0% found this document useful (0 votes)

122 views174 pages

Science

Q: What novel spindle assembly mechanisms have been identified in human oocytes compared to other mammalian species, and what genetic components are crucial for this process?

In human oocytes, a distinct spindle assembly mechanism has been identified wherein the nucleation of spindle microtubules is initiated from kinetochores after nuclear envelope breakdown (NEBD), contrasting with other mammals like rodents. The novel structure observed, termed the human oocyte microtubule organizing center (huoMTOC), comprises components such as TACC3, CKAP5, CCP110, and DISC1, which are crucial for spindle microtubule nucleation and subsequent assembly . This differs from other species, which typically use acentriolar microtubule organizing centers (aMTOCs). These findings highlight the uniqueness of human oocyte spindle mechanisms and underscore the importance of understanding specific proteins and their roles in oocyte maturation and fertility treatments .

Q: Discuss how the geometry of CNT transistors impacts their scalability and performance. What contact structures have been proposed to overcome these challenges?

The geometry of CNT transistors significantly influences their scalability and overall performance. Scaled geometries, such as bottom-gate or top-gate configurations, directly impact how short the channel lengths can be, with measurements as short as 5 nm reported . Contact structures, like edge contacts, offer ideal scalability by minimizing the contact length while maintaining efficient charge injection. Palladium contacts have achieved near-theoretical quantum limits at 10-nm contact lengths, but reproducibility remains a challenge . End-bonded contacts, such as those using Mo-CNT carbide structures, have demonstrated potential for sub-10-nm contact lengths, further pushing the boundaries of scalability and performance . These contacts reduce resistance and enhance carrier injection, essential for both p-type and n-type configurations, and are crucial for maintaining performance as devices are miniaturized .

Q: Describe the challenges associated with achieving scalable contact lengths in CNT transistors and the solutions proposed to address degradation issues at sub-30-nm lengths.

Achieving scalable contact lengths in CNT transistors is challenging due to performance degradation issues that arise at sub-30-nm lengths. As contact length scales down, maintaining efficient charge injection and low resistance becomes increasingly difficult. It has been observed that some processes encounter severe degradation at these scales . End-bonded structures, such as those involving Mo-reacted carbide contacts, are proposed solutions that offer potential for ideal scalability without significant performance loss . This edge-contact structure focuses on reducing contact resistance and enhancing scalability while maintaining low processing temperatures and minimizing material and transport degradation . Additionally, optimizing material combinations and processing techniques can further alleviate barriers to achieving consistent high-performance contacts at these reduced dimensions.

Q: What are the advantages and challenges of using complementary metal-oxide-semiconductor (CMOS) technology in relation to CNT transistors for digital logic applications?

The primary advantage of using CMOS technology with CNT transistors is the potential to surpass conventional silicon technology in terms of performance and energy efficiency. CNT transistors, particularly when arranged in gate-all-around configurations with doped extensions, can achieve energy-delay product (EDP) benefits as much as seven times greater than silicon nanosheets at technology nodes like 2 nm . However, a significant challenge is the compatibility of CNT transistor fabrication processes with existing CMOS fabrication methods. Although some metal contact processes, such as liftoff, are not scalable, alternate patterning processes tend to be slow, indicating a need for further development to make the transition feasible . Another challenge involves achieving consistently high-quality, scalable contacts, particularly for n-type CNT transistors .

Q: How does the interaction between cancer cells and T cells reflect the impact of genetic differences such as those between FMRP-WT and FMRP-KO cell lines?

The interaction between cancer cells and T cells is significantly affected by genetic differences, such as those between FMRP-WT and FMRP-KO cell lines. FMRP-WT cells tend to exhibit an immunosuppressive environment, marked by the regulation of immune-modulatory genes like IL-33, which stimulates the production of Tregs and enhances immunosuppression . In contrast, FMRP-KO cells are characterized by increased CD8 T cell infiltration due to up-regulated chemokines like CCL7, which recruit immune effector cells, highlighting a less immunosuppressive microenvironment . The differential expression of cytokines and chemokines between these genetic states underscores how inherent cellular genetic differences dictate the tumor immune microenvironment, thereby influencing tumor growth dynamics, infiltration by immune cells, and the efficacy of potential antitumor immune responses .

Q: Evaluate the role of FMRP in the immune response of cancer cells and its implications for cancer treatment strategies.

FMRP plays a crucial role in modulating the immune response of cancer cells by regulating the expression of cytokines that influence immune cell behavior. In FMRP-WT cancer cells, immune-modulatory genes such as IL-33 are up-regulated, which stimulates the production of Tregs and promotes an immunosuppressive microenvironment, aiding tumor evasion from immune attack . Conversely, FMRP-KO cells up-regulate chemokines like CCL7, which enhance CD8 T cell recruitment, facilitating a more pro-inflammatory and less immunosuppressive milieu . These differential gene expression profiles suggest that targeting FMRP-related pathways could offer novel therapeutic strategies, either by enhancing immune system recruitment and cytotoxic activity in FMRP-KO phenotypes or mitigating immunosuppression in tumors with high FMRP expression .

Q: How do the environmental conditions within tumors, shaped by FMRP expression, influence the recruitment and activity of T cells and macrophages?

The environmental conditions within tumors, influenced by FMRP expression, significantly impact the recruitment and activity of T cells and macrophages. In FMRP-KO tumors, there is an up-regulation of chemokines like CCL7, leading to enhanced recruitment of CD8 T cells, creating a more inflammatory microenvironment and potentially increasing antitumor immunity . Meanwhile, decreased levels of IL-33 in FMRP-KO tumors compromise the abundance of Tregs, which are typically associated with immunosuppression . Despite these changes, FMRP-KO tumors do not show alterations in TAM programming, indicating that factors beyond CCL7 influence macrophage behavior . These findings suggest that manipulating FMRP-regulated pathways could be a strategic way to modulate the tumor immune environment and enhance immunotherapeutic outcomes .

Q: What are the critical factors in ensuring the effective integration of CNT technology in large-scale thin-film transistor (TFT) applications?

The effective integration of CNT technology into large-scale thin-film transistor (TFT) applications depends on several critical factors, including compatibility with large substrate sizes and maintaining low fabrication costs. TFTs are often used in display backplanes and need to be manufactured at a commodity cost level, meaning materials and processes should be scalable to accommodate large panels . While device-level performance is important, the relaxed constraints on performance in TFTs compared to nanoscale FETs place greater emphasis on manufacturing scalability and cost-effectiveness . Thus, identifying and implementing materials and processes that allow for high-throughput manufacturing without compromising device functionality is key for the widespread adoption of CNT technology in TFT applications.

Q: Compare the potential applications of CNT transistors based on their cost and complexity as observed in different integration scales.

CNT transistors exhibit a broad range of potential applications, largely influenced by their performance relative to cost and complexity. At the microscale, CNT transistors are utilized in thin-film devices such as printed electronics and biosensors, where lower complexity and cost are prioritized. Conversely, at the nanoscale, they are integrated into high-performance field-effect transistors (FETs) for applications like low-voltage very-large-scale integration (VLSI), where performance drives design even at higher costs . Three-dimensional (3D) integration at the back end of line (BEOL) level—such as heterogeneous 3D layers on silicon CMOS—represents another application area where integration complexity increases with corresponding performance enhancements . Thus, as the desired application shifts from cost-efficiency to performance optimization, CNT transistors facilitate a scalable approach to meet those needs across diverse technological landscapes.

Q: How does the purity of semiconducting carbon nanotubes (CNTs) affect the performance of digital transistors, and what purity levels are typically targeted?

The purity of semiconducting CNTs directly affects the performance of digital transistors, with higher purity leading to better device performance. For high-performance digital transistors, achieving greater than 99.9999% purity is desired because any metallic CNTs present could degrade the transistor's electrical properties. This level of purity minimizes unwanted residuals, enhances electrical contact, and improves gating efficiency and carrier transport in CNT transistors . Techniques like gel chromatography and aqueous two-phase extraction have helped achieve purities close to these targets, making CNTs more viable for advanced applications .

Uploaded by

piotr reysner

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

122 views174 pages

Science

Uploaded by

piotr reysner

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

A Texas-size defense against Short-lived menstruation changes Phloem flow controls

storms and sea-level rise p. 698 after COVID-19 vaccination p. 704 root outgrowth p. 762

$15
18 NOVEMBER 2022
SPECIAL ISSUE
[Link]

75 YEARS OF
TRANSISTORS
[Link]/journal/sciimmunol

READY TO
PUT THE
SPOTLIGHT
ON YOUR
RESEARCH?
Submit your research:
[Link]

Twitter: @SciImmunology
Facebook: @ScienceImmunology

[Link] 678 11/10/22 8:21 AM

Produced by the Science

Advertorial

Zhejiang Lab’s research team members discuss chip design and testing plans.

Intelligent computing connects everything

At Zhejiang Lab, scientists develop new tools and techniques that will expand our computer interaction,” Shi says. “Driven by the intelligent data reactor engine, it
understanding of the universe, improve health care, and open up new avenues schedules different computing tasks with the optimal computing resources, adapts
of research. the best computing methods, and forms the optimal results.”
For FAST and the future FAST array, the Astronomical Big Data Intelligent
The Internet of Everything (IoE) will connect people, processes, and devices. Computing and Service Platform will support, as several examples, discoveries
Scientists at Zhejiang Lab in Hangzhou, China see the IoE as a symbol of the new era of pulsars, fast radio bursts, and extragalactic galaxies. It’s already been used to
of digital civilization, with intelligent computing encompassing computing methods propose a unified mechanism that can explain the evolution of the polarization
and systems complemented by other technical capabilities that will drive the IoE. At frequency of repeated fast radio bursts.
Zhejiang Lab, scientists work on chips, algorithms, software, and novel computing In addition, this facility at Zhejiang Lab will advance other fields. As Shi says, “It
architectures for intelligent computing. will provide new methods, new tools, and new paradigms for scientific discovery,
““We take advantage of intelligent computing and novel research paradigms social governance, and the digital economy, and promote work in health care and
to revolutionarily advance the research in pertinent disciplines and develop new other fields.”
research models,” says Tuo Shi, research specialist at Zhejiang Lab’s Research
Center for Intelligent Computing Hardware. Making a material genome
To drive this research, Zhejiang Lab created its Intelligent Computing Data “Although many people know about the genetic genome and other ’omes, the
Reactor. This facility collects virtually any type of data in real time and analyzes it material genome refers to the acceleration of the discovery and design of novel
with intelligent computing. As Shi explains, the reactor “efficiently uses intelligent- materials using high value datasets and computational technologies,” Shi explains.
computing clusters, intelligent supercomputers, brain-like computers, as well as “Based on reliable experimental results obtained from characterization platforms
intelligent-computing software and hardware.” and their databases involving chemical compositions, crystal structures, and
PHOTOS: LEFT: © ZHENGZAISHURU/[Link]; RIGHT: PROVIDED BY ZHEJIANG LAB

That collection of intelligent-computing technology can be applied to a wide physical properties, the relation between a material’s structure and device
range of fields including astronomy, genetics, materials and medical treatments. performance can be established via theoretical predictions and simulations, leading
to the discovery of new materials for innovative commercial purposes.”
Fostering FAST advances Zhejiang Lab’s Material Intelligent Computing Platform combines big data,
In Guizhou, a mountainous region in southwest China, scientists built a radio artificial intelligence, and domain knowledge to develop new approaches to creating
telescope known as FAST (Five-hundred-meter Aperture Spherical radio Telescope). intelligent materials. For example, artificial proteins can be developed for advanced
With a diameter of 500 meters, which is more than 1.5 times the height of the Eiffel cancer treatments and other disease treatments.
Tower, FAST is the world’s largest single-dish radio telescope, but that’s just the The future of connecting people and technology will change how we see and
beginning. As Shi notes, “The future ‘FAST array’ will be a large-scale telescope interact with the world. Intelligent computing will create the foundation of
cluster consisting of up to five FASTs.” This array will be used to explore the early those advances.
universe, detect gravitational waves, and more. Sponsored by
To analyze data from projects like FAST, Zhejiang Lab developed its Astronomical
Big Data Intelligent Computing and Service Platform. “The service platform takes
the computing facility and intelligent technology as the base, and then uses
data, algorithms, models, and knowledge to build a public knowledge base, and
six engines, which include operation management, collaborative computing,
knowledge construction, simulation deduction, data processing, and human–

[Link] 679 11/10/22 2:55 PM

CONTENTS
1 8 N O V E M B E R 2 0 2 2 • VO LU M E 3 7 8 • I S S U E 6 6 2 1

698
Houston’s petrochemical plants are at risk from hurricane-driven flooding.

NEWS 697 Billionaire’s crypto company

collapse strands scientists
As FTX files for bankruptcy, the grantees
709 Potential bias in genetic correlations
Mating patterns across two traits can inflate
estimates of genetic overlap
its foundations supported may not see all By A. D. Grotzinger and M. C. Keller
IN BRIEF of their pledged money By R. F. Service RESEARCH ARTICLE p. 754

690 News at a glance FEATURES

710 Adding functions to pyridines
IN DEPTH 698 Shelter from the storm Chemical reactions break a pyridine ring to
A plan to wall off Houston and nearby allow its modification By J. M. Joo
692 Science community braces for industry from flooding caused by REPORTS pp. 773 & 779
divided government hurricanes will cost tens of billions of
Split control of Congress could produce fiery dollars. Will it be enough?
hearings and limit new funding By J. Mervis By W. Cornwall
712 Using mirrors to control molecular
dynamics
693 Tumors can teem with microbes. An optical cavity mixes molecular vibrations
But what are they doing there?
New study suggests microbiomes can
promote cancer by suppressing immune
INSIGHTS with light and changes chemical reactivity
By L. Chuntonov
REPORT p. 790
response and seeding metastases By G. Sinha
PERSPECTIVES POLICY FORUM
694 Whistleblower finds possible

PHOTO: WITOLD SKRYPCZAK/ALAMY STOCK PHOTO

misconduct in his own papers 704 COVID-19 vaccination and 713 Research under China’s personal
Matthew Schrag confronts a mentor after menstruation information law
their joint work is flagged on the PubPeer COVID-19 vaccination causes small changes The new law may present obstacles to some
website By C. Piller to menstruation that quickly resolve kinds of research By X. Li et al.
By V. Male
696 Booming trade in mammoth ivory may BOOKS ET AL.
be bad news for elephants 706 Defining the onset of the
Paleontologists are urged to take a stand Anthropocene 716 The Gaia letters
against a market that may provide cover for Twelve sites are considered for defining the Contextualized correspondence traces the
continued poaching By M. Price Anthropocene geological epoch emergence of a provocative hypothesis
PODCAST By C. N. Waters and S. D. Turner By P. Falkowski

680 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118 TOC_16198450.indd 680 11/15/22 5:36 PM

717 Aquatic life and its parallels
Blurring the personal and the scientific, SPECIAL SECTION
an author probes the beauty, terror, and
richness of the natural world By J. Berwald 75 YEARS OF TRANSISTORS
LETTERS
INTRODUCTION ON THE COVER
718 Editorial Retraction 720 From one transistor… A hand holds a portable radio origi-
By H. H. Thorp nally presented at the 21st Radio and
PERSPECTIVES Television Show at the Porte de
718 Brazil’s preventable bridge disasters Versailles Exhibition Center in Paris,
By W. E. Magnusson 722 Moore’s law: The journey ahead
France, on 9 September 1959. The
M. S. Lundstrom and M. A. Alam
invention of transistors 75 years ago
718 Welcoming Taiwan’s diaspora spurred innovations that affected
724 Toward gallium oxide power
scientists communication, computation, control
electronics M. J. Tadjer systems, instrumentation, and other
By J. Huang
REVIEWS technologies. Today, the development of
719 Life in Science: The elephant new transistor tech-
in the inbox 726 Carbon nanotube transistors: nologies continues
By C. T. A. Lewis Making electronics from molecules on many fronts.
A. D. Franklin et al. See the special
section beginning

RESEARCH
733 Toward attoJoule switching energy on page 720. Photo:
in logic transistors S. Datta et al. Keystone-France/
Gamma-Rapho via
SEE ALSO EDITORIAL p. 683 Getty Images
IN BRIEF
741 From Science and other journals
747 Solar cells Organic chemistry
RESEARCH ARTICLES Initializing film homogeneity to retard phase 773 Halogenation of the 3-position
segregation for stable perovskite solar cells of pyridines through Zincke imine
744 Developmental biology Y. Bai et al. intermediates B. T. Boyle et al.
Hippo signaling instructs ectopic but not 779 Radical and ionic meta-C–H
normal organ growth W. Kowalczyk et al. 754 Human genomics functionalization of pyridines, quinolines,
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT: Cross-trait assortative mating is widespread and isoquinolines H. Cao et al.
[Link]/10.1126/SCIENCE.ABG3679
and inflates genetic correlation estimates PERSPECTIVE p. 710
R. Border et al.
745 Human fertility PERSPECTIVE p. 709 785 Quantum physics
The mechanism of acentrosomal spindle Noise-resilient edge modes on a chain
assembly in human oocytes T. Wu et al. 762 Plant science of superconducting qubits
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT: Hydraulic flux–responsive hormone X. Mi et al.
[Link]/10.1126/SCIENCE.ABQ7361
redistribution determines root branching
P. Mehra et al. 790 Chemical physics
746 Cancer Cavity-enabled enhancement of ultrafast
Aberrant hyperexpression of the RNA binding intramolecular vibrational redistribution over
REPORTS
protein FMRP in tumors mediates immune pseudorotation T.-T. Chen et al.
evasion Q. Zeng et al. 768 3D printing PERSPECTIVE p. 712
RESEARCH ARTICLE SUMMARY; FOR FULL TEXT: Mechanical nanolattices printed using
[Link]/10.1126/SCIENCE.ABL7207 nanocluster-based photoresists Q. Li et al.
DEPARTMENTS
683 Editorial
746 Shockley was a racist and eugenicist
By H. H. Thorp
75 YEARS OF TRANSISTORS SECTION p. 720

798 Working Life

Silent no longer By W. P. Suza

Science Staff ............................................. 682

Killer T cells (red) attack susceptible cancer cells (green). Science Careers .........................................796
IMAGE: JEREMY GUILLOT

SCIENCE (ISSN 0036-8075) is published weekly on Friday, except last week in December, by the American Association for the Advancement of Science, 1200 New York Avenue, NW, Washington, DC 20005. Periodicals mail
postage (publication No. 484460) paid at Washington, DC, and additional mailing offices. Copyright © 2022 by the American Association for the Advancement of Science. The title SCIENCE is a registered trademark of the AAAS. Domestic
individual membership, including subscription (12 months): $165 ($74 allocated to subscription). Domestic institutional subscription (51 issues): $2212; Foreign postage extra: Air assist delivery: $98. First class, airmail, student, and
emeritus rates on request. Canadian rates with GST available upon request, GST #125488122. Publications Mail Agreement Number 1069624. Printed in the U.S.A.
Change of address: Allow 4 weeks, giving old and new addresses and 8-digit account number. Postmaster: Send change of address to AAAS, P.O. Box 96178, Washington, DC 20090–6178. Single-copy sales: $15 each plus shipping and
handling available from [Link]; bulk rate on request. Authorization to reproduce material for internal or personal use under circumstances not falling within the fair use provisions of the Copyright Act can be obtained
through the Copyright Clearance Center (CCC), [Link]. The identification code for Science is 0036-8075. Science is indexed in the Reader’s Guide to Periodical Literature and in several specialized indexes.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 681

1118 TOC_16198450.indd 681 11/15/22 5:36 PM

Editor-in-Chief Holden Thorp, hthorp@[Link] BOARD OF REVIEWING EDITORS (Statistics board members indicated with S)
Executive Editor Valda Vinson Erin Adams, U. of Chicago Daniel Haber, Mass. General Hos. Ana Pêgo, U. do Porto
Editor, Research Jake S. Yeston Editor, Insights Lisa D. Chong Managing Editor Lauren Kmec Takuzo Aida, U. of Tokyo Sharon Hammes-Schiffer, Yale U. Samuel Pfaff, Salk Inst.
DEPUTY EDITORS Gemma Alderton (UK),Stella M. Hurtley (UK), Phillip D. Szuromi, Sacha Vignieri SR. EDITORS Caroline Ash (UK), Leslie Aiello, Wenner-Gren Fdn. Wolf-Dietrich Hardt, ETH Zürich Julie Pfeiffer,
Michael A. Funk, Brent Grocholski, Pamela J. Hines, Di Jiang, Priscilla N. Kelly, Marc S. Lavine (Canada), Mattia Maroso, Deji Akinwande, UT Austin Louise Harra, UCL UT Southwestern Med. Ctr.
Yevgeniya Nusinovich, Ian S. Osborne (UK), L. Bryan Ray, Seth Thomas Scanlon (UK), H. Jesse Smith, Keith T. Smith (UK), Judith Allen, U. of Manchester Carl-Philipp Heisenberg, Philip Phillips, UIUC
Jelena Stajic, Peter Stern (UK), Valerie B. Thompson, Brad Wible ASSOCIATE EDITORS Bianca Lopez, Corinne Simonti, Yury V. Marcella Alsan, Harvard U. IST Austria Matthieu Piel, Inst. Curie
James Analytis, UC Berkeley Janet G. Hering, Eawag Kathrin Plath, UCLA
Suleymanov, Ekeoma Uzogara LETTERS EDITOR Jennifer Sills LEAD CONTENT PRODUCTION EDITORS Chris Filiatreau, Harry Jach
Paola Arlotta, Harvard U. Christoph Hess, Martin Plenio, Ulm U.
SR. CONTENT PRODUCTION EDITOR Amelia Beyna CONTENT PRODUCTION EDITORS Julia Haber-Katris, Nida Masiulis, Abigail Shashikanth,
Delia Baldassarri, NYU U. of Basel & U. of Cambridge Katherine Pollard, UCSF
Suzanne M. White SR. EDITORIAL MANAGERS Carolyn Kyle, Beverly Shields SR. PROGRAM ASSOCIATE Maryrose Madrid EDITORIAL
Nenad Ban, ETH Zürich Heather Hickman, NIAID, NIH Elvira Poloczanska,
ASSOCIATE Joi S. Granger SR. EDITORIAL COORDINATORS Aneera Dobbins, Jeffrey Hearn, Lisa Johnson, Jerry Richardson, Hilary
Christopher Barratt, Hans Hilgenkamp, U. of Twente Alfred-Wegener-Inst.
Stewart (UK), Alice Whaley (UK), Anita Wynn EDITORIAL COORDINATORS Maura Byrne, Alexander Kief, Ronmel Navas, Isabel
U. of Dundee Janneke Hille Ris Lambers, Julia Pongratz,
Schnaidt, Qiyam Stewart, Brian White RESEARCH & DATA ANALYST Jessica L. Slater ASI DIRECTOR, OPERATIONS Janet Clements (UK)
Nandita Basu, U. of Waterloo ETH Zürich Ludwig Maximilians U.
ASI SR. OFFICE ADMINISTRATOR Jessica Waldock (UK)
Franz Bauer, Kai-Uwe Hinrichs, U. of Bremen Philippe Poulin, CNRS
News Editor Tim Appenzeller Pontificia U. Católica de Chile Deirdre Hollingsworth, Lei Stanley Qi, Stanford U.
NEWS MANAGING EDITOR John Travis INTERNATIONAL EDITOR Martin Enserink DEPUTY NEWS EDITORS Shraddha Chakradhar, Elizabeth Ray H. Baughman, UT Dallas U. of Oxford Trevor Robbins, U. of Cambridge
Culotta, Lila Guterman, David Grimm, Eric Hand (Europe), David Malakoff SR. CORRESPONDENTS Daniel Clery (UK), Jon Cohen, Jeffrey Carlo Beenakker, Leiden U. Randall Hulet, Rice U. Joeri Rogelj, Imperial Coll. London
Mervis, Elizabeth Pennisi ASSOCIATE EDITORS Jeffrey Brainard, Michael Price, Kelly Servick NEWS REPORTERS Adrian Cho, Jennifer Yasmine Belkaid, NIAID, NIH Auke Ijspeert, EPFL Amy Rosenzweig, Northwestern U.
Couzin-Frankel, Jocelyn Kaiser, Rodrigo Pérez Ortega (Mexico City), Robert F. Service, Erik Stokstad, Paul Voosen, Meredith Wadman Philip Benfey, Duke U. Gwyneth Ingram, ENS Lyon Mike Ryan, UT Austin
INTERNS Zack Savitsky, Viviana Flores, Katherine Irving CONTRIBUTING CORRESPONDENTS Warren Cornwall, Andrew Curry (Berlin),
Kiros T. Berhane, Columbia U. Darrell Irvine, MIT Miquel Salmeron,
Joseph J. Berry, NREL Akiko Iwasaki, Yale U. Lawrence Berkeley Nat. Lab
Ann Gibbons, Sam Kean, Eli Kintisch, Kai Kupferschmidt (Berlin), Andrew Lawler, Mitch Leslie, Eliot Marshall, Virginia Morell, Dennis
Alessandra Biffi, Harvard Med. Stephen Jackson, Nitin Samarth, Penn State U.
Normile (Tokyo), Elisabeth Pain (Careers), Charles Piller, Gabriel Popkin, Michael Price, Joshua Sokol, Richard Stone, Emily
Chris Bowler, USGS & U. of Arizona Erica Ollmann Saphire,
Underwood, Gretchen Vogel (Berlin), Lizzie Wade (Mexico City) CAREERS Rachel Bernstein (Editor), Katie Langin (Associate Editor)
École Normale Supérieure Erich Jarvis, Rockefeller U. La Jolla Inst.
COPY EDITORS Julia Cole (Senior Copy Editor), Morgan Everett, Cyra Master (Copy Chief) ADMINISTRATIVE SUPPORT Meagan Weiland
Ian Boyd, U. of St. Andrews Peter Jonas, IST Austria Joachim Saur, U. zu Köln
Creative Director Beth Rakouskas Malcolm Brenner, Baylor Coll. Johanna Joyce, U. de Lausanne Alexander Schier, Harvard U.
DESIGN MANAGING EDITOR Chrystal Smith GRAPHICS MANAGING EDITOR Chris Bickel PHOTOGRAPHY MANAGING EDITOR Emily Petersen of Med. Matt Kaeberlein, U. of Wash. Wolfram Schlenker, Columbia U.
MULTIMEDIA MANAGING PRODUCER Kevin McLean WEB CONTENT STRATEGY MANAGER Kara Estelle-Powers DESIGN EDITOR Marcy Atarod Emily Brodsky, UC Santa Cruz William Kaelin Jr., Dana-Farber Susannah Scott, UC Santa Barbara
DESIGNER Christina Aycock SENIOR SCIENTIFIC ILLUSTRATOR Valerie Altounian SCIENTIFIC ILLUSTRATORS Austin Fisher, Kellie Holoski, Ron Brookmeyer, UCLA (S) Daniel Kammen, UC Berkeley Anuj Shah, U. of Chicago
Ashley Mastin INTERACTIVE GRAPHICS EDITOR Kelly Franklin SENIOR GRAPHICS SPECIALISTS Holly Bishop, Nathalie Cary SENIOR PHOTO Christian Büchel, UKE Hamburg Kisuk Kang, Seoul Nat. U. Vladimir Shalaev, Purdue U.
EDITOR Charles Borst PHOTO EDITOR Kaitlyn Dolan SENIOR PODCAST PRODUCER Sarah Crespi VIDEO PRODUCER Meagan Cantwell Dennis Burton, Scripps Res. Sabine Kastner, Princeton U. Jie Shan, Cornell U.
SOCIAL MEDIA STRATEGIST Jessica Hubbard SOCIAL MEDIA PRODUCER Sabrina Jenkins WEB DESIGNER Jennie Pajerowski Carter Tribley Butts, UC Irvine V. Narry Kim, Seoul Nat. U. Beth Shapiro, UC Santa Cruz
György Buzsáki, Robert Kingston, Harvard Med. Jay Shendure, U. of Wash.
NYU School of Med. Nancy Knowlton, Steve Sherwood,
Chief Executive Officer and Executive Publisher Sudip Parikh Mariana Byndloss, Smithsonian Institution U. of New South Wales
Publisher, Science Family of Journals Bill Moran Vanderbilt U. Med. Ctr. Etienne Koechlin, Brian Shoichet, UCSF
DIRECTOR, BUSINESS SYSTEMS AND FINANCIAL ANALYSIS Randy Yi DIRECTOR, BUSINESS OPERATIONS & ANALYSIS Eric Knott DIRECTOR OF Annmarie Carlton, UC Irvine École Normale Supérieure Robert Siliciano,
ANALYTICS Enrique Gonzales MANAGER, BUSINESS OPERATIONS Jessica Tierney MANAGER, BUSINESS ANALYSIS Cory Lipman Simon Cauchemez, Inst. Pasteur Alex L. Kolodkin, Johns Hopkins U. JHU School of Med.
BUSINESS ANALYSTS Kurt Ennis, Maggie Clark FINANCIAL ANALYST Isacco Fusi BUSINESS OPERATIONS ADMINISTRATOR Taylor Fisher Ling-Ling Chen, SIBCB, CAS LaShanda Korley, U. of Delaware Lucia Sivilotti, UCL
SENIOR PRODUCTION MANAGER Jason Hillman SENIOR MANAGER, PUBLISHING AND CONTENT SYSTEMS Marcus Spiegler Wendy Cho, UIUC Julija Krupic, U. of Cambridge Emma Slack, ETH Zürich & U.
CONTENT OPERATIONS MANAGER Rebecca Doshi SENIOR CONTENT & PUBLISHING SYSTEMS SPECIALIST Jacob Hedrick Ib Chorkendorff, Denmark TU Paul Kubes, U. of Calgary of Oxford
SENIOR PRODUCTION SPECIALIST Kristin Wowk PRODUCTION SPECIALISTS Kelsey Cartelli, Audrey Diggs DIGITAL PRODUCTION MANAGER Karlene Cimprich, Stanford U. Chris Kuzawa, Northwestern U. Richard Smith, UNC (S)
Lisa Stanford CONTENT SPECIALIST Kimberley Oster ADVERTISING PRODUCTION OPERATIONS MANAGER Deborah Tompkins DESIGNER, James J. Collins, MIT Laura Lackner, Northwestern U. John Speakman, U. of Aberdeen
CUSTOM PUBLISHING Jeremy Huntsinger SR. TRAFFIC ASSOCIATE Christine Hall SPECIAL PROJECTS ASSOCIATE Sarah Dhere Robert Cook-Deegan, Gabriel Lander, Scripps Res. (S) Tara Spires-Jones, U. of Edinburgh
Arizona State U. Mitchell A. Lazar, UPenn Allan C. Spradling,
ASSOCIATE DIRECTOR, BUSINESS DEVELOPMENT Justin Sawyers GLOBAL MARKETING MANAGER Allison Pritchard DIGITAL MARKETING
Virginia Cornish Columbia U. Hedwig Lee, Duke U. Carnegie Institution for Sci.
MANAGER Aimee Aponte JOURNALS MARKETING MANAGER Shawana Arnold MARKETING ASSOCIATES Aaron Helmbrecht, Ashley Carolyn Coyne, Duke U. Luis Liz-Marzán, CIC biomaGUNE V. S. Subrahmanian,
Hylton, Mike Romano, Lorena Chirinos Rodriguez, Jenna Voris SENIOR DESIGNER Kim Huynh Roberta Croce, VU Amsterdam Omar Lizardo, UCLA Northwestern U.
DIRECTOR AND SENIOR EDITOR, CUSTOM PUBLISHING Sean Sanders ASSISTANT EDITOR, CUSTOM PUBLISHING Jackie Oberst Christina Curtis, Stanford U. Jonathan Losos, Ira Tabas, Columbia U.
PROJECT MANAGER Melissa Collins Ismaila Dabo, Penn State U. Wash. U. in St. Louis Eriko Takano, U. of Manchester
DIRECTOR, PRODUCT & PUBLISHING DEVELOPMENT Chris Reid DIRECTOR, BUSINESS STRATEGY AND PORTFOLIO MANAGEMENT Sarah Whalen Jeff L. Dangl, UNC Ke Lu, Inst. of Metal Res., CAS A. Alec Talin, Sandia Natl. Labs
ASSOCIATE DIRECTOR, PRODUCT MANAGMENT Kris Bishop PRODUCT DEVELOPMENT MANAGER Scott Chernoff PUBLISHING TECHNOLOGY Chiara Daraio, Caltech Christian Lüscher, U. of Geneva Patrick Tan, Duke-NUS Med. School
Nicolas Dauphas, U. of Chicago Jean Lynch-Stieglitz, Sarah Teichmann,
MANAGER Michael Di Natale SR. PRODUCT ASSOCIATE Robert Koepke PRODUCT ASSOCIATE Caroline Breul, Anne Mason SPJ ASSOCIATE
Frans de Waal, Emory U. Georgia Inst. of Tech. Wellcome Sanger Inst.
MANAGER Samantha Bruno Fuller SPJ ASSOCIATE Casey Buchta
Claude Desplan, NYU David Lyons, U. of Edinburgh Rocio Titiunik, Princeton U.
MARKETING MANAGER Kess Knight BUSINESS DEVELOPMENT MANAGER Rasmus Andersen SENIOR INSTITUTIONAL LICENSING MANAGER Sandra DÍaz, Fabienne Mackay, Shubha Tole,
Ryan Rexroth INSTITUTIONAL LICENSING MANAGER Marco Castellan, Claudia Paulsen-Young SENIOR MANAGER, INSTITUTIONAL LICENSING U. Nacional de CÓrdoba QIMR Berghofer Tata Inst. of Fundamental Res.
OPERATIONS Judy Lillibridge SENIOR OPERATIONS ANALYST Lana Guz SYSTEMS & OPERATIONS ANALYST Ben Teincuff FULFILLMENT ANALYST Samuel Díaz-Muñoz, UC Davis Zeynep Madak-Erdogan, UIUC Maria-Elena Torres Padilla,
Aminta Reyes Ulrike Diebold, TU Wien Anne Magurran, U. of St. Andrews Helmholtz Zentrum München
DIRECTOR, GLOBAL SALES Tracy Holmes US EAST COAST AND MID WEST SALES Stephanie O'Connor US MID WEST, MID ATLANTIC AND Stefanie Dimmeler, Ari Pekka Mähönen, U. of Helsinki Kimani Toussaint, Brown U.
SOUTH EAST SALES Chris Hoag US WEST COAST SALES Lynne Stickrod ASSOCIATE DIRECTOR, ROW Roger Goncalves SALES REP, ROW Goethe-U. Frankfurt Asifa Majid, U. of Oxford Barbara Treutlein, ETH Zürich
Sarah Lelarge SALES ADMIN ASSISTANT, ROW Victoria Glasbey DIRECTOR OF GLOBAL COLLABORATION AND ACADEMIC PUBLISHING RELATIONS, Hong Ding, Inst. of Physics, CAS Oscar Marín, King’s Coll. London Jason Tylianakis, U. of Canterbury
ASIA Xiaoying Chu ASSOCIATE DIRECTOR, INTERNATIONAL COLLABORATION Grace Yao SALES MANAGER Danny Zhao MARKETING MANAGER Dennis Discher, UPenn Charles Marshall, UC Berkeley Wim van der Putten, Netherlands
Kilo Lan ASCA CORPORATION, JAPAN Rie Rambelli (Tokyo), Miyuki Tani (Osaka) Jennifer A. Doudna, Christopher Marx, U. of Idaho Inst. of Ecology
UC Berkeley David Masopust, U. of Minnesota Matthew Vander Heiden, MIT
DIRECTOR, COPYRIGHT, LICENSING AND SPECIAL PROJECTS Emilie David RIGHTS AND PERMISSIONS ASSOCIATE Elizabeth Sandler
Ruth Drdla-Schutting, Geraldine Masson, CNRS Ivo Vankelecom, KU Leuven
LICENSING ASSOCIATE Virginia Warren CONTRACT SUPPORT SPECIALIST Michael Wheeler
Med. U. Vienna C. Robertson McClung, Judith Varner, UC San Diego
Raissa M. D'Souza, UC Davis Dartmouth Henrique Veiga-Fernandes,
MAIN HEADQUARTERS EDITORIAL PRODUCT ADVERTISING AAAS BOARD OF DIRECTORS Bruce Dunn, UCLA Rodrigo Medellín, Champalimaud Fdn.
Science/AAAS science_editors@[Link] & CUSTOM PUBLISHING CHAIR Susan G. Amara William Dunphy, Caltech U. Nacional Autónoma de México Reinhilde Veugelers, KU Leuven
1200 New York Ave. NW [Link]/ Scott Edwards, Harvard U. C. Jessica Metcalf, Princeton U. Bert Vogelstein, Johns Hopkins U.
NEWS PRESIDENT Gilda A. Barabino
Washington, DC 20005 science_news@[Link] products-services PRESIDENT-ELECT Keith Yamamoto
Todd Ehlers, U. of TÜbingen Baoxia Mi, UC Berkeley Julia Von Blume, Yale School of Med.
science_advertising@[Link] Nader Engheta, UPenn Tom Misteli, NCI, NIH David Wallach, Weizmann Inst.
SCIENCE INTERNATIONAL INFORMATION FOR AUTHORS TREASURER Carolyn N. Ainslie
Karen Ersche, U. of Cambridge Alison Motsinger-Reif, Jane-Ling Wang, UC Davis (S)
Clarendon House [Link]/authors/ CLASSIFIED ADVERTISING CHIEF EXECUTIVE OFFICER Beate Escher, UFZ & U. of Tübingen NIEHS, NIH (S) Jessica Ware,
Clarendon Road science-information-authors [Link]/ Sudip Parikh Barry Everitt, U. of Cambridge Danielle Navarro, Amer. Mus. of Natural Hist.
Cambridge, CB2 8FH, UK REPRINTS AND PERMISSIONS science-careers BOARD Cynthia M. Beall Vanessa Ezenwa, U. of Georgia U. of New South Wales David Waxman, Fudan U.
[Link]/help/ advertise@[Link] Ann Bostrom Toren Finkel, U. of Pitt. Med. Ctr. Daniel Nettle, Newcastle U. Chris Wikle, U. of Missouri (S)
SCIENCE CHINA
reprints-and-permissions Janine Austin Clayton Gwenn Flowers, Simon Fraser U. Daniel Neumark, UC Berkeley Terrie Williams, UC Santa Cruz
Room 1004, Culture Square JOB POSTING CUSTOMER SERVICE
[Link] Kaye Husbands Fealing
MEDIA CONTACTS Natascha Förster Schreiber, Beatriz Noheda, U. of Groningen Ian A. Wilson, Scripps Res. (S)
No. 59 Zhongguancun St. Maria M. Klawe MPI Extraterrestrial Phys. Helga Nowotny, Hao Wu, Harvard U.
scipak@[Link]
Haidian District, Beijing, 100872 support@[Link] Jane Maienschein Peter Fratzl, MPI Potsdam Vienna Sci. & Tech. Fund Li Wu, Tsinghua U.
MULTIMEDIA CONTACTS
SCIENCE JAPAN MEMBERSHIP AND INDIVIDUAL Robert B. Millard Elaine Fuchs, Rockefeller U. Pilar Ossorio, U. of Wisconsin Wei Xie, Tsinghua U.
SciencePodcast@[Link]
SUBSCRIPTIONS Jay Gallagher, U. of Wisconsin Andrew Oswald, U. of Warwick Benjamin Youngblood, St. Jude
ASCA Corporation ScienceVideo@[Link] Babak Parviz
[Link]/subscriptions Daniel Geschwind, UCLA Isabella Pagano, Yu Xie, Princeton U.
Sibaura TY Bldg. 4F, 1-14-5 INSTITUTIONAL SALES William D. Provine
Ramon Gonzalez, Istituto Nazionale di Astrofisica Jan Zaanen, Leiden U.
Shibaura Minato-ku AND SITE LICENSES MEMBER BENEFITS Juan S. Ramírez Lugo U. of South Florida Elizabeth Levy Paluck, Kenneth Zaret,
Tokyo, 108-0073 Japan [Link]/librarian [Link]/membership/benefits Susan M. Rosenberg Sandra González-Bailón, UPenn Princeton U. UPenn School of Med.
Science serves as a forum for discussion of important issues related to the advancement of science by publishing material on Gillian Griffiths, U. of Cambridge Jane Parker, MPI Cologne Lidong Zhao, Beihang U.
which a consensus has been reached as well as including the presentation of minority or conflicting points of view. Accordingly, Nicolas Gruber, ETH Zürich Giovanni Parmigiani, Bing Zhu, Inst. of Biophysics, CAS
all articles published in Science—including editorials, news and comment, and book reviews—are signed and reflect the individual Hua Guo, U. of New Mexico Dana-Farber (S) Xiaowei Zhuang, Harvard U.
views of the authors and not official points of view adopted by AAAS or the institutions with which the authors are affiliated. Taekjip Ha, Johns Hopkins U. Daniel Pauly, U. of British Columbia Maria Zuber, MIT

682 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

EDITORIAL

Shockley was a racist and eugenicist

T
his week’s issue on the 75th anniversary of the overtly in favor of eugenics as was Shockley, but he was a
transistor describes a triumph of both basic and strong advocate for genetic determinism, even claiming
applied science. What started out as studies on at the behest of the cigarette industry that tobacco itself
the fundamental physics of silicon led to the was not harmful because genetics determined whether
device that makes it possible to read this article smokers would ultimately contract lung cancer.
online. The coinventor of the transistor, William Following Shockley’s death in 1989, Nature correctly
Shockley, who along with John Bardeen and Wal- called out his racism in an obituary, but then published
ter Brattain won the 1956 Nobel Prize in Physics, is cor- a letter from Seitz defending Shockley and claiming H. Holden Thorp
rectly recognized as a primary architect of the comput- that the reason Shockley became a eugenicist was be-
Editor-in-Chief,
er age. Gordon Moore (cofounder of Intel Corporation) cause of physical trauma he experienced in a near-fatal
Science journals.
famously said that Shockley put the silicon in “Silicon car accident. When Science wrote about this dustup,
hthorp@[Link];
Valley.” Appallingly, Shockley devoted the latter part of it referred to Shockley’s ideas as merely “unpopular”
his life to promoting racist views, arguing that higher and “extremely controversial.” It then ran a letter from @hholdenthorp
IQs among Blacks were correlated with higher extents an even more notorious eugenicist, J. Philippe Rush-
of Caucasian ancestry, and advo- ton, who argued that by merely
cating for voluntary sterilization covering the disagreement at
of Black women. At the time, Nature, Science was deliver-
Science did not condemn Shock-
ley for what he was: a charlatan
“The process ing an “ad hominem attack.” In
addition to an ill-advised deci-
who used his scientific creden-
tials to advance racist ideology.
of science is one of sion to publish Rushton’s letter,
Science posted a response saying,
The failure of Science to con-
demn Shockley began in 1968,
continual revision, “no criticism of Shockley was in-
tended.” Yikes.
when it published a letter la-
menting the fact that he was
but it’s also Looking back, it’s clear that
what was intended as an attempt
prohibited from speaking at the
Polytechnic Institute of Brook-
one that must have to make room for dissent and
discussion only served to abet
lyn. The letter repeated the fa-
miliar trope that Shockley was a conscience.” Shockley and his cohorts in their
effort to build support for eugen-
simply asking questions about ics. Science gave them a platform
the role of race in intelligence. and inadequate scorn. The les-
But Shockley had no scientific basis for doing so, he was son is that we at Science need to make more effort to
not submitting peer-reviewed papers on the topic, and think about everything that we do, not only from the
most importantly, he was using his ideas as the basis standpoint of communicating science to the public, but
for promoting eugenics. Such a debate had no place in also as an organization that above all, supports all of
this journal. humanity. The process of science is one of continual
Shockley was part of a cadre of physicists who ad- revision, but it’s also one that must have a conscience.
vanced ideas outside of their area of expertise to pro- It was only a few months ago, in a commentary on
mote a right-wing agenda. He was a close friend of racism in science by Ebony Omotola McGee, that Shock-
Frederick Seitz—president of both the National Acad- ley was described in our pages in the terms he deserved.
emy of Sciences and Rockefeller University—who, But as recently as 2001, Science described him simply as
following a career in physics, became a purveyor of mis- a “transistor inventor and race theorist.” That won’t cut
information on tobacco, nuclear weapons, and climate it anymore. As of today, a link to this editorial will ap-
change. Like Shockley, Seitz carried out his nonphysics pear along with any mention of Shockley in this journal.
work through op-eds and conservative think tanks, not Make no mistake. Shockley was a racist. Shockley was
through the accepted mechanism of peer review that he a eugenicist. That’s all.
used in doing physics. Seitz was not, at least publicly, as –H. Holden Thorp
PHOTO: CAMERON DAVIDSON

10.1126/science.adf8117

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 683

1118Editorial_16203146.indd 683 11/15/22 4:03 PM

Advertorial

The multi(dimension)verse of computing

Communications technology is changing the way we live at an ever-faster pace. announced its “Q-LEAP” initiative, committing hundreds of millions more into
Massive amounts of data can be transferred around the world in a flash, and research on hardware and software related to quantum computing. All this is in
mathematical problems that have eluded the best minds can now be solved in anticipation of a market size that is estimated to increase almost 10-fold in the
mere months by supercomputing. The arrival of quantum computing, however, next 5 years.
will make these technologies archaic. If its expected potential is realized, then The impressive contributions by Japanese government and industry
quantum computing will be as superior to supercomputing as the works of are matched by the discoveries made through Japan’s academia–industry
Shakespeare compared to the output of a monkey at a typewriter. collaborations. Building on its recognized strengths in semiconductors
Supercomputing remains the pinnacle of digital computing; it works by taking and high-performance computing, Japan has achieved a number of firsts in
classical bits of information and, with enough energy, making an extraordinary quantum computing. Quantum key distribution and quantum annealing are
number of calculations to find a solution to complex problems. However, just two of the areas it excels in as it advances into the commercial market,
whereas a classical bit only takes one of two states (either a 0 or a 1), a qubit, pushed by companies such as Fujitsu, Toshiba, and NEC, among many others.
the information basis of quantum computing, can represent a superposition Commercialization is also being facilitated by organizations like the Quantum ICT
of the two states. This distinction allows quantum computers to analyze Forum, which is bringing the best minds in the field to translate technologies and
multidimensional spaces to recognize patterns that simply cannot be observed guide policy.
by classical computers. Not only does this new perspective reduce the time Quantum computers will not be in commission for another decade or two
needed to calculate a solution, but it also expands by leaps and bounds the types because of the many functional challenges that remain, such as the short
of problems that can be solved. coherence time of qubits and the large error rates in quantum operations.
The applications are endless and will impact every area of our lives. These Ironically, these problems will be solved iteratively, much in the way digital
computers will be better at designing drugs, planning global supply chains, computing works. In the meantime, Japan is already producing quantum-

PHOTO: © DMITRIY RYBIN/[Link]

reducing financial risk, and predicting the weather. They could even reveal the inspired digital technologies that solve many of the combinatorial optimization
elusive secrets of the universe. problems eluding even the best supercomputers. Although not as fast as true
Quantum computing may not get the same attention as the space race did quantum computers, these quantum-based devices can reduce problem-
many decades ago, but countries of all sizes recognize the implications of this solving time from months to hours and from days to seconds. Moreover, they are
technology and are competing to take the lead. Included in this list is Japan, available right now.
which aims to announce four new quantum computing research centers by the For decades, Japan has been a herald of the future. From its bullet trains
end of the fiscal year. Japanese government funding agencies have invested to its robot-staffed hotels, the country has played a major role in advancing
more than USD $250 million in this field in the past 15 years, and Japanese technological progress on the world stage. Now its next performance is about to
industry is matching these amounts. Moreover, in 2018, the government debut, with a new star—quantum computing.

[Link] 684 11/10/22 8:21 AM

Produced by the Science

Director of the Quantum ICT Forum Akihisa Tomita is helping to usher in quantum technology.

Gathering the best minds in quantum technology

The strength of Japan’s research in quantum information and communication Bringing more people to quantum ICT
technology (ICT) can be seen all across the country. In terms of R&D, Japan’s Officially, the Quantum ICT Forum was founded in 2019, but its history can be
universities are at the top of the field and its companies are at the forefront of traced to 2001, when Japanese researchers were rapidly making several major
developing key technologies such as superconducting qubits, quantum annealing, discoveries in the field. Since then, the Forum has undergone a number of
and optical lattice clocks. However, so long as this progress remains scattered, changes to reflect how quantum ICT has progressed, and it now comprises more
its benefits to society will be limited. It was in response to this problem that the than 100 members, including both individuals and companies, who lead the
Quantum ICT Forum was founded, to quicken Japan’s adoption of quantum ICT. quantum-related R&D. The Forum hopes to increase its membership, and many
History has shown that societies have gone through four great transformations, of its activities—from its regular public seminars to interviews with leaders in the
beginning with the hunter-gatherer society and moving through the agricultural, field—are dedicated to doing so.
industrial, and most recently, information societies. Japan is now preparing for The Forum is an ideal organization for people with creative ideas seeking
Society 5.0, which aims to connect cyberspace with physical space. This will networking opportunities in Japan, looking for research partners, and pursuing
only happen through a new generation of ICT, one that will depend heavily on funding for their ventures. It also provides an opportunity to have real influence
quantum-based technologies. on the direction of quantum ICT. For example, the Forum’s input is requested at
“Our goal is to bring researchers and industry together so that we can use the task force meeting on the Cross-Ministerial Strategic Innovation Promotion
quantum technologies for a healthier, more productive society,” says Akihisa Program, a Japanese government initiative that has identified key technology
Tomita, director of the Quantum ICT Forum. fields—such as quantum ICT—that have a strong potential to grow the economy.
However, unlike academic societies, the Forum does more than invite Tomita stresses that although the Forum is focused currently on Japan, it
researchers to present their studies. It is designed to promote commercialization also provides a place for international partners to understand the talent and
so that Japan stays at the forefront of ICT. dedication that Japan has to offer to quantum ICT.
In particular, the Forum has identified three sectors—high-speed “It is important for Japan not only to have a strong environment for quantum
computers, highly sensitive and accurate sensors, and faster and more secure ICT; we also want the world to know how much investment the country is making
communications—as having the biggest impact on ICT in the near future. Thus, it in this field and how dedicated we are to it,” he says.
has formed three committees: one on quantum computing, another on quantum
Sponsored by
sensing, and the last on quantum key distribution.
These committees serve many purposes that benefit academic, industrial,
PHOTO: PROVIDED BY QUANTUM ICT FORUM

and public interests. First and foremost, they promote collaborative research.
Second, they bring consensus to standards and provide guidance on government
policy. Finally, they promote public awareness through information dissemination,
whether in the form of exhibitions, seminars, newsletters, or other mediums.
“Each committee is looking at best practices to keep Japan at the leading
edge of quantum ICT. The most important efforts involve translating research
into valuable customer products and assisting with policy making so that society
benefits,” says Tomita.

[Link] 685 11/10/22 8:21 AM

Advertorial

Scientific discoveries for technological advances

Throughout its history, Toshiba technology has been found everywhere, of parallel processing distinguishes simulated bifurcation machines from other
ranging from everyday items in the home to social infrastructure to protect simulated annealers used to solve optimization problems.
entire nations. This diversity reflects not only the high quality of its products, “In simulated annealers, we have to update variables one by one. But with
but also the commitment and excellence of its research, which pushes for this algorithm we can update the variables in parallel, which is a tremendous
discovery in both applied technology and pure science. Now, Toshiba’s scientific advantage,” he added. Indeed, his algorithm has set world records for
breakthroughs are impacting computer science, setting the stage for faster, computation times, and problems that would take more than a year to solve
more accurate computational solutions to businesses’ most difficult problems. using standard central processing unit (CPU)-based simulated annealers can be
completed in less than an hour if using the simulated bifurcation algorithm.7
Protecting data Thus, industries anticipating the benefits of quantum computing—including
Toshiba is a global company driving innovation in its many research centers finance, logistics, and pharmaceuticals—no longer need to wait when working
across the world, including its Cambridge Research Laboratory in the United with Toshiba.
Kingdom, founded more than 30 years ago. It was there that Toshiba made its QKD and the simulated bifurcation algorithm are just two examples of
first significant innovation in quantum cryptography.1 With quantum computing Toshiba’s efforts in quantum computing and another reminder of how its
becoming a reality and the protection of data becoming more urgent, Toshiba research breaks new ground in science and technology. The drive for this
is leveraging its initial investment in the field to generate estimated revenue progress, however, comes because Toshiba is committed to people and to the
exceeding USD 3 billion by 2030. future.
“Data vulnerability is a major concern in many sectors. Our quantum key
distribution [QKD] platform is being used by governments, health care centers, References
and financial companies to protect data,” says Yoshimichi Tanizawa, chief 1. Z. Yuan et al., Science 295, 102–105, [Link]
science.1066790.
research scientist at Toshiba.
2. [Link]
Using its QKD, Toshiba has repeatedly set world records for the fastest 3. [Link]
data transmission time over long distances, drawing attention from many Toshiba-and-Ciena-Build-the-First-Quantum-Key-Distribution-Network-Used-to-Secure-
organizations for data encryption. Mission-Critical-Blockchain-Application/[Link]
4. [Link]
In Japan, Toshiba, Tohoku University, and NICT demonstrated a secure genome
5. [Link]
analysis data backup to multiple sites with Toshiba’s QKD system. 2 In the United 6. H. Goto, K. Tatsumura, A. R. Dixon, Sci. Adv. 5, eaav2372 (2019), [Link]
States, JPMorgan Chase, Toshiba and Ciena demonstrated QKD system securing sciadv.aav2372
a peer-to-peer blockchain network.3 In London, BT and Toshiba, along with EY, 7. H. Goto et al., Sci. Adv. 7, eabe7953 (2021), [Link]
launched the trial of world first commercial quantum-secured metro network,
Sponsored by
helping secure information transmission across the city’s data networks.4 And
in South Korea, Toshiba group and KT is developing a long
distance hybrid QKD network, enabling the safe transfer of data
between the country’s biggest cities, Seoul and Busan.5
Sponsored by
Changing science and technology
While Toshiba research is designed to push new frontiers in
technologies, it is also changing our understanding of the
science behind these technologies. Such is the case with its
simulated bifurcation machine. In their efforts to develop a
quantum annealer, which is considered the gold standard
for solving combinatorics optimization problems, Toshiba
scientists discovered that the principle of quantum annealing
called “adiabatic evolution” could be applied to classical
nonlinear Hamiltonian systems. The result was the simulated
bifurcation algorithm,6 which incorporates bifurcation
phenomena, adiabatic evolution, and chaos theory, allowing
PHOTO: PROVIDED BY TOSHIBA

digital computers to quickly solve the same complex problems

thought exclusive to the domain of quantum computing.
“This is a new idea in computer science. This algorithm can
be used by classical computers to achieve high performance
from parallel processing,” said Toshiba chief research scientist
Hayato Goto, who discovered the algorithm. The incorporation
Simulated bifurcation machine

[Link] 686 11/10/22 8:21 AM

Produced by the Science

A company of firsts in quantum technology

Throughout its history, NEC—originally founded as the Nippon Electric The next critical step for quantum technology is quantum computing. In
Company more than 100 years ago—has proudly accomplished a number 1999, NEC succeeded in demonstrating the basic operation of the world’s first
of firsts in science and engineering. While its initial innovations were in superconducting-based quantum bits suitable for integration, and has continued
telecommunication technologies, its R&D has positioned the company as a to research and develop quantum computers for more than 20 years. 2 Currently,
global leader in communications, with developments in semiconductors, mobile we are developing a quantum annealing machine dedicated to combinatorial
communications, and now the next great frontier—quantum technology. optimization problems using Josephson parametric oscillators, as well as a
Despite being a multinational company with more than 100,000 employees, fault-tolerant quantum computer that no one has been able to produce yet. Since
NEC conducts research with an academic approach in which curiosity and 2019, we have advanced quantum technology by starting to verify the application
exploration are paramount. It is this attitude that has led NEC to numerous of solving customers’ combinatorial optimization problems with simulated
quantum-computing firsts—most notably, in 1999, being the first in the world to annealing using vector machines, which are supercomputers that NEC has
report a solid-state qubit. accumulated over many years.

The evolution of quantum technology at NEC Great gains through great ambition
Progress in quantum technology is happening in stages. The first is quantum NEC’s work demonstrates the incremental process required to achieve true
key distribution (QKD), which will revolutionize cryptography, securing data quantum technology. Its competence is on par with those who are developing
that implements this technology while enabling safety guaranteed by quantum the “quantum internet.” However, as with many of the grander problems in
mechanics and information theory. science today, no one organization can solve them all. While NEC is at the
NEC’s quantum cryptography efforts have been dedicated on the forefront of the evolution of quantum technology, the company is therefore
commercialization and social implementation of the BB84 system, which partnering with many university laboratories and companies, building a wide
excels in medium-to-long-distance communication, and currently center on network of quantum technology experts who are dedicated to realizing this great
researching the continuous variable QKD (CV-QKD) system for widespread use, leap in quantum ability.
with the aim of further expanding its applications. NEC has been conducting This strategy, Nakamura explains, is in place because NEC’s ultimate goal
demonstration experiments in various fields, such as health care and finance, goes beyond individual applications such as QKD or quantum computing. “Few
with a secure encryption method that will not be deciphered even when companies have the diversity in quantum technology that we have. And few
quantum computers eventually appear. Many enterprises with a long-term companies have the partners we have,” he says. “Ultimately, our technology
perspective on their business have been interested in NEC’s quantum aims to solve today’s most difficult challenges and meet society’s most pressing
cryptography technology as a powerful measure to protect their critical needs. Quantum technology definitely has such potential.”
information in the future.
“As a company with strong research and development, we are committed to References
innovation and translating that innovation into customer solutions,” says Yuichi 1. K. Matsumoto et al., Phys Rev. A 105, 023110 (2022).
Nakamura, an NEC executive professional and former vice president. 2. Y. Nakamura, Y. A. Pashkin, J. S. Tsai, Nature 398, 786–788, [Link]
Another essential step toward quantum technology is to control quantum articles/19718.
states with accurate clock timing despite environmental disturbances. This PHOTO: PROVIDED BY POSTECH
PHOTO: PROVIDED BY NEC

requires quantum clocks that operate accurately against external turbulence.

Researching the construction of these clocks, NEC has reported specific Sponsored by
coherent-population-trapping resonances.1 Advances in this technology will
have profound effects on the robustness of our communications systems by
protecting them against natural disasters and malicious attacks.

[Link] 687 11/10/22 8:21 AM

Advertorial

The complete array of advanced computing

Since its founding, the Tokyo-based information company Fujitsu has been using Operational at room temperature and using traditional racks, Fujitsu’s Digital
communications technologies to build a more sustainable world. Critical to this Annealer is already being employed in multiple industries, bringing radical
aim is staying at the forefront of computing. In the past 20 years, this has meant solutions and great savings. Successful use cases include new drug design that
putting intensive efforts into high-performance computing (HPC) and quantum reduces the delivery of candidate molecules from several months to several
computing in order to offer the most comprehensive toolkit available for problems weeks, warehouse management that reduces the distances travelled for parts by
demanding computational solutions. 45%, and traffic optimization that reduces car congestion by 40%.
Its versatility has also been demonstrated
Collaboration is key in response to COVID-19. A critical problem in
Fujitsu recognizes that progress comes faster by managing the pandemic has been the effective
working with partners who hold complementary distribution of personal protective equipment,
strengths. Its supercomputer Fugaku—which it such as masks and N95 respirators, which
completed last year in collaboration with RIKEN, were suddenly in major demand throughout
Japan’s largest research institute—held the top the world when the pandemic started, leading
spot in HPC rankings for four consecutive terms. to huge supply shortfalls. It was validated
Several groups in Japan are using Fujitsu’s by a CRADA report that the Digital Annealer
supercomputing technologies to interpret optimized resource allocation and minimized
tsunami data for evacuation responses and the time data traveled to emerging hotspots
protein bindings for drug design. in the United States in mere seconds, as
In the promising field of quantum computing, compared to the days or weeks required by
Fujitsu has teamed with RIKEN to establish traditional computing algorithms. It was also
the RIKEN RQC-Fujitsu Collaboration Center, used to maximize the number of people who
which aims to develop quantum computers could attend professional football matches
using superconducting qubits to solve even in Germany while respecting social distance
more difficult societal problems. Parallel to policies, allowing communities to return to
Superconducting 64-qubit chip IMAGE: © RIKEN CENTER FOR QUANTUM COMPUTING
this, a collaboration with Delft University of normal life faster.
Technology in the Netherlands is looking at the advantages of diamond-spin
qubits.1 Where superconducting qubits have excellent scalability so far, diamond- The future of computing
spin qubits can hold quantum information relatively long. A major concern in Quantum computing will have profound effects that resonate throughout society.
quantum computing is the sensitivity of qubits to environmental interference, However, as powerful as this technology may one day become, it will not solve all
including noise. Fujitsu’s collaboration with Delft University also recently revealed the problems by itself. In other words, quantum computing will not replace HPC,
a solution to this problem by demonstrating a fault-tolerant qubit operation using but as a very important new tool, work with other computing technologies to solve
a quantum processor based on spin qubits in diamond. 2 difficult societal problems in the future.
While its projects with RIKEN and Delft are focused on hardware, Fujitsu is “Quantum computing is just one type of computing at which we excel. Our
teaming with Osaka University in Japan to develop software for fault-tolerant research is motivated by society’s needs. In the future, our offerings will include
quantum computing in combination with Fujitsu hardware for higher-level supercomputing, quantum-inspired computing, and quantum computing,”
quantum computers. The collaboration will take the form of a new research says Sato.
division on the university campus.
While working on superconducting and diamond-spin quantum computers, References
Fujitsu also developed the world’s fastest 36-qubit quantum computer simulator 1. R. Ishihara et al, 2021 IEEE International Electron Devices Meeting (IEDM),
using its supercomputing technologies. [Link]
“In quantum computing, there is no best technology. Fujitsu believes it must 2. M. H. Abobeih et al., Nature 606, 884–889 (2022), [Link]
explore multiple channels to reach the fullest potential,” says Shintaro Sato, head
of the Quantum Laboratory at Fujitsu. Sponsored by

Quantum-inspired computers
Fujitsu surveys have found that a large majority of businesses are eager to
benefit from quantum computing. However, many experts predict that the first
commercially practical quantum computers are still years away. As an alternative,
researchers are developing quantum-inspired computing technologies. The best
example at Fujitsu is its Digital Annealer.

[Link] 688 11/10/22 8:21 AM

Subscribe to News from Science for unlimited
access to authoritative, up-to-the-minute news
on research and science policy.

[Link]/NewsFromScience

[Link] 689 11/10/22 8:21 AM

NEWS
IN BRIEF
Edited by
Jeffrey Brainard

CLIMATE POLICY

Carbon emissions increase—as do ways to track them

C
arbon dioxide emissions from burning fossil track, verify, and regulate greenhouse gases. One tool,
fuels are on track to rise 1% this year from the developed by the Climate TRACE coalition, uses satel-
2021 level, making it harder for many nations lite imagery and machine-learning algorithms to detect
to reach their goal of achieving net-zero emis- and measure emissions from 72,000 sources, including
sions by 2050, scientists from the Global Carbon power plants. Separately, the United Nations unveiled
Project said last week. They cited an easing of the Methane Alert and Response System, which will use
pandemic precautions, including increased air data from new satellites capable of spotting large leaks
travel, as one reason for the rise. Most researchers say of methane, a powerful greenhouse gas. The announce-
the world is unlikely to meet the net-zero ments came as politicians at the U.N. climate
Transmission lines carry
goals and limit global warming to 1.5°C by electricity from
conference in Egypt debated whether and
2050. But two new tools announced last week a coal-fired power plant how wealthy countries should pay for climate-
will aid efforts by improving the ability to in Weisweiler, Germany. related damages to low-income nations.

PHOTO: INA FASSBENDER/AFP/GETTY IMAGES

database on U.S. water management and on administrative leave starting in 2018. Her
Wrongful dismissal case is settled providing information to a Chinese govern- case mobilized the Chinese American com-
RESEARCH SECURITY | The U.S. government official while working at the National munity, which saw it as racial profiling years
ment has agreed to pay hydrologist Xiafen Weather Service. Chen denied acting before former President Donald Trump’s
“Sherry” Chen $1.8 million to settle a wrong- improperly, and the government dropped administration’s China Initiative against
ful dismissal lawsuit that stemmed from a the charges in 2015. The Department of Chinese espionage drew similar criti-
failed federal prosecution alleging threats to Commerce, which includes the weather cism. In the 10 November settlement, the
national security. In 2014, the government service, fired her in 2016, but after a review Department of Commerce does not admit
accused her of tapping into a restricted board ruled the termination illegal, put her any wrongdoing but agrees to meet with

690 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NewsInBrief_16210255.indd 690 11/15/22 5:23 PM

Chen to discuss how she was treated and DEMOGRAPHY
to issue a letter praising her work. Chen
agreed to retire by the end of this year. Her World population hits 8 billion as growth slows
legal team called the settlement “a great

E
arth’s population has reached a milestone by surpassing 8 billion people, the United
blow … against bigotry and for the rights Nations said this week. But the rate of increase is falling, and global population may
of Asian Americans.” begin to decline late in the century after topping out at about 10.4 billion, according to
the U.N. Population Division. Its World Population Prospects 2022 report notes that two-
thirds of the global population already lives in a country or area where lifetime fertility
China trims some COVID-19 rules is below 2.1 births per woman, roughly the level required for zero growth for a population
| China last week
P U B L I C H E A LT H with low mortality. More than half of the projected increase in global population between
announced 20 revisions of pandemic control now and 2050 will be concentrated in just eight countries: the Democratic Republic of the
and prevention measures that somewhat Congo, Egypt, Ethiopia, India, Nigeria, Pakistan, the Philippines, and Tanzania.
ease the burden on its weary, frustrated pop-
ulation. Changes include cutting required 2.5 %
stays in designated quarantine facilities from
Predicted 95% prediction range
7 days to 5 days for international travelers

Annual rate of population change

2
and close contacts of infected people, end-
ing tracing of contacts of patients’ contacts,
and restricting mass testing to situations 1.5
where the source of infection is unclear.
Local governments retain responsibility for 1
setting the timing, location, and duration of
lockdowns, which have disrupted industry
0.5
and sparked increasingly angry protests;
on 14 November, residents of Guangzhou
defied a lockdown by crashing barriers and 0
marching through the streets. The policy
changes come as COVID-19 is surging again –0.5
in China: The National Health Commission 1950 1975 2000 2025 2050 2075 2100
reported 17,909 new cases on 14 November,
the most since the spring. Most of the new 2.5 billion 4 billion 6 billion 8 billion 10 billion 10.4 billion
cases were asymptomatic.

each $150,000 to further develop the protec- adhesive to make a tight fit on differently
Roche Alzheimer’s drug flops tive wear. The Mask Innovation Challenge, shaped faces. Both masks are already on the
| An antibody that
C L I N I CA L R E S E A R C H bankrolled by the Biomedical Advanced market. Despite the wins, BARDA says it has
pharmaceutical giant Roche designed to Research and Development Authority no plans to purchase either to stockpile for
treat Alzheimer’s disease by targeting beta (BARDA), tested the masks for ability to health emergencies.
amyloid, a protein that builds up in patients’ filter out airborne particles as small as
brains, has failed in two large, phase 3 viruses and for breathability, comfort, and
clinical trials. Compared with a placebo, looks. The contest began in March 2021 and Big satellite vexes astronomers
injections of gantenerumab slowed cogni- attracted 1448 entrants. One first-place win- | Scientists are alarmed that
S PAC E P O L I C Y
tive decline on standard tests by just 6% or ner, called Airgami and made by Air99, uses the glow from the largest commercial com-
8% in trials enrolling nearly 2000 people an origami shape that provides a big breath- munications satellite could interfere with
CREDITS: (GRAPHIC) C. BICKEL/SCIENCE; (DATA) UNITED NATIONS, DEPARTMENT OF ECONOMIC

with mild dementia due to Alzheimer’s, ing space, making it comfortable to wear ground-based observations. The satellite,
Roche announced on 13 November. That for long periods. The other, ReadiMask, by BlueWalker 3, unfurled its 64-square-
AND SOCIAL AFFAIRS, POPULATION DIVISION, WORLD POPULATION PROSPECTS 2022

reduction was not statistically significant. Global Safety First, has no straps and uses meter antenna this week, which made it
The drug removed less beta amyloid than among the brightest satellites in the sky.
expected, which some scientists suggest BlueWalker 3 is a prototype for the world’s
explains its failure. The setback follows THEY SAID IT first space-based broadband network,
positive results earlier this year for an anti- planned by the company AST SpaceMobile,
It just tells us how terrible
“
amyloid antibody called lecanemab, made which would deploy a constellation of 168
by Biogen and Eisai. More detailed results even larger satellites. Astronomers worry
on several antibody drugs are expected at our culture is becoming, it could blot out objects such as explod-
the Clinical Trials on Alzheimer’s Disease that we can’t have an honest ing stars or Earth-bound asteroids. Radio
meeting later this month. astronomers are also troubled because the
scientific debate.
Improved masks win U.S. contest
| Two small companies that
PA N D E M I C S
Georges Benjamin, executive director of
the American Public Health Association, in MedPage
Today, on the decision by public health specialist
” satellites will operate at radio frequencies
that could infringe on parts of the spectrum
traditionally reserved for the ground-based
observatories. Astronomers were already
make innovative face masks designed to and commentator Leana Wen not to speak anxious that communications satellites
thwart the spread of pathogens tied for at its annual meeting about harassment of public launched by the corporation SpaceX, which
first place this week in a U.S. government– health officials, after she received criticism plans a network of thousands, are obstruct-
sponsored competition that awarded them and threats over her views about COVID-19. ing observations.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 691

1118NewsInBrief_16210255.indd 691 11/15/22 5:23 PM

IN DEP TH

U.S. ELECTIONS

Science community braces for divided government

Split control of Congress could produce fiery hearings and limit new funding

By Jeffrey Mervis with China’s push to become a technological see NIH’s budget keep pace with inflation—
superpower, how the National Institutes of or exceed it. Several higher education lob-

T
he U.S. Congress returned this week Health (NIH) responded to the COVID-19 byists think the Democrats’ stronger-than-
after tumultuous midterm elections pandemic, and whether a laboratory leak in expected showing at the polls will result
that left the Republicans likely to China led to the catastrophe. They will also in 2023 budgets for research agencies
retake control of the House of Repre- likely challenge the Biden administration’s that are close to the generous requests the
sentatives and the Democrats beating efforts to curb climate change by moving White House submitted to Congress earlier
expectations by retaining the Senate. away from fossil fuels. Although such hear- this year.
A divided Congress could mean a bumpy ings will allow Republicans to showcase Some research advocates fear those
ride for the U.S. research community over political arguments that resonate with their numbers could represent a high-water
the next 2 years. In the short run, however, supporters, they are unlikely to lead to sub- mark, however. Under one scenario, if the
science advocates hope the looming pros- stantive policy changes. Democrat-led Senate and the Republican
pect of gridlock could propel legislators to Science advocates are hopeful the parti- House fail to agree on new spending levels
approve bigger research budgets before the san battles and gridlock won’t undermine in 2024 and 2025, budgets could end up es-
current session ends next month. Once the the traditional bipartisan support for re- sentially frozen at or near the 2023 levels.
new Congress begins its 2-year session in search funding. An early test began this “Many Republicans, and some Demo-
January 2023, funding increases could be- week as the current Congress took up a crats, view the massive expansion of the
come more difficult to secure. massive piece of legislation that would set national debt in the last 3 years as un-
“I believe that the days of generous sci- spending levels for all federal agencies in sustainable,” Atkinson says. Those lawmak-
ence funding increases are over,” says Robert fiscal year 2023, which began on 1 October. ers “will likely try to limit the growth of
Atkinson, president of the Information (Federal agencies are now under a spending nonentitlement, nondefense spending to at
Technology and Innovation Foundation. freeze that expires on 16 December, raising most the rate of inflation.”

PHOTO: MICHAEL CIAGLO/GETTY IMAGES

The 8 November elections didn’t gener- the prospect of a government shutdown un- The fate of future research budgets will
ate a “red wave” that would have given Re- less the freeze is extended or replaced with be partly in the hands of the lawmakers who
publicans the sheer numbers to roll back a yearlong appropriation.) end up leading the appropriations panels in
large parts of President Joe Biden’s agenda. A new law, the CHIPS and Science Act, both chambers. Senator Patty Murray (D–
Instead, neither party will likely be able to calls for double-digit annual funding boosts WA) is expected to succeed retiring Senator
advance major new policy initiatives. for several research agencies, including the Patrick Leahy (D–VT) atop the full Senate
House Republicans have said they will ag- National Science Foundation (NSF), and spending panel, with subcommittee chairs
gressively investigate whether the Biden ad- science advocates hope lawmakers will fol- still in flux. A Republican House is expected
ministration is doing enough to keep pace low through on that plan. They’d also like to to elevate Representative Kay Granger

692 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NID_16198183.indd 692 11/15/22 5:25 PM

NE WS

(R–TX) to lead the spending panel, re- CANCER RESEARCH

placing Representative Rosa DeLauro
(D–CT). Representative Robert Aderholt
(R–AL) is in line to lead the panel oversee-
ing NSF, NASA, and several other research
Tumors can teem with microbes.
agencies, while Representative Tom Cole
(R–OK) would get the appropriations sub-
committee overseeing NIH.
But what are they doing there?
Cole has been a reliable advocate for New study suggests microbiomes can promote cancer
biomedical research spending. And one re-
tired Republican congressman thinks his
by suppressing immune response and seeding metastases
former colleagues will continue to back
hefty budgets for NIH. “Who wants to fight By Gunjan Sinha antibiotics. And because each type of can-
with their constituents when they come to cer appears to come with a unique micro-

O
Washington to demand the government do ur bodies harbor countless biome, researchers are exploring whether
more to find a cure for this or that disease?” microbes—and so do our tumors, it microbes could be used as a diagnostic tool
says Charlie Dent, who now serves on the turns out. Over the past 5 years, re- to detect cancer early in a blood sample.
board of Research!America, an advocacy searchers have shown cancer tissue Until recently, most cancer researchers
group for biomedical research. At the same contains entire communities of bac- believed tumors were sterile, says Ravid
time, Dent says, the retirement of Senator teria and fungi. Now, it appears some Straussman, a cancer researcher at the
Roy Blunt (R–MO) means NIH needs a new of the bacteria may be cancer’s accomplices. Weizmann Institute of Science. But about
champion in the Senate. In a paper in Nature this week, a team a decade ago, as a postdoc at the Broad
The House science committee, which led by Susan Bullman of the Fred Hutchin- Institute, Straussman accidentally dis-
helps set policy at major nonhealth research son Cancer Center reports that in oral covered that human pancreatic and colo-
agencies, would also get a new leader: Rep- and colorectal tumors, bacteria live inside rectal cancer cells grown in the lab stopped
resentative Frank Lucas (OK). Currently the cancer cells and boost their production of responding to a cancer drug named gem-
panel’s top Republican, Lucas worked closely proteins known to suppress immune re- citabine when Mycoplasma bacteria
with the committee’s outgoing chair, Repre- sponses. The microbial interlopers may set were present in the culture. The bacteria,
sentative Eddie Bernice Johnson (D–TX), to off a chain reaction that prevents the im- he discovered, “protected” the cells by
craft broadly bipartisan bills. That includes mune system from killing cancerous cells, producing an enzyme that breaks down
CHIPS, although Lucas reluctantly voted and they may also help cancer metastasize gemcitabine.
against it after Republican leaders decided to to other parts of the body. Straussman found he could render gem-
enforce party discipline for political reasons. The study doesn’t entirely clinch the citabine ineffective in mice with colon
The science committee is expected to case for a bacterial role in cancer, but it cancer by injecting the animals with other
look at how the Biden administration is is very suggestive, says Laurence Zitvogel, types of bacteria, including an Escherichia
implementing several popular provisions in a tumor immunologist at the Gustave coli strain, and that treating them with
CHIPS, notably programs to spread federal Roussy Institute. “It shows that bacteria antibiotics restored the drug’s effective-
research spending to regions of the country in colorectal and oral tumors can actively ness. When he studied 113 human pan-
that traditionally receive little of it and to disturb the immune equilibrium,” she says. creatic cancer samples, he found bacteria
accelerate the commercialization of basic Confirmation that microbes can cause tu- that produced the drug-chewing enzymes
research discoveries, creating new indus- mors to grow or spread could open up new in 76% of them—raising the question of
tries and lots of well-paying jobs. Issues ways to make cancer treatment more effec- whether they contributed to drug resis-
important to Lucas’s rural district are also tive, for instance by killing bacteria with tance in human cancers. Straussman and
high on his agenda, including reauthoriza-
tion of a major bill governing U.S. agricul-
tural research policy, weather programs,
and the regulation of drones.
Given the economic and fiscal strug-
gles facing the nation, U.S. research-
ers shouldn’t expect to get everything
they want from the new Congress, says
John Culberson, a Texas Republican who
chaired the House spending panel that
oversees NSF and NASA before losing his
IMAGE: JORGE GALEANO NINO/BULLMAN LAB

House seat as part of a Democratic wave

in 2018. But Culberson, now a lobbyist for
Federal Science Partners, believes Republi-
cans who are likely to occupy key positions
in the next Congress “understand that in-
creased support for basic science and space
exploration are good for the economy and
important to the nation. And they will
fund as much science as the country—and Researchers cultured cancer spheroids without bacteria (left) and with Fusobacterium nucleatum (shown
taxpayers—can afford.” j in pink, right). After 20 hours, individual cancer cells containing bacteria migrated away from the spheroid.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 693

1118NID_16198183.indd 693 11/15/22 5:25 PM

NEWS | I N D E P T H

his colleagues are now planning a clinical outcome. Now, Bullman and her colleagues SCIENTIFIC INTEGRITY
trial to test whether antibiotics can im- have addressed the question by studying
prove pancreatic cancer treatment.
Soon afterward, Gregory Sepich-Poore, a
doctoral student in the lab of microbiome
eight tumors removed from patients with
oral cancer and 19 others from colorectal
cancer patients. Mapping the distribution
Whistleblower
researcher Rob Knight at the University of
California, San Diego (UCSD), was hunting
for ways to diagnose pancreatic cancers
of the microbes showed they only colonized
specific areas of the tumors. These infected
regions had high levels of proteins known
finds possible
early. He was motivated by his grandmother’s
death from the cancer, which is often di-
agnosed too late for treatment to be effec-
to suppress cancer-fighting T cells or fuel
cancer growth. T cells amassed outside
these regions, the researchers found, but
misconduct in
tive. Inspired by Straussman’s 2017 paper,
Sepich-Poore began to scour the Cancer
few were found inside. (Instead, the regions
contained neutrophils—a type of immune
his own papers
Genome Atlas, a large DNA database of hu- cell that fights infections, among other
man cancers, for snippets of genetic material jobs.) “It’s conceivable that the bacteria are Matthew Schrag confronts
from microbes. somehow causing the T cells to move away a mentor after their
In March 2020, he, Knight, and colleagues from the tumor,” Blaser says.
reported that microbial RNA and DNA was Using a technique called single cell se- joint work is flagged on
present in each of the 33 types of cancers quencing, the researchers found bacte- the PubPeer website
they studied, and that each cancer type had ria preferentially infect cancer epithelial
a unique microbiome. The team also found cells—which line the inside surface of
those distinct microbial signatures in blood organs—and that only cells in which Fuso- By Charles Piller
samples from cancer patients. Based on bacterium and Treponema bacteria were

W
their findings, Sepich-Poore and dominant tended to show both hen Vanderbilt University neuro-
Knight co-founded San Diego– immunosuppressive and can- scientist and physician Matthew
based Micronoma, a startup “This paper cer promoting characteristics. Schrag went public earlier this
that aims to identify early-stage
cancer in blood samples—a so-
has taken the “This paper fills a critical
gap” by showing that bacteria
year with concerns about appar-
ently doctored images in scores
called liquid biopsy. field a big inside cancer cells may alter of Alzheimer’s papers—including
Later in 2020, Straussman the cells’ behavior, says George seminal research underpinning one aspect
and his colleagues confirmed step forward.” Miller, a cancer doctor and re- of the dominant amyloid hypothesis of the
that many tumors have distinc- Ravid Straussman, searcher at Trinity Health of disease—he anticipated that his motives and
tive populations of microbes Weizmann New England. analyses would be dissected. “I also expected
and found they mostly reside Institute of Science Bullman and her colleagues every project that I ever participated in to
inside cancer and immune also co-cultured Fusobacterium be carefully scrutinized, and that my work
cells, rather than between those cells. species with colon cancer spheroids—small would stand up to that scrutiny,” he says.
Fungi often take up residence in tumors models of human cancers—embedded in a So Schrag assumed there would be an
as well. In a study of 17,000 tumors, pub- matrix that contained neutrophils, and innocent explanation when, a few weeks
lished in Cell in September, the UCSD and compared them with bacteria-free spher- after a Science investigation reported
Weizmann groups found fungal species re- oids. With the bacteria present, neutro- his disturbing findings of apparent mis-
siding in each of 35 cancer types. Again, phils tended to move toward the cancer conduct, he received automated emails
each cancer type was associated with a dis- cells, just as they did in the patient tumor from PubPeer, a web forum where scien-
tinct combination of species, which could samples. And the researchers saw infected tific wrongdoing charges are often leveled.
help refine Micronoma’s diagnostic tools. cancer cells breaking off the spheroids and They notified him that two of his own arti-
(Straussman now sits on the company’s sci- migrating, which Bullman thinks may be a cles from more than 15 years ago had been
entific advisory board.) sign that they are metastasizing. flagged as containing dubious images.
The paper reported another striking Zitvogel says the paper paints a plausi- On close examination, Schrag had to
finding: Certain combinations of fun- ble picture of how microbes could hamper confront an unnerving prospect: that a
gal species correlated with lower odds of the body’s defenses against cancer. Still, co-author, neuropharmacologist Othman
survival in several types of cancers, most the spheroid model “is a reductionist ap- Ghribi, Schrag’s first mentor and still a
strongly in ovarian and breast cancer. In proach,” she cautions; the human body, trusted friend, might have also engaged
October, another group reported some- which has a varied arsenal of immune cells in misconduct. The papers, published in
thing similar in Cancer Cell: The presence and a diverse and largely beneficial micro- 2006 when Schrag was an undergraduate
of a particular bacterial signature seemed biome, may have other mechanisms that working in Ghribi’s lab at the University
to hasten death in pancreatic cancer. The keep cancers from metastasizing. of North Dakota (UND), covered research
probability of surviving 2 years after treat- The study was small and only included on several factors related to amyloid pro-
ment doubled in patients that did not have two types of cancers, Straussman adds, teins in rabbit brains. (Many Alzheimer’s
the signature. “That’s an eyebrow-raising which leaves plenty of work to do. But, researchers believe the disease is caused by
finding,” says co-author Martin Blaser, a “Bullman’s research has shown us how we amyloid’s effects on brain cells.)
cancer microbiome researcher at Rutgers should be exploring the tumor microbiome,” Schrag soon discovered that the suspect
University, Piscataway, who also sits on he says. “This paper has taken the field a big work in the two papers fit a large pattern
Micronoma’s scientific advisory board. step forward.” j of questionable research spanning much of
But none of these findings showed just Ghribi’s career, both before and after they
how fungi or bacteria might lead to a worse Gunjan Sinha is a science journalist in Berlin. worked together.

694 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NID_16198183.indd 694 11/15/22 5:25 PM

According to Schrag, in a phone con-
versation the senior scientist emotion-
ally acknowledged “problems” in many of
his papers, including those two flagged
on PubPeer, and accepted responsibility.
During their discussion, which Schrag re-
counted to Science, Ghribi maintained to
his former student that the underlying
findings were correct but admitted to ex-
aggerating data. “I’m nauseated talking
about it,” Schrag says.
After initially agreeing to an in-person
interview with Science, Ghribi backed out
a few days before the appointment, cit-
ing a pending UND investigation. He did
not respond to a request to verify Schrag’s
account of their talk or to comment on a
dossier of suspect images in his papers
that Schrag has compiled with other foren-
sic experts. But Ghribi told Science via a
5 October email that he wanted to “exoner-
ate [Schrag] of any wrongdoing with the
manuscripts he co-authored in my lab.”
Back in August, Schrag contacted a UND
official to see whether any of the original
images for the two suspect papers could be
found. The official said he could not iden- Matthew Schrag (left) and Othman Ghribi enjoying better times at Schrag’s Ph.D. graduation ceremony.
tify any relevant documents. But the uni-
versity found that Schrag’s concerns about prepared by Ghribi for a third joint paper, the matter, but in an email to Schrag it said
the papers, along with PubPeer comments which appeared in 2008 in the journal it would look into the concerns.)
on those and other publications by Ghribi, Hippocampus, also warrant a retraction or The three found possible scientific mis-
warranted the inquiry, according to a 10 Oc- correction; he has contacted the journal to conduct in five other Ghribi papers in the
tober email to Ghribi, obtained by Science. share his concerns. Journal of Neurochemistry. Lawrence and
“We are taking all reasonable steps to After reviewing additional comments Prado said those cases “will be also care-
secure records related to the research in posted to PubPeer about other Ghribi pa- fully evaluated in a timely way.” Four pa-
question,” John Mihelich, a UND vice pres- pers, Schrag says he recently resolved to pers with suspect images, co-authored
ident, told Science in an email. “We are in take a deeper look at his mentor’s work, by Ghribi, appeared in the Journal of Al-
communication with the appropriate fed- including all their joint papers. He enlisted zheimer’s Disease. UT San Antonio pro-
eral research offices and sponsors.” help from microbiologist and forensic im- fessor George Perry, chief editor of the
Resolving the matter might prove chal- age analyst Elisabeth Bik and another im- journal, reviewed the full dossier and says
lenging. Ghribi has moved to the Univer- age sleuth—a nonscientist who uses the he was surprised to find many images “sus-
sity of Texas (UT), Rio Grande Valley. And pseudonym Cheshire, in part to minimize picious,” although he emphasizes that he
Schrag says during the recent phone call legal risks. (Science confirmed Cheshire’s has no expertise in forensic image analysis.
Ghribi said that, out of remorse, he had identity and agreed to keep it confidential.) He describes Ghribi as a “likeable,” rela-
discarded his scholarly awards and purged In recent years, Bik and Cheshire have iden- tively established researcher whom he has
his computers and files of raw experimen- tified thousands of apparently manipulated always considered “a person of integrity.”
tal data. If so, a review of Ghribi’s papers or duplicated images in many papers, often Perry says he contacted Ghribi to ask for
could be hampered, because proof of im- leading to retractions or corrections. more information on the suspect papers
age manipulation sometimes requires un- In their assessment of Ghribi’s work and discuss Ghribi’s role as a member of
cropped original images for comparison from 2001 through 2019, the three found the journal’s editorial board. The journal
with published images. suspect images in 33 papers, including the will assess the claims, he adds, and issue
Schrag recently asked two journals to three co-authored with Schrag. Ghribi was errata or retractions if warranted.
retract his suspect papers with Ghribi. the only author shared by all the papers, Schrag acknowledges Ghribi’s early in-
The publisher of Experimental Neurology and he was usually in a prominent posi- fluence in his career and feels sad about
said via email he was awaiting comment tion as either first or last author. The prob- his former mentor’s apparent misdeeds.
from Ghribi. Andrew Lawrence and Marco lematic images included Western blots (a But Schrag says he faced an imperative
Prado, editors of the Journal of Neuro- common method for displaying proteins in to correct the scientific record, including
chemistry, said in an email to Science tissue samples) and micrographs of brain work he played a part in. “You have to have
PHOTO: MATTHEW SCHRAG

they “applied forensic analysis and con- tissue. Their joint 67-page dossier, which a near-religious commitment to research
cluded there are issues with the pub- Schrag provided to Science, shows more integrity. If the rules apply to others they
lication.” They added that all authors, than 100 apparently problematic images, have to apply to [all of ] us,” he says. j
including Ghribi, agreed the paper should many from work funded by grants from
be retracted, which will occur soon. Schrag the National Institutes of Health. (The This story was supported by the Science Fund for
says he’s also evaluating whether images agency declined to comment to Science on Investigative Reporting.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 695

1118NID_16198183.indd 695 11/15/22 5:25 PM

CONSERVATION Woolly mammoth tusks, such as these from Wrangel
Island in Russia, may be sustaining

Booming trade in mammoth ivory the global market for illegal elephant ivory.

ephant ivory, Huynh said. “In order to

may be bad news for elephants make up for decreasing supply, organized
crime has turned to using mammoth ivory.”
According to Huynh, criminal organiza-
Paleontologists are urged to take a stand against a market tions in Russia pay good money for private
that may provide cover for continued poaching tusk hunters to find and extract mammoth
ivory from the melting permafrost. “The
rest of the skeleton is either destroyed on
By Michael Price, in Toronto living elephants. He urged paleontologists the spot or lost to erosion,” he said.
to raise their voices against the fossil ivory He noted a 2014 report from wildlife

I
n 2015, Andy Huynh was accompany- trade—and avoid dealing with unscrupu- conservation researchers Lucy Vigne of
ing wildlife guards in Kenya’s Maasai lous collectors who might be involved in it. Oxford Brookes University and Esmond
Mara National Reserve to help ward off Some researchers question whether there Bradley Martin, which found that since
poachers. Fresh off a decade of service are enough data to prove ancient ivory really 2002, sales of mammoth ivory had grown
in the Middle East with U.S. Special Op- is buoying the demand for elephant tusks, from almost nothing to about 40% of all
erations Forces, he thought there was but others at the meeting welcomed his call ivory items sold in Beijing and nearly 70%
little that could faze him. But when he saw to action. “Andy did a great job of making in Shanghai. Huynh also presented more
his first poached rhinoceros, with half of its his case,” says Thomas Holtz, a paleonto- recent, unpublished data from a 2018–19
face sawed away for the horn, he turned and logist at the University of Maryland, College joint operation conducted by Interpol
threw up. “I knew then and there I wanted Park. “Plenty of SVP talks address mass ex- and the United Nations Office on Drug
to dedicate my life to stopping wildlife tinction or even the death of an individual, and Crime that suggested both elephant
crime,” Huynh said. but those are separated from us by an im- and mammoth ivory are continuing to
He began to work with various wildlife mensity of time. This dealt with death and enter Vietnamese and Chinese markets,
protection nonprofits, then joined a series suffering happening right now.” primarily on Russian shipping contain-
of U.N. and Interpol undercover operations Ivory from African and Asian elephants ers that at times also contain illicit drugs
in China and Vietnam to bust up the il- commands up to $3000 per kilogram on and weapons.

PHOTO: SERGEY GORSHKOV/MINDEN PICTURES

legal trade in elephant ivory. Now, he has the black market, primarily in South- The Convention on International Trade
extended his definition of wildlife to the east Asia, where it is mixed into tradi- in Endangered Species of Wild Fauna and
distant past: the great, tusked mammoths tional medicine as a powder and carved Flora monitors and regulates the trade of
and mastodons of the ice age. into status-signifying statues and other items such as elephant ivory and rhinoceros
At the annual meeting of the Society of trinkets. About 55 African elephants are horn, but it has no explicit protections for ex-
Vertebrate Paleontology (SVP) here last killed every day for their tusks. But stricter tinct animals such as mammoths or woolly
week, Huynh argued that the growing trade poaching laws and China’s closing of le- rhinos. “So, the trade for mammoth ivory is
in ivory from the ancient carcasses found gal ivory carving facilities in 2018 have left unchecked and growing … and the global
in thawing Arctic permafrost is sustaining made it harder for suppliers—sometimes criminal network has clearly taken advan-
a global market that leads to the death of tied to criminal syndicates—to source el- tage of this loophole,” Huynh said.

696 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NID_16198183.indd 696 11/15/22 5:25 PM

NE WS | I N D E P T H

The harm extends to today’s elephants, PHILANTHROPY

he argued. He noted that although mam-
moth ivory has surged into the market, the
overall demand for elephant ivory appears
to have remained constant or even grown.
Billionaire’s crypto company
His conclusion: Ancient ivory isn’t replac-
ing elephant ivory, but is instead sustaining
the market’s appetite for it. Some traffickers
collapse strands scientists
may even be passing off elephant ivory as As FTX files for bankruptcy, the grantees its foundations
legal mammoth ivory, he says, noting that
it can be hard to distinguish between the
supported may not see all of their pledged money
materials in smaller worked pieces, such as
trinkets and beads. By Robert F. Service avoiding use of its Future Fund grant
Vigne, who did not attend the conference money for now, except to pay the salaries

L
but watched a recording of Huynh’s talk, ast week’s collapse of the crypto- of the three newly hired people.
says it’s unclear just how mammoth ivory currency exchange FTX is sending “We don’t think it is right that anyone
is affecting demand for elephant ivory in aftershocks through the scientific should lose their jobs over a financial ca-
mainland China. Some research hints that community. An undergraduate phys- lamity totally unrelated to the excellent
the rise in mammoth ivory has led to less ics major at the Massachusetts In- work they are doing,” Esvelt says.
elephant poaching. Vigne’s own visits to stitute of Technology (MIT) who SecureBio is now scrambling to secure
mainland Chinese ivory markets suggest founded FTX and quickly became a bil- emergency funding. The organization just
“mammoth ivory has certainly helped re- lionaire, 30-year-old Sam Bankman-Fried last week released a white paper that re-
duce elephant ivory, but it has also pro- backed philanthropic organizations that views new technologies, such as improved
vided a route to [continue to sell] elephant supported a wide variety of science- personal protective equipment and germi-
ivory,” she says. “So, it’s a tricky one.” related causes designed to improve human cidal lights, that could cope with or halt
What is needed, Vigne says, is more infor- well-being. pandemics. “The events of last week put
mation on just how much worked elephant Now, with FTX in bankruptcy and un- this critical work in jeopardy,” Esvelt says.
ivory is being trafficked under the guise of der investigation for misuse of investors’ FTX’s collapse was unthinkable just
legal mammoth ivory. Victoria Herridge, a money, his formerly flush foundations are days earlier. The company, which serves
paleontologist at the Natural History Mu- suddenly strapped for cash and much of as an online trading platform for crypto-
seum in London, agrees. “You need passion- that work is at risk. currency, had assets between $10 billion
ate, activist voices [like Huynh’s], but you One foundation, the Future Fund, was and $50 billion, according to bankruptcy
also need data.” just launched in February. But by the end documents. But it was brought to its knees
Thomas Carr, a paleontologist at Car- of June, its officials reported awarding by an old-style run on the bank, as inves-
thage College and Huynh’s undergraduate 262 grants and “investments” totaling tors tried to withdraw their money after
mentor, notes that paleontology also has a $132 million. It’s unclear how much of that doubts were raised about FTX’s financial
stake in limiting the trade: The sale of fos- money has been distributed. But on 10 No- health. The company declared bankruptcy
sil ivory reduces the number of mammoth vember, five senior Future Fund officials re- on 11 November, and just hours later more
carcasses available for study. “If fossils are signed and announced in a statement, “We than $500 million was reportedly stolen
being destroyed for tusks, you lose so much are devastated to say that it looks likely that from it by hackers.
data,” he says. “It’s a loss for science and it’s there are many committed grants that the Just what will happen to awards the Fu-
a loss for society.” Future Fund will be unable to honor.” ture Fund and the similar FTX Foundation
SVP and the broader paleontological “It’s definitely a mess,” says Josh have already made remains unclear. FTX
community can help, Huynh says, by issu- Morrison, who heads 1Day Sooner, a pan- owes billions of dollars to creditors and is
ing public statements and putting pressure demic preparedness research and advocacy now being investigated by the U.S. Securi-
on elected officials and international regu- organization that received $375,000 from ties and Exchange Commission and the De-
lating bodies. Scientists should also avoid the Future Fund and the FTX Foundation. partment of Justice, according to The Wall
obtaining samples from unscrupulous tusk During the pandemic, 1Day Sooner became Street Journal.
hunters, he says, as they might be collabo- known for advocating so-called human chal- Writing in an online forum hosted by the
rating with collectors who are employed lenge trials, which deliberately infected vol- Center for Effective Altruism, to which the
by criminal organizations. Holtz agrees unteers with SARS-CoV-2 to test vaccines. Future Fund pledged nearly $14 million,
paleontologists should do more, saying, “I Other notable science recipients of the Molly Kovite, legal operations manager for
don’t think most of us understood that fos- Future Fund’s money include Sherlock Bio- the Open Philanthropy foundation, noted
sils were being run in the same shipments sciences, which was awarded $2 million that FTX’s creditors could try to “claw
as, for example, heroin.” for CRISPR-based infectious disease diag- back” their investments during bankruptcy
Jessica Theodor, a paleontologist at the nostics; HelixNano, which was awarded proceedings. If grantees received awards
University of Calgary and SVP’s outgoing $10 million for research on a vaccine effec- after 11 August, which is 90 days prior to
president, says SVP has for decades issued tive against all different coronaviruses; and the bankruptcy filing, “the bankruptcy pro-
statements decrying the loss of scientifi- SecureBio, which was given $1.2 million to cess will probably ask you, at some point,
cally significant fossils to commercial trade. develop better pandemic defenses, such as to pay all or part of that money back,”
In the wake of Huynh’s talk, though, she an early warning system that screens waste- she predicts.
says the society will establish a task force to water for pathogen genetic material. That has grantees wondering how they
investigate what more it can do to protect SecureBio’s co-founder, Kevin Esvelt, will pay the bills. “Everyone is obviously
mammoth ivory. j a biologist at MIT, says the nonprofit is really worried,” Morrison says. j

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 697

1118NID_16198183.indd 697 11/15/22 5:25 PM

NEWS

FEATURES

PHOTO: CREDIT GOES HERE AS SHOWN; CREDIT GOES HERE AS SHOWN

698 00 MONTH 2022 • VOL XXX ISSUE XXXX [Link] SCIENCE

1118NewsFeature_16177465.indd 698 11/14/22 5:32 PM

NE WS

SHELTER FROM THE STORM

A plan to wall off Houston and nearby industry from flooding
caused by hurricanes will cost tens of billions of dollars. Will it be enough?

