitro
(Drayman et al., 2019; Shnayder et al., 2018) and in vivo infection
models (Steuerman et al., 2018), no general computational frame-
work has been developed to detect viruses and analyze host-viral
maps in clinical samples. Here, we present a new computational
tool, called Viral-Track, that is designed to systematically scan
for viral RNA in scRNA-seq data of physiological viral infections us-
ing a direct mapping strategy. Viral-Track performs comprehen-
sive mapping of scRNA-seq data onto a large database of known
viral genomes, providing precise annotation of the cell types asso-
ciated with viral infections. Integrating these data with the host
transcriptome enables transcriptional sorting and differential
profiling of the viral-infected cells compared to bystander cells. Us-
ing a new statistical approach for differential gene expression be-
tween infected and bystander cells, we are able to recover virus-
induced programs and reveal key host factors required for viral
replication. Viral-Track is able to annotate the viral program with
high accuracy and sensitivity, as we demonstrate in several in vivo
mouse models of infection, as well as human samples of hepatitis
B virus (HBV) infection. Applying Viral-Track on bronchoalveolar
lavage (BAL) samples from moderate and severe COVID-19 pa-
tients, we reveal the infection landscape of SARS-CoV-2 and its
interaction with the host tissue. Our analysis shows a dramatic
impact of the SARS-CoV-2 virus on the immune system of severe
patients, compared to mild cases, including replacement of the tis-
sue-resident alveolar macrophages with recruited inflammatory
monocytes, neutrophils, and macrophages and an altered CD8+
T cell cytotoxic response. We find that SARS-CoV-2 mainly infects
the epithelial and macrophage subsets. In addition, Viral-Track de-
tects an unexpected co-infection of the human metapneumovirus
in one of the severe patients. This study establishes Viral-Track as
a broadly applicable tool for dissecting mechanisms of viral infec-
tions, including identification of the cellular and molecular signa-
tures involved in virus-induced pathologies.
RESULTS
Viral-Track: An Unsupervised Pipeline for
Characterization of Viral Infections in scRNA-Seq Data
All scRNA-seq computational packages implement a pipeline that
initially aligns the sequenced reads to the expressed part of a
1476 Cell 181, 1475–1488, June 25, 2020
Article
reference host genome of the relevant profiled organism. Irrele-
vant reads, representing other organisms, primers, adaptors, tem-
plate switching oligonucleotides, and other contaminants are then
commonly discarded. We reasoned that during infection, and
likely many other pathological processes, these reads can poten-
tially carry valuable information about viral RNA that is discarded
in this filtering step. In order to efficiently detect viral reads from
raw scRNA-seq data in an unsupervised manner, we developed
Viral-Track, an R-based computational pipeline (Figure 1A;
STAR Methods). Briefly, Viral-Track relies on the STAR aligner
(Dobin et al., 2013) to map the reads of scRNA-seq data to both
the host reference genome and an extensive list of high-quality
viral genomes (Stano et al., 2016). Because viral reads are highly
repetitive and generate substantial sequencing artifacts, the viral
genomes identified in Viral-Track with a sufficient number of map-
ped reads are then filtered, based on read mapping quality, nucle-
otide composition, sequence complexity, and genome coverage,
to limit the occurrence of false-positives (STAR Methods). Due to
the lack of high-quality viral genome annotations, Viral-Track in-
cludes de novo transcriptome assembly of the identified viruses
using StringTie (Pertea et al., 2015). Finally, viral reads are demul-
tiplexed, quantified using unique molecular identifiers (UMI), and
assigned to unique viral transcripts and cells (Figures 1A and
S1A). The Viral-Track algorithm has been designed to robustly
handle various types of scRNA-seq datasets, as illustrated below,
and is publicly accessible at https://github.com/PierreBSC/
Viral-Track.
In order to evaluate the specificity and sensitivity of Viral-
Track, we benchmarked Viral-Track on several scRNA-seq
datasets (Table S1). These datasets include a large number of
experiments we conducted, as well as published studies, that
span several tissues (lung, spleen, liver, and lymph node) and a
wide range of viruses: influenza A, lymphocytic choriomeningitis
virus (LCMV), vesicular stomatitis virus (VSV), herpes simplex vi-
rus 1 (HSV-1), human immunodeficiency virus (HIV), and HBV.
We first evaluated mouse lungs infected in vivo by influenza A vi-
rus and sequenced using MARS-seq2.0 (Keren-Shaul et al.,
2019; Steuerman et al., 2018). Viral-Track analysis specifically
detected the 8 distinct influenza A viral segments (NC_002016
to NC_002023 Refseq nucleotide sequences) from the specific
infecting strain (H1N1 Puerto Rico 8 strain) (Figure 1B). We per-
formed transcriptome assembly to test the feasibility of recon-
structing the viral transcriptome from 30 -enriched scRNA-seq
data. The results were highly coherent with the current knowl-
edge of influenza A transcriptome, exemplified by Viral-Track’s
ability to identify documented spliced transcript structures with
single-nucleotide precision. For instance, we identified the exact
location of the key splicing site on segment 7 that gives rise to M2
transcript and links nucleotides 51 and 740 (Dubois et al., 2014)
(Figure 1C). Quantification of the number of viral reads across
different experimental conditions was consistent with current
knowledge of the disease, with lung stomal cells of non-immune
lineages (CD45) exhibiting a significantly higher viral load
compared to immune cells (CD45+) (p = 0.039, two-tailed
Welch’s t test) (Figure 1D).
As inbred mice lack the influenza-specific restriction factor
Mx1, influenza A infection is extremely virulent in inbred mice
(Haller et al., 1980). Moreover, all influenza A mRNA are capped
 ll
Article

B C
D

E F

G H

(legend on next page)

Cell 181, 1475–1488, June 25, 2020 1477

 ll
Article

and polyadenylated, making them an optimal substrate for
scRNA-seq isolation and amplification protocols. We therefore
evaluated the sensitivity and specificity of Viral-Track in a more
challenging dataset. In this model, photoactivatable-GFP (PA-
GFP) mice were infected with LCMV (Armstrong acute strain),
a virus lacking strong poly(A) mRNA signals (Burrell et al.,
2017), via injection to the footpad. 72 h post-infection, CD45+
splenic immune cells from different spatial niches (T zone, B
zone, marginal zone, and total spleen) were profiled using the
NICHE-seq technology (Medaglia et al., 2017). Even though
the LCMV viral mRNAs are not polyadenylated, we detected
mRNA molecules that converted to cDNA through priming of
the MARS-seq oligo(dt) RT primer, and Viral-Track successfully
identified the two viral segments (LCMV segment L
[NC_004291] and S [NC_004294]) (Figure S1B), albeit the num-
ber of detected reads was an order of magnitude lower than
the number observed in influenza A infection (Figure 1E). We de-
tected viral reads in samples from the marginal zone, B zone, and
the total spleen, but not in T zone samples, and marginal zone
samples exhibited significantly higher viral load compared to B
zone and total spleen samples (Figure 1E; p = 0.0067 and
0.0083 respectively, two-tailed Welch’s t test). This observation
is in line with the biology of LCMV, which primarily infects mac-
rophages and lymphocytes from the marginal zone of the spleen
(Müller et al., 2002).
We next evaluated whether Viral-Track is sensitive to barcode
swapping during Illumina-based scRNA-seq (Griffiths et al.,
2018), which, in the case of viral RNA detection, can lead to the
false assignment of viral reads to uninfected cells. To this end,
we infected mice with one of two different viruses, LCMV and
VSV, and performed MARS-seq2.0 on CD45+CD19CD3 non-
B/T cells from the auricular draining lymph node 1 day after infec-
tion (STAR Methods). All samples were sequenced concurrently
to test for cross-sample viral read contamination. For both viruses,
Viral-Track was able to identify the correct viral segments (Figures
S1C and S1D), with no cross-contamination, evident by the
absence of VSV reads detected in the LCMV-infected cells and
vice versa (Figure S1E). We further generalized Viral-Track for
commonly used scRNA-seq technologies and non-RNA viruses.
We applied Viral-Track to scRNA-seq data from a recently publi-
cation of human primary cells infected ex vivo with HSV-1, a linear
double-stranded DNA virus, generated by the Drop-seq platform
(Drayman et al., 2019; Macosko et al., 2015). We found that Viral-
Track detected and identified correctly HSV-1 RNA specifically in
the infected samples but not in the controls (NC_001806 Refseq
nucleotide sequences) (Figures S1F and S1G). Finally, we
analyzed scRNA-seq data of CD4+ T cells infected ex vivo with
HIV-1 (Bradley et al., 2018), generated using the droplet-based
chromium platform (Zheng et al., 2017). Viral-Track successfully
identified HIV as the unique virus present in the infected samples
(Figures S1H and S1I), but detected significant amounts of HIV-1
viral reads in one control samples probably due to ambient
contamination (Yang et al., 2020).
Defining the Host Viral Interactions of HBV Using
Viral-Track
We further tested Viral-Track’s applicability for detecting viral
reads in human clinical samples. For this purpose, we generated
scRNA-seq data from a liver biopsy of an untreated hepatitis B
patient and analyzed the data using Viral-Track. Viral-Track suc-
cessfully identified HBV as the only virus present in the sample
(Figure 1F) with 18,420 reads assigned to the HBV genome
(NC_003977 Refseq sequence). Coverage analysis revealed a
strong peak located at the 50 end of the C gene, encoding for
the main core protein, suggesting that the HBV virus is actively
producing virions (Figure 1G). We then overlaid the viral data
on the host transcriptome to identify infected and bystander
populations. A total of 13,803 cells passed a lenient quality con-
trol, permitting apoptotic signals that may arise from viral infec-
tion. We identified several non-immune cell types (Figure S1J),
including hepatocytes (expressing ALB and APOA2), as well as
hepatocytes showing apoptotic signatures (ALB with high
expression of mitochondrial genes), sinusoidal endothelial cells
(FCN2), and epithelial cells (KRT7). We also observed several
subsets of immune cells such as B cells (MS4A1), plasma cells
(MZB1), conventional dendritic cells 1 (cDC1; XCR1),

Figure 1. Viral-Track Retrieves Viral Reads in a Variety of Tissues, Viral Strains, and Sequencing Platforms
(A) Schematics of the Viral-Track approach. Single-cell sequencing data of cells from an infected tissue, containing infected and bystander cells are analyzed by
Viral-Track. Viral-Track maps the sequenced reads to both the host reference genome and a database of viral genomes, overlaying infection status on top of the
host transcriptional landscape.
(B) Results of Viral-Track analysis on scRNA-seq data from influenza A PR8-infected mouse lungs. For each viral segment, represented by a dot, the complexity of
the sequences (measured by entropy, i.e., how repetitive are the mapped sequences) and the percentage of the segment that is mapped are plotted. Dark red
dots correspond to viral segments of the influenza A PR8 strain and yellow dots to segments belonging to other H1N1 influenza strains. Viral segments with more
than 50 mapped reads are plotted.
(C) Coverage plot of the influenza A segment NC_002016 (influenza A PR8 segment 7), M2 transcript location estimated using StringTie is shown below with the
splicing site position.
(D) Quantification of the number of reads assigned to influenza viral segments across experimental settings. Each dot corresponds to a technical replicate (384-
well plate). Two-tailed Welch’s t test was used to compare viral load betwen CD45 and CD45+ cells (p = 0.039).
(E) Quantification of the number of reads assigned to LCMV viral segments in the different zones of the spleen. Each dot corresponds to a technical replicate (384-
well plate). Two-tailed Welch’s t test was used to compare viral load between cells from the infected marginal zone to cells from the B zone or the whole speen (p =
0.0067 and 0.0083 respectively).
(F) Result of Viral-Track analysis on scRNA-seq data from a HBV patient. For each viral segment, represented by a dot, the entropy of the sequence and the
percentage of the segment that is mapped is plotted. Green dots correspond to viral segments that passed quality control. Viral segments with more than 50
mapped reads are plotted.
(G) Coverage plot of the HBV genome. Locations of the different viral genes from NCBI database are depicted at the bottom.
(H) Enrichment of infected cells across hepatic cell subsets (left panel); red line corresponds to an enrichment of one. Distribution of the number of HBV UMIs per
cell in each cell subset (right panel).
See also Figure S1.

1478 Cell 181, 1475–1488, June 25, 2020

 ll
Article

A B

D E

Figure 2. Viral-Track Identifies Virus-Modified Transcription in Infected Cell Subsets

(A) Distribution of vUMI+ and GFP+ cells across cells types found in the spleen.
(B) Distribution of the Pearson Correlation between GFP+ cells, vUMI+, and bystander (GFPvUMI) cells. Two-tailed Kruskal-Wallis test.
(C) Number of differentially expressed genes between bystander and infected cells in MZB cells, monocytes, and macrophages.
(D) Top 10 enriched terms identified by Gene Ontology enrichment analysis.
(E) Mean expression of four top differentially expressed genes in bystander and infected MZB cells.
See also Figure S2.

plasmacytoid dendritic cells (pDCs) (TCF4), and three different
macrophage subsets (expressing TREM2, CD163, and FCN1,
respectively). We observed a large diversity among the lympho-
cyte compartment with CD8+ T cells (CD8A), Th17 cells (CCR6,
IL23A), gd T cells (TRGC1), activated CD4 T cells (LEF1, OX40),
natural killer (NK) cells (NKG7), and a distinct cluster of activated
CD8+ T cells (CSF2 and TOX2). We analyzed infected cells using
automated thresholding over the viral signal (Figure S1J; STAR
Methods). As expected, hepatocytes and apoptotic hepatocytes
were strongly enriched among the infected cells (Figures 1H and
S1K). Interestingly, we also detected viral reads in non-hepato-
cyte clusters, including two subsets of macrophages (CD163+
and TREM2+ populations, respectively), the cDC1 subset
(XCR1+), as well as endothelial (OIT3+ cells) and epithelial cells
(KRT7+) (Figures 1H and S1K). Infection of non-hepatocyte clus-
ters, although with relatively low viral load, is coherent with
several studies, reporting active infection of macrophages
(Faure-Dupuy et al., 2019).
Together, this extensive list of validations demonstrate that
Viral-Track is a sensitive and accurate method to detect and
identify, in an unsupervised manner, virus strains in diverse
scRNA-seq samples, in different tissues, and at varying viral
types and loads. Importantly, Viral-Track can be applied to hu-
man clinical samples to extract valuable insight into the biology
of the host-virus interactions.
Viral-Track Identifies Infected versus Bystander Cells
and Uncovers Virus-Induced Pathways
To further evaluate the accuracy of Viral-Track against a well-es-
tablished model for tracking infection in single cells, we infected
mice with a GFP-expressing LCMV virus (LCMV-GFP virus) (Med-
aglia et al., 2017). We performed MARS-seq on GFP+ splenocytes

Cell 181, 1475–1488, June 25, 2020 1479

 ll
Article

A B

D E F

G I

H J

(legend on next page)

1480 Cell 181, 1475–1488, June 25, 2020

 ll
Article

and total spleen cells 72 h post-infection and analyzed the
sequenced cells (Figures S2A and S2B; STAR Methods). GFP+
cells were enriched for vUMI+ cells compared to total spleen (Fig-
ure S2A). We then calculated whether the cells positive for the
LCMV-GFP signal (GFP+ cells) were similar to the ones desig-
nated by Viral-Track as containing viral UMIs (vUMI+). Following
clustering and annotation, we observed similar proportions of
GFP+ and vUMI+ cells across cell clusters (Figures 2A and S2C;
R = 0.95, p = 9.0 * 1012), with monocytes, marginal zone B cells
(MZBs), and macrophages being the major infected cell types. We
then evaluated the transcriptional signatures within these two sets
of cells by computing the Pearson correlation between each pair
of cells. We observed similar distribution of Pearson correlation
within the GFP+ and vUMI+ monocyte cells (Figure 2B) that was
significantly higher (median correlation of 0.65, 0.64, and 0.51,
respectively) than the correlation observed between GFP vUMI
bystander monocytes. We conclude that Viral-Track correctly
identifies a homogeneous set of infected cells from in vivo
scRNA-seq samples similar to the one identified by conventional
reporter viruses, even in the more difficult scenario in which viral
transcripts are poorly polyadenylated.
We next evaluated the ability of Viral-Track to detect host fac-
tors associated with virus replication. For this purpose, we devel-
oped a statistical method that detects differentially expressed
genes based on data binarization and complementary log-log
regression (STAR Methods; Methods S1). We used this approach
to test for transcriptional differences between bystander and in-
fected cells during spleen LCMV infection across the three main
infected cell types: macrophages, MZB cells, and monocytes.
We observed that MZB cells were the most influenced by the viral
infection, compared to monocytes and macrophages (107, 42,
and 3 genes upregulated, respectively, Z score >3) (Figure 2C).
We performed Gene Ontology enrichment analysis on the upregu-
lated genes in MZB cells and observed a significant enrichment in
several pathways, including ‘‘chromosome organization,’’ ‘‘DNA
replication,’’ and ‘‘cell cycle,’’ suggesting that LCMV triggers cell
division in MZB cells (Figure 2D). Indeed, LCMV-infected MZB
cells exhibited higher levels of cell cycle-related genes such as
Smc2 (required for chromatin condensation), Cdc6 (regulator of
DNA replication), and Stmn1 (regulator of mitotic spindle) (Figures
2E and S2D), but also fibrillarin (Fbl), a host factor whose expres-
sion is required by several viruses (Deffrasnes et al., 2016) (Fig-
ure 2E). This is in line with a previous report highlighting the ability
of LCMV to trigger an abortive form of cell division blocked in the
G1 phase (Beier et al., 2015). Altogether, our results show that
Viral-Track is sufficient to detect infected cells in in vivo scRNA-
seq data and infer the differential gene expression in infected
versus bystander cells.
A Single-Cell Map of SARS-CoV-2 Infection in Mild and
Severe Patients
COVID-19 is a viral disease caused by SARS-CoV-2 infection,
which has recently been recognized as the cause for a pandemic
(Wang et al., 2020a). Little is currently known about the course of
the disease and how the virus interacts with the host immune
system in its mild and severe manifestations. To gain insights
on the infection course in humans, we performed scRNA-seq
and Viral-Track analysis on BALF samples from three mild and
six severe COVID-19 patients (Liao et al., 2020). In total, 50,615
cells passed quality control and were analyzed using the MetaCell
algorithm (Baran et al., 2019) (Figure 3A; STAR Methods). Meta-
cell analysis coarsely grouped the metacells into the myeloid,
lymphoid, and epithelial lineages, and each lineage was further
subdivided into smaller subsets (Figures 3A, 3B and S3A). Among
epithelial cells, we identified epithelial progenitors (expressing
SOX4), type II alveolar cells (AT2, expressing SFTPB), ciliated
cells (FOXJ1), ionocytes (CFTR), goblet cells (MUC5B), and club
cells (SCGB1A1; Figure S3B). Lymphoid cells consisted several
subtypes of CD4+ T cells, including naive CD4+ T cells (express-
ing CCR7), regulatory T cells (Treg, expressing FOXP3), and T
follicular helper cells (Tfh, expressing CXCL13 and PDCD1), but
also diverse CD8+ subsets, such as NK cells (NCAM1), resident
memory CD8+ T cells (Trm, CD8A, and ZNF683), effector CD8+
T cells (GZMA and GZMK), and cytotoxic CD8+ T cells (GNLY,
PRF1), as well as B cells (CD79A; Figure S3C). The myeloid
compartment exhibited a high diversity of cell states, including
neutrophils (FCGR3B), mast cells (CPA3), alveolar macrophages
(FABP4), dendritic cells (DCs; FSCN1), and plasmacytoid DCs
(pDC; TCF4) as well as a large diversity of monocytes (FCN1)
and monocyte-derived macrophages (SPP1) sub-populations
(Figure S3D). These results were robust across different analysis
platforms (Liao et al., 2020).
Comparison of the cellular landscape of mild and severe
patients revealed key differences in the composition of BAL

Figure 3. scRNA-Seq of 6 COVID-19 Samples Reveals Myeloid Remodeling in Severe Patients

(A) A 2-dimensional visualization of 50,615 single cells from three mild and six severe COVID-19 patients, generated by the MetaCell algorithm. Colors indicate
grouping of cells into 27 subsets, based on transcriptional similarity (Figure S3A).
(B) Quantification of the three main compartments, myeloid, lymphoid, and epithelial, across the three mild (M1–M3) and six severe (S1–S6) patients.
(C) Density plots depicting projection of cells from the mild (left) and severe (right) patients on the 2D map shown in (A).
(D–F) Quantification of the frequency of specific cell subsets in the myeloid (D), lymphoid (E), and epithelial (F) compartments, across the nine patients. Diamond
marks patient S1, co-infected with the human metapneumovirus (Figures 4D–4H). Horizontal lines indicate mean frequency.
(G) Percentage of proliferating cells (determined by thresholding over a cell-cycle-related gene module, detailed in Table S3) in each of 455 metacells, projected
on the 2D map shown in (A).
(H) Quantification of the type I interferon response gene module across 455 metacells, projected on the 2D map shown in (A). Color scale represents log2 fold
change over the median expression of the module across all metacells.
(I) Differential gene expression analysis. Each panel compares pooled gene expression between naive and non-naive CD4+ T cells (left) and effector and cytotoxic
CD8+ T cells (right) cell subsets.
(J) Differential gene expression analysis between cells belonging to AM (left) and SPP1hiC1Qhi macrophages (right) from mild (x axis) and severe (y axis) patients. (I
and J) Values represent log2 size-normalized expression (transcripts per 1,000 UMI).
See also Figure S3.

Cell 181, 1475–1488, June 25, 2020 1481

 ll
Article

A B

C D

F
E

(legend on next page)

1482 Cell 181, 1475–1488, June 25, 2020

 ll
Article

samples (Figures 3B and 3C). We found changes to each of the
three compartments (Figures 3D–3F and S3E–S3G). While alve-
olar macrophages and pDC where enriched in the myeloid
compartment in the mild patients, the severe patients’ myeloid
cells were characterized by a patient-specific diversity associ-
ated with accumulation of neutrophils, FCN1+ monocytes, and
monocyte-derived SPP1+ macrophages (Figures 3D and S3E).
Additionally, NK cells and naive CCR7+ CD4+ T cells were
consistently enriched across severe patients BAL, while
ZNF683hi CD8+ Trm cells were specific to mild patients (Figures
3E and S3F). We also observed changes in the epithelial
compartment, as severe patients exhibited higher numbers of
club cells and AT2 cells (Figures 3F and S3G). By investigating
expression patterns of shared gene expression programs, we
observed that cytotoxic CD8+ cells and the CD4+ Tfh cells are
the most proliferative compartments (Figure 3G), while a broad
interferon type I response, a hallmark of viral response, is mainly
expressed by neutrophils and, to a lesser extent, FCN1+ mono-
cytes (Figure 3H). We next performed in-depth differential gene
expression analysis between subsets characteristic of mild or
severe patients. We found that CD4+ T cells in the severe pa-
tients exhibit a more naive phenotype, expressing higher levels
of IL7R, CCR7, S1PR1, and LTB. The CD8+ Trm cells signatures
are restricted to the mild patients and have higher levels of the
effector molecules XCL1, ITGAE, CXCR6, and ZNF683 (Fig-
ure 3I). Comparing gene expression differences in myeloid types
between severe and mild patients revealed disease severity-
associated upregulation of inflammatory chemokine genes in
SPP1+ monocyte-derived macrophages populations (CCL2,
CCL3, CCL4, CCL7, and CCL8; Figure 3J), as well as genes
associated with hypoxia or oxidative stress (HMOX1 and
HIF1A), and downregulation of MHC class II (HLA-A and HLA-
DQA1) and type I IFN genes (IFIT1 and OAS1). Alveolar macro-
phages displayed a severity-associated signature, including
upregulation of the chemokines CCL18 and CCL4L2 and the
cathepsins CTSL and CTSB (Figure 3J). Together, we identified
dramatic differences between the mild and severe COVID-19
patients, including an inflammatory signature and a perturbed
immune response associated with the severe manifestation
of the COVID-19 disease. These also highlight potential
immunotherapy treatment of the severe patients by targeting
the hyper inflammatory response that is activated by inflamma-
tory cytokines such as interleukin (IL)-6 and IL-8 (Liu et al.,
2019) (Figure S3H).
Viral-Track Identifies Co-infection of SARS-CoV-2 with
the Human Metapneumovirus
To characterize the in vivo crosstalk of SARS-CoV-2 with its human
host, we applied Viral-Track on the data generated from the nine
SARS-CoV-2 patients and the rich cellular landscape we identi-
fied. SARS-CoV-2 transcripts were detected in all six severe sam-
ples in variable amounts, ranging from less than 400 transcripts to
more than 15,000 (Figures 4A and S4A). In contrast, no viral reads
were detected in the three mild patients (Figure 4A). Coverage
analysis revealed that the majority of the viral reads mapped to
the 30 end of the viral segment and corresponded to positive-
stranded RNA (Figure 4B). This is in agreement with the coronavi-
rus transcription: due to a nested transcription process all genomic
and subgenomic RNA molecules share the same 30 end (Masters,
2006). We then analyzed the enrichment of vUMIs in the cell pop-
ulations represented in the BAL samples. We observed a strong
enrichment of viral reads in the ciliated and epithelial progenitor
population, two known cellular targets of the virus, which express
the main receptor of the SARS-CoV-2 virus ACE2, as well as
TMPRSS2, a protease essential for SARS-CoV-2 entry (Figures
4C and S4B; Table S2) (Hoffmann et al., 2020). We also observed
enrichment of SARS-CoV-2 reads in the SPP1+ macrophage pop-
ulation, suggesting either that SARS-CoV-2 can infect immune
cells from the myeloid compartment or that SPP1+ macrophages
phagocytose infected cells or viral particles. Differential gene
expression analysis between vUMI+ infected and vUMI
bystander SPP1+ macrophages in the patients with the highest
viral load, revealed that infected macrophages have a higher
expression of chemokines (CCL7, CCL8, and CCL18) and APOE,
and a lower expression of TAOK1, a serine/threonine-protein ki-
nase in the p38 MAPK cascade (Figure S4C). Interestingly,
CD147 (also known as BSG), a potential new SARS-CoV-2
receptor (Wang et al., 2020b), is expressed by all cell types,
including immune cells, suggesting alternative routes for the virus
to infect these cells.
Often in cases of infectious diseases, the specific infecting vi-
rus is not known, or may be accompanied by co-infection with
additional unknown viruses. Viral-Track applies an unsupervised
mapping strategy and is optimally designed to systematically
profile the source of infection or co-infections in human clinical
samples. To our surprise, Viral-Track analysis of data from one
of the severe patients (S1) revealed the presence of a second vi-
rus, the human metapneumovirus (hMPV) (NC_039199 Refseq
sequence, Figure 4D) with more than one million reads mapped

Figure 4. Viral-Track Reveals Infection Specificity and a Co-infection in Severe COVID-19

(A) Total number of viral reads mapped to the SARS-CoV-2 viral genome in the profiled COVID-19 patients.
(B) Coverage plot of the SARS-CoV-2 viral genome.
(C) Enrichment of viral UMIs over expected values across 361 metacells, projected on the 2D map shown in Figure 4A. Color scale indicates log2 observed/
expected vUMIs. Only metacells with more than one expected UMI are plotted.
(D) Result of Viral-Track analysis on patient S1. For each viral segment, represented by a dot, the entropy of the sequence (how repetitive are the mapped
sequences) and the percentage of the segment that is mapped is plotted. Green dots correspond to viral segments that have passed quality control. Viral
segments with more than 50 mapped reads are plotted.
(E) Coverage plot of the human metapneumovirus (hMPV) genome.
(F) Distribution of hMPV UMIs across patient S1 sequenced cells. Red dashed line indicates automatic thresholding of vUMI+ cells.
(G) Enrichment of vUMI+ cells over expected values across 297 metacells, projected on the 2D map shown in Figure 4A. Color scale indicates log2 observed/
expected. Only metacells with more than one expected vUMI+ cell are plotted.
(H) Volcano plot showing the relative expression between infected and bystander monocytes of patient S1. Differentially expressed (>1 log2 fold change) and
statistically significant (p value <0.01) are colored in orange.
See also Figure S4.

Cell 181, 1475–1488, June 25, 2020 1483

 ll
to hMPV in this specific patient. hMPV is a non-segmented, sin-
gle-stranded, and negative-sense RNA virus that is responsible
for upper and lower respiratory tract infections in mostly young
(<5 years) children but can also target elderly as well as im-
muno-compromised patients (Panda et al., 2014). hMPV has
been implicated as a possible source of co-infection with the
original SARS-CoV virus (Chan et al., 2003).
Coverage analysis revealed that most reads fall into the N, P, M,
F, M2, SH, G, but not L, genes of hMPV (Figure 4E). We observed a
typical pattern of biased scRNA-seq coverage, indicating that the
N, P, M, F, M2, SH, and G genes are actively transcribed, and sug-
gesting that the hMPV was active and replicating at the time of
sample collection. Analysis of the viral UMI distribution across
cells revealed a substantial viral load in a large subset of the cells,
spanning hundreds to thousands vUMIs per infected cell (Fig-
ure 4F), independently of the total host UMIs in that cell (Fig-
ure S4D). We mapped the infected cells and characterized their
distribution across cell types. The infected patient is characterized
by high levels of monocytes and CD4+ T cells (Figure S4E). Unlike
the SARS-CoV-2 virus infection map, hMPV-infected cells were
highly enriched in the monocyte compartment but not in the
epithelial and SPP1+ macrophage compartments (Figure 4G).
We tested whether the hMPV could alter the function of the in-
fected monocytes, and therefore influence the course of the dis-
ease. Using Viral-Track, we detected a large number of up- and
downregulated genes in infected monocytes compared to
bystander monocytes (Figure 4H). Interestingly, several key recep-
tor genes required for monocyte activation such as CD16
(FCGR3B), G-CSF receptor (CSF3R), and the formyl peptide recep-
tor (FRP1) were downregulated in the infected compared to the
bystander cells. Moreover, we observed a dramatic downregulation
of type I Interferon signaling and interferon stimulated genes (ISGs),
including viral restriction factors, (e.g., IFIT3). A gene set enrichment
analysis (Figure S4F) revealed a strong enrichment of interferon
response genes in the downregulated gene set, suggesting that
the hMPV is strongly downregulating the IFN response pathway.
Several anti-inflammatory genes were upregulated, including
LILRB4 (a potent inhibitor of monocyte activation) (Lu et al., 2009)
and MITF, a transcription factor known to be a critical suppressor
of innate immunity (Harris et al., 2018). Last, we observed a positive
and significant association between total number of hMPV UMIs
and production of type I IFN, highlighting that while hMPV dampens
the response to type I IFN, production of this signal is highly
restricted to a rare (~1%) population of cells with a high viral load
(Figure S4G). Altogether, our analysis described the distribution of
SARS-CoV-2-infected cells in patient’s BAL and revealed the pres-
ence of a viral co-infection by the hMPV that dampens the immune
activation of the monocyte compartment in the infected patient.
Further large-scale analyses of mild versus severe patients need
to be conducted to better understand if the co-infection is corre-
lated or even causative in SARS-CoV-2 pathology.
DISCUSSION
The virosphere contains hundreds of thousands of species that
constantly interact with their host cells. Over the years, several
genomic techniques have been developed to detect virus-derived
sequences in human samples. For instance, deep sequencing as-
1484 Cell 181, 1475–1488, June 25, 2020
Article
says are unbiased and sensitive in their ability to detect extremely
rare viral sequences (Moustafa et al., 2017), but do not provide in-
formation about the infected cells and the cellular changes
induced by the infection. Alternatively, it is possible to combine
DNA probes with scRNA-seq to enrich for viral sequences and in-
crease the sensitivity of the assay, but this requires prior knowl-
edge of the viruses present in each sample (Zanini et al., 2018).
Here, we present Viral-Track, a robust and unsupervised compu-
tational pipeline that can detect viral RNA in any scRNA-seq data-
set without the need for experimental modifications or prior knowl-
edge of the infecting agent. Viral-Track was benchmarked on data
originating from various tissues, infected by viruses with marked
differences in their RNA properties, and generated with different
scRNA-seq platforms. We demonstrate that Viral-Track can
readily provide essential information on infection status in clinical
samples, identify infected cells, probe viral-induced transcrip-
tional alterations, and reveal cases of co-infection.
In practice, only 70%–85% of scRNA-seq reads map to the host
genome and represent polyadenylated exonic host transcripts,
whereas the remainder of the data is usually overlooked in analysis.
We show that these unmapped scRNA-seq reads, in pathological
human samples, potentially contain valuable information on viral
infection and can be effectively used for viral genome assembly.
Viral-Track can resolve complex cellular ecosystems perturbed
by viral infection and provide an unbiased map of the infected cells,
as well as the transcriptional perturbations induced by the virus at
the single cell level. We combine Viral-Track with a novel statistical
approach to detect differentially expressed genes from scRNA-
seq data, therefore allowing the detection of gene expression
changes triggered by viral infection and differentiating them from
the more abundant bystander effects, such as type I IFN signaling,
at the single cell level. Further advances will focus on applying Viral-
Track on largescale datasets containing scRNA-seq data from
dozens of samples, leading to robust single-cell viral metagenomic
studies that characterize the viral evolution and interactions of vi-
rus-induced disease mechanisms with host genetics.
Here, we applied scRNA-seq and Viral-Track analysis to
COVID-19 patient-derived samples to provide a cellular and viral
atlas of the BAL lung cells from COVID-19 patients. This analysis
revealed the diversity of the immune responses across COVID-
19 patients and between mild and severe patients. We expect
that as the pandemic keeps spreading and global research ef-
forts grow, additional scRNA-seq samples from COVID-19 pa-
tients will be generated, including patients treated with emerging
immunotherapies (Liu et al., 2019). Such an approach might help
to solve key questions including the contribution of the humoral
response (Iwasaki and Yang, 2020), the role of the IL6 pathway
(Herold et al., 2020), and the immune memory induced by the vi-
rus (Prompetchara et al., 2020). Viral-Track can contribute to the
global effort to identify the different cellular compartments that
are targeted and affected by COVID-19 and other viruses and
to detect possible co-infection by unexpected viruses. Co-infec-
tions are gaining recognition in the scientific and medical com-
munity as critical factors in disease prognosis (Zhang et al.,
2020). So far, research focused mainly on co-infections of bac-
terial sources or of well-known viruses such as influenza A (Wu
et al., 2020). Understanding the diversity of viral co-infections
and their mechanisms of immune suppression at the cellular

Article
and molecular level could therefore provide highly valuable infor-
mation and lead toward possible therapeutic targets, especially
for severe patients, whose treatment options are limited.
Limitations
Viral-Track is a new and powerful tool to decipher host-viral in-
teractions. However, its impact is dependent on several factors,
the most critical one being the biochemical and pathophysiolog-
ical properties of the virus. The absence of a poly(A) tail at the
end of viral RNA molecules can significantly decrease their cap-
ture rate efficiency in current scRNA-seq techniques, as shown
by the LCMV example. This may hinder Viral-Track’s ability to
robustly identify infected cells or discern differential expression
between infected and bystander cells in such viruses. Other
properties of the viral RNA molecules, absence/presence of 50
capping, nucleotide composition, or dependence on RNA bind-
ing proteins, may also affect capture efficiency, and as the tech-
nology develops, further research will focus on the classification
of molecular features that facilitate or prevent virus identification
by scRNA-seq. Notably, non poly(A)-based scRNA-seq tech-
niques, such as RamDA-seq (Hayashi et al., 2018), can be poten-
tially used when profiling these datasets.
Another limiting factor for Viral-Track’s applicability is the
potential scarcity of viral reads and infected cells in the sam-
ple. As shown in our analysis of SARS-CoV-2-infected sam-
ples, only a limited number of viral reads are detected in
some of the samples. This may be due to the specific stage
of the disease (He et al., 2020), or sampling biases favoring
mainly the lung immune populations, with lower representa-
tion of non-immune cells that are the primary targets of the vi-
rus. Therefore, future COVID-19 scRNA-seq studies should
consider this limitation in their experimental design and aim
for a better representation of the upper respiratory tissue
and the lung parenchyma. Alternative approaches may rely
on index sorting and single-cell transcriptome-trained sorting
to design optimal gating strategies for capturing and enriching
the stromal populations.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Mice
B LCMV/VSV infections
B Subjects
d METHOD DETAILS
B Lymph Node MARS-seq data generation
B Influenza MARS-seq data generation
B LCMV spleen MARS-seq data generation
B 10X HBV liver data generation
B 10X COVID-19 data generation
ll
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Read mapping/alignment
B Viral database and STAR Index building
B Processing and filtering of the BAM files
B Transcript reconstruction
B MARS-seq data demultiplexing and UMI count
B Drop-seq and 10X data download, pre-processing and
demultiplexing
B Analysis of the MARS-seq spleen LCMV dataset
B Analysis of the 10X HBV liver dataset
B Analysis of the COVID-19 BAL dataset
B Testing for infection specificity in COVID-19 BAL
dataset
B Dichotomized differential gene expression analysis
B Automate thresholding to detect HBV and hMPV in-
fected cells
B Gene set enrichment analysis
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.006.
ACKNOWLEDGMENTS
We thank Dr. Noam Stern Ginossar and Dr. Yoav Golan for careful evaluation of
the manuscript; Dr. Etienne Simon-Loriere, Thomas Jacquemont, and Alice
Balfourier for valuable advices; Tali Wiesel from the Scientific Illustration unit
of the Weizmann Institute for artwork; and members of the Amit laboratory
for discussions. I.A. is an Eden and Steven Romick Professorial Chair and sup-
ported by Merck KGaA (Darmstadt, Germany), the Chan Zuckerberg Initiative
(CZI), an HHMI International Scholar award, the European Research Council
Consolidator Grant (ERC-COG) 724471-HemTree2.0, an SCA award from
the Wolfson Foundation and Family Charitable Trust, the Thompson Family
Foundation, an MRA Established Investigator Award (509044), the Israel Sci-
ence Foundation (703/15), the Ernest and Bonnie Beutler Research Program
for Excellence in Genomic Medicine, the Helen and Martin Kimmel award for
innovative investigation, a NeuroMac DFG/Transregional Collaborative
Research Center grant, an International Progressive MS Alliance/NMSS (PA-
1604 08459), and an Adelis Foundation grant. P.B. is supported by a PhD
scholarship from the Ecole Normale Supérieure, Paris. B.S. has received fund-
ing from the French Government’s Investissement d’Avenir program, Labora-
toire d’Excellence ‘‘Integrative Biology of Emerging Infectious Diseases’’
(ANR-10-LABX-62-IBEID). Z.Z. and Y.L. were supported by fundings from
the National Natural Science Foundation of China (91442127 to Z.Z.；
81700540 to Y.L.).
AUTHOR CONTRIBUTIONS
P.B. designed and developed Viral-Track, performed various computational
analyses, and wrote the manuscript. A.G. performed computational analyses
and wrote the manuscript. Y.L., G.X., and S.Z. designed and performed exper-
iments. Y.B. developed Viral-Track. E.D. contributed to data processing and
analysis. R.B.-G., M.C., and C.M. designed and performed experiments.
H.L. contributed to data analysis. A.D. contributed to data analysis, manu-
script writing, and scientific communication. B.S. contributed to development
of computational methods and bioinformatic analysis and wrote the manu-
script. Z.Z. conceived, designed, and analyzed experiments and wrote the
manuscript. I.A. directed the project, conceived, designed, and analyzed ex-
periments, and wrote the manuscript.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Cell 181, 1475–1488, June 25, 2020 1485
 ll
Received: March 30, 2020
Revised: April 16, 2020
Accepted: May 1, 2020
Published: May 8, 2020
REFERENCES
Bailey, J.R., Barnes, E., and Cox, A.L. (2019). Approaches, Progress, and
Challenges to Hepatitis C Vaccine Development. Gastroenterology 156,
418–430.
Baran, Y., Bercovich, A., Sebe-Pedros, A., Lubling, Y., Giladi, A., Chomsky,
E., Meir, Z., Hoichman, M., Lifshitz, A., and Tanay, A. (2019). MetaCell: anal-
ysis of single-cell RNA-seq data using K-nn graph partitions. Genome Biol.
20, 206.
Beier, J.I., Jokinen, J.D., Holz, G.E., Whang, P.S., Martin, A.M., Warner, N.L.,
Arteel, G.E., and Lukashevich, I.S. (2015). Novel mechanism of arenavirus-
induced liver pathology. PLoS ONE 10, e0122839.
Benjamini, Y., and Hochberg, Y. (1995). Controlling the false discovery rate: a
practical and powerful approach to multiple testing. J. R. Stat. Soc. B 57,
289–300.
Blecher-Gonen, R., Bost, P., Hilligan, K.L., David, E., Salame, T.M., Rous-
sel, E., Connor, L.M., Mayer, J.U., Bahar Halpern, K., Tóth, B., et al.
(2019). Single-Cell Analysis of Diverse Pathogen Responses Defines a
Molecular Roadmap for Generating Antigen-Specific Immunity. Cell
Syst. 8, 109–121.
Blondel, V.D., Guillaume, J.-L., Lambiotte, R., and Lefebvre, E. (2008). Fast un-
folding of communities in large networks. J. Stat. Mech. Theory Exp. 2008,
P10008.
Bradley, T., Ferrari, G., Haynes, B.F., Margolis, D.M., and Browne, E.P. (2018).
Single-Cell Analysis of Quiescent HIV Infection Reveals Host Transcriptional
Profiles that Regulate Proviral Latency. Cell Rep. 25, 107–117.
Burrell, C.J., Howard, C.R., and Murphy, F.A. (2017). Fenner and White’s Med-
ical Virology (Academic Press).
Burton, D.R. (2019). Advancing an HIV vaccine; advancing vaccinology. Nat.
Rev. Immunol. 19, 77–78.
Chan, P.K.S., Tam, J.S., Lam, C.-W., Chan, E., Wu, A., Li, C.-K., Buckley, T.A.,
Ng, K.-C., Joynt, G.M., Cheng, F.W.T., et al. (2003). Human metapneumovirus
detection in patients with severe acute respiratory syndrome. Emerg. Infect.
Dis. 9, 1058–1063.
De Baets, S., Verhelst, J., Van den Hoecke, S., Smet, A., Schotsaert, M.,
Job, E.R., Roose, K., Schepens, B., Fiers, W., and Saelens, X. (2015). A
GFP expressing influenza A virus to report in vivo tropism and protection
by a matrix protein 2 ectodomain-specific monoclonal antibody. PLoS
ONE 10, e0121491.
Deffrasnes, C., Marsh, G.A., Foo, C.H., Rootes, C.L., Gould, C.M., Grusovin,
J., Monaghan, P., Lo, M.K., Tompkins, S.M., Adams, T.E., et al. (2016).
Genome-wide siRNA screening at biosafety level 4 reveals a crucial role for fi-
brillarin in henipavirus infection. PLoS Pathog. 12, e1005478.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Drayman, N., Patel, P., Vistain, L., and Tay, S. (2019). HSV-1 single-cell anal-
ysis reveals the activation of anti-viral and developmental programs in distinct
sub-populations. eLife 8, e46339.
Dubois, J., Terrier, O., and Rosa-Calatrava, M. (2014). Influenza viruses and
mRNA splicing: doing more with less. MBio 5, e00070-14.
Faure-Dupuy, S., Delphin, M., Aillot, L., Dimier, L., Lebossé, F., Fresquet, J.,
Parent, R., Matter, M.S., Rivoire, M., Bendriss-Vermare, N., et al. (2019). Hep-
atitis B virus-induced modulation of liver macrophage function promotes he-
patocyte infection. J. Hepatol. 71, 1086–1098.
Finak, G., McDavid, A., Yajima, M., Deng, J., Gersuk, V., Shalek, A.K.,
Slichter, C.K., Miller, H.W., McElrath, M.J., Prlic, M., et al. (2015). MAST: a
flexible statistical framework for assessing transcriptional changes and char-
1486 Cell 181, 1475–1488, June 25, 2020
Article
acterizing heterogeneity in single-cell RNA sequencing data. Genome Biol.
16, 278.
Griffiths, J.A., Richard, A.C., Bach, K., Lun, A.T.L., and Marioni, J.C. (2018).
Detection and removal of barcode swapping in single-cell RNA-seq data.
Nat. Commun. 9, 2667.
Hafemeister, C., and Satija, R. (2019). Normalization and variance stabilization
of single-cell RNA-seq data using regularized negative binomial regression.
Genome Biol. 20, 296.
Haller, O., Arnheiter, H., Lindenmann, J., and Gresser, I. (1980). Host gene in-
fluences sensitivity to interferon action selectively for influenza virus. Nature
283, 660–662.
Harris, M.L., Fufa, T.D., Palmer, J.W., Joshi, S.S., Larson, D.M., Incao, A., Gil-
dea, D.E., Trivedi, N.S., Lee, A.N., Day, C.-P., et al.; NISC Comparative
Sequencing Program (2018). A direct link between MITF, innate immunity,
and hair graying. PLoS Biol. 16, e2003648.
Hayashi, T., Ozaki, H., Sasagawa, Y., Umeda, M., Danno, H., and Nikaido, I.
(2018). Single-cell full-length total RNA sequencing uncovers dynamics of
recursive splicing and enhancer RNAs. Nat. Commun. 9, 619.
He, X., Lau, E.H.Y., Wu, P., Deng, X., Wang, J., Hao, X., Lau, Y.C., Wong, J.Y.,
Guan, Y., Tan, X., et al. (2020). Temporal dynamics in viral shedding and trans-
missibility of COVID-19. Nat. Med. 26, 672–675.
Herold, T., Jurinovic, V., Arnreich, C., Hellmuth, J.C., von Bergwelt-Baildon,
M., Klein, M., and Weinberger, T. (2020). Level of IL-6 predicts respiratory fail-
ure in hospitalized symptomatic COVID-19 patients. MedRxiv. https://doi.org/
10.1101/2020.04.01.20047381.
Hoffmann, M., Kleine-Weber, H., Schroeder, S., Krüger, N., Herrler, T., Erich-
sen, S., Schiergens, T.S., Herrler, G., Wu, N.-H., Nitsche, A., et al. (2020).
SARS-CoV-2 Cell Entry Depends on ACE2 and TMPRSS2 and Is Blocked by
a Clinically Proven Protease Inhibitor. Cell 181, 271–280.
Itzhaki, R.F. (2018). Corroboration of a Major Role for Herpes Simplex Virus
Type 1 in Alzheimer’s Disease. Front. Aging Neurosci. 10, 324.
Iwasaki, A., and Yang, Y. (2020). The potential danger of suboptimal antibody
responses in COVID-19. Nat. Rev. Immunol. Published online April 21, 2020.
https://doi.org/10.1038/s41577-020-0321-6.
Jaitin, D.A., Kenigsberg, E., Keren-Shaul, H., Elefant, N., Paul, F., Zaretsky, I.,
Mildner, A., Cohen, N., Jung, S., Tanay, A., and Amit, I. (2014). Massively par-
allel single-cell RNA-seq for marker-free decomposition of tissues into cell
types. Science 343, 776–779.
Jaitin, D.A., Adlung, L., Thaiss, C.A., Weiner, A., Li, B., Descamps, H., Lundg-
ren, P., Bleriot, C., Liu, Z., Deczkowska, A., et al. (2019). Lipid-Associated
Macrophages Control Metabolic Homeostasis in a Trem2-Dependent Manner.
Cell 178, 686–698..
Chen, J., Liu, D., Liu, L., Liu, P., Xu, Q., Xia, L., Ling, Y., Huang, D., Song,
S., Zhang, D., et al. (2020). A pilot study of hydroxychloroquine in treatment
of patients with common coronavirus disease-19 (COVID-19). J. Zhejiang
Univ. Medical Sci. 49 https://doi.org/10.3785/j.issn.1008-9292.2020.
03.030.
Keren-Shaul, H., Spinrad, A., Weiner, A., Matcovitch-Natan, O., Dvir-Sztern-
feld, R., Ulland, T.K., David, E., Baruch, K., Lara-Astaiso, D., Toth, B., et al.
(2017). A Unique Microglia Type Associated with Restricting Development of
Alzheimer’s Disease. Cell 169, 1276–1290.
Keren-Shaul, H., Kenigsberg, E., Jaitin, D.A., David, E., Paul, F., Tanay, A., and
Amit, I. (2019). MARS-seq2.0: an experimental and analytical pipeline for in-
dexed sorting combined with single-cell RNA sequencing. Nat. Protoc. 14,
1841–1862.
Kharchenko, P.V., Silberstein, L., and Scadden, D.T. (2014). Bayesian
approach to single-cell differential expression analysis. Nat. Methods 11,
740–742.
Lake, B.B., Chen, S., Sos, B.C., Fan, J., Kaeser, G.E., Yung, Y.C., Duong, T.E.,
Gao, D., Chun, J., Kharchenko, P.V., and Zhang, K. (2018). Integrative single-
cell analysis of transcriptional and epigenetic states in the human adult brain.
Nat. Biotechnol. 36, 70–80.

Article
Li, H., Handsaker, B., Wysoker, A., Fennell, T., Ruan, J., Homer, N., Marth, G.,
Abecasis, G., and Durbin, R.; 1000 Genome Project Data Processing Sub-
group (2009). The Sequence Alignment/Map format and SAMtools. Bioinfor-
matics 25, 2078–2079.
Li, H., van der Leun, A.M., Yofe, I., Lubling, Y., Gelbard-Solodkin, D., van
Akkooi, A.C.J., van den Braber, M., Rozeman, E.A., Haanen, J.B.A.G.,
Blank, C.U., et al. (2019). Dysfunctional CD8 T Cells Form a Proliferative,
Dynamically Regulated Compartment within Human Melanoma. Cell 176,
775–789.
Liao, M., Liu, Y., Yuan, J., Wen, Y., Xu, G., Zhao, J., Cheng, L., Li, J., Wang, X.,
Wang, F., et al. (2020). The landscape of lung bronchoalveolar immune cells in
COVID-19 revealed by single-cell RNA sequencing. medRxiv. https://doi.org/
10.1101/2020.02.23.20026690.
Liberzon, A., Birger, C., Thorvaldsdóttir, H., Ghandi, M., Mesirov, J.P., and
Tamayo, P. (2015). The Molecular Signatures Database (MSigDB) hallmark
gene set collection. Cell Syst. 1, 417–425.
Lienenklaus, S., Cornitescu, M., Zie˛tara, N., qyszkiewicz, M., Gekara, N., Ja-
b1ónska, J., Edenhofer, F., Rajewsky, K., Bruder, D., Hafner, M., et al.
(2009). Novel reporter mouse reveals constitutive and inflammatory expres-
sion of IFN-b in vivo. J. Immunol. 183, 3229–3236.
Liu, L., Wei, Q., Lin, Q., Fang, J., Wang, H., Kwok, H., Tang, H., Nishiura, K.,
Peng, J., Tan, Z., et al. (2019). Anti-spike IgG causes severe acute lung injury
by skewing macrophage responses during acute SARS-CoV infection. JCI
Insight 4, 123158.
Lu, H.K., Rentero, C., Raftery, M.J., Borges, L., Bryant, K., and Tedla, N.
(2009). Leukocyte Ig-like receptor B4 (LILRB4) is a potent inhibitor of Fcgam-
maRI-mediated monocyte activation via dephosphorylation of multiple ki-
nases. J. Biol. Chem. 284, 34839–34848.
Macosko, E.Z., Basu, A., Satija, R., Nemesh, J., Shekhar, K., Goldman, M., Tir-
osh, I., Bialas, A.R., Kamitaki, N., Martersteck, E.M., et al. (2015). Highly Par-
allel Genome-wide Expression Profiling of Individual Cells Using Nanoliter
Droplets. Cell 161, 1202–1214.
Masters, P.S. (2006). The molecular biology of coronaviruses. Adv. Virus Res.
66, 193–292.
McInnes, L., Healy, J., and Melville, J. (2018). Umap: Uniform manifold
approximation and projection for dimension reduction. ArXiv,
ArXiv1802.03426.
Medaglia, C., Giladi, A., Stoler-Barak, L., De Giovanni, M., Salame, T.M.,
Biram, A., David, E., Li, H., Iannacone, M., Shulman, Z., et al. (2017). Spatial
reconstruction of immune niches by combining photoactivatable reporters
and scRNA-seq. Science 358, 1622–1626.
Moustafa, A., Xie, C., Kirkness, E., Biggs, W., Wong, E., Turpaz, Y., Bloom, K.,
Delwart, E., Nelson, K.E., Venter, J.C., and Telenti, A. (2017). The blood DNA
virome in 8,000 humans. PLoS Pathog. 13, e1006292.
Müller, S., Hunziker, L., Enzler, S., Bühler-Jungo, M., Di Santo, J.P., Zinker-
nagel, R.M., and Mueller, C. (2002). Role of an intact splenic microarchitec-
ture in early lymphocytic choriomeningitis virus production. J. Virol. 76,
2375–2383.
Münz, C., Lünemann, J.D., Getts, M.T., and Miller, S.D. (2009). Antiviral im-
mune responses: triggers of or triggered by autoimmunity? Nat. Rev. Immunol.
9, 246–258.
Otsu, N. (1979). A Threshold Selection Method from Gray-Level Histograms.
IEEE Trans. Syst. Man Cybern. 9, 62–66.
Panda, S., Mohakud, N.K., Pena, L., and Kumar, S. (2014). Human metapneu-
movirus: review of an important respiratory pathogen. Int. J. Infect. Dis.
25, 45–52.
Pertea, M., Pertea, G.M., Antonescu, C.M., Chang, T.-C., Mendell, J.T., and
Salzberg, S.L. (2015). StringTie enables improved reconstruction of a tran-
scriptome from RNA-seq reads. Nat. Biotechnol. 33, 290–295.
Pierson, T.C., and Diamond, M.S. (2018). The emergence of Zika virus and its
new clinical syndromes. Nature 560, 573–581.
ll
Prompetchara, E., Ketloy, C., and Palaga, T. (2020). Immune responses in
COVID-19 and potential vaccines: Lessons learned from SARS and MERS
epidemic. Asian Pac. J. Allergy Immunol. 38, 1–9.
Shnayder, M., Nachshon, A., Krishna, B., Poole, E., Boshkov, A., Binyamin, A.,
Maza, I., Sinclair, J., Schwartz, M., and Stern-Ginossar, N. (2018). Defining the
Transcriptional Landscape during Cytomegalovirus Latency with Single-Cell
RNA Sequencing. MBio 9, e00013-18.
Silverman, J.D., Roche, K., Mukherjee, S., and David, L.A. (2018). Naught all
zeros in sequence count data are the same. bioRxiv. https://doi.org/10.
1101/477794.
Smith, T., Heger, A., and Sudbery, I. (2017). UMI-tools: modeling sequencing
errors in Unique Molecular Identifiers to improve quantification accuracy.
Genome Res. 27, 491–499.
Stano, M., Beke, G., and Klucar, L. (2016). viruSITE-integrated database for
viral genomics. Database J. Biol. Databases Curation 2016. https://doi.org/
10.1093/database/baw162.
Steuerman, Y., Cohen, M., Peshes-Yaloz, N., Valadarsky, L., Cohn, O., David,
E., Frishberg, A., Mayo, L., Bacharach, E., Amit, I., and Gat-Viks, I. (2018).
Dissection of Influenza Infection In Vivo by Single-Cell RNA Sequencing. Cell
Syst. 6, 679–691.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gil-
lette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., and Me-
sirov, J.P. (2005). Gene set enrichment analysis: a knowledge-based approach
for interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA
102, 15545–15550.
Svensson, V. (2020). Droplet scRNA-seq is not zero-inflated. Nat. Biotechnol.
38, 147–150.
Svensson, V., da Veiga Beltrame, E., and Pachter, L. (2019). Quantifying the
tradeoff between sequencing depth and cell number in single-cell RNA-seq.
bioRxiv. https://doi.org/10.1101/762773.
Townes, F.W., Hicks, S.C., Aryee, M.J., and Irizarry, R.A. (2019). Feature selec-
tion and dimension reduction for single-cell RNA-Seq based on a multinomial
model. Genome Biol. 20, 295.
Wang, C., Horby, P.W., Hayden, F.G., and Gao, G.F. (2020a). A novel corona-
virus outbreak of global health concern. Lancet 395, 470–473.
Wang, K., Chen, W., Zhou, Y.-S., Lian, J.-Q., Zhang, Z., Du, P., Gong, L.,
Zhang, Y., Cui, H.-Y., Geng, J.-J., et al. (2020b). SARS-CoV-2 invades host
cells via a novel route: CD147-spike protein. BioRxiv. https://doi.org/10.
1101/2020.03.14.988345.
Wu, X., Cai, Y., Huang, X., Yu, X., Zhao, L., Wang, F., Li, Q., Gu, S., Xu, T., Li, Y.,
et al. (2020). Co-infection with SARS-CoV-2 and Influenza A Virus in Patient
with Pneumonia, China. Emerg. Infect. Dis. 26 https://doi.org/10.3201/
eid2606.200299.
Yang, S., Corbett, S.E., Koga, Y., Wang, Z., Johnson, W.E., Yajima, M., and
Campbell, J.D. (2020). Decontamination of ambient RNA in single-cell RNA-
seq with DecontX. Genome Biol. 21, 57.
Yofe, I., Dahan, R., and Amit, I. (2020). Single-cell genomic approaches for
developing the next generation of immunotherapies. Nat. Med. 26,
171–177.
Young, L.S., and Rickinson, A.B. (2004). Epstein-Barr virus: 40 years on. Nat.
Rev. Cancer 4, 757–768.
Zanini, F., Robinson, M.L., Croote, D., Sahoo, M.K., Sanz, A.M., Ortiz-
Lasso, E., Albornoz, L.L., Rosso, F., Montoya, J.G., Goo, L., et al. (2018).
Virus-inclusive single-cell RNA sequencing reveals the molecular signature
of progression to severe dengue. Proc. Natl. Acad. Sci. USA 115, E12363–
E12369.
Zhang, F., Wei, K., Slowikowski, K., Fonseka, C.Y., Rao, D.A., Kelly, S.,
Goodman, S.M., Tabechian, D., Hughes, L.B., Salomon-Escoto, K., et al.;
Accelerating Medicines Partnership Rheumatoid Arthritis and Systemic
Lupus Erythematosus (AMP RA/SLE) Consortium (2019). Defining inflam-
matory cell states in rheumatoid arthritis joint synovial tissues by inte-
grating single-cell transcriptomics and mass cytometry. Nat. Immunol.
20, 928–942.
Cell 181, 1475–1488, June 25, 2020 1487
 ll
Zhang, H., Wang, X., Fu, Z., Luo, M., Zhang, Z., Zhang, K., He, Y., Wan, D.,
Zhang, L., Wang, J., et al. (2020). Potential Factors for Prediction of Disease
Severity of COVID-19 Patients. MedRxiv. https://doi.org/10.1101/2020.03.
20.20039818.
Zheng, G.X.Y., Terry, J.M., Belgrader, P., Ryvkin, P., Bent, Z.W., Wilson, R.,
Ziraldo, S.B., Wheeler, T.D., McDermott, G.P., Zhu, J., et al. (2017).
1488 Cell 181, 1475–1488, June 25, 2020
Article
Massively parallel digital transcriptional profiling of single cells. Nat. Com-
mun. 8, 14049.
Zhu, N., Zhang, D., Wang, W., Li, X., Yang, B., Song, J., Zhao, X., Huang, B.,
Shi, W., Lu, R., et al.; China Novel Coronavirus Investigating and Research
Team (2020). A Novel Coronavirus from Patients with Pneumonia in China,
2019. N. Engl. J. Med. 382, 727–733.
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
santi-mouse TCRb biotin (clone H57-597) Biolegend Cat#:109203; RRID:AB_313426
anti-mouse CD3 biotin (clone 17A2) Biolegend Cat#:100243; RRID:AB_2563946
anti-mouse CD19 biotin (clone 6D5) Biolegend Cat#:115503; RRID:AB_313638
Bacterial and Virus Strains
Vesicular Stomatitis Virus (VSV) Indiana In house N/A
Strain
Lymphocytic choriomeningitis virus In house N/A
(LCMV)- Armstrong (Arm) strain
LCMV-Arm-eGFP In house N/A
Biological Samples
COVID-19 BAL samples Shenzhen Third People’s Hospital N/A
HBV liver sample Shenzhen Third People’s Hospital N/A
Chemicals, Peptides, and Recombinant Proteins
Liberase TL Roche Cat#:5401020001
Dnase I, grade II Roche Cat#:10104159001
Critical Commercial Assays
Chromium Single Cell 3ʹ Reagent Kit (v3 10X Genomics 1000075
chemistry)
Chromium Single Cell V(D)J Reagent Kits 10X Genomics 1000006
(v1 Chemistry)
Deposited Data
Raw data files for the 10X COVID-19 and This paper GEO: GSE145926
HBV patients
Raw data files for the LCMV/VSV single-cell This paper GEO: GSE149443
RNA-seq
Experimental Models: Organisms/Strains
Mouse: C57BL/6 WT Jackson laboratories RRID:IMSR_JAX:000664
Software and Algorithms
R (3.5.0) The R project https://www.r-project.org
Python (3.6.5) Python software foundation https://www.python.org
STAR (2.7.0) Dobin et al., 2013 https://github.com/alexdobin/STAR
Samtools (1.4.0) Li et al., 2009 http://www.htslib.org/download/
StringTie (1.3.5) Pertea et al., 2015 https://ccb.jhu.edu/software/stringtie/
UMI-tools (1.0.0) Smith et al., 2017 https://umi-tools.readthedocs.io/en/latest/
Pagoda2 (0.1.0) Lake et al., 2018 https://github.com/hms-dbmi/pagoda2/
MetaCell (0.3.41) Baran et al., 2019 https://github.com/tanaylab/metacell
Cell Ranger (3.1.0) N/A https://support.10xgenomics.com/
single-cell-gene-expression/software/
pipelines/latest/what-is-cell-ranger
Other
MARS-seq reagents Jaitin et al., 2014 N/A

Cell 181, 1475–1488.e1–e6, June 25, 2020 e1

 ll
Article

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact Ido Amit
(ido.amit@weizmann.ac.il).

Materials Availability
This study did not generate new unique reagents.

Data and Code Availability

The whole Viral-Track pipeline is freely available at https://github.com/PierreBSC/Viral-Track. The datasets generated during this
study were deposited to the Gene Expression Omnibus (GEO) repository with accession codes GEO: GSE145926 and GSE149443.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Mice
C57BL/6 mice were purchased from Jackson Laboratories and bred and housed at the Weizmann Institute of Science animal facility,
under specific pathogen-free conditions. Female mice, 6-8 weeks of age, were used for all experiments. Experimental protocols were
approved by the Weizmann Institute of Science Ethics Committee and were performed according to institutional guidelines.

LCMV/VSV infections
For LCMV infection, 1x105 Focus-Forming Units (FFUs) of the LCMV-Arm strain were injected. For VSV, 1x105 Plaque-Forming Units
(PFUs) of the VSV Indiana strain were used. Mice were anesthetized and viruses administered by intradermal injection into the ear
pinna. 24h later, mice were sacrificed and auricular LN were harvested.

Subjects
This study was conducted according to the principles expressed in the Declaration of Helsinki. Ethical approval was obtained from
the Research Ethics Committee of Shenzhen Third People’s Hospital. All participants provided written informed consent for sample
collection and subsequent analyses.

METHOD DETAILS

Lymph Node MARS-seq data generation

To prepare single cell suspensions for MARS-seq and flow cytometry, auricular LNs were digested in IMDM containing 100mg/mL
Liberase TL and 100mg/mL DNase I (both from Roche, Germany) for 20 minutes at 37C. In the last 5 minutes of incubation, EDTA was
added at a final concentration of 10mM. Cells were collected, filtered through a 70mm cell strainer, washed with IMDM and main-
tained strictly at 4C. Cells were sorted with FACSARIA-FUSION (BD Biosciences, San Jose, CA). Prior to sorting, all samples
were filtered through a 70-mm nylon mesh. Isolated cells were single cell sorted into 384-well cell capture plates containing 2 mL
of lysis solution and barcoded poly(T) reversetranscription (RT) primers for single-cell RNA-seq (Jaitin et al., 2014). Four empty wells
were kept in each 384-well plate as a no-cell control for data analysis. Immediately after sorting, each plate was spun down to ensure
cell immersion into the lysis solution, snap frozen on dry ice, and stored at –80C until processing Single-cell RNA-seq libraries were
prepared as previously described (Jaitin et al., 2014). In brief, mRNA from single cells sorted into capture plates were barcoded and
converted into cDNA and then pooled using an automated pipeline. The pooled sample was linearly amplified by T7 in vitro transcrip-
tion, and the resulting RNA was fragmented and converted into a sequencing-ready library by tagging the samples with pool barc-
odes and Illumina sequences during ligation, RT, and PCR. Each pool of cells was tested for library quality and concentration as
described previously (Jaitin et al., 2014).

Influenza MARS-seq data generation

Full description of the protocol used to generate the Influenza A lung data can be found in Steuerman et al. (2018). Influenza PR8
H1N1 influenza virus (A/Puerto Rico/8/34) was cultivated in hen egg anion. 40mL of diluted virus (6x103 PFU per mouse) were inoc-
ulated intranasaly to the mice, or 40mL of PBS for the control mice. Mice were killed 48 or 72h post infection and the lung perfused.
Immune and none-immune cells were then extracted using two different extraction protocols before being single-cell sorted in 384-
well plates and sequenced using the original MARS-seq protocol (Jaitin et al., 2014).

LCMV spleen MARS-seq data generation

Description for the full protocol used to generate the NICHE-Seq spleen data can be found in Medaglia et al. (2017). Briefly female
mice received 1x106 FFU of LCMV-Arm or LCMV-Arm-eGFP in the footpad. 72 hours after injection, spleens were harvested and
forced through a 70mm mesh to form a single-cell suspension. Cells were then single-cell sorted using a SORP-aria into 384-well

e2 Cell 181, 1475–1488.e1–e6, June 25, 2020

 ll
Article

plates containing lysis buffer before processing the plate according to the MARS-seq protocol (Jaitin et al., 2014). All infectious work
was performed in designated Biosafety Level 2 (BSL-2) and BSL-3 workspaces in accordance with institutional guidelines

10X HBV liver data generation

The approximately 1 cm long Liver biopsy was homogenized by mincing with scissors into smaller pieces (~0.5 mm2 per piece). Then
the tissue was transferred into 10 mL of enzyme mix consisting of 0.3 mg/ml collagenase type IV (Sigma, C9891) and DNase I (Sigma,
D5025) for mild enzymatic digestion for 1 h at 37 C while shaking. 5 mL of Dulbecco’s phosphate-buffered saline (DPBS, Thermo,
14190250) supplemented with 5% FBS was added to interrupt digestion and dissociated cells in suspension were passed through a
40 mm strainer and centrifuged at 300 g for 5 min at 4 C. Erythrocytes were lysed using Ammonium-Chloride-Potassium (ACK,
Thermo, A1049201), and finally cells were re-suspended in DPBS supplemented with 1% FBS at the concentration of 2, 000
cells/ml for scRNA-Seq. The single-cell capturing and downstream library constructions were performed using the Chromium Single
Cell 30 V3 library preparation kit according to the manufacturer’s protocol (10x Genomics). Full-length cDNA along with cell-barcode
identifiers were PCR-amplified and sequencing libraries were prepared and normalized to 3 nM. The constructed library was
sequenced on BGI MGISEQ-2000 platform. The Cell Ranger Software Suite (Version 3.1.0) was then used to perform sample de-mul-
tiplexing, barcode processing and single-cell 30 UMI counting with human GRCh38 as the reference genome

10X COVID-19 data generation

20 mL of BALF was obtained and placed on ice. BALF was processed within 2 hours and all operations were performed in BSL-3
laboratory. By passing BALF through a 100 mm nylon cell strainer to filter out lumps, the supernatant was centrifuged and the cells
were re-suspended in the cooled RPMI 1640 complete medium. Then the cells were counted in 0.4% trypan blued, centrifuged and
re-suspended at the concentration of 2 3 106 /ml for further use. Total 11 ml of single cell suspension and 40 ml barcoded Gel Beads
were loaded to Chromium Chip A to generate single-cell gel bead-in-emulsion (GEM). The poly-adenylated transcripts were reverse-
transcribed later. The single-cell capturing and downstream library constructions were performed using the Chromium Single Cell 50
library preparation kit according to the manufacturer’s protocol (10x Genomics). Full-length cDNA along with cell-barcode identifiers
were PCR-amplified and sequencing libraries were prepared and normalized to 3 nM. The constructed library was sequenced on BGI
MGISEQ-2000 platform. Each sample was sequenced on a different sequencing run to avoid contamination between samples. The
Cell Ranger Software Suite (Version 3.1.0) was then used to perform sample de-multiplexing, barcode processing and single-cell 50
UMI counting with human GRCh38 as the reference genome. A more extensive description of the data generation process can be
found in Liao et al. (2020).

QUANTIFICATION AND STATISTICAL ANALYSIS

Read mapping/alignment
Reads were aligned using STAR 2.7.0 (Dobin et al., 2013) in the two-pass mode using the following parameters:–runThreadN was set
to 14,–outSAMattributes to ‘NH HI AS nM NM XS’,–outSAMtype to ‘BAM SortedByCoordinate’,–outFilterScoreMinOverLread to
0.6,–outFilterMatchNminPverLread to 0.6, and–twopassMode to ‘Basic’.

Viral database and STAR Index building

As STAR performance drastically dropped when the reference index contains more than 10.000 scaffold/chromosomes, we decided
to base our analysis on the limited, but high-quality, viruSITE database (Stano et al., 2016), derived from the NCBI Refseq database.
The corresponding FASTA file was downloaded from the viruSITE website (http://www.virusite.org/archive/2019.1/genomes.fasta.
zip). STAR indexes were build for both human and mouse samples using respectively the GRCh38 (hg38) and GRCm38 (mm10) refer-
ence genomes in addition with the whole viruSITE database. Both reference genomes were downloaded at http://www.ensembl.
org//useast.ensembl.org/info/data/ftp/index.html?redirectsrc=//www.ensembl.org%2Finfo%2Fdata%2Fftp%2Findex.html.
For the analysis of COVID-19 patients we added the official SARS-CoV-2 reference genome from the Refseq database
(NC_045512.2) as it has not been added to the viruSITE database yet. In total this database contains 11988 viral segments from
9431 different viruses.

Processing and filtering of the BAM files

We empirically observed that viral genome sequences can contain highly repetitive subsequences and can therefore create false
positive signal. Moreover, some viral genes can share a significant similarity with host genes and also generate mapping artifacts.
To remove those, we implemented a strict filtering approach where for each viral segment, a list of mapping features are measured
and used to estimate the quality of the mapping.
Following the alignment, the resulting BAM files were processed using the samtools toolbox (Li et al., 2009): first the BAM files were
indexed using the samtools index command. Virus segment with more than 50 mapped reads were detected using the samtools idx-
stats command and a unique bam file was then created for each of the viral segment using the samtools view command.
Each viral bam files were then loaded into an R environment using the readGAlignments() function from the GenomicAlignments
package. Various features were then extracted to assess the quality of the mapping:

Cell 181, 1475–1488.e1–e6, June 25, 2020 e3

 ll
Article
d The length of the longest mapped contig computed using the coverage() function.
d The percentage of the viral segment that is mapped, also computed using the coverage() function.
d The mean sequencing quality of the mapped reads.
d The number and percentage of uniquely mapped reads.
d The mean sequenced entropy of the mapped reads defined as follows: for each mapped read each nucleotide frequency was
extracted using the alphabetFrequency() function of the Biostring package and averaged over the reads. Then the correspond-
ing Shanon entropy was computed using napierian logarithm.

Empirically we determined that a mean sequence entropy bigger than 1.2, a coverage bigger than 5% and the longest contig bigger
than three times the mean read length is sufficient to consider a viral segment to be present. This filter configuration eliminated all
manually identified artifacts in the various benchmarked datasets and was used unchanged in the HBV and COVID-19 patient
data analysis.
When using this strategy, we observed two different kinds of ‘contamination’:
-
d the first one consists of the detection of retroviruses specific to the sequenced host species: this is likely due to the expression
of host endogenous retro-viral elements that highly similar to ‘real’ retroviruses.
-
d the second is the presence of a plant virus, the Tomato brown rugose fruit virus: this is an emerging virus that infects tomatoes
and peppers and is endemic in Israel and Jordan. It is highly contagious and spreads easily. We detected this virus only in sam-
ples sequenced in Rehovot (Israel) suggesting that it was due to an airborne contamination.

To improve computation speed, this step was parallelised using the doParallel R package.

Transcript reconstruction
As viral genomes are poorly annotated, we decided to systemically reconstruct the transcriptome of each viral segment detected
using the transcript assembler StringTie (Pertea et al., 2015). StringTie was used with default parameter except the minimum isoform
abundance parameter -f which was set to 0.01 to detect lowly abundant transcripts and the minimal distance between two transcript
-g set to 5.

MARS-seq data demultiplexing and UMI count

In order to have a UMI-counting procedure adapted to viral genomes, i.e that distinguish spliced and un-spliced RNA molecule, we
developed an in-house R script based on the GenomicRanges, GenomicAlignments and GenomicFeatures packages that used the
same strategy as the commercial CellRanger toolkit. Briefly cell barcodes were extracted and compared with a cell barcode whitelist
provided by the MARS-seq2 demultiplexing pipeline (Keren-Shaul et al., 2019): cell barcode that belong to the whitelist were kept
while cell barcodes that did not belong to the whitelist but that has a highly similar barcode (Hamming distance equal to one,
computed using the stringdist() function from the stringdist package) were corrected and kept. UMIs were also extracted and
mono-nucleotide UMIs filtered out. Hamming distances between UMIs assigned to the same cell and the same gene were then
computed similarly to cell barcodes and UMIs with a Hamming distance equal to one were aggregated and considered as redundant
UMIs. Lastly the mapping file was loaded using the readGAlignments() function from the GenomicAlignments package and reads
were assigned to a specific viral gene using the findOverlaps() function from the same package. In case the read mapped to a given
viral transcript but was not assigned to any viral gene, it was considered as coming from an un-spliced viral RNA molecule.

Drop-seq and 10X data download, pre-processing and demultiplexing

Fastq files were downloaded through the SRA Explorer tool (https://sra-explorer.info/#). Identification and correction of cellular bar-
code, as well as UMI demultiplexing was performed using UMI-tools 1.0.0 (Smith et al., 2017). First, cell barcodes were extracted and
a putative whitelist computed using the umi_tools whitelist command with the parameters ‘–stdin —bc-pattern =
CCCCCCCCCCCCCCCCNNNNNNNNNN–log2stderr ’ for the 10X data. For Drop-Seq data the same command is used except
the–bc-pattern option set to CCCCCCCCCCCCNNNNNNNN. Collapsing of the UMIs is performed using the command umi_tools
extract with parameters ‘—bc-pattern = CCCCCCCCCCCCCCCCNNNNNNNNNN —stdin —filter-cell-barcode’ on the 10X data
and with the same command for Drop-seq data except for the–bc-pattern option set to ‘CCCCCCCCCCCCNNNNNNNN’. Following
the mapping of the reads to viral genomes and transcript assembly, the mapped reads were assigned to transcripts using the R pack-
age Rsubread through the function featureCounts() with default parameters. The command umi_tools count is then used to compute
the final expression table with the following parameters:–per-gene–gene-tag = XT–assigned-status-tag = XS–per-cell.

Analysis of the MARS-seq spleen LCMV dataset

High-level analysis were performed using the R-based Pagoda2 pipeline (https://github.com/hms-dbmi/pagoda2/) (Lake et al., 2018)
in addition to an in-house R script. Briefly UMI table were loaded and cells with less than 350 UMIs were removed. Lowly abundant
genes (less than 100 UMIs) were also removed from analysis. Analysis of the filtered dataset was then performed similarly to our pre-
vious paper (Blecher-Gonen et al., 2019) by using the 1500 most variant genes and 100 PCs for dimensionality reduction. kNN graph
was build with a parameter K equal to 30 and Louvain’s method used for clustering. Cluster marker genes were computed by using

e4 Cell 181, 1475–1488.e1–e6, June 25, 2020

 ll
Article

the getdiffGenes function with default parameters. Data were visualized using UMAP (McInnes et al., 2018) implemented by the uwot
package.

Analysis of the 10X HBV liver dataset

High-level analysis were performed using the R-based Pagoda2 pipeline (https://github.com/hms-dbmi/pagoda2/) (Lake et al., 2018)
in addition to an in-house R script. Briefly UMI table were loaded and cells with less than 1000 UMIs were removed. Lowly abundant
genes (less than 50 UMIs) were also removed from analysis. Analysis of the filtered dataset was then performed similarly to our pre-
vious paper (Blecher-Gonen et al., 2019) by using the 1000 most variant genes and 100 PCs for dimensionality reduction. kNN graph
was build with a parameter K equal to 30 and Louvain’s method used for clustering. Cluster marker genes were computed by using
the getdiffGenes function with default parameters. Data were visualized using UMAP (McInnes et al., 2018) implemented by the uwot
package.

Analysis of the COVID-19 BAL dataset

Upstream processing of reads was done with the CellRanger toolkit, resulting in a UMI table of 75,790 cells with a median UMI count
of 2,442, and a median of 868 genes per cell. Cells with less than 500 UMI, or more than 50% mitochondrial genes were excluded.
We used the MetaCell package (Baran et al., 2019) to group single cells from all patients into groups of transcriptionally homoge-
neous groups, termed metacells . We first removed mitochondrial genes, ERCC, and the diverse immunoglobulin genes (IGH, IGK,
and IGL).
Gene features for metacell covers were selected using the parameter Tvm = 0.4, total umi > 30, and more than 4 UMI in at least 3
cells (using the functions mcell_gset_filter_varmean, and mcell_gset_filter_cov). We excluded gene features associated with the cell
cycle, stress response, type I interferon, and batch-specific genes via a clustering approach (using the functions mcell_mat_rpt_cor_-
anchors and mcell_gset_split_by_dsmat). To this end we first identified all genes with a correlation coefficient of at least 0.1 for one of
the anchor genes TOP2A, MKI67, PCNA, MCM4, UBE2C, STMN1 (cell cycle), HSPA1B, HSPA1A, DNAJB1, HSPB1, HSPA6, FOS,
JUN, CCL4, CCL4L2, MT1E, MT1X, MT1F, TYMS, GADPH, DUT, HMGB2 (stress and batch effect), IFIT1, IFIT3, OASL, IRF7, IRF1,
STAT1, and STAT3 (type I IFN). We then hierarchically clustered the correlation matrix between these genes (filtering genes with low
coverage and computing correlation using a down-sampled UMI matrix) and selected the gene clusters that contained the above
anchor genes. We thus retained 402 genes as features (Table S3). We used metacell to build a kNN graph, perform boot-strapped
co-clustering (500 iterations; resampling 70% of the cells in each iteration), and derive a cover of the co-clustering kNN graph (K =
100). Outlier cells featuring gene expresssion higher than 4-fold than the geometric mean in the metacells in at least one gene were
discarded.
Annotation of the metacell model was done using the metacell confusion matrix and analysis of marker genes. Detailed annotation
within the myeloid, lymphoid and epithelial compartments was performed using hierarchical clustering of the metacell confusion ma-
trix (Figure S3A) and supervised analysis of enriched genes. Metacells enriched for markers from more than one lineage (either T
(TRBC2), myeloid (S100A8, C1QB), epithel (KRT18), and plasma cells (XBP1)) were marked as doublets and discarded from further
analysis. We additionally discarded metacells of erythrocytes or plasma cells from further analysis.
To derive cell cycle and type I interferon response co-expressed gene modules, we used a clustering-approach as described in the
previous paragraphs (using the functions mcell_mat_rpt_cor_anchors and mcell_gset_split_by_dsmat) on a set of cell cycle and
interferon genes. We clustered, and manually inspected the resulting clusters, retrieving 72 cell-cycle related and 65 interferon
related genes (Table S3).
To extract proportion of proliferating cells (Figure 3G), we calculated for each cells the number of cell-cycle related transcripts per
1,000 UMI. Cells with more than 8 transcripts were determined proliferating.

Testing for infection specificity in COVID-19 BAL dataset

To test for SARS-CoV-2 infection specificity in different cell populations, we computed for each metacell the total number of host
UMIs (hUMI) and viral UMIs (vUMI) in the three severe patients (S1-3). We then computed for each metacell its expected vUMI
cout, based on its total UMI count (hUMI + vUMI) and the total vUMI proportion across all cells. Figure 4C shows log2 fold change
between the observed and expected UMI in each metacell, after adding a regularization factor ( = 5) for each factor. Log2 fold change
for the 27 subsets in Figure 3A, and calculated for each severe patient separately is shown in Table S2.
Testing for hMPV infection specificity was done in a similar manner. However, since UMI distribution across cells was abundant
and heavy-tailed, we computed for each metacell the expected number of vUMI+ cells instead of its total vUMI count. A cell was
determined vUMI+ if it had more than 10 viral UMI, as determined by automatic thresholding (Figure 4F). Figure 4G shows log2
fold change between the observed and expected vUMI+ cells in each metacell, after adding a regularization factor ( = 5) for each
factor.

Dichotomized differential gene expression analysis

ScRNA-seq data are intrinsically noisy data with a large proportion of zeros values (previously called dropouts) due to limited sam-
pling of the initial mRNA molecule pool. In addition, cell library size is a major cofounder variable, even after common normalization

Cell 181, 1475–1488.e1–e6, June 25, 2020 e5

 ll
Article

procedures such as TPM, especially for lowly expressed genes (Hafemeister and Satija, 2019). We therefore improved the method
used in our former paper (Blecher-Gonen et al., 2019) that was based on logistic regression.
Briefly our method is based on the global trend of the field that consists in sequencing large amounts of cells but with a limited
sequencing depth. Such approach will produce mostly ‘binary’ data and seem to be represent the best compromise on a cost/effi-
ciency point of view (Svensson et al., 2019). So far, several statistical models have been used to model and analyze scRNA-seq count
data, most of them being based on the zero-inflated negative-binomial (ZINB) distribution (Finak et al., 2015; Kharchenko et al., 2014).
However, recent studies suggested that those models are too complex and introduce artificial complexity (Silverman et al., 2018;
Svensson, 2020; Townes et al., 2019). We hypothesize that with such binary data, current models will not fit properly and more suited
ones need to be developed.
We therefore developed a new approach based on the binomial complementary Log-log regression (cloglog model): once a given
group of cells has been isolated, through Louvain’s clustering for instance (Blondel et al., 2008), we first dichotomized gene expres-
sion (if the normalized expression is bigger than 0 the gene is considered as expressed) and then computed a binomial Generalized
Linear Model (GLM) with a complementary log log link function (cloglog) using the glm() R function. To mitigate the variation of the
library size as well as the global effect of the infection (bystander effect), we include both variables in the regression model. The cor-
responding p value are then computed using a Likelihood Ratio Test (LRT) and then corrected using Benjamini Hochberg correction
(Benjamini and Hochberg, 1995).
For a more comprehensive description of the approach please see Methods S1.

Automate thresholding to detect HBV and hMPV infected cells

In the case of the HBV and hMPV infections, we observed that cells could contain from one to several thousands UMIs. In order to
know which cells were really infected and which one contain viral UMIs due to ambient contamination, we decided to apply Otsu’s
thresholding after logarithmic transformation. Otsu’s method was implemented using an in-house R script (Otsu, 1979).

Gene set enrichment analysis

Gene set enrichment analysis was performed using the online GSEA tool https://www.gsea-msigdb.org/gsea/index.jsp (Liberzon
et al., 2015; Subramanian et al., 2005). The enrichment analysis was performed using the Hallmark and Gene Ontology biological
process databases. False detection rate was set to 0.05. Only the top 10 most enriched terms were reported.

e6 Cell 181, 1475–1488.e1–e6, June 25, 2020

 ll
Article

Supplemental Figures

(legend on next page)

 ll
Article

Figure S1. Benchmarking of Viral-Track on Diverse Infection Models, Related to Figure 1

A. Graph chart representing the different steps of the Viral-Track pipeline. B-D. Results of Viral-Track analysis performed on LCMV spleen, LCMV lymph node and
VSV lymph node datasets, respectively. Viral segments with more than 50 mapped reads are plotted. (E). Number of detected LCMV (left panel) and VSV (right
panel) reads in the different samples from the lymph node experiment. F. Results of Viral-Track analysis performed on the in-vitro HSV-1 data. G. Quantification of
the number of HSV-1 reads in HSV-1 infected and control samples. (H). Results of Viral-Track analysis performed on the in-vitro HIV data. I. Quantification of the
number of HIV reads in HIV infected and control samples. J. UMAP plot of the liver HBV data, dots are colored by cell subset assignment based on Louvain
clustering. K. UMAP plot of the liver HBV data. infected cells are colored in orange and bystander cells in gray.
 ll
Article

Figure S2. Comparison of Viral-Track Performance to Fluorescence Tagging Techniques, Related to Figure 2
A. Proportion of vUMI+ cells from total spleen and the LCMV-GFP+ population B. UMAP plot of the spleen LCMV data, spots are colored based on Louvain
clustering. C. UMAP plot of the spleen LCMV data, bystander cells are colored in gray, vUMI+ cells are colored in red and GFP+ cells in green. D. Mean gene
expression in bystander and infected MZB cells. Genes with a log2FC bigger than 1 or lower than 1 and a corrected p value lower than 0.01 are colored in
orange.
ll
Article

(legend on next page)

 ll
Article

Figure S3. Detailed Molecular and Cellular Profiling of COVID-19 BAL Samples, Related to Figure 3
A. The confusion matrix of the MetaCell model shown in Figure 3A. Entries denote for each pair of metacells the propensity of cells from both metacells to be
clustered together in a bootstrap analysis. B-D. Gene expression profiles of cells belonging to the epithelial (B), lymphoid (C), and myeloid (D). In A-D, color bars
indicate association to 27 cell subsets depicted in Figure 3A. E-G. Quantification of the frequency of specific cell subsets in the myeloid (E), lymphoid (F), and
epithelial (G) compartments, across the nine patients. Diamond marks patient S1, co-infected with the human Metapneumovirus (Figures 4D-4H). Horizontal lines
indicate mean frequency. (H). Projection of IL6 and IL8 (CXCL8) expression on the 2D map shown in Figure 4A. Colors represent expression quantiles.
 ll
Article

Figure S4. Viral-Track Performance on COVID-19 BAL Samples, Related to Figure 4

A. Results of Viral-Track analysis performed on samples with highest viral load (patients S2 and S3). B. Mean normalized expression of ACE2, TMPRSS2 and BSG
across the 27 cell subsets C. Log2 fold change between vUMI+ and vUMI- SPP1+ monocyte-derived macrophages in patient S2 (x axis) and patient S3 (y axis). D.
Relation between total human and viral UMIs in cells from patient S1. E. Projection of cells from patient S1, co-infected with hMPV, on the metacell map from
Figure 3A. F. Enrichment analysis of the downregulated genes in hMPV infected monocytes. G. Number of hMPV UMIs in cells producing type I IFN or not. P value
was computed by fitting a logistic regression predicting if a cell would produce type I IFN using total host and viral UMIs.
 Article

Targets of T Cell Responses to SARS-CoV-2

Coronavirus in Humans with COVID-19 Disease and
Unexposed Individuals
Graphical Abstract Authors
Alba Grifoni, Daniela Weiskopf,
Sydney I. Ramirez, ..., Davey M. Smith,
Shane Crotty, Alessandro Sette

Correspondence
shane@lji.org (S.C.),
alex@lji.org (A.S.)

In Brief
An analysis of immune cell responses to
SARS-CoV-2 from recovered patients
identifies the regions of the virus that is
targeted and also reveals cross-reactivity
with other common circulating
coronaviruses

Highlights
d Measuring immunity to SARS-CoV-2 is key for
understanding COVID-19 and vaccine development

d Epitope pools detect CD4+ and CD8+ T cells in 100% and

70% of convalescent COVID patients

d T cell responses are focused not only on spike but also on M,

N, and other ORFs

d T cell reactivity to SARS-CoV-2 epitopes is also detected in

non-exposed individuals

Grifoni et al., 2020, Cell 181, 1489–1501

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.015 ll
 ll

Article
Targets of T Cell Responses to SARS-CoV-2
Coronavirus in Humans with COVID-19
Disease and Unexposed Individuals
Alba Grifoni,1 Daniela Weiskopf,1 Sydney I. Ramirez,1,2 Jose Mateus,1 Jennifer M. Dan,1,2
Carolyn Rydyznski Moderbacher,1 Stephen A. Rawlings,2 Aaron Sutherland,1 Lakshmanane Premkumar,3
Ramesh S. Jadi,3 Daniel Marrama,1 Aravinda M. de Silva,3 April Frazier,1 Aaron F. Carlin,2 Jason A. Greenbaum,1
Bjoern Peters,1,2 Florian Krammer,4 Davey M. Smith,2 Shane Crotty,1,2,5,* and Alessandro Sette1,2,5,6,*
1Center for Infectious Disease and Vaccine Research, La Jolla Institute for Immunology, La Jolla, CA 92037, USA
2Department of Medicine, Division of Infectious Diseases and Global Public Health, University of California, San Diego, La Jolla, CA
92037, USA
3Department of Microbiology and Immunology, University of North Carolina School of Medicine, Chapel Hill, NC 27599-7290, USA
4Department of Microbiology, Icahn School of Medicine at Mount Sinai, New York, NY, USA
5These authors contributed equally
6Lead Contact

*Correspondence: shane@lji.org (S.C.), alex@lji.org (A.S.)

https://doi.org/10.1016/j.cell.2020.05.015

SUMMARY

Understanding adaptive immunity to SARS-CoV-2 is important for vaccine development, interpreting coro-
navirus disease 2019 (COVID-19) pathogenesis, and calibration of pandemic control measures. Using HLA
class I and II predicted peptide ‘‘megapools,’’ circulating SARS-CoV-2-specific CD8+ and CD4+ T cells
were identified in 70% and 100% of COVID-19 convalescent patients, respectively. CD4+ T cell responses
to spike, the main target of most vaccine efforts, were robust and correlated with the magnitude of the anti-
SARS-CoV-2 IgG and IgA titers. The M, spike, and N proteins each accounted for 11%–27% of the total CD4+
response, with additional responses commonly targeting nsp3, nsp4, ORF3a, and ORF8, among others. For
CD8+ T cells, spike and M were recognized, with at least eight SARS-CoV-2 ORFs targeted. Importantly, we
detected SARS-CoV-2-reactive CD4+ T cells in 40%–60% of unexposed individuals, suggesting cross-
reactive T cell recognition between circulating ‘‘common cold’’ coronaviruses and SARS-CoV-2.

INTRODUCTION different depending on whether SARS-CoV-2 infection creates

COVID-19 is a worldwide emergency. The first cases occurred in
December 2019, and now more than 240,000 deaths and
3,000,000 cases of SARS-CoV-2 infection have been reported
worldwide as of May 1st (Dong et al., 2020; Wu and McGoogan,
2020). Vaccines against SARS-CoV-2 are just beginning devel-
opment (Amanat and Krammer, 2020; Thanh Le et al., 2020).
An understanding of human T cell responses to SARS-CoV2 is
lacking, due to the rapid emergence of the pandemic. There is
an urgent need for foundational information about T cell
responses to this virus.
The first steps for such an understanding are the ability to quan-
tify the virus-specific CD4+ and CD8+ T cells. Such knowledge is
of immediate relevance, as it will provide insights into immunity
and pathogenesis of SARS-CoV-2 infection, and the same knowl-
edge will assist vaccine design and evaluation of candidate vac-
cines. Estimations of immunity are also central to epidemiological
model calibration of future social distancing pandemic control
measures (Kissler et al., 2020). Such projections are dramatically
substantial immunity, and whether any cross-reactive immunity
exists between SARS-CoV-2 and circulating seasonal ‘‘common
cold’’ human coronaviruses. Definition and assessment of human
antigen-specific SARS-CoV2 T cell responses are best made
with direct ex vivo T cell assays using broad-based epitope pools
and assays capable of detecting T cells of any cytokine polariza-
tion. Herein, we have completed such an assessment with blood
samples from COVID-19 patients.
There is also great uncertainty about whether adaptive im-
mune responses to SARS-CoV-2 are protective or pathogenic,
or whether both scenarios can occur depending on timing,
composition, or magnitude of the adaptive immune response.
Hypotheses range the full gamut (Peeples, 2020), based on avail-
able clinical data from severe acute respiratory disease
syndrome (SARS) or Middle East respiratory syndrome (MERS)
(Alshukairi et al., 2018; Wong et al., 2004; Zhao et al., 2017) or an-
imal model data with SARS in mice (Zhao et al., 2009, 2010,
2016), SARS in non-human primates (NHPs) (Liu et al., 2019; Ta-
kano et al., 2008) or feline infectious peritonitis virus (FIPV) in cats

Cell 181, 1489–1501, June 25, 2020 ª 2020 Elsevier Inc. 1489
 ll
(Vennema et al., 1990). Protective immunity, immunopathogene-
sis, and vaccine development for COVID-19 are each briefly dis-
cussed below, related to introducing the importance of defining
T cell responses to SARS-CoV-2.
Based on data from SARS patients in 2003–2004 (caused by
SARS-CoV, the most closely related human betacoronavirus to
SARS-CoV-2), and based on the fact that most acute viral infec-
tions result in development of protective immunity (Sallusto et al.,
2010), a likely possibility has been that substantial CD4+ T cell,
CD8+ T cell, and neutralizing antibody responses develop to
SARS-CoV-2, and all contribute to clearance of the acute infec-
tion, and, as a corollary, some of the T and B cells are retained
long term (i.e., multiple years) as immunological memory and
protective immunity against SARS-CoV-2 infection (Guo et al.,
2020b; Li et al., 2008). However, a contrarian viewpoint is also
legitimate. While most acute infections result in the development
of protective immunity, available data for human coronaviruses
suggest the possibility that substantive adaptive immune re-
sponses can fail to occur (Choe et al., 2017; Okba et al., 2019;
Zhao et al., 2017) and robust protective immunity can fail to
develop (Callow et al., 1990). A failure to develop protective im-
munity could occur due to a T cell and/or antibody response of
insufficient magnitude or durability, with the neutralizing anti-
body response being dependent on the CD4+ T cell response
(Crotty, 2019; Zhao et al., 2016). Thus, there is urgent need to un-
derstand the magnitude and composition of the human CD4+
and CD8+ T cell responses to SARS-CoV-2. If natural infection
with SARS-CoV-2 elicits potent CD4+ and CD8+ T cell responses
commonly associated with protective antiviral immunity, COVID-
19 is a strong candidate for rapid vaccine development.
Immunopathogenesis in COVID-19 is a serious concern (Cao,
2020; Peeples, 2020). It is most likely that an early CD4+ and
CD8+ T cell response against SARS-CoV-2 is protective, but an
early response is difficult to generate because of efficient innate
immune evasion mechanisms of SARS-CoV-2 in humans
(Blanco-Melo et al., 2020). Immune evasion by SARS-CoV-2 is
likely exacerbated by reduced myeloid cell antigen-presenting
cell (APC) function or availability in the elderly (Zhao et al., 2011).
In such cases, it is conceivable that late T cell responses may
instead amplify pathogenic inflammatory outcomes in the pres-
ence of sustained high viral loads in the lungs, by multiple hypo-
thetical possible mechanisms (Guo et al., 2020a; Li et al., 2008;
Liu et al., 2019). Critical (ICU) and fatal COVID-19 (and SARS) out-
comes are associated with elevated levels of inflammatory cyto-
kines and chemokines, including interleukin-6 (IL-6) (Giamarel-
los-Bourboulis et al., 2020; Wong et al., 2004; Zhou et al., 2020)
Vaccine development against acute viral infections classi-
cally focuses on vaccine-elicited recapitulation of the type of
protective immune response elicited by natural infection.
Such foundational knowledge is currently missing for
COVID-19, including how the balance and the phenotypes of
responding cells vary as a function of disease course and
severity. Such knowledge can guide selection of vaccine
strategies most likely to elicit protective immunity against
SARS-CoV-2. Furthermore, knowledge of the T cell responses
to COVID-19 can guide selection of appropriate immunolog-
ical endpoints for COVID-19 candidate vaccine clinical trials,
which are already starting.
1490 Cell 181, 1489–1501, June 25, 2020
Article
Limited information is also available about which SARS-CoV-2
proteins are recognized by human T cell immune responses. In
some infections, T cell responses are strongly biased toward
certain viral proteins, and the targets can vary substantially be-
tween CD4+ and CD8+ T cells (Moutaftsi et al., 2010; Tian
et al., 2019). Knowledge of SARS-CoV-2 proteins and epitopes
recognized by human T cell responses is of immediate rele-
vance, as it will allow for monitoring of COVID-19 immune re-
sponses in laboratories worldwide. Epitope knowledge will also
assist candidate vaccine design and facilitate evaluation of vac-
cine candidate immunogenicity. Almost all of the current COVID-
19 vaccine candidates are focused on the spike protein.
A final key issue to consider in the study of SARS-CoV-2 im-
munity is whether some degree of cross-reactive coronavirus
immunity exists in a fraction of the human population, and
whether this might influence susceptibility to COVID-19 disease.
This issue is also relevant for vaccine development, as cross-
reactive immunity could influence responsiveness to candidate
vaccines (Andrews et al., 2015).
In sum, the ability to measure and understand the human CD4+
and CD8+ T cell responses to SARS-CoV-2 infection is a major
knowledge gap currently impeding COVID-19 vaccine develop-
ment, interpretation of COVID-19 disease pathogenesis, and
calibration of future social distancing pandemic control
measures.
RESULTS
SARS-CoV-2 Peptides and Predicted Class I and Class II
Epitopes
We recently predicted SARS-CoV-2 T cell epitopes utilizing
the Immune Epitope Database and Analysis Resource (IEDB)
(Dhanda et al., 2019; Vita et al., 2019). Utilizing bioinformatic
approaches, we identified specific peptides in SARS-CoV-2
with increased probability of being T cell targets (Grifoni
et al., 2020). We previously developed the megapool (MP)
approach to allow simultaneous testing of large numbers of
epitopes. By this technique, numerous epitopes are solubi-
lized, pooled, and re-lyophilized to avoid cell toxicity problems
(Carrasco Pro et al., 2015). These MPs have been used in hu-
man T cell studies of a number of indications, including al-
lergies (Hinz et al., 2016), tuberculosis (Lindestam Arlehamn
et al., 2016), tetanus, pertussis (Bancroft et al., 2016; da Silva
Antunes et al., 2017), and dengue virus, for both CD4+ and
CD8+ T cell epitopes (Grifoni et al., 2017; Weiskopf et al.,
2015). Here, we generated MPs based on predicted SARS-
CoV-2 epitopes. Specifically, one MP corresponds to 221 pre-
dicted HLA class II CD4+ T cell epitopes (Grifoni et al., 2020)
covering all proteins in the viral genome, apart from the spike
(S) antigen (CD4_R MP). The prediction strategy utilized is
geared to capture 50% of the total response (Dhanda
et al., 2018; Paul et al., 2015) and was designed and validated
to predict dominant epitopes independently of ethnicity and
HLA polymorphism. This approach takes advantage of the
extensive cross-reactivity and repertoire overlap between
different HLA class II loci and allelic variants to predict pro-
miscuous epitopes, capable of binding many of the most
common HLA class II prototypic specificities (Greenbaum
 ll
Article

Table 1. Participant Characteristics mind in terms of comparison of the magnitude of the CD4+
T cell responses to those pools.
Unexposed (n = 20) COVID-19 (n = 20)
In the case of CD8 epitopes, since the overlap between
20–66 20–64
different HLA class I allelic variants and loci is more limited to
(median = 31, (median = 44,
Age (years) IQR = 21) IQR = 9)
specific groups of alleles, or supertypes (Sidney et al., 2008),
we targeted a set of the 12 most prominent HLA class I A and
Gender
B alleles, which together allow broad coverage (>85%) of the
Male (%) 35% (7/20) 45% (9/20)
general population. Two class I MPs were synthesized based
Female (%) 65% (13/20) 55% (11/20) on epitope predictions for those 12 most common HLA A and
Residency B alleles (Grifoni et al., 2020), which collectively encompass
California (%) 95% (19/20) 100% (20/20) 628 predicted HLA class I CD8+ T cell epitopes from the entire
USA, Non-California 5% (1/20) 0% (0/20) SARS-CoV-2 proteome (CD8 MP-A and MP-B).
(%)
Sample Collection March 2015– March–April 2020 Immunological Phenotypes of Recovered COVID-19
Date March 2018 Patients
SARS-CoV-2 PCR N/A 100% (16/16 tested) To test for the generation of SARS-CoV-2 CD4+ and CD8+ T cell
Positivity responses following infection, we initially recruited 20 adult pa-
Antibody Test N/A 90% (18/20)
tients who had recovered from COVID-19 disease (Table 1).
Positivitya We also utilized peripheral blood mononuclear cell (PBMC) and
plasma samples from local healthy control donors collected in
Disease Severityb
2015–2018 (see STAR Methods). Blood samples were collected
Mild N/A 70% (14/20)
at 20–35 days post-symptoms onset from non-hospitalized
Moderate N/A 20% (4/20) COVID-19 patients who were no longer symptomatic. SARS-
Severe N/A 10% (2/20) CoV-2 infection was determined by swab test viral PCR during
Critical N/A 0% (0/20) the acute phase of the infection. Verification of SARS-CoV-2
Symptoms exposure was attempted both by lateral flow serology and
Cough N/A 79% (15/19) SARS-CoV-2 spike protein receptor binding domain (RBD)
ELISA (Stadlbauer et al., 2020), using plasma from the convales-
Fatigue N/A 42% (8/19)
cence stage blood draw. Most patients were confirmed positive
Fever N/A 37% (7/19)
by lateral flow immunoglobulin (Ig) tests (Table 1). All patients
Anosmia N/A 21% (4/19) were confirmed COVID-19 cases by SARS-CoV-2 RBD ELISA
Dyspnea N/A 16% (3/19) (Figures 1 and S1). All cases were IgG positive; anti-RBD IgM
Diarrhea N/A 5% (1/19) and IgA was also detected in the large majority of cases (Figures
Days Post Symptom N/A 20–36 (18/20) 1 and S1).
Onset at Collection (median = 26, IQR = 7) We defined a 21-color flow cytometry panel of mononuclear
Past Medical History leukocyte lineage and phenotypic markers (Table S2) to broadly
No known N/A 65% (13/20) assess the immunological cellular profile of recovered COVID-19
patients (Figures 1 and S2). The frequency of CD3+ cells was
Hyperlipidemia N/A 15% (3/20)
slightly increased in recovered COVID-19 patients relative to
Hypertension N/A 10% (2/20)
non-exposed controls, while no significant differences overall
Asthma N/A 10% (2/20) were observed in the frequencies of CD4+ or CD8+ T cells
Known or suspected N/A 75% (15/20) between the two groups. Frequencies of CD19+ cells were
sick contact/exposure somewhat decreased, while no differences were observed in
a
Commercial skin prick lateral flow assay. the frequencies of CD3–CD19– cells or CD14+CD16– monocytes
b
WHO criteria. (Figures 1 and S2). No evidence of general lymphopenia was
observed in the convalescing patients, consistent with the litera-
ture. Next, we utilized the SARS-CoV-2 MPs to probe CD4+ and
et al., 2011; O’Sullivan et al., 1991; Sidney et al., 2010a, CD8+ T cell responses.
2010b; Southwood et al., 1998).
For the spike protein, to ensure that all T cell reactivity Identification and Quantitation of SARS-CoV-2-Specific
against this important antigen can be detected, we generated CD4+ T Cell Responses
a separate MP covering the entire antigen with 253 15-mer We utilized T cell receptor (TCR) dependent activation induced
peptides overlapping by 10-residues (MP_S, Table S1). As marker (AIM) assays to identify and quantify SARS-CoV-2-spe-
stated above, the MP used to probe the non-spike regions is cific CD4+ T cells in recovered COVID-19 patients. Initial defini-
expected to capture 50% of the total response. The use of tion and assessment of human antigen-specific SARS-CoV-2
overlapping peptides spanning entire open reading frames T cell responses are best made with direct ex vivo T cell assays
(ORFs) instead allows for a more complete characterization using broad-based epitope pools, such as MPs, and
but also requires more cells. This factor should be kept in assays capable of detecting T cells of unknown cytokine

Cell 181, 1489–1501, June 25, 2020 1491

 ll
Article

A B C Figure 1. SARS-CoV-2 IgM, IgA, and IgG Re-

107 107 sponses of Recovered COVID-19 Patients
107 **** ****

SARS-CoV2 spike RBD IgM

SARS-CoV2 spike RBD IgA

****
SARS-CoV2 spike RBD IgG

(A–C) Plasma ELISA titers to SARS-CoV-2 spike RBD. (A)

IgG. (B) IgM. (C) IgA. Neg, unexposed donors from 2015–
106 106 106
2018 (n = 20); COVID, convalescing COVID-19 patients (n =
20). All data are shown as ELISA titers based on a standard.
105 105 105 The dotted line indicates limit of detection. Geometric
mean titers with geometric SDs are indicated.

104 104 104 (D–I) Immunophenotyping of mononuclear leukocytes.

Frequency of (D) CD3+ total T cells, (E) CD4+ T cells
(CD4+CD3+), (F) CD8+ T cells (CD8+CD3+), (G) CD19+ B
103 103 103
Neg COVID Neg COVID cells (CD19+CD3–), (H) CD3–CD19– cells, and (I)
Neg COVID
CD14+CD16– monocytes (CD3–CD19–CD56–) from the
D E F PBMCs of unexposed donors (Neg, n = 13) or convalescing
80 * 80 60 ns
ns COVID-19 patients (COVID, n = 14). Data were analyzed
using the Mann-Whitney test with mean and standard de-
70
60 viation shown.

% CD8+ T cells
% CD4+ T cells
% CD3+ cells

40 *p < 0.05, ****p < 0.0001. See also Figures S1 and S2.
60
40
50
20
20
40
surements in independent experiments was
0 30 0 high (p < 0.0002, Figure S3D). To assess func-
Neg COVID Neg COVID Neg COVID
tionality and polarization of the SARS-CoV-2-
G H I specific CD4+ T cell response, we measured cy-
30 * 80 ns 80 ns
tokines secreted in response to MP stimulation.
% CD14+CD16- monocytes

The SARS-CoV-2-specific CD4+ T cells were

% CD3-CD19- cells

60 70
functional, as the cells produced IL-2 in
% CD19+ cells

20
40
10
20
0 0
Neg COVID
polarization and functional attributes. AIM assays are cytokine-
independent assays to identify antigen-specific CD4+ T cells
(Havenar-Daughton et al., 2016; Reiss et al., 2017). AIM assays
have been successfully used to identify virus-specific, vaccine-
specific, or tuberculosis-specific CD4+ T cells in a range of
studies (Dan et al., 2016, 2019; Herati et al., 2017; Morou
et al., 2019).
We stimulated PBMCs from 10 COVID-19 cases and 11
healthy controls (SARS-CoV-2 unexposed, collected in 2015–
2018) with a spike MP (MP_S) and the class II MP covering the
remainder of the SARS-CoV-2 orfeome (‘‘non-spike,’’ MP
CD4_R). A CMV MP was used as a positive control, while
DMSO was used as the negative control (Figures 2 and S3).
SARS-CoV-2 spike-specific CD4+ T cell responses
(OX40+CD137+) were detected in 100% of COVID-19 cases
(p < 0.0001 versus unexposed donors spike MP, Figures 2A
and 2B. p = 0.002 versus DMSO control, Figure 2C). CD4+
T cell responses to the remainder of the SARS-CoV-2 orfeome
were also detected in 100% of COVID-19 cases (p < 0.0079
versus unexposed donors non-spike MP, Figures 2A and 2B.
p = 0.002, non-spike versus DMSO control, Figure 2C). The
magnitude of the SARS-CoV-2-specific CD4+ T cell responses
measured was similar to that of the CMV MP (Figure S3C). The
concordance between SARS-CoV-2-specific CD4+ T cell mea-
response to non-spike and spike MPs (Fig-
60 ure 2D). Polarization of the cells appeared to be
a classic TH1 type, as substantial interferon
50 (IFN)-g was produced (Figure 2E), while little to
no IL-4, IL-5, IL-13, or IL-17a was expressed
40 (Figures S3G–S3J).
Neg COVID Neg COVID
Thus, recovered COVID-19 patients consis-
tently generated a substantial CD4+ T cell
response against SARS-CoV-2. Similar conclusions were
reached using stimulation index as the metric (Figures S3E
and S3F). In terms of total CD4+ T cell response per donor (Fig-
ure 2A), on average 50% of the detected response was
directed against the spike protein, and 50% was directed
against the MP representing the remainder of the SARS-CoV-
2 orfeome (Figure 2A). This is of significance, since the SARS-
CoV-2 spike protein is a key component of the vast majority
of candidate COVID-19 vaccines under development. Of
note, given the nature of the MP_R peptide predictions, the
actual CD4+ T cell response to be ascribed to non-spike
ORFs was likely to be higher, addressed in further experiments
below.
Identification and Quantitation of SARS-CoV-2-Specific
CD8+ T Cell Responses
To measure SARS-CoV-2-specific CD8+ T cells in the recov-
ered COVID-19 patients, we utilized two complementary meth-
odologies, AIM assays and intracellular cytokine staining (ICS).
The two SARS-CoV-2 class I MPs were used, CD8-A and CD8-
B, with CMV MP and DMSO serving as positive and negative
controls, respectively (Figures 3 and S4). CD8+ T cell responses
were detected by AIM (CD69+CD137+) in 70% of COVID-19
cases (p < 0.0011 versus unexposed donors ‘‘CD8 total,’’

1492 Cell 181, 1489–1501, June 25, 2020

 ll
Article

A B

C D E

Figure 2. SARS-CoV-2-Specific CD4+ T Cell Responses of Recovered COVID-19 Patients

(A) SARS-CoV-2-specific CD4+ T cells measured as percentage of AIM+ (OX40+CD137+) CD4+ T cells after stimulation of PBMCs with peptide pools encom-
passing spike only (Spike) MP or the CD4_R MP representing all the proteome without spike (Non-spike). Data were background subtracted against DMSO
negative control and are shown with geometric mean and geometric standard deviation. Samples were from unexposed donors (Unexposed, n = 11) and
recovered COVID-19 patients (COVID-19, n = 10).
(B) Fluorescence-activated cell sorting (FACS) plot examples, gated on total CD4+ T cells.
(C) AIM+ CD4+ T cell reactivity in COVID-19 cases between the negative control (DMSO) and antigen-specific stimulations.
(D and E) Cytokine levels in the supernatant of PBMCs from COVID-19 donors after stimulation with peptide pools (Spike and Non-spike) or the negative control
(DMSO). (D) IL-2. (E) IFN-g.
Statistical comparisons across cohorts were performed with the Mann-Whitney test. Pairwise comparisons (C–E) were performed with the Wilcoxon test. **p <
0.01; ***p < 0.001. See also Figure S3 and Table S6.

Figures 3A and 3B; p = 0.002, CD8-A or CD8-B versus DMSO
control, Figure S4B). MP CD8-A contains spike epitopes,
among epitopes to other proteins. The magnitude of the
SARS-CoV-2 reactive CD8+ T cell responses measured by
AIM was somewhat lower than the CMV MP (Figure S4C).
Similar conclusions were reached using stimulation index (Fig-
ures S3D and S3E).
Independently, ICS assays detected IFN-g+ SARS-CoV-2-
specific CD8+ T cells in the majority of COVID-19 cases (Figures
3C and 3D). The majority of IFN-g+ cells co-expressed granzyme
B (Figures 3D and 3E). A substantial fraction of the IFN-g+
cells expressed tumor necrosis factor (TNF) but not IL-10 (Fig-
ure 3D). Thus, the majority of recovered COVID-19 patients
generated a CD8+ T cell response against SARS-CoV-2.
Relationship between SARS-CoV-2-Specific CD4+ T Cell
Responses and IgG and IgA Titers
Most protective antibody responses are dependent on CD4+
T cell help. Therefore, we assessed whether stronger SARS-
CoV-2-specific CD4+ T cell responses were associated with
higher antibody titers in COVID-19 cases. Given that spike is
the primary target of SARS neutralizing antibodies, we exam-
ined spike-specific CD4+ T cells. Spike-specific CD4+ T cell
responses correlated well with the magnitude of the anti-spike
RBD IgG titers (R = 0.81; p < 0.0001; Figure 4A). Similar re-
sults were obtained using stimulation index (Figure S5A).
The non-spike SARS-CoV-2-specific CD4+ T cell response
did not correlate as well with anti-spike RBD IgG titers (Fig-
ures 4B and S5B), consistent with a common requirement
for intramolecular CD4+ T cell help (Sette et al., 2008). Anti-
spike IgA titers also correlated with spike-specific CD4+
T cells (p < 0.0002, Figure S5). Thus, COVID-19 patients
make anti-spike RBD antibody responses commensurate
with the magnitude of their spike-specific CD4+ T cell
response. We then assessed the relationship between the
CD4+ and CD8+ T cell responses to SARS-CoV-2. SARS-
CoV-2-specific CD4+ and CD8+ T cell responses were well
correlated (R = 0.62. p = 0.0025, Figures 4C and S5). Thus,
antibody, CD4+, and CD8+ T cell responses to SARS-CoV-2
were generally well correlated.

Cell 181, 1489–1501, June 25, 2020 1493

 ll
Article

A B

C D E

Figure 3. SARS-CoV-2-Specific CD8+ T Cell Responses by Recovered COVID-19 Patients

(A) SARS-CoV-2-specific CD8+ T cells measured as percentage of AIM+ (CD69+CD137+) CD8+ T cells after stimulation of PBMCs with class I MPs (CD8-A, CD8-
B, and the combined data [Total]). Data were background subtracted against DMSO negative control and are shown with geometric mean and geometric
standard deviation. Samples were from unexposed donors (Unexposed, n = 11) and recovered COVID-19 patients (COVID-19, n = 10).
(B) FACS plot examples.
(C) Percentage of CD8+ T cells producing IFN-g in response to SARS-CoV-2 MPs, or CMV MP, in PBMCs from COVID-19 and unexposed donors after back-
ground subtraction. Data are shown with geometric mean and geometric standard deviation.
(D) Functional profile of IFN-g+CD8+ T cells producing granzyme B (GzB), TNF-a (TNF), or IL-10 in response to SARS-CoV-2 MPs. Mean and SD are shown.
(E) FACS plot examples of IFN-g and granzyme B co-expression.
Statistical comparisons across cohorts were performed with the Mann-Whitney test. *p < 0.05; **p < 0.01.; ns not significant. See also Figure S4 and Table S6.

Pre-existing Cross-Reactive Coronavirus-Specific
T Cells
While spike- and non-spike-specific CD4+ T cell responses were
detectable in all COVID-19 cases, cells were also detected in un-
exposed individuals (Figures 3A and 3B). These responses were
statistically significant for non-spike-specific CD4+ T cell reac-
tivity (non-spike, p = 0.039; spike, p = 0.067; Figures 5A and
5B). Non-spike-specific CD4+ T cell responses were above the
limit of detection in 50% of donors based on stimulation index
(SI) (Figure S3E). All of the donors were recruited between
2015 and 2018, excluding any possibility of exposure to SARS-
CoV-2. Four human coronaviruses are known causes of sea-
sonal ‘‘common cold’’ upper-respiratory tract infections:
HCoV-OC43, HCoV-HKU1, HCoV-NL63, and HCoV-229E. We
tested the SARS-CoV-2 unexposed donors for seroreactivity to
HCoV-OC43 and HCoV-NL63 as a representative betacoronavi-
rus and alphacoronavirus, respectively. All donors were IgG
seropositive to HCoV-OC43 and HCoV-NL63 RBD, to varying
degrees (Figure 5C), consistent with the endemic nature of these
viruses (Gorse et al., 2010; Huang et al., 2020; Severance et al.,
2008). We therefore examined whether these represented true
pan-coronavirus T cells capable of recognizing SARS-CoV-2
epitopes.
SARS-CoV-2 ORF Targets of CD4+ and CD8+ T Cells
A most pressing, yet unresolved, set of issues in understanding
SARS-CoV-2 immune responses is what antigens are targeted
by CD4+ and CD8+ T cells, whether the corresponding antigens

1494 Cell 181, 1489–1501, June 25, 2020

 ll
Article

A B C

Figure 4. Correlations between SARS-CoV-2-Specific CD4+ T Cells, Antibodies, and CD8+ T Cells
(A) Correlation between SARS-CoV-2 spike-specific CD4+ T cells (%) and anti-spike RBD IgG.
(B) Correlation between SARS-CoV-2 non-spike-specific CD4+ T cells (%) and anti-spike RBD IgG.
(C) Correlation between SARS-CoV-2-specific CD4+ T cells and SARS-CoV-2-specific CD8+ T cells. Total MP responses per donor were used in each case
(‘‘Non-spike’’ + ‘‘spike’’ (CD4_R + MP_S) for CD4+ T cells, CD8_A + CD8_B for CD8+ T cells).
Statistical comparisons were performed using Spearman correlation. See also Figure S5.

are the same or different, and how do they reflect the antigens
currently considered for COVID-19 vaccine development. We
synthesized sets of overlapping peptides spanning the entire
sequence of SARS-CoV-2 and pooled them separately so that
each pool would represent one antigen (with the exception of
nsp3, for which two pools were made; Table S1).
In the case of CD4+ T cell responses, no obvious pattern of
antigen specificity was observed based on SARS-CoV-2
genome organization; however, coronaviruses increase protein
synthesis of certain ORFs in infected cells via subgenomic
RNAs. Accounting for the relative abundance of subgenomic
RNAs (Figure 6A) (Irigoyen et al., 2016; Snijder et al., 2003;
Xie et al., 2020), the ORFs were re-ordered based on predicted
protein abundance (Figure 6B). A clear hierarchy of SARS-CoV-
2-specific CD4+ T cell targets was then apparent, with the ma-
jority of the CD4+ T cell response in COVID-19 cases directed
against highly expressed SARS-CoV-2 ORFs spike, M, and N.
On average, these antigens accounted for 27%, 21%, and
11% of the total CD4+ T cell response, respectively. Most
COVID-19 cases also had CD4+ T cells specific for SARS-
CoV-2 nsp3, nsp4, and ORF8 (Figure 6B), on average each ac-
counting for 5% of the total CD4+ T cell response (Figure 6C).
E, ORF6, hypothetical ORF10, and nsp1 are all small antigens
(or potentially not expressed, in the case of ORF10) and were
most likely predominantly unrecognized as a result. These re-
sults are somewhat unexpected, because data for other coro-
naviruses, from 27 different studies curated in the IEDB, re-
ported that spike accounted for nearly two-thirds of reported
CD4+ T cell reactivity (Table S3). N accounted for most of the
remaining epitopes in the published literature, although human
N-specific CD4+ T cell responses were not observed in one of
the most comprehensive studies of human SARS-CoV-1
T cell responses (Li et al., 2008). Coronavirus M has not previ-
ously been described as a prominent target of CD4+ T cell re-
sponses (Table S3). In sum, these results, fully scanning the
SARS2 orfeome, demonstrate a pattern of robust and diverse
SARS-CoV-2-specific CD4+ T cell reactivity in convalescing
COVID-19 cases that correlated largely with predicted viral pro-
tein abundance in infected cells.
When examining the non-exposed donors, the pattern of CD4+
T cell targets changed. While S was still a relatively prominent
target (23% of total, on average), there was no, or marginal, reac-
tivity against SARS-CoV-2 N and M. Among donors with detect-
able CD4+ T cells, a shift in reactivity was observed toward
SARS-CoV-2 nsp14 (25%), nsp4 (15%) and nsp6 (14%) (Figures
6B and 6C). SARS-CoV-2-reactive CD4+ T cells were detected in
at least six different unexposed donors, demonstrating that the
cross-reactivity is relatively widely distributed (Figure S6A).
Having scanned the full SARS-CoV-2 orfeome for CD4+ T cell
reactivity in multiple donors, it was possible to assess whether
the epitope prediction MP approach successfully enriched for
SARS-CoV-2 epitopes targeted by human CD4+ T cells. When
the total reactivity observed with the CD4_R MP was plotted
versus the sum total of all antigen pools (excluding spike, given
that spike predictions were not included in the CD4_R MP), a sig-
nificant correlation was observed (p < 0.0002, Figure S6C). The
single MP-R captured 50% (44% +/ range 28%–80%) of
the non-spike response per COVID-19 donor, demonstrating
the success of the prediction approach, which, as mentioned
above, was devised to attempt to capture approximately 50%
of the total response (Dhanda et al., 2018; Paul et al., 2015).
In the case of CD8+ T cell responses, the data in the literature
from other coronaviruses (57 different studies curated in the
IEDB; Table S3) reported spike accounting for 50% and N ac-
counting for 36% of the defined epitopes. In a large study of hu-
man SARS-CoV-1 responses, spike was reported as essentially
the only target of CD8+ T cell responses (Li et al., 2008), while in a
study of MERS CD8+ T cells, responses were noted for spike, N
and a pool of M/E peptides (Zhao et al., 2017). Few epitopes
have been reported from other coronavirus antigens (Table
S3). Here, we scanned the full SARS-CoV-2 orfeome for CD8+
T cell recognition. Our data indicate a somewhat different pattern
of immunodominance for SARS-CoV-2 CD8+ T cell reactivity
(Figures 6D and 6E), with spike protein accounting for 26%
of the reactivity, and N accounting for 12%. Significant reac-
tivity in COVID-19 recovered subjects was derived from other
antigens, such as M (22%), nsp6 (15%), ORF8 (10%), and
ORF3a (7%) (Figures 6D and 6E). In unexposed donors, SARS-

Cell 181, 1489–1501, June 25, 2020 1495

 ll
A B
Figure 5. SARS-CoV-2 Epitope Reactivity in Unexposed Individuals
(A) SARS-CoV-2-reactive CD4+ T cells measured as percentage of AIM+ (OX40+CD137+) CD4+ T cells in unexposed (n = 11) donors.
(B) FACS plot examples, gated on total CD4+ T cells.
(C) Plasma IgG ELISAs for seroreactivity to RBD of HCoV-OC43 or HCoV-NL63. Data are expressed as geometric mean and geometric SD.
Pairwise statistical comparisons (A) were performed with the Wilcoxon test. *p < 0.05; ns, not significant.
CoV-2-reactive CD8+ T cells were detected in at least four
different donors (Figure S7), with less clear targeting of specific
SARS-CoV-2 proteins than was observed for CD4+ T cells, sug-
gesting that coronavirus CD8+ T cell cross-reactivity exists but is
less widespread than CD4+ T cell cross-reactivity.
DISCUSSION
There is a critical need for foundational knowledge about T cell
responses to SARS-CoV-2. Here, we report functional validation
of predicted epitopes when arranged in epitope MPs, utilizing
PBMCs derived from convalescing COVID-19 cases. The exper-
iments also used protein-specific peptide pools to determine
which SARS-CoV-2 proteins are the predominant targets of hu-
man SARS-CoV-2-specific CD4+ and CD8+ T cells generated
during COVID-19 disease. Importantly, we utilized the exact
same series of experimental techniques with blood samples
from healthy control donors (PBMCs collected in the 2015–
2018 time frame), and substantial cross-reactive coronavirus
T cell memory was observed.
Our results demonstrate that the epitope MPs are reagents
well suited to analyze and detect SARS-CoV-2-specific T cell
responses with limited sample material. We also developed
and tested peptide pools corresponding to each of the 25 pro-
teins encoded in the SARS-CoV-2 genome. Data from both
the epitope MPs and protein peptide pool experiments can
be interpreted in the context of previously reported T cell
response immunodominance patterns observed for other co-
ronaviruses, particularly the SARS and MERS viruses, which
have been studied in humans, HLA-transgenic mice, wild-
type mice, and other species. In the case of CD4+ T cell re-
sponses, data for other coronaviruses found that spike ac-
counted for nearly two-thirds of reported CD4+ T cell reac-
tivity, with N and M accounting for limited reactivity, and no
reactivity in one large study of human SARS-CoV-1 responses
(Li et al., 2008). Our SARS-CoV-2 data reveal that the pattern
of immunodominance in COVID-19 is different. In particular,
M, spike, and N proteins were clearly co-dominant, each
1496 Cell 181, 1489–1501, June 25, 2020
Article
C
recognized by 100% of COVID-19 cases studied here. Signif-
icant CD4+ T cell responses were also directed against nsp3,
nsp4, ORF3s, ORF7a, nsp12, and ORF8. These data suggest
that a candidate COVID-19 vaccine consisting only of SARS-
CoV-2 spike would be capable of eliciting SARS-CoV-2-spe-
cific CD4+ T cell responses of similar representation to that
of natural COVID-19 disease, but the data also indicate that
there are many potential CD4+ T cell targets in SARS-CoV-2,
and inclusion of additional SARS-CoV-2 structural antigens
such as M and N would better mimic the natural SARS-CoV-
2-specific CD4+ T cell response observed in mild to moderate
COVID-19 disease.
Regarding SARS-CoV-2 CD8+ T cell responses, the pattern of
immunodominance found here differed from the literature for
other coronaviruses. However, stringent comparisons are not
possible, as some earlier studies were not similarly comprehen-
sive and did not utilize the same experimental strategy. The spike
protein was a target of human SARS-CoV-2 CD8+ T cell re-
sponses, but it is not dominant. SARS-CoV-2 M was just as
strongly recognized, and significant reactivity was noted for
other antigens, mostly nsp6, ORF3a, and N, which comprised
nearly 50% of the total CD8+ T cell response, on average.
Thus, these data indicate that candidate COVID-19 vaccines
endeavoring to elicit CD8+ T cell responses against the spike
protein will be eliciting a relatively narrow CD8+ T cell response
compared to the natural CD8+ T cell response observed in mild
to moderate COVID-19 disease. An optimal vaccine CD8+
T cell response to SARS-CoV-2 might benefit from additional
class I epitopes, such as the ones derived from the M, nsp6,
ORF3a, and/or N.
There have been concerns regarding vaccine enhancement of
disease by certain candidate COVID-19 vaccine approaches, via
antibody-dependent enhancement (ADE) or development of a
TH2 responses (Peeples, 2020). Herein, we saw predominant
TH1 responses in convalescing COVID-19 cases, with little to
no TH2 cytokines. Clearly more studies are required, but the
data here appear to predominantly represent a classic TH1
response to SARS-CoV-2.
 ll
Article

B C

D E

Figure 6. Protein Immunodominance of SARS-CoV-2-Specific CD4+ and CD8+ T Cells in COVID-19 Cases and Unexposed Donors
(A) SARS-CoV-2 genome organization and predicted viral protein abundance in infected cells.
(B) SARS-CoV-2 antigen-specific CD4+ T cells (AIM+, OX40+CD137+) quantified by stimulation index, using a peptide pool for each viral protein (with two ex-
ceptions, see Table S1). COVID-19 cases (top, in blue. n = 10) and unexposed donors (bottom, in white. n = 10). Data are expressed as geometric mean and
geometric SD.
(C) Fraction of SARS-CoV-2 proteins recognized by CD4+ T cells in COVID-19 cases (top) and unexposed donors (bottom).
(D) SARS-CoV-2 antigen-specific CD4+ T cells (AIM+, OX40+CD137+) quantified by stimulation index, using a peptide pool for each viral protein (with two ex-
ceptions, see Table S1). COVID-19 cases (top, in red. n = 10) and unexposed donors (bottom, in gray. n = 10). Data are expressed as geometric mean and
geometric SD.
(E) Fraction of SARS-CoV-2 proteins recognized by CD8+ T cells in COVID-19 cases (top) and unexposed donors (bottom).
See also Figures S6 and S7 and Table S6.

Cell 181, 1489–1501, June 25, 2020 1497

 ll
While it was important to identify antigen-specific T cell re-
sponses in COVID-19 cases, it is also of great interest to under-
stand whether cross-reactive immunity exists between corona-
viruses to any degree. A key step in developing that
understanding is to examine antigen-specific CD4+ and CD8+
T cells in COVID-19 cases and in unexposed healthy controls,
utilizing the exact same antigens and series of experimental
techniques. CD4+ T cell responses were detected in 40%–60%
of unexposed individuals. This may be reflective of some degree
of cross-reactive, preexisting immunity to SARS-CoV-2 in some,
but not all, individuals. Whether this immunity is relevant in influ-
encing clinical outcomes is unknown—and cannot be known
without T cell measurements before and after SARS-CoV-2
infection of individuals—but it is tempting to speculate that the
cross-reactive CD4+ T cells may be of value in protective immu-
nity, based on SARS mouse models (Zhao et al., 2016). Clear
identification of the cross-reactive peptides, and their sequence
homology relation to other coronaviruses, requires deconvolu-
tion of the positive peptide pools, which is not feasible with the
cell numbers presently available, and time frame of the pre-
sent study.
Regarding the value of cross-reactive T cells, influenza (flu)
immunology in relationship to pandemics may be instructive. In
the context of the 2009 H1N1 influenza pandemic, preexisting
T cell immunity existed in the adult population, which focused
on the more conserved internal influenza viral proteins (Green-
baum et al., 2009). The presence of cross-reactive T cells was
found to correlate with less severe disease (Sridhar et al.,
2013; Wilkinson et al., 2012). The frequent availability of cross-
reactive memory T cell responses might have been one factor
contributing to the lesser severity of the H1N1 flu pandemic
(Hancock et al., 2009). Cross-reactive immunity to influenza
strains has been modeled to be a critical influencer of suscepti-
bility to newly emerging, potentially pandemic, influenza strains
(Gostic et al., 2016). Given the severity of the ongoing COVID-
19 pandemic, it has been modeled that any degree of cross-pro-
tective coronavirus immunity in the population could have a very
substantial impact on the overall course of the pandemic, and
the dynamics of the epidemiology for years to come (Kissler
et al., 2020).
Limitations and Future Directions
Caveats of this study include the sample size and the focus on
non-hospitalized COVID-19 cases. Sample size was limited by
expediency. The focus on non-hospitalized cases of COVID-19
is a strength, in that these donors had uncomplicated disease
of moderate duration, and thus it was encouraging that substan-
tial CD4+ T cell and antibody responses were detected in all
cases, and CD8+ T cell responses in the majority of cases. Com-
plementing these data with MP T cell data from acute patients
and patients with complicated disease course will also be of
clear value, as will studies on the longevity of SARS-CoV-2
immunological memory. Additionally, lack of detailed informa-
tion on common cold history or matched blood samples pre-
exposure to SARS-CoV-2 prevents conclusions regarding the
abundance of cross-reactive coronavirus T cells before expo-
sure to SARS-CoV-2 and any potential protective efficacy of
such cells. Finally, full epitope mapping in the future will add
1498 Cell 181, 1489–1501, June 25, 2020
Article
important detailed resolution of the human coronavirus-specific
T cell responses.
In sum, we measured SARS-CoV-2-specific CD4+ and CD8+
T cells responses in COVID-19 cases. Using multiple
experimental approaches, SARS-CoV-2-specific CD4+ T cell
and antibody responses were observed in all COVID-19 cases,
and CD8+ T cell responses were observed in most. Importantly,
pre-existing SARS-CoV-2-cross-reactive T cell responses were
observed in healthy donors, indicating some potential for pre-ex-
isting immunity in the human population. ORF mapping of T cell
specificities revealed valuable targets for incorporation in candi-
date vaccine development and revealed distinct specificity pat-
terns between COVID-19 cases and unexposed healthy
controls.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Human Subjects
d METHOD DETAILS
B Peptide Pools
B PBMC isolation
B SARS-CoV-2 RBD ELISA
B OC43 and NL63 coronavirus RBD ELISA
B Flow Cytometry
B Cytokine bead assays
B Identification of coronavirus epitopes and associated
literature references
d QUANTIFICATION AND STATISTICAL ANALYSIS
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.015.
ACKNOWLEDGMENTS
We would like to thank Cheryl Kim, director of the LJI flow cytometry core fa-
cility for outstanding expertise. We thank Prof. Peter Kim, Abigail Powell, PhD,
and colleagues (Stanford) for RBD protein synthesized from Prof. Florian
Krammer (Mt. Sinai) constructs. J.M. was supported by PhD student fellow-
ships from the Departamento Administrativo de Ciencia, Tecnologia e Innova-
cion (COLCIENCIAS), and Pontificia Universidad Javeriana. This work was
funded by the NIH NIAID under awards AI42742 (Cooperative Centers for Hu-
man Immunology) (S.C. and A.S.), National Institutes of Health contract Nr.
75N9301900065 (A.S. and D.W.), and U19 AI118626 (A.S. and B.P.). The BD
FACSymphony purchase was partially funded by the Bill and Melinda Gates
Foundation and LJI Institutional Funds (S.C. and A.S.). This work was addition-
ally supported in part by the Johnathan and Mary Tu Foundation (D.M.S.), the
NIAID under K08 award AI135078 (J.D.), and UCSD T32s AI007036 and
AI007384 Infectious Diseases Division (S.I.R. and S.A.R.).

Article
AUTHOR CONTRIBUTIONS
Conceptualization, A.G., D.W., S.C., and A.S.; Investigation, A.G., D.W., J.M.,
C.R.M., J.M.D., D.M., L.P., R.S.J., A.S., and D.W.; Formal Analysis, A.G., D.W.,
C.R.M., J.M.D., J.M., and S.C.; Resources, S.I.R., S.A.R., D.M.S., A.F.C., F.K.,
S.C., and A.S.; Data Curation, J.A.G. and B.P.; Writing, S.C., A.S., A.G., and
D.W.; Supervision, B.P., A.M.d.S., S.C., and A.S.; Project Administration,
A.F.; Funding Acquisition, S.C., A.S., D.W., D.S., and J.D.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: April 20, 2020
Revised: May 4, 2020
Accepted: May 7, 2020
Published: May 14, 2020
REFERENCES
Alshukairi, A.N., Zheng, J., Zhao, J., Nehdi, A., Baharoon, S.A., Layqah, L., Bo-
khari, A., Al Johani, S.M., Samman, N., Boudjelal, M., et al. (2018). High Prev-
alence of MERS-CoV Infection in Camel Workers in Saudi Arabia. MBio 9,
e01985–e01918.
Amanat, F., and Krammer, F. (2020). SARS-CoV-2 Vaccines: Status Report.
Immunity 52, 583–589.
Andrews, S.F., Huang, Y., Kaur, K., Popova, L.I., Ho, I.Y., Pauli, N.T., Henry Du-
nand, C.J., Taylor, W.M., Lim, S., Huang, M., et al. (2015). Immune history pro-
foundly affects broadly protective B cell responses to influenza. Sci. Transl.
Med. 7, 316ra192.
Bancroft, T., Dillon, M.B., da Silva Antunes, R., Paul, S., Peters, B., Crotty, S.,
Lindestam Arlehamn, C.S., and Sette, A. (2016). Th1 versus Th2 T cell polari-
zation by whole-cell and acellular childhood pertussis vaccines persists upon
re-immunization in adolescence and adulthood. Cell. Immunol. 304-
305, 35–43.
Blanco-Melo, D., Nilsson-Payant, B.E., Liu, W.-C., Møller, R., Panis, M.,
Sachs, D., Albrecht, R.A., and tenOever, B.R. (2020). SARS-CoV-2 launches
a unique transcriptional signature from in vitro, ex vivo, and in vivo systems.
bioRxiv. https://doi.org/10.1101/2020.03.24.004655.
Callow, K.A., Parry, H.F., Sergeant, M., and Tyrrell, D.A. (1990). The time
course of the immune response to experimental coronavirus infection of
man. Epidemiol. Infect. 105, 435–446.
Cao, X. (2020). COVID-19: immunopathology and its implications for therapy.
Nat. Rev. Immunol. 20, 269–270.
Carrasco Pro, S., Sidney, J., Paul, S., Lindestam Arlehamn, C., Weiskopf, D.,
Peters, B., and Sette, A. (2015). Automatic Generation of Validated Specific
Epitope Sets. J. Immunol. Res. 2015, 763461.
Choe, P.G., Perera, R.A.P.M., Park, W.B., Song, K.H., Bang, J.H., Kim, E.S.,
Kim, H.B., Ko, L.W.R., Park, S.W., Kim, N.J., et al. (2017). MERS-CoV Antibody
Responses 1 Year after Symptom Onset, South Korea, 2015. Emerg. Infect.
Dis. 23, 1079–1084.
Crotty, S. (2019). T Follicular Helper Cell Biology: A Decade of Discovery and
Diseases. Immunity 50, 1132–1148.
da Silva Antunes, R., Paul, S., Sidney, J., Weiskopf, D., Dan, J.M., Phillips, E.,
Mallal, S., Crotty, S., Sette, A., and Lindestam Arlehamn, C.S. (2017). Definition
of Human Epitopes Recognized in Tetanus Toxoid and Development of an
Assay Strategy to Detect Ex Vivo Tetanus CD4+ T Cell Responses. PLoS
ONE 12, e0169086.
Dan, J.M., Lindestam Arlehamn, C.S., Weiskopf, D., da Silva Antunes, R., Ha-
venar-Daughton, C., Reiss, S.M., Brigger, M., Bothwell, M., Sette, A., and
Crotty, S. (2016). A Cytokine-Independent Approach To Identify Antigen-Spe-
cific Human Germinal Center T Follicular Helper Cells and Rare Antigen-Spe-
cific CD4+ T Cells in Blood. J. Immunol. 197, 983–993.
ll
Dan, J.M., Havenar-Daughton, C., Kendric, K., Al-Kolla, R., Kaushik, K., Ro-
sales, S.L., Anderson, E.L., LaRock, C.N., Vijayanand, P., Seumois, G., et al.
(2019). Recurrent group A Streptococcus tonsillitis is an immunosusceptibility
disease involving antibody deficiency and aberrant TFH cells. Sci. Transl. Med.
11, eaau3776.
Dhanda, S.K., Karosiene, E., Edwards, L., Grifoni, A., Paul, S., Andreatta, M.,
Weiskopf, D., Sidney, J., Nielsen, M., Peters, B., and Sette, A. (2018). Predict-
ing HLA CD4 Immunogenicity in Human Populations. Front. Immunol. 9, 1369.
Dhanda, S.K., Mahajan, S., Paul, S., Yan, Z., Kim, H., Jespersen, M.C., Jurtz,
V., Andreatta, M., Greenbaum, J.A., Marcatili, P., et al. (2019). IEDB-AR: im-
mune epitope database-analysis resource in 2019. Nucleic Acids Res. 47
(W1), W502–W506.
Dong, E., Du, H., and Gardner, L. (2020). An interactive web-based dashboard
to track COVID-19 in real time. Lancet Infect. Dis. 20, 533–534.
Giamarellos-Bourboulis, E.J., Netea, M.G., Rovina, N., Akinosoglou, K., Anto-
niadou, A., Antonakos, N., Damoraki, G., Gkavogianni, T., Adami, M.E., Kat-
saounou, P., et al. (2020). Complex Immune Dysregulation in COVID-19 Pa-
tients with Severe Respiratory Failure. Cell Host Microbe. Published online
April 17, 2020. https://doi.org/10.1016/j.chom.2020.04.009.
Gorse, G.J., Patel, G.B., Vitale, J.N., and O’Connor, T.Z. (2010). Prevalence of
antibodies to four human coronaviruses is lower in nasal secretions than in
serum. Clin. Vaccine Immunol. 17, 1875–1880.
Gostic, K.M., Ambrose, M., Worobey, M., and Lloyd-Smith, J.O. (2016). Potent
protection against H5N1 and H7N9 influenza via childhood hemagglutinin
imprinting. Science 354, 722–726.
Greenbaum, J.A., Kotturi, M.F., Kim, Y., Oseroff, C., Vaughan, K., Salimi, N.,
Vita, R., Ponomarenko, J., Scheuermann, R.H., Sette, A., and Peters, B.
(2009). Pre-existing immunity against swine-origin H1N1 influenza viruses in
the general human population. Proc. Natl. Acad. Sci. USA 106, 20365–20370.
Greenbaum, J., Sidney, J., Chung, J., Brander, C., Peters, B., and Sette, A.
(2011). Functional classification of class II human leukocyte antigen (HLA) mol-
ecules reveals seven different supertypes and a surprising degree of repertoire
sharing across supertypes. Immunogenetics 63, 325–335.
Grifoni, A., Angelo, M.A., Lopez, B., O’Rourke, P.H., Sidney, J., Cerpas, C.,
Balmaseda, A., Silveira, C.G.T., Maestri, A., Costa, P.R., et al. (2017). Global
Assessment of Dengue Virus-Specific CD4+ T Cell Responses in Dengue-
Endemic Areas. Front. Immunol. 8, 1309.
Grifoni, A., Sidney, J., Zhang, Y., Scheuermann, R.H., Peters, B., and Sette, A.
(2020). A Sequence Homology and Bioinformatic Approach Can Predict
Candidate Targets for Immune Responses to SARS-CoV-2. Cell Host Microbe
27, 671–680.
Guo, T., Fan, Y., Chen, M., Wu, X., Zhang, L., He, T., Wang, H., Wan, J., Wang,
X., and Lu, Z. (2020a). Cardiovascular Implications of Fatal Outcomes of Pa-
tients With Coronavirus Disease 2019 (COVID-19). JAMA Cardiol. Published
online March 27, 2020. https://doi.org/10.1001/jamacardio.2020.1017.
Guo, X., Guo, Z., Duan, C., Chen, Z., Wang, G., Lu, Y., Li, M., and Lu, J.
(2020b). Long-Term Persistence of IgG Antibodies in SARS-CoV Infected
Healthcare Workers. medRxiv. https://doi.org/10.1101/2020.02.12.20021386.
Hancock, K., Veguilla, V., Lu, X., Zhong, W., Butler, E.N., Sun, H., Liu, F., Dong,
L., DeVos, J.R., Gargiullo, P.M., et al. (2009). Cross-reactive antibody re-
sponses to the 2009 pandemic H1N1 influenza virus. N. Engl. J. Med. 361,
1945–1952.
Havenar-Daughton, C., Reiss, S.M., Carnathan, D.G., Wu, J.E., Kendric, K.,
Torrents de la Peña, A., Kasturi, S.P., Dan, J.M., Bothwell, M., Sanders,
R.W., et al. (2016). Cytokine-Independent Detection of Antigen-Specific
Germinal Center T Follicular Helper Cells in Immunized Nonhuman Primates
Using a Live Cell Activation-Induced Marker Technique. J. Immunol. 197,
994–1002.
Herati, R.S., Muselman, A., Vella, L., Bengsch, B., Parkhouse, K., Del Alcazar,
D., Kotzin, J., Doyle, S.A., Tebas, P., Hensley, S.E., et al. (2017). Successive
annual influenza vaccination induces a recurrent oligoclonotypic memory
response in circulating T follicular helper cells. Sci Immunol 2, eaag2152.
Cell 181, 1489–1501, June 25, 2020 1499
jixiansheng
ll
Hinz, D., Seumois, G., Gholami, A.M., Greenbaum, J.A., Lane, J., White, B.,
Broide, D.H., Schulten, V., Sidney, J., Bakhru, P., et al. (2016). Lack of allergy
to timothy grass pollen is not a passive phenomenon but associated with the
allergen-specific modulation of immune reactivity. Clin. Exp. Allergy 46,
705–719.
Huang, A.T., Garcia-Carreras, B., Hitchings, M.D.T., Yang, B., Katzelnick, L.,
Rattigan, S.M., Borgert, B., Moreno, C., Solomon, B.D., Rodriguez-Barraquer,
I., et al. (2020). A systematic review of antibody mediated immunity to
coronaviruses: antibody kinetics, correlates of protection, and association
of antibody responses with severity of disease. medRxiv,
2020.2004.2014.20065771.
Irigoyen, N., Firth, A.E., Jones, J.D., Chung, B.Y., Siddell, S.G., and Brierley, I.
(2016). High-Resolution Analysis of Coronavirus Gene Expression by RNA
Sequencing and Ribosome Profiling. PLoS Pathog. 12, e1005473.
Jurtz, V., Paul, S., Andreatta, M., Marcatili, P., Peters, B., and Nielsen, M.
(2017). NetMHCpan-4.0: Improved Peptide-MHC Class I Interaction Predic-
tions Integrating Eluted Ligand and Peptide Binding Affinity Data.
J. Immunol. 199, 3360–3368.
Kissler, S.M., Tedijanto, C., Goldstein, E., Grad, Y.H., and Lipsitch, M. (2020).
Projecting the transmission dynamics of SARS-CoV-2 through the postpan-
demic period. Science, eabb5793.
Li, C.K., Wu, H., Yan, H., Ma, S., Wang, L., Zhang, M., Tang, X., Temperton,
N.J., Weiss, R.A., Brenchley, J.M., et al. (2008). T cell responses to whole
SARS coronavirus in humans. J. Immunol. 181, 5490–5500.
Lindestam Arlehamn, C.S., McKinney, D.M., Carpenter, C., Paul, S., Rozot, V.,
Makgotlho, E., Gregg, Y., van Rooyen, M., Ernst, J.D., Hatherill, M., et al.
(2016). A Quantitative Analysis of Complexity of Human Pathogen-Specific
CD4 T Cell Responses in Healthy M. tuberculosis Infected South Africans.
PLoS Pathog. 12, e1005760.
Liu, L., Wei, Q., Lin, Q., Fang, J., Wang, H., Kwok, H., Tang, H., Nishiura, K.,
Peng, J., Tan, Z., et al. (2019). Anti-spike IgG causes severe acute lung injury
by skewing macrophage responses during acute SARS-CoV infection. JCI
Insight 4, 123158.
Morou, A., Brunet-Ratnasingham, E., Dubé, M., Charlebois, R., Mercier, E.,
Darko, S., Brassard, N., Nganou-Makamdop, K., Arumugam, S., Gendron-
Lepage, G., et al. (2019). Altered differentiation is central to HIV-specific
CD4+ T cell dysfunction in progressive disease. Nat. Immunol. 20, 1059–1070.
Moutaftsi, M., Tscharke, D.C., Vaughan, K., Koelle, D.M., Stern, L., Calvo-
Calle, M., Ennis, F., Terajima, M., Sutter, G., Crotty, S., et al. (2010). Uncover-
ing the interplay between CD8, CD4 and antibody responses to complex path-
ogens. Future Microbiol. 5, 221–239.
O’Sullivan, D., Arrhenius, T., Sidney, J., Del Guercio, M.F., Albertson, M., Wall,
M., Oseroff, C., Southwood, S., Colón, S.M., Gaeta, F.C., et al. (1991). On the
interaction of promiscuous antigenic peptides with different DR alleles. Identi-
fication of common structural motifs. J. Immunol. 147, 2663–2669.
Okba, N.M.A., Raj, V.S., Widjaja, I., GeurtsvanKessel, C.H., de Bruin, E., Chan-
dler, F.D., Park, W.B., Kim, N.J., Farag, E.A.B.A., Al-Hajri, M., et al. (2019). Sen-
sitive and Specific Detection of Low-Level Antibody Responses in Mild Middle
East Respiratory Syndrome Coronavirus Infections. Emerg. Infect. Dis. 25,
1868–1877.
Paul, S., Lindestam Arlehamn, C.S., Scriba, T.J., Dillon, M.B., Oseroff, C., Hinz,
D., McKinney, D.M., Carrasco Pro, S., Sidney, J., Peters, B., and Sette, A.
(2015). Development and validation of a broad scheme for prediction of HLA
class II restricted T cell epitopes. J. Immunol. Methods 422, 28–34.
Paul, S., Sidney, J., Sette, A., and Peters, B. (2016). TepiTool: A Pipeline for
Computational Prediction of T Cell Epitope Candidates. Curr. Protoc. Immu-
nol. 114, 18.19.1–18.19.24.
Peeples, L. (2020). News Feature: Avoiding pitfalls in the pursuit of a COVID-19
vaccine. Proc. Natl. Acad. Sci. USA 117, 8218–8221.
Reiss, S., Baxter, A.E., Cirelli, K.M., Dan, J.M., Morou, A., Daigneault, A., Bras-
sard, N., Silvestri, G., Routy, J.P., Havenar-Daughton, C., et al. (2017).
Comparative analysis of activation induced marker (AIM) assays for sensitive
identification of antigen-specific CD4 T cells. PLoS ONE 12, e0186998.
1500 Cell 181, 1489–1501, June 25, 2020
Article
Sallusto, F., Lanzavecchia, A., Araki, K., and Ahmed, R. (2010). From vaccines
to memory and back. Immunity 33, 451–463.
Sette, A., Moutaftsi, M., Moyron-Quiroz, J., McCausland, M.M., Davies, D.H.,
Johnston, R.J., Peters, B., Rafii-El-Idrissi Benhnia, M., Hoffmann, J., Su, H.P.,
et al. (2008). Selective CD4+ T cell help for antibody responses to a large viral
pathogen: deterministic linkage of specificities. Immunity 28, 847–858.
Severance, E.G., Bossis, I., Dickerson, F.B., Stallings, C.R., Origoni, A.E., Sul-
lens, A., Yolken, R.H., and Viscidi, R.P. (2008). Development of a nucleo-
capsid-based human coronavirus immunoassay and estimates of individuals
exposed to coronavirus in a U.S. metropolitan population. Clin. Vaccine Immu-
nol. 15, 1805–1810.
Sidney, J., Peters, B., Frahm, N., Brander, C., and Sette, A. (2008). HLA class I
supertypes: a revised and updated classification. BMC Immunol. 9, 1.
Sidney, J., Steen, A., Moore, C., Ngo, S., Chung, J., Peters, B., and Sette, A.
(2010a). Divergent motifs but overlapping binding repertoires of six HLA-DQ
molecules frequently expressed in the worldwide human population.
J. Immunol. 185, 4189–4198.
Sidney, J., Steen, A., Moore, C., Ngo, S., Chung, J., Peters, B., and Sette, A.
(2010b). Five HLA-DP molecules frequently expressed in the worldwide human
population share a common HLA supertypic binding specificity. J. Immunol.
184, 2492–2503.
Snijder, E.J., Bredenbeek, P.J., Dobbe, J.C., Thiel, V., Ziebuhr, J., Poon, L.L.,
Guan, Y., Rozanov, M., Spaan, W.J., and Gorbalenya, A.E. (2003). Unique and
conserved features of genome and proteome of SARS-coronavirus, an early
split-off from the coronavirus group 2 lineage. J. Mol. Biol. 331, 991–1004.
Southwood, S., Sidney, J., Kondo, A., del Guercio, M.F., Appella, E., Hoffman,
S., Kubo, R.T., Chesnut, R.W., Grey, H.M., and Sette, A. (1998). Several com-
mon HLA-DR types share largely overlapping peptide binding repertoires.
J. Immunol. 160, 3363–3373.
Sridhar, S., Begom, S., Bermingham, A., Hoschler, K., Adamson, W., Carman,
W., Bean, T., Barclay, W., Deeks, J.J., and Lalvani, A. (2013). Cellular immune
correlates of protection against symptomatic pandemic influenza. Nat. Med.
19, 1305–1312.
Stadlbauer, D., Amanat, F., Chromikova, V., Jiang, K., Strohmeier, S., Arunku-
mar, G.A., Tan, J., Bhavsar, D., Capuano, C., Kirkpatrick, E., et al. (2020).
SARS-CoV-2 Seroconversion in Humans: A Detailed Protocol for a Serological
Assay, Antigen Production, and Test Setup. Curr. Protoc. Microbiol. 57, e100.
Takano, T., Kawakami, C., Yamada, S., Satoh, R., and Hohdatsu, T. (2008).
Antibody-dependent enhancement occurs upon re-infection with the identical
serotype virus in feline infectious peritonitis virus infection. J. Vet. Med. Sci. 70,
1315–1321.
Thanh Le, T., Andreadakis, Z., Kumar, A., Gómez Román, R., Tollefsen, S., Sa-
ville, M., and Mayhew, S. (2020). The COVID-19 vaccine development land-
scape. Nat. Rev. Drug Discov. 19, 305–306.
Tian, Y., Grifoni, A., Sette, A., and Weiskopf, D. (2019). Human T Cell Response
to Dengue Virus Infection. Front. Immunol. 10, 2125.
Vennema, H., de Groot, R.J., Harbour, D.A., Dalderup, M., Gruffydd-Jones, T.,
Horzinek, M.C., and Spaan, W.J. (1990). Early death after feline infectious peri-
tonitis virus challenge due to recombinant vaccinia virus immunization. J. Virol.
64, 1407–1409.
Vita, R., Mahajan, S., Overton, J.A., Dhanda, S.K., Martini, S., Cantrell, J.R.,
Wheeler, D.K., Sette, A., and Peters, B. (2019). The Immune Epitope Database
(IEDB): 2018 update. Nucleic Acids Res. 47 (D1), D339–D343.
Weiskopf, D., Angelo, M.A., de Azeredo, E.L., Sidney, J., Greenbaum, J.A.,
Fernando, A.N., Broadwater, A., Kolla, R.V., De Silva, A.D., de Silva, A.M.,
et al. (2013). Comprehensive analysis of dengue virus-specific responses sup-
ports an HLA-linked protective role for CD8+ T cells. Proc. Natl. Acad. Sci. USA
110, E2046–E2053.
Weiskopf, D., Cerpas, C., Angelo, M.A., Bangs, D.J., Sidney, J., Paul, S., Pe-
ters, B., Sanches, F.P., Silvera, C.G., Costa, P.R., et al. (2015). Human CD8+ T-
Cell Responses Against the 4 Dengue Virus Serotypes Are Associated With
Distinct Patterns of Protein Targets. J. Infect. Dis. 212, 1743–1751.

Article
Wilkinson, T.M., Li, C.K.F., Chui, C.S.C., Huang, A.K.Y., Perkins, M., Liebner,
J.C., Lambkin-Williams, R., Gilbert, A., Oxford, J., Nicholas, B., et al. (2012).
Preexisting influenza-specific CD4+ T cells correlate with disease protection
against influenza challenge in humans. Nat. Med. 18, 274–280.
Wong, C.K., Lam, C.W., Wu, A.K., Ip, W.K., Lee, N.L., Chan, I.H., Lit, L.C., Hui,
D.S., Chan, M.H., Chung, S.S., and Sung, J.J. (2004). Plasma inflammatory cy-
tokines and chemokines in severe acute respiratory syndrome. Clin. Exp. Im-
munol. 136, 95–103.
Wu, Z., and McGoogan, J.M. (2020). Characteristics of and Important Lessons
From the Coronavirus Disease 2019 (COVID-19) Outbreak in China: Summary
of a Report of 72 314 Cases From the Chinese Center for Disease Control and
Prevention. JAMA 323, 1239–1242.
Xie, X., Muruato, A., Lokugamage, K.G., Narayanan, K., Zhang, X., Zou, J., Liu,
J., Schindewolf, C., Bopp, N.E., Aguilar, P.V., et al. (2020). An Infectious cDNA
Clone of SARS-CoV-2. Cell Host Microbe 27, 841–848.
Zhao, J., Zhao, J., Van Rooijen, N., and Perlman, S. (2009). Evasion by stealth:
inefficient immune activation underlies poor T cell response and severe dis-
ease in SARS-CoV-infected mice. PLoS Pathog. 5, e1000636.
ll
Zhao, J., Zhao, J., and Perlman, S. (2010). T cell responses are required for
protection from clinical disease and for virus clearance in severe acute respi-
ratory syndrome coronavirus-infected mice. J. Virol. 84, 9318–9325.
Zhao, J., Zhao, J., Legge, K., and Perlman, S. (2011). Age-related increases in
PGD(2) expression impair respiratory DC migration, resulting in diminished
T cell responses upon respiratory virus infection in mice. J. Clin. Invest. 121,
4921–4930.
Zhao, J., Zhao, J., Mangalam, A.K., Channappanavar, R., Fett, C., Meyerholz,
D.K., Agnihothram, S., Baric, R.S., David, C.S., and Perlman, S. (2016). Airway
Memory CD4(+) T Cells Mediate Protective Immunity against Emerging Respi-
ratory Coronaviruses. Immunity 44, 1379–1391.
Zhao, J., Alshukairi, A.N., Baharoon, S.A., Ahmed, W.A., Bokhari, A.A., Nehdi,
A.M., Layqah, L.A., Alghamdi, M.G., Al Gethamy, M.M., Dada, A.M., et al.
(2017). Recovery from the Middle East respiratory syndrome is associated
with antibody and T-cell responses. Sci. Immunol. 2, eaan5393.
Zhou, F., Yu, T., Du, R., Fan, G., Liu, Y., Liu, Z., Xiang, J., Wang, Y., Song, B.,
Gu, X., et al. (2020). Clinical course and risk factors for mortality of adult inpa-
tients with COVID-19 in Wuhan, China: a retrospective cohort study. Lancet
395, 1054–1062.
Cell 181, 1489–1501, June 25, 2020 1501
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
M5E2 (V500) [anti-CD14] Becton Dickinson 561391 (RRID:AB_10611856)
HIB19 (V500) [anti-CD19] Becton Dickinson 561121 (RRID:AB_10562391)
FN50 (BV605) [anti-CD4] Becton Dickinson 562989 (RRID:AB_2737935)
RPA-T8 (BV650) [anti-CD8] BioLegend 301042 (RRID:AB_2563505)
FN50 (PE-CF594) [anti-CD69] Becton Dickinson 562617 (RRID:AB_2737680)
Ber-ACT35 (PE-Cy7) [anti-OX40] Biolegend 350012 (RRID:AB_10901161)a
4B4-1 (APC) [anti-CD137] BioLegend 309810 (RRID:AB_830672)
OKT3 (AF700) [anti-CD3] Biolegend 317340 (RRID:AB_2563408)
G043H7 (BV421) [anti-CD45RA] BioLegend 353207 (RRID:AB_10915137)
4S.B3 (FITC) [anti-IFNg] Thermo Fisher Scientific 11-7319-82 (RRID: AB_465415)
GB11 (PE) [anti-Granzyme B] Thermo Fisher Scientific 12-8899-41 (RRID: AB_1659718)
Mab11 (PeCy7) [anti-TNFa] ebioscience 25-7349-82 (RRID:AB_469686)
JES3-19F1 (APC) [anti-IL-10] BioLegend 506807 (RRID:AB_315457)
3D12 (APC ef780) [anti-CCR7] eBioscience 47-1979-42 (RRID:AB_1518794)
B56 (FITC) [anti-KI67] Becton Dickinson 556026 (RRID:AB_396302)
SK3 (percp efluor710) [anti-CD4] Invitrogen 46-0047-42 (RRID:AB_1834401)
GB11 (af647) [anti-GzmB] Biolegend 515406 (RRID:AB_2566333)
MHM-88 (af700) [anti-IgM] Biolegend 314538 (RRID:AB_2566615)
O323 (APC cy7) [anti-CD27] Biolegend 302816 (RRID:AB_571977)
IA6-2 (PE) [anti-IgD] Becton Dickinson 555779 (RRID:AB_396114)
HCD56 (PE Dazzle) [anti-CD56] Biolegend 318348 (RRID:AB_2563564)
HIB19 (Cy5) [anti-CD19PE] Biolegend 302210 (RRID:AB_314240)
HIT2 (PECy7) [anti-CD38] Invitrogen 25-0389-42 (RRID:AB_1724057)
J252D4 (bv421) [anti-CXCR5] Biolegend 356920 (RRID:AB_2562303)
63D3 (bv510) [anti-CD14] Biolegend 367123 (RRID:AB_2716228)
HI100 (bv570) [anti-CD45RA] Biolegend 304132 (RRID:AB_2563813)
G025H7 (bv605) [anti-CXCR3] Biolegend 353728 (RRID:AB_2563157)
2H7 (bv650) [anti-CD20] Biolegend 302336 (RRID:AB_2563806)
G043H7 (bv711) [anti-CCR7] Biolegend 353228 (RRID:AB_2563865)
EH12.2H7 (bv786) [anti-PD-1] Biolegend 329930 (RRID:AB_2563443)
UCHT1 (buv395) [anti-CD3] Becton Dickinson 563546 (RRID:AB_2744387)
Live/dead (UV) [Zombie] Biolegend 423108
11A9 (buv496) [anti-CCR6] Becton Dickinson 612948 (RRID:AB_2833076)
3G8 (buv737) [anti-CD16] Becton Dickinson 612786 (RRID:AB_2833077)
SK1 (buv805) [anti-CD8] Becton Dickinson 612889 (RRID:AB_2833078)
LEGENDplex 13-plex kit Biolegend 740809
Biological Samples
Healthy donor blood samples Carter BloodCare http://www.carterbloodcare.org/
Healthy donor blood samples LJI Clinical Core https://www.iedb.org/
Convalescent donor blood samples UC San Diego Health http://www.health.ucsd.edu/
Chemicals, Peptides, and Recombinant Proteins
Synthetic peptides Synthetic Biomolecules (aka A&A) http://www.syntheticbiomolecules.com/
SARS-CoV-2 Receptor Binding Domain Stadlbauer et al., 2020 N/A
(RBD) protein
(Continued on next page)

e1 Cell 181, 1489–1501.e1–e6, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Critical Commercial Assays
CoronaCheck COVID-19 Rapid Antibody 20/20 BioResponse https://coronachecktest.com/
Test Kit
Deposited Data
Wuhan-Hu-1 RNA isolate NCBI nuccore database GenBank: MN_908947
ORF10 protein NCBI protein database NCBI: YP_009725255.1
Nucleocapsid phosphoprotein NCBI protein database NCBI: YP_009724397.2
ORF8 protein NCBI protein database NCBI: YP_009724396.1
ORF7a protein NCBI protein database NCBI: YP_009724395.1
ORF6 protein NCBI protein database NCBI: YP_009724394.1
membrane glycoprotein NCBI protein database NCBI: YP_009724393.1
envelope protein NCBI protein database NCBI: YP_009724392.1
ORF3a protein NCBI protein database NCBI: YP_009724391.1
surface glycoprotein NCBI protein database NCBI: YP_009724390.1
orf1ab polyprotein NCBI protein database NCBI: YP_009724389.1
Software and Algorithms
IEDB Vita et al., 2019 https://www.iedb.org
IEDB-AR (analysis resource) Dhanda et al., 2019 http://tools.iedb.org
NetMHCpan EL 4.0 Jurtz et al., 2017 http://tools.iedb.org/mhci/
IEDB Vita et al., 2019 https://www.iedb.org
Tepitool Paul et al., 2016; Paul et al., 2015 http://tools.iedb.org/tepitool/
FlowJo 10 FlowJo https://www.flowjo.com/
GraphPad Prism 8.4 GraphPad https://www.graphpad.com/
LEGENDplex v8.0 Biolegend https://www.biolegend.com/

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Dr. Ales-
sandro Sette (alex@lji.org).

Materials Availability
Aliquots of synthesized sets of peptides utilized in this study will be made available upon request. There are restrictions to the avail-
ability of the peptide reagents due to cost and limited quantity.

Data and Code Availability

The published article includes all data generated or analyzed during this study, and summarized in the accompanying tables, figures
and Supplemental materials.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human Subjects
Healthy Unexposed Donors
Samples from healthy adult donors were obtained by the La Jolla Institute for Immunology (LJI) Clinical Core or provided by a
commercial vendor (Carter Blood Care) for prior, unrelated studies between early 2015 and early 2018. These samples were
considered to be from unexposed controls, given that SARS-CoV-2 emerged as a novel pathogen in late 2019, more than one
year after the collection of any of these samples. These donors were considered healthy in that they had no known history of
any significant systemic diseases, including, but not limited to, autoimmune disease, diabetes, kidney or liver disease, congestive
heart failure, malignancy, coagulopathy, hepatitis B or C, or HIV. An overview of the characteristics of these unexposed donors is
provided in Table 1.

Cell 181, 1489–1501.e1–e6, June 25, 2020 e2

 ll
Article

The LJI Institutional Review Board approved the collection of these samples (LJI; VD-112). At the time of enrollment in the
initial studies, all individual donors provided informed consent that their samples could be used for future studies, including
this study.
Convalescent COVID-19 Donors
The Institutional Review Boards of the University of California, San Diego (UCSD; 200236X) and La Jolla Institute (LJI; VD-214)
approved blood draw protocols for convalescent donors. All human subjects were assessed for capacity using a standardized
and approved assessment. Subjects deemed to have capacity voluntarily gave informed consent prior to being enrolled in the study.
Individuals did not receive compensation for their participation.
Study inclusion criteria included subjects over the age of 18 years, regardless of disease severity, race, ethnicity, gender, preg-
nancy or nursing status, who were willing and able to provide informed consent, or with a legal guardian or representative willing
and able to provide informed consent when the participant could not personally do so. Study exclusion criteria included lack of
willingness or ability to provide informed consent, or lack of an appropriate legal guardian or representative to provide informed
consent.
Blood from convalescent donors was obtained at a UC San Diego Health clinic. Blood was collected in acid citrate dextrose (ACD)
tubes and stored at room temperature prior to processing for PBMC isolation and plasma collection. A separate serum separator
tube (SST) was collected from each donor. Samples were de-identified prior to analysis. Other efforts to maintain the confidentiality
of participants included referring to specimens and other records via an assigned, coded identification number.
Prior to enrollment in the study, donors were asked to provide proof of positive testing for SARS-CoV-2, and screened for clinical
history and/or epidemiological risk factors consistent with the World Health Organization (WHO) or Centers for Disease Control and
Prevention (CDC) case definitions of COVID-19 or Persons Under Investigation (PUI) (https://www.who.int/emergencies/diseases/
novel-coronavirus-2019/technical-guidance/surveillance-and-case-definitions, https://www.cdc.gov/coronavirus/2019-nCoV/hcp/
clinical-criteria.html). Per CDC and WHO guidance, clinical features consistent with COVID-19 included subjective or measured
fever, signs or symptoms of lower respiratory tract illness (e.g., cough or dyspnea). Epidemiologic risk factors included close contact
with a laboratory-confirmed case of SARS-CoV-2 within 14 days of symptom onset or a history of travel to an area with a high rate of
COVID-19 cases within 14 days of symptom onset.
Disease severity was defined as mild, moderate, severe or critical based on a modified version of the WHO interim guidance,
‘‘Clinical management of severe acute respiratory infection when COVID-19 is suspected’’ (WHO Reference Number: WHO/2019-
nCoV/clinical/2020.4). Mild disease was defined as an uncomplicated upper respiratory tract infection (URI) with potential non-
specific symptoms (e.g., fatigue, fever, cough with or without sputum production, anorexia, malaise, myalgia, sore throat, dys-
pnea, nasal congestion, headache; rarely diarrhea, nausea and vomiting) that did not require hospitalization. Moderate disease
was defined as the presence of lower respiratory tract disease or pneumonia without the need for supplemental oxygen, without
signs of severe pneumonia, or a URI requiring hospitalization (including observation admission status). Severe disease was
defined as severe lower respiratory tract infection or pneumonia with fever plus any one of the following: tachypnea (respiratory
rate > 30 breaths per minute), respiratory distress, or oxygen saturation less than 93% on room air. Critical disease was defined as
the need for ICU admission or the presence of acute respiratory distress syndrome (ARDS), sepsis, or septic shock, as defined in
the WHO guidance document.
Convalescent donors were screened for symptoms prior to scheduling blood draws, and had to be symptom-free and approxi-
mately 3 weeks out from symptom onset at the time of the initial blood draw. Following enrollment, whole blood from convalescent
donors was run on a colloidal-gold immunochromatographic ‘lateral flow’ assay to evaluate for prior exposure to SARS-CoV-2. This
assay detects IgM or IgG antibodies directed against recombinant SARS-CoV-2 antigen labeled with a colloidal gold tracer (20/20
BioResponse CoronaCheck). Ninety percent of convalescent donors tested positive for IgM or IgG to SARS-CoV-2 by this assay
(Table 1).
Convalescent donors were California residents, who were either referred to the study by a health care provider or self-referred. The
majority (75%) of donors had a known sick contact with COVID-19 or suspected exposure to SARS-CoV-2 (Table 1). The most com-
mon symptoms reported were cough, fatigue, fever, anosmia, and dyspnea. Seventy percent of donors experienced mild illness.
Donors were asked to self-report any known medical illnesses. Of note, 65% of these individuals had no known underlying medical
illnesses.

METHOD DETAILS

Peptide Pools
Epitope MegaPool (MP) design and preparation
SARS-CoV-2 virus-specific CD4 and CD8 peptides were synthesized as crude material (A&A, San Diego, CA), resuspended in
DMSO, pooled and sequentially lyophilized as previously reported (Carrasco Pro et al., 2015). SARS-CoV-2 epitopes were predicted
using the protein sequences derived from the SARS-CoV-2 reference (GenBank: MN908947) and IEDB analysis-resource as previ-
ously described (Dhanda et al., 2019; Grifoni et al., 2020). Specifically, CD4 SARS-CoV-2 epitope prediction was carried out using a
previously described approach in Tepitool resource in IEDB (Paul et al., 2015; Paul et al., 2016), to select peptides with median
consensus percentile % 20, similar to what was previously described, but removing the resulting spike glycoprotein epitopes

e3 Cell 181, 1489–1501.e1–e6, June 25, 2020

 ll
Article

from this prediction (CD4-R (remainder) ‘‘Non-spike’’ MP, n = 221). This approach takes advantage of the extensive cross-reactivity
and repertoire overlap between different HLA class II loci and allelic variants to predict promiscuous epitopes, capable of binding
across the most common HLA class II prototypic specificities (Greenbaum et al., 2011; O’Sullivan et al., 1991; Sidney et al.,
2010a, b; Southwood et al., 1998). The algorithm utilizes predictions for seven common HLA-DR alleles (DRB1*03:01,
DRB1*07:01, DRB1*15:01, DRB3*01:01, DRB3*02:02, DRB4*01:01 and DRB5*01:01) empirically determined to allow coverage of
diverse populations and for different pathogens and antigen systems (Dhanda et al., 2018; Paul et al., 2015).
To investigate in-depth spike-specific CD4 T cells, 15-mer peptides (overlapping by 10 amino acids) spanning the entire antigen
have been synthesized and pooled separately (CD-4 S (spike) MP, n = 253).
In the case of CD8 epitopes, since the overlap between different HLA class I allelic variants and loci is more limited to specific
groups of alleles, or supertypes (Sidney et al., 2008), we targeted a set of the 12 most prominent HLA class I A and B alleles
(A*01:01, A*02:01, A*03:01, A*11:01, A*23:01, A*24:02, B*07:02, B*08:01, B*35:01, B*40:01, B*44:02, B*44:03), which have been
shown to allow broad coverage of the general population. CD8 SARS-CoV-2 epitope prediction was performed as previously re-
ported, using NetMHC pan EL 4.0 algorithm (Jurtz et al., 2017) for the top 12 more frequent HLA alleles and selecting the top 1
percentile predicted epitope per HLA allele clustered with nested/overlap reduction (Grifoni et al., 2020). The 628 predicted CD8 epi-
topes were split in two CD8 MPs containing 314 peptides each (CD8-A and CD8-B). The CMV MP is a pool of previously reported
class I and class II epitopes (Carrasco Pro et al., 2015).
Protein peptide pools
In the case of the protein pools, peptides of 15 amino acid length overlapping by 10 spanning each entire protein sequence were
tested in a single MP (6-253 peptides per pool). Table S1 lists the number of peptides pooled for each of the viral proteins. Upon
request we are prepared to make these MP available to the scientific community for use in a diverse set of investigations.

PBMC isolation
For all samples whole blood was collected in ACD tubes (COVID-19 donors) or heparin coated blood bag (healthy unexposed do-
nors). Whole blood was then centrifuged for 15 min at 1850 rpm to separate the cellular fraction and plasma. The plasma was
then carefully removed from the cell pellet and stored at 20C.
Peripheral blood mononuclear cells (PBMC) were isolated by density-gradient sedimentation using Ficoll-Paque (Lymphoprep,
Nycomed Pharma, Oslo, Norway) as previously described (Weiskopf et al., 2013). Isolated PBMC were cryopreserved in cell recovery
media containing 10% DMSO (GIBCO), supplemented with 10% heat inactivated fetal bovine serum, depending on the processing
laboratory, (FBS; Hyclone Laboratories, Logan UT) and stored in liquid nitrogen until used in the assays.

SARS-CoV-2 RBD ELISA

SARS-CoV-2 Receptor Binding Domain (RBD) protein was obtained courtesy of Florian Krammer and Peter Kim (Stadlbauer et al.,
2020). Corning 96-well half-area plates (ThermoFisher 3690) were coated with 1mg/mL SARS-CoV-2 RBD overnight at 4 C. ELISA
protocol generally followed that of the Krammer lab, which previously demonstrated specificity (Stadlbauer et al., 2020). Plates
were blocked the next day with 3% milk (Skim Milk Powder ThermoFisher LP0031 by weight/volume) in Phosphate Buffered Saline
(PBS) containing 0.05% Tween-20 (ThermoScientific J260605-AP) for 2 hours at room temperature. Plasma was then added to the
plates and incubated for 1.5 hours at room temperature. Prior to plasma addition to the plates, plasma was heat inactivated at 56 C
for 30-60 minutes. Plasma was diluted in 1% milk in 0.05% PBS-Tween 20 starting at a 1:3 dilution and diluting each sample at by 1:3.
Plates were then washed 5 times with 0.05% PBS-Tween 20. Secondary antibodies were diluted in 1% milk in 0.05% Tween-20 and
incubated for 1 hour. For IgG, anti-human IgG peroxidase antibody produced in goat (Sigma A6029) was used at a 1:5000 dilution. For
IgM, anti-human IgM peroxidase antibody produced in goat (Sigma A6907) was used at a 1:10,000 dilution. For IgA, anti-human IgA
horseradish peroxidase antibody (Hybridoma Reagent Laboratory HP6123-HRP) was used at a 1:1,000 dilution. Plates were washed
5 times with 0.05% PBS-Tween 20. Plates were developed with TMB Substrate Kit (ThermoScientific 34021) for 15 minutes at room
temperature. The reaction was stopped with 2M sulfuric acid. Plates were read on a Spectramax Plate Reader at 450 nm using Soft-
Max Pro, and ODs were background subtracted. A positive control standard was created by pooling plasma from six convalescing
COVID-19 patients. Positive control standard was run on each plate and was used to calculate titers (relative units) for all samples
using non-linear regression interpolations, done to quantify the amount of anti-RBD IgG, anti-RBD IgM, and anti-RBD IgA present in
each specimen. Titers were plotted for each specimen and compared to COVID-19 negative specimens. As a second analytical
approach, Area under the curve was also calculated for each specimen to compare COVID-19 to negative specimens, using a base-
line of 0.05 for peak calculations.

OC43 and NL63 coronavirus RBD ELISA

An in-house ELISA at UNC was performed by coating with recombinant S RBD antigens (SARS-CoV-2, SARS-CoV, OC43-CoV and
NL63-CoV) in TBS for 1 h at 37 C. After blocking, we added 1:20 diluted serum and incubated at 37 C for 1 h. Antigen-specific
antibodies (Ig) were measured at 405 nm by using alkaline phosphatase conjugated goat anti-human IgG, IgA and IgM Abs and
4-Nitrophenyl phosphate.

Cell 181, 1489–1501.e1–e6, June 25, 2020 e4

 ll
Article

Flow Cytometry
Direct ex vivo PBMC immune cell phenotyping
For the surface stain, 1x106 PBMCs were resuspended in 100 ml PBS with 2% FBS (FACS buffer) and stained with antibody cocktail
for 1 hour at 4 C in the dark. Following surface staining, cells were washed twice with FACS buffer. Cells were then fixed/permea-
bilized for 40min at 4C in the dark using the eBioscience FoxP3 transcription factor buffer kit (ThermoFisher Scientific, Waltham, MA).
Following fixation/permeabilization, cells were washed twice with 1x permeabilization buffer, resuspended in 100 ml permeabilization
buffer and stained with intracellular/intranuclear antibodies for 1 hour at 4 C in the dark. Samples were washed twice with 1x per-
meabilization buffer following staining. After the final wash, cells were resuspended in 200ml FACS buffer. All samples were acquired
on a BD FACSymphony cell sorter (BD Biosciences, San Diego, CA). A list of antibodies used in this panel can be found in Table S2.
T cell stimulations
For all flow cytometry assays of stimulated T cells, cryopreserved cells were thawed by diluting them in 10 mL complete RPMI 1640
with 5% human AB serum (Gemini Bioproducts) in the presence of benzonase [20ul/10mL]. All samples were acquired on a ZE5 Cell
analyzer (Bio-rad laboratories), and analyzed with FlowJo software (Tree Star, San Carlos, CA).
Activation induced cell marker assay
Cells were cultured for 24 hours in the presence of SARS-CoV-2 specific MPs [1 mg/ml] or 10 mg/mL PHA in 96-wells U bottom plates
at 1x106 PBMC per well. A stimulation with an equimolar amount of DMSO was performed as negative control, phytohemagglutinin
(PHA, Roche, 1 mg/ml) and stimulation with a combined CD4 and CD8 cytomegalovirus MP (CMV, 1 mg/ml) were included as positive
controls. Supernatants were harvested at 24 hours post-stimulation for multiplex detection of cytokines. Antibodies used in the AIM
assay are listed in Table S4. AIM assays shown in Figures 2 and 3 and AIM assays shown in Figure 6 had five COVID-19 donors in
common and nine Unexposed donors. Full raw data is listed in Table S6.
Intracellular cytokine staining assay
For the intracellular cytokine staining, PBMC were cultured in the presence of SARS-CoV-2 specific MPs [1 mg/ml] for 9 hours. Golgi-Plug
containing brefeldin A (BD Biosciences, San Diego, CA) and monensin (Biolegend, San Diego, CA) were added 3 hours into the culture.
Cells were then washed and surface stained for 30 minutes on ice, fixed with 1% of paraformaldehyde (Sigma-Aldrich, St. Louis, MO) and
kept at 4 C overnight. Antibodies used in the ICS assay are listed in Table S5. The gates applied for the identification of IFNg, GzB, TNFa,
or IL-10 production on the total population of CD8+ T cells were defined according to the cells cultured with DMSO for each individual.

Cytokine bead assays

Supernatants were collected from 24-hour stimulation cultures of the AIM assays and stored in 96 well plates at 20 C. Cytokines in
cell culture supernatants of the same samples used for AIM were quantified using a human Th cytokine panel (13-plex) kit (LEGEND-
plex, Biolegend) according to the manufacturer’s instruction. Supernatants were mixed with beads coated with capture antibodies
specific for IL-5, IL-13, IL-2, IL-6, IL-9, IL-10, IFNg, TNFa, IL-17a, IL-17F, IL-4, IL-21 and IL-22 and incubated on a 96 well filter plate
for 2 hours. Beads were washed and incubated with biotin-labeled detection antibodies for 1 hour, followed by a final incubation with
streptavidin-PE. Beads were analyzed by flow cytometry using a FACS Canto cytometer. Analysis was performed using the
LEGENDplex analysis software v8.0, which distinguishes between the 13 different analytes on basis of bead size and internal dye.

Identification of coronavirus epitopes and associated literature references

To identify coronavirus epitopes and associated references, the IEDB was searched (on April 16, 2020) utilizing the following queries.
A first query was run to identify references associated with class I restricted CD8 epitopes, which utilized the criteria settings ‘‘An-
tigen’’: Organism = Coronavirus (taxonomy ID 11118); ‘‘Assay’’: Positive assays only; ‘‘Assay’’: T cell assay; ‘‘MHC restriction’’ = MHC
Class II; no parameters were defined for ‘‘Host’’ or ‘‘Disease.’’ This query identified 57 references, which are listed and displayed
under the ‘‘References’’ tab on the results page.
A second query was run to identify references associated with class II restricted CD4 epitopes which utilized the criteria settings
‘‘Antigen’’: Organism = Coronavirus (taxonomy ID 11118); ‘‘Assay’’: Positive assays only; ‘‘Assay’’: T cell assay; ‘‘MHC restriction’’ =
MHC Class II; no parameters were defined for ‘‘Host’’ or ‘‘Disease.’’ This query identified 27 references, which are listed and
displayed under the ‘‘References’’ tab on the results page.
A third query was run to specifically capture epitopes and map them back to the antigen of origin using the setting; ‘‘Antigen’’: Or-
ganism = Coronavirus (taxonomy ID 11118); ‘‘Assay’’: Positive assays only; ‘‘Assay’’: T cell assay; no parameters were defined for
‘‘MHC restriction,’’ ‘‘Host’’ or ‘‘Disease.’’ Results were exported as csv files, and then examined in Excel to tabulate the number
of CD4 and CD8 epitopes recognized in humans, mice, transgenic mice and other hosts associated with each respective antigen.

QUANTIFICATION AND STATISTICAL ANALYSIS

Data and statistical analyses were done in FlowJo 10 and GraphPad Prism 8.4, unless otherwise stated. The statistical details of the
experiments are provided in the respective figure legends. Data plotted in linear scale were expressed as Mean + Standard Deviation
(SD). Data plotted in logarithmic scales were expressed as Geometric Mean + Geometric Standard Deviation (SD). Correlation an-
alyses were performed using Spearman, while Mann-Whitney or Wilcoxon tests were applied for unpaired or paired comparisons,
respectively. Details pertaining to significance are also noted in the respective legends. T cell data have been calculated as

e5 Cell 181, 1489–1501.e1–e6, June 25, 2020

 ll
Article

background subtracted data or stimulation index. Background subtracted data were derived by subtracting the percentage of AIM+
cells after SARS-CoV-2 stimulation from the DMSO stimulation. Stimulation Index was calculated instead by dividing the percentage
of AIM+ cells after SARS-CoV-2 stimulation with the percentage of AIM+ cells derived from DMSO stimulation. If the AIM+ cells per-
centage after DMSO stimulation was equal to 0, the minimum value across each cohort was used. When two stimuli were combined
together, the percentage of AIM+ cells after SARS-CoV-2 stimulation was combined and either subtracted twice or divided by twice
the value of the percentage of AIM+ cells derived from DMSO stimulation. Additional data analysis techniques are described in the
STAR Methods sections above.

Cell 181, 1489–1501.e1–e6, June 25, 2020 e6

 ll
Article

Supplemental Figures
A B C
anti-RBD IgG anti-RBD IgM anti-RBD IgA
1.5 2.0 3
1565 AK4 1565 AK4 1565 AK4
1570 AK1 1570 AK1 1570 AK1
1592 AK2 1.5 1592 AK2 1592 AK2
1.0 2079 AK1 2 2079 AK1
2079 AK1

OD450
OD450
2095 AK1
OD450

2095 AK1 2095 AK1

1.0
4745 4745 4745
0.5 3SJD1 3SJD1 1 3SJD1
4747 0.5 4747 4747
4748 4748 4748
0.1 4750 0.1 4750 0.1 4750
0.0 0.0 0
1 10 100 1000 1 10 100 1000 1 10 100 1000

D E F
104 104 104
**** **** ****
SARS-CoV-2 spike RBD IgM (AUC)
SARS-CoV-2 spike RBD IgG (AUC)

SARS-CoV-2 spike RBD IgA (AUC)

103 103
103

102 102
102
101 101

101
100 100

10-1 10-1 100

Neg COVID Neg COVID Neg COVID

Figure S1. SARS-CoV-2 Spike Protein RBD Serology, Related to Figure 1

(A-C) ELISA curves for (A) IgG, (B) IgM, and (C) IgA from 10 representative donors. Five COVID-19 cases and
(D-F) Area under the curve (AUC) SARS-CoV-2 spike protein RBD (D) IgG (E) IgM, and (F) IgA, ELISA quantitation, from the same donors and experiments shown
in Figure 1. Geometric mean titers with geometric SDs are indicated. P values are two-tailed Mann-Whitney tests.
 ll
Article

Figure S2. Phenotyping Flow Cytometry, Related to Figure 1

Representative gating of CD3+ T cells, CD19+ B cells, CD3-CD19- cells, CD4+ T cells, CD8+ T cells and CD14+ monocytes from donor PBMCs is shown. Briefly,
mononuclear cells were gated out of all events followed by subsequent singlet gating. Live cells are gated as Zombie UV-. Cells were then gated as CD19-PE-
Cy5+, CD3-buv395+ or CD19-CD3- cells. T cells were further subdivided into either CD8-buv805+ or CD4-PerCPefluor710+ populations. CD3-CD19- cells were
defined as CD56-PE-Dazzlebright NK cells, CD56dimCD-16buv737+ NK cells or CD56- monocytes. Monocytes were further classified on differential expression of
CD14-bv510 and CD16.
 ll
Article

Figure S3. SARS-CoV-2-Specific CD4+ T Cell Responses of Recovered COVID-19 Patients, Related to Figure 2
(A) Example flow cytometry gating strategy.
(B) FACS plot examples for controls. DMSO negative control, CMV positive control, PHA positive control.
(C) CMV-specific CD4+ T cells as percentage of AIM+ (OX40+CD137+) CD4+ T cells after stimulation of PBMCs with CMV peptide pool. Data were background
subtracted against DMSO negative control and are shown with geometric mean and geometric standard deviation. Samples were from unexposed donors
(‘‘Unexposed,’’ n = 11) and recovered COVID-19 patients (‘‘COVID-19,’’ n = 10).
(D) Spearman correlation of SARS-CoV-2 spike specific CD4+ T cells (AIM+ (OX40+CD137+) CD4+ T cells, background subtracted) after stimulation with spike
pool run on the same donors in two independent experiment series run on different dates. COVID-19 patient samples shown in blue. Unexposed donor samples
shown in black.
(E-F) Stimulation index quantitation of AIM+ (OX40+CD137+) CD4+ T cells; the same samples as in Figure 2 and Figure S3C were analyzed.
(G-H) Cytokine levels in the supernatants of AIM assays after stimulation with (G) Spike MP (MP_S), or (H) CD4-R (‘‘Non-spike’’). Data are shown in comparison to
the negative control (DMSO), per donor.

(legend continued on next page)

 ll
Article

(I-J) Cytokine production by CD4+ T cells in response to Non-spike (CD4-R MP) or Spike (MP_S) peptide pools (‘‘CoV antigen (Ag)’’) was confirmed by analyzing
cytokine secretion from the subset of COVID-19 donors determined to have low or negative CD8+ T cell responses (< 0.1% by AIM) to the same peptide pool
determined positive for SARS-CoV-2 specific CD4+ T cells by AIM. (I) IL-2. (J) IFNg.
Statistical comparisons across cohorts were performed with the Mann-Whitney test, while paired sample comparisons were performed with the Wilcoxon test.
**p < 0.01; ***p < 0.001. ns not significant.
 ll
Article

Figure S4. SARS-CoV-2-Specific CD8+ T Cell Responses of Recovered COVID-19 Patients, Related to Figure 3
(A) Flow cytometry gating strategy.
(B) SARS-CoV-2 specific CD8+ T cells as determined by AIM+ (CD69+CD137+) CD8+ T cells. Response of PBMCs from COVID-19 cases between the negative
control (DMSO) and antigen specific stimulation.
(C) CMV-specific CD8+ T cells as percentage of AIM+ (CD69+CD137+) CD8+ T cells after stimulation of PBMCs with CMV peptide pool. Data were background
subtracted against DMSO negative control and are shown with geometric mean and geometric standard deviation. Samples were from unexposed donors
(‘‘Unexposed,’’ n = 11) and recovered COVID-19 patients (‘‘COVID-19,’’ n = 10).
(D-E) Stimulation index quantitation of AIM+ (CD69+CD137+) CD8+ T cells; the same samples as in Figure 2 and Figure S4C were analyzed.
Statistical comparisons across cohorts were performed with the Mann-Whitney test, while paired sample comparisons were performed with the Wilcoxon test.
**p < 0.01; ***p < 0.001. ns not significant.
 ll
Article

Figure S5. Correlations between SARS-CoV-2-Specific CD4+ T Cells, Antibodies, and CD8+ T Cells, Related to Figure 4
(A) Correlation between SARS-CoV-2 spike specific CD4+ T cells and anti-spike RBD IgG, using CD4+ T cell stimulation index.
(B) Correlation between SARS-CoV-2 non-spike specific CD4+ T cells and anti-spike RBD IgG, using CD4+ T cell stimulation index.
(C) Correlation between SARS-CoV-2 spike specific CD4+ T cells (%) and anti-spike RBD IgA.
(D) Correlation between SARS-CoV-2 spike specific CD4+ T cells (%) and anti-spike RBD IgA.
(E) Correlation between SARS-CoV-2 specific CD4+ T cells and SARS-CoV-2 specific CD8+ T cells, using stimulation index. Total MP responses per donor
were used in each case (‘‘Non-spike’’ + ‘‘spike’’ (CD4_R + MP_S) for CD4+ T cells, CD8-A + CD8-B for CD8+ T cells).
Statistical comparisons were performed using Spearman correlation.
 ll
Article

Figure S6. Protein Immunodominance of SARS-CoV-2 Specific CD4+ T Cells in Recovered COVID-19 Patients and Unexposed Donors,
Related to Figure 6
(A) The same data as Figure 6B, but with each unexposed donor color coded.
(B) The same experiment as Figure 6B, but with SARS-CoV-2 specific CD4+ T cells measured as percentage of AIM+ (OX40+CD137+) CD4+ T cells, after
background subtraction. COVID-19 cases (top, in blue. n = 10) and unexposed donors (bottom, in white. n = 10).

(legend continued on next page)

 ll
Article

(C) Correlation of SARS-CoV-2 specific CD4+ T cells detected using the epitope prediction approach (CD4_R MP) compared against the sum total of all antigen
pools of overlapping peptides (excluding spike), run with samples from the same donors in two different experiment series. Dotted line indicates 1:1 concordance.
Statistical comparison was performed using Spearman correlation.
 ll
Article

Figure S7. Protein Immunodominance of SARS-CoV-2-Specific CD8+ T Cells in Recovered COVID-19 Patients and Unexposed Donors,
Related to Figure 6
(A) The same data as Figure 6D, but with each unexposed donor color coded.
(B) The same experiment as Figure 6D, but with SARS-CoV-2 specific CD8+ T cells measured as percentage of AIM+ (CD69+CD137+) CD8+ T cells, after
background subtraction. COVID-19 cases (top, in red. n = 10) and unexposed donors (bottom, in gray. n = 10).
 Article

Hybrid Gene Origination Creates Human-Virus

Chimeric Proteins during Infection
Graphical Abstract Authors
Jessica Sook Yuin Ho, Matthew Angel,
IAV mRNA synthesis Yixuan Ma, ..., Jonathan W. Yewdell,
Edward Hutchinson, Ivan Marazzi
5’ of Host mRNA Cap During infection

Cleavage of host mRNA Cap Correspondence

edward.hutchinson@glasgow.ac.uk
Annealing to viral genomic RNA vPol viral RNA
(E.H.),
Initiation of viral RNA transcription ivan.marazzi@mssm.edu (I.M.)
Generation of chimeric host-virus mRNA viral mRNA
In Brief
The process by which RNA viruses, such
Translation
as influenza virus, cleave capped host
Known mechanism Novel mechanism transcripts to drive viral mRNA
Viral proteins Chimeric proteins production leads to the translation of host
Cap
Cap UTR and viral RNA to make hybrid proteins
AUG AUGGAA ...
UTR
{

AUGGAA ... that then generate T cell responses and

{

In-frame Off-frame
contribute to virulence.
Met E ... Met X X Met E ... Met - X X X X X X ...

Canonical viral proteins Viral proteins with Upstream Frankenstein ORF (UFO)
host-encoded extensions Novel host-virus encoded proteins

Highlights
d A mechanism of hybrid gene birth is employed by many
families of RNA viruses

d Human RNA and viral RNA encode new genes together

d Hybrid genes either make extensions of viral proteins or

novel proteins (UFOs)

d Human-virus genes and proteins play roles in pathogenesis

and are conserved

Ho et al., 2020, Cell 181, 1502–1517

June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.035 ll
 ll
OPEN ACCESS

Article
Hybrid Gene Origination Creates
Human-Virus Chimeric Proteins during Infection
Jessica Sook Yuin Ho,1,26 Matthew Angel,2,26 Yixuan Ma,1,26 Elizabeth Sloan,14,26 Guojun Wang,1,12,25
Carles Martinez-Romero,1,12,13 Marta Alenquer,15 Vladimir Roudko,6,7,8,9 Liliane Chung,16 Simin Zheng,1 Max Chang,4
Yesai Fstkchyan,1 Sara Clohisey,16 Adam M. Dinan,17 James Gibbs,2 Robert Gifford,14 Rong Shen,20 Quan Gu,14
Nerea Irigoyen,17 Laura Campisi,1 Cheng Huang,19 Nan Zhao,1 Joshua D. Jones,17,22 Ingeborg van Knippenberg,14,23
Zeyu Zhu,1 Natasha Moshkina,1 Léa Meyer,14 Justine Noel,1 Zuleyma Peralta,5 Veronica Rezelj,14,24 Robyn Kaake,3
Brad Rosenberg,1 Bo Wang,16 Jiajie Wei,2 Slobodan Paessler,19 Helen M. Wise,16 Jeffrey Johnson,1,3
Alessandro Vannini,20,21 Maria João Amorim,15 J. Kenneth Baillie,16 Emily R. Miraldi,10,11 Christopher Benner,4
Ian Brierley,17 Paul Digard,16 Marta quksza,5 Andrew E. Firth,17 Nevan Krogan,3 Benjamin D. Greenbaum,6,7,8,9
Megan K. MacLeod,18 Harm van Bakel,5 Adolfo Garcı̀a-Sastre,1,12,13 Jonathan W. Yewdell,2 Edward Hutchinson,14,27,*
and Ivan Marazzi1,12,27,28,*
1Department of Microbiology, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
2Laboratory of Viral Diseases, National Institute of Allergy and Infectious Diseases, NIH, Bethesda, MD 20892, USA
3Department of Cellular and Molecular Pharmacology, University of California, San Francisco, San Francisco, CA 94158, USA
4Department of Medicine, School of Medicine, University of California San Diego, La Jolla, CA 92037, USA
5Department of Genetics and Genomic Sciences, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
6Tisch Cancer Institute, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
7Department of Medicine, Hematology and Medical Oncology, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
8Department of Oncological Sciences, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
9Department of Pathology, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
10Divisions of Immunobiology and Biomedical Informatics, Cincinnati Children’s Hospital, Cincinnati, OH 45229, USA
11Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH 45257, USA
12Global Health and Emerging Pathogens Institute, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
13Division of Infectious Diseases, Department of Medicine, Icahn School of Medicine at Mount Sinai, New York, NY 10029, USA
14MRC-University of Glasgow Centre for Virus Research, Glasgow G61 1QH, UK
15Instituto Gulbenkian de Ciência, 2780-156 Oeiras, Portugal
16The Roslin Institute, University of Edinburgh, Edinburgh EH25 9PS, UK
17Division of Virology, Department of Pathology, University of Cambridge, Cambridge CB2 0SP, UK
18Centre for Immunobiology, Institute of Infection, Immunity and Inflammation, University of Glasgow, Glasgow G12 8QQ, UK
19Department of Pathology, the University of Texas Medical Branch, Galveston, TX 77555, USA
20Division of Structural Biology, The Institute of Cancer Research, London SW7 3RP, UK
21Fondazione Human Technopole, Structural Biology Research Centre, 20157 Milan, Italy
22Present address: Infection Medicine, Edinburgh Medical School: Biomedical Sciences, University of Edinburgh, Edinburgh, UK
23Present address: Department of Learning and Teaching Enhancement, Sighthill Court, Edinburgh Napier University, Edinburgh, UK
24Present address: Viral Populations and Pathogenesis Unit, Department of Virology, Institut Pasteur, CNRS UMR 3569, Paris, France
25Present address: The State Key Laboratory of Reproductive Regulation and Breeding of Grassland Livestock, College of Life Sciences,

Inner Mongolia University, Hohhot, China

26These authors contributed equally
27Senior author
28Lead Contact

*Correspondence: edward.hutchinson@glasgow.ac.uk (E.H.), ivan.marazzi@mssm.edu (I.M.)

https://doi.org/10.1016/j.cell.2020.05.035

SUMMARY

RNA viruses are a major human health threat. The life cycles of many highly pathogenic RNA viruses like influ-
enza A virus (IAV) and Lassa virus depends on host mRNA, because viral polymerases cleave 50 -m7G-capped
host transcripts to prime viral mRNA synthesis (‘‘cap-snatching’’). We hypothesized that start codons within
cap-snatched host transcripts could generate chimeric human-viral mRNAs with coding potential. We report
the existence of this mechanism of gene origination, which we named ‘‘start-snatching.’’ Depending on the
reading frame, start-snatching allows the translation of host and viral ‘‘untranslated regions’’ (UTRs) to create
N-terminally extended viral proteins or entirely novel polypeptides by genetic overprinting. We show that
both types of chimeric proteins are made in IAV-infected cells, generate T cell responses, and contribute
to virulence. Our results indicate that during infection with IAV, and likely a multitude of other human, animal
and plant viruses, a host-dependent mechanism allows the genesis of hybrid genes.

1502 Cell 181, 1502–1517, June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Article
INTRODUCTION
In eukaryotes, ribosomes typically recognize mRNAs with a ter-
minal 50 cap structure followed by an untranslated region (UTR),
which can be tens to hundreds of nucleotides in length (Decroly
et al., 2011; Kochetov et al., 2008; Leppek et al., 2018). However,
a growing body of work has shown that translation can initiate in
the 50 UTRs of a large proportion of eukaryotic mRNAs, some-
times extremely close to the 50 cap, resulting in upstream open
reading frames (uORFs) (Andreev et al., 2015; Calvo et al.,
2009; Dikstein, 2012; Elfakess and Dikstein, 2008; Haimov
et al., 2017; Johnstone et al., 2016; Kochetov et al., 2008; Young
and Wek, 2016).
A large subphylum of RNA viruses, the segmented negative
strand RNA viruses (sNSVs), makes direct use of the 50 termini
of host mRNAs when transcribing their own genes. The sNSVs
include the families Arenaviridae, Peribunyaviridae, and Ortho-
myxoviridae. Highly contagious human and animal viruses like
influenza A virus (IAV) and Lassa virus (LASV) belong to these
families and are responsible for significant levels of morbidity
and mortality worldwide. In sNSVs, viral mRNA synthesis is
primed using short 50 methyl-7-guanosine (m7G) capped RNA
sequences, which the viral polymerase cleaves from host RNA
polymerase II (RNAPII) transcripts in a process known as ‘‘cap-
snatching’’ (Dias et al., 2009; Plotch et al., 1981; Reich et al.,
2014; Rialdi et al., 2017). Cap-snatching creates viral transcripts
that are genetic hybrids of host and viral sequences, with the
host-derived 50 sequences being highly diverse (Gu et al.,
2015; Koppstein et al., 2015; Rialdi et al., 2017; Sikora et al.,
2017). Once made, viral mRNAs are translated by the host
machinery.
In this manuscript, we hypothesized that by appropriating 50
terminal mRNA sequences from their hosts, sNSVs could obtain
functional upstream start codons (uAUGs), a mechanism we
termed ‘‘start-snatching.’’ Translation from host-derived up-
stream start codons in chimeric host-viral transcripts would ac-
cess upstream viral ORFs (uvORFs). Depending on the frame of
the uAUG relative to that of the canonical viral protein, two novel
chimeric types of protein in infected cells could be generated:
canonical viral proteins with host and viral UTR-derived N-termi-
nal extensions, and previously uncharacterized proteins read
from ORFs that are out-of-frame with, and overprinted on,
canonical viral ORFs. Below, we report on how we tested this hy-
pothesis using genomics, cell biology, virology, and phyloge-
netic analyses.
RESULTS
IAV Cap-Snatches Sequences Containing uAUGs
IAV gene transcription is initiated by cap-snatching from a host
mRNA (Figure 1A). This process generates an IAV mRNA with
a 50 end portion derived from the host. This mechanism is used
to express viral genes that encode canonical viral proteins (Fig-
ure 1B, OUTCOME 1). We hypothesized that AUGs within host
sequences could generate upstream host-virus chimeric ORFs
with coding potential. Depending on the reading frame, a host-
derived uAUG might initiate the synthesis of two novel chimeric
genes encoding for an N-terminally extended viral protein (Fig-
ll
OPEN ACCESS
ure 1B, OUTCOME 2, upper panel) or alternatively, an entirely
novel protein overprinted on the canonical viral ORF (Figure 1B,
OUTCOME 2, lower panel). These outcomes are contingent on
two assumptions: (1) uAUGs are present in cap-snatched host
sequences and can enable translation initiation, and (2) the 50
mRNA transcribed from the viral UTR should lack stop codons.
Furthermore, the absence of stop codons interrupting UTRs or
the downstream ORFs should be evolutionarily conserved.
To address the first point, we determined the abundance of
uAUGs in cap-snatched host sequences archived in a Decap
and 50 end sequencing (DEFEND-seq) dataset (Rialdi et al.,
2017) that we had previously generated from A549 cells infected
with the IAV A/Puerto Rico/8/34(H1N1) (PR8) (Figure 1C). AUG-
containing, host-derived capped sequences (Figure 1C, red
bars) ranged from 7–20 nt, with a median length of 11 nt, similar
to the distribution obtained for all cap-snatched sequences (Fig-
ure 1C, gray bars). Host-derived oligonucleotides with AUG co-
dons were present at similar ratios in all eight genome segments
of the virus and were present in all three reading frames, consti-
tuting 12% of all cap-snatched sequences (Figures 1D and
S1A). Similar results were also obtained when we performed
cap analysis of gene expression (CAGE) on primary human
monocyte-derived macrophages infected with a different strain
of IAV (A/Udorn/72(H3N2); Udorn) (Figure S1B; Table S1). These
results indicate that, upon infection, neither the virus nor the host
cells appear to prevent the formation of chimeric RNAs with
hybrid coding potential.
IAV 50 UTRs Are Translatable
We next performed a bioinformatic analysis to determine if stop
codons were absent from IAV sequences within the 50 UTRs and,
if so, whether this was evolutionarily conserved across IAV
strains. First, we analyzed the nucleotide sequence variability
of the 50 UTRs of all eight segments, using all IAV H1N1 strains
available from the NCBI Influenza Virus database (Zhang et al.,
2017). 50 UTRs of each individual segment are highly conserved
within each individual segment, as shown by the positional
weight matrices (Figure S2, top panels) and sequence alignment
(Figure S2, lower panels). We then translated the 50 UTR of each
genome segment in silico in all possible frames (Figure 2A, upper
panels) This revealed that the 50 UTR of every IAV genome
segment can maintain a reading frame in at least one frame (Fig-
ure 2A, upper panels, stop codons indicated by red boxes).
We found that the 50 UTRs of five out of the eight genome seg-
ments (PB2, HA, NP, NA, and NS) lacked upstream stop codons
in-frame with the major ORF (Figure 2A, upper panels, major ORF
start codons indicated by green boxes). These segments thus
have the potential to code for N-terminally extended viral pro-
teins. Stop codons were also absent from the 50 UTRs of six of
the eight genome segments when these were read out of frame
with the major ORF (Figure 2A; segments PB2, PB1, PA, NA, M,
and HA). This suggested the intriguing possibility that, in the
presence of a host-donated start codon, these genome seg-
ments could make novel genes encoding hybrid polypeptides.
To probe the length of uvORFs, we translated viral sequences
that had cap-snatched uAUGs in our dataset in silico. The result
of these analyses (Figure 2A, lower panels) indicated the general
propensity to create chimeric ORFs, with half of the viral genome
Cell 181, 1502–1517, June 25, 2020 1503
 ll
OPEN ACCESS Article

A Host mRNA viral mRNA (vmRNA)

Infection vRNA
cap cap cap cap cap
1 - Cleavage 2 - Annealing to viral 3 - Primed viral RNA
genomic RNA transcription

B OUTCOME 1. Known mechanism Viral Protein

AUGGAA...
Met Glu ... canonical viral protein
{

cap UTR

OUTCOME 2. Novel mechanism Chimeric Protein

Depending M X M D V N ... Viral protein with human-encoded extension

AUGGAA... on frame
G
AU
{

cap UTR M X X X X X X... New human-virus encoded protein

C D
40
11 nt Percent of all unique CS sequences 15
Percentage of CS sequences

that contain uAUGs

30

10

20

5
10

0 0
0 5 10 15 20 PB2 PB1 PA HA NP NA M NS
CS length (nt) Segment
All CS Sequences CS containing AUGs

Figure 1. Upstream AUGs Are Present in Host-Derived Sequences of Viral mRNAs

(A) Schematic of cap-snatching during the transcription of a segmented negative sense RNA virus (sNSV) such as influenza A virus (IAV).
(B) Schematic showing how the presence of upstream AUGs (uAUGs) in host-derived cap-snatched RNA sequences may drive the formation of novel host-viral
chimeric proteins.
(C) Histograms showing the length distributions of all cap-snatched (CS) sequences (gray bars) or only CS sequences containing uAUGs (red bars) in A549 cells
infected with IAV (strain PR8) for 4 h, as determined by DEFEND-seq.
(D) Bar plots showing the percentages of uAUG containing CS sequences in each IAV genome segment.

segments predicted to make sizable products (>30 aa) (Fig-
ure 2B). These ORFs overlap with canonical viral genes but are
read in different frames (overprinted). They range from over 40
residues (HA) to nearly 80 residues (PB1). Where N-terminal
extensions of the major ORF were possible, these ranged from
8–21 aa in length (Figure 2B).
Thus, uvORFs are present in all genome segments and, if
licensed by host-derived uAUG-containing RNAs, could
generate polypeptides of varying length (Figure 2B).
Host-Virus mRNA Chimeras Associate with Elongating
Ribosomes
If cap-snatched host uAUGs did initiate translation of viral 50
UTRs, the 50 termini of viral mRNAs would be bound by initiating
ribosomes. We therefore performed ribosomal profiling of IAV in-
fected cells, in the presence of harringtonine, which blocks elon-
gation of de novo assembled 80S initiation complexes but not of
those already engaged in elongation. Ribosome-protected frag-
ments (RPFs) were mapped to both the human and viral ge-
nomes (Figures 3A and S3A–S3C). Mapping of RPF sequences
revealed an accumulation of ribosomes at the canonical initiation
site in mRNAs transcribed from all eight genome segments
(Figure 3B; main ORF AUG), consistent with previous reports
(Machkovech et al., 2019). As well as observing ribosomes accu-
mulating at the canonical initiation sites, we also observed RPFs
mapping to the host-derived sequence upstream of the 50 UTR,
suggesting that translation initiated in this region (Figure 3B, in-
sets). The total number of RPF reads mapping to host-derived
sequences for each segment was 5%–20% of the reads map-
ping to the canonical start codon (Figure 3C), broadly consistent
with the proportion of cap-snatched sequences containing
uAUGs (Figures 1D and S1B).

1504 Cell 181, 1502–1517, June 25, 2020

 ll
Article OPEN ACCESS

Figure 2. IAV 50 UTRs Are Conserved and Translatable

(A) Sequence analysis of all unique 50 UTR sequences from each segment of 10,904 H1N1 subtype IAV genomes (coding sense), showing (upper panels) the
translation of the 50 UTR in all three reading frames; and (lower panels) the predicted amino acid length (aa) distributions of N-terminal extensions to the major
gene product and of overprinted new ORFs. This is calculated from the distribution of uAUG positions in DEFEND-seq data and (for overprinted new ORFs) from
the position of stop codons in the IAV PR8.
(B) The numbers of translatable products that could be accessed from uAUGs in each genome segment of IAV.

Cell 181, 1502–1517, June 25, 2020 1505

 ll
OPEN ACCESS Article

A mRNA 1
C
mRNA 2

Proportion of total CS Reads relative

to Reads at the Main ORF AUG
Ribo 1
0.20
Ribo 2
Ribo + Harr 1 0.15
Ribo + Harr 2
0.10
0 25 50 75 100
Aligned reads (%)
0.05
B
PB2 Segment PB1 Segment
Normalized Read Counts

1500 20
0.00
1500 100 Canonical
CS Reads Canonical CS Reads
15 AUG 80 AUG PB2 PB1 PA HA NP NA M NS
1000 10 1000
60 Segment
40
5 20
500 0 500 0
-20 -15 -10 -5 0 -20 -15 -10 -5 0

0 0
-20 -15 -10 -5 0 5 10 15 20 25 -20 -15 -10 -5 0 5 10 15 20

PA Segment HA Segment D
40
Normalized Read Counts

800 20 50
CS Reads Canonical 25000
CS Reads

Sequences with AUG

15 AUG 40 Canonical
600 20000 AUG

Percentage CS
30
10 30
15000 20
400 5
10
0
10000 DMSO
200 -20 -15 -10 -5 0 0
5000 -20 -15 -10 -5 0 20 Harringtonine
0 mRNA
-20 -15 -10 -5 0 5 10 15 20 0
-20 -10 0 10 20 30
10
NP Segment NA Segment
Normalized Read Counts

10000 200 8000 50

CS Reads CS Reads Canonical
40
8000 150
Canonical 6000 30 AUG 0
6000
100 AUG All Segments
20
50 4000
4000 10
0 2000 0
2000 -20 -15 -10 -5 0 -20 -15 -10 -5 0

0 0
-20 -10 0 10 20 30 40 -20 -15 -10 -5 0 5 10 15 20

M Segment NS Segment
Normalized Read Counts

25000 250 15000 250

CS Reads Canonical CS Reads
200 AUG 200
20000
150 150 Canonical
10000 AUG
15000 100 100

10000 50 50
5000 0
0
-20 -15 -10 -5 0 -20 -15 -10 -5 0
5000

0 0
-20 -15 -10 -5 0 5 10 15 20 -20
20 -15
1 -10
10 -5 0 5 10 15 20 25

Distance from first viral nucleotide Distance from first viral nucleotide

Figure 3. IAV mRNAs Can Be Translated from Host-Derived AUGs

(A) Proportion of reads that align to viral and human transcripts for the indicated experimental conditions.
(B) 50 end mapping of ribosome protected fragments (RPFs) in harringtonine-treated A549 cells infected with the IAV PR8 at 8 h post-infection, showing for each
segment of the IAV genome the distribution of reads in the cap-snatched regions (shown in insets) and virally encoded mRNA up to 10 nt after the canonical start
codon. The x axis is shown relative to the first virally encoded nucleotide.
(C) For each IAV genome segment, the number of ribosome-protected fragments (RPFs) upstream of the canonical AUG as a proportion of those mapping to the
canonical AUG is shown. Data are shown as the mean ± SD.
(D) Barplots showing the percentages of RPFs that contain an AUG when cells were treated with DMSO (black bars) or harringtonine (gray bars) immediately prior
to harvest, or from total mRNA-seq (white bars). Results from two sequencing replicates are shown as points, with bars showing the mean.

Precisely mapping initiation sites very close to the cap is chal- segment), and less frequent toward the 50 end of the host-derived
lenging, because many of the heterogeneous 50 mRNA ends would sequence (Figure S3D). As well as inferring upstream ribosome
be too short to extrude from the ribosome, making P-site phasing initiation by mapping RPFs to protected uAUGs, we could test for
problematic by standard Ribo-seq analysis. To address this, we it directly by comparing ribosomal profiles with and without harring-
used the location of AUGs within the RPF to identify the reading tonine arrest. Harringtonine increased the proportion of RPFs from
frame being translated. This suggested that initiation occurred in cap-snatched sequences that contained an AUG, indicating trans-
all three reading frames (Figure S3D). uAUG codons were more lation was initiating on uAUGs in these host-derived sequences
frequently close to the start of the viral UTR sequences, peaking (Figure 3D). Taken together, our data show that translation initiates
at the 2 position of mRNAs from all genome segments (numbered from cap-snatched host-derived uAUGs in viral mRNA chimeras,
from the first position in the coding sense of the viral genome albeit at lower frequencies than at canonical start codons.

1506 Cell 181, 1502–1517, June 25, 2020

 ll
Article OPEN ACCESS

A B

C D

E F

(legend on next page)

Cell 181, 1502–1517, June 25, 2020 1507

 ll
OPEN ACCESS Article

Host-Virus Protein Chimeras Are Expressed during
Infection, Recognized by T Cells, and Affect Virulence
To demonstrate that chimeric proteins are expressed during
infection, we performed mass spectrometry analyses of cell ly-
sates from infected cells. We also checked whether any chimeric
proteins could be integrated into viral progeny by analyzing pu-
rified virions (Figures 4A, S4A, and S4B).
There are limitations to this approach, as the likelihood of a
tryptic digest generating peptides that can be detected by the
mass spectrometer is lower for short proteins. This issue
reduces the chance of finding peptides derived from small over-
printed uvORFs (<30 aa), or that map to short N-terminal exten-
sions. Nevertheless, we were able to identify at least 2 distinct
peptides that were derived from the two long overprinted uvORFs
in the PB1 and PB2 segments, which we named PB1-UFO and
PB2-UFO, respectively (for ‘‘Upstream Frankenstein ORF’’). In
addition, we detected a UTR-encoded N-terminal extension of
NP, which we named NP-extension (NP-ext) (Figures 4A, 4B,
S4A, and S4B; Table S2A). Peptides from all three proteins were
present in PR8 IAV infected cell lysates (Figures 4B, left panels,
and S4A; Table S2A). These novel viral peptides were not de-
tected in uninfected controls (Figure S4A). We were also able to
identify peptides derived from the PB1-UFO protein when we
re-analyzed three previously published proteomic datasets of
IAV infection (Heaton et al., 2016) (Figure S4C; Table S2C). Only
NP-ext was detected in virions (Figure S4B; Table S2B), presum-
ably because influenza virions specifically package hundreds of
copies of NP, while there is no known mechanism to specifically
package other uvORF-encoded proteins (Hutchinson et al., 2014).
Quantification of the PB1-UFO, PB2-UFO, and NP-ext pro-
teins indicated that, although they are less abundant than the
major viral proteins, they are expressed at detectable levels
within an infected cell. When quantified, tryptic peptides from
these proteins were found between the 20th and 40th percentile
of normalized peptide intensities, including both host and viral
proteins, within our samples (Figures S4A and S4B). Taken
together, our data show that N-terminal extensions and over-
printed uvORFs are synthesized during IAV infection and are pre-
sent at a moderate abundance within infected cells.
We next asked whether chimeric host-viral proteins could
be recognized by the host’s immune system. To test this,
we created modified IAVs containing insertions of a class
I-restricted epitope of ovalbumin (Porgador et al., 1997). Based
on the uvORFs predicted from our in silico analyses, we inserted
the epitope (OVAI; OVA 257-264; SL8; SIINFEKL) in frame with
the longest uvORF (PB1 frame 3 uvORF; PB1-UFO(SIIN) (Fig-
ure 4C) and one of the shortest uvORFs (NS, frame 2 uvORF;
NS-UFO(SIIN) (Figure 4D). In the case of PB1 segment, we inte-
grated sequences encoding OVAI directly into the UTR, placing
the epitope within the uvORF encoding PB1-UFO (Figure 4C, top
panels). For the NS segment, we used synonymous mutations in
the canonical viral gene to delete five naturally occurring stop co-
dons in the uvORF; we then inserted OVAI into the extended
uvORF, positioning the insertion in a flexible ‘‘linker’’ region of
the major viral gene NS1 (Thulasi Raman and Zhou, 2016). This
genetic configuration was chosen to ascertain whether uvORFs
are translated by default provided that they are not interrupted by
stop codons (Figure 4D, top panels).
Mouse DC2.4 cells infected with PB1-UFO(SIIN) activated trans-
genic OT-I CD8+ T cells (that are highly specific for mouse H-2 Kb
class I molecule complexed with SIINFEKL; Kb-SIIN) (Hogquist
et al., 1994) as determined by upregulation of CD25 and CD69 (Fig-
ure 4C, lower panels). Recombinant IAV expressing SIIN(PB1-Ub-
SIIN) at high levels (Wei et al., 2019) was used as a positive control
(Figure 4C, right panels). No upregulation of CD25 and CD69 was
observed in mock treated samples. Similar results were obtained
with the NS-UFO(SIIN) virus. Here, OT-I CD8+ T cells were acti-
vated when incubated with bone marrow-derived dendritic cells
(BMDCs) infected with the NS-UFO(SIIN) virus (Figure 4D, right
panels). This was comparable to the activation seen in a control
experiment using a virus in which OVAI was inserted into the
stem of the viral NA protein (NA-SIIN) (Figure 4D, middle panels)
(Bottermann et al., 2018). Again, noo upregulation was observed
during mock infection. Taken together, our data with both the
PB1-UFO(SIIN) and the NS-SIIN viruses indicate that, unless
blocked by stop codons, uvORFs are translated and expressed
during infection, and T cell immunosurveillance extends to pep-
tides encoded by uvORFs.
Next, to probe if the expression of chimeric host-viral proteins
has an impact on viral pathogenesis, we generated a battery of
recombinant viruses, in which specific N-terminal extensions
or uvORFs were knocked out through the introduction of
premature stop codons (NP-Dext and UFOD, respectively). The
viruses were generated either in the PR8 (Figures 4E, 4F, and

Figure 4. uvORFs Are Expressed during Infection and Can Contribute to Virulence
(A) The number of upstream viral open reading frames (uvORFs) that could be translated for each segment of the IAV genome (empty circles), highlighting those
detected in infected cell lysates by mass spectrometry (filled red circles).
(B) Tryptic peptides that map to translated uvORFs, detected by mass spectrometry across multiple experiments (summarizing data in Figures S4A and S4C).
(C) Schematic showing the generation of the PB1-UFO(SIIN) virus. DC2.4 cells were infected with the indicated viruses and co-cultured with OT-I CD8+ T cells.
OT-1 activation, assessed by CD69 and CD25 expression, was assayed by flow cytometry at 24 h post co-culture. vmRNA, viral mRNA.
(D) Schematic showing the generation of the NS-SIIN virus. Red bars indicate stop codons mutated to permit uninterrupted NS1-UFO translation. Mouse BMDC
cells were incubated with IAV antigen presentations, and co-cultured with OT1-CD8+ T cells. OT-I activation, assessed by CD69 and CD25 expression, was
assayed by flow cytometry of CD69 and CD25 expression at 24 h post co-culture.
(E) Upper panel: schematic showing mutations that truncate NP-ext (NP-DEXT) and control mutations (NP-SYN), as engineered into the IAV PR8. Wild-type PR8 is
also shown. Lower panel: weight loss and survival curves of 6- to 8-week-old BALB/c mice infected with 15 plaque-forming unit (PFU)/mouse of the indicated
viruses. Data are an aggregate of 2 independent experiments of n = 3 mice, using 2 independently plaque purified clones of the NP-DEXT or PR8;NP-SYN viruses
(total n = 6/condition). *p < 0.05; data are shown as the mean ± SEM.
(F) Upper panel: schematic showing mutations that knocked out PB1-UFO (PB1-UFOD) and control mutations (PB1-UFOSYN), as engineered into the IAV PR8.
Wild-type PR8 is also shown. Lower panel: weight loss and survival curves of 6- to 8-week-old BALB/c mice infected with the indicated dose (per mouse) of the
indicated viruses. n = 10 mice/condition. *p < 0.05. Data are shown as the mean ± SEM.