P
lans for one of the world’s biggest By Warren Cornwall, in Galveston, Texas government. In the process, each country or
and most expensive flood barriers city is confronting similar trade-offs: between
were born in a second-floor apart- the U.S. Army Corps of Engineers, the na- protecting people and preserving ecosystems,
ment here in this city on the Gulf tion’s builder of mammoth water infrastruc- between technically sensible designs and aes-
of Mexico, as water 4 meters deep ture. The state of Texas has embraced the thetically acceptable ones, between maximal
filled the street below. In Septem- idea, creating a taxing district to help pay its protection and what’s affordable.
ber 2008, Bill Merrell, an oceano- share. In July, Congress authorized the Corps In Galveston, Merrell and some other
grapher at Texas A&M University, to proceed—though it has yet to appropriate scientists think the Corps hasn’t struck
Galveston, was trapped with his money for construction. the right balance. Merrell warns that its
wife, daughter, grandson, and “two annoy- The project would be the costliest ever built plan—a scaled-down version of his original
ing chihuahuas” in the historic building he by the agency. By some measures it would blueprint—is destined to fail, perhaps cata-
owns. Outside, 180-kilometer- strophically. “It’s too weak—[the
per-hour winds generated by defenses] would only stand up
Hurricane Ike rattled windows to like a 30-year storm,” he says.
and drove water from the gulf “Essentially you don’t have any
and Galveston Bay into the city. protection” against the more ex-
As saltwater swirled through treme storms that have already
the shops and restaurants left deep scars on Galveston—
downstairs, Merrell sat in his and are likely, as climate change
office and sketched plans for a advances, to leave more.
project he hoped would put an
end to the storm-driven flood- LIKE A GYMNAST on a balance
ing that had repeatedly devas- beam, Galveston perches on a
tated this part of Texas. slender ridge of sand, precari-
It was an ambitious vision: ous and exposed. To the north,
Seventy kilometers of seawalls behind that barrier island, lies
rising 5 meters above sea level Galveston Bay, an estuary half
would stretch the length of the size of Rhode Island, teem-
Galveston Island and beyond. ing with shrimp and birds. The
Enormous gates would span Oceanographer Bill Merrell, next to a statue memorializing a deadly hurricane that hit bay is so shallow, locals joke that
PHOTOS: MELISSA PHILLIP/© HOUSTON CHRONICLE; (OPPOSITE PAGE) SMILEY N. POOL/AP PHOTO

the 3-kilometer-wide channel Galveston, Texas, in 1900, developed a plan for protecting the region from storms. if you fall out of a boat, just stand
through which ships pass in up. To the south, Galveston faces
and out of Galveston Bay. The defensive pe- dwarf anything else in the world. The sea the Gulf of Mexico, whose warm waters fuel
rimeter would seal off not just Galveston, gates meant to block gulf waters from Galves- hurricanes nearly every year.
but the whole bay, with Houston at its far ton Bay would span a gap bigger than the Merrell’s 1870 building in the Strand
end, protecting more than 6 million people famous pivoting Maeslant barriers that hold Historic District has survived a number of
and the country’s largest collection of chem- back the North Sea near Rotterdam, Nether- them, including the all-time worst. In Sep-
ical plants and oil refineries. lands. “Everything is bigger in Texas,” quips tember 1900, a monster Category 4 hurri-
Though Merrell had spent decades study- Bas Jonkman, a civil engineer and water- cane with gusts topping 210 kilometers per
ing ocean currents and storm surges, he control expert at the Delft University of Tech- hour blasted the area. Driven by the wind,
had no engineering experience. But as he nology (TU Delft). flood waters nearly 5 meters deep surged
watched the murky waters soak the city, Dutch experts like Jonkman are in high into the city from both the bay and the gulf,
including his own carefully restored 19th demand these days. Around the world, from crushing thousands of buildings. More than
century landmark, he New York City to Singapore, governments 8000 people died. It’s still the deadliest natu-
In 2008, winds decided there had to be are planning massive seawalls and other ral disaster in U.S. history.
and flooding from a better way. “The Dutch measures to ward off the rising seas and In the aftermath, local leaders and the fed-
Hurricane Ike would never put up with intensifying storms expected from climate eral government erected a 6-kilometer-long,
leveled shorefront this,” he said to his wife. change. “A lot of countries are really think- 5-meter-high seawall along the gulf, filling
homes along Today, that first brain- ing through now: ‘How should we defend the space behind it with a deep sand layer
the Gulf of Mexico storm has morphed into ourselves?’” says Marc Walraven, a senior that sloped gently toward the bay. On that
in Texas. a $31 billion plan from adviser on storm surge barriers to the Dutch raised ground, they rebuilt their city.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 699

1118NewsFeature_16177465.indd 699 11/14/22 5:32 PM

NEWS | F E AT U R E S

When Hurricane Ike made landfall a cen-

Houston tury later, Galveston’s seawall held. But bay
waters flooded central Galveston, including
Trinity the Strand, from the less protected north-
Defensive measures Bay ern side. Low-lying neighborhoods nearby
Deadly flooding from wind-driven storm washed away. Storm surge inside Galveston
surges has repeatedly devastated Texas Bay drenched towns dozens of kilometers
coastal communities around Galveston 2 from the gulf. When the water receded, it left
Galveston
and Houston. To reduce the threat, the U.S. Bay Bolivar behind $37.5 billion in damage and 74 dead.
Army Corps of Engineers has proposed Texas Peninsula Galveston got lucky. At the last minute,
building a network of walls, gates, and sand City 1 the storm veered east, sending the eye over
dunes. The $31 billion plan would be the the city and up the bay. In the Northern
largest single project ever for the Corps, Galveston Gulf of Hemisphere, where cyclones rotate counter-
Mexico
and has drawn extensive criticism. clockwise, the most destructive winds blow
on the eastern side, known as the “dirty side.”
Sam Brody, a coastal planner at Texas A&M
San Luis Galveston, calls Ike a “near miss.”
Pass
A MONTH AFTER the storm, in a small online
Galveston seawall Bolivar Roads gate
Texas newsletter, Merrell wrote a column unveiling
Extended seawall Path of Hurricane
Houston his idea for a wall that would barricade the
0 20 Dune surge barriers Ike (2008)
barrier islands and Galveston Bay during a
At right km Houston ship channel Petrochemical facility storm. He called it the “Ike Dike.”
Although the name caught on, the idea
Flood sector
didn’t at first. Some derided it as a monu-
1 Bolivar Roads gate mental overreach in a place that had relied
The centerpiece of the project, a system of gates, gates
would span the 3-kilometer-wide channel that To prevent storm- for years on more modest seawalls to pro-
is the main connection between the Gulf of Mexico driven waters tect towns like Galveston, and on perching
from surging into Water
and Galveston Bay. houses on stilts. But eventually the Corps,
Bolivar Peninsula the bay, huge flow
gates would close which has a regional office on the outskirts
channels that of Galveston, agreed to study the matter. In
allow freighters 2021, it signed off on the current plan, which
New Panamax ship Gulff of
Gul of and tankers to Congress authorized this summer.
3 km M ic
Me
Mex ico
ico reach Houston’s
port and many The blueprint has key features in com-
chemical plants. mon with the original Ike Dike. At Boli-
Surge blocked
var Roads, the chief waterway joining the
gulf and the bay, four swinging gates, each
Vertical lift more than 100 meters long, would guard
Galves
Galves
vve
eston
t n gates two openings big enough to fit some of the
Bay
ay Barriers flanking world’s largest freighters, including Pana-
the channels Gate
max ships (see graphic, left). In Galveston,
would remain
open except the existing seawall would be extended and
during a storm, improved to encircle much of the city. Over-
to allow regular Gates all, analysts at the Corps forecast that the
Shipping channel tidal currents lowered system could reduce storm surge damage to
openings are to flow back
the region by $62 billion over 50 years, or
Galveston 200 meters wide and forth.
and 18 meters deep. twice its estimated cost.
But the plan’s flood-stopping powers are
less than what the Corps—and Merrell—first
envisioned. In 2018, the agency had proposed
2 Dune system New beach Seawall (5.2 meters above sea level) a more conventional concrete seawall along
The Corps has proposed profile the beaches and highways at the gulf’s edge,
protecting the coastline
by constructing a double the length of Galveston Island and the Boli-
Dunes Surge water
row of dunes with tops (4.2 meters
var Peninsula, much like Merrell’s Ike Dike.
level
up to 4.25 meters above sea That proposal did not fare well with island
above sea level. Some level) residents. People living near the beaches ob-
critics fear the dunes
won’t be tall or strong
jected that the wall would be an eyesore,
enough to withstand even as it left many homes unprotected. GRAPHIC: C. BICKEL/SCIENCE

large storm surges, and “The public went ballistic,” says Kelly Burks-
would prefer a taller Copes, an ecologist who led the crafting of
concrete seawall or taller Sea level the plan. “We had some really harsh, really
dunes built over a
hardened rock core. confrontational public meetings.”
Recently, standing on a 50-meter-wide
Old beach profile beach in front of the houses that crowd

700 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NewsFeature_16177465.indd 700 11/14/22 5:32 PM

Galveston Island’s gulf shore, Burks-Copes tropical cyclones. But there is broad agree- Michel Bechtel, mayor of Morgan’s Point, a
explained how the Corps adapted to the crit- ment that the most severe storms—such as town on the northwestern shore of Galves-
icisms. In its final plan, the agency replaced Category 4 and 5 hurricanes—will become ton Bay. Bechtel is president of the board for
much of the seawall with two parallel dunes more intense, says Gabriel Vecchi, a Princ- the Gulf Coast Protection District, which the
built from sand dredged offshore, each eton University climate scientist. And an state created to collect taxes to help pay for
roughly 2.5 meters tall. One dune would increasing number of storms in the Gulf the project. Bechtel questions why any bar-
crest at 3.7 meters above sea level, and a of Mexico will be Category 3 or bigger, pre- rier on the islands would be built lower than
second, farther up the sloping beach, at dicted a 2017 study by scientists at the Na- the nearby Galveston city seawall and the
4.25 meters—almost 1 meter shorter than tional Center for Atmospheric Research. gates spanning the water channel.
the earlier proposed wall. The Corps has understated the risks “Why are we spending billions of dollars
The new plan didn’t trigger an outcry from posed by such storms, Merrell and his col- to do that and have it overtopped?” he says.
locals, but it also doesn’t offer them as much laborators warn, because it overstates the “Let’s look at it with common sense.”
protection. The dunes are sized to withstand protection the new dunes will offer. “I know
the kind of storm that, in the historic record, from my own experience that those dunes THERE IS AN ARGUMENT for spending many
has come along every 50 years on average. will be completely obliterated by any signif- billions more than the Corps has proposed
The surge from a storm like Ike (which an icant hurricane,” says Bruce Ebersole, a civil on protecting Galveston Bay. It rests on a
agency engineer pegged at about an 80-year engineer who until 2011 oversaw storm and worst-case scenario that envisions hurricane-
storm) would flood over the embankments, flood protection research at the agency’s driven floodwaters battering Houston, the
the Corps acknowledges. laboratory in Vicksburg, Mississippi. nation’s fourth largest city, some 80 kilo-
meters away. Modeling by both Ebersole
and Clint Dawson, who studies coastal
ocean dynamics at the University of Texas,
Austin, shows a Category 4 hurricane could
drive a storm surge topping 7 meters into
the northern tendrils of the bay, where the
shore is lined with sprawling oil refineries
and chemical plants.
That’s Jim Blackburn’s nightmare. Earlier
this year, the veteran environmental attorney,
co-director of Rice University’s Severe Storm
Prediction, Education, & Evacuation from
Disasters Center (SSPEED), pictured that sce-
nario as he stood on a bluff looking across the
placid waters of the bay near Buffalo Bayou.
At his back, a concrete tower topped by a sin-
gle enormous star marked the site of the 1836
Battle of San Jacinto, which paved the way for
Texas’s independence from Mexico. Ahead,
Blackburn faced an industrial landscape of
metal towers and white tanks that sprouted
like giant mushrooms along the waterfront.
After Hurricane Ike, officials moved quickly to shore up dunes and seawalls damaged by the storm. “Every one of these tanks could be hit.
That’s what I’m horrified about,” he said.
But Burks-Copes said the agency will be Ebersole, who recently retired from a re- “This complex is incredibly vulnerable.”
able to blunt some of the flooding by clos- search position at Jackson State University, About 30% of the country’s crude oil and
ing the Bolivar Road gates at low tide be- worked with Merrell to craft the Ike Dike 25% of its natural gas are processed in coastal
fore a big storm arrives, leaving room in plan and to vet the proposal from the Corps. Texas, much of it around Galveston Bay. The
the estuary for some of the overflow. Most He says agency modelers assume the dunes region’s refineries can churn out 2.5 million
beachfront homes, meanwhile, are now less will act like a solid 3.5-meter wall during a barrels a day, nearly 15% of the U.S. capacity.
vulnerable than when Ike arrived, Burks- storm. But in a major hurricane, “They’re go- Oil is stored in more than 4600 huge tanks,
Copes noted, because they’ve been elevated ing to be eroded and destroyed quite early,” many near Houston.
PHOTO: NICK DE LA TORRE/HOUSTON CHRONICLE/AP PHOTO

on pilings. “Unless the surge is over the first he says. “Their effect on storm surge,” he If a 500-year storm hit, Rice scientists
floor,” she said, “they’re not impacted.” suspects, will be “negligible.” have found, the surge and winds could dam-
To stand up to a 100-year storm, col- age some 700 tanks, spilling as much as
EVEN UNDER ITS PLAN, the Corps estimates leagues calculated the dunes would need 470 million liters of oil—almost as much as
that damages from multiple storms over to be far more massive: nearly 7 meters tall the 2010 Deepwater Horizon spill in the Gulf
50 years could still reach $30 billion or more in with a crest 45 meters across, almost as wide of Mexico, the largest in U.S. history. The wa-
the Galveston Bay region. The agency’s esti- as a U.S. football field. A more realistic solu- ters could also inundate Houston neighbor-
mate of how often a storm of a certain size is tion, Ebersole and Merrell argue, would be hoods kilometers upstream of the tanks. “The
likely to strike is based, however, on historical building the original Ike Dike seawall, or an toxic goo that might be produced could set
patterns. It does not take into account how a artificial dune made of sand spread over a back not only the region, but the country, for
warmer future might alter that equation. hardened core of rocks or concrete. decades,” says Houston City Council Member
Debate and uncertainty surround the Their view has won support from some David Robinson, an architect who chairs the
many ways a warmer planet could influence important local political leaders, such as council’s infrastructure committee.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 701

1118NewsFeature_16177465.indd 701 11/14/22 5:32 PM

NEWS | F E AT U R E S

For Robinson, some version of the Ike York Harbor to fend off a repeat of 2012’s There’s no reason different regions should
Dike is the cornerstone of preventing that Hurricane Sandy. That option beat an ear- adopt the same standard, says engineer
nightmare. But he worries it’s not enough. lier proposal for a single gated wall guarding Gregory Baecher of the University of Mary-
That’s because, even if the defenses succeed the whole harbor, which earned widespread land, College Park, who served on a National
in reducing a storm surge originating from criticism in part for its $119 billion price tag. Academy of Sciences panel in the 2010s that
the gulf, a major hurricane could still gener- Also in September, the agency said it would studied how the Corps evaluates risks. In the
ate a surge inside the bay. reconsider its proposal to build 10 kilo- Netherlands, he notes, surges from a 1000-
Houston, surrounding Harris County, the meters of seawalls and gates, 6 meters high year storm aren’t much deeper than those
Port of Houston, and a local entrepreneur in some places, around downtown Miami. caused by a 100-year storm, because even
have committed $1 million to SSPEED for Local officials and environmental groups very rare storms there aren’t much more
studies. Blackburn and his colleagues have had strongly objected to the plan, fearing it severe. Gulf Coast hurricanes, however, can
already touted one solution. They propose could further degrade Biscayne Bay and ruin dwarf European storms. By one estimate,
building a string of artificial islands in the the city’s famous shoreline. the surge from a 100-year storm in the gulf
bay—in effect a second wall guarding its In Singapore, officials announced plans is roughly equal to a once-in-a-millennium
northwestern reaches, including Houston. earlier this year to study the feasibility of storm in the Netherlands. So, building to
“The Dutch talk about multiple lines of de- storm barriers. Indonesia is considering a the more conservative Dutch-level standard
fense,” Blackburn says. “That’s what we’re re- giant sea wall to protect Jakarta, its sinking along the gulf would be an enormous task.
ally talking about here.” capital, even as it builds a new capital city “Everybody loves the Dutch,” Baecher says.
One obstacle is cost: Blackburn estimates elsewhere. In 2020, Venice, Italy, after de- But, “They’ve got an easier problem.”
building the islands would add $3 billion to cades of debate and delay, finally inaugurated Merrell says the split also reflects a dif-
$6 billion to the price of the plan advanced a system of 72 mobile walls to seal off its la- ferent attitude. The Dutch build to prevent
by the Corps. Another issue is the potential goon and protect the flood-prone city from flooding, whereas the United States is more
environmental damage caused by yet an- rising seas and extreme high tides. The sys- willing to accept flood damage and then
other huge engineering scheme. rebuild. But allowing another
Regulators “are just not going to Ike-scale flood to hit the Texas
let some massive island be built coast, Merrell says, would be
in the middle of Galveston Bay,” “insane. … We have to protect
says Bob Stokes, president of the it. We can’t recover it anymore.
Galveston Bay Foundation, an It’s just too expensive.”
environmental group. The Corps says federal law re-
Environmental concerns have quires it to build projects that,
already prompted the Corps over 50 years, will produce ben-
to ditch one part of the Ike efits greater than the cost. If a
Dike—a proposed gate system region wants something more
across the shallow San Luis robust, it must be willing to pay
Pass, a smaller channel join- the higher price—but it might
ing the gulf to the bay’s west- be a lot higher, warns agency
ern end. Putting a barrier there engineer Mike Braden. “How
would “cause irreparable envi- close to 100% [protection] are
ronmental impacts” by inter- we trying to get?” he asked ear-
fering with ecologically rich lier this year as he stood on the
tidelands, Burks-Copes says. beach next to Burks-Copes. “As
Stokes fears that even the Hurricane Harvey’s intense rains flooded this chemical facility in 2017. we try to approach that 99%,
remaining gates on the deeper A wind-driven storm surge could be more destructive. 100% solution, the costs expo-
Bolivar Roads channel could nentially go up.”
meddle with tidal currents and wildlife— tem is working—but environmentalists say it Braden, who arrived in Galveston earlier
including commercially valuable shrimp. threatens to destroy salt marshes. this year, has a title out of a Marvel super-
As a result, his group has withheld judg- Underlying many of the debates about the hero movie: He’s chief of the Mega Project
ment on the agency’s plan. “It’s hard to say costs and impacts of flood barriers are deeper Division at the Corps. That means he’s in
we’re for or against it,” Stokes says, “until questions: How much risk is tolerable, and charge of shepherding the Texas project
we actually have a deeper level of environ- what is worth preserving? Do you build for a through the remainder of the planning
mental analysis.” once-in-a-lifetime or a once-in-a-millennium process and breaking ground. If everything
All these protection plans come an impor- storm? And how do you account for the un- goes smoothly—if Congress appropriates the
tant caveat: They would do little to prevent certain effect climate change will have on money, and the project survives challenges
the destruction caused by storms like 2017’s those odds? from opponents—it will take decades to com-
Hurricane Harvey, which brought record in- In the Netherlands, key coastal defenses plete. The Corps now predicts the system

PHOTO: CHARLIE RIEDEL/AP PHOTO

land rains—not a wind-driven storm surge— are built to withstand freakishly rare could be finished by 2043.
that flooded Houston neighborhoods. 10,000-year storms. That partly reflects Braden is open to tweaking the design.
the existential threat of flooding to a na- But he also hears the clock ticking. His big-
OTHER CITIES in the United States and else- tion in which one-third of the land is below gest fear is that a massive storm rivaling
where are wrestling with similar issues as sea level, including much of Rotterdam. In Hurricane Ike or the 1900 disaster will arrive
they consider new defenses against storm the United States, by contrast, the Corps before the work is done. If we don’t want to
surges. In September, the Corps issued plans frequently designs projects for a storm ex- repeat those catastrophes, he said, “We need
for a $52 billion network of barriers in New pected to hit once in a century. to go, go, go, go.” j

702 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118NewsFeature_16177465.indd 702 11/14/22 5:32 PM

CALL FOR PAPE RS

The Journal of Remote Sensing is an online-only Open Access Science Partner Journal published in affiliation with
Aerospace Information Research Institute, Chinese Academy of Sciences (AIR-CAS) and distributed by the
American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner
Journal program, the Journal of Remote Sensing is editorially independent from the Science family of journals and
AIR-CAS is responsible for all content published in the journal. This journal covers multiple research areas that include
theory, science, technology of remote sensing, and interdisciplinary research with earth science and information
science. Particular topics of interest within the journal include radiative transfer modeling, biogeosciences remote
sensing, remote sensing of energy, and more.

Submit your research to the Journal of Remote Sensing today!

Learn more at [Link]/remotesensing

The Science Partner Journal (SPJ) program was established by the American Association for the Advancement of Science (AAAS), the
nonprofit publisher of the Science family of journals. The SPJ program features high-quality, online-only, Open Access publications produced
in collaboration with international research institutions, foundations, funders and societies. Through these collaborations, AAAS furthers its
mission to communicate science broadly and for the benefit of all people by providing top-tier international research organizations with the
technology, visibility, and publishing expertise that AAAS is uniquely positioned to offer as the world’s largest general science membership society.
Visit us at [Link]

@SPJournals @SPJournals

A R T I C L E P R O C E S S I N G C H A R G E S WA I V E D U N T I L J U LY 2 0 2 3

[Link] 703 11/14/22 10:10 AM

INSIGHTS

Various studies show that

COVID-19 vaccination
temporarily alters
menstrual cycles, but the
PERSPECTIVES underlying mechanisms
require more research.

VIEWPOINT: COVID-19

COVID-19 vaccination and menstruation

COVID-19 vaccination causes small changes to menstruation that quickly resolve

By Victoria Male been reported in association with a variety However, these systems are not designed to
of vaccines, including those against patho- detect increased rates of nonserious events

T
he rapid development of safe and ef- gens other than severe acute respiratory that occur commonly. Because menstrual
fective vaccines against COVID-19 has syndrome coronavirus 2 (SARS-CoV-2), so a cycles vary naturally, a particular challenge
been a triumph of medical science, secondary aim of this work is to understand was determining the extent (if any) to which
but vaccines only work if people take the mechanisms by which vaccine-associated the changes reported could be attributed
them. Although there is extensive menstrual changes could occur. to COVID-19 vaccination rather than back-
evidence that COVID-19 vaccination By April 2022, the Vaccine Adverse Event ground variation.
does not affect fertility, misinformation that Reporting System (VAERS) in the United Recent studies illustrate the extent to
it could has been a major source of vaccine States had received more than 11,000 reports which menstrual cycles vary in the absence
hesitancy among young women. As the vac- of menstrual changes and unexpected vagi- of vaccination. The Norwegian Institute of
cination program was rolled out to younger nal bleeding after COVID-19 vaccination (1). Public Health (NIPH) studied a cohort of
age groups, some people noticed menstrual Yellow Card, the equivalent scheme in the 5688 females aged 18 to 30 years who had
changes after COVID-19 vaccination, and United Kingdom, had received more than been recruited to examine other side effects

PHOTO: LUKAS BARTH/REUTERS PICTURES

many members of the public found these 50,000 reports (1). These schemes are effec- of COVID-19 vaccination. Participants were
reports concerning. Research was needed to tive at detecting patterns of serious, but rare, asked to recall their prevaccination and
generate robust data to inform health care adverse events associated with vaccination: postvaccination menstrual cycles: 37.8% de-
professionals and the public about these po- Yellow Card detected the rare clotting dis- scribed at least one aspect as different from
tential side effects. Menstrual changes have order vaccine-induced immune thrombotic their usual experience, even in prevaccina-
thrombocytopenia (VITT) that is associated tion cycles (2). Further, a US study that exam-
with adenovirus-vectored vaccines, and ined changes to cycle length using data from
Department of Metabolism, Digestion
and Reproduction, Imperial College London, VAERS identified myocarditis as a rare ad- a menstrual cycle tracking app reported a
London, UK. Email: [Link]@[Link] verse event associated with mRNA vaccines. within-individual standard deviation in cycle

704 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_16124035.indd 704 11/11/22 4:25 PM

length of 4.2 days in unvaccinated individu- this resolved in subsequent follow-ups (6). extent of side effects and changes to cycle
als (1). This underlines the need for formal The NIPH study found that heavi- length or flow (5). Conversely, a survey of
approaches, including unvaccinated compar- er-than-normal bleeding was most com- 27,143 menstruating individuals found that
ison groups, to identify vaccine-associated monly associated with vaccination, with those who experienced fever or fatigue post-
cycle changes. 13.6% of participants reporting this for the vaccination were more likely to experience a
The findings of these formal approaches period after they were vaccinated compared heavier-than-usual period (8). Each of these
have been remarkably consistent: COVID-19 with 7.6% for the period before vaccination approaches has weaknesses: The UK cohort
vaccination is associated with a small in- (2). The UK study of 79 participants was un- is potentially underpowered to detect an as-
crease in menstrual cycle length, but this rap- able to detect an increase in menstrual flow sociation, whereas the survey relies on partic-
idly resolves. In a study of data from 3959 US- associated with vaccination (5), although ipants accurately recalling their experiences
resident users of the menstrual cycle tracking this could be a result of the smaller cohort and is potentially influenced by respondents
app Natural Cycles, of whom 2403 were vac- size. These conflicting results could also re- being more likely to recall other side effects if
cinated and 1556 were unvaccinated, there flect the increased risk of recall bias in the they noticed a menstrual change.
was no effect of the first vaccine dose, the sec- NIPH study. Approaches that combine large Two biologically plausible mechanisms
ond dose was associated with an increase in cohorts and data collection in real time are by which immune stimulation might cause
cycle length of 0.45 days, and administration needed to resolve this. menstrual changes have been proposed:
of both doses of vaccine in the same cycle was Given this evidence that COVID-19 vac- Innate immune responses could transiently
associated with an increase of 2.32 days (3). cination does alter menstrual cycles, albeit interfere with the hormones that drive the
A follow-up study using the same approach temporarily, the obvious question is how? menstrual cycle, or they could affect mac-
in a global cohort of 19,622 people, of whom The type of vaccine does not affect the chance rophages and natural killer cells in the lining
14,936 were vaccinated, reported similar that an individual will experience a change of the uterus, which control the breakdown
results (4). Notably, in both studies, cycle to menstrual timing (1, 5–7) or flow (5, 7, 8), and regeneration of this tissue through the
length returned to normal within two cycles. suggesting that the effect is a result of the cycle (see the figure). In support of the hy-
Consistently, data from 9652 US residents immune response to vaccination rather than pothesis that the effects are hormonally
who tracked their cycles as part of the Apple a specific vaccine component. In support mediated, individuals in whom the ovarian
Women’s Health Study identified an increase of this, menstrual changes have previously hormones estrogen and progesterone are
in cycle length of 0.5 days after the first vac- been reported with typhoid (9), hepatitis B supplied exogenously by combined hormo-
cine dose and 0.39 days after the second (10), and human papillomavirus (HPV) (11) nal contraception are less likely to experi-
dose, with cycle length returning to normal vaccines. Indeed, as early as 1549, the doc- ence menstrual changes after vaccination (5,
the next cycle (1). In the UK, a prospectively tor Wan Chhüan observed that inoculation 7). Furthermore, the timing of vaccination
recruited cohort of 79 people who recorded against smallpox could bring on menstru- within the menstrual cycle affects whether
their cycles in real time also found a small, ation unexpectedly (12). Two studies have cycle length increases. The menstrual cycle
but significant, increase in cycle length only addressed the hypothesis that menstrual is divided into two phases: the follicular
in cycles in which a vaccine dose was given changes after COVID-19 vaccination are as- phase, which occurs before ovulation and
(5). Among 3858 North American nurses tak- sociated with activation of the immune re- can be prolonged by hormonal alterations,
ing part in the Nurses’ Health Study 3, which sponse, but with conflicting results. The 79 and the luteal phase, which occurs after ovu-
collects data from its participants biannually, participants in the prospectively recruited lation and is more consistent in length. If
vaccination was associated with increased UK cohort recorded their experience of com- menstrual changes are mediated by immune
odds (1.48) of reporting a longer menstrual mon immune-mediated vaccine side effects, effects on the control of ovarian hormones,
cycle in the next follow-up questionnaire, but and no association was found between the vaccination would be expected to prolong

Possible mechanisms by which vaccines affect menstrual cycles

Cytokines produced in response to COVID-19 vaccination could affect the hormonal dialogue between the hypothalamus, pituitary gland, and ovaries (HPO), lengthening
the cycle. Cytokines could also affect macrophages and natural killer cells in the lining of the uterus, which mediate tissue repair, resulting in heavier menstrual flow.

Effects on the HPO axis

Menstruation Follicular phase Luteal phase
Hypothalamus
Pituitary gland

FSH Estrogen Typical After Increased

LH Progesterone 28-day cycle vaccination cycle length

Ovary

Effects on immune cells in the endometrium

GRAPHIC: A. MASTIN/SCIENCE

Macrophages and Increased

natural killer cells menstrual flow
control regeneration
Uterus of the endometrium
FSH, follicle-stimulating hormone;
LH, luteinizing hormone.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 705

1118Perspectives_16124035.indd 705 11/11/22 4:25 PM

INSIGHTS | P E R S P E C T I V E S

the follicular phase, but this can only occur if menstruate at some time in their lives, yet GEOLOGY
vaccines are administered during this phase. data about effects on menstruation are
Indeed, the Apple Women’s Health Study
found that cycle-length increases are only
associated with vaccination in the follicular
rarely collected in vaccine trials. This must
change—not least to offer reassurance that
this area of public health is also taken seri-
Defining the
phase of the cycle (2).
In support of the possibility that COVID-19
vaccination affects immune cells in the
ously by vaccine developers. Furthermore,
some information can only be collected in
randomized controlled trials. Approaches
onset of the
uterine lining, the survey of 27,143 men-
struating individuals found that increasing
age was associated with an increased risk of
using menstrual cycle tracking apps have
proved powerful because large volumes of
data are available and data collection in real
Anthropocene
heavier bleeding (8). This could suggest that time mitigates recall and recruitment bias. Twelve sites are
altered tissue repair, which is mediated by
immune cells in the uterus and may be less
However, app users are not representative
of the global population because they live
considered for defining
effective in older people, is the mechanism mainly in high-income countries and young, the Anthropocene
by which COVID-19 vaccination increases
menstrual flow.
white, educated individuals are overrepre-
sented (1, 3, 4). App users are also aware of
geological epoch
The evidence for the underlying mecha- being vaccinated, and this may affect their
nism is therefore mixed and could be con- perceptions of aspects of menstruation that By Colin N. Waters1 and Simon D. Turner2
sistent with effects mediated are partially or wholly subjec-

E
by both ovarian hormones (af- tive, such as menstrual flow, arth’s geological history is divided into
fecting cycle length) and endo- “...data about pain, and premenstrual syn- chronostratigraphic units that distin-
metrial repair (affecting men- drome (PMS) symptoms. The guish phases in the planet’s evolution
strual flow). Further research effects on inclusion of a blinded control by summarizing complex biotic, geo-
is required to definitively iden-
tify the pathways involved, but
menstruation group, as in vaccine trials, pre-
cludes these problems.
chemical, and climatic changes. Over
the past century, many components
now that there is evidence that are rarely But there are also opportu- of the Earth system have changed so much
COVID-19 vaccination is asso- nities. The recent advances in that they no longer occur within the ranges
ciated with menstrual changes, collected in defining vaccine effects on the evident during the Holocene—the geologi-
more-involved studies that
track blood hormone levels
vaccine trials.” menstrual cycle open several
avenues of inquiry. Vaccination
cal epoch that represents the past ~11,700
years. There are also distinct geological traces
before and after vaccination is planned and occurs at a sin- that warrant recognition as a new geologic
and studies of immune cells isolated from gle point in time, and thus, importantly, epoch: the Anthropocene. The Anthropocene
endometrial biopsies or menstrual fluid individuals who are already planning to re- Working Group (AWG), a task group of the
are justified. ceive a vaccine dose can be recruited, which Subcommission on Quaternary Stratigraphy
Does SARS-CoV-2 infection affect men- circumvents the ethical challenges of giving (SQS) of the International Commission on
struation? The nature of COVID-19 vaccina- participants an immune stimulus purely Stratigraphy (ICS), have been working to de-
tion makes it amenable to studies that track for experimental reasons. These individuals cide precisely when the Anthropocene began,
menstrual cycle parameters before and after can participate in studies not only to fully with a focus around the mid-20th century.
exposure, but this is harder for infection define how immune stimulation affects The definition will need to identify specific
because it is unpredictable, it may last for female reproductive parameters but also physical properties in sediment layers, or
days or weeks, and many people may be un- to address the converse question: How do strata, that capture the effects of recent in-
aware that they have been infected, making menstrual cycle phase and the use of hor- creases in human population; unprecedented
it more difficult to define an uninfected con- monal contraception affect the immune re- industrialization and globalization; and
trol group. In studies early in the pandemic, sponse? There is an opportunity, finally, to changes imposed on the landscape, climate,
between 15 and 25% of individuals reported start making real progress in an area that and biosphere (1–7).
changes to their periods after SARS-CoV-2 has historically been understudied. j The definitions of chronostratigraphic
infection, although one study was on individ- units form the basis of the International
R EF ERENCES AND NOTES
uals who were hospitalized with COVID-19, Chronostratigraphic Chart (ICC) (8) and
1. E. A. Gibson et al., medRxiv 2022.07.07.22277371 (2022).
one was on individuals with Long Covid, and 2. L. Trogstad, SSRN ssrn.3998180 (2022). standardize the geological time scale—e.g.,
none included an uninfected control group, 3. A. Edelman et al., Obstet. Gynecol. 139, 481 (2022). when specific periods, epochs, and ages be-
so these are likely to be overestimates (13– 4. A. Edelman et al., BMJ Med. 1, e000297 (2022). gin and end and how they can be identified
5. A. Alvergne et al., Front. Reprod. Health 4, 952976 (2022).
15). More recently, the Nurses’ Health Study 6. S. Wang et al., Am. J. Obstet. Gynecol. 227, 739.e1 (2022). in strata. Protocols established by the ICS to
3, which compared menstrual cycle length 7. A. Alvergne et al., medRxiv 2021.11.23.21266709 (2022). formalize chronostratigraphic units require
and regularity self-reported in 2011 to 2016 8. K. M. N. Lee et al., Sci. Adv. 8, eabm7201 (2022). the definition of a global boundary strato-
9. A. R. Lamb, Arch. Intern. Med. XII, 565 (1913).
to that self-reported in 2021 found no effect 10. T. Shingu et al., Kurume Med. J. 29, 123 (1982). type section and point (GSSP). This requires
of SARS-CoV-2 infection, although the tim- 11. S. Suzuki, A. Hosono, Papillomavirus Res. 5, 96 (2018). the selection of a single reference point for
ing of the questionnaires (not immediately 12. J. Needham, East. Horiz. 19, 6 (1980). defining the base (the lowermost part) of a
13. K. Li et al., Reprod. Biomed. Online 42, 260 (2021).
pre- and post-exposure) and the relatively 14. S. M. Khan et al., Am. J. Obstet. Gynecol. 226, 270 (2022).
chronostratigraphic unit, from which cor-
coarse detail in which participants could 15. H. E. Davis et al., EClinicalMedicine 38, 101019 (2021). relation of an isochronous boundary (i.e.,
respond limit the ability of this approach to
ACKNOWL EDGMENTS 1
School of Geography, Geology and the Environment,
detect small or temporary effects (6).
The author receives financial support from Borne, University of Leicester, Leicester, UK. 2Department
There are important lessons to be learned. a charity that funds preterm-birth research. of Geography, University College London, London, UK.
More than half of the world’s population 10.1126/science.ade1051 Email: cw398@[Link]

706 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_16124035.indd 706 11/14/22 2:08 PM

one that is the same age everywhere) for Candidate sites for defining the Anthropocene
the Anthropocene can be achieved globally Core samples from the 12 candidate sites contain various markers of geochemical change and anthropogenic
(9). Given the diversity of geological modi- influence. Globally important stratigraphic markers from each core sample include spheroidal carbonaceous
fication that has occurred in different envi- particles (SCPs), plutonium isotopes, carbon-14 (14C), and d15N [based on the ratio of nitrogen-14 (14N)/15N
ronments, it is challenging to find a single isotopes]. Lines show the full range of each marker’s presence in the core. The point of rapid change of multiple
GSSP that can best represent a global refer- markers provides chronological correlation between the sites.
ence for the Anthropocene.
A candidate Anthropocene GSSP site
must have at least one primary stratigraphic 1
marker to be used as a reference to corre- 11
late the boundary at the selected GSSP site 7 6 12 8
9 10
to other locations across the planet (see the 2
Lake 3
figure). These markers include those that
Coral
indicate a geochemical change—e.g., plu-
Estuary or coastal
tonium isotopes, carbon-14 (14C), and d15N
Anthropogenic
[based on the ratio of nitrogen-14 (14N)/15N
Speleothem (stalagmite) 4
isotopes]—or the appearance of anthropo-
genic particles—e.g., spheroidal carbona- Anoxic marine basin
ceous particles (SCPs), which are a type of Peat
fly ash only produced by high-temperature Ice sheet
combustion of coal or fuel oil (see the photo).
5
Ideally, the site should have precisely dat-
able strata with enough resolution to fix the
onset of the Anthropocene to a specific year Markers: SCPs Plutonium isotopes 14
C d15N Rapid change Peak
(10). The strata, mainly extracted as cores
East Gotland Basin 1
at the sites, must represent a complete time
record, with no gaps or disturbances in the Beppu Bay 2
critical Holocene-Anthropocene boundary West Flower Garden Bank 3
interval. A continuous part of the entire
Flinders Reef 4
core will need to be conserved (for future
reference), and the site should preferably be Palmer Ice Sheet 5
sufficiently accessible that additional cores Ernesto Cave 6
can be extracted if needed.
Building on many years of previous re- Crawford Lake 7
search (10), 12 candidate sites have been pro- Sihailongwan Maar 8
posed for consideration by the AWG. In ad-
Searsville Lake 9
dition to the globally important stratigraphic
markers (plutonium isotopes, 14C, d15N, and San Francisco Bay 10
SCPs), numerous regional markers have been Śnieżka, Sudetes 11
analyzed, including microplastics, heavy met-
als, and even biological proxies such as the Vienna 12
population trends of introduced species (neo- 1830 1850 1870 1890 1910 1930 1950 1970 1990 2010
biota). AWG voting members are tasked with
selecting one candidate GSSP site. several other Anthropocene markers, includ- ter lake with a seasonally stratified water
Two of the candidate sites are in ma- ing d13C (based on the ratio of 13C/12C), d15N, column with a lowermost anoxic layer; and
rine sediments—one in the Baltic Sea and SCPs, microplastics, and pesticides. Searsville Lake in the US, created by a dam
one in the coastal Beppu Bay of Japan. For Two candidate sites are from coral reefs— built in 1892. The Crawford Lake core indi-
the Anthropocene, suitable marine sedi- the West Flower Garden Bank (Gulf of cates ecological effects from indigenous oc-
ments should exhibit sufficient accumula- Mexico) and the Flinders Reef (Coral Sea, cupation followed by European colonization
tion rates, be in contact with anoxic bottom Australia). Here, the skeletons of long-lived and shows major ecological changes in its
waters to limit bioturbation (disturbance corals preserve visible seasonal growth bands microfossils from the 1950s onward, with
by living organisms), and be unaffected by that contain geochemical markers of envi- carbon and nitrogen cycle perturbations
“anthroturbation”—i.e., disturbance by hu- ronmental change, which can be resolved and SCP influx. Sihailongwan Lake is more
man activities, such as trawler fishing (10). precisely back to the 1750s and 1700s, respec- remote and lacks obvious anthropogenic sig-
Both the Baltic Sea and Beppu Bay sites con- tively, allowing for suitable comparison of nals during the Holocene but shows the in-
tain dark, carbon-rich silts and clays tens of Anthropocene and late Holocene successions. creased concentrations of soot (particulates
centimeters thick, with the Beppu Bay site Both sites show clear effects of increased formed as a high-temperature condensate),
GRAPHIC: K. FRANKLIN/SCIENCE

also showing seasonal layering. At both sites, global fossil fuel combustion driving d13C to SCPs, heavy metals, and polycyclic aromatic
these strata lie above the paler, weakly lami- more negative values and proxies indicating hydrocarbons resulting from fossil fuel burn-
nated Holocene sediments, and human activ- warming sea surface temperatures. ing and industrial emissions. Searsville Lake
ity–driven eutrophication (i.e., nutrient en- Three lake sites have been proposed— has silts impounded through seasonal storm
richment causing phytoplanktonic blooms) is Crawford Lake in Canada, which is meromic- events with a layer-count chronology tied to
reflected in the prominent textural and com- tic (i.e., its water layers do not intermix); local pollution and two earthquake events. At
positional changes in the strata along with Sihailongwan Lake in China, which is a cra- Searsville Lake, SCPs, polychlorinated biphe-

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 707

1118Perspectives_16124035.indd 707 11/14/22 2:08 PM

INSIGHTS | P E R S P E C T I V E S

nyls, and pesticides mark the Anthropocene.

The Palmer ice core from the Antarctic
Peninsula has also been proposed. Ice cores
have previously been selected to define a
GSSP for two Holocene subdivisions (11). The
length of the core from the Palmer site that
corresponds to the Anthropocene is by far the
longest among the candidates, at ~32 m. This
provides more material to analyze per year,
which is important given the low concentra-
tions of some markers in this very remote set-
ting. The core displays seasonally resolvable
laminae and records snowfall extremes dur-
ing the Anthropocene, acceleration of atmo-
spheric methane concentrations, and even
small amounts of SCPs.
A core from a raised ombrotrophic peat
bog—i.e., one that receives nutrients and wa-
ter from the atmosphere for growth—near
the Śnieżka peak in the Sudetes Mountains,
Poland, has been proposed. The peat from 5mm
the site shows an increase of fly ash from
thermal power stations, coinciding with A scanning electron microscopy image shows an ~40-µm-diameter spheroidal carbonaceous particle (SCP)
the appearance of pollen from introduced from Crawford Lake, Canada.
plants and perturbations in d13C and d15N,
respectively, that align with the 1950s. signals from industrial and agricultural the ICC. Each of these votes is required
Three proposals have proved less favor- processes with decade-scale variations in to pass by a 60% majority and may fail at
able as the GSSP site: the Ernesto Cave in timing, global signals provide clear in- any stage, in which case it cannot be sub-
Italy, the San Francisco Bay estuary, and a stances of time that can be linked between sequently modified for at least 10 years.
section in anthropogenic deposits in cen- the sites. Hence, the base of strata mark- If approved, the Anthropocene will have
tral Vienna. The Ernesto Cave site, based on ing the beginning of the Anthropocene can a fixed scientific definition with a precise
a core of a stalagmite only 4 mm thick that be recognized and correlated between all of start date that is characterized by a section
accumulated during the Anthropocene, the sites, whether on land, in ice, or beneath in the strata with well-defined markers.
has 14C and sulfur records that lag behind the sea. For instance, one such marker is This will help focus studies on the scale
monitored atmospheric records by 1 to 2 the acute influx of plutonium isotopes cre- and effect of recent human activity on the
decades as a result of vegetation and soil ated by a series of aboveground hydrogen planet, contrasting with the relative stabil-
delaying atmospheric signals reaching the bomb test detonations commencing in late ity of the preceding Holocene. j
cave. The site was discounted because of 1952 (12). The plutonium provides a sharper
REF ERENCES AND NOTES
this delay. The San Francisco Bay estuary global marker than those produced by
1. W. Steffen, W. Broadgate, L. Deutsch, O. Gaffney,
was originally proposed because it has one larger and longer-lasting drivers of Earth C. Ludwig, Anthr. Rev. 2, 81 (2015).
of the best-resolved historical records of system change—e.g., increased burning of 2. M. J. Head et al., Episodes 10.18814/epiiugs/2021/021031
(2021).
neobiotic species, more than half of which fossil fuels and expansion of agriculture— 3. C. N. Waters et al., Science 351, eaad2622 (2016).
were introduced after 1960. The core re- and consequently has been selected by 4. J. Zalasiewicz, C. N. Waters, M. Williams, C. P.
cords increases in mercury and SCPs in the many of the sites as the primary marker. Summerhayes, Eds., The Anthropocene as a Geological
Time Unit: A Guide to the Scientific Evidence and Current
1960s and a paleontological succession of The AWG will select their preferred can- Debate (Cambridge Univ. Press, 2019).
five neobiotic species in the 1980s, with didate GSSP site based on comparison of 5. J. Syvitski et al., Commun. Earth Environ. 1, 32 (2020).
6. J. Zalasiewicz, C. Waters, M. Williams, in A Geologic Time
their stratigraphic appearances mirroring the robustness of the core chronology and Scale 2020, F. M. Gradstein, J. G. Ogg, M. D. Schmitz, G. M.
their known arrival dates. However, be- clarity of physical markers around the on- Ogg, Eds. (Elsevier, 2020), pp. 1257–1280.
cause the age of the succession is relatively set of the Anthropocene. The chosen site 7. M. Williams et al., Palaeontology 65, e12618 (2022).
8. K. M. Cohen, S. C. Finney, P. L. Gibbard, J.-X. Fan,
poorly constrained and the core stopped will be used to compose the age name. For The ICS International Chronostratigraphic Chart
before reaching the key interval at the example, if the Crawford Lake is chosen, (IUGS, 2022); [Link]
[Link].
base of the Anthropocene, this site was then the current age may be named the 9. A. Salvador, International Stratigraphic Guide: A Guide to
also discounted as a GSSP. Finally, a sec- Crawfordian Age, which would be the first Stratigraphic Classification, Terminology, and Procedure
tion of anthropogenic deposits in central age of the Anthropocene epoch. The chosen (IUGS and Geological Society of America, 1994).
10. C. N. Waters et al., Earth Sci. Rev. 178, 379 (2018).
Vienna reflects the city’s development and GSSP level at the site will be used to set the 11. M. Walker et al., Episodes 41, 213 (2018).
can be dated by artifacts, including debris precise beginning of the Anthropocene. 12. C. N. Waters et al., Bull. At. Sci. 71, 46 (2015).
from World War II. It provides a near-200- The AWG will vote on the site in late ACKNOWL EDGMENTS
year history of pronounced urban changes. 2022, which is only the initial stage in a C.N.W. and S.D.T. are, respectively, the Chair and Secretary
However, because of notable gaps during process. It will be followed by a proposal of the AWG. The authors thank J. Zalasiewicz and M. Head for
PHOTO: SARAH ROBERTS

this time interval, the Vienna site is also to be voted on by the SQS, then voted their comments; the Haus der Kulturen der Welt, Berlin, for
its support and collaboration with the AWG and for the many
not suitable as a GSSP. on by the ICS, and finally ratified by the teams involved in compiling the results for the 12 sites; and
Together, the 12 sites provide a diverse International Union of Geological Sciences S. Roberts for supplying the SCP image.
perspective on the geological reality of the (IUGS). Only then will the Anthropocene
Anthropocene. Although showing regional be formally approved as a component of 10.1126/science.ade2310

708 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_16124035.indd 708 11/14/22 2:08 PM

GENETICS

Potential bias in genetic correlations

Mating patterns across two traits can inflate estimates of genetic overlap

By Andrew D. Grotzinger1,2 and both traits. If this were the only issue, the the true sense of the term: The inflation of rg
Matthew C. Keller1,2 interpretation of rg as strictly an index of is a statistical artifact that distorts the corre-
shared biology would require updating, but lation between polygenic scores. Somewhat

G
enome-wide association studies rg would still be a valuable metric with etio- reassuringly, the first effect of xAM on rg will
(GWASs) identify genetic variants as- logical and clinical implications for under- typically be greater than the influence of the
sociated with a trait. Most traits are standing risk for trait Y given trait Z. second effect, particularly when genomic
associated with thousands of vari- The second effect that xAM has on rg is REML is used.
ants, and many variants are pleiotro- potentially more problematic. Commonly In the worst-case scenario, xAM for two
pic, meaning they are associated with used methods for estimating rg, such as link- traits with no overlapping causal variants
multiple traits. Pervasive pleiotropy makes it age disequilibrium (LD) score regression will produce rg estimates that are not due to
impractical to assess genetic overlap between (5), Haseman-Elston regression (6), and ge- pleiotropy and are upwardly biased. Border
two traits by tallying the shared variants. For nomic residual maximum likelihood (REML) et al. use simulation findings and empirical
example, traits such as major depression and (7), assume that there is no long-range cor- rg and cross-mate, cross-trait correlation esti-
anxiety are likely associated with a shared set relation across causal variants. Border et al. mates to show that, for example, rg estimates
of thousands of variants. Genetic correlation find that xAM violates this assumption by between alcohol use disorder–schizophrenia
(rg) estimated from GWASs of a pair of traits inducing long-range correlations between or height–educational attainment may fall
is typically interpreted as an overall measure causal variants that are on average positive into this worst-case scenario. However, for
of genetic overlap, providing a useful met- or negative (i.e., sign-consistent). This causes other pairs of traits, the rg estimates are too
ric for quantifying shared biology between an upward bias in rg estimates after even high relative to the estimated levels of xAM
traits. On page 754 of this issue, Border et a single generation of xAM. Unlike in the to realistically reflect a by-product of xAM
al. (1) report simulation-based and empirical first case where only the interpretation of rg alone. Within psychiatric phenotypes, these
findings that challenge this interpretation. would shift, this second issue causes bias in trait pairs with large rg estimates include
Border et al. investigate the ef- bipolar disorder–schizophrenia and
fect on rg of cross-trait assortative major depressive disorder–anxiety.
mating (xAM), which occurs when Measuring genetic correlations In addition, estimates of rg between
individuals scoring highly for trait Y Cross-trait assortative mating could result in genetic correlations traits that show little or no xAM,
mate with partners who score high with certain traits that are explained by three scenarios, two of such as certain diseases or late-
(or low) for a separate trait, Z. The which inflate the score, rg, upward even in the absence of pleiotropy. onset disorders, are still likely to re-
level of xAM will vary depending on flect pleiotropy. Even for traits that
the pair of traits being considered Cross-trait
Cro
oss-trait do show evidence for xAM, rg will
and is quantified in their study us- assortative
assortati
ative mating typically reflect some combination
ing cross-trait correlations across of three alternative contributions—
spouses. They show that xAM can in- pleiotropy, correlated polygenic
crease rg in two ways (see the figure). scores not due to pleiotropy, and
Trait Y Trait Z
They demonstrate that xAM will in- bias due to long-range correlation.
crease rg due to genetic variants for Therefore, the findings of Border et
trait Y being coinherited with the Child inherit
inherits two
al. should not be interpreted as im-
variants for trait Z (and vice versa). copies of shared plying that rg estimates are wholly
This first potential effect of xAM on causal variant for unrelated to pleiotropy.
rg estimates has been understood for traits Y and Z Their results for case-control
decades as an interpretive caveat in along with variant traits, which compare individuals
family-based studies (2–4), although for trait Y and with a disorder (a case) to those who
variant for trait Z.
it is perhaps not widely discussed. do not have the disorder (a control),
The authors point out that this first should also be interpreted with cau-
If assortative mating occurs in the population,
population the measure
effect of xAM does not cause an ac- of genetic correlation, rg, would reflect three scenarios: tion. As is standard, Border et al.
tual bias in rg estimates because the quantify xAM between case-control
polygenic scores—the cumulative set traits using tetrachoric correlations,
of genetic variants that affect each which estimate the correlation be-
trait—are legitimately correlated. tween the continuous risk scores
GRAPHIC: V. ALTOUNIAN/SCIENCE

This is despite there being no pleio- assumed to underlie two disorders.

tropic genetic variants that affect Because patterns of xAM across all
levels of genetic risk affect rg, this
Shared causal Separate causal variants Upward bias in
1
Institute for Behavioral Genetics, University approach assumes that the same
of Colorado at Boulder, Boulder, CO, USA. variant reflects are inherited together. rg score owing
2 typical pleiotropy rg increases owing to to long-range degree of xAM inferred from mat-
Department of Psychology and Neuroscience,
University of Colorado at Boulder, Boulder, CO, interpretation of r g
. correlated genetic scores correlation ing patterns for cases with extreme
USA. Email: [Link]@[Link] but not pleiotropy. between variants. risk holds across all levels of risk,

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 709

1118Perspectives_16124035.indd 709 11/11/22 4:25 PM

INSIGHTS | P E R S P E C T I V E S

including for mates who are “subthreshold” ORGANIC CHEMISTRY

for a disorder. However, increased mating
between individuals who receive a diagnosis
may occur for reasons, such as an elevated
chance of meeting a partner in a clinical set-
Adding functions to pyridines
ting, that have no bearing on mating that oc- Chemical reactions break a pyridine ring
curs between individuals with subthreshold
risk. Thus, tetrachoric correlations between
to allow its modification
disorders, and perhaps especially between
psychiatric disorders, may overstate the in- By Jung Min Joo oms themselves are electron poor and thus
fluence of xAM on rg estimates. do not react easily with electrophiles. Fur-

P
The findings of Border et al. make it clear yridine (C5H5N) is a launch point for thermore, the nitrogen atom reacts prefer-
that more realistic models for why mates creating a wide range of chemicals, entially with electrophilic reagents to form
correlate within and between multiple traits including those used in drug discov- the corresponding pyridinium salts, which
need to be developed and tested. It may be ery, catalysis, and materials science are more electron poor than pyridines and
that xAM occurs only on a few traits of cen- (1). It consists of a hexagonal ring of thereby less likely to react with electrophilic
tral importance to mating and that most five carbon-hydrogen (C–H) pairs and halogen species. As a result, halogenation
other mate correlations are consequences one nitrogen (N) atom. The synthesis of pyr- requires harsh reaction conditions, usually
of the phenotypic and genetic correlations idine derivatives bearing substituents other involving strong acids and elevated temper-
between those central traits and the traits than hydrogen requires custom-designed atures, to generate very strong electrophiles
that are actually measured (indirect AM). processes. Among them, substitutions of that can enable the reaction. This limits the
It is also possible that some xAM estimates one substituent for another are frequently range of reactants that can be used and the
are due to mates becoming more similar performed while maintaining the integrity products that can be created.
over time (convergence). Assortment can of the ring. However, the efficiency of these Positional selectivity is another impor-
also reflect social homogamy, which oc- methods and the diversity of pyridine com- tant consideration in preparing pyridine
curs when mate choice is based on envi- pounds that they yield derivatives. Although the
ronmental components of the trait that are need improvement (2). most relatively electron-
not due to genetics. This could occur, for On pages 773 and 779 of “...new pathways... rich positions on the
example, if mates are chosen on the basis
of religious beliefs and religion is both due
this issue, Boyle et al. (3)
and Cao et al. (4), respec- generate a wider variety ring (positions 3 and 5)
allow electrophilic sub-
to the environment and affects the traits
being studied. Convergence and social ho-
tively, report different
approaches to breaking
of agrochemicals, stitution, it is difficult to
control which of these
mogamy are not expected to affect rg esti- a pyridine ring, replac- pharmaceuticals, positions is substituted
mates. Deriving more complete models of ing the hydrogens, and or whether one or both of
the mechanisms through which observed then restoring the ring. and material compounds.” the positions are substi-
levels of xAM manifest will be important These techniques present tuted. In addition, some
for obtaining a better understanding of possible new pathways to generate a wider substituents could alter the electronic and
downstream effects on rg. variety of agrochemicals, pharmaceuticals, structural characteristics of the pyridine
Border et al. demonstrate that xAM is and material compounds. ring in ways that enable chemical functions
likely to be pervasive and affect many rg The effects of substituents that replace but may decrease the positional selectivity
estimates to some degree. This bias will in- the hydrogens in a pyridine is an important of the reaction on the ring (7). There are also
herently carry forward to results from any topic in organic chemistry. Halogenation— certain products that require the pyridine
methods that use genetic correlations as the replacement of these hydrogens with to react while positioned alongside other
input, such as Mendelian randomization a halogen atom—is useful as an interme- aromatic rings (such as those with delocal-
(8) or genomic structural equation model- diate step for making pyridine derivatives ized electrons) that would compete with the
ing (9). Therefore, complex trait genetics because the installed halogen can be much aromatic pyridine ring for halogenation.
ignores these problems at its peril. These more readily substituted by another atom or Different strategies addressing the halo-
issues can be addressed by increased care group of atoms as compared with hydrogen genation of pyridines present advantages
in interpreting rg as well as through the de- (5). In addition, the halogen substituents and disadvantages. For example, the use
velopment of methods that can disentangle can facilitate intermolecular interactions of bases instead of acids to mediate halo-
the various contributions to rg estimates. j that are essential for certain desired func- genation tends to generate more reactive
tions such as binding to target proteins (6). intermediates (8). Some indirect methods
REFERENCES AND NOTES
Halogenation of pyridines can be accom- use more reactive and accessible intermedi-
1. R. Border et al., Science 378, 754 (2022).
2. L. J. Eaves, A. C. Heath, N. G. Martin, Behav. Genet. 14, plished through their reaction with elec- ates to avoid the use of strong electrophiles
371 (1984). trophilic chemicals (7). These electrophiles (9, 10). However, the requirements of these
3. M. C. Keller et al., PLOS Genet. 9, e1003451 (2013). are electron poor and can form chemical other strategies can limit the range of pos-
4. G. P. Vogler, J. C. DeFries, Behav. Genet. 15, 111 (1985).
5. B. Bulik-Sullivan et al., Nat. Genet. 47, 1236 (2015). bonds by accepting an electron pair from sible substituents. Thus, more versatile
6. J. K. Haseman, R. C. Elston, Behav. Genet. 2, 3 (1972). electron-rich chemicals. However, because means are needed to prepare halogenated
7. J. Yang, S. H. Lee, M. E. Goddard, P. M. Visscher, Am. J. the nitrogen atom attracts electrons from pyridines from simple pyridines or complex
Hum. Genet. 88, 76 (2011).
8. F. P. Hartwig, N. M. Davies, G. Davey Smith, Genet. carbon atoms of the ring, the carbon at- substituted pyridines.
Epidemiol. 42, 608 (2018). Boyle et al. developed a strategy that
9. A. D. Grotzinger et al., Nat. Hum. Behav. 3, 513 (2019). transforms the electron-deficient pyridine
Department of Chemistry and Chemistry Institute for
Functional Materials, Pusan National University, Busan into an electron-rich molecule, enabling
10.1126/science.ade8002 46241, South Korea. Email: jmjoo@[Link] it to react with electrophilic halogenat-

7 10 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_16124035.indd 710 11/11/22 4:25 PM

ing reagents with high site selectivity (see A sequence of open and closed rings
the figure). The authors were inspired by Generating halogen derivatives of pyridine is typically a slow and inefficient process. However, two approaches
a century-old method that creates reac- that break open the pyridine ring allow it to readily react with a reagent and with high positional selectivity on
tive intermediates by opening the pyridine the ring. The ring can then be restored. Both pathways allow halogenation, and it may be possible to explore
ring. The process, called a Zincke reaction, other ring modifications. C3 and C5 are positions on the ring.
was originally devised to convert pyridines
to pyridinium salts by substituting at the Existing substituent E Electrophile X Electrophilic halogen
nitrogen site (11). The intermediate com-
1 Nitrogen (N) reacts with an
pound during this reaction, known as the C5 C3 E X
X electrophile (E), which hampers
Zincke imine intermediate, has one of the further reactions with an
N Slow
carbon-nitrogen (C–N) bonds broken, thus N additional electrophile (X) such
E N
opening the ring. Boyle et al. adopted the Pyridine as a halogen.
Zincke reaction for installing substituents Pyridinium salt
at carbon atoms of pyridines by using an
electrophilic reagent, trifluoromethanesul-
fonic anhydride [(CF3SO2)2O], and an amine X
additive, dibenzylamine [(C6H5CH2)2NH)], X 2 The pyridine ring can be broken
R R to allow modification with an
to produce electron-rich Zincke imine inter- N N N N electrophilic halogenating reagent.
mediates. The resultant intermediates are R E R E
even more electron rich than conventional
Zincke imine
electron-rich aromatic rings and can out-
compete the latter for electrophilic haloge-
X
nation. Subsequent ring closure to regener- X
ate the pyridine ring was also unexpectedly 3 The aromaticity of the pyridine
O N O N ring can also be broken by
easy and did not cause the highly reactive R" R" attaching a readily removable
Zincke imine intermediates to decompose R' R' R' R' oxazino ring. This allows
through undesirable side reactions. This re- R' R' modification with an electrophilic
action sequence of ring-opening, halogena- Oxazino pyridine halogen.
tion, and ring-closing allowed the authors to
synthesize several different C3-halogenated bromine, and iodine), nitro (NO2), sulfanyl to produce pyridine derivatives with vary-
pyridine derivatives. (SC6H5), and selenyl (SeC6H5) group doing substituents. Apart from electrophilic
Cao et al. took a different path to produce nors. After the substitutions, the attached halogenating reagents, other electrophiles
the same result by transforming pyridines oxazino ring was removed by means of acid that are not typically amenable to pyridine
to electron-rich oxazino (C4N1O1) pyridine treatment to yield the final product. functionalization may now be considered.
intermediates. The oxazino pyridine can The approaches developed by Boyle et al. The ease of producing reactive intermedi-
react with radical species (compounds with and Cao et al. can be performed in a “one- ates with a wide range of installable substit-
at least one unpaired electron) as well as pot” process that only requires the sequen- uents should optimize the creation of drug
with electrophilic reagents to substitute hy- tial addition of reagents without changing candidates and organic functional materi-
drogen with another atom or group of at- the reaction vessel. Both methods exhibit als that were not achievable in the past. j
oms. Similar to the intermediates described high selectivity for the pyridine ring, which
REF ERENCES AND NOTES
by Boyle et al., the formation of oxazino allows the selective halogenation of the ring
1. C. H. McAteer, R. Murugan, J. H. Yamamoto, in
pyridine intermediates is well studied (12), without halogenating other aromatic rings. Comprehensive Heterocyclic Chemistry IV, D. S. Black, J.
but the potential for their use in pyridine They also have high selectivity for the C3 Cossy, C. V. Stevens, Eds. (Elsevier, 2022), pp. 217.
functionalization has not been explored. position within the pyridine ring. Although 2. K. Murakami, S. Yamada, T. Kaneda, K. Itami, Chem. Rev.
117, 9302 (2017).
Cao et al. used inexpensive reagents, in- the two electron-rich positions, C3 and C5, 3. B. T. Boyle, J. N. Levy, L. de Lescure, R. S. Paton, A.
cluding dimethyl acetylenedicarboxylate exist in the intermediates of both synthe- McNally, Science 378, 773 (2022).
(CH3OCOC≡CCO2CH3) and methyl pyruvate ses, C3 is preferentially substituted over 4. H. Cao, Q. Cheng, A. Studer, Science 378, 779 (2022).
5. I. J. S. Fairlamb, Chem. Soc. Rev. 36, 1036 (2007).
(CH3COCO2CH3), to break the aromatic- C5 except in a few cases, such as when a 6. G. Cavallo et al., Chem. Rev. 116, 2478 (2016).
ity of the pyridine ring and make it more substituent is already present at C3. Both 7. M. R. Grimmett, in Advances in Heterocyclic Chemistry,
reactive. This approach is similar to other approaches also allow sequential haloge- A. R. Katritzky, Ed. (Academic Press, 1993), vol. 58,
pp. 271.
known reactions that break the aromatic- nation, which produces pyridines with two 8. M. Balkenhohl, P. Knochel, SynOpen 2, 78 (2018).
ity of the pyridine ring (13). However, these halogen atoms in the ring—first at C3 and 9. S. P. Zucker, F. Wossidlo, M. Weber, D. Lentz, C. C.
other pathways tend to generate intermedi- then at C5. Tzschucke, J. Org. Chem. 82, 5616 (2017).
10. J. S. Wright, P. J. H. Scott, P. G. Steel, Angew. Chem. Int.
ates that are either too stable and do not The techniques of Boyle et al. and Cao et
GRAPHIC: N. CARY/SCIENCE BASED ON J. M. JOO

Ed. 60, 2796 (2021).

allow restoration of the pyridine ring after al. offer opportunities to explore pyridine 11. T. Zincke, G. Heuser, W. Möller, Justus Liebigs Ann. Chem.
substitution or are too unstable and restore functionalization. The preferential haloge- 333, 296 (1904).
the ring before the desired substitutions nation of pyridines over other electron-rich 12. R. Huisgen, M. Morikawa, K. Herbig, E. Brunn, Chem. Ber.
100, 1094 (1967).
can happen. By attaching a readily remov- aromatic rings, with high positional selec- 13. X.-Y. Zhou, M. Zhang, Z. Liu, J.-H. He, X.-C. Wang, J. Am.
able oxazino ring, Boyle et al. created stable tivity, removes the constraints of having to Chem. Soc. 144, 14463 (2022).
intermediates compared with these other generate halogen-containing pyridines at an
ACKNOWL EDGMENTS
approaches. They demonstrated the com- early stage of synthesis to prevent halogena-
The National Research Foundation of Korea (NRF-
patibility of their method using a range of tion of all aromatic rings. Furthermore, se- 2022R1A2C2008629) is gratefully acknowledged.
reagents, including trifluoromethyl (CF3), lective halogenation can be combined with
perfluoroalkyl (CnF2n+1), halogen (chlorine, the diversification of carbon-halogen bonds 10.1126/science.ade5501

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 7 11

1118Perspectives_16124035.indd 711 11/14/22 11:26 AM

INSIGHTS | P E R S P E C T I V E S

POLARITON CHEMISTRY

Using mirrors to control molecular dynamics

An optical cavity mixes molecular vibrations with light and changes chemical reactivity

By Lev Chuntonov and intermediates affect reactions, identify- ments for all the molecules in solution. The
ing the corresponding mechanisms in detail authors observed that vibrational dynamics

P
olariton chemistry is an emerging is difficult. Learning how coupling to the were different only when the polariton state
field that explores whether the prod- cavity changes the way vibrational energy was excited, whereas upon excitation of the
ucts and rates of chemical reactions is dissipated can help in developing cavity- reservoir, the molecules behaved the same in
can be controlled by placing a reac- catalyzed chemistry for other reactions. the cavity as outside the cavity.
tion mixture between a pair of mir- Highly symmetrical Fe(CO)5 undergoes an Chemists quantify the likelihood of reac-
rors (an optical cavity), allowing light ultrafast process of pseudo-rotation in which tions by the associated change in enthalpy
to interact with the molecules through mo- the CO moieties slightly change their angles and entropy. Synchronous vibrations of the
lecular vibrational excitation. In this setup, in such a way that the molecule appears to polariton states presumably reduce the en-
molecules absorb energy from the light and have rotated by 90°, even though it did not tropy of reactants with a potentially large
release it back to the cavity many times over. rotate at all (8). Chen et al. measured the impact on the reaction rate compared with
Despite recent demonstrations (1–3), the de- directional change of the C–O stretching vi- relatively small changes in molecular vibra-
tails of this phenomenon are poorly under- bration using infrared spectroscopy (9) and tions caused by coupling to the cavity (12).
stood (4). On page 790 of this issue, Chen et found that the strong coupling between the However, in contrast to other reported reac-
al. (5) describe the molecular dynamics of C–O vibrations and the cavity slowed down tions (1–3), Chen et al. did not observe sub-
iron pentacarbonyl [Fe(CO)5] in an optical the pseudo-rotation rate by about one-fourth stantial changes in the entropy and enthalpy
cavity. By using ultrafast infrared spectros- of the corresponding value for reactions that of the pseudo-rotation and IVR upon cou-
copy with subpicosecond time resolution, occur outside the cavity. At the same time, the pling of Fe(CO)5 to the cavity. It is not clear
the authors observed how the cavity affects rate of energy redistribution between differ- how these traditional chemical concepts can
the intramolecular energy dissipation and ent C–O vibrations nearly doubled at room be used to interpret polariton chemistry.
rearrangement of CO moieties around the Fe Chen et al. simultaneously excited mole-
atom. The findings add to our understanding cules to the first excited vibrational state and
of how polariton chemistry might be used to “...polariton chemistry from the first to the second excited state. To
perform molecular transformations that are
intractable by conventional synthesis.
might be used to perform simplify the theoretical treatment, research-
ers frequently make approximations by con-
Chemical reactions depend on how a mol-
ecule shares energy between different vibra-
molecular transformations sidering only the two lowest-energy quantum
states of the molecule. This leads to a theo-
tions (imagine a molecule changing the way that are intractable retical model where the coupling strength be-
it vibrates from stretching a chemical bond to tween the molecules and the cavity depends
bending) (6). Such intramolecular vibrational by conventional synthesis.” on the number of molecules being excited
relaxation (IVR) can be important when (13). Such an approach, however, cannot de-
considering polariton chemistry (7). When temperature (25°C) and increased fourfold at scribe experiments where dynamics of higher
the vibrations of many molecules interact 80°C. These findings apparently contradict a excited states must be considered (14). More
with light trapped inside an optical cavity, recent conclusion that the pseudo-rotation fundamental research is needed to bridge be-
the molecules and the cavity are said to be reaction coordinate and the C-O stretching tween different interpretations of the experi-
strongly coupled because the energy excess vibrations do not interact strongly (10), and mental observations.
in one component can be readily distributed emphasize the need for tools that measure Polariton chemistry is in its infancy, and
to the other. This coupling can be described internal vibrational coupling in molecules. the fundamental principles driving uncharac-
by a quantum-mechanical state shared by the When molecules are strongly coupled to terized phenomena are yet to be understood.
combined system (the molecules and the cav- the cavity, each contributes a small vibrational The work of Chen et al. is a step toward a bet-
ity) known as a vibrational polariton. amplitude to the collective polariton state in ter understanding of such principles. j
Reaction coordinates describe the path- which all the molecules vibrate in sync. The
REF ERENCES AND NOTES
ways that reactants take as they transform remaining vibrational amplitude contributes
1. A. Thomas et al., Science 363, 615 (2019).
into products. Vibrations and the reaction to the so-called reservoir states, whose prop- 2. J. Lather et al., Chem. Sci. 13, 195 (2022).
coordinate may interact directly, for example, erties are not entirely understood. When all 3. A. Sau et al., Angew. Chem. Int. Ed. 60, 5712 (2021).
when the strong vibration of a chemical bond the molecules in solution are surrounded by 4. B. S. Simpkins et al., J. Phys. Chem. C 125, 19081 (2021).
5. T.-T. Chen et al., Science 378, 790 (2022).
leads to its cleavage (1), or indirectly, when similar microscopic environments, reservoir 6. S. Karmakar, S. Keshavamurthy, Phys. Chem. Chem. Phys.
the vibrational energy in one bond is trans- states involve unsynchronized vibrations of 22, 11139 (2020).
ferred through other bonds to the one that is individual molecules, akin to molecules out- 7. D. S. Wang et al., J. Phys. Chem. Lett. 13, 3317 (2022).
8. J. F. Cahoon et al., Science 319, 1820 (2008).
cleaved (3). Although such intramolecular vi- side the cavity. However, when the environ- 9. M. C. Thielges, M. D. Fayer, Acc. Chem. Res. 45, 1866 (2012).
brational couplings in the reagents, products, ments slightly differ from one molecule to 10. P. Portius et al., Organometallics 38, 4288 (2019).
another, reservoir states can involve synchro- 11. B. Cohn et al., J. Phys. Chem. Lett. 13, 8369 (2022).
12. G. D. Scholes et al., J. Phys. Chem. Lett. 11, 6389 (2020).
nized vibrations of multiple molecules (11). 13. R. Houdré et al., Phys. Rev. B 52, 7810 (1995).
Schulich Faculty of Chemistry and Solid-State
Institute, Technion–Israel Institute of Technology, Chen et al. used laser pulses to excite Fe(CO)5 14. C. A. DelPo et al., J. Phys. Chem. Lett. 11, 2667 (2020).
Haifa, Israel. Email: chunt@[Link] in solvents, which creates similar environ- 10.1126/science.ade9815

7 12 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_16124035.indd 712 11/11/22 4:25 PM

P OLICY FORUM

DATA REGULATION

Research under China’s personal information law

The new law may present obstacles to some kinds of research

By Xiaojie Li1,2, Yali Cong1,3, Ruishuang Liu1,3 freedom in data processing (1). We provide We suggest a few potential explanations
here a general overview of the effect of the for why, unlike under GDPR, the PIPL does

T
he Personal Information Protection PIPL on ongoing and future transnational not adequately address the needs of con-
Law (PIPL) that came into effect in scientific research through a legal analysis temporary scientific research. Most legis-
the Peoples Republic of China (PRC) of its provisions, and we suggest feasible lators believed that China was particularly
in November 2021 is in line with many solutions for challenges brought by the insufficient in personal information protec-
international standards because it PIPL to transnational scientific research. tion, including in scientific research; thus,
was designed by referring to jurisdic- The PIPL covers personal information of they felt that much stricter provisions were
tions of other nations, especially the provi- people in China, regardless of the location needed to protect personal information
sions of the European Union (EU) General of the processing of information. In the re- rights (2). Also, legislators did not fully con-
Data Protection Regulation (GDPR). Thus, search context, its provisions apply across sider the issues and implications around
the fundamental principles of the PIPL— the entire span of research processing, from scientific research in the big data era, espe-
such as lawfulness, fairness, transparency, the moment of data or biospecimens collec- cially when the secondary use of personal
accuracy, and purpose limitation—are now tion until the disclosure of research results. information can be quite important. In part,
broadly in line with the GDPR. However, All personal information, except through this is because there was a relative lack of
China has not used “academic derogation” anonymization processing that does not re- voices from the scientific community dur-
ILLUSTRATION: DAVIDE BONAZZI/SALZMAN ART

or “academic exemptions” as have been late to an identified or identifiable person, ing the legislation process, a result of the
used in European nations, the United States, is included in the regulation scope of the legislation system in China.
and Australia to reconcile the protection of PIPL, no matter whether it was processed With the law having taken effect only a
personal data rights with scientific research by scientific researchers, government agen- year ago, there has not yet been a great deal of
cies, companies, or others and no matter systematic monitoring and analysis of the im-
1
Department of Medical Ethics and Law, Peking University whether it was provided knowingly by peo- plementation and impacts. Anecdotally, how-
Health Science Center, Beijing, China. 2Department of ple or collected without the knowledge of ever, there is an impression that many ethical
Situation and Policy, University of International Business those whose information was collected (for review boards and researchers are still in con-
and Economics, Beijing, China. 3National Institute of Health
Data Science, Peking University, Beijing, China. example, by cameras in public spaces and fusion or unclear about what is considered
Email: lrs1187@[Link] cookies on web browsers). legal under the PIPL. Many researchers may

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 7 13

1118PolicyForum_16127759.indd 713 11/11/22 4:26 PM

INSIGHTS | P O L I C Y F O RU M

still be using broad consent in their research SEPARATE CONSENT which consent is sought, the corresponding
and may not know how PIPL lacks space for Like the GDPR, the PIPL requires “specific types of information should be listed under
this. Some researchers may be eager to con- consent,” which means obtaining consent each processing item, and consent should be
sider ways by which broad consent for big for a specific purpose, such as developing obtained separately for each option.
data research may be made compatible with a drug for diabetes. But unlike the GDPR,
the PIPL. PIPL also requires “separate consent,” SPECIMENS AND DATA TRANSPORT
meaning that the authorization of multiple ACROSS BOUNDARIES
PURPOSE LIMITATION general types or features of processing (for Compared with the relatively low limita-
The PIPL requires that personal informa- example, processing of genetic information, tion on free transfer of data and specimens
tion shall be processed for explicit and rea- processing of health care information, or under the GDPR, transport of data and
sonable purposes and be directly related to change of information processors) cannot be specimens across boundaries in China is
the purpose of processing in a manner that bundled together to fall under a single um- subject to stricter requirements under the
has a minimum impact on the rights and brella authorization of individual consent. PIPL. Personal information, reaching up to
interests of individuals (article 6). If there Instead, if there are multiple general types a specified quantity collected and generated
is any change in the purpose or method of or features of sensitive personal informa- within PRC territory by critical information
the processing of personal information, or tion processing for which consent is sought, infrastructure (CII) operators and personal
the category of processed personal infor- consent options should be set separately for information processors must be stored do-
mation, the individual is to be informed each (4). The PIPL stipulates five scenarios mestically (article 40) (for example, trans-
and their consent obtained again. The PIPL in which obtaining separate consent is report abroad of genetic information on 500
expects researchers to gain consent from quired, three of which are particularly rel- people or more needs administrative permis-
the participant even if they have been dei- evant for transnational scientific research: sion). If scientific researchers require trans-
dentified, except by using “anonymization (i) when processing sensitive personal infer of personal information beyond the PRC,
processing” (article 4). This severely limits formation, such as biometric identification, they must obtain separate consent from the
the conditions of processing sensitive infor- participants and also meet one of the follow-
mation without specific consent from the ing conditions: (i) pass a security assessment
participant. Primary investigators (the pro- “A repeated informed-consent organized by the Cyberspace Administra-
cessors) are required to seek voluntary and
explicit consent from participants for the
process…may dilute tion of China (CAC) in accordance with the
provisions of article 40; (ii) be awarded per-
use of data or samples obtained from them the purpose of research and sonal information protection certification
by specified groups for a specific purpose by a professional institution in accordance
that is necessary for the original study. For what it can achieve…” with provisions of the CAC; (iii) enter into a
example, medical research institutions are contract with the foreign recipient in accor-
typically not allowed to collect genetic in- health care, specific identification, financial dance with the standard contract formulated
formation for research use unless they have accounts, religious beliefs, personal loca- by the CAC, stipulating the rights and obliga-
a specific purpose (such as a multinational tions, and personal information of minors tions of both parties; or (iv) adhere to other
clinical trial) and sufficient necessity and under the age of fourteen (articles 28 and conditions stipulated by laws, administrative
take strict protective measures (article 28). 29); (ii) when personal information is pro- regulations, or the CAC (article 38).
The process of scientific research and inno- vided to any processor outside the territory The presence of article 38(2) offers
vation in this era of big data is dynamic, full of of the PRC (article 39); and (iii) when the prospects for the free flow of data across
uncertainties and possibilities. Determining information processors change (article 22). boundaries under international treaties
all potential future specific purposes of data Draft regulations on the management of and agreements concluded or acceded to by
processing at the time of data collection is online data security define separate consent China that stipulate the conditions for pro-
often not possible. If specific and explicit con- as “the data processor obtains individual viding personal information outside China
sent is required before the scientific research consent for each item of personal informa- (for example, if China acceded to an agree-
process begins, it may be almost impossible tion when carrying out specific data pro- ment with other countries similar to the
for many researchers to lawfully process the cessing activities, excluding the consent for EU-US and Swiss-US Privacy Shield for com-
information. It may hinder advances in sci- multiple personal information and multiple mercial purposes). Furthermore, scientific
ence and technology and the free flow of data. processing activities at one time” (5). We be- researchers are to take necessary measures
Furthermore, obtaining specific and explicit lieve that separate consent refers to obtain- to ensure that the processing of personal
consent for secondary research often needs a ing personal consent for each type—such as information by foreign recipients meets the
disproportionate amount of effort, and find- biometric identification, financial accounts, personal information protection standards
ing the sources of data and specimens could and personal information of minors under stipulated in the PIPL under article 38(3). No
be difficult or impossible. Such specification the age of 14—rather than obtaining personal specifications are available for determining
and identification of each type of process- consent for each item. To define the scope of whether the foreign receptions have met the
ing without any derogation of personal data “separate,” look to three points from article PIPL’s standard, which leads to uncertainty
rights can substantially burden scientific re- 23 of the PIPL: (i) The data processor has to and requires the CAC to publish a standard
searchers with additional bureaucratic work. perform specific data processing behavior contract for that purpose (6).
A repeated informed-consent process that or collect a specific type of data, (ii) the data
protects participants may dilute the purpose processor has to inform each participant in- INDIVIDUAL RIGHTS AND DEROGATIONS
of research and what it can achieve while dividually, and (iii) the process of obtaining FOR TRANSNATIONAL SCIENTIFIC
bringing unnecessary disturbances to par- informed consent has to be independent of RESEARCH
ticipants (3). An exception from the purpose other authorizations or should not be con- One key tenet of the PIPL, outlined in article
limitation principle may promote the use of founded with the provision of other services. 1, is that it prioritizes the rights of the data
data under many research scenarios. If multiple types of information are used for subjects (2). Accordingly, to ensure legiti-

18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118PolicyForum_16127759.indd 714 11/11/22 4:26 PM

macy and to comply with PIPL, researchers medical Research Involving Human Subjects To ensure that research is considered legal
should ensure that subjects understand sev- allow the donated sample and related infor- under the PIPL, we suggest some key steps.
eral rights when obtaining consent: the rights mation to be used for all kinds of medical re- First and foremost, awareness of the require-
to be informed, to make decisions about the search once the biological sample donor has ments of PIPL must be increased, and the
processing of their personal information, to signed the informed consent form (7). In the law enforced, in transnational scientific re-
access their information, to copy their in- absence of restrictions, this kind of “blanket search efforts. Institutions such as the Peking
formation, to correct and supplement their consent” allows samples and data to be used University Health Science Center–Michigan
information, as well as the right to erasure in future studies that may conflict with the Medicine (PKUHSC-MM) Joint Institute for
(to be forgotten), and the right to portability individual’s fundamental values. However, Translational and Clinical Research should
(to move data from A to B). Procedures for according to article 88 of the Legislation Law update informed consent forms and process
realizing these rights are elaborated, and re- (8), this provision became invalid after the personal information within the process-
quirements for processing sensitive personal Regulations on the Administration of Hu- ing purpose and scope. Institutional review
information such as genetic data are stricter: man Genetic Resources (HGR regulation) boards must be in place to ensure that the
“Only when there is a specific purpose, suffi- and the PIPL took effect. Draft Ethical Re- PIPL has been followed through training,
cient necessity and strict protection measures view Measures for Life Sciences and Medical guidance, and ethical review. Additionally,
are taken, the processor can handle sensitive Research Involving Human Subjects clearly foreign researchers can promote the formu-
personal information” (article 28). support broad consent by allowing second- lation of a treaty with China on transnational
However, there are legal bases for a certain ary research to be carried out without recon- research or join a common international
degree of derogation from the data subject’s sent under a wide range of circumstances (9). convention in accordance with article 38(2).
right to be informed and make decisions However, these provisions are contradictory Research experts should work toward con-
(article 44: otherwise provided by laws and to the PIPL, and the draft finally did not pass. sensus and provide proposals to the National
administrative regulation), the right to delete Under PIPL, scientific researchers may be People’s Congress to characterize what con-
(article 47: the circumstances specified in granted exceptions from the requirement stitutes scientific research and to legitimize
article 18, paragraph 1, and article 35), and to collect consent when (i) performing legal exemption from portions of the PIPL for sci-
the right to access and copy (article 45: if the duties or legal obligations; (ii) responding entific research exemptions, such as the der-
retention period stipulated by laws and ad- to public health emergencies or providing ogation of rights, the substitution of broad
ministrative regulations has not expired, or protection to the life, health, and property of consent for specific consent under specific
it is technically difficult to delete personal individuals in emergencies; (iii) conducting conditions, and the relaxation of purpose lim-
information), although these derogations are news reports, public opinion supervision, and itation. The advantages and disadvantages of
not as detailed as those in the GDPR. other acts for the public interest and process- applying PIPL should be evaluated as this law
ing personal information within a reasonable is being implemented. We believe that legis-
LACK OF SPACE FOR BROAD CONSENT scope; (iv) processing personal information lative adjustments should be made to ensure
Broad consent is an alternative to specific that has been disclosed by individuals them- better balance among the rights and interests
consent only in cases in which it is not fea- selves or that has been legally disclosed in of data subjects, the wellness of the human-
sible to obtain specific consent for scientific accordance with the provisions of this law, kind, and the interests of researchers. j
research (such as the situation in which the within a reasonable scope; or (v) there are
REF ERENCES AND NOTES
purpose of future research is uncertain and other circumstances stipulated by laws and
1. M. Mourby et al., Int. Data Priv. Law 9, 192 (2019).
the secondary use of personal information is administrative regulations (article 13). In the 2. C. Xiao, Understanding and Application of Personal Infor-
required in the future). Unlike blanket con- above scenario, the PIPL allows processing mation Protection Law (China Legal Publishing House,
2021).
sent or general consent, broad consent must of personal information without consent but 3. D. Nicol et al., Science 364, 445 (2019).
inform potential research subjects of certain does not deprive individuals of their right to 4. X. Zhang, Interpretation of “China Personal Information
matters concerning the long-term storage their personal information. But these scenar- Protection Law” (People’s Press, (2021), p. 245.
5. CAC,“Regulations on the management of online data
of biological samples or future studies with ios would not likely apply to most transna- security (draft for comments),” November 2021; http://
uncertain purposes: risks, benefits, confiden- tional scientific research, which means that [Link]/2021-11/14/c_1638501991577898.htm.
6. A. Lee et al.,“Seven major changes in China’s finalized
tiality, right to refuse participation without the PIPL lack space for broad consent. Fortu- personal information protection law” (Brookings, 2021);
reason, right to withdraw at any time, per- nately, article 13(7) points out the direction of [Link]
sonal information processor information, efforts in relation to scientific research: new changes-in-chinas-finalized-personal-information-
protection-law.
whether biological samples are used for com- laws and administrative regulations can be 7. The National Health Commission of the People’s Republic
mercial benefit-sharing or whole-genome se- put forward in the future to support the legit- of China (NHC),“The ethical review measures for biomedi-
cal research involving human beings (October 2016);
quencing, a brief description of the field or imacy and utility of broad consent. [Link]
type of research involved, storage duration, [Link]?id=84b33b81d8e747eaaf048f68b174f829.
and research contacts and contact informa- SUGGESTIONS FOR THE FUTURE 8. The Ninth National People’s Congress, The Legislation Law
of the People’s Republic of China (March,2000); http://
tion. It is worth emphasizing that if specific Although PIPL emphasizes personal infor- [Link]/xinwen/2015-03/18/content_2835648.
consent is feasible after the purpose of future mation rights and interests, it does not fully htm.
9. NHC, The National Science and Technology Ethics Com-
research is determined, then specific consent consider the special circumstances of trans- mittee Measures for ethical review of life sciences and
should still be preferred. national scientific research in the big data medical research involving human beings (March 2021);
Unlike the GDPR, which includes provi- era. It does not recognize how processing of [Link]
[Link].
sions that seem to offer clear support for personal data for research is different from 10. J. Bovenberg et al., Science 370, 40 (2020).
broad consent in scientific research, the PIPL other uses such as commercial purposes,
ACKNOWL EDGMENTS
requires specific consent for secondary use of neither do they notice that transnational
All authors contributed equally to this work. Partial support
personal information and specimens, which scientific research can benefit society while came from the PKUHSC-MM Joint Institute for Translational
is inconsistent with research traditions and balancing data subjects’ rights (10). These and Clinical Research.
current ethical review guidelines. For ex- concerns have been raised in the scientific
ample, the Ethical Review Measures for Bio- research community. 10.1126/science.abq7402

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 7 15

1118PolicyForum_16127759.indd 715 11/11/22 4:26 PM

B O OKS et al .
HISTORY OF SCIENCE

The Gaia letters

Contextualized correspondence traces the emergence of a provocative hypothesis

By Paul Falkowski 1970, when Margulis, on the advice of James Lovelock and Lynn Margulis grappled with
her ex-husband Carl Sagan, first wrote to the possibility that Earth’s life and its geosphere are

W
riting Gaia, edited by Bruce Clarke Lovelock, seeking his assistance in fleshing intertwined in a single, synergistic system.
and Sébastien Dutreuil, is a fasci- out research questions related to the contri-
nating read that reproduces and bution of biological entities to the planetary neers in a field that would become geobiol-
contextualizes a four-decade-long atmosphere. (That letter, unfortunately, is ogy. Vernadsky published Biosfera in Russian
conversation between environ- not included in this collection.) in 1926, which was informally translated into
mental scientist James Lovelock The first letter in the book, dated English in the 1980s but not formally pub-
and evolutionary biologist Lynn Margulis 11 September 1970, from Lovelock states, lished in English until 1998.
from which emerged the provocative Gaia “I am in the course of writing a paper on the In the foreword of the 1998 edition,
hypothesis, which posits that Earth and Earth’s atmosphere as a biological cybernetic Margulis wrote, “[Vernadsky] illuminates the
all its inhabitants can be thought system.” The editors explain that difference between an inanimate, mineralog-
of as a single, synergistic, self- by 1969, Lovelock had begun to ical view of Earth’s history, and an endlessly
regulating system. think about Earth as a planetary dynamic picture of Earth as the domain and
The book opens with a foreword ecosystem, and the term “biologi- product of life, to a degree not yet well under-
by the late Lovelock, in which he cal cybernetic system” became a stood.” Lovelock disagreed; to him, Gaia was
discusses a conversation he had “shorthand trademark” for Gaia. original. In June 1986, he wrote of Vernadsky
in 1967 with William Golding— Although much of the cor- to Margulis, saying, “he was a middle weight
the Nobel Prize–winning author respondence is relatively trivial, expressing his ideas in a vague and all-
of Lord of the Flies—about Erwin Writing Gaia detailing, for example, when and inclusive manner and with the support of
Schrödinger’s book What Is Life? Bruce Clarke and where they would meet, the editors little or no testable evidence.”
“According to this book,” observes Sébastien Dutreuil, Eds. did an exceptional job of culling Regardless of its originality, Lovelock per-
Lovelock, “life was a process that Cambridge University the myriad letters and—with help sisted in developing Gaia as a theory and, to
Press, 2022.
reduced the entropy of a system 510 pp.
from students and colleagues—in- his death, defended the notion of a positive
while excreting entropy to the en- terpreting the correspondence and feedback between the biosphere and geo-
vironment.” Lovelock wanted to understand revealing how the Gaia hypothesis evolved. sphere. He and Margulis thought the concept
how Mars, with an atmosphere composed By the early 1970s, we learn, Margulis had was as seminal as Darwin’s theory of evolu-
almost entirely of carbon dioxide, is of “high become a major force for developing the idea tion. Indeed, in 1995, Lovelock wrote, “Where
entropy” and therefore probably lifeless, that Earth systems function through feed- organisms affected their personal environ-
whereas Earth, containing both methane backs between life and the geosphere and ment then the tendency could be inherited
and oxygen, is of “low entropy” and sustains gave Lovelock increasing confidence in the and could become extensive, even global.”
life. “If you intend to put forward an idea like notion that “Gaia has the equivalent of a cen- At almost 500 pages, Writing Gaia is a
that,” Golding replied, “you had better give tral nervous system.” This notion would even- weighty read. The more casual reader can
the low-entropy system that is our planet a tually lead to heated debates among many skim the correspondence and read the edi-

PHOTO: TUI DE ROY/MINDEN PICTURES

proper name, and I suggest the name Gaia.” scientists, and the concept was, and still is, tors’ excellent summaries. Others will find
The book goes on to document the challenged, if not downright dismissed as an the details of the correspondence a fascinat-
four-decade-long correspondence between untestable hypothesis. ing illustration of how ideas between scien-
Lovelock and Margulis, which began in Arguably, one of the key questions sur- tists evolve. I am sure this tome will be used
rounding the Gaia hypothesis is how original by many students and scholars to under-
The reviewer is at the Department of Earth and Planetary it was. The editors address this issue, discuss- stand the history of geobiology, writ large,
Sciences and the Department of Marine and Coastal
Sciences, Rutgers University, New Brunswick, NJ, USA. ing the contributions of Vladimir Vernadsky, in the 20th century. j
Email: falko@[Link] a Soviet scientist who was one of the pio- 10.1126/science.ade8916

7 16 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118_Books_16142600.indd 716 11/11/22 4:27 PM

INSI GHTS

SCIENCE AND SOCIETY How Far the

Light Reaches:

Aquatic life and its parallels A Life in Ten

Sea Creatures
Sabrina Imbler
Little, Brown, 2022.
Blurring the personal and the scientific, an author probes 320 pp.

the beauty, terror, and richness of the natural world

where, along hydrothermal vents, the yeti
By Juli Berwald death of a relationship with a lover. crab dances with its claws in the air, bring-
As a writer, Imbler is a terrific talent, ing sulfur-laden water to the chemosynthetic

I
n 1996, Scientific American staff writer at once commanding and questioning, un- bacteria that sustain it. In the dark nights and
John Horgan published The End of flinching and vulnerable, factual and po- gay bars of Seattle, Imbler, too, learns what
Science, which posited that all the great etic, polished and raw, specific and general. it means to dance to be alive, finally among
discoveries had been made and science The book opens with the words “The truth a community of people who make them feel
was reaching its twilight (1). But from is,” which the reader should take seriously. seen and known.
within those pages, a loophole emerged. Imbler bares their soul and leaves no room The theme that pulls these essays to-
In one chapter, physicist David Bohm pointed for platitudes. gether is that the distinctive characteristics
out that art and science, now separate pur- Some of the book’s most powerful essays that make the author who they are, which
suits, have not always been so. During the revolve around Imbler’s growing under- are often framed as outliers and differences,
Renaissance, for example, the two intermin- standing and recognition of their queerness. are all reflected throughout the natural
gled, and great progress was made. They contemplate the biology of a goldfish, world. Nature’s greatest gift is its creativity
But as science became more objective, confined to a bowl growing ever more pol- and diversity.
it also became more erudite and less relat-
able. Facts, although powerful, eventually
cave under the weight of emotion. Vaccine
hesitancy and climate denial, to name two
big examples, suggest the limits of sci-
ence, which is constrained not by all pos-
sible discoveries having been made but by
the fact that humans tend to comprehend
ideas yet act on feelings.
What would happen if art, whose main
purpose is to make us feel, again merged
with science? Writer and science journalist
Sabrina Imbler’s second book, How Far the
Light Reaches, points toward an answer. The
book’s 10 essays, each musing on a different
aquatic creature, upend paradigms of culture
and history and force readers to confront
their assumptions about society.
Imbler’s pieces often start in one place
and end up somewhere else entirely. “Beware
the Sand Striker” describes what it means
to be preyed upon, bringing together pain-
ful stories of Imbler’s sexual encounters with Like a goldfish in a bowl, we can survive in stifling environments, but we thrive when our needs are fully met.
men and the whip-fast strike of a polychaete
worm that hides in the sandy seafloor, invis- luted. The same chapter tells of Imbler’s But there is something else going on in
ible to the fish and crabs above. It also tells struggle as an overachieving rule-follower Imbler’s work. Like Horgan did in 1996,
the true story of Lorena Bobbitt, whose mar- in high school. And then, years later, their Imbler may make readers uncomfortable.
ried surname has been exploited as a nick- first gay encounter with someone unex- Not because they portend a dismal future,
name for the same worm. pectedly from their hometown. “Release a but because, with brutal candor and elegant
In a chapter called “Hybrids,” Imbler goldfish, and it will never look back,” Imbler metaphor, How Far the Light Reaches reveals
explores “The Question” of being biracial, writes. “Nothing fully lives in a bowl; it only the gap between where we are today and a
intertwining it with the story of a hybrid learns to survive it.” truly inclusive and connected world. In so
PHOTO: MATT LEE/ALAMY STOCK PHOTO

butterflyfish collected in Australia in the In another essay, contemplating the seem- doing, it also threads the loophole, weaving
1970s, in the process turning a searing eye ingly boundless morphing of a cuttlefish’s the outlines of a future where art and science
on a celebrated ichthyologist. In “How to color and texture to avoid predation, Imbler amplify one other. j
Draw a Sperm Whale,” Imbler describes asks, “But once the danger passes, you may
REF ERENCES AND NOTES
whale necropsy—the animal equivalent of feel tempted to seize this metamorphic power
1. J. Horgan, The End of Science: Facing the Limits
an autopsy—and then uses it to dissect the for means beyond escape. When you are not of Knowledge in the Twilight of the Scientific Age
fleeing, what will you become?” (Broadway Books, 1996).
The reviewer is the author of Life on the Rocks:
Building a Future for Coral Reefs (Riverhead Books, Part of a solution is discovered in the es-
2022). Contact: [Link] say “Pure Life,” which looks into the deep sea 10.1126/science.ade9267

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 7 17

1118_Books_16142600.indd 717 11/11/22 4:27 PM

INSIGHTS

infrastructure. The legislation designates never in the public interest to destroy

LET TERS Permanent Protection Areas around water Permanent Protection Areas and risk pro-
bodies to guarantee the health of riparian voking disasters.
areas that can buffer floods. According William E. Magnusson
Edited by Jennifer Sills to the law, construction is forbidden in Instituto Nacional de Pesquisas da Amazônia,
Manaus, Amazonas, Brazil.
these important ecosystems. The law also
Email: wemagnusson@[Link]
Editorial Retraction protects riparian systems in the Amazon,
which has heavy rainfall, fragile soils, and REF ERENCES AND NOTES
On 21 July 2017, Science published the often superficial water tables. 1. H. Cristaldo, “Amazonas registra segundo desabamento
de ponte na BR-319,” Agência Brasil (2022); https://
Report “Chiral Majorana fermion modes Earth-moving companies are the pri-
[Link]/geral/noticia/2022-10/
in a quantum anomalous Hall insula- mary beneficiaries of the law’s loophole. amazonas-registra-segundo-desabamento-de-ponte-
tor–superconductor structure” by Q. L. The companies argue that the cost sav- na-br-319 [in Portuguese].
He et al. (1). Readers who failed to repro- ings of bulldozing earthen access ramps 2. C. Rosa et al., Aquat. Conserv. Mar. Freshw. Ecosys. 31,
1548 (2021).
duce the findings requested raw data files across protected riparian areas justifies 3. “Queda de pontes e viadutos no Brasil,” Folha de São
from the authors, which they provided. the added risks. However, the embank- Paulo (2018); [Link]
Subsequently, the provenance of the raw ments they build funnel the main flow galerias/1617195713692955-queda-de-pontes-e-
viadutos [in Portuguese].
data came into question; additionally, an of flood waters around the pylons of 4. Câmara dos Deputados, “Legislação Informatizado
analysis of the raw and published data bridges, increasing the pressure and the – Lei n° 12.651, de 25 de maio de 2012 – publicação
revealed serious irregularities and dis- likelihood of collapse. In some cases, the original” (2012); [Link]
fed/lei/2012/lei-12651-25-maio-2012-613076-publica-
crepancies. These issues have caused the bridge holds up and only the access ramps [Link] [in Portuguese].
editors at Science to lose all confidence in are washed away, after which the earth-
the conclusions of the paper, and we are moving companies are paid to restore 10.1126/science.adf2868
therefore proceeding with an Editorial the road. The companies benefit, perhaps
Retraction. Authors A. L. Stern, J. Wang, unwittingly, but the economy and local
and B. Lian agree with this decision.
Authors Q. L. He, L. Pan, X. Che, G. Yin, E.
population suffer.
The Forest Code is good, evidence-based
Welcoming Taiwan’s
S. Choi, K. Murata, X. Kou, Z. Chen, T. Nie,
Q. Shao, Y. Fan, K. Liu, J. Xia, and K. L.
policy, but it must be enforced wisely to
maximize its impact. The loophole was
diaspora scientists
Wang disagree with this decision. Authors included so that infrastructure truly in the Academia Sinica, the highest research
E. C. Burks and Q. Zhou did not respond. public good, such as bridges required for academy in Taiwan, recently announced
Author S.-C. Zhang is deceased. access, could be constructed in the ripar- 19 newly elected Members and 3 Honorary
This Retraction replaces the Editorial ian zone. The policy was not designed Members (1) after a 2-year pandemic delay.
Expression of Concern posted on 16 to license the wholesale destruction of Because in 2020, the Taiwanese authority
December 2021 (2). protected areas, along with their essential declared that only Taiwanese citizens could
H. Holden Thorp ecosystem functions in flood mitigation. be nominated as Members of Academia
Editor-in-Chief Brazil must acknowledge that it is almost Sinica (2), another five candidates remain
REFERENCES AND NOTES
1. Q. L. He et al., Science 357, 294 (2017).
2. H. H. Thorp, Science 374, 1454 (2021).

10.1126/science.adf7575

Brazil’s preventable
bridge disasters
In October, a second bridge collapsed
within 10 days on Brazil’s controversial
BR 319 highway (1), as biologists predicted
could happen (2). Bridges collapse regu-
larly in Brazil, and the events are often
associated with flooding (3). Brazil’s envi-
ronmental legislation theoretically protects
locations that are likely to flood from
development, but a legal loophole allows
infrastructure projects to proceed if they
PHOTO: BRUNO KELLY/REUTERS
are considered to be in the public interest.
If Brazil’s government does not take action
to restrict use of the loophole, preventable
disasters will continue to occur.
Brazilian Law 12,651, known as the
Forest Code (4), was enacted to pre-
vent natural disasters and protect A loophole has allowed construction in Brazil’s protected areas, contributing to multiple bridge collapses.

7 18 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Letters_16183348.indd 718 11/14/22 4:41 PM

pending (1). This decision undermines the
inclusiveness of Academia Sinica and the
scientific community in Taiwan.
Many small countries and territories—
including Taiwan—have suffered from
“brain drain” (3) as scientists leave to
study and work elsewhere. Elections of
diaspora scientists into the academies
of their home countries and territories
provide an opportunity to keep them
connected to their homelands, in some
cases even influencing their decision to
return (4–6). The Hungarian Academy of
Sciences has External Memberships for
diaspora Hungarians, as well as Honorary
Memberships for non-Hungarian foreign-
ers. In Colombia (7), Israel (8), mainland
China (9), and African countries (10), The Asian elephant’s
efforts to connect with diaspora scien- anatomy could hold the
tists include research scholarships and key to questions in a
cooperation initiatives that facilitate posi- variety of scientific fields.
tive interactions between overseas scien-
tists and scientific developments in their
native countries.
LIFE IN SCIENCE
A small homeland like Taiwan can ben-
efit from enhancing the local influence
of scientists who have left (11). The new The elephant in the inbox
citizenship requirement for member elec- “We’ll need to get you some boots,” the veterinarian said to me, pointing toward
tions at Academia Sinica in Taiwan could my clean, shiny dress shoes. “Excuse me?” I replied naively, beginning to regret my
instead decrease the engagement of sci- decision to check my emails at 5 p.m. on a Friday. I had come to work dressed for an
entists abroad (12). Given the challenges after-work dinner date, not for what I was about to experience.
Taiwan faces, the Taiwanese authority An hour before, I had been shutting down the microscope and finishing up
should adapt its policy to be more wel- what had been a normal week as a postdoc in a muscular metabolism lab here
coming of its diaspora scientists. In in Copenhagen. Looking forward to what I thought would be a romantic evening, I
addition to piloting an External Member scanned my inbox one last time. Seven new emails caught my attention, all sharing a
system, similar to those in Hungary and title: Re: Elephant.
Columbia, Taiwan should counter brain Copenhagen Zoo houses more than 3000 animals, and despite receiving excellent
drain by increasing the opportunities care, one of them occasionally requires euthanasia. The tragic death of such a unique
available to overseas scientists and animal can provide a valuable opportunity for scientists. The Asian elephant (Elephas
adopting more flexible citizenship maximus) is one of the largest existing land mammals, and
requirements. powerful animals require powerful muscles. By studying how
Call for submissions an elephant’s muscles function, we may be able to better
Jianan Huang
Nanyang Centre for Public Administration, Life in Science is
an occasional feature understand how our own muscles work.
Nanyang Technological University, 639798
Singapore and National Institute of Education,
highlighting some In my career, I have worked with mice, rats, and worms. Yet
of the humorous or when I clicked on the emails, I found a request to cycle to the
Nanyang Technological University, 639798
unusual day-to-day
Singapore. Email: jhuang067@[Link] local zoo immediately with some ice and collect a sample of
realities that face
REFERENCES AND NOTES
our readers. Can you elephant muscle from the front gate.
top this? Submit So there I was, in my best shoes, clutching my box of ice and
1. Academia Sinica, Academia Sinica Newsletter (Eng. Ver.) your story to www.
1769 (2022). [Link]. suddenly realizing why boots were going to be required. There
2. Legislative Yuan of ROC, Legislative Yuan Bulletin 72, 109 would be no quick handoff at the zoo entrance as the emails
(2020) [in Chinese].
3. M. U. M. Anas, S. I. Wickremasinghe, Sci. Public Pol. 5, 37
had implied. Over the next 4 hours, I assisted a team of veterinarians in the post-
(2010). mortem of this gigantic animal, as a forklift and heavy machinery harvested organs
4. Y. Wang et al., Bull. Chin. Acad. Sci. (Chin. Ver.) 1, 34 to send to teams of researchers across Europe. I learned that tissue would be used in
(2019) [in Chinese]. research for vaccines, cardiac disease, diabetes, and infectious disease, as well as for
PHOTO: WPHIL MCLEAN/FLPA/SCIENCE SOURCE

5. C. Cao, China’s Scientific Elite (Routledge, 2004).

6. C. Cao, Science 278, 785 (1997). my lab’s research into muscular weakness.
7. J.-B. Meyer et al., Sci. Technol. Soc. 2, 285 (1997). It was dark by the time I stored the tissue. Five hours late for my dinner, I left the
8. O. Rabinowitz, Y. Abramson, Soc. Stud. Sci. 2, 52 (2022). boots behind and went to beg forgiveness. Fortunately, my date forgave me. We now
9. D. Zweig, S. F. Chung, D. Han, Sci. Technol. Soc. 1, 13
(2008). live together, and we have a no-email rule for Friday nights.
10. H. Pellerin, B. Mullings, Rev. Int. Polit. Econ. 20, 89
(2013). Christopher T. A. Lewis
11. Y. Shain, A. Barth, Int. Org. 57, 449 (2003). Xlab, Center for Healthy Aging, Department of Biomedical Sciences, Faculty of Health and Medical
12. Legislative Yuan of ROC, Legislative Yuan Bulletin 83, 111 Sciences, University of Copenhagen, Copenhagen, Denmark.
(2022) [in Chinese].
10.1126/science.adf5850
10.1126/science.adf3523

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 7 19

1118Letters_16183348.indd 719 11/14/22 4:41 PM

720

1118SpecialIssueIntro_16151220.indd 720
CREDIT: (PHOTO) MIT MUSEUM/CC BY-NC; (PAGE ELEMENTS) ADAPTED FROM KYOSHINO/[Link] AND SAMGRANDY/[Link]

11/11/22 4:32 PM
SPECIAL SECTION

PERSPECTIVES
Moore’s law: The journey ahead p. 722

Toward gallium oxide power

electronics p. 724

REVIEWS
Carbon nanotube transistors: Making
electronics from molecules p. 726
Toward attojoule switching energy in
logic transistors p. 733

1118SpecialIssueIntro_16151220.indd 721 11/11/22 4:32 PM

SPECIAL SEC TION T RA N S I S T O R S

DEVICE TECHNOLOGY

Moore’s law: The journey ahead

High-performance electronics will focus on increasing the rate of computation

By Mark S. Lundstrom and off current ratio to allow practical opera- The number of transistors on a chip
Muhammad A. Alam tion and suppress leakage current to reduce is still increasing, but the rate of scaling
wasted power. In 2003, strained silicon has slowed because smaller transistors

T
he transistor was invented 75 years was introduced as channel material, and do not function very well. Specifically, the
ago, and the integrated circuit (IC) it increased the on-current by increasing length of the channel (the region between
soon thereafter. The progress in the velocity of electrons (3), and in 2004, the source and drain electrode where the
making transistors smaller also led gate insulators with a high dielectric con- gate acts as a switch) is now ~10 nm. At
to them becoming cheaper, which stant decreased the off-state gate-leakage shorter channel lengths, excessive quan-
was famously noted as Moore’s law current. In 2011, the FinFET, a nonplanar tum-mechanical tunneling degrades tran-
(1). Today’s sophisticated processor chips transistor structure that increases the elec- sistor action. Key performance metrics,
contain more than 100 billion transistors, trostatic control of the energy barrier by such as on-current (which should be high
but the pace of downsizing (“scaling”) the gate electrode (and thereby improves for high-speed operation), off-current
has slowed and it is no longer the only or the on-off current ratio), was introduced (which should be low to minimize standby
even main design goal for improv- power), and power supply voltage
ing performance in particular ap- (which should be low to minimize
plications. How can Moore’s law Three platforms forward the power consumed), all degrade
continue on a path forward? New Two-dimensional (2D) nanoelectonics, three-dimensional (3D) terascale simultaneously. Silicon MOSFETS
approaches include three-dimen- integration, and functional integration can all extend Moore’s law, but all are now about as small as they can
sional (3D) integration that will face substantial challenges and fundamental limits. get, and the 2D chips are about as
focus on increasing the rate of in- large as they can be made, so new
formation processing, rather than Platforms Challenges Limits ways to advance performance must
on increasing the density of tran- be found.
2D nanoelectronics • Heat
sistors on a chip. dissipation
Performance is being enhanced
Although Moore’s law predicted Although other design challenges
• Process
Electron by moving from general-purpose,
a rate for the decrease in cost per can be met, smaller transistors, even integration tunneling “commodity chips” to ones that
ones enabled by advanced surround-
transistor, it is popularly viewed gate design, will eventually hit the • Lithography accelerate specific functions. For
in terms of transistor size, which electron tunneling limit. example, hardware acceleration
for two-dimensional (2D) chip ar- offloads specific tasks to spe-
rays translates into an areal size or 3D terascale integration • Process cialized chips such as graphics
“footprint.” During the last 75 years, integration processing units or an application-
Transistor count can increase • 3D design Heat
as the footprint has decreased from through 3D monolithic integration dissipation specific IC. Companies such as Ap-
micrometer to nanometer scales, or stacking of logic, memory, and
• Reliability ple now design their own chips to
issues with implementing new fab- power chips. The approach, however, • Lab-to-Fab meet their specific requirements,
rication technologies have raised faces several design challenges and as will all of the major automobile
heat dissipation limits.
concerns several times about the manufacturers. Computing is the
“end of Moore’s Law.” Twenty years Functional integration • Application- limiting factor for machine learn-
ago, a pessimistic outlook pre- specific design ing, and companies such as Google
vailed regarding the development Integrating intelligent sensing, Unknown now design their own artificial in-
• Developing
actuation, and data analytics sensors and
of several difficult technologies for would improve functional perform- telligence (AI) accelerator chips.
edge analytics
scaling to continue. In this con- ance by sending information Custom chip designs can increase
text, one of the authors (M.S.L.) instead of raw data. performance by orders of magni-
predicted that instead of slowing tude, but just as the cost of chip
down, the scaling of metal-oxide-semicon- in commercial ICs. Gate all-around transis- manufacturing facilities (“fabs”) has mul-
ductor field-effect transistors (MOSFETs) tors that further improve the electrostatic tiplied (from ~$1 billion in 2000 to ~ $20
below the so-called 65-nm node, which was control of the gate are now in development billion for a leading-edge fab), so has the
state-of-the-art in 2003, would continue (4). The size of transistors that can be fabri- cost of leading-edge design. The design of a
unabated for at least a decade before the cated is limited by patterning and etching. leading-edge chip can cost $0.5 billion and
scaling limit was reached (2). Patterning is done by a process known as require a team of ~1000 engineers. Lower-
Scaling indeed continued from about photolithography, in which a photoreactive ing the cost of leading-edge, custom-chip
100 million transistors per chip in 2003 to polymer creates a mask on the chip for etch- design (possibly by using machine learning GRAPHIC: K. HOLOSKI/SCIENCE

as many as 100 billion transistors per chip ing steps. The minimum size of the pattern techniques) will be a key challenge for the
today. One approach was to improve the on- is determined by the wavelength of the light next era of electronics.
used. The recent emergence of extreme ul- Continued progress will also require
traviolet lithography (EUV) made it possi- advances in the underlying technology.
School of Electrical and Computer Engineering, Purdue
University, West Lafayette, IN 47907-1971, USA. ble for Moore’s law to continue beyond the Despite the sharp increase in the number
Email: lundstro@[Link] 7-nm node (5). of transistors on chips (both by decreasing

722 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_Special_Issue_16126706.indd 722 11/11/22 5:00 PM

their size and increasing the 2D chip area), developed. Stacking already-processed 2D experimental loops that maximize learning
until recently one aspect of the design had chips to achieve 3D systems has its own are needed.
remained largely unchanged. Individual set of material and processing challenges, Thermal issues will define the limits of
chips are packaged and combined with such as maintaining interconnect align- 3D terascale integration, just as tunnel-
other chips and other components (such ment over distances of ~1 to 5 mm. Hetero- ing limits have stymied 2D scaling. This
as inductors) laterally on printed circuit geneous integration of components such as requirement need not herald an end of
boards. Sending signals on- and off-chip Si high- and low-voltage logic and memory Moore’s law. The goal of computing is not
increases delays and power consumption. transistors, and compound semiconduc- operations per second but information per
An emerging design theme is to exploit tor-based power and high-frequency tran- second. In that regard, biology offers a
the third (vertical) dimension to enable sistors, presents another set of complex guide. Human senses process information
terascale integration (TSI), with trillions integration challenges. locally before forwarding it to the brain.
of transistors integrated into monolithic or Transistors generate heat when they op- Empowering the sensors at the edge that
stacked chips and terabits per second per erate, and removing the heat is a serious interfaces the analog world, supported by
millimeter communication speed for elec- issue in electronics today (7, 10). Indeed, local memory and data processing (edge
trical or optical interconnections (“per mil- thermal cross-talk among logic, memory, analytics), could prevent the data deluge
limeter” refers to the communication link power transistors, and inductors in a het- from overwhelming the computer.
distance between the chips). For example, a erogeneous IC creates unprecedented de- Electronics is at an inflection point. For
3D NAND flash memory (which is based on sign challenges. New ways to remove heat, 75 years, it has been possible to make tran-
NAND logic gates and retains its states with perhaps mimicking the thermoregulation sistors smaller, but that will not be the driv-
the power off ) can have nearly 200 layers of organisms, and thermally aware design ing force for progress in the decades ahead.
of devices and half a trillion memory tran- will be critical when trillions of transistors If Moore’s law is understood to refer to the
sistors (6). Emerging logic transistors with are placed in close proximity. increasing number of transistors per inte-
new channel materials (such as transition The reliability of electronic systems must grated system (not necessarily per chip),
metal dichalcogenides and indium oxide) be guaranteed for a minimum time, typi- then the end of Moore’s law is not in sight
that can be processed at low temperatures cally 10 years, but decades for some applica- (see the figure). The increasing number of
and embedded within the interconnect tions. Ensuring a failure rate between 1 and transistors will not come by making them
stacks offer further opportunities. 10 parts per million for ICs with 100 billion smaller, but by stacking them vertically or
The third dimension also opens up the transistors each requires predicting the reli- combining them laterally in sophisticated
possibility of vertical heterogeneous inability of quintillion (~1018) transistors. In packages, and eventually in monolithic 3D
tegration of logic, memory, and power practice, reliability is determined through chips and adding functionality.
transistors. With “through-Si vias,” metal short-term accelerated testing of no more Shifting from nanoelectronics (focused
wires that connect vertically from the chip, than a few thousand transistors. Thus, the on reducing the transistor dimension) to
already-processed chips can be stacked to reliability physics of the wear-out and cata- terascale electronics (driven by increas-
put them in close physical proximity to strophic failure modes of these new systems ing transistor count and related function-
minimize signal delays and reduce power need to be understood with unprecedented ality) defines the paradigm shift and core
consumption (7). Vertically stacked logic precision. When so many devices are inter- research challenges of the future. It will
and memory chips also enable new com- connected and placed in close proximity, require fundamental advances in materi-
puting paradigms, such as “compute-in- new phenomena will emerge, and these als, devices, processing, and the design and
memory.” Monolithic 3D ICs would consist must be managed or exploited. manufacture of the most complex systems
of layers of active devices, such as 2D logic Future terascale systems will be funda- humans have ever built. Someday the elec-
transistors, magnetoresistive and resistive mentally different from today’s gigascale trical tunneling and thermal bottleneck will
random-access memories, and ferroelectric systems in that understanding the building define the limits of 3D integration. Until
FETs, along with the metal lines that inter- blocks of a system do not inform how these then, Moore’s law will likely continue as re-
connect them (8). blocks interact and lead to new phenom- searchers address the challenges of these ex-
Recent packaging innovations, such as ena (11). Chip design is already complex traordinarily complex electronic systems. j
silicon-interposer and multi-die silicon and expensive, but algorithms or tools to
REF ERENCES AND NOTES
bridges, inserted between the 3D chips place devices for 3D design and routing the
1. G. E. Moore, Electronics (Basel) no. 8 (19 April 1965)
and the substrate, create denser lateral interconnections among them are not yet (1965).
interconnection and faster communica- available. These design tools must model 2. M. Lundstrom, Science 299, 210 (2003).
tion among the chips. Advanced packaging the complexity of the process and package 3. K. Mistry et al., in IEEE International Electron Devices
brings together logic, memory, power man- integration, thermal cross-talk among 3D Meeting, pp. 247250 (2007).
4. D. Jang et al., IEEE Trans. Electron Dev. 64, 2707 (2017).
agement, communications, and photon- ICs, and operation-specific variability and 5. WIRED, “The $150 million machine keeping Moore’s law
ics through side-by-side integration. The reliability of the packaged system. alive,” 30 August 2021; [Link]
proximity of integration now rivals that in When new materials and processing asml-extreme-ultraviolet-lithography-chips-moores-
stacked and monolithic 3D ICs (8, 9). techniques are developed in research, they law/.
6. A. Goda, Electronics 10, 3156 (2021).
Monolithic 3D integration will require must be translated to large-scale manufac-
7. R. W. Keyes, Proc. IEEE 60, 225 (1972).
that the growth or deposition steps do not turing. Translating advances achieved with 8. S. Iyer, IEEE Trans. Compon. Packaging Manuf. Technol. 6,
affect the already-processed layers. For ex- research-grade equipment to large-scale 973 (2016).
ample, the transistors embedded within manufacturing with different and more so- 9. C. H. Douglas et al., in 2021 IEEE International Electron
Devices Meeting, pp. 3–7.
the interconnect stack must be deposited phisticated state-of-the-art manufacturing
10. M. A. Alam et al., IEEE Trans. Electron Dev. 66, 4556
at a low-enough temperature not to disturb equipment presents a serious “lab to fab” (2019).
the dopant profiles of the Si transistors un- challenge. The research community will 11. P. W. Anderson, Science 177, 393 (1972).
derneath. The needed materials are often need access to advanced processing facili-
incompatible unless special processes are ties, and short “conceive-conduct-analyze” 10.1126/science.ade2191

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 723

1118Perspectives_Special_Issue_16126706.indd 723 11/11/22 5:00 PM

SPECIAL SEC TION T RA N S I S T O R S

DEVICE TECHNOLOGY

Toward gallium oxide power electronics

Ultrawide-bandgap semiconductors show promise for high-power transistors

By Marko J. Tadjer device would also reduce switching losses into their development cycle, UWBG de-
(proportional to capacitance) and provide vice architectures are still actively ex-

E
fficient, ultrahigh-voltage power- a platform for higher-speed electronics plored by researchers. A vertical field-ef-
conversion electronics (with voltages without sacrificing output power. Such fect transistor (FET), such as the FinFET
>20 kV) require semiconductors with high-speed power transistors would be dis- (see the figure, left), can in theory block
an energy gap much larger than that ruptive in the power electronics industry very high fields but is more susceptible to
of silicon. The wide-bandgap (WBG) because system volume is inversely propor- extended defects in the epitaxial layer. A
semiconductor silicon carbide (SiC) tional to frequency. lateral transistor, such as the heterojunc-
has matured into a commercial techno- Out of the six crystalline Ga2O3 phases, tion FET (see the figure, right), could po-
logical platform for power electronics (1), the low-symmetry monoclinic b-Ga2O3 is tentially switch faster and more efficiently
but ultrawide-bandgap (UWBG) (band- furthest along in its development cycle because of its smaller capacitances and
gaps >4.5 eV) semiconductor devices could because of its thermal stability at high shorter transit times, and it could also
potentially enable substantially higher- temperatures (>650°C) (3), and the dis- use ternary alloys of Ga2O3, in this case
voltage electronics. Candidate UWBG cussion below refers to this phase. Unlike b-(AlxGa1-x)2O3, to boost power performance
semiconductors include aluminum nitride any other WBG or UWBG semiconductor, even further.
(AlN), cubic boron nitride, and diamond, melt growth methods originally developed The existence of shallow energy donors
but during the past decade, the greatest infor silicon substrates have been adapted and acceptors (charged impurities) has
crease in research activity has likely been
directed at gallium oxide (Ga2O3). This in-
terest is driven in part by its large bandgap Gallium oxide (b-Ga2O3) devices
of ~4.85 eV and breakthroughs in crystal The availability of wafer substrates and growth processes enables fabrication of devices for power electronics.
growth that led to the first Ga2O3 transis-
tor demonstration in 2012 (2). Ga2O3 has High-field operation Thermal management
promise as a platform for power electron- Schematic of a vertical b-Ga2O3 Schematic of near-junction integration of
ics, but there are challenges in bringing fin field-effect transistor (FinFET) nanocrystalline diamond with b-Ga2O3
with 4.2-kV breakdown voltage. transistor for top-side heat extraction.
this UWBG semiconductor into commer-
cial use within the next decade. Source n+ layer Gate metal Gate oxide
The electrification process that has
captured the attention of many indus- Heat Heat
tries could be disruptively accelerated if Field oxide Gate oxide
ultrahigh-voltage electronics penetrate Gate metal Source
Diamond
Drain
into applications such as next-generation Dielectric
power-grid control and protection, ultra- (Al0.21Ga0.79)2O3
n+ layer n+ layer
fast electric vehicle chargers, or efficient Silicon-doped layer
point-of-load converters with size, weight, Unintentionally doped b-Ga2O3 buffer
and power advantages. Although SiC de-
vices are higher in cost compared with b-Ga2O3 drift layer
conventional silicon power electronics, at n+ b-Ga2O3 substrate
Drain Semi-insulating (010) b-Ga2O3 substrate
the system level, those costs are expected
to be offset by savings given simpler cir-

GRAPHIC: A. FISHER/SCIENCE BASED ON NOVEL CRYSTAL TECHNOLOGY, INC.

cuitry requirements. to commercialize Ga2O3 substrates (4). plagued all UWBG semiconductors be-
Power conversion at very high voltages b-Ga2O3 wafers have reached the 4-inch cause an increasingly wider energy gap
and high switching speeds may become at- (100-mm) commercial milestone and are typically leaves an extrinsic impurity to
tainable beyond 20 kV if a viable UWBG on track to become available at the 6-inch reside farther from the conduction (or va-
technological platform emerges. Even at (150-mm) size by 2027 (5). In parallel, the lence) band. However, for Ga2O3, silicon is
10 kV, it is difficult to increase the switch- infrastructure for high-quality epitaxy is an excellent extrinsic shallow donor that
ing frequency of a power converter beyond being scaled to keep up with the increasing has enabled a very wide range of con-
~10 kHz without sacrificing circuit effi- Ga2O3 substrate size. Methods for Ga2O3 trolled conductivities, from below 1014 cm-3
ciency. UWBG semiconductors inherently epitaxial growth, such as chemical vapor to above 1020 cm-3. The controllable n-type
require thinner device layers that lead to deposition (CVD), molecular beam epitaxy, conductivity extends even to the ternary
reduced conduction losses (proportional and halide vapor phase epitaxy among alloy (AlxGa1-x)2O3, which has an even wider
to channel resistance). The reduced car- others, are being extensively researched bandgap that is also tunable depending
rier transit time through a smaller UWBG with the goal of producing the highest- on phase and Al concentration (6). Also,
quality material. CVD-grown homoepitaxial b-Ga2O3 has a
United States Naval Research Laboratory, Washington, DC Although the basic infrastructure build- purity surpassed only by silicon. Recently,
20375, USA. Email: [Link]@[Link] ing blocks for UWBG technology are well extremely high low-temperature mobility

724 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118Perspectives_Special_Issue_16126706.indd 724 11/11/22 5:00 PM

(23,000 cm2 V-1 s-1) in homoepitaxial CVD in device fabrication processes to pattern portance of this new technology. The recently
Ga2O3 was enabled by ultralow levels of conducting and insulating surface regions. enacted US CHIPS and Science Act will not
background donor compensation by unin- Dry etching—a common processing step for only fund chip fabrication facilities but will
tentional acceptors (2 × 1013 cm-3, ~0.06% creating such patterns, which can introduce also direct $13 billion to the US Department
donor compensation), likely originating surface defects that affect device reliability— of Commerce and the US Department of
from unintentional formation of point de- could potentially be eliminated altogether Defense toward semiconductor and micro-
fects in the lattice (7). if patterning could be implemented solely electronics research and development. These
However, achieving growth of very thick through ion implantation. Unlike any other investments should spur additional funding
(>30 µm) epitaxial b-Ga2O3 at this level of UWBG material, Ga2O3 can even be wet of research into UWBG semiconductors and
purity is extremely challenging, and its de- etched in phosphoric acid and etched using related materials in the coming years, with
velopment is required to be competitive with gas-phase Ga, both of which eliminate chemi- the expectation that a diverse portfolio of
SiC for the ultrahigh-power switching appli- cal and mechanical damage from plasma heterogeneously integrated semiconductor
cations space. The understanding of defects etching that always introduces defects at the modules will overcome the drawbacks of
in Ga2O3 epitaxy will have to advance over etched surface (10, 11). Developing Ga2O3- chips made using one particular semicon-
the next several years before high-voltage specific fabrication processes in parallel with ductor. Furthermore, a higher-frequency
Ga2O3 devices could become commercially high-quality thick epilayers could speed up device will be beneficial at the system level
viable. Point defects, such as vacancies and Ga2O3 device commercialization during the only if the passive components can keep up.
their related complexes (such as vacancy-in- next decade, at least at the scale of a two- Advances in magnetic materials could also
terstitial defects) as well as extended defects terminal device (such as a diode) (5). help prevent components such as inductors
in thick epitaxial layers, currently inhibit The extremely low thermal conductivity and transformers from becoming too lossy at
Ga2O3 device dimensions. In general, defect of Ga2O3 (11 to 27 W m-1 K-1) must be care- higher frequencies.
characterization in Ga2O3 promises to be a fully considered (5). Cooling of Ga2O3 tran- Exploratory research into Ga2O3 as well as
rich research field that will also enable any sistors will be even more critical than that of other UWBG semiconductors will likely con-
Ga2O3 power electronics commercial enter- GaN transistors, which also suffer from self- tinue given that fundamental questions are
prise hoping to break the 20-kV barrier with heating effects (12). Although Ga2O3 devices still being resolved (15). For Ga2O3, the larg-
a useful device size to do so. still output about an order of magnitude less est hurdles, such as scaling epilayer thick-
For power electronics, the development of power during operation compared with their ness without sacrificing material quality and
p-type (hole-carrier) materials is necessary GaN counterparts, the top- and substrate- managing ultralow thermal conductivity
because holes in Ga2O3 form localized polar- side cooling approaches developed for GaN and unipolar conductivity, will require new
ons that lead to a self-trapping phenomenon could be applied to Ga2O3 (12). Indeed, cap- physics to be solved as well as the develop-
that limits their conduction (8). The absence ping a lateral transistor with AlN or nano- ment of new or modified materials. The
of p-type conductivity in Ga2O3 presents a crystalline diamond has allowed for the continued pace of innovative engineering
challenge for high–electric field management achievement of 5– to 6–W mm-1 DC output of Ga2O3 could bring commercialization of
regardless of device geometry, and any practi- power with Ga2O3, which is similar to early GaO3 power diodes within the next decade,
cal solution will require innovation in hetero- results from GaN high–electron mobility but only if substantial investments in this in-
geneous integration not faced for previously transistors in the 1990s. Heterogeneous interesting semiconductor continue. j
developed semiconductors. tegration with high–thermal conductivity,
REF ERENCES AND NOTES
Unlike p-type semiconductors, such as SiC, WBG p-type semiconductors, such as SiC,
1. C. R. Eddy Jr., D. K. Gaskill, Science 324, 1398 (2009).
gallium nitride (GaN), or diamond, WBG p- GaN, and even diamond, is of particular 2. M. Higashiwaki, K. Sasaki, A. Kuramata, T. Masui, S.
type nickel oxide (NiO) can be sputtered at interest for p-n and junction barrier Yamakoshi, Appl. Phys. Lett. 100, 013504 (2012).
room temperature and is thus benign for in- Schottky rectifiers. 3. S. J. Pearton et al., Appl. Phys. Rev. 5, 011301 (2018).
4. Novel Crystal Technology, Inc., Current commercial
tegration with Ga2O3 devices. Recent studies, Looking to early commercialization efforts β-Ga2O3 wafer suppliers; [Link]/.
such as the 8-kV NiO/Ga2O3 p-n diode dem- of WBG semiconductors, the success of SiC 5. A. J. Green et al., APL Mater. 10, 029201 (2022).
onstration by Zhang et al. (9), have shown was driven in part by substantial government 6. J. A. Spencer et al., Appl. Phys. Rev. 9, 011315 (2022).
7. G. Seryogin et al., Appl. Phys. Lett. 117, 262101 (2020).
that the lack of p-type conductivity in Ga2O3 investment and scientific research efforts 8. J. B. Varley, A. Janotti, C. Franchini, C. G. Van de Walle,
could potentially be managed by integrat- that continue its innovation. Solving the is- Phys. Rev. B 85, 081109 (2012).
ing heterojunctions with the purpose of field sues of micropipe and basal plane dislocation 9. J. Zhang et al., Nat. Commun. 13, 3900 (2022).
10. H.-C. Huang, Z. Ren, C. Chan, X. Li, J. Mater. Res. 36, 4756
management and charge balance in these de- defects in SiC relied on advanced character- (2021).
vices. The prospects of Ga2O3 as a power elec- ization techniques, such as ultraviolet pho- 11. N. K. Kalarickal et al., Appl. Phys. Lett. 119, 123503
tronics material would be greatly enhanced if toluminescence imaging and spectroscopy. (2021).
12. M. J. Tadjer, T. J. Anderson, Eds., Thermal Management of
robust heterogeneous integration with p-type Material scientists have continued to develop
Gallium Nitride Electronics (Elsevier, 2022).
WBG semiconductors (such as GaN or AlN) their understanding of defects in the ever- 13. N. A. Mahadik, R. E. Stahlbush, W. Sung, J. Appl. Phys.
are developed. Such development could lead larger-diameter SiC wafers (13). 131, 225702 (2022).
to a reliable junction barrier Schottky recti- Similar efforts will be needed to under- 14. M. D. McCluskey, J. Appl. Phys. 127, 101101 (2020).
15. J. Y. Tsao et al., Adv. Electron. Mater. 4, 1600501 (2018).
fier being commercialized, as was the case stand and control both point and extended
for SiC. defects in thick (>30 µm) Ga2O3 epitaxial lay- ACKNOWL EDGMENTS
Effective electric field termination at sur- ers (14). Government funding would be cru- The views expressed in this article do not necessarily rep-
faces is a key requirement for the use of cial in the early support of such efforts. The resent the views of the US Department of Defense or the
United States. Research at the Naval Research Laboratory
UWBG materials in practical high-voltage small-business technology transfer program
was supported by the Office of Naval Research. K. Hobart, C.
electronic devices. The nitrogen deep ac- initiated by the US Office of Naval Research Eddy Jr., R. Stahlbush, A. Jacobs, H. Masten, J. S. Lundh, and
ceptor has been effective in rendering Ga2O3 in 2017 with the purpose of initiating the de- T. Anderson; epitaxial growth by Novel Crystal Technology
nearly insulating and creating an effective di- velopment of b-Ga2O3 CVD has resulted in (Sayama, Japan); and Agnitron Technology (Chanhassen,
MN) are gratefully acknowledged.
electric layer that can reduce electric fields. the commercialization of this capability be-
Selective ion implantation could be useful fore the program’s end and highlights the im- 10.1126/science.add2713

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 725

1118Perspectives_Special_Issue_16126706.indd 725 11/11/22 5:00 PM

T R A N S I S T O RS

REVIEW Advances in materials for CNT transistors

Carbon nanotube transistors:

Exploiting the advantages of semiconducting
CNTs requires overcoming several materials

Making electronics from molecules

science hurdles. Just as silicon must be pu-
rified and doped to be a useful channel mate-
rial, as-synthesized CNTs can be either metallic
Aaron D. Franklin1,2*, Mark C. Hersam3,4,5, H.-S. Philip Wong6,7 or semiconducting and must be purified into
semiconducting-only for use in transistors.
Semiconducting carbon nanotubes are robust molecules with nanometer-scale diameters that can be Whether CNTs are metallic or semiconducting
used in field-effect transistors, from larger thin-film implementation to devices that work in depends on how the hexagonal lattice wraps
conjunction with silicon electronics, and can potentially be used as a platform for high-performance into a tube. This structure is most easily vis-
digital electronics as well as radio-frequency and sensing applications. Recent progress in the ualized by rolling a rectangular section of the
materials, devices, and technologies related to carbon nanotube transistors is briefly reviewed. sp2-bonded hexagonal carbon lattice of atom-
Emphasis is placed on the most broadly impactful advancements that have evolved from ically thin graphene into a 1D cylinder, where
single-nanotube devices to implementations with aligned nanotubes and even nanotube thin the resulting diameter is ~1 nm and the length
films. There are obstacles that remain to be addressed, including material synthesis and processing is 102 to 108 nm. The vector that defines the
control, device structure design and transport considerations, and further integration demonstrations width of the rectangular section with respect
with improved reproducibility and reliability; however, the integration of more than 10,000 devices to the graphene lattice is commonly referred
in single functional chips has already been realized. to as the chiral vector and ultimately deter-
mines the diameter, helicity, and conductiv-

T
ity properties of the CNT (5).
ransistors are electronic switching devices thus have limited utility as transistor channels In addition to specifying the physical struc-
that enable digital computation based (5). Throughout this review, CNTs will imply ture of the CNT, the chiral vector also imposes
on their on-state (binary 1) and off-state single-walled nanotubes. Semiconducting well-defined quantum-mechanical boundary
(binary 0) operation. In the earliest days CNTs are composed of a cylindrical shell of conditions on the electronic band structure,
of integrated circuits, it became clear hexagonally arranged carbon with a diameter which implies that for random tube closure,
that scaling down the size of the transistors of ~1 nm. Electrons travel only forward or back- ~33% of CNT chiralities are metallic and ~67%
would drive better chip-level performance, ward with a wave function wrapping around are semiconducting. Moreover, among the
which is now known as Moore’s law. One of the the nanotube to create a one-dimensional (1D) semiconducting chiralities, the bandgap is
most important dimensions for such scaling semiconductor with an energy bandgap of a approximately inversely proportional to the
is the semiconducting channel length, which few hundred milli–electron volts (5). These ma- CNT diameter. Because CNT transistors re-
is the distance that electrical current flows terials are stable in air and can be manipulated quire semiconducting channels, preferably
or is controlled by a gate electric field to turn the through a variety of processing methods that with a well-defined and uniform bandgap, the
device on and off. Although the initial channel are commonly used in the semiconductor inability to scalably synthesize and isolate CNTs
lengths were many microns in size, proposals dustry. The early demonstrations of field-effect with atomically precise chiral vector control is
to scale the semiconducting channel to the transistors (FETs) by draping a semiconducting the ultimate goal for high-performance CNT
ultimate limit of molecular dimensions (frac- CNT over metal electrodes (3, 4) have led to integrated circuits.
tions of a nanometer) date to the mid-1970s (1). continued research activity with the goal of
Decades of study on the transfer of electrons creating reproducible, scalable, and high- Controlled synthesis of CNTs
through conjugated organic molecules, con- performance devices integrated into dense cir- CNTs can be synthesized by introducing a
sidered to replace the silicon channel, high- cuitry using processing steps similar to those carbonaceous feedstock with a metal catalyst
lighted several important challenges for such used to create silicon electronics. (usually Fe or Ni) into a growth chamber,
molecular transistors. The foremost issues in- The widespread interest in semiconducting where energy is added through heat, light, or
cluded low stability and the difficulty of effec- CNTs has also inspired intense and ongoing plasma excitation. Because CNT growth typ-
tively gating, and making reliable electrical exploration of other nanomaterials, including ically occurs at temperatures at which these
contact to, the molecules (2). semiconducting nanowires (6), 2D graphene catalysts undergo substantial restructuring,
To meet or exceed the performance of silicon (7), transition metal dichalcogenides (8), and it is difficult to control the chiral vector, and a
electronics, it became clear that new channel Xenes (9). Despite the growing number of range of CNT diameters and both electronic
materials must have similar stability. Among nanomaterial options, CNTs stand out in offer- types are produced; much effort has been ex-
the molecular options, semiconducting single- ing stability, bandgap, and superb electrical pended to gain control over CNT chirality (10).
walled carbon nanotubes (CNTs) have several and thermal properties that are unmatched These approaches include the use of refrac-
advantages (3, 4). Nested multiwalled CNTs are by other candidates. Here, we review recent tory catalyst particles such as W-Co alloys with
effectively metallic at room temperature and material, device, and technology advances for well-defined size and shape that remain struc-
CNT transistors, establishing both the sub- turally invariant at the growth temperature
stantial promise and the remaining challenges and thus can drive predictable nucleation of
1
Department of Electrical and Computer Engineering, Duke
for this molecular transistor. Progress in the targeted CNT chiralities (Fig. 2A) (11), the
University, Durham, NC, USA. 2Department of Chemistry, field will be related to the foremost potential addition of molecular seeds that have a struc-
Duke University, Durham, NC, USA. 3Department of Materials applications for CNT transistors, highlighted ture that closely matches the targeted CNT
Science and Engineering, Northwestern University, Evanston,
in Fig. 1. Two of the most prominent potential chirality (12), or the deployment of CNTs them-
IL, USA. 4Department of Chemistry, Northwestern University,
Evanston, IL, USA. 5Department of Electrical and Computer applications are high-performance (HP) com- selves as seeds in “CNT cloning” (13). Although
Engineering, Northwestern University, Evanston, IL, USA. puting chips and thin-film transistors (TFTs) for tailored catalysts or seeds help control synthetic
6
Department of Electrical Engineering, Stanford University, display backplanes and the internet of things outcomes, many other growth parameters also
Stanford, CA, USA. 7Stanford SystemX Alliance, Stanford
University, Stanford, CA, USA. (IoT); a few of the target performance metrics play a role—including temperature, pressure,
*Corresponding author. Email: [Link]@[Link] for these applications are summarized in Table 1. flow rates, and applied electric fields (14)—and

726 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

thus growth optimization entails the search of
L ch < 100 nm L ch < 10 nm
a broad parameter space. In an effort to accel- Source Gate
erate this exploration, autonomous growth 3D Drain
using closed-loop iterative experimentation Memory
is showing promise for rapid identification
of synthesis conditions that minimize CNT L ch > 10 µm 3D logic
structural polydispersity (15). Scaled, HP FETs
CMOS
Separation of semiconducting CNTs
Because the most-optimized CNT growth pro- 3D integration

Performance
cedures still lack sufficient monodispersity
for wafer-scale transistor applications, post-
synthetic separation methods are required
to sort as-grown CNTs by diameter, chirality, Low-voltage VLSI
Thin-film devices
and electronic type. Fortunately, CNTs have RF transistors for telecom
sizes and shapes that are comparable to those
of biological macromolecules, which has al- Heterogeneous 3D chips
lowed many CNT separation methods to be Chemical and biomedical sensors
adapted from ones that have already been
developed for biochemistry. In density gra- Paper-based printed electronics
dient ultracentrifugation (DGU), CNTs are
first dispersed and encapsulated with mix-
Cost and complexity
tures of surfactants that show selectivity for
different CNT separation targets (including Fig. 1. Broad range of potential applications for CNT transistors. Illustration of device performance
chiral vector, chiral handedness, electronic versus cost and complexity for some of the foremost potential applications of CNT transistors. Applications
type, and diameter) and then separated by range from microscale thin-film devices (e.g., printed electronics, biosensors) to three-dimensionally
buoyant density in aqueous density gradients integrated BEOL devices (such as heterogeneous 3D layers integrated onto silicon CMOS) and scaled
(16). Although DGU has sufficient scalability high-performance (HP) FETs [such as low-voltage very-large-scale integration (VLSI)], with increasing
to be viable commercially, other strategies performance corresponding with increased cost and complexity of integration. Lch, channel length.
from biochemistry have also been heavily
developed, including gel chromatography
(17) and dielectrophoresis. The latter method Table 1. A few of the target metrics for two prominent CNT transistor applications. Values are
has the added benefit of enabling aligned as- approximations based on achieving optimal performance. Notably, although some of these targets
sembly of CNTs between prepatterned elec- have been achieved, one of the foremost challenges is to achieve them simultaneously (e.g., high
trodes (18). on-state current with low subthreshold swing, which is a measure of how much gate voltage is
Methods from polymer chemistry have required to modulate the current by one decade). High-performance FETs are used in applications
also been used for CNT separations, includ- such as central processing units (CPUs) for servers, and TFTs are thin-film transistors such as those
ing aqueous two-phase extraction (19) and used in the backplane electronics of displays.
selective dispersion of targeted CNT chiralities
with structure-discriminating polymers that Target for
Metric Target for TFTs
wrap the nanotubes (Fig. 2B) (20). In all cases, high-performance FETs
purities of semiconducting CNTs have reached CNT semiconducting purity >99.9999% >99.9%
.....................................................................................................................................................................................................................
the detectable limits of optical spectroscopic Parallel CNTs, Thin-film CNTs in an
characterization (~99.9%) and begun to pro- CNT array alignment
consistent pitch unaligned network
.....................................................................................................................................................................................................................
vide sufficient monodispersity for many CNT >200 CNTs per micrometer >50 CNTs per square
transistor applications. The ultimate goal for CNT array density
(linear density) micrometer (aerial density)
.....................................................................................................................................................................................................................
high-performance digital transistors is to Channel length <12 nm >10 μm
.....................................................................................................................................................................................................................
achieve >99.9999% pure semiconducting CNTs Contact length <10 nm >1 μm
.....................................................................................................................................................................................................................
(see Table 1)—the higher the purity, the bet- On-state current > 0.5 mA μm−1 at 0.6 V >100 μA mm−1 at 3 V
.....................................................................................................................................................................................................................
ter the corresponding performance. In addi- Contact resistance <50 ohm·μm per side <20 kilohm·μm per side
.....................................................................................................................................................................................................................
tion, any molecular wrapper (e.g., surfactant or <200 mV per decade
polymer) should ideally be completely removed Subthreshold swing <70 mV per decade
(application dependent)
.....................................................................................................................................................................................................................
after deposition of the CNTs because this pre-
sents an unwanted residue that can hamper
electrical contact, gating efficiency, and trans-
port in the CNT transistor. transistors (21). However, the requirement based on metal work function, such as Pd for
for complementary p-type and n-type transis- p-type injection and Sc for n-type injection,
Other material considerations tors in digital circuits implies that controlled also enable complementary CNT transistors
A transistor also requires electrical contacts, n-type injection and/or doping is required. (24, 25). Beyond the metal selection, interfacial
doping, and dielectrics. Because contacts from Electron-donating adsorbates, such as organo- material considerations and overall contact struc-
commonly used metals (e.g., Au, Pd) tend to rhodium compounds (22), coupled with atomic ture also play a role (see Fig. 2D for an exemplary
yield Fermi-level alignment near the valence layer–deposited encapsulation layers (23), en- end-bonded contact structure using Mo). Ex-
band of CNTs, p-type behavior from the in- able the fabrication of highly stable n-type CNT tension regions of a metal-oxide-semiconductor
jection of holes is readily achieved for CNT transistors (Fig. 2C). Charge-selective contacts FET (MOSFET), which are between the source

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 727

T R A N S I S T O RS

A B Initial SFM 1st

Templated growth
3 Recycling
tion
struc

Absorbance (cm–1)
Ethanol CVD 1st
Recon
, ( H2
2) 2nd E11
(1) O 2
2 3rd

Nanocrystal
1 E22

Molecular cluster
Interface 0
500 1000 1500
W Co C (n,m) SWNT Wavelength (nm)

C D Mo
Ni
SiO2

Al2O3 SiNx
Au

SiO2 Lch
Lc
Nanotube
Si
Silicon

Fig. 2. Examples of materials for high-performance CNT transistors, on the right. E11 and E22, absorption peaks; SFM, shear force mixing. [Adapted
including synthesized CNTs, purified CNT mixtures, doping strategies, and from (20) with permission from Elsevier]. (C) Electron-donating organorhodium
contact metals. (A) Templated CNT growth of targeted chiralities using compounds encapsulated with atomic layerÐdeposited alumina enabling stable
refractory W-Co nanocrystal catalysts. CVD, chemical vapor deposition; SWNT, n-type CNT transistors. Black is the CNT layer, orange is the dopant layer,
single-walled nanotube. [Adapted by permission from Springer Nature Customer and red is a seeding layer for dielectric growth. [Adapted with permission
Service Center GmbH, Springer Nature (11), copyright (2014)]. (B) Selective from (22). Copyright 2016 American Chemical Society]. (D) When reacted to
polymer dispersion enabling scalable isolation of targeted CNT chiralities form end-bonded carbides, molybdenum contacts to CNT transistors can be
from as-grown polydisperse mixtures as verified by absorption spectrum is scaled down to sub-10-nm dimensions in contact length (Lc) while retaining
shown on the left; a photo of the resultant bottle of sorted CNTs is shown efficient charge injection. [Adapted with permission from (54)].

or drain and the gated semiconducting chan- ity (29). Although devices with an individual Recent progress is encouraging, including a
nel, require stable doping with well-controlled nanotube channel are still of interest for sensing small-scale demonstration with a controlled CNT
doping levels that are optimized for the trade- applications, they are no longer considered suit- pitch of ~10 nm using DNA-directed assembly
off between series resistance and parasitic able for digital or radio-frequency (RF) electron- (34). There are also wafer-scale, high-throughput
capacitances (26)—a feat yet to be reliably ac- ics based on the need for higher current flow strategies that use various forms of solution-
complished for CNT transistors. For the gate than a single CNT can deliver. Although the phase assembly (also referred to as dimension-
dielectric layer, specific materials such as current-carrying capacity for CNTs is astonish- limited self-alignment or liquid crystalline
Y2O3 have exhibited nearly ideal properties ing [~109 A cm−2 (30)], they are only ~1 nm in interfacial assembly), which achieved a ~20-nm
with a high dielectric constant κ and conformal diameter, which yields only ~10 mA per CNT. pitch (Fig. 3, C and D) in one report (35) and a
dielectric coating on CNTs after oxidation of Hence, recent work has predominantly focused 5- to 10-nm pitch in another (36). The primary
deposited yttrium (27). A more conventional on having multiple CNTs in the channel. differences in the two studies were the polymer
approach that uses atomic layer deposition of used to wrap the CNTs and the solution-phase
Al3O3 and HfO2 bilayer dielectrics has enabled Aligned arrays of CNTs technique of depositing the CNTs into arrays
transistors that have a 10-nm gate length with Ideally, the CNTs in a transistor channel would on the substrate. Nevertheless, these ap-
a gate leakage current commensurate with be perfectly aligned in a parallel array with a proaches still require further work to remove
state-of-the-art Si transistors (28). Upon inte- controlled pitch of ~2 to 5 nm (31), similar to unwanted residue from the solution-phase pro-
gration of all these optimized materials, CNT how fins of silicon are arranged in modern cessing along with more consistent, controlled
transistors have been shown to exceed the transistor technologies (FinFETs). Realizing alignment (without bundling) in all directions
performance of incumbent silicon integrated such arrays continues to be a challenge. If the with uniform spacing.
circuit technology, as will be discussed in the CNTs are too close (or bundled), it can create
subsequent sections. cross-talk (electric field screening) and effective Thin films of CNTs
gating issues (32). If the CNTs are too far apart, The difficulty of achieving aligned arrays with
CNT transistor design current density (current per transistor width) controlled pitch has led some researchers to
The initial focus for CNT transistor research will be insufficient. For digital systems with a use unaligned CNT networks or thin films
was on the use of a single CNT as the channel high density of CNT transistors, variations in the (Fig. 3, E to H). Although these unaligned films
(see Fig. 3, A and B) and demonstration of bal- pitch between CNTs also deleteriously affects are less favorable for carrier transport, as well
listic transport (21) and digital circuit operabil- the overall energy, delay, and noise margin (33). as for contacting and gating the nanotubes,

728 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

unaligned CNT networks have achieved high thin-film channel are not long enough to trans- ness and permittivity of both the gate dielectric
performance in nanoscale transistors (37, 38). verse the channel and instead operate as perco- and the semiconducting channel. A generally
Moreover, CNT thin films can be deposited by lating networks in which electrons travel from accepted approximation is that a channel length
using printing techniques, including roll-to- CNT to CNT in transit from source to drain greater than 3λ will ensure that deleterious
roll (39) and direct-write (40, 41) approaches (Fig. 3G) (44, 45). Compared with long-studied short-channel effects are avoided.
(Fig. 3H), which makes them attractive for organic semiconductor TFTs (46), CNT-TFTs Given their intrinsically small size, CNTs
TFTs. The application space for these larger have considerably higher mobilities (10 to offer advantages for aggressively scaled de-
(approximately tens of micrometers) TFTs is 100 cm2 V−1 s−1) and stability under bias, in air, vices. Although it is ideal for an FET to have
distinct from high-performance nanoscale or both. a gate-all-around geometry to minimize λ,
FETs and includes sensors, flexible electron- and demonstrations of such gate structures
ics, IoT, and display backplanes (42). For TFT Advanced gating structures for CNTs have been reported (48, 49), studies
applications, CNT thin films compete well In addition to the density and arrangement of have shown that channel lengths that are
against incumbent semiconductor options such nanotubes in the channel, the gate configura- much less than 10 nm (as short as 5 nm) can
as organics and polymers, metal oxides, and tion in a CNT transistor has advanced in many be achieved in either bottom-gate (50, 51)
low-temperature polysilicon (LTPS) (43). ways. For nanoscale FETs, the primary goal is or top-gate (51, 52) geometries. Although gate
When CNT thin films are used in FETs with to maximize the gate control of the CNT energy geometry does vary for TFTs, it is less critical
nanoscale channel lengths (<100 nm), most of bands in the channel, which is achieved through and mostly limited by the gate dielectric mate-
the nanotubes bridge the entire channel, even if strong gate coupling that is typically expressed rial and the application needs.
they are not perfectly aligned (Fig. 3F). In the as a small scale length, λ (47). The scale length
microscale lengths of TFTs, nanotubes in the depends on the gate geometry and on the thick- Source-drain contact structures
For highly scaled CNT transistors with small
A Single CNT B footprints, not only does the channel length
Pt SIO2 Pt SIO2 Pt need to be at the nanometer scale but the
source and drain contacts also need to have
minimal dimensions while still providing
efficient ohmic charge injection. Palladium
contacts have achieved the quantum limit of
C D 6.5 kilohm per CNT at a 10-nm contact length
CNT array
for a p-type side contact, where the metal rests
on top of a CNT without any chemical bond-
Source ing (53), though this needs to be realized with
higher yield and reproducibility. Alternatively,
Drain an edge-contact structure would offer ideal
scalability and has been demonstrated by re-
E CNT network F Pad CNT film
acting Mo with CNTs to yield a carbide end-
500 nm Wch
V 1 = 4.987 µm bonded contact with sub-10-nm contact lengths
Source H 1 = 208.4 nm (Fig. 2D) (54). Regardless of the geometry,
Gate

Drain L ch contacts to CNTs are a leading factor in deter-

1 µm
mining overall performance, and the combi-
nation of material, structure, and processing
must be further refined to yield contacts for
H
G CNT thin film both p- and n-type carrier injection with high
consistency and low resistance.

Technology demonstrations
5 mm High-performance, energy-efficient digital logic
Although many applications can benefit from
Fig. 3. Variations in CNT transistor structures. (A) Illustration of a single CNT channel with metallic source the properties of CNTs, digital logic applica-
and drain contacts. (B) Atomic force microscopy image of the first-reported CNT transistor, which had a single tions have received the greatest attention
nanotube. [Reprinted by permission from Springer Nature Customer Service Center GmbH, Springer Nature (Fig. 4) because they have the potential to sur-
(3), copyright (1998)]. (C and D) Illustration of an aligned array of CNTs as the channel (C) and a corresponding pass incumbent Si technology in performance
recent example of a transistor with such an array (D), including a scanning electron microscopy (SEM) image and energy efficiency. Such exemplary high-
of the aligned CNTs (left) and schematic of the solution-phase assembly process (right). [(D) is reprinted with performance devices from aligned arrays of
permission from (35); Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC). CNTs can achieve high on-state currents at rel-
[Link] (E and F) Illustration of a CNT network (not aligned) atively low voltages (Fig. 4, A to C). As shown
used as channel for nanoscale FET (E) with a corresponding recent example of a high-performance transistor (F), in Fig. 4D, gate-all-around CNT transistors
including a SEM image of the high-density film (left) and a top-view schematic of the device structure (right). with doped extensions and multiple layers of
Note that most nanotubes directly bridge the source and drain in this nanoscale configuration. H1 and V1, measurement high-density CNTs are projected to show up to
markers of CNT film area height and channel length, respectively; Wch, width of CNT channel region. [(F) is adapted seven times the energy-delay product (EDP)
by permission from Springer Nature Customer Service Center GmbH, Springer Nature (38), copyright (2018)]. benefits compared with Si nanosheets at the
(G and H) Illustration of a CNT thin film used in a thin-film transistor (dimensions of tens of micrometers) (G) with a 2-nm technology node (the EDP, or switching
corresponding recent example of an aerosol-jet printed CNT-TFT on a flexible plastic substrate (H), including a energy, is the product of the time and the
SEM image of the printed thin film (left) and a picture of the printed CNT-TFTs with a schematic of the printing power consumption for an on-off cycle, and
technique (right). [(H) is reprinted with permission from (41). Copyright 2019 American Chemical Society]. a measure of energy efficiency) (26). As noted

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 729

T R A N S I S T O RS

A 10–3 B 1.4
Vgs from –2.5 V to 1 V C Source D 3
Gate Drain
1.2 (0.5-V step)

Total energy (aJ/cycle)

10–4 Si nanosheet
10–5 1.0

Ids (A/µm)

Ids (A/µm)
10–6 190 0.8
mV/dec 7 x EDP benefits
10–7 0.6
Vds =–0.1 V
Si nanosheet Source Gate
10–8 Vds =–0.3 V 0.4 Drain
Vds =–0.5 V
10–9 Vds =–0.8 V 0.2 CNT aligned array
Vds =–1 V Preferred corner
10–10 0.0 0
–2 –1 0 1 2 –1.2 –0.8 –0.4 0.0 0 600
Vgs (V) Vds (V) CNT aligned array Frequency (GHz)

Fig. 4. High-performance CNT transistors for digital logic applications. schematics for a Si nanosheet transistor with two stacked channels
(A and B) Subthreshold (A) and output characteristics (B) of CNT transistors and a CNT aligned array transistor. (D) Projected energy versus frequency
fabricated with aligned arrays of ~150 CNTs per micrometer that achieve pareto curves for Si nanosheet and CNT transistors at the 2-nm technology
an on-state current of >1 mA μm−1. Ids, drain current; Vds, drain-source voltage; node for an inverter ring oscillator. [© 2021. Adapted, with permission,
Vgs, gate-source voltage. [Adapted with permission from (36)]. (C) Device from (26)].

earlier, because of their ultrathin body (~1 nm), resistance of 6.5 kilohm per CNT (53), sub- RF electronics
CNT transistors offer excellent electrostatic 10-nm gate length (52), multiple stacked CNT Although digital electronics remains the dom-
control even at aggressively scaled gate lengths, channel layers (61), and doped source or drain inant focus in the field, CNT transistors also
limited only by direct source-to-drain tunnel- extensions (62). This MOSFET-like CNT struc- hold great promise for high-frequency RF
ing. Parasitic capacitance, a key detractor of ture with 35-nm contacted gate pitch and 20-nm transistors, which are relevant to telecommu-
speed and energy efficiency, accounts for >70% active width is projected to have a performance nications applications (68, 69). Many of the
of the total capacitance of modern Si tran- that far exceeds that of Si transistors for a 2-nm material and device needs for digital CNT
sistors. Because of the ultrathin body, CNT node logic technology. transistors also apply to RF electronics, with
transistors have substantially lower parasitic some relaxing of the semiconducting purity
gate-to-source or gate-to-drain capacitance. These 3D integration needs and an enhanced need for high trans-
two key attributes of CNTs, along with the Future semiconductor chips will go beyond 2D conductance and linearity, which translates
high transport and injection velocities, are the device miniaturization and instead will have to low distortion when amplifying a signal.
physical basis for high-performance, energy- 3D layers of active devices (63). Because logic Recent progress on RF CNT transistors from
efficient digital logic. device layers in 3D must be thin and fabricated aligned arrays of nanotubes shows the abil-
As noted earlier, many fundamental build- at temperatures that are compatible with back- ity to operate at frequencies up to hundreds
ing blocks of a CNT transistor technology have end-of-line (BEOL) wiring layers (typically of gigahertz with attractively low power con-
already been demonstrated. At the circuit or <400°C), CNT transistors are particularly well- sumption and high versatility for integration
system level, a fully functional static random- suited for 3D integration because of the low in system-on-chip applications (70).
access memory (SRAM) array (55), a monolithic device-fabrication temperature and thin device
3D imager (56), and a 16-bit RISC-V (where layer. Starting from the first demonstration of Printed electronics
RISC is reduced instruction set computer) an all-CNT transistor computer almost a decade The ability to purify a solution-phase dispersion
processor with >14,000 transistors (Fig. 5B) ago (64), progress has occurred not only in the of semiconducting CNTs also enables printing
(57) have been fabricated entirely from CNT level of integration but also in the variety of into thin-film devices (Fig. 2H). Many reports
transistors. What’s more, wafer-scale fabrica- devices as well as maturation of the technology have shown that fully printed CNT-TFTs can
tion of CNT transistors has been demonstrated from university laboratories to industry. be used in digital logic circuitry to illustrate the
in an industrial foundry using 200-mm wafer A four-layer monolithically integrated chip ability of these devices to deliver computational
processing techniques (Fig. 5A) (58). The fab- comprising a silicon transistor layer, a CNT functionality (71–73). However, given the low
rication and design of CNT transistors with transistor memory read-out circuit layer, a cost of legacy-node silicon transistor technol-
the same tools and infrastructure as commer- resistive switching metal-oxide random-access ogies, the likelihood that printed CNT-TFT
cial semiconductor technologies helps lower memory (RRAM) layer, and a CNT transistor circuitry will be of widespread use is low. More
the barrier for the introduction of CNT devices sensor layer on the top illustrates the benefits encouraging is the use of printed CNT-TFTs
into mass production. of monolithic integration (Fig. 5, C to E) (65). for the backplane control of displays (74) or
At the individual device level, recent work This 3D chip can process information from the for custom biosensing systems (75). Recent
shows short gate length (10 nm), complemen- sensors to the memory cells to the transistors studies also reveal the recyclability of CNT
tary p- and n-channel devices with near ideal in parallel at rates of terabytes per second. An- thin films (76), which shows promise for en-
subthreshold swing for single-CNT transis- other example is an end-to-end brain-inspired abling a fully printed, paper-based electronic
tors (59), and high on-state current per width hyperdimensional computing nanosystem that system in which all core materials are able to
for aligned CNTs with a density of 50 CNTs is effective for cognitive tasks such as language be recaptured and reused (77).
per micrometer (60). In the near future, it recognition, which was realized with monolithic
will be possible to integrate the following ele- 3D integration of CNT transistors and RRAM, Future developments and perspectives
ments (already shown separately) in a single enabling fine-grained and dense vertical con- Materials outlook
device demonstration: gate-all-around geom- nections between computation and storage Advances in materials are anticipated to be
etry (48, 49), >250 CNTs per micrometer in layers using BEOL interlayer vias (66). The CNT central to future advances in CNT transistors.
highly aligned arrays (36), 3-nm oxide dielectric transistor fabrication process not only has been Improving the purity of semiconducting CNTs
(target oxide capacitance = 2.94 × 10−10 F m−1) shown on full 200-mm wafers (58) but also is critical for all device use cases. In this regard,
(28), sub-10-nm p-type contacts with a contact has 3D integration with RRAM (67). one of the largest impediments to minimizing

730 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

metallic CNT impurities down to concentra- Ultimately, growth conditions encompass mentary metal-oxide-semiconductor (CMOS)
tions of parts per million or billion is the lack of such a vast parameter space that methods for fabs. Most metal-contact formation processes
high-throughput analytical methods for detect- efficiently searching for and identifying opti- rely on liftoff, which is not considered a scala-
ing ultralow concentrations of metallic CNTs. mal growth conditions are needed. Emerging ble process, and the liftoff-free alternatives also
Most high-throughput optical detection meth- artificial intelligence and machine-learning tend to rely on slow patterning processes (78).
ods for CNTs (such as photoluminescence optimization approaches coupled with high- The scalability of the contact length, which
spectroscopy) are less sensitive, if not com- throughput experimental screening hold prom- is an equally important parameter as the gate
pletely insensitive, to metallic species. Indeed, ise for next-generation synthetic efforts (15). length for overall transistor scaling, needs fur-
the only established method for quantifying Similarly, the discovery, optimization, and in- ther consideration. Some studies show severe
ultralow concentrations of metallic CNTs is tegration of the many other materials in a CNT degradation at sub-30-nm contact lengths (52),
to fabricate massive arrays of individual CNT transistor (including dopants, contacts, gate whereas others have shown less degradation at
transistors and then electrically probe them electrodes, and dielectrics) can also likely be scaled lengths but have not yet realized them at
one by one in search of short circuits. This accelerated by machine learning coupled with high yield (53). Such contact-length scaling chal-
approach is extremely time consuming and high-throughput experimental screening. lenges are common to all transistors (79), but
only gets worse as the semiconducting purity discovering a solution that allows for aggressively
increases. Thus, most CNT separation meth- Device outlook scaled contacts without degrading the device
ods have only been optimized to the detection Although much has been learned about es- would be a critical advance. End-bonded or edge
limits of optical spectroscopy (~99.9%). tablishing interfaces to CNTs, including gate contacts present one such possibility (54), al-
Another unresolved issue for semiconducting structure and contacts, challenges remain. The though further work is required to reduce the
CNTs is the need for a scalable and sustainable roles of material selection and purification processing temperature and to understand trans-
manufacturing approach to produce sufficient (discussed earlier), methods of fabrication, port and performance limits. In addition, realiz-
quantities of ultrahigh-purity semiconducting and doping control continue to be elucidated in ing an equally high-quality and scalable contact to
CNTs to meet the potentially large market rep- an expansive volume of reports. Indeed, one of n-type CNT transistors remains to be addressed.
resented not only by high-performance inte- the foremost challenges moving forward is de- Regarding TFTs from CNTs, much of the
grated circuits but also by high-volume printed termining what combination of materials and knowledge gained from nanoscale FET devices
electronics. Most solution-based separation processes (of the thousands reported) is most is applicable. The foremost exceptions are that
methods do not possess fundamental barriers appropriate to use. More systematic studies are a TFT technology should ideally be compatible
to scalability, but the yields of these processes needed that explore certain contact and gate- with large substrate sizes and have exception-
are ultimately limited by the quality of the input stack material configurations for their impact ally low cost. Because one of the primary appli-
raw material. Improvements in synthesis that on device performance, yield, reproducibility, cations for TFTs is in display backplanes, the
minimize impurities and maximize semicon- and stability. For example, it is clear that CNT materials and processes should be scalable to
ducting purity with narrow CNT diameter dis- channels are scalable to sub-10-nm lengths in large panels. Although device-level performance
tributions are needed to improve the yield of a variety of configurations, but it is not clear and size matter, TFTs have relaxed constraints,
downstream separation. An enticing option which device structure is superior (e.g., top-gate with more emphasis given to fabrication cost
would be to refine cloning (13) to the point that versus gate-all-around, side contact versus edge because these devices will be used in com-
iterative separation and amplification could be contact) and whether top-performing options modity applications (such as backplanes) or
achieved in a manner analogous to the poly- also have fabrication processes that are com- disposable applications (such as IoT). The
merase chain reaction (PCR) in biochemistry. patible with relevant manufacturing in comple- recent demonstration of recyclable printed

A B C
CNT logic and
CNT transistor RV16X-NANO

200-mm Si wafer sensors

of CNT RRAM
transistors
CNT logic
Silicon logic

CNTs
D E

S D
High-k G
20 µm 1 µm

Fig. 5. Wafer-scale and 3D integration of CNT transistors. (A) A 200-mm Si [Adapted by permission from Springer Nature Customer Service Center GmbH,
wafer with CNT transistors processed in a commercial silicon foundry. An image of a Springer Nature (57), copyright (2019)]. (C to E) Image and schematic of a 3D N3XT
single die or chip from the wafer is shown at the bottom left, and a schematic of the chip with monolithic integration of CNT transistors and RRAM memory layers on top of
CNT transistor structure is shown at the bottom right. D, drain; G, gate; k, relative silicon logic (C); cross-sectional TEM image showing the bottom Si logic layer, the
permittivity; S, source. [Adapted by permission from Springer Nature Customer RRAM memory layer, and the two CNT transistor layers [carbon nanotube field-effect
Service Center GmbH, Springer Nature (58), copyright (2020)]. (B) Optical image transistor (CNFET), logic, and sensors] (D); and scanning electron microscopy
of a RISC-V processor realized with CMOS CNT transistors (RV16X-NANO), including images of a CNT circuit and devices in the top layer of the 3D N3XT chip (E) (scale bar,
higher magnification images showing details of CNT circuits (false colors represent 500 nm). [Adapted by permission by Springer Nature Customer Service Center
different metal layers) and a single CNT device (CNTs are highlighted in yellow). GmbH, Springer Nature (65), copyright (2017)].

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 731

T R A N S I S T O RS

CNT-TFTs on paper substrates suggests sus- and surfaces without dangling bonds), it also 39. P. H. Lau et al., Nano Lett. 13, 3864–3869 (2013).
tainable implementations (77). Improvements retains all the benefits of an FET, which in- 40. P. Chen et al., Nano Lett. 11, 5301–5308 (2011).
41. S. Lu et al., ACS Nano 13, 11263–11272 (2019).
in CNT-TFT yield and stability will be critical, clude a well-developed circuit or system design 42. A. D. Franklin, Science 349, aab2750 (2015).
particularly the role of tube-tube contacts in ecosystem and a mature manufacturing tech- 43. J. A. Cardenas, J. B. Andrews, S. G. Noyce, A. D. Franklin, Nano
percolating networks. nology, with the further potential to integrate Futures 4, 012001 (2020).
44. N. F. Zorn, J. Zaumseil, Appl. Phys. Rev. 8, 041318 (2021).
in three dimensions for chips with increasing 45. Q. Cao et al., Nature 454, 495–500 (2008).
Technology outlook device count and connectivity. These benefits 46. S. Park, S. H. Kim, H. H. Choi, B. Kang, K. Cho, Adv. Funct.
The realization of a CNT transistor technology are projected to eventually outweigh all the Mater. 30, 1904590 (2020).
47. R. Yan, A. Ourmazd, K. F. Lee, IEEE Trans. Electron Dev. 39,
that meets high-volume manufacturing needs limitations because the power of incumbency 704–710 (1992).
has many remaining hurdles that require a and scalability in three dimensions cannot be 48. Z. H. Chen et al., IEEE Electron Device Lett. 29, 183–185
concerted effort from academia and industry understated. Between the opportunities in high- (2008).
49. A. D. Franklin et al., Nano Lett. 13, 2490–2495 (2013).
to surmount. Regarding semiconducting CNT performance digital logic with the potential for
50. A. D. Franklin et al., Nano Lett. 12, 758–762 (2012).
purity, although the highest possible purity 3D integration and the possibilities for printed 51. Q. Cao, J. Tersoff, D. B. Farmer, Y. Zhu, S. J. Han, Science 356,
remains ideal for EDP, logic design techniques and even recyclable thin-film electronics, CNT 1369–1372 (2017).
can be used to relax the requirement for cer- transistors warrant a renewed and even re- 52. C. Qiu et al., Science 355, 271–276 (2017).
53. G. Pitner et al., Nano Lett. 19, 1083–1089 (2019).
tain applications by about 100 times (from doubled effort from academic, government, and 54. Q. Cao et al., Science 350, 68–72 (2015).
99.9999 to 99.99%), without imposing addi- industry contributors. These molecular transis- 55. P. S. Kanhaiya, C. Lau, G. Hills, M. D. Bishop, M. M. Shulaker,
tional processing steps or redundancy (65). tor technologies are within reach, but only if the IEEE Trans. Electron Dev. 66, 5375–5380 (2019).
56. T. Srimani, G. Hills, C. Lau, M. Shulaker, in 2019 Symposium on
For high-performance digital systems, device remaining challenges are surmounted by the VLSI Technology (TEL, 2019), pp. T24–T25.
variations play an important role in determin- scientific and engineering communities. 57. G. Hills et al., Nature 572, 595–602 (2019).
ing the overall EDP and noise margin of the 58. M. D. Bishop et al., Nat. Electron. 3, 492–501 (2020).
59. L. Peng, Z. Zhang, C. Qiu, Nat. Electron. 2, 499–505 (2019).
system. CNT-specific sources of variation in- RE FERENCES AND NOTES
60. G. J. Brady et al., Sci. Adv. 2, e1601240 (2016).
clude CNT density and pitch (distance between 1. A. Aviram, M. A. Ratner, Chem. Phys. Lett. 29, 277–283 (1974). 61. Y. Xie, Z. Zhang, D. Zhong, L. Peng, Nano Res. 12, 1810–1816
2. M. Ratner, Nat. Nanotechnol. 8, 378–381 (2013). (2019).
CNTs in a multi-CNT transistor), CNT band- 3. S. J. Tans, A. R. M. Verschueren, C. Dekker, Nature 393, 49–52 62. C. Lau, T. Srimani, M. D. Bishop, G. Hills, M. M. Shulaker, ACS
gap (determined by the chirality and diame- (1998). Nano 12, 10924–10931 (2018).
ter), and extreme sensitivity to random fixed 4. R. Martel, T. Schmidt, H. Shea, T. Hertel, P. Avouris, Appl. 63. S. Salahuddin, K. Ni, S. Datta, Nat. Electron. 1, 442–450
Phys. Lett. 73, 2447–2449 (1998). (2018).
charges in the surroundings (which is also the
5. M. S. Dresselhaus, G. Dresselhaus, P. Avouris, Carbon 64. M. M. Shulaker et al., Nature 501, 526–530 (2013).
reason why CNTs are ultrasensitive sensors). Nanotubes: Synthesis, Structure, Properties, and Applications 65. M. M. Shulaker et al., Nature 547, 74–78 (2017).
The transistor width (perpendicular to the di- (Springer, 2001). 66. T. F. Wu et al., in 2018 IEEE International Solid - State Circuits
6. C. Jia, Z. Lin, Y. Huang, X. Duan, Chem. Rev. 119, 9074–9135 Conference (ISSCC) (IEEE, 2018), pp. 492–494.
rection of current flow) of logic technologies is 67. T. Srimani et al., in 2020 IEEE Symposium on VLSI Technology
(2019).
on the order of 20 to 40 nm. For a CNT density 7. D. Akinwande et al., Nature 573, 507–518 (2019). (IEEE, 2020), pp. 1–2.
of 250 CNTs per micrometer, there will only be 8. S. Das et al., Nat. Electron. 4, 786–799 (2021). 68. Y. Cao et al., ACS Nano 10, 6782–6790 (2016).
69. M. Schroter, M. Claus, P. Sakalas, M. Haferlach, W. Dawei,
5 to 10 CNTs in the channel; hence, variations 9. L. Zhang et al., Adv. Funct. Mater. 31, 2005471 (2021).
Electron Devices Soc. IEEE J. 1, 9–20 (2013).
10. R. Rao et al., ACS Nano 12, 11756–11784 (2018).
of the CNT density and the CNT pitch will lead 70. H. Shi et al., Nat. Electron. 4, 405–415 (2021).
11. F. Yang et al., Nature 510, 522–524 (2014).
71. M. Ha et al., ACS Nano 4, 4388–4395 (2010).
to substantial variations in the current drive. 12. J. R. Sanchez-Valencia et al., Nature 512, 61–64 (2014). 72. S. Lu, A. D. Franklin, Nanoscale 12, 23371–23390 (2020).
Design solutions that mitigate such varia- 13. B. Liu et al., Nano Lett. 13, 4416–4421 (2013). 73. K. Hong et al., Adv. Mater. 26, 7032–7037 (2014).
14. J. Wang et al., Nat. Catal. 1, 326–331 (2018). 74. K. Schnittker, M. Tursunniyaz, J. B. Andrews, J. Inf. Disp. 22,
tions are essential as part of a co-design process 15. P. Nikolaev et al., npj Comput. Mater. 2, 16031 (2016). 193–209 (2021).
for technology development (80). For example, 16. M. S. Arnold, A. A. Green, J. F. Hulvat, S. I. Stupp, 75. Z. Lin, G. Wu, L. Zhao, K. W. C. Lai, IEEE Nanotechnol. Mag. 13,
CNT bandgap variation directly translates into M. C. Hersam, Nat. Nanotechnol. 1, 60–65 (2006). 4–14 (2019).
17. H. Liu, D. Nishide, T. Tanaka, H. Kataura, Nat. Commun. 2, 309 76. S. Wang, J. Zhao, Q. Wang, D. Zhang, ACS Sustain. Chem. Eng.
variation in off-state leakage current through (2011). 10, 3851–3861 (2022).
the threshold voltage and band-to-band tun- 18. Q. Cao, S. J. Han, G. S. Tulevski, Nat. Commun. 5, 5071 77. N. X. Williams, G. Bullard, N. Brooke, M. J. Therien,
neling at the drain. Band-to-band tunneling (2014). A. D. Franklin, Nat. Electron. 4, 261–268 (2021).
19. J. A. Fagan et al., Adv. Mater. 26, 2800–2804 (2014). 78. A. J. M. Mackus et al., Appl. Phys. Lett. 110, 013101 (2017).
leakage varies exponentially with bandgap 79. S. K. Su et al., in 2022 IEEE Symposium On VLSI Technology
20. A. Graf et al., Carbon 105, 593–599 (2016).
and sets a minimum achievable leakage cur- 21. A. Javey, J. Guo, Q. Wang, M. Lundstrom, H. Dai, Nature 424, and Circuits (IEEE, 2022), pp. 403–404.
rent (81)—a boundary for trading on-state cur- 654–657 (2003). 80. G. Hills et al., IEEE Trans. Comput. Aided Des. Integrated Circ.
22. M. L. Geier, K. Moudgil, S. Barlow, S. R. Marder, M. C. Hersam, Syst. 34, 1082–1095 (2015).
rent with off-state leakage current by tuning 81. Q. Lin et al., IEEE Electron Device Lett. 43, 490–493 (2022).
Nano Lett. 16, 4329–4334 (2016).
the threshold voltage. Direct source-to-drain 82. G. Hills et al., IEEE Trans. NanoTechnol. 17, 1259–1269
23. M. L. Geier et al., Nat. Nanotechnol. 10, 944–948 (2015).
(2018).
tunneling current also depends exponentially 24. Z. Chen, J. Appenzeller, J. Knoch, Y. Lin, P. Avouris, Nano Lett.
on the bandgap and sets the limit for gate- 5, 1497–1502 (2005).
AC KNOWLED GME NTS
25. C. Wang, K. Ryu, A. Badmaev, J. Zhang, C. Zhou, ACS Nano 5,
length scaling. The choice of the CNT diameter 1147–1153 (2011). Funding: A.D.F. acknowledges support from the National Science
(bandgap) is faced with the same trade-off as 26. C. Gilardi et al., in 2021 IEEE International Electron Devices Foundation (NSF) (NSF ECCS-1915814). M.C.H. acknowledges
Meeting (IEDM) (IEEE, 2021), pp. 27.3.1–27.3.4. support from the NSF Materials Research Science and Engineering
other FETs. Small-bandgap CNTs have lower
27. Z. Wang et al., Nano Lett. 10, 2024–2030 (2010). Center at Northwestern University (NSF DMR-1720139). H.-S.P.W.
effective masses and higher on-state currents, 28. G. Pitner et al., in 2020 IEEE International Electron Devices acknowledges support in part from the Defense Advanced Research
and large-bandgap CNTs have lower tunneling Meeting (IEDM) (IEEE, 2020), pp. 3.5.1–3.5.4. Projects Agency (DARPA) Electronics Resurgence Initiative (ERI)
off-state leakage currents and can scale down 29. Z. Chen et al., Science 311, 1735 (2006). Three Dimensional Monolithic System on a Chip (3DSoC), the NSF
30. B. Q. Wei, R. Vajtai, P. M. Ajayan, Appl. Phys. Lett. 79, (award 1640060), and the Stanford SystemX Alliance, as well as
further in gate length and maintain higher 1172–1174 (2001). collaboration and discussion with G. Pitner and colleagues at Taiwan
operating voltages for high speed. The optimal 31. A. D. Franklin, Nature 498, 443–444 (2013). Semiconductor Manufacturing Company (TSMC), M. Shulaker at
choice is necessarily application-dependent and 32. F. Léonard, Nanotechnology 17, 2381–2385 (2006). MIT, and S. Mitra at Stanford. Competing interests: The authors
33. J. Zhang, N. P. Patil, A. Hazeghi, H.-S. P. Wong, S. Mitra, IEEE declare no competing interests. License information: Copyright ©
must be co-designed, given the target comput-
Trans. Comput. Aided Des. Integrated Circ. Syst. 30, 1103–1113 2022 the authors, some rights reserved; exclusive licensee
ing workloads (82). (2011). American Association for the Advancement of Science. No claim to
Although the CNT transistor inherits all the 34. W. Sun et al., Science 368, 874–877 (2020). original US government works. [Link]
limitations of a MOSFET (electrostatics and 35. K. R. Jinkins et al., Sci. Adv. 7, eabh0640 (2021). science-licenses-journal-article-reuse
36. L. Liu et al., Science 368, 850–856 (2020).
transport physics) and has all the challenges of 37. C. Zhao et al., Adv. Funct. Mater. 29, 1808574 (2019). Submitted 15 June 2022; accepted 7 September 2022
a low-dimensional channel material (contacts 38. D. Zhong et al., Nat. Electron. 1, 40–45 (2018). 10.1126/science.abp8278

732 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

REVIEW power while in a stable logical state, as it dis-
sipated power only during state transition (2).
Toward attoJoule switching energy In its infancy CMOS was viewed by circuit
designers as too slow compared with bipolar
in logic transistors transistors for use in high-performance pro-
cessors and was relegated to low-power and
Suman Datta1,2*, Wriddhi Chakraborty2, Marko Radosavljevic3 lightweight processing engines meant for data
terminals, calculators, and avionics. Advances
Advances in the theory of semiconductors in the 1930s in addition to the purification of germanium and silicon in high-resolution lithography techniques for
crystals in the 1940s enabled the point-contact junction transistor in 1947 and initiated the era of transferring fine patterns to the integrated
semiconductor electronics. Gordon Moore postulated 18 years later that the number of components in an circuit and in ion implantation and dopant
integrated circuit would double every 1 to 2 years with associated reductions in cost per transistor. Transistor activation techniques for forming shallow
density doubling through scaling—the decrease of component sizes—with each new process node continues source-drain junctions and tailoring channel
today, albeit at a slower pace compared with historical rates of scaling. Transistor scaling has resulted in doping profiles led to the design of highly scaled
exponential gain in performance and energy efficiency of integrated circuits, which transformed computing and fast-switching metal-oxide-silicon field-
from mainframes to personal computers and from mobile computing to cloud computing. Innovations in effect transistors (MOSFETs), which can have
new materials, transistor structures, and lithographic technologies will enable further scaling. Monolithic 3D p- or n-type channels (PMOS and NMOS, re-
integration, design technology co-optimization, alternative switching mechanisms, and cryogenic operation spectively), CMOS used complementary and
could enable further transistor scaling and improved energy efficiency in the foreseeable future. symmetrical pairs of these transistors. During
the next four decades, transistor scaling would

I
provide steady metronomic improvement for
n electronics, new inventions that reduce logic (1) to deliver two billion floating point CMOS circuit performance. Gordon Moore’s
energy consumption and enhance integra- operations per second (gigaFLOPs) of peak prediction that transistor count would double
tion capacity tend to become the dominant performance. The fast clock speed and dense every 2 years (3) proved true (Fig. 1). Demon-
device platform. For example, the triode in packaging led to such high heat dissipation strating clear advantages in energy efficien-
a vacuum tube was used to amplify signals that the machine had to be continuously cooled cy and device density over bipolar transistors,
for almost half a century but engineers recog- through circulation of Fluorinert liquid (sup- MOSFET would become the primary choice
nized drawbacks such as high power consump- plied by 3M) through the processor circuits. The for high-performance digital circuits for a broad
tion resulting in high heat dissipation as well as visible heat exchange water tank that came with range of technologies from general-purpose
overall lack of robustness. Energy consumption the liquid cooling system earned the Cray-2 the processors to domain-specific accelerators for
issues in telephone signal transmission drove nickname “Bubbles.” cloud-enabled data centers and from desktop
researchers at Bell Laboratories to invent a During the late 1980s, there was a wide- and mobile client computers to low-power em-
solid-state semiconductor device as a more re- spread design transition from bipolar to com- bedded processors for wearable and internet-of-
liable and energy-efficient replacement for the plementary metal oxide semiconductor (CMOS) things applications.
vacuum tube amplifier. The bipolar point-contact transistor technology. CMOS, which had been The operating voltage of the MOSFETs de-
was made from two isolated strips of gold foil on in development since the 1960s and was fa- creased progressively as the lithographic di-
a plastic wedge, and these strips were pushed vored by Japanese manufacturers, allowed for mensions shrank, leading to improved energy
down to make contact with a slab of n-doped a much-needed reprieve from power challenges. efficiency, faster speed, and lower cost per
germanium. This was dubbed the transistor and Unlike bipolar transistors, complementary pull transistor. The lower energy consumption per
was commercialized by their Western Electric up and pull down transistor configuration of switching event per transistor meant that the
subsidiary for radio transmission applications, CMOS technology consumed negligible standby integrated circuit could accommodate more
although this configuration was eventually re-
placed by the easier-to-manufacture and more Geometrical scaling
reliable silicon bipolar junction transistor. Equivalent scaling 103
10 Å
For three decades, the silicon bipolar tran- 10 9 Hyper scaling 14 Å
sistor would remain the device of choice in the 2 nm
design of both discrete and integrated circuits. FinFET 6T, 2 Fins
Number of transistors (mm–2 )

3 nm

Switching energy (attoJoule)

The performance of the bipolar transistor im- 8 7.5T, 3 Fins 5 nm Stacked
10 nano 102
proved with scaling but also led to increased 9T, 4 Fins 7 nm
sheet
self-heating and lower breakdown voltage. In
10 nm Gate all
the early 1980s, the increasing count of bipolar
transistors on the integrated circuit reached 10 7 High k/ 14 nm
around (GAA)
Metal gate nano sheet
an unacceptable level of power density. Man- 22 nm 101
aging power delivery and heat dissipation be- Strain 32 nm
came a formidable challenge. For example, the 106 45 nm
65 nm
iconic Cray-2 supercomputer of this era used 90 nm
dense packaging of multiple processors and fast 130 nm
180 nm
100
bipolar transistor logic called emitter-coupled 105
1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 2021 2023 2025 2027 2029
1
School of Electrical and Computer Engineering, Georgia Year
Institute of Technology, Atlanta, GA, USA. 2Department of
Electrical Engineering, University of Notre Dame, Notre Fig. 1. Metronomic progress in CMOS transistor density and switching energy. Improvements in
Dame, IN, USA. 3Components Research, Logic Technology
Development, Intel Corporation, Hillsboro, OR, USA. transistor density and switching energy are shown for geometrical, equivalent, and hyperscaling approaches.
*Corresponding author. Email: sdatta68@[Link] Examples for each approach are illustrated in Fig. 2.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 733

T R A N S I S T O RS

logic gates that switched faster while main- millimeter, and the switching energy per tran- degenerate conduction band valleys, whereas
taining the same operating power budget. The sistor fell below 1 picojoule (note that terms the hole mobility increases with compressive
power dissipated is energy consumed per switch- such as “90 nm process” refer to the design strain as a result of the lifting of degeneracy
ing event divided by the temporal duration of goals of the International Technology Road- between the light and heavy hole bands. A
the switching event. In theory, it is conceivable map for Semiconductors that assessed the transistor structure with epitaxial source-drain
to keep the power consumption of the chip research challenges for each size reduction). regions can be designed to impart strain of
constant but in reality the power budget is Operating voltage scaling was responsible for both types to the channel (Fig. 2A). After the
often exceeded to meet the performance spe- the reduction in the switching energy, whereas formation of gate stack and sidewall spacers as
cification, especially for high-performance com- gate length scaling increased switching speed. well as a silicon recess etch, selective hetero-
puting applications. During the era of geometric scaling, the pro- epitaxy can be performed to regrow a strained
Complicating the matter further, modern cessor clock frequency rose three orders of material in the source-drain regions adjacent
processor chips use a heterogenous structure in magnitude from 2 MHz in the case of the Intel to the channel. If the lattice spacing of this
which diverse functional blocks work with vary- 8008 processor (used to control traffic light sig- material is larger (or smaller) than silicon, uni-
ing activity factors (activity factors refer to how nals) to 3 GHz in the case of the Intel Pentium D axial compressive (or tensile) strain is induced
often the transistors undergo state transitions), 64-bit dual-core processor (used to power client in the channel. The transistor designers at Intel
resulting in non-uniform dissipation of power desktop computers). engineered uniaxial compressive strain in the
on the chip. This spatial nonuniformity is ac- It is implicit in Dennard’s scaling law that channel of PMOS transistors using selective
companied by temporal nonuniformity as circ- the threshold voltage of the transistor, VT, will silicon-germanium heteroepitaxy in the source-
uit blocks switch modes from idle to fully active scale proportionally with the operating voltage drain regions to boost hole mobility, and in-
at different points in time. Modern power- to ensure that there is sufficient voltage over- troduced uniaxial tensile strain in the channel
aware design techniques (e.g., power-down drive to provide high drive current. After three of the NMOS transistors using a tensile silicon
sleep modes, clock gating, dynamic voltage, decades of transistor downsizing, VT dropped nitride capping layer to increase electron mo-
and frequency scaling) exacerbate the spatio- so low that the subthreshold leakage current bility (4). The widespread use of strained chan-
temporal nonuniformity, creating “hot spots” (the current that flows through the transistor nel CMOS transistors in volume production at
on the chip. Failure to remove these hot spots even in its standby or off state) was now high the 90 nm and 65 nm nodes heralded the era of
in an expeditious manner not only affects chip enough to make standby power impose a con- equivalent scaling.
performance but also causes errors in logic straint in addition to the dissipated dynamic
states, accelerates aging of the transistors, and power. A second implicit assumption in Dennard’s Gate dielectrics
in extreme cases results in premature failure scaling law is the continued scaling of the phys- Toward the end of the geometric scaling era,
of the integrated circuit. ical gate oxide thickness, which provides electri- the gate dielectric SiO2 layers were so thin that
Reflecting on the innovations that have en- cal insulation of the gate electrode from the further decreases would effectively run out of
abled CMOS transistor scaling and elucidated current carrying transistor channel. After three atomic layers. There were two critical challenges
ongoing efforts to improve energy efficiency, decades of scaling, the thickness of the gate to replacing the polycrystalline silicon (polySi)
performance, and density of transistors in the dielectric—which used a nitrogen-containing gate electrode and SiO2 with polySi and higher
future, this Review revisits past innovations silicon dioxide—was reduced to 1.2 nm (4 mono- permittivity dielectrics such as hafnium diox-
and highlights future advances required for layers thick) at the 90 nm node. With such a thin ide (HfO2). First, HfO2 and polySi were incom-
the transistor to continue its scaling trajectory dielectric, the gate leakage current caused by patible because charge sharing between the
and approach the milestone of attoJoule (aJ) direct quantum mechanical tunneling became polySi and defect states within the HfO2 created
energy consumption per switching event and a noticeable fraction of the standby power. undesirable interface dipoles that led to un-
density of over three billion transistors per These unacceptable levels of subthreshold and acceptably high VT in the transistors. Second,
square millimeter of silicon. gate leakage current finally ended the era of polySi/HfO2 transistors exhibited severely de-
geometric scaling and began that of equivalent graded channel mobility caused by scattering
Geometric Scaling of Transistors following scaling of transistors. of the carriers in the channel by the soft op-
DennardÕs Law tical (SO) phonons arising from the polariza-
Robert Dennard and his colleagues at IBM Equivalent Scaling of transistors tion of the HfO2.
T. J. Watson Research Center established the The era of equivalent scaling of transistors is Metal gate electrodes proved effective in
MOSFET scaling rules that result in simul- defined by innovations in new materials and screening the high-κ SO phonons from coupling
taneous improvement in transistor density, transistor structures in addition to dimen- to the channel electrons and holes, whereas
switching speed, and power dissipation (2). sional scaling. In the absence of physical scaling strain engineering in the channel provided fur-
Each new generation of CMOS process tech- of the SiO2 gate oxide, transistor researchers ther boosts in both electron and hole mobility
nology was expected to reduce the minimum pioneered three complementary approaches to (5, 6). Transistor researchers successfully engi-
transistor dimension from X in the current improving transistor performance, energy effi- neered the gate metal electrodes to exhibit the
generation to 0.7X in the next, which then ciency, and scalability. Channel mobility was correct work functions (4.1 eV for NMOS and
led to a 0.49X reduction in transistor area enhanced through the introduction of strain. 5.1 eV for PMOS) using a replacement metal
and thus a 2X increase in transistor density. The electrical gate oxide thickness was scaled gate process flow. Also called the gate-last flow,
The industry embraced Dennard’s scaling and by replacing SiO2 with an alternative dielectric the gate metal electrode and the high-κ dielec-
between 1974 (when the idea was introduced) with high electrical permittivity κ. Planar, single- tric are deposited after the high-temperature
and 2003 (when the 90 nm process, also called gate architectures were replaced with nonplanar, activation anneal associated with source-drain
the 90 nm node, was introduced), the physical multigate structures to improve electrostatic dopant activation. This sequence preserves
gate length of the transistor was successfully integrity. the targeted threshold voltage, mobility, and
scaled from 1 mm to 35 nm. Additionally, the reliability at scaled electrical oxide thicknesses.
operating voltage was lowered from 4 V to Strain engineering High-performance high-κ/metal gate CMOS
1.2 V, the transistor density increased from a Electron mobility in silicon increases with ten- with negligible gate leakage, targeted thresh-
few hundred to 6 × 105 transistors per square sile strain as a result of splitting of the sixfold old voltages, low electrical oxide thickness, and

734 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

A B

sile L
Ten CES

te te GeX te te GeX
Ga in Ga Si1-X n Ga in Ga Si1-X n
Dra i Dra i
Dra Dra
GeX -k GeX
urc
e Si1-X e e
urc High Si1-X e h-k
So u rc ve So u rc Hig ive
sile
So ssi sile
So
res
s
Tenain pre n Tenain mp
Co rain
m
str Co strai str st

NMOS PMOS NMOS PMOS

C D Gate
E Gate
Gate in in
Dra Dra
in
Dra n
Gate
o Gate
t i in
e dir
ec Sou Dra n-CFET So
urc w rce urc
Dra
in
So nt flo in e
rre Dra n
Cu t i o
ire c
e Sou
urc w d
So nt flo rce p-CFET So
rre urc
Cu nFET e
NMOS

PMOS pFET

F
FinFET 9T 4F FinFET 7.5T 3F

Vss Vdd Vss Vdd

Metal pitch
NMOS PMOS NMOS PMOS

CFET (sheet) BPR 4T (3T)

FinFET BPR 5T 1F Nanosheet FET BPR 5T

PMOS
Buried
power NMOS PMOS
NMOS PMOS
rails
NMOS

Fig. 2. Platforms enabling transistor improvements. Schematics show transistor confinement will further improve with gate-all-around nanosheet transistors.
architectures that have enabled geometrical and equivalent hyperscaling. (A) Strain (E) To further increase density, p-channel nanosheet transistors will be stacked on
engineering of silicon used vertical tensile strain to improve electron mobility in n-channel nanosheet transistors or vice versa. (F) DTCO provides further improvement
NMOS devices, whereas compressively strained selective silicon-germanium in the power performance area for each node through codesign of the physical
heteroepitaxy improved hole mobility in PMOS devices. (B) Metal gate electrodes layout of the logic standard cell and specific technology choices. The design parameters
and high k gate dielectrics allowed scaling beyond the limits of SiO2 as a gate include the overall transistor size as measured in track height T, fin height, spacing,
dielectric. (C) The first nonplanar gate architecture, the FinFET, used vertical fins to and count (where F refers to the number of paired fins), the number of nanosheets
create tri-gate structures with improved electrostatic confinement. (D) Electrostatic stacked (which can be complementary in the CFET), and use of BPR.