1508 Cell 181, 1502–1517, June 25, 2020

 ll
Article OPEN ACCESS

chemistry
A Acidic Basic Hydrophobic Neutral Polar

4
Bits

2
0

100
Conservation

75
50
25
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77

amino acid position

B H1N1 H3N2 H5N1

5.92% 0.35% PB1-UFO lengths
3.36% 2.00% 0.20% 14.58%
14.43%
>70 aa
50-70 aa
30-50 aa
< 30 aa

90.72% 83.37% 85.07%

C
Positive selection
Propagator Model Analysis G(X) : Chance that a class of mutation Mutation class likely to be propagated
in PB1-UFO reaches frequency X
g(X) : quantifies the relative likelihood that a
test class of mutations reaches frequency > x G0(X): Chance that a neutral class of mutations

g(X)
1 Neutral / Heterogenous selection
when compared to a neutral class of mutations reaches the same frequency X.
G X Neutral class defined as synonymous mutations
g X = that occur in PB1 reading frame Negative selection
G0 X and do not overlap with PB1-UFO Mutation class not likely to be propagated
Frequency, X

D Consider all nucleotides E Consider all nucleotides F Consider all nucleotides

R1 R2 R3 R1 R2 R3 R1 R2 R3
PB1-UFO Frame Test PB1-UFO Frame Test PB1-UFO Frame

PB1 Frame Neutral class PB1 Frame Neutral class PB1 Frame Test Neutral class

1.0 1.6 1.5

1.5 1.4
0.9 1.4 1.3
1.3 1.2
Propagator ratio, g(X)

Propagator ratio, g(X)

Propagator ratio, g(X)

0.8
1.2 1.1
0.7 1.1 1.0
1.0 0.9
0.6
0.9 0.8
0.5 0.8
0.7
0.7
0.4 0.6
0.6
0.5
0.3 0.5
0.4 0.4
0.2 0.3 0.3
0.2 0.2
0.1
0.1 0.1
0.0 0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Frequency, X Frequency, X Frequency, X

G 15 H
Position of epitopes within aa sequence of PB1-UFO
10
number of unique # of unique
identities, log2 HLA-epitope pairs
5
(Kd < 500nM)
0 0 77
100

HLA-B5801 10
Unique PB1 sequences
Percent identity across

HLA-B4001

HLA-B3901

HLA-B2705
50

HLA-B1501

HLA-B0801
5
HLA-B0702

HLA-A2601

HLA-A2402
0

HLA-A0301

HLA-A0201 0 50 100 150 200 232

HLA-A0101
0 Nucleotide Position

(legend on next page)

Cell 181, 1502–1517, June 25, 2020 1509

 ll
OPEN ACCESS Article

S4D), A/WSN/33(H1N1) (WSN) (Figure S4E), or mouse-adapted
A/California/04/2009(H1N1) (Cal09) (Figure S4F) backgrounds.
We also generated the reciprocal control viruses carrying synon-
ymous mutations (NPSYN;UFOSYN). Both genomic configurations
of control and knockout viruses maintained intact the canonical
viral ORFs (Table S3).
The mutant viruses did not display gross alterations in viral
growth in vitro (Figures S4D–S4F). This was independent of
viral background and also of the cell type infected (Figures
S4D–S4F). To determine if interrupting upstream translation
had effects in vivo, we focused on the NP-Dext and PB1-
UFOD viruses in the PR8 background. The strategy used to
generate these viruses is shown in the top panels of Figures
4E and 4F.
We found that the NP-Dext viruses were less virulent in mice
compared to the control NP-SYN viruses (Figure 4E), suggest-
ing that NP-ext expression contributes to virulence. A similar
role for NP-ext was recently proposed for the pandemic
2019 IAV (pdm2009) strain, in which an extended NP protein
was found to contribute to virulence in mice and pigs (Wise
et al., 2019). Importantly, however, pdm2009 viruses translate
NP-ext from a uAUG encoded in the 50 UTR of NP, but no cor-
responding uAUG is encoded by the PR8 virus used in
our study.
The PB1-UFOD viruses displayed increased virulence when
compared to the PB1-UFOSYN viruses in vivo, although in this
case an effect was only observed at high infectious doses (Fig-
ure 4F). Gene expression analyses suggested that there were
distinct transcriptomic signatures in the lungs of mice infected
with high doses of the PB1-UFOD or PB1-UFOSYN viruses
(Figures S4G and S4H; Table S4A). Gene Ontology analysis of
differentially expressed genes indicated changes in a number
of pathways, including leukocyte activation and pro-inflamma-
tory cytokine secretion (Figure S4I; Table S5). Immune cell dys-
regulation may therefore be at least partially responsible for the
differences in morbidity and mortality during infection with the
PB1-UFOD or PB1-UFOSYN viruses.
Together, these functional data show that uvORFs are ex-
pressed during IAV infections, can be detected by the adap-
tive immune system, and can modulate the severity of
infection.
Chimeric Host-IAV Proteins Are Conserved
We next asked if NP-ext and PB1-UFO are conserved across
different strains. The ability to express NP-ext without interrup-
tion by stop codons in the 50 UTR was maintained in 99.9% of
IAV isolates present in the NCBI Influenza database (Zhang
et al., 2017) (Figures S2, S5A, and S5B). Sequence analysis of
the translated 50 UTR also suggested that N-terminally extended
sequences would be similar within IAV subtypes (Figure S5C).
There are many reasons why these sequences are conserved,
including constraints imposed by RNA structure and the require-
ment to interact with the viral polymerase complex (Fodor, 2013).
Whatever the primary selective pressure, the result of the con-
servation of the 50 UTR sequence is that the ability to express
NP-ext is nearly universal among IAV strains.
The ability to express PB1-UFO requires not only a lack of stop
codons in the appropriate frame of the 50 UTR, but also the main-
tenance of a uvORF overprinted on the canonical PB1 ORF. We
first analyzed sequences of the IAV subtypes H1N1, H3N2, and
H5N1. We found that PB1-UFO is conserved within each of these
three virus subtypes (Figure 5A), and stop codons resulting in
PB1-UFO proteins <77 aa long were infrequent (Figure 5B).
To understand the factors that contribute to the maintenance
of PB1-UFO ORF length and amino acid sequence composition
within the IAV, we first looked at the probability that an ORF
similar in length to PB1-UFO could have arisen stochastically
in the IAV PB1 segment. We used a sequence randomization
model (Figure S5D) on the H3N2 subtype of IAV, the subtype
for which the greatest number of complete sequences were
available. We found that 77% of the sequences in the NCBI
Influenza database (Zhang et al., 2017) encoded a 77-aa PB1-
UFO (Figure S5E) that is significantly longer than the 15–30
aa long ORFs expected by chance (Figures S5E–S5G). We
also found that these predicted ORFs would require multiple
(30–70) additional synonymous mutations in order to generate
an ORF that is of similar length to PB1-UFO (Figure S5H).
The above analysis does not take into account constraints
imposed by nucleotide biases in the viral UTR or canonical
PB1 ORF or from viral RNA structure. To examine their roles in
the maintenance of the PB1-UFO ORF we used the frequency
propagator method (Luksza and Lässig, 2014; Strelkowa and
Lässig, 2012) (Figures 5C and S6A). This method can determine

Figure 5. uvORFs Are Conserved

(A) Conservation analysis of PB1-UFO protein sequences across all IAV subtypes.
(B) Pie charts showing percentages of sequences in H1N1, H3N2, and H5N1 IAV subtypes that have a PB1-UFO that is 77 aa long (blue), 50–77 aa long (gray), 30–
50 aa long (orange), and <30 aa long (yellow).
(C) Outline of the propagator model analysis. Diagrams describe possible outcomes and interpretations of calculated g(x) ratios
(D) Frequency propagator ratios of the indicated classes of mutations occurring in PB1-UFO relative to the PB1 open reading frame of H3N2 viruses. Top: regions
used for the test (G(x); yellow), and neutral class (G0(X); blue) ratios are shown. The test class is the region of PB1-UFO ORF that overlaps only with the viral 50 UTR;
the neutral class consists of synonymous mutations in the PB1 ORF that do not overlap with PB1-UFO. All nucleotides positions were considered. Error bars
indicate sampling uncertainties. See also Figure 5C for interpretations
(E) Frequency propagator ratios, as in (D), but with the test class comprising the C-terminal region of the PB1-UFO ORF.
(F) Frequency propagator ratios, as in (D), but with the test class comprising the region in the main PB1 ORF overlapping the PB1-UFO reading frame.
(G) Number of predicted PB1-UFO epitope-allele interactions for frequent 11 human HLA alleles. Heatmaps show number of PB1-UFO epitopes derived from all
possible unique identities and predicted to bind selected MHC-I alleles. Number of unique identities (i.e., unique influenza A virus sequences) encoding predicted
epitopes are shown in histograms, next to the heatmaps.
(H) Locations of PB1-UFO peptides that are predicted to result in strong (Kd <500 nM) unique interacting HLA-epitope pairs across the PB1-UFO reading frame.
This plot is juxtaposed with percent identity plot of PB1-UFO (lower panel) across 3,140 unique PB1-UFO sequences taken from the NCBI Influenza Database
(Zhang et al., 2017).

1510 Cell 181, 1502–1517, June 25, 2020


Article
if these factors imposed constraints on the PB1-UFO amino acid
sequence. The model and its possible outcomes are shown and
discussed in detail in Figures 5C and S6A and the STAR
Methods.
Briefly, mutations that occur in the viral UTR region, which
encodes the N-terminal part of PB1-UFO, undergo negative se-
lection (Figure 5D; g < 1). This indicates that mutations in the viral
UTR, should they occur, have a low probability of being propa-
gated down the IAV strain tree. On the other hand, when we
consider the nucleotide sequences that encode the overlapping
regions of PB1-UFO and the canonical PB1 ORF, we see that
there is heterogeneous/neutral selection occurring on mutations
in the PB1-UFO ORF (g z 1). This is most likely shaped by the
requirement to maintain the main PB1 ORF sequence, as muta-
tions that maintain the PB1 ORF aa sequence (synonymous mu-
tations in PB1 ORF) are more likely to be fixed in the population
(Figure 5E; red line; g < 1). Mutations that change the PB1 amino
acid sequence instead undergo negative selection (Figure 5F;
blue line; g < 1) and are unlikely to be propagated down the strain
tree, consistent with PB1 ORF being fixed and essential for IAV.
Selection in these regions is unlikely to be dominated by RNA
structural constraints because similar effects are observed when
RNA secondary structure is taken into account for our analysis
(Figures S6B–S6D). Overall, our analyses suggest that PB1-
UFO conservation is largely dictated by the need to preserve
both the viral UTR nucleotide sequence and the amino acid
sequence of the main PB1 ORF. Taken together, this suggests
that the evolution of the PB1-UFO ORF is heavily constrained
by converging selective pressures.
Because we had shown that peptides derived from PB1-UFO
could be presented to the immune system (Figures 4C and 4D),
we asked whether epitope-HLA class I interactions could play a
role in shaping PB1-UFO sequence. We found that multiple
unique PB1-UFO peptides were predicted to bind to and interact
with various HLA types (Figure 5G; Table S6). Notably, high-affin-
ity (<500 nM) HLA-epitope pairs were concentrated in regions of
PB1-UFO where conservation was low, suggesting that immune
pressure on PB1-UFO may lead to some diversifying selection
on the protein (Figure 5H).
Chimeric Host-Virus Proteins of Other Viruses
Finally, we asked whether our finding that start-snatching gener-
ates novel ORFs could be generalized from IAV to other sNSVs.
We began by looking at another member of the Orthomyxoviri-
dae family, influenza B virus (IBV), by performing DEFEND-seq
on A549 cells infected with IBV. The host-derived sequences
that IBV obtains by cap-snatching had comparable median
lengths to those appropriated by IAV (Figure S7A). Sequence
analysis indicates that uAUG-initiated translation could read
through the 50 UTR of every IBV genome segment in at least
one frame and predicted at least two long overprinted new
ORFs (PA and NA segments) (Figures 6A and 6B), as well as
N-terminal extensions of six of the eight major viral proteins (Fig-
ures 6A and S7B).
Next, we looked at other families of sNSVs. We performed
CAGE analysis on cells infected by Lassa virus (LASV), a member
of the family Arenaviridae and an emerging virus that in the past
decade has caused several epidemics of hemorrhagic fever.
ll
OPEN ACCESS
LASV genomes comprise two ambisense segments. The median
cap-snatched length of LASV mRNAs was seven nucleotides
(Figure S7C) in agreement with structural prediction of the
LASV polymerase (Wallat et al., 2014). Sequence analysis indi-
cates that these uAUGs could lead to the translation of N-termi-
nal extensions of the GPC protein, as well as the formation of two
overprinted new ORFs of 50 and 80 aa from the viral mRNAs
encoding the nucleoprotein (N) and Z proteins of LASV (Fig-
ure 6C, 6D, and S7D). The proportions of uAUGs detected in
cap-snatched sequences from IBV and LASV were dependent
on viral segments and ranged between 4% and 12% (Table S7).
We also tested the hypothesis that translation of UTR-derived
sequences could occur in other sNSVs by using minireplicon as-
says encoding a luciferase reporter to a member of the Phenui-
viridae (Heartland banyangvirus; L segment UTRs). By mutating
the canonical AUG, we identified low but readily detectable
levels of upstream translation (Figure S7E).
Overall, these data suggest that generation of chimeric virus-
host ORFs is a common feature of sNSVs. To quantify the poten-
tial pervasiveness of this mechanism and the likelihood of novel
ORFs being conserved and functionalized into new genes, we
analyzed RNA virus genomic sequences for their propensity to
generate novel proteins by performing in silico analyses of their
genomes. Although the exact levels of upstream translation will
depend on a range of factors, including the intrinsic properties
of viral polymerase complexes and, potentially, mechanisms
that modulate upstream AUG translation, our results indicate
the genomic potential of start-snatching (Figure 7). Given that
viral mRNA and proteins are among the most highly expressed
biotypes in infected cells, our data support the idea that all
cap-snatching virus could expand their proteome by start-
snatching uAUGs from their hosts.
DISCUSSION
In this manuscript, we describe the existence of a mechanism em-
ployed by sNSVs to generate chimeric host-virus genes. This
mechanism, ‘‘start-snatching,’’ involves the co-opting of start co-
dons from host mRNA sequences to expand the viral proteome.
This mechanism appears to be accessible to all sNSVs, including
major human pathogens such as IAV and LASV. Start-snatching
allows the translation of proteins from cryptic uvORFs, either as
canonical viral proteins with N-terminal extensions, or as UFO pro-
teins overprinted on the canonical viral ORF. In this study, we have
identified examples of both types of uvORF in IAV infections. We
have shown that translation can initiate on uAUGs in the host-
derived sequence of viral mRNAs, and that this leads to the
expression of chimeric host-virus proteins that can be detected
in infected cells. In our hands, the ablation of uvORFs did not
impact viral replication in vitro but had a moderate effect in vivo,
which would be consistent with uvORFs encoding accessory pro-
teins. We found that uvORFs can be recognized by the immune
system, and we modeled the contribution of different evolutionary
forces at play on uvORFs by characterizing viral-intrinsic and host-
immune features that contribute to their evolution. Finally, we
showed experimentally and by sequence analysis that the capa-
bility to express uvORFs through start-snatching is widespread
among the sNSVs.
Cell 181, 1502–1517, June 25, 2020 1511
 ll
OPEN ACCESS Article

A C

B D

Figure 6. uvORFs Are Encoded by Cap-Snatching Viruses from Diverse Families

(A) The number of host-virus chimeric protein species potentially encoded by influenza B virus (IBV; B/Wisconsin/01/2010).
(B) Sequence analysis of the PA and NA segments of IBV, showing the translation of the 50 UTR in all three reading frames and the predicted length distributions of
N-terminal extensions to the main ORF and of overprinted new ORFs, calculated from uAUG positions in DEFEND-seq data.
(C) The number of host-virus chimeric protein species potentially encoded by the ambisense genome of Lassa virus (LASV; Josiah strain), in both forward and
reverse senses. The ORF encoded by the segment is indicated in the square brackets.
(D) Sequence analysis of L and S segments of LASV in the indicated orientations, showing a schematic of genome organization, the translation of the 50 UTR in all
three reading frames, and the predicted length distributions of overprinted new ORFs, calculated from uAUG positions in CAGE-seq data.

Chimeric mRNAs Encode Novel Viral Proteins
We hypothesized that cap-snatching of sNSVs could generate
ORFs that are encoded by two genomes (human and virus).
Consistent with this, our analyses indicate that roughly 10% of
IAV mRNA contains host-derived uAUGs (Figures 1D and S1B).
Furthermore, uvORFs are translated in at least three of the eight
IAV genome segments, generating NP-ext, PB2-UFO, and PB1-
UFO (Figures 2 and 4). Genetic evidence suggested that many
other uvORF proteins are also likely to be expressed, although
we did not detect them in our current study, potentially due to

1512 Cell 181, 1502–1517, June 25, 2020

 ll
Article OPEN ACCESS

25 Canonical ORFs both of these effects (Figures 5D–5F). Despite this, we

New ORFs (>30 aa) also observe that (1) nonsense mutations do not occur
20 New Extensions frequently in the population (Figures S5D–S5H), and (2)
Number of ORFs

missense mutations that do eventually accumulate in

15 the PB1-UFO ORF tend to be those that change poten-
tially immunogenic epitopes (Figures 5G and 5H).
10
This information, and the mere fact that full-length PB1-UFO is
5
present in more than 75% of all IAV isolates and NP extensions
Viral Family are present in more than 99% of IAV isolates, suggests that mul-
0
Orthomyxoviridae tiple forces at the host-virus interface drive the virus to maintain
Influenza A
Influenza B
Influenza C
Thogoto
LCMV
Lassa
Tacaribe
Bunyamwera
La Crosse
Herbert

Hantaan
EMARV

Arenaviridae the full-length proteins in their sequences. The relative contribu-

Peribunyaviridae
tions of distinct evolutionary forces in maintaining these proteins
Fimoviridae
Hantaviridae
are not yet clear.
An important point to be made about uvORFs is that conserva-
Figure 7. Start-Snatching Increases the Number of Potential ORFs
tion and/or expression does not equate to functionality. While
in sNSVs some uvORFs might have gained functions, we predict others
The increase in number of potential ORFs in cap-snatching viruses when will exist as afunctional, evolutionary spandrels. Such uvORFs
uvORFs are considered. Black, number of canonical ORFs; yellow, number of are stuck in a place where they have to be made but suffer too
new overprinted ORFs >30 aa; red, number of new extensions. LCMV, lym- many external constraints to productively sample evolutionary
phocytic choriomeningitis virus; EMARV, European mountain ash ringspot- space for functionalization. All things considered, we can fairly
associated emaravirus.
surmise that any cost the virus might incur through uvORFs be-
ing made is outweighed by the fitness benefits of maintaining a
the substantial sequence overlap of N-terminal extensions with
genetic architecture that allows for their expression. The aware-
canonical viral proteins and the short lengths of many over-
ness of uvORF existence, and their pervasiveness in the viral
printed ORFs. Overall, our analysis indicates that multiple fam-
world, is thus critical for our understanding of viral biology, viral
ilies of viruses can generate chimeric RNAs and could produce
evolution, and host immune surveillance.
proteins via this mechanism (Figures 6 and 7).

Conservation and Function of uvORFs Gene Origination through Overprinting and the Mis-
Our analysis shows that most sNSV infections lead to expression naming of "UTRs"
of chimeric genes and uvORFs. Because they are host and virally Genetic overprinting typically occurs when a pre-existing
encoded, it is therefore reasonable to ask who benefits from their reading frame acquires mutations that enable translation in alter-
expression. Key considerations in this regard, and based on our native reading frames while maintaining the function of the
analysis are: ancestral frame. This is an important mechanism for the creation
of new proteins, especially in the context of compact genomes
(1) Epitopes encoded in uvORFs are recognized by the adap- (viral, prokaryotic, and eukaryotic organelles) with little coding
tive immune system. MHC I presentation of uvORF- capacity (Keese and Gibbs, 1992; Kovacs et al., 2010; Poulin
derived peptides poses the risk of an adaptive immune et al., 2003; Sabath et al., 2012).
response against cells infected with sNSVs, analogous While genetic overprinting could be selectively advantageous
to the risks posed to IAV by the presentation of alternative for some organisms, the evolution of overprinted genes is prob-
reading frames (ARFs) and defective ribosomal products lematic. Any evolution of the overprinted ORF will be constrained
(DRiPs) (Dolan et al., 2010; Wei et al., 2019; Wei and Yew- by the effects of mutations in the underlying ORF. In addition, es-
dell, 2017, 2019; Zanker et al., 2019). Indeed, the risks tablished overlapping ORFs typically have dedicated mecha-
posed to the virus by the presentation of uvORFs are nisms for their expression, such as ribosomal scanning or frame-
potentially even higher due to the high conservation of shifting, which allow for efficient and regulated expression
these sequences. patterns. Exploring the limited evolutionary space that satisfies
(2) Two uvORFs considered here (NP-ext and PB1-UFO) are all of these constraints presumably requires the overprinted
both highly conserved across multiple strains of IAV. gene to provide a strong selective advantage.
However, merely assessing conservation is insufficient, Start-snatching exposes the 50 coding regions of sNSV ge-
as other forms of selection also act on IAV genome se- nomes to low levels of non-specific out-of-frame translation.
quences. In particular, genome packaging signals in the This ‘‘genetic feature’’ could facilitate the evolution of novel
primary RNA sequence are concentrated in the terminal genes through genetic overprinting, without having to evolve a
regions of each genome segment (Dadonaite et al., dedicated method to express an overprinted ORF before that
2019; Gog et al., 2007; Hutchinson et al., 2010), resulting ORF could provide a selective advantage.
in a suppression of synonymous codon usage (Gog et al., A similar argument applies to the evolution of alternative up-
2007; Jagger et al., 2012). In overprinted regions, like stream translation mechanisms for N-terminally extended pro-
PB1-UFO, there is also selective pressure conferred by teins: if an N-terminal extension provided by start-snatching was
the sequence encoding the canonical ORF. We observe selectively advantageous, the virus could evolve to directly

Cell 181, 1502–1517, June 25, 2020 1513

 ll
OPEN ACCESS
encode an uAUG in the UTR and make the generation of extended
protein host-independent and heritable. In this respect, it is inter-
esting to note that some recent strains of IAV have evolved to
encode a uAUG in the UTR of NP that allows it to express an
N-terminally extended protein that can modulate virulence (Wise
et al., 2019). In essence, start-snatching might simply be a way
to increase the chances of UTR translation by outsourcing
uAUG to non-viral genomic material.
The translation of 50 UTRs (that implies their misnaming) oc-
curs frequently in eukaryotic genes. uORFs are, in fact, perva-
sively expressed, with some functioning as short biologically
active polypeptides (Andrews and Rothnagel, 2014; Calvo
et al., 2009; Combier et al., 2008; Sendoel et al., 2017; Starck
et al., 2016; Wang and Rothnagel, 2004; Wen et al., 2009).
uORFs are abundantly expressed in cancer cells (Sendoel
et al., 2017) and activated T cells (Starck et al., 2016). Overall,
future work will be needed to redefine what, in reality, a gene is.
Lessons for Other Viruses
The capacity of a pathogen to overcome host barriers and estab-
lish infection is based on the expression of pathogen-derived
proteins. To understand how a pathogen antagonizes the host
and establishes infection we need to have a clear understanding
of what proteins a pathogen encodes, how they function, and the
manner in which they contribute to virulence. The current dogma
about many life-threatening pathogens is that they encode a
small repertoire of proteins because of their limited genome
size. RNA viruses, such as IAV, are a prime example of this.
Here, we have shown that IAV, IBV, LASV, and likely most, if
not all, other sNSVs, can use host RNA to expand their genetic
repertoire. Similar to novel human genes which originated from
other mechanisms and contributed to organismal evolution
(Kaessmann, 2010; Ohno, 1970), we expect chimeric genes to
shape (and have shaped) host-virus relationships.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Cells cultures
B Mice
B Virus Strains
d METHOD DETAILS
B Growth kinetics of Viruses in Cell Culture
B Quantification of IAV titers by Plaque Assays
B Ribosome profiling and analysis
B Mass Spectrometry experiments (in infected cell
lysates)
B Mass Spectrometry experiments (in virions)
B DEFEND sequencing of IBV infected cells
B Preparation of CAGE libraries from LASV infected cells
1514 Cell 181, 1502–1517, June 25, 2020
Article
B Mouse Infection studies
B Preparation of RNA sequencing Libraries (In-
fected Mice)
B SIINFEKL expression analysis
B Minireplicon Assays
d QUANTIFICATION AND STATISTICAL ANALYSES
B Mouse Infection Studies
B Quantitative qPCR assays
B CAGE sequencing of WSN IAV virus infected cells
B Ribosome sequencing analyses
B RNA sequencing Analyses
B LASV CAGE sequencing Analyses
B Sequence Randomization Model for PB1-UFO length
B Frequency Propagator Ratio Analysis
B Epitope predictions for PB1-UFO
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.035.
ACKNOWLEDGMENTS
We thank the Genomics and Mouse facility at Icahn School of Medicine at
Mount Sinai, the Global Health and Emerging Pathogens Institute (GHEPI) at
Mount Sinai, and the entire Marazzi Lab team. The authors would also like to
thank Ervin Fodor, University of Oxford for support and critical comments on
the project; Svenja Hester, Benjamin Thomas, and Shabaz Mohammed of
the Advanced Proteomics Facility, University of Oxford for proteomics; Elly
Gaunt, University of Edinburgh and Michael Goodin, University of Kentucky,
for critical reading of the manuscript; and staff within the Institute of Infection,
Immunity, and Inflammation Flow Cytometry Facility, the Central Research Fa-
cility at the University of Glasgow, Thomas Purnell of the Institute of Infection,
Immunity and Inflammation, University of Glasgow and Dimitris Athineos of the
Beatson Institute for technical assistance and discussion. I.M. is supported by
the Burroughs Wellcome Fund (United States; 1017892) and by the Chan
Zuckerberg Initiative (United States; 2018-191895). I.M. and H.v.B. are sup-
ported by the NIH (United States; R01AI113186). A.G.-S. and I.M. are sup-
ported by the NIH (U19AI135972 FLUOMICS). E.H., E.S., and Q.G. are sup-
ported by an MRC (United Kingdom) Career Development Award (MR/
N008618/1), and E.H. carried out proteomics work when funded by an MRC
Programme Grant to Prof. Ervin Fodor, University of Oxford (MR/K000241/
1). V.R. and I.v.K. were supported by a Wellcome Trust (United Kingdom) Se-
nior Investigator award (099220/Z/12/Z) awarded to Prof. Richard M. Elliott,
University of Glasgow. R.G. and Q.G. were supported by the MRC
(MC_UU_12014/12). J.K.B. was supported by a Wellcome Trust Intermediate
Clinical Fellowship (103258/Z/13/Z), a Wellcome-Beit Prize (103258/Z/13/A),
and the UK Intensive Care Society. J.K.B. and S.C. acknowledge the BBSRC
(United Kingdom) Institute Strategic Programme Grant to the Roslin Institute.
B.W. was supported by a SHIELD (MR/N02995X/1) Edinburgh Global
Research Scholarship. A.E.F. was supported by the Wellcome Trust
(106207) and European Research Council (European Union; 646891). I.B.,
P.D., and H.M.W. were supported by an MRC project grant (MR/M011747/
1). P.D. was supported by the BBSRC Institute Strategic Programme (BB/
J004324/1 and BB/P013740/1). J.D.J. was funded by a Wellcome Trust PhD
scholarship. M. Alenquer and M.J.A. were supported by the FCT (Portugal)
award PTDC/BIA-CEL/32211/2017 and IF/00899/2013, respectively. M.K.M.
was supported by a Wellcome Investigator Award (210703/Z/18/Z).
AUTHOR CONTRIBUTIONS
Conceptualization, I.M., E.H., A.G.-S., J.W.Y., and E.S.; Methodology, I.M.,
J.W.Y., Y.M., M.A., G.W., and J.S.Y.H.; Formal Analysis, Y.M., M.A., G.W.,
J.S.Y.H., N.Z., J.N., N.M., J.G., J.W., J.J., M.C., Z.P., H.v.B., M.L., E.R.M.,

Article
and A.E.F.; Investigation, J.S.Y.H., M.A., Y.M., E.S., G.W., C.M.-R., M.A., V.R.,
L. Campisi, S.Z., M.C., Y.F., S.C., A.M.D., J.G., R.G., R.S., Q.G., N.I., L. Chung,
N.Z., J.D.J., I.v.K., Z.Z., N.M., L.M., J.N., Z.P., V.R., R.K., B.R., B.W., J.W.,
H.W., J.J., A.V., M.J.A., E.R.M., C.B., I.B., P.D., M.L., A.E.F., N.K., B.D.G.,
M.K.M., and H.v.B.; Data Curation, M.A., Y.M., H.v.B., R.G., and J.J.; Re-
sources, Y.M., M.A., J.J., M.C., H.v.B., E.R.M., A.G.-S., S.P., and C.H.; Writing
– Original Draft, I.M. and E.H.; Writing – Review & Editing, E.H., I.M., J.W.Y.,
Y.M., J.S.Y.H., M.A., E.S., A.M.D., H.v.B., M.L., B.D.G., E.R.M., A.G.-S.,
P.D., and A.E.F.; Visualization, E.H., Y.M., M.A., G.W., J.H., J.J., M.C., Z.P.,
E.R.M., and Z.Z.; Funding Acquisition, A.G.-S., I.M., J.K.B., A.E.F., I.B., P.D.,
M.J.A., E.H., and M.K.M.; Project Administration, I.M. and E.H.; Supervision,
E.H. and I.M.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: March 29, 2019
Revised: February 26, 2020
Accepted: May 18, 2020
Published: June 18, 2020
REFERENCES
Andreatta, M., and Nielsen, M. (2016). Gapped sequence alignment using arti-
ficial neural networks: application to the MHC class I system. Bioinformatics
32, 511–517.
Andreev, D.E., O’Connor, P.B., Fahey, C., Kenny, E.M., Terenin, I.M., Dmitriev,
S.E., Cormican, P., Morris, D.W., Shatsky, I.N., and Baranov, P.V. (2015).
Translation of 50 leaders is pervasive in genes resistant to eIF2 repression. eLife
4, e03971.
Andrews, S.J., and Rothnagel, J.A. (2014). Emerging evidence for functional
peptides encoded by short open reading frames. Nat. Rev. Genet. 15,
193–204.
Bottermann, M., Foss, S., van Tienen, L.M., Vaysburd, M., Cruickshank, J.,
O’Connell, K., Clark, J., Mayes, K., Higginson, K., Hirst, J.C., et al. (2018).
TRIM21 mediates antibody inhibition of adenovirus-based gene delivery and
vaccination. Proc. Natl. Acad. Sci. USA 115, 10440–10445.
Buchholz, U.J., Finke, S., and Conzelmann, K.K. (1999). Generation of bovine
respiratory syncytial virus (BRSV) from cDNA: BRSV NS2 is not essential for vi-
rus replication in tissue culture, and the human RSV leader region acts as a
functional BRSV genome promoter. J. Virol. 73, 251–259.
Calvo, S.E., Pagliarini, D.J., and Mootha, V.K. (2009). Upstream open reading
frames cause widespread reduction of protein expression and are polymor-
phic among humans. Proc. Natl. Acad. Sci. USA 106, 7507–7512.
Clohisey, S., Parkinson, N., Wang, B., Bertin, N., Wise, H., Tomoiu, A., Sum-
mers, K.M., Hendry, R.W., Carninci, P., Forrest, A.R.R., et al.; FANTOM5 Con-
sortium (2020). Comprehensive characterisation of transcriptional activity dur-
ing influenza A virus infection reveals biases in cap-snatching of host RNA
sequences. J. Virol. 94, e01720-19.
Combier, J.P., de Billy, F., Gamas, P., Niebel, A., and Rivas, S. (2008). Trans-
regulation of the expression of the transcription factor MtHAP2-1 by a uORF
controls root nodule development. Genes Dev. 22, 1549–1559.
Cox, J., and Mann, M. (2008). MaxQuant enables high peptide identification
rates, individualized p.p.b.-range mass accuracies and proteome-wide pro-
tein quantification. Nat. Biotechnol. 26, 1367–1372.
Dadonaite, B., Gilbertson, B., Knight, M.L., Trifkovic, S., Rockman, S., Laeder-
ach, A., Brown, L.E., Fodor, E., and Bauer, D.L.V. (2019). The structure of the
influenza A virus genome. Nat. Microbiol. 4, 1781–1789.
de Wit, E., Spronken, M.I., Bestebroer, T.M., Rimmelzwaan, G.F., Osterhaus,
A.D., and Fouchier, R.A. (2004). Efficient generation and growth of influenza vi-
rus A/PR/8/34 from eight cDNA fragments. Virus Res. 103, 155–161.
ll
OPEN ACCESS
Decroly, E., Ferron, F., Lescar, J., and Canard, B. (2011). Conventional and un-
conventional mechanisms for capping viral mRNA. Nat. Rev. Microbiol.
10, 51–65.
Dias, A., Bouvier, D., Crépin, T., McCarthy, A.A., Hart, D.J., Baudin, F., Cusack,
S., and Ruigrok, R.W. (2009). The cap-snatching endonuclease of influenza vi-
rus polymerase resides in the PA subunit. Nature 458, 914–918.
Dikstein, R. (2012). Transcription and translation in a package deal: the TISU
paradigm. Gene 491, 1–4.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Dolan, B.P., Li, L., Takeda, K., Bennink, J.R., and Yewdell, J.W. (2010). Defec-
tive ribosomal products are the major source of antigenic peptides endoge-
nously generated from influenza A virus neuraminidase. J. Immunol. 184,
1419–1424.
Edgar, R.C. (2004). MUSCLE: multiple sequence alignment with high accuracy
and high throughput. Nucleic Acids Res. 32, 1792–1797.
Elfakess, R., and Dikstein, R. (2008). A translation initiation element specific to
mRNAs with very short 5’UTR that also regulates transcription. PLoS ONE
3, e3094.
Fodor, E. (2013). The RNA polymerase of influenza a virus: mechanisms of viral
transcription and replication. Acta Virol. 57, 113–122.
Fodor, E., Devenish, L., Engelhardt, O.G., Palese, P., Brownlee, G.G., and Gar-
cı́a-Sastre, A. (1999). Rescue of influenza A virus from recombinant DNA.
J. Virol. 73, 9679–9682.
Forrest, A.R., Kawaji, H., Rehli, M., Baillie, J.K., de Hoon, M.J., Haberle, V.,
Lassmann, T., Kulakovskiy, I.V., Lizio, M., Itoh, M., et al.; FANTOM Consortium
and the RIKEN PMI and CLST (DGT) (2014). A promoter-level mammalian
expression atlas. Nature 507, 462–470.
Gaush, C.R., and Smith, T.F. (1968). Replication and plaque assay of influenza
virus in an established line of canine kidney cells. Appl. Microbiol. 16, 588–594.
Gog, J.R., Afonso, Edos.S., Dalton, R.M., Leclercq, I., Tiley, L., Elton, D., von
Kirchbach, J.C., Naffakh, N., Escriou, N., and Digard, P. (2007). Codon conser-
vation in the influenza A virus genome defines RNA packaging signals. Nucleic
Acids Res. 35, 1897–1907.
Gruber, A.R., Lorenz, R., Bernhart, S.H., Neuböck, R., and Hofacker, I.L.
(2008). The Vienna RNA websuite. Nucleic Acids Res. 36, W70-4.
Gu, W., Gallagher, G.R., Dai, W., Liu, P., Li, R., Trombly, M.I., Gammon, D.B.,
Mello, C.C., Wang, J.P., and Finberg, R.W. (2015). Influenza A virus preferen-
tially snatches noncoding RNA caps. RNA 21, 2067–2075.
Haimov, O., Sinvani, H., Martin, F., Ulitsky, I., Emmanuel, R., Tamarkin-Ben-
Harush, A., Vardy, A., and Dikstein, R. (2017). Efficient and Accurate Transla-
tion Initiation Directed by TISU Involves RPS3 and RPS10e Binding and Differ-
ential Eukaryotic Initiation Factor 1A Regulation. Mol. Cell. Biol. 37, e00150-17.
Heaton, N.S., Moshkina, N., Fenouil, R., Gardner, T.J., Aguirre, S., Shah, P.S.,
Zhao, N., Manganaro, L., Hultquist, J.F., Noel, J., et al. (2016). Targeting Viral
Proteostasis Limits Influenza Virus, HIV, and Dengue Virus Infection. Immunity
44, 46–58.
Hoffmann, E., Neumann, G., Kawaoka, Y., Hobom, G., and Webster, R.G.
(2000). A DNA transfection system for generation of influenza A virus from eight
plasmids. Proc. Natl. Acad. Sci. USA 97, 6108–6113.
Hogquist, K.A., Jameson, S.C., Heath, W.R., Howard, J.L., Bevan, M.J., and
Carbone, F.R. (1994). T cell receptor antagonist peptides induce positive se-
lection. Cell 76, 17–27.
Hutchinson, E.C., and Stegmann, M. (2018). Purification and Proteomics of
Influenza Virions. Methods Mol. Biol. 1836, 89–120.
Hutchinson, E.C., Curran, M.D., Read, E.K., Gog, J.R., and Digard, P. (2008).
Mutational analysis of cis-acting RNA signals in segment 7 of influenza A virus.
J. Virol. 82, 11869–11879.
Hutchinson, E.C., von Kirchbach, J.C., Gog, J.R., and Digard, P. (2010).
Genome packaging in influenza A virus. J. Gen. Virol. 91, 313–328.
Cell 181, 1502–1517, June 25, 2020 1515
 ll
OPEN ACCESS
Hutchinson, E.C., Charles, P.D., Hester, S.S., Thomas, B., Trudgian, D., Mar-
tı́nez-Alonso, M., and Fodor, E. (2014). Conserved and host-specific features
of influenza virion architecture. Nat. Commun. 5, 4816.
Jagger, B.W., Wise, H.M., Kash, J.C., Walters, K.A., Wills, N.M., Xiao, Y.L.,
Dunfee, R.L., Schwartzman, L.M., Ozinsky, A., Bell, G.L., et al. (2012). An over-
lapping protein-coding region in influenza A virus segment 3 modulates the
host response. Science 337, 199–204.
Johnstone, T.G., Bazzini, A.A., and Giraldez, A.J. (2016). Upstream ORFs are
prevalent translational repressors in vertebrates. EMBO J. 35, 706–723.
Kaessmann, H. (2010). Origins, evolution, and phenotypic impact of new
genes. Genome Res. 20, 1313–1326.
Keese, P.K., and Gibbs, A. (1992). Origins of genes: ‘‘big bang’’ or continuous
creation? Proc. Natl. Acad. Sci. USA 89, 9489–9493.
Kim, D., Langmead, B., and Salzberg, S.L. (2015). HISAT: a fast spliced aligner
with low memory requirements. Nat. Methods 12, 357–360.
Kochetov, A.V., Ahmad, S., Ivanisenko, V., Volkova, O.A., Kolchanov, N.A.,
and Sarai, A. (2008). uORFs, reinitiation and alternative translation start sites
in human mRNAs. FEBS Lett. 582, 1293–1297.
Koppstein, D., Ashour, J., and Bartel, D.P. (2015). Sequencing the cap-snatch-
ing repertoire of H1N1 influenza provides insight into the mechanism of viral
transcription initiation. Nucleic Acids Res. 43, 5052–5064.
Kovacs, E., Tompa, P., Liliom, K., and Kalmar, L. (2010). Dual coding in alter-
native reading frames correlates with intrinsic protein disorder. Proc. Natl.
Acad. Sci. USA 107, 5429–5434.
Langmead, B., Trapnell, C., Pop, M., and Salzberg, S.L. (2009). Ultrafast and
memory-efficient alignment of short DNA sequences to the human genome.
Genome Biol. 10, R25.
Leppek, K., Das, R., and Barna, M. (2018). Functional 50 UTR mRNA structures
in eukaryotic translation regulation and how to find them. Nat. Rev. Mol. Cell
Biol. 19, 158–174.
Liao, Y., Smyth, G.K., and Shi, W. (2014). featureCounts: an efficient general
purpose program for assigning sequence reads to genomic features. Bioinfor-
matics 30, 923–930.
Love, M.I., Huber, W., and Anders, S. (2014). Moderated estimation of fold
change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
Luksza, M., and Lässig, M. (2014). A predictive fitness model for influenza. Na-
ture 507, 57–61.
Machkovech, H.M., Bloom, J.D., and Subramaniam, A.R. (2019). Comprehen-
sive profiling of translation initiation in influenza virus infected cells. PLoS
Pathog. 15, e1007518.
Martin, M. (2011). Cutadapt removes adapter sequences from high-
throughput sequencing reads. EMBnet.journal 17, 10–12.
Masella, A.P., Bartram, A.K., Truszkowski, J.M., Brown, D.G., and Neufeld,
J.D. (2012). PANDAseq: paired-end assembler for illumina sequences. BMC
Bioinformatics 13, 31.
McGlincy, N.J., and Ingolia, N.T. (2017). Transcriptome-wide measurement of
translation by ribosome profiling. Methods 126, 112–129.
Ohno, S. (1970). Evolution by Gene Duplication (Springer).
Plotch, S.J., Bouloy, M., Ulmanen, I., and Krug, R.M. (1981). A unique
cap(m7GpppXm)-dependent influenza virion endonuclease cleaves capped
RNAs to generate the primers that initiate viral RNA transcription. Cell 23,
847–858.
Porgador, A., Yewdell, J.W., Deng, Y., Bennink, J.R., and Germain, R.N.
(1997). Localization, quantitation, and in situ detection of specific peptide-
MHC class I complexes using a monoclonal antibody. Immunity 6, 715–726.
Poulin, F., Brueschke, A., and Sonenberg, N. (2003). Gene fusion and overlap-
ping reading frames in the mammalian genes for 4E-BP3 and MASK. J. Biol.
Chem. 278, 52290–52297.
Price, M.N., Dehal, P.S., and Arkin, A.P. (2010). FastTree 2–approximately
maximum-likelihood trees for large alignments. PLoS ONE 5, e9490.
Reich, S., Guilligay, D., Pflug, A., Malet, H., Berger, I., Crépin, T., Hart, D., Lu-
nardi, T., Nanao, M., Ruigrok, R.W., and Cusack, S. (2014). Structural insight
1516 Cell 181, 1502–1517, June 25, 2020
Article
into cap-snatching and RNA synthesis by influenza polymerase. Nature 516,
361–366.
Rezelj, V.V., Mottram, T.J., Hughes, J., Elliott, R.M., Kohl, A., and Brennan, B.
(2019). M Segment-Based Minigenomes and Virus-Like Particle Assays as an
Approach To Assess the Potential of Tick-Borne Phlebovirus Genome Reas-
sortment. J. Virol. 93, e02068-18.
Rialdi, A., Hultquist, J., Jimenez-Morales, D., Peralta, Z., Campisi, L., Fenouil,
R., Moshkina, N., Wang, Z.Z., Laffleur, B., Kaake, R.M., et al. (2017). The RNA
Exosome Syncs IAV-RNAPII Transcription to Promote Viral Ribogenesis and
Infectivity. Cell 169, 679–692.
Rosenfeld, J., Capdevielle, J., Guillemot, J.C., and Ferrara, P. (1992). In-gel
digestion of proteins for internal sequence analysis after one- or two-dimen-
sional gel electrophoresis. Anal. Biochem. 203, 173–179.
Sabath, N., Wagner, A., and Karlin, D. (2012). Evolution of viral proteins origi-
nated de novo by overprinting. Mol. Biol. Evol. 29, 3767–3780.
Sagulenko, P., Puller, V., and Neher, R.A. (2018). TreeTime: Maximum-likeli-
hood phylodynamic analysis. Virus Evol. 4, vex042.
Schwanhäusser, B., Busse, D., Li, N., Dittmar, G., Schuchhardt, J., Wolf, J.,
Chen, W., and Selbach, M. (2011). Global quantification of mammalian gene
expression control. Nature 473, 337–342.
Sendoel, A., Dunn, J.G., Rodriguez, E.H., Naik, S., Gomez, N.C., Hurwitz, B.,
Levorse, J., Dill, B.D., Schramek, D., Molina, H., et al. (2017). Translation
from unconventional 50 start sites drives tumour initiation. Nature 541,
494–499.
Shu, Y., and McCauley, J. (2017). GISAID: Global initiative on sharing all influ-
enza data - from vision to reality. Euro Surveill. 22, 30494.
Sikora, D., Rocheleau, L., Brown, E.G., and Pelchat, M. (2017). Influenza A vi-
rus cap-snatches host RNAs based on their abundance early after infection.
Virology 509, 167–177.
Stamatakis, A. (2014). RAxML version 8: a tool for phylogenetic analysis and
post-analysis of large phylogenies. Bioinformatics 30, 1312–1313.
Starck, S.R., Tsai, J.C., Chen, K., Shodiya, M., Wang, L., Yahiro, K., Martins-
Green, M., Shastri, N., and Walter, P. (2016). Translation from the 50 untrans-
lated region shapes the integrated stress response. Science 351, aad3867.
Strelkowa, N., and Lässig, M. (2012). Clonal interference in the evolution of
influenza. Genetics 192, 671–682.
Stuller, K.A., Cush, S.S., and Flaño, E. (2010). Persistent gamma-herpesvirus
infection induces a CD4 T cell response containing functionally distinct
effector populations. J. Immunol. 184, 3850–3856.
Takahashi, H., Lassmann, T., Murata, M., and Carninci, P. (2012). 50 end-
centered expression profiling using cap-analysis gene expression and next-
generation sequencing. Nat. Protoc. 7, 542–561.
Thulasi Raman, S.N., and Zhou, Y. (2016). Networks of Host Factors that
Interact with NS1 Protein of Influenza A Virus. Front. Microbiol. 7, 654.
Tilston-Lunel, N.L., Shi, X., Elliott, R.M., and Acrani, G.O. (2017). The Potential
for Reassortment between Oropouche and Schmallenberg Orthobunyavi-
ruses. Viruses 9, 220.
Tyanova, S., Temu, T., and Cox, J. (2016). The MaxQuant computational plat-
form for mass spectrometry-based shotgun proteomics. Nat. Protoc. 11,
2301–2319.
Wallat, G.D., Huang, Q., Wang, W., Dong, H., Ly, H., Liang, Y., and Dong, C.
(2014). High-resolution structure of the N-terminal endonuclease domain of
the Lassa virus L polymerase in complex with magnesium ions. PLoS ONE
9, e87577.
Wang, X.Q., and Rothnagel, J.A. (2004). 50 -untranslated regions with multiple
upstream AUG codons can support low-level translation via leaky scanning
and reinitiation. Nucleic Acids Res. 32, 1382–1391.
Wei, J., and Yewdell, J.W. (2017). Autoimmune T cell recognition of alternative-
reading-frame-encoded peptides. Nat. Med. 23, 409–410.
Wei, J., and Yewdell, J.W. (2019). Flu DRiPs in MHC Class I Immunosurveil-
lance. Virol. Sin. 34, 162–167.

Article
Wei, J., Kishton, R.J., Angel, M., Conn, C.S., Dalla-Venezia, N., Marcel, V., Vin-
cent, A., Catez, F., Ferre, S., Ayadi, L., et al. (2019). Ribosomal Proteins Regu-
late MHC Class I Peptide Generation for Immunosurveillance. Mol. Cell 73,
1162–1173.
Wen, Y., Liu, Y., Xu, Y., Zhao, Y., Hua, R., Wang, K., Sun, M., Li, Y., Yang, S.,
Zhang, X.J., et al. (2009). Loss-of-function mutations of an inhibitory upstream
ORF in the human hairless transcript cause Marie Unna hereditary hypotricho-
sis. Nat. Genet. 41, 228–233.
Westerhof, L.M., McGuire, K., MacLellan, L., Flynn, A., Gray, J.I., Thomas, M.,
Goodyear, C.S., and MacLeod, M.K. (2019). Multifunctional cytokine produc-
tion reveals functional superiority of memory CD4 T cells. Eur. J. Immunol. 49,
2019–2029.
Wise, H.M., Gaunt, E., Ping, J., Holzer, B., Jasim, S., Lycett, S.J., Murphy, L.,
Livesey, A., Brown, R., Smith, N., et al. (2019). An alternative AUG codon that
produces an N-terminally extended form of the influenza A virus NP is a viru-
lence factor for a swine-derived virus. bioRxiv. https://doi.org/10.1101/
738427.
Ye, J., Sorrell, E.M., Cai, Y., Shao, H., Xu, K., Pena, L., Hickman, D., Song, H.,
Angel, M., Medina, R.A., et al. (2010). Variations in the hemagglutinin of the
ll
OPEN ACCESS
2009 H1N1 pandemic virus: potential for strains with altered virulence pheno-
type? PLoS Pathog. 6, e1001145.
Young, S.K., and Wek, R.C. (2016). Upstream Open Reading Frames Differen-
tially Regulate Gene-specific Translation in the Integrated Stress Response.
J. Biol. Chem. 291, 16927–16935.
Zanker, D.J., Oveissi, S., Tscharke, D.C., Duan, M., Wan, S., Zhang, X., Xiao,
K., Mifsud, N.A., Gibbs, J., Izzard, L., et al. (2019). Influenza A Virus Infection
Induces Viral and Cellular Defective Ribosomal Products Encoded by Alterna-
tive Reading Frames. J. Immunol. 202, 3370–3380.
Zhang, Y., Aevermann, B.D., Anderson, T.K., Burke, D.F., Dauphin, G., Gu, Z.,
He, S., Kumar, S., Larsen, C.N., Lee, A.J., et al. (2017). Influenza Research
Database: An integrated bioinformatics resource for influenza virus research.
Nucleic Acids Res. 45 (D1), D466–D474.
Zhou, Y., Zhou, B., Pache, L., Chang, M., Khodabakhshi, A.H., Tanaseichuk,
O., Benner, C., and Chanda, S.K. (2019). Metascape provides a biologist-ori-
ented resource for the analysis of systems-level datasets. Nat. Commun.
10, 1523.
Cell 181, 1502–1517, June 25, 2020 1517
 ll
OPEN ACCESS Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
anti-CD25-APC (PC61.5) Thermo Fisher Cat #17-0251-82; RRID: AB_469366
anti-CD8-Alexaflor488 (53-6.7) Thermo Fisher Cat #53-0081-82; RRID: AB_469897
anti-Va2-E450 (B20.1) Thermo Fisher Cat #48-5812-82; RRID: AB_10804752
anti-Vb5-PE (MR9-4) BD Biosciences Cat # 553190; RRID: AB_394698
anti-CD44-PerCp-Cyanine5.5 (IM7) Thermo Fisher Cat #45-0441-82; RRID: AB_925746
anti-CD69-PE-Cy7 (H1.2F3) Thermo Fisher Cat #25-0691-82; RRID: AB_469637
Anti-NP antibody BioRad Cat # MCA400; RRID: AB_2151884
m-IgGk BP-HRP Santa Cruz Cat # sc-516102; RRID: AB_2687626
Bacterial and Virus Strains
A/Puerto Rico/8/34 (H1N1) (PR8) de Wit et al., 2004 N/A
PR8; PB1-UFOD This study N/A
PR8; PB1-UFOSYN This study N/A
PR8; PB1-EXT+ This study N/A
PR8; NP-EXTD This study N/A
PR8; NP-EXTSYN This study N/A
A/California/04/09(H1N1) (Cal09) Ye et al., 2010 N/A
Cal09; PB1-UFOD This study N/A
Cal09; PB1-UFOSYN This study N/A
Cal09; PB2-UFOD This study N/A
Cal09; PB2-UFOSYN This study NA
Cal09; PA-UFOD This study NA
SYN
Cal09; PB1-UFO This study NA
Cal09; HA-UFOD This study NA
Cal09; HA-UFOSYN This study NA
A/WSN/33(H1N1) (WSN) Hoffmann et al., 2000 N/A
WSN; PB1-UFOD This study N/A
WSN; PB1-UFOSYN This study N/A
WSN; PB2-UFOD This study N/A
WSN; PB2-UFOSYN This study N/A
WSN; PA-UFOD This study N/A
WSN; PA-UFOSYN This study N/A
WSN; HA-UFOD This study N/A
WSN; HA-UFOSYN This study N/A
PB1-UFO(SIIN) This study N/A
NS-UFO(SIIN) This study N/A
PB1-SIIN Wei et al., 2019 N/A
NA-SIIN MRC-University of Glasgow Centre for Virus N/A
Research; As Bottermann et al., 2018
A/Udorn/72(H3N2) The Roslin Institute, University of N/A
Edinburgh; As Clohisey et al., 2020
LASV (Josiah strain) Department of Pathology, the University of N/A
Texas Medical Branch
B/Wisconsin/01/2010 Department of Microbiology, Icahn School N/A
of Medicine at Mount Sinai
(Continued on next page)

e1 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Biological Samples
Primary CD14+ human monocytes The Roslin Institute, University of N/A
Edinburgh; As Clohisey et al., 2020
Chemicals, Peptides, and Recombinant Proteins
Dulbecco’s Modified Eagle Thermo Fisher / GIBCO Cat#11965175
Medium (DMEM)
Minimum Essential Medium (MEM) Sigma-Aldrich Cat# 51411C
Purified Agar Oxoid Cat #: LP0028
Trypsin from bovine pancreas, TPCK- Sigma-Aldrich Cat #: T1426-500MG
treated
Protease Inhibitor Cocktail Set III, EDTA- EMD Millipore Cat# 539134-10ML
Free - Calbiochem
Trypsin Sigma-Aldrich Cat# T8802-100MG
TRIzol Reagent Thermo Fisher Scientific Cat#15596018
SimplyBlueTM SafeStain Thermo Fisher Scientific Cat# LC6060
NuPage 412% BT Gel 1.5mm 12w 10 Thermo Fisher Scientific Cat# NP0322BOX
Per Box
MG-132 Sigma-Aldrich Cat# M7449-1ML
NuPAGE MOPS SDS Running Buffer (20X) Thermo Fisher Scientific Cat# NP0001
Ovalbumin (257-264) chicken Sigma-Aldrich Cat# S7951
LT-1 transfection reagent Mirius Cat# MIR 2304
recombinant human colony-stimulating A gift from Chiron, Emeryville, CA, US; As N/A
factor 1 Clohisey et al., 2020
Lys-C lysyl endopeptidase Wako 121-05063
Harringtonine LKT biochemicals H0169
Cycloheximide Sigma-Aldrich Cat# C7698
Sequencing grade modified trypsin Promega 9PIV511
Critical Commercial Assays
Dual Luciferase Reporter Assay System Promega Cat#E1910
CD8a+ T Cell Isolation Kit Miltenyi Biotec Cat#130-104-075
EasySep Mouse CD8+ T Cell Isolation Kit StemCell Technologies Cat# 19853
PureLink RNA Mini Kit 250 Reactions Thermo Fisher Scientific Cat# 12183025
PureLink DNase Set Thermo Fisher Scientific Cat# 12185010
miRNeasy Mini Kit QIAGEN Cat# 217004
Q5 site directed mutagenesis kit NEB Cat# E0554S
Ribo-Zero Gold rRNA Removal Kit Illumina Cat# MRZG12324
(Human/Mouse/Rat)
SMARTer total RNA Pico kit Clontech Cat# 634411
TruSeq Stranded Total RNA Library Prep Kit Illumina Cat # 20020596
Deposited Data
CAGE sequencing of WSN IAV virus Clohisey et al., 2020 https://fantom.gsc.riken.jp/5/data/
infected cells
DEFEND seq of PR8 IAV infected A549 cells Rialdi et al., 2017 GEO: GSE96677
DEFEND seq of IBV infected A549 cells This study GEO: GSE85474
Ribosome Profiling of PR8 IAV infected cells This study GEO: GSE148245
CAGE sequencing of LASV infected This study GEO: GSE148122
vero cells
RNA seq of PR8; PB1-UFOD and PR8;PB1- This study GEO: GSE128519
UFOSYN infected mouse lungs
(Continued on next page)

Cell 181, 1502–1517.e1–e11, June 25, 2020 e2

 ll
OPEN ACCESS Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
GISAID Database Shu and McCauley, 2017 https://www.gisaid.org
NCBI Influenza Virus Database Zhang et al., 2017 http://www.ncbi.nlm.nih.gov/genomes/
FLU/Database
Mass spectrometry Data: PR8 IAV infected This study Table S2A
A549 and 293T cells
Mass spectrometry Data: WSN IAV Virions Hutchinson et al., 2014 https://massive.ucsd.edu/ProteoSAFe/
datasets.jsp using the MassIVE ID
MSV000078740; Table S2B
Mass spectrometry Data: Heaton et al., 2016 Table S2C
Immunoprecipitation of PR8 IAV RdRp
Experimental Models: Cell Lines
Dog: MDCK ATCC CCL-34; RRID: CVCL_0422
Human: A549 ATCC CCL-185; RRID: CVCL_0023
Human: 293T ATCC CRL-3216; RRID: CVCL_0063
Cow: MDBK Sigma 90050801-1VL; RRID: CVCL_0421
Monkey: Vero ATCC CCL-81; RRID: CVCL_0059
Mouse: DC2.4 Sigma-Aldrich Cat# SCC142; RRID: CVCL_J409
Hamster: BSR-T7/5 Buchholz et al., 1999 N/A
Experimental Models: Organisms/Strains
Mouse: BALB/cJ (6-8 weeks) Jackson Laboratories 00651
Chicken: Specific Pathogen Free Charles River Cat #: 10100329
Fertile Eggs
Mouse: OT-I: C57BL/6-Tg(TcraTcrb) The Jackson Laboratory / in-house; Cat# 003831; RRID: IMSR_JAX:003831
1100Mjb/J Hogquist et al., 1994
Mouse: C57BL/6 (10-14 weeks) Envigo N/A
Oligonucleotides
DEFEND-seq cDNA synthesis–3’ primer Rialdi et al., 2017 N/A
qPCR Primers This Study Table S4B
Recombinant DNA
PR8 pDUAL plasmids A kind gift of Prof Ron Fouchier; de Wit N/A
et al., 2004
Cal09 pDP2002 plasmids A kind gift of Prof Daniel Perez.; Ye N/A
et al., 2010
pT7HRTMRen(-) MRC-University of Glasgow Centre for Virus N/A
Research; Rezelj et al., 2019
pTMHRTN MRC-University of Glasgow Centre for Virus N/A
Research; Rezelj et al., 2019
pTMHRTL MRC-University of Glasgow Centre for Virus N/A
Research; Rezelj et al., 2019
pTM1-FFLuc MRC-University of Glasgow Centre for Virus N/A
Research; Rezelj et al., 2019
pRL-TK Promega E2241
Software and Algorithms
DESeq2 Love et al., 2014 https://bioconductor.org/packages/
release/bioc/html/DESeq2.html
Bowtie Langmead et al., 2009 http://bowtie-bio.sourceforge.net/
index.shtml
MaxQuant Cox and Mann, 2008 https://www.biochem.mpg.de/5111795/
maxquant
Cutadapt Martin, 2011 https://cutadapt.readthedocs.io/en/stable/
(Continued on next page)

e3 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
STAR Dobin et al., 2013 https://github.com/alexdobin/STAR
FlowJo Treestar N/A
Metascape Zhou et al., 2019 https://metascape.org/gp/index.html#/
main/step1
Vienna RNA Webserver Gruber et al., 2008 http://rna.tbi.univie.ac.at
FastTree Price et al., 2010 http://www.microbesonline.org/fasttree
RAxML Stamatakis, 2014 https://cme.h-its.org/exelixis/web/
software/raxml/index.html
TreeTime Sagulenko et al., 2018 https://github.com/neherlab/treetime
PANDASeq Masella et al., 2012 https://github.com/neufeld/pandaseq
NetMHC (v3.4 and v4.0) Andreatta and Nielsen, 2016 https://services.healthtech.dtu.dk/service.
php?NetMHC-4.0
MUSCLE Edgar, 2004 https://www.drive5.com/muscle/
HISAT2 Kim et al., 2015 http://daehwankimlab.github.io/hisat2
Prism 8 Graphpad N/A

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for reagents may be directed to and will be fulfilled by Lead Contact Ivan Marazzi (ivan.marazzi@
mssm.edu).

Materials Availability
All unique/stable reagents generated in this study are available from the Lead Contact with a completed Materials Transfer
Agreement.

Data and Code Availability

The datasets for CAGE sequencing of A/Udorn/72 (H3N2) IAV virus infected cells are reported in Clohisey et al. (2020) deposited in
https://fantom.gsc.riken.jp/5/data/. Datasets for DEFEND-seq of PR8-IAV infected A549 cells were taken from a pre-existing dataset
[GEO: GSE96677] (Rialdi et al., 2017). DEFEND-seq of IBV infected cells were generated in this study and deposited in GEO:
GSE85474. Ribosome profiling profile of PR8 IAV infected cells were generated in this study and deposited in GEO: GSE148245.
The datasets for CAGE sequencing of LASV infected Vero cells were generated in this study and deposited in GEO: GSE148122.
RNA seq of PR8; PB1-UFOD and PR8;PB1-UFOSYN infected mouse lungs was generated in this study and deposited in
GSE128519. Mass spectrometry data for PR8 infected IAV infected A549 and 293 cells was generated in this study and presented
in Table S2A. Mass spectrometry of WSN IAV virions was analyzed from datasets generated in Hutchinson et al. (2014), and taken
from https://massive.ucsd.edu/ProteoSAFe/datasets.jsp using the MassIVE ID MSV000078740. Tables are also found in Table S2B.
Mass spectrometry data for PB1-UFO interactions with IAV polymerase subunits was analyzed using datasets from Heaton et al.
(2016) and presented in Table S2C.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cells cultures
Madin–Darby Canine Kidney (MDCK) cells, A549 human lung epithelial cells, Vero (ATCC-CCL81) and 293T human embryonic kidney
cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM; GIBCO) supplemented with 10% fetal bovine serum (FBS;
GIBCO). Madin-Darby Bovine Kidney (MDBK) cells were cultured in Minimum Essential Medium (MEM; Sigma) supplemented
with 2 mM L-glutamine and 10% fetal calf serum (FCS). BSR-T7/5 golden hamster cells (Buchholz et al., 1999) were cultured in Glas-
gow Minimal Essential Medium (GMEM) supplemented with 10% FCS and 10% tryptose phosphate broth under G418 selection. All
cells were maintained at 37C and 5% CO2.

Mice
For infection studies: Six to eight-week-old female BALB/c mice were obtained from Jackson Laboratories (Bar Harbor, ME). All mice
infection procedures were performed following protocols approved by the Icahn School of Medicine at Mount Sinai Institutional

Cell 181, 1502–1517.e1–e11, June 25, 2020 e4

 ll
OPEN ACCESS Article

Animal Care and Use Committee (IACUC). Animal studies were carried out in strict accordance with the recommendations in the
Guide for the Care and Use of Laboratory Animals of the National Research Council.
For antigen presentation experiments: female OTI (Hogquist et al., 1994) mice were bred in-house on a mixed genetic background.
Animals were kept in dedicated barrier facilities, proactive in environmental enrichment under the EU Directive 2010 and Animal (Sci-
entific Procedures) Act (UK Home Office license number 70/8645) with ethical review approval (University of Glasgow). Animals were
cared for by trained and licensed individuals and humanely sacrificed using Schedule 1 methods.
For BMDC Isolation: 10-14 week old naive female C57BL/6 mice, purchased from Envigo (UK) and maintained at the University of
Glasgow under standard animal husbandry conditions in accordance with UK home office regulations and approved by the local
ethics committee.

Virus Strains
Wild-type viruses
A/Puerto Rico/8/34(H1N1) (PR8) virus was generated by reverse genetics and propagated in 9-11 day old embryonated chicken eggs
(Charles River, Cat # 10100329). Mouse-adapted A/California/04/09(H1N1) (Cal09) was generated by reverse genetics (Ye et al.,
2010) and propagated on MDCK cells in the presence of 1 mg/ml TPCK-trypsin, as described previously (Hutchinson et al., 2008).
The influenza virus A/WSN/33(H1N1) (WSN) (Hoffmann et al., 2000) was propagated on MDBK cells. A/Udorn/72(H3N2) (Udorn)
was propagated on MDCK cells in the presence of 1 mg/ml TPCK-trypsin, as described previously (Clohisey et al., 2020; Hutchinson
et al., 2014). Plaque assays were carried out in MDCK cells and visualized by immunocytochemistry or staining with crystal violet or
Coomassie blue, as previously described (Gaush and Smith, 1968) (See below also for method details).
Mutant viruses
All mutant and control viruses were generated using a plasmid-based reverse genetics system (Fodor et al., 1999; Ye et al., 2010),
using either the A/Puerto Rico/8/1934 (PR8), A/WSN/33 (WSN) or mouse-adapted A/California/4/09 (Cal09) strains as the backbone.
Plasmids used for reverse genetics were the PR8 pDUAL plasmids (de Wit et al., 2004) and the Cal09 pDP2002 plasmids (Ye et al.,
2010) (a kind gift of Prof Daniel R. Perez (University of Georgia, USA). Site-directed mutagenesis of plasmids was performed using the
Q5 site-directed mutagenesis kit (QIAGEN); the edited NS segment sequence required for the PR8-NS.F3.SIIN mutant virus
(described in Figure 4) was synthesized by Genewiz.
PB1-UFO(SIIN) virus
OVA257-264 (SIINFEKL) epitope was inserted into the 50 UTR of the PB1 segment of the influenza A virus (IAV) genome at position 1
before the PB1 start codon. This insertion did not result in an N-terminal extension of or mutations in the PB1 protein, but results in the
insertion of the OVA257-264 antigenic epitope in frame with the PB1-UFO protein.
NS-UFO(SIIN) virus
The OVA257-264 (SIINFEKL) epitope was inserted into frame 2 of the NS segment of the IAV genome, in a region corresponding to the
linker sequence of the NS1 protein (encoded in frame). This effectively replaced codons 79-84 of NS1, while retaining the sequence of
NEP. The replacement sequence was flanked by two upstream nucleotides and one downstream nucleotide to introduce a frameshift
into frame 2. Premature stop codons in frame 2 were also mutated at positions 4, 27, 32, 74 and 77, relative from the start codon of
NS1, to generate a 106 amino acid long NS-UFO sequence, extending it from the original 4 amino acid long uvORF in reading frame 2.
PB1-SIIN virus and NA-SIIN viruses
These viruses have been described in Wei et al. (2019) and Bottermann et al. (2018) respectively.
PR8; PB1-UFOD, PR8; PB1-UFOSYN, PR8; PB1-EXT+ viruses
PR8; PB1-UFOD contains a C to T nucleotide substitution 9 nucleotides after the start of PB1 open reading frame. This generates a
premature stop codon in the PB1-UFO ORF. Its control virus, PR8; PB1-UFOSYN, contains a C to G nucleotide substitution at the
same position. Both viruses retain the amino acid sequence of the PB1 ORF. PR8; PB1-EXT+ contains a T to C nucleotide substi-
tution three nucleotides before the start of PB1 open reading frame. This disrupts a conserved stop codon (‘‘TGA’’) in frame with
PB1 ORF, resulting in the N-terminal extension of the PB1-ORF. PB1-UFO ORF is maintained in this virus. Mutations were confirmed
by sequencing both plasmids and viruses. All viruses were expanded in 9-11 day old embryonated chicken eggs after rescue. The
stock virus titers were calculated from the average of three independent experiments.
PR8; NP-EXTD, PR8; NP-EXTSYN viruses
PR8; NP-EXTD contains an A to T nucleotide substitution 6 nucleotides before the start of the NP open reading frame. This generates
an in-frame stop codon that results in the loss of the N-terminal NP-extension. Its control virus, PR8; NP-EXTSYN, bears an A to G
nucleotide substitution at the same position in the UTR, preserving the NP-extension. Mutations were confirmed by sequencing
both plasmids and viruses, and 3 independent plaque purified clones of each virus, grown on MDCK cells, were used in subsequent
experiments. Stock virus titers were calculated from the average of three independent experiments.
WSN; PB1-UFOD, WSN; PB1-UFOSYN, Cal09; PB1-UFOD, Cal09; PB1-UFOSYN viruses
WSN; PB1-UFOD and Cal09; PB1-UFOD viruses contain C to U nucleotide substitutions 9 nucleotides after the start of PB1 open
reading frame. This generates a premature stop codon in the PB1-UFO ORF. Their control viruses, WSN; PB1-UFOSYN and
Cal09; PB1-UFOSYN respectively, contain C to G nucleotide substitutions at the same positions. All the viruses retain the amino
acid sequence of the PB1 ORF.

e5 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

WSN; PB2-UFOD, WSN; PB2-UFOSYN, Cal09; PB2-UFOD, Cal09; PB2-UFOSYN viruses

WSN; PB2-UFOD and Cal09; PB2-UFOD viruses contain A to T nucleotide substitutions 12 nucleotides after the start of PB2 open
reading frame. This generates a premature stop codon in the PB2-UFO ORF. Their control viruses, WSN; PB2-UFOSYN and
Cal09; PB2-UFOSYN respectively, contain a A to C nucleotide substitutions at the same position. All the viruses retain the amino
acid sequence of the PB2 ORF.
WSN; PA-UFOD, WSN; PA-UFOSYN, Cal09; PA-UFOD, Cal09; PA-UFOSYN viruses
WSN; PA-UFOD and Cal09; PA-UFOD viruses contain C to T nucleotide substitutions 42 nucleotides after the start of PA open reading
frame. This generates a premature stop codon in the PA-UFO ORF. Their control viruses, WSN; PA-UFOSYN and Cal09; PA-UFOSYN
respectively, contain C to A nucleotide substitutions at the same position. All the viruses retain the amino acid sequence of the
PA ORF.
WSN; HA-UFOD, WSN; HA-UFOSYN viruses
WSN; HA-UFOD viruses contain A to T nucleotide substitutions 45 nucleotides after the start of HA open reading frame. This gener-
ates a premature stop codon in the HA-UFO ORF. Their control viruses, WSN; HA-UFOSYN and Cal09; PA-UFOSYN respectively,
contain A to C nucleotide substitutions at the same position. All the viruses retain the amino acid sequence of the HA ORF.
Cal09; HA-UFOD, Cal09; HA-UFOSYN viruses
Cal09; HA-UFOD viruses contain G to T nucleotide substitutions 52 nucleotides after the start of HA open reading frame. This gen-
erates a premature stop codon in the HA-UFO ORF. Its control virus, Cal09; HA-UFOSYN contains an G to C nucleotide substitution at
the same position. All the viruses retain the amino acid sequence of the HA ORF.
Primary CD14+ human monocytes
Primary CD14+ human monocytes were isolated from whole blood samples under ethical approval from Lothian Research Ethics
Committee (11/AL/0168). Cells were obtained from blood donated by 4 anonymous healthy volunteers. Volunteers were not treated
with any drugs. Some volunteers have donated blood used in multiple experiments outside this study. Health status is not assessed.
Plasmids
Plasmids used for HRTV minireplicon assays were the Renilla-luciferase-encoding pT7HRTMRen(–); the viral-gene-encoding
pTMHRTN and pTMHRTL and the firefly-luciferase-encoding control plasmid pTM1-FFluc (Rezelj et al., 2019).

METHOD DETAILS

Growth kinetics of Viruses in Cell Culture

A549 or MDCK cells were infected with the indicated viruses at a multiplicity of infection (MOI) of 0.001 and incubated for one hour at
37 C. Infected cells were washed twice, and then cultured with Opti-MEM and TPCK-treated trypsin at 37 C for 72 h. Supernatants
were collected at the indicated time points. Viral titers were determined by plaque assays.

Quantification of IAV titers by Plaque Assays

Plaque assay in MDCK cells were performed as described previously (Gaush and Smith, 1968). Briefly, serially diluted culture super-
natants of infected cells were adsorbed on layers of confluent MDCK cells for 1 hour. Infected cells were then overlaid with 2ml of
DMEM, 25mM HEPES, 2mM glutamine, 100ug/ml penicillin-streptomycin, 1ug/ml TPCK-trypsin and 0.8% Oxoid Agar. Plates
were incubated for 48-72h until plaques were observed. Plaque were then fixed in 4% formaldehyde and visualized through staining
with 1% crystal violet solution. Alternatively, MDCK cells were overlaid with DMEM mixed 1:1 with 2% (w/v) low gelling temperature
agarose in PBS and supplemented with 1ug/ml TPCK-trypsin, incubated for 48-72h until plaques were observed, and then either
fixed and stained directly (with 0.2% (w/v) Coomassie Brilliant Blue R in 7.5% (v/v) acetic acid and 50% (v/v) ethanol) or fixed in
80% chilled acetone and visualized by immunocytochemistry (permeabilized in 1% Triton X-100 in PBS, blocked in 10% FBS in
PBS, immunostained with mouse anti-NP (BioRad: Cat# MCA400) and peroxidase-conjugated rabbit anti-mouse IgG (Santa
Cruz; Cat # sc-516102) and visualized with True Blue Peroxidase).

Ribosome profiling and analysis

A549 cells were infected in a 10cm dish with A/Puerto Rico/8/1934 (H1N1, PR8) at a MOI of 3. At 8h post infection, ribosome profiling
libraries were prepared as previously described (McGlincy and Ingolia, 2017) with the following exceptions. Infected cells were
treated with either DMSO or 5mg/mL harringtonine for 15 minutes. Cell lysis was performed by flash freezing in liquid nitrogen prior
to the addition of ice-cold lysis buffer. rRNA removal was performed as previously described (Wei et al., 2019). Sequencing was per-
formed two lanes of a HiSeq using a 2x150 bp configuration.

Mass Spectrometry experiments (in infected cell lysates)

A549 or HEK293T cells were infected with PR8 virus stock at multiplicities of infection of 3 and 5 respectively. At 8h or 24h post infec-
tion, cells were scraped, washed twice in PBS with protease inhibitors (Calbiochem), before being snap-frozen in liquid nitrogen.
Where indicated, MG132 was added to the cell culture media 4h prior to sample collection. Mock infected samples were included
as negative controls. To prepare cell lysates for mass spectrometry, cell pellets were lysed in lysis buffer (50mM Tris pH8, 1% NP-40,
100mM NaCl, protease inhibitors) on ice. NaCl concentration was then brought up to 500mM by adding salt drop-wise into the

Cell 181, 1502–1517.e1–e11, June 25, 2020 e6

 ll
OPEN ACCESS Article

solution while agitating. Lysates were rotated for 30min at 4 C before an equal volume of water was added to the sample to bring
NaCl concentration back to 250mM. Samples were then centrifuged at full speed for 15 min at 4 C. 4x Laemmli buffer (200mM
Tris-HCl pH6.8, 8% SDS, 40% glycerol, 0.588M B-mercaptoethanol, 50mM EDTA and 0.08% Bromophenol Blue) was then added
to the supernatant to 1x concentration, and 5ml of the lysate was loaded on a 4%–12% Bis-Tris gel (Novex). Gels were run under a
hood for 150V for 1h15min in 1X MOPS running buffer and stained in SimplyBlueTM SafeStain (Invitrogen), following the manufac-
turer’s recommended protocol. Once stained, gel bands corresponding to 40-60kDa and < 15kDa were excised. Gel slices were sub-
ject to in-gel tryptic digests as previously described (Rosenfeld et al., 1992).
Digested samples were analyzed on a Thermo Fisher Orbitrap Fusion mass spectrometry system equipped with an Easy nLC 1200
ultra-high pressure liquid chromatography system interfaced via a Nanospray Flex nanoelectrospray source. Samples were injected
on a C18 reverse phase column (25 cm 3 75 mm packed with ReprosilPur C18 AQ 1.9 mm particles). Peptides were separated by an
organic gradient from 5% to 30% ACN in 0.1% formic acid over 70 minutes at a flow rate of 300 nL/min. The MS continuously ac-
quired spectra in a data-dependent manner throughout the gradient, acquiring a full scan in the Orbitrap (at 120,000 resolution with
an AGC target of 200,000 and a maximum injection time of 100 ms) followed by as many MS/MS scans as could be acquired on the
most abundant ions in 3 s in the dual linear ion trap (rapid scan type with an intensity threshold of 5000, HCD collision energy of 29%,
AGC target of 10,000, a maximum injection time of 35 ms, and an isolation width of 1.6 m/z). Singly and unassigned charge states
were rejected. Dynamic exclusion was enabled with a repeat count of 1, an exclusion duration of 20 s, and an exclusion mass width of
± 10 ppm. Raw mass spectrometry data were assigned to human protein sequences and MS1 intensities extracted with the Max-
Quant software package (version 1.6.8) (Cox and Mann, 2008). Data were searched against the SwissProt human protein database
(downloaded on October 10, 2019) and a custom influenza A virus database comprising all six open-reading frames greater than 10
amino acids for the IAV (strain PR-8) genomic sequence. Variable modifications were allowed for N-terminal protein acetylation,
methionine oxidation, and lysine acetylation. A static modification was indicated for carbamidomethyl cysteine. All other settings
were left using MaxQuant default settings.

Mass Spectrometry experiments (in virions)

The purification of influenza virions and collection of mass spectra by LC-MS/MS has been described previously (Hutchinson et al.,
2014), and followed previously-described protocols for purification, mass spectrometry and data analysis (Hutchinson and Stegmann,
2018). Briefly, the IAV WSN was propagated on MDBK cells. Six viral stocks were prepared, of which half were subjected to haemad-
sorption on chicken red blood cells to stringently remove non-viral material. Virus particles were then purified by sucrose gradient ul-
tracentrifugation, lysed in urea, reduced, alkylated and digested with trypsin and LysC. Tryptic peptides were analyzed by liquid chro-
matography and tandem mass spectrometry (LC-MS/MS) using an Ultimate 3000 RSLCnano HPLC system (Dionex, Camberley, UK)
run in direct injection mode and coupled to a Q Exactive mass spectrometer (Thermo Electron, Hemel Hempstead, UK) in ‘Top 10’ data-
dependent acquisition mode. Raw files describing these mass spectra have been deposited at the Mass spectrometry Interactive Vir-
tual Environment (MassIVE; Center for Computational Mass Spectrometry at University of California, San Diego) and can be accessed at
https://massive.ucsd.edu/ProteoSAFe/datasets.jsp using the MassIVE ID MSV000078740. For the purposes of this project, data were
re-analyzed using MaxQuant 1.5.8.3 analysis software (Tyanova et al., 2016) using standard settings and the following parameters: la-
bel-free quantitation and the iBAQ algorithm (Schwanhäusser et al., 2011) enabled; enzyme: trypsin/P; variable modifications: oxidation
(M) and acetyl (Protein N-ter); and fixed modifications: carbamidomethyl (C); digestion mode: semi-specific free N terminus. Peptide
spectra were matched to custom databases containing the IAV WSN proteome (including full-length translations of all six reading
frames), an edited version of the Bos taurus proteome (UP000009136; retrieved from UniProt on 16/05/2017) in which all instances
of the ubiquitin sequence had been deleted, and a single repeat of the ubiquitin protein sequence.

DEFEND sequencing of IBV infected cells

DEFEND-seq was performed as previously described (Rialdi et al., 2017). Briefly, RNA was extracted from A549 cells infected with
influenza B virus (B/Wisconsin/01/2010) for 8 hours using Trizol (Invitrogen) and subjected to DNase treatment (QIAGEN). 5mg of
DNase treated RNA was then incubated with 10U of Tobacco Acid Phosphotase (Epicentre; 37 C, 1.5h) to remove mRNA 50 caps.
Sodium periodate was then added (to 500mM) into the reaction to block the 30 OH. The reaction was then allowed to proceed for
1.5h at 4 C, before being blocked by the addition of 1/10 volume 1M L-lysine, and incubating for an additional 10min at room tem-
perature. RNA was purified with 1.8X AMPure XP beads (Beckman Coulter). Barcoded with RNA adapters were then ligated to the
50 ends of RNAs overnight at 16 C. Adaptor-ligated RNA was purified using 1.8X volume of AMPure XP beads. Ribosomal RNAs were
removed using the Ribo-Zero Gold rRNA Removal Kit (Human/Mouse/Rat) (Illumina), according to the manufacturer’s protocol.
cDNA synthesis was performed using a custom 30 primer ((50 -AGA CGT GTG CTC TTC CGA TCT N*N*N*N*N*N*-30 , Bioo Scientific,
N* = randomized bases) for 2 min at 65 C. Illumina adapters were added by PCR, and products were size-selected (200-400bp) using
BluePippin 2% M1 gels (Sage Scientific). The library was validated on the Agilent Bioanalyzer, and samples were sequenced on the
Illumina HiSeq 2500 platform in a 100bp SE read run format.

Preparation of CAGE libraries from LASV infected cells

Vero cells (ATCC-CCL81) grown on T75 flask were infected with recombinant LASV (Josiah strain) at MOI 0.1. At 2 days post infection,
cells were lysed in Trizol (Invitrogen). The infection work with pathogenic Lassa virus and RNA lysate preparation were performed at the

e7 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

BSL4 facilities in Galveston National Laboratory in the University of Texas Medical Branch in accordance with institutional health and
safety guidelines and federal regulations. Total RNA from the trizol-treated lysates was isolated and DNase treated using the Purelink
RNA Minikit (Invitrogen). The purified RNA was then submitted for CAGE-sequencing at Kabushiki Kaisha DNAFORM, Japan.

Mouse Infection studies

All mice infection procedures were performed following protocols approved by the Icahn School of Medicine at Mount Sinai Institu-
tional Animal Care and Use Committee (IACUC). Animal studies were carried out in strict accordance with the recommendations in
the Guide for the Care and Use of Laboratory Animals of the National Research Council. Six to eight-week-old female BALB/c mice
were obtained from Jackson Laboratories (Bar Harbor, ME). Mice were anesthetized by intraperitoneal injection of a mixture of 85mg/
kg ketamine and 12.5mg/kg xylazine before infection before being inoculated intranasally with 50ml virus re-suspended in PBS. Mice
were monitored daily for clinical signs of illness and weight loss after infection. Upon reaching 75% of initial body weight, animals
were humanely euthanized with carbon dioxide (CO2) as per the IACUC protocol.

Preparation of RNA sequencing Libraries (Infected Mice)

3 mice were intranasally (i.n.) infected with 100 plaque-forming units (PFU) of viruses in a volume of 50 mL and euthanized at 6 days post-
inoculation (d.p.i.). The middle lobe of the lung was collected for total RNA extraction, and the post-caval lobes of the lung was collected
to determine virus titers by plaque assay on MDCK cells. Lung tissue was then homogenized in Trizol (Invitrogen), and RNA was ex-
tracted as per manufacturer’s guidelines. Libraries were constructed using the Illumina TruSeq Stranded Total RNA Library Prep Kit.

SIINFEKL expression analysis

For T cell activation assays with PB1-UFO(SIIN) and PB1-SIIN viruses, OT-I T cells were harvested from the spleen and lymph nodes
of OTI transgenic mice and purified on the AutoMACS with the CD8a+ T Cell Isolation Kit (Miltenyi, Germany). DC2.4 cells were in-
fected with influenza A viruses for 18 hours, and then co-cultured with OTI T cells. T cells were stained with anti-CD25 and anti-CD28
labeled antibodies at 24 hours post co-culture for activation assays. T cell proliferation assays were conducted at 48 hours post infec-
tion by measuring CellTrace Violet staining by flow cytometry.
For T cell activation assays with the NS-UFO(SIIN) and NA-SIIN viruses, IAV antigen was propagated by infecting MDCK cells with
IAV PR8 wild-type, PR8 containing an NS segment with SIINFEKL inserted into frame 3 (PR8-NS.F3.SIIN) or PR8 containing an NA
segment with SIINFEKL inserted into frame 1 (PR8-NA.SIIN) (Bottermann et al., 2018). The IAV antigen preparations were prepared as
described (Stuller et al., 2010; Westerhof et al., 2019). Briefly, MDCK cells were infected for 48 h with each IAV stain and then centri-
fuged, resuspended in 0.1 M glycine buffer containing 0.9% NaCl (pH 9.75), and shaken at 4 C for 20 min. Preparations were
sonicated 4 times at 10 s intervals before centrifugation, and the supernatant stored at 80 C.
Bone marrow was then taken from 10-14 week old naive female C57BL/6 mice, purchased from Envigo (UK) and maintained at the
University of Glasgow under standard animal husbandry conditions in accordance with UK home office regulations and approved by the
local ethics committee. Bone marrow derived dendritic cells (BMDCs) were prepared as previously described (Westerhof et al., 2019).
Briefly, the tibias and femurs were flushed to obtain bone marrow cells. Red blood cells were lysed. Cells were then cultured in RPMI
with 10% FCS, 100ug/ml penicillin-streptomycin and 2mM L-glutamine, in the presences of GM-CSF (prepared from X-63 supernatant),
for 7 days, with media supplemented on day 2 and replaced on day 5. DCs were then harvested and incubated overnight with IAV an-
tigen preparations. Control BMDCs were incubated with SIINFEKL peptide (Ovalbumin (257-264), chicken, Sigma-Aldrich) for 1 h
at 37 C.
Lymph nodes (LN) (inguinal, brachial, axillary and cervical) and spleen were obtained from OTI mice sacrificed at weeks 12-13. CD8
T cells were negatively selected from LN and spleen using EasySep Mouse CD8+ T Cell Isolation Kit (Stemcell technologies).
BMDCs that had been exposed to viral antigen were co-cultured with CD8+ OTI T cells for 24 h. Activated T cells were detected by
immunostaining with antibodies against Va2-E450 (Thermo Fisher), Vb5-PE (M59-4 BD Biosciences), CD8-Alexaflor488 (53-6.7
Thermo Fisher), CD25-APC (PC61.3 Thermo Fisher), CD44-PerCpC5.5 (IM7 Thermo Fisher), and CD69-PerCy7 (H1.2F3 Thermo
Fisher). Data were acquired with a BD Fortessa cell analyzer and analyzed by FlowJo (BD, version 10).

Minireplicon Assays
Minireplicon assays were performed as previously described (Rezelj et al., 2019; Tilston-Lunel et al., 2017). Briefly, and using the
plasmids indicated above, LT-1 transfection reagent (Mirus) was used to transfect sub-confluent BSR-T7/5 cells. After 24 h cells
were processed using a Dual-Luciferase Reporter Assay System (Promega), with luciferase measured using Glowmax 20/20 lumin-
ometer (Promega).

QUANTIFICATION AND STATISTICAL ANALYSES

Mouse Infection Studies

Statistical significance between survival curves were compared using Log-rank (Mantel-Cox) test using Graphpad Prism 8.0 soft-
ware. Two tailed Student’s t tests under the assumption of equal variances between groups were used to compare weight loss in
mice from different groups for each day post infection. Data are shown as +/- SEM.

Cell 181, 1502–1517.e1–e11, June 25, 2020 e8

 ll
OPEN ACCESS Article

Quantitative qPCR assays

qPCR assays were done with 4 biological replicates (4 infected mice/condition). Statistical significance in gene expression was
calculated with Graphpad Prism 8.0 software, and determined using one-tailed Student’s t test under the assumption of equal var-
iances between groups. Data are shown as mean +/- SEM.

CAGE sequencing of WSN IAV virus infected cells

The sequencing of cap-snatched leader sequences was described in detail in a recent publication (Clohisey et al., 2020). Briefly,
primary CD14+ human monocytes were isolated from 4 volunteer donors under ethical approval from Lothian Research Ethics Com-
mittee (11/AL/0168) and cultured in the presence of 100 ng/ml (104 U/ml) recombinant human colony-stimulating factor 1 (a gift from
Chiron, USA) for 8 days to differentiate them into macrophages. Monocyte-derived macrophages were then infected with influenza
(Udorn) at an MOI of 5, harvested at 0, 2, 7 and 24 hours post-infection (times defined as starting after a 1h adsorption step), and
processed for RNA extraction using a miRNeasy Mini Kit (QIAGEN). Cap analysis of gene expression (CAGE) was performed as
part of the FANTOM5 project, following the procedure of (Takahashi et al., 2012). Data were processed as in (Forrest et al., 2014)
using custom Python scripts available at https://github.com/baillielab/influenza_cage ’ATG analysis.’ The datasets analyzed during
the current study are available in the Fantom5 repository, https://fantom.gsc.riken.jp/5/data/

Ribosome sequencing analyses

Footprints were obtained by first removing the AGATCGGAAGAGC linker and filtering for low quality sequences with Cutadapt (Mar-
tin, 2011). Contigs were then generated from the paired end reads with PANDASeq (Masella et al., 2012) using default parameters.
Concurrent demultiplexing of the libraries by sample ID and UMI extraction was then performed. Reads were then aligned against
rRNA and tRNA sequences with Bowtie (Langmead et al., 2009) to remove these contaminating sequences. Unmapped reads
were aligned against a custom reference containing the human genome (hg38) and the eight genome segments of PR8 with HISAT2
(Kim et al., 2015). Host primer sequences were extracted from this alignment as well as unmapped reads by searching for a match to
conserved nucleotides at the 50 end of the influenza mRNA (GC[GA]AAAGCAGG). These reads were kept if the sequences could be
extended to unambiguously assign it to a segment. Finally, 50 end mapping was performed on these and all reads mapping to PR8.

RNA sequencing Analyses

After adaptor removal with cutadapt (Martin, 2011) and base-quality trimming to remove 30 read sequences if more than 20 bases with
Q < 20 were present, paired-end reads were mapped to the mouse (mm10) reference genome with STAR (Dobin et al., 2013), and
gene-count summaries were generated with featureCounts (Liao et al., 2014). DESeq2 (Love et al., 2014) was used to variance-
normalize the data before a 1-factor model (gene ConditionTimeMutant) was applied to identify differentially expressed genes.
Differentially expressed genes were identified as genes that had a 2-fold difference, with an adjusted p .value < 0.01. RNA-seq
raw data are deposited in GEO: GSE128519. Gene ontology analysis was performed using Metascape (Zhou et al., 2019).

LASV CAGE sequencing Analyses

Unique chimeric host-virus reads were extracted from the resulting FASTQ files by searching for a match to conserved nucleotides at
the 50 end of the LASV (Josiah Strain) mRNAs (GCAC[M]G[N]GGATCCT), allowing for a maximum of 1 mismatch, and removing all
reads with ambiguously mapped nucleotides. The reference genome of LASV was obtained from UniProt (Accessions: J04324
and U73034). Reads were kept if at least 60 nucleotides could be mapped and assigned unambiguously to the viral reference se-
quences. Each read was then split into host derived or virus derived sequences based on the sequences of viral 50 end (GCAC[M]
G[N]GGATCCT). To calculate potential uvORF length, each read was extended bioinformatically, based off the mapped genome
segment and coding sense, and translated from the first AUG found in the read.

Sequence Randomization Model for PB1-UFO length

Influenza A PB1 nucleotide sequences were obtained from the NCBI database (Zhang et al., 2017). Only unique sequences contain-
ing complete 50 UTR regions were included. Sequences containing ambiguous nucleotides were excluded. Multiple sequence align-
ment was then performed by using MUSCLE (Edgar, 2004).
We then constructed a codon usage table for each individual nucleotide sequence. To run the random sequence model, each
nucleotide sequence was translated into two protein sequences in the two translation reading frames of interest: the canonical
PB1 open reading frame (Pr-ORF) and PB1-UFO frame (Pr-UFO). Pr-UFO was considered as the observed protein sequence. Based
on the frequencies of synonymous codons within a codon usage table, each Pr-ORF was reverse translated into multiple random
nucleotide sequences in the open reading frame (Nt-ORFs) 1,000 times. 1,000 Nt-ORFs were then translated into proteins in the
UFO frame (Pr-UFOs) which were considered as the expected protein sequences and their protein lengths were computed. We
used the length of observed Pr-UFO and the lengths of expected Pr-UFOs to calculate the z-score for each nucleotide sequence.
In total, 3140 unique IAV PB1 (H3N2 only: 499) sequences were included in the analysis. From the z-scores, P values were calculated
for the Pr-UFOs occurrence biases. A threshold of p < 0.05 was used for the prediction of the likelihood of IAV PB1 sequences that
were able to be translated. Similar analyses were also performed for other genome segments.

e9 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

Frequency Propagator Ratio Analysis

Sequence dataset
Our study was based on a dataset of 26,742 human influenza A/H3N2 sequences available from the GISAID database (Shu and
McCauley, 2017), which contains 6,244 unique PB1 strains. For downstream analyzes, we included only sequences that are had
a complete 50 and 30 UTR.
Prediction of RNA secondary structure
We used the most abundant unique, full length PB1 nucleotide sequence as an input to predict RNA secondary structure. RNA
secondary structure was predicted using RNAfold from the ViennaRNA Webserver (version 2.4.13) (Gruber et al., 2008), using the
default settings to calculate the minimum free energy (MFE) structure of the PB1 segment RNA. The output structure was saved
in a dot-bracket format, and used to partition nucleotides into probable loop and stem regions for downstream analyses.
Strain tree reconstruction
Our analysis was based on an ensemble of strain trees obtained from the PB1 sequence dataset described above. Such trees
describe the genealogy of influenza strains resulting from an evolutionary process under selection (Strelkowa and Lässig, 2012).
Trees were constructed with maximum-likelihood phylogenies using FastTree (Price et al., 2010). We used a general time-reversible
model. We further refined the tree topology with RAxML (Stamatakis, 2014). Given the output topology, we reconstructed maximum-
likelihood sequences and timing of internal nodes with the TreeTime package (Sagulenko et al., 2018).
Frequency Propagator Ratio analysis
A detailed discussion of this method has previously been presented in Strelkowa and Lässig (2012) and Luksza and Lässig (2014).
Briefly, for a given polymorphism time-series, the frequency propagator GðxÞ can be used as a statistical measure of selection.
GðxÞ is defined as the conditional probability that a mutation class of interest, with an initial frequency of xi , reaches a frequency
of x > xi at a later point in time. This is estimated in our dataset as
nðxÞ
GðxÞ =
n

where nðxÞ is the number of mutations that reach frequency x

and n is the total number of mutations
Data availability might vary, depending on the year of sequence collection (fewer data points are available in the earlier years). As
such, to attain a more robust measure of selection, we use the ratio of propagators between our mutation class of interest, GðxÞ,
against a neutral reference class of mutations, G0 ðxÞ, to calculate
GðxÞ
gðxÞ =
Go ðxÞ

where,

GðxÞ is the likelihood a mutation in a given class reaches frequency, x

Go ðxÞ is the likelihood a mutation in the neutral reference class of mutations reaches the same frequency x.

The frequency propagator ratio takes into account both numbers and histories of the mutation class of interest. It is a robust mea-
sure of selection because it is (a) largely independent of data entry frequency, and (b) insensitive to clonal expansion of mutations.
At the limit x = 1, the propagator ratio gðxÞ reduces to g, where
d=n
g=
d0 =n0

and,

d is the # of mutations in our class of interest that reach fixation

n is the total # of mutations in the same class of interest
d0 is the # of mutations in a neutral reference mutation class that reaches fixation
n0 is the total # of mutations in the same neutral reference mutation class.

Selection on a mutation class of interest can be inferred from the value of g. g < 1 suggests evolutionary constraints (negative se-
lection) on the mutation class of interest relative to the reference class, where a fraction ð1  gÞ of the mutations are under negative
selection. g > 1 suggests that fixation of the mutation class of interest undergoes positive selection, and that at least a fraction
ðg 1Þ=g of that mutation class is beneficial. gz1 suggests weak or heterogenous selection acting on the mutation class of interest,
relative to that of the neutral reference class.
To quantify selection occurring across the PB1-UFO frame, we calculated mutation frequencies in the set of codons derived from
the following three regions (R1-R3)

Cell 181, 1502–1517.e1–e11, June 25, 2020 e10

 ll
OPEN ACCESS Article
R1: sequences that encode the N-terminal of PB1-UFO and the viral 50 UTR
R2: sequences that encode the C-terminal of PB1-UFO and overlap with the N-terminal of PB1
R3: Sequences that encode for the C-terminal region of the main PB1 ORF and do not overlap with PB1-UFO.

We chose to use synonymous mutations in the main PB1 ORF (reading frame) in R3 as our neutral reference class to calculate
G0 ðxÞ, as we reasoned that the majority of such mutations evolve near neutrality.
To quantify selection on the N-terminal of PB1-UFO in R1, we calculated the GðxÞ for two classes of mutations: Those that changed
(non-synonymous in PB1-UFO) or did not change (synonymous in PB1-UFO) the amino acid sequence of PB1-UFO. We used
synonymous mutations occurring the PB1 ORF in R3 as our neutral reference class (G0 ðxÞ). We found that g < 1 for both cases,
suggesting that mutations occurring in this region of PB1-UFO were not likely to be fixed over time, and mostly undergo negative
selection, relative to our reference class.
To quantify selection on the C-terminal of PB1-UFO in R2, we again calculated the GðxÞ for mutations that changed (non-synon-
ymous in PB1-UFO) or did not change (synonymous in PB1-UFO) the amino acid sequence of PB1-UFO. We used synonymous
mutations occurring the PB1 ORF in R3 as our neutral reference class ðG0 ðxÞÞ. We found that gz1 for both cases, suggesting
that mutations occurring in this region of PB1-UFO underwent heterogenous selection, relative to that of the reference class.
Since R2 mutations in PB1-UFO appear to undergo heterogeneous selection, we asked if selection occurring on the main PB1 ORF
was a contributing factor. To do so, we calculated the GðxÞ for mutations that changed (non-synonymous in PB1) or did not change
(synonymous in PB1) the amino acid sequence of PB1 in R2. Synonymous mutations occurring the PB1 ORF in R3 as our neutral
reference class ðG0 ðxÞÞ. Here we found that g > 1 for synonymous mutations and g < 1 for non-synonymous mutations, suggesting
that mutations that do NOT alter the amino acid sequence of PB1 are preferentially fixed over time. This suggests to us that part of the
reason why PB1-UFO is undergoing heterogeneous selection in R2 is that there is a requirement to maintain the protein sequence of
PB1. This is not surprising, given that PB1 is an integral part of the viral RNA dependent RNA polymerase complex.
Finally, to interrogate the effect of RNA structure, we classified nucleotides as pairing or non-pairing based on the MFE structure
(discussed above) calculated by RNAFold. We masked nucleotides that were predicted to base pair (‘‘stem-forming’’) from down-
stream analyses as we reasoned that mutations in these nucleotides are likely to affect both RNA structure AND protein sequence,
thus confounding later interpretations of the data. Regions that were not predicted to base pair (‘‘loop nucleotides’’) were then used
for downstream calculations of frequency propagator ratios. Mutation frequencies were calculated in the same regions (R1, R2 and
R3) and reading frames (PB1-UFO versus PB1) as described above. We found that similar effects to before were found, suggesting
that RNA structure was not a major contributor to the maintenance of the PB1-UFO frame.
Note: The absolute number of polymorphism histories that reach a given frequency are finite (since the tree is constructed over a
defined period of time). This can give rise to sampling fluctuations. These sampling uncertainties are reported as error bars in
our figures.

Epitope predictions for PB1-UFO

Analyses were done using NetMHC3.4 and NetMHC4.0 (Andreatta and Nielsen, 2016). Binders were filtered using KD threshold of
500 nM. The collection of viral MHC-I epitopes was downloaded from IEDB database and preformatted for BLAST usage (makeblastdb
-in iedb.fasta -parse_seqids -dbtype prot). Predicted epitopes from PB1-UFO were BLASTed against IEDB and the human proteome.
For comparison with viral antigens we used the following commands: blastp -db iedb.fasta -query antigens.fasta -outfmt ‘‘6
qseqid sseqid pident ppos positive mismatch gapopen length qlen slen qstart qend sstart send qseq sseq evalue bitscore’’ -word_size
3 -gapopen 32767 -gapextend 32767 -evalue 1 -max_hsps_per_subject 1 -matrix BLOSUM62 -max_target_seqs 10000000 -out anti-
gens.iedb.blast.out. For comparison with human proteome we used the command: blastp -db human.proteome.fasta -query anti-
gens.fasta -outfmt ‘‘6 qseqid sseqid pident ppos positive mismatch gapopen length qlen slen qstart qend sstart send qseq sseq evalue
bitscore’’ -word_size 3 -gapopen 32767 -gapextend 32767 -evalue 1 -max_hsps_per_subject 1 -matrix BLOSUM62 -max_target_seqs
10000000 -out antigens.human.proteome.blast.out. To find perfect matches between predicted epitopes and human proteome or viral
antigens, we used the last command. First, we preformatted the human proteome (ensemble archive from December 2016): lastdb -p
human.proteome human.proteome.fasta. Then we used following command to compare epitopes to this database: lastal -f MAF -r 2 -q
1 -m 100000000 -a 100000 -d 15 -l 4 -k 1 -j1 -P 10 human.proteome antigens.netMHC.score.fasta > antigens.human.last.out. Finally,
obtained results were processed with bash and python and finally analyzed in PRISM 8. Similar processing was performed with viral
antigens.

e11 Cell 181, 1502–1517.e1–e11, June 25, 2020

 ll
Article OPEN ACCESS

Supplemental Figures

(legend on next page)

 ll
OPEN ACCESS Article

Figure S1. uAUGs Are Present in Viral mRNAs, Related to Figure 1

(A) Incorporation of host transcript sequences increases the diversity of putative alternative start codons. For each viral genome segment, the frequency and
position of alternative start codons is shown relative to native start of the viral genes. For each reading frame, the frequency and location of the first in-frame stop
codon are indicated.
(B) Percentages of cap-snatched sequences that contain AUG codons, as identified by CAGE. Data are shown relative to all the viral reads from the specified
genome segments.
 ll
Article OPEN ACCESS

Figure S2. Viral 50 UTRs Are Conserved, Related to Figure 2

Multiple sequence alignments of unique H1N1 IAV 50 UTRs per genome segment (n = 10904). The overall distribution of each unique nucleotide sequence is
indicated on the left, and the consensus sequence of each UTR is indicated below each alignment. The top panels show the positional weight matrix of each
nucleotide across the UTRs.
 ll
OPEN ACCESS Article

(legend on next page)

 ll
Article OPEN ACCESS

Figure S3. IAV mRNAs Can Be Translated from Host-Derived AUGs, Related to Figure 3
(A) Length distribution of ribosome profiling reads that aligned to human (left panel) and viral (right panel) transcripts in DMSO (Ribo) or harringtonine (Ribo + Harr)
treated samples.
(B) Metagene alignment of average P site density around annotated start codons in human (left panel) or viral (right panel) transcripts in DMSO treated samples.
(C) Metagene alignment of average P site density around annotated start codons in human (left panel) or viral (right panel) transcripts in harringtonine treated
samples.
(D) Frequency of AUG codons by position relative to the viral transcription initiation site. Bars show the mean frequency and are color coded according to frame.
Error bars indicate the standard deviation.
 ll
OPEN ACCESS Article

(legend on next page)

 ll
Article OPEN ACCESS

Figure S4. uvORFs Are Expressed during Infection and Can Contribute to Virulence, Related to Figure 4
(A) Plots showing the position of uvORF peptides found in lysates of cells (A549 or 293) infected with A/PR/8/34 virus at 8 or 24h post infection. The specific cell
lysates they were found in are indicated on the right. 1: MG132 treated, 2: DMSO treated. Peptide locations are drawn relative to uvORFs (gray regions) and
canonical ORFs (blue regions) and are colored by the log10 of their intensities, relative to the sample median.
(B) Same as in (A), but for uvORF peptides found within purified A/WSN/33 virions.
(C) Same as in (A), but for uvORF peptides found from an independent, previously published dataset.
(D) In vitro growth curves of the indicated mutant (UFOD) and control (UFOSYN) viruses made in the PR8 background, and performed on MDCK cells. Error bars
indicate the standard deviation of 3 replicates.
(E) In vitro growth curves of the indicated mutant (UFOD) and control (UFOSYN) viruses made in the WSN/33 background, and performed on MDCK cells.
(F) In vitro growth curves of the indicated mutant (UFOD) and control (UFOSYN) viruses made in the Cal/09 background, and performed on A549 cells. Error bars
indicate the standard deviation of 3 replicates.
(G) Heatmap of differentially expressed genes (Fold Change > 2, p < 0.01) found in the lungs of mice infected with 100PFU of either the PR8;PB1-UFOD or
PR8;PB1-UFOSYN viruses at day 6 post infection.
(H) qPCR validation of four significantly changed genes identified in (G) (highlighted with green text). Each dot represents the lung of one mouse infected with
100PFU of the indicated viruses, collected at day 6 post infection. P values were calculated through a one tailed t test. *p < 0.05
(I) Gene ontology analysis of genes shown in (G).
 ll
OPEN ACCESS Article

(legend on next page)

 ll
Article OPEN ACCESS

Figure S5. uvORFs Are Conserved, Related to Figure 5

(A) Bar plot showing the number of unique NP sequences that give rise to the full length, extended NP protein of 514aa, or those that result in truncated (non-
extended) uvORFs.
(B) Percentages of unique NP sequences that preserve the propensity to code for NP-extension.
(C) Top five most common NP extension protein sequences in three types of influenza A strains, H1N1, H3N2 and H5N1.
(D) Schematic showing the model used to calculate the expected versus observed PB1-UFO sequence lengths.
(E) Density plot of predicted length of H3N2 PB1-UFO protein sequences. Sequences predicted to generate a protein of 77aa are shown in medium blue, shorter
than 77aa in light blue, and those longer than 77aa are in dark blue. Sequences predicted not to generate PB1-UFO protein are shown in gray.
(F) P value distribution/volcano plot of H3N2 PB1-UFO protein sequence length. Each dot represents the difference between observed length and expected
length of each individual sequence.
(G) Density plot showing the distribution of expected lengths of H3N2 PB1-UFO proteins, based on random codon-shuffled sequences.
(H) Line plot showing the number of synonymous mutations in frame of WT H3N2 PB1 (x axis) that are required to generate stop codons in frame of H3N2 PB1-
UFO (y axis).
 ll
OPEN ACCESS Article

Figure S6. Controls Related to Propagator Analysis, Related to Figures 5C–5F

(A) Schematic of analysis steps taken to quantify selection occurring on synonymous and non-synonymous mutations in the PB1-UFO ORF. Propagator model
analyses were done by either not taking (Figure 5B and 5D) or taking the RNA structure of IAV PB1 segment into account (Figures 5C–5E).
(B) Frequency propagator ratios of the indicated classes of mutations occurring in PB1-UFO relative to the PB1 open reading frame of H3N2 viruses. The region
used to calculate the test class ratio (G(X)) is indicated in yellow, and the region used to calculate the neutral class ratio (G0(X)) is indicated in blue in the top
schematic. Here, the test class is the region of the PB1-UFO ORF that overlaps only with the virally-encoded 50 UTR; the neutral class consists of synonymous
mutations in the PB1 ORF that do not overlap with PB1-UFO. Only nucleotides within predicted loop regions (i.e., non-pairing) positions were considered. Error
bars indicate sampling uncertainties. gðxÞ < 1: negative selection, gðxÞz1: weak/heterogeneous selection; gðxÞ > 1: positive selection; see also Figure 5C)
(C) Frequency propagator ratios, as in (B), but with the test class comprising the C-terminal region of the PB1-UFO ORF.
(D) Frequency propagator ratios, as in (B), but with the test class comprising the region in the main PB1 ORF overlapping the PB1-UFO reading frame.
 ll
Article OPEN ACCESS

Figure S7. DEFEND-Seq and CAGE Analysis of Other Cap-Snatching Viruses, Related to Figure 6
(A) Distribution of lengths for cap-snatched sequences found in IBV, as determined by DEFEND-seq.
(B) Host derived uAUGs give rise to long uvORFs (> 30aa). (Upper panels) Predicted peptide sequences derived upon translation of all three ribosome reading
frames in the indicated IBV genome segments. (Lower panels) Predicted distribution of the lengths of new ORF and extension peptides generated from each
reading frame of the viral 50 UTR. Peptide lengths are calculated based on AUG positions obtained through DEFEND-sequencing.
(C) Distribution of lengths for cap-snatched sequences found in LASV infected cells, as determined by CAGE-seq.
(D) Host derived uAUGs enable reverse sense genome segments of Lassa virus L and S to give rise to uvORFs and extensions. (Upper panels) Schematic of
proteins encoded in the indicated reading frames in either the L or S segment. Lassa virus RNA is ambisense. (Middle panels) Predicted peptide sequences
derived upon translation of all three reading frames in the reverse sense L and S segments. (Lower panels) Predicted distribution of the lengths of new ORFs and
extension peptides generated from each reading frame of the viral 50 UTR. Peptide lengths are calculated based on AUG positions obtained through CAGE.
(legend continued on next page)
 ll
OPEN ACCESS Article

(E) (Left panels) Schematic showing (in coding sense) the 50 termini of viral reporter RNAs, in which a viral untranslated region (UTR) flanks a luciferase (Luc)
reporter gene. Reporter RNAs were used to assess upstream translation in the mRNAs of Heartland virus (HRTV). The 50 terminus of the mRNAs consisted of cap-
snatched sequence from host mRNAs (cap), followed by a viral 50 UTR (50 UTR) and the reporter gene (Luc). Cap structures are indicated as circles, the most
N-terminal AUG as a triangle, AUG mutations as crosses and stop codons as lines. (Right panels) Luc expression when these reporters were included in min-
ireplicon assays, as a percentage of expression with the WT construct, showing the means and s.d. of 3 repeats compared to WT-STOP by Student’s 2-tailed t
test (n.s.: p R 0.05, *p < 0.05, ***p % 0.0005).
 Article

A Dual-Mechanism Antibiotic Kills Gram-Negative

Bacteria and Avoids Drug Resistance
Graphical Abstract Authors
James K. Martin II, Joseph P. Sheehan,
Benjamin P. Bratton, ...,
Mikhail M. Savitski, Maxwell Z. Wilson,
Zemer Gitai

Correspondence
zgitai@princeton.edu

In Brief
A compound that kills both Gram-positive
and Gram-negative bacteria through two
independent mechanisms may provide a
platform for the development of future
antibiotics.

Highlights
d SCH-79797 kills Gram-negative and Gram-positive bacteria
with undetectable resistance

d It works by simultaneously targeting folate metabolism and

membrane integrity

d SCH’s dual-targeting is synergistic, but only when on the

same chemical scaffold

d Irresistin-16, an SCH derivative, effectively treats mouse

N. gonorrhoeae infection

Martin et al., 2020, Cell 181, 1518–1532

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.005 ll
 ll

Article
A Dual-Mechanism Antibiotic Kills Gram-Negative
Bacteria and Avoids Drug Resistance
James K. Martin II,1,7 Joseph P. Sheehan,1,7 Benjamin P. Bratton,1,2,7 Gabriel M. Moore,1 André Mateus,3
Sophia Hsin-Jung Li,1 Hahn Kim,4,5 Joshua D. Rabinowitz,2,4 Athanasios Typas,3 Mikhail M. Savitski,3
Maxwell Z. Wilson,1,6 and Zemer Gitai1,8,*
1Department of Molecular Biology, Princeton University, Princeton, NJ 08544, USA
2Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, NJ 08544, USA
3European Molecular Biology Laboratory, Genome Biology Unit, 69117 Heidelberg, Germany
4Department of Chemistry, Princeton University, Princeton, NJ 08544, USA
5Princeton University Small Molecule Screening Center, Princeton University, Princeton, NJ 08544, USA
6Department of Molecular, Cellular, and Developmental Biology, Center for BioEngineering, University of California, Santa Barbara, Santa

Barbara, CA 93106, USA

7These authors contributed equally
8Lead Contact

*Correspondence: zgitai@princeton.edu
https://doi.org/10.1016/j.cell.2020.05.005

SUMMARY

The rise of antibiotic resistance and declining discovery of new antibiotics has created a global health crisis.
Of particular concern, no new antibiotic classes have been approved for treating Gram-negative pathogens in
decades. Here, we characterize a compound, SCH-79797, that kills both Gram-negative and Gram-positive
bacteria through a unique dual-targeting mechanism of action (MoA) with undetectably low resistance fre-
quencies. To characterize its MoA, we combined quantitative imaging, proteomic, genetic, metabolomic,
and cell-based assays. This pipeline demonstrates that SCH-79797 has two independent cellular targets,
folate metabolism and bacterial membrane integrity, and outperforms combination treatments in killing
methicillin-resistant Staphylococcus aureus (MRSA) persisters. Building on the molecular core of SCH-
79797, we developed a derivative, Irresistin-16, with increased potency and showed its efficacy against Neis-
seria gonorrhoeae in a mouse vaginal infection model. This promising antibiotic lead suggests that
combining multiple MoAs onto a single chemical scaffold may be an underappreciated approach to targeting
challenging bacterial pathogens.

INTRODUCTION monas aeruginosa, is more effective than other fluoroquinolones

More than twenty unique classes of antibiotics were characterized
in the 30 years following the discovery of penicillin in 1929 (Coates
et al., 2011; Davies, 2006). However, a combination of scientific
and economic factors have slowed the discovery and develop-
ment of these life-saving molecules to the extent that only six
new classes of antibiotics have been approved in the past 20
years, none of which are active against Gram-negative bacteria
(Butler et al., 2017). This decline in the discovery of new antibiotic
classes, coupled with the evolution of multi-drug resistant bacte-
ria and horizontal transfer of resistance mechanisms, has created
a public health crisis that is predicted to only escalate in the com-
ing years (Culyba et al., 2015; Hofer, 2019; O’Neill, 2014).
Recent efforts have begun to reinvigorate antibiotics research,
but most of this work has resulted in compounds that function via
similar mechanisms to those of traditional antibiotics. For
example, finafloxacin, a fluoroquinolone antibiotic that was
recently approved to treat ear infections caused by Pseudo-
because it maintains its potency in acidic environments (McKe-
age, 2015). However, finafloxacin is still susceptible to the same
resistance mechanisms that affect other fluoroquinolones
(Randall et al., 2016). The recent discovery of the natural product
teixobactin suggests that it is possible to find compounds that
selectively kill bacteria without being prone to resistance (Ling
et al., 2015). However, teixobactin is only functional against
Gram-positive bacteria. Another natural product, darobactin,
was recently found to specifically target Gram-negative bacteria,
but resistance to darobactin was relatively easy to achieve
through mutations in bamA (Imai et al., 2019). Thus, there is still
a strong need for characterizing new classes of antibiotics with
distinct mechanisms of action (MoA), especially those that target
Gram-negatives with low resistance frequency.
An ideal antibiotic would be hard to develop resistance
against, able to kill both Gram-positive and Gram-negative bac-
teria, and easy to access. It is important to note that while anti-
biotics that are not prone to resistance are attractive clinically,

1518 Cell 181, 1518–1532, June 25, 2020 ª 2020 Elsevier Inc.

Article
selecting for resistant mutants is the most common method for
characterizing MoA, making the characterization of new anti-
biotic MoAs without resistance mutants a significant challenge.
Phenotypic methods, such as macromolecular synthesis as-
says, have been previously used in such cases, as was done
for teixobactin (Ling et al., 2015). However, these assays only
allow the classification of molecules with previously described
MoAs (King and Wu, 2009). Thus, there is also a need for resis-
tance-independent approaches for the de novo characterization
of antibiotic MoA.
Here, we describe a compound, SCH-79797, which is bacte-
ricidal toward both Gram-negative and Gram-positive bacteria,
including clinically significant bacterial pathogens such as meth-
icillin resistant Staphylococcus aureus (MRSA), Enterococcus
faecalis, Neisseria gonorrhoeae, and Acinetobacter baumannii,
with no signs of resistance. In an animal host model, SCH-
79797 blocked infection by A. baumannii with low toxicity to
the host at the dose required for effective antibiotic activity. To
rapidly and efficiently classify the MoA of SCH-79797, we used
a variant of a recently described quantitative imaging-based
approach known as bacterial cytological profiling (BCP) (None-
juie et al., 2013). This effort showed that SCH-79797 functions
through a mechanism distinct from that of most known classes
of antibiotics. In the absence of being able to evolve resistant
mutants, we used thermal proteome profiling (Savitski et al.,
2014), CRISPRi genetic sensitivity (Peters et al., 2016), and me-
tabolomic profiling (Kwon et al., 2008; Kwon et al., 2010) to char-
acterize the MoA of SCH-79797. Using this multi-dimensional,
systems-level approach, we identified the candidate targets of
SCH-79797 as dihydrofolate reductase and the bacterial mem-
brane. Classical enzymology and membrane permeability and
polarization assays confirmed the targets identified by our
high-throughput approaches. By analyzing derivatives of the
SCH-79797 structure, we demonstrated that the two pharmaco-
phores of this compound can be distinguished. Finally, we
describe a derivative of SCH-79797, Irresistin-16 (IRS-16), with
improved potency that demonstrates efficacy in a mouse vaginal
Neisseria gonorrhoeae model. Thus, our findings identify and
characterize a promising antibiotic candidate and provide a po-
tential roadmap for future antibiotic discovery efforts.
RESULTS
SCH-79797 Is a Broad-Spectrum, Bactericidal Antibiotic
With the aim of finding antibiotics with novel mechanisms of ac-
tion (MoA), we began with an unbiased, whole-cell screening
approach. To include antibiotics that target both Gram-negative
and Gram-positive bacteria, we screened for compounds that in-
hibited the growth of E. coli lptD4213, which has a compromised
outer membrane that makes it partially permeable to antibiotics
that would otherwise have difficulty penetrating the Gram-nega-
tive lipopolysaccharide (Ruiz et al., 2006). We screened a small
molecule library of 33,000 unique compounds and one of our
most potent hits was SCH-79797, a compound that had been
previously reported as a human PAR-1 antagonist (Ahn et al.,
2000). This finding was surprising because there are no PAR-1
homologs in bacteria. A recent report suggested that SCH-
79797 increases the ability of neutrophils to kill bacteria, perhaps
ll
by directly functioning as an antibiotic (Gupta et al., 2018). Given
that studies focusing on characterizing its anticoagulant activ-
ities suggested that at least 5 mg/kg SCH-79797 can be safely
tolerated in animals (Gobbetti et al., 2012; Strande et al., 2007)
and its emergence as a potential antimicrobial with no known
bacterial target (Gupta et al., 2018), we decided to further char-
acterize SCH-79797 as a candidate antibiotic.
To assess the spectrum of bacterial species susceptible to
SCH-79797, we measured the minimal inhibitory concentration
(MIC) of SCH-79797 against several clinically relevant patho-
gens, including the ESKAPE pathogens (Boucher et al., 2009).
In this study, we define MIC as the concentration of drug that re-
sults in no visible bacterial growth after 14 h of growth at 37 C.
We found that SCH-79797 significantly hindered the growth of
multiple Gram-negative and Gram-positive pathogens including
Neisseria gonorrhoeae, two clinical isolates of Acinetobacter
baumannii, Enterococcus faecalis, and Staphylococcus aureus
(Figure 1A; Table S1). SCH-79797 also exhibited potent activity
against several antibiotic-resistant pathogen strains including
multi-drug-resistant WHO-L N. gonorrhoeae and MRSA
S. aureus. Using the E. coli lptD4213 strain from our original
screen, SCH-79797 exhibited potent and rapid bactericidal ac-
tivity (Figures 1B and 1C). SCH-79797 also exhibited similar
bactericidal activity against a clinical isolate of S. aureus MRSA
USA300 (Tenover and Goering, 2009) suggesting that its bacte-
ricidal activity is not species-specific (Figure S1A).
SCH-79797 Is Effective In Vivo and Has a Low Frequency
of Resistance
Given SCH-79797’s promising ability to kill bacteria, we sought to
determine if it can function as an effective antibiotic in vivo. To test
its antibiotic activity in the context of an animal host infection, we
focused on A. baumannii as it has emerged as an important Gram-
negative pathogen that is targeted by relatively few available anti-
biotics, and has a well-established host animal model in the wax
worm, Galleria mellonella (Gebhardt et al., 2015; Peleg et al.,
2009). We first established that injecting G. mellonella with SCH-
79797 at concentrations four times higher than the MIC of SCH-
79797 toward A. baumannii did not result in higher host toxicity
than the solvent-only control (Figures S1E and S1F; Table S2).
We next tested the ability of SCH-79797 to treat infection of
G. mellonella with a lethal dose of A. baumannii AB17978. Treat-
ment with SCH-79797 significantly prolonged the survival of
A. baumannii-infected G. mellonella (p < 0.001) (Figures 1D,
S1G, and S1H). The survival rate of G. mellonella treated with
SCH-79797 was similar to the control antibiotics meropenem,
rifampicin, and gentamicin (Figures 1D, S1G, and S1H) (Karlowsky
et al., 2003; Viehman et al., 2014).
To further characterize its promise as an antibiotic, we attemp-
ted to determine the frequency with which bacteria develop
resistance toward SCH-79797. Because spontaneous suppres-
sors can restore E. coli lptD4213’s membrane barrier function-
ality, we focused our resistance studies on S. aureus MRSA
USA300 (Tenover and Goering, 2009). We were unable to isolate
stable SCH-79797-resistant mutants upon plating 108 CFU of
MRSA USA300 onto agar containing 25 mg/mL SCH-79797 (4X
MIC). We were also unable to isolate SCH-79797-resistant mu-
tants upon plating 108 colony-forming units (CFUs) of
Cell 181, 1518–1532, June 25, 2020 1519
 ll
Article

A B

D E

Figure 1. SCH-79797 Is a Broad-Spectrum, Bactericidal Antibiotic that Is Effective in an Animal Model and Has a Low Frequency of
Resistance
(A) The MIC of SCH-79797 against Gram-negative (red) and Gram-positive (black) bacteria. The MICs of E. cloacae and P. aeruginosa were greater the maximal
drug concentrations tested. See also Table S1.
(B) The relative growth of E. coli lptD4213 after treatment with SCH-79797. Bacterial growth was measured as the optical density at 600 nm (OD600) 14 h following
inoculation. Each data point represents 2 biological replicates. Mean ± SD are shown.
(C) Colony forming units (CFUs mL 1) after 3-h treatment of E. coli lptD4213 with 1% DMSO (solvent control), 6.2 mg/mL SCH-79797 (23 MIC), 0.12 mg/mL
ampicillin (23 MIC), or 0.48 mg/mL novobiocin (43 MIC). Data points at 1 3 102 CFU mL 1 are below the level of detection. Each data point represents 3 biological
replicates. Mean ± SD are shown.
(D) The percent survival of G. mellonella wax worms infected with A. baumannii and concomitantly treated with 2 mL/larva 100% DMSO, 67 mg/larva SCH-79797,
67 mg/larva gentamicin, 67 mg/larva merapenem, or 67 mg/larva rifampicin. Data represents a typical cohort (n = 12) from a biological triplicate. p values are
determined from a Mantel-Cox test using Prism (**p < 0.01; ***p < 0.001), and the pooled results are presented in the supplemental material (Figure S1G). For other
Mantel-Cox comparisons, see Table S2.
(E) Fold increase in resistance of S. aureus MRSA USA300 to SCH-79797, novobiocin, trimethoprim, or nisin after 25 days of serial passaging in each drug. Data
represents one biological replicate and the data for a second replicate is shown in Figure S1B.

B. subtilis, suggesting that the difficulty in developing resistant dent cultures of S. aureus MRSA USA300 in sub-lethal concen-
mutants is not species-specific. To address resistance rates trations of SCH-79797, as well as three control antibiotics: novo-
more quantitatively, we serial-passaged 2 biologically indepen- biocin, trimethoprim, and nisin. Over the course of 25 days, we

1520 Cell 181, 1518–1532, June 25, 2020


Article
A Figure 2. Bacterial Cytological Profiling
B biologically
successfully isolated mutants resistant to all the control antibi-
otics while no SCH-79797-resistant mutants emerged (Figures
1E and S1B). For novobiocin, trimethoprim, and nisin, resistance
gradually increased throughout the experiment, while the resis-
tance level remained constant for SCH-79797 (Figures 1E and
S1B), indicating that these bacteria did not even acquire partial
resistance to SCH-79797. In addition, the mutants that evolved
increased resistance to antibiotics like trimethoprim and nisin
ll
Indicates that SCH-79797 Functions by a Mech-
anism Distinct from Known Classes of Antibi-
otics
(A) Fluorescent images of E. coli lptD4213 cells treated
with antibiotics representative of 5 different antibiotic
classes. Cells were treated for 2 h with 53 MIC of each
drug. Merged image channels are phase contrast
(gray), FM4-64 (red), DAPI (blue), and SYTOX (green).
Scale bar, 1 mm.
(B) Comparison of cytological profiles of known anti-
biotics with the cytological profile of SCH-79797.
Single-linkage clustered dendrogram from one-way
MANOVA comparisons between antibiotic treatment
groups compared to all other antibiotic treatment
groups. Inset: structure of SCH-79797.
did not demonstrate cross-resistance to
SCH-79797 (Figure S1D). To extend these
findings to a Gram-negative species, we
repeated our serial passaging study with 2
independent cultures of
A. baumannii AB17978 (Figure S1C).
A. baumannii resistance remained constant
for SCH-79797 but increased for all other an-
tibiotics, including gentamycin, supporting
the conclusion that the lack of resistance to
SCH-79797 is not species-specific.
A Variant of Bacterial Cytological
Profiling Suggests that SCH-79797
Has a Unique MoA
The inability to isolate SCH-79797-resistant
mutants makes SCH-79797 an appealing
candidate antibiotic but poses a challenge
for determining its MoA. As a result, we
used a quantitative imaging-based
approach to determine if the MoA of SCH-
79797 is similar to that of any previously
characterized antibiotics. Specifically, we
modified a single-cell, high-content imaging
methodology, known as BCP (Nonejuie
et al., 2013). The logic of BCP is that antibi-
otics with similar MoA result in similar death
phenotypes such that by quantifying how
bacteria appear upon death, we can gain
insight into the cause of death (much like a
bacterial autopsy). Here, we applied our
BCP analysis to a training set of 37 distinct
antibiotics with known MoA as well as to
SCH-79797. For each compound, we treated E. coli lptD4213
with 5X MIC of an antibiotic for 2 h, stained with three dyes
that report on nucleoid morphology (DAPI), membrane
morphology (FM4-64), and membrane integrity (SYTOX Green),
and imaged the cells at high resolution (Figure 2A). For each
condition we imaged 100 cells and quantified 14 parameters
reflecting various morphological and fluorescence features
(Table S3).
Cell 181, 1518–1532, June 25, 2020 1521
 ll
Article
Figure 3. Thermal Proteome Profiling Sug-
A gests that SCH-79797 Binds Dihydrofolate
Reductase
(A) Schematic of the thermal shift assay that
+ compares the thermal stability of the entire pro-
Drug teome with and without drug treatment. Protein
samples are aliquoted, and each aliquot is heated
to a different temperature. The relative fraction of
soluble and insoluble proteins is then determined
for each aliquot by ultracentrifugation and mass
spectrometry.
(B and C) The relative thermal stability of the sol-
uble E. coli lptD4213 proteome after treatment of
whole cell and cell lysate samples with SCH-
79797 (B) or trimethoprim (C). Changes in thermal
Increasing Temperature Increasing Temperature stability were determined by measuring changes
in the abundance of soluble protein across 10
B C different temperatures ranging from 42 C–72 C
SCH-79797 Trimethoprim
and 4 drug concentrations and a vehicle control.
mild effect mild effect
2.5 For each point, the color indicates the maximal
maximum log2 fold-change in whole cell

maximum log2 fold-change in whole cell

effect at multiple effect at multiple

2
2.0 temperatures temperatures
FolA effect size across all temperatures and the largest
moderate and moderate and
1.5 consistent effect FolA 1 consistent effect change in abundance across all concentrations.
1.0 Squares represent the proteins with a change in
0.5
0 abundance of at least 25% at three or more
0.0
temperatures. To be considered consistent, the
-1
change in abundance of a protein had to show the
-0.5 2
average log2 fold-change

average log2 fold-change

-2
same sign at least 90% of the time and have an
-1.0 1 1
effect size of at least 2-fold in either whole cells or
-1.5 0 0
-3 cell lysates. Triangles represent a milder effect
-2.0 -1 -1
where at least one temperature had a change in
-2 -4 -2
-2.5 abundance of at least 25% in both whole cell and
-2.5 -2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0 2.5 -4 -3 -2 -1 0 1 2 3 cell lysate treatments.
maximum log2 fold-change in cell lysate maximum log2 fold-change in cell lysate

Because we had gold standards of the BCP results of SCH- SCH-79797 targets. Specifically, we used thermal proteome
79797 and of several antibiotics representing different classes profiling, an assay that uses mass spectrometry to compare
and sub-groups within classes, we applied a machine learning the thermal stability of the entire proteome with and without
approach to classify the BCP data. Using a one-way MANOVA, drug treatment (schematized in Figure 3A) (Mateus et al., 2018;
we performed dimensionality reduction to remove the influence Savitski et al., 2014). Briefly, intact cells or cell lysate samples
of naturally covarying metrics, such as cell length and cell treated with a range of compound concentrations are heated
perimeter. We then used single-linkage clustering to cluster to a series of increasing temperatures and the soluble proteins
treatment groups by their neighborhood representation vec- at each temperature are collected (Becher et al., 2016). Proteins
tors, such that samples whose neighborhoods were similar that bind to the drug are thermally stabilized, which leads to a
would be clustered together. This analysis indicated that shift in the temperature at which those proteins precipitate (Fig-
SCH-79797 resulted in a phenotypic death-state that was ure 3A). Using E. coli lptD4213, we treated intact cells and cell ly-
different from the other antibiotics tested (Figure 2B). We also sates with SCH-79797 and found that it significantly shifted the
tested an additional method to assess whether this result thermal stability of dihydrofolate reductase (the DHFR homolog
was robust to multiple statistical methods. Using UMAP for in E. coli is known as FolA) (Figures 3B and S3A). The fact that
dimensionality reduction on Z score normalized data (McInnes the same result was observed with both intact cells and cell ly-
et al., 2018), we found that SCH-79797-treated cells formed a sates (Figures 3B and S3A) suggests that SCH-79797 enters
distinct peak separated from other treatments (Figure S2). E. coli cells and directly binds to FolA. As a positive control,
This result supports our conclusion that SCH-79797 is dissim- we used a well-characterized antibiotic that targets DHFR,
ilar from most antibiotics tested and therefore likely possesses trimethoprim, and found that it also thermally stabilizes its known
a MoA distinct from that of any of the antibiotics in our target, the E. coli DHFR, FolA (Figures 3C and S3B) (Gleckman
training set. et al., 1981).
To test both the physiological significance and species-spec-
Thermal Profiling and CRISPRi Genetics Demonstrate ificity of the suggestion that SCH-79797 binds to DHFR, we took
that SCH-79797 Targets Dihydrofolate Reductase advantage of a collection of B. subtilis essential gene CRISPR-
In the absence of resistant mutants or similarity to antibiotics interference (CRISPRi) knockdown mutants (Peters et al.,
with known MoA by BCP, we turned to a high-throughput prote- 2016). In each of these mutants, an essential gene is targeted
omics-based approach for de novo identification of candidate by CRISPRi to reduce its expression 3-fold. A strain with

1522 Cell 181, 1518–1532, June 25, 2020


Article
reduced levels of the SCH-79797 target should be sensitized to
sub-lethal doses of SCH-79797. Given the thermal profiling
result, we focused on mutants in the folate biosynthesis pathway
(schematized in Figure 4A). As a negative control, we confirmed
that CRISPRi knockdowns of genes unrelated to folate meta-
bolism are not sensitized to SCH-79797 (Figure S4). As a positive
control for our assay, we again utilized trimethoprim. We
confirmed that dihydrofolate reductase (dfrA in B. subtilis) and
dihydrofolate synthase (folC, an enzyme that acts upstream of
DfrA) knockdowns are hypersensitive to trimethoprim, while
knockdowns of enzymes that function downstream of DHFR,
folD and glyA, are not (Figure 4B). SCH-79797 exhibited the
same genetic sensitivity pattern as trimethoprim in that both
dfrA and folC, but not folD and glyA knockdowns, were sensi-
tized to SCH-79797 (Figure 4B).
SCH-79797 Inhibits DHFR Activity in Cells and in Purified
Enzymatic Assays
To determine how SCH-79797 affects folate metabolism in living
cells, we used mass spectrometry to measure the relative abun-
dance of folate metabolite pools in E. coli NCM3722 treated with
SCH-79797. E. coli NCM3722 was used because these bacteria
lack mutations that disrupt primary metabolism in other lab
strains of E. coli (Soupene et al., 2003). E. coli NCM3722 cells
were grown in Gutnick Minimal Media and treated with
13.9 mg/mL SCH-79797 (13 MIC) for 15 min (Kwon et al.,
2008, 2010). In response to SCH-79797 treatment, the levels of
the DHFR substrate, 7,8-dihydrofolate (DHF), rose 10-fold
compared to untreated cells, while the levels of folate metabo-
lites downstream of DHF dropped significantly (Figure 4C).
This metabolic response is characteristic of dihydrofolate reduc-
tase (FolA in E. coli) inhibition as we observed a similar pattern
upon treatment with trimethoprim, a known inhibitor of DHFR
(Figure 4C) (Gleckman et al., 1981).
To determine whether SCH-79797 inhibits DHFR directly, we
obtained purified E. coli FolA protein and measured its enzymatic
activity in the presence of increasing concentrations of SCH-
79797. We found that SCH-79797 has an IC50 of 8.6 ± 3 mM
against FolA (Figure 4D). We also measured the initial velocity
of FolA activity at various DHF substrate concentrations to
establish if SCH-79797 acts competitively or non-competitively.
A Michaelis-Menten fit to the data demonstrated that 8.6 mM
SCH-79797 (the IC50) increases the Km from 32 ± 25 mM to 100
± 80 mM. These results indicate that SCH-79797 functions at
least partially as a competitive inhibitor of FolA’s activity on its
DHF substrate (Figure 4E). Likewise, the Michaelis-Menten fits
for the established FolA inhibitor trimethoprim show a very
similar effect IC50 = 15 ± 4 nM (Figure 4D), consistent with previ-
ous measurements of the tight binding between trimethoprim
and E. coli FolA (Cammarata et al., 2017).
SCH-79797 Also Disrupts Bacterial Membrane Potential
and Permeability Barrier
The similarities between SCH-79797 and trimethoprim with
respect to FolA inhibition helped confirm DHFR as a target of
SCH-79797 but were also surprising because these two com-
pounds did not generate similar profiles in our BCP analysis
(Figure 2B). One potential explanation is that SCH-79797 has
ll
additional targets that are not shared with trimethoprim. If this
was the case, we would expect that cells resistant to trimetho-
prim would still be susceptible to SCH-79797. Previous studies
demonstrated that resistance to trimethoprim can be achieved
by deleting thyA and supplementing the media with thymine
(Amyes and Smith, 1975). We confirmed that deleting thyA
from E. coli lptD4213 in the presence of excess thymine led
to trimethoprim resistance (Figure 5A). We also found that
thymine supplementation decreased the sensitivity of E. coli
lptD4213 to SCH-79797 (comparing ‘‘WT’’ of Figure 5A to Fig-
ure 1B; Table S1). The findings that reducing cellular depen-
dence on DHFR activity (by thymine supplementation) makes
cells less sensitive to SCH-79797 treatment, while increasing
reliance on DHFR activity (by dfrA CRISPRi) makes cells more
sensitive to SCH-79797, together indicate that DHFR inhibition
is a physiologically important target of SCH-79797. Neverthe-
less, comparing cells with and without thyA in the presence of
thymine showed no change in sensitivity to SCH-79797 (Fig-
ure 5A), suggesting that SCH-79797 is likely to have a second,
folate-independent MoA.
To obtain clues about the potential additional MoA of SCH-
79797, we revisited our fluorescent BCP images of E. coli
lptD4213 cells treated with SCH-79797. We observed SYTOX
Green staining in some of the bacteria (Figure 2A), suggesting
that SCH-79797 compromises the integrity of the bacterial mem-
brane. To directly quantify the effect of SCH-79797 on bacterial
membrane integrity, we used flow cytometry to measure the
membrane potential and permeability of E. coli lptD4213 in the
presence of the fluorescent dyes, DiOC2(3) and TO-PRO-3.
DiOC2(3) is a cationic dye that accumulates in the cytoplasm of
cells with an active membrane potential and shifts its fluores-
cence from red to green in these cells, providing a measure of
membrane potential (Figure 5B) (Novo et al., 1999). TO-PRO-3
is a nucleic acid stain that only accumulates in cells with compro-
mised membranes, providing an independent measure of mem-
brane permeability (Figure 5B) (Novo et al., 2000). As positive
controls, we observed the expected shifts in both DiOC2(3)
and TO-PRO-3 staining using CCCP, a membrane-decoupler
that affects membrane potential but not permeability (Novo
et al., 2000), and two compounds that disrupt both membrane
potential and permeability: nisin, a pore-forming antibacterial
peptide (Prince et al., 2016; Wiedemann et al., 2001), and poly-
myxin B, a small lipopeptide membrane destabilizer (Warren
et al., 1957) (Figure 5C). As negative controls, we confirmed
that antibiotics that do not target the membrane, including ampi-
cillin, rifampicin, and novobiocin, do not shift DiOC2(3) or TO-
PRO-3 staining (Figures S5A–S5D). After 15 min of treatment
with SCH-79797, subsequent DiOC2(3) and TO-PRO-3 staining
revealed significant defects in both membrane polarization and
permeability (Figure 5C). These effects on the membrane are
not secondary consequences of DHFR inhibition, as trimetho-
prim-treated E. coli showed no significant changes in DiOC2(3)
and TO-PRO-3 staining (Figure 5C). The membrane-targeting ef-
fect of SCH-79797 is also not species-specific, as similar results
were seen with SCH-79797-treated B. subtilis 168 (Figures S5E–
S5G). These findings indicate that independent of its ability to
inhibit DHFR activity, SCH-79797 disrupts both membrane po-
tential and permeability barrier.
Cell 181, 1518–1532, June 25, 2020 1523
 ll
Article

A B

D E

Figure 4. SCH-79797 Targets Folate Metabolism by Competitively Inhibiting Dihydrofolate Reductase

(A) A partial representation of the folate synthesis pathway. Where the E. coli and B. subtilis differ, the E. coli names are listed in parentheses.
(B) The growth of CRISPRi B. subtilis knockdown mutants (Peters et al., 2016) involved in folate synthesis relative to a DMSO-treated control after SCH-79797 or
trimethoprim treatment. OD600 of each condition 14 h after inoculation was plotted against drug concentration. Each data point represents 2 biological replicates.
Mean ± SD are shown.
(C) Metabolomic analysis of E. coli NCM3722 cells treated with 13.9 mg/mL SCH-79797 (13 MIC) or 0.15 mg/mL trimethoprim (13 MIC). Samples were taken 0, 5,
10, and 15 min after drug treatment. Folate metabolite abundance at each time point was quantified relative to the DMSO-treated control samples at the initial
time point. Each data point represents 3 independent replicates. Mean ± SD are shown.
(D and E) E. coli dihydrofolate reductase (FolA) activity is reduced in the presence of SCH-79797, IRS-10, or trimethoprim. Activity is relative to the standard
condition of 60 mM NADPH and 100 mM DHF. (D) IC50 values were derived from fits to the Hill equation for reactions performed at 60 mM NADPH and 100 mM DHF.
(E) FolA activity as a function of dihydrofolate concentration in the presence of 8.6 mM SCH-79797 (IC50), 0.065 mM IRS-10 (IC50), or 0.015 mM trimethoprim (IC50).
Kinetic parameters are derived from fits to the Michaelis-Menten model.

1524 Cell 181, 1518–1532, June 25, 2020

 ll
Article

A B

Figure 5. SCH-79797 Is Distinct from Other Dihydrofolate Reductase Inhibitors and Disrupts Membrane Integrity
(A) The growth of wild-type (WT) and DthyA E. coli lptD4213 relative to a DMSO-treated control after SCH-79797 and trimethoprim treatment. Bacterial growth
was measured for 14 h and the final OD600 of each condition was plotted against drug concentration. Each data point represents 2 biological replicates. Mean ±
SD are shown.
(B) Schematic of flow cytometry data showing the expected results for each class of polarized, depolarized, permeable, and impermeable bacteria.
(C) Flow cytometry analysis of the membrane potential and permeability of E. coli lptD4213 cells after 15 min incubation with 1% DMSO (solvent control), 5 mM
CCCP, 25 mg/mL nisin (23 MIC), 0.8 mg/mL polymyxin B (23 MIC), 12.5 mg/mL SCH-79797 (23 MIC), or 2 mg/mL trimethoprim (103 MIC). The limits for the
depolarized region were defined by comparing the values in the CCCP and solvent only controls. The limits for the permeabilized region were defined by
comparing the nisin and solvent only controls.

SCH-79797 Treatment Can Kill Bacteria in Contexts
Where Combination Therapy Fails
Having established that SCH-79797 disrupts both folate meta-
bolism and membrane integrity, we sought to determine if these
two targets can together explain how SCH-79797 kills bacteria.
To address this question, we used BCP analysis to compare the
cell morphology of bacteria treated with SCH-79797 to that of
bacteria treated with a combination of two different antibiotics,
one of which targets DHFR and one of which targets membrane
integrity. Qualitative inspection suggested that SCH-79797-
treated E. coli appeared similar to E. coli lptD4213 cells treated
with a combination of trimethoprim and nisin (Figure 6A). Quan-
tification of the images confirmed that SCH-79797 clusters with
the co-treatment of trimethoprim and nisin (Figure 6A). Co-treat-
ment with polymyxin B and trimethoprim similarly clustered with
SCH-79797, suggesting that this effect is due to membrane
perturbation and not specific to the complex MoA of nisin (Has-
per et al., 2006; Prince et al., 2016; Wiedemann et al., 2001)

Cell 181, 1518–1532, June 25, 2020 1525

 ll
Article

Figure 6. SCH-79797 Mimics Co-treatment with Folate Metabolism and Membrane Integrity Disruptors but Can Be More Effective Than Their
Combination
(A) BCP analysis of E. coli lptD4213 cells after 30 min of treatment with 1% DMSO, 6.3 mg/mL SCH-79797 (13 MIC), 2 mg/mL trimethoprim (103 MIC), 25 mg/mL
nisin (23 MIC), or the combination of 2 mg/mL trimethoprim (103 MIC) and 25 mg/mL nisin (23 MIC). Cytological profiles were clustered by the first three principal
components that account for at least 90% of the variance between samples. Cells were stained with DAPI, FM4-64, and SYTOX Green. Shown here are the
merged images of DAPI (blue) and FM4-64 (red). Scale bar, 2 mm.
(B)The viability of E. coli lptD4213 cells measured in CFU mL 1 after 2 h of treatment with 1% DMSO (solvent control), 2 mg/mL trimethoprim (103 MIC), 25 mg/mL
nisin (23 MIC), the combination of 2 mg/mL trimethoprim (103 MIC) and 25 mg/mL nisin (23 MIC), 0.8 mg/mL polymixin B (23 MIC), the combination of 2 mg/mL
trimethoprim (103 MIC) and 0.8 mg/mL polymixin B (23 MIC), or 3.1 mg/mL SCH-79797 (13 MIC). Each bar represents 3 biological replicates. Mean ± SD
are shown.
(C) Viability of S. aureus MRSA USA300 persister cells measured in CFU mL 1 after 2 h of treatment with 1% DMSO (solvent control), 63 mg/mL trimethoprim (103
MIC), 100 mg/mL nisin (23 MIC), the combination of 63 mg/mL trimethoprim (103 MIC) and 50 mg/mL nisin (23 MIC), 63 mg/mL daptomycin (23 MIC), the
combination of 63 mg/mL trimethoprim (103 MIC) and 63 mg/mL daptomycin (23 MIC), or 6.3 mg/mL SCH-79797 (13 MIC). Each bar represents 3 biological
replicates. Mean ± SD are shown.

1526 Cell 181, 1518–1532, June 25, 2020

 ll
Article

B C

D E

F G

Figure 7. Derivates of SCH-79797 Show Increased Potency and the Ability to Help Clear Infection in Mouse Vaginal N. gonorrhea Model
(A) Structures of SCH-79797, the pyrroloquinazolinediamine core lacking the side chains (IRS-10), and the pyrroloquinazolinediamine core with a biphenyl
decoration (IRS-16).
(B) The MICs of SCH-79797, IRS-10, and IRS-16 against a few selected species. For the MICs against additional strains, see Table S1.
(C) The growth of CRISPRi B. subtilis knockdown mutants involved in folate synthesis relative to a DMSO-treated control after treatment with IRS-10 or IRS-16.
Bacterial growth was measured for 14 h and the final optical density (OD600) of each condition was plotted against drug concentration. Each data point represents
2 biological replicates. Mean ± SD are shown.
(D) Flow cytometry analysis of the membrane potential and permeability of E. coli lptD4213 cells after 15 min incubation with 0.4 mg/mL IRS-10 (13 MIC) or
0.02 mg/mL IRS-16 (13 MIC).
(E) Therapeutic index of SCH-79797 and IRS-16 was calculated by dividing the MIC of each drug for the indicated mammalian cell line by its MIC against E. coli
lptD4213. The MIC of IRS-16 against PBMC was greater than the maximal drug concentrations tested.

(legend continued on next page)

Cell 181, 1518–1532, June 25, 2020 1527

 ll
(Figure S6). The fact that SCH-79797 clusters more closely to the
co-treatments than to the individual treatments with trimetho-
prim or nisin/polymyxin B reinforces the conclusion that SCH-
79797 kills bacteria by targeting both DHFR and the membrane.
There are no other antibiotics that have been shown to target
both folate metabolism and membrane integrity, indicating that
SCH-79797 represents an antibiotic with a unique MoA. This
result also explains why SCH-79797 failed to cluster with any
of the known antibiotics in our BCP analysis (Figure 2B).
Combination antibiotic therapy has been suggested as a po-
tential means of circumventing the rise of antibiotic resistance
(Tamma et al., 2012; Tyers and Wright, 2019) but it has remained
unclear whether it is better to combine multiple activities on the
same molecule. To probe this issue, we measured the synergy of
co-treatment with one antibiotic targeting dihydrofolate reduc-
tase and another targeting the membrane and compared their
combined effectiveness to that of SCH-79797. Interestingly,
when E. coli lptD4213 cells were co-treated with trimethoprim
and nisin, or co-treated with trimethoprim and polymixin B, the
two antibiotics antagonized one another’s activity, resulting in
a greater number of viable cells remaining after 2 h of co-treat-
ment (Figure 6B). MRSA USA300 persister cells (Kim et al.,
2018) are resistant to treatment with the membrane-disrupting
daptomycin (Chen et al., 2014; Taylor and Palmer, 2016). Treat-
ing MRSA USA300 persister cells with 63 mg/mL daptomycin (13
MIC) for 2 h did not reduce the number of CFUs remaining in the
culture (Figure 6C). However, SCH-79797 treatment of MRSA
robustly killed these persister cells while the combination of
trimethoprim with either nisin or daptomycin could not (Fig-
ure 6C). These results suggest that the combination of two
different antibacterial activities on the same molecular scaffold
can, at least in the case of SCH-79797, produce a more potent
antibacterial effect than co-treating with two antibiotics with
the two separate targeting activities.
The Chemical Basis of the Two MoAs of SCH-79797
SCH-79797 consists of a pyrroloquinazolinediamine core that is
substituted with an isopropylphenyl group on one side and a cy-
clopropyl moiety on the other. In order to test the function of the
pyrroloquinazolinediamine core on the antibiotic activity of SCH-
79797, we synthesized a derivative of SCH-79797 (Irresistin-10
or IRS-10) that lacks both side groups (Figure 7A). When
compared to the parent molecule SCH-79797, removing the iso-
propylphenyl and cyclopropyl groups increased the potency
against E. coli lptD4213 but decreased the potency against
B. subtilis 168, MRSA USA300, and A. baumannii AB17978 (Fig-
ure 7B; Table S1). To determine whether the pyrroloquinazoline-
diamine core of SCH-79797 is specifically involved in targeting
folate metabolism or membrane integrity, we assessed the activ-
ity of IRS-10 using the dfrA and folC CRISPRi hypersensitivity
assay and the quantitative flow cytometry membrane integrity
assay. The CRISPRi hypersensitivity assay indicated that IRS-
10 maintains the ability to inhibit folate metabolism, suggesting
(F) The stability of IRS-16 was measured following incubation with mouse liver microsomes. Each data point represents 2 biological replicates. Mean ± SD
are shown.
(G) Treatment of mice with IRS-16 (10 mg/kg, i.v., twice a day [b.i.d.]) reduces the vaginal burden of N. gonorrhoeae 24 h after inoculation. p value from one-
factor ANOVA.
1528 Cell 181, 1518–1532, June 25, 2020
Article
that the pyrroloquinazolinediamine core is sufficient to target
DHFR (Figure 7C). However, unlike SCH-79797, DiOC2(3) and
TO-PRO-3 staining showed that IRS-10 does not disrupt mem-
brane polarity or permeability (Figure 7D). We also confirmed
that IRS-10 directly inhibits the enzymatic activity of purified
E. coli FolA (Figures 4D and 4E). IRS-10 proved to be a more
potent inhibitor of FolA than SCH-79797 (IRS-10 IC50 = 65 ±
19 nM) (Figure 4D), suggesting that its increased efficacy against
E. coli in cell culture can be explained by increased activity to-
ward DHFR. Together, these findings suggest that the pyrrolo-
quinazolinediamine core of SCH-79797 targets DHFR.
We next sought to determine if the isopropylbenzene side
group is responsible for the membrane-targeting properties of
SCH-79797. We thus obtained isopropylbenzene alone (also
known as cumene) (Figure S7A) and determined its effects on
membrane integrity and folate biosynthesis. DiOC2(3) and TO-
PRO-3 staining showed that ispropylbenzene disrupts both
membrane polarity and permeability (Figure S7B). Meanwhile,
reduction of dfrA or folC levels by CRISPRi had no effect on
the sensitivity of bacteria to isopropylbenzene (Figure S7C).
These results support the conclusion that SCH-79797 is a
dual-targeting compound where the pyrroloquinazolinediamine
core specifically targets folate metabolism while the isopropyl-
benzene group specifically targets membrane integrity.
As a further test of whether the hydrophobic isopropylbenzene
chain functions to target the membrane, we generated another
small molecule, Irresistin-16 (IRS-16), in which we decorated
the pyrroloquinazolinediamine core with a biphenyl group that
is even more hydrophobic than isopropylbenzene (Figure 7A).
As predicted from the inclusion of both folate-targeting and
membrane-targeting moieties, dfrA and folC CRISPRi mutants
proved hypersensitive to IRS-16 and IRS-16 disrupted mem-
brane permeability and polarity by DiOC2(3) and TO-PRO-3
staining (Figures 7C and 7D). IRS-16 also had more potent anti-
biotic activity than IRS-10 against most bacteria tested (Fig-
ure 7B; Table S1), suggesting that targeting both membrane
integrity and folate biosynthesis is more powerful than targeting
folate metabolism alone.
The SCH-79797 Derivative IRS-16 Is Efficacious in a
Mouse Infection Model
An effective antibiotic needs to be able to target pathogenic bac-
teria without killing mammalian hosts. To determine the concen-
trations required to inhibit the growth of mammalian cells, we
treated several mammalian cell lines with both SCH-79797 and
IRS-16, the derivative with the most potent antibiotic activity.
SCH-79797 showed promising results with PBMC cells, as
they required more than 10-fold higher doses for growth inhibi-
tion than those required for killing E. coli lptD4213 (Figure 7E).
However, SCH-79797 inhibited the growth of other mammalian
cell lines, including HK-2, HEK293, and HLF, at doses compara-
ble to the doses needed to kill bacteria. In contrast, IRS-16 killed
bacteria at 100–1,000-fold lower doses than those required to

Article
affect mammalian cells in all cell lines tested (Figure 7E). In light
of this larger therapeutic window, we focused our further in vivo
analysis efforts on IRS-16.
Because IRS-16 preferentially killed bacteria in culture
models, we proceeded to characterize its in vivo effects on
mice. We determined the maximal tolerated dose (MTD) of
IRS-16 to be 15 mg/kg when administered intravenously. A phar-
macokinetic analysis revealed that at this MTD, the plasma con-
centration of IRS-16 peaked at 1.4 mg/mL with a half-life of 15.8 h
(Figures S7D and S7E). Consistent with this robust in vivo stabil-
ity, a mouse liver microsome study showed that IRS-16 is
extremely stable as compared to a control drug, verapamil
(Figure 7F).
Finally, we determined whether IRS-16 has antibiotic activity in
a mouse bacterial infection model. For this purpose, we focused
on N. gonorrhoeae as it is a Gram-negative pathogen for which
there is an acute need for new antibiotics due to widespread
resistance toward existing drugs (CDC, 2019). There is also
a well-validated mouse vaginal infection model for
N. gonorrhoeae (Jerse et al., 2002; Song et al., 2008). In bacterial
culture, IRS-16 showed robust activity toward N. gonorrhoeae,
with an MIC of 0.03 mg/mL (Table S1). Our pharmacokinetic anal-
ysis indicated that IRS-16 should persist in mice at concentra-
tions above this MIC for nearly 48 h (Figures S7D and S7E). To
test its in vivo efficacy, we inoculated the vaginal tracts of
BALB/c mice with 1.85 3 106 CFU/mouse of N. gonorrhoeae
ATCC 700825, treated with intravenous (i.v.) doses of either
10 mg/kg IRS-16 or vehicle control at 2 and 14 h post-infection,
and assayed vaginal N. gonorrhoeae CFUs at 26 h post-infec-
tion. IRS-16 significantly reduced the vaginal load of
N. gonorrhoeae (p < 0.05) as compared to the vehicle control
(Figure 7G). Consistent with its favorable therapeutic index and
pharmacokinetic profile, this result confirms that IRS-16 can
function as an effective antibiotic in an in vivo mouse gonorrhea
infection model.
DISCUSSION
Due to the rise in resistance to known antibiotics, there is an
acute need for new antibiotics with the key features of having
unique MoAs, potency toward Gram-negatives, and reduced
susceptibility to resistance. Here, we describe a promising com-
pound, SCH-79797, and its derivative, IRS-16, that are effective
in animals and address these key criteria with a unique dual-tar-
geting MoA, the ability to kill both Gram-negative and Gram-pos-
itive pathogens, and an undetectably low frequency of
resistance. We also describe a systems-level pipeline that com-
bines independent orthogonal approaches to characterize the
MoA of SCH-79797 in the absence of resistant mutants. Specif-
ically, we used BCP classification to categorize the MoA of SCH-
79797 as distinct from those of 37 known antibiotics (Figure 2B).
We then used thermal proteome profiling to identify DHFR as a
candidate binding partner of SCH-79797 and confirmed that
SCH-79797 inhibits folate metabolism through metabolomic
analysis and CRISPRi genetic hypersensitivity (Figures 3B, 4B,
and 4C). We confirmed that SCH-79797 directly inhibits DHFR
activity by acting competitively toward its DHF substrate (Fig-
ures 4D and 4E). The BCP images also alerted us to a second po-
ll
tential target for SCH-79797, the bacterial membrane. Quantita-
tive flow cytometry with dyes that report on membrane
permeability and polarity confirmed that SCH-79797 has a
folate-independent effect on bacterial membrane integrity (Fig-
ure 5C). Together, these assays constitute a pipeline that can
be used in the future to rapidly characterize antibiotic MoAs de
novo. Such a pipeline is especially important for compounds
such as SCH-79797 that are not prone to resistance and do
not mimic known MoAs. BCP, thermal proteome profiling, me-
tabolomics, CRISPRi sensitivity, and flow cytometry are all as-
says that can be performed in small volumes, such that they
can be readily scaled without the need for synthesizing large
amounts of the compound in question. The orthogonal nature
of the assays enables the independent identification of multiple
MoAs, which may help in the discovery of unique antibiotic
classes.
Both of the targets of SCH-79797 are relevant for its function
as an antibiotic. The CRISPRi and metabolomic studies
demonstrate that SCH-79797 actively disrupts folate meta-
bolism in multiple bacterial species in a manner that is rate-
limiting for growth (Figures 4B and 4C). Meanwhile, the flow cy-
tometry assay demonstrates that SCH-79797 simultaneously
disrupts membrane integrity even though folate inhibition itself
has no effect on the membrane (Figure 5C). The ability of SCH-
79797 to disrupt membrane integrity is particularly interesting
given that membrane-disruptors are often selective for either
Gram-positive or Gram-negative bacteria (Ling et al., 2015;
Taylor and Palmer, 2016; Warren et al., 1957), while SCH-
79797 proved potent against both Gram-positive pathogens
like S. aureus and E. faecalis as well as Gram-negative patho-
gens like A. baumannii, N. gonorrhoeae, and pathogenic
E. coli (Figure 1A). Host toxicity is often a concern for mem-
brane-targeting antibiotics, and while SCH-79797 was well
tolerated by some animal cells like G. mellonella wax worms
and PBMC cells, it killed other mammalian cell lines at doses
similar to those at which it functions as an antibiotic. Mean-
while, IRS-16, a derivative of SCH-79797, increased antibiotic
activity without increasing mammalian toxicity, thereby
increasing its therapeutic window >100-fold. The ability of
IRS-16 to selectively target bacteria is consistent with a recent
study of retinoid derivatives that provided proof-of-principle
that small molecules can preferentially target bacterial mem-
branes (Kim et al., 2018). Future biophysical characterization
and medicinal chemistry will help to further increase potency
and reduce toxicity.
The undetectably low frequency of resistance to SCH-79797
could result from its two distinct targets. Specifically, we were
successful in isolating resistance mutants for mimics of each
of its two individual targets, trimethoprim and nisin, but not for
SCH-79797 (Figure 1E). The average mutation rate in E. coli is
2.1 3 10 7 per gene per generation (Chen and Zhang, 2013).
If E. coli required 2 mutations to acquire resistance to
SCH-79797, the number of bacteria that would be necessary
to find a resistant mutant would be in the range of 1014. Even if
that represents an overestimate, humans are estimated to carry
roughly 4 3 1013 bacteria in total, so such low frequencies of
resistance would be unlikely to result in resistant mutants in a
clinical context.
Cell 181, 1518–1532, June 25, 2020 1529
 ll
Article

Our studies suggest that SCH-79797 is more potent than com- B Materials Availability
bination treatment with antibiotics that mimic its two activities, the B Data and Code Availability
DHFR-inhibitor trimethoprim and the membrane-disruptor nisin. B Bacterial Cytological Profiling Code
Similarly, co-treatment with trimethoprim and polymyxin B B Thermal Proteome Profiling Data
showed antagonistic interactions (Figure 6B), while MRSA B Flow Cytometry Data
persister cells were killed by SCH-79797 but not by combined B Additional Raw Data
treatment with trimethoprim and daptomycin (Figure 6C). A poten- d EXPERIMENTAL MODEL AND SUBJECT DETAILS
tial explanation for the potency of SCH-79797 is that recruiting a B Bacterial strains and growth conditions
DHFR inhibitor to the membrane could increase its effective con- B Mammalian cell lines
centration or potentiate its inactivation of DHFR by sequestering it. B Animal models
Permeabilizing the membrane could also enhance the access of d METHOD DETAILS
SCH-79797 to its cytoplasmic DHFR target. The difference be- B Minimum inhibitory concentration assays
tween SCH-79797 and the combination treatments could also B Compound library
be based on non-primary target effects such as differences in B Galleria mellonella killing assay
localized synergistic drug concentrations, drug uptake or efflux. B Colony forming units assay
Membrane-targeting molecules can act either synergistically or B S. aureus MRSA persister cell assay
antagonistically with antibiotics with different MoAs (Brochado B Serial passaging assay to evolve resistance
et al., 2018). Because trimethoprim and various membrane dis- B Bacterial cytological profiling
ruptors antagonize each other separately, but DHFR inhibition B Thermal proteome profiling
and membrane disruption synergize in the context of SCH- B Metabolomics
79797, combining antibiotic activities onto the same molecule B Dihydrofolate reductase activity assay
could present a solution for bypassing this antagonistic effect. In B Membrane potential and permeability assay
any event, our results suggest that despite the promise of combi- B Mammalian cell cytotoxicity
nation antibiotic therapies (Brochado et al., 2018; Tyers and B Mouse liver microsomal stability
Wright, 2019), an even more powerful approach could be to B Pharmacokinetic analysis
combine different targeting moieties onto the same chemical B Neisseria gonorrhea vaginal infection model
scaffold. d QUANTIFICATION AND STATISTICAL ANALYSIS
Discovering the MoA of SCH-79797 also enabled us to design B Center, spread, and statistical significance
derivatives that improve its efficacy. Our most promising deriva- B Bacterial cytological profiling
tive currently is IRS-16, in which we replaced the isopropyl- B Thermal proteome profiling
phenyl group with a biphenyl group (with the idea to increase B Flow cytometry analysis
the membrane-targeting activity) and removed the cyclopropyl B Pharmacokinetic analysis
side chain (to enhance the DHFR inhibition). IRS-16 showed
improved ability to kill bacteria, with significantly lower MICs SUPPLEMENTAL INFORMATION
than SCH-79797. More importantly, IRS-16 did not similarly
Supplemental Information can be found online at https://doi.org/10.1016/j.
enhance the growth inhibition of mammalian cells. Thus, IRS-
cell.2020.05.005.
16 exhibited a promising therapeutic index, as reducing the con-
centration necessary to kill bacteria without affecting the con- ACKNOWLEDGMENTS
centration necessary to kill mammalian cells resulted in a com-
pound that is >100-fold more potent toward bacteria than The B. subtilis CRISPR knockdown library was a kind gift from Jason M. Pe-
hosts. IRS-16 was also stable in mice and tolerated at doses ters. Flow cytometry was performed in collaboration with Christina DeCoste
(Princeton University Flow Cytometry Resource Facility [FCRF]). We appre-
significantly above the MIC for several hours. Finally, we
ciate the support and feedback from lab members in the Gitai and Shaevitz
confirmed that IRS-16 significantly reduced the burden of labs. Funding was provided in part by NIH (DP1AI124669 to Z.G., J.P.S.,
N. gonorrhoeae in a mouse vaginal infection model. B.P.B., and J.K.M. and T32 GM007388 to J.K.M. and G.M.M.), as well as
N. gonorrhoeae is a Gram-negative pathogen with some of the Princeton DFR Innovation Funds for New Ideas in Science (to J.K.M. and
highest rates of drug resistance for any pathogen. The acute J.P.S.). Additional funding provided by the National Science Foundation
need for new antibiotics to treat N. gonorrhoeae makes IRS-16 (NSF PHY-1734030 to B.P.B.) and for the FCRF by the National Cancer Insti-
tute (NCI-CCSG P30CA072720-5921). The opinions, findings, and conclu-
a particularly promising small molecule candidate for future
sions or recommendations expressed in this material contents are solely the
development. responsibility of the authors and do not necessarily represent the official views
of the NIH or the National Science Foundation.
STAR+METHODS
AUTHOR CONTRIBUTIONS

Detailed methods are provided in the online version of this paper Conceptualization, Z.G., J.K.M., M.Z.W., and H.K.; Methodology, Z.G., B.P.B.,
and include the following: M.Z.W., A.M., A.T., M.M.S., J.R., S.H.-J.L., and H.K.; Software, M.Z.W.,
J.K.M., B.P.B., A.T., and M.M.S.; Validation, J.P.S.; Formal Analysis, B.P.B.
d KEY RESOURCES TABLE and G.M.M.; Investigation, M.Z.W., J.K.M., J.P.S., G.M.M., A.M., A.T.,
d RESOURCE AVAILABILITY M.M.S., S.H.-J.L., and B.P.B.; Resources, J.P.S. and M.Z.W.; Writing – Orig-
B Lead Contact inal Draft, Z.G. and J.K.M.; Writing – Reviewing & Editing, Z.G., B.P.B.,

1530 Cell 181, 1518–1532, June 25, 2020


Article
J.P.S., H.K., and C.D.; Visualization, J.P.S., J.K.M., and B.P.B.; Supervision,
Z.G., H.K., J.R., A.T., and M.M.S.; Funding Acquisition, Z.G., J.R., A.T.,
and M.M.S.
DECLARATION OF INTERESTS
A patent application describing the use of SCH-79797 as an antibiotic, as well
as the pharmaceutical composition and use as antibiotic of derivatives is
currently pending.
Received: June 6, 2019
Revised: February 24, 2020
Accepted: May 1, 2020
Published: June 3, 2020
REFERENCES
Ahn, H.S., Foster, C., Boykow, G., Stamford, A., Manna, M., and Graziano, M.
(2000). Inhibition of cellular action of thrombin by N3-cyclopropyl-7-[[4-(1-
methylethyl)phenyl]methyl]-7H-pyrrolo[3, 2-f]quinazoline-1,3-diamine (SCH
79797), a nonpeptide thrombin receptor antagonist. Biochem. Pharmacol.
60, 1425–1434.
Amyes, S.G., and Smith, J.T. (1975). Thymineless mutants and their resistance
to trimethoprim. J. Antimicrob. Chemother. 1, 85–89.
Becher, I., Werner, T., Doce, C., Zaal, E.A., Tögel, I., Khan, C.A., Rueger, A.,
Muelbaier, M., Salzer, E., Berkers, C.R., et al. (2016). Thermal profiling reveals
phenylalanine hydroxylase as an off-target of panobinostat. Nat. Chem. Biol.
12, 908–910.
Bell-Pedersen, D., Galloway Salvo, J.L., and Belfort, M. (1991). A transcription
terminator in the thymidylate synthase (thyA) structural gene of Escherichia coli
and construction of a viable thyA:Kmr deletion. J. Bacteriol. 173, 1193–1200.
Boucher, H.W., Talbot, G.H., Bradley, J.S., Edwards, J.E., Gilbert, D., Rice,
L.B., Scheld, M., Spellberg, B., and Bartlett, J. (2009). Bad bugs, no drugs:
no ESKAPE! An update from the Infectious Diseases Society of America.
Clin. Infect. Dis. 48, 1–12.
Brochado, A.R., Telzerow, A., Bobonis, J., Banzhaf, M., Mateus, A., Selkrig, J.,
Huth, E., Bassler, S., Zamarreño Beas, J., Zietek, M., et al. (2018). Species-
specific activity of antibacterial drug combinations. Nature 559, 259–263.
Butler, M.S., Blaskovich, M.A., and Cooper, M.A. (2017). Antibiotics in the clin-
ical pipeline at the end of 2015. J. Antibiot. (Tokyo) 70, 3–24.
Cammarata, M., Thyer, R., Lombardo, M., Anderson, A., Wright, D., Ellington,
A., and Brodbelt, J.S. (2017). Characterization of trimethoprim resistant E. coli
dihydrofolate reductase mutants by mass spectrometry and inhibition by
propargyl-linked antifolates. Chem. Sci. (Camb.) 8, 4062–4072.
CDC (2019). Antibiotic Resistance Threats in the United States (U.S. Depart-
ment of Health and Human Services).
Chen, X., and Zhang, J. (2013). No gene-specific optimization of mutation rate
in Escherichia coli. Mol. Biol. Evol. 30, 1559–1562.
Chen, Y.F., Sun, T.L., Sun, Y., and Huang, H.W. (2014). Interaction of daptomy-
cin with lipid bilayers: a lipid extracting effect. Biochemistry 53, 5384–5392.
Chen, L., Ducker, G.S., Lu, W., Teng, X., and Rabinowitz, J.D. (2017). An LC-
MS chemical derivatization method for the measurement of five different one-
carbon states of cellular tetrahydrofolate. Anal. Bioanal. Chem. 409,
5955–5964.
Coates, A.R., Halls, G., and Hu, Y. (2011). Novel classes of antibiotics or more
of the same? Br. J. Pharmacol. 163, 184–194.
Culyba, M.J., Mo, C.Y., and Kohli, R.M. (2015). Targets for Combating the Evo-
lution of Acquired Antibiotic Resistance. Biochemistry 54, 3573–3582.
Davies, J. (2006). Where have All the Antibiotics Gone? Can. J. Infect. Dis.
Med. Microbiol. 17, 287–290.
Gebhardt, M.J., Gallagher, L.A., Jacobson, R.K., Usacheva, E.A., Peterson,
L.R., Zurawski, D.V., and Shuman, H.A. (2015). Joint Transcriptional Control
of Virulence and Resistance to Antibiotic and Environmental Stress in Acineto-
bacter baumannii. MBio 6, e01660-15.
ll
Gleckman, R., Blagg, N., and Joubert, D.W. (1981). Trimethoprim: mecha-
nisms of action, antimicrobial activity, bacterial resistance, pharmacokinetics,
adverse reactions, and therapeutic indications. Pharmacotherapy 1, 14–20.
Gobbetti, T., Cenac, N., Motta, J.-P., Rolland, C., Martin, L., Andrade-Gordon,
P., Steinhoff, M., Barocelli, E., and Vergnolle, N. (2012). Serine protease inhi-
bition reduces post-ischemic granulocyte recruitment in mouse intestine.
Am. J. Pathol. 180, 141–152.
Gupta, N., Liu, R., Shin, S., Sinha, R., Pogliano, J., Pogliano, K., Griffin, J.H.,
Nizet, V., and Corriden, R. (2018). SCH79797 improves outcomes in experi-
mental bacterial pneumonia by boosting neutrophil killing and direct antibiotic
activity. J. Antimicrob. Chemother. 73, 1586–1594.
Gutnick, D., Calvo, J.M., Klopotowski, T., and Ames, B.N. (1969). Compounds
which serve as the sole source of carbon or nitrogen for Salmonella typhimu-
rium LT-2. J. Bacteriol. 100, 215–219.
Hasper, H.E., Kramer, N.E., Smith, J.L., Hillman, J.D., Zachariah, C., Kuipers,
O.P., de Kruijff, B., and Breukink, E. (2006). An alternative bactericidal mech-
anism of action for lantibiotic peptides that target lipid II. Science 313,
1636–1637.
Hofer, U. (2019). The cost of antimicrobial resistance. Nat. Rev. Microbiol.
17, 3.
Imai, Y., Meyer, K.J., Iinishi, A., Favre-Godal, Q., Green, R., Manuse, S., Ca-
boni, M., Mori, M., Niles, S., Ghiglieri, M., et al. (2019). A new antibiotic selec-
tively kills Gram-negative pathogens. Nature 576, 459–464.
Jerse, A.E., Crow, E.T., Bordner, A.N., Rahman, I., Cornelissen, C.N., Moench,
T.R., and Mehrazar, K. (2002). Growth of Neisseria gonorrhoeae in the female
mouse genital tract does not require the gonococcal transferrin or hemoglobin
receptors and may be enhanced by commensal lactobacilli. Infect. Immun. 70,
2549–2558.
Karlowsky, J.A., Draghi, D.C., Jones, M.E., Thornsberry, C., Friedland, I.R.,
and Sahm, D.F. (2003). Surveillance for antimicrobial susceptibility among clin-
ical isolates of Pseudomonas aeruginosa and Acinetobacter baumannii from
hospitalized patients in the United States, 1998 to 2001. Antimicrob. Agents
Chemother. 47, 1681–1688.
Kim, W., Zhu, W., Hendricks, G.L., Van Tyne, D., Steele, A.D., Keohane, C.E.,
Fricke, N., Conery, A.L., Shen, S., Pan, W., et al. (2018). A new class of syn-
thetic retinoid antibiotics effective against bacterial persisters. Nature 556,
103–107.
King, A.C., and Wu, L. (2009). Macromolecular synthesis and membrane
perturbation assays for mechanisms of action studies of antimicrobial agents.
Curr. Protoc. Pharmacol. Chapter 13, Unit 13A.17.
Kwon, Y.K., Lu, W., Melamud, E., Khanam, N., Bognar, A., and Rabinowitz,
J.D. (2008). A domino effect in antifolate drug action in Escherichia coli. Nat.
Chem. Biol. 4, 602–608.
Kwon, Y.K., Higgins, M.B., and Rabinowitz, J.D. (2010). Antifolate-induced
depletion of intracellular glycine and purines inhibits thymineless death in
E. coli. ACS Chem. Biol. 5, 787–795.
Ling, L.L., Schneider, T., Peoples, A.J., and Spoering, A.L. (2015). A new anti-
biotic kills pathogens without detectable resistance. Nature 517, 455–459.
Mateus, A., Bobonis, J., Kurzawa, N., Stein, F., Helm, D., Hevler, J., Typas, A.,
and Savitski, M.M. (2018). Thermal proteome profiling in bacteria: probing pro-
tein state in vivo. Mol. Syst. Biol. 14, e8242.
McInnes, L., Healy, J., and Melville, J. (2018). UMAP: Uniform Manifold
Approximation and Projection for Dimension Reduction. arXiv,
arXiv:1802.03426.
McKeage, K. (2015). Finafloxacin: first global approval. Drugs 75, 687–693.
Nonejuie, P., Burkart, M., Pogliano, K., and Pogliano, J. (2013). Bacterial cyto-
logical profiling rapidly identifies the cellular pathways targeted by antibacte-
rial molecules. Proc. Natl. Acad. Sci. USA 110, 16169–16174.
Novo, D., Perlmutter, N.G., Hunt, R.H., and Shapiro, H.M. (1999). Accurate
flow cytometric membrane potential measurement in bacteria using diethylox-
acarbocyanine and a ratiometric technique. Cytometry 35, 55–63.
Novo, D.J., Perlmutter, N.G., Hunt, R.H., and Shapiro, H.M. (2000). Multipa-
rameter flow cytometric analysis of antibiotic effects on membrane potential,
Cell 181, 1518–1532, June 25, 2020 1531
E23G671
ll
membrane permeability, and bacterial counts of Staphylococcus aureus and
Micrococcus luteus. Antimicrob. Agents Chemother. 44, 827–834.
O’Neill, J. (2014). AMR Review Paper–Tackling a Crisis for the Health and
Wealth of Nations (AMR Review Paper).
Peleg, A.Y., Jara, S., Monga, D., Eliopoulos, G.M., Moellering, R.C., Jr., and
Mylonakis, E. (2009). Galleria mellonella as a model system to study Acineto-
bacter baumannii pathogenesis and therapeutics. Antimicrob. Agents Chemo-
ther. 53, 2605–2609.
Peters, J.M., Colavin, A., Shi, H., Czarny, T.L., Larson, M.H., Wong, S., Haw-
kins, J.S., Lu, C.H.S., Koo, B.-M.M., Marta, E., et al. (2016). A Comprehensive,
CRISPR-based Functional Analysis of Essential Genes in Bacteria. Cell 165,
1493–1506.
Prince, A., Sandhu, P., Ror, P., Dash, E., Sharma, S., Arakha, M., Jha, S.,
Akhter, Y., and Saleem, M. (2016). Lipid-II Independent Antimicrobial Mecha-
nism of Nisin Depends On Its Crowding And Degree Of Oligomerization. Sci.
Rep. 6, 37908.
Randall, L.B., Georgi, E., Genzel, G.H., and Schweizer, H.P. (2016). Finafloxa-
cin overcomes Burkholderia pseudomallei efflux-mediated fluoroquinolone
resistance. J. Antimicrob. Chemother. 72, 1258–1260.
Ruiz, N., Kahne, D., and Silhavy, T.J. (2006). Advances in understanding bac-
terial outer-membrane biogenesis. Nat. Rev. Microbiol. 4, 57–66.
Savitski, M.M., Reinhard, F.B.M., Franken, H., Werner, T., Savitski, M.F., Eber-
hard, D., Martinez Molina, D., Jafari, R., Dovega, R.B., Klaeger, S., et al. (2014).
Tracking cancer drugs in living cells by thermal profiling of the proteome. Sci-
ence 346, 1255784.
Song, W., Condron, S., Mocca, B.T., Veit, S.J., Hill, D., Abbas, A., and Jerse,
A.E. (2008). Local and humoral immune responses against primary and repeat
Neisseria gonorrhoeae genital tract infections of 17beta-estradiol-treated
mice. Vaccine 26, 5741–5751.
Soupene, E., van Heeswijk, W.C., Plumbridge, J., Stewart, V., Bertenthal, D.,
Lee, H., Prasad, G., Paliy, O., Charernnoppakul, P., and Kustu, S. (2003). Phys-
1532 Cell 181, 1518–1532, June 25, 2020
Article
iological studies of Escherichia coli strain MG1655: growth defects and
apparent cross-regulation of gene expression. J. Bacteriol. 185, 5611–5626.
Strande, J.L., Hsu, A., Su, J., Fu, X., Gross, G.J., and Baker, J.E. (2007). SCH
79797, a selective PAR1 antagonist, limits myocardial ischemia/reperfusion
injury in rat hearts. Basic Res. Cardiol. 102, 350–358.
Tamma, P.D., Cosgrove, S.E., and Maragakis, L.L. (2012). Combination ther-
apy for treatment of infections with gram-negative bacteria. Clin. Microbiol.
Rev. 25, 450–470.
Taylor, S.D., and Palmer, M. (2016). The action mechanism of daptomycin.
Bioorg. Med. Chem. 24, 6253–6268.
Tenover, F.C., and Goering, R.V. (2009). Methicillin-resistant Staphylococcus
aureus strain USA300: origin and epidemiology. J. Antimicrob. Chemother. 64,
441–446.
Tyers, M., and Wright, G.D. (2019). Drug combinations: a strategy to extend
the life of antibiotics in the 21st century. Nat. Rev. Microbiol. 17, 141–155.
Ursell, T., Lee, T.K., Shiomi, D., Shi, H., Tropini, C., Monds, R.D., Colavin, A.,
Billings, G., Bhaya-Grossman, I., Broxton, M., et al. (2017). Rapid, precise
quantification of bacterial cellular dimensions across a genomic-scale
knockout library. BMC Biol. 15, 17.
Viehman, J.A., Nguyen, M.H., and Doi, Y. (2014). Treatment options for carba-
penem-resistant and extensively drug-resistant Acinetobacter baumannii in-
fections. Drugs 74, 1315–1333.
Warren, G.H., Gray, J., and Yurchenco, J.A. (1957). Effect of polymyxin on the
lysis of Neisseria catarrhalis by lysozyme. J. Bacteriol. 74, 788–793.
Werner, T., Sweetman, G., Savitski, M.F., Mathieson, T., Bantscheff, M., and
Savitski, M.M. (2014). Ion coalescence of neutron encoded TMT 10-plex re-
porter ions. Anal. Chem. 86, 3594–3601.
Wiedemann, I., Breukink, E., van Kraaij, C., Kuipers, O.P., Bierbaum, G., de
Kruijff, B., and Sahl, H.-G. (2001). Specific binding of nisin to the peptidoglycan
precursor lipid II combines pore formation and inhibition of cell wall biosyn-
thesis for potent antibiotic activity. J. Biol. Chem. 276, 1772–1779.
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Bacterial and Virus Strains
E. coli lptD4213 DthyA This paper ZG1599
B. subtilis 168 CRISPRi Library Jason Peters and Carol Gross Peters et al., 2016
Detailed bacterial strain information, N/A N/A
including strain identifiers, growth media
and temperature, is provided in Table S1
Chemicals, Peptides, and Recombinant Proteins
SCH-7979 Tocris Biosciences 1592
Ampicillin MP Biomedicals 190148
Cumene Sigma PHR1210
Daptomycin TCI D4229
Gentamicin sulfate Sigma G1914
Meropenem trihydrate TCI M2279
Nisin MP Biomedicals 155839
Novobiocin sodium salt MP Biomedicals 0215595705
Polymyxin B sulfate salt Sigma P1004
Purified E. coli dihydrofolate Genscript N/A
reductase (FolA)
DAPI ThermoFisher D1306
SYTOX Green Dead Cell Stain ThermoFisher S34860
FM4-64 Dye ThermoFisher T13320
Critical Commercial Assays
BacLight Bacterial Membrane Potential kit ThermoFisher B34950
Colorimetric Dihydrofolate Reductase Sigma CS0340
Assay Kit
Deposited Data
Proteomics ProteomeXchange Consortium PXD013673
Flow cytometry FlowRespository FR-FCM-Z2JD
Raw images and Metabolomics Data This Paper; DataSpace at Princeton N/A
University
Experimental Models: Cell Lines
HEK293 ATCC CRL-1573
HK2 ATCC CRL-2190
HLF Cell Applications 506K-05a
PBMC TPCS PB010C
Experimental Models: Organisms/Strains
Male CD-1 mice Pharmaron N/A
Female ovariectomized BALB/c mice Pharmacology Discovery Services Taiwan N/A
Galleria mellonella PetCo 2336624
Software and Algorithms
MATLAB MathWorks https://www.mathworks.com/products/
matlab.html
Fiji Eliceiri/LOCI group at University of https://imagej.net/Fiji
Washington
R Studio R Studio https://rstudio.com/products/rstudio/
(Continued on next page)

Cell 181, 1518–1532.e1–e6, June 25, 2020 e1

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
FlowJo FlowJo, LLC https://www.flowjo.com/solutions/flowjo/
downloads
Other
Synergy HT microplate reader BioTek N/A
InfiniteM200 Pro microplate reader Tecan N/A
27G x 0.5 inch needle BD Biosciences 305109
QuantaMaster 40 Spectrophotometer HORIBA Instruments N/A

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Zemer
Gitai (zgitai@princeton.edu).

Materials Availability
The unique strain (E. coli lptD4213 DthyA) generated in this study are available from the Lead Contact.

Data and Code Availability

Other than specific datasets and analysis code listed below, the published article includes all datasets generated or analyzed during
this study.

Bacterial Cytological Profiling Code

The BCP analysis code is available under a BSD 3-clause license at https://github.com/PrincetonUniversity/gitai-bacterialAutopsy
and archived at https://doi.org/10.5281/zenodo.3758582.

Thermal Proteome Profiling Data

The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository
with the dataset identifier PRIDE:PXD013673.

Flow Cytometry Data

Flow cytometry data has been deposited to FlowRepository with the dataset identifier FR-FCM-Z2JD https://flowrepository.org/id/
FR-FCM-Z2JD.

Additional Raw Data

The raw microscopy images used for bacterial cytological profiling and raw metabolomics data are available at Princeton DataSpace
https://dataspace.princeton.edu/jspui/handle/88435/dsp01cr56n3903.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Bacterial strains and growth conditions

Detailed bacterial strain information, including growth media and temperature, is provided in Table S1. Where listed, growth media
were prepared according manufacturer recommendations: LB Broth and LB Broth supplemented with 0.3 mM thymine (BD Biosci-
ences 244610, Alfa Aesar A15879), Brain Heart Infusion (BD Biosciences 237500), Tryptic Soy Broth (BD Biosciences 211825),
Nutrient Broth (BD Biosciences 234000), and Gutnick Minimal Media (1.0 g/L K2SO4, 13.5 g/L K2HPO4, 4.7 g/L KH2PO4, 0.1 g/L
of MgSO4-7H2O, 10 mM NH4Cl as a nitrogen source, and 0.4% w/v glucose as a carbon source) (Gutnick et al., 1969), Terrific Broth
(BD Biosciences 243820).
E. coli lptD4213 DthyA was constructed by moving DthyA::Kmr (Bell-Pedersen et al., 1991) into E. coli lptD4213 by P1 transduction.

Mammalian cell lines

Detailed cell line information is provided in Table S1. Mammalian cell line toxicity assays were performed by Pharmaron, Inc.
(Beijing, ROC).

e2 Cell 181, 1518–1532.e1–e6, June 25, 2020

 ll
Article

Animal models
Pharmacokinetics determination
Care and handling of male CD-1 mice approximately 6-8 weeks old conformed to institutional animal care and use policies as carried
out at Pharmaron, Inc. (Beijing, ROC).
Neisseria gonorrhoeae infection model
Care and handling of 5-week old ovariectomized BALB/c mice conformed to institutional animal care and use policies as carried out
at Pharmacology Discovery Services Taiwan, Ltd. (New Taipei City, TW).

METHOD DETAILS

Minimum inhibitory concentration assays

The minimum inhibitory concentration (MIC) of each antibiotic was defined as the lowest concentration of antibiotic that resulted in no
visible growth. MICs were measured by diluting overnight cultures 1:150 and then adding 2-fold dilutions of each antibiotic in 96-well
plates and grown with shaking at 37 C. Cell growth was monitored by measuring the OD600. Assays performed by Pharmacology
Discovery Services (New Taipei City, TW) and Pharmaron Inc. (Beijing, ROC) are indicated in Table S1. For MIC measurements
done in-house, OD600 readings were measured by either a BioTek Synergy HT (Winooski, VT) or Tecan InfiniteM200 Pro (Männedorf,
CH) microplate reader. In locations where drug concentrations are listed, they are included both as a relative fold change compared
to the MIC for that bacterial strain and drug combination (1X MIC, 2X MIC, etc) as well as an absolute concentration (1 mg/mL, 2 mg/
mL, etc).

Compound library
Compounds were sourced from commercial vendors: MicrosourceDiversity, Aldrich, Selleckchem, Chiromics, and Chembridge.
Each compound dissolved in DMSO at 50 mM and then was screened for antibiotic activity against E. coli lptD4213 growing in Terrific
Broth. After normalizing for plate-to-plate variation, an OD600 of half the median plate OD600 was used as a cutoff, below which any
compound was assumed to have inhibited the growth of E. coli lptD4213 and above which compounds were assumed to be ineffec-
tive. Compounds that either had not been previously identified as antibiotics or had unknown or ambiguous mechanisms of action
were chosen for further investigation and their MIC’s were measured using the microdilution method described above.

Galleria mellonella killing assay

All Galleria mellonella larvae were obtained from Vita-Bugsª, distributed through PetCoª (2336624, San Diego, CA), and kept in a
20 C chamber. All injections were administered using a sterile 1 mL syringe with a 27G x 0.5 in needle (BD 305109) attached to a
syringe pump (KD Scientific, 78-8210) delivered at a rate of 250 mL/min into the fourth leg of the worm. Prior to injection, the site
was sterilized with ethanol. A. baumannii AB17978 (105 CFU/larva) and drug dissolved in DMSO (SCH-79797 at 66.6 mg/larva,
gentamicin at 66.6 mg/larva, merapenem at 66.6 mg/larva) were pre-mixed prior to injection. The viability of each injected larva
was determined by prodding each larva with a dowel and observing whether there was subsequent movement.

Colony forming units assay

Overnight E. coli lptD4213 or S. aureus MRSA USA300 cultures were diluted 1:100 in fresh media and grown to mid exponential phase
(OD600 = 0.4-0.6). Each culture was then diluted 1:10 into fresh media and then treated with the indicated concentration of each anti-
biotic. Each time point was taken by removing 150 mL and performing 10-fold serial dilutions. Six dilutions of each condition were then
plated in the absence of antibiotic and grown at 37 C overnight. Colony forming units (CFU’s) were measured by counting the result-
ing number of colonies the next day.

S. aureus MRSA persister cell assay

Stationary phase S. aureus cells have the same antibiotic tolerant properties of persister cells (Kim et al., 2018). Thus, overnight
MRSA USA300 cultures were used to measure the effectiveness of SCH-79797 against persister cells. An overnight S. aureus
MRSA USA300 culture was diluted 1:100 in PBS and then then treated with the indicated concentration of each antibiotic. After a
2 hour incubation with shaking at 37 C, 150 mL from each treatment condition was taken and 10-fold serial dilutions were performed.
Six dilutions of each condition were then plated in the absence of antibiotic and grown at 37 C overnight. CFU’s were measured by
counting the resulting number of colonies the next day.

Serial passaging assay to evolve resistance

MICs of two biological replicates of S. aureus MRSA USA300 or A. baumannii AB17978 were obtained against the indicated antibi-
otics. Cells that grew in 0.5X MIC of the indicated antibiotics were subsequently cultured in the absence of antibiotics, and the MIC
was then remeasured. This is a single passage and was repeated for 5-30 passages. Cells from each passage were stored as a frozen
stock. Resistance was confirmed by comparing the MICs of resistant mutants against S. aureus MRSA USA300 or A. baumannii
AB17978 that had not been previously exposed to the indicated antibiotics.

Cell 181, 1518–1532.e1–e6, June 25, 2020 e3

 ll
Article

Bacterial cytological profiling

Compound library
Following the protocol of Nonejuie et al. (2013), overnight E. coli lptD4213 cultures were diluted 1:100 in LB and grown at 30 C for
90 minutes on a roller drum. Each culture was then treated with 5X the MIC and incubated at 30 C with slight agitation. Following
antibiotic treatment, cells were stained with 0.5 mM SYTOX Green, 1 mg/mL FM4-64, and 2 mg/mL DAPI for incubated for 10 minutes.
Each stained culture was then centrifuged at 3220 rcf. for 40 s and resuspended in 1/10 volume of LB. Cells spotted onto a 1.2%
agarose pad in 20% LB medium for imaging. Images were collected on a Nikon 90i upright microscope equipped with a 100X 1.4
N.A. objective (Nikon Instruments Inc., Melville, NY) and a RoleraXR (Photometrics, Tucson, AZ) camera. Microscope control and
image acquisition were performed in NIS Elements.
Dual-treatment
In dual-treatment experiments (Figures 6A and S6), overnight E. coli lptD4213 cultures were diluted 1:100 and grown to early-mid
exponential phase (OD600 = 0.4-0.6). Each culture was then diluted 1:10 into fresh LB and treated with the desired concentration
of antibiotic for 10 minutes. Following antibiotic treatment, cells were stained with 0.5 mM SYTOX Green, 1 mg/mL FM4-64, and
2 mg/mL DAPI. Each stained culture was then spotted onto a 1.5% agar pad supplemented with casamino acids and glucose in
M63 (15 mM (NH4)2SO4, 100 mM KH2PO4, 1.7 mM FeSO4, 0.5% glucose, 0.2% casamino acids, 1 mM MgSO4). Images were
collected on either Nikon 90i upright microscope equipped with a 100X 1.4 N.A. objective (Nikon) or a Nikon Ti-E inverted microscope
equipped with a 100X 1.4 NA objective. Both microscopes utilize an Orca Flash4 camera (Hamamatsu, Bridgewater, NJ). Microscope
control and image acquisition were performed in NIS Elements (Nikon).
Image analysis
Following imaging, the E. coli cells were segmented (Ursell et al., 2017) and single-cell features were extracted using custom MAT-
LAB code (Mathworks, Natwick, MA). Clustering was performed using the single-linkage method with MANOVA (Figure 2, MATLAB
functions manova1, manovacluster). Before applying the UMAP dimensionality reduction technique (McInnes et al., 2018), each of
the 14 cytological features was normalized into a z-score by subtracting off the mean and dividing by the standard deviation of that
feature (Figure S2). Principal component analysis was performed using the prcomp function in R and clustering was performed using
the single-linkage method (Figure 6).

Thermal proteome profiling

Thermal proteome profiling experiments were performed following Mateus et al. (2018). Whole cell samples were treated with 0.6,
1.1, 2.2, 4.4 mg/mL SCH-79797 or 0.1, 0.5, 2.3, 11.6 mg/mL trimethoprim. Cell lysate samples were treated with 0.4, 1.8, 8.9,
44.4 mg/mL SCH-79797 or 0.1, 0.5, 2.3, 11.6 mg/mL trimethoprim. After heat treatment, the soluble fraction was collected, digested
with trypsin and peptides were labeled with tandem mass tags (Werner et al., 2014). Samples were subjected to two-dimensional
liquid chromatography and analyzed on a Q Exactive Plus mass spectrometer (Thermo Fisher Scientific). While the data collected
is proteome wide, three different cutoffs were used to define three classes of potential targets. The color of the point indicates
the signal maximal effect size across all temperatures and the largest change in abundance across all concentrations was selected
for continued analysis (Dots, Figure 3.) A mild effect indicates at least one temperature had a change in abundance of at least 25% in
both whole cell and cell lysate treatments (Triangles, Figure 3) Proteins that had a change in abundance of at least 25% at three or
more temperatures (Squares, Figure 3). To be considered a consistent effect, the change in abundance of the protein had to show the
same sign at least 90% of the time and have an effect size of at least 1 (2-fold) in either whole cells or cell lysates.

Metabolomics
Overnight E. coli NCM3722 cultures were grown and diluted 1:100 in Gutnick Minimal Media and grown to early-mid exponential
phase (OD600 = 0.4-0.6). Cultures were treated with either 13.9 mg/mL SCH-79797 (1X MIC) or 0.15 mg/mL trimethoprim (1X MIC)
for 15 minutes. Folates were extracted by vacuum filtering 15 mL of treated cells using 0.45 mm HNWP Millipore nylon membranes
and immediately placing filters into an ice-cold quenching solution containing 40:40:20 Methanol:acetonitrile:25 mM NH4OAc + 0.1%
sodium Ascorbic in HPLC grade water. The resulting solution was then centrifuged at 16,000 3 g for 1.5 min at 4 C and the super-
natant saved for mass spectrometry analysis. Mass spectrometry analysis was performed as described in Chen et al. (2017).

Dihydrofolate reductase activity assay

E. coli dihydrofolate reductase (FolA) was purified by Genscript (Piscataway, NJ). The enzymatic activity of FolA was measured using
a colorimetric dihydrofolate reductase assay kit (Sigma CS0340) using a QuantaMaster 40 Spectrophotometer (Photon Technology
International Inc., Edison, NJ). The kit measures FolA activity by monitoring the change in sample absorbance at 340 nm due to FolA-
dependent NADPH consumption. For each condition, 1 mL purified enzyme in storage buffer (1X PBS, 10% glycerol, pH 7.4) was
thawed on ice and added to 1000 mL 1X assay buffer immediately before assaying. After mixing the reaction components together,
the 100 mL reaction volume was pipetted into a 1 mM pathlength quartz cuvette and the transmitted light intensity at 340 nm was
measured for 100 s at 1 kHz sampling. These high frequency readings were downsampled by averaging to 1 Hz and the activity
of each sample was calculated from the slope (b) a linear regression of the log transformed intensity measurements using MATLAB.
Because the enzyme was only moderately stable in assay buffer (half-time 20 minutes), all enzymatic measurements were

e4 Cell 181, 1518–1532.e1–e6, June 25, 2020

 ll
Article

normalized to a standard condition (60 mM NADPH and 100 mM DHF) that was measured immediately before the sample of interest.
The relative activity was calculated from (bsample – bnoEnzyme)/(bstandard – bnoEnzyme).

Membrane potential and permeability assay

Overnight E. coli lptD4213 and B. subtilis 168 cultures were diluted 1:100 and grown to early-mid exponential phase (OD600 = 0.4-0.6)
at 37 C. Each culture was then diluted 1:10 into PBS and treated with the desired concentration of antibiotic for 15 minutes. Cells
were then stained with the BacLight Bacterial Membrane Potential kit (ThermoFisher B34950). This kit uses DiOC2(3) to measure
a cell’s membrane potential as a ratio of green (488 nm ex, 525/50 nm em) to red (488 nm ex, 610/20 nm em) (Novo et al., 1999).
Membrane integrity was measured by staining cells with TO-PRO-3, a dye that is excluded from cells with an intact membrane
(640 nm ex, 670/30 nm em). The LSRII flow cytometer (BD Biosciences) at the Flow Cytometry Resource Facility, Princeton Univer-
sity, was used to measure the fluorescent intensities of both dyes in response to antibiotic treatment. 100,000 events were recorded
for each data file. Data was analyzed using FlowJo v10 software (FlowJo LLC, Ashland, OR).

Mammalian cell cytotoxicity

In a white, opaque, 384-well plate, either HEK293 (500 cells/well), HK-2 (500 cells/well), HLF cells (500 cells/well) or PBMC (3000
cells/well) cells were seeded overnight. After 24 hours, DMSO or compounds were added in 3-fold dilutions and incubated for 72
hours. CyQUANT Detection Reagent was added in equal volume and incubated for 1 hour. Fluorescence was read from the bottom
of plate with standard green filter set (508/527 nm ex). Cell toxicity was evaluated by Pharmaron, Inc. (Beijing, ROC).

Mouse liver microsomal stability

The metabolic stability of IRS-16 was tested in mouse liver microsomes with NADPH and UDPGA over a period of 1 hour. IRS-16
(2 mM) and Verapamil (2 mM, positive control) were added to 0.5 mg/mL microsomes. Aliquots of 50 mL were taken from the reaction
solution at 0.5, 5, 15, 30, 45 and 60 minutes. The reaction was stopped by the addition of 5 volumes of cold acetonitrile with IS (100 nM
alprazolam, 200 nM caffeine, and 100 nM tolbutamide). Samples were centrifuged at 3,220xg for 40 minutes. Aliquots of 100 mL of the
supernatant were mixed with 100 mL of ultra-pure H2O and then used for LC-MS/MS analysis. Mouse liver microsomal stability was
evaluated by Pharmaron, Inc. (Beijing, ROC).

Pharmacokinetic analysis
6-8 week old male CD-1 mice were injected intravenously with a single dose of IRS-16 at 15 mg/kg in 5% DMSO + 95% (20% HP-
b-CD in water, W/V) and showed no adverse effects. Plasma measurements were averaged from 2 mice at the indicated time points
following administration. Concentration was determined by LC/MS. Pharmacokinetic assay was performed by Pharmaron, Inc. (Bei-
jing, ROC).

Neisseria gonorrhea vaginal infection model

Groups of 5 ovariectomized BALB/c mice 5 weeks of age were used. Animals were subcutaneously injected with a water soluble form
of estradiol at 0.23 mg/mouse on day 2, 0, 2, 4 and 6.Starting from day 2, and following inoculation to the study end, mice were
dosed twice daily with streptomycin (1.2 mg/mouse) and vancomycin (0.6 mg/mouse) by IP injection to minimize the vaginal flora.
Trimethoprim sulfate (0.04 g/100 mL) was administered in drinking water. At day 0, mice were anesthetized by ketamine (100 mg/
kg) and xylazine (10 mg/kg) through IP injection, and then inoculated intravaginally with 1.0-2.0 3 106 CFU of N. gonorrhoeae
(FA1090, ATCC 700825). Before inoculation, the mouse vagina was first rinsed with 30 mL of 50 mM HEPES (pH 7.4) followed by inoc-
ulation with 20 mL gonococci suspension in PBS containing 0.5 mM CaCl2 and 1 mM MgCl2. IRS-16 or vehicle were administered 2 h
and 12 h after inoculation. Mice are euthanized with CO2 asphyxiation after 2 h and 24 h. Lavage was performed with 400 mL GC broth
containing 0.05% saponin to recover vaginal bacteria. The bacterial suspensions were plated onto chocolate agar to determine the
N. gonorrhoeae counts. N. gonorrhoeae vaginal infection model was performed by Pharmacology Discovery Services (New Taipei
City, TW).

QUANTIFICATION AND STATISTICAL ANALYSIS

Center, spread, and statistical significance

For all assays except the Neisseria gonorrhea vaginal infection model, we use the arithmetic mean and standard deviation across
multiple biological replicates as our measures of center and spread. The particular numbers of replicates for each experiment
type are included in the respective figure legends. For the Neisseria gonorrhea vaginal infection model, we display median, Q1,
Q3, and the values for each of the five individual mice in each sample cohort are (Figure 7G).
Statistical significance for Galleria mellonella killing assays were determined from Mantel-Cox test using Prism 8.1.2 (GraphPad,
San Diego, CA). Each of three independent cohorts started with twelve animals. p values can be found in Table S2. For the Neisseria
gonorrhea vaginal infection model, significance (p < 0.05) was determined from a one-factor ANOVA.

Cell 181, 1518–1532.e1–e6, June 25, 2020 e5

 ll
Article

Bacterial cytological profiling

Hierarchical-clustering of the cytological features was performed using the single-linkage method with MANOVA (Figure 2, MATLAB
functions manova1, manovacluster). Before applying the UMAP dimensionality reduction technique (McInnes et al., 2018), each of
the 14 cytological features was normalized into a z-score by subtracting off the mean and dividing by the standard deviation of that
feature (Figure S2). Principal component analysis was performed using the prcomp function in R and clustering was performed using
the single-linkage method (Figure 6). Details of the dimensionality reduction technique and hierarchical clustering method are
included in the respective figure legends.

Thermal proteome profiling

While the data collected is proteome wide, three different cutoffs were used to define three classes of potential targets. The color of
the point indicates the signal maximal effect size across all temperatures (nTemp = 10) and the largest change in abundance across
all concentrations (nConc = 4) was selected for continued analysis (Dots, Figure 3.) A mild effect indicates at least one temperature
had a change in abundance of at least 25% in both whole cell and cell lysate treatments (Triangles, Figure 3) Proteins that had a
change in abundance of at least 25% at three or more temperatures (Squares, Figure 3). To be considered a consistent effect,
the change in abundance of the protein had to show the same sign at least 90% of the time and have an effect size of at least 1
(2-fold) in either whole cells or cell lysates. The details for these symbols and cutoffs are included in the respective figure legend.

Flow cytometry analysis

100,000 events were recorded for each flow cytometry sample and analyzed using FloJo v10. Gates used to quantify the fraction of
events with depolarized or permeabilized membranes were determined from control samples treated with CCCP or nisin respec-
tively. These details are included in the respective figure legends (Figures 5C, 7D, S5, and S7B).

Pharmacokinetic analysis
Plasma measurements were averaged from 2 mice at the indicated time points following administration. After the rapid initial
approach to pseudoequilibrium, filled data symbols were used as the input for terminal half-life determination (Figure S7D). Pharma-
cokinetic parameters estimated of a non-compartmental model of IRS-16 serum levels (Figure S7D). These details are included in the
respective figure legends.

e6 Cell 181, 1518–1532.e1–e6, June 25, 2020

 ll
Article

Supplemental Figures

Figure S1. SCH-79797 Is Bactericidal against Staphylococcus aureus, Exhibits Undetectably Low Rates of Resistance, and Is an Effective
Antibiotic in an Infection Model of Galleria mellonella by Acinetobacter baumannii, Related to Figure 1
A. Colony forming units (CFU mL-1) after 3-hour treatment of S. aureus MRSA USA300 with solvent only (1% DMSO), and 6.3 mg/mL SCH-79797 (1X MIC), or
4.0 mg/mL novobiocin (5X MIC). Each data point represents 3 independent samples and 3 technical replicates. Mean ± SD are shown. B. Fold increase in
resistance of S. aureus MRSA USA300 to SCH-79797, novobiocin, trimethoprim, or nisin after 25 days of serial passaging in each drug. C. Fold increase in
resistance of A. baumannii AB17978 to SCH-79797 and gentamicin after 5 days of serial passaging in each drug. D. Fold increase in the susceptibility of S. aureus
MRSA USA300 mutants to the indicated antibiotics. Trimethoprim and nisin resistant mutants were obtained from serial passaging in respective antibiotics. E-F.
The percent survival of non-infected G. mellonella wax worms after treatment with 2 mL/larva of 100% DMSO, 67 mg/larva SCH-79797, 67 mg/larva gentamicin,
(legend continued on next page)
 ll
Article

67 mg/larva rifampicin, or 67 mg/larva meropenem. Data in (E) represents a typical cohort (n = 12) from a biological triplicate and the pooled results are presented in
(F). p values are determined from a Mantel-Cox test using Prism (n.s., pR0.05; *p < 0.05). G. The percent survival of G. mellonella wax worms infected with
A. baumannii (AB 17978) and concomitantly treated with 67 mg/larva SCH-79797, 67 mg/larva gentamicin, 67 mg/larva rifampicin, or 67 mg/larva meropenem. Data
represents the pooled results from a biological triplicate. H. The relative survival of drug-treated of G. mellonella wax worms infected with A. baumannii (AB 17978)
and treated concomitantly with antibiotics relative to larvae treated with antibiotics only without infection. Data represents the pooled results from a biological
triplicate.
 ll
Article

Figure S2. Bacterial Cytological Profiling of SCH-79797 and Antibiotics with Known Mechanisms of Action, Related to Figure 2
Dimensionality reduction of all treatments using umap were replotted using two color channels, one for only SCH-79797 treatments and one for the remaining
conditions. Density smoothed with a s = 0.1 width Gaussian kernel.
 ll
Article

Figure S3. Thermal Stability of Dihydrofolate Reductase Increases after SCH-79797 and Trimethoprim Treatment, Related to Figure 3
A-B. The relative thermal stability of E. coli dihydrofolate reductase (FolA) after treatment of whole cell and cell lysate samples with (A) SCH-79797 and (B)
trimethoprim. Changes in thermal stability were determined by measuring changes in the abundance of FolA across 10 different temperatures ranging from 42-
72 C and 4 drug concentrations and a vehicle control.
 ll
Article

Figure S4. CRISPRi Mutants Not Involved in Folate Metabolism Are Not Sensitized to SCH-79797, Related to Figure 4
A. The growth of CRISPRi B. subtilis knockdown mutants relative to a DMSO-treated control after SCH-79797 treatment. Bacterial growth was measured for 14 h
and the final optical density (OD600) of each condition was plotted against drug concentration. Each data point represents 2 independent replicates. Mean ± SD
are shown.
 ll
Article

Figure S5. Treatment with Ampicillin, Rifampicin, and Novobiocin Does Not Disrupt Membrane Integrity while SCH-79797 Disrupts Bacillus
subtilis 168 Membrane Integrity, Related to Figure 5
A-D. Flow cytometry analysis of the membrane potential and permeability of E. coli lptD4213 cells after 15 minute incubation with (A) 0.06 mg/mL ampicillin (2X
MIC), (B) 0.002 mg/mL rifampicin (2X MIC), (C) 3.1 mg/mL SCH-79797 (1X MIC), or (D) 0.12mg/mL novobiocin (2X MIC). The limits for the depolarized region were
defined by comparing the values in the CCCP and solvent only controls. The limits for the permeabilized region were defined by comparing the nisin and solvent
only controls. E-G. Flow cytometry analysis of the membrane potential and permeability of B. subtilis 168 cells after 15 minute incubation with (E) 1% DMSO, (F)
3.1 mg/mL SCH-79797 (1X MIC), or (G) 6.2 mg/mL SCH-79797 (2X MIC). The limits for the regions were defined from the solvent only controls.
 ll
Article

Figure S6. SCH-79797 Mimics Co-treatment with Folate Metabolism and Alternate Membrane Integrity Disruptors, Related to Figure 6
BCP analysis of E. coli lptD4213 cells after 30 minutes of treatment with 1% DMSO, 6.3 mg/mL SCH-79797 (1X MIC), 2 mg/mL trimethoprim (10X MIC), 0.8 mg/mL
polymyxin B (2X MIC), or the combination of 2 mg/mL trimethoprim (10X MIC) and 0.8 mg/mL polymyxin B (2X MIC). Cytological profiles were clustered by the first
three principal components that account for at least 90% of the variance between samples. Cells were stained with DAPI, FM4-64, and SYTOX Green. Shown
here are the merged images of DAPI (blue) and FM4-64 (red). Scale bar is 2 mm.
 ll
Article

Figure S7. Cumene (Isopropylbenzene) Disrupts Membrane Integrity with No Additional Sensitivity in Folate-Metabolism Mutants and
Pharmacokinetic Analysis of IRS-16 Stability, Related to Figure 7
(A) Structure of cumene. (B) Flow cytometry analysis of the membrane potential and permeability of E. coli lptD4213 cells after 15 minute incubation with
17000 mg/mL cumene (1X MIC). The limits for the depolarized region were defined by comparing the values in the CCCP and solvent only controls. The limits for
the permeabilized region were defined by comparing the nisin and solvent only controls. (C) The growth of CRISPRi B. subtilis knockdown mutants relative to a
DMSO-treated control after cumene treatment. Bacterial growth was measured for 14 h and the final optical density (OD600) of each condition was plotted against
drug concentration. (D) Plasma concentrations of IRS-16 were measured following a single 15000 mg/kg IV dose. After the rapid initial approach to pseudoe-
quilibrium, filled data symbols were used as the input for terminal half-life determination. (E) Pharmacokinetic parameters estimated of a non-compartmental
model of IRS-16 serum levels.
 Article

Interpersonal Gut Microbiome Variation Drives

Susceptibility and Resistance to Cholera Infection
Graphical Abstract Authors
Salma Alavi, Jonathan D. Mitchell,
Jennifer Y. Cho, Rui Liu, John C. Macbeth,
Ansel Hsiao

Correspondence
ansel.hsiao@ucr.edu

In Brief
Differences in the gut microbiome
between individuals determine resistance
to cholera infection through the effects on
the activity of a bile salt enzyme.

Highlights
d Interpersonal human gut microbiome variation confers
variable infection resistance

d Microbiome-dependent infection resistance can be restored

through co-transplantation

d Colonization resistance is mediated through the bile salt

hydrolase enzyme activity

d Bile salt hydrolase abundance in human microbiomes

correlates to final infection

Alavi et al., 2020, Cell 181, 1533–1546

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.036 ll
 ll

Article
Interpersonal Gut Microbiome Variation Drives
Susceptibility and Resistance to Cholera Infection
Salma Alavi,1,5 Jonathan D. Mitchell,1,5 Jennifer Y. Cho,1,2 Rui Liu,1,3 John C. Macbeth,1,4 and Ansel Hsiao1,6,*
1Department of Microbiology and Plant Pathology, University of California, Riverside, Riverside, CA, USA
2Department of Biochemistry, University of California, Riverside, Riverside, CA, USA
3Graduate Program in Genetics, Genomics, and Bioinformatics, University of California, Riverside, Riverside, CA, USA
4Division of Biomedical Sciences, School of Medicine, University of California, Riverside, Riverside, CA, USA
5These authors contributed equally
6Lead Contact

*Correspondence: ansel.hsiao@ucr.edu
https://doi.org/10.1016/j.cell.2020.05.036

SUMMARY

The gut microbiome is the resident microbial community of the gastrointestinal tract. This community is high-
ly diverse, but how microbial diversity confers resistance or susceptibility to intestinal pathogens is poorly
understood. Using transplantation of human microbiomes into several animal models of infection, we
show that key microbiome species shape the chemical environment of the gut through the activity of the
enzyme bile salt hydrolase. The activity of this enzyme reduced colonization by the major human diarrheal
pathogen Vibrio cholerae by degrading the bile salt taurocholate that activates the expression of virulence
genes. The absence of these functions and species permits increased infection loads on a personal micro-
biome-specific basis. These findings suggest new targets for individualized preventative strategies of
V. cholerae infection through modulating the structure and function of the gut microbiome.

INTRODUCTION terotoxigenic Escherichia coli, V. cholerae, and rotavirus infec-

Gastrointestinal infections represent a major global health
concern. One major human diarrheal pathogen is Vibrio chol-
erae, the etiologic agent of the severe disease cholera that af-
fects millions of people annually (Clemens et al., 2017).
V. cholerae cycles between aquatic reservoirs and the small in-
testine, requiring the coordinated regulation of environmental
fitness genes versus virulence genes including the attachment
factor toxin co-regulated pilus (TCP) and cholera toxin (CT) (Her-
rington et al., 1988; Miller et al., 1987). In this and other patho-
gens, regulation depends on the chemical state of the gut,
shaped by the gut microbiome, the dense resident gut microbial
community (Eckburg et al., 2005) that varies dramatically across
host species and across individuals as a function of diet, geog-
raphy, and environmental insults (Yatsunenko et al., 2012). In
cholera-endemic areas, the gut microbiome is subject to mutu-
ally reinforcing pressures: malnutrition, leading to reduced host
infection resistance, repeated diarrhea, and poorly controlled
antimicrobial usage in an attempt to mitigate the resulting
sequelae. Previous 16S ribosomal gene studies of the gut micro-
biome in these areas demonstrate that these factors are able to
drive the gut microbiome into a characteristic dysbiotic state,
dominated by Streptococci such as Streptococcus salivarius,
Enterococci, and Enterobacteriaceae. This configuration has
been shown to be inducible by malnutrition (Subramanian
et al., 2014), and diarrhea irrespective of etiology, including en-
tion (David et al., 2015; Hsiao et al., 2014; Kieser et al., 2018).
Generalized model ‘‘healthy’’ microbial communities of the hu-
man gut have been shown to be resistant to V. cholerae infection
(Hsiao et al., 2014), and associative metagenomic studies have
examined how the microbiome differs between cholera patients
and household contacts that did not exhibit disease symptoms
(Midani et al., 2018). Yet, few studies have mechanistically
explored how interpersonal microbiome variation can drive path-
ogen susceptibility. Here, we show that the post-malnutrition/
post-diarrheal dysbiotic community is highly vulnerable to sub-
sequent infection. Moving beyond a dichotomous ‘‘normal’’
versus ‘‘dysbiotic’’ comparison, we show that microbiome differ-
ences among healthy humans drive striking differences in sus-
ceptibility. We show that fecal studies in animals and potentially
humans may have limited utility for studies of community interac-
tions with pathogens of the small intestine, and that microbiome-
dependent infection susceptibility at the small intestine can be
rescued by microbiome transplantation. In order to identify com-
mensals that strongly interact with enteropathogens across
many community contexts, we established an efficient unbiased
experimental pipeline that revealed that the commensal species
Blautia obeum can suppress virulence. We identify an enzymatic
mechanism in B. obeum that degrades the host-produced viru-
lence-inducing molecule taurocholate (TC), which B. obeum
uses alongside other mechanisms (Hsiao et al., 2014) to sup-
press V. cholerae virulence gene activation and colonization.

Cell 181, 1533–1546, June 25, 2020 ª 2020 Elsevier Inc. 1533
 ll
Article

A B

C D

Figure 1. Model Human Gut Microbiomes Replicate Structure of Communities Affected by Diarrhea-Induced Dysbiosis
(A) Defined human gut communities.
(B) Composition of healthy US human donor fecal microbiomes.
(C) Principal coordinates analysis of defined and complete human gut microbiomes based on weighted UniFrac distance, % variance explained shown in pa-
rentheses. Ellipses show 95% confidence intervals.
(D) Weighted UniFrac distance to indicated defined human model microbiomes of fecal samples from cholera patients at the end of diarrhea (left) and healthy
human donors (right) *p < 0.05, ****p < 0.0001, Mann-Whitney U-test. Boxplots show inter-quartile range, whiskers minimum to maximum.

RESULTS of defined gut communities using cultured human isolates (Fig-

Dysbiotic Microbiomes Are Susceptible to V. cholerae
Colonization, and Pathogen Resistance Can Be Rescued
by Microbiome Transplantation
We took a two-pronged approach to study the effects of micro-
biome variation on pathogen resistance, both involving reconsti-
tuting human gut microbiomes in animal models of V. cholerae
colonization and virulence. The first involved the construction
ure 1A), and the second involved studies with complete human
fecal microbiomes (Figure 1B).
As the basis for designing defined model communities, we
compared fecal microbiomes of a healthy adult volunteer cohort
in the United States (Figure 1C; Table S2C) and previously pub-
lished 16S ribosomal RNA gene sequencing of Bangladeshi
adults (Hsiao et al., 2014; Subramanian et al., 2014). Previous
studies in Bangladesh revealed that cholera drives the human

1534 Cell 181, 1533–1546, June 25, 2020

 ll
Article

A B C D

E F

G H

(legend on next page)

Cell 181, 1533–1546, June 25, 2020 1535

 ll
gut microbial community to a highly dysbiotic, low-diversity state
dominated by Streptococci, which recovers to a configuration
similar to non-diarrheal individuals over the course of weeks after
the cessation of acute disease (Hsiao et al., 2014). This has also
been observed in other diarrheal infections, such as enterotoxi-
genic E. coli and rotavirus (David et al., 2015; Kieser et al., 2018),
and other gut pathologies, such as severe malnutrition (Subra-
manian et al., 2014). Principal coordinates analysis (PCoA) of a
human cohort from Bangladesh (Hsiao et al., 2014) displays
the dysbiosis caused by cholera (Figure 1C, Diarrhea (start))
and community structure weeks after the cessation of diarrhea
(Diarrhea (end) to Recovery (end)), when the microbiome be-
comes similar to that of individuals in the same area not suffering
from acute malnutrition or diarrhea (Subramanian et al., 2014).
Interpersonal microbiome variation in Bangladesh was far higher
than among healthy US individuals sampled as part of this study;
indeed, some Bangladeshi ‘‘healthy’’ microbiomes closely
resemble cholera-diarrheal communities. As infectious diarrhea
and malnutrition are frequent in cholera endemic areas, we hy-
pothesized that the distinctive dysbiotic microbiome structure
observed during recovery from multiple sources of environ-
mental insult to the gut may be a recurring window of vulnera-
bility to cholera.
We then used human-derived isolates to assemble defined gut
communities (Figure 1A) based on these metagenomic analyses.
One model microbiome (‘‘CR’’) was based on metagenomic sur-
veys of healthy individuals, characterized by high taxonomic di-
versity but commonly including members of the genera Bacter-
oides, Clostridium, and Blautia (Arumugam et al., 2011; Qin
et al., 2010; Yatsunenko et al., 2012). Another (‘‘DS’’) model mi-
crobiome is characteristic of the dysbiotic state found in cholera-
endemic areas, comprising Streptococci, Enterococcus faecalis,
and E. coli. 16S sequencing analysis confirmed that the CR com-
munity is more similar to healthy human microbiomes than dys-
biotic diarrheal microbiomes, while the DS model community
was more similar to microbiomes at the conclusion of cholera
(Figures 1C and 1D).
We grew bacterial species from each defined group in pure
culture and used culture optical density (OD600) to introduce
equivalent amounts of each member species with V. cholerae
to germfree (GF) adult C57BL/6 mice by intra-gastric gavage.
Overall, microbial load during infection was equivalent as
measured by qPCR of 16S gene levels (Figure 2F). Mice that
received the CR microbiome at infection were resistant to
V. cholerae colonization, compared to animals receiving DS mi-
crobes (Figures 2A and 2B). We observed these colonization
Figure 2. Gut Microbiome Composition Contributes to V. cholerae Infection Resistance
(A and B) V. cholerae colonization in germfree adult mice harboring defined model communities co-gavaged with V. cholerae in (A) feces and (B) small intestines
4 days post-infection.
(C) Fecal V. cholerae colonization in GF adult mice harboring defined communities for 2 weeks and then gavaged with V. cholerae. Mix: 10 days DS colonization,
followed by SR microbes 4 days prior to V. cholerae infection.
(D) Intestinal V. cholerae colonization of GF suckling mice co-gavaged with model communities and V. cholerae.
(E) Intestinal V. cholerae colonization of antibiotic-treated CD-1 suckling mice co-gavaged with indicated communities.
(F) 16S gene abundance in small intestine.
(G) Expression of tcpA in infected mice with model human microbiomes.
(H) T6SS effects on small intestine colonization in antibiotic-treated CD-1 pups. *p < 0.05, **p < 0.01, ***p < 0.001 (Mann-Whitney U-test); n.s., not significant. Error
bars represent mean ± SEM. n = 6–12 animals for all experiments.
1536 Cell 181, 1533–1546, June 25, 2020
Article
phenotypes in both feces and in the medial and distal thirds of
the small intestine. In prior studies in adult mice, small intestinal
colonization by V. cholerae required antibiotics (Freter, 1955,
1956) and ketamine anesthesia (Olivier et al., 2009). Our results
with human, as opposed to murine, gut bacteria suggest that mi-
crobiome differences across host species and inter-individual
variation within host species both play key roles in determining
pathogen susceptibility. Significantly, we could restore coloniza-
tion resistance by mixing the CR and DS bacteria, suggesting
that susceptibility is reversible through microbiome modification
(Figures 2A and 2B). In ‘‘Mix’’ groups, where mice were adminis-
tered a 1:1 mixture of CR and DS, V. cholerae levels were signif-
icantly less compared to that in DS mice, in fact dropping below
the level of CR mice 2 days post-infection.
We also observed increased colonization susceptibility of DS
microbiomes when compared to a simplified model healthy mi-
crobiome (‘‘SR’’) when GF mice were colonized with defined
communities for 2 weeks prior to introduction of V. cholerae (Fig-
ure 2C). The SR community consisted of three species repre-
senting major phylogenetic lineages commonly found in the
healthy human gut. To model an attempted microbiome restora-
tion of a fully established and dense gut community, we also
introduced DS microbes for 10 days, followed by a gavage of
SR microbes 4 days prior to infection with V. cholerae. In this
Mix group, pathogen colonization was strongly inhibited
compared to DS-colonized mice, suggesting that microbiome
modification could be used to restore colonization resistance
even to entrenched dysbiotic communities.
We then profiled gut microbiome structure during infection in
feces and small intestines of gnotobiotic animals with different
communities (Figures 3A–3C and 3G). In these samples, the
CR and DS communities were distinct and the CR community
more similar to complete fecal microbiomes of healthy US do-
nors, while co-inoculation of CR and DS led to an intermediate
final microbiome.
Non-dysbiotic Human Microbiomes Reduce Virulence
Gene Expression and Colonization of V. cholerae in a
Suckling Mouse Model of Infection
While gnotobiotic adult mice serve as a useful microbial-interac-
tion model, we extended our studies to suckling mice, where the
pathology and virulence gene expression observed during
V. cholerae infection is closer to that of humans (Klose, 2000).
First, we recapitulated the CR and DS resistance phenotypes
in suckling GF C57BL/6 animals (Figure 2D). We then con-
structed a more accessible and scalable model of microbiome-
 ll
Article

A B C

D E F

G H

Figure 3. Addition of the CR Model Human Microbiome to Mice Hosting DS Microbes Yields a Community Structure Closer to Complete
Fecal Communities of Healthy Human Volunteers
(A–F) Principal coordinates analysis (PCoA) of microbial community diversity based on weighted UniFrac distance, % variance explained shown in parentheses
for each axis. Ellipses show 95% confidence intervals. (A) PCoA of fecal samples and distal third of small intestine of GF mice with model communities during
V. cholerae infection compared to healthy US donor fecal samples, with (B) PC1 positions and (C) all pairwise weighted UniFrac distances to healthy US donor
fecal samples. (D) PCoA of model communities and healthy human donor communities in suckling mice, with (E) PC1 positions and (F) all pairwise weighted
UniFrac distances to healthy US donor fecal samples.