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 735

T R A N S I S T O RS

channel mobilities for high-performance logic ratory of Hitachi had previously proposed a FinFETs operated at lower supply voltages yet
applications were developed with gate-last nonplanar, self-aligned double-gate transistor exhibited less leakage as a result of superior
flow (Fig. 2B) (7). In the gate-first approach, structure called the FinFET (Fig. 2C). This de- electrostatics compared with their planar
developed by Sematech and the IBM-led Al- sign, which operated in the fully depleted mode counterpart, with a marked reduction in both
liance, the high-k and metal gate stack goes and suppressed the short channel effect, pro- dynamic and standby energy consumption. Be-
through source-drain activation anneal and vided a path for further scaling (9). cause conduction occurs on all three surfaces
incorporates ultrathin capping layers—Al2O3 It would take more than a decade of effort of the fin, the device could drive more current
for the PMOS and LaOx for the NMOS—to for practitioners to combine the electrostatic through a given area of silicon than a planar
create dipoles that set the threshold voltage benefit of the fully depleted FinFET transistor transistor. The FinFET technology offered at the
of the device (8). For low power or DRAM ap- structure with high-k gate dielectric, work- 5 nm node remains the workhorse of leading-
plications in which threshold voltage and elec- function engineered metal gate electrodes edge logic technology (12).
trical oxide thickness requirements are relaxed, and channel strain engineering in a triple-
gate-first became a viable and promising option gate nonplanar geometry to demonstrate a Hyperscaling of transistors
for integrating a cost effective high-k and metal high-performance transistor called the tri-gate As the contacted gate pitch (the smallest space
gate CMOS solution. transistor (10). The piezoresistive coefficients between the gate electrodes of neighboring tran-
of silicon change their sign and magnitude as sistors) continues to scale and the transistor
Nonplanar multigate architectures the function of the direction of the crystallo- dimensions shrink, researchers are exploring
High-k dielectric and metal gate technology graphic planes (<100> versus <110>) for the top structures beyond FinFETs to further improve
provided two generations of transistor scaling fin surface versus the fin sidewalls. The channel electrostatics. They are also exploring the ver-
at the 45 and 32 nm nodes. Further scaling of strain, high-k gate dielectric, low channel doping, tical direction (out of the silicon plane) for
the transistor gate length scaled with a single gate electrode work functions, and selective epi- increasing the transistor density with gate-
gate planar structure was no longer adequate taxial source-drain regions were co-optimized all-around (GAA) architectures where transistor
to maintain electrostatic robustness and switch together with the fin dimensions (width and channels, often called nanosheets (NSs) or
off the transistor effectively. The electrostatic height) and the fin shape (trapezoidal versus nanoribbons, are completely surrounded by
integrity of the MOSFET is quantified by two rectangular) to demonstrate high-performance the gate dielectric and gate electrode (Fig. 3,
key metrics: the subthreshold slope (SS) and the NMOS and PMOS tri-gate transistors. D and E). This physical confinement provides
drain-induced barrier lowering (DIBL). Both The transition from single-gate planar tran- even stronger electrostatic control of the charge
SS and DIBL degraded as the channel length sistors to triple-gate nonplanar transistors was carriers in the channel compared with the
was further scaled beyond the 32 nm node. a major milestone in the journey of transistor FinFET. This effect is illustrated in Fig. 3, A
Researchers from the University of California scaling, and FinFET was introduced by Intel and B, where despite having the same critical
at Berkeley and the Central Research Labo- for mainstream logic at the 22 nm node (11). channel parameters (fin width in case of FinFET
and channel thickness in case of GAA), the NS
A FinFET GAA nanosheet B FinFET structure provides better electrostatics and sup-
80 GAA nanosheet
Footprint Footprint TNS
ports a shorter gate length at matched DIBL.
48 nm 70
48 nm 6 nm Design technology co-optimization (DTCO)—
TNS 60 which is of paramount importance in the era
DIBL (mV/V)

6 nm
50 of hyperscaling—optimizes together the stan-
Fin 40 dard logic cell and process technology features
Hfin
45 nm width 30 to improve performance, energy efficiency, den-
6 nm
WNS 20 sity, and cost (Fig. 2F). Process technologists
25 nm 10 collaborate with standard cell designers from
Fin pitch the very outset of each process node as they
0
24 nm Weff = 1.00x Weff = 1.30x 10 12 14 16 balance innovative design choices with process
Gate Length (nm)
capability. In the early days of the technology
C D 2D CMOS transition between planar MOSFET to non-
nanosheet inverter
planar FinFET, standard cells were designed
Gate
Signal wires with four fins per FinFET to deliver a drive
strength equivalent to planar devices, result-
Gate ing in standard cell tracks that are 9 tracks
contact 3D stacked CMOS (9 T) high. In subsequent nodes, fin depopu-
Gate nanosheet inverter lation (three fins per FinFET) enabled tech-
S/D contact
n-Drain Gate nologists to reduce track height to 7.5 T and
NMOS
p-Drain further maximize power, performance, and
PMOS density gain. In the future, use of buried power
rail (BPR) will lead to more aggressive track
Buried
power-rails
height reduction to 5 T.
In addition to scalability, NS transistors can
Fig. 3. Hyperscaling of transistors with stacked NSs. Compared with the FinFET design, a GAA transistor has be stacked several sheets high and provide
(A) effectively higher effective gate width (Weff) for the same geometric dimensions (“footprint”) and (B) superior the flexibility required in DTCO. In contrast
electrostatic gate control to enable further gate length scaling (14). Here, Tsus is nanosheet spacing, and TNS to FinFETs where the width of the transistors
and Wns are the NS thickness and width, respectively. (C) Side-by-side stacked NS PMOS and NMOS transistors comes in quanta of periphery of single fins
will be replaced by stacking NMOS on top of PMOS transistors [from (37) with permission]. (D) The additional (twice the fin height plus fin width), the width
density scaling resulting from vertical stacking is shown; Vss and Vcc are source and drain supply voltage, respectively in NS transistors can be more readily tuned by
[from (15) with permission]. lithography and optimized to provide higher

736 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

A B
Antiferroelectric (AFE)
Ferroelectric (FE)

Landau free energy (a.u.)

Mixed Ferroic phase Gate kT −Vg
SS = 2.3
Gate (FE+AFE) q −Cs
Mixed ferroic
phase dielectric Cins kT Cs
Cs SS= 2.3 1+
Source Dielectric (DE) Drain q Cins
n+ n+ High permittivity

Semiconductor
Cs Cins >> Cs SS~60 mV/dec

Polarization (a.u.)
C D

VG = 0V VG > 0V
X
Gate X EFS
EFS
—
X
Gate oxide Available states
—
Source Drain
p+ n+ Filled
Tunneling f(E) states
Semiconductor

E Dirac-Source cold FET F

Drain
E E
Exponential tail VG = 0V VG > 0V
Source EDirac
n2D~ exp((EF–E)/kBT)
Dirac-source Cold electron injection
SS , 60 mV/dec
DOS~EDirac –E
Fb EC
Semiconductor Fb
EF
EC
Gate oxide
DOS (E) n (E)
Gate
Superexponential tail
nDirac ~(EDirac –E) exp((EF–E)/kBT)

Fig. 4. Embodiments of unconventional transistors that can overcome valence and conduction bands. Here, f(E) is Fermi function, the probability of
the Boltzmann supply voltage limit of 500 mV. (A and B) Negative an electron to occupy a state with energy E, as a function of the density of
capacitance FETs use a mixed FE and AFE layer on top of the conventional states (DOS), and EFS is the energy of the highest filled state. (E and F) Dirac
dielectric to enhance the permittivity and maintain SS at ~ 60 mV per decade. source cold FETs make use of semiconductors such as graphene that have
(C and D) Tunnel FETs allow current to flow at positive gate voltages from a narrower super-exponential Boltzmann distribution of hot carriers that limits
immobile valence band states to mobile conduction band states [green region for thermal injection. The number of 2D and Dirac carriers are n2D and nDirac,
(D)]. Electron tunneling occurs through states created in the bandgap when the respectively, EF and EC denote the Fermi energy and conduction energy band,
bias voltages changes the band structure and narrows the gap between the respectively, and ϕB is the Schottky barrier.

overall current density per unit area (Fig. 3A). leverages back-side power delivery options using this density scaling approach does not de-
Transistors are designed with optimally de- BPR (15). pend on any specific transistor performance
signed widths to achieve a better balance of The concept of device stacking was proposed or scaling boosters, and as such system power-
power, performance, and cost to meet circuit decades ago for co-integrating heterogeneous performance balance will come from design
and product design specifications (13). All of devices such as silicon (Si) and germanium technology co-optimization effort.
these benefits are being harnessed now as (Ge) (16) or Si and compound semiconductors
major semiconductor manufacturers race to (such as GaAs) for additional functionality. Unconventional transistors operating at
introduce GAA stacked NS transistors in the However, for the purpose of density improve- extremely low supply voltages
3 nm nodes and beyond (14). Transistor den- ment the stacking must be done in an effici- Density scaling can be enabled with GAA-stacked
sity can be increased further with more ad- ent manner. In the proposal shown in Fig. 3C, NS transistors along with design technology
vanced stacking techniques that place NMOS electrical coupling of stacked devices is facili- co-optimization approaches such as self-aligned
and PMOS NS transistors on top of each other tated by constructing interconnects on both source-drain contacts, contacts landing direct-
rather than side-by-side (Fig. 3C). Nominal- the front and back side of the wafer (17). Stack- ly on active gates, buried supply rails, nano-
ly, by having two layers of transistors, density ing devices in this monolithic 3D fashion is an scale through-silicon-vias, and direct backside
could be improved by as much as 50% (Fig. area of intense research worldwide. This archi- source-drain contacts. However, the scaling of
3D) although more careful DTCO is needed tecture is referred to as complementary FET (or the supply voltage of operation of future tran-
that accounts for signal routing needs and CFET) (18). In addition to being very promising, sistors may slow down because of limits imposed

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 737

T R A N S I S T O RS

A B C
10–2 1.2 100
Iso-ION
300K 50
1.0
77K FinFET@Iso-speed

Metal line resistance (a.u.)

VDD scaling
10–4 –33%

Switching energy (aJ)

Drain current (A/mm)
at matched 0.8 20
ION ,IOFF
–42% 10
0.6 ~18x
10–6 5
0.4 Reduced
300K CMOS interconnect
delay 2
77K Cryo-CMOS 0.2
10–8
Iso-IOFF 1 aJ
1
0.0
0.0 0.2 0.4 0.6 0.8 20 100 200 300K 77K
Gate voltage (V) BEOL Metal line width (nm) CMOS Cryo-CMOS

D E F
State 0 Cryogenic cache memory 300K 6T SRAM 77K 6T SRAM
Metal 1T floating-body (FB) AM 77K 1T FBRAM
Poly-Si 1.00
TiN DS
Cores
Cache

n+ p+

Relative performance
Source High-k Drain 0.75
Drivers Box
Read1
State 1 0.50
Metal
DRead
Poly-Si
0.25
TiN
Off chip DRAM n +
p + Read0
Source + + + Drain 0.00
Box Injected holes VGS Cell area Latency Energy

Fig. 5. Benefits of cryogenic operation. (A) Reduced back-end-of-line (BEOL) metal interconnect resistance compared with room temperature CMOS (31).
(B) Cryogenic CMOS enables superior transistor performance attributed to key cryogenic boosters, such as higher channel mobility, steeper switching slope, and
lower supply voltage operation. (C) Cryogenic CMOS shows potential for scaling of VDD from 0.8 V at 300 K to 0.18 V at 77 K, hence approaching the aJ switching energy
level. (D to F) Cryogenic floating-body DRAM provides low latency and high bandwidth access to high-capacity on-die memory to augment cryogenic SRAM, providing
additional means to mitigate the memory bottleneck problem (35, 36).

by the so-called “Boltzmann tyranny” as CMOS 300 mV beyond the threshold voltage to deliver layer, CIns is now much larger than CS and SS is
technology reaches its thermal and reliability sufficient on-state current sets the minimum op- now effectively constant. The conventional FE
limits. The MOSFET fundamentally operates erating voltage of the MOSFET at 500 mV. double-well energy landscape can be effective-
as a barrier-controlled device where the gate The inability to scale supply voltage below ly flattened by the depolarization field arising
electrode electrostatically modulates the po- 500 mV will prevent CMOS technology from from the AFE phase through electrostatic and
tential barrier between the source and the ever reaching aJ switching energies and set a elastic interactions. The energy landscape flat-
channel to either impede or allow charge car- practical limit of ~10 aJ at the proposed limits tening has the effect of enhancing the permit-
rier injection into the channel. of hyperscaling. Doing otherwise will require tivity of the overall stack and allows the use of
However, the carriers follow Boltzmann sta- revolutionary changes in the fundamental a thinner electrical gate oxide of <6.5 Å. Re-
tistics and there exists a high-energy tail in the switching mechanisms of MOSFETs. sumption of gate dielectric thickness scaling in
distribution of carriers that enables injection future generations of transistors will enhance
of a small fraction of the charge over the bar- Negative capacitance control of the gate over the channel charge
rier. Additionally, the fraction increases with One such approach is negative capacitance and allow a transistor operating voltage of
temperature T and hence the MOSFET oper- FET (Fig. 4A), which boosts gate capacitance <500 mV.
ates at the thermionic limit of kBT/q (where kB by stabilizing competing phases of ferroelec-
is Boltzmann’s constant and q is the charge of tric (FE) and antiferroelectric (AFE) order in a Tunneling
an electron). The gate voltage must swing by at HfO2- ZrO2 superlattice (19). Increasing gate Tunnel FET (TFET) is another concept that
least 60 mV at room temperature to induce an capacitance allows the SS to remain at 60 mV replaces the conventional p-n junction formed
order of magnitude change in source-to-drain per decade even at lower operating voltages. by n-type sources and p-type channels in an
current (the so-called SS voltage), and in prac- This stability results because SS scales as (1 + NMOS transistor with a reverse biased inter-
tice imposes a lower limit on the minimum VT CS/CIns), where CS is the capacitance of the un- band tunnel junction using a p-type source and
of ~200 mV to keep the off-state leakage cur- derlying semiconductor and CIns is the capac- n-type channel (Fig. 4C). The switching slope in
rent at an acceptable level. The requirement of itance of the entire dielectric layer (Fig. 4B). such a device is not limited to 60 mV per decade.
a minimum gate overdrive voltage of around However, with the addition of the mixed ferroic Here, the thermal tail of the electrons in the

738 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

valence band of the p+ source is abruptly Thermal management of transistors at 77 K, which approaches the aJ switching
truncated by the energy gap of the source Energy-efficient scaling of CMOS technology energy level. The bulk resistivity of copper (Cu)
and can lead to steep switching character- while still delivering on performance in the interconnects also improves at low temperature
istics as the TFET is turned on (Fig. 4D). A era of hyperscaling remains a challenge. New (31) (Fig. 5C) and results in reduced intercon-
large indirect bandgap and relatively heavy computational workloads such as training nect delay and higher computing throughput.
carrier mass in silicon means that the drive of large-scale deep neural networks involve Limited on-die cache memory capacity [a
current in silicon TFET is very low. a colossal amount of data movement. In such static random-access memory (SRAM) situated
High-quality epitaxial growth of compound a situation, CMOS logic transistor perform- on the CPU itself] will still require a cryogenic
semiconductors and their incorporation into ance needs to be balanced with high capacity, processor core to access off-chip dynamic random-
TFETs show promise from drive current per- low latency (fast switching speeds), and high- access memory (DRAM) for data-intensive work-
spective as a result of low bandgap and carrier bandwidth cache memory [near or on the cen- loads (Fig. 4D) and will limit performance and
mass. Staggered bandgap or even broken band- tral processing unit (CPU)] to avoid the so-called energy efficiency gains. Augmenting an on-die
gap heterojunctions tend to have the highest memory-wall bottleneck (30). Low-resistance cryogenic SRAM cache with high-performance
TFET drive current (20, 21). However, demon- metal line interconnects are also necessary embedded memory will be critical to realize
stration of TFETs with both high drive current to facilitate faster data movement between the full benefit of cryogenic CMOS technology.
and steep switching characteristics but oper- the logic and memory subsystems to maximize Cache memory is subcategorized into L1, L2,
ating voltages <500 mV remain elusive due to overall system performance (Fig. 5A) (31). Post- and L3 cache levels depending on its relative
limitations of gate-dielectric quality for these CMOS device technology with substantially distance from the logic core. Because the L1
newly introduced materials and difficulties in improved energy-delay product (EDP) metrics cache is placed near the processor and requires
creating defect-free tunnel junctions (22). remain to be developed (32). extremely low latency, cryogenic 6T-SRAM with
Cryogenic CMOS—that is, CMOS operating 20% reduced latency and energy than 300 K
Dirac source at liquid nitrogen temperatures (77 K)—may SRAM (35, 36) is a potential memory candidate.
The Dirac source FET is another candidate ca- provide a path toward scaling devices and re- Because the latency constraint can be some-
pable, in principle, of switching with subther- what relaxed for upper-level cache memory
mal slope and could potentially operate at low (L2 and L3), single transistor floating-body
supply voltage (Fig. 4E). The charge carriers in
the 2D gapless graphene source obey a linear
“Energy-efficient scaling of CMOS RAM (FBRAM) is being explored for cryogenic
L2/L3 cache. Cryogenic one-transistor FBDRAM
energy dispersion relation mimicking mass- technology while still delivering has been demonstrated with ~5-ns program-
less relativistic Dirac particles. This property ming speed and 1.4 V programming voltage at
results in a narrower carrier density distribu- on performance in the era of 77 K (36). The floating-body effect in a single
tion around the Fermi level than that of the
Boltzmann distribution of carriers in conven- hyperscaling remains a challenge. MOS transistor performs memory operations
by injecting majority carriers into the “body”
tional MOSFETs (Fig. 4F). A Dirac source FET
with a carbon nanotube channel was demon-
New computational workloads such (the gate region) (Fig. 5E) to perform the write
operation and are later expunged during an
strated with an average switching slope of as training of large-scale deep erase operation. The presence and absence of
40 mV per decade over four decades of change excess majority carriers in the floating body
in source-to-drain current at room tempera- neural networks involve a colossal modulate the transistor source barrier and the
ture and a modest on-state current (23).
The intrinsic performance of future transis-
amount of data movement.” drain current during the read operation (Fig.
5E). The low temperature operation enables
tors benefits from new switching mechanisms higher transistor read current, suppresses
discussed above. Still, source-drain (S/D) ex- ducing dynamic energy consumption to the Shockley-Read-Hall recombination rate and
ternal resistance (REXT) may become a funda- aJ level. Cryogenic operation inherently en- produces longer data retention time. The sim-
mental bottleneck for achieving maximized ables steep SS CMOS logic transistors because pler 1T structure allows 8.33X higher capacity
transistor performance. At the 5 nm node, SS scales linearly with temperature. Thresh- than 6T SRAM, along with 2.5X lower latency
contact resistance (RCON) contributes to 40% old voltage can now be further reduced by and 2.7X lower energy than 6T SRAM array
of REXT with epitaxial S/D extension resistance using gate electrode work-function engineer- (Fig. 5F) (36) and provides a superior core-to-
(REPI) being the other limiting factor (24). This ing while maintaining iso-leakage as room cache balance, reduced power consumption,
problem increases as the contact area con- temperature (300 K) CMOS (Fig. 5B). Improved and faster computational throughput.
tinues to scale. Included among various ap- carrier mobility and lower source/drain ex-
proaches investigated for RCON reduction is trinsic resistance (33, 34) at cryogenic tem- Outlook
Schottky contact barrier modulation through peratures further contributes to higher drive During the next decade, CMOS transistor tech-
insertion of an atomically thin interlayer (TiO2) current and could enable aggressive supply nology is poised to reach the unprecedented
at the metal-Si interface (25), implementation voltage (VDD) scaling. integration complexity of 2 billion transistors
of dual-layer metal stack (NiPt-Ti) (26), or even Taking into account the energy cost of cool- packed into one millimeter square of silicon with
maximization of contact area through wrap ing to maintain the computing system at low each transistor consuming only a few attoJoules
around contact (27). By contrast, as dopant (P temperatures, it is possible to demonstrate a of energy during switching operations. With
and B) concentration in Si S/D approaches the net energy delay product (EDP) benefit. Low an average projected switching time of less
solid-solubility limit (~1020 atoms/cm3), high VDD operation enables a pathway for equiv- than a picosecond and an activity factor of a
activation of dopants through metastable Si:P alent oxide thickness scaling at cryogenic tem- few percent, such an integrated circuit will con-
alloying with 4 atomic % phosphorus concen- peratures while maintaining iso-gate leakage, sume only 20 watts of power. However, in order
tration (28), solid phase epitaxially regrowth which opens up a broader design space for EDP to achieve these goals, a multitude of innova-
(SPER) of S/D through millisecond or nano- optimization that takes into account device-to- tions will be needed in materials, 3D transistor
second laser annealing (29) can lead to lower device variation. A cryogenic-FinFET could po- structures, interconnects, backside power de-
contact resistivity (ρc) as well as REPI. tentially scale VDD from 0.8 V at 300 K to 0.18 V livery networks, and monolithic 3D integration

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 739

T R A N S I S T O RS

techniques, It will also require low latency ac- 13. G. Bae et al., “3nm GAA Technology featuring Multi-Bridge- 28. C.-N. Ni et al., in 2016 International Symposium on VLSI
cess to on-die high-capacity memory beyond Channel FET for Low Power and High Performance Technology, Systems and Application (VLSI-TSA), (IEEE, 2016),
Applications,” 2018 IEEE International Electron Devices Meeting pp. 1–2.
SRAM and fine-grained thermal manage- (IEDM), 2018, pp. 28.7.1-28.7.4, doi: 10.1109/ 29. Y. R. Yang et al., in 2016 EEE Symposium on VLSI Technology,
ment features. IEDM.2018.8614629. (IEEE, 2016), pp. 1–2.
14. J. Zhang et al., “Full Bottom Dielectric Isolation to Enable 30. H. L. Chiang et al., in 2021 Symposium on VLSI Technology,
Stacked Nanosheet Transistor for Low Power and High (IEEE, 2021), pp. 1–2.
RE FE RENCES AND N OT ES Performance Applications,” 2019 IEEE International Electron 31. H. L. Chiang et al., in 2020 IEEE Symposium on VLSI
1. Fairchild Semiconductor, “CMOS, the Ideal Logic Family”, Devices Meeting (IEDM), 2019, pp. 11.6.1-11.6.4, doi: 10.1109/ Technology, (IEEE, 2020), pp. 1–2.
Application Note 77 at the Wayback Machine, 1983. IEDM19573.2019.8993490. 32. C. Pan, A. Naeemi, IEEE J. Explor. Solid-State Comput. Devices
2. R. H. Dennard et al., IEEE J. Solid-State Circuits SC-9, 256–268 15. C.-Y. Huang et al., in IEEE International Electron Devices Circuits 3, 101–110 (2017).
(1974). Meeting (IEDM) Technical Digest (IEEE, 2020), 33. W. Chakraborty et al., in Digest of Technical Papers, IEEE
3. G. E. Moore, Electronics 38, 114–117 (1965). pp. 20.6.1–20.6.4. Symposium on VLSI Technology, (IEEE, 2021), pp. 1–2.
4. K. Mistry et al., in Digest of Technical Papers at Symposium on 16. S. Subramanian et al., in Digest of Technical Papers, IEEE 34. W. Chakraborty et al., in IEEE International Electron Devices Meeting
VLSI Technology (IEEE, 2004), pp. 50–51. Symposium on VLSI Technology (IEEE, 2020), pp. 1–2. (IEDM) Technical Digest, (IEEE, 2019), pp. 39.4.1–39.4.4.
5. S. Datta et al., in IEEE International Electron Device Meeting 17. W. Rachmady et al., in IEEE International Electron 35. W. Chakraborty et al., in IEEE International Electron Devices
(IEDM) Technical Digest (IEEE, 2003), pp. 653–656. Devices Meeting (IEDM) Technical Digest, (IEEE, 2019), Meeting (IEDM) Technical Digest, (IEEE, 2021), pp. 40.1.1–40.1.4.
6. R. Chau et al., IEEE Electron Device Lett. 25, 408–410 (2004). pp. 29.7.1–29.7.4. 36. W. Chakraborty et al., in Digest of Technical Papers, IEEE
7. K. Mistry et al., in IEEE International Electron Devices Meeting 18. J. Ryckaert et al., in 2018 IEEE Symposium on VLSI Technology Symposium on VLSI Technology and Circuits, (IEEE, 2022),
(IEDM) Technical Digest (IEEE, 2007), pp. 247–250. (IEEE, 2018), pp. 141–142. pp. 302–303.
8. K. Henson, in IEEE International Electron Device Meeting (IEDM) 19. S. Cheema et al., Nature 604, 65–71 (2022). 37. L. Liebmann et al., in 2021 IEEE International Electron Devices
Technical Digest (IEEE, 2008), pp. 645–648. 20. D. K. Mohata et al., in IEEE Electron Devices Meeting (IEDM) Meeting (IEDM), (IEEE, 2021), pp. 3.1.1–3.1.4.
9. D. Hisamoto et al., IEEE Trans. Electron Dev. 47, 2320–2325 (2000). Technical Digest (IEEE, 2011), pp. 33.5.1–33.5.4.
10. J. Kavalieros et al., in Digest of Technical Papers, IEEE 21. G. Dewey et al., in IEEE Electron Devices Meeting (IEDM) AC KNOWLED GME NTS
Symposium on VLSI Technology (IEEE, 2006), pp 62–63. Technical Digest (IEEE, 2011), pp. 33.6.1–33.6.4. Funding: W.C. and S.D. would like to acknowledge the support by
11. C. Auth et al., “A 22nm High Performance and Low-Power 22. Y. Qiu, R. Wang, Q. Huang, R. Huang, IEEE Trans. Electron Dev. the Defense Advanced Research Projects Agency (DARPA) Low
CMOS Technology Featuring Fully-Depleted Tri-Gate 61, 1284–1291 (2014). Temperature Logic Technology (LTLT) Program. Competing
Transistors, Self-Aligned Contacts and High Density MIM 23. C. Qiu et al., Science 361, 387–392 (2018). interests: The authors declare no competing interests. License
Capacitors”, Digest of Technical Papers, IEEE Symposium on 24. N. Breil, in 2021 IEEE International Interconnect Technology information: Copyright © 2022 the authors, some rights reserved;
VLSI Technology, 2012. Conference (IITC) (IEEE, 2021), pp. 1–3. exclusive licensee American Association for the Advancement of
12. J. C. Liu et al., “A Reliability Enhanced 5nm CMOS Technology 25. A. Agarwal et al., Appl. Phys. Lett. 104, 112101 (2014). Science. No claim to original US government works. [Link]
Featuring 5th Generation FinFET with Fully-Developed EUV and 26. P. Adusumilli et al., in 2016 IEEE Symposium on VLSI [Link]/about/science-licenses-journal-article-reuse
High Mobility Channel for Mobile SoC and High Performance Technology, (IEEE, 2016), pp. 1–2
Computing Application”, IEEE International Electron Devices 27. S.-A. Chew et al., in 2017 IEEE International Interconnect Submitted 6 September 2022; accepted 12 October 2022
Meeting (IEDM) Technical Digest, 2020. Technology Conference (IITC), (IEEE, 2017), pp. 1–3. 10.1126/science.ade7656

740 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RESEARCH
IN S CIENCE JOURNAL S
Edited by Michael Funk

3D PRINTING

Metal nanoclusters
improve printing

T
o enhance the properties of
two-photon three-dimensional
(3D)–printed polymeric struc-
tures, photoinitiators are added to
increase the formation of radicals
and to improve the polymerization pro-
cess. However, each polymer may require
a specific initiator, and the organic mol-
ecules typically used do not enhance the
properties of the finished product. Li et al.
report an improvement in two-photon 3D
printing in which atomic metal clusters
are used as two-photon absorbers to
enhance the mechanical properties of
the printed structures. The authors reveal
the formation of a complex architecture
with tunable and hierarchical porosity
showing high mechanical strength and
stability. —MSL
Science, abo6997, this issue p. 768

The addition of metal nanoclusters enables 3D

printing of intricate and durable polymer lattices.

SOLAR CELLS Iodide-based solar cells could into and through dry pockets in the effect of vibrational strong
retain 90% of their efficiency soil. In moist conditions, water coupling (VSC), a promising
Staying distributed under illumination after 3000 flows from the root surface route to manipulating chemi-
with selenophene hours at 45°C. —PDS into the phloem, but in a dry cal dynamics in condensed
In perovskite solar cells, Science, abn3148, this issue p. 747 spot, water flows instead from phases. Previous studies of
mixtures of cations and anions the phloem out into the root VSC-modified chemistry
can be used to tune band tissues. This reverse flow car- used traditional chemical
gaps and increase stability, PLANT SCIENCE ries along the phloem-derived kinetics tools that cannot
but these mixtures are prone hormone abscisic acid, which properly address ultrafast
to unwanted segregation that
Lateral root closes off intercellular pores, vibrational processes in their
degrades performance. Bai development blocking the ability of the hor- natural time scales. Using
et al. found that segregation Plant roots are most effective mone auxin to initiate lateral ultrafast two-dimensional
can be slowed by creating in pockets of soil containing root development. —PJH infrared spectroscopy, Chen
an initially homogeneous the water that the plant needs. Science, add3771, this issue p. 762 et al. showed that polaritons
distribution of cesium and Water is not uniformly distrib- (delocalized superpositions of
formamidinium cations in the uted throughout the soil, and vibrations and electromagnetic
colloidal film precursors. The the xerobranching response, in CHEMICAL PHYSICS cavity modes) can switch the
additive selenophene main- which lateral root formation is rates of two vibrational energy
tained homogeneity during suppressed, ensures that roots
Spectroscopy for pathways of a metal carbonyl
PHOTO: LI ET AL.

film processing and device are not uniformly distributed polaritonic chemistry compound under VSC, making
operation for both single- and either. Mehra et al. show what There is currently consider- intramolecular vibrational
mixed-halide perovskites. happens as plant roots grow able interest in understanding redistribution more favorable

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 741

1118ISIO_16181047.indd 741 11/14/22 4:17 PM

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

over pseudorotation (the CORONAVIRUS

opposite is true outside of a
cavity) (see the Perspective
Casting a wide NET IN OTHER JOURNALS Edited by Caroline Ash
and Jesse Smith
by Chuntonov). The authors on long COVID
also clarified the role of dark A substantial number of
modes in VSC, a longtime individuals who recover from
question that has been heavily COVID-19 still present with
debated but had lacked direct long-term sequelae. George
experimental evidence. —YS et al. followed individuals after
Science, add0276, this issue p. 790; recovery from severe COVID-19
see also ade9815, p. 712 to identify features that
distinguished those who had
evidence of long-term pul-
QUANTUM PHYSICS
monary sequelae from those
Tough edges who made a full recovery.
The dynamics of quantum They found that a neutrophil-
many-body systems can be associated inflammatory
profoundly affected by their phenotype was apparent
interaction with the environ- in those who had persis-
ment. This includes systems tent pulmonary symptoms,
that have topological protec- and evidence of neutrophil
tion from certain kinds of extracellular traps, or NETs,
perturbations due to sym- was found in the blood of
metry. Mi et al. studied the these individuals. These data
interplay between symmetry highlight a potential role for
and noise using a chain of 47 neutrophils in pulmonary long
superconducting qubits. They COVID. —CSM
implemented a periodically Sci. Transl. Med. 14,
driven transverse Ising spin eabo5795 (2022).
model, and found that the
system’s edge modes were
HUMAN GENOMICS
surprisingly resilient to some
types of symmetry-breaking Nonrandom selection
noise. —JS Many studies have exam-
Science, abq5769, this issue p. 785 ined correlations between
complex traits, assuming
that correlations implied a HOST AND PATHOGEN be replicated by injecting tick
NEURODEGENERATION genetic connection even when saliva, along with Borrelia spiro-
there was no clear biological
Ex vivo model ticks chetes, into human skin explants,
Protecting the brain overlap. It had been proposed all the right boxes providing a potential model to
from prions that overlapping genes with Lyme disease, which is caused by study human-tick interactions ex
Various neurodegenera- pleiotropic effects contribute Borrelia burgdorferi, is transmit- vivo. —STS
tive diseases are caused by to multiple different psychi- ted by ticks of the genus Ixodes. J. Clin. Invest. 10.1172/
prions or prion-like misfolded atric disorders or even across Although tick saliva is generally JCI161188 (2022).
proteins that aggregate disease categories such as known to be immunosuppressive,
and propagate through psychiatric and metabolic the effects of tick feeding and the
VIROLOGY
the brain. Because loss of conditions. By combining accompanying tissue damage
acetylcholine signaling is analysis of phenotype data on local and systemic immunity Some viruses mix it up
associated with cognitive from two large, population- are not well understood. Strobl Respiratory viruses such
deficits in patients, Dwomoh based cohorts with in silico et al. recruited individuals with as influenza virus (IAV) and
et al. investigated the effect simulations, Border et al. recent tick bite history and took respiratory syncytial virus (RSV)
of enhancing acetylcholine demonstrated that many skin biopsies from both the site are common among children
signaling in a mouse model correlations between human of the bite and a control area. and may cause coinfections,
of prion disease. Systemic traits can be explained Neutrophils, B cells, and T cells sometimes with exacerbated

PHOTO: FRANS LEMMENS/ALAMY STOCK PHOTO

administration of a positive instead by cross-trait (particularly tissue-resident symptoms. Such infections are
allosteric modulator targeting assortative mating, which is memory T cells) were increased likely to have different dynamics
the M1 acetylcholine receptor an individual’s tendency to at the site of the bite, whereas from single infections. Haney et
reduced the levels of prion- choose a mate with specific other immune cell populations al. examined IAV and RSV coin-
induced molecular markers characteristics that have no such as Langerhans cells were fections in human lung cells in
in the hippocampus, restored genetic relationship (see the decreased. Cytokine produc- vitro and observed hybrid virus
various cognitive functions, Perspective by Grotzinger and tion by T cells was impaired, and particles with surface glycopro-
and slowed disease progres- Keller). —YN innate lymphoid cells at bite teins and ribonucleoproteins
sion in the mice. —LKF Science, abo2059, this issue sites were also decreased. The from both viruses. Although RSV
Sci. Signal. 15, eabm3720 (2022). p. 754; see also ade8002, p. 709 immunomodulatory effects could tends to be at a disadvantage in

742 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118ISIO_16181047.indd 742 11/14/22 4:17 PM

METALLOPROTEINS
BEHAVIOR A design that bucks
Too hot to shoal the trend
Although metal sites in

C
limate change is altering a wide array
of environmental conditions, and we proteins can be somewhat
know little about how these changes selective, they often fol-
may affect various traits such as low a predictable trend in
social behavior. In fishes, research which copper ions are the
has shown that sociability can be lower in strongest binders. Choi et al.
warmer temperatures, a possible result of redesigned an artificial metal-
selection for reduced disease transmis- binding protein such that two
sion or resource competition. To test for a molecules were linked by a
direct influence of warmer waters, Pilakouta disulfide bond but were still
et al. took advantage of a unique natural slightly flexible and able to
system in which adjacent and overlapping adopt a variety of conforma-
populations of three-spine stickleback live tions. This construct bound
in either cold pools or those warmed by zinc, nickel, and copper ions
geothermal outflow. They found that the fish and displayed selectivity for
from isolated cold pools were more sociable zinc in competition experi-
than those from warm pools, but there was ments, possibly enabled by
no evidence of plastic responses to water binding three ions instead of
temperature in adult fish. Common garden one. Structural and biophysi-
rearing experiments, however, revealed that cal analysis of the binding
sociability was heritable and that the fish mode reveals how tetrahedral
reared in warmer waters were less sociable, coordination geometry guides
suggesting that temperature changes could metal selectivity in this sys-
alter fundamental behaviors. —SNV tem. —MAF
J. Am. Chem. Soc.
Glob. Chang. Biol. 10.1111/gcb.16451 (2022).
144, 18090 (2022).

Three-spined sticklebacks tend to form shoals

in cool water, but not in warm conditions. HISTORY OF SCIENCE
A fragment of
Hipparchus’ catalog
coinfections with IAV, the hybrids et al. report a trapping reagent where there was no evidence of The second-century BCE
could apparently evade anti- that reacts efficiently with a wide dairy consumption and others Hellenistic astronomer
influenza antibodies by using variety of radicals by the displace- with low levels. Their dating esti- Hipparchus is known to have
the RSV fusion glycoprotein to ment of a nitroxide leaving group mates are earlier than those of produced a star catalog, but
get into cells from which IAV on a carbon adjacent to an olefin. previous studies and indicate that no copy of it has survived.
receptors had been removed. Its reactivity toward non-carbon- dairy consumption was adopted Gysembergh et al. have reex-
Proteins from both viruses were centered radicals is a particular more widely and less gradually amined previously recorded
found to colocalize on the api- improvement over prior probe than was previously believed. multispectral imaging of a
cal side of bronchial epithelial compounds. —JSY —CNS palimpsest, a parchment
cells. The authors did not test J. Am. Chem. Soc. 144, 15969 (2022). Proc. Natl. Acad. Sci. U.S.A. that was incompletely erased
whether these hybrid viruses are 119, e2109325118 (2022). before being reused for a
transmissible between animals different text. They identi-
ANTHROPOLOGY
in vivo. —CA fied about 100 words of faint
Nat. Microbiol. 7, 1879 (2022). Updating carbon dating Greek writing describing the
Lactose tolerance is a classic constellation Corona Borealis
example of selection in humans and four of its brightest stars.
ORGANIC CHEMISTRY
that substantially expanded The coordinate system and
Trapping radicals available food sources with positions identify it as a
Many chemical reactions proceed the advent of agriculture. copied part of Hipparchus’
PHOTO: WOLFGANG SAUBER/ CC BY SA 3.0

through radicals, highly reac- However, timing the rise of catalog. Although brief, the
tive intermediates bearing an dairy consumption is difficult fragment is sufficient to
unpaired electron. Because they from archaeological sources demonstrate that Hipparchus’
are often short-lived, radicals can alone. Casanova et al. used a catalog was more precise
be hard to detect, but verifying new method of carbon dating than the one produced by
their presence is important to directly on the dairy fat residues An example of a Linearbandkeramik Claudius Ptolemy three cen-
optimize conditions and, in some found in pottery. They examined vessel, in which traces of dairy fat turies later. —KTS
cases, to avoid hazardous unan- pottery vessels from across cen- enabled carbon dating of Early J. Hist. Astron. 10.1177/
ticipated chain reactions. Williams tral Europe and found locations Neolithic farming activities in Europe. 00218286221128289 (2022).

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 743

1118ISIO_16181047.indd 743 11/14/22 4:17 PM

RESEARCH

ALSO IN SCIENCE JOURNALS Edited by Michael Funk

CORONAVIRUS how human activity has affected and identified several of its 3-position. Recyclization then
the planet. —GKA constituent proteins. They then yielded a selectively halogenated
Measuring menstruation Science, ade2310, this issue p. 706 showed that mutations in one of ring. A complementary approach
after vaccines these proteins are responsible by Cao et al. relies on a reversible
Menstrual changes, particu- for the clinical infertility associ- reaction of the pyridines with an
larly longer cycles and heavier DEVELOPMENTAL BIOLOGY ated with oocyte maturation alkyne and ester, disrupting the
bleeding, have been reported arrest in human patients. —YN electronic structure to activate
in association with various
Rethinking Hippo Science, abq7361, this issue p. 745 the 3-position for halogenation
vaccines, including those for signaling as well as trifluoromethylation
COVID-19. These observa- Mutations that cause dys- (see the Perspective by Joo).
tions have led to widespread regulation of the Hippo signaling CANCER —JSY
misinformation that resulted pathway are known to cause Science, add8980 and ade6029,
in COVID-19 vaccine hesitancy excessive growth of organs,
FMRP and tumor this issue, pp. 773 and 779;
among young women. Various which has led many research- immunity see also ade5501, p. 710
studies have since found that ers to think of this pathway as Many tumors have developed
although menstruation can be a master regulator of organ mechanisms rendering them
affected by COVID-19 vaccina- growth. Studying fruit fly resistant to attack and destruc- SKIN INFLAMMATION
tion, these changes are transient eye discs and mouse livers, tion by the immune system.
and resolve within a few months. Kowalczyk et al. found instead Zeng et al. report that fragile
Sustenance for
In a Perspective, Male discusses that Hippo signaling does not X mental retardation protein skin-resident TH17 cells
studies looking at how COVID-19 instruct normal growth. The (FMRP) is highly expressed The T helper 17 (TH17) cell subset
vaccination and infection affect Hippo-induced overgrowth in human cancers, and they provides host protection from
menstruation and the possible phenotypes appear to be caused propose that it is involved in extracellular pathogens but
underlying mechanisms. These by Hippo signaling activating antitumor immunity. FMRP is also contributes to autoim-
effects are not routinely moni- abnormal genetic programs. best known as an RNA-binding mune diseases. Interleukin-23
tored in vaccine trials, and this is These findings challenge a protein that regulates the stabil- (IL-23) promotes cutaneous
a missed opportunity that could long-standing idea about the ity and translation of neuronal TH17 responses, but the source
reveal feedback between the role of Hippo signaling in organ RNAs. By genetically inactivating of the IL-23 that supports
immune system and the female growth and suggest the need to the FMRP gene in mouse cancer skin-resident memory TH17
reproductive system. —GKA re-evaluate our understanding cells, the researchers found cells has been unclear. Whitley
Science, ade1051, this issue p. 704 of its function in other contexts, that FMRP-deficient tumors et al. characterized the effects
such as in cancer and regenera- had reduced growth and were of blocking IL-23–mediated
tion. —SMH and BAP more susceptible to attack by T signaling in mice on TH17 func-
GEOLOGY Science, abg3679, this issue p. 744 lymphocytes. Tumor cells lacking tion in defending against skin
FMRP showed remodeling of infection by Candida albicans
Defining the the tumor microenvironment, and in promoting inflammatory
Anthropocene epoch HUMAN FERTILITY macrophage polarization, and dermatitis. Dermal myeloid
A new geological epoch is upon up-regulation of the chemokines cells were identified as the
us. Moving beyond the Holocene,
Organizing for involved in effector CD8+ T cell critical source of IL-23 needed
we are now in the Anthropocene, meiotic success recruitment. —PNK to sustain skin-resident memory
which is proposed to have Proper organization of the Science, abl7207, this issue p. 746 TH17 cells. These findings provide
begun around the mid-20th microtubules is vital for ensuring a deeper understanding of how
century. The definition of the that daughter cells end up with blocking antibodies targeting
Anthropocene involves the iden- the appropriate complement of ORGANIC CHEMISTRY either IL-23 or its receptor can
tification of specific markers in chromosomes in both mitosis hold aberrant cutaneous TH17
sediment layers that capture the and meiosis. In somatic cells
Activating pyridine’s responses in check, providing
influence of increased human undergoing mitosis, this task 3-position a durable therapeutic benefit
population on the planet. In a is performed by centrosomes. Numerous pharmaceutical to patients with TH17-mediated
Perspective, Waters and Turner By contrast, meiosis does not compounds contain aromatic inflammatory skin diseases such
discuss 12 proposed sites that involve centrosomes in many heterocycle components such as psoriasis. —IRW
could be used as the reference animal species, but the spe- as pyridine. Chemists therefore Sci. Immunol. 7, eabq3254 (2022).
for defining the beginning of the cific methods of organizing the prize reactions that selectively
Anthropocene. Cores from these microtubules differ between append substituents to particu-
sites contain various markers animals. In particular, the mech- lar sites on these rings. Boyle et
that allow precise dating to fix anism of spindle organization al. report that upon opening pyri-
the onset of the Anthropocene. in human oocytes has not been dine rings to form linear imines,
These candidate sites will be previously understood. Wu et al. the distinct characteristics of
voted on in late 2022. If a site detected a protein structure that these intermediates strongly
is eventually selected, it will be they named the human oocyte favored halogenation at the car-
used to help focus studies on microtubule organizing center bon in the otherwise unreactive

743-B 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118ISIO_16181047.indd 741-B 11/14/22 4:18 PM

Publish your research in the Science family of journals
The Science family of journals (Science, Science Advances, Science Immunology, Science
Robotics, Science Signaling, and Science Translational Medicine) are among the most highly-
regarded journals in the world for quality and selectivity. Our peer-reviewed journals are
committed to publishing cutting-edge research, incisive scientific commentary, and insights
on what’s important to the scientific world at the highest standards.

Submit your research today!

Learn more at [Link]/journals
RES EARCH

◥ tral regulator of organ growth should fulfil three

RESEARCH ARTICLE SUMMARY requirements: (i) normal growth does not start
or terminate properly without its input, (ii) its
DEVELOPMENTAL BIOLOGY activity correlates with the spatial and tempo-

Hippo signaling instructs ectopic

ral pattern of cell proliferation, and (iii) when
hyperactivated, it produces larger but otherwise

but not normal organ growth

normal organs. We thus investigated whether
the Hippo pathway fits these criteria for an
organ growth control mechanism.
W. Kowalczyk†, L. Romanelli†, M. Atkins, H. Hillen, C. Bravo González-Blas, J. Jacobs, J. Xie,
S. Soheily, E. Verboven, I. M. Moya, S. Verhulst, M. de Waegeneer, L. Sansores-Garcia, RESULTS: To address these hypotheses, we
L. van Huffel, R. L. Johnson, L. A. van Grunsven, S. Aerts, G. Halder* studied two commonly used models for organ
growth control, the Drosophila eye and the
mouse liver, in which we examined the effects
INTRODUCTION: Accurate growth control is fun- tors Yap/Taz/Yki, can drive cell proliferation of loss of Hippo signaling output on growth,
damental to ensure proper organ size and and cause substantial organ overgrowth and its pattern of activity during normal develop-
animal health. Many signaling pathways can tumor development. ment, and the genetic program induced by the
regulate cell proliferation and apoptosis, which experimental hyperactivation of Yap/Taz/Yki.
in turn determine the size of an organ. How- RATIONALE: The Hippo pathway integrates di- Our data show that Hippo pathway output
ever, mechanisms that can sense when a de- verse biochemical and mechanical cell-cell and is not required for the developmental prolifer-
veloping organ has reached its proper size and cell–extracellular matrix signals to regulate ation of liver precursor and Drosophila retinal
instruct the cessation of further growth remain gene expression through Yap/Taz/Yki. In a cells. Furthermore, the transcriptional activity
elusive. Such mechanisms may also orchestrate current model of growth control, the activity of of Yap/Taz/Yki did not correlate with cell pro-
regeneration and induce tumorigenesis when Yap/Taz/Yki directs the right amount of cell liferation during development, and their ex-
defective. In this context, the Hippo signaling proliferation during development such that perimental hyperactivation in hepatocytes and
pathway has attracted much attention, because they are active when organs grow and in- Drosophila retinal cells, which caused over-
experimental hyperactivation of its down- activated when they stop growing. We reasoned growth, did not reactivate progenitor programs.
stream effectors, the transcriptional coactiva- that a signaling pathway functioning as a cen- Rather, Yap/Taz/Yki were active and required
in specific cell types, namely Drosophila squa-
Cell-cell & ECM-cell Hippo signaling in organ growth mous peripodial epithelial cells and mouse
signals cholangiocytes, and their hyperactivation in-
Previous model Experimental evidence
duced aberrant gene expression profiles that
Hippo-Mst1/2 Hippo partly resembled the normal programs of these
signaling cells. Finally, a functional screen in Drosophila
Warts-Lats1/2 pathway identified several Hippo pathway target genes
that were required for ectopic overgrowth but
Yki-Yap/Taz not for normal growth.
Hippo Developmental time Developmental time
Sd-Tead1-4 output
Organ size Organ growth Hippo output Cell type specification CONCLUSION: Our data show that Hippo sig-
naling does not instruct normal organ growth,
Role of Hippo output in Gene programs Histology and that the classic Hippo mutant overgrowth
organ growth control? phenotypes represent Yap/Taz/Yki gain-of-
function situations that involve ectopic ex-
1 Pattern of activity? pression of genes that are normally expressed
Single cell Genetic screens in Hippo-dependent cell types. Cells that re-
2 Loss of function? quire endogenous Hippo pathway output, such
as squamous cells, have in common being
3 Gain of function? mechanically strained. Thus, Hippo signaling is
involved in cellular responses to mechanical
1 Endogenous Hippo output 2 Loss of Hippo output 3 Ectopic Hippo output strain and in the differentiation and homeosta-
sis of morphologically challenged cells. These
Active Yap Cell type not specified findings correct a long-standing misconcep-
tion about the role of Hippo signaling in organ
growth and reveal the need to re-evaluate our
understanding of its function in other contexts,
such as in cancer and regeneration.
▪
The list of author affiliations is available in the full article online.
Proliferating cell Proliferating cell Ectopic proliferation *Corresponding author. Email: [Link]@[Link]
These authors contributed equally to this work.
Functional analysis of the Hippo pathway in organ growth control. The regulation of the Hippo signaling
Cite this article as W. Kowalczyk et al., Science 378,
pathway has been thought to determine the amount of organ growth during development. However, analysis of the eabg3679 (2022). DOI: 10.1126/science.abg3679
endogenous activity and effects of genetic manipulations of the pathway in the developing mouse liver and
Drosophila eye show that Hippo signaling does not instruct normal organ size. In the Hippo pathway schematic READ THE FULL ARTICLE AT
(top left), Drosophila proteins are shown on the left and mouse proteins on the right, separated by a hyphen. [Link]

744 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH

◥ and others have previously reported that co-

RESEARCH ARTICLE deletion of Yap and Taz from hepatoblasts, the
embryonic liver progenitor cells that give rise
DEVELOPMENTAL BIOLOGY to hepatocytes and cholangiocytes (also known

Hippo signaling instructs ectopic

as biliary epithelial cells), using Albumin-Cre
(Alb-Cre) did not affect liver size (23, 24). How-

but not normal organ growth

ever, because Alb-Cre triggers recombination
only after embryonic day 15.5 (E15.5), when
hepatoblasts are already proliferating (29), we
W. Kowalczyk1†, L. Romanelli1†, M. Atkins1,2, H. Hillen1, C. Bravo González-Blas3, J. Jacobs3, J. Xie1, used Alfp-Cre, which is expressed earlier, be-
S. Soheily1, E. Verboven1, I. M. Moya1,4, S. Verhulst5, M. de Waegeneer3, L. Sansores-Garcia1, fore day E10.5 (30), at the onset of hepatoblast
L. van Huffel1, R. L. Johnson6, L. A. van Grunsven5, S. Aerts3, G. Halder1* specification and before hepatoblasts undergo
extensive proliferation (31). Alfp-Cre efficiently
The Hippo signaling pathway is widely considered a master regulator of organ growth because of the recombined a R26-Lox-STOP-Lox-tdTomato re-
prominent overgrowth phenotypes caused by experimental manipulation of its activity. Contrary to porter transgene starting at E10.5 (fig. S1B) and
this model, we show here that removing Hippo transcriptional output did not impair the ability of the knocked out Yap/Taz, as verified by immuno-
mouse liver and Drosophila eyes to grow to their normal size. Moreover, the transcriptional activity histochemistry (fig. S1D). Alfp-Cre;Yap/Taz dou-
of the Hippo pathway effectors Yap/Taz/Yki did not correlate with cell proliferation, and hyperactivation ble knockout mice (Alfp-Cre;Yap/TazKO) were
of these effectors induced gene expression programs that did not recapitulate normal development. born at Mendelian ratios but presented yellow
Concordantly, a functional screen in Drosophila identified several Hippo pathway target genes that ears and paws and had high aspartate trans-
were required for ectopic overgrowth but not normal growth. Thus, Hippo signaling does not instruct aminase (AST), alanine aminotransferase
normal growth, and the Hippo-induced overgrowth phenotypes are caused by the activation of abnormal (ALT), and total bilirubin serum levels, indi-
genetic programs. cating liver damage (fig. S1E). However, gross
liver morphology, liver to body weight ratio,

T
and hepatocyte size and numbers were normal
he Hippo pathway is a conserved signal expression of genes that induce cell prolifer- (Fig. 1, A to C, and fig. S1C). Rather, mutant
transduction pathway that regulates gene ation, cell survival, and tissue growth (6, 11–20). mice lacked bile ducts similar to but stronger
expression through its downstream effec- Accordingly, the central hypothesis of the Hippo than the phenotype of Yap (and Taz) knockout
tors, the transcriptional coactivators Yap pathway growth control model is that Yap/ by Alb-Cre (23, 24, 32). This was because few
and Taz [Yorkie (Yki) in Drosophila] that Taz/Yki are active and drive cell proliferation cholangiocytes were specified in Alfp-Cre;Yap/
bind to TEAD [Scalloped (Sd) in Drosophila] during the growth phase of an organ, but are TazKO embryos (Fig. 1D and fig. S1D). There-
and other transcription factors (1–5). The core then inactivated when the organ has reached fore, Yap and Taz are not required in hepato-
of the pathway comprises a kinase cascade in its proper size. Thus, in this model, the Hippo blasts for their proliferation but are required
which the Mst1/2 kinases [Hippo (Hpo) in pathway acts instructively in that it directs for their differentiation into cholangiocytes.
Drosophila] activate the Lats1/2 kinases [Warts cells when to proliferate and when to stop
(Wts) in Drosophila], which then inhibit Yap/ proliferation. Drosophila eye discs can grow without Hippo
Taz/Yki by phosphorylation (fig. S1A). Hippo This model makes three key predictions: pathway transcriptional output
pathway activity is modulated by diverse bio- (i) Hippo pathway output is required for Unlike the ability of the mouse liver to grow
chemical and mechanical cell-cell and extra- normal growth; (ii) Hippo pathway activity in the absence of Yap and Taz, deletion of
cellular matrix (ECM)–cell signals integrating is modulated during development and corre- yki in Drosophila imaginal disc cells results in
information about cell shape, cell adhesion, lates with cell proliferation; and (iii) the ge- smaller eyes and wings (21). However, although
and tissue integrity (1–5). netic program that drives overgrowth in Hippo this phenotype shows that Yki is required for
The Hippo pathway is thought to be a master mutants resembles that of normal growth. Con- imaginal disc development, such small tissue
regulator of organ growth (5–10). This model sistent with these predictions, loss of Yki results phenotypes result not only from mutations
is largely based on the observation that mu- in smaller eyes and wings (21), and hyperac- in genes that control growth, but also from
tations in Hippo pathway components can tivation of Yap/Taz/Yki can drive excessive cell mutations in genes required for basic cellular
lead to organ overgrowth in flies and mice. In proliferation (6, 12–20). However, some find- functions (i.e., housekeeping genes). There-
particular, the mouse liver and Drosophila ings do not fit the model, namely that Yap/Taz fore, the yki loss-of-function phenotype does
imaginal disc–derived structures such as eyes mutant hepatocytes and sd mutant eye disc not reveal whether Hippo signaling acts instruc-
and wings show substantial overgrowth upon cells can proliferate (22–24). We therefore tively to command when and where cells pro-
Hippo pathway deregulation. There, loss-of- re-evaluated the function of Hippo signaling liferate or permissively to simply enable cells
function mutations in the core kinases or over- during organ growth. We studied Drosophila to function properly. We therefore turned to
expression of Yap/Taz/Yki up-regulates the imaginal discs and the mouse liver, because Sd, the main transcription factor binding to Yki.
these are major developmental paradigms in Sd is required for hyperactivated Yki to cause
1
VIB Center for Cancer Biology and KU Leuven Department which the Hippo pathway and organ growth overgrowth (22, 33, 34), but unlike yki mutant
of Oncology, KU Leuven, Leuven, Belgium. 2Department of
Biological Sciences, Sam Houston State University,
have been investigated (10, 25, 26). cells, sd-null mutant cells survive and can pro-
Huntsville, TX, USA. 3VIB Center for Brain and Disease liferate (22, 33) (Fig. 1E). As previously reported,
Research and KU Leuven Department of Neurosciences, KU Yap/Taz are not required in liver progenitor this difference in phenotype is explained be-
Leuven, Leuven, Belgium. 4Facultad de Ingeniería y Ciencias cells for liver growth cause Sd acts as a transcriptional repressor in
Aplicadas, Universidad de Las Américas, Quito, Ecuador.
5
Department for Cell Biology, Faculty of Medicine and We started by analyzing the phenotype of the absence of Yki (35). Thus, yki mutant cells
Pharmacy, Vrije Universiteit Brussel, Brussel-Jette, Belgium. mouse livers in which we removed Hippo path- die because Sd represses cell survival genes
6
The University of Texas MD Anderson Cancer Center, way transcriptional output. We conditionally in the absence of Yki, but sd single or yki;sd
Houston, TX, USA.
*Corresponding author. Email: [Link]@[Link] deleted Yap and Taz during embryonic liver double mutant cells are able to proliferate
These authors contributed equally to this work. development using floxed alleles (27, 28). We because the repressive activity of Sd is removed

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 1 of 11

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

A B C βCat HNF4α DAPI D HNF4αSox9DAPI Sox9

Liver to body weight ratio

control
control

control
n.s.
8% n.s.

n.s. n.s.
4%

Yap/Taz KO
Yap/Taz KO

Yap/Taz KO
2% control
Yap/Taz KO
0%
P5 P10 2m 6m

E F H ctrl sd - Yki3SA
Yki3SA
sd -
GFPDAPI GFPDAPI
3
2
1
0
-1
hpo - hpo - sd - ctrl GFP+ -2
-3

Yki3SA GFP+
n.s.
sd - GFP+ G
ctrl sd - ****
100%
rel. mutant area

80% n.s.
60%
40%
20%
0%
ctrl sd - Yki3SA Yki3SA
yki - yki - sd - Yki3SA sd - GFP+ sd -

I ctrl sd - Yki3SA Yki3SA sd - J Yki3SA vs. ctrl K Yki3SA sd - vs. sd - L sd - vs. ctrl
6 6 6
PC2: 10% of variance

10
log2 fold change

3 3 3
0
0 0 0
-10

-20 -3 -3 -3
-30
-6 -6 -6
-60 -40 -20 0 20 0 5 10 15 0 5 10 15 0 5 10 15
PC1: 72% of variance log2 mean expression log2 mean expression log2 mean expression

Fig. 1. Loss of Hippo signaling output does not affect developmental images and quantification of clone size of eye imaginal discs with clones of cells
growth. (A) Control and Aflp-Cre;Yap/Taz KO whole liver pictures from adult of the indicated genotypes marked by GFP expression. (H) Heatmap showing
mice (6 months old; scale bars, 1 cm). (B) Average liver to body weight ratios relative expression levels of genes differentially expressed between Yki3SA
from postnatal day 5 to adulthood (N = 3). (C and D) Liver sections of indicated and control eye discs (adjusted P < 0.05; |log2FC|>1) of the indicated genotypes.
genotypes stained for nuclei [4′,6-diamidino-2-phenylindole (DAPI), blue], (I) Principal components analysis plot showing all samples of indicated
adherens junctions [b-catenin, green in (C)], hepatocytes [HNF4a, red in (C), genotypes. (J to L) MA plots for the indicated comparisons; up-regulated genes
green in (D)] and cholangiocytes [Sox9, red in (D)]. (E) Lateral views of mosaic (adjusted P < 0.05; log2FC > 1) are shown in red, and down-regulated genes
adult Drosophila eyes with mutant clones of the indicated genotypes; mutant (adjusted P < 0.05; log2FC < –1) are shown in blue. ****P < 0.0001; n.s.,
cells are marked by the absence of red pigment. (F and G) Immunofluorescent not significant. Error bars indicate SD.

(Fig. 1E, bottom). Therefore, sd mutants provide deletion of Sd reduces the ability of Yki to the fate of wing disc cells in addition to its
a tool to eliminate Hippo pathway output with- drive gene expression and thus Hippo path- role in the Hippo pathway (38). We induced
out killing the cells. However, because Yki can way transcriptional output. mutant clones early in eye development, be-
also interact with other transcription factors To this end, we studied eye imaginal discs fore the major growth phase, using the eyeless-
(36, 37), we first determined to what extent rather than wing discs, because Sd controls Flipase (ey-Flp)/FRT MARCM recombination

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 2 of 11

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

system, and inspected phenotypes in adult flies peripodial epithelium (Fig. 2, B and C), and strongly enriched in early (proliferating) hep-
or third instar discs, in which most of the eye knockdown of sd specifically in peripodial atoblasts and in a small cluster of proliferating
disc cell proliferation has already occurred. As cells with C311-Gal4 produced eye discs with- cholangiocytes (Fig. 3C and fig. S5A). By con-
reported previously (22, 33, 34), deletion of sd out peripodial epithelium but with folded trast, the Yap signature was not enriched in
suppressed the overgrowth caused by loss of retinal fields and pharate adults with severe hepatoblasts but was enriched in all cholangio-
hpo (Fig. 1E) and the immense overgrowth of defects in head eversion (40), typical for defects cytes (Fig. 3D and fig. S5, B to D). Yap activity
cell clones that overexpressed Yki3SA, a con- in peripodium function (Fig. 2, G to I). The was comparable in cycling and noncycling
stitutively active form of Yki that has all three lack of peripodial cells in C311>sdRNAi eye discs cholangiocytes, and a single-cell scatter plot
Warts phosphorylation sites mutated to ala- was caused by defects in cell proliferation and further showed a general lack of correlation
nine (39) (Fig. 1, F and G). At the level of the survival rather than a transformation into between Yap signature expression and cell pro-
transcriptome, deleting sd completely inhib- retinal cells, because coexpression of the anti- liferation (Fig. 3E and fig. S5C). The pattern of
ited the ability of Yki3SA overexpression to apoptotic protein p35 partially rescued the Yap signature expression was highly robust
cause changes in gene expression (Fig. 1, H to loss of the peripodial cells, which, however, and driven by most of the Yap signature genes
K; fig. S2, A to C; and table S1). Thus, although produced a clump of morphologically abnor- (fig. S5, B, D, and E). Therefore, the pattern of
Yki3SA overexpression caused differential ex- mal cells (fig. S3B). Yap activity during liver development mir-
pression of 1632 genes in wild-type third- At the transcriptional level, the expression rors the phenotype of Yap/Taz mutants in
instar eye discs [log2-fold change (log2FC) > 1; profile of eye discs with sd mutant clones was cholangiocyte specification and cements that
adjusted P < 0.05] (Fig. 1J and fig. S2B), these not significantly different from that of control Yap/Taz drive cholangiocyte development but
effects were fully suppressed when Yki3SA was discs (Fig. 1L). Thus, loss of Sd did not affect do not control hepatoblast proliferation and
overexpressed in sd mutant clones (Fig. 1K and the expression of cell cycle genes or the basal thus liver growth.
fig. S2, A to C). Such double mutant discs dif- expression levels of Hippo target genes. Cor-
ferentially expressed only the rosy gene, which respondingly, loss of Sd reverted the massive Yki activity during eye disc development
is an eye color gene that does not affect growth. increase in Yki target gene expression caused We also compared the pattern of Yki activity
Therefore, Yki hyperactivation has no effect on by Yki3SA overexpression back to wild-type and cell proliferation during Drosophila imagi-
gene expression and cell proliferation anymore levels in sd;Yki3SA double mutants (fig. S2, A nal disc development. To measure Yki activ-
when Sd is not present. Therefore, these results to C). We did not observe effects on peripodial ity, we generated a Yki target gene signature
show that deletion of sd abolishes the tran- gene expression because these discs did not by assembling a list of high-quality Sd-binding
scriptional output of the Hippo pathway and have sd clones in the peripodial epithelium regions in the genome by overlapping Sd chro-
provides a model with which to study imaginal (Fig. 2, B and C, and materials and methods). matin immunoprecipitation-sequencing (ChIP-
disc growth control in the absence of Hippo Altogether, Sd, and thus transcriptional out- seq) peaks and Sd-binding motifs] and then
output. put from the Hippo pathway, is not measur- selecting nearby genes that were up-regulated
We then analyzed sd mutant phenotypes in ably required for the growth of the eye disc by Yki activation in eye and wing discs (fig. S6,
more detail in third-instar eye discs because proper but is required for the development of A and B, and materials and methods). The
they allow visualization of cell proliferation, the peripodial epithelium. resulting gene signature contained 52 genes,
tissue patterning, and cellular differentiation. including the Yki target genes ex, kibra, and
Third-instar eye discs are flat, sac-like epithe- Yap/Taz activity during liver development wts (table S2), and the signature’s expression
lial structures in which one side, the disc proper, The second prediction of the model was that level correlated with the strength of Yki activ-
is a columnar epithelium (of ~40,000 cells) that Hippo activity is modulated during develop- ity in different Hippo pathway mutants (fig. S7,
gives rise to the adult eye, antenna, and head ment such that Yap/Taz/Yki activity correlates A and B). Therefore, the Yki signature reports
cuticle, whereas the other side, the peripodial with cell proliferation. To test this, we mea- Yki activity in imaginal discs.
epithelium, comprises large squamous cells sured the activity of Yap/Taz by analyzing the We then analyzed the expression of the Yki
(~2000) that are required for disc eversion expression pattern of a stringent Yap/Taz tar- signature using an RNA-seq dataset of discs at
during metamorphosis (Fig. 2A) (40–42). sd- get gene signature (for simplicity “Yap signa- different time points during development span-
null mutant clones in the disc proper had nor- ture”) of 22 genes (44). The Yap signature was ning from mid-second-instar [72 hours after
mal numbers and distribution of cells in the strongly up-regulated in response to Yap hyper- egg laying (AEL)] to late-third-instar (120 hours
S phase and grew to normal size and without activation in adult hepatocytes, confirming that AEL) just before puparium formation. The
developmental delay (Fig. 2, B to D). They it reports Yap activity (fig. S4, A to E). In ad- expression of >40% of the Yki signature genes
also had normal patterning and neuronal spe- dition, we mapped the signature onto an ex- increased at later time points, but, overall, the
cification (Fig. 2D). Similarly, knockdown of sd tensive single-cell RNA-sequencing (RNA-seq) expression patterns were not uniform and
specifically in the growing retinal field by atlas of the adult mouse liver that included poorly correlated with time (fig. S7C and table
expressing upstream activator sequence RNA parenchymal cells, endothelial cells, fibroblasts, S3), suggesting that Yki activity is not signif-
interference (UAS-RNAi) transgenes using immune cells, and other cell types (45). The Yap icantly down-regulated toward the end of disc
the optix-Gal4 driver did not reduce eye size signature was preferentially expressed in chol- growth. However, a potential down-regulation
or affect morphology (Fig. 2E and fig. S3A), angiocytes, endothelial cells, and fibroblasts, of Yki activity may have been obscured, because
nor did it increase the variation in eye size which are precisely the cell types in the liver cells in the posterior eye disc had already started
between the left and right eyes of individual that require Yap/Taz for their development differentiating while more anterior cells were
flies, a measure of fluctuating asymmetry that (fig. S4, F and G) (46). We then mapped the still proliferating (Fig. 2D, ctrl, and fig. S7D),
evaluates the robustness of growth control (43) Yap signature and a canonical cell cycle signa- thereby never providing a stage at which all
(Fig. 2F). optix>sdRNAi knockdown was effec- ture onto a single-cell RNA-seq dataset of em- cells synchronously ceased proliferation. To cir-
tive because coexpressing sd-RNAi reversed the bryonic liver development that encompasses the cumvent this problem, we fed larvae a special
small eye phenotype of yki knockdown (Fig. stages when hepatoblasts proliferate and start to diet that lacks precursor molecules neces-
2E). Thus, loss of sd did not detectably affect the differentiate into hepatocytes and cholangio- sary for the synthesis of the molting hormone
development of the disc proper. By contrast, cytes (E10.5 to E17.5) (Fig. 3, A and B) (47). As ecdysone, which is required for pupation, thus
sd mutant clones were not recovered in the expected, expression of the cell cycle genes was keeping animals in the larval stage (48). The

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 3 of 11

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

A Eye B Disc proper Peripodium C Disc Peri-

Peripodium proper podium
lateral

ctrl GFP+

Percentage GFP+ area

Head vertex 40%
Antenna
A1
Peripod 30%
A2 Disc
medial
A3 proper
20%

sd - GFP +
10%
PRs, Inter-
Eye progenitors ommatidial
0%
AMF MF PMF ctrl sd - ctrl sd -

D GFP EdU EdU GFP ELAV Ato Ato ELAV E optix>w RNAi optix>sd RNAi
ctrl GFP -

optix>yki RNAi optix>ykiRNAi sd RNAi

sd - GFP -
hpo - GFP -

F 3.0%

Asymmetry
2.0%
hpo - sd - GFP -

1.0%

0.0%
ctrl sd -

G C311>wRNAi H C311>sdRNAi I
C311>wRNAi ELAVDAPI
ELAVDAPI ELAVDAPI

C311>sdRNAi ELAVDAPI

Fig. 2. Sd is required for the development of the peripodial epithelium but not Atonal (Ato; blue) and anti-ELAV (red) staining illustrates neuronal patterning and
for the disc proper. (A) Top view and cross section schematic representation of photoreceptor differentiation. (E) Lateral views of adult Drosophila eyes of the
a wild-type third-instar eye antennal disc showing the location of its major cell types. indicated genotypes. (F) Quantification of asymmetry (difference in the number
PR, photoreceptor; MF, morphogenetic furrow; AMF, anterior to the MF; PMF, of ommatidia between the two eyes of the same animal) in control and sd– mutant
posterior to the MF. (B and C) Representative images and size quantification of retinae (N = 10, P = 0.25). (G and H) Lateral views of an adult head and a pupa
control and sd– mutant clones in disc proper and peripodium. Clones are marked lacking a head and confocal images of eye imaginal discs of the indicated genotypes
by GFP (green) (N = 10). (D) Eye imaginal discs with mutant clones of the stained with DAPI (blue) and anti-ELAV (red) to label nuclei and photoreceptors,
indicated genotypes marked by the absence of GFP (green) expression. EdU (red) respectively. (I) Cross-section view of the eye imaginal discs shown in (G) and (H).
incorporation labels cells in the S phase. Details of eye imaginal discs with anti- Arrow and arrowheads point to the peripodium and disc proper, respectively.

imaginal discs in these larvae grew until their with discs at the same developmental stage that down-regulated in growth-completed discs
terminal size was reached but did not proceed terminated proliferation after reaching their from 11-day-old nonpupating larvae (Fig. 3,
with differentiation because this requires a proper size. Indeed, 5-ethynyl-2′-deoxyuridine F and G, and fig. S8, A to D). However, of the
pulse of ecdysone. This method thus allowed (EdU) incorporation was nearly absent, and Yki signature genes, only eight (15%) were
us to directly compare actively growing discs RNA-seq showed that cell cycle genes were down-regulated and >40% (23) were actually

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 4 of 11

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

A Development B Cell type C Cell cycle D Yap signature

E10.5
HB early 1.5
E11.5 1.0 1
E12.5
HB late 0.5 0
Hep 0
E13.5 -0.5 -1
Chol
E14.5
E15.5
E17.5

E F G Cell cycle H Yki signature

Corr. Yap - cell cycle EdU DAPI Growing Compl log2 FC Growing Compl log2 FC
vih 0.5
0.6 mars Tie

Growing
Jupiter 0 1.5
ald 0.5 brk
0.3 serp 1
msps -0.5
Nek2 Cht6
1 0
-1 0 0.5
Cell cycle

scra REPTOR
-0.3 Bub1 -1.5 ftz-f1 0
CycE Atf3
0 -0.6 Psf2 -2
-0.5 -0.5
Klp61F dMyc
Mcm5
glu NetB
-1 Cdc6 boi
EdU DAPI sub sog
ex
fzy
wts

Completed
-1 0 1 Cks30A
Yap signature CycB cher
geminin kibra
Top2 Imp
Mcm6 jbug
HB early Hep tld
CDC45L Chd64
HB late Chol dally
Fen1
RnrL chinmo
Mcm7
PCNA rgn
Claspin fng

I Eye disc cell fates J Cell cycle Yki signature

Antenna A3 EdU DAPI KibraZ DAPI
Antenna A2
Antenna A1
Disc proper
Head vertex
Peripodium lat
Peripodium med
Progenitors
Morpho Furrow
Interommatidia
Photorec early
Photorec late
UMAP

K Yki signature in wing discs L Open chromatin 0.50 0.50

regions 0.25 0.25
0.6 0.00 0.00
0.4 -0.25 -0.25

0.2
EdU DAPI KibraZ DAPI
0.0 Sd motif
Peripodium

-0.2
-0.4
353 1175
um
ge

n
ad

gi
iu
in

ar
od

bl
H

m
rip

g
in

g
Pe

in
W

Fig. 3. Yap/Taz/Yki activity does not correlate with cell proliferation. (I) UMAP plot of single-cell RNA-seq data with 2596 cells colored by assigned
(A and B) Uniform manifold approximation and projection (UMAP) plot of cell type together with a schematic visualization of an eye imaginal disc. lat,
single-cell RNA-seq data from embryonic murine livers colored by developmental lateral; med, medial; morpho, morphogenetic; photorec, photoreceptor. (J) Left
time point (A) and cell type (B). HB, hepatoblast; Hep, hepatocyte; Chol, column: cell proliferation in an eye disc shown by EdU incorporation (yellow) in
cholangiocyte. (C and D) Average expression of cell cycle genes (C) and Yap the disc proper (top) and peripodium (bottom) and average expression of cell
signature genes (D). (E) Scatter plot showing the correlation between Yap cycle genes on the UMAP (middle). Right column: Yki activity shown by kibra-lacZ
signature and cell cycle gene expression in single cells colored by cell identity. staining (green) in the disc proper (top) and peripodium (bottom) and average
(F) Actively growing (day 5 AEL) and growth-completed (day 11 AEL) eye expression of Yki signature genes on the UMAP (middle). The arrows point
imaginal discs from animals feed a pupation-preventing diet, stained for nuclei to corresponding groups of cells. (K) Violin plot showing average expression of
(DAPI, blue) and EdU (yellow) marking cells in S-phase. (G and H) Heatmaps Yki signature genes in single cells of the wing imaginal disc grouped by cell type.
showing relative expression and log2FCs (green sidebar marks statistically (L) Venn diagram showing the proportion of all chromatin regions specifically
significant changes; adjusted P < 0.05) in cell cycle genes (G) and Yki signature accessible in peripodial cells overlapping with those that also show enrichment
genes (H) between actively growing and growth completed eye imaginal discs. for Sd-binding motifs.

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 5 of 11

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

up-regulated in growth-completed discs (Fig. binding regions were accessible in peripodial We then investigated to what extent the
3H; fig. S8, E and F; and table S3). Moreover, cells, and 25% were exclusively accessible in Yaphyper program resembled that of embryonic
knockdown of yki reduced the expression of peripodial cells (fig. S11G). In addition, >20% liver development. We mapped the Yaphyper
the Yki reporter gene ex-lacZ in growing as of all peripodial-specific accessible chromatin program onto the single-cell RNA-seq atlas
well as growth-completed discs (fig. S9). These regions were bound by Sd (Fig. 3L). Therefore, of liver development (47). The average expres-
results show that Yki is still active in growth- peripodial specific enhancers are strongly sion of Yaphyper program genes was highest in
completed discs and not inactivated when enriched for Sd binding, and Sd-bound en- proliferating cholangiocytes, followed by non-
growth ceases. This was also observed using hancers are preferentially accessible in peri- cycling cholangiocytes and early (proliferating)
a similar approach in wing discs (49). podial cells. Altogether, the data presented hepatoblasts (Fig. 4B). However, this pattern
In a second approach, we separated cell in this section show that the activity of Yki- was a composite of the expression patterns
proliferation from differentiation by using a Sd does not correlate with the rate of cell of four gene clusters with distinct expression
single-cell gene expression atlas of third instar proliferation, but does correlate with the de- patterns (Fig. 4C). The first cluster of 247 genes
eye discs (50). This atlas reproduces the switch velopment of peripodial cells, which is con- (13% of the Yaphyper program) contained genes
from proliferating precursor cells in the an- sistent with the loss-of-function phenotype of that were expressed in all proliferating cells,
terior to differentiating photoreceptor and sd mutants. mostly early hepatoblasts and a few prolif-
interommatidial support cells in the posterior erating cholangiocytes. The second cluster of
of the eye disc (Fig. 3I). The cell cycle signature Yap hyperactivation induces an ectopic 312 genes (17% of the Yaphyper program) was
was preferentially expressed in proliferating program in hepatocytes expressed specifically in all cholangiocytes
antennal and retinal precursor cells and in The third prediction of the model positing that regardless of the cell cycle. The other two clus-
cells in the second mitotic wave, which is con- the Hippo pathway is a master regulator of ters had uniform or overall low expression in
sistent with the proliferation pattern observed organ growth is that the overgrowth induced embryonic liver cells. Thus, although the expres-
by EdU incorporation (Fig. 3J, left, and fig. by Yap/Taz/Yki hyperactivation recapitulates sion of Yaphyper genes in hepatoblasts showed
S10A). By contrast, expression of the Yki signa- growth during normal development. To test that Yap induces parts of the hepatoblast pro-
ture was highest in peripodial cells and showed this prediction, we compared the genetic program, this was not because Yap induced pro-
no correlation with cell proliferation (Fig. 3J, gram active during Yap/Taz/Yki–driven over- genitor cell characteristics but rather because
right, and fig. S10, B to E). Likewise, the Yki growth with that of normal development. We it induced general cell cycle genes. Indeed,
signature was also preferentially expressed studied how ectopic Yap activation causes the Yaphyper program contained <5% of the
in peripodial cells of wing discs (Fig. 3K and liver overgrowth in mice. We conditionally hepatoblast specific gene set (fig. S12K). Con-
fig. S11, A to F). activated Yap in hepatocytes of adult mice versely, Yap activated genes that are specif-
To further scrutinize our genomics findings, using two different methods to ensure the ically expressed in embryonic cholangiocytes.
we examined Yki activity in situ. Yki intra- highest robustness of our results. We deleted However, the Yaphyper program represented
cellular localization, assayed by a functional the Lats1/2 kinases by injecting AAV8-Cre into only ~15% of all cholangiocyte markers (fig.
Yki–green fluorescent protein (Yki-GFP) ge- Lats1/2 double-floxed mice (Lats1/2KO) and S12K). It follows that Yap hyperactivation in
nomic knock-in construct, showed a tendency overexpressed YAP by feeding doxycycline to adult hepatocytes did not simply activate the
for enhanced nuclear localization in peripodial mice that expressed the tetracycline-activated genetic program of embryonic hepatoblasts
cells (fig. S11H), but this assay was not very sen- transactivator (rtTA) under the hepatocyte- but did induce a general cell proliferation pro-
sitive (51). Conversely, the expression of ex-lacZ specific ApoE promoter and human YAP from gram and partially transdifferentiated hepato-
was high in peripodial cells, but it showed a rtTA-responsive TetO promoter (hereafter cytes toward cholangiocyte fate.
Sd independent basal expression in disc proper Apo>YAP) (6). Adult hepatocytes are normally Another possibility was that Yap hyperacti-
cells (22), which prevented its use to compare quiescent, but Yap activation triggered cell vation induced a program that is used during
Yki activity in peripodial and disc proper cells cycle re-entry and caused the liver to more than liver regeneration, when hepatocytes re-enter
(fig. S11I). As previously observed and shown double its size within 2 weeks in both mutants, the cell cycle. To test this, we performed RNA-
in our RNA-seq data, kibra expression strongly as previously described (fig. S12A) (6). One week seq on regenerating hepatocytes that we pu-
correlated with Yki activity (fig. S11J) (52). We after Yap activation, purified Lats1/2KO and rified 48 hours after acute injury by injection
thus monitored kibra expression with a lacZ Apo>YAP hepatocytes had massive changes in of the liver toxin CCl4, when hepatocyte pro-
enhancer trap insertion into the kibra gene. gene expression that were highly correlated liferation peaks. Regenerating hepatocytes
kibra-lacZ expression was strongly induced (R = 0.84; Fig. 4A and fig. S12, B to F). The up-regulated 1021 genes, of which more than
upon Yki activation by wts knockdown (fig. S11, overlap comprised 1819 up-regulated and 372 half were also present in the Yaphyper program
K and L), indicating that it is regulated by Yki. down-regulated genes (log2FC > 1, adjusted (fig. S14A). Thus, Yap hyperactivation indeed
In wild-type animals, kibra-lacZ expression P < 0.05; table S4). Functions of up-regulated induced a large portion of the liver regener-
was absent in disc proper cells but high in genes (hereafter called the Yaphyper program) ation program (Fig. 4, D and G, and fig. S13A).
peripodial cells, even when cell proliferation were enriched for cell cycle regulation (113 genes, However, endogenous Yap/Taz were not re-
stopped after disc growth was complete (Fig. 6%), ECM regulation, cell motility and angio- quired for hepatocyte proliferation (23) or
3J, right, and fig. S11M). Thus, kibra-lacZ is a genesis (306 genes, 17%), and immune func- the induction of the regeneration program
new Yki-Sd activity reporter that reports positions (417 genes, 23%), (fig. S12, G and H), after CCl4 injection (Fig. 4, E and F; fig. S13,
tive Yki-Sd output, and its expression pattern whereas the down-regulated genes were en- B to E; and table S5). This conundrum can be
confirmed the predominantly peripodial activ- riched for hepatocyte metabolic functions (fig. resolved because Yap hyperactivation and
ity of Yki-Sd. S12, I and J). The Yaphyper program is different the resulting overgrowth caused liver dam-
The preferential activity of Yki-Sd in peri- from the Yap signature in that it included all age, as revealed by elevated serum levels of
podial cells was also observed at the chroma- 1819 genes that were either directly or indirectly AST (fig. S13G), which would then induce the
tin level. Analysis of our single-cell assay for up-regulated in hepatocytes in response to Yap regeneration program as a secondary effect.
transposase-accessible chromatin sequencing hyperactivation, whereas the Yap signature is Therefore, the Yaphyper program comprises
(ATAC-seq) dataset of accessible chromatin in a very select set of 22 genes that are broadly direct effects caused by Yap hyperactivation
eye disc cells (50) showed that >90% of the Sd- activated in response to Yap. and indirect effects triggered by liver damage.

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 6 of 11

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

A UP-regulated C HB early HB late Chol Hep D Yap vs. CCl4

Cell Type
Cell Cycle

log2FC ctrl+CCl4
Apo>YAP Lats1/2 KO Time 6 R=0.62

Cell Cycle
1455 1819 171 3
12
0
Yaphyper Cell Type
10
HB early -3

Chol
DOWN-regulated 8 HB late -6
Chol -5 0 5
Apo>YAP Lats1/2 KO 6 Hep log2FC Apo>YAP
1714 372 44
4 Cell Cycle Cell cycle Yap signature

General
S/M
2 G1

B
0 Time
E10.5
E CCl4 ctrl vs. CCl4 Y/T KO
UP in Yaphyper
E11.5
E12.5 6

log2FC CCl4 Y/T KO

0.3 R=0.82
0.0 E13.5
3

Not expressed
-0.3 E14.5
-0.6
E15.5
0
E17.5
Cluster -3
Cell Cycle
Chol -6
General -6 -3 0 3 6
Not expr log2FC CCl4 ctrl

F G CCl4 Y/T KO H I Yap-CCl4 K

CCl4 Y/T KO

Lats1/2 KO
Lats1/2 KO

log2FC
Yap- Yap-
Apo>YAP
Apo>YAP

1.0 0.0
CCl4 ctrl
CCl4 ctrl

spec CCl4
100% 0.5 -0.2
log2FC
0.0 -0.4
6 8
GO terms -0.5 -0.6
Ccn2 5 6
Myof 4
Axl 3 4
Ankrd1 75% Cell cycle
2 2
Ccn1 ECM
1
Tgfb2 0
0
Yap-hyper

Signaling
Amotl2
Ptpn14 -1 -2 Angiogenesis
J 1.2 L
Crim 50% 0.00
Ccdc80
-4 Cell movement 0.8

Yap-CCl4
Asap1 -6 Chemotaxis 0.4 -0.25
Nuak2
CCl4-Yap

Rbms3 Multiple 0.0

Arhgef17 -0.50
Igfbp3 25% -0.4
F3
Gadd45a -1 0 1 -1 0 1
Lats2
CCl4

Dock5
Cell cycle Yap signature
Nt5e 0% HB early HB late Hep Chol

M log10 CPM N Extracellular matrix O ECM cluster gene expression

Extracellular structure
4 Cell migration
ECM

Angiogenesis Fibroblasts
3 Cluster 144
Vascular development
ECM Response to protozoan ECM Cluster
Imm

2 10 55 76
Immune Leukocyte activation
Cell activation immune 41 18
1 Signaling 24
T-cell activation
Signaling

36 336
212 21
0 Mononuclear cell diff.
Detection bacteria
Detection other organisms 161

Regulation of peptidase Cholangiocytes

T
Hep
Chol
HPC
Fibro
Endo
Kupff
Mono
Neutr
Baso
cDC1
cDC2
mcDC
pDC
B
NK
ILC1

49
8 Interleukin-1b production Endothelial cells

Fig. 4. Yap hyperactivation drives an ectopic growth program in hepatocytes. (G) log2FC of genes (rows) induced by Yap hyperactivation or CCl4 injection for
(A) Overlap between the up-regulated (top) and down-regulated (bottom) genes comparisons (columns) between the indicated conditions and their respective controls.
in hepatocytes with YAP overexpression and Lats1/2 KO. (B) Average expression of (H) Barplot showing the percentage of genes in the Yap-specific or Yap-CCl4
Yaphyper program genes in the developing liver. (C) Clustering of Yaphyper program subsets of the Yaphyper program that belong to the given Gene Ontology (GO) term
genes by their normal expression pattern during liver development. Cells (columns) are categories. (I to L) Average expression of the Yap-CCl4 (I) and Yap-specific (K) parts of
grouped by cell type, cell cycle phase, and developmental time; genes (rows) are the Yaphyper program in single cells of the embryonic liver and its correlation with
grouped by similar expression pattern. (D) Correlation of changes in gene expression the expression of cell cycle (J) and Yap signature (L) genes, respectively. (M and
between Apo>YAP and CCl4 (compared with their relative controls). Cell cycle and Yap N) Subclustering of the Yap-specific genes by their expression in different cell types of
signature genes are marked in green and red, respectively, and their distribution is the adult liver. (M) Heatmap showing the absolute expression [log10 counts per million
shown by the density plot on the right side. (E) Correlation of changes in hepatocyte (CPM)] in single cells aggregated by cell type. (N) Treeplot showing the top five GO
gene expression induced by CCl4 injection into control (ctrl) and Yap/Taz KO terms enriched for each cluster. Dot size corresponds to the number of genes, and dot
animals. (F) log2FCs of Yap signature genes for comparisons between the indicated color reflects the proportion of genes in each cluster. (O) Overlap between genes in
conditions and their respective controls; values are rounded to the nearest integer. the ECM cluster and marker genes of fibroblasts, cholangiocytes, and endothelial cells.

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 7 of 11

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

Therefore, we divided it into a general re- replication, whereas the down-regulated genes broadly and reflected the pattern of cell pro-
generation part that comprised the genes were involved in retinal cell differentiation liferation (Fig. 5, G, H, K, and L). Conversely,
induced by both Yap and CCl4 (Yap-CCl4) and (fig. S15D). This phenotype raised the possi- the Ykihyper program resembled abnormal
a “Yap-specific” part that comprised the genes bility that hyperactivation of Yki during eye growth, here defined as genes up-regulated in
that were induced by Yap but not by CCl 4 development trapped cells in an early devel- RasV12G–driven neoplastic tumors (RasV12G
(Fig. 4G and table S6). Genes in the Yap-CCl4 opmental stage, thereby producing an excess scribRNAi tumors) (55). Such tumors have hy-
set predominantly function in the cell cycle and of precursor cells. However, three lines of evi- peractive Yki (54), and their gene expression
immune system, are the most active in early dence refute this hypothesis. profile significantly overlapped with the Ykihyper
hepatoblasts and proliferating cholangiocytes, First, we compared the Ykihyper program, program and was enriched in peripodial cells
and their expression is poorly correlated with defined as the 222 genes overlapping between (Fig. 5, I to K, and fig. S17, A to G). Altogether,
endogenous Yap activity (Fig. 4, H to J). This the Yki3SA and wts– mutant profiles, with that ectopic Yki activation did not simply stimulate
set thus constitutes the regenerative response of wild-type eye discs from different develop- extra normal growth, but partially reprog-
to the liver damage caused by Yap-induced mental stages (53). Forty-four percent (n = 98) rammed disc proper cells toward a peripodial
liver overgrowth. By contrast, genes in the Yap- of the Ykihyper program genes were expressed cell phenotype (Fig. 5M and fig. S17H).
specific set were normally expressed in embry- at higher levels at early stages compared with Next, we investigated which Yki-Sd target
onic cholangiocytes (Fig. 4, K and L), as well as later stages of eye development (Fig. 5, B and genes were required for Yki-driven overgrowth.
in endothelial cells, fibroblasts, and cholangio- C). However, these genes represented <4% of To this end, we performed a genetic screen in
cytes of adult livers (fig. S14). These are the the genes that characterize early eye develop- which we assessed the role of the 322 predicted
cells with the highest Yap activity and they ment (Fig. 5B and fig. S15E). Therefore, Yki Sd target genes plus an additional 143 low
require Yap/Taz for their development (46). hyperactivation did not simply extend early confidence targets (table S10 and materials
To understand the biological functions of development. and methods). We knocked down each of
the Yap-specific genes, we classified them ac- Second, mapping the Ykihyper program onto these genes in eye discs that overexpressed
cording to their expression patterns in normal the single-cell expression atlas of third-instar Yki1SA under the control of the GMR-Gal4
adult liver. A first cluster of genes (ECM cluster; eye discs showed that Yki hyper cells down- driver (GMR>Yki1SA), which causes massively
Fig. 4, M and N) was expressed in endothelial regulated genes normally expressed in differ- overgrown adult eyes (22). We thus crossed
cells, fibroblasts, and cholangiocytes and was entiating photoreceptor cells, consistent with GMR>Yki 1SA flies to a collection of 1673 fly
strongly enriched for functions in angiogenesis their lack of retinal differentiation (Fig. 5D). stocks that each expressed an RNAi construct
and ECM. A second cluster was predominantly The up-regulated genes, however, were not en- targeting one of the candidate genes under the
expressed in immune cells and enriched for riched for genes expressed in progenitor cells control of a Gal4 responsive UAS-promoter
immune functions (immune cluster; Fig. 4, but rather for genes expressed in peripodial (on average, 3.6 different RNAi lines per
M and N). A third cluster had uniform but low cells (Fig. 5, E and F). Expression in peripodial gene). Knockdown of 44 genes suppressed
expression in normal livers (signaling cluster; cells was the predominant pattern, with few of GMR>Yki1SA overgrowth, whereas knockdown
Fig. 4, M and N). The ECM cluster comprised the Ykihyper program genes being broadly or of 40 enhanced overgrowth, and knockdown
>50% of the endothelial cell specific mark- lowly expressed in normal eye discs (Fig. 5F). of the remaining 381 did not noticeably affect
ers and 33% and 18% of the fibroblast and Because ey-Flp induces clones not only in the the adult eye phenotype (Fig. 5N and table
cholangiocyte-specific markers, respectively, disc proper but also in peripodial cells, this S10). We then screened all 465 genes for re-
with several genes expressed in all three cell result may simply be the consequence of hy- quirement during normal growth using the
types (Fig. 4O). Altogether, these data show peractivating Yki in peripodial cells causing optix-Gal4 driver, which is active in disc proper
that Yap hyperactivation in adult hepatocytes their overproliferation. However, driving Yki3SA cells during eye development. Here, knock-
does not cause liver overgrowth because it ac- overexpression specifically in disc proper cells down of 70 genes reduced eye size, whereas
tivates a normal progenitor (hepatoblast) cell by optix-Gal4 caused essentially the same knockdown of eight resulted in larger eyes
program. Rather, Yap hyperactivation ectopi- phenotype and triggered the induction of (Fig. 5N and table S10). Overlapping the results
cally induces parts of the genetic programs of peripodial cell–specific genes (fig. S15, F and of the two screens showed that 25 (57%) of the
endothelial cells, cholangiocytes, and fibro- G). Likewise, in wing discs, wts mutant clones 44 Yki1SA-overgrowth suppressors were also
blasts, as well as other genes that are not nor- and Yki overexpression induced a genetic pro- required for normal growth, but knockdown
mally expressed in hepatocytes. gram that was most similar to that of peripodial of the other 19 (43%) suppressors did not
cells (fig. S16). However, Yki induced less than affect normal growth (Fig. 5N). Although RNAi
Ectopic Yki activates an abnormal genetic half of the peripodial cell markers (fig. S15, H knockdown can be incomplete and may not
program in imaginal discs and I). Therefore, Yki activation did not sim- reveal full loss-of-function phenotypes, our
Next, we investigated how ectopic Yki drives ply cause an expansion of retinal progenitor results indicate that the GMR>Yki1SA pheno-
tissue overgrowth in Drosophila imaginal discs. cells and inhibit their differentiation, but ec- type is preferentially sensitive to knockdown
As we did for the liver, we overexpressed the topically activated a portion of the peripodial of these genes. Thus, these results further
constitutively active Yki3SA and deleted the cell program. support that Yki-driven overgrowth does not
Lats1/2 homolog wts in clones of cells. Both Third, we compared the Ykihyper program simply recapitulate normal growth.
conditions produced eye discs that were nearly with a gene set representative of normal growth Among the Yki1SA overgrowth suppressors
entirely composed of mutant cells and lacked that we generated as the difference in gene ex- that were also required for normal growth
ommatidial patterning and photoreceptor dif- pression between actively growing and growth were RNAi lines targeting Yki as expected;
ferentiation (Fig. 1F). Discs had substantial completed eye discs collected from larvae fed epidermal growth factor receptor (EGFR) and
changes in gene expression that were highly the ecdysone precursor–deficient food (fig. its ligands, Vn and Spi; the cell cycle regulators
correlated between the two mutants (Fig. 5A; S15J). There was minimal overlap between E2f1, CyclinE, and Stg; the cell growth pro-
fig. S15, A to C; and table S8). Consistent with the Ykihyper and normal growth programs moter dMyc; and components of Notch signal-
the histological phenotype, induced genes were and, unlike the Ykihyper program, the normal ing, Dl, Gp150, Uif, and E(spl)malpha (Fig. 5O).
mainly involved in ribosome biogenesis (in- growth program was not specifically expressed Overexpression of these genes can cause ec-
dicating accelerated cell growth) and DNA in peripodial cells, but rather was expressed topic cell proliferation, although overexpression

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 8 of 11

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

A wts- vs. Yki3SA B Ykihyper during development D Down in Ykihyper F Ykihyper in cell types log10 CPM
R=0.75 Time 3
72h 2.5
96h 2
120h
5 1.5

log2FC wts-
11days 0.2
0.0
1
0 0.5
4 0
-5
-2.5 0.0 2.5 5.0 2
E Up in Ykihyper 0.4
log2FC Yki3SA 0.2
0.0
C up at 72h 0 -0.2

Ykihyper

Peripod med
Head vertex
0.50

Antenna A3
Antenna A2
Antenna A1

Morpho Fur
Progenitors

Photo early
Peripod lat
1948

Photo late
-2 0.25

Interom
124 98 0.00

G Normal growth H 0.75 Antenna A3

I RasG12V-scribRNAi tumors J
R=0.71 R=0.86
Antenna A2 0.4
0.50 Antenna A1
Cell cycle

Ykihyper
Head vertex
0.25 0.2
Peripodium lateral
Peripodium medial
0.00 Progenitors 0.2 0.0
0.4 0.1
Morphogenetic Furrow
0.2 -0.25 0.0 -0.2
Interommatidial
0.0 Photoreceptors early
-0.2 0.0 0.2 0.4 Photoreceptors late
-0.2 0.0 0.2
UP in normal growth UP in tumor

K Ykihyper
100
M Correlations between genetic programs N Genetic screen for growth suppressors
th
5 108
Knockdown of 465 genes
ow

m
up in 9
e

iu
gr

r
l

pe
normal
yc

od
al

or
up in tumor i hy
lc

rip
rm

m
el

394
Pe
No

Tu
C

546 1 44 suppress 70 suppress

GMR>Yki1SA normal growth
Cell cycle 100 71 15 6 2 0.8 overgrowth 19 25 45
L 0.6

O
R=0.24
0.4
Completed growth

Normal growth 71 100 58 35 32

5.0
0.2 Pathway GMR>Yki1SA suppressors (general, )
2.5
0 Hippo Yki, Sd
Peripodium 15 58 100 78 68
Cell Cycle CycE, E2F1, Stg, DNA2, TTC28, dMyc
0.0
EGFR/RTK EGFR, Spi, Vn, Pdk1
Notch Dl, Gp150, Uif, E(spl)ma
-2.5 Ykihyper 6 35 78 100 86
Gene expr. Osa, Heph, Pan3, pAbp, RtcB, Shep, CG6026
-2.5 0.0 2.5 5.0
Cytoskeleton Drak, Chd64, Arpc5, Klar, RhoGAP18B, PHACTR1, Mlc1
Yki3SA vs. ctrl
ECM Dlp, Ccn, Mtgo, Spn55B, CG11357
Yki signature Cell cycle Tumor 2 32 68 86 100 Other Bor, NetB, IP3K2, ELVOL1, CAHbeta, CG31689
Other Unknown CG9733, CG15646, CG43759

Fig. 5. Yki hyperactivation driven overgrowth resembles tumor but not of the genes up-regulated in actively growing compared with growth-completed
normal growth. (A) Correlation of changes in gene expression between Yki3SA eye discs (G) and genes up-regulated in RasG12V scribRNAi tumors (I) and its
overexpression and wts KO (compared with their respective controls). Ykihyper correlation with the expression of cell cycle (H) and Yki signature (J) genes,
program genes are shown in green. (B) Relative expression levels of the Ykihyper respectively. (K) Overlap between the Ykihyper program and genes up-regulated in
program genes at different time points of larval growth (72, 96, and 120 hours actively growing compared with growth-completed eye discs and genes up-regulated
AEL) and after growth completion (11 days AEL). (C) Overlap between the Ykihyper in RasG12V scribRNAi tumors (adjusted P < 0.05, log2FC > 1). (L) Correlation of
program and the genes up-regulated (adjusted P < 0.05, log2FC > 1) at 72 hours log2FCs in gene expression between Yki3SA overexpression relative to control and
AEL (compared with 120 hours AEL, which is the time point of Ykihyper samples growth completed discs relative to growing discs. Cell cycle and Yki signature
and their controls). (D and E) Average expression of the overlap of the genes genes are marked in blue and red, respectively, and their distribution along the
down-regulated in Yki3SA overexpression and wts KO (D) and the genes in Ykihyper axes is shown by the density plots. (M) Correlation between the average
program (E) in the single cells of imaginal eye discs. (F) Expression of the Ykihyper expression of the indicated gene sets in single cells of the eye disc. “Peripodium”
program in eye disc cell types. Heatmap (top) shows the absolute expression refers to the marker genes (adjusted P < 0.05; log2FC > 0.5) of all peripodial
(log10 CPM) in single cells aggregated by cell type. Violin plot (bottom) shows the cells. (N) Schematic representation of the results of the in vivo RNAi screens.
average expression in single cells grouped by cell type. (G to J) Average expression (O) Suppressors of Yki-induced overgrowth listed in nine functional categories.

of any of them alone does not phenocopy the general growth gene) and several genes en- and actin regulator 1 (PHACTR1, CG32264), the
amount of overgrowth seen in Hippo mutants coding components functioning in cytoskele- endopeptidase Spn55B, Miles to go (Mtgo, a
(25). Conversely, the 17 Yki1SA overgrowthÐ ton and ECM regulation such as RhoGAP18B fibronectin type III domain containing pro-
specific suppressors included Sd (confirming (a GTPase activating protein for Rho GTPases), tein), and Netrin B (NetB, a secreted ligand
that it is an overgrowth-specific gene, not a Myosin alkali light chain 1 (Mlc1), phosphatase regulating axon outgrowth). Thus, ectopic Yki

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 9 of 11

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

causes eye disc overgrowth through mecha- a stiff cell culture substrate (60) and mediate to re-evaluate our understanding of Yap/Taz’s
nisms that include effects on the cytoskeleton a cell’s response to mechanical strain such as function as oncogenes and in regeneration.
and on cell-cell and cell-ECM interactions. the reinforcement of the cytoskeleton and cell Investigating Yap/Taz as drivers of a response
adhesion. A major function of Yap/Taz/Yki-Sd to mechanical strain will identify new mech-
Discussion thus appears to be a homeostatic function in anisms that promote cancer cell phenotypes
Here, we evaluated the role of the Hippo path- morphologically and mechanically challenged and initiate regenerative responses.
way in organ growth control. We examined fly cells. Consistent with this, we found that Yap
eye imaginal discs and the mouse liver, two and Yki hyperactivation induced factors in- Methods summary
commonly used model systems to study organ volved in cytoskeleton and ECM regulation A full description of the methods can be found
growth. Our data show that Hippo pathway and activated genetic programs that were most in the supplementary materials. This includes
output is not required for the proliferation of similar to the programs of cells in which Yap/ full mouse and Drosophila genotypes and treat-
liver precursor cells and disc proper cells, the Taz/Yki-Sd are active and required for their ment protocols, descriptions of sample pro-
main parenchymal cell types whose prolifera- development. We thus hypothesize that ectopic cessing for immunofluorescence staining and
tion embodies the growth of the liver and eye activation of these programs changes how cells RNA-seq, statistical tests, and bioinformatics
discs. We conclude that the Hippo pathway interact with their neighbors and how they analyses.
does not instruct cells when to proliferate and receive and interpret growth control signals.
REFERENCES AND NOTES
when to stop proliferation during the normal In this way, hyperactivation of Hippo pathway
1. Z. Wu, K. L. Guan, Hippo signaling in embryogenesis and
development of these organs. output can cause ectopic cell proliferation and development. Trends Biochem. Sci. 46, 51–63 (2021).
How can this conclusion be consistent with organ overgrowth even though it is not di- doi: 10.1016/[Link].2020.08.008; pmid: 32928629
the fact that loss of the Hippo core kinases rectly involved in normal growth control. 2. Y. Zheng, D. Pan, The Hippo signaling pathway in development
and disease. Dev. Cell 50, 264–282 (2019). doi: 10.1016/
causes severe overgrowth of liver and imaginal Although Hippo output does not play a role [Link].2019.06.003; pmid: 31386861
discs, thus demonstrating that proper activity in growth control in liver and eye discs, it is 3. I. M. Moya, G. Halder, Hippo-YAP/TAZ signalling in organ
of the kinases is required for normal organ required for the growth of other organs. In regeneration and regenerative medicine. Nat. Rev. Mol. Cell
Biol. 20, 211–226 (2019). doi: 10.1038/s41580-018-0086-y;
growth? This seeming paradox is resolved flies, modulation of Yki is required to fine-tune pmid: 30546055
when considering that the Hippo kinase mu- the growth of the presumptive wing tissue, the 4. J. R. Davis, N. Tapon, Hippo signalling during development.
tant overgrowth phenotypes are the result of wing pouch of the wing disc, where Yki is Development 146, dev167106 (2019). doi: 10.1242/dev.167106;
pmid: 31527062
ectopic Yap/Taz/Yki activation and therefore activated in peripheral cells because of the
5. G. Halder, R. L. Johnson, Hippo signaling: Growth control and
represent gain-of-function phenotypes. Yap/ mechanical strain exerted by the growth of beyond. Development 138, 9–22 (2011). doi: 10.1242/
Taz and Yki-Sd are normally not active in cells in the center (61). In the mouse, Yap/Taz dev.045500; pmid: 21138973
hepatocytes and disc proper cells, but their are required for the growth of the embryonic 6. J. Dong et al., Elucidation of a universal size-control
mechanism in Drosophila and mammals. Cell 130, 1120–1133
ectopic activation can induce the overexpres- heart, whereas hyperactivation of Yap causes (2007). doi: 10.1016/[Link].2007.07.019; pmid: 17889654
sion of a battery of genes that then cause over- heart enlargement and can trigger cardiomyo- 7. D. Rogulja, C. Rauskolb, K. D. Irvine, Morphogen control of wing
growth. This means that the role of the Hippo cyte proliferation in adults (62). In these tis- growth through the Fat signaling pathway. Dev. Cell 15,
309–321 (2008). doi: 10.1016/[Link].2008.06.003;
core kinases in hepatocytes and disc proper sues, Yap/Taz/Yki appear to be activated by pmid: 18694569
cells is to keep Yap/Taz/Yki in check to prevent mechanical stress (e.g., heartbeat) and their 8. T. Aegerter-Wilmsen et al., Integrating force-sensing
them from triggering abnormal overgrowth. activity may affect tissue growth by regulating and signaling pathways in a model for the regulation of wing
imaginal disc size. Development 139, 3221–3231 (2012).
It follows that Hippo signaling is permissive the mechanical properties of the cells. Thus, doi: 10.1242/dev.082800; pmid: 22833127
but not instructive for organ growth control the function of Yap/Taz/Yki in cardiomyocytes 9. K. D. Irvine, K. F. Harvey, Control of organ growth by patterning
during the normal development of Drosophila and wing pouch cells may be comparable to and hippo signaling in Drosophila. Cold Spring Harb. Perspect.
Biol. 7, a019224 (2015). doi: 10.1101/cshperspect.a019224;
eye discs and mouse liver. that in cholangiocytes, peripodial cells, and pmid: 26032720
So, what do Yap/Taz/Yki normally do and other mechanically challenged cells and does 10. I. K. Hariharan, Organ size control: Lessons from Drosophila.
how does ectopic Yap/Taz/Yki induce organ not reflect a role of the Hippo pathway as a Dev. Cell 34, 255–265 (2015). doi: 10.1016/j.
devcel.2015.07.012; pmid: 26267393
overgrowth? Yap/Taz and Yki-Sd are tran- master regulator of organ growth. Similarly,
11. M. Kango-Singh et al., Shar-pei mediates cell proliferation arrest
scriptionally active in and required for the Yap/Taz/Yki can be activated by mechanical during imaginal disc growth in Drosophila. Development 129,
development of cholangiocytes and peripodial strain after injury and then promote tissue 5719–5730 (2002). doi: 10.1242/dev.00168; pmid: 12421711
cells, respectively. Others have reported that repair and regeneration (2, 3). In the adult fly 12. N. Tapon et al., salvador Promotes both cell cycle exit
and apoptosis in Drosophila and is mutated in human cancer
Yap/Taz/Yki-Sd are also required in endothe- gut, for example, Yki is not required for nor- cell lines. Cell 110, 467–478 (2002). doi: 10.1016/
lial cells, activated fibroblasts, lung AT1 cells, mal homeostasis, but injury activates Yki in S0092-8674(02)00824-3; pmid: 12202036
the squamous retinal pigment epithelial cells differentiated enterocytes, which then pro- 13. K. F. Harvey, C. M. Pfleger, I. K. Hariharan, The Drosophila Mst
ortholog, hippo, restricts growth and cell proliferation and
of zebrafish, and the squamous follicular cells of duce multiple cytokines and EGFR ligands promotes apoptosis. Cell 114, 457–467 (2003). doi: 10.1016/
Drosophila egg chambers (51, 55–58). These that induce regenerative proliferation of intes- S0092-8674(03)00557-9; pmid: 12941274
cells have in common that they are either flat tinal stem cells (63). Thus, also in these cases, 14. S. Wu, J. Huang, J. Dong, D. Pan, hippo encodes a Ste-20
family protein kinase that restricts cell proliferation and
cells that are under mechanical or morpho- Yki function is permissive but not instructive promotes apoptosis in conjunction with salvador and warts. Cell
logical stress or that they have extra strong for organ growth, and its effect on growth in- 114, 445–456 (2003). doi: 10.1016/S0092-8674(03)00549-X;
cell-ECM adhesion. Consistent with this theme, volves a specialized response that is not essen- pmid: 12941273
15. R. S. Udan, M. Kango-Singh, R. Nolo, C. Tao, G. Halder, Hippo
when we analyzed expression of the Yki-Sd tial for normal development. promotes proliferation arrest and apoptosis in the Salvador/
reporter kibra-lacZ, we found strong expres- Our findings correct a long-standing mis- Warts pathway. Nat. Cell Biol. 5, 914–920 (2003).
sion in glial cells, in particular, in the two carpet conception that the Hippo pathway is a master doi: 10.1038/ncb1050; pmid: 14502294
16. D. Zhou et al., Mst1 and Mst2 maintain hepatocyte quiescence
cells of developing eye discs, two huge, flat cells regulator of organ growth. Hippo signaling as
and suppress hepatocellular carcinoma development through
that each covers half of the eye disc (59). There- a master regulator of growth control has pro- inactivation of the Yap1 oncogene. Cancer Cell 16, 425–438
fore, there is a strong correlation between vided a straightforward and convenient expla- (2009). doi: 10.1016/[Link].2009.09.026; pmid: 19878874
cell morphology and Hippo pathway activity. nation for its function in regeneration and 17. H. Song et al., Mammalian Mst1 and Mst2 kinases play
essential roles in organ size control and tumor suppression.
Indeed, Yap/Taz are activated by mechanical cancer. Our new model for how Hippo sig- Proc. Natl. Acad. Sci. U.S.A. 107, 1431–1436 (2010).
stress, for example, when cells stretch out on naling affects organ growth reveals the need doi: 10.1073/pnas.0911409107; pmid: 20080598

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 10 of 11

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

18. F. D. Camargo et al., YAP1 increases organ size and expands 38. G. Halder et al., The Vestigial and Scalloped proteins act 679–688 (2015). doi: 10.1016/[Link].2015.04.011;
undifferentiated progenitor cells. Curr. Biol. 17, 2054–2060 together to directly regulate wing-specific gene expression in pmid: 25908270
(2007). doi: 10.1016/[Link].2007.10.039; pmid: 17980593 Drosophila. Genes Dev. 12, 3900–3909 (1998). doi: 10.1101/ 57. J. B. Miesfeld et al., Yap and Taz regulate retinal pigment
19. L. Lu et al., Hippo signaling is a potent in vivo growth and gad.12.24.3900; pmid: 9869643 epithelial cell fate. Development 142, 3021–3032 (2015).
tumor suppressor pathway in the mammalian liver. Proc. Natl. 39. H. Oh, K. D. Irvine, In vivo analysis of Yorkie phosphorylation pmid: 26209646
Acad. Sci. U.S.A. 107, 1437–1442 (2010). doi: 10.1073/ sites. Oncogene 28, 1916–1927 (2009). doi: 10.1038/ 58. L. B. Nantie et al., Lats1/2 inactivation reveals Hippo function
pnas.0911427107; pmid: 20080689 onc.2009.43; pmid: 19330023 in alveolar type I cell differentiation during lung transition
20. J. H. Lee et al., A crucial role of WW45 in developing epithelial 40. T. Zhang, Q. Zhou, F. Pignoni, Yki/YAP, Sd/TEAD and Hth/ to air breathing. Development 145, 10.1242/dev.163105 (2018).
tissues in the mouse. EMBO J. 27, 1231–1242 (2008). MEIS control tissue specification in the Drosophila eye disc pmid: 30305289
doi: 10.1038/emboj.2008.63; pmid: 18369314 epithelium. PLOS ONE 6, e22278 (2011). doi: 10.1371/ 59. M. Silies et al., Glial cell migration in the eye disc. J. Neurosci. 27,
21. J. Huang, S. Wu, J. Barrera, K. Matthews, D. Pan, The Hippo [Link].0022278; pmid: 21811580 13130–13139 (2007). doi: 10.1523/JNEUROSCI.3583-07.2007;
signaling pathway coordinately regulates cell proliferation and 41. M. C. Gibson, G. Schubiger, Peripodial cells regulate pmid: 18045907
apoptosis by inactivating Yorkie, the Drosophila Homolog of proliferation and patterning of Drosophila imaginal discs. Cell 60. S. Dupont et al., Role of YAP/TAZ in mechanotransduction.
YAP. Cell 122, 421–434 (2005). doi: 10.1016/ 103, 343–350 (2000). doi: 10.1016/S0092-8674(00)00125-2; Nature 474, 179–183 (2011). doi: 10.1038/nature10137;
[Link].2005.06.007; pmid: 16096061 pmid: 11057906 pmid: 21654799
22. S. Wu, Y. Liu, Y. Zheng, J. Dong, D. Pan, The TEAD/TEF family 42. K. D. McClure, G. Schubiger, Developmental analysis and 61. C. Rauskolb, S. Sun, G. Sun, Y. Pan, K. D. Irvine, Cytoskeletal
protein Scalloped mediates transcriptional output of the squamous morphogenesis of the peripodial epithelium tension inhibits Hippo signaling through an Ajuba-Warts
Hippo growth-regulatory pathway. Dev. Cell 14, 388–398 in Drosophila imaginal discs. Development 132, 5033–5042 complex. Cell 158, 143–156 (2014). doi: 10.1016/
(2008). doi: 10.1016/[Link].2008.01.007; pmid: 18258486 (2005). doi: 10.1242/dev.02092; pmid: 16236766 [Link].2014.05.035; pmid: 24995985
23. E. Verboven et al., Regeneration defects in Yap and Taz mutant 43. D. M. Vallejo, S. Juarez-Carreño, J. Bolivar, J. Morante, 62. S. Liu, J. F. Martin, The regulation and function of the Hippo
mouse livers are caused by bile duct disruption and M. Dominguez, A brain circuit that synchronizes growth and pathway in heart regeneration. Wiley Interdiscip. Rev. Dev. Biol.
cholestasis. Gastroenterology 160, 847–862 (2021). maturation revealed through Dilp8 binding to Lgr3. Science 8, e335 (2019). doi: 10.1002/wdev.335; pmid: 30169913
doi: 10.1053/[Link].2020.10.035; pmid: 33127392 350, aac6767 (2015). doi: 10.1126/science.aac6767; 63. P. Zhang, B. A. Edgar, Insect gut regeneration. Cold Spring
24. L. Lu, M. J. Finegold, R. L. Johnson, Hippo pathway coactivators pmid: 26429885 Harb. Perspect. Biol. 14, a040915 (2022). doi: 10.1101/
Yap and Taz are required to coordinate mammalian liver 44. Y. Wang et al., The Cancer Genome Atlas Research Network, cshperspect.a040915; pmid: 34312250
regeneration. Exp. Mol. Med. 50, e423 (2018). pmid: 29303509 F. Camargo, H. Liang, Comprehensive molecular characterization
25. R. M. Neto-Silva, B. S. Wells, L. A. Johnston, Mechanisms of of the Hippo signaling pathway in cancer. Cell Rep. 25, AC KNOWLED GME NTS
growth and homeostasis in the Drosophila wing. Annu. Rev. Cell 1304–1317.e5 (2018). doi: 10.1016/[Link].2018.10.001; We thank D. Pan for providing the sd, hpo, and yki mutant fly
Dev. Biol. 25, 197–220 (2009). doi: 10.1146/annurev. pmid: 30380420 stocks and ApoE>hYAP mice; K. D. Irvine for the yki3SA stock;
cellbio.24.110707.175242; pmid: 19575645 45. M. Guilliams et al., Spatial proteogenomics reveals distinct and H. Bellen for the gp-anti-Ato antibody; the Transgenic RNAi
26. B. Z. Stanger, The biology of organ size determination. evolutionarily conserved hepatic macrophage niches. Cell 185, Project (TRiP) at Harvard Medical School, the Vienna Drosophila
Diabetes Obes. Metab. 10 (Suppl 4), 16–22 (2008). 379–396.e38 (2022). doi: 10.1016/[Link].2021.12.018; Resource Center, and the Bloomington Drosophila Stock Center
doi: 10.1111/j.1463-1326.2008.00938.x; pmid: 18834429 pmid: 35021063 for providing transgenic RNAi and other fly stocks; and N. Alves
27. M. Xin et al., Hippo pathway effector Yap promotes cardiac 46. S. Manmadhan, U. Ehmer, Hippo signaling in the live – A long Souza Carvalhais, E. Kesikiadou, and J. Mach for help with
regeneration. Proc. Natl. Acad. Sci. U.S.A. 110, 13839–13844 and ever-expanding story. Front. Cell Dev. Biol. 7, 33 (2019). experiments. Funding: This work was supported by The Research
(2013). doi: 10.1073/pnas.1313192110; pmid: 23918388 doi: 10.3389/fcell.2019.00033; pmid: 30931304 Foundation Flanders (FWO Odysseus type I grant and project
28. M. Xin et al., Regulation of insulin-like growth factor signaling 47. L. Yang et al., A single-cell transcriptomic analysis reveals grants G.0954.16N and G.030616N to G.H. and an FWO
by Yap governs cardiomyocyte proliferation and embryonic precise pathways and regulatory mechanisms underlying doctoral fellowship to W.K.) and by the Cancer Prevention and
heart size. Sci. Signal. 4, ra70 (2011). doi: 10.1126/ hepatoblast differentiation. Hepatology 66, 1387–1401 (2017). Research Institute of Texas (grant RP200240 to R.J.). Author
scisignal.2002278; pmid: 22028467 doi: 10.1002/hep.29353; pmid: 28681484 contributions: G.H. conceived the project. G.H., W.K., L.R.,
29. C. M. Weisend, J. A. Kundert, E. S. Suvorova, J. R. Prigge, 48. T. Katsuyama, R. Paro, Innate immune cells are dispensable I.M.M., and M.A. wrote the manuscript. G.H., W.K., L.R., and M.A.
E. E. Schmidt, Cre activity in fetal albCre mouse hepatocytes: for regenerative growth of imaginal discs. Mech. Dev. 130, designed the experiments and analyzed the data. W.K. and
Utility for developmental studies. Genesis 47, 789–792 (2009). 112–121 (2013). doi: 10.1016/[Link].2012.11.005; L.R. organized and performed most of the experiments. M.A.
pmid: 19830819 pmid: 23238120 contributed to Drosophila experiments. W.K., C.B.G.-B., J.J.,
30. C. Kellendonk, C. Opherk, K. Anlag, G. Schütz, F. Tronche, 49. K. Strassburger, M. Lutz, S. Müller, A. A. Teleman, Ecdysone M.d.W., and S.A. performed the bioinformatics analyses. H.H.,
Hepatocyte-specific expression of Cre recombinase. Genesis regulates Drosophila wing disc size via a TORC1 dependent I.M.M., E.V., S.S., S.V., L.S.-G., L.v.H., R.J., and L.A.v.G. contributed
26, 151–153 (2000). doi: 10.1002/(SICI)1526-968X(200002) mechanism. Nat. Commun. 12, 6684 (2021). doi: 10.1038/ to the mouse experiments. Competing interests: The authors
26:2<151:AID-GENE17>[Link];2-E; pmid: 10686615 s41467-021-26780-0; pmid: 34795214 declare no competing interests. Data and materials availability:
31. M. Gordillo, T. Evans, V. Gouon-Evans, Orchestrating liver 50. C. Bravo González-Blas et al., Identification of genomic All sequencing data generated in this study have been deposited
development. Development 142, 2094–2108 (2015). enhancers through spatial integration of single-cell in the Gene Expression Omnibus (GEO) under the accession
doi: 10.1242/dev.114215; pmid: 26081571 transcriptomics and epigenomics. Mol. Syst. Biol. 16, e9438 number GSE211461 with the following subseries: GSE211457
32. N. Zhang et al., The Merlin/NF2 tumor suppressor functions (2020). doi: 10.15252/msb.20209438; pmid: 32431014 (Drosophila – nonpupating larvae), GSE211458 (Drosophila – Yki
through the YAP oncoprotein to regulate tissue homeostasis in 51. G. C. Fletcher et al., Mechanical strain regulates the Hippo hyperactivation), GSE211459 (liver – Yap hyperactivation), and
mammals. Dev. Cell 19, 27–38 (2010). doi: 10.1016/ pathway in Drosophila. Development 145, dev.159467 (2018). GSE211460 (liver – CCl4 in Yap/Taz KO). Processed data and
[Link].2010.06.015; pmid: 20643348 doi: 10.1242/dev.159467; pmid: 29440303 analyses results are provided in the supplementary tables. License
33. L. Zhang et al., The TEAD/TEF family of transcription factor 52. A. Genevet, M. C. Wehr, R. Brain, B. J. Thompson, N. Tapon, information: Copyright © 2022 the authors, some rights
Scalloped mediates Hippo signaling in organ size control. Dev. Kibra is a regulator of the Salvador/Warts/Hippo signaling reserved; exclusive licensee American Association for the
Cell 14, 377–387 (2008). doi: 10.1016/[Link].2008.01.006; network. Dev. Cell 18, 300–308 (2010). doi: 10.1016/ Advancement of Science. No claim to original US government
pmid: 18258485 [Link].2009.12.011; pmid: 20159599 works. [Link]
34. Y. Goulev et al., SCALLOPED interacts with YORKIE, the nuclear 53. M. Torres-Oliva, J. Schneider, G. Wiegleb, F. Kaufholz, journal-article-reuse
effector of the hippo tumor-suppressor pathway in N. Posnien, Dynamic genome wide expression profiling of
Drosophila. Curr. Biol. 18, 435–441 (2008). doi: 10.1016/ Drosophila head development reveals a novel role of
[Link].2008.02.034; pmid: 18313299 Hunchback in retinal glia cell development and blood-brain SUPPLEMENTARY MATERIALS
35. L. M. Koontz et al., The Hippo effector Yorkie controls normal barrier integrity. PLOS Genet. 14, e1007180 (2018).
[Link]/doi/10.1126/science.abg3679
tissue growth by antagonizing scalloped-mediated default doi: 10.1371/[Link].1007180; pmid: 29360820
Materials and Methods
repression. Dev. Cell 25, 388–401 (2013). doi: 10.1016/ 54. M. Atkins et al., An ectopic network of transcription factors
Figs. S1 to S17
[Link].2013.04.021; pmid: 23725764 regulated by hippo signaling drives growth and invasion of a
Tables S1 to S10
36. H. Oh, K. D. Irvine, Cooperative regulation of growth by Yorkie malignant tumor model. Curr. Biol. 26, 2101–2113 (2016).
References (64–79)
and Mad through bantam. Dev. Cell 20, 109–122 (2011). doi: 10.1016/[Link].2016.06.035; pmid: 27476594
MDAR Reproducibility Checklist
doi: 10.1016/[Link].2010.12.002; pmid: 21238929 55. X. Wang et al., YAP/TAZ orchestrate VEGF signaling
37. H. W. Peng, M. Slattery, R. S. Mann, Transcription factor choice during developmental angiogenesis. Dev. Cell 42, 462–478.e7
in the Hippo signaling pathway: Homothorax and yorkie (2017). doi: 10.1016/[Link].2017.08.002;
regulation of the microRNA bantam in the progenitor domain pmid: 28867486 Submitted 31 December 2020; resubmitted 9 August 2022
of the Drosophila eye imaginal disc. Genes Dev. 23, 2307–2319 56. I. Mannaerts et al., The Hippo pathway effector YAP controls Accepted 12 October 2022
(2009). doi: 10.1101/gad.1820009; pmid: 19762509 mouse hepatic stellate cell activation. J. Hepatol. 63, 10.1126/science.abg3679

Kowalczyk et al., Science 378, eabg3679 (2022) 18 November 2022 11 of 11

An estate gift
to AAAS
Going all the way back to 1848, our founding year, the American “As a teacher and instructor, I bear responsibility for the

Association for the Advancement of Science (AAAS) has been younger generations. If you have extra resources, concentrate
them on organizations, like AAAS, that are doing work for all.”
deeply committed to advancing science, engineering and
—Prof. Elisabeth Ervin-Blankenheim, 1848 Society member
innovation around the world for the benefit of all people.

Today, we are dedicated to advocating for science and scientific

If you intend to include AAAS in your estate plans, provide this
evidence to be fully and positively integrated into public policy and information to your lawyer or financial adviser:
for the community to speak with one voice to advance science and
Legal Name: American Association for the Advancement of Science
engineering in the United States and around the world.
Federal Tax ID Number: 53-0196568

By making AAAS a beneficiary of your will, trust, retirement plan Address: 1200 New York Avenue, NW, Washington, DC 20005

or life insurance policy, you will become a member of our 1848 If you would like more information on making an estate gift to
Society and will help fuel our work on behalf of science and society AAAS, cut out and return the form below or send an email to
– including publishing the world’s most promising, innovative philanthropy@[Link]. Additional details are also available

research in the Science family of journals and engaging in the online at [Link]/1848Society.

issues that matter locally, nationally and around the world.

cut here

Yes, I would like more information about joining the AAAS 1848 Society.
PLEASE CONTACT ME AT:

Name:

Address:

City: State: Zip code: Country:

Email: Phone:

RETURN THIS FORM TO:

AAAS Office of Philanthropy and Strategic Partnerships • 1200 New York Avenue, NW • Washington, DC 20005 USA
RESEAR CH

◥ bly impaired spindle microtubule nucleation

RESEARCH ARTICLE SUMMARY and spindle assembly in human oocytes. This
structure was not detected in the oocytes of
HUMAN FERTILITY other mammalian species such as mice and pigs.
We finally identified two oocyte maturation
The mechanism of acentrosomal spindle arrest patients with compound heterozygous
mutations in the key huoMTOC component
assembly in human oocytes TACC3. All mutations disrupted the normal
function of TACC3, resulting in the absence
Tianyu Wu†, Jie Dong†, Jing Fu†, Yanping Kuang†, Biaobang Chen, Hao Gu, Yuxi Luo, Ruihuan Gu, of the huoMTOC structure and completely im-
Meiling Zhang, Wen Li, Xi Dong, Xiaoxi Sun*, Qing Sang*, Lei Wang* paired spindle assembly in the patients’ oocytes.

CONCLUSION: Our study shows that human

INTRODUCTION: Spindle assembly is essential ponents of microtubule nucleators in a cohort oocytes possess an aMTOC-like structure, the
for ensuring accurate chromosome transmission of 1394 infertile female patients characterized huoMTOC, that serves as a major site of mi-
in both meiosis and mitosis. In somatic cells, by oocyte maturation arrest. crotubule nucleation and is required for spin-
mitotic spindle assembly is mediated by dupli- dle assembly. The huoMTOC shows drastically
cated centrosomes, but canonical centrosomes RESULTS: First, we found that in human oocytes different characteristics in terms of number,
are absent in the oocytes of many species. In the nucleation of spindle microtubules is in- localization, and composition compared with
rodents, acentriolar microtubule organizing itiated from kinetochores from 2 to 4 hours aMTOCs in mouse oocytes. These findings
centers (aMTOCs) are responsible for meiotic after NEBD. We showed the process of spindle suggest that a distinct mechanism for the in-
spindle assembly, but it has long been sup- microtubules nucleating from kinetochores itiation of microtubule nucleation and spindle
posed that human oocytes lack prominent in human oocytes. We then found that there assembly has evolved in human oocytes. We
aMTOCs on the meiotic spindle, and the exact are 43 proteins localized in the meiotic spin- found that mutations in TACC3 cause defects
mechanism of acentrosomal spindle assembly dle, among which four proteins—centriolar in spindle assembly by disrupting the struc-
in human oocytes has remained unclear. coiled-coil protein 110 (CCP110), cytoskeleton- ture of the huoMTOC, which leads to clinical
associated protein 5 (CKAP5), disrupted in oocyte maturation arrest. This suggests that
RATIONALE: Microtubule nucleation and en- schizophrenia 1 (DISC1), and transforming acid- the huoMTOC might be an important bio-
suring spindle assembly are core events reg- ic coiled-coil–containing protein 3 (TACC3)— marker for evaluating the quality of human
ulating oocyte nuclear maturation. To identify exhibited both kinetochore and spindle micro- oocytes.
the potential proteins driving spindle microtubule localization. The localization of the four Our discovery of huoMTOC provides in-
tubule nucleation in human oocytes, we sys- proteins was notably different from their local- sights into the physiological mechanism of
tematically localized 86 human centrosome and ization in human mitotic cells and in mouse microtubule nucleation and spindle assembly
microtubule-related proteins by immunofluo- oocytes. Together, the four proteins formed an in human oocytes. These findings also im-
rescence or three-dimensional high-resolution unusual structure that was surrounded by mi- prove our understanding of the pathological
live cell imaging in more than 2000 human
oocytes. We then tracked the dynamic migra-
crotubules in human germinal vesicle (GV)
oocytes just before NEBD. We refer to this po-
mechanisms of oocyte maturation arrest.
▪
tion of identified microtubule nucleators at tential nucleating structure as the human oo-
different time points before and after nuclear cyte microtubule organizing center (huoMTOC). The list of author affiliations is available in the full article online.
envelope breakdown (NEBD). We further down- We found that a single huoMTOC is formed at *Corresponding author. Email: wangleiwanglei@[Link] (L.W.);
regulated corresponding proteins to confirm the cortex of human GV oocytes and migrates sangqing@[Link] (Q.S.); xiaoxi_sun@[Link] (X.S.)
These authors contributed equally to this work.
their role in microtubule nucleation and spin- to the nuclear envelope before NEBD. After
Cite this article as T. Wu et al., Science 378, eabq7361
dle assembly. Given that spindle microtubule NEBD, the huoMTOC becomes fragmented (2022). DOI: 10.1126/science.abq7361
nucleation defects result in impaired spindle and is recruited to kinetochores to initiate spin-
assembly and abnormal oocyte maturation, we dle microtubule nucleation. Down-regulation READ THE FULL ARTICLE AT
screened for mutations in genes encoding com- of huoMTOC components caused considera- [Link]

Merge Tubulin TACC3

20 µm 20 µm

The huoMTOC structure in a human oocyte. The human GV oocyte shown here was matured for ~5 hours and fixed for immunofluorescence before NEBD. The huoMTOC
(TACC3, magenta) was surrounded by numerous microtubules (green) on the nuclear envelope. The dashed square shows the magnification region. The arrow highlights the huoMTOC.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 745

RES EARCH

◥ nucleation in live human oocytes by using

RESEARCH ARTICLE three-dimensional (3D) high-resolution time-
lapse imaging (Fig. 1A). Human GV oocytes
HUMAN FERTILITY were co-injected with mRNA encoding fluo-
rescently fused histone H2B (mCherry) and
The mechanism of acentrosomal spindle centromere protein B (CENPB) (mClover3)
to visualize chromosomes and kinetochores,
assembly in human oocytes respectively. Combined with fluorescent pro-
teins, SiR-tubulin (a dye with far-red fluores-
Tianyu Wu1†, Jie Dong1†, Jing Fu2†, Yanping Kuang3†, Biaobang Chen4, Hao Gu1, Yuxi Luo1, cence) was used to label microtubules before
Ruihuan Gu2, Meiling Zhang5, Wen Li5, Xi Dong6, Xiaoxi Sun2*, Qing Sang1*, Lei Wang1* oocyte maturation (24). According to our
three-channel time-lapse images, a micro-
Meiotic spindle assembly ensures proper chromosome segregation in oocytes. However, the mechanisms tubule cluster (dashed square) derived from a
behind spindle assembly in human oocytes remain largely unknown. We used three-dimensional GV oocyte was observed proximal to chromo-
high-resolution imaging of more than 2000 human oocytes to identify a structure that we named the somes upon NEBD and disassembled in the
human oocyte microtubule organizing center (huoMTOC). The proteins TACC3, CCP110, CKAP5, and first few hours of meiosis I (Fig. 1A). Apart
DISC1 were found to be essential components of the huoMTOC. The huoMTOC arises beneath the oocyte from the microtubules (light-gray curve, bot-
cortex and migrates adjacent to the nuclear envelope before nuclear envelope breakdown (NEBD). tom panel) derived from the GV oocyte, a few
After NEBD, the huoMTOC fragments and relocates on the kinetochores to initiate microtubule nascent microtubules (dark-gray curve, bottom
nucleation and spindle assembly. Disrupting the huoMTOC led to spindle assembly defects and oocyte panel) were detected by microscopy after NEBD
maturation arrest. These results reveal a physiological mechanism of huoMTOC-regulated spindle (Fig. 1A).
assembly in human oocytes. To confirm that the nascent microtubules
were polymerized from kinetochores, we eval-

S
uated the tubulin intensity around kineto-
pindle assembly is essential for ensuring Mouse aMTOCs lack centrioles but contain chores over time after NEBD. The intensity of
accurate chromosome transmission in partial centrosomal proteins such as pericen- microtubules (tubulin) and chromosomes (H2B)
both mitosis and meiosis (1, 2). In somat- trin (14), g-tubulin (20), CEP192 (18), and NEDD1 was measured around the representative kineto-
ic cells, mitotic spindle assembly is me- (21). Upon meiotic resumption after prophase I chores (Fig. 1A, arrows). The nascent micro-
diated via duplicated centrosomes, which arrest, multiple aMTOCs initiate microtubule tubules (dark-gray curve) were hardly observed
consist of two centrioles surrounded by the nucleation around the nuclear envelope in at the beginning of meiosis in human oocytes
pericentriolar material (PCM) (3, 4). Centro- mouse germinal vesicle (GV) oocytes. After (0 to 2 hours after NEBD). With the disassem-
somes are the major microtubule organizing nuclear envelope breakdown (NEBD), aMTOCs bly of the derived microtubule cluster (light-
centers (MTOCs) in mitotic cells, and they are are clustered and then concentrated at the gray curve), the fluorescent tubulins began
responsible for microtubule nucleation and spindle poles for bipolar spindle organiza- to concentrate at the kinetochores starting at
spindle pole organization of the centrosomal tion (15, 18, 19). ~2 hours after NEBD (Fig. 1, A and B). Accord-
spindle (5, 6). A mechanistic understanding of meiotic spin- ing to three-channel time-lapse imaging, the
Unlike somatic cells, canonical centrosomes dle assembly in human oocytes, on the other fluorescence of nascent microtubules (dark-
are absent in the oocytes of many species hand, remains elusive (2). It is only known gray curve) overlapped primarily with kineto-
(7–11). Instead, acentriolar MTOCs (aMTOCs) that the spindle microtubule nucleation in chores (red curve) from 2 to 2.5 hours after
are observed in frog (12) and mouse (13–15) human oocytes is mediated by chromosomes NEBD and consistently throughout the begin-
oocytes, whereas no MTOC structures are and promoted by guanosine triphosphate (GTP)– ning of meiosis I (2 to 4 hours after NEBD)
found in the Drosophila (16) or Caenorhabditis bound Ran (RanGTP) (22). Subsequently, a (Fig. 1A). However, the fluorescence of nascent
elegans (17) oocytes. The difference suggests multipolar spindle is assembled as an interme- microtubules (dark-gray curve) did not consis-
that the mechanisms of female meiotic spindle diate, and then the spindle poles are merged tently overlap with chromosomes (blue curve),
assembly are not conserved between species. to form a bipolar spindle (22). However, hu- suggesting that the nascent microtubules pri-
The aMTOC-directed meiotic spindle assembly man oocytes lack detectable aMTOCs at the marily polymerized from the kinetochores rath-
was only elucidated in mouse oocytes (15, 18, 19). meiotic spindle poles (22, 23), and thus the er than other regions of the chromosomes
exact mechanism of acentrosomal spindle as- (Fig. 1A). Along with meiotic maturation, mi-
1
sembly in human oocytes remains unclear. crotubules were nucleated slowly but con-
Institute of Pediatrics, Children’s Hospital of Fudan
University, State Key Laboratory of Genetic Engineering, tinuously (Fig. 1, A and B), and the stable
Institutes of Biomedical Sciences, Shanghai Key Laboratory
Results microtubules with high tubulin intensity could
of Medical Epigenetics, Fudan University, Shanghai 200032, Spindle microtubules are nucleated from be observed on kinetochores starting at 3 hours
China. 2Shanghai Ji Ai Genetics and IVF Institute, Obstetrics kinetochores in human oocytes
and Gynecology Hospital, Fudan University, Shanghai
after NEBD (Fig. 1, A and B). These results
200011, China. 3Department of Assisted Reproduction, A previous study implied that spindle micro- suggest that in human oocytes, the nucleation
Shanghai Ninth People’s Hospital, Shanghai Jiao Tong tubules emanate from the kinetochores in hu- of spindle microtubules is initiated from kinet-
University School of Medicine, Shanghai 200011, China.
4
NHC Key Lab of Reproduction Regulation, Shanghai
man oocytes (22), and this phenomenon was ochores starting at 2 to 4 hours after NEBD.
Institute for Biomedical and Pharmaceutical Technologies, also observed in our immunofluorescent re- Next, to further confirm the nucleation of
Fudan University, Shanghai 200032, China. 5Center for sults in early prometaphase oocytes (fig. S1). spindle microtubules on kinetochores, we
Reproductive Medicine and Fertility Preservation Program,
This suggests that kinetochores may serve as treated live human metaphase I (MI) oocytes
International Peace Maternity and Child Health Hospital,
School of Medicine, Shanghai Jiao Tong University, Shanghai the microtubule nucleation sites in human with a reversible microtubule inhibitor (noco-
200030, China. 6Reproductive Medicine Center, Zhongshan oocytes. However, the dynamic process of how dazole) and tracked the dynamic recovery
Hospital, Fudan University, Shanghai 200032, China. spindle microtubules were originally nucleated of microtubule nucleation immediately after
*Corresponding author. Email: wangleiwanglei@[Link] (L.W.);
sangqing@[Link] (Q.S.); xiaoxi_sun@[Link] (X.S.) from kinetochores was not delineated. Thus, nocodazole was washed out (Fig. 1C). Micro-
†These authors contributed equally to this work. we first observed the process of microtubule tubules and kinetochores were marked by

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 1 of 12

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

Fig. 1. The dynamic process of A

spindle microtubules nucleating NEBD 1:00 2:00 2:30 3:00 3:30 4:00

from kinetochores in human

Histone H2B
oocytes. (A) Representative time-
lapse images showing microtubule
nucleation in human oocytes
after NEBD. Blue, chromosomes
(H2B-mCherry); gray, microtubules
(SiR-tubulin); magenta, kineto-

CENPB
chores (mClover-CENPB). The
microtubule cluster was marked at
0 to 2 hours after NEBD (dashed
squares). Representative kineto-
chores of each time point are shown
in some z-sections. The intensity of

Tubulin
chromosomes and microtubules
was measured and plotted around
representative kinetochores. The
graphs in the bottom panel are the
fluorescence profiles of H2B, tubu-
Merge
lin, and CENPB across the repre-
sentative kinetochores along the
direction of the arrows in the
merged images at each time point. 1.0 1.0
1.0 1.0 1.0 1.0 1.0
The microtubules derived from GV
Intensity (A.U.)

oocytes (light-gray curves) and

0.5 0.5 0.5 0.5 0.5 0.5 0.5
nascent microtubules (dark-gray
curves) were distinguished at
0 0 0 0 0 0 0
2 hours after NEBD. The plotted 0 1 2 3 4 0 1 2 3 4 0 1 2 3 4 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
Derived from GV oocyte Nascent Distance (µm)
intensity curves indicate the relative
location of each fluorescent protein.
B 1 .0
C Nocodazole
CENPB Tubulin Mature for 10 h 1h Wash out and imaging
Blue, chromosomes; gray, micro-
Tubulin Intensity on kinetochore (A.U.)

P<0.01
0:05 0:10 0:15 0:20
tubules; magenta, kinetochores.
2h

0 .8

A.U., arbitrary units. Scale bar,

CENPB
5 mm. (B) Representative images 0 .6
3h

showing kinetochore nucleated

microtubules at different time points. 0 .4

The intensity of the microtubules on

3.5 h

kinetochores measured in (A) was 0 .2

MAP4

compared between 2 and 4 hours after

NEBD (n > 24 kinetochores, signifi- 0 .0
4h

cance of the differences in intensity Time from NEBD (hours

was calculated by ANOVA and indi- D E F
cated on the graph). Scale bar, 2 mm. 0:05 0:10 0:15 0:20
1.5
P<0.0001
(C) Representative time-lapse P=0.002
z-section

Tubulin intensity (A.U.)

images (z-stack) showing microtubule P 0.0001

1.0
nucleation after nocodazole washout
in live human oocytes. Gray, micro-
Intensity (A.U.)

1.0 1.0 1.0 1.0

tubules (GFP-MAP4); magenta, 0.5
2- 4 hours
kinetochores (mScarlet-CENPB). Flow 0.5 0.5 0.5 0.5
Time from NEBD
diagram shows the sequence of 0.0 0.0 0.0 0.0 0.0
experiments. Arrow indicates time- 0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3
5 10 15 20 Chromosome Kinetochore Tubulin
lapse imaging initiation. Time is given Distance (µm) Time from release (min

as hours:minutes after nocodazole

washout. Scale bar, 5 mm. (D) Single-slice images of boxed areas in (C). Time is given as hours:minutes after nocodazole washout. Scale bar, 2 mm. Intensity of kinetochores
and microtubules is indicated in the graphs. The bottom graphs are the fluorescence profiles of MAP4 (gray) and CENPB (magenta) proximal to the representative kinetochores
along the direction of the yellow arrow. (E) The intensity of microtubules on kinetochores in (D) was measured and compared after nocodazole washout (significance was
calculated by ANOVA). (F) Mechanistic model for microtubule nucleation in human oocytes. Time is after NEBD.

fluorescent microtubule-associated protein polymerized microtubules were not detected on and some of them were emanating outward
4 (MAP4) and CENPB proteins, respectively. kinetochores until ∼10 min after nocodazole was (Fig. 1C). The trajectory of microtubules was
Initially, nocodazole completely disrupted the washed out. At that time, several nascent mi- also shown in a single slice (Fig. 1D), and the
microtubules in human MI oocytes, and re- crotubules were observed near the kinetochores, representative kinetochore was nucleating

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 2 of 12

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

microtubules slowly and continuously, suggest- cells (21, 25, 26), but their localization in chore and spindle microtubule localization
ing that nascent microtubules were nucleated human oocytes was largely unknown. These (Fig. 2B), and such localization was consistent
from kinetochores (Fig. 1, D and E). Together, proteins included 34 centrosomal proteins, until the MII stage (fig. S4), which was notably
these results show the dynamic process of 25 microtubule-associated proteins, 12 dynein- different from their localization in human mi-
acentrosomal spindle microtubules nucleating related proteins, five regulatory kinases or totic cells and in mouse oocytes (6, 21, 27–30).
from kinetochores in human oocytes (Fig. 1F). substrates, five spindle assembly factors, and CCP110 and DISC1 are centrosomal proteins
five nuclear pore–related proteins (fig. S2). A that are involved in spindle assembly at the
Discovery of a specific microtubule nucleator in total of 36 proteins were specifically local- centrosomes of human mitotic cells and
human oocytes ized on spindle microtubules, two proteins aMTOCs of mouse oocytes (6, 21, 27). CKAP5
To identify the specific factors driving spindle were concentrated on spindle poles, and one and TACC3 are microtubule-associated pro-
microtubule nucleation from kinetochores in protein showed spindle periphery localiza- teins concentrated at centrosomes and micro-
human oocytes, we localized 86 human centro- tion (Fig. 2A and fig. S3). Unexpectedly, four tubules in human mitotic cells during mitosis
some and microtubule-related proteins by per- proteins—centriolar coiled-coil protein 110 (28–30). Importantly, both CKAP5 and TACC3
forming immunofluorescence in >1000 fixed (CCP110), cytoskeleton-associated protein 5 have been reported to nucleate microtubules
human oocytes (Fig. 2A and figs. S2 and S3). (CKAP5), disrupted in schizophrenia 1 (DISC1), in vitro and in mitotic cells, respectively (31–34).
These proteins were classified according and transforming acidic coiled-coil–containing These observations suggested that CCP110,
to their function or localization in somatic protein 3 (TACC3)—exhibited both kineto- CKAP5, DISC1, and TACC3 are potential

Fig. 2. Identification of a micro- A B

tubule nucleator in human
Merge CCP110 Tubulin Chromosome
oocytes. (A) Schematic of
the metaphase I spindle
in human oocytes (chromosomes
in blue, microtubules in green,
spindle poles in magenta, kineto-
chores in yellow, and spindle
Merge CKAP5 Tubulin Chromosome
periphery in gray). (B) Immuno-
fluorescence images of
metaphase I spindle in human
MI oocytes showing protein
localization relative to both
microtubules and chromosomes. Spindle microtubules Merge DISC1 Tubulin Chromosome
Gray, chromosomes; green,
AKAP450 AURKA BUGZ CETN2 CETN3 CENPJ
microtubules. Scale bar, 5 mm. CEP120 CEP192 CEP250 CLIP1 CLTC DCTN1
(C) Immunofluorescence images DCTN2 DLGAP5 DYNLT1 GTSE1 HAUS4 HAUS6
of the nucleus in human GV HAUS8 HMMR KANSL3 KIF11 KIF20A KIZ
oocytes before NEBD. The yellow LIS1 MYO10 NDE1 NDEL1 NEK2 NUSAP1
PCM1 PLK1 PLK4 PRC1 TPX2 TUBG1
squares indicate the newly Merge TACC3 Tubulin Chromosome
identified structure surrounded Kinetochores and spindle microtubules
by microtubules. Scale bar, CCP110 CKAP5 DISC1 TACC3
10 mm. The yellow arrows indicate
intensity measurements in the Spindle poles Spindle periphery
images at the top. The localization NUMA1 KIF2A HOOK3

of huoMTOC and microtubule C D

distribution of human GV oocytes CCP110 CKAP5 DISC1 TACC3 Merge CCP110 TACC3
are shown below. Gray, chromatin
(Hoechst); green, microtubules
(tubulin); magenta, huoMTOC.
(D) Representative images of live
human GV oocytes. mScarlet- Tubulin
Hoechst
TACC3 was used as the standard Merge CKAP5 TACC3
Live human GV oocytes

for colocalization. Green, TACC3;

magenta, CCP110, CKAP5, or
DISC1; gray, chromosomes
(Hoechst). The yellow arrows
indicate the colocalization. Scale
bar, 10 mm. 1.0 1.0 1.0 1.0 Merge DISC1 TACC3
Intensity (A.U.)

0.5 0.5 0.5 0.5

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
Distance (µm)

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 3 of 12

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

candidates for regulating microtubule nucle- drastically diminished, implying the essen- most obvious effects on the microtubule nu-
ation and polymerization in human oocytes. tial role of the huoMTOC for microtubule cleation and spindle assembly (Fig. 4, F and
Subsequently, these proteins were monitored nucleation in human oocytes (Fig. 3A). In ad- G), further suggesting that TACC3 plays a
by immunofluorescence in human GV oocytes dition, different degrees of disruption in the leading role in the huoMTOC. To directly test
just before NEBD (fig. S5 and movie S1). Nota- asymmetrical distribution of microtubules whether huoMTOC is essential for spindle as-
bly, each of the four proteins showed an un- around the nuclear envelope were also ob- sembly, the huoMTOC marked by fluorescent
usual structure surrounded by microtubules. served after specific depletion of endogenous TACC3 was disrupted by laser ablation (fig.
This structure was proximal to the nuclear en- CCP110, CKAP5, or DISC1 (Fig. 3, B and C, and S10, A and B). Similar to TACC3 depletion, the
velope in human GV oocytes ∼0 to 2 hours fig. S8). Of note, depletion of TACC3 resulted spindle microtubule polymerization and spin-
before NEBD (Fig. 2C), which was consistent in the most severe disruption of microtubule dle assembly were significantly impaired by
with our observations of a microtubule clus- asymmetrical distribution in most of the ana- the laser ablation of huoMTOC (P = 0.015,
ter proximate to chromatin in human oocytes lyzed human GV oocytes (Fig. 3, B and C). Fisher’s exact test) (fig. S10, C and D). It has
after NEBD (movie S2). To test whether these These results indicate that the huoMTOC is been demonstrated that RanGTP is also re-
four proteins were colocalized, we overex- essential for the asymmetrical nucleation of quired for spindle assembly in human oocytes
pressed TACC3 (labeled with mScarlet) and the microtubules around the nuclear envelope (22). Disruption of both TACC3 and RanGTP
other three proteins (labeled with mClover3). and that all four proteins are indispens- aggravated the spindle microtubule polymer-
As indicated in Fig. 2D, CCP110, CKAP5, and able for maintaining normal function of the ization defects (fig. S11), indicating combined
DISC1 were all colocalized with TACC3, imply- huoMTOC, in which TACC3 presumably plays effects of TACC3 and RanGTP on microtubule
ing that the four proteins belong to the same a leading role. nucleation and spindle assembly. These re-
structure. To examine the interactions among sults suggest that the huoMTOC is required
these proteins, coimmunoprecipitation was The huoMTOC is fragmented and recruited to for spindle microtubule polymerization and
performed in human embryonic kidney 293T kinetochores for the initiation of spindle spindle assembly of human oocytes.
(HEK293T) cells transfected with plasmids assembly in human oocytes
containing the corresponding genes. As a re- Unlike the mechanism in mouse MI oocytes in huoMTOC deficiency interrupts normal
sult, CCP110, CKAP5, and DISC1 all interacted which aMTOCs were aggregated on the spin- spindle assembly and causes clinical oocyte
directly with TACC3, whereas the negative dle poles during spindle assembly (15), the maturation arrest
control GDF9 had no interaction with TACC3 components of the huoMTOC in human MI Oocyte maturation requires microtubule nu-
(fig. S6). This suggests that the four proteins oocytes were localized on kinetochores. To cleation and spindle assembly (22). In the
are all components of the same structure. In reveal the dynamic process of huoMTOC- clinic, a number of infertile patients with re-
addition, a dense microtubule cluster was ob- regulated spindle assembly, human GV oocytes current failed in vitro fertilization (IVF) or in-
served around this structure in human GV were fixed for immunofluorescence at NEBD tracytoplasmic sperm injection (ICSI) attempts
oocytes, which caused asymmetrical microtu- or at 2, 4, or 6 hours before NEBD. Initially, the have been diagnosed with oocyte maturation
bule distribution around the nuclear envelope huoMTOC appeared beneath the oocyte cor- arrest. Considering the key role of the huoMTOC
prior to NEBD (Fig. 2C). Given these features, tex. Slowly, the huoMTOC migrated from the in human spindle assembly, we hypothesized
we refer to this potential nucleating structure cortex to the nuclear envelope of human GV that a disrupted huoMTOC resulting from mu-
as the human oocyte microtubule organizing oocytes (Fig. 4, A and B). The huoMTOC then tations in CCP110, CKAP5, DISC1, or TACC3
center (huoMTOC). This structure was not expanded, and its nucleated microtubules grew may cause impaired spindle assembly and
detected in the oocytes of other mammalian rapidly during the resumption of meiosis (Fig. abnormal oocyte maturation in patients.
species such as mice and pigs (fig. S7), sug- 4, A and C). We thus screened for likely pathogenic mu-
gesting the specificity of this structure in We next visualized the dynamic changes of tations in a cohort of 1394 infertile female
human oocytes. the huoMTOC after NEBD by 3D time-lapse patients characterized by oocyte maturation
imaging. At the beginning of NEBD, the arrest by analyzing their whole exome sequenc-
The huoMTOC is essential for huoMTOC localized proximal to the chromo- ing datasets (data deposited in the Genome
microtubule nucleation somes and became fragmented within the first Variation Map of the National Genomics Data
According to the transcriptional landscape hour (fig. S9A and movie S3). The huoMTOC Center under accession number GVM000402).
data of human oocytes in public databases microtubules were disrupted and were barely Each of these patients had undergone several
(23), TACC3 shows overwhelmingly higher observable in the first few hours after NEBD IVF attempts, all of which failed because of
expression (more than 29-fold) than CCP110, (Fig. 4D and movie S4). The fragmented oocyte maturation arrest. We identified two
CKAP5, and DISC1, implying its potential key huoMTOC was then relocated to chromosomes, patients with compound heterozygous muta-
role in the huoMTOC. We therefore tried to primarily to kinetochores (Fig. 4D, fig. S9B, tions in the key huoMTOC component TACC3
disrupt the huoMTOC by knocking down and movie S5). The spindle microtubules were (Fig. 5A and table S1). According to the Ge-
TACC3 by injecting the corresponding short- initially observed when the huoMTOC was re- nome Aggregation Database (gnomAD) and
interfering RNAs (siRNAs) into human GV built (Fig. 4, D and E; fig. S9C; and movie S4). our in-house controls (data deposited in the
oocytes (fig. S8). The integrity of the huoMTOC Next, to determine the role of the huoMTOC Genome Variation Map of the National Geno-
was evaluated by immunofluorescence for in spindle assembly, each huoMTOC compo- mics Data Center under accession number
CCP110, CKAP5, DISC1, and TACC3. As ex- nent was down-regulated in human GV oocytes GVM000394), all four TACC3 variants from
pected, no nucleating structures were detected, that were cultured to the MI stage. Down- patients are rare variants (table S2). Impor-
demonstrating the complete disruption of the regulation of these components significantly tantly, these two patients showed very similar
huoMTOC upon TACC3 depletion (Fig. 3A). impaired spindle microtubule nucleation and phenotypes, in which most of their retrieved
To determine whether huoMTOC is the main spindle assembly in human MI oocytes com- oocytes were immature after in vitro matu-
microtubule nucleator, we assessed the mi- pared with the control group (P < 0.001, Fisher’s ration (table S1), and polarization microscopy
crotubule distribution in oocytes under the exact test), and stable microtubules were great- images showed no visible spindles in the live
condition of huoMTOC deficiency. The asym- ly decreased (Fig. 4, F and G). Compared to oocytes (Fig. 5B). Immunofluorescence in fixed
metrical distribution of microtubules was other components, TACC3 depletion had the NEBD oocytes also demonstrated that the

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 4 of 12

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

A
TACC3 CKAP5 CCP110 DISC1
Control siRNA TACC3 siRNA Control siRNA TACC3 siRNA Control siRNA TACC3 siRNA Control siRNA TACC3 siRNA

Hoechst Tubulin
1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0
Intensity (A.U.)

0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5

0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0

0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50
Length (µm)
B C
Control siRNA TACC3 siRNA CKAP5 siRNA CCP110 siRNA DISC1 siRNA P<0.001

Asymmetrical microtubule distribution (%)

100
(16)
Merge

Hoechst
60
Tubulin

40
(13)
(17) (15)
20
(13)
1.0 1.0 1.0 1.0 1.0
Intensity (A.U.)

1
0.5 0.5 0.5 0.5 0.5

tro

C
P1
KA
C

IS
on

D
C
C

C
0.0 0.0 0.0 0.0 0.0
0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 0 10 30 50 siRNAs
Length (µm)

Fig. 3. The huoMTOC is required for microtubule nucleation in human oocytes. specific siRNAs. Green, microtubules (tubulin); gray, chromosomes (Hoechst). The
(A) Immunofluorescence images of human GV oocytes injected with TACC3 yellow arrows indicate the direction of the tubulin intensity measurements shown
siRNAs. Green, microtubules (tubulin); magenta, huoMTOC (yellow squares); at the bottom. Scale bar, 10 mm. (C) The percentages of human oocytes with
blue, chromatin. Scale bar, 10 mm. The measurements of tubulin intensity along the asymmetrical microtubule distribution in (B). P < 0.001, FisherÕs exact test. Data
direction of the yellow arrows are shown at the bottom. Green, tubulin; blue, were from three independent experiments. The number of oocytes analyzed is
chromosomes. (B) Immunofluorescence images of human GV oocytes injected with specified in parentheses.

meiotic spindle was completely disrupted depletion, whereas supplementing the mutant bule nucleation and the initiation of acentro-
(Fig. 5C). We also determined the stability of TACC3 mRNAs could not rescue the phenotype somal spindle assembly.
the huoMTOC and microtubule distribution (Fig. 5, E and F), suggesting that the mutations
in the patients’ GV oocytes. Compared to normal had loss-of-function effects on TACC3. Thus, Discussion
human GV oocytes, the huoMTOC was miss- TACC3 deficiency caused female infertility and Here, we report a structure that we named the
ing, and the asymmetrical distribution of mi- oocyte maturation arrest by disrupting the in- huoMTOC, which serves as a major site of
crotubules around the nucleus was impaired tegrity of the huoMTOC. These results suggest microtubule nucleation and is required for
in the patients’ GV oocytes (Fig. 5D). that disruption of the huoMTOC impaired the spindle assembly in human oocytes. A single
In addition, supplementing the wild-type nucleation of microtubules in the GV oocytes huoMTOC is formed near the cortex of human
TACC3 mRNA successfully rescued the pheno- of patients, further highlighting the critical GV oocytes at the time of meiosis resumption,
type of spindle disruption resulting from TACC3 role of the huoMTOC in regulating microtu- and it migrates to the nuclear envelope before

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 5 of 12

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

Fig. 4. The huoMTOC is A B

recruited from the

Distance to nucleus (µm)

80
Early prophase (-6 h) Prophase (-4 h) Late prophase (-2 h) NEBD (0 h)
cortex to kinetochores
60
for spindle microtubule
polymerization in 40

Merge
human oocytes. 20 n=3
(A) Immunofluorescence 0
images of human GV -6 -4 -2 0
Time from NEBD (hours)
oocytes at NEBD and at
C
2 hours (late prophase),
1.0 Tubulin

Intensity (A.U.)
4 hours (prophase), and TACC3
6 hours (early prophase)

Tubulin
0.5
before NEBD. Green,
n=3
microtubules (tubulin);
0.0
blue, chromatin -6 -4 -2 0
(Hoechst); magenta, Time from NEBD (hours)

huoMTOC (TACC3). Green E 1.2 Tubulin

Intensity (A.U.)
arrows indicate the TACC3
microtubule cluster, and 0.6
TACC3

magenta arrows indicate

n=3
the huoMTOC. Scale bar,
0
20 mm. (B) The distance 3 4 5 6
between the huoMTOC Time from NEBD (hours)
and the nucleus was D
measured in (A). Time is 3:10 4:00 5:00 5:50 7:00 8:20 11:20 11:40 12:10
after NEBD. (C) The
Chromosome

intensity of the micro-

tubule cluster (tubulin)
and the huoMTOC
(TACC3) was measured
at different time points in
TACC3

(A). Time is after NEBD.

(D) Representative time-
lapse images showing
the relationship between
Tubulin

the huoMTOC and micro-

tubule nucleation in
live human oocytes.
Green, microtubules (SiR- F G
tubulin); blue, chromo- siRNAs P<0.001
somes (H2B-mClover); Control TACC3 CKAP5 CCP110 DISC1 100
(24)
magenta, huoMTOC

Spindle assembly (%)

(mScarlet-TACC3). Scale 80
bar, 5 mm. (E) The inten-
Merge

60
sity of microtubules
(SiR-tubulin) and
Hoechst 40 (18) (16)
huoMTOC (mClover- (18)
TACC3) close to chromo- 20
(20)
somes was measured in
Tubulin

(D). Time is after NEBD. 0

The number of oocytes tro
l
C
3
P5 10 1
on
C
KA P1 IS
C
analyzed in experiments C TA C C
C D

is indicated. The mean siRNAs

and standard error were
calculated on the basis of two independent experiments. Error bars are standard deviations. (F) Immunofluorescence images of human MI spindles from control
and TACC3, CKAP5, CCP110, and DISC1 siRNA-injected human oocytes. Green, microtubules (tubulin); gray, chromosomes (Hoechst). Scale bar, 5 mm. (G) The spindle
assembly percentage measured and collected from (F). The number of oocytes analyzed in three independent experiments is indicated. P < 0.001, FisherÕs exact test.

NEBD. After NEBD, the huoMTOC becomes recruited to the spindle microtubules (Fig. 6 components of the huoMTOC and that muta-
fragmented and is recruited to chromosomes and fig. S9C). Ablation of the huoMTOC re- tions in TACC3 cause clinical oocyte maturation
and kinetochores for spindle microtubule nucle- sults in microtubule loss and defective spindle arrest and female infertility.
ation (Fig. 6). With the microtubule poly- assembly. In addition, we demonstrated that Distinct aMTOCs have been identified and
merization, the huoMTOC proteins are also TACC3, CCP110, CKAP5, and DISC1 are essential investigated in mouse oocytes as microtubule

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 6 of 12

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

Fig. 5. Disruption of A C
the huoMTOC in human Family 1 Family 2
Normal Family 1 II-1 Family 2 II-1
oocytes impairs micro- c.530G>C [p.Ser177Thr] c.673_708del [p.Lys225_Cys236del]
c.1184C>G [p.Pro395Arg] c.1892G>T [p.Gly631Val]
tubule nucleation and
I I

Merge
spindle assembly. 1 2
1 2
(A) Pedigrees of the two [c.673_708del];[WT] [c.1892G>T];[WT]
[c.1184C>G];[WT] [c.530G>C];[WT]
families with TACC3 II
II
mutations with Sanger 1
1 2
[c.530G>C];[c.1184C>G]
sequencing confirmation. [c.673_708del];[c.1892G>T] [c.673_708del];[WT]

TACC3
Squares denote male c.530G>C c.1184C>G c.673_708del c.1892G>T
family members, circles [p.Ser177Thr] [p.Pro395Arg] [p.Lys225_Cys236del] [p.Gly631Val]
denote female members,
black solid circles denote I-1 I-1
A A A G C G G A G A C C A CCG G A CC T A
probands, and the equal
T G G C A G C C C T G C C C C C C/G C A T G C
CGG C A C G G T G G Hoechst Tubulin

sign denotes infertility.

(B) Human oocytes from
I-2
T G G C A G/C C C C T G C C C C C C C A T G C
I-2
A A A G C G G A G A C C A C C G G/T A C C T A
D
donors (normal) and II-1 Normal Family 1 II-1 Family 2 II-1
T G G C A G/C C C C T G C C C C C C/G C A T G C
II-1
patients (family 1 II-1, A A A G C G G A G A C C A C C G G/T A C C T A
CGG C A C G G T G G
family 2 II-1) were

Merge
examined by light and II-2
A A A G C G G A G A C CACC G G ACC T A
polarization microscopy. B
CGG C A C G G T G G

The arrow indicates an MI

Normal Family 1 II-1 Family 2 II-1
spindle. (C) Immuno-
fluorescence images of

TACC3
human MI oocytes from
donors (normal) and
Light

patients (family 1 II-1,

family 2 II-1). Green,
microtubules (tubulin); Hoechst Tubulin

blue, chromosomes
(Hoechst); magenta, 1.0 1.0 1.0
TACC3. Scale bar, 5 mm.

Intensity (A.U.)
Polarization

(D) Immunofluorescence 0.5 0.5

0.5
images of human GV
oocytes from donors
0.0 0.0 0.0
(normal) and patients 0 10 30 50 0 10 30 50 0 10 30 50
(family1 II-1, family2 II-1). Length (µm)

Green, microtubules
E F
(tubulin); blue, chromo- TACC3 siRNA 100 P<0.05
somes (Hoechst);
H2O WT S177T P395R K225_C236del G631V (10)
magenta, TACC3. Scale 80
bar, 10 mm. The yellow

Maturation rate (%)

square shows the
Merge

60
huoMTOC in normal (6)

human oocyte. The

40 (9)
microtubule distribution
(9)
was measured as
20 (7)
TACC3

previously described. (11)

(E) Immunofluorescence
0
images of human MI

1V
2O

23 R
S1 T
25 P3 T

G l
e
W
77
_C 95
6d
H

63
oocytes injected with H2B Tubulin
TACC3 siRNAs and wild K2

type or patient-derived
mutant mRNAs. Green,
microtubules (tubulin); blue, chromosomes (Hoechst), magenta, FLAG-TACC3. Scale bar, 5 mm. (F) The percentage of human oocyte maturation measured in (E).
(FisherÕs exact test, P < 0.05). The number of oocytes analyzed is specified in parentheses.

nucleators for meiotic spindle assembly (15). the huoMTOC is not concentrated on the spin- and it is fragmented immediately after NEBD,
However, the primary spindle microtubule nu- dle poles in human MI and MII oocytes (22). making it difficult to capture in human oo-
cleator of human oocytes has remained un- (ii) The classic aMTOC marker pericentrin is cytes by immunofluorescence.
known. The following factors might be reasons not among the components of the huoMTOC In previous investigations, spindle assem-
why the huoMTOC was not identified in pre- and therefore cannot label the structure. (iii) Only bly of human oocytes was reported to be me-
vious investigations: (i) Unlike mouse aMTOCs, a single huoMTOC is formed in late prophase, diated by chromosomes and dependent on

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 7 of 12

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

tubule polymerization have not yet been deter-

mined in either mitotic or meiotic cells.
Apart from the four huoMTOC components,
we also identified 39 other proteins that
showed spindle-related localization. We there-
fore cannot exclude the possibility that some
of these proteins may also play a role in
Before NEBD ~ -6 hours ~ -4 hours ~ -2 hours ~ 0 hours huoMTOC formation and function. The spe-
cific components of huoMTOC and the mech-
Early prophase Prophase Late prophase NEBD
anism for its nucleation of microtubules are
worth investigating in the future. In addition,
according to our observations, the localization
of most centrosome and microtubule-related
proteins in the MI spindle of human oocytes is
obviously different from that in mouse oocytes
(21). Thus, future investigations on the func-
tions of these proteins in human oocytes should
After NEBD ~ 0-2 hours ~ 3~5 hours ~ 6-8 hours ~ 9-12 hours shed more light on the mechanism of human
oocyte spindle assembly.
In a recent study, a distinct mechanism of
Nucleus Chromosome Microtubule huoMTOC Kinetochore
spindle pole organization was discovered in
human oocytes, suggesting that loss of kinesin
Fig. 6. Mechanistic model for huoMTOC migration and microtubule nucleation in human oocytes. superfamily protein C1 (KIFC1) induces meiotic
The huoMTOC (magenta) assembles near the cortex of GV oocytes and migrates to the nuclear envelope spindle instability (23). In addition, our pre-
before NEBD. The huoMTOC expands and the surrounding microtubules (green) keep growing until NEBD. After vious investigations suggest that tubulin beta 8
NEBD, the huoMTOC fragments, and the surrounding microtubules are disassembled. The fragmented huoMTOC is class VIII (TUBB8) is the main isotype of spin-
then recruited to kinetochores and initiates microtubule nucleation for meiotic spindle assembly. dle b-tubulin in human oocytes but is not found
in mice or other nonprimate species (50). These
findings suggest that a distinct mechanism for
RanGTP (22). It has been demonstrated that formed in human MI oocytes by systematic the initiation of microtubule nucleation and
RanGTP inhibition impairs spindle assembly in immunofluorescent staining. We identified spindle assembly has evolved in human oocytes,
human oocytes (22). In mouse or Drosophila oo- the components (TACC3, CKAP5, CCP110, and which may contribute to a series of physio-
cytes, spindles have defects but still assemble if DISC1) of the huoMTOC on kinetochores in logical characteristics including increasing
RanGTP is inhibited, suggesting that the RanGTP MI oocytes and then verified their localiza- chromosome segregation errors, high spindle
pathway is not essential for meiotic spindle as- tion in GV oocytes using high-resolution imag- instability, and aneuploidy in human oocytes.
sembly in oocytes of these two species (35, 36). ing in live human oocytes. Emerging evidence In the clinic, many infertility patients ex-
In the mouse oocytes lacking centrosomes, the suggests that the human TACC3 plays an im- perience recurrent failure of IVF or ICSI at-
processes of spindle assembly and spindle mi- portant role in microtubule growth and spin- tempts owing to oocyte maturation arrest.
crotubule nucleation are mainly achieved by dle assembly during mitosis (30, 40–42). The However, the genetic factors involved remain
aMTOCs and facilitated by liquid-like meiotic microtubule nucleation is blocked in ovarian unknown for most patients. In this study, we
spindle domain (21, 37). It has been argued for cancer cells when TACC3 expression is affect- found that mutations in TACC3 cause defects
a long time that, unlike mouse oocytes, human ed, suggesting that TACC3 is required for in spindle assembly by disrupting the struc-
oocytes lack prominent aMTOCs (2, 22, 38). centrosome-involved microtubule nucleation ture of the huoMTOC, which leads to oocyte
However, in the present study, the spindle mi- (43). TACC3 depletion impairs the g-tubulin maturation arrest in patients. It is worth per-
crotubule in human oocytes was observed to ring complex assembly (44) and centrosome forming mutational screening for both known
be primarily polymerized from the kinetochores integrity (45) of human somatic cells. CKAP5 and potential genes that might participate in
by an aMTOC-like structure that we named was previously identified as a microtubule nu- the maintenance of huoMTOC integrity. The
huoMTOC. In the Drosophila oocytes that also cleation factor in vitro (32, 33, 46). CKAP5 was results of such screening will help in precision
lack prominent aMTOCs, the spindle assembly observed functioning synergistically with the diagnosis for these patients and will provide
is dominated by the chromosomes, which re- g-tubulin ring complex for de novo microtu- therapeutic targets for future clinical treat-
cruit and/or nucleate the microtubules (39). bule nucleation (32) and catalyzing numer- ments. Our findings provide not only insights
The spindle microtubules are organized around ous rounds of tubulin subunit addition at the into the physiological mechanism of micro-
the chromosomes into fibers of two types, microtubule plus-end for spindle microtubule tubule nucleation and spindle assembly in
the interpolar and kinetochore microtubules, assembly in vitro (33). In addition, CKAP5 was human oocytes but also improve our under-
to assemble the meiotic spindle of Drosophila also implicated in the importin-regulated mi- standing of pathophysiological mechanisms of
(39). Although the spindle assembly is directed crotubule nucleation as a microtubule poly- human oocyte maturation arrest.
by chromosomes in both human and Drosophila merase in vitro (46). In mitosis, TACC3 and
oocytes, the mechanisms involved could be CKAP5 are RanGTP-regulated spindle assem- Materials and methods
different. The meiotic spindle assembly in bly factors for spindle microtubule nucleation Human oocyte collection and culture
Drosophila oocytes requires the chromosomal and stabilization (47–49). Collectively, we in- Human GV oocytes were donated by patients
passenger complex (16), but that in human ferred that TACC3 and CKAP5 may act as undergoing ICSI as part of their assisted re-
oocytes requires the huoMTOC. organizers of huoMTOC assembly and micro- production treatment at Shanghai Ji Ai Genetics
In our study, the localization of 86 centro- tubule nucleation in human oocytes. The and IVF Institute affiliated with the Obstetrics
some and spindle-related proteins was per- functions of CCP110 and DISC1 in micro- and Gynecology Hospital of Fudan University,

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 8 of 12

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

Center for Reproductive Medicine and Fertility (Sigma-Aldrich) and used at a concentration 1:50), anti-DLGAP5 antibody (A13575; ABclonal;
Preservation program affiliated with International of 0.1% in G-MOPS. 1:50), anti-DYNLT1 antibody (11954-1-AP; Pro-
Peace Maternity and Child Health Hospital of teintech; 1:50), anti-GTSE1 antibody (A302-
Shanghai Jiao Tong University. Only imma- Immunofluorescence microscopy 425A; Bethyl Laboratories; 1:50), anti-HAUS4
ture oocytes that were unable to be used for Human GV oocytes were fixed for immunofluo- antibody (20104-1-AP; Proteintech; 1:50), anti-
assisted reproduction treatment were col- rescence at 2, 4, or 6 hours after milrinone HAUS6 antibody (A4797; ABclonal; 1:50), anti-
lected for this research, and the use of these washout. Human MI oocytes were fixed at 5 to HAUS8 antibody (PA5-21331; Thermo Fisher
human GV oocytes was clearly explained to 12 hours after NEBD (prometaphase or meta- Scientific; 1:100), anti-HMMR antibody (15820-
the patient donors. All these female donors phase). Human oocytes were fixed for 30 min 1-AP; Proteintech; 1:50), anti-HOOK2 antibody
were receiving ICSI treatment because of male in phosphate-buffered saline (PBS) containing (ab133691; Abcam; 1:50), anti-HOOK3 anti-
factor-induced infertility. The age range of the 2% formaldehyde and 0.1% Triton-X at 37°C body (A15536; ABclonal; 1:50), anti-HOOK3
donors was 25 to 38 and their mean age was on a heat block and were then permeabilized antibody (15457-1-AP; Proteintech; 1:50), anti-
34.3. The BMI range of these donors was 20 to in PBS containing 0.5% Triton-X (PBT) at 4°C KANSL3 antibody (HPA035018; Merck; 1:100),
22. The collected human oocytes were mixed overnight. Oocytes were extensively washed anti-KIF11 antibody (HPA010568; Merck; 1:100),
thoroughly and distributed randomly to the with PBS between stages and then blocked anti-KIF20A antibody (15911-1-AP; Proteintech;
control and experimental groups. Only mor- at room temperature in a blocking buffer of 1:50), anti-KIF20A antibody (CL594-67190; Pro-
phologically normal human GV oocytes were 3% bovine serum albumin (BSA) in 0.3% PBT. teintech; 1:50), anti-KIF22 antibody (A19881;
used in this investigation. G-MOPS medium All antibody incubations were performed in ABclonal; 1:50), anti-KIF2A antibody (13105-
(Vitrolife) with milrinone (2 mM, HY-14252, blocking buffer at 4°C overnight for primary 1-AP; Proteintech; 1:50), anti-KIF2B antibody
MedChemExpress) was used to maintain hu- antibodies and at 37°C for 1 hour for secondary (A6480; ABclonal; 1:20), anti-KIFC1 antibody
man oocyte prophase arrest. The human GV antibodies. Primary antibodies were anti- (A3304; ABclonal; 1:20), anti-KIZ antibody (21177-
oocytes were matured in Multipurpose Hand- centromere antibody (HCT-0100; Immuno- 1-AP; Proteintech; 1:50), anti-LIS1 antibody
ling Medium-Complete (MHM-C) (FUJIFILM Vision; 1:500), anti-AKAP450 antibody (611518; (H00005048-M03; Abnova; 1:50), anti-LMNB
Irvine Scientific) or G-MOPS medium at 37°C BD Biosciences; 1:50), anti-ASPM antibody antibody (A1910; ABclonal; 1:20), anti-LRRC45
on a heating block. (26223-1-AP; Proteintech; 1:50), anti-AURKA antibody (PA5-54777; Thermo Fisher Scientif-
antibody (NBP2-50041; Novus Biological; 1:50), ic; 1:100), anti-MCRS1 antibody (HPA039057;
Animals and oocyte culture anti-AURKA antibody (10297-1-AP; Protein- Merck; 1:100), anti-MYO10 antibody (sc-23137;
The porcine ovaries were collected from local tech; 1:20), anti-AURKB antibody (ab45145; SantaCruz Biotechnology; 1:50), anti-NDE1 anti-
slaughterhouses and transported in warm Abcam; 1:50), anti-BBS4 antibody (12766-1-AP; body (10233-1-AP; Proteintech; 1:50), anti-NDEL1
0.9% NaCl. Porcine GV oocytes were collected Proteintech; 1:50), anti-beta tubulin antibody antibody (H00081565-D01P; Abnova; 1:50), anti-
in TCM199 medium at 39°C, and milrinone (ab204686; Abcam; 1:50), anti-beta tubulin anti- NEDD1 antibody (13993-1-AP; Proteintech; 1:50),
(2 mM) was added to the medium to main- body (ab11309; Abcam; 1:100), anti-BUGZ anti- anti-NEK2 antibody (14233-1-AP; Proteintech;
tain oocyte prophase arrest. Only fully grown body (A20177; ABclonal; 1:20), anti-CAMSAP2 1:50), anti-NIN antibody (A8215; ABclonal; 1:50),
oocytes were used in our experiments. For antibody (17880-1-AP; Proteintech; 1:50), anti- anti-NUMA1 antibody (A0527; ABclonal; 1:50),
maturation, GV oocytes were washed free from CCP110 antibody (12780-1-AP; Proteintech; 1:50), anti-NUP107 antibody (A13110; ABclonal;
milrinone and cultured in fresh TCM199 anti-CCP110 antibody (PA5-58775; Thermo Fisher 1:50), anti-NUP160 antibody (PRS4707; Merck;
medium at 39°C and 5% CO2. Porcine GV Scientific; 1:100), anti-CDK5RAP2 antibody 1:100), anti-NUP62 antibody (13916-1-AP; Pro-
and MI oocytes were fixed for immunofluo- (06-1398; Merck; 1:100), anti-CENPJ antibody teintech; 1:50), anti- NUP85 antibody (19370-1-
rescence at 6 or 15 hours after oocyte isola- (11517-1-AP; Proteintech; 1:50), anti-CEP57 anti- AP; Proteintech; 1:50), anti-NUSAP1 antibody
tion, respectively. body (24957-1-AP; Proteintech; 1:50), anti-CEP63 (12024-1-AP; Proteintech; 1:50), anti-ODF2 anti-
Female C57Bl/6 mice (3- to 4-week-old) were antibody (16268-1-AP; Proteintech; 1:50), antibody (12058-1-AP; Proteintech; 1:50), anti-PCNT
purchased from Beijing Vital River Laboratory CEP72 antibody (19928-1-AP; Proteintech; 1:50), antibody (611815; BD Biosciences; 1:200), anti-
Animal Technology Co., Ltd. (Beijing, China). anti-CEP120 antibody (PA5-55985; Thermo Fisher PCM1 antibody (19856-1-AP; Proteintech; 1:50),
All mice were used in accordance with insti- Scientific; 1:100), anti-CEP135 antibody (24428- anti-PLK1 antibody (A2548; ABclonal; 1:50),
tutional guidelines, and the experiments were 1-AP; Proteintech; 1:50), anti-CEP152 antibody anti-PLK3 antibody (10977-1-AP; Proteintech;
approved by the Animal Care and Use Com- (21815-1-AP; Proteintech; 1:50), anti-CEP164 1:50), anti-PLK4 antibody (12952-1-AP; Pro-
mittee of Fudan University, China. Mouse GV antibody (22227-1-AP; Proteintech; 1:50), anti- teintech; 1:50), anti-PRC1 antibody (15617-1-
oocytes were released from the ovaries at 44 CEP170 antibody (27325-1-AP; Proteintech; AP; Proteintech; 1:50), anti-RAE1 antibody
to 52 hours after injection of 10 IU pregnant 1:50), anti-CEP192 antibody (18832-1-AP; (20491-1-AP; Proteintech; 1:50), anti-RAN anti-
mare serum gonadotropin. For maturation, Proteintech; 1:50), anti-CEP250 antibody (14498- body (10469-1-AP; Proteintech; 1:50), anti-
denuded GV oocytes were cultured in fresh 1-AP; Proteintech; 1:50), anti-CEP290 antibody SASS6 antibody (21377-1-AP; Proteintech; 1:50),
M2 medium at 37°C on a heat block. The pH (22490-1-AP; Proteintech; 1:50), anti-CETN2 anti- anti-SKAP2 antibody (12926-1-AP; Proteintech;
value of culture medium was between 7.2 and body (A5397; ABclonal; 1:20), anti-CETN3 1:50), anti-SNF2H antibody (A2000; ABclonal;
7.4 (S210, Mettler Toledo). Mouse GV and MI antibody (A8111; ABclonal; 1:50), anti-CKAP5 1:50), anti-SPDL1 antibody (PA5-99285; Thermo
oocytes were fixed for immunofluorescence at antibody (PA5-59150; Thermo Fisher Scienti- Fisher Scientific; 1:100), anti-SSX2IP antibody
0.5 or 6 hours after in vitro maturation initia- fic; 1:100), anti-CKAP5 antibody (CL488-67631; (13694-1-AP; Proteintech; 1:50), anti-TACC3
tion, respectively. Proteintech; 1:50), anti-CLIP1 antibody (23839- antibody (A18641; ABclonal; 1:50), anti-TACC3
1-AP; Proteintech; 1:50), anti-CLTC antibody antibody (ab134154; Abcam; 1:100), anti-TOP2A
Inhibitor treatment (610500; BD Biosciences; 1:50), anti-CNTROB antibody (20233-1-AP; Proteintech; 1:50), anti-
For microtubule depolymerization, nocodazole antibody (26880-1-AP; Proteintech; 1:50), anti- TPX2 antibody (A18327; ABclonal; 1:20), anti-
(10 mM, HY-13520, MedChemExpress) was DCTN1 antibody (55182-1-AP; Proteintech; 1:50), TUBG1 antibody (A9657; ABclonal; 1:20).
added to human oocytes at ∼10 hours after anti-DCTN2 antibody (A2200; ABclonal; 1:50), Secondary antibodies were Alexa Fluor 647-
NEBD and 1 hour before time-lapse imaging. anti-DYNC1H1 antibody (12345-1-AP; Proteintech; conjugated anti-rabbit IgG (4414S; Cell Sig-
The drug was dissolved in dimethyl sulfoxide 1:50), anti-DISC1 antibody (A4678; ABclonal; naling; 1:200), Alexa Fluor 647-conjugated

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 9 of 12

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

anti-mouse IgG (4410S; Cell Signaling; 1:200), visualized by SiR-tubulin (1 mM) staining in immunoglobulin G (IgG) (1:5000 dilution;
Cy3-conjugated anti-rabbit IgG (AS008; ABclonal; G-MOPS. To reduce background noise, some Abmart) or goat anti-mouse IgG (1:5000 di-
1:500), and Atto 488-conjugated anti-human images were passed through a Gaussian filter lution; Abmart) conjugated to horseradish
IgG (52526; Merck; 1:500). Chromatin was of 2 sigma in Fiji (NIH). peroxidase.
briefly counterstained with Hoechst 33342
(20 mg·ml−1, HY-15559, MedChemExpres) be- Fluorescent intensity measurement Clinical samples
fore imaging. The samples were in PBS and The pattern of microtubule distribution was A cohort of 1394 infertile female patients with
imaged with an LSM 880 confocal laser scan- defined by the intensity of tubulin around the oocyte maturation arrest recruited from the
ning microscope (ZEISS) with a 63×/1.4 NA nuclear envelope. The direction of intensity Ninth Hospital affiliated with Shanghai Jiao
Plan Apochromat oil immersion lens at room measurements is shown in the figures. The Tong University and Shanghai Ji Ai Genetics
temperature. intensity was measured by Image J (NIH). The and IVF Institute affiliated with the Obstetrics
integrated intensity of kinetochore foci was and Gynecology Hospital of Fudan University
RNA interference measured with the Foci_Picker3D plugin in participated in this study. Written informed
The siRNAs were provided by GenePharma or Image J. The same threshold was applied to consent was provided by patients. The recruit-
Tsingke Biotechnology. The sequences of siRNA each focus within an oocyte. ment of patients was performed as follows: (i)
for TACC3 down-regulation were 5′-GGU UCG female patients were younger than 45 years
AAG AGG UUG UGU A-3′ and 5′-GCA UGC Laser ablation old, failing to conceive after 1 year (or longer)
ACG GUG CAA AUG A-3′. The siRNA sequen- To examine the role of huoMTOC in microtu- of regular unprotected sex; (ii) had undergone
ces against CKAP5 were 5′-GGA AAT AGC TGT bule polymerization of human oocytes, the ≥2 failed attempts of IVF/ICSI, characterized
TCA CAT A-3′ and 5′-GGC CAA AGC TCC AGG huoMTOC was directly disrupted by laser in by oocyte maturation arrest; (iii) female patients
ATT A-3′. The siRNA sequences targeting CCP110 live human GV oocytes. The human GV oo- with other known causes of infertility, includ-
were 5′-CAC UCU ACU GCA GCA AAG C-3′ and cytes expressing mClover3-TACC3 were ro- ing male factors, chromosome anomalies, radio-
5′-AUG UUC UUC UCC AAG GUG C-3′. The tated with an unbroken microinjection pipette therapy, or chemotherapy, were excluded.
siRNA sequence targeting DISC1 was 5′-GGA to obtain huoMTOC. The square regions of Peripheral blood samples were taken for DNA
UUU GAG AAU AGU UUC A-3′. The negative interest were marked and photobleached using extraction.
control siRNAs were provided by the same com- a 488-nm laser line at the maximum power. The GV and MI oocytes from two patients
pany. To increase the efficiency of RNA inhibi- The laser ablation was performed at 37°C. with compound heterozygous mutations in
tion, mixed siRNAs were microinjected into TACC3 were obtained as part of their assisted
human GV oocytes. The final concentration of Cell culture and transfection reproduction treatment at the Shanghai Ninth
siRNAs was 40 mM. HEK293T cells were obtained from the Cell Bank Hospital affiliated to Shanghai Jiao Tong
of Shanghai Institute for Biological Sciences, University.
Microinjection of human oocytes the Chinese Academy of Sciences (Shanghai, This study was approved by the Ethics Com-
Human GV oocytes were microinjected in China). Cells were cultured in Dulbecco’s mod- mittee of the Medical College of Fudan Univ-
G-MOPS with milrinone (2 mM) on the stage of ified Eagle’s medium supplemented with 10% ersity and the Reproductive Study Ethics
an inverted microscope (Leica) with microma- fetal bovine serum and 1% penicillin-streptomycin Committees of the hospitals.
nipulators (Eppendorf). A 0.1 to 0.3% volume of (Gibco, Waltham, MA, USA) in an atmosphere
mRNA was injected using a timed pulse, and of 5% CO2 at 37°C to between 70 and 80% con- Genetic studies
the final concentration of mRNA was 1 mg/ml. fluence. Plasmids were transfected into HEK293T Genomic DNA was extracted from peripheral
The injected GV oocytes were arrested in pro- cells using the PolyJet In Vitro DNA Transfec- blood using the QIAamp DNA Blood Mini Kit
phase for 2–4 hours for mRNA expression. tion Reagent (SignaGen) according to the man- (Qiagen). Whole-exome capture was performed
ufacturer’s instructions. using the SeqCap EZ Exome Kit (Roche), and
mRNA synthesis sequencing was performed on the Illumina
mRNA was transcribed in vitro from purified Immunoblots and immunoprecipitation NovaSeq 6000 platform (Illumina). Sequenc-
linear double-stranded DNA templates. HEK293T cells were harvested after transfec- ing analysis was compared with the human
mMessage T7 or T3 RNA polymerase kits tion for 36 hours and washed with PBS. Cells reference sequence (NCBI Genome build
(New England Biolabs) were used for the were lysed in radioimmunoprecipitation assay GRCh37). Mutations were annotated with
in vitro transcription reaction. The constructs lysis buffer (Shanghai Wei AO Biological Tech- GRCh37 and the dbSNP (version 138) and
of mClover3-CENPB, mScarlet-CENPB, mClover3- nology, Shanghai, China) with 1% protease in- gnomAD along with our in-house exome data-
TACC3, mScarlet-TACC3, mClover3-CKAP5, hibitor cocktail (Bimake, Houston, TX, USA). base (data deposited in public database, acces-
mClover3-CCP110, and mClover3-DISC1 were After quantification with the bicinchoninic acid sion numbers GVM000402 and GVM000394).
made and used for mRNA production. assay (Shanghai Biocolor BioScience & Tech-
nology Co.), the supernatant was subjected to Statistical analysis
Live cell imaging immunoprecipitation with affinity beads (Sigma). The investigators were not blinded during ex-
For high-resolution time-lapse imaging, time After incubation at 4°C for 4 hours, the beads periments and outcome assessment. The ex-
points were acquired at 10-min intervals using were washed with lysis buffer four times. periments were not randomized. The statistical
an LSM 880 confocal laser scanning micro- The bead-bound proteins were eluted using methods were not used to determine sample
scope (ZEISS) fitted with sensitive detectors, sodium dodecyl sulfate (SDS) sample buffer, size. Sample means were compared with either
an environmental chamber set to 37°C, and a resolved by SDS–polyacrylamide gel electro- Student’s t test or one-way analysis of variance
long-distance 40×/1.1 NA C-Apochromat water phoresis (SDS-PAGE), transferred to nitrocellu- (ANOVA) with a post-hoc test as stated (two-
immersion lens. A volume of 30 mm by 30 mm lose membranes (Pall Corporation), and probed sided). Dichotomous data were compared
by 15 mm centered around the chromosomes with rabbit anti-FLAG (1:3000 dilution; Cell using Fisher’s exact test (two-tailed). All data
was typically imaged. The chromosomes were Signaling Technology) or mouse anti-vinculin are from at least two independent experiments.
tracked automatically by MyPiC on LSM880. (1:5000 dilution; Sigma-Aldrich) antibodies. All tests were performed using GraphPad
Microtubules in live human oocytes were The secondary antibodies were goat anti-rabbit Prism 7 (GraphPad Software).