(legend continued on next page)

Cell 181, 1533–1546, June 25, 2020 1537

 ll
pathogen interaction by clearing the native murine flora of CD-1
pups with streptomycin before introduction of human-associ-
ated species. Using this system, we observed similar micro-
biome-dependent infection outcomes; competitive CR/DS
transplantation yielded a dominant CR-like phenotype (Fig-
ure 2E), while CR and DS colonization load did not differ in
non-Vibrio-infected animals (Figure 2F). This pattern was re-
flected in 16S sequencing data (Figures 3D–3F; Table S3). During
infection, pups receiving CR microbes had very different com-
munity structure (with Vibrio reads filtered) compared to animals
with DS microbes, and animals receiving a mixed inoculum
(CR+DS) closely resembled CR mice.
Total microbial diversity was not a dominant factor in infection
resistance, because we restored colonization resistance in DS
mice to almost full CR levels by transplanting only a small subset
of CR species (SR). Expression levels of the key colonization fac-
tor tcpA were reduced
9.7-fold in SR compared to DS animals
(Figure 2G). We did not observe significant microbiome-depen-
dent differences in cholera toxin gene expression, diarrhea, or
fluid accumulation in these animals (Figure S1).
Recent studies have shown type VI secretion system (T6SS)
killing of murine commensal E. coli acts to drive increased viru-
lence in infection of suckling mice (Zhao et al., 2018). As our
DS model community contains E. coli, albeit a different strain,
we tested the effects of T6SS on V. cholerae colonization and
E. coli levels in our experimental system. A T6SS DvasK mutant
was deficient for colonization compared to wild-type V. cholerae
in antibiotic-cleared suckling mice (Figure 2H). However, T6SS
activity did not significantly alter levels of a co-inoculated strep-
tomycin-resistant K12 E. coli, and we observed E. coli at compa-
rable levels in DS and Mix (DS+CR) communities in the small
intestine (Figure 3G). These differences may be E. coli strain-
specific, or due to the much higher levels of V. cholerae used
in previous T6SS studies.
Together, our data suggested that the mechanism for
improved colonization resistance of CR/SR microbes lay in
T6SS-independent manipulation of V. cholerae virulence gene
expression.
Inter-individual Variation in Pathogen Colonization
Resistance of Human Gut Microbiomes
Our microbiome transplantation system in suckling mice al-
lowed us to screen numerous intact human fecal microbiomes
collected from healthy adult volunteers without malnutrition or
recent antibiotic usage or diarrhea for effects on V. cholerae
(Figure 1B). These complete fecal communities were taxonom-
ically similar to the CR, but not DS microbiomes in both original
microbial content and community structure upon transplanta-
tion (Figures 3D–3F). This was unsurprising, because the CR
model community represented up to 73% of genus-level diver-
sity by total relative abundance in these samples, while
members of the DS community only represented <1% of the
total (Table S4).
(G) Microbiome structure during infection with V. cholerae and host reads filtered (left) and in antibiotic-treated suckling mice without V. cholerae (right).
(H) B. obeum abundance in adult GF mice containing indicated microbiomes during V. cholerae infection (4 days post-infection). *p < 0.05, ****p < 0.0001; n.s. not
significant (Mann-Whitney U-test). Error bars represent mean ± SEM.
1538 Cell 181, 1533–1546, June 25, 2020
Article
We normalized fecal slurries and transplanted these samples
into antibiotic-treated suckling mice with V. cholerae (Figure 4),
with dramatically different effects on V. cholerae colonization,
even though these fecal communities colonized suckling animals
at similar efficiencies and density (Figure 2F). We observed an

1.5-log10 range of V. cholerae colonization depending on the
human donor (Figure 4), suggesting wide variation in infection
outcomes based on individual gut microbiome structure. The
higher basal microbiome diversity in Bangladesh (Figure 1C)
suggests that interpersonal variations in infection resistance in
endemic areas could be substantially higher.
A Pipeline for Randomization of Microbiome Members
Identifies Commensal Species Consistently Correlated
with V. cholerae Infection Outcome
We hypothesized that the CR species best able to colonize intes-
tines with the DS community might be prophylactic for infection,
because transplantation of CR microbes into DS communities
reduced V. cholerae colonization. Therefore, we examined gut
microbiome structure during V. cholerae infection in GF animals
with a mixed CR+DS community (Figures 3A and 3G). The CR
community in the small intestine was quite distinct from that in
feces, but all animals with this community were consistently
colonized by Blautia and Bacteroides species. The DS commu-
nity was consistent in feces and small intestine, and dominated
by Streptococci. The lack of generalizability of fecal data to other
gut compartments suggested that fecal sampling studies may
mask important differences for pathogenesis in the proximal
gastrointestinal tract.
In the small intestine, B. obeum maintained its relative abun-
dance in gnotobiotic CR+DS and CR-only animals (Figure 3H),
suggesting that it may play a role in CR infection resistance
and in transmitting this phenotype to DS animals. However,
these findings had potentially limited translational applicability
given the basal inter-personal diversity in humans; the ability to
displace one dysbiotic microbiome and resist V. cholerae coloni-
zation may not be representative across diverse individuals and
microbial communities. To identify Vibrio-antagonistic microbes
across many different microbiome contexts, we generated
random, unique, 5-member combinations drawn from CR and
DS strains (Figures 5A and 5B) and established OD600-normal-
ized mixtures of these bacteria in antibiotic-cleared suckling an-
imals with and without Vibrio infection. We reasoned that species
whose presence/absence or abundance consistently correlated
against V. cholerae colonization in many different communities
would be excellent putative targets for anti-Vibrio interventions.
We identified several species consistently associated with
pathogen levels across multiple microbiome combinations.
Higher levels of B. obeum were significantly associated with
reduced V. cholerae colonization (Figure 5C), but did not directly
correlate with V. cholerae abundance. This is consistent with the
effects of the mixed CR/DS microbiome on infection; B. obeum
consistently established in the DS small intestine, but high levels

Article
were not required to affect V. cholerae. That this effect was seen
across numerous randomized communities suggests that the
inhibitory activity of B. obeum on V. cholerae infection may be
broadly generalizable across many gut microbiomes. Except
for B. obeum, we found no statistically significant effects on
V. cholerae colonization based on the presence or absence of
an SR or CR species. In contrast, levels during infection of DS mi-
crobes (Streptococcus, E. faecalis, and E. coli) all positively and
significantly correlated with higher Vibrio levels (Figure 5D). In
mice with combinations with both B. obeum and
S. thermophilus, V. cholerae colonization was comparable to
mice with combinations including B. obeum but not Strepto-
coccus, again supporting the observation that B. obeum’s ef-
fects on pathogenesis are dominant across diverse microbiomes
(Figure 5C).
Because SR microbes largely recapitulated the colonization
resistance of the full CR microbiome, we performed similar ana-
lyses looking for whether combinations of different SR microbes
with B. obeum yielded lower V. cholerae colonization than when
those species were present in isolation. We observed no statis-
tically significant additive effects on V. cholerae colonization of
adding either Bacteroides vulgatus or Clostridium scindens to
B. obeum in defined communities (Figure S2).
Intestinal Signals That Induce V. cholerae Virulence
Gene Expression Are Depleted by B. obeum
Having identified a candidate driver of V. cholerae infection resis-
tance, we began to search for a molecular mechanism for these
effects. Numerous host and some commensal microbial cues
regulate V. cholerae gene expression in vivo (Gupta and Chowd-
hury, 1997; Kovacikova et al., 2010; Yang et al., 2013). Prior
studies have identified a role for a B. obeum-produced AI-2 auto-
inducer in downregulating the expression of TCP biogenesis
genes during infection (Hsiao et al., 2014). Consistent with this
finding, we observed reduced tcpA expression during infection
of mice with microbiomes containing B. obeum (Figure 2G).
In vivo conditions for virulence gene regulation can be
mimicked ex vivo using microaerophilic/anaerobic growth of
V. cholerae with intestinal tissue from mice (Yang et al., 2013).
We took homogenates of suckling mouse intestine and incu-
ll
Figure 4. Gut Microbiome Composition
Contributes to Inter-personal Differences
in V. cholerae Infection Resistance
Intestinal V. cholerae colonization of antibiotic-
treated CD-1 pups colonized with complete fecal
microbiomes from healthy US human volunteers.
n = 5–7 animals for all experiments. Error bars
represent mean ± SEM.
bated them anaerobically with a
V. cholerae lacZ:PtcpA -sh ble zeocin
resistance reporter. As expected, intesti-
nal homogenates induced tcpA expres-
sion, while pre-treatment of intestinal ho-
mogenates with B. obeum ablated tcpA
induction (Figure 6A). As a control, we
boiled intestinal homogenates that had
been incubated with B. obeum cultures in order to remove AI-
2, which is heat labile (Figure S3). Strikingly, homogenates incu-
bated with B. obeum remained unable to induce tcpA even after
boiling, in contrast to boiled homogenates alone or homoge-
nates incubated with S. salivarius and then boiled (Figure 6A).
These data suggested that B. obeum can deplete virulence-acti-
vating signals within the gut as well as produce virulence-sup-
pressing signals.
The Bile Salt Taurocholate Acts as a Potent Virulence
Gene Activator and Is More Efficiently Degraded by
Commensal Microbes in Healthy but Not Dysbiotic
Communities
One abundant heat-resistant molecule present in the small intes-
tine is bile. In humans and mice, bile acids are synthesized in the
liver from cholesterol and stored in the gallbladder. These pri-
mary bile acids are then secreted into the duodenum, where
they, typically in their sodium salt form, act to aid in the emulsi-
fication and absorption of dietary fats. More than 95% of
secreted bile acids are actively absorbed by the terminal ileum
and sent back to the liver in a process known as enterohepatic
circulation (Di Ciaula et al., 2017). Bile is a highly complex
mixture, although bile acids dominate the dry weight of biliary
bile (Muraca et al., 1991). Many prior studies have focused on
crude extracts from varying sources, including ruminants, con-
taining poorly defined mixtures of bile molecules. Human bile
acids secreted into the small intestine are predominantly conju-
gated to taurine (taurocholic acid/sodium taurocholate) and
glycine (glycocholic acid/sodium glycocholate), while taurine-
conjugated forms predominate in mice (Li and Dawson, 2019;
Sayin et al., 2013). Commensal microbial action is important
for bile acid metabolism; bacterial enzymes (bsh, bile salt hydro-
lases) are able to remove the conjugated amino acids from
secreted bile acids in the small intestine, for example converting
glycocholate (GC) and taurocholate (TC) to cholate/cholic acid
(CA) (Jones et al., 2008; Ridlon et al., 2006; Song et al., 2019).
Indeed, in GF mice, the bile pool in the small intestine is almost
exclusively taurine-conjugated (Sayin et al., 2013).
Previous studies have shown that TC, one of the most abun-
dant bile molecules in humans and mice, can induce tcp
Cell 181, 1533–1546, June 25, 2020 1539
 ll
Article

C D

Figure 5. An Unbiased Combinatorial Strategy for Identifying Commensal Correlates of Protection Against and Susceptibility to V. cholerae
colonization
(A) Combinatorial strategy.
(B) 5-member microbiome embodiments randomly generated using CR/DS members (left). Resulting V. cholerae infection of antibiotic-treated suckling mice
containing defined microbiome embodiments are shown at right.
(C) Mean V. cholerae colonization in suckling mice bearing communities containing B. obeum or Streptococcus species alone and in combination. Data
normalized across experiments as fold colony-forming unit (CFU)-gavaged V. cholerae recovered after infection.
(D) Abundance of DS member species in randomized microbiomes correlated to resulting V. cholerae abundance after infection. Points represent mice receiving
different microbiome embodiments. *p < 0.05 (Mann-Whitney U-test); n.s., not significant. Error bars represent mean ± SEM.

expression under anaerobic conditions through modulating the
structure and activity of the upstream virulence activator TcpP
(Yang et al., 2013). Similarly, we saw that TC activated PtcpA,
while CA was not an efficient inducer (Figure 6B). Intestinal ho-
mogenate effects on tcpA was bile-specific; pre-treatment of ho-
mogenates with the bile-sequestering resin cholestyramine ab-
lated their ability to activate PtcpA (Figure 6A).
We next screened the DS and CR species for their effects on
TC. We incubated TC solutions at a physiologically relevant con-
centration with pure cultures of these microbes, heat-treated

1540 Cell 181, 1533–1546, June 25, 2020

 ll
Article

A B

C D E

Figure 6. B. obeum Exerts Effects on V. cholerae Colonization through Degradation of the In Vivo Virulence Gene Activating Signal Taur-
ocholate (TC)
PtcpA activity normalized to tcpA induction by 125 mM TC unless noted.
(A) Modulation of tcpA-activating signals in suckling CD-1 mouse intestinal homogenates by pure cultures of B. obeum and S. salivarius, with heat treatment.
(B) Bile effects on tcp gene expression in vitro.
(C) Effects of CR and DS pure cultures on TC activation of virulence in vitro.
(D) Effects of B. obeum bsh enzyme expression on TC-mediated tcp activation in vitro.

(legend continued on next page)

Cell 181, 1533–1546, June 25, 2020 1541

 ll
and filter-sterilized the resulting supernatants, and measured
their ability to induce PtcpA (Figure 6C). We observed dramatic
differences in the ability of these strains to affect TC virulence in-
duction, with members of the CR/SR microbiomes in general
being better able to prevent tcp activation. The ability to ablate
TC-mediated induction of tcp expression varied at genus level,
with Blautia torques unable to affect tcp induction by TC in
comparison to B. obeum, and Streptococcus infantarius able
to process TC in contrast to other DS Streptococci. Of SR
species, B. vulgatus, but not C. scindens, showed efficient TC
processing.
Because the SR community largely recapitulated the CR
colonization resistance phenotype, and B. vulgatus demon-
strated high activity against TC in vitro, we wanted to examine
the relative contribution of B. obeum and B. vulgatus on TC
levels in the distal small intestine. We colonized adult GF
mice with CR members, or CR species without B. obeum,
and measured TC and CA in the distal third of the small intes-
tine 2 days post-colonization, compared to GF mice (Fig-
ure S4A). As expected given the absence of microbial bsh,
GF distal small intestine showed a high TC/CA ratio, while the
presence of CR microbes efficiently processed TC to CA,
yielding low TC/CA ratios. Strikingly, the removal of B. obeum
restored the TC/CA ratio in the distal small intestine to a level
not statistically significantly different from GF animals, although
there was a trend toward more TC in GF animals. This sug-
gested that although other CR organisms can contribute to
TC processing to CA, B. obeum is particularly well suited to
manipulating the level of this bile acid in the distal small intes-
tine. This agrees with our findings that B. obeum efficiently col-
onizes the small intestine (Figures 3 and 5), and the presence of
B. vulgatus and B. obeum together does not significantly
improve the ability of a microbiome to exclude V. cholerae
(Figure S2).
A Bile Salt Hydrolase Enzyme Encoded by B. obeum Is
Able to Degrade Virulence-Activating Signals in the Gut
To determine the molecular basis for B. obeum’s effect on
TC-dependent virulence induction, we examined genetic
determinants of bile acid processing. The B. obeum genome en-
codes for a hypothetical choloylglycine hydrolase (EC 3.5.1.24).
Such bile salt hydrolase (bsh) enzymes catalyze the removal of
the conjugated amino acids of bile salts, for example the removal
of taurine from TC to form CA.
Putative bsh genes are broadly distributed across gut micro-
bial species, because the ability to survive the inhibitory effects
of bile are extremely important for enteric commensals (De
Smet et al., 1995). A recent study classed bacterial bsh genes
into several phylotypes based on sequence similarity and
showed that bsh phylotypes have variable and substrate-
dependent deconjugation activity (Song et al., 2019). We
binned predicted bsh genes in the CR and DS genomes into
these phylotypes (Table S5). By sequence alignment and pre-
(E) Mass spectrometry measurement of TC in suckling CD-1 mouse intestines after incubation with pure cultures of indicated strains.
(F) V. cholerae infection of suckling CD-1 mice after 1-day of colonization with indicated E. coli strains. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 (unpaired
Student’s t test). Error bars represent mean ± SEM. n = 3–10 for all experiments.
1542 Cell 181, 1533–1546, June 25, 2020
Article
dicted structure, B. obeum encodes for predicted type 1 bsh
enzymes, which are highly effective at deconjugating TC, GC,
glycodeoxycholate, and taurodeoxycholate (Song et al.,
2019), the strongest activators of tcpA expression in
V. cholerae (Figure S5). All DS members except S. infantarius
lacked annotated bsh genes, while many CR species encoded
bsh homologs. Although S. infantarius showed high activity
against TC in vitro, despite bearing bsh homologs to phylotypes
with poor predicted TC activity, we observed no statistically
significant difference in effects on V. cholerae colonization
compared to S. salivarius (Figure S4B). This suggests that there
may be differences in in vivo regulation of these enzymes in
S. infantarius. Conversely, B. torques, despite encoding several
putative bsh genes, was not able to prevent tcpA induction by
TC. B. vulgatus was an efficient TC processor in vitro, but
despite encoding, 3 putative bsh genes could not drive signifi-
cantly lower levels of TC to CA processing in the distal small in-
testine in the absence of B. obeum (Figure S4A), further sug-
gesting that enzyme expression or function may diverge in
in vitro versus in vivo.
V. cholerae also encodes a putative bsh enzyme (VCA0877).
However, this is a predicted phylotype 6 bsh, which has poor
activity against TC but higher activity against rarer bile acids
that do not participate in regulation of virulence but may be
bacteriostatic in vivo (Table S5). V. cholerae cannot prevent
TC-mediated tcp activation in vitro, suggesting that V. cholerae
has limited bsh activity against TC in comparison to B. obeum
(Figure S4C).
We constitutively expressed the B. obeum bsh
RUMOBE_000028 by cloning this locus downstream of a
constitutive PLtet-O1 promoter in E. coli, generating strain
bshC. This bshC strain efficiently reduced levels of TC and
tcp activation compared to the isogenic vector control in
both pure TC solutions (Figure 6D) and intestinal homogenates
(Figure 6E). Significantly, given the dominant effect of
B. obeum on Vibrio resistance, pure cultures of either
B. obeum or bshC reduce tcp activation by TC (Figure 6D)
and TC levels in intestinal homogenates (Figure 6E) in the pres-
ence of S. salivarius. The activity of this B. obeum enzyme is
able to affect V. cholerae in vivo, because V. cholerae is unable
to colonize mice gavaged with bshC as effectively as mice with
the vector control (Figure 6F).
Bile Salt Hydrolase Levels in Human Gut Microbial
Communities Are Correlated to V. cholerae Infection
Outcome
To determine the distribution of bsh phylotypes in human gut mi-
crobiomes predicted to be dysbiotic or healthy, we re-examined
an existing deeply sequenced shotgun metagenomic dataset of
human cholera patients in Bangladesh (Table S2D) (David et al.,
2015). Importantly, data was available from patients at presenta-
tion at clinic for cholera (‘‘Diarrhea (d0)’’) without any prior anti-
biotic usage and from patients that received oral antibiotics as

Article
A B
C
part of the standard of care for cholera (‘‘Diarrhea + abx’’). These
data showed that bsh levels were already affected by diarrhea
prior to any clinical intervention. We found that several bsh
phylotypes followed diarrhea-dependent patterns in comparison
to a healthy Bangladeshi cohort. Phylotypes 1, 3, 4, and 5 were
significantly depleted in dysbiotic samples compared to healthy
controls (Figures 7A and S6). Of these, phylotypes 1, 3, and 4
were shown to be highly active against TC (Song et al., 2019)
(summarized in Table S5). These data suggested that a charac-
teristic of healthy human microbiomes that may modulate
V. cholerae susceptibility is their ability to deconjugate TC into
non virulence-inducing forms.
We then assayed whether complete healthy US fecal com-
munities were differentially able to convert TC to non-tcp-acti-
vating forms. We took the first six healthy US donors and
anaerobically cultured bacteria from their fecal samples
in vitro and were able to recover species representing 66%–
99% relative abundance of the original sample (Table S4C).
We inoculated these complex fecal specimens in media and
ll
Figure 7. Levels of B. obeum bsh Enzymes
in Human Samples Correlate to Infection
Outcome and Can Independently Modulate
V. cholerae Colonization
(A) Levels of phylotype 1 bsh enzymes in meta-
genomic sequencing of fecal microbiomes of
cholera patients pre- (‘‘Diarrhea (d0)’’) and post-
antibiotic (‘‘Diarrhea + abx’’) treatment, compared
to healthy adults in Bangladesh.
(B) Mass spectrometry measurement of bile levels
in 125-mM solutions of TC incubated with indicated
cultured human fecal communities in vitro. y, TC
not detected.
(C) B. obeum bsh levels in intestines of antibiotic-
cleared suckling CD-1 mice colonized with com-
plex human fecal samples without V. cholerae,
compared to V. cholerae colonization of antibiotic-
cleared suckling animals bearing human donor
microbiomes. **p < 0.01, ***p < 0.001 (Mann-
Whitney U-test). Error bars represent mean ± SEM.
used standardized amounts of the result-
ing mixed cultures to treat TC solutions,
which we then used for virulence re-
porter assays. Strikingly, we observed
that microbiomes (subjects B and D)
that allowed higher V. cholerae levels
when transplanted in suckling animals
were also unable to completely remove
TC from solution after 24 h, whereas
communities exhibiting stronger coloni-
zation resistance (A, C, and E) reduced
TC to undetectable levels (Figure 7B).
Because species in genus Blautia
demonstrated differences in bsh activity
in vitro, as well as association with cholera
patients and uninfected family members
(Midani et al., 2018), we assayed for the
level of the bsh gene of B. obeum specif-
ically in total DNA extracted from human fecal samples by real-
time PCR. This also served as a function-specific measure
of B. obeum abundance in these complex fecal mixtures.
We found that communities associated with higher V. cholerae
colonization had lower levels of B. obeum bsh (Figure 7C), sug-
gesting that B. obeum abundance and specifically the presence
of the bsh activity is associated with resistance to V. cholerae
infection.
DISCUSSION
A role for gastrointestinal microbes in resistance to enteropath-
ogens such as V. cholerae has been recognized for some time
(Freter, 1955, 1956). However, this colonization resistance has
often been examined in dichotomous terms, either germfree or
conventional, or ‘‘normal’’ and damaged by specific factors
such as antibiotics. Our results suggest that beyond extreme
fluctuations in structure, such as those due to diarrhea and se-
vere malnutrition, diversity even within otherwise ‘‘healthy’’
Cell 181, 1533–1546, June 25, 2020 1543
 ll
human populations can serve as predictive markers for infection
resistance.
A key difficulty in identifying taxa that can drive infection
susceptibility is the complexity of animal gut microbiomes,
compounded by dramatic differences in the taxonomic diver-
sity across host species (Seedorf et al., 2014). The limited
taxonomic and functional resolution of many commonly em-
ployed metagenomic techniques is also a barrier for converting
observations in large population studies to mechanistic in-
sights on microbial effects on pathogenesis and other pheno-
types. Findings in this study and others (Hsiao et al., 2014), in
which single genes encoded by specific microbiome members
are able to affect V. cholerae colonization in isolation from
other functions, suggest that correlative studies have distinct
limits in their abilities to provide mechanistic insights to
microbiome-pathogen interactions. For instance, a recent
sequencing-based study that sampled gut microbes of cholera
patients and healthy household contacts identified microbes
from the same genus as associated with both individuals
with cholera and individuals with putative exposure but non-
progression to disease (Midani et al., 2018). Thus, genus-,
and possibly even species-level data may be insufficient to
identify clear targets for future mechanistic studies in the
absence of experimental manipulations.
Recent developments in high-throughput sequencing,
anaerobic microbiology, and gnotobiotic animal systems
have allowed for mechanistic studies of the interaction of hu-
man commensals and human pathogens in animal models of
colonization and disease. Specific target taxa identified by
multi-omics approaches can be established in animals, with
species and gene content defined before introduction of
pathogens. These types of approaches allowed us to identify
B. obeum as a key member of the human gut microbiome
that drove infection resistance and microbial interactions
with bile acids as a driver of virulence gene regulation
in V. cholerae and a mechanism of protection against infection.
We hypothesize that microbial bile metabolism most affects
V. cholerae pathogenesis during early infection. V. cholerae
tightly regulates gene expression in response to host
environmental signals such as bile. Bile acids are thought to
stabilize the structure of the key virulence regulator TcpP
(Yang et al., 2013), which drives the activation of toxT tran-
scription. ToxT then causes full induction of tcp and cholera
toxin, with TCP-dependent colonization thought to begin prior
to toxin expression (Lee et al., 1999). Following colonization,
the activity of bile becomes less clear. Some studies have
demonstrated that the unsaturated fatty acids in bile are
able to modulate the binding of ToxT to target promoters,
reducing CT and TCP expression (Plecha and Withey, 2015).
The deconjugated bile salt sodium deoxycholate promotes
interaction between the virulence activators ToxR and ToxS
and subsequent activity (Midgett et al., 2017). However,
ToxRS likely does not directly activate toxT, but rather acts
to boost the activity of TcpP at the toxT promoter (Krukonis
et al., 2000). Both conjugated and deconjugated bile acids
are also able to induce ToxT-independent activation of
cholera toxin dependent on ToxRS (Hung and Mekala-
nos, 2005).
1544 Cell 181, 1533–1546, June 25, 2020
Article
Our results suggest a model where, at the initial point of
colonization in the small intestine, tcp gene expression and
thus TCP biogenesis is determined by the balance of bile acids
that is modulated by commensal microbes with differential bsh
activity. Differences in early tcp gene expression, and thus
colonization, may have disproportionate impact on the pro-
gression of V. cholerae infections; variation in microbiome
bsh activity may help determine whether infection proceeds
to fulminant diarrhea or low or temporary colonization that
leads to mild or asymptomatic infections that are common in
cholera endemic areas (King et al., 2008). Once severe diar-
rhea has begun, the commensal community and existing
lumenal bile is depleted, and any future bile secretion is
predominantly conjugated primary species that stimulate
virulence.
Although B. obeum bile modification is an important regu-
lator of V. cholerae colonization, there may be additive effects
on pathogen behavior of multiple community members and
through other mechanisms. Removal of B. obeum was suffi-
cient to raise TC levels in the mouse distal small intestine,
but although constitutive expression of B. obeum bsh yielded
a 1 log drop in V. cholerae colonization, the full CR micro-
biome yielded almost 2 log-fold differences in colonization
compared to DS microbiomes. In V. cholerae, virulence
gene expression is negatively regulated by several different
quorum sensing systems involving the sensing of specific
autoinducers (Jung et al., 2015). Prior studies identified
B. obeum-produced autoinducer AI-2 as a suppressor of
tcpA, functioning through a pathway bypassing the canonical
AI-2 receptor LuxP and involving the regulator VqmA (Hsiao
et al., 2014). VqmA has been shown to activate the master
quorum-sensing regulator HapR, which leads to repression
of tcpP (Liu et al., 2006; Zhu et al., 2002). Thus, AI-2 expres-
sion and bsh activity by B. obeum may synergize to reduce
the level and activity of TcpP during infection. Additional
studies will be required to determine how these two inputs
exert their effects on regulation of virulence determinants of
V. cholerae in vivo. Non-B. obeum CR members may also
impact V. cholerae colonization, both at the level of virulence
regulation and possibly metabolic competition for colonization
niches. Host diet may play a role in colonization; bile is
secreted from the gallbladder in response to food ingestion,
and this varies by the fat content of the meal (Marciani
et al., 2013). Diet is also a potent driver of microbiome struc-
ture (Faith et al., 2011), but the effect of different dietary com-
positions on driving the microbiome to infection resistant or
susceptible states has not been well studied. Ingestion of
food has been shown to dramatically reduce the infectious
dose of V. cholerae due to buffering of stomach pH, but
also possibly by raising the levels of virulence-activating
conjugated bile acids secreted into the small intestine.
Taken together, our results suggest that variation in
human gut microbiomes are a significant contributor to
V. cholerae infection risk, and this can be modulated through
introduction of a human gut commensal with multiple molec-
ular effects on V. cholerae, able to affect levels of both
virulence-activating and virulence-suppressing signals at
the site of infection. This suggests that targeted microbiome
 ll
Article

modification can be a promising prophylactic target against REFERENCES

cholera.
Arumugam, M., Raes, J., Pelletier, E., Le Paslier, D., Yamada, T., Mende, D.R.,
Fernandes, G.R., Tap, J., Bruls, T., Batto, J.-M., et al.; MetaHIT Consortium
STAR+METHODS (2011). Enterotypes of the human gut microbiome. Nature 473, 174–180.
Bassler, B.L., Wright, M., and Silverman, M.R. (1994). Multiple signalling sys-
Detailed methods are provided in the online version of this paper tems controlling expression of luminescence in Vibrio harveyi: sequence and
and include the following: function of genes encoding a second sensory pathway. Mol. Microbiol. 13,
273–286.
d KEY RESOURCES TABLE Caporaso, J.G., Kuczynski, J., Stombaugh, J., Bittinger, K., Bushman, F.D.,
d RESOURCE AVAILABILITY Costello, E.K., Fierer, N., Peña, A.G., Goodrich, J.K., Gordon, J.I., et al.
B Lead Contact (2010). QIIME allows analysis of high-throughput community sequencing
B Materials Availability data. Nat. Methods 7, 335–336.
B Data and Code Availability Clemens, J.D., Nair, G.B., Ahmed, T., Qadri, F., and Holmgren, J. (2017).
d EXPERIMENTAL MODEL AND SUBJECT DETAILS Cholera. Lancet 390, 1539–1549.
B Human studies Dalia, A.B., McDonough, E., and Camilli, A. (2014). Multiplex genome editing
B Animal studies by natural transformation. Proc. Natl. Acad. Sci. USA 111, 8937–8942.
B Bacterial strains and growth conditions David, L.A., Weil, A., Ryan, E.T., Calderwood, S.B., Harris, J.B., Chowdhury,
d METHOD DETAILS F., Begum, Y., Qadri, F., LaRocque, R.C., and Turnbaugh, P.J. (2015). Gut mi-
B 16S library preparation crobial succession follows acute secretory diarrhea in humans. MBio 6,
e00381-15.
B Human gut microbiome 16S meta-analysis
B Metagenomic analysis of bsh phylotypes De Smet, I., Van Hoorde, L., Vande Woestyne, M., Christiaens, H., and Ver-
straete, W. (1995). Significance of bile salt hydrolytic activities of lactobacilli.
B Preparation of bacteria for animal studies
J. Appl. Bacteriol. 79, 292–301.
B Measurement of fluid accumulation
Di Ciaula, A., Garruti, G., Lunardi Baccetto, R., Molina-Molina, E., Bonfrate, L.,
B Assessment of T6SS killing by V. cholerae in vivo
Wang, D.Q., and Portincasa, P. (2017). Bile Acid Physiology. Ann. Hepatol. 16
B Quantitative real-time PCR
(Suppl 1 ), S4–S14.
B Culturing of complex human fecal communities
Eckburg, P.B., Bik, E.M., Bernstein, C.N., Purdom, E., Dethlefsen, L., Sargent,
B AI-2 heat-stability assay
M., Gill, S.R., Nelson, K.E., and Relman, D.A. (2005). Diversity of the human in-
B In vitro bile-dependent tcp induction testinal microbial flora. Science 308, 1635–1638.
B BSH structure comparisons
Faith, J.J., McNulty, N.P., Rey, F.E., and Gordon, J.I. (2011). Predicting a hu-
B Commensal effects on tcp-activating signals man gut microbiota’s response to diet in gnotobiotic mice. Science 333,
B Quantification of bile salts 101–104.
d QUANTIFICATION AND STATISTICAL ANALYSIS Freter, R. (1955). The fatal enteric cholera infection in the guinea pig, achieved
by inhibition of normal enteric flora. J. Infect. Dis. 97, 57–65.
SUPPLEMENTAL INFORMATION Freter, R. (1956). Experimental enteric Shigella and Vibrio infections in mice
and guinea pigs. J. Exp. Med. 104, 411–418.
Supplemental Information can be found online at https://doi.org/10.1016/j. Gupta, S., and Chowdhury, R. (1997). Bile affects production of virulence fac-
cell.2020.05.036. tors and motility of Vibrio cholerae. Infect. Immun. 65, 1131–1134.
Herrington, D.A., Hall, R.H., Losonsky, G., Mekalanos, J.J., Taylor, R.K., and
ACKNOWLEDGMENTS Levine, M.M. (1988). Toxin, toxin-coregulated pili, and the toxR regulon are
essential for Vibrio cholerae pathogenesis in humans. J. Exp. Med. 168,
We thank Jun Zhu and Gary Dunny for the kind gifts of V. cholerae and 1487–1492.
E. faecalis strains, respectively. We thank the Metabolomics Core Facility at
Hsiao, A., Ahmed, A.M.S., Subramanian, S., Griffin, N.W., Drewry, L.L., Petri,
University of California, Riverside for mass spectrometry studies. This work
W.A., Jr., Haque, R., Ahmed, T., and Gordon, J.I. (2014). Members of the hu-
was supported by the National Institute of General Medical Sciences
man gut microbiota involved in recovery from Vibrio cholerae infection. Nature
(R35GM124724 to A.H.).
515, 423–426.
Humbert, L., Maubert, M.A., Wolf, C., Duboc, H., Mahé, M., Farabos, D., Sek-
AUTHOR CONTRIBUTIONS
sik, P., Mallet, J.M., Trugnan, G., Masliah, J., and Rainteau, D. (2012). Bile acid
profiling in human biological samples: comparison of extraction procedures
All authors helped to design and analyze experiments. S.A., J.D.M., J.C.M.,
and application to normal and cholestatic patients. J. Chromatogr. B Analyt.
R.L., and J.Y.C. performed experiments. S.A., J.D.M., and A.H. wrote
Technol. Biomed. Life Sci. 899, 135–145.
the paper.
Hung, D.T., and Mekalanos, J.J. (2005). Bile acids induce cholera toxin expres-
sion in Vibrio cholerae in a ToxT-independent manner. Proc. Natl. Acad. Sci.
DECLARATION OF INTERESTS
USA 102, 3028–3033.

The authors declare no competing interests. Jones, B.V., Begley, M., Hill, C., Gahan, C.G., and Marchesi, J.R. (2008). Func-
tional and comparative metagenomic analysis of bile salt hydrolase activity in
Received: December 5, 2019 the human gut microbiome. Proc. Natl. Acad. Sci. USA 105, 13580–13585.
Revised: March 16, 2020 Jung, S.A., Chapman, C.A., and Ng, W.L. (2015). Quadruple quorum-sensing
Accepted: May 18, 2020 inputs control Vibrio cholerae virulence and maintain system robustness. PLoS
Published: June 16, 2020 Pathog. 11, e1004837.

Cell 181, 1533–1546, June 25, 2020 1545

 ll
Kelley, L.A., Mezulis, S., Yates, C.M., Wass, M.N., and Sternberg, M.J. (2015).
The Phyre2 web portal for protein modeling, prediction and analysis. Nat. Pro-
toc. 10, 845–858.
Kieser, S., Sarker, S.A., Sakwinska, O., Foata, F., Sultana, S., Khan, Z., Islam,
S., Porta, N., Combremont, S., Betrisey, B., et al. (2018). Bangladeshi children
with acute diarrhoea show faecal microbiomes with increased Streptococcus
abundance, irrespective of diarrhoea aetiology. Environ. Microbiol. 20,
2256–2269.
Kim, D., Langmead, B., and Salzberg, S.L. (2015). HISAT: a fast spliced aligner
with low memory requirements. Nat. Methods 12, 357–360.
King, A.A., Ionides, E.L., Pascual, M., and Bouma, M.J. (2008). Inapparent in-
fections and cholera dynamics. Nature 454, 877–880.
Klose, K.E. (2000). The suckling mouse model of cholera. Trends Microbiol. 8,
189–191.
Kovacikova, G., Lin, W., and Skorupski, K. (2010). The LysR-type virulence
activator AphB regulates the expression of genes in Vibrio cholerae in
response to low pH and anaerobiosis. J. Bacteriol. 192, 4181–4191.
Krukonis, E.S., Yu, R.R., and Dirita, V.J. (2000). The Vibrio cholerae ToxR/
TcpP/ToxT virulence cascade: distinct roles for two membrane-localized tran-
scriptional activators on a single promoter. Mol. Microbiol. 38, 67–84.
Lee, S.H., Hava, D.L., Waldor, M.K., and Camilli, A. (1999). Regulation and
temporal expression patterns of Vibrio cholerae virulence genes during infec-
tion. Cell 99, 625–634.
Li, J., and Dawson, P.A. (2019). Animal models to study bile acid metabolism.
Biochim. Biophys. Acta Mol. Basis Dis. 1865, 895–911.
Liu, Z., Hsiao, A., Joelsson, A., and Zhu, J. (2006). The transcriptional regulator
VqmA increases expression of the quorum-sensing activator HapR in Vibrio
cholerae. J. Bacteriol. 188, 2446–2453.
Liu, Z., Miyashiro, T., Tsou, A., Hsiao, A., Goulian, M., and Zhu, J. (2008).
Mucosal penetration primes Vibrio cholerae for host colonization by repressing
quorum sensing. Proc. Natl. Acad. Sci. USA 105, 9769–9774.
Marciani, L., Cox, E.F., Hoad, C.L., Totman, J.J., Costigan, C., Singh, G.,
Shepherd, V., Chalkley, L., Robinson, M., Ison, R., et al. (2013). Effects of
various food ingredients on gall bladder emptying. Eur. J. Clin. Nutr. 67,
1182–1187.
Midani, F.S., Weil, A.A., Chowdhury, F., Begum, Y.A., Khan, A.I., Debela, M.D.,
Durand, H.K., Reese, A.T., Nimmagadda, S.N., Silverman, J.D., et al. (2018).
Human Gut Microbiota Predicts Susceptibility to Vibrio cholerae Infection.
J. Infect. Dis. 218, 645–653.
Midgett, C.R., Almagro-Moreno, S., Pellegrini, M., Taylor, R.K., Skorupski, K.,
and Kull, F.J. (2017). Bile salts and alkaline pH reciprocally modulate the inter-
action between the periplasmic domains of Vibrio cholerae ToxR and ToxS.
Mol. Microbiol. 105, 258–272.
Miller, V.L., Taylor, R.K., and Mekalanos, J.J. (1987). Cholera toxin transcrip-
tional activator toxR is a transmembrane DNA binding protein. Cell 48,
271–279.
1546 Cell 181, 1533–1546, June 25, 2020
Article
Muraca, M., Vilei, M.T., Miconi, L., Petrin, P., Antoniutti, M., and Pedrazzoli, S.
(1991). A simple method for the determination of lipid composition of human
bile. J. Lipid Res. 32, 371–374.
Olivier, V., Queen, J., and Satchell, K.J. (2009). Successful small intestine colo-
nization of adult mice by Vibrio cholerae requires ketamine anesthesia and
accessory toxins. PLoS ONE 4, e7352.
Pettersen, E.F., Goddard, T.D., Huang, C.C., Couch, G.S., Greenblatt, D.M.,
Meng, E.C., and Ferrin, T.E. (2004). UCSF Chimera–a visualization system
for exploratory research and analysis. J. Comput. Chem. 25, 1605–1612.
Plecha, S.C., and Withey, J.H. (2015). Mechanism for inhibition of Vibrio chol-
erae ToxT activity by the unsaturated fatty acid components of bile.
J. Bacteriol. 197, 1716–1725.
Qin, J., Li, R., Raes, J., Arumugam, M., Burgdorf, K.S., Manichanh, C., Nielsen,
T., Pons, N., Levenez, F., Yamada, T., et al.; MetaHIT Consortium (2010). A hu-
man gut microbial gene catalogue established by metagenomic sequencing.
Nature 464, 59–65.
Ridlon, J.M., Kang, D.J., and Hylemon, P.B. (2006). Bile salt biotransforma-
tions by human intestinal bacteria. J. Lipid Res. 47, 241–259.
Sayin, S.I., Wahlström, A., Felin, J., Jäntti, S., Marschall, H.U., Bamberg, K.,
Angelin, B., Hyötyläinen, T., Oresic
, M., and Bäckhed, F. (2013). Gut micro-
biota regulates bile acid metabolism by reducing the levels of tauro-beta-mur-
icholic acid, a naturally occurring FXR antagonist. Cell Metab. 17, 225–235.
Seedorf, H., Griffin, N.W., Ridaura, V.K., Reyes, A., Cheng, J., Rey, F.E., Smith,
M.I., Simon, G.M., Scheffrahn, R.H., Woebken, D., et al. (2014). Bacteria from
diverse habitats colonize and compete in the mouse gut. Cell 159, 253–266.
Song, Z., Cai, Y., Lao, X., Wang, X., Lin, X., Cui, Y., Kalavagunta, P.K., Liao, J.,
Jin, L., Shang, J., and Li, J. (2019). Taxonomic profiling and populational pat-
terns of bacterial bile salt hydrolase (BSH) genes based on worldwide human
gut microbiome. Microbiome 7, 9.
Subramanian, S., Huq, S., Yatsunenko, T., Haque, R., Mahfuz, M., Alam, M.A.,
Benezra, A., DeStefano, J., Meier, M.F., Muegge, B.D., et al. (2014). Persistent
gut microbiota immaturity in malnourished Bangladeshi children. Nature 510,
417–421.
Yang, M., Liu, Z., Hughes, C., Stern, A.M., Wang, H., Zhong, Z., Kan, B., Fen-
ical, W., and Zhu, J. (2013). Bile salt-induced intermolecular disulfide bond for-
mation activates Vibrio cholerae virulence. Proc. Natl. Acad. Sci. USA 110,
2348–2353.
Yatsunenko, T., Rey, F.E., Manary, M.J., Trehan, I., Dominguez-Bello, M.G.,
Contreras, M., Magris, M., Hidalgo, G., Baldassano, R.N., Anokhin, A.P.,
et al. (2012). Human gut microbiome viewed across age and geography. Na-
ture 486, 222–227.
Zhao, W., Caro, F., Robins, W., and Mekalanos, J.J. (2018). Antagonism to-
ward the intestinal microbiota and its effect on Vibrio cholerae virulence. Sci-
ence 359, 210–213.
Zhu, J., Miller, M.B., Vance, R.E., Dziejman, M., Bassler, B.L., and Mekalanos,
J.J. (2002). Quorum-sensing regulators control virulence gene expression in
Vibrio cholerae. Proc. Natl. Acad. Sci. USA 99, 3129–3134.
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Bacterial and Virus Strains
Vibrio cholerae C6706 El Tor Hsiao Lab stock C6706
DvasK V. cholerae This paper N/A
Vibrio harveyi AI-2 bioassay strain Bassler et al., 1994 BB170
lacZ::PtcpA-sh ble zeocin resistance Liu et al., 2008 PtcpA-sh ble
reporter V. cholerae C6706 El Tor strain
pZE21 in V. cholerae C6706 This paper C6706-KMR
luxS- E. coli strain BW30045D(araD-araB) E. coli Genetic Resources at CGSC#8227
567, DlacZ4787(::rrnB-3), l-, DluxS1368, Yale. CGSC, The Coli Genetic
D(rhaD-rhaB)568, hsdR514 Stock Center
Escherichia coli Hsiao Lab stock DH5a- lpir
E. coli expressing B. obeum luxS This paper BW30045_RO_AI-2
Constitutive B. obeum bsh expression This paper bshC
strain: E. coli DH5alpir pZE21-BSH
Vector control for bshC This paper DH5a-lpir pZE21
Bacteroides caccae ATCC ATCC 43185
Streptococcus salivarius subsp. salivarius ATCC ATCC 13419
Collinsella aerofaciens ATCC ATCC 25986
Dorea formicigenerans ATCC ATCC 27755
Blautia torques ATCC ATCC 27756
Blautia obeum ATCC ATCC 29174
Eubacterium rectale ATCC ATCC 33656
Clostridium scindens ATCC ATCC 35704
Bacteroides vulgatus ATCC ATCC 8482
Bacteroides uniformis ATCC ATCC 8492
Streptococcus infantarius subsp. ATCC ATCC BAA-102
infantarius
Dorea longicatena DSMZ DSM 13814
Faecalibacterium prausnitzii DSMZ DSM 17677
Bifidobacterium longum subsp. longum DSMZ DSM 20219
Streptococcus salivarius subsp. DSMZ DSM 20617
thermophilus
Enterococcus faecalis Dunny Lab stock (University OG1RF
of Minnesota)
Bacteroides thetaiotaomicron Hsiao Lab stock VPI-5482
Biological Samples
Human volunteer donor fecal sample This paper Subject A
Human volunteer donor fecal sample This paper Subject B
Human volunteer donor fecal sample This paper Subject C
Human volunteer donor fecal sample This paper Subject D
Human volunteer donor fecal sample This paper Subject E
Human volunteer donor fecal sample This paper Subject F
Human volunteer donor fecal sample This paper Subject G
Human volunteer donor fecal sample This paper Subject H
Human volunteer donor fecal sample This paper Subject I
(Continued on next page)

Cell 181, 1533–1546.e1–e7, June 25, 2020 e1

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Human volunteer donor fecal sample This paper Subject J
Human volunteer donor fecal sample This paper Subject K
Human volunteer donor fecal sample This paper Subject L
Human volunteer donor fecal sample This paper Subject M
Human volunteer donor fecal sample This paper Subject N
Human volunteer donor fecal sample This paper Subject O
Human volunteer donor fecal sample This paper Subject P
Chemicals, Peptides, and Recombinant Proteins
Zeocin Research Products International Cat# 1006-33-0
Corporation
Sodium taurocholate hydrate Sigma Aldrich Cat# 86339
Sodium glycocholate hydrate Sigma Aldrich Cat# G7132
Cholic acid Alfa Aesar Cat# A1125714
Sodium taurodeoxycholate hydrate Sigma Aldrich Cat# T0557
Sodium glycodeoxycholate Sigma Aldrich Cat# G9910
Deoxycholic acid MP Biomedicals Cat# 0210149610
Tauro-b-muricholic acid Steraloids Inc. Cat# C1899-000
b-muricholic acid Steraloids Inc. Cat# C1895-000
Cholestyramine Sigma Aldrich Cat# C4650
Critical Commercial Assays
iQ SYBR Green Supermix Biorad Cat# 170882
Platinum Hot Start PCR Master Mix Thermo Scientific Cat# 13000013
SuperScript IV First-Strand Synthesis Invitrogen Cat# 18091200
System
Gibson Master Mix New England Biolabs Cat# E2611S
Deposited Data
Short-read sequencing data This paper European Nucleotide
Archive (ENA) PRJEB31497
Short-read sequencing data for meta- European Nucleotide Archive See Table S2 for accession
analysis numbers
Experimental Models: Organisms/Strains
Mouse: C57BL/6 UCR gnotobiotic facility N/A
Mouse: CD-1 IGS Charles River Laboratories N/A
Oligonucleotides
All primers for study, see Table S6. This paper N/A
Recombinant DNA
B. obeum codon optimized LuxS placed This paper N/A
downstream of the PLtet-O-1 constitutive
promoter sequence derived from the
plasmid vector pZE21 vector
(pMK_B. obeum_luxS)
Plasmid: Constitutive expression construct This paper N/A
for B. obeum bsh (bshc)
Software and Algorithms
QIIME Caporaso et al., 2010 http://qiime.org/
Phyre2 Kelley et al., 2015 http://www.sbg.bio.ic.ac.uk/

phyre2/html/page.cgi?id=index
Chimera Pettersen et al., 2004 https://www.cgl.ucsf.edu/chimera/
(Continued on next page)

e2 Cell 181, 1533–1546.e1–e7, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
HISAT2 Kim et al., 2015 http://ccb.jhu.edu/software/hisat2/
manual.shtml
Other
Lab diet Newco Distributors Cat# 5K52

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Ansel Hsiao (ansel.
hsiao@ucr.edu).

Materials Availability
Unique plasmids, strains, and reagents generated in this study are available from the Lead Contact with a completed Materials Trans-
fer Agreement.

Data and Code Availability

The accession number for the Illumina sequencing data reported in this paper is European Nucleotide Archive (ENA): PRJEB31497.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human studies
All human samples were part of a study approved by the UCR Institutional Review Board and followed NIH guidelines. We collected
intact fecal samples from a cohort of healthy adult volunteers at the University of California, Riverside. Inclusion criteria were: 1) age
between 18 and 40 years, 2) must be able to provide signed and dated informed consent, 3) must be willing and able to provide stool
specimen. Exclusion criteria were: 1) systemic antibiotic usage (oral, intramuscular, or intravenous) in the 2 weeks prior to sampling;
2) acute disease at time of enrollment (presence of moderate or severe illness with or without fever); 3) diarrhea (liquid or very loose
stools not associated with a change in diet) in the 2 weeks prior to sampling; 4) active uncontrolled GI disorders or diseases including
Inflammatory bowel disease (IBD), ulcerative colitis, Crohn’s disease, or indeterminate colitis, persistent, infectious gastroenteritis,
colitis, or gastritis, and chronic constipation; 5) Major surgery of the GI tract, excluding cholecystectomy and appendectomy, but
including major bowel resection at any time. Age inclusion criteria were chosen to avoid age-related microbiome differences, which
are strongest in early life (Yatsunenko et al., 2012). Fecal samples were collected aseptically from each person at UCR and
immediately preserved at 80 C until processing for DNA extraction, culturing, and animal colonization. Stocks of fecal slurries
for subsequent experiments were prepared by resuspending samples at 1:3 weight/volume in sterile reduced PBS and adding sterile
glycerol to a final concentration of 25% volume/volume.

Animal studies
All animal experiments used protocols approved by the Institutional Animal Care and Use Committee of the University of California,
Riverside (UCR) and followed NIH guidelines. All CD-1 suckling animals were purchased from Charles River Laboratories. Suckling
and adult germfree C57BL/6 mice were reared at the UCR gnotobiotic facility. No distinction was made between male and female
animals for bacterial studies. Adult animals were used at > 3 weeks of age. Germfree suckling mice used at 5-6 days of age. Animals
were checked for signs of moribund condition prior to use in experiments, and used for one experimental procedure only. Adult an-
imals were co-housed in cages without mixing sex. Male mice were separated except in cases of littermates.
For the antibiotic-cleared suckling mouse model, 4-day old suckling CD-1 animals were fasted for 1.5 hours, then orally dosed with

1mg/g body weight streptomycin using 30-gauge plastic tubing, after which the animals were placed with a lactating dam for 1 day.
After 24 hours, mice received microbial communities with V. cholerae in a maximum gavage volume of 50 ml. At 18 hours post-infec-
tion, animals were sacrificed, and relevant sections of intestinal tissue dissected and homogenized for CFU numeration and nucleic
acid extraction.
Germ-free C57BL/6 mice were bred and maintained in plastic gnotobiotic isolators at University of California, Riverside. Mice were
fed an autoclaved, low-fat plant polysaccharide-rich mouse chow (Lab Diet 5K52) and were 5–8 weeks old at time of gavage. Bac-
terial cultures were prepared as described above. Mice were fasted for 30 minutes prior to introduction of bacteria, and stomach pH
was buffered by intra-gastric gavage of 100 mL 1M NaHCO3, followed by gavage with 150uL of balanced defined microbial libraries.
Fecal samples were collected across the course of the experiment. Mice were sacrificed 4 days post gavage and small intestine

Cell 181, 1533–1546.e1–e7, June 25, 2020 e3

 ll
Article

collected and cut to three equal (proximal, medial, distal) sections by length. Samples were homogenized and used for CFU enumer-
ation of bacteria on LB agar containing 200 mg/mL streptomycin.
For establishing defined microbiomes prior to V. cholerae infection, germ-free C57BL/6 mice were maintained as mentioned above
and used at 6-13 weeks of age. Mice were fasted and given NaHCO3 as previously described and either given 150 mL of the simple
resistant (SR), or dysbiotic (DS) communities. For the Mix group, mice were initially given the DS microbiome embodiment. 10 days
after microbiome introduction and 4 days prior to V. cholerae infection, the SR community was introduced into the Mix group animals
by gavage. 2 weeks after human commensal colonization, each group was infected with
5 3 109 CFU V. cholerae O1 El Tor C6706.
Fecal samples were suspended in 500 mL of PBS and homogenized using a bead beater (BioSpec) at 1,400 RPM for 30 s. CFU
enumeration of V. cholerae was done on LB agar containing 200 mg/mL streptomycin.
We used the antibiotic-treated infant mouse model described above to determine which members of the healthy human micro-
biome contribute most to resistance to V. cholerae. We made 18 random combinations of human gut microbiome strains and intro-
duced them to suckling mice along with V. cholerae. Each combination included five unique strains, and each gavage contained the
equivalent total microbial mass of 300 mL of OD600 = 0.4 culture, divided evenly across all constituent strains. After introducing human
microbiome and V. cholerae to suckling mice, V. cholerae levels in homogenized intestines were determined by plating on selective
agar. The absolute abundance of each species was determined with a combination of 16S rRNA gene qPCR and 16S rRNA
sequencing.

Bacterial strains and growth conditions

All human gut commensal strains used are listed in Table S1. Unless otherwise noted, human gut strains were propagated in LYH-BHI
liquid medium (BHI supplemented to 5g/L yeast extract, 5mg/L hemin, 1mg/mL cellobiose, 1mg/mL maltose and 0.5mg/mL cysteine-
HCl). Cultures were then propagated in a Coy anaerobic chamber (5% H2, 20% CO2, balance N2) or aerobically at 37 C.
All V. cholerae strains were derived from the C6706 El Tor pandemic isolate, including the lacZ:PtcpA:sh-Ble zeocin resistance re-
porter strain (Liu et al., 2008), and propagated in LB media with appropriate antibiotics at 37 C. To construct a strain resistant to kana-
mycin, the plasmid pZE21 was cloned into V. cholerae C6706 (C6706-KMR) and propagated in LB with kanamycin sulfate (Fisher
Scientific, 50 mg/ml) and streptomycin sulfate salt (100 mg/ml). Vibrio harveyi BB170 was propagated in LM medium (Bassler
et al., 1994) aerobically at 37 C.
To construct bshC, a strain constitutively expressing a bile salt hydrolase found in B. obeum, the RUMOBE_00028 locus was ampli-
fied from B. obeum genomic DNA (primers: 50 -GTCGACGGTATCGATAATGCTTATGTGTACAGCTGC-30 and 50 -GCAGGAATTCGA-
TATCACTAATTCTGAAAATGAATCTGC-30 ). All cloning and amplification primers are listed in Table S6. This amplicon was then
cloned downstream of the constitutive PLtet-O1 promoter of plasmid pZE21 through digestion of the vector backbone with HindIII fol-
lowed by Gibson assembly (New England Biolabs). The resulting plasmid was then introduced by electroporation into E. coli DH5al
pir to generate bshC. Strains were propagated aerobically in LB with kanamycin (50 mg/ml) at 37 C.
A strain overexpressing the AI-2 signal of B. obeum (BW30045_RO_AI-2) was constructed by constitutively expressing the
B. obeum luxS AI-2 synthase into E. coli BW30045 (DluxS). The B. obeum luxS coding region (from genome position 33,305-
33,784) was codon-optimized for expression in E. coli, placed downstream of the PLtet-O-1 constitutive promoter sequence derived
from the plasmid vector pZE21 vector, and the construct cloned into vector pMK using the GeneArt Subcloning & Express Cloning
Service (ThermoFisher). This expression construct was then amplified and inserted into the endA gene of the E. coli genome using
primers (forward: 50 -CCAAAACAGCTTTCGCTACGTTGCTGGCTCGTTTTAACACGGAGTAAGTGTTAGAAAAATTCATCCAGCA-3 0 ,
reverse: 50 -GGTTGTACGCGTGGGGTAGGGGTTAACAAAAAGAATCCCGCTAGTGTAGGCGGGCAGTGAAAGGAAGGCC-30 ).
We used natural transformation (Dalia et al., 2014) to construct a DvasK V. cholerae. Fragments of flanking genomic DNA upstream
(forward: 50 -GAACTTTCGTCACGTAAGTC-30 , reverse: 50 -GTCGACGGATCCCCGGAATCATGAATTGTGTCCTTGTTTAC-30 ) and
downstream (forward: 50 -GAACTTTCGTCACGTAAGTC-30 , reverse: 50 -GTCGACGGATCCCCGGAATCATGAATTGTGTCCTTGTT
TAC-30 ) of vasK and an antibiotic resistant gene cassette (forward: 50 -ATTCCGGGGATCCGTCGAC-30 , reverse: 50 -TGTAGGCTG
GAGCTGCTTC-30 ) were amplified from V. cholerae genomic DNA. Amplicons were then joined by overlapping PCR and introduced
into C6706 via natural transformation. Resistant colonies were then selected on trimethoprim-containing agar (10ug/ml) and insertion
confirmed via PCR.

METHOD DETAILS

16S library preparation

For DNA from human fecal samples,
200 mg (average wet weight) fecal sample was suspended in 600 mL PBS. For mouse intestinal
samples, tissues were dissected, homogenized in 5mL PBS, and 500 mL of the homogenate used for DNA extraction.
500 mL 0.1mm
glass beads (BioSpec), 210 mL SDS %20, and 500 mL neutral phenol:chloroform:isoamyl-alcohol (24:24:1, Fisher Scientific) were
added to each sample, and samples were lysed by bead-beating followed by ethanol precipitation (Hsiao et al., 2014).
The V4 variable region of bacterial 16S ribosomal RNA genes was amplified in 25 mL total volume reactions comprising 1 mL of
extracted DNA as template, 10 mL Platinum Hot Start PCR Master Mix (ThermoFisher), 13 mL PCR-grade water and 0.5 mL of forward
and reverse primers (10 mM). Cycling conditions were 94 C for 3 min, followed by 33 cycles (94 C for 45 s, 50 C for 60 s, 72 C for 90 s),
and 72 C for 10 min. An equal amount of each amplicon (
240ng) was pooled into libraries, which were then purified using QIAquick

e4 Cell 181, 1533–1546.e1–e7, June 25, 2020

 ll
Article

PCR purification columns (QIAGEN) and subjected to sequencing using the Illumina MiSeq platform. Paired-end 150nt reads were
assembled, de-multiplexed, rarefied to > 900 reads per sample, and analyzed using the QIIME 1.9.1 software package (Caporaso
et al., 2010). Sequencing run results are summarized in Tables S2A and S3.

Human gut microbiome 16S meta-analysis

For the analysis of bacterial composition between the human gut microbial communities and our artificial communities, the
sequencing data of the V4 region of the 16S rRNA gene from published studies and samples collected for this study. For the different
phases of V. cholerae infection, we used the first and last time points of diarrhea, and the last time of recovery (Hsiao et al., 2014). The
last time point of fecal samples collected from parents of malnourished Bangladesh children were selected as the healthy adult
Bangladesh control (Subramanian et al., 2014). See Table S2C for sequence accession numbers. Defined community inputs were
designed on the basis that all the strains in the specific community are evenly distributed (CR: 1000 reads/species; SR: 3000
reads/species; DS: 2000 reads/species). All of the sequencing data were collected together and analyzed using QIIME as described
above.

Metagenomic analysis of bsh phylotypes

The protein sequences of the eight representative BSH were obtained from Song et al. (2019). Deep metagenomic sequencing data
was obtained from David et al. (2015) (see Table S2D for sequence accession numbers). The microbial community DNA was aligned
to the protein sequences using blastx. For queries that hit multiple protein sequences, the one that had the highest hit score was
selected. The relative abundance of each type of the BSHs = the BSH reads count / (total reads count – human reads count).
Host reads were determined by mapping metagenomic DNA to Homo sapiens reference genome (assembly GRCh38.p13) using
HISAT2 (Kim et al., 2015).

Preparation of bacteria for animal studies

Each human gut bacterium was cultured from glycerol stocks in LYH-BHI media for 48 hours at 37 C, and then diluted (1:50) in fresh
LYH-BHI media. After growth for an additional 48 hours, cultures were normalized for density by OD600. For inoculation into suckling
mice, the equivalent of a total of 300 mL of 0.4 OD600 culture divided evenly across strains by community was pooled, pelleted by
centrifugation, and resuspended in fresh LYH-BHI. Each mouse received this mass of bacterial cells in a maximum gavage volume
of 50 mL per pup. In mice containing multiple defined communities, normalized mixtures were prepared so that 300 mL of OD600 = 0.4
equivalent of each community was represented in the final gavage. In mice receiving V. cholerae, the total resuspension volume of
commensal strains was 25 ml, with the remaining 25 mL containing 1x104-1x105 CFU V. cholerae in PBS.
Microbial levels in human fecal slurries were estimated via real-time PCR using universal 16S primers as described below, and
samples were normalized to so that each suckling animal received slurries containing the equivalent of
20 mg of microbial genomic
DNA.

Measurement of fluid accumulation

Suckling mice were treated as described above. The fluid accumulation ratio percentage was determined as: [weight of intestines
(Large and Small) / mouse body weight] x 100.

Assessment of T6SS killing by V. cholerae in vivo

CD-1 suckling animals were gavaged with antibiotics as described above. 3-5 day old infant mice were orally inoculated with total of

1x109 CFU E. coli TB1 (lacZ-) and
1x104 CFU V. cholerae (lacZ+) together in 50 mL LB. Animals were sacrificed after 16 hr of infec-
tion. Mouse intestines were dissected and homogenized in 5 mL of PBS, and 10 mL of the homogenate used for enumeration of bac-
teria via serial dilution on LB agar containing streptomycin and X-gal.

Quantitative real-time PCR

Total bacterial load in fecal samples and intestinal homogenates was determined by using real-time quantitative PCR. Reactions
comprised 2 mL of extracted DNA (200ng/reaction) as template, 12.5 mL SYBRGreen Master Mix (BioRad), 10 mL PCR-grade water,
and 0.25 mL of forward and reverse primers at 10 mM (forward: 50 -CTCCTACGGGAGGCAGCAG-30 , reverse: 50 -TTACCGCGG
CTGCTGGCAC-30 ). Cycle conditions were 95 C for 3 min, followed by 39 cycles (95 C for 10 s, 55 C for 30 s, 95 C for 10 s, 65 C
for 5 s, 95 C for 5 s).
Levels of the B. obeum bsh enzyme (RUMOBE_00028) were determined by real-time PCR as described above, using the primers
50 -GCGATCAGATTACGATCACTC-30 and 50 -GCCATGCCAACACCTTTTTC-30 . 200ng of purified DNA from intestinal homogenates
of CD-1 mice colonized with complex human fecal samples were used as template for each reaction.
Levels of tcpA and ctxA expression in antibiotic-cleared CD-1 suckling animals containing SR and DS microbiomes were
measured using real-time PCR (tcpA: primers 50 -GAAGAAGTTTGTAAAAGAAGAACACG-30 and 50 -CGCTGAGACCACACCCATA-
30 , ctxA: primers 50 -CACTAAGTGGGCACTTCTCA-30 and 50 -TGATCATGCAAGAGGAACTCA-30 ), with recA (primers 50 -ATTGA
AGGCGAAATGGGCGATAG-30 and 50 -TACACATACAGTTGGATTGCTTGAGG-30 ) as a control. Templates were generated by Trizol
(Ambion) extraction of total RNA from intestines of SR and DS-colonized mice infected with V. cholerae, followed by cDNA synthesis

Cell 181, 1533–1546.e1–e7, June 25, 2020 e5

 ll
Article

with the SuperScript IV First-Strand Synthesis System (Invitrogen) following manufacturers’ instructions. Real-time PCR was per-
formed using conditions 95 C for 3 min, followed by 39 cycles (95 C for 10 s, 55 C for 30 s, 95 C for 10 s, 65 C for 5 s, 95 C for
5 s).

Culturing of complex human fecal communities

Fecal slurries of complex human fecal samples were prepared as described above, and spread on LYH-BHI agar and incubated aero-
bically and anaerobically at 37 C. All colonies recovered were gathered by scraping, and DNA extracted and 16S rRNA genes ampli-
fied for sequencing as described.

AI-2 heat-stability assay

Cultures of BW30045_RO_AI-2, BB170 and C6706 were grown overnight. BW30045_RO_AI-2 was subcultured 1:100 into 12ml of LB
and grown in a shaker at 37 C for until OD600 z0.22, centrifuged, and the supernatant filter sterilized. Aliquots of supernatant were
then heated at 100 C for 30 minutes and cooled to room temperature. AI-2 activity was assessed using the BB170 bioassay (Bassler
et al., 1994). Briefly, overnight cultures of reporter strain BB170 in LM medium were diluted at 1:1000 in AB medium, and 10 mL of cell-
free supernatant or heat-treated cell-free supernatant were then added to 90ml of BB170 dilution. Luminescence and OD600 of each
sample were measured immediately and after
3.5hrs growth at 30 C with agitation.

In vitro bile-dependent tcp induction

PtcpA-sh-ble was grown as overnight culture, diluted 1:1000 in fresh LB, and grown for
2 hours at 37 C. Each reaction was prepared
in 40ml 0.5X pH 8.5 LB medium, with sodium taurocholate hydrate (TC, Sigma-Aldrich), sodium glycocholate hydrate (Sigma Aldrich),
cholic acid (CA, Alfa Aesar), sodium taurodeoxycholate hydrate (Sigma-Aldrich), sodium glycodeoxycholate (Sigma Aldrich), deox-
ycholic acid (DCA, MB Biomedicals), tauro-b-muricholic acid (Steraloids Inc.), or b-muricholic acid (Steraloids Inc.) added to a final
concentration of 125mM. 2ml of reporter strain subculture was then added, and samples incubated anaerobically at 37 C for 4hrs. 2ml
of each reaction was then added to 200ml of 0.5X LB pH 8.5 ± 10mg/ml of zeocin (Sigma Aldrich), incubated for 30 minutes aerobically
at 37 C with agitation, and then serially diluted and plated onto LB agar plates with streptomycin to determine survival rates. Induc-
tion represents percentage of PtcpA-sh-ble reporter cells surviving treatment with zeocin after incubation with indicated samples un-
der anaerobic conditions, defined as (zeocin-treated sample survival/average of no-zeocin controls)*100. Where noted, survival rates
were normalized to that induced by 125mM taurocholate.

BSH structure comparisons

The potential 3D models of unknown structure were produced by Phyre2 (Kelley et al., 2015), based on amino acid sequence align-
ment to known protein structure. The 3D structure comparison and root-mean-square distance (RMSD) per-column spatial variation
among structures were calculated using UCSF developed Chimera (V1.14) (Pettersen et al., 2004).

Commensal effects on tcp-activating signals

Commensal isolate cultures were grown for 48 hr, and then subcultured at 3:100 for 48hrs. Growth was measured by OD600 and
cultures normalized to 1.5mL of OD600 = 0.4 culture. bshc and the corresponding vector strain were grown overnight in LB with kana-
mycin, subcultured at 1:100 for 24hr, and normalized as above. All cultures were clarified, the aqueous layer removed, and the pellets
resuspended in sterile PBS with TC to a final concentration of 125uM. Cultures grown with antibiotics were washed one additional
time with 1 volume of PBS prior to addition of TC. Samples were then incubated anaerobically for 24hr at 37 C followed by heat treat-
ment at 100 C for 30 minutes, cooled to room temperature, centrifuged and the supernatant filter-sterilized with a 0.22 mm filter.
These samples were then used to induce PtcpA-sh-ble, and percent survival following zeocin treatment determined as
described above.
Removal of tcp activating signals in the gut was assayed as above, replacing pure TC solution in PBS with homogenate. Tissues
were collected from 5-6-day-old CD-1 suckling animals in 2.5ml sterile H2O, disrupted with a tissue homogenizer, pooled, and centri-
fuged to clear tissue debris. The resulting aqueous layer was heat-treated and filter sterilized as described above. The resulting
sample was desiccated using a Savant Integrated speedvac system (Fisher Scientific) and resuspended in one-fifth volume of sterile
water. Four volumes of acetonitrile (Sigma Aldrich) were then added and sample was incubated at room temperature for 20 minutes
for deproteinization (Humbert et al., 2012). Samples were clarified, with the aqueous layer filter sterilized, desiccated and
resuspended in one-fifth original volume sterile H2O as described above. To sequester bile salts, 12.5mg of cholestyramine
(Sigma-Aldrich) was added to 0.5ml of de-proteinized sample, and the mixture incubated at 1 hour at 37 C with agitation followed
by passage through a 3kDa protein concentrator (Pierce PES Protein Concentrators).
Human complex fecal sample TC processing was assayed in vitro. 100 mL of fecal slurries in glycerol prepared as described
above was inoculated into 5ml LYBHI and incubated anaerobically for 2 days at 37 C. Cells were then pelleted, normalized to

OD600 = 0.4 in 1.5ml sterile PBS with 125 mM TC, and incubated anaerobically at 37 C for 24hrs. Supernatants were then collected
via centrifugation, heat-treated and filter sterilized as described above, and submitted for mass spectroscopy (see below).
The effects of B. obeum bsh on V. cholerae colonization was assayed by introducing bshC and the isogenic vector control into
suckling animals prior to infection with Vibrio. 4-day old CD-1 suckling mice were treated with 50 mL of 75mg/mL kanamycin,

e6 Cell 181, 1533–1546.e1–e7, June 25, 2020

 ll
Article

then returned to lactating dams. Overnight cultures of bshC and vector strains were normalized to the equivalent 300 mL of OD600 = 0.4
culture, and cells pelleted and resuspended in fresh LYH-BHI. 50 mL of this was then introduced via intra-gastric gavage into anti-
biotic-treated suckling mice that had been fasted for 1.5 hours. Pups were then returned to a lactating dam. After 1 day of pre-infec-
tion colonization with E. coli, animals were gavaged with V. cholerae as described above.

Quantification of bile salts

All standards (TC, CA, and DCA) were submitted as 10mM solutions. LC-MS analysis of bile acids was performed on a Synapt G2-Si
quadrupole time-of-flight mass spectrometer (Waters) coupled to an I-class UPLC system (Waters). Separations were carried out on
a CSH phenyl-hexyl column (2.1 3 100mm, 1.7mM) (Waters). The mobile phases were (A) water with 0.1% formic acid and (B) aceto-
nitrile with 0.1% formic acid. The flow rate was 250 mL/min and the column was held at 40 C. The gradient was as follows: 0min, 1%
B; 1min, 1% B; 8min, 40% B; 13min, 58.8% B; 13.5min, 100% B; 15.5min, 100% B; 16min, 1% B; 18min, 1% B. Flow rate was
ramped to 600 mL/min at 13.5 min to speed up column flushing and re-equilibration.
The MS was operated in positive ion mode (50 to 1200 m/z) with a 100ms scan time. Source and desolvation temperatures were
150 C and 600 C, respectively. Desolvation gas was set to 1100L/hr and cone gas to 150L/hr. All gases were nitrogen except the
collision gas, which was argon. Capillary voltage was 1kV. Injection volume was 1 mL for all samples. The identity of bile acids in sam-
ples was confirmed by mass, retention time, and MS/MS as compared to authentic standards. Samples were analyzed in random
order and injected in duplicate. Leucine enkephalin was infused and used for mass correction. Data processing (peak integration)
was performed using QuanLynx software (Waters). Accuracy of peak integrations was checked manually.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical tests were performed in the GraphPad Prism software package. Results are representative of two independent experi-
ments. Statistical parameters for studies are reported in relevant figure legends and tables.

Cell 181, 1533–1546.e1–e7, June 25, 2020 e7

 ll
Article

Supplemental Figures

Figure S1. V. cholerae Pathology and Gut Distribution in Different Microbiome Contexts, Related to Figure 2
All experiments are in antibiotic-cleared suckling CD-1 mice. (A) Expression of ctxA in intestinal tissues of infected mice containing defined model human mi-
crobiomes. (B) Fluid accumulation in intestines of infected mice containing defined model human microbiomes. (C) Distribution of V. cholerae in infected mice
containing defined model human microbiomes. n.s. not significant (Mann-Whitney U-test). Error bars represent mean ± SEM.
 ll
Article

Figure S2. Mean V. cholerae Colonization in Antibiotic-Cleared Suckling CD-1 Mice Bearing Communities Containing B. obeum in Combi-
nations of SR Species, Related to Figure 2
Normalized colonization across experiments reported as fold CFU V. cholerae gavaged recovered after infection. n.s. not significant (Mann-Whitney U-test).
 ll
Article

Figure S3. Induction of BB170 AI-2 Reporter by Indicated Cell-Free Supernatants, Related to Figure 6
**p < 0.01, (Mann-Whitney U-test). Error bars represent mean ± SEM.
 ll
Article

Figure S4. Contribution of Different Microbial Species to TC Levels, Related to Figure 6

(A) Mass spectrometry measurement of taurocholate (TC) to cholic acid (CA) ratio in distal third of small intestine of adult germfree C57BL/6 mice 2 days after
colonization with pure cultures of indicated strains. (B) Amount of V. cholerae recovered during co-infection of suckling CD-1 mice with either S. infantarius or
S. salivarius, normalized to input CFU V. cholerae. (C) Ability of B. obeum and V. cholerae to interfere with TC activation of virulence in reporter V. cholerae in vitro
after 5 hours incubation, normalized to induction by 125 mM TC. *p < 0.05, **p < 0.01 (Mann-Whitney U-test), n.s. not-significant. Error bars represent
mean ± SEM.
 ll
Article

Figure S5. Comparison of Bile Salt Hydrolases, Related to Figure 6

(A) Predicted structure of B. obeum bile salt hydrolase RUMOBE_00028. (B) Predicted structure of consensus phylotype 1 BSH. (C). Amino acid alignment of
RUMOBE_00028, phylotype 1 BSH, and phylotype 6 BSH using Chimera. Header in gray shows the spatial variation per column. Colored boxes shows structural
similarity between regions, with coloring of one-letter code amino acids using Clustal X, dependent on both residue type and the pattern of conservation across
aligned sequences.
 ll
Article

Figure S6. Levels of Different Phylotypes of Microbial bsh Enzymes in Metagenomic Sequencing of Fecal Microbiomes of Cholera Patients
Pre- (d0) and Post- (+abx) Antibiotic Treatment Compared to Healthy Individuals in Bangladesh, Related to Figure 7
*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001 (Mann-Whitney U-test), n.s. not-significant. Error bars represent mean ± SEM.
 Article

Neuronal Inactivity Co-opts LTP Machinery to Drive

Potassium Channel Splicing and Homeostatic Spike
Widening
Graphical Abstract Authors
Boxing Li, Benjamin S. Suutari,
Simon D. Sun, ..., Thomas A. Neubert,
Gordon Fishell, Richard W. Tsien

Correspondence
liboxing@mail.sysu.edu.cn (B.L.),
richard.tsien@nyulangone.org (R.W.T.)

In Brief
Silencing neuronal activity triggers similar
molecular mechanisms as activating
neurons during long-term potentiation,
demonstrating Hebbian mechanisms of
homeostatic spike regulation.

Highlights
d Chronic spike blockade with tetrodotoxin causes
homeostatic spike broadening

d Alternative splicing of BK channels by exclusion of a specific

exon is responsible

d Synaptic homeostasis starts CaM kinase signaling to drive

nuclear exit of Nova-2

d Chronic inactivity and hyperactivity can initiate similar LTP-

like events

Li et al., 2020, Cell 181, 1547–1565

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.013 ll
 ll

Article
Neuronal Inactivity Co-opts LTP Machinery
to Drive Potassium Channel Splicing
and Homeostatic Spike Widening
Boxing Li,1,2,* Benjamin S. Suutari,2,3,7 Simon D. Sun,2,3,7 Zhengyi Luo,4,7 Chuanchuan Wei,1,7 Nicolas Chenouard,2
Nataniel J. Mandelberg,2 Guoan Zhang,5 Brie Wamsley,2,6 Guoling Tian,2 Sandrine Sanchez,2 Sikun You,1
Lianyan Huang,1 Thomas A. Neubert,5 Gordon Fishell,2,6 and Richard W. Tsien2,3,8,*
1Neuroscience Program, Guangdong Provincial Key Laboratory of Brain Function and Disease, Zhongshan School of Medicine and The Fifth
Affiliated Hospital, Sun Yat-sen University, Guangzhou 510810, China
2Department of Neuroscience and Physiology, Neuroscience Institute, NYU Grossman Medical Center, New York, NY 10016, USA
3Center for Neural Science, New York University, New York, NY 10003, USA
4Guangdong Provincial Key Laboratory of Malignant Tumor Epigenetics and Gene Regulation, RNA Biomedical Institute, Sun Yat-sen

Memorial Hospital, Zhongshan School of Medicine, Sun Yat-sen University, Guangzhou 510120, China
5Department of Biochemistry and Molecular Pharmacology and Skirball Institute, NYU Grossman Medical Center, New York, NY 10016, USA
6Stanley Center for Psychiatric Research, The Broad Institute, Cambridge, MA 02142, USA
7These authors contributed equally
8Lead Contact

*Correspondence: liboxing@mail.sysu.edu.cn (B.L.), richard.tsien@nyulangone.org (R.W.T.)

https://doi.org/10.1016/j.cell.2020.05.013

SUMMARY

Homeostasis of neural firing properties is important in stabilizing neuronal circuitry, but how such plasticity
might depend on alternative splicing is not known. Here we report that chronic inactivity homeostatically in-
creases action potential duration by changing alternative splicing of BK channels; this requires nuclear
export of the splicing factor Nova-2. Inactivity and Nova-2 relocation were connected by a novel synapto-nu-
clear signaling pathway that surprisingly invoked mechanisms akin to Hebbian plasticity: Ca2+-permeable
AMPA receptor upregulation, L-type Ca2+ channel activation, enhanced spine Ca2+ transients, nuclear trans-
location of a CaM shuttle, and nuclear CaMKIV activation. These findings not only uncover commonalities be-
tween homeostatic and Hebbian plasticity but also connect homeostatic regulation of synaptic transmission
and neuronal excitability. The signaling cascade provides a full-loop mechanism for a classic autoregulatory
feedback loop proposed
25 years ago. Each element of the loop has been implicated previously in neuro-
psychiatric disease.

INTRODUCTION basis of neuronal homeostasis. This is surprising because AS

Neurons use multiple forms of plasticity to allow circuits to be
flexible yet stable: Hebbian plasticity (positive feedback to
strengthen already active circuit elements, as in long-term
potentiation [LTP]) and homeostatic plasticity (negative feed-
back to boost neuronal elements deprived of activity or depress
those undergoing hyperactivity). Homeostatic plasticity helps
keep neuronal circuits away from extremes of silence or runaway
excitation (Turrigiano and Nelson, 2004) and occurs at multiple
levels of synapses, neurons, and circuits (Desai et al., 1999;
Kim and Tsien, 2008; O’Leary et al., 2014; Turrigiano, 2008). Ho-
meostatic plasticity is known to rely on changes in gene tran-
scription (Schaukowitch et al., 2017), mRNA translation (Magh-
soodi et al., 2008; Penney et al., 2012; Schanzenbächer et al.,
2016), and protein stability (Ehlers, 2003). In contrast, alternative
splicing (AS) of precursor mRNA has received little attention as a
is a critical intermediary between gene transcription and
translation and engenders the enormous complexity of neuronal
proteomes. By regulating exon inclusion, AS diversifies synaptic
components, ion channels, and other key neuronal proteins
(Furlanis and Scheiffele, 2018; Licatalosi and Darnell, 2010;
Vuong et al., 2016). Dysregulation of AS contributes to brain dis-
orders (Licatalosi and Darnell, 2006; Parikshak et al., 2016),
possibly via altered homeostasis (Mullins et al., 2016; Ramocki
and Zoghbi, 2008; Styr and Slutsky, 2018; Wondolowski and
Dickman, 2013).
We asked whether AS supports homeostatic responses
induced by long-term inactivity. Would this be converse to
neuronal responses to hyperactivity (Ding et al., 2017; Iijima
et al., 2011; Li et al., 2007; Mauger et al., 2016; Xie and Black,
2001), or would inactivity engage its own distinct mecha-
nisms? In turn, how does AS affect neuronal activity?

Cell 181, 1547–1565, June 25, 2020 ª 2020 Elsevier Inc. 1547
 ll
Answering these questions would delineate a classic homeo-
static feedback loop
D Firing / DCa2 + signaling
[ Y
D Cellular proteins )D Gene expression
proposed more than 25 year ago at Brandeis (LeMasson et al.,
1993; Marder et al., 1996; Siegel et al., 1994). This scheme em-
bodies the feedback principles of neuronal homeostasis but has
never been worked out in a full loop for any aspect of neuronal
function. We focused on action potential duration (APD), funda-
mental in tuning neurons, synapses, and circuits. Even small
changes in APD greatly affect Ca2+ influx (Borst and Sakmann,
1998; Geiger and Jonas, 2000; Llinás et al., 1982) and, thus,
neurotransmitter release, gene expression, and overall neuronal
function (Byrne and Kandel, 1996; Deng et al., 2013; Jackson
et al., 1991; Matthews et al., 2009; Sabatini and Regehr, 1997).
APD is generally prolonged by inactivity (Kim and Tsien, 2008;
Trasande and Ramirez, 2007), an appropriate response to home-
ostatically restore Ca2+ entry, but remarkably little is known
about underlying mechanisms.
Here we report that inactivity-induced homeostatic regulation
of APD in excitatory neurons is controlled by AS of the large-
conductance Ca2+-activated potassium channel (BK channel).
We deciphered the underlying signaling mechanism: a specific
change in BK AS driven by a novel cascade involving Ca2+-
permeable glutamate receptors, voltage-gated Ca2+ channels,
calcium/calmodulin-dependent (CaM) kinase kinase b
(bCaMKK), CaM kinase IV (CaMKIV), and the splicing factor
Nova-2. We find that complete silencing of an excitatory neuron
triggers signaling similar to that activated during direct depolar-
ization. Strikingly, each signaling player is encoded by a gene
implicated previously in neuropsychiatric disease.
RESULTS
Homeostatic AP Broadening Driven by BK Current
Attenuation
Although homeostasis of action potential (AP) frequency is well
studied (Desai et al., 1999; Lee and Chung, 2014; Maffei and Tur-
rigiano, 2008), how APD is regulated remains mysterious. This
distinction is critical because spike width regulates neuronal
excitability, Ca2+ influx, and neurotransmission and is governed
by its own set of ion channels operating in parallel with channels
controlling spike frequency (Kimm et al., 2015).
To examine homeostasis of APD, we blocked spiking of
cultured cortical neurons by chronic (24- or 48-h) sodium chan-
nel blockade with tetrodotoxin (TTX) and examined APD after
TTX removal. Chronically silenced neurons had a significantly
longer APD than mock-treated controls (Figures 1A and S1),
and AP-induced Ca2+ transients were elevated 1.5- to 2-fold in
amplitude and duration (Figure 1B), as expected from APD-
dependent prolongation of Ca2+ channel activation (Bischof-
berger et al., 2002). Thus, chronic inactivity drove compensatory
1548 Cell 181, 1547–1565, June 25, 2020
Article
increases in APD and Ca2+ entry as negative feedback
autoregulation.
The slowing of repolarization that drove APD prolongation
was associated with blunted afterhyperpolarization (AHP) (Fig-
ure 1A). Both changes would ensue if potassium channel activity
was reduced. We considered the BK channel because it regu-
lates APD and AHP (Hu et al., 2001; Lee and Cui, 2010; Sausbier
et al., 2004; Shao et al., 1999). Indeed, inhibiting BK channels
with a specific blocker, iberiotoxin (IbTx), lengthened AP half-
width (Figures 1C and 1D). To test whether BK channels are
involved in inactivity-induced APD widening, we applied IbTx
to TTX-silenced cortical cultures (48 h) just before AP recording.
Strikingly, IbTx did not further widen APs (Figures 1C and 1D);
the effect of inactivity largely occluded IbTx’s effect on APD.
Evidently, neuronal silencing and IbTx converge on a common
pathway—inhibiting BK activity in widening the AP.
Chronic Inactivity Alters BK Channel AS
Chronic inactivity might inhibit BK channels via gene expression
and membrane trafficking. However, decreases in transcription,
mRNA translation, or protein trafficking were ruled out by mea-
surements of BK mRNA, total protein, and surface membrane
expression (Figures S2A–S2D). BK channels also undergo AS,
with profound effects on activity (Fodor and Aldrich, 2009; Shel-
ley et al., 2013; Shipston and Tian, 2016; Xie and McCobb, 1998;
Zarei et al., 2001). We assessed the expression of exons pre-
dicted to undergo splicing (X1–X6; Figure 1E) via RT-PCR (Pietr-
zykowski et al., 2008). In cultured cortical neurons, BK under-
went AS in X3, X4, and X6 but not in X1, X2, and X5 (Figures
1F and S3A). Notably, only X6 (called E29 hereafter) splicing
decreased after chronic TTX treatment (Figures 1F); AS of other
exons remained unchanged (Figure S3B), narrowing our search
for the basis of BK modulation.
We also tested the effect of enhanced activity by depolariza-
tion with K+-rich culture media (20, 40, and 60 mM K+;
Figure S3C). Surprisingly, chronic depolarization decreased
E29 exon inclusion (Figures 1G and S3D); the magnitude of shift
in AS varied with strength and duration of depolarization (Fig-
ure S3D), but always in the same direction as chronic inactivity.
We return later to resolve the paradox of how chronic silencing
and depolarization produce similar changes in BK AS.
Depolarization- or seizure-induced changes in AS involve
activation of CaM-dependent kinases (Xie and Black, 2001). In
contrast, little is known about AS induced by inactivity, leading
us to focus on its functional effect and underlying mechanisms.
Altered E29 Inclusion Dampens K+ Current and
Broadens Spikes
E29 encodes 27 amino acids strategically located within the
regulator of K+ conductance (RCK) domain, near the Ca2+
bowl, one of the Ca2+ binding sites promoting BK activation (Fig-
ure 1E). E29 inclusion affects BK responses to Ca2+ (Ha et al.,
2000) and to alcohol-induced microRNAs (miRNAs) (Pietrzykow-
ski et al., 2008). We expressed BK channels in HEK293 cells
to find out whether E29 inclusion influences BK currents. BK
currents were smaller with DE29 than with E29 (Figures 1H and
1I), whereas the membrane expression level was no different
(Figures S2E–S2G). Thus, the reduction of E29 inclusion after
 ll
Article

Figure 1. Chronic Spike Blockade-Induced BK Channel AS Is Responsible for Homeostatic Prolongation of AP Duration
(A) Action potentials (APs) recorded from sham control- or chronic TTX-treated (48 h) neurons (left); AP duration (APD) was measured as half-width (right, n = 10).
Scale bars, 20 mV, 1 ms.
(B) Single AP-elicited Ca2+ transients from control or TTX neurons transfected with GCaMP6s (left) and DF/F (right, n = 6).
(C) APs recorded from control (left) and TTX (right) neurons with (red) or without (black) IbTx. Scale bar, 20 mV and 1 ms.

(legend continued on next page)

Cell 181, 1547–1565, June 25, 2020 1549

 ll
TTX treatment likely contributes to the dampening of BK outward
current.
We returned to cultured cortical neurons to test directly
whether a drop in E29 inclusion leads to APD prolongation.
Upon overexpression of the DE29 BK construct, APD was
longer, in line with lower BK conductance, relative to that
observed for its E29 counterpart. Furthermore, action potentials
of E29-expressing neurons failed to broaden with chronic
silencing. Likewise, the already widened AP of DE29-expressing
showed no further prolongation (Figures 1J and 1K). Thus, inclu-
sion or exclusion of E29 in overexpressed channels overrides
the regulation engaged by chronic inactivity, suggesting that
such regulation involves E29 splicing. Clinching this calls for
understanding how the splicing is controlled.
Nova-2 Binds to the Intron Downstream of E29
AS of BK channels has been intensely studied, but how E29
inclusion is controlled remains unknown. Generally, splicing
regulatory factors favor exon inclusion when binding to down-
stream introns but inhibit it when binding to upstream introns
or the exon itself (Black, 2003; Ule et al., 2006). Our bioinformat-
ics analysis of the intronic sequence past E29 revealed YCAY
(Y=C/U) clusters (Figure 2A), consensus sequences for binding
of Nova proteins, a well-studied class of neuron-specific RNA-
binding proteins that regulate AS of neuronal proteins (Buckano-
vich and Darnell, 1997; Ule et al., 2005). Mice lacking Nova-2, the
predominant isoform in the neocortex (Yang et al., 1998), are
deficient in E29 inclusion (Ule et al., 2005).
To see whether E29 splicing is directly regulated by Nova-2, we
tested whether Nova-2 binds to synthetic RNA oligonucleotides
spanning the putative binding region in the downstream intron
(probe A1; Figure 2A). Pull-down assays showed that probe A1
binds efficiently to endogenous Nova-2 from mouse cortical ly-
sates. In contrast, no binding to Nova-2 was detected with a probe
with mutations in putative Nova-2-binding (YCAY) sites (probe A2)
or a control probe spanning a 35-bp intronic stretch upstream of
A1 without YCAY motifs (probe B; Figures 2A and 2B). We asked
whether Nova-2 also bound to the intron downstream of E29
in vivo, subjecting cortical lysates to RNA immunoprecipitation
(IP) with Nova-2 antibodies and assaying with RT-PCR using
primers flanking the YCAY sites. RT-PCR product was detected
following IP with Nova-2 antibodies but not with control immuno-
globulin G (IgG) (Figure 2C). No product from the Nova-2 immuno-
precipitate was detected without reverse transcriptase, excluding
an effect of genomic DNA contamination (Figure 2C). Thus, in vivo
and in vitro experiments show that Nova-2 binds directly to the
YCAY sites in the intron downstream of E29.
(D) Half-widths of APs depicted in (C) (n = 8).
(E) Diagram of BK channel pore-forming a subunit and AS sites (X1–X6, arrowheads) (Pietrzykowski et al., 2008). Amino acid sequences of the sixth exon X6, E29
(purple, underlined), near the ‘‘Ca2+ bowl’’ (orange, underlined).
(F) RT-PCR products from control or TTX cultures with or without E29 inclusion (E29 or DE29, left) and percentage of products containing E29 (right; n = 8).
(G) E29 splicing in control or 60 mM K+ (60K)-treated cultures (24 h) and percentage of products containing E29 (right, n = 6).
(H) Whole-cell recordings from HEK293 cells overexpressing BK channel constructs with or without E29.
(I) Corresponding current–voltage (I-V) curves.
(J) APs from control or TTX cortical neurons transfected with plasmids encoding BK channel isoforms with or without E29. Scale bar, 20 mV and 1 ms.
(K) Pooled half-widths of APs depicted in (J) (n = 7).
For (A), (B), (D), (F), (G), (I), and (K), data are represented as mean ± SEM; *p < 0.05; **p < 0.01; N.S. represents p > 0.05.
See also Figures S1–S3.
1550 Cell 181, 1547–1565, June 25, 2020
Article
Nova-2 Is Required for E29 Splicing
Testing the necessity of Nova-2 for E29 inclusion, we knocked
down Nova-2 in cultured cortical neurons using short hairpin
RNAs (shRNAs) that targeted its coding sequence (CDS), or its
untranslated region (UTR) (Figure S4A); the UTR-directed shRNA
spared an exogenous Nova-2 construct lacking the UTR (Nova-
2(R); Figure S4B). E29 inclusion was sharply reduced by lentiviral
delivery of UTR-targeting shRNA but not of scrambled shRNA
and not of exogenous Nova-2(R) (Figure 2D). This match to the
pattern of Nova-2 knockdown and rescue (Figure S4B) sug-
gested that Nova-2 is necessary for the AS event.
Changes in Nuclear Localization of Nova-2 Can Regulate
E29 Splicing
Because Nova-2 acts outside of the nucleus (Racca et al.,
2010), not just within it, we assessed the effect of cellular locale
by comparing the effects of Nova-2 lacking its nuclear localiza-
tion signal (DNLS) and with the NLS intact (wild type [WT]).
Although WT Nova-2 was concentrated in the nucleus, DNLS-
Nova-2 was mainly cytoplasmic (Figure 2E). We verified that
Nova-2 directly regulates E29 inclusion using a splice reporter
(Stoilov et al., 2008) containing E29 and partial flanking intron
sequences (Figure 2F) in HEK293 cells (largely lacking endoge-
nous Nova-2). In controls, E29 inclusion was knocked down by
Nova-2 shRNA and rescued by Nova2(R) (Figure S4C). Critically,
DNLS-Nova-2 largely failed to induce E29 inclusion (Figures 2F
and S4C), supporting the importance of nuclear localization.
Nova-2 Is Necessary and Sufficient for TTX-Induced E29
Exclusion
To confirm in neurons that Nova-2 is critical for chronic inac-
tivity-induced E29 reduction, we assayed the effects of chronic
TTX after knockdown of Nova-2 (Figure 2G). Although scram-
bled lentiviral shRNA spared the reduction of E29 inclusion
following 48-h TTX treatment, knockdown of Nova-2 mimicked
and occluded the effect of chronic inactivity on E29 inclusion
(Figure 2G). These experiments showed that Nova-2 binds
directly to the intron downstream of E29 and is necessary and
sufficient for regulation of E29 inclusion by inactivity.
Chronic Inactivity Reduces Nova-2 Nuclear Localization
In Vitro
This sets up the question of how Nova-2 effectiveness is linked to
chronic inactivity. Because Nova-2 must be in the nucleus to
control splicing, its cellular relocalization is a potential control
point. We assessed Nova-2 localization in cultured cortical neu-
rons after 48-h treatment with TTX. In TTX-treated neurons,

Article
Nova-2 nuclear intensity was lower than control cells, whereas
cytosolic Nova-2 intensity increased (Figure 3A), yielding a rela-
tive 2-fold drop in the nuclear/cytoplasmic ratio (Figure 3B, left).
We verified these immunocytochemical findings by nuclear frac-
ll
Figure 2. Nova-2 Regulates E29 AS
(A) Diagram of E29 (red box) and flanking introns (top)
and biotinylated RNA probes (A1, A2, and B) created,
spanning downstream intronic YCAY sequences (bot-
tom).
(B) RNA pull-down of proteins from adult mouse brains,
using the probes in (A) or empty beads with immunoblot
of Nova-2 protein (n = 3).
(C) RT-PCR products from IP of Nova-2 from mouse
cortical lysates. IP with IgG and an assay without
reverse transcription (No RT) are controls.
(D) E29 splicing in neurons transfected with scrambled
shRNA (Scr.), shRNA against the UTR of the Nova-2
gene (shRNA(UTR)), or shRNA(UTR) with exogenous
shRNA(UTR)-resistant Nova-2 (Nova-2 (R)).
(E) Diagram of Nova-2 with the NLS present (WT) or
absent (DNLS), with RNA binding domains (top: KH1,
KH2, and KH3) and immunofluorescence of FLAG-
tagged Nova-2 with or without the NLS. Nuclei are
indicated by dashes. Scale bar, 10 mm.
(F) Diagram of the splicing reporter containing consti-
tutive exons (black rectangles), E29 (red rectangle), and
partial flanking intron sequences (horizontal lines) (top)
and RT-PCR products from HEK293 cells transfected
with the splicing reporter with control vector (sham), WT
Nova-2 (WT), or DNLS-Nova-2 co-expression.
(G) E29 splicing in cultured cortical neurons infected by
lentiviral Scr. or shRNAs against Nova-2 gene exposed
to a sham control (Con) or 48 h TTX.
See also Figure S4.
tionation and immunoblot analysis (Figure 3C).
After TTX exposure, nuclear Nova-2 levels fell,
cytoplasmic Nova-2 rose, and total Nova-2
expression remained unchanged (Figures 3C
and 3D), indicating that Nova-2 translocates
from the nucleus to the cytosol upon silencing
of cultured neurons. A similar shift in the nu-
cleus/cytoplasmic ratio was seen upon
chronic depolarization (Figure 3B, right);
thus, the paradoxically similar regulation of
E29 inclusion arises at or before Nova-2
relocation.
Chronic Inactivity Reduces Nova-2
Nuclear Localization In Vivo
We next asked whether silencing neocortical
neurons in vivo caused similar changes as in
dissociated cultures. Tonic inputs from the
thalamus to the neocortex, which excite
ongoing activity of cortical neurons, were
chronically eliminated in mice by expressing
tetanus toxin light chain (TeTox) in Olig3-pos-
itive thalamic neurons (Figure 3E). Nova local-
ization was assayed in cortical neurons inner-
vated by thalamic inputs. Nova was largely
nuclear in Olig3 Cre (control) mice but excluded from the nucleus
in Olig3 Cre; TeToxf/+mice (Figures 3F and 3G). Chronic removal
of thalamic input led to Nova translocation from the nucleus to
the cytosol of cortical neurons.
Cell 181, 1547–1565, June 25, 2020 1551
 ll
Article

(legend on next page)

1552 Cell 181, 1547–1565, June 25, 2020

 ll
Article

Inactivity-induced Nova-2 translocation was also observed
with a monocular deprivation (MD) paradigm (Wiesel and Hubel,
1963). A lid suture was applied in juvenile mice for 5 days during
the critical period of visual development (post-natal days 26–31)
(Figure 3H), and subcellular localization of Nova-2 in monocular
V1 was assayed by immunostaining. The nuclear intensity of
Nova-2 in the deprived hemisphere (contralateral to the closed
eye) was less than in the non-deprived hemisphere (ipsilateral);
opposite differences were seen in cytoplasmic intensity (Figures
3I and 3J). Thus, MD-induced inactivity also induced in vivo
Nova-2 translocation from the nucleus to the cytosol. Strikingly,
MD also reduced E29 inclusion and increased APD in the vision-
deprived hemisphere compared with its non-deprived counter-
part (Figures 3K–3N). Thus, chronic inactivity in vivo leads to
nucleus-to-cytosol Nova-2 redistribution, reduction of E29 inclu-
sion, and increased APD.
sayed Nova-2 localization. In vector control-transfected neurons,
Nova-2 was predominantly located in the nucleus (Figure 4D, top
row). In contrast, in CA-CaMKIV-transfected neurons, Nova-2
nuclear intensity was lower, and cytoplasmic intensity was higher
(Figure 4D, second row), causing a significant drop in nuclear/
cytoplasmic (nuc/cyt) ratio (Figure 4E). Evidently, activation of
CaMKIV is sufficient to drive Nova-2 translocation. Continuous
activation of CaMKIV requires its binding to Ca2+/CaM. Accord-
ingly, we overexpressed a nucleus-localized Ca2+/CaM trap,
CaMBP4(nuc) (Cohen et al., 2016; Wang et al., 1995), before
examining Nova-2 location. This intervention abolished the
chronic inactivity-induced nuclear export of Nova-2 (Figures
4D, fourth row, and 4E), supporting the idea that Nova-2 translo-
cation is regulated by activated CaMKIV. These experiments indi-
cate that CaMKIV activation is sufficient and necessary to induce
Nova-2 translocation from the nucleus to the cytosol.

CaMKIV Is Activated by Chronic Inactivity CaMKIV Interaction with Nova-2

While cellular control mechanisms of Nova-2 localization have
not been reported, we considered scenarios including regula-
tion by phosphorylation. CaMKIV was a likely candidate as a
robust protein kinase, highly concentrated within the nucleus
(Bito et al., 1996; Nakamura et al., 1995); CaMKIV regulates
other splicing factors, such as heterogeneous nuclear ribonu-
cleoprotein L (hnRNP L) (Liu et al., 2012), and supports ho-
meostatic adjustments (Ibata et al., 2008). To test for involve-
ment of CaMKIV, we looked at its status in the nucleus after
chronic inactivity, monitoring phosphorylation on Thr196, its
dominant activation site (Selbert et al., 1995; Tokumitsu and
Soderling, 1996). Thr196 phosphorylation was higher in the
nucleus of TTX-treated neurons relative to the control (Fig-
ure 4A), resulting in a 1.5-fold higher nuclear/cytoplasmic ratio
(Figure 4B). Elevated activation of nuclear CaMKIV was also
seen in vivo upon immunoblot analysis of samples from
cortical V1, comparing lysates from monocularly deprived
and non-deprived hemispheres (Figure 4C).
Both Nova-2 and CaMKIV are mostly located in the nucleus (Fig-
ure 5A), but whether they directly interact is unknown. To explore
this, we co-expressed constructs encoding FLAG-tagged Nova-
2 and GFP-tagged CaMKIV in HEK293 cells. CoIP and western
blotting with anti-FLAG or anti-GFP antibodies were performed
on protein lysates to detect protein-protein interaction. FLAG-
tagged Nova-2 was found with IP using anti-GFP antibody;
GFP-tagged CaMKIV was seen upon IP with anti-FLAG antibody
(Figure 5B). No reciprocal interaction was detected upon IP of
lysates of mock-transfected cells or when they were expressed
individually (Figure 5B).
The interaction was also evident even at endogenous levels of
protein expression in neurons. Nova-2 was immunoprecipitated
with anti-CaMKIV antibody; CaMKIV was brought down with
anti-Nova-2 antibody, but no binding interactions were seen
with IgG controls (Figure 5C). These results indicate that Nova-
2 and CaMKIV engage in protein-protein interaction, studied as
exogenous or endogenous proteins.

CaMKIV Signaling Drives Nuclear Nova-2 Export CaMKIV Phosphorylates Nova-2 and Regulates Its
Is mimicking CaMKIV activation sufficient to induce Nova-2 Nuclear Localization
translocation? To test this, we expressed constitutively activated To find out whether CaMKIV phosphorylates Nova-2, we co-ex-
CaMKIV (CA-CaMKIV) (Cruzalegui and Means, 1993) and as- pressed CA-CaMKIV (or GFP as a control) with FLAG-tagged

Figure 3. Chronic Spike Blockade Reduces Nova-2 Nuclear Localization and E29 Inclusion
(A) Immunofluorescence of Nova-2 from neurons treated with the sham Con or 48 h TTX. Nuclei are indicated by dashes. Scale bar, 10 mm.
(B) Quantification of the nuc/cyt ratio of Nova-2 immunofluorescence intensity from neurons sham- or 48 h TTX-treated (left) and sham- or 24 h 60 mM K+ solution-
treated (right) (n = 30 cells).
(C) Western blots from nuclear (left), cytosolic (center), or whole-cell fractions (right), with Nova-2, Lamin B, and Glyceraldehyde 3-phosphate dehydrogenase
(GAPDH) levels.
(D) Quantification of (C) (for each condition, n = 4).
(E) Diagram of eliminated thalamic inputs to the cortex.
(F) Nova (red) localization in cortical neurons from Olig3 Cre or Olig3 Cre; TeToxf/+ mice. Scale bar, 20 mm.
(G) Quantified nuc/cyt ratio of Nova immunofluorescence intensity from the groups in (F) (n = 20 cells).
(H) Diagram of visual pathways involved in monocular deprivation (MD), with the monocular region of the primary visual cortex (V1mono) indicated (boxes).
(I) Nova-2 (red) localization in the V1mono region, contralateral (Contra) or ipsilateral (Ipsi) to visual deprived eye (MD). Scale bar, 10 mm.
(J) Quantified nuc/cyt ratio of Nova-2 immunofluorescence intensity from the groups in (I) (n = 20 cells).
(K) E29 splicing in V1mono Contra or Ipsi to the visually deprived eye (MD).
(L) Quantification of E29 splicing in (K) (n = 4).
(M) AP waveforms recorded from the V1mono region, Contra or Ipsi to the visually deprived eye (MD).
(N) Half-width of APs in (M) (n = 12–14).
For (B), (D), (G), (J), (L), and (N), data are represented as mean ± SEM; **p < 0.01; N.S. represents p > 0.05.

Cell 181, 1547–1565, June 25, 2020 1553

 ll
Nova-2 in HEK293 cells and probed Nova-2 phosphorylation
with mass spectrometry (liquid chromatography-tandem mass
spectrometry [LC-MS/MS]) (Figures S5A–S5D and 5D5G).
Compared with the control, CA-CaMKIV drove Nova-2 phos-
phorylation at three sites: serine 25 (site 1), threonine 27 (site
2), and serine 194 (site 3). Sites 1 and 2 lie within or next to the
NLS in Nova-2, predictive of influence on nuclear localization
(Harreman et al., 2004); site 3 resides in the KH2 domain, one
of the RNA binding domains (Yang et al., 1998). To test whether
phosphorylation affects Nova-2 location in cultured neurons, we
probed the distribution of FLAG-tagged Nova-2 constructs with
the sites mutated to glutamic acid (E) to mimic the negative
charge phosphorylation confers or to alanine (A) to prevent phos-
phorylation. In contrast to the mostly nuclear positioning of WT
Nova-2, Nova-2 with phosphomimetic mutations at sites 1 and
2 (1E2E) was mostly cytoplasmic (Figures 5H and 5I). Single mu-
tations (1E or 2E) reduced Nova-2 nuclear localization, even
when the other site was mutated to alanine (1E2A or 1A2E) (Fig-
ures 5H and 5I). Conversely, Nova-2 with single or double muta-
tions to alanine (2A and 1A2A) was predominantly nuclear, like
the WT (Figures 5H and 5I). In contrast, glutamate mutations at
site 3 had no effect on Nova-2 localization (Figures 5H and 5I)
or on Nova-2 binding to RNA (Figure S5E). These results showed
that active CaMKIV reduces Nova-2 nuclear localization by
phosphorylating sites 1 and 2 (S25, T27).
To test whether CaMKIV phosphorylation of Nova-2 controlled
its ability to regulate E29 splicing, we co-expressed the 1E2E
double mutant Nova-2 with the E29 splicing reporter in
HEK293 cells. Unlike WT Nova-2, the 1E2E variant was unable
to induce E29 inclusion, like Nova-2 lacking its NLS (DNLS; Fig-
ure 5J). Evidently, CaMKIV binds to Nova-2 in the nucleus,
primed to phosphorylate serine 25 and threonine 27 sites near
the NLS of Nova-2; the resulting Nova-2 nuclear exit prevents
its action in splicing, reducing exon 29 inclusion.
1554 Cell 181, 1547–1565, June 25, 2020
Article
Figure 4. Inactivity-Induced CaMKIV Activa-
tion Leads to Nuclear Nova-2 Export
(A) Immunofluorescence of phosphorylated CaM-
KIV (pCaMKIV) from neurons treated with sham
Con or 48 h TTX. The nucleus is indicated (dashed
white). Scale bar, 10 mm.
(B) Quantified nuc/cyt ratio of pCaMKIV immuno-
fluorescence intensity from groups in (A) (n = 20).
(C) Western blots indicating pCaMKIV and CaMKIV
levels in V1mono Contra or Ipsi to the visually
deprived eye (MD).
(D) Immunofluorescence of neurons transfected
with Con vectors, FLAG-tagged CA-CaMKIV (top
two rows, sham Con treatment), or FLAG-tagged
nuclear CaMBP4 (bottom two rows, 48 h TTX).
Scale bar, 10 mm.
(E) Quantified nuc/cyt ratio of Nova-2 immunofluo-
rescent intensity in (D) (n = 20).
For (B) and (E), data are represented as mean ±
SEM; **p < 0.01.
Spontaneous Spine Depolarizations Detected with a
Membrane Voltage Probe
How does chronic elimination of AP firing lead to activation of
CaMKIV and nuclear exit of Nova-2? Activity blockade with
TTX abolishes evoked vesicle release but spares spontaneous,
AP-independent vesicle release. We asked whether sponta-
neous synaptic transmission might be sufficient to activate
signaling to CaMKIV activation and Nova-2 translocation.
To test this, we monitored the membrane potential in den-
dritic spines of cortical neurons using the genetically encoded
voltage indicator ASAP1 (St-Pierre et al., 2014). ASAP1 fluores-
cence was seen in the plasma membrane of somata, dendritic
shafts, and spines (Figure 6A). Decreases in ASAP1 fluores-
cence were linearly related to membrane depolarizations (St-
Pierre et al., 2014) imposed by K+-rich external solution (Fig-
ure 6B, gray symbols). In cortical cultures acutely exposed to
TTX, dendritic spines exhibited dips in relative fluorescence in-
tensity (DF/F), reflecting spontaneous excitatory postsynaptic
potentials (EPSPs) (Figure 6C, green exemplar trace, and 6B,
pooled data). After chronic TTX, spontaneous spine depolariza-
tions grew by
10 mV (Figure 6C, red trace), exceeding depo-
larization attained with 20 mM K+ (Figure 6B, pooled data), suf-
ficient to recruit L-type Ca2+ channel- and NMDA receptor
(NMDAR)-mediated Ca2+ influx (Helton et al., 2005; Mayer
et al., 1984; Nowak et al., 1984). For comparison, even larger
DF/F transients were observed in dendritic spines of control
neurons during backpropagating dendritic APs (bAPs) or so-
matic action potentials recorded without TTX (Figures 6C,
blue and violet exemplar traces, and 6B, pooled data).
Chronic Inactivity Enhances Spontaneous Spine Ca2+
Transients
To study spontaneous Ca2+ transients in dendritic spines,
we expressed the fluorescent Ca2+ indicator GCaMP6s
 ll
Article

Figure 5. CaMKIV Phosphorylates Nova-2 and Regulates Its Nuclear Localization

(A) Immunofluorescence of CaMKIV and Nova-2. Scale bar, 10 mm.
(B) Immunoblots of coIP lysate from HEK293 cells transfected with Nova-2-FLAG, CaMKIV-GFP, or both. Sham-transfected cells were used as Con.
(C) Immunoblots of CaMKIV and Nova-2 coIP lysate from cortical neurons. IgG was used as Con.
(D) Location of S25, T27, and S194 (red); amino acid sequences of the NLS (blue, underlined) and the partial KH2 domain are indicated.
(E–G) Quantification of phosphorylation levels of sites S25 (E), T27 (F), and S194 (G) of Nova-2 induced by CaMKIV, normalized to cells co-expressing Nova-2 and
GFP (red dotted lines) (n = 3).
(H) Micrographs of FLAG-tagged wild-type (WT) and mutant Nova-2. Nuclei are indicated by white dotted circles. Scale bar, 10 mm.
(I) Quantification of the nuc/cyt ratio of the fluorescence intensity of WT or mutant Nova-2 from (H) (n = 20–62). **p < 0.01.
(J) RT-PCR products from HEK293 cells expressing the E29 splicing reporter and WT Nova-2, DNLS Nova-2, or S25E/T27E Nova-2.
For (E)–(G) and (I), data are represented as mean ± SEM. See also Figure S5.

(Figures 6D–6G). Although the soma of cortical neurons was treated with TTX than in controls (Figure 6H). Likewise, the Ca2+
totally silent in the presence of TTX (Figures S6A and S6B), den- transients were taller, broader, and larger in area (Figures 6I–6K),
dritic spines remained active despite spike blockade (Figures whereas fluorescent punctum size and event frequency per
6D–6G and S6C). More spines were active in neurons chronically punctum were no different (Figures S6D and S6E). Thus, chronic

Cell 181, 1547–1565, June 25, 2020 1555

 ll
Article

(legend on next page)

1556 Cell 181, 1547–1565, June 25, 2020

 ll
Article

blockade of APs enhances spontaneous Ca2+ transients in
spines, in line with optical voltage recordings.
Contributions of Various Ca2+ Pathways during
Spontaneous Transmission
We characterized the elevated synaptic Ca2+ transients, mindful
of increases in Ca2+-permeable, GluA1-containing AMPA recep-
tors (AMPARs) following activity blockade (Kim and Ziff, 2014;
Thiagarajan et al., 2005). Live-labeled surface GluA1 increased
in neurons that had undergone chronic inactivity (Figures S6F
and S6G). Elevated surface GluA1 contributed to the enlarged
Ca2+ transient, indicated by a sharp drop in synaptic Ca2+ tran-
sients upon acute exposure to the GluA1 antagonist philantho-
toxin (PhTx) (Figure 6L). Importantly, PhTx completely blocked
chronic inactivity-induced AP prolongation (Figures 6M and
6N), indicating that GluA1 activation was critical for homeostatic
regulation of APD.
Following chronic inactivity, elevated Ca2+ transients were
also inhibited by blockade of NMDAR with ((2R)-amino-5-
phosphonovaleric acid; (2R)-amino-5-phosphonopentanoate)
APV and depletion of internal Ca2+ stores with thapsigargin
(Figure 6L). This aligns with AMPAR activation recruiting
Ca2+ delivery via NMDAR and intracellular Ca2+ stores (Empt-
age et al., 1999). Voltage-dependent L-type Ca2+ channels
(CaV1) also contributed to spine Ca2+ transients during
chronic inactivity, as judged by partial reduction with nimodi-
pine (Figure 6L). A glutamate-gated cation current would
create a voltage drop across the spine neck resistance (Har-
nett et al., 2012; Palmer and Stuart, 2009), giving rise to
directly measured depolarizations and CaV1- and NMDA-
mediated Ca2+ influx (Figure 6L).
We verified inactivity-induced engagement of Ca2+ signaling
by testing for activation of CaMKII and CaMKI in dendritic
spines, which, respectively, undergo autophosphorylation or
CaM kinase kinase (CaMKK)-mediated phosphorylation in
response to local Ca2+ signals (Wayman et al., 2008). Using
site-specific anti-phospho-Thr antibodies, we showed CaMKII
Thr286 autophosphorylation and CaMKI Thr177/178 phos-
phorylation, respectively (Figures S6H–S6K). Chronic inactivity
augmented the intensity of both markers relative to the control
(Figures S6H–S6K), providing independent biochemical evi-
dence that local Ca2+ signaling in spines is enhanced by
chronic inactivity.
The CaV1-CaMKK-CaMKIV Pathway Drives Reduced
E29 Inclusion
Intensified minis and spine Ca2+ signaling could link inactivity
to nuclear AS. To find out whether this involves a classical
CaV1-CaMKK-CaMKIV cascade, like that engaged by depolari-
zation (West et al., 2002), we monitored phosphorylation of
CaMKIV as a pivotal step. Inactivity induced elevation of phos-
phorylated CaMKIV (pCaMKIV) but not in the presence of nimo-
dipine (a CaV1 blocker), KN93 (a CaMK inhibitor), or STO-609
(a CaMKK blocker). This pattern was consistent in western
blots of nuclear extracts (Figures 7A and 7B) and in nuc/cyto ra-
tios obtained by immunocytochemistry (Figures S7A and S7B).
Nova-2 localization showed a reciprocal pattern (Figures 7C,
7D, S7C, and S7D) which was mirrored by TTX-induced reduc-
tion on E29 inclusion (Figures 7E and 7F). These results sup-
ported a chain of signaling events emanating from spontane-
ously active spines whereby a CaV1-CaMKK-CaMKIV pathway
drives reductions in nuclear Nova-2 and in E29 inclusion.
Activation of a CaV1-CaMKK-CaMKIV pathway by chronic
inactivity appears surprising because it seemingly recapitulates
effects of hyperactivity (Deisseroth et al., 2003; Ma et al.,
2014). However, we found that directly imposed depolarization
caused similar effects as chronic TTX on Nova-2 translocation
and BK E29 exclusion (Figures S7E and S7F), with the latter ef-
fect completely prevented by nimodipine or KN93 (Figures 7G,
7H, and S7F). Thus, activation of CaV1 channels and CaMKs
drives Nova-2 relocation and E29 splicing, irrespective of how
the pathway is engaged.
bCaMKK Translocation Is Required for E29 Inclusion
A remaining question is how, without spiking, CaV1 activation at
dendritic sites causes activation of CaMKIV in the nucleus. In
excitation-transcription coupling (E-T coupling) following acute
depolarization, signaling to nuclear CaMKIV involves transloca-
tion of Ca2+/calmodulin via different shuttle proteins: gCaMKII
in cortical, hippocampal, and sympathetic neurons (Ma et al.,
2014) and gCaMKI in parvalbumin-positive inhibitory neurons
(Cohen et al., 2016). In seeking a translocator for excitation-AS
coupling, we looked for an increase in nuclear level paired
with a drop in cytoplasmic level during chronic inactivity. This
pattern was not evident for aCaMKI, bCaMKI, gCaMKI, dCaMKI,
gCaMKII, aCaMKK, and CaM itself (Figures 7J, S7G, and S7H).
In contrast, levels of bCaMKK rose in the nucleus and fell in

Figure 6. Chronic Spike Blockade Leads to Elevated Depolarization and Ca2+ Transients in Dendritic Spines
(A) Micrograph of a neuron expressing ASAP1. Scale bar, 10 mm.
(B) Quantification of membrane potential (ordinate, left) of somata and spines induced by APs (AP in a soma, purple; bAP in a spine, blue) or by AP-independent
synaptic transmission (spine depolarization in sham Con cultures, Vspine green; spine depolarization in 48 h TTX cultures, Vspine red), plotted against the cor-
responding change in ASAP1 fluorescence from the respective events (DF/F). See also STAR Methods.
(C) Example traces of ASAP1 fluorescence intensity (DF/F) in the spines and somata in (B).
(D and F) GCaMP6s expression in Con (D) and 48 h TTX neurons (F). Active spines with Ca2+ transients (during 5 min of imaging) are indicated (red dots). Scale bar,
10 mm. See also STAR Methods.
(E and G) GaMP6s signal (spontaneous Ca2+ signals were recorded in the presence of TTX) from active spines (y axis, red dots in D and F) in Con (E) and 48 h TTX
neurons (G).
(H–K) Comparison of synaptic Ca2+ transients between Con (black) and TTX (red) neurons. Shown are (H) number of active puncta (putative active spines) during
5 min of recording, (I) mean amplitude, (J) mean duration, and (K) normalized total fluorescence in Con and TTX (48 h) neurons (n = 14–15).
(L) Quantification of Ca2+ transient amplitudes in dendritic spines with wash on PhTx, APV, thapsigargin, or nimodipine (n = 8).
(M and N) AP waveforms (M) and pooled half-widths (N) after sham (black), 48 h TTX (red), or 48 h TTX with co-application of PhTx (green). Scale bars, 20 mV, 1 ms.
For (H)–(L) and (N), data are represented as mean ± SEM; *p < 0.05; **p < 0.01; N.S. represents p > 0.05. See also Figure S6.

Cell 181, 1547–1565, June 25, 2020 1557

 ll
Article

(legend on next page)

1558 Cell 181, 1547–1565, June 25, 2020

 ll
Article

the cytosol; its nuc/cyto ratio increased by more than 50%
(Figures 7I and 7J). bCaMKK can phosphorylate CaMKIV and
is thus a plausible mediator of cytosol-to-nucleus signaling and
nuclear CaMKIV activation. We probed bCaMKK’s involvement
by shRNA knockdown. This completely blocked inactivity-
induced reduction of E29 inclusion, whereas concomitantly ex-
pressing shRNA-resistant bCaMKK fully rescued it (Figure 7K
and 7L). Thus, bCaMKK translocates to the nucleus and is
necessary for chronic inactivity-induced E29 splicing.
DISCUSSION
We found an unexpected mechanism by which excitatory neu-
rons modify their APs in responding homeostatically to chronic
inactivity (Figure 7M). The adaptation arises from a well-defined
change in splicing and is mediated by a novel signaling cascade
that mobilizes Hebbian-type signaling, even in the absence of
spikes. Remarkably, each element of this signaling pathway
has been genetically implicated in neuropsychiatric disease
(see below).
Regulation of BK Channel Splicing Selectively
Controls APD
We found that splicing of BK channels is necessary and sufficient
for inactivity-induced lengthening of APD. Under typical circum-
stances, regulatory changes in one repolarizing current (e.g., BK)
would be offset by compensatory changes in others (e.g., KV2)
(Kimm et al., 2015). Here, however, the key findings are null ef-
fects (Figures 1C, 1D, 1J, 1K, 6M, and 6N); without changes in
spike waveform, altered recruitment of other voltage-gated
channels is not expected. Kimm et al. (2015) have further shown
that blockade of BK channels with IbTx spares the current-fre-
quency relationship. Thus, mechanisms other than BK regulation
are expected and seen for homeostatic adjustments of firing fre-
quency (Lee and Chung, 2014; Lee et al., 2015). For APD, we
cannot exclude ancillary changes enabled by E29 inclusion,
such as BK channel phosphorylation or modified auxiliary sub-
units. However, the brain-dominant BK auxiliary subunit b4 is
not significantly altered by TTX treatment (Lee et al., 2015).
CaV1-CaMKIV Signaling Is Engaged by Chronic Inactivity
and Acute Depolarization
Elegant work shows how AS can be regulated by depolarization
or activity elevation (Ding et al., 2017; Iijima et al., 2011; Mauger
et al., 2016; Xie and Black, 2001). Here we show that AS is also
controlled by chronic inactivity and dominates homeostatic
regulation of AP shape. Strikingly, chronic inactivity-induced
AS of BK, although homeostatic in outcome, relies on CaV1-
CaMKK-CaMKIV signaling, like acute depolarization-induced
splicing. Activation of Ca2+ signaling by silencing neuronal
spiking is counterintuitive; the expectation is that Ca2+ entry
would be dampened (Bridi et al., 2018). We resolved the paradox
by optically tracking transient depolarizations in dendritic spines
of ‘‘inactive’’ neurons (Figures 6A–6C), which were strong
enough to activate the CaV1 channels present in dendritic spines
(Stanika et al., 2016). This demystified the recruitment of CaV1
signaling, already indicated pharmacologically.
A potent set of synaptic events combine to drive homeostatic
readjustment of APD. Chronic spike blockade drives enhanced
spontaneous presynaptic vesicle release (Jakawich et al.,
2010; Lindskog et al., 2010) and incorporation of postsynaptic,
high-conductance, PhTx-sensitive GluA1 receptors (Kim and
Ziff, 2014; Thiagarajan et al., 2005). This increases glutamate-
gated cation influx, drives greater spine depolarization, and re-
cruits L-type Ca2+ channels in spine heads (Obermair et al.,
2004; Yasuda et al., 2003), Mg2+-unblocked NMDARs, and
Ca2+ release from internal stores (Dittmer et al., 2017). Our sce-
nario (Figure 7M) accounts for L-type channel participation in
responses triggered by chronic TTX; L-type channel involvement
in responses to chronic depolarization is well known (O’Leary
et al., 2010). Ca2+ channel-triggered signaling in response to
chronic inactivity and depolarization runs counter to conven-
tional expectations of diametrically opposing effector actions.
Our data suggest that L-type-dependent signaling can be

Figure 7. CaV1-CaMK Signaling Is Required for Chronic Spike Blockade-Induced Nova-2 Translocation and BK Channel AS
(A) Expression level of pCaMKIV and CaMKIV in neurons treated with nimodipine, KN93, or STO-609.
(B) Quantification of CaMKIV activation from (A) (n = 3).
(C) Western blot indicating nuclear Nova-2 levels from neurons treated as indicated.
(D) Quantification of the nuclear Nova-2 level from (C) (n = 3).
(E) E29 splicing after 48 h TTX with other treatments as indicated.
(F) Quantification of E29 splicing from (E) (n = 4).
(G) E29 splicing after chronic depolarization (60K for 24 h) with or without nimodipine or KN93.
(H) Quantification of E29 splicing from (G) (n = 4).
(I) bCaMKK immunostaining (green) with MAP2 (red) and DAPI (blue). Nuclei are indicated by white dashed circles. Scale bar, 10 mm.
(J) Fold changes of aCaMKK and bCaMKK expression in the nucleus, cytosol, and nuc/cyt ratio after TTX (48 h) relative to Con neuron levels (n = 30).
(K) E29 splicing from neurons expressing viral shRNA constructs against endogenous bCaMKK with or without co-expression of shRNA-resistant bCaMKK
(bCaMKK (R)). Scrambled shRNA was used as Con.
(L) Quantification of E29 splicing from (K) (n = 4).
(M) Diagram of the homeostatic signaling loop regulating APD. (1) Chronic spike blockade leads to upregulation of synaptic GluA1 (synaptic scaling). (2) Activation
of GluA1 mediates excessive cation influx, induces membrane depolarization in dendritic spines, and facilitates opening of the NMDAR and CaV1 channels. (3)
CaV1 opening leads to activation of CaMKs, including CaMKI, CaMKII, and bCaMKK. (4) bCaMKK translocates from the cytosol to nucleus and activates CaMKIV.
(5) Activated CaMKIV phosphorylates nuclear Nova-2. (6) Phosphorylated Nova-2 translocates to the cytosol. (7) This leads to reduced E29 inclusion in BK
channel pre-mRNA. (8) BK channel activity is inhibited when lacking E29, broadening APDs, the observed homeostatic response to chronic spike blockade.
(N) The autoregulatory feedback loop of neuronal excitability (LeMasson et al., 1993; Marder et al., 1996; Siegel et al., 1994) with components described in this
study. Various genes implicated in neuropsychiatric diseases are highlighted.
For (B), (D), (F), (H), (J), and (L), data are represented as mean ± SEM; **p < 0.01; N.S. represents p > 0.05. See also Figure S7.

Cell 181, 1547–1565, June 25, 2020 1559

 ll
mobilized to achieve the appropriate response irrespective of the
direction of the initial perturbation.
Strikingly, the CaV1 blocker nimodipine completely abolished
TTX-induced CaMK activation, Nova-2 translocation, and E29
splicing but only partially inhibited the Ca2+ transient in spines
induced by chronic TTX. We suggest that other Ca2+ influx path-
ways contribute to Ca2+ flux (Wheeler et al., 2012), whereas only
nimodipine-sensitive CaV1 channels provide critical voltage-
dependent conformational (VDC) signaling (Li et al., 2016) to
help trap activated CaMKII (Wang et al., 2017).
Multiple CaM Translocators Support Different Forms of
Surface-to-Nucleus Communication
Translocation of a CaMK is a recurrent theme in cytonuclear
signaling in neurons, but the identity of the kinase varies (Co-
hen et al., 2016; Ma et al., 2014). In the present case of
chronic inactivity-induced signaling, sensitivity to the selective
CaMKK inhibitor STO-609 indicated that a CaMKK was
involved. These findings resembled studies in C. elegans
where cytonuclear signaling relies on a monomeric CaMKK
(CKK-1) that translocates across the nuclear membrane (Ki-
mura et al., 2002). Precedent exists for regulated cytonuclear
distribution of bCaMKK (Cao et al., 2011; Karacosta et al.,
2012; Nakamura et al., 2001), but its molecular mechanism
needs elucidation. bCaMKK lacks a classic NLS, but it might
rely on regulation of its nuclear export signal (NES) (Xu
et al., 2012) through interaction with a resident nuclear
protein.
CaMKIV was, as expected, the target of bCaMKK activation,
as judged by elevation of pCaMKIV and its blockade by STO-
609. Finding that CaMKIV was prebound to its substrate
Nova-2 opens up the possibility of local signaling events. This
would minimize the perturbation of nuclear Ca2+ regulation
overall, desirable for signaling extending over hours rather
than seconds to minutes. To be activated by CaMKK, CaMKIV
must be bound to Ca2+/CaM. In one scenario, bCaMKK could
locally transfer Ca2+/CaM to CaMKIV while retaining most of
its catalytic activity, a known feature of this enzyme (Anderson
et al., 1998; Edelman et al., 1996; Tokumitsu and Soderling,
1996). Consistent with such an intranuclear handoff, buffering
of nuclear free Ca2+/CaM by CaMBP4nuc completely inhibited
chronic inactivity-induced Nova-2 translocation (Figures 4D
and 4E).
Roles of CaMKIV in Diverse Aspects of Homeostatic
Plasticity
CaMKIV has been identified previously as a key player in ho-
meostatic plasticity of excitatory neurons by Ibata et al. (2008)
and Joseph and Turrigiano (2017), who found that the synap-
tic response to 4-h TTX treatment, a relatively brief period of
inactivity, could be mimicked by CaMKIV inhibition using
STO-609 or overexpression of a dominant-negative CaMKIV.
Those findings might seem at odds with enhanced pCaMKIV
in neurons undergoing chronic inactivity (Figures 4 and 7).
However, reconciliation may be possible based on the dy-
namics of the homeostatic response. During TTX treatment
for 24–48 h, Kim and Ziff (2014) found initial inhibition of
Ca2+ signaling for at least 6 h, followed by late activation at
1560 Cell 181, 1547–1565, June 25, 2020
Article
48 h, indicating that homeostatic modulation occurs in two
phases. Perhaps early inhibition of CaMKIV, together with in-
hibition of calcineurin, contributes to up-scaling of AMPARs,
critical for activation of synaptic NMDARs and CaV1; with
further inactivity, enhanced synaptic events lead to recruit-
ment of calcium signaling and enhanced nuclear pCaMKIV, vi-
tal for regulation of APD.
Regulation of AS by CaMKIV has been studied previously
in the context of acute depolarization, acting at conserved CaM-
KIV-responsive RNA element (CaRRE) sequences in target pre-
cursor mRNAs (pre-mRNAs) (Iijima et al., 2011; Lee et al.,
2007; Liu et al., 2012; Xie and Black, 2001). We looked for regu-
lation of inclusion of stress-axis-regulated (STREX) exon (X4
in our study) and for CaRRE sequences flanking E29 but did
not find either (Figure S3). This is not surprising in light of evi-
dence that multiple activity-dependent splicing factors (hnRNP
L, SAM68, and related STAR [signal transduction and activation
of RNA] family members) have specific effects on individual pre-
mRNA targets (Iijima et al., 2011). For Nova-2 and other splicing
factors, puzzles remain regarding how selective control could be
exerted if the splicing events are controlled by the same regu-
lator, CaMKIV. Possible differences in the dynamics and locali-
zation of specific splicing events need further study.
Similar to Rbfox1 (Lee et al., 2009), Nova function is regulated
by cytonuclear relocation (Racca et al., 2010). Nova-2 posi-
tioning is a U-shaped function of activity level, with nuclear exit
favored by hyperactivity (seizure induction [Eom et al., 2013] or
sustained depolarization; Figure S7E) or chronic inactivity. This
echoes the U-shaped relationship of CaMKIV activation to activ-
ity. Although we focused on Nova-2 actions in the nucleus,
Nova-2 has also been found in dendrites, co-localized with its
target mRNAs (Racca et al., 2010). Outside of the nucleus,
splicing factors might facilitate translocation of their binding
partners (e.g., CaMKIV) and regulate the stability and translation
of their target mRNAs (see Lee et al., 2016, for such a role of
Rbfox1).
Implications for Autism, Schizophrenia, and Other
Neuropsychiatric Diseases
In the specific feedback loop we circumnavigated (Figures 7M
and 7N), each element is genetically implicated in autism
spectrum disorder (ASD), schizophrenia, or other neuropsychi-
atric disorders. First, L-type Ca2+ channels have been impli-
cated repeatedly in ASD and schizophrenia (Bhat et al., 2012;
Purcell et al., 2014; Splawski et al., 2004, 2005) and affect
downstream signaling of many forms (Deisseroth et al., 2003).
Second, bCaMKK (gene name CAMKK2) exhibits sporadic
mutations in individuals with schizophrenia with severe
biochemical effects (Luo et al., 2014; O’Brien et al., 2017) and
affect multiple target kinases (Marcelo et al., 2016). Third,
Nova-2 mediates splicing of hundreds of pre-mRNAs, encod-
ing proteins prominent at synaptic sites (Saito et al., 2016;
Ule et al., 2005) whose patterns are altered in Fragile X, a
form of ASD (Kuwano et al., 2011; Lewis et al., 2000). Finally,
KCNMA1, encoding the BK channel, is causally involved in
certain sporadic forms of ASD (Laumonnier et al., 2006), mental
retardation, schizophrenia, and epilepsy (Du et al., 2005; Hig-
gins et al., 2008; Zhang et al., 2006); inclusion of BK exon E29

Article
is significantly lower in ASD samples than in matched controls
(Parikshak et al., 2016). Together, these findings suggest that
individual steps in the adaptive feedback pathway not only
contribute to homeostatic regulation but might also go awry
in neuropsychiatric disorders (Mullins et al., 2016; Figure 7N).
In pinpointing players in E-AS coupling that seem to support
pathogenesis, our results align with generally altered splicing
in ASD (Parikshak et al., 2016), possibly arising from this kind
of regulatory loop or side branches from it. Thus, correction
of faulty E-AS coupling merits consideration as a therapeutic
strategy (Hébert et al., 2014).
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d LEAD CONTACT AND MATERIALS AVAILABILITY
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Cell lines
B Primary cell cultures
B Animals
d METHOD DETAILS
B Constructs
B Transfection and treatment of cortical neurons
B Lentiviral transduction of cortical neurons
B Immunocytochemistry and image acquisition and
analysis
B Transfection and electrophysiology of HEK293 cells
B Electrophysiological recording of action potential half-
width in cultured cortical neurons
B Recording Current-Voltage relationships in HEK cells
transfected with BK channels
B Ca
2+
imaging on dendrite and soma of cultured
neurons
B Voltage imaging of cultured neurons
B Immunoprecipitation
B Protein sample preparation and western blot
B Oligo-RNA pull-down
B RNA immunoprecipitation
B RT-PCR and Realtime qPCR
B Thalamic input elimination and immunohistochemistry
B Monocular deprivation
B Protein sample preparation and mass spectrometry
analysis
B Phosphopeptide identification and quantitation
d QUANTIFICATION AND STATISTICAL ANALYSIS
d DATA AND CODE AVAILABILITY
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.013.
ACKNOWLEDGMENTS
We thank Dr. Robert B. Darnell for providing human anti-Nova serum and Dr.
Peter Stoilov and Dr. Douglas Black for providing AS reporter constructs. We
ll
thank Xiaohan Wang and other Tsien lab members for advice and comments
on the manuscript. This work was supported by research grants from the
NIGMS (GM058234), NINDS (NS24067), NIMH (MH071739), NIDA
(DA040484), Druckenmiller Foundation, Simons Foundation, Mathers Founda-
tion, and Burnett Family Foundation (to R.W.T.); the National Key R&D Program
of China (2018YFA0108300 to B.L.); the National Natural Science Foundation
of China (81622016 and 31571034 to B.L. and 81871048 and 81741063 to
L.H.); the Guangdong Natural Science Foundation (Grants for Distinguished
Young Scholars 2015A030306019 to B.L. and 2018B030311034 to L.H.);
and the Guangdong Provincial Key R&D Programs (Key Technologies for
Treatment of Brain Disorders 2018B030332001 and Development of New
Tools for Diagnosis and Treatment of Autism 2018B030335001 to L.H.
and B.L.).
AUTHOR CONTRIBUTIONS
B.L. and R.W.T. conceived the project. B.S.S., S.D.S., and Z.L. performed
electrophysiology experiments. B.L. and C.W. performed subcellular fraction-
ation and immunostaining. N.C. provided analysis tools for calcium and
voltage imaging. N.J.M. and S.D.S. performed immunostaining and analysis
for GluA1. G.Z. and T.A.N. performed mass spectrometry. B.W. and G.F. per-
formed thalamic input elimination and immunohistochemistry for Nova. G.T.,
S.S., and S.D.S. performed cell culture. S.Y. and L.H. performed qRT-PCR
and immunostaining. B.L. performed all other experiments and data analysis.
B.L., R.W.T., and S.D.S. wrote the manuscript with advice from L.H. and G.F.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: April 24, 2019
Revised: January 28, 2020
Accepted: May 4, 2020
Published: June 2, 2020
REFERENCES
Anderson, K.A., Means, R.L., Huang, Q.H., Kemp, B.E., Goldstein, E.G., Sel-
bert, M.A., Edelman, A.M., Fremeau, R.T., and Means, A.R. (1998). Compo-
nents of a calmodulin-dependent protein kinase cascade. Molecular cloning,
functional characterization and cellular localization of Ca2+/calmodulin-depen-
dent protein kinase kinase b. J. Biol. Chem. 273, 31880–31889.
Bhat, S., Dao, D.T., Terrillion, C.E., Arad, M., Smith, R.J., Soldatov, N.M., and
Gould, T.D. (2012). CACNA1C (Cav1.2) in the pathophysiology of psychiatric
disease. Prog. Neurobiol. 99, 1–14.
Bischofberger, J., Geiger, J.R., and Jonas, P. (2002). Timing and efficacy of
Ca2+ channel activation in hippocampal mossy fiber boutons. J. Neurosci.
22, 10593–10602.
Bito, H., Deisseroth, K., and Tsien, R.W. (1996). CREB phosphorylation and
dephosphorylation: a Ca(2+)- and stimulus duration-dependent switch for hip-
pocampal gene expression. Cell 87, 1203–1214.
Black, D.L. (2003). Mechanisms of alternative pre-messenger RNA splicing.
Annu. Rev. Biochem. 72, 291–336.
Borst, J.G., and Sakmann, B. (1998). Calcium current during a single action po-
tential in a large presynaptic terminal of the rat brainstem. J. Physiol. 506,
143–157.
Bridi, M.C.D., de Pasquale, R., Lantz, C.L., Gu, Y., Borrell, A., Choi, S.Y., He,
K., Tran, T., Hong, S.Z., Dykman, A., et al. (2018). Two distinct mechanisms for
experience-dependent homeostasis. Nat. Neurosci. 21, 843–850.
Buckanovich, R.J., and Darnell, R.B. (1997). The neuronal RNA binding protein
Nova-1 recognizes specific RNA targets in vitro and in vivo. Mol. Cell. Biol. 17,
3194–3201.
Byrne, J.H., and Kandel, E.R. (1996). Presynaptic facilitation revisited: state
and time dependence. J. Neurosci. 16, 425–435.
Cell 181, 1547–1565, June 25, 2020 1561
 ll
Cao, W., Sohail, M., Liu, G., Koumbadinga, G.A., Lobo, V.G., and Xie, J. (2011).
Differential effects of PKA-controlled CaMKK2 variants on neuronal differenti-
ation. RNA Biol. 8, 1061–1072.
Cohen, S.M., Ma, H., Kuchibhotla, K.V., Watson, B.O., Buzsáki, G., Froemke,
R.C., and Tsien, R.W. (2016). Excitation-transcription coupling in parvalbumin-
positive interneurons employs a novel CaM kinase-dependent pathway
distinct from excitatory neurons. Neuron 90, 292–307.
Cruzalegui, F.H., and Means, A.R. (1993). Biochemical characterization of the
multifunctional Ca2+/calmodulin-dependent protein kinase type IV expressed
in insect cells. J. Biol. Chem. 268, 26171–26178.
Deisseroth, K., Mermelstein, P.G., Xia, H., and Tsien, R.W. (2003). Signaling
from synapse to nucleus: the logic behind the mechanisms. Curr. Opin. Neuro-
biol. 13, 354–365.
Deng, P.Y., Rotman, Z., Blundon, J.A., Cho, Y., Cui, J., Cavalli, V., Zakharenko,
S.S., and Klyachko, V.A. (2013). FMRP regulates neurotransmitter release and
synaptic information transmission by modulating action potential duration via
BK channels. Neuron 77, 696–711.
Desai, N.S., Rutherford, L.C., and Turrigiano, G.G. (1999). Plasticity in the
intrinsic excitability of cortical pyramidal neurons. Nat. Neurosci. 2,
515–520.
Ding, X., Liu, S., Tian, M., Zhang, W., Zhu, T., Li, D., Wu, J., Deng, H., Jia,
Y., Xie, W., et al. (2017). Activity-induced histone modifications govern
Neurexin-1 mRNA splicing and memory preservation. Nat. Neurosci. 20,
690–699.
Dittmer, P.J., Wild, A.R., Dell’Acqua, M.L., and Sather, W.A. (2017). STIM1
Ca2+ Sensor Control of L-type Ca2+-Channel-Dependent Dendritic Spine
Structural Plasticity and Nuclear Signaling. Cell Rep. 19, 321–334.
Du, W., Bautista, J.F., Yang, H., Diez-Sampedro, A., You, S.A., Wang, L., Ko-
tagal, P., Lüders, H.O., Shi, J., Cui, J., et al. (2005). Calcium-sensitive potas-
sium channelopathy in human epilepsy and paroxysmal movement disorder.
Nat. Genet. 37, 733–738.
Edelman, A.M., Mitchelhill, K.I., Selbert, M.A., Anderson, K.A., Hook, S.S., Sta-
pleton, D., Goldstein, E.G., Means, A.R., and Kemp, B.E. (1996). Multiple
Ca(2+)-calmodulin-dependent protein kinase kinases from rat brain. Purifica-
tion, regulation by Ca(2+)-calmodulin, and partial amino acid sequence.
J. Biol. Chem. 271, 10806–10810.
Ehlers, M.D. (2003). Activity level controls postsynaptic composition and
signaling via the ubiquitin-proteasome system. Nat. Neurosci. 6,
231–242.
Emptage, N., Bliss, T.V., and Fine, A. (1999). Single synaptic events evoke
NMDA receptor-mediated release of calcium from internal stores in hippocam-
pal dendritic spines. Neuron 22, 115–124.
Eom, T., Zhang, C., Wang, H., Lay, K., Fak, J., Noebels, J.L., and Darnell, R.B.
(2013). NOVA-dependent regulation of cryptic NMD exons controls synaptic
protein levels after seizure. eLife 2, e00178.
Fodor, A.A., and Aldrich, R.W. (2009). Convergent evolution of alternative
splices at domain boundaries of the BK channel. Annu. Rev. Physiol.
71, 19–36.
Furlanis, E., and Scheiffele, P. (2018). Regulation of neuronal differentiation,
function, and plasticity by alternative splicing. Annu. Rev. Cell Dev. Biol. 34,
451–469.
Geiger, J.R., and Jonas, P. (2000). Dynamic control of presynaptic Ca(2+)
inflow by fast-inactivating K(+) channels in hippocampal mossy fiber boutons.
Neuron 28, 927–939.
Green, M.F., Anderson, K.A., and Means, A.R. (2011a). Characterization of the
CaMKKb-AMPK signaling complex. Cell. Signal. 23, 2005–2012.
Green, M.F., Scott, J.W., Steel, R., Oakhill, J.S., Kemp, B.E., and Means, A.R.
(2011b). Ca2+/Calmodulin-dependent protein kinase kinase b is regulated by
multisite phosphorylation. J. Biol. Chem. 286, 28066–28079.
Ha, T.S., Jeong, S.Y., Cho, S.W., Jeon, Hk., Roh, G.S., Choi, W.S., and Park,
C.S. (2000). Functional characteristics of two BKCa channel variants differen-
tially expressed in rat brain tissues. Eur. J. Biochem. 267, 910–918.
1562 Cell 181, 1547–1565, June 25, 2020
Article
Harnett, M.T., Makara, J.K., Spruston, N., Kath, W.L., and Magee, J.C. (2012).
Synaptic amplification by dendritic spines enhances input cooperativity. Na-
ture 491, 599–602.
Harreman, M.T., Kline, T.M., Milford, H.G., Harben, M.B., Hodel, A.E., and Cor-
bett, A.H. (2004). Regulation of nuclear import by phosphorylation adjacent to
nuclear localization signals. J. Biol. Chem. 279, 20613–20621.
Hébert, B., Pietropaolo, S., Même, S., Laudier, B., Laugeray, A., Doisne, N.,
Quartier, A., Lefeuvre, S., Got, L., Cahard, D., et al. (2014). Rescue of fragile
X syndrome phenotypes in Fmr1 KO mice by a BKCa channel opener molecule.
Orphanet J. Rare Dis. 9, 124.
Helton, T.D., Xu, W., and Lipscombe, D. (2005). Neuronal L-type calcium
channels open quickly and are inhibited slowly. J. Neurosci. 25,
10247–10251.
Higgins, J.J., Hao, J., Kosofsky, B.E., and Rajadhyaksha, A.M. (2008). Dysre-
gulation of large-conductance Ca2+-activated K+ channel expression in non-
syndromal mental retardation due to a cereblon p.R419X mutation. Neuroge-
netics 9, 219–223.
Hodgkin, A.L., and Huxley, A.F. (1952). A quantitative description of membrane
current and its application to conduction and excitation in nerve. J. Physiol.
117, 500–544.
Hu, H., Shao, L.R., Chavoshy, S., Gu, N., Trieb, M., Behrens, R., Laake, P.,
Pongs, O., Knaus, H.G., Ottersen, O.P., and Storm, J.F. (2001). Presynaptic
Ca2+-activated K+ channels in glutamatergic hippocampal terminals and their
role in spike repolarization and regulation of transmitter release. J. Neurosci.
21, 9585–9597.
Ibata, K., Sun, Q., and Turrigiano, G.G. (2008). Rapid synaptic scaling induced
by changes in postsynaptic firing. Neuron 57, 819–826.
Iijima, T., Wu, K., Witte, H., Hanno-Iijima, Y., Glatter, T., Richard, S., and
Scheiffele, P. (2011). SAM68 regulates neuronal activity-dependent alternative
splicing of neurexin-1. Cell 147, 1601–1614.
Jackson, M.B., Konnerth, A., and Augustine, G.J. (1991). Action potential
broadening and frequency-dependent facilitation of calcium signals in pituitary
nerve terminals. Proc. Natl. Acad. Sci. USA 88, 380–384.
Jakawich, S.K., Nasser, H.B., Strong, M.J., McCartney, A.J., Perez, A.S., Ra-
kesh, N., Carruthers, C.J., and Sutton, M.A. (2010). Local presynaptic activity
gates homeostatic changes in presynaptic function driven by dendritic BDNF
synthesis. Neuron 68, 1143–1158.
Jiang, M., and Chen, G. (2006). High Ca2+-phosphate transfection efficiency in
low-density neuronal cultures. Nat. Protoc. 1, 695–700.
Joseph, A., and Turrigiano, G.G. (2017). All for one but not one for all: excitatory
synaptic scaling and intrinsic excitability are coregulated by CaMKIV, whereas
inhibitory synaptic scaling is under independent control. J. Neurosci. 37,
6778–6785.
Karacosta, L.G., Foster, B.A., Azabdaftari, G., Feliciano, D.M., and Edelman,
A.M. (2012). A regulatory feedback loop between Ca2+/calmodulin-dependent
protein kinase kinase 2 (CaMKK2) and the androgen receptor in prostate can-
cer progression. J. Biol. Chem. 287, 24832–24843.
Kim, J., and Tsien, R.W. (2008). Synapse-specific adaptations to inactivity in
hippocampal circuits achieve homeostatic gain control while dampening
network reverberation. Neuron 58, 925–937.
Kim, S., and Ziff, E.B. (2014). Calcineurin mediates synaptic scaling via syn-
aptic trafficking of Ca2+-permeable AMPA receptors. PLoS Biol. 12,
e1001900.
Kimm, T., Khaliq, Z.M., and Bean, B.P. (2015). Differential regulation of action
potential shape and burst-frequency firing by BK and Kv2 channels in substan-
tia nigra dopaminergic neurons. J. Neurosci. 35, 16404–16417.
Kimura, Y., Corcoran, E.E., Eto, K., Gengyo-Ando, K., Muramatsu, M.A., Ko-
bayashi, R., Freedman, J.H., Mitani, S., Hagiwara, M., Means, A.R., and Toku-
mitsu, H. (2002). A CaMK cascade activates CRE-mediated transcription in
neurons of Caenorhabditis elegans. EMBO Rep. 3, 962–966.
Kuwano, Y., Kamio, Y., Kawai, T., Katsuura, S., Inada, N., Takaki, A., and Ro-
kutan, K. (2011). Autism-associated gene expression in peripheral leucocytes

Article
commonly observed between subjects with autism and healthy women having
autistic children. PLoS ONE 6, e24723.
Laumonnier, F., Roger, S., Guérin, P., Molinari, F., M’rad, R., Cahard, D., Bel-
hadj, A., Halayem, M., Persico, A.M., Elia, M., et al. (2006). Association of a
functional deficit of the BKCa channel, a synaptic regulator of neuronal excit-
ability, with autism and mental retardation. Am. J. Psychiatry 163,
1622–1629.
Lee, K.Y., and Chung, H.J. (2014). NMDA receptors and L-type voltage-gated
Ca2+ channels mediate the expression of bidirectional homeostatic intrinsic
plasticity in cultured hippocampal neurons. Neuroscience 277, 610–623.
Lee, U.S., and Cui, J. (2010). BK channel activation: structural and functional
insights. Trends Neurosci. 33, 415–423.
Lee, J.A., Xing, Y., Nguyen, D., Xie, J., Lee, C.J., and Black, D.L. (2007). Depo-
larization and CaM kinase IV modulate NMDA receptor splicing through two
essential RNA elements. PLoS Biol. 5, e40.
Lee, J.A., Tang, Z.Z., and Black, D.L. (2009). An inducible change in Fox-1/
A2BP1 splicing modulates the alternative splicing of downstream neuronal
target exons. Genes Dev. 23, 2284–2293.
Lee, K.Y., Royston, S.E., Vest, M.O., Ley, D.J., Lee, S., Bolton, E.C., and
Chung, H.J. (2015). N-methyl-D-aspartate receptors mediate activity-depen-
dent down-regulation of potassium channel genes during the expression of
homeostatic intrinsic plasticity. Mol. Brain 8, 4.
Lee, J.A., Damianov, A., Lin, C.H., Fontes, M., Parikshak, N.N., Anderson, E.S.,
Geschwind, D.H., Black, D.L., and Martin, K.C. (2016). Cytoplasmic Rbfox1
regulates the expression of synaptic and autism-related genes. Neuron 89,
113–128.
LeMasson, G., Marder, E., and Abbott, L.F. (1993). Activity-dependent regula-
tion of conductances in model neurons. Science 259, 1915–1917.
Lewis, H.A., Musunuru, K., Jensen, K.B., Edo, C., Chen, H., Darnell, R.B., and
Burley, S.K. (2000). Sequence-specific RNA binding by a Nova KH domain: im-
plications for paraneoplastic disease and the fragile X syndrome. Cell 100,
323–332.
Li, Q., Lee, J.A., and Black, D.L. (2007). Neuronal regulation of alternative pre-
mRNA splicing. Nat. Rev. Neurosci. 8, 819–831.
Li, B., Jie, W., Huang, L., Wei, P., Li, S., Luo, Z., Friedman, A.K., Meredith, A.L.,
Han, M.H., Zhu, X.H., and Gao, T.M. (2014). Nuclear BK channels regulate
gene expression via the control of nuclear calcium signaling. Nat. Neurosci.
17, 1055–1063.
Li, B., Tadross, M.R., and Tsien, R.W. (2016). Sequential ionic and conforma-
tional signaling by calcium channels drives neuronal gene expression. Science
351, 863–867.
Licatalosi, D.D., and Darnell, R.B. (2006). Splicing regulation in neurologic dis-
ease. Neuron 52, 93–101.
Licatalosi, D.D., and Darnell, R.B. (2010). RNA processing and its regulation:
global insights into biological networks. Nat. Rev. Genet. 11, 75–87.
Lindskog, M., Li, L., Groth, R.D., Poburko, D., Thiagarajan, T.C., Han, X., and
Tsien, R.W. (2010). Postsynaptic GluA1 enables acute retrograde enhance-
ment of presynaptic function to coordinate adaptation to synaptic inactivity.
Proc. Natl. Acad. Sci. USA 107, 21806–21811.
Liu, G., Razanau, A., Hai, Y., Yu, J., Sohail, M., Lobo, V.G., Chu, J., Kung, S.K.,
and Xie, J. (2012). A conserved serine of heterogeneous nuclear ribonucleo-
protein L (hnRNP L) mediates depolarization-regulated alternative splicing of
potassium channels. J. Biol. Chem. 287, 22709–22716.
Llinás, R., Sugimori, M., and Simon, S.M. (1982). Transmission by presynaptic
spike-like depolarization in the squid giant synapse. Proc. Natl. Acad. Sci. USA
79, 2415–2419.
Luo, X.J., Li, M., Huang, L., Steinberg, S., Mattheisen, M., Liang, G., Donohoe,
G., Shi, Y., Chen, C., Yue, W., et al.; MooDS SCZ Consortium (2014). Conver-
gent lines of evidence support CAMKK2 as a schizophrenia susceptibility
gene. Mol. Psychiatry 19, 774–783.
Ma, W.P., Li, Y.T., and Tao, H.W. (2013). Downregulation of cortical inhibition
mediates ocular dominance plasticity during the critical period. J. Neurosci.
33, 11276–11280.
ll
Ma, H., Groth, R.D., Cohen, S.M., Emery, J.F., Li, B., Hoedt, E., Zhang, G.,
Neubert, T.A., and Tsien, R.W. (2014). gCaMKII shuttles Ca2+/CaM to the nu-
cleus to trigger CREB phosphorylation and gene expression. Cell 159,
281–294.
Maffei, A., and Turrigiano, G.G. (2008). Multiple modes of network homeosta-
sis in visual cortical layer 2/3. J. Neurosci. 28, 4377–4384.
Maghsoodi, B., Poon, M.M., Nam, C.I., Aoto, J., Ting, P., and Chen, L. (2008).
Retinoic acid regulates RARalpha-mediated control of translation in dendritic
RNA granules during homeostatic synaptic plasticity. Proc. Natl. Acad. Sci.
USA 105, 16015–16020.
Marcelo, K.L., Means, A.R., and York, B. (2016). The Ca(2+)/Calmodulin/
CaMKK2 axis: nature’s metabolic CaMshaft. Trends Endocrinol. Metab. 27,
706–718.
Marder, E., Abbott, L.F., Turrigiano, G.G., Liu, Z., and Golowasch, J. (1996).
Memory from the dynamics of intrinsic membrane currents. Proc. Natl.
Acad. Sci. USA 93, 13481–13486.
Matthews, E.A., Linardakis, J.M., and Disterhoft, J.F. (2009). The fast and slow
afterhyperpolarizations are differentially modulated in hippocampal neurons
by aging and learning. J. Neurosci. 29, 4750–4755.
Mauger, O., Lemoine, F., and Scheiffele, P. (2016). Targeted intron retention
and excision for rapid gene regulation in response to neuronal activity. Neuron
92, 1266–1278.
Mayer, M.L., Westbrook, G.L., and Guthrie, P.B. (1984). Voltage-dependent
block by Mg2+ of NMDA responses in spinal cord neurones. Nature 309,
261–263.
Mullins, C., Fishell, G., and Tsien, R.W. (2016). Unifying views of autism spec-
trum disorders: a consideration of autoregulatory feedback loops. Neuron 89,
1131–1156.
Nakamura, Y., Okuno, S., Sato, F., and Fujisawa, H. (1995). An immunohisto-
chemical study of Ca2+/calmodulin-dependent protein kinase IV in the rat cen-
tral nervous system: light and electron microscopic observations. Neurosci-
ence 68, 181–194.
Nakamura, Y., Okuno, S., Kitani, T., Otake, K., Sato, F., and Fujisawa, H.
(2001). Immunohistochemical localization of Ca(2+)/calmodulin-dependent
protein kinase kinase b in the rat central nervous system. Neurosci. Res. 39,
175–188.
Nowak, L., Bregestovski, P., Ascher, P., Herbet, A., and Prochiantz, A. (1984).
Magnesium gates glutamate-activated channels in mouse central neurones.
Nature 307, 462–465.
O’Brien, M.T., Oakhill, J.S., Ling, N.X., Langendorf, C.G., Hoque, A., Dite, T.A.,
Means, A.R., Kemp, B.E., and Scott, J.W. (2017). Impact of genetic variation
on human CaMKK2 regulation by Ca2+-calmodulin and multisite phosphoryla-
tion. Sci. Rep. 7, 43264.
O’Leary, T., van Rossum, M.C., and Wyllie, D.J. (2010). Homeostasis of
intrinsic excitability in hippocampal neurones: dynamics and mechanism of
the response to chronic depolarization. J. Physiol. 588, 157–170.
O’Leary, T., Williams, A.H., Franci, A., and Marder, E. (2014). Cell types,
network homeostasis, and pathological compensation from a biologically
plausible ion channel expression model. Neuron 82, 809–821.
Obermair, G.J., Szabo, Z., Bourinet, E., and Flucher, B.E. (2004). Differential
targeting of the L-type Ca2+ channel a 1C (CaV1.2) to synaptic and extrasy-
naptic compartments in hippocampal neurons. Eur. J. Neurosci. 19,
2109–2122.
Palmer, L.M., and Stuart, G.J. (2009). Membrane potential changes in dendritic
spines during action potentials and synaptic input. J. Neurosci. 29,
6897–6903.
Parikshak, N.N., Swarup, V., Belgard, T.G., Irimia, M., Ramaswami, G., Gan-
dal, M.J., Hartl, C., Leppa, V., Ubieta, L.T., Huang, J., et al. (2016). Genome-
wide changes in lncRNA, splicing, and regional gene expression patterns in
autism. Nature 540, 423–427.
Penney, J., Tsurudome, K., Liao, E.H., Elazzouzi, F., Livingstone, M., Gonza-
lez, M., Sonenberg, N., and Haghighi, A.P. (2012). TOR is required for the
Cell 181, 1547–1565, June 25, 2020 1563
 ll
retrograde regulation of synaptic homeostasis at the Drosophila neuromus-
cular junction. Neuron 74, 166–178.
Pietrzykowski, A.Z., Friesen, R.M., Martin, G.E., Puig, S.I., Nowak, C.L.,
Wynne, P.M., Siegelmann, H.T., and Treistman, S.N. (2008). Posttranscrip-
tional regulation of BK channel splice variant stability by miR-9 underlies neu-
roadaptation to alcohol. Neuron 59, 274–287.
Purcell, S.M., Moran, J.L., Fromer, M., Ruderfer, D., Solovieff, N., Roussos, P.,
O’Dushlaine, C., Chambert, K., Bergen, S.E., Kähler, A., et al. (2014). A poly-
genic burden of rare disruptive mutations in schizophrenia. Nature 506,
185–190.
Racca, C., Gardiol, A., Eom, T., Ule, J., Triller, A., and Darnell, R.B. (2010). The
Neuronal Splicing Factor Nova Co-Localizes with Target RNAs in the Dendrite.
Front. Neural Circuits 4, 5.
Ramocki, M.B., and Zoghbi, H.Y. (2008). Failure of neuronal homeostasis re-
sults in common neuropsychiatric phenotypes. Nature 455, 912–918.
Rappsilber, J., Mann, M., and Ishihama, Y. (2007). Protocol for micro-purifica-
tion, enrichment, pre-fractionation and storage of peptides for proteomics us-
ing StageTips. Nat. Protoc. 2, 1896–1906.
Sabatini, B.L., and Regehr, W.G. (1997). Control of neurotransmitter release by
presynaptic waveform at the granule cell to Purkinje cell synapse. J. Neurosci.
17, 3425–3435.
Saito, Y., Miranda-Rottmann, S., Ruggiu, M., Park, C.Y., Fak, J.J., Zhong, R.,
Duncan, J.S., Fabella, B.A., Junge, H.J., Chen, Z., et al. (2016). NOVA2-medi-
ated RNA regulation is required for axonal pathfinding during development.
eLife 5, e14371.
Sausbier, M., Hu, H., Arntz, C., Feil, S., Kamm, S., Adelsberger, H., Sausbier,
U., Sailer, C.A., Feil, R., Hofmann, F., et al. (2004). Cerebellar ataxia and Pur-
kinje cell dysfunction caused by Ca2+-activated K+ channel deficiency. Proc.
Natl. Acad. Sci. USA 101, 9474–9478.
Schanzenbächer, C.T., Sambandan, S., Langer, J.D., and Schuman, E.M.
(2016). Nascent Proteome Remodeling following Homeostatic Scaling at Hip-
pocampal Synapses. Neuron 92, 358–371.
Schaukowitch, K., Reese, A.L., Kim, S.K., Kilaru, G., Joo, J.Y., Kavalali, E.T.,
and Kim, T.K. (2017). An Intrinsic Transcriptional Program Underlying Synaptic
Scaling during Activity Suppression. Cell Rep. 18, 1512–1526.
Schneider, C.A., Rasband, W.S., and Eliceiri, K.W. (2012). NIH Image to Im-
ageJ: 25 years of image analysis. Nat. Methods 9, 671–675.
Selbert, M.A., Anderson, K.A., Huang, Q.H., Goldstein, E.G., Means, A.R., and
Edelman, A.M. (1995). Phosphorylation and activation of Ca(2+)-calmodulin-
dependent protein kinase IV by Ca(2+)-calmodulin-dependent protein kinase
Ia kinase. Phosphorylation of threonine 196 is essential for activation. J. Biol.
Chem. 270, 17616–17621.
Shao, L.R., Halvorsrud, R., Borg-Graham, L., and Storm, J.F. (1999). The
role of BK-type Ca2+-dependent K+ channels in spike broadening during
repetitive firing in rat hippocampal pyramidal cells. J. Physiol. 521,
135–146.
Shelley, C., Whitt, J.P., Montgomery, J.R., and Meredith, A.L. (2013). Phos-
phorylation of a constitutive serine inhibits BK channel variants containing
the alternate exon ‘‘SRKR’’. J. Gen. Physiol. 142, 585–598.
Shipston, M.J., and Tian, L. (2016). Posttranscriptional and posttranslational
regulation of BK channels. Int. Rev. Neurobiol. 128, 91–126.
Siegel, M., Marder, E., and Abbott, L.F. (1994). Activity-dependent current
distributions in model neurons. Proc. Natl. Acad. Sci. USA 91,
11308–11312.
Splawski, I., Timothy, K.W., Sharpe, L.M., Decher, N., Kumar, P., Bloise, R.,
Napolitano, C., Schwartz, P.J., Joseph, R.M., Condouris, K., et al. (2004).
Ca(V)1.2 calcium channel dysfunction causes a multisystem disorder including
arrhythmia and autism. Cell 119, 19–31.
Splawski, I., Timothy, K.W., Decher, N., Kumar, P., Sachse, F.B., Beggs, A.H.,
Sanguinetti, M.C., and Keating, M.T. (2005). Severe arrhythmia disorder
caused by cardiac L-type calcium channel mutations. Proc. Natl. Acad. Sci.
USA 102, 8089–8096, discussion 8086–8088.
1564 Cell 181, 1547–1565, June 25, 2020
Article
St-Pierre, F., Marshall, J.D., Yang, Y., Gong, Y., Schnitzer, M.J., and Lin, M.Z.
(2014). High-fidelity optical reporting of neuronal electrical activity with an ul-
trafast fluorescent voltage sensor. Nat. Neurosci. 17, 884–889.
Stanika, R., Campiglio, M., Pinggera, A., Lee, A., Striessnig, J., Flucher, B.E.,
and Obermair, G.J. (2016). Splice variants of the CaV1.3 L-type calcium chan-
nel regulate dendritic spine morphology. Sci. Rep. 6, 34528.
Stoilov, P., Lin, C.H., Damoiseaux, R., Nikolic, J., and Black, D.L. (2008). A
high-throughput screening strategy identifies cardiotonic steroids as
alternative splicing modulators. Proc. Natl. Acad. Sci. USA 105,
11218–11223.
Styr, B., and Slutsky, I. (2018). Imbalance between firing homeostasis and syn-
aptic plasticity drives early-phase Alzheimer’s disease. Nat. Neurosci. 21,
463–473.
Thiagarajan, T.C., Lindskog, M., and Tsien, R.W. (2005). Adaptation to synap-
tic inactivity in hippocampal neurons. Neuron 47, 725–737.
Tokumitsu, H., and Soderling, T.R. (1996). Requirements for calcium and
calmodulin in the calmodulin kinase activation cascade. J. Biol. Chem. 271,
5617–5622.
Trasande, C.A., and Ramirez, J.M. (2007). Activity deprivation leads to sei-
zures in hippocampal slice cultures: is epilepsy the consequence of homeo-
static plasticity? J. Clin. Neurophysiol. 24, 154–164.
Turrigiano, G.G. (2008). The self-tuning neuron: synaptic scaling of excitatory
synapses. Cell 135, 422–435.
Turrigiano, G.G., and Nelson, S.B. (2004). Homeostatic plasticity in the devel-
oping nervous system. Nat. Rev. Neurosci. 5, 97–107.
Ule, J., Ule, A., Spencer, J., Williams, A., Hu, J.S., Cline, M., Wang, H., Clark,
T., Fraser, C., Ruggiu, M., et al. (2005). Nova regulates brain-specific splicing
to shape the synapse. Nat. Genet. 37, 844–852.
Ule, J., Stefani, G., Mele, A., Ruggiu, M., Wang, X., Taneri, B., Gaasterland, T.,
Blencowe, B.J., and Darnell, R.B. (2006). An RNA map predicting Nova-depen-
dent splicing regulation. Nature 444, 580–586.
Vuong, C.K., Black, D.L., and Zheng, S. (2016). The neurogenetics of alterna-
tive splicing. Nat. Rev. Neurosci. 17, 265–281.
Wang, J., Campos, B., Jamieson, G.A., Jr., Kaetzel, M.A., and Dedman,
J.R. (1995). Functional elimination of calmodulin within the nucleus by tar-
geted expression of an inhibitor peptide. J. Biol. Chem. 270,
30245–30248.
Wang, X., Marks, C.R., Perfitt, T.L., Nakagawa, T., Lee, A., Jacobson,
D.A., and Colbran, R.J. (2017). A novel mechanism for Ca2+/calmodulin-
dependent protein kinase II targeting to L-type Ca2+ channels that initi-
ates long-range signaling to the nucleus. J. Biol. Chem. 292,
17324–17336.
Wayman, G.A., Lee, Y.S., Tokumitsu, H., Silva, A.J., and Soderling, T.R. (2008).
Calmodulin-kinases: modulators of neuronal development and plasticity.
Neuron 59, 914–931.
West, A.E., Griffith, E.C., and Greenberg, M.E. (2002). Regulation of transcrip-
tion factors by neuronal activity. Nat. Rev. Neurosci. 3, 921–931.
Wheeler, D.G., Groth, R.D., Ma, H., Barrett, C.F., Owen, S.F., Safa, P., and
Tsien, R.W. (2012). Ca(V)1 and Ca(V)2 channels engage distinct modes of
Ca(2+) signaling to control CREB-dependent gene expression. Cell 149,
1112–1124.
Wiesel, T.N., and Hubel, D.H. (1963). Effects of visual deprivation on
morphology and physiology of cells in the cats lateral geniculate body.
J. Neurophysiol. 26, 978–993.
Wondolowski, J., and Dickman, D. (2013). Emerging links between homeo-
static synaptic plasticity and neurological disease. Front. Cell. Neurosci.
7, 223.
Xie, J., and Black, D.L. (2001). A CaMK IV responsive RNA element mediates
depolarization-induced alternative splicing of ion channels. Nature 410,
936–939.
Xie, J., and McCobb, D.P. (1998). Control of alternative splicing of potassium
channels by stress hormones. Science 280, 443–446.

Article
Xu, D., Farmer, A., Collett, G., Grishin, N.V., and Chook, Y.M. (2012). Sequence
and structural analyses of nuclear export signals in the NESdb database. Mol.
Biol. Cell 23, 3677–3693.
Yang, Y.Y., Yin, G.L., and Darnell, R.B. (1998). The neuronal RNA-binding pro-
tein Nova-2 is implicated as the autoantigen targeted in POMA patients with
dementia. Proc. Natl. Acad. Sci. USA 95, 13254–13259.
ll
Yasuda, R., Sabatini, B.L., and Svoboda, K. (2003). Plasticity of calcium chan-
nels in dendritic spines. Nat. Neurosci. 6, 948–955.
Zarei, M.M., Zhu, N., Alioua, A., Eghbali, M., Stefani, E., and Toro, L. (2001). A
novel MaxiK splice variant exhibits dominant-negative properties for surface
expression. J. Biol. Chem. 276, 16232–16239.
Zhang, L., Li, X., Zhou, R., and Xing, G. (2006). Possible role of potassium
channel, big K in etiology of schizophrenia. Med. Hypotheses 67, 41–43.
Cell 181, 1547–1565, June 25, 2020 1565
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Rabbit pCaMKII Cell Signaling Technology Cat#3361;RRID:AB_10015209
Rabbit pCaMKIV Abcam Cat# ab59424;RRID:AB_2068253
Rabbit pCaMKI Santa Cruz Cat#sc-28438-R;RRID:AB_667968
Goat anti-bCaMKK Santa Cruz Cat#sc-9629;RRID:AB_2243844
Rabbit anti-bCaMKK Santa Cruz Cat#sc-50341;RRID:AB_2068532
Mouse anti-CaMKIV Santa Cruz Cat#sc-55501;RRID:AB_2243836
Mouse anti-CaMKIV Santa Cruz Cat#sc-136249;RRID:AB_2275109
Mouse anti-flag (DYKDDDDK Tag) Cell Signaling Technology Cat#8146;RRID:AB_10950495
Rabbit anti-PSD95 Synaptic Systems Cat#124002; RRID:AB_887760
Mouse anti-PSD95 Synaptic Systems Cat#124011; RRID:AB_10804286
Goat anti-Nova-2 Santa Cruz Cat#sc-10546;RRID:AB_2151558
Human anti-pan NOVA anti-Nova paraneoplastic Saito et al., 2016
human serum
Mouse anti-MAP2 (HM-2) Sigma Cat#M9942;RRID:AB_477256
Mouse anti-CaM Millipore Cat#05-173;RRID:AB_309644
Goat anti-gCaMKII Santa Cruz Cat#sc-1541;RRID:AB_2068234
Mouse anti-CaMKI Santa Cruz Cat#sc-377418;RRID:AB_2069999
Goat anti-bCaMKI Santa Cruz Cat#sc-131452;RRID:AB_2243992
Rabbit anti-dCaMKI Santa Cruz Cat#sc-134638;RRID:AB_2070115
Mouse anti-gCaMKI Abcam Cat#ab77046;RRID:AB_1565944
Rabbit anti-gCaMKI Thermo Scientific Cat#PA5-19661;RRID:AB_10981875
Mouse anti-aCaMKK Santa Cruz Cat#sc-17827;RRID:AB_2275110
Rabbit anti-aCaMKK Santa Cruz Cat#sc-11370;RRID:AB_2068406
Rabbit anti-GluR1 Calbiochem Cat# 04-855;RRID: AB_1977216
Rabbit anti-BK channel (extracellular epitope) Alomone labs Cat#APC151;RRID:AB_10915895
Mouse anti-PSD-95 UC Davis/NIH NeuroMab facility N/A
Guinea anti-MAP2 Synaptic Systems Cat# 188004;RRID:AB_2138181
Rabbit anti-GFP Abcam Cat# ab290;RRID:AB_303395
Rabbit anti-pCaMKII PhosphoSolutions Cat# p1005-286; RRID:AB_2492051
Rabbit anti-Lamin-B1 Cell Signaling Technology Cat#13435;RRID:AB_2737428
Rabbit anti-GAPDH Cell Signaling Technology Cat#5174;RRID:AB_10622025
Chemicals, Peptides, and Recombinant Proteins
Tetrodotoxin (TTX) Ascent Scientific Cat#4368-28-9
KN93 Tocris Cat#5215
STO-609 Tocris Cat#1551
PhTx Tocris Cat#2770
APV Tocris Cat#0106
nimodipine Abcam Cat#ab120138
thapsigargin Tocris Cat#1138
Critical Commercial Assays
Phusion Site-Directed Mutagenesis Kit Thermo Scientific Cat# F541
Dynabeads Protein G Immunoprecipitation Kit Life Technologies Cat#10007D
Magnetic RNA-Protein Pull-Down Kit Thermo Scientific Cat#20164
(Continued on next page)

e1 Cell 181, 1547–1565.e1–e8, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Magna RIP RNA-Binding Protein Sigma Cat#17-700
Immunoprecipitation Kit
Pierce Cell Surface Protein Isolation Kit Thermo Scientific Cat#89881
Pierce Nuclear Protein Extraction Kit Thermo Scientific Cat#78833
CelLytic M Cell Lysis Reagent Sigma Cat#C3228
Experimental Models: Organisms/Strains
Mouse: Olig3-Cre mouse gift from Y. Nakagawa, N/A
University of Minnesota
Mouse: R26floxstopTeNT gift from M. Goulding at the Salk N/A
Institute for Biological Studies
Oligonucleotides
Probes and shRNAs, see Table S1. This paper N/A
Primers for RT-PCR and qPCR, see Table S1 This paper N/A
Primers to insert E29 and flanking introns to the This paper N/A
splicing reporter: F: CCGGAATTCCGGC
TATGTGGCAACCCTAC
Primers to insert E29 and flanking introns to the This paper N/A
splicing reporter: R: CGCGGATCCGCGT
CTCCTTTGACTTCCTCT
Primers to measure E29 splicing in the splicing This paper N/A
reporter: F: GGAGAAGTCTGCCGTTACTGCCC
TGTG (DY-782 labeled)
Primers to measure E29 splicing in the splicing This paper N/A
reporter: R: CCGTCGTCCTTGAAGAAGATGGTGC
Recombinant DNA
mouse bCaMKK construct: Lentiviral CaMKKbeta Green et al., 2011b Addgene Plasmid #33322;
RRID:Addgene_33322
rat bCaMKK 1-460 construct: pSG5-FLAG- Green et al., 2011a Addgene Plasmid #33324;
CaMKKbeta rat 1-460 RRID:Addgene_33324
Human Nova-2 ORF Origene Cat# RC216200L1V, RC216200L2V
pcDNA3-BK-GFP Li et al., 2014 N/A
pCKII-GFP Li et al., 2016 N/A
BK channel without E29 This paper N/A
BK channel with E29 This paper N/A
splice reporter pFlare5 vector Stoilov et al., 2008 N/A
pGFP-C-shLenti bCaMKK shRNA constructs Origene Cat#TL711303
pGFP-C-shLenti Nova2 shRNA constructs Origene Cat#TL508674
pcDNA3.1/Puro-CAG-ASAP1 St-Pierre et al., 2014 Addgene Plasmid # 52519;
RRID:Addgene_52519
AAV-CaMKIIa-GCaMP6s-P2A-nls-tdTomato Gift from Jonathan Ting Addgene Plasmid #51086;
(unpublished) RRID:Addgene_51086
CaMKIV-GFP construct Gift from Haruhiko Bito N/A
CA-CaMKIV construct Gifts from Tian-Ming Gao N/A
CaMBP4 construct Gifts from Tian-Ming Gao N/A
Software and Algorithms
pClamp 9 Molecular Devices https://www.moleculardevices.com/
Prism GraphPad https://www.graphpad.com/
MATLAB Mathworks https://www.mathworks.com/
ImageJ Schneider et al., 2012 https://imagej.nih.gov/ij/

Cell 181, 1547–1565.e1–e8, June 25, 2020 e2

 ll
Article

LEAD CONTACT AND MATERIALS AVAILABILITY

Further information and requests for resources and reagents should be addressed to Lead Contact, Richard W. Tsien (richard.tsien@
nyulangone.org)
All unique reagents generated in this study are available from the Lead Contact with a completed Materials Transfer Agreement.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cell lines
HEK293 cell lines (human, female) were purchased from the American Type Culture Collection (ATCC). Cells were cultured in
standard Dulbeco’s modified Eagle’s medium (DMEM) supplemented with 10% FBS, 100 U/mL penicillin, 100 mg/mL streptomycin,
2 mM L-glutamine. All cells used were negative for mycoplasma. The cell lines were authenticated by the source repository.

Primary cell cultures

Cortical neurons were cultured from postnatal day 0 (P0) male and female Sprague-Dawley rat pups as previously described (Li et al.,
2016; Ma et al., 2014). Mixed cohorts of female and male pups were used for all experiments to minimize gender effects. The frontal
cortex was isolated and washed twice in ice-cold modified HBSS (4.2 mM NaHCO3 and 1 mM HEPES, pH 7.35, 300 mOsm)
containing 20% fetal bovine serum (FBS; Hyclone, Logan, UT). Samples were washed and digested for 30 min in a papain solution
(2.5 mL HBSS + 145 U papain + 40 mL DNase) at 37 C with gentle shaking every 5 min. Digestion was stopped by adding 5 mL of
modified HBSS containing 20% fetal bovine serum. After additional washing, the tissue was dissociated using Pasteur pipettes of
decreasing diameter. The cell suspension was pelleted twice, filtered with a 70 mm nylon strainer, and plated on 10 mm coverslips
coated with poly-D-lysine. The cultures were maintained in NbActiv4 (BrainBits, Springfield, IL), at 37 C in a 5% CO2 incubator. Half
of the media was changed at 7 days, and once per week thereafter.

Animals
C57BL/6 male and female mice (postnatal day 26), purchased from the Charles River Laboratory, were used for monocular depriva-
tion experiments. The Olig3Cre mouse line was a gift from Y. Nakagawa, University of Minnesota. R26floxstop-TeNT (tetoxf/f) mouse
line was a gift from M. Goulding at the Salk Institute for Biological Studies. These two mouse strains were maintained on a mixed
background (Swiss Webster and C57/ B16). All mice were housed with a 12 hour light-dark cycle. Mixed cohorts of female and
male mice were used for all experiments to minimize gender effects. Animal protocols were performed in accordance with NIH
guidelines and approved by the Institutional Animal Care and Use Committee at New York University and Sun Yat-sen University.

METHOD DETAILS

Constructs
Nova-2 mutants were produced by Phusion Site-Directed Mutagenesis Kit (Thermo Scientific). BK channel without E29 was
first cloned from pcDNA3-BK-GFP (Li et al., 2014) via PCR and inserted into pCKII-GFP construct with an aCaMKII promoter to
restrict expression to pyramidal cells (Li et al., 2016). E29 sequence was synthesized and inserted by Phusion Site-Directed Muta-
genesis Kit (Thermo Scientific). The splice reporter containing E29 and partial flanking intron sequences were cloned by PCR
from genomic DNA obtained from rat brain tissue and inserted into pFlare5 vector (Stoilov et al., 2008). Other constructs, see Key
Resources Table.

Transfection and treatment of cortical neurons

Neurons were transfected 7 to 9 days after plating using a high efficiency Ca2+- phosphate transfection method (Jiang and Chen,
2006). Experiments were performed 12-14 days after plating. For chronic spike blockade, neurons were treated with 1 mM TTX on
DIV12 for 48 hours to prevent action potential. Where indicated, additional antagonists (5 mM nimodipine, 4 mM KN93 or 3 mM
STO-609) were added with TTX. For chronic depolarization, neurons were treated with 20 mM K+ (20K), 40K or 60K solution (to depo-
larize cells; Na+ adjusted to maintain osmolarity) for 12 or 24 h. For action potential recording by patch clamp, coverslips with control
or TTX-treated neurons were removed from the culture medium, washed and recorded in TTX-free artificial cerebrospinal fluid
(ACSF). For imaging the action potential-independent Ca2+ transients and voltage changes in the dendrites, control or TTX-treated
neurons were transferred to a TTX-containing Tyrode’s solution consisting of (in mM): 150 NaCl, 4 KCl, 1 MgCl2, 2 CaCl2, 10 HEPES,
10 glucose, pH 7.4 with 1 mM TTX. Where indicated, antagonists were added by manual pipetting to achieve a final concentration
of 10 mM PhTx, 10 mM APV, 5 mM nimodipine or 1 mM thapsigargin in TTX-containing Tyrode’s solution.

Lentiviral transduction of cortical neurons

The production of lentivirus was performed as previously described (Ma et al., 2014), pGFP-C-shLenti constructs encoding
shRNAs against Nova-2 or bCaMKK (OriGene) were transfected into 293T cells along with the packaging plasmid psPAX2 and
the envelope plasmid pMD2.g. After 16 h, the medium was changed and the supernatant was collected 24 h later and cleared of

e3 Cell 181, 1547–1565.e1–e8, June 25, 2020

 ll
Article

cell debris by filtering through a 0.45 mm filter. The viral particles were concentrated by centrifuging the filtrate at 70,000 3 g for 2 hr at
4 C using a Beckman SW28 rotor. The viral pellet was then resuspended in sterile PBS, aliquoted, and stored at 80 C. Lentivirus
particles (0.5-1 mL of viral stock diluted in 20 mL of PBS per coverslip) were added to cortical cultures containing 500 mL of medium on
DIV7. The experiments with overexpressed proteins or shRNAs were performed on DIV14.

Immunocytochemistry and image acquisition and analysis

Cells were fixed in ice-cold 4% paraformaldehyde in phosphate buffer with 20 mM EGTA and 4% sucrose; permeabilized with 0.1%
Triton X-100; blocked with 10% normal goat serum (or donkey serum); and incubated overnight at 4 C in primary antibodies. For
surface staining of GluR1 or BK channel, coverslips were fixed and blocked in 7.5% normal donkey serum for 30 minutes in the
absence of Triton X-100. Surface staining with primary antibodies (Key Resources Table) was then performed for 1 hour at room tem-
perature. Cells were then permeabilized in 0.1% Triton X-100, and stained with anti-PSD-95 and anti-MAP2 (Figure S6) overnight at
4 C. The next day, cells were washed with PBS, incubated at RT for 60 min with Alexa secondary antibodies (1:1000, Molecular
Probes), washed again and mounted with ProLong Gold + DAPI (Invitrogen).
Fixed cells were imaged with a 40X or 60X oil objective on a Zeiss LSM 800 confocal microscope. Intensity quantification was per-
formed with custom scripts in MATLAB (Mathworks) or ImageJ (NIH). For all analyses, a region of interest lacking cells was selected in
each field of view as an ‘off-cell’ background, and the mean intensity was subtracted from all cellular regions of interest for each color
channel. The following types of analyses were performed from at least three independent cultures in each case:

(1) Analysis of Nova-2, pCaMKIV, bCaMKK and flag-tagged Nova-2 (e.g., Figures 3 and 4). Nuclear and cytosolic regions of in-
terest were manually drawn while viewing only DAPI or MAP2 channels, but blinded to the Nova-2 color channel. Background-
subtracted mean intensity was quantified and normalized to control conditions.
(2) Analysis of pCaMKII, pCaMKI and surface GluR1 intensity (e.g., Figure S6) was restricted to MAP2-positive regions of interest
containing a proximal dendrite (
60 mm in length,
15 mm in width) at least 30 mm from cell soma. For GluR1, PSD-95 puncta
were identified, and the average intensity of GluR1 signal intensity measured in the region of those puncta. For pCaMKII and
pCaMKI, the regions of interests were manually drawn while viewing MAP2 channel but blinded to the pCaMKII color channel.
The background was subtracted from pCaMKII intensity. Data represents mean ± SEM over 20 such dendrites, normalized to
control conditions.

Transfection and electrophysiology of HEK293 cells

HEK293 cells were transfected via a high efficiency Ca2+- phosphate transfection method (Jiang and Chen, 2006), with constructs
encoding CMV promoter-driven BK channel with (E29) or without (DE29) E29, or with splicing reporter constructs co-expressed with
or without WT or mutant Nova-2, or with CaMKIV-GFP and flag-Nova-2 constructs, or with CA-CaMKIV and flag-Nova-2 constructs.
Experiments were done 2-3 days later. Whole-cell recordings were obtained at room temperature (Axopatch 200B, Molecular De-
vices). Electrodes were pulled borosilicate glass capillaries (World Precision Instruments, MTW 150-F4), with 5–8 MU resistances,
before 80% series resistance compensation.

Electrophysiological recording of action potential half-width in cultured cortical neurons

DIV 10-14 cultured cortical neurons were used to measure homeostatic changes in action potential waveforms. Cultured cortical
neuron coverslips were placed into a submerged recording chamber and perfused with artificial cerebrospinal fluid (ACSF) at
4 mL/min held at room temperature (20-25 C) and bubbled with 95%/5% O2/CO2. All cells were measured in current clamp
mode using borosilicate patch electrodes with tip resistance of 2-4 MU. Recordings were not corrected for the liquid junction poten-
tial. Data were sampled at 10 kHz and low pass filtered at 2 kHz using a Bessel filter in pClamp 9 software. Recordings were not used
if the access resistance was above 25 MU or changed significantly throughout the recording (> 20%).
The ACSF contained (in mM): 3 KCl, 10 D-Glucose, 122 NaCl, 1.25 NaH2PO4, 1.3 MgCl2, 2 CaCl2, 26 NaHCO3. The internal pipette
solution contained (in mM): 130 K Gluconate, 1 MgCl2, 10 HEPES, 0.3 EGTA, 10 Tris-Phosphocreatine, 4 Mg-ATP, 0.3 Na-GTP.
To minimize changes in action potential duration due to depolarizing current steps, all action potentials were evoked as a rebound
spike (anodal break response) (Hodgkin and Huxley, 1952). Briefly, a 500 ms long pulse between 500 pA and 800 pA was injected
into the neuron before it was allowed to rebound to resting potential. Upon return to resting potential a single action potential was
usually evoked and sweeps where an action potential was successfully evoked were used for analysis. Using this method allows
an action potential to be evoked from a membrane potential negative enough to maximize removal of inactivation and minimize
the influence of variations in resting potential.
Action potential half width was measured using custom MATLAB scripts. Briefly, the action potential amplitude (baseline to
action potential peak value) was measured and the time between the half peak amplitude on the rising and falling phases was taken
to be the action potential half width.

Cell 181, 1547–1565.e1–e8, June 25, 2020 e4

 ll
Article

Recording Current-Voltage relationships in HEK cells transfected with BK channels

For recordings in HEK cells a symmetrical solution (containing equal ionic concentrations in the internal pipette solution and external
solution) was used containing (in mM): 10 K Gluconate, 2 KCl, 2 MgCl2, 10 HEPES, 9.93 CaCl2, 10 EGTA, 10 Glucose. A current-
voltage (I-V) relationship was measured in voltage clamp mode with 250 ms long voltage steps between 100 mV and 100 mV
in 10 mV increments. Custom MATLAB scripts were used offline to take current measurements when the membrane current
stabilized after voltage steps.

Ca2+ imaging on dendrite and soma of cultured neurons

Neurons were transfected with GCaMP6s on DIV 7-9 and Ca2+ transients were imaged on DIV13 to 15. Images were collected with
a 40X oil objective on a Zeiss 710 confocal microscope, 300 ms per frame. Intensity quantification was performed with custom
scripts in MATLAB (Mathworks).
For imaging of Ca2+ transients in soma, the 488 nm fluorescence changes were quantified using regions of interest in cell soma
(Figure 1B). Ca2+ transients were imaged in TTX-free Tyrode’s solution, consisting of (in mM): 150 NaCl, 4 KCl, 1 MgCl2, 2 CaCl2,
10 HEPES, 10 glucose, pH 7.4. The amplitude of single action-induced Ca2+ transient was defined by the minimal amplitude of single
Ca2+ transient that could propagate to the entire cell (soma and dendritic tree). The Ca2+ transients that arose locally in the soma
but did not propagate to the dendrite was regarded as local under-threshold depolarization-induced Ca2+ transients and was elim-
inated from the quantification. For imaging the action potential-independent Ca2+ transients in the soma, control or TTX-treated
neurons were imaged in a TTX-containing Tyrode’s solution.
For Ca2+ transients in dendrite, a
200 mm long dendrite located at least 30 mm from cell soma was chosen. Ca2+ transients
were imaged in TTX-containing Tyrode’s solution. Where indicated, antagonists were added by manual pipetting to achieve a
final concentration of 10 mM PhTx, 10 mM APV, 5 mM nimodipine or 1 mM thapsigargin in TTX-containing Tyrode’s solution.
In dendritic trees, sites of spontaneous Ca2+ influx were identified as regions of interest (ROIs) which showed sparse peaks of
GCaMP6s fluorescence through time, many of which spatially overlapped with spine-like protrusions in baseline GCaMP6s images
highlighting dendritic trees structure (Figures 6D and 6F). ROI shape and number were obtained by automatically computing a series
of operations. First, temporal stacks of fluorescence images were de-noised and the baseline GCaMP6s fluorescence was
subtracted pixel-wise. Noise elimination was performed using a custom method that kept intact the non-random correlation between
neighboring pixels showing coordinated activity in order to preserve localized signals from Ca2+ influx. An overly large set of putative
ROIs were then identified by using the separable non-negative matrix factorization (sNMF) algorithm. sNMF works by identifying
strong but temporally sparse signals and preventing the temporal redundancy between extracted signals - a good model for
spontaneous Ca2+ signals in dendrites. The final ROI set was obtained by unbiasedly selecting only those with cumulated temporal
!
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P 2
fluorescence FðtÞ greater than the average + 3 3 the standard deviation of cumulated fluorescence values drawn outside
t

candidate ROIs.

Voltage imaging of cultured neurons

Cultured pyramidal neurons were transfected via a high efficiency Ca2+- phosphate method 7-9 days after plating (Jiang and Chen,
2006), with 1 mg pcDNA3.1/Puro-CAG-ASAP1 plasmid (St-Pierre et al., 2014). Experiments were done 5-7 days later. In voltage im-
aging data in Figures 6A–6C, the spontaneous AP in soma and bAP in dendritic spines were monitored by simultaneously measuring
the DF/F of voltage sensor ASAP1 in soma and dendritic spines. We reasoned that 1) only action potential, instead of subthreshold
synaptic current, could backpropagate from soma to dendritic trees and elicit simultaneous voltage changes in soma, different den-
dritic branches and dendritic spines; 2) due to the high on and off kinetics of ASAP1 (
2ms), high speed recording (2.87 ms / frame in
our experiments) of the DF/F changes could reliably detect single spontaneous action potential; DF/F changes induced by single ac-
tion potential were comparable in both amplitude and duration; DF/F changes induced by burst of action potentials were much larger
in amplitude and duration and were discarded in our analysis. Accordingly, we named the simultaneous DF/F changes (voltage
changes) in soma and spines as ‘‘AP in soma’’ and ‘‘bAP in spine,’’ respectively, and named the DF/F changes (voltage changes)
that took place only in spines as ‘‘Vspine.’’ When measuring Vspine, TTX was acutely applied to eliminate action potentials. A narrow
region of interest containing the plasma membrane of dendritic spines (30-80 mm from soma) and cell soma were chosen, and time
lapse imaging with a speed of 2.87 ms per frame was performed on a Zeiss LSM 710 confocal microscope, utilizing GFP filter sets.
Background-subtracted fluorescence intensity were analyzed in ImageJ (NIH). ASAP1 waveforms were plotted as -DF/F (to account
for the negative change in ASAP1 fluorescence upon depolarization) and corrected for bleaching. The noise was filtered with
LOWESS smoothing. Identical cell-selection, bleaching correction and noise-filtering parameters were used for all datasets in
Figure 6. Data is plotted as mean ± SEM of 14-15 cells for each condition.
To calculate the standard fluorescence-voltage (DF/F-DV) curve for ASAP1, neurons were perfused high K+ Tyrode’s solutions with
Na+ adjusted to maintain osmolarity. The change of fluorescence intensity (-DF/F) was plot against voltage according to 4 mM K+
(4K), 60 mV; 20K, 37 mV; 40K, 19 mV; 60K, 9 mV; 90K, 0 mV (Wheeler et al., 2012). The membrane potentials in different
conditions were then calculated according to the peak values of -DF/F.

e5 Cell 181, 1547–1565.e1–e8, June 25, 2020

 ll
Article

Immunoprecipitation
Immunoprecipitation was performed with Dynabeads Protein G Immunoprecipitation Kit (Life Technologies) following the kit protocol
as described (Li et al., 2016). Briefly, for binding the antibody with the beads, 5 mg of antibody or control IgG was crosslinked and
incubated at room temperature with 100 mL Dynabeads. The supernatant was removed with the help of a magnet to retain the
bead-antibody complex, which was then washed with Ab Binding & Washing Buffer. For lysate preparation, cultured cortical neurons
or HEK293 were washed with PBS and lysed on ice with IP Lysis Buffer (Pierce) containing protease and phosphatase inhibitors. The
lysate was centrifuged at 16,000 g for 10 min at 4 C to pellet the cell debris. The resulting supernatant was incubated with bead-anti-
body complex with rotation for 30 min at room temperature. The bead-antibody-antigen complex was then washed 3 times using
washing buffer. The protein complex was eluted with Elution buffer and the eluent was mixed and heated with 2 3 SDS sample buffer
containing b-mercaptoethanol at 90 C for 10 min.

Protein sample preparation and western blot

Nuclear and cytosolic protein fraction was performed as described (Li et al., 2014) . Briefly, nuclear and cytosolic protein was ex-
tracted using the Pierce Nuclear Protein Extraction Kit (Thermo), following the manual. Neurons were collected by centrifugation
at 500 g for 3 min at 4 C and were resuspended in 200 mL precooled CER I by 15 s of intense vortexing. The solution was then incu-
bated on ice for 10 min. CER II (11 ml) was precooled and added to the solution. After 5 s of intense vortexing, the solution was incu-
bated on ice for 1 min. After another 5 s of intense vortexing, the solution was then centrifuged at 16,000 g for 5 min at 4 C and the
supernatant collected as cytosolic protein. NER (100 ml) was added to the precipitate and the sample vibrated intensively for 15 s. The
solution was then incubated on ice for 10 min. The vibration and incubation were repeated four times, and then the precipitate was
centrifuged at 16,000 g for 10 min at 4 C. The supernatant represented purified nuclear protein.
Plasma membrane protein extraction was performed with Pierce Cell Surface Protein Isolation Kit as described (Li et al., 2014). For
surface biotinylation, neurons were chilled on ice, washed twice with ice-cold PBS, and then incubated with PBS dissolved sulfo-
NHS-SS-biotin for 30 min at 4 C. Unreacted biotin was quenched by washing cells with Quenching Solution. Cultures were harvested
in Lysis Buffer. Homogenates were centrifuged at 10,000 g for 2 min at 4 C. The resulting supernatant was rotated 1 hr at room tem-
perature with NeutrAvidin Agarose. The beads were washed with Wash Buffer and analyzed by immunoblotting with each antibody.
Whole cell lysate was performed as previously described (Li et al., 2016). Cells were lysed with CelLytic M Cell Lysis Reagent
(Sigma). Protein concentration was measured by the BCA Protein Assay Kit (Thermo). The lysates were combined with 2 3 SDS
loading buffer, then boiled for 10 min.
Protein samples were loaded onto 4%–20% gradient precast or 10% SDS-PAGE gel (Bio-Rad). Proteins were transferred to a
PVDF membrane (Millipore) for 1 h at 350 mA, and membranes were incubated in Odyssey Blocking Buffer (Li-Cor) for 1 h at
room temperature. The membrane was incubated overnight at 4 C with primary antibodies. The blots were washed three times in
PBS containing 0.1% Tween-20 for 15 min and then incubated with IRDye-conjugated secondary antibody for 1 h in PBS, 0.1%
Tween-20 at room temperature. Immunoreactivity was detected by an Odyssey imaging system (Li-Cor).

Oligo-RNA pull-down
RNA pull-down was performed with Magnetic RNA-Protein Pull-Down Kit (Thermo) following the manual. Briefly, synthesized 50 -bio-
tinylated 20 -OMe-RNA oligonucleotides (Eurofins) (Key Resources Table) were bound to streptavidin magnetic beads (Thermo)
in RNA Capture Buffer at room temperature with agitation. Cleared brain lysates from 1-month old mouse cortex (Figure 2B) or
transfected HEK cells (Figure S5E) prepared in IP Lysis Buffer (Pierce) containing protease and phosphatase inhibitors mixed with
Protein-RNA Binding Buffer, and then incubated with the packed beads at 4 C for 1 hr. Beads were washed three times with
wash buffer and the precipitate was subjected to immunoblot analysis. Signal intensities were quantified by an Odyssey imaging
system (Li-Cor).

RNA immunoprecipitation
RNA immunoprecipitation was performed with Magna RIP RNA-Binding Protein Immunoprecipitation Kit (Sigma) following
the manual. Briefly, 5 mg of anti-Nova-2 antibody or control IgG were incubated with magnetic beads in RIP wash buffer at room
temperature for 30 mins. The antibody prebound beads were then washed 3 times with RIP wash buffer at room temperature. For
lysate preparation, mouse cortical lysates were prepared with complete RIP Lysis Buffer. The lysates were mixed with prebound
beads in RIP Immunoprecipitation Buffer containing RIP wash buffer, EDTA and RNase inhibitor, overnight with rotation at 4 C.
The magnetic beads antibody-RNA binding protein complex was washed with wash buffer and then re-suspended in proteinase
K buffer at 55 C for 30 minutes. Phenol:chloroform:isoamyl alcohol and chloroform were used to separate the phase and the
RNA was precipitated with Salt Solution I, Salt Solution II, Precipitate Enhancer and absolute ethanol at 80 C overnight.
The RNA precipitation was centrifugated and washed with 80% ethanol solution. The pellets were dried and resuspended in
RNase-free water and used for further RT-PCR.

RT-PCR and Realtime qPCR

Total RNA was isolated from cortical neurons or HKE 293 cells with RNeasy Mini kit (QIAGEN) according to manufacturer’s
instructions. Extracted RNA was reverse transcribed into first strand cDNA using oligo dT and Superscript III (Invitrogen). PCR

Cell 181, 1547–1565.e1–e8, June 25, 2020 e6

 ll
Article

was performed with AmpliTaq Gold 360 Master Mix with DY-682 labeled primers (for X6). The PCR products were either resolved
on agarose gel (X1-X5) or denatured and resolved on 6% polyacrylamide/8 M urea denaturing gels (X6). The band density was
analyzed by an Odyssey imaging system (Li-Cor). qPCR was performed with SYBR-green PCR master mix (Fermentas) with primers
(Table S1) using the DNA engine Opticon 2 (Bio-Rad).

Thalamic input elimination and immunohistochemistry

The Olig3Cre mouse line was a gift from Y. Nakagawa, University of Minnesota. R26floxstop-TeNT (tetoxf/f) was a gift from M. Goulding at
the Salk Institute for Biological Studies. Neurotransmission is selectively blocked from the thalamic cells as expression of Cre re-
moves a stop codon at the R26 locus (removing the flox sites) permitting the expression of the tetanus toxin light chain subunit
(TeNT) only within thalamic Olig3+ neurons. Olig3Cre mice were bred with animals that were homozygous for tetoxf/f. Pups from
the subsequent litters were genotyped for Cre and presence of the TeNT allele in the R26 locus. Olig3Cre-only animals were used
as controls and Olig3Cre plus tetoxf/+ heterozygous animals as mutants. All mouse strains were maintained on a mixed background
(Swiss Webster and C57/ B16).
Mice were perfused inter cardiac with 4% PFA after being anesthetized either on ice or using Sleepaway IP administration. Brains
were post-fixed for 30 minutes and cryopreserved in 30% sucrose following the perfusion and brain harvest. 16mm coronal sections
were obtained using Cryostat (Leica Biosystems) and collected on super-frost coated slides, then allowed to dry overnight and stored
at 20 C until use. For immunofluorescence, cryosections were thawed and allowed to dry for 5-10 min and rinsed twice in 1x PBS.
They were incubated at room temperature in a blocking solution of PBST (PBS-0.1%Tx-100) and 10% normal donkey serum (NDS)
for 60min, followed by incubation with primary antibodies in PBS-T and 1% NDS at 4 C overnight. Samples were then washed 3 times
with PBS-T and incubated with fluorescence conjugated secondary Alexa antibodies (Life Technologies) in PBS-T with 1% NDS at
room temperature for 60-90min. Slides were then incubated for 30 s with DAPI, washed 3 times with PBS-T and once with PBS.
Slides were mounted with Fluoromount G (Southern Biotech) and imaged on a Zeiss LSM 510 Laser scanning microscope.

Monocular deprivation
The monocular deprivation was performed following previous study (Ma et al., 2013). Briefly, the mice (postnatal day 26) were anes-
thetized with the mixture of 100 mg/kg ketamine hydrochloride and 10 mg/kg xylazine hydrochloride intraperitoneally. Under micro-
scope and illuminator, the eyelashes and the edge of eyelids were cut using the spring scissors. The eyelids of the right eye were
sutured with three mattress sutures. A thin layer of xylocaine 2% Jelly and bacitracin zinc ointment was applied to the sutured eyelids.
After 5 days monocular deprivation, mice brains were sliced for immunostaining or RNA isolation. Nova-2 localization or E29 splicing
in the monocular region of primary cortex V1, both contralateral and ipsilateral to the visual deprived eye, was analyzed.

Protein sample preparation and mass spectrometry analysis

Protein sample preparation and mass spectrometry analysis were perform as described (Ma et al., 2014). Briefly, flag-tagged Nova-2
were co-transfected with CA-CaMKIV or GFP into HEK293 cells. 48 h after transfection, the cells were washed with ice-cold PBS and
lysed with IP Lysis buffer (Thermo Fisher) containing a protease inhibitor cocktail tablet (Roche, Indianapolis, IN). The lysates were
mixed with CaM kinase reaction buffer (Enzo Life Sciences) and incubated with anti-flag antibody (Cell Signaling Technology) and
Protein A/G Sepharose beads. The immunoprecipitated complex was separated by SDS-PAGE. Corresponding areas showing
Nova-2 bands were cut out from the Coomassie blue-stained SDS-PAGE gels. Gel pieces underwent in-gel digestion and the phos-
phopeptides were enriched by titanium dioxide (GL Sciences Inc., Japan) based on a previously published protocol (Rappsilber et al.,
2007). Resulting peptides were subjected to Nano LC-MS/MS analysis using a Thermo Scientific EASY-nLC 1000 coupled to a Q
Exactive mass spectrometer (Thermo Fisher Scientific). A self-packed 75mm 3 25 cm reversed phase column (Reprosil C18,
3mm, Dr. Maisch GmbH, Germany) was used for peptide separation. Peptides were eluted by a gradient of 3%–30% acetonitrile
in 0.1% formic acid over 60 min at a flow rate of 250 nL/min. The Q Exactive was operated in the data-dependent mode with survey
scans acquired at a resolution of 50,000 at m/z 400. Up to the top 10 most abundant precursors from the survey scan were selected
with an isolation window of 1.6 Thomsons and fragmented by higher energy collisional dissociation with normalized collision energies
of 27. The maximum ion injection times for the survey scan and the MS/MS scans were 20 and 60 ms, respectively, and the ion target
value for both scan modes were set to 1,000,000. Each sample was processed and analyzed in triplicate.

Phosphopeptide identification and quantitation

Phosphopeptide identification and quantitation was performed as described (Ma et al., 2014). Briefly, the raw files were processed
using the MaxQuant computational proteomics platform version 1.2.7.0 for peptide identification and quantitation. The fragmentation
spectra were searched against the Uniprot protein database allowing up to two missed tryptic cleavages. Carbamidomethylation of
cysteine was set as a fixed modification, and phosphorylation of serine/threonine/tyrosine, oxidation of methionine and protein N-ter-
minal acetylation were used as variable modifications for database searching. The precursor and fragment mass tolerances were
set to 7 and 20 ppm, respectively. Quantitation of the phosphopeptides was performed using the ion intensity values calculated
by MaxQuant for the quintuply charged peptides. Normalization of phosphopeptide intensities to global peptide intensities had no
effect on calculated ratios.

e7 Cell 181, 1547–1565.e1–e8, June 25, 2020

 ll
Article

QUANTIFICATION AND STATISTICAL ANALYSIS

All statistical analyses were performed using Prism (GraphPad Software). Means of two groups were compared using Student’s t
test. One-way ANOVA was used to compare group means between more than two groups, followed by either Dunnett’s multiple
comparisons test when all other groups were compared to the control group, or Tukey’s multiple comparisons test when
means of each pair of groups were compared. All comparisons are two-sided. Statistical significance was determined as follows:
*p < 0.05, **p < 0.01. Data are shown as mean ± SEM. The statistical details of all experiments, including the number of samples
and p values, can be found in figures and figure legends.

DATA AND CODE AVAILABILITY

All data supporting the findings of this study and custom MATLAB code for calcium imaging are available upon request from the
Lead Contact.

Cell 181, 1547–1565.e1–e8, June 25, 2020 e8

 ll
Article

Supplemental Figures

Figure S1. APD Increased after Chronic TTX Treatment, Related to Figure 1
(A) Representative waveforms of action potential recorded after TTX treatment (24 or 48 h). (B) Quantification of half width of action potentials in the groups in (A)
(n = 10. Data for control and 48 h groups is from Figure 1A).
 ll
Article

Figure S2. Effects of Inactivity and E29 Splicing on Trafficking and Expression of BK Channels, Related to Figure 1
(A) Expression level of BK channel mRNA was assayed by RT-PCR (n = 4). (B) BK channel expression in the plasma membrane was assayed by western blotting
after biotinylation-mediated membrane protein fractionation. (C) Immunofluorescence of surface BK channel. Surface BK channel was probed with antibodies
targeting extracellular N terminus. (D) Quantification of intensity of surface BK channel (n = 10). (E) Overexpression of GFP-tagged BK channel isoforms in HEK293
cells. (F) Quantification of surface fluorescence intensity of GFP-tagged BK channel isoforms. (G) Expression level of GFP-tagged BK channel isoforms was
assayed by western blotting after biotinylation-mediated membrane protein fractionation. Na/K ATPase, Lamin B and GAPDH are markers for plasma membrane,
nucleus and cytosol.
 ll
Article

Figure S3. Effects of Chronic Inactivity and Depolarization on AS of the BK Channel, Related to Figure 1
(A) Impact of chronic inactivity on alternative splicing of X1 to X5. Cultured cortical neurons were sham- or TTX-treated for 48 h. Separation of RT-PCR products
obtained with specific primers flanking each alternatively spliced site (X1 to X5). Closed arrows indicate the longer RT-PCR products obtained if the alternative
exon was included, open arrows indicate the shorter products if that exon was excluded. (B) Quantification of inclusion or exclusion of X1 to X6 by quantitative RT-
PCR. (C) High K+ solution maintained long-term depolarization in resting membrane potential. (D) Impact of sustain depolarization on E29 inclusion. Cultured
cortical neurons were sham- or high K+-treated (20mM, 40mM and 60mM, respectively) for 12 or 24 h. E29 inclusion was examined by RT-PCR.
 ll
Article

Figure S4. Nova-2 Is Responsible for E29 AS, Related to Figure 2

(A) Validation of the knockdown of Nova-2 expression by shRNAs. Cortical neurons were transfected with shRNAs either targeting the untranslated region (UTR)
of the Nova-2 gene or targeting the Nova-2 coding sequence (CDS). Both shRNAs efficiently knocked down endogenous Nova-2 expression. (B) UTR-lacking
flag-Nova-2 plasmids (Nova-2 (R)) were co-expressed with shRNAs targeting CDS (targeting three different regions) or shRNAs targeting UTR in HEK293 cells.
The shRNAs targeting CDS efficiently knocked down flag-Nova-2 expression, whereas shRNAs targeting UTR did not. (C) E29 inclusion could only be induced by
wild-type (WT) Nova-2, but not DNLS-Nova-2. WT-Nova-2-induced E29 inclusion was prevented by the shRNAs targeting the Nova-2 coding sequence (CDS),
but not by shRNAs targeting the untranslated region (UTR) of the Nova-2 gene.
 ll
Article

Figure S5. Identification of CaMKIV-Mediated Nova-2 Phosphorylation Sites by Mass Spectrometry, Related to Figure 5
(A) Schematic diagram of Nova-2 protein as in Figure 7. (B to D) Mass spec results of phosphorylation level of S25, T27 and S194 sites. (E) WT and S194E Nova-2
are no different in RNA binding. Flag-tagged WT Nova-2 or S194E Nova-2 was overexpressed in HEK293 cells. RNA pull-down assay was performed from the cell
lysates by probe A1 (Figure 2). The RNA binding ability of WT and S194E Nova-2 was assayed by western blotting with anti-flag antibody.
 ll
Article

Figure S6. Ca2+ Transients, Surface GluA1 Expression, and Activation of CaMKs in Con or TTX-Treated Neurons, Related to Figure 6
(A) Ca2+ transients were imaged from the soma of control neurons, with no TTX present. (B) Ca2+ imaging, from the soma of control neurons (upper) or TTX-treated
(48 h) neurons, was performed in the acute presence of TTX. (C) Ca2+ transients induced by action potential-independent synaptic transmission (red, imaged with
TTX present) or by back-propagating action potentials (bAP) (blue, imaged without TTX). (D and E) Puncta area (indicating active spines, D) and the number of
Ca2+ transients (per minute, E) in each puncta (active spine) in control and TTX-treated (48 h) neurons (n = 14 to 15). (F) The influence of chronic inactivity on
surface GluA1 expression. Surface-labeling of GluA1 in the dendrites from control and TTX-treated (48 h) neurons. PSD95 as synaptic marker, MAP2 as dendrite
marker. Scale bar 10 mm. (G) Quantification of surface GluA1 immunofluorescent intensity in the dendrites from control and TTX-treated (48 h) neurons. (H) The
activation of CaMKII in control and TTX-treated neurons. Immunofluorescence of phosphorylated CaMKII on Thr286 site, indicating the level of autophos-
phorylated CaMKII. (I) Quantification of dendritic pCaMKII intensity in control and TTX groups. (J) The activation of CaMKI in control and TTX-treated neurons.
Immunofluorescence of phosphorylated CaMKI on Thr177/178 site, indicating the activation of CaMKI by CaMKK. (K) Quantification of dendritic pCaMKII in-
tensity in control and TTX groups. MAP2 as dendrite marker. Scale bar 10 mm.
 ll
Article

Figure S7. Regulation of pCaMKIV Activation, Nova-2 Localization, E29 Splicing, and CaMK Localization by Chronic Inactivity and CaV1-
CaMK Blockers or Inhibitors, Related to Figure 7
(A) The activation of nuclear CaMKIV in different groups as indicated. Immunofluorescence of phosphorylated CaMKIV on Thr196 site indicating the level of
CaMKIV activation by CaMKK. (B) Normalized nuclear/cytosolic ratio of pCaMKIV immunofluorescence intensity in each group as indicated. (C) The localization
of Nova-2 in different groups as indicated. (D) Normalized nuclear/cytosolic ratio of Nova-2 immunofluorescence intensity in each group as indicated. (E) 12 h and
24 h high potassium media treatment (20mM, 40mM and 60mM, respectively) led to Nova-2 translocation from nucleus to cytosol. (F) Long-term depolarization
reduced the inclusion of E29, which could be blocked by CaV1 blocker nimodipine and CaM kinases blocker KN-93. (G and H) Fold changes in abundance of
CaM, aCaMKI, bCaMKI, gCaMKI, dCaMKI and gCaMKII in the nucleus, cytosol and nucleus/cytosol ratio. (G) Representative micrographs of different proteins in
control or TTX-treated neurons. (H) Ratios of levels after TTX treatment (48 h) relative to levels in control neurons. (I) Validation of the knockdown of bCaMKK
expression by shRNAs. flag-tagged bCaMKK was co-expressed with shRNAs targeting different locations of rat bCaMKK CDS. The efficiency of shRNAs was
assayed by western blotting.
 Article

An Autonomous Oscillation Times and Executes

Centriole Biogenesis
Graphical Abstract Authors
Mustafa G. Aydogan,
Thomas L. Steinacker,
Mohammad Mofatteh, ..., Alain Goriely,
Michael A. Boemo, Jordan W. Raff

Correspondence
mustafa.aydogan@path.ox.ac.uk
(M.G.A.),
mb915@cam.ac.uk (M.A.B.),
jordan.raff@path.ox.ac.uk (J.W.R.)

In Brief
Feedback-driven oscillations in centriolar
Plk4 kinase levels—normally entrained by
the cell-cycle oscillator but capable of
running autonomously—trigger and time
centriole biogenesis to ensure that
daughter centrioles grow at the right time
and to the right size.

Highlights
d Centriolar Plk4 levels oscillate and act as a switch for
centriole biogenesis

d Oscillations may be generated via an Asl/Plk4 delayed

negative feedback loop

d Plk4 oscillations are entrained and phase-locked by the Cdk/

Cyclin oscillator (CCO)

d Plk4 oscillations can drive centriole biogenesis even when

the CCO is perturbed

Aydogan et al., 2020, Cell 181, 1566–1581

June 25, 2020 ª 2020 The Author(s). Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.018 ll
 ll

Article
An Autonomous Oscillation
Times and Executes Centriole Biogenesis
Mustafa G. Aydogan,1,5,* Thomas L. Steinacker,1,5 Mohammad Mofatteh,1 Zachary M. Wilmott,1,2 Felix Y. Zhou,3
Lisa Gartenmann,1 Alan Wainman,1 Saroj Saurya,1 Zsofia A. Novak,1 Siu-Shing Wong,1 Alain Goriely,2
Michael A. Boemo,1,4,* and Jordan W. Raff1,6,*
1Sir William Dunn School of Pathology, University of Oxford, Oxford OX1 3RE, UK
2Mathematical Institute, University of Oxford, Oxford OX2 6GG, UK
3Ludwig Institute for Cancer Research, University of Oxford, Oxford OX3 7DQ, UK
4Present address: Department of Pathology, University of Cambridge, Cambridge CB2 1QP, UK
5These authors contributed equally
6Lead Contact

*Correspondence: mustafa.aydogan@path.ox.ac.uk (M.G.A.), mb915@cam.ac.uk (M.A.B.), jordan.raff@path.ox.ac.uk (J.W.R.)

https://doi.org/10.1016/j.cell.2020.05.018

SUMMARY

The accurate timing and execution of organelle biogenesis is crucial for cell physiology. Centriole biogenesis is
regulated by Polo-like kinase 4 (Plk4) and initiates in S-phase when a daughter centriole grows from the side of a
pre-existing mother. Here, we show that a Plk4 oscillation at the base of the growing centriole initiates and times
centriole biogenesis to ensure that centrioles grow at the right time and to the right size. The Plk4 oscillation is
normally entrained to the cell-cycle oscillator but can run autonomously of it—potentially explaining why cen-
trioles can duplicate independently of cell-cycle progression. Mathematical modeling indicates that the Plk4
oscillation can be generated by a time-delayed negative feedback loop in which Plk4 inactivates the interaction
with its centriolar receptor through multiple rounds of phosphorylation. We hypothesize that similar organelle-
specific oscillations could regulate the timing and execution of organelle biogenesis more generally.

INTRODUCTION we recently examined living syncytial Drosophila embryos where

Albert Claude’s landmark paper (Claude, 1943) challenged the
idea that cells are a mere bag of enzymes whose contents
grow freely in the cytoplasm with no active regulation. We now
appreciate the diverse and compact nature of the many organ-
elles in the cytoplasm (Marsh et al., 2001), yet the physical mech-
anisms that regulate the number and size of these organelles
remain largely unknown (Marshall, 2016). For most organelles
in the cell, however, this question has been difficult to address,
as the variation in their numbers and 3D-shape has made it chal-
lenging to monitor their growth—or to even determine which
parameter (e.g., their surface area, volume, or perhaps the
amount of a limiting component) best defines their size.
Centrioles are highly structured organelles that form centro-
somes and cilia (Bettencourt-Dias et al., 2011; Nigg and Holland,
2018; Nigg and Raff, 2009). Their linear structure and tightly
controlled pattern of duplication makes them an attractive model
with which to study organelle biogenesis (Goehring and Hyman,
2012; Marshall, 2016). Most cells are born with a single pair of
centrioles that duplicate precisely once during S-phase, when
a daughter centriole grows out orthogonally from the base of
each mother until it reaches the same size as its mother (Banterle
and Gönczy, 2017; Fırat-Karalar and Stearns, 2014; Nigg and
Holland, 2018). To monitor the dynamics of centriole growth,
we could follow the assembly of hundreds of centrioles as they
duplicate in near-synchrony in a common cytoplasm (Aydogan
et al., 2018). These studies revealed that centriole growth in
these embryos is homeostatic: when centrioles grow slowly,
they grow for a longer period; when centrioles grow quickly,
they grow for a shorter period. As a result, centrioles grow to a
consistent size.
Polo-like kinase 4 (Plk4) is the master regulator of centriole
biogenesis and it is initially recruited to a ring around the mother
centriole, but this ring resolves into a single focus on the side of
the mother, defining the site of daughter centriole assembly (Ban-
terle and Gönczy, 2017; Fırat-Karalar and Stearns, 2014; Leda
et al., 2018; Nigg and Holland, 2018; Takao et al., 2019). Unex-
pectedly, we found that Plk4 not only determines the position of
this site, but also helps to establish the inverse relationship be-
tween the rate and period of daughter centriole growth (Aydogan
et al., 2018). Plk4 presumably influences the rate of centriole
growth, at least in part, by phosphorylating Ana2/STIL to promote
its interaction with Sas-6 and, consequently, the assembly of the
central cartwheel (Dzhindzhev et al., 2014; Kratz et al., 2015; Ohta
et al., 2014), the 9-fold symmetric structure that forms the back-
bone of the growing daughter centriole (Kitagawa et al., 2011;
van Breugel et al., 2011, 2014). It is less clear, however, how
Plk4 might influence the period of centriole growth.

1566 Cell 181, 1566–1581, June 25, 2020 ª 2020 The Author(s). Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
 ll
Article

A B

E F

Figure 1. Plk4 Levels Oscillate at the Centriole in a Process Entrained by the CCO
(A) Top panel: micrograph shows an image from a time-lapse movie of an embryo expressing Plk4-NG. Middle panels: micrographs illustrate the centriolar Plk4-
NG oscillation during nuclear cycle 12—obtained by superimposing all the Plk4-NG foci (n = 60) at each time point (see STAR Methods). Bottom panel:
quantification of centriolar Plk4-NG levels during nuclear cycles 11–13 in a single embryo (red arrows highlight equivalent time points in the middle panels).
(B) Graphs show the mathematical regression of centriolar Plk4-NG dynamics during S-phase of cycles 11–13 (regression mean ± SEM). R2 values indicate
goodness-of-fit. N R 15 embryos; n = 24, 37, and 53 centrioles (mean) per embryo over cycles 11–13, respectively.
(C) The bar charts quantify the oscillation parameters—derived from the data shown in (B). Data are presented as mean ± SD. Statistical significance was as-
sessed using an ordinary one-way ANOVA test (for Gaussian-distributed data) or a Kruskal-Wallis test (***p < 0.001; ****p < 0.0001; ns, not significant).
(D) Micrographs show, and pie charts quantify, the distribution of Plk4-NG at centrioles assessed by 3D-SIM at the indicated phases of the nuclear cycle (see
STAR Methods). N = 6 embryos per cell-cycle stage; n = 20 centrioles per embryo; all images were scored blindly by 3 assessors and the mean score is shown
(scale bar, 0.5 mm).
(legend continued on next page)

Cell 181, 1566–1581, June 25, 2020 1567

 ll
Recent studies have shown that Plk4 localizes to centrioles in
a cyclical manner in both fly embryos (Aydogan et al., 2018) and
human cultured cells (Takao et al., 2019), but the functional sig-
nificance of this localization pattern is unclear. Here, we show
that a Plk4 oscillation at the base of the growing centriole initiates
and times centriole biogenesis in fly embryos.
RESULTS AND DISCUSSION
Plk4 Levels Oscillate at the Base of Growing Daughter
Centrioles
To investigate the cyclical recruitment of Plk4 to the centrioles,
we generated flies transgenically expressing Plk4-mNeonGreen
(Plk4-NG) under the control of its own promoter in a Plk4 mutant
background. We monitored centriolar Plk4-NG levels in living
Drosophila syncytial embryos, where the duration of S-phase
gradually elongates over nuclear cycles 11–13 (Figures 1A, S1,
and S2A; Video S1). Centriolar Plk4-NG levels oscillated during
each cycle: levels started to rise in M-phase, peaked in early-
mid S-phase, and were minimal by the next M-phase (Figures
1A and S2A). We fit the S-phase oscillations in individual em-
bryos (Figures S1C and S1D) to derive an average S-phase oscil-
lation for each cycle (Figure 1B).
Not surprisingly, the Plk4 oscillations appeared to be entrained
by the core Cdk/Cyclin oscillator as their period increased as nu-
clear cycles slowed during cycles 11–13 (Figure 1C). Moreover,
genetically altering the duration of the nuclear cycles elicited cor-
responding alterations in the Plk4 oscillation period (Figures 1E
and 1F). Interestingly, however, the Plk4 oscillation exhibited
adaptive behavior: as the period (T) of the oscillation tended to
increase at successive cycles, its amplitude (A) tended to
decrease, so that the total amount of Plk4 recruited to centri-
oles—i.e., the area under the S-phase oscillation curve (area un-
der the curve [U])—remained relatively constant (Figure 1C).
Plk4 is initially recruited to a ring around the mother centriole
that resolves into a single hub that defines the site of daughter
centriole assembly (Banterle and Gönczy, 2017; Fırat-Karalar
and Stearns, 2014; Nigg and Holland, 2018). To examine how
this localization related to the Plk4 oscillations, we used 3D-
structured illumination super-resolution microscopy (3D-SIM)
to assess the centriolar localization of Plk4 during the nuclear cy-
cles in living embryos. Plk4-NG was only very briefly detectable
in a ring during late-mitosis; at all other stages it appeared largely
as a single hub (Figure 1D). Thus, the recruitment and loss of Plk4
from the centriole wall is not responsible for the S-phase oscilla-
tion we observe in these embryos; instead, centriolar Plk4-NG
levels oscillate at the base of the growing daughter centriole.
(E) Graph shows the mean regression of Plk4-NG oscillations in nuclear cycle 12 of WT embryos (green), or in embryos where the genetic dose of either cyclin B
(CycB1/2; blue) or grapes (Drosophila Chk1) (grp1/2; red) has been halved to slow or speed-up the nuclear cycles, respectively. Dashed lines mark the center (peak)
of the Plk4-NG oscillations (denoted with C), and dotted lines indicate the time of NEB (denoted with N) for each genotype. N R 14 embryos for each condition; n =
55, 43, and 44 centrioles (mean) per embryo in WT, CycB1/2, and grp1/2 embryos, respectively. To clearly illustrate the phase shift in the oscillations, the highest
mean fluorescence signal for each group was normalized to 1.
(F) Bar charts quantify the time at which the Plk4-NG oscillations peaked, the length of S-phase, and the ratio between them (C/N)—derived from the data shown
in (E). Data are presented as mean ± SD. Statistical significance was assessed using an ordinary one-way ANOVA test (for Gaussian-distributed data) or a
Kruskal-Wallis test (**p < 0.01; ***p < 0.001; ns, not significant).
See also Figures 6, S1, and S2.
1568 Cell 181, 1566–1581, June 25, 2020
Article
Plk4 Oscillations Time and Execute Centriole
Biogenesis
To test whether the Plk4 oscillations were important for centriole
biogenesis, we generated flies co-expressing Plk4-NG (in a Plk4
mutant background) and the centriole cartwheel component
Sas-6-mCherry, which is irreversibly incorporated into the base
of the growing daughter centriole cartwheel and can be used
to monitor centriole growth in fly embryos (Aydogan et al.,
2018). These flies laid embryos that often failed to hatch (Fig-
ure S3C), but we simultaneously measured Plk4 oscillations
and centriole growth in those embryos that appeared to be
developing normally (Figures 2A, S3A, and S3B; Video S2). The
mother centrioles in these embryos were often slightly delayed
in initiating daughter centriole growth (Figures 2A, S3D, and
S3E), allowing us to measure the amount of Plk4 at the centrioles
when daughter centrioles either started or stopped growing (Fig-
ure 2A, colored dotted lines).
Strikingly, the centriolar levels of Plk4 at which centriole
growth initiated at each cycle (‘‘Start’’; Figures 2A and 2B)
were not significantly different than the levels at which centriole
growth stopped (‘‘Stop’’; Figures 2A and 2B). This suggests that
at each cycle there is a threshold level of centriolar Plk4 that is
required to support centriole growth: above this threshold the
centrioles can grow, below this threshold they cannot. If the
threshold concept is correct, then mother centrioles that failed
to recruit sufficient Plk4 should not grow a daughter. We
observed that the centrioles in a fraction of the embryos express-
ing both Plk4-NG and Sas-6-mCherry (mostly at nuclear cycle
13) separated at the start of S-phase but did not detectably
incorporate Sas-6-mCherry, indicating that daughter centrioles
did not grow (Figures S3D and S3E)—a defect that may explain
why many of these embryos failed to hatch (Figure S3C). Intrigu-
ingly, centriolar Plk4 levels continued to oscillate in these
embryos, but the average amplitude of these oscillations was
lower than in the embryos in which centrioles continued to dupli-
cate—and it was almost always below the average threshold at
which centriole growth was normally initiated (Figure 2C).
Together, these results suggest that the Plk4 oscillations initiate,
and determine the duration of, centriole growth.
Mathematical Modeling of the Plk4 Oscillation
Oscillations in biology are often generated by delayed feedback
circuits (Tsai et al., 2008). In Drosophila, Plk4 is recruited to cen-
trioles by Asterless (Asl), which also activates Plk4, allowing it to
phosphorylate both itself and Asl at multiple sites (Boese et al.,
2018; Dzhindzhev et al., 2010; Klebba et al., 2015). Human Asl
(Cep152) also binds, and is phosphorylated by, Plk4 in vitro (Ciz-
mecioglu et al., 2010; Hatch et al., 2010). We realized that this
 ll
Article

B C

Figure 2. Plk4 Oscillations Initiate and Time Centriole Biogenesis

(A) Graphs show the mean regression of Plk4-NG oscillations (red, green, and blue lines for cycles 11–13, respectively) and centriole growth (monitored by Sas-6-
mCherry incorporation, black lines) measured simultaneously in embryos during S-phase of cycles 11–13. For ease of presentation, the SEM for these data are
not shown, but are presented in Figures S3A and S3B. Dotted lines indicate the centriolar Plk4 levels at which centrioles ‘‘start’’ or ‘‘stop’’ growing. N = 17 embryos
(cycles 11 and 12), and 8 embryos (cycle 13); n = 19, 31, and 45 centrioles (mean) per embryo in cycles 11–13, respectively. See STAR Methods for an explanation
of data normalization and scaling.
(B) Bar charts quantify the centriolar Plk4-NG threshold levels at which centrioles start and stop growing during cycles 11–13—derived from the data shown in (A).
Data are presented as mean ± SD. Statistical significance was assessed using an unpaired t test with Welch’s correction (for Gaussian-distributed data) or an
unpaired Mann-Whitney test (ns, not significant).
(C) Eight embryos in which the centrioles did not grow (Figures S3D and S3E) were excluded from the analysis shown in (A) and (B); the scatter graph shown here
illustrates how the mean amplitude of the Plk4 oscillations in each of these eight embryos (red dots) tended to be lower than the mean amplitude (±SEM) of the
Plk4 oscillations in the embryos where the centrioles did grow.
See also Figure S3.

system could form a time-delayed negative feedback network
capable of generating Plk4 oscillations if the activation of Plk4
by Asl eventually led to the inhibition of their interaction.
A simple version of such a scenario is illustrated in Figures 3A
and 3B. At the start of each oscillation cycle, we envisage that
unphosphorylated Asl receptors on the mother centriole recruit
Plk4 to the site of daughter centriole assembly with high affinity
(Figure 3A, (i)). Binding activates Plk4, allowing it to phosphory-
late itself (Cunha-Ferreira et al., 2013; Holland et al., 2010;
Klebba et al., 2013), Ana2/STIL (Dzhindzhev et al., 2014; Kratz
et al., 2015; McLamarrah et al., 2018; Ohta et al., 2014) and
Asl/Cep152 (Boese et al., 2018; Hatch et al., 2010) at multiple
sites (Figure 3A, (ii)). The phosphorylated Ana2 promotes cart-
wheel assembly, potentially explaining why a threshold level of
Plk4 is required to promote centriole growth—but in our model
this reaction is not important for the Plk4 oscillation per se, so
we do not consider it further. We speculate that the phosphory-
lation of Asl at multiple sites reduces its affinity for Plk4, so that
the bound Plk4 molecules are released, leaving behind the phos-
phorylated Asl-receptor that can no longer recruit Plk4 (Fig-
ure 3A, (iii)) (see the end of this section for how this network
can be reset to trigger subsequent rounds of oscillations).
This network (Figure 3B; see mathematical model 1 in STAR
Methods) maps onto a set of coupled linear ordinary differential
equations, which we solved analytically. Solutions to this first
model (model 1 in STAR Methods) fit the discrete Plk4 oscillation
data from each S-phase of nuclear cycles 11–13 very well (Fig-
ure 3C; R2 > 0.99). Although the model may overfit the data,

Cell 181, 1566–1581, June 25, 2020 1569

 ll
Article

A B

D E

Figure 3. A Simple Mathematical Model of the Plk4 Oscillation and Experimental Investigations to Test Its Predictions
(A) Diagram of the model. (i) During mitosis, Asl receptors (red) on the surface of the mother centriole start to bind Plk4 (green) with high affinity (k1). (ii) Once bound,
Plk4 is activated, and it starts to phosphorylate itself, Ana2 (black) and Asl (k2) at multiple sites (indicated by dotted black arrows and black dots). (iii) We speculate
that, after several rounds of phosphorylation, Asl is converted to a state with low affinity for Plk4, so phosphorylated Plk4 is released (k3)—and likely degraded.
These Asl receptors are now inactivated and can no longer bind Plk4 to promote centriole growth.
(B) Schematic depicts the topology of the mathematical model (see STAR Methods for full details of the model). Asl-p and Plk4-p indicate phosphorylated
proteins. Bold arrows indicate the dominant direction of the reactions. This model discretely examines centriolar Plk4-NG levels only during S-phase of each
cycle. We speculate that a phosphatase normally removes the phosphate groups from Asl during mitosis to reset the system for the next oscillation (red arrow; k4),
and we extend the model to include this step elsewhere (Figures S4A and S4B).
(C) Graphs show the Lorentzian fit of the Plk4-NG oscillation data during S-phase of cycles 11–13 (solid lines) overlaid with analytical solutions to the model
(dotted lines). R2 values indicate goodness-of-fit.
(D) Bar charts quantify the average cytosolic concentration and centriolar fluorescence of Asl-GFP at the start of S-phase of each cycle. Cytosolic Asl-GFP was
measured using FCS; each data point represents the average of 4–6 10-s recordings from a single embryo (Figure S5). Centriolar Asl-GFP was measured using
confocal microscopy, as described in STAR Methods. N R 14 embryos for each cell cycle; n = 48, 70, and 130 centrioles (mean) per embryo in cycles 11–13,
respectively. Data are presented as mean ± SEM. Statistical significance was assessed using an ordinary one-way ANOVA test (ns, not significant).

(legend continued on next page)

1570 Cell 181, 1566–1581, June 25, 2020


Article
these solutions were within a reasonable and generally narrow
parameter space (Figures 3C and S5; Data S1, first and fourth
charts). Nevertheless, we believe this model is likely to be over-
simplified. Plk4’s ability to phosphorylate itself, for example,
could help to generate the oscillation by promoting Plk4 degra-
dation (Cunha-Ferreira et al., 2009; Guderian et al., 2010;
Holland et al., 2010; Rogers et al., 2009) or lowering the affinity
of the Asl::Plk4 interaction—as has recently been demonstrated
(Park et al., 2019). Moreover, the model considers the behavior
of only Asl and Plk4, when other factors, such as Ana2/STIL,
are likely to modulate the systems behavior (Arquint and Nigg,
2016; Gönczy and Hatzopoulos, 2019). Finally, the model does
not consider the possibility that Plk4 bound to one receptor
could phosphorylate nearby receptors, or the Plk4 bound to
nearby receptors, to influence their behavior—a concept that
may be important when considering how Plk4 ultimately local-
izes to only a single site on the side of the mother centriole
(Leda et al., 2018; Takao et al., 2019).
In order to demonstrate how this network could be reset for the
next oscillation, we extended our model (model 2 in STAR
Methods) to allow a protein phosphatase (PPTase) to be acti-
vated during M-phase to dephosphorylate Asl (Figures 3B, red
arrow, and S4A). This resetting is biologically plausible, because
the activities of several PPTases are regulated during the cell cy-
cle (Nilsson, 2019). This model can be solved exactly, and its so-
lutions generate robust centriolar Plk4 oscillations within the
context of a system that, like the early Drosophila embryo, alter-
nates between periods of S- and M-phases (Figure S4B). Thus,
our minimal model illustrates that a classical ‘‘time delayed nega-
tive-feedback’’ network (Novák and Tyson, 2008) can generate
Plk4 oscillations, although the precise molecular details of this
system remain to be fully elucidated.
Testing Predictions of the Mathematical Models
A key feature of our models is that the phosphorylation of Asl by
Plk4 reduces their affinity (although, as discussed above, Plk4’s
ability to phosphorylate itself, and other factors, could also help
to generate the oscillation). To test the plausibility of this idea, we
mutated 13 potential Plk4 phosphorylation sites in Asl to Ala (Asl-
13A) (Figure S4C). These sites were selected based on their con-
servation, their similarity to known Plk-family consensus sites
(Leung et al., 2007), their proximity to the N- and C-terminal re-
gions of Asl that are thought to interact with Plk4 (Boese et al.,
2018), and a previous analysis of sites in the Asl N-terminal re-
gion that are either phosphorylated by Plk4 kinase domain
in vitro or have been shown to be phosphorylated in cultured
Drosophila cells (Boese et al., 2018). If some of these sites are
normally phosphorylated by Plk4 to reduce the affinity of the
Asl::Plk4 interaction, we would predict that expressing Asl-13A
in the presence of endogenous, unlabeled Asl would lead to an
increase in centriolar Plk4-NG levels—because the Plk4 should
unbind from the mutant Asl receptors less efficiently. Although
(E) Bar chart shows the relative abundance of Plk4-NG at the start of each nuclear cycle measured by PeCoS (see STAR Methods; Figure S6). Each data point
represents a single 180-s recording from a single embryo. Statistical significance was assessed using a Kruskal-Wallis test (****p < 0.0001). Data are presented as
mean ± SD.
See also Figures S4, S5, and S6 and Data S1 (first and fourth charts as well as the Monte Carlo analysis).
ll
Asl-13A-mKate2 localized to centrioles less efficiently than Asl-
WT-mKate2 (Figure S4D), expressing untagged Asl-13A
increased the amplitude of the Plk4-NG oscillation (Figure S4E),
consistent with our idea that phosphorylating Asl can reduce its
affinity for Plk4 (Figures 3A and 3B).
An inspection of the parameters generated by our model re-
vealed that the reduction in the amplitude of the Plk4 oscillation
at successive nuclear cycles was driven primarily by a reduction
in the cytosolic concentration of Plk4 (that determines k1, the rate
at which Plk4 binds to Asl), while total levels of the Asl receptor
(Atot) remain relatively constant (Data S1, first chart). To test if
this was the case, we first used fluorescence correlation spec-
troscopy (FCS) (Figure S5) to examine the cytosolic concentra-
tion of Asl-GFP. Although the number of centrioles assembled
doubles at each successive cycle, the average cytosolic con-
centration of Asl-GFP, and the average centriolar levels of Asl-
GFP, remained relatively constant at the start of each successive
cycle (Figure 3D), as predicted by our model. Unfortunately, the
cytosolic concentration of Plk4-NG was too low to be measured
by conventional FCS, so we developed a new method, peak
counting spectroscopy (PeCoS), to measure relative protein
abundance at lower concentrations (see STAR Methods) (Fig-
ure S6). This revealed that, in contrast to Asl-GFP, the cytosolic
levels of Plk4-NG tended to decrease at successive nuclear cy-
cles (Figure 3E), as predicted by the model.
Why do cytosolic Plk4 levels decrease at successive nuclear cy-
cles? Our modeling suggests that if total Plk4 levels in the devel-
oping embryo remain constant (i.e., the rate of Plk4 degradation
and synthesis are balanced), then the doubling of centriole
numbers at each cycle can lead to the depletion of cytosolic
Plk4—particularly during later nuclear cycles—as an increasing
fraction of the protein is sequestered by the increasing number
of centrioles (Figure S4F). Alternatively (or additionally), Plk4 mole-
cules that are activated by binding to Asl may be more likely to
phosphorylate themselves to stimulate their degradation, ensuring
that more Plk4 is degraded at each cycle as the number of centri-
oles increase. Interestingly, in either of these scenarios, increasing
centriole numbers lead to more Plk4 depletion from the cytosol,
potentially allowing embryos to effectively ‘‘count’’ their centrioles.
The Plk4 Oscillation Can Adapt to Changes in Plk4
Levels to Maintain a Constant Centriole Size
Our finding that cytosolic levels of Asl remain constant at succes-
sive cycles while cytosolic Plk4 levels decrease suggests a ratio-
nale for why centriole biogenesis may be regulated by an oscilla-
tory system. In our models, Asl effectively functions as an
integrator (Ferrell, 2016; Somvanshi et al., 2015) whose levels are
kept constant so that it can measure changes in the input (cytosolic
Plk4 levels) and adapt the oscillation to maintain a constant output
(centriole size). If this interpretation is correct, then the Plk4 oscil-
lation should adapt to maintain a constant centriole size when
Plk4 levels change, but not when Asl levels change. To test this,
Cell 181, 1566–1581, June 25, 2020 1571
 ll
Article

Figure 4. The Plk4 Oscillator Can Adapt to Changes in Plk4 Concentration but Not to Changes in Asl Concentration
(A and B) Graphs show the regression data (solid lines) and mathematical solutions (dotted lines) for Plk4-NG oscillations in cycle 12 for experiments where either
(A) the genetic dose of Plk4-NG was halved (Plk4-NG1/2), or (B) the genetic dose of asl was halved (asl1/2) (gray lines) compared to controls (green lines). (A) N R 11
embryos for each condition; n = 47 and 42 centrioles (mean) per embryo in control or Plk4-NG1/2 groups, respectively. (B) N = 18 embryos for each condition; n =
44 and 43 centrioles (mean) per embryo in control or asl1/2 groups, respectively. Data are presented as mean ± SEM. Bar charts quantify oscillation parameters,
as indicated; data are presented as mean ± SD.

(legend continued on next page)

1572 Cell 181, 1566–1581, June 25, 2020


Article
we monitored Plk4-NG oscillations in embryos laid by mothers
where we genetically halved the dose of either Plk4-NG (hereafter
Plk4-NG1/2 embryos) or asl (hereafter asl1/2 embryos). Centrioles
appeared to duplicate normally in both sets of embryos, but the
Plk4 oscillation parameters were altered: in Plk4-NG1/2 embryos,
A decreased but there was a compensatory increase in T, so U re-
mained relatively constant (Figures 4A and S7A); in asl1/2 embryos,
A decreased, but there was no compensatory change in T, so U
decreased (Figures 4B and S7B–S7D).
Our mathematical model (model 1) could fit both sets of data
well (Figures 4A and 4B; R2 > 0.99), generating a reasonable range
of parameters (Data S1, second and third charts), several of which
we again validated experimentally (Figure S7; see mathematical
modeling section in STAR Methods). Interestingly, if we took the
normal parameters derived from our model and simply adjusted
the amount of Asl or Plk4 in the model to the levels we experimen-
tally measured in the half-dose embryos, the model fit the data
less well (not shown). This suggests that changing the concentra-
tion of one component is likely to influence the concentration and/
or behavior of other components so that several parameters of the
Plk4 oscillation are altered. This seems plausible, as the core
centriole duplication proteins are known to interact with and influ-
ence each other in multiple ways (Arquint and Nigg, 2016; Gönczy
and Hatzopoulos, 2019; Nigg and Holland, 2018).
Consistent with our observation that the Plk4-NG oscillations
adapt in Plk4-NG1/2 embryos by reducing A and increasing T to
maintain a relatively constant U, we previously showed that halving
the genetic dose of Plk4 led to the centrioles growing slowly, but for
a longer period of time, to maintain a constant size (Aydogan et al.,
2018). In contrast, we would predict that daughter centrioles in
asl1/2 embryos should grow more slowly (as A is decreased), but
for a normal period (as T is unchanged), and so centrioles would
be too short (as U decreases). We measured the parameters of
daughter centriole growth in asl1/2 embryos and confirmed that
this was the case (Figure 4C). Together, these experiments sug-
gest that the Plk4 oscillatory network functions to maintain a con-
stant centriole size even when Plk4 levels vary.
Plk4 Oscillations Can Execute Centriole Duplication
Independently of a Robust Cdk/Cyclin Cell-Cycle
Oscillator
Although the Plk4 oscillations in fly embryos are normally en-
trained by the cell-cycle oscillator (CCO) (Figures 1E and 1F), it
has long been known that centrioles can continue to duplicate
in many systems even when several other aspects of cell-cycle
progression are blocked (Balczon et al., 1995; Gard et al., 1990;
Sluder et al., 1990). We wondered whether this might be because
Plk4 oscillations can continue to drive centriole biogenesis even in
the absence of a robust CCO. To test this possibility, we injected
embryos with double-stranded RNAs (dsRNAs) targeting the three
embryonic mitotic cyclins: A, B, and B3. These embryos arrest in
(C) Graph quantifies the parameters of cartwheel growth—as measured by Sas-6-GFP fluorescence incorporation (Aydogan et al., 2018)—in WT and asl1/2
embryos; data are presented as mean ± SEM. Bar charts quantify growth parameters presented as mean ± SD. N = 17 embryos for each condition; n = 77 and 72
centrioles (mean) per embryo in WT or asl1/2 groups, respectively. Statistical significance was assessed using an unpaired t test with Welch’s correction (for
Gaussian-distributed data) or an unpaired Mann-Whitney test (*p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001; ns, not significant). R2 values indicate goodness-
of-fit for the mathematical solutions.
See also Figure S7 and Data S1 (second and third charts).
ll
an interphase-like state with intact nuclei that do not duplicate
their DNA, but where centrosomes can continue to duplicate
(McCleland and O’Farrell, 2008). We initially injected embryos in
nuclear cycles 7–8 and monitored Plk4-NG behavior
30 min
later. In all such embryos, we observed an initial synchronous
round of centriole duplication without NEB (indicating that the
CCO was perturbed), followed by one or more rounds of less syn-
chronous centriole duplication (Figures 5A and S2B; Video S3).
Strikingly, a normal Plk4-NG oscillation was associated with the
first, synchronous, round of centriole duplication, but subsequent
oscillations were more variable (Figures 5A and S2B).
We reasoned that any residual Plk4-NG oscillations in these
embryos might be triggered by residual CCO oscillations that
could trigger centriole duplication, but not DNA synthesis or
NEB. While one can never rule out the possibility of residual
CCO activity, we tried to overcome this potential problem by
examining centriole behavior in embryos in which the CCO was
likely to be more fully suppressed by injecting the embryos earlier
(nuclear cycles 2–4) and monitoring them later (after
90 min). The
centrioles in these embryos were now completely dissociated
from the non-dividing nuclei and they appeared to divide stochas-
tically, with some centrioles duplicating one or more times, and
others not duplicating at all (Figure S8; Video S4). The CCO coor-
dinates cell-cycle events in normal early embryos by spreading as
a chemical trigger wave (Chang and Ferrell, 2013; Deneke et al.,
2016), but duplicating centrioles did not detectably trigger the
duplication of nearby centrioles (Figure S8F). Thus, the ‘‘decision’’
to duplicate in these CCO-suppressed embryos appears to be
largely intrinsic to each individual centriole.
To test whether these stochastic centriole duplications were
triggered by Plk4 oscillations, we measured Plk4-NG fluores-
cence levels at individual centrioles. The raw intensity data were
noisy, but duplicating ‘‘fertile’’ centrioles appeared to exhibit
more prominent Plk4-NG oscillations than non-duplicating ‘‘ster-
ile’’ centrioles (Figure 5B). Moreover, the average centriolar
Plk4-NG fluorescence level (expressed as signal-to-noise ratio
[SNR]) was significantly higher at fertile centrioles (Figure S8B),
and Plk4-NG SNR values could distinguish fertile and sterile cen-
trioles, correctly predicting centriole fertility or sterility
74% and

71% of the time, respectively (Figures S8C and S8D).
Upon filtering the raw oscillation data, we found that the peaks
of the Plk4-NG oscillations (see STAR Methods for a description
of peak-calling methodology) were often associated with
centriole duplication events (Figure 5B). An unbiased computa-
tional analysis of all the 45 fertile centrioles that we observed in
3 different embryos revealed that the predicted Plk4-NG oscilla-
tion peaks predicted centriole duplication events with high preci-
sion (40/49 Plk4-NG peaks were associated with a duplication
event that occurred within ±5 min of the peak) and recall (40/
52 duplication events occurred within ±5 min of a Plk4-NG oscil-
lation peak) (Figures 5C and 5D). Computer simulations revealed
Cell 181, 1566–1581, June 25, 2020 1573
 ll
Article

A B

C D E F

Figure 5. Plk4 Levels Can Continue to Oscillate and Promote Centriole Duplication Even When the CCO is Perturbed
(A) Graph shows the Plk4-NG oscillations in an embryo injected with dsRNA against cyclin A-B-B3; the schema above the graph illustrates the experimental
protocol. The nuclei in this embryo arrest in interphase, but centrioles go through an additional round of division—centriole separation (CS)—accompanied by a
Plk4 oscillation. See Figure S2B for additional examples; n = 30 centrioles (mean) per embryo.
(B) Graphs show the raw (black lines) and filtered (red lines) fluorescence intensity data of 3 individual ‘‘fertile’’ centrioles and 3 individual ‘‘sterile’’ centrioles within
the same cyclin-depleted embryo. The fertile centrioles duplicate (black dotted lines), and these events were often closely associated with computed Plk4
oscillation peaks (red dotted lines) (see STAR Methods for further details of the peak calling methodology).
(C) An unbiased computational analysis of all 45 fertile centrioles in 3 embryos reveals that >80% of the computationally detected Plk4 oscillation peaks occur
within 5 min of an experimentally observed duplication event. A simulation with randomly distributed centriole duplication events and Plk4 oscillation peaks
showed a mean time separation of 10.5 min (data not shown).
(D) Venn diagram shows how, using a 5-min window, the oscillation peaks can be used to predict duplication events with both high precision and high recall (40/49
Plk4 oscillation peaks are associated with a duplication event, and 40/52 duplication events are associated with a Plk4 oscillation peak).
(E) Graph shows the ability of Plk4 oscillation peaks to ‘‘retrieve’’ centriole duplication events across all peak prominences. All detected oscillation peaks were
ranked in order of their peak prominence from high to low (black dots) and assigned uniquely to a duplication event if within a 5-min time window. The graph then
plots the precision and recall values if the threshold for calling a peak were set as the peak prominence value of each peak (in descending order). Below the
detected peak that is associated with a peak prominence threshold of 0.12, the precision dramatically drops, suggesting the existence of a minimum peak
amplitude for centriole duplication. At this threshold, precision and recall are jointly optimized. Note, if there were no overall correlation between Plk4 peaks and a
duplication event, the integrated area under the curve across all peak prominences or average precision (AP) for the 5-min time window (AP5min) would be ~50%
(given by # duplications/(# duplications + # peaks)); so the score of ~75% indicates a meaningful correlation.
(F) Graph shows the correlation between the time of the computationally determined Plk4 peaks and their respective experimentally observed duplication events.
Correlation strength was examined using Pearson’s correlation coefficient (r < 0.40 weak; 0.40 < r < 0.60 moderate; r > 0.60 strong); significance of correlation
was determined by the p value (p < 0.05) (see STAR Methods for a full description of this analysis).
See also Figures S2 and S8.

that a random distribution of the duplication events lead to an based on amplitude revealed that the higher the amplitude of
average time of >10 min between the peaks and duplication the oscillation, the more likely it was to be associated with a
events, indicating that the observed association was not centriole duplication event (Figure 5E), while plotting the relative
random. Moreover, a rank ordering of the Plk4-NG oscillations timing of the Plk4-NG oscillations and the centriole duplication

1574 Cell 181, 1566–1581, June 25, 2020


Article
events revealed a strong positive correlation (Figure 5F; Pearson
r = 0.9580, p < 0.0001). We conclude that individual centrioles
can organize autonomous Plk4 oscillations that can drive
centriole duplication even in the absence of a robust CCO.
This potentially explains how centrioles can continue to dupli-
cate independently of many other cell-cycle events.
The CCO Can Phase-Lock the Plk4 Oscillation to
Coordinate Centriole Duplication with Other Cell-Cycle
Events
It is widely believed that the CCO acts primarily as a ‘‘ratchet’’
whose activity increases over the cell cycle to trigger the sequen-
tial execution of cell-cycle events such as DNA replication,
centriole duplication, nuclear envelope breakdown (NEB), and
spindle assembly (Stern and Nurse, 1996; Swaffer et al., 2016,
2018). An interesting alternative possibility is that the CCO could
act as a ‘‘phase-locker’’ whose function is simply to entrain the
phase of a network of autonomous oscillations, each of which
is responsible for the execution of a specific cell-cycle event
(Lu and Cross, 2010). The Plk4 oscillation appears to time and
execute centriole biogenesis, and it can trigger centriole duplica-
tion independently of a robust CCO, so it is an excellent candi-
date for such an autonomous oscillation.
To better understand how the CCO might entrain the Plk4
oscillation, we measured the average period of the stochastic
Plk4 oscillations in cyclin-depleted embryos (20.5 ± 4.6 min)
and compared this to the average period of the Plk4 oscillations
in cycles 11–12 (11.7 ± 0.7 min) and 12–13 (14.9 ± 1.7 min). The
natural period of the autonomous Plk4 oscillation in these early
embryos is therefore similar to, but slightly slower than, the
period of the Plk4 oscillations normally enforced by the CCO,
indicating that the CCO could entrain the Plk4 oscillation by
speeding up a phase of its natural cycle.
To examine which phase this might be, we tested for correla-
tions between various parameters of the Plk4 oscillation and the
length of S- or M-phase. During cycles 11–13, we observed a sig-
nificant correlation between the timing of the Plk4-NG oscillation
trough in M-phase and the duration of M-phase (Figure 6, lower
scatterplots in the light yellow panel), suggesting that the CCO en-
trains the Plk4 oscillation by speeding it up during M-phase. This is
consistent with our minimal model, in which the CCO entrains the
Plk4 oscillation by ensuring the rapid and coordinated dephos-
phorylation of Asl during M-phase (Figures S4A and S4B).
We also noticed an additional correlation between the peak of
the Plk4-NG oscillation and S-phase length in cycle 13 (Figure 6,
upper rightmost scatterplot in the light yellow panel). This is not
surprising, as a Wee1-dependent checkpoint dramatically slows
the CCO—and many other aspects of S-phase progression—
particularly during nuclear cycle 13 (Deneke et al., 2016; Stumpff
et al., 2004). Moreover, in Wee1/ embryos, the correlation be-
tween the Plk4-NG oscillation trough and M-phase length was
maintained (Figure 6, lower rightmost scatterplot in the light yel-
low panel), while the correlation between the Plk4-NG oscillation
peak and S-phase length was lost (Figure 6; upper rightmost
scatterplot in the light yellow panel), demonstrating that Wee1
can influence the Plk4 oscillation in S-phase. Interestingly, the
cytosolic levels of Plk4-NG were essentially the same in wild-
type (WT) and Wee1/ embryos (Figure S6E), indicating that
ll
cell-cycle regulators can influence the Plk4 oscillation without
changing Plk4’s cytosolic concentration. This supports the
model prediction that the drop in cytosolic Plk4 levels at succes-
sive nuclear cycles (Figure 3E) is not, on its own, sufficient to
account for the change in Plk4 oscillation parameters we
observe from cycles 11–13. This presumably explains why the
model requires several parameters to change slightly at each
successive cycle to best fit the data (Data S1, first chart).
Taken together, our observations are consistent with the
phase-locker model of cell-cycle regulation (Lu and Cross,
2010). We propose that the Plk4 oscillation may be an exemplar
of an autonomously oscillating system that can independently
drive a cellular event (centriole duplication), but that is normally
phase-locked by the CCO to ensure its proper coordination
with other biological events and with cell division.
A Model to Generate Autonomous Plk4 Oscillations in
the Absence of a CCO
How can a Plk4 oscillation be generated independently of the
CCO? Our mathematical model (model 2 in STAR Methods) cannot
explain this, as it requires a PPTase to reset the system specifically
during M-phase (Figures S4A and S4B). Interestingly, if we extend
the model to allow the PPTase to have a constant low-level of ac-
tivity (
10% of the level normally required to reset the system in M-
phase) (Figure S8G) this new model (model 3 in STAR Methods) re-
capitulates several features of centriole duplication in the cyclin-
depleted embryos (Figure S8H). This model predicts that after a
last round of mitosis the centrioles in the cyclin-depleted embryos
will undergo a single synchronous Plk4 oscillation (as all of the Asl
receptors start this first cyclin-depleted cycle in a dephosphory-
lated state), but subsequent Plk4 oscillations rapidly dampen as
the individual Asl receptors lose synchrony, and the system tends
toward a steady state—where some of the centriolar Asl receptors
are Plk4-bound and being phosphorylated, while others are not
Plk4-bound and are being dephosphorylated (Figure S8H). Intrigu-
ingly, the inherent noise in the system generated stochastic Plk4
oscillations that could plausibly drive centriole duplication (Fig-
ure S8H)—potentially mimicking the stochastic Plk4 oscillations
and centriole duplication events that we observe in the cyclin-
depleted embryos (Figure 5B).
In this model, each Asl receptor effectively behaves as an in-
dependent oscillator—alternating between a Plk4-bound form
that is being phosphorylated and a non-Plk4-bound form that
is being dephosphorylated. In the presence of the CCO, the
Asl receptors generate coordinated Plk4 oscillations because
the CCO synchronizes them every nuclear cycle by providing a
coordinated burst of PPTase activity during mitosis.
Plk4 Oscillations Are Detectable in Non-dividing Mouse
Liver Cells and Can Be Entrained by the Circadian Clock
In species as distant as cyanobacteria and mammals, the CCO can
be entrained to the circadian clock (Matsuo et al., 2003; Yang et al.,
2010). We wondered, therefore, whether the autonomous Plk4
oscillation could also be entrained by the circadian clock. We
examined a recently published diurnal proteome from non-regen-
erating mouse liver (Wang et al., 2018), where hepatocytes, the ma-
jor building blocks of the liver, are largely quiescent (Friedman,
2000). Several key cell-cycle regulators (such as Cdk1, cyclin E,
Cell 181, 1566–1581, June 25, 2020 1575
 ll
Article

Figure 6. The CCO Phase-Locks the Plk4 Oscillations in Mitosis of Cycles 11–13 Independently of Wee1 and in Interphase of Cycle 13 in a
Wee1-Dependent Manner
Scatterplots illustrate correlations between various parameters of the Plk4 oscillation and the length of S- or M-phase in nuclear cycles 11–13 in WT and Wee1/
embryos (see Plk4-NG smooth curve fitting and parameter extraction in STAR Methods for details of how these parameters [along with their descriptions] were
obtained in an unbiased way). During cycles 11–13, there is a significant correlation between the timing of the Plk4 oscillation trough in M-phase and the duration
of M-phase (lower scatterplots in the light yellow panel), suggesting that the CCO entrains the Plk4 oscillation during M-phase. This entrainment is not altered in
the Wee1/ embryos. During nuclear cycle 13, there is an additional correlation between the peak of the Plk4 oscillation and S-phase length that is lost in the
Wee1/ embryos (upper rightmost scatterplot in the light yellow panel). The plots for the WT group were generated with the data obtained from Figures 1A and
S2A, as well as 5 additional embryos of the same genotype. N = 10 embryos; n = 23 centrioles (mean; starting from cycle 11) per embryo in Wee1/ group.
Correlation strength was examined using Pearson’s correlation coefficient (r < 0.40 weak; 0.40 < r < 0.60 moderate; r > 0.60 strong); significance of correlation
was determined by the p value (p < 0.05).
See also Figures S4 and S6E.

cyclin B1, and Plk1) were not detectable at any stage of the diurnal
cycle, confirming that these cells were largely quiescent. In
contrast, Plk4 protein (but not transcript) levels exhibited a striking
oscillation that was entrained to the light/dark cycles (Figures 7A–
7C). We presume that this oscillation is sub-threshold for centriole
biogenesis—because centrioles should not be duplicating in these
non-dividing cells—and simply reflects the ability of the Plk4 sys-
tem to oscillate in a way that can be entrained by the circa-
dian clock.
In our model, a mitotic PPTase that dephosphorylates Asl-re-
ceptors out of phase with Plk4 is required to generate Plk4 oscil-
lations (Figures S4A and S4B). We therefore used the mouse
dataset to examine the behavior of the mouse homologs of all
the mitotic PPTase subunits that function in flies (Chen et al.,
2007). Among the 27 PPTase subunits examined, only PPP2CB
exhibited a clear oscillatory behavior that is similar to Plk4, and
the period of these oscillations was precisely out of phase with
the Plk4 oscillation (Figure 7D, highlighted with a red dotted
frame). Intriguingly, PPP2CB is the homolog of Mts, the catalytic
subunit of PP2A in Drosophila that localizes to centrosomes spe-
cifically during mitosis in fly cells, and its knockdown leads to
centrosome duplication defects (Dobbelaere et al., 2008).
Thus, PP2A is an excellent candidate for the PPTase that may
normally dephosphorylate centriolar Asl during mitosis.

1576 Cell 181, 1566–1581, June 25, 2020

 ll
Article

A C

(legend on next page)

Cell 181, 1566–1581, June 25, 2020 1577

 ll
Article

Remarkably,
8% of the
6,800 proteins in the mouse data-
set exhibited a 24 h-entrained oscillatory behavior. It is unclear
why so many proteins oscillate in this way, or whether any of
these oscillations are of functional significance. Nevertheless,
these observations indicate that there are many other proteins,
and so perhaps many different biological processes, that have
a largely under-appreciated ability to oscillate.
Concluding Remarks
There is great interest in determining the physical and molecular
principles that cells use to regulate the biogenesis of their organ-
elles (Liu et al., 2018; Mukherji and O’Shea, 2014). The idea that
an organelle-specific oscillation could time and execute organelle
biogenesis has, to our knowledge, not been proposed previously.
We suggest that the Plk4 centriole oscillation could be a paradigm
for a general mechanism describing the regulation of organelle
biogenesis: oscillations in the levels/activity of key regulatory fac-
tors essential for organelle biogenesis could precisely time the initi-
ation and duration of the growth process, ensuring that organelles
grow at the right time and to the appropriate size. In such a model,
the CCO and circadian clocks could act simply as ‘‘phase-lockers’’
(Lu and Cross, 2010; Morgan, 2010), whose function is to entrain
the phase of a network of autonomous oscillators to ensure that
biological processes occur in a coordinated manner.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B D. melanogaster stocks and husbandry
d METHOD DETAILS
B Hatching experiments
B Synthesis of double-stranded RNA
B Embryo collections and dsRNA injections
B Immunoblotting
B Image acquisition, processing, and analysis
B Analysis of centriole ‘‘fertility’’ in embryos injected with
dsRNA against cyclin A-B-B3
B Spatiotemporal heatmap of centriole duplications
B Spatial clustering assessment of centriole duplications
B Plk4-NG smooth curve fitting and parameter extraction
B 3D-Structured Illumination Microscopy (3D-SIM)
B Mathematical modeling and its experimental validation
B Model 2: Generating robust Plk4 oscillations entrained
by the CCO
B Model 3: Stochastic duplications
B Fluorescence Correlation Spectroscopy (FCS)
B FCS background corrections
B Data restriction
B Peak Counting Spectroscopy (PeCoS)
d QUANTIFICATION AND STATISTICAL ANALYSIS
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.018.
ACKNOWLEDGMENTS
We are grateful to Laura Hankins, Fabio Echegaray Iturra, Marjorie Fournier,
Christoffer Lagerholm, and Bela Novak for advice and discussion and Alissa
M. Kleinnijenhuis and members of the Raff laboratory for critically reading
the manuscript. Microscopy was performed at the Micron Oxford Advanced
Bioimaging Unit, funded by a Strategic Award from the Wellcome Trust
(107457). The research was funded by a Wellcome Trust Senior Investigator
Award (104575 to T.L.S., M.M., Z.M.W., A.W., S.S., and Z.A.N.), Edward Pen-
ley Abraham Scholarships (to M.G.A. and L.G.), a Cancer Research UK Oxford
Centre Prize DPhil Studentship (C5255/A23225 to S.-S.W.), a Balliol Jason Hu
Scholarship, (to S.-S.W.), a Clarendon Scholarship (to S.-S.W.), and Ludwig
Institute for Cancer Research funding (to F.Y.Z.). M.A.B. was supported by a
Biotechnology and Biological Sciences Research Council grant (BB/
N016858/1) and the St. Cross Emanoel Lee Junior Research Fellowship.
AUTHOR CONTRIBUTIONS
This study was conceptualized by M.G.A., T.L.S., M.A.B., and J.W.R. Investi-
gation was done by M.G.A., T.L.S., M.M., Z.M.W., L.G., A.W., S.S., and M.A.B.
Data were analyzed by M.G.A., T.L.S., Z.M.W., L.G., F.Y.Z., and M.B.A. Meth-
odology was developed by M.G.A., T.L.S., M.M., Z.M.W., S.-S.W., F.Y.Z.,
M.A.B., and J.W.R. Project was administered by M.G.A., M.A.B., and J.W.R.
Resources were shared/made by M.G.A., M.M., L.G., A.W., S.S., S.-S.W.,
Z.A.N., and M.A.B. Software development was carried out by M.G.A., T.L.S.,
Z.M.W., S.-S.W., F.Y.Z., and M.A.B. Overall supervision was done by
M.G.A., A.G., and J.W.R. Validation experiments/analyses were carried out
by M.G.A., A.W., S.S., and J.W.R. M.G.A., T.L.S., A.G., M.A.B., and J.W.R.
wrote and edited the draft with significant input from all authors.

Figure 7. Plk4 Levels May Autonomously Oscillate in Mouse Liver Cells Entrained by the Circadian Clock Where Levels of PP2A Catalytic
Subunit Oscillates Precisely out of Phase
(A) Diagram shows the workflow used by Wang et al. (2018) to obtain a diurnal proteome of the whole liver of light/dark-entrained mice.
(B) Graphs reproduced from Wang et al. (2018) show the relative diurnal expression of the circadian clock transcripts Bmal1 and Per1 as internal controls. We re-
analyzed the diurnal proteome produced in this study—comprising a matrix of Z scores for 6,780 proteins identified during 2 circadian cycles (supplemental
dataset 9 from Wang et al. [2018]).
(C) Graphs we derived show the relative protein levels of Plk4 and the cartwheel component STIL (Ana2 in flies). Plk4 levels strongly spike in a periodic manner
every circadian cycle, whereas STIL levels appear to randomly fluctuate and show neither a discernible pattern of oscillation nor any entrainment to the circadian
clock. Because these cells are generally not proliferating, centrioles should not be duplicating, so the Plk4 oscillations are presumably sub-threshold for centriole
biogenesis. Thus, Plk4 oscillations are detectable in non-dividing mammalian cells, where they are entrained by the circadian clock.
(D) Graphs examine in the non-dividing liver cells the behavior of mouse homologs of all the mitotic PPTase subunits that function in flies (Chen et al., 2007).
Among the 27 PPTase subunits examined, only PPP2CB (highlighted with a red dotted frame) exhibited a clear oscillatory behavior that is similar to Plk4, and the
period of these oscillations was precisely out of phase with the Plk4 oscillation.

1578 Cell 181, 1566–1581, June 25, 2020


Article
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: April 5, 2019
Revised: December 19, 2019
Accepted: May 8, 2020
Published: June 11, 2020
REFERENCES
Alvarez-Rodrigo, I., Steinacker, T.L., Saurya, S., Conduit, P.T., Baumbach, J.,
Novak, Z.A., Aydogan, M.G., Wainman, A., and Raff, J.W. (2019). Evidence
that a positive feedback loop drives centrosome maturation in fly embryos. eL-
ife 8, D430.
Arquint, C., and Nigg, E.A. (2016). The PLK4-STIL-SAS-6 module at the core of
centriole duplication. Biochem. Soc. Trans. 44, 1253–1263.
Aydogan, M.G., Wainman, A., Saurya, S., Steinacker, T.L., Caballe, A., Novak,
Z.A., Baumbach, J., Muschalik, N., and Raff, J.W. (2018). A homeostatic clock
sets daughter centriole size in flies. J. Cell Biol. 217, 1233–1248.
Balczon, R., Bao, L., Zimmer, W.E., Brown, K., Zinkowski, R.P., and Brinkley,
B.R. (1995). Dissociation of centrosome replication events from cycles of DNA
synthesis and mitotic division in hydroxyurea-arrested Chinese hamster ovary
cells. J. Cell Biol. 130, 105–115.
Ball, G., Demmerle, J., Kaufmann, R., Davis, I., Dobbie, I.M., and Schermelleh,
L. (2015). SIMcheck: a Toolbox for Successful Super-resolution Structured
Illumination Microscopy. Sci. Rep. 5, 15915.
Banterle, N., and Gönczy, P. (2017). Centriole Biogenesis: From Identifying
the Characters to Understanding the Plot. Annu. Rev. Cell Dev. Biol.
33, 23–49.
Basto, R., Brunk, K., Vinadogrova, T., Peel, N., Franz, A., Khodjakov, A., and
Raff, J.W. (2008). Centrosome amplification can initiate tumorigenesis in flies.
Cell 133, 1032–1042.
Baumbach, J., Novak, Z.A., Raff, J.W., and Wainman, A. (2015). Dissecting the
function and assembly of acentriolar microtubule organizing centers in
Drosophila cells in vivo. PLoS Genet. 11, e1005261.
Bettencourt-Dias, M., Hildebrandt, F., Pellman, D., Woods, G., and Godinho,
S.A. (2011). Centrosomes and cilia in human disease. Trends Genet. 27,
307–315.
Blachon, S., Gopalakrishnan, J., Omori, Y., Polyanovsky, A., Church, A., Ni-
castro, D., Malicki, J., and Avidor-Reiss, T. (2008). Drosophila asterless and
vertebrate Cep152 Are orthologs essential for centriole duplication. Genetics
180, 2081–2094.
Boese, C.J., Nye, J., Buster, D.W., McLamarrah, T.A., Byrnes, A.E., Slep, K.C.,
Rusan, N.M., and Rogers, G.C. (2018). Asterless is a Polo-like kinase 4 sub-
strate that both activates and inhibits kinase activity depending on its phos-
phorylation state. Mol. Biol. Cell 29, 2874–2886.
Chang, J.B., and Ferrell, J.E., Jr. (2013). Mitotic trigger waves and the spatial
coordination of the Xenopus cell cycle. Nature 500, 603–607.
Chen, F., Archambault, V., Kar, A., Lio’, P., D’Avino, P.P., Sinka, R., Lilley,
K., Laue, E.D., Deak, P., Capalbo, L., and Glover, D.M. (2007). Multiple
protein phosphatases are required for mitosis in Drosophila. Curr. Biol.
17, 293–303.
Cizmecioglu, O., Arnold, M., Bahtz, R., Settele, F., Ehret, L., Haselmann-
Weiss, U., Antony, C., and Hoffmann, I. (2010). Cep152 acts as a scaffold
for recruitment of Plk4 and CPAP to the centrosome. J. Cell Biol. 191,
731–739.
Claude, A. (1943). The constitution of protoplasm. Science 97, 451–456.
Conduit, P.T., Wainman, A., Novak, Z.A., Weil, T.T., and Raff, J.W. (2015). Re-
examining the role of Drosophila Sas-4 in centrosome assembly using two-
colour-3D-SIM FRAP. eLife 4, 1032.
Cunha-Ferreira, I., Rodrigues-Martins, A., Bento, I., Riparbelli, M., Zhang, W.,
Laue, E., Callaini, G., Glover, D.M., and Bettencourt-Dias, M. (2009). The SCF/
ll
Slimb ubiquitin ligase limits centrosome amplification through degradation of
SAK/PLK4. Curr. Biol. 19, 43–49.
Cunha-Ferreira, I., Bento, I., Pimenta-Marques, A., Jana, S.C., Lince-Faria, M.,
Duarte, P., Borrego-Pinto, J., Gilberto, S., Amado, T., Brito, D., et al. (2013).
Regulation of autophosphorylation controls PLK4 self-destruction and
centriole number. Curr. Biol. 23, 2245–2254.
Deneke, V.E., Melbinger, A., Vergassola, M., and Di Talia, S. (2016). Waves of
Cdk1 Activity in S Phase Synchronize the Cell Cycle in Drosophila Embryos.
Dev. Cell 38, 399–412.
Dzhindzhev, N.S., Yu, Q.D., Weiskopf, K., Tzolovsky, G., Cunha-Ferreira, I., Ri-
parbelli, M., Rodrigues-Martins, A., Bettencourt-Dias, M., Callaini, G., and
Glover, D.M. (2010). Asterless is a scaffold for the onset of centriole assembly.
Nature 467, 714–718.
Dobbelaere, J., Josué, F., Suijkerbuijk, S., Baum, B., Tapon, N., and Raff, J.
(2008). A genome-wide RNAi screen to dissect centriole duplication and
centrosome maturation in Drosophila. PloS. Biol. 6, e224.
Dzhindzhev, N.S., Tzolovsky, G., Lipinszki, Z., Schneider, S., Lattao, R.,
Fu, J., Debski, J., Dadlez, M., and Glover, D.M. (2014). Plk4 phosphory-
lates Ana2 to trigger Sas6 recruitment and procentriole formation. Curr.
Biol. 24, 2526–2532.
Eilers, P.H., and Boelens, H.F. (2005). Baseline Correction with Asymmetric
Least Squares Smoothing. Leiden University Medical Centre report, 2005.
Ferrell, J.E., Jr. (2016). Perfect and Near-Perfect Adaptation in Cell Signaling.
Cell Syst. 2, 62–67.
Fırat-Karalar, E.N., and Stearns, T. (2014). The centriole duplication cycle.
Philos. Trans. R. Soc. Lond. B Biol. Sci. 369, 20130460.
Friedman, S.L. (2000). Molecular regulation of hepatic fibrosis, an integrated
cellular response to tissue injury. J. Biol. Chem. 275, 2247–2250.
Gard, D.L., Hafezi, S., Zhang, T., and Doxsey, S.J. (1990). Centrosome dupli-
cation continues in cycloheximide-treated Xenopus blastulae in the absence
of a detectable cell cycle. J. Cell Biol. 110, 2033–2042.
Goehring, N.W., and Hyman, A.A. (2012). Organelle growth control through
limiting pools of cytoplasmic components. Curr. Biol. 22, R330–R339.
Gönczy, P., and Hatzopoulos, G.N. (2019). Centriole assembly at a glance.
J. Cell Sci. 132, jcs228833.
Guderian, G., Westendorf, J., Uldschmid, A., and Nigg, E.A.E. (2010). Plk4
trans-autophosphorylation regulates centriole number by controlling be-
taTrCP-mediated degradation. J. Cell Sci. 123, 2163–2169.
Hatch, E.M., Kulukian, A., Holland, A.J., Cleveland, D.W., and Stearns, T.
(2010). Cep152 interacts with Plk4 and is required for centriole duplication.
J. Cell Biol. 191, 721–729.
Holland, A.J., Lan, W., Niessen, S., Hoover, H., and Cleveland, D.W. (2010).
Polo-like kinase 4 kinase activity limits centrosome overduplication by autore-
gulating its own stability. J. Cell Biol. 188, 191–198.
Jacobs, H.W., Knoblich, J.A., and Lehner, C.F. (1998). Drosophila Cyclin B3 is
required for female fertility and is dispensable for mitosis like Cyclin B. Genes
Dev. 12, 3741–3751.
Jones, E., Oliphant, T., and Peterson, P. (2001). SciPy: Open Source Scientific
Tools for Python. http://www.scipy.org/.
Kitagawa, D., Vakonakis, I., Olieric, N., Hilbert, M., Keller, D., Olieric, V.,
Bortfeld, M., Erat, M.C., Flückiger, I., Gönczy, P., and Steinmetz, M.O.
(2011). Structural basis of the 9-fold symmetry of centrioles. Cell 144,
364–375.
Klebba, J.E., Buster, D.W., Nguyen, A.L., Swatkoski, S., Gucek, M., Rusan,
N.M., and Rogers, G.C. (2013). Polo-like kinase 4 autodestructs by generating
its Slimb-binding phosphodegron. Curr. Biol. 23, 2255–2261.
Klebba, J.E., Galletta, B.J., Nye, J., Plevock, K.M., Buster, D.W., Hollings-
worth, N.A., Slep, K.C., Rusan, N.M., and Rogers, G.C. (2015). Two Polo-like
kinase 4 binding domains in Asterless perform distinct roles in regulating ki-
nase stability. J. Cell Biol. 208, 401–414.
Koppel, D.E. (1974). Statistical accuracy in fluorescence correlation spectros-
copy. Phys. Rev. A 10, 1938–1945.
Cell 181, 1566–1581, June 25, 2020 1579
 ll
Kratz, A.-S., Bärenz, F., Richter, K.T., and Hoffmann, I. (2015). Plk4-dependent
phosphorylation of STIL is required for centriole duplication. Biol. Open 4,
370–377.
Leda, M., Holland, A.J., and Goryachev, A.B. (2018). Autoamplification and
Competition Drive Symmetry Breaking: Initiation of Centriole Duplication by
the PLK4-STIL Network. iScience 8, 222–235.
Leung, G.C., Ho, C.S.W., Blasutig, I.M., Murphy, J.M., and Sicheri, F. (2007).
Determination of the Plk4/Sak consensus phosphorylation motif using peptide
spots arrays. FEBS Lett. 581, 77–83.
Liu, T.-L., Upadhyayula, S., Milkie, D.E., Singh, V., Wang, K., Swinburne, I.A.,
Mosaliganti, K.R., Collins, Z.M., Hiscock, T.W., Shea, J., et al. (2018).
Observing the cell in its native state: Imaging subcellular dynamics in multicel-
lular organisms. Science 360, eaaq1392.
Lu, Y., and Cross, F.R. (2010). Periodic cyclin-Cdk activity entrains an auton-
omous Cdc14 release oscillator. Cell 141, 268–279.
Markow, T.A., Beall, S., and Matzkin, L.M. (2009). Egg size, embryonic devel-
opment time and ovoviviparity in Drosophila species. J. Evol. Biol. 22,
430–434.
Marsh, B.J., Mastronarde, D.N., Buttle, K.F., Howell, K.E., and McIntosh, J.R.
(2001). Organellar relationships in the Golgi region of the pancreatic beta cell
line, HIT-T15, visualized by high resolution electron tomography. Proc. Natl.
Acad. Sci. USA 98, 2399–2406.
Marshall, W.F. (2016). Cell Geometry: How Cells Count and Measure Size.
Annu. Rev. Biophys. 45, 49–64.
Matsuo, T., Yamaguchi, S., Mitsui, S., Emi, A., Shimoda, F., and Okamura, H.
(2003). Control mechanism of the circadian clock for timing of cell division
in vivo. Science 302, 255–259.
McCleland, M.L., and O’Farrell, P.H. (2008). RNAi of mitotic cyclins in Drosophila
uncouples the nuclear and centrosome cycle. Curr. Biol. 18, 245–254.
McLamarrah, T.A., Buster, D.W., Galletta, B.J., Boese, C.J., Ryniawec,
J.M., Hollingsworth, N.A., Byrnes, A.E., Brownlee, C.W., Slep, K.C., Ru-
san, N.M., and Rogers, G.C. (2018). An ordered pattern of Ana2 phos-
phorylation by Plk4 is required for centriole assembly. J. Cell Biol. 217,
1217–1231.
Morgan, D.O. (2010). The hidden rhythms of the dividing cell. Cell 141,
224–226.
Mukherji, S., and O’Shea, E.K. (2014). Mechanisms of organelle biogenesis
govern stochastic fluctuations in organelle abundance. eLife 3, e02678.
Nigg, E.A., and Holland, A.J. (2018). Once and only once: mechanisms of
centriole duplication and their deregulation in disease. Nat. Rev. Mol. Cell
Biol. 19, 297–312.
Nigg, E.A., and Raff, J.W. (2009). Centrioles, centrosomes, and cilia in health
and disease. Cell 139, 663–678.
Nilsson, J. (2019). Protein phosphatases in the regulation of mitosis. J. Cell
Biol. 218, 395–409.
Novák, B., and Tyson, J.J. (2008). Design principles of biochemical oscillators.
Nat. Rev. Mol. Cell Biol. 9, 981–991.
Novak, Z.A., Conduit, P.T., Wainman, A., and Raff, J.W. (2014). Asterless li-
censes daughter centrioles to duplicate for the first time in Drosophila em-
bryos. Curr. Biol. 24, 1276–1282.
Ohta, M., Ashikawa, T., Nozaki, Y., Kozuka-Hata, H., Goto, H., Inagaki, M.,
Oyama, M., and Kitagawa, D. (2014). Direct interaction of Plk4 with STIL en-
sures formation of a single procentriole per parental centriole. Nat. Commun.
5, 5267.
Park, J.-E., Zhang, L., Bang, J.K., Andresson, T., DiMaio, F., and Lee, K.S.
(2019). Phase separation of Polo-like kinase 4 by autoactivation and clustering
drives centriole biogenesis. Nat. Commun. 10, 4959.
Petrásek, Z., and Schwille, P. (2008). Precise measurement of diffusion coef-
ficients using scanning fluorescence correlation spectroscopy. Biophys. J. 94,
1437–1448.
1580 Cell 181, 1566–1581, June 25, 2020
Article
Price, D., Rabinovitch, S., O’Farrell, P.H., and Campbell, S.D. (2000).
Drosophila wee1 Has an Essential Role in the Nuclear Divisions of Early
Embryogenesis. Genetics 155, 159–166.
Ripley, B.D. (1976). The second-order analysis of stationary point processes.
J. Appl. Probab. 13, 255–266.
Rogers, G.C., Rusan, N.M., Peifer, M., and Rogers, S.L. (2008). A multi-
component assembly pathway contributes to the formation of acentroso-
mal microtubule arrays in interphase Drosophila cells. Mol. Biol. Cell 19,
3163–3178.
Rogers, G.C., Rusan, N.M., Roberts, D.M., Peifer, M., and Rogers, S.L. (2009).
The SCF Slimb ubiquitin ligase regulates Plk4/Sak levels to block centriole
reduplication. J. Cell Biol. 184, 225–239.
Rüttinger, S., Buschmann, V., Krämer, B., Erdmann, R., Macdonald, R., and
Koberling, F. (2008). Comparison and accuracy of methods to determine the
confocal volume for quantitative fluorescence correlation spectroscopy.
J. Microsc. 232, 343–352.
Schönle, A., Von Middendorff, C., Ringemann, C., Hell, S.W., and Eggeling, C.
(2014). Monitoring triplet state dynamics with fluorescence correlation spec-
troscopy: bias and correction. Microsc. Res. Tech. 77, 528–536.
Schwarz, G. (1978). Estimating the Dimension of a Model. Ann. Stat. 6,
461–464.
Shaner, N.C., Lambert, G.G., Chammas, A., Ni, Y., Cranfill, P.J., Baird, M.A.,
Sell, B.R., Allen, J.R., Day, R.N., Israelsson, M., et al. (2013). A bright mono-
meric green fluorescent protein derived from Branchiostoma lanceolatum.
Nat. Methods 10, 407–409.
Shcherbo, D., Murphy, C.S., Ermakova, G.V., Solovieva, E.A., Chepurnykh,
T.V., Shcheglov, A.S., Verkhusha, V.V., Pletnev, V.Z., Hazelwood, K.L., Roche,
P.M., et al. (2009). Far-red fluorescent tags for protein imaging in living tissues.
Biochem. J. 418, 567–574.
Sibon, O.C., Stevenson, V.A., and Theurkauf, W.E. (1997). DNA-replication
checkpoint control at the Drosophila midblastula transition. Nature
388, 93–97.
Sluder, G., Miller, F.J., Cole, R., and Rieder, C.L. (1990). Protein synthesis and
the cell cycle: centrosome reproduction in sea urchin eggs is not under trans-
lational control. J. Cell Biol. 110, 2025–2032.
Somvanshi, P.R., Patel, A.K., Bhartiya, S., and Venkatesh, K.V. (2015). Imple-
mentation of integral feedback control in biological systems. Wiley Interdiscip.
Rev. Syst. Biol. Med. 7, 301–316.
Stern, B., and Nurse, P. (1996). A quantitative model for the cdc2 control of S
phase and mitosis in fission yeast. Trends Genet. 12, 345–350.
Stumpff, J., Duncan, T., Homola, E., Campbell, S.D., and Su, T.T. (2004).
Drosophila Wee1 kinase regulates Cdk1 and mitotic entry during embryogen-
esis. Curr. Biol. 14, 2143–2148.
Swaffer, M.P., Jones, A.W., Flynn, H.R., Snijders, A.P., and Nurse, P. (2016).
CDK Substrate Phosphorylation and Ordering the Cell Cycle. Cell 167,
1750–1761.
Swaffer, M.P., Jones, A.W., Flynn, H.R., Snijders, A.P., and Nurse, P. (2018).
Quantitative Phosphoproteomics Reveals the Signaling Dynamics of Cell-Cy-
cle Kinases in the Fission Yeast Schizosaccharomyces pombe. Cell Rep. 24,
503–514.
Takao, D., Watanabe, K., Kuroki, K., and Kitagawa, D. (2019). Feedback loops
in the Plk4-STIL-HsSAS6 network coordinate site selection for procentriole
formation. Biol. Open 8, bio047175.
Tinevez, J.-Y., Perry, N., Schindelin, J., Hoopes, G.M., Reynolds, G.D.,
Laplantine, E., Bednarek, S.Y., Shorte, S.L., and Eliceiri, K.W. (2017).
TrackMate: An open and extensible platform for single-particle tracking.
Methods 115, 80–90.
Tsai, T.Y.-C., Choi, Y.S., Ma, W., Pomerening, J.R., Tang, C., and Ferrell, J.E.,
Jr. (2008). Robust, tunable biological oscillations from interlinked positive and
negative feedback loops. Science 321, 126–129.
van Breugel, M., Hirono, M., Andreeva, A., Yanagisawa, H.-A., Yamaguchi, S.,
Nakazawa, Y., Morgner, N., Petrovich, M., Ebong, I.-O., Robinson, C.V., et al.

Article
(2011). Structures of SAS-6 suggest its organization in centrioles. Science 331,
1196–1199.
van Breugel, M., Wilcken, R., McLaughlin, S.H., Rutherford, T.J., and Johnson,
C.M. (2014). Structure of the SAS-6 cartwheel hub from Leishmania major. eL-
ife 3, e01812.
Waithe, D., Clausen, M.P., Sezgin, E., and Eggeling, C. (2016). FoCuS-point:
software for STED fluorescence correlation and time-gated single photon
counting. Bioinformatics 32, 958–960.
ll
Wang, Y., Song, L., Liu, M., Ge, R., Zhou, Q., Liu, W., Li, R., Qie, J., Zhen, B.,
Wang, Y., et al. (2018). A proteomics landscape of circadian clock in mouse
liver. Nat. Commun. 9, 1553.
Yamamoto, S., and Kitagawa, D. (2019). Self-organization of Plk4 regulates
symmetry breaking in centriole duplication. Nat. Comm. 10, 1810.
Yang, Q., Pando, B.F., Dong, G., Golden, S.S., and van Oudenaarden, A.
(2010). Circadian gating of the cell cycle revealed in single cyanobacterial cells.
Science 327, 1522–1526.
Cell 181, 1566–1581, June 25, 2020 1581
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mouse anti-GFP Roche RRID: AB_390913
Mouse anti-Actin Sigma RRID: AB_476730
HRPO-linked anti-mouse IgG Sigma / GE Healthcare Cat# GENA931
Chemicals, Peptides, and Recombinant Proteins
QuikChange II XL mutagenesis kit Agilent Technologies Cat# 200521
Q5 Site Directed Mutagenesis kit New England Biolabs Cat# E0554S
Voltalef grade H10S oil Arkema N/A
Alexa Fluor 488 NHS Ester Thermo Fisher Scientific Cat# A20000
Experimental Models: Organisms/Strains
D. melanogaster: Plk4-mNeonGreen This paper N/A
D. melanogaster: Plk4Aa74 (Plk4 null mutant) Aydogan et al., 2018 FlyBase ID: FBab0049012
D. melanogaster: Asl-mKate2 This paper N/A
D. melanogaster: Sas-6-mCherry Rogers et al., 2008 N/A
D. melanogaster: CycB2 Jacobs et al., 1998 FlyBase ID: FBal0094855
D. melanogaster: grpfsA4 Sibon et al., 1997 FlyBase ID: FBal0062815
D. melanogaster: Asl-GFP Blachon et al., 2008 FlyBase ID:FBtp0040947
D. melanogaster: aslB46 Baumbach et al., 2015 FlyBase ID: FBal0343439
D. melanogaster: Plk4-GFP Aydogan et al., 2018 FlyBase ID: FBal0343977
D. melanogaster: Asl-mCherry Conduit et al., 2015 FlyBase ID: FBal0343645
D. melanogaster: Sas-6-GFP Aydogan et al., 2018 FlyBase ID: FBtp0131375
D. melanogaster: Asl-13A-mKate2 This paper N/A
D. melanogaster: Asl-13A This paper N/A
D. melanogaster: Asl This paper N/A
D. melanogaster: wee1* (Homozygous viable mutant derived from N/A
Price et al., 2000; courtesy of Prof. Shelagh
Campbell)
D. melanogaster: Plk4-mNeonGreen, This paper N/A
Plk4Aa74 / Plk4-mNeonGreen, Plk4Aa74
D. melanogaster: Asl-mKate2 / Cyo; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: Sas-6-mCherry / +; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: CycB2 / +; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: grpfsA4 / +; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: Asl-GFP / Asl-GFP; This paper N/A
aslB46 / aslB46
D. melanogaster: Oregon-R (Wild-type Kyoto Stock Center FlyBase ID: FBst0324696
strain)
D. melanogaster: Asl-mKate2, aslB46 / + This paper N/A
(Continued on next page)

e1 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
D. melanogaster: Plk4-GFP / Cyo; Plk4 Aa74
/ Aydogan et al., 2018 N/A
Plk4Aa74
D. melanogaster: Plk4-mNeonGreen, This paper N/A
Plk4Aa74 / Plk4Aa74
D. melanogaster: Asl-mCherry / +; This paper N/A
Plk4Aa74 / +
D. melanogaster: Plk4-mNeonGreen / +; This paper N/A
Plk4-mNeonGreen, Plk4Aa74 / Plk4Aa74
D. melanogaster: Plk4-mNeonGreen / +; This paper N/A
Plk4-mNeonGreen, Plk4Aa74 / aslB46,
Plk4Aa74
D. melanogaster: Asl-GFP / +; aslB46 / aslB46 This paper N/A
D. melanogaster: Plk4-GFP / Cyo; aslB46, This paper N/A
Plk4Aa74 / Plk4Aa74
D. melanogaster: Sas-6-GFP / +; aslB46 / + This paper N/A
D. melanogaster: Asl-13A-mKate2 / Asl- This paper N/A
13A-mKate2; aslB46 / aslB46
D. melanogaster: Asl-mKate2 / Asl-mKate2; This paper N/A
aslB46 / aslB46
D. melanogaster: Asl-13A / +; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: Asl / +; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
D. melanogaster: wee1* / wee1*; Plk4- This paper N/A
mNeonGreen, Plk4Aa74 / Plk4-
mNeonGreen, Plk4Aa74
Oligonucleotides
Primers to introduce the NheI restriction Invitrogen, Thermo Fisher Scientific N/A
enzyme sites into the mCherry C-terminal
Gateway vector, see Table S1.
Primers to replace the mCherry tag with Invitrogen, Thermo Fisher Scientific N/A
mNeonGreen by homologous
recombination on the destination vector,
see Table S1.
Primers to replace the mCherry tag with Invitrogen, Thermo Fisher Scientific N/A
mKate2 by homologous recombination on
the destination vector, see Table S1.
Primers to remove the NheI restriction Invitrogen, Thermo Fisher Scientific N/A
enzyme sites from the destination vector via
site-directed mutagenesis (mNeonGreen
vector), see Table S1.
Primers to remove the NheI restriction Invitrogen, Thermo Fisher Scientific N/A
enzyme sites from the destination vector via
site-directed mutagenesis (mKate2 vector),
see Table S1.
Primers to amplify Cyclin A, B or B3, see Invitrogen, Thermo Fisher Scientific N/A
Table S1.
Primers to introduce various site directed Invitrogen, Thermo Fisher Scientific N/A
mutations for Asl-13A construct, see
Table S1.
(Continued on next page)

Cell 181, 1566–1581.e1–e14, June 25, 2020 e2

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Primers to delete mKate2 to generate Invitrogen, Thermo Fisher Scientific N/A
endogenous Asl-13A construct without a
fluorescent tag, see Table S1.
Primers to generate endogenous Asl Invitrogen, Thermo Fisher Scientific N/A
construct without a fluorescent tag, see
Table S1.
Recombinant DNA
mCherry C-terminal Gateway vector Basto et al., 2008 N/A
pDONR-Zeo vector Thermo Fisher Scientific Cat# 12535035
mNeonGreen vector Shaner et al., 2013 N/A
mKate2 vector Shcherbo et al., 2009 N/A
Asl-mKate2 P-element transformation This study N/A
vector
Software and Algorithms
Fiji (ImageJ) National Institutes of Health https://imagej.nih.gov/ij/
TrackMate Tinevez et al., 2017 https://imagej.net/TrackMate
Prism 7 and 8 GraphPad https://www.graphpad.com/
scientific-software/prism/
Scipy’s find_peaks function Jones et al., 2001 https://docs.scipy.org/doc/scipy/
reference/generated/scipy.signal.
find_peaks.html
Asymmetric baseline smoothing Eilers and Boelens, 2005 N/A
Zen Black Software Zeiss https://www.zeiss.com/microscopy/us/
products/microscope-software/zen.html
FoCuS-Point Software Waithe et al., 2016 N/A
The equations used for mathematical This paper https://github.com/RaffLab/
modeling and regressions centriole_oscillator_model
Python script to automate PeCoS analysis This paper https://github.com/RaffLab/
centriole_oscillator_model

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Jordan W.
Raff (jordan.raff@path.ox.ac.uk).

Materials Availability
All unique/stable reagents generated in this study are available from the Lead Contact without restriction, unless for commercial
application, in which case a completed Materials Transfer Agreement will be requested. There is restriction to the availability of
dsRNA cocktails produced in this study, as they only last for
6 months without degradation (if preserved at conditions indicated
in the STAR Methods), and therefore these cocktails are recommended to be made fresh using the protocol described in the
STAR Methods. Fly alleles and plasmids (with original source species) generated in this study will be requested by FlyBase admin-
istration to deposit onto FlyBase public archives within 6 months following the publication of this study. Compound and recombinant
flies are deposited to the Lead Contact’s laboratory stocks (without direct public access), but are available without restriction upon
request.

Data and Code Availability

The codes generated to perform mathematical modeling and regressions are available in the following web link: < https://github.com/
RaffLab/centriole_oscillator_model >. The code generated to automate PeCoS analysis procedure is available in the following web
link: < https://github.com/RaffLab/PeCoS >. Source 3D time-lapse spinning-disk confocal micrographs and SIM reconstruction da-
tasets supporting the current study are of sizes between 10 and 20GB for each experiment (exceeding the current upload limits of
public repositories) and therefore have been deposited in Open Microscopy Environment (OMERO) repository. These are available

e3 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

without restriction, via file transfer systems, when requested from the Lead Contact – unless for commercial application, in which
case a completed Materials Transfer Agreement will be requested.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

D. melanogaster stocks and husbandry

The specific D. melanogaster stocks used, generated and/or tested in this study are listed in Key Resources Table. To generate Plk4-
mNeonGreen and Asl-mKate2 constructs: 1) NheI restriction enzyme sites were introduced into an mCherry C-terminal Gateway vec-
tor (Basto et al., 2008), using the Quikchange II XL mutagenesis kit (Agilent Technologies). 2) The mCherry tag was replaced with
either mNeonGreen (Shaner et al., 2013) (Allele Biotechnology) or mKate2 (Shcherbo et al., 2009) tags by homologous recombination
via In-fusion Cloning (TaKaRa). 3) NheI restriction enzyme sites were removed via site-directed mutagenesis, using the Quikchange II
XL mutagenesis kit (Agilent Technologies). These vectors were recombined via Gateway technology to pDONR-Zeo vectors (Thermo
Fisher Scientific) where the genetic regions of either Plk4 (Aydogan et al., 2018) or asl (Novak et al., 2014) were previously cloned from
2 kb upstream of the start codon up to (but excluding) the stop codon. Similary, to generate an Asl construct without any fluorescent
tag, the endogeneous Asl stop codon was introduced to the Asl pDONR by site-directed mutagenesis, using the Quikchange II XL
mutagenesis kit.
To generate the Asl-13A-mKate2 construct, 13 point mutations (Figure S4C) were introduced into the endogenous Asl-mKate2 P-
element transformation vector (this study) through NEB Q5 Site Directed Mutagenesis (NEB #E0554S) in four sequential mutagenesis
steps. To generate the endogenous Asl-13A construct without any fluorescent tag, the mKate2 coding sequence was removed, and a
stop codon was introduced immediately following the Asl coding sequence in the eAsl-13A-mKate2 vector (described above) using
NEB Q5 Site Directed Mutagenesis.
Primer sequences used to generate these constructs are listed in Table S1. Transgenic lines were generated using standard P-
element mediated transformation by the Fly Facility in the Department of Genetics, University of Cambridge (Cambridge, England,
UK) or BestGene Inc. (USA). Flies were maintained at 18 C or 25 C on Drosophila culture medium (0.77% agar, 6.9% maize, 0.8%
soya, 1.4% yeast, 6.9% malt, 1.9% molasses, 0.5% propionic acid, 0.03% ortho-phosphoric acid, and 0.3% nipagin) in vials or
bottles.

METHOD DETAILS

Hatching experiments
To measure embryo hatching rates, 0-3 h embryos were collected and aged for 24 h, and the % of embryos that hatched out of their
chorion was calculated.

Synthesis of double-stranded RNA

Double-stranded RNAs (dsRNAs) against cyclins A, B and B3 were synthesized essentially as described previously (McCleland and
O’Farrell, 2008). Primer sequences used for gene amplification are listed in Table S1. The resulting RNA was precipitated with 8 mL of
3M Na-Acetate and 220 mL of 100% ethanol before washing with 70% cold ethanol. The RNA pellets were air-dried and resuspended
in 30 mL of RNase-free diethylpyrocarbonate-treated water (Thermo Fisher Scientific). To generate double-stranded molecules,
RNAs were placed in a 67.5 C water bath for 30 min, and allowed to cool to room temperature over 90 min. Unincorporated
UTPs were removed using CHROMA SPIN-100-DEPC-H2O columns (Clontech) according to the manufacturer’s instructions. To
confirm the synthesis of the correct RNA product, 3 mL of the final reaction was subjected to electrophoresis on a 1.5% agarose
gel using 2xRNA loading buffer (Thermo Fisher Scientific). A 1:1 mix of RNA and loading buffer was heated to 65 C for 5 min and
then placed on ice to denature any secondary structure of RNA.

Embryo collections and dsRNA injections

For embryo collections, 25% cranberry-raspberry juice plates (2% sucrose and 1.8% agar with a drop of yeast suspension) were
used. Embryos for imaging experiments were collected for 1h at 25 C, and aged at 25 C for
45–60 min. Embryos were dechorio-
nated by hand, mounted on a strip of glue on a 35-mm glass-bottom Petri dish with 14 mm micro-well (MatTek), and were left to
desiccate for 1 min at 25 C. After desiccation, the embryos were covered with Voltalef grade H10S oil (Arkema). Embryos for dsRNA
injection experiments were treated in the same way except that the desiccation period was increased to 5-6 min. Embryos were in-
jected with dsRNA at a needle concentration of 0.6–0.8 mg/ml.

Immunoblotting
Immunoblotting was performed as described previously (Aydogan et al., 2018). Primary antibodies used in this study are as follows:
mouse anti-GFP (Roche; RRID: AB_390913) and mouse anti-Actin (Sigma; RRID: AB_476730). Both the antibodies were used at
1:500 dilution in blocking solution (Aydogan et al., 2018). For all blots, 10, 20 or 30 staged early embryos were boiled in sample buffer
and loaded in each lane. The incubation period for primary antibodies was 1 h (or overnight at 4 C). Membranes were quickly washed
3x in TBST (TBS and 0.1% Tween 20) and then incubated with HRPO-linked anti-mouse IgG (both GE Healthcare) diluted 1:3,000 in

Cell 181, 1566–1581.e1–e14, June 25, 2020 e4

 ll
Article

blocking solution for 45 min. Membranes were washed 3x15min in TBST and then incubated in SuperSignal West Femto Maximum
Sensitivity Substrate (Thermo Fisher Scientific). Membranes were exposed to film using exposure times that ranged from < 1 to 60s.

Image acquisition, processing, and analysis

Spinning disk confocal microscopy
Living embryos were imaged at room temperature using a system equipped with an EM-CCD Andor iXon+ camera on a Nikon Eclipse
TE200-E microscope using a Plan-Apochromat 60x/1.42-NA oil DIC lens, controlled with Andor IQ2 software. Confocal sections of 17
slices at 0.5mm intervals were collected every 30 s. A 488nm laser was used to excite mNeonGreen and GFP, and a 568nm laser was
used to excite mCherry and mKate2. Emission discrimination filters were applied when mNeonGreen and mCherry were imaged
together.
Post-acquisition image processing was performed using Fiji (National Institutes of Health). Maximum-intensity projections of the
images were first bleach-corrected with Fiji’s exponential fit algorithm, and background was subtracted using the subtract back-
ground tool with a rolling ball radius of 10 pixels. Plk4-NG, Sas-6-mCherry or -GFP, and Asl-mCherry or -GFP were tracked using
the Fiji plug-in TrackMate (Tinevez et al., 2017) with a track spot diameter size of 1.1 mm. When Plk4-NG was continuously monitored
over cycles 11-13, the maximum intensity projections were limited to ± 5 slices from the central plane of the nuclei, as the nuclei and
centrosomes progressively get closer to the embryo cortex at successive cycles. This processing more accurately compares the
dynamics of centriolar Plk4-NG at successive cycles by avoiding fluctuations due to the varying depths of the centrioles in the em-
bryo. The regressions for the centriole growth curves (Sas-6-GFP or -mCherry) were calculated in Prism 7 (GraphPad Software), as
described previously (Aydogan et al., 2018). The regressions for the Plk4 oscillation curves (Plk4-NG) were calculated using the
nonlinear regression (curve fit) function in Prism 7. Discrete Plk4 oscillation curves in S-phase were initially fitted against four different
functions to assess the most suitable regression model: 1) Lorentzian, 2) Gaussian, 3) Increase – Constant – Decrease, and 4)
Increase – Decrease. Among these models, Lorentzian best fit the data (Figure S1D), so all the discrete Plk4 oscillation curves in
S-phase were regressed using this function. The Lorentzian and Gaussian functions are described in Prism 7, while the latter two
functions are in-house algorithms (Alvarez-Rodrigo et al., 2019).
In order to plot the dynamics of Plk4-NG and Sas-6-mCherry together (Figures 2, S3A, and S3B), the highest mean fluorescence
signal for each tag was normalized to 1 and was accordingly scaled across cycles 11-13 (the scaling factor for Plk4-NG was calcu-
lated from the data shown in Figures 1B and 1C). Note that the amplitude of the Plk4 oscillation does not appear to decrease signif-
icantly between nuclear cycles 11-13 in the data shown in Figure 2A—in contrast to the Plk4 oscillations shown in Figure 1B. This is
not because of the scaling procedure applied to the data shown in Figure 2A (described above), but rather because embryos that
failed to grow their centrioles were excluded from the analysis shown in Figure 2A. The amplitude of the Plk4 oscillations was lower
in these embryos (Figure 2C), so embryos with low amplitude Plk4 oscillations were effectively excluded from the analysis shown in
Figure 2A. Almost all of these excluded embryos were at nuclear cycle 13, so the ‘‘average’’ oscillation at nuclear cycle 13 is in reality
an average of only those embryos that had a relatively high amplitude Plk4 oscillation.
In all the imaging experiments, the beginning of S-phase was taken as the time at which the old and new mother centrioles were first
detected to separate from each other (termed ‘‘centrosome separation’’ or ‘‘CS’’). Entry into mitosis was taken as the time of nuclear
envelope breakdown (NEB), which could be determined in our movies by adjusting the contrast to visualize when the cytosolic pool of
the fluorescent protein was first observed to enter into the nucleus.

Analysis of centriole ‘‘fertility’’ in embryos injected with dsRNA against cyclin A-B-B3
In experiments where we depleted embryos of mitotic cyclins during early rounds of nuclear division, we observed qualitatively that
‘‘fertile’’ centrioles exhibited distinct Plk4-NG fluorescence peaks that often appeared to correlate with centriole duplication events,
while ‘‘sterile’’ centrioles exhibited no obvious peaks (Figure 5B). To test if we could more quantitatively distinguish between fertile
and sterile centrioles, we computationally analyzed all 81 centrioles that we could track throughout the observation period in 3
different embryos. We first assessed the average signal-to-noise ratio (SNR) of Plk4-NG fluorescence of each centriole over the entire
observation period and found that fertile centrioles exhibited a significantly higher SNR than sterile centrioles—assessed using a t
test assuming equal variance (Figure S8B). The distribution of SNR within sterile and fertile centriole signals was unimodal and sym-
metrically distributed (Figure S8C), so we attempted to classify centrioles in an unbiased way by thresholding the SNR. Based on the
bimodality of the SNR, an automatic threshold was determined from the data using Otsu thresholding (red dashed line Figure S8C);
the classification performance was summarized in a visual confusion matrix, which shows the proportion of correctly and falsely clas-
sified signals (Figure S8D). This unbiased computational method successfully classified
74% of the fertile centrioles and
71% of
the sterile centrioles.
Peak Calling
We next tested whether computationally identified peaks in the Plk4-NG signal were correlated with centriole duplication events.
Plk4-NG peaks were only called on signals whose fluctuation (as measured by signal-to-noise ratio, SNR) was greater than a certain
defined threshold (0.1, see below). A peak was defined only as a local maximum in intensity. To call a peak, the Plk4-NG signal in-
tensity was compared to the signal intensity at neighboring times. Here an unbiased distance = 1 was set, that is, an intensity at time t
is a peak only if the intensity is higher than those at both t  1 and t + 1. To filter noise detections, a threshold of 0.1 was placed on the
peak prominence. Peak prominence measures the extent to which a detected peak stands out from its surrounding – it is defined as

e5 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

the vertical distance between the peak and its lowest contour line (Scipy’s find_peaks function) (Jones et al., 2001). The choice of 0.1
as a threshold was guided by comparing a peaks predictive power given no cut-off with the (ground truth) duplication time. This anal-
ysis indicated that the optimal peak prominence cut-off—i.e., the point at which the power of the peaks to predict duplication events
(see below) sharply drops off—was 0.12 (green dot, Figure 5E). The observed steep drop-off in predictive power below this threshold
supports the view that that there is likely to be a minimal amount of centriolar Plk4-NG that is necessary to trigger duplication under
these conditions. Moreover, this unbiased computational approach identified
4x as many peaks in the fertile centrioles when
compared to the sterile centrioles.
To determine whether the filtered Plk4-NG peaks were predictive of centriole duplication, we determined all the peaks above the
0.1 threshold for the fertile centriole signals and assessed whether these peaks could be used to ‘‘retrieve’’ the real or relevant time
points for centriole duplication. The performance of such retrieval can be evaluated using ‘‘precision’’ (the number of relevant re-
trievals among all retrieved instances—in this case the number of Plk4 peaks associated with a centriole duplication event divided
by the total number of Plk4 peaks) and ‘‘recall’’ (the number of relevant instances retrieved of the total relevant instances—in this case
the number of Plk4 peaks associated with a centriole duplication event divided by the total number of centriole duplication events), as
defined below.
Number ofðRelevant & RetrievedÞ
Precision = ;
Number of ðRetrievedÞ

Number ofðRelevant & RetrievedÞ

Recall =
Number ofðRelevantÞ

The evaluation of such a system naturally depends on the cut-off to call a positive match between the Plk4-NG signal peak and its
corresponding centriole duplication time. Too small a cut-off (e.g., 0 minutes) is unrealistic: no system can predict time perfectly;
while too large a cut-off (e.g., 15 min) is too lenient and non-specific. Figure 6C plots the precision evaluated over all centrioles
for different temporal cut-offs attempting to uniquely match Plk4-NG peaks to the nearest duplication time within a given time win-
dow. The elbow point (red dashed line) at 5 min was selected as an appropriate cut-off with a precision of
80%. (Note that the recall
was not plotted in this graph, but it exhibits a similar behavior to the precision: the Number of (Relevant) = 52, while the Number of
(Retrieved) = 49). This temporal cut-off can also be interpreted as an estimate of the temporal accuracy to which Plk4-NG peak time
associates with centriole duplication time. For comparison, we also derived the mean temporal separation distance of peaks and
duplication events, if the same number of experimental centriole duplication times were randomly distributed over the same time in-
terval for each embryo. 1000 simulations were run per embryo to produce a distribution. Across all embryos, an average temporal
separation distance (for randomly distributed duplication times) was 10.5 minutes (data not shown), twice as long as the chosen
5 min cut-off, thus the association is not coincidental.
In addition, we assessed the precision and recall performance over different possible threshold values (based on peak prominence)
used to call a Plk4 ‘peak’. To do this, we computed the precision-recall curve. All detected Plk4 peaks (Black dots; Figure 5E) were
ranked according to their peak prominence from high to low and were assigned uniquely to a duplication event according to a 5 min
time window for determining a positive match. Peaks that could not be uniquely assigned in such a manner were regarded as ‘neg-
atives’. The graph then plots the precision, recall values if the threshold for calling a peak were set as the peak prominence value of
each peak in descending order. Beyond the detected peak associated with a peak prominence of 0.12 (i.e., points right of this point),
the precision drops sharply. At this threshold, precision and recall are jointly optimized. This suggests that a minimum level of Plk4-
NG peak fluorescence intensity is required to predict duplication. The ability of Plk4 peaks to predict duplication across all peak
prominences (over the selected time window of 5 min) is quantified by the integrated area under the curve or average precision
(AP5min). If there were no overall correlation between a Plk4 peak and a duplication event, AP5min would be 51.5% (given by # dupli-
cations / (# duplications + # peaks)); the score of
75% indicates a strong overall correlation (Figure 5E).
Finally, the correlation between the Plk4-NG peaks and times of centriole duplication was examined (Figure 5F), which provided an
alternative accuracy test. Plk4-NG peaks were uniquely matched to the nearest centriole division times without using a temporal cut-
off over individual centrioles from three independent embryos. Pearson correlation r, R2 and P values are reported as goodness of fit.
The fitted regression line, y = 0:87x + 3:69. Together, these unbiased computational analyses indicate that the Plk4 oscillations at
individual centrioles are highly correlated with the time at which these centrioles duplicate.

Spatiotemporal heatmap of centriole duplications

To visually assess whether there is bias as to where and when centriole duplications happen, the (x,y) position of all duplicating cen-
trioles were overlaid on the (x,y) positions of all centrioles (duplicating and non-duplicating at all times; black dots) (Figure S8E). To
grade the temporal sequence of centriole separations, duplication points were colored blue to red. To enable comparison across
embryos relative to the same geometric reference, the embryonic width and anterior-posterior were set as x- and y-axes, respec-
tively, by applying a principal component analysis on the extracted (x,y) positions of all centrioles.

Cell 181, 1566–1581.e1–e14, June 25, 2020 e6

 ll
Article

Spatial clustering assessment of centriole duplications

To statistically measure whether centriole duplications are enriched at particular spatial regions over time, Ripley K statistics was
used (Ripley, 1976). In the field of spatial statistics, given a set of (x,y) points, the Ripley K statistic detects deviations from spatial
homogeneity at different distances between points or spatial scales (e.g., such analyses are heavily used in geophysics to map
out the spatial distribution of natural disasters and in crime statistics for detecting high incidence areas). For a dataset of n points,
the Ripley K statistic, K Ripley ðdÞ is the mean spatial occurrence of two points, point i and j having a separation distance, dij less than
the search distance threshold of d:
1 XIðdij < dÞ
K Ripley ðdÞ =
l isj n

l is the average density of points (estimated as n=A where the number of total points, n is divided by the area of the region containing
all points, A), and I is the indicator counting function (I = 1 if its operand is true, 0 otherwise). Thus, if points are homogeneously spread
in 2D, the Ripley K statistic should vary quadratically as pd2 . The basic test assumes (x,y) points occur at any spatial position contin-
uously in the image. However, centrioles only duplicate in certain discrete positions within fly embryos. Thus, to examine evidence of
spatial clustering from the natural distribution of centrioles, we assessed difference in the Ripley K statistics computed from the (x,y)
positions of all duplicating centrioles and the (x,y) positions of all centrioles accumulated over time.

Plk4-NG smooth curve fitting and parameter extraction

To enable accurate extraction of signal parameters from the Plk4-NG oscillations, yðtÞ the signals were robustly fit to a smooth 1D
function. Here a mixture of N Gaussian and linear trend was used with the following functional form to enable local modeling of the
peak and troughs of signals:

X
N ðtmi Þ2

s2i
yðtÞ = ðA + BtÞ + Ci e
i=1

where t is time, A; B are the constant and slope of a linear trend line, and Ci ; mi ; si are the amplitude, time and temporal duration of the
ith Gaussian, respectively. This function was fit in two steps. In the first step A; B was fit by applying least-squares linear regression on
the baseline trend line that is extracted from asymmetric baseline smoothing (Eilers and Boelens, 2005). In the second step, peak and
trough positions were first detected on the de-trended signal, y0 ðtÞ after subtraction of the fitted trend line in the first step from the
original signal, yðtÞ, so as to determine the number N of Gaussians to fit. The mixture of Gaussians was then fit by iterative non-linear
regression using a robust Cauchy loss function. From the fit signal, yðtÞfitted , peak and trough positions were re-detected, and the
following signal parameters were extracted (Figure 6):

d Acceleration rate ðDFluo: A:U: =secsÞ: The maximum rate of increase in fluorescence between successive time points during
a trough to peak oscillation phase.
d Deceleration rate ðDFluo: A:U: =secsÞ: The maximum rate of decrease in fluorescence between successive time points during
a peak to trough oscillation phase.
d Oscillation peak time ðsÞ: The time point corresponding to the maximum fluorescence.
d Oscillation trough time ðsÞ into mitosis: The end point of a trough after which the fluorescence begins to accelerate upward.

The method described here was also used to determine the period of Plk4 oscillations (measuring peak-to-peak time; see Results
and Discussion) both in normal embryos and in embryos where dsRNA was injected against cyclins A, B and B3 to halt the progres-
sion of cell cycle.

3D-Structured Illumination Microscopy (3D-SIM)

Living embryos were imaged at room temperature using a DeltaVision OMX V3 Blaze microscope (GE Healthcare). The system was
equipped with a 60x/1.42-NA oil UPlanSApo objective (Olympus Corp.), 488nm and 593nm diode lasers, and Edge 5.5 sCMOS cam-
eras (PCO). Spherical aberration was reduced by matching the refractive index of the immersion oil (1.514) to that of the embryos. 3D-
SIM image stacks consisting of six slices at 0.125mm intervals were acquired in five phases and from three angles per slice. The raw
acquisition was reconstructed using softWoRx 6.1 (GE Healthcare) with a Wiener filter setting of 0.006 and channel-specific optical
transfer functions (OTFs). Filters used for the green and red channels were a 540/80 center band pass filter and a 605-long pass filter,
respectively. For two-color 3D-SIM, images from green and red channels were registered with the alignment coordination information
obtained from the calibrations using 0.2mm-diameter TetraSpeck beads (Thermo Fisher Scientific) in the OMX Editor software. The
SIMCheck plug-in in ImageJ (National Institutes of Health) was used to assess the quality of the SIM reconstructions (Ball et al., 2015);
only images that passed this test were used.

e7 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

Mathematical modeling and its experimental validation

Model 1: A simple mathematical model of discrete S-phase Plk4 oscillations
Figures 3A and 3B, specify a regulatory network where Plk4 binds to an Asl receptor with high affinity; this activates Plk4, allowing it to
phosphorylate itself and Asl multiple times. After a certain number of phosphorylations, Asl switches to a new state that binds Plk4
with low affinity. As a result, Plk4 unbinds, leaving Asl in a phosphorylated, low affinity state. We suspect that Asl is normally dephos-
phorylated by a phosphatase in M-phase, which ‘‘resets’’ it to a high-affinity state in preparation for the next oscillation in S-phase
(this additional step is considered in Model 2 below). In this first model, the gradual conversion of centriolar Asl to a low affinity binding
state forms a time-delayed negative feedback loop wherein Asl effectively activates Plk4 to gradually promote its own inhibition. After
making assumptions about the chemical kinetics of the system and imposing suitable initial conditions, the behavior of this regulatory
network can be simulated by mapping it onto a set of coupled ordinary differential equations.
In the model, it is assumed that the diffusion of Plk4 is sufficiently fast such that it remains well-mixed in the cytoplasm and that
centrioles are large macromolecular structures; this implies that Plk4 and Asl receptors on the centriole follow mass-action kinetics.
Let ½P and ½A0  denote the concentrations of unbound cytosolic Plk4 and unbound centriolar Asl, (per unit volume and area, respec-
tively). Plk4 binds to Asl with the fixed rate constant k, and the rate constant of the reverse reaction is sufficiently small that any
unbinding is ignored. Once Plk4 is bound to Asl, it can only unbind after it has phosphorylated Asl a certain number of times (N).
We denote by ½A0  the concentration of Asl receptor that has bound Plk4, but has not yet been phosphorylated, and by ½Ai  the con-
centration of Asl receptors that have been phosphorylated i times. Throughout this paper, we use an asterisk superscript to denote an
Asl receptor that has bound Plk4 and a numerical subscript to denote the phosphorylation state of the receptor. Each phosphory-
lation of Asl by Asl-bound Plk4 has rate constant k 2 and, following N phosphorylations, the Asl is switched to a state that binds
Plk4 with very low affinity ½AN . Once Asl has been converted to this low affinity state, Plk4 unbinds at rate k3 . The rate constant
of the reaction where Plk4 binds to Asl receptor in this low affinity state is assumed to be sufficiently small that this reaction is ignored
in the model.
Intuitively, k scales the affinity with which Plk4 binds to an Asl receptor. By mass-action kinetics, the rate of this reaction is given by
k½P½A0 . It is assumed in the model that Plk4 is abundant enough in the cytoplasm that its concentration does not decrease over the
few minutes of a single S-phase cycle; this assumption means that ½P remains constant over that time (see below for further discus-
sion of this assumption). Therefore, the number of parameters in the model is reduced by introducing a new rate constant k1 = k½P.
Using the assumptions above, the regulatory network in Figures 3A and 3B is simulated using the following set of ordinary differ-
ential equations which are solved over the time domain 0 % t % S, where S is the length of S-phase:
 
d A0  
= k1 ½A0   k2 A0 (1)
dt

 
d A1    
= k2 A0  k2 A1 (2)
dt

 
d AN    
= k2 AN1  k3 AN (3)
dt

d½A0 
=  k1 ½A0  (4)
dt

Appropriate initial conditions at t = 0 are,

       
b ; ½A0  = A
A0 = A b0 ; A = A = / = A = 0: (5)
0 1 2 N


In Equation 5, the positive constant A b is the initial amount of Plk4-bound Asl at the centriole at the start of each S-phase, which is
0
determined experimentally for each cell cycle using the techniques described in the Image acquisition, processing and analysis sec-
tion of STAR Methods. The constant A b is the initial amount of unbound Asl, so the total amount of Asl in the system is given by Atot =

b b
A0 + A0.
Since the model specified by Equations 1–5 is a system of linear differential equations with constant coefficients, it has an analytical
solution that can be expressed as a sum of exponentials. Values for the parameters A b0 , k1 , k2 , and k3 can then be determined by fitting
  
the curve ½A0 ðtÞ + ½A1 ðtÞ + / + ½AN ðtÞ to the experimentally measured data for the amount of Asl-bound Plk4 (i.e., the Plk4 that is
recruited to the centriole) over time. Fitting was done using a trust-region algorithm to optimize a nonlinear least-squares penalty
function.

Cell 181, 1566–1581.e1–e14, June 25, 2020 e8

 ll
Article

Parameter fitting
The fitting was constrained to enforce that all parameters were positive, and k1 and k2 were taken to be less than 1. Each cycle was
fitted individually using the discrete Plk4-NG oscillation data from S-phase of cycles 11, 12 and 13 (Figure 3C). Parameter values are
shown in Data S1 (First, second and third charts). As explained below, the solutions to this model are very insensitive to variations in k3
(see Data S1, the Monte Carlo analysis), so in the solutions presented here k3 was kept at a constant value of 0.06906, which was the
best-fit parameter value for cycle 12 (Data S1; first, second and third charts).
Picking the value of N, the number of phosphorylation sites:
In the model, we assumed that Asl had to be phosphorylated by Plk4 N times before it switched to a low-affinity state—indicated by
variables ½A0 ;/; ½AN . We tested the effect of the number of phosphorylation sites on the model solution by using N = 1, 4, 9, 14, or 16.
The best fit curves for ½A0 ðtÞ + ½A1 ðtÞ + / + ½AN ðtÞ suggested that the model is a good fit for the data for any value of N > 4 (N = 1 (R2 =
0.9152), N = 4 (R2 = 0.9886), N = 9 (R2 = 0.9996), N = 14 (R2 = 0.9962) or N = 16 (R2 = 0.9931)). So, we use N = 9 corresponding to 10
phosphorylation sites of Asl (1 unbound, 9 bound with various stages of phosphorylation) in all subsequent modeling, although we
note that any value above 4 works essentially equally well.
Data S1, first chart, shows that the trust-region algorithm finds a very good fit (R2 > 0.99) for the model to the experimental data
(Figure 3C), but this provides little information about uniqueness of the fit as there may be other subsets of the parameter space that
also provide a good fit to the data. To see if any such regions could be detected, the parameter space was further explored by using a
Metropolis-Hastings Markov chain Monte Carlo algorithm. Four Markov chains were started at the positions in the parameter space
specified in Data S1 (fourth chart). The Monte Carlo analysis in Data S1 shows the six two-dimensional traces of the four-dimensional
parameter space. For clarity, only points that provided a good fit to the cycle 12 data (R2 > 0.95) are shown.
The results in Data S1 (the Monte Carlo analysis) reveal how sensitive the model is to changes in each parameter value. The model
only fit the data well for a relatively narrow range of values for k1 , k2 , and A b0 . In contrast, the fits are mostly insensitive to k3 . This is
likely, because the rate of phosphorylating Asl at multiple sites is relatively slow compared to the rate at which Plk4 is subsequently
released from the multiply phosphorylated Asl—so the rate of release is not limiting. The Monte Carlo simulations also reveal corre-
lations between k1 and k2 , k1 and A b0 , and k2 and A b0 . For example, these results show that if A b0 (the initial amount of unbound Asl
receptor at the start of S-phase) is reduced, the model can still fit the data well if k2 is decreased and k1 is increased. While these
results suggest that there is a single, continuous region of the parameter space that provides a good fit to the data, it is still possible
that there are other such regions that the Markov chains in Data S1 (the Monte Carlo analysis) did not explore. However, the results in
Data S1 (the Monte Carlo analysis, panel B) show that the points, which are identified at the center of the parameter region, provide
the best fit to the data. This suggests the nonlinear least-squares minima found by the trust-region fitting is insensitive to the
initial seed.
Interestingly, the best-fit parameters for cycles 11–13 showed that the biggest difference between the parameters of the Plk4 os-
cillations at each cycle is in k1—the rate at which Plk4 binds to Asl (which is dependent on the cytosolic concentration of Plk4).
Although our model assumes that the cytosolic concentration of Plk4 remains constant during the S-phase period within each cycle,
if the phosphorylated Plk4 molecules that are released from the Asl receptor are ultimately degraded—and there is good evidence
that Asl activates Plk4 to promote Plk4 degradation (Klebba et al., 2015)—there could be a wave of phosphorylated-Plk4 degradation
in the cytoplasm toward the end of S-phase. If so, the cytosolic levels of Plk4 would get successively lower at the start of each suc-
cessive cycle, as our PeCoS analysis indicates is the case (Figure 3E).
The effects of reducing the genetic dose of Plk4 by half (Plk4-NG1/2 embryos—see Results and Discussion for details) were
analyzed. Our PeCoS analysis indicated that there was a
45% drop in the cytosolic levels of the Plk4-NG protein in the Plk4-
NG1/2 embryos (Figure S6C). When the model was fit to Plk4-NG1/2 oscillation, the best-fit (R2 = 0.996) parameter had a k1 value
that was
39% of the control value (Data S1, second chart), so in reasonable agreement with the 45% drop in cytosolic Plk4 levels
we measured experimentally. These parameter values also suggested that the total amount of centriolar Asl ðAtot Þshould remain rela-
tively unchanged between the Plk4-NG and Plk4-NG1/2 conditions (Data S1, second chart). Centriolar Asl levels were analyzed in em-
bryos expressing Asl-mCherry in either WT versus Plk41/2 conditions, and our findings showed that this was indeed the case
(Figure S7A).
Next, the effects of reducing the genetic dose of asl by half (asl1/2 embryos—see Results and Discussion for details) were analyzed.
Interestingly, the best-fit parameter values (R2 = 0.999) predicted that the total centriolar Asl levels (Atot) would be reduced by only

28% in asl1/2 embryos (Data S1, third chart). This value was therefore directly measured in embryos expressing either one or two
copies of Asl-GFP (under the control of its own promoter in an asl mutant background). Encouragingly our findings showed that
reducing the genetic dose of Asl-GFP by half led to a reduction of only
30% in centriolar Asl-GFP levels (Figure S7B). Moreover,
the parameter values suggested that the concentration of Plk4 (incorporated in the k1 term) should not vary significantly between
WT and asl1/2 conditions (Data S1, third chart). We confirmed this prediction using western blotting for Plk4-GFP and PeCoS for
Plk4-NG (both transgenically expressed from their own promoters in a Plk4 mutant background) in control and asl1/2 embryos (Fig-
ures S7C and S7D).
Taken together, these analyses indicate that our model can robustly describe the Plk4-NG oscillations under normal conditions
(Figure 3C) and when the levels of either Plk4 or Asl are perturbed experimentally (Figures 4A and 4B). Moreover, the model makes
several plausible predictions about the relative levels of these proteins in the perturbed conditions that are close to the levels that we
measured experimentally.

e9 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

Finally, the best-fit value of k2 (reflecting the kinase activity of individual Plk4 molecules) decreased slightly between cycles 11 to 12
and decreased more significantly between cycles 12 to 13 (by
9% and
37%, respectively); k2 also decreased when levels of Plk4
were genetically reduced in Plk4-NG1/2 embryos (by
25%)—but not when Asl levels were genetically reduced in asl1/2 embryos
(Data S1; first, second and third charts). The molecular basis for this inferred decrease in kinase activity remains unknown, but we
believe it is biologically plausible. We previously suggested that centriolar Plk4 was likely to integrate several inputs at the start of
each cycle (from, for example, cell cycle regulators, or its activator Ana2/STIL) and adjust its kinase activity in response to the length-
ening of S-phase during successive nuclear cycles (Aydogan et al., 2018). Moreover, our finding that Wee1 kinase, an important cell
cycle regulator, can influence the Plk4 oscillation parameters in S-phase strongly supports this hypothesis (Figure 6).

Model 2: Generating robust Plk4 oscillations entrained by the CCO

The network described above was further extended to test the possibility of generating robust oscillations in centriolar Plk4 levels. To
do so, we allow the Asl receptors to be dephosphorylated by a phosphatase at rate k 4 (Figure S4A). Subject to the constraint that
there is no Plk4 bound to the Asl receptors initially (i.e., at the start of cycle 1), we model multiple cycles by allowing the phosphatase
to be active only during mitosis, so that k 4 is nonvanishing only in this period. Therefore, the system reads:
 
d A0  
= k1 ½A0   k2 A0 (6)
dt

 
d A1      
= k2 A0  k2 A1  k4 A1 (7)
dt

 
d AN      
= k2 AN1  k3 AN  k4 AN (8)
dt

d½A0 
= k4 ½A1   k1 ½A0  (9)
dt

d½A1 
= k4 ½A2   k4 ½A1  (10)
dt

d½AN   
= k3 AN  k4 ½AN  (11)
dt

subject to the initial conditions,

   
½A0  = 1; ½A1  = / = ½AN  = 0; A0 = / = AN = 0: (12)

It is further assumed that the embryo is in mitosis for 30% of the total time in each nuclear cycle and all cycle times are kept constant.
Hence, k4 = 0 for 0 < t mod T < 0:7T and a positive constant for 0:7T < t mod T < T, where T is the period of the cell cycle. Values for
the rate constants are determined by fitting the exact analytical solution in the S-phase of cycle 12 to the Lorentzian regression of the
experimental data (R2 = 0.9870). In the first instance, we assumed that the cytosolic concentration of Plk4 remains constant over the
nuclear cycles (see below). We plot the exact solution for the percentage of Asl-bound Plk4 molecules for a total of 14 nuclear cycles,
as is the case in fly embryos. This minimal model was sufficient to generate sustained oscillations in centriolar Plk4 levels (Figure S4B;
k4 = 0.0708).
As an alternative to the assumption that the cytosolic concentration of Plk4 is constant over the cycles, we also considered the
case where the total number of Plk4 molecules in the embryo is kept constant (so any Plk4 degradation is balanced by new synthesis).
In this model, as the number of centrioles, NC ðtÞ, increases at successive cycles so the number of available Plk4 molecules in the
cytosol initially decreases during S-phase (as Plk4 binds to centriolar Asl receptors), and then increases (as Plk4 unbinds from Asl
receptors). We estimate that there are NP = 105 molecules of Plk4 in an embryo of 0.01mm3 volume (Markow et al., 2009); a concen-
tration of
10nM, in agreement with that measured in human cells (Yamamoto and Kitagawa, 2019), but potentially higher than we
infer from our observation that Plk4 levels are too low to be measured by FCS (as we cannot infer absolute protein concentration from
our PeCoS experiments). To simulate the effect of centriole duplications, we double the number of centrioles each cycle and assume
that, at each centriole duplication, the bound Plk4 (attached to Asl receptors) is equally split between the mother and separating
daughter. To consider the Cdk/Cyclin trigger wave that sweeps through embryos (Deneke et al., 2016), it is assumed that the

Cell 181, 1566–1581.e1–e14, June 25, 2020 e10

 ll
Article

duplicated centrioles separate nearly synchronously over the last 10% of the time-window in each cycle. Based on our 3D-SIM mi-
croscopy data, we assume that each centriole has
30 Asl receptors, as we essentially only need to consider Plk4 binding to Asl at
the site of centriole assembly (the model works well for between
20-80 receptors), and that there is a single centriole in cycle 1. With
these modifications to the model, the system reads
 
d A0   1 dNC
= k1 ½P½A0   k2 A0  A0 (13)
dt NC dt

 
d A1       1 dNC
= k2 A0  k2 A1  k4 A1  A1 (14)
dt NC dt

 
d AN       1 dNC
= k2 AN1  k3 AN  k4 AN  AN (15)
dt NC dt

d½A0  1  dNC
= k4 ½A1   k1 ½A0  + A (16)
dt NC 0 dt

d½A1  1  dNC
= k4 ½A2   k4 ½A1  + A (17)
dt NC 1 dt

d½AN    1  dNC
= k3 AN  k4 ½AN  + A (18)
dt NC N dt

NR NC ðtÞ XN
½P = 1  An (19)
NP n=0

subject to the initial conditions (12).

We plot the solution of the model for the percentage of Asl-bound Plk4 molecules, as well as the percentage of Plk4 molecules that
remain in the cytoplasm, over 14 nuclear cycles (Figure S4F; R2 = 0.9871 and k4 = 0.0612).
We observe a small spike as the centrioles begin to separate at the end of each mitosis (Figure S4F). The spike is small, due to the
slight asynchrony in centriole separations; if the centrioles were to separate all simultaneously, the concentration would instantly
halve at this point, since the number of receptors would all double. Interestingly, these small spikes are consistent with our exper-
imental observations (Figures 1A and S2A). Moreover, we emphasize that, for the first few nuclear cycles, almost all of the Plk4 re-
mains in the cytoplasm since there are only a few centrioles. In the later cycles, however, the amount of Plk4 sequestered by the Asl
receptors increases exponentially, as the number of centrioles increase by a factor of 2 in each cycle. Therefore, the rate at which the
Asl receptors are able to recruit Plk4 from the cytoplasm decreases, resulting in a reduction in the amplitude of the Plk4 oscillation
(Figure S4F). This feature of the model is also consistent with our experimental observations (Figures 1B and 1C).

Model 3: Stochastic duplications

Finally, we have also developed a discrete mathematical model, which is analogous to Model 2 described previously, in order to
consider the possibility of stochastic centriole duplication in non-cycling embryos (as in those injected with dsRNA against cyclins
A, B and B3; Figure S8G). As before, we assume that unphosphorylated Asl receptors bind Plk4 with high affinity until they become
fully phosphorylated and Plk4 unbinds, and we initially assume that the Asl receptors can be dephosphorylated during mitosis. We
model this system stochastically by defining the state vector for each receptor to be
 
V = A0 ; /; AN ; A0 ; /; AN (20)

At any given time, precisely one entry of V is equal to unity, corresponding to the state which the receptor is in at that moment, and all
other entries are equal to zero. We allow the receptor to change state over time according to the transition matrix:

e11 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article
2 3
Q1 0 0 / 0 P1 0 / / 0
6 P4 Q4 0 1 0 0 0 1 1 « 7
6 7
6 0 P4 Q4 1 0 « « 1 1 « 7
6 7
6 « 1 1 1 « « « 1 1 « 7
6 7
6 0 / 0 P4 Q4 0 0 1 1 « 7
M= 6
6 0
7 (21)
6 / / 0 0 Q2 P2 0 1 « 7 7
6 « 1 1 « « P4;2 Q2;4 P2;4 0 « 7
6 7
6 « 1 1 « « 0 1 1 1 0 7
6 7
4 « 1 1 0 0 « 1 P4;2 Q2;4 P2;4 5
0 / / 0 P3;4 0 / 0 P4;3 Q3;4

where
ki  
Pi = 1  eki ; Pi;j = 1  eðki + kj Þ ; (22)
ki + kj

Qi = eki ; Qi;j = eðki + kj Þ ; (23)

describe the probabilities of a receptor changing state and remaining in the same state which arise in our model. We may allow the
cytosolic Plk4 concentration to vary in this model by making the substitution k1 / k1 ½P and using (19) to compute ½P (evaluating the
sum over all receptors being simulated). We also assume that, if a receptor is in a Plk4-bound state ðAn Þ during the last 10% of a cycle,
the Plk4 will unbind ðAn / An Þ with 50% probability during that time period in order to simulate mother-daughter separation.
In this model, each Asl receptor behaves as an independent oscillator—alternating between a Plk4-bound form that is being phos-
phorylated, and an unbound form that is being dephosphorylated. In the presence of the CCO, the individual Asl receptors generate
coordinated oscillations because the CCO effectively synchronizes them every cycle by ensuring a coordinated burst of PPTase ac-
tivity during mitosis. This activity is lost in the absence of the CCO, but instead we allow the PPTase to be active at a low, but constant,
level (10% of the mitotic activity in cycling embryos). We plot the Asl-bound Plk4 levels for a total of 10 centrioles (each with 30 re-
ceptors as assumed above; Figure S8H). We observe that the centrioles are initially synchronized, since they all start in an unbound
state, and display a single round of Plk4 binding. However, as time progresses, the Asl receptors lose synchrony, and each centriole
exhibits stochastic, low-amplitude oscillations. Such oscillations may be sufficient to trigger duplications at individual centrioles, as
evident from our experimental observations (Figures 5B–5F and S8B–S8F; Video S4).
All the equations used for mathematical modeling and regressions are available in the following web link: < https://github.com/
RaffLab/centriole_oscillator_model >.

Fluorescence Correlation Spectroscopy (FCS)

FCS setup and measurements
Point FCS measurements were performed on a confocal Zeiss LSM 880 (Argon laser excitation at 488 nm and GaASP detector) with
the Zen Black Software. A C-Apochromat 40x/1.2 W objective and a pinhole setting of 1AU were used. A laser power of 10 mW was
used, and no photobleaching was observed during the measurements. The microscope was kept at 25 C using the Zeiss inbuilt heat-
ing insert P and the heating unit XL. A schematic overview of the methodology used is shown in Figure S5A, and a comparison of the
average autocorrelation curves generated at the start of S-phase of nuclear cycles 11-14 is shown in Figure S5B.
The effective volume of the imaging setup was estimated to be 0.28 fL by averaging the estimate obtained by three independent
methods, as described previously (Rüttinger et al., 2008): 1) Measuring the concentration of a soluble Alexa Fluor 488 NHS Ester
dilution series (100 nM, 10 nM, 1 nM and 0.1 nM); 2) Measuring the diffusion time for Alexa Fluor 488 NHS Ester (same concentrations)
in water at 25 C. The measured diffusion time was then compared to a previously reported diffusion coefficient for the Alexa Fluor 488
NHS Ester (Petrásek and Schwille, 2008); 3) Imaging subresolution beads (FluoSpheres Carboxylate-Modified Microspheres, 0.1 mm)
and determining the effective volume via Gaussian fitting with the line tool and Z-axis profile in ImageJ (Bethasda, USA).
Embryo collections (from mother flies expressing Asl-GFP under the control of its own promoter in an asl mutant background) were
as described above, with the exception of using high precision 35 mm, high Glass Bottom m-dishes (ibidi). Before every measure-
ment, spherical aberrations were adjusted on the correction collar of the objective by maximizing the count-rate per molecule
(CPM). At the beginning of S-phase in each cell cycle (when the old and new mother centrioles were separating), consecutive cyto-
solic measurements were made 6x for 10 s each at the centriolar plane of the embryo. Individual recordings where centrioles moved
through the measurement spot, based on the highly erratic shape of the correlation curve (< 2% of all recordings), were discarded.

Cell 181, 1566–1581.e1–e14, June 25, 2020 e12

 ll
Article

Autocorrelation analysis and post-acquisition curve fitting

The autocorrelation function, G(t), was calculated during each measurement in the Zen Black software using the following equation:
hdIðtÞ$dIðt + t Þi
Gðt Þ = D E
dIðtÞ2

where CD denotes a time average, dIðtÞ describes the intensity fluctuation at the time point t, and t states the lag time of the
autocorrelation.
All 10 s-recordings were then fitted with 8 different 3D diffusion models using the software FoCuS-Point (Waithe et al., 2016) with
the following equation:

XDs  ak !1   1=2

t t
G3D ðt Þ = Ak 1 + 1+
k =1
txyk AR2 txyk

where Ak defines the fraction of a diffusing species for which the sum of all diffusing species equals 1, txy describes the average resi-
dence time of the diffusing species in Veff, a accounts for anomalous subdiffusion within the cytoplasm, and AR is a structural param-
eter that describes the relationship among the x, y and z-axes of the excitation volume.
Dark states of the fluorophore were fitted with the following formula:
XTs Tj
GT ðtÞ = 1 + $et=tTj
j = 11  Tj

where T depicts the triplet population, and tT states the triplet correlation time during which the fluorophore stays in the dark state
(Schönle et al., 2014).
The data was fitted within the boundaries of 4x104 ms and 1.5x103 ms, and the dark states were restricted to 10-300 ms for the
blinking state, and 1-10 ms for the triplet state. The models (Ms) were defined as the following: M1) 1 diffusing species (ds) 0 blinking
states (bs) 0 triplet states (ts); M2) 1 ds 1bs 0ts; M3) 1 ds 0bs 1ts; M4) 1 ds 1bs 1ts; M5) 2 ds 0bs 0ts; M6) 2 ds 1bs 0ts; M7) 2 ds 0bs
1ts; M8) 2 ds 1bs 1ts. In all models, the structural parameter AR and the anomalous subdiffusion parameter a were kept constant at 5
and 0.7, respectively.
In order to avoid over-fitting the data, the most plausible model to describe the autocorrelation functions was selected using the
Bayesian Information Criterion (BIC), which is based on the likelihood function, but introduces a penalty term for the complexity (num-
ber of variables) for the models (Schwarz, 1978). In this study, M4 was the preferred model to describe Asl-GFP diffusion
(Figure S5A(iv)). The concentration was calculated from the FoCuS-point fit data of the preferred model:
1 hNi
hNi = ; conc: =
G0 Veff

where N states the average number of particles within the effective volume Veff, and G0 represents the height of the autocorrelation
function at t = 0.

FCS background corrections

In order to estimate the contribution of the background noise, 22 wild-type embryos were measured with the same laser intensity
(10 mW) and in roughly the same plane and developmental stage as the Asl-GFP embryos. Despite no observable correlated back-
ground, the uncorrelated background contributed
30% of the total photon count rate, presumably due to the low concentration of
cytosolic Asl-GFP and the high autofluorescence of the embryo itself (Figure S5(v)). Background corrections were performed after the
autocorrelation analysis by calculating the correction factor c2 using the following formula (Koppel, 1974):
1 1
=
c2 ð1 + hbi=hf iÞ2

1
hNi =
c2 G0

where hbi denotes the average background and hf i states the average count rate of the sample.

Data restriction
In some FCS measurements a sudden drop in CPM was observed, possibly due to movements within the embryo or the embryo
drifting away from the measurement plane. When this happened, a strong, often unreasonable increase in concentration was

e13 Cell 181, 1566–1581.e1–e14, June 25, 2020

 ll
Article

observed. These outliers were therefore discarded based on a ROUT outlier test (with the aggression factor Q = 1%), which was per-
formed on all 10 s-long concentration measurements (the red data points in Figure S5(vi)). Only the embryos with at least 4x10 s re-
cordings (after discarding outliers and erratically shaped ACFs) were included in the final analysis.

Peak Counting Spectroscopy (PeCoS)

FCS was not sensitive enough to investigate the cytosolic concentration of Plk4-NG, presumably because its concentration was too
low. We therefore developed a new method that we term Peak Counting Spectroscopy (PeCoS) that allows the relative concentration
of low abundance proteins to be measured accurately (Figure S6). PeCoS uses the same set up as the point FCS protocol described
above, but it differs in terms of its data acquisition and analysis. In PeCoS, the intensity peaks, which are generated by a fluorophore
moving through the effective volume, are counted as a proxy for concentration. Due to the low cytosolic concentration of the fluo-
rescently-tagged protein of interest (e.g., Plk4-NG in this study), spherical aberrations could not be corrected within the same embryo
where the measurements were taken. Therefore, embryos that express a bright fluorescent centriolar marker were positioned next to
the experimental embryos on the same imaging dish and these were used for correction collar adjustment (Figure S6A(i)). Experi-
mental recordings were then captured for 180 s (instead of 6x10 s), as the number of particles that pass through the field of view
was usually very low. Before every measurement, the observation region was pre-bleached with the same laser intensity (10 mW)
for 3 s to bleach away any potential immobile fraction.
Instead of autocorrelation analysis, the resulting intensity traces (Figure S6A(iii)) were quantified for their number of peaks, which
originate from a fluorophore moving through the excitation volume and causing a detectable burst of photons. In order to determine
the cut-off threshold, which was used to subtract the background noise, 40 control embryos were measured (Figure S6A(iii)). These
control embryos were from mothers expressing Asl-mKate2 to allow measurements at the centriolar plane and at the right nuclear
cycle stage (beginning of S-phase). ‘‘Mean + n*SD’’ (where n = 1,2,3,.) of all control recordings was subtracted from each control
recording, and the threshold that resulted in an average of less than five peaks (per 180 s control measurement) was subtracted from
all intensity traces (Figure S6A(iv-vi)). This threshold was found to be a good compromise for minimizing the background noise without
discarding too much information. A Python script was written in order to automate this procedure, which is available via < https://
github.com/RaffLab/PeCoS >. The subtraction of ‘‘Mean + 8*SD’’ resulted in an average peak count of 3.25 for the control record-
ings, and this was used for the background subtraction in all in vivo measurements. In the peak detection algorithm above, a peak
was defined as any consecutive value (photon count) that surpasses the subtracted threshold (Figure S6A(vi)).
In order to assess the effective concentration range of the PeCoS methodology, two-fold dilution series of Alexa488 NHS Ester
were measured, and for every sample, both the ACF and the number of peaks were calculated using FCS and PeCoS (where Back-
ground = Mean ± 23*SD (water)), respectively. As expected, PeCoS did not perform well at high concentrations, where, presumably,
too many particles move simultaneously through the excitation volume; at lower particle concentrations, however, (where FCS was
no longer accurate) the number of peaks decreased in a nearly linear fashion (Figure S6B). To test the sensitivity of PeCoS under
in vivo conditions, the cytosolic concentration of Plk4-NG was measured at the beginning of nuclear cycle 12 in embryos expressing
either one (1x) or two (2x) copies of Plk4-NG in the Plk4 mutant background. PeCoS analysis indicated a 90% increase in the number
of peaks (per minute) in the 2x embryos compared to the 1x embryos, indicating the effectiveness of PeCoS in measuring the relative
cytosolic concentration of low abundance proteins (Figure S6C).

QUANTIFICATION AND STATISTICAL ANALYSIS

The details for quantification, statistical tests, sample numbers, definitions of center, and the measures for dispersion and precision
are described in the main text, relevant figure legends, or relevant sections of STAR Methods. Significance in statistical tests was
defined by p < 0.05. To determine whether the data values were normally distributed, a D’Agostino–Pearson omnibus normality
test was applied. Prism 7 and 8 were used for all the modeling and statistical analyses.

Cell 181, 1566–1581.e1–e14, June 25, 2020 e14

 ll
Article

Supplemental Figures

(legend on next page)

 ll
Article

Figure S1. Summary of the Protocol for Image Acquisition, Processing, and Analysis of the Plk4-NG Oscillations, Related to Figure 1
(A) Diagram illustrates the centrioles in ~2 h old embryo expressing Plk4-NG being imaged on a spinning-disk confocal system.
(B) Micrograph shows a typical image of the tracks of the Plk4-NG centrioles in S-phase of cycle 12, tracked using the ImageJ plugin, TrackMate.
(C) Graphs show the Plk4-NG oscillation during cycle 12 in a single embryo quantified from the tracks of either several individual centriole pairs (i), or the Mean ±
SD oscillation calculated from the tracks of > 90 centriole pairs (ii). The data for each embryo was then regressed using a Lorentzian equation (red line, iii)—see (D)
for an explanation of the rationale for choosing this function. This process was repeated for multiple embryos to calculate a Mean ± SEM regression for nuclear
cycle 12 (iv). R2 values indicate the goodness-of-fit (Mean ± SD) of the regression. CS = time of centrosome separation (set to 0); NEB = time of nuclear envelope
breakdown.
(D) Table shows the various models that were tested to fit the Plk4-NG oscillation data. R2 and SSAbs (absolute sum of squares) values indicate the goodness of fit.
The Lorentzian function was the best fit for the majority of embryos, so it was used for all further analyses.
Further details of these models are provided in STAR Methods.
 ll
Article

Figure S2. Plk4-NG Oscillations in Individual Embryos, Related to Figures 1 and 5

(A) Graphs show the Mean ± SD centriolar fluorescence intensity of Plk4-NG (two copies of a transgene expressed from its own promoter in a Plk4 null mutant
background) during nuclear cycles 11-13 in 5 different embryos imaged on a spinning-disk confocal system. n = 26 centrioles (mean) tracked starting from cycle
11 per embryo. See STAR Methods for full details of image acquisition and data analysis.
(B) Same as in (A), but showing the Plk4-NG oscillation in 5 embryos arrested in interphase by the injection of dsRNAs against Cyclin A, B and B3.
See Figure 5A for further details on sample numbers and experimental protocol.
ll
Article

(legend on next page)

 ll
Article

Figure S3. Simultaneously Measuring Centriole Growth and the Plk4 Oscillation in the Same Embryos, Related to Figure 2
(A and B) Graphs show the same data presented in Figure 2A, but with the SEM included (as these error bars were omitted from Figure 2A for ease of pre-
sentation). CS = centrosome separation and NEB = nuclear envelope breakdown. R2 values indicate the goodness of fit.
(C) Graph quantifies the embryo hatching frequency in embryos laid by either wild-type (Oregon-R) females or females simultaneously expressing Sas-6-mCherry
and Plk4-NG in a Plk4 mutant background (all mated with WT males). At least 4 technical repeats were carried out over several days, and a total of at least 400
embryos were analyzed.
(D) Cartoon graphs (i.e., imaginary data) illustrate the three different centriole growth phenotypes we observed in the Plk4 mutant embryos that simultaneously
express 2 copies of Plk4-NG and one copy of Sas-6-mCherry. In our previous analysis of centriole growth kinetics (Aydogan et al., 2018) almost all embryos
started to incorporate Sas-6-GFP at the very start of S-phase (‘‘Growth on time,’’ left graph). In the embryos analyzed here (with a more complicated genotype,
and expressing Sas-6-mCherry rather than Sas-6-GFP), some of the embryos exhibited a clear delay in initiating the incorporation of Sas-6-mCherry (‘‘Late
growth,’’ middle graph), while others did not appear to incorporate significant amounts of Sas-6-mCherry at all (‘‘No growth,’’ right graph).
(E) Pie charts quantify the percentage of embryos exhibiting each centriole growth phenotype at each nuclear cycle. Note that embryos exhibiting the ‘‘No
growth’’ phenotype were excluded from the analysis shown in (A) and (B) and in Figure 2A, although the amplitude of the Plk4 oscillations in these embryos was
analyzed separately (Figure 2C): we observed 8 embryos in total that exhibited the ‘‘No growth’’ phenotype (1 in cycle 12, and 7 in cycle 13). Centriolar Plk4-NG
levels continued to oscillate in these embryos, and the scatter graph shown in Figure 2C plots the peak amplitude of the Plk4-NG oscillations in these 8 embryos
overlaid on the average ‘‘threshold’’ level of Plk4-NG at which centrioles started to grow in the population of embryos that did exhibit Sas-6-mCherry incor-
poration. This threshold was very similar at cycle 12 and 13, so the threshold shown in Figure 2C is taken from cycle 13 embryos (as 7 of the 8 embryos shown here
were at cycle 13). The Plk4-NG oscillation in all but one of the 8 embryos failed to reach the average ‘‘threshold’’ level that would normally initiate centriole growth
in these embryos.
ll
Article

(legend on next page)

 ll
Article

Figure S4. Theoretical and Experimental Assessment of Several Assumptions Made in the Mathematical Model, Related to Figure 3
(A) Our mathematical model depicted in Figures 3A and 3B only discretely examines the Plk4-NG oscillation during S-phase of each nuclear cycle. The schematic
here shows our speculation that a phosphatase normally removes the phosphate groups (dotted circles) from Asl (red) during mitosis to reset the system for the
next oscillation at rate k4 (dotted black arrow).
(B) We implemented this step to extend the original model and plotted the mathematical solution for the percentage of Asl-bound Plk4 molecules (black curve) for
a total of 14 nuclear cycles. For simplicity we kept the length of S-phase and mitosis constant through all 14 cycles (see STAR Methods for further details of this
extended model).
(C) Schematic shows the Serine (S) and Threonine (T) residues (in bold) that were mutated to Alanine in the Asl-13A construct. Dark gray boxes show the relative
positions of the previously mapped Plk4-interacting regions within the N-terminal (Dzhindzhev et al., 2010) and C-terminal (Klebba et al., 2015) regions of Asl.
(D) Micrographs show images from time-lapse movies of embryos expressing Asl-WT-mKate2 and Asl-13A-mKate2 (under the control of their own promoters in
an asl mutant background), respectively.
(E) Graphs show the regression data (solid lines) for Plk4-NG oscillations in cycle 12 in embryos expressing either Asl-WT (green) or Asl-13A (dark gray) (both
without any fluorescent tag) simultaneously with Plk4-NG. N R 25 embryos for each condition; n = 71 and 68 centrioles (mean) per embryo in Asl-WT or Asl-13A,
respectively (collection of two trials performed by two independent researchers, blinded for each other’s data). Data are presented as Mean ± SEM R2 values
indicate goodness-of-fit for the regressions. CS = Centrosome separation; NEB = Nuclear envelope breakdown.
(F) In (B) it is assumed that the cytosolic concentration of Plk4 is kept constant over all cycles. The graph here plots an alternative model where the total number of
Plk4 molecules in the embryo is kept constant at all cycles. The number of centrioles doubles each cycle, and the mathematical solution for the percentage of Asl-
bound Plk4 molecules (black curve), and the percentage of Plk4 molecules that remain in the cytoplasm (red curve), is depicted over 14 nuclear cycles (see STAR
Methods for further details and implications of this model). For the first few nuclear cycles, almost all of the Plk4 remains in the cytoplasm since there are only a few
centrioles. In the later cycles, however, the amount of Plk4 sequestered by the Asl receptors increases exponentially, as the number of centrioles increase by a
factor of 2 in each cycle. Therefore, the rate at which the Asl receptors are able to recruit Plk4 from the cytoplasm decreases, resulting in a reduction in the
amplitude of the Plk4 oscillation. This aspect of the model is consistent with our experimental observations that the amplitude of the Plk4 oscillation decreases at
later cycles (Figure 1), as does the cytosolic concentration of Plk4 (Figure 3E). An alternative, or additional, mechanism that might explain these observations is
that the Plk4 molecules activated by binding to Asl may be more likely to autophosphorylate to stimulate their degradation, so ensuring that more Plk4 is degraded
at each cycle as the number of centrioles increase. Interestingly, in either of these scenarios, increasing centriole numbers leads to increasing Plk4 depletion from
the cytosol, potentially allowing embryos to effectively ‘‘count’’ their centrioles.
 ll
Article

Figure S5. FCS Analysis of Cytosolic Asl Levels, Related to Figure 3

(A) Schematic workflow describes the acquisition and analysis of point Fluorescence Correlation Spectroscopy (FCS) measurements (see STAR Methods for
further details). The 488nm laser beam is positioned at the centriolar plane in embryos expressing 2 copies of Asl-GFP (under the control of its own promoter in an
asl mutant background). (i) At the beginning of every cycle, when the old and new mother centrioles have just separated (white arrows), 6x 10 s FCS mea-
surements were taken at a point in the cytosol maximally distant from the centrioles (center of red crosshairs). (ii) This generated 6 autocorrelation functions
(ACFs) (a typical example is shown here). (iii) In the FoCuS-point software, 8 different models were fitted to each ACF. (iv) The model that best fitted the majority of
the data (#4 in this case) was chosen based on the Bayesian information criterion, and all ACFs were then fitted to this model. (v) The fitted ACFs were corrected
for background noise which was determined by measurements in WT embryos. (vi) The ACFs used for further analysis were then restricted by excluding individual
outlier measurements based on a ROUT-outlier test (Q = 1%) (these outlier measurements usually had a poor signal-to-noise ratio and gave concentrations that
were often biologically unrealistic, and were presumably generated when a centriole or non-specific fluorescent structure passed through the analyzed volume).
(B) Graph shows the average ACFs (represented as Mean ± SEM) for nuclear cycles 11-14 before background corrections. All individual ACFs were used to
calculate the cytosolic concentration data shown in Figure 3D.
(C) Western blot shows the protein levels of the Asl-GFP in either the early or late cell cycles from embryos of the same genotype used in (A) and (B). This supports
the results obtained from the FCS measurements, and suggests that total Asl levels do not change significantly during the development of the syncytial embryo.
Early and late embryos were separated based on their distinct morphology (judged by eye using a dissection microscope). Actin is shown as a loading control. A
representative blot is shown from two technical repeats.
 ll
Article

Figure S6. Peak Counting Spectroscopy Analysis of Cytosolic Plk4 Levels, Related to Figures 3 and 6
(A) Schematic workflow describes the acquisition and analysis of Peak Counting Spectroscopy (PeCoS) measurements. (i) In addition to embryos expressing
Plk4-NG under its own endogenous promoter, embryos of two other genotypes were placed on the same imaging dish. One expressing a green-fluorescent
centriole marker to allow correction of the spherical aberration caused by coverslip thickness variation, the other expressing Asl-mKate2 to determine the au-
tofluorescence background threshold for the Plk4-NG expressing embryos—Asl-mKate2 allows one to determine the correct plane (containing the centrioles;
white arrows) for background measurement, while the mKate2 fluorophore does not interfere with the PeCoS measurements. (ii) As for FCS (see Figure S5), a
488nm laser beam is positioned near the cortex of embryos, and the measurements are taken at a single point in the cytosol (red crosshairs) at the beginning of S-
phase, but for 1x 180 s, in both control and Plk4-NG expressing embryos (iii). Afterward, (iv) an appropriate threshold is calculated from the control embryos, so
that the background contributes less than 5 peaks on average during each recording. Following background subtraction, (v and vi) the number of peaks is
quantified.
(B) To compare the effective linear concentration range of FCS and PeCoS we assessed a two-fold dilution series of the Alexa488 dye. At high dye concentrations,
FCS (black symbols) exhibits a near-linear response, while PeCoS (gray symbols) is saturated—presumably because there are too many fluorophores in the
effective volume (Veff) for them to be measured as individual peaks. At intermediate dye concentrations, both methods exhibit a near linear response. At low
concentrations (~ < 0.2nM), however, FCS becomes unreliable while PeCoS continues to have a near-linear response.
(C) The bar chart shows the in vivo validation of PeCoS. A significant difference in the number of peaks per minute was observed between embryos expressing
either 1x or 2x copies of Plk4-NG (under the control of its endogenous promoter), which were measured at the beginning of S-phase in nuclear cycle 12. Each data
point represents a 180 s recording from a single embryo. Statistical significance was assessed using Mann-Whitney test (***p < 0.001). Data are presented as
Mean ± SD
(D) Western blot analysis of Plk4-GFP (arrow) levels in early and late embryos supports the conclusion from the PeCoS analysis (Figure 3E) that cytosolic Plk4
levels are lower in late embryos than in early embryos. Prominent non-specific bands are indicated (*). A representative blot is shown from two technical repeats.
(E) The bar chart compares the cytosolic levels of Plk4-NG (under the control of its endogenous promoter; at the beginning of S-phase in Cycle 13) between WT
and Wee1/ embryos (the same genotypes as in Figure 6). Statistical significance was assessed using an ordinary unpaired t test (ns, not significant). Data are
presented as Mean ± SD.
 ll
Article

Figure S7. Quantification of Centriolar Asl and Cytosolic Plk4 Levels When the Genetic Dose of asl or Plk4 Is Halved, Related to Figure 4
(A) Micrograph shows an image of Asl-mCherry at centrioles in an embryo in early S-phase (just after centrosome separation). Bar charts quantify the average
centriolar Asl-mCherry levels in early S-phase in either WT embryos (WT) or in embryos where the genetic dose of Plk4 has been halved (Plk41/2). N = 17 embryos
for each condition; n = 67 and 58 centrioles (mean) per embryo in WT or Plk41/2 groups, respectively. Average centriolar Asl levels do not change significantly
when the genetic dosage of Plk4 is halved, in agreement with the prediction of our model (see Data S1; first, second and third charts).
(B) Same schema as (A), but showing the localization of Asl-GFP, and quantifying the centriolar levels of Asl-GFP in asl mutant embryos expressing either 1 (Asl-
GFP1x) or 2 (Asl-GFP2x) copies of Asl-GFP. N = 10 embryos for each condition; n = 59 and 54 centrioles (mean) per embryo in Asl-GFP1x or Asl-GFP2x groups,
respectively. This analysis reveals that centriolar Asl-GFP levels drop by ~30% when the genetic dosage of Asl-GFP is halved, in good agreement with the
prediction of our model (see Data S1; first, second and third charts). Data are represented as Mean ± SEM. Statistical significance was assessed using an
unpaired t test with Welch’s correction (for Gaussian-distributed data) or an unpaired Mann-Whitney test (**p < 0.01; ns, not significant).
(C) Western blot compares the protein levels of Plk4-GFP (arrow) (expressed under the control of its own promoter in a Plk4 mutant background) in otherwise WT
embryos or in embryos in which the genetic dosage of asl has been halved. This analysis reveals that Plk4-GFP levels in the embryo do not change dramatically
when the genetic dosage of asl is halved, in agreement with the prediction of our model. WT embryos (Lane 1) are shown as a negative control to demonstrate that
the Plk4-GFP band is only detected in embryos expressing Plk4-GFP. Prominent non-specific bands are indicated (*). Actin is shown as a loading control. A
representative blot is shown from two technical repeats.

(legend continued on next page)

 ll
Article

(D) The bar chart compares the number of Plk4-NG peaks per minute that was observed between normal embryos (WT) or embryos where the genetic dose of Asl
was halved (asl1/2). Measurements were performed at the beginning of S-phase in nuclear cycle 12. Each data point represents a 180 s recording from a single
embryo. Statistical significance was assessed using an ordinary unpaired t test (for Gaussian-distributed data) or a Mann-Whitney test (ns, p > 0.05). Data are
presented as Mean ± SD.
 ll
Article

Figure S8. The Average Centriolar Plk4-NG Level on Individual Centrioles Can Be Used to Predict Stochastic Centriole Duplications in
Embryos Arrested in Interphase by Mitotic Cyclin Depletion, Related to Figures 5 and S2
(A) The pie chart quantifies the percentage of centrioles that continued to duplicate in embryos where cyclin A-B-B3 dsRNA was injected into embryos at nuclear
cycle 2-4, and centriole behavior assessed ~90 min later. Ambiguous (gray) indicates the fraction of centrioles whose duplication state could not be unam-
biguously determined due to their drifting out of focus during imaging.
(B) Bar chart shows the mean signal-to-noise ratio (SNR) of Plk4-NG fluorescence signals from sterile and fertile centrioles (red and green, respectively) through
the entire period of observation. Data are presented as Mean ± SD. Statistical significance of SNR was tested using a t test assuming equal variance (***p < 0.001).
(legend continued on next page)
 ll
Article

(C) Heatmap histogram of all SNR values from sterile and fertile centrioles. Red dashed line shows the unbiased threshold, determined automatically from Otsu
thresholding for distinguishing sterile and fertile centrioles. Heatmap (Red: Sterile and Green: Fertile) indicates the fraction of fertile/sterile centrioles in each
column. Note that, the higher the SNR, the more fertile the centrioles are.
(D) Confusion matrix shows the classification performance of sterile versus fertile centriole Plk4-NG signals using the Otsu threshold in (C) as a proportion of the
total number of signals, n = 81 centrioles from 3 embryos.
(E) Heatmap plots demonstrate the spatial (x,y) coordinates of all centriole duplication events in a representative cycling embryo (left; at the beginning of cycle 13
when centrioles are separating over the course of ~3.5 min) and non-cycling embryo (right; captured over ~60 min), as each duplication event colored light blue to
dark red to represent early and late time points, respectively. The black points plot the observed spatial (x,y) positions of all centrioles, duplicating and non-
duplicating, at all time points. Note that the duplications in the cycling embryo are spatially and temporally coordinated (tending to divide first at the top of the
embryo and later at the bottom of the embryo), while the duplications in the non-cycling embryo occur over a longer time-scale and do not appear to be co-
ordinated in space or time.
(F) To test more rigorously whether the centriole duplication events in non-cycling embryos are largely stochastic, we calculated Ripley K statistics for all the non-
cycling embryos used in (A–D). This statistic provides a measure of whether the temporal duplication events have spatial preference by measuring the average
number of events that occur as a function of distance from individual centrioles. Curves were computed from the (x,y) coordinates of only the duplicating
centrioles (denoted Kdivpoint , red line) and of all centrioles at all times (denoted Kallpoints , black line). The sigmoidal increase in the statistic as a function of distance in
both cases suggests that duplication events do not cluster spatially at short distances (< 50-100 pixels). The trend and amplitudes of red and black lines (mean ±
SD) are very similar and fall in each other’s statistical confidence range, indicating that duplicating centrioles do not exhibit additional spatial clustering above the
natural spatial distribution of centrioles.
(G) Schematic depicts the topology of the mathematical model that illustrates how Plk4 oscillations at individual centrioles could be generated to trigger sto-
chastic duplications in non-cycling embryos. Briefly, we no longer assume that a PPTase acts in discrete bursts during mitosis and instead assume a continuous
low-level PPTase activity (10% of the activity in cycling embryos). We allow individual Asl receptors to bind Plk4 and be phosphorylated until they release Plk4 (as
in our original model), and to be continuously be slowly dephosphorylated by the PPTase. Asl-p and Plk4-p indicate phosphorylated proteins. Bold arrows
indicate the dominant direction of the reactions.
(H) Graph shows how the percentage of Asl receptors that are bound to Plk4 changes over time at 10 individual centrioles, two of which have been colored red or
green. The centrioles are initially synchronized, as their Asl receptors all start in a dephosphorylated, unbound, state and so exhibit a coordinated pulse of Plk4
binding. As time progresses, however, the Asl receptors lose synchrony (as their dephosphorylation is no longer entrained by the CCO), and so each centriole
exhibits low amplitude stochastic oscillations. These oscillations may be sufficient to trigger centriole duplication under these conditions of interphase arrest with
low CCO activity, as evident from our experimental observations (Figures 5B–5F and Video S4; see STAR Methods for full details of the model).
 Article

A Unified Model for the Function of YTHDF Proteins in

Regulating m6A-Modified mRNA
Graphical Abstract Authors
Sara Zaccara, Samie R. Jaffrey

Correspondence
srj2003@med.cornell.edu

In Brief
Analysis of the transcriptome-wide
effects of m6A-mRNA effectors, known as
YTHDF proteins, demonstrates that they
act redundantly to induce degradation of
the same subset of mRNAs, with no
evidence of a direct role in promoting
translation.

Highlights
d YTHDF proteins function together to mediate degradation of
m6A-mRNAs

d YTHDF proteins show identical binding to all m6A sites

in mRNAs

d Each YTHDF paralog can compensate for the function of the

other YTHDF paralogs

d Depletion of all three YTHDF proteins promotes

differentiation of leukemia cells

Zaccara & Jaffrey, 2020, Cell 181, 1582–1595

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.012 ll
 ll

Article
A Unified Model for the Function of YTHDF Proteins
in Regulating m6A-Modified mRNA
Sara Zaccara1 and Samie R. Jaffrey1,2,*
1Department of Pharmacology, Weill Cornell Medicine, Cornell University, New York, NY 10065, USA
2Lead Contact
*Correspondence: srj2003@med.cornell.edu
https://doi.org/10.1016/j.cell.2020.05.012

SUMMARY

N6-methyladenosine (m6A) is the most abundant mRNA nucleotide modification and regulates critical as-
pects of cellular physiology and differentiation. m6A is thought to mediate its effects through a complex
network of interactions between different m6A sites and three functionally distinct cytoplasmic YTHDF
m6A-binding proteins (DF1, DF2, and DF3). In contrast to the prevailing model, we show that DF proteins
bind the same m6A-modified mRNAs rather than different mRNAs. Furthermore, we find that DF proteins
do not induce translation in HeLa cells. Instead, the DF paralogs act redundantly to mediate mRNA degrada-
tion and cellular differentiation. The ability of DF proteins to regulate stability and differentiation becomes
evident only when all three DF paralogs are depleted simultaneously. Our study reveals a unified model of
m6A function in which all m6A-modified mRNAs are subjected to the combined action of YTHDF proteins
in proportion to the number of m6A sites.

INTRODUCTION (24%). Based on these studies, each DF paralog mediates the

Methylation of adenosine in mRNA to form N6-methyladenosine
(m6A) has important roles in cellular physiology. Initial studies in
Arabidopsis and yeast showed defective seed development
and sporulation, respectively, upon genomic deletion of the
m6A-forming methyltransferase (Clancy et al., 2002; Zhong
et al., 2008). More recent studies have further documented the
connection between m6A and cellular differentiation in embry-
onic and hematopoietic stem cells, acute myeloid leukemia cells,
as well as other cell types (Barbieri et al., 2017; Cui et al., 2017;
Geula et al., 2015; Lee et al., 2019; Vu et al., 2017; Wang et al.,
2014b). Together, these studies show that m6A affects cellular
physiology and that these effects likely reflect modulation of
mRNA fate by m6A.
The effects of m6A in cytosolic transcripts are thought to be
mediated by a poorly understood and complex network of
interactions between specific m6A sites and specific members
of the YT521-B homology Domain-containing Family (YTHDF)
family of m6A-binding proteins (Shi et al., 2017, 2019). The
YTHDF family includes three paralogs (YTHDF1, YTHDF2, and
YTHDF3; also called DF1, DF2, and DF3, respectively), each of
which has different reported functions; DF1 enhances mRNA
translation, DF2 promotes mRNA degradation, and DF3 en-
hances translation and degradation (Li et al., 2017; Shi et al.,
2017; Wang et al., 2014a, 2015). According to the prevailing
model (Shi et al., 2017, 2019), the largest fraction of m6A-modi-
fied mRNAs (m6A-mRNAs; ~44%) bind a single DF paralog,
but some m6A-mRNAs (~32%) bind two DF paralogs. m6A-
mRNAs that bind all three DF paralogs are relatively rare
effects of m6A by targeting specific cohorts of m6A-mRNAs,
which has led to the major concept that different DF paralogs
control distinct physiologic processes (Anders et al., 2018; Han
et al., 2019; Hesser et al., 2018; Paris et al., 2019; Shi
et al., 2018).
Although the ability of the DF paralogs to bind different m6A
sites is fundamental to explain the different biology of the DF pa-
ralogs, it remains unclear how they achieve selective binding to
different m6A sites. Additionally, the mechanistic basis of the
different functions of DF proteins remains unknown, especially
in light of their high sequence identity (Patil et al., 2018; Wang
and He, 2014).
Here we present a fundamentally different model to explain the
effects of m6A on mRNA and how DF proteins mediate these ef-
fects. In contrast to the prevailing view that different m6A sites
bind largely different DF paralogs, we find that all m6A sites
bind all three DF paralogs in an essentially similar manner.
Thus, all DF paralogs regulate the same mRNAs based on the
presence of m6A sites. We also find that DF paralogs are not
translation enhancers, as currently thought for DF1 and DF3.
Instead, we show that the major effect of the three DF paralogs
is to promote mRNA degradation in a largely redundant manner.
The redundant effect of DF depletion has not been detected
previously because these proteins are typically depleted sepa-
rately rather than simultaneously. Depletion of only one protein
allows for varying degrees of compensation by the other paral-
ogs. Our finding that DF paralogs have similar functions rather
than different functions is supported by their essentially identical
m6A-binding properties, their similar set of associated binding

1582 Cell 181, 1582–1595, June 25, 2020 ª 2020 Elsevier Inc.
 ll
Article

A B

Figure 1. DF Proteins Bind the Same m6A Sites throughout the Transcriptome
(A) The amino acids that contact m6A and the m6A-proximal nucleotides are conserved in the DF1, DF2, and DF3 YTH domain. Conserved (gray) and non-
conserved (blue) amino acids are shown on the YTH domain, rendered from the DF1 m6A-RNA structure (Xu et al., 2015) (PDB: 4RCJ). Conserved is defined as
amino acids identical in all three DF paralogs, whereas non-conserved is defined as amino acids that are different in at least one DF paralog. Shown are a front
view (left) and back view (rotated 180
, right).
(B) DF1, DF2, and DF3 have similar binding preferences for different m6A submotifs. Shown is the prevalence of different binding sites recognized by DF1, DF2,
and DF3 based on iCLIP binding data. For comparison, the percentage of each DRACH motif identified by miCLIP is shown. The three DF paralogs have similar
binding site preferences. Their binding preferences are similar to the prevalence of the m6A sequence motifs.
(C) Each m6A site in the transcriptome binds DF1, DF2, and DF3. DF1, DF2, and DF3 binding at 4,182 m6A mRNAs was based on DF1, DF2, and DF3 iCLIP
datasets in HEK293T cells (Patil et al., 2016). m6A sites are plotted as points in which the x and y coordinates represent the number of normalized DF iCLIP reads

(legend continued on next page)

Cell 181, 1582–1595, June 25, 2020 1583

 ll
proteins, and their shared subcellular localizations. Furthermore,
we show that the effects of m6A on cellular differentiation can be
explained by the combined action of the three DF proteins rather
than individual DF proteins. Overall, these studies reveal a new
unified model of m6A function in which m6A predominantly influ-
ences mRNA degradation via the combined action of the three
largely redundant DF proteins.
RESULTS
Structural Analysis of Different YTH Domains Reveals
Their Similar RNA-Binding Properties
An important concept in m6A-mediated gene regulation is that
the different DF proteins bind distinct subsets of m6A residues
in the transcriptome (Han et al., 2019; Liu et al., 2018; Shi et al.,
2017, 2018, 2019; Wang et al., 2015). As a result, m6A is thought
to influence mRNAs in different ways, depending on which DF pa-
ralog it binds. Based on this, a set of DF1, DF2, or DF3 ‘‘unique’’
m6A sites is commonly used for analysis of the function of each
DF paralog (Han et al., 2019; Liu et al., 2018; Park et al., 2019;
Shi et al., 2017, 2018, 2019). Because this m6A-specific binding
behavior of the DF paralogs ultimately determines how m6A
regulates cellular processes, a major goal is to understand the
basis of selective m6A recognition by the DF paralogs.
To address this, we first asked whether differences in the
YT521-B homology (YTH) domains might account for the
different binding preferences of the DF proteins. The YTH
domain comprises ~134 amino acids and mediates selective
binding to m6A (Li et al., 2014; Patil et al., 2018; Xu et al.,
2015). We used the crystal structures of the DF1 and DF2 YTH
domains bound to m6A-containing RNA (Li et al., 2014; Xu
et al., 2015) to annotate the amino acids that recognize m6A
and the adjacent nucleotides. In the case of DF3, an RNA-bound
structure has not yet been reported; however, each of the amino
acids that bind m6A and the adjacent nucleotides in DF1 and DF2
is conserved in DF3 (Figure S1A), suggesting that the structural
mechanism of m6A binding is the same for all three DF proteins.
It remains possible that the few amino acids that differ be-
tween the YTH domains (Figure S1A) could affect RNA binding.
However, these amino acids map on the surface opposite the
RNA-binding pocket (Figure 1A), suggesting that they might
not affect YTH-RNA interactions. Furthermore, each full-length
DF protein shows a similar binding affinity for an m6A-containing
RNA oligonucleotide (Figure S1B), consistent with previous
studies (Arguello et al., 2019; Wang et al., 2014a; Xu et al.,
2015). Overall, the RNA-binding surfaces for the three YTH
domains appear to be identical.
The DF Paralogs Show Equivalent Binding to All m6A
Sites throughout the Transcriptome
The different transcriptome-wide binding properties reported
for the DF paralogs (Shi et al., 2017) could reflect the ability of
overlapping that site (log2 normalized). We did not find m6A sites that preferentially bound either DF1, DF2, or DF3. The high Pearson correlation coefficients (r)
show that DF paralogs have highly similar binding preferences. Similar results were found in HeLa cells using DF PAR-CLIP datasets (Figure S1E).
(D) DF1, DF2, and DF3 iCLIP reads show a similar distribution in mRNAs that resembles the miCLIP read distribution. Shown are representative examples of DF1,
DF2, and DF3 iCLIP read distribution and miCLIP read distribution on HNRNPF and MDM2. The iCLIP and miCLIP data shown here were obtained in HEK293T
cells and were similar to data obtained in HeLa cells (Figures S1F–S1H).
1584 Cell 181, 1582–1595, June 25, 2020
Article
the different DF paralogs to bind different m6A sequence motifs.
m6A residues are almost exclusively found in a single highly
conserved sequence motif: DR-m6A-CH (D = A, G, U; R = A,
G; H = A, C, U) (Canaani et al., 1979; Dimock and Stoltzfus,
1977; Linder et al., 2015; Wei et al., 1976). We reasoned that
the differential association of DF paralogs with distinct sets of
m6A sites could be explained by their preferences for certain
DR-m6A-CH submotifs, such as GG-m6A-CU versus AG-
m6A-CU.
To determine the binding preference for each DF paralog,
we examined our recently generated transcriptome-wide
iCLIP (individual-nucleotide-resolution UV crosslinking and
immunoprecipitation) maps of the binding sites of endogenously
expressed DF1, DF2, and DF3 in HEK293T cells (Patil et al.,
2016). We calculated the percentage of each DR-m6A-CH
sequence submotif at each DF-binding site identified by iCLIP
and ranked the most common sequence submotifs bound by
each DF paralog. These analyses showed a nearly identical fre-
quency of each DR-m6A-CH submotif bound by each DF paralog
(Figure 1B). The rank order of the prevalence of the submotifs
recognized by the DF paralogs was similar to the overall preva-
lence of each m6A submotif in the transcriptome (Linder et al.,
2015; Figure 1B). Thus, each DF paralog shows the same binding
preferences, which largely correlate with the prevalence of m6A
motifs, rather than a paralog-specific sequence preference.
Although the DF paralogs seem to bind m6A in proportion
to the prevalence of the m6A motif, each paralog may be targeted
to a different set of m6A sites, as suggested by previous studies
(Shi et al., 2017, 2019; Figure S1C). Thus, we sought to identify
the m6A sites uniquely bound by each DF paralog.
To do this, we quantified the number of DF1, DF2, and DF3
reads from the iCLIP datasets performed in HEK293T cells (Patil
et al., 2016) that map to each HEK293T m6A site (Linder et al.,
2015) or, similarly, the FLAG-DF1 and FLAG-DF2 reads from
the PhotoActivatable Ribonucleoside-enhanced Crosslinking
and ImmunoPrecipitation (PAR-CLIP) datasets performed in
HeLa cells (Shi et al., 2017; Wang et al., 2014a, 2015) and map-
ped at each HeLa m6A site (Ke et al., 2017). We could not use the
FLAG-DF3 PAR-CLIP dataset (Shi et al., 2017) because it con-
tained a low number of reads mappable to cytosolic mRNAs
(Figure S1D).
Using this m6A-centric approach, we found a strong linear cor-
relation in pairwise comparisons between DF1, DF2, and DF3
reads at each m6A site. This correlation was seen in the PAR-
CLIP datasets in HeLa cells mapped onto HeLa m6A sites (Fig-
ures S1E–S1F), as well as the iCLIP datasets in HEK293T cells
mapped onto HEK293T m6A sites (Figure 1C). Notably, in
contrast to the previous model where many m6A sites show DF
paralog-specific binding, we found no preferential enrichment
of iCLIP or PAR-CLIP reads for any paralog at any m6A site.
Overall, these results suggest that each m6A site is bound by
DF1, DF2, and DF3 in similar proportions.
 ll
Article

C D

(legend on next page)

Cell 181, 1582–1595, June 25, 2020 1585

 ll
Article

Because we were surprised by the lack of m6A sites uniquely
bound by each of the DF proteins, we reexamined the previously
reported DF1 and DF2 unique RNA-binding sites (Shi et al.,
2017). We asked whether DF1 binding, as measured by PAR-
CLIP signal, was missing at DF2 unique sites and vice versa.
To test this, we plotted DF2 binding at each DF1 unique site
and vice versa. This analysis shows that there is a comparable
level of DF1 and DF2 binding at each DF1 unique site and a com-
parable level of DF1 and DF2 binding at each DF2 unique site
(Figures S1G and S1H). Thus, the previously described unique
sites instead appear to be indistinguishable in terms of their
DF1 and DF2 binding.
Next, we directly examined the PAR-CLIP reads on individual
transcripts proposed to be DF1 unique or DF2 unique. After
extensive visual inspection, we found essentially no differences
in the distribution of DF1 and DF2 PAR-CLIP reads for any of
these transcripts (Figure S1J). Notably, the PAR-CLIP reads
correlated with the location of m6A-called sites, supporting the
idea that DF binding is due to m6A. A similar effect was seen
when using the iCLIP datasets for DF1, DF2, and DF3 (Figure 1D).
These data further suggest that the sites that were previously
described as being DF paralog unique are not, in fact, unique.
Our reexamination of the iCLIP and PAR-CLIP studies contrast
with previous reports showing that each DF paralog has only par-
tial overlap with m6A sites and with each other (Shi et al., 2017;
Wang et al., 2014a, 2015; Figure S1C). In the PAR-CLIP studies,
DF paralog-binding sites were determined based on a threshold
number of misincorporation-containing reads at any individual
site in the transcriptome. However, using a threshold approach
in independent experiments can often produce false negative
sites. These arise because of the arbitrary nature of a threshold,
which can cause some sites to be missed when they fall just
beneath that threshold. For this reason, the threshold approach
is not generally used for comparative analysis of datasets (Chak-
rabarti et al., 2018; Guertin et al., 2018; Landt et al., 2012).
Instead, comparative analysis is often performed by selecting a
set of specific transcriptomic sites, such as sequence motifs,
and determining whether one CLIP dataset or another shows
preferential binding to any of these sites (Dominguez et al.,
2018; Wheeler et al., 2018). Our approach is similar; by
comparing DF paralogs binding at mapped m6A sites, we
find that the RNA binding of DF paralogs is essentially identical,
even when using different cell lines and different CLIP methodol-
ogies (iCLIP and PAR-CLIP).
Overall, these diverse lines of evidence, including structural
similarity, motif analysis, and analysis of DF binding at each
m6A site in the transcriptome using CLIP datasets from two
cell lines, suggest that the DF paralogs do not exhibit different
patterns of m6A-binding in the transcriptome. Instead, they
appear to have similar binding preferences and affinities.
DF Proteins Exhibit Similar Protein Interaction Networks
Although we find that the DF paralogs bind the same m6A sites,
an unresolved question is how these nearly identical proteins
exert different molecular effects on m6A-mRNAs, especially in
the case of DF1 (translation) compared with DF2 (degradation).
A compelling possibility for divergent functions of DF paralogs
might lie within their effector domain, an ~40-kDa low-
complexity domain that comprises the remainder of the protein
outside of the YTH domain (Patil et al., 2018). Recruitment of
a distinct set of proteins to the effector domain of each DF pa-
ralog may allow the DF paralogs to exert different effects on
m6A-mRNAs. We thus analyzed the differences in the effector
domains between DF paralogs as well as their protein interaction
partners.
Examination of the three effector domains shows them to be
superficially similar. Each is a proline and glutamine-rich low-
complexity domain with ~60% amino acid identity and 70%
amino acid similarity (Figure S2A). Furthermore, the hydropho-
bicity and charge distribution along the length of the low-
complexity domains are highly similar. Additionally, the positions
of the disordered regions within the low-complexity domains are
similar for each paralog (Figure 2A).
The different DF proteins may have different protein interaction
partners to mediate their different functions. We therefore exam-
ined a recent comprehensive Bio-ID study in which 139 proteins
were BirA tagged, and interacting proteins were detected based
on their proximity-induced biotinylation (Youn et al., 2018). In
these experiments, 937–1,360 proteins were detected as
possible DF interactors, of which 63–103 were identified as
high-confidence interactors (63 of 937 DF1 interactors, 100 of
1,270 DF2 interactors, and 103 of 1,360 DF3 interactors).
To determine which interactors are preferentially bound by
each DF paralog, we performed a scatterplot analysis in which

Figure 2. DF Proteins Have Similar Binding Partners and Intracellular Localization

(A) The three DF paralogs have similar intrinsically disordered regions, hydrophobicity, and amino acid charge distribution along the length of their effector
domains. A schematic representation of each DF protein is shown, with the YTH domain indicated. Green indicates disordered regions (STAR Methods). Each of
these physiochemical parameters has high similarity among all DF paralogs, suggesting that these proteins may have similar properties.
(B) DF paralogs show similar binding preferences for their high-confidence interacting proteins. Shown are pairwise comparisons of the average spectral counts
of peptides derived from each protein identified as a DF paralog interactor in vivo (Youn et al., 2018). Each protein interactor is plotted as a circle in which x and y
coordinates represent the average spectral counts calculated based on a Bio-ID study of each DF paralog (log2-normalized values). The diameter is proportional
to the average probability of interaction (AvgP). As indicated by the orange circles (high-confident interactors with AvgP > 0.95 for both DF paralogs), there is a
high level of correlation between each pair of DF paralogs in the spectral counts and AvgP for each DF high-confidence interactor. These interactors are enriched
in proteins related to mRNA degradation pathways (shown in red). Proteins with a function in mRNA translation, including eIF3A and eIF3B, have a low AvgP and,
thus, are considered low-confidence or non-specific interactors.
(C) DFs are predicted to have similar subcellular localization. Shown is the ‘‘non-negative matrix factorization’’ probability assigned to each DF paralog in each
subcellular compartment, calculated based on the protein-protein interaction network (STAR Methods). Each DF paralog shows a similar probability of being
classified as a P-body protein and high probability of being enriched in ribonucleoprotein (RNP) granules or stress granules.
(D) Structured illumination microscopy (SIM) images indicate that the DF paralogs (red) are localized to small punctate structures (<150 nm) throughout the
cytoplasm and P-bodies (150–200 nm), based on colocalization with EDC4 (green). Scale bar is indicated.

1586 Cell 181, 1582–1595, June 25, 2020


Article
each biotinylated interactor was plotted as a circle, with the x
and y axis position each reflecting the average spectral counts
for one of the three DF paralogs. In this pairwise analysis, we
found similar proteins bound by all three paralogs. Additionally,
we found a marked linear correlation in the spectral counts for
each high-confidence protein interactor when examined in
pairwise comparisons of DF1, DF2, and DF3 (Figure 2B). Thus,
the DF binding partners seem to be shared and bind with the
same overall rank order preference for all three DF paralogs.
Notably, the top 25 interactors of all three DF paralogs
were highly similar and included high-scoring interactions
with components of the Carbon Catabolite Repression-
Negative On TATA-less (CCR4-NOT) RNA degradation complex,
such as CNOT1, CNOT7, and CNOT10 (Figures 2B and S2B).
Other high-confidence interactors of all three DF paralogs
included RNA degradation proteins such as PATL1, XRN1,
LSM12, and DDX6. In addition, all three DF paralogs interact
with protein components of stress granules (Figures 2B and
S2B). Stress granule proteins interact even in the absence of
stress (Markmiller et al., 2018; Youn et al., 2018). Notably, all of
these interactors have been seen in other studies. For example,
CNOT1 immunoprecipitation studies have identified all three DF
paralogs (Du et al., 2016). Additionally, an engineered ascorbate
peroxidase (APEX)-based proteomics analysis of the G3BP1
stress granule protein found all three DF paralogs to be interac-
tors (Markmiller et al., 2018). Thus, all three DF paralogs show
similar patterns of binding proteins, and the major interactions
are seen across independent proteomic datasets.
Because DF1 is thought to regulate translation through its
interactions with the eIF3A and eIF3B translation initiation fac-
tors (Shi et al., 2017; Wang et al., 2015), we wanted to determine
whether DF1 shows selective and robust interactions with these
proteins. The Bio-ID proteomics analyses (Figure 2B) shows
weak interactions of eIF3A and eIF3B with all three DF paralogs,
and in each case, the probability of interaction was low. Previous
analysis of DF1-binding partners identified eIF3A and eIF3B
based on mass spectrometry analysis of DF1 immunoprecipi-
tates (Wang et al., 2015). However, in that study, eIF3A and
eIF3B were among the lowest-scoring DF1 interactors (Fig-
ure S2C). The weak binding seen in immunoprecipitates is
consistent with the low probability seen in the Bio-ID studies
(Figure 2B). Thus, each proteomic study suggests that all three
DF paralogs exhibit nonspecific or low-level binding to eIF3A
and eIF3B in cells.
Overall, the protein-protein interactions of all three DF paral-
ogs are very similar, rather than different, with high-confidence
interactions with RNA degradation machinery and low-confi-
dence interactions with translation machinery.
DF Proteins Exhibit Similar Intracellular Localizations
To further understand the different functions of the DF paralogs,
we examined their subcellular localizations because distinct
subcellular localizations might be expected for proteins with
different functions. In the Bio-ID study (Youn et al., 2018), a
non-negative matrix factorization classification analysis was
developed to predict the subcellular localization of 139 RNA-
binding proteins based on their interaction partners. The
prominent localizations predicted for all three DF paralogs
ll
were P-bodies and cytoplasmic mRNA ribonucleoprotein
(mRNP) complexes (Figure 2C). Thus, DF paralogs are predicted
to have similar subcellular localizations.
To further test this, we examined the localization of each DF
paralog by immunofluorescence. Here we found that all three
DF paralogs showed essentially identical localization throughout
the cytoplasm in small punctate structures and, to a lower
extent, in larger punctate structures. Although the smaller ones
resemble the granule-like particles detected previously under
unstressed conditions (Youn et al., 2018), the larger punctate
structures were identified as P-bodies based on their colocaliza-
tion with EDC4, a P-body marker (Figure 2D). This is consistent
with an independent proteomic analysis using P-body markers
that also identified all three DF paralogs in P-bodies (Youn
et al., 2018). Thus, rather than exhibiting distinct localizations
in cells, DF proteins have similar punctate cytoplasmic localiza-
tions, some of which include localization to P-bodies. Overall,
these data suggest that DF proteins have highly similar se-
quences, functional domains, protein binding partners, and
intracellular localizations.
The Combined Activity of DF Proteins Leads to
Degradation of m6A-Modified mRNA
Because all DF paralogs show high-confidence interactions
with RNA degradation pathway proteins (Youn et al., 2018; Fig-
ure 2B), we considered the possibility that each paralog may
function to mediate degradation of m6A-mRNA.
Previous comparative studies of the DF paralogs found that
some of them induce mRNA degradation when artificially teth-
ered to a reporter transcript (Kennedy et al., 2016; Shi et al.,
2017; Tirumuru et al., 2016; Wang et al., 2014a, 2015). How-
ever, because these studies used heterologously overex-
pressed DF proteins, we were concerned about the potential
for these proteins to aggregate into stress granule-like struc-
tures when overexpressed. This phenomenon is seen in pro-
teins, like the DF paralogs, that contain low-complexity do-
mains (Alberti et al., 2019). Indeed, we observed variable
levels of DF granule-like structures after transfecting cells
with DF-expressing plasmids (Figure S2D). Thus, DF overex-
pression experiments may be misleading because of the
variable degrees of DF aggregation, which could sequester
proteins and potentially suppress their function. We therefore
re-examined the role of DF paralogs using knockdown ap-
proaches instead.
We first examined mRNA abundance using RNA sequencing
(RNA-seq) after small interfering RNA (siRNA)-mediated deple-
tion of DF1, DF2, and DF3 in HeLa cells. We used siRNA that
we validated for selective high-efficiency DF1, DF2, and DF3
knockdown (Patil et al., 2016; Figure S3A–S3C). Because m6A
sites on mRNAs bind all the three DF paralogs (Figures 1 and
S1), we examined how m6A-mRNAs were affected by DF
depletion and whether the effects were correlated with the num-
ber of m6A sites (Table S1). As in a previous study of DF1 (Wang
et al., 2015), we found no effect of DF1 depletion on m6A-mRNA
abundance compared with non-methylated mRNAs (Figure 3A).
In contrast, as reported previously (Wang et al., 2014a), deple-
tion of DF2, the most highly expressed DF paralog in HeLa cells
(Patil et al., 2018; Figure S3D), was associated with a small but
Cell 181, 1582–1595, June 25, 2020 1587
 ll
Article

A E

Figure 3. DF Proteins Redundantly Control the Abundance and Stability of m6A-mRNAs

(A–D) m6A-mRNAs show a substantially higher increase in expression upon simultaneous silencing of DF1, DF2, and DF3 in HeLa cells. The abundance of each
mRNA (based on RNA-seq counts) was compared between the control and the conditions of DF paralog silencing, as indicated. mRNAs were binned based on
the number of m6A sites. The expression distribution of each of m6A-mRNA subgroup is quantified in the boxplots. The center of each box represents the median
fold change (log2), the boundaries contain genes within a quartile of the median, and whiskers represent 1.53 interquartile ranges. The width of the boxplot is
proportional to the number of genes in each category. m6A-mRNAs do not change in expression upon silencing of DF1 (A) or DF3 (C) compared with non-
methylated mRNAs. However, m6A-mRNAs show a small increase in expression upon silencing of DF2 (B). m6A-mRNAs show a substantially higher increase in
expression upon simultaneous silencing of DF1, DF2, and DF3 (D). Only significant p values are shown in each graph (two-tailed Mann-Whitney test). n = 3
replicates; n = 2,425 mRNAs for 0, n = 894 mRNAs for 1, n = 1,521 mRNAs for 2–4, and n = 2,022 mRNAs for 5+ m6A sites.
(E) Quantification of the results in (A)–(D). The fold increase in mRNA expression was quantified for each indicated knockdown condition. Only mRNAs with 5 or
more annotated m6A sites were used in this analysis and compared with mRNAs with 0 annotated m6A sites. Although no effect on m6A mRNA expression is seen
with depletion of DF1 or DF3, a small increase is seen when these two are knocked down together. Triple knockdown shows a significantly larger effect compared
with knockdown of each DF paralog alone. ****p % 2.2e16, **p % 0.001, two-tailed Mann-Whitney test.
(F) m6A mRNA stability is increased upon depletion of DF paralogs. The stability of m6A-mRNAs and non-methylated mRNAs was determined by quantifying
mRNA levels before and 2 h after actinomycin D treatment. Shown is the fold change in mRNA levels, used as a measure of stability change upon silencing of the
DF paralog(s) (STAR Methods). The increase in mRNA stability is most apparent when all three DF paralogs were knocked down and is proportional to the number
of m6A sites per mRNA. n = 456 mRNAs for 0, n = 96 mRNAs for 1–5, n = 140 mRNAs for more than 5 annotated m6A sites. ***p = 2.079e6, **p % 0.01, *p % 0.02,
two-tailed Mann-Whitney test, n = 2 replicates.

1588 Cell 181, 1582–1595, June 25, 2020


Article
statistically significant increase in the abundance of m6A-
mRNAs (Figure 3B). This effect was more pronounced for
mRNAs with high numbers of annotated m6A sites. No increase
in m6A-mRNA abundance was observed following DF3 depletion
(Figure 3C), which, like DF1, is more lowly expressed in HeLa
cells (Patil et al., 2018; Figure S3D). Overall, our data confirm
the idea that DF2, but not DF1 or DF3, destabilizes m6A-mRNAs.
We then considered the possibility that the DF paralogs may
be functionally redundant and that our inability to detect a
stability-regulatory effect was due to compensation by the other
paralogs. Additionally, we noticed that knockdown of any of
the DF paralogs was associated with a compensatory increase
in the expression of the other paralogs in HeLa cells (Figures
S3A–S3B). Thus, compensatory upregulation of the other DF pa-
ralogs could further mask an effect of knockdown.
We therefore tested simultaneous knockdown of different
combinations of two DF paralogs or all three DFs. Knocking
down any two DF paralogs increased the overall abundance of
m6A-mRNA (Figures S3E–S3I). Importantly, the selective in-
crease in m6A-mRNA expression was largest upon triple knock-
down (Figure 3D), and, in each case, it was directly correlated
with the number of m6A sites per mRNA (Figures 3D and S3F–
S3H). Double knockdown and triple knockdown exhibited
greater increases in m6A-mRNA abundance than DF2 knock-
down alone (Figure 3E). These data are consistent with a
model in which any of the DF proteins can promote degradation
of m6A-mRNAs, but m6A-mRNA degradation is most effective
when all three DF paralogs are available, resulting in maximal
DF-dependent m6A-mRNA degradation.
To determine whether the expression levels seen upon
triple DF knockdown are correlated with changes in mRNA
stability levels, we measured the levels of m6A-mRNAs
after transcription inhibition with actinomycin D. Because
m6A-mRNAs tend to have a short half-life (Ke et al., 2017;
Schwartz et al., 2014), we treated cells with actinomycin D
for 2 h and measured the amount of each mRNA remaining
using RNA-seq. mRNA stability was relatively unaffected
upon depletion of each DF paralog individually, except in
the case of DF2, where a small stabilizing effect was seen
(Figure 3F). However, when all three DF paralogs were
knocked down, a substantial increase in mRNA stability was
seen for m6A-containing mRNAs compared with non-m6A-
mRNAs (Figure 3F). These effects were also seen in qRT-
PCR experiments examining the stability of specific mRNAs
annotated to contain high numbers of m6A sites or annotated
to lack m6A (Figure S3J).
Overall, these data further suggest that the most efficient
degradation of m6A-mRNAs occurs when all three DF paralogs
are present. When DF1 or DF3 is knocked down alone, the ef-
fects are nearly undetectable, but knockdown of DF1 and
DF3 together or in combination with DF2 resulted in readily
detectable increases in m6A-mRNA mRNA expression (Fig-
ure 3E). This suggests that the other DF paralogs can partially
compensate for the loss of another DF. The ability of any DF
protein to compensate is likely to be influenced by its expres-
sion level, with the lower expression of DF1 and DF3 (Fig-
ure S3D) making them less able to compensate for DF2 deple-
tion. In this model, when all three DF paralogs are depleted,
ll
compensation cannot occur, and m6A-mRNA stability is
increased most robustly.
DF Paralogs Do Not Affect the Translation of m6A-
Modified mRNAs
Although our data strongly indicate that DF paralogs act
together to destabilize m6A-mRNA, it remains possible that
one or more of these proteins could also control m6A-mRNA
translation. Importantly, DF1 and DF3 have been described to
be enhancers of m6A-mRNA translation (Shi et al., 2017; Wang
et al., 2015). Therefore, we examined the role of the DF paralogs
in translational regulation of m6A-mRNAs.
DF1 has been proposed to enhance translation by binding
m6A and recruiting eIF3 to 30 UTRs. This subsequently facilitates
loading of eIF3 onto mRNA caps (Wang et al., 2015). This is
reminiscent of the poly(A)-binding protein PABPC1, which also
binds 30 UTRs, binds an initiation factor, and promotes formation
of initiation complexes at mRNA caps (Le et al., 1997; Ozoe et al.,
2013; Wells et al., 1998). The ability of PABPC1 to enhance
mRNA translation is evident based on its enrichment in the
polysome fractions corresponding to highly translated mRNAs
(Arava et al., 2003).
We therefore examined whether DF1 is similarly enriched with
highly translated mRNAs. However, all three DF paralogs are
mostly excluded from the polysome fraction (Figure 4A and
S4A) and highly enriched in the fractions at the beginning of
the gradients, which represent cytoplasmic mRNPs (Arava
et al., 2003). This result is not consistent with a model in which
any DF protein is stably bound to mRNA 30 UTRs, enhancing
their translation.
Despite the absence of DF paralogs from the highly translating
mRNA pool, we wanted to examine whether translation of m6A-
mRNAs is enhanced by any of the DF paralogs.
We first examined the published ribosome profiling data that
revealed the translation-enhancing effect of DF1 (Wang et al.,
2015). In this study, processed data were provided, listing the
abundance of ribosome-protected fragments that map to each
gene. When we examined the processed data obtained after
DF1 silencing, it was evident that m6A-mRNAs were less trans-
lated than in control siRNA-transfected samples (Figure S4B).
Thus, consistent with the previous study (Wang et al., 2015),
but in contrast to what was expected from the polysome analysis
(Figure S4A), DF1 silencing causes a reduction in translation of
m6A-mRNAs. This analysis was performed using the average
of the two ribosome profiling replicates, as described in the orig-
inal study.
However, we found different results when the processed data
of the two DF1 siRNA replicates were analyzed separately. Repli-
cate 1 showed a prominent decrease in translation of m6A-
mRNAs upon DF1 depletion, consistent with the idea that DF1
enhances translation of m6A-mRNA (Figure S4C). However,
replicate 2 showed no change in translation of m6A-mRNAs
upon DF1 silencing (Figure S4D). Thus, the overall reduction in
m6A-mRNA translation we saw in the averaged data (Figure S4B)
was driven by the large drop in m6A-mRNA translation seen in
replicate 1.
Because we were surprised by the different roles of DF1
implied by the two different replicates, we re-processed the raw
Cell 181, 1582–1595, June 25, 2020 1589
 ll
Article

A Figure 4. Depletion of DF Proteins Does Not Affect the Translation

Efficiency of m6A mRNAs in HeLa Cells
(A) DF paralogs are not highly associated with actively translated mRNAs. DF1,
DF2, and DF3 were not enriched in polysomal fractions. Instead, they were
predominantly associated with the mRNA ribonucleoprotein (mRNP) complex
fraction. The ribosomal protein RPS6 was used as a marker for ribosome-
enriched fractions.
(B–E) The translation efficiency of m6A-mRNAs is not reduced upon silencing
of DF1 or any DF paralog. The number of ribosome-protected fragments
bound to each mRNA was normalized to the abundance of the respective
mRNA to calculate translation efficiency (TE). The TE was then compared
between control and cells knocked down for the indicated DF paralog(s) (DF1
in B, DF2 in C, DF3 in D). mRNAs were binned based on the number of m6A
sites. The distribution of each of m6A mRNA subgroups is quantified in the
boxplots. The center of each box represents the median fold change (log2), the
boundaries contain genes within a quartile of the median, and whiskers
represent 1.53 interquartile ranges. The width of the boxplot is proportional to
the number of genes in each category. m6A-mRNAs do not decrease their TE
B upon single or triple silencing of DF1, DF2, or DF3 (E) compared with non-
methylated mRNAs. Only significant p values are shown in each graph, and the
exact p values are reported (two-tailed Mann-Whitney test). n = 3 replicates;
n = 2,425 mRNAs for 0, n = 894 mRNAs for 1, n = 1,521 mRNAs for 2–4, n =
2,022 mRNAs for 5+ m6A sites.

sequencing data generated in the Wang et al. (2015) study. Upon

reanalysis using the standard pipeline for alignment and calcula-
tion of ribosome-protected fragments (Calviello et al., 2016;
McGlincy and Ingolia, 2017), both replicates showed the same
C result: DF1 depletion did not affect m6A-mRNA translation effi-
ciency (Figures S4E–S4G; Tables S2 and S3), similar to the effect
seen with the previously processed replicate 2 but not replicate 1.
To understand the basis of the inconsistencies between the
original replicate 1 and 2, we generated correlation coefficients
for the ribosome-protected fragments between replicates.
Here we saw only modest correlation in the number of ribo-
some-protected fragments per m6A-mRNA between the original
replicates (Figure S4H). In contrast, the re-processed data
showed high correlation (Figure S4I), as expected for replicates.
D
We therefore considered that a bioinformatic rather than a bio-
logical artifact may have caused variability in the original re-
ported analysis. Overall, based on one of the original replicates,
and based on both replicates after re-processing of the raw
ribosome profiling data from Wang et al. (2015), DF1 does not
regulate translation of m6A-mRNAs.
We next examined DF3, which has also been proposed previ-
ously to promote translation (Shi et al., 2017). In this previous
analysis, DF3-bound mRNAs were identified by PAR-CLIP, and
E the translation efficiency of these mRNAs was analyzed after
DF3 knockdown (Shi et al., 2017). However, we found that the
DF3 PAR-CLIP dataset did not contain a sufficient number of
mappable reads, as described above (Figure S1D). We therefore
decided to reanalyze the effect of DF3 depletion on mRNAs
based on the number of m6A sites because m6A sites correlate
with DF3 binding (Figure 1). Here we saw that DF3 depletion
does not affect the translation efficiency of m6A-mRNAs
(Figure S4J).
Because of the various inconsistencies, we wanted to inde-
pendently confirm the effects of depleting each DF paralog on
the translation efficiency of m6A-mRNAs using ribosome
profiling. In each experiment, we generated three replicates for

1590 Cell 181, 1582–1595, June 25, 2020


Article
A B
Figure 5. DF Proteins Redundantly Contribute to Suppressing
Differentiation of Leukemia Cells
(A) TNFRSF1B mRNA expression is increased upon depletion of all DF pa-
ralogs. TNFRSF1B mRNA levels were measured by qRT-PCR in MOLM-13
cells after knockdown of each paralog. DF2 silencing lead to increased levels
of TNFRSF1B. However, a larger increase was seen upon silencing all three DF
paralogs. TNFRSF1B levels were normalized to RPS28 expression levels, a
non-methylated mRNA (Vu et al., 2017). Non-parametric ANOVA test, ****p %
0.0001, n = 3, mean ± SEM.
(B) Expression of the CD14 differentiation marker in MOLM-13 cells is strongly
induced by silencing all three DF paralogs. Shown is the percentage of CD14+
cells after silencing of the indicated DF paralog(s). For each DF paralog, two
different short-hairpin RNAs (shRNAs) were tested in two biological replicates,
with the two different shRNAs indicated as a closed circle and an open circle.
Upon triple knockdown, the percentage of CD14+ cells is significantly
increased. Non-parametric ANOVA test; ****p < 0.0001, ***p = 0.0010, **p =
0.0012, *p = 0.03.
knockdown and control (Figures S3A, S3B, and S5A–S5C). We
found no reduction in translation efficiency of m6A-mRNAs
compared with non-methylated mRNAs in DF1-deficient HeLa
cells (Figure 4B). Similarly, neither DF2 nor DF3 depletion was
associated with a change in translation in m6A-mRNAs by ribo-
some profiling (Figures 4C and 4D). As a second approach,
we quantified the levels of specific m6A-mRNAs at each
position along the polysome gradient. We focused on specific
m6A-mRNAs reported to show the largest drop in translation af-
ter DF1 depletion (Wang et al., 2015). However, we saw no
decrease in translation upon DF1 depletion (Figure S5D).
Because all three DF proteins redundantly mediate mRNA
degradation, we considered the possibility that they may also
redundantly regulate mRNA translation. We therefore performed
ribosome profiling after knockdown of all three paralogs (Figures
S5E and S5F). Here we still saw no reduction but a slight
enhancement in the translation efficiency of m6A-containing
transcripts (Figure 4E).
Overall, multiple independent datasets demonstrate that DF1
has no translation-promoting role in HeLa cells, including the
original data used to show the translation-promoting effect of
DF1. Thus, none of the DF paralogs directly enhance mRNA
ll
translation. Instead, their major effect is to mediate m6A-mRNA
degradation.
The Combined Activity of DF Proteins Suppresses
Differentiation of Leukemia Cells
One of the most well-studied examples of m6A-regulated cellular
differentiation is in acute myeloid leukemia cell lines, where m6A
maintains cells in an undifferentiated blast-like state (Barbieri
et al., 2017; Vu et al., 2017). Knockdown of METTL3 induces
expression of markers such as CD14, a transmembrane protein
that reflects differentiation and commitment to the myeloid line-
age. Depletion of DF2 affects expression of m6A-RNAs whose
repression may contribute to leukemogenesis, such as
TNFRSF1B (Paris et al., 2019). Attempts to identify the m6A
reader that mediate the differentiation-suppressive effects of
m6A have not yet been successful. Recent studies have sug-
gested that DF2 does not mediate the differentiation-suppres-
sive effects of m6A because DF2 knockout does not induce dif-
ferentiation (Paris et al., 2019), which contrasts the effect of
METTL3 depletion (Barbieri et al., 2017; Vu et al., 2017). There-
fore, it has not been possible to ascribe the differentiation-sup-
pressive effects of m6A to a specific m6A reader.
We therefore asked whether suppression of TNFRSF1B is
mediated by the combined action of the three DF proteins. For
these experiments, we knocked down each DF transcript alone
or in combination in MOLM-13 leukemia cells, which, as we
and others showed previously, were maintained in an undifferen-
tiated state by METTL3 (Barbieri et al., 2017; Vu et al., 2017; Fig-
ure S5G). TNFRSF1B mRNA expression levels were slightly
affected by knockdown of each of the DF paralogs individually,
with the largest increase seen with DF2 knockdown (Figure 5A).
However, upon knockdown of all three DF paralogs, TNFRSF1B
expression was markedly enhanced (Figure 5A). Overall, these
data further support the idea that the DF proteins act redundantly
to control the expression levels of m6A-mRNAs in MOLM-
13 cells.
We next examined the role of DF proteins in suppressing dif-
ferentiation. DF2 depletion on its own does not induce differen-
tiation, as measured by the CD14 differentiation marker (Paris
et al., 2019). We therefore asked whether the combined action
of the DF paralogs suppresses differentiation. To test this idea,
we measured CD14 5 days after knockdown of DF paralogs us-
ing flow cytometry. Similar to the previous study (Paris et al.,
2019), we did not find induction of CD14 upon knockdown of
DF2 (Figure 5B). However, upon triple knockdown, CD14
expression was markedly increased. Overall, these data are
consistent with the idea that the three DF paralogs function
together to mediate the differentiation-suppressing effects
of m6A.
DISCUSSION
The YTHDF family proteins DF1, DF2, and DF3 are major cyto-
solic m6A-binding proteins and are thought to mediate the
effects of m6A in the cytosol (Patil et al., 2018). In contrast
to the prevailing model, where each DF paralog binds to
distinct subsets of mRNAs, we show that the DF paralogs
bind proportionately to each m6A site throughout the
Cell 181, 1582–1595, June 25, 2020 1591
 ll
transcriptome. Additionally, rather than exerting different ef-
fects on different m6A-mRNAs, we show that the three DF
proteins function together to mediate degradation of m6A-
containing mRNAs. Although depletion of single DF paralogs
leads to mild or no effects on mRNA abundance and stability,
depleting all three DF proteins leads to robust stabilization of
m6A-mRNAs, suggesting that each paralog can fully or
partially compensate for the function of the other DF paralogs.
Lastly, we find that previous studies that linked DF proteins to
mRNA translation were affected by bioinformatic and tech-
nical issues, which led to the incorrect view that a major func-
tion of DF proteins was to promote translation. Instead, we
show that DF paralogs do not appear to directly enhance
translation of m6A-mRNAs in HeLa cells. Our comprehensive
analysis of DF paralog function reveals a unified model of
DF protein binding and function, with the major effect of
m6A to mediate mRNA degradation through the combined ef-
fects of all three DF paralogs.
The issue of whether DF paralogs bind different mRNAs or
the same mRNAs is a central question for understanding m6A
function in cells. If each DF paralog binds different mRNAs,
then each DF would affect different cellular pathways and pro-
cesses. This prevailing model has created the impetus for
knockout studies focusing on separate analysis of each DF pa-
ralog, with the goal of understanding the specific functions of
each. However, this model lacks a clear mechanism to explain
how the DF paralogs could bind different m6A sites.
Our analysis supports the opposite conclusion. We find that
all m6A sites bind all DF proteins in an essentially indistinguish-
able manner, with the main determinant of DF paralog binding
simply being the presence of the m6A site. The level of DF bind-
ing is likely to be correlated with m6A stoichiometry, which would
positively correlate with the degradation effect. It should be
noted that some m6A sites may bind DF paralogs poorly when
they are obscured by local RNA structure or by binding of nearby
RNA-binding proteins that limit access to m6A.
Another central question is that of establishing the function of
the major cytoplasmic m6A readers DF1, DF2, and DF3. This is
arguably the most important step in understanding m6A biology
because it can explain how the effects of m6A can be rationalized
in diverse m6A-dependent processes. The current concept is
that m6A, through the action of different DF proteins, can induce
mRNA translation, mRNA degradation, or both and that this ef-
fect is transcript specific (Shi et al., 2017, 2019).
In contrast to this prevailing model, we find diverse lines of ev-
idence supporting the idea that DF proteins have similar rather
than different functions. In addition to the high sequence identity
in the RNA-binding and effector domains, DF proteins have
similar cytoplasmic localizations, including P-body localization,
and similar protein interactors, which are hallmarks of proteins
with similar functions. Most notable among the binding partners
for all three DF proteins are CCR4-NOT deadenylation complex
proteins, supporting the idea that all DF paralogs have a com-
mon role in mRNA degradation. A recent study showed that all
three DF proteins, when heterologously expressed, bind
CNOT1 and elicit mRNA deadenylation (Du et al., 2016). The abil-
ity of all three DF proteins to mediate deadenylation supports the
overall finding that the major function of these proteins is to
1592 Cell 181, 1582–1595, June 25, 2020
Article
mediate degradation of m6A-mRNAs. Further studies may also
address whether the position of m6A along the transcript body
has an effect on DF-mediated mRNA degradation.
Our data do not support the idea that DF paralogs have a
direct role in regulating mRNA translation efficiency. Our reanal-
ysis of previously published data, together with our independent
ribosome profiling datasets and analysis of individual m6A-
mRNAs by polysome fractionation analysis performed in the
same cell line, shows no evidence of decreased translation of
m6A-mRNAs upon depletion of DF proteins. This conclusion is
consistent with the previously published protein polysome frac-
tionation analysis (Wang et al., 2015) as well as the new one pre-
sented here, which do not show an enrichment of DF proteins in
the fractions containing highly translated mRNAs. Thus, DF pro-
teins, including DF1, do not behave like other translation-
enhancing RNA-binding proteins that bind the 30 UTR (Le et al.,
1997; Ozoe et al., 2013; Wells et al., 1998). Additionally, protein
interaction analysis shows a lack of high-confidence interactions
of the DF paralogs with translation initiation factors in cells (Youn
et al., 2018). Overall, these diverse lines of evidence suggest that
DF proteins do not promote translation. Although we cannot
exclude a role of DF proteins in translation enhancement in other
cell lines or conditions, our current findings are not consistent
with a role of any DF paralog in regulating m6A-mRNA transla-
tion, at least in the original cell line and conditions where which
DF1 was originally linked to translational regulation.
Although DF proteins do not promote translation, m6A can
promote translation through DF-independent mechanisms.
m6A can affect 30 UTR length (Ke et al., 2017), which can indi-
rectly affect translation. m6A may directly bind eIF3 when m6A
is in the 50 UTR (Meyer et al., 2015), or m6A may be associated
with bound METTL3, which may facilitate translation (Choe
et al., 2018). However, because eIF3 and METTL3 are thought
to bind just a small subset of m6A sites, the role of eIF3 and
METTL3 binding in the overall cellular effects of m6A is likely to
be limited to specific transcripts (Zaccara et al., 2019).
Our finding that all DF paralogs contribute to mRNA destabili-
zation reconciles findings made by diverse groups where
depletion of DF1 was not associated with reduced mRNA
translation. For example, DF1 depletion does not affect transla-
tion of a heterologously expressed m6A-mRNA in MCF7 cells
(Slobodin et al., 2017) and does not affect translation of m6A-
mRNAs in neurons when using a DF1 tethering system unless
a stimulus is added (Shi et al., 2018). Thus, the lack of translation
regulation seen upon DF1 depletion can now be explained by
the new model of DF function presented here.
In light of our current findings, phenotypes seen upon DF
depletion likely derive from upregulation of m6A-mRNAs.
Depletion of any single DF protein can affect mRNA levels to
a small degree, which can still result in clear cellular effects.
For example, recent studies show that m6A suppresses inter-
feron beta (IFNB1) mRNA levels and that knockdown of
different DF paralogs can cause a small increase in the abun-
dance and translation of this transcript (Winkler et al., 2019).
Because cells are sensitive to small increases in IFNB1, single
DF knockdown can lead to readily detectable cellular effects.
These hypomorphic phenotypes may be different from
METTL3 depletion because METTL3 depletion would cause

Article
more a substantial increase in the levels of highly m6A-modi-
fied mRNAs and affect multiple cellular pathways, potentially
leading to very different phenotypes.
Although DF proteins appear to have similar functions,
depletion of different DF paralogs can have different effects.
This is because DF proteins exhibit markedly different expres-
sion levels (Patil et al., 2018). Therefore, depletion of a low-
abundance DF paralog, such as DF3, is likely to only affect
a small number of highly sensitive mRNAs, whereas depletion
of a higher-abundance DF paralog would affect a larger sub-
set of mRNAs, causing a different phenotype. Additionally,
because DF paralogs may be expressed at different levels in
different tissues, single DF paralog depletion can result in tis-
sue-specific phenotypes (Nishizawa et al., 2017). DF paralogs
may also be phosphorylated in a paralog-specific manner
(Patil et al., 2018; Zaccara et al., 2019) or selectively induced
in response to specific stimuli, as seen with p63-dependent
induction of DF3 (Birkaya et al., 2007; Shi et al., 2017). These
pathways and subtle DF-paralog amino acid sequence differ-
ences may confer additional modes of regulation that can
affect the ability of DF paralogs to induce m6A-mRNA
degradation.
Notably, the functions of the DF paralogs may differ de-
pending on the cellular context. For example, in cell stress,
DF proteins bind m6A-mRNA and relocalize to stress granules,
but the mRNA is not targeted for degradation (Ries et al.,
2019). In neurons, DF proteins have also been localized
to transport granules and may therefore have roles in traf-
ficking m6A-mRNA to dendrites and, thus, indirectly promote
translation in neurons (Merkurjev et al., 2018; Ries et al.,
2019). However, for cell types that exhibit the classic
mRNA-destabilization effect of m6A, which has been seen in
diverse cell types (Geula et al., 2015; Ke et al., 2017; Vu
et al., 2017) and was originally described in 1978 (Sommer
et al., 1978), the likely mediator of this m6A-dependent
destabilization effect is the combined action of DF1, DF2,
and DF3.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Material Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Cell Culture
B siRNA silencing in HeLa cells
B shRNA silencing in MOLM-13 cells
B DF1, DF2 and DF3 overexpression and imaging
B Protein expression and purification
d METHOD DETAILS
B Flow cytometry
B Antibody staining and 3D-SIM imaging.
B Polysome
ll
B Generation of the ribosome profiling and RNA seq li-
braries upon DFs silencing
B Actinomycin D treatment
B Quantitative PCR analysis
B Microscale Thermophoresis (MST)
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Protein-protein interactome analysis
B Analysis of the physiochemical properties of each DF
paralog.
B Reanalysis of publicly available PAR-CLIP data of DF1,
DF2 and DF3 in HeLa cells
B Comparison of the coverage of DF proteins at each
m6A site on mRNAs throughout the transcriptome.
B Calculation of coverage at each DF1 and DF2
unique sites
B Analysis of the ribosome profiling and RNA seq data
upon silencing of DF proteins
B Reanalysis of publicly available ribosome profiling data
upon silencing of DF1 in HeLa cells
6
B Analysis of m A mRNA stability upon DF paralog
depletion.
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.012.
ACKNOWLEDGMENTS
We thank members of the Jaffrey laboratory for comments and suggestions, in
particular V. Despic, and D. Patil for early contributions to this project. We
thank members of the Epigenomics and Flow Cytometry Weill Cornell Cores.
We thank A. North and the Rockefeller Bio-Imaging Resource Center for assis-
tance with SIM microscopy. We thank the Rockefeller High-Throughput
Screening Resource Center for assistance with the MST experiments. We
thank Y. Cheng and M. Kharas for advice regarding the experiments using leu-
kemia cells. This work was supported by NIH grants R35NS111631 and
R01CA186702 (to S.R.J.) and an American-Italian Cancer Foundation fellow-
ship (to S.Z.).
AUTHOR CONTRIBUTIONS
S.Z. and S.R.J. designed the experiments. S.Z. performed the experiments.
S.Z. and S.R.J. wrote the manuscript.
DECLARATION OF INTERESTS
S.R.J. is scientific founder of, advisor to, and owns equity in Gotham
Therapeutics.
Received: July 23, 2019
Revised: January 14, 2020
Accepted: May 4, 2020
Published: June 2, 2020
REFERENCES
Alberti, S., Gladfelter, A., and Mittag, T. (2019). Considerations and Challenges
in Studying Liquid-Liquid Phase Separation and Biomolecular Condensates.
Cell 176, 419–434.
Anders, M., Chelysheva, I., Goebel, I., Trenkner, T., Zhou, J., Mao, Y., Ver-
zini, S., Qian, S.-B., and Ignatova, Z. (2018). Dynamic m6A methylation
Cell 181, 1582–1595, June 25, 2020 1593
 ll
facilitates mRNA triaging to stress granules. Life Sci. Alliance 1,
e201800113.
Arava, Y., Wang, Y., Storey, J.D., Liu, C.L., Brown, P.O., and Herschlag, D.
(2003). Genome-wide analysis of mRNA translation profiles in Saccharomyces
cerevisiae. Proc. Natl. Acad. Sci. USA 100, 3889–3894.
Arguello, A.E., Leach, R.W., and Kleiner, R.E. (2019). In Vitro Selection
with a Site-Specifically Modified RNA Library Reveals the Binding Prefer-
ences of N6-Methyladenosine Reader Proteins. Biochemistry 58,
3386–3395.
Bailey, T.L., Boden, M., Buske, F.A., Frith, M., Grant, C.E., Clementi, L., Ren,
J., Li, W.W., and Noble, W.S. (2009). MEME SUITE: tools for motif discovery
and searching. Nucleic Acids Res. 37, W202-8.
Barbieri, I., Tzelepis, K., Pandolfini, L., Shi, J., Millán-Zambrano, G., Robson,
S.C., Aspris, D., Migliori, V., Bannister, A.J., Han, N., et al. (2017). Promoter-
bound METTL3 maintains myeloid leukaemia by m6A-dependent translation
control. Nature 552, 126–131.
Birkaya, B., Ortt, K., and Sinha, S. (2007). Novel in vivo targets of DeltaNp63 in
keratinocytes identified by a modified chromatin immunoprecipitation
approach. BMC Mol. Biol. 8, 43.
Calviello, L., Mukherjee, N., Wyler, E., Zauber, H., Hirsekorn, A., Selbach, M.,
Landthaler, M., Obermayer, B., and Ohler, U. (2016). Detecting actively trans-
lated open reading frames in ribosome profiling data. Nat. Methods 13,
165–170.
Canaani, D., Kahana, C., Lavi, S., and Groner, Y. (1979). Identification and
mapping of N6-methyladenosine containing sequences in simian virus 40
RNA. Nucleic Acids Res. 6, 2879–2899.
Chakrabarti, A.M., Haberman, N., Praznik, A., Luscombe, N.M., and Ule, J.
(2018). Data Science Issues in Studying Protein–RNA Interactions with CLIP
Technologies. Annu. Rev. Biomed. Data Sci. 1, 235–261.
Choe, J., Lin, S., Zhang, W., Liu, Q., Wang, L., Ramirez-Moya, J., Du, P.,
Kim, W., Tang, S., Sliz, P., et al. (2018). mRNA circularization by METTL3-
eIF3h enhances translation and promotes oncogenesis. Nature 561,
556–560.
Clancy, M.J., Shambaugh, M.E., Timpte, C.S., and Bokar, J.A. (2002). Induc-
tion of sporulation in Saccharomyces cerevisiae leads to the formation of N6-
methyladenosine in mRNA: a potential mechanism for the activity of the IME4
gene. Nucleic Acids Res. 30, 4509–4518.
Cui, Q., Shi, H., Ye, P., Li, L., Qu, Q., Sun, G., Sun, G., Lu, Z., Huang, Y.,
Yang, C.-G., et al. (2017). m6A RNA Methylation Regulates the Self-
Renewal and Tumorigenesis of Glioblastoma Stem Cells. Cell Rep. 18,
2622–2634.
Dimock, K., and Stoltzfus, C.M. (1977). Sequence specificity of internal
methylation in B77 avian sarcoma virus RNA subunits. Biochemistry 16,
471–478.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Dodt, M., Roehr, J.T., Ahmed, R., and Dieterich, C. (2012). FLEXBAR-Flexible
Barcode and Adapter Processing for Next-Generation Sequencing Platforms.
Biology (Basel) 1, 895–905.
Dominguez, D., Freese, P., Alexis, M.S., Su, A., Hochman, M., Palden, T., Ba-
zile, C., Lambert, N.J., Van Nostrand, E.L., Pratt, G.A., et al. (2018). Sequence,
Structure, and Context Preferences of Human RNA Binding Proteins. Mol. Cell
70, 854–867.e9.
Du, H., Zhao, Y., He, J., Zhang, Y., Xi, H., Liu, M., Ma, J., and Wu, L. (2016).
YTHDF2 destabilizes m(6)A-containing RNA through direct recruitment of
the CCR4-NOT deadenylase complex. Nat. Commun. 7, 12626.
Geula, S., Moshitch-Moshkovitz, S., Dominissini, D., Mansour, A.A., Kol, N.,
Salmon-Divon, M., Hershkovitz, V., Peer, E., Mor, N., Manor, Y.S., et al.
(2015). Stem cells. m6A mRNA methylation facilitates resolution of naı̈ve plu-
ripotency toward differentiation. Science 347, 1002–1006.
1594 Cell 181, 1582–1595, June 25, 2020
Article
Guertin, M.J., Cullen, A.E., Markowetz, F., and Holding, A.N. (2018). Parallel
factor ChIP provides essential internal control for quantitative differential
ChIP-seq. Nucleic Acids Res. 46, e75.
Han, D., Liu, J., Chen, C., Dong, L., Liu, Y., Chang, R., Huang, X., Liu, Y.,
Wang, J., Dougherty, U., et al. (2019). Anti-tumour immunity controlled
through mRNA m6A methylation and YTHDF1 in dendritic cells. Nature
566, 270–274.
Hesser, C.R., Karijolich, J., Dominissini, D., He, C., and Glaunsinger, B.A.
(2018). N6-methyladenosine modification and the YTHDF2 reader
protein play cell type specific roles in lytic viral gene expression during
Kaposi’s sarcoma-associated herpesvirus infection. PLoS Pathog. 14,
e1006995.
Holehouse, A.S., Das, R.K., Ahad, J.N., Richardson, M.O.G., and Pappu, R.V.
(2017). CIDER: Resources to Analyze Sequence-Ensemble Relationships of
Intrinsically Disordered Proteins. Biophys. J. 112, 16–21.
Ke, S., Pandya-Jones, A., Saito, Y., Fak, J.J., Vågbø, C.B., Geula, S.,
Hanna, J.H., Black, D.L., Darnell, J.E., Jr., and Darnell, R.B. (2017). m6A
mRNA modifications are deposited in nascent pre-mRNA and are not
required for splicing but do specify cytoplasmic turnover. Genes Dev. 31,
990–1006.
Kennedy, E.M., Bogerd, H.P., Kornepati, A.V., Kang, D., Ghoshal, D., Marshall,
J.B., Poling, B.C., Tsai, K., Gokhale, N.S., Horner, S.M., and Cullen, B.R.
(2016). Posttranscriptional m(6)A Editing of HIV-1 mRNAs Enhances Viral
Gene Expression. Cell Host Microbe 19, 675–685.
Lancaster, A.K., Nutter-Upham, A., Lindquist, S., and King, O.D. (2014).
PLAAC: a web and command-line application to identify proteins with prion-
like amino acid composition. Bioinformatics 30, 2501–2502.
Landt, S.G., Marinov, G.K., Kundaje, A., Kheradpour, P., Pauli, F., Batzoglou,
S., Bernstein, B.E., Bickel, P., Brown, J.B., Cayting, P., et al. (2012). ChIP-seq
guidelines and practices of the ENCODE and modENCODE consortia.
Genome Res. 22, 1813–1831.
Lauria, F., Tebaldi, T., Bernabò, P., Groen, E.J.N., Gillingwater, T.H., and Viero,
G. (2018). riboWaltz: Optimization of ribosome P-site positioning in ribosome
profiling data. PLoS Comput. Biol. 14, e1006169.
Le, H., Tanguay, R.L., Balasta, M.L., Wei, C.C., Browning, K.S., Metz, A.M.,
Goss, D.J., and Gallie, D.R. (1997). Translation initiation factors eIF-iso4G
and eIF-4B interact with the poly(A)-binding protein and increase its RNA bind-
ing activity. J. Biol. Chem. 272, 16247–16255.
Lee, H., Bao, S., Qian, Y., Geula, S., Leslie, J., Zhang, C., Hanna, J.H., and
Ding, L. (2019). Stage-specific requirement for Mettl3-dependent m6A
mRNA methylation during haematopoietic stem cell differentiation. Nat. Cell
Biol. 21, 700–709.
Li, F., Zhao, D., Wu, J., and Shi, Y. (2014). Structure of the YTH domain of hu-
man YTHDF2 in complex with an m(6)A mononucleotide reveals an aromatic
cage for m(6)A recognition. Cell Res. 24, 1490–1492.
Li, A., Chen, Y.-S., Ping, X.-L., Yang, X., Xiao, W., Yang, Y., Sun, H.-Y., Zhu, Q.,
Baidya, P., Wang, X., et al. (2017). Cytoplasmic m6A reader YTHDF3 promotes
mRNA translation. Cell Res. 27, 444–447.
Linder, B., Grozhik, A.V., Olarerin-George, A.O., Meydan, C., Mason, C.E., and
Jaffrey, S.R. (2015). Single-nucleotide-resolution mapping of m6A and m6Am
throughout the transcriptome. Nat. Methods 12, 767–772.
Liu, J., Eckert, M.A., Harada, B.T., Liu, S.M., Lu, Z., Yu, K., Tienda, S.M., Chry-
plewicz, A., Zhu, A.C., Yang, Y., et al. (2018). m6A mRNA methylation regulates
AKT activity to promote the proliferation and tumorigenicity of endometrial
cancer. Nat. Cell Biol. 20, 1074–1083.
Madeira, F., Park, Y.M., Lee, J., Buso, N., Gur, T., Madhusoodanan, N., Basut-
kar, P., Tivey, A.R.N., Potter, S.C., Finn, R.D., and Lopez, R. (2019). The EMBL-
EBI search and sequence analysis tools APIs in 2019. Nucleic Acids Res. 47
(W1), W636–W641.
Markmiller, S., Soltanieh, S., Server, K.L., Mak, R., Jin, W., Fang, M.Y., Luo, E.-
C., Krach, F., Yang, D., Sen, A., et al. (2018). Context-Dependent and Disease-
Specific Diversity in Protein Interactions within Stress Granules. Cell 172, 590–
604.e13.

Article
McCarthy, D.J., Chen, Y., and Smyth, G.K. (2012). Differential expression anal-
ysis of multifactor RNA-Seq experiments with respect to biological variation.
Nucleic Acids Res. 40, 4288–4297.
McGlincy, N.J., and Ingolia, N.T. (2017). Transcriptome-wide measurement of
translation by ribosome profiling. Methods 126, 112–129.
Merkurjev, D., Hong, W.-T., Iida, K., Oomoto, I., Goldie, B.J., Yamaguti, H.,
Ohara, T., Kawaguchi, S.-Y., Hirano, T., Martin, K.C., et al. (2018). Synaptic
N6-methyladenosine (m6A) epitranscriptome reveals functional partitioning of
localized transcripts. Nat. Neurosci. 21, 1004–1014.
Meyer, K.D., Patil, D.P., Zhou, J., Zinoviev, A., Skabkin, M.A., Elemento, O.,
Pestova, T.V., Qian, S.B., and Jaffrey, S.R. (2015). 50 UTR m(6)A Promotes
Cap-Independent Translation. Cell 163, 999–1010.
Nishizawa, Y., Konno, M., Asai, A., Koseki, J., Kawamoto, K., Miyoshi, N., Ta-
kahashi, H., Nishida, N., Haraguchi, N., Sakai, D., et al. (2017). Oncogene c-
Myc promotes epitranscriptome m6A reader YTHDF1 expression in colorectal
cancer. Oncotarget 9, 7476–7486.
Oates, M.E., Romero, P., Ishida, T., Ghalwash, M., Mizianty, M.J., Xue, B.,
Dosztányi, Z., Uversky, V.N., Obradovic, Z., Kurgan, L., et al. (2013). D2P2:
database of disordered protein predictions. Nucleic Acids Res. 41,
D508–D516.
Olarerin-George, A.O., and Jaffrey, S.R. (2017). MetaPlotR: a Perl/R pipeline
for plotting metagenes of nucleotide modifications and other transcriptomic
sites. Bioinformatics 33, 1563–1564.
Ozoe, A., Sone, M., Fukushima, T., Kataoka, N., Arai, T., Chida, K., Asano, T.,
Hakuno, F., and Takahashi, S. (2013). Insulin receptor substrate-1 (IRS-1)
forms a ribonucleoprotein complex associated with polysomes. FEBS Lett.
587, 2319–2324.
Paris, J., Morgan, M., Campos, J., Spencer, G.J., Shmakova, A., Ivanova,
I., Mapperley, C., Lawson, H., Wotherspoon, D.A., Sepulveda, C., et al.
(2019). Targeting the RNA m6A Reader YTHDF2 Selectively Compromises
Cancer Stem Cells in Acute Myeloid Leukemia. Cell Stem Cell 25,
137–148.e6.
Park, O.H., Ha, H., Lee, Y., Boo, S.H., Kwon, D.H., Song, H.K., and Kim, Y.K.
(2019). Endoribonucleolytic Cleavage of m6A-ContainingRNAs by RNase P/
MRP Complex. Mol. Cell 74, 494–507.e8.
Patil, D.P., Chen, C.-K., Pickering, B.F., Chow, A., Jackson, C., Guttman, M.,
and Jaffrey, S.R. (2016). m(6)A RNA methylation promotes XIST-mediated
transcriptional repression. Nature 537, 369–373.
Patil, D.P., Pickering, B.F., and Jaffrey, S.R. (2018). Reading m6A in the Tran-
scriptome: m6A-Binding Proteins. Trends Cell Biol. 28, 113–127.
Quinlan, A.R., and Hall, I.M. (2010). BEDTools: a flexible suite of utilities for
comparing genomic features. Bioinformatics 26, 841–842.
Ramı́rez, F., Ryan, D.P., Grüning, B., Bhardwaj, V., Kilpert, F., Richter, A.S.,
Heyne, S., Dündar, F., and Manke, T. (2016). deepTools2: a next generation
web server for deep-sequencing data analysis. Nucleic Acids Res. 44 (W1),
W160-5.
Ries, R.J., Zaccara, S., Klein, P., Olarerin-George, A., Namkoong, S.,
Pickering, B.F., Patil, D.P., Kwak, H., Lee, J.H., and Jaffrey, S.R. (2019).
m6A enhances the phase separation potential of mRNA. Nature 571,
424–428.
Risso, D., Ngai, J., Speed, T.P., and Dudoit, S. (2014). Normalization of RNA-
seq data using factor analysis of control genes or samples. Nat. Biotechnol.
32, 896–902.
Schwartz, S., Mumbach, M.R., Jovanovic, M., Wang, T., Maciag, K., Bushkin,
G.G., Mertins, P., Ter-Ovanesyan, D., Habib, N., Cacchiarelli, D., et al. (2014).
Perturbation of m6A writers reveals two distinct classes of mRNA methylation
at internal and 50 sites. Cell Rep. 8, 284–296.
Shi, H., Wang, X., Lu, Z., Zhao, B.S., Ma, H., Hsu, P.J., Liu, C., and He, C.
(2017). YTHDF3 facilitates translation and decay of N6-methyladenosine-
modified RNA. Cell Res. 27, 315–328.
ll
Shi, H., Zhang, X., Weng, Y.-L., Lu, Z., Liu, Y., Lu, Z., Li, J., Hao, P., Zhang, Y.,
Zhang, F., et al. (2018). m6A facilitates hippocampus-dependent learning and
memory through YTHDF1. Nature 563, 249–253.
Shi, H., Wei, J., and He, C. (2019). Where, When, and How: Context-Depen-
dent Functions of RNA Methylation Writers, Readers, and Erasers. Mol. Cell
74, 640–650.
Slobodin, B., Han, R., Calderone, V., Vrielink, J.A.F.O., Loayza-Puch, F., Elkon,
R., and Agami, R. (2017). Transcription Impacts the Efficiency of mRNA Trans-
lation via Co-transcriptional N6-adenosine Methylation. Cell 169, 326–
337.e12.
Sommer, S., Lavi, U., and Darnell, J.E., Jr. (1978). The absolute frequency of
labeled N-6-methyladenosine in HeLa cell messenger RNA decreases with la-
bel time. J. Mol. Biol. 124, 487–499.
Tirumuru, N., Zhao, B.S., Lu, W., Lu, Z., He, C., and Wu, L. (2016). N(6)-meth-
yladenosine of HIV-1 RNA regulates viral infection and HIV-1 Gag protein
expression. eLife 5, e15528.
Vu, L.P., Pickering, B.F., Cheng, Y., Zaccara, S., Nguyen, D., Minuesa, G.,
Chou, T., Chow, A., Saletore, Y., MacKay, M., et al. (2017). The N6-meth-
yladenosine (m6A)-forming enzyme METTL3 controls myeloid differentia-
tion of normal hematopoietic and leukemia cells. Nat. Med. 23,
1369–1376.
Wang, X., and He, C. (2014). Reading RNA methylation codes through methyl-
specific binding proteins. RNA Biol. 11, 669–672.
Wang, X., Lu, Z., Gomez, A., Hon, G.C., Yue, Y., Han, D., Fu, Y., Parisien, M.,
Dai, Q., Jia, G., et al. (2014a). N6-methyladenosine-dependent regulation of
messenger RNA stability. Nature 505, 117–120.
Wang, Y., Li, Y., Toth, J.I., Petroski, M.D., Zhang, Z., and Zhao, J.C. (2014b).
N6-methyladenosine modification destabilizes developmental regulators in
embryonic stem cells. Nat. Cell Biol. 16, 191–198.
Wang, X., Zhao, B.S., Roundtree, I.A., Lu, Z., Han, D., Ma, H., Weng, X., Chen,
K., Shi, H., and He, C. (2015). N(6)-methyladenosine Modulates Messenger
RNA Translation Efficiency. Cell 161, 1388–1399.
Wei, C.M., Gershowitz, A., and Moss, B. (1976). 50 -Terminal and internal
methylated nucleotide sequences in HeLa cell mRNA. Biochemistry 15,
397–401.
Wells, S.E., Hillner, P.E., Vale, R.D., and Sachs, A.B. (1998). Circulariza-
tion of mRNA by eukaryotic translation initiation factors. Mol. Cell 2,
135–140.
Wheeler, E.C., Van Nostrand, E.L., and Yeo, G.W. (2018). Advances and chal-
lenges in the detection of transcriptome-wide protein-RNA interactions. Wiley
Interdiscip. Rev. RNA 9, e1436.
Winkler, R., Gillis, E., Lasman, L., Safra, M., Geula, S., Soyris, C., Nachshon,
A., Tai-Schmiedel, J., Friedman, N., Le-Trilling, V.T.K., et al. (2019). m6A modi-
fication controls the innate immune response to infection by targeting type I in-
terferons. Nat. Immunol. 20, 173–182.
Wu, C.C.-C., Zinshteyn, B., Wehner, K.A., and Green, R. (2019). High-Res-
olution Ribosome Profiling Defines Discrete Ribosome Elongation States
and Translational Regulation during Cellular Stress. Mol. Cell 73,
959–970.e5.
Xu, C., Liu, K., Ahmed, H., Loppnau, P., Schapira, M., and Min, J. (2015). Struc-
tural Basis for the Discriminative Recognition of N6-Methyladenosine RNA by
the Human YT521-B Homology Domain Family of Proteins. J. Biol. Chem. 290,
24902–24913.
Youn, J.-Y., Dunham, W.H., Hong, S.J., Knight, J.D.R., Bashkurov, M., Chen,
G.I., Bagci, H., Rathod, B., MacLeod, G., Eng, S.W.M., et al. (2018). High-Den-
sity Proximity Mapping Reveals the Subcellular Organization of mRNA-Asso-
ciated Granules and Bodies. Mol. Cell 69, 517–532.e11.
Zaccara, S., Ries, R.J., and Jaffrey, S.R. (2019). Reading, writing and erasing
mRNA methylation. Nat. Rev. Mol. Cell Biol. 20, 608–624.
Zhong, S., Li, H., Bodi, Z., Button, J., Vespa, L., Herzog, M., and Fray, R.G.
(2008). MTA is an Arabidopsis messenger RNA adenosine methylase and inter-
acts with a homolog of a sex-specific splicing factor. Plant Cell 20, 1278–1288.
Cell 181, 1582–1595, June 25, 2020 1595
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mouse anti-Edc4 (H-12) Santa Cruz Biotechnology Cat# sc-376382; RRID: AB_10988077
Rabbit anti-YTHDF1 Proteintech Group Cat# 17479-1-AP; RRID: AB_2217473
Rabbit anti-YTHDF2 Aviva System Biology Cat# ARP67917_P050
Rabbit anti-YTHDF3 Abcam Cat# ab103328; RRID: AB_10710895
Rabbit anti-GAPDH Santa Cruz Biotechnology Cat# sc-365062; RRID: AB_10847862
Rabbit anti-S6 (5G10) Cell Signaling Technology Cat# 2217; RRID: AB_331355
Alexa Fluor 488 goat anti-mouse IgG (H+L) Thermo Fisher Scientific Cat# A-11001; RRID: AB_2534069
Alexa Fluor 568 goat anti-rabbit IgG (H+L) Thermo Fisher Scientific Cat# A-11011; RRID: AB_143157
Mouse anti-CD14 (M5E2) BD Bioscience Cat# 555398; RRID: AB_395799
Rabbit anti-PABP Abcam Cat# ab21060; RRID: AB_777008
Bacterial and Virus Strains
Rosetta2 DE3 Novagen Cat# 70954
Chemicals, Peptides, and Recombinant Proteins
Human 6xHis-YTHDF1 Patil et. al., 2016 https://doi.org/10.1038/nature19342
Human 6xHis-YTHDF2 Patil et. al., 2016 https://doi.org/10.1038/nature19342
Human 6xHis-YTHDF3 Patil et. al., 2016 https://doi.org/10.1038/nature19342
Critical Commercial Assays
Talon Metal Affinity Resin Clontech Cat# PT1320-1
Monolith Protein Labeling Kit RED-NHS 2nd Generation Nanotempertech Cat# MO-L011
NEBNext Ultra II Directional RNA Library Prep Kit New England Biolabs Cat# E7760L
for Illumina
NEBNext rRNA Depletion Kit (Human/Mouse/Rat) New England Biolabs Cat# E6310X
Deposited Data and Code
RNA-seq, siDF1-siDF2-siDF3 in HeLa cells This paper GSE134380
Ribo-seq, siDF1-siDF2-siDF3 in HeLa cells This paper GSE134380
RNA-seq upon Actinomycin D treatment, This paper GSE134380
siDF1-siDF2-siDF3 in HeLa cells
Code use to analyze the Ribo-seq datasets This paper https://doi.org/10.17632/nsz68ywvvx.1
iCLIP, DF1-DF2-DF3 in HEK293T cells Patil et al., 2016 GSE78030
PARCLIP, DF1 in HeLa cells Wang et al., 2015 GSE63591
PARCLIP, DF2 in HeLa cells Wang et al., 2014a GSE49339
PARCLIP, DF3 in HeLa cells Shi et al., 2017 GSE86214
Ribo-seq and RNA-seq, siDF1 in HeLa cells Wang et al., 2015 GSE63591
Ribo-seq and RNA-seq, siDF2 in HeLa cells Wang et al., 2014a GSE49339
Ribo-seq and RNA-seq, siDF3 in HeLa cells Shi et al., 2017 GSE86214
m6A-CLIP-seq in HeLa cells Ke et. al., 2017 GSE86336
miCLIP-seq in HEK293T Linder et al., 2015 GSE63753
Experimental Models: Cell Lines
Human female origin: HeLa cells ATCC Cat# CCL-2
Human male origin: MOLM13 cells Gift from M. Kharas’s lab NA
(Continued on next page)

e1 Cell 181, 1582–1595.e1–e8, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Oligonucleotides
Single m6A-DRACH 10-nt oligo: rUrCrCrGrG/ Xu et al., 2015 https://doi.org/10.1074/jbc.M115.680389
iN6Me-rA/rCrUrGrU
Human YTHDF2 shRNA Sigma Cat# TRCN000254410
Human YTHDF3 shRNA Sigma Cat# TCN000365173
Human YTHDF1, YTHDF2 and YTHDF3 shRNA Patil et al., 2016 https://doi.org/10.1038/nature19342
ERCC RNA Spike-in Mix Thermo Fisher Scientific Cat# 4456740
Recombinant DNA
pcDNA3-FLAG-HA Patil et. al., 2016 https://doi.org/10.1038/nature19342
pcDNA3-FLAG-HA-YTHDF1 Patil et. al., 2016 https://doi.org/10.1038/nature19342
pcDNA3-FLAG-HA-YTHDF2 Patil et. al., 2016 https://doi.org/10.1038/nature19342
pcDNA3-FLAG-HA-YTHDF3 Patil et. al., 2016 https://doi.org/10.1038/nature19342
Software and Algorithms
GraphPad Prism 8 GraphPad Software https://www.graphpad.com/scientific-
software/prism/
RStudio (1.0.153) RStudio https://rstudio.com/
Fiji (ImageJ 2.0.0-rc-68/1.52h) NIH https://fiji.sc/
Imaris Bitplane https://imaris.oxinst.com
FLEXBAR Dodt et. al., 2012 https://github.com/seqan/flexbar/wiki
riboWaltz Lauria et al., 2018 https://github.com/
LabTranslationalArchitectomics/
riboWaltz
STAR Dobin et al., 2013 https://github.com/alexdobin/STAR
BEDTools Quinlan and Hall, 2010 https://github.com/arq5x/bedtools2
MEME Suite Bailey et al., 2009 http://meme-suite.org/index.html
DeepTools Ramı́rez et al., 2016 https://deeptools.readthedocs.io/en/
develop/

RESOURCE AVAILABILITY

Lead Contact
Information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Samie R. Jaffrey
(srj2003@med.cornell.edu).