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 10 of 12

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

RE FE RENCES AND N OT ES 21. C. So et al., A liquid-like spindle domain promotes 379–388 (2008). doi: 10.1016/[Link].2008.06.005;
1. S. L. Prosser, L. Pelletier, Mitotic spindle assembly in acentrosomal spindle assembly in mammalian oocytes. Science pmid: 18656360
animal cells: A fine balancing act. Nat. Rev. Mol. Cell Biol. 364, eaat9557 (2019). doi: 10.1126/science.aat9557; 42. D. G. Booth, F. E. Hood, I. A. Prior, S. J. Royle, A TACC3/ch-TOG/
18, 187–201 (2017). doi: 10.1038/nrm.2016.162; pmid: 31249032 clathrin complex stabilises kinetochore fibres by inter-microtubule
pmid: 28174430 22. Z. Holubcová, M. Blayney, K. Elder, M. Schuh, Error-prone bridging. EMBO J. 30, 906–919 (2011). doi: 10.1038/
2. B. Mogessie, K. Scheffler, M. Schuh, Assembly and positioning chromosome-mediated spindle assembly favors emboj.2011.15; pmid: 21297582
of the oocyte meiotic spindle. Annu. Rev. Cell Dev. Biol. 34, chromosome segregation defects in human oocytes. 43. R. Yao et al., A small compound targeting TACC3 revealed its
381–403 (2018). doi: 10.1146/annurev-cellbio-100616-060553; Science 348, 1143–1147 (2015). doi: 10.1126/science. different spatiotemporal contributions for spindle assembly in
pmid: 30028643 aaa9529; pmid: 26045437 cancer cells. Oncogene 33, 4242–4252 (2014). doi: 10.1038/
3. S. Doxsey, Re-evaluating centrosome function. Nat. Rev. 23. C. So et al., Mechanism of spindle pole organization and onc.2013.382; pmid: 24077290
Mol. Cell Biol. 2, 688–698 (2001). doi: 10.1038/35089575; instability in human oocytes. Science 375, eabj3944 (2022). 44. P. Singh, G. E. Thomas, K. K. Gireesh, T. K. Manna, TACC3
pmid: 11533726 doi: 10.1126/science.abj3944; pmid: 35143306 protein regulates microtubule nucleation by affecting g-tubulin
4. S. Petry, Mechanisms of mitotic spindle assembly. Annu. Rev. 24. T. Wu, S. I. R. Lane, S. L. Morgan, F. Tang, K. T. Jones, ring complexes. J. Biol. Chem. 289, 31719–31735 (2014).
Biochem. 85, 659–683 (2016). doi: 10.1146/annurev-biochem- Loss of centromeric RNA activates the spindle assembly doi: 10.1074/jbc.M114.575100; pmid: 25246530
060815-014528; pmid: 27145846 checkpoint in mammalian female meiosis I. J. Cell Biol. 45. T. V. Suhail, P. Singh, T. K. Manna, Suppression of centrosome
5. A. Vertii, H. Hehnly, S. Doxsey, The centrosome, a multitalented 220, e202011153 (2021). doi: 10.1083/jcb.202011153; protein TACC3 induces G1 arrest and cell death through
renaissance organelle. Cold Spring Harb. Perspect. Biol. 8, pmid: 34379093 activation of p38-p53-p21 stress signaling pathway. Eur. J.
a025049 (2016). doi: 10.1101/cshperspect.a025049; 25. J. R. A. Hutchins et al., Systematic analysis of human protein Cell Biol. 94, 90–100 (2015). doi: 10.1016/[Link].2014.12.001;
pmid: 27908937 complexes identifies chromosome segregation proteins. pmid: 25613365
6. J. L. Badano, T. M. Teslovich, N. Katsanis, The centrosome in Science 328, 593–599 (2010). doi: 10.1126/science.1181348; 46. J. Roostalu, N. I. Cade, T. Surrey, Complementary activities of
human genetic disease. Nat. Rev. Genet. 6, 194–205 (2005). pmid: 20360068 TPX2 and chTOG constitute an efficient importin-regulated
doi: 10.1038/nrg1557; pmid: 15738963 26. H. V. Goodson, E. M. Jonasson, Microtubules and microtubule- microtubule nucleation module. Nat. Cell Biol. 17, 1422–1434
7. D. Szollosi, P. Calarco, R. P. Donahue, Absence of centrioles in associated proteins. Cold Spring Harb. Perspect. Biol. 10, (2015). doi: 10.1038/ncb3241; pmid: 26414402
the first and second meiotic spindles of mouse oocytes. a022608 (2018). doi: 10.1101/cshperspect.a022608; 47. P. R. Clarke, C. Zhang, Spatial and temporal coordination of
J. Cell Sci. 11, 521–541 (1972). doi: 10.1242/jcs.11.2.521; pmid: 29858272 mitosis by Ran GTPase. Nat. Rev. Mol. Cell Biol. 9, 464–477
pmid: 5076360 27. Z. Chen, V. B. Indjeian, M. McManus, L. Wang, B. D. Dynlacht, (2008). doi: 10.1038/nrm2410; pmid: 18478030
8. A. T. Hertig, E. C. Adams, Studies on the human oocyte and CP110, a cell cycle-dependent CDK substrate, regulates 48. O. J. Gruss et al., Ran induces spindle assembly by reversing
its follicle. I. Ultrastructural and histochemical observations on centrosome duplication in human cells. Dev. Cell 3, 339–350 the inhibitory effect of importin a on TPX2 activity. Cell
the primordial follicle stage. J. Cell Biol. 34, 647–675 (1967). (2002). doi: 10.1016/S1534-5807(02)00258-7; 104, 83–93 (2001). doi: 10.1016/S0092-8674(01)00193-3;
doi: 10.1083/jcb.34.2.647; pmid: 4292010 pmid: 12361598 pmid: 11163242
9. T. Mikeladze-Dvali et al., Analysis of centriole elimination 28. F. Gergely et al., The TACC domain identifies a family of 49. P. Kalab, R. Heald, The RanGTP gradient – a GPS for the
during C. elegans oogenesis. Development 139, centrosomal proteins that can interact with microtubules. mitotic spindle. J. Cell Sci. 121, 1577–1586 (2008).
1670–1679 (2012). doi: 10.1242/dev.075440; Proc. Natl. Acad. Sci. U.S.A. 97, 14352–14357 (2000). doi: 10.1242/jcs.005959; pmid: 18469014
pmid: 22492357 doi: 10.1073/pnas.97.26.14352; pmid: 11121038 50. R. Feng et al., Mutations in TUBB8 and human oocyte meiotic
10. A. Pimenta-Marques et al., A mechanism for the elimination of 29. F. Gergely, V. M. Draviam, J. W. Raff, The ch-TOG/XMAP215 arrest. N. Engl. J. Med. 374, 223–232 (2016). doi: 10.1056/
the female gamete centrosome in Drosophila melanogaster. protein is essential for spindle pole organization in human NEJMoa1510791; pmid: 26789871
Science 353, aaf4866 (2016). doi: 10.1126/science.aaf4866; somatic
pmid: 27229142 cells. Genes Dev. 17, 336–341 (2003). doi: 10.1101/gad.245603;
11. G. Manandhar, H. Schatten, P. Sutovsky, Centrosome reduction pmid: 12569123 AC KNOWLED GME NTS
during gametogenesis and its significance. Biol. Reprod. 72, 30. C. H. Lin, C. K. Hu, H. M. Shih, Clathrin heavy chain mediates We thank all the patients and volunteers who participated in
2–13 (2005). doi: 10.1095/biolreprod.104.031245; TACC3 targeting to mitotic spindles to ensure spindle stability. this study. We thank J. Dai from the Shanghai Academy of
pmid: 15385423 J. Cell Biol. 189, 1097–1105 (2010). doi: 10.1083/ Agricultural Sciences for providing the porcine ovaries. We
12. D. L. Gard, Microtubule organization during maturation of jcb.200911120; pmid: 20566684 thank C. Zhang from Peking University for the gift of the human
Xenopus oocytes: Assembly and rotation of the meiotic 31. S. Petry, R. D. Vale, Microtubule nucleation at the centrosome TACC3 construct. We thank K. Qiao from the Cell Biological
spindles. Dev. Biol. 151, 516–530 (1992). doi: 10.1016/0012- and beyond. Nat. Cell Biol. 17, 1089–1093 (2015). doi: 10.1038/ Imaging Core Facility of IMIB at Fudan University for assistance
1606(92)90190-R; pmid: 1601183 ncb3220; pmid: 26316453 with imaging. We are grateful to X. Li (ZEISS) for technical
13. M. Łuksza, I. Queguigner, M. H. Verlhac, S. Brunet, Rebuilding 32. A. Thawani, R. S. Kadzik, S. Petry, XMAP215 is a microtubule support with confocal live-oocyte time-lapse imaging. Funding:
MTOCs upon centriole loss during mouse oogenesis. Dev. Biol. nucleation factor that functions synergistically with This work was supported by the National Key Research and
382, 48–56 (2013). doi: 10.1016/[Link].2013.07.029; the g-tubulin ring complex. Nat. Cell Biol. 20, 575–585 Development Program of China (2021YFC2700100), the Basic
pmid: 23954884 (2018). doi: 10.1038/s41556-018-0091-6; pmid: 29695792 Science Center Program of the National Natural Science
14. M. J. Carabatsos, C. M. H. Combelles, S. M. Messinger, 33. G. J. Brouhard et al., XMAP215 is a processive microtubule Foundation of China (82288102), the National Natural Science
D. F. Albertini, Sorting and reorganization of centrosomes polymerase. Cell 132, 79–88 (2008). doi: 10.1016/ Foundation of China (81725006, 32130029, 82101737,
during oocyte maturation in the mouse. Microsc. Res. [Link].2007.11.043; pmid: 18191222 82171643, 81971450, and 81971382), the Shanghai Municipal
Tech. 49, 435–444 (2000). doi: 10.1002/(SICI)1097- 34. W. Fu et al., Self-assembly and sorting of acentrosomal Science and Technology Major Project (2017SHZDZX01), the
0029(20000601)49:5<435::AID-JEMT5>[Link];2-H; microtubules by TACC3 facilitate kinetochore capture during Project of the Shanghai Municipal Science and Technology
pmid: 10842370 the mitotic spindle assembly. Proc. Natl. Acad. Sci. U.S.A. Commission (21XD1420300), the Capacity Building Planning
15. M. Schuh, J. Ellenberg, Self-organization of MTOCs replaces 110, 15295–15300 (2013). doi: 10.1073/pnas.1312382110; Program for the Shanghai Women and Children’s Health Service,
centrosome function during acentrosomal spindle assembly in pmid: 24003142 and the collaborative innovation center project construction
live mouse oocytes. Cell 130, 484–498 (2007). doi: 10.1016/ 35. J. Cesario, K. S. McKim, RanGTP is required for meiotic spindle for Shanghai Women and Children’s Health. Ethics statement:
[Link].2007.06.025; pmid: 17693257 organization and the initiation of embryonic development in The studies involving human participants were reviewed and
16. S. J. Radford, A. L. Nguyen, K. Schindler, K. S. McKim, The Drosophila. J. Cell Sci. 124, 3797–3810 (2011). doi: 10.1242/ approved by the ethics committee of the Shanghai Ji Ai Genetics
chromosomal basis of meiotic acentrosomal spindle jcs.084855; pmid: 22100918 and IVF Institute (JIAI E2020-23 and JIAI E2022-05), the
assembly and function in oocytes. Chromosoma 126, 36. J. Dumont et al., A centriole- and RanGTP-independent spindle ethics committee of the International Peace Maternity and Child
351–364 (2017). doi: 10.1007/s00412-016-0618-1; assembly pathway in meiosis I of vertebrate oocytes. Health Hospital (GKLW 2021-18), and the ethics committee of
pmid: 27837282 J. Cell Biol. 176, 295–305 (2007). doi: 10.1083/jcb.200605199; Fudan University (FE21198 and 2020-012). The patients and
17. I. D. Wolff, M. V. Tran, T. J. Mullen, A. M. Villeneuve, pmid: 17261848 participants provided their written informed consent to
S. M. Wignall, Assembly of Caenorhabditis elegans 37. C. So, S. Cheng, M. Schuh, Phase separation during participate in this study. The animal study was reviewed and
acentrosomal spindles occurs without evident microtubule- germline development. Trends Cell Biol. 31, approved by the animal welfare and ethics group of the
organizing centers and requires microtubule sorting by 254–268 (2021). doi: 10.1016/[Link].2020.12.004; Department of Experimental Animal Science of Fudan University
KLP-18/kinesin-12 and MESP-1. Mol. Biol. Cell 27, pmid: 33455855 (2017 1350 A265). Author contributions: T.W., J.D., J.F., and
3122–3131 (2016). doi: 10.1091/mbc.e16-05-0291; 38. A. Webster, M. Schuh, Mechanisms of aneuploidy in human Y.K. contributed equally to this work. L.W., Q.S., and X.S.
pmid: 27559133 eggs. Trends Cell Biol. 27, 55–68 (2017). doi: 10.1016/ conceived of and designed the research study. T.W. designed
18. D. Clift, M. Schuh, A three-step MTOC fragmentation [Link].2016.09.002; pmid: 27773484 the experiments and the methods for data analysis. T.W., J.D.,
mechanism facilitates bipolar spindle assembly in mouse 39. S. Doubilet, K. S. McKim, Spindle assembly in the oocytes of Y.K., and H.G. performed the experiments. Y.L. and R.G. helped
oocytes. Nat. Commun. 6, 7217 (2015). doi: 10.1038/ mouse and Drosophila—similar solutions to a problem. with the molecular work and genome sequencing. J.F., M.Z.,
ncomms8217; pmid: 26147444 Chromosome Res. 15, 681–696 (2007). doi: 10.1007/s10577- and W.L. helped with human oocyte sample collection and
19. T. Wu, S. I. R. Lane, S. L. Morgan, K. T. Jones, Spindle tubulin 007-1148-8; pmid: 17674154 manipulation. X.D. helped with human oocyte in vitro
and MTOC asymmetries may explain meiotic drive in oocytes. 40. I. Peset et al., Function and regulation of Maskin, a TACC maturation. B.C. organized the medical records and analyzed
Nat. Commun. 9, 2952 (2018). doi: 10.1038/s41467-018- family protein, in microtubule growth during mitosis. J. Cell the whole-exome data. L.W., Q.S., and T.W. drafted and revised
05338-7; pmid: 30054463 Biol. 170, 1057–1066 (2005). doi: 10.1083/jcb.200504037; the manuscript. Competing interests: The authors declare no
20. C. Gueth-Hallonet et al., g-Tubulin is present in acentriolar pmid: 16172207 competing interests. Data and materials availability: All data
MTOCs during early mouse development. J. Cell Sci. 105, 41. I. Peset, I. Vernos, The TACC proteins: TACC-ling microtubule in the paper are present in the paper or the supplementary
157–166 (1993). doi: 10.1242/jcs.105.1.157; pmid: 8360270 dynamics and centrosome function. Trends Cell Biol. 18, materials. In addition, the variation data reported in this paper

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 11 of 12

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

have been deposited in the Genome Variation Map of the to original US government works. [Link] MDAR Reproducibility Checklist
National Genomics Data Center, China National Center for about/science-licenses-journal-article-reuse Movies S1 to S5
Bioinformation/Beijing Institute of Genomics, Chinese Academy
of Sciences, under accession numbers GVM000402 (patients) SUPPLEMENTARY MATERIALS
and GVM000394 (controls). License information: Copyright © [Link]/doi/10.1126/science.abq7361 Submitted 27 April 2022; resubmitted 24 August 2022
2022 the authors, some rights reserved; exclusive licensee Figs. S1 to S11 Accepted 14 October 2022
American Association for the Advancement of Science. No claim Tables S1 and S2 10.1126/science.abq7361

Wu et al., Science 378, eabq7361 (2022) 18 November 2022 12 of 12

RES EARCH

◥ deficient hosts, indicating that FMRP was not

RESEARCH ARTICLE SUMMARY involved in stimulating tumor growth per se.
By contrast, tumor growth was impaired and
CANCER survival extended in immunocompetent hosts,
implicating the adaptive immune system. Indeed,
Aberrant hyperexpression of the RNA binding protein FMRP-expressing WT tumors were largely devoid
of T cells, whereas FMRP-KO tumors were highly
FMRP in tumors mediates immune evasion inflamed. Depletion of CD8 and CD4 T cells re-
stored tumor growth and reduced survival, impli-
Qiqun Zeng†, Sadegh Saghafinia†, Agnieszka Chryplewicz†, Nadine Fournier, Lucine Christe, cating FMRP in immune evasion in WT tumors.
Yu-Qing Xie, Jeremy Guillot, Simge Yucel, Pumin Li, José A. Galván, Eva Karamitopoulou, Inti Zlobec, WT and FMRP-KO tumors were profiled by
Dalya Ataca, Fleuriane Gallean, Peng Zhang, José Antonio Rodriguez-Calero, Mark Rubin, single-cell RNA sequencing, revealing marked
Mélanie Tichet, Krisztian Homicsko, Douglas Hanahan* differences in genome-wide transcription and
abundance of cancer cells, macrophages, and
T cells. To elucidate the effects of this multifaceted
INTRODUCTION: Cancer biology and therapy and metastatic. We investigated the expression regulatory protein, we performed several func-
have been transformed by knowledge about of FMRP in human tumors, further assessed its tional perturbations, revealing that: FMRP-
immunoregulatory mechanisms that govern tumor-promoting functions in mouse models expressing cancer cells produce the chemokine
adaptive immunity. Although some forms of of cancer, and evaluated its association with interleukin-33 (IL-33), which induces regulatory
treatment resistance are related to the inten- prognosis for human cancer patients. T cells, as well as tumor-secreted protein S
tionally transitory operations of the adaptive (PROS1) ligand and exosomes that elicit tumor-
immune system, others reflect more subtle re- RESULTS: When human tumor tissue microar- promoting (M2) macrophages. Both cell types
quirements to modulate the immune system in rays were immunostained for expression of are immunosuppressive, collectively contribut-
different contexts. In this work, we identified FMRP, a majority of tumors expressed FMRP, ing to the barrier against T cell attack. By
an immunoregulatory mechanism involving the whereas cognate normal tissues did not. To in- contrast, FMRP-KO cancer cells down-regulate
neuronal RNA binding protein fragile X mental vestigate the functional significance of this broad all three factors and up-regulate C-C motif
retardation protein (FMRP), which broadly reg- up-regulation, the FMR1 gene was ablated through chemokine ligand 7 (CCL7), which helps recruit
ulates protein translation and mRNA stability CRISPR-Cas9 gene editing (FMRP-KO, where and activate T cells. Additionally, immunostimu-
and is aberrantly up-regulated in multiple forms KO indicates knockout) in mouse cancer cell latory macrophages develop in this context that
of cancer. lines that were inoculated into both immuno- express three proinflammatory chemokines—
deficient and syngeneic immunocompetent mice CCL5, CXCL9, and CXCL10—which cooperate
RATIONALE: This study was motivated by re- to establish tumors in parallel with wild-type with CCL7 in recruiting T cells. Finally, neither
ports that cancer cells naturally overexpressing (WT) FMRP-expressing cell lines. Mice bearing FMR1 mRNA nor FMRP protein levels were
FMRP, whose loss of expression in developing FMRP-KO tumors had similar survival com- sufficient to predict outcomes in cohorts of
neurons causes cognitive defects, were invasive pared with isogenic WT tumors in immuno- cancer patients. Recognizing FMRP’s function
as an RNA binding protein that modulates
mRNA stability and hence levels in transcrip-
Il-33 Tregs tome datasets, a gene signature reflecting
(no) CD8 T cells
FMRP’s cancer regulatory activity (involving
Cancer cell 156 genes) was developed by comparing FMRP-
hyperexpressing expressing versus FMRP-deficient cancer cells,
FMRP both in culture and within tumors. Our FMRP
Pros1 cancer activity signature was prognostic for
+ M2-inducing
survival across multiple human cancers; anti-
exosomes
correlated with the intensity of T cell infil-
Immunosuppressive Excluded
and suppressed tration in different tumor types, consistent
macrophages
CD8 T cells with FMRP’s immunosuppressive effects;
and was associated with comparatively poor
CD8 T cells
responses to immune checkpoint inhibitors
Ccl7 and immune-dependent chemotherapy in se-
Cancer cell Ccl5
with low or Cxcl9 lected cohorts.
Cxcl10
absent FMRP
Down-regulation CONCLUSION: FMRP is revealed as a regulator
of Pros1 and of a network of genes and cells in the tumor
M2-inducing
microenvironment that contribute to the capa-
exosomes
Immunostimulatory
macrophages
Recruited
and activated
bility of tumors to evade immune destruction.
▪
CD8 T cells
The list of author affiliations is available in the full article online.
FMRP enables tumors to evade being attacked by the immune system. Up-regulated FMRP expression *Corresponding author. Email: [Link]@[Link]
Cite this article as Q. Zeng et al., Science 378, eabl7207
and activity are involved in an immunosuppressive program in cancer cells (upper left) that renders tumors
(2022). DOI: 10.1126/science.abl7207
impenetrable, creating so-called immune deserts (upper right). By contrast, its absence is associated These authors contributed equally to this work.
with a reprogrammed tumor microenvironment that recruits and activates T lymphocytes, producing inflamed
tumors, including CD8 T cells (red immunostain; lower right) with consequently beneficial immune READ THE FULL ARTICLE AT
destruction. Tregs, regulatory (immunosuppressive) T cells. [Link]

746 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH

◥ mouse cancer cell lines that could be estab-

RESEARCH ARTICLE lished as tumors in immunocompetent mice
to evaluate the effects both of FMRP expres-
CANCER sion and of its attenuation or absence on the
heterotypic cell types populating the TME, in-
Aberrant hyperexpression of the RNA binding protein cluding potential interfaces with the adaptive
immune system.
FMRP in tumors mediates immune evasion Results
Qiqun Zeng1,2†, Sadegh Saghafinia1,2†, Agnieszka Chryplewicz1,3†, Nadine Fournier4, Lucine Christe5, To further substantiate the previously reported
Yu-Qing Xie6, Jeremy Guillot1,3, Simge Yucel1,3, Pumin Li1,3,7, José A. Galván5, Eva Karamitopoulou5, up-regulation of FMRP in tumors (15–19), we
Inti Zlobec5, Dalya Ataca2, Fleuriane Gallean2, Peng Zhang8, José Antonio Rodriguez-Calero9, assessed FMRP expression in human tumor tis-
Mark Rubin9, Mélanie Tichet1,3,10, Krisztian Homicsko1,3,10,11,12, Douglas Hanahan1,3,10,12* sue microarrays (TMAs) and in cognate de novo
mouse tumors from the pancreas, colon, breast,
Many human cancers manifest the capability to circumvent attack by the adaptive immune system. In and melanomas compared with normal tissue.
this work, we identified a component of immune evasion that involves frequent up-regulation of fragile X Consistent with the aforementioned studies,
mental retardation protein (FMRP) in solid tumors. FMRP represses immune attack, as revealed by FMRP is up-regulated in major subsets of hu-
cancer cells engineered to lack its expression. FMRP-deficient tumors were infiltrated by activated T cells man pancreatic ductal adenocarcinoma (PDAC),
that impaired tumor growth and enhanced survival in mice. Mechanistically, FMRP’s immunosuppression colon carcinoma, and triple-negative breast
was multifactorial, involving repression of the chemoattractant C-C motif chemokine ligand 7 (CCL7) cancer (TNBC) (Fig. 1, A, C, and E) as well as in
concomitant with up-regulation of three immunomodulators—interleukin-33 (IL-33), tumor-secreted genetically engineered mouse models of PDAC,
protein S (PROS1), and extracellular vesicles. Gene signatures associate FMRP’s cancer network with colon cancer, breast cancer, and melanoma
poor prognosis and response to therapy in cancer patients. Collectively, FMRP is implicated as a (Fig. 1, B, D, and F, and fig. S1, A and B). By
regulator that orchestrates a multifaceted barrier to antitumor immune responses. contrast, it is lowly or not detectably expressed
in the cognate normal human and mouse tis-

C
sues (left panels of Fig. 1, A to G; fig. S1, A and B;
ancers frequently develop the ability to translational regulatory protein, identified as and fig. S2). FMRP is also up-regulated in a
evade destruction by the immune sys- a regulator of synaptic plasticity and architec- majority of human hepatocellular carcinomas
tem. Therapeutic strategies aimed at dis- ture, which is an essential component of neu- (HCCs) and all analyzed primary and brain-
rupting barriers to tumor immunity are ronal function (4). metastatic prostate cancers, in contrast to
producing noteworthy results and, in FMRP is expressed at high levels in neuro- normal or benign tissue (Fig. 1, G and H). An
some cases, cures for patients with certain can- nal synapses and at variably lower levels in independent investigation of cancer patients
cers. Most patients, however, do not respond to organs throughout the body ([Link] by the Human Protein Atlas (20) reported that
immunotherapies, or the benefit is only tran- [Link]/ENSG00000102081-FMR1/ FMRP is variably up-regulated in most forms
sient (1–3), which highlights the importance tissue#top). Its genetic or epigenetic inactiva- of human solid tumors (fig. S1C). In particular,
of understanding the cellular and molecular tion is the primary cause of the neurodevel- we found that FMRP was highly up-regulated
mechanisms underlying immune responses opmental disorder fragile X syndrome (FXS), in most but not all tumor samples represented
to cancer. which has instigated extensive investigations in the analyzed TMAs (boxplots in Fig. 1, A, C, E,
We describe the discovery and charac- into the molecular biology of this protein. His- G, and H). To substantiate the magnified fields
terization of an immune barrier that is di- torically, FMRP had been established as a trans- shown in Fig. 1, the entirety of representative
versely erected in the tumor microenvironments lational repressor (5–7) and, more recently, had 0.6-mm cores from the analyzed human TMAs
(TMEs) of human and experimental tumors as been recognized as a regulator of gene expres- are shown in fig. S2, illustrating a normal tis-
a consequence of up-regulating the fragile X sion, in particular by modulating abundance sue core, three FMRP-positive tumor cores, and
mental retardation protein (FMRP) (encoded (through transcription and stability) of hun- one FMRP-negative tumor core. Analysis of the
by the FMR1 gene). FMRP is an RNA binding dreds of cellular mRNAs that play roles in largest patient cohort profiled by the Cancer
neuronal function (7, 8). Although mice and Genome Atlas (TCGA) (21) revealed no evidence
humans lacking FMRP are viable and osten- that the observed up-regulation of FMRP in
1
Swiss Institute for Experimental Cancer Research (ISREC), sibly normal beyond prominent defects in the human tumors commonly involves mutation
School of Life Sciences, Swiss Federal Institute of central nervous system (9–11), the role(s) and or amplification of the FMR1 gene that en-
Technology Lausanne (EPFL), 1015 Lausanne, Switzerland.
2
Opna Bio SA, Biopole, 1066 Epalinges, Lausanne, functional importance of lower-level FMRP codes it (fig. S1, D and E). This might suggest
Switzerland. 3Agora Cancer Research Center, 1011 Lausanne, expression in other organs are not well de- that FMRP’s up-regulation is directly or in-
Switzerland. 4Swiss Institute of Bioinformatics (SIB), 1015 fined. Notably, there have been descriptive clues directly mediated by the neoplastic state.
Lausanne, Switzerland. 5Institute of Pathology, University of
Bern, 3008 Bern, Switzerland. 6Institute of Bioengineering,
(12–14) but no substantive documentation that We have previously shown that glutamate-
School of Engineering, Swiss Federal Institute of Technology FMRP might play an immunoregulatory role stimulated N-methyl-D-aspartate receptor (NMDAR)
Lausanne (EPFL), 1015 Lausanne, Switzerland. 7Department for T cells. signaling was one such inducer of FMRP ex-
of Computational Biology, University of Lausanne, 1015
Lausanne, Switzerland. 8Beijing Pediatric Research Institute,
Several studies have previously described pression (15). However, the NMDARs are only
Beijing Children’s Hospital, Capital Medical University, the involvement of FMRP in tumor progres- up-regulated in a subset of tumor types com-
National Center for Children’s Health, Beijing 100045, China.
9
sion, in particular stimulating invasion and pared with FMRP (fig. S3, A and B), which im-
Department for BioMedical Research, University of Bern,
metastasis in certain tumor types (15–17), plicates other regulatory mechanisms in cancers.
3008 Bern, Switzerland. 10Lausanne Branch, Ludwig Institute
for Cancer Research, 1011 Lausanne, Switzerland. which led us to further assess its functional To explore whether transcriptional regulation
11
Department of Oncology, University Hospital of Lausanne roles through gene knockout (KO) and gene of FMRP is involved in its up-regulation in can-
(CHUV), 1011 Lausanne, Switzerland. 12Swiss Cancer Center knockdown (KD) in cultured cancer cells and cer, the ENCODE database (22) of chromatin
Leman (SCCL), 1011 Lausanne, Switzerland.
*Corresponding author. Email: [Link]@[Link] tumors growing in mice. Notably, our research immunoprecipitation sequencing (ChIP-seq)
†These authors contributed equally to this work. design included the use of FMRP-expressing analyses for transcription factors (TFs) binding

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 1 of 23

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

A Human Pancreas B Mouse Pancreas

P < 0.0001 Normal PDAC P < 0.0001
100

FMRP expression score

FMRP+ ductal / tumor

Normal PDAC 250
200 80

cells (%)
150 60
100 40
50 20
0 0
Normal PDAC Normal PDAC
(n=100) (n=108)
C Human Colon 400 P < 0.0001 D Mouse Colon

FMRP expression score

P < 0.0001
Normal CRC Normal APC 100

FMRP+ colon / tumor

300 80

cells (%)
200 60
40
100
20
0 0
Normal CRC Normal APC
(n=153) (n=118)

E Human Breast
FMRP expression score 400 F
P < 0.0001 Mouse Breast

FMRP+ Breast / TNBC

Normal TNBC 150
Normal TNBC
300 P < 0 .0001
100

cells (%)
200

50
100
0
0 TNBC
Normal TNBC
(n=5) (n=241)

G H Human Prostate

Benign
Human Liver
Normal HCC 300 400 P < 0.0001

FMRP expression score

P < 0.0001
P < 0.0001
300
200
200

100 Prostate Cancer Brain metastasis 100

0 0
Normal HCC ) ) )
=9 =7 11
(n=3) (n=115) (n y (n ( n=
n ar s
nig im et
Be Pr in
m
a
Br

Fig. 1. FMRP is up-regulated in diverse human tumors and cognate de novo (D) Mouse colon tumors and normal colon from the ApcD/D; CDX2 Cre ERT2 (APC)
mouse tumors. Representative IHC images and quantification of FMRP mouse model in the FVB/N background (83) For each mouse (N = 5 per group),
expression in the indicated human and mouse tumor tissues and normal tissue five fields from two tissue sections from one independent tumor were analyzed.
controls. For the human TMAs, the FMRP expression score = [percentage of Statistics by unpaired t test. Scale bar, 50 mm. (E) Human triple-negative breast
FMRP-positive cells in an entire core section (0 to 100) × staining intensity cancer (TNBC) in a TNBC TMA (N = 241) and normal breast tissue sections
(0 to 3)]. See the Materials and methods for details; scoring was performed (N = 5, from reduction mammaplasties), analyzed as in (A). Statistics by Mann-
independently by two individuals and reconciled. Data are presented as means ± Whitney test. Scale bar, 50 mm. (F) Mouse breast and TNBC-like breast tumors
SEMs. (A) Human PDAC tumors and tumor-adjacent normal human arising in the C3-Tag mouse model (84) in the FVB/N background. For each
pancreas in 0.6-mm-diameter core sections of a TMA. N = 108 PDAC tumors mouse (N = 5 per group), five fields from two tissue sections from one independent
and N = 100 normal pancreases were analyzed. Statistics were by Mann-Whitney tumor were analyzed. Statistics by unpaired t test. Scale bar, 50 mm. (G) Human
test. Scale bar, 50 mm. (B) Mouse PDAC tumors and normal pancreas from hepatocellular carcinoma (HCC) and normal liver in a TMA; N = 115 HCC tumor cores
the P48-cre; LSL-KrasG12D; P53R172H/+ PDAC mouse model in the FVB/N and N = 3 normal livers cores; analyzed as in (A). Statistics by one-sample Wilcoxon
background (82). N = 3 mice per group. For each mouse, five fields from two test. Scale bar, 50 mm. (H) Human primary prostate carcinoma, brain-metastatic
tissue sections from one independent tumor were analyzed. Statistics by prostate carcinoma, and normal or benign human prostate tissue in a TMA.
unpaired t test. Scale bar, 50 mm. (C) Human colorectal tumors (CRCs) in tumor- N = 7 primary prostate tumor cores, N = 11 brain-metastatic prostate tumor cores,
adjacent normal colon and in a TMA, analyzed as in (A). N = 118 CRCs and and N = 8 normal pancreas cores; analyzed as in (A). Statistics by one-way analysis
N = 153 normal colons. Statistics by Mann-Whitney test. Scale bar, 50 mm. of variance (ANOVA) with Kruskal-Wallis test. Scale bar, 50 mm.

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 2 of 23

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

to DNA elements was queried for ChIP-seq parallel into syngeneic and immunodeficient bearing PDAC FMRP-KO2 tumors were treat-
peaks in the FMR1 promoter region (see Ma- mice. Tail vein injections were performed to ed with monoclonal antibodies that deplete
terials and methods). The MYC TF, one of seed the lungs followed by survival analysis CD8 or CD4 T cells (Fig. 3G), which markedly
the most widely up-regulated oncogenes across (Fig. 2A), and subcutaneous (s.c.) inoculations reduced the respective abundance of CD8
different human cancers (23) (fig. S3C), as well were used to assess tumor growth (Fig. 2D). and CD4 T cells in the blood (fig. S6, C and D)
as its transcriptional regulatory partner MAX Unexpectedly, differences in phenotypes were and of CD8 T cells in the spleen and tumor
were among the top TFs with significant peaks observed in immunodeficient versus immuno- (fig. S6, E and F). FMRP-KO tumor growth was
at the FMR1 promoter in a variety of cancer competent mice bearing experimental lung markedly increased in the mice treated with
cell lines (fig. S3, D and E). We next queried a metastases and primary s.c. tumors. The WT2 the depleting anti-CD8 or anti-CD4 antibodies,
recently published dataset (24) involving a and FMRP-KO2 PDAC cells had similar meta- and more so in combination (Fig. 3H), whereas
switchable genetic mouse model of PDAC, static survival (Fig. 2B) and similar primary their survival was commensurately reduced
whereby Myc expression can be activated in tumor growth (Fig. 2E) in immunodeficient (Fig. 3I), which functionally implicates CD8 and
indolent KrasG12D-induced pancreatic intraepi- SCID/beige mice. In marked contrast, both CD4 T cells in the FMRP-KO tumor phenotype.
thelial neoplasias (PanINs), which results in metastatic and primary tumor survival were In marked contrast, the T cell–depleting anti-
the progression of PanIN to adenocarcinoma extended (Fig. 2, C and G) and lung metastatic bodies had no effect on survival or growth of
(fig. S3F). Consistently, upon Myc activation, tumor burden impaired (Fig. 2F and fig. S5, WT2 PDAC tumors (fig. S6G).
Fmr1 mRNA was significantly up-regulated in A to D) for the FMRP-KO cells inoculated Additional evidence for the functional im-
PDAC tumors (fig. S3, G and H), encouraging into syngeneic, immunocompetent FVBN mice. portance of CD8 T cells came from blocking
the hypothesis that MYC is a regulator of FMRP To address the possibility that these differences the programmed cell death protein–1 (PD-1)
expression in cancer. Additionally, to further were KO clone specific, we analyzed a second immune checkpoint, based on the observa-
assess this possibility, we evaluated Fmr1 mRNA pair of WT (WT3) and FMRP-KO (KO8) PDAC tion that WT and FMRP-KO cells and tumors
as well as FMRP protein expression levels upon clones, which revealed similarly impaired tu- similarly expressed the checkpoint ligand pro-
Myc KD with small interfering RNA (siRNA) mor growth in immunocompetent mice (Fig. 2, grammed cell death ligand–1 (PD-L1) (fig. S6,
in an FMRP-expressing mouse PDAC cancer H and I). Analogous phenotypic differences in H and I). Consistent with the dearth of infil-
cell line, further described below. Concordant- primary tumor growth were also observed com- trating CD8 T cells in mice bearing WT2 tu-
ly, Myc KD resulted in a significant reduction paring the WT versus FMRP-KO CT26 colon mors, treatment with anti–PD-1 had no effect
of FMRP expression both at mRNA and pro- cancer cells (Fig. 2, J to N). Additionally, an (Fig. 3J), phenocopying the lack of response in
tein levels (fig. S3, I to K). Collectively, these Fmr1-KO was engineered into the B16-OVA human PDAC (26, 27). By contrast, tumor growth
data implicate the MYC oncogene in the broad melanoma cell line (fig. S5E) and the 4T1 breast was further impaired when mice bearing FMRP-
up-regulation of FMRP across the spectrum of cancer cell line (fig. S5I). Paired isogenic WT KO2 tumors were treated with anti–PD-1 anti-
solid tumors (fig. S3). and FMRP-KO single-cell clones were eval- body, concomitant with inoculation of the
uated, again revealing impaired tumor growth cancer cells to seed tumors (Fig. 3K), or after
Assessing FMRP functions by its genetic in immunocompetent mice for the FMRP-KO macroscopic solid tumors had formed (Fig. 3L).
deletion in cancer cells clones in these cases (fig. S5, F to H and J to L). To assess possible effects of FMRP-KO can-
Aiming to extend upon previous functional Collectively, the data demonstrate that tumors cer cells on the clonal expansion of T cells in
studies involving short hairpin RNA (shRNA) derived from FMRP-KO cancer cells grow more tumors, DNA was extracted from WT2 and
KD of FMR1 in cancer cells (15–19), we used slowly in immunocompetent but not in immu- FMRP-KO2 PDAC tumors and subjected to se-
CRISPR-Cas9 technology to genetically delete nodeficient mice, which suggests an involve- quencing analysis for the T cell receptor (TCR)
the FMR1 gene in a series of single-cell cloned ment of the adaptive immune system. Vß variable region, much as recently described
mouse cancer cell lines, beginning with the (28). The analysis revealed both clonal expan-
PDAC line 4361.12 (15) and the colon cancer Evaluating the effects of the adaptive immune sion and markedly increased diversity of TCRs
line CT26 (25). For the analyses, we selected system in WT versus FMRP-KO tumors in the FMRP-KO tumors (fig. S7A), consistent
two PDAC Fmr1-KO clones and one CT26 Fmr1- To assess the possibility that FMRP loss was with a productive antitumor immune response.
KO clone that were devoid in expression of triggering an adaptive immune response, we Additionally, we assessed the relative immuno-
FMRP protein (fig. S4, A, B, and I). The FMRP- analyzed WT and FMRP-KO tumors for the genicity of a neoantigen—gp70—expressed in
KO had no discernable effect on proliferative presence of CD8 and CD4 T lymphocytes. The the CT26 colon cancer cell line (29). Splenocytes
phenotypes of the PDAC or CT26 cancer cells WT PDAC tumors (clones WT2 and WT3) and collected from mice bearing tumors derived by
in culture compared with their otherwise iso- the WT CT26 colon tumors (WT12) were im- s.c. inoculation of WT and FMRP-KO CT26 cell
genic wild-type (WT) progenitors (fig. S4, C mune deserts, largely devoid of CD8 T cells clones were stimulated in culture with gp70
to E, J, and K). The levels of cell viability (fig. (Fig. 3, A and B). By contrast, the isogenic FMRP- protein, followed by FACS analysis with a gp70-
S4D) and cell proliferation (fig. S4E) were sim- KO tumors were highly inflamed by CD8 T cells specific tetramer (30, 31). The FMRP-KO tumors
ilar; the modestly increased rate of apoptosis (Fig. 3, A and B). Fluorescence-activated cell elicited increased numbers of tetramer-positive
seen in culture with the KO2 line compared sorting (FACS) analysis of tumors derived from antigen-specific CD8 T cells (fig. S7, B and C),
with the WT and KO8 lines (fig. S4F) is not re- the two sets of WT and FMRP-KO PDAC cell further documenting the immunostimulatory
capitulated in tumor growth assays (fig. S4G) lines confirmed the differential abundance of effects of the TME created by cancer cells lack-
and hence was judged inconsequential. Next, both CD8 and CD4 T cells (Fig. 3C and fig. S6A). ing FMRP.
confirming the results of the Fmr1 siRNA KD Notably, the CD8 T cells infiltrating the FMRP- In addition to its expression in neurons and
analysis performed previously (15), the FMRP- KO tumors expressed markers indicative of many cancer cells, FMRP is also expressed in
KO2 line proved to be severely impaired in a their activation as cytotoxic T lymphocytes a variety of normal nonneuronal cell types in
transwell migration assay, a metric of invasive- (CTLs) (Fig. 3, D to F, and fig. S6B). many tissues ([Link]
ness (fig. S4H). Given that the FMRP-KO tumors were highly ENSG00000102081-FMR1/tissue), albeit at gen-
To probe the functional effects of the FMRP- inflamed by activated CTLs, we next evaluated erally lower levels than in neurons. Therefore,
KO on tumor phenotypes in vivo, the WT2 and their functional contributions to the impaired we assessed possible FMRP expression originat-
FMRP-KO2 PDAC cell lines were inoculated in tumor growth. First, immunocompetent mice ing from the stromal compartment of inflamed

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 3 of 23

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

Fig. 2. Genetic deletion of FMRP A Experimental design B Survival of WT vs. FMRP-KO PDAC tumors C Survival of WT vs. FMRP-KO PDAC tumors
in cancer cells impairs tumor Tail-vein injection SCID/beige FVB/N
100 WT2 (n=5)
growth and improves survival in 100 WT2 (n=5)
FMRP KO2 FMRP KO2
immunocompetent but not (n=5)

Percent survival
Percent survival
(n=5)
Survival
immunodeficient mice. (A) Sche-
matic of the lung metastasis assay 50 50
for PDAC WT and FMRP-KO cancer IV
P = 0.0018 P = 0.0027
cells (clones WT2 and KO2 from
2X105 4361.12
PDAC cell line 4361.12). In total, 2 × WT2 or FMRP KO2 cells 0 0
0 50 100 150 0 50 100 150
105 cells in 200 ml of PBS were Days after injection Days after IV injection
injected into the tail vein of immu-
nodeficient SCID/beige or immu- D Experimental design E Time course of PDAC tumor growth F Time course of PDAC tumor growth
S.C. injection WT vs. FMRP-KO WT vs. FMRP-KO
nocompetent FVB/N mice. Mice 1500 WT2 (n=5) 2000 WT2 (n=5)
PDAC SCID/beige FVB/N
FMRP KO2

Tumor Volume (mm³)

were monitored twice per week for (n=5) FMRP KO2
1500 (n=5)
long-term survival. IV, intravenous. 1000
P = 0.0454
(B and C) Overall survival (OS) S.C.
1000
Day 25 P = 0.0016
of immunodeficient SCID/beige (B) 500
or immunocompetent FVB/N (C) 500

mice bearing lung metastases 5X105 4361.12

WT2 or FMRP KO2 cells 0 0
seeded with PDAC WT or FMRP KO 0 5 10 15 20 25 0 5 10 15 20 25
cells. N = 5 mice per group; one Days after injection Days after injection
independent experiment. Statistics G Survival of WT vs. FMRP-KO PDAC tumors H Tumor burden for PDAC tumors I Tumor burden for PDAC tumors
by Kaplan-Meier test. (D) Sche- FVB/N WT vs. FMRP-KO, NSG WT vs. FMRP-KO, FVB/N
matic of the s.c. primary tumor WT2 (n=6)
P = 0.5749
P = 0.5339 P = 0.0188
growth assay after inoculation of 100 FMRP KO2 (n=10) P = 0.0039
2.0 P = 0.1318 2.0
Percent survival

PDAC FMRP-WT and FMRP-KO P = 0.8338

Tumor weight (g)

cancer cells (clones WT2 and KO2

Tumor weight (g)

1.5 1.5
from line 4361.12). In total, 5 × 105 50 P < 0.0001
1.0 1.0
cells in 100 ml of PBS were s.c.
injected into the right flank of 0.5 0.5
0
immunodeficient SCID/beige or 0 20 40 60 80 100
0.0 0.0
immunocompetent FVB/N mice. On Days after injection WT2 WT3 KO2 KO8 WT2 WT3 KO2 KO8
day 25, all mice were euthanized,
and tumors were collected for J Experimental design K Time course of colon tumor growth L Time course of colon tumor growth
S.C. injection WT vs. FMRP-KO WT vs. FMRP-KO
2000
further analysis. (E and F) Tumor CRC NSG WT17 (n=5) Balb/c WT17 (n=9)
1500

Tumor Volume (mm3)

growth curves of resultant WT and FMRP KO12 FMRP KO12
(n=5) 1500 (n=10)
FMRP-KO PDAC tumors in SCID/
Tumor Volume (mm3)

S.C.
Day 28 or P = 0.3168
beige (E) or FVB/N (F) mice. N = 5 Day 14
1000
1000
mice per group. Statistics by P = 0.0084
unpaired t test. Tumor volumes 5X105 CT26
500
500
were measured twice per week and WT17 or FMRP KO12 cells
0 0
calculated using the standard 0 5 10 15 0 10 20 30
formula, V = 0.52 × length × width . 2
Days after injection Days after injection
Data are representative of two M Tumor burden for colon tumors N Tumor burden for colon tumors
independent experiments and are WT vs. FMRP-KO, NSG
4
WT vs. FMRP-KO, Balb/c
3
presented as means ± SEMs. P = 0.9047 P = 0.0252
Tumor weight (g)

Tumor weight (g)

(G) OS of immunocompetent FVB/N 3

2
mice bearing s.c. tumors seeded
with PDAC WT2 or FMRP KO2 cells. 2
N = 6 mice for the WT group and 1
1
N = 11 for the FMRP KO2 group;
one independent experiment. Log- 0 0
rank test was used for survival WT17 FMRP KO12 WT17 FMRP KO12
analyses. (H and I) Tumor weights
are shown for PDAC WT tumors (formed by two independent clones, WT2 and WT3) and for FMRP-KO tumors (formed by two independent clones, KO2 and
KO8) growing in NSG (H) or FVB/N (I) mice at day 25 after s.c. injection. N = 4 to 10 mice per group; one independent experiment. Data are presented as means ±
SEMs. Statistics by unpaired t test. (J) Schematic of primary tumor growth assay involving FMRP-WT and FMRP-KO CT26 colon cancer cells. In total, 5 × 105
cells in 100 ml of PBS were s.c. injected into the right flank of immunodeficient NSG or immunocompetent Balb/c mice. On day 14 (for NSG mice) and day 28
(for Balb/c mice), all the mice were euthanized, and tumors were collected for further analysis. (K and L) Growth curves of resultant WT and FMRP-KO CT26 tumors
(from clones WT17 or FMRP KO12) in NSG (K) or Balb/c (L) mice. N = 9 mice for the WT17 group and N = 10 for the FMRP-KO12 group. Statistics by unpaired
t test. Tumor volumes were measured twice per week, as above. Data are representative of two independent experiments and are presented as means ± SEMs.
(M and N) Tumor weights for CT26 WT17 versus FMRP-KO12 cells transplanted into NSG (M) or Balb/c (N) mice. Cells were injected s.c., as described in Fig. 2G.
N = 5 NSG mice for both the WT17 and FMRP-KO12 groups, N = 8 Balb/c mice for the WT group, and N = 9 Balb/c mice for the FMRP-KO12 group. Statistics
by unpaired t test. Data are representative of two independent experiments and are presented as means ± SEMs.

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 4 of 23

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

A Quantification of CD8 T-cell infiltration in PDAC tumors

B Quantification of CD8 T-cell infiltration in colon tumors
PDAC: DAPI/CD8 CRC: DAPI/CD8
WT2 WT3 WT17

Number of CD8+ cells per field

Number of CD8+ T cells per filed

80 100

80
60
60
KO2 KO8 40 KO12
40
20 20

0 0
WT17 FMRP KO12

8
KO

KO
W

W
PDAC TME characterization using FACS
C CD8 D IFN E GZMB F TNF

50 60 60 40

40
30

%TNF +/CD8+
%GZB+/CD8+
%IFN +/CD8+
%CD8+/Live

40 40
30
20
20
20 20
10
10

0 0 0 0

8
T2

T3
T2

KO
KO

W
W

G Experimental design H Time course of PDAC tumor growth

I Survival of PDAC tumors
S.C. injection treated with CD4/CD8-depleting antibodies treated with CD4/CD8-depleting antibodies
PDAC
KO2 IgG
1200 100 WT2 IgG
WT2 IgG
Tumor Volume (mm3)

Antibody 1000 KO2 aCD4 KO2 IgG

Percent survival
treatment KO2 aCD4 aCD8 KO2 aCD4
800
S.C. KO2 aCD8 KO2 aCD4 aCD8
Day 28 600 50 KO2 aCD8
400

5X105 WT2 or 200

FMRP KO2 cells 0
10 20 30 40 0 5 10 15 20 25 30 35 40 45
Days after injection Days after injection

Time course of PDAC tumor growth upon anti-PD1 therapy

J K L KO2 IgG
WT PDAC FMRP KO PDAC 1000
800 IgG (n=7) KO2 anti-PD1
150
Tumor Volume (mm3)

anti-PD1 (n=11) WT2 IgG

800
Tumor Volume (mm3)

WT2 anti-PD1
Tumor Volume (mm3)

600
100 600
400 p <0.0001
400 **
200
50
** *
IgG (n=5) 200
*
anti-PD1 (n=5)
0 0 0
0 10 20 30 0 10 20 30 0 10 20 30
Days after injection Days after enrollment into trials
Days after injection

Fig. 3. Tumors lacking FMRP expression in cancer cells elicit inflammation each group. For each mouse, five fields from two tissue sections from one
by CD8 and CD4 T cells that are responsible for impaired growth and independent tumor were analyzed. Data are presented as means ± SEMs.
improved survival. (A) Representative IF images and quantification of CD8 Statistics by unpaired t test. (C to F) FACS analysis of the frequency of
T cells in primary s.c. tumors formed by WT and FMRP-KO PDAC cells (clones CD11b−CD3+CD8+ T cells (C) and IFN-g+– (D), GZMB+- (E), and TNF-a+– (F)
WT2, WT3, FMRP-KO2, and FMRP-KO8). Scale bar, 10 mm. For each mouse, five expressing CD8 T cells infiltrating WT and FMRP KO PDAC tumors from FVB/N
fields from two tissue sections from one independent tumor of each mouse were mice. N = 5 mice for all groups. Data are representative of two independent
analyzed. Data are presented as means ± SEMs. Statistics by one-way ANOVA experiments and are presented as means ± SEMs. Statistics by one-way ANOVA
with Tukey’s multiple comparison analysis. DAPI, 4′,6-diamidino-2-phenylindole. with Tukey’s multiple comparison analysis. (G) Schematic of the assay to assess the
(B) Representative IF images and quantification of CD8 T cells in primary s.c. functional importance of CD8 T cells for impaired tumor growth. Mice were
tumors formed by CT26 WT17 and KO12 cells. Scale bar, 10 mm. N = 3 mice for injected s.c. with WT2 or FMRP-KO2 PDAC cells, treated with control rat IgG or

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 5 of 23

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

rat anti-mouse CD8 or CD4 antibodies, 250 mg per mouse, started on the twice per week. Mice were euthanized on day 28 after the cancer cell inoculation.
same day as s.c. injection, through intraperitoneal (IP) injection, twice per week. N = 5 for both IgG and anti–PD-1 cohorts of WT2 PDAC tumors. N = 7 for the IgG
(H and I) Tumor growth (H) and survival analysis (I) for FVB/N mice bearing and N = 11 for the anti–PD-1 cohorts of FMRP-KO2 PDAC tumors. Statistics
WT2 or FMRP KO2 PDAC tumors, treated with rat IgG or CD4- and/or CD8- by unpaired t test, Data are presented as means ± SEMs of one independent
depleting antibodies. Mice were euthanized when the tumors reached 1000 mm3 experiment. (L) Intervention trial with the anti–PD-1 antibody in WT2 and FMRP
of volume. N = 6 for all groups. Tumor volumes were measured twice per week. KO2 tumors. Mice were subjected to IgG or anti–PD-1 trials beginning when
Data are represented as means ± SEMs of one independent experiment. the tumor volume reached 100 mm3. Mice were euthanized when the tumors
Statistical analysis by Mantel-Cox test. (J and K) Assessing the response of WT reached 1000 mm3 of volume or became necrotic. N = 5 for FMRP-KO2 IgG and
versus FMRP-KO2 tumors to anti–PD-1 immunotherapy. FVB/N mice bearing anti–PD-1 arms and N = 4 for WT2 IgG and anti–PD-1 arms. Statistical analysis
WT2 (J) or FMRP-KO2 PDAC tumors (K) were treated with control rat IgG by Mann-Whitney test. Data are represented as means ± SEMs of one independent
antibody or rat anti-mouse PD-1 antibody, and tumor growth was measured experiment. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

tumors growing in immunocompetent mice which were markedly smaller than those in Congruently, the abundance of infiltrating lym-
derived from inoculation of FMRP-deficient WT control tumors. Thus, we infer that both phocytes was markedly increased (Fig. 4D), sub-
cancer cells. Very low levels of FMRP protein WT and FMRP-KO cancer cells can be killed stantiating the aforementioned analysis that
were detected in tumor sections of FMRP-KO by CTLs, a result further substantiated below revealed the infiltration of CD8 and CD4 T cells
PDAC and colon tumors (fig. S7, D and E), in a cell culture bioassay. We also assessed mu- (Fig. 3A) as well as their functional involvement
whereas no FMRP expression was detected tational burden and neoantigen load of WT in the impaired growth of FMRP-KO tumors
in FMRP-KO breast tumors (fig. S7F), which versus FMRP-KO PDAC cells after 30, 60, and (Fig. 3, H, I, K, and L).
indicated that the principal source of FMRP 150 days in cell culture by exome sequencing Cluster analysis (Fig. 4, A and B, and fig. S8)
detected in WT tumors (Fig. 1 and figs. S1 and bioinformatic analysis (Materials and revealed less-pronounced variations in other
and S2) is in fact the cancer cells. methods), which revealed no significant differ- annotated cell types populating the distinc-
The evident differences in immune pheno- ences in mutational load or neoantigen abun- tive TMEs (fig. S9A), which led us to focus on
types between WT and FMRP-KO tumors dance (fig. S6, J and K) and hence in inferred further characterization of lymphocytes and to
raised an ancillary question: Does the resist- immunogenicity. identify subtypes that are differentially abun-
ance versus susceptibility to immune infiltra- Finally, to substantiate the FMRP-KO can- dant in WT versus FMRP-KO tumors. To this
tion and attack reflect dominant characteristics cer cell phenotype described in the various end, a computational algorithm for character-
of FMRP-WT or FMRP-KO cancer cells? To s.c. transplant models, we evaluated the effects izing T cell states and clusters based on refer-
address this question, we performed mixing of FMRP in the orthotopic context by implant- ence T cell atlases (ProjecTILs) (33) was applied
experiments whereby WT and FMRP-KO can- ing FMRP-KO and WT PDAC cells directly to the lymphocyte cluster in the scRNA-seq data.
cer cells were differentially labeled with red into the pancreas of immunocompetent mice. In addition to the markedly different overall
and green fluorescent proteins (RFP and GFP) Congruent with the s.c. model, FMRP-KO tu- abundance of lymphocytes (Fig. 3A and Fig.
and coinoculated into syngeneic mice at vary- mors were significantly smaller and highly in- 4D), there were appreciable qualitative differ-
ing ratios. Subsequently, tumor growth, CD8 filtrated with CD8 T cells (fig. S6, L to N). ences in the fractional representation of lym-
T cell infiltration, and the localization and phocyte subtypes (Fig. 4, E and F, and fig. S9B),
abundance of WT and FMRP-KO cells were Profiling the cellular constitutions of most notably the increased fractional (and over-
assessed in tumor tissue sections. Impairment FMRP-WT versus FMRP-KO tumors through all) abundances of effector CD8 T cell subtypes
in tumor growth or burden was evident and single-cell RNA sequencing and the reduction in CD4 regulatory T cells
scaled with the proportion of FMRP-KO can- Having established in functional genetic as- (Tregs) in FMRP-KO tumors. Motivated by these
cer cells in the inoculum (fig. S7G), and it was says that inflammation by CD8 and CD4 T cells differences in abundance and in subtype repre-
consistently correlated with increased CD8 of tumors lacking FMRP expression in cancer sentation, we sought to investigate cell-to-cell
T cell infiltration (fig. S7H). We focused on cells was central to the impaired tumor growth, signaling from WT and FMRP-KO cancer cells
the 5:5 mixture with intermediate growth we sought to investigate the cellular and mo- to lymphocytes that might explain these immu-
impairment, reasoning that this would best lecular mechanisms underlying the immuno- nological phenotypes.
reveal differential and potentially conflict- suppression evident in WT tumors and the
ing cellular phenotypes of the two popula- adaptive immunity that is triggered in FMRP- Charting cell-cell communications
tions of cancer cells. Tumors were collected KO tumors. As a first step, WT and FMRP-KO in tumors with and without FMRP
7 days after inoculation, and tissue sections PDAC tumors were subjected to single-cell RNA expression in cancer cells
were analyzed for green and red fluorescence sequencing (scRNA-seq), and clustering anal- Using CellChat, an algorithm for mapping cell-
to visualize the WT and FMRP-KO cancer cells ysis was performed to identify the various cell cell communications in scRNA-seq datasets
and immunostained with anti-CD8 to iden- populations in the TMEs. Using published cell (34), we assessed differentially increased ligand-
tify CD8 T cells. Tumors were highly infiltrated type–specific gene signatures (32), clusters with receptor pairs in tumors, wherein cancer cells
with CD8 T cells in close proximity to both red- similar expression patterns were annotated were the “sender cells” expressing genes en-
and green-labeled cancer cells, revealing that (fig. S8 and Materials and methods), which coding secreted ligands and the noncancer
both WT and FMRP-KO cancer cells were as- revealed marked differences in the cellular cells in the TME were the receptor-expressing
sociated with the recruited CD8 T cells (fig. S7I). constitution of WT versus FMRP-KO tumors “recipient cells.” Notably, despite the lower
The data indicate that the immunosuppres- (Fig. 4, A and B). The abundance of cancer abundance of cancer cells in FMRP-KO tumors
sion evident in pure WT tumors cannot be cells was reduced in FMRP-KO PDAC tumors (Fig. 4C), FMRP-KO cancer cells were implicated
fully conveyed to tumors containing a mixed (Fig. 4C), which, given the unimpaired ability by this analysis to have a higher cumulative
population of WT and FMRP-KO cancer cells, of FMRP-KO PDAC cancer cells to proliferate number of cell-cell communications with other
given the intensive infiltration of CD8 T cells in culture (fig. S4, C to E) and to grow as tu- cell types in the TME compared with WT cancer
and the similar abundances and intersper- mors in immunodeficient mice (Fig. 2, B and cells (fig. S9C). Given the aforementioned
sion of WT and KO cancer cells in such tumors, E), implicated the cytotoxicity of T lymphocytes. data that indicate profound differences in the

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 6 of 23

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

Fig. 4. The absence of up- A scRNA-seq profile for FMRP-WT tumors B scRNA-seq profile for FMRP-KO tumors
regulated FMRP in cancer
Proliferating CAFs Proliferating CAFs
cells reprograms the immune iCAFs
plasmacytoid
iCAFs
plasmacytoid
Neutrophils Neutrophils
TME, as revealed by scRNA- DCs DCs
Monocytes 2 Monocytes 2
seq. (A and B) tSNE plots from Monocytes 1 myCAFs myCAFs
cDC1 B cells cDC1 Monocytes 1 B cells
scRNA-seq analysis for WT Pericytes Pericytes
tumors (A) and FMRP-KO PDAC Proliferating M -2 M -2
Proliferating
tumors (B). The combination Myeloids Myeloids

of two biological replicates for

each population is shown; 18 cell
clusters were identified. One M -3 Endothelial M -3 Endothelial
M -1 cells M -1 cells
independent tumor was analyzed
Migratory DCs Migratory DCs
for each cell line (i.e., WT2, WT3, Lymphocytes Lymphocytes
KO2, KO8). CAF, cancer-associated

tSNE 2

tSNE 2
fibroblast; DC, dendritic cell.
(C) Comparative abundance of Cancer cells Cancer cells
tSNE 1 tSNE 1
the cancer cells in WT (red
column) versus FMRP-KO (blue
C Cancer cells Total abundance of Lymphocytes &
fractional representation of Lymphocyte subtypes G Upregulated communications
column) PDAC tumors, extracted
from the scRNA-seq data.
Abundance
D Total Lymphocytes E Lymphocyte subtypes
FMRP-WT tumors F Lymphocyte subtypes
FMRP-KO tumors
Cancer cells to Lymphocytes; FMRP-KO vs. WT tumors
Cxcr6
30
Ccr2
(D) Comparative abundance of

Fractional frequency (%)

30 Ccr2 Receiving cells
30 20 Cxcr6
(receptors)
Frequency (%)

the lymphocytes in WT (red)

Frequency (%)
20
Ccr2 Ccr2 DC8 Eff. Mem.
20
20 CD8 Tex
versus FMRP-KO (blue) PDAC 10
Ccr5 Tpex
Tfh
10
tumors from scRNA-seq data. 10 10 Cxcr6 Th1
Treg

(E and F) Fractional representation 0

0 0 0
Ccl6
Sender cell
of lymphocyte subtypes in WT Ccl27a

A e

8 x
K x
e
ve .
D e

Tr 1
eg

A e

K x
e
ve .

C 8 e
8 x

Tr 1
eg
lls

lls
T

(ligands)

ai m

ai m
Ccl8

D Te
N Tpe

D Te
N Tpe
Ef ly ik

C Lik

Ef ly ik

C Lik
D or iv

Tf
Th

D or iv

Tf
Th
W

ce
N Me

N Me
T

8 ar e L

8 ar e L
C ect ct

C ect ct
C 8
P-

W
P-

D
E v

E v
(E) (red) versus FMRP-KO (F) P- Cancer cells
R

P-

8 ai

8 ai
R

D N

D N
FM
FM

R
R

C 4

8
f

f
FM

Ccl7
FM

D
C

C
(blue) PDAC tumors, as revealed

D
C

C
Ccl2 Cxcl16

by ProjecTILs, an algorithm
leveraging reference datasets for H C 7 protein expression I C 7 protein expression J PDAC sc-injection in FVBN mice
in cultured cells’ supernatant FMRP/C 7 double KO FMRP-KO vs FMRP-KO-C 7-KO cancer cells
T cell subtypes (fig. S8 and 1500
Tumor volume Tumor weight CD8+ T cells
Materials and methods). Tfh, 50 2.5 0.25 15

T follicular helper cell; Th1,

Secreted CCL7 (pg/ml)

CD8+ T cell / Live cells

Normalized tumor voulme

2.0 0.20
Secreted CCL7 (pg/ml)

1000

Tumor weight [g]

T helper 1 cell. (G) CellChat analysis 10
30 1.5 0.15
of scRNA-seq data implicating
20 1.0 0.10
signaling circuits that are up- 500
ns 5

regulated in FMRP-KO tumors, in 10 0.5 0.05

particular involving cancer cells 0 0 0.0 0.00 0

KO
KO
KO
as ligand-expressing sender cells

2
T2

2
2
T2

KO
KO
KO
KO
KO

7-
7
W

7-
7-
W

C
C
and lymphocytes as receptor-
C
C
2-

2-
2-
2-

KO
KO
KO
KO
expressing receiving cells (see
table S1 for specifics). (H) ELISA K IL33 protein expression L IL33 protein expression M PDAC sc-injection in FVBN mice
analysis of the levels of CCL7 FMRP-WT vs. KO cultured cells FMRP-KO2 with I 33 overexpression FMRP-KO vs FMRP-KO-IL33-OE cancer cells

protein secreted from WT and Mock IL33-OE

Tumor volume Tumor weight Treg abundance
WT2 KO2 KO8 10 60
0.4
FMRP-KO PDAC cancer cells 1.0 0.31 0.22 1.0 3.38
Normalized tumor volume

%FOXP3+CD4+/CD4 +
in culture. In brief, one million cells IL33
IL33 8
0.3
Tumor weight [g]

40
were seeded in 10-mm culture 6
0.2
plates with complete medium. HSP90 4
HSP90 20
After 48 hours, culture medium 2
0.1

was replaced with 6 ml of 0 0.0 0

serum-free medium. Medium was
2

2
2

E
E
KO

KO
KO
-O

-O
-O
33

33
33

collected 48 hours later for

I
I
2-

2-
2-
KO

KO
KO

cytokine profiling and ELISA

analysis. N = 3 samples for each group. Data are representative of two independent experiments and are presented as means ± SEMs. (I) Absence of CCL7 protein expression
in FMRP-KO cancer cells engineered to carry a genetic deletion of Ccl7. In brief, 2 × 105 cells were seeded in 2 ml of complete medium in one well of a 6-well plate. After
48 hours, medium was collected for ELISA analysis of CCL7 protein level. N = 3 samples for each group. Data are representative of three independent experiments and are presented
as means ± SEMs. (J) Comparison of FMRP-KO versus Fmr1/CCL7-DKO s.c. tumors. The panels represent tumor volume (left), tumor weight (middle), and FACS analysis of CD8
T cells (right). Mice were euthanized on day 14 after cell implantation. Tumor volumes were normalized between days 7 and 14. CD8 T cells are shown as the proportion of live
cells. N = 7 mice for both groups; one independent experiment. Statistical analysis was performed with Mann-Whitney test. Data are presented as means ± SEMs. (K) Protein blot of
IL-33 expression in WT versus FMRP-KO PDAC cell lines. Data are representative of three independent experiments. (L) Protein blot demonstrating genetically rescued levels of
IL-33 expression in an FMRP-KO cell line. Data are representative of three independent experiments. (M) Functional effects of IL-33 overexpression in FMRP-KO tumors. The graphs show
means ± SEMs of tumor volume (left), tumor weight (middle), and FACS analysis of Tregs (right). Mice were euthanized on day 14 after s.c. injection. Tumor volumes were normalized
between days 7 and 14. Tregs were defined as FOXP3+CD4+ cells and shown as a percentage of CD4+ T cells. N = 5 for FMRP-KO2 group, N = 4 for KO2–IL-33–OE group, and N = 3
for WT2 group; one independent experiment. Statistical analysis by Mann-Whitney or one-way ANOVA tests. ns, not significant; *P < 0.05; **P < 0.01; ****P < 0.0001.

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 7 of 23

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

lymphocyte abundance and subtype constitution, for differentially expressed cytokines (table S2), alone, the WT and siRNA-control (WT) cancer
we sought to chart up-regulated communica- and interleukin 33 (Il-33), a chemokine that is cells were potently inhibitory for CD8 and
tions from FMRP-KO cancer cells to lympho- known to stimulate the production and activ- CD4 T cell proliferation (Fig. 5, B and C). By
cytes. Most prominently, a group of chemokines ity of Tregs (37), was identified as the most contrast, the FMRP-KO and FMRP-KD PDAC
and their receptors were identified that were prominent. IL-33 mRNA is down-regulated cells only modestly inhibited proliferation of
more frequently expressed in FMRP-KO can- in FMRP-KO cancer cells (fig. S9I), and IL-33 CD8 (Fig. 5B) and CD4 (Fig. 5C) T cells. Ad-
cer cells and lymphocytes in the scRNA-seq protein level is decreased both in KO and siRNA ditionally, we compared PDAC WT2 cells and
dataset, respectively (Fig. 4G and table S1). KD cancer cells (Fig. 4K and fig. S9J). Addi- KO2 cells with the double KO of FMRP and
Among the five up-regulated ligands in the tionally, CLIP-seq analysis indicated that Il-33 CCL7, which revealed that the FMRP-CCL7
scRNA-seq dataset (fig. S9D), only C-C motif mRNA binds FMRP (fig. S9K) and is therefore DKO cells were similarly inhibitory toward
chemokine ligand 7 (Ccl7) demonstrated sig- implicated as a direct target of FMRP regu- CD8 T cell proliferation as WT cancer cells
nificant up-regulation in cultured FMRP-KO lation. To follow up on these observations, an (Fig. 5D), which further established CCL7 as a
cancer cells as both an mRNA (fig. S9E) and FMRP-KO cancer cell line was engineered that key FMRP-regulated factor involved in counter-
a secreted protein (Fig. 4H); congruently, an overexpressed IL-33 protein (FMRP-KO–IL-33– acting the direct immunosuppressive effect of
siRNA KD of Fmr1 in WT cancer cells also OE) (Fig. 4L). After transplantation into synge- WT cancer cells on CD8 T cells.
elevated CCL7 secretion (fig. S9F). Moreover, neic immunocompetent mice, tumor growth To extend on these results, we also com-
Ccl7 scored significantly in a previously pub- was examined and characterized. Compared pared the effects on proliferation of activated
lished FMRP cross-linking immunoprecipita- with FMRP-KO tumors, the FMRP-KO–IL-33– antigen-nonspecific T cells with matched pairs
tion sequencing (CLIP-seq) analysis (35) (fig. OE tumors had increased volume and weight of WT versus FMRP-KO B16-OVA melanoma
S9G) that is indicative of direct binding of and increased abundance of Tregs, comparable cells and CT26 colon cancer cells. FMRP-KO
mRNAs to FMRP (albeit in neurons), consistent to WT tumors (Fig. 4M), revealing yet another B16-OVA cells similarly evidenced reduced
with the implication that FMRP represses Ccl7 FMRP-regulated factor involved in modulat- inhibitory activity for both CD8 and CD4 T cell
mRNA translation (and mRNA stability) in ing antitumor immunity. proliferation compared with WT B16-OVA
WT tumors—a repression that is released in Thus, we have identified and functionally cancer cells (Fig. 5, E and F). Concordantly,
cancer cells lacking FMRP expression. Nota- validated two direct targets of FMRP expres- both FMRP-KO and siRNA Fmr1 (FMRP-KD)
bly, CCL7 is a proinflammatory cytokine cap- sion in cancer cells that are involved through parental CT26 cells showed reduced inhibition
able of contributing to the recruitment of their differential expression in the distinctive compared with WT or siRNA Ctrl CT26 cells
lymphocytes and thereby triggering antitumor effects of WT versus FMRP-KO cancer cells on (Fig. 5, G and H).
immune responses (36). To assess the possible the immune TME. CCL7 is up-regulated in KO Next, we assessed antigen-specific killing by
role of up-regulated CCL7 in the immune phe- tumors, leading to the recruitment of CD8 T cells, CD8 T cells in a classical CTL assay, involving
notype of FMRP-KO cancer cells, we used whereas the concomitant reduction in IL-33 ex- OVA-specific OT1 CD8 T cells (isolated from
CRISPR-Cas9 technology to produce a double pression compared with that in WT tumors OT1 transgenic mice) cocultured with WT or
knockout (DKO) of the Fmr1 and Ccl7 genes in led to fewer Tregs, consistent with the marked FMRP-KO B16-OVA cancer cells, both express-
PDAC cancer cells such that CCL7 protein is differences in lymphocyte abundance and sub- ing similar levels of OVA antigen (Fig. 5I and
not expressed (Fig. 4I). Then, we audited these type representation that distinguishes WT and fig. S9L). A T cell titration assay was performed,
DKO cancer cells upon transplantation com- FMRP-KO tumors (Fig. 4, D to F, and fig. S9B). which revealed the sensitivity of both WT and
pared with FMRP-KO cells. In the absence of FMRP-KO cancer cells to T cell killing, con-
cancer cell–derived CCL7, tumor volume and Assessing interactions of T cells with WT sistent with the aforementioned mixing exper-
weight were increased compared with those of versus FMRP-KO cancer cells in coculture iment (Fig. 3, G, H, and I). The FMRP-KO cells
FMRP-KO tumors, and the abundance of CD8 Given the results on cancer cell–to-lymphocyte were more susceptible in this assay to T cell
T cells was reduced (Fig. 4J). We also observed cell-cell communication, we next sought to killing at all ratios (Fig. 5J). The differential
a modest increase in the abundance of Tregs further assess the implication that WT and sensitivity to T cell killing was confirmed in
but no change in the expression of markers that FMRP-KO cancer cells have direct and differ- a second FMRP-KO B16-OVA cell line as well
indicate the immunosuppressive versus proin- ential interactions with CD8 and CD4 T cells. as with matched pairs of siRNA control and
flammatory phenotypes of tumor-associated To this end, a coculture assay was established siRNA FMRP-KD transfected B16-OVA cancer
macrophages (TAMs) (fig. S9H), which suggests (Fig. 5A and Materials and methods) involv- cells, as shown comparatively at the 1:1 ratio
that CCL7 is modulating the inflammation by ing ex vivo activated and tumor antigen– (Fig. 5K).
T cells but not the programming of TAMs. As nonspecific splenic T cells, which were combined Thus, FMRP-KO (and KD) cancer cells are
such, the FMRP-KO phenotype is evidently with WT versus FMRP-KO cells, or parental can- (i) markedly less inhibitory toward T cell pro-
not singularly modulated by CCL7 because it is cer cells transfected with control siRNA versus liferation, in part resulting from expression of
not affecting TAMs. Therefore, additional fac- siRNA targeting Fmr1. Subsequently, T cell CCL7, and (ii) somewhat more susceptible to
tors expressed by KO cancer cells—including proliferation without the confounding com- CD8 T cell killing. The results clearly establish
the other four up-regulated chemokines (fig. plexity of T cell–mediated killing of the cancer a direct mechanistic link between the FMRP-
S9D)—may also be involved, warranting future cells was assessed through FACS analysis. No- deficient cancer cells and the adaptive immune
investigation. tably, because the anti-CD3 plus anti-CD28 response catalyzed in FMRP-KO tumors.
As a complement to the CellChat analysis of bead-activated T cells are already highly pro-
the scRNA-seq, bulk RNA-seq was performed liferative, this assay only reports upon poten- Distinctive phenotypes of macrophages
for cultured WT and FMRP-KO cancer cells. In tial inhibitory effects of cancer cells on T cell populating WT versus KO tumors
this analysis, we focused on the converse ques- proliferation. Two pairs of WT and FMRP-KO We next considered another possible parameter
tion: Namely, are well-recognized immuno- PDAC cancer cells were assessed, as were two that might distinguish the immunosuppressed
modulatory genes comparatively up-regulated pairs of parental PDAC cells treated with two WT versus the T cell–inflamed FMRP-KO tumors,
in the immunosuppressive FMRP-WT cells? independent control siRNAs or siRNAs to namely TAMs and myeloid cells, collectively
To address this question, the transcriptome Fmr1 (FMRP-KD), each in coculture with CD8 defined by expression of the cell surface marker
datasets for cultured cancer cells were scored or CD4 T cells. Compared with T cells cultured CD11b. These innate immune cells are well

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 8 of 23

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

A Co-culture assay B PDAC cancer cell co-culture with CD8 T-cells

C PDAC cancer cell co-culture with CD4 T-cells
Cancer cells and CD8/4 T-cells

FMRP KO or WT
Activated cancer cells 100 100 100 100
CD8/4 T-cells

(of labeled CD4 T cells)

(of labeled CD8 T cells)
(of labeled CD8 T cells)

(of labeled CD4 T cells)

80 80 80 80

% of CFSE low
% of CFSE low
% of CFSE low

% of CFSE low
60 60 60 60

40 40 40 40

20 20 20 20

0 0 0 0

si L1
L2

r2
4

T2
T3

2
8
e
8
e
T2
T3

2
8

trl

trl
D CD

KO
KO
on
Fm

Fm
D CD
on

KO
KO

Fm
TR

TR
D

W
W

C
W
W

al
al

si
si

si
C

ed
ed

4
al
8

al
si

at
at

nt
C
C

ul
ul

re
im
im

pa
st
st

un
un
Cancer cell co-culture B16-Ova cancer cells co-cultured with T cells CT26 cancer cells co-cultured with CD8 T-cells

D with CD8 T cells

E CD8 T-cells F CD4 T-cells G H
100
60 40 40 50
(of labeled CD8 T cells)

(of labeled CD8 T cells)

(of labeled CD4 T cells)
80

(of labeled CD8 T cells)

40
% of CFSE low

% of CFSE low
% of CFSE low
30 30

% of CFSE low

% of CFSE low
60 40
30
20 20
40 20
20
20 10 10 10

0 0 0 0 0

-1

-2
T2

72
D D8

2- O2

trl

trl
T2

r1
on

T1
7K

C
KO

KO
W

KO
W
C

Fm
W
al

si
cl
ed

si
at
C
ul

KO
im
st
un

I T-cell killing assay J T-cell killing titration assay K

Cancer cells with Cancer cells with antigen-specific CD8 T-cells B16-OVA cancer cell co-culture with CD8 T-cells
antigen-specific CD8 T-cells Two-way Anova p-value < 0.0001
1.5
1.25 FMRP-WT 1.5
Normalized number of

FMRP KO or WT

Normalized number of
FMRP-KO

Normalized number of
B16-OVA cells
live cancer cells

OT1 1.00

live cancer cells

live cancer cells
CD8 T cells 1.0
1.0
0.75

0.50
0.5 0.5

0.25

0 0.0 0.0
1

-1

-2

-1

2
5:

-
r1
trl
trl

r1
KO

KO
0.

C
C

Fm
si

si
E:T ratio

si
Fig. 5. FMRP-KO cells directly stimulate T cell proliferation and are more or parental CT26 cells transfected with siCtrl or siFmr1 RNAs (right). The graphs
sensitive to killing by antigen-specific CD8 T cells. (A) Schematic representa- represent the fraction of proliferating (CFSE-low) cells as a fraction of all labeled CD8
tion of the ex vivo T cell coculture assay. Splenic CD4 or CD8 T cells were or CD4 T cells. N = 3 samples for each group. Data are representative of two
prestained with carboxyfluorescein diacetate succinimidyl ester (CFSE), activated independent experiments and represent means ± SEMs. Statistics by one-way
with the CD3 or CD28 Dynabeads, and cocultured with FMRP-KO or WT PDAC cells. ANOVA with Tukey’s multiple comparison analysis. (I) Schematic representation of
After 72 hours of coculture, cells were harvested and processed for flow cytometry an in vitro T cell killing assay. B16-OVA WT and FMRP-KO cells were cocultured in
analysis. (B and C) FACS analysis of CD8 (B) and CD4 (C) T cell proliferation in different ratios with CD8+ T cells isolated from OVA-specific T cell receptor
cocultures with WT or FMRP-KO PDAC cancer cells (left), or WT2 cells transfected transgenic (OT1) mice that had been preactivated by OVA peptide and IL-2 plus IL-7
with siCtrl or siFmr1 RNAs (right). The graphs represent the fraction of proliferating treatment. The number of viable B16-OVA cancer cells at the end point of the assay
(CFSE-low) cells as a fraction of all labeled CD8 or CD4 T cells. N = 3 samples was determined and reported. (J) Titration assay involving coculture at different
for each group. Data are representative of three independent experiments and are effector-to-target ratios (E:T ratios) assessing sensitivity to cytotoxic T cell killing of
presented as means ± SEMs. Statistics by one-way ANOVA with Tukey’s multiple B16-OVA WT and FMRP-KO cells; after 48 hours, FACS analysis revealed the
comparison analysis. (D) FACS analysis of CD8 T cell proliferation in cocultures with numbers of live cancer cells, which were normalized to the WT cancer cell number
WT, FMRP-KO2 PDAC, and KO2-CCL7KO (DKO) PDAC cancer cells. The graph in the 0.5:1 group. N = 4 samples for each group. Statistics by two-way ANOVA
represents the fraction of proliferating (CFSE-low) cells as a fraction of all labeled test. Data are representative of two independent experiments and represent means ±
CD8 T cells. N = 3 samples for each group. Data are representative of three SEMs. (K) Further validation of the T cell killing assay using OT1 CD8 T cells and
independent experiments and are shown as means ± SEMs. Statistics by one-way B16-OVA cancer cells at a 1:1 ratio. (Left) Two independent B16-OVA FMRP KO clones
ANOVA with Tukey’s multiple comparison analysis. (E and F) FACS analysis of CD8 (KO51 and KO72) after 48 hours of coculture; FACS analysis revealed the numbers of
(E) and CD4 (F) T cell proliferation in cocultures with WT or FMRP-KO B16-OVA remaining cocultured cancer cells, which were normalized to the WT cancer cell
cancer cells. The graphs represent the fraction of proliferating (CFSE-low) cells over number. (Right) B16-OVA WT cells transfected with siCtrl or siFmr1 RNAs were
all labeled CD8 or CD4 T cells. N = 3 samples for each group. Data are means ± analyzed after 24 hours. FACS analysis revealed the numbers of cocultured cancer
SEMs and are representative of two independent experiments. Statistics by one-way cells, normalized to the siCtrl1 group. Statistical analysis by one-way ANOVA. N = 3
ANOVA with Tukey’s multiple comparison analysis. (G and H) FACS analysis of samples for each group. Data are representative of three independent experiments and
CD8 T cell proliferation in cocultures with WT or FMRP-KO CT26 cancer cells (left), represent means ± SEMs. ****P < 0.0001.

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 9 of 23

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

known to populate PDAC and other solid immune deserts that characterized FMRP-WT cilitates the generation of immunostimulatory
tumors, in particular TAMs displaying a so- tumors (Fig. 3, A and B). M1-like macrophages.
called M2-like immunosuppressive phenotype Given the evident M2-like TAM program- Upon curating the scRNA-seq data for other
(38, 39). To begin, we modified the T cell pro- ming activity of FMRP-WT cancer cells, we potentially immunosuppressive factors in WT
liferation assay (Fig. 5A), replacing the cancer sought to identify differentially expressed fac- tumors, Cd63 mRNA was identified as one of
cells with myeloid cells isolated by immuno- tors that might contribute to their induction. the most differentially expressed genes in WT
magnetic selection for CD11b-positive cells Returning to the RNA-seq comparison of WT versus FMRP-KO cancer cells (Fig. 6K and
from tumors established with the two matched and FMRP-KO cancer cells in culture (table table S2); notably, Cd63 encodes a protein that
pairs of PDAC WT and FMRP-KO cells. The S2), a mechanistically logical gene—Pros1—was is displayed on and involved in the secretion
CD11b+ myeloid cells isolated from both WT found to be preferentially expressed in WT of exosomal vesicles (EVs) (41). Investigation
tumors suppressed CD8 and CD4 T cell pro- cancer cells. Tumor-secreted protein S (PROS1), of exosome secretion from cultured WT versus
liferation, whereas those from FMRP KO tumors a ligand for the MER/TYRO3 receptor that is FMRP-KO cancer cells revealed that the WT
did not discernibly inhibit T cell proliferation, expressed on macrophages, has been shown cells secreted more exosomes in both the PDAC
similar to controls without added myeloid to repress expression of immunostimulatory and B16-OVA matched pairs (Fig. 6L and fig.
cells (Fig. 6, A and B). Additionally, we further genes in M1-like macrophages, and its gene- S10I), consistent with the higher levels of CD63
assessed the phenotype of tumoral macro- tic ablation in cancer cells enhances antitumor mRNA. Notably, a recent study revealed a di-
phages by FACS from one matched set of WT immune responses (40). Our investigation re- rect role for FMRP in the process of exocyto-
and FMRP-KO tumors (PDAC WT2 and KO2) vealed that PROS1 protein and mRNA are more sis during inflammation (42), which suggests
for expression of CD86, a marker of M1-like highly expressed in WT versus FMRP-KO can- that the regulation of exosome formation and
immunostimulatory macrophages, and argi- cer cells (Fig. 6H and fig. S10C), and CLIP-seq secretion might be another mechanism by
nase 1 (ARG1), a marker of M2-like immuno- analysis indicated that Pros1 mRNA can bind which FMRP mediates its immunosuppressive
suppressive macrophages. Their converse levels specifically to FMRP, implicating it as a trans- effects.
of expression suggested that WT tumors had lational (and transcriptional) regulatory target Therefore, we reasoned that the secreted
more abundant M2-like macrophages, whereas (fig. S10D). To assess its possible functional WT cancer cell–derived exosomes might be in-
FMRP-KO tumors favored M1-like TAMs (Fig. involvement, we first cocultured M1-polarized volved in macrophage programming, as has
6C). These results collectively motivated in- BMDMs in the transwell assay with WT ver- been suggested in previous studies (43, 44).
vestigation of the possibility that the WT versus FMRP-KO PDAC cancer cells in the pres- Accordingly, EVs were isolated from WT and
sus FMRP-KO cancer cells were differentially ence of recombinant mouse PROS1 protein. FMRP-KO PDAC cancer cells (fig. S10J) and
programming TAMs. FMRP-KO cells, which weakly induced immu- added in equal amounts to the cultured M1-
A transwell coculture assay was established nosuppressive ARG1 compared with WT can- polarized BMDMs. WT exosomes markedly
(Fig. 6D) and used to assess the effects of WT cer cells (Fig. 6E and fig. S10A), elevated ARG1 up-regulated mRNAs for four immunosup-
versus FMRP-KO cancer cells on bone marrow– expression in macrophages to a level compa- pressive genes characteristically expressed in
derived macrophages (BMDMs), which were rable to that of WT cancer cells in coculture M2-like TAMs (Fig. 6M), whereas FMRP-KO
programmed ex vivo to adopt an M1-like phe- with M1-polarized BMDMs when both were exosomes had little or no effect (Fig. 6M), which
notype (Materials and methods). WT cancer cells similarly treated with PROS1 protein (Fig. 6I). was also apparent at the protein level for ARG1
induced the expression of two immunosuppres- Concordantly, the addition of PROS1 to the (fig. S10K). To further assess the differential
sive proteins—ARG1 and IL-10—prototypically coculture medium with FMRP-KO cancer cells effects of WT versus FMRP-KO exosomes, we
expressed in M2-like macrophages, whereas served to inhibit expression of the aforemen- polarized BMDMs to an M2 phenotype and
the FMRP-KO cancer cells slightly induced tioned M1-like marker genes, Cd86 and Il-1a, cultured them in the presence of exosomes.
ARG1 and did not affect IL-10 (Fig. 6E). FACS to the same level as with WT cells (Fig. 6J). WT exosomes further increased the expres-
analysis of the M1-polarized BMDMs from the Next, we overexpressed PROS1 protein in sion of Chil3 and Arg1 mRNA in M2-polarized
transwell assay confirmed the more-pronounced FMRP-KO cancer cells (fig. S10E) and assessed BMDMs (fig. S10L) and slightly increased the
induction of ARG1 by WT versus FMRP-KO phenotypic effects in both the transwell assay already appreciable ARG1 protein levels (fig.
cancer cells and further revealed that WT can- with M1-polarized BMDMs and tumors. Forced S10M); by contrast, FMRP-KO exosomes had
cer cells down-regulated expression of the M1- expression of PROS1 protein in FMRP-KO can- no effect on Chil3 mRNA, modestly increased
like marker CD86, whereas the FMRP-KO cells cer cells resulted in marked up-regulation of Arg1 mRNA, and nevertheless reduced ARG1
had no effect (Fig. 6F), in agreement with the ARG1 protein (fig. S10F) and down-regulation protein levels (fig. S10, L and M). Therefore,
analysis of these two markers in bona fide of Cd86 and Il-1a mRNA (fig. S10G) in the the exosomes secreted by FMRP-KO cancer
TAMs (Fig. 6C). Additionally, mRNA levels of transwell assay, congruent with the results in- cells were both less abundant and functionally
four immunosuppressive genes expressed in volving the addition of PROS1 protein (Fig. 6, I distinct given that—even when applied in equal
M2-like TAMs were preferentially up-regulated and J). Moreover, when FMRP-KO (KO2) and amounts—the EVs from FMRP-KO cancer cells
in M1-polarized BMDMs by WT cancer cells PROS1-overexpressing FMRP KO (KO2ÐPros1- had markedly reduced capability to induce im-
in the transwell coculture assay (Fig. 6G). The OE) PDAC cancer cells were transplanted into munosuppressive gene expression in pheno-
results with ARG1 as a biomarker of the M2- syngeneic mice, up-regulation of PROS1 pro- typically polarized BMDMs.
like TAM state induced by WT cancer cells tein led to modestly increased tumor growth, Both PROS1 and exosomes produced by
were further corroborated in the analysis of a tumor weight, and ARG1 expression as well as FMRP-expressing cancer cells are directly reg-
second pair of WT versus FMRP-KO cancer reduced expression of CD86 in TAMs (fig. S10H). ulated by FMRP and implicated in the program-
cells (fig. S10A) and in a comparison of siRNA These results suggest another mechanistic com- ming of an immunosuppressive TME through
control versus siRNA FMRP-KD cells (fig. S10B). ponent of FMRP’s mode of action in immune the induction of M2-like TAMs. In the absence
Collectively, these bioassays reveal that FMRP- evasion, whereby FMRP directly regulates PROS1 of FMRP, the TAMs are evidently biased toward
expressing WT cancer cells are differentially active expression in WT cancer cells, thereby contrib- an M1-like immunostimulatory phenotype that is
in their capability to reprogram M1-polarized uting to an immunosuppressive TME through expected to contribute to the above-documented
macrophages into an M2-like immunosuppres- its induction of M2-like TAMs, such that its recruitment and activation of CD8 and CD4
sive phenotype, consistent with the distinctive down-regulation in FMRP-KO cancer cells fa- T cells.

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 10 of 23

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

PDAC tumoral CD11b cell co-culture with CD8 or CD4 T cells C FACS characterization of D Transwell co-culture: E Macrophage characterization
Tumor-associated Cancer cells and Macrophages FMRP-WT vs -KO cancer cells
A CD8 T-cells
B CD4 T-cells macrophages
150 150 80 M1 WT KO

(of labeled CD8 T cells)

cancer cells 1.0 11.6 6.2

(of labeled CD4 T cells)

M2-like marker genes

% of CFSE low

over macrophages
%C 86 or ARG1+
60 ARG1

% of CFSE low
100 100

40
50 50 IL10

20 M1-polarized BMDMs
H 90
0 0
0

Tu al C 1b + 2

b+ 2

b+ 3
8

al 1b + 2

b+ 3
8
on

or D11 O

T
KO

T
KO
W

W
K
Al

Al
Tu ral 1b +

b+
Tu ral 1b +

1
86
11

11
1
o D1

or D 1
o D1

1
m CD

A
D
m CD
C
Tu al C

C
Tu al C
al

a l
or

or
or
m

m
m

m
Tu

Tu
Tu
F Macrophage characterization (FACS analysis) G Transwell co-culture of M1-polarized BMDMs with
FMRP-WT vs FMRP-KO cancer cells FMRP-WT vs. FMRP-KO cancer cells

A 1 CD86 Arg1 Il10 Adm Cd163

50 5 30 4 25
300000

mRNA Expression level

(normalized to 18S)
40 4
3
%CD86+/CD45 +
MFI ARG1

200000 20
30 3 15
2
20 2 10
100000 10
1
10 1 5

0 0 0 0 0 0

T
1

KO
T

T
1

M
W

W
M

M
W

+
+

1
1

M
M

H P 1 protein expression
I M1 Macrophage reprogramming J M1 Macrophage reprogramming (M1-like marker genes)
FMRP-WT vs -KO cancer cells FMRP-WT vs FMRP-KO cancer cells
FMRP-WT vs. KO cultured cells
M1 cells + Cancer cells +
Cd86 Il1
WT2 KO2 KO8 +P 1 6
2.0
1.0 0.10 0.09

mRNA Expression level

FMRP WT FMRP KO

mRNA Expression level

M0 M1

(normalized to 18S)
(normalized to 18S)
1.0 12.3 13.2 14.6 12.2 12.2 17.7 1.5
P 1 4

A 1 1.0

2
HSP90 0.5
H 90

0.0 0

s1
T

os
W
os

os
W

o
Pr

Pr
Pr

+
+

1
1

+
+

M
M

KO
T

W
W

+
+
+
+

1
1
1
1

M
M
M
M

K Cd63 expression L Abundance of EVs M Co-culture of M1-polarized BMDMs with cancer cell-derived exosomes
scRNA-seq analysis WT vs. FMRP-KO cancer cells FMRP-WT vs. FMRP-KO cancer cells
4 2.0
Chil3 Arg1 Il10 Il13
Protein concentration [mg/ml]

3 2.5 2.0 2.5

3 1.5
mRNA Expression level
Expression Level

(normalized to 18S)

2.0 2.0
1.5
2
2 1.0 1.5 1.5
1.0
1.0 1.0
1 1
0.5
0.5
0.5 0.5

0
0.0 0 0.0 0.0 0.0
T

1
s

s
O
T

M
EV

EV
-W

-K
-W

-K

RP
RP

KO
W

W
FM
FM

FM
FM

+
+

+
1

1
1

1
M

M
M

Fig. 6. Immunosuppressive M2-like macrophages are abundant in the TME expression in macrophages from WT and FMRP-KO PDAC tumors. N = 4 for WT
of WT but not FMRP-KO tumors, programmed by PROS1 and cancer cellÐ tumors and N = 3 for KO tumors. Data are presented as means ± SEMs.
derived exosomes that are both more highly expressed by WT cancer Statistical analysis by Mann-Whitney test; one independent experiment.
cells. (A and B) FACS analysis of immunosuppressive effects of tumoral CD11b (D) Schematic representation of a transwell coculture assay involving ex vivo
cells—isolated by FACS from WT (WT2 and WT3) or FMRP-KO (KO2 and KO8) programmed M1-polarized BMDMs combined with WT or FMRP-KO cancer cells.
PDAC tumors—on CD8 (A) or CD4 (B) T cell proliferation after 72 hours of (E) Protein blot analysis of M2 TAM–like marker (ARG1 and IL-10) expression in
coculture. Proliferation is shown as the percentage of CFSE-low cells compared M1-polarized BMDMs cocultured with WT or FMRP-KO cancer cells. Data are
with the total number of CD8 or CD4 T cells. N = 2 to 3 samples for each group. representative of three independent experiments. (F) FACS analysis for ARG1
Statistical analysis by one-way ANOVA. Data are representative of two and CD86 (M1 TAM–like marker) protein expression for M1-polarized BMDMs
independent experiments and are presented as means ± SEMs. (C) FACS cocultured with WT or FMRP-KO cancer cells. N = 3 samples for each group.
analysis of CD86 (M1 TAM–like marker) or ARG1 (M2 TAM–like marker) Data are representative of two independent experiments and are shown as

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 11 of 23

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

means ± SEMs. (G) RT-PCR quantification of M2 TAM–like marker gene (Chil3, the media. N = 3 samples for each group. Data are representative of two
Arg1, Il-10, Il-13) mRNA expression in M1-polarized BMDMs cocultured with independent experiments and are presented as means ± SEMs. (K) Violin plot
cancer cell–derived exosomes from WT or FMRP-KO cancer cells. N = 3 samples showing the distribution of mRNA expression for the prototypical extracellular
for each group. Data are representative of two independent experiments. vesicle marker, Cd63, in WT versus FMRP-KO tumors. See table S2 for the fold
Error bars represent means ± SEMs. (H) Protein blotting analysis of PROS1 change and level of significance. (L) Comparative abundance of isolated
expression in WT versus FMRP-KO PDAC cancer cells. Data are representative of extracellular vesicles from WT and FMRP-KO PDAC cancer cells in culture. Pooled
three independent experiments. (I) Western blotting analysis of ARG1 protein data of six independent isolation experiments with six paired WT and FMRP KO
expression in M1-like macrophages cocultured with WT or FMRP-KO cancer cells, samples. Statistical analysis by paired t test. (M) RT-PCR quantification of M2 TAM–
with or without addition of recombinant PROS1 protein to the culture media. like marker gene (Chil3, Arg1, Il-10, Il-13) mRNA expression in M1-polarized BMDMs
N = 3 samples for each group. Data are representative of two independent cocultured with cancer cell–derived exosomes from WT or FMRP-KO PDAC
experiments and are shown as means ± SEMs. (J) RT-PCR quantification of M1 cancer cells. Statistical analysis by one-way ANOVA. N = 3 samples for each
TAM–like marker gene (Cd86, Il-1a) mRNA expression in M1-polarized BMDMs group. Data are representative of two independent experiments and are shown
cocultured with WT or FMRP-KO cancer cells, with or without PROS1 protein in as means ± SEMs. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

Contributions of macrophages to tumor increased the number of TAMs expressing CCL5 and L), which suggests that their expression is
immunity in FMRP-KO tumors and induced CXCL9 expression in TAMs, where- insufficient to elicit T cell recruitment in the
Given these implications, we next sought to as CXCL10 was unchanged compared with the highly immunosuppressive FMRP-WT TME,
assess indirectly regulated components of the TAMs from tumors in the immunodeficient consistent with the low levels of infiltrating
FMRP-KO immunostimulatory phenotype, in mice (Fig. 7F). The expression of these three T cells (Fig. 3, A and B and Fig. 4D) and the
particular the putative contributions of M1- proinflammatory chemokines was also docu- absence of antitumoral T cell immunity (Fig.
like immunostimulatory TAMs. First, we per- mented in CT26 tumors (fig. S11D). 3, I and J, and fig. S6G). As such, these three
formed a CellChat analysis of the scRNA-seq Next, aiming to further substantiate the in- chemokines, up-regulated in the absence of
datasets to identify potential cell-to-cell commu- ferred mechanism of T cell–dependent amplifi- FMRP, are established as indirectly regulated
nications preferentially up-regulated in FMRP- cation of TAMs expressing these proinflammatory secondary components of the FMRP-KO im-
KO tumors in the form of genes for secreted chemokines, PDAC KO2 tumors were treated munostimulatory phenotype, operating as
ligands that were expressed in myeloid cells with anti–IFN-g antibody, which, when com- downstream effectors of the recruitment and
and macrophages (as senders) versus cognate pared with sham-treated PDAC-KO tumors, re- antitumoral efficacy of T cells; notably, the
receptors that were expressed in lymphocytes sulted in enhanced tumor weight and reduced chemokine signaling through the CCR5 and
(as receivers) (table S1). Among the compar- CD8 T cell abundance (Fig. 7G) as well as lower CXCR3 receptors is functionally independent
atively elevated signaling interactions in FMRP- CTL activity reflected in a FACS analysis that and complimentary, such that their combined
KO tumors, most belonged to the CCL and assessed granzyme B (GZMB) and tumor necrosis activity is required for robust CD8 T cell inflam-
CXCL families of cytokines (Fig. 7A), which was factor–a (TNF-a) expression (fig. S11E). Consist- mation and consequent antitumor immunity.
confirmed by differential expression analysis ent with the expression pattern in immuno- Collectively, the results presented above elu-
for myeloid cells and macrophages comparing deficient mice (Fig. 7F), the neutralizing antibody cidate the distinctive immune phenotypes con-
FMRP-KO with WT tumors (table S3). Of these, to IFN-g reduced CCL5 and CXCL9 expression sequent to FMRP expression versus its absence
three proinflammatory chemokines known to in tumor-infiltrating macrophages, whereas in tumors, as summarized in Fig. 7I. FMRP-
recruit CD8 T cells (45–50) stood out: Ccl5, Cxcl9, CXCL10 expression was evidently independent expressing cancer cells induce immunosup-
and Cxcl10, which were broadly up-regulated in of IFN-g–expressing CD8 T cells (fig. S11F). pressive Tregs and M2-like TAMs through the
myeloid cells and macrophages from FMRP-KO Having established the direct and indirect expression of IL-33 and of PROS1 plus exo-
tumors (Fig. 7, B to D, and table S3). The prin- up-regulation of three proinflammatory chemo- somes, respectively, producing a TME essen-
cipal sources of these three cytokines in the kines in TAMs populating FMRP-KO tumors, we tially devoid of T cells. The down-regulation of
TME were monocytes and macrophages, with assessed their functional importance for the these three immunosuppressive regulatory
the exception of Ccl5, which is also highly ex- T cell–dependent tumor immunity by pharma- factors in tumors with genetic KOs or KDs
pressed in lymphocytes (fig. S11, A to C). Given cologically inhibiting their receptors. Thus, mice of FMRP in cancer cells, in concert with up-
that these chemokines are known to be up- bearing FMRP-KO2 tumors were treated with a regulation of immunostimulatory CCL7, leads
regulated in myeloid cells by infiltrating CD8 small molecule inhibitor—maraviroc (51)—of to the recruitment and activation of T cells
T cell–induced interferon-g (IFN-g) secretion, the receptor for CCL5 (fig. S11G) or with a block- that are responsible for the impaired tumor
we first considered whether expression of these ing antibody for CXCR3 (52), the receptor for growth and improved survival of mice bearing
genes in TAMs and myeloid cells in FMRP-KO CXCL9 and CXCL10 (fig. S11H). The two in- tumors lacking this regulator of translation
tumors was regulated directly by FMRP-KO hibitors, most notably in combination (Fig. 7H), and transcription.
cancer cells or indirectly by IFN-g–secreting reduced the recruitment of activated CD8 T cells
T cells. To this end, WT and FMRP-KO cancer and rescued the otherwise impaired growth Associating FMRPÕs cancer network with
cells were inoculated into immunodeficient rate of the FMRP-KO tumors compared with prognosis of human cancers
mice, which enabled the evaluation of chemo- vehicle or immunoglobulin G (IgG)–treated con- Having uncovered and functionally charac-
kine expression in the absence of an adaptive trols (Fig. 7H and fig. S11, I and J). The three terized the unprecedented activity of FMRP
immune system. Analysis of FACS-purified TAMs chemokines CCL5, CXCL9, and CXCL10 were through functional manipulations in mice,
revealed that CCL5 and CXCL10 were expressed also modestly expressed in TAMs from WT tu- we sought to assess the translational applica-
in FMRP-KO tumors in the absence of T cells mors detected in the FACS analysis (Fig. 7F and bility, to human cancers, of the realization that
(Fig. 7E), consistent with direct, inductive com- fig. S11D), although mRNA levels were very low FMRP can limit the recruitment of CD8 T cells
munication from macrophages to lymphocytes. (Fig. 7B). However, treatment of mice bearing into tumors and thereby facilitate tumor growth.
The same analysis was then performed in im- WT tumors with the inhibitors of the CCL5, To begin, we assessed the predictive value of
munocompetent mice, which revealed that CXCL9, and CXCL10 receptors had no effect FMR1 mRNA expression and failed to associ-
the presence of an adaptive immune system on either tumor growth or survival (fig. S11, K ate comparatively elevated levels with overall

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 12 of 23

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

A Upregulated communications B Ccl5 mRNA expression C Cxcl9 mRNA expression

D Cxcl10 mRNA expression
Macrophages/Monocytes to Lymphocytes 6
5
FMRP-KO vs. WT tumors
4
4
Tpex Tfh
CD8 Tex

Expression Level
Th1 4 3
CD8 Naive 3
CD8
Eff. Mem. Treg 2 2
CD8 2
EarlyAct.
1 1
IL1 Sender cell
CXCL
(ligands)
CCL 0 0 0
CCL
Prolif. Myeloids
CXCL

s
1

2
1

3
IL1 Mono. 1

id
o.

o.
oi

lo
M

M
Mono. 2

on
l

l
CXCL

ye
M 1

M
CCL

M
CCL M 2

if.

if.
ol

ol
IL1 CXCL M 3

Pr
CXCL CXCL CCL
FMRP-WT FMRP-KO

E PDAC sc-injection in SCID mice F PDAC sc-injection in FVBN mice

C 5 C 9 C 10
C 5 C 9 C 10
40 40 ns 100 40 40 100

%CXCL10 / Macrophages
%CXCL10 / Macrophages

%CXCL9 / Macrophages
%CXCL9 / Macrophages
%CCL5 / Macrophages

%CCL5 / Macrophages
80 80
30 30 30 30

60 60
20 20 20 20
40 40
10 10 10 10
20 20

0 0 0 0 0 0
T2

2
T2

2
KO

KO
KO
KO

KO
W

W
W

W
W
G PDAC sc-injection in FVBN mice + anti-IFN
H Dual inhibition of CXCR3 and CCR5
Tumor weight CD8+ / Live Tumor burden Tumor volume CD8+ T cells
800 40 8 1000 8

Normalized tumor volume

Tumor weight [mg]

800

Tumor weight mg

%CD3+CD8+/LD
600 30 6 6
%CD8+/Live

600
400 20 4 4
400
200 10 2 2
200

0 0 0 0 0
e

c
T

KO
T

c
i ro

iro
cl

cl
N
FN

iro
W
W

cl
hi

hi
IF

hi
av

av
I

ve
ti-
ti-

ar
an
an

+
m

m
G

G
+
+

+
Ig

Ig
KO
KO

3
R

R
XC

XC
C

C
ti-

ti-

ti-
an

an
I
High FMRP cancer network activity Low FMRP cancer network activity

Il33 Tregs Immunosuppressive Pro-inflammatory

factors - Arg1 et al. chemokines
Ccl7
Cancer Cell Cancer Cell Ccl5
hyper-expressing with low/absent Cxcl9
FMRP FMRP Cxcl10
Pros1 downregulation
+ M2-inducing of Pros1 & M2-
exosomes inducing
exosomes
Immunosuppressive Excluded Immunostimulatory Recruited
macrophages and suppressed macrophages and activated
CD8 T cells CD8 T cells

Fig. 7. Proinflammatory chemokines directly and indirectly up-regulated in Ly6C−/Ly6G−—within WT versus FMRP-KO tumors derived from inoculation of
macrophages of FMRP-KO tumors orchestrate inflammation by T cells that PDAC cancer cells growing as tumors in SCID mice (E) or FVBN mice (F).
mediate tumor immunity. (A) Inferred intercellular communication networks Statistical significance was assessed with the Mann-Whitney test. (E) N = 5 mice
in FMRP-KO tumors from myeloid cells as senders of ligands to lymphocytes as for all arms. (F) For CCL5, N = 8 for the WT cohort and N = 9 mice for the
receivers expressing cognate receptors in FMRP-KO tumors, revealing up- KO2 cohort; for CCL9, N = 6 for WT and 7 for KO cohorts; and for CXCL10, N = 6
regulation of CCL and CXCL pathways in FMRP-KO myeloid cells (see also table for WT and N = 10 for KO cohorts; one independent experiment. Data are
S1 for specifics). (B to D) Violin plot showing the distribution of mRNA represented as means ± SEMs. (G) Quantification of tumor burden (left) and
expression for differentially up-regulated ligands in myeloid cells comparing CD8 T cell infiltration (right) in PDAC tumors at day 14 postinoculation (s.c.) of
FMRP-KO versus WT tumors: Ccl5 (B), Cxcl9 (C), and Cxcl10 (D). (E and F) FACS WT and FMRP-KO cancer cells, as well as FMRP-KO cancer cells treated with
analysis of CCL5 (left), CXCL9 (middle), and CXCL10 (right) ligands expression an IFN-g neutralizing antibody. One-way ANOVA was used for multiple group
in the overarching macrophage population—identified as being CD45+/CD11b+/ comparisons. N = 5 mice for each group; one independent experiment. Data are

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 13 of 23

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

presented as means ± SEMs. (H) Quantification of tumor burden (left), tumor inflammation by CD8 and CD4 T lymphocytes (left). By contrast, cancer cells
volume (middle), and FACS analysis of CD8 T cell infiltration (right) of FMRP-KO lacking FMRP recruit CD8 and CD4 T lymphocytes and promote their
tumors in mice treated with the combination the CCR5 inhibitor maraviroc proliferation through up-regulation in cancer cells of CCL7 and down-regulation
and anti-CXCR3 antibody. P values were assessed with the Mann-Whitney test. of IL-33, along with down-regulation of PROS1 and exosomes, whose absence
N = 12 mice for each group; one independent experiment. Data are presented as leads to indirect activation of proinflammatory cytokines (CCL5, CXCL9,
means ± SEMs. (I) Schematic representation of FMRP-mediated immuno- and CXCL10) in reprogrammed immunostimulatory macrophages; these profound
suppressive mechanisms in tumors. Cancer cells hyperexpressing FMRP modulate differences in the TME collectively trigger antitumor immunity, leading to impaired
Tregs and immunosuppressive macrophages cells through the up-regulation of tumor growth and improved survival. *P < 0.05; **P < 0.01; ***P < 0.001;
IL-33 and PROS1 and exosomes, respectively, which prevents appreciable ****P < 0.0001.

survival (OS), with disease-free survival (DFS), or on samples collected after 48 hours. Upon over- versus low FMRP signature scores (Fig. 8, E
with reduced CD8 T cell infiltration (estimated laying the FMRP CC subsignature onto the to G). Again, no meaningful associations with
by a CD8 T cell signature; see Materials and siRNA FMRP-KD transcriptome profile, we ob- tumor aggressiveness (estimated by tumor stage;
methods) in a pan-cancer analysis of 31 human served a significant down-regulation of the fig. S12, M to O) were observed. Congruently,
tumor datasets from TCGA (fig. S12, A to C). FMRP CC signature score in FMRP-KD cancer similar inverse associations with CD8 infiltration
This negative result is potentially consistent cells (fig. S12G), consistent with its veracity as a scores were observed with the FMRP CC sub-
with the knowledge that FMRP is not only a reporter for FMRP’s regulatory network, which signature in these three tumor types (fig. S12, P to
regulator of translation and transcription but we therefore applied in the analyses below. R), substantiating the implication that FMRP
is also itself posttranscriptionally regulated in In notable contrast to the nonassociation of transcriptional network activity is modulating
neurons, such that the mRNA levels, as well as FMR1 mRNA itself, the FMRP cancer signa- the immune response in human tumors.
protein levels, do not necessarily correlate linear- ture revealed a statistically significant associ- Furthermore, the FMRP cancer signature
ly with its downstream translational network ation with OS and DFS across the same cohort showed a significant correlation with progression-
activity (53, 54). of 31 cancer datasets, such that patients with free survival (PFS) in colorectal cancer (Fig. 8H,
Given that the level of FMR1 mRNA was not a higher FMRP cancer signature score have left panel), wherein lymphocytic reaction is
prognostically informative, we developed an worse OS and DFS (Fig. 8A and fig. S12H). More- an important prognostic factor (55). Notably,
overarching gene signature that is indicative over, a high FMRP cancer signature score this association exists for microsatellite stable
of FMRP’s regulation of a downstream tran- demonstrated a statistically significant anti- (MSS) but not for microsatellite instable (MSI)
scriptional network in cancer cells (henceforth correlation with a CD8 T cell signature that is colorectal patients (Fig. 8H, middle and right
called the FMRP cancer signature), which re- diagnostic of CTL abundance in human tumors panels). Additionally, the FMRP CC subsigna-
flects genes in the network that are directly (Fig. 8B). Notably, the FMRP cancer signature ture showed a similarly significant association
and indirectly controlled by FMRP at the score exhibits similar distributions across four with PFS in colorectal cancer, again only for
transcriptional level. The FMRP cancer signa- clinical tumor stages reflective of malignant MSS tumors (fig. S13A). To assess the effect
ture was generated by combining the differ- progression of different cancer types (fig. S12I), of tumor grade on the observed association,
entially expressed genes in two subsignatures, which suggests that the observed associations the samples were separated into lower-grade
derived by comparing FMRP-WT versus FMRP- with prognosis are not driven by FMRP’s reported (stages I to III) and high-grade (stage IV) tu-
KO PDAC cell lines in culture [FMRP cancer cell roles in stimulating invasion and metastasis. mors. A similar trend was observed for the as-
(CC) subsignature] and by comparing FMRP- To substantiate these results, we further in- sociation of the FMRP cancer signature with
WT versus FMRP-KO tumors (FMRP tumor vestigated the correlation between patient sur- PFS in lower-grade MSS (but not MSI) tumor
subsignature) (fig. S12D, table S4, and Mate- vival and the FMRP CC subsignature, which samples (fig. S13B). The high-grade colorectal
rials and methods). Functional enrichment only involves differentially expressed genes in cancers represented in the available dataset
analysis linked the FMRP cancer signature bulk RNA-seq analysis of the WT versus KO comprise only MSS tumor samples, for which
with multiple pathways (fig. S12E and table PDAC cell lines in culture and therefore eval- the FMRP activity signature again showed a
S4) relating to inflammation and immune uates FMRP’s downstream transcriptional net- significant correlation with PFS (fig. S13C). No-
responses as well as epithelial-mesenchymal work without the indirect effects of the TME tably, MSI patients with high mutational burden
transition (EMT) and other tumor pheno- (fig. S12D and table S4). Consistently, the FMRP often demonstrate better response to immuno-
types potentially reflective of FMRP’s previ- CC subsignature also revealed a significant cor- therapy and tend to have better prognosis (56).
ously reported capability to stimulate invasion relation with patient OS and DFS across the The less-immunogenic MSS tumors, however,
and metastasis (15–17). Notably, the correla- 31 cancer types (Fig. 8C and fig. S12J) as well as can evade the immune response and metasta-
tion of differentially expressed genes with the a significant anticorrelation with CD8 T cell size in >80% of cases with advanced colorectal
apoptosis pathway is only evident in the FMRP infiltration in tumor samples (Fig. 8D), which cancer (57). The association of the FMRP signa-
tumor subsignature, derived from scRNA-seq was again independent of tumor stage (fig. ture scores with the prognosis of MSS patients
data of WT versus FMRP-KO tumors, con- S12, K and L). These results further suggest suggests that high FMRP transcriptional net-
sistent with ongoing apoptosis induced by the that FMRP transcriptional network activity work activity may contribute to the poor re-
observed CTL inflammation and killing, and per se is limiting the recruitment of CD8 T cells sponses in this and other tumor types largely
not by intrinsic differences in the WT versus into human tumors and influencing cancer pa- refractory to immune checkpoint inhibition
FMRP-KO cancer cells. To assess the specificity tients’ prognosis. (see below).
of the FMRP CC subsignature and to ensure The association of the FMRP cancer signa- Breast cancers have been histologically char-
that it is directly linked to FMRP expression ture with reduced OS and lower CD8 T cell in- acterized as immune deserts, largely devoid of
(and not, for example, to adaptations occurring filtration is further illustrated for three human inflammation by CD8 T cells (58). Despite this
during outgrowth of the single-cell Fmr1-KO tumor types, namely endometrial carcinoma, overall poor immunogenicity, human breast
clones), Fmr1 was knocked down in a PDAC melanoma, and head and neck squamous cell tumor samples with a low FMRP signature score
cancer cell line, using siRNA transfection (fig. carcinoma, which have particularly provocative are characterized by a significantly higher CD8
S12F), and RNA-seq analysis was performed discriminations between tumors with high T cell infiltration score in a TCGA cohort of

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 14 of 23

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

Association of FMRP cancer signature Association of FMRP cancer cell sub-signature

across 31 human cancers across 31 human cancers

A B C D
Disease free survival FMRP signature vs. T-cell infiltration Disease free survival FMRP signature vs. T-cell infiltration
1.00 Low FMRP score (lower 25%) Cor.: - 0.34 1.00 Low FMRP score (lower 25%) Cor.: - 0.31
Medium FMRP score P-value: 1.1 e-95 Medium FMRP score P-value: 4.5 e-20
0.6 0.6
High FMRP score (upper 25%) High FMRP score (upper 25%)

0.75 0.75

Proportional Survival

CD8+ T−cell infiltration

CD8+ T−cell infiltration
0.4 0.4

0.50 0.50

0.2 0.2
0.25 0.25

COX model p-value; COX model p-value;

Medium vs. low: 0.23 0.0 Medium vs. low: 0.02
0.00 High vs. low: 0.0007 0.0
0.00 High vs. low: 1.5 e-4
0 100 200 −0.2 −0.1 0.0 0.1 0.2 0.3 0 100 200 0.0 0.2 0.4
Months FMRP cancer signature score Months FMRP cancer cell sub-signature score

E Uterine Corpus Endometrial Carcinoma

F Melanoma G Head and Neck Squamous Cell Carcinoma
Overall survival T-cell infiltration Overall survival T-cell infiltration Overall survival T-cell infiltration
0.6
1.00 0.16 p-val: 5.5 e-4 1.00 p-val: 3.5 e-17 1.00 p-val: 1.24 e-8

CD8+ T−cell infilteration

CD8+ T−cell infiltration
Proportional Survival

CD8+ T−cell infiltration

Proportional Survival

Proportional Survival
0.75 0.75 0.75
0.12 0.4 0.2

0.50 0.50 0.50

0.08
0.2 0.1
0.25 0.25 0.25
0.04
0.00 Log-rank p-val: 2.14 e-6 0.00 Log-rank p-val: 0.001 0.00 Log-rank p-val: 0.02
0.0

re
0 50 100 150 200 0 100 200 300 0 50 100 150 200
re

% atu

% atu
% atu

% atu

% atu
Months Months Months

25 gn

25 gn
25 gn

25 gn

)
)

er si

er si
er si

er si

w P

pp P
w P

pp P

w P

pp P

(lo MR

(u R
(lo MR

(u MR

(lo MR

(u R

FM
FM

F
F

h
w

ig
ig

Lo
Lo

H
H

H
H Colon adenocarcinoma
FMRP cancer signature score associations (Progression-free survival) I Breast Cancer
FMRP cancer signature vs. Immunophenotype
Immunophenotype TCR diversity
i All samples ii MSS samples iii MSI samples i TCGA cohort ii GSE177043
iii GSE177043
1.00 1.00 1.00 0.25 p-val: 2.0 e-6 p-val: 0.003
0.4 0.4

CD8+ T−cell infiltration

Proportional Survival

FMRP signature score

FMRP signature score
0.75 0.75 0.75 0.20
0.3 0.3

0.15
0.50 0.50 0.50 0.2 0.2

0.10 0.1 0.1

0.25 0.25 0.25
0.05 0.0 0.0
0.00 Log-rank p-val: 0.005 0.00 Log-rank p-val: 0.001 0.00 Log-rank p-val: 0.65 p-val: 0.02
0 50 100 150 0 50 100 150 0 20 40 60

) y

) y
re

% sit

% sit
% atu
% atu

Months Months Months

50 ver

50 ver
fla
cl
25 gn
25 gn

er di

er di
Ex

In
)
)

er si
er si

w R

pp R
pp P
w P

(lo TC

(u TC
Low FMRP signature (lower 25%) High FMRP signature (upper 25%)
(u MR
(lo MR

h
F
F

ig
Lo
h
w

H
ig
Lo

J K
FMRP cancer cell sub-signature in responses to anti-PD1/PDL1 therapy FMRP cancer cell sub-signature in chemotherapy responses
i Melanoma - anti-PD1 ii Lung cancer - anti-PD1/PDL1 iii Urothelial Cancer - anti-PD-L1 i Breast Cancer - Taxane ii Lung Cancer - chemotherapy
Overall survival Progression-free Survival Overall Survival Disease-free survival Progression-free survival
1.00 + 1.00 1.00 + 1.0 1.00
Log-rank p-val: 0.01 ++

+
Proportional Survival

Proportional Survival
Proportional Survival

+
Proportional Survival

Proportional Survival
+
+ +++ +
+

0.75 0.75 0.75 +

0.8 0.75
+ +

+
+

0.50
+

0.50 0.50 ++ 0.6 0.50

++
+
++ +++
+
+++ ++ +++
++
++ ++++++++
+
+ ++ + +++++ + ++++ ++ ++
+
+ + + ++
+

0.25 0.25 0.25

++
+
++ 0.25
+ ++ ++ + +++
+++++++ + +++++++ 0.4
+

0.00 Log-rank p-val: 0.11 0.00 0.00 Log-rank p-val: 0.003 COX model p-val: 0.005 0.00 COX model p-val: 0.02
0 250 500 750 1000 0 200 400 600 0 5 10 15 20 25 0 2 4 6 0 1000 2000 3000
Days Days Months Years Days
Low FMRP signature (lower 50%) Low FMRP signature (lower 25%)
High FMRP signature (upper 50%) High FMRP signature (upper 25%)

Fig. 8. FMRP transcriptional network signatures are negatively associated in the TCGA human pan-cancer dataset, as in (A). (D) Anticorrelation of the
with both CD8 T cell infiltration and more favorable prognosis in multiple FMRP CC subsignature score with a CD8 T cell infiltration signature, as in (B).
human tumors. (A) Anticorrelation of the FMRP cancer signature score with Only up-regulated genes in FMRP activity CC subsignature were used in deriving
DFS in the TCGA human pan-cancer dataset. Low FMRP activity, quantile 1, blue the signature score in this analysis. (E to G) Anticorrelation of the FMRP cancer
curve; medium FMRP activity, quantiles 2 and 3, light blue curve; high FMRP signature score with PFS (left) and CD8 T cell infiltration signature (right;
activity, quantile 4, red curve. The COX model was used, considering the tumor boxplots showing median, lower, and upper quantiles) in endometrial carcinoma
type as a covariate to estimate the significance of correlation. (B) Anticorrelation (E), melanoma (F), and head and neck squamous cell carcinoma (G). Log-rank
of the FMRP cancer signature score with a CD8 T cell infiltration signature, test was used for survival analyses, and Wilcoxon two-tailed test was used
estimated by the xCell package, in human pan-cancers. Linear regression model for the CD8 T cell association analyses. (H) Anticorrelation of the FMRP cancer
with tumor type as a covariate was used to estimate the significance of signature score with PFS for human colorectal cancer. Significant correlation
correlation. (C) Anticorrelation of the FMRP CC subsignature score with DFS was found in all colorectal samples (i) and for MSS colorectal samples (ii), but

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 15 of 23

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

with a notable lack of correlation for MSI samples (iii); see also fig. S12, A to C. FMRP CC subsignature score with comparatively poor OS in melanoma patients
(I) FMRP cancer signature score in human breast cancer, showing boxplots with treated with anti–PD-1. (ii) Association of a high FMRP CC subsignature score with
median, lower, and upper quantiles for the signature scores. (i) Comparison of CD8 worse PFS of lung cancer patients treated with anti–PD-1 or anti–PD-L1. (iii)
T cell infiltration score in high versus low FMRP cancer signature–scored tumor Association of a high FMRP CC subsignature score with poorer OS in urothelial cancer
samples. (ii) Comparison FMRP cancer signature scores in immune-excluded versus patients treated with anti–PD-L1. (K) Correlation of a high FMRP CC subsignature
inflamed breast cancer tumors (cohort: GSE177043). (iii) Comparison FMRP score with comparatively poor prognosis for cancer patients in response to different
cancer signature scores in low– versus high–TCR diversity breast cancer tumors forms of chemotherapy. (i) Association of a high FMRP CC subsignature score
(cohort: GSE177043). Statistics by Wilcoxon two-tailed test. (J) Correlation of the with reduced DFS of breast cancer patients treated with taxane. (ii) Association of
FMRP CC subsignature score (fig. S13K) with poor prognosis for different cancer a high FMRP CC subsignature score with poorer DFS of lung cancer patients
patients in response to immune checkpoint inhibitor therapy. (i) Association of a high treated with cisplatin, carboplatin, or paclitaxel.

breast cancer patients (Fig. 8I, panel i), which point inhibitors and biopsied for transcriptome models, including many tumor types that are
is again independent of tumor stage (fig. S13D). analysis. Notably, tumors with a high FMRP infrequently responsive to immune checkpoint
To further investigate the relevance of the FMRP CC signature score responded poorly or failed immunotherapy. A gene signature diagnostic of
signature, we leveraged a recently published to respond to anti–PD-1 or anti–PD-L1 thera- the cancer transcriptional network regulated
cohort of 43 primary breast cancer patients pies (fig. S13K), and a high FMRP CC sub- by FMRP was associated with worse DFS and
(59) that includes transcriptional profiling of signature score correlated significantly with OS and reduced inflammation by CD8 T cells in
tumor samples with matched clinical, im- poorer prognosis of treated patients (Fig. 8J). a pan-cancer analysis and in selected human
munohistochemistry and multiplexed immu- It is notable that the FMRP CC subsignature tumor types, consistent with the hypothesis
nofluorescence data collectively describing a was used in this analysis, which suggests that that FMRP is a cancer cell–intrinsic immuno-
multiparametric immunophenotype. Consistent the observed associations were solely based suppressor that contributes to immune eva-
with the results presented above, a high FMRP on intrinsic FMRP transcriptional cancer sion and the pathogenesis of human cancer.
cancer signature score was associated with network activity, without effects of the TME An intriguing question is why should a protein
low numbers of tumor-infiltrating lymphocytes involving indirectly induced immune response involved in synaptic transmission have this
(immune-excluded immunophenotype; Fig. 8I, genes. capability? One possibility is that FMRP serves
panel ii) and low TCR diversity in tumors (Fig. Finally, recognizing that optimal efficacy for to raise the bar to protect the brain against
8I, panel iii). Likewise, the FMRP CC subsig- certain forms of chemotherapy is now appre- precocious induction of T cell inflammation,
nature score showed a significant correlation ciated to require an adaptive immune response as suggested in a recent study in a rat model of
with this immune-excluded phenotype in (60–63), we assessed the FMRP CC subsigna- neuroinflammation (13).
the tumor samples of this dataset (fig. S13E). ture score in several cohorts of cancer patients Notably, we and others have previously im-
Although the transcriptomics data are a rich treated with chemotherapy. Consistent with plicated FMRP as a promoter of invasion and
resource for estimating CD8 T cell infiltration FMRP’s demonstrable role in programing an metastasis (15–17), which has not been a fo-
in tumor samples, to further validate our re- immunoevasive TME, a high FMRP cancer cus in this study. However, our bioinformatic
sults, we sought to evaluate possible associations cell–derived transcriptional cancer network studies did not significantly associate high sig-
at the protein level. To this end, a TMA cohort of signature had a statistically significant asso- nature scores for the FMRP cancer regulatory
breast cancer patients from Sweden Cancerome ciation with poorer prognosis in breast and lung network with tumor grade or stage in multiple
Analysis Network - Breast (SCAN-B) was used cancer patients in the context of treatment with forms of human cancer (fig. S12, I and K to O).
to assess FMRP protein expression level along- different forms of immune-dependent chemo- FMRP is an RNA binding protein that has
side CD8 T cell abundance through immunotherapy (Fig. 8K). These associations were again diverse functions in neurons (8, 64), of which
staining for both (fig. S13F). Using the paired independent of tumor stage (fig. S13L), em- the most prominent is the translational regu-
transcriptomics data for the TMA samples, we phasizing the role of the FMRP regulatory net- lation of hundreds of only partially overlap-
first revealed a significantly positive correla- work as an immunosuppressor throughout ping proteins in different cell types (7, 65–67).
tion between FMR1 mRNA and FMRP protein tumor progression. In this case again, the FMRP Additionally, FMRP binds to and potentially
levels (fig. S13G). Next, the TMA samples were CC subsignature was used, whereby the FMRP modulates the functions of multiple cellular
categorized as high or low for CD8 infiltration, regulatory network in cultured cancer cells proteins involved in diverse functions (66),
based on the immunostaining quantification. was queried for its collective diagnostic ef- and it also directly and indirectly regulates
Congruent with the previous results, no meaning- fects in the treated human tumors without the mRNA levels of hundreds of genes through
ful associations—in particular anticorrelations— potentially confounding effects of the immune- transcription, mRNA splicing, and mRNA
were observed between CD8 T cell abundance inflamed TME. stability (7, 8). Although the constellation of
and FMRP mRNA or protein expression lev- FMRP’s translational regulatory targets in
els (fig. S13, H and I). In marked contrast, the Discussion cancer cells has not yet been elaborated, we
FMRP cancer signature score for the SCAN-B Our findings suggest that FMRP limits the have shown by transcriptome profiling that
TMA samples again revealed a significant in- recruitment and expansion of CD8 and CD4 FMRP significantly modulates the mRNA
verse correlation with CD8 T cell abundance T cells, which allows solid tumors to evade levels of more than a hundred genes in WT
(fig. S13J), which further supports the role of immune destruction. This capability was re- versus FMRP-KO cancer cells and indirectly
FMRP transcriptional network activity in mod- vealed by experimental tumor models based affects a similar number of genes in the cell
ulating the immune response in human breast on engineered gene KOs of the Fmr1 gene types populating the TME, which supports
cancer and substantiates the relevance and (encoding the FMRP protein) in a series of the proposition that it is a key regulator of
potential utility of the FMRP cancer network cancer cell lines of different origins, all of which transcriptional and translational networks
signatures in the clinical setting. expressed endogenous FMRP at a high level. in cancer cells.
Next, we assessed datasets of melanoma, lung, The aberrant up-regulation of FMRP was evi- Our investigation suggests that FMRP cre-
and urothelial cancer patients that were treated dent at varying frequencies in a variety of hu- ates a barrier to immune recruitment that is
with and scored for responses to immune check- man cancers and genetically engineered mouse multifactorial, as are the immunostimulatory

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 16 of 23

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

effects in its absence. As described above, we cells of immunostimulatory macrophages ex- in accordance with the Swiss Federal Human
have begun to dissect the functional effectors pressing the proinflammatory chemokines Research Act (HFG 2014). All tissues were re-
of these dichotomous immune phenotypes. Ac- CCL5, CXCL9, and CXCL10, which indirectly trieved from the archives of the Institute of
cordingly, analysis of the TME in FMRP-WT contributed to the recruitment of CD8 T cells Pathology, University of Bern, Switzerland.
versus FMRP-KO tumors by scRNA-seq and that in turn amplified chemokine expression Tissue blocks were used for the construction
cellular annotation (Fig. 4) has revealed sig- and T cell inflammation through the secre- of the TMAs. Permission to use the patient
nificant alterations of the adaptive and in- tion of IFN-g. material and accompanying patient-related
nate immune compartments and has led to The molecular mechanisms governed by data for this study were approved by the Ethics
the identification and functional validation of FMRP and its absence that we have function- committee of the Canton of Bern and were
molecular determinants involved in the im- ally implicated in the immunosuppressed ver- collected in accordance with the 2014 Human
mune evasion versus tumor immunity orches- sus inflamed tumors are summarized in Fig. 7I. Research Act/Law (KEK#200/2014). Samples
trated by FMRP up-regulation versus its absence, It is evident, e.g., from the scRNA-seq analysis, were coded after they were retrieved from the
respectively (Fig. 7I). that FMRP—much like in neurons—is a key archives of the Institute of Pathology, Univer-
The immunoevasive TME in FMRP-expressing regulator of a multiplicity of genes, operating sity of Bern, and matched to the correspond-
tumors involves both Tregs and immunosup- both directly (and indirectly) in cancer cells and ing data. The details of the construction of
pressive (M2-like) macrophages (and poten- indirectly in cells of the TME. As such, we ex- those TMAs and have been described previ-
tially additional classes of myeloid cells), each pect that other components remain to be iden- ously (68, 69).
elicited by factors secreted by FMRP-expressing tified and functionally elucidated. The HCC TMA was constructed at the Insti-
cancer cells. The Tregs were induced by IL-33, Finally, consistent with the knowledge base tute of Pathology, University of Bern, Switzerland.
which we have implicated through CLIP-seq about FMRP’s complex posttranscriptional and Tissue blocks were collected for the construc-
as a direct target of FMRP translational con- posttranslational regulation in neurons (53), tion of the TMA. Permission to use patient ma-
trol. The importance of IL-33 was confirmed we have not been able to associate the levels of terial for this study was approved by the Ethics
in functional studies, wherein it was overex- FMR1 mRNA (or in the case of TNBC, the lev- committee of the Canton of Bern (KEK#07/10/
pressed in FMRP-KO cancer cells. Concordant- els of FMRP protein) with prognosis in human 13). A total of 119 patients diagnosed with HCC
ly, immunosuppressive M2-like macrophages cancers (fig. S12, A to C, and fig. S13, F to I), were included from 1991 to 2008. Multiple—
were induced by PROS1 secretion, implicated which implicates analogous regulatory mech- up to seven—core biopsies (0.6 mm) of HCC
by CLIP-seq as another direct target of FMRP anisms operative in cancer cells that warrant tumors were included for every patient. Clin-
activity, again functionally validated in co- future investigation. This disconnect motivated ical data except for the diagnosis of HCC are
culture assays and through overexpression in the development of gene signatures diagnos- not available.
FMRP-KO cancer cells. Concomitantly, cancer tic of FMRP’s regulation of a cancer network The prostate cancer brain metastasis TMA
cell–derived exosomes have been implicated (fig. S12D and table S4), which have revealed was constructed as part of the larger prostate
as yet another primary mechanistic effector FMRP’s downstream transcriptional network cancer brain metastases cohort (PCBM co-
of FMRP expression in cancer cells, whereby as a negative prognostic factor for many forms hort; N = 51) (unpublished; currently in peer-
these exosomes demonstrably induced im- of human cancer (Fig. 8; fig. S12, H and J; and review). The TMA contains 204 core biopsies
munoinhibitory gene expression in macro- fig. S13, A to C and K). Evidence that immune (each 1 mm in diameter) from prostate cancer
phages as part of their programming to become evasion is an important prognostic compo- brain metastases (PCBMs) as well as—in some
immunosuppressive. nent, congruent with the mechanistic studies cases—matched primary tumors (PTs) and
By contrast, the inflammatory TME com- presented herein, comes from the demonstra- benign tissue (BT), comprising 108, 66, and
posed of activated CD8 and CD4 T cells that tions that high FMRP transcriptional network 30 core biopsies, respectively. Multiple core
develops in FMRP-deficient tumors has sev- activity was associated with comparatively poor biopsies from both PCBM and matched pri-
eral mechanistic components originating in responses to immune checkpoint inhibitors in maries covering intratumoral heterogeneous
cancer cells lacking FMRP expression. IL-33 several patient cohorts for which transcriptome areas were included in the TMA, identified by
was down-regulated, which reduces the abun- datasets are available (Fig. 8J and fig. S13K) and a previous review protocol based on histomor-
dance of immunosuppressive Tregs, as did the with poor responses to certain chemotherapies phological features and immunohistochemical
up-regulation of otherwise FMRP-repressed (Fig. 8K and fig. S13L) for which the activity of expression. Additionally, core biopsies from BT
CCL7, which concomitantly contributed to the adaptive immune system has been impli- were incorporated when available. All analyses
the recruitment and proliferative capability of cated as an instrumental part of therapeutic were carried out in accordance with proto-
CD8 and CD4 T cells and the T cell–dependent efficacy (61). As such, it is possible that the cols approved by the Ethical Committee Bern
impairment of tumor growth. Additionally, described gene signatures for FMRP cancer (project ID: 2019 – 00328). One of the 5 TMA
PROS1 was down-regulated in FMRP-deficient network expression (or protein sets derived blocks constructed was used for the FMRP anal-
cancer cells, which contributed to the attenu- therefrom) could have utility in selecting pa- ysis. This block contains tissue from 15 different
ation of immunosuppressive macrophages, as tients more likely to respond to such immune- patients including 6 patients harboring all tis-
did the quantitative reduction and qualitative dependent therapies. Overall, the widespread sue types, 6 patients with PCBM and BT, and
modification of cancer cell–secreted exosomes. expression of FMRP in solid tumors, concom- 3 patients with PT and BT. Note that this is a
Although it is conceivable that unidentified itant with induction of its cancer regulatory cohort of a special subtype of prostate cancers
M1-inducing factors are up-regulated in FMRP- network, constitutes a previously unappreciated that develop brain metastasis, which may not
deficient cancer cells, these two effectors are mechanism whereby tumors evade immune reflect the subtypes of prostate cancers with-
demonstrably instrumental in inducing im- destruction. out brain metastasis.
munosuppressive M2-like macrophages, thereby The human TNBC TMA was constructed at
allowing, consequent to their down-regulation Materials and methods the University of Lund, in accordance with ap-
in FMRP-deficient cancer cells, the develop- Patient tumor TMA analysis propriate ethical approvals, as described previ-
ment of immunostimulatory (M1-like) macro- The human PDAC and colon cancer TMAs were ously (70). Sections on glass slides from this
phages. The net result is the appearance in constructed at the Institute of Pathology, Uni- TMA were sent to the Institute of Pathology
tumors populated by FMRP-deficient cancer versity of Bern. (Patient samples were collected in Bern for immunohistochemical analysis of

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 17 of 23

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

FMRP expression. Normal breast tissues were tech, France, lenti-rfpluc, lenti-gfpluc) accord- CGAGAACAGUUGAAACACA, purchased from
collected from women undergoing reduction ing to the manufacturer’s instructions. Sigma-Aldrich.
mammaplasties with no history of breast can-
cer, as previously described (71), and tissue sec- Engineering the Ccl7-KO and FMRP-KO PDAC cells Cell proliferation and viability assays
tions provided by the Brisken laboratory (EPFL, The Ccl7 gene was deleted in mouse PDAC Cell proliferation was measured using an MTT
Lausanne) were used for immunostaining to 4361.12 FMRP KO2 cells using the Alt-R CRISPR- cell proliferation kit (Roche, catalog no. 11 465
detect FMRP expression. Cas9 system (Integrated DNA Technologies), 007 001). Briefly, 1 × 104 cells were plated in
to produce so-called FMRP/Ccl7 DKO cells. triplicate in 96-well plates. 72 hours later, 10 ml
Generation of CRISPR-edited Fmr1-null cancer Cultured cancer cells were cotransfected with of MTT labeling reagent was added to each
cell lines Cas9 protein and CRISPR RNA (crRNA) tar- well and then incubated for 4 hours at 37°C,
Mouse PDAC 4361.12 cells, a single clone– geting Ccl7, and transactivating CRISPR RNA followed by the addition of 100 ml MTT solu-
derived cell line from the mouse PDAC cell line (tracrRNA) labeled with ATTO 550 (tracrRNA- bilization reagent overnight. Absorbance was
4361 (15) were cultured in Dulbecco’s mini- ATTO 550; Integrated DNA Technologies, catalog measured at 59 nm on a plate reader (Tecan
mum essential medium (DMEM) with 10% no. 1075927) using electroporation (Nepagene, Safire). For the CellTiter-Glo Luminescent Cell
fetal bovine serum (FBS) (Gibco, catalog no. NEPA21 Super Electroporator). Single cells Viability Assay (Promega, catalog no. G7570),
10270-106) with 100 U penicillin / 0.1 mg/ml were distributed into 96-well culture plates after 1 × 104 cells were seeded in one well of a 96-well
streptomycin (Thermo Fisher Scientific, cata- cells were recovered from the electroporation plate. 24 hours later, cell viability was measured
log no. 15140122). The Fmr1 gene was deleted chamber. Ccl7 KO clones were identified by according to the manufacturer’s protocol.
in cancer cells using the CRISPR-Cas9 system, enzyme-linked immunosorbent assay (ELISA)
to produce so-called FMRP-KO cells. Cultured analysis of supernatants for the lack of CCL7 se- Cell apoptosis assay
cancer cells were transiently cotransfected with cretion. The crRNA sequences used for Ccl7 are 1.5 × 105 cells were seeded into each well of a
Cas9 and single-guide RNA (sgRNA) expres- 5′-AAGCGTGGCAGAGATCCTCA-3′, 5′-CGTGG- 6-well tissue culture plate. 24 hours later, Annexin-
sion plasmids (72) and selected in blasticidin CAGAGATCCTCATGG-3′, and 5′-AAGCAGCG- V+/7-AAD− apoptotic cells were measured by flow
(10 ug/ml; Thermo Fisher Scientific, catalog no. CCTATGAGCAGC-3′. cytometry (Attune NxT, Thermo Fisher Scien-
A1113903) for 5 days. This transient CRISPR tific), according to the manufacturer’s protocol
strategy for the deletion of the Fmr1 gene Engineering the expression vectors to restore (BioLegend, catalog no. 640922).
avoids potential unspecific effects consequent expression of Pros1 and Il-33
to the stable integration of Cas9/sgRNA into the The overexpression plasmid pLenti-C-Myc- Cell migration assay
genome. The guide sequences used for Fmr1 DDK-P2A-Puro (OriGene, catalog no. PS100092), 5000 cells in 50 ml basal medium were seeded
were 5′-GTGGAAGTGCGGGGCTCCAA-3′ and as well as mouse Pros1 cDNA (OriGene, catalog into the top inserts of a 96-well Boyden chamber
5′- GAGCTGGTGGTGGAAGTGCG-3′. Single cells no. MR224303) and Il-33 cDNA (OriGene, catalog (Corning, catalog no. 3374); 200 ml of complete
were distributed into 96-well culture plates with- no. MR203528), were purchased from OriGene. medium was added to each bottom well. 18 hours
out blasticidin. Fmr1 KO clones were identified The insertion of the Pros1 and Il-33 cDNAs later, nonmigratory cells growing on the top
by immunoblotting of protein lysates for the into the overexpression plasmid used the Pre- membrane of the inserts were removed with a
lack of FMRP protein using two FMRP anti- cision Shuttling system from OriGene. The cotton swab infused with 70% ethanol, and mi-
bodies (table S5). The obtained Fmr1-deleted construction of lentiviruses using the over- gratory cells underneath the membrane were
(FMRP-KO) cells remained sensitive to blasticidin, expression plasmid followed the lentiviral pack- fixed, stained with crystal violet, and counted.
indicating the CRISPR-Cas9 system had been aging protocol provided by OriGene. FMRP KO2
only transiently expressed in those cells. Two cells were infected with lentivirus produced with Reverse transcription polymerase chain
independent FMRP-KO clones and two con- mock and Pros1 or Il-33 rescue vectors were se- reaction (RT-PCR)
trol WT clones (transfected with a Cas9 vector lected using Puromycin (2 ug/ml; Thermo Fisher RNA was isolated with the miRNeasy Mini Kit
and the empty sgRNA vector) were analyzed as Scientific, catalog no. A1113803). (Qiagen, catalog no. 217084) or RNeasy Plus
described in the results section. Micro Kit (Qiagen, catalog no. 74034). A total
Similarly, the Fmr1 gene encoding FMRP siRNA transfection of 500 ng of RNA was reverse transcribed into
was deleted in mouse CT26 colon cancer cells, Parental cells were seeded at 100,000 cells per cDNA using the PrimeScript RT Master Mix
4T1 breast cancer cells, and B16-OVA mela- well into a 6-well plate and transfected the next (Takara, catalog no. RR036A). cDNA levels were
noma cells. In brief, cultured cancer cells were day with 20 mM siRNA (5 ml per well, 100 pmol quantified using the SYBR green method in a
transiently cotransfected with Cas9 and an final) using RNAiMAX lipofectamine (5 ml QuantStudio 6 Flex Qpcr system in a 384-well
sgFmr1 expression plasmid or the empty sgRNA per well) in OptiMEM Reduced Serum Media format. 18S ribosomal RNA was used to equil-
vector and selected in blasticidin for 5 days. (Thermo Fisher Scientific, catalog no. 13778075). ibrate sample loading. The primers used in
Single cells were then distributed into 96-well The transfected cells were then used for in these studies are listed in table S6.
plates, and the FMRP KO clones were validated various bioassays, including the CD4/CD8 T
by immunoblotting. One WT and one or two coculture assay, the M1 macrophage coculture Protein (Western) blotting
FMRP-KO clones were analyzed as indicated for assay, and the CTL cell killing assay, 48 hours Cells were lysed in RIPA Lysis and Extraction
the described experiments. posttransfection. MISSION siRNA Univer- Buffer (Thermo Fisher Scientific, catalog no.
CT26 and 4T1 cells were cultured in RPI- sal Negative Control no. 1 and no. 2 (Sigma- 89900). Protein concentration was assessed using
1640, whereas B16-OVA cells were cultured in Aldrich, catalog nos. SIC001 and SIC002) were the BCA Protein Assay Kit (Thermo Fisher Sci-
DMEM, all with 10% FBS with 100 U penicillin / used as the siCtrl1 and siCtrl2 constructs, entific, catalog no. 23225), or the Bradford 1X
0.1 mg/ml streptomycin. All cell lines were respectively. The sequences used for siRNAs dye reagent (Bio-Rad). Samples were resolved
tested as mycoplasma negative. targeting Fmr1 (Stealth siRNA) were GAGUU- using Mini-Protean precast gels and transferred
For the mixed inoculation experiment, to CAAGGCAGCUUGCCUCAAGA (siFmr1) and to PDVF or NC membranes. Membranes were
fluorescently label the WT2 PDAC cells with GAGCUAGUUCUAGACCACCACCAAA (siFmr2), blocked with 5% BSA/TBST for 1 hour at room
RFP and FMRP-KO2 cells with GFP, cells were purchased from Thermo Fisher Scientific. temperature and incubated overnight at 4°C
infected with an RFP or GFP lentivirus (GEG- For siRNA targeting Myc, the sequence was with antibodies recognizing ARG1, IL-10, Hsp90,

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 18 of 23

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

FMRP, PROS1, Alix, CD81, GAPDH, Myc, and trial of anti–PD-1, which was started when the and KO2-GFP cells were mixed at 1:1 ratio; in
Ovabumin (table S5). Membranes were then tumors reached a volume of 100 mm3. total 5 × 105 cells suspended in 100 ml PBS were
probed with HRP-linked secondary antibodies injected s.c. On day 7, mice were anesthetized
and developed with WesternBright Sirius Analysis of tumor-infiltrating lymphocytes and perfused with 20 ml PBS and 20 ml 4% PFA.
(Advansta) or the PierceTM ECL Protein Blot- by flow cytometry Tumors were collected and fixed in 4% PFA
ting substrate (Thermo Fisher Scientific) using For intracellular cytokine staining, mice were for 4 more hours, and then embedded in 2%
Fusion FX7. injected 6 hours before euthanasia with 250 mg agarose in PBS and sectioned onto 100 mm
Brefaldin A (Sigma-Aldrich, catalog no. B7651). slides with a Leica Biosystems vibratome. Slides
ELISA for CCL7 To obtain single-cell suspensions, harvested were incubated with primary CD8 antibody
1 × 106 PDAC WT2, WT3, FMRP-KO2, and tumors were chopped into fragments, digested (table S5) at 4°C for 2 days, and then second-
FMRP-KO8 cells were seeded into a 100 mm in a mix of DNase I (144 U/mL, Roche, catalog ary antibody (goat anti-rat Alexa Fluor 647;
tissue culture plate with DMEM complete me- no. 11284932001), Dispase II (0.85 U/mL, Roche, Thermo Fisher Scientific, catalog no. A-21247)
dium on day 1. On day 3, medium was removed catalog no. 04 942 078 001) and Collagenase was applied overnight.
and replaced with 6 ml of serum-free DMEM A (0.33 U/mL, Roche, catalog no. 10 103 578 001), Images were variously acquired with a Leica
medium. On day 5, medium was collected into and passed through a 70-mm cell strainer. Cell DM5500B microscope, a Zeiss LSM 700 upright
15 ml tubes, and centrifuged at 2000 rpm in suspensions were then blocked with CD16/32 confocal microscopes, or an Olympus VS120
centrifuge at 4°C for 10 min. Supernatants were (clone 93, BioLegend) and labeled with live/ whole-slide scanner, and analyzed with Image
collected and diluted at 1:2 ratio. For the ELISA dead reagent (The LIVE/DEAD Fixable Red J. The antibodies used are shown in table S5.
analysis of CCL7 in WT2, KO2, and KO2-CCL7 Dead Cell Stain Kit, Thermo Fisher Scientific,
KO cells, 0.1 × 106 cells were seeded into one catalog no. L34971, or LIVE/DEAD Fixable Blue Ex vivo T cell coculture assays
well of 6-well plate in DMEM complete me- Dead Cell Stain Kit, for UV excitation, Thermo CD8 and CD4 T cells were isolated from a spleen
dium, and supernatants were collected 48 hours Fisher Scientific, catalog no. L12105). Surface cell suspension following the EasySep Mouse
later. The secretion of CCL7 in the supernatants staining was performed for 20 min on ice. CD8 or CD4 T cell isolation kits (StemCell
was analyzed using the mouse MCP3 ELISA Kit To detect biotinylated antibodies, samples Technologies, catalog no. 19853 or 19851, re-
(CCL7) (Abcam, catalog no. ab205571), following were incubated with fluorophore-conjugated spectively). Cells were then stained with CFSE
the manufacturer’s protocol. streptavidin. Cells were then fixed and perme- (BioLegend, catalog no. 422701) or Cell tracer
abilized with the Foxp3/Transcription Factor violet (Thermo Fisher Scientific, catalog no.
Animal studies Staining Buffer Set Kit (Invitrogen, catalog no. C34571) at 1 mM for 6 min and resuspended
All experiments using animals were performed 00-5523-00), and intracellular staining was in the RPMI media containing 10% FBS, 1%
in accordance with protocols approved by the performed in a Perm/Wash buffer overnight at PS, NEAA (Thermo Fisher Scientific, catalog
local animal experimentation committee of the 4°C. Samples were variously analyzed on Gallios no. 11140050) and b-mercaptoethanol (Thermo
Canton de Vaud (license no. 3214 and 3214.x). or Cytoflex (Beckman Coulter), LSRII Fortessa Fisher Scientific, catalog no. 21985023). T cells
FVB/N, Balb/c, C57Bl/6, NSG, or SCID/beige (BD), or Attune NxT (Thermo Fisher Scientific) were then activated with the CD3/CD28 Dyna-
mice were used at 8 weeks of age. For the lung flow cytometers. Data analysis was performed beads (Thermo Fisher Scientific, catalog no.
metastasis assay, 2 × 105 mouse PDAC WT or using FlowJo software. The antibodies used in 11456D) and plated at 200,000 cells per well in
FMRP-KO cells suspended in 200 ml phosphate- this analysis are shown in table S5. a 96-well plate. Cancer cells were then added
buffered saline (PBS) were injected into the at a 1:1 ratio.
lateral tail vein of the mice. For the PT growth Immunohistochemical and For the CD11b cocultures, tumors were har-
assay, 5 × 105 cells suspended in 100 ml PBS immunofluorescent staining vested and processed as described in the flow
were injected s.c. For the cell mixing injection Harvested mouse tissues were fixed in 4% para- cytometry section. CD11b cells were then iso-
experiment, PDAC WT-RFP and FMRP-KO- formaldehyde (PFA) overnight, embedded in lated from the cell suspension following the
GFP cells engineered to express RFP or Turbo paraffin, and sectioned using a microtome EasySep Mouse CD11b Positive Selection Kit II
GFP protein (as described above) were mixed (Leica). Antigen retrieval was performed in a (StemCell, catalog no. 18970). After isolation,
at different ratios; in total 5 × 105 cells sus- citrate buffer (pH = 6.0) in a water bath at cells were plated in a 1:1 ratio with T cells.
pended in 100 ml PBS were injected s.c. Mice 95°C for 20 min, or in a tris-EDTA buffer (pH = 72 hours later, the CSFE -low CD4 or CD8
were monitored twice per week and euthan- 8.0) in a water bath at 95°C for 10 min. Primary T cells were counted by flow cytometry (Gallios,
ized at the days indicated in the figure legends. antibodies were incubated at 4°C overnight. Beckman Coulter).
For the orthotopic tumor growth assay in For immunohistochemical (IHC) staining,
the pancreas, 10-week-old FVBn mice were secondary antibodies (ImmPRESS HRP re- In vitro T cell killing assay
injected with 200,000 cells in a total volume agent kit, anti-rabbit, catalog no. MP-7401, and 3 × 103 B16-OVA WT and FMRP-KO cells were
of 50 ml into the parenchyma of the pancreas anti-rat, catalog no. MP-7444) were incubated seeded into a 96-well plate for 48 hours (or
(in close proximity to the spleen). Mice were at room temperature for 45 min and finally vi- 24 hours). OT1 CD8+ T cells were isolated from
euthanized 2 weeks later and tumors collected sualized with the peroxidase substrate DAB the splenocytes of OVA-specific T cell receptor
for analysis. (Sigma-Aldrich, catalog no. D5637-1G) for the transgenic OT-1 mice, and then activated by
For pharmacological trials using function- same amount of time (maximum 10 min) at OVA peptide plus IL-2 and IL-7 (10 ng/ml;
blocking antibodies (table S5), mice were treated room temperature. The stained tissue sections BioLegend, catalog nos. 575406 and 577806)
intraperitoneally with 250 mg of a given anti- were counterstained with Meyer’s hematoxylin. for 5 days following a published protocol (73).
body per mouse twice per week. The CCR5 in- For immunofluorescent (IF) staining, second- 3 × 103 preactivated OT1 CD8 T cells were then
hibitor Maraviroc (Selleckchem, catalog no. ary antibodies (Alexa Fluor 488, 568, and 647; added to cancer cell culture at different ratios,
S2003) was prepared in PBS and administered Thermo Fisher Scientific, catalog nos. A-11006, with IL-2 (10 ng/ml) included to maintain
daily by intraperitoneal injection at 10 mg/kg. A-11077, and A-21247) were incubated at room their activity. 24 hours later, the viable cancer
All treatments were initially started on the same temperature for 45 min. cells were counted by flow cytometry (Attune
day as the s.c. cancer cell inoculation, and con- For the IF staining of fresh tumor slides from NxT, Thermo Fisher Scientific, or Cytoflex,
tinued for 2 weeks, except for the intervention the mixing injection experiment, the WT-RFP Beckman Coulter).

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 19 of 23

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

BMDM isolation 69504). Immunosequencing of the Mouse TCRb primary cancer cells for each tumor were loaded
Bone marrow extraction was performed as complementary determining region 3 (CDR3) into the 10X Genomics Chromium platform.
previously described (74). Cells were plated was performed using the ImmunoSEQ Assay Samples were processed following the manu-
in RPMI with 10% FBS, 1% Pen/Strep and (Adaptive Biotechnologies, Seattle, WA). In brief, facturer’s protocol with Single Cell 3′ v2 re-
50 ng/mL mouse macrophage colony-stimulating the mouse CDR3 regions of T cells from the agent and then sequenced using an Illumina
factor (BioLegend, catalog no. BLG-576406- whole-tumor genomic DNAs were amplified NextSeq sequencer. Raw data were mapped to
100 uG). On day 7, BMDMs were polarized through bias-controlled multiplex PCRs using mouse genome mm10-3.0.0 with cellranger-
using cytokines and cancer cells. For M1-like primers specific to TCR-Vb and TCR-Jb seg- 3.1.0. Downstream analyses were performed
polarization, 200 U/mL IFN-g (Proteintech, ments and then subjected to high-throughput in R (v.3.6.1) using Seurat (v.3.2.2) and scran
catalog no. 315-05-20UG) and 100 ng/mL LPS sequencing. The resulting raw sequences were (v.1.14.6) following the guidelines given in
(Sigma-Aldrich, catalog no. L4516-1MG) were filtered to identify and quantitate the absolute [Link]
used. To induce M2-like polarization, 20 ng/mL abundance of each individual TCR rearrange- We used standard Seurat settings for normal-
Il-4 (Sigma-Aldrich, catalog no. I1020-5UG) was ment in each sample. N = 3 mice for each group. ization, principal components analysis (PCA),
used. For the coculture experiments, PDAC WT t-distributed stochastic neighbor embedding
or FMRP KO cancer cells were seeded at 1 × 105 Mutational load and neoantigen prediction (tSNE) analysis, and clustering. For differen-
cells/mL into a 0.4-mm insert a day before the analysis in cell lines tial expression analysis, the Wilcoxon rank
polarization experiment. The next day, inserts WT and FMRP KO PDAC cells (WT2, WT3, sum test was applied, using 1.5 for the average
containing cancer cells were transferred to a KO2, and KO8) were cultured in DMEM with fold change in expression level, and 0.05 as the
6-well plate seeded with M1- or M2-polarized 10% FBS (Gibco) and 100 U penicillin / 0.1 mg/ml cutoff for adjusted P value.
BMDMs. Cells were harvested 48 hours after streptomycin (Gibco) for 30, 60, and 150 days;
polarization for further analyses. For FACS the cells were spit when 90% confluent and scRNA-seq cluster annotation
analyses, 1.5 × 105 cancer cells in total were plated passaged every 3 days. Genomic DNAs was To annotate different clusters of scRNA-seq
directly onto polarized BMDMs for 48 hours extracted from the cells at these different time analysis, we used cell type–specific gene sig-
and then collected and stained as described in points (Qiagen, catalog no. 69504), and whole- nature lists derived from multiple references
the flow cytometry section of the Materials and exome sequencing was performed using an (33, 77, 78), applying the singscore R package
methods. For the experiments involving Pros1, Illumina HiSeq 4000. (79) to derive signature scores for each gene-
macrophages were treated with 1 mg/mL mouse Exome sequences were mapped to the mouse list. Subsequently, scRNA-seq clusters were
recombinant Pros1 protein (R&D Systems, cata- genome using BWA (v.0.7.15), and variants were annotated according to the highest gene-list
log no. 9740-PS). The exosome cocultures fol- called using FreeBayes (v.1.1.0) against mm10. signature scores (fig. S7). Cell type–specific
lowed a previously described protocol (75, 76). Germline variants were filtered out using the gene signature lists were used as follows—
mouse strain FVBN/J germline variations [single- cancer cells: Krt18, Krt8, and Krt19; B cells:
Preparation of EVs nucleotide polymorphisms (SNPs) and indels] Ms4a1, Bank1, Cd79b, Fcer2a, Pax5, Cd79a, Tcl1,
Isolation of EVs was performed as previously stored in the Mouse Genome Project (www. Stap1, Fcrl5, Pou2af1, Ptprc, and Cd19; lympho-
reported (75, 76). Briefly, PDAC cells were [Link]/science/data/mouse-genomes- cytes: Cd5, Ubash3a, Il7r, Itk, Cd28, Themis,
plated in 15-cm cell culture dishes. When cells project). Bcl11b, Cd3d, Cd3e, and Cd3g; DCs: Itgax, Cst3,
reached 50% confluency, complete medium Variant annotation was performed with and Cd74; migDC: Ccr7, Ccl22, and Fscn1; cDC1:
was replaced with 5% EV-depleted FBS me- VariantAnnotation R package. Amino acid Itgae, Clec9a, Ccl17, and Xcr1; pDC: Bst2, Ly6d,
dium. The EVs were collected after ~2 to 3 days coding changes were predicted for the nonsyn- Siglech, and Tcf4; neutrophils: Pglyrp1, Csf3r,
using sequential centrifugation, as described. onymous variants; the SIFT (sorting intolerant Cyp4f18, and Itgam; myeloid cells: Kit, Cd34,
Quantification was performed with the BCA from tolerant) and PolyPhen (polymorphism Ly6a, Kcnip3, Med21, Sgms2, Tfec, Itgam, and
Protein Assay Kit. phenotyping) algorithms provided predictions Ly6g; monocytes: Ms4a6c, Ms4a6b, Ms4a6d,
of how severely the coding changes might be and Ly6c2; macrophages: Adgre1, Cd68, Csf1r,
Analyzing ENCODE ChIP-seq data for TFs affecting protein function. The SIFT method Fcgr1, Cd14, C1qb, C1qa, Flt3, Cx3cr1, and Vcam1;
binding to the FMR1 promoter region uses sequence homology and the physical prop- CAFs: Col1a1, Col1a2, Pdpn, and Dcn; myCAFs:
All human MAX and MYC ChIP-seq data from erties of amino acids to make predictions about Tagln, Thy1, Col12a1, and Thbs2; iCAFs: Clec3b,
ENCODE database were downloaded. The fol- protein function. PolyPhen uses sequence-based Col14a1, Has1, and Il6; pericytes: Acta2, Higd1b,
lowing two types of files are used: (i) .bed files features and structural information charac- Rgs5, and Pdgfrb; and endothelial cells: Pecam1.
were used to find TFs’ binding location on terizing the substitution to make predictions After annotating each cluster, to ensure that
FMR1’s promoter, for which the peak calling about the structure and function of the pro- the clusters within each cell type had similar
algorithm has already been applied, and (ii). tein. The remaining variants were remapped distributions of the total number of features
bigWig files were used for representing the to the FVBN/J_v1 genome for neoantigen pre- (reads), we filtered out the clusters that had a
data (fig. S3). diction. MuPeXi was used to predict all mu- low number of reads (average < 1000 read)
We have also used the Ensembl database for tated peptides (neopeptides) of lengths 8 to such that their distributions were significantly
selecting FMR1’s promoter region: Chromo- 11, using FVBN/J-specific peptide and cDNA lower (3 standard deviations lower) than the
some X: 147,911,000-147,915,201, whose stable references. rest of the clusters with the same annotations.
ID is ENSR00000249258. The result for the Therefore, one cancer cell cluster and two mye-
significant peak is presented in table S7. scRNA-seq analysis loid cell clusters with very low number of reads
WT (WT2, WT3) and FMRP-KO (KO2, KO8) were excluded from the subsequent analysis.
TCR sequencing PDAC cancer cells (5 × 105) were separately
WT2 or FMRP-KO2 PDAC cancer cells (5 × injected s.c. into the right flank of FVBN mice. Lymphocyte analysis, ProjecTILs
105) were s.c. injected into the right flank of 14 days later, PDAC tumors were dissected and All clusters annotated as lymphocytes were
FVBn mice. 14 days later, PDAC tumors were digested by collagenase II, IV (0.25%), and subselected, and further analyzed for lympho-
dissected, and genomic DNA extracted (using DNase I (0.05%) in DMEM/F12 medium for cyte subtype annotations. To this end, we used
Qiagen DNeasy Blood & Tissue Kit, catalog no. 30 min at 37°C. Cells were counted, and 8 × 103 the ProjecTILs pipeline, implemented in an R

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 20 of 23

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

package (33), which is an algorithm that accu- using the Illumina Pipeline Software version and FMR1 expression with OS. We applied the
rately overlays scRNA-seq data onto reference 1.82. Adapter sequences and low-quality ends large-sample Chi-square test (log-rank test) to
T cell atlases, and characterizes different cell were removed with cutadapt,1.4.2, trimming determine the associations between predictor
states and clusters. for TrueSeq and polyA sequences. Reads were variables and to obtain adjusted hazard-ratios.
To assess natural killer (NK) cells, which aligned to mouse genome build mm10 using These analyses were performed with the R
are not incorporated into ProjecTILs, we used HISAT2 aligner. Normalization was performed package “survival.”
scGate (80), an algorithm that identifies scRNA- using DESeq2 pipeline, by normalizing with
seq populations based on both positive and neg- the size factor. This method is implemented in Quantification and statistical analysis
ative markers. We applied scGate onto the whole the R Bioconductor package DESeq2. Differ- All statistical analyses described above were per-
lymphocyte subpopulation using NCAM1pos, ential expression analysis was performed by formed either with R, or with Prism 7 (GraphPad
KLRD1pos, CD3Dneg as the virtual gating strategy, DESeq pipeline (DESeq2 R package). The cut- Software).
and annotated the positive cells as the NK cell offs for gene selection were adjusted P < 0.05 REFERENCES AND NOTES
population. and fold change > 1.5. 1. J. Gill, V. Prasad, A reality check of the accelerated approval of
immune-checkpoint inhibitors. Nat. Rev. Clin. Oncol. 16,
Cell-cell communication analysis in scRNA-seq data Development of gene signatures for FMRP 656–658 (2019). doi: 10.1038/s41571-019-0260-y;
pathway activity pmid: 31383994
To infer the outgoing communication patterns 2. R. W. Jenkins, D. A. Barbie, K. T. Flaherty, Mechanisms of
between different cell types in the scRNA-seq To identify a gene signature reflecting FMRP resistance to immune checkpoint inhibitors. Br. J. Cancer 118,
data, we applied the CellChat pipeline (34), downstream transcriptional network, we de- 9–16 (2018). doi: 10.1038/bjc.2017.434; pmid: 29319049
3. A. J. Schoenfeld, M. D. Hellmann, Acquired Resistance to
implemented in the R package ([Link] veloped three different signature lists, as fol- Immune Checkpoint Inhibitors. Cancer Cell 37, 443–455
com/sqjin/CellChat), following the pipeline lows (fig. S11D): (2020). doi: 10.1016/[Link].2020.03.017; pmid: 32289269
provided by the developers. Accordingly, the 1) Differentially expressed genes (fold change > 4. M. S. Sidorov, B. D. Auerbach, M. F. Bear, Fragile X mental
retardation protein and synaptic plasticity. Mol. Brain 6, 15
cell-cell communication patterns between cell 1.5, adjusted P < 0.05) were identified, compar- (2013). doi: 10.1186/1756-6606-6-15; pmid: 23566911
types within each condition, namely WT and ing WT and FMRP-KO cell lines. Because we 5. J. C. Darnell et al., FMRP stalls ribosomal translocation on
FMRP-KO tumors, were separately inferred, had two baches of FMRP-KO2 clones, only dif- mRNAs linked to synaptic function and autism. Cell 146, 247–261
(2011). doi: 10.1016/[Link].2011.06.013; pmid: 21784246
using the “computeCommunProb” function. ferentially expressed genes in both batches were 6. Y. Feng et al., FMRP associates with polyribosomes as an
Then, the respective ligand-receptor networks selected, and then the union of these genes with mRNP, and the I304N mutation of severe fragile X syndrome
from the two conditions were merged using the differentially expressed genes in the FMRP- abolishes this association. Mol. Cell 1, 109–118 (1997).
“mergeCellChat” function. Subsequently, to KO8 clone constituted the list of genes in FMRP doi: 10.1016/S1097-2765(00)80012-X; pmid: 9659908
7. M. Li et al., Identification of FMR1-regulated molecular
identify the up- and down-regulated signaling CC network subsignature. networks in human neurodevelopment. Genome Res. 30,
ligand-receptor pairs between WT and FMRP- 2) Differentially expressed genes (fold change > 361–374 (2020). doi: 10.1101/gr.251405.119; pmid: 32179589
KO tumors, we used the “netVisual_bubble” func- 1.5, adjusted P < 0.05) between WT and FMRP- 8. J. D. Richter, X. Zhao, The molecular biology of FMRP: New
insights into fragile X syndrome. Nat. Rev. Neurosci. 22,
tion with the default parameters in the package. KO tumors from the scRNA-seq data were iden- 209–222 (2021). doi: 10.1038/s41583-021-00432-0;
To further filter the identified interactions, the tified and constitute the list of genes in FMRP pmid: 33608673
differentially expressed genes (comparing WT tumor network subsignature. In this analysis, 9. C. Malecki, B. D. Hambly, R. W. Jeremy, E. N. Robertson,
The RNA-binding fragile-X mental retardation protein and its
with FMRP-KO tumors) were extracted for we included all the cancer cells detected using role beyond the brain. Biophys. Rev. 12, 903–916 (2020).
the source cell type (the cell population express- Krt8 and Krt18 marker genes, regardless of the doi: 10.1007/s12551-020-00730-4; pmid: 32654068
ing the ligands), and then only the interac- total number of features. 10. E. J. Mientjes et al., The generation of a conditional Fmr1 knock
out mouse model to study Fmrp function in vivo. Neurobiol.
tions for which the ligands are within this list 3) The union of these two gene lists constitutes Dis. 21, 549–555 (2006). doi: 10.1016/[Link].2005.08.019;
of differentially expressed genes were selected the main FMRP cancer network signature. pmid: 16257225
and plotted as the up- or down-regulated sig- 11. The Dutch-Belgian Fragile X Consortium, Fmr1 knockout mice:
HALLMARK terms enrichment analysis A model to study fragile X mental retardation. Cell 78, 23–33
naling ligand-receptor pairs between two cell (1994). doi: 10.1016/0092-8674(94)90569-X; pmid: 8033209
populations. HALLMARK enrichment analysis was per- 12. E. Donnard, H. Shu, M. Garber, Single cell transcriptomics
formed using Molecular Signatures Database reveals dysregulated cellular and molecular networks in a
mRNA sequencing analysis fragile X syndrome model. PLOS Genet. 18, e1010221 (2022).
(MSigDB), available through an online service. doi: 10.1371/[Link].1010221; pmid: 35675353
We performed RNA-seq analysis for two inde- False discovery rate (FDR)–adjusted P < 0.05, 13. P.-Y. Nie et al., miR-142 downregulation alleviates rat PTSD-like
pendent KO clones and WT clones. For one retrieving a maximum of 20 terms. behaviors, reduces the level of inflammatory cytokine
expression and apoptosis in hippocampus, and upregulates
clone of each condition (KO2 and WT2), we the expression of fragile X mental retardation protein.
performed the RNA-seq in two separate baches. Gene signature enrichment analysis
J. Neuroinflammation 18, 17 (2021). doi: 10.1186/s12974-020-
As for PDAC WT2 cells with siCTLR and siFMR1 Single-sample gene set enrichment analysis 02064-0; pmid: 33407653
14. D. Prilutsky et al., Gene expression analysis in Fmr1KO mice
transfection, we performed the RNA-seq in one (ssGSEA) (81) was used to calculate an expression identifies an immunological signature in brain tissue and
batch. RNA-seq libraries were generated from score for gene expression signatures. The method mGluR5-related signaling in primary neuronal cultures.
double-stranded cDNA derived from 10 ng total was implemented in the R package GSVA. Mol. Autism 6, 66 (2015). doi: 10.1186/s13229-015-0061-9;
pmid: 26697163
RNA with the NuGEN Ovation RNA-Seq Sys- 15. L. Li et al., GKAP Acts as a Genetic Modulator of NMDAR Signaling
tem V2 (NuGEN Technologies, San Carlos, CA). Correlation analysis
to Govern Invasive Tumor Growth. Cancer Cell 33, 736–751.e5
100 ng of the cDNA was fragmented to 350 bp To evaluate the correlation between two con- (2018). doi: 10.1016/[Link].2018.02.011; pmid: 29606348
using Covaris S2 (Covaris, Woburn, MA). Se- tinuous variables, “cor” function was used in R 16. R. Lucá et al., The fragile X protein binds mRNAs involved in
cancer progression and modulates metastasis formation.
quencing libraries were prepared from the to develop the Pearson correlation coefficient. EMBO Mol. Med. 5, 1523–1536 (2013). doi: 10.1002/
fragmented cDNA with the Illumina TruSeq To estimate the significance of the correlation, emmm.201302847; pmid: 24092663
Nano DNA Library Prep Kit (Illumina, San we used the “lm” function in R for linear mod- 17. F. Zalfa et al., The fragile X mental retardation protein
regulates tumor invasiveness-related pathways in melanoma
Diego, CA). Cluster generation was performed eling and retrieved FDR-adjusted P values. cells. Cell Death Dis. 8, e3169 (2017). doi: 10.1038/
with the libraries using the Illumina TruSeq cddis.2017.521; pmid: 29144507
SR Cluster Kit v4 reagents and sequenced on Survival analysis 18. A. Di Grazia et al., The Fragile X Mental Retardation Protein
Regulates RIPK1 and Colorectal Cancer Resistance to
the Illumina HiSeq 2500 with TruSeq SBS Kit Kaplan-Meier survival analysis was used to Necroptosis. Cell. Mol. Gastroenterol. Hepatol. 11, 639–658
v4 reagents. Sequencing data were processed assess the relationship of the signature scores (2021). doi: 10.1016/[Link].2020.10.009; pmid: 33091622

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 21 of 23

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

19. G. Pedini et al., FMRP modulates the Wnt signalling pathway in 764222 (2021). doi: 10.3389/fmolb.2021.764222; 64. I. D’Annessa, F. Cicconardi, D. Di Marino, Handling FMRP and
glioblastoma. Cell Death Dis. 13, 719 (2022). doi: 10.1038/ pmid: 34722637 its molecular partners: Structural insights into Fragile X
s41419-022-05019-w; pmid: 35982038 44. Z. Jiang et al., Cancer derived exosomes induce macrophages Syndrome. Prog. Biophys. Mol. Biol. 141, 3–14 (2019).
20. M. Uhlén et al., A human protein atlas for normal and cancer immunosuppressive polarization to promote bladder cancer doi: 10.1016/[Link].2018.07.001; pmid: 30905341
tissues based on antibody proteomics. Mol. Cell. Proteomics 4, progression. Cell Commun. Signal. 19, 93 (2021). doi: 10.1186/ 65. E. Pasciuto, C. Bagni, SnapShot: FMRP mRNA targets and
1920–1932 (2005). doi: 10.1074/mcp.M500279-MCP200; s12964-021-00768-1; pmid: 34521440 diseases. Cell 158, 1446–1446.e1 (2014). doi: 10.1016/
pmid: 16127175 45. M. T. Chow et al., Intratumoral Activity of the CXCR3 [Link].2014.08.035; pmid: 25215498
21. F. Sanchez-Vega et al., Oncogenic Signaling Pathways in The Chemokine System Is Required for the Efficacy of Anti-PD- 66. M. S. Taha et al., Novel FMRP interaction networks linked to
Cancer Genome Atlas. Cell 173, 321–337.e10 (2018). 1 Therapy. Immunity 50, 1498–1512.e5 (2019). doi: 10.1016/ cellular stress. FEBS J. 288, 837–860 (2021). doi: 10.1111/
doi: 10.1016/[Link].2018.03.035; pmid: 29625050 [Link].2019.04.010; pmid: 31097342 febs.15443; pmid: 32525608
22. Y. Luo et al., New developments on the Encyclopedia of DNA 46. I. G. House et al., Macrophage-Derived CXCL9 and CXCL10 67. C. R. Hale et al., FMRP regulates mRNAs encoding distinct
Elements (ENCODE) data portal. Nucleic Acids Res. 48, Are Required for Antitumor Immune Responses Following Immune functions in the cell body and dendrites of CA1 pyramidal neurons.
D882–D889 (2020). doi: 10.1093/nar/gkz1062; Checkpoint Blockade. Clin. Cancer Res. 26, 487–504 (2020). eLife 10, e71892 (2021). doi: 10.7554/eLife.71892; pmid: 34939924
pmid: 31713622 doi: 10.1158/[Link]-19-1868; pmid: 31636098 68. C. Schafroth et al., VE1 immunohistochemistry predicts
23. S. Pelengaris, M. Khan, G. Evan, c-MYC: More than just a 47. N. Karin, CXCR3 Ligands in Cancer and Autoimmunity, BRAF V600E mutation status and clinical outcome in
matter of life and death. Nat. Rev. Cancer 2, 764–776 (2002). Chemoattraction of Effector T Cells, and Beyond. Front. colorectal cancer. Oncotarget 6, 41453–41463 (2015).
doi: 10.1038/nrc904; pmid: 12360279 Immunol. 11, 976 (2020). doi: 10.3389/fimmu.2020.00976; doi: 10.18632/oncotarget.6162; pmid: 26496026
24. N. M. Sodir et al., MYC Instructs and Maintains Pancreatic pmid: 32547545 69. M. Wartenberg et al., Integrated Genomic and Immunophenotypic
Adenocarcinoma Phenotype. Cancer Discov. 10, 588–607 48. J. Liu et al., Local production of the chemokines CCL5 and Classification of Pancreatic Cancer Reveals Three Distinct
(2020). doi: 10.1158/[Link]-19-0435; pmid: 31941709 CXCL10 attracts CD8+ T lymphocytes into esophageal Subtypes with Prognostic/Predictive Significance. Clin. Cancer
25. J. C. Castle et al., Immunomic, genomic and transcriptomic squamous cell carcinoma. Oncotarget 6, 24978–24989 (2015). Res. 24, 4444–4454 (2018). doi: 10.1158/[Link]-17-
characterization of CT26 colorectal carcinoma. BMC Genomics 15, doi: 10.18632/oncotarget.4617; pmid: 26317795 3401; pmid: 29661773
190 (2014). doi: 10.1186/1471-2164-15-190; pmid: 24621249 49. P. M. Marcovecchio, G. Thomas, S. Salek-Ardakani, 70. J. Staaf et al., Whole-genome sequencing of triple-negative
26. J. R. Brahmer et al., Safety and activity of anti-PD-L1 antibody CXCL9-expressing tumor-associated macrophages: New breast cancers in a population-based clinical study. Nat.
in patients with advanced cancer. N. Engl. J. Med. 366, players in the fight against cancer. J. Immunother. Cancer 9, Med. 25, 1526–1533 (2019). doi: 10.1038/s41591-019-0582-4;
2455–2465 (2012). doi: 10.1056/NEJMoa1200694; e002045 (2021). doi: 10.1136/jitc-2020-002045; pmid: 31570822
pmid: 22658128 pmid: 33637602 71. G. Sflomos, M. Shamsheddin, C. Brisken, An ex vivo model to
27. N. Pu, W. Lou, J. Yu, PD-1 immunotherapy in pancreatic 50. T. J. Zumwalt, M. Arnold, A. Goel, C. R. Boland, Active study hormone action in the human breast. J. Vis. Exp. 95,
cancer: Current status. J. Pancreatol. 2, 6–10 (2019). secretion of CXCL10 and CCL5 from colorectal cancer e52436 (2015). doi: 10.3791/52436; pmid: 25590412
doi: 10.1097/JP9.0000000000000010 microenvironments associates with GranzymeB+ CD8+ T-cell 72. P. Mali et al., RNA-guided human genome engineering via Cas9.
28. T. R. Medler et al., Complement C5a Fosters Squamous infiltration. Oncotarget 6, 2981–2991 (2015). doi: 10.18632/ Science 339, 823–826 (2013). doi: 10.1126/science.1232033;
Carcinogenesis and Limits T Cell Response to Chemotherapy. oncotarget.3205; pmid: 25671296 pmid: 23287722
Cancer Cell 34, 561–578.e6 (2018). doi: 10.1016/[Link].2018. 51. G. G. Xu, J. Guo, Y. Wu, Chemokine receptor CCR5 antagonist 73. Y. Guo et al., Metabolic reprogramming of terminally exhausted
09.003; pmid: 30300579 maraviroc: Medicinal chemistry and clinical applications. CD8+ T cells by IL-10 enhances anti-tumor immunity. Nat.
29. A. Y. Huang et al., The immunodominant major Curr. Top. Med. Chem. 14, 1504–1514 (2014). doi: 10.2174/ Immunol. 22, 746–756 (2021). doi: 10.1038/s41590-021-
histocompatibility complex class I-restricted antigen of a 1568026614666140827143745; pmid: 25159165 00940-2; pmid: 34031618
murine colon tumor derives from an endogenous retroviral 52. R. Uppaluri et al., Prolongation of cardiac and islet allograft 74. G. Galliverti et al., Myeloid Cells Orchestrate Systemic
gene product. Proc. Natl. Acad. Sci. U.S.A. 93, 9730–9735 survival by a blocking hamster anti-mouse CXCR3 monoclonal Immunosuppression, Impairing the Efficacy of
(1996). doi: 10.1073/pnas.93.18.9730; pmid: 8790399 antibody. Transplantation 86, 137–147 (2008). doi: 10.1097/ Immunotherapy against HPV+ Cancers. Cancer Immunol. Res.
30. T. Moroishi et al., The Hippo Pathway Kinases LATS1/2 TP.0b013e31817b8e4b; pmid: 18622291 8, 131–145 (2020). doi: 10.1158/[Link]-19-0315:
Suppress Cancer Immunity. Cell 167, 1525–1539.e17 (2016). 53. M. Prieto, A. Folci, S. Martin, Post-translational modifications of pmid: 31771984
doi: 10.1016/[Link].2016.11.005; pmid: 27912060 the Fragile X Mental Retardation Protein in neuronal function 75. C. Cianciaruso et al., Molecular Profiling and Functional
31. L. M. Kranz et al., Systemic RNA delivery to dendritic cells exploits and dysfunction. Mol. Psychiatry 25, 1688–1703 (2020). Analysis of Macrophage-Derived Tumor Extracellular Vesicles.
antiviral defence for cancer immunotherapy. Nature 534, 396–401 doi: 10.1038/s41380-019-0629-4; pmid: 31822816 Cell Rep. 27, 3062–3080.e11 (2019). doi: 10.1016/
(2016). doi: 10.1038/nature18300; pmid: 27281205 54. C. M. Rodriguez et al., A native function for RAN translation [Link].2019.05.008; pmid: 31167148
32. X. Zhang et al., CellMarker: A manually curated resource of cell and CGG repeats in regulating fragile X protein synthesis. 76. I. Keklikoglou et al., Chemotherapy elicits pro-metastatic
markers in human and mouse. Nucleic Acids Res. 47, D721–D728 Nat. Neurosci. 23, 386–397 (2020). doi: 10.1038/s41593-020- extracellular vesicles in breast cancer models. Nat. Cell Biol.
(2019). doi: 10.1093/nar/gky900; pmid: 30289549 0590-1; pmid: 32066985 21, 190–202 (2019). doi: 10.1038/s41556-018-0256-3;
33. M. Andreatta et al., Interpretation of T cell states from single-cell 55. S. Ogino et al., Lymphocytic reaction to colorectal cancer is pmid: 30598531
transcriptomics data using reference atlases. Nat. Commun. 12, associated with longer survival, independent of lymph node 77. E. Elyada et al., Cross-Species Single-Cell Analysis of
2965 (2021). doi: 10.1038/s41467-021-23324-4; pmid: 34017005 count, microsatellite instability, and CpG island methylator Pancreatic Ductal Adenocarcinoma Reveals Antigen-Presenting
34. S. Jin et al., Inference and analysis of cell-cell phenotype. Clin. Cancer Res. 15, 6412–6420 (2009). Cancer-Associated Fibroblasts. Cancer Discov. 9, 1102–1123
communication using CellChat. Nat. Commun. 12, 1088 doi: 10.1158/[Link]-09-1438; pmid: 19825961 (2019). doi: 10.1158/[Link]-19-0094; pmid: 31197017
(2021). doi: 10.1038/s41467-021-21246-9; pmid: 33597522 56. S. Kang et al., The significance of microsatellite instability in 78. S. Cheng et al., A pan-cancer single-cell transcriptional
35. T. Maurin et al., HITS-CLIP in various brain areas reveals colorectal cancer after controlling for clinicopathological atlas of tumor infiltrating myeloid cells. Cell 184, 792–809.e23
new targets and new modalities of RNA binding by fragile X factors. Medicine 97, e0019 (2018). doi: 10.1097/ (2021). doi: 10.1016/[Link].2021.01.010; pmid: 33545035
mental retardation protein. Nucleic Acids Res. 46, 6344–6355 MD.0000000000010019; pmid: 29489646 79. M. Foroutan et al., Single sample scoring of molecular
(2018). doi: 10.1093/nar/gky267; pmid: 29668986 57. G. Golshani, Y. Zhang, Advances in immunotherapy for phenotypes. BMC Bioinformatics 19, 404 (2018). doi: 10.1186/
36. Y. Liu, Y. Cai, L. Liu, Y. Wu, X. Xiong, Crucial biological colorectal cancer: A review. Therap. Adv. Gastroenterol. 13, s12859-018-2435-4; pmid: 30400809
functions of CCL7 in cancer. PeerJ 6, e4928 (2018). 1756284820917527 (2020). doi: 10.1177/1756284820917527; 80. M. Andreatta, A. J. Berenstein, S. J. Carmona, scGate:
doi: 10.7717/peerj.4928; pmid: 29915688 pmid: 32536977 Marker-based purification of cell types from heterogeneous
37. A. Hatzioannou et al., An intrinsic role of IL-33 in Treg cell– 58. S. Bayraktar, S. Batoo, S. Okuno, S. Glück, Immunotherapy single-cell RNA-seq datasets. Bioinformatics 38,
mediated tumor immunoevasion. Nat. Immunol. 21, 75–85 in breast cancer. J. Carcinog. 18, 2 (2019). doi: 10.4103/ 2642–2644 (2022). doi: 10.1093/bioinformatics/btac141;
(2020). doi: 10.1038/s41590-019-0555-2; pmid: 31844326 jcar.JCar_2_19; pmid: 31160888 pmid: 35258562
38. D. G. DeNardo, B. Ruffell, Macrophages as regulators of 59. D. Hammerl et al., Spatial immunophenotypes predict response 81. D. A. Barbie et al., Systematic RNA interference reveals that
tumour immunity and immunotherapy. Nat. Rev. Immunol. 19, to anti-PD1 treatment and capture distinct paths of T cell oncogenic KRAS-driven cancers require TBK1. Nature 462,
369–382 (2019). doi: 10.1038/s41577-019-0127-6; evasion in triple negative breast cancer. Nat. Commun. 12, 5668 108–112 (2009). doi: 10.1038/nature08460; pmid: 19847166
pmid: 30718830 (2021). doi: 10.1038/s41467-021-25962-0; pmid: 34580291 82. S. R. Hingorani et al., Trp53R172H and KrasG12D cooperate
39. S. Yan, G. Wan, Tumor-associated macrophages in immunotherapy. 60. L. Galluzzi, J. Humeau, A. Buqué, L. Zitvogel, G. Kroemer, to promote chromosomal instability and widely metastatic
FEBS J. 288, 6174–6186 (2021). doi: 10.1111/febs.15726; Immunostimulation with chemotherapy in the era of immune pancreatic ductal adenocarcinoma in mice. Cancer Cell 7,
pmid: 33492779 checkpoint inhibitors. Nat. Rev. Clin. Oncol. 17, 725–741 469–483 (2005). doi: 10.1016/[Link].2005.04.023;
40. E. Ubil et al., Tumor-secreted Pros1 inhibits macrophage M1 (2020). doi: 10.1038/s41571-020-0413-z; pmid: 32760014 pmid: 15894267
polarization to reduce antitumor immune response. J. Clin. 61. J. W. Opzoomer, D. Sosnowska, J. E. Anstee, J. F. Spicer, 83. M. Germann et al., Neutrophils suppress tumor-infiltrating
Invest. 128, 2356–2369 (2018). doi: 10.1172/JCI97354; J. N. Arnold, Cytotoxic Chemotherapy as an Immune Stimulus: T cells in colon cancer via matrix metalloproteinase-mediated
pmid: 29708510 A Molecular Perspective on Turning Up the Immunological activation of TGFb. EMBO Mol. Med. 12, e10681 (2020).
41. M. S. Pols, J. Klumperman, Trafficking and function of the Heat on Cancer. Front. Immunol. 10, 1654 (2019). doi: 10.3389/ doi: 10.15252/emmm.201910681; pmid: 31793740
tetraspanin CD63. Exp. Cell Res. 315, 1584–1592 (2009). fimmu.2019.01654; pmid: 31379850 84. I. G. Maroulakou, M. Anver, L. Garrett, J. E. Green, Prostate and
doi: 10.1016/[Link].2008.09.020; pmid: 18930046 62. Y. H. Park et al., Chemotherapy induces dynamic immune mammary adenocarcinoma in transgenic mice carrying a
42. A. L. Wozniak et al., The RNA binding protein FMR1 controls responses in breast cancers that impact treatment outcome. rat C3(1) simian virus 40 large tumor antigen fusion gene.
selective exosomal miRNA cargo loading during inflammation. Nat. Commun. 11, 6175 (2020). doi: 10.1038/s41467-020- Proc. Natl. Acad. Sci. U.S.A. 91, 11236–11240 (1994).
J. Cell Biol. 219, e201912074 (2020). doi: 10.1083/ 19933-0; pmid: 33268821 doi: 10.1073/pnas.91.23.11236; pmid: 7972041
jcb.201912074; pmid: 32970791 63. L. Zitvogel, L. Apetoh, F. Ghiringhelli, G. Kroemer, Immunological 85. Q. Zeng et al., Aberrant hyperexpression of the RNA binding
43. Q. Chen et al., Exosome-Mediated Crosstalk Between Tumor aspects of cancer chemotherapy. Nat. Rev. Immunol. 8, 59–73 protein FMRP in tumors mediates immune evasion, Zenodo
and Tumor-Associated Macrophages. Front. Mol. Biosci. 8, (2008). doi: 10.1038/nri2216; pmid: 18097448 (2022); [Link]

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 22 of 23

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

ACKN OW LEDG MEN TS Paulsson Foundations (Lund) and from Goran Grosskopf (Lund), all Competing interests: Q.Z., S.S., D.A., and F.G. are currently
We wish to thank M. De Palma (EPFL), J. Schlessinger (Yale administered by the CHUV Foundation, Lausanne. Author employees of Opna Bio SA (Lausanne), which has licensed a patent
University), and G. Bollag (Opna Bio SA) for insightful discussions; contributions: Q.Z. contributed to the conception, experimental application from EPFL describing FMRP as a cancer cellÐintrinsic
M. De Palma (EPFL) for providing the OT1 mice and B16-OVA cells; design, and implementation throughout; prepared figures; immunosuppressor and potential therapeutic target, in which
G. Ciriello (University of Lausanne) for constructive comments and edited the manuscript. S.S. performed bioinformatic analyses Q.Z. and D.H. are coinventors. D.H. is a scientific founder of Opna
and advice on the bioinformatic analyses; Y. Mohammadzadeh and of RNA-seq and scRNA-seq datasets, developed FMRP activity Bio, and D.H., Q.Z., S.S., D.A., and F.G. have stock or stock
B. Torchia (De Palma laboratory, EPFL) for helpful suggestions signatures and their use in human association studies, prepared options in the company. EPFL has filed a second patent application
on the immune-related assays and exosome isolation, respectively; figures, and edited the manuscript. A.C. assessed and functionally covering the prognostic FMRP cancer network signatures described
the patients, clinicians, and hospital staff members who validated adaptive and innate immune cell populations and herein, in which S.S. and D.H. are coinventors. No other authors have
contributed to the construction of the human tumor TMAs in Lund their regulation, performed the macrophage reprogramming competing interests. Data and materials availability: Materials
(TNBC, SCAN-B) and Bern (all other TMAs) as well as the staff at experiments, and prepared the figures and illustrations. N.F. are available, subject to material transfer agreement (MTA) requests
the central SCAN-B laboratory at Division of Oncology, Lund contributed to bioinformatic analyses of RNA-seq and scRNA-seq submitted to D.H. All data are available in the manuscript or the
University, the Swedish National Quality Register for Breast Cancer datasets and the analyzing of TCR data. J.G. performed the mixing supplementary materials. Bulk RNA-seq and scRNA-seq, TCR
(NKBC), Regional Cancer Center South, and the South Swedish experiment shown in fig. S6. S.Y. contributed to the Myc KD sequencing, and whole-exome sequencing data in this study are
Breast Cancer Group (SSBCG) for providing the SCAN-B TMA experiment shown in fig. S3, I to K. P.L. queried the Encode ChIP-seq deposited on Zenodo (85). License information: Copyright © 2022
and associated transcriptome dataset, and in particular S. Lehn database for Fmr1 as a target gene for MYC and MAX binding in the authors, some rights reserved; exclusive licensee American
and P. Bolivar for liaison; L. Noti (University of Bern) for assisting cancer cell lines versus normal cell lines shown in fig. S3, D and Association for the Advancement of Science. No claim to original US
in the analysis of human tumor TMAs; A. Ayyanan and C. Brisken E. L.C. analyzed human TMAs stained to detect FMRP expression government works. [Link]
(EPFL) for providing normal human breast tissue samples; shown in Fig. 1 and fig. S2. J.A.G. immunostained multiple human journal-article-reuse
A. Bowler (Radtke laboratory, EPFL) for providing mouse colon tumor TMAs and tissue sections for FMRP and participated in
adenocarcinoma samples; L. Tang (EPFL) for support of Y.-Q.X.; their analysis for Fig. 1. E.K. generated the PDAC TMA and contributed SUPPLEMENTARY MATERIALS
M.-W. Peng, M. Gaveta, and E. Drori (EPFL) for technical support; to its initial analysis. I.Z. supervised the analysis of multiple human
[Link]/doi/10.1126/science.abl7207
all the members of the Flow Cytometry Core (especially V. Glutz tumor TMAs and tissue sections for FMRP expression shown in
Figs. S1 to S13
for helping set up the FACS panels), the Bioimaging and Optics Fig. 1 and fig. S2. D.A. and F.G. contributed to the data shown in
Tables S1 to S7
Core Facility, the Gene Expression Core Facility, the Histology fig. S3, I to K; fig. S9H; and fig. S10, F and H. P.Z. participated in initial
References (86, 87)
Facility, and the Animal Care Facility (School of Life Sciences, EPFL) bioinformatic analysis of transcriptome datasets. Y.-Q.X. performed
MDAR Reproducibility Checklist
for respective technology support; and all the members of the the CTL assay involving B6-OVA cancer cells and OT1 T cells. J.A.R.-C.
Hanahan laboratory for discussions. Funding: The research was and M.R. produced and provided the unpublished prostate cancer
supported by grants from the Swiss National Science Foundation, TMA analyzed in Fig. 1 and fig. S2. M.T. participated in FACS profiling
by the Ludwig Institute for Cancer Research (New York), by a grant of immune cell populations. K.H. contributed to the scRNA-seq. D.H. Submitted 4 August 2021; resubmitted 20 April 2022
from the Biltema Foundation (Lund) administered by the ISREC contributed to the conception, experimental design, project Accepted 17 October 2022
Foundation (Lausanne), and by grants from the Cancera and supervision, and the writing and preparation of the manuscript. 10.1126/science.abl7207

Zeng et al., Science 378, eabl7207 (2022) 18 November 2022 23 of 23

Your Legacy to Science
A N E STAT E G I F T TO T H E
A M E R I CA N A S S O C I AT I O N FO R T H E A DVA N C E M E N T O F S C I E N C E

Since 1848, our founding year, the American Association for the “As a teacher and instructor, I bear responsibility for the

Advancement of Science (AAAS) has been deeply committed to younger generations. If you have extra resources, concentrate
them on organizations, like AAAS, that are doing work for all.”
advancing science, engineering and innovation around the world
—Prof. Elisabeth Ervin-Blankenheim, 1848 Society member
for the benefit of all people.