Material Availability
This study did not generate new unique reagents.

Data and Code Availability

All datasets generated during this study have been deposited to the NCBI Gene Expression Omnibus under accession number GEO:
GSE134380.
The code used to analyze the ribosome profiling datasets in this study is available at Mendeley Data: https://doi.org/10.17632/
nsz68ywvvx.1.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cell Culture
HeLa (ATCC CCL-2) cells of female origin were maintained in 1x DMEM (11995-065, Life Technologies) with 10% FBS, 100 U ml-1
penicillin and 100 mg ml-1 of streptomycin under standard culture conditions. Cells were split with TrypLE Express (Life Technologies)
according to the manufacturer’s instructions. HeLa cells were authenticated. MOLM-13 (man origin) were maintained in 1x RPMI with
10% FBS, 100 U ml-1 penicillin and 100 mg ml-1 of streptomycin under standard culture condition. MOLM-13 cells were a kind dona-
tion from the laboratory of M. Kharas and were not authenticated during this study.

Cell 181, 1582–1595.e1–e8, June 25, 2020 e2

 ll
Article

siRNA silencing in HeLa cells

Cells were seeded in 6-well plates. After 24 h, 30 nM siRNA was transfected using Pepmute transfection reagent (Signagen) accord-
ing to the manufacturer’s instructions. 48 h after the first transfection, a second transfection was performed. Cells were maintained at
70%–80% confluency and when necessary expanded in a 10-cm dish. Cells were then collected 4 days after the first transfection.
When silencing multiple DF proteins at the same time, the total concentration of siRNA in the final transfection mixture was main-
tained constant (i.e., silencing of DF1, DF2, and DF3 was performed by adding 10 nM of siRNAs targeting DF1, 10 nM of siRNAs tar-
geting DF2 and 10 nM of siRNAs targeting DF3 in the same transfection mixture, i.e., silencing of DF1 and DF2 was performed by
adding 15 nM of siRNA for DF1 and 15 nM of siRNA for DF2 in the same transfection mixture). siRNA sequences were previously
reported and validated (Patil et al., 2016). Knockdown was also additionally validated prior to all experiments reported in this study
by western blot. In each silencing condition, the siRNA sequences were specific to the DF paralog of interest.

shRNA silencing in MOLM-13 cells

MOLM-13 cells at a concentration of 500,000 cells ml-1 were transduced with lentiviruses expressing shRNAs against DF1, DF2,
and DF3 or expressing a control scrambled shRNA sequence as performed previously (Vu et al., 2017). To increase the virus inte-
gration efficiency, cells were spinoculated at 1400 RPM for 1 h. After 24 h from the spinoculation (Day 1), cell media was changed.
Cells were then collected 5 days after the transduction. shRNAs sequences against DF1, DF2, and DF3 were previously reported
(Ries et al., 2019). In order to obtain at least two shRNAs for each DF, additional MISSION shRNA plasmid DNA sequences were
purchased from Sigma (DF2 shRNA: TRCN000254410, DF3 shRNA: TCN000365173). EGFP/TGFP expression was used to select
efficiently transduced cells by FACS sorting.

DF1, DF2 and DF3 overexpression and imaging

HeLa cells were transfected with FLAG-3xHA-tagged DF1, FLAG-3xHA-tagged DF2 and FLAG-3xHA-tagged DF3-expressing
plasmids using Fugene HD Transfection Reagent (E2311, Promega) according to the manufacturer’s instructions. As control, a
FLAG-3xHA-expressing plasmid was used as control. Briefly, cells were plated on 35-mm glass-bottom dishes coated with 1%
gelatin and allowed to reach 40%–60% confluency before transfection. 24 h after transfection, plates were fixed, and immunostain-
ing was performed as described. To visualize the overexpressed DF, staining using an antibody against each DF paralog and the HA
tag was performed. Areas where the HA staining overlaps with the DF staining indicate the localization of the FLAG-3xHA-tagged DF.
After antibody staining, images were acquired using a Nikon TE-2000 inverted microscope with a 40x oil objective.

Protein expression and purification

For in vitro binding affinity studies of m6A-modified RNA to DF paralogs, each full-length DF paralog was purified in bacteria as pre-
viously described (Patil et al., 2016; Ries et al., 2019). Briefly, DF proteins were expressed in Escherichia coli Rosetta2 (DE3) (Nova-
gen) using pProEx HTb (Invitrogen) with checking every 20 min at 600nm (OD600) until the culture reached log phase (04-0.6 OD600).
DF protein expression was induced with 1 mM isopropyl b-D-1-thiogalactopyranoside (IPTG) for 16 h at 18
C. Cells were collected,
pelleted and then resuspended in the following buffer: 50 mM NaH2PO4 pH 7.2, 300 mM NaCl, 20 mM imidazole at pH 7.2 and sup-
plemented with EDTA-free protease inhibitor cocktail (05892791001, Roche) according to the manufacturer’s instructions. The cells
were lysed by sonication and then centrifuged at 10,000g for 20 min. The soluble protein was purified using Talon Metal Affinity Resin
(Clontech) and eluted in the following buffer: 50 mM NaH2PO4 pH 7.2, 300 mM NaCl, 250 mM imidazole-HCl at pH 7.2. Further con-
centration and buffer exchange were performed using Amicon Ultra-4 spin columns (Merck-Millipore) when proteins were not directly
labeled. Recombinant protein was stored in the following buffer: 20 mM HEPES pH 7.4, 300 mM KCl, 6 mM MgCl2, 0.02% NP40, 50%
glycerol at 80
C or 20% glycerol at 20
C. All protein purification steps were performed at 4
C. The purified protein was quan-
tified using a ND-2000C NanoDrop spectrophotometer (NanoDrop Technologies) with OD 280 and verified by Coomassie staining.

METHOD DETAILS

Flow cytometry
After 5 days of DF silencing, MOLM-13 cells were tested for the expression of the CD14 differentiation marker using the following
protocol. Cells were collected by centrifugation, washed once with 1X PBS+2% FBS and then counted using a hemocytometer.
One million cells were used to perform cellular staining using PE-CD14 antibody (BD PharMingen). The antibody dilution was per-
formed according to the manufacturer’s instructions. Staining was performed on ice for 1 h. After the staining, excess unbound anti-
body was removed by two washes with 1x PBS. DAPI was added prior to analysis. Cells were analyzed on a BD FACS ARIA instru-
ment. To calculate the percentage of CD14 positive cells, data were processed using FlowJo.

Antibody staining and 3D-SIM imaging.

HeLa cells were seeded on coated no 1.5H (170mm ± 5mm) coverslip (474030-9000-000, Carl Zeiss) in 6 well-plates. Coverslips
were coated with gelatin prior to seeding. 24 h after seeding, cells were washed twice with PBS at 25
C and then fixed with 4% form-
aldehyde in 1X PBS for 15 min at 25
C. After two washes with 1X PBS to remove the excess of formaldehyde, cells were
permeabilized with permeabilization/blocking buffer (1x FBS, 1% Triton-X in 1x PBS) at room temperature for 30 min. Following

e3 Cell 181, 1582–1595.e1–e8, June 25, 2020

 ll
Article

permeabilization/blocking, cells were incubated with the primary antibody in a humidified chamber for 2 h. Cells were then washed
with 1x PBS for three times and then incubated with the appropriate secondary antibody (anti-rabbit Alexa Fluor 568, anti-mouse
Alexa Fluor 488) for 1 h at room temperature. Following the incubation, cells were washed as before. Hoechst staining was performed
for 5 min. After additional washes, coverslips were mounted in mounting media (Prolong Diamond, P36961, Life Technologies) and
quickly sealed with nail polish. Stained cells were imaged by super-resolution 3D-SIM on OMX Blaze 3D-SIM super-resolution
microscope (Applied Precision) equipped with a 100x/1.40 UPLSAPO oil objective. To reduce spherical aberrations, an oil with
the optimal refractive index was first identified at the beginning of every acquisition session. Image reconstruction and alignment
was performed using SoftWoRx. Because all acquired images have AF-568 staining (red staining), Optimal-Red transfer functions
(OTFs) were used during the image processing. Further processing was performed on Imaris.

Polysome
HeLa cells were seeded on 10-cm dishes. At 70%–80% confluency, cells were treated with 100 mg/mL cycloheximide for 10 min at
37
C and then collected. Briefly, cells were washed with PBS in mild cycloheximide condition and then lysed directly on the dish us-
ing 400 ml of lysis buffer (20 mM Tris HCl pH 7.4 100 mM KCl, 5 mM MgCl2, 1% Triton X-100, 100 mg/ml cycloheximide, 2 mM DTT, 1x
cOmplete no EDTA protease inhibitor cocktail). The lysate was then left on ice for 10 min and then a centrifugation at 12,000 g x 15 min
was performed to clear the lysate. The cleared lysates were then loaded in 15%–50% linear sucrose gradients, ultra-centrifuged
and fractionated with an automated fraction collector. Proteins were extracted from each fraction using trichloroacetic acid. Sodium
deoxycholate was added as a carrier to assist in the protein precipitation. After acetone addition and washes, proteins were resus-
pended in protein loading buffer, denatured at 95
C for 10 min, and used for detecting DF1, DF2, DF3 and RPS6 by western blot. RNA
was extracted from each fraction using TRIzol-LS reagent (Invitrogen). To account for differences in extraction efficiency, 2 ng of a
spike-in luciferase mRNA (TriLink Biotechnologies) was added to each fraction before RNA extraction. Briefly, RNA extraction was
performed using a ratio of 2:1 between the volume of TRIzol LS reagent and the sample. After TRIzol addition, RNA isolation was
performed as indicated by the manual’s instructions.

Generation of the ribosome profiling and RNA seq libraries upon DFs silencing
Ribosome profiling was performed following the original protocol (McGlincy and Ingolia, 2017) with minor modifications. Briefly, HeLa
cells were plated on a 10-cm dish and each DF or combination of DF transcripts was silenced as described. After the silencing, to
inhibit ribosome run-off during the collection and lysis, cells were rapidly washed once with ice-cold PBS containing 50 mg/ml cyclo-
heximide and the plates were immersed in liquid nitrogen and placed in dried ice as previously described (Calviello et al., 2016). After
allowing cells to reach 4
C by keeping cells on ice, cells were immediately lysed in 400 mL of cell lysis buffer (20 mM Tris pH 7.4,
150 mM NaCl, 5 mM MgCl2, 1 mM DTT, 100 mg/mL cycloheximide, 1% Triton, 25 U DNase I) and by triturating the sample through
a 26-gauge needle ten times to further lyse the cellular material. 5% of the total cellular extract was used for RNaseq library prepa-
ration. 10% of the total cellular extract was used for Western Blot validation of the silencing. Lysate was then clarified by performing a
centrifugation step at 20,000 x g for 10 min at 4
C. Supernatant was collected, flash-frozen and stored at 80
C.
For the ribosome profiling library preparation, the RNA concentration in the cleared lysate was quantified using RiboGreen. 30 mg of
RNA was digested using RNase I (Epicenter) as previously described (McGlincy and Ingolia, 2017). After RNase digestion, lysates
were loaded on a sucrose gradient and centrifuged at 100,000 rpm (TLA 100.3 rotor) for 1 h to pellet ribosomes. RNA from the re-
suspended ribosomal pellet was purified using Trizol and run on a gel to selectively excise foot-printed RNAs (from 17 nt to 34 nt
in length). To reduce ribosomal contamination in the library preparation steps, we then performed Ribo Zero Gold depletion of the
foot-printed RNA. The rRNA-depleted RNA fragments were dephosphorylated, and the adaptor was ligated. To specifically deplete
unligated linker, yeast 50 -deadenylase and RecJ exonuclease digestion was performed. At this point, the library preparation steps
were performed essentially as described previously (McGlincy and Ingolia, 2017). Briefly, reverse transcription was performed using
SuperScript III. To avoid untemplated nucleotide addition, reverse transcription was carried out at 57
C, as described (McGlincy and
Ingolia, 2017). cDNA purification, circularization, and amplification were performed as previously described (Linder et al., 2015). Us-
ing this method, every ribosomal footprint represents a unique ribosome-binding event. Libraries were sequenced with a single end
50 bp run using an Illumina Hiseq2500 platform.
For the RNaseq library preparation, the RNA was extracted from the 5% of the lysate before the RNase digestion using Trizol LS.
The RNA quality was assessed by Bioanalyzer analysis. 1 mg of total RNA was used for library preparation using the NEBNext Ultra
Directional RNA Library Prep Kit. Ribosomal RNA was removed using NEBNext rRNA Depletion Kit. The libraries were sequenced on
the Illumina HiSeq 2500 instrument, in single read mode, 50 bases per read.

Actinomycin D treatment
HeLa cells were plated on a 10-cm dish and each DF or combination of DF proteins was silenced as described. After five days of
silencing, cells were treated with 5 mg/ml of actinomycin D or vehicle (DMSO) for 2 h to inhibit transcription. To ensure that the treat-
ment did not affect cell viability, cells were counted before collection. Total RNA was extracted using TRIzol and 1 mg of total RNA was
used to perform RNA-seq library preparation using the NEBNext Ultra Directional RNA Library Prep Kit. Ribosomal RNA was removed

Cell 181, 1582–1595.e1–e8, June 25, 2020 e4

 ll
Article

using NEBNext rRNA Depletion Kit. Prior to ribosomal depletion, 2 mL of the 1:100 dilution of the ERCC (external RNA control
consortium) RNA Spike-in Control Mix 1 was added to each 1 mg of total RNA sample, as suggested by the manufacturer’s
instructions.

Quantitative PCR analysis

Total RNA was isolated from cells using TRIzol according to the manufacturer’s instructions. For each condition, the same amount of
total RNA was reverse transcribed to cDNA using the SuperScript IV First-Strand kit. When the transcript abundance was tested upon
actinomycin D treatment in HeLa cells, oligo-dT primers were used during the cDNA synthesis step. This allowed us to selectively
convert to cDNA the amount of intact RNA still present in cells upon actinomycin D treatment, while avoiding the conversion of frag-
mented RNA to cDNA. When transcript abundance was tested in MOLM13 cells, random hexamers were used. Quantitative qPCR
was performed using the iQ SYBR Green Supermix with 200nM primers in 10 mL final reaction mix. For the amplification, the following
protocol was used: 98
C for 3 min, 40 cycles of 95
C for 15 s and 60
C for 45 s. To test primer specificity, melting curves were per-
formed at the end of the 40 cycles of amplification. A delta cycle threshold (Ct) was calculated using the average Ct values across
technical triplicates, by subtracting the geometric mean of a control non-m6A gene (RPS28). To test the distribution of the mRNA
levels along the sucrose gradient, for each fraction, equal volume of extracted RNA was reverse transcribed to cDNA. The list of
primers used for qPCR is presented in Table S4.

Microscale Thermophoresis (MST)

10 mM of each recombinant DF paralog expressed in E. coli was labeled with RES-NHS dye using the Monolith Protein Labeling Kit
RED-NHS 2nd Generation (Amine Reactive) following the protocol instructions. Prior to each MST run, protein aggregation was
minimized by centrifuging the protein solutions at 20,820 3 g for 10 min. A constant concentration (50 nM) of each DF paralog
was mixed with increasing concentration of the m6A-modified RNA (rUrCrCrGrG/iN6Me-rA/rCrUrGrU, IDT) in MST buffer (50 mM
HEPES, 100 mM NaCl, 0.05% Tween-20, pH 7.4) and incubated for 30 min at room temperature. After incubation, samples were
loaded onto Premium Coated capillaries (NanoTemper Technologies) for MicroScale Thermophoresis (MST) acquisition. The MST
measurements were performed at room temperature using a fixed IR-laser power of 40% for 20 s per capillary, using an LED power
of 20% in a red laser equipped Monolith NT.115 instrument (NanoTemper Technologies) (HTSRC, Rockefeller University). The Mono-
Temper analysis software (NanoTemper Technologies) was used to determine KD values.

QUANTIFICATION AND STATISTICAL ANALYSIS

Protein-protein interactome analysis

Data related to Figure 2B were previously reported in a recent Bio-ID interactome study in which the DF proteins were analyzed
among 139 other proteins (Youn et al., 2018). This analysis was performed using two C-terminal labeling experiments for DF1,
two N-terminal and two C-terminal labeling-experiments for DF2 and DF3. ‘‘Table S1’’ of the Bio-ID interactome study was used
to select the number of spectra counts, the number of performed replicates, the level of confidence of each DF1, DF2 and DF3 in-
teractor. As determined in the study (Youn et al., 2018), only proteins with an average probability of interaction of 0.95 were consid-
ered as high-confidence interactors. In the Youn et al. (2018) study, 0.95 was considered as an optimal cut-off for high-confidence
interactors as it was well above the Bayesian 1% FDR estimation (corresponding to an AvgP of 0.91) and as decreasing the threshold
resulted in no worthwhile gains in BioGRID-curated interactions for the drop-in sensitivity. The list of top 25 interacting proteins is also
provided in Table S1 of the same Bio-ID interactome study. ‘‘Table S7’’ of the same Bio-ID interactome study (Youn et al., 2018) was
used to define the NMF analysis-based association of each DF protein presented in Figure 2C. The 14 subcellular compartments are:
1. RNP granule (cytoplasmic ribonucleoprotein granule), 2. chromosome, 3. ER (Endoplasmic reticulum), 4. RNP complex (ribonu-
cleoprotein complex), 5. nucleolus, 6. plasma membrane, 7. mitochondrion, 8. P-body (CCR4-NOT complex/P-body), 9. stress
granules, 10. cytoskeleton, 11. nucleus, 12. vesicles, 13. MOC (microtubule organizing center), 14. snRNPs (small nuclear ribonu-
cleoprotein complex/spliceosomal complex).

Analysis of the physiochemical properties of each DF paralog.

The disordered region score was calculated using D2P2 (Oates et al., 2013). The prion-like likelihood ratio was calculated
using PLAAC (Lancaster et al., 2014). The net charge per residue, and the probability of charged residues, were calculated in a sliding
window of 10 amino acids using CIDER (Holehouse et al., 2017). Hydrophobicity analysis was performed using ProtScale (ExPASy) in
a sliding window of 10 amino acids.

Reanalysis of publicly available PAR-CLIP data of DF1, DF2 and DF3 in HeLa cells
Data from the DF3 PAR-CLIP (GEO: GSE86214) were downloaded and used for analysis in this study. Reads processing was per-
formed using Flexbar (Dodt et al., 2012). First, the adaptor sequence reported on the GEO submission was removed using the
following set-up: flexbar -r SRR509926X.fastq -f i1.8 -t SRR509926X.noadapter–max-uncalled 1 -a adapters_par-clip.fasta–pre-
trim-phred 20 -n 20–min-read-length 1. After the adaptor removal, quality read analysis check was performed using Fastqc
(https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). According to the Fastqc results, 40% of the reads of each DF3

e5 Cell 181, 1582–1595.e1–e8, June 25, 2020

 ll
Article

PAR-CLIP library replicate was represented by the following sequence: CTCAACACCCACTACCTAAAAAATCCCAAACATATAACT-

GAACTCCTCAC. By using the BLAT analysis tool, we found that the sequence had 100% sequence identity to MT-RNR2, the mi-
tochondrially encoded 16S ribosomal RNA. This sequence is likely to represent a contaminant of the library preparation procedure.
The removal of this sequence and any other PCR duplicate was performed using the CIMS/fastq2collapse.pl script. Reads shorter
than 18 nucleotides were then removed by using Flexbar with the default–min-read-length option (Dodt et al., 2012). The remaining
reads were aligned to the human genome (H. sapiens, NCBI GRCh38) using Bowtie version 1 and the reported parameters in the GEO
submission data processing description (-v 2 -m 5).
Data from the DF1 and DF2 PAR-CLIP datasets were analyzed similarly. For DF1, the original PAR-CLIP sequencing data were
downloaded from GEO: GSE63591. For DF2, the original PAR-CLIP sequencing data were download from GEO: GSE49339. In
this case, when we performed the quality check using Fastqc, we did not find any over-represented sequences representing
more than 5% of the library. Thus, after removing low-quality nucleotides, the adaptor sequence, short reads, and duplicates, the
remaining reads were aligned to the human genome (H. sapiens, NCBI GRCh38) using Bowtie version 1 and the reported parameters
in the GEO submission data processing description (-v 2 -m 5).
To visualize the aligned reads and comparing them across the different PAR-CLIP datasets, we calculated the number of reads
counts normalized to the millions of mapped reads at every coordinate on the human genome. In this case, every single read
does not represent a unique event of RNA-protein interaction since libraries were prepared using a Truseq small RNA sample prep-
aration kit without the use of unique molecular identifiers. Therefore, PCR duplicates cannot be distinguished from distinct binding
events.

Comparison of the coverage of DF proteins at each m6A site on mRNAs throughout the transcriptome.
To compare iCLIP coverage at each m6A site, we calculated the number of normalized DF iCLIP read counts at each m6A site using a
previously described approach (Patil et al., 2016). We first calculated the number of unique read counts per million mapped reads
obtained from the CITS analysis of each iCLIP dataset (Patil et al., 2016). In this approach, each iCLIP read represents a unique
RNA-bound protein. Only m6A mapping to mRNAs were used in this analysis. A 5-nucleotide window from the m6A site genomic co-
ordinate was selected and the number of iCLIP reads per million uniquely mapped reads was calculated at every m6A coordinate
using BEDTools (Quinlan and Hall, 2010). The following formula was used: (r*106)/R where r = number of unique CLIP reads at the
m6A window, R = total number of uniquely mapped unique iCLIP reads in the whole iCLIP library. We refer to the resulting number
as ‘‘coverage at each m6A site’’ throughout the manuscript. For comparing the reanalyzed DF1 and DF2 PAR-CLIP coverage at each
m6A site, we used a similar approach. For each DF1, DF2 and DF3 iCLIP and miCLIP, the enriched motif presented in Figure 1 was
obtained using MEME suite (Bailey et al., 2009) as previously described (Patil et al., 2016).
For representation of iCLIP, PAR-CLIP and miCLIP tracks in Figure 1 and Figure S1, the normalized number of read counts from
iCLIP, PAR-CLIP and miCLIP datasets at the indicated genomic coordinate is presented. The reported annotated m6A sites in the
respective GEO submission were downloaded (GEO: GSE86336, GSE63753). In order to calculate the number of m6A sites per tran-
script, m6A sites were assigned to the gene transcript using MetaplotR (Olarerin-George and Jaffrey, 2017).

Calculation of coverage at each DF1 and DF2 unique sites

To select the previously described DF1-uniquely bound sites and DF2-uniquely bound sites, we first obtained the list of genes consid-
ered unique DF1 and DF2 targets from Shi et al., 2017. Next, the subset of called DF1 or DF2 sites on these targets were selected from
the entire list of called sites reported in supplementary file ‘‘GSE63591_A-Y1-PARCLIP.xlsx’’ (GEO: GSE63591), and in the supple-
mentary file ‘‘GSE49339_A-PARCLIP.xlsx.gz’’ (GEO: GSE49339). The number of DF1 and DF2 normalized coverage at each site was
calculated using deepTools bamCoverage (parameters:–binSize 50–effectiveGenomeSize–normalizeUsing BPM). To compare the
two calculated coverage, deepTools bamCompare was used and visualized using deepTools plotHeatmap (Ramı́rez et al., 2016).

Analysis of the ribosome profiling and RNA seq data upon silencing of DF proteins
After the sequencing of the ribosome profiling libraries performed in this study, reads were quality-based trimmed and reads below
16 nucleotides were excluded. The adaptor was removed using Flexbar (Dodt et al., 2012). The demultiplexing was performed based
on the experimental barcode using the pyBarcodeFilter.py script. The second part of the random barcode was then moved to the
read header. After removal of the PCR duplicates based on the UMI sequence, ribosomal RNA reads were removed using STAR
aligner (Dobin et al., 2013). As previously reported, a high percentage of reads in the library is represented by ribosomal RNA. To avoid
possible contamination of the reads representing the Ago-complex recruited on the siRNA target sequence, reads were also mapped
to the siRNA sequences to remove these sequences. siRNA sequences are previously described and reported in the GEO submis-
sion related to this previous study. Reads with no acceptable alignment to ribosomal and siRNA sequences were then mapped to the
hg38 transcriptome. The STAR genome index was built using the annotation obtained from GENCODE (version 26). The following
parameters were used to align to the genome: STAR–runThreadN 20–genomeDir ./star_hg38–readFilesIn reads_rRNA_siRNA_Un-
mapped.fastq–outSAMtype BAM SortedByCoordinate–sjdbGTFfile gencode.v26.protein_coding.annotation.gtf–outFilterMulti-
mapNmax 8–limitBAMsortRAM 10000000000–alignEndsType EndtoEnd–outWigType bedGraph–outWigStrand Stranded–
outWigNorm None–quantMode TranscriptomeSAM GeneCounts–outFilterMismatchNmax 8–outSAMattributes All–
outFilterIntronMotifs RemoveNoncanonicalUnannotated. These parameters were previously used (Calviello et al., 2016). The map-

Cell 181, 1582–1595.e1–e8, June 25, 2020 e6

 ll
Article

ped reads that represent ribosome-protected fragments (reads longer than 21 nucleotides, but shorter than 32 nucleotides) with a
40% minimum level of periodicity were considered for the subsequent analysis. We specifically considered only ribosome-protected
fragments reads mapped to the coding sequence to avoid any possible contamination coming from the untranslated area of the
genome. To measure the degree to which transcripts are translated independently from the initiation and termination rate, we
excluded reads mapping to the first 15 nucleotides from the start codon and the 9 nucleotides before the stop codon of each coding
sequence, as recommended in previous studies (Lauria et al., 2018). To precisely determine the localization of each ribosome with
respect to the start and stop codon, the P-site offset was estimated as previously described using riboWaltz (Lauria et al., 2018).
Reads matching all these criteria were considered to determine the number of ribosome-protected fragments per gene. Given the
level of variability and the possibility of contamination of the ribosome profiling library preparation, all these criteria have been re-
ported as essential analysis steps to better estimate the number of ribosome-protected fragments mapped to each transcript (Cal-
viello et al., 2016; McGlincy and Ingolia, 2017). Only the longest splice isoform of each gene was considered. Genes with less than 10
ribosome-protected fragments were excluded from the analysis. For each DF paralog silencing experiment, TMM normalization,
empirical Bayes estimate of the negative binomial dispersion, and measurement of the level of change in translation (log2 Fold
Change) compared to the control condition was performed using edgeR (McCarthy et al., 2012). To account for sample-to-sample
variability, all replicates were analyzed at the same time to define a unique log2 Fold Change measurement.
After the sequencing of the RNA-seq libraries, reads with low quality were discarded and read lengths shorter than 18 nt were dis-
carded. Ribosomal reads were removed using STAR aligner. The remaining reads were mapped to hg38 genome using STAR and the
data were used to normalize the Riboseq data and derive the change in RNaseq expression. For each DF paralog silencing, TMM
normalization, empirical Bayes estimate of the negative binomial dispersion, and measurement of the level of change in translation
(log2 Fold Change) compared to the control condition was performed using edgeR. To account for sample to sample variability, all
replicates were analyzed at the same time to define a unique log2 Fold Change measurement.
For the translational efficiency measurement, the Riboseq change was compared to the RNaseq change.

Reanalysis of publicly available ribosome profiling data upon silencing of DF1 in HeLa cells
To analyze the public available ribosome profiling data upon DF1 silencing, we used two independent approaches. First, we down-
loaded the Table ‘‘GSE63591_C-Y1-Ribosome_profiling.xlsx’’ available at the GEO: GSE63591 and assigned to each reported tran-
script a number of m6A sites. To define whether m6A mRNAs were affected by the DF1 silencing, we used the translational efficiency
values reported in the table without performing any further processing step. Secondarily, the original sequencing data of the GEO:
GSE63591were downloaded and processed as follow. Quality check was performed using Flexbar. After quality check, sequences
longer than 21 nt were first mapped to the siRNA sequences reported in the published study (Wang et al., 2015), to the ribosomal
RNA, and then to the hg38 transcriptome. All mapping steps were performed using STAR as mentioned above. Reads mapping
to the coding sequence that represent ribosome protected fragments and have high level of periodicity were considered for the anal-
ysis steps as described above. We excluded reads mapping to the first 5 codons from the start codon and the 3 codons before the
stop codon of each coding sequence as described above.
We noticed a major difference between the original public available table and the new tables generated by our reanalysis. In the
case of Replicate 2, the number of ribosome-protected fragments assigned to the DF1 mRNA in the processed table by Wang et al.
(2015) showed a significant reduction in the number of ribosome-protected fragments for the DF1-silenced sample compared to the
matching control sample. This is expected for the DF1 knockdown condition. When we reanalyzed the original sequences reads from
GEO: GSE63591, we reproduced this loss of ribosome-protected fragments assigned to DF1 upon DF1 silencing. In contrast, the
number of ribosome-protected fragments assigned to the DF1 mRNA in Replicate 1, as shown in the processed table
‘‘GSE63591_C-Y1-Ribosome_profiling.xlsx,’’ shows a substantially higher number of ribosome-protected fragments for the DF1-
silenced sample compared to the matching control sample (15-fold increase). This would indicate a condition of overexpression
instead of knockdown. We therefore examined the raw sequencing reads for Replicate 1 to generate a new table of ribosome-pro-
tected fragments. Upon reanalysis of the Replicate 1 ribosome-profiling data (Table S2), the number of ribosome-protected frag-
ments mapped to DF1 mRNA was significantly lower than the number indicated in the published processed table ‘‘GSE63591_C-
Y1-Ribosome_profiling.xlsx.’’ In the reanalysis, we observed 98% less ribosome-protected fragments mapped to DF1 than the num-
ber of ribosome-protected fragments mapped to DF1 in the matching control sample. This is consistent with the raw data represent-
ing a DF1 knockdown condition rather than an overexpression condition.
We used a standard protocol for determining ribosome-protected reads, which inherently discards any reads that derive from the
siRNA itself because these fragments would be too short to be a ribosome-protected fragment, as described above (Calviello et al.,
2016; McGlincy and Ingolia, 2017). It is possible that the large number of ribosome-protected fragments that map to DF1 in the pub-
lished processed table for Replicate 1 derive from DF1-specific siRNA molecules that were cloned into the library but not removed, if
there was no appropriate size filter used in the alignment protocol. The length, sequence and alignment position of each ribosome
protected fragment is not provided in the processed table. Alternatively, the ribosome-protected reads assigned to DF1 could derive
from a sample which contains overexpressed DF1 mRNA. Because of this uncertainty, we generated new DF1 knockdown ribosome
profiling datasets in the same cell line. The pipeline used for the reanalysis of these datasets is reported (see ‘‘Data and Code
Availability’’).

e7 Cell 181, 1582–1595.e1–e8, June 25, 2020

 ll
Article

Analysis of m6A mRNA stability upon DF paralog depletion.

The stability of m6A-modified mRNAs and non-methylated mRNAs was determined by quantifying mRNA levels immediately before
and 2 h after actinomycin D treatment. After sequencing, libraries were checked for their quality as previously described for all RNA-
seq libraries in this study. Reads were then mapped to the transcriptome using STAR aligner. A read count table was generated using
the STAR aligner as well. ERCC Spike-in reads were also mapped using STAR aligner. Thus, the resulting read counts table contains
reads mapping to the ERCC RNA Spike-in as well. To remove unwanted variation, the RUVg estimation package was used (Risso
et al., 2014). Additionally, edgeR was used to calculate ERCC spike-in RNA size factors for each sample, which were then applied
to normalize for library size changes in each replicate. We thus obtained a final number of normalized reads mapped to each mRNA.
Since m6A mRNAs are known to have a short half-life (2 to 4 h) (Geula et al., 2015; Ke et al., 2017; Schwartz et al., 2014), m6A mRNA
expressions are often substantially reduced even at 2 h after actinomycin D treatment. We focused on mRNAs in which the reduction
in expression level could be accurately measured, i.e., those mRNAs that showed a reduction in expression level by at least 40% after
2 h of actinomycin D treatment. The percentage of mRNA abundance reduction was calculated in both the control samples (control
siRNA) and in the samples subjected to DF paralog knockdown. In some cases, the stabilization of the mRNA was very substantial
upon depletion of all three DF paralogs, and the mRNA expression reduction was less than 10%. Since these mRNAs no longer show
a drop in mRNA expression, it is difficult to precisely establish the fold change in mRNA stability. Therefore, we simply assigned the
reduction at 10% for these mRNAs in the calculations used in Figure 3F.

Cell 181, 1582–1595.e1–e8, June 25, 2020 e8

 ll
Article

Supplemental Figures

Figure S1. DF Paralogs Bind the Same m6A Sites throughout the Transcriptome, Related to Figure 1
(A) The YTH domain of DF1, DF2 and DF3 exhibit high sequence homology. Shown is a detailed representation of the aligned amino acid sequence for the YTH
domains of DF1, DF2 and DF3. Amino acids previously described to be essential for the m6A binding based on the DF1- m6A RNA and DF2- m6A RNA crystal
structures (Li et al., 2014; Xu et al., 2015) (PDB: 4RCJ, 4RDN) are highlighted. Pink indicates the three tryptophan residues that form the aromatic cage
(legend continued on next page)
 ll
Article

surrounding m6A. Green and Blue indicate amino acids that make additional points of contact between the YTH domain and the nucleotides adjacent to m6A. Of
note, each YTH-DF domain shares each of these amino acids. Shown below in a yellow color code scheme is the level of conservation of every amino acid among
the three YTHDF proteins with a range from 1 (low conservation) to 10 (high conservation) as predicted by the Clustal Omega alignment algorithm (Madeira et al.,
2019). Most residues (88%) are fully conserved residues across all DFs paralogs as indicated by the conservation score of 10. Importantly, amino acids essential
for recognizing m6A and adjacent residues in RNA are fully conserved. As indicated in Figure 1A, amino acids with a conservation score lower than 10 are located
on the opposite side of the YTH domain away from the RNA-binding pocket, suggesting that they might not affect the YTH-m6A RNA interaction.
(B) DF1, DF2 and DF3 have similar binding affinity to m6A-modified mRNA. Previous studies suggest that DF1, DF2 and DF3 bind with similar affinity to m6A-
modified mRNA (Patil et al., 2018; Wang et al., 2014a; Xu et al., 2015). To compare the affinities side-by-side, full length DF1, DF2 and DF3 were prepared as
recombinant proteins in E. coli and the affinity to an m6A-modified RNA was measured by microscale thermophoresis (MST). Shown is the percentage of protein
bound (fraction bound) at increasing concentration of the m6A-RNA. These data show that the three DFs have similar in vitro binding affinity to m6A. Data are
represented as mean ± s.d. (n = 2 replicates).
(C) The previously reported (Shi et al., 2017) level of overlap of the DF1, DF2, and DF3 targets is summarized. In these previous studies, the targets of each DF
paralog were first identified by PAR-CLIP (Shi et al., 2017; Wang et al., 2014a, 2015). Then, to determine which of these mRNAs represent valid DF-mRNA in-
teractions, RIP (immunoprecipitation of the DF paralog, followed by reverse transcription and sequencing of the bound RNA) was performed. An overlap
approach was used to identify targets of each DF paralog, and subsequently to define DF1, DF2 or DF3 unique targets or common targets. This analysis resulted
in the conclusion that only a minority of mRNAs (23.98%) are bound by all DF paralogs, as shown by the pie chart. Most of the mRNAs were annotated to be either
unique targets of one DF or shared by two DFs. The finding that different DF proteins bind distinct subsets of m6A-containing mRNAs in the transcriptome has
been the foundation of the current model of how DFs differentially regulate different m6A mRNA cohorts (Han et al., 2019; Liu et al., 2018; Park et al., 2019; Shi
et al., 2017, 2018, 2019).
(D) Read quality analysis of the previously performed DF1, DF2 and DF3 PAR-CLIP datasets. For each PAR-CLIP dataset (Shi et al., 2017; Wang et al., 2014a,
2015), reads are classified based on the following quality parameters: read length (yellow, whether a read is shorter than 18 nucleotides, and thus cannot be
aligned to the genome), presence of PCR duplicates (blue, duplicates category), and mappability to the genome using Bowtie and the previously reported
parameters (categorized into ‘‘mappable’’ (green) or ‘‘fail to map’’ (light blue) reads). The DF3 PAR-CLIP dataset was unusual for several reasons. First, 44.11% of
all reads mapped to a single site in a single gene, MT-RNR2 (STAR Methods). These reads were identical, consistent with a PCR duplication event and were not
found in the DF1 or DF2 PAR-CLIP datasets or any miCLIP datasets. Second, of the remaining reads, only a small percentage of these were mappable (indicated
in green in the 10 3 10 dot plot). Overall, the DF3 PAR-CLIP dataset lacks sufficient read depth to detect endogenous DF3-binding sites in the transcriptome as
efficiently as the other PAR-CLIP datasets.
(E) Schematic representation of the method used to identify m6A sites bound by one DF or shared by two or more DFs. Scatterplots are used to represent the
pairwise comparison of the number of normalized DF reads calculated from the iCLIP or PAR-CLIP datasets at each m6A site. Yellow areas indicate where m6A
sites that uniquely bind one DF should be located. For instance, DF1-unique sites with a high level of coverage for DF1 (high number on x axis) and a low level of
coverage for DF2 (low number on the y axis) will be shown in the lower right area of the scatterplot. m6A sites bound by two or more DFs with a similar level of
binding for each DF will be found in the indicated light red area.
(F) Essentially all m6A sites show highly correlated binding of DF1 and DF2 based on previous PAR-CLIP datasets (Shi et al., 2017; Wang et al., 2014a, 2015). As
described in (E), to identify m6A sites that preferentially bind DF1 or DF2, their level of binding was quantified at each m6A site in the transcriptome (Ke et al., 2017).
In this experiment, we used the previously reported PAR-CLIP datasets prepared in HeLa cells (Shi et al., 2017; Wang et al., 2014a, 2015). According to the
previous analysis (Shi et al., 2017; Wang et al., 2014a, 2015), nearly 50% of m6A sites should be uniquely bound by one DF. However, few if any m6A sites show
evidence of disproportionate binding of one DF paralog compared to the other. Additionally, the high Pearson correlation coefficients (r) show that DF1 and DF2
have highly similar binding preferences for each m6A site in the transcriptome of HeLa cells. Similar results are shown in Figure 1C using iCLIP data for DF1, DF2
and DF3 in HEK293T cells (Linder et al., 2015; Patil et al., 2016). Thus, the binding of all m6A sites to each DF paralog is not cell-type specific, and not specific to
the type of CLIP assay used (PAR-CLIP versus iCLIP).
(G, H) DF1-unique sites show extensive levels of bound DF2, and vice versa. Previous studies show that nearly half of all m6A sites are uniquely bound by only one
DF paralog (Shi et al., 2017; Wang et al., 2014a, 2015. To confirm whether DF1-unique mRNAs are indeed uniquely bound by DF1, we asked if there are markedly
fewer DF2 PAR-CLIP reads than DF1 PAR-CLIP reads located at DF1-unique sites. For this analysis, we selected the previously described DF1-uniquely bound
sites and DF2-uniquely bound sites (Shi et al., 2017; Wang et al., 2014a, 2015). Shown is the DF1 and DF2 PAR-CLIP coverage at DF1-unique sites (G) and at DF2-
unique sites (H). For each site, PAR-CLIP DF1 or DF2 coverage is shown, with each row representing the coverage for a different unique site and its surrounding
area. The rows are ordered based on the degree of DF1 PAR-CLIP coverage (G) or DF2 PAR-CLIP coverage (H). In principle, DF1-unique sites should show higher
level of DF1 PAR-CLIP coverage than DF2. However, the PAR-CLIP coverage for both DF1 and DF2 appears largely identical at both DF1-unique sites and DF2-
unique sites. We were unable to define the DF3 coverage level at each site given the low number of mappable reads of the DF3 PAR-CLIP dataset, as discussed in
Figure S1C.
(J) DF1 and DF2 PAR-CLIP read tracks are highly similar even on transcripts thought to be DF1-unique or DF2-unique (Shi et al., 2017; Wang et al., 2014a, 2015).
DF1-unique and DF2-unique transcripts should show unique peaks or patterns of DF1 and DF2 binding. However, we found that transcripts generally appear to
show identical PAR-CLIP read coverage along the entire length of the transcript body. When inspecting different mRNAs, differences in PAR-CLIP coverage were
only seen if read coverage was low and therefore susceptible to read noise. Shown is the normalized read distribution of DF1 PAR-CLIP (in blue) and DF2 PAR-
CLIP (in green) on OGT, an mRNA previously thought to be a DF1-unique target, and on DUSP1, an mRNA previously thought to be a DF2-unique target. The
‘‘called’’ DF peaks were previously reported by Wang et al. (2015) and Wang et al. (2014a) based on a statistical peak finding algorithm PARalyzerv1.1 and are
presented below their respective tracks. Each row represents called peaks in a different PAR-CLIP replicate. Notably, the location of peaks and the relative
heights of each peak is similar for DF1 and DF2 on both transcripts. All the peaks overlap substantially with called m6A sites in HeLa cells (Ke et al., 2017). In the
original DF1, DF2, and DF3 mapping studies, the PAR-CLIP called sites were overlapped with a list of target mRNAs immunoprecipitated using the RIP protocol
(Shi et al., 2017; Wang et al., 2014a, 2015). Compared to the total number of m6A-containing mRNAs or the total number of mRNAs containing DF1 or DF2 PAR-
CLIP peaks, only a few mRNAs were successfully immunoprecipitated using this method (1747 for DF1, 1592 for DF2, 2080 for DF3). Therefore, the use of this
approach makes the final list of DF-specific mRNA targets highly dependent on this low efficiency mRNA immunoprecipitation method.
 ll
Article

(legend on next page)

 ll
Article

Figure S2. DF Proteins Have Similar Effector Domains and Protein-Protein Interactions, Related to Figure 2
(A) DF1, DF2 and DF3 exhibit generally high sequence identity and similarity in their effector domain. In Figure S1, we examined the high sequence identity of all
DFs in their YTH domain. Shown here is a schematic representation of the aligned amino acid sequence of the remaining part of the DF1, DF2 and DF3 protein.
There was overall similarity along the entire length of the effector domain, with small regions where there were amino acid differences. These small differences
may account for the previously described differences of DF function, which are examined in this study. Residues are classified according to the ClustalW Multiple
Sequence Alignment score (Madeira et al., 2019) based on whether the residue is exactly identical among all DFs, or ‘‘strongly and weakly similar’’ residues (the
amino acids share higher or lower level of similarity in their physicochemical properties among all DFs). Shown in the table is the overall percentage of sequence
identity and similarity calculated by the pairwise sequence EMBOSS Needle aligner (Madeira et al., 2019). The pairwise comparisons of the DF amino acid
sequences confirms the high sequence similarity and identity of all DFs.
(B) High-confidence interactors of each DF paralog generally show high-confidence interactions with the other DF paralogs. As described in Figure 2B, high-
confidence interactors of each DF paralog in vivo were identified previously using a Bio-ID proteomic approach (Youn et al., 2018). The top 25 interacting proteins
for each DF was previously determined based on their length-normalized spectral counts and the average probability of interaction calculated across different
replicates (Youn et al., 2018). The heatmap show the AvgP, average probability of interaction, for each of the top 25 interactors of DF1, DF2 and DF3. The majority
of the top 25 interactors have a high average probability of interaction with all three paralogs. These proteins include the main components of the CCR4-NOT
RNA-degradation complex and proteins previously identified as stress granule components. A recent study showed that all three DF proteins bind CNOT1 and all
three can elicit mRNA deadenylation when tethered in cells to a reporter mRNA (Du et al., 2016), supporting the overall finding that the major function of the DF
proteins is to mediate degradation of m6A-mRNAs. Interestingly, interaction with components, or repressor of the translation process may suggest a possible
involvement of DF proteins in the translational repression of m6A mRNAs. This repression may be then interconnected to the observed DF-dependent regulation
of m6A-mRNA stability. Shown at the bottom is the AvgP of interaction of each DF with the other DF paralogs.
(C) The Bio-ID DF1 interactome data (Youn et al., 2018) did not identify eIF3A and 3B as high confidence DF1 interactors. However, in Wang et al., 2015, eIF3A and
eIF3B were identified in FLAG-tagged DF1 immunoprecipitates by mass spectrometry (Wang et al., 2015). Shown is the reported score and number of unique
peptides mapped to each putative DF1 interactor from Wang et al. (2015). However, even in this study, eIF3A and eIF3B are not among the top DF1 interactors
when ranking all the interactors by the reported score. Moreover, as shown in red, the number of unique peptides is less than 5. This low level of interaction is
consistent with the Bio-ID interactome studies presented in Figure 2B. Thus, eIF3A and eIF3B are not seen as DF1 interactors in the Bio-ID study (Youn et al.,
2018) and are weak interactors in the Wang et al. (2015) study.
(D) Heterologous expression of DF proteins causes formation of DF stress granule-like structures. Plasmids expressing FLAG-3xHA tagged DF proteins, or a
control FLAG-3xHA construct were transfected in HeLa cells and staining was performed after 24 h of transfection. Areas where the HA staining (green) overlaps
with the DF staining (red) indicate the FLAG-3xHA tagged DF localization (yellow). As indicated by the arrows, the FLAG-3xHA tagged protein formed granule-like
structures of different dimension in some, but not all cells. This phenomenon occurs with proteins with low-complexity domains (Alberti et al., 2019). As shown by
the absence of these structures in the control-transfected cells (FLAG-3xHA), the granule-like structures are not an artifact of transfection. Thus, expressing DF
proteins may lead to uninterpretable results due to the variable formation of protein aggregates that could act to sequester DF proteins and mask their function.
Scale bar, 20 mm.
 ll
Article

(legend on next page)

 ll
Article

Figure S3. Analysis and Validation of RNA-Seq Data Obtained upon Single, Double, and Triple DF Protein Silencing in HeLa Cells, Related to
Figure 3
(A) Validation of DF1, DF2 and DF3 knockdown upon silencing of each of the DF paralogs alone or together. Western blotting was used to confirm knockdown
using the indicated DF paralog-specific antibodies on each sample used to perform RNA seq and Ribosome profiling analysis. HeLa cells were transfected with
the same total concentration (30 nM) of siRNA in each condition. Western blotting was performed 4 days after transfection. In each silencing condition, the siRNA
sequences were specific to the DF paralog of interest, as seen by the selective knockdown seen with each siRNA. GAPDH was used as loading control.
(B) Relative quantification of DF1, DF2 and DF3 protein band intensity shown in A. western blot chemiluminescent signal of every band was measured using
ImageLab Image tools and normalized to the GAPDH loading control. Knockdown of any of the DF paralogs was associated with a compensatory increase in the
expression of the other paralogs. The number of dots indicates the number of Western Blot replicates for each condition. Error bars indicate standard deviation.
(C) Reproducibility of the RNA-seq library replicates. For each silencing condition tested (DF1, DF2, DF3 and triple silencing), three independent replicates were
performed. The Pearson correlation coefficient of the normalized number of mapped reads across replicates was calculated and presented in each heatmap.
(D) DF1, DF2 and DF3 RNA and protein levels in HeLa cells measured by RNA-seq and Ribo-seq. The normalized counts of reads mapped to each paralog after
RNaseq (mRNA) and ribosome profiling (RPFs, ribosome protected fragments) were used as a proxy of the mRNA expression and protein expression levels,
respectively. As shown by the heatmap, DF1 has mRNA levels similar to DF2. However, DF2 is the most highly translated DF paralog. This will result in a higher
amount of DF2 protein in HeLa cells. Thus, at least in HeLa cells, DF2 is likely to be the most abundant DF paralog.
(E) Validation of the simultaneous knockdown of DF1 and DF2, DF1 and DF3, DF2 and DF3 four days after siRNA transfection. Cells were transfected with siRNA
on Day 1 and Day 3, and then western blotted to detect levels of DF1, DF2, or DF3 on Day 4. Knockdown of two DF paralogs was associated with a modest
compensatory increase in the expression of the remaining DF paralog. GAPDH was used as loading control.
(F-H) Cumulative distribution plots related to Figure 3E. In Figures 3A–3D, the cumulative distribution plots were shown only for the single knockdown experiments
and the triple knockdown experiment. Here we show the cumulative distribution blots for the double knockdown experiments. The abundance of each mRNA
(based on RNA-seq counts) was compared between the control and the indicated double DF silenced condition. mRNAs were binned based on the number of
annotated m6A sites. m6A mRNAs show higher expression levels compared to nonmethylated mRNA upon silencing any two DF paralogs. In F-H, each m6A
binned group was compared to the non-methylated mRNA group using a two-tailed Mann–Whitney test. Only significant p values are shown in each graph and
the exact p values are reported. n = 3 biological replicates.
(I) Reproducibility of the RNA-seq library replicates. For each silencing condition tested (DF1-DF2, DF1-DF3, DF2-DF3), three independent replicates were
performed. The Pearson correlation coefficient of the normalized number of reads across replicates was calculated and presented in each heatmap.
(J) RT-qPCR validation of the increase in stability of m6A mRNAs upon the silencing of the three paralogs. Highly methylated mRNAs (ZNF503 – 4 annotated m6A
sites, SGK1 – 4 annotated m6A sites, ID3 – 7 annotated m6A sites) were selected for validation. Transcripts lacking annotated m6A sites (PP1E3C and RPS28)
were chosen as controls. The stability of m6A-modified mRNAs and non-methylated mRNAs was determined by quantifying by RT-qPCR the mRNA levels
immediately before (0 h), 1 h and 2 h after actinomycin D treatment. The amount of mRNA detected at each time point is shown as percentage of the total mRNA
amount quantified at time 0 and normalized on the RPS28 abundance. The increase in mRNA stability is most apparent when all three DF paralogs were knocked
down. (n = 2 biological replicates ± s.d.).
 ll
Article

(legend on next page)

 ll
Article

Figure S4. Reanalysis of Previously Performed Ribosome Profiling Data Shows that DF1 and DF3 Do Not Control the Translational Efficiency
of m6A-mRNAs, Related to Figure 4
(A) Unlike PABPC1, DF1 is not enriched in the polysomal fraction. DF1 has been proposed to enhance translation by facilitating loading of eIF3 to mRNA 50 UTRs.
In this model, DF1 binds to an mRNA 30 UTR, and binds eIF3. This can create an mRNA loop in which the 30 UTR and 50 UTR are connected via the DF1-eIF3
complex. This is similar to other PABPC1, which also binds to mRNA 30 UTRs and binds a translation initiation factor, which results in mRNA looping. PABPC1
function in enhancing mRNA translation is evidenced by its enrichment in the sucrose fractions occupied by mRNAs containing high levels of actively translating
ribosomes (polysomes). We therefore tested if DF1 is also enriched in the polysome fractions by treating HeLa cells with cycloheximide and fractionating
polysomes on a sucrose gradient. The RNA fractions indicated here are related to the UV trace as shown in Figure 4A. DF1 is not enriched in these fractions based
on paralog-specific western blotting (also see Figure 4A). In contrast, PABPC1 is readily detected in the actively translating ribosomal fractions. The ribosomal
protein RPS6 was used to confirm the efficiency in recovering the translating ribosomes from the polysomal sucrose fractions. Thus, these data do not support a
model in which DF1 enhances the translation of m6A-mRNAs by binding to highly translated mRNAs.
(B) m6A-mRNAs are downregulated upon depletion of DF1, based on ribosome profiling datasets provided in Wang et al. (2015). Here, we used two ‘‘processed’’
datasets provided in the GEO submission for Wang et al. (2015) (GSE63591_C-Y1-Ribosome_profiling.xlsx). These datasets are referred to as processed since
the number of ribosome-protected fragments, RNA-Seq data, and translational efficiency (TE) are listed for each transcript based on the authors’ calculations and
analysis of their next-generation sequencing data. We used this processed data to calculate the change in translational efficiency of mRNAs upon DF1 silencing
based on the number of annotated m6A sites. As can be seen, m6A-modified mRNAs show reduced translation upon DF1 knockdown. This result is essentially
identical to the cumulative distribution plots presented in Wang et al. (2015). Notably, the data presented here is based on the average of two replicates provided
by Wang et al. (2015), designated as Replicate 1 and Replicate 2.
(C) Replicate 1 shows a more prominent reduction in m6A-mRNA translation upon depletion of DF1. The effect seen in Replicate 1 is more pronounced than the
effect seen in (A), which is the average of Replicate 1 and Replicate 2. Only significant p values are shown in each graph and the exact p values are reported.
(D) Replicate 2 shows no reduction in m6A-mRNA translation efficiency upon depletion of DF1. Surprisingly, and in contrast to Replicate 1 (shown in (B), Replicate
2 does not show a drop in m6A mRNA translation efficiency upon depletion of DF1. The effect seen in (A) was seen since it was the average of Replicate 1 and
Replicate 2, which show different effects on m6A mRNA. In B-D, each m6A group was compared to the non-methylated mRNA group using a two-tailed Mann–
Whitney test. Only significant p values are shown in each graph and the exact p values are reported.
(E) The reanalyzed ribosome profiling datasets do not show reduced m6A-mRNA translation efficiency upon depletion of DF1. In this experiment, we accessed the
original next-generation sequencing data generated by Wang et al. (2015) (GEO: GSE63591) and generated a new table of ribosome-protected fragments using a
standard ribosome-profiling analysis pipeline (McGlincy and Ingolia, 2017). We referred to these reanalyzed datasets ‘‘Replicate R1’’ and ‘‘Replicate R2.’’ mRNAs
were binned based on the number of annotated m6A sites. In contrast to the results presented in (B), analysis of the two replicates (presented as an average)
shows that m6A mRNAs do not show reduced translation efficiency upon DF1 silencing.
(F) The reanalyzed ribosome profiling sample, i.e., Replicate R1, no longer shows a link between m6A mRNA translation and DF1. In contrast to the original
processed data presented in (C), reanalysis of the underlying next-generation sequencing data from the ribosome profiling data in Wang et al. (2015), no longer
shows decreased m6A-mRNA translation upon depletion of DF1.
(G) Replicate R2, like the original Replicate 2 shown in (D), shows no change in m6A-mRNA translation efficiency upon depletion of DF1. In E-G, each m6A binned
group was compared to the non-methylated mRNA group using a two-tailed Mann–Whitney test. Only significant p values are shown in each graph and the exact
p values are reported.
(H) The number of ribosome-protected fragments per mRNA reported by Wang et al. (2015) for Replicate 1 only modestly correlates to the number of ribosome-
protected fragments of Replicate 2. Both the control replicates, and the DF1-depleted replicates were compared. For this analysis, we used the number of
ribosome-protected fragments reported for the two replicates in the GEO submission for Wang et al. (2015) (GSE63591_C-Y1-Ribosome_profiling.xlsx). These
datasets are referred to as ‘‘processed’’ since the number of ribosome-protected fragments are listed based on the authors’ analysis of their next-generation
sequencing raw data. For each tested condition (transfection with a control siRNA or with siRNAs targeting DF1), a plot showing the comparison between the
number of ribosome-protected fragments calculated for Replicate1 and Replicate 2 is presented. A Pearson correlation was calculated for each comparison. To
reduce the complexity of the analysis, only the ribosome-protected fragments mapped to the m6A-mRNAs are presented.
(I) Upon our re-examination of the raw data reported by Wang et al. (2015), the number of ribosome-protected fragments per mRNA in Replicate 1 strongly
correlates with the number of ribosome-protected fragments of Replicate 2. This analysis was performed for both the control and DF1-depleted samples. Here,
we accessed the original next-generation sequencing data generated by Wang et al. (2015) (GEO: GSE63591) and generated a new table of ribosome-protected
fragments per gene using a standard ribosome-profiling analysis pipeline (McGlincy and Ingolia, 2017). We referred to these as ‘‘Reanalysis of the data from Wang
et al. (2015).’’ As in (H), a plot showing the comparison between the number of ribosome-protected fragments calculated for Replicate1 and Replicate 2 is
presented for each tested condition. In contrast to (H), there is high level of correlation between replicates independently from the tested condition. This result is
generally seen when analyzing replicates of the same condition giving higher level of confidence in interpreting the possible effect of DF1 on the translation of
m6A genes.
(J) m6A-mRNAs do not show a reduction in translation efficiency upon depletion of DF3. We asked if m6A-mRNA translation is affected upon depletion of DF3,
using the ribosome profiling datasets generated by Shi et al. (2017). Unlike in the DF1 knockdown dataset, the processed data provided by Shi et al. (2017)
presented the ribosome-protected fragments from two replicates as a log2-fold-change relative to a single control ribosome profiling experiment. mRNAs were
binned based on the number of annotated m6A sites. Here, we saw no change in translation of m6A-annotated mRNAs using the two knockdown replicates
provided in Shi et al. (2017).
 ll
Article

(legend on next page)

 ll
Article

Figure S5. Quality Control Analysis of Ribosome Profiling Datasets and Validation of DF Paralog Silencing in MOLM-13 Cells, Related to
Figures 4 and 5
(A) Reproducibility of the ribosome profiling library replicates. For each silencing condition tested (DF1, DF2, DF3), three independent replicates were performed.
The Pearson correlation coefficient of the normalized number of ribosome-protected fragments reads across replicates was calculated and presented in each
heatmap.
(B) Validation of ribosome profiling datasets. Distribution of read length in the ribosome profiling experiments in each tested condition (DF1, DF2 and DF3
silencing). Most of the reads are 29-30 nt in length, which reflects the expected size of the ribosome footprint when a ribosome is translating on the mRNA and the
A site is occupied (McGlincy and Ingolia, 2017; Wu et al., 2019). The samples were prepared in the absence of cycloheximide as recommended in the most current
ribosome profiling protocols (McGlincy and Ingolia, 2017). The percentage of P-sites assigned to the coding sequence, 30 UTR, or the 50 UTR is presented as a bar
plot for each condition. As expected, most of the reads map to the coding region. (lower right) The position of ribosome footprints relative to the reading frame was
determined for each dataset. Overall, these analyses indicate that a high fraction of the reads obtained in these datasets represent ribosome-protected fragments
that reflect the position of the translating ribosomes in cells.
(C) Validation of the DF1, DF2 and DF3 knockdown. To confirm knockdown of the DF paralogs, we quantified the number of ribosome-protected fragments (RPFs)
that mapped to each transcript before and after the silencing of each DF paralog. Upon the silencing, the reduction in the RPFs confirms that DF1, DF2 and DF3
are less translated, consistent with efficient knockdown. These results are consistent with the western blot validation presented in Figures S3A and S3B.
(D) Effect of DF1 depletion on the distribution of m6A-modified mRNAs (OSGIN2, CDC27 and YY1) and non-m6A-modified mRNA (RPS28) along the sucrose
gradient. To independently test whether m6A-mRNAs are less translated, we performed polysome fractionation on a sucrose gradient of DF1-silenced and
control siRNA-treated cytoplasmic lysates (upper graph). The level of m6A-mRNAs reported to be exhibit markedly reduced in translation upon DF1 silencing by
(YY1 and OSGIN2, 5.6 log2 fold change; CDC27, 3.66 log2 fold change; Wang et al., 2015), were tested in each fraction by qRT-PCR. The quantity of each
mRNA per fraction was normalized to the amount of spike-in luciferase mRNA in each fraction and presented as percentage of the total amount measured in all
fractions. As shown in the lower graphs, DF1 silencing does not affect the distribution of m6A-mRNAs along the gradient. Thus, the effect previously shown by the
ribosome profiling for the top-regulated m6A-mRNAs cannot be independently validated using polysome fractionation and qPCR analysis. Data are means ±
s.e.m. (n = 2)
(E) Reproducibility of the DF1, DF2, and DF3 triple knockdown ribosome profiling dataset. As shown in (A), the Pearson correlation coefficient of the normalized
number of ribosome-protected fragments across replicates was calculated and presented in a heatmap.
(F) Validation of the ribosome profiling dataset obtained upon depletion of DF1, DF2, and DF3. As presented in (B), shown are the distribution of the ribosome
profiling read lengths, the percentage of P-sites assigned to the coding sequence, 30 UTR, and 50 UTR, and the position of ribosome footprints relative to the
reading frame. Overall, these analyses indicate that a high fraction of the reads obtained in this dataset represent ribosome-protected fragments that reflect the
position of the translating ribosomes in cells.
(G) Confirmation of DF1, DF2 and DF3 knockdown in MOLM-13 cells. Western Blot was performed after 5 days from the shRNA transduction. GAPDH was used
as loading control.
 Article

Metabolic Fingerprinting Links Oncogenic PIK3CA

with Enhanced Arachidonic Acid-Derived
Eicosanoids
Graphical Abstract Authors
Nikos Koundouros, Evdoxia Karali,
Aurelien Tripp, ..., Robert C. Glen,
Zoltan Takats, George Poulogiannis

Correspondence
z.takats@imperial.ac.uk (Z.T.),
george.poulogiannis@icr.ac.uk (G.P.)

In Brief
Metabolic fingerprinting using the iKnife
offers near real-time diagnosis of PIK3CA
mutant breast cancers and connects
oncogenic PIK3CA with enhanced
arachidonic acid metabolism. cPLA2
inhibition shows remarkable synergy with
dietary fat restriction to restore tumoral
immune cell infiltration and inhibit growth
of mutant PIK3CA-bearing breast tumors.

Highlights
d The iKnife offers near real-time diagnosis of PIK3CA mutant
breast cancers

d Oncogenic PIK3CA promotes enhanced arachidonic acid via

mTORC2-PKCz-cPLA2 signaling

d Mutant PIK3CA regulates proliferation beyond a cell

autonomous manner

d cPLA2 inhibition and dietary fat restriction suppress

PIK3CA-induced tumorigenicity

Koundouros et al., 2020, Cell 181, 1596–1611

June 25, 2020 ª 2020 The Author(s). Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.053 ll
 ll
OPEN ACCESS

Article
Metabolic Fingerprinting Links Oncogenic PIK3CA
with Enhanced Arachidonic Acid-Derived Eicosanoids
Nikos Koundouros,1,2 Evdoxia Karali,1 Aurelien Tripp,1 Adamo Valle,1,3,4,5 Paolo Inglese,2 Nicholas J.S. Perry,1
David J. Magee,1,6 Sara Anjomani Virmouni,1,7 George A. Elder,1,8 Adam L. Tyson,9,10 Maria Luisa Dória,2
Antoinette van Weverwijk,11,12 Renata F. Soares,2 Clare M. Isacke,11 Jeremy K. Nicholson,2,13 Robert C. Glen,2,14
Zoltan Takats,2,* and George Poulogiannis1,2,15,*
1Signalling and Cancer Metabolism Team, Division of Cancer Biology, The Institute of Cancer Research, 237 Fulham Road, London SW3

6JB, UK
2Division of Systems Medicine, Department of Metabolism Digestion and Reproduction, Imperial College London, London SW7 2AZ, UK
3Energy Metabolism and Nutrition, Research Institute of Health Sciences (IUNICS), University of Balearic Islands, 07122 Palma de

Mallorca, Spain
4Health Research Institute of the Balearic Islands (IdISBa), University of Balearic Islands, 07120 Palma de Mallorca, Spain
5Biomedical Research Networking Center for Physiopathology of Obesity and Nutrition (CIBERobn CB06/03/0043), Instituto de Salud Carlos

III, 28029 Madrid, Spain

6Pain Medicine Department, The Royal Marsden Hospital, London, UK
7Department of Life Sciences, College of Health and Life Sciences, Brunel University London, Uxbridge UB8 3PH, UK
8School of Biological and Chemical Sciences, Queen Mary University of London, London E1 4NS, UK
9Flow Cytometry and Light Microscopy Core Facility, Division of Cancer Biology, The Institute of Cancer Research, 237 Fulham Road, London

SW3 6JB, UK
10Sainsbury Wellcome Centre for Neural Circuits and Behaviour, University College London, 25 Howland Street, London W1T 4JG, UK
11Breast Cancer Now Research Centre, The Institute of Cancer Research, 237 Fulham Road, London SW3 6JB, UK
12Division of Tumor Biology and Immunology, the Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, the Netherlands
13The Australian National Phenome Centre, Health Futures Institute, Murdoch University, Perth WA6150, WA, Australia
14Centre for Molecular Informatics, Department of Chemistry, University of Cambridge, Lensfield Road, Cambridge CB2 1EW, UK
15Lead Contact

*Correspondence: z.takats@imperial.ac.uk (Z.T.), george.poulogiannis@icr.ac.uk (G.P.)

https://doi.org/10.1016/j.cell.2020.05.053

SUMMARY

Oncogenic transformation is associated with profound changes in cellular metabolism, but whether tracking
these can improve disease stratification or influence therapy decision-making is largely unknown. Using the
iKnife to sample the aerosol of cauterized specimens, we demonstrate a new mode of real-time diagnosis,
coupling metabolic phenotype to mutant PIK3CA genotype. Oncogenic PIK3CA results in an increase in
arachidonic acid and a concomitant overproduction of eicosanoids, acting to promote cell proliferation
beyond a cell-autonomous manner. Mechanistically, mutant PIK3CA drives a multimodal signaling network
involving mTORC2-PKCz-mediated activation of the calcium-dependent phospholipase A2 (cPLA2).
Notably, inhibiting cPLA2 synergizes with fatty acid-free diet to restore immunogenicity and selectively
reduce mutant PIK3CA-induced tumorigenicity. Besides highlighting the potential for metabolic phenotyping
in stratified medicine, this study reveals an important role for activated PI3K signaling in regulating arachi-
donic acid metabolism, uncovering a targetable metabolic vulnerability that largely depends on dietary fat
restriction.

INTRODUCTION tracking in diagnosis and disease monitoring (Vander Heiden

In addition to the routine use of histological subtyping in cancer
diagnosis and therapy decision-making, a renewed interest has
emerged in the identification of novel predictive factors to
improve patient stratification and therapeutic response. Malig-
nant transformation and disease progression involve multiple
changes in biosynthetic and energy production pathways, offer-
ing opportunities to exploit the clinical utility of metabolic
and DeBerardinis, 2017). Cells can also reprogram pathways
of nutrient acquisition and processing as a result of oncogenic
activation of commonly perturbed signaling pathways (Tar-
rado-Castellarnau et al., 2016). A characteristic example is the
activation of the phosphatidylinositol 3-kinase (PI3K)/Akt/
mammalian target of rapamycin (mTOR) pathway, which
regulates a wide range of transcriptional and post-translational
programs to support the anabolic and catabolic requirements

1596 Cell 181, 1596–1611, June 25, 2020 ª 2020 The Author(s). Published by Elsevier Inc.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Article
of proliferating cells (Carracedo and Pandolfi, 2008; Dibble and
Manning, 2013; Fruman et al., 2017; Lien et al., 2016).
These observations raise some fundamental questions:
whether we can use metabolic tracking for more effective
screening of the molecular features underlying tumor pathogen-
esis and, ultimately, whether this information can be translated
into better and more efficacious treatment strategies for each
patient. Indeed, the concept of metabotyping has been widely
applicable in characterizing functionally distinct traits that have
the power to influence clinical decision-making (Gavaghan
et al., 2000; Holmes et al., 2008; Nicholson et al., 2002, 2012).
Here, we used rapid evaporative ionization mass spectrometry
(REIMS), which coupled to the intelligent surgical device, also
known as iKnife, allows for instantaneous chemical analysis of
the aerosol generated during electrosurgical tissue ablation
and cauterization, in the form of gas-phase ionic species. Unlike
other technologies that are commonly used for metabolite
profiling, such as liquid chromatography-mass spectrometry
(LC-MS), REIMS analysis requires no sample preparation and al-
lows for near real-time (1–2 s) lipidomic analysis and tissue
recognition, based on multivariate classification analysis of
spectral libraries of reference mass spectra.
The iKnife/REIMS can be used both in the intraoperative and
biopsy collection settings to differentiate cancerous from non-
cancerous tissues with very high precision, based on their lipido-
mic composition (Alexander et al., 2017; Balog et al., 2013; St
John et al., 2017). However, the potential of using this technology
beyond the pattern-level identification of tissues, to reveal the
biological mechanisms underlying unique metabolic signatures,
or identify which patients will likely benefit from a given treat-
ment, has not yet been explored.
RESULTS
Metabolic Phenotyping Using REIMS Predicts Molecular
Markers Including Oncogenic Mutations in PIK3CA
We first examined whether REIMS-detected lipid signatures
correlate with any established molecular markers of breast can-
cer of known prognostic and therapeutic value (Figure 1A). For
this, we selected a panel of 43 breast cancer cell lines, 18 pa-
tient-derived xenograft (PDX), and 12 primary breast tumors
that are well characterized for their estrogen (ER), progesterone
(PR), and HER2 receptor status. REIMS profiling of cell lines
consistently classified ER, HER2, and triple negative status
(TN) with area under the curve (AUC) accuracies between 0.8–
0.9, and 0.6–0.7 for PR (Figure 1B).
Consistent with previous studies (Hilvo et al., 2011), the most
striking differences in lipid profiles were observed between ER-
positive (+ve) and -negative (
ve) breast cancer cell lines (Fig-
ures 1B and S1A; Table S1) and tumor specimens (Figure S1B).
A surrogate marker for ER positivity, aside from its routine deter-
mination by immunohistochemistry (IHC), is expression of the
estrogen receptor 1 (ESR1) gene. We built a regression model
to predict ESR1 expression based on the spectral profiles ob-
tained by REIMS and tested this in representative ER+ve cell lines
treated with or without 4-hydroxy-tamoxifen (4-OHT). Of note,
the predicted ESR1 expression was significantly reduced
following 4-OHT treatment as compared to untreated controls
ll
OPEN ACCESS
(Figures 1C and S1C), suggesting that the modulation of ER
signaling induces distinct lipidomic alterations, which are detect-
able by REIMS and are reversible by ER inhibition.
With a robust lipidomic profile obtained using REIMS, we
next performed unsupervised hierarchical clustering to partition
all breast cancer cell lines on the basis of their spectral similar-
ities measured over 872 lipid species. This analysis revealed
two subtypes with distinctive signatures, in which REIMS-de-
tected lipid species were significantly enriched (black) or
depleted (gray) (Figure 1D). The observed clusters were also
confirmed using a consensus non-negative matrix factorization
(NMF) (Figure S1D).
To shed light on the mechanism that is driving this unique
metabolic classification, we examined mutational enrichment
of the cells between the two subtypes. Out of the top 150 genes
that are frequently (>20%) mutated in these cell lines, oncogenic
mutation in PIK3CA was the only one to be significantly (Fisher’s
test, p value = 0.019) overrepresented in the lipid-enriched clus-
ter (Figure 1E; Table S2). In accordance with this finding, analysis
of isogenic MCF10A PIK3CA wild-type (WT) and mutant (MUT)
(E545K and H1047R) cell lines also revealed clustering of the
latter in the lipid-enriched group, both when cells were cultured
in 2D, or 3D as spheroids (Figures 1E and S1E). Consistent with
this stratification, performing gene and functional pathway
enrichment analyses revealed KEGG-pathway ontologies
relating to metabolic pathways that were significantly associated
with the lipid-enriched subtype (Figures S2A and S2B). Specific
overexpressed genes included FASN and ELOVL6, which are
involved in de novo lipogenesis, and LDLRAP1, which facilitates
exogenous lipid uptake (Figures S2C–S2E). Indeed, PIK3CA
MUT cells displayed elevated induction of the de novo lipogen-
esis transcriptional regulator SREBP1 (Figure 1F) and higher
exogenous FA uptake capacity (Figure 1G), suggesting that
both could contribute to the lipid-enriched metabotype.
Most importantly, the observed metabolic stratification was
also evident among PIK3CA WT and MUT breast cancer PDXs
(Figure 1H) and primary tumors (Figure 1I; Table S3). Among
the PDX tumors assessed (n = 18), only one was misclassified
(BR5017), and this harbored a rare I391M mutation that has ac-
tivity reminiscent of WT PIK3CA (Dan et al., 2010) (Figure 1H).
Overall, PIK3CA mutation status in both PDX and primary tumors
could be classified with an accuracy of 90% using all measurable
lipid species (Figure S1F), suggesting that the iKnife/REIMS
could be used for near real-time diagnosis of PIK3CA MUT
breast cancers by MS analysis of aerosolized tissue material.
mTORC2 Signaling Downstream of Oncogenic PIK3CA
Drives the Lipid-Enriched Phenotype
The effects of PI3K/Akt/mTOR signaling on lipid metabolism have
been observed on numerous levels (Dibble and Manning, 2013;
Lien et al., 2016; Saxton and Sabatini, 2017). To elucidate the spe-
cific mechanisms underlying the observed phenotype, we treated
PIK3CA MUT cells with inhibitors targeting the activity of key no-
des in the PI3K pathway (Figure 2A), albeit at concentrations that
do not affect cell viability (Figure S3A). PI3K (BYL719, BKM120)
and mTOR (rapamycin, torin 1) inhibition dramatically reduced
relative phospholipid levels, but surprisingly, Akt inhibition with
either MK2206 or GSK690693 did not (Figure 2B). Similar results
Cell 181, 1596–1611, June 25, 2020 1597
 ll
OPEN ACCESS Article

A B C

D E

F H I

Figure 1. REIMS Analysis Predicts Breast Cancer Molecular Markers Including Oncogenic Mutations in PIK3CA
(A) Schematic overview of sample preparation for REIMS analysis.
(B) Area under the curve (AUC) classification accuracies for ER, PR, HER2 receptor, and triple negative status of 43 breast cancer (BC) cell lines (median intensity
of n = 3 biological replicates) following feature selection for phospholipids in the m/z range 600–900, and leave-one-out cross validation.
(C) Immunoblot analysis of estrogen inducible protein pS2 and predicted ESR1 expression in ER+ve MCF7 cells following treatment with 0.1% DMSO or indicated
concentrations of 4-OHT for 72 h.
(D) Unsupervised hierarchical clustering of 872 lipid species detected by REIMS across 43 BC cell lines.
(E) Dendrogram of BC cell lines and isogenic MCF10A cells harboring either WT or MUT (E545K or H1047R) PIK3CA.
(F) Immunoblot analysis of mature SREBP1 transcription factor expression in nuclear extracts of the MCF10A PIK3CA isogenic panel.
(G) Relative exogenous fatty acid uptake in MCF10A PIK3CA WT and MUT cells following serum starvation for 1 h and supplementation with fluorescently labeled
dodecanoic acid (n = 5 replicates).
(H and I) Unsupervised hierarchical clustering of 9 PIK3CA WT and 9 MUT breast PDX tumors (H) and (I) 5 WT and 7 MUT primary breast tumors. Individual rows in
the heatmaps in (D), (H) and (I) correspond to scaled Z score phospholipid intensities (n = 3 biological replicates). Error bars represent ± SEM. n.s., not significant;
*p % 0.05; **p % 0.01; ***p % 0.001. p values in (C, bottom panel) and (G) were calculated with one-way ANOVA, followed by unpaired, two-tailed Student’s t test
with Bonferroni correction.

1598 Cell 181, 1596–1611, June 25, 2020

 ll
Article OPEN ACCESS

A B C D

Figure 2. Oncogenic PIK3CA Drives the Lipid-Enriched Phenotype via mTORC2 Signaling
(A) MCF10A PIK3CA MUT cells were treated with BYL719 or BKM120 (100 nM), MK2206, or GSK690693 (150 nM) for 72 h, or rapamycin (20 nM) for 4 h, or
rapamycin or torin 1 (20 nM) for 72 h.
(B) Unsupervised hierarchical clustering of MCF10A E545K and H1047R MUT cells treated with PI3K, AKT, and mTOR inhibitors.
(C and D) Immunoblot analysis (C) and unsupervised hierarchical clustering (D) of MCF10A E545K and H1047R cells transfected with RAPTOR, RICTOR, or mTOR
siRNA. Individual rows in the heatmaps in (B) and (D) correspond to scaled Z score phospholipid intensities (n = 3 biological replicates).

were observed in a panel of 5 PIK3CA MUT breast cancer cell
lines, with the exception of MCF7 cells that also responded to
MK2206 (Figure S3B).
Interestingly, we did not observe an effect on relative phos-
pholipid abundances following acute exposure to rapamycin
for 4 h, despite inhibition of mTORC1 (Figures 2A, bottom left
panel, and 2B). Because extended rapamycin treatment inhibits
mTORC1 and mTORC2 (Figure 2A, bottom right panel) (Sarbas-
sov et al., 2006), and both impinge upon lipogenesis (Düvel et al.,
2010; Griffiths et al., 2013; Guri et al., 2017; Lee et al., 2017;
Ricoult et al., 2016), we sought to investigate which of these
complexes might contribute to the regulation of the observed
phenotype. Knockdown of RICTOR or mTOR, but not RAPTOR,
led to a significant reduction in relative phospholipid abun-
dances (Figures 2C and 2D), pointing to a PIK3CA- and
mTORC2-dependent metabolic phenotype that is largely inde-
pendent of mTORC1 or Akt inhibition.
Oncogenic PIK3CA Drives Enhanced Arachidonic Acid
Metabolism, thereby Promoting Cell Proliferation
beyond a Cell-Autonomous Manner
Given that global lipidomic profiles could stratify breast cancer
cell lines and tumors based on PIK3CA mutation status, we
next aimed to characterize specific lipid alterations that are
associated with oncogenic PIK3CA. Fatty acids (FAs), which
are the main constituents of phospholipids and have additional
effector functions in cancer pathogenesis, were profiled in the
PIK3CA isogenic panel using REIMS. Among the most abundant
FAs in PIK3CA MUT compared to WT cells were palmitoleate
(FA16:1), palmitic acid (FA16:0), and oleic acid (FA18:1), all of
which are established products of elevated lipogenesis and in
line with the observed lipid enriched phenotype (Figure 1D; Table
S4). Interestingly, the second most significantly elevated FA after
palmitoleate was arachidonic acid (AA) (FA20:4), an omega-6 FA
which is predominantly found in animal fats and is of particular
relevance as a major regulator of pro-inflammatory responses
in cancer, through the production of bio-active lipids known as
eicosanoids (Wang and Dubois, 2010) (Figure 3A; Table S4).
Importantly, in addition to the cell lines, significant elevations in
AA were also observed in all the PIK3CA MUT breast PDX and
primary tumors (Figures 3B and 3C) and across tumors of other
tissue types including ovarian, pancreatic, and sarcomas (Fig-
ure 3D). In agreement with our REIMS findings, AA and down-
stream eicosanoids were also found to be significantly elevated
in both PIK3CA MUT cells using LC-MS (Figures 3E, S4A,
and S4B).
To measure FAs that are secreted from cells, as opposed to
those that might already exist in serum-supplemented media,
cells were grown under FA-deprived conditions. Pro-inflamma-
tory derivatives of AA were significantly increased in the media
of PIK3CA MUT cells (Figure 3F), signifying a potential role for
these bio-active lipids in tumor microenvironment (TME)
interactions.
Next, to ascertain the functional consequences of elevated
eicosanoid metabolism, the effects of PIK3CA MUT-derived
conditioned media (CM) were assessed. PIK3CA WT cells dis-
played dramatically increased proliferative rate following incuba-
tion with CM obtained from MUT cells, and this was effectively
rescued by depleting the lipids from the media (Figures 3G,
S4C, and S4D). Moreover, the proliferation of both PIK3CA

Cell 181, 1596–1611, June 25, 2020 1599

 ll
OPEN ACCESS Article

A B C

D E

F G H

Figure 3. Oncogenic PIK3CA Drives Enhanced Arachidonic Acid Metabolism

(A and B) Arachidonic acid (AA) levels measured by REIMS in MCF10A PIK3CA WT and MUT cells cultured under full 5% horse serum or fatty acid-free (FAF)
conditions for 72 h (n = 3 biological replicates) (A). AA levels of 18 breast PDX tumors (n = 9 PIK3CA WT and n = 9 MUT) (left) (B). Three sections corresponding to
different tumor regions were analyzed with REIMS. Data are summarized in the boxplot to the right.
(C) 12 primary breast tumors (n = 5 PIK3CA WT and n = 7 MUT) (left). Data are summarized in the boxplot to the right.
(D) Breast, ovarian, pancreatic, sarcoma, and colorectal PDX tumors (n = 5 PIK3CA WT and MUT tumors for breast, pancreatic, sarcoma, and colorectal tissues,
and n = 4 PIK3CA WT and MUT ovarian PDX tumors).
(E) Heatmap and Venn diagram summarizing the intracellular eicosanoids that were significantly different between MCF10A PIK3CA WT and MUT cells. Rows
correspond to the Z score scaled eicosanoid intensities detected by LC-MS (n = 3 biological replicates).
(F) Heatmap and Venn diagram summarizing the eicosanoids of the conditioned media (CM) that were significantly different between MCF10A PIK3CA WT and
MUT cells.
(G) Cell proliferation assays of MCF10A PIK3CA WT cells cultured in CM derived from WT or H1047R MUT cells before or after lipid depletion (LD), with or without
the supplementation of 25 mM AA, palmitate, or palmitoleate.
(H) Cell proliferation assays of MCF10A PIK3CA H1047R MUT cells before or after LD, with or without the supplementation of 25 mM AA, palmitate, or palmi-
toleate. Sulforhodamine B (SRB) protein staining was used in (G) and (H) to measure cell proliferation over 5 days (replicates from n = 3 wells). Error bars in (G) and
(H) represent mean ± SEM for each time point. *p % 0.05; **p % 0.01; ***p % 0.001. p values in (A)–(D) were calculated with unpaired, two tailed Student’s t test.
Two-way ANOVA was used for (G) and (H).

MUT cell lines was significantly reduced following incubation
with their respective lipid-deprived CM (Figures 3H and S4E),
whereas supplementation of lipid-deprived CM with AA, but
not palmitate or palmitoleate, restored proliferation in both WT
and MUT cells (Figures 3G, 3H, S4D, and S4E). Together, these
data support not only a critical role for oncogenic PIK3CA in influ-
encing autocrine- and paracrine-mediated cell proliferation, but
also point to AA as an easily measured metabolic biomarker that
could help with the diagnosis and treatment of PIK3CA MUT
tumors.

1600 Cell 181, 1596–1611, June 25, 2020

 ll
Article OPEN ACCESS

A B C D

E F G H

J K M

Figure 4. Oncogenic PIK3CA Signaling Triggers cPLA2-Induced Arachidonic Acid Production

(A) Enzymatic activity of cPLA2, iPLA2, and sPLA2 in the MCF10A PIK3CA isogenic panel.
(B–D) cPLA2 activity (B) and AA levels (C and D) measured by REIMS in MCF10A H1047R PIK3CA MUT cells following RAPTOR or RICTOR siRNA-mediated
knockdown (C), or treatment with 100 nM ASB14780, 1 mM each of PKCa, b, ε, or z peptide inhibitors, 250 mM GSK650394, or 150 nM MK2206 for 72 h (D). Cells
were grown under exogenous FAF conditions.
(E) cPLA2 activity following PKCz inhibition with 1 mM peptide inhibitor for 72 h.

(legend continued on next page)

Cell 181, 1596–1611, June 25, 2020 1601

 ll
OPEN ACCESS Article

Oncogenic PIK3CA Promotes Enhanced Arachidonic
Acid Production via mTORC2-PKCz-cPLA2 Signaling
To better understand the mechanism by which PIK3CA MUT
cells have elevated AA, we assessed various pathways which
contribute to its cellular pool, including: direct exogenous up-
take, synthesis from linoleic acid, hydrolysis from diacylglycerol
(DAG), or endogenous release from membrane phospholipid
through phospholipase (PLAs) activity. Curiously, we noted a
persistent increase in AA in PIK3CA MUT isogenic panel even
when cells were cultured with FA-free media (Figure 3A). Addi-
tionally, DAG levels—that can be in part generated from the hy-
drolysis of phosphatidylinositol-4,5-bisphosphate (PIP2)—were
significantly reduced in PIK3CA MUT cells (Figures S4F
and S4G).
These results pointed to a potential role for phospholipases
(PLAs), of which three main classes predominate: cytosolic/cal-
cium-dependent (cPLA2), calcium-independent (iPLA2), and
secretory (sPLA2) phospholipase A2 (Burke and Dennis, 2009).
Among these, only cPLA2 displayed significantly higher enzy-
matic activity (Figure 4A), as well as elevated total protein levels
and stability (Figures S5A and S5B) in the presence of oncogenic
PIK3CA, whereas expression of PLA2G4A—the gene encoding
cPLA2a—remained unchanged (Figure S5C).
Consistent with the predominant role of mTORC2 in driving the
lipid enriched phenotype in PIK3CA MUT cells (Figures 2C and
2D), RICTOR, but not RAPTOR, silencing rescued cPLA2 activity
(Figure 4B), and led to a concomitant reduction in AA and pros-
taglandin E2 (PGE2) levels (Figures 4C and S5D–S5F). Impor-
tantly, mTORC2-specific inhibition was also accompanied by
decreased cPLA2 stability (Figure S5G). To elucidate the mech-
anism underlying this observation, known substrates of
mTORC2 including serum/glucocorticoid-regulated kinase 1
(SGK-1) and the protein kinase C (PKC) isoforms were inhibited
in PIK3CA MUT cells. ASB14780—an indole-based compound
that inhibits both cPLA2 translocation to membrane compart-
ments and the interaction between phospholipid substrates
with the enzyme active site (McKew et al., 2008; Tomoo et al.,
2014)—was used as a positive control. Importantly, a significant
reduction in AA and cPLA2’s enzymatic activity, reminiscent of
that observed upon ASB14780 inhibition, was only observed
following pharmacological and small interfering RNA (siRNA)-
mediated inhibition of PKCz (Figures 4D, 4E, and S5H–S5K).
Previous studies have shown that phosphorylation of cPLA2
on S505 regulates its activity and stability and that this is medi-
ated, at least in part, by the p38 MAPK/ERK signaling pathway
(Kramer et al., 1996; Lin et al., 1993). The contribution of PI3K-
mTORC2 signaling to this process is unknown, as is the role of
PKCz in the regulation of cPLA2. Consistent with PKCz being a
direct substrate of PDK-1 (Chou et al., 1998) and mTORC2 phos-
phorylation (Li and Gao, 2014), it was found to be hyperphos-
phorylated in PIK3CA MUT cells (Figure 4F) and breast PDX
tumors (Figure S5L). Moreover, its inhibition led to a marked
reduction in MAPK/ERK signaling (Figure 4G), as well as p38
MAPK phosphorylation and active GTP-bound Rac-1 (Fig-
ure 4H), culminating in reduced cPLA2 phosphorylation at the
S505 site (Figures 4G and S5L).
In addition to S505 phosphorylation, an increase in intracel-
lular calcium levels is essential for sustained phospholipase ac-
tivity and liberation of AA by cPLA2 (Ambs et al., 1995; Clark
et al., 1995). Given that elevated PIP3 levels, induced by onco-
genic PIK3CA, promote the activation of phospholipase C
gamma 1 (PLCg1), leading to an increase in cytosolic calcium
via generation of inositol-1,4,5-trisphosphate (IP3) (Rameh
et al., 1998), we hypothesized that this signaling node could
also play a role in regulating cPLA2 activity and AA release down-
stream of active PI3Ka. In line with this premise, we observed
higher phosphorylation of PLCg1 (Figure S6A) and significantly
elevated intracellular calcium levels in PIK3CA MUT cells (Fig-
ure S6B). Although genetic and pharmacological (U73122) inhibi-
tion of PLCg1 led to a significant reduction in calcium flux in both
PIK3CA WT and MUT cells (Figures S6C–S6E), an inhibitory ef-
fect on cPLA2 activity and AA levels was only observed in the
MUT cells (Figures S6F–S6H), highlighting the importance of
PLCg1-mediated Ca2+ flux in sustaining elevated cPLA2 activity
in the context of oncogenic PIK3CA.
Finally, because cPLA2 has a predicted PKCz phosphorylation
site (T376), we tested the possibility that it could serve as a direct
substrate for PKCz. In vitro kinase assays using purified PKCz
and cPLA2 suggested a direct interaction and phosphorylation
(Figure 4I), and this was confirmed following immunoprecipita-
tion (Figure 4J) and proximity ligation activity (PLA) assays (Fig-
ures S6I and S6J). To further evaluate cPLA2 as a candidate sub-
strate for PKCz, we developed a custom antibody recognizing
the cPLA2 T376 phosphorylation site. Specificity was validated
in both serum starved/stimulated samples (Figure S6K, left
panel), following PKCz inhibition (Figures S5K and S6K, middle),
and in cPLA2 CRISPR knockout cells overexpressing a phos-
phoresistant MUT (T376A) cPLA2 (Figure S6K, right panel).
Importantly, increased T376 phosphorylation was observed in
the presence of oncogenic PIK3CA and this was reduced to
levels equivalent to PIK3CA WT cells upon PKCz inhibition
(Figure 4K).

(F and G) Immunoblot analysis of the MCF10A PIK3CA isogenic panel following growth factor deprivation for 16 h and 30 min stimulation with serum and growth
factors (F) or PKCz inhibition with 1 mM peptide inhibitor for 72 h (G).
(H) Immunoblot analysis of activated Rac-1 and p38 MAPK in the MCF10A PIK3CA isogenic panel following PKCz inhibition with 1 mM peptide inhibitor for 72 h.
(I) In vitro kinase assay of 100 ng and 0.5 mg/mL purified PKCz and cPLA2 proteins, respectively.
(J) Immunoblot analysis of anti-HA immunoprecipitates derived from HA-tagged cPLA2 transfected MCF10A PIK3CA WT and MUT cells.
(K) Immunoblot analysis of anti-HA immunoprecipitates derived from HA-tagged cPLA2 transfected MCF10A PIK3CA WT and MUT cells treated where indicated
with 1 mM PKCz peptide inhibitor for 48 h.
(L) AA levels across H1047R MUT cells with CRISPR knockout of PLA2G4A reconstituted with WT or phosphoresistant cPLA2 isoforms.
(M) Diagram summarizing the proposed model for PI3K-mTORC2-PKCz and calcium-dependent activation of cPLA2, leading to a concomitant increase in AA
and downstream eicosanoids. Data are presented as the mean ± SEM of n = 3–6 biological replicates and are representative of at least two independent ex-
periments. n.s., not significant; *p % 0.05; **p % 0.01; ***p % 0.001. p values in (A) were calculated with unpaired, two tailed Student’s t test, and in (B)–(E), (I), and
(L) with one-way ANOVA, followed by unpaired, two-tailed Student’s t test with Bonferroni correction.

1602 Cell 181, 1596–1611, June 25, 2020

 ll
Article OPEN ACCESS

A B C D

F G

Figure 5. Genetic and Pharmacological Inhibition of cPLA2 Selectively Reduces Oncogenic PIK3CA-Mediated Tumorigenicity
(A–D) Cell viability of (A) PIK3CA WT and MUT MCF10A cells, and (B) breast cancer cell lines (PIK3CA WT: MDAMB134, Hs578T, AU565; PIK3CA MUT: MCF-7,
CAL-51, MDAMB453) following treatment with increasing concentrations (20 nM–10 mM) of ASB14780 under full serum conditions for 72 h. The same treatments
were also performed under fatty acid-free conditions in (C) and (D), in the presence or absence of exogenous supplementation of 25 mM AA.
(E) Clonogenic assays of MCF10A PIK3CA WT and MUT cells treated with increasing concentrations of ASB14780 as in (A)–(D). Treatments were performed
under fatty acid-free conditions, with or without the supplementation of 25 mM AA.
(F) Immunoblot analysis confirming specific knockdown of cPLA2 using two independent constitutive shRNAs (sh1 and sh5) (left) and reduction in AA levels in
MCF10A E545K/H1047R MUT cells using REIMS.
(G) Proliferation of MCF10A H1047R MUT cells expressing shGFP, cPLA2-sh1, or cPLA2-sh5 under exogenous FAF conditions. Sulforhodamine B (SRB) protein
staining was used to measure cell proliferation over 5 days.
(legend continued on next page)

Cell 181, 1596–1611, June 25, 2020 1603

 ll
OPEN ACCESS
To ascertain the functional significance of these two phos-
phorylation sites (S505 and T376), endogenous cPLA2 in
PIK3CA WT and H1047R MUT cells was reconstituted with WT
or phosphoresistant MUT of cPLA2 (S505A or T376A) (Figure 4L).
Knockout of cPLA2 in H1047R cells reduced AA to levels equiv-
alent to the PIK3CA WT background, and this could be rescued
with ectopic expression of the WT, but not MUT (S505A or
T376A) cPLA2 (Figure 4L). Interestingly, the activity of exoge-
nously expressed WT or MUT cPLA2 largely mirrored the trends
in AA that were previously detected by REIMS (Figure S6L), sug-
gesting that cPLA2 activity is highly dependent on the phosphor-
ylation of S505 and T376 in the context of oncogenic PIK3CA.
Overall, our model provides a unifying framework for several pre-
viously unconnected components of PI3K signaling, which
converge on cPLA2 activation and enhanced AA metabolism
(Figure 4M).
cPLA2 Inhibition and Dietary Fat Restriction Suppress
PIK3CA-Induced Tumorigenicity and Restore Anti-
cancer Immune Responses
If cPLA2 is required for PIK3CA oncogenicity, then targeting this
enzyme could represent an attractive therapeutic strategy. For
this, we used the inhibitor ASB14780, which displays excellent
oral bioavailability and higher specificity for the cPLA2 isoform
than other commonly used compounds such as Efipladib and
Ecopladib (Lee et al., 2007; Tomoo et al., 2014). Furthermore,
ASB14780 has been shown to ameliorate inflammatory pathol-
ogies including non-alcoholic fatty liver disease (NAFLD) and
chronic obstructive pulmonary disease (COPD) through sup-
pression of AA and prostaglandin synthesis (Kanai et al., 2016;
Tomoo et al., 2014). However, the potential anti-neoplastic prop-
erties of this inhibitor remain obscure. Of note, PIK3CA mutation
sensitized cells to pharmacological inhibition of cPLA2 with
ASB14780 (Figures 5A and 5B), and this effect was more prom-
inent under exogenous FA-free conditions (Figures 5C and 5D),
alleviating any compensatory mechanisms to obtain AA. In addi-
tion to viability, cPLA2 inhibition significantly reduced the clono-
genicity of PIK3CA MUT cells, and, importantly, both could be
rescued by exogenous supplementation of AA (Figures 5C, 5D,
and 5E). Near-identical results were obtained following genetic
knockdown of cPLA2 using two constitutive short-hairpin
RNAs (shRNAs, denoted as cPLA2-sh1 and cPLA2-sh5) (Figures
5F, 5G, and S6M), while WT cells were unaffected, suggesting
that cPLA2 is dispensable in this setting (Figure S6N). To further
confirm the importance of cPLA2-induced AA metabolism in
PIK3CA MUT cancers, we suppressed cPLA2 in the isogenic
panel and assessed their ability to form colonies under FA-free
conditions. Although no significant difference in colony forma-
tion was detected in PIK3CA WT cells, there was a marked
reduction in the number of colonies following cPLA2 knockdown
in the MUT cells that was restored in the presence of AA
(Figure 5H). Furthermore, using an inducible shRNA against
(H) Clonogenic assays of MCF10A PIK3CA WT and MUT cells expressing shGFP, cPLA2-sh1, or cPLA2-sh5 under FAF conditions, supplemented with or without
25 mM AA. Data in (A)–(H) are presented as the mean ± SEM of n = 3–4 biological replicates and are representative of at least two independent experiments. Data
in (D) are presented as the mean viability of three PIK3CA MUT (MCF-7, CAL-51, MDAMB453) and WT (MDAMB134, Hs578T, AU565) measured in triplicate wells.
n.s., not significant; *p < 0.05; **p < 0.01; ***p < 0.001; p values in (A)–(D) and (G) were calculated using two-way ANOVA. For (E, right), (F, right), and (H, right), one-
way ANOVA followed by unpaired, two-tailed Student’s t test with Bonferroni correction was applied.
1604 Cell 181, 1596–1611, June 25, 2020
Article
cPLA2, we demonstrated that PIK3CA MUT, but not WT cells,
rely on cPLA2 to form epithelial acini in 3D culture and sustain
their proliferation (Figures 6A–6C).
To further evaluate the therapeutic effect of cPLA2 inhibition in
primary breast cancers, we treated triple negative breast cancer
(TNBC) (Figure 6D) PIK3CA WT (Figure 6E) and MUT (Figure 6F)
PDX-bearing mice with the ASB14780 inhibitor in conjunction
with a normal or near-isocaloric fat-free diet. A significant reduc-
tion in tumor weight was only observed in the PIK3CA MUT PDX
model when both the inhibitor and a fat-free diet were adminis-
tered in combination (Figure 6F). Corroborating these observa-
tions, histological analysis did not reveal any changes in tumor
area for the PIK3CA WT BR1458 model (Figures 6G, left panels,
and 6H), while a striking reduction in viable tumor regions was
observed in PIK3CA MUT PDX-bearing mice treated with
ASB14780 in fat-free diet (Figures 6G, right panels, and 6I).
The concomitant increase in necrotic regions (as indicated by
areas of pale eosinophilic cytoplasms, in addition to loss of
nuclei and karyolysis), evidently contributed to a substantial pro-
portion of the overall weight of the PIK3CA MUT-bearing tumor
that was left after treatment of ASB14780 in fat-free diet (Fig-
ure 6G, bottom right panel). In the interest of measuring AA levels
at the end of the treatment regime (Figure 6D), resected tumors
were analyzed directly with REIMS. Although a fat-free diet alone
led to a modest, yet significant reduction in the AA levels of the
PDX tumors, the decrease was significantly more pronounced
when accompanied with cPLA2 inhibition (Figures 6J and 6K),
highlighting the role of dietary AA restriction in therapy response
(Figures 6E–6I).
The observation that cPLA2 inhibition and fat-free diet selec-
tively reduces the tumorigenicity of PIK3CA MUT cells raises
the interesting prospect of modulating this response by altering
dietary fat content. Indeed, it is becoming increasingly appreci-
ated that the diet of Western populations commonly contains
an excess of pro-inflammatory omega-6 to omega-3 FAs by up
to 50 times (popularly referred to as the ‘‘Western’’ diet), and
this may be implicated in the progression of breast and colo-
rectal cancers (Patterson et al., 2012; Simopoulos, 2008). To
further explore this premise, we injected triple negative CAL51
(PIK3CA MUT) and Hs578T (PIK3CA WT) cell lines stably ex-
pressing either control shGFP, or two independent shRNAs tar-
geting cPLA2 (cPLA2-sh1 or sh5) into the mammary fat pad of
BALB/c nude mice that had been preconditioned on either fat-
free, balanced (omega6:omega3 = 1:1) or ‘‘Western’’ (omega6:o-
mega3 = 50:1) diets (Figure 7A). Consistent with pharmacolog-
ical inhibition, knockdown of cPLA2 significantly impaired the
growth of PIK3CA MUT tumor xenografts under fat-free diet con-
ditions (Figure 7B), and this therapeutic effect was completely
reversed when animals were fed the AA-enriched ‘‘Western’’
diet (Figures 7C and 7D). Although there was a trend toward a
reduction in overall tumor weights under a balanced diet, this
did not reach statistical significance (Figures S7A and S7B). In
 ll
Article OPEN ACCESS

A B C

E F G

H I J K

(legend on next page)

Cell 181, 1596–1611, June 25, 2020 1605

 ll
OPEN ACCESS Article

accordance with our model, PIK3CA WT tumors were unaffected
by these treatments (Figures 7E–7G, S7C, and S7D). Further
corroborating our findings, genetic inhibition of cPLA2 only
reduced viable tumor area in animals with PIK3CA MUT tumors
when those were fed a fat-free diet, while no anti-neoplastic
benefit was conferred under a ‘‘Western’’ diet (Figures 7H–7J).
REIMS profiling of excised tumors revealed that AA levels
were significantly altered in concordance with targeting cPLA2
and dietary fat intake (Figures 7K and S7E). Interestingly, a
more substantial AA reduction was observed in PIK3CA MUT
xenograft tumors following cPLA2-knockdown and administra-
tion of fat-free diet (40%–50% decrease), as compared to WT tu-
mors (20% decrease) (Figures 7K and 7L, S7E, and S7F). In line
with our previous in vivo study (Figure 6), these findings demon-
strate that the modulation of dietary fat content, and supplemen-
tation of omega-6 FAs either in a balanced ratio with omega-3
FAs, or to a much larger extent in the ‘‘Western’’ diet, completely
abolishes the therapeutic benefit of cPLA2 inhibition in PIK3CA
MUT tumors.
In addition to promoting growth and proliferation through both
autocrine and paracrine mechanisms, AA and its downstream me-
tabolites have also been implicated in the metabolic remodeling of
the tumor microenvironment. One of their major consequences is
inhibition of the anti-cancer immune responses, ultimately leading
to immune evasion and tumor progression (Böttcher et al., 2018;
Zelenay et al., 2015). Although the BALB/c nude mice used in
this study lack adaptive immunity in the form of T cells, these an-
imals mount robust innate immune responses predominantly
mediated by natural killer (NK) cells (Lee et al., 2015; Okada
et al., 2019). We therefore sought to investigate how the various
therapeutic and dietary regimes that were used in this study
impact NK cell responses. To do this, the levels of type I inter-
feron-induced chemokines (CCL5 and CX3CL1) and expression
of a major NK cell-activating receptor (NKp46) were measured
in tumors from different treatment groups.
Coinciding with the largest reduction in AA levels (Figure 6K), a
marked increase in CCL5 and CX3CL1 was only observed in the
BR1282 (PIK3CA MUT) PDX tumor following co-administration
of ASB14780 and fat-free diet (Figures S7G and S7I). Similar re-
sults were obtained in the CAL51 (PIK3CA MUT)-derived xeno-
graft tumor model, where the increase in chemokine levels was
rescued when mice were fed the AA-enriched ‘‘Western’’ diet
(Figures S7H and S7J).
In agreement with the chemokine analysis, tumor infiltration of
NK cells was significantly increased in PIK3CA MUT tumors by
dietary and therapeutic interventions, with dual inhibition (either
with ASB14780 or shRNA) of cPLA2 and fat-free diet leading to
the largest increase in NKp46 staining (Figures S7K, S7L, S7N,
and S7O). It is also noteworthy that PIK3CA WT PDX and cell
line-derived xenograft tumors contained relatively higher base-
line levels of CCL5, CX3CL5, and NKp46 expression, as
compared to MUT tumors, and these were not significantly
altered by cPLA2 inhibition and/or changes in the diet (Figures
S7G–S7J, S7M, and S7P). Considered together, these data sug-
gest that oncogenic PIK3CA might suppress BC immunoge-
nicity, at least in part, through regulation of AA metabolism,
and this can be reversed through co-administration of cPLA2 in-
hibition and dietary fat restriction.
DISCUSSION
Unraveling the interplay between genotype and metabolic
phenotype, as well as their complex interactions with nutrient
availability, unequivocally plays a major role in understanding
disease pathogenesis and identifying novel therapeutic interven-
tions. Here, we demonstrate that the iKnife/REIMS enables close
to real-time prediction of clinically relevant tumor features based
on their metabolic fingerprints, offering a novel repertoire for
cancer diagnosis and therapy decision-making. Among these,
is oncogenic PIK3CA, which triggers almost ‘‘the perfect storm’’
of signaling events, culminating in the overproduction of AA and
downstream eicosanoids via the activation of cPLA2.
We have evidence of the central role of AA and eicosanoids in
a wide range of disorders including cancer, obesity, diabetes,
asthma, and autoimmune disorders (Dennis and Norris, 2015;
Sonnweber et al., 2018; Wang and Dubois, 2010). However,
the signal transduction pathways behind their activation, as
well as the molecular cues that link these bio-active lipids with
growth factor-independent cell proliferation have remained
largely obscure. Our results demonstrate that mTORC2 down-
stream of oncogenic PIK3CA acts as a pivotal signaling hub for
driving enhanced AA metabolism to sustain cell proliferation
beyond a cell autonomous manner. This is particularly interesting
in light of evidence obtained from in situ single-cell analysis of
primary breast tumors, showing often that only a small fraction
of cancer cells within a tumor carry PIK3CA mutations, while

Figure 6. Oncogenic PIK3CA Serves as a Defining Biomarker for Sensitivity of Pre-clinical Models to cPLA2 Inhibition
(A) Immunoblot analysis confirming inducible knockdown of cPLA2 following induction with 2 mg/mL doxycycline.
(B) 3D acini formation of MCF10A PIK3CA WT and MUT cells following doxycycline-induced cPLA2-sh1 or shGFP expression. Cells were stained for Ki-67 (pink,
Alexa Fluor 546), F-actin (red, Phalloidin 633), and DAPI (blue).
(C) Quantification of Ki-67 staining from treatments in (B).
(D) Schematic of in vivo experimental design and tumor profiling with REIMS.
(E and F) Tumor weights of (E) PIK3CA WT (BR1458) and (F) C420R MUT (BR1282) breast PDX tumors treated with 100 mg/kg of the cPLA2a pharmacological
inhibitor ASB14780 under FAF diet (n = 8 mice for the BR1458 model, and n = 7 mice for the BR1282 for both the vehicle- and ASB14780-treated groups).
(G) Representative images of H&E staining from resected tumors in (E) and (F). The black masks in (G) represent viable tumor area, while unshaded regions
correspond to necrotic tissue.
(H and I) Quantification of viable tumor area from (H) PIK3CA WT (BR1458) and (I) PIK3CA MUT (BR1282) tumor sections based on the analysis depicted in (G).
(J and K) AA levels measured by REIMS in (J) PIK3CA WT (BR1458) and (K) MUT (BR1282) tumors excised and snap frozen 2 h after the final dosing. Error bars in
(C), (J), and (K) represent mean ± SEM, with data in (J) and (K) corresponding to tumor REIMS measurements from n = 7–8 mice. n.s., not significant;*p < 0.05;
**p < 0.01; ***p < 0.001; p values in (E), (F), and (H)–(K) were calculated using one-way ANOVA followed by unpaired, two-tailed Student’s t test with Bonferroni
correction.

1606 Cell 181, 1596–1611, June 25, 2020

 ll
Article OPEN ACCESS

B C D

E F G

H I J

K L

(legend on next page)

Cell 181, 1596–1611, June 25, 2020 1607

 ll
OPEN ACCESS Article

many of the neighboring cancer and stromal cells are WT (Janis-
zewska et al., 2015). Given that AA has been shown to induce
both PI3K (Hughes-Fulford et al., 2006) and MAPK (Alexander
et al., 2006) signaling, PIK3CA MUT cells could trigger a snow-
ball effect, through overproduction of AA, affecting not only their
own signaling and proliferation, but also that of their adjacent
PIK3CA WT cells. Moreover, in light of the role of prostaglandins
in lymphangiogenesis (Lala et al., 2018), the paracrine effects of
AA could be of further relevance to the activity of PI3Ka in endo-
thelial cells (Okkenhaug et al., 2016; Wang and Dubois, 2010).
Eicosanoids no longer represent the missing link between
inflammation and cancer (Greene et al., 2011). Elevated tumor-
derived PGE2 contributes to immune evasion by preventing the
interferon gamma (IFNg)-dependent upregulation of ICAM-1
that is pertinent for complete CD8(+) T cells activation (Basingab
et al., 2016). In addition, autocrine PGE2 impairs NK cell viability
and chemokine production and leads to downregulation of the
chemokine receptors of cDC1 that promote their recruitment
into tumors (Böttcher et al., 2018). Importantly, there is a strong
positive correlation between the gene signature of cDC1 and
NK cells and better overall survival in melanoma and breast can-
cers (Böttcher et al., 2018), suggesting that monitoring the immu-
nomodulatory functions of prostaglandins via PI3K/Akt pathway
inhibition could have important clinical implications. Indeed, we
have shown that modulation of AA levels in PIK3CA MUT tumors
through cPLA2 inhibition in combination with dietary fat restriction
increases intra-tumor infiltration of NK cells and their associated
chemokines, while this can be reversed by the ‘‘Western’’ diet,
which contains an excess of omega6-FAs. NK cell markers
were largely unaffected in PIK3CA WT tumors, and this could
reflect their lower intra-tumor AA levels that are likely attributable
to reduced cPLA2 activity and FA uptake, as compared to
PIK3CA MUT cells. In light of recent evidence showing that block-
ing PI3K signaling with the pan-PI3K inhibitor BKM120 increases
tumor-immune infiltrate and renders PIK3CA MUT mouse bladder
tumors more susceptible to PD-1 blockade (Borcoman et al.,
2019), our model raises important considerations for how immu-
notherapies may be successfully applied to oncogenic PIK3CA
MUT tumors that may be inherently less immunogenic, at least
in part, due to enhanced AA production.
Another way to relieve the immunosuppressive effects of tu-
mor cells is by inhibiting COX activity via the use of non-steroidal
anti-inflammatory drugs (NSAIDs), such as aspirin. Notably,
oncogenic PIK3CA has been shown to sensitize cancer cells to
aspirin (Henry et al., 2017; Liao et al., 2012) and, based on our
findings, it is tempting to speculate that this connection could
be true, in part because of the heightened capacity of PIK3CA
MUT cells for high AA production. However, further studies are
needed to ascertain this connection, because evidence sug-
gests that the growth inhibitory effect of aspirin in PIK3CA
MUT cells is likely to be COX-2-independent (Henry et al., 2017).
Although PI3K pathway inhibitors have shown some efficacy in
treating advanced solid tumors, the majority has been associ-
ated with only partial tumor remission and they are often accom-
panied by severe side effects (Fruman et al., 2017; Li et al., 2018).
Recent evidence suggests that one way to enhance their efficacy
is by suppressing their insulin feedback through adoption of a
ketogenic diet (Hopkins et al., 2018). Indeed, diet could play a
much more significant role in therapy response than previously
anticipated. Our data suggest that a diet rich in FAs limits the ef-
ficacy of the cPLA2 inhibitor, as PIK3CA MUT tumors likely
depend on their high flux of extracellular FA intake to compen-
sate for the loss of AA. This observation raises the possibility
that adopting a diet without meat and dairy products (major sour-
ces of AA) could dramatically improve the sensitivity of the
cPLA2 inhibitor and help restore tumor immunogenicity, sug-
gesting a novel path for future clinical trials where nutrition will
play a major role in disease management and treatment.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Human samples
B Animals
B Cell culture
d METHOD DETAILS
B Experimental design
B Mass spectrometry analysis

Figure 7. Dietary Supplementation of Arachidonic Acid Reverses the Sensitivity of PIK3CA Mutant Tumors to cPLA2 Inhibition
(A) Schematic of in vivo experimental design and profiling of breast cancer cell line xenografts with REIMS.
(B and C) Relative tumor growth of CAL-51 (PIK3CA MUT)-derived xenografts stably expressing control shGFP or two independent shRNAs targeting cPLA2
(cPLA2-sh1 and cPLA2-sh5) under (B) fat-free or (C) ‘‘Western’’ diets.
(D) Weights of tumors excised at the end of the experiments (B) and (C).
(E and F) Relative tumor growth of Hs578T (PIK3CA WT)-derived xenografts stably expressing control shGFP or two independent shRNAs targeting cPLA2 under
(E) fat-free or (F) ‘‘Western’’ diets.
(G) Weights of tumors excised at the end of the experiments (E) and (F).
(H) Representative images of H&E staining from resected tumors in (D) and (G). The black masks in (H) represent viable tumor area, while unshaded regions
correspond to necrotic tissue.
(I and J) Quantification of viable tumor area from (I) PIK3CA MUT (CAL-51) and (J) PIK3CA WT (Hs578T) tumor sections based on the analysis depicted in (H).
(K and L) AA levels measured by REIMS in (K) PIK3CA MUT (CAL-51) and (L) PIK3CA WT (Hs578T) snap frozen excised tumors. AA intensities are reported as
scaled values to the appropriate shGFP-fat-free diet condition. Data in (B), (C), (E), (F), (K), and (L) represent the mean ± SEM of relative tumor growth or tumor
REIMS measurements from n = 3–5 mice. n.s., not significant; *p < 0.05; **p < 0.01; ***p < 0.001; p values in (B), (C), (E), and (F) were calculated using two-way
ANOVA, and one-way ANOVA followed by unpaired, two-tailed Student’s t test with Bonferroni correction was used in (D), (G), and (I)–(L).

1608 Cell 181, 1596–1611, June 25, 2020


Article
B Metabolomics data pre-processing and analysis
B Transfections and site directed mutagenesis
B Cell based assays
B Three-dimensional cell culture
B Confocal microscopy
B Enzymatic assays
B Immunoblot analysis
B Immunoprecipitation analysis
B Immunohistochemistry analysis
B Proximity ligation assay
B Fluorescent calcium assay
B Quantitative RT-PCR and PIK3CA mutation analysis
B In vitro kinase assay
B Chemokine assays
B Lipid extraction and eicosanoid profiling
B Bioinformatic analysis
d QUANTIFICATION AND STATISTICAL ANALYSIS
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.053.
A video abstract is available at https://doi.org/10.1016/j.cell.2020.05.
053#mmc8.
ACKNOWLEDGMENTS
We thank Naomi Guppy and Farzana Noor (Breast Cancer Now Histopatholo-
gy, ICR, London, UK) and Elena Miranda and Adriana Resende Alves (Pathol-
ogy Core Facility, University College London Cancer Institute) for support with
immunohistochemistry, hematoxylin, and eosin analysis; and Champions
Oncology (London, UK) for kindly providing breast, pancreatic, ovarian, sar-
coma, and colorectal cancer PDX tumor specimens. We would also like to
thank Edward St. John for enabling access to primary breast tumor samples,
Verena M. Horneffer-van der Sluis for assistance with eicosanoid analysis, and
the Biological Services Unit staff at the Institute of Cancer Research (Chelsea
site) for their assistance with in vivo experiments. N.K. was supported by an
ICR PhD studentship. The work described and the laboratory of G.P. was sup-
ported by the Institute of Cancer Research and a Cancer Research UK Grand
Challenge award (C59824/A25044). Work in the Z.T. lab was supported by the
European Research Council (MASSLIP Consolidator grant), Cancer Research
UK Grand Challenge award (C59824/A25044), and the National Institute for
Health Research (Imperial Biomedical Research Centre). A.V. was funded by
the Ministry of Education, Culture and Sport under the Program for Promoting
and Hiring of Talent and its Employability (Subprogram for Mobility ‘‘José Cas-
tillejo’’) of the Spanish Government and by Comunitat Autònoma de les Illes
Balears, Direcció General d’lnnovació i Recerca (AAEE003/2017) and Fons Eu-
ropeu de Desenvolupament Regional de la Unió Europea (FEDER).
AUTHOR CONTRIBUTIONS
N.K., Z.T., and G.P. designed the study with contributions from J.K.N. and
R.C.G. Z.T. and G.P. directed the project and supervised data analysis. N.K.
performed and analyzed most experiments. E.K. performed 3-D acini, prox-
imity ligation assays, calcium flux analyses, the CRISPR knockdown genera-
tion, the site-directed mutagenesis constructs, and assisted with xenograft
studies. A.T. performed in vitro kinase assays, ELISAs, and assisted with xeno-
graft studies and REIMS analysis of tumor samples. A.V. assisted with estro-
gen receptor signaling experiments, the generation of dox-inducible cPLA2-
knockdown cell lines, and lipid extractions for LC-MS analysis. P.I. developed
essential platforms for pre-processing and interpretation of REIMS data.
N.J.S.P. assisted with xenograft studies. D.J.M. assisted with chemokine as-
says. S.A.V. and G.A.E. cultured and analyzed cell lines grown under different
microenvironment conditions. A.L.T. carried out image analysis. M.L.D. assis-
ll
OPEN ACCESS
ted with LC-MS profiling of eicosanoids. A.v.W. and C.M.I. helped with animal
experiments. R.F.S. provided essential technical support for sample process-
ing with REIMS. N.K. and G.P. wrote the manuscript. All authors edited the
manuscript.
DECLARATION OF INTERESTS
N.K. and G.P. are inventors on a patent application covering new methods and
compositions useful in the treatment of cancers with PIK3CA mutation (appli-
cation number GB 2005874.9).
Received: August 5, 2019
Revised: March 7, 2020
Accepted: May 28, 2020
Published: June 18, 2020
REFERENCES
Alexander, L.D., Ding, Y., Alagarsamy, S., Cui, X.L., and Douglas, J.G. (2006).
Arachidonic acid induces ERK activation via Src SH2 domain association with
the epidermal growth factor receptor. Kidney Int. 69, 1823–1832.
Alexander, J., Gildea, L., Balog, J., Speller, A., McKenzie, J., Muirhead, L.,
Scott, A., Kontovounisios, C., Rasheed, S., Teare, J., et al. (2017). A novel
methodology for in vivo endoscopic phenotyping of colorectal cancer based
on real-time analysis of the mucosal lipidome: a prospective observational
study of the iKnife. Surg. Endosc. 31, 1361–1370.
Ambs, P., Baccarini, M., Fitzke, E., and Dieter, P. (1995). Role of cytosolic
phospholipase A2 in arachidonic acid release of rat-liver macrophages: regu-
lation by Ca2+ and phosphorylation. Biochem. J. 311, 189–195.
Balog, J., Sasi-Szabó, L., Kinross, J., Lewis, M.R., Muirhead, L.J., Veselkov,
K., Mirnezami, R., Dezso } , B., Damjanovich, L., Darzi, A., et al. (2013). Intraoper-
ative tissue identification using rapid evaporative ionization mass spectrom-
etry. Sci. Transl. Med. 5, 194ra93.
Basingab, F.S., Ahmadi, M., and Morgan, D.J. (2016). IFNg-Dependent Inter-
actions between ICAM-1 and LFA-1 Counteract Prostaglandin E2-Mediated
Inhibition of Antitumor CTL Responses. Cancer Immunol. Res. 4, 400–411.
Bligh, E.G., and Dyer, W.J. (1959). A rapid method of total lipid extraction and
purification. Can. J. Biochem. Physiol. 37, 911–917.
Borcoman, E., De La Rochere, P., Richer, W., Vacher, S., Chemlali, W.,
Krucker, C., Sirab, N., Radvanyi, F., Allory, Y., Pignot, G., et al. (2019). Inhibi-
tion of PI3K pathway increases immune infiltrate in muscle-invasive bladder
cancer. OncoImmunology 8, e1581556.
Böttcher, J.P., Bonavita, E., Chakravarty, P., Blees, H., Cabeza-Cabrerizo, M.,
Sammicheli, S., Rogers, N.C., Sahai, E., Zelenay, S., and Reis E Sousa, C.
(2018). NK Cells Stimulate Recruitment of cDC1 into the Tumor Microenviron-
ment Promoting Cancer Immune Control. Cell 172, 1022–1037.e14.
Burke, J.E., and Dennis, E.A. (2009). Phospholipase A2 structure/function,
mechanism, and signaling. J. Lipid Res. 50 (Suppl ), S237–S242.
Carracedo, A., and Pandolfi, P.P. (2008). The PTEN-PI3K pathway: of feed-
backs and cross-talks. Oncogene 27, 5527–5541.
Chambers, M.C., Maclean, B., Burke, R., Amodei, D., Ruderman, D.L., Neu-
mann, S., Gatto, L., Fischer, B., Pratt, B., Egertson, J., et al. (2012). A cross-
platform toolkit for mass spectrometry and proteomics. Nat. Biotechnol. 30,
918–920.
Chou, M.M., Hou, W., Johnson, J., Graham, L.K., Lee, M.H., Chen, C.S.,
Newton, A.C., Schaffhausen, B.S., and Toker, A. (1998). Regulation of protein
kinase C zeta by PI 3-kinase and PDK-1. Curr. Biol. 8, 1069–1077.
Clark, J.D., Schievella, A.R., Nalefski, E.A., and Lin, L.L. (1995). Cytosolic
phospholipase A2. J. Lipid Mediat. Cell Signal. 12, 83–117.
Dan, S., Okamura, M., Seki, M., Yamazaki, K., Sugita, H., Okui, M., Mukai, Y.,
Nishimura, H., Asaka, R., Nomura, K., et al. (2010). Correlating phosphatidyli-
nositol 3-kinase inhibitor efficacy with signaling pathway status: in silico and
biological evaluations. Cancer Res. 70, 4982–4994.
Cell 181, 1596–1611, June 25, 2020 1609
 ll
OPEN ACCESS
Debnath, J., Muthuswamy, S.K., and Brugge, J.S. (2003). Morphogenesis and
oncogenesis of MCF-10A mammary epithelial acini grown in three-dimen-
sional basement membrane cultures. Methods 30, 256–268.
Dennis, E.A., and Norris, P.C. (2015). Eicosanoid storm in infection and inflam-
mation. Nat. Rev. Immunol. 15, 511–523.
Dibble, C.C., and Manning, B.D. (2013). Signal integration by mTORC1 coor-
dinates nutrient input with biosynthetic output. Nat. Cell Biol. 15, 555–564.
Düvel, K., Yecies, J.L., Menon, S., Raman, P., Lipovsky, A.I., Souza, A.L., Tri-
antafellow, E., Ma, Q., Gorski, R., Cleaver, S., et al. (2010). Activation of a
metabolic gene regulatory network downstream of mTOR complex 1. Mol.
Cell 39, 171–183.
Fruman, D.A., Chiu, H., Hopkins, B.D., Bagrodia, S., Cantley, L.C., and
Abraham, R.T. (2017). The PI3K Pathway in Human Disease. Cell 170,
605–635.
Gavaghan, C.L., Holmes, E., Lenz, E., Wilson, I.D., and Nicholson, J.K. (2000).
An NMR-based metabonomic approach to investigate the biochemical conse-
quences of genetic strain differences: application to the C57BL10J and Alp-
k:ApfCD mouse. FEBS Lett. 484, 169–174.
Gibb, S., and Strimmer, K. (2012). MALDIquant: a versatile R package for the
analysis of mass spectrometry data. Bioinformatics 28, 2270–2271.
Greene, E.R., Huang, S., Serhan, C.N., and Panigrahy, D. (2011). Regulation of
inflammation in cancer by eicosanoids. Prostaglandins Other Lipid Mediat.
96, 27–36.
Griffiths, B., Lewis, C.A., Bensaad, K., Ros, S., Zhang, Q., Ferber, E.C., Konisti,
S., Peck, B., Miess, H., East, P., et al. (2013). Sterol regulatory element binding
protein-dependent regulation of lipid synthesis supports cell survival and tu-
mor growth. Cancer Metab. 1, 3.
Guri, Y., Colombi, M., Dazert, E., Hindupur, S.K., Roszik, J., Moes, S., Jenoe,
P., Heim, M.H., Riezman, I., Riezman, H., and Hall, M.N. (2017). mTORC2 Pro-
motes Tumorigenesis via Lipid Synthesis. Cancer Cell 32, 807–823.e12.
Henry, W.S., Laszewski, T., Tsang, T., Beca, F., Beck, A.H., McAllister, S.S.,
and Toker, A. (2017). Aspirin Suppresses Growth in PI3K-Mutant Breast Can-
cer by Activating AMPK and Inhibiting mTORC1 Signaling. Cancer Res. 77,
790–801.
Hilvo, M., Denkert, C., Lehtinen, L., Müller, B., Brockmöller, S., Seppänen-
Laakso, T., Budczies, J., Bucher, E., Yetukuri, L., Castillo, S., et al. (2011).
Novel theranostic opportunities offered by characterization of altered mem-
brane lipid metabolism in breast cancer progression. Cancer Res. 71,
3236–3245.
Holmes, E., Wilson, I.D., and Nicholson, J.K. (2008). Metabolic phenotyping in
health and disease. Cell 134, 714–717.
Hopkins, B.D., Pauli, C., Du, X., Wang, D.G., Li, X., Wu, D., Amadiume, S.C.,
Goncalves, M.D., Hodakoski, C., Lundquist, M.R., et al. (2018). Suppression
of insulin feedback enhances the efficacy of PI3K inhibitors. Nature 560,
499–503.
Huang, W., Sherman, B.T., and Lempicki, R.A. (2009). Systematic and integra-
tive analysis of large gene lists using DAVID bioinformatics resources. Nat.
Protoc. 4, 44–57.
Hughes-Fulford, M., Li, C.F., Boonyaratanakornkit, J., and Sayyah, S. (2006).
Arachidonic acid activates phosphatidylinositol 3-kinase signaling and in-
duces gene expression in prostate cancer. Cancer Res. 66, 1427–1433.
Janiszewska, M., Liu, L., Almendro, V., Kuang, Y., Paweletz, C., Sakr, R.A.,
Weigelt, B., Hanker, A.B., Chandarlapaty, S., King, T.A., et al. (2015). In situ sin-
gle-cell analysis identifies heterogeneity for PIK3CA mutation and HER2 ampli-
fication in HER2-positive breast cancer. Nat. Genet. 47, 1212–1219.
Kanai, S., Ishihara, K., Kawashita, E., Tomoo, T., Nagahira, K., Hayashi, Y., and
Akiba, S. (2016). ASB14780, an Orally Active Inhibitor of Group IVA Phospho-
lipase A2, Is a Pharmacotherapeutic Candidate for Nonalcoholic Fatty Liver
Disease. J. Pharmacol. Exp. Ther. 356, 604–614.
Kramer, R.M., Roberts, E.F., Um, S.L., Börsch-Haubold, A.G., Watson, S.P.,
Fisher, M.J., and Jakubowski, J.A. (1996). p38 Mitogen-activated protein ki-
nase phosphorylates cytosolic phospholipase A2 (cPLA2) in thrombin-stimu-
lated platelets. Evidence that proline-directed phosphorylation is not required
1610 Cell 181, 1596–1611, June 25, 2020
Article
for mobilization of arachidonic acid by cPLA2. J. Biol. Chem. 271,
27723–27729.
Lala, P.K., Nandi, P., and Majumder, M. (2018). Roles of prostaglandins in tu-
mor-associated lymphangiogenesis with special reference to breast cancer.
Cancer Metastasis Rev. 37, 369–384.
Lee, K.L., Foley, M.A., Chen, L., Behnke, M.L., Lovering, F.E., Kirincich, S.J.,
Wang, W., Shim, J., Tam, S., Shen, M.W., et al. (2007). Discovery of Ecopladib,
an indole inhibitor of cytosolic phospholipase A2alpha. J. Med. Chem. 50,
1380–1400.
Lee, S.J., Kang, W.Y., Yoon, Y., Jin, J.Y., Song, H.J., Her, J.H., Kang, S.M.,
Hwang, Y.K., Kang, K.J., Joo, K.M., and Nam, D.H. (2015). Natural killer (NK)
cells inhibit systemic metastasis of glioblastoma cells and have therapeutic ef-
fects against glioblastomas in the brain. BMC Cancer 15, 1011.
Lee, G., Zheng, Y., Cho, S., Jang, C., England, C., Dempsey, J.M., Yu, Y., Liu,
X., He, L., Cavaliere, P.M., et al. (2017). Post-transcriptional Regulation of De
Novo Lipogenesis by mTORC1-S6K1-SRPK2 Signaling. Cell 171, 1545–1558.
Li, X., and Gao, T. (2014). mTORC2 phosphorylates protein kinase Cz to regu-
late its stability and activity. EMBO Rep. 15, 191–198.
Li, X., Dai, D., Chen, B., Tang, H., Xie, X., and Wei, W. (2018). Efficacy of PI3K/
AKT/mTOR pathway inhibitors for the treatment of advanced solid cancers: A
literature-based meta-analysis of 46 randomised control trials. PLoS ONE 13,
e0192464.
Liao, X., Lochhead, P., Nishihara, R., Morikawa, T., Kuchiba, A., Yamauchi, M.,
Imamura, Y., Qian, Z.R., Baba, Y., Shima, K., et al. (2012). Aspirin use, tumor
PIK3CA mutation, and colorectal-cancer survival. N. Engl. J. Med. 367,
1596–1606.
Lien, E.C., Lyssiotis, C.A., and Cantley, L.C. (2016). Metabolic Reprogramming
by the PI3K-Akt-mTOR Pathway in Cancer. Recent Results Cancer Res.
207, 39–72.
Lin, L.L., Wartmann, M., Lin, A.Y., Knopf, J.L., Seth, A., and Davis, R.J. (1993).
cPLA2 is phosphorylated and activated by MAP kinase. Cell 72, 269–278.
McKew, J.C., Lee, K.L., Shen, M.W., Thakker, P., Foley, M.A., Behnke, M.L.,
Hu, B., Sum, F.W., Tam, S., Hu, Y., et al. (2008). Indole cytosolic phospholipase
A2 alpha inhibitors: discovery and in vitro and in vivo characterization of 4-{3-
[5-chloro-2-(2-{[(3,4-dichlorobenzyl)sulfonyl]amino}ethyl)-1-(diphenylmethyl)-
1H-indol-3-yl]propyl}benzoic acid, efipladib. J. Med. Chem. 51, 3388–3413.
Nicholson, J.K., Connelly, J., Lindon, J.C., and Holmes, E. (2002). Metabo-
nomics: a platform for studying drug toxicity and gene function. Nat. Rev.
Drug Discov. 1, 153–161.
Nicholson, J.K., Holmes, E., Kinross, J.M., Darzi, A.W., Takats, Z., and Lindon,
J.C. (2012). Metabolic phenotyping in clinical and surgical environments. Na-
ture 491, 384–392.
Okada, S., Vaeteewoottacharn, K., and Kariya, R. (2019). Application of Highly
Immunocompromised Mice for the Establishment of Patient-Derived Xeno-
graft (PDX) Models. Cells 8, 889.
Okkenhaug, K., Graupera, M., and Vanhaesebroeck, B. (2016). Targeting PI3K
in Cancer: Impact on Tumor Cells, Their Protective Stroma, Angiogenesis, and
Immunotherapy. Cancer Discov. 6, 1090–1105.
Otsu, N. (1979). A Threshold Selection Method from Gray-Level Histograms.
IEEE Trans. Syst. Man Cybern. 9, 62–66.
Patterson, E., Wall, R., Fitzgerald, G.F., Ross, R.P., and Stanton, C. (2012).
Health implications of high dietary omega-6 polyunsaturated Fatty acids.
J. Nutr. Metab. 2012, 539426.
Rameh, L.E., Rhee, S.G., Spokes, K., Kazlauskas, A., Cantley, L.C., and Cant-
ley, L.G. (1998). Phosphoinositide 3-kinase regulates phospholipase
Cgamma-mediated calcium signaling. J. Biol. Chem. 273, 23750–23757.
Ricoult, S.J.H., Yecies, J.L., Ben-Sahra, I., and Manning, B.D. (2016). Onco-
genic PI3K and K-Ras stimulate de novo lipid synthesis through mTORC1
and SREBP. Oncogene 35, 1250–1260.
Sanjana, N.E., Shalem, O., and Zhang, F. (2014). Improved vectors and
genome-wide libraries for CRISPR screening. Nat. Methods 11, 783–784.

Article
Sarbassov, D.D., Ali, S.M., Sengupta, S., Sheen, J.H., Hsu, P.P., Bagley, A.F.,
Markhard, A.L., and Sabatini, D.M. (2006). Prolonged rapamycin treatment in-
hibits mTORC2 assembly and Akt/PKB. Mol. Cell 22, 159–168.
Saxton, R.A., and Sabatini, D.M. (2017). mTOR Signaling in Growth, Meta-
bolism, and Disease. Cell 168, 960–976.
Simopoulos, A.P. (2008). The importance of the omega-6/omega-3 fatty acid
ratio in cardiovascular disease and other chronic diseases. Exp. Biol. Med.
(Maywood) 233, 674–688.
Sonnweber, T., Pizzini, A., Nairz, M., Weiss, G., and Tancevski, I. (2018).
Arachidonic Acid Metabolites in Cardiovascular and Metabolic Diseases. Int.
J. Mol. Sci. 19, 3285.
St John, E.R., Balog, J., McKenzie, J.S., Rossi, M., Covington, A., Muirhead,
L., Bodai, Z., Rosini, F., Speller, A., Shousha, S., et al. (2017). Rapid evapora-
tive ionisation mass spectrometry of electrosurgical vapours for the identifica-
tion of breast pathology: towards an intelligent knife for breast cancer surgery.
Breast Cancer Res. 19.
Tarrado-Castellarnau, M., de Atauri, P., and Cascante, M. (2016). Oncogenic
regulation of tumor metabolic reprogramming. Oncotarget 7, 62726–62753.
Tomoo, T., Nakatsuka, T., Katayama, T., Hayashi, Y., Fujieda, Y., Terakawa,
M., and Nagahira, K. (2014). Design, synthesis, and biological evaluation of
ll
OPEN ACCESS
3-(1-Aryl-1H-indol-5-yl)propanoic acids as new indole-based cytosolic phos-
pholipase A2a inhibitors. J. Med. Chem. 57, 7244–7262.
Vander Heiden, M.G., and DeBerardinis, R.J. (2017). Understanding the Inter-
sections between Metabolism and Cancer Biology. Cell 168, 657–669.
Wang, D., and Dubois, R.N. (2010). Eicosanoids and cancer. Nat. Rev. Cancer
10, 181–193.
Wiederschain, D., Wee, S., Chen, L., Loo, A., Yang, G., Huang, A., Chen, Y.,
Caponigro, G., Yao, Y.M., Lengauer, C., et al. (2009). Single-vector inducible
lentiviral RNAi system for oncology target validation. Cell Cycle 8, 498–504.
Wolfer, A.M., Gaudin, M., Taylor-Robinson, S.D., Holmes, E., and Nicholson,
J.K. (2015). Development and Validation of a High-Throughput Ultrahigh-Per-
formance Liquid Chromatography-Mass Spectrometry Approach for
Screening of Oxylipins and Their Precursors. Anal. Chem. 87, 11721–11731.
Yu, G., and He, Q.-Y. (2016). ReactomePA: an R/Bioconductor package for re-
actome pathway analysis and visualization. Mol. Biosyst. 12, 477–479.
Zelenay, S., van der Veen, A.G., Böttcher, J.P., Snelgrove, K.J., Rogers, N.,
Acton, S.E., Chakravarty, P., Girotti, M.R., Marais, R., Quezada, S.A., et al.
(2015). Cyclooxygenase-Dependent Tumor Growth through Evasion of Immu-
nity. Cell 162, 1257–1270.
Cell 181, 1596–1611, June 25, 2020 1611
 ll
OPEN ACCESS
STAR+METHODS
KEY RESOURCES TABLE
REAGENT or RESOURCE
Antibodies
Rabbit polyclonal anti-beta-actin
Rabbit polyclonal anti-Akt antibody
Rabbit monoclonal anti-phospho-Akt
(Ser473)
Rabbit polyclonal anti-PRAS40
Rabbit monoclonal anti-phospho-PRAS40
(Thr246)
Mouse monoclonal anti-S6 Ribosomal
Protein
Rabbit polyclonal anti-phospho-S6
Ribosomal Protein (Ser235/236)
Rabbit polyclonal anti-NDRG1
Rabbit monoclonal anti-phospho-NDRG1
(Thr376)
Rabbit polyclonal anti-Raptor
Rabbit polyclonal anti-Rictor
Rabbit monoclonal anti-mTOR
Rabbit polyclonal anti-P70 S6 Kinase
Rabbit polyclonal anti-phospho-p70 S6
Kinase (Thr389)
Rabbit polyclonal anti-4E-BP1 (53H11)
Rabbit polyclonal anti-phospho-4E-
BP1 (Ser65)
Rabbit monoclonal anti-Lamin B1 (D4Q4Z)
Rabbit polyclonal anti-phospho-PKC (pan)
(bII Ser660)
Rabbit polyclonal anti-PKCdelta
Rabbit polyclonal anti-phospho-PKCdelta/
theta (Ser643/676)
Rabbit polyclonal anti-RKIP (G38)
Rabbit polyclonal anti-c-Raf
Rabbit monoclonal anti-phospho-c-Raf
(Ser338)
Rabbit polyclonal anti-MEK1/2
Rabbit polyclonal anti-phospho-MEK1/2
(Ser217/Ser221)
Rabbit polyclonal anti-p44/42 MAPK
(Erk1/2)
Rabbit polyclonal anti-phospho-p44/42
MAPK (Thr202/Tyr204)
Rabbit polyclonal anti-p38 MAPK
Rabbit polyclonal anti-phospho-p38 MAPK
(Thr180/Tyr182)
Rabbit polyclonal anti-cPLA2
e1 Cell 181, 1596–1611.e1–e16, June 25, 2020
Article
SOURCE IDENTIFIER
Cell Signaling Technology Cat# 4967; RRID: AB_10695744
Cell Signaling Technology Cat# 9272; RRID: AB_329827
Cell Signaling Technology Cat# 4060; RRID: AB_2315049
Cell Signaling Technology Cat# 2610; RRID: AB_916206
Cell Signaling Technology Cat# 2997; RRID: AB_2258110
Cell Signaling Technology Cat# 2317; RRID: AB_2238583
Cell Signaling Technology Cat# 2211; RRID: AB_331679
Cell Signaling Technology Cat# 5196; RRID: AB_10626626
Cell Signaling Technology Cat# 5482; RRID: AB_10693451
Cell Signaling Technology Cat# 2280; RRID: AB_561245
Cell Signaling Technology Cat# 2140; RRID: AB_2179961
Cell Signaling Technology Cat# 2983; RRID: AB_2105622
Cell Signaling Technology Cat# 9202; RRID: AB_331676
Cell Signaling Technology Cat# 9205; RRID: AB_330944
Cell Signaling Technology Cat# 9644; RRID: AB_2097841
Cell Signaling Technology Cat# 9451; RRID: AB_330947
Cell Signaling Technology Cat# 12586; RRID: AB_2650517
Cell Signaling Technology Cat# 9371; RRID: AB_2168219
Cell Signaling Technology Cat# 2058; RRID: AB_10694655
Cell Signaling Technology Cat# 9376; RRID: AB_2168834
Cell Signaling Technology Cat# 5060; RRID: AB_1904081
Cell Signaling Technology Cat# 9422; RRID: AB_390808
Cell Signaling Technology Cat# 9427; RRID: AB_2067317
Cell Signaling Technology Cat# 9122; RRID: AB_823567
Cell Signaling Technology Cat# 9121; RRID: AB_331648
Cell Signaling Technology Cat# 9102; RRID: AB_330744
Cell Signaling Technology Cat# 9101; RRID: AB_331646
Cell Signaling Technology Cat# 9212; RRID: AB_330713
Cell Signaling Technology Cat# 9211; RRID: AB_331641
Cell Signaling Technology Cat# 2832; RRID: AB_2164442
(Continued on next page)

Article
Continued
REAGENT or RESOURCE
Rabbit polyclonal anti-phospho-cPLA2
(Ser505)
Rabbit monoclonal anti-HA-Tag
Rabbit monoclonal anti-GAPDH
Rabbit polyclonal anti-PLCgamma1
Rabbit polyclonal anti-phospho-
PLCgamma1 (Tyr783)
Rabbit polyclonal anti-phospho-threonine
Rabbit polyclonal anti-IkBalpha
Rabbit monoclonal anti-phospho-
IkBalpha (Ser32)
Rabbit monoclonal anti-Stat3
Rabbit polyclonal anti-phospho-Stat3
(Ser727)
Rabbit monoclonal anti-PKD/PKCm
Rabbit polyclonal anti-phospho-PKD/PKCm
(Ser744/748)
Mouse monoclonal anti-Rb (4H1)
Rabbit monoclonal anti-estrogen inducible
protein pS2
Rabbit polyclonal anti-PKCzeta
Rabbit monoclonal anti-phospho-PKCzeta
(Thr560)
Rabbit polyclonal anti-PKCepsilon
Rabbit polyclonal anti-phospho-
PKCepsilon (Ser729)
Rabbit monoclonal anti-secretory
phospholipase A2
Mouse monoclonal anti-PKCbeta II (F-7)
Mouse monoclonal anti-PKCzeta (B-7)
Mouse monoclonal anti-phospho-RKIP
(Ser153)
Normal rabbit IgG
Mouse monoclonal anti-phospho-Rb
(Thr821/826)
Rabbit polyclonal anti-phospholipase
A2 (iPLA2)
Rabbit monoclonal anti-phospho-PKCzeta
(Thr410)
Mouse monoclonal anti-SREBP1
Goat anti-rabbit IgG (H+L)-HRP conjugate
Goat anti-mouse IgG (H+L)-HRP conjugate
Mouse polyclonal anti-Nkp46/ncr1
Rabbit monoclonal anti-Ki-67 (D3B5)
Goat anti-Mouse IgG (H+L) secondary
antibody, Alexa Fluor 546
Phalloidin-iFluor 633
Rabbit polyclonal anti-phospho-cPLA2
(Thr376) (Peptide name: PLA2G4A-
369:383-pT376)
SOURCE
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Cell Signaling Technology
Abcam
Abcam
Abcam
Abcam
Abcam
Abcam
Santa Cruz Biotechnology
Santa Cruz Biotechnology
Santa Cruz Biotechnology
Santa Cruz Biotechnology
Santa Cruz Biotechnology
Sigma-Aldrich
Thermo Fisher Scientific
BD Biosciences
Bio-Rad
Bio-Rad
R and D systems
Cell Signaling Technology
Thermo Fisher Scientific
Abcam
This paper (produced by Thermo Fisher
Scientific)
Cell 181, 1596–1611.e1–e16, June 25, 2020 e2
ll
OPEN ACCESS
IDENTIFIER
Cat# 2831; RRID: AB_2164445
Cat# 3724; RRID: AB_1549585
Cat# 2118; RRID: AB_561053
Cat# 2822; RRID: AB_2163702
Cat# 2821; RRID: AB_330855
Cat# 9381; RRID: AB_330301
Cat# 9242; RRID: AB_331623
Cat# 2859; RRID: AB_561111
Cat# 4904; RRID: AB_331269
Cat# 9134; RRID: AB_331589
Cat# 90039; RRID: AB_2800149
Cat# 2054; RRID: AB_2172539
Cat# 9309; RRID: AB_823629
Cat# ab92377; RRID: AB_10562122
Cat# ab59364; RRID: AB_944858
Cat# ab62372; RRID: AB_946309
Cat# ab63638; RRID: AB_1142276
Cat# ab63387; RRID: AB_1142277
Cat# ab139692
Cat# sc-13149; RRID: AB_628144
Cat# sc-393218
Cat# sc-135779; RRID: AB_2163163
Cat# sc-2027; RRID: AB_737197
Cat# sc-271930; RRID: AB_670923
Cat# SAB4200129; RRID: AB_11129638
Cat# MA5-15060; RRID: AB_10983263
Cat#557036; RRID: AB_396559
Cat#170-6515; RRID: AB_11125142
Cat#170-6516; RRID: AB_11125547
Cat#AF2225; RRID: AB_355192
Cat# 9129; RRID: AB_2687446
Cat# A-11003; RRID: AB_2534071
Cat# ab176758
Cat# UE1820P-T-AB1792
(Continued on next page)
 ll
OPEN ACCESS Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER

Bacterial and Virus Strains

MAX Efficiency DH5a Competent Cells Thermo Fisher Scientific Cat#18258012
Biological Samples
Primary human breast tissue Imperial College Tissue Bank Cat# IKB180
Primary human breast tissue Imperial College Tissue Bank Cat# IKB83
Primary human breast tissue Imperial College Tissue Bank Cat# IKB092
Primary human breast tissue Imperial College Tissue Bank Cat# IKB519
Primary human breast tissue Imperial College Tissue Bank Cat# IKB512
Primary human breast tissue Imperial College Tissue Bank Cat# IKB367
Primary human breast tissue Imperial College Tissue Bank Cat# IKB541
Primary human breast tissue Imperial College Tissue Bank Cat# IKB383
Primary human breast tissue Imperial College Tissue Bank Cat# IKB323
Primary human breast tissue Imperial College Tissue Bank Cat# IKB210
Primary human breast tissue Imperial College Tissue Bank Cat# IKB334
Primary human breast tissue Imperial College Tissue Bank Cat# IKB283
Breast PDX model CrownBioscience Cat# BR6695
Breast PDX model CrownBioscience Cat# BR5020
Breast PDX model CrownBioscience Cat# BR5009
Breast PDX model CrownBioscience Cat# BR5022
Breast PDX model CrownBioscience Cat# BR1282
Breast PDX model CrownBioscience Cat# BR5011
Breast PDX model CrownBioscience Cat# BR5014
Breast PDX model CrownBioscience Cat# BR5012
Breast PDX model CrownBioscience Cat# BR5017
Breast PDX model CrownBioscience Cat# BR5013
Breast PDX model CrownBioscience Cat# BR1474
Breast PDX model CrownBioscience Cat# BR3267
Breast PDX model CrownBioscience Cat# BR5010
Breast PDX model CrownBioscience Cat# BR5015
Breast PDX model CrownBioscience Cat# BR5337
Breast PDX model CrownBioscience Cat# BR1115
Breast PDX model CrownBioscience Cat# BR1283
Breast PDX model CrownBioscience Cat# BR1458
Breast PDX model Champions Oncology Cat# CTG1059
Breast PDX model Champions Oncology Cat# CTG1350
Breast PDX model Champions Oncology Cat# CTG1941
Breast PDX model Champions Oncology Cat# CTG2308
Breast PDX model Champions Oncology Cat# CTG0033
Breast PDX model Champions Oncology Cat# CTG0012
Breast PDX model Champions Oncology Cat# CTG0017
Breast PDX model Champions Oncology Cat# CTG0018
Breast PDX model Champions Oncology Cat# CTG0437
Breast PDX model Champions Oncology Cat# CTG0473
Ovarian PDX model Champions Oncology Cat# CTG0253
Ovarian PDX model Champions Oncology Cat# CTG1423
Ovarian PDX model Champions Oncology Cat# CTG1602
Ovarian PDX model Champions Oncology Cat# CTG1627
(Continued on next page)

e3 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Ovarian PDX model Champions Oncology Cat# CTG0252
Ovarian PDX model Champions Oncology Cat# CTG0258
Ovarian PDX model Champions Oncology Cat# CTG0259
Ovarian PDX model Champions Oncology Cat# CTG0486
Pancreatic PDX model Champions Oncology Cat# CTG0292
Pancreatic PDX model Champions Oncology Cat# CTG0381
Pancreatic PDX model Champions Oncology Cat# CTG0391
Pancreatic PDX model Champions Oncology Cat# CTG1485
Pancreatic PDX model Champions Oncology Cat# CTG2205
Pancreatic PDX model Champions Oncology Cat# CTG0282
Pancreatic PDX model Champions Oncology Cat# CTG0283
Pancreatic PDX model Champions Oncology Cat# CTG0284
Pancreatic PDX model Champions Oncology Cat# CTG0285
Pancreatic PDX model Champions Oncology Cat# CTG0286
Sarcoma PDX model Champions Oncology Cat# CTG0886
Sarcoma PDX model Champions Oncology Cat# CTG1084
Sarcoma PDX model Champions Oncology Cat# CTG1116
Sarcoma PDX model Champions Oncology Cat# CTG1255
Sarcoma PDX model Champions Oncology Cat# CTG1628
Sarcoma PDX model Champions Oncology Cat# CTG0142
Sarcoma PDX model Champions Oncology Cat# CTG0143
Sarcoma PDX model Champions Oncology Cat# CTG0241
Sarcoma PDX model Champions Oncology Cat# CTG0242
Sarcoma PDX model Champions Oncology Cat# CTG0243
Colorectal PDX model Champions Oncology Cat# CTG0083
Colorectal PDX model Champions Oncology Cat# CTG0129
Colorectal PDX model Champions Oncology Cat# CTG0360
Colorectal PDX model Champions Oncology Cat# CTG0706
Colorectal PDX model Champions Oncology Cat# CTG0799
Colorectal PDX model Champions Oncology Cat# CTG0058
Colorectal PDX model Champions Oncology Cat# CTG0062
Colorectal PDX model Champions Oncology Cat# CTG0063
Colorectal PDX model Champions Oncology Cat# CTG0065
Colorectal PDX model Champions Oncology Cat# CTG0066
Chemicals, Peptides, and Recombinant Proteins
Hydrocortisone Sigma-Aldrich Cat# H-0888
Insulin-transferrin-selenium GIBCO Cat# 41400-045
Epidermal growth factor (EGF) Peprotech Cat# AF-100-15
Cholera toxin Sigma-Aldrich Cat# C-8052
Insulin Sigma-Aldrich Cat# I-1882
Fatty acid free bovine serum albumin Sigma-Aldrich Cat# A8806
PKC zeta pseudosubstrate inhibitory Sigma-Aldrich Cat# P1614
peptide
PKC beta II peptide inhibitor Sigma-Aldrich Cat# P-0102
FIPI hydrochloride hydrate Sigma-Aldrich Cat# F5808
4-hydroxytamoxifen Sigma-Aldrich Cat# T176
Bromoenol lactone Sigma-Aldrich Cat# B1552
Cycloheximide Sigma-Aldrich Cat# C7698
(Continued on next page)

Cell 181, 1596–1611.e1–e16, June 25, 2020 e4

 ll
OPEN ACCESS Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Doxycycline hyclate Sigma-Aldrich Cat# D9891
Arachidonic acid Sigma-Aldrich Cat# 10931
Palmitoleate Sigma-Aldrich Cat# P9417
Palmitate Sigma-Aldrich Cat# P0500
PKC alpha (C2-4) inhibitor peptide Santa Cruz Biotechnology Cat# sc-304
PKC epsilon inhibitor peptide Cambridge Bioscience Cat# CAY17476
Rapamycin Selleckchem Cat# S1039
Torin 1 Selleckchem Cat#S2827
BYL719 Selleckchem Cat# S2814
BKM120 Selleckchem Cat# 2247
MK2206 Selleckchem Cat# S1078
GSK690693 Selleckchem Cat# S1113
GSK650394 Tocris Bioscience Cat# 3572
U73122 Tocris Bioscience Cat# 1268
ASB14780 Axon Medchem Cat# 2578
DharmaFECT-1 transfection reagent Dharmacon Cat# T-2001-02
FuGENE HD Transfection reagent Promega Cat# E2311
Lipid Removal Adsorbent Supelco Cat# 13358
Matrigel Corning Cat# 354230
Paraformaldehyde Sigma-Aldrich Cat# 158127
Triton X-100 Sigma-Aldrich Cat# X100
DAPI (4’,6-Diamidino-2-Phenylindole, Thermo Fisher Scientific Cat# D1306
Dihydrochloride)
RIPA buffer Thermo Fisher Scientific Cat# 89900
Leupeptin Sigma-Aldrich Cat# L2884
Pepstatin Sigma-Aldrich Cat# P5318
Na3VO4 Sigma-Aldrich Cat# 450243
DL-Dithiothreitol Sigma-Aldrich Cat# 646653
Calyculin A Cell Signaling Technology Cat# 9902
Beta-glycerophosphate Sigma-Aldrich Cat# G9422
PMSF protease inhibitor Cell Signaling Technology Cat# 8553
Bradford reagent Bio-Rad Cat# 5000006
ALLN protease inhibitor Merck-Millipore Cat# 208719
2x Laemmli sample buffer Bio-Rad Cat# 161-0737
4x Laemmli sample buffer Bio-Rad Cat# 161-0747
10X Cell lysis buffer Cell Signaling Technology Cat# 9803
Recombinant human protein kinase C zeta Insight Biotechnology Cat# TP302472
Recombinant human phospholipase A2, Insight Biotechnology Cat# TP320972
group IVA
Magnesium chloride Sigma-Aldrich Cat# M8266
Bovine serum albumin Sigma-Aldrich Cat# A2153
Puromycin Invivogen Cat# ant-pr-1
Blasticidin Invivogen Cat# ant-bl-1
Critical Commercial Assays
QuikChange Lightning Site-Directed Agilent Cat# 210518
Mutagenesis Kit
CellTiter96 Aqueous Non-radioactive (MTS) Promega Cat# G5430
cell proliferation assay
(Continued on next page)

e5 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Fatty Acid Uptake Kit Sigma-Aldrich Cat# MAK156
Cytosolic Phospholipase A2 Assay Kit Abcam Cat# ab133090
Secretory Phospholipase A2 Assay Kit Abcam Cat# ab133089
Diacylglycerol (DAG) Assay Kit Cell Biolabs Inc Cat# MET-5028
Active Rac1 Detection Kit Cell Signaling Technology Cat# 8815
Duolink In Situ Detection Reagents Red Kit Sigma-Aldrich Cat# DUO92008
Minus and Plus PLA probes Sigma-Aldrich Cat# DUO92004 and DUO92002
Fluo-4 Direct Calcium Assay Kit Thermo Fisher Scientific Cat# F10471
RNeasy Plus Mini Kit QIAGEN Cat# 74134
QuantiTect Reverse Transcription Kit QIAGEN Cat#205311
SYBR Select Master Mix Thermo Fisher Scientific Cat# 4472908
QIAamp DNA mini kit QIAGEN Cat# 51304
PNAClamp PIK3CA Mutation Detection Kit Panagene Cat# PNAC-4001
ADP-Glo Kinase Assay Promega Cat# V6930
Arachidonic Acid ELISA Kit Generon Cat# CEB098Ge
Prostaglandin E2 ELISA Kit Enzo Life Sciences Cat# ADI-900-001
Mouse RANTES (CCL5) ELISA Kit Abcam Cat# ab100739
Mouse Fractalkine (CX3CL1) ELISA Kit Abcam Cat# ab100683
Pierce BCA Protein Assay Kit Thermo Fisher Scientific Cat# 23225
Deposited Data
Custom script for quantification of proximity This manuscript Github: https://github.com/adamltyson/
ligation assay images foci2D
Custom script for quantification of acini This manuscript Github: https://github.com/adamltyson/
images cell-coloc-3D
Custom script for quantification of This manuscript Github: https://github.com/adamltyson/
calcium flux CalciumAnalysis
REIMS data for Figure 1D This manuscript Mendeley Data; https://doi.org/10.17632/
xcgc5kpntm.1
REIMS data for Figure 1E This manuscript Mendeley Data; https://doi.org/10.17632/
xcgc5kpntm.1
REIMS data for Figure 1H This manuscript Mendeley Data; https://doi.org/10.17632/
xcgc5kpntm.1
REIMS data for Figure 3B This manuscript Mendeley Data; https://doi.org/10.17632/
xcgc5kpntm.1
REIMS data for Figure 3D This manuscript Mendeley Data; https://doi.org/10.17632/
xcgc5kpntm.1
Significantly altered phospholipids across This manuscript Table S1
breast cancer cell lines of different receptor,
or triple negative status
Significantly different fatty acids between This manuscript Table S4
MCF10A PIK3CA wild-type and E545K/
H1047R mutant cells
Experimental Models: Cell Lines
Human PIK3CA (H1047R/+) MCF10A Horizon Discovery Cat# HD 101-011
Human PIK3CA (E545K/+) MCF10A Horizon Discovery Cat# HD 101-002
AU565 (human breast carcinoma) ATCC Cat# CRL-2351; RRID: CVCL_1074
BT20 (human breast carcinoma) ATCC Cat# HTB-19; RRID: CVCL_0178
BT474 (human breast carcinoma) ATCC Cat# HTB-20; RRID: CVCL_0179
BT549 (human breast carcinoma) ATCC Cat# HTB-122; RRID: CVCL_1092
CAL51 (human breast carcinoma) DSMZ Cat# ACC-302; RRID: CVCL_1110
(Continued on next page)

Cell 181, 1596–1611.e1–e16, June 25, 2020 e6

 ll
OPEN ACCESS
Continued
REAGENT or RESOURCE SOURCE
CAMA1 (human breast carcinoma) ATCC
EFM19 (human breast carcinoma) DSMZ
Hs578T (human breast carcinoma) ATCC
JIMT1 (human breast carcinoma) DSMZ
KPL1 (human breast carcinoma) DSMZ
MCF7 (human breast carcinoma) ATCC
MDAMB134 (human breast carcinoma) ATCC
MDAMB157 (human breast carcinoma) ATCC
MDAMB231 (human breast carcinoma) ATCC
MDAMB361 (human breast carcinoma) ATCC
MDAMB436 (human breast carcinoma) ATCC
MDAMB453 (human breast carcinoma) ATCC
MDAMB468 (human breast carcinoma) ATCC
MFM223 (human breast carcinoma) DSMZ
S68 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
SKBR3 (human breast carcinoma) ATCC
T47D (human breast carcinoma) ATCC
UACC812 (human breast carcinoma) ATCC
VP229 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
BT483 (human breast carcinoma) ATCC
HCC1143 (human breast carcinoma) ATCC
HCC1395 (human breast carcinoma) ATCC
HCC1428 (human breast carcinoma) ATCC
HCC1500 (human breast carcinoma) ATCC
HCC1569 (human breast carcinoma) ATCC
HCC1937 (human breast carcinoma) ATCC
HCC1954 (human breast carcinoma) ATCC
HCC202 (human breast carcinoma) ATCC
HCC38 (human breast carcinoma) ATCC
HCC70 (human breast carcinoma) ATCC
SUM52 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
ZR751 (human breast carcinoma) ATCC
ZR7530 (human breast carcinoma) ATCC
SUM44 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
SUM159 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
SUM149 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
SUM225 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
SUM229 (human breast carcinoma) Breast Cancer Now (Institute of Cancer
Research)
HEK293T (human embryonic kidney) ATCC
Human MCF10A PIK3CA WT CRISPR This manuscript
control cell line
e7 Cell 181, 1596–1611.e1–e16, June 25, 2020
Article
IDENTIFIER
Cat# HTB-21; RRID: CVCL_1115
Cat# ACC-231; RRID: CVCL_0253
Cat# HTB-126; RRID: CVCL_0332
Cat# ACC-589; RRID: CVCL_2077
Cat# ACC-317; RRID: CVCL_2094
Cat# HTB-22; RRID: CVCL_0031
Cat# HTB-23; RRID: CVCL_0617
Cat# HTB-24; RRID: CVCL_0618
Cat# HTB-26; RRID: CVCL_0062
Cat# HTB-27; RRID: CVCL_0620
Cat# HTB-130; RRID: CVCL_0623
Cat# HTB-131; RRID: CVCL_0418
Cat# HTB-132; RRID: CVCL_0419
Cat# ACC-422; RRID: CVCL_1408
RRID: CVCL_5585
Cat# HTB-30; RRID: CVCL_0033
Cat# HTB-133; RRID: CVCL_0553
Cat# CRL-1897; RRID: CVCL_1781
RRID: CVCL_2754
Cat# HTB-121; RRID: CVCL_2319
Cat# CRL-2321; RRID: CVCL_1245
Cat# CRL-2324; RRID: CVCL_1249
Cat# CRL-2327; RRID: CVCL_1252
Cat# CRL-2329; RRID: CVCL_1254
Cat# CRL-2330; RRID: CVCL_1255
Cat# CRL-2336; RRID: CVCL_0290
Cat# CRL-2338; RRID: CVCL_1259
Cat# CRL-2316; RRID: CVCL_2062
Cat# CRL-2314; RRID: CVCL_1267
Cat# CRL-2315; RRID: CVCL_1270
RRID: CVCL_3425
Cat# CRL-1500; RRID: CVCL_0588
Cat# CRL-1504; RRID: CVCL_1661
RRID: CVCL-3424
RRID: CVCL_5423
RRID: CVCL_3422
RRID: CVCL_5593
RRID: CVCL_5594
Cat# CRL-3216; RRID: CVCL_0063
N/A
(Continued on next page)
 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Human MCF10A PIK3CA WT cPLA2 This manuscript N/A
CRISPR cell line
Human MCF10A PIK3CA H1047R (+/) This manuscript N/A
CRISPR control cell line
Human MCF10A PIK3CA H1047R (+/
) This manuscript N/A
cPLA2 CRISPR cell line
Experimental Models: Organisms/Strains
Mouse: BALB/c nude (female, age: 6– Beijing Anikeeper Biotech (Beijing, China) N/A
8 weeks)
Mouse: BALB/c nude (female, age: 7– Envigo N/A
9 weeks)
Oligonucleotides
Primers for cPLA2 shRNA amplification Sigma-Aldrich N/A
(Forward 50 -
GTGGAAAGGACGAAACACCGGT-30 ,
Reverse 50 -
TTTGTCTCGAGGTCGAGAATTC-30
Mutagenesis primers to generate cPLA2 Sigma-Aldrich N/A
S505A (Forward 50 -
GCAAAGTCACTCAAAGGAGCCAGTG
GATAAGATGTATTG-30 , Reverse 50 -
CAATACATCTTATCCACTGGCTCCTTT
GAGTGACTTTGC-30
Mutagenesis primers to generate Sigma-Aldrich N/A
cPLA2 T376A (Forward 50 -TTCTTC
ATACTTCTTAACGACTGCTCCCATAAAAA
ATTTGCTTCCAA-30 ,
Reverse 50 -TTGGAAGCAAATTT
TTTATGGGAGCAGTCGTTAAG
AAGTATGAAGAA-30
qPCR primers for human cPLA2 (PLA2G4A) Sigma-Aldrich N/A
(Forward 50 -
GATGAAACTCTAGGGACAGCAAC-30 ,
Reverse 50 -
CTGGGCATGAGCAAACTTCAA-30
qPCR primers for human beta-actin Sigma-Aldrich N/A
(Forward 50 -
GACCCAGATCATGTTTGAGACC-30 ,
Reverse 50 -
CTTCATGAGGTAGTCAGTCAGG-30 )
Recombinant DNA
pLKO-Tet-On Vector Wiederschain et al., 2009 Addgene ID: 21915
pCMV3-HA-PLA2G4A Sino Biological Cat# HG13126-NY
RICTOR ON-TARGETplus SMARTPool Dharmacon Cat# L-016984-00-0005
human siRNA
RAPTOR ON-TARGETplus SMARTPool Dharmacon Cat# L-004107-00-0005
human siRNA
FRAP1 ON-TARGETplus SMARTPool Dharmacon Cat# L-003008-00-0005
human siRNA
PLCg1 ON-TARGETplus SMARTPool Dharmacon Cat# L-003559-00-0005
human siRNA
PRKCZ ON-TARGETplus SMARTPool Dharmacon Cat# L-003526-00-0005
human siRNA
Non-targeting siRNA control Dharmacon Cat# D-001810-01-05
(Continued on next page)

Cell 181, 1596–1611.e1–e16, June 25, 2020 e8

 ll
OPEN ACCESS Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
TRC Lentiviral eGFP shRNA positive control Dharmacon Cat# RHS4459
TRC Lentiviral Human PLA2G4A shRNA 1 Dharmacon Cat# TRCN0000050263
TRC Lentiviral Human PLAG2G4A shRNA 5 Dharmacon Cat# TRCN0000050267
LentiCRISPR v2 Sanjana et al., 2014 Addgene ID: 52961
PLA2G4A sgRNA CRISPR/Cas9 All-in-One Applied Biological Materials Cat# K1659207
Lentivector Target 2
pCMV-HA-PLA2G4A-S505A This manuscript N/A
pCMV-HA-PLA2G4A-T376A This manuscript N/A
Inducible pLKO-Tet-On-TRC Lentiviral This manuscript N/A
Human PLA2G4A shRNA 1
Software and Algorithms
R statistical software (version 3.5.1) The R Project https://www.r-project.org/
ProteoWizard MsConvert software (version Chambers et al., 2012 http://proteowizard.sourceforge.net/
3.0.11781) download.html
MALDIquant package Gibb and Strimmer, 2012 https://cran.r-project.org/web/packages/
MALDIquant/index.html
ReactomePA package Yu and He, 2016 https://bioconductor.org/packages/
release/bioc/html/ReactomePA.html
MATLAB (2014a, version 8.3.0.532) Mathworks https://www.mathworks.com/products/
matlab.html
GraphPad Prism (version 8.0.1) GraphPad https://www.graphpad.com/
scientific-software/prism/
Database for Annotation, Visualization and Huang et al., 2009 https://david.ncifcrf.gov/
Integrated Discovery (DAVID, version 6.8)
TargetLynx software Waters Corporation https://www.waters.com/waters/
home.htm
Image Lab Software (version 5.2.1) Bio-Rad https://www.bio-rad.com/en-uk/product/
image-lab-software?ID=KRE6P5E8Z
ImageJ (version 1.51) NIH https://imagej.nih.gov/ij/download.html
Other
Protein G Sepharose beads Sigma-Aldrich Cat# P3296
Dulbecco’s Modified Eagle’s GIBCO Cat# 41965-039
Medium (DMEM)
Roswell Park Memorial Institute GIBCO Cat# 10220-106
(RPMI) 1640
Ham’s F12 media Thermo Fisher Scientific Cat# 11765054
DMEM/F-12 GIBCO Cat# 31330-038
Horse Serum Thermo Fisher Scientific Cat# 16050-122
4–15% Criterion TGX Precast Midi Bio-Rad Cat# 5671084
Protein Gel
4–15% Mini-PROTEAN TGX Precast Bio-Rad Cat# 4561083
Protein Gel
Electrosurgical bipolar forceps Erbe Elektromedizin (Germany) N/A
ForceTriad electrosurgical unit Covidien (Ireland) N/A
Thermo Exactive orbitrap instrument Thermo Scientific N/A
Waters HSS T3 UPLC column Waters Corporation Cat# 186005614
Waters Xevo TQ-S triple quadrupole mass Waters Corporation N/A
spectrometer
Normal diet (for PDX study) Keaoxieli Feed (Beijing, China) Cat# 2152
Fat free diet (for PDX study) Xietong Organism (Beijing, China) Cat# RD17112401
(Continued on next page)

e9 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Western diet (omega-3/omega-6 = 1:50) Research Diets Cat# D19032707
(For cell line xenograft study)
Balanced diet (omega-3/omega-6 = 1:1) Research Diets Cat# D19032708
(For cell line xenograft study)
Fat free diet (For xenograft study) Research Diets Cat# D19032705
Precellys Lysing Soft tissue Precellys Cat# P000912-LYSK0
homogenizing kit
Precellys24 homogenizer Bertin Instruments Cat# P000669-PR240-A

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, George
Poulogiannis (george.poulogiannis@icr.ac.uk).

Materials Availability
All unique/stable reagents generated in this study are available from the Lead Contact without restriction.

Data and Code Availability

The code generated during this study are available at GitHub using the following accessions: https://github.com/adamltyson/
CalciumAnalysis, https://github.com/adamltyson/cell-coloc-3D, and https://github.com/adamltyson/foci2D. These accessions are
also provided in the Key Resources Table. The published article includes all REIMS m/z values and putative annotations for signif-
icantly different lipids between various receptor subtypes and MCF10A PIK3CA isogenics in the Supplementary Information in Tables
S1 and S4, respectively. Original/source data of REIMS profiles for Figures 1D, 1E, 1H, 3B, and 3D in the paper corresponding to
breast cancer cell lines and tumors is available through Mendeley Data (https://doi.org/10.17632/xcgc5kpntm.1)

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human samples
12 primary breast cancer samples from female patients (> 18 years of age) who consented to utilization of tissue for research were
provided by the Imperial College Healthcare NHS Trust Tissue Bank. Other investigators may have received samples from these
same tissues. The research was supported by the National Institute for Health Research (NIHR) Biomedical Research Centre based
at Imperial College Healthcare NHS Trust and Imperial College London. The views expressed are those of the author(s) and not
necessarily those of the NHS, the NIHR or the Department of Health. Human samples used in this research project were obtained
with evaluation and approval from the Wales Research Ethics Committee Reference 17/WA/0161 (Imperial College Healthcare Tissue
Bank Human Tissue Authority license: 12275; Project number R18024), the East of England – Cambridge East Research Ethics Com-
mittee Reference 14/EE/0024, and the project was registered under the Imperial College Tissue Bank.

Animals
Mouse PDX experiments were performed by Crown Bioscience in accordance with approved Institutional Animal Care and Use Com-
mittee (IACUC) protocols and ethical guidelines, and in strict accordance with the Crown Bioscience Guidelines and Standard Oper-
ating Procedures. Two primary human triple-negative breast cancer PDX corresponding to PIK3CA WT (BR1458) or PIK3CA C420R
MUT (BR1282) tumor fragments (2-3 mm in diameter) were inoculated subcutaneously into the breast pad of 7-9-week-old female
immunodeficient BALB/c nude mice weighing 17-23 g and which had not received previous treatments or procedures. Once tumors
reached a volume of 100-200 mm3, mice were randomized into four groups corresponding to either a normal or fatty acid free (FAF)
diet and administered with vehicle (0.5% hydroxypropyl cellulose in sterile water) or 100 mg/kg cPLA2 inhibitor ASB14780 (2578,
Axon Medchem) daily through oral gavage for 21 days. Animals were housed in a specific pathogen free facility in individually vented
cages and provided with diets and distilled water ad libitum. Room temperature was monitored and maintained at 20-25 C with the
light cycle set at 12 hours. All animals were checked daily for signs of ill health, as well as for any effects of tumor growth and
treatments on behavior such as mobility, food and water consumption, and body weight gain/loss. Researchers were not blinded
to treatment groups. Tumors were excised 2 hours after the final dosing and snap frozen in preparation for metabolomics/REIMS
processing and histopathological assessment. The mean tumor area as a percent of the total tissue area was initially assessed by
an independent histopathologist, and subsequently quantified using ImageJ version 1.51. Normal and FAF diets were purchased

Cell 181, 1596–1611.e1–e16, June 25, 2020 e10

 ll
OPEN ACCESS Article

from Keaoxieli Feed (2152) and Xietong Organism (RD17112401), respectively, and their compositions are summarized in Table S6.
All animal work undertaken at the Institute of Cancer Research was carried out under UK Home Office Project Licenses
P6AB1448A (Establishment License, X702B0E74 70/2902) and was approved by the Animal Welfare and Ethical Review Body at
the ICR. For cell line-derived xenograft studies, 7-9 week old female immunodeficient BALB/c nude mice weighing 18-25 g and which
had not received previous treatments or procedures were initially pre-conditioned on fat free, balanced (omega3/omega6 1:1) or
Western (omega3/omega6 1:50) diets for 2 weeks to assess tolerability. In order for the animal feed to be controlled, cages were ran-
domized to treatment groups rather than individual mice, and this occurred prior to orthotopic injections. Animals were subsequently
injected with 2.5x106 triple negative CAL51 (PIK3CA mutant) or Hs578T (PIK3CA WT) cells expressing either control shGFP or two
independent shRNAs targeting cPLA2 (cPLA2-sh1 or cPLA2-sh5) in 100 mL PBS:matrigel (50:50) into the right mammary fat pad. An-
imals were housed in a specific pathogen free facility in individually vented cages (no more than 4 mice per cage) and provided with
diets and distilled water ad libitum. Room temperature was monitored and maintained at 20-25 C with the light cycle set at 12 hours.
All animals were checked daily for signs of ill health, as well as for any effects of tumor growth and treatments on behavior such as
mobility, food and water consumption, and body weight gain/loss. Owing to the nature of the diets, blinding was not possible. Tumor
measurements were taken twice weekly in three dimensions (width, length and depth), and presented as relative tumor growth
normalized to first measurement day. Mice were excluded from the analysis if the primary tumor engrafted subcutaneously or into
the peritoneum instead of the mammary fat pad. On the final day of the experiment, tumors were excised and immediately snap
frozen in preparation for metabolomics/REIMS processing and histopathological assessment. The mean tumor area was quantified
as a percent of the total tissue area using ImageJ version 1.51. Fat free, balanced and Western diets were purchased from Research
Diets Inc. (D19032705, D19032708, D19032707, respectively). The composition of the diets used for the cell line xenograft study are
summarized in Table S6
Fresh frozen breast patient-derived xenograft (PDX) tumors were obtained from Crown Bioscience, and additional breast, ovarian,
pancreatic, sarcoma and colorectal PDX tumors were kindly provided by Champions Oncology. These are described in the Key Re-
sources Table. The PIK3CA mutational status for all 66 PDX tumor samples is summarized in Table S5.

Cell culture
Human female breast carcinoma cell lines AU565 BT20, BT474, BT549, CAL51, CAMA1, EFM19, Hs578T, JIMT1, KPL1, MCF7,
MDAMB134, MDAMB157, MDAMB231, MDAMB361, MDAMB436, MDAMB453, MDAMB468, MFM223, S68, SKBR3, T47D,
UACC812 and VP229 cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM) (GIBCO, 41965-039) and BT483,
HCC1143, HCC1395, HCC1428, HCC1500, HCC1569, HCC1937, HCC1954, HCC202, HCC38, HCC70, SUM52, ZR751 and
ZR7530 cells were cultured in Roswell Park Memorial Institute (RPMI) 1640 media (Sigma-Aldrich, R8758), both supplemented
with 10% fetal bovine serum (FBS) (GIBCO, 10220-106). SUM44 (human, female), SUM159 (human, female), SUM149 (human, fe-
male), SUM225 (human, female) and SUM229 (human, female) were cultured in Ham’s F12 media (Thermo Fisher Scientific,
11765054) supplemented with 5% FBS, 0.5 mg/ml hydrocortisone (Sigma-Aldrich, H-0888) and 0.4% insulin-transferrin-selenium
(GIBCO, 41400-045) and MCF10A (human, female) cells (including the PIK3CA MUT isogenic panel) were cultured in DMEM/F-12
(GIBCO, 31330-038) supplemented with 5% horse serum (Thermo Fisher Scientific, 16050-122), 20 ng/ml epidermal growth factor
(EGF) (Peprotech, AF-100-15), 100 ng/ml cholera toxin (Sigma-Aldrich, C-8052), 0.5 mg/ml hydrocortisone (Sigma-Aldrich, H-0888)
and 10 mg/ml insulin (Sigma-Aldrich, I1882). For cell culture conditions free of exogenous sources of fatty acids, 10% FBS or 5%
horse serum was replaced with 1% fatty acid bovine serum albumin (BSA) (Sigma-Aldrich, A8806). All cell lines were maintained
at 37 C, 5% CO2. All cell lines were authenticated by short tandem repeat analysis (Eurofins Scientific) and were tested and
confirmed to be negative for mycoplasma infection.

METHOD DETAILS

Experimental design
Experiments were repeated multiple times across different cell line and tumor models with similar results as indicated in the figure leg-
ends. Key findings from in vivo experiments were reproduced using orthogonal approaches including cell line xenograft models and
genetic inhibitions. Animals were randomized into treatment groups either individually following PDX engraftment and growth to
100-200 mm3, or in cages following diet preconditioning. Throughout the study researchers were not blinded as data analysis required
prior knowledge of the sample annotation. For in vitro and in vivo experiments, sample size was chosen based on preliminary exper-
iments and previous experience with protocols. No completed data were excluded from the analysis performed in this manuscript.

Mass spectrometry analysis

Cell lines were cultured in 10 cm2 plates and harvested in biological triplicate for REIMS analysis when 80% confluent. After replacing
the media one hour prior to harvesting, the plates were washed twice with ice-cold PBS, and immediately flash frozen in liquid nitro-
gen. 1 mL PBS was added to the cells and cell extracts were scraped and collected in eppendorf tubes and centrifuged for 5 min at
5000rpm. Pellets were flash frozen and stored at
80 C. For fresh frozen PDX and primary breast tumors, three separate regions
were sectioned for analysis.

e11 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

REIMS analysis was performed with commercially available electrosurgical bipolar forceps (Erbe Elektromedizin, Germany) con-
nected to a ForceTriad electrosurgical unit (Covidien, Ireland) programmed in Macro bipolar setting using 4 W or 30 W power for cell
lines and tumors, respectively. Bipolar forceps were connected to the inlet capillary of a Thermo Exactive orbitrap instrument (Thermo
Scientific) using PTFE tubing, allowing for the direct suction of aerosol generated from rapid biomass heating to the mass spectrom-
eter (set up is shown in Figure 1A). The mass spectrometer settings used for phospholipid and fatty acid profiling are summarized in
Table S7.

Metabolomics data pre-processing and analysis

Raw mass spectrometric files were converted to mzXML format using the ProteoWizard MSConvert software (Version 3.0.11781)
(Chambers et al., 2012) and imported into RStudio (Version 3.4.4). Data were pre-processed using the MALDIquant package
(Gibb and Strimmer, 2012). Prior to pre-processing, high quality spectra corresponding to the maximum total ion (TIC) count for
each measurement were selected from each individual sample runs. Spectra were normalized using median scaling determined
from the non-zero intensities across the full scan range (m/z 150-2000), and square root transformed. Peaks were aligned using
the locally weighted scatterplot smoothing (LOWESS) function and detected using median absolute deviation (MAD) with a signal
to noise ratio (SNR) equal to 3. Peak matching was subject to a maximum peak shift of 5 ppm. For further analysis, the median in-
tensity of all biological and technical replicates was calculated, such that each sample is represented by a single spectrum. The mass
ranges of m/z 600-900 and m/z 150-500 were recorded for the detection of phospholipids and fatty acids, respectively. m/z values of
interest which discriminated significantly between experimental conditions were annotated based on previously performed tandem
mass spectrometry (MS/MS) fragmentation analysis and data available from Lipid Maps (https://www.lipidmaps.org). To assess in-
strument reproducibility, the MDAMB468 cells were included in every REIMS run as a quality control. The spectra obtained were
compared among different batches using unsupervised or guided principal component analysis (gPCA) and no significant batch ef-
fect was observed. Classification and feature-selection was performed using leave-one-out or 3-fold cross validation with random
forest as the classifier, applying the ‘caret’ and ‘randomForest’ packages, while unsupervised analysis was performed using hierar-
chical clustering (Euclidean distance, complete linkage), or non-negative matrix factorization (NMF) using the ‘NMF’ package. Pre-
diction of continuous variables such as ESR1 gene expression was performed using random forest regression analysis with the Tree-
Bagger function in MATLAB (2014a, version 8.3.0.532).

Transfections and site directed mutagenesis

Sub-cloning of cPLA2 sh1 into the pLKO-Tet-On inducible vector was done by first amplifying the shRNA sequence of interest with
the following primers: Forward 50 -GTGGAAAGGACGAAACACCGGT-30 and Reverse 50 -TTTGTCTCGAGGTCGAGAATTC-30 , and
subsequently introducing compatible AgeI and EcoRI restriction sites in the shRNA oligonucleotides and pLKO-Tet-On vector
(TET-pLKO puro was a gift from D. Wiederschain, Addgene plasmid 21915). For lentiviral gene knockdown, PLA2G4A or GFP
shRNAs were transfected in 293T cells using FuGENE HD Transfection reagent (Promega, E2311), according to the manufacturer’s
instructions and infected cells were selected in the presence of 1 mg/ml puromycin. For siRNA knockdown, MCF10A PIK3CA WT and
E545K/H1047R MUT cells were transfected with 25 nM ON-TARGETplus SMARTpool human siRNAs targeting RAPTOR, RICTOR,
FRAP1, PLCg1 or PRKCZ using DharmaFECT-1 Transfection Reagent for 48 hours. HA-tagged cPLA2 was transiently overex-
pressed using 9 mg pCMV3-HA-PLA2G4A vector DNA and 18 mL FuGENE HD Transfection reagent, and experiments were per-
formed 48 hours post transfection.
For generating CRISPR knockouts, MCF10A PIK3CA WT and H1047R cells were transfected with PLA2G4A sgRNA CRISPR/Cas9
All-in-One Lentivector Target 2 (Applied Biological Materials, K1659207) or control LentiCRISPR v2 (lentiCRISPR v2 was a gift from
Feng Zhang, Addgene plasmid #52961) according to the manufacturer’s instructions. Immunoblotting and DNA sequencing were
used to validate the isolation of PLA2G4A deleted clones.
Phosphoresistant isoforms of cPLA2 were generated using the QuikChange Lightning Site-Directed Mutagenesis Kit (Agilent,
210518) according to the manufacturer’s instructions. Briefly, 100 ng of pCMV3-HA-PLA2G4A vector DNA was incubated with
125 ng of the appropriate mutagenesis primers:

S505A Forward 50 -GCAAAGTCACTCAAAGGAGCCAGTGGATAAGATGTATTG-30 , Reverse 50 -CAATACATCTTATCCACTGGC

TCCTTTGAGTGACTTTGC-30 ;
T376A
Forward 50 -TTCTTCATACTTCTTAACGACTGCTCCCATAAAAAATTTGCTTCCAA-3 0 Reverse 50 -TTGGAAGCAAATTTTTTATGGG
AGCAGTCGTTAAGAAGTATGAAGAA-30 ;

PCR cycling parameters were as follows: initial denaturing at 95 C for 2 min, followed by 18 cycles of 20 s denaturing (95 C), 10 s
annealing (60 C) and 4.5 min elongation (68 C). A final elongation step occurred for 5 min (68 C). To assess the effects of these phos-
phomutants on cPLA2 activity or arachidonic acid (AA) levels, cells were transiently transfected with 9 mg pCMV3-HA-PLA2G4A,
pCMV3-HA-PLA2G4A-S505A or pCMV3-HA-PLA2G4A-T376A vector DNA and 18 mL FuGENE HD Transfection reagent, and ex-
periments were performed 48 hours post transfection.

Cell 181, 1596–1611.e1–e16, June 25, 2020 e12

 ll
OPEN ACCESS Article

Cell based assays

For colony formation assays, single cell suspensions containing 5x102 MCF10A PIK3CA WT or E545K/H1047R MUT cells were
seeded in 6 well plates and allowed to adhere for 18 hours. Wells were subsequently washed twice with PBS and DMEM-F/12 media
containing growth factors and 1% FAF BSA (herein called ‘FAF media’), supplemented either with 0.1% DMSO or 25 mM AA. After
14 days, colonies were stained with crystal violet, and quantified using ImageJ. For 3D spheroid culture, 5x103 cells were seeded in
round bottom, low attachment 96 well plates (Corning, 7007) and incubated for 14 days. Media was replaced every 3 days. For
ASB14780 viability assays, 5x103 PIK3CA WT (MCF10A, MDAMB134, AU565, Hs578T) or PIK3CA MUT (MCF10A E545K/
H1047R, CAL51, MCF7, MDAMB453) were seeded in 96 well plates in full media and allowed to adhere for 18 hours. Wells were
washed twice with PBS, and FAF media containing either 0.1% DMSO, 20 nM-10 mM ASB14780 or the inhibitor supplemented
with 25 mM AA was added. Viability was assessed after 72 hours using the CellTiter 96 Aqueous Non-Radioactive (MTS) cell prolif-
eration assay (Promega, G5430).
To measure cell proliferation, 5x103 MCF10A isogenic cells expressing GFP control or cPLA2 shRNAs (sh1 and sh5) were seeded
in 96 well plates, left to adhere for 18 hours, washed twice with PBS, and incubated with FAF media. Cell number was determined on
Days 0, 1, 3 and 5 using the Sulforhodamine B (SRB) assay.
To assess eicosanoid-mediated autocrine and paracrine effects on cell proliferation, conditioned media were derived from
MCF10A PIK3CA WT or MUT isogenics. Cells were grown to 80% confluency in full media, after which residual media were washed
off twice with PBS, and cells were cultured with FAF media for 48 hours. The conditioned media were centrifuged at 2000 g for 20 min
to remove cell debris and incubated with Lipid Removal Adsorbent (LRA, Supelco, 13358) at 0.4 g of LRA per 10 mL media to deplete
secreted lipids. Following overnight incubation with gentle shaking, lipid-deprived media were centrifuged at 2000 g for 20 min and
supernatants were collected. Proliferation was assessed over 5 days using the SRB assay.
To assess fatty acid uptake, 5x104 MCF10A PIK3CA WT or MUT cells were seeded in 100 mL full medium containing 5% horse
serum in a 96-well plate and incubated for 24 hours at 37 C, 5% CO2. Cells were washed twice with PBS and serum deprived for
1 hour before adding 100 mL TF2-C12 Fatty Acid Stock Solution (Sigma-Aldrich, MAK156). After incubating the cells for 1 hour at
37 C, the fluorescence signal was measured at Ex/Em = 485/515 nm.

Three-dimensional cell culture

For 3D acini formation, MCF10A cells were grown in Matrigel (Corning, 354230) as described previously (Debnath et al., 2003). Briefly,
PIK3CA isogenic MCF10A cells expressing pLKO-Tet-On inducible shRNA against cPLA2 or GFP were induced with 2 mg/ml doxy-
cycline for 48 hours prior to seeding. A single cell suspension (3x103 cells per well) was plated in assay medium containing 1% FAF
BSA onto eight-well chamber slides coated with a layer of pure growth factor–reduced Matrigel. Cells were then overlaid with a
300 mL of assay medium containing 2% Matrigel and cultured for 10 days. Growth media was replaced every 3 days.

Confocal microscopy
For immunofluorescence, cells were fixed with 4% PFA at room temperature, permeabilized with pre-chilled 0.5% Triton X-100 for
10 min prior to blocking with 2% BSA/PBS for 1 hour. After blocking, cells were incubated overnight at 4 C with primary Ki-67 anti-
body (Cell Signaling Technology, 9129) in a humidified chamber. The following day cells were incubated with fluorescently labeled
secondary antibodies Alexa Fluor 546-conjugated secondary antibody (Thermo Fischer Scientific, A-11003) and Phalloidin 633
(Abcam, ab176758) to visualize Ki-67 and F-actin, respectively, 1% BSA/PBS for 1 hour. Slides were mounted using Prolong
Gold anti-fade reagent with DAPI (Invitrogen, D1306). Images were captured using a Zeiss AxioObserver microscope equipped
with a Yokogawa CSU-W1 spinning disk unit (Intelligent Imaging Innovations) and a 40x oil objective. Serial z stacks of the acini struc-
tures were acquired at 5 um intervals (usually 10-15 sections per field), and then analyzed with a custom MATLAB script (2017b, The
Mathworks Inc.). Images were resampled to isotropic resolution and each spheroid was manually segmented. The DAPI signal was
thresholded using Otsu’s method (Otsu, 1979) following intensity depth correction and smoothing. Holes were then filled and small,
non-cellular objects were removed. The resulting binary nuclei image was used as a mask to measure cellular proliferation.

Enzymatic assays
cPLA2, iPLA2 and sPLA2 activities were measured using commercially available assays (Abcam, ab133089 or ab133090) according
to the manufacturers instructions. Briefly, total cell lysates were obtained using 1x Cell Lysis Buffer (Cell Signaling Technology, 9803)
under non-denaturing conditions. For cPLA2 activity, 10 mL lysate, 5 mL Assay buffer and 200 mL substrate solution containing arach-
idonoyl Thio-PC were incubated at room temperature for one hour. For iPLA2 activity, cell lysates were either untreated (measuring
both cPLA2 and iPLA2 activity) or treated with 5 mM of the iPLA2 specific inhibitor bromoenol lactone (BEL), and activity was deter-
mined as follows: iPLA2 activity = (Activity without BEL) – (Activity with BEL). For sPLA2 activity, conditioned media was obtained
from MCF10A PIK3CA WT and MUT cells and concentrated using a centrifugal vacuum evaporator (‘‘SpeedVac’’). Dried samples
were resuspended in 100 mL Assay Buffer, and 10 mL of this was used for the assay. Cellular diacylglycerol (DAG) levels were
measured using a DAG assay kit according to the manufacturer’s instructions (Cell Biolabs Inc., MET-5028). Briefly, 1x107 cells
were harvested by scraping with 1 mL cold PBS, and pellets were obtained after centrifugation at 1500 g for 10 min. Lipids were
extracted following sonication and incubation with methanol, sodium chloride and chloroform. The lower chloroform phase was

e13 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

washed twice with pre-equilibrated upper phase (PEU) and dried under a stream of nitrogen. 50 mL of assay buffer was used to re-
suspend the dried sample, 20 mL of which was used for the assay.

Immunoblot analysis
Cells were washed with ice-cold PBS and lysed on ice for 30 min with cell lysis buffer containing RIPA buffer (Thermo Scientific,
89900) supplemented with 4 mg each of leupeptin (Sigma-Aldrich, L2884) and pepstatin (Sigma-Aldrich, P5318), 2 mM Na3VO4
(Sigma-Aldrich, 450243), 1 mM DL-Dithiothreitol (Sigma-Aldrich, 646653), 10 mM Calyculin A (Cell Signaling Technology, 9902),
250 mM b-glycerophosphate (Sigma-Aldrich, G9422) and 400 mM PMSF protease inhibitor (Cell Signaling Technology, 8553). Lysates
were subjected to centrifugation at 12,000 g for 30 min at 4 C, and protein concentrations were determined using the Bradford assay
(Bio-Rad, 5000006). Nuclear isolation for mature SREBP probing was performed using a Nuclear Extract Kit (Active Motif, 40010) with
10 mg/ml of the protease inhibitor ALLN (Millipore, 208719). Protein lysates were boiled for 10 min and subjected to SDS-PAGE elec-
trophoresis using 4%–15% precast gels (Bio-Rad, 567-1084). Densitometry was calculated using the Image Lab Software 5.2.1 (Bio-
Rad). Affinity purified custom antibodies for phosphorylated cPLA2 on Thr376 were developed by Thermo Fisher Scientific using the
PLA2G4A-369:383 peptide antigen and used at a dilution of 1:500 in 5% BSA/TBST. All other primary antibodies were used at a dilu-
tion of 1:1000 in 5% BSA/TBST solution, and secondary antibodies at 1:5000 in 5% milk/TBST.

Immunoprecipitation analysis
MCF10A PIK3CA WT and MUT isogenics were transiently transfected with HA-tagged cPLA2 (pCMV3-HA-PLA2G4A, Sino Biological,
HG13126-NY) and lysed with 1X Cell Lysis Buffer (Cell Signaling Technologies, 9803) 48 hours post transfection. Equal volume of diluted
cell lysates containing 4 mg of soluble protein were incubated with 2 mg rabbit anti-HA (Cell Signaling Technology, 3724) or 2 mg normal
rabbit IgG (Santa-Cruz Biotechnology, sc-2027) as a negative control, and were pre-coupled with protein G Sepharose beads (Sigma-
Aldrich, P3296) following incubation at 4 C for 4 hours rotating. 1X Cell Lysis Buffer was used to wash the beads four times, after which
samples were resuspended in 30 mL 2x Laemmli sample buffer (Bio-Rad, 161-0737) and boiled for 10 min to release bound proteins.
To detect GTP-bound Rac1, the Active Rac1 Detection Kit (Cell Signaling Technology, 8815) was used, following the manufac-
turer’s instructions. Briefly, cell lysates were harvested under non-denaturing conditions using 1X Cell Lysis Buffer. To affinity pre-
cipitate activated G-protein, spin cups were incubated with 100 mL of glutathione resin and subsequently washed with 1X Cell Lysis
Buffer. 20 mg of GST-PAK1-PBD was added to spin cups containing glutathione resin, after which 700 mL of the cell lysate containing
1 mg total protein was added, and the mix was incubated at 4 C for 1 hour with gentle shaking. Following three washes with 400 mL 1X
Cell Lysis Buffer, samples were eluted by adding 2X reducing sample buffer containing 200 mM DTT in 2X SDS sample buffer, and
subsequently heated for 5 min at 95 C. 25 mL of eluted samples were loaded onto an SDS-PAGE gel for subsequent immunoblot
analysis using the provided Rac1 mouse monoclonal antibody (1:1000 dilution).

Immunohistochemistry analysis
10 mM thick sections were obtained from formalin-fixed and paraffin-embedded tumors and stained using an anti-phospho PKCz
Thr560 antibody (Abcam, ab62372) at a dilution of 1:100. For breast PDX and cell line-derived xenograft tumors, fresh frozen tumor
pieces were mounted on Optimal Cutting Temperature (OCT) compound and sectioned to 10 mM thickness. To assess the presence
of activated natural killer cells, slides were stained using goat anti-mouse NKp46/ncr1 polyclonal antibody (R and D systems, 170-
6516) at a final concentration of 3 mg/ml. Images were captured on a Hamamatsu NanoZoomer, and staining was quantified using the
IHC Profiler plug-in for ImageJ.

Proximity ligation assay

For the proximity ligation assay (PLA), all incubations were performed in a humidity chamber as per the manufacturer’s instructions
using a Duolink In Situ Detection Reagents Red kit (Sigma-Aldrich, DUO92008). Briefly, MCF10A PIK3CA isogenic cells were seeded
on glass coverslips, serum-starved overnight, and stimulated with full serum and growth factors for 30 min; control cells were un-
treated. They were then fixed with 4% paraformaldehyde (PFA), blocked and incubated with primary antibodies (anti-cPLA2 and
anti-phospho PKCz Thr560) for 1 hour at room temperature. Samples were then incubated with Minus and Plus PLA probes
(Sigma-Aldrich, DUO92004 and DUO92002) for 1 hour at 37 C, followed by a ligation step for 30 min at 37 C, and an amplification
step for 100 min at 37 C. Finally, the coverslips were mounted with the Duolink PLA mounting media and DAPI, and fluorescence was
visualized with a Zeiss 710 Confocal microscope with a 40x oil objective. Images were analyzed using a custom MATLAB (2017b)
script. Briefly, the DAPI signal was segmented by taking a maximum intensity projection, smoothing, and then thresholding (Otsu,
1979). Borders between cells were estimated by finding the midpoints between nuclei using a watershed approach. A summed pro-
jection was calculated for the probe signal, this was smoothed (Gaussian kernel sigma of 1 pixel), and thresholded. The threshold
calculated from a positive control image was applied to all images. The segmented image of the probe was used to calculate the
number of foci on a per cell basis.

Fluorescent calcium assay

Calcium flux was measured using the Fluo-4 DirectTM Calcium Assay Kit (Thermo Fisher Scientific, F10471). Full MCF10A growth
medium was removed and replaced with FBS-free medium for 18 hours. Control cells were treated with equal volume of the culture

Cell 181, 1596–1611.e1–e16, June 25, 2020 e14

 ll
OPEN ACCESS Article

media of the 2x Fluo-4 DirectTM calcium assay reagent solution for 30 min in 37 C, 5% CO2, without removing the assay media. The
induced cells were stimulated with full media for 30 min, while incubated with the calcium assay reagent solution. Cells were then
analyzed by monitoring the fluorescence of the Fluo-4 dye using an ImageXpress Micro Confocal High-Content Imaging System
(Molecular Devices) and a 20x objective with 488 nm excitation. The fluorescence per field before and after induction was calculated
using a MetaXpress Software Custom Module.
To assess the effects of PLCg1 inhibition on calcium flux, cells were treated with either 2 mM U73122 for 24 hours, or transiently
transfected with 25 nM ON-TARGETplus SMARTpool human siRNA targeting PLCg1 for 48 hours. For the final 18 hours, full MCF10A
growth medium was removed and replaced with FBS-free medium. Immediately prior to the assay, cells were treated with equal vol-
ume culture media and 2x Fluo-4 DirectTM calcium assay reagent, and baseline fluorescence was monitored using an ImageXpress
Micro Confocal High-Content Imaging System (Molecular Devices). After 300 ms, cells were induced by adding 1 mL full MCF10A
growth medium, and fluorescence of the Fluo-4 dye was monitored using a 20x objective with 488 nm excitation. Calcium flux
was calculated by first normalizing fluorescence readings to baseline measurements using the formula: dF = F(t) – F(0)/F(0), where
F = fluorescence at 488 nm, t = time, and F(0) represents the average baseline readings from 1-299 ms. Fluorescence values were
subsequently normalized to a unit interval between 0 and 1 and presented as the time required post stimulation to reach a maximal
calcium intensity.

Quantitative RT-PCR and PIK3CA mutation analysis

RNA was extracted using the RNeasy Plus Mini Kit (QIAGEN, 74134) and 1 mg of total RNA was used to synthesize complementary
DNA (cDNA) using the QuantiTect Reverse Transcription Kit (QIAGEN, 205311). Primer sequences used are as follows: Human
cPLA2, forward: 50 -GATGAAACTCTAGGGACAGCAAC-30 ,

reverse: 50 -CTGGGCATGAGCAAACTTCAA-30 , Human b-actin: forward: 50 -GACCCAGATCATGTTTGAGACC-30 ,

reverse: 50 -CTTCATGAGGTAGTCAGTCAGG. Reactions were performed with SYBR Select Master Mix (ThermoFisher, 4472908)
using the TProfessional Thermocycler from Biometra and analyzed with the qPCRsoft version 3.1 (Thistle Scientific/Analytik Jena).
The following cycle reactions were used: pre-denaturation for 3 min at 95 C, followed by 45 cycles of 5 s at 95 C (denaturation), 5 s at
55.8 C (annealing) and 15 s at 72 C (elongation).
To detect PIK3CA mutations in primary breast tumor samples, DNA was extracted from 10 mg of tissue using the QIAamp DNA
mini kit (QIAGEN, 51304). Mutations in exon 9 (helical domain) and exon 20 (kinase domain) of PIK3CA were assessed using the PNA-
Clamp PIK3CA Mutation Detection Kit (Panagene, PNAC-4001), according to the manufacturer’s instructions. Briefly, reactions were
performed using 10 ng DNA with a SYBR Green PCR reaction premix and primer premixes detecting E542, E545, Q546 and H1047
mutations using the TProfessional Thermocycler. The following cycle reactions were used: pre-denaturation for 5 min at 94 C, fol-
lowed by 40 cycles of 30 s at 94 C (denaturation), 20 s at 70 C (peptide nucleic acid clamping), 30 s at 63 C (annealing) and 30 s
at 72 C (extension). PIK3CA mutations were assessed based on manufacturer’s instructions.

In vitro kinase assay

In vitro kinase assays were performed using the ADP-Glo Kinase Assay (Promega, V6930), according to the manufacturer’s instruc-
tions. Briefly, 100 ng recombinant human protein kinase C zeta (PKCz, Insight Biotechnology, TP302472) and 0.5 mg/ml phospholi-
pase A2, group IVA (cPLA2, Insight Biotechnology, TP320972) were incubated in kinase buffer comprising 40 mM TRIS buffer pH 7.5,
20 mM MgCl2 (Sigma-Aldrich, M8266), 0.1 mg/ml BSA (Sigma-Aldrich, A2153) and 50 mM DTT (Sigma-Aldrich, 646563), and supple-
mented with 20 mM ATP. Reactions were incubated at 30 C for 3 hours, after which luminescence was measured following an inte-
gration time of 1 s.

Chemokine assays
Tumor samples were placed in 2 mL lysing tubes prefilled with 1.4 mm ceramic (zirconium oxide) beads and 1 mL of chilled PBS.
Samples were homogenized with a Precellys24 homogenizer programmed with three 30 s cycles at 6,500 Hz and 4-min pause times.
At the end of the homogenization cycle, samples were centrifuged at 5,000rpm for 10 min at 4 C, and supernatants were transferred
to fresh Eppendorf tubes
To measure CCL5 (RANTES) chemokine levels, the CCL5 Mouse ELISA kit (Abcam, ab100739) was used according to the man-
ufacturer’s instructions. Briefly, 100 mL of standards prepared in Assay Diluent A and sample supernatants were added to appropriate
wells and incubated at room temperature for 2.5 hours. Wells were subsequently washed 4 times with 1X Wash Solution, and 100 mL
of 1X Biotinylated CCL5 (RANTES) Detection Antibody was added and incubated for 1 hour at room temperature with gentle shaking.
Following the incubation with the detection antibody, wells were washed 4 times and incubated with 100 mL HRP-Streptavidin so-
lution for 45 min at room temperature with gentle shaking. Following a final set of 4 washes, 100 mL of TMB One-Step Substrate Re-
agent was added to each well and incubated for 30 min at room temperature in the dark with gentle shaking. 50 mL of stop solution
was subsequently added, and measurements were taken immediately at 450 nm.
To measure CX3CL1 chemokine levels, the CX3CL1 Mouse ELISA kit (Abcam, ab100683) was used according to the manufac-
turer’s instructions. Briefly, 100 mL of standards prepared in Assay Diluent C and sample supernatants were added to appropriate

e15 Cell 181, 1596–1611.e1–e16, June 25, 2020

 ll
Article OPEN ACCESS

wells and incubated at room temperature for 2.5 hours. Wells were subsequently washed 4 times with 1X Wash Solution, and 100 mL
of 1X Biotinylated CX3CL1 Detection Antibody was added and incubated for 1 hour at room temperature with gentle shaking.
Following the incubation with the detection antibody, wells were washed 4 times and incubated with 100 mL HRP-Streptavidin so-
lution for 45 min at room temperature with gentle shaking. Following a final set of 4 washes, 100 mL of TMB One-Step Substrate Re-
agent was added to each well and incubated for 30 min at room temperature in the dark with gentle shaking. 50 mL of stop solution
was subsequently added, and measurements were taken immediately at 450 nm. All measurements were normalized to total protein
content as determined using the BCA protein assay (Thermo Fisher Scientific, 23225).

Lipid extraction and eicosanoid profiling

Total lipids were extracted based on the Bligh and Dyer method (Bligh and Dyer, 1959) and analyzed as described previously (Wolfer
et al., 2015). Briefly, cell lines were harvested when 80% confluent, and media was replaced one hour prior to lipid extraction. Cell
pellets were resuspended in 1ml water, after which 3.75 mL chloroform/methanol mixture (1:2 v/v) was added and samples were
vortexed and incubated for at least 30 min on ice. 1.25 mL chloroform was added, followed by 1.25 mL mili-Q water and after vortex-
ing and centrifugation at 1000rpm for 10 min at 4 C, the organic bottom phase was separated and transferred to a 15 mL amber vial to
get dried under a nitrogen flow. For conditioned media profiling, cells were grown to 80% confluency in full media, after which residual
media was washed off twice with PBS, and cells were cultured with FAF media for 48 hours. The conditioned media was centrifuged
at 2000 g for 20 min to remove cell debris. Samples were concentrated using a centrifugal vacuum evaporator. The lipid extraction
was then performed as above.
Eicosanoid levels were determined using liquid chromatography with a Waters HSS T3 UPLC column connected to a Waters Xevo
TQ-S triple quadrupole mass spectrometer with an electrospray ionization (ESI) source. TargetLynx (Waters, Manchester, UK) was
used for peak detection and integration. Data was normalized to protein concentration.
To measure PGE2 (Enzo Life Sciences, ADI-900-001) and AA (Generon, CEB098Ge) concentrations, enzyme-linked immunosor-
bent assays were used, following the manufacturer’s instructions.

Bioinformatic analysis
Gene-centric RMA-normalized expression data and mutation status for available cell lines were obtained from the Cancer Cell Line
Encyclopedia (CCLE). 150 frequently mutated genes were identified based on a mutation frequency of R 20% or higher in n = 35
available cell lines, and significant enrichment in observed clusters was assessed using Fisher’s exact test. A functional annotation
analysis using the Database for Annotation, Visualization and Integrated Discovery (DAVID, v6.8) (Huang et al., 2009) was used to
identify biologically relevant pathways represented by a set of 512 genes, which were significantly upregulated in the PIK3CA mu-
tation- and lipid-enriched cluster. Interactions between genes involved in the identified pathways were visualized by functional
gene networks using the ‘ReactomePA’ package in R (Yu and He, 2016).

QUANTIFICATION AND STATISTICAL ANALYSIS

Student t test, Fisher’s exact test, and one- or two-way ANOVA were used to evaluate statistical significance as indicated in the
respective figure legends. The Shapiro-Wilk test was used to assess normality. The ‘N’ for each experiment can be found in the figure
legends and represents independently generated samples for in vitro experiments, wells for cell-based assays, or mice for in vivo
experiments. Bar graphs present the mean ± standard error of the mean (SEM). Significance was defined as p < 0.05, and denoted
by asterisks throughout the figures as follows: n.s (not significant), *(p < 0.05), **(p < 0.01), ***(p < 0.001). No statistical methods were
used to determine sample size, and no completed data were excluded from the analysis performed in this manuscript. Statistical
tests were performed using GraphPad Prism (version 8.0.1) and R statistical software (version 3.5.1).

Cell 181, 1596–1611.e1–e16, June 25, 2020 e16

 ll
Article OPEN ACCESS

Supplemental Figures
A B C

E F

(legend on next page)

 ll
OPEN ACCESS Article

Figure S1. Related to Figure 1

(A) Volcano plots of significantly altered phospholipids between receptor positive and negative cell lines. Black dots: not significantly altered; Red dots:
significantly upregulated; Green dots: significantly downregulated phospholipids. (B) Area under the curve (AUC) classification accuracies for estrogen (ER),
progesterone (PR), HER2 receptor and triple negative status of 30 primary and PDX breast tumors (median intensity of n = 3 separate sections per tumor) following
feature selection for phospholipids in the m/z range 600-900 and leave-one-out cross validation. (C) Immunoblot analysis of estrogen inducible protein pS2 (top)
and prediction of ESR1 expression (bottom) in ER+ve T47D cells following treatment with 0.1% DMSO or indicated concentrations of 4-OHT for 72 hours using
REIMS. (D) NMF consensus maps summarizing the clustering of cell lines used in Figure 1D. The color map represents the correlation between cell lines in the
same cluster when samples are divided into 2-6 groups. The highest cophenetic score was obtained for two clusters. (E) REIMS analysis of MCF10A PIK3CA WT
and MUT cells cultured as 3D spheroids for 10 days. Clustering was performed as in Figure 1D using the median lipid intensities of 3 biological replicates. (F)
Overall, precision and recall classification accuracies for PIK3CA mutation status in primary and PDX breast tumors (n = 30 in total), using all detectable lipid
features (n = 1147) following 3-fold cross validation repeated 100 times with random forest as a classifier. n.s., not significant; *p % 0.05; **p % 0.01; ***p % 0.001.
P values in (C, bottom panel) were calculated with one-way ANOVA, followed by unpaired, two-tailed Student’s t test with Bonferroni correction.
 ll
Article OPEN ACCESS

A B

Figure S2. Related to Figures 1 and 2

(A) Pathway enrichment analysis of genes corresponding to the three most significantly enriched KEGG pathways determined from the 512 significantly upre-
gulated genes in the lipid-enriched cluster from Figure 1D. (B) Gene interaction networks corresponding to the 55 genes encompassing the KEGG pathways in (A).
(C) FASN, (D) ELOVL6, and (E) LDLRAP1 mRNA expression between cell lines in the lipid-enriched (n = 19) and depleted (n = 15) clusters. *p % 0.05; **p % 0.01;
***p % 0.001. P values in (C), (D), and (E) were calculated with unpaired, two-tailed Student’s t test.
 ll
OPEN ACCESS Article

B C

Figure S3. Related to Figure 2

(A) Cell viability of MCF10A PIK3CA MUT cells following treatment with increasing concentrations of rapamycin, torin 1, BYL-719, BKM120, MK2206 or
GSK690693 for 72 hours. (B) Unsupervised hierarchical clustering of the median phospholipid intensities of 5 PIK3CA MUT breast cancer cell lines (MCF7, T47D,
MDAMB361, MDAMB453 and BT474) treated with 20 nM rapamycin, 100 nM BYL-719 and 150 nM MK2006 for 72 hours. (C) Immunoblot analysis of mTORC1
and mTORC2 signaling in the PIK3CA MUT isogenic panel. Cells were serum and growth-factor starved for 16 hours and subsequently stimulated with 5% horse
serum, 20 ng/ml EGF, 0.5 mg/ml hydrocortisone and 10 mg/ml insulin for 30 min. Data in (A) are presented as the mean ± SEM of n = 4 biological replicates and are
representative of at least two independent experiments. n.s., not significant; *p % 0.05; **p % 0.01. P values in (A) were calculated with unpaired, two-tailed
Student’s t test.
 ll
Article OPEN ACCESS

Figure S4. Related to Figures 3 and 4

Intracellular levels of (A) Arachidonic acid (AA) and (B) PGE2 levels as measured by LC-MS profiling. (C) AA levels from the conditioned media (CM) of indicated
cells before and after lipid depletion (LD). (D) Cell proliferation assays of MCF10A PIK3CA WT cells cultured in CM derived from WT or E545K MUT cells before or
after LD, with or without the supplementation of 25 mM AA, palmitate or palmitoleate. (E) Cell proliferation assays of MCF10A PIK3CA E545K MUT cells before or
after LD, with or without the supplementation of 25 mM AA, palmitate or palmitoleate. (F) Diacylglycerol (DAG) levels in MCF10A PIK3CA WT and MUT cells. (G)
Diagram summarizing DAG contribution to AA production. Sulforhodamine B (SRB) protein staining was used in (D) and (E) to measure cell proliferation over
5 days. Data are presented as the mean ± SEM of n = 3-4 biological replicates and are representative of at least two independent experiments. *p % 0.05; **p %
0.01; ***p % 0.001. P values in (A), (B), (C) and (F) were calculated using one-way ANOVA followed by unpaired, two-tailed Student’s t test with Bonferroni
correction, and in (D) and (E) with two-way ANOVA.
 ll
OPEN ACCESS Article

(legend on next page)

 ll
Article OPEN ACCESS

Figure S5. Related to Figure 4

(A) Immunoblot of phospholipases in MCF10A PIK3CA WT and MUT cells following serum and growth-factor deprivation for 16 hours and stimulation with serum
and growth factors for 30 min. (B) Immunoblot analysis of cPLA2 protein decay following treatment with 50 mM cycloheximide (CHX) for the indicated times. Image
and quantification is from one experiment. (C) Real-time quantitative PCR of PLA2G4A expression in the MCF10A PIK3CA isogenic panel. (D) AA levels measured
by REIMS in MCF10A E545K PIK3CA MUT cells following RAPTOR and RICTOR siRNA-mediated knockdown 48 hours post transfection under exogenous FAF
conditions. ELISA analysis of (E) AA and (F) PGE2 in the MCF10A PIK3CA isogenic panel 48 hours post RICTOR siRNA-mediated knockdown. (G) Immunoblot
analysis of cPLA2 protein decay (top) and quantification (bottom) following RICTOR siRNA-mediated knockdown and treatment with 50 mM cycloheximide for the
indicated times. Image and quantification is from one experiment. (H) Immunoblot analysis of substrates of conventional PKCa/b (p-IkBa Ser32 and p-RB Thr821/
826), novel PKCε (p-STAT3 Ser727 and p-PKD Ser744/748) and atypical PKCz (p-RKIP Ser153 and p-cPLA2 T376) isoforms following treatment of MCF10A
E545K and H1047R MUT cells with 1 mM of each PKCa, b, ε, and z peptide inhibitors for 72 hours. (I) Enzymatic activity of cPLA2, iPLA2, and sPLA2 in the MCF10A
PIK3CA isogenic panel following treatment with 100 nM ASB14780 for 72 hours. (J) AA levels measured by REIMS in MCF10A E545K MUT cells treated with
100 nM ASB14780, 1 mM each of PKCa, b, ε, and z peptide inhibitors, 250 mM GSK650394, or 150 nM MK2206 for 72 hours under exogenous FAF conditions. (K)
Immunoblot (right) of total PKCz and phospho-S505 and T376 cPLA2 of MCF10A PIK3CA WT and MUT cells following PKCz siRNA-mediated knockdown, and
cPLA2 activity (left) following 48 hours post-transfection. (L) Representative phospho-PKCz Thr560 immunoreactivity images (left) of 9 PIK3CA MUT (blue) and 9
WT (red) breast PDX tumors. Scale bar = 250 mm. Quantification of percent positive regions (right) was performed using the IHC profiler plug-in for ImageJ. Data
are presented as the mean ± SEM of n = 3-5 biological replicates and are representative of at least two independent experiments. n.s., not significant, *p % 0.05;
**p % 0.01. P values in (C), (D), (E), (F), (I), (J) and (K, right) with one-way ANOVA followed by unpaired, two-tailed Student’s t test with Bonferroni correction, and in
(L) with unpaired, two-tailed Student’s t test.
 ll
OPEN ACCESS Article

Figure S6. Related to Figures 4 and 5

(A) Immunoblot analysis of phospho-Tyr783 PLCg1 in MCF10A PIK3CA isogenics following serum and growth factor deprivation for 16 hours, and stimulation
with serum and growth factors for 30 min. Densitometry values are either scaled to unstimulated or stimulated (bold) WT samples. (B) Measurement of intracellular
calcium flux in MCF10A PIK3CA isogenics following serum and growth factor deprivation and stimulation for 30 min. (C) Immunoblot of MCF10A PIK3CA WT and
MUT cells following siRNA-mediated knockdown of PLCg1. (D) Intracellular calcium flux of MCF10A PIK3CA WT (top), E545K (middle) and H1047R (bottom) cells
(legend continued on next page)
 ll
Article OPEN ACCESS

48 hours post transfection with siPLCg1, or (E) treatment with 2 mM U73122 for 24 hours. For the final 18 hours of the treatments, cells were serum and growth
factor deprived, and stimulated with full media immediately prior to the assay. cPLA2 activity in MCF10A PIK3CA WT and MUT following (F) siRNA-mediated
knockdown of PLCg1 for 48 hours, or (G) treatment with 2 mM U73122 for 24 hours. (H) AA levels measured by REIMS in MCF10A PIK3CA isogenics following
treatment with 2 mM U73122 for 24 hours. (I) Representative confocal images and (J) quantification of in situ proximity ligation assay (PLA) between cPLA2 and
phospho-Thr560 PKCz in MCF10A PIK3CA WT and MUT cells. (K) Immunoblot analysis of phospho-cPLA2 (T376) custom antibody in the MCF10A isogenic panel
following serum and growth factor deprivation for 18 hours and subsequent stimulation for 30 min (left), treatment with 1 mM PKCz peptide inhibitor for 72 hours
(middle), and in MCF10A H1047R cPLA2 CRISPR knockout cells overexpressing a phosphoresistant mutant (T376A) cPLA2 (right). (L) Activity of cPLA2 in
MCF10A PIK3CA WT or H1047R cPLA2 CRISPR knockout cells transfected with 9 mg of either WT-cPLA2, or S505A/T376A phosphoresistant mutant cPLA2
constructs. Activity was measured 48 hours post-transfection. Cell proliferation of MCF10A (M) PIK3CA WT and (N) E545K MUT cells expressing control shGFP,
cPLA2-sh1 or sh5 under exogenous FAF conditions. Sulforhodamine B (SRB) protein staining was used to measure cell proliferation over 5 days. Data in (B), (D),
(E), (F), (G), (H), (L), (M) and (N) are presented as the mean ± SEM of n = 3-6 biological replicates and are representative of at least two independent experiments.
n.s., not significant; *p < 0.05; **p < 0.01; ***p < 0.001; P values in (B), (D), (E), (M) and (N) were calculated using two-way ANOVA. One-way ANOVA followed by
Student’s t test with Bonferroni correction was used for (F), (G), (H), (J)and (L).
 ll
OPEN ACCESS Article

(legend on next page)

 ll
Article OPEN ACCESS

Figure S7. Related to Figures 6 and 7

(A) Relative tumor growth and (B) tumor weights of CAL51 (PIK3CA MUT)-derived xenografts stably expressing control shGFP or two independent shRNAs
targeting cPLA2 (cPLA2-sh1 and cPLA2-sh5) in mice fed a balanced omega3:omega6 diet. (C) Relative tumor growth and (D) tumor weights of Hs578T (PIK3CA
WT)-derived xenografts stably expressing control shGFP or cPLA2-sh1 or cPLA2-sh5 in mice fed a balanced omega3:omega6 diet. AA levels measured by
REIMS in (E) PIK3CA MUT (CAL51) and (F) PIK3CA WT (Hs578T) snap frozen excised tumors. AA intensities are reported as scaled values to the appropriate
shGFP-fat free diet control. Quantification of CCL5 from excised tumors derived from (G) PDX and (H) cell line-derived xenograft studies. Quantification of
CX3CL1 from excised tumors derived from (I) PDX and (J) cell line-derived xenograft studies. Concentrations of chemokines in (G-J) were determined from whole
tumor lysates using ELISA, and normalized to protein content. (K) Representative immunohistochemical staining of the activated NK cell marker NKp46 in BR1282
(PIK3CA MUT) and BR1458 (PIK3CA WT) PDX tumors. (L) and (M) Quantification of positively immunostained areas from (K). (N) Representative immunohis-
tochemical staining of NKp46 in shGFP, cPLA2-sh1 and cPLA2-sh5 expressing CAL51 (PIK3CA MUT) and Hs578T (PIK3CA WT)-derived xenograft tumors under
fat free or ‘Western’ diets. (O) and (P) Quantification of positively immunostained areas from (N). Data in (A), (C), and (E) to (J) are presented as the mean ± SEM of
n = 3–5 mice for cell line xenograft or n = 7–8 mice for PDX studies. n.s., not significant; *p < 0.05; **p < 0.01; ***p < 0.001; P values in (A) and (C) were calculated
using two-way ANOVA, and one-way ANOVA followed by unpaired, two tailed Student’s t test with Bonferroni correction was used in (B), (D), (E), (F), (G), (H), (I), (J),
(L), (M), (O), (P).
 Article

Intratumoral CD4+ T Cells Mediate Anti-tumor

Cytotoxicity in Human Bladder Cancer
Graphical Abstract Authors
David Y. Oh, Serena S. Kwek,
Siddharth S. Raju, ...,
Terence W. Friedlander, Chun Jimmie Ye,
Lawrence Fong

Correspondence
jimmie.ye@ucsf.edu (C.J.Y.),
lawrence.fong@ucsf.edu (L.F.)

In Brief
Single-cell RNA and paired T cell receptor
sequencing highlights enrichment of
cytotoxic CD4+ T cells rather than CD8+
T cells in human bladder cancer. These
CD4+ T cells are capable of killing
autologous tumor cells and are subjected
to inhibition by Tregs.

Highlights
d Human bladder tumors contain multiple clonally expanded
cytotoxic CD4+ T cell states

d Cytotoxic CD4+ T cells can kill autologous tumors in an MHC

class II-dependent fashion

d Autologous regulatory T cells can inhibit the activity of

cytotoxic CD4+ T cells

d A cytotoxic CD4+ gene signature predicts response to anti-

PD-L1 in bladder cancer

Oh et al., 2020, Cell 181, 1612–1625

June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.017 ll
 ll

Article
Intratumoral CD4+ T Cells Mediate Anti-tumor
Cytotoxicity in Human Bladder Cancer
David Y. Oh,1,11 Serena S. Kwek,1,11 Siddharth S. Raju,3,6,8 Tony Li,1 Elizabeth McCarthy,3 Eric Chow,4 Dvir Aran,5
Arielle Ilano,1 Chien-Chun Steven Pai,1,12 Chiara Rancan,1 Kathryn Allaire,1 Arun Burra,1 Yang Sun,3
Matthew H. Spitzer,2,6,8 Serghei Mangul,9 Sima Porten,7 Maxwell V. Meng,7 Terence W. Friedlander,1
Chun Jimmie Ye,2,3,5,10,* and Lawrence Fong1,2,13,*
1Division of Hematology/Oncology, Department of Medicine, University of California, San Francisco, San Francisco, CA 94143, USA
2Parker Institute for Cancer Immunotherapy, University of California, San Francisco, San Francisco, CA 94143, USA
3Division of Rheumatology, Department of Medicine; Department of Epidemiology and Biostatistics; and Institute for Human Genetics,

University of California, San Francisco, San Francisco, CA 94143, USA

4Department of Biochemistry and Biophysics, Center for Advanced Technologies, University of California, San Francisco, San Francisco, CA

94143, USA
5Bakar Computational Health Sciences Institute, University of California, San Francisco, San Francisco, CA 94143, USA
6Department of Microbiology and Immunology, University of California, San Francisco, San Francisco, CA 94143, USA
7Department of Urology, University of California, San Francisco, San Francisco, CA 94143, USA
8Department of Otolaryngology – Head and Neck Surgery, University of California, San Francisco, San Francisco, CA 94143, USA
9Department of Computer Science, University of California, Los Angeles, Los Angeles, CA 90095, USA
10Chan Zuckerberg Biohub, San Francisco, CA, USA
11These authors contributed equally
12Present address: Department of Oncology Biomarker Development, Genentech, 1 DNA Way, South San Francisco, CA 94080, USA
13Lead Contact

*Correspondence: jimmie.ye@ucsf.edu (C.J.Y.), lawrence.fong@ucsf.edu (L.F.)

https://doi.org/10.1016/j.cell.2020.05.017

SUMMARY

Responses to anti-PD-1 immunotherapy occur but are infrequent in bladder cancer. The specific T cells that
mediate tumor rejection are unknown. T cells from human bladder tumors and non-malignant tissue were as-
sessed with single-cell RNA and paired T cell receptor (TCR) sequencing of 30,604 T cells from 7 patients. We
find that the states and repertoires of CD8+ T cells are not distinct in tumors compared with non-malignant
tissues. In contrast, single-cell analysis of CD4+ T cells demonstrates several tumor-specific states, including
multiple distinct states of regulatory T cells. Surprisingly, we also find multiple cytotoxic CD4+ T cell states
that are clonally expanded. These CD4+ T cells can kill autologous tumors in an MHC class II-dependent
fashion and are suppressed by regulatory T cells. Further, a gene signature of cytotoxic CD4+ T cells in tu-
mors predicts a clinical response in 244 metastatic bladder cancer patients treated with anti-PD-L1.

INTRODUCTION cytes (TILs) and their differential ability to confer a therapeutic

Immunotherapies have changed the landscape of cancer
treatment by producing durable and long-lasting responses
through triggering anti-tumor cell-mediated immunity. In
particular, checkpoint inhibitors (CPIs) targeting the immune
inhibitory molecules CTLA-4 and PD-1 in T lymphocytes
have been approved based on responses and improved over-
all survival in multiple malignancies, particularly those with a
high mutational burden (Hodi et al., 2010; Herbst et al.,
2014; Powles et al., 2014; Robert et al., 2015; Martincorena
and Campbell, 2015; Cancer Genome Atlas Research
Network, 2008). However, in specific malignancies, such as
transitional cell carcinoma (TCC) of the bladder, CPIs as
monotherapies are efficacious in only
20% of patients
(Powles et al., 2014; Hargadon et al., 2018). This could be
partly due to the heterogeneity of tumor-infiltrating T lympho-
benefit upon treatment.
Currently, cytotoxic CD8+ T cells are the main focus of efforts
to understand how immunotherapy elicits anti-tumor immunity.
In melanoma, expression and chromatin state signatures of
cytotoxicity and exhaustion (Tirosh et al., 2016; Philip et al.,
2017; Ayers et al., 2017; Herbst et al., 2014) and the presence
of CD8+ T cells at the tumor-invasive margin pre-treatment (Tu-
meh et al., 2014) are significantly correlated with subsequent re-
sponses to PD-1-directed therapy. However, in metastatic
bladder TCC, where response rates to PD-1 blockade are

15%–20% in platinum chemotherapy-refractory patients and
more than 20% in frontline platinum-ineligible patients, predic-
tive biomarkers of response are unclear, including PD-L1
expression (Koshkin and Grivas, 2018). Recently, bulk RNA
sequencing (RNA-seq) of the pre-treatment tumor microenviron-
ment in TCC found that a higher score of CD8+ gene signature

1612 Cell 181, 1612–1625, June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Article
and tumor mutational burden and, conversely, a lower score of
transforming growth factor b (TGF-b) gene signature, particularly
in immune-excluded tumors, were associated with response to
the anti-PD-L1 agent atezolizumab (Mariathasan et al., 2018).
However, the importance of heterogeneous subsets of TILs in
TCC beyond canonical CD8+ cytotoxic and exhausted pheno-
types in response to PD-1 blockade remains unexplored. In
particular, the role of CD4+ T cells in controlling or enhancing
TCC tumor growth remains largely unknown. Although regulato-
ry CD4+ T cells (Tregs) in the TCC environment have been asso-
ciated with adverse outcomes (Baras et al., 2016), and a CD4+
subset expressing the inducible costimulator (ICOS) that pro-
duces interferon-gamma (IFNg) in response to anti-CTLA-4 ther-
apy has been described in human bladder tumors (Liakou et al.,
2008), the presence of other CD4+ T cell subsets that directly
promote cell-mediated immunity through other effector mecha-
nisms remains unclear. Detailed characterization of the T lym-
phocytes in the tumor is critically needed for precisely mapping
the cells responsible for tumor recognition and control and
defining predictive markers of response to CPI in bladder cancer.
To address these points, we interrogated the tumor microen-
vironment of patients with localized muscle-invasive bladder
TCC who received or did not receive neoadjuvant anti-PD-L1
immunotherapy (atezolizumab, Genentech) prior to surgical
resection. Droplet single-cell RNA-seq (dscRNA-seq) and paired
T cell receptor sequencing (TCR-seq) of more than 30,000 CD4+
and CD8+ T cells from paired tumors and adjacent non-malig-
nant tissues revealed heterogeneity in known CD4+ states,
such as regulatory T cells, which were also enriched and clonally
expanded in tumors. In addition, several states of cytotoxic
CD4+ T cells expressing cytolytic effector proteins were identi-
fied, some of which are enriched in tumors. Cytotoxic CD4+
T cells were clonally expanded in tumors and could kill autolo-
gous tumors ex vivo. Cytotoxic CD4+ T cells existed in discrete
proliferating and non-proliferating states in tumors. A gene
signature of cytotoxic CD4+ T cells was predictive of a response
to PD-1 blockade in an orthogonal RNA-seq dataset of metasta-
tic bladder cancer patients treated with anti-PD-L1. Overall,
these findings highlight the importance of CD4+ T cell heteroge-
neity and the relative balance between activation of cytotoxic
CD4+ effectors and inhibitory regulatory cells for killing autolo-
gous tumors.
RESULTS
Canonical CD8+ T Cell States Were Not Enriched in the
Bladder Tumor Microenvironment
To assess the T cell composition of the tumor environment, we
profiled T cells from dissociated bladder tumors and adjacent
uninvolved bladder tissues using single-cell RNA and TCR
sequencing (see schematic in Figure S1A). We used the 10X Ge-
nomics Chromium platform (Zheng et al., 2017b) to sequence
8,833 tumor-derived and 1,929 non-malignant tissue-derived
CD8+ T cells from 7 patients (Table S1). All samples were mus-
cle-invasive bladder cancer (MIBC) from 2 standard-of-care-un-
treated patients (‘‘untreated’’), 1 chemotherapy-treated patient
(gemcitabine + carboplatin, ‘‘chemo’’), and 4 anti-PD-L1-treated
patients (‘‘anti-PD-L1’’) with detailed clinical annotations (Table
ll
S1). To assess the shared heterogeneity of T cells across sam-
ples, we restricted the analysis to highly variable genes and
used an empirical Bayes approach (ComBat; Johnson et al.,
2007; Büttner et al., 2019) to account for preparation batch
among individual samples. We subsequently used Leiden clus-
tering (Traag et al., 2019) to define clusters that were visualized
using uniform manifold approximation and projection (UMAP)
(McInnes and Healy, 2018). Tumor- and non-malignant-derived
CD8+ T cells form 11 clusters, each populated by cells from all
samples suggestive of shared states in TCC regardless of the
treatment regimen (Figure 1A; Figure S2A). Differential expres-
sion analyses comparing each cluster with all other cells com-
bined identified 1,067 differentially expressed genes in at least
one cluster (adjusted P value (Padj) < 0.05, |log2(fold change,
FC)| > 0.5) (Table S2). The identified states include known CD8+
subtypes (Figures 1B and 1C): cells expressing HAVCR2 (TIM-
3), LAG3, ENTPD1, as well as the chemokine CXCL13
(CD8ENTPD1: log2(FC) = 1.4–3.7), described previously as tumor-
reactive CD8+ T cells (Duhen et al., 2018); effector cells express-
ing FGFBP2 and GNLY, a granule-associated pore-forming pro-
tein known to function in pathogen killing (Krensky and Clay-
berger, 2009) (CD8FGFBP2: log2(FC) = 3.6–5.3); naive cells
expressing CCR7 and GZMK (CD8NAIVE: log2(FC) = 0.9–2.8); cen-
tral memory cells expressing CCR7 and SELL (L-selectin)
(CD8CM: log2(FC) = 1.5–1.7); and mucosal-associated invariant
T (MAIT) cells expressing KLRB1 (CD8MAIT: log2(FC) = 2.7) that
preferentially use the known semi-invariant TCR a chains
TRAV1-2 and/or TRAJ33 (Kurioka et al., 2016; Figure 1D). Of
note, we also found MKI67+ proliferating cells (CD8PROLIF: log2(-
FC) = 6.5) as well as cells expressing the chemokines XCL1/2
(CD8xcl: log2(FC) = 5.2–5.6). Similar states were also identified
in the tumor environment of hepatocellular carcinoma based on
scRNA-seq (Zheng et al., 2017a). Surprisingly, although the fre-
quency of CD8ENTPD1 cells was higher in tumors, none of the
CD8+ states displayed statistically significant differences in fre-
quency between the tumor and non-malignant bladder (exact
permutation test; Figure 1E; density plots in Figure 1F).
Tregs Included Heterogeneous States that Are Enriched
in Bladder Tumors
Given the lack of tumor enrichment of CD8+ states and the higher
frequency of CD4+ over CD8+ T cells in bladder tumors (Fig-
ure S1B), we investigated CD4+ T cell heterogeneity in a similar
fashion to determine their contribution to anti-tumor responses.
We sequenced and analyzed 16,995 tumor- and 2,847 non-ma-
lignant tissue-infiltrating CD4+ T cells isolated from the same pa-
tients. Tumor-derived and non-malignant tissue-derived CD4+
T cells formed 11 clusters each with representation from all indi-
vidual patients (Figure 2A; Figure S2B). We identified 1,511
differentially expressed genes in at least one cluster (Padj <
0.05, |log2(FC)| > 0.5; Table S2; Figures 2B and 2C) defining
several canonical CD4+ T cell states. These include CCR7+ cells,
which demonstrated a central memory phenotype (CD4CM)
based on parallel flow cytometry data showing that these were
CD45RA (see below) as well as cells expressing high levels of
CXCL13 and IFNG (CD4CXCL13: log2(FC) = 5.9 and 1.4), which
were also likely to be exhausted based on overexpression of
TOX (log2(FC) = 1.9) (Yao et al., 2019) and whose presence has
Cell 181, 1612–1625, June 25, 2020 1613
 ll
Article

A B C

D E
F

Figure 1. Bladder Cancer Contains Canonical CD8+ T Cell States

(A) Uniform manifold approximation and projection (UMAP) plots of 10,762 single sorted CD3+ CD8+ T cells obtained from bladder tumors and adjacent non-
malignant tissue (N = 7 patients). Phenotypic clusters are represented in distinct colors.
(B) Relative intensity of expression of select genes superimposed on the UMAP projections in (A).
(C) Violin plots showing the relative expression of select differentially expressed genes (columns) for each cluster shown in (A) (rows) (all Padj < 0.05).
(D) The frequency of cells expressing MAIT-associated TRAV1-2/TRAJ33+ TCRs within each defined CD8+ phenotypic cluster.
(E) The frequency of cells in individual clusters shown as a proportion of total CD8+ cells within tumor or non-malignant compartments across all patients (orange,
tumor; blue, non-malignant). For each cluster, a box and whisker plot is shown with the median, interquartile range (IQR, a box with lower and upper bounds
representing 25th and 75th percentiles, respectively), and 1.5 times the IQR (whiskers). Outlier points are shown if more than 1.5 times the IQR beyond the lower
and upper quartiles. Statistical testing was done using an exact permutation test.
(F) Density plots showing distribution of cells in tumor or non-malignant samples.

been associated with improved outcomes in breast, gastric, and
microsatellite-unstable colorectal carcinoma, which is an im-
mune-responsive tumor (Schmidt et al., 2018; Gu-Trantien
et al., 2013, 2017; Wei et al., 2018; Zhang et al., 2018). Other
states included Th17 cells expressing IL17A (CD4TH17: log2(FC) =
4.7), which represented important anti-tumor effectors (Kryczek
et al., 2011); activated cells expressing CD69 (CD4ACTIVATED:
log2(FC) = 2.2) but not FOXP3 (log2(FC) < 0.5) (Figures 2B and
2C); as well as several important additional states described in
further detail below. Notably, some of these states were selec-
tively enriched in specific compartments. CD4CXCL13 demon-
strated significant enrichment in tumor compared with non-ma-
lignant tissue (tumor versus non-malignant: 6.5% versus 3.0%,
p = 0.015, exact permutation test), whereas states enriched in
non-malignant tissue included CD4CM (tumor versus non-malig-
nant: 30% versus 42%, p = 0.008) and CD4ACTIVATED (tumor
versus non-malignant: 7.5% versus 10%, p = 0.02) (density plots
in Figure 2D; tumor and non-malignant frequencies in Figure 2E).
Tregs were abundant constituents of the bladder tumor micro-
environment with demonstrated heterogeneity. We found two
states of Tregs (CD4IL2RAHI and CD4IL2RALO), together consti-
tuting 26% ± 1.9% (mean ± SEM) of tumor-infiltrating CD4+ cells,
which co-expressed FOXP3 (CD4IL2RAHI: log2(FC) = 2.7;
CD4IL2RALO: log2(FC) = 1.2) and known immune checkpoints,
including IL2RA, TIGIT, TNFRSF4/9/18, and CD27 (Philip et al.,
2017; Zheng et al., 2017a; Plitas et al., 2016; De Simone et al.,
2016) (CD4IL2RAHI and CD4IL2RALO: log2(FC) > 0.65; Figures 2B,
2C, and 3A). With the exception of TIGIT, these immune check-
points are minimally expressed by other CD4+ states, such as
CD4CM (Figure 3A). The two Treg states were distinguished by
higher expression of IL2RA, TNFRSF4, TNFRSF9, and
TNFRSF18 in CD4IL2RAHI cells (CD4IL2RAHI: log2(FC) = 2.5–3.6;
CD4IL2RALO: log2(FC) = 0.4–1.6; Figure 3A; Table S2). Of note,
both Treg states were significantly enriched in tumor compared
with adjacent non-malignant tissue (CD4IL2RAHI: 14.3% versus
4.6%, p = 0.002; CD4IL2RALO: 11.1% versus 6.7%, p = 0.002;
exact permutation test; Figure 2E). We confirmed, by flow cytom-
etry from 7 additional bladder tumors, that multiple tumors con-
tained distinct regulatory states that expressed graded protein
levels of IL2RA and co-expressed significantly different levels
of immune checkpoints, such as TNFRSF18 (p < 0.05 for
TNFRSF18 expression in FOXP3+ CD25low versus CD25hi

1614 Cell 181, 1612–1625, June 25, 2020

 ll
Article

A B

D E

Figure 2. CD4+ T Cells in Bladder Tumors Are Composed of Multiple Distinct Functional States
(A) UMAP plots of 19,842 single sorted CD3+ CD4+ T cells obtained from bladder tumors and adjacent non-malignant tissue (N = 7 patients). Each distinct
phenotypic cluster identified using Leiden clustering is identified with a distinct color. Annotation of each unbiased cluster was performed by manual inspection of
the highest-ranked differentially expressed genes for each cluster and using reference signature-based correlation methods (SingleR) as described in the text.
(B) Relative intensity of expression of select genes superimposed on the UMAP projections shown in (A).
(C) Violin plot showing relative expression of select differentially expressed genes (columns) for each cluster shown in (A) (rows) (all Padj < 0.05).
(D) Density plots showing distribution of cells in tumor or non-malignant samples.
(E) The frequency of cells in individual CD4+ T cell states defined by scRNA-seq clustering is shown as a proportion of total CD4+ cells within either tumor or non-
malignant compartments across all patients (orange, tumor; blue, non-malignant). A box and whisker plot is shown with formatting as in Figure 1E. *p < 0.05,
**p < 0.01 by exact permutation test.

populations by Wilcoxon signed-ranked test; Figure 3B; gating
strategy in Figures S1C–S1D). This heterogeneity may be conse-
quential because Tregs expressing higher levels of immune
checkpoints have been shown to be correlated with poorer out-
comes in non-small cell lung cancer (Guo et al., 2018). Both reg-
ulatory states also demonstrated a common tumor-specific gene
expression program that included several heat shock proteins
compared with non-malignant tissue (Figure S2C; Table S2).
Tregs Are Clonally Expanded in Bladder Tumors
To query the TCR sequence in the same single cells for which
we obtained whole-transcriptome data, we PCR-amplified and
sequenced to saturation the complementarity-determining re-
gion 3 (CDR3) of the TCR alpha (TRA) and beta (TRB) loci
from the barcoded full-length cDNA library (primers in Table
S3). After filtering for matching whitelisted cell barcodes (Cell
Ranger), this approach yielded 11,081 CD4+ T cells and 5,779
CD8+ T cells with paired TRA and TRB CDR3 sequences
(49% and 47% recovery, respectively; summary in Table S3).
These results are consistent with expected frequencies based
on the average recovery of individual TRA (CD4+, 54%; CD8+,
50%) and TRB (CD4+, 68%; CD8+, 67%) sequences across
whitelisted cells. Overall, the TCR repertoire was more
restricted in the tumor microenvironment than in adjacent
non-malignant tissue based on two analyses. First, in intratu-
moral CD4+ T cells, 10.8% ± 1.6% of unique clonotypes are
shared by 2 or more cells; this degree of sharing was signifi-
cantly greater than in the non-malignant compartment (5.1%
± 1.6%, unpaired t test, p = 0.033) and is not seen in blood
from healthy donors (0.12%–0.16%) or from publicly available
reference circulating CD4+ T cell data (0%) (Figure S3A). Sec-
ond, we observed skewing of the intratumoral CD4+ T cell

Cell 181, 1612–1625, June 25, 2020 1615

 ll
Article

A B

Figure 3. Regulatory CD4+ T Cells Are Heterogeneous, Enriched, and Clonally Expanded in Bladder Tumors
(A) Heatmap showing the expression of select regulatory T cell marker genes (rows) for individual single cells (columns) within the CD4IL2RAHI and CD4IL2RLO
clusters compared with the CD4CM cluster. Cells were grouped based on their annotations by tissue (tumor or non-malignant), treatment, and patient. Log2-
transformed expression of each gene was row scaled.
(B) Flow cytometry staining of CD4+ FOXP3+ TILs from a bladder tumor, showing the gating strategy for CD25neg, CD25low, and CD25hi (top left), and histograms of
TNFRSF18 staining from each CD25 gate (top right). Mean fluorescence intensity of TNFRSF18 and percent TNFRSF18+ from the parental gate are shown for
CD25 gates across samples (N = 7 tumors, mean ± SEM). *p < 0.05 by Wilcoxon paired t test.
(C) Gini coefficients for regulatory populations (CD4ILRA2HI and CD4IL2RALO, red labels at far left) and other CD4+ T cell populations within tumor and non-malignant
compartments across all samples. For each cluster, a box and whisker plot is shown with the median, IQR (box), and 1.5 times the IQR (whiskers), with outliers
exceeding 1.5 times the IQR beyond lower and upper quartiles. *p < 0.05, **p < 0.01 by exact permutation test. N = 7 tumor samples and 6 non-malignant
samples.
(D) Left: single cells expressing the top 3 most expanded clonotypes found in the combined regulatory populations (CD4ILRA2HI and CD4IL2RALO) are shown in red
in the same UMAP space as in Figure 2A. The regions composed of regulatory, cytotoxic, and proliferating T cells are outlined and superimposed on the UMAP
projection. Right: density plots for total CD4+ T cell distribution within tumor and non-malignant compartments are reproduced from Figure 2D for ease of visual
comparison.

repertoire toward an increased cumulative frequency of clono-
types over fewer cells (Figure S3B) and a corresponding higher
Gini coefficient (0.21 for tumor versus 0.05 for non-malignant
tissue, Wilcoxon signed-rank test with Benjamini-Hochberg
correction, p = 0.009; Figure S3C) compared with the non-ma-
lignant compartment and healthy controls.
When we assigned TCR sequences to cells with cluster
identities (9,770 CD4+ and 5,151 CD8+ T cells with a paired
TRA/TRB had an assigned phenotypic cluster or 49% and
48% of all T cells with assigned clusters, respectively; merged
TCR sequences and phenotypic clusters for CD4+ and CD8+
T cells in Table S4), we found that clonal expansion of Tregs

1616 Cell 181, 1612–1625, June 25, 2020

 ll
Article

B D

C E

(legend on next page)

Cell 181, 1612–1625, June 25, 2020 1617

 ll
Article

contributes to intratumoral CD4+ T cell repertoire restriction.
Compared with paired non-malignant tissue, both regulatory
states exhibited increased Gini coefficients in tumors
(CD4IL2RAHI: Ginitumor 0.17 versus Gininormal 0, p = 0.003;
CD4IL2RALO: Ginitumor 0.06 versus Gininormal 0.003, p = 0.009;
exact permutation test; Figure 3C). The most expanded clono-
types within the Tregs were specific to regulatory cells but not
other cell states (all single cells expressing the top expanded
regulatory clonotypes are shown in Figure 3D). The CXCL13-
expressing state CD4cxcl13 (discussed in greater detail below)
was also restricted in tumors (Ginitumor 0.07 versus Gininormal
0, p = 0.02, exact permutation test; Figure 3C). Gini coeffi-
cients for CD4+ states did not differ significantly by anti-PD-
L1 treatment (Figure S3G). In contrast, although repertoire
restriction was also seen in CD8+ T cells from the same
samples, this was observed in both tumor (percent unique clo-
notypes shared between cells: 15.1% ± 1.1%; Ginitumor:
0.36% ± 0.04%) and non-malignant compartments (percent
unique clonotypes shared between cells: 14.6% ± 0.2%;
Gininormal: 0.39% ± 0.06; Figures S3D–S3F). Furthermore, no
significant increase in Gini coefficient in tumor over non-ma-
lignant tissue was seen for any CD8+ state, including with
anti-PD-L1 treatment (Figures S3H and S3I). Hence, an impor-
tant contributor to increased repertoire restriction of tumor-
infiltrating CD4+ over non-malignant tissue, which was not
seen in the CD8+ compartment, involved clonal expansion of
several distinct regulatory T cell states that differed in their
levels of immune checkpoint expression, which may be
driven by tumor-associated antigens and the tumor-specific
microenvironment.
Bladder Tumors Possessed Multiple Cytotoxic CD4+ T
Cell States
In addition to regulatory states, we also found 2 distinct states of
cytotoxic CD4+ T cells in all samples constituting 15% ± 0.9% of
tumor-infiltrating CD4+ T cells. CD4GZMB and CD4GZMK cytotoxic
cells expressed a core set of cytolytic effector molecules (log2(-
FC) > 0.5, Padj < 0.05): GZMA (granzyme A), GZMB (granzyme B),
and NKG7 (a granule protein that translocates to the surface of
natural killer (NK) cells following target cell recognition; Medley
et al., 1996) (Figures 2B, 2C, and 4A; Table S2). Each cytotoxic
CD4+ state was distinguished by the expression of specific
effector molecules. CD4GZMB cells co-expressed high levels of
GZMB, the pore-forming protein PRF1 (perforin), and the
granule-associated proteins GNLY and NKG7 (CD4GZMB: log2(-
FC) = 5.7, 3.4, 5.1, and 4.4, respectively), whereas CD4GZMK cells
co-expressed high levels of the distinct granzyme GZMK and
lower levels of NKG7 (CD4GZMK: log2(FC) = 6.3 and 3.9) (Fig-
ure 4A; Table S2). These shared cytolytic molecules were not ex-
pressed by other CD4+ states, including regulatory and central
memory T cells (Figure 4A). Cytotoxic CD4+ cells co-expressed
additional molecules, which may further contribute to anti-tumor
effector function. Notably, IFNG was expressed by both cyto-
toxic states, which may contribute to tumor cell death, including
ferroptosis (Wang et al., 2019) (CD4GZMB and CD4GZMK: log2(-
FC) = 2.1). Of note, the minority of CD4GZMB cells that expressed
IFNG appeared to also express TNF as well as specific immune
checkpoints, such as PDCD1, LAG3, and HAVCR2 (TIM3) (Fig-
ure 4A). A larger proportion of CD4GZMB cells expressed
CXCR6 (CD4GZMB: log2(FC) = 1.3; Figure 4A). This chemokine
is expressed in regulatory and non-regulatory CD4+ TILs from
colorectal carcinoma, nasopharyngeal carcinoma, and renal
cell carcinoma and, together with its ligand CXCL16, can
mediate TIL chemotaxis (Löfroos et al., 2017; Parsonage et al.,
2012; Oldham et al., 2012). Finally, CD4GZMB and CD4GZMK cells
did not express high levels of other checkpoints associated with
regulatory T cells, such as IL2RA, TIGIT, or TNFRSF4/9/18
(log2(FC) < 0.5; Figure 4A), nor did they express the exhaustion
marker TOX (Table S2). Similar states were found with unbiased
clustering without batch correction for paired tumor- and

Figure 4. Multiple Cytotoxic CD4+ T Cell States Are Enriched and Clonally Expanded in Bladder Tumors and Possess Lytic Capacity against
Tumors
(A) Heatmap showing the expression of select cytotoxic or regulatory T cell marker genes (rows) for individual single cells (columns) within the cytotoxic CD4GZMB
and CD4GZMK clusters compared with regulatory (CD4IL2RAHI and CD4IL2RLO) and CD4CM clusters. Cells were grouped based on their annotations by tissue (tumor
or non-malignant), treatment, and patient. Log2-transformed expression of each gene was row scaled.
(B) Flow cytometry staining of GZMB, perforin, or GZMK in CCR7 CD4+ FOXP3 T cells.
(C) Percentage of cells expressing GZMB, GZMK, or perforin from CCR7 CD4+ FOXP3 T cells by flow cytometry (left) and the percentage of cells co-expressing
perforin within GZMB+ or GZMK+ CCR7 CD4+ FOXP3 T cells (right) (N = 7 tumors, mean + SEM).
(D) Representative flow cytometry staining of IFNg and TNF-a expression in GZMB+ or GZMK+ CCR7 CD4+ FOXP3 T cells stimulated with PMA and ionomycin.
(E) Percentages of cells expressing IFNg, TNF-a, or both from GZMB+ or GZMK+ CCR7 CD4+ FOXP3 T cells with and without stimulation (N = 11 tumors,
mean + SEM).
(F) Multiplex immunofluorescent staining of DAPI (blue), CD4 (immunohistochemistry, red), GZMK (RNAscope probe, green), and GZMB (RNAscope probe, white)
and overlay without DAPI from a cystectomy tumor region from a patient with parallel scRNA-seq and TCR-seq data (anti-PD-L1 C, top row) and from a cor-
responding tumor field with negative control staining (bottom row). CD4+ cells that co-express GZMK (arrows) or GZMB (arrowhead) are indicated. Scale
bar, 10 mm.
(G) The ratio of abundances of all regulatory T cell populations (CD4ILRAHI and CD4IL2RALO) to all cytotoxic CD4+ populations (CD4GZMB and CD4GZMK) across all
tumor and non-malignant samples (mean + SEM shown; *p < 0.05 by unpaired t test, assuming unequal variance).
(H) Gini coefficients for each of the cytotoxic CD4+ populations within tumor and non-malignant compartments across all samples (box and whisker plot is shown
with formatting as in Figure 3C; *p < 0.05, **p < 0.01, exact permutation test, N = 7 tumor samples and 6 non-malignant samples).
(I) Left panel: quantitation of Annexin V+ apoptotic cells over time from a time-lapse cytotoxicity experiment with tumor cells cultured alone or with bulk CD4+ TILs
(CD4total) or CD4+ TILs depleted of regulatory T cells (CD4eff) at a 30:1 effector:target ratio. Right panel: CD4eff TILs and tumor cells (30:1 effector:target ratio) were
co-cultured with a pan-anti-MHC class II antibody or isotype control. All traces were from the same culture and cytotoxicity assay from the same patient. All traces
show relative change in cell death from time point 0. Cytotoxicity with CD4eff is representative of independent experiments from 4 different patients. Mean ± SEM
from multiple technical replicates for each experiment is shown.

1618 Cell 181, 1612–1625, June 25, 2020


Article
non-malignant-derived CD4+ cells from individual patients (Fig-
ures S2D and S2E).
We validated the presence and functional heterogeneity of
cytotoxic CD4+ T cells using several orthogonal and comple-
mentary methods. Using flow cytometry, the presence of cyto-
toxic CD4+ T cells with an effector memory (CCR7 CD45RA)
or effector (CCR7 CD45RA+) phenotype that express GZMB,
GZMK, and perforin protein was confirmed by flow cytometry
in tumors from multiple independent replicate samples (N = 7 tu-
mors; Figure 4B; gating strategy in Figures S1C and S1D).
Across this sample set, 9% ± 2.9% (mean ± SEM) of CD4+
FOXP3 CCR7 cells expressed GZMB, whereas 16% ± 4.5%
expressed GZMK and 5.3% ± 2.6% expressed perforin (Fig-
ure 4C, left panel), at lower frequencies than CCR7 CD8+ cyto-
toxic cells from the same patients (Figures S1E–S1G). Impor-
tantly, 25.9% ± 8.7% of GZMB+ CD4+ FOXP3 CCR7 and
8.6% ± 3.5% of GZMK+ CD4+ FOXP3 CCR7 cytotoxic
T cells showed co-expression of perforin with granzymes, in
agreement with the scRNA-seq data (Figure 4C, right panel);
these frequencies of granzyme and perforin co-expression
were lower than those of CCR7 CD8+ cytotoxic cells from the
same patients (Figures S1F and S1G). Importantly, CD45
bladder tumor cells express multiple major histocompatibility
complex (MHC) class II molecules (data not shown), which would
allow antigen recognition by TCRs expressing CD4 as a co-re-
ceptor. Flow cytometry of a separate set of 11 muscle-invasive
bladder tumors confirms the functional capacity of cytotoxic
CD4+ T cells to produce multiple cytokines. In agreement with
the scRNA-seq data, 56% ± 4.8% (mean ± SEM) of CD4+
CCR7 cells were polyfunctional and could produce both IFNg
and tumor necrosis factor alpha (TNF-a), whereas a minority of
these cells only secrete IFNg alone or TNF-a alone after stimula-
tion and, therefore, may demonstrate signs of exhaustion (IFNg+
TNF-a: 2.0% ± 0.76%; IFNg TNF-a+: 19% ± 3.3%) (Figures
4D and 4E). The frequency of polyfunctional cytotoxic CD4+
T cells was similar to stimulated CD8+ CCR7 T cells from the
same patients (IFNg+ TNF-a+: 55% ± 6.3%), although CD8+
CCR7 T cells that were monofunctional demonstrated an
increased trend toward preferential IFNg production alone over
TNF-a production compared with cytotoxic CD4+ T cells
(IFNg+ TNF-a: 14% ± 4.7%; IFNg TNF-a+: 7.2% ± 2.1%)
(Figure S1H).
As further validation of the cytotoxic CD4+ T cell phenotype in
tissue, multiplex immunofluorescence tissue staining of bladder
tumor tissue from a patient in the scRNA-seq dataset demon-
strated CD4+ T cells that also expressed GZMB or GZMK (Fig-
ure 4F, top row; tissue staining from an additional patient in Fig-
ure S1I) at levels not seen with negative control staining
(Figure 4F, bottom row).
Overall annotation of clusters from the scRNA-seq data was
supported by an independent analysis that assigns each single
cell to the best-known published immune subset profiled by
bulk expression analysis after sorting (SingleR) (Aran et al.,
2019). This corroborated the identification of Tregs (90% and
78% of CD4IL2RAHI and CD4IL2RALO cells are assigned to Treg an-
notations, respectively) and further demonstrated that both cyto-
toxic CD4+ states are most similar to CD8+ effector memory
T cells (37% and 45% of CD4GZMB and CD4GZMK cells, respec-
ll
tively, are assigned to effector memory CD8+ cell annotations),
reinforcing their cytotoxicity profile (Figure S2F). Finally, an inter-
nal comparison of the transcriptional profiles from CD4+ and
CD8+ TIL clusters from our scRNA-seq data indicated that,
although most CD4+ clusters are most similar to other CD4+ clus-
ters, cytotoxic CD4+ T cells are an exception. CD4GZMB cytotoxic
cells were most correlated with tumor-specific CD8ENTPD1 cells
(Pearson correlation coefficient = 0.92), whereas CD4GZMK cyto-
toxic cells were most correlated with CD8CM and CD8NAIVE cells
(Pearson correlation coefficient = 0.98 for both) (Figure S2G). The
tumor-specific gene expression program of these cytotoxic
CD4+ cells was marked by heat shock protein expression in
both states as well as tumor overexpression of CXCL13 and
numerous immune checkpoints (TNFRSF18/LAG3/TIGIT/
HAVCR2) as well as ENTPD1 within CD4GZMB cells (Figure S2C;
Table S2).
Cytotoxic CD4+ T Cells Were Enriched and Clonally
Expanded in Bladder Tumors
Of the 2 cytotoxic CD4+ states, CD4GZMK cells were significantly
enriched in abundance in tumors (CD4GZMK in tumor versus non-
malignant tissues: 7.2% ± 0.5% versus 5.0% ± 0.5%, exact per-
mutation test, p = 0.01; Figure 2E). Overall, the CD4+ compart-
ment exhibited a bias toward regulatory over cytotoxic CD4+
T cells in tumors (regulatory CD4+/cytotoxic CD4+ ratio = 1.8 ±
0.2) compared with non-malignant tissues, where proportions
of regulatory and cytotoxic CD4+ T cells were more balanced
(regulatory CD4+/cytotoxic CD4+ ratio = 1.1 ± 0.2, t test, p =
0.04; Figure 4G). Cytotoxic CD4+ T cell states contributed to in-
tratumoral CD4+ repertoire restriction. Both cytotoxic CD4+
T cell states have significantly increased Gini coefficients in tu-
mor compared with non-malignant tissues, with CD4GZMB repre-
senting the more restricted cytotoxic state in tumors (CD4GZMB:
Ginitumor 0.21 versus Gininormal 0.06; CD4GZMK: Ginitumor 0.12
versus Gininormal 0; exact permutation test, p = 0.04 for CD4GZMB
and p = 0.002 for CD4GZMK; Figure 4H). Hence, unbiased
dscRNA-seq revealed that heterogeneous cytotoxic CD4+
T cells, a subset of which are closely related to conventional
cytotoxic CD8+ T cells based on their functional program, are un-
expected but frequent constituents of the bladder tumor micro-
environment, some of which are quantitatively enriched in tu-
mors. The tumor-specific clonal expansion of both cytotoxic
CD4+ states suggests that their restricted repertoire may result
from recognition of MHC class II cognate antigens that may
include bladder tumor antigens.
Cytotoxic CD4+ T Cells Possessed Lytic Capacity
against Autologous Tumor Cells that Was Restricted by
Autologous Tregs
To validate the functional relevance of cytotoxic CD4+ in bladder
tumors, we isolated CD4+ TILs by fluorescence-activated cell
sorting (FACS) and then cultured the cells ex vivo with inter-
leukin-2 (IL-2). We then co-cultured these cells with autologous
tumor cells in an imaging-based time-lapse cytotoxicity assay,
assessing for cell death with Annexin V. We found that expanded
CD4+ TILs were cytotoxic and could trigger increased tumor
apoptosis (‘‘CD4total:tumor,’’ Figure 4I, left panel). However,
when we performed the same co-cultures but with CD4+ TILs
Cell 181, 1612–1625, June 25, 2020 1619
 ll
Article

A B

C D

E
F

Figure 5. Proliferating CD4+ T Cells Contain Regulatory and Cytotoxic Cell States
(A) Heatmap showing expression of select cytotoxic, regulatory, and proliferating marker genes (rows) for individual single cells (columns) within the CD4PROLIF
cluster. Samples were hierarchically clustered. Log2-transformed expression of each gene was row scaled.
(B) Representative flow cytometry staining from a bladder tumor showing expression of CD25, GZMB, GZMK, and Ki67.
(C) Single cells expressing the top 3 most expanded clonotypes found in the CD4PROLIF T cell population are shown in red in the same UMAP space as in
Figure 2A. The regions composed of proliferating, regulatory, and cytotoxic T cells are outlined and superimposed on the UMAP projection for visualization.
(D) Left panel: pseudotime trajectories derived from all tumors (N = 7 samples) and non-malignant samples (N = 6 samples). Cells with expanded TCRs from the
proliferating (CD4PROLIF, green), regulatory (CD4IL2RAHI and CD4IL2RALO, shades of red), and cytotoxic (CD4GZMB and CD4GZMK, shades of purple) states were
used for this analysis. Specific branches corresponding to proliferating cytotoxic cells (top right), non-proliferating cytotoxic cells (bottom right), proliferating
regulatory cells (top left), and non-proliferating regulatory cells (bottom left) are labeled. Right panel: branches are color-coded according to the above prolif-
erating or non-proliferating identities. Also labeled are branch points that discriminate proliferating and non-proliferating cytotoxic CD4+ T cells (branch point 1)
and proliferating and non-proliferating regulatory T cells (branch point 2).
(legend continued on next page)

1620 Cell 181, 1612–1625, June 25, 2020

 ll
Article

from the same patient that were depleted of Tregs, we found that
killing was increased (‘‘CD4eff:tumor,’’ Figure 4I, left panel), indi-
cating that autologous Tregs can inhibit the activity of cytotoxic
CD4+ T cells. Significant tumor death was seen in co-cultures
with CD4eff TILs compared with tumors alone (Figure 4I, left
panel; representative of 3 independent experiments from
different patients). Furthermore, the cytotoxic activity of CD4eff
was at least partially dependent on MHC class II recognition
because tumor apoptosis was inhibited with pre-incubation
with a pan-anti-MHC class II antibody that was not seen with
an isotype control antibody (Figure 4I, right panel; representative
of 2 independent patients). Independent experiments with an
alternative death indicator (Cytotox Red) confirmed increased
autologous tumor killing with tumor/CD4eff co-cultures (Fig-
ure S4A), MHC class II dependence of CD4eff killing (Figure S4B),
as well as similar MHC class I-dependent autologous tumor
killing with expanded CD8+ T cells (Figures S4C and S4D).
Hence, flow cytometry and functional analyses from multiple in-
dependent patients confirmed not only that cytotoxic CD4+
T cells expressed cytolytic proteins, such as granzymes and per-
forin, in tumor tissue but that these cells can recognize bladder
tumor antigens in an MHC class II-dependent fashion and
were functionally competent to lyse autologous tumor cells in a
manner that can be suppressed by autologous Tregs.
Proliferating CD4+ T Cells Contained Regulatory and
Cytotoxic Cells
Induction of proliferating T cells can be beneficial for anti-tumor
immune responses. Proliferating CD4+ T cells are rapidly
induced in the periphery within weeks of initiating checkpoint
blockade in prostate cancer patients (Kavanagh et al., 2008)
and in separate cohorts of thymic epithelial tumors and non-
small cell lung cancer treated with anti-PD-1; a higher fold
change in Ki67+ cells among PD-1+ CD8+ T cells in the periph-
ery after a week was predictive of durable clinical benefit, pro-
gression-free survival, and (in the non-small cell lung cancer
cohorts) overall survival (Kim et al., 2019). Within our tumor-
infiltrating CD4+ T cell compartment in TCC, we also identified
proliferating cells (CD4PROLIF) expressing MKI67, microtubule-
associated markers (e.g., STMN1/TUBB), and DNA-binding
proteins associated with cell cycle progression, such as
PCNA, HMGB1, and HMGB2, which were expressed at lower
levels in regulatory and cytotoxic CD4+ T cells (CD4PROLIF:
log2(FC) > 2.1; Figure 2C; Table S2). A similar signature was
also seen in the CD8+ compartment (CD8PROLIF; Figure 1C; Ta-
ble S2). Higher-resolution clustering revealed that this prolifer-
ating state is comprised of discrete groups of cells co-express-
ing regulatory or cytotoxic genes but not both simultaneously
(Figure 5A). Flow cytometry analysis of separate TCC samples
confirmed the presence of discrete regulatory or cytotoxic pop-
ulations of Ki67+ CD4+ T cells that co-expressed CD25, GZMB,
or GZMK (Figure 5B). Across multiple independent samples,
4.7% ± 1.0% (mean ± SEM) of CD4+ FOXP3+ cells co-ex-
pressed Ki67 and CD25, whereas 1.2% ± 0.5% of CD4+
FOXP3 CCR7 cells co-expressed Ki67 and GZMB, and
1.0% ± 0.1% of CD4+ FOXP3 CCR7 cells co-expressed
Ki67 and GZMK (N = 7 tumors; Figure S1J). Proliferating
Ki67+ GZMB+ cells are also seen, using flow cytometry, within
the CD8+ compartment of TCC patients (Figure S1K). Examina-
tion of exact TCR clonotype sharing of the most expanded
CD4PROLIF clones identified sharing with regulatory and cyto-
toxic CD4+ T cells, further underscoring the contribution of
each state to CD4PROLIF cells (Figure 5C).
Given that regulatory and cytotoxic CD4+ T cells were heterog-
enous and composed of cells that were proliferating to a different
extent, existing clusters may fail to resolve the separate contri-
bution of specific expression programs from subsets with
different proliferative capacity. Hence, we used pseudotime
analysis to separate regulatory and cytotoxic cells into prolifer-
ating and non-proliferating components (Qiu et al., 2017). This
analysis divided CD4PROLIF cells into two groups, each lying
along a branch specific for proliferating regulatory or cytotoxic
CD4+ T cells, with separate branches for non-proliferating regu-
latory and cytotoxic cells (Figure 5D). This underscored that reg-
ulatory and cytotoxic CD4+ T cells consist of distinct proliferating
and non-proliferating states in TCC, based on transcriptomic
and clonotypic analyses.
A Signature of Cytotoxic CD4+ T Cells Predicts Clinical
Response to Anti-PD-L1
To assess the importance of the specific proliferating and non-
proliferating cytotoxic CD4+ T cell states for patient outcomes,
we performed branched expression analysis modeling (BEAM)
to identify all genes that were differentially expressed between
branches at branchpoint 1 of the pseudotime trajectory. This
branchpoint divided proliferating cytotoxic CD4+ T cells, non-
proliferating cytotoxic CD4+ T cells, and all other regulatory cells
(Figure 5D, right panel). Hierarchical clustering identified genes
upregulated preferentially in the proliferating cytotoxic branch
(cluster 7) or the non-proliferating cytotoxic branch (cluster 4)
but not in regulatory branches within this analysis (all genes
with q < 0.05; heatmap of clusters and branches in Figure 5E;
branch-specific signatures in Table S5). We developed a gene
signature from this analysis consisting of genes that were upre-
gulated specifically in proliferating or non-proliferating cytotoxic
CD4+ T cells (from cluster 7: ABCB1; from cluster 4: APBA2,
SLAMF7, GPR18, and PEG10; Figure 5E) but were not upregu-
lated in any of the CD8+ T cell states from our scRNA-seq

(E) Heatmap showing all differentially expressed genes (columns) between branches for branch point 1 across cells in the pseudotime analysis (rows). Cells are
grouped by their proliferating or non-proliferating branch assignments, color-coded at the right of the heatmap and corresponding to colors in (D). Genes are
grouped by color-coded clusters (1–8) shown at the top of the plot, which result from hierarchical clustering based on co-regulation in specific branches.
(F) Cytotoxic CD4+ T cell gene signature scores were plotted in clinical responders (complete response or partial response) versus non-responders (stable
disease or progressive disease) from baseline metastatic biopsies from bladder cancer patients with inflamed tumors on the IMvigor210 clinical trial (N = 62
tumors). The signature score was obtained from the IMvigor210 bulk RNA-seq dataset for the cytotoxic CD4+ T cell-specific genes derived from non-proliferating
(cluster 4) and proliferating (cluster 7) cytotoxic CD4+ clusters from the pseudotime analysis shown below the heatmap in (E). Median ± SEM is shown; *p = 0.037
by two-tailed t test.

Cell 181, 1612–1625, June 25, 2020 1621

 ll
analysis (Table S2). We then tested this gene signature’s ability
to predict treatment response using bulk RNA-seq data from
pre-treatment tumors from a separate phase 2 trial of atezolizu-
mab for metastatic bladder cancer (IMvigor210; Mariathasan
et al., 2018). In 244 metastatic bladder cancer patients with
pre-treatment RNA-seq data, immunohistochemistry (IHC) infor-
mation regarding immune phenotype (immune desert, immune-
excluded, or inflamed), and information regarding clinical
response, this gene signature was significantly correlated with
clinical response to anti-PD-L1 therapy in inflamed samples
(p = 0.037, two-tailed t test, N = 62 inflamed samples; Figure 5F),
which was not seen in samples with an immune-excluded or im-
mune desert phenotype. Hence, we used a composite signature
containing genes that discriminated proliferating and non-prolif-
erating cytotoxic CD4+ T cells to assess the specific contribu-
tions of these discrete states and found that this signature is
associated with response to PD-1 blockade in a large cohort of
TCC patients. This result highlights the potential clinical impor-
tance of possessing intratumoral cytotoxic CD4+ T cell activity
in response to anti-PD-L1 treatment.
DISCUSSION
Current efforts to dissect the mechanism of tumor immune sur-
veillance and enhance the efficacy of cancer immunotherapies
have primarily focused on conventional cytotoxic CD8+ T cell-
mediated responses. However, given the known functional di-
versity of CD4+ T cell effector responses and emerging data
that CD4+ T cell recognition may be important for anti-tumor re-
sponses (for instance, in the context of a neoantigen vaccine; Ott
et al., 2017; Sahin et al., 2017), the role of specific CD4+ states in
enhancing or suppressing immune responses in the tumor
microenvironment and how these are modulated by systemic
therapies, including immunotherapy, remains unknown. Here
we use unbiased massively parallel genotypic and phenotypic
profiling of the T cell compartment in localized bladder tumors
and the adjacent non-malignant compartment, including those
treated with anti-PD-L1 immunotherapy, as a tool to finely
dissect heterogeneity in CD4+ T cell subsets. We identified spe-
cific CD4+ T cell states with functional relevance for response to
immunotherapy and clinical outcomes. We not only confirmed
the presence of CD4+ T cell states with known contributions to
anti-tumor immune responses, such as CXCL13+ CD4+ T cells
(Schmidt et al., 2018; Gu-Trantien et al., 2013, 2017; Wei et al.,
2018; Zhang et al., 2018) as well as Th17 cells (Kryczek et al.,
2011), we also uncovered insights into the contribution of
CD4+ TILs to tumor control by the immune system in bladder
cancer.
First we identified distinct states of Tregs that differed based
on the level of expression of IL2RA and immune checkpoints,
such as TNFRSF18, which was confirmed at the protein level.
These Tregs possessed a private repertoire with no detected
clonotype sharing with other T cell states, which would suggest
that these are not induced Tregs. Because a gene signature from
checkpoint-high Tregs is associated with worse outcome in non-
small cell lung cancer (Guo et al., 2018), it is possible that these
regulatory cells are responsible for setting a basal state of more
potent immunosuppression and adverse outcomes in TCC.
1622 Cell 181, 1612–1625, June 25, 2020
Article
Second, we identified heterogenous states of cytotoxic CD4+
T cells that were unexpected and differed in their expression of
canonical cytolytic effector molecules (GZMB, GZMK, and
PRF1 [perforin]) as well as other granule-associated proteins
(GNLY [granulysin] and NKG7) that may have roles in target
cell killing. These were distinct populations based on scRNA-
seq and orthogonal validation by flow cytometry and multiplex
immunofluorescence tissue staining. Our annotation using Sin-
gleR indicated that effector states such as cytotoxic CD4+
T cells found in the tumor microenvironment may not yet be an-
notated, and, based on ‘‘best fit’’ comparisons with external
reference data and transcriptional correlation within our own
data, these cells were, in fact, most similar to conventional
effector memory cytotoxic CD8+ T cells. The functional similarity
between cytotoxic CD4+ T cells and conventional CD8+ T cells
was underscored by our finding that CD4GZMB TILs were actually
most similar to tumor-specific CD8ENTPD1 cells (Duhen et al.,
2018), based on transcriptional data, whereas CD4GZMK TILs
were most similar to CD8CM and CD8NAIVE cells. Although these
were distinct cell types, based on separate CD4 and CD8 co-re-
ceptor expression, this may indicate shared modes of tumor
recognition and tumor clearance by cytotoxic CD4+ and CD8+
T cells. Although cytotoxic CD4+ T cells are present in non-small
cell lung and hepatocellular carcinoma (Zheng et al., 2017a; Guo
et al., 2018), circulate with ipilimumab treatment in metastatic
melanoma (Kitano et al., 2013), and also are present in an infec-
tious context, where they represent a clonally expanded dengue
virus-specific effector subset (Patil et al., 2018), the extent of
their heterogeneity in other solid tumors (including bladder can-
cer) and whether these cells are important for systemic immuno-
therapy has remained unclear prior to this work. We found that
cytotoxic CD4+ subsets in bladder tumors were clonally
expanded, which may be the result of recognition of cognate
bladder tumor antigens. Their functional importance was
confirmed by their ability to kill autologous tumors ex vivo. The
mechanism by which these cells kill target tumor cells involves
contact-dependent mechanisms based on inhibition of killing
by anti-MHC class II antibodies, although other mechanisms
may also contribute. We documented that these cytotoxic
CD4+ T cells are polyfunctional and secrete multiple, such as
TNF-a and IFNg; the latter may contribute to tumor death as
well through ferroptosis (Wang et al., 2019) in addition to con-
tact-dependent cytotoxicity. Of note, apart from the subset of
cells that co-express TNF-a, IFNg, PDCD1, LAG3, and HAVCR2,
cytotoxic CD4+ T cells are found to generally lack surface
expression of many immune checkpoints currently being tested
with therapeutic antibodies in pre-clinical and clinical testing,
suggesting that these effector cells may have distinct require-
ments for activation.
Importantly, a gene signature derived from single-cell analysis
of proliferating and non-proliferating cytotoxic CD4+ T cells is
predictive of the response to anti-PD-L1 therapy in a separate
set of 62 patients with inflamed metastatic bladder cancer.
Most of these genes have been previously implicated in the
biology of cytotoxic effector cells or specific human CD4+
T cell responses in pathogenesis or autoimmunity, including hu-
man cytotoxic CD4+ T cells (Arlehamn et al., 2014; Burel et al.,
2018; Campbell et al., 2018, Imbeault et al., 2012; Mattoo

Article
et al., 2016; Sumida and Cyster, 2018; Wang et al., 2014). Over-
all, the predictive value of this cytotoxic CD4+ T cell-specific
signature in a large cohort of anti-PD-L1-treated metastatic
TCC patients highlights how anti-PD-L1 therapy may alter the
immune microenvironment to favor activation of cytotoxic
CD4+ effectors, particularly in patients with pre-existing cyto-
toxic CD4+ T cell activity.
The importance of the relative balance between regulatory and
effector T cells is well known for conventional effectors; the reg-
ulatory CD4+:cytotoxic CD8+ ratio has been associated with
improved survival or response to therapy in several cancers,
including bladder (Preston et al., 2013; Sato et al., 2005; Baras
et al., 2016; Takada et al., 2018). This work identifies the biolog-
ical importance of another axis involving the relative balance of
regulatory T cells and these cytotoxic CD4+ effectors for anti-tu-
mor activity: removal of regulatory T cells enhanced tumor killing
by cytotoxic CD4+ T cells. Our findings suggest that manipu-
lating the balance between cytotoxic CD4+ and regulatory
T cell states can lead to therapeutic benefit in TCC.
Finally, the origin of cytotoxic CD4+ T cell effectors within tu-
mors remains unclear. We do not find direct evidence of plas-
ticity or interconversion of regulatory T cells into cytotoxic
CD4+ T cells, based on clonotype sharing. Cytotoxic CD4+
T cells do share clones with the proliferating CD4+ state, raising
the possibility that these cells may arise from activation of other
CD4+ subsets, whether within the tumor milieu itself or as a result
of tumor homing of precursors, which are first activated outside
of the tumor in the peripheral circulation.
There are important limitations to the interpretation of this
study. The size of the sample set used for single-cell discovery
of T cell heterogeneity was limited to 7 patients; hence, larger-
scale single-cell sequencing efforts in bladder cancer will help
to validate these findings. The treatments administered before
collection were also heterogeneous; as a result, given the limita-
tions in sample size, our ability to directly assess modulation of
T cell subsets, such as cytotoxic CD4+ T cells, by immuno-
therapy in this dataset is limited. Finally, the scope of our findings
in this dataset is limited to patients with MIBC; further efforts will
be needed to assess the immune context in other bladder cancer
disease states. Nonetheless, the robustness of our findings
across the individual patients in this dataset highlight conserved
CD4+ heterogeneity across patients, which is important for im-
mune recognition of bladder tumors.
This work lays an important conceptual foundation for ef-
forts to enhance bladder tumor immunotherapy. We identify
cytotoxic CD4+ effectors whose distinct expression of cyto-
lytic molecules and other marker genes will lead to further ef-
forts to isolate and enhance the activity of specific cytotoxic
subsets as well as to discover the bladder tumor antigens
they are recognizing. At the same time, this work points to
specific regulatory T cell states that may be more suppressive
in bladder cancer and therefore represent ideal targets for
parallel approaches to inhibit their activity. Collectively, our
findings point to the importance of understanding multiple
axes that balance suppressive regulatory T cell activity with
effector function of anti-tumor immune subsets in TCC to
enhance our ability to effectively manipulate these with thera-
peutic approaches to enhance tumor control.
ll
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
d METHOD DETAILS
B Tissue processing
B Flow cytometry/FACS
B Single cell RNA sequencing
B TCR sequencing
B Expression analysis
B TCR analysis
B Tumor infiltrating lymphocyte (TIL) isolation and
culturing
B Cytotoxic T lymphocyte (CTL) killing assay
B Pseudotime analysis
B Gene signature analysis
B RNAscope/tissue immunofluorescence staining
d QUANTIFICATION AND STATISTICAL ANALYSIS
d ADDITIONAL RESOURCES
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.017.
ACKNOWLEDGMENTS
We thank the patients who volunteered to participate in these studies; the
UCSF Genitourinary Medical Oncology and Urology providers involved in
screening, enrollment, and clinical care of these patients; the UCSF Bio-
specimen Resources Program for assistance with tissue acquisition; and the
Institute for Human Genetics Core for assistance with sequencing. This work
was supported by the Parker Institute for Cancer Immunotherapy. D.Y.O. is
supported by NIH T32CA177555 and K08AI139375, the Harry F. Bisel, MD En-
dowed Young Investigator Award from the Conquer Cancer Foundation of the
American Society of Clinical Oncology, the Bladder Cancer Advocacy Network
Palm Beach New Discoveries Young Investigator Award, and the Prostate
Cancer Foundation Young Investigator Award. S.S.K. is supported by the Pe-
ter Michael Foundation. L.F. is support by the Parker Institute of Cancer Immu-
notherapy; the Prostate Cancer Foundation; and NIH R01CA194511,
R01CA223484, U01CA233100, and U01CA244452.
AUTHOR CONTRIBUTIONS
Study concept and design, D.Y.O., S.S.K., and L.F.; Acquisition of Patients,
Samples, and Data, D.Y.O., S.S.K., A.I., C.-C.S.P., C.R., K.A., A.B., S.P.,
M.V.M., and T.W.F.; Data Analysis and Interpretation, D.Y.O., S.S.K., S.S.R.,
T.L., E.M., E.C., D.A., A.I., K.A., A.B., Y.S., M.H.S., S.M., C.J.Y., and L.F.;
Writing of the Manuscript, D.Y.O., S.S.K., C.J.Y., and L.F.; Study Oversight,
C.J.Y. and L.F.
DECLARATION OF INTERESTS
D.Y.O. has received research support from Roche/Genentech and Merck and
has served as a paid consultant for Maze Therapeutics. L.F. has received
research support from Roche/Genentech, Abbvie, Bavarian Nordic, Bristol
Cell 181, 1612–1625, June 25, 2020 1623
 ll
Myers Squibb, Janssen, and Merck. C.J.Y. is a co-founder of Dropprint
Genomics.
Received: August 2, 2019
Revised: February 21, 2020
Accepted: May 8, 2020
Published: June 3, 2020
REFERENCES
Aran, D., Looney, A.P., Liu, L., Wu, E., Fong, V., Hsu, A., Chak, S., Naikawadi,
R.P., Wolters, P.J., Abate, A.R., et al. (2019). Reference-based analysis of lung
single-cell sequencing reveals a transitional profibrotic macrophage. Nat. Im-
munol. 20, 163–172.
Arlehamn, C.L., Seumois, G., Gerasimova, A., Huang, C., Fu, Z., Yue, X., Sette,
A., Vijayanand, P., and Peters, B. (2014). Transcriptional profile of tuberculosis
antigen-specific T cells reveals novel multifunctional features. J. Immunol. 193,
2931–2940.
Ayers, M., Lunceford, J., Nebozhyn, M., Murphy, E., Loboda, A., Kaufman,
D.R., Albright, A., Cheng, J.D., Kang, S.P., Shankaran, V., et al. (2017). IFN-
g-related mRNA profile predicts clinical response to PD-1 blockade. J. Clin.
Invest. 127, 2930–2940.
Baras, A.S., Drake, C., Liu, J.J., Gandhi, N., Kates, M., Hoque, M.O., Meeker,
A., Hahn, N., Taube, J.M., Schoenberg, M.P., et al. (2016). The ratio of CD8 to
Treg tumor-infiltrating lymphocytes is associated with response to cisplatin-
based neoadjuvant chemotherapy in patients with muscle invasive urothelial
carcinoma of the bladder. OncoImmunology 5, e1134412.
Bolotin, D.A., Poslavsky, S., Mitrophanov, I., Shugay, M., Mamedov, I.Z., Pu-
tintseva, E.V., and Chudakov, D.M. (2015). MiXCR: software for comprehen-
sive adaptive immunity profiling. Nat. Methods 12, 380–381.
Burel, J.G., Lindestam Arlehamn, C.S., Khan, N., Seumois, G., Greenbaum,
J.A., Taplitz, R., Gilman, R.H., Saito, M., Vijayanand, P., Sette, A., and Peters,
B. (2018). Transcriptomic Analysis of CD4+ T Cells Reveals Novel Immune Sig-
natures of Latent Tuberculosis. J. Immunol. 200, 3283–3290.
Büttner, M., Miao, Z., Wolf, F.A., Teichmann, S.A., and Theis, F.J. (2019). A test
metric for assessing single-cell RNA-seq batch correction. Nat. Methods
16, 43–49.
Campbell, K.S., Cohen, A.D., and Pazina, T. (2018). Mechanisms of NK Cell
Activation and Clinical Activity of the Therapeutic SLAMF7 Antibody, Elotuzu-
mab in Multiple Myeloma. Front. Immunol. 9, 2551.
Cancer Genome Atlas Research Network (2008). Comprehensive genomic
characterization defines human glioblastoma genes and core pathways. Na-
ture 455, 1061–1068.
De Simone, M., Arrigoni, A., Rossetti, G., Gruarin, P., Ranzani, V., Politano, C.,
Bonnal, R.J.P., Provasi, E., Sarnicola, M.L., Panzeri, I., et al. (2016). Transcrip-
tional Landscape of Human Tissue Lymphocytes Unveils Uniqueness of Tu-
mor-Infiltrating T Regulatory Cells. Immunity 45, 1135–1147.
Duhen, T., Duhen, R., Montler, R., Moses, J., Moudgil, T., de Miranda, N.F.,
Goodall, C.P., Blair, T.C., Fox, B.A., McDermott, J.E., et al. (2018). Co-expres-
sion of CD39 and CD103 identifies tumor-reactive CD8 T cells in human solid
tumors. Nat. Commun. 9, 2724.
Gu-Trantien, C., Loi, S., Garaud, S., Equeter, C., Libin, M., de Wind, A., Ravoet,
M., Le Buanec, H., Sibille, C., Manfouo-Foutsop, G., et al. (2013). CD4+ follic-
ular helper T cell infiltration predicts breast cancer survival. J. Clin. Invest. 123,
2873–2892.
Gu-Trantien, C., Migliori, E., Buisseret, L., de Wind, A., Brohée, S., Garaud, S.,
Noël, G., Chi, V.L.D., Lodewyckx, J.N., Naveaux, C., et al. (2017). CXCL13-
producing TFH cells link immune suppression and adaptive memory in human
breast cancer. JCI Insight 2, 91487.
Guo, X., Zhang, Y., Zheng, L., Zheng, C., Song, J., Zhang, Q., Kang, B., Liu, Z.,
Jin, L., Xing, R., et al. (2018). Global characterization of T cells in non-small-cell
lung cancer by single-cell sequencing. Nat. Med. 24, 978–985.
1624 Cell 181, 1612–1625, June 25, 2020
Article
Hargadon, K.M., Johnson, C.E., and Williams, C.J. (2018). Immune checkpoint
blockade therapy for cancer: An overview of FDA-approved immune check-
point inhibitors. Int. Immunopharmacol. 62, 29–39.
Herbst, R.S., Soria, J.C., Kowanetz, M., Fine, G.D., Hamid, O., Gordon, M.S.,
Sosman, J.A., McDermott, D.F., Powderly, J.D., Gettinger, S.N., et al. (2014).
Predictive correlates of response to the anti-PD-L1 antibody MPDL3280A in
cancer patients. Nature 515, 563–567.
Hodi, F.S., O’Day, S.J., McDermott, D.F., Weber, R.W., Sosman, J.A., Haanen,
J.B., Gonzalez, R., Robert, C., Schadendorf, D., Hassel, J.C., et al. (2010).
Improved survival with ipilimumab in patients with metastatic melanoma.
N. Engl. J. Med. 363, 711–723.
Imbeault, M., Giguère, K., Ouellet, M., and Tremblay, M.J. (2012). Exon level
transcriptomic profiling of HIV-1-infected CD4(+) T cells reveals virus-induced
genes and host environment favorable for viral replication. PLoS Pathog. 8,
e1002861.
Johnson, W.E., Li, C., and Rabinovic, A. (2007). Adjusting batch effects in mi-
croarray expression data using empirical Bayes methods. Biostatistics 8,
118–127.
Kavanagh, B., O’Brien, S., Lee, D., Hou, Y., Weinberg, V., Rini, B., Allison, J.P.,
Small, E.J., and Fong, L. (2008). CTLA4 blockade expands FoxP3+ regulatory
and activated effector CD4+ T cells in a dose-dependent fashion. Blood 112,
1175–1183.
Kim, K.H., Cho, J., Ku, B.M., Koh, J., Sun, J.M., Lee, S.H., Ahn, J.S., Cheon, J.,
Min, Y.J., Park, S.H., et al. (2019). The First-week Proliferative Response of Pe-
ripheral Blood PD-1+CD8+ T Cells Predicts the Response to Anti-PD-1 Therapy
in Solid Tumors. Clin. Cancer Res. 25, 2144–2154.
Kitano, S., Tsuji, T., Liu, C., Hirschhorn-Cymerman, D., Kyi, C., Mu, Z., Allison,
J.P., Gnjatic, S., Yuan, J.D., and Wolchok, J.D. (2013). Enhancement of tumor-
reactive cytotoxic CD4+ T cell responses after ipilimumab treatment in four
advanced melanoma patients. Cancer Immunol. Res. 1, 235–244.
Koshkin, V.S., and Grivas, P. (2018). Emerging Role of Immunotherapy in
Advanced Urothelial Carcinoma. Curr. Oncol. Rep. 20, 48.
Krensky, A.M., and Clayberger, C. (2009). Biology and clinical relevance of
granulysin. Tissue Antigens 73, 193–198.
Kryczek, I., Zhao, E., Liu, Y., Wang, Y., Vatan, L., Szeliga, W., Moyer, J., Klimc-
zak, A., Lange, A., and Zou, W. (2011). Human TH17 cells are long-lived
effector memory cells. Sci. Transl. Med. 3, 104ra100.
Kurioka, A., Walker, L.J., Klenerman, P., and Willberg, C.B. (2016). MAIT cells:
new guardians of the liver. Clin. Transl. Immunology 5, e98.
Liakou, C.I., Kamat, A., Tang, D.N., Chen, H., Sun, J., Troncoso, P., Logothetis,
C., and Sharma, P. (2008). CTLA-4 blockade increases IFNgamma-producing
CD4+ICOShi cells to shift the ratio of effector to regulatory T cells in cancer pa-
tients. Proc. Natl. Acad. Sci. USA 105, 14987–14992.
Löfroos, A.B., Kadivar, M., Resic Lindehammer, S., and Marsal, J. (2017).
Colorectal cancer-infiltrating T lymphocytes display a distinct chemokine re-
ceptor expression profile. Eur. J. Med. Res. 22, 40.
Mariathasan, S., Turley, S.J., Nickles, D., Castiglioni, A., Yuen, K., Wang, Y.,
Kadel, E.E., III, Koeppen, H., Astarita, J.L., Cubas, R., et al. (2018). TGFb atten-
uates tumour response to PD-L1 blockade by contributing to exclusion of
T cells. Nature 554, 544–548.
Martincorena, I., and Campbell, P.J. (2015). Somatic mutation in cancer and
normal cells. Science 349, 1483–1489.
Mattoo, H., Mahajan, V.S., Maehara, T., Deshpande, V., Della-Torre, E., Wal-
lace, Z.S., Kulikova, M., Drijvers, J.M., Daccache, J., Carruthers, M.N., et al.
(2016). Clonal expansion of CD4(+) cytotoxic T lymphocytes in patients with
IgG4-related disease. J. Allergy Clin. Immunol. 138, 825–838.
McInnes, L., and Healy, J. (2018). UMAP: Uniform Manifold Approximation and
Projection for Dimension Reduction. arXiv, arXiv: 1802.03426. https://arxiv.
org/abs/1802.03426.
Medley, Q.G., Kedersha, N., O’Brien, S., Tian, Q., Schlossman, S.F., Streuli,
M., and Anderson, P. (1996). Characterization of GMP-17, a granule mem-
brane protein that moves to the plasma membrane of natural killer cells
following target cell recognition. Proc. Natl. Acad. Sci. USA 93, 685–689.

Article
Monaco, G., Lee, B., Xu, W., Mustafah, S., Hwang, Y.Y., Carré, C., Burdin, N.,
Visan, L., Ceccarelli, M., Poidinger, M., et al. (2019). RNA-Seq Signatures
Normalized by mRNA Abundance Allow Absolute Deconvolution of Human Im-
mune Cell Types. Cell Rep. 26, 1627–1640.e7.
Oldham, K.A., Parsonage, G., Bhatt, R.I., Wallace, D.M., Deshmukh, N.,
Chaudhri, S., Adams, D.H., and Lee, S.P. (2012). T lymphocyte recruitment
into renal cell carcinoma tissue: a role for chemokine receptors CXCR3,
CXCR6, CCR5, and CCR6. Eur. Urol. 61, 385–394.
Ott, P.A., Hu, Z., Keskin, D.B., Shukla, S.A., Sun, J., Bozym, D.J., Zhang, W.,
Luoma, A., Giobbie-Hurder, A., Peter, L., et al. (2017). An immunogenic per-
sonal neoantigen vaccine for patients with melanoma. Nature 547, 217–221.
Parsonage, G., Machado, L.R., Hui, J.W., McLarnon, A., Schmaler, T., Balaso-
thy, M., To, K.F., Vlantis, A.C., van Hasselt, C.A., Lo, K.W., et al. (2012). CXCR6
and CCR5 localize T lymphocyte subsets in nasopharyngeal carcinoma. Am. J.
Pathol. 180, 1215–1222.
Patil, V.S., Madrigal, A., Schmiedel, B.J., Clarke, J., O’Rourke, P., de Silva,
A.D., Harris, E., Peters, B., Seumois, G., Weiskopf, D., et al. (2018). Precursors
of human CD4+ cytotoxic T lymphocytes identified by single-cell transcriptome
analysis. Sci. Immunol. 3, eaan8664.
Philip, M., Fairchild, L., Sun, L., Horste, E.L., Camara, S., Shakiba, M., Scott,
A.C., Viale, A., Lauer, P., Merghoub, T., et al. (2017). Chromatin states define
tumour-specific T cell dysfunction and reprogramming. Nature 545, 452–456.
Plitas, G., Konopacki, C., Wu, K., Bos, P.D., Morrow, M., Putintseva, E.V., Chu-
dakov, D.M., and Rudensky, A.Y. (2016). Regulatory T Cells Exhibit Distinct
Features in Human Breast Cancer. Immunity 45, 1122–1134.
Powles, T., Eder, J.P., Fine, G.D., Braiteh, F.S., Loriot, Y., Cruz, C., Bellmunt,
J., Burris, H.A., Petrylak, D.P., Teng, S.L., et al. (2014). MPDL3280A (anti-PD-
L1) treatment leads to clinical activity in metastatic bladder cancer. Nature
515, 558–562.
Preston, C.C., Maurer, M.J., Oberg, A.L., Visscher, D.W., Kalli, K.R., Hart-
mann, L.C., Goode, E.L., and Knutson, K.L. (2013). The ratios of CD8+
T cells to CD4+CD25+ FOXP3+ and FOXP3- T cells correlate with poor clinical
outcome in human serous ovarian cancer. PLoS ONE 8, e80063.
Qiu, X., Mao, Q., Tang, Y., Wang, L., Chawla, R., Pliner, H.A., and Trapnell, C.
(2017). Reversed graph embedding resolves complex single-cell trajectories.
Nat. Methods 14, 979–982.
Robert, C., Long, G.V., Brady, B., Dutriaux, C., Maio, M., Mortier, L., Hassel,
J.C., Rutkowski, P., McNeil, C., Kalinka-Warzocha, E., et al. (2015). Nivolumab
in previously untreated melanoma without BRAF mutation. N. Engl. J. Med.
372, 320–330.
Sahin, U., Derhovanessian, E., Miller, M., Kloke, B.P., Simon, P., Löwer, M.,
Bukur, V., Tadmor, A.D., Luxemburger, U., Schrörs, B., et al. (2017). Personal-
ized RNA mutanome vaccines mobilize poly-specific therapeutic immunity
against cancer. Nature 547, 222–226.
Sato, E., Olson, S.H., Ahn, J., Bundy, B., Nishikawa, H., Qian, F., Jungbluth,
A.A., Frosina, D., Gnjatic, S., Ambrosone, C., et al. (2005). Intraepithelial
CD8+ tumor-infiltrating lymphocytes and a high CD8+/regulatory T cell ratio
are associated with favorable prognosis in ovarian cancer. Proc. Natl. Acad.
Sci. USA 102, 18538–18543.
Schmidt, M., Weyer-Elberich, V., Hengstler, J.G., Heimes, A.S., Almstedt, K.,
Gerhold-Ay, A., Lebrecht, A., Battista, M.J., Hasenburg, A., Sahin, U., et al.
ll
(2018). Prognostic impact of CD4-positive T cell subsets in early breast cancer:
a study based on the FinHer trial patient population. Breast Cancer Res. 20, 15.
Sumida, H., and Cyster, J.G. (2018). G-Protein Coupled Receptor 18 Contrib-
utes to Establishment of the CD8 Effector T Cell Compartment. Front. Immu-
nol. 9, 660.
Takada, K., Kashiwagi, S., Goto, W., Asano, Y., Takahashi, K., Takashima, T.,
Tomita, S., Ohsawa, M., Hirakawa, K., and Ohira, M. (2018). Use of the tumor-
infiltrating CD8 to FOXP3 lymphocyte ratio in predicting treatment responses
to combination therapy with pertuzumab, trastuzumab, and docetaxel for
advanced HER2-positive breast cancer. J. Transl. Med. 16, 86.
Tirosh, I., Izar, B., Prakadan, S.M., Wadsworth, M.H., 2nd, Treacy, D., Trom-
betta, J.J., Rotem, A., Rodman, C., Lian, C., Murphy, G., et al. (2016). Dissect-
ing the multicellular ecosystem of metastatic melanoma by single-cell RNA-
seq. Science 352, 189–196.
Traag, V.A., Waltman, L., and van Eck, N.J. (2019). From Louvain to Leiden:
guaranteeing well-connected communities. Sci. Rep. 9, 5233.
Tumeh, P.C., Harview, C.L., Yearley, J.H., Shintaku, I.P., Taylor, E.J., Robert,
L., Chmielowski, B., Spasic, M., Henry, G., Ciobanu, V., et al. (2014). PD-1
blockade induces responses by inhibiting adaptive immune resistance. Nature
515, 568–571.
Wang, X., Sumida, H., and Cyster, J.G. (2014). GPR18 is required for a normal
CD8aa intestinal intraepithelial lymphocyte compartment. J. Exp. Med. 211,
2351–2359.
Wang, W., Green, M., Choi, J.E., Gijón, M., Kennedy, P.D., Johnson, J.K., Liao,
P., Lang, X., Kryczek, I., Sell, A., et al. (2019). CD8+ T cells regulate tumour fer-
roptosis during cancer immunotherapy. Nature 569, 270–274.
Wei, Y., Lin, C., Li, H., Xu, Z., Wang, J., Li, R., Liu, H., Zhang, H., He, H., and Xu,
J. (2018). CXCL13 expression is prognostic and predictive for postoperative
adjuvant chemotherapy benefit in patients with gastric cancer. Cancer Immu-
nol. Immunother. 67, 261–269.
Wolf, F.A., Angerer, P., and Theis, F.J. (2018). SCANPY: large-scale single-cell
gene expression data analysis. Genome Biol. 19, 15.
Yao, C., Sun, H.W., Lacey, N.E., Ji, Y., Moseman, E.A., Shih, H.Y., Heuston,
E.F., Kirby, M., Anderson, S., Cheng, J., et al. (2019). Single-cell RNA-seq re-
veals TOX as a key regulator of CD8+ T cell persistence in chronic infection.
Nat. Immunol. 20, 890–901.
Zemmour, D., Zilionis, R., Kiner, E., Klein, A.M., Mathis, D., and Benoist, C.
(2018). Single-cell gene expression reveals a landscape of regulatory T cell
phenotypes shaped by the TCR. Nat. Immunol. 19, 291–301.
Zhang, L., Yu, X., Zheng, L., Zhang, Y., Li, Y., Fang, Q., Gao, R., Kang, B.,
Zhang, Q., Huang, J.Y., et al. (2018). Lineage tracking reveals dynamic rela-
tionships of T cells in colorectal cancer. Nature 564, 268–272.
Zheng, C., Zheng, L., Yoo, J.K., Guo, H., Zhang, Y., Guo, X., Kang, B., Hu, R.,
Huang, J.Y., Zhang, Q., et al. (2017a). Landscape of Infiltrating T Cells in Liver
Cancer Revealed by Single-Cell Sequencing. Cell 169, 1342–1356.e16.
Zheng, G.X., Terry, J.M., Belgrader, P., Ryvkin, P., Bent, Z.W., Wilson, R., Zir-
aldo, S.B., Wheeler, T.D., McDermott, G.P., Zhu, J., et al. (2017b). Massively
parallel digital transcriptional profiling of single cells. Nat. Commun. 8, 14049.
Cell 181, 1612–1625, June 25, 2020 1625
 ll
Article

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Brilliant Violet 605 CD25, clone BC96 Biolegend Cat# 302632
Brilliant Violet 786 CD127, clone A019D5 Biolegend Cat# 351330
Brilliant Violet 421 CD4, clone OKT4 Biolegend Cat# 317434
Brilliant Violet 650 CD3, clone UCHT1 Biolegend Cat# 300468
Brilliant Ultraviolet 395 CD45, clone H130 Becton Dickinson Cat# 563792
Alexa Fluor 647 CD8, clone SK1 Biolegend Cat# 344726
FITC GZMK, clone GM26E7 Biolegend Cat# 370508
PerCP-Cy5.5 HLA-DR, clone L243 Biolegend Cat# 307630
APC-R700 CCR7, clone 3D12 Becton Dickinson Cat# 565867
Brilliant Violet 480 CD3, clone UCHT1 Becton Dickinson Cat# 566105
Brilliant Violet 510 GZMB, clone GB11 Becton Dickinson Cat# 563388
Brilliant Violet 605 Ki67, clone Ki-67 Biolegend Cat# 350522
Brilliant Violet 650 CD45RA, clone HI100 Biolegend Cat# 304136
Brilliant Violet 786 CD25, clone BC96 Biolegend Cat# 302638
Brilliant violet 711 TNFSRF18, clone 108-17 Biolegend Cat# 371212
Brilliant ultraviolet 395 CD4, clone RPA-T4 Becton Dickinson Cat# 564724
Brilliant ultraviolet 496 CD8, clone RPA-T8 Becton Dickinson Cat# 564808
Brilliant ultraviolet 805 CD45, clone HI30 Becton Dickinson Cat# 564914
PE-CF594 FoxP3, clone 259D/C7 Becton Dickinson Cat# 562421
PE-Cy7 Perforin, clone B-D48 Biolegend Cat# 353316
Alexa Fluor 647 IFNg, clone 4S.B3 Biolegend Cat# 502516
PE anti-human TNFa, clone Mab11 Biolegend Cat# 502909
CD4, clone SP35 Cell Marque Cat# 104R-18
Alexa Fluor 555 goat anti-rabbit IgG(H+L) Invitrogen Cat# 4050-32
Chemicals, Peptides, and Recombinant Proteins
Liberase TL, research grade Millipore Sigma Cat# 5401020001
Draq7 Biolegend Cat# 424001
Live/dead fixable Near-IR dead cell stain Invitrogen Cat# L34975
FluoroFix buffer Biolegend Cat# 422101
Recombinant human IL-2 Peprotech Cat# 200-02
IncuCyte Annexin V Red reagent Essen Bioscience Cat# 4641
IncuCyte Cytotox Red reagent Essen Bioscience Cat# 4632
RNAscope probe, Homo sapiens, GZMB Advanced Cell Diagnostics Cat# 445971-C2
(channel 2)
RNAscope probe, Homo sapiens, GZMK Advanced Cell Diagnostics Cat# 475901-C1
(channel 1)
L15 media, 15 mM HEPES, 600 mg% UCSF Cell Culture Facility N/A
glucose
Fetal bovine serum Omega Scientific Cat# FB-01
RPMI-1640 UCSF Cell Culture Facility N/A
ImmunoCult XF complete medium STEMCELL Technologies Cat# 10981
(Medium + 10% FCS + 1% penicillin /
streptomycin)
(Continued on next page)

e1 Cell 181, 1612–1625.e1–e6, June 25, 2020

 ll
Article

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Critical Commercial Assays
GentleMACS Miltenyi Biotec Cat# 130-093-235
FoxP3/transcription factor staining eBioscience Cat# 00-5523-00
buffer set
Cell stimulation cocktail eBioscience Cat# 00-4975
Chromium Single Cell 30 Library, Gel Bead & 10X Genomics Cat# 120233 (discontinued)
Multiplex Kit
Chromium Single Cell 30 Chip Kit 10X Genomics Cat# 120232 (discontinued)
Dynabeads Human T-Activator CD3/ GIBCO Cat# 11162D
CD28/CD137
Opal 7-color manual IHC kit Perkin Elmer Cat# NEL811001KT
Deposited Data
Processed data This study NCBI GEO: GSE149652
Healthy human donor TCR data for CD4+ 10X Genomics https://support.10xgenomics.com/
peripheral blood mononuclear cells single-cell-vdj/datasets/2.2.0/
vdj_v1_hs_cd4_t
Healthy human donor TCR data for CD8+ 10X Genomics https://support.10xgenomics.com/
peripheral blood mononuclear cells single-cell-vdj/datasets/2.2.0/
vdj_v1_hs_cd8_t
Human reference genome, build hg19 10X Genomics http://software.10xgenomics.com/
Oligonucleotides
TCR sequencing primers Table S3 N/A
Software and Algorithms
Cell Ranger v1.1 10X Genomics http://software.10xgenomics.com/
Scanpy v1.4.3 Wolf et al., 2018 https://scanpy.readthedocs.io/en/stable/
index.html
miXCR v2.1.12 Bolotin et al., 2015 https://mixcr.readthedocs.io/en/latest/
Monocle v2.10.1 Qiu et al., 2017 Bioconductor
SingleR v1.1.9 Aran et al., 2019 Bioconductor
FlowJo TreeStar N/A

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Lawrence
Fong (lawrence.fong@ucsf.edu).

Materials Availability
Primer sequences for TCR sequencing are enumerated in Table S3. All unique reagents generated in this study are available from the
Lead Contact upon request.

Data and Code Availability

Processed single-cell RNA sequencing and TCR sequencing data that support this study have been deposited in the NCBI GEO data-
base under accession GSE149652. Raw sequencing data will be deposited in dbGaP. All software algorithms used for analysis are
available for download from public repositories which are listed in the Key Resources Table.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Tissues were obtained from patients with localized bladder transitional cell carcinoma (TCC) who either received 1-2 doses of neo-
adjuvant atezolizumab as part of an ongoing clinical trial (UCSF IRB# 14-15423, patients were accrued sequentially to receive
increasing numbers of atezolizumab doses), or standard of care treatments recommended by their treating physician including

Cell 181, 1612–1625.e1–e6, June 25, 2020 e2

 ll
Article

chemotherapy (gemcitabine/carboplatin) or no systemic therapy prior to planned cystectomy (these patients were consented for tis-
sue collection under a separate protocol, UCSF IRB# 10-04057). All studies with patients and patient samples were conducted with
appropriate institutional IRB approval and oversight. Patient demographics, including age, gender, disease state (all were localized
muscle-invasive bladder cancer), neoadjuvant treatment, and presence of tumor and pathologic staging at the time of surgery are
provided in Table S1. No formal sample size calculations were conducted for this particular collection.

METHOD DETAILS

Tissue processing
Tissues were obtained from patients with localized bladder transitional cell carcinoma (TCC) who either received neoadjuvant ate-
zolizumab, standard of care chemotherapy (gemcitabine/carboplatin), or no systemic therapy as per standard of care prior to
planned cystectomy. Cystectomy surgical specimens were obtained fresh from the operating field, and dissected in surgical pathol-
ogy where grossly apparent tumor or adjacent bladder not grossly affected by tumor (‘‘non-malignant’’) were isolated, minced, and
transported at room temperature immersed in L15 media with 15 mM HEPES and 600 mg% glucose. Once received, these were
digested using Liberase TL as well as mechanical dissocation with heat (gentleMACS) using standard protocols. Single cell suspen-
sions were obtained and counted for viability before staining for FACS. Healthy donor blood was separately collected, processed by
gradient centrifugation to peripheral blood mononuclear cells (PBMCs), and cryopreserved to be thawed later for control
experiments.

Flow cytometry/FACS
Freshly dissociated TILs and previously frozen healthy donor PBMCs were used for sorting. Samples were stained with designated
panels for 30 minutes at 4 C and washed twice with FACS buffer (PBS, 2% FBS, 1mM EDTA). Cells were incubated with Draq7 (Bio-
legend, Cat# 424001) for 5 mins at room temperature to stain dead cells. Samples were sorted on a FACSAria Fusion (Becton Dick-
inson) using FACSDiva software with single channel compensation controls acquired on the same day.
For RNA sequencing flow validation, previously frozen TILs were thawed into in complete media (RPMI, 10% heat inactivated FBS,
1% non-essential amino acids solution, 10 uM HEPES, 1mM sodium pyruvate, 2 mM L-glutamine, 100 U/ml penicillin-streptomycin)
and washed once with PBS. Live/dead fixable Near-IR dead cell stain (Invitrogen, Cat# L34975) was incubated with cells for 30 mi-
nutes at room temperature and washed once with FACS buffer. Samples were stained with designated panels for 30 minutes at 4 C
and washed twice with FACS buffer. Cells requiring intracellular staining were fixed and permeabilized with eBioscience FoxP3/ Tran-
scription factor staining buffer set (Cat# 00-5523-00) according to the manufacturer’s protocol. Intracellular staining with antibodies
was carried out for 30 minutes at room temperature and washed twice with FACS wash. Cells were fixed with FluoroFix buffer (Bio-
legend, Cat# 422101) and washed once with FACS buffer. Cells were acquired the next day on a FACSymphony (Becton Dickinson)
using FACSDiva with single channel compensation controls acquired on the same day. Data was analyzed offline using FlowJo anal-
ysis software (FlowJo, LLC).
For cytokine expression, cells were resuspended in complete media and divided into two T25 flasks. One flask was activated with
cell stimulation cocktail (eBioscience, cat# 00-4975 containing phorbol 12-myristate 13-acetate, ionomycin, brefeldin A and monen-
sin at a final concentration of 81 nM, 1.34 nM, 10.6 mM and 2 mM respectively) and both flasks were incubated upright for 3 h in a CO2
incubator at 37 C. Cells were collected and washed once with PBS prior to Live/Dead Fixable Near-IR dead cell staining and surface
and intracellular flow staining as described above.
Antibodies used for sorting were Brilliant Violet 605 CD25 (Biolegend, clone BC96, Cat# 302632), Brilliant Violet 786 CD127 (Bio-
legend, clone A019D5, Cat# 351330), Brilliant Violet 421 CD4 (Biolegend, clone OKT4, Cat# 317434), Brilliant Violet 650 CD3 (Bio-
legend, clone UCHT1, Cat# 300468), Brilliant Ultraviolet 395 CD45 (Becton Dickinson, clone H130, Cat# 563792), and Alexa Fluor
647 CD8 (Biolegend, clone SK1, Cat# 344726). Antibodies used for RNA sequencing flow validation were FITC GZMK (Biolegend,
clone GM26E7, Cat# 370508), PerCP-Cy5.5 HLA-DR (Biolegend, clone L243, Cat# 307630), APC-R700 CCR7 (Becton Dickinson,
clone 3D12, Cat# 565867), Brilliant Violet 480 CD3 (Becton Dickinson, clone UCHT1, Cat# 566105), Brilliant Violet 510 GZMB (Becton
Dickinson, clone GB11, Cat# 563388), Brilliant Violet 605 Ki67 (Biolegend, clone Ki-67, Cat# 350522), Brilliant Violet 650 CD45RA
(Biolegend, clone HI100, Cat# 304136), Brilliant Violet 786 CD25 (Biolegend, clone BC96, Cat# 302638), Brilliant violet 711 TNFSRF18
(Biolegend, clone 108-17, Cat# 371212), Brilliant ultraviolet 395 CD4 (Becton Dickinson, clone RPA-T4, Cat# 564724), Brilliant ultra-
violet 496 CD8 (Becton Dickinson, clone RPA-T8, Cat# 564808), Brilliant ultraviolet 805 CD45 (Becton Dickinson, clone HI30, Cat#
564914), PE-CF594 FoxP3 (Becton Dickinson, clone 259D/C7), PE-Cy7 Perforin (Biolegend, clone B-D48, Cat# 353316).
Antibodies used for cytokine staining in addition to those used above were Alexa Fluor 647 IFNg (Biolegend, clone 4S.B3, Cat#
502516) and PE anti-human TNFa (Biolegend, clone Mab11, Cat# 502909).

Single cell RNA sequencing

Droplet-based single-cell RNA sequencing (dscRNA-seq) was performed using the 10X Genomics Chromium Single Cell 30 platform,
version 1, according to manufacturer’s instructions. CD3+CD4+ and CD3+CD8+ T cells were sorted from digested tumor and

e3 Cell 181, 1612–1625.e1–e6, June 25, 2020

 ll
Article

non-malignant tissues, or Ficoll-purified and previously cryopreserved healthy control PBMCs, into 500 ul of PSA/0.04% BSA for
loading onto 10X. Following library preparation, sequencing was performed on an Illumina HiSeq 2500 (Rapid Run mode). Paired
samples from the same experiment and patient were processed in parallel during library preparation, and sequenced on the
same flowcell to minimize batch effects.

TCR sequencing
In brief, approximately 10% of the barcoded cDNA from the 10X workflow was utilized for TCR analysis. Primers used for TCR
sequencing are listed in Table S3. cDNA were first amplified with 6-12 amplification cycles using a template switching oligonucleotide
(TSO) and P7 primers. A pool of forward Va and Vb primers containing the TruSeq Read 1 primer sequence were then used in
conjunction with a reverse P7 primer to amplify CDR3 sequences from the TCR alpha and beta loci. An additional amplification
step using forward primers containing the Illumina P5, i5 and Truseq Read 1 sequences was used with reverse P7 primer to create
final TCR libraries for sequencing. Deep sequencing was done on an Illumina NovaSeq S1 with separate lanes for the TCR alpha and
TCR beta sequencing. Read 1 contained 280 bp of the TCR alpha or beta CDR3 sequence, and the i7 read contained the 14 bp 10X
barcode.

Expression analysis
After 10X sequencing data was processed through the Cell Ranger pipeline (version 1.1, hg19 genome assembly) with default set-
tings, filtered gene-barcode matrices for single tumors were analyzed using the scanpy toolkit (Wolf et al., 2018). Genes that were
detected in less than three cells were filtered out, and cells were filtered out with greater than ten percent of mitochondrial genes
and with fewer than 100 or greater than 1200 detected genes. Cells that were annotated as red blood cells (HBB) or macrophage
(CD14, CD68, CD163) were also excluded from downstream analyses. The gene expression values were log2 plus one transformed
and normalized to 10,000 counts per cell. The resulting matrix was batch corrected by regressing out total UMI counts and percent
mitochondrial genes using the built-in scanpy function followed by using the scanpy implementation of ComBat (Johnson et al., 2007)
with each well acting as a batch (13 wells total). The adjusted matrix was scaled to a mean of zero and variance of 1. Highly variable
genes were selected using the embedded scanpy function followed by principal component analysis (PCA), leiden clustering and
UMAP plotting with default settings with the exception of using a resolution of 1.5 for CD4+ T cells and 1.0 for CD8+ T cells for the
leiden clustering. This yielded 19 clusters which were collapsed to 11 cell types based on manual gene annotations (for CD4+ cells),
and 11 clusters (for CD8+ cells). We performed differential expression to identify marker genes that were upregulated in each indi-
vidual cluster relative to the combination of all other single cells (regardless of tumor or non-malignant tissue origin), or genes that
were upregulated in tumor versus non-malignant compartments. We compared the gene lists to known literature to label the clusters,
as well as using SingleR (Aran et al., 2019) to map the expression signature for each cluster to the best correlated candidate immune
reference signature, using the Monaco bulk RNA-seq reference of sorted human immune cell populations described within (Monaco
et al., 2019). Significant differences between the cell type abundances for the normal and tumor tissue samples were assessed using
an exact permutation test on the abundances.
Correlation analysis between gene expression from distinct clusters was performed by restricting to genes expressed across all
clusters being tested, and then correlating the scaled expression of the multidimensional vector of shared genes between pairs of
clusters and computing the Pearson correlation coefficient.

TCR analysis
TRA and TRB CDR3 nucleotide reads were demultiplexed by matching reads to 10X barcodes from cells with existing expression
data that passed filtering in the Cell Ranger pipeline, excluding cell barcodes that overlapped between multiple samples. Following
demultiplexing of the TRA and TRB CDR3s, reads were aligned against known TRA/TRB CDR3 sequences then assembled into clo-
notype families using miXCR (Bolotin et al., 2015) with similar methodologies to a previous study (Zemmour et al., 2018). For any given
10X barcode, the most abundant TRA or TRB clonotype was accepted for further analysis; if 2 TRA or TRB clonotypes were equally
abundant for a given 10X barcode, the clonotype with the highest sequence alignment score was used for further analysis. Detailed
sequencing statistics and saturation analysis are provided in Table S3. Only cells with paired TRA and TRB were used for further
downstream analysis. Analysis utilizing TCR data only (number of unique cells sharing a specific TRA/TRB clonotype sequence,
Gini coefficient) utilized cells both with and without a specific functional population that had been assigned by clustering. Analysis
involving both TCR clonotype and function was restricted to cells with both a mapped TRA/TRB and a functional population from
clustering. Statistical comparisons of Gini coefficients across compartments was performed using Wilcoxon signed-rank test with
Benjamini-Hochberg correction for multiple testing; statistical testing of differences in Gini coefficients between tumor and non-ma-
lignant compartments across all phenotypic clusters was performed using exact permutation testing.

Tumor infiltrating lymphocyte (TIL) isolation and culturing

Single-cell suspensions from processed and digested bladder tumors were viably frozen at 80 C and stored prior to culture setup.
To sort the tumor-infiltrating lymphocytes, frozen cancer cell aliquots were thawed, washed once with PBS, and counted by Vicell.

Cell 181, 1612–1625.e1–e6, June 25, 2020 e4

 ll
Article

Cells were subsequent stained and sorted by FACS. CD4 TIL (Draq7-CD45+CD3+CD4+ that were not CD25+CD127lo) and CD8 TIL
(Draq7-CD45+CD3+CD8+) were sorted into ImmunoCult XF complete medium (Medium + 10% FCS + 1% penicillin/streptomycin;
STEMCELL Technologies #10981). T cells were pooled together for culturing. After centrifugation, T cells were suspended in Immu-
noCult XF complete medium, and Dynabeads Human T-Activator CD3/CD28/CD137 (GIBCO #11162D) were added to the culture per
manufacturer’s protocol. T cells were cultured in 96 well U-bottom plates, and briefly centrifuged to ensure cell contact with Dyna-
beads. T cell expansion was managed in two phases. For the first week of T cell expansion, TILs were maintained with ImmunoCult XF
complete medium + 200 IU/ml of human recombinant IL-2 (Peprotech #200-02). From the second week onward, IL-2 concentration
was gradually increased from 200 IU/ml to 2000 IU/ml based on cell growth kinetics (which varied by patient sample). T cells were
harvested between 5-8 weeks for functional killing assays.

Cytotoxic T lymphocyte (CTL) killing assay

After expansion, TILs were again sorted for either CD4 or CD8 as distinct effector populations. Primary cancer cells from frozen
aliquots were freshly thawed and sorted on CD45-Draq7- as target cells. To achieve various effector-to-target (E:T) ratios, 3000
cancer cell targets were suspended in ImmunoCult XF complete medium and seeded into each well. Different ratios of TILs
were serially diluted and added to the corresponding wells. Each well contained 200 mL of medium supplemented with 1 ml/
well of IncuCyte Annexin V Red reagent (Essen Bioscience #4641). For MHCI and MHCII blockade, 10 mg of blocking antibody
(or isotype control matched to the anti-MHCII antibody) was added into wells containing cancer cells and cultured at 37 C for
1 hour prior to co-culture with TILs. Cell culture was monitored by the IncuCyte Zoom system (Essen Bioscience) at 15-30 minute
intervals for up to 36 h when needed. Experiments with Annexin V were carried out with samples from 3 independent patients.
Additional experiments were performed using 0.25 mL of IncuCyte Cytotox Red reagent (Essen Bioscience #4632) instead of An-
nexin V; 2 independent experiments were performed with Cytotox Red using distinct aliquots from the same patient. Analysis was
performed using the IncuCyte Zoom software, using images background subtracted with a tophat filter. For Annexin V experi-
ments, background death was subtracted, with all traces displayed as relative change in cell death from time point 0. For Cytotox
Red experiments, tumor cells were larger than TIL based on inspection of wells with tumor cells alone or free TILs in wells con-
taining TILs; based on this, the number of dying tumor cells per mm2 was determined using a minimum area threshold of 75 mm2.
Out of focus frames were discarded, as were any wells where the first time frame was out of focus precluding accurate
normalization.

Pseudotime analysis
Pseudotime analysis, including branched expression analysis modeling (BEAM) to identify all genes with branch-dependent differ-
ential expression followed by unbiased clustering of genes based on patterns of co-expression in specific branches, was performed
using Monocle v2.10.1 as described (Qiu et al., 2017), for the combination of proliferating (CD4PROLIF, regulatory (CD4IL2RAHI,
CD4IL2RALO) and cytotoxic (CD4GZMB, CD4GZMK) states from scRNA-seq clustering.

Gene signature analysis

Genes were selected based on their specific upregulation in proliferating or non-proliferating cytotoxic CD4+ T cell branches, but not
in regulatory T cell branches, from the pseudotime analysis. Genes that overlapped with differentially expressed genes in any of the
CD8+ T cell states from our scRNA-seq analysis (Table S2) were removed. The resulting gene list was used to construct a composite
signature consisting of genes that distinguish either proliferating or non-proliferating cytotoxic CD4+ T cells; a signature score was
computed as previously published for the IMvigor210 dataset (Mariathasan et al., 2018), by z-score transforming the expression of
each gene in the signature, and then using the first component (PC1) of a principal component analysis as the gene signature score.
Immune subtypes for the IMvigor210 samples for this analysis were previously assigned based on CD8 immunohistochemistry stain-
ing (Mariathasan et al., 2018).

RNAscope/tissue immunofluorescence staining

RNAscope (Advanced Cell Diagnostics) in situ hybridization and immunofluorescence staining were performed on 4 mm FFPE sec-
tions from cystectomy specimens with existing scRNA-seq and TCR-seq data. Tissues were pre-treated with target retrieval re-
agents and protease to improve target recovery based on guidelines provided in the RNAscope Multiplex Fluorescent Reagent
Kit v2 Assay protocol. Probes for human GZMB and GZMK mRNA (ACD) were incubated at 1:700 dilution for 2 hr 40 min. The probe
was then hybridized with Opal 7-Color Manual IHC Kit (PerkinElmer), with detection of GZMB and GZMK using Opal 620 and Opal
540 respectively. Samples were then immunofluorescence stained for human CD4 (Cell Marque) which was detected using Alexa
Fluor 555-conjugated goat anti-rabbit IgG secondary antibody (Invitrogen). Tissues were counterstained with DAPI. Slides were
imaged using a TCS SP8 X white light laser inverted confocal microscope (Leica Microsystems). No staining was seen with negative
control probes (for RNAscope) or with secondary antibody alone (for immunofluorescence) for tumor tissue from cystectomy spec-
imens (shown in Figure 4F) or healthy tonsil tissue (data not shown).

e5 Cell 181, 1612–1625.e1–e6, June 25, 2020

 ll
Article

QUANTIFICATION AND STATISTICAL ANALYSIS

Specific statistical tests and metrics (median, mean, standard error) used for comparisons, along with sample sizes, are described in
the Results and figure legends. The chemotherapy sample was included in unbiased clustering, testing for conserved marker genes
and tumor versus non-malignant testing, but was excluded from analyses of treatment effect (anti-PD-L1 versus untreated).

ADDITIONAL RESOURCES

The clinical trial of neoadjuvant atezolizumab prior to planned cystectomy for localized bladder cancer is registered under
clinicaltrials.gov (NCT02451423).

Cell 181, 1612–1625.e1–e6, June 25, 2020 e6

 ll
Article

Supplemental Figures

(legend on next page)

 ll
Article

Figure S1. Flow Cytometry and Immunofluorescence Validation of T Cell Phenotypes in Bladder Tumors, Related to Figures 1, 2, 3, 4, and 5
(A) Schematic of processing for paired tumor and adjacent non-malignant tissue from either anti-PD-L1-treated, or standard-of-care (untreated/chemotherapy-
treated) cystectomy patients. FACS-sorted CD4+ or CD8+ T cells were subjected to droplet-based single-cell RNA sequencing (dscRNA-seq) with paired T cell
receptor (TCR) sequencing as described in the text. (B) Parallel flow cytometry data from the same single-cell digest used for dscRNA-seq from 4 anti-PD-L1-
treated tumors, showing the percentage of CD4+ or CD8+ T cells from total CD3+ cells. (C) Gating strategy for flow cytometric analysis of populations in CD4+ and
CD8+ T cells from RNA-seq. CD4+ and CD8+ populations were gated out of CD3+ CD45+ single live cells. CD4+ cells were further gated as FoxP3- and FoxP3+.
Treg cells are gated as FOXP3+ CD25+ cells. FOXP3- CD4+ and CD8+ cells were gated into central memory (CM, CCR7+ CD45RA-), and CCR7- cells (a com-
bination of effector memory CCR7- CD45RA- and effector CCR7- CD45RA+). Boolean gating of CCR7- cells was used to obtain GZMK+, GZMB+ and Ki67+
populations for further marker analysis. Plots are shown here to demonstrate the presence of these populations. (D) Representative gates shown for each marker
for CD4+ and CD8+ T cells were used for Boolean gating for the populations described above. (E) Flow cytometry staining of GZMB, GZMK, or perforin versus CD3
in CCR7- CD8+ T cells. Gates used for Boolean analysis are shown. (F) Flow cytometry staining of GZMB or GZMK co-expression with perforin in CCR7- CD8+
T cells. (G) Percentage of cells expressing GZMB, GZMK, or perforin from CCR7- CD8+ T cells by flow cytometry (left), and the percentage of cells co-expressing
perforin within GZMB+ or GZMK+ CCR7- CD8+ T cells (right), are shown (N = 7 tumors, mean + SEM). (H) Percentages of cells expressing IFNg, TNFa, or both from
GZMB+ or GZMK+ CCR7- CD8+ T cells with and without stimulation (N = 11 tumors, mean + SEM). (I) Multiplex immunofluorescent staining of DAPI (blue), CD4
(red), GZMK (green), GZMB (white) and overlay without DAPI are shown from a cystectomy tumor region from an additional patient with parallel scRNA-seq and
TCR-seq data (anti-PD-L1 D). CD4+ cells that co-express GZMK (arrows) or GZMB (arrowhead) are indicated. Scale bar, 10 mm. (J) Percentage of cells co-
expressing Ki67 and either GZMB or GZMK from CCR7- CD4+ FOXP3- T cells (left), or Ki67 and CD25 from CD4+ FOXP3+ T cells (right), by flow cytometry are
shown, with dots for values from individual tumors (N = 7 tumors, mean ± SEM). (K) Flow cytometry staining showing co-expression of GZMB and Ki67, or GZMK
and Ki67, from CCR7- CD8+ T cells.
 ll
Article

(legend on next page)

 ll
Article

Figure S2. Clustering, Differential Expression, Annotation, and Correlation Analysis of T Cell Transcriptional Phenotypes, Related to Figures
1, 2, 3, and 4
(A-B) UMAP plots showing cluster representation for CD8+ (A) and CD4+ (B) TIL from individual patients. (C) Volcano plots showing adjusted P values versus
log2(FC) for differential testing of genes between tumor and non-malignant compartments for regulatory T cell populations (top, CD4IL2RAHI, CD4IL2RALO) and
cytotoxic CD4+ populations (bottom, CD4GZMB, CD4GZMK). Genes whose expression is significantly different between compartments with Padj < 0.05 and |
log2(FC) > 1.4| are shown in red. (D-E) Unbiased clustering of CD4+ T cells from tumor and adjacent non-malignant tissue from a single patient (anti-PD-L1 C). (D)
UMAP plot showing individual cells coded by cluster or by tissue of origin. (E) Violin plot showing top 5 differentially expressed marker genes for each unbiased
cluster. (F) Annotations of single CD4+ T cells from tumor and adjacent non-malignant tissue using SingleR. (G) Correlation matrix of all CD4+ and CD8+ pop-
ulations from tissue (combined tumor and non-malignant tissues) based on expression of shared genes. Pearson correlation coefficient is shown. Populations
were arranged based on hierarchical clustering using Euclidean distance metric.
 ll
Article

Figure S3. T Cell Receptor Repertoire Analysis of CD4+ and CD8+ Bladder Tumor- and Non-malignant Tissue-Infiltrating T Cells, Related to
Figures 3 and 4
(A) The percentage of unique paired TRA and TRB CDR3 nucleotide sequences that are expressed by one cell (blue), shared by two cells (green), or shared by three or
more cells (red) is indicated for CD4+ T cells from individual tumor (darker shades) and non-malignant tissues (lighter shades) from anti-PD-L1-treated (‘‘PD-L1’’),
untreated, and chemotherapy-treated (‘‘chemo’’) patients. Triplicate control samples from a single healthy donor’s CD4+ T cells sorted from peripheral blood and
processed for scRNA-seq and TCR in identical fashion in separate sequencing runs is shown (‘‘healthy 1-3’’), as well as reference publicly available data from peripheral
blood CD4+ from a healthy donor. (B) Lorenz curves showing the cumulative frequency distributions for unique CD4+ T cells and unique CD4+ T cell clonotypes for
tumor, non-malignant tissues, and healthy donor blood. Mean ± SD is shown. (C) Gini coefficients for CD4+ T cell clonotypes from tumor, non-malignant tissues, and
healthy donor blood, calculated from the Lorenz curves in (D); p = 0.009 by Wilcoxon with Benjamini-Hochberg correction for tumor versus non-malignant tissues. For

(legend continued on next page)

 ll
Article

(D) and (E): N = 7 tumor samples; 6 non-malignant samples, 4 healthy donor samples (3 triplicates from one healthy donor, 1 dataset from 10X Genomics). (D-F) Paired
TRA/TRB clonotype sharing between cells, Lorenz curves, and Gini coefficients for CD8+ clonotype data as in (A-C). (G) Gini coefficients for tissue-infiltrating CD4+ in
individual populations, separated by treatment type. (H-I) Gini coefficients for CD8+ T cells in individual populations, separated by tumor versus non-malignant tissue (H)
and treatment type (I). All box and whisker plots are formatted as in Figure 3C.
 ll
Article

Figure S4. Autologous MHC-Dependent Killing of Bladder Tumors by CD4+ and CD8+ TIL, Related to Figure 5
Analysis of the increase in the number of dead cells over time from the same killing assay for CD4eff TIL (ie cultures with Tregs sorted out during expansion) at 30:1
effector:target ratio (A), CD4eff TIL at 30:1 effector:target ratio with a pan-anti-MHCII antibody (B), CD8+ TIL at 30:1 effector:target ratio (C), or CD8+ TIL at 30:1
effector:target ratio with a pan-anti-MHCI antibody (D), are shown. Control traces from separate wells with tumor only are included. All traces were normalized to
the number of dead cells per mm2 at time point 0. Experiments were done using Cytotox Red. The observation of autologous tumor killing by CD4+ and CD8+ TIL
above the background level of spontaneous death is representative of 2 independent experiments involving distinct aliquots from the same patient.
 Resource

Single-Cell Mapping of Human Brain Cancer Reveals

Tumor-Specific Instruction of Tissue-Invading
Leukocytes
Graphical Abstract Authors
Ekaterina Friebel, Konstantina Kapolou,
Susanne Unger, ..., Sonia Tugues,
Marian Christoph Neidert,
Burkhard Becher

Correspondence
becher@immunology.uzh.ch

In Brief
High-parametric single-cell mapping of
the tumor microenvironment of patients
with primary brain tumors or brain
metastases reveals that the immune
response to cancer in the brain is shaped
by cancer type, with metastases favoring
T cell and monocyte-derived
macrophage invasion and gliomas
characterized by activated microglia.

Highlights
d Leukocyte invasion is higher in brain metastasis than in CNS-
endogenous cancers

d The tumor type shapes the differentiation of monocyte-

derived macrophages

d Brain metastases harbor a high frequency of regulatory

T cells

d Both activation and exhaustion are prevalent in lymphocytes

of the metastatic TME

Friebel et al., 2020, Cell 181, 1626–1642

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.04.055 ll
 ll

Resource
Single-Cell Mapping of Human Brain Cancer
Reveals Tumor-Specific Instruction
of Tissue-Invading Leukocytes
Ekaterina Friebel,1,5 Konstantina Kapolou,2,5 Susanne Unger,1 Nicolás Gonzalo Núñez,1 Sebastian Utz,1
Elisabeth Jane Rushing,4 Luca Regli,3 Michael Weller,2 Melanie Greter,1 Sonia Tugues,1 Marian Christoph Neidert,2,3,6
and Burkhard Becher1,6,7,*
1Instituteof Experimental Immunology, University of Zurich, Zurich 8057, Switzerland
2Laboratory of Molecular Neuro-Oncology, Department of Neurology, Clinical Neuroscience Center, University Hospital Zurich and University
of Zurich, Zurich 8091, Switzerland
3Department of Neurosurgery, Clinical Neuroscience Center, University Hospital Zurich and University of Zurich, Zurich 8091, Switzerland
4Department of Neuropathology, University Hospital Zurich and University of Zurich, Zurich 8091, Switzerland
5These authors contributed equally
6These authors contributed equally
7Lead Contact

*Correspondence: becher@immunology.uzh.ch
https://doi.org/10.1016/j.cell.2020.04.055

SUMMARY

Brain malignancies can either originate from within the CNS (gliomas) or invade from other locations in the
body (metastases). A highly immunosuppressive tumor microenvironment (TME) influences brain tumor
outgrowth. Whether the TME is predominantly shaped by the CNS micromilieu or by the malignancy itself
is unknown, as is the diversity, origin, and function of CNS tumor-associated macrophages (TAMs). Here,
we have mapped the leukocyte landscape of brain tumors using high-dimensional single-cell profiling (Cy-
TOF). The heterogeneous composition of tissue-resident and invading immune cells within the TME alone
permitted a clear distinction between gliomas and brain metastases (BrM). The glioma TME presented pre-
dominantly with tissue-resident, reactive microglia, whereas tissue-invading leukocytes accumulated in BrM.
Tissue-invading TAMs showed a distinctive signature trajectory, revealing tumor-driven instruction along
with contrasting lymphocyte activation and exhaustion. Defining the specific immunological signature of
brain tumors can facilitate the rational design of targeted immunotherapy strategies.

INTRODUCTION
Malignant brain tumors can be either primary tumors that arise in
the brain (e.g., gliomas) or secondary tumors of extracranial
origin (metastases), most frequently invading from non-small-
cell lung carcinoma (NSCLC), breast cancer, and melanoma
(Quail and Joyce, 2013). Both primary and metastatic brain ma-
lignancies have a very poor prognosis, mainly due to limitations
of standard therapies (e.g., surgery, radio/chemotherapy) (Al-
dape et al., 2019). In this scenario, immunotherapy poses a
promising treatment option; however, programmed death (PD)-
1 blockade has not shown a survival benefit in recurrent glioblas-
toma (NCT 02017717). Only a selective group of patients with te-
mozolomide-induced hypermutations in gliomas have been
shown to benefit from immunotherapy (Daniel et al., 2019; Jo-
hanns et al., 2016; Bouffet et al., 2016). In contrast, a consider-
able number of patients with brain metastases (BrM) respond
to checkpoint blockade with concordant responses in extracere-
bral disease and brain lesions (Goldberg et al., 2016; Kluger
et al., 2019; Tawbi et al., 2018). These divergent clinical re-
sponses to immunotherapy may be explained by tumor cell
genomic (intrinsic) properties regulating T cell antigen recogni-
tion and immune responses. Alternatively, tumor-extrinsic fea-
tures such as the tumor microenvironment (TME) may represent
resistance pathways to immune-mediated interventions (Keenan
et al., 2019). Thus, in-depth knowledge of the immune cell
composition within the TME across different types of brain tu-
mors may be critical to predict not only how tumors will progress,
but also their immunotherapeutic outcomes.
Uniquely, the CNS is both a site of immune privilege and a
complex leukocyte landscape (Keren-Shaul et al., 2017; Forres-
ter et al., 2018; Mrdjen et al., 2018; Van Hove et al., 2019). An
abundant cellular component of the brain TME are tumor-asso-
ciated macrophages (TAMs), which possess both tumor-pro-
moting and immunosuppressive capacities (Quail and Joyce,
2013; Takenaka et al., 2019). This dichotomy is further compli-
cated by the heterogeneity of TAMs. TAM populations can be
subdivided ontogenetically into at least two major populations:

1626 Cell 181, 1626–1642, June 25, 2020 ª 2020 Elsevier Inc.
 ll
Resource

Figure 1. Mass Cytometry Analysis Reveals Unique Changes in the Leukocyte Composition of Brain Tumors
For a Figure360 author presentation of Figure 1, see https://doi.org/10.1016/j.cell.2020.04.055.
(A) Experimental approach.
(B) MDS plot showing the mean antigen expression across leukocytes. The clinical groups are indicated by varying point shapes and colors.
(C) A representative UMAP map showing the FlowSOM-guided meta-clustering of CD45+ cells.

(legend continued on next page)

Cell 181, 1626–1642, June 25, 2020 1627

 ll
(1) tissue-resident microglia and macrophages of embryonic
origin, and (2) tissue-invading monocyte-derived macrophages
(hereafter termed MDM) (Croxford et al., 2015; Kiss et al.,
2018). Tissue-resident populations comprise microglia and
border-associated macrophages (BAMs), representing the ma-
jority of phagocytes in the healthy steady-state brain. However,
tumor progression induces the recruitment of blood-borne
monocytes, which differentiate into MDMs (Bowman et al.,
2016; Ginhoux et al., 2010; Goldmann et al., 2016; Mrdjen
et al., 2018). Concomitant with the expansion of MDMs, CNS-
resident macrophages undergo marked phenotypical and func-
tional changes, which further complicates the identification of
the origin of TAMs (Hambardzumyan et al., 2016; Bowman
et al., 2016; Quail and Joyce, 2017). Importantly, the majority
of studies assessing the contribution of TAMs in brain cancer
have been performed in mouse models of malignant glioma or
restricted cohorts of glioblastoma patients, with only a few at sin-
gle-cell resolution (Darmanis et al., 2017; Quail et al., 2016; Ven-
teicher et al., 2017).
Brain tumors also harbor variable lymphocyte infiltrates (TILs)
(Bien kowski and Preusser, 2015). Both CD8 and CD4 T cell sub-
sets (including regulatory T cells [Tregs]) increase with tumor
grade (Jacobs et al., 2010; Kuppner et al., 1988). Paradoxically,
prognostically more favorable diffuse gliomas with mutations in
the isocitrate dehydrogenases 1 and 2 are associated with
reduced T cell abundance, presumably due to the effect of the
oncometabolite (R)-2-hydroxyglutarate on the TME (Bunse
et al., 2018; Kohanbash et al., 2017). The extent of T cell infiltra-
tion in brain tumors of extracranial origin has also been examined
in some studies. In melanoma BrM, for instance, high densities of
TILs have been observed mainly within the tumor stroma and
surrounding brain (Amit et al., 2013). Nevertheless, none of these
reports have specifically investigated T cell and other lympho-
cyte infiltration in relation to the overall immune landscape of
the TME.
To provide a more detailed insight into the nature of the brain
TME with a focus on the immune cell composition, we performed
single-cell mass cytometry (CyTOF) on ex vivo surgical resec-
tions of brain tumors and non-tumor controls. Our in-depth im-
mune cell phenotyping has revealed fundamental differences in
cellular frequencies and phenotypes within primary and second-
ary brain tumor TMEs. Our findings may explain the divergent re-
sults of immunotherapy in patients with brain tumors and allow
for the data-driven design of novel therapeutic interventions.
RESULTS
Mass Cytometry Analysis Reveals Unique Changes in
the Leukocyte Composition of Brain Tumors
The overall strategy involved harvesting freshly resected tissue
from 38 patients undergoing neurosurgery for the treatment of
(D) Heatmap displaying the median antigen intensity of markers used to generate (C).
(E) The relative frequencies of immune populations of brain tumors and non-tumor control. Only statistically significant p values were displayed (p < 0.05, Mann-
Whitney-Wilcoxon test, Benjamini-Hochberg correction). Error bars define an interval of max/min value ± SD.
(F) Circos plots showing the multiple correlation matrix between the leukocyte frequencies. Results are plotted in order to display statistical significance (p < 0.05).
Here, correlation coefficients (R) higher than 0.6 are represented in red, and those lower than 0.6 in blue.
See also Figures S1 and S2 and Table S1.
1628 Cell 181, 1626–1642, June 25, 2020
Resource
gliomas, BrM (metastatic melanoma, NSCLC, and other tumors)
and epilepsy (serving as a quasi-steady-state control) (Fig-
ure 1A). Gliomas were further characterized according to IDH1
R132H mutation (IDH1mut and IDH1wt) and methylation status
of the O6-methylguanine DNA methyltransferase (MGMT) pro-
moter. Some patients with BrM had a long disease history and
underwent either chemotherapy or treatment with immune
checkpoint inhibitors. Clinical information and treatment regi-
mens prior to surgery are summarized in Table S1.
To map the complexity of the leukocyte compartment of the
TME, we designed two CyTOF panels together, measuring 74
parameters at the single-cell level. The first panel was myeloid-
focused (Table S2). Here, we combined antibodies to capture
the entire spectrum of phagocytes in the brain TME, together
with lineage-identifying markers to map the major leukocyte
populations and their relative cellular frequencies. The second
panel was lymphoid-focused (Table S3) and was designed to
deeply interrogate the lymphoid compartment. We started the
analysis of leukocytes obtained from the single-cell suspension
of brain tissues with the myeloid-focused panel and estimated
similarities between glioma, BrM, and non-tumor control sam-
ples. To do so, we applied multidimensional scaling (MDS) to
the total leukocytes (Figure 1B), where the distances, or relative
similarities, between samples were calculated using the mean
value of antigen expressions (Buja et al., 2008). The first dimen-
sion featured an unambiguous separation between the glioma
and BrM TME, although the single epilepsy sample was mapped
closer to the glioma samples. The second dimension showed
relative variety between samples within the tumor group (Fig-
ure 1B). In order to better understand the basis of the group sep-
aration in the MDS analysis, we visualized the mean antigen
expressions across leukocytes using a heatmap with unsuper-
vised hierarchical clustering (Figure S1) (Nowicka et al., 2017).
The most informative differentially expressed markers were
those commonly expressed by most phagocytes (CD11c,
CD64, HLA-DR, and CX3CR1), as well as common T cell-
restricted antigens (CD3, PD-1, and CD38).
To visualize every single immune population isolated from the
different brain tumor samples, we created a two-dimensional
graph using the dimensionality reduction algorithm uniform
manifold approximation and projection (UMAP) (Figures 1C
and S2A) (Mcinnes et al., 2018). To compute UMAP, we specified
the lineage markers, listed in Figure 1D to be considered for the
estimation of cell similarity. By reducing the high-dimensional
data into two dimensions, we could present the quantification
of all measured marker expressions on all cellular subtypes,
simultaneously (Figure S2A). Next, we categorized the various
embedded cell clusters using self-organizing maps (FlowSOM)
(Van Gassen et al., 2015; Hartmann et al., 2016). This strategy al-
lowed us to create a map of diverse immune cells including,
CNS-resident and invading TAMs/monocytes (CD64+, CD11c+,
 ll
Resource

Figure 2. The Brain TME Harbors a Heterogeneous Mononuclear Phagocyte Population

(A) CD49d expression overlaid onto the Scaffold map.
(B) Heatmap displaying the median antigen intensity of landmark nodes of (A).
(C and D) P2Y12 and CD49d expression of TAMs/monocytes isolated from glioma IDH1wt (C) and epilepsy (D), overlaid onto the reference map (A).

(legend continued on next page)

Cell 181, 1626–1642, June 25, 2020 1629

 ll
and CD11b+), neutrophils (CD66b+ and CD16+), two subsets of
dendritic cells (CD141+ and CADM1+ for cDC1 and CD1c+ for
cDC2), T cells (CD3+), natural killer (NK) cells (CD56+CD16+),
B cells (CD19+ and HLA-DR+), and plasma cells (CD19+ and
CD38high) (Figures 1C and 1D). Notably, TAMs and monocytes
comprised up to 80% (±18%) of leukocytes in the IDH1mut or
IDH1wt gliomas similar to the epilepsy non-tumor control, while
T cells represented only 13% (±10%) (Figures 1E and S2B).
Further stratification of IDH1wt glioblastoma patients according
to the MGMT promoter methylation status showed comparable
frequencies of major immune populations (Figure S2C). Interest-
ingly, we observed the opposite situation in BrM, such as mela-
noma and carcinoma (Figure 1E), with significantly higher relative
frequencies of T cells (up to 50% ± 16%) and lower frequencies
of TAMs (up to 40% ± 18%) in the TME. Additionally, the meta-
static TME in both melanoma and carcinoma showed signifi-
cantly higher frequencies of plasma cells, which were absent in
gliomas (Figure 1E).
To determine whether therapy (including immunotherapy) was
driving these inter-tumor variations, we tracked treated patients
in our analysis. However, we could not find a significant treat-
ment effect across the samples, and hence all samples were
analyzed together and not stratified across therapeutic interven-
tions (Figure 1E). The relative distribution across leukocytes and
their relationship within brain tumors differed (Figure 1F). In gli-
oma, an increase in the relative frequencies of T cells, neutro-
phils, and pDCs correlated negatively with TAM/monocyte
frequencies, whereas T cell frequencies positively correlated
with pDCs and cDCs frequencies. In BrM, on the other hand,
we only observed a negative correlation between T cell and
TAM/monocyte frequencies. Taken together, our results show
that gliomas and BrM shape TMEs with a distinct immune cell
composition: TAMs dominate the TME in gliomas whereas TILs
dominate BrM. A close analysis of the immune compartment of
TME alone allowed for a clear separation into cancers of CNS
origin or extracranial CNS-invading metastasis.
The Brain TME Harbors a Heterogeneous Mononuclear
Phagocyte Population
TAMs within the CNS may originate from the CNS-resident mi-
croglia and/or from blood-derived monocytes that invade the
TME and transform into MDMs. To study the relative contribution
of CNS-resident versus invading TAMs across gliomas and BrM,
we re-applied FlowSOM to CD64+, CD11c+, CD11b+, CD1c ,
and CD66b cells (Figures 1C, 1D) identified in the TME and
non-tumor control. This time, however, we essentially consid-
ered the expression of the macrophage activation markers and
initially clustered cells into 100 nodes, where each node repre-
sents a group of cells. For visualization, we chose a single-cell
analysis by fixed force- and landmark-directed (Scaffold) layout
(Figure 2A) (Spitzer et al., 2015). Here, one cluster node (cellular
population) represents one FlowSOM node and will be mapped
(E) Sall1 and CD49d expression of mouse TAMs/monocytes, overlaid onto the reference map (A).
(F) TAMs/monocytes of glioma and BrM overlaid onto the reference map (A).
(G) Relative frequencies of the three TAM populations among CD64+ cells. Only statistically significant p values were displayed (p < 0.05, Mann-Whitney-Wil-
coxon test, Benjamini-Hochberg correction), whiskers within 1.5x IQR.
See also Figure S3 and Table S1.
1630 Cell 181, 1626–1642, June 25, 2020
Resource
closer to proximal nodes indicative of shared and similar fea-
tures. The size of the node represents the number of cells within
the group. In addition, the Scaffold layout allowed for the intro-
duction of manually gated data, which were used as landmark
populations for the reference map. This permitted a direct com-
parison of single-cell data from different species and/or single-
cell maps acquired using different techniques. This way, CyTOF
and fluorescence-activated cell sorting (FACS) data obtained
from patients and mice could be mapped side-by-side. To build
the reference map, we used equally normalized pooled data from
glioma and BrM samples, to ensure that all TAM subsets were
present in our reference sample. Here, we focused on the
expression of the integrin alpha 4 (ITGA4/CD49d), which was
previously reported to specifically mark CNS-invading macro-
phages (Bowman et al., 2016) (Figure 2A). Last, we created the
reference map for the Scaffold layout, considering three
major cellular clusters: (1) CNS-resident microglia (CD49d ,
Mertk+, CX3CR1+, CD11c+, and CD64+), (2) monocytes
(CD14+, CCR2+, and CD11b+), and (3) MDMs (CD49d+, Mertk+,
CD163+, and CD64+) (Figures 2A and 2B).
Of note, our study revealed additional markers (CD45RA,
CD141, and ICAM), which were differentially expressed by
invading MDMs compared to CNS-resident microglia (Figure 2B).
Cellular identification by computational tools was confirmed us-
ing manual gating of the CyTOF data (Figure S3A). The CyTOF
spectrum of CNS TAMs was complemented by 24-parameter
fluorescence cytometry including the microglia-specific protein
P2Y12 (Butovsky et al., 2014) in epilepsy, glioma, and BrM sam-
ples. We repeated the same strategy using FlowSOM of the
CD11b+CD11c+CD66b lin cells measured by FACS. Over-
laying FACS data onto the reference map from the CyTOF data
showed the expected position of nodes around the fixed land-
mark populations (Figures 2C, 2D). FACS data confirmed both
the phenotype and frequencies of the CyTOF populations and
categorized the exclusive expression of CD49d on CNS-
invading and P2Y12 on CNS-resident macrophages. As ex-
pected, we identified only P2Y12+ cells in epilepsy, as an
approximation to the steady-state CNS, whereas glioma and
BrM contained both populations, CD49d+ invading MDMs and
P2Y12+ CNS-resident microglia.
Even though P2Y12 is one of the markers most commonly
used to identify mouse and human microglia, it has been re-
ported that P2Y12 expression is reduced after microglial activa-
tion (Haynes et al., 2006). Therefore, to solidify the notion that
CD49d expression distinguishes MDMs from microglia even
within the TME, we used a preclinical model of glioma combined
with genetic fate mapping Sall1 expression, a transcriptional
regulator exclusive to microglia within the hematopoietic system
(Figure 2E) (Buttgereit et al., 2016; Mrdjen et al., 2018). This
model showed that Sall1 expression by microglia remains high
even during inflammatory conditions, and this marker is not ex-
pressed by any other CNS-invading cell types or BAMs. Mouse
 ll
Resource

Figure 3. TAM Instruction Is Driven by the Type of Tumor Rather Than the Local Tissue Microenvironment
(A) Two-dimensional dot-plots displaying cellular markers among microglia.
(B) Boxplots quantify the mean antigen intensity of data shown in (A), whiskers within 1.5x IQR.
(C) A force-directed graph displays MDMs/monocytes from glioma and BrM. Individual plots are overlaid with the most differentially expressed markers
among MDMs.

(legend continued on next page)

Cell 181, 1626–1642, June 25, 2020 1631

 ll
CD64+ phagocytes were analyzed using the same reference
Scaffold map as for human CyTOF data, excluding markers
that are differentially expressed in human and mouse (CD11c)
or were not used in both panels (Ly6C and CD14). The map faith-
fully separated mouse microglia (Sall1 YFP+CD49d ) from other
phagocytes (Sall1 YFP CD49d+) and invading phagocytes and
confirmed that CD49d is preferentially expressed among CNS-
invading phagocytes (MDMs) (Figure 2E).
Using a combination of experimental approaches, we could
reliably identify the origin of each TAM population stemming
either from the embryonically derived CNS-resident microglia
or from blood-derived MDMs. Glioma TME predominantly con-
tained TAMs of microglial origin, whereas BrM had a higher
invasion of MDMs, with a particular extension in the TME of car-
cinoma BrM (Figures 2F and 2G). MDM accumulation was also
observed in the IDH1wt glioma TME, albeit to a lesser extent,
whereas the macrophage composition of the TME of IDH1mut gli-
omas—associated with a better prognosis—did not differ signif-
icantly from the epilepsy case (Figures 2F and 2G). Monocytes
were present at similarly low frequencies across all samples (Fig-
ures 2F and 2G). Overall, the data show that the TME of IDH1mut
tumors of neuronal derivation is dominated by TAMs of microglial
origin. The more aggressive gliomas (IDH1wt) showed an
increased invasion by MDM TAMs, independently of the methyl-
ation status of the MGMT promoter (Figure S3B). Along this tra-
jectory, brain tumors of extracranial origin are predominantly
invaded by MDMs, which represent the majority of the TAM
population.
TAM Instruction Is Driven by the Type of Tumor Rather
Than the Local Tissue Microenvironment
TAM plasticity and polarization in the TME across various cancer
entities are subjects of intense investigation (Kiss et al., 2018).
Here, we explored the expression of a large number of markers
used for macrophage phenotyping and polarization (Table S2) on
human CNS-resident microglia with single-cell resolution. Of
those, for instance, CD169, CD206, and CD209 were virtually ab-
sent from CNS-resident tumor-associated microglia (Figure 2B).
We also examined the expression level of markers, which were
previously used to describe the ‘‘reactivity’’ of CNS-resident mi-
croglia in different pathological settings (Figure 3A) (Hopperton
et al., 2018; Keren-Shaul et al., 2017; Mrdjen et al., 2018; Walker
and Lue, 2015). Microglia in IDH1mut glioma had a comparable
expression of HLA-DR to microglia in IDH1wt glioma and BrM
(Figure 3B). However, in both IDH1wt gliomas and BrM, but not
in IDH1mut gliomas, these cells upregulated CD14 and CD64
(Figure 3B). The ‘‘reactive’’ phenotype of tumor-associated mi-
croglia in IDH1wt glioblastomas and BrM was corroborated by
co-staining of brain sections with Iba1 and CD163, revealing
an amoeboid microglial morphology of these cells. In contrast,
(D–F) FlowSOM-guided metaclustering overlaid on (C). (E) Heatmap displaying the median antigen intensity of markers used to generate (D). (F) A force-directed
graph display changes in the brain TME (D and F). The overlaid lines represent the results of the slingshot pseudotime analysis.
(G) Bar plots representing the relative frequencies of monocytes and MDMs in the brain TME. The IDH1mut glioma TME composed of monocytes 74.5% ± 3.3%,
MDM1 13.8% ± 3.3%, MDM2 1.9% ± 0.6%, MDM3 5.7% ± 1.5%, and MDM4 4.1% ± 2.1%; the IDH1wt glioma: monocytes 20.4% ± 21.4%, MDM1 26.9% ±
12.8%, MDM2 17.3% ± 15.5%, MDM3 7.4% ± 10.2%, and MDM4 27.9% ± 20.7%; the melanoma BrM: monocytes 24.8% ± 18.1%, MDM1 22.8% ± 6.9%,
MDM2 17.5% ± 8.2%, MDM3 33.7% ± 27.9%, and MDM4 1.2% ± 1.3%; carcinoma BrM: monocytes 14.2% ± 5.8%, MDM1 26.1% ± 5.9%, MDM2 34.6% ±
7.8%, MDM3 22.7% ± 12.3%, and MDM4 2.4% ± 0.5%.
1632 Cell 181, 1626–1642, June 25, 2020
Resource
microglia in IDH1wt diffuse astrocytoma preserved a more rami-
fied shape (Figure S4A).
Next, we investigated the monocyte-to-macrophage transi-
tion, which is initiated by the invasion of the monocytes from
the blood into the tissue and is primarily dictated by cues in
the local tissue microenvironment (Ginhoux and Guilliams,
2016; Okabe and Medzhitov, 2016). The fundamental question
addressed was whether the phenotypic and functional features
of MDMs are driven by the nature of the pathogenic insult (i.e.,
the tumor type growing within the brain or instead by the CNS tis-
sue itself), which is generally deficient in blood-borne leukocytes
during the steady state. In order to chart the monocyte/MDM
developmental trajectory in relation to either gliomas versus
BrM, we built a force-directed graph using Vortex (Figure 3C)
(Good et al., 2019; Samusik et al., 2016). The algorithm was
applied to a combined CyTOF dataset of CNS-invading TAMs
of glioma and BrM TMEs, focusing on the expression level of
monocyte/macrophage markers (Figure 3C). The resulting graph
depicts a continuous process of monocyte/MDM differentiation,
which was characterized by the downregulation of CCR2 and
CD33, concomitant with the upregulation of CD163 and Mertk
(Figure 3C). Of note, among MDMs, we observed a trifurcation
in the developmental trajectory displayed in three branches on
the force-directed map, which was driven by the differential
expression of CD169, CD206, CD209, CD38, and PD-L1. To infer
cell lineages and pseudotimes, we computed the slingshot algo-
rithm (Figure 3D) (Street et al., 2018). The algorithm modeled the
developmental trajectory by connecting adjacent clusters, using
the monocyte cluster as a starting point to construct branching
curves (Figure 3C). To explore different subsets of MDM, we per-
formed FlowSOM analysis using the differentially expressed an-
tigens (Figure 3C). Our analysis revealed the presence of four
MDM clusters and one monocyte cluster (Figures 3D and 3E).
Next, we separated the force-directed map of invading
MDMs/monocytes for IDH1mut and IDH1wt gliomas versus mela-
noma and carcinoma BrM to analyze the trajectories and relative
frequencies of different MDM subsets (Figure 3F). Importantly,
we found that the developmental traces of the monocyte-
to-MDM transition is not a random feature across brain tumors,
but that each tumor entity clearly dictates the development of its
own MDM type with a distinctive and tumor type-specific pheno-
typic signature. For instance, CD163+ CX3CR1+ CADM1+ MDMs
(MDM 4) were found almost exclusively in the IDH1wt glioma TME
(Figures 3F and 3G). MDMs expressing CD163, CD206, and
CD169 (MDM 2) showed higher relative frequencies in carci-
noma BrM, whereas MDMs with higher levels of CD209, CD38,
PD-L1, and PD-L2 (MDM 3) were equally infiltrating the mela-
noma and carcinoma BrM (Figures 3F and 3G). CNS-invading
phagocytes in the TME of IDH1mut glioma were predominantly
composed of monocytes and low frequencies of MDMs. These

Resource
findings clearly showed that the monocyte to MDM transition is
dictated by tumor-specific education of the TAM compartment
rather than a general feature of monocyte behavior within CNS
tissue (Figures 3F and 3G).
The localization of both CNS-resident microglia and invading
MDMs within the TME of brain tumors was further investigated
using immunofluorescence and immunohistochemistry. Based
on our single-cell map (Figures 2B, 2C, and 2D), we used the
pan-myeloid marker Iba1 in combination with either CD163 or
P2RY12 to distinguish Iba1+CD163+ MDMs from Iba1+CD163
and Iba1+P2Y12+ microglia (Figures 4A, 4B, and S4A). Iba1+
CD163 microglia were diffusely scattered throughout gliomas,
but they were often confined to tumor border areas and absent
from the tumor core in BrM (Figures 4A and S4A). Moreover,
we found Iba1+CD163+ MDMs and CD206+, CD169+, and
CD209+ subsets of MDMs in close proximity to VE-cadherin-
positive blood vessels in glioma and BrM (Figures 4B and
S4A–S4D). Frequencies of MDMs were higher in samples with
NSCLC BrM, mirroring the results of the CyTOF analysis.
To interrogate the clinical significance of MDMs and CNS-resi-
dent microglia infiltrating the glioma TME, we correlated the fre-
quencies of identified TAM populations with overall survival (OS)
or time of follow-up in days (for patients alive at the time of anal-
ysis). For this purpose, we used our cohort of IDH1wt glioma as
well as two publicly available TCGA databases (TCGA-GBM
and TCGA-LGG). For the group of patients analyzed by CyTOF,
we selected only patients with newly diagnosed IDH1wt glioma
(excluding recurrent cases) due to higher variability of disease
stage and treatment among patients with IDH1mut glioma and
BrM. In the IDH1wt glioma cohort, we did not observe a
correlation between the frequencies of microglia or MDMs
with patient outcome. Only the frequencies of monocytes corre-
lated with a trend of longer OS (Figure S4E). Further, we deter-
mined the expression levels of the pan-microglial marker
CX3CR1 and the pan-MDM marker CD163 in the TCGA-GBM
database (Glioblastoma Multiforme) (Figure 5A). Interestingly,
the MDM marker CD163, in contrast to the microglia marker
CX3CR1, showed a negative correlation with OS, suggesting
that this feature has the potential to stratify patient clinical
outcomes (Figure 5A). We then determined whether the
composite TAM signatures could improve the correlation
with clinical disease development and survival. First, we
analyzed the relation of the microglial signature expression
ll
Figure 4. Localization of CNS-Resident and
Invading Macrophages In Situ
(A and B) Representative immunofluorescence
images of microglia (A) and MDMs (B) in gliomas.
See also Figures S4A–S4D and Table S1.
(ITGAX(CD11c), FCGR1A(CD64), CX3CR1,
and P2RY12(P2Y12)) and the OS in the
TCGA-LGG (Brain Lower Grade Glioma;
WHO II/III grade) and TCGA-GBM data-
bases, which again did not reveal any sig-
nificant correlation (Figures 5B and 5C).
Conversely, applying the different gene
signatures of MDMs demonstrated a significant positive correla-
tion of the MDM2/3 signature: CD163, MRC1 (CD206), CD14,
and CD68, with patient survival (Figures 5D, 5E). The data sug-
gest that CD206+ MDM2/3 infiltration contributes to tumor sup-
pression, particularly in WHO II and III grade glioma.
Taken together, our findings reveal a diversity of TAM subsets
in the brain TME besides CNS-resident and invading macro-
phages. We propose that TAMs play a distinct role in the TME,
following the strong variation across all brain tumor types in their
phenotype, localization, and correlation with clinical outcomes.
In addition, we observed tumor-specific changes, with resident
phagocytes predominating in glioma and invading phagocytes
in the metastatic brain TME. Moreover, the TME site, rather
than the CNS site, drives TAM polarization, resulting in unique
phenotypes that correlate with clinical outcomes. Importantly,
our results suggest that CD206+ MDMs favor tumor suppression
in WHO II and III grade glioma patients.
Tregs Accumulation and T Cell Exhaustion Characterize
the TME of Brain Metastases
The nature and frequencies of TILs are emerging as a prognostic
marker for several cancer types (Binnewies et al., 2018). Previ-
ous reports have shown a correlation between the number of
TILs and the IDH1 mutational status of gliomas (Bunse et al.,
2018). Here, using CyTOF with a focus on the lymphoid compart-
ment (Table S3), we identified several TIL phenotypes and
described their activation status in glioma and BrM. Upon
t-distributed stochastic neighbor embedding (t-SNE) dimension-
ality reduction in conjunction with FlowSOM clustering, we iden-
tified CD4 (CD3+CD4+FoxP3 ) and CD8 (CD3+CD8+) T cells,
Tregs (CD3+ CD4+FoxP3+CTLA4+), gdT cells (CD3+TCRgd+),
NK cells (CD56+), B cells (CD19+CD20+), and plasma cells
(CD19+ and CD38high) (Figures S5A–S5C). Overall, the relative
frequencies across all lymphocytes and samples were compara-
ble between primary glioma and metastatic samples (Figures 6A
and S5D). Also, we did not observe changes in the lymphocyte
frequencies associated with MGMT promoter methylation status
(Figure S5E). With few exceptions, IDH1wt gliomas had higher
frequencies of Tregs compared to IDH1mut (Figure 6A). However,
the highest relative frequencies of Tregs were found in the BrM,
especially within the carcinoma TME (Figure 6A).
Next, we closely interrogated the phenotype of the CD8 T cells
using categorical One-SENSE analysis (Cheng et al., 2016;
Cell 181, 1626–1642, June 25, 2020 1633
 ll
Resource

Figure 5. Overall Glioma Patient Survival Correlates with the Presence of MDM Signatures
(A) Kaplan-Meier curve shows OS in glioblastoma patients (TCGA-GBM database) with high and low CD163 or CX3CR1 gene expression.
(B–E) Kaplan-Meier survival analysis in patient groups of (B) TCGA-LGG and (C) TCGA-GBM databases with high and low microglial signature. (D and E) Kaplan-
Meier survival analysis in patient groups of (D) TCGA-LGG and (E) TCGA-GBM databases correlated with MDM2/3 signature. (B–E) Heatmaps show the selected
groups of patients according to the gene expression.
See also Figure S4E and Table S1.

1634 Cell 181, 1626–1642, June 25, 2020

 ll
Resource

Figure 6. Tregs Accumulation and T Cell Exhaustion Characterize the TME of Brain Metastases
(A) Relative frequencies of the main T cell populations among lymphocytes. Only statistically significant p values were displayed (p < 0.05, Mann-Whitney-
Wilcoxon test, Benjamini-Hochberg correction), whiskers within 1.5x IQR.
(B) One-SENSE analysis comparing the lineage and activation profiles of CD8 T cells.
(C) Representative histograms showing differentially expressed markers on CD8 RM and CD8 EM subsets.
See also Figures S5, S6, and S7A and Table S1.

Cell 181, 1626–1642, June 25, 2020 1635

 ll
Laurens van der Maaten, 2008). Here, the x-axis represents a
naive/memory profile and the y-axis the activation profile,
including expression of co-stimulatory and co-inhibitory recep-
tors (Figure 6B). Based on the expression of five markers
(x-axis: CD45RA, CD45RO, CCR7, CD127, and CD103), T cells
can be differentiated into naive, central memory (CM), effector
memory (EM), terminally differentiated effector memory
(TEMRA), and non-circulating tissue-resident (RM) T cells (Smol-
ders et al., 2018) (Figures 6B and S6A). The majority of T cells
analyzed were memory T cells, without statistically significant
variations in terms of relative frequencies among total T cell
numbers between the TME of gliomas versus BrM (Figure S6B).
However, we observed a positive correlation of CD4 CM (p =
0.021) and CD8 CM (p = 0.077) T cell frequencies with OS/
number of follow up days for patients with IDH1wt glioma
(Figure S6C).
Further analysis of additional T cell markers demonstrated a
higher expression of co-stimulatory receptors (ICOS, CD27,
and CD137), co-inhibitory receptors (2B4, TIGIT, and PD-1),
the activation marker CD38, effector (CD57 and granzyme B)
and proliferation functions (Ki-67) among RM and EM CD8
T cells in melanoma BrM (Figures 6B, 6C, and S7A). The same
phenotype, albeit with lower frequencies, was identified in carci-
noma BrM. Of note, CD8 RM and EM T cells in IDH1mut glioma
had lower expression of proliferation and activation markers
compared to DH1wt glioma. The same analysis strategy was
applied to CD4 T cells; however, here we did not observe major
differences between glioma and BrM.
Altogether, the data revealed the accumulation of Tregs in the
TME of IDH1wt gliomas and BrM, with the highest frequencies in
the carcinoma BrM. In IDH1wt gliomas, the accumulation of CD4
CM and CD8 CM in the TME may be a positive prognostic indi-
cator of better OS. In addition, the expansion of the analyzed
markers revealed unique phenotypic features of EM and RM
CD8 T cells. The metastatic TME was composed of activated/
exhausted T cells, whereas the glioma samples showed a lower
expression of activation markers.
Immature NK Cells Accumulate in Glioblastoma
In view of the potential role of innate lymphoid cells in anti-tumor
immunity (Tugues et al., 2019), we next used the lymphoid panel
(Table S3) dataset to interrogate NK cells (CD56+CD3-) (Fig-
ure 7A). We assigned cells mainly corresponding to the combina-
torial expression of CD16 and CD56 and identified two major
populations of CD56int/brightCD16 and CD56intCD16+, which
correspond to the immature and the high cytotoxic populations
of NK cells, respectively (Simoni et al., 2017). Among
CD56int/brightCD16 cells, we also found a major population of
CD69+CD103+CD56+ cells, which closely resemble intraepithe-
lial ILC1-like cells (ie-ILC1-like cells) (Figure 7A) (Simoni et al.,
2017). Next, we compared the frequencies of these three
above-described CD56+ subsets in gliomas and BrM of different
origins. In the IDH1wt gliomas, we observed the enrichment of
immature CD56int/bright CD16 NK cells among lymphocytes,
whereas in both the IDH1mut gliomas and BrM, predominantly
CD56int/bright CD16+ NK cells (Figures 7B and S7B) accumulated.
Further splitting of the IDH1wt glioma cohort according to the
methylation status of the MGMT promoter indicated a trend
1636 Cell 181, 1626–1642, June 25, 2020
Resource
(p = 0.1) toward a higher proportion of CD56int/brightCD16+ NK
cells in the unmethylated cases (Figure S7C). We also correlated
the frequencies of CD56+ subsets with OS or number of follow up
days in the IDH1wt cohort, which suggested ie-ILC1-like cell
accumulation as a potential marker of OS (Figure S7D).
We next characterized the activation status of the three sub-
sets of CD56+ cells, represented in the y-axis of the One-SENSE
map (Cheng et al., 2016; Laurens van der Maaten, 2008) (Fig-
ure 7C). Detailed analysis of the CD56int/brightCD16+ population
across patients revealed a higher trend of 2B4, CD38, and
Ki-67 in BrM (Figures 7C and 7D). The CD56int/brightCD16+ pop-
ulation in IDH1wt gliomas was distinguished by CD57 and TIM3
expression, with the latter also observed in the IDH1mut gliomas.
The profile of metastatic and glioma CD56int/brightCD16+ popula-
tion was relatively homogeneous. Taken together, NK cell pro-
files also display tumor specificity, highlighted by the proportion
of infiltrating cytotoxic and immature NK cells, with the preva-
lence of the latter in the IDH1wt gliomas. Additionally, we
observed heterogeneity of infiltrating CD56int/brightCD16 NK
cells between glioma and BrM.
DISCUSSION
In this study, we combined extensive single-cell proteome anal-
ysis, immunofluorescence imaging, and genetic fate-mapping to
interrogate the leukocyte landscape of the TME in brain tumors.
Our analyses revealed major changes in the CNS-resident and
invading leukocyte populations, which were dictated by the
type of brain tumor. We found resident phagocytes to be abun-
dant in the glioma TME, whereas invading leukocytes dominate
the immune landscape of BrM.
Previous studies have shown that TAMs are the largest popu-
lation of leukocytes in the glioma TME. Despite the correlation of
TAMs with clinical prognosis and grade of glioma (Quail and
Joyce, 2017; Venteicher et al., 2017), it appears that cellular phe-
notypes are more indicative of clinical outcomes than the mere
number of infiltrating macrophages (Müller et al., 2017; Pyonteck
et al., 2013). For instance, blocking the colony-stimulating factor
receptor 1 (CSF-1R) resulted in a significant reduction of tumor
growth in a preclinical glioma model, which was associated
with a ‘‘re-education’’ of TAMs toward a pro-inflammatory tu-
mor-suppressive phenotype (Coniglio et al., 2012; Pyonteck
et al., 2013). Despite the general interest in targeting TAMs for
anti-glioma therapy, only a few studies performed in preclinical
models have taken into account the dual origin of the TAM pop-
ulation (Bowman et al., 2016; Chen et al., 2017). Discrimination of
microglia and the blood-derived MDM revealed the predomi-
nance of ‘‘reactive’’ microglia-derived TAMs in glioma lesions,
which is in line with a recent report by Sankowski et al. (2019).
This phenotype likely results from interferon (IFN)-g produced
by TILs (both T and NK cells), a notion that can be further inves-
tigated in preclinical mouse models.
Of note, we observed high relative frequencies of MDMs (not
measured in Sankowski et al. [2019]) in IDH1wt glioma and
BrM. In contrast to microglia frequencies, MDMs significantly
correlated with clinical outcome of glioma patients and particu-
larly LGG. Interestingly, CX3CR1+CADM1+ MDMs observed in
the IDH1wt glioma TME were phenotypically different from those
 ll
Resource

Figure 7. Immature NK Cells Accumulate in Glioblastoma

(A) Gating strategy to identify ILC (CD56+) cell subsets.
(B) Relative frequencies of ILC populations among CD56+ cells.
(C) One-SENSE analysis comparing the lineage and activation profiles of CD56+ cells.
(D) The antigen median expression of the CD56int/brightCD16neg subset, which showed a statistically significant difference between glioma and BrM. (C and D) Only
statistically significant p values were displayed (p < 0.05, Mann-Whitney-Wilcoxon test, Benjamini-Hochberg correction), whiskers within 1.5x IQR.
See also Figures S7B–S7D and Table S1.

Cell 181, 1626–1642, June 25, 2020 1637

 ll
identified in BrM. Recently developed preclinical glioma models
will likely fuel the study of novel roles as well as unique
trajectories of MDMs. For instance, the model mimicking the
overproduction of (R)-2-hydroxyglutarate driven by IDH mutation
(Amankulor et al., 2017) or those recreating various molecular
types of gliomas (Miyai et al., 2017) can be used to further inves-
tigate differentially matured MDMs upon extrinsic tumor factors.
Another attractive approach to study unique combination of
TAMs for personalized therapy is the engraftment of patient-
derived glioblastoma organoids, which represent the molecular
heterogeneity among gliomas (Jacob et al., 2020; Mansour
et al., 2018).
In the BrM TME, we found a unique subpopulation of MDMs
expressing CD206, CD209, CD169, and CD163 as well as high
levels of CD38, PDL-1, and PDL-2, which is reminiscent of a tu-
mor population of TAMs described by Ries et al. (2014) in
NSCLC. MDMs expressing CD38 were also described in vitro
on IFN-g stimulation, whereas CD38 expression was strongly
associated with phagocytosis (Schulz et al., 2019). In our study,
we did not observe a clear separation of TAMs based on the M1/
M2 model of macrophage polarization (Murray et al., 2014),
which underestimates the dynamic nature of pro- or anti-inflam-
matory properties across macrophages (Xue et al., 2014).
Instead, the unique signature of cytokines, growth factors, and
mutational landscapes that instruct TAMs in each brain tumor
type may trigger the phenotypic differences observed.
In most cases, we found MDMs localized near blood vessels in
glioma and BrM. Previous preclinical glioma models showed a
close connection of MDMs and glioma stem cells, which were
mainly found in perivascular niches and reported to secrete peri-
ostin to attract TAMs (Zhou et al., 2015). In turn, TAMs are a rich
source of soluble mediators (e.g., interleukin [IL]-6, IL-10, and
transforming growth factor-b1 and pleiotrophin) capable of sus-
taining malignant stem cells (Shi et al., 2017). Regarding BrM, we
hypothesize that MDMs localize near blood vessels to establish
the metastatic niche. However, further studies using autochtho-
nous mouse models of BrM (An et al., 2017) are required to reveal
potential targets for immunotherapy.
Regarding the lymphocyte compartment, our analyses re-
vealed that Tregs preferentially accumulate in BrM rather than
in gliomas. Importantly, the presence of PD-L1+ TAMs has
been correlated with the Treg frequencies in several solid tissue
tumors (Harter et al., 2015) as well as IDH1wt gliomas (Berghoff
et al., 2017). In turn, Tregs secrete IL-10, IL-4, and IL-13, which
may trigger the development of TAMs with immunosuppressive
properties (Mantovani et al., 2017). An increased number of
Tregs might also result in the suppression of cytotoxic CD8
T cell responses. Brain tumors establish an immunosuppressive
TME that leads to T cell dysfunction (Quail and Joyce, 2017).
T cells infiltrating glioblastoma, for instance, were found to ex-
press high amounts of multiple immune checkpoints (e.g.,
PD-1, Lag-3, TIM-3, or TIGIT), which correlated with a loss of
effector function (Woroniecka and Fecci, 2018). PD-1 expression
has been found in a high percentage of TILs in melanoma BrM
(Berghoff et al., 2015). However, we found the greatest changes
in T cell activation in BrM, with CD8 TRM (CD103+; CD69+) and
TEM (CCR7 ; CD45RA ) displaying an activated phenotype
characterized by high amounts of both co-stimulatory and co-
1638 Cell 181, 1626–1642, June 25, 2020
Resource
inhibitory molecules, as well as proliferation markers. How this
activated phenotype is translated into certain functional proper-
ties is not yet known and will require further investigation.
However, the map provided here represents an important
resource for the informed design of novel therapeutic avenues.
A similar trend was observed in the NK cell compartment
where the proportion of infiltrating cytotoxic and immature NK
cells decreased with disease severity. In gliomas, the impaired
lymphocyte populations are in line with the TCGA data that char-
acterize IDH1wt gliomas as lymphocyte-depleted and IDH1mut
gliomas as immunologically silent (Kiss et al., 2018; Thorsson
et al., 2018). In neuro-oncology, the IDH1 status is leveraged
as an important classifier (Louis et al., 2016), with IDH1mut pa-
tients showing a favorable survival (Eckel-Passow et al., 2015;
Unruh et al., 2019). It is becoming more evident that IDH1 status
is not only prognostically relevant but represents a marker for a
distinct disease entity within gliomas. The distinct molecular
characteristics of IDH1mut glioma and the distinct clinical
behavior are now also well characterized by a distinct TME: im-
mune cells show the least amount of activation across all brain
tumor samples.
Taken together, we show that the immune cell compartment of
brain TMEs is mainly shaped by the specific tumor type rather
than by the CNS as an ‘‘immune-privileged niche.’’ We observed
a continuous increase in bone marrow-derived infiltrates, such
as MDMs and T cells, along the axis of IDH1mut, IDH1wt gliomas,
and BrM, and increasing dominance of CNS-resident cells, such
as microglia, in the glioma counterparts. The richness and acti-
vation of the BrM TMEs regarding cellular subtypes and fre-
quencies as well as functional states parallels their favorable
clinical response to checkpoint inhibitors. In-depth knowledge
of the specific immunological TME signatures across brain tu-
mors is a major step forward for the rational design of targeted
immunotherapy strategies.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Human Brain Tissue Samples
B Animal Models
B Cell Lines
d METHOD DETAILS
B Processing of Human Samples for Cytometry Analysis
B Tissue collection for Immunofluorescence
B Orthotopic Glioma Cell Injection
B In Vivo Bioluminescent Imaging
B Harvesting and Processing of Mouse Brain Samples
B Mass-Tag Cellular Barcoding
B Metal-Isotope-Tagged Antibodies
B Cell Surface Staining for Cytometry

Resource
B Intracellular Cytokine Staining for Cytometry
B Cell Preparation and Mass Cytometry Acquisition
B Flow Cytometry Acquisition
B Immunohistochemistry
B Immunofluorescence
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Preprocessing of Cytometry Data
B Automated Population Identification
B Survival Analysis
B Statistical Analysis
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.04.055.
ACKNOWLEDGMENTS
Sall1CreER mice were kindly provided by R. Nishinakamura (Kumamoto Uni-
versity). We thank the Mass- and Flow Cytometry Facility (University of Zurich)
for technical assistance, the Department of Neuropathology (University Hospi-
tal Zurich and University of Zurich) for performing immunohistochemistry stain-
ing, and Helen Pickersgill of Life Science Editors for critical review and editing
of the manuscript. This work was supported by grants from the Swiss Cancer
League (KFS-4431-02-2018), the Swiss National Science Foundation (733
310030_170320, 310030_188450, and CRSII5_183478) to B.B., the European
Union H2020 Project iPC (826121 to B.B.), the Clinical Research Priority Pro-
gram ImmunoCure (to B.B., M.W., and M.C.N.), and the University Research
Priority Program Translational Cancer Research of the University of Zurich
(to B.B. and M.W.).
AUTHOR CONTRIBUTIONS
Conceptualization, B.B., E.F., M.W., and M.C.N.; Methodology, E.F. and K.K.;
Software, E.F.; Formal Analysis, E.F. and N.G.N.; Investigation, E.F., K.K., S.
Unger, S. Utz, and N.G.N.; Resources, B.B., M.C.N., M.W., and L.R.; Data Cu-
ration, E.F., S. Unger, N.G.N., K.K., M.C.N., and E.J.R.; Writing – Original Draft,
E.F., B.B., and S.T.; Writing – Review & Editing, E.F., B.B., S.T., K.K., M.C.N.,
E.J.R., M.W., and M.G.; Visualization, E.F.; Funding Acquisition and Supervi-
sion, M.C.N., M.W., and B.B.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: November 27, 2019
Revised: March 11, 2020
Accepted: April 28, 2020
Published: May 28, 2020
REFERENCES
Aldape, K., Brindle, K.M., Chesler, L., Chopra, R., Gajjar, A., Gilbert, M.R., Got-
tardo, N., Gutmann, D.H., Hargrave, D., Holland, E.C., et al. (2019). Challenges
to curing primary brain tumours. Nat. Rev. Clin. Oncol. 16, 509–520.
Amankulor, N.M., Kim, Y., Arora, S., Kargl, J., Szulzewsky, F., Hanke, M., Mar-
gineantu, D.H., Rao, A., Bolouri, H., Delrow, J., et al. (2017). Mutant IDH1 reg-
ulates the tumor-associated immune system in gliomas. Genes Dev. 31,
774–786.
Amit, M., Laider-Trejo, L., Shalom, V., Shabtay-Orbach, A., Krelin, Y., and Gil,
Z. (2013). Characterization of the melanoma brain metastatic niche in mice and
humans. Cancer Med. 2, 155–163.
An, J., Wang, L., Zhao, Y., Hao, Q., Zhang, Y., Zhang, J., Yang, C., Liu, L.,
Wang, W., Fang, D., et al. (2017). Effects of FSTL1 on cell proliferation in breast
ll
cancer cell line MDA-MB-231 and its brain metastatic variant MDA-MB-231-
BR. Oncol. Rep. 38, 3001–3010.
Bastian, M., Heymann, S., and Jacomy, M. (2009). Gephi: An open source soft-
ware for exploring and manipulating networks. ICWSM 8, 361–362.
Bendall, S.C., Simonds, E.F., Qiu, P., Amir, A.D., Krutzik, P.O., Finck, R.,
Bruggner, R.V., Melamed, R., Trejo, A., Ornatsky, O.I., et al. (2011). Single-
cell mass cytometry of differential immune and drug responses across a hu-
man hematopoietic continuum. Science 332, 687–696.
Berghoff, A.S., Kiesel, B., Widhalm, G., Rajky, O., Ricken, G., Wöhrer, A., Die-
ckmann, K., Filipits, M., Brandstetter, A., Weller, M., et al. (2015). Programmed
death ligand 1 expression and tumor-infiltrating lymphocytes in glioblastoma.
Neuro-oncol. 17, 1064–1075.
Berghoff, A.S., Kiesel, B., Widhalm, G., Wilhelm, D., Rajky, O., Kurscheid, S.,
Kresl, P., Wöhrer, A., Marosi, C., Hegi, M.E., and Preusser, M. (2017). Correla-
tion of immune phenotype with IDH mutation in diffuse glioma. Neuro-oncol.
19, 1460–1468.
Bienkowski, M., and Preusser, M. (2015). Prognostic role of tumour-infiltrating
inflammatory cells in brain tumours: literature review. Curr. Opin. Neurol. 28,
647–658.
Binnewies, M., Roberts, E.W., Kersten, K., Chan, V., Fearon, D.F., Merad, M.,
Coussens, L.M., Gabrilovich, D.I., Ostrand-Rosenberg, S., Hedrick, C.C., et al.
(2018). Understanding the tumor immune microenvironment (TIME) for effec-
tive therapy. Nat. Med. 24, 541–550.
Bouffet, E., Larouche, V., Campbell, B.B., Merico, D., de Borja, R., Aronson,
M., Durno, C., Krueger, J., Cabric, V., Ramaswamy, V., et al. (2016). Immune
Checkpoint Inhibition for Hypermutant Glioblastoma Multiforme Resulting
From Germline Biallelic Mismatch Repair Deficiency. J. Clin. Oncol. 34,
2206–2211.
Bowman, R.L., Klemm, F., Akkari, L., Pyonteck, S.M., Sevenich, L., Quail, D.F.,
Dhara, S., Simpson, K., Gardner, E.E., Iacobuzio-Donahue, C.A., et al. (2016).
Macrophage Ontogeny Underlies Differences in Tumor-Specific Education in
Brain Malignancies. Cell Rep. 17, 2445–2459.
Buja, A., Swayne, D.F., Littman, M.L., Dean, N., Hofmann, H., and Chen, L.
(2008). Data Visualization with Multidimensional Scaling Introduction.
J. Comput. Graph. Stat. 17, 444–472.
Bunse, L., Pusch, S., Bunse, T., Sahm, F., Sanghvi, K., Friedrich, M., Alansary,
D., Sonner, J.K., Green, E., Deumelandt, K., et al. (2018). Suppression of anti-
tumor T cell immunity by the oncometabolite (R)-2-hydroxyglutarate. Nat.
Med. 24, 1192–1203.
Butovsky, O., Jedrychowski, M.P., Moore, C.S., Cialic, R., Lanser, A.J., Ga-
briely, G., Koeglsperger, T., Dake, B., Wu, P.M., Doykan, C.E., et al. (2014).
Identification of a unique TGF-b-dependent molecular and functional signature
in microglia. Nat. Neurosci. 17, 131–143.
Buttgereit, A., Lelios, I., Yu, X., Vrohlings, M., Krakoski, N.R., Gautier, E.L.,
Nishinakamura, R., Becher, B., and Greter, M. (2016). Sall1 is a transcriptional
regulator defining microglia identity and function. Nat. Immunol. 17,
1397–1406.
Chen, Z., Feng, X., Herting, C.J., Garcia, V.A., Nie, K., Pong, W.W., Rasmus-
sen, R., Dwivedi, B., Seby, S., Wolf, S.A., et al. (2017). Cellular and molecular
identity of tumor-associated macrophages in glioblastoma. Cancer Res. 77,
2266–2278.
Cheng, Y., Wong, M.T., van der Maaten, L., and Newell, E.W. (2016). Categor-
ical Analysis of Human T Cell Heterogeneity with One-Dimensional Soli-
Expression by Nonlinear Stochastic Embedding. J. Immunol. 196, 924–932.
Colaprico, A., Silva, T.C., Olsen, C., Garofano, L., Cava, C., Garolini, D., Sabe-
dot, T.S., Malta, T.M., Pagnotta, S.M., Castiglioni, I., et al. (2016). TCGAbio-
links: an R/Bioconductor package for integrative analysis of TCGA data. Nu-
cleic Acids Research 44, e71.
Coniglio, S.J., Eugenin, E., Dobrenis, K., Stanley, E.R., West, B.L., Symons,
M.H., and Segall, J.E. (2012). Microglial stimulation of glioblastoma invasion in-
volves epidermal growth factor receptor (EGFR) and colony stimulating factor
1 receptor (CSF-1R) signaling. Mol. Med. 18, 519–527.
Cell 181, 1626–1642, June 25, 2020 1639
 ll
Croxford, A.L., Lanzinger, M., Hartmann, F.J., Schreiner, B., Mair, F., Pelczar,
P., Clausen, B.E., Jung, S., Greter, M., and Becher, B. (2015). The Cytokine
GM-CSF Drives the Inflammatory Signature of CCR2+ Monocytes and Li-
censes Autoimmunity. Immunity 43, 502–514.
Daniel, P., Sabri, S., Chaddad, A., Meehan, B., Jean-Claude, B., Rak, J., and
Abdulkarim, B.S. (2019). Temozolomide Induced Hypermutation in Glioma:
Evolutionary Mechanisms and Therapeutic Opportunities. Front. Oncol. 9, 41.
Darmanis, S., Sloan, S.A., Croote, D., Mignardi, M., Chernikova, S., Samgha-
babi, P., Zhang, Y., Neff, N., Kowarsky, M., Caneda, C., et al. (2017). Single-
Cell RNA-Seq Analysis of Infiltrating Neoplastic Cells at the Migrating Front
of Human Glioblastoma. Cell Rep. 21, 1399–1410.
de Boer, J., Williams, A., Skavdis, G., Harker, N., Coles, M., Tolaini, M., Norton,
T., Williams, K., Roderick, K., Potocnik, A.J., and Kioussis, D. (2003). Trans-
genic mice with hematopoietic and lymphoid specific expression of Cre.
Eur. J. Immunol. 33, 314–325.
Eckel-Passow, J.E., Lachance, D.H., Molinaro, A.M., Walsh, K.M., Decker,
P.A., Sicotte, H., Pekmezci, M., Rice, T., Kosel, M.L., Smirnov, I.V., et al.
(2015). Glioma Groups Based on 1p/19q, IDH, and TERT Promoter Mutations
in Tumors. N. Engl. J. Med. 372, 2499–2508.
Eisenring, M., vom Berg, J., Kristiansen, G., Saller, E., and Becher, B. (2010).
IL-12 initiates tumor rejection via lymphoid tissue-inducer cells bearing the
natural cytotoxicity receptor NKp46. Nat. Immunol. 11, 1030–1038.
Ellis, B., Haaland, P., Hahne, F., Le Meur, N., Gopalakrishnan, N., Spidlen, J.,
Jiang, M., and Finak, G. (2019). flowCore: Basic structures for flow cytometry
data. https://rdrr.io/bioc/flowCore/.
Field, A., Miles, J., and Field, Z. (2013). Discovering Statistics Using R (Sage).
Finak, G. (2018). flowWorkspaceData: A data package containing two
flowJo, one diva xml workspace and the associated fcs files as well
as three GatingSets for testing the flowWorkspace, openCyto and
CytoML packages. https://bioconductor.org/packages/release/data/
experiment/html/flowWorkspaceData.html.
Finck, R., Simonds, E.F., Jager, A., Krishnaswamy, S., Sachs, K., Fantl, W.,
Pe’er, D., Nolan, G.P., and Bendall, S.C. (2013). Normalization of mass cytom-
etry data with bead standards. Cytometry A 83, 483–494.
Forrester, J.V., McMenamin, P.G., and Dando, S.J. (2018). CNS infection and
immune privilege. Nat. Rev. Neurosci. 19, 655–671.
Ginhoux, F., and Guilliams, M. (2016). Tissue-Resident Macrophage Ontogeny
and Homeostasis. Immunity 44, 439–449.
Ginhoux, F., Greter, M., Leboeuf, M., Nandi, S., See, P., Gokhan, S., Mehler,
M.F., Conway, S.J., Ng, L.G., Stanley, E.R., et al. (2010). Fate mapping analysis
reveals that adult microglia derive from primitive macrophages. Science 330,
841–845.
Goldberg, S.B., Gettinger, S.N., Mahajan, A., Chiang, A.C., Herbst, R.S.,
Sznol, M., Tsiouris, A.J., Cohen, J., Vortmeyer, A., Jilaveanu, L., et al. (2016).
Pembrolizumab for patients with melanoma or non-small-cell lung cancer
and untreated brain metastases: early analysis of a non-randomised, open-la-
bel, phase 2 trial. Lancet Oncol. 17, 976–983.
Goldmann, T., Wieghofer, P., Jordão, M.J.C., Prutek, F., Hagemeyer, N., Fren-
zel, K., Amann, L., Staszewski, O., Kierdorf, K., Krueger, M., et al. (2016).
Origin, fate and dynamics of macrophages at central nervous system inter-
faces. Nat. Immunol. 17, 797–805.
Good, Z., Borges, L., Vivanco Gonzalez, N., Sahaf, B., Samusik, N., Tibshirani,
R., Nolan, G.P., and Bendall, S.C. (2019). Proliferation tracing with single-cell
mass cytometry optimizes generation of stem cell memory-like T cells. Nat.
Biotechnol. 37, 259–266.
Gu, Z., Gu, L., Eils, R., Schlesner, M., and Brors, B. (2014). circlize Implements
and enhances circular visualization in R. Bioinformatics 30, 2811–2812.
Hahne, F., Gopalakrishnan, N., Khodabakhshi, A.H., Wong, C., and Lee, K.
(2020). flowStats: Statistical methods for the analysis of flow cytometry data.
R package version 4.0.0. http://www.github.com/RGLab/flowStats.
Hambardzumyan, D., Gutmann, D.H., and Kettenmann, H. (2016). The role of
microglia and macrophages in glioma maintenance and progression. Nat.
Neurosci. 19, 20–27.
1640 Cell 181, 1626–1642, June 25, 2020
Resource
Harrell, F.E. (2020). Hmisc: Harrell Miscellaneous. http://biostat.mc.vanderbilt.
edu/Hmisc. https://github.com/harrelfe/Hmisc.
Harter, P.N., Bernatz, S., Scholz, A., Zeiner, P.S., Zinke, J., Kiyose, M., Blasel,
S., Beschorner, R., Senft, C., Bender, B., et al. (2015). Distribution and prog-
nostic relevance of tumor-infiltrating lymphocytes (TILs) and PD-1/PD-L1 im-
mune checkpoints in human brain metastases. Oncotarget 6, 40836–40849.
Hartmann, F.J., Bernard-Valnet, R., Quériault, C., Mrdjen, D., Weber, L.M.,
Galli, E., Krieg, C., Robinson, M.D., Nguyen, X.-H., Dauvilliers, Y., et al.
(2016). High-dimensional single-cell analysis reveals the immune signature
of narcolepsy. J. Exp. Med. 213, 2621–2633.
Haynes, S.E., Hollopeter, G., Yang, G., Kurpius, D., Dailey, M.E., Gan, W.-B.,
and Julius, D. (2006). The P2Y12 receptor regulates microglial activation by
extracellular nucleotides. Nat. Neurosci. 9, 1512–1519.
Hopperton, K.E., Mohammad, D., Trépanier, M.O., Giuliano, V., and Bazinet,
R.P. (2018). Markers of microglia in post-mortem brain samples from patients
with Alzheimer’s disease: a systematic review. Mol. Psychiatry 23, 177–198.
Inoue, S., Inoue, M., Fujimura, S., and Nishinakamura, R. (2010). A mouse line
expressing Sall1-driven inducible Cre recombinase in the kidney mesen-
chyme. Genesis 48, 207–212.
Jacob, F., Salinas, R.D., Zhang, D.Y., Nguyen, P.T.T., Schnoll, J.G., Wong,
S.Z.H., Thokala, R., Sheikh, S., Saxena, D., Prokop, S., et al. (2020). A Pa-
tient-Derived Glioblastoma Organoid Model and Biobank Recapitulates Inter-
and Intra-tumoral Heterogeneity. Cell 180, 188–204.
Jacobs, J.F.M., Idema, A.J., Bol, K.F., Grotenhuis, J.A., de Vries, I.J.M., Wes-
seling, P., and Adema, G.J. (2010). Prognostic significance and mechanism of
Treg infiltration in human brain tumors. J. Neuroimmunol. 225, 195–199.
Johanns, T.M., Miller, C.A., Dorward, I.G., Tsien, C., Chang, E., Perry, A., Up-
paluri, R., Ferguson, C., Schmidt, R.E., Dahiya, S., et al. (2016). Immunoge-
nomics of Hypermutated Glioblastoma: A Patient with Germline POLE Defi-
ciency Treated with Checkpoint Blockade Immunotherapy. Cancer Discov.
6, 1230–1236.
Keenan, T.E., Burke, K.P., and Van Allen, E.M. (2019). Genomic correlates of
response to immune checkpoint blockade. Nat. Med. 25, 389–402.
Keren-Shaul, H., Spinrad, A., Weiner, A., Matcovitch-Natan, O., Dvir-Sztern-
feld, R., Ulland, T.K., David, E., Baruch, K., Lara-Astaiso, D., Toth, B., et al.
(2017). A Unique Microglia Type Associated with Restricting Development of
Alzheimer’s Disease. Cell 169, 1276–1290.
Kiss, M., Van Gassen, S., Movahedi, K., Saeys, Y., and Laoui, D. (2018).
Myeloid cell heterogeneity in cancer: not a single cell alike. Cell. Immunol.
330, 188–201.
Kluger, H.M., Chiang, V., Mahajan, A., Zito, C.R., Sznol, M., Tran, T., Weiss,
S.A., Cohen, J.V., Yu, J., Hegde, U., et al. (2019). Long-Term Survival of Pa-
tients With Melanoma With Active Brain Metastases Treated With Pembrolizu-
mab on a Phase II Trial. J. Clin. Oncol. 37, 52–60.
Kohanbash, G., Carrera, D.A., Shrivastav, S., Ahn, B.J., Jahan, N., Mazor, T.,
Chheda, Z.S., Downey, K.M., Watchmaker, P.B., Beppler, C., et al. (2017). Iso-
citrate dehydrogenase mutations suppress STAT1 and CD8+ T cell accumula-
tion in gliomas. J. Clin. Invest. 127, 1425–1437.
Kolde, R. (2019). pheatmap: Pretty Heatmaps. https://rdrr.io/cran/pheatmap/.
Kuppner, M.C., Hamou, M.-F., Bodmer, S., Fontana, A., and de Tribolet, N.
(1988). The glioblastoma-derived T-cell suppressor factor/transforming
growth factor beta 2 inhibits the generation of lymphokine-activated killer
(LAK) cells. Int. J. Cancer 42, 562–567.
Laurens van der Maaten, G.H. (2008). Visualizing Data using t-SNE. J. Mach.
Learn. Res. 9, 2579–2605.
Louis, D.N., Perry, A., Reifenberger, G., von Deimling, A., Figarella-Branger, D.,
Cavenee, W.K., Ohgaki, H., Wiestler, O.D., Kleihues, P., and Ellison, D.W.
(2016). The 2016 World Health Organization Classification of Tumors of the
Central Nervous System: a summary. Acta Neuropathol. 131, 803–820.
Mansour, A.A., Gonçalves, J.T., Bloyd, C.W., Li, H., Fernandes, S., Quang, D.,
Johnston, S., Parylak, S.L., Jin, X., and Gage, F.H. (2018). An in vivo model of
functional and vascularized human brain organoids. Nat. Biotechnol. 36,
432–441.

Resource
Mantovani, A., Marchesi, F., Malesci, A., Laghi, L., and Allavena, P. (2017).
Tumour-associated macrophages as treatment targets in oncology. Nat.
Rev. Clin. Oncol. 14, 399–416.
Mcinnes, L., Healy, J., and Melville, J. (2018). UMAP: Uniform Manifold
Approximation and Projection for Dimension Reduction. arXiv,
arXiv:1802.03426.
Mei, H.E., Leipold, M.D., and Maecker, H.T. (2016). Platinum-conjugated anti-
bodies for application in mass cytometry. Cytometry A 89, 292–300.
Miyai, M., Tomita, H., Soeda, A., Yano, H., Iwama, T., and Hara, A. (2017). Cur-
rent trends in mouse models of glioblastoma. J. Neurooncol. 135, 423–432.
Mrdjen, D., Hartmann, F.J., and Becher, B. (2017). High Dimensional Cytome-
try of Central Nervous System Leukocytes During Neuroinflammation.
Methods Mol. Biol. 1559, 321–332.
Mrdjen, D., Pavlovic, A., Hartmann, F.J., Schreiner, B., Utz, S.G., Leung, B.P.,
Lelios, I., Heppner, F.L., Kipnis, J., Merkler, D., et al. (2018). High-Dimensional
Single-Cell Mapping of Central Nervous System Immune Cells Reveals
Distinct Myeloid Subsets in Health, Aging, and Disease. Immunity 48, 380–395.
Müller, S., Kohanbash, G., Liu, S.J., Alvarado, B., Carrera, D., Bhaduri, A.,
Watchmaker, P.B., Yagnik, G., Di Lullo, E., Malatesta, M., et al. (2017). Sin-
gle-cell profiling of human gliomas reveals macrophage ontogeny as a basis
for regional differences in macrophage activation in the tumor microenviron-
ment. Genome Biol. 18, 234.
Murray, P.J., Allen, J.E., Biswas, S.K., Fisher, E.A., Gilroy, D.W., Goerdt, S.,
Gordon, S., Hamilton, J.A., Ivashkiv, L.B., Lawrence, T., et al. (2014). Macro-
phage activation and polarization: nomenclature and experimental guidelines.
Immunity 41, 14–20.
Noble, W.S. (2009). How does multiple testing correction work? Nat. Bio-
technol. 27, 1135–1137.
Nowicka, M., Krieg, C., Crowell, H.L., Weber, L.M., Hartmann, F.J., Guglietta,
S., Becher, B., Levesque, M.P., and Robinson, M.D. (2017). CyTOF workflow:
differential discovery in high-throughput high-dimensional cytometry data-
sets. F1000Res. 6, 748.
Okabe, Y., and Medzhitov, R. (2016). Tissue biology perspective on macro-
phages. Nat. Immunol. 17, 9–17.
Pyonteck, S.M., Akkari, L., Schuhmacher, A.J., Bowman, R.L., Sevenich, L.,
Quail, D.F., Olson, O.C., Quick, M.L., Huse, J.T., Teijeiro, V., et al. (2013).
CSF-1R inhibition alters macrophage polarization and blocks glioma progres-
sion. Nat. Med. 19, 1264–1272.
Quail, D.F., and Joyce, J.A. (2013). Microenvironmental regulation of tumor
progression and metastasis. Nat. Med. 19, 1423–1437.
Quail, D.F., and Joyce, J.A. (2017). The Microenvironmental Landscape of
Brain Tumors. Cancer Cell 31, 326–341.
Quail, D.F., Bowman, R.L., Akkari, L., Quick, M.L., Schuhmacher, A.J., Huse,
J.T., Holland, E.C., Sutton, J.C., and Joyce, J.A. (2016). The tumor microenvi-
ronment underlies acquired resistance to CSF-1R inhibition in gliomas. Sci-
ence 352, aad3018.
R Development Core Team (2008). R: A language and environment for statis-
tical computing (Vienna, Austria: R Foundation for Statistical Computing).
RStudio Team (2015). RStudio: Integrated Development for R (Boston, MA:
RStudio, Inc.).
Ries, C.H., Cannarile, M.A., Hoves, S., Benz, J., Wartha, K., Runza, V., Rey-
Giraud, F., Pradel, L.P., Feuerhake, F., Klaman, I., et al. (2014). Targeting tu-
mor-associated macrophages with anti-CSF-1R antibody reveals a strategy
for cancer therapy. Cancer Cell 25, 846–859.
Samusik, N., Good, Z., Spitzer, M.H., Davis, K.L., and Nolan, G.P. (2016). Auto-
mated mapping of phenotype space with single-cell data. Nat. Methods 13,
493–496.
Sankowski, R., Böttcher, C., Masuda, T., Geirsdottir, L., Sagar, Sindram, E.,
Seredenina, T., Muhs, A., Scheiwe, C., Shah, M.J., et al. (2019). Mapping mi-
croglia states in the human brain through the integration of high-dimensional
techniques. Nat. Neurosci. 22, 2098–2110.
ll
Schindelin, J., Arganda-Carreras, I., Frise, E., Kaynig, V., Longair, M., Pietzsch,
T., Preibisch, S., Rueden, C., Saalfeld, S., Schmid, B., et al. (2012). Fiji: an
open-source platform for biological-image analysis. Nat. Methods 9, 676–682.
Schulz, D., Severin, Y., Zanotelli, V.R.T., and Bodenmiller, B. (2019). In-Depth
Characterization of Monocyte-Derived Macrophages using a Mass Cytome-
try-Based Phagocytosis Assay. Sci. Rep. 9, 1925.
Shi, Y., Ping, Y.-F., Zhou, W., He, Z.-C., Chen, C., Bian, B.-S.-J., Zhang, L.,
Chen, L., Lan, X., Zhang, X.-C., et al. (2017). Tumour-associated macrophages
secrete pleiotrophin to promote PTPRZ1 signalling in glioblastoma stem cells
for tumour growth. Nat. Commun. 8, 15080.
Simoni, Y., Fehlings, M., Kløverpris, H.N., McGovern, N., Koo, S.L., Loh, C.Y.,
Lim, S., Kurioka, A., Fergusson, J.R., Tang, C.L., et al. (2017). Human Innate
Lymphoid Cell Subsets Possess Tissue-Type Based Heterogeneity in Pheno-
type and Frequency. Immunity 46, 148–161.
Smolders, J., Heutinck, K.M., Fransen, N.L., Remmerswaal, E.B.M., Hom-
brink, P., Ten Berge, I.J.M., van Lier, R.A.W., Huitinga, I., and Hamann, J.
(2018). Tissue-resident memory T cells populate the human brain. Nat. Com-
mun. 9, 4593.
Spitzer, M.H., Gherardini, P.F., Fragiadakis, G.K., Bhattacharya, N., Yuan,
R.T., Hotson, A.N., Finck, R., Carmi, Y., Zunder, E.R., Fantl, W.J., et al.
(2015). IMMUNOLOGY. An interactive reference framework for modeling a dy-
namic immune system. Science 349, 1259425.
Street, K., Risso, D., Fletcher, R.B., Das, D., Ngai, J., Yosef, N., Purdom, E.,
and Dudoit, S. (2018). Slingshot: cell lineage and pseudotime inference for sin-
gle-cell transcriptomics. BMC Genomics 19, 477.
Takasato, M., Osafune, K., Matsumoto, Y., Kataoka, Y., Yoshida, N., Meguro,
H., Aburatani, H., Asashima, M., and Nishinakamura, R. (2004). Identification of
kidney mesenchymal genes by a combination of microarray analysis and
Sall1-GFP knockin mice. Mech. Dev. 121, 547–557.
Takenaka, M.C., Gabriely, G., Rothhammer, V., Mascanfroni, I.D., Wheeler,
M.A., Chao, C.C., Gutiérrez-Vázquez, C., Kenison, J., Tjon, E.C., Barroso,
A., et al. (2019). Control of tumor-associated macrophages and T cells in glio-
blastoma via AHR and CD39. Nat. Neurosci. 22, 729–740.
Tawbi, H.A., Forsyth, P.A., Algazi, A., Hamid, O., Hodi, F.S., Moschos, S.J.,
Khushalani, N.I., Lewis, K., Lao, C.D., Postow, M.A., et al. (2018). Combined
Nivolumab and Ipilimumab in Melanoma Metastatic to the Brain. N. Engl. J.
Med. 379, 722–730.
Thorsson, V., Gibbs, D.L., Brown, S.D., Wolf, D., Bortone, D.S., Ou Yang,
T.-H., Porta-Pardo, E., Gao, G.F., Plaisier, C.L., Eddy, J.A., et al.; Cancer
Genome Atlas Research Network (2018). The Immune Landscape of Cancer.
Immunity 48, 812–830.
Tugues, S., Ducimetiere, L., Friebel, E., and Becher, B. (2019). Innate lymphoid
cells as regulators of the tumor microenvironment. Semin. Immunol. 41,
101270.
Uhl, M., Aulwurm, S., Wischhusen, J., Weiler, M., Ma, J.Y., Almirez, R., Man-
gadu, R., Liu, Y.-W., Platten, M., Herrlinger, U., et al. (2004). SD-208, a Novel
Transforming Growth Factor b Receptor I Kinase Inhibitor, Inhibits Growth and
Invasiveness and Enhances Immunogenicity of Murine and Human Glioma
Cells In vitro and In vivo. Cancer Res. 64, 7954–7961.
Unruh, D., Zewde, M., Buss, A., Drumm, M.R., Tran, A.N., Scholtens, D.M., and
Horbinski, C. (2019). Methylation and transcription patterns are distinct in IDH
mutant gliomas compared to other IDH mutant cancers. Sci. Rep. 9, 8946.
Van Gassen, S., Callebaut, B., Van Helden, M.J., Lambrecht, B.N., Demeester,
P., Dhaene, T., and Saeys, Y. (2015). FlowSOM: Using self-organizing maps for
visualization and interpretation of cytometry data. Cytometry A 87, 636–645.
Van Hove, H., Martens, L., Scheyltjens, I., De Vlaminck, K., Pombo Antunes,
A.R., De Prijck, S., Vandamme, N., De Schepper, S., Van Isterdael, G., Scott,
C.L., et al. (2019). A single-cell atlas of mouse brain macrophages reveals
unique transcriptional identities shaped by ontogeny and tissue environment.
Nat. Neurosci. 22, 1021–1035.
Venteicher, A.S., Tirosh, I., Hebert, C., Yizhak, K., Neftel, C., Filbin, M.G., Hov-
estadt, V., Escalante, L.E., Shaw, M.L., Rodman, C., et al. (2017). Decoupling
Cell 181, 1626–1642, June 25, 2020 1641
 ll
genetics, lineages, and microenvironment in IDH-mutant gliomas by single-
cell RNA-seq. Science 355, eaai8478.
Walker, D.G., and Lue, L.-F. (2015). Immune phenotypes of microglia in human
neurodegenerative disease: challenges to detecting microglial polarization in
human brains. Alzheimers Res. Ther. 7, 56.
Warnes, G.R., Bolker, B., Bonebakker, L., Gentleman, R., Liaw, W.H.A., Lum-
ley, T., Maechler, M., Magnusson, A., Moeller, S., Schwartz, M., et al. (2019).
gplots: Various R Programming Tools for Plotting Data. https://rdrr.io/cran/
gplots/.
Wickham, H., François, R., Henry, L., and Müller, K. (2019a). dplyr: A Grammar
of Data Manipulation. https://dplyr.tidyverse.org/.
Wickham, H., Chang, W., Henry, L., Pedersen, T.L., Takahashi, K., Wilke, C.,
Woo, K., and Yutani, H. (2019b). ggplot2: Create Elegant Data Visualisations
Using the Grammar of Graphics. https://ggplot2.tidyverse.org/reference/
ggplot2-package.html.
1642 Cell 181, 1626–1642, June 25, 2020
Resource
Woroniecka, K., and Fecci, P.E. (2018). T-cell exhaustion in glioblastoma. On-
cotarget 9, 35287–35288.
Xue, J., Schmidt, S.V., Sander, J., Draffehn, A., Krebs, W., Quester, I., De
Nardo, D., Gohel, T.D., Emde, M., Schmidleithner, L., et al. (2014). Transcrip-
tome-based network analysis reveals a spectrum model of human macro-
phage activation. Immunity 40, 274–288.
Zhou, W., Ke, S.Q., Huang, Z., Flavahan, W., Fang, X., Paul, J., Wu, L., Sloan,
A.E., McLendon, R.E., Li, X., et al. (2015). Periostin secreted by glioblastoma
stem cells recruits M2 tumour-associated macrophages and promotes malig-
nant growth. Nat. Cell Biol. 17, 170–182.
Zunder, E.R., Finck, R., Behbehani, G.K., Amir, A.D., Krishnaswamy, S., Gon-
zalez, V.D., Lorang, C.G., Bjornson, Z., Spitzer, M.H., Bodenmiller, B., et al.
(2015). Palladium-based mass tag cell barcoding with a doublet-filtering
scheme and single-cell deconvolution algorithm. Nat. Protoc. 10, 316–333.
 ll
Resource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
Mass cytometry myeloid panel
anti-human Anti-SynCAM (TSLC1/CADM1), purified MBL Life Science Cat# CM004-3; RRID:AB_592783
anti- Biotin (1D4-C5), purified Biolegend Cat# 409002; RRID: AB_10642032
anti-human CCR2 (K036C2), purified Biolegend Cat# 357202;RRID: AB_2561851
anti-human CD100 (A8), purified Biolegend Cat# 328401;RRID: AB_1236386
anti-human CD11b (ICRF44), 209 Bi Fluidigm Cat# 3209003B; RRID: AB_2687654
anti-human CD11c (Bu15), 147-Sm Fluidigm Cat# 3147008B; RRID: AB_2687850
anti-human CD123 (6H6), purified Biolegend Cat# 306002; RRID: AB_314576
anti-human CD14 (M5E2), 160-Gd Fluidigm Cat# 3160001B; RRID: AB_2687634
anti-human CD141 (M80), purified Biolegend Cat# 344102; RRID: AB_2201808
anti-human CD16 (3G8), 165-Ho Fluidigm Cat# 3165001B; RRID: AB_2802109
anti-human CD163 (GHI/61), purified Biolegend Cat# 333602; RRID: AB_1088991
anti-human CD169 (7-239), purified Biolegend Cat# 346002; RRID: AB_2189031
anti-human CD19 (HIB19), 142-Nd Fluidigm Cat# 3142001B; RRID: AB_2651155
anti-human CD1c (L161), purified Biolegend Cat# 331502; RRID: AB_1088995
anti-human CD206 (15-2), purified Biolegend Cat# 321102; RRID: AB_571923
anti-human CD209 (DCN46), purified BD Cat# 551186; RRID: AB_394087
anti-human CD3 (UCHT1), 170-Er Fluidigm Cat# 3170001B; RRID: AB_2661807
anti-human CD33 (WM53), 169-Tm Fluidigm Cat# 3169010B; RRID: AB_2802111
anti-human CD38 (HIT2), 167-Er Fluidigm Cat# 3167001B; RRID: AB_2802110
anti-human CD45 (HI30), 89-Y Fluidigm Cat#3089003B; RRID: AB_2661851
anti-human CD45RA (HI100), 153-Eu Fluidigm Cat# 3153001B; RRID: AB_2802108
anti-human CD49d (9F10), 141-Pr Fluidigm Cat# 3141004B; RRID: N/A
anti-human CD5 (UCHT2), purified Biolegend Cat# 300602; RRID: AB_314088
anti-human CD56 (NCAM16), purified BD Cat# 559043, RRID: AB_397180
anti-human CD64 (10.1), 146-Nd Fluidigm Cat# 3146006B; RRID: AB_2661790
anti-human CD66b (80H3), 152-Sm Fluidigm Cat# 3152011B; RRID: AB_2661795
anti-human CD68 (Y1/82A), purified Biolegend Cat# 333802; RRID: AB_1089058
anti-human CD86 (IT2.2), 156-Gd Fluidigm Cat# 3156008B; RRID: AB_2661798
anti-human CD88 (S5/2), purified Biolegend Cat# 344302; RRID: AB_2259318
anti-human CX3CR1 (2A9-1), purified Biolegend Cat# 341602; RRID: AB_1595422
anti-human FCER1 (AER-37 (CRA1)), biotin eBioscience Cat# 13-5899-82; RRID: AB_466786
anti-human HLA-DR (TU36), 150-Nd Fluidigm Cat# 3150028B; RRID: N/A
anti-human ICAM-1(HA58), purified Biolegend Cat#353102 ; RRID: AB_11204426
anti-human LAP (TW4-2F8), purified Biolegend Cat# 349602; RRID: AB_10645476
anti-human MERTK (125518), purified R&D Cat# MAB8912; RRID: AB_2143588
anti-human PD-1 (EH12.2H7), 174-Yb Fluidigm Cat# 3174020B; RRID: N/A
anti-human PD-L1 (29E.2A3),175-Lu Fluidigm Cat# 3175017B; RRID: AB_2687638
anti-human PD-L2 (24F.10C12), purified Biolegend Cat# 329602; RRID: AB_1089010
anti-human SOX2 (14A6A34), purified Biolegend Cat# 656102 ; RRID: AB_2562246
Mass cytometry lymphoid panel
anti-human 2B4 (C1.7), purified Biolegend Cat# 329502, RRID: AB_1279194
anti-Biotin (1D4-C5), purified Biolegend Cat# 409002; RRID: AB_10642032
(Continued on next page)

Cell 181, 1626–1642.e1–e11, June 25, 2020 e1

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
anti-human CCR2 (K036C2), purified Biolegend Cat# 357202;RRID: AB_2561851
anti-human CCR4 (205410), Gd-158 Fluidigm Cat# 3158006A, RRID: AB_2687647
anti-human CCR6 (11A9), Pr-141 Fluidigm Cat# 3141014A, RRID: N/A
anti-human CCR7 (G043H7), Er-167 Fluidigm Cat# 3167009A, RRID: N/A
anti-human CD103 (Ber-ACT8), Eu-151 Fluidigm Cat# 3151011B, RRID: AB_2756418
anti-human CD127 (A019D5), Ho-165 Fluidigm Cat# 3165008B, RRID: N/A
anti-human CD137 (4B4-1), purified Biolegend Cat# 309802, RRID: AB_314781
anti-human CD16 (3G8), Bi-209 Fluidigm Cat# 3209002B, RRID: AB_2756431
anti-human CD19 (HIB19), Nd-142 Fluidigm Cat# 3142001B, RRID: AB_2651155
anti-human CD25 (2A3), Sm-149 Fluidigm Cat# 3149010B, RRID: AB_2756416
anti-human CD27 (L127), Gd-155 Fluidigm Cat# 3155001B, RRID: AB_2687645
anti-human CD28 (CD28.2), purified Biolegend Cat# 302902, RRID: AB_314304
anti-human CD3 (UCHT1), Sm-154 Fluidigm Cat# 3154003B, RRID: AB_2687853
anti-human CD38 (HIT2), purified Biolegend Cat# 303502, RRID: AB_314354
anti-human CD4 (RPA-T4), Nd-145 Fluidigm Cat# 3145001B, RRID: AB_2661789
anti-human CD45 (HI30)purified Biolegend Cat# 304002, RRID: AB_314390
anti-human CD45RA (HI100), Eu-153 Fluidigm Cat# 3153001B, RRID: AB_2802108
anti-human CD45RO (UCHL1), Dy-164 Fluidigm Cat# 3164007B, RRID: AB_2811092
anti-human CD45RO (UCHL1), purified Biolegend Cat# 304202, RRID: AB_314418
anti-human CD56 (NCAM16), purified BD Cat# 559043, RRID: AB_397180
anti-human CD57 (hCD57), Yb-172 Fluidigm Cat# 3172009B, RRID: N/A
anti-human CD69 (FN50), Nd-144 Fluidigm Cat# 3144018, RRID: AB_2687849
anti-human CD8 (RPA-T8), Nd-146 Fluidigm Cat# 3146001B, RRID: AB_2687641
anti-human CD90 (5E10), Tb-159 Fluidigm Cat# 3159007B, RRID: N/A
anti-human CD95 (DX2), Dy-164 Fluidigm Cat# 3164008B, RRID: N/A
anti-human CRTH2 (BM16), purified Biolegend Cat# 350102, RRID: AB_10639863
anti-human CTLA-4 (14D3), Dy-161 Fluidigm Cat# 3161004B, RRID: AB_2687649
anti-human CXCR3 (G025H7), Gd-156 Fluidigm Cat# 3156004B, RRID: AB_2687646
anti-human Granzyme B (GB11), Yb-171 Fluidigm Cat# 3171002B, RRID: AB_2687652
anti-human HLA-DR (TU36), Nd-150 Fluidigm Cat# 3150028B, RRID: N/A
anti-human ICOS (C398.4A), purified Biolegend Cat# 313502, RRID: AB_416326
anti-human Ki-67 (B56), Er-168 Fluidigm Cat# 3168007B, RRID: AB_2800467
anti-human KLRG1 (14C2A07), purified Biolegend Cat# 368602, RRID: AB_2566256
anti-human LAG-3 (11C3C65), Er-168 Fluidigm Cat# 3165037B, RRID: AB_2810971
anti-human Nkp44 (P44-8), purified Biolegend Cat# 325102, RRID: AB_756094
anti-human CD134 (OX-40; Ber-ACT35 (ACT35)), purified Biolegend Cat# 350002, RRID: AB_10639951
anti-human PD-1 (EH12.2H7), Lu-175 Fluidigm Cat# 3175008, RRID: AB_2687629
anti-human SOX2 (14A6A34), purified Biolegend Cat# 656102 ; RRID: AB_2562246
anti-human TCRgd (11F8), Sm-152 Fluidigm Cat# 3152008B, RRID: AB_2687643
anti-human TIGIT (MBSA43), Tm-169 eBioscience Cat# 16-9500-82, RRID: AB_10718831
anti-human Tim3 (F38-2E2), Biotin Miltenyi Cat# 130-098-945, RRID: N/A
anti-human Tim3 (F38-2E2), VioBright FITC Miltenyi Cat# 130-104-646, RRID: N/A
Flow cytometry panel, human brain samples
anti-human CCR2 (K036C2), BV605 Biolegend Cat# 357213; RRID:AB_2562702
anti-human CD11b (ICRF-44), FITC Biolegend Cat# 301330 ; RRID: AB_2561703
anti-human CD11c (B-ly6), BV570 BD Cat# 624298; RRID: N/A
anti-human CD123 (6H6), BV711 Biolegend Cat# 306029; RRID: AB_2566353
anti-human CD14 (6H6), BUV737 BD Cat# 564444; RRID:AB_2744285
(Continued on next page)

e2 Cell 181, 1626–1642.e1–e11, June 25, 2020

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
anti-human CD141 (1A4), BB700 BD Cat# 742245; RRID:AB_2740668
anti-human CD16 (3G8), BUV496 BD Cat# 321102; RRID: N/A
anti-human CD163 (GHI/61), BV650 BD Cat# 563888; RRID:AB_2738468
anti-human CD169 (7-239), PE/Dazzle 594 Biolegend Cat# 346015; RRID:AB_2750265
anti-human CD19 (REA675), APC-Vio770 Myltenyi Cat# 130-110-252; RRID: N/A
anti-human CD1c (F10/21A3), BB660 BD Cat# 624295; RRID: N/A
anti-human CD206 (15-2), BV421 Biolegend Cat# 321126; RRID: AB_2563839
anti-human CD235A (REA175), APC-Vio770 Myltenyi Cat# 130-100-264; RRID:AB_2656505
anti-human CD273 (MIH18), AF700 BD Cat# 565189; RRID:AB_2739102
anti-human CD274 (MIH1), PE-Cy7 BD Cat# 558017; RRID:AB_396986
anti-human CD3(REA613), APC-Vio770 Myltenyi Cat# 130-109-543; RRID:AB_2657072
anti-human CD33 (WM53), BUV395 BD Cat# 740293; RRID:AB_2740032
anti-human CD38 (HIT2), BV785 Biolegend Cat# 303529; RRID: AB_2561368
anti-human CD45 (HI-30), BUV805 BD Cat# 564915; RRID: N/A
anti-human CD45RA (MEM-56),Pe-Cy5.5 LifeTechnologies, Thermo Cat# MHCD45RA18; RRID:AB_10372221
Fisher Scientific
anti-human CD49d (9F10), Biotin Biolegend Cat# 304334; RRID: AB_2749896
anti-human CD56 (HCD56), APC-C7 Biolegend Cat# 318332; RRID: AB_10896424
anti-human CD66b (G10F5), BB790-P BD Customized
anti-human CD86 (IT2.2), PE-Cy5 Biolegend Cat# 305407; RRID: AB_314527
anti-human CD235A (REA175), APC-Vio770 Myltenyi Cat# 130-100-264; RRID:AB_2656505
anti-human CD273 (MIH18), AF700 BD Cat# 565189; RRID:AB_2739102
anti-human CX3CR1 (2A9-1), BV480 BD Cat# 746723; RRID:AB_2743987
anti-human HLA-DR (G46-6), BUV661 BD Cat# 565073; RRID:AB_2722500
anti-human MERTK (125518), APC R&D Cat# FAB8912A; AB_357213
anti-human P2Y12 (S16001E), PE Biolegend Cat# 392104; RRID: AB_2716007
Flow cytometry panel, mouse brain samples
anti-mouse MHCII (M5/114.15.2), BB700 BD Cat# 746197; RRID:AB_2743544
anti-mouse CCR2 (SA203G11), BV650 BD Cat# 747968; RRID: N/A
anti-mouse CD11b (M1/70), BUV737 BD Cat# 564443; RRID:AB_2738811
anti-mouse CD11c (N418), BV570 Biolegend Cat# 117331; RRID: AB_10900261
anti-mouse CD163 (TNKUPJ), biotin eBioscience Cat# 14-1631-82; RRID:AB_2716934
anti-mouse CD206 (C068C2), AF700 Biolegend Cat# 141734; RRID: AB_2629637
anti-mouse CD209a (MMD3), PE Biolegend Cat# 833004; RRID: AB_2721637
anti-mouse CD38 (90), PE-Dazzle594 Biolegend Cat# 102729; RRID: AB_2632890
anti-mouse CD45 (30-F11), BUV395 BD Cat# 564279; RRID:AB_2651134
anti-mouse CD49d (R1-2), PE-Cy7 BioLegend Cat# 103618; RRID:AB_2563700
anti-mouse CD64 (X54-5/7.1), BV421 Biolegend Cat# 139309; RRID: AB_2562694
anti-mouse CX3CR1 (SA011F11), BV510 Biolegend Cat# 149025; RRID: AB_2565707
anti-mouse F4/80 (BM8), PE-Cy5 Biolegend Cat# 123112; RRID: AB_893482
anti-mouse Ly6C (HK1.4), BV711 Biolegend Cat# 128037; RRID: AB_2562630
anti-mouse Ly6G (1A8), BUV 563 BD Cat# 565707; RRID:AB_2739334
anti-mouse MerTK (DS5MMER), SuperBright 780 eBioscience Cat# 78-5751-82; RRID: AB_2762814
anti-mouse PD-L1 (10F.9G2), BV605 Biolegend Cat# 124321; RRID: AB_2563635
anti-mouse Siglec1 (3D6.112), APC Biolegend Cat# 142417; RRID: AB_2565640
Streptavidin, BUV661 BD Customized
anti-mouse CD64 (X54-5/7.1), BV421 Biolegend Cat# 139309; RRID: AB_2562694
anti-mouse CX3CR1 (SA011F11), BV510 Biolegend Cat# 149025; RRID: AB_2565707
(Continued on next page)

Cell 181, 1626–1642.e1–e11, June 25, 2020 e3

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Immunohistochemistry
CD163 (163C01/10D6) NeoMarkers / Lab Vision Cat# MS-1103-S, RRID:AB_64138
Corporation
Iba1 Wako Chemicals Cat# 019-19741, RRID:AB_839504
Immunofluorescence
VE-cadherin (C-19) (polyclonal) Santa Cruz Cat# sc-6458, RRID:AB_2077955
anti-human P2RY12 (S16001E), PE Biolegend Cat# 392104; RRID:AB_2716007
anti-human CD163 (GHI/61), PE Biolegend Cat# 333606; RRID:AB_1134002
anti-human CD169 (7-239), PE-Dazzle594 Biolegend Cat# 346015; RRID:AB_2750265
anti-human CD206 (15-2), PE-Dazzle594 Biolegend Cat# 321106; RRID:AB_571911
anti-human CD209 (DCS-8C1), PE Biolegend Cat# 343004; RRID:AB_2074328
Biological Samples
Brain tumor (glioma and brain metastases) University Hospital Zurich N/A
and non-tumor brain tissue (epilepsy) samples
Chemicals, Peptides, and Recombinant Proteins
16% Paraformaldehyde aqueous solution Electron Microscopy Sciences/ Cat#15710; RRID: N/A
LucernaChem
2-methylbutane Sigma Aldrich Cat# M32631-25.L; RRID: N/A
Antibody Stabilizer PBS Candor Bioscience Cat# 131 050; RRID: N/A
Antifading Mounting Medium with DAPI Dianova Cat# SCR-038448; RRID: N/A
Bambanker LubioScience GmbH Cat# 523303 (BB02); RRID: N/A
Benzonase nuclease Sigma-Aldrich Cat# E1014-25KU; RRID: N/A
Bovine Serum Albumin (BSA) Sigma-Aldrich Cat# B2064; RRID: N/A
Bromoacetamidobenzyl-EDTA (BABE) Dojindo Laboratories Cat# B437-10; RRID: N/A
Cell-IDTM Intercalator-Ir Fluidigm Cat# 201192B; RRID: N/A
Cell-ID Cisplatin Fluidigm Cat# 201064; RRID: N/A
CO2-Independent Medium Thermo Fisher Scientific Cat# 18045-070; RRID: N/A
Collagenase from Clostridium histolyticum, type IV Sigma-Aldrich Cat# C5138; RRID: N/A
Cryo Embedding Medium Medite Cat# 41-3011-00; RRID: N/A
Dead Cell Removal Kit Miltenyi Cat# 30-090-101; RRID: N/A
Deoxyribonuclease I from bovine pancreas Sigma-Aldrich Cat# DN25-1G; RRID: N/A
D-Luciferin Perkin Elmer Cat# 122799; RRID: N/A
DMSO Sigma-Aldrich Cat# D2438; RRID: N/A
DMSO; Dimethyl Sulfoxide, anhydrous, 99.7% Fischer Bioreagents Cat# BP231-1; RRID: N/A
EDTA StemCell Technologies, Inc Cat# EDS-100G; RRID: N/A
EQ Four Element Calibration Fluidigm Cat# 201078; RRID: N/A
Formaldehyde 4.0% PanReac Cat# 252931.1211; RRID: N/A
Foxp3 / Transcription Factor Staining Buffer Set eBioscience Cat# 00-5523-00; RRID: N/A
HBSS ThermoFisher Scientific Cat# 14175095; RRID: N/A
Human TruStain FcX Biolegend Cat# 422302; RRID:AB_2818986
Indium (115In) Trace Sciences International N/A
Iridium (191Ir, 193Ir) Fluidigm Cat# 201192A; RRID: N/A
Isoflurane Minrad N/A
Maleimido-mono-amide-DOTA (mDOTA) Macrocyclics Cat# B-272; RRID: N/A
Maxpar X8 Multimetal Labeling Kit Fluidigm Cat# 201300; RRID: N/A
Maxpar Fix and Perm Buffer Fluidigm Cat# 201067; RRID: N/A
Normal goat serum ThermoFisher Scientific Cat# PCN5000; RRID: N/A
Palladium (104Pd,105Pd, 106Pd, 108Pd, 110Pd) Trace Sciences International N/A
(Continued on next page)

e4 Cell 181, 1626–1642.e1–e11, June 25, 2020

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Percoll GE Cat# P4937; RRID: N/A
Percoll Sigma-Aldrich Cat# GE17-0891; RRID: N/A-0
Phosphate-buffered saline Homemade N/A
Phosphate-buffered saline, DPBS, NO CALCIUM, Life Technologies Cat# 14190094; RRID: N/A
NO MAGNESIUM
RPMI 1640 Bioswisstec; Seraglob Cat# M 3413; RRID: N/A
RPMI 1640 Medium, HEPESS, no glutamine Life Technologies Cat# 42401042; RRID: N/A
Saponin Sigma-Aldrich Cat# S7900; RRID: N/A
Sudan black B Sigma-Aldrich Cat# 199664; RRID: N/A
Deposited Data
TCGA-LGG - Harmonized The Cancer Genome Atlas https://portal.gdc.cancer.gov -
mRNA gene quantification
HTSeq - Counts
TCGA-GBM - Legacy The Cancer Genome Atlas https://portal.gdc.cancer.gov/legacy-
archive - mRNA gene expression and
quantification HT_HG-U133A
TCGA-GBM - Harmonized The Cancer Genome Atlas https://portal.gdc.cancer.gov -
mRNA gene quantification
HTSeq - Counts
Mass- and flow cytometry data this study https://data.mendeley.com/datasets/
jk8c3c3nmz/draft?a=c0a9d8dc-
8ac2-4942-baf9-208de7a8c310
Experimental Models: Cell Lines
GL-261 cells A. Fontana, Experimental N/A
Immunology, University of Zurich,
Zurich, Switzerland
Experimental Models: Organisms/Strains
Sall1CreER/+ Ryuchi Nishinakamura RRID:MGI:4818961
(Kumamoto University)
R26YFP The Jackson Laboratory RRID:IMSR_JAX:00 6148
Software and Algorithms
MATLAB R2016a N/A https://www.mathworks.com/
Normalizer Finck et al., 2013 https://github.com/nolanlab/bead-
normalization/releases
FlowJo V10.6.1.1 Tree Star https://www.flowjo.com/
R version 3.5 R Development Core Team, https://www.r-project.org/
2008
R Studio RStudio Team, 2015 https://www.rstudio.com/
FlowSOM Van Gassen et al., 2015 https://github.com/SofieVG/FlowSOM
Circlize Gu et al., 2014 https://cran.r-project.org/web/
packages/circlize/index.html
MCF-data-analysis Hartmann et al., 2016 https://github.com/hartmannfj/
MCF-data-analysis
CyTOF workflow Nowicka et al., 2017 https://f1000research.com/articles/6-748#
UMAP Mcinnes et al., 2018 https://github.com/lmcinnes/umap
SCAFFoLD Spitzer et al., 2015 https://github.com/nolanlab/scaffold
VorteX Samusik et al., 2016 https://github.com/nolanlab/vortex/
wiki/Getting-Started
Slingshot Street et al., 2018 https://bioconductor.org/packages/
release/bioc/html/slingshot.html
t-SNE Laurens van der Maaten, 2008 https://github.com/jkrijthe/Rtsne
(Continued on next page)

Cell 181, 1626–1642.e1–e11, June 25, 2020 e5

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
One-SENSE Cheng et al., 2016 N/A
flowStats Hahne et al., 2020 https://www.bioconductor.org/
packages/release/bioc/html/
flowStats.html
pheatmap Kolde, 2019 https://cran.r-project.org/web/
packages/pheatmap/index.html
dplyr Wickham et al., 2019a https://cran.r-project.org/web/
packages/dplyr/index.html
ggplot2 Wickham et al., 2019b https://cran.r-project.org/web/
packages/ggplot2/index.html
gplots Warnes et al., 2019 https://cran.r-project.org/web/
packages/gplots/index.html
Hmisc Harrell, 2020 https://cran.r-project.org/web/
packages/Hmisc/index.html
flowWorkspaceData Finak, 2018 N/A
flowCore Ellis et al., 2019 N/A
TCGAbiolinks Colaprico et al., 2016 https://bioconductor.org/packages/
release/bioc/html/TCGAbiolinks.html
Gephi Bastian et al., 2009 https://gephi.org/
Fiji Schindelin et al., 2012 https://imagej.net/Fiji
Living Image 2.5 Caliper Life Sciences N/A
Leica Bond III N/A https://www.leicabiosystems.com
Adobe Illustrator CS6 Adobe https://www.adobe.com/ch_de/
products/illustrator.html
Other
gentleMACS Octo Dissociator with Heaters Miltenyi Biotec Cat# 130-096-427
Bone wax (Aesculap B. Braun N/A
FACSymphony BD N/A
Gentle MACS C-tubes Miltenyi Biotec Cat# 130-096-334
Hamilton 75N syringe Sigma-Aldrich Cat# 28613-U
Hamilton syringe 5 ml Sigma-Aldrich Cat# 26286
Helious CyTOF2 Fluidigm N/A
Hyrax C60 Cryostat Zeiss N/A
MACS columns MS Miltenyi Biotec Cat# 130-042-201
microinjection pump (UMP-3) World precision Instruments N/A
Olympus IX81 Olympus N/A
Stereotactic frame David Kopf Instruments N/A
Tissue glue Indermil; Henkel N/A
Xenogen IVIS 100 Caliper Life Sciences N/A

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources should be directed to the Lead Contact, Burkhard Becher (becher@immunology.
uzh.ch).

Materials Availability
This study did not generate new unique reagents.

Data and Code Availability

Mass and flow cytometry data generated during this study are available at https://data.mendeley.com/datasets/jk8c3c3nmz/draft?
a=c0a9d8dc-8ac2-4942-baf9-208de7a8c310

e6 Cell 181, 1626–1642.e1–e11, June 25, 2020

 ll
Resource

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human Brain Tissue Samples

Human brain tissue samples were collected in the Department of Neurosurgery at the University Hospital Zurich after written
informed patient consent following the local ethical requirements and the declaration of Helsinki. All analyses were performed ac-
cording to the guidelines of the local ethics committees (KEK-ZH-Nr. 2015-0163). All the collected samples were anonymized before
processing. Clinical information related to patients was collected during diagnosis and summarized in Table S1. A total of 45 patients
(including glioma, brain metastases, and epilepsy cases) from both male and female subjects between the ages of 3 - 80 years old
were included in the present study.

Animal Models
Mice were bred in house (see Key Resources Table): R26YFP (de Boer et al., 2003). Sall1CreER were kindly provided by R. Nishina-
kamura (Kumamoto University) (Inoue et al., 2010; Takasato et al., 2004). All ‘Cre’ and ‘CreER’ strains were used as heterozygotes.
Six-week-old mice (female and male) were used for all experiments of the glioma preclinical model. All animal experiments performed
in this study were approved by the Swiss Veterinary Office. All mice were on a C57BL/6 background and kept in individually ventilated
cages under specific-pathogen-free conditions. Animals were monitored once a week to assess weight loss and physical/neurolog-
ical abnormalities. In vivo measurements to assess tumor growth were also performed once a week.

Cell Lines
GL-261 cells, which are syngeneic in C57BL/6 mice, were stably transfected with pGl3-ctrl and pGK-Puro (Promega) and selected
with puromycin (Sigma-Aldrich) to generate luciferase-stable GL-261 cells. A single clone was isolated by limiting dilution and
passaged in vivo by intracranial tumor inoculation. Subsequently, cells were transfected with pCEP4-mIgG3, pCEP4-mIl-
12mIgG3, or pCEP4-mIl23mIgG3, and cytokine production was detected by ELISA and RT-PCR, as previously described (Eisenring
et al., 2010). SMA-560 spontaneous murine astrocytoma cells were characterized previously (Uhl et al., 2004).

METHOD DETAILS

Processing of Human Samples for Cytometry Analysis

Fresh resected human brain tissue samples were taken within 2 hours to the Laboratory of Molecular Neuro-Oncology to start tissue
dissection and processing. First, tissue samples were thoroughly washed with phosphate-buffered saline (PBS) to remove visible
blood clots and to reduce blood leukocytes contamination. After tissue was minced using scalpels into approximately 3 to 5 mm
diameter pieces and digested (1 mg/mL Collagenase IV (Sigma-Aldrich), 10 mg/mL DNase (Sigma-Aldrich), 10% Fetal Bovine Serum
(FBS) (Thermo Fisher Scientific), RPMI 1640 (Seraglob)) at 37 C for 45 minutes using the gentle MACS Octo Dissociator with Heaters
(Miltenyi Biotech) and continuous shaking. The enzymatic reaction was stopped by adding EDTA 2 mM (StemCell Technologies, Inc)
in PBS to a double volume of the sample. Afterward, the homogenate was filtered through a 100 mm cell strainer and centrifuged at
400 3 g for 8 minutes at 4 C to pellet the cells and myelin. This was followed by myelin removal step by gradient centrifugation with
30% Percoll (Sigma-Aldrich) in PBS (1,592 x g for 30 minutes at 4 C; without brakes during deceleration) using a 50 mL tube with a lid
for a fixed angle rotor fitting in a centrifuge. After myelin (the top white layer) separation, the middle transparent layer without the bot-
tom layer of red blood cells was collected and filtered once more through a 100 mm cell strainer. The single-cell suspension was
washed in PBS and centrifuged at 400 3 g for 8 minutes at 4 C to pellet the cells. Next, cells were ready for flow cytometry analysis
or storing samples by freezing. To freeze cells, we divided the single-cell suspension into two parts. One half was viably frozen in
Bambanker (LubioScience GmbH). The second half of the cell suspension was stained for viability with Cell-ID Cisplatin (Fluidigm)
(3 minutes at 4 C), followed by washing in PBS and centrifugation at 400 3 g for 5 minutes at 4 C to pellet the cells. Next, cells
were fixed with 1.6% Paraformaldehyde aqueous solution (PFA; Electron Microscopy Sciences) for 15 minutes at room temperature
(RT) (Zunder et al., 2015). After fixation, cells were washed twice in PBS, centrifuged at 600 3 g for 5 minutes at 4 C to pellet the cells
and cryopreserved in FACS buffer (EDTA 2mM (StemCell Technologies, Inc.), FBS 0.5% (ThermoFisher Scientific), PBS) comple-
mented with 10% DMSO (Sigma-Aldrich). Cell fixation before cryopreservation allowed us to preserve the myeloid compartment
(including microglia and neutrophils) and cell frequencies. Cryopreserved samples were stored at 80 C (maximum 30 days) or in
liquid nitrogen 160 C until analysis.

Tissue collection for Immunofluorescence

In some cases (Table S1), a portion of the tissue was collected for Immunofluorescence. The samples were cut into pieces of max 1 cm x
0.5 cm x 0.5 cm and embedded in cryomolds filled with Cryo Embedding Medium (Medite). Freezing was performed in a beaker filled
with 2-methylbutane (Sigma-Aldrich) and dry ice or on dry ice directly and stored at 80 C until further processing.