By making AAAS a beneficiary of your will, trust, retirement plan or

If you intend to include AAAS in your estate plans, provide this
life insurance policy, you become a member of our 1848 Society, information to your lawyer or financial adviser:

joining Thomas Edison, Alexander Graham Bell and the many

Legal Name: American Association for the Advancement of Science
distinguished individuals whose vision led to the creation of AAAS Federal Tax ID Number: 53-0196568
and our world-renowned journal, Science, so many years ago. Address: 1200 New York Avenue, NW, Washington, DC 20005

Unlike many of its peers, Science is not for-profit. Your estate gift If you would like more information on making an estate gift to
AAAS, cut out and return the form below or send an
would provide long-term financial stability and durable annual
email to philanthropy@[Link]. Additional
income that will support operations and competitive innovation
details are also available online at
for years to come. This support is vital. [Link]/1848Society.

cut here

Yes, I would like more information about joining the AAAS 1848 Society.
PLEASE CONTACT ME AT:

Name:

Address:

City: State: Zip code: Country:

Email: Phone:

RETURN THIS FORM TO:

AAAS Office of Philanthropy and Strategic Partnerships • 1200 New York Avenue, NW • Washington, DC 20005 USA
RESEAR CH

◥ retained >80% of its initial PCE after 340 hours

RESEARCH ARTICLES at maximum power point (MPP) tracking. It
indicates that this model is applicable to various
SOLAR CELLS perovskite formulations such as mixed halide
perovskites, which demonstrates the generality
Initializing film homogeneity to retard phase of the Schelling model for halide perovskites.

segregation for stable perovskite solar cells Cation aggregation and phase transition
The a-phase FAxCs1–xPbI3 (FACs) perovskites
Yang Bai1,2†, Zijian Huang3†, Xiao Zhang1†, Jiuzhou Lu1†, Xiuxiu Niu1†, Ziwen He4, Cheng Zhu1, have an optimal bandgap and high thermal
Mengqi Xiao1, Qizhen Song1, Xueyuan Wei1, Chenyue Wang1, Zhenhua Cui1, Jing Dou1, Yihua Chen1, stability from the resultant crystal structure,
Fengtao Pei1, Huachao Zai3, Wei Wang4, Tinglu Song1, Pengfei An5, Jing Zhang5, Juncai Dong5, with an ideal Goldschmidt tolerance factor
Yiming Li6, Jiangjian Shi6, Haibo Jin1, Pengwan Chen7, Yuchao Sun8, Yujing Li1, Haining Chen9, (11, 21, 28–31). However, reported devices based
Zhongming Wei10, Huanping Zhou3, Qi Chen1,2* on FACs perovskite absorbers have had limited
lifetimes, resulting from material degradation
The mixtures of cations and anions used in hybrid halide perovskites for high-performance that is largely the result of phase separation
solar cells often undergo element and phase segregation, which limits device lifetime. We (32). In Fig. 1, A, C, and D, we show the two-
adapted SchellingÕs model of segregation to study individual cation migration and found that dimensional (2D) photoluminescence (PL)
the initial film inhomogeneity accelerates materials degradation. We fabricated perovskite films mapping by wavelength, intensity, and full
(FA 1ÐxCsx PbI 3; where FA is formamidinium) through the addition of selenophene, which led to width at half-maximum (FWHM), respectively,
homogeneous cation distribution that retarded cation aggregation during materials processing of the FA0.9Cs0.1PbI3 perovskite film. The emis-
and device operation. The resultant devices achieved enhanced efficiency and retained >91% sion spectra display peak variability between
of their initial efficiency after 3190 hours at the maximum power point under 1 sun illumination. the selected regions. Moreover, although the
We also observe prolonged operational lifetime in devices with initially homogeneous emission of region 1 was concentrated at
FACsPb(Br 0.13I0.87 )3 absorbers. 795 nm (attributed to FA0.9Cs0.1PbI3) (12), both
regions 2 and 3 exhibited splitting of the emis-

W
sion peak (Fig. 1B). This feature was indicative
ith an endless compositional space structural stability and extra functionality to of phase segregation within the perovskite film.
that is almost continuously tailorable strengthen chemical stability in the resultant We investigated the segregation evolution
in the full region, halide perovskite absorber thin films (9–11). However, mixed of films aged under continuous light illumi-
materials present tunable electronic perovskite absorbers often undergo element nation for 20 min at 80°C at the same location in
and optical properties (1–5). Although and phase segregation that can decrease de- ambient atmosphere. As shown in the mapping
perovskite solar cells can have high power con- vice efficiency and lifetime (12, 13). spectra (Fig. 1E), the phase segregation of the
version efficiency (PCE), their limited device Most studies of phase segregation of mixed film intensified with increasing areas of both
lifetime remains a substantial challenge to perovskites focus on film aging to understand the yellow domains (which represent the
commercialization (6–8). The use of mixed cation and anion migration, the formation and longer-wavelength, FA-rich composition) and
cations and anions has led to a more appropri- development of nanoscale clusters (14–16), red domains (which represent the shorter-
ate tolerance factor of the crystal for improved thermodynamic driving forces (17, 18), and wavelength, Cs-rich composition), which indi-
their impacts on film properties and device cated that the local Cs/FA ratio deviated from
performance (12, 19–23). Effective strategies that observed in the pristine films. Region 1
1
Beijing Key Laboratory of Construction Tailorable Advanced such as relaxing residual strain and incorpo- showed peak splitting at 799 and 789 nm,
Functional Materials and Green Applications, MIIT Key rating low-dimension perovskites can retard with increased peak intensity at longer wave-
Laboratory for Low-dimensional Quantum Structure and
phase segregation by suppressing ion migra- lengths. By contrast, for regions 2 and 3, the
Devices, Experimental Center of Advanced Materials, School
of Materials Science and Engineering, Beijing Institute of tion (5, 24, 25). However, relevant metrics are peak intensity increase was observed at shorter
Technology, Beijing 100081, P. R. China. 2Advanced needed to investigate the atomistic aggrega- wavelengths (783 and 781 nm, respectively).
Research Institute of Multidisciplinary Science, Beijing tion of individual ions to correlate their col- Moreover, the FWHM increased notably (Fig.
Institute of Technology, Beijing 100081, P. R. China.
3
Department of Materials Science and Engineering, College lective behavior that led to film degradation. 1H), which indicates the broadening of the
of Engineering, Peking University, Beijing 100871, P. R. We adapted a physical analog of Schelling’s PL emission peak after aging. After compar-
China. 4Center for Research on Intelligent Perception and model, which was used to illustrate agent (such ison with the pure FAPbI3 film, for which no
Computing, National Laboratory of Pattern Recognition,
Institute of Automation, Chinese Academy of Sciences, as isolated particles) segregation caused by obvious change on the PL mapping was ob-
Beijing 100190, P. R. China. 5Beijing Synchrotron Radiation even low levels of individual preference in the served after aging (fig. S1), we hypothesized
Facility, Institute of High Energy Physics, Chinese Academy context of social economics (26) and physics that Cs+ cations in region 1 migrated to re-
of Sciences, Beijing 100049, P. R. China. 6Key Laboratory for
Renewable Energy, Beijing Key Laboratory for New Energy
(27). It bridges the microscale analysis of cat- gions 2 and 3 in the mixed-cations composition
Materials and Devices, Institute of Physics, Chinese Academy ion aggregation in perovskites and macro- of the perovskite with a single halide. This
of Sciences, Beijing 100190, P. R. China. 7State Key scale observations of their phase separation hypothesis was confirmed through time-of-
Laboratory of Explosion Science and Technology, Beijing
Institute of Technology, Beijing 100081, P. R. China. 8Auner
and film degradation. In the simulation, the flight secondary-ion mass spectrometry (TOF-
Technology Co., Ltd., Beijing 100084, P. R. China. 9School of initial film homogeneity affected cation ag- SIMS) mapping (fig. S2).
Materials Science and Engineering, Beihang University, gregation duration and thus device lifetime. Upon extending the aging duration to
Beijing 100191, P. R. China. 10State Key Laboratory of
On the basis of these findings, we grew ho- 1000 hours (supplementary materials), we ob-
Superlattices and Microstructures, Institute of
Semiconductors, Chinese Academy of Sciences and Center mogeneous a-phase FA1–xCsxPbI3 (FA, for- served crystal phase transformation in the
of Materials Science and Optoelectronics Engineering, mamidinium) that resulted in a solar cell with film. In addition to the cubic perovskite peak
University of Chinese Academy of Sciences, Beijing 100083, enhanced efficiency and prolonged operational (14.0°), two extra diffraction peaks at 9.7° and
P. R. China.
*Corresponding author. Email: qic@[Link] lifetime. Furthermore, for the wide-bandgap 11.6° were dominant in the x-ray diffraction
These authors contributed equally to this work. FACsPb(Br0.13I0.87)3, the corresponding device (XRD) patterns (figs. S3 and S4), which indicates

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 747

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

Fig. 1. Degradation mechanisms for binary FACs perovskites. (A to H) 2D PL FWHM mapping spectra of (A) and (E). a.u., arbitrary units. (I to K) The (I) Cs
mapping spectra (wavelength) of perovskite film before (A) and after (E) and (J) FA intensity mapping for the degraded film measured with TOF-SIMS and
degradation. [(B) and (F)] PL spectra of selected regions from (A) and (E), (K) the overlay mapping spectra. Scale bar, 10 mm. (L to N) 3D topographic of Cs
respectively. [(C), (D), (G), and (H)] The corresponding 2D PL intensity and and FA for the degraded films measured by TOF-SIMS.

the formation of nonphotoactive d-CsPbI3 regation (Fig. 1J). The separated Cs/FA do- gregated in a manner similar as that at the
and d-FAPbI3 phases (18, 33, 34). To cor- mains were also spatially complementary as surface (fig. S5).
relate the composition of each region to its expected (Fig. 1K), given the binary cation
crystal structure, we conducted 2D TOF- nature of the perovskite system. Moreover, Cation segregation dynamics by the
SIMS mapping on the fully degraded film. depth-dependent TOF-SIMS measurement Schelling model
The Cs distribution (Fig. 1I) showed domains (Fig. 1, L to N) showed that vertical direc- On the basis of the XRD, PL mapping, and
at the scale of several to several tens of micro- tion distribution was relatively uniform, so TOF-SIMS results, we argue that the degra-
meters, which occurred in parallel to FA seg- cations at different depths of the bulk ag- dation of the binary system follows the route

748 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

Fig. 2. Schelling model simulations of cation segregation dynamics. (A) Energy distribution (case I, ISP = 27%). (C) An initially less segregated distribution
change of the degradation processes from a uniform distribution to phase (case II, ISP = 2%). (D) A 1D and 3D (or 2D and 3D) fused system (case III,
segregation to d-phases, by means of DFT. (B to D) Simulations of segregation ISP = 2%). (E and F) U index evolution of red particles (indicates Cs) derived
process on the basis of the Schelling model. (B) An initially segregated from (E) cases I to III and (F) cases with different moving tendencies.

from cation aggregation to phase transition energy change of −9.427 kJ mol−1 acted as the to a 2D case (35, 36) with a 100 by 100 lattice to
(Fig. 2A). Local aggregation of cations initiates energy sink corresponding for the irreversible accommodate binary cations of Cs (Fig. 1, red
phase transition that leads to the complete pro- segregation. Our observations showed that particles) and FA (Fig. 1, blue particles) that
cess of phase segregation. The thermodynamic cation aggregation occurred before phase tran- could randomly swap positions with neighbors.
driving force for the phase segregation was sition, which provides an opportunity to inter- In the simulation for any step, each single
investigated by means of density functional theory vene before d-phase formation occurs (20, 21). cation has a probability to exchange its posi-
(DFT) calculations (fig. S6). DFT results identi- The Schelling model of segregation can be tion with neighbors. We defined this proba-
fied that the FA0.89Cs0.11PbI3 perovskite exhib- applied to isolated particles (such as molecules bility as moving tendency that is the product
ited an energy change of only −0.133 kJ mol−1 or ions) on the basis of a few simple physical of the thermodynamic parameter of the phase
from the homogeneous distribution to cation assumptions (see supplementary text S1) (27). transition (hT) and kinetic parameter of cation
aggregation (Fig. 2A, stage 1), which is possibly We adapted the Schelling model to study the aggregation (hK ). DFT results revealed the
due to the reduction in interfaces between the cation dynamics in perovskite films by means phase formation energy to identify hT in the
FA and Cs domains. However, in the subse- of a Monto Carlo algorithm (fig. S7). Given that Schelling model. Particularly, hT values were
quent process, the segregated domains expe- cations at different depths of the film aggre- assigned to be 1 for all spontaneous processes
rienced phase transition to form d-FAPbI3 and gated in the same manner as at the surface with a negative value for energy change. The
d-CsPbI3 phases (Fig. 2A, stage 2), in which the (Fig. 1, I to N), we simplified the simulation kinetic parameter hK was determined by the

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 749

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

Fig. 3. Film homogeneity of mixed cation perovskite and the impacts. with a relative humidity of 40% and the corresponding statistics diagrams.
(A and B) 2D TOF-SIMS mapping of Cs+ for (A) the first step of PbI2/CsI film (G) 2D TOF-SIMS mappings of perovskite films (Cs) after continuous
fabricated by reference and selenophene-modified precursor solutions and thermal treatment in an N2-filled glovebox at 85°C for 144 hours with the
(B) perovskite films fabricated by reference and modified precursor solutions, corresponding statistics diagrams. (H) Time-dependent PCE measurements
as well as the corresponding intensity distribution. (C) Colloid size distribution under continuous light illumination (LED source with 100 mW/cm2) held at the
in the perovskite precursor solutions obtained from DLS. (D) J-V curves of open circuit in an N2 atmosphere. (I and J) MPP tracking of devices under
the champion devices from homogeneous film with an active area of 1 cm2. 100 mW/cm2 LED source in an N2-filled glovebox without encapsulation
(E and F) In situ 2D PL mapping spectra evolution of perovskite films (E) before at (I) 45° ± 5°C and (J) elevated temperatures of 65° ± 5°C and 85° ± 5°C.
and (F) after aging under 80°C thermal treatment in an ambient environment MPPT, MPP tracking.

750 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

Fig. 4. Film homogeneity of mixed halide perovskite and the impacts. with severe iodide segregation, under light illumination for 20 min in an
(A to D) 2D TOF-SIMS mapping of [(A) and (C)] I− in the mixed halide ambient environment with relative humidity of 60%. (I and J) Simulation
perovskite films (A) without and (C) with initial segregation and [(B) and of segregation process of perovskite films in (I) an initially segregated
(D)] perovskite films aged under continuous light illumination at 45° ± 5°C distribution (ISP = 28%) and (J) an initially less segregated distribution
in an N2-filled glovebox for 24 hours (B) without and (D) with severe initial (ISP = 3%). (K) U index evolution of red particles (Br) derived from simulation
iodide segregation. (E to H) In situ 2D PL mapping spectra (wavelength) cases. (L) MPP tracking under 100 mW/cm2 LED illumination in an N2-filled
evolution of perovskite films [(E) and (F)] without and [(G) and (H)] glovebox without encapsulation.

ion migration activation energy, depending on of interest, the heavier segregation can be re- stration (Fig. 2B, case I). After 200 steps, the
the cation properties and their local chemi- flected on the smaller U index (Fig. 2, B and E). U index decreased from 1.00 to 0.94. After
cal environment. In a particular scenario for The initial state of cation aggregation affected 500 and 1000 steps, the U index dropped to
simulation, we simply normalized the migra- the phase segregation trajectories. We inves- 0.77 and 0.59, respectively, and heavier aggre-
tion activation energy for different cations tigated phase segregation with different initial gation and enlarged domains were observed
from DFT or experimental results to obtain segregation percentages (ISPs), which stands (movie S1).
hK ranging from 0 to 1. To provide a quan- for the ratio of initially segregated Cs to the We compared the initially more homoge-
titative measure of the film homogeneity, we total amount of Cs. The values of other core neous film (case II, ISP of 2%) with the in-
defined unsegregated index (U index) as the parameters are summarized in tables S1 and homogeneous one (case I) through simulation.
ratio of presently unsegregated Cs ions to the S2. The cation aggregation with an arbitrary After 200 steps, the U index for case II only
initially unsegregated ones. Regarding any case ISP of 27% was simulated for clear demon- dropped to 0.98 (Fig. 2C, middle left), in which

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 751

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

the aggregation is negligible as compared with added, the CsI/PbI2 was converted to perovskite smaller clusters effectively diluted the aggre-
the initial state. After 500 and 1000 steps, the films when annealed at 150°C. The inhomoge- gation of Cs, which mitigated the heterogene-
U index remained at 0.95 and 0.88, respec- neous Cs+ distribution was still observed with ity in the resultant film.
tively (movie S2). With linear regression, we TOF-SIMS mapping (Fig. 3B) and the PL mea- A series of devices, with the configura-
obtained the U index decay rate of 0.012 per surements (fig. S14). tion of indium-doped tin oxide (ITO)/SnO2/
100 steps for case II, whereas that for case I For less inhomogeneity, we modulated the Cs 0.1 FA 0.9 PbI 3 /2,2′,7,7′-tetrakis[N,N-di(4-
is 0.041 per 100 steps. The initial homo- Pb-solvent interactions by introducing a donor methoxyphenyl)amino]-9,9′-spirobifluorene
geneous film clearly exhibited slower seg- molecule to reduce the CsI–PbI2 colloid size (Spiro-OMeTAD)/Au, were fabricated by tuning
regation (Fig. 2, B and C, and fig. S8), which (41). Common Lewis bases (such as dimethyl the initial film homogeneity (fig. S22), and
suggests a longer lifetime in the corresponding sulfoxide) lead to strong coordination with the current-voltage (J-V) curves of champion
devices. Pb and form crystalline intermediates (42). To devices are shown in fig. S23 with the key pa-
Increasing the activation energy of ion mi- achieve a moderate coordination with Pb and rameters summarized in table S4. On the basis
gration is proven to prolong device lifetime disturb those crystalline intermediates, two of the enhanced initial film homogeneity with
(10), which can be reflected as a lower hK in weak donors, thiophene and selenophene, were selenophene, the device performance improved
our model. Simulations for the group of films separately added to the PbI2/CsI precursor solu- from 20.9 to 23.4% (0.08 cm2, average PCE
with the same initial homogeneity (ISP of 27%) tions to grow films. The 2D TOF-SIMS mapping from 20.1 to 22.6%). We observed improved
but decreased hK (1, 0.8, 0.6, and 0.4) showed of both films revealed a more homogeneous photoelectronic properties in the resultant
reduced decay rates of 0.041, 0.022, 0.017, and distribution of Cs than that of the reference device with suppressed carrier recombina-
0.007 per 100 steps in their linear decay re- sample (Fig. 3A and fig. S15 for selenophene tion (fig. S24), lower trap density (fig. S25),
gions, respectively (Fig. 2F, fig. S9, and table S2). and thiophene, respectively), although the enhanced mobility (fig. S25), and better charge
The initial increase of U index could be ob- selenophene was more homogeneous. extraction (fig. S26). With further optimiza-
served in some cases, in which a large number The intensity distribution analysis of CsI– tion, the device with an active area of 1.00 cm2
of small segregation domains were dissolved PbI2 film (Fig. 3A) showed a reduced FWHM showed a certified PCE of 23.7% (reverse scan),
to form larger ones (supplementary materials). for the selenophene-modified film (0.04 versus with an open-circuit voltage (Voc) of 1.14 V, a
It reveals the determinant impact of the initial 0.08 for the reference), which indicates the short-circuit current density (Jsc) of 24.7 mA
state on the segregation dynamics, which were smaller intensity difference of Cs at each scan- per cm2, and a fill factor (FF) of 0.84 (Fig. 3D).
in good agreement with the experimental data ning point (~0.1 by 0.1 mm2). In addition, the The device was also subjected to continuous
reported elsewhere (23). x-ray photoelectron spectroscopy (XPS) map- illumination by holding at 1.03 V and achieved
We also simulated the aggregation dynam- ping (fig. S16) showed a narrow intensity a stabilized efficiency of 22.4% after 300 s in
ics of ternary cation perovskites (fig. S10) and variation in the modified film, which also ambient atmosphere without encapsulation
1D and 3D fused perovskites (Fig. 2, D and E, demonstrated an enhanced film homogeneity. (fig. S27).
and fig. S11). For the fused perovskites (case III), No phase structure difference was detected The absorber films with improved initial
the immobile large cations created barriers from the XRD patterns (fig. S17A). As expected, homogeneity (with selenophene) exhibited
between each sublattice and retarded cation the resultant perovskite film (selenophene enhanced stability as compared with the ref-
migration. According to previous observations modified) was more uniform than the reference erence film. We conducted in situ PL mapping
(25, 37), we reasonably assumed the hK of large (Fig. 3B), with the FWHM of the perovskite measurements to the films during aging under
cations (yellow) to be zero. These boundaries peak from XRD patterns at 14.0° decreasing thermal treatment. After 20 min of thermal
retarded phase segregation (movie S3), and from 0.19 to 0.14 (fig. S17B). No obvious phase aging in an ambient atmosphere at 80°C, the
the decay rate further declined to 0.006 per segregation was detected in the PL mapping reference sample exhibited two emissive re-
100 steps, which was just half of that of case II measurement for the modified film (fig. S14). gions, of which the ones at ~785 nm grew larger
and in agreement with reported experimental We studied the interaction between seleno- (Fig. 3, E and F). This change was most likely
data (37, 38). The stability enhancement in 1D phene and Pb in the precursor solution with caused by the formation of Cs-rich and Cs-poor
and 3D perovskites is generally attributed to several methods. Carbon nuclear magnetic domains, in which the local Cs-FA stoichiome-
the prevention of moisture by large hydro- resonance (13C-NMR) revealed that the chem- try deviated from that in the pristine films. By
phobic cations, but our findings suggest that ical shift of the a carbon in selenophene at contrast, the perovskite film with enhanced
retarded phase segregation is caused in part 131.95 parts per million (ppm) moved down- initial homogeneity exhibited a constant and
by partially immobilized cations. field to 132.05 ppm, which is indicative of uniform emission at 794 nm with retarded
chemical bonding between selenium (Se) and Cs aggregation. These results were confirmed
Initializing film homogeneity and PbI2 (fig. S19). Extended x-ray absorption fine by means of TOF-SIMS mapping (Fig. 3G) of
device performance structure (EXAFS) spectra of the Pb L3-edge films after aging for 144 hours. Scanning elec-
We modulated the Cs aggregation in FACs (fig. S20 and table S3) showed that the coor- tron microscopy (SEM) images also revealed
perovskite thin films with a two-step deposi- dination numbers of Pb in Pb-I before and that the initially segregated film underwent
tion process to introduce CsI when depositing after the introduction of selenophene were substantial morphology changes after the ther-
the PbI2 film in the first step (fig. S12) (39). The 4.2 ± 0.6 and 3.6 ± 0.5, respectively. The EXAFS mal treatment, whereas no obvious changes
Coulomb interaction between the Cs+ cation results are not conclusive because of its limited were observed for the selenophene-treated film
and the PbI42−-based complex led to the forma- sensitivity but are in line with the 13C-NMR (fig. S28).
tion of CsI–PbI2 colloids in solution (fig. S13) results, which is indicative of an interaction The homogeneous films also exhibited en-
and subsequent Cs-rich aggregates in the pre- between Pb and Se in the precursor (43). Dy- hanced light stability. After aging under contin-
cursor film after thermal annealing at 70°C namic light scattering (DLS) showed that the uous illumination for 144 hours in an N2-filled
(40). The top-view mapping of Cs+ on the PbI2/ mean colloidal size decreased from 622 nm in glovebox, notable morphology changes occurred
CsI films that were characterized by TOF-SIMS the reference to 552 nm (Fig. 3C). We argue in the reference at the grain boundaries as
(Fig. 3A) revealed the inhomogeneous distri- that selenophene interacted with Pb to disturb shown in SEM images (fig. S29, A and B), and
bution of Cs-rich domains on the microme- the formation of large colloid clusters in the the d-FAPbI3 phase was identified by means
ter scale. After the organic components were precursor. As compared with large clusters, of XRD (fig. S29, C and D). By contrast, no

752 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

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

obvious morphology change occurred on the distribution (Fig. 4, A and B). However, for the 7. E. H. Jung et al., Nature 567, 511–515 (2019).
homogeneous films, and we observed a re- film with heavier initial segregation, the larger 8. H. Min et al., Nature 598, 444–450 (2021).
9. M. Abdi-Jalebi et al., Nature 555, 497–501 (2018).
tarded formation of d-FAPbI3. The 2D PL map- I−-deficient regions (~5 mm) were identified on 10. N. X. Li et al., Nat. Energy 4, 408–415 (2019).
ping measurements show that the PL peak of the resultant film (Fig. 4, C and D). 11. S. H. Turren-Cruz, A. Hagfeldt, M. Saliba, Science 362,
reference film exhibited blue shift from 790 In addition, we conducted in situ PL map- 449–453 (2018).
12. N. X. Li et al., Joule 4, 1743–1758 (2020).
to 785 nm with the emergence of nonpho- ping as films were being aged under light illu- 13. J. P. Correa-Baena et al., Science 363, 627–631 (2019).
toactive regions, possibly owing to the for- mination. Similar to cation segregation, the film 14. T. A. S. Doherty et al., Nature 580, 360–366 (2020).
mation of d-FAPbI3. However, the film with with initially homogeneous halide distribution 15. S. Macpherson et al., Nature 607, 294–300 (2022).
16. K. Frohna et al., Nat. Nanotechnol. 17, 190–196 (2022).
improved initial homogeneity showed only showed better stability under continuous illu- 17. J. Barrier et al., Energy Environ. Sci. 14, 6394–6405 (2021).
slight changes of the peak position and dis- mination (Fig. 4, E and F). By contrast, the film 18. L. T. Schelhas et al., Energy Environ. Sci. 12, 1341–1348
tribution, which illustrates the enhanced light with heavier initial halide segregation showed (2019).
19. D. B. Straus, S. Guo, A. M. M. Abeykoon, R. J. Cava, Adv. Mater.
stability (fig. S30). a marked change on the PL mapping spectra
32, e2001069 (2020).
The DFT simulations suggested that the (Fig. 4, G and H, and fig. S38). We observed 20. S. Kim, T. Eom, Y. S. Ha, K. H. Hong, H. Kim, Chem. Mater. 32,
remaining selenophene can benefit the sup- that the width of longer-wavelength regions 4265–4272 (2020).
pressing of Cs migration with increased acti- (Fig. 4, G and H, yellow) was increased from 21. C. Y. Yi et al., Energy Environ. Sci. 9, 656–662 (2016).
22. C. A. R. Perini, T. A. S. Doherty, S. D. Stranks,
vation energy (fig. S31). XPS data showed that ~3 to 6 mm. Moreover, the emission peak was J.-P. Correa-Baena, R. L. Z. Hoye, Joule 5, 1024–1030 (2021).
after being stored in the ambient environment clearly redshifted from 739 to 751 nm (Fig. 4, 23. L. E. Mundt et al., ACS Energy Lett. 7, 471–480 (2022).
for 12 hours, the perovskite films still exhibited G and H, circled region), which indicates the 24. Y. H. Lin et al., Science 369, 96–102 (2020).
25. Y. Lin et al., ACS Energy Lett. 2, 1571–1572 (2017).
the symmetric peaks of Pb (II) at 138.5 and fast halide segregation. The Schelling model 26. T. C. Schelling, J. Math. Sociol. 1, 143–186 (1971).
143.4 eV that are attributed to Pb 4f7/2 and simulation results are in good agreement with 27. D. Vinkovic, A. Kirman, Proc. Natl. Acad. Sci. U.S.A. 103,
Pb 4f5/2, respectively. It indicates that the the experimental observation. As shown in 19261–19265 (2006).
28. M. Saliba et al., Science 354, 206–209 (2016).
film surface was protected from oxidation, Fig. 4, I to K, the segregated case (ISP = 28%), 29. Z. Li et al., Chem. Mater. 28, 284–292 (2016).
which may have originated from the coordi- presented a decay rate of 0.040 per 100 steps, 30. M. M. Hao et al., Nat. Energy 5, 79–88 (2020).
nation of residual selenophene (fig. S32). How- which is more than four times higher than 31. S. Kawachi et al., J. Phys. Chem. Lett. 10, 6967–6972
(2019).
ever, the remaining amount is small according that of the initially homogeneous film of 0.008 32. K. Ho, M. Wei, E. H. Sargent, G. C. Walker, ACS Energy Lett. 6,
to the TOF-SIMS measurement (fig. S33), so the per 100 steps (ISP = 3%). 934–940 (2021).
enhanced film stability could be mainly attrib- We further fabricated devices based on these 33. Y. Liu et al., Angew. Chem. Int. Ed. 59, 15688–15694
(2020).
uted to the increased initial film homogeneity. two absorbers, which showed PCEs of 18.78%
34. S. S. Xiang et al., ACS Energy Lett. 3, 1824–1831 (2018).
Selenophene-treated perovskite devices were (homogeneous films) and 17.93% (reference), 35. H. Pan et al., Science 374, 100–104 (2021).
aged under various conditions according to respectively, as shown in fig. S39. The mixed 36. W. Mao et al., Nat. Mater. 20, 55–61 (2021).
the modified International Summit on Organic halide device with initially homogeneous halide 37. M. Parashar, R. Singh, K. Yoo, J.-J. Lee, ACS Appl. Energy
Mater. 4, 2751–2760 (2021).
Photovoltaic Stability (ISOS) stability protocols distribution achieved an impressive opera- 38. Z. Wang et al., Nat. Energy 2, 17135 (2017).
(44) and presented enhanced stability than tional stability during the MPP test at 45° ± 39. J. H. Im, I. H. Jang, N. Pellet, M. Grätzel, N. G. Park, Nat.
that of reference devices. The selenophene- 5°C in N2, which retained 80% of its initial PCE Nanotechnol. 9, 927–932 (2014).
40. M. Qin et al., Adv. Mater. 32, e2004630 (2020).
treated perovskite device showed good sta- after 340 hours (168 hours for the reference 41. J. C. Hamill Jr., J. Schwartz, Y.-L. Loo, ACS Energy Lett. 3,
bility without PCE degradation more than device). To date, this stability result is among 92–97 (2018).
2000 hours after being stored in the dark in an the best device operational lifetimes (T80) based 42. J.-W. Lee, H.-S. Kim, N.-G. Park, Acc. Chem. Res. 49, 311–319
(2016).
N2 atmosphere (fig. S34). The devices that on mixed halide perovskites (46). In addition,
43. A. Sharenko, C. Mackeen, L. Jewell, F. Bridges, M. F. Toney,
were aged at 85°C in N2 in the dark (fig. S35) re- at an elevated temperature of 85° ± 5°C, the Chem. Mater. 29, 1315–1320 (2017).
tained ~80% of their initial PCE after 1750 hours. mixed halide perovskite device with improved 44. M. V. Khenkin et al., Nat. Energy 5, 35–49 (2020).
When aged over 1-sun continuous illumination homogeneity achieved a T80 lifetime of 80 hours 45. N. Li et al., Science 373, 561–567 (2021).
46. A. Al-Ashouri et al., Science 370, 1300–1309 (2020).
[light-emitting diode (LED) source, 100 mW/ (fig. S40).
47. Z. He, shaguopohuaizhe/Schelling_simulation: fix a small bug
cm2] at open circuit, the devices could retain We have demonstrated that the Schelling of 1.0.0. Zenodo (2022); doi:10.5281/zenodo.7187438.
~80% of their initial PCE after 960 hours (Fig. model is a powerful tool to bridge theoretical
3H and fig. S36). For the MPP test at 45° ± 5°C in analysis of perovskites at the atomic scale and AC KNOWLED GME NTS

an N2 atmosphere, the device retained 91% of macroscale observations of their phase sepa- We thank Y. Zhang, N. Li, K. Li, Y. Wu, and H. Liu (Peking University)
and N. Yang, S. Ma, G. Yuan, and Y. Ma (Beijing Institute of
its highest PCE after 3190 hours. In addition, ration and film degradation. From simula- Technology) for assistance with device fabrication; R. Si (Sun
the MPP test was performed at elevated tem- tion and experiment results, we found that Yat-sen University) and J. Liao (Institute of High Energy Physics)
peratures in the N2 atmosphere, and the de- the initial film homogeneity, in terms of ele- for the assistance in EXAFs analysis; Z. Liu (Northwestern
Polytechnical University) for the discussion on Schelling model; and
vices retained 91 and 89% of their highest PCE mental distribution, has shown substantial Y. Wu (Auner Technology Co.) for assistance with device
after 500 hours at 65° ± 5°C and 85° ± 5°C, influence on film and device stability. Bene- measurements. Funding: This work was supported by the National
respectively (Fig. 3J and fig. S37). fiting from homogeneous films by tailoring Natural Science Foundation of China (U21A20172, 21975028,
52172182, 22011540377, 62125404, U1932201, and 11705225) and
For mixed-halide perovskites, the main de- precursor chemistry with selenophene, we de- Beijing Municipal Natural Science Foundation (JQ19008). Author
vice issue has been the poor light stability. We veloped high-performance devices based on contribution: Y.B. and Q.C. conceived the idea. X.Z., J.L., and Y.B.
fabricated FACsPb(Br0.13I0.87)3 of a bandgap of mixed perovskites that achieved a long-term prepared the samples and carried out XPS, NMR, and optical
spectroscopy characterizations. Y.B. and T.S. carried out the
1.66 eV with different initial halide segrega- stability during the MPP test even at elevated TOF-SIMS characterization. J.L. and H.C. carried out the DLS
tion. We compared 2D TOF-SIMS mapping on temperatures. measurements. Y.B., [Link], Q.C., and W.W. worked on the
the absorber films, which were subjected to modification of the Schelling Model code. Yujing Li, P.A., J.Z.,
and J. Dong carried out the EXAFS measurement and data
continuous light illumination for 24 hours at RE FERENCES AND NOTES analysis. J.L., X.W., C.W., Z.C., J. Dou, N.X., and Y.C. carried out the
45° ± 5°C in an N2-filled glovebox. The film 1. L. Gu et al., Nature 581, 278–282 (2020). other materials characterizations. M.X., Q.S., and Z.W. carried out
without severe initial segregation that was 2. K. Lin et al., Nature 562, 245–248 (2018). the DFT computation. Y.B., Yiming Li, and J.S. carried out the
fabricated under the protocol that is reported 3. Z. Xiao et al., Nat. Mater. 14, 193–198 (2015). characterization of TPC and TPV measurement and data analysis.
4. M. Jeong et al., Science 369, 1615–1620 (2020). Y.B., [Link], X.Z., J.L., X.N., C.Z., [Link]., F.P., and Y.S. supported
elsewhere (45) (supplementary materials, mate- 5. G. Kim et al., Science 370, 108–112 (2020). the fabrication and characterization of perovskite solar cells. H.J.
rials and methods) retained a homogeneous I− 6. S. Bai et al., Nature 571, 245–250 (2019). and P.C. assisted in the manuscript writing. Y.B., [Link]., X.Z., J.L.,

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 753

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

X.N., [Link]., and Q.C. carried out data analysis and wrote the licensee American Association for the Advancement of Science. No Figs. S1 to S40
manuscript. All authors discussed the results and commented on claim to original US government works. [Link] Tables S1 to S4
the manuscript. Y.B., [Link]., X.Z., J.L., and X.N. contributed equally about/science-licenses-journal-article-reuse References (48, 49)
to this work. Competing interests: The authors declare no Movies S1 to S3
competing interests. Data and materials availability: All data are
available in the main text or the supplementary materials. The SUPPLEMENTARY MATERIALS Submitted 17 November 2021; resubmitted 1 July 2022
Schelling Model code is available at (47). License information: [Link]/doi/10.1126/science.abn3148 Accepted 19 October 2022
Copyright © 2022 the authors, some rights reserved; exclusive Materials and Methods 10.1126/science.abn3148

HUMAN GENOMICS Mendelian randomization (12) and association

studies (13).
Cross-trait assortative mating is widespread and We set out to systematically assess the im-
pact of xAM on genetic correlation estimates,
inflates genetic correlation estimates first compiling a large atlas of cross-mate cor-
relations across a broad array of previously
Richard Border1,2,3*, Georgios Athanasiadis4,5, Alfonso Buil4,6, Andrew J. Schork4,6,7, Na Cai8, studied phenotypes using two large population-
Alexander I. Young9,10, Thomas Werge4,6,11, Jonathan Flint12, Kenneth S. Kendler13, based samples (n = 81,394; n = 746,566). We find
Sriram Sankararaman2,10,14, Andy W. Dahl15, Noah A. Zaitlen1,10,14 that these phenotypic cross-mate correlations
explain a major portion of empirical marker-
The observation of genetic correlations between disparate human traits has been interpreted as based genetic correlation estimates for the
evidence of widespread pleiotropy. Here, we introduce cross-trait assortative mating (xAM) as an same trait pairs (R2 = 74% across samples). We
alternative explanation. We observe that xAM affects many phenotypes and that phenotypic cross-mate next demonstrate that xAM biases genetic cor-
correlation estimates are strongly associated with genetic correlation estimates (R2 = 74%). We relation estimates and yields nontrivial esti-
demonstrate that existing xAM plausibly accounts for substantial fractions of genetic correlation mates even among traits with uncorrelated
estimates and that previously reported genetic correlation estimates between some pairs of psychiatric genetic effects. We use a simulation-based ap-
disorders are congruent with xAM alone. Finally, we provide evidence for a history of xAM at the proach to evaluate the extent to which empir-
genetic level using cross-trait even/odd chromosome polygenic score correlations. Together, our results ical levels of xAM alone might plausibly explain
demonstrate that previous reports have likely overestimated the true genetic similarity between genetic correlation estimates among previously
many phenotypes. studied traits, finding that, for many trait pairs,
substantial fractions of empirical genetic cor-

M
relation estimates are congruent with expecta-
ethods that use summary statistics notable: Many trait pairs, even those with tions for etiologically independent traits subject
from genome-wide association studies limited phenotypic similarity, display non- to xAM. At the same time, we observe that par-
(GWAS) to investigate genetic overlap trivial genetic correlations [for example, 0.209 ticular phenotype pairs, such as schizophrenia
across phenotypes have become a fun- (SE = 0.042) for attention-deficit hyperac- and bipolar disorders, evidence substantially
damental statistical tool across many tivity disorder (ADHD) and body mass in- larger genetic correlation estimates than can be
domains of human complex trait genetics dex (BMI) in (1)]. These findings have been plausibly attributed to xAM-induced artifact.
(1–5). The results of these analyses have been broadly interpreted as evidence for widespread Lastly, we utilize correlations between even
pleiotropy across the phenome (6–8) and, in versus odd chromosome-specific polygenic
the case of psychiatric disorders, have raised scores (PGS) to detect genetic signatures of
1
Department of Neurology, David Geffen School of Medicine, concerns about the suitability of the existing xAM, extending a previous approach (11). We
University of California, Los Angeles, Los Angeles, CA 90095, nosology given evidence for shared genetic find that cross-trait even/odd PGS correla-
USA. 2Department of Computer Science, University of
California, Los Angeles, Los Angeles, CA 90095, USA. bases (1, 9). tions mirror cross-mate phenotypic correlation
3
Department of Epidemiology, Harvard T.H. Chan School of Here, we consider an overlooked source of patterns and, through this association, explain
Public Health, Boston, MA 02115, USA. 4Institute of Biological potential bias in these findings: cross-trait substantial variation in empirical genetic cor-
Psychiatry, Mental Health Center Sct Hans, Copenhagen
University HospitalÐMental Health Services CPH, 2100 assortative mating (xAM), the phenomenon relation estimates.
Copenhagen, Denmark. 5Department of Evolutionary Biology, whereby mates display cross-correlations across
Ecology, and Environmental Sciences, University of Barcelona, distinct traits. There are several reasons to be Results
08028 Barcelona, Spain. 6Globe Institute, University of Genetic correlation estimates mirror cross-mate
Copenhagen, 1350 Copenhagen, Denmark. 7Neurogenomics concerned with this potential oversight: First,
Division, The Translational Genomics Research Institute, the single-trait linear mixed model, which ge- phenotypic correlations
Phoenix, AZ 85004, USA. 8Helmholtz Pioneer Campus, netic correlation estimators generalize, is mis- We begin by quantifying the extent to which
Helmholtz Zentrum München, 85764 Neuherberg, Germany.
9
Anderson School of Management, University of California, Los
specified under single-trait assortative mating empirical genetic correlation estimates align
Angeles, Los Angeles, CA 90095, USA. 10Department of (sAM) and overestimates single-nucleotide with cross-trait spousal correlations across a
Human Genetics, University of California, Los Angeles, Los polymorphism (SNP) heritability (10). Second, broad array of phenotypes: a set of 20 previous-
Angeles, CA 90095, USA. 11Department of Clinical Medicine,
University of Copenhagen, 1165 Copenhagen, Denmark.
sAM is widespread across multiple domains ly studied traits measured in the UK Biobank
12
Center for Neurobehavioral Genetics, University of California, for which substantial genetic correlations have (UKB) (14) and a collection of six psychiatric
Los Angeles, Los Angeles, CA 90095, USA. 13Department of been observed, including anthropometric, psy- disorder diagnoses ascertained from Danish
Psychiatry, Virginia Institute for Psychiatric and Behavioral chosocial, and disease traits (1, 7, 8). Third, civil registry data (15). We estimated cross-mate
Genetics, Virginia Commonwealth University, Richmond, VA
23298, USA. 14Department of Computational Medicine, David recent work has provided genetic-level evidence correlations for 40,697 spousal pairs within the
Geffen School of Medicine, University of California, Los for a history of sAM with respect to some of UKB sample and 373,283 mate pairs random-
Angeles, Los Angeles, CA 90095, USA. 15Section of Genetic these same phenotypes (11). Fourth, xAM is ly selected from the Danish population. For a
Medicine, Department of Medicine, University of Chicago,
Chicago, IL 60637, USA. known to generate spurious results for other pair of phenotypes Y and Z, there are three
*Corresponding author. Email: [Link]@[Link] marker-based inference procedures, including cross-mate correlation parameters: ryy and

754 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

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

Fig. 1. Cross-mate phenotypic correlation and genetic correlation estimates. Superdiagonal entries are previously reported LDSC correlation estimates (23).
(A) Correlations among previously studied UK Biobank (UKB) phenotypes. (C) Association between empirical cross-mate phenotypic correlation and genetic
Diagonal and subdiagonal heatmap entries correspond to cross-mate phenotype correlation estimates (meta-analytic R2 ≈ 74%). Error lines indicate 95%
correlation estimates derived from 40,697 putative spouse pairs in the UKB. confidence intervals, and the purple dashed line displays the line of best fit across
Superdiagonal entries correspond to empirical linkage disequilibrium score all points. All numbers have been rounded to two decimal places. The model
regression (LDSC) correlation estimates among unrelated European ancestry UKB for bone mineral density (BMD) and subjective happiness failed to converge and is
participants. (B) Cross-mate correlation and genetic correlation estimates for omitted. ADHD: attention-deficit hyperactivity disorder; ALC: alcohol use disorders;
psychiatric disorders. Diagonal and subdiagonal entries reflect cross-mate ANX: anxiety disorders; BIP: bipolar disorders; BMI: body mass index; HDL/LDL:
tetrachoric correlations among 373,283 spousal pairs sampled from the Danish high/low-density lipoprotein; IQ: intelligence quotient; MDD: major depressive
population, all of which were significantly greater than zero (maximum p = 1.69e-5). disorder; SCZ: schizophrenia.

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 755

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

Table 1. Notions of genetic similarity and their relationship to genetic correlation estimators. Under random mating, score correlations and effect
correlations are equal in expectation, imply the existence of pleiotropic loci, and are well captured by widely used genetic correlation estimators. Under xAM,
however, substantial score correlations can arise in the absence of effect correlation or even pleiotropy, and genetic correlation estimators overestimate both
rb and r‘.

Metric of genetic similarity Relation to shared etiology and xAM

Pleiotropy is present at a particular Reflects shared etiology. Substantial numbers of pleiotropic loci imply that overlapping genetic
locus when it influences both phenotypes. variants affect both traits, though their effects may not be consistent.
..............................................................................................................................................................................................................................................................................................................................................
Effect correlation (rb) refers to the Reflects shared etiology. rb > 0 implies that an overlapping set of variants (at pleiotropic loci) influence
correlation between standardized genetic effects. both traits with similar effects on average.
..............................................................................................................................................................................................................................................................................................................................................
Reflects shared etiology, or xAM-induced population structure, or both. Roughly equal to rb under
Score correlation (r‘) refers to the
random mating but larger than rb under xAM owing to long-range sign-consistent LD. r‘ > 0 does
correlation between true polygenic scores.
not necessarily imply biological similarity or even the existence of pleiotropic loci.
..............................................................................................................................................................................................................................................................................................................................................
Genetic correlation estimators Reflect shared etiology under random mating but produce estimates substantially
(^r b ), such as bivariate LD score regression, greater than both rb and r‘ under xAM, even when rb ¼ 0 or in the complete
are commonly interpreted as estimates absence of pleiotropy.
of the effect correlation.
..............................................................................................................................................................................................................................................................................................................................................

rzz , the correlations between mates on Y and genetic correlation estimates: Considering where the b vectors include all variants, causal
Z, respectively, and ryz , the cross-mate cross- only trait pairs with estimated genetic corre- or otherwise, and thus may contain elements
trait correlation; we generically denote these lations below 0.50 in magnitude yielded R2 = equal to zero. A value of rb > 0 implies both
quantities rmate and present these estimates 70.94%; further restricting to those below 0.30 the existence of pleiotropic loci and that such
in the diagonal and subdiagonal entries of in magnitude, yielded R2 = 67.88%. This sug- loci have similar effects on average, and we
Fig. 1, A and B. We also compiled linkage gests that the observed association does not term a pair of traits genetically orthogonal
disequilibrium score regression (LDSC) ge- merely reflect sAM on genetically homoge- when rb ¼ 0. Effect correlation is distinct from
netic correlation estimates, denoted ^ r b;LDSC , neous factors. the classical definition of genetic correlation as
for each pair of phenotypes, which we present the correlation between the heritable compo-
in the superdiagonal entries of Fig. 1, A and Defining genetic correlation nents of two traits (17), which we refer to as the
B. All pairwise estimates are provided in Having established that a large degree of the score correlation:
table S1. variance in genetic correlation estimates can
Cross-mate correlation structures were di- be predicted from phenotypic mating corre- r‘ ¼ corð‘y ; ‘z Þ
verse across the trait pairs that we examined lations, we now provide theoretical intuition
(fig. S1). Whereas cross-mate single-trait and as to why this might occur [see supplementary as it reflects the correlation between the
cross-mate cross-trait correlations were sim- text for further details (16)]. We start by de- true PGS.
ilar for some trait pairs (for example,^ryy = 0.26, fining three distinct notions of genetic sim- Within the standard linear mixed model
^rzz = 0.20, and ^ryz = 0.20 for BMI and hip cir- ilarity between phenotypes. These definitions framework, r‘ and rb are equivalent and hence
cumference), these quantities were of oppos- are summarized in Table 1. seldom discussed as separate quantities [though
ing signs for others (for example, ^ryy ¼ 0:33 We consider a pair of phenotypes Y, Z, with genetic correlation estimates are commonly
and ^rzz ¼ 0:22 versus ^ryz ¼ 0:09 for years heritable components ‘y , ‘z reflecting the interpreted as estimates of rb (4, 18, 19)]. How-
of education and regular smoking). In gen- additive effects of m standardized haploid ever, traits with uncorrelated effects can un-
eral, cross-mate correlation structures were variants X1 ; …; Xm with phenotype-specific intuitively have correlated PGS. Under xAM,
not consistent with sAM alone. When the cross- effect vectors by , bz . For simplicity, we assume all causal variants affecting trait Y become
mate correlation rzz for a secondary trait Z is that causal variants are initially unlinked and correlated with all causal variants affecting
fully mediated through sAM on Y, we expect that both phenotypes have unit variance under trait Z, and these correlations are direction-
that ^rzz ≈ ^r yy^syz2 ; this model fit the data poorly random mating (panmixis), such that the ally consistent with their respective effects
(fig. S2). panmictic heritabilities are h2y;pan ¼ b⊺y by and [see supplementary text (16)]. As we will
Estimates of ^ r b;LDSC were strongly associ- h2z;pan ¼ b⊺z bz . demonstrate in the following section, this
ated with rmate estimates across both samples Pleiotropy is present when a locus influ- results in nonzero score correlations in the
[Fig. 1C; meta-analytic R2 = 74.32%; 95% confi- ences two or more phenotypes. Thus, locus Xi direction of the cross-mate cross-trait pheno-
dence interval (CI): 67.02% 81.62% in a linear is pleiotropic with respect to Y and Z when typic correlation, even for genetically orthogo-
model; R2 = 76.69%; 95% CI: 73.94% 79.45% both by;i ≠ 0 and bz;i ≠ 0, though these effects nal traits.
in a Bayesian model accounting for hetero- might differ substantially in magnitude or
skedasticity and estimation error]. The regres- direction. By contrast, the correlation between The impact of xAM in simulations
sion slope did not significantly differ across effects, which we refer to as the effect corre- We ran a series of forward-time simulations
the UKB and psychiatric phenotypes in either lation rb , indexes the similarity of variant ef- using realistic genotype data to investigate the
model (for example, p = 0.16 for a sample-by- fects on two phenotypes: impact of xAM on multiple measures of ge-
rmate interaction term in the linear model). netic correlation. At each generation, individ-
The strength of this association largely per- uals (consisting of a set of genotypes together
sisted when excluding trait pairs with large rb ¼ cor by ; bz with two phenotypes) were matched to achieve

756 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

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

A B

C D

Fig. 2. Impact of xAM on genetic correlation estimates in forward-time further overestimate rb , and the magnitude of this bias increases over
simulations. Score correlation (r‘) and genetic correlation estimates (^r b) for two subsequent generations. (B) After three generations of xAM, ^r b estimates are
phenotypes with true effect correlation rb , panmictic heritabilities h2pan , and all upwardly biased for genetically distinct phenotypes. (C) The impact of three
cross-mate correlations set to rmate. (A) xAM increases the true score correlation generations of xAM increases with the cross-mate correlation. (D) The impact of
among genetically orthogonal phenotypes. HE, LDSC, and REML estimators all three generations of xAM increases with the panmictic heritabilities.

target cross-mate correlation parameters, after investigated the impact of xAM on GWAS Fig. 2A demonstrates the increase in the true
which we estimated genetic correlations (^ r b) effect estimates and GWAS-based methods for score correlation across multiple genera-
using LDSC [denoted ^ r b;LDSC (6)], Haseman- identifying pleiotropic SNPs (figs. S7 to S10), tions of xAM for a pair of traits with rb ¼ 0,
Elston regression [HE, denoted ^ r b;HE (20)], and genetic correlation estimates for binary phe- rmate ¼ 0:5, and h2pan ¼ 0:5. Across simulation
residual maximum likelihood [REML; ρ ^ b;REML notypes subject to misdiagnosis (fig. S11), replicates, the average score correlation was
(18)]. We also computed true score correlations partitioned genetic correlation estimates 0.11 after a single generation of xAM, which
(r‘ ), which is possible when the true genetic (fig. S12), and genetic covariance estimators increased to 0.24 after three generations.
effects are known. We performed sensitivity (fig. S13). Notably, this increase in score correlation
analyses to confirm that our results did not induced by xAM does not represent bias: The
depend on simulation parameters, including xAM induces nonzero score correlations population-level correlation between the her-
the number of causal variants (fig. S3), mate among genetically orthogonal traits itable components of the phenotypes truly does
selection algorithm (fig. S4), recombination We confirmed that xAM induces substantial increase under xAM. However, as we demon-
scheme (fig. S5), and whether causal vari- score correlations across a broad array of simu- strate next, genetic correlation estimators be-
ants with orthogonal genetic effects arose lation parameters. This is perhaps most pro- come misspecified under xAM and yield biased
on overlapping loci (fig. S6). We additionally nounced for traits with orthogonal effects: estimates. Still, even unbiased estimates of

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 757

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

score correlation can be driven by either shared However, spurious effect estimates will remain ferent from zero), the inverse variance weighted
biology, or xAM, or both, further complicating attenuated (figs. S7 to S9), implying that meth- average^g estimate was 0.25 (SE = 0.005). Figure
interpretation (Table 1). ods for identifying cross-trait heterogeneity in 3B presents the first 20 pairs in descending
GWAS estimates may have the potential to order of ^g , and Fig. 3C presents the raw pro-
Effect correlation estimates are differentiate trait-specific signal from xAM- jected and empirical effect correlation estimates
biased upward induced artifacts. across all 132 pairs (see table S3 and figs. S14
For genetically orthogonal traits, after a sin- and S15 for detailed results spanning five gen-
gle generation of xAM, the REML estimator xAM alone can plausibly explain erations of xAM). Finally, Fig. 3D displays aver-
yielded ^r b = 0.15 and the HE and LDSC es- substantial variance in empirical genetic age ^g values within and between qualitative
timators, which are closely related (21, 22), correlation estimates phenotypic domains.
both yielded average estimates of ^ r b = 0.21, We next sought to quantify the extent to which
all of which were greater than the true value of empirical ρ ^ β estimates for previously studied Expected effect correlation estimates
rb = 0.00. After three generations of xAM, this trait pairs could be explained by xAM alone, among psychiatric disorders in the
upward bias became more pronounced, with assuming genetic orthogonality. We proceeded absence of pleiotropy
REML and LDSC yielding estimates of ^ rb = with a simulation-based approach, which we We next estimated ρ ^ xAM for a collection of six
0.30 and ^ r b = 0.44, respectively. present for the UKB and psychiatric pheno- psychiatric disorders, using correlations esti-
types in turn. mated in spousal pairs randomly selected from
The quantities ρ‘ , and ρ ^ β are monotonically Within each sample, we identified mate pairs the Danish population (table S4). We then com-
related to ρβ , rmate , and h2pan and estimated phenotypic cross-mate correla- pared these projections to the LDSC genetic
Many trait pairs subject to xAM will truly tions. We then used these estimates, together correlation estimates reported by Grotzinger
have correlated genetic effects. Figure 2B with empirical heritability estimates, as inputs and colleagues (23).
illustrates the relationship between rb , r‘ , to a forward-time simulation where separate, Across all pairwise combinations of dis-
and ^ r b for two traits with h2pan = 0.5. Except- nonoverlapping collections of causal variants orders, we observed an average ratio of ^g =
ing the case of rb = 1.0 (genetically identical were assigned to each phenotype, such that rb= 0.29 (SE = 0.016; Fig. 4, A and B) after five
phenotypes), results remained consistent with 0.0. At each generation, we estimated the effect generations of xAM. Some trait pairs evidenced
the genetically orthogonal case: rb was lower correlation, rb, using method of moments. These considerably greater empirical genetic correla-
than r‘ , which was in turn lower than the up- projected effect correlation estimates, which we tion estimates than might be explained by xAM
wardly biased ^ r b estimates provided by REML, denote ^ r xAM , can be interpreted as expected alone (for anxiety disorders and major depres-
HE, and LDSC. For example, when rb = 0.25, LDSC genetic correlation estimates for genet- sion,^g = 0.21; 95% CI: 0.17 0.25), whereas for
LDSC yielded ^ r b = 0.62 after three generically orthogonal traits under xAM consistent other pairs this discrepancy was modest (for
ations of xAM. We note that whereas the true with empirical spousal correlations. alcohol use disorder and schizophrenia, ^g =
effect correlation varies in Fig. 2B, the cross- We next compared ^ r xAM to empirical LDSC 0.83; 95% CI: 0.59 1.24; see table S5 and fig.
mate correlations remain fixed, demonstrat- estimates derived in real data, which we de- S16 for complete results).
ing that the potential for xAM-induced bias is note ^r emp . To simplify discussion, we define
present even when cross-mate cross-trait cor- the ratio xAM exacerbates bias due to misdiagnosis
relations partially reflect shared genetic bases. After additional simulations demonstrated that
The impact of xAM on both r‘ and ^ r b was greater ^g ¼ ^
r xAM =^
r emp xAM further inflates genetic correlation es-
for traits under stronger xAM (Fig. 2C) and for timates in the context of diagnostic errors
traits with greater heritabilities (Fig. 2D). which measures the projected LDSC effect (Fig. S11), we extended our method for estimat-
correlation estimate due to xAM-induced arti- ing ^ r xAM to incorporate misdiagnosis. Re-
xAM biases annotation- and fact relative to the empirical LDSC effect cor- sults were heterogeneous across disorder pairs
locus-level analyses relation estimate for a given phenotype pair (fig. S17). For example, whereas moderate rates
Partitioned genetic correlation estimators evi- (Fig. 3A). of diagnostic errors (5%), together with three
denced similar biases as genome-wide esti- generations of xAM, yielded genetic correla-
mators under xAM, even when supplied with Expected effect correlation estimates for tion estimates for ADHD and major depres-
annotations directly relevant to bivariate UKB phenotypes in the absence sion on par with published estimates (^g = 0.97;
genetic architecture. Further, this bias was of pleiotropy 95% CI: 0.73 1.22), substantial diagnostic er-
greatest at regions relevant to only one of the We restricted our attention to 132 (of 190 pos- rors (15%) after five generations of xAM yielded
two phenotypes (fig. S12). sible) pairs of UKB phenotypes with nominally estimates well below previously published esti-
In association studies, GWAS effect estimates significant (p < 0.05) LDSC genetic correlation mates for bipolar disorders and major depres-
for SNPs causal for the focal trait were biased estimates (table S1). We first obtained pedigree- sion (^g = 0.37; 95% CI: 0.12 0.62). Figure 4C
upward in magnitude, whereas those causal based heritability estimates for each of the highlights the potential impacts of xAM and
for a secondary, unrelated trait under xAM traits of interest from the literature, using diagnostic errors on four selected trait pairs,
with the first were biased toward their effects estimates derived in demographically compa- and fig. S17 presents results for all pairs.
on that trait (figs. S7 to S9). These biases were rable samples (table S2). Together with the
asymptotically non-negligible (fig. S9). As a phenotypic mating correlations (Fig. 1A), these Genetic evidence for xAM recapitulates empirical
result, xAM increased the likelihood of reject- comprised inputs to forward time simulations cross-mate correlations
ing the null hypothesis of no association at all that we used to compute ^ r xAM . Cross-mate phenotypic correlation estimates
SNPs causal for either trait, increasing both Across 132 trait pairs, 42 evidenced ^g val- (rmate ) explained substantial variance in the
statistical power and false-positive rates (fig. ues significantly greater than zero (their 95% cross-chromosome even/odd PGS correlations
S10). Eventually, all variants affecting a sec- credible intervals did not include zero) after a r ‘;eo ; R2 = 47.66%; Fig. 5A).
in a linear model (^
ondary phenotype subject to xAM with the single generation of xAM, which increased to This association, which is congruent with ex-
GWAS phenotype will reach genome-wide 74 trait pairs after three generations. Across all pectations under phenotypically mediated
significance as sample size becomes large. trait pairs (including those not significantly dif- xAM [supplementary text (16)], persisted when

758 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

Fig. 3. Empirical and expected genetic correlations among UK Biobank phenotype pairs with nominally significant (p < 0.05) r^ emp values. (C) Projected
phenotypes. (A) We computed the expected LDSC genetic correlation estimate versus empirical LDSC estimates for all UKB phenotypes with nominally
in the absence of pleiotropy and after a given number of generations of xAM significant genetic correlation estimates. (D) Inverse variance weighted average
(^r xAM ), which we compared to empirical LDSC estimates (^r emp ) to obtain the ^g estimates within and between qualitatively similar phenotypic domains. Error
ratio ^g ¼ ^r xAM =^r emp . (B) The top 20 ^g estimates across previously studied UKB bars throughout represent 95% credible intervals.

accounting for measurement error and hetero- ated with empirical LDSC genetic correlation biological overlap. Further, regressing ^ r b;LDSC
skedasticity, and across PGS p-value thresh- r b;LDSC; R2 = 34.81%; Fig. 5B). This is
estimates (^ r ‘;eo and ^r yz simultaneously revealed that
on ^
olds (fig. S18). consistent with the hypothesis that empirical the association between ^ r b;LDSC and ^ r ‘;eo is
Additionally, cross-trait even/odd chromo- effect correlation estimates are capturing ad- mediated via ^r yz (DR2 < 0.001; partial effect
some PGS correlations were positively associ- ditional structure beyond the signatures of r ‘;eo versus p < 5e-8 for ^r yz). Thus,
p = 0.48 for ^

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 759

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

Fig. 4. Empirical and expected genetic correlations among psychiatric versus empirical LDSC estimates across psychiatric phenotype pairs. (C) The
phenotypes. (A) Ratios (^g ¼ ^r xAM =^r emp ) of projected LDSC genetic correlation potential combined impacts of bidirectional errors in diagnosis and xAM on
estimates under xAM alone relative to empirical genetic correlation estimates genetic correlation estimates for selected psychiatric disorder pairs. The red
(23) for six psychiatric disorders with 95% credible intervals. (B) Projected dashed line corresponds to ^g ¼ 1 across all panels.

alternative sources of structure independent PGS coincide with expectations after one or centrality of genetic correlation estimates in
from xAM do not appear to explain the positive more generations of xAM. We conclude that modern statistical genetics.
association between ^ ^ b;LDSC .
r ‘;eo and ρ xAM comprises a source of systematic bias in With this in mind, we comment on poten-
the study of genetic similarity across complex tial approaches to disentangling true effect
Discussion traits, one that subverts the widespread inter- correlation from xAM-induced artifact. First,
Nonzero effect correlation estimates have been pretation of genetic correlation as a direct in- family-based designs for addressing xAM (25)
widely interpreted as evidence for overlapping dex of biological similarity. are increasingly being applied to molecular ge-
genetic bases. It is therefore surprising that Our findings mirror recent results regard- netic data with promising results (26). Second,
substantial variation in genetic correlation ing the potential impacts of assortative mating we conjecture that approaches aimed at char-
estimates can be explained by cross-mate phe- across other areas of statistical genetics, in- acterizing effect heterogeneity across multiple
notypic correlations. Given the strength of this cluding marker-based heritability estimation phenotypes may provide a viable means for
association, the consequences of the random (10) and Mendelian randomization (12). Our identifying trait-specific loci: Although all trait-
mating assumption implicit in all commonly results also complicate the interpretation of a specific loci for either of two traits will achieve
used genetic correlation estimators warrant number of multivariate analytic frameworks. genome-wide significance in large sample
critical attention. For example, genomic structural equation mod- GWAS of either trait, effect estimates will re-
Our results show that cross-mate phenotypic eling (24), which takes marker-based genetic main substantially larger at causal loci. Finally,
correlations among many pairs of phenotypes correlation estimates as inputs, will propagate we propose that directly modeling the depen-
are strong enough that one or more gen- xAM-induced biases. This does not mean such dence between genotypes and their effects will
erations of xAM would substantially inflate methods are fundamentally flawed, but instead allow the differentiation of effect correlations
genetic correlation estimates. Further, the cor- demonstrates the importance of developing and score correlations in samples of unrelated
relation structures of even/odd chromosome unbiased effect correlation estimators given the individuals.

760 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

A B

Fig. 5. Genetic level evidence consistent with xAM in the UK Biobank. For a pair of traits Y, Z, the vertical axis reflects a single parameter to which
(A) Correlation between even and odd chromosome-specific polygenic scores the correlations between ^‘ y;even , ^‘ z;odd and between ^‘ y;odd , ^‘ z;even are both
(PGS) as a function of the cross-mate phenotypic correlation. For a single trait, constrained, and the horizontal axis reflects the cross-mate cross-trait correlation.
the vertical axis reflects the correlation between even and odd chromosome (B) Cross-trait even/odd PGS correlations as a function of empirical LD score
scores ^‘ even , ^‘ odd and the horizontal axis reflects the cross-mate correlation. regression genetic correlation estimates.

There are several limitations to the current ancestrally homogeneous groups (27), a broader AC KNOWLED GME NTS
investigation. Foremost among these are the characterization of population structure and We thank the editors and referees as well as A. L. Price and M. C. Keller
numerous assumptions about population dy- methods for addressing such structure will for their thorough and helpful feedback. Funding: Chan Zuckerberg
Initiative grant CZF2019-002449 (N.A.Z., R.B.), NIH grant
namics required to model xAM, including but likely be necessary to generate results that R01AG042568 (A.I.Y.), NIH grant R01CA227237 (N.A.Z., R.B.),
not limited to panmictic heritabilities, stability are maximally clinically relevant and can be NIH grant R01CA227466 (N.A.Z., R.B.), NIH grant R01ES029929
of cross-mate phenotypic correlation structures applied equitably. (N.A.Z., R.B.), NIH grant R01GM142112 (N.A.Z., R.B.), NIH grant
R01HG006399 (N.A.Z., R.B.), NIH grant R01HG011345 (N.A.Z., R.B.),
over succeeding generations, stability and NIH grant R01HL151152 (N.A.Z., R.B.), NIH grant R01HL155024
transmissability of environmental factors, and RE FERENCES AND NOTES (N.A.Z., R.B.), NIH grant R01MH125252 (N.A.Z., R.B.), NIH grant
the extent to which mating patterns reflect 1. V. Anttila et al., Science 360, eaap8757 (2018). R01MH122688 (N.A.Z., R.B.), NIH grant R01MH130581 (A.W.D., J.F.,
2. P. H. Lee, Y. A. Feng, J. W. Smoller, Biol. Psychiatry 89, 20–31 K.S.K., N.A.Z., N.C.), NIH grant R24AG065184 (A.I.Y.), NIH grant
social versus genetic homogamy. We proceeded (2021). R35GM125055 (S.S.), NIH grant R35GM133531 (N.A.Z., R.B.), NIH
under the tractable dynamical framework of 3. O. Weissbrod, J. Flint, S. Rosset, Am. J. Hum. Genet. 103, grant T32NS048004 (R.B.), NIH grant U01MH126798 (N.A.Z., R.B.),
two additive phenotypes subject to primary- 89–99 (2018). NIH grant U01HG009080 (N.A.Z., R.B.), National Science Foundation
4. B. K. Bulik-Sullivan et al., Nat. Genet. 47, 291–295 (2015). award CAREER-1943497 (S.S.), Open Philanthropy grant 010623-
phenotypic xAM with constant cross-mate
5. Y. Wu et al., Nat. Commun. 10, 1891 (2019). 00001 (A.I.Y.), Wellcome Trust grant 200176/Z/15/Z (J.F., K.S.K.).
correlations, stable nonheritable sources of vari- 6. B. Bulik-Sullivan et al., Nat. Genet. 47, 1236–1241 (2015). Author contributions: Conceptualization: A.W.D., J.F., K.S.K., N.A.Z.,
ation, and no vertical transmission. Though 7. J. Zheng et al., Bioinformatics 33, 272–279 (2017). R.B., T.W. Methodology: A.B., A.I.Y., A.J.S., A.W.D., G.A., N.A.Z., N.C.,
8. W. van Rheenen, W. J. Peyrot, A. J. Schork, S. H. Lee, R.B., S.S., T.W. Investigation: G.A., R.B. Visualization: R.B. Funding
each of these assumptions is likely untenable
N. R. Wray, Nat. Rev. Genet. 20, 567–581 (2019). acquisition: A.W.D., J.F., K.S.K., N.A.Z., N.C., S.S. Software: R.B.
for particular trait pairs, thereby compromising 9. P. H. Lee et al. Cell 179, 1469–1482.e11 (2019). Supervision: A.W.D., J.F., K.S.K., N.A.Z., S.S. Writing – original draft:
the accuracy of our projections, we hypothesize 10. R. Border et al., Nat. Commun. 13, 660 (2022). A.W.D., N.A.Z., R.B. Writing – review and editing: A.I.Y., A.J.S., A.W.D.,
that the qualitative phenomenon whereby xAM 11. L. Yengo et al., Nat. Hum. Behav. 2, 948–954 (2018). J.F., N.A.Z., R.B. Competing interests: The authors declare that
12. F. P. Hartwig, N. M. Davies, G. Davey Smith, Genet. Epidemiol. they have no competing interests. Data and materials availability:
inflates genetic correlation estimates will per- 42, 608–620 (2018). Cross-mate correlation estimates and all other results are provided in
sist for many traits. Nonetheless, we caution 13. A. Kong et al., Science 359, 424–428 (2018). tables S1 to S6 (16). Code used to perform simulations is publicly
that these projections are contingent upon 14. C. Bycroft et al., Nature 562, 203–209 (2018). available (28). UK Biobank data were accessed under application
15. O. Plana-Ripoll et al., JAMA Psychiatry 76, 259–270 (2019). 33127 and are available through the UK Biobank Access Management
multiple consequential decisions. Constructing 16. Materials and methods are available as supplementary materials. System [Link] Danish data are stored
a generative model that reconciles the associa- 17. D. S. Falconer, Introduction to Quantitative Genetics (Oliver and in a national high performance computing facility in Denmark.
tion between empirical mating patterns and Boyd, 1960). Access to data may be granted upon application to Statistics
18. J. Yang, S. H. Lee, M. E. Goddard, P. M. Visscher, Am. J. Hum. Genet. Denmark in accordance with Danish Law as detailed online (29).
genetic correlation estimates is an ill-posed 88, 76–82 (2011). License information: Copyright © 2022 the authors, some
inverse problem for which there are multiple 19. P.-R. Loh, G. Kichaev, S. Gazal, A. P. Schoech, A. L. Price, rights reserved; exclusive licensee American Association for the
solutions and of which we have only explored Nat. Genet. 50, 906–908 (2018). Advancement of Science. No claim to original US government
20. J. K. Haseman, R. C. Elston, Behav. Genet. 2, 3–19 (1972). works. [Link]
a subset. At the same time, existing methods 21. B. Bulik-Sullivan, Relationship between LD Score and journal-article-reuse
are only able to sidestep these decisions by Haseman-Elston Regression. bioRxiv [Preprint].
20 April 2015. [Link] SUPPLEMENTARY MATERIALS
making the strong (and often incorrect) as-
22. X. Zhou, Ann. Appl. Stat. 11, 2027–2051 (2017). [Link]/doi/10.1126/science.abo2059
sumption that mating is random. 23. A. D. Grotzinger et al., Nat. Genet. 54, 548–559 (2022). Materials and Methods
Lastly, we remark that xAM is, in essence, a 24. A. D. Grotzinger et al., Nat. Hum. Behav. 3, 513–525 (2019). Supplementary Text
25. M. C. Keller et al., PLOS Genet. 9, e1003451 (2013). Figs. S1 to S20
form of population structure not captured by 26. L. J. Howe et al., Nat. Genet. 54, 581–592 (2022). Tables S1 to S7
conventional principal-component– or mixed- 27. J. J. Berg et al., eLife 8, e39725 (2019).
References (30–46)
model–based correction. Given the increasing 28. R. Border, xAM_and_gen_corr, Zenodo (2022). [Link]
MDAR Reproducibility Checklist
10.5281/zenodo.7065696.
evidence that existing methods fail to complete- 29. Statistics Denmark, Data for research; [Link] Submitted 21 January 2022; accepted 28 September 2022
ly address structural factors, even in ostensibly TilSalg/Forskningsservice. 10.1126/science.abo2059

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 761

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

PLANT SCIENCE xerobranching stimulus (Fig. 1H). Simulations

reveal that an inward water flux dominates
Hydraulic fluxÐresponsive hormone redistribution under normal externally hydrated conditions,
and phloem-derived water is limited to close
determines root branching to companion cells (Fig. 1H). However, after
a xerobranching stimulus (when the exter-
Poonam Mehra1*, Bipin K. Pandey1, Dalia Melebari1, Jason Banda1, Nicola Leftley1, Valentin Couvreur2, nal water source is temporarily lost), root-tip
James Rowe3, Moran Anfang4, Hugues De Gernier5,6, Emily Morris1 , Craig J. Sturrock1, tissues rely on an outward flux of phloem-
Sacha J. Mooney1, Ranjan Swarup1, Christine Faulkner7, Tom Beeckman5,6, Rishikesh P. Bhalerao8, derived water to maintain growth, requiring
Eilon Shani4, Alexander M. Jones3, Ian C. Dodd9, Robert E. Sharp10, Ari Sadanandom11, ~8 hours for outer root tissues to receive phloem
Xavier Draye2, Malcolm J. Bennett1* water (Fig. 1H). Phloem-derived ABA is likely
to travel with this outward flow of water during
Plant roots exhibit plasticity in their branching patterns to forage efficiently for heterogeneously a xerobranching response.
distributed resources, such as soil water. The xerobranching response represses lateral root formation
when roots lose contact with water. Here, we show that xerobranching is regulated by radial movement ABA moves outward after a
of the phloem-derived hormone abscisic acid, which disrupts intercellular communication between xerobranching stimulus
inner and outer cell layers through plasmodesmata. Closure of these intercellular pores disrupts the To visualize whether, when, and where ABA
inward movement of the hormone signal auxin, blocking lateral root branching. Once root tips regain movement occurs after a xerobranching stim-
contact with moisture, the abscisic acid response rapidly attenuates. Our study reveals how roots ulus, we used a sensitive Förster resonance
adapt their branching pattern to heterogeneous soil water conditions by linking changes in hydraulic flux energy transfer (FRET)–based hormone bio-
with dynamic hormone redistribution. sensor (18), nlsABACUS2, which displays in-
creased fluorescence emission ratio upon ABA

R
binding (fig. S6). Seedlings expressing a
oot branching determines foraging capac- in an air gap), which causes branching to nuclear-localized nlsABACUS2 biosensor revealed
ity of crop plants (1, 2). Roots have evolved temporarily cease until roots reenter moist soil dynamic changes in ABA distribution and levels
plasticity in their branching patterns to (Fig. 1A). To discover the mechanistic basis of during our agar-based xerobranching bioassay
maximize access to soil water resources, xerobranching, an agar-based xerobranching (Fig. 1, I to K, and figs. S7 and S8). Spatially,
which are often distributed heteroge- bioassay was developed (fig. S1) that pheno- nlsABACUS2 accumulates in epidermal cells
neously (3, 4). Root branching is affected both copies the original x-ray computed tomography in the basal meristem and elongation zones
by extreme weather (such as drought or flood- (CT) soil-based bioassay (Fig. 1A) in experi- (Fig. 1J). Quantifying temporal changes in emis-
ing) (5, 6) and by transient or local spatial ments with Arabidopsis and tomato seedlings. sion ratios revealed that ABA levels initially
differences in soil moisture (7, 8). Negative As levels of the abiotic stress signal abscisic increase in root epidermal tissues >12 hours
effects of water scarcity on modern agriculture acid (ABA) increase in root tips during tran- after a xerobranching stimulus. Considering
will be exacerbated as climate change affects sient water stress (7), we tested whether tomato the sensitivity of the biosensor, these results
hydrological cycles and soil water resources ABA biosynthesis mutants (flacca and notabilis) are also consistent with our hydraulic simula-
(9, 10). Increased resilience in the face of climate (11, 12) are disrupted in xerobranching response. tions that predict the outward flux of phloem-
change will require better insight into how Both soil- and agar-based bioassays revealed derived water after ~8 hours of xerobranching
plant roots sense and adapt to fluctuating that, unlike wild-type (WT) controls, primary stimulus (Fig. 1H). Once root tips reconnect
water availability. roots of both flacca and notabilis continue to with the agar (recovery phase), ABA levels
branch when growing across an air gap (Fig. 1, A fall rapidly in epidermal cells (Fig. 1, J and K,
Dissecting rootsÕ water sensing to D, and fig. S2). Similarly, the ABA-deficient and fig. S8).
using xerobranching maize mutant vp14 (13) disrupted xerobranch- To directly visualize whether ABA moves
Xerobranching (7) provides an experimental ing in both paper-based and soil-based bio- radially outward (“piggy-backing” the redi-
model to study root adaptive responses to assays (figs. S3 and S4). Hence, ABA is a widely rection of water flux) after a xerobranching
transient water stress. A xerobranching re- conserved regulator that represses lateral root stimulus, we generated cross-sectional images
sponse is triggered when growing root tips development during a xerobranching response of nlsABACUS2 root-tip tissues (Fig. 2) from
temporarily lose contact with moist soil (e.g., in these monocot and eudicot plant models. three key zones. These zones included meri-
From which tissue(s) does ABA originate to stem and early elongation zone tissues (zone 1),
1
trigger a xerobranching response? ABA bio- mid- to late elongation and early differentiation
Plant and Crop Sciences, School of Biosciences, University
of Nottingham, Nottingham, UK. 2Earth and Life Institute, synthesis genes, such as ABA2, are expressed zone tissues (zone 2), and fully differentiated
Université catholique de Louvain, 1348 Louvain-la-Neuve, in phloem-related root vascular tissues (14). root tissues (zone 3). Radial images of these
Belgium. 3Sainsbury Laboratory, University of Cambridge, Arabidopsis aba2-1 mutant roots exhibit a three zones (taken at different time points
Cambridge, UK. 4School of Plant Sciences and Food Security,
Tel Aviv University, Tel Aviv, Israel. 5Department of Plant
xerobranching defect, which can be rescued during the xerobranching response) revealed
Biotechnology and Bioinformatics, Ghent University, 9052 by expressing WT ABA2 using the phloem contrasting patterns of ABA redistribution.
Ghent, Belgium. 6Center for Plant Systems Biology, companion cell–expressed SUC2 promoter For example, an increase in ABA levels can
VIB-UGent, 9052 Ghent, Belgium. 7Crop Genetics, John Innes
Centre, Norwich Research Park, Norwich, UK. 8Umeå Plant
(Fig. 1, E to G, and fig. S5A). A SUC2:YFP-ABA2 initially be detected in inner root tissues that
Science Centre, Department of Forest Genetics and Plant translation fusion confirmed that the ABA2 is later seen in outer tissues in zone 2 (com-
Physiology, Swedish University of Agricultural Sciences, protein can act in a phloem companion cell– pare nlsABACUS2 signals in 3- versus 4-mm
SE-901 87 Umeå, Sweden. 9Lancaster Environment Centre,
autonomous manner (fig. S5B). Phloem pro- root-tip sections for zone 2 in Fig. 2), consistent
Lancaster University, Lancaster, UK. 10Division of Plant
Science and Technology, University of Missouri, Columbia, vides water for growing roots in the temporary with ABA movement from inner to outer root
MO, USA. 11Department of Biosciences, University of absence of an external supply (15). Hydraulic tissues. By contrast, radial cross sections through
Durham, Durham, UK. modeling (16, 17) predicts that radial water nlsABACUS2 zone 1 root tissues revealed an
*Corresponding author. Email: [Link]@[Link] (P.M.);
[Link]@[Link] (M.J.B.) fluxes within root tissues change direction increase in ABA in outer (but not inner) tissues,
Present address: Oxford University Innovation, Oxford, UK. from predominately inward to outward after a which suggests that this signal originates from

762 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

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

phloem unloading in zone 2 and then undergoes

redistribution through symplastic continuity.
ABA accumulated much less in outer root tissues
of zone 3 versus zones 1 and 2 after a xero-
branching stimulus (Fig. 2). The results are
consistent with the phloem unloading patterns
of solutes described earlier (19), which suggests
that zone 2 overlaps with the protophloem
unloading zone, where solutes are transferred
laterally through symplastic continuity. By con-
trast, zone 3 is active in solute translocation
through phloem but inactive in phloem unload-
ing (19). Another parallel study also reports
similar expression patterns of nlsABACUS when
exogenously applied ABA unloads through
phloem (18). The pattern of elevated ABA in
outer root tissues of zones 1 and 2 also coin-
cides with the root-tip region, where new lateral
root primordia initiate (termed the basal meri-
stem or oscillatory zone) (20, 21). Elevated levels
of ABA in zones 1 and 2 rapidly decrease once
root-tip tissues contact a new moisture source
during the recovery phase (Fig. 2). Endogenous
ABA levels in cotyledons of nlsABACUS2 plants
did not change after a xerobranching stimulus,
which suggests that ABA accumulation is locally
confined to root tips that are exposed to air gaps
(figs. S7 to S9). Hence, during a xerobranching
response, ABA accumulates in the root zone that
is associated with the earliest stage of lateral root
development.
To test the function of the observed changes
in ABA after a xerobranching stimulus, we
examined the impact of mutations in key
components of the ABA signaling machinery
for this adaptive response. ABA-insensitive
mutants disrupting ABA receptors (pyr/pyl)
and ABA-activated SnRK2 kinases (snrk2.2/
2.3, snrk2.2/2.3/2.6) exhibited lateral root branch-
ing in our xerobranching bioassay (Fig. 3A and
figs. S10 and S11), whereas ABA-hypersensitive
pp2c mutants failed to produce lateral roots in
the air gap (fig. S12). Detailed analysis with
snrk triple, septuple, and nonuple mutants as
well as snrk2.2 complementation lines revealed Fig. 1. ABA functions as a repressive signal during xerobranching response. (A to D) Representative
that SnRK2.2 is required for xerobranching (Fig. CT images [(A) to (C)] and a bar graph (D) showing xerobranching defect in ABA-deficient tomato mutants
3D and fig. S11). Roots of a snrk nonuple mutant (notabilis and flacca) versus WT (Ailsa Craig). Air gap, ~1.5 cm. Scale bar, 1 cm. n.d., not detected.
(which lacked 9 of 10 members of the SnRK2 (E) Agar-based air-gap assay system showing xerobranching response in Arabidopsis WT (Col-0) (air gap,
family except SnRK2.2) behaved like WT in ~5 mm). (F) Arabidopsis ABA biosynthetic mutant aba2 produces lateral roots in air gap. (G) Expression of
our xerobranching bioassay (fig. S11). Further- ABA2 in phloem companion cells rescues xerobranching defect of aba2. (H) The MECHA hydraulic model
more, a green fluorescent protein (GFP)–tagged predicts phloem as the main source of water in Arabidopsis root elongation zone only when located in an air
SnRK2.2 reporter pSnRK2.2:SnRK2.2-GFP re- gap. Simulated water advection maps show dominating phloem water in root epidermal cells within
vealed that a xerobranching stimulus induced 8 hours. (I) For studying temporal response of reporters during xerobranching, primary root tips were
the fusion protein in the basal root meristem grown to different lengths in air-gap (0 to 5 mm) and recovery (up to 2 mm in moist agar) conditions.
(Fig. 3, B and C, and fig. S13), which overlaps with (J and K) Emission ratio images (J) and box plots (K) of nlsABACUS2-400n reveal dynamic changes in
the site of largest emission ratio changes of the endogenous ABA levels as root tips grow across air gap and after recovery. Color scale represents
ABA biosensor nlsABACUS2 (Fig. 1J). Hence, emission ratios. *P ≤ 0.05; ***P ≤ 0.001; ns, nonsignificant changes.
SnRK2.2 regulates ABA (and in turn, the xero-
branching) response in Arabidopsis root tissues.
source to outer root tissues during a xero- ing response. To determine this, we attempted
ABA regulates xerobranching by branching response (Fig. 1, E to H and J; Fig. to rescue the xerobranching defect of the ABA
closing plasmodesmata 2; and Fig. 3B). However, it remains unclear response mutant snrk2.2/2.3 by expressing the
Genetic and biosensor-based studies revealed which root tissue(s) ABA targets to repress WT SnRK2.2 sequence under a series of root-
that ABA moves from its inner root (phloem) lateral root development during a xerobranch- tissue and/or zone-specific promoters (Fig. 3,

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 763

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

E to J). Expressing SnRK2.2 in either root epi-

dermal, cortical, endodermal, or basal meristem
(but not lateral root cap) tissues rescued the
snrk2.2/2.3 xerobranching defect (Fig. 3, E to
J). This result led us to ask, is there a common
target important for xerobranching that the
ABA signaling pathway regulates in each of
these outer root tissues? ABA modulates hy-
draulic conductivity by inducing aquaporin
expression to affect water movement out of
the vasculature (22). To investigate possible
roles of aquaporins in ABA-dependent repres-
sion of branching in air gaps, we analyzed a pip
(plasma membrane intrinsic proteins) quad-
ruple mutant in our xerobranching bioassay
system. However, analysis of the pip mutant
disrupting the most highly expressed PIP genes
in roots did not show any significant difference
compared with the WT control (fig. S14).
ABA has been reported to trigger reversible
closure of intercellular pores, termed plasmo-
desmata (PD) (23). Does ABA regulate the xero-
branching response by gating PD between root
tissues? To address this, the symplastic fluores-
cent tracer carboxyfluorescein (CF) [derived
from carboxyfluorescein diacetate (CFDA)]
was used to monitor whether PD were open
or shut in root tissues during a xerobranching
response. Within plant cells, nonfluorescent
CFDA is cleaved by cellular esterases to release
membrane-impermeable fluorescent CF that
travels exclusively through the PD. After CFDA
application to cotyledons, we observed decreased
protophloem unloading in Arabidopsis WT root
tips that were exposed to a xerobranching
stimulus (fig. S15). Delayed unloading was
also observed in root tips of a pSUC2:GFP
line that expresses free GFP in phloem com-
panion cells (fig. S16). Collectively, both lines
of evidence reveal that root tips experience
decreased PD permeability after a xerobranch-
ing stimulus.
To assess the role of ABA in root PD gating
during xerobranching, CFDA was locally ap-
plied to root tips. WT roots exhibited reduced
intercellular fluxes of the tracer, which indi-
cated closed PD after a xerobranching stimulus
(Fig. 4, A and B). By contrast, CFDA treatment of
root tips of the ABA response mutant snrk2.2/
2.3 revealed that PD remained open after a
xerobranching stimulus (fig. S17). PD aperture
Fig. 2. ABA moves radially outward after a xerobranching stimulus. Ratio images of Arabidopsis is regulated by modifying the size of a callose
roots expressing nlsABACUS2-400n reveal spatiotemporal distribution of ABA as the roots grow ring encircling each end of the intercellular
across an air gap. Each image is a representative maximum projection ratio map generated from a pore. The callose stain aniline blue revealed
stack of several radial cross sections. Roots were imaged at different time points (indicated with hours) higher callose deposition in basal meristem
as they exit moist agar (control), grow in a 5-mm air gap, and reconnect with agar (recovery). cells of WT roots compared with snrk2.2/2.3
Length (in millimeters) on the vertical axis indicates the length of root tips in air-gap and recovery after a xerobranching stimulus (fig. S18). Taken
conditions. For each time point, cross sections were generated from different zones of roots marked together, the evidence indicates that ABA regu-
as zone 1 (early elongation and meristem), zone 2 (mid- to late elongation and early differentiation), and lates gating of PD in key root tissues during a
zone 3 (differentiation). Radial cross sections from zone 1 (compare cross sections from 1 to 5 mm) xerobranching response.
reveal an increase in ABA in outer (but not inner) tissues, which appears to originate from phloem To determine the function of PD gating
unloading in zone 2 after a xerobranching stimulus. Color scale represents FRET emission ratios. during a xerobranching response, Arabidopsis
Higher ratios correspond to higher ABA concentrations. Experiment was performed with at least four mutants no longer able to gate their PD were
replicates per time point per zone (n = 4). characterized (figs. S19 and S20). PD aperture

764 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

ability to close PD disrupts a root’s ability to

activate a xerobranching response.
How does ABA trigger PD closure after a
xerobranching stimulus? Transgenic lines ex-
pressing GFP reporters that were fused to either
PDLP2 or PDLP3 proteins under their native
promoters were up-regulated in response to a
xerobranching stimulus (Fig. 4, C and D, and
figs. S21 to S23). In the case of PDLP3, its
mRNA and PDLP3:PDLP3-GFP reporter were
up-regulated in root tip and, specifically, basal
meristem cells during a xerobranching response,
exhibiting a punctate pattern in apical, basal,
and side walls, consistent with its PD regulatory
function (Fig. 4, C and D, and figs. S21 and S23).
Treating roots of PDLP3:PDLP3-GFP with low
levels of ABA (50 nM) phenocopied a xero-
branching stimulus (fig. S24), in agreement with
the presence of several ABA-responsive elements
(ABREs) in its promoter sequence (table S1).
We also noted that many other PDLP and callose
synthase genes contained ABREs in their pro-
moter sequences (tables S1 and S2). Comparing
spatiotemporal response curves of nlsABACUS2
and PDLP3-GFP (during xerobranching and
subsequent recovery) reveals that ABA accu-
mulation precedes PDLP3 induction (Fig. 1K
and fig. S23). To test the functional impor-
tance of ABA-regulated PDLP up-regulation
during xerobranching, we characterized the
ABA-insensitive mutant abi1-1. The dominant
negative abi1-1 allele disrupts ABA-dependent
PD closure and maintains high frequencies of
open PD (26). Unlike WT, abi1-1 roots exhibit
branching in air gaps (fig. S25). However, ectopic
expression of PDLP1 (using 35S rather than its
native promoter) rescued the xerobranching
defect of abi1-1 (fig. S25). Hence, ABA triggers
PD closure after a xerobranching stimulus by
up-regulating expression of PDLPs. Additionally,
GAL4-mediated transactivation of abi1-1 in root
ground tissues (cortex and endodermis) led to
maximal disruption of xerobranching response
as compared with induction of abi1-1 in single
root cell layers (fig. S26). Hence, ABA-mediated
PD closure in root ground tissues is essential for
Fig. 3. Xerobranching is dependent on ABA perception in Arabidopsis primary root transition zone tissues. a xerobranching response.
(A) ABA signaling mutant snrk2.2/2.3 disrupts xerobranching in agar-based air-gap assay system. Air gap, ~5 mm.
(B and C) Elevated response of pSnRK2.2:SnRK2.2-GFP in root tips subjected to air-gap versus control conditions. PD closure blocks inward auxin movement
Scale bar, 100 mm. (D) Expression of SnRK2.2 under its own promoter rescues the xerobranching defect of snrk2.2/ How does ABA-mediated PD gating disrupt
2.3. (E to G) Functional complementation of snrk2.2/2.3 by tissue-specific expression of SnRK2.2 in epidermis, lateral root branching? Our reporter studies
cortex, and endodermis using the WER, CO2, and SCR promoters, respectively. (H) Root zone–specific expression of reveal that closing PD during a xerobranching
SnRK2.2 in root meristem and transition zone (using RCH promoter) rescues the xerobranching defect of snrk2.2/ response blocks the inward (symplastic) move-
2.3. (I) Expression of SnRK2.2 in columella and lateral root cap using SMB promoter in snrk2.2/2.3 background. ment of another plant hormone signal, auxin,
(J) Number of lateral roots produced by WT (Col-0), snrk2.2/2.3, and different tissue- and/or zone-specific rescue which is required to initiate lateral root branch-
lines in the air-gap region versus the corresponding region of roots grown under control conditions. Orange color in ing. Exposing roots of the DR5:VENUS auxin
root schematics defines the tissue and/or zone of expression of selected promoters used for complementation. reporter line to a xerobranching stimulus results
Significant changes with respect to control (C) or WT (J) were calculated by Student’s t test. ***P ≤ 0.001. in an elevated auxin response in epidermal cells
within the root basal meristem zone (Fig. 4, E
is regulated through the action of enzymes, in- family resulted in nine of these lines exhibiting and F). The altered DR5:VENUS reporter pattern
cluding a family of callose synthases and as- a xerobranching defect, including CALS7, which is consistent with xerobranching-induced PD
sociated regulatory proteins such as PD-located is primarily associated with phloem sieve pores closure in basal meristem epidermal cells tran-
proteins (PDLPs) (24). Mutating individual mem- (fig. S19) (25), as did mutating both PDLP2 and siently blocking the radially inward symplastic
bers of the Arabidopsis callose synthase gene PDLP3 genes (fig. S20). Hence, blocking the movement of auxin (Fig. 4, E and F, and fig. S27).

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 765

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

Fig. 4. ABA blocks radial

auxin flow by reducing PD
permeability causing
xerobranching. (A) Slower
migration of symplastic tracer
CF in WT Arabidopsis (Col-0)
primary root tips subjected to
xerobranching stimulus versus
control conditions. CF migra-
tion was indistinguishable in
ABA signaling mutant snrk2.2/
2.3 with or without the xero-
branching stimulus. Root tips
were treated with CFDA before
imaging. (B) CF fluorescence
intensity in WT and snrk2.2/
2.3 root tips exposed to air-
gap versus control conditions.
(C) Xerobranching stimulus
induces expression of pPDLP3:
PDLP3-GFP. (D) Box plots
showing significantly enhanced
response of pPDLP3:PDLP3-
GFP in air-gap versus control
conditions. (E and F) Auxin
response reporter DR5:VENUS
shows higher fluorescence
signal in epidermis of root
tips exposed to air-gap
versus control and recovery
conditions. (G and H) ABA
treatments phenocopy
DR5:VENUS response during
xerobranching. Scale bar,
100 mm. *P ≤ 0.05; **P ≤
0.01; ***P ≤ 0.001; ns,
nonsignificant (StudentÕs
t test).

The auxin reporter DII-VENUS also revealed symplastic auxin movement that is required for Collectively, multiple lines of evidence estab-
higher auxin abundance in root tips during lateral root induction. lish a role for ABA-mediated PD gating of radial
exposure to a xerobranching stimulus (figs. Root auxin distribution was originally thought symplastic auxin flow to regulate xerobranch-
S27 and S28). The effect of a xerobranching to be exclusively determined by the combined ing. Roots appear to sense water availability by
stimulus on DR5:VENUS and DII-VENUS can activities of AUX1/LAX, ABCB, and PIN auxin comobilizing water and hormones through
be mimicked by exogenous application of ABA influx and efflux carriers (27). To determine the PD. If water, ABA, and auxin moved through
to roots (Fig. 4, G and H, and fig. S29). ABA- impact of xerobranching on carrier-mediated substrate-specific carriers (aquaporins, ABCG,
mediated PD gating during a xerobranching auxin transport, reporter lines for key auxin and PINs), their comobilization would be broken.
response blocks radial inward auxin movement influx (pAUX1:AUX1-YFP) or efflux (pPIN2: Instead, because PD are large intercellular pores
and lateral root induction. Consistent with this PIN2-GFP) carriers expressed in the root elonga- that are nonselective, radial hydraulic fluxes
model, switching off ABA signaling in the tion zone were analyzed (fig. S32). No spatial comobilize hormones either inward (such as
abi1-1 mutant (which ensured that PD remained changes were detected in AUX1-YFP or PIN2- auxin) or outward (such as ABA) when water
constitutively open) blocked any change in DII- GFP expression in the outermost root tissues is available or not, respectively. Thus, auxin
VENUS auxin response after a xerobranching after a xerobranching stimulus, consistent with and ABA can be considered as hydrosignals
stimulus (fig. S30). Suppression of lateral root the observed elevated auxin response (Fig. 4, E when comobilized through PD with water.
initiation in roots exposed to a xerobranching and F, and fig. S27) being the result of gating Our study has revealed that xerobranching
stimulus was visualized by using the DR5:LUC of hormone movement through PD rather than is a widely conserved, ABA-dependent root
reporter, which normally demarks groups of through modulation of apoplastic auxin trans- adaptive response to localized loss of contact
lateral root stem cells in the basal meristem port (fig. S32). Ectopic expression of either AUX1 with soil water in eudicot and monocot
(or oscillation zone) that go on to form new or PIN2 in every elongation zone tissue, to create species. We also report the molecular and
branches in control roots, but these DR5:LUC a synthetic radial apoplastic auxin pathway, suc- cellular mechanisms that enable plant roots to
oscillations were blocked in roots exposed to cessfully bypassed the PD-mediated restriction block lateral root formation after experiencing
a xerobranching stimulus (fig. S31). Hence, the of symplastic auxin flow and led to lateral root a xerobranching stimulus (see schematic in
xerobranching response blocks radially inward branching in air gaps (fig. S33). Fig. 5). Under optimal moisture conditions,

766 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

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

Fig. 5. Dynamic hormone redistribution determines root branching or transient absence of external water, phloem-derived water comobilizes ABA
xerobranching response to external water availability. Schematics illustrate from inner (stele) to outer root cells (epidermis) through PD (i). ABA induces
cellular auxin, ABA, and water fluxes in root basal meristem (highlighted closure of PD through PDLPs and CALs (ii) disrupting symplastic radial
cell files in transverse root section) exposed to water or air. (A) Under normal movement of auxin to pericycle (iii). (C) Ectopic expression of auxin carriers
moisture conditions, roots cotransport water and auxin radially inward through such as AUX1 in all elongation zone tissue creates a synthetic radial apoplastic
PD, which delivers the key hormone signal (auxin) to pericycle cells, where auxin pathway, which bypasses the disrupted symplastic auxin flow and thus
lateral root initiation is triggered. (B) Xerobranching stimulus triggers rise in ABA causes root branching in air gap. Ct, cortex; En, endodermis; Ep, epidermis;
levels in inner root tissues (stele), which suppresses root branching. In the Pe, pericycle; St, stele.

roots cotransport auxin and water inward be- stimulus levels. These dynamic regulatory 7. B. Orman-Ligeza et al., Curr. Biol. 28, 3165–3173.e5 (2018).
tween root basal meristem tissues through PD events—coupling changes in hydraulic fluxes 8. Y. Bao et al., Proc. Natl. Acad. Sci. U.S.A. 111, 9319–9324
(2014).
and thus provide a symplastic pathway to with hormone redistribution, which we term 9. D. B. Lobell, S. M. Gourdji, Plant Physiol. 160, 1686–1697
radially mobilize this key hormone signal with hydrosignaling—enable plant roots to calibrate (2012).
10. Y. Grusson, I. Wesström, E. Svedberg, A. Joel, Agric. Water
hydraulic fluxes from epidermal to pericycle spatial root branching responses in heteroge-
Manage. 249, 106766 (2021).
cells, where auxin triggers lateral root initia- neous soil environments, which are the norm 11. A. J. Thompson et al., Plant Cell Environ. 27, 459–471 (2004).
tion. However, when the external source of rather than the exception. Climate change is 12. I. B. Taylor, R. S. T. Linforth, R. J. Al‐Naieb, W. R. Bowman,
B. A. Marples, Plant Cell Environ. 11, 739–745 (1988).
water is transiently unavailable, root tips rely likely to exacerbate water scarcity (9, 10); thus, 13. B. C. Tan, S. H. Schwartz, J. A. D. Zeevaart, D. R. McCarty,
on an internal (phloem) source of water to that water in soils will be even more heteroge- Proc. Natl. Acad. Sci. U.S.A. 94, 12235–12240 (1997).
continue growing (15) and couple its outward neously distributed in the future than at present 14. T. Kuromori, E. Sugimoto, K. Shinozaki, Plant Physiol. 164,
1587–1592 (2014).
radial flow with movement of the hormone is highly likely. Insight into the molecular and 15. B. S. Wiegers, A. Y. Cheer, W. K. Silk, Plant Physiol. 150,
ABA. This key water stress signal reduces PD cellular mechanisms through which roots accli- 2092–2103 (2009).
aperture, disrupting radial movement of the mate under these conditions may improve the 16. V. Couvreur et al., Plant Physiol. 178, 1689–1703 (2018).
17. F. C. Pascut et al., Nat. Commun. 12, 4682 (2021).
branching signal auxin to lateral root “stem climate resilience of future crops. 18. J. Rowe et al., bioRxiv 2022.10.19.512731 [Preprint] (2022).
cells” in the pericycle. Restoring an external 19. T. J. Ross-Elliott et al., elife