Orthotopic Glioma Cell Injection

8–12-week-old mice were anesthetized with 2%–5% Isoflurane (Minrad) in an induction chamber. Anesthesia was maintained on the
stereotactic frame (David Kopf Instruments). A blunt-ended syringe (Hamilton; 75N, 26 s/2’’/2, 5 ml; Sigma-Aldrich) was injected

Cell 181, 1626–1642.e1–e11, June 25, 2020 e7

 ll
Resource

1.5 mm lateral and 1 mm frontal from the bregma. A 5ml syringe (Hamilton; Sigma-Aldrich) was injected with a depth of 4 mm below
the skull and retracted 1 mm, forming a reservoir. Using a microinjection pump (UMP-3; World precision Instruments Inc.), 5 3 104
GL-261 cells were injected in a volume of 2 ml at 1 ml/minute. After resting the needle for 2 minutes, it was retracted at a speed of
1 mm/minute. The injection hole was closed with bone wax (Aesculap; B. Braun), and the scalp wound was sealed with tissue
glue (Indermil; Henkel).
Every animal got an ID code. ‘‘C’’ in the animal ID marked the control animals, the one did not have the injection of tumor cells, but
had Sall1/YFP expression. Samples LH60, LV57, L57, RV58, RH57, RH58 and RV57 developed a big tumor, and the rest had an in-
termediate to a small tumor. Samples RV58, LH60, L57 and RH60 were wild-type control animals (YFP-) with the injection of tu-
mor cells.

In Vivo Bioluminescent Imaging

Tumor-bearing mice were injected with D-Luciferin (150 mg/kg body weight; Perkin Elmer). Animals were transferred to the dark
chamber of a Xenogen IVIS 100 (Caliper Life Sciences) imaging system and luminescence was detected. Data were subsequently
analyzed using Living Image 2.5 software (Caliper Life Sciences).

Harvesting and Processing of Mouse Brain Samples

Mice were euthanized with CO2 and perfused with PBS through the left ventricle of the heart using a 25-G butterfly needle attached to
a 50 mL syringe (Mrdjen et al., 2017). The collected complete brain sample was dissected into approximately 1 to 3 mm diameter
pieces using scissors and digested (0.4 mg/mL collagenase IV (Sigma-Aldrich), 10 mg/mL Deoxyribonuclease I from (DNase;
Sigma-Aldrich), 10% FBS (Thermo Fisher Scientific), RPMI 1640 (Seraglob)) at 37 C for 30 minutes applying continuous shaking.
The enzymatic reaction was stopped by adding EDTA (StemCell Technologies, Inc) in PBS to a final concentration 5 mM. To homog-
enize the sample, it was repeatedly aspirated and ejected using a 5 mL syringe with a 20-G needle until a uniform homogenate was
formed (Mrdjen et al., 2017). Afterward, the homogenate was filtered through a 70 mm cell strainer and centrifuged at 400 3 g for
8 minutes at 4 C to pellet the cells and myelin. This was followed by myelin removal step by gradient centrifugation with 30% Percoll
(Sigma-Aldrich) in PBS (1,592 x g for 30 minutes at 4 C; without brakes during deceleration) using a 50 mL tube with a lid for a fixed
angle rotor fitting in a centrifuge. After myelin (the top white layer) separation, the middle transparent layer without the bottom layer of
red blood cells was collected and filtered through a 70 mm cell strainer. The single-cell suspension was washed in PBS and centri-
fuged at 400 3 g for 8 minutes at 4 C to pellet the cells. Cells were then ready for flow cytometry analysis.

Mass-Tag Cellular Barcoding

To minimize inter-sample staining variation, we applied two strategies of mass-tag barcoding to fixed (myeloid focused panel) (Zun-
der et al., 2015) and viably frozen cells (lymphoid focused panel) (Mei et al., 2016).
Intracellular (fixed cell) barcoding consisted of a ten sample barcoding scheme composed with unique combinations of three out of
five palladium metals (104Pd, 105Pd, 106Pd, 108Pd, 110Pd) (Trace Sciences International). Palladium isotopes were conjugated to Bro-
moacetamidobenzyl-EDTA (Dojindo Laboratories). The concentrations were adjusted to 100 nM for all metals. After thawing
samples in a cold water bath, cells were washed once with Cell Staining Medium (CSM: PBS, 0.5% Bovine Serum Albumin (BSA)
(Sigma-Aldrich) and 0.02%NaN3), once with PBS and once with 0.03% Saponin (Sigma-Aldrich) in PBS and centrifuged at 600 3
g for 5 minutes at 4 C to pellet the cells. After the last washing step, 100 mL residual volume was left in each tube. Next, 1x1000 diluted
barcoding reagent in PBS with 0.03% saponin was thoroughly and quickly mixed with the sample and incubated at RT for 15 minutes.
Cells were then washed three times in CSM, and ten samples were pooled for antibody surface staining.
Live cell (viable frozen) barcoding comprised a twenty sample barcoding scheme with unique combinations of three out of five
palladium metals (104Pd, 105Pd, 106Pd, 108Pd, 110Pd) (Trace Sciences International) and one Yttrium (preconjugated anti-human
CD45 89Y) (Fluidigm). Palladium metals were conjugated with purified anti-human CD45 antibody (Biolegend) in house using the
Maxpar X8 chelating polymer kit (Fluidigm) according to the manufacturer’s instructions. Viable frozen cells were thawed in a water
bath at 37 C for 2 minutes. Cells were immediately resuspended in CO2-Independent Medium (Thermo Fisher Scientific) comple-
mented with 5% FBS (ThermoFisher Scientific) + Benzonase Nuclease (100 000 x) (Sigma-Aldrich) and centrifuged 400 3 g for 7 mi-
nutes at RT to pellet the cells. Cells were resuspended in RPMI-1640 complemented with 5% FBS and proceeded with a protocol of
dead cells were removal by negative selection using MACS columns MS (Miltenyi Biotech) and dead cell removal kit (Miltenyi Biotech)
according to the manufacturer’s instructions. Next, cells were incubated on ice in RPMI-1640 complemented with 5% FCS and a
unique combination of metal-tagged CD45 antibodies. Samples were then washed twice in PBS, and twenty samples were pooled
for antibody surface staining.

Metal-Isotope-Tagged Antibodies
All anti-human antibodies, corresponding clone, and tagged metal isotope for mass cytometry analysis are listed in Tables S2 and S3
and Key Resources Table. Preconjugated antibodies to metal isotope were purchased from Fluidigm or commercial suppliers in
purified form and conjugated in house using the Maxpar X8 chelating polymer kit (Fluidigm) according to the manufacturer’s
instructions.

e8 Cell 181, 1626–1642.e1–e11, June 25, 2020

 ll
Resource

Cell Surface Staining for Cytometry

To avoid nonspecific binding of antibodies, the sample was incubated at 4 C for 15 minutes in Human TruStain FcX (Fc Receptor
Blocking Solution; Biolegend). Without washing, the cells were spun down, resuspended in the antibody mixture (Key Resources
Table) in PBS, and incubated at 4 C for 30 minutes. To optimize antibody staining for chemokine receptors, see Table S3, the sample
was incubated at 37 C first 10 minutes, and at 4 C followed 20 minutes. After staining the lymphoid focused panel (Table S3), we
added Cell-ID Cisplatin (Fluidigm) to the sample for 3 minutes to discriminate viable/dead cells. Then, the sample was washed
once in PBS and centrifuged to pellet the cells.

Intracellular Cytokine Staining for Cytometry

Cells were permeabilized using Foxp3/Transcription Factor Staining Buffer Set (eBioscience) according to the manufacturer’s in-
structions for 45 minutes at 4 C. Subsequently, the sample was washed once in Perm/Wash buffer (eBioscience) and incubated
in the antibody mixture (Key Resources Table) in Perm/Wash buffer (including transcriptional factors, intracellular antigens, anti-
streptavidin, anti-PE) for 30 minutes at 4 C. The sample was washed once in Perm/Wash buffer (eBioscience) and centrifuged to pel-
let the cells. The cells were ready for flow cytometry acquisition or processed before CyTOF acquisition.

Cell Preparation and Mass Cytometry Acquisition

After cell surface and intracellular antibody staining the cells were incubated in 4% PFA (Electron Microscopy Sciences) overnight.
Prior to acquisition the cells were pelleted without washing and resuspended in up to 1 mL of diluted 1:3000 Cell-IDTM Intercalator-Ir
(Fluidigm) + Maxpar Fix and Perm Buffer (Fluidigm) for 1.5-3 hours. After the sample was washed twice in PBS and twice in ddH2O,
diluted to 1.5–106 cells/ml in ddH2O containing 10% EQ Four Element Calibration Beads (Fluidigm) and filtered through a 40-mm filter-
cap FACS tube. Samples were analyzed with a Helious CyTOF2 (Fluidigm). Quality control and tuning processes on the Helios
CyTOF2 (Fluidigm) were performed following the guidelines for the daily instrument operation. Data were collected as FCS files.

Flow Cytometry Acquisition

Samples stained with antibodies coupled to fluorophores (Key Resources Table) were analyzed with a FACSymphony (BD Biosci-
ences). Data were collected as FCS files. Before acquisition, PMT voltages were adjusted using single-stained samples and
controlled using the fully stained sample. FCS files were named accordingly the patient ID or animal ID.

Immunohistochemistry
Immunohistochemistry was performed on 4-mm-thick tissue sections. Double immunostaining for Iba1 (Wako Chemicals USA, 019-
19741), pretreatment with TrisEDTABorat antigen retrieval, 32 minutes, followed by 32 minutes incubation, dilution 1/1000, OptiView
DAB detection kit (Ventana) was performed on an automated Ventana Benchmark Ultra, and CD163 (NeoMarkers / Lab Vision Cor-
poration, MS-1103-S), prediluted, incubation 32 minutes, Bond Polymer Refine Detection kit, was performed on a Leica Bond III. In
total we analyzed 27 glioma, 7 BrM and 2 epilepsy samples from the CyTOF cohort.

Immunofluorescence
Frozen tissues were cryosectioned (10-mm thick) using a Hyrax C60 Cryostat (Zeiss) and stored at 20  C. Sections were fixed with
4% PFA (PanReac), washed in PBS, and incubated with a blocking solution consisting of PBS supplemented with 1% BSA (Sigma-
Aldrich) and 0.3% Triton X-100 (Sigma-Aldrich). Subsequently, sections were incubated with the following primary antibodies (diluted
in blocking solution) at 4  C overnight: rabbit anti-Iba1 antibody (Wako; polyclonal; 1:500); goat anti-VE-cadherin antibody (Santa
Cruz; polyclonal; 1:00). Sections were then washed with blocking solution and incubated with AF647-labeled donkey anti-rabbit,
AF488-labeled donkey anti-goat secondary antibodies (Life Technologies; 1:500) and one of the following directly-labeled antibodies
(all diluted in blocking solution) at room temperature for 2h: P2RY12-PE (Biolegend; clone S16001E; 1:100), CD163-PE (Biolegend;
clone GHI/61; 1:200), CD169-PE (Biolegend; clone 7-239; 1:100), CD206-PE-Dazzle594 (Biolegend, clone 15-2; 1:100) or CD209-PE
(Biolegend; clone DCS-8C1; 1:100). Sections were washed with blocking solution and incubated in Sudan black B (Sigma-Aldrich)
dissolved in 70% ethanol to reduce autofluorescence of the tissues at RT for 30 minutes. Finally, sections were washed with HBSS
(ThermoFisher Scientific) and mounted with Immunoselect Anti fading Mounting Medium with DAPI (Dianova). Fluorescence photo-
micrographs were captured with an Olympus IX81 microscope using the 40 3 objective and images were processed with Fiji soft-
ware (Schindelin et al., 2012). A Gaussian filter s = 1 was applied to the P2RY12 image before merging with the other channels.

QUANTIFICATION AND STATISTICAL ANALYSIS

Preprocessing of Cytometry Data

Raw mass cytometry data were normalized using the MATLAB version of the Normalizer tool (Finck et al., 2013). Cells were
assigned by manually gating on Event length and DNA (191Ir and 193Ir) channels, followed by the dead cell discrimination analyzing
195
Pt expression using FlowJo Software (Tree Star). Doublets were excluded using Gaussian discrimination channels (Center, Offset,
Width, Residual). Next, data were concatenated and de-barcoded using Boolean gating in FlowJo software (Tree Star). The normal-
ized data containing living cells from every individual patient were manually exported from FlowJo Software (Tree Star) and imported

Cell 181, 1626–1642.e1–e11, June 25, 2020 e9

 ll
Resource

into R studio of R using the R packages ‘‘flowCore’’ and ‘‘flowWorkspaceData’’ (R Foundation for Statistical Computing) (Ellis et al.,
2019; Finak, 2018). Before automated high-dimensional data analysis, the mass cytometry data were transformed with a cofactor in
the range of 5 and 60 using an inverse hyperbolic sine (arcsinh) function (Bendall et al., 2011).
For flow cytometry data, the compensation matrix was corrected using FlowJo software (Tree Star). After live, single, CD45
positive and compensated cells were exported and imported into R Studio. Before automated high-dimensional data analysis,
flow cytometry data were transformed using an inverse hyperbolic sine (arcsinh) function with a cofactor in the range of between
300 and 600.
Additionally, all cytometry data were normalized between 0 and 1 to the 99-999th percentile of the merged sample in each
batch. To control for the batch effect, we used the same clinical sample in two acquisition rounds. For the mass cytometry data,
the marker expression distributions were verified between two batches of the acquisition applying R package ‘‘flowStats’’ (Hahne
et al., 2020).

Automated Population Identification

To identify myeloid and lymphoid immune cell populations accurately, we first carried out a step of FlowSOM clustering to generate a
starting point of 100 nodes, on pre-processed and combined mass/ or flow cytometry datasets (Van Gassen et al., 2015; Hartmann
et al., 2016). This was then followed by expert-guided manual metaclustering, using parameters listed in each figure legend. The
respective k-value was manually chosen (in the range of between 20 and 30); identified clusters were annotated and merged based
on a similarity of antigen expression in order to uphold the biological relevance of the dataset. Manually-annotated clusters were used
to calculate the relative frequencies of immune populations. Heatmaps display median expression levels of all markers per merged
population and plotted using the R package ‘‘pheatmap’’ (Figures 1D, 2B, 3E, and S5B) (Kolde, 2019). From both mass cytometry
datasets, we pre-selected major populations and performed additional FlowSOM analysis to identify smaller cell subsets. For Figures
3B, 7D, and S7A, we calculated the median antigen expression among selected cell types of the second mass cytometry batch using
the R package ‘‘dplyr’’ (Wickham et al., 2019a).
For data visualization, we applied various dimensionality reduction techniques. For a complex overview of the immune compart-
ment, we used Uniform Manifold Approximation and Projection (UMAP) (Mcinnes et al., 2018). To create a UMAP of isolated leuko-
cytes (Figures 1C and S2A), we pooled equally proportioned 120,000 CD45+ cells from the glioma and BrM datasets from the second
CyTOF batch. To create a UMAP of lymphocytes (Figures S5A, S5C), we pooled equally proportioned 80,000 CD45highHLA-DR-lin+
cells of glioma and BrM datasets from the second CyTOF batch. To separate populations with a similar phenotypic spectrum, we
generated a force-directed layout integrating the SCAFFoLD algorithm (Figure 2) (Spitzer et al., 2015) and VorteX graphical environ-
ment (Figures 3C and 3D) (Samusik et al., 2016). To define landmark populations for the SCAFFoLD layout, we analyzed a combined
dataset of TAMs/monocytes of the second CyTOF batch and defined three myeloid populations (CNS-resident microglia, CNS-
invading MDMs and monocytes) using FlowSOM as described above. Each SCAFFoLD map was generated using 100 FlowSOM
initial nodes. The subsequent results were saved as GRAPHML files, which were later displayed with the open graph viz platform
Gephi (Bastian et al., 2009). SCAFFoLD maps allowed us to align and compare mass and flow cytometry, human and mouse flow
cytometry data. The force-directed graph (x-shift Vortex, Figure 3C, 3D) includes pooled 80,000 MDMs/monocytes from glioma
and BrM. The cell position matrix generated using VorteX was exported as CSV file and then imported into the R environment to over-
lay calculated FlowSOM results, using the R package ‘‘ggplot2’’ (Wickham et al., 2019b). Next, we carried out a pseudo-time align-
ment analysis using markers listed in Figure 3E and cell positions calculated using Vortex, to reveal the trifurcation of cellular subsets
from a monocytic phenotype toward a macrophage one (Street et al., 2018). Categorical ONE-Sense analysis generated one-dimen-
sional t-SNE (Laurens van der Maaten and Hinton, 2008) of equally pooled 64,000 CD8 T cells (Figure 6B) or 36,000 NK cells (Fig-
ure 7C) from glioma and BrM of the second CyTOF batch, where axis was calculated using lineage or activation markers. The
one-dimensional t-SNEs were aligned with two heatmaps, displaying lineage or activation cell profile using the R package ‘‘gplots’’
(Warnes et al., 2019).

Survival Analysis
For overall Survival (OS) data, we used The Cancer Genome Atlas (TCGA) from 164 (harmonized data, RNaseq) and 558 (legacy data,
Affymetrix) glioblastoma samples (TCGA-GBM). In order to analyze OS in low grade gliomas (TCGA-LGG), we used 516 samples
(harmonized data, RNaseq). To obtain the signature of genes associated with outcome, data were extracted from TCGAbiolinks
Rstudio and the gene list obtained from CyTOF. For harmonized data (TCGA-GBM/LGG), genes were plotted as a heatmap, and
selected according to high or low expression (gene groups). For legacy data, median levels were used to segregate cancer patients
according to OS outcome.

Statistical Analysis
P values were calculated to compare the relative frequencies of leukocytes or median antigen expression of an immune cell types
between glioma IDH1mut versus IDH1wt versus melanoma BrM versus carcinoma BrM using nonparametric Mann-Whitney-Wilcoxon
tests, and controlled for multiple testing by using the Benjamini-Hochberg test (Field et al., 2013; Noble, 2009). The relative fre-
quencies of individual patients’ leukocyte subsets from the first and second mass cytometry batches were combined to perform
the statistical analysis. Reported p-values were below 0.05, considered statistically significant and displayed on the corresponding

e10 Cell 181, 1626–1642.e1–e11, June 25, 2020

 ll
Resource

graph. In Figures 1E, S2C, S3B, and S5E error bars define an interval of max/min value ± SD, horizontal line indicates the mean value.
In Figures 2G, 6A, 7B, 7D, S5D, S6B, and S7A–S7C boxplots represent the interquartile range (IQR) 50% and whiskers 25%. Listed
figures showing the relative frequencies of leukocytes or median antigen expression of an immune cell types generated using the R
package ggplot2 (Wickham et al., 2019b). The Pearson’s correlation matrix between the relative frequencies of immune populations
was calculated with the R environment (‘‘Hmisc’’ R package) and include p-value (p) and correlation coefficients (R) (Harrell, 2020).
Correlations were considered statistically significant if the p-value was below 0.05 and R value was below 0.6 or above 0.6 and
visualized using the R package ‘‘circlize’’ (Gu et al., 2014).

Cell 181, 1626–1642.e1–e11, June 25, 2020 e11

 ll
Resource

Supplemental Figures

Figure S1. Mass Cytometry Analysis Reveals Unique Changes in the Leukocyte Composition of Brain Tumors, Related to Figure 1
Heatmap of the mean expression of all markers in the myeloid panel, calculated across CD45+ leukocytes (value range 0-1 post-transformation/normalization).
 ll
Resource

Figure S2. Mass Cytometry Analysis Reveals Unique Changes in the Leukocyte Composition of Brain Tumors, Related to Figure 1
(A) Individual UMAP plots are overlaid with all markers from the myeloid panel. (B) Composition of major immune populations found in the brain samples. Selected
bars show the reference sample included in both CyTOF acquisition rounds, which allowed normalization and batch correction. Samples were ordered according
to patients’ clinical diagnosis. (C) Frequencies of the main immune populations among CD45+ cells in the glioblastoma cohort stratified according to methylation
status of the MGMT promoter. Error bars define an interval of max/min value ± SD, horizontal line indicates the mean value. P-values were calculated using a non-
parametric Mann-Whitney-Wilcoxon test. P-values of less than 0.05 were considered statistically non-significant and were not displayed.
 ll
Resource

Figure S3. The Brain TME Harbors a Heterogeneous Mononuclear Phagocyte Population, Related to Figure 2
(A) Manual gating strategy to validate the identification of the major immune populations by unsupervised machine-learning algorithms. (B) Relative frequencies of
three TAM populations among CD64+ cells in the glioblastoma cohort stratified according to methylation status of the MGMT promoter. Error bars define an
interval of max/min value ± SD, horizontal line indicates the mean value. P-values were calculated using a non-parametric Mann-Whitney-Wilcoxon test. P-values
of less than 0.05 were considered statistically non-significant and were not displayed.
 ll
Resource

Figure S4. TAM Instruction Is Driven by the Type of Tumor Rather Than the Local Tissue Microenvironment, Related to Figures 4 and 5
(A) Representative immunohistochemistry images of CD163 (brown) and Iba1 (red) co-staining for glioma and BrM samples. Selected areas showing Iba1+
CD163- microglia with amoeboid (1) and ramified (2) morphology. Errors depicting the blood vessel. (B) Representative immunofluorescence image of CD209+
CNS phagocytes in recurrent anaplastic oligodendroglioma (ZH927). (C) Representative immunofluorescence image of CD169+ CNS phagocytes in the

(legend continued on next page)

 ll
Resource

melanoma BrM (ZH879). (D) Representative immunofluorescence images of CD163+, CD206+ and CD169+ CNS phagocytes in NSCLC BrM (ZH968). (E) The
Pearson correlation between TAM frequencies and OS in the IDH1wt glioma group. Relative frequencies of microglia were calculated among CD64+ cells. Relative
frequencies of monocytes and MDM subsets were calculated among parental MDM/monocyte population. Patients that have survived till day of analysis (hence
past 400 days) have been highlighted using an asterisk (*) alongside their Patient ID.
 ll
Resource

Figure S5. Preferential Treg Accumulation in the TME of Brain Metastases, Related to Figure 6
(A) A representative UMAP map displaying 120,000 singlets, live, TILs (CD3, CD19, CD56, CD90 expressing cells) equally proportioned from glioma and
metastasis samples. Individual UMAP plots are overlaid with all markers in the lymphoid panel. (B) Heatmap displaying the median antigen intensity of markers

(legend continued on next page)

 ll
Resource

used to generate part (C). (C) UMAP map overlaid with FlowSOM-guided manual meta-clusters. (D) Relative frequencies of the main TILs populations identified
among lymphocytes in the brain tumors. (E) Relative frequencies of the main TILs populations identified in patients with glioblastoma stratified according to
methylation status of the MGMT promoter. (D, E) Only statistically significant p-values were displayed (p < 0.05, Mann-Whitney-Wilcoxon test, Benjamini-
Hochberg correction).
 ll
Resource

Figure S6. Characterization of T cell memory formation and correlation with Gliobastoma patient survival. Related to Figure 6
(A) Gating strategy of naive/memory T cells. (B) Relative frequencies of main T cell subsets among CD3+ cells. (C) The Pearson correlation between T cell subset
frequencies and OS in the IDH1wt glioma group. Relative frequencies of naive/memory populations were calculated among T cells. Patients that have survived till
day of analysis (hence past 400 days) have been highlighted using an asterisk (*) alongside their Patient ID.
 ll
Resource

Figure S7. T Cell Exhaustion Characterize the TME of Brain Metastases, whereas Immature NK Cells Accumulate in Glioblastoma, Related to
Figures 6 and 7
(A) Median expression of antigens derived from the most differentially expressed genes among CD8 RM and CD8 EM. Boxplots quantify the mean antigen
intensity of data shown in Figure 6C. Point shapes differentiate patients who received therapy prior to surgery (including immune-, radio- and chemotherapy). Only
statistically significant p-values were displayed (p < 0.05, Mann-Whitney-Wilcoxon test, Benjamini-Hochberg correction). (B) Relative frequencies of three CD56
expressing populations identified in the brain samples among lymphocytes. Point shapes differentiate patients who received therapy prior to surgery (including
immune-, radio- and chemotherapy). (C) Relative frequencies of ILC populations among CD56+ cells identified in the glioblastoma cohort stratified according to
methylation status of the MGMT promoter. (B, C) Only statistically significant p-values were displayed (p < 0.05, Mann-Whitney-Wilcoxon test, Benjamini-
Hochberg correction). (D) The Pearson correlation between ILC cell subset frequencies and OS in the IDH1wt glioma group. Relative frequencies of ILC
populations were calculated among CD56+ cells. Patients that have survived till day of analysis (hence past 400 days) have been highlighted using an asterisk
(*) alongside their Patient ID.
 Resource

Interrogation of the Microenvironmental Landscape

in Brain Tumors Reveals Disease-Specific
Alterations of Immune Cells
Graphical Abstract Authors
Florian Klemm, Roeltje R. Maas,
Robert L. Bowman, ..., Roy T. Daniel,
Monika E. Hegi, Johanna A. Joyce

Correspondence
johanna.joyce@unil.ch

In Brief
High-dimensional, multi-omics
characterization of the brain tumor
microenvironment, including
comparisons of gliomas and brain
metastases, suggests that education of
immune cell types in the TME depends on
tumor origin and IDH mutational status.

Highlights
d Flow cytometry, RNA-seq, and protein and image analyses
reveal brain TME complexity

d Glioma IDH mutation status and brain metastasis primary

tumors shape the brain TME

d Microglia and monocyte-derived macrophages exhibit

multifaceted activation

d TME immune cells show disease- and cell-type-specific

expression patterns

Klemm et al., 2020, Cell 181, 1643–1660

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.007 ll
 ll

Resource
Interrogation of the Microenvironmental Landscape
in Brain Tumors Reveals Disease-Specific
Alterations of Immune Cells
Florian Klemm,1,2 Roeltje R. Maas,1,2,3,4 Robert L. Bowman,5 Mara Kornete,1,2 Klara Soukup,1,2 Sina Nassiri,1,2,6
Jean-Philippe Brouland,7 Christine A. Iacobuzio-Donahue,8 Cameron Brennan,9 Viviane Tabar,9 Philip H. Gutin,9
Roy T. Daniel,4 Monika E. Hegi,3,4 and Johanna A. Joyce1,2,10,*
1Department of Oncology, University of Lausanne, Lausanne, Switzerland
2Ludwig Institute for Cancer Research, University of Lausanne, Lausanne, Switzerland
3Neuroscience Research Center, Centre Hospitalier Universitaire Vaudois, Lausanne, Switzerland
4Department of Neurosurgery, Centre Hospitalier Universitaire Vaudois, Lausanne, Switzerland
5Memorial Sloan Kettering Cancer Center, New York, NY, USA
6Bioinformatics Core Facility, SIB Swiss Institute of Bioinformatics, Lausanne, Switzerland
7Department of Pathology, Centre Hospitalier Universitaire Vaudois, Lausanne, Switzerland
8Department of Pathology, Memorial Sloan Kettering Cancer Center, New York, NY, USA
9Department of Neurosurgery, Memorial Sloan Kettering Cancer Center, New York, NY, USA
10Lead Contact

*Correspondence: johanna.joyce@unil.ch
https://doi.org/10.1016/j.cell.2020.05.007

SUMMARY

Brain malignancies encompass a range of primary and metastatic cancers, including low-grade and high-
grade gliomas and brain metastases (BrMs) originating from diverse extracranial tumors. Our understanding
of the brain tumor microenvironment (TME) remains limited, and it is unknown whether it is sculpted
differentially by primary versus metastatic disease. We therefore comprehensively analyzed the brain TME
landscape via flow cytometry, RNA sequencing, protein arrays, culture assays, and spatial tissue character-
ization. This revealed disease-specific enrichment of immune cells with pronounced differences in propor-
tional abundance of tissue-resident microglia, infiltrating monocyte-derived macrophages, neutrophils,
and T cells. These integrated analyses also uncovered multifaceted immune cell activation within brain ma-
lignancies entailing converging transcriptional trajectories while maintaining disease- and cell-type-
specific programs. Given the interest in developing TME-targeted therapies for brain malignancies, this
comprehensive resource of the immune landscape offers insights into possible strategies to overcome
tumor-supporting TME properties and instead harness the TME to fight cancer.

INTRODUCTION mors, the incidence of BrMs significantly exceeds that of

Brain malignancies include tumors that arise within the brain,
such as low-grade gliomas and glioblastomas, and brain metas-
tases (BrMs), which originate from extracranial primary tumors,
including melanoma, breast, and lung cancers (Cagney et al.,
2017). Gliomas mutant for the metabolic enzymes isocitrate de-
hydrogenase 1 and 2 (IDH mut) are generally low grade (II or III)
and have a significantly better prognosis than IDH wild-type
(WT) tumors, which are typically grade IV glioblastomas. Despite
standard of care treatment comprising surgery followed by radi-
ation and temozolomide (Stupp et al., 2005), median survival
rates for glioblastoma patients remain stubbornly low (Aldape
et al., 2019). Patient survival following BrM diagnosis can be
even lower, with rates typically measured in months (Cagney
et al., 2017; Ceccarelli et al., 2016), and among all adult brain tu-
gliomas.
Given the current limited treatment options for these pa-
tients, a key question to address is whether a deep compre-
hensive understanding of how primary and metastatic cancers
develop within the brain tumor microenvironment (TME) could
reveal promising new targets for therapeutic intervention.
Although diverse TME cell types can critically regulate cancer
progression and response to therapy across a broad range
of extracranial tumors (Klemm and Joyce, 2015), we cannot
simply extrapolate findings from these cancers to the singular
brain TME, given its unique cell types, including astrocytes,
neurons, and microglia (MG); the immune-suppressive environ-
ment of this organ; and the challenges presented for cells and
drugs to cross the blood-brain barrier (BBB) (Quail and
Joyce, 2017).

Cell 181, 1643–1660, June 25, 2020 ª 2020 Elsevier Inc. 1643
 ll
Resource

A C

E F

(legend on next page)

1644 Cell 181, 1643–1660, June 25, 2020

 ll
Resource

Immune checkpoint blockade (ICB), adoptive cell therapy, and
vaccines represent treatments targeted against immune cells
within the TME and systemically. The success of immunother-
apies in certain extracranial cancers has led to clear motivation
for their evaluation in brain malignancies. However, although
they show some clinical efficacy in a subset of BrM patients
(Hendriks et al., 2019; Long et al., 2018; Tawbi et al., 2018),
ICB has only resulted in responses in isolated cases of primary
gliomas to date (Lim et al., 2018; Schalper et al., 2019). Beyond
tumor cell-intrinsic effects, this may be attributed in part to im-
mune-suppressive components of the brain TME, including tu-
mor-associated macrophages (TAMs), which have emerged as
prominent players in brain cancers (Gutmann and Kettenmann,
2019; Quail and Joyce, 2017).
Lineage-tracing experiments in mice revealed that brain TAMs
can originate from tissue-resident MG or monocyte-derived mac-
rophages (MDMs) recruited from the peripheral circulation
(Bowman et al., 2016; Chen et al., 2017). TAMs are highly plastic
cells that integrate input from cytokines, growth factors, and other
stimuli, resulting in diverse activation states and cellular pheno-
types, including promotion of invasion, angiogenesis, metastasis,
and immune suppression (Mantovani et al., 2017; Noy and Pollard,
2014). This plasticity and their position at the nexus between ma-
lignant cells and tumor-infiltrating T cells makes TAMs a promising
target of TME-directed therapies in different cancers. Indeed,
studies in mice showed that phenotypic alteration of TAMs results
in anti-tumor efficacy in glioblastoma (Pyonteck et al., 2013; Quail
et al., 2016; Yan et al., 2017), whereas TAM depletion prevents
BrM outgrowth (Qiao et al., 2019).
Despite these preclinical studies, the precise contribution of
the two ontogenetically distinct TAM cell types in human brain
malignancies is unclear, which hinders clinical translation. For
example, previous studies interrogating the role of TAMs in pa-
tient brain tumors did not distinguish between MG and MDMs
based on use of lineage tracing-derived markers (Gabrusiewicz
et al., 2016; Sankowski et al., 2019; Szulzewsky et al., 2016) or
focused solely on gliomas (Müller et al., 2017; Venteicher et al.,
2017). We therefore interrogated the TME landscape in gliomas
and BrMs, with an emphasis on exploring TAMs, while also
investigating their relation to other immune cells and structures
in the TME. We leveraged this multimodal resource to address
a number of questions. Do tumors arising within the brain shape
their TME differently than cancers that metastasize from extra-
cranial sites? Does IDH mutation status affect the TME? How
do distinct TME compositions potentially modulate the activa-
tion states of immune cells? By integrating the answers to these
questions, we provide insights into potential strategies to
harness the brain TME in the fight against these deadly diseases.
RESULTS
Tumor Origin and IDH Mutational Status Influence the
Immune Composition of Brain Malignancies
We first determined the broad immune cell abundance in the
brain TME by analyzing the pan-leukocyte marker CD45 through
immunofluorescence (IF) staining of whole-tissue sections and
flow cytometry (FCM) analyses of non-tumor brain tissue, IDH
mut low-grade and IDH WT high-grade gliomas, and BrMs orig-
inating from different primaries, including breast cancer, lung
cancer, and melanoma (Figures 1A, 1B, and S1A). This showed
a leukocyte abundance from 20%–40% across the cancer
samples. Stratification of CD45+ cells into myeloid and lymphoid
lineages revealed a significant increase in myeloid cells in IDH
mut and IDH WT gliomas and of lymphocytes in IDH WT tumors
and BrMs compared with non-tumor tissue (Figure 1B; p < 0.05,
one-sided Student’s t test). We used multicolor fluorescence-
activated cell sorting (FACS) to analyze 14 major immune cell
populations across 100 clinical samples (Figure S1A; Tables
S1 and S2) and collected cells for RNA sequencing (RNA-seq)
from 48 patients (Table S3; full clinical annotation).
By incorporating cell lineage tracing and mouse models of
high-grade gliomas and BrM, we previously identified the cell
surface marker integrin alpha 4, ITGA4/CD49D, as a means to
discriminate tumor-associated MG (T-MG) from tumor-associ-
ated MDMs (T-MDMs) (Bowman et al., 2016), which we inte-
grated here into clinical sample analyses. This enabled sorting
of CD45 non-immune cells, CD49Dlow MG, CD49Dhigh MDMs,
neutrophils, and CD4+ and CD8+ T cells (Figure S1A; Tables S2
and S3A) for transcriptome analysis by RNA-seq. We assessed
sorting fidelity by FCM re-analysis of the sorted CD49Dlow and
CD49Dhigh TAM populations (purity, 98.4%–99.8%) and by
investigating the frequency of the canonical IDH codon 132
missense mutation in the RNA-seq reads from CD45 cells
and CD49Dlow and CD49Dhigh TAM populations. Although we

Figure 1. The Immune Cell Composition of Brain Malignancies

(A) Quantification of immunofluorescence (IF) staining of non-immune (CD45 ) and immune cells (CD45+) in sections of non-tumor brain tissue (n = 6), gliomas
(nIDH mut = 16, nIDH WT = 16), and brain metastases (BrMs, nbreast = 12, nlung = 5, nmelanoma = 7). Data are represented as mean ± SEM.
(B) Flow cytometry (FCM) quantification of non-immune cells (CD45 ), myeloid cells (CD45+, CD11B+), and lymphocytes (CD45+, CD11B ) in non-tumor tissue
(n = 6), gliomas (nIDH mut = 17, nIDH WT = 40), and BrMs (nbreast = 13, nlung = 16, nmelanoma = 8). Data are represented as mean ± SEM.
(C) Gene set variation analysis (GSVA) normalized enrichment score (NES) of MG and MDM ontogeny-specific core gene signatures in CD49Dlow MG and
CD49Dhigh MDMs from non-tumor and tumor tissues.
(D) Heatmap of immune cell proportions in relation to all CD45+ cells (MG, microglia; MDM, monocyte-derived macrophage; CD14low/CD16+, CD14low/CD16+
monocyte; CD14+/CD16+, CD14+/CD16+ monocyte; CD16 Gran., CD16 granulocyte; iMC, immature myeloid cell; DC, dendritic cell; Treg, regulatory T cell;
DNT, double-negative T cell) across the cohort (nnon-tumor = 6, nglioma = 57, nBrM = 37). Cluster assignment, disease type, IDH mutation status, and BrM primary
tumor are annotated per column (for clinical information, see Table S1).
(E) Principal component (PC) biplot of FCM data with sample scores and top 5 loadings of the first two PCs (n = 100 clinical samples, proportion of variance shown
on PC axes).
(F) Mean of immune cell populations in non-tumor tissue (n = 6), gliomas (nIDH mut = 17, nIDH WT = 40), and BrMs (nbreast = 13, nlung = 16, nmelanoma = 8) as percentage
of CD45+ cells.
See also Figure S1 and Tables S1 and S2.

Cell 181, 1643–1660, June 25, 2020 1645

 ll
observed a mean mutated allele frequency of 0.43 in CD45 cells
from IDH mut gliomas (range, 0.3–0.61), this was very rare in
TAMs (mean, 0.01; range, 0.0–0.09), indicating reliable separa-
tion of cell populations. In a t-distributed stochastic neighbor
embedding (t-SNE) visualization of sorted populations, samples
clustered mostly by cell type (Figure S1B), with gliomas and
BrMs discernible as separate groups in the CD45 population.
In this global expression analysis in the context of the other
major brain TME components, CD49Dlow and CD49Dhigh TAM
populations clustered closely, suggesting broad transcriptomic
similarity. We thus further interrogated the utility of CD49D to
differentiate between TAM populations by analyzing association
of MG- and MDM-specific ontogeny core gene sets, identified
previously from lineage-tracing studies (Bowman et al., 2016),
in human CD49Dlow and CD49Dhigh cells sorted from non-malig-
nant and brain cancer tissues. This revealed enrichment of
ontogeny core gene sets in the corresponding cell type (Fig-
ure 1C), demonstrating our ability to accurately distinguish MG
and MDMs in human samples across different disease entities.
Interestingly, these core signatures were influenced within
certain tumor types, with T-MDMs showing an increased MG
core gene set signal in IDH mut gliomas and T-MG acquiring
MDM features in BrMs, suggesting tissue-dependent transcrip-
tional programming of these cells, as further interrogated below.
We next assessed the landscape of intratumoral immune cell
populations (Figure S1A; Table S2) using clustering analysis to
identify patterns of cellular abundance (Figure 1D; chi-square
test for independence, p < 0.0001). This revealed three major
clusters: (1) non-tumor samples and IDH mut gliomas character-
ized by dominance of MG with low numbers of other immune
cells; (2) IDH WT gliomas and several BrMs with an influx of
MDMs and, to some extent, neutrophils into the tumor while
mostly excluding lymphocytes; and (3) predominantly BrMs and
few IDH WT gliomas exhibiting the most diverse immune cell
landscape with substantial infiltration of T cells and neutrophils.
Certain tumors contained CD14low/CD16+ non-classical mono-
cytes, CD14+/CD16+ intermediate monocytes, CD16 granulo-
cytes, dendritic cells (DCs), or immature myeloid cells. Across
all samples, the lymphocyte compartment was mostly composed
of T cells with fewer natural killer (NK) cells and B cells.
Principal-component analysis (PCA) of the relative abundance
of all investigated populations confirmed that MG, MDMs, neu-
trophils, and CD4+ and CD8+ T cells are the major immune cell
determinants of the brain TME landscape (Figure 1E). Principal
component 1 (PC1) separated non-tumor tissue and IDH mut gli-
omas from IDH WT gliomas and BrMs, whereas PC2 distin-
guished IDH WT gliomas and BrMs. Further analysis stratifying
for IDH status in gliomas and the primary tumor site in BrMs veri-
fied a substantially higher proportion of lymphocytes in BrMs
(Figure 1F; meanlymphocytes %CD45+ = 46.23%, SEM = 4.15, t
test, p < 0.0001). Melanoma BrMs exhibited the most abundant
lymphocyte infiltrate with a sizeable CD8+ T cell fraction
(meanCD8+ %CD45+ = 33.01%, SEM = 5.82, one-way ANOVA,
p < 0.01). Regulatory T cells (Tregs) were detected in certain
BrMs (meanTreg %CD45+ = 1.2%, SEM = 0.36) but were rare in gli-
omas (meanTreg %CD45+ = 0.25 %, SEM = 0.05, t test, p < 0.05).
Because of the prominence of T-MG and T-MDMs in the
myeloid compartment of brain malignancies, we used IF staining
1646 Cell 181, 1643–1660, June 25, 2020
Resource
and deconvolution analyses to independently validate their pres-
ence. Commonly employed MG markers, such as P2RY12,
TMEM119, and SALL1, and MDM-associated genes, such as
AHR and VDR, showed varying RNA expression levels across
different brain malignancies while maintaining their cell type
specificity (Figure S2A) in a similar manner as observed for the
ontogeny core gene sets (Figure 1C). An equivalent pattern
was observed at the protein level (Figure S2B), where P2RY12
showed the highest expression in non-tumor tissue, and CD68
was most abundant in BrM-TAM populations. This necessitated
use of both markers complemented by CD49D to reliably identify
MG and MDMs in IF analyses (Figure S2C). We used this strategy
to interrogate a cohort of non-tumor, glioma, and BrM samples
by whole-section quantification, confirming MDM accumulation
in IDH WT gliomas and BrMs (Figures 2A–2C). Furthermore,
comparison of tissue processed independently for IF and FCM
from the same individual samples demonstrated significant
concordance (Figure S2D).
We queried the sorted cell populations for T-MG- and T-MDM-
specific differentially expressed genes (DEGs) that separate
these two populations from the most abundant other cell types;
i.e., CD45 cells, neutrophils, and T cells (Figure S2E). Several of
the genes highly expressed in T-MG are well-established MG
markers (P2RY12, TMEM119, and TAL1), whereas genes highly
expressed in T-MDMs include markers of alternative macro-
phage polarization (FCGR2B and CLEC10A) and DC-like pheno-
types (CD1C, CD1B, and CD207) with increased phagocytic and
antigen cross-presentation ability (CD209). These gene sets also
allowed us to utilize a publicly available integrated dataset (Vivian
et al., 2017) containing bulk expression data of healthy cortical
brain tissue from the Genotype-Tissue Expression project
(GTEx; GTEx Consortium, 2013) and low- and high-grade glioma
samples from The Cancer Genome Atlas (TCGA; Ceccarelli et al.,
2016) in a bulk tissue transcriptome deconvolution approach
(Racle et al., 2017). The estimates obtained of MG and MDM pro-
portions in this external dataset (n = 711 samples) verified the
prevalence of MG in IDH mut gliomas and MDM enrichment in
IDH WT gliomas (Figure 2D).
MG and MDMs Exhibit a Multifaceted Polarization
Phenotype in Brain Malignancies
We next employed PCA to specifically focus on TAMs and
analyze genes whose expression was influenced by tissue type
(i.e., reference MDMs, non-tumor brain, gliomas, and BrMs)
and cell type (i.e., MG and MDMs) (Figure 3A). Within the first
two PCs, MG and MDMs projected into different spaces, with
in vitro differentiated MDMs distinct from tissue-derived sam-
ples. We observed a gradient across PC1 with non-tumor brain
tissue at one end, traversing IDH mut and IDH WT gliomas,
and ending with BrMs. Thus, TAM transcriptomic changes are
influenced by the brain TME per se and also by the specific
type of malignancy.
We contrasted T-MG and T-MDMs from BrMs or gliomas
(regardless of IDH mutation status) with MG from non-tumor
brain or in vitro differentiated MDMs from healthy donors,
respectively (Figure S3A; Tables S3A and S4). This revealed pro-
found expression changes in both populations, with T-MDMs ex-
hibiting a higher magnitude in their transcriptional response
 ll
Resource

C D

Figure 2. Analysis of MG and MDM Abundance

(A and B) Representative IF images (A) and corresponding cell type identification (B) of MG (CD45+, P2RY12+/CD68+, CD49D ), MDMs (CD45+, P2RY12+/CD68+,
CD49D+), and non-immune (CD45 ) and non-TAM immune cells (CD45+, P2RY12 /CD68 , CD49D /+) in non-tumor brain tissue, IDH mut and IDH WT gliomas,
and BrMs. Scale bars, 100 mm. Insets show quantification per field of view (FOV).
(C) IF quantification of MG and MDM abundance in non-tumor brain tissue (n = 6), IDH mut (n = 16) and IDH WT (n = 16) gliomas, and BrMs (n = 24).
(D) Deconvolution of merged GTEx and TCGA glioma datasets, showing relative abundance of MG, MDMs, and non-TAMs (‘‘other cells’’) in healthy frontal cortex
and IDH mut and IDH WT gliomas.
Wilcoxon rank-sum test was used for statistical analysis: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. See also Figure S2.

compared with T-MG (Figure S3A). The intersect of DEGs in gli-
omas and BrMs was highest in T-MDMs (Figure S3B), potentially
reflecting the greater changes experienced by these cells upon
entering the completely foreign environment of a brain tumor.
This was also evident when focusing on genes upregulated in gli-
omas and BrMs that are exclusive to T-MG or T-MDMs (Fig-
ure 3B). In T-MG and T-MDMs, the number of shared genes
was higher across different diseases than between these two
cell populations within the same tumor type. Consequently,
only a small number of genes (n = 137) showed concordant up-
regulation across a comparison of all diseases and TAM types
(Figure 3B).
To explore the underlying biological processes conserved in
gliomas and BrMs, we examined the intersect of upregulated
genes (Figure S3B) in T-MG or T-MDMs using gene set over-
representation analysis (ORA). In the Molecular Signature Data-
base (MSigDB; Liberzon et al., 2015) ‘‘hallmark’’ collection of
major biological categories, T-MG and T-MDMs showed
pathway enrichment in (1) modeling of the TME (‘‘Angiogenesis,’’
‘‘Hypoxia’’), (2) inflammation (‘‘Inflammatory Response,’’ ‘‘Allo-
graft Rejection’’), and (3) immune cell activation states (‘‘TNF⍺
Signaling via NFkB,’’ ‘‘Interferon ⍺ Response,’’ ‘‘Interferon g
Response,’’ ‘‘IL2 STAT5 Signaling,’’ and ‘‘IL6 JAK STAT3
Signaling’’) (Figure S3C).
We also assessed the M1 and M2 polarization status of T-MG
and T-MDMs using a panel of marker genes (Murray et al., 2014).
However, no evident pattern emerged of a defined M1 versus M2
phenotype in glioma or BrM T-MG or T-MDMs (Figure S3D). To
further explore the activation state of T-MG and T-MDMs, we
subjected their respective upregulated genes to ORA of
macrophage stimulus-specific programs (Xue et al., 2014). This
revealed a multifaceted response (Figure 3C) incorporating ca-
nonical M1 (interferon g [IFNg]) and M2 polarization (inter-
leukin-4 [IL-4]), including expression changes associated with
chronic inflammatory stimuli (tumor necrosis factor alpha
[TNF-⍺] + prostaglandin E [PGE2] and TNF⍺ + PGE2 +
Pam3CysSerLys4 [TPP]) and exposure to free fatty acids (oleic
acid [OA] and palmitic acid [PA]), which have been implicated
in modulating myeloid cell function (Thapa and Lee, 2019). This
indicates diverse transcriptional programming of T-MG and
T-MDMs in gliomas and BrMs extending beyond simple M1
versus M2 polarization.
To understand which processes are linked to and potentially
driving these responses, we identified the gene set enrichment

Cell 181, 1643–1660, June 25, 2020 1647

 ll
Resource

A B C

Figure 3. MG and MDMs Exhibit a Multifaceted Polarization Phenotype in Brain Malignancies

(A) PC biplot of MG and MDM transcriptome data from non-tumor brain tissue, IDH mut and IDH WT gliomas, and BrMs (for clinical information, see Table S3A;
reference = in-vitro-generated MDMs; proportion of variance shown on PC axes).
(B) Visualization of intersects of the conserved sets of significantly upregulated genes in MG and MDMs. Intersects between sets are shown in the combination
matrix. ngenes found uniquely in a gene set or intersect is indicated above individual bars.
(C) Stimulus-specific macrophage gene expression modules overrepresented (within conserved differentially expressed genes [DEGs] versus respective ref-
erences) in tumor associated MG (T-MG) and tumor-associated MDMs (T-MDMs). Bar heights and color indicate significance level. GC, glucocorticoid; IFNg,

(legend continued on next page)

1648 Cell 181, 1643–1660, June 25, 2020


Resource
analysis (GSEA; Subramanian et al., 2005) leading-edge genes in
T-MG and T-MDMs in gliomas and BrMs and clustered them into
leading-edge metagenes (LEMs) with non-negative matrix
factorization (Godec et al., 2016). This identified up to 5 distinct
LEMs per cell type and comparison that were tested for signifi-
cant overlap in a pairwise fashion (Figure S3E) and annotated us-
ing Gene Ontology (GO) terms (Figure 3D). LEMs associated with
mitosis and cell proliferation were present in T-MG and T-MDMs
in gliomas and BrMs (Figure 3D, group 1). The biological validity
of these LEMs were verified by staining for Ki-67, a marker of cell
proliferation, in non-tumor, glioma, and BrM tissue sections (Fig-
ure 3E), showing increased proliferation in T-MG and T-MDMs in
IDH WT gliomas and BrMs and in T-MG in IDH mut gliomas.
Interestingly, LEMs enriched for type I IFN signaling were de-
tected in glioma and BrM T-MDMs and in BrM T-MG but not in
glioma T-MG (Figure 3D, group 2). Sustained type I IFN signaling
has been implicated in mediating immune suppression and ICB
resistance (Benci et al., 2016). The stringency of these group 2
LEMs was validated by building a protein-protein interaction
(PPI) network of the shared LEM genes (Figure S3F). Beyond
their role in antiviral responses, the genes highlighted at the cen-
ter of the PPI network (Figure S3F, red nodes) have been impli-
cated in a variety of tumor-promoting and -suppressing roles
(Benci et al., 2016). Similarly, the more peripheral network nodes
IL15 and TNFSF10 are potentially able to modulate an effective
immunological anti-tumor response or induce apoptosis in can-
cer cells, respectively (Bouralexis et al., 2005; Santana Carrero
et al., 2019). We asked whether these genes were directly
induced by secreted factors in the brain TME and established
cell-based assays to expose MDMs to TME conditioned medium
(CM) generated from single-cell suspensions of freshly isolated
glioma or BrM samples in culture. All genes analyzed were upre-
gulated by BrM-TME-CM and to a lesser extent by glioma TME-
CM (Figure 3F). We also detected induction of inflammation- and
nuclear factor kB (NFkB) signaling-associated LEMs in BrM-MG,
glioma MDMs, and BrM-MDMs (Figure 3D, group 3). LEMs that
point toward a Th17 response (group 4) and recruitment of im-
mune cells and interactions between different immune cell com-
partments were exclusively detected in MDMs (group 5). Collec-
tively, these analyses reveal acquisition of a multifaceted
activation state of MG and MDMs upon their integration into
the TME of brain malignancies.
IDH Mutation Status Associated with Changes in Glioma
TAM Activation
We next asked whether MG and MDMs occupy distinct regions
within the TME of IDH WT gliomas. Spatial analysis of tissue sec-
tions showed significant enrichment of both populations in the
interferon gamma; LA, lauric acid; LiA, linoleic acid; OA, oleic acid; PA, palmitic acid; PGE2, prostaglandin E2; sLPS, standard lipopolysaccharide; TNF-a = tumor
necrosis factor alpha; TPP, TNFa + PGE2 + Pam3CysSerLys4; IL-10, interleukin 10.
(D) Heatmap of GO overrepresentation analysis of leading-edge metagenes (LEMs) in MG and MDMs from gliomas and BrMs. Tile fill indicates significance
(hypergeometric test, -log10 (adjusted p value), terms were filtered by significance).
(E) IF quantification of the proportion of proliferating Ki67+ MG and MDMs in non-tumor tissue (n = 5), IDH mut (n = 10) and IDH WT gliomas (n = 9), and BrMs (n = 8).
Means were compared with one-tailed t test: *p < 0.05.
(F) qRT-PCR of type I IFN LEM marker genes from group 2 (Figure S3F) in in-vitro-generated MDMs stimulated with the indicated TME culture-conditioned
medium (TME CM). Fold changes were calculated relative to colony-stimulating factor-1 (CSF-1)-treated MDM baseline (one-way ANOVA, p < 0.1, nMDM = 4–11).
Data are represented as mean ± SEM.
See also Figure S3 and Table S4.
ll
perivascular niche (Figures 4A and S4A). Analysis of their distri-
bution relative to CD31+ vascular structures showed a closer
proximity of T-MDMs compared with T-MG (Figures 4B and
S4A). Interrogation of anatomical transcriptome data from the
Ivy Glioblastoma Atlas Project (Ivy GAP) study (Puchalski et al.,
2018) also demonstrated enrichment of T-MDMs in the micro-
vascular compartment (Figure S4B). This enrichment coincided
with CD4+ and CD8+ T cells, indicating further spatial TME orga-
nization in IDH WT gliomas.
We assessed whether the distinct T-MG and T-MDM distribu-
tions and cell numbers are paralleled by their activation state. In
the LEM analysis, we had detected a type I IFN response in gli-
oma MDMs but not MG (Figure 3D); we therefore queried the
FCM data to analyze levels of major histocompatibility complex
(MHC) class II human leukocyte antigen-DR isotype (HLA-DR)
expression. This showed significantly increased HLA-DR in
T-MDMs compared with T-MG in IDH mut and IDH WT tumors
(Figure 4C). We screened the associated RNA-seq data for anti-
gen processing and presentation pathway gene sets using GSEA
and gene set variation analysis (GSVA) (Figure 4D). Interestingly,
we found evidence of increased expression of MHC class II an-
tigen presentation gene sets in IDH WT glioma MDMs and also
antigen processing-associated pathways (Figure S4C) and
MHC class I presentation gene sets (Figure 4D). Although these
findings suggest the potential of TAMs, particularly T-MDMs, to
initiate an immune response, this potential is generally not real-
ized in the glioma TME, based on the current status of ICB trials
in this disease, and we thus asked whether there was also evi-
dence of pro-tumor states in these cell populations.
We compared T-MG and T-MDMs from IDH WT gliomas with
T-MG from IDH mut gliomas because they constitute the most
abundant TME cell types in these tumors, respectively (Figures
1F and 2C; Table S5). This revealed 489 DEGs in T-MG (Fig-
ure 4E; Table S5; 406 up- and 83 downregulated), and 1,478
DEGs in T-MDMs (Figure 4F; Table S5; 903 up- and 575 down-
regulated). Although these gene lists were generated by
comparing T-MDMs from IDH WT gliomas with T-MG from IDH
mut gliomas, they similarly separated T-MDMs in IDH mut versus
IDH WT disease in a clustering analysis (Figure 4F), indicating
that they indeed reflect T-MDM alterations based on the IDH sta-
tus of the tumor. 421 genes exhibit a similar pattern across both
TAM cell types (343 up- and 78 downregulated), suggesting that
T-MG and T-MDMs can also acquire a common transcriptional
pattern in IDH WT tumors. Among the shared genes were
several encoding extracellular matrix (ECM) proteins (Figure 4G,
FN1 and VCAN) and ECM-associated matricellular proteins
(THBS1, TGFBI, LGALS3, and ANGPTL4) that regulate the avail-
ability of ECM-sequestered ligands, angiogenesis, and tumor
Cell 181, 1643–1660, June 25, 2020 1649
 ll
Resource

A B C

E G H

I J

(legend on next page)

1650 Cell 181, 1643–1660, June 25, 2020

 ll
Resource

immunity (Mushtaq et al., 2018). This suggests that TAMs help
shape the composition and effector functions of ECM proteins
in IDH WT tumors. We also found the anti-inflammatory mole-
cules ANXA1 and GPNMB (Figure 4G), previously implicated in
pro-tumorigenic macrophage polarization and inhibition of
T cell activation (Kobayashi et al., 2019; Ripoll et al., 2007), to
be upregulated in T-MG and T-MDMs.
We next investigated inflammation mediators within the
CD45 population of IDH WT tumors in parallel with their corre-
sponding receptors in TAMs. TGFB2 expression was elevated
compared with IDH mut CD45 cells, and the accessory trans-
forming growth factor b (TGF-b) receptor ENG was highly ex-
pressed in IDH WT TAMs (Figure 4H). TGFB2 has pleiotropic ef-
fects in inflammation and tissue remodeling during wound
healing and has been implicated in an autocrine signaling loop
in glioblastoma cells (Rodón et al., 2014). The neuroinflammatory
cytokine MDK, which modulates TAM polarization to a M2-like
phenotype in glioma (Meng et al., 2019), was upregulated in
CD45 cells from IDH WT tumors, and its receptors SDC4 and
ITGA4/CD49D were differentially expressed in T-MDMs versus
T-MG (Figure 4H), suggesting cell-type-specific effects of this in-
ferred signaling loop.
We asked whether a T-MDM-specific gene set generated from
IDH WT gliomas was associated with a survival difference in pa-
tients. By logistic regression, we derived a representative signa-
ture consisting of 36 genes (Figure S4D) from the total number of
genes upregulated in TAMs in brain malignancies (Figure 3B).
This included the macrophage marker RUNX3; the atypical che-
mokine receptor ACKR3, which can regulate CXCL12-CXCR4
signaling; the endoplasmic reticulum (ER) stress protein
HERPUD1 and the inhibitory Fc receptor FCGR2B, which can
modulate macrophage activation (Bournazos et al., 2016; Li
et al., 2018); and the cytokine IL19, which affects angiogenesis
and macrophage polarization (Richards et al., 2015). The signature
was used to classify patients in a merged TCGA dataset of low-
and high-grade gliomas (Figures 4I and S4E). In IDH mut patients,
a decrease in median overall survival was associated with enrich-
ment of the T-MDM IDH WT signature, whereas IDH WT patients
with a low enrichment score showed increased survival. This
was confirmed in a multivariate Cox proportional hazard model
that included the transcriptomic glioma subtypes (as annotated
in the TCGA dataset) and IDH status (Figure 4J). To verify that
this effect did not simply reflect changes in T-MDM number, we
classified the TCGA cohort based on enrichment of the T-MDM-
specific gene set used for deconvolution, which showed a low ef-
fect on survival (Figure S4F).
In light of disappointing outcomes from PD1 or PDL1 ICB trials
in glioblastoma to date, we queried whether the abundant T-MG
and T-MDMs could contribute to the limited therapeutic efficacy.
We performed ORA of a panel of 20 gene sets previously asso-
ciated with innate anti-PD1 resistance (IPRES; Hugo et al.,
2016) in the TAM DEGs of IDH WT gliomas and found a sizeable
fraction to be upregulated in T-MG and T-MDMs (Figure S4G).
We then included the CD45 population and interrogated enrich-
ment of IPRES gene sets on the single-sample level by GSVA
(Figure S4H). This yielded a diverse picture with tumor cells
and TAMs enriched for IPRES gene sets to varying degrees.
Therefore, TAMs and CD45 cells from IDH WT gliomas may
contribute to mediating innate ICB resistance.
The Immune Contexture Influences the TME on a
Global Level
Through integrated analysis of protein and gene expression
data, we next explored the effect of immune cell infiltration.
Of 200 inflammation-associated proteins assessed, 55 were
differentially detected in our sample cohort (for clinical infor-
mation, see Table S3B). Unsupervised clustering analysis re-
vealed distinct clusters with abundant inflammatory proteins in
tumors (Figure 5A). The profile of IDH WT gliomas and BrMs
showed a sizeable overlap (protein cluster 1), encompassing
angiogenic factors (VEGFA and ANG), growth factors (PDGFA,
TGFB1, SPP1, and GDF15), several proteases and protease in-
hibitors (SERPINE1, CTSS, and TIMP1), the proteolysis cascade
regulator PLAUR, and the cytokines CCL2 and CCL5 (Figures
5A and S5A). However, we also found distinct protein
patterns between gliomas and BrMs. The neurotrophic growth
factor FGF2 and neuronal cell adhesion molecules, including
ALCAM, which regulates immune cell infiltration during

Figure 4. IDH Mutation Status Shapes TAM Activation in Gliomas

(A) Number of MG and MDMs per square millimeter in the perivascular niche (PVN) or distant from the PVN (non-PVN) in IDH WT gliomas by IF staining. Means
were compared with Wilcoxon signed-rank test: ***p < 0.001.
(B) Distance of MG and MDMs to the nearest vessel in IDH WT gliomas (nsamples = 14, nMG = 88,781, and nMDM = 92,969 cells counted).
(C) Boxplot of HLA-DR geometric mean fluorescence intensity measured by FCM in MG and MDMs in IDH mut and IDH WT gliomas. MG and MDMs from the
same samples are connected by lines (nIDH mut = 17, nIDH WT = 39; Wilcoxon signed-rank test: ***p < 0.001, ****p < 0.0001).
(D) GSVA of antigen processing and presentation pathways from the Molecular Signatures Database (MSigDB) ‘‘Canonical Pathways’’ collection with significant
differential enrichment between MG and MDMs in IDH WT tumors and in MG and MDMs across IDH mut and IDH WT samples. Columns are ordered by IDH
mutation status and cell type, and rows (Z score) are hierarchically clustered.
(E and F) Expression heatmap of T-MG (E) and T-MDM (F) DEGs (compared with T-MG in IDH mut gliomas) in IDH mut and IDH WT glioma samples. Columns and
rows (Z score) are hierarchically clustered.
(G) Normalized counts of selected genes in MG and MDMs in gliomas stratified by IDH status. Data are represented as mean ± SEM.
(H) Relative expression in CD45 MG and MDM cells of ligands and receptors upregulated in CD45 cells in IDH WT versus IDH mut samples and their matching
counterparts. Variance-stabilized expression values were scaled to the expression range.
(I) Kaplan-Meier estimator of survival in the TCGA glioma cohort based on enrichment for the MDM IDH WT signature, assessed by GSVA in IDH mut and IDH WT
gliomas from the combined TCGA cohort. GSVA scores were separated into tertiles across the combined IDH mut and IDH WT sample set. Pairwise p values were
calculated using a log rank test.
(J) Hazard ratios of a multivariate Cox proportional hazards model with transcriptomic subtype (TCGA annotation), IDH status (TCGA annotation), and T-MDM IDH
WT GSVA score as covariates for overall survival within the TCGA glioma cohort (PN, proneural; NE, neural; CL, classical; ME, mesenchymal subtype).
See also Figure S4 and Table S5.

Cell 181, 1643–1660, June 25, 2020 1651

 ll
Resource

B C

Figure 5. The Immune Contexture Influences the TME on a Global Level

(A) Inflammation-associated bulk tissue protein concentration heatmap subset on 55 proteins with significantly different concentrations between non-tumor
brain, gliomas, and BrMs in an ANOVA (p < 0.1, nnon-tumor = 3, nglioma = 14, nBrM = 12; concentrations were log10-transformed and Z scored). Rows and columns
(legend continued on next page)

1652 Cell 181, 1643–1660, June 25, 2020


Resource
neuroinflammation (Lécuyer et al., 2017), were highly expressed
in non-tumor brain, IDH mut, and IDH WT samples (protein clus-
ter 3; Figure S5A). Conversely, BrM samples had abundant im-
mune-regulatory molecules affecting myeloid and lymphocytic
cells and their heterotypic signaling (protein cluster 2; Fig-
ure S5A), such as CD40L, IL6R, INHBA, and AREG (Morianos
et al., 2019; Zaiss et al., 2015), possibly reflecting the greater im-
mune cell diversity in BrMs. This orthogonal dataset reinforces
the RNA-seq analyses showing that inflammatory signaling path-
ways are highly enriched in brain tumors.
We integrated the cell-type-specific RNA-seq data and bulk
protein data to distinguish proteins with more restricted expres-
sion versus those that are expressed across a range of cell types.
Transcriptome data from CD45 cells, TAMs, neutrophils, and
CD4+ and CD8+ T cells from all tumor samples were clustered us-
ing a self-organizing map (SOM). This yielded 6 SOM spots (i.e.,
metagenes of co-expressed genes; Figure 5B) that recapitulated
the respective cell lineages (Figure S5B). The CD45 populations
were assigned to three distinct spots that were associated with
more aggressive IDH WT gliomas and BrMs (spot VI) or reflected
the brain-intrinsic or -extrinsic tumor origin (spots I and V). These
cell-type-associated SOM spots overlapped considerably with
the protein data (30 of 55 proteins, Fisher’s exact test, p <
0.0001; Figure 5C). Although VEGFA, ANG, and TGFB1 were ex-
pressed by diverse cell types in gliomas and BrMs, other genes,
such as GDF15 and IGFBP2, showed more CD45 cell-restricted
expression (Figure 5D). The significant contribution of TAMs to
production of key inflammatory proteins, including SPP1 and
IHNBA, is reflected by TAM SOM spot III, constituting the largest
group of proteins with cell type-specific expression (Figures 5C
and 5D).
Myeloid Cells Show a Distinct Phenotype in BrMs
Our global analysis juxtaposing the expression patterns of
TAMs in gliomas (regardless of IDH status) with BrMs showed
disease- and cell-type-specific transcriptomic changes. We
thus explored BrM-specific alterations by focusing on genes
upregulated only in relation to the corresponding reference
and to IDH WT gliomas (Figures S6A and S6B; Table S6).
Various cytokines, chemokines, and pro-inflammatory mole-
cules were elevated in BrM-MG and BrM-MDMs (Figure 6A),
including the potent mediators of autoimmune neuroinflamma-
tion CSF2 and IL23A (Zhao et al., 2017) and the pattern
recognition receptor MARCO. Intriguingly, antibody-mediated
MARCO targeting in extracranial tumors increases M1-like po-
larization of TAMs and enhances ICB efficacy (Georgoudaki
et al., 2016). These effects relied on interaction of the antibody
with FCGR2B, which is also part of the T-MDM IDH WT signa-
ture (Figures S2E and S4C). Finally, RETN, which is involved in
are hierarchically clustered. Clinical data are annotated per row; column annotation reflects the major protein clusters (further information can be found in
Table S3B).
(B) Self-organizing map (SOM) of RNA expression data of major cell populations in glioma and BrM samples. SOM spots are highlighted, numbered with Roman
numerals, and annotated with their cell type association.
(C) Overlap of individual proteins and SOM spot metagenes; tile color fill reflects protein cluster membership from (A).
(D) RNA-seq counts (normalized, scaled to expression range) of proteins from (A) across major cell types in IDH mut and IDH WT gliomas and BrMs. SOM spot
membership of individual genes is indicated per row.
See also Figure S5.
ll
systemic inflammatory disorders (Filková et al., 2009), was up-
regulated in BrM-TAMs (Figure 6A).
Analysis of individual BrM-TAM populations uncovered
distinct expression patterns. BrM-MG showed restricted upre-
gulation of IL6 (Figure 6A), which exerts immunosuppressive ef-
fects on T cell function in cancer and mediates ICB resistance
(Tsukamoto et al., 2018), and the receptor TREM1, which mod-
ulates pro-inflammatory responses in MG and systemically in
myeloid cells during neuroinflammation (Liu et al., 2019; Xu
et al., 2019). Among the upregulated chemokines, we found in-
creases in both TAM cell types (CCL23) and BrM-MG-restricted
(CXCL5 and CXCL8) or BrM-MDM-restricted increases (CCL8,
CCL13, CCL17, and CCL18) (Figure 6A). These results reveal
distinct contributions of TAM populations to the inflammatory
TME milieu in a disease-specific manner.
GSEA identified additional cell-type-specific enrichment pat-
terns. BrM-MG showed evidence of IL-6 pathway activity (Fig-
ure S6C), and in BrM-MDMs, the ‘‘Naba core matrisome’’ gene
set was significantly enriched (Figure S6D). This prompted us
to assess expression of genes encoding ECM and matricellular
proteins in BrM-MDMs versus BrM-MG, which revealed genes
encoding matrix proteins, including type III and IV collagens,
FN1, the proteoglycans LUM and OGN, and the matricellular
proteins ECM1, SPARC, and SPARCL1 as highly expressed in
BrM-MDMs (Figure 6B). Although ECM remodeling has been
implicated in tumor progression, LUM, OGN, SPARCL, and
SPARCL1 exhibit pro- and anti-metastatic properties, which un-
derscores the complex context-dependent role of the ECM (Kai
et al., 2019). We also found high expression of the cathepsin pro-
teases CTSB and CTSW in BrM-MDMs (Figure 6B), which partic-
ipate in multiple tumor-promoting processes, including invasion
and metastasis (Olson and Joyce, 2015). The hyaluronan recep-
tor HMMR, involved in macrophage chemotaxis and fibrosis in
lung injury (Cui et al., 2019), was also higher in BrM-MDMs (Fig-
ure 6B). Together, these data suggest that the ECM is not only
shaped by macrophages at the primary site (Afik et al., 2016)
but that T-MDMs may also play a pivotal role in ECM niche con-
struction in BrM that is distinct from IDH WT gliomas (Figure 4G).
Given the upregulation of CXCL8, a key neutrophil chemoattrac-
tant, by BrM-MG (Figure 6A), we explored the TME contribution to
recruitment of neutrophils, which were highly abundant in BrM
(Figure 1F). Analysis of major neutrophil-recruiting chemokines
and their receptors showed broad expression across all interro-
gated myeloid cells (Figure S6E). To explore the phenotype of
BrM-associated neutrophils, we queried the RNA-seq data, which
revealed BrM-specific upregulation of ITGA3 (Figure 6C), which is
involved in neutrophil tissue infiltration in sepsis, and CXCL17, pre-
viously implicated in neutrophil and macrophage recruitment in
cancer (Li et al., 2014). We also observed upregulation of the
Cell 181, 1643–1660, June 25, 2020 1653
 ll
B C
Figure 6. Myeloid Cells Show Distinct Transcriptional Changes in BrMs
(A) Normalized counts of the indicated genes in MG and MDMs in non-tumor or reference, IDH WT gliomas, and BrMs. Data are represented as mean ± SEM.
(B) Expression heatmap of Extracellular matrix-associated genes differentially expressed between MG and MDMs in BrMs. Rows are Z-scored and manually
sorted, and columns are ordered by cell type.
(C) Expression of the indicated BrM-specific genes in neutrophils from unmatched healthy blood, IDH WT gliomas, and BrMs. Data are represented as mean
± SEM.
See also Figure S6 and Table S6.
adenosine receptor ADORA2A (Figure 6C), which attenuates the
phenotype of pro-inflammatory neutrophils (Barletta et al., 2012).
Furthermore, we found increased expression of CD177 (Fig-
ure 6C), a cell surface receptor that modulates neutrophil migra-
tion and activation and serves as a marker for PR3-positive neutro-
phils, which, in turn, negatively affect T cell proliferation (Yang
et al., 2018). Notably, MET, which has been linked to recruitment
of immunosuppressive neutrophils in cancer (Glodde et al.,
2017), was upregulated in neutrophils in a BrM-specific manner
(Figure 6C). In sum, we have uncovered multiple disease-specific
alterations of myeloid cells extending beyond BrM-TAMs to neu-
trophils, which has potential implications for the recruitment and
activation of other cell types within the TME, including T cells.
TAMs Are Poised toward an Immunomodulatory
Phenotype in BrMs
Although we found a significant accumulation of CD4+ and CD8+
T cells in BrMs versus IDH WT gliomas by FCM, this analysis of
dissociated tissue samples lacks structural information. We thus
performed neighborhood analysis of IF-phenotyped IDH WT and
BrM tissue sections to elucidate whether there is a spatial rela-
tionship between TAMs and CD3+ T cells in BrMs. In IDH WT gli-
1654 Cell 181, 1643–1660, June 25, 2020
Resource
omas, T-MG and T-MDMs mostly neighbored homotypic cells
while lacking T cells in their close vicinity (Figures 7A, 7B, and
S7A), possibly reflecting the general T cell sparseness in these
tumors. In contrast, both TAM populations neighbored T cells
far more frequently in BrMs, indicating the potential for interac-
tion (Figures 7A, 7B, and S7A).
We thus investigated the T cell activation state in BrMs in rela-
tion to unmatched healthy donor blood and also juxtaposed
them to the corresponding populations from IDH WT gliomas.
Compared with controls, CD4+ T cells from BrM showed evi-
dence of a hyporesponsive, anergic phenotype (Figure 7C),
whereas CD8+ T cells exhibited an exhaustion signature (Fig-
ure 7D), which usually occurs upon chronic activation, resulting
in upregulation of inhibitory receptors. These defective T cell
states can be caused by aberrant activation or T cell inhibition
by tumor cells and antigen-presenting cells in the TME and are
a major obstacle in treating cancers.
To delineate putative mechanisms in the BrM TME that may
drive these alterations, we probed the RNA-seq data from
CD45 cells, TAMs, and T cells (Figure 7E) for expression of acti-
vating and inhibitory immunomodulatory signals (Wei et al.,
2018). This revealed upregulation of various canonical T cell
 ll
Resource

A B

C D

E F

(legend on next page)

Cell 181, 1643–1660, June 25, 2020 1655

 ll
Resource

activators and co-activators but also mediators of inhibition in
T cells (PDCD1/PD1, CD28, and CTLA4), whereas T cell-inhibit-
ing and activating signals were detected in both TAM popula-
tions (CD274/PD-L1 and PDCD1LG2/PD-L2). Notably, we found
an upregulation of CD80, which has diverse roles in T cell activa-
tion because it heterodimerizes with CD274, provides co-stimu-
latory signals to T cells via CD28 and exerts inhibitory effects via
interaction with CTLA4 (Zhao et al., 2019), in both TAM popula-
tions compared with their normal references and IDH WT tumor
populations (Figure 7E). The potential contribution of TAMs to
metabolic immune evasion is also suggested by high expression
of IDO1 and IDO2 (Zhai et al., 2018) in BrM (Figure 7E).
We investigated additional immunomodulatory mediators us-
ing weighted gene correlation network analysis (WGCNA; Lang-
felder and Horvath, 2008) and correlated the resulting expres-
sion patterns with paired FCM abundance of CD4+ and CD8+
cells in a disease- and cell-type-specific manner. We identified
15 unique co-expression modules showing significant correla-
tion (p < 0.05) of their eigengenes (i.e., the first PC of the module
expression data) with any of the provided sample traits (Fig-
ure S7B). Among these, the ‘‘brown’’ WGCNA module correlated
with a specific BrM-MDM annotation and CD4+ T cell abun-
dance. ORA of this module revealed signals for pathways such
as coagulation and ECM modulation (Figure S7C) that affect
the availability and activity of growth factors and cytokines within
the TME (Mohan et al., 2020). We ranked genes by module mem-
bership strength and correlation with CD4+ T cell abundance,
which identified several factors with opposing immunomodula-
tory functions (Figure 7F). Although the receptors CD300E and
BST1 promote monocyte motility and survival (Isobe et al.,
2018; Ortolan et al., 2019), we also detected effectors of immu-
nosuppression, such as the actin-associated regulatory protein
CNN2, which negatively regulates macrophage motility and
phagocytic activity (Huang et al., 2008). The leukocyte immuno-
globulin-like receptor subfamily B members LILRB2 and LILRB3,
which attenuate myeloid cell activation (van der Touw et al.,
2017), are also highly ranked genes within this module. Interest-
ingly, LILRB2 has been identified as a novel myeloid immune
checkpoint that limits antitumor immunity (Chen et al., 2018).
We also found evidence of effects on T cells; CD52, which, in
its soluble form, inhibits T cell function, was among the BrM-
MDM module genes. The notion that BrM-MDMs undergo dis-
ease-specific alterations distinct from the primary extracranial
tumor is supported by upregulation of these genes (Figure 7G)
in our analysis of an external cohort of BrM samples compared
with their matched primary tumor tissue (Vare slija et al., 2019).
DISCUSSION
Brain tumors, including glioblastoma and BrMs, confer some of
the poorest prognoses for patients with cancer, with survival
rates often measured in just months. Given the current dearth
of effective therapeutic options for these patients and the
modest effects of the various immunotherapies evaluated to
date, it is of critical urgency to identify novel targets for future
clinical evaluation. One potentially rich source of therapeutic tar-
gets is the TME. However, even though the TME is now widely
accepted as an important regulator of cancer progression and
therapeutic response, our knowledge of the brain TME is
restricted to individual brain tumor types or cellular compart-
ments and lacks comprehensive and integrative analysis.
In this study, we leveraged a diverse panel of analyses to
deeply interrogate the immune landscape of primary and meta-
static brain cancers. Through integration of multiparameter
FCM analyses, RNA-seq data, TME cell culture assays, protein
arrays, and spatial tissue characterization, we uncovered critical
insights into the composition and transcriptomes of the most
abundant immune cell populations in patient samples from IDH
mut and WT gliomas and BrMs originating from distinct extracra-
nial primary tumors.
By exploring the broad immune landscape, we uncovered
several pronounced differences between gliomas and BrMs
when directly compared side by side. In brain tumors, TAMs
are composed of tissue-resident MG and recruited MDMs, and
we found a significant shift in the ratio of MG to MDMs between
IDH mut and IDH WT gliomas. Additionally, gliomas contain an
abundance of TAMs, whereas T cells were much fewer, particu-
larly in IDH mut tumors. This confirms the notion that gliomas are
immunologically cold tumors (Jackson et al., 2019). Although
T cell sequestration in the bone marrow has been observed in gli-
oma mouse models and following intracranial implantation of
brain-extrinsic tumors (Chongsathidkiet et al., 2018), our clinical

Figure 7. TAMs Have a Wide Range of Immunomodulatory Functions in BrMs

(A) Representative IF images and corresponding cell type identification of non-immune cells (CD45 ), MG (CD45+, P2RY12/CD68+, CD49D ), MDMs (CD45+,
P2RY12/CD68+, CD49D+), CD3+ (CD45+, P2RY12/CD68 , CD49D /+, CD3+) and CD45+ other cells (CD45+, P2RY12/CD68 , CD49D /+, CD3 ) in IDH WT gli-
omas and BrMs. Scale bars, 50 mm. Insets show quantifications per FOV.
(B) Neighborhood analyses of IDH WT glioma and BrM IF tissue sections. Rows show the mean proportion of each neighboring cell type per frequency of
observed nneighbors in the vicinity of MG or MDMs (nIDH WT = 9, nBrMs = 13).
(C and D) Gene set enrichment analysis (GSEA) of a T cell anergy gene set in CD4+ T cells (C) and a T cell exhaustion gene set in CD8+ T cells (D) from the MSigDB
C2 collection.
(E) Gene expression heatmap of antigen-presenting cell (APC) and T cell activating and inhibitory signaling mediators (left panels, scaled to the expression range
of variance-stabilized counts across all cell types in IDH WT glioma and BrMs) and corresponding fold changes (right panels, BrMs versus non-tumor/reference
and IDH WT glioma versus BrMs, absolute log2(fold change) > 1, adjusted p value < 0.05) in CD45 MG and MDMs and CD4+ and CD8+ T cells in IDH WT gliomas
and BrMs. Gray tiles indicate expression below the threshold (normalized counts < 10); white tiles correspond to a non-significant fold change.
(F) Scatterplot of module membership (correlation of expression to the module eigengene) and gene significance (correlation of expression to CD4+ T cell
abundance) of genes from the BrM-MDM-related gene co-expression network. Highly connected genes with immunomodulatory functions are annotated.
(G) Expression of the indicated genes in matched bulk primary breast cancer and BrM tissues using the Vareslija et al. (2019) dataset (Wilcoxon signed-rank test:
***p < 0.001, ****p < 0.0001).
See also Figure S7.

1656 Cell 181, 1643–1660, June 25, 2020


Resource
BrM samples showed pronounced accumulation of lymphocytes
and neutrophils. This indicates that tumors that arise within the
brain indeed shape their TME differently than cancers that
metastasize from extracranial sites. Moreover, when exploring
BrMs that originate from distinct primary tumors, there were
additional differences; for example, in melanoma BrM samples,
the combined abundance of CD4+ and CD8+ T cells represented
the major immune compartment, whereas breast BrM samples
showed the highest neutrophil infiltration. These key differences
in the TME landscape, which are evident only when directly
juxtaposing different brain malignancies, mirror the efficacy of
immunotherapies that show promising efficacy in melanoma pa-
tients for controlling BrMs but with very modest effects to date in
treating T cell-excluded glioblastoma (Schalper et al., 2019).
We also uncovered complex multifaceted phenotypes for
TAMs across different brain tumors that extend beyond their nu-
merical abundance. T-MG and T-MDMs showed distinct tran-
scriptomic profiles and shared expression signatures, which
are additionally influenced by the underlying disease type (IDH
mut versus IDH WT glioma versus BrMs). A T-MDM signature
derived from IDH WT gliomas, consisting of macrophage activa-
tion markers, chemokine receptors, and cytokines, proved to
also be a predictor of patient survival in IDH mut gliomas. More-
over, analyses of T-MDMs indicated that even though these re-
cruited cells have the potential to process and present antigens,
and can be located proximally to T cells in BrMs, this potential is
evidently not sufficiently utilized within the brain TME. Orthog-
onal analyses from the diverse panel of experimental assays
used in this study reveal additional insights into potential mech-
anisms of immune suppression. These included our findings that
different TAM populations produced pro-inflammatory mole-
cules, negative regulators of myeloid cell activation, factors
associated with IPRES, IDO1 and IDO2 immune checkpoint in-
hibitors, and specific ECM components and proteases that
may collectively help sculpt an immune-suppressive niche.
Therefore, therapeutic strategies that alter the multifaceted phe-
notypes of TAMs (Kowal et al., 2019), rather than aiming to sim-
ply deplete all of these cells with potentially opposing functions,
should be considerably more effective.
Looking beyond TAMs, it will also be critical to assess the roles
of neutrophils, particularly in BrMs, where we found them to be
highly abundant, because they can act as potent immune-sup-
pressive cells, as indicated by studies of other organs (Coffelt
et al., 2016). Given the highly complex and multifaceted immune
landscape of brain cancers revealed in this study, it is clear that
rational combinations of TME-targeted agents will be critical to
avoid the emergence of adaptive resistance, incorporating pre-
clinical studies to help determine optimal combinations (Quail
et al., 2016). In sum, this rich resource is available for further inter-
rogation by the research community so that we can work collec-
tively to uncover novel therapeutic strategies that unleash the po-
tential of diverse cells in the TME to combat different brain
malignancies.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
ll
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
d METHOD DETAILS
B Clinical sample processing, flow cytometry (FCM) and
fluorescence activated cell sorting (FACS)
B Tumor microenvironment-conditioned medium (TME-
CM) generation
B In vitro generation of monocyte-derived macrophages
(MDM) and TME-CM stimulation
B RNA isolation, cDNA synthesis and quantitative real-
time PCR
B Immunofluorescence staining and microscopy image
acquisition
B Image analysis and cell type identification
B Protein isolation and enzyme-linked immunosorbent
assay (ELISA)
B RNA sequencing (RNA-seq)
B Bioinformatic analysis environment
B Gene set-centered analyses
B Deconvolution of Toil-RNA sequencing data
B Leading edge metagene (LEM) analysis
B Protein-Protein-Interaction network building
B Nearest neighbor distance measurements and neigh-
borhood analysis of IF data
B Cell type abundance estimation in spatial Ivy Glioblas-
toma Atlas Project (GAP) data
B Survival analysis of the IDH wt MDM-specific gene
signature in gliomas
B Self-organizing map (SOM) clustering
B Weighted gene correlation network analysis (WGCNA)
B Expression analysis of external dataset of matched pri-
mary breast cancer and BrMs
B Plotting and graph generation
d QUANTIFICATION AND STATISTICAL ANALYSIS
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.007.
ACKNOWLEDGMENTS
We thank Prof. Ron Stoop, Dr. Nathalie Piazzon, and the Neurosurgery/Neuro-
oncology clinical and nursing teams at CHUV and MSKCC for excellent infra-
structural support; the Joyce lab members for insightful discussions; the Hegi
lab members for technical help during sample processing; and Vladimir Wisch-
newski for critical manuscript review. We thank the UNIL and MSKCC Flow Cy-
tometry Core Facilities for exceptional technical assistance, especially Romain
Bedel. Finally, we convey our immense gratitude to all patients who volun-
teered to participate in this study. Research in the Joyce lab is supported by
the Swiss Cancer League, a Swiss Bridge award, the Ludwig Institute for Can-
cer Research, the University of Lausanne, the Breast Cancer Research Foun-
dation, and Cancer Research UK. F.K. was supported in part by the German
Research Foundation (DFG, KL2491/1-1) and Fondation Medic and K.S. by
the Austrian Science Fund (FWF, J4343-B28). The results shown here are in
part based on data generated by the TCGA Research Network (https://
www.cancer.gov/tcga).
Cell 181, 1643–1660, June 25, 2020 1657
 ll
AUTHOR CONTRIBUTIONS
F.K., R.L.B., and J.A.J. designed the study. F.K., R.R.M., R.L.B., M.K., and K.S.
performed experiments and analyzed data. F.K., R.R.M., R.L.B., and S.N. per-
formed computational analyses. C.A.I.-D., C.B., V.T., P.H.G., R.T.D., and
M.E.H. provided clinical material. J.-P.B. provided histopathological reviews.
F.K. and R.R.M. prepared the figures. F.K. and J.A.J. wrote the manuscript.
All authors edited or commented on the manuscript.
DECLARATION OF INTERESTS
The authors declare no competing interests.
Received: November 29, 2019
Revised: April 1, 2020
Accepted: May 1, 2020
Published: May 28, 2020
REFERENCES
Afik, R., Zigmond, E., Vugman, M., Klepfish, M., Shimshoni, E., Pasmanik-
Chor, M., Shenoy, A., Bassat, E., Halpern, Z., Geiger, T., et al. (2016). Tumor
macrophages are pivotal constructors of tumor collagenous matrix. J. Exp.
Med. 213, 2315–2331.
Aldape, K., Brindle, K.M., Chesler, L., Chopra, R., Gajjar, A., Gilbert, M.R., Got-
tardo, N., Gutmann, D.H., Hargrave, D., Holland, E.C., et al. (2019). Challenges
to curing primary brain tumours. Nat. Rev. Clin. Oncol. 16, 509–520.
Baddeley, A.D., Eysenck, M.W., and Anderson, M.C. (2015). Spatial Point Pat-
terns: Methodology and Applications with R (London: Chapman and Hall/
CRC Press).
Barletta, K.E., Ley, K., and Mehrad, B. (2012). Regulation of neutrophil function
by adenosine. Arterioscler. Thromb. Vasc. Biol. 32, 856–864.
Bates, D., Machler, M., Bolker, B.M., and Walker, S.C. (2015). Fitting Linear
Mixed-Effects Models Using lme4. J. Stat. Softw. 67, 1–48.
Benci, J.L., Xu, B., Qiu, Y., Wu, T.J., Dada, H., Twyman-Saint Victor, C., Cu-
colo, L., Lee, D.S.M., Pauken, K.E., Huang, A.C., et al. (2016). Tumor Interferon
Signaling Regulates a Multigenic Resistance Program to Immune Checkpoint
Blockade. Cell 167, 1540–1554.e12.
Bouralexis, S., Findlay, D.M., and Evdokiou, A. (2005). Death to the bad guys:
targeting cancer via Apo2L/TRAIL. Apoptosis 10, 35–51.
Bournazos, S., Wang, T.T., and Ravetch, J.V. (2016). The Role and Function of
Fcg Receptors on Myeloid Cells. Microbiol. Spectr. 4 https://doi.org/10.1128/
microbiolspec.MCHD-0045-2016.
Bowman, R.L., Klemm, F., Akkari, L., Pyonteck, S.M., Sevenich, L., Quail, D.F.,
Dhara, S., Simpson, K., Gardner, E.E., Iacobuzio-Donahue, C.A., et al. (2016).
Macrophage Ontogeny Underlies Differences in Tumor-Specific Education in
Brain Malignancies. Cell Rep. 17, 2445–2459.
Cagney, D.N., Martin, A.M., Catalano, P.J., Redig, A.J., Lin, N.U., Lee, E.Q.,
Wen, P.Y., Dunn, I.F., Bi, W.L., Weiss, S.E., et al. (2017). Incidence and prog-
nosis of patients with brain metastases at diagnosis of systemic malignancy: a
population-based study. Neuro-oncol. 19, 1511–1521.
Ceccarelli, M., Barthel, F.P., Malta, T.M., Sabedot, T.S., Salama, S.R., Murray,
B.A., Morozova, O., Newton, Y., Radenbaugh, A., Pagnotta, S.M., et al.; TCGA
Research Network (2016). Molecular Profiling Reveals Biologically Discrete
Subsets and Pathways of Progression in Diffuse Glioma. Cell 164, 550–563.
Chen, Z., Feng, X., Herting, C.J., Garcia, V.A., Nie, K., Pong, W.W., Rasmus-
sen, R., Dwivedi, B., Seby, S., Wolf, S.A., et al. (2017). Cellular and Molecular
Identity of Tumor-Associated Macrophages in Glioblastoma. Cancer Res. 77,
2266–2278.
Chen, H.M., van der Touw, W., Wang, Y.S., Kang, K., Mai, S., Zhang, J., Alsina-
Beauchamp, D., Duty, J.A., Mungamuri, S.K., Zhang, B., et al. (2018). Blocking
immunoinhibitory receptor LILRB2 reprograms tumor-associated myeloid
cells and promotes antitumor immunity. J. Clin. Invest. 128, 5647–5662.
1658 Cell 181, 1643–1660, June 25, 2020
Resource
Chongsathidkiet, P., Jackson, C., Koyama, S., Loebel, F., Cui, X., Farber, S.H.,
Woroniecka, K., Elsamadicy, A.A., Dechant, C.A., Kemeny, H.R., et al. (2018).
Sequestration of T cells in bone marrow in the setting of glioblastoma and other
intracranial tumors. Nat. Med. 24, 1459–1468.
Coffelt, S.B., Wellenstein, M.D., and de Visser, K.E. (2016). Neutrophils in can-
cer: neutral no more. Nat. Rev. Cancer 16, 431–446.
Colaprico, A., Silva, T.C., Olsen, C., Garofano, L., Cava, C., Garolini, D., Sabe-
dot, T.S., Malta, T.M., Pagnotta, S.M., Castiglioni, I., et al. (2016). TCGAbio-
links: an R/Bioconductor package for integrative analysis of TCGA data. Nu-
cleic Acids Res. 44, e71.
Cui, Z., Liao, J., Cheong, N., Longoria, C., Cao, G., DeLisser, H.M., and Savani,
R.C. (2019). The Receptor for Hyaluronan-Mediated Motility (CD168) promotes
inflammation and fibrosis after acute lung injury. Matrix Biol. 78-79, 255–271.
Dobin, A., Davis, C.A., Schlesinger, F., Drenkow, J., Zaleski, C., Jha, S., Batut,
P., Chaisson, M., and Gingeras, T.R. (2013). STAR: ultrafast universal RNA-seq
aligner. Bioinformatics 29, 15–21.
Filková, M., Haluzı́k, M., Gay, S., and Senolt, L. (2009). The role of resistin as a
regulator of inflammation: Implications for various human pathologies. Clin.
Immunol. 133, 157–170.
Friedman, J., Hastie, T., and Tibshirani, R. (2010). Regularization Paths for
Generalized Linear Models via Coordinate Descent. J. Stat. Softw. 33, 1–22.
Gabrusiewicz, K., Rodriguez, B., Wei, J., Hashimoto, Y., Healy, L.M., Maiti,
S.N., Thomas, G., Zhou, S., Wang, Q., Elakkad, A., et al. (2016). Glioblas-
toma-infiltrated innate immune cells resemble M0 macrophage phenotype.
JCI Insight 1, e85841.
Gaujoux, R., and Seoighe, C. (2010). A flexible R package for nonnegative ma-
trix factorization. BMC Bioinformatics 11, 367.
Georgoudaki, A.M., Prokopec, K.E., Boura, V.F., Hellqvist, E., Sohn, S., Ös-
tling, J., Dahan, R., Harris, R.A., Rantalainen, M., Klevebring, D., et al.
(2016). Reprogramming Tumor-Associated Macrophages by Antibody Target-
ing Inhibits Cancer Progression and Metastasis. Cell Rep. 15, 2000–2011.
Glodde, N., Bald, T., van den Boorn-Konijnenberg, D., Nakamura, K., O’Don-
nell, J.S., Szczepanski, S., Brandes, M., Eickhoff, S., Das, I., Shridhar, N., et al.
(2017). Reactive Neutrophil Responses Dependent on the Receptor Tyrosine
Kinase c-MET Limit Cancer Immunotherapy. Immunity 47, 789–802.e9.
Godec, J., Tan, Y., Liberzon, A., Tamayo, P., Bhattacharya, S., Butte, A.J., Me-
sirov, J.P., and Haining, W.N. (2016). Compendium of Immune Signatures
Identifies Conserved and Species-Specific Biology in Response to Inflamma-
tion. Immunity 44, 194–206.
GTEx Consortium (2013). The Genotype-Tissue Expression (GTEx) project.
Nat. Genet. 45, 580–585.
Gutmann, D.H., and Kettenmann, H. (2019). Microglia/Brain Macrophages as
Central Drivers of Brain Tumor Pathobiology. Neuron 104, 442–449.
Hänzelmann, S., Castelo, R., and Guinney, J. (2013). GSVA: gene set variation
analysis for microarray and RNA-seq data. BMC Bioinformatics 14, 7.
Hendriks, L.E.L., Henon, C., Auclin, E., Mezquita, L., Ferrara, R., Audigier-Val-
ette, C., Mazieres, J., Lefebvre, C., Rabeau, A., Le Moulec, S., et al. (2019).
Outcome of Patients with Non-Small Cell Lung Cancer and Brain Metastases
Treated with Checkpoint Inhibitors. J. Thorac. Oncol. 14, 1244–1254.
Huang, Q.Q., Hossain, M.M., Wu, K., Parai, K., Pope, R.M., and Jin, J.P. (2008).
Role of H2-calponin in regulating macrophage motility and phagocytosis.
J. Biol. Chem. 283, 25887–25899.
Hugo, W., Zaretsky, J.M., Sun, L., Song, C., Moreno, B.H., Hu-Lieskovan, S.,
Berent-Maoz, B., Pang, J., Chmielowski, B., Cherry, G., et al. (2016). Genomic
and Transcriptomic Features of Response to Anti-PD-1 Therapy in Metastatic
Melanoma. Cell 165, 35–44.
Isobe, M., Izawa, K., Sugiuchi, M., Sakanishi, T., Kaitani, A., Takamori, A.,
Maehara, A., Matsukawa, T., Takahashi, M., Yamanishi, Y., et al. (2018). The
CD300e molecule in mice is an immune-activating receptor. J. Biol. Chem.
293, 3793–3805.
Jackson, C.M., Choi, J., and Lim, M. (2019). Mechanisms of immunotherapy
resistance: lessons from glioblastoma. Nat. Immunol. 20, 1100–1109.

Resource
Kai, F., Drain, A.P., and Weaver, V.M. (2019). The Extracellular Matrix Modu-
lates the Metastatic Journey. Dev. Cell 49, 332–346.
Klemm, F., and Joyce, J.A. (2015). Microenvironmental regulation of therapeu-
tic response in cancer. Trends Cell Biol. 25, 198–213.
Kobayashi, M., Chung, J.S., Beg, M., Arriaga, Y., Verma, U., Courtney, K.,
Mansour, J., Haley, B., Khan, S., Horiuchi, Y., et al. (2019). Blocking Monocytic
Myeloid-Derived Suppressor Cell Function via Anti-DC-HIL/GPNMB Antibody
Restores the In Vitro Integrity of T Cells from Cancer Patients. Clin. Cancer
Res. 25, 828–838.
Kowal, J., Kornete, M., and Joyce, J.A. (2019). Re-education of macrophages
as a therapeutic strategy in cancer. Immunotherapy 11, 677–689.
Langfelder, P., and Horvath, S. (2008). WGCNA: an R package for weighted
correlation network analysis. BMC Bioinformatics 9, 559.
Lécuyer, M.A., Saint-Laurent, O., Bourbonnière, L., Larouche, S., Larochelle,
C., Michel, L., Charabati, M., Abadier, M., Zandee, S., Haghayegh Jahromi,
N., et al. (2017). Dual role of ALCAM in neuroinflammation and blood-brain bar-
rier homeostasis. Proc. Natl. Acad. Sci. USA 114, E524–E533.
Li, L., Yan, J., Xu, J., Liu, C.Q., Zhen, Z.J., Chen, H.W., Ji, Y., Wu, Z.P., Hu, J.Y.,
Zheng, L., and Lau, W.Y. (2014). CXCL17 expression predicts poor prognosis
and correlates with adverse immune infiltration in hepatocellular carcinoma.
PLoS ONE 9, e110064.
Li, Y., Xie, Y., Hao, J., Liu, J., Ning, Y., Tang, Q., Ma, M., Zhou, H., Guan, S.,
Zhou, Q., and Lv, X. (2018). ER-localized protein-Herpud1 is a new mediator
of IL-4-induced macrophage polarization and migration. Exp. Cell Res. 368,
167–173.
Liberzon, A., Birger, C., Thorvaldsdóttir, H., Ghandi, M., Mesirov, J.P., and
Tamayo, P. (2015). The Molecular Signatures Database (MSigDB) hallmark
gene set collection. Cell Syst. 1, 417–425.
Lim, M., Xia, Y., Bettegowda, C., and Weller, M. (2018). Current state of immu-
notherapy for glioblastoma. Nat. Rev. Clin. Oncol. 15, 422–442.
Liu, Q., Johnson, E.M., Lam, R.K., Wang, Q., Ye, B.H., Wilson, E.N., Minhas,
P.S., Liu, L., Swarovski, M.S., Tran, S., et al. (2019). Peripheral TREM1 re-
sponses to brain and intestinal immunogens amplify stroke severity. Nat. Im-
munol. 20, 1023–1034.
Löffler-Wirth, H., Kalcher, M., and Binder, H. (2015). oposSOM: R-package for
high-dimensional portraying of genome-wide expression landscapes on bio-
conductor. Bioinformatics 31, 3225–3227.
Long, G.V., Atkinson, V., Lo, S., Sandhu, S., Guminski, A.D., Brown, M.P., Wil-
mott, J.S., Edwards, J., Gonzalez, M., Scolyer, R.A., et al. (2018). Combination
nivolumab and ipilimumab or nivolumab alone in melanoma brain metastases:
a multicentre randomised phase 2 study. Lancet Oncol. 19, 672–681.
Love, M.I., Huber, W., and Anders, S. (2014). Moderated estimation of fold
change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
Mantovani, A., Marchesi, F., Malesci, A., Laghi, L., and Allavena, P. (2017).
Tumour-associated macrophages as treatment targets in oncology. Nat.
Rev. Clin. Oncol. 14, 399–416.
Meng, X., Duan, C., Pang, H., Chen, Q., Han, B., Zha, C., Dinislam, M., Wu, P.,
Li, Z., Zhao, S., et al. (2019). DNA damage repair alterations modulate M2 po-
larization of microglia to remodel the tumor microenvironment via the p53-
mediated MDK expression in glioma. EBioMedicine 41, 185–199.
Mohan, V., Das, A., and Sagi, I. (2020). Emerging roles of ECM remodeling pro-
cesses in cancer. Semin. Cancer Biol. 62, 192–200.
Morianos, I., Papadopoulou, G., Semitekolou, M., and Xanthou, G. (2019). Ac-
tivin-A in the regulation of immunity in health and disease. J. Autoimmun. 104,
102314.
Müller, S., Kohanbash, G., Liu, S.J., Alvarado, B., Carrera, D., Bhaduri, A.,
Watchmaker, P.B., Yagnik, G., Di Lullo, E., Malatesta, M., et al. (2017). Sin-
gle-cell profiling of human gliomas reveals macrophage ontogeny as a basis
for regional differences in macrophage activation in the tumor microenviron-
ment. Genome Biol. 18, 234.
Murray, P.J., Allen, J.E., Biswas, S.K., Fisher, E.A., Gilroy, D.W., Goerdt, S.,
Gordon, S., Hamilton, J.A., Ivashkiv, L.B., Lawrence, T., et al. (2014). Macro-
ll
phage activation and polarization: nomenclature and experimental guidelines.
Immunity 41, 14–20.
Mushtaq, M.U., Papadas, A., Pagenkopf, A., Flietner, E., Morrow, Z., Chaudh-
ary, S.G., and Asimakopoulos, F. (2018). Tumor matrix remodeling and novel
immunotherapies: the promise of matrix-derived immune biomarkers.
J. Immunother. Cancer 6, 65.
Noy, R., and Pollard, J.W. (2014). Tumor-associated macrophages: from
mechanisms to therapy. Immunity 41, 49–61.
Olson, O.C., and Joyce, J.A. (2015). Cysteine cathepsin proteases: regulators
of cancer progression and therapeutic response. Nat. Rev. Cancer 15,
712–729.
Ortolan, E., Augeri, S., Fissolo, G., Musso, I., and Funaro, A. (2019). CD157:
From immunoregulatory protein to potential therapeutic target. Immunol.
Lett. 205, 59–64.
Puchalski, R.B., Shah, N., Miller, J., Dalley, R., Nomura, S.R., Yoon, J.G.,
Smith, K.A., Lankerovich, M., Bertagnolli, D., Bickley, K., et al. (2018). An
anatomic transcriptional atlas of human glioblastoma. Science 360, 660–663.
Pyonteck, S.M., Akkari, L., Schuhmacher, A.J., Bowman, R.L., Sevenich, L.,
Quail, D.F., Olson, O.C., Quick, M.L., Huse, J.T., Teijeiro, V., et al. (2013).
CSF-1R inhibition alters macrophage polarization and blocks glioma progres-
sion. Nat. Med. 19, 1264–1272.
Qiao, S., Qian, Y., Xu, G., Luo, Q., and Zhang, Z. (2019). Long-term character-
ization of activated microglia/macrophages facilitating the development of
experimental brain metastasis through intravital microscopic imaging.
J. Neuroinflammation 16, 4.
Quail, D.F., and Joyce, J.A. (2017). The Microenvironmental Landscape of
Brain Tumors. Cancer Cell 31, 326–341.
Quail, D.F., Bowman, R.L., Akkari, L., Quick, M.L., Schuhmacher, A.J., Huse,
J.T., Holland, E.C., Sutton, J.C., and Joyce, J.A. (2016). The tumor microenvi-
ronment underlies acquired resistance to CSF-1R inhibition in gliomas. Sci-
ence 352, aad3018.
R Core Team (2018). R: A Language and Environment for Statistical Computing
(R Foundation for Statistical Computing).
Racle, J., de Jonge, K., Baumgaertner, P., Speiser, D.E., and Gfeller, D. (2017).
Simultaneous enumeration of cancer and immune cell types from bulk tumor
gene expression data. eLife 6, e26476.
Rau, A., Gallopin, M., Celeux, G., and Jaffrézic, F. (2013). Data-based filtering
for replicated high-throughput transcriptome sequencing experiments. Bioin-
formatics 29, 2146–2152.
Richards, J., Gabunia, K., Kelemen, S.E., Kako, F., Choi, E.T., and Autieri, M.V.
(2015). Interleukin-19 increases angiogenesis in ischemic hind limbs by direct
effects on both endothelial cells and macrophage polarization. J. Mol. Cell.
Cardiol. 79, 21–31.
Ripoll, V.M., Irvine, K.M., Ravasi, T., Sweet, M.J., and Hume, D.A. (2007).
Gpnmb is induced in macrophages by IFN-gamma and lipopolysaccharide
and acts as a feedback regulator of proinflammatory responses. J. Immunol.
178, 6557–6566.
Robinson, J.T., Thorvaldsdóttir, H., Wenger, A.M., Zehir, A., and Mesirov, J.P.
(2017). Variant Review with the Integrative Genomics Viewer. Cancer Res. 77,
e31–e34.
Rodón, L., Gonzàlez-Juncà, A., Inda, Mdel.M., Sala-Hojman, A., Martı́nez-
Sáez, E., and Seoane, J. (2014). Active CREB1 promotes a malignant TGFb2
autocrine loop in glioblastoma. Cancer Discov. 4, 1230–1241.
Sankowski, R., Böttcher, C., Masuda, T., Geirsdottir, L., Sagar, Sindram, E.,
Seredenina, T., Muhs, A., Scheiwe, C., Shah, M.J., et al. (2019). Mapping mi-
croglia states in the human brain through the integration of high-dimensional
techniques. Nat. Neurosci. 22, 2098–2110.
Santana Carrero, R.M., Beceren-Braun, F., Rivas, S.C., Hegde, S.M., Gangad-
haran, A., Plote, D., Pham, G., Anthony, S.M., and Schluns, K.S. (2019). IL-15 is
a component of the inflammatory milieu in the tumor microenvironment pro-
moting antitumor responses. Proc. Natl. Acad. Sci. USA 116, 599–608.
Schalper, K.A., Rodriguez-Ruiz, M.E., Diez-Valle, R., López-Janeiro, A., Por-
ciuncula, A., Idoate, M.A., Inogés, S., de Andrea, C., López-Diaz de Cerio,
Cell 181, 1643–1660, June 25, 2020 1659
 ll
A., Tejada, S., et al. (2019). Neoadjuvant nivolumab modifies the tumor im-
mune microenvironment in resectable glioblastoma. Nat. Med. 25, 470–476.
Stupp, R., Mason, W.P., van den Bent, M.J., Weller, M., Fisher, B., Taphoorn,
M.J., Belanger, K., Brandes, A.A., Marosi, C., Bogdahn, U., et al.; European
Organisation for Research and Treatment of Cancer Brain Tumor and Radio-
therapy Groups; National Cancer Institute of Canada Clinical Trials Group
(2005). Radiotherapy plus concomitant and adjuvant temozolomide for glio-
blastoma. N. Engl. J. Med. 352, 987–996.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gil-
lette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., and Me-
sirov, J.P. (2005). Gene set enrichment analysis: a knowledge-based approach
for interpreting genome-wide expression profiles. Proc. Natl. Acad. Sci. USA
102, 15545–15550.
Szklarczyk, D., Morris, J.H., Cook, H., Kuhn, M., Wyder, S., Simonovic, M.,
Santos, A., Doncheva, N.T., Roth, A., Bork, P., et al. (2017). The STRING data-
base in 2017: quality-controlled protein-protein association networks, made
broadly accessible. Nucleic Acids Res. 45 (D1), D362–D368.
Szulzewsky, F., Arora, S., de Witte, L., Ulas, T., Markovic, D., Schultze, J.L.,
Holland, E.C., Synowitz, M., Wolf, S.A., and Kettenmann, H. (2016). Human
glioblastoma-associated microglia/monocytes express a distinct RNA profile
compared to human control and murine samples. Glia 64, 1416–1436.
Tawbi, H.A., Forsyth, P.A., Algazi, A., Hamid, O., Hodi, F.S., Moschos, S.J.,
Khushalani, N.I., Lewis, K., Lao, C.D., Postow, M.A., et al. (2018). Combined
Nivolumab and Ipilimumab in Melanoma Metastatic to the Brain. N. Engl. J.
Med. 379, 722–730.
Thapa, B., and Lee, K. (2019). Metabolic influence on macrophage polarization
and pathogenesis. BMB Rep. 52, 360–372.
Tsukamoto, H., Fujieda, K., Miyashita, A., Fukushima, S., Ikeda, T., Kubo, Y.,
Senju, S., Ihn, H., Nishimura, Y., and Oshiumi, H. (2018). Combined Blockade
of IL6 and PD-1/PD-L1 Signaling Abrogates Mutual Regulation of Their Immu-
nosuppressive Effects in the Tumor Microenvironment. Cancer Res. 78,
5011–5022.
Van, P., Jiang, W., Gottardo, R., and Finak, G. (2018). ggCyto: next generation
open-source visualization software for cytometry. Bioinformatics 34,
3951–3953.
van der Touw, W., Chen, H.M., Pan, P.Y., and Chen, S.H. (2017). LILRB recep-
tor-mediated regulation of myeloid cell maturation and function. Cancer Immu-
nol. Immunother. 66, 1079–1087.
Vareslija, D., Priedigkeit, N., Fagan, A., Purcell, S., Cosgrove, N., O’Halloran,
P.J., Ward, E., Cocchiglia, S., Hartmaier, R., Castro, C.A., et al. (2019). Tran-
scriptome Characterization of Matched Primary Breast and Brain Metastatic
Tumors to Detect Novel Actionable Targets. J. Natl. Cancer Inst. 111, 388–398.
1660 Cell 181, 1643–1660, June 25, 2020
Resource
Venteicher, A.S., Tirosh, I., Hebert, C., Yizhak, K., Neftel, C., Filbin, M.G., Hov-
estadt, V., Escalante, L.E., Shaw, M.L., Rodman, C., et al. (2017). Decoupling
genetics, lineages, and microenvironment in IDH-mutant gliomas by single-
cell RNA-seq. Science 355, eaai8478.
Vivian, J., Rao, A.A., Nothaft, F.A., Ketchum, C., Armstrong, J., Novak, A., Pfeil,
J., Narkizian, J., Deran, A.D., Musselman-Brown, A., et al. (2017). Toil enables
reproducible, open source, big biomedical data analyses. Nat. Biotechnol. 35,
314–316.
Wei, S.C., Duffy, C.R., and Allison, J.P. (2018). Fundamental Mechanisms of
Immune Checkpoint Blockade Therapy. Cancer Discov. 8, 1069–1086.
Wickham, H. (2016). ggplot2: Elegant Graphics for Data Analysis (Springer).
Xu, P., Zhang, X., Liu, Q., Xie, Y., Shi, X., Chen, J., Li, Y., Guo, H., Sun, R.,
Hong, Y., et al. (2019). Microglial TREM-1 receptor mediates neuroinflamma-
tory injury via interaction with SYK in experimental ischemic stroke. Cell Death
Dis. 10, 555.
Xue, J., Schmidt, S.V., Sander, J., Draffehn, A., Krebs, W., Quester, I., De
Nardo, D., Gohel, T.D., Emde, M., Schmidleithner, L., et al. (2014). Transcrip-
tome-based network analysis reveals a spectrum model of human macro-
phage activation. Immunity 40, 274–288.
Yan, D., Kowal, J., Akkari, L., Schuhmacher, A.J., Huse, J.T., West, B.L., and
Joyce, J.A. (2017). Inhibition of colony stimulating factor-1 receptor abrogates
microenvironment-mediated therapeutic resistance in gliomas. Oncogene 36,
6049–6058.
Yang, T.H., St John, L.S., Garber, H.R., Kerros, C., Ruisaard, K.E., Clise-
Dwyer, K., Alatrash, G., Ma, Q., and Molldrem, J.J. (2018). Membrane-Associ-
ated Proteinase 3 on Granulocytes and Acute Myeloid Leukemia Inhibits T Cell
Proliferation. J. Immunol. 201, 1389–1399.
Young, M.D., Wakefield, M.J., Smyth, G.K., and Oshlack, A. (2010). Gene
ontology analysis for RNA-seq: accounting for selection bias. Genome Biol.
11, R14.
Zaiss, D.M.W., Gause, W.C., Osborne, L.C., and Artis, D. (2015). Emerging
functions of amphiregulin in orchestrating immunity, inflammation, and tissue
repair. Immunity 42, 216–226.
Zhai, L., Ladomersky, E., Lenzen, A., Nguyen, B., Patel, R., Lauing, K.L., Wu,
M., and Wainwright, D.A. (2018). IDO1 in cancer: a Gemini of immune check-
points. Cell. Mol. Immunol. 15, 447–457.
Zhao, J., Sun, L., and Li, X. (2017). Commanding CNS Invasion: GM-CSF. Im-
munity 46, 165–167.
Zhao, Y., Lee, C.K., Lin, C.H., Gassen, R.B., Xu, X., Huang, Z., Xiao, C., Bonor-
ino, C., Lu, L.F., Bui, J.D., et al. (2019). PD-L1:CD80 Cis-Heterodimer Triggers
the Co-stimulatory Receptor CD28 While Repressing the Inhibitory PD-1 and
CTLA-4 Pathways. Immunity 51, 1059–1073.e9.
 ll
Resource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
FCM: AF700 mouse monoclonal anti-human CD45 BioLegend Cat#304024; RRID:AB_493761
(clone HI30)
FCM: BV421 rat monoclonal anti-mouse/human CD11B BioLegend Cat#101251; RRID:AB_2562904
(clone M1/70)
FCM: PE mouse monoclonal anti-human CD66B BioLegend Cat#305106; RRID:AB_2077857
(clone G10F5)
FCM: AF488 mouse monoclonal anti-human CD14 BioLegend Cat#325610; RRID:AB_830683
(clone HCD14)
FCM: BUV737 mouse monoclonal anti-human CD16 BD Cat#612786; RRID:AB_2833077
(clone 3G8)
FCM: APC mouse monoclonal anti-human CD49D BioLegend Cat#304308; RRID:AB_2130041
(clone 9F10)
FCM: BV605 mouse monoclonal anti-human CD11C BioLegend Cat#301636; RRID:AB_2563796
(clone 3.9)
FCM: BV711 mouse monoclonal anti-human anti HLA-DR BioLegend Cat#307644; RRID:AB_2562913
(clone L243)
FCM: PerCP/Cy5.5 mouse monoclonal anti-human CD3 BioLegend Cat#300328; RRID:AB_1575008
(clone HIT3a)
FCM: BV 650 mouse monoclonal anti-human anti CD4 BioLegend Cat#317436; RRID:AB_2563050
(clone OKT4)
FCM: PE mouse monoclonal anti-human CD25 (clone BC96) BioLegend Cat#302606; RRID:AB_314276
FCM: BV510 mouse monoclonal anti-human CD127 (clone BioLegend Cat#351332; RRID:AB_2562304
A019D5)
FCM: PE/Cy7 mouse monoclonal anti-human CD8A BioLegend Cat#300914; RRID:AB_314118
(clone HIT8a)
FCM: BUV563 mouse monoclonal anti-human CD20 BD Cat#748456
(clone 2H7)
FCM: BUV563 mouse monoclonal anti-human CD19 (clone BD Cat#612916
SJ25C1)
FCM: PE/Dazzle mouse monoclonal anti-human CD56 BioLegend Cat#318348; RRID:AB_2563564
(clone HDC56)
FCM: PE mouse monoclonal anti-human P2RY12 (clone BioLegend Cat#392103; RRID:AB_2716006
S16001E)
FCM: PE/Cy7 Mouse monoclonal anti-human CD68 (clone BioLegend Cat#333816; RRID:AB_2562936
Y1/82A)
IF: Mouse monoclonal anti-human CD68 (clone KP1), 1:100 Abcam Cat#ab955; RRID:AB_307338
dilution
IF: Rat monoclonal anti-human CD49D (clone PS/2), 1:100 Abcam Cat#ab25247
dilution
IF: Rabbit polyclonal anti-human P2RY12, 1:600 dilution Sigma-Aldrich Cat#HPA014518; RRID:AB_2669027
IF: Goat polyclonal anti-human CD45, 1:100 dilution LSBio Cat#LS-B14248-300
IF: AF488 mouse monoclonal anti-human CD45 (clone HI30), BioLegend Cat#304019; RRID:AB_493033
1:100 dilution
IF: AF488 mouse monoclonal anti-human CD3 (clone UCHT1), BioLegend Cat#300406; RRID:AB_314060
1:100 dilution
IF: Sheep polyclonal anti-human CD31, 1:200 dilution R&D Cat#AF806; RRID:AB_355617
IF: APC rat monoclonal anti Ki-67 (clone SolA15), 1:100 Thermo Fisher Scientific Cat#17-5698-82
dilution
(Continued on next page)

Cell 181, 1643–1660.e1–e7, June 25, 2020 e1

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
IF: AF555 donkey anti-rabbit IgG 1:1000 dilution Thermo Fisher Scientific Cat#A31572, RRID:AB_162543
IF: AF555 donkey anti-mouse IgG, 1:500 dilution Thermo Fisher Scientific Cat#A32773; RRID:AB_2762848
IF: AF488 donkey anti-rat IgG, 1:500 dilution Thermo Fisher Scientific Cat#A21208; RRID:AB_141709
IF: AF647 donkey anti-rat IgG, 1:500 dilution abcam Cat#ab150155; RRID:AB_2813835
IF: DyLight755 donkey anti-goat IgG, 1:500 dilution Thermo Fisher Scientific Cat# SA5-10091; RRID:AB_2556671
IF: AF555 donkey anti-sheep IgG, 1:500 dilution Thermo Fisher Scientific Cat#A21436; RRID:AB_2535857
Biological Samples
Non-tumor, glioma and brain metastasis tissue Centre Hospitalier Universitaire N/A
Vaudois, Lausanne, Switzerland
Non-tumor, glioma and brain metastasis tissue Memorial Sloan Kettering N/A
Cancer Center, New York,
NY, USA
Healthy donor blood Transfusion Interrégionale N/A
Croix-Rouge Suisse, Epalinges,
Switzerland
Healthy donor blood New York Blood Bank, N/A
New York, NY, USA
Chemicals, Peptides, and Recombinant Proteins
DMEM-F12 (1:1), GlutaMAX GIBCO Cat#31331028
DMEM, high glucose, GlutaMAX, pyruvate GIBCO Cat#31966021
Penicillin/Streptomycin GIBCO Cat#15140122
Human recombinant CSF-1 R&D Systems Cat#216-MC-025
Ficoll-Paque Premium GE Cat#17-5442-02
Trizol Thermo Fisher Scientific Cat#15596018
Trizol LS Thermo Fisher Scientific Cat#10296028
Tween 20 Applied Chemicals Cat#A4974
Triton X-100 Applied Chemicals Cat#A4975
TNB Blocking Reagent Perkin Elmer Cat#FP1020
Fluorescence Mounting Medium Dako Cat#S302380
Critical Commercial Assays
Brain Tumor Dissociation Kit (P) Miltenyi Cat#130-095-942
Tumor Dissociation Kit, human Miltenyi Cat#130-095-929
Myelin Removal Beads Miltenyi Cat#130-096-733
CD14 MicroBeads, human Miltenyi Cat#130-050-201
Human TruStain FcX BioLegend Cat#422302
ZombieNIR Fixable Viability Kit BioLegend Cat#423106
High Capacity cDNA Reverse Transcription Kit Applied Biosystems Cat#4368814
TaqMan Universal PCR Master Mix Applied Biosystems Cat#4304437
Quantibody Array Q4000 ELISA Raybiotech Cat#QAH-CAA-4000-1
Deposited Data
RNAseq count data This paper https://joycelab.shinyapps.io/
braintime/
Human reference genome, hg38 Genomics Data Common https://gdc.cancer.gov/about-
data/data-harmonization-and-
generation/gdc-reference-files
TCGA LGG and GBM datasets Genomics Data Common https://portal.gdc.cancer.gov/
TOIL TGCA TARGET GTEx datasets Vivian et al., 2017 https://xenabrowser.net/datapages/
Ivy Glioblastoma Atlas Project RNA sequencing fata Puchalski et al., 2018 https://glioblastoma.alleninstitute.
org/static/download.html
(Continued on next page)

e2 Cell 181, 1643–1660.e1–e7, June 25, 2020

 ll
Resource

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
STRING Protein-Protein-Interaction database, version 10.5 Szklarczyk et al., 2017 https://version-10-5.string-db.org/
cgi/download.pl
Molecular Signatures Database gene set collection Liberzon et al., 2015; https://www.gsea-msigdb.org/
Subramanian et al., 2005 gsea/msigdb/
RNA sequencing count matrix from matched breast cancer Vare
slija et al., 2019 https://github.com/npriedig
primaries and brain metastases
Oligonucleotides
See Table S7 N/A
Software and Algorithms
FlowJo, version 10.4 BD https://www.flowjo.com/
BBDuk, version 38.12 Joint Genome Institute https://jgi.doe.gov/data-and-tools/
bbtools/
STAR aligner, version 2.5.2b Dobin et al., 2013 https://github.com/alexdobin/STAR
R environment, version 3.5.0 R Core Team, 2018 https://www.r-project.org/
VIS Image Analysis, version 2019.7 Visiopharm https://www.visiopharm.com/
Other
gentleMACS Octo Dissociator Miltenyi Cat#130-095-937
gentleMACS C Tubes Miltenyi Cat#130-096-334
LS Columns Miltenyi Cat#130-042-401
SepMate-50 StemCell Cat#85450
PermaLife Cell Culture Bags OriGen Biomedical Cat#PL30-2G
LSR II flow cytometer BD N/A
Fortessa flow cytometer BD N/A
FACSAria III, flow cytometer & cell sorter BD N/A
Axio Scan.Z1 slide scanner Zeiss N/A
QuantStudio 6 Flex Applied Biosystems N/A
Omni Tissue Homogenizer (TH) Omni International Cat#TH220

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources should be directed to the Lead Contact, Johanna Joyce (johanna.joyce@unil.ch).

Materials Availability
This study did not generate new unique reagents.

Data and Code Availability

RNA-seq count expression data generated during this study can be visualized and downloaded at https://joycelab.shinyapps.io/
braintime/. Due to patient privacy protection, the raw RNA-seq data will be made available upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional
and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical
standards.
Informed consent was obtained from all individual participants included in this study. The collection of non-tumor and tumor tissue
samples at the Centre Hospitalier Universitaire Vaudois (CHUV, Lausanne, Switzerland) was approved by the Commission cantonale
d’éthique de la recherche sur l’être humain (CER-VD, protocol PB 2017-00240, F25 / 99). Sample collection at Memorial Sloan Ket-
tering Cancer Center (MSKCC, New York, NY, USA) was approved by the institutional review board (IRB, protocols #IRB #06-107,
#14-230). Non-tumor samples of cerebral cortex tissues were collected at CHUV during medically indicated surgical treatment of

Cell 181, 1643–1660.e1–e7, June 25, 2020 e3

 ll
Resource

refractory epilepsy patients, and at MSKCC in normal brain distant from the tumor in patients with low-grade glioma or from post-
mortem samples collected through the rapid autopsy program with no history of brain malignancy.
Tissue specimens were immediately collected from the operating room and processed as described below. All patient-related data
and unique identifiers were removed so that human samples were anonymized before any further processing.
Pathological review and molecular analysis of tumor samples was performed as part of standard clinical care at the respective
locations (CHUV or MSKCC). In all glioma samples subjected to RNA sequencing, the IDH1 and IDH2 mutation status was verified
by inspection of the reads from the CD45- population aligning to the IDH1 and IDH2 loci with the Integrative Genomics Viewer (IGV;
Robinson et al., 2017). For immunofluorescence sections the tumor diagnosis was confirmed independently, for all non-tumor sam-
ples, the absence of malignancy was equally confirmed by a pathologist.
Peripheral blood and buffy coats were obtained from the Transfusion Interrégionale, Croix-Rouge Suisse (Epalinges/Lausanne,
Switzerland), the New York Blood Center (New York, NY, USA), and healthy donors.

METHOD DETAILS

Clinical sample processing, flow cytometry (FCM) and fluorescence activated cell sorting (FACS)
Tissue specimens were washed in HBSS and macro-dissected under sterile conditions. Parts of the tissue were either immediately
frozen by submerging the sample in liquid nitrogen-cooled 2-methyl butane (Sigma-Aldrich) or OCT-embedded (Tissue-Tek) before
freezing for subsequent sectioning and immunofluorescence staining. OCT embedding was performed by placing the sample in a
freezing mold filled with OCT and then submerging the mold in 2-methyl butane cooled with dry ice.
The remaining tissue was further processed with either the Brain Tumor Dissociation Kit (Miltenyi) for non-tumor tissue and gli-
omas, or the Tumor Dissociation Kit for BrMs (Miltenyi) using the gentleMACS Octo Dissociator (Miltenyi). Myelin debris in cell sus-
pensions from non-tumor and glioma tissues was removed by incubating the cells with Myelin Removal Beads (Miltenyi) and mag-
netic-activated cell sorting (MACS) using LS columns (Miltenyi) according to the manufacturer’s instructions. All tissue suspensions
were filtered through a 40 mm filter and underwent red blood cell lysis (BioLegend). Single cell suspensions were stained with a fixable
live-dead stain (Zombie NIR, BioLegend), FC-blocked for 10 min (Human TruStain FcX, BioLegend) and then incubated with direct
fluorophore-conjugated antibodies for 20 min at 4 C. All FCM antibodies were titrated in a lot-specific manner. Antibody details are
listed in the Key Resources Table. Cells were washed with PBS +2% fetal bovine serum (FBS) +0.5 mM EDTA and stored at 4 C in the
dark until FAC-sorting.
All FCM acquisition was completed on either a BD Fortessa or a BD LSR II device (BD), and cell sorting was performed on a
FACSAria III (BD) using FACSDiva (BD). Cells were sorted directly into Trizol LS (Thermo Fisher Scientific) and immediately snap
frozen with liquid nitrogen. Analysis of FCM data was performed with FlowJo (BD).

Tumor microenvironment-conditioned medium (TME-CM) generation

Single cell suspensions from whole tumor samples were resuspended in DMEM-F12 (1:1) +Glutamax (GIBCO) +10% FBS +1% peni-
cillin/streptomycin (P/S, GIBCO) and adjusted to a concentration of 2 3 106 cells/ml with 2 ml plated into each well of a 6-well plate
(TPP). The supernatant of these tissue cultures, containing cancer cells, immune cells etc. from the complex brain TME, was har-
vested at 24 hours after initial seeding, spun down to remove debris (300 g, 10 min) and stored at 80 C until further use.

In vitro generation of monocyte-derived macrophages (MDM) and TME-CM stimulation

Peripheral blood mononuclear cells were isolated from buffy coats of healthy donors with a Ficoll (GE) gradient using SepMate tubes
(StemCell) and monocytes selected by MACsorting with CD14 MicroBeads (Miltenyi). Monocytes were differentiated into macro-
phages by culture in Teflon-coated bags (OriGen) for 7 days in DMEM +GlutaMAX (GIBCO) +10% FBS +1% P/S with the addition
of 10 ng/ml recombinant human CSF-1 (R&D Systems).
Differentiated MDMs were plated at a density of 1 3 106 cells/well of a 6-well plate in DMEM +10% FBS +1% P/S +10 ng/ml CSF-1.
After cell attachment, MDMs were cultured in serum free medium for 6 hours before stimulation with TME-CM for 24 hours.

RNA isolation, cDNA synthesis and quantitative real-time PCR

TME-CM-stimulated MDMs were lysed with Trizol (Thermo Fisher Scientific), RNA was purified with Direct-zol columns (Zymo
Research), DNase treated and 1.0 mg of RNA was used for cDNA synthesis using the High Capacity cDNA Reverse Transcription
Kit (Applied Biosystems). An amount of cDNA equivalent to 5 ng total RNA was used for real-time PCR. For primer and probe details
see Table S6. Assays were run in triplicate on a QuantStudio 6 Flex instrument (Applied Biosystems) using the TaqMan Universal PCR
Master Mix (Applied Biosystems) and expression was normalized to the average expression of Ubiquitin C (UBC) and Ribosomal Pro-
tein L19 (RPL19) for each sample.

Immunofluorescence staining and microscopy image acquisition

10 mm cryostat sections were thawed, air-dried and fixed with ice-cold 100% methanol for 5 minutes. After rehydration with PBS,
sections were washed twice in PBS +0.2% Tween 20 (Applied Chemicals), permeabilized with PBS +0.2% Triton X-100 (Applied
Chemicals) for 3 hours and washed again with PBS +0.2% Tween 20. Blocking was performed with PBS +0.5% Tween 20 +1%

e4 Cell 181, 1643–1660.e1–e7, June 25, 2020

 ll
Resource

TNB Blocking Reagent (Perkin Elmer), followed by incubation with primary antibody in the same buffer overnight at 4 C. Primary anti-
body information and dilutions are listed in the Key Resources table. Sections were washed with PBS +0.2% Tween 20 before incu-
bation with fluorophore-conjugated secondary antibodies at a dilution of 1:500 in PBS +0.5% Tween 20 +1% TNB Blocking
Reagent +1 mg/ml DAPI at room temperature. Directly-conjugated primary antibodies were employed where indicated after an initial
round of primary and secondary antibody staining, to avoid potential for cross reactivity. Finally, sections were washed with
PBS +0.2% Tween and mounted with Fluorescence Mounting Medium (Dako).
Stained tissue sections were imaged with an Axio Scan.Z1 slide scanner (Zeiss) equipped with a Colibri 7 LED light source (Zeiss)
using a Plan-Apochromat 20x/0.8 DIC M27 coverslip-corrected objective (Zeiss). All slides from the same staining panel were digi-
talized using identical acquisition settings.

Image analysis and cell type identification

Image quantification was performed using the VIS Image Analysis software (Visiopharm). For each staining panel a specific applica-
tion was created using the software’s authoring module. The tissue outline was detected after applying a 21 pixel mean filter. The
edges of the derived regions of interest were smoothened with the built-in function ‘‘close’’ and holes in the mask were filled using
the ‘‘fill holes’’ command. Aberrant signals resulting from e.g., dust particles, tissue folds or air bubbles were manually excluded from
these regions of interest. Nuclear classification was based on the watershed signal of the DAPI staining and filtered by area to exclude
incomplete nuclei. The obtained nuclear label was expanded by 5 pixels to capture both nuclear and adjacent cytoplasmic fluores-
cent signal. Cell types were identified using a hierarchical decision tree with manually set thresholds. Finally, a representation of the
cytoplasm was created using the inbuilt growth algorithm with a maximum distance of 15 pixels from the nucleus. Vessel segmen-
tation was performed by creating a separate classifier based on pixel intensity of the CD31 signal. Nuclear classifiers were excluded a
priori and incorporated in the vessel label only when exceeding the threshold for CD31. Perivascular niches (PVNs) were established
by generating an ROI around vessels at a distance of 20 mm. All object-based phenotyping result tables were exported as csv files for
further analysis within the R environment.

Protein isolation and enzyme-linked immunosorbent assay (ELISA)

Frozen tissues were weighed and homogenized on ice with an Omni Tissue Homogenizer (Omni International) in 10 mL of RIPA lysis
buffer (Thermo Fisher Scientific) +cOmplete Protease Inhibitor (Roche Diagnostics) per mg of tissue. The homogenate was gently
agitated on ice for 10 minutes, centrifuged at 10.000 g for 5 minutes at 4 C and the supernatant collected. The protein concentration
was determined using a Bradford assay (Bio-Rad) and adjusted to 1 mg/ml. Samples were shipped to Raybiotech (Peachtree Corners)
for quantitative analysis with the multiplexed Quantibody Array Q4000 ELISA.

RNA sequencing (RNA-seq)

RNA was isolated by chloroform extraction and isopropanol precipitation. RNA sequencing libraries were generated with the
SMART-Seq preparation kit (CloneTech) and fragmented with the Nextera XT kit (Illumina). Paired end, 100 or 150 base pair, and
single end, 100 base pair, sequencing was performed by Genewiz (South Plainfield, New Jersey, USA) on an Illumina HiSeq 2500
(Illumina).
Reads were adaptor trimmed and quality clipped using BBDuk (version 38.12; https://sourceforge.net/projects/bbmap/). Trimmed
reads were mapped to the Genomic Data Commons (GDC) GRCh38.d1.vd1 reference sequence using the STAR aligner (version
2.5.2b, Dobin et al., 2013) in two-pass mode with parameters corresponding to the GDC RNA-seq alignment workflow. Transcript
abundance was estimated using the corresponding GDC reference gtf file. A raw count matrix was produced and differential
gene expression was assessed with DESeq2 using an absolute log2 fold change of 1 and a false discovery rate of 0.01 when con-
trasting to reference samples, and 0.05 for within tumor contrasts (Love et al., 2014).

Bioinformatic analysis environment

All bioinformatic analyses were performed within the R environment (version 3.5.0, R Core Team 2018).

Gene set-centered analyses

The Molecular Signatures Database (MSigDB, version 6.1, Liberzon et al., 2015; Subramanian et al., 2005) was used as the main
source for gene set-based analyses.
Over-representation was assessed with the goseq R package (Young et al., 2010) for differentially expressed genes to correct for
gene length bias, otherwise the hypergeometric test was employed. For individual samples, gene set enrichment was estimated with
the Gene Set Variation Analysis R package (GSVA, Hänzelmann et al., 2013) using the ‘‘gsva’’ function. Gene set enrichment analysis
(GSEA) was evaluated with the R package fgsea (https://github.com/ctlab/fgsea) using the maximum likelihood log fold changes
determined by DESeq2 as the ranking metric.

Deconvolution of Toil-RNA sequencing data

Toil-processed (Vivian et al., 2017), DESeq2-standardized gene expression data and matching phenotype data from the TCGA and
Genotype-Tissue Expression Project (GTEx) databases were downloaded from the UCSC Xena platform and filtered to include only

Cell 181, 1643–1660.e1–e7, June 25, 2020 e5

 ll
Resource

low-grade glioma ‘‘TCGA-LGG’’ and high-grade "TCGA-GBM’’ and ‘‘frontal cortex’’ GTEx samples to integrate bulk glioma expres-
sion data with unmatched non-tumor samples. MG- and MDM-specific marker genes were derived by identifying differentially ex-
pressed genes in these two populations versus all other sorted populations in a pairwise fashion, determining the intersect and
ranking the resulting genes by their fold change versus the CD45- population. The 20 highest ranked genes were then used as
cell type-specific marker genes. Deconvolution of MG- and MDM-proportions in tumor and non-tumor sample expression data
was done with the EPIC R package (Racle et al., 2017) using these marker genes and providing the expression data from the sorted
populations as reference profiles. As the exact amount of RNA within the estimated cell types is not known, this parameter was set to
1 when running the deconvolution.

Leading edge metagene (LEM) analysis

To capture biologically meaningful patterns of gene expression within the differentially expressed genes the LEM approach (Godec
et al., 2016) was employed: (a) GSEA was performed using the MSigDB C7 collection as described above, (b) the leading edge genes
of the significant gene sets were arranged into a genes by gene sets matrix with the shrunken fold changes as the entries, (c) this
matrix was clustered using non-negative matrix factorization with the R package NMF (Gaujoux and Seoighe, 2010), (d) genes
with a small coefficient in each metagene were filtered based on the 95th quantile of a fitted exponential distribution of the coefficients
and (e) each gene with a coefficient above the threshold was assigned to the metagene where it had the highest coefficient.

Protein-Protein-Interaction network building

Version 11.0 of the STRING database (Szklarczyk et al., 2017) was downloaded from the consortium’s website and gene identifiers
from RNA-seq were mapped to Ensembl Protein IDs using the provided accessory data. The resulting interaction data was filtered to
contain only interactions with a high confidence STRING combined score (i.e., > 700). For network layout calculation the combined
score was used as an edge weight.

Nearest neighbor distance measurements and neighborhood analysis of IF data

Nearest neighbor distances from MG and MDM to vessels in IDH wt glioma samples were calculated using the spatstat R package
(Baddeley et al., 2015). Statistical significance was assessed by fitting a mixed effects model with the cell type as the fixed effect, and
the clinical sample ID as the random effect using the R package lme4 (Bates et al., 2015).
Neighbors for each individual cell were determined based on their occurrence within a range of 5 mm outside of the radius of the cell
(calculated based on the area). This was used to tabulate the number of neighbors and their cell type for each cell within the tissue
section.

Cell type abundance estimation in spatial Ivy Glioblastoma Atlas Project (GAP) data
The micro-dissected Ivy GAP (Puchalski et al., 2018) RNA-seq RSEM count data and sample annotation containing anatomical loca-
tion were downloaded from the Ivy GAP website (https://glioblastoma.alleninstitute.org/static/download.html) and normalized using
DESeq2. The relative abundance of cell types was estimated by deriving marker genes through a multinomial logistic regression
model on the normalized expression data of the FAC-sorted cell types of interest in IDH WT tumors and then computing the
GSVA enrichment scores in the Ivy GAP samples.

Survival analysis of the IDH wt MDM-specific gene signature in gliomas

The harmonized TCGA low-grade and high-grade HTSeq hg38 count data and clinical data was accessed from the GDC repository
using the TCGAbiolinks R package (Colaprico et al., 2016). Datasets were pre-processed to remove outliers and normalized using the
functions provided by TCGAbiolink before merging. Subsequent analyses were performed including only samples where an anno-
tation of the IDH mutation status was available. Cell type-specific gene signatures were derived by training a multinomial logistic
regression model with an elastic-net penalty to separate between MG and MDMs along IDH status with the ‘‘glmnet’’ R package
(Friedman et al., 2010). A mean-centered expression matrix of all MG and MDMs expression data in gliomas and BrMs, subset by
genes that were upregulated in tumors versus non-tumor tissue or healthy controls, served as the input matrix. The strength of
the penalty was determined by a 10-fold cross-validation of the l parameter. For survival analysis, GSVA enrichment scores of these
cell type-specific gene signatures were estimated and used to divide samples into tertiles. Kaplan-Meier survival curves were
computed using the ‘‘survfit’’ function. Survival curves were compared with a log-rank test between the individual levels and multi-
variate Cox regression analysis was performed with the ‘‘coxph’’ function.

Self-organizing map (SOM) clustering

Variance stabilized counts from sorted populations of interest from IDH mut, IDH WT glioma and BrM samples were filtered with the R
package HTSFilter (Rau et al., 2013) to ensure removal of genes with a low, constant expression. The resulting matrix of genes and
samples was used as input for the SOM neural network building, which was performed with the oposSOM R package (Löffler-Wirth
et al., 2015) with a map space of 50 3 50. To investigate associations between the sample phenotype and the SOM metagenes, the
tumor type and cell type were provided as group labels.

e6 Cell 181, 1643–1660.e1–e7, June 25, 2020

 ll
Resource

Weighted gene correlation network analysis (WGCNA)

The WGCNA (Langfelder and Horvath, 2008) R package was used to identify co-regulated genes associated with a MG- or MDM-BrM
phenotype. A variance stabilized, batch-corrected count matrix of MG and MDM samples was filtered with the R package HTSFilter
(Rau et al., 2013) yielding input expression data with 15826 genes and 56 samples. WGCNA standard parameters were changed as
follows: the soft-thresholding power was raised to 7, the minModuleSize was increased to 50, ‘‘bicor’’ was used to calculate the cor-
relation, the network type was set to ‘‘signed hybrid’’ and a dendrogram cut height of 0.25 was used for module merging. This yielded
20 modules whose eigengene, i.e., the first principal component, was tested for correlation to the provided sample information, i.e.,
tumor- and cell-type and abundance as determined by FCM.

Expression analysis of external dataset of matched primary breast cancer and BrMs
RNA-seq raw count data from patient-matched primary breast tumors and corresponding BrMs (Vare slija et al., 2019) were down-
loaded (https://github.com/npriedig/jnci_2018) and transformed using DESeq2. The statistical significance of gene expression
changes between primary tumors and BrMs was assessed with a two-tailed Wilcoxon signed-rank test on the variance-stabilized
counts.

Plotting and graph generation

Plots were created using the ggplot2 R package (Wickham, 2016) and the ggpubr (https://cran.r-project.org/web/packages/
ggpubr/), survminer (https://cran.r-project.org/web/packages/survminer/), ggraph (https://cran.r-project.org/web/packages/
ggraph/) and ggcyto extensions (Van et al., 2018). Annotated heatmaps were drawn with the pheatmap R package (https://cran.
r-project.org/web/packages/pheatmap/).

QUANTIFICATION AND STATISTICAL ANALYSIS

Summary data are presented as mean ± standard error of the mean (SEM) or Tukey boxplots using ‘‘ggplot2.’’ Numerical data was
analyzed using the statistical tests noted within the corresponding sections of the article. Hierarchical clustering was performed using
Ward’s method with 1-Pearson correlation coefficient as the distance metric unless noted otherwise. P values were annotated as
follows: * < 0.05, ** < 0.01, *** < 0.001, **** < 0.0001, ns > 0.05.

Cell 181, 1643–1660.e1–e7, June 25, 2020 e7

 ll
Resource

Supplemental Figures

Figure S1. FACS of Cell Populations and RNA-Seq, Related to Figure 1

(A) Flow cytometry (FCM) plots illustrating the gating strategy employed during FAC-sorting of immune cell populations in non-tumor and tumor tissue (for cell
type markers, see Table S2). (B) tSNE plot of gene expression data (500 most variable genes) from all sorted cell populations (n = 226) across the complete clinical
cohort (MG = microglia, MDM = monocyte-derived macrophages, reference = unmatched healthy blood and in vitro generated MDMs).
See also Table S2.
ll
Resource

(legend on next page)

 ll
Resource

Figure S2. MG and MDM Marker Expression, Related to Figure 2

(A) Normalized counts (log10 transformed) of MG and MDM marker genes in sorted CD49Dlow MG and CD49Dhigh MDM populations across both non-tumor and
tumor tissues (reference = healthy donor in vitro generated MDMs). (B) Percentage of CD49Dlow MG and CD49Dhigh MDMs positive for P2RY12 and CD68 as
determined by FCM in relation to the total number of MG/MDMs in non-tumor (n = 8) and tumor tissue (nIDH mut = 6, nIDH WT = 6, nBrM = 9). (C) Single channel and
merged immunofluorescence (IF) images of CD45, CD68, P2RY12 and CD49D stainings which were employed to delineate MG and MDMs. The last column
shows the resulting Visiopharm cell type assignments for quantitative analyses (MG (CD45+, P2RY12+/CD68+, CD49D-), MDM (CD45+, P2RY12+/CD68+,
CD49D+), non-immune cells (CD45-) and non-TAM-immune cells (CD45+, P2RY12-/CD68-, CD49D-/+). Scale bars represent 100mm. (D) Scatterplots of the
abundance of MG and MDMs as determined by IF versus FCM in non-tumor (n = 4) and tumor tissues (nIDH mut = 13, nIDH WT = 14, nBrM = 18) processed
independently from the same individual samples. Pearson’s correlation coefficient and significance are indicated at the top of each plot. (E) Heatmap of human
MG- and MDM-specific gene set expression used for deconvolution across FAC-sorted population samples from all disease types.
ll
Resource

(legend on next page)

 ll
Resource

Figure S3. Analysis of DEGs and TAM Activation Patterns, Related to Figure 3
(A) Summary of contrasts applied when performing differential gene expression (DEG) analysis in MG and MDMs in gliomas (regardless of IDH status) and BrMs
(from all primaries) in comparison to normal controls (non-tumor brain MG and in vitro differentiated MDMs respectively) with the corresponding log2(fold-change)
versus -log10(adjusted p value) volcano plots. (B) Euler plot of the number of differentially expressed genes (DEG, log2(fc) > 1, p.adj < 0.01) that overlap in MG and
MDMs as shown in (A). (C) Molecular Signatures Database (MSigDB) ‘‘Hallmark’’ gene set collection overrepresentation analysis (ORA) in genes upregulated in
both gliomas and BrMs versus non-tumor brain tissue or healthy donors in MDMs and MG in MDMs and MG. Dot sizes reflect the fraction of gene set members
found within the analyzed DEGs, and dot color indicates cell type. (D) Heatmap of fold changes of macrophage M1 and M2 polarization marker genes (absolute
log2(fc) > 1, p.adj < 0.05) in MDMs and MG in gliomas and BrMs. Blank tiles indicate the lack of significant fold change. Genes are annotated with their canonical
stimuli and the associated polarization phenotype. (GC = glucocorticoid, Ic = immune complexes, IFNg = Interferon gamma, IL10 = interleukin 10, IL4 = interleukin
4, LPS = lipopolysaccharide, TGFb = transforming growth factor beta). (E) Overlap between leading edge metagenes (LEMs) in MG and MDMs in gliomas and
BrMs. Tile fill color indicates significance of overlap determined by hypergeometric testing (-log10(p.adj)). (F) String-DB protein-protein-interaction network of the
intersect from IFN Type-1 group 2 modules from LEMs ‘‘BrM-MG 1,’’ ‘‘Glioma MDM 1’’ and ‘‘BrM-MDM 4.’’ Genes selected for validation through qRT-PCR are
highlighted in red (corresponding data shown in Figure 3E). Node size indicates the centrality, while edge width corresponds to the String-DB interaction score
(only scores > 700, i.e., with a high degree of confidence have been included).
ll
Resource

(legend on next page)

 ll
Resource

Figure S4. IDH WT-Specific Alterations in TAMs, Related to Figure 4

(A) Representative IF image and cell type quantification below of non-immune cells (CD45-), non-TAM immune cells (CD45+, P2RY12/CD68-, CD49D+/-), MG
(CD45+, P2RY12/CD68+, CD49D) and MDM (CD45+, P2RY12/CD68+, CD49D+) and vessels (CD31+) in IDH WT glioma. Dashed line indicates the border of the
perivascular niche (PVN), scale bar represents 100mm. (B) Heatmap of cell-type gene set variation analysis (GSVA) enrichment scores of micro-dissected Ivy
Glioblastoma Atlas Project samples (dataset from Puchalski et al., 2018). Columns are ordered by anatomical location, rows have been z-scored. (C) Gene set
enrichment analysis (GSEA) results of MSigDB ‘‘C2’’ antigen processing and cross-presentation associated pathways in T-MDMs versus T-MG in IDH WT glioma.
(D) Heatmap of MDM IDH WT gene set expression in sorted MG and MDMs from IDH mut and WT glioma samples. Columns are ordered by IDH status and cell
type, expression values have been z-scored. (E) Plot of z-scored MDM IDH WT signature scores in the TCGA glioma dataset. Subjects are ranked by their
enrichment score (small amount of random variation added for readability) and the IDH status is indicated by color. (F) Kaplan-Meier estimator of survival in the
combined TCGA glioma cohort based on the enrichment for a cell type-specific T-MDM signature (see Figure S2E). (G) ORA of ‘‘innate anti-PD-1 resistance’’
(IPRES) signatures within DEG from MG- and MDMs in IDH WT gliomas DEGs (versus MG from IDH mut tumors) with tile fill indicating the -log10 of the adjusted p
value. (H) GSVA of IPRES signatures in CD45- cells, MG, and MDMs from IDH mut and IDH WT gliomas. Columns are ordered by cell type, rows (z-score) have
been hierarchically clustered.
 ll
Resource

Figure S5. Protein Concentration in Bulk Tumor Tissues and Relation to Cell-Type-Associated SOM Spots, Related to Figure 5
(A) Bulk tissue protein concentrations of indicated proteins in non-tumor brain (n = 3), gliomas (n = 14) and BrMs (n = 12). Color indicates disease type and IDH
status. (B) Heatmap of self-organizing map (SOM) spot metagene expression across the analyzed samples. Rows were z-scored and have been hierarchically
clustered, columns were ordered by cell type, disease type and IDH mutation status.
 ll
Resource

Figure S6. Gene Expression Analysis in BrM-TAMs, Related to Figure 6

(A) Overlap of the number of differentially expressed genes (DEG, log2(fc) > 1, p.adj < 0.05) in MG and (B) MDMs in the indicated comparisons. BrM-specific gene
sets are highlighted in gray within each cell type. The intersect of highlighted BrM-MG and BrM-MDM sets contains 87 genes. (C) GSEA of the ‘‘Biocarta IL-6
pathway’’ in BrM-MG versus -MDM and the (D) ‘‘Naba core matrisome’’ gene set from the MSigDB ‘‘C2’’ collection in BrM-MDM versus -MG. (E) Expression
(log10-transformed normalized counts) of neutrophil-recruiting chemokines and receptors in sorted MG, MDMs and neutrophil populations from IDH WT and BrM
samples.
 ll
Resource

Figure S7. Correlation of WGCNA Modules with External Traits and Module Pathway ORA, Related to Figure 7
(A) Representative immunofluorescence images in IDH WT gliomas and BrMs. Scale bars = 100mm, boxed area is shown in higher magnification in Figure 7A. (B)
Heatmap of the weighted gene correlation network analysis (WGCNA) module eigengene (= first principal component of expression data, columns, module
columns are labeled with a color code) correlation to the traits (rows, cell type and disease, abundance of CD4+ or CD8+ T cells in % of CD45+). Values inside the
cells state Pearson’s r and the associated p value. (C) ‘‘Brown’’ BrM-MDM module MSigDB ‘‘C2CP’’ ORA results (p value < 0.01) enrichment map network
visualization. Node size represents p value, edge thickness reflects overlap of genes between gene sets.
 Resource

Personalized Mapping of Drug Metabolism by the

Human Gut Microbiome
Graphical Abstract Authors
Bahar Javdan, Jaime G. Lopez,
Pranatchareeya Chankhamjon, ...,
Xiaojuan Wang, Seema Chatterjee,
Mohamed S. Donia

Correspondence
donia@princeton.edu

In Brief
Each human has a diverse gut
microbiome, which can metabolize drugs
differently. In this resource, Javdan et al.
present a way to capture and grow much
of the unique diversity of human
microbiomes in culture and also a way to
detect many of our microbiome-derived
metabolites. Together, they use these
unique gut communities and the
metabolomics pipeline to see how
personalized microbiomes metabolize
drugs in different ways.

Highlights
d Development of subject-personalized ex vivo batch cultures
of the gut microbiome

d Discovery of diverse drug-microbiome interactions using

MDM-Screen

d MDM-Screen quantifies drug metabolism by personalized

gut microbial communities

d Functional genomic and metagenomic screens identify drug-

metabolizing enzymes

Javdan et al., 2020, Cell 181, 1661–1679

June 25, 2020 ª 2020 Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.001 ll
 ll

Resource
Personalized Mapping of Drug Metabolism
by the Human Gut Microbiome
Bahar Javdan,1,4 Jaime G. Lopez,2,4 Pranatchareeya Chankhamjon,1 Ying-Chiang J. Lee,1 Raphaella Hull,1 Qihao Wu,1
Xiaojuan Wang,1 Seema Chatterjee,1 and Mohamed S. Donia1,3,5,*
1Department of Molecular Biology, Princeton University, Princeton, NJ 08544, USA
2Lewis-Sigler Institute for Integrative Genomics, Princeton University, Princeton, NJ 08544, USA
3Department of Chemical and Biological Engineering, Princeton University, Princeton, NJ 08544, USA
4These authors contributed equally
5Lead Contact

*Correspondence: donia@princeton.edu
https://doi.org/10.1016/j.cell.2020.05.001

SUMMARY

The human gut microbiome harbors hundreds of bacterial species with diverse biochemical capabilities.
Dozens of drugs have been shown to be metabolized by single isolates from the gut microbiome, but the
extent of this phenomenon is rarely explored in the context of microbial communities. Here, we develop a
quantitative experimental framework for mapping the ability of the human gut microbiome to metabolize
small molecule drugs: Microbiome-Derived Metabolism (MDM)-Screen. Included are a batch culturing sys-
tem for sustained growth of subject-specific gut microbial communities, an ex vivo drug metabolism screen,
and targeted and untargeted functional metagenomic screens to identify microbiome-encoded genes
responsible for specific metabolic events. Our framework identifies novel drug-microbiome interactions
that vary between individuals and demonstrates how the gut microbiome might be used in drug development
and personalized medicine.

INTRODUCTION
The oral route is the most common route for drug administration.
Upon exiting the stomach, drugs can be absorbed in the small
and/or large intestine to reach systemic circulation and eventu-
ally the liver or directly transported there via the portal vein. In
the liver, drugs may be metabolized and secreted back to the in-
testines through bile—via enterohepatic circulation (Kimura
et al., 1994; Li and Jia, 2013). Even parenterally administered
drugs and their metabolites can reach the intestines through
biliary secretion. Thus, whether prior to or after absorption,
some administered drugs will spend a considerable amount of
time in the small and large intestines, where our human gut
microbiome resides. It is therefore important to study gut micro-
biome composition and function, specifically as it relates to drug
interactions, while accounting for the significant variability be-
tween individuals (Falony et al., 2016).
Broadly speaking, the microbiome interacts with drugs both
directly and indirectly. Indirect interactions include competition
between microbiome-derived metabolites and administered
drugs for the same host metabolizing enzymes (Clayton et al.,
2009), microbiome effects on the immune system in anticancer
immunotherapy (Iida et al., 2013; Sivan et al., 2015; Vétizou
et al., 2015), microbiome reactivation of secreted inactive me-
tabolites of the drug (Wallace et al., 2010), and overall micro-
biome effects on the levels of metabolizing enzymes in the liver
and intestine (Meinl et al., 2009). Direct interactions include the
partial or complete biochemical transformation of a drug into
more or less active metabolites by microbiome-derived enzymes
(termed herein: Microbiome-Derived Metabolism, or MDM).
The human gut microbiome harbors hundreds of bacterial
species, encoding an estimated 100 times more genes than
the human genome (Qin et al., 2010). This enormous diversity
and richness represent a repertoire of yet-uncharacterized
biochemical activities capable of metabolizing ingested chemi-
cals (Bäckhed et al., 2005; Koppel et al., 2017). Although MDM
has been observed in dozens of examples for the past 50 years,
this process is still mostly overlooked in the drug development
pipeline where little to no effort is spent on determining the spe-
cific role of MDM in pharmacokinetics (Ilett et al., 1990; Li and
Jia, 2013; Scheline, 1973; Spanogiannopoulos et al., 2016).
This is because of the vast complexity of the microbiome, and
overwhelming technical challenge of testing hundreds of drugs
against thousands of cultured isolates under multiple conditions.
In contrast to liver-derived metabolism, we lack a systematic and
standardized map of MDM, hindering our ability to reliably pre-
dict and eventually interfere with undesired microbiome effects
on drug pharmacokinetics and pharmacodynamics.
To address this gap in knowledge, we developed a
quantitative experimental workflow for mapping MDM of orally
administered drugs using personalized gut microbiome-derived
microbial communities (MDM-Screen). The methods and

Cell 181, 1661–1679, June 25, 2020 ª 2020 Elsevier Inc. 1661
 ll
findings reported here provide a framework for discovering and
characterizing novel cases of MDM and for potentially incorpo-
rating an ‘‘MDM module’’ in the drug development pipeline.
RESULTS
Mapping the Capacity of a Single Subject’s Microbiome
to Metabolize Hundreds of Drugs
A major challenge in studying the capacity of the human gut mi-
crobiome to metabolize orally administered drugs is the diversity
of bacterial species and strains involved (Almeida et al., 2019;
Lloyd-Price et al., 2017; Nayfach et al., 2019; Pasolli et al.,
2019; Qin et al., 2010). Because it is impractical to systematically
screen thousands of isolated strains against hundreds of drugs,
previous studies have relied mainly on monocultures of a
selected set of representative species. However, gene expres-
sion and biochemical transformation profiles vary dramatically
between a strain grown in monoculture versus in a mixed com-
munity. To address these challenges, we sought to develop an
optimized ex vivo mixed culturing system that supports the
growth of a large proportion of the species from a given micro-
biome sample and is amenable to high-throughput (HT)
biochemical screens.
We began our screening efforts focused on a single micro-
biome donor (pilot donor, PD). To identify the medium and
culturing period that can support the growth of a batch culture
whose composition is maximally similar to the original PD micro-
biome, freshly collected and glycerol-stocked human feces from
PD were cultured in 14 different media and sampled daily for four
days. We then extracted DNA from all samples, amplified the V4
region of the bacterial 16S rRNA gene, and deeply sequenced
the amplicons (Figure 1A). From the sequencing results, ampli-
con sequence variants (ASVs) were inferred and the taxonomic
composition at different levels was determined for each sample
(see STAR Methods) (Bokulich et al., 2018; Bolyen et al., 2018;
Callahan et al., 2016; McDonald et al., 2012). We then quantified
the differences between the various media and PD at the family
level (using the Jensen-Shannon divergence, DJS), and the
variant recovery from PD at the single ASV level.
As expected, we observed a great level of variation in both the
taxonomic composition and diversity between the different me-
dia and culturing periods. Some media led to highly diverse com-
munities that captured portions of the original fecal diversity,
while others became dominated almost exclusively by a single
family. Among the 14 media commonly used in gut microbiome
cultivation efforts (Rettedal et al., 2014), we identified one, modi-
fied Gifu Anaerobic Medium (mGAM), that supported the growth
of a bacterial community most similar in composition and diver-
sity to PD’s (Figures 1B and S1A). At the family level, mGAM
cultures largely match the composition of PD, differing primarily
in a commonly observed expansion of the facultative anaerobes,
Enterobacteriaceae, at the expense of the obligate anaerobes,
Ruminococcaceae (McDonald et al., 2018). Among all tested
media, mGAM cultures showed the lowest DJS divergence
from PD, becoming increasingly similar to the original sample
as growth proceeds (Figure S1A).
Even at the single ASV level, mGAM cultures capture much of
the diversity in PD (mGAM cultures have the highest Shannon di-
1662 Cell 181, 1661–1679, June 25, 2020
Resource
versity across all media, and the closest one to PD) (Figure 1C). In
PD, there are 33 ASVs present above a relative abundance of
1%, 26 (79%) of which are present in mGAM day two culture.
Overall, total shared ASVs between PD and mGAM day two ac-
count for 70% of the PD composition (by relative abundance),
indicating that the mGAM culture recapitulates the bulk of the
original community. Taken together, and consistent with previ-
ous reports showing that mGAM can support the growth of a
wide variety of gut microorganisms in monoculture (Rettedal
et al., 2014; Tramontano et al., 2018), our results support the
use of mGAM day two cultures as a viable ex vivo batch culturing
model for the PD microbiome.
With an optimized ex vivo culturing system for PD in hand, we
next developed a combined biochemical and analytical chemis-
try approach to map the capacity of PD-derived microbial com-
munities to metabolize clinically used, orally administered drugs
(MDM-Screen) (Figure 2A). Three samples were prepared per
drug of interest: (1) a 24-h mGAM ex vivo culture of PD, incu-
bated with the drug of interest (final concentration 33 mM, in
line with estimates of drug concentrations in the gastrointestinal
tract) (Maier et al., 2018), (2) a similar culture incubated with a
vehicle control (DMSO), and (3) an equal volume of sterile
mGAM, incubated with the same drug concentration. Cultures
and controls were then incubated for an additional 24 h at
37
C in an anaerobic chamber, chemically extracted, and
analyzed using high performance liquid chromatography
coupled with mass spectrometry (HPLC-MS). To verify the
reproducibility, the entire procedure was repeated three consec-
utive times. We tested a diverse library of 575 orally administered
drugs, and although the majority of the drugs in this library are
currently being used in the clinic, less than 10% of them had
been previously explored with respect to MDM (Table S1A). A
drug was deemed MDM+ when (1) a new metabolite was
observed when incubated with PD culture or (2) the drug was
no longer detected when incubated with PD culture, indicating
that it is either consumed entirely or metabolized into a molecule
that fails our detection, and (3) the drug was metabolized in the
same manner during at least two of three independent
experiments.
MDM-Screen Identifies Known and Novel Drug-
Microbiome Interactions
Among the 575 drugs, we successfully analyzed 438 (76%); the
remaining 137 failed MDM-Screen because of issues related to
drug stability or incompatibilities with the extraction or chroma-
tography methods employed (see Discussion). Of the success-
fully analyzed drugs, we identified 57 (13%) as MDM+. These
spanned 28 pharmacological classes and even more based on
their chemical structure (Figure 2B; Table S1B; Data S1). As ex-
pected, several previously reported MDM cases were identified.
These include the nitroreduction of the muscle relaxant dantro-
lene (Kuroiwa et al., 1985), the antiepileptic clonazepam (Elmer
and Remmel, 1984; Zimmermann et al., 2019b), and the antihy-
pertensive drug nicardipine (Kuroiwa et al., 1986); hydrolysis of
the isoxazole moiety in the antipsychotic risperidone (Mannens
et al., 1993; Meuldermans et al., 1994); and azoreduction of
the anti-inflammatory prodrug sulfasalazine (Azadkhan et al.,
1982; Peppercorn and Goldman, 1972).
 ll
Resource

B C

Figure 1. Development of an Ex Vivo Batch-Culturing System for the PD Microbiome

(A) Schematic representation of the media selection procedure.
(B) Family level bacterial composition of the original fecal sample (far left), as well as that of PD ex vivo cultures, grown anaerobically in 14 different media over two
days (.01 and .02). See STAR Methods for full media names. 16S rRNA gene sequences that could not be classified at the family level and families with less than
1% relative abundance in all samples are grouped into ‘‘Other.’’ Cultures are ordered according to their Jensen-Shannon (DJS) divergence from the original PD
sample (upper axes, computed at the family level in base e).
(C) ASV level bacterial composition of the original PD fecal sample, and that of day two ex vivo cultures of PD, where each square represents one sample. Rainbow
colored dots represent the relative abundance of individual ASVs that are above 1% in PD, while gray dots represent the combined relative abundance of all ASVs
below 1% in PD. Samples are ordered by their Shannon diversity (H) at the ASV level, computed in base 2 and shown above each square.
See also Figure S1; Table S2.

More importantly, we identified a suite of novel MDM cases (45
cases, 80% of the MDM+ drugs): ten resulted from full depletion
of the parent drug (or full conversion to a metabolite that evades
our detection), while 35 resulted from the appearance of a new
metabolite (Figures 2C and 2D; Table S1B; Data S1A). In most
cases, the new metabolites showed a high-resolution mass spec-
trometry (HRMS) profile within a small difference from their parent
drugs and/or a similar tandem MS fragmentation (HRMS/MS)
pattern, indicating that they are derivatives (Wang et al., 2016) (Ta-
ble S1B; Data S1B). An aggregate statistical analysis of MDM+ and
MDM drugs revealed specific structural features that are signifi-
cantly enriched in MDM+ drugs (e.g., a steroidal skeleton, nitro
groups, ketones, among others) (see STAR Methods; Table S1C).
Although HRMS and HRMS/MS analyses can narrow down
the number of possibilities for the molecular structure of a given
metabolite, they are not sufficient for full structural determina-
tion. Thus, we selected seven MDM+ examples for detailed
characterization of their resulting metabolites: spironolactone
(anti-hypertensive), tolcapone (anti-Parkinson’s), misoprostol
(anti-ulcer), mycophenolate mofetil (immunosuppressant),
capecitabine (anticancer), and finally, hydrocortisone and hydro-
cortisone acetate (two steroidal anti-inflammatory drugs that
produced an identical MDM metabolite). In all but one example,
no direct drug-microbiome interactions had been previously re-
ported. The exception was hydrocortisone where several metab-
olites had been previously reported from individual gut isolates

Cell 181, 1661–1679, June 25, 2020 1663

 ll
Resource

B C

(legend on next page)

1664 Cell 181, 1661–1679, June 25, 2020


Resource
(Ridlon et al., 2013; Winter et al., 1982), but the identity of the
MDM metabolite we observed could not be accurately matched
to any of them based on MS alone.
To unequivocally determine the structure of the resulting
metabolites, we isolated them from scaled-up biochemical incu-
bations with PD cultures and elucidated their structures using
nuclear magnetic resonance (NMR) and/or comparison to an
authentic standard (STAR Methods; Data S2). For hydrocorti-
sone, we determined that MDM results in the reduction of the
ketone group at C20, producing 20b-dihydrocortisone. For hy-
drocortisone acetate, the same modification occurs but is
accompanied with deacetylation of the C21 hydroxyl group (Fig-
ure S2). While C20 ketone reduction was previously reported for
hydrocortisone (to produce either 20b-dihydrocortisone or
20a-dihydrocortisone depending on the gut isolate incubated
with it; Ridlon et al., 2013; Winter et al., 1982), neither MDM de-
acetylation nor C20 ketone reduction were reported for hydro-
cortisone acetate. For capecitabine, we show that MDM results
in complete deglycosylation; for misoprostol and mycopheno-
late mofetil, we observed an ester hydrolysis transformation;
and in the case of spironolactone, a thioester hydrolysis one.
None of these MDM transformations were previously reported
for these drugs. Finally, for tolcapone, we observed two consec-
utive transformations, a typical nitroreduction followed by a
relatively uncommon N-acetylation—neither of which had been
previously linked to the microbiome for this drug (Figure 2D).
Taken together, these results establish MDM-Screen as a viable
method for identifying both known and novel biochemical mod-
ifications of structurally and pharmacologically diverse drugs by
the gut microbiome.
Interestingly, based on already known pharmacokinetic
studies in humans, some of these new MDM cases may have
direct consequences on the activation or toxicity of the drugs
involved. For example, in the case of spironolactone we
observed the production of 7a-thiospironolactone, a postulated
intermediate en route to the drug’s main active metabolite,
7a-thiomethylspironolactone (Gardiner et al., 1989; Sica,
2005). For the prodrugs misoprostol and mycophenolate mofetil,
we observed the production of their active metabolites miso-
prostol acid (Schoenhard et al., 1985; Tsai et al., 1991) and
mycophenolic acid (Bullingham et al., 1998), respectively. Inter-
estingly, mycophenolic acid has been linked to the clinically
observed gastrointestinal toxicity associated with mycopheno-
late mofetil use (Taylor et al., 2019), albeit generated via a
different route—hydrolysis of a biliary secreted glucuronide con-
jugate by gut microbiome-derived b-glucuronidases. Finally, in
the case of tolcapone, N-acetylamino-tolcapone has been de-
tected systemically in humans post-tolcapone administration
and was suggested to be involved in liver toxicity observed clin-
Figure 2. Screening of the PD Microbiome against Orally Administered Drugs Identifies Novel Drug-Microbiome Interactions
(A) Schematic representation of MDM-Screen. A drug was considered MDM+ if a new metabolite is produced (e.g., drug 3) or if the drug is no longer detectable
(e.g., drug 5) after incubation with the microbiome, as compared with abiotic media controls.
(B) A bar graph showing the pharmacological classes of MDM+ drugs discovered by MDM-Screen with the PD microbiome. ‘‘Others’’ include one drug each from
14 additional classes.
(C) Examples of MDM+ drugs where the drug is no longer detectable after incubation with the PD microbiome.
(D) Examples of MDM+ drugs where a new metabolite is discovered by MDM-Screen and fully characterized in this study.
See also Table S1; Data S1 and S2.
ll
ically following tolcapone use, yet the mechanism of its produc-
tion remains unknown (Jorga et al., 1999; Smith et al., 2003). Our
discovery that the same metabolite can be produced via MDM
provides a possible explanation, and a potential link between
the gut microbiome and tolcapone toxicity. In all four cases,
additional experiments need to be performed to differentiate
the contribution of human- and microbiome-derived metabolism
to the observed drug pharmacokinetics and/or toxicity in
humans.
Expanding MDM-Screen to Multiple Subjects
Next, we sought to expand our framework to accommodate mul-
tiple subjects. To accomplish this goal, we needed to first design
a generalizable quantitative metric for assessing the best
culturing medium for microbiome samples (Figures 3A and 3B).
In our analysis of PD’s ex vivo cultures, we applied a variety of
metrics and found a medium that was the best trade-off between
richness, evenness, and compositional similarity. However, this
approach is not scalable to a large number of donors and also
ignores the role of community biomass, which may lead to sub-
optimal media selection. Therefore, we developed a metric
called Expected Number of Detectable Strains (ENDS), a cor-
rected richness metric where the contribution of each ASV is
weighed by the probability that its metabolite can be detected
while considering total biomass (STAR Methods; Figure 3B;
Methods S1). The core idea is that we desire a medium that sup-
ports the highest number of different bacterial ASVs, while
ensuring that the metabolic contributions of these ASVs are
detectable by our experimental method. ENDS utilizes two
data inputs related to the ex vivo culture composition: relative
abundance at a given taxonomic level and total community
biomass; and it also utilizes two inputs related to the instrument
detection sensitivity: a model of instrument background noise
and a model of instrument measurement noise. Using this
information, a simple mechanistic model of MDM metabolite
production, and estimations of statistical power, we compute
the probabilities that metabolic reactions performed by each
strain will be detected in the ex vivo culture (see STAR Methods
for a detailed mathematical description of ENDS).
With this quantitative framework in hand, we collected addi-
tional fresh fecal samples from 20 healthy donors (D1-20) and
processed them in the same manner as PD. We then cultured
each sample in nine representative media and used 16S rRNA
gene sequencing to determine the composition of the cultured
communities as previously described (Figures 3A–3C and S1;
Table S2A). To measure community biomass, one mL of each
culture was pelleted and weighed (STAR Methods; Tables S2C
and S2D). We observed a wide variation in culture characteris-
tics, with the richness ranging from 20–135 ASVs and mean
Cell 181, 1661–1679, June 25, 2020 1665
 ll
Resource

D E F

G H

(legend on next page)

1666 Cell 181, 1661–1679, June 25, 2020

 ll
Resource

biomass density ranging from 2–27.9 g/L (Figure 3D). mGAM and
Bryant and Burkey (BB) media consistently performed well with
all 20 donors. Interestingly, mGAM had moderate ASV richness
and high biomass, while BB yielded a much lower biomass
with high richness and did not suffer from the Enterobacteri-
aceae expansion observed in mGAM. We calculated that a 70/
30 BB/mGAM mixture would yield an optimal medium with mod-
erate biomass, high richness, and a reduced Enterobacteriaceae
expansion (Methods S1B), thus we included this mixture (named
BG) as a 10th medium in our culturing trials.
Next, we wondered whether the ex vivo cultured communities
are truly personalized per subject, an important prerequisite if
cultured communities are to be used for assessing inter-individual
variability in MDM. Personalization between cultures was clearly
observed at the ASV level, with clear specific patterns unique to
individual donors and their cultures (Figure 3C). We found signifi-
cantly more ASVs shared between donor feces and their self
ex vivo cultures than non-self (47.1 versus 27.5 ASVs, p < 0.001,
permutation test), partially recapitulating the inherent personaliza-
tion between the donor fecal samples (Figures 3E and 3F). More-
over, we identified 167 ASVs that were unique to one of the 20 do-
nors in their fecal samples (8.4 ASVs per donor on average) and
were concordantly unique to the same donor in their ex vivo cul-
tures (Table S2F). Finally, we grew and sequenced multiple repli-
cates of mGAM and BG ex vivo cultures from different donors (all
20 donors, three replicates each for BG, and eight donors, six rep-
licates each for mGAM). The analysis of these replicates, whether
cultured from the same or separate glycerol stock aliquots, re-
vealed a high correlation between ASV abundances of replicates
from the same donor and ensured that the community assembly
process was replicable (Pearson correlation coefficient >0.9)
(Methods S1A; Figure S1). These analyses confirm that our
approach results in personalized and replicable microbial
communities.
To select a single medium that would be on average optimal for
use in a 20-donor screen, we computed the average ENDS of all
media across a range of reaction rates (Figure 3G). We found
that BG is on average an optimal medium at reaction rates we esti-
mated to be the most physiologically relevant (Methods S1D). In
addition, BG recapitulates a large portion of the microbial commu-
nity in the original fecal sample. On average, BG cultures recover
76.6% of the ASVs above 1% in the original fecal sample, which
translates to 84.7%, 88.3%, and 92.7% recovery rate on the spe-
cies, genus, and family levels of taxa above 1% in the original sam-
ple, respectively (Figure 3H). In terms of recovery of all elements,
BG recovers 43.3%, 57.3%, 60.6%, and 62.8% on the ASV, spe-
cies, genus, and family levels, respectively. BG is also, on average,
the closest in composition to the original sample (DJS = 0.16,
computed at the family level in base e), has the highest average di-
versity (H = 4.3, computed at the ASV level in base 2), and shows
the highest ASV richness (ranging from 34-133 ASVs with an
average of 77). We therefore selected BG as the medium to use
in our 20-donor screen.
We next sought to develop a HT, quantitative metabolomic
approach to assess MDM inter-individual variability with a sub-
set of drugs. Several improvements were made to the original
drug metabolism screen. All experimental steps including incu-
bation, chemical extraction, and HPLC-HRMS analysis were
performed in microtiter 96-well plates instead of individual tubes,
at a 400 mL volume instead of 3 mL. This lowered the amount of
drug used per incubation, allowed us to perform triplicated reac-
tions simultaneously, and streamlined our chemical extraction
and analysis procedures. We also spiked a known concentration
of an internal standard prior to the chemical extraction, which al-
lowed us to precisely quantify partial, in addition to complete,
drug depletion.
We chose a 23-drug subset to test the ability of our quantita-
tive approach to reveal potential inter-individual variability in
MDM under the MDM-Screen conditions. 13 drugs had at least
one defined metabolite with a known chemical structure, allow-
ing us to unambiguously compare their levels between samples
(Figure S3; Table S3). For all 20 donors, ex vivo cultures (in BG
medium) were incubated in triplicates in a 96-well microtiter plate
with each of the 23 drugs at a final concentration of 33 mM, or
with DMSO (Figure 4A). In addition, an abiotic medium-drug
plate as well as a heat-killed-microbiome-drug (HKM-drug) plate

Figure 3. Identifying the Optimal Medium for Multi-Donor MDM-Screen

(A) Schematic representation of the media selection procedure for D1-20.
(B) Schematic representation of the ENDS metric. Using 16S rRNA gene sequencing and biomass measurements, absolute abundances of ASVs (orange and
gray strains) in different ex vivo communities are measured and metabolite production from each member of the community is estimated using a simple
mathematical model. Using instrument noise properties, distributions of metabolite measurements from each ASV (orange and gray distributions) are estimated
and compared to instrument noise (white distribution). Statistical power estimation is then used to compute metabolism detection probabilities for each ASV and
the condition maximizing ENDS (the sum of these probabilities) is selected.
(C) ASV abundance heatmap of the original fecal samples and ex vivo microbial communities for each donor. Each box corresponds to samples from a single
donor, with the original fecal sample shown on the far left followed by different ex vivo media in the order specified above the heatmap. Only ASVs above 5% in at
least one sample are shown, with all remaining ASVs aggregated into ‘‘Other.’’ The taxonomic classification of each ASV (on the order level) is indicated by the
color bar on the left.
(D) Histogram of ex vivo community biomass for all donors in different media conditions.
(E) Comparison of shared ASVs within (self, i.e., the ASV richness) and between (non-self) donor fecal samples. ***p < 0.001, permutation test.
(F) Comparison of shared ASVs between donor fecal samples and ex vivo cultures originating from the same donor (self) versus ones originating from other donors
(non-self). ***p < 0.001, permutation test.
(G) Average ENDS of different media conditions at varying metabolite production rates (quantified as AUC normalized to an internal standard). ENDS was
computed for each ex vivo culture assuming a p value significance cutoff of 0.01 and three replicates. For each media condition, ENDS was averaged across all
donors.
(H) Average fractional recovery of different taxa in BG ex vivo communities as a function of relative abundance in the original donor fecal sample. The fractional
recovery was calculated for all donors and then averaged.
See also Figure S1; Table S2.

Cell 181, 1661–1679, June 25, 2020 1667

 ll
Resource

B C

D F

(legend on next page)

1668 Cell 181, 1661–1679, June 25, 2020

 ll
Resource

were prepared in the same manner. After 24 h incubation, culture
and control plates were chemically extracted and analyzed using
HPLC-HRMS (STAR Methods).
We then devised a targeted metabolomics strategy to quantify
MDM. We calculated percent drug remaining and metabolite
level, to assess drug depletion and metabolite production in
the presence of microbiome cultures, respectively. Both metrics
were calculated using area under the curve (AUC) integration
with normalization to the internal standard (see STAR Methods;
Tables S3J–M). We determined statistical significance for
metabolite production using one-sided Welch’s t tests
between the donor-drug condition and the donor-DMSO,
medium-drug, and HKM-drug conditions and corrected the re-
sulting p values for multiple hypotheses using the Benjamini-
Hochberg method, requiring that tests against all three control
conditions have a false discovery rate (FDR) corrected p < 0.01
(Benjamini and Hochberg, 1995). For drug depletion, we used
the same method with the donor-drug and HKM-drug conditions
as controls and included an additional fold-change cutoff of two
(Figures 4B and 4C; Tables S3N–S3P). We also performed untar-
geted metabolomics analyses for new metabolite discovery by
identifying unique molecular features from all samples, deter-
mining statistical significance using similar methods as for
the targeted metabolomics and verifying the metabolite’s
relationship to the parent drug based on their HRMS/MS frag-
mentation pattern (Wang et al., 2016) (STAR Methods; Tables
S3C and S3D). All verified metabolites from the untargeted me-
tabolomics approach were then quantified using the same tar-
geted metabolomics workflow described above (Figure 4C;
Tables S3N–S3P).
We observed cases of consistently negative MDM across do-
nors (ketoconazole, praziquantel, ropinirole, and torsemide),
consistently positive MDM in either drug depletion (misoprostol,
nicardipine, and spironolactone), metabolite production (tolca-
pone and vorinostat), or both (clonazepam, risperidone, and sul-
fasalazine), and variable MDM (Figures 4B–4D). This variability
was in drug depletion (ketoprofen and levonorgestrel), metabo-
lite production (misoprostol, nicardipine, and spironolactone),
or both (capecitabine, clofazimine, digoxin, hydrocortisone,
lovastatin, mycophenolate mofetil, sulindac, and vorinostat).
We quantified this variability by computing the Shannon entropy
(in base 2) of the distribution of metabolizers and non-metaboliz-
ers, denoted as HV. This metric is maximal (HV = 1) when half of
donors metabolize the drug and is minimal when the drug is
either always or never metabolized (HV = 0).
The observed variability ranged widely from 1/20 to 19/20 do-
nors deemed MDM+ for a given type of drug depletion or metab-
olite production. In the case of digoxin, for example, 3/20 donors
(HV = 0.61) produced the known metabolite dihydrodigoxin in
statistically significant amounts (Figures 4C and 4E). Inter-indi-
vidual variability in digoxin MDM has been clinically known for
decades, where significant reduction of the drug into dihydrodi-
goxin and related metabolites occurs in only a subset of patients
(Lindenbaum et al., 1981). These results demonstrate that our
screen can quantitatively assess the inter-individual variability
of MDM between personalized gut microbial communities
cultured under identical ex vivo conditions. Follow-up studies
will need to be performed to evaluate whether our screening re-
sults directly correlate with clinical outcomes.
Next, we sought to determine whether the depletion of drugs
in our screen can be explained by the production of associated
metabolites. If changes in drug levels are primarily because of
conversion to a detected metabolite, there should exist a strong
negative correlation between depletion and metabolite produc-
tion, corresponding to a stoichiometric mass balance. The
absence of such a correlation, on the other hand, would suggest
additional events that are not accounted for (e.g., the production
of additional unknown or undetectable metabolites, the conver-
sion of the initial metabolite into a second one, or bacterial con-
sumption of the parent drug). For drugs with variable MDM (HV >
0.5 for at least one metabolite or the parent drug), we computed
the Pearson correlation coefficient of the drug signal and the sum
of known metabolite signals in all donor-drug ex vivo samples.
We then determined whether a drug has statistically significant
correlation by performing t tests, correcting p values using the
Benjamini-Hochberg method, and requiring FDR corrected p <
0.01. For vorinostat and digoxin, for example, we found a signif-
icant negative correlation between metabolite production and
drug depletion (Pearson correlation coefficient of 0.91 and
0.79, respectively), suggesting that the majority of drug

Figure 4. A HT, Quantitative Metabolomic Approach to Assess Inter-Individual Variability in MDM Using Personalized Microbial Communities
(A) Schematic representation of quantitative MDM-Screen with 20 donors and 23 selected drugs.
(B) Heatmap of drug depletion showing the mean fraction of drug remaining after 24 h for each donor-drug combination. The fraction remaining is computed
relative to the medium-drug control, and fractions above 1 are truncated to 1 for simplicity.
(C) Heatmap of metabolite production showing the mean level of metabolite after 24 h, normalized to the maximum level of that metabolite across all donors.
Metabolites in red were discovered using the untargeted metabolomics approach, while ones in black were discovered previously or by MDM-Screen with the PD
microbiome (Table S3B). In (B) and (C), *statistically significant metabolism in the donor condition as compared with controls. The upper inset axes represent
inter-individual variability in MDM using the Shannon entropy (calculated in base 2) of the distribution of donors with significant and non-significant metabolism.
(D) Cumulative histogram of the number of significant donors for both metabolite production and parent drug depletion. For parent drugs, the y axis is normalized
to the total number of drugs tested (23), and for metabolite production, it is normalized to the total number of metabolites produced (32).
(E) Levels of metabolite production (measured by HPLC-HRMS in AUC normalized to an internal standard) for four drugs, with the variability entropy indicated
above. Filled data points indicate that the replicates are significantly higher than control conditions, while hollow data points indicate that they are not.
(F) The upper three scatterplots show significant negative correlation between drug depletion and metabolite production, with the Pearson correlation coefficient
indicated above. The line shown is a linear regression fit of the data. The lower bar plot indicates the Pearson correlation coefficient between remaining drug levels
and total metabolite production for all computed cases. *FDR corrected two-sided t test p < 0.01. For drugs with multiple metabolites, we sum the normalized
AUC of all metabolites.
(G) Correlation between drug depletion and metabolite production for nicardipine before and after inclusion of metabolites discovered by untargeted
metabolomics.
See also Table S3.

Cell 181, 1661–1679, June 25, 2020 1669

 ll
depletion can be explained by the production of the quantified
metabolite (Figure 4F). Nicardipine, on the other hand, exhibited
a very poor correlation initially (Pearson correlation coefficient of
0.22), implying that additional unknown factors are at play.
Interestingly, our untargeted metabolomics pipeline detected
11 additional metabolites of nicardipine, which upon inclusion
in the analysis resulted in a stronger negative correlation (Pear-
son correlation coefficient of 0.6, FDR corrected p = 0.0102)
(Figure 4G; Table S3E). Since our screen is based on microbial
communities and not individual strains, it provides a powerful
platform to discover interacting factors that influence drug and
metabolite levels under realistic conditions—as exemplified by
the varying number of nicardipine metabolites observed per
personalized community.
Next, we assessed whether we could predict MDM using
taxonomic data. We computed Spearman correlations between
absolute abundances of taxonomic elements (at different
levels) in the BG ex vivo cultures and measured drug and
metabolite levels in matching donors but found no significant
correlations—even in specific cases of MDM where meta-
bolism has been previously attributed to a single species
(e.g., digoxin reduction by Eggerthella lenta) (Haiser et al.,
2013). This is likely due to a combination of two factors. First,
as has been previously observed (Haiser et al., 2013; Maini Re-
kdal et al., 2019), taxonomic classifications may not reflect the
presence or absence of gene variants that encode strain-spe-
cific drug-metabolizing enzymes, even at the ASV level. Sec-
ond, the observed level of MDM may not be monotonically
dependent on a single taxon’s abundance if confounding com-
munity effects are at play. Examples of such effects include the
contribution of several community members to the production
of the metabolite(s), the consumption of the drug or metabo-
lite(s), or the inhibition of the metabolite-producing or drug-
depleting bacterium or enzyme (for a mathematical analysis
of the impact of these factors on the correlation, see Methods
S1C). These results emphasize the importance of considering
whole community effects in MDM. While our ex vivo commu-
nities may not fully recapitulate all possible community effects
that occur in humans, they represent an important step toward
identifying and quantifying them.
Linking MDM to Specific Genes in the Human
Microbiome
Next, we sought to link the observed biochemical transforma-
tions to specific microbiome-derived enzymes. We picked two
representative cases of MDM transformations: MDM deglycosy-
lation of capecitabine into deglycocapecitabine and C20 ketone
reduction of hydrocortisone into 20b-dihydrocortisone. Three
main approaches had been previously employed to identify
genes responsible for a specific MDM transformation: compara-
tive transcriptomics, which assumes that the expression of
metabolizing enzymes is induced in the presence of their sub-
strates (e.g., digoxin) (Haiser et al., 2013; Koppel et al., 2018), ho-
mology-based discovery, which assumes that related classes of
enzymes metabolize similar substrates (e.g., levodopa) (Maini
Rekdal et al., 2019; van Kessel et al., 2019; Zimmerman et al.,
2019b)), and HT mutagenesis screens, which identify metabo-
lizing enzymes by isolating loss-of-function mutant strains (Zim-
1670 Cell 181, 1661–1679, June 25, 2020
Resource
mermann et al., 2019b). For capecitabine deglycosylation, we
elected to use a homology-based approach.
Characterizing the Genetic Basis of MDM
Deglycosylation Using a Homology-Based Approach
To identify a specific microbiome-derived isolate where a homol-
ogy-based approach can be employed, we explored the ability
of a limited panel of bacterial isolates to deglycosylate capecita-
bine, including strains isolated originally from PD. Interestingly,
capecitabine deglycosylation was mainly performed by Proteo-
bacteria (including Escherichia coli), and one of two tested
Bacteroidetes: Parabacteroides distasonis, providing genetically
tractable organisms for functional studies (Figure S4). In hu-
mans, thymidine phosphorylase (TP) and uridine phosphorylase
(UP), both part of the pyrimidine salvage pathway, catalyze the
deglycosylation of 50 -deoxy-5-fluorouridine (a late metabolite
of capecitabine) to yield 5-fluorouracil (5-FU) (Temmink et al.,
2007). To test whether bacterial homologs of human TP and/or
UP are responsible for the observed MDM deglycosylation of ca-
pecitabine, we generated strains of E. coli BW25113 that are
knockouts for TP (DdeoA), UP (Dudp), or both, and compared
their ability to metabolize capecitabine with that of wild-type
(WT) E. coli (Figure 5A). While WT E. coli efficiently deglycosy-
lates capecitabine (30% conversion rate), the deglycosylating
activity of the Dudp and DdeoA/Dudp strains is significantly
diminished (less than 4% conversion rate, p value < 0.001,
two-tailed t test) (Figure 5B). Surprisingly, the DdeoA strain
showed a significant increase in its deglycosylating activity in
comparison with WT (50% conversion rate, p value < 0.01,
two-tailed t test), possibly because of a compensating mecha-
nism (e.g., overexpression of udp) in the absence of deoA. These
results indicate that microbiome-derived UP is, at least in part,
responsible for the deglycosylation of capecitabine.
Capecitabine is one of several generations of antimetabolite
chemotherapeutic agents, many of which are prodrugs for
5-FU, and are known collectively as the oral fluoropyrimidines
(FPs) (Lamont and Schilsky, 1999; Longley et al., 2003). Impor-
tantly, oral FPs’ bioavailability and toxicity vary widely among pa-
tients (Cleary et al., 2017; Zampino et al., 1999), but the human
gut microbiome’s contribution to this variability had not been
explored. To determine whether deglycosylation occurs with
other FPs, and whether the same enzymes are involved, we
investigated the MDM of two additional oral FPs (doxifluridine
and trifluridine) using WT and mutant E. coli. Unlike with capeci-
tabine, almost complete deglycosylation was observed for both
drugs with WT E. coli, and the activity was dependent on both TP
and UP (Figures S4 and S5). These results indicate a level of de-
glycosylation specificity for TP/UP among the FPs (Figure 5C).
Remarkably, the consequences of the same modification differ
depending on the tested drug. For trifluridine, the resulting
metabolite (trifluorothymine) is inactive (Figures 5D and S4):
trifluridine is typically incorporated intact into DNA to cause cyto-
toxicity (Cleary et al., 2017; Lenz et al., 2015). Such a premature
intestinal inactivation by the microbiome may thus be an un-
known contributor to the established low bioavailability of triflur-
idine, in addition to the known contribution of human TP (Cleary
et al., 2017). For doxifluridine, however, the resulting metabolite
is the active 5-FU (Figures 5E and S5). This premature activation
 ll
Resource

B C

D E

F G

Figure 5. Genetic Basis and Widespread Nature of MDM Deglycosylation among the FPs and in Human Gut Metagenomes
(A) Genetic organization of the udp and deoA loci in the genome of E. coli BW25113.
(B) A bar graph indicating percent conversion of capecitabine to deglycocapecitabine by wild-type (WT) E. coli BW25113 and Dudp, DdeoA, and DdeoA/Dudp
mutants (each tested in triplicate). ***p value < 0.001, while **p value < 0.01, two-tailed t test. Error bars represent the standard deviation.
(C) Biochemical reaction catalyzed by thymidine and uridine phosphorylases on their natural substrates.
(D) MDM deglycosylation of the oral anticancer drug trifluridine leads to its premature inactivation, since trifluorothymine is no longer active.
(E) MDM deglycosylation of the anticancer prodrug doxifluridine leads to its premature activation, since 5-FU is the intended active metabolite.
(F) Heatmaps indicating the prevalence (in percent of subjects harboring the gene of interest) and abundance (in median RPKM for all positive subjects) of E. coli-
derived deoA and udp across six gut metagenomic cohorts. The number of subjects analyzed in each cohort is indicated on the right. For cohorts with multiple
sub-cohorts, the values reported are the averages of the sub-cohort values.
(G) Jitter plots of E. coli-derived deoA and udp abundances (in RPKM) for all positive subjects in the same cohorts.
See also Figures S4 and S5; Table S4.

Cell 181, 1661–1679, June 25, 2020 1671

 ll
of the prodrug may therefore lead to gastrointestinal toxicity—
again, a side effect commonly associated with oral doxifluridine
(Kim et al., 2001; Min et al., 2000). Additional studies are neces-
sary to directly correlate the level of MDM deglycosylation of
different FPs in humans to their clinically observed pharmacoki-
netics and/or toxicity.
Because capecitabine was significantly and variably metabo-
lized into deglycocapecitabine in 17/20 donors, we sought to
examine the representation of FP deglycosylating enzymes in the
gut microbiome of the human population at large. We specifically
focused on enzymes that we experimentally verified to have a
role in FP deglycosylation: E. coli-derived TP and UP. Overall, we
analyzed six large and diverse human cohorts: the Human Micro-
biome Project (HMP-1-1 and HMP-1-2, 299 subjects from the
USA) (Human Microbiome Project Consortium, 2012; Lloyd-Price
et al., 2017), the Metagenomics of the Human Intestinal Tract Con-
sortium (MetaHIT, 219 subjects from Spain and 176 subjects from
Denmark) (Nielsen et al., 2014), a Chinese cohort (194 subjects)
(Qin et al., 2012), and a Fijian cohort (Fijicomp, 192 subjects) (Brito
et al., 2016). We mapped fecal metagenomic reads from each of
the cohort samples to the DNA sequence of deoA and udp, and
calculated two metrics: prevalence, i.e., the percent of subjects
from each cohort that are positive for a given gene, and abundance
of the gene among positive samples (calculated in reads per Kbp
per million of sequenced reads, or RPKM) (STAR Methods; Tables
S4B and S4C). Interestingly, we found that both genes were most
prevalent in non-Western cohorts (Chinese 74%/76% positive
subjects, and Fijicomp 64%/70% positive subjects for deoA/udp,
respectively) in comparison with Western ones (24%/26% on
average, for deoA/udp, respectively) and that their abundance
per positive samples varies widely within and between cohorts
(from 101 to 102 RPKM) (Figures 5F and 5G; Tables S4B and
S4C). These results indicate that FP deglycosylating enzymes are
both widespread and variable in the gut microbiome of diverse hu-
man cohorts (even when considering the contribution of a single
bacterial species, E. coli) and further highlight the importance of
considering MDM deglycosylation of FPs in clinical studies.
An Untargeted Functional Metagenomic Screening
Approach for Identifying Metabolizing Enzymes
Although the homology-based approach was relatively straight-
forward in identifying responsible species and enzymes for the
deglycosylation of FPs, it is not widely applicable. Unlike pyrim-
idine phosphorylases, oxidoreductases (the enzyme class likely
responsible for hydrocortisone reduction to 20b-dihydrocorti-
sone) are extremely diverse and typically substrate specific,
with numerous homologs found per bacterial genome. More-
over, homology-based discovery, as well as comparative tran-
scriptomics and HT mutagenesis screens, typically require the
identification of an isolated strain that performs the modification
of interest and its use as the basis for genetic manipulations and/
or functional analyses. These two limitations motivated us to
employ an orthogonal strategy that is not reliant on enzymatic
homology nor isolated strains.
While no human gut microbiome-derived enzymes had previ-
ously been deemed responsible for converting hydrocortisone
into 20b-dihydrocortisone, a cat microbiome-derived enzyme
had: a 20b-hydroxysteroid dehydrogenase (20b-HSDH) from
1672 Cell 181, 1661–1679, June 25, 2020
Resource
the cat fecal isolate Butyricicoccus desmolans ATCC 43058 (De-
vendran et al., 2017). Neither this enzyme nor close homologs
thereof (at 60% protein sequence identity or above) could be
identified in a deep metagenomic sequencing dataset that we
generated from the PD fecal DNA (STAR Methods). We therefore
decided to use this example as a test case for developing an un-
targeted functional metagenomic screening strategy for metab-
olizing enzymes. In typical functional metagenomic screens,
metagenomic DNA is cloned into a vector that replicates in
E. coli and functional screens are performed in either a selective
manner (e.g., for antibiotic resistance or an engineered circuit for
survival) (Genee et al., 2016; Sommer et al., 2009; Uribe et al.,
2019) or a visual readout (e.g., a colorimetric or antibacterial
one) (Brady et al., 2002; Cohen et al., 2015; Gillespie et al.,
2002; Rondon et al., 2000). Here, we use a functional metage-
nomic screen where the readout is a specific MDM transforma-
tion that is detected by MS. For metagenomic genes that are
successfully expressed and produce functional gene products
in E. coli, this approach would allow access to enzymes encoded
by cultured and not-yet cultured members of the microbiome,
and to ones that share no close homology with previously
characterized enzymes. Two major technical challenges in this
strategy, however, are to produce a large-enough metagenomic
library that captures the majority of the genetic content in the
complex microbiome, and to develop a HT analytical chemistry
approach that permits the screening of such a library.
We isolated metagenomic DNA from PD and used it to
construct a 3 3106-member clone library (PD-CL) in an
E. coli expression vector (insert size 2–4 Kbp) (Figure 6A). To
determine whether PD-CL is truly representative of the genetic
content in PD, we deeply sequenced a representative pool that
contains 105 unique clones (PD-CL-100) and compared it
with the deeply sequenced PD fecal metagenome. We mapped
metagenomic reads from either PD or PD-CL-100 to assembled
scaffolds from the PD metagenome (25,529 scaffolds R 2 Kbp).
Satisfyingly, reads from PD-CL-100 (which represents only 3%
of the full PD-CL) map to 21% of the PD scaffolds, including
ones that originate from all major bacterial phyla and varying
coverages in the PD microbiome (Figure 6B; Table S4A). These
results indicate that PD-CL represents a large component of
the genetic content in PD and that it is adequate for use in func-
tional metagenomics screens.
During the construction of PD-CL, we split it into 80 pools of 2–
6 3 104 unique clones (UCs) each and preserved them in corre-
sponding glycerol stocks (see STAR Methods). We tested each
of these pools for the ability to convert hydrocortisone into
20b-dihydrocortisone and identified six that showed significant
metabolism. To reach a single functional clone, we performed
10-fold serial dilutions of a selected positive pool of 2 3 104
UCs, by following positive sub-pools at the 2 3 103, 2 3 102,
and 2 3 101 UC levels. We then plated the 20-UCs positive
sub-pool and screened individual clones in a 96-well plate
format to reach a single positive clone: Hyd-red-1 (Figures 6C
and S6). Sequencing of Hyd-red-1 revealed that it likely origi-
nated from a Bifidobacterium sp. Analysis of the genetic context
of Hyd-red-1 in a PD scaffold revealed a single putative oxidore-
ductase in the cloned insert (Figure 6D). We then cloned and het-
erologously expressed this single gene and showed that it is
 ll
Resource

C D

G
F

(legend on next page)

Cell 181, 1661–1679, June 25, 2020 1673

 ll
Resource

indeed a 20b-HSDH (Figures 6C and 6D). A second round of
screening of PD-CL performed in a similar manner revealed a
different clone, Hyd-red-2, harboring the same gene and con-
firming our findings (Figure 6D). These results indicate that
combining MDM-Screen with a functional metagenomics
approach is a valid strategy to link MDM transformations to
metabolizing enzymes from diverse bacteria without the need
for bacterial isolation.
We then sought to further probe the biological relevance of the
discovered 20b-HSDH. Because it was discovered by heterolo-
gous expression of PD-derived DNA in E. coli, we wondered
whether it is actually expressed under host colonization condi-
tions. To answer this question, we isolated RNA from PD, sub-
jected it to deep metatranscriptomic sequencing, and mapped
resulting reads to the PD scaffold harboring the 20b-HSDH
gene (see STAR Methods). We observed robust expression of
the 20b-HSDH gene in PD-derived metatranscriptomic data
but not of neighboring genes, suggesting that it is expressed
individually and not as part of a gene cluster (Figure 6E). To
determine whether the identified 20b-HSDH is unique to PD or
widespread in the human population, we mapped fecal metage-
nomic reads from the same six human cohorts mentioned above
to the DNA sequence of its gene and to that of a previously iden-
tified 20a-HSDH gene from the gut microbiome isolate Clos-
tridium scindens ATCC 35704 for comparison (which converts
hydrocortisone to 20a-dihydrocortisone) (Ridlon et al., 2013).
While the C. scindens-derived 20a-HSDH gene was rare (present
in only 0.6% of subjects, on average), the PD-derived 20b-HSDH
gene was widespread in all cohorts (present in 36% of subjects,
on average), and its abundance varied widely between subjects
and cohorts (Figures 6F and 6G; Tables S4B and S4C).
Although Bifidobacterium adolescentis had been known to
convert hydrocortisone into 20b-dihydrocortisone for almost
40 years (Winter et al., 1982), no responsible enzymes have
been identified from it. Interestingly, while this manuscript was
under revision, a different study published the crystal structure
of a 20b-HSDH from B. adolescentis L2-32 (which is 98% iden-
tical to the 20b-HSDH we identified from the PD microbiome),
further corroborating our findings (Doden et al., 2019). As
mentioned above (Figure 2D), we also observed the production
of 20b-dihydrocortisone from hydrocortisone acetate when
incubated with PD. This transformation would require two steps:
deacetylation at the C21 hydroxyl, by a yet-unidentified enzyme
and reduction of the ketone at C20 by a 20b-HSDH. Interestingly,
when we incubated hydrocortisone acetate with either
P. distasonis or C. bolteae, it was deacetylated to yield hydrocor-
tisone but not further reduced, implying that the two metabolic
steps at play here can be uncoupled and performed by different
members of the microbiome in a sequential manner (Figure S2).
MDM Deglycosylation Occurs In Vivo
Although MDM-Screen is able to uncover novel microbiome-drug
interactions, it is unclear whether these results (observed ex vivo)
can be recapitulated within the gastrointestinal tract of a live
mammalian host (in vivo). To address this question, we sought
to monitor one MDM transformation, MDM deglycosylation of
FPs, in an in vivo pharmacokinetic study that is performed in a mi-
crobiome-dependent manner. Capecitabine was among the initial
hits that resulted from MDM-Screen and its modification yields a
novel metabolite (deglycocapecitabine) that has not been previ-
ously reported in humans or animals; we selected its MDM degly-
cosylation as a test case for in vivo studies and a proxy for other
FPs. We treated two groups of C57BL/6 mice with a cocktail of an-
tibiotics for 14 days to eliminate their native microbiome, then
colonized one group with PD while the control group remained
non-colonized (see STAR Methods). The two groups were then
treated with a single human-equivalent oral dose of capecitabine
(755 mg/kg), and blood and feces were collected from each
mouse at 0, 20, 40, 60, 120, and 240 min post-drug administration
(Figures 7A and 7B). We then quantified capecitabine and its me-
tabolites in the serial fecal and blood samples using HPLC-HRMS.
In blood samples, capecitabine and its major liver-derived metab-
olite (50 -deoxy-5-fluorocytidine), but not deglycocapecitabine,
were readily detected and showed no significant differences be-
tween the two groups (Figure S7). In fecal samples, however, de-
glycocapecitabine was detected from animals colonized with PD

Figure 6. A Functional Metagenomic Screening Approach to Identify a Metabolizing Enzyme

(A) Schematic representation of the functional metagenomic screening approach.
(B) A scatterplot comparing the coverage of assembled PD scaffolds (R2 Kbp, in RPKM) in the two metagenomic datasets (PD and PD-CL-100). Dots repre-
senting PD metagenomic scaffolds are colored and sized on the basis of their phylum-level taxonomic assignments and lengths, respectively, and as indicated in
the key on the right. For ease of visualization, only scaffolds with RPKM values %10 are shown in this plot (97% of all scaffolds R2 Kbp) (see also Table S4A for the
entire dataset).
(C) Functional metagenomic screening of the PD-CL library. Beginning with pools containing 2–6 3104 unique clones, pools were selected and further sub-pooled
based on their functional ability to convert hydrocortisone to 20b-dihydrocortisone. Produced 20b-dihydrocortisone levels were quantified using HPLC-HRMS as
AUC normalized to an internal standard. For each round, the pool producing the highest normalized signal of 20b-dihydrocortisone (signified by a red dot with a
black outline) was selected for further sub-pooling, until a unique clone encoding a 20b-HSDH activity was identified. A single 20b-HSDH gene from the positive
metagenomic clone was further verified by heterologous expression in E. coli, when cloned as the native sequence (cloned) or synthesized as codon-optimized
for E. coli (synth.), in comparison with an empty-vector control (empty vector).
(D) Genetic organization of the inserts from two unique clones identified using functional metagenomic screening for the 20b-HSDH activity (PD-CL-Hyd-red-1
and PD-CL-Hyd-red-2) in comparison with their corresponding scaffold assembled from the PD metagenome.
(E) A bar graph indicating the count of PD fecal metatranscriptomic reads that mapped to the discovered 20b-HSDH gene (red) and its flanking genes (gray).
(F) Heatmaps indicating the prevalence (in percent of subjects harboring the gene of interest) and abundance (in median RPKM for all positive subjects) of
20a-HSDH (from C. scindens) and 20b-HSDH (from the PD metagenome) across six gut metagenomic cohorts. The number of subjects analyzed in each cohort is
indicated on the right. For cohorts with multiple sub-cohorts, the values reported are the averages of the sub-cohort values.
(G) Jitter plots of 20a-HSDH gene (from C. scindens) and 20b-HSDH gene (from the PD metagenome) abundances (in RPKM) for all positive subjects from the
same cohorts.
See also Figure S6; Table S4.

1674 Cell 181, 1661–1679, June 25, 2020


Resource
A Figure 7. MDM Deglycosylation Occurs
B C
as early as 20 min after dosing and was almost completely absent
in non-colonized ones (Figure 7C). These results indicate that—at
least in the case of FP deglycosylation—MDM transformations
observed ex vivo by MDM-Screen are recapitulated in vivo (i.e.,
in mice); establishing the same results in humans awaits further
studies. They also suggest that MDM deglycosylation of certain
FPs (e.g., doxifluridine, which is prematurely activated into 5-FU
upon deglycosylation) should be investigated as a potential
contributor to their undesired intestinal toxicity observed in the
clinic, although future in vivo studies with different dosing regi-
mens and a variety of FPs need to be performed.
DISCUSSION
In the current study, we developed a quantitative experimental
workflow for assessing the ability of the human gut microbiome
to directly metabolize orally administered drugs, using a combi-
nation of microbial community cultivation, small-molecule struc-
tural analysis, quantitative metabolomics, functional genomics
and metagenomics, and mouse colonization assays. Several
key differences set our approach apart from previous studies
in this area. First, instead of relying on single isolates in perform-
ing the initial screen, we use well-characterized, subject-person-
alized microbial communities. Despite the technical challenges
associated with characterizing and maintaining stable microbial
communities in batch cultures, three main advantages make this
strategy worth pursuing: (1) the extent of a biochemical transfor-
mation performed by single isolates cultured individually may be
different than that performed by the same isolates when cultured
as part of a complex community; (2) the net result of several
members of the microbiome acting on the same drug can only
be identified in mixed communities and not in single-isolate ex-
periments, unless all pairwise and higher order permutations
are tested; and (3) our strategy is ‘‘personalized.’’ The results
ll
In Vivo
(A) Schematic representation of the microbiome-
dependent pharmacokinetic experiment per-
formed here.
(B) Design of the capecitabine pharmacokinetic
experiment. Mice are treated with antibiotics for
14 days, then colonized with PD (n = 6) or left non-
colonized (n = 6). On the pharmacokinetic experi-
ment day, a single human-equivalent dose is
administered to mice using oral gavage, and serial
sampling of blood (B) and feces (F) is performed at
0, 20, 40, 60, 120, and 240 min post dosing.
(C) HPLC-HRMS based quantification of deglyco-
capecitabine in fecal samples from mice colonized
with PD in comparison to non-colonized ones.
Metabolite AUC per gram of feces is normalized by
the AUC of the internal standard (see STAR
Methods). Error bars represent the standard error of
the mean. The difference between the two condi-
tions is significant (p < 0.01, determined by testing
the intersection null hypothesis with marginal two-
tailed t tests using the Bonferroni correction to
control family-wise error rate).
See also Figure S7.
obtained here—including the extent and type of certain modifi-
cations—are specific to the strain-level composition of each
donor’s microbiome. MDM-Screen thus has a good potential
for assessing inter-individual variability in MDM.
Second, most previous studies have focused on certain combi-
nations of drugs and species that have historically been deemed
important (e.g., have been readily observed in humans) or that
are manageable experimentally. By default, our microbial-com-
munity setup allows us to screen a wider range of combinations,
which enabled us to expand in either the drug or subject spaces
and to discover drug-microbiome interactions never reported
before. Notably, while this manuscript was under revision, an
elegant study reported the screening of 271 orally administered
drugs against 76 bacterial isolates of the human gut microbiome
(Zimmermann et al., 2019a). Two thirds of the tested drugs were
shown to be significantly depleted by at least one of the tested iso-
lates, further emphasizing the great potential of gut microbes to
metabolize orally administered small-molecule drugs. We view
these two approaches as complementary: while screening drugs
against optimized, well-characterized, donor-derived microbial
communities in MDM-Screen provides a personalized view of
drug metabolism that takes into account strain-level and commu-
nity-wide contributions, screening drugs against a set of repre-
sentative gut isolates streamlines the identification and character-
ization of specific taxon-drug and gene-drug interactions.
Combined together, the results from the two approaches serve
as a valuable resource for the scientific community to further study
the mechanistic details and pharmacological consequences of
newly discovered drug-microbiome interactions.
Despite these advances, our approach is still subject to several
limitations. First, 24% of the drugs tested failed to be analyzed us-
ing the general analytical chemistry workflow described in MDM-
Screen. These drugs fell into one or more of three main categories:
unstable after overnight incubation in no-microbiome controls,
Cell 181, 1661–1679, June 25, 2020 1675
 ll
could not be extracted using ethyl acetate, or could not be
analyzed using reverse phase chromatography. An alternative
chemical analysis method will need to be developed for these
molecules in order to assess their MDM. Second, we focused
initially on oral drugs, yet several parenteral drugs and their
liver-derived metabolites may be subject to important MDM trans-
formations after biliary secretion. Third, even in our most diverse
ex vivo cultures, we fail to support the growth of 100% of the com-
munity in the original sample. This limitation can potentially be
overcome by utilizing multiple distinct media conditions that
each captures unique portions of the community. ENDS provides
the theoretical framework for selecting an optimal ensemble of
media conditions, and we show in STAR Methods how to
compute a version of ENDS that estimates the number of detect-
able strains gained by testing additional media.
We developed our screen in two stages. We began with a sin-
gle human sample, PD, and incubated its ex vivo culture with 575
drugs. We then transitioned into a HT format with more rigorous
methods for media selection, drug and metabolite quantification,
and metabolite discovery and used these methods to screen
ex vivo cultures from 20 human donors against 23 drugs. A simul-
taneous expansion into hundreds of drugs and hundreds of
donor samples is necessary to reveal the complete biochemical
potential of MDM: it is very likely that the types of MDM transfor-
mations observed here are an underestimation of all possible
ones. With the HT experimental approach and automatic tar-
geted and untargeted metabolomic analyses developed here,
we have laid the groundwork for this expansion. Finally, and
most relevant from a clinical stand point, a direct comparison be-
tween drug metabolism outcomes in humans and in MDM-
Screen for the same cohort of donors is important to establish
which MDM transformations can be observed in humans, and
to quantify the magnitude by which inter-individual variability in
MDM-Screen recapitulates that which occurs in humans. Our
quantitative framework—on both the microbial community and
metabolomic angles—provides the necessary tools to perform
such comparison.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Human subject samples
B Bacterial strains and conditions
B Mice
d METHOD DETAILS
B Fecal sample processing for PD and D1-20
B ex vivo culture of PD
B ex vivo culture of D1-20
B 16S rRNA gene amplicon sequencing and analysis
B Measurement of biomass for cultured D1-20
1676 Cell 181, 1661–1679, June 25, 2020
Resource
B ex vivo screening of the drug library with PD
B Structural elucidation of selected metabolites
B Molecular networking analysis in PD screen
B Enrichment analysis for drugs in PD screen
B Gene abundance analysis in metagenomic cohorts
B ENDS (Expected Number of Detectable Strains)
B High-throughput screen with D1-20
B Targeted quantitative metabolomics analysis
B Untargeted metabolomics analysis
B Isolate screen for capecitabine
B Metagenomic library construction
B Functional screening of the metagenomic library
B Heterologous expression of PD-derived 20b-HSDH
B Metagenomic and metatranscriptomic analyses
B TP and UP gene deletions in E. coli BW25113
B MDM-Screen of capecitabine using E. coli mutants
B MDM-Screen of other FPs using E. coli mutants
B Microbiome-dependent pharmacokinetic experiment
d QUANTIFICATION AND STATISTICAL ANALYSIS
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.001.
ACKNOWLEDGMENTS
We would like to thank Wei Wang and the Lewis Sigler Institute sequencing
core facility for assistance with HT sequencing; Matthew Cahn and Abhishek
Biswas for assistance with sequencing data analysis; Shuo Wang for assis-
tance with the functional group analysis; Joseph Koos, A. James Link, and
Yuki Sugimoto for assistance with Mass Spectrometry; Riley Skeen-Gaar for
assistance with statistical analysis; Joseph Sheehan and Zemer Gitai for assis-
tance with obtaining the Keio library mutants; the Laboratory Animal Re-
sources at Princeton University for assistance with mouse studies; Janie
Kim for illustrating the graphical abstract; and members of the Donia lab for
useful discussions. We are grateful to the 21 anonymous donors who provided
the fecal samples that made this project possible. Figure S6 and a part of
Figure 7 were created with BioRender.com. Funding for this project has
been provided by an Innovation Award from the Department of Molecular
Biology, Princeton University and an NIH Director’s New Innovator Award
(1DP2AI124441), both to M.S.D. B.J. is funded by a New Jersey Commission
on Cancer Research Pre-doctoral award (DFHS18PPC056), Y.-C.J.L. is
funded by a training grant from the National Institute of General Medicine Sci-
ences (T32GM007388), and J.G.L. is funded by a National Science Foundation
Graduate Research Fellowship (2017249408).
AUTHOR CONTRIBUTIONS
M.S.D. conceived and directed the research, designed the study, and ob-
tained funding. P.C. and B.J. performed the PD culturing and screening exper-
iments. B.J. performed the D1-D20 culturing and screening experiments. P.C.,
B.J., Q.W., X.W., and M.S.D. analyzed the PD screening data and character-
ized the new metabolites. J.G.L. performed the statistical, computational, me-
tabolomic, and quantitative analyses. R.H. and Y.-C.J.L. performed the func-
tional metagenomic screening experiments. S.C. isolated metagenomic DNA
and RNA and prepared them for sequencing. M.S.D, B.J., and J.G.L. wrote
the manuscript, with input from all authors.
DECLARATION OF INTERESTS
M.S.D. is a member of the scientific advisory board of DeepBiome Therapeu-
tics. A patent is being filed by Princeton University for the use of quantitative
MDM-Screen to measure inter-individual variability in drug metabolism.

Resource
Received: March 18, 2019
Revised: January 7, 2020
Accepted: April 29, 2020
Published: June 10, 2020
REFERENCES
Almeida, A., Mitchell, A.L., Boland, M., Forster, S.C., Gloor, G.B., Tarkowska,
A., Lawley, T.D., and Finn, R.D. (2019). A new genomic blueprint of the human
gut microbiota. Nature 568, 499–504.
Azadkhan, A.K., Truelove, S.C., and Aronson, J.K. (1982). The disposition and
metabolism of sulphasalazine (salicylazosulphapyridine) in man. Br. J. Clin.
Pharmacol. 13, 523–528.
Baba, T., Ara, T., Hasegawa, M., Takai, Y., Okumura, Y., Baba, M., Datsenko,
K.A., Tomita, M., Wanner, B.L., and Mori, H. (2006). Construction of Escheri-
chia coli K-12 in-frame, single-gene knockout mutants: the Keio collection.
Mol Syst Biol 2. https://doi.org/10.1038/msb4100050.
Bäckhed, F., Ley, R.E., Sonnenburg, J.L., Peterson, D.A., and Gordon, J.I.
(2005). Host-bacterial mutualism in the human intestine. Science 307,
1915–1920.
Bankevich, A., Nurk, S., Antipov, D., Gurevich, A.A., Dvorkin, M., Kulikov, A.S.,
Lesin, V.M., Nikolenko, S.I., Pham, S., Prjibelski, A.D., et al. (2012). SPAdes: a
new genome assembly algorithm and its applications to single-cell
sequencing. J. Comput. Biol. 19, 455–477.
Benjamini, Y., and Hochberg, Y. (1995). Controlling the False Discovery Rate:
A Practical and Powerful Approach to Multiple Testing. J. R. Stat. Soc. B 57,
289–300.
Bokulich, N.A., Kaehler, B.D., Rideout, J.R., Dillon, M., Bolyen, E., Knight, R.,
Huttley, G.A., and Gregory Caporaso, J. (2018). Optimizing taxonomic classi-
fication of marker-gene amplicon sequences with QIIME 2’s q2-feature-clas-
sifier plugin. Microbiome 6, 90.
Bolyen, E., Rideout, J.R., Dillon, M.R., Bokulich, N.A., Abnet, C., Al-Ghalith,
G.A., Alexander, H., Alm, E.J., Arumugam, M., Asnicar, F., et al. (2018). QIIME
2: Reproducible, interactive, scalable, and extensible microbiome data sci-
ence. PeerJ Preprints 6, e27295v27292.
Brady, S.F., Chao, C.J., and Clardy, J. (2002). New natural product families
from an environmental DNA (eDNA) gene cluster. J. Am. Chem. Soc. 124,
9968–9969.
Brito, I.L., Yilmaz, S., Huang, K., Xu, L., Jupiter, S.D., Jenkins, A.P., Naisilisili,
W., Tamminen, M., Smillie, C.S., Wortman, J.R., et al. (2016). Mobile genes in
the human microbiome are structured from global to individual scales. Nature
535, 435–439.
Bullingham, R.E., Nicholls, A.J., and Kamm, B.R. (1998). Clinical pharmacoki-
netics of mycophenolate mofetil. Clin. Pharmacokinet. 34, 429–455.
Callahan, B.J., McMurdie, P.J., Rosen, M.J., Han, A.W., Johnson, A.J., and
Holmes, S.P. (2016). DADA2: High-resolution sample inference from Illumina
amplicon data. Nat. Methods 13, 581–583.
Caporaso, J.G., Lauber, C.L., Walters, W.A., Berg-Lyons, D., Huntley, J., Fi-
erer, N., Owens, S.M., Betley, J., Fraser, L., Bauer, M., et al. (2012). Ultra-
high-throughput microbial community analysis on the Illumina HiSeq and
MiSeq platforms. ISME J. 6, 1621–1624.
Clayton, T.A., Baker, D., Lindon, J.C., Everett, J.R., and Nicholson, J.K. (2009).
Pharmacometabonomic identification of a significant host-microbiome meta-
bolic interaction affecting human drug metabolism. Proc. Natl. Acad. Sci. USA
106, 14728–14733.
Cleary, J.M., Rosen, L.S., Yoshida, K., Rasco, D., Shapiro, G.I., and Sun, W.
(2017). A phase 1 study of the pharmacokinetics of nucleoside analog trifluri-
dine and thymidine phosphorylase inhibitor tipiracil (components of TAS-102)
vs trifluridine alone. Invest. New Drugs 35, 189–197.
Cohen, L.J., Kang, H.S., Chu, J., Huang, Y.H., Gordon, E.A., Reddy, B.V., Ter-
nei, M.A., Craig, J.W., and Brady, S.F. (2015). Functional metagenomic discov-
ery of bacterial effectors in the human microbiome and isolation of commen-
damide, a GPCR G2A/132 agonist. Proc. Natl. Acad. Sci. USA 112,
E4825–E4834.
ll
Datsenko, K.A., and Wanner, B.L. (2000). One-step inactivation of chromo-
somal genes in Escherichia coli K-12 using PCR products. Proc. Natl. Acad.
Sci. USA 97, 6640–6645.
Devendran, S., Méndez-Garcı́a, C., and Ridlon, J.M. (2017). Identification and
characterization of a 20b-HSDH from the anaerobic gut bacterium Butyricicoc-
cus desmolans ATCC 43058. J. Lipid Res. 58, 916–925.
Doden, H.L., Pollet, R.M., Mythen, S.M., Wawrzak, Z., Devendran, S., Cann, I.,
Koropatkin, N.M., and Ridlon, J.M. (2019). Structural and biochemical charac-
terization of 20b-hydroxysteroid dehydrogenase from Bifidobacterium adoles-
centis strain L2-32. J. Biol. Chem. 294, 12040–12053.
Elmer, G.W., and Remmel, R.P. (1984). Role of the intestinal microflora in clo-
nazepam metabolism in the rat. Xenobiotica 14, 829–840.
Falony, G., Joossens, M., Vieira-Silva, S., Wang, J., Darzi, Y., Faust, K., Kuril-
shikov, A., Bonder, M.J., Valles-Colomer, M., Vandeputte, D., et al. (2016).
Population-level analysis of gut microbiome variation. Science 352, 560–564.
Gardiner, P., Schrode, K., Quinlan, D., Martin, B.K., Boreham, D.R., Rogers,
M.S., Stubbs, K., Smith, M., and Karim, A. (1989). Spironolactone metabolism:
steady-state serum levels of the sulfur-containing metabolites. J. Clin. Phar-
macol. 29, 342–347.
Genee, H.J., Bali, A.P., Petersen, S.D., Siedler, S., Bonde, M.T., Gronenberg,
L.S., Kristensen, M., Harrison, S.J., and Sommer, M.O. (2016). Functional min-
ing of transporters using synthetic selections. Nat. Chem. Biol. 12, 1015–1022.
Gillespie, D.E., Brady, S.F., Bettermann, A.D., Cianciotto, N.P., Liles, M.R.,
Rondon, M.R., Clardy, J., Goodman, R.M., and Handelsman, J. (2002). Isola-
tion of antibiotics turbomycin a and B from a metagenomic library of soil micro-
bial DNA. Appl. Environ. Microbiol. 68, 4301–4306.
Goodman, A.L., Kallstrom, G., Faith, J.J., Reyes, A., Moore, A., Dantas, G., and
Gordon, J.I. (2011). Extensive personal human gut microbiota culture collec-
tions characterized and manipulated in gnotobiotic mice. Proc. Natl. Acad.
Sci. USA 108, 6252–6257.
Haiser, H.J., Gootenberg, D.B., Chatman, K., Sirasani, G., Balskus, E.P., and
Turnbaugh, P.J. (2013). Predicting and manipulating cardiac drug inactivation
by the human gut bacterium Eggerthella lenta. Science 341, 295–298.
Harrison, D.A., and Brady, A.R. (2004). Sample Size and Power Calculations
using the Noncentral t-distribution. Stata J. 4, 142–153.
Human Microbiome Project Consortium (2012). Structure, function and diver-
sity of the healthy human microbiome. Nature 486, 207–214.
Hunter, J.D. (2007). Matplotlib: A 2D Graphics Environment. Comput. Sci. Eng.
9, 90–95.
Iida, N., Dzutsev, A., Stewart, C.A., Smith, L., Bouladoux, N., Weingarten, R.A.,
Molina, D.A., Salcedo, R., Back, T., Cramer, S., et al. (2013). Commensal bac-
teria control cancer response to therapy by modulating the tumor microenvi-
ronment. Science 342, 967–970.
Ilett, K.F., Tee, L.B., Reeves, P.T., and Minchin, R.F. (1990). Metabolism of
drugs and other xenobiotics in the gut lumen and wall. Pharmacol. Ther.
46, 67–93.
Jorga, K., Fotteler, B., Heizmann, P., and Gasser, R. (1999). Metabolism and
excretion of tolcapone, a novel inhibitor of catechol-O-methyltransferase.
Br. J. Clin. Pharmacol. 48, 513–520.
Karmarkar, D., and Rock, K.L. (2013). Microbiota signalling through MyD88 is
necessary for a systemic neutrophilic inflammatory response. Immunology
140, 483–492.
Kearse, M., Moir, R., Wilson, A., Stones-Havas, S., Cheung, M., Sturrock, S.,
Buxton, S., Cooper, A., Markowitz, S., Duran, C., et al. (2012). Geneious Basic:
an integrated and extendable desktop software platform for the organization
and analysis of sequence data. Bioinformatics 28, 1647–1649.
Kim, N.K., Min, J.S., Park, J.K., Yun, S.H., Sung, J.S., Jung, H.C., and Roh, J.K.
(2001). Intravenous 5-fluorouracil versus oral doxifluridine as preoperative
concurrent chemoradiation for locally advanced rectal cancer: prospective
randomized trials. Jpn. J. Clin. Oncol. 31, 25–29.
Kimura, T., Sudo, K., Kanzaki, Y., Miki, K., Takeichi, Y., Kurosaki, Y., and Na-
kayama, T. (1994). Drug absorption from large intestine: physicochemical fac-
tors governing drug absorption. Biol. Pharm. Bull. 17, 327–333.
Cell 181, 1661–1679, June 25, 2020 1677
 ll
Koppel, N., Maini Rekdal, V., and Balskus, E.P. (2017). Chemical transforma-
tion of xenobiotics by the human gut microbiota. Science 356, eaag2770.
Koppel, N., Bisanz, J.E., Pandelia, M.E., Turnbaugh, P.J., and Balskus, E.P.
(2018). Discovery and characterization of a prevalent human gut bacterial
enzyme sufficient for the inactivation of a family of plant toxins. eLife 7, e33953.
Kuroiwa, M., Inotsume, N., Iwaoku, R., and Nakano, M. (1985). [Reduction of
dantrolene by enteric bacteria]. Yakugaku Zasshi 105, 770–774.
Kuroiwa, M., Inotsume, N., and Nakano, M. (1986). [Reduction of nicardipine,
calcium antagonist, with enteric bacteria]. Yakugaku Zasshi 106, 698–702.
Lamont, E.B., and Schilsky, R.L. (1999). The oral fluoropyrimidines in cancer
chemotherapy. Clin. Cancer Res. 5, 2289–2296.
Langmead, B., and Salzberg, S.L. (2012). Fast gapped-read alignment with
Bowtie 2. Nat. Methods 9, 357–359.
Lenz, H.J., Stintzing, S., and Loupakis, F. (2015). TAS-102, a novel antitumor
agent: a review of the mechanism of action. Cancer Treat. Rev. 41, 777–783.
Li, H., and Jia, W. (2013). Cometabolism of microbes and host: implications for
drug metabolism and drug-induced toxicity. Clin. Pharmacol. Ther. 94,
574–581.
Lindenbaum, J., Tse-Eng, D., Butler, V.P., Jr., and Rund, D.G. (1981). Urinary
excretion of reduced metabolites of digoxin. Am. J. Med. 71, 67–74.
Lloyd-Price, J., Mahurkar, A., Rahnavard, G., Crabtree, J., Orvis, J., Hall, A.B.,
Brady, A., Creasy, H.H., McCracken, C., Giglio, M.G., et al. (2017). Strains,
functions and dynamics in the expanded Human Microbiome Project. Nature
550, 61–66.
Longley, D.B., Harkin, D.P., and Johnston, P.G. (2003). 5-fluorouracil: mecha-
nisms of action and clinical strategies. Nat. Rev. Cancer 3, 330–338.
Maier, L., Pruteanu, M., Kuhn, M., Zeller, G., Telzerow, A., Anderson, E.E., Bro-
chado, A.R., Fernandez, K.C., Dose, H., Mori, H., et al. (2018). Extensive
impact of non-antibiotic drugs on human gut bacteria. Nature 555, 623–628.
Maini Rekdal, V., Bess, E.N., Bisanz, J.E., Turnbaugh, P.J., and Balskus, E.P.
(2019). Discovery and inhibition of an interspecies gut bacterial pathway for
Levodopa metabolism. Science 364, eaau6323.
Mannens, G., Huang, M.L., Meuldermans, W., Hendrickx, J., Woestenborghs,
R., and Heykants, J. (1993). Absorption, metabolism, and excretion of risper-
idone in humans. Drug Metab. Dispos. 21, 1134–1141.
McDonald, D., Price, M.N., Goodrich, J., Nawrocki, E.P., DeSantis, T.Z.,
Probst, A., Andersen, G.L., Knight, R., and Hugenholtz, P. (2012). An improved
Greengenes taxonomy with explicit ranks for ecological and evolutionary ana-
lyses of bacteria and archaea. ISME J. 6, 610–618.
McDonald, D., Hyde, E., Debelius, J.W., Morton, J.T., Gonzalez, A., Acker-
mann, G., Aksenov, A.A., Behsaz, B., Brennan, C., Chen, Y., et al.; American
Gut Consortium (2018). American Gut: an Open Platform for Citizen Science
Microbiome Research. mSystems 3, e00031-18.
McKinney, W.G. (2010). Data structures for statistical computing in python.
Proceedings of the 9th Python in Science Conference. 445, 52–56.
Meinl, W., Sczesny, S., Brigelius-Flohé, R., Blaut, M., and Glatt, H. (2009).
Impact of gut microbiota on intestinal and hepatic levels of phase 2 xenobi-
otic-metabolizing enzymes in the rat. Drug Metab. Dispos. 37, 1179–1186.
Meuldermans, W., Hendrickx, J., Mannens, G., Lavrijsen, K., Janssen, C.,
Bracke, J., Le Jeune, L., Lauwers, W., and Heykants, J. (1994). The meta-
bolism and excretion of risperidone after oral administration in rats and
dogs. Drug Metab. Dispos. 22, 129–138.
Min, J.S., Kim, N.K., Park, J.K., Yun, S.H., and Noh, J.K. (2000). A prospective
randomized trial comparing intravenous 5-fluorouracil and oral doxifluridine as
postoperative adjuvant treatment for advanced rectal cancer. Ann. Surg. On-
col. 7, 674–679.
Nayfach, S., Shi, Z.J., Seshadri, R., Pollard, K.S., and Kyrpides, N.C. (2019).
New insights from uncultivated genomes of the global human gut microbiome.
Nature 568, 505–510.
Nielsen, H.B., Almeida, M., Juncker, A.S., Rasmussen, S., Li, J., Sunagawa, S.,
Plichta, D.R., Gautier, L., Pedersen, A.G., Le Chatelier, E., et al.; MetaHIT Con-
sortium; MetaHIT Consortium (2014). Identification and assembly of genomes
1678 Cell 181, 1661–1679, June 25, 2020
Resource
and genetic elements in complex metagenomic samples without using refer-
ence genomes. Nat. Biotechnol. 32, 822–828.
Northfield, T.C., and McColl, I. (1973). Postprandial concentrations of free and
conjugated bile acids down the length of the normal human small intestine. Gut
14, 513–518.
O’Boyle, N.M., Banck, M., James, C.A., Morley, C., Vandermeersch, T., and
Hutchison, G.R. (2011). Open Babel: An open chemical toolbox.
J. Cheminform. 3, 33.
Oliphant, T.E. (2006). A guide to NumPy (USA: Trelgol Publishing).
Pasolli, E., Asnicar, F., Manara, S., Zolfo, M., Karcher, N., Armanini, F., Beghini,
F., Manghi, P., Tett, A., Ghensi, P., et al. (2019). Extensive Unexplored Human
Microbiome Diversity Revealed by Over 150,000 Genomes from Metage-
nomes Spanning Age, Geography, and Lifestyle. Cell 176, 649–662.
Peppercorn, M.A., and Goldman, P. (1972). The role of intestinal bacteria in the
metabolism of salicylazosulfapyridine. J. Pharmacol. Exp. Ther. 181, 555–562.
Planer, J.D., Peng, Y., Kau, A.L., Blanton, L.V., Ndao, I.M., Tarr, P.I., Warner,
B.B., and Gordon, J.I. (2016). Development of the gut microbiota and mucosal
IgA responses in twins and gnotobiotic mice. Nature 534, 263–266.
Qin, J., Li, R., Raes, J., Arumugam, M., Burgdorf, K.S., Manichanh, C., Nielsen,
T., Pons, N., Levenez, F., Yamada, T., et al.; MetaHIT Consortium (2010). A hu-
man gut microbial gene catalogue established by metagenomic sequencing.
Nature 464, 59–65.
Qin, J., Li, Y., Cai, Z., Li, S., Zhu, J., Zhang, F., Liang, S., Zhang, W., Guan, Y.,
Shen, D., et al. (2012). A metagenome-wide association study of gut micro-
biota in type 2 diabetes. Nature 490, 55–60.
Rettedal, E.A., Gumpert, H., and Sommer, M.O. (2014). Cultivation-based
multiplex phenotyping of human gut microbiota allows targeted recovery of
previously uncultured bacteria. Nat. Commun. 5, 4714.
Ridlon, J.M., Ikegawa, S., Alves, J.M., Zhou, B., Kobayashi, A., Iida, T., Mita-
mura, K., Tanabe, G., Serrano, M., De Guzman, A., et al. (2013). Clostridium
scindens: a human gut microbe with a high potential to convert glucocorticoids
into androgens. J. Lipid Res. 54, 2437–2449.
Rondon, M.R., August, P.R., Bettermann, A.D., Brady, S.F., Grossman, T.H.,
Liles, M.R., Loiacono, K.A., Lynch, B.A., MacNeil, I.A., Minor, C., et al.
(2000). Cloning the soil metagenome: a strategy for accessing the genetic
and functional diversity of uncultured microorganisms. Appl. Environ. Micro-
biol. 66, 2541–2547.
Scheline, R.R. (1973). Metabolism of foreign compounds by gastrointestinal
microorganisms. Pharmacol. Rev. 25, 451–523.
Schmieder, R., and Edwards, R. (2011). Quality control and preprocessing of
metagenomic datasets. Bioinformatics 27, 863–864.
Schoenhard, G., Oppermann, J., and Kohn, F.E. (1985). Metabolism and phar-
macokinetic studies of misoprostol. Dig. Dis. Sci. 30 (11, Suppl), 126S–128S.
Shannon, P., Markiel, A., Ozier, O., Baliga, N.S., Wang, J.T., Ramage, D., Amin,
N., Schwikowski, B., and Ideker, T. (2003). Cytoscape: a software environment
for integrated models of biomolecular interaction networks. Genome Res. 13,
2498–2504.
Sica, D.A. (2005). Pharmacokinetics and pharmacodynamics of mineralocorti-
coid blocking agents and their effects on potassium homeostasis. Heart Fail.
Rev. 10, 23–29.
Sivan, A., Corrales, L., Hubert, N., Williams, J.B., Aquino-Michaels, K., Earley,
Z.M., Benyamin, F.W., Lei, Y.M., Jabri, B., Alegre, M.L., et al. (2015).
Commensal Bifidobacterium promotes antitumor immunity and facilitates
anti-PD-L1 efficacy. Science 350, 1084–1089.
Smith, K.S., Smith, P.L., Heady, T.N., Trugman, J.M., Harman, W.D., and Mac-
donald, T.L. (2003). In vitro metabolism of tolcapone to reactive intermediates:
relevance to tolcapone liver toxicity. Chem. Res. Toxicol. 16, 123–128.
Sommer, M.O.A., Dantas, G., and Church, G.M. (2009). Functional character-
ization of the antibiotic resistance reservoir in the human microflora. Science
325, 1128–1131.

Resource
Spanogiannopoulos, P., Bess, E.N., Carmody, R.N., and Turnbaugh, P.J.
(2016). The microbial pharmacists within us: a metagenomic view of xenobiotic
metabolism. Nat. Rev. Microbiol. 14, 273–287.
Storey, J.D. (2002). A Direct Approach to False Discovery Rates. J. R. Stat.
Soc. Series B Stat. Methodol. 64, 479–498.
Sugimoto, Y., Camacho, F.R., Wang, S., Chankhamjon, P., Odabas, A., Bis-
was, A., Jeffrey, P.D., and Donia, M.S. (2019). A metagenomic strategy for har-
nessing the chemical repertoire of the human microbiome. Science 366,
eaax9176.
Taylor, M.R., Flannigan, K.L., Rahim, H., Mohamud, A., Lewis, I.A., Hirota,
S.A., and Greenway, S.C. (2019). Vancomycin relieves mycophenolate mofe-
til-induced gastrointestinal toxicity by eliminating gut bacterial b-glucuroni-
dase activity. Sci Adv 5, eaax2358.
Temmink, O.H., de Bruin, M., Turksma, A.W., Cricca, S., Laan, A.C., and Pe-
ters, G.J. (2007). Activity and substrate specificity of pyrimidine phosphory-
lases and their role in fluoropyrimidine sensitivity in colon cancer cell lines.
Int. J. Biochem. Cell Biol. 39, 565–575.
Tramontano, M., Andrejev, S., Pruteanu, M., Klünemann, M., Kuhn, M., Galar-
dini, M., Jouhten, P., Zelezniak, A., Zeller, G., Bork, P., et al. (2018). Nutritional
preferences of human gut bacteria reveal their metabolic idiosyncrasies. Nat.
Microbiol. 3, 514–522.
Tsai, B.S., Kessler, L.K., Stolzenbach, J., Schoenhard, G., and Bauer, R.F.
(1991). Expression of gastric antisecretory and prostaglandin E receptor bind-
ing activity of misoprostol by misoprostol free acid. Dig. Dis. Sci. 36, 588–593.
Uribe, R.V., van der Helm, E., Misiakou, M.A., Lee, S.W., Kol, S., and Sommer,
M.O.A. (2019). Discovery and Characterization of Cas9 Inhibitors Dissemi-
nated across Seven Bacterial Phyla. Cell Host Microbe 25, 233–241.
van Kessel, S.P., Frye, A.K., El-Gendy, A.O., Castejon, M., Keshavarzian, A.,
van Dijk, G., and El Aidy, S. (2019). Gut bacterial tyrosine decarboxylases
ll
restrict levels of levodopa in the treatment of Parkinson’s disease. Nat. Com-
mun. 10, 310.
Vétizou, M., Pitt, J.M., Daillère, R., Lepage, P., Waldschmitt, N., Flament, C.,
Rusakiewicz, S., Routy, B., Roberti, M.P., Duong, C.P., et al. (2015). Anticancer
immunotherapy by CTLA-4 blockade relies on the gut microbiota. Science
350, 1079–1084.
Wallace, B.D., Wang, H., Lane, K.T., Scott, J.E., Orans, J., Koo, J.S., Venka-
tesh, M., Jobin, C., Yeh, L.A., Mani, S., and Redinbo, M.R. (2010). Alleviating
cancer drug toxicity by inhibiting a bacterial enzyme. Science 330, 831–835.
Wang, M., Carver, J.J., Phelan, V.V., Sanchez, L.M., Garg, N., Peng, Y.,
Nguyen, D.D., Watrous, J., Kapono, C.A., Luzzatto-Knaan, T., et al. (2016).
Sharing and community curation of mass spectrometry data with Global Nat-
ural Products Social Molecular Networking. Nat. Biotechnol. 34, 828–837.
Winter, J., Cerone-McLernon, A., O’Rourke, S., Ponticorvo, L., and Bokken-
heuser, V.D. (1982). Formation of 20 b-dihydrosteroids by anaerobic bacteria.
J. Steroid Biochem. 17, 661–667.
Wood, D.E., and Salzberg, S.L. (2014). Kraken: ultrafast metagenomic
sequence classification using exact alignments. Genome Biol. 15, R46.
Zampino, M.G., Colleoni, M., Bajetta, E., Stampino, C.G., Guenzi, A., and de
Braud, F. (1999). Pharmacokinetics of oral doxifluridine in patients with colo-
rectal cancer. Tumori 85, 47–50.
Zimmermann, M., Zimmermann-Kogadeeva, M., Wegmann, R., and
Goodman, A.L. (2019a). Mapping human microbiome drug metabolism by
gut bacteria and their genes. Nature 570, 462–467.
Zimmermann, M., Zimmermann-Kogadeeva, M., Wegmann, R., and
Goodman, A.L. (2019b). Separating host and microbiome contributions to
drug pharmacokinetics and toxicity. Science 363, eaat9931.
Cell 181, 1661–1679, June 25, 2020 1679
 ll
Resource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT OR RESOURCE SOURCE IDENTIFIER

Chemicals, Media, and Reagents
7a-Thiospironolactone Toronto Research Chemicals Cat#T375000
10mM dNTP mix Life Technologies Cat#18427-088
2-Propanol Fisher Scientific Cat#A451-4
20-alpha-dihydrocortisol MuseChem Cat#M122174
20-beta-dihydrocortisol MuseChem Cat#R060042
Acetonitrile Fisher Scientific Cat#A998-4
Allopurinol Sigma Cat#A8003
Ammonium Chloride Sigma-Aldrich Cat#A4514-500 g
Ampicillin Sigma-Aldrich Cat#A1593-25G
Antarctic Phosphatase New England Biolabs (NEB) Cat#M0289S
Aspartame Sigma Cat#47135
Azathioprine Sigma Cat#A4638
Bacto Agar Fisher Scientific Cat#214010
BD Bacto Peptone Fisher Scientific Cat#S71604
Beef Extract Fisher scientific Cat#S25661A
Bisacodyl Sigma Cat#B1390
Brain Heart Infusion VWR Cat#90003-032 (EA)
Brewer Thioglycollate Medium Sigma Cat#B2551-500 g
Bryant and Burkey Medium Sigma Cat#91903-500 g
Calcium Chloride, dihydrate Sigma Cat#C8106-500 g
Capecitabine Sigma Cat#PHR1405
Clofazimine Sigma Cat#C8895
Clonazepam Sigma Cat#C1277
Cooked Meat Broth, Microbiology Sigma Cat#60865-500 g
D-(+)-Glucose SIGMA Cat#G8270-1KG
Digoxin Sigma Cat#D6003
Dihydrodigoxin Toronto Research Chemicals Cat#D452680
Dimethyl sulfoxide Sigma Cat#D8418-100ML
Dolasetron Sigma Cat#CDS021594
Doxifluridine Sigma Cat#F8791-100MG
Dpn1 New England Biolabs (NEB) Cat#R0176S
EDTA, disodium salt Sigma-Aldrich Cat#E5134-1kg
Ethyl Acetate Fisher Scientific Cat#E195-4
Famciclovir Sigma Cat#F7932
Formic acid Fisher Scientific Cat#A117-50
GAM Broth Modified HyServe Cat#5433
Glycerol Fisher Scientific Cat#BP2291
Hemin BioXtra Sigma-Aldrich Cat#51280-5G
Hydrocortisone Sigma Cat#H4001
IPTG, dioxane-free Thermo Scientific Cat#R0393
Kanamycin Sigma Cat#K4000-5G
Ketoconazole Sigma Cat#K1003
Ketoprofen Sigma Cat#K1751
(Continued on next page)

e1 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

Continued
REAGENT OR RESOURCE SOURCE IDENTIFIER
L-arabinose Sigma Cat#A3256-25G
L-Cysteine Fisher Scientific Cat#ICN19464625
LB Broth Sigma Cat#L3522-1Kg
Levonogestrel Santa Cruz Biotech Cat#SC205731
Liver broth Sigma Cat#61724-500 g
Lovastatin Sigma Cat#PHR1285
M17 BROTH Fisher Scientific Cat#OXCMO817B
Magnesium Sulfate, heptahydrate Sigma Cat#230391-500 g
Meat extract Sigma-Aldrich Cat#70164-500G
MegaX DH10b Electrocompetent Cells Thermo Fisher scientific Cat#C640003
Methanol Fisher Scientific Cat#A452-4
Metronidazole Sigma Cat#M3761
Milli-q water Millipore Corporation N/A
Misoprostol Sigma Cat#M6807
Misoprostol free acid Sigma Cat#M6932
MRS Broth Sigma/Millipore Cat#69966-500G
Mycophenolic acid Sigma Cat#Cat#M5255
Mycophenolate mofetil Sigma Cat#SML0284
Neomycin Sigma-Aldrich Cat#N1876-25G
Nicardipine Sigma Cat#N7510
Nitrofurantoin Sigma Cat#N7878
Pellet Paint NF Co-Precipitant EMD Millipore Cat#70748
pGFPuv Clontech Laboratories Cat#632312
Phusion High-Fidelity DNA Polymerase New England Biolabs (NEB) Cat#M0530S
Plasmid pCP20 encoding the FLP Gitai lab, Princeton University (Datsenko and Wanner, 2000)
recombinase
Plasmid pKD46 expressing the Lambda Fischbach lab, Stanford University (Datsenko and Wanner, 2000)
Red recombinase
Potassium phosphate monobasic Sigma Cat#P5655-500G
Praziquantel Sigma Cat#P4668
Reinforced Clostridial Medium Fisher Scientific Cat#DF1808-17-3
Resazurin Santa Cruz Biotechnology Cat#62758-13-8
Risperidone Sigma Cat#PHR1631
Ropinirole Sigma Cat#R2530
Screen-Well FDA approved drug library Enzo Life Sciences Cat#BML-2843-0100
Sodium Bicarbonate Sigma Cat#S6014-500 g
Sodium chloride Sigma-Aldrich Cat#S7653-5KG
Sodium hydroxide Sigma-Aldrich Cat#795429-500G
Spironolactone Sigma Cat#S3378
Sulfamethoxazole Sigma Cat#S7507
Sulfasalazine Sigma Cat#S0883
Sulindac Cayman Cat#38194-50-2
T4 DNA ligase NEB Cat#M0202T
Terrific Broth Fisher Scientific Cat#DF0438-17
Thioglycolate Broth Sigma Cat#70157-500 g
Tinidazole Santa Cruz Cat#sc-205862
Tolcapone Sigma Cat#SML0150
Trace Mineral Supplement ATTC ATCC MD-TMS
(Continued on next page)

Cell 181, 1661–1679.e1–e15, June 25, 2020 e2

 ll
Resource

Continued
REAGENT OR RESOURCE SOURCE IDENTIFIER
Trifluridines Sigma Cat#T2255-100MG
Tryptic Soy Broth Fisher Scientific Cat#DF0370 17 3
Tryptic Soy Broth Thermo Fischer Cat#DF0370-17-3
Trypticase Peptone VWR Cat#90000-434 (EA)
Tween 80 Sigma-Aldrich Cat#P1754-500ml
Vancomycin Sigma-Aldrich Cat#V2002-1G
Vitamin K1 Sigma-Aldrich Cat#95271-250MG
Vitamin Mix ATCC ATCC MD-VS
Voriconazole Cayman Cat#15633
Vorinostat Sigma Aldrich Cat#SML0061
Yeast extract VWR Cat#90000-026 (EA)
Zidovudine Sigma Cat#PHR1292
Experimental models: Organisms/Strains
Anaerococcus prevotii Fischbach lab, Stanford University N/A
Anaerostipes caccae DSM 14662 DSMZ DSM 14662
Clostridium bolteae ATCC BAA-613 ATCC ATCC BAA-613
E. coli BW25113 Keio collection (Baba et al., 2006)
E. coli BW25113 mutants harboring Keio collection (Baba et al., 2006)
replacement of deoA or udp with a
kanamycin resistance gene
E. coli BW25113 clean mutants of deoA or This study Methods
udp or both
E.coli BL21 NEB Cat#C2527I
E.coli Stellar Clontech Laboratories Cat#636763
Enterococcus faecalis TYG’11 This paper Methods
Escherichia coli TYG1 This paper Methods
Escherichia coli TYG2 This paper Methods
Lactobacillus gasseri JV-V03 BEI Resources HM-104
Parabacteroides distasonis CL09T03C24 Comstock lab, Harvard Medical School N/A
Prevotella bivia ATCC 29303 ATCC ATCC 29303
Salmonella enterica SL484 Ravel lab, UMB N/A
Serratia marcescens Fischbach lab, Stanford University N/A
Oligonucleotides
Primers used in this study This study Methods
Software and Algorithms
Adobe Illustrator Adobe N/A
Agilent Profinder 8.0 Agilent N/A
Agilent Qualitative Analysis 10.0 Agilent N/A
Agilent Quantiative Analysis 9.0 Agilent N/A
Bowtie2 (Langmead and Salzberg, 2012) http://bowtie-bio.sourceforge.net/bowtie2/
index.shtml
Blast 2.7.1+ NCBI https://blast.ncbi.nlm.nih.gov/Blast.cgi
ChemDraw Professional 16.0 PerkinElmer N/A
Cytoscape (Shannon et al., 2003) https://cytoscape.org/
ENDS This study https://github.com/donia-lab/
personalized_community_MDM_screen
Geneious R9 Geneious https://www.geneious.com/
Global Natural Products Social Molecular (Wang et al., 2016) https://gnps.ucsd.edu
Networking
(Continued on next page)

e3 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

Continued
REAGENT OR RESOURCE SOURCE IDENTIFIER
Illumina HiSeq Control Software Illumina N/A
Kraken-1.1.1 (Wood and Salzberg, 2014) http://ccb.jhu.edu/software/kraken
MATLAB R2017b Mathworks https://www.mathworks.com
Matplotlib 2.1.0 (Hunter, 2007) https://matplotlib.org/
MestreNova 10.0 Mestrelab Research N/A
NumPy 1.14.2 (Oliphant, 2006) https://www.numpy.org/
Open Babel (O’Boyle et al., 2011) http://openbabel.org
Pandas 0.22.0 (McKinney, 2010) https://pandas.pydata.org/
PRINSEQ-lite v0.20.4B (Schmieder and Edwards, 2011) http://prinseq.sourceforge.net/
ProteoWizard N/A http://proteowizard.sourceforge.net/
Python 3.6.3 Python Core Team https://www.python.org/
QIIME2 version 2018.6 (Bolyen et al., 2018) https://qiime2.org
SPAdes v3.11.0 (Bankevich et al., 2012) http://cab.spbu.ru/software/spades/
Targeted and untargeted metabolomics This study https://github.com/donia-lab/
pipeline personalized_community_MDM_screen
Deposited Data
Sequencing data This study NCBI BioProject: PRJNA593062
Metabolomics data This study MassIVE: MSV000084641
Critical Commercial Assays
DNeasy Power Soil Kit(100) QIAGEN Cat#12888-100
End-It DNA End-Repair Kit Epicenter Cat#ER0720
In-Fusion HD Cloning System (50RXNS) TAKARA Cat#639646
Qiaprep Spin mini prep(250) kit QIAGEN Cat#27106
Qiaquick Gel Extraction Kit,250 QIAGEN Cat#8706
QIAquick PCR Purification kit QIAGEN Cat#28106
RNAlater Stabilization Solution Fisher Scientific Cat#AM7021
Zymoclean Large Fragment DNA Zymo Research Cat#D4046
Recovery Kit
Animal Models
C57BL/6 mice Jackson Laboratories (Jax-West facility) N/A
Others
96-well plate, 1.0 mL, round wells, U shape, Agilent Cat#5043-9305
polypropylene, 32 mm, 50/pk
Agilent 1290 Infinity II LC System Agilent N/A
Agilent 2100 Bioanalyzer Agilent N/A
Agilent 6120 quadrupole mass Agilent N/A
spectrometer
Agilent 6530 Q-TOF LC/MS equipment Agilent N/A
Agilent 6545 Q-TOF LC/MS equipment Agilent N/A
Agilent Eclipse Plus C18 RRHD column Agilent Part Number 959757-902
1.8 mM (2.1 3 50 mm)
Agilent Poroshell 120 EC-C18 column Agilent Part Number 695775-902
2.7um (2.1x100mm)
Agilent Poroshell 120 EC-C18 column Agilent Part Number 695975-902
2.7um (4.6x100mm)
Agilent Poroshell 120 EC-C18 column Agilent Part Number 699975-902
2.7um (4.6x50mm)
Anaerobic chamber COY N/A
EquaVAP 96-Well Blowdown Evaporator Fisher Analytical Sales & Services Cat#23096
(Continued on next page)

Cell 181, 1661–1679.e1–e15, June 25, 2020 e4

 ll
Resource

Continued
REAGENT OR RESOURCE SOURCE IDENTIFIER
g-TUBE Covaris N/A
Mega Bond Elut-C18 10 g Agilent Part Number 12256031
NanoDrop 2000 Thermo Fisher Scientific N/A
Nunc 96-Well Cap Mats, case of 50 ThermoFisher Cat#276000
Nunc 96-Well Polypropylene DeepWell FisherScientific Cat#278743
Storage Plates, case of 60, 2mL
Reservoir w/Lid 75 mL 25/Pk RV-L25 Mettler-Toledo Rainin LLC Cat#17007886
Sealing mat, 96 wells, round, preslitted, Agilent Part Number 5043-9317
silicone, 50/pk

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Mohamed
S. Donia donia@princeton.edu)

Materials Availability
All unique/stable reagents generated in this study are available from the Lead Contact, but we may require a completed Materials
Transfer Agreement if there is potential for commercial application.

Data and Code Availability

The sequencing datasets generated during this study are available in Table S2 and at NCBI BioProject: PRJNA593062. The metabolo-
mics datasets generated during this study are available in Tables S1 and S3, and at MassIVE: MSV000084641. The code generated dur-
ing this study is available at GitHub (https://github.com/donia-lab/personalized_community_MDM_screen).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human subject samples

Fecal samples were collected under Princeton University IRB#11606 at the Princeton University Department of Molecular Biology.
Healthy volunteers were recruited via e-mails sent to the departmental listserv as well as flyer advertisements. Volunteers gave
informed consent prior to sample collection. Eligibility criteria included age (18 and above) and health status (feeling well at time
of sample collection, no diabetes, gastrointestinal, oral, or skin infections, diseases, malignancies, or antibiotic use three months
prior to or during sample collection).

Bacterial strains and conditions

The following media were pre-reduced by incubation in the anaerobic chamber for 24 h before inoculation with the
corresponding isolate’s glycerol stock: (PYG for Anaerostipes caccae and Clostridium bolteae; RCM for Prevotella bivia, Parabacter-
oides distasonis; mGAM for Serratia marcescens, Enterococcus faecalis TYG’11, Anaerococcus prevotii, Escherichia coli
TYG’1 and Escherichia coli TYG’2; BHI for Lactobacillus gasseri; LB for Salmonella enterica and Escherichia coli BL21). See
below the complete media names. Cultures were grown overnight at 37
C in the same anaerobic chamber (70% N2, 25% CO2,
5% H2).
MegaX DH10b E. coli (used for the metagenomic library construction) and BL21-DE3 E. coli (used for the heterologous expression
of the discovered 20b-HSDH gene) were cultured aerobically in LB medium, at 37
C.
E. coli BW25113 wild type and mutants that harbor a replacement of deoA or udp with a kanamycin resistance gene were
obtained from the Keio collection (Baba et al., 2006) and cultured in LB medium at 37
C. Clean TP knockout (DdeoA), UP knockout
(Dudp), and TP/UP double knockout (DdeoA/Dudp) strains were generated as explained below, and also cultured in LB medium
at 37
C.

Mice
8-to-10-week-old male and female mice (25–30 g) C57BL/6 mice were purchased from Jackson laboratories. All animals
were housed and maintained in a certified animal facility and all experiments were conducted according to USA Public Health
Service Policy of Humane Care and Use of Laboratory Animals. All protocols were approved by the Institutional Animal Care
and Use Committee, protocol 2087-16 (Princeton University). The sex and number of animals are specified for the pharmacokinetic
study.

e5 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

METHOD DETAILS

Fecal sample processing for PD and D1-20

Freshly collected human fecal material from the healthy donors (30 min from collection for PD, transported on ice; < 15 min from
collection for the rest of the donors, transported without ice) was brought into an anaerobic chamber (70% N2, 25% CO2, 5% H2). One
gram of the sample was suspended in 15 mL of sterile phosphate buffer supplemented with 0.1% L-cysteine (PBSc) in a 50 mL sterile
falcon tube. The suspension was left standing still for 5 min to let insoluble particles settle. The supernatant was mixed with an equal
volume of 40% glycerol in PBSc. Aliquots (1 mL) of this suspension were placed in sterile cryogenic vials and frozen at 80
C until
use (Goodman et al., 2011). Samples were assigned de-identifying numbers (PD, and D1-D20), and stored in the Donia laboratory.

ex vivo culture of PD
A small aliquot (20 mL) from a PD glycerol stock was used to inoculate 10 mL of 14 different media: Liver Broth (Liver), Brewer Thiogly-
colate Medium (BT), Bryant and Burkey Medium (BB), Cooked Meat Broth (Meat), Thioglycolate Broth (TB), Luria-Bertani Broth (LB) (ob-
tained from Sigma Aldrich, USA), Brain Heart Infusion (BHI), MRS (MRS), Reinforced Clostridium Medium (RCM), M17 (M17) (obtained
from Becton Dickinson, USA), modified Gifu Anaerobic Medium (mGAM) (obtained from HyServe, Germany), Gut Microbiota Medium
(GMM (Goodman et al., 2011)), TYG, and a 1:1 mix of each (BestMix), and cultures were incubated at 37
C in an anaerobic chamber.
One mL was harvested from each culture each day for 4 consecutive days, and centrifuged to recover the resulting bacterial pellets.

ex vivo culture of D1-20

30 mL from each donor glycerol stock was used to inoculate 3 mL of 10 different pre-reduced media: Liver Broth (Liver), Bryant and
Burkey Medium (BB), Thioglycolate Broth (TB), Luria-Bertani Broth (LB) (obtained from Sigma Aldrich, USA), Brain Heart Infusion
(BHI), MRS (MRS), Reinforced Clostridium Medium (RCM), modified Gifu Anaerobic Medium (mGAM) (obtained from HyServe, Ger-
many), Gut Microbiota Medium (GMM (Goodman et al., 2011)), and a 70:30 mix of BB:GAM (BG), and cultures were incubated at 37
C
in an anaerobic chamber. One mL was harvested from each culture after 48 h and centrifuged to recover the resulting bacterial
pellets.

16S rRNA gene amplicon sequencing and analysis

DNA was extracted from all pellets using the Power Soil DNA Isolation kit (Mo Bio Laboratories, USA, now QIAGEN), the 16S rRNA
gene was amplified (250 bp, V4 region), and Illumina sequencing libraries were prepared from the amplicons according to a pre-
viously published protocol and primers (Caporaso et al., 2012). Libraries were further pooled together at equal molar ratios and
sequenced on an Illumina HiSeq 2500 Rapid Flowcell (PD samples) or MiSeq (D1-D20 samples) as paired-end reads. For PD sam-
ples, these reads were 2X175 bp with an average depth of 100,000 reads, while for D1-D20 samples the reads were 2X150bp with
an average depth of 30,000 reads. Also included were 8 bp Index reads, following the manufacturer’s protocol (Illumina, USA). Raw
sequencing reads were filtered by Illumina HiSeq Control Software to generate Pass-Filter reads for further analysis. Different sam-
ples were de-multiplexed using the index reads. Amplicon sequencing variants (ASVs) were then inferred from the unmerged paired-
end sequences using the DADA2 plugin within QIIME2 version 2018.6 (Bolyen et al., 2018; Callahan et al., 2016). For PD samples, the
forward reads were trimmed at 165 bp and the reverse reads were trimmed at 140 bp. For D1-20 samples, the forward reads were
trimmed at 150bp and the reverse reads trimmed at 140bp. All other settings within DADA2 were default. Taxonomy was assigned to
the resulting ASVs with a naive Bayes classifier trained on the Greengenes database version 13.8 (Bokulich et al., 2018; McDonald
et al., 2012). Only the target region of the 16S rRNA gene was used to train the classifier. Downstream analyses were performed in
either MATLAB or Python (Hunter, 2007; McKinney, 2010; Oliphant, 2006). See Table S2.

Measurement of biomass for cultured D1-20

30 mL from each donor glycerol stock was cultured in 10 different pre-reduced media as previously described (in duplicates). One mL
was harvested from each culture after 48 h and centrifuged to recover the resulting bacterial pellets. The pellets were weighed in
Eppendorf tubes and the mass was subtracted from that of the empty tube prior to pellet collection. See Table S2.

ex vivo screening of the drug library with PD

In an anaerobic chamber, a small volume (100 mL) of a PD glycerol stock was diluted in 1 mL of mGAM, then 20 mL of this solution
was used to inoculate 3 mL of mGAM in culture tubes. Cultures were grown for 24 h at 37
C in an anaerobic chamber. After 24 h, 10 mL
of each drug (575 total drugs, a subset of the SCREEN-WELL FDA approved drug library, Enzo Life Sciences, Inc. with each mole-
cule having a concentration of 10 mM in DMSO) or of a DMSO control was added to the growing microbial community. In addition,
10 mL of each drug was also incubated similarly in a no-microbiome, mGAM control. The no-drug control distinguishes microbiome-
derived small molecules from ones that result from MDM, and the no-microbiome control distinguishes cases of passive drug degra-
dation or faulty chemical extraction from those of active MDM. PD-DMSO control pellets from several batches of the screen were
analyzed using high-throughput 16S rRNA gene sequencing as described above to ensure the maintenance of a similarly diverse
microbial composition. Experiments and controls were allowed to incubate under the same conditions for a second 24 h period. After
incubation, cultures were extracted with double the volume of ethyl acetate and the organic phase was dried under vacuum using a

Cell 181, 1661–1679.e1–e15, June 25, 2020 e6

 ll
Resource

rotary evaporator (Speed Vac). This extraction method recovers organic molecules from both cells and broths of the cultures, and
therefore is not affected by cases of bacterial sequestration of the parent drugs. The dried extracts were suspended in 250 mL
MeOH, centrifuged at 15000 rpm for 5 min to remove any particulates, and analyzed using HPLC-MS (Agilent Single Quad, column:
Poroshell 120 EC-C18 2.7mm 4.6 3 50mm, flow rate 0.8 mL/min, 0.1% formic acid in water (solvent A), 0.1% formic acid in acetonitrile
(solvent B), gradient: 1 min, 0.5% B; 1-20 min, 0.5%–100% B; 20-25 min, 100% B). We tested each drug twice, along with matching
no-drug and no-microbiome controls. If drugs were deemed positive for MDM in one or both of the two trials they were tested a third
time and analyzed using both HPLC-MS and HR-HPLC-MS/MS (Agilent QTOF, column: Poroshell 120 EC-C18 2.7mm 2.1x100 mm,
flow rate 0.25 mL/min, 0.1% formic acid in water (solvent A), 0.1% formic acid in acetonitrile (solvent B), gradient: 1 min, 0.5% B;
1-20 min, 0.5%–100% B; 25-30 min, 100% B). For selected molecules, cultures were scaled up and metabolites were isolated
and their structures were elucidated using NMR and/or comparison to an authentic standard obtained commercially using HPLC-
HRMS/MS (see below). See Data S1 for the chromatograms of all MDM+ metabolites.

Structural elucidation of selected metabolites

One mL of PD glycerol stock was used to inoculate 100 mL of mGAM medium and cultured for 24 h at 37
C in an anaerobic chamber.
After 24 h, 2 mL of 10 mM of either capecitabine, hydrocortisone, tolcapone, or misoprostol solutions was added to the PD culture
and incubated for another 24 h. After the second 24 h, the cultures were extracted with double the volume of ethyl acetate and the
organic solvent layer was dried under vacuum in a rotary evaporator. The dried extract was then suspended in MeOH and partitioned
by reversed phase flash column chromatography (Mega Bond Elut-C18 10 g, Agilent Technology, USA) using the following mobile
phase conditions: solvent A, water with 0.01% formic acid; solvent B acetonitrile with 0.01% formic acid, gradient, 100% A to 100% B
in 20% increments. Fractions containing the metabolites of interest were identified by HPLC-MS, and reverse phase HPLC was used
to purify each metabolite using a fraction collector. The purified metabolites were subjected to NMR and HR-MS/MS analysis. For
misoprostol, hydrocortisone, spironolactone, and mycophenolate mofetil, detailed HPLC-HRMS/MS comparisons with authentic
standards were also performed. Structural elucidation details of capecitabine, hydrocortisone, tolcapone, spironolactone, misopros-
tol, and mycophenolate mofetil metabolites are detailed in Data S2.

Molecular networking analysis in PD screen

Raw data files were converted to the .mzXML format using ProteoWizard and uploaded to the Global Natural Products Social Mo-
lecular Networking (GNPS) online platform (https://gnps.ucsd.edu) (Wang et al., 2016). The data was first filtered, removing MS/MS
peaks within ± 17 Da of the precursor m/z. MS/MS spectra were window filtered by choosing only the top 6 peaks in the ± 50 Da
window throughout the spectrum. Before networking, the data was dereplicated using MS-cluster with a parent mass and MS/
MS fragment mass tolerance of 0.1 Da and minimum fragment intensity of 1000. Following this consensus spectra were removed
if they contained less than 2 spectra. Molecular ion networking was then performed, requiring that two ions have a cosine similarity
of 0.5 and share at least 3 peaks in order to be linked. Connections were removed if the ions did not appear in each other’s top 10
most similar ions. Molecular networks were visualized and mined using Cytoscape (Shannon et al., 2003). We call the two
compounds (parent drug and metabolite) related if they are in the same connected component of the graph. In the cases where either
the metabolite or the parent drug or both were not picked up in the molecular ion networking analysis, we deem the linkage ‘‘unde-
termined.’’ There are several reasons why the metabolites or drugs are not picked up in the analysis, including the abundance of the
ions and the number and abundance of fragment ions. See Data S1B for figures of all molecular ion networks of linked metabolites
and parent drugs, and Table S1 for the GNPS web links of all molecular ion networking analyses.

Enrichment analysis for drugs in PD screen

The results of our screen against 438 drugs allow for an aggregate analysis of MDM by the PD microbiome. We hypothesized that
members of the microbiome would be more likely to metabolize natural compounds or derivatives thereof due to a higher probability
of prior exposure. To test this hypothesis, we first annotated each of the MDM+ or MDM- drugs to one of three categories: naturally
occurring molecules (i.e., molecules directly derived from humans, plants, or microbes; e.g., hydrocortisone; N = 30), derivatives of
naturally occurring molecules (i.e., a semisynthetic derivative or a close structural mimic of a natural product, e.g., hydrocortisone
acetate; N = 90), and synthetic molecules (e.g., nicardipine; N = 318). By comparing the fraction of MDM+ drugs in the first two cat-
egories (natural + derivative, 26 out of 120, 21.6%) to that of the third category (synthetic, 31 out of 318, 10%), we revealed a signif-
icant difference (p < 0.001, two-tailed proportions z-test, n based on the number of molecules with and without the classification).
Intrigued, we decided to examine differences in MDM at lower levels of drug classification. We observed a significantly higher hit
rate among steroids (steroids: 16 out of 28, 57.1%; non-steroid: 41 out of 410, 10%, p < 0.001, two-tailed proportions z-test),
including hormonal steroids, corticosteroids, bile acids, and derivatives thereof. In fact, the high hit rate of the steroid class is the
major contributor to the observed difference between the hit rates of natural/derivative and synthetic groups, which is abolished
upon exclusion of the steroids (non-steroid natural/derivative: 10 out of 94, 10.6%; non-steroid synthetic: 31 out of 316, 9.8%) (Table
S1). The high hit rate among steroids is in-line with the idea that the microbiome is more likely to metabolize compounds it frequently
encounters, as steroids (e.g., bile acids) are normally present in the gut at high concentrations (Northfield and McColl, 1973). The fact
that 10% of fully synthetic molecules tested in our screen are metabolized by PD indicates the presence of a yet-unexplored range
of biochemical activities encoded by the gut microbiome that are capable of recognizing foreign substrates.

e7 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

Other than natural and synthetic classifications, we also looked more closely at functional groups that are enriched or depleted in
MDM+ drugs. We generated a list of 94 common functional groups and structural features and searched for them within all of the
drugs tested in our screen. To determine whether certain functional groups are enriched in MDM positive drugs, we aggregated
the SMARTS of common functional groups and the SMILES of all drugs within our screen. We then searched for these functional
groups within the drugs using the obgrep function within Open Babel (O’Boyle et al., 2011). We then tested for enrichment or deple-
tion of these groups within MDM+ drugs using two-sided proportion z-tests, correcting the resulting p values using the Benjamini-
Hochberg method and requiring that the false discovery rate (FDR) corrected p value is less than 0.01. The n for these tests is based
on the number of molecules with and without the functional group. Not surprisingly, we observed an enrichment of the following func-
tional groups in MDM+ drugs: nitro groups (FDR corrected p = 3e-16), ketones (FDR corrected p = 3e-8), carbonyl groups with one
carbon attachment (FDR corrected p = 8e-4), azo groups (FDR corrected p = 0.001), imines (FDR corrected p = 0.002), and alkenes
(FDR corrected p = 0.001). These results are consistent with common reduction and hydrolysis reactions often performed by gut bac-
teria. On the other hand, we observed a general depletion of arenes and nitrogen atoms in MDM+ drugs (FDR corrected p = 7e-5 and
FDR corrected p = 1e-7, respectively). However, when we excluded steroids and repeated the analysis we found that the depletions
in arenes and nitrogen atoms were no longer statistically significant (corrected p = 0.7 and corrected p = 0.7, respectively). This in-
dicates that the original statistically significant depletions were the result of steroids being a highly modified class that generally does
not contain these functional groups, rather than the functional groups themselves being important predictors for the lack of meta-
bolism (Figure 2D and Table S1). The exclusion of steroids, on the other hand, did not affect the observed enrichments we found
for nitro groups, imines, azo groups and ketones (FDR corrected p < 0.01). It is important to note that the results of our analysis
of MDM enrichment are based on a single subject’s microbiome, and should be repeated in the future with data from a much larger
set of donors.

Gene abundance analysis in metagenomic cohorts

The following datasets were used for the metagenomic analysis of the genes of interest in this study: HMP-1-1 (Human Microbiome
Project Consortium, 2012), HMP-1-2 (Lloyd-Price et al., 2017), MetaHIT (Nielsen et al., 2014), Chinese (Qin et al., 2012), and Fijicomp
(Brito et al., 2016). Raw sequencing data were obtained using the accession numbers of the associated manuscripts, and pre-pro-
cessed as previously described (Sugimoto et al., 2019). Quality-filtered reads were mapped to each gene using Bowtie2 (–end-to-
end,–fast,–score-min L,-0.6,-0.3) (Langmead and Salzberg, 2012), and gene abundance (in Reads per Kbp per Million reads, RPKM)
as well as gene breadth coverage (in percent of gene length) were calculated.
We only considered the gene as ‘‘present’’ if the reads cover greater than 50% of the gene length, otherwise the quantification
RPKM is considered zero. For datasets with multiple samples per subject, we aggregated the quantifications. If one or more samples
corresponding to a single subject met the coverage threshold, we present the average RPKM of these samples as the RPKM of the
subject. If no samples met the threshold, we consider the overall RPKM of the subject to be zero. See Table S4 for tabulated results of
this analysis.

ENDS (Expected Number of Detectable Strains)

A metric (‘‘ENDS’’) was developed to estimate the number of strains for which MDM reactions can be experimentally detected in
ex vivo cultures. ENDS answers the following question: if all of the ASVs in the ex vivo culture performed an MDM reaction at a given
rate r (in units of normalized metabolite signal per unit biomass per time), how many ASVs’ reactions will be detected in our screen
given the culture composition and measurement instrument? This framework is needed to incorporate the potentially confounding
impact of community biomass in the media selection process. For example, if biomass is not considered, a high-diversity low-density
community where bacterial load is too low to produce detectable metabolite levels would be favored over a lower diversity commu-
nity with high enough bacterial load to produce detectable metabolite levels. Formally, we are computing the expected value of the
number of the reactions detected:
X
n
E½Ns  = Bðxi Þ
i=1

Where E½Ns  is the expected number of detectable microbes, and Bðxi Þ is the probability of microbe i’s reaction being detected with
an absolute population of size xi :
How can we construct Bðxi Þ? In statistical terms, Bðxi Þ is equivalent to the power (the probability of deciding there is a reaction
when the reaction is actually present) of the hypothesis testing method used to analyze the data. In this case, we are using a one-
sided unequal variances t test with cutoff a. Now that we have a framework for calculating Bðxi Þ; we must relate a given xi to a
null and alternative distribution of measurements.
We assume that the metabolite measurements are composed of two types of signal: background noise, X and compound signal, Y.
Both signals are assumed to be normally distributed (we show in Methods S1F that an empirically estimated power function provides
similar results). The background noise X  Nðm1 ; s1 Þ is the signal present when no actual metabolite is present. The compound signal
Y  Nðm2 ; fðm2 ÞÞ is the portion of the signal due to measurement of an actual metabolite. The measurements in the control condition
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
are modeled by X while the measurements in the experimental conditions are modeled as Z = X + Y  Nðm1 + m2 ; s21 + fðm2 Þ2 Þ

Cell 181, 1661–1679.e1–e15, June 25, 2020 e8

 ll
Resource

We will now relate the population abundance xi to the mean of Y, m1 . In real terms, m1 is the average level of a metabolite produced
by xi . We assume the production metabolite is governed by the dynamics ðd½M =dtÞ = rxi where ½M is the concentration of the metab-
olite and r is the rate of metabolite production per cell. If we assume the drug is added at stationary phase such that ðdxi =dtÞ = 0 for all
t after drug addition, the total amount of metabolite produced is ½M = trxi where t is the incubation time. The r can vary widely, and to
account for this it can be set using the rate of a known MDM reaction (see Methods S1D).
We must now estimate the distribution of X. We estimate this by computing the mean and standard deviation of spurious peaks
detected when samples not containing the compound being measured are quantified.
Now we must define the standard deviation of Y, fðm2 Þ. This is clearly dependent on the instrument being used. By plotting the
standard deviations of triplicate measurements from our machine against their mean, we can estimate fðm2 Þ. To ensure we are
capturing only measurement signal, we based our model only on measurements that largely composed of measurement signal
(more than three standard deviations above the mean of the null distribution). We have found that a power law fðm2 Þ = amb2 fits the
data well. With the distributions of X and Z, we then estimate the Bðxi Þ using existing methods (Harrison and Brady, 2004).
What if we want to create an ensemble of multiple culture conditions to detect even more microbial reactions? To do this we can
define the expected number of new detectable microbes gained if we include another media E½DNs . For each ASV in all media, we
take the product of the probability that the reaction is not detected in the existing media and the probability that the reaction is
detected in the new media.
X
n Y
m
  
E½DNs  = Bðxi Þ 1  B yij
i=1 j=1

Where m is the number of media in the existing ensemble, and yij is the abundance of ASV i in existing media j.
In our actual computation of ENDS, we exclude samples with less than 10,000 reads and take the optimal medium as the one with
the highest ENDS averaged across all twenty donors.

High-throughput screen with D1-20

A new medium, BG, was formulated (see Methods S1-B). BG was made as follows: BB powder and mGAM powder were reconsti-
tuted in water and autoclaved as per manufacturer instructions (Sigma and Hyserv, respectively). Liquid BB medium was mixed with
liquid mGAM in a 70:30 ratio, respectively. For each donor (D1-D20), 500 mL of a donor’s glycerol stock was used to inoculate 50 mL
of pre-reduced BG medium and incubated for 24 h in an anaerobic chamber at 37
C. The culture was transferred to a sterile Nunc 96-
well plate (Fisher Scientific) with each well containing 400 mL of culture. 13.2 mL of each drug stock solution (1mM in DMSO), or 13.2 mL
of DMSO as a vehicle-only control, was pipetted and resuspended in 4 adjacent wells of the 96-well plate (quadruplicates) (see Table
S3). Selected drugs included MDM+ drugs from our original screen (N = 15), MDM- drugs randomly (N = 4) or rationally (N = 2)
selected, and drugs reported in the literature to be MDM+ but were deemed negative in our screen (N = 2). The plate was incubated
for 24 h anaerobically at 37
C. Before chemical extraction, 10 mL of the internal standard (voriconazole, 1mM) was pipetted into the
wells. For chemical extraction, 800 mL of ethyl acetate was pipetted into the wells and resuspended several times. 400 mL of the
organic layer was pipetted into an Agilent 96-well plate and dried under Nitrogen with a 96-well blow-down evaporator (Fisher Analyt-
ical). For HPLC-HRMS analysis, the plate was resuspended in 300 mL methanol and left for 10 min at room temperature prior to centri-
fugation at 3900 RPM for 10 min at 4
C. 60 mL was carefully decanted from the top into a new plate with 60 mL methanol and run on an
Agilent 6545 LC/QTOF machine (0.5 mL injection, only three of the four replicates for each drug were run on the machine). The remain-
ing 240 mL were dried and stored at 20
C for future runs. An abiotic BG-drug plate and a heat-killed-microbiome-drug plate were
prepared and analyzed using the same method, serving as controls to estimate non-enzymatic drug degradation or metabolite pro-
duction. To prepare the heat-killed-microbiome plate, 500 mL of D20 glycerol stock was used to inoculate 50 mL of pre-reduced BG
media and incubated for 24 h in an anaerobic chamber at 37
C. The culture was heat-killed at 100
C for 30 min while keeping the flask
containing it sealed (and therefore maintaining the anaerobic conditions). Drug incubation, chemical extraction, and HPLC-HRMS
analysis for the control plates were performed as previously described for the donor-drug plates.

Targeted quantitative metabolomics analysis

HPLC-HR-MS analysis was performed on an Agilent 6545 LC/QTOF machine. The Autosampler compartment was kept at 7
C, and
the column was kept at 25
C. Reverse phase chromatography was performed using an Agilent Eclipse Plus C18 RRHD column
1.8 mM (2.1 3 50 mm) column (Agilent, USA) with the gradient 95%A, 5%B to 5%A, 95%B in 12 min, then 95%B for 2 min, followed
by initial conditions (95%A, 5%B) for 3 min to re-equilibrate the column (A = 0.1% formic acid in water and B = 0.1% formic acid in
Acetonitrile). The flow rate was 0.4 mL/min. The samples in this study were run in one of two modes: a high resolution 4GHz mode and
a high dynamic range 2GHz mode. MS acquisition parameters for the 4GHz mode were set as follows: positive ion polarity, 0.5min
delay before MS measurement, 325
C gas temperature, 10 L/min drying gas flow rate, 20 psi nebulizer pressure, 325
C sheath gas
temperature, 12 L/min sheath gas flow rate, 4000 V capillary voltage, 500 V nozzle voltage, 135 V fragmentor voltage, 45 V skimmer
voltage, MS and MS/MS mass range of 100 - 1700 m/z, acquisition of 5 MS1 spectra/s, acquisition of 3 MS2 spectra/s, 20 eV collision
energy, a maximum of 2 precursors per cycle, and a precursor selection threshold of 200 counts absolute or 0.01% relative. The sys-
tem was run in auto MS/MS mode. For the 2GHz mode, the parameters were the same as the 4GHz mode with the following changes:

e9 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

acquisition of 8 MS1 spectra/s, acquisition of 6 MS2 spectra/s, maximum of 5 precursors per cycle, precursor selection threshold of
2000 counts.
To verify that the concentration of the internal standard (voriconazole) used in the screen was below the saturation limit of the ma-
chine, we constructed a standard curve of voriconazole. 12 mL of 1 mM voriconazole was added to 228 mL of methanol and serial
dilutions were performed by a factor of three to cover the concentration ranges of 40 mM to 0.165 mM. These samples were analyzed
using the 2GHz setting described above to match the setting used for drug quantification in the multi-donor screen.
Drugs and their detected metabolites were quantified in the MS1 of all samples using MassHunter Quantitative Analysis with the
Agile2 integrator. The metabolites quantified here were either ones that we previously discovered during the PD screen, ones that
were previously reported in the literature, or novel metabolites from the multi-donor screen identified using untargeted metabolomics
and verified by molecular networking (See below, Table S3, Figure S3). For quantification of dihydrodigoxin, we required a highly
sensitive integration method to distinguish between dihydrodigoxin and the isotopes of digoxin, since parent and metabolite eluted
at similar retention times. To do this, we performed integrations within MassHunter Qualitative, specifying a mass range of 805.43 -
805.44 m/z for dihydrodigoxin. We verified this method could differentiate the two compounds using authentic standards of dihydro-
digoxin and digoxin, showing that it could accurately quantify dihydrodigoxin and that it did not detect dihydrodigoxin when only
digoxin was present.
Following quantification, all further data processing was performed in MATLAB. For each plate, we remove any samples whose
internal standard AUC was greater than three interquartile ranges above the third quartile or below the first quartile. In order to correct
for differences in extraction efficiency, all peak areas in a given sample were divided by the corresponding internal standard area. This
ratio was then used for hypothesis testing and all other downstream analyses. For drug depletion, unadjusted p values were obtained
by one-sided Welch’s t tests testing whether drug levels are significantly lower in the donor-drug conditions than in controls; p values
were computed for tests against controls where the drug was incubated with BG medium (medium-drug) and incubated with the
heat-killed-microbiome (HKM-drug) controls. For all metabolite quantification statistical tests, n is the number of replicates passing
quality control (maximum 3). For metabolite production, unadjusted p values were obtained for one-sided Welch’s t tests testing
whether metabolite levels are significantly higher in donor-drug conditions than in control conditions; p values were computed for
tests against medium-drug, HKM-drug, as well as donor-DMSO (where the cultures are incubated with only the vehicle, DMSO) con-
trols. Correction for multiple hypotheses was performed using the Benjamini-Hochberg procedure (Benjamini and Hochberg, 1995).
Metabolite production and drug depletion p values were adjusted separately. For depletion to be considered significant, we required
FDR corrected p value < 0.01 for both the medium-drug and HKM-drug adjusted p values and also depletion was required to be
greater than 50% relative to both controls. For metabolite production to be considered significant, we required that FDR corrected
p value < 0.01 for the medium-drug, HKM-drug and donor-DMSO p values.
We then looked for correlation between the results of the targeted metabolomics and the compositions of the BG communities, We
restricted the drugs and metabolites tested by requiring that they must have at least one associated compound (parent drug or
metabolite) with an inter-individual variability entropy > 0.5, and that there exists at least one sample with more than 20% of the
drug remaining relative to medium-drug controls. We tested only taxonomic elements present in at least three samples with a
biomass of at least 1 mg/L in at least one sample, and corrected the resulting p values for multiple hypotheses at each taxonomic
level using the Benjamini-Hochberg method. The n for these tests is based on the number of observations used to compute the cor-
relations. Spearman correlation was computed for this analysis. The mean BG community composition for each donor was used for
the correlation analysis. We tested taxonomic elements at the ASV, species, genus, and family levels.

Untargeted metabolomics analysis

In order to extract all features in a given sample, we used the batch recursive feature extraction method for small molecules and pep-
tides within Profinder 8 (Agilent). We changed the following settings in the method from default: compound ion count threshold set to
two ions, alignment retention time tolerance set to 0.2 min, minimum MFE score set to 90, minimum file prevalence set to 2, expected
retention time set to ± 1 min, retention time contribution to matching score set to 90, expected MS1 mass variation set to 10 ppm,
expected retention time tolerance set to 0.2 min, absolute height threshold for EIC integration set to 2500 counts, and final absolute
height threshold set to 5000. We then analyzed the resulting feature abundances in MATLAB. We remove any sample whose internal
standard AUC is less than 106. We then perform hypothesis testing using similar statistical methods as for metabolite production in
the targeted metabolomics, except that we utilize the multiple hypothesis correction of Storey (Storey, 2002) in place of the Benja-
mini-Hochberg method and require a fold-change cut-off of two relative to all controls. We combine features if their retention times
and estimated molecular weights differ by less than 0.2 min and 0.01 Da, respectively. We then remove all ions already quantified in
our screen and all features that are statistically significant for more than one drug.
For the remaining statistically significant features, we use molecular ion networking to verify the metabolite’s relationship to the
parent drug based on their HR-MS/MS fragmentation pattern. In order to gather data with a large enough number of MS2 spectra
per parent ion we reran samples of interest on our QTOF HPLC-HRMS/MS instrument (Agilent) using the same column and condi-
tions, and the 4GHz settings listed above (instead of the 2GHz one used in the multi-donor screen). In order to minimize the number of
samples run a second time, we identified the minimal set of 81 samples that would allow us to detect and perform molecular ion
networking on all novel metabolites found in the original 1380 donor-drug samples. For molecular ion networking, we used the Global
Natural Product Social Molecular Networking (GNPS) server using the same settings used for the PD molecular networking. In order

Cell 181, 1661–1679.e1–e15, June 25, 2020 e10

 ll
Resource

to determine whether a metabolite is linked to its parent by the molecular ion networking, we first identify whether the drug and
metabolite are present in the network. For this, we require that the mass and retention time found in the molecular ion networking
differ by less than 0.2 min and 0.02 Da, respectively, from the properties reported by the initial donor-drug stage of the pipeline.
We call the two compounds related if they are in the same connected component of the graph. In the cases where either the metab-
olite or the parent drug or both were not picked up in the molecular ion networking analysis, we deem the linkage ‘‘undetermined.’’
There are several reasons why the metabolites or drugs are not picked up in the analysis, including the abundance of the ions and the
number and abundance of fragment ions.

Isolate screen for capecitabine

Three-mL overnight seed cultures of the following bacteria in their respective media were obtained as described above: Anaerostipes
caccae, Clostridium bolteae, Prevotella bivia, Parabacteroides distasonis, Serratia marcescens, Enterococcus faecalis TYG’11,
Anaerococcus prevotii, Escherichia coli TYG’1, Escherichia coli TYG’2, Lactobacillus gasseri, Salmonella enterica, and Escherichia
coli BL21. This panel was selected from three of the most abundant Phyla that normally inhabit the gut microbiome (Firmicutes, Bac-
teroidetes, and Proteobacteria), spans 10 bacterial genera (12 strains in total), and includes three strains isolated originally from PD
using standard techniques (Enterococcus faecalis TYG11, Escherichia coli TYG1, and Escherichia coli TYG2). 20 mL of these seed
cultures was inoculated into 3 mL of the same selected pre-reduced medium, and incubated at 37
C under the same anaerobic con-
ditions for an additional 24 h. After 24 h, 10 mL of the 10 mM drug solution in DMSO, or of a DMSO control was added to the growing
microbial culture and incubated for another 24 h. In addition, 10 mL of the drug was incubated for 24 h under the same conditions in a
no-bacterium, medium-only control. After incubation, cultures were extracted with ethyl acetate and the organic phase was dried
under vacuum in a rotary evaporator. Extracts were suspended in 250 mL of MeOH and analyzed using HPLC-MS as described above
in the PD ex vivo screen. See Figure S4.

Metagenomic library construction

Metagenomic DNA was extracted from approximately 0.25 g of PD stool (stored at 80
C in RNAlater (Thermo Fisher Scientific,
Massachusetts, USA)) using the PowerSoil DNA Isolation kit (MO BIO Laboratories California, now QIAGEN, USA) according to man-
ufacturer’s instructions. DNA was quantified using a NanoDrop 2000 (Thermo Fisher Scientific, Massachusetts, USA).
The vector pGFPuv (Clonetch Laboratories) was used as the parent vector for the metagenomic library. The pGFPuv plasmid was
linearized by using PCR and Phusion High-Fidelity DNA Polymerase (New England Biolabs, Massachusetts, USA) and the 2.5 Kbp
product was excised and extracted from the gel after gel electrophoresis and using the QIAquick Gel Extraction kit according to man-
ufacturer’s instructions. PCR primers for linearization were: forward, pGFP-IF-F = TAATGAATTCCAACTGA GCGCCGG and reverse,
pGFP-IF-R = CATAGCTGTTTCCTGT GTGAAATTG. DNA was quantified using a NanoDrop 2000 and incubated with Dpn1 (New En-
gland BioLabs, Massachusetts, USA) at 37
C for one h followed by incubation at 80
C for 20 min to heat inactivate the Dpn1 enzyme.
To prevent re-ligation of the vector, Antarctic Phosphatase (New England Biolabs, Massachusetts, USA) was added to the Dpn1
treated mixture and incubated at 37
C for 15 min and 70
C for five minutes to heat inactivate the phosphatase. The product was
then purified using the QIAquick PCR Purification kit (QIAGEN, Maryland, USA) according to manufacturer’s instructions and quan-
tified with a NanoDrop 2000.
The extracted DNA sample was brought to 150 mL in Milli-Q water. DNA was sheared via a g-TUBE (Covaris, Massachusetts,
USA) according to manufacturer’s instructions using an accuSpin Micro 17R Microcentrifuge (Thermo Fisher Scientific, Massachu-
setts, USA). DNA fragment size was validated using the Agilent 2100 Bioanalyzer (Agilent Technologies, California, USA).
Sheared DNA was subjected to gel electrophoresis, and the 2-4 Kbp product was excised from the gel. Gel extraction was
performed using Zymoclean Large Fragment DNA Recovery Kit (Zymo Research, California, USA) according to the manufacturer’s
instructions.
Sheared DNA was end-repaired using the End-It DNA End-Repair Kit (Epicenter (Illumina), California, USA) according to manufac-
turer’s instructions. The end-repaired DNA was then purified by isopropanol precipitation. Blunt-end ligation of DNA with linearized
pGFPuv was performed with T4 DNA ligase (New England Biolabs, Massachusetts, USA) according to manufacturer’s instructions at
16
C overnight using 100 ng vector and 360 ng insert to achieve a 1:3 molar ratio. Ligated plasmids were purified by precipitation with
Pellet Paint NF Co-Precipitant (EMD Millipore, Massachusetts, USA) and brought to 4 mL in Milli-Q water. 2 mL ligated plasmid were
transformed into MegaX DH10b Electrocompetent Cells (Thermo Fisher Scientific, Massachusetts, USA) according to manufac-
turer’s instructions, and recovered in 1 mL of LB medium.
The 2 3 1 mL transformation products from each ligation were combined and brought up to 20.2 mL in LB/ampicillin and divided
into 20 3 1 mL pools. Each 1 mL pool was cultured in a shaking incubator at 30
C overnight and mixed with 1 mL 50% glycerol the
following day, gently vortexed, and stored at 80
C for future screening. Serial dilutions were made from the remaining 200 mL of the
transformation product and spread on LB/ampicillin plates which were incubated overnight at 37
C, to calculate the library titer of
unique clones the following day. Random clones were picked, checked for the presence of insert, and used to estimate insert
size by PCR (most clones harbored inserts in the 2-4 Kbp range). This procedure was repeated 4 times, with a yield of 2-6 X 104
unique clones per pool (80 total pools) and a total of 3 X 106 unique clones.

e11 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

Functional screening of the metagenomic library

2 mL from each of the initial 80 pools containing 2-6 X 104 unique clones were added to 3 mL of LB-carbenicillin (LB-carb, 100 mg/mL)
in glass culture tubes and grown at 37
C for 1 h. After 1 h, 10 mL of 10 mM hydrocortisone (in DMSO) was added to each culture and
incubated at 37
C for 20 h.
After a total of 20 h of growth post hydrocortisone addition, cultures were chemically extracted as follows: 5 mL of 10 mM vorico-
nazole was added to the cultures, followed by the addition of 7 mL of ethyl acetate solvent using a glass pipette. The resulting solution
was vortexed on a medium-high setting twice, allowing for the mixture to separate between vortexing steps. After the last vortexing
step, the mixture was left undisturbed for 5 min. 5 mL of the top organic layer of solvent was transferred to a 4 mL glass tube using a
chemically resistant 1 mL tip and pipette. The samples were then dried using a Labconco CentriVap Concentrator for 1.5 h. Dried
samples are then resuspended in 250 mL of methanol and sonicated for 5 min. Samples are then transferred to a 1.5 mL Eppendorf
tube and centrifuged. 180 mL of the clear solution was then deposited in glass LCMS vial and analyzed using HPLC-HRMS on an
Agilent 6545 LC/QTOF instrument using the same column and gradient as described for the D1-D20 plates, and the 4GHz MS set-
tings. An identical quantitative workflow was employed to identify the pools with the highest signal for the hydrocortisone metabolite/
internal standard ratio.
Selected pools were then plated for colony forming units (CFU) counting. CFU counting was done as follows: 5 mL of each glycerol
stock was taken and diluted into 95 mL of LB-carb by pipetting up and down. 5 mL was taken out from the first tube and moved to a
second tube containing also 95 mL of LB-carb, with pipetting to ensure proper mixing. Serial dilutions were done four more times for a
total of five dilutions. 50 mL of each serial dilution was plated onto LB-carb agar plates and spread evenly across the entire surface.
Plates were incubated at 37
C overnight and checked for colonies the next day. The number of colonies on a plate were counted and
used to calculate CFUs for the original positive sample glycerol stock.
The positive glycerol stock sample was then diluted to obtain sub-pools of 2,000 clones (based on the CFU counts) per sub-pool
with a total of 20 sub-pools to be screened for each positive sample. These 20 sub-pools were then tested for hydrocortisone meta-
bolism and glycerol stocked following the same protocol outlined above for the 20,000 level sub-pools, except that hydrocortisone
was added after 4 h of culture instead of 1 h and that a glycerol stock was made from each sub-pool 12 h after the culture was initi-
ated. Glycerol stock from the top 2,000-clone sub-pool that produced the hydrocortisone metabolite was then subjected to CFU
counting, further sub-pooled and tested at the 200-clone level. The top 200-clone sub-pool was treated the same way and tested
at the 20-clone level.
Glycerol stock from the sub-pool positive for hydrocortisone metabolite at the 20-clone level was then serially diluted and plated
onto five LB-carb plates and incubated overnight at 37
C. The next day, the plates were retrieved and single colonies were picked
and grown in 4 X 96 deep-well plates with 500 mL LB-carb per well. The deep-well plates were then grown at 37
C at 220 RPM shaking
overnight. The next day, 100 mL of culture was saved from each well into another 96-well plate for glycerol stock using a multichannel
pipette.
10 mL from each well of each row were pooled before inoculating a new glass culture tube containing 3 mL LB-carb. The resulting
32 culture tubes corresponding to the 32 pooled rows were grown at 37
C for 4 h before the addition of 10 mL of 10 mM hydrocor-
tisone, after which samples were grown at 37
C for 20 h.
After 20 h, 5 mL of 10 mM voriconazole was added and chemical extraction with ethyl acetate was performed as described pre-
viously and analyzed using HPLC-HRMS. The positive hit indicated which row the hydrocortisone metabolite originated from. 5 mL of
each well in the positive row was used to inoculate a new glass culture tube with 3 mL LB-carb, which was then grown for 4 h at 37
C,
after which hydrocortisone was added and the metabolism screen proceeded as previously described. A single well from the positive
row was then identified as a positive hit. See Figure S6.
The glycerol stock of the positive well was then used to inoculate a 5 mL LB-carb culture that was grown overnight at 37
C and
shaken at 220 RPM. The next day, the plasmid containing metagenomic DNA was isolated from the culture using a QIAprep 
Spin Miniprep Kit (QIAGEN, USA) according to manufacturer’s instructions and subjected to DNA sequencing (Sanger) to identify
the metagenomic insert on the plasmid. End-sequences were compared to the assembled PD metagenome using BLASTn.

Heterologous expression of PD-derived 20b-HSDH

The 20b-HSDH gene identified from the functional metagenomic screening of PD metagenomic DNA was codon optimized for
expression in E. coli and ordered as a gBlock from IDT (USA). Primers annealing to the gBlock insert were designed to contain over-
hang restriction enzyme cut sites that would recognize restriction digested pet28a vector. The forward primer contains the cut site for
NdeI, and the reverse primer contains the cut site for NotI. Primer sequences are as follows (50 to 30 ):
Forward: CTGGTGCCGCGCGGCAGCCATATGGCAGATGAATCATCGAAGATTCC
Reverse: GGTGGTGCTCGAGTGCGGCCTTAGAAAACTGAATACCCACCGTCC
PCR product was gel extracted and cloned into a double-digested pet28a vector using the InFusion cloning kit (Takara). InFusion
product was transformed into chemically competent BL21-DE3 E. coli cells. Plasmid was recovered using a QIAprep  Spin Miniprep
Kit (QIAGEN, USA) and sequenced (Sanger) to confirm the presence of the correct insert.
To clone the native-sequence of the 20b-HSDH gene identified from the functional metagenomic screening of PD metagenomic
DNA, primers designed with restriction cut sites (as described above) were used to clone it from the metagenomic clone into a dou-
ble-digested pet28a vector. Primer sequences are as follows (50 to 30 ):

Cell 181, 1661–1679.e1–e15, June 25, 2020 e12

 ll
Resource

Forward: CTGGTGCCGCGCGGCAGCCATATGGCAGACGAATCATCGAAGATTC
Reverse: GGTGGTGCTCGAGTGCGGCCCAAATGGGGTACGGTGATTAGAAGAC
Plasmid was recovered using a QIAprep  Spin Miniprep Kit (QIAGEN, USA) and sequenced (Sanger) to confirm the presence of
the correct insert.
Cultures of BL21-DE3 E. coli harboring either the codon optimized or native sequence of the 20b-HSDH gene were grown from
glycerol stock overnight at 37
C, 220 RPM. The next day, cultures were back diluted to an OD600 of 0.05, grown at 37
C,
220RPM until OD600 was 0.4 and induced with a final concentration of 1mM IPTG. After four h of growth (from the time of back dilu-
tion), 10 mL of 10 mM hydrocortisone was added. All cultures were then grown for 20 h at 37
C, 220 RPM, chemical extraction and
HPLC-HRMS analysis were performed as described above.

Metagenomic and metatranscriptomic analyses

PD metagenomic sequencing: metagenomic DNA of PD, prepared as described above, was sheared to a mean size of 500 bps
using a Covaris E220 sonicator (Covaris, USA). An Illumina sequencing library was prepared from the sheared DNA using the auto-
mated Apollo 324TM NGS Library Prep System and the PrepX DNA library kit (WaferGen, USA) according to the manufacturer’s
protocol. This step included DNA end repairing, A-tailing, adaptor ligation, and limited PCR amplification. After examination on Bio-
analyzer (Agilent, CA) DNA HS chips for size distribution, and quantification by Qubit fluorometer (Invitrogen, CA), the library was
sequenced on Illumina HiSeq 2500 Rapid Flowcell as paired-end reads, along with 8 bps Index reads, following the manufacturer’s
protocol (Illumina, USA). Raw sequencing reads were filtered by Illumina HiSeq Control Software to generate Pass-Filter reads for
further analysis.
PD metatranscriptomic sequencing: Total nucleic acid (DNA and RNA) was prepared from the RNAlater-preserved PD stool sam-
ple using the AllPrep PowerViral DNA/RNA Kit (QIAGEN, USA). DNA was digested and removed using Turbo DNase (ThermoFisher
Scientific, USA) following the manufacturer’s instructions, and remaining DNA-free RNA concentration and quality were measured
using a Nano Drop 2000 and a Bioanalyzer (Agilent, USA). This RNA was further subjected to ribosomal RNA depletion using the
Ribo-Zero Gold rRNA Removal Kit (Epidemiology) (Illumina, USA), and a strand-specific Illumina RNA-seq library was prepared
from it as described for the PD metagenomic DNA sample.
PD-CL-100 metagenomic library sequencing: plasmids were isolated using the QIAprep Spin Miniprep Kit (QIAGEN, USA) from two
sub-pools containing a total 100,000 Unique Clones. An Illumina sequencing library was prepared and sequenced from the resulting
DNA as described above, except that it was sequenced on a NovaSeq 6000 instead of the HiSeq 2500.
PD metagenomic sequencing yielded 35,527,955 reads (2 X 175 bps), PD metatranscriptomic sequencing yielded 30,796,174
reads (2 X 150 bps), and PD-CL-100 yielded 69,173,413 reads (2 X151 bps).
Raw Illumina reads were filtered using PRINSEQ, according to the following parameters: minimum average quality score of 30,
maximum percentage of undetermined reads of 2%. Trimming on each end was implemented until a minimum average quality score
of 30 is reached. Trimmed reads that are shorter than half of the original read length were discarded (Schmieder and Edwards, 2011).
SPAdes was used to assemble the resulting pairs and singletons of filtered reads, using default parameters (Bankevich et al., 2012).
For PD-CL-100, Bowtie2 was used to identify and remove reads mapping to the pGFP-UV plasmid backbone using the following
settings:–end-to-end–sensitive. Reads that do not map to the backbone were then aligned to the SPAdes scaffolds produced from
the PD metagenomic dataset, also using Bowtie2 (–very-sensitive-local) (Langmead and Salzberg, 2012). Reads from the PD meta-
genomic dataset itself were also aligned to their corresponding SPAdes-produced scaffolds using the same settings. RPKM values
and coverage breadths for each PD scaffold equal or longer than 2 Kbp were calculated on the basis of the Bowtie2 alignment results
for each the two read datasets (PD metagenome and PD-CL-100). A PD scaffold was considered ‘‘present’’ in PD-CL-100 if meta-
genomic reads from PD-CL-100 covered at least 50% of the scaffold’s length, otherwise the quantification RPKM is considered zero.
To infer the taxonomy of the PD scaffolds, we ran kraken-1.1.1 with standard settings and the ‘‘MiniKraken DB_8GB’’ pre-built data-
base constructed from complete bacterial, archaeal, and viral genomes in RefSeq (as of Oct. 18, 2017) (Wood and Salzberg, 2014).
See Table S4.
Metratranscriptomic analysis: we first used BLASTn to map a database of quality filtered PD metatranscriptomic reads to a query
of the PD scaffold harboring the 20b-HSDH gene, while specifying an e-value cutoff of 1e-30. Next, we matched these BLASTn-iden-
tified reads to either the 20b-HSDH gene or neighboring genes that we annotated in 5 Kbp windows upstream and downstream of it.
For this step, we used the Geneious assembler in Geneious (Kearse et al., 2012) with the following parameters: minimum overlap:
50 bp, minimum percent identity at overlap: 90%, and maximum percentage of mismatch per read: 20%. Matched reads per
gene were counted and used to construct the bar graph in Figure 6E.

TP and UP gene deletions in E. coli BW25113

E. coli BW25113 mutants that harbor a replacement of deoA or udp with a kanamycin resistance gene were obtained from the Keio
collection (Baba et al., 2006). Since the kanamycin resistance gene is flanked by FLP recognition target sites, we decided to excise it
and obtain in-frame deletion mutants. Plasmid pCP20, encoding the FLP recombinase, was transformed to each of the mutants by
electroporation, and transformants were selected on Ampicillin at 30
C for 16 h. 10 transformants from each mutant were then picked
in 10 mL LB medium with no selection, and incubated at 42
C for 8 h to cure them from the temperature-sensitive pCP20 plasmid.
Each growing colony was then streaked on three plates (LB-ampicillin, LB-kanamycin, and LB with no selection). Mutants that could

e13 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

only grow on LB, but not on LB-ampicillin (confirming the loss of the pCP20 plasmid), nor on LB-kanamycin (confirming the excision of
the kanamycin resistance gene) were confirmed to harbor the correct deletion using PCR and DNA sequencing. Primers deoA-
Check-F: 50 -CGCATCCGGCAAAAGCCGCCTCATACTCTTTTCCTCGGGAGGTTACCTTG-30 , deoA-Check-R: 50 - CAAATTTAAA
TGATCAGATCAGTATACCGTTATTCGCTGATACGGCGATA-3 0 , udp-Check-F: 50 -CGCGTCGGCCTTCAGACAGGAGAAGAGAA
TTACAGCAGACGACGCGCCGC-30 , and udp-Check-R: 50 -TGTCTTTTTGCTTCTTCTGACTAAACCGATTCACAGAGGAGTTGT
ATATG-30 were used in PCR experiments to confirm the deletion of the deoA or upd genes and the kanamycin resistance gene re-
placing them (Baba et al., 2006). To construct the DdeoA/Dudp double knockout, the in-frame Dudp knockout obtained above was
used as a starting point. Plasmid pKD46 expressing the Lambda Red recombinase was transformed to it using electroporation (Dat-
senko and Wanner, 2000) and transformants were selected on LB-Ampicillin at 30
C for 16 h. One Ampicillin-resistant transformant
was then cultured at 30
C in 50 mL of LB-Ampicillin, with an added 50 mL of 1 M L-arabinose to induce the expression of the recom-
binase. At an optical density of 0.4-0.6, electrocompetent cells were prepared from the growing culture by serial washes in ice cold
10% glycerol, and 300 ng of a linear PCR product were transformed to it by electroporation. This PCR product was prepared by
using the deoA-Check-F and deoA-Check-R primers on a template DNA prepared from the deoA mutant of the Keio library, in which a
kanamycin resistance gene replaces deoA. After electroporation, transformants were selected on LB-kanamycin at 37
C to induce
the loss of the temperature sensitive pKD46 plasmid, cultured in LB-kanamycin overnight at 37
C, and checked by PCR to confirm
the correct recombination position. Finally, the kanamycin resistance gene was excised from the deoA locus by the FLP recombinase
using the same strategy explained above, resulting in the final DdeoA/Dudp mutant

MDM-Screen of capecitabine using E. coli mutants

Wild type E. coli BW25113, and corresponding TP knockout (DdeoA), UP knockout (Dudp), and TP/UP double knockout (DdeoA/
Dudp) strains were cultured overnight in LB medium (aerobically, shaking at 37
C, 50 mL each). Triplicates of 3 mL for each strain
were incubated with 10 mL of 10 mM capecitabine (in DMSO) for an additional 24 h in an anaerobic chamber along with bacteria-
only and media-only controls. Cultures were then extracted and analyzed as previously described in the PD screen, except for
the addition of 20 mL of 0.25 mg/mL of an internal standard (voriconazole) prior to the extraction.

MDM-Screen of other FPs using E. coli mutants

Wild type E. coli BW25113, and corresponding TP knockout (DdeoA), UP knockout (Dudp), and TP/UP double knockout (DdeoA/
Dudp) strains were cultured overnight in LB medium (aerobically, shaking at 37
C, 50 mL each). Aliquots (100 mL) of each strain
were used to inoculate 3 mL of M9 medium, which were grown again overnight (aerobically, shaking at 37
C). 10 mL of 10 mM doxi-
fluridine (in DMSO) or trifluridine (in methanol) were incubated with each culture for an additional 24 h in an anaerobic chamber, along
with bacteria-only and medium-only controls. Cultures were spun down and collected supernatants were lyophilized. The dried res-
idues were then resuspended in 500 mL methanol and analyzed by HPLC-MS (Agilent Single Quad; column: Poroshell 120 EC-C18
2.7mm 4.6 3 100mm; flow rate: 0.6 mL/min; solvent A: 0.1% formic acid in water: solvent B: 0.1% formic acid in acetonitrile) and the
following gradient: 1 min, 0.5% B; 1-20 min, 0.5%–35% B; 25-30 min, 35%–100% B; 30-35 min, 100% B. The structures of all re-
sulting metabolites were confirmed by comparison to authentic standards. See Figure S4 and S5.

Microbiome-dependent pharmacokinetic experiment

Twelve C57BL/6 mice were treated with a commonly used cocktail of antibiotics (1 g/L of ampicillin, neomycin, metronidazole and 0.5
g/L vancomycin) in drinking water for 14 days (Planer et al., 2016). The antibiotic solution was supplemented with 5 g/L aspartame to
make it more palatable (Karmarkar and Rock, 2013). During these two weeks, the gut microbiome composition was monitored by
collecting feces from each mouse and performing molecular and microbiological analyses to make sure the microbiome is being
cleared by the antibiotic treatment. On day 15, no antibiotics were administered for 24 h (a washout period). On day 16, mice
were separated into the two groups, 6 per group (3 males and 3 females). In group 1, mice remained non-colonized. In group 2,
mice were administered 200 mL of freshly thawed PD glycerol stock using oral gavage. On day 17, the oral gavage was repeated
the same way to ensure the colonization of the administered bacteria (fecal samples were collected on days 16 and 17 and cultured
anaerobically to ensure colonization). On day 18, the pharmacokinetic experiment was performed by monitoring the fate of capeci-
tabine in mouse blood and feces over time. A capecitabine dose equivalent to a single human dose and adjusted to the weight of the
mice was administered by oral gavage (755 mg / kg, as a solution in 50 mL DMSO), then serial sampling of tail vein blood (by tail snip-
ping), as well as fecal collection were performed at these time points (zero, 20 min, 40 min, 60 min, 2 h, and 4 h post dosage). Blood for
each time point (30 mL) was collected using a 30 mL capillary tube and bulb dispenser (Drummond Microcaps, Drummond Scientific),
quickly dispensed in 60 mL EDTA to prevent blood coagulation, and stored on ice for up to 4 h and then frozen at 80
C until further
analysis. Feces were also collected at the same time points (even though defecation was left at will, we succeeded in collecting feces
for most time points), stored on ice for up to 4 h and then frozen in 80
C until further analysis. After the 4 h pharmacokinetic time
point, mice were euthanized.
For chemical extraction, 2 mL of an internal standard solution (0.5 mg / mL of voriconazole) were added to the blood / EDTA solution
mentioned above, and the sample was mixed using a vortex mixer. Next, 500 mL of ethyl acetate was added and mixed. The sample
was then centrifuged briefly at 15000 rpm, and the organic layer was transferred to a glass tube and evaporated under vacuum using
rotary evaporation (Speed Vac). The dried residue was dissolved in 100 mL of MeOH, and the solution was centrifuged at 15000 rpm

Cell 181, 1661–1679.e1–e15, June 25, 2020 e14

 ll
Resource

and transferred to an autosampler vial for HPLC-HR-MS analysis. For fecal samples, pellets were weighed (for later normalizations),
and suspended in 500 mL sterile Milli-Q water (Millipore Corporation, USA). 2 mL of an internal standard solution (0.5 mg / mL of vor-
iconazole) were added to the sample, and the mixtures were extracted with 500 mL 1:1 ethyl acetate: MeOH. Fecal debris were then
spun down and collected supernatants were dried under vacuum using a rotary evaporator (Speed Vac). The dried residues were
suspended in 100 mL MeOH. The final solutions were centrifuged at 15000 rpm and transferred to autosampler vials.
The prepared samples were analyzed by HPLC-HR-MS (Agilent QTOF). Chromatography separation was carried out on a Poros-
hell 120 EC-C18 2.7 mm 2.1 3 100 mm column (Agilent, USA) with the gradient: 99.5% A, 0.5% B to 100% B in 20 min and a flow rate
of 0.25 mL/min, where A = 0.1% formic acid in water and B = 0.1% formic acid in acetonitrile. A 10 mL aliquot of the reconstituted
extract was injected into the HR-HPLC-MS system, and the Area Under the Curve (AUC) was integrated for each metabolite and
normalized by the internal standard’s AUC. Peak identities were confirmed by accurate mass, and by comparison of chromato-
graphic retention time and MS/MS spectra to those of authentic standards.

QUANTIFICATION AND STATISTICAL ANALYSIS

All statistical analyses were performed in MATLAB. p values less than 0.01 (after correction for multiple hypotheses, if applicable)
were considered significant. For comparisons of the means of two populations, Welch’s t test was generally used. In cases where
the independence assumption of this test was not met (as determined by the form of the null hypothesis), permutation tests were
used instead. Comparison of multiple means was done via ANOVA. Comparisons of two proportions was done via a proportions
z-test. For all analyses, the meaning and value of n and the measures of center, dispersion, and precision used can be found in
the relevant main text or in Method Details.

e15 Cell 181, 1661–1679.e1–e15, June 25, 2020

 ll
Resource

Supplemental Figures

Figure S1. Expanded Bacterial Community Composition Analysis for PD, D1-20 and Their Ex Vivo Cultures, Related to Figures 1 and 3
A) Four-day time course of PD cultured ex vivo in various media. Family level bacterial composition of the original PD fecal sample (far left), as well as that of PD
ex vivo cultures grown anaerobically in 14 different media over four days (.01, .02, .03, .04). 16S rRNA gene amplicon sequences that could not be classified at the
family level, and families with less than 1% relative abundance in all samples are grouped into ‘‘Other.’’ Cultures are ordered according to their Jensen-Shannon
divergence (DJS) from the original PD sample (upper axes, computed at the family level), where lower values indicate higher similarity to PD. Note that cultures
grown in mGAM are the most similar to PD. B) Family level bacterial composition of the original D1-20 fecal samples and corresponding BG cultures. ‘‘Other’’
includes 16S rRNA gene amplicon sequences not classified at the family level and taxa that are below 5% abundance in all sequenced samples. Replicate BG
cultures for each donor labeled as ‘‘BG_1,’’ ‘‘BG_2,’’ and ‘‘BG_3.’’ See also STAR Methods, Table S2.
 ll
Resource

Figure S2. Sequential Metabolism of Hydrocortisone Acetate by the PD Microbiome, Related to Figure 2
HPLC-MS analysis of hydrocortisone acetate (1) and hydrocortisone (2) incubated with PD mGAM.02 culture, mGAM.02 broth, P. distasonis culture, or C. bolteae
culture. An HPLC chromatogram at an absorbance of 254 nm is shown for all samples, indicating the conversion of hydrocortisone acetate to hydrocortisone by
both P. distasonis and C. bolteae, and the conversion of both hydrocortisone acetate (in two steps) and hydrocortisone (in one step) to 20b-dihydrocortisone (3) in
the presence of the PD microbiome.
 ll
Resource

Figure S3. Structures of All Drugs Quantified in the D1-20 MDM-Screen and Their Known or Predicted Metabolites, Related to Figure 4
 ll
Resource

Figure S4. MDM Deglycosylation of the Anticancer Drugs Capecitabine and Trifluridine, Related to Figure 5
A) A heatmap indicating the ability of each of 12 tested bacterial strains to perform capecitabine deglycosylation (d-G) in order to identify candidate species for the
homology-based gene finding approach. E. faecalis TYG11, E. coli TYG1, and E. coli TYG2 were originally isolated from PD. B. HPLC-MS analysis of trifluridine (1)
incubated with wild type E. coli BW25113 (WT), and Dudp, DdeoA, and DdeoA/Dudp mutants in M9 medium. An HPLC chromatogram at an absorbance of
250 nm is shown for all samples, indicating the conversion of trifluridine (1) to trifluorothymine (2) by wild type E. coli BW25113 (WT), Dudp, and DdeoA, but not the
DdeoA/Dudp mutant.
 ll
Resource

Figure S5. MDM Deglycosylation of the Anticancer Prodrug Doxifluridine, Related to Figure 5
HPLC-MS analysis of doxifluridine (1) incubated with wild type E. coli BW25113 (WT), and Dudp, DdeoA, and DdeoA/Dudp mutants in M9 medium. Extracted Ion
Chromatograms for both doxifluridine (1) and its resulting MDM metabolite 5-fluorouracil (2) are shown for all samples, indicating the complete conversion of
doxifluridine (1) to 5-fluorouracil (2) by wild type E. coli BW25113 (WT), Dudp, and DdeoA, but not the DdeoA/Dudp mutant.
 ll
Resource

Figure S6. Functional Metagenomic Screening for Hydrocortisone Metabolizing Enzymes in the PD Microbiome, Related to Figure 6
Schematic representation of the screening scheme followed to identify the 20b-HSDH gene from the PD microbiome metagenomic clone library.
 ll
Resource

Figure S7. Microbiome-Dependent Pharmacokinetics of Capecitabine, Related to Figure 7

HPLC-HR-MS based quantification of capecitabine and its human metabolite (50 -deoxy-5-fluorocytidine) in fecal and blood samples from mice colonized with PD
in comparison to non-colonized ones. No significant difference in the levels of capecitabine were observed between the two groups. (p > 0.05, significance was
determined by testing the intersection null hypothesis with marginal two-tailed t tests using the Bonferroni correction to control family-wise error rate). Error bars
represent the standard error of the mean.
 Resource

Metabolic Dynamics and Prediction of Gestational

Age and Time to Delivery in Pregnant Women
Graphical Abstract Authors
Liang Liang,
Marie-Louise Hee Rasmussen,
Brian Piening, ..., Hanyah Zackriah,
Michael Snyder, Mads Melbye

Correspondence
mpsnyder@stanford.edu (M.S.),
mmelbye@stanford.edu (M.M.)

In Brief
Identification of blood metabolites in
pregnant women that can accurately
predict gestational age and provide
insights into pregnancy variations
undetected by ultrasound.

Highlights
d Weekly metabolome of maternal blood changes dynamically
through healthy pregnancy

d A metabolic clock of five blood metabolites accurately

predicts gestational age

d Two to three metabolites identify labor onset within two, four,

and eight weeks

d Women with metabolic clocks that outpaced ultrasound

evaluation tend to deliver earlier

Liang et al., 2020, Cell 181, 1680–1692

June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.cell.2020.05.002 ll
 ll

Resource
Metabolic Dynamics and Prediction of Gestational
Age and Time to Delivery in Pregnant Women
Liang Liang,1 Marie-Louise Hee Rasmussen,2 Brian Piening,1,8 Xiaotao Shen,1 Songjie Chen,1 Hannes Röst,1,9
John K. Snyder,3 Robert Tibshirani,4 Line Skotte,2 Norman CY. Lee,3 Kévin Contrepois,1 Bjarke Feenstra,2
Hanyah Zackriah,5 Michael Snyder,1,10,* and Mads Melbye2,6,7,*
1Department of Genetics, Stanford University School of Medicine, Stanford, CA 94305, USA
2Department of Epidemiology Research, Statens Serum Institut, Copenhagen, 2300, Denmark
3Department of Chemistry and the Chemical Instrumentation Center, Boston University, Boston, Massachusetts 02215, USA
4Department of Statistics and Biomedical Data Science, Stanford University, Stanford, CA 94305, USA
5Department of Molecular and Cell Biology, University of California, Berkeley, Berkeley, CA 94720, USA
6Department of Clinical Medicine, University of Copenhagen, Copenhagen, 2200, Denmark
7Department of Medicine, Stanford University School of Medicine, Stanford, CA 94305, USA
8Present address: Earle A. Chiles Research Institute, Providence Portland Medical Center, Portland, OR, 97213 USA
9Present address: Donnelly Centre for Cellular and Biomolecular Research, University of Toronto, Toronto, ON M5S 3E1, Canada
10Lead Contact

*Correspondence: mpsnyder@stanford.edu (M.S.), mmelbye@stanford.edu (M.M.)

https://doi.org/10.1016/j.cell.2020.05.002

SUMMARY

Metabolism during pregnancy is a dynamic and precisely programmed process, the failure of which can bring
devastating consequences to the mother and fetus. To define a high-resolution temporal profile of metabo-
lites during healthy pregnancy, we analyzed the untargeted metabolome of 784 weekly blood samples from
30 pregnant women. Broad changes and a highly choreographed profile were revealed: 4,995 metabolic fea-
tures (of 9,651 total), 460 annotated compounds (of 687 total), and 34 human metabolic pathways (of 48 total)
were significantly changed during pregnancy. Using linear models, we built a metabolic clock with five me-
tabolites that time gestational age in high accordance with ultrasound (R = 0.92). Furthermore, two to three
metabolites can identify when labor occurs (time to delivery within two, four, and eight weeks, AUROC R
0.85). Our study represents a weekly characterization of the human pregnancy metabolome, providing a
high-resolution landscape for understanding pregnancy with potential clinical utilities.

INTRODUCTION (Committee on Obstetric Practice, the American Institute of Ul-

Pregnancy is one of the most critical periods for mother and child
(Alkema et al., 2016; Say et al., 2014). It involves a tremendous
flow of physiological changes and metabolic adaptations week
by week, and even small deviations from the norm might have
detrimental consequences at different pregnancy stages. For
example, approximately 20% of all pregnancies end in miscar-
riage (< 20 weeks), and around 10% end in preterm birth (<
37 weeks) (Blencowe et al., 2013; Wang et al., 2003). The latter
is the leading cause of global neonatal morbidity and mortality
(Blencowe et al., 2013). Of 200 million annual pregnancies,
300,000 pregnancy- and birth-related maternal deaths and 7
million perinatal deaths occur worldwide (GBD 2013 Mortality
and Causes of Death Collaborators, 2015; Sedgh et al., 2014).
With a better understanding of how pregnancy is regulated,
even small improvements in obstetric health care can enhance
the well-being of many women and children.
An accurate estimation of the timing of pregnancy and birth is
important for many clinical decisions in obstetrics, including
determination of preterm birth and related treatment regimens
trasound in Medicine, and the Society for Maternal-Fetal Medi-
cine, 2017). The current clinical method of determining the
gestational age and due date is based on information about
the last menstruation date, which can be imprecise, or ultra-
sound imaging, which depends on accessibility at early preg-
nancy (Committee on Obstetric Practice, the American Institute
of Ultrasound in Medicine, and the Society for Maternal-Fetal
Medicine, 2017). Missing the time window is common even in
developed countries: in the United States, approximately
900,000 pregnancies annually do not have a prenatal visit before
the second or third trimester (Martin et al., 2018).
The maternal circulatory system connects with the fetal circu-
latory system through the placenta, carrying bioactive molecules
and biomarkers such as steroid hormones, micronutrients, and
circulating nucleic acids, whose concentrations alter as gesta-
tion progresses (King, 2000; Koh et al., 2014; Tulchinsky et al.,
1972; Wang et al., 2016). Recent work on cell-free RNA suggests
that markers in maternal blood can be used to estimate gesta-
tional age, but sequencing can be expensive and time-
consuming, and the accuracy, at present, is not ideal (Ngo

1680 Cell 181, 1680–1692, June 25, 2020 ª 2020 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
 ll
Resource

Table 1. Demographics and Birth Characteristics of the The study identified a large number of pregnancy-related metab-
Discovery and Validation Cohorts olites and metabolic pathways offering a comprehensive view of
Discovery Test Set 1 Test Set 2 the metabolite changes during healthy pregnancy and the post-
partum period. Leveraging the high-resolution datasets, we built
N = 21 N=9 N=8
a metabolic clock that not only predicts gestational age in high
Demographics
accordance with the first-trimester ultrasound, the clinical gold
Maternal age at birth, 29.8 ± 3.1 29.7 ± 3.3 31.4 ± 1.0 standard, but also recovers personal pregnancy variations unde-
years
tected by ultrasound but capable of affecting delivery time.
Previous births, No. (%)
0 13 (61.9) 6 (66.7) 4 (50.0) RESULTS
1 8 (38.1) 2 (22.2) 3 (37.5)
R2 0 (0) 1 (11.1) 1 (12.5) Danish Pregnancy Cohort: A Study of Normal Pregnancy
with High-Density Sampling
Pre-pregnancy BMI, 22.1 ± 2.9 21.2 ± 3.4 21.1 ± 1.6
kg/m2 To capture the highly dynamic pregnancy process at high reso-
lution, we established a multi-year single-center Danish normal
Smoking during pregnancy, No. (%)
pregnancy cohort and a design of high-density blood sampling.
Yes 0 (0) 0 (0) 1 (12.5) Consenting female participants submitted weekly blood draws
No 18 (85.7) 9 (100) 6 (75.0) beginning in week 5 of pregnancy and ending in the postpartum
Missing 3 (14.3) 0 (0) 1 (12.5) period. A total of 30 women with weekly blood samples were as-
Alcohol during pregnancy, No. (%) signed to a discovery (N = 21) and a validation (test set 1, N = 9)
Yes 5 (23.8) 1 (11.1) 1 (12.5)
cohort (Table 1; Figures 1A and S1A). The samples were
analyzed in two separate years. In addition, another separate
Average number of 0.80 1.0 0.25
set of women (N = 8) was included as the secondary validation
units per week
cohort. These samples were analyzed independently three years
No 13 (61.9) 8 (88.9) 6 (75.0)
apart from the discovery cohort (test set 2) (Table 1).
Missing 3 (14.3) 0 (0) 1 (12.5)
Birth characteristics Weekly Pregnancy Progression Is Precisely Ordered by
Gestational age, days 281 ± 8.4 280.7 ± 8.3 279.3 ± 9.5 Metabolites
Mode of delivery, No. (%) We randomized the 784 samples from the first 30 subjects within
each cohort (Discovery and test set 1), processed them by using
Spontaneous vaginal birth 10 (47.6) 5 (55.6) 4 (50.0)
a standardized protocol (Contrepois et al., 2015), and analyzed
Induced vaginal birth 7 (33.3) 1 (11.1) 3 (37.5)
them by liquid chromatography-mass spectrometry (LC-MS)
C-section before onset 1 (4.8) 3 (33.3) 1 (12.5) for untargeted metabolomics across two separate years. After
of labor
quality control, data filtering, and normalization (see STAR
C-section during labor 3 (14.3) 0 (0) 0 (0) Methods) (N = 30), we identified 9,651 metabolic features across
Birth weight, grams 3,638 ± 500 3,803 ± 662 3,362 ± 493 the different samples. Of these, 4,995 features (51.7%) were
Birth length, centimeters 52.4 ± 2 53.3 ± 2 51 ± 2.3 altered during pregnancy and/or the postpartum period (false
Gender of child, No. (%) discovery rate [FDR] < 0.05), suggesting extensive metabolic
Male 9 (42.9) 5 (55.6) 5 (62.5)
changes occur during pregnancy. We examined the data glob-
ally with principal component analysis (PCA), in which the sam-
Female 12 (57.1) 4 (44.4) 3 (37.5)
ples were distributed on the basis of the first two principal com-
Values are means (SDs) or numbers (percentages).
ponents according to their gestational stages (Figure 1B; Scree
plot in Figure S1B and the partial least-squares discriminant
et al., 2018). Therefore, a more accurate and cost-effective analysis [PLS-DA] results in Figure S1C), regardless of individual
method for estimating gestational age and delivery time, variation and batches (Figures S1D and S1E). Interestingly, we
possibly using blood metabolites, is needed. In addition, current found that metabolites with uni-directional behaviors dominated
clinical tests often only focus on a few markers, whereas the features, and over half of them increased across pregnancy
research covering more molecules often examines the profiles until reaching their peaks immediately before labor (Figures S1F
at one or a few time points during pregnancy (Bahado-Singh and S1G).
et al., 2012; Chan et al., 2003; Dudzik et al., 2014; Gagnon To understand the potential function of pregnancy-related
et al., 2008; Kenny et al., 2010; Koh et al., 2014; López-Hernán- metabolites, we first annotated metabolic features by using an
dez et al., 2019; Romero et al., 2010; Sachse et al., 2012; Soldin in-house library and a combined public spectral database (see
et al., 2005). Thus, a high-resolution landscape of pregnancy- details in STAR Methods). A total of 952 metabolic features
related metabolites during healthy pregnancy and the post- were mapped to 687 compounds, which include plasma metab-
partum period is still poorly understood. olites with important functions in humans. We then applied
Here, we use untargeted metabolomics (Kaddurah-Daouk significance analysis for microarrays (SAM) to examine the cor-
et al., 2008) to systematically profile blood metabolites relation between the abundance of each compound and the
throughout pregnancy with weekly sampling of maternal blood. reported gestational age of a woman at blood sampling. Among

Cell 181, 1680–1692, June 25, 2020 1681

 ll
Resource

B C
Gestational
Age/ weeks
10 20 30 40 Postpartum
Estriol-16-Glucuronide
60 Pr Estrone 3-sulfate
eg 0.15 THDOC
Af na N-Acetyl-D-glucosamine

Weekly log2(intensity) change

te
rc
nc Progesterone
hi
y 17alpha-Hydroxyprogesterone
ld LPC(P-18:0)
bi LPC(20:5) 3-Acetoxypyridine
30 rth 0.10 MG(14:1) 5-Pregnane-3,17-diol-20-one 3-sulfate
LPC(P-18:1) 7alpha,24-Dihydroxy-4-cholesten-3-one
PC2 (3.9 %)

PE(P-16:0e/0:0) Theophylline
DHEA-S 1-Methylxanthine
Oleoylcarnitine C16PAF (Platelet-activating factor)
beta-Glycyrrhetinic acid Cortisone
0 0.05 LPC(17:0) Docosadienoic acid
LPC(P-16:0)
MG(24:1)
MG(24:0)
LPE(22:2)
MG(22:2)
0.00 MG(20:0)
−30 Top increased
Top decreased
−0.05 All others

−60 −30 0 30 60 0 100 200

PC1 (5.1%) Rank

D E

Log2(Intensity) Log2(Intensity)
up
up

ro

Norm to 1st Trimester Norm to 1st Trimester

ro

G
G

Estriol−16−Glucuronide 8,9−DHET
Estrone 3−sulfate 4 Hexadecadienoylcarnitine 2
LPE(22:2)
Progesterone LPC(20:5)
2 1
17alpha−Hydroxyprogesterone LPC(P−18:0)
THDOC LPC(P−16:0)
0 Oleoylcarnitine 0
C16 PAF (Platelet−activating factor) LPE(22:4)
Taurochenodeoxycholate −2 LPC(18:2) −1
17,18−EpETE LPE(20:1)
Cortisone −4 LPC(17:0) −2
LPE(22:1)
Cortisol LPE(20:0)
Theobromine LPE(20:3)
Tetracosapentaenoic acid Sinapyl alcohol Group
Tetracosatetraenoic acid 2−Phenylbutyric acid Amino acid metabolism
Isobutyryl−L−carnitine Bile acid biosynthesis
1−Methylxanthine LPC(24:0) Ca eine metabolism
Docosadienoic acid PC(18:1(9Z)e/2:0) Fatty acid metabolism
Erucic acid Tricosanoic acid Phospholipid metabolism
Hydroxybupropion Steroid hormone biosynthesis
Caffeine Others
MG(18:1)
Theophylline Glycyrrhetinic acid
N−Acetyl−D−glucosamine PE(P-16:0e/0:0)
5−Pregnane−3,17−diol−20−one 3−sulfate DHEA-S
3−Acetoxypyridine LPC(P−18:1)
Valylhistidine
7alpha,24−Dihydroxy−4−cholesten−3−one Glycochenodeoxycholate
Dodecanoylcarnitine Androsterone sulfate
Androstane−3,17−diol 3−Hydroxyoleylcarnitine
Pregnenolone sulfate PC(22:1/22:1) (Lecithin)
beta−Glycyrrhetinic acid
Corticosterone MG(14:1)
Sphingosine MG(24:0)
7−Methylguanine MG(24:1)
Ketoisovaleric acid
Cyclo(leucylprolyl) MG(20:0)
Tetracosahexaenoic acid MG(22:2)
16
18
20
22
24
26
28
30
32
34
36
38
40
PP

16
18
20
22
26
24
28
30
32
34
36
38
40
PP

Adj. GA (weeks) Adj. GA (weeks)

Figure 1. Untargeted Metabolomics Cluster the Weekly Plasma Samples Precisely According to Gestational Age
(A) Sampling scheme. Note that validation cohort refers to test set 1 in Table 1.
(B) Principal component analysis (PCA) distributed individual samples according to pregnancy stages (based on 9,651 features). The two PCs explaining the
largest part of the variation are shown.
(C) Plot shows the top 15 increased (red) and decreased (blue) metabolites (with MSI level 1 or 2 identification) in pregnancy.
(D and E) Heatmap displays the metabolite signal intensity averaged across individuals, showing the top 68 altered metabolites (D) increased and (E) decreased
by the end of pregnancy. Abbreviations are as follows: PP, postpartum. The gestational ages (GAs) were calculated by scaling delivery events to 40 weeks. The

(legend continued on next page)

1682 Cell 181, 1680–1692, June 25, 2020


Resource
the 687 annotated compounds, 460 compounds were signifi-
cantly associated with pregnancy (67.0%; FDR < 0.05, SAM).
In addition, 264 compounds were identified with a metabolomics
standards initiative (MSI) level 1 or 2 identification (Viant et al.,
2017), among which 176 compounds (66.7%) were significantly
associated with pregnancy, as determined by linear regression
with gestational age (FDR < 0.05, SAM).
Our dense sampling revealed detailed temporal patterns of
molecular changes. Among the top 68 metabolites (of the 176)
that changed over 50% during whole pregnancy, those that
increased (N = 30) included steroid hormones estriol-16-glucu-
ronide, estrone 3-sulfate, and tetrahydrodeoxycorticosterone
(THDOC). All three increased more rapidly than the well-known
steroids progesterone and 17a-hydoxyprogesterone (FDR <
0.05, SAM) (Figures 1C and 1D). By contrast, the top metabolites
that decreased during pregnancy (N = 38) were mostly lipids or
lipid-like molecules, such as monoacylglycerides (MGs), lyso-
phosphatidylcholines (LPC or lysoPC), and oleoylcarnitine (Fig-
ures 1C and 1E). Hierarchical clustering of the weekly samples
on the basis of the top 68 altered metabolites revealed a week-
order mostly consistent with the actual progression of gesta-
tional age (Table S1; Figures 1D and 1E). Intriguingly, most of
these metabolite changes rapidly returned to baseline after
childbirth (postpartum) (Figures 1B and 1D and 1E). Together,
these results suggest a dramatic and programmed change of hu-
man blood metabolites at a system level during pregnancy.
Metabolite Groups Altered during Pregnancy
To detect the functional groups of metabolites that change during
pregnancy, we performed correlation analysis on the temporal in-
tensity profiles of the top 68 pregnancy-related compounds
mentioned above. In Figure S2, metabolites that were significantly
increased or decreased tended to cluster together. Using existing
structural and biological information, we first categorized the top
changing compounds in pregnancy into seven groups. Interest-
ingly, compounds of the same groups tended to cluster together
in the correlation matrix. On the basis of the correlation relation-
ship, we constructed a regularized partial correlation network us-
ing all pregnancy-related compounds to explore the potential reg-
ulatory relationships (Figures 2A and S2B). The topology of the
network indicates that different metabolite groups occupied
different positions; dense interactions occurred between both in-
ter- and intra- metabolite groups with the densest interactions be-
tween central steroid hormones (Figure 2A). These findings high-
light that even though the amount of each compound dynamically
changes during pregnancy, a highly coordinated metabolite reg-
ulatory network underlies the pregnancy process.
We next examined the main clusters that were present in the
correlation analysis. Three main clusters emerged from the hier-
archical clustering of metabolites (Figure S2A), with a steroid
cluster (e.g., antrostane-3,17-diol, estriol-16-glucuronide, pro-
gesterone, 17a-hydroxyprogesterone, and THDOC) sitting be-
tween the large clusters of lipids and non-lipid molecules.
Compared with the other steroids in this cluster that slowly but
week order, which mostly coincides with the actual order, was ordered by hierarchical clustering on the basis of Manhattan distances. The intensities averaged
before 14 weeks of all women were used as the baseline.
See also Figure S1 and Table S1.
ll
steadily increased throughout pregnancy, estriol-16-glucuro-
nide exhibited a rapid increase before week 24, (Figures 1D
and 2B). Nearly all upregulated metabolites positively correlated
with this cluster of steroids, whereas all downregulated metabo-
lites negatively correlated with this cluster (Figure S2A). This
result suggests that different steroid hormones might regulate
global metabolome dynamics during pregnancy.
Within the lipid cluster, intra-correlation was relatively high. The
largest cluster was composed of LysoPCs (Figures 2A and S2A), a
class of phospholipids. LysoPCs gradually decreased during
pregnancy and increased after childbirth in a pattern that highly
correlates with the steroid dehydroepiandrosterone sulfate
(DHEA-S) (Figure 2C). LysoPCs are bioactive pro-inflammatory
lipids that have been linked with organismal oxidative stress
and inflammation (Sevastou et al., 2013). The second-largest
cluster of lipids included several free fatty acids that were highly
correlated within the cluster (Figures 2A and S2A). Long-chain
fatty acids showed intricate dynamics in their amounts revealed
by the dense sampling. Hexadecadienoylcarnitine and tetracosa-
hexaenoic acid (THA) decreased at the beginning of pregnancy,
followed by waves of increased amounts similar to other fatty
acids in the late second and third trimesters (Figure 2D). After
childbirth, the amounts of most long-chain fatty acids decreased,
except for hexadecadienoylcarnitine (Figure 2D).
Within the non-lipid cluster, one sub-cluster included five high-
ly correlated metabolites belonging to the same caffeine meta-
bolism pathway (Figures 2A and S2A). All five metabolites were
consistently elevated during pregnancy, and caffeine reached
a concentration three times higher at the end of pregnancy
than at the beginning (Figure 2E). This elevation might be due
to a slower caffeine metabolism in pregnant women rather than
an increase in coffee intake (Knutti et al., 1981). Overall, among
the 68 top-altered metabolites in pregnancy, functional metabo-
lite groups (e.g., steroids, LysoPCs, fatty acids, and caffeine
metabolites) were altered in an orchestrated manner during
pregnancy, and individual compounds within each group
showed inter-correlation to each other (Figure 2A).
Orchestrated Metabolome Reconfigurations Span
Multiple Pathways during Pregnancy
Next, we longitudinally examined the global pathway changes of
all 687 annotated compounds during normal pregnancy. Among
the 48 mapped Kyoto Encyclopedia of Genes and Genomes
(KEGG) pathways, 34 showed significant changes (70.8%,
adjusted FDR < 0.05, global test) (Goeman et al., 2004; Xia and
Wishart, 2010a) through metaboanalystR (Figure 3A) (Chong
et al., 2018), suggesting large-scale pathway changes of meta-
bolism in pregnancy. To quantify the pathway activities through
gestational age, we calculated the average intensity of metabo-
lites in the pathways (Figure 3B; see STAR Methods). Among the
top altered pathways (Figure 3A), steroid hormone biosynthesis
showed elevated activity precisely timed to gestation, peaking
before the end of pregnancy and then declining sharply shortly
after delivery (Figure 3B). Along with the essential roles of steroid
Cell 181, 1680–1692, June 25, 2020 1683
 ll
Resource

A Ketoisovaleric acid
LPC(P−18:1)
LPE(22:2)
Theobromine
LPC(P−16:0)
1−Methylxanthine
Theophylline PE(P−16:0e/0:0)
LPE(22:1)
Caffeine
LPE(20:0) LPC(P−18:0) Group
LPC(17:0) 2−Phenylbutyric acid
Amino acid metabolism
LPC(18:2) LPE(20:1) LPE(22:4)
Bile acid biosynthesis
Cyclo(leucylprolyl)
LPE(20:3) Caffeine metabolism
Taurochenodeoxycholate Fatty acid metabolism
3−Acetoxypyridine Oleoylcarnitine Phospholipid metabolism
PC(22:1/22:1) (Lecithin)
Glycochenodeoxycholate PC(18:1(9Z)e/2:0) 3−Hydroxyoleylcarnitine Steroid hormone biosynthesis
N−Acetyl−D−Pregnenolone sulfate Others
glucosamine MG(14:1)
Androsterone sulfate
ate
te
e LPC(20:5)
Sinapyl alcohol
Estrone 3−sulfate
Androstane−3,17−diol
Sphingosine Tetracosahexaenoic acid
5−Pregnane−3,17
−diol−20−one 3−sulfate THDOC 7−Methylguanine Tetracosapentaenoic acid
Estriol−16−
Glucuronide Hexadecadienoylcarnitine
Tetracosatetraenoic acid
Progesterone 17alpha− DHEA−S Erucic acid
Hydroxyprogesterone
Dodecanoylcarnitine
Tricosanoic acid Docosadienoic acid
7alpha,24−Dihydroxy−
Cortisone C16 PAF
4−cholesten−3−one Valylhistidine
Cortisol
17,18−EpETE LPC(24:0)
MG(24:1) MG(20:0)

Corticosterone beta−Glycyrrhetinic acid

Isobutyryl−L−carnitine MG(24:0) MG(22:2)

Hydroxybupropion
Glycyrrhetinic acid MG(18:1)

8,9−DHET

B C
Steriod hormone biosynthesis Phospholipids and DHEA-S
1.0
Mean log2(Intensity)

5.0
Mean log2(Intensity)

0.5 PC (22:1/22:1) (Lecithin)

LPC (18:2)
2.5 LPC (17:0)
Estriol−16−Glucuronide 0.0
THDOC DHEA−S
0.0 17alpha−Hydroxyprogesterone PE (P−16:0e/0:0)
Progesterone −0.5 LPE (22:2)
5−Pregnane−3,17−diol−20−
−2.5 one 3−sulfate −1.0
Androstane−3,17−diol

1416 24 32 40 PP 1416 24 32 40 PP
Gestational age (weeks) Gestational age (weeks)

D E
Long-chain fatty acids Caffeine metabolism
Mean log2(Intensity)

Mean log2(Intensity)

1.5
1.0
C16 PAF (Platelet−activating factor)
Docosadienoic acid 1.0
0.5
Tetracosapentaenoic acid Caffeine
Tetracosatetraenoic acid 0.5 Theophylline
0.0 Sphingosine 1−Methylxanthine
Erucic acid Theobromine
−0.5 Dodecanoylcarnitine 0.0 Cyclo (leucylprolyl)
Tetracosahexaenoic acid (THA)
Hexadecadienoylcarnitine
−1.0
−0.5
1416 24 32 40 PP 14 16 24 32 40 PP
Gestational age (weeks) Gestational age (weeks)

Figure 2. Functional Metabolite Groups Altered during Pregnancy

(A) Regularized partial correlation network of top altered compounds in pregnancy. Here, each node represents a compound, and each edge represents the
strength of partial correlation between two compounds after conditioning on all other compounds in the datasets. Edge weights represent the partial correlation
coefficients. Note that the seven nodes with red circles with central positions were also the predictors in the models of Figures 4 and 5.
(B–E) The average levels of the metabolite changes against the gestational progression in the clusters of steroid hormone biosynthesis (B), phospholipids and
DHEA-S (C), long-chain fatty acids (D), and caffeine metabolism (E). The intensities were normalized to the baseline, which was defined by averaging all samples
before 14 weeks. The standard errors, derived from 30 subjects, are shown. The GAs were standardized by scaling delivery events to 40 weeks. Abbreviation is as
follows: PP, postpartum. Note that the y axis scale is much larger for steroids than for other compounds.
See also Figure S2.

1684 Cell 181, 1680–1692, June 25, 2020

 ll
Resource

A B

Figure 3. System-Wide Reconfiguration of Metabolic Pathways during Pregnancy

(A) Metabolic pathways undergoing significant changes during pregnancy. Red dots denote pregnancy-related pathways with FDR < 0.05, which were further
analyzed in (B). The topological pathway effects were quantified by using published methods (Xia and Wishart, 2010a).
(B) Heatmap shows the temporal changes of pregnancy-related pathway activities during pregnancy and postpartum (PP). To quantify pathway activity, the
average intensity of metabolites in each pathway at each time window was calculated. Note that although some pathways contained mainly the metabolites
increasing or decreasing during pregnancy, many pregnancy-related pathways contained both metabolites increasing and decreasing. Thus, their average
values would not show large changes in the heatmap. For each pathway, the average values from samples earlier than 14 weeks (marked as week 14) were used
as the baseline.
(C) Human disease states that correlated with pregnancy-related metabolites on the basis of published metabolomics data (Chong et al., 2018).
See also Figure S3.

hormones in maintaining pregnancy and later inducing parturi-
tion (Mendelson, 2009), we observed an orchestrated elevation
of many components centered on progesterone, including
some less well-characterized hormones (Figure S3A). Consistent
with known sources of pregnancy metabolites (e.g., hormones)
(Maltepe and Fisher, 2015), metabolite set enrichment analysis
(MSEA) (Xia and Wishart, 2010b) revealed that the adrenal cor-
tex, gonad, and placenta were among the top origins of preg-
nancy-related metabolites (Figure S3B). The ability to recognize
many well-known and less-characterized steroid hormone
changes across pregnancy validates our approach.
In addition to the steroid pathway, we observed a dynamic
pattern of metabolite changes with pregnancy in other pathways,
such as the arachidonic acid metabolism pathway (Figures 3A,

Cell 181, 1680–1692, June 25, 2020 1685

 ll
3B, and S3C). We observed 20-HETE amounts increased until
week 34; 20-HETE is potentially linked to the regulation of blood
pressure and renal function during pregnancy (Wang et al.,
2002; Wu et al., 2014) (Figures S3C and S3D). By contrast,
5-HETE amounts generally decreased during pregnancy, poten-
tially associated with its regulation of the uterus (Figures S3C
and S3D) (Edwin et al., 1997; Pearson et al., 2010). Thus, beyond
energy metabolism and hormones, a system-wide reconfiguration
of the metabolome occurs as the mother adapts to pregnancy. In
addition, based on MSEA analysis, many pregnancy-related me-
tabolites are implicated in human disease states, including obesity
and prepartum depression (Figure 3C) (Xia and Wishart, 2010b).
The Metabolic Clock of Normal Pregnancy Identified by
Machine Learning
We next determined whether we can build a metabolic clock
based on the high-resolution profile to predict gestational age
for individual plasma samples. In the discovery cohort (samples
n = 507, subjects N = 21), we applied feature selection (lasso
[least absolute shrinkage and selection operator]) with all 9,651
features to build the linear regression model that shows optimal
cross-validation performance for predicting a given phenotype in
this cohort. We then ran the validation cohort data (test set 1,
samples n = 245, subjects N = 9) through the model established
in the discovery cohort to measure the independent perfor-
mance of our model (Figure 4A; see STAR Methods).
We first tested whether the metabolome change can quantita-
tively determine the gestational age in normal pregnant women.
Feature selection in the discovery cohort yielded a linear model
that included 42 metabolic features (Figure S4A; Table S2). In
the cross-validation test of 507 samples in the discovery cohort,
the metabolic model predicted gestational age in weeks
(GAmetabolic) that correlated with gestational age estimated by
the first-trimester ultrasound (GAultrasound, in compliance with
the clinical standard of care) with a Pearson correlation coeffi-
cient (R) of 0.96 (R2 = 0.93, p < 1 X 10
100, root mean squared
error [RMSE]= 2.49) (Figure S4B). In the independent-validation
cohort, the model yielded a similar R of 0.95 (R2 = 0.91, p < 1 X
10
100, RMSE = 2.76, test set 1) (Figure S4C). This indicates
metabolic features can accurately predict the gestational age
on the basis of a blood sample from a pregnant woman.
For potential clinical use, we next tested whether we can use
the annotated compounds in blood to predict the gestational
age in pregnant women. We performed feature selection in dis-
covery cohort by using the 264 level 1 and level 2 compounds
identified in the Human Metabolome Database (HMDB) in the dis-
covery cohort (Table S3). This yielded a linear model including five
compounds (Figures S4D and 4D) that together are highly predic-
tive. We first evaluated the performance of the model in a 10-fold
cross-validation (CV) test in the discovery cohort, in which sam-
ples were distributed into folds by subject instead of by sample
to prevent person-specific information cross-over between the
training folds and the test fold. In the CV test, the metabolic-clock
model produced a result (GAmetabolic) that correlated with the
gestational age estimated by the first-trimester ultrasound
(GAultrasound) with a Pearson correlation coefficient (R) of 0.92
(R2 = 0.85, p = 8e
222, RMSE = 3.67) (Figure 4B). To avoid the
hyperparametric selection bias, we further evaluated the perfor-
1686 Cell 181, 1680–1692, June 25, 2020
Resource
mance of our model in two independent validation cohorts (test
set 1 and test set 2). In test set 1, the model yielded an R of
0.89 (R2 = 0.80, p = 8e
93, RMSE = 4.11) (Figure 4C). The model,
including four steroids and one lipid (Figure 4D), was further veri-
fied in a second independent-validation cohort of eight individuals
with R of 0.91 (R2 = 0.83, RMSE = 3.05, samples n = 32, test set 2)
(Table 1; Figure 4E). The compound identifications were
confirmed by chemical standards (Figures 4F–4H, S4E, and
S4F; Table S4; see STAR Methods). We noted that four of these
five compounds are among the central steroid cluster forming a
dense correlation network with one another (Figure 2A).
As pregnancy progresses toward term, clinical classifications
and decisions often need to be made based on the timing of preg-
nancy (e.g., < 37 weeks for preterm birth). Babies born before
37 weeks are considered preterm, those born before 20 weeks
are considered a miscarriage, and those born before 24 weeks
have low survival. Therefore, for clinical action it is important to
accurately classify the gestational age by clinical cutoff points at
weeks 20–37. As a proof-of-principle, we tested the potential use
of the metabolome data to classify the normal pregnancy samples
as before or after 20, 24, 28, 32, and 37 gestational weeks (Fig-
ure S5A). First, using only samples from the third-trimester (>
28 weeks of gestation), the time window where women were
more susceptible to preterm delivery, we determined whether the
identified maternal blood metabolites can distinguish the sample
gestational age as before or after 37 weeks. Both the discovery
and the validation prediction yielded an area under the receiver
operating characteristics (AUROC) over or close to 0.90 (Fig-
ure S5B; see STAR Methods). Remarkably, the prediction model
contained only three metabolites, and the abundance range of
each individual metabolite separated the > 37 week samples
from the < 37 week samples for all but one to two validation sub-
jects (Figures S5C–S5F). Similarly, using samples across the whole
pregnancy, we found that metabolites can also accurately distin-
guish pregnancy samples before or after other important gesta-
tional age cutoffs, such as 20, 24, 28, and 32 gestational weeks
(Figures S5A and S5G–S5J).
Personal Metabolic Clock of Pregnancy Linked with
Timing of Delivery and Fetal Growth
Next, we examined the metabolic clock prediction performance
in individuals. First, we noted that for most individuals, our model
produced predictions consistently aligned with the gestational
age estimated by the first-trimester ultrasound (Figures 5A and
5B). In both cross validations for Discovery and test set 1, the
prediction deviation (measured by RMSE) in individuals centered
around 3 weeks (Figures S6A and S6B). However, in each data-
set, there is a small population of individuals with higher predic-
tion deviation. When we examined these individuals (e.g.,
subjects 1, 2, and 4), we found the predictions were not more
randomly scattered than other individuals. Rather, in the majority
of them, predictions shifted away from the actual gestational
ages in a portion of the pregnancy duration (Figures 5A and
5B), suggesting effects from non-random causes.
We hypothesized that some of these large prediction deviations
might arise from biological causes, particularly from the maternal-
fetal interaction. It is reported that the fetoplacental unit secretes
hormones in conjunction with fetal growth and development
 ll
Resource

B C

D E

F G H

Figure 4. Metabolic Clock Of Pregnancy: Five Metabolites Selected by Machine Learning Can Accurately Predict the Timing of Normal
Pregnancy Progression in Both a Discovery and Two Validation Cohorts
(A) Design of the analytical pipeline.
(B and C) Gestational age (GA) predicted by the linear model consisting of five identified metabolites (GAmetabolic, y axis) highly correlates with clinical values
determined by the standard of care (by first-trimester ultrasound [GAultrasound] x axis) in the Discovery (B) and the validation cohort (test set 1) (C). Note that two
samples presented as outliers in the validation cohort, possibly because of occasional mass-spectrometry signal instability in given samples. The 95% confi-
dence interval for the linear regression is represented by the gray area.
(D) Contribution of the five metabolites to the gestational age prediction model.
(E) Gestational age predicted by the five metabolites (GAmetabolic, y axis, scaled) correlates with clinical values determined by the standard of care (by first-
trimester ultrasound [GAultrasound] x axis) in the test set 2 cohort. The 95% confidence interval for the linear regression is represented by the gray area.
(F–H) Confirmation of the metabolites predicting gestational age in the metabolic clock model by standard compounds, THDOC (F), estriol-16-glucuronide (G),
and progesterone (H) (see two additional compounds PE(P-16:0e/0:0) and DHEA-S in Figures S4E and S4F). Measured MS/MS spectral fragmentation profiles
(top, in black) matching chemical standards (bottom, in red). Note that the discovery results were from the 10-fold CV to avoid over-fitting (see STAR Methods).
See also Figures S4 and S5 and Tables S2–S4.

(Murphy et al., 2006). Indeed, we noted that the average prediction gestational age estimation determined by first-trimester ultrasound
deviation strongly correlated with adjusted infant birth weight (Fig- in mothers with a heavier fetus while being delayed in the mothers
ure 5C, adjusted for gestational length; see Figure S6C and STAR with a lighter fetus. The finding suggests that fetal growth appears
Methods). Thus, the overall metabolic clock tends to outpace the to be one of the inputs read by the metabolic clock.

Cell 181, 1680–1692, June 25, 2020 1687

 ll
Resource

A B
RMSE= 2.6 RMSE= 4.2 RMSE= 3.1 RMSE= 2.4 RMSE= 2.1 RMSE= 3.2 RMSE= 3.4 RMSE= 4.3 RMSE= 3.0 RMSE= 2.9
R2= 0.88 R2= 0.90 R2= 0.93 R2= 0.94 R2= 0.88 R2= 0.93 R2= 0.89 R2= 0.88 R2= 0.92 R2= 0.90
40 p= 5e−11 40 p= 2e−13 40 p= 3e−14 40 p= 2e−19 40 p= 3e−09 40 p= 2e−09 40 p= 7e−12 40 p= 9e−14 40 p= 2e−17 40 p= 1e−12
30 30 30 30 30 30 30 30 30 30
20 20 20 20 20 20 20 20 20 20
10 10 10 10 10 10 10 10 10 10
GAmetabolic (weeks)

GAmetabolic (weeks)
Subject 18 Subject 11 Subject 15 Subject 5 Subject 16 Subject 7 Subject 19 0 Subject 30 0 Subject 25 Subject 26
0
0 0 0
0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40

RMSE= 2.7 RMSE= 5.7 RMSE= 3.3 RMSE= 4.8 RMSE= 2.0 RMSE= 4.6 RMSE= 4.5 RMSE= 3.6 RMSE= 3.0 RMSE= 4.6
R2= 0.96 R2= 0.82 R2= 0.93 R2= 0.90 R2= 0.94 R2= 0.90 R2= 0.89 R2= 0.91 R2= 0.88 R2= 0.82
40 p= 2e−10 40 p= 3e−12 40 p= 3e−16 40 p= 2e−17 40 p= 7e−16 40 p= 3e−14 40 p= 5e−15 40 p= 2e−13 40 p= 6e−12 40 p= 3e−14
30 30 30 30 30 30 30 30 30 30
20 20 20 20 20 20 20 20 20 20
10 10 10 10 10 10 10 10 10 10
0 Subject 21 Subject 4 0 Subject 17 0 Subject 1 Subject 10 Subject 2 Subject 3 Subject 28 Subject 22 Subject 24
0 0 0 0 0 0
0
0 10 20 30 40 0 10 20 30 40 0 10 20 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
RMSE= 3.3 RMSE= 4.1 RMSE= 2.6 RMSE= 3.5 RMSE= 2.5 RMSE= 4.1 RMSE= 3.0
RMSE= 2.6 RMSE= 3.3 RMSE= 7.5
R2= 0.88 R2= 0.89 R2= 0.97 R2= 0.93 R2= 0.94 R2= 0.87 R2= 0.84
R2= 0.90 R2= 0.86 R2= 0.56
40 p= 3e−11 40 p= 3e−15 40 p= 1e−08 40 p= 4e−17 40 p= 3e−14 40 p= 1e−08 40 p= 5e−08
40 p= 9e−13 40 p= 7e−12 40 p= 3e−05
30 30 30 30 30 30 30
30 30 30
20 20 20 20 20 20 20
20 20 20
10 10 10 1030 10 10 10
10 10 10
0 Subject 12 Subject 8 0 Subject 20 0 Subject 14 Subject 13 Subject 9 Subject 6 Subject 29 Subject 27 Subject 23
0 0 0 0 0 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40

GAultrasound (weeks) GAultrasound (weeks)

C D E
P = 0.01 P = 0.02
R2= 0.22 R2= 0.28
( (GAmetabolic- GAultrasound)) (weeks)

42
2
GA at delivery (weeks)
Average discrepencies

41

0
Weeks to delivery: 8 4 2
40
AUC in validation: 0.91 0.87 0.86
THDOC 1 2 1
−2 39 Androstane-3,17-diol 2
Estriol-16-glucuronide 1 3
N= 29 N= 18 17alpha-Hydroxyprogesterone 2
with birthweight info.
38
with natural labor onsets PE(P−16:0e/0:0) 3
−4
−1000 −500 0 500 1000 −4 −2 0 2
Adjusted brithweight deviation Average discrepencies
from the population mean (g) ( (GAmetabolic- GAultrasound)) (weeks)

F Weeks to delivery (WD) < 2w G H

1.00 Weeks to delivery predictors THDOC
WD > 2w
WD < 2w
0.5 13.5
True positive rate

0.75
Contribution

0.4
log2(Intensity)

13.0
0.3
0.50 12.5
0.2
AUC: 0.88
Discovery CV N= 128 0.1 12.0
0.25 95% CI: 0.82-0.95 0.0
AUC: 0.86 11.5
C
O

Test Set 1 N= 65
D

0.00
TH

95% CI: 0.80-0.94

28

29

30
22

25

26

0.00 0.25 0.50 0.75 1.00

Subject ID
False positive rate

Figure 5. Personal Metabolic Clock of Pregnancy Linked with Timing of Delivery and Fetal Growth
(A and B) Highly correlated patterns of the metabolic-clock-predicted gestational age (GAmetabolic) of the five-metabolite model with the gestational age estimated
by the first-trimester ultrasound (GAultrasound) at the individual level in the cross validation (A) and test set 1 (B). Note that the outlier sample with negative prediction
value in Figure 4C belonged to the last subject of the test set 1 and did not show in the current plot with the y axis scale limitation.
(C) The average discrepancies between metabolic-clock-predicted gestational age and ultrasound-estimated gestational age (D(GAmetabolic-GAultrasound)) were
significantly correlated with the fetal growth deviation from the population by person. All 29 subjects who had baby birth weight information are included here. The
95% confidence interval for the linear regression is represented by the gray area.
(D) Average discrepancies between GAmetabolic and GAultrasound (D(GAmetabolic-GAultrasound)) were negatively correlated with the actual delivery weeks (by ultra-
sound-estimation). All 18 subjects who had natural labor onset are included here. Dashed lines marked the ultrasound estimated GA at 40 weeks (due date,
black), GAmetabolic one week earlier than the GAultrasound (blue), and GAmetabolic one week later than the GAultrasound (red). The 95% confidence interval for the linear
regression is represented by the gray area.
(E) Summary of prediction models of 2, 4, and 8 weeks approaching delivery, using two to three metabolites. The contribution rank of each predictor in every
model is listed as number 1, 2, and 3. The weeks to delivery were built using samples of the third trimester (> 28 weeks). AUCs in the validation cohort (test set 1)
are listed.

(legend continued on next page)

1688 Cell 181, 1680–1692, June 25, 2020


Resource
In addition, within the 18 women with natural labor onset (i.e.,
excluding women with induction before labor onset and sched-
uled cesarean-section), we found that the women whose overall
metabolic clock of pregnancy outpaced ultrasound evaluation
tended to deliver earlier, whereas a delay in metabolic clock
correlated with a delayed time to child delivery compared to ul-
trasound estimated due date (Figure 5D). Interestingly, five out of
six women (83%) with a metabolic-clock-predicted gestational
age more than one week later than the ultrasound-estimated
gestational age had natural labor onset after their due date (esti-
mated by ultrasound, marked in red in Figure 5D), and four out of
five women (80%) with metabolic-clock-predicted gestational
age more than one week earlier than the ultrasound-estimated
gestational age had natural labor onset before their due date
(marked in blue in Figure 5D). These results suggest the meta-
bolic clock of pregnancy with maternal metabolites contains in-
formation on the timing of delivery in normal pregnancy.
Prediction for Timing of Delivery
We then tested whether the maternal blood metabolites can also
predict the timing of a normal delivery event within a defined period
(2, 4, and 8 weeks from delivery) approaching the labor events (in the
third trimester). We first examined whether metabolites can predict
a delivery within 2 weeks (weeks to delivery [WD] < 2w). To predict
delivery triggered naturally without outside procedures (such as
scheduled cesarean-section), we only included delivery events
naturally triggered (subjects N = 18, samples n = 193). With just three
metabolites, the metabolome accurately predicted an upcoming
delivery event within 2 weeks in both discovery and validation co-
horts with AUROC close to 0.9 (Figures 5E–5H, S6D, and S6E;
see STAR Methods). Similarly, identified metabolites can also be
used to predict the timing of a normal delivery event within 4 and
8 weeks (Figures 5E and S6F–S6I). Intriguingly, the panels of metab-
olites partially overlapped between the models, whereas the individ-
ual metabolites contributed differently to the models (Figure 5E). All
of the metabolite markers were identified as steroids, except for
phospholipid PE(P-16:0e/0:0), and most of them (three out of five
in total) also appeared in the aforementioned metabolic clock for
gestational age (Figure 5E; Table S4). These results demonstrate
that we can precisely categorize critical pregnancy stages in normal
subjects by using a small number of maternal blood metabolites,
which can be further validated in larger and independent cohorts.
DISCUSSION
In this study, we performed untargeted metabolomics profiling
and identified highly dynamic temporal regulation of metabolic
changes in human pregnancy: more than half of the measured
metabolites and metabolic pathways changed during preg-
nancy. We were able to detect many of the pregnancy-associ-
ated metabolite profiles revealed in previous targeted studies
(F) The logistic regression model based on three metabolites can accurately identify the third-trimester plasma samples approaching the delivery (weeks to
delivery [WD] < 2w; only women with natural labor onset included).
(G) Contribution of the three metabolites to the prediction model of 2 weeks approaching delivery.
(H) Metabolite THDOC showed abundance separations before or after 2 weeks approaching the delivery, except in one subject. See Figure S6 for other me-
tabolites in the model. Note that the discovery results were from the 10-fold CV instead of direct fitting to avoid over-fitting.
See also Figure S6 and Table S4.
ll
(Tulchinsky et al., 1972; Wang et al., 2016) (such as progester-
one, 17-hydroxyprogesterone, and the linoleic acid pathway),
validating our approach. At the same time, we also noted that
a large portion of pregnancy-related metabolites identified in
our study was less well-studied. For example, over 95% of the
pregnancy-related metabolites identified in our study were not
recovered from a targeted metabolic profiling study on preg-
nancy (Wang et al., 2016), demonstrating the power of unbiased
and hypothesis-independent profiling. Among the changing me-
tabolites, the major class that increased was steroids, including
progesterone, which interacts with the hypothalamic-pituitary-
adrenal axis (HPA axis) (Chrousos et al., 1998), and estriol-16-
glucuronide produced by the placenta (Levitz et al., 1984).
Here, the detailed differences in their temporal profiles were re-
vealed by the weekly sampling design of the study (Figures 1D
and 2B). In addition, we discovered less well-studied steroids
in pregnancy, such as the neurosteroid THDOC, an allosteric
modulator of the GABAA receptor that potentially affects stress
and depression in human pregnancy (Hosie et al., 2006; Reddy,
2003). Intriguingly, many pregnancy-related metabolites that
changed, including steroids, quickly returned to the maternal
non-pregnant state after childbirth (Figures 1B and 1D and 1E).
In addition, we also identified a wide variety of non-steroid
hormones whose abundance altered during pregnancy
progression.
These metabolite changes presumably accommodate and/or
reflect important maternal biological physiology during preg-
nancy and fetal growth (Bispham et al., 2003; Prentice and Gold-
berg, 2000). For maternal nutrient metabolism, one of the
decreased carnitines, oleoylcarnitine (Figure 1C), accumulates
during certain metabolic conditions, including fasting (Hoppel
and Genuth, 1980; Minkler et al., 2005). Also, one phosphatidyl-
choline that functions as a micronutrient, lecithin, increased in
pregnancy, suggesting a systematic change in the maternal
nutritional status during gestation. Within molecules reflecting
pregnancy-related physiological changes, consistent with
decreased blood pressure (Hermida et al., 1997), the antihyper-
tensive molecule 20-HETE of the arachidonic acid metabolism
pathway is elevated during pregnancy until the early third
trimester, and its synthesis is regulated in a renal-specific
manner (Wang et al., 2002; Wu et al., 2014). This reveals the high-
ly dynamic temporal regulation of 20-HETE in blood pressure
and kidney function during pregnancy. In contrast, compared
with early pregnancy and postpartum, the amount of 5-HETE
in the same pathway was generally lower in the second and third
trimesters with an increasing trend right before the childbirth,
consistent with previous findings that 5-HETE elevates in the
uterus and amniotic fluid at the onset of human labor (Edwin
et al., 1997; Pearson et al., 2010). In the developing fetus,
changes in hexadecadienoylcarnitine amounts are associated
with congenital heart defects (Bahado-Singh et al., 2014).
Cell 181, 1680–1692, June 25, 2020 1689
 ll
Here, we revealed that the amount of hexadecadienoylcarnitine
in the blood decreased continuously until week 24, then steadily
increased thereafter (Figure 2D). In addition, the amount of long-
chain fatty acids in maternal blood samples is associated with
childhood metabolic health (Maslova et al., 2018). Here, the
omega-3 fatty acid THA decreased during early pregnancy and
gradually increased before childbirth (Figure 2D), suggesting
gestation-related changes in the formation of docosahexaenoic
acid (DHA) (Moore et al., 1995). Our findings are robust even
without a requirement for prior fasting. It will be interesting to
validate these findings in cohorts that have dietary information
and detailed clinical measurements, to define critical ‘‘nutritional
time-zones’’ for micronutrient amounts and further understand
the metabolite changes that are important for physiologic
changes across pregnancy.
Our high-density sampling scheme allowed us to study the tem-
poral alteration of metabolite levels at weekly resolution. For
example, even though many steroid metabolites were elevated
during pregnancy, our profiling was able to show that there were
at least two different behaviors: an early wave (such as progester-
one and 17a-hydroxyprogesterone) and a second wave (such as
estriol-16-glucuronide). These temporal changes of steroids
across pregnancy and after childbirth are at least partially regulated
by the fetoplacental unit, including both maternal adrenal gland
and placenta and fetal adrenals and liver (Diczfalusy, 1953; Frand-
sen and Stakemann, 1961; Raeside, 2017). Further investigation
into the interaction of fetal-maternal contribution will be necessary
for understanding the temporal regulation of these metabolites.
Untargeted metabolome and high-density sampling enabled us
to identify a broad set of high-resolution temporal profiles of me-
tabolites during pregnancy. We hypothesized that this information
might help us to understand the underlying metabolic clock that
times the progression of pregnancy. We found that solely using
the abundance of five compounds, without any other inputs from
clinical features, we can precisely determine the gestational age
of a healthy pregnant woman. The precision surpasses the recent
cell-free RNA model by using maternal blood (Ngo et al., 2018).
Similarly, with two to three compounds, we can categorically pre-
dict many pregnancy cutoff times with high AUC: we can deter-
mine whether a woman has reached 20, 24, 28, 32, or 37 weeks
(clinical cutoffs for miscarriage, age of viability, extremely preterm,
very preterm, and prematurity, respectively) into her pregnancy
(Figure S5A), or whether a woman will enter into labor within the
next two, four, or eight weeks (Figure 5E). The proof-of-principle
study suggested that metabolome bears rich quantitative informa-
tion about pregnancy progression. However, our study has its lim-
itations. The studied population consisted of healthy Caucasian
pregnant women with small variations in clinical characteristics.
In the future, we need to test the models in a larger cohort with
diverse ethnicities and complications. Meanwhile, targeted chem-
ical assays need to be developed on the small panels of identified
metabolite markers that were discovered by untargeted metabolo-
mics to measure the metabolite concentration independent of
batches. Intriguingly, we found the metabolic clock of pregnancy
to be robust in general, but small personal deviations can be
observed, most likely affected by the fetal growth (Figure 5C).
Lastly, we also found that the discrepancies between metabolic
timing and ultrasound suggested biological significance: the
1690 Cell 181, 1680–1692, June 25, 2020
Resource
women that had advanced metabolic clock tended to deliver
earlier than predicted by ultrasound, whereas a delay in metabolic
clock correlated with a delayed time to child delivery (Figure 5D).
In summary, combining untargeted metabolome and high-den-
sity sampling revealed the landscape of metabolome changes dur-
ing pregnancy and the postpartum period with high resolution. The
data itself can serve as a resource for future research. As a proof-
of-principle, we also demonstrated that the temporal abundance
information of metabolome can be used to predict gestational
age with high accuracy in a cohort of healthy women. There is a
great need for accurate timing of pregnancy: in the US alone,
900,000 women annually missed their first-trimester ultrasound
(Martin et al., 2018), currently the only accurate timing method
for pregnancy (Committee on Obstetric Practice, the American
Institute of Ultrasound in Medicine, and the Society for Maternal-
Fetal Medicine, 2017). In low and middle-income countries, acces-
sibility to ultrasound is even more scarce, complicating many preg-
nancies and fetal care down-stream (e.g., identify imminent labor,
manage complications, etc.). Our study demonstrated that the
development of clinical tools with a few metabolites in maternal
blood to time pregnancy is promising. Testing of blood drawn
from the pregnant woman would likely be limited to once or a
few times to be informative and have the potential to benefit preg-
nant women in both developed and developing worlds.
STAR+METHODS
Detailed methods are provided in the online version of this paper
and include the following:
d KEY RESOURCES TABLE
d RESOURCE AVAILABILITY
B Lead Contact
B Materials Availability
B Data and Code Availability
d EXPERIMENTAL MODEL AND SUBJECT DETAILS
B Pregnancy cohort
d METHOD DETAILS
B Plasma sample preparation
B Chemical materials for untargeted metabolomics
B MS acquisition
B Chromatographic conditions
d QUANTIFICATION AND STATISTICAL ANALYSIS
B Section 1: Metabolomics Data Processing
B Section 2: Metabolic Features Identification
B Section 3: Identify Significantly Altered
Features/Compounds
B Section 4: Regularized Partial Correlation Network
B Section 5: Pathway Analysis
B Section 6: Machine Learning for Pregnancy Timing
B Section 7: Analyze the discrepancies between meta-
bolic clock (GA prediction model) and first-trimester ul-
trasound estimations
SUPPLEMENTAL INFORMATION
Supplemental Information can be found online at https://doi.org/10.1016/j.
cell.2020.05.002.

Resource
ACKNOWLEDGMENTS
We thank the dedicated women who made this study possible by generously
participating every week during their pregnancy. We thank R. Jian, L. Tian, and
C. Yeh for technical assistance; S.M.S. Rose, G. Chen, and Y. Zhang for re-
viewing the manuscripts. This work was supported by the Bill & Melinda Gates
Foundation, United States (OPP1128928). M.S. was supported by the Center
for Personal Dynamic Regulomes (2RM1HG00773506). M.M. and B.F. were
supported by the Oak Foundation, Switzerland (OCAY-18-598). M.L.H.R.
was supported by the Independent Research Fund Denmark (6120-000998).
L.S. was supported by a Carlsberg Foundation Postdoctoral Fellowship
(CF15-0899).
AUTHOR CONTRIBUTIONS
M.-L.H.R. and M.M. organized and contributed to the collection of pregnancy
samples. L.L., B.P., B.F., M.S., and M.M. conceptualized the study. L.L., M.S.,
and M.M. created the analysis plan. L.L. processed and analyzed the samples.
L.L. and H.R. analyzed the data. L.L., X.S., S.C., J.S., and N.L. contributed to
the spectral analysis. L.L., M.S., and M.M. wrote the manuscript, and all au-
thors contributed to reviewing and editing the manuscript.
DECLARATION OF INTERESTS
M.S. is a co-founder and member of the scientific advisory boards of the
following: Personalis, SensOmics, Filtricine, Qbio, January, Mirvie, and Ora-
lome. He is a member of the scientific advisory board of Jungla. M.M. is a
co-founder of Mirvie. L.L., M.S., and M.M. are inventors on the patent applica-
tion PCT/US2019/052515 related to this work.
Received: September 4, 2018
Revised: March 11, 2020
Accepted: April 29, 2020
Published: June 25, 2020
REFERENCES
Alkema, L., Chou, D., Hogan, D., Zhang, S., Moller, A.B., Gemmill, A., Fat,
D.M., Boerma, T., Temmerman, M., Mathers, C., and Say, L.; United Nations
Maternal Mortality Estimation Inter-Agency Group collaborators and technical
advisory group (2016). Global, regional, and national levels and trends in
maternal mortality between 1990 and 2015, with scenario-based projections
to 2030: a systematic analysis by the UN Maternal Mortality Estimation Inter-
Agency Group. Lancet 387, 462–474.
Altman, N.S. (1992). An Introduction to Kernel and Nearest-Neighbor Nonpara-
metric Regression. Am. Stat. 46, 175–185.
Bahado-Singh, R.O., Akolekar, R., Mandal, R., Dong, E., Xia, J., Kruger, M.,
Wishart, D.S., and Nicolaides, K. (2012). Metabolomics and first-trimester pre-
diction of early-onset preeclampsia. J Matern Fetal Neonatal Med 25,
1840–1847.
Bahado-Singh, R.O., Ertl, R., Mandal, R., Bjorndahl, T.C., Syngelaki, A., Han,
B., Dong, E., Liu, P.B., Alpay-Savasan, Z., Wishart, D.S., et al. (2014). Metab-
olomic prediction of fetal congenital heart defect in the first trimester. A J Ob-
stet Gynecol 211, e1–e14.
Baumann, D., and Baumann, K. (2014). Reliable estimation of prediction errors
for QSAR models under model uncertainty using double cross-validation.
J. Cheminform. 6, 47.
Bispham, J., Gopalakrishnan, G.S., Dandrea, J., Wilson, V., Budge, H., Keisler,
D.H., Broughton Pipkin, F., Stephenson, T., and Symonds, M.E. (2003).
Maternal endocrine adaptation throughout pregnancy to nutritional manipula-
tion: consequences for maternal plasma leptin and cortisol and the program-
ming of fetal adipose tissue development. Endocrinology 144, 3575–3585.
Blencowe, H., Cousens, S., Chou, D., Oestergaard, M., Say, L., Moller, A.B.,
Kinney, M., Lawn, J., and Born Too Soon Preterm Birth Action, G.; Born Too
ll
Soon Preterm Birth Action Group (2013). Born too soon: the global epidemi-
ology of 15 million preterm births. Reprod. Health 10, S2.
Chambers, M.C., Maclean, B., Burke, R., Amodei, D., Ruderman, D.L., Neu-
mann, S., Gatto, L., Fischer, B., Pratt, B., Egertson, J., et al. (2012). A cross-
platform toolkit for mass spectrometry and proteomics. Nat. Biotechnol. 30,
918–920.
Chan, L.Y., Chiu, P.Y., and Lau, T.K. (2003). Cord blood thyroid-stimulating
hormone level in high-risk pregnancies. Eur. J. Obstet. Gynecol. Reprod.
Biol. 108, 142–145.
Chong, J., Soufan, O., Li, C., Caraus, I., Li, S., Bourque, G., Wishart, D.S., and
Xia, J. (2018). MetaboAnalyst 4.0: towards more transparent and integrative
metabolomics analysis. Nucleic Acids Res. 46, W486–W494.
Chrousos, G.P., Torpy, D.J., and Gold, P.W. (1998). Interactions between the
hypothalamic-pituitary-adrenal axis and the female reproductive system: clin-
ical implications. Ann. Intern. Med. 129, 229–240.
Committee on Obstetric Practice, the American Institute of Ultrasound in Med-
icine, and the Society for Maternal-Fetal Medicine (2017). Committee Opinion
No 700: Methods for Estimating the Due Date. Obstet. Gynecol. 129,
e150–e154.
Contrepois, K., Jiang, L., and Snyder, M. (2015). Optimized Analytical Proced-
ures for the Untargeted Metabolomic Profiling of Human Urine and Plasma by
Combining Hydrophilic Interaction (HILIC) and Reverse-Phase Liquid Chroma-
tography (RPLC)-Mass Spectrometry. Mol. Cell. Proteomics 14, 1684–1695.
Diczfalusy, E. (1953). Chorionic gonadotrophin and oestrogens in the human
placenta. Acta Endocrinol. Suppl. (Copenh.) 11, 1–175.
Donahue, S.M., Kleinman, K.P., Gillman, M.W., and Oken, E. (2010). Trends in
birth weight and gestational length among singleton term births in the United
States: 1990-2005. Obstet. Gynecol. 115, 357–364.
Dudzik, D., Zorawski, M., Skotnicki, M., Zarzycki, W., Kozlowska, G., Bibik-
Malinowska, K., Vallejo, M., Garcı́a, A., Barbas, C., and Ramos, M.P. (2014).
Metabolic fingerprint of Gestational Diabetes Mellitus. J. Proteomics
103, 57–71.
Edwin, S.S., Mitchell, M.D., and Dudley, D.J. (1997). Action of immunoregula-
tory agents on 5-HETE production by cultured human amnion cells. J. Reprod.
Immunol. 36, 111–121.
Epskamp, S., and Fried, E.I. (2018). A tutorial on regularized partial correlation
networks. Psychol. Methods 23, 617–634.
Frandsen, V.A., and Stakemann, G. (1961). The site of production of oestro-
genic hormones in human pregnancy. Hormone excretion in pregnancy with
anencephalic foetus. Acta Endocrinol. (Copenh.) 38, 383–391.
Gagnon, A., and Wilson, R.D.; Society of Obstetricians and Gynaecologists of
Canada Genetics Committee (2008). Obstetrical complications associated
with abnormal maternal serum markers analytes. Journal d’obstetrique et gy-
necologie du Canada 30, 918–932.
Goeman, J.J., and Bühlmann, P. (2007). Analyzing gene expression data in
terms of gene sets: methodological issues. Bioinformatics 23, 980–987.
Goeman, J.J., van de Geer, S.A., de Kort, F., and van Houwelingen, H.C.
(2004). A global test for groups of genes: testing association with a clinical
outcome. Bioinformatics 20, 93–99.
Hermida, R.C., Ayala, D.E., Mojón, A., Fernández, J.R., Silva, I., Ucieda, R.,
and Iglesias, M. (1997). High sensitivity test for the early diagnosis of gesta-
tional hypertension and preeclampsia. IV. Early detection of gestational hyper-
tension and preeclampsia by the computation of a hyperbaric index. J. Perinat.
Med. 25, 254–273.
Hochreiter, S., and Schmidhuber, J. (1997). Long short-term memory. Neural
Comput. 9, 1735–1780.
Hoppel, C.L., and Genuth, S.M. (1980). Carnitine metabolism in normal-weight
and obese human subjects during fasting. Am. J. Physiol. 238, E409–E415.
Hosie, A.M., Wilkins, M.E., da Silva, H.M., and Smart, T.G. (2006). Endogenous
neurosteroids regulate GABAA receptors through two discrete transmem-
brane sites. Nature 444, 486–489.
Cell 181, 1680–1692, June 25, 2020 1691
 ll
Kaddurah-Daouk, R., Kristal, B.S., and Weinshilboum, R.M. (2008). Metabolo-
mics: a global biochemical approach to drug response and disease. Annu.
Rev. Pharmacol. Toxicol. 48, 653–683.
Kenny, L.C., Broadhurst, D.I., Dunn, W., Brown, M., North, R.A., McCowan, L.,
Roberts, C., Cooper, G.J., Kell, D.B., and Baker, P.N.; Screening for Preg-
nancy Endpoints Consortium (2010). Robust early pregnancy prediction of
later preeclampsia using metabolomic biomarkers. Hypertension 56, 741–749.
King, J.C. (2000). Physiology of pregnancy and nutrient metabolism. Am. J.
Clin. Nutr. 71 (5, Suppl), 1218S–1225S.
Knutti, R., Rothweiler, H., and Schlatter, C. (1981). Effect of pregnancy on the
pharmacokinetics of caffeine. Eur. J. Clin. Pharmacol. 21, 121–126.
Koh, W., Pan, W., Gawad, C., Fan, H.C., Kerchner, G.A., Wyss-Coray, T., Blu-
menfeld, Y.J., El-Sayed, Y.Y., and Quake, S.R. (2014). Noninvasive in vivo
monitoring of tissue-specific global gene expression in humans. Proc. Natl.
Acad. Sci. USA 111, 7361–7366.
Levitz, M., Kadner, S., and Young, B.K. (1984). Intermediary metabolism of es-
triol in pregnancy. J. Steroid Biochem. 20 (4B), 971–974.
López-Hernández, Y., Herrera-Van Oostdam, A.S., Toro-Ortiz, J.C., López,
J.A., Salgado-Bustamante, M., Murgu, M., and Torres-Torres, L.M. (2019). Uri-
nary Metabolites Altered during the Third Trimester in Pregnancies Compli-
cated by Gestational Diabetes Mellitus: Relationship with Potential Upcoming
Metabolic Disorders. Int. J. Mol. Sci. 20, 1186.
Maltepe, E., and Fisher, S.J. (2015). Placenta: the forgotten organ. Annu. Rev.
Cell Dev. Biol. 31, 523–552.
Martin, J.A., Hamilton, B.E., Osterman, M.J.K., Driscoll, A.K., and Drake, P.
(2018). Births: Final Data for 2017. Natl. Vital Stat. Rep. 67, 1–50.
Maslova, E., Rifas-Shiman, S.L., Olsen, S.F., Gillman, M.W., and Oken, E.
(2018). Prenatal n-3 long-chain fatty acid status and offspring metabolic health
in early and mid-childhood: results from Project Viva. Nutr. Diabetes 8, 29.
Mayr, A., Klambauer, G., Unterthiner, T., Steijaert, M., Wegner, J.K., Ceule-
mans, H., Clevert, D.A., and Hochreiter, S. (2018). Large-scale comparison
of machine learning methods for drug target prediction on ChEMBL. Chem.
Sci. (Camb.) 9, 5441–5451.
Mendelson, C.R. (2009). Minireview: fetal-maternal hormonal signaling in preg-
nancy and labor. Mol. Endocrinol. 23, 947–954.
Minkler, P.E., Kerner, J., North, K.N., and Hoppel, C.L. (2005). Quantitation of
long-chain acylcarnitines by HPLC/fluorescence detection: application to
plasma and tissue specimens from patients with carnitine palmitoyltransfer-
ase-II deficiency. Clin. Chim. Acta 352, 81–92.
Moore, S.A., Hurt, E., Yoder, E., Sprecher, H., and Spector, A.A. (1995). Doco-
sahexaenoic acid synthesis in human skin fibroblasts involves peroxisomal
retroconversion of tetracosahexaenoic acid. J. Lipid Res. 36, 2433–2443.
GBD 2013 Mortality and Causes of Death Collaborators (2015). Global,
regional, and national age-sex specific all-cause and cause-specific mortality
for 240 causes of death, 1990-2013: a systematic analysis for the Global
Burden of Disease Study 2013. Lancet 385, 117–171.
Murphy, V.E., Smith, R., Giles, W.B., and Clifton, V.L. (2006). Endocrine regu-
lation of human fetal growth: the role of the mother, placenta, and fetus. En-
docr. Rev. 27, 141–169.
Ngo, T.T.M., Moufarrej, M.N., Rasmussen, M.H., Camunas-Soler, J., Pan, W.,
Okamoto, J., Neff, N.F., Liu, K., Wong, R.J., Downes, K., et al. (2018). Nonin-
vasive blood tests for fetal development predict gestational age and preterm
delivery. Science 360, 1133–1136.
Pearson, T., Zhang, J., Arya, P., Warren, A.Y., Ortori, C., Fakis, A., Khan, R.N.,
and Barrett, D.A. (2010). Measurement of vasoactive metabolites (hydroxyei-
cosatetraenoic and epoxyeicosatrienoic acids) in uterine tissues of normal
and compromised human pregnancy. J. Hypertens. 28, 2429–2437.
Prentice, A.M., and Goldberg, G.R. (2000). Energy adaptations in human preg-
nancy: limits and long-term consequences. Am. J. Clin. Nutr. 71 (5, Suppl),
1226S–1232S.
1692 Cell 181, 1680–1692, June 25, 2020
Resource
Raeside, J.I. (2017). A Brief Account of the Discovery of the Fetal/Placental
Unit for Estrogen Production in Equine and Human Pregnancies: Relation to
Human Medicine. Yale J. Biol. Med. 90, 449–461.
Reddy, D.S. (2003). Is there a physiological role for the neurosteroid THDOC in
stress-sensitive conditions? Trends Pharmacol. Sci. 24, 103–106.
Romero, R., Mazaki-Tovi, S., Vaisbuch, E., Kusanovic, J.P., Chaiworapongsa,
T., Gomez, R., Nien, J.K., Yoon, B.H., Mazor, M., Luo, J., et al. (2010). Metab-
olomics in premature labor: a novel approach to identify patients at risk for pre-
term delivery. The journal of maternal-fetal & neonatal medicine 23,
1344–1359.
Sachse, D., Sletner, L., Mørkrid, K., Jenum, A.K., Birkeland, K.I., Rise, F., Pieh-
ler, A.P., and Berg, J.P. (2012). Metabolic changes in urine during and after
pregnancy in a large, multiethnic population-based cohort study of gestational
diabetes. PLoS ONE 7, e52399.
Say, L., Chou, D., Gemmill, A., Tunçalp, Ö., Moller, A.B., Daniels, J., Gülmezo-
glu, A.M., Temmerman, M., and Alkema, L. (2014). Global causes of maternal
death: a WHO systematic analysis. Lancet Glob. Health 2, e323–e333.
Sedgh, G., Singh, S., and Hussain, R. (2014). Intended and unintended preg-
nancies worldwide in 2012 and recent trends. Stud. Fam. Plann. 45, 301–314.
Sevastou, I., Kaffe, E., Mouratis, M.A., and Aidinis, V. (2013). Lysoglycero-
phospholipids in chronic inflammatory disorders: the PLA(2)/LPC and ATX/
LPA axes. Biochim. Biophys. Acta 1831, 42–60.
Shen, X., Wang, R., Xiong, X., Yin, Y., Cai, Y., Ma, Z., Liu, N., and Zhu, Z.J.
(2019). Metabolic reaction network-based recursive metabolite annotation
for untargeted metabolomics. Nat. Commun. 10, 1516.
Soldin, O.P., Guo, T., Weiderpass, E., Tractenberg, R.E., Hilakivi-Clarke, L.,
and Soldin, S.J. (2005). Steroid hormone levels in pregnancy and 1 year post-
partum using isotope dilution tandem mass spectrometry. Fertil. Steril. 84,
701–710.
Stein, S.E., and Scott, D.R. (1994). Optimization and testing of mass spectral
library search algorithms for compound identification. J. Am. Soc. Mass Spec-
trom. 5, 859–866.
Tulchinsky, D., Hobel, C.J., Yeager, E., and Marshall, J.R. (1972). Plasma
estrone, estradiol, estriol, progesterone, and 17-hydroxyprogesterone in hu-
man pregnancy. I. Normal pregnancy. Am. J. Obstet. Gynecol. 112,
1095–1100.
Tusher, V.G., Tibshirani, R., and Chu, G. (2001). Significance analysis of micro-
arrays applied to the ionizing radiation response. Proc. Natl. Acad. Sci. USA
98, 5116–5121.
Viant, M.R., Kurland, I.J., Jones, M.R., and Dunn, W.B. (2017). How close are
we to complete annotation of metabolomes? Curr. Opin. Chem. Biol.
36, 64–69.
Wang, M.H., Zand, B.A., Nasjletti, A., and Laniado-Schwartzman, M. (2002).
Renal 20-hydroxyeicosatetraenoic acid synthesis during pregnancy. Am. J.
Physiol. Regul. Integr. Comp. Physiol. 282, R383–R389.
Wang, X., Chen, C., Wang, L., Chen, D., Guang, W., and French, J. (2003).
Conception, early pregnancy loss, and time to clinical pregnancy: a popula-
tion-based prospective study. Fertil. Steril. 79, 577–584.
Wang, Q., Würtz, P., Auro, K., Mäkinen, V.P., Kangas, A.J., Soininen, P., Tiai-
nen, M., Tynkkynen, T., Jokelainen, J., Santalahti, K., et al. (2016). Metabolic
profiling of pregnancy: cross-sectional and longitudinal evidence. BMC Med.
14, 205.
Wu, C.C., Gupta, T., Garcia, V., Ding, Y., and Schwartzman, M.L. (2014). 20-
HETE and blood pressure regulation: clinical implications. Cardiol. Rev.
22, 1–12.
Xia, J., and Wishart, D.S. (2010a). MetPA: a web-based metabolomics tool for
pathway analysis and visualization. Bioinformatics 26, 2342–2344.
Xia, J., and Wishart, D.S. (2010b). MSEA: a web-based tool to identify biolog-
ically meaningful patterns in quantitative metabolomic data. Nucleic Acids
Res. 38, W71–W77.
 ll
Resource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Biological Samples
Plasma samples This paper N/A
Chemicals, Peptides, and Recombinant Proteins
MS-grade water Fischer Scientific Cat#7732-18-5
MS-grade methanol Fischer Scientific Cat#A456-500
MS-grade acetonitrile Fischer Scientific Cat#A9554
MS-grade acetone Fischer Scientific Cat#67-64-1
MS-grade acetic acid Sigma Aldrich Cat#64-19-7
Progesterone Sigma Aldrich Cat#P-069-1ML
THDOC Sigma Aldrich Cat#P2016-5MG
Estriol-16-glucuronide Sigma Aldrich Cat#E1877-10MG
DHEA-S Sigma Aldrich Cat#D-066-1ML
PE(P-16:0e/0:0) Avanti Polar Lipids Cat#852470
Androstane-3,17-diol Sigma Aldrich Cat#A7755-100MG
17a-Hydroxyprogesterone Sigma Aldrich Cat#H-085-1ML
Deposited Data
Raw data This paper https://www.metabolomicsworkbench.org;
Project ID PR000918; Project https://
doi.org/10.21228/M81H58.
Software and Algorithms
Progenesis QI Software Nonlinear Dynamics http://www.nonlinear.com/progenesis/qi/
ProteoWizard version 3.0.19095- Chambers et al., 2012 http://proteowizard.sourceforge.net
938eda31a
the k-Nearest Neighbor algorithm Altman, 1992 N/A
Forward Dot–Product algorithm Stein and Scott, 1994 N/A
MetDNA Shen et al., 2019 http://metdna.zhulab.cn/
MetaboAnalystR Chong et al., 2018; https://github.com/xia-lab/
Xia and Wishart, 2010a; MetaboAnalystR
Xia and Wishart, 2010b
R R Core Team https://www.R-project.org
MS/MS identification pipeline This paper https://jaspershen.github.io/metID/
index.html
Other
Zorbax SB columns (2.1 X 50mm, 1.8 Agilent Technologies 827700-914
Micron, 600 Bar)

RESOURCE AVAILABILITY

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Mike
Snyder (mpsnyder@stanford.edu).

Materials Availability
This study did not generate new unique reagents.

Cell 181, 1680–1692.e1–e5, June 25, 2020 e1

 ll
Resource

Data and Code Availability

Original data have been deposited to the NIH Common Fund’s National Metabolomics Data Repository (NMDR) website (supported
by NIH grant U2C-DK119886), the Metabolomics Workbench https://www.metabolomicsworkbench.org, Project ID PR000918.
https://doi.org/10.21228/M81H58.
The code for the MS/MS identification pipeline used is available on github at https://jaspershen.github.io/metID/index.html https://
github.com/jaspershen/metID.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Pregnancy cohort
We recruited pregnant women through family doctors and advertisements (Danish IRB number H-3-2014-004). At enrollment, all
women were screened to ensure that they were healthy at baseline, without chronic conditions, and without medication intake of
any kind (ages 23 to 36 at giving birth). From each woman, non-fasting blood samples were collected weekly during pregnancy
and one sample was collected after pregnancy (2x9 mL EDTA tube and 1xPaxGene RNA tube).

METHOD DETAILS

Plasma sample preparation

Within Discovery and Test Set 1 cohorts, 784 normal pregnancy samples from 30 women were completely randomized and analyzed
in 12 batches across two years. The 32 normal pregnancy samples in Test Set 2 were randomized and analyzed three years later.
Plasma was prepared from whole blood treated with anti-clot EDTA, aliquoted, and stored at
80 C. Plasma (200 mL) was treated
with four volumes (800 mL) of an acetone:acetonitrile:methanol (1:1:1, v/v) solvent mixture with internal standards (i.e., Acetyl-d3-
carnitine, Phenylalanine-3,3-d2, Tiapride, Trazodone, Reserpine, Phytosphingosine, and Chlorpromazine), mixed for 15 min at
4 C, and incubated at
20 C for 2 h to allow for protein precipitation. The supernatant was collected after centrifugation at
10,000 rpm for 10 min at 4 C and evaporated under nitrogen to dryness (Biotage Turbovap). The dry extracts were reconstituted
with 200 mL 1:1 methanol:water before analysis. A quality control (QC) sample was generated by pooling all the plasma samples
from 10 women and injected between every 10–15 sample injections to monitor the consistency of the retention time and the signal
intensity. The QC sample was also diluted two, four, and eight times to determine the linear-dilution effect of metabolic features.

Chemical materials for untargeted metabolomics

MS-grade water (7732-18-5), methanol (A456-500), acetonitrile (A9554), and acetone (67-64-1) were purchased from Fischer Scien-
tific (Morris Plains, NJ, USA). MS-grade acetic acid (64-19-7) was purchased from Sigma Aldrich (St. Louis, MO, USA). Analytical
grade chemical standards were purchased [Progesterone (Sigma-Aldrich, P-069-1ML), THDOC (Sigma-Aldrich, P2016-5MG), Es-
triol-16-Glucuronide (Sigma-Aldrich, E1877-10MG), DHEA-S (Sigma-Aldrich, Dehydroepiandrosterone-D5-3-sulfate (DHEAS-D5)
(2,2,3,4,4,-D5) sodium salt solution, D-066-1ML), PE(P-16:0e/0:0) (Avanti Polar lipids, 852470), Androstane-3,17-diol (Sigma-Aldrich,
A7755-100MG), 17a-Hydroxyprogesterone (Sigma-Aldrich, H-085-1ML)] and prepared in methanol, except PE(P-16:0e/0:0), which
was prepared in chloroform/methanol (8:2).

MS acquisition
Metabolic extracts were analyzed by reversed-phase liquid chromatographic (RPLC)-mass spectrometry (MS) in both positive and
negative ionization modes. Thermo Q Exactive Hybrid Quadrupole-Orbitrap plus and Q Exactive mass spectrometers (Xcalibur,
Thermo Scientific, San Jose, CA, USA) were operated in full MS-scan mode for data acquisition (acquisition from m/z 500 to
2,000) with a scan rate of approximately 4 Hz and a resolution set at 30,000 (at m/z 400). The MS/MS spectra of the QC sample
were acquired under different fragmentation energy (25 NCE and 50 NCE) of the top 10 parent ions. The resulting mass spectra
were exported into Progenesis QI Software (Nonlinear Dynamics, Durham, NC, USA) for further processing.

Chromatographic conditions
RPLC separation was performed using Zorbax SB columns (2.1 X 50mm, 1.8 Micron, 600 Bar; 827700-914) purchased from Agilent
Technologies (Santa Clara, CA, USA). Mobile phases for RPLC consisted of 0.06% acetic acid in water (phase A) and 0.06% acetic
acid in MeOH (phase B). Metabolites were eluted from the column at a flow rate of 0.6 mL/min, leading to a backpressure of 220–
280 bar at 99% phase A. A linear 1%–80% phase B gradient was applied over 9–10 min. The oven temperature was set to 60 C,
and the sample injection volume was 5 mL.

QUANTIFICATION AND STATISTICAL ANALYSIS

Section 1: Metabolomics Data Processing

Metabolomic features were extracted with a unique mass/charge ratio and retention time, then aligned and quantified with the
Progenesis QI software (Nonlinear Dynamics, Durham, NC, USA, http://www.nonlinear.com/progenesis/qi/). Peak deconvolution

e2 Cell 181, 1680–1692.e1–e5, June 25, 2020

 ll
Resource

was performed under default settings in Progenesis QI. Acquired data were processed using an analysis pipeline written in R (https://
www.R-project.org). Progenesis QI output was then processed by removing all metabolites that were quantified in less than 30% of
the samples or had a median intensity of less than twofold signal over the noise threshold (S/N < 2). The noise threshold was esti-
mated by using the median signal across all the blank runs (if no quantitation was reported in any of the blank runs, the feature
was also included in the analysis, as it likely had good S/N characteristics). Then the data were log-transformed and normalized.
For each run, the median of all features was centered to correct for variation in the sample amount. Then for each analyte, a linear
correction was applied per batch to correct for any linear decrease or increase in abundance during the acquisition of a batch. In
short, for each analyte and each batch, a linear model was fitted with the log-abundance of the analyte as the dependent variable
and the acquisition number [run order (randomized)] as the independent variable. The model prediction was interpreted as an under-
lying drift in mass spectrometric sensitivity and subtracted from the analyte level to yield within-batch normalized abundances.
Finally, for each analyte, the abundances were median centered by batch to correct for sensitivity differences between batches.
The positive- and negative-mode features were then concatenated for downstream analysis. In total, 9,651 features were included
in the final analysis. In addition, for samples with more than 50% of the values missing, the sample was removed (one sample in total).
The remaining missing values were imputed by the nearest 10 neighbors using the k-Nearest Neighbor algorithm (Altman, 1992). Note
that Discovery and Test Set 1 were normalized together, while samples of Test Set 2 were normalized independently.
We applied principal component analysis (PCA) to examine the overall distribution of the sample data (with all 9,651 features) and
check the run quality. The gestational ages (based on first-trimester ultrasound measurements) were superimposed to facilitate the
analysis. During the analysis, the vast majority of the samples were separated by pre- and postpartum in PCA space defined by two
components, which explained the largest variations (PC1 and 2, Figure 1B), while two samples of a same subject (last two in her
collection, before and after childbirth) displayed irregular behavior in PCA and unsupervised clustering analysis. The two samples
were treated as outliers and excluded from further analysis. We also performed partial least-squares discriminant analysis (PLS-
DA) according to the categories of gestational age (by the mixOmics package).

Section 2: Metabolic Features Identification

Metabolite identification was performed using a two-step approach. First, to identify compounds, we used our in-house metabolite
library, which contains chemical standards and a manually curated compound list based on accurate mass (m/z, ± 5 ppm), retention
time and spectral patterns. Second, further metabolites were identified based on accurate mass, isotope pattern and MS/MS spectra
against public databases, including HMDB, MoNA, MassBank, METLIN, and NIST.
Specifically, tandem mass spectrometry (MS/MS) data of QC samples were acquired using a Thermo Q Exactive plus mass spec-
trometers. The raw MS data (.raw format) were converted to .mgf format files using ProteoWizard (Chambers et al., 2012) (Version
3.0.19095-938eda31a, http://proteowizard.sourceforge.net). Using the metabolic features table (from Waters Progenesis QI) and QC
MS/MS data (.mgf format), the metabolic features and MS/MS spectra were matched according to their accurate masses (±25 ppm),
and RT values (±30 s) (Shen et al., 2019). If one metabolic feature matched multiple MS/MS spectra, then all matched MS/MS spectra
were used for the identification.
Next, the generated MS1/MS2 pairs were automatically searched in the public databases: HMDB (http://www.hmdb.ca/), MoNA
(http://mona.fiehnlab.ucdavis.edu/), and MassBank (http://www.massbank.jp/). The MS/MS spectra similarity score was calculated
using the forward dot-product algorithm (Stein and Scott, 1994), which considers both fragments and intensities. The similarity score
cutoff was set as 0.5.
Furthermore, the metabolic features with MS/MS spectra and not matched in download public databases were searched in the
online public databases, METLIN (https://metlin.scripps.edu) and NIST (https://www.nist.gov/). Then the MS/MS spectra match
was manually checked to confirm the identifications, which was considered a level 2 identification according to MSI (Viant et al.,
2017). In addition, the metabolic peaks with MS/MS spectra that were not matched in public databases were analyzed by MetDNA
(Shen et al., 2019) and given a MSI level 4 identification.
Finally, predictors from the machine-learning models were further confirmed with chemical standards by matching the accurate
masses (±5 ppm), retention time (±30 s), and the MS/MS spectra for a MSI level 1 identification (Viant et al., 2017).
In the rare cases, when a given metabolic feature was matched differently between different matching methods, we choose the
matching based on the identification level: standards > MS/MS > MetDNA.

Section 3: Identify Significantly Altered Features/Compounds

A statistical method specialized for multi-testing, SAM (Significance Analysis of Microarrays) (Tusher et al., 2001) was applied to iden-
tify metabolic features/compounds altered significantly in metabolome-wide analysis. Specifically, we used SAM to examine the cor-
relation between abundance of each compounds and the gestational age of each sample in Discovery and Test Set 1 cohorts. For all
SAM analyses, distribution-independent ranking tests (based on the Wilcoxon test) and the sample-wise permutation (default by the
samr package) were used to ascertain significance (false discovery rate, FDR < 0.05). The adjusted gestational ages were included in
a number of plots to present the changes in metabolites among individuals, which were calculated by scaling all delivery event timing
to 40 weeks. The populational baseline was calculated by taking the mean intensity values of all women with samples before 14 (20
out of 30 women).

Cell 181, 1680–1692.e1–e5, June 25, 2020 e3

 ll
Resource

To identify top changed compounds with abundance increases or decreases more than 50% during the whole pregnancy
(40 weeks), we performed a linear regression between log2 abundance and the gestational weeks of samples, and only those com-
pounds with absolute slope larger than log2(1.5)/40 weeks = 0.015 were chosen.

Section 4: Regularized Partial Correlation Network

The regularized partial correlation network captures the remaining association between two nodes after controlling all other informa-
tion (indirect correlations) in the network (Epskamp and Fried, 2018). Namely, each node represents a compound, and each edge
represents the strength of partial correlation between two nodes after conditioning on all other variables in the datasets. Edge weights
represent the partial correlation coefficeients. Lasso (least absolute shrinkage and selection operator) was used to shrink small
association coefficient to zero and thus limit spurious correlations in the network. To perform the lasso-based regularized partial cor-
relation, we used qgraph package in R. The tuning parameter gamma(g), which controls the complexity of the network, was set to 0.5
as suggested (Epskamp and Fried, 2018). Three measures, strength (the sum of absolute edge weight connected to each metabolite),
closeness (inverse of the sum of distances from one metabolite to all others), and betweenness (how often one metabolite is in the
shortest paths between other metabolites), indicated how important metabolites are in the network.

Section 5: Pathway Analysis

The compounds identified by the methods mentioned above were pooled together. We utilized MetaboAnalystR (Chong et al., 2018)
(https://github.com/xia-lab/MetaboAnalystR) to perform the metabolite set enrichment analysis (MSEA) (Xia and Wishart, 2010b) as
well as metabolic pathway analysis (MetPA) (Xia and Wishart, 2010a) on all identified metabolites. For the potential location/organ
analysis on metabolites, we excluded male organ/cell types for MSEA. To quantify pathway activity, we averaged the intensities
of all identified metabolites for each pathway that includes no less than three identified metabolites and plotted them on the heatmap
(Figure 3B). The pathway activity before 14 weeks were averaged across all available samples and subtracted from all later time
points. The statistical significance of the changes in a pathway’s activity across pregnancy was evaluated by global testing (Goeman
and Bühlmann, 2007), the default method used by MetaboAnalystR. The topological pathway impacts were quantified using
published method (Xia and Wishart, 2010a), with MetaboAnalystR. Human desease states that correlated with pregnancy-related
metabolites were calculated based on published metabolomics data (Chong et al., 2018).

Section 6: Machine Learning for Pregnancy Timing

Three cohorts of data collected and run at different years but from the same center were used to establish Discovery (subjects N = 21,
samples n = 507), Test Set 1 (subjects N = 9, samples n = 245), and Test Set 2 (subjects N = 8, samples n = 32) datasets, excluding
non-pregnant (postpartum) samples. We applied lasso (R package: glmnet) in the Discovery dataset to select compounds/metabolic
features to build the linear regression model to predict gestational age. A 10-fold cross validation was performed to choose optimal
lambda (penalty for the number of features), which determines the performance of the lasso model (number of features included in the
model and prediction deviations). For the practical utility of a signature in potential clinical settings, in the identified compound-pre-
diction models, if the number of predictors exceeded five under a given optimal lambda, we increased the lambda value so that the
number of predictors is no more than five in the final models. We then used two different methods to evaluate the prediction deviation
of our lasso model produced under a given lambda value: 1) A 10-fold cross validation within the Discovery cohort, in which the
optimal lambda was used (Baumann and Baumann, 2014; Mayr et al., 2018). In the CV, samples were distributed into folds by subject
instead of by samples to prevent person-specific information cross-over between the training folds and the test fold. 2) Validation
tests in the separate Test Set 1 and Test Set 2 cohorts, in which independent subjects were included and samples were analyzed
one year and three years from the Discovery cohort. We built the model using the optimized lambda and full discovery datasets.
This model was applied to the validation cohort for prediction and verification. A linear fitting from the two above evaluations
were performed, between the predicted value and the actual values, with Pearson correlation coefficient (R), R2, and RMSE reported.
The contribution of each predictor (metabolite) in each prediction model is defined by:
Contribution = absðcoefficienti Þ=sumðabsðcoefficienti ÞÞ
i: the metabolite included in the linear model
Unlike cross-validation in the discovery dataset, validation tests are not prone to hyperparametric selection bias for lambda value.
Since samples from Test Set 2 cohort were normalized independently from other samples, a scaling was done in the end. Note that
since the sequential nature of the data were not used in the machine-learning methods, other statistical tools, such as recurrent neu-
ral networks (e.g., LSTM (Hochreiter and Schmidhuber, 1997; Mayr et al., 2018)), may be explored to improve the model.
For samples collected after 28 weeks (third-trimester samples), we started with 264 level 1 and 2 compounds, and we used a
similar discovery and validation pipeline described for predicting gestational age (above) to build logistic regression models predict-
ing the categorical labels of gestational age > 32 and 37 weeks or delivery within 2, 4, 8 weeks. The prediction models for > 20, 24, and
28 weeks were built using samples from all three trimesters. For the prediction on delivery within 2, 4, and 8 weeks, only the 18 women
(out of 30) with natural labor onset were included, excluding subjects with induction before labor onset and scheduled cesarean-sec-
tion (induction by oxytocin/membrane strip after the onset is allowed). To estimate the confidence interval for each AUROC, we per-
formed bootstrapping by person (instead of by samples) for 1000 times, and calculate the 95% confidence interval for AUROC.

e4 Cell 181, 1680–1692.e1–e5, June 25, 2020

 ll
Resource

Section 7: Analyze the discrepancies between metabolic clock (GA prediction model) and first-trimester ultrasound
estimations
We evaluated the individual correlation between the predictions made by the metabolic clock and the estimations from first-trimester
ultrasound: We first examined the correlation between metabolic clock predictions and the gestational age based on first-trimester
ultrasound in individual persons. Each correlation was evaluated by Pearson’s correlation. We then performed meta-analysis across
the persons to generate a summary p value, using Fisher’s method, to describe the overall correlation in each cohort (cross-validation
in the Discovery, Independent validation of Test Set 1).
Previous literature (Donahue et al., 2010) and our own observations suggest that birth weight and gestational length are positively
correlated; later delivery is associated with a heavier absolute birth weight of an infant. To determine whether an infant’s birth weight
falls above or below the group mean, we performed a linear regression between the two parameters and took the residuals to repre-
sent the birth weight deviation adjusted for delivery timing.
Average D(GAmetabolic- GAultrasound): For each person, at each time point, we examined the differences between the metabolic clock
and first-trimester ultrasound estimation of gestational age. These values were averaged for each person to represent the overall
relative pace of metabolic clock compared to the first-trimester ultrasound estimation. We then examined the correlation between
delivery timing adjusted birth weights and average D(GAmetabolomic- GAultrasound) (Figure 5C).
To examine whether an accelerated metabolic clock (compared to the first-trimester ultrasound estimation) associates with
advanced delivery, we performed the correlation between average D(GAmetabolomic- GAultrasound) and delivery timing, only in women
with a natural labor onset (Figure 5D).

Cell 181, 1680–1692.e1–e5, June 25, 2020 e5

 ll
Resource

Supplemental Figures
A Postpartum samples B
Pregnancy samples
Child birth event

250 350 450

Variances
150
1 2 3 4 5 6 7 8 9 10
Component number
Subjects

C
PLS−DA
GA:
Under 10

Component 2 (2%)
50
10−20
20−30
25 Over 30
PP
0

−25
−30 0 30
Component 1 (5%)

5 14 28 37 40
Gestational age (weeks)

D E
60 60
Batches
Discovery
30 30
Validation
PC2 (3.9 %)

PC2 (3.9 %)

0 0

−30 −30

−60 −30 0 30 60 −60 −30 0 30 60

PC1 (5.1%) PC1 (5.1%)

F G

1000
log10(Frequency)

Log2(Intensity)

10

0.1 0.2
−0.1 0.0
Linear fitting slope
ln(Intensity) ~ Gestational age

Gestational age (weeks)

Figure S1. Untargeted Metabolomics for Longitudinal Pregnancy Samples, Related to Figure 1
(A) High-density longitudinal sampling of pregnancies.
(B) The Scree plot of the principal component analysis.
(C) The PLS-DA result according to the categories of gestational age. GA: gestational age; PP: postpartum.
(legend continued on next page)
 ll
Resource

(D and E) Principal component analysis based on all 9,651 features shows that the samples do not separate according to the 30 subjects (D) samples from
individual subjects are represented by different colors or experimental batches of Discovery and Validation (Test Set 1) analyzed across two different years (E)
samples of the discovery cohort are presented in red; samples of the validation cohort (Test Set 1) are presented in blue.
(F) Histogram shows the distribution of slopes in the linear fitting model of the 9,651 features (intensities against the gestational ages).
(G) For each of the 30 women, the intensities of an example metabolic feature are shown over the course of gestation, which reveals consistent increases in
abundance according to gestational age among 30 subjects, despite individual differences.
 ll
Resource

A
Pearson correlation coefficient
1 0.5 0 −0.5 −1

Taurochenodeoxycholate
Cyclo(leucylprolyl)
Theophylline
Caffeine Pregnancy alteration
Theobromine
1−Methylxanthine
Pregnenolone sulfate
Decrease
Estrone 3−sulfate
Corticosterone
Increase
Cortisone
17,18−EpETE
Cortisol
3−Acetoxypyridine Group
N−Acetyl−D−glucosamine
Androstane−3,17−diol Amino acid metabolism
5−Pregnane−3,17−diol−20−one 3−sulfate
Estriol−16−Glucuronide
Bile acid biosynthesis
Progesterone Caffeine metabolism
17alpha−Hydroxyprogesterone Fatty acid metabolism
THDOC
7−Methylguanine Phospholipid metabolism
7alpha,24−Dihydroxy−4−cholesten−3−one
Hexadecadienoylcarnitine Steroid hormone biosynthesis
Dodecanoylcarnitine Others
C16 PAF (Platelet−activating factor)
Tetracosahexaenoic acid
Tetracosapentaenoic acid
Sphingosine
Tetracosatetraenoic acid
Docosadienoic acid
Erucic acid
Androsterone sulfate
Glycochenodeoxycholate
Sinapyl alcohol
Isobutyryl−L−carnitine
8,9−DHET
2−Phenylbutyric acid
DHEA-S
LPC(18:2)
LPC(17:0)
PE(P-16:0e/0:0)
LPE(22:2)
PC(22:1/22:1) (Lecithin)
LPE(22:4)
LPE(20:3)
LPC(24:0)
LPE(22:1)
LPC(P−18:0)
PC(18:1(9Z)e/2:0)
LPE(20:0)
LPE(20:1)
MG(20:0)
MG(22:2)
MG(18:1)
Tricosanoic acid
MG(24:1)
MG(24:0)
MG(14:1)
Oleoylcarnitine
3−Hydroxyoleylcarnitine
LPC(20:5)
LPC(P−18:1)
LPC(P−16:0)
Hydroxybupropion
Ketoisovaleric acid
Valylhistidine
Glycyrrhetinic acid
beta−Glycyrrhetinic acid
Taurochenodeoxycholate
Cyclo(leucylprolyl)
Theophylline
Caffeine
Theobromine
1−Methylxanthine
Pregnenolone sulfate
Estrone 3−sulfate
Corticosterone
Cortisone
17,18−EpETE
Cortisol
3−Acetoxypyridine
N−Acetyl−D−glucosamine
Androstane−3,17−diol
5−Pregnane−3,17−diol−20−one 3−sulfate
Estriol−16−Glucuronide
Progesterone
17alpha−Hydroxyprogesterone
THDOC
7−Methylguanine
7alpha,24−Dihydroxy−4−cholesten−3−one
Hexadecadienoylcarnitine
Dodecanoylcarnitine
C16 PAF (Platelet−activating factor)
Tetracosahexaenoic acid
Tetracosapentaenoic acid
Sphingosine
Tetracosatetraenoic acid
Docosadienoic acid
Erucic acid
Androsterone sulfate
Glycochenodeoxycholate
Sinapyl alcohol
Isobutyryl−L−carnitine
8,9−DHET
2−Phenylbutyric acid
DHEA-S
LPC(18:2)
LPC(17:0)
PE(P-16:0e/0:0)
LPE(22:2)
PC(22:1/22:1) (Lecithin)
LPE(22:4)
LPE(20:3)
LPC(24:0)
LPE(22:1)
LPC(P−18:0)
PC(18:1(9Z)e/2:0)
LPE(20:0)
LPE(20:1)
MG(20:0)
MG(22:2)
MG(18:1)
Tricosanoic acid
MG(24:1)
MG(24:0)
MG(14:1)
Oleoylcarnitine
3−Hydroxyoleylcarnitine
LPC(20:5)
LPC(P−18:1)
LPC(P−16:0)
Hydroxybupropion
Ketoisovaleric acid
Valylhistidine
Glycyrrhetinic acid
beta−Glycyrrhetinic acid

Pregnancy Alteration
Group

B
Strength Closeness Betweenness
Androsterone sulfate
Androstane−3,17−diol
THDOC
Dehydroisoandrosterone sulfate (DHEA−S)
PC(22:1/22:1) (Lecithin)
Estrone 3−sulfate
Estriol−16−Glucuronide
17alpha−Hydroxyprogesterone
LPC(20:5)
7−Methylguanine
Pregnenolone sulfate
LPC(18:2)
PC(18:1(9Z)e/2:0)
LPC(17:0)
3−Hydroxyoleylcarnitine
Taurochenodeoxycholate
LPE(20:1)
LPE(22:4)
Progesterone
Sphingosine
Tetracosatetraenoic acid
5−Pregnane−3,17−diol−20−one 3−sulfate
Cyclo(leucylprolyl)
LPE(20:3)
Valylhistidine
Cortisol
LPE(22:1)
Tetracosapentaenoic acid
Erucic acid
C16 PAF (Platelet−activating factor)
Cortisone
17,18−EpETE
Oleoylcarnitine
LPE(20:0)
Tricosanoic acid
Glycochenodeoxycholate
Tetracosahexaenoic acid
PE(P−16:0e/0:0)
N−Acetyl−D−glucosamine
Docosadienoic acid
LPC(P−18:0)
3−Acetoxypyridine
LPC(24:0)
Sinapyl alcohol
Hexadecadienoylcarnitine
MG(24:1)
MG(14:1)
Corticosterone
Caffeine
MG(24:0)
Dodecanoylcarnitine
7alpha,24−Dihydroxy−4−cholesten−3−one
beta−Glycyrrhetinic acid
Theophylline
LPC(P−18:1)
LPE(22:2)
1−Methylxanthine
MG(20:0)
Glycyrrhetinic acid
Theobromine
Isobutyryl−L−carnitine
MG(18:1)
Hydroxybupropion
MG(22:2)
2−Phenylbutyric acid
LPC(P−16:0)
−2 −1 0 1 2 −2 −1 0 1 2 −1 0 1 2 3

(legend on next page)

 ll
Resource

Figure S2. Functional Metabolite Groups Altered during Pregnancy, Related to Figure 2
(A) Correlation matrix colored by the Pearson correlation coefficient of each pair of pregnancy-related compounds across samples.
(B) The strength, closeness, and betweenness of metabolites in the regularized partial correlation network indicate how important the metabolites are in the
network. Metabolite names are listed on the left side ranked by the closeness, with the names of the seven compounds in the prediction models of Figure 4 and
Figure 5 (bold).
 ll
Resource

Figure S3. Pregnancy-Related Metabolic Pathways and Metabolite Origin Analysis, Related to Figure 3
(A) Steroid hormone biosynthesis pathway, with metabolite increases (in red) or decreases (in blue) over the course of gestation.
(B) Numerous metabolites in plasma that were altered during pregnancy can be traced back to organs by metabolite set enrichment analysis (MSEA).
(C) Arachidonic acid metabolism pathway, with metabolite increases (in red) or decreases (in blue) over the course of gestation.
(D) The average levels of the 20-HETE and 5-HETE changes against the gestational progression. The intensities were normalized to the baseline, which was
defined by averaging all samples before 14 weeks. The standard errors, derived from 30 subjects, are shown. The gestational ages were adjusted by scaling
delivery events to 40 weeks. PP, postpartum.
 ll
Resource

A Model reduction
256
Number of predictors

32

4 6 8
Cross validation: RMSE

B C
Discovery Test Set 1
2
40 2
R = 0.93 R = 0.91
GAmetabolic (weeks)

40
P< 1X10-100 N= 245
N= 507
30 30

20 20

10 10

10 20 30 40 10 20 30 40
GAultrasound (weeks) GAultrasound (weeks)

D Model reduction
256
Number of predictors

32

3 4 5 6 7 8 9
Cross validation: RMSE

E F

1.0 1.0
m/z: 438.2974 m/z: 367.1583
RT error (second): 0.6 RT error (second): 9

0.5 0.5
Relative intensity
Relative intensity

0.0 0.0

-0.5 -0.5

O
O
H3 C P NH 2
O O
OH
-1.0 Standard: PE(P-16:0e/0:0) HO H -1.0 Standard: DHEA-S(Dehydroepiandrosterone sulfate)
100 200 300 400 100 200 300 400
Mass to charge ratio (m/z) Mass to charge ratio (m/z)

Figure S4. Metabolites Predict Gestational Age in Machine-Learning Models, Related to Figure 4
(A) Feature selection for predicting gestational age (GA) using metabolomic features.
(B and C) GA predicted by metabolic features (GAmetabolic, y axis) highly correlates with clinical values determined by standard of care (by first-trimester ultra-
sound, GAultrasound, x axis) in the Discovery (B) and the validation cohort (Test Set 1) (C). The 95% confidence interval for the linear regression is represented by the
gray area.
(D) Feature selection for predicting GA using identified metabolites.
(E and F) Measured MS/MS fragmentation profiles (upper) matching of PE(P-16:0e/0:0) (E) and DHEA-S (F) with the MS/MS of standard compounds (lower). GA,
gestational age.
 ll
Resource

Child birth
Gestational age (weeks): 20 24 28 32 37
AUC in validation: 0.98 0.98 0.97 0.89 0.87
THDOC 2 1 1 2
Progesterone 2 1 2
Androstane-3,17-diol 3
Estriol-16-glucuronide 1 3 2 1

B Gestational age (GA) > 37w C D

1.00 GA predictors
Estriol−16−Glucuronide
GA < 37w
True positive rate

0.75 0.4
GA > 37w

Contribution

log2(Intensity)
12
0.50 0.2
AUC: 0.91
Discovery N= 222 0.0 11
0.25 95% CI: 0.87-0.96
ne TH e

ol
,1 C
id

di
−3 DO
AUC: 0.87
n

7−
ro

10
cu

Test Set 1 N= 98
lu
G

0.00 95% CI: 0.79-0.93

6−

ta
−1

os

0.00 0.25 0.50 0.75 1.00

28

29

30
22

23

24

25

26

27
ol

dr
tr i

An
Es

False positive rate Subject ID

E THDOC GA < 37w F Androstane-3,17-diol GA < 37w

GA > 37w GA > 37w
13.5
13.5
13.0
log2(Intensity)

13.0
12.5
12.5
12.0

11.5 12.0
28

29

30

28

29

30
22

23

24

25

26

27

22

23

24

25

26

27
Subject ID Subject ID
G H
Prediction: Gestational age (GA) > 20w Prediction: Gestational age (GA) > 24w
1.00 1.00

0.75 0.75
True positive rate

True positive rate

0.50 0.50

AUC: 0.99
Discovery N= 507 AUC: 0.97
Discovery N= 507
95% CI: 0.97-0.99
0.25 0.25 95% CI: 0.96-0.98
AUC: 0.98 AUC: 0.98
Validation N= 245 Validation N= 245
95% CI: 0.96-0.99 95% CI: 0.97-0.99
0.00 0.00

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
False positive rate False positive rate

I Prediction: Gestational age (GA) > 28w J Prediction: Gestational age (GA) > 32w
1.00 1.00

0.75 0.75
True positive rate

True positive rate

0.50 0.50

AUC: 0.97 AUC: 0.90

Discovery N= 507 Discovery N= 222
0.25 0.25
95% CI: 0.96-0.98 95% CI: 0.84-0.96
AUC: 0.98 AUC: 0.89
Validation N= 245 Validation N= 98
0.00 95% CI: 0.97-0.99 0.00 95% CI: 0.83-0.95

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
False positive rate False positive rate

(legend on next page)

 ll
Resource

Figure S5. Metabolites Selected by Machine Learning Can Accurately Predict Gestational Age before or after 20, 24, 28, 32, and 37 Weeks in
Both the Discovery and Validation Cohort (Test Set 1), Related to Figure 4
(A) Summary of prediction models of gestational age (GA) before or after 20, 24, 28, 32, and 37 weeks, using two to three metabolites. Note that the prediction
models for 20, 24, and 28 gestational weeks were built using samples from all three trimesters and the ones for late pregnancy (32 and 37 weeks) were build using
third-trimester samples. The contribution rank of each predictor in every model is listed as number 1, 2, and 3. Area under the curves (AUCs) in the validation
cohort (Test Set 1) are listed.
(B) The logistic regression model based on three metabolites can accurately distinguish the third-trimester plasma samples before or after 37 weeks.
(C) Contribution of the three metabolites to the prediction model of gestational age before or after 37 weeks.
(D) Estriol-16-Glucuronide shows intensity range separations before and after 37 weeks.
(E and F) THDOC and androstane-3,17-diol show intensity range separations before/after 37 weeks.
(G–J) The logistic regression models can accurately distinguish pregnancy samples before or after 20 (G) 24 (H), and 28 (I) weeks, and the third trimester plasma
samples before or after 32 weeks (J). GA, gestational age.
 ll
Resource

A B C 5000
●
●

Discovery Test Set 1 R2= 0.11

4500 ●
●

10.0 5
●
● ●
●

●
●

Birth weight (g)

●
●

7.5
4000
Frequency

Frequency
●
●

3 ●
●

5.0 ●
●

●
●

2
3500
●
●

2.5
1
●
●
●
●

● ●

0.0 0 3000 ●
●
●
●
●
●
● ●
●

2 4 6 8 10 2 4 6 8 10
●
●

●
●

Prediction deviation: RMSE Prediction deviation: RMSE

2500 ●
● N= 29
38 39 40 41 42
Birth gestational age (weeks)

D
Androstane-3,17-diol WD > 2w E Estriol−16−Glucuronide WD > 2w
WD < 2w WD < 2w
13.5
log2(Intensity)

13.0

12.5

12.0
28

29

30
22

25

26

28

29

30
22

25

26
Subject ID Subject ID
F G

1.0 1.0
m/z: 257.226076 m/z: 331.226241
RT error (second): 2.4 RT error (second): 29.4

0.5 0.5
Relative intensity
Relative intensity

0.0 0.0

-0.5 -0.5

-1.0 Standard: Androstane-3,17-diol -1.0 Standard: 17α-Hydroxyprogesterone

100 150 200 250 100 200 300

Mass to charge ratio (m/z) Mass to charge ratio (m/z)

H Prediction: Weeks to Delivery (WD) < 4w I Prediction: Weeks to Delivery (WD) < 8w

1.00 1.00

0.75 0.75
True positive rate

True positive rate

0.50 0.50

AUC: 0.96 AUC: 0.90

Discovery N= 128 Discovery N= 128
0.25 0.25
95% CI: 0.93-0.99 95% CI: 0.84-0.96

AUC: 0.87 AUC: 0.91

Validation N= 65 Validation N= 65
0.00 95% CI: 0.78-0.95 0.00 95% CI: 0.80-0.97

0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00

False positive rate False positive rate

(legend on next page)

 ll
Resource

Figure S6. Identified Compounds Predict Gestational Age and 4 and 8 Weeks Approaching Delivery, Related to Figure 5
(A and B) Histogram shows the distribution of prediction deviation (RMSE) in the cross-validation of the discovery cohort (A) and the validation cohort (B) Test
Set 1.
(C) The baby birth weight shows correlation with the gestational length (gestational age at childbirth). All 29 subjects who had baby birth weight information are
included here. The 95% confidence interval for the linear regression is represented by the gray area.
(D and E) Androstane-3,17-diol (D) and estriol-16-Glucuronide (E) show intensity range separations before or after 2 weeks approaching the delivery.
(F and G) Measured MS/MS fragmentation profiles (upper) matching of androstane-3,17-diol (F) and 17a-hydroxyprogesterone (G) with the MS/MS of standard
compounds (lower).
(H and I) The logistic regression models can accurately identify the third trimester plasma samples approaching delivery (weeks to delivery, WD < 4w (H), WD < 8w
(I); only includes women with natural labor onset). Note that the discovery results were from the 10-fold cross-validation (CV) instead of direct fitting to avoid
over-fitting.
 ll

Correction
An Engineered CRISPR-Cas9 Mouse Line
for Simultaneous Readout of Lineage Histories
and Gene Expression Profiles in Single Cells
Sarah Bowling, Duluxan Sritharan, Fernando G. Osorio, Maximilian Nguyen, Priscilla Cheung, Alejo Rodriguez-Fraticelli,
Sachin Patel, Wei-Chien Yuan, Yuko Fujiwara, Bin E. Li, Stuart H. Orkin, Sahand Hormoz,* and Fernando D. Camargo*
*Correspondence: sahand_hormoz@hms.harvard.edu (S.H.), fernando.camargo@childrens.harvard.edu (F.D.C.)
https://doi.org/10.1016/j.cell.2020.06.018

(Cell 181, 1410–1422.e1–e27; June 11, 2020)

Due to a production error, Figures S5 and S6 were not included with the article when it initially published. In addition, in the following
equation, equal ( = ) signs were incorrectly included between the six union (u) signs and the curly brackets.

   
. s .
Mj;k = mj1;k1 m j ; mj1;k1 ˛Mj1;k1
rk
   
. s .
W = d j1;k1 m j ; d j1;k1 ˛Dj1;k1
rk
   
. s .
W = i j1;k1 m j ; i j1;k1 ˛I j1;k1
rk
   
. B .
Dj;k = mj;k1 m ; mj;k1 ˛Mj;k1
rk
       
. B . . B .
W = d j;k1 m ; d j;k1 ˛Dj;k1 W = i j;k1 m ; i j;k1 ˛I j;k1
rk rk
   
. s
I j;k = mj1;k m j ; mj1;k ˛Mj1;k
B
       
. s . . s .
W = d j1;k m j ; d j1;k ˛Dj1;k W = i j1;k m j ; i j1;k ˛I j1;k
B B

Incorrect equation

Cell 181, 1693–1694, June 25, 2020 ª 2020 Elsevier Inc. 1693
 ll

   
. s .
Mj;k = mj1;k1 m j ; mj1;k1 ˛Mj1;k1
rk
   
. s .
W d j1;k1 m j ; d j1;k1 ˛Dj1;k1
rk
   
. s .
W i j1;k1 m j ; i j1;k1 ˛I j1;k1
rk
   
. B .
Dj;k = mj;k1 m ; mj;k1 ˛Mj;k1
rk
   
. B .
W d j;k1 m ; d j;k1 ˛Dj;k1
rk
   
. B .
W i j;k1 m ; i j;k1 ˛I j;k1
rk
   
. s
I j;k = mj1;k m j ; mj1;k ˛Mj1;k
B
   
. s .
W d j1;k m j ; d j1;k ˛Dj1;k
B
   
. s .
W i j1;k m j ; i j1;k ˛I j1;k
B

Correct equation
These errors have now been corrected online, and we apologize for the inconvenience.

1694 Cell 181, 1693–1694, June 25, 2020

 ll

Retraction
Retraction Notice to:
A Monoclonal Antibody that Targets
a NaV1.7 Channel Voltage Sensor
for Pain and Itch Relief
Jun-Ho Lee, Chul-Kyu Park, Gang Chen, Qingjian Han, Rou-Gang Xie, Tong Liu, Ru-Rong Ji,* and Seok-Yong Lee*
*Correspondence: ru-rong.ji@duke.edu (R.-R.J.), sylee@biochem.duke.edu (S.-Y.L.)
https://doi.org/10.1016/j.cell.2020.06.019

(Cell 157, 1393–1404; June 5, 2014)

In this publication, using in vitro and in vivo approaches, we described and characterized a monoclonal antibody (mAb) that binds to
the voltage sensor of the sodium channel subtype Nav1.7 and inhibits channel function. A follow-up study by Liu et al. reported that a
similar but distinct recombinant mAb targeting Nav1.7 did not show significant in vitro activities (Liu et al., 2016, F1000 Res. 5, 2764,
https://doi.org/10.12688/f1000research.9918.1), which prompted us to re-examine our previous results. In the process, we found
irregularities in some of the raw data used for the published in vitro results, and we notified our institution, Duke University, about
the irregularities. The institution subsequently appointed an ad hoc committee that concluded that the first author, Jun-Ho Lee, fabri-
cated and/or falsified the results in Figures 3A, 3C, 3D, and 4. While subsequent work has both clarified the distinct in vitro activities of
the two antibodies and confirmed the in vivo activity of our antibody (Bang et al., 2018, Neurosci. Bull. 34, 22–41, https://doi.org/10.
1007/s12264-018-0203-0), we feel that the responsible course of action is to retract our paper because it contains falsified data. We
sincerely apologize to the scientific community for the inconvenience and confusion that we have caused.

The first author, Jun-Ho Lee, did not respond to the request to sign this retraction.

Cell 181, 1695, June 25, 2020 ª 2020 Elsevier Inc. 1695
SnapShot: JAK-STAT Signaling II
Alejandro V. Villarino,1 Massimo Gadina,1 John J. O’Shea,1 and Yuka Kanno1
1
National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS),
National Institutes of Health (NIH), Bethesda, MD 20892, USA

JAK STAT Ligand

JH
FERM domain 6-7 ND N-terminal domain
Receptor
CCD Coiled-coil domain PLASMA MEMBRANE
SH2 like domain 3-5
DBD DNA-binding domain
Pseudokinase 2 USP18
domain SH2 Src homology 2 domain JAK JAK CYTOPLASM
Y PY PTP SOCS
Kinase domain 1 ATP
CD
C-terminal
PY PY Cytoplasmic
PY Y STAT pool Monomer
JAKi trans-activation domain
PS S JAKi PY PY
STAT STAT

Dimer
FERM 4.1 protein, ezrin,
radixin moesin
Parallel SK
JH JAK homology (1-7) IRF9 Tetramer

Protein tyrosine
STAT2 STAT1 PS ETC
PTP phosphatase PY PY
PY PY PY PY PY PY PY Y Y
PS PS
PS PS PS PS PS PS PS
PS
I II IV ATP
Suppressor of
SOCS cytokine signaling
III V Ca2+
Heterotrimer Tetramer Heterodimer Homodimer Antiparallel STAT3
Ubiquitin specific
USP18
peptidase 18
Mitochondrion

SK Serine kinase

ATP Adenosine triphosphate

JAKi JAK inhibitor

ETC Electron transport chain Proximal enhancer Gene-coding region

PTP
Gene transcritption Intronic
enhancer
TSS Transcription start site
HAT
TF Other transcription factors Pol II
TF
Pol II RNA polymerase II TSS Pol II

HAT Histone acetyltransferase Promotor mRNA

MT Methyltransferase PIAS IL-12-induced STAT4 binding motif
T
TTCC GGAAGAA
G
eRNA enhancer RNA A T
C
T
C G
A
GA
G
TC ATC C

PIAS Protein inhibitor of eRNA transcritption also STAT1(IFN-γ), STAT3(IL-6), STAT5(IL-2)

activated STAT

IRF9 HAT
TF Pol II HAT

MT
IFN-β-induced STAT2 binding motif IL-4-induced STAT6 binding motif TF
ATA
GT
TC A
CA
AGTTTC
AGTTTCC
C
CCGG
T TC
T
AGG C
CC
GG
A TTCC A
G
TT
TCG
CGT
A AG AC
GAA
T C

NUCLEUS also STAT1(IFN-β)

Distal enhancer Repressed chromatin

Mutations of JAK Mutations of STAT

Mouse STAT Mouse knockout Human LOF Human GOF
JAK Human LOF Human GOF
knockout STAT1
Impaired IFN signaling, viral
Primary immunodeficiency
Fungal susceptibility,
and bacterial susceptibility autoimmunity, aneurysms
Primary immunodeficiency, Systemic immune Interferonopathy,
Perinatal lethality, atypical mycobacterial dysregulation, STAT2 Viral susceptibility Viral susceptibility
autoinflammatory disease
JAK1 T−B+NK− SCID infections; somatic hypereosinophilic
mutations in cancers syndrome Lethal; multiple defects Primary immunodeficiency, Systemic autoimmunity,
STAT3 revealed by conditional hyperimmunoglobulin anemia, leukemia,
deletion E syndrome lymphomas other cancers
Myeloproliferative
Embryonic lethal,
JAK2 Not reported neoplasms,
Tuberculosis, atypical
Variants associated with
anemia leukemia, lymphoma rheumatoid arthritis, Sjogren’s
STAT4 Impaired type I responses mycobacterial and fungal
syndrome and systemic
susceptibility, Kaposi sarcoma
lupus erythematosus

JAK3 T−B+NK− SCID
Viral susceptibility,
TYK2 diminished responses
to type I IFNs, IL-12, IL-23
T−B+NK− SCID Leukemia Impaired mammary gland Leukemia, lymphomas
STAT5A development Not reported other cancers
Somatic mutations and
Impaired NK cell function and eosinophilia, urticaria, leukemia,
STAT5B sexual dimorphic growth Dwarfism, autoimmunity
Primary immunodeficiency, lymphomas, other cancers
variants protect against Not reported
autoimmunity Impaired type II
STAT6 immune responses
Not reported Lymphoma

See online version for

1696 Cell 181, June 25, 2020 © 2020 Published by Elsevier Inc.  DOI https://doi.org/10.1016/j.cell.2020.04.052 legends and references
SnapShot: Jak-STAT Signaling II
Alejandro V. Villarino,1 Massimo Gadina,1 John J. O’Shea,1 and Yuka Kanno1
1
National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS),
National Institutes of Health (NIH), Bethesda, MD 20892, USA

The discovery of the Janus kinase (JAK) and signal transducer and activator of transcription (STAT) pathway arose from investigation of interferon (IFN) signaling (Darnell et
al., 1994; Leonard and O’Shea, 1998). Canonical JAK-STAT signaling begins with extracellular apposition of members of a family of structurally related cytokines, interleukins,
interferons, colony-stimulating factors, and some hormones with their corresponding structurally related transmembrane receptors. This enables trans-activation of receptor-
bound JAKs that catalyze tyrosine phosphorylation (p-Tyr) of receptors and STATs, resulting in the formation of homodimers and/or heterodimers that accumulate in the
nucleus and instruct gene transcription. Interferons also induce a complex of STAT1, STAT2, and IRF9.
In mammals, there are 4 JAKs sharing 4 major structural domains: (1) the FERM (band-four-point-one ezrin radixin moesin) domain, which mediates interaction with
receptors and promotes kinase function, (2) the SH2-like domain, which mediates interaction with receptors, (3) the pseudokinase domain, which regulates kinase activity,
and (4) the kinase domain. Germline loss of function (LOF) mutations of JAK3 and TYK2 underlie human primary immunodeficiency disorders (Tangye et al., 2017). Somatic
LOF mutations of JAK1 can arise in tumor cells and are associated with resistance to IFN-γ and cancer evasion. Gain of function (GOF) mutations underlie systemic autoim-
munity (JAK1), polycythemia vera (JAK2 kinase-like domain), leukemias, lymphomas, and other malignancies. JAKs were recognized as pivotal drug targets and multiple JAK
inhibitors (jakinibs) have been approved for the treatment of myeloproliferative neoplasms, graft versus host disease, rheumatoid arthritis, psoriatic arthritis, and inflammatory
bowel disease (Gadina et al., 2020). Most jakinibs target the JAK kinase domain, but newer agents act via allosteric mechanisms (e.g., target the kinase-like domain). Despite
the clinical success, questions remain regarding the optimal degree of JAK inhibition for specific cell types in various tissues and disorders. Compared with first-generation
pan-jakinibs, selective jakinibs may provide advantages in terms of reduced toxicity. For example, selective targeting of JAK1, TYK2, and JAK3 would avoid disrupting actions
of JAK2-dependent cytokines involved in hematopoiesis (e.g., erythropoietin). In some circumstances, though, selectivity might result in reduced efficacy.
There are seven mammalian STAT family members bearing five major structural domains. In addition to JAKs, STATs may be phosphorylated by receptor tyrosine kinases,
src family kinases, and bacterial and parasite enzymes. STAT expression and availability are subject to a range of intrinsic (e.g., cellular lineage) and extrinsic (e.g., cytokines)
factors. They are regulated at transcriptional, post-transcriptional, and post-translational levels, including by protein tyrosine phosphatases (PTPs), which “de-activate” STATs
and mediate degradation or recycling. STATs are subject to serine phosphorylation (p-Ser), which influences both p-Tyr-dependent and p-Tyr-independent functions. The latter
are particularly relevant for “unphosphorylated” STATs (uSTATs), which are arranged as anti-parallel dimers (unlike conventional p-Tyr-dependent parallel dimers). uSTATs can
modulate the pool of cytoplasmic STATs and mediate gene transcription (Stark et al., 2018; Stark and Darnell, 2012). STATs are present in mitochondria and positively regulate
of complexes I, II, and V of the electron transport chain, elevating mitochondrial membrane potential and favoring the mitochondrial respiration (Garama et al., 2016).
STATs are classical transcription factors (TFs) that engage DNA regulatory elements (DREs) bearing a defined sequence motif and instruct transcription of protein-coding
(mRNAs) or non-coding genes (microRNAs, long non-coding RNAs). This is achieved through a combination of proximal DREs located close to the transcriptional start sites
(TSS) or within gene bodies, distal elements which can physically interact via chromatin looping. STATs bind broadly throughout the genome at promoters, even more so at
enhancers. STATs often congregate at both conventional and super-enhancers; notably, the genes encoding STATs themselves bear super-enhancers, in line with the idea that
they are tightly transcriptionally regulated.
STATs typically engage multiple DREs associated with a given gene, and genes are often bound by multiple STAT family members, along with other TFs, such as IRF, AP-1,
and NF-κB family proteins, creating a platform for multi-molecular networks to control gene expression (Villarino et al., 2017). STATs also influence chromatin remodeling by
recruiting histone-modifying enzymes. Although originally identified as transcriptional activators, STAT binding is associated with both induction and repression of genes.
Germline STAT mutations are associated with primary immunodeficiency and autoimmunity, whereas somatic GOF STAT mutations are associated with cancer. STATs do
not have enzymatic activity and are more challenging than JAKs as therapeutic targets. Three potential strategies have been employed: (1) inhibitory peptides, which seques-
ter STATs from upstream receptors and kinases, (2) small-molecule inhibitors, which impede STAT activation and/or function, and (3) decoy oligonucleotides, which sequester
STATs away from genomic binding sites.
The JAK-STAT pathway is negatively regulated in multiple ways. Aside from dephosphorylation, suppressor of cytokine signaling (SOCS) proteins are induced by cytokine
and IFNs and inhibit proximal receptor signaling (Morris et al., 2018). Ubiquitin specific peptidase 18 (USP18) is induced by IFNs and binds to STAT2, negatively regulating its
function; mutations that interfere with USP18 action are associated with lethal interferonopathy (Basters et al., 2018). STATs are also targeted and degraded by viruses.
In summary, the JAK-STAT pathway has emerged as a paradigm for membrane-to-nucleus signaling, and building on the legacy of the first quarter century of JAK-STAT
research, the field continues to deliver transformative, clinically relevant insights on the nature of intercellular communication, gene expression, and signal-dependent
transcription factors. Still, there is much to learn. Detailed molecular and cell-biologic understanding of cytokine receptor/JAK structure-function relationships will be further
enhanced by advanced real-time measurement of signaling and gene transcription by imaging. These advances are eagerly awaited and will no doubt offer many new trans-
lational insights and therapeutic opportunities.

ACKNOWLEDGMENTS
This work was supported by the NIAMS Intramural Research Program.

REFERENCES
Basters, A., Knobeloch, K.P., and Fritz, G. (2018). USP18 - a multifunctional component in the interferon response. Biosci. Rep. 38, BSR20180250.
Darnell, J.E., Jr., Kerr, I.M., and Stark, G.R. (1994). Jak-STAT pathways and transcriptional activation in response to IFNs and other extracellular signaling proteins. Science 264,
1415–1421.
Gadina, M., Chisolm, D.A., Philips, R.L., McInness, I.B., Changelian, P.S., and O’Shea, J.J. (2020). Translating JAKs to Jakinibs. J. Immunol. 204, 2011–2020.
Garama, D.J., White, C.L., Balic, J.J., and Gough, D.J. (2016). Mitochondrial STAT3: Powering up a potent factor. Cytokine 87, 20–25.
Leonard, W.J., and O’Shea, J.J. (1998). Jaks and STATs: biological implications. Annu. Rev. Immunol. 16, 293–322.
Morris, R., Kershaw, N.J., and Babon, J.J. (2018). The molecular details of cytokine signaling via the JAK/STAT pathway. Protein Sci. 27, 1984–2009.
Stark, G.R., and Darnell, J.E., Jr. (2012). The JAK-STAT pathway at twenty. Immunity 36, 503–514.
Stark, G.R., Cheon, H., and Wang, Y. (2018). Responses to Cytokines and Interferons that Depend upon JAKs and STATs. Cold Spring Harb. Perspect. Biol. 10, a028555.
Tangye, S.G., Pelham, S.J., Deenick, E.K., and Ma, C.S. (2017). Cytokine-Mediated Regulation of Human Lymphocyte Development and Function: Insights from Primary Immuno-
deficiencies. J. Immunol. 199, 1949–1958.
Villarino, A.V., Kanno, Y., and O’Shea, J.J. (2017). Mechanisms and consequences of Jak-STAT signaling in the immune system. Nat. Immunol. 18, 374–384.

1696.e1 Cell 181, June 25, 2020 © 2020 Published by Elsevier Inc.  DOI https://doi.org/10.1016/j.cell.2020.04.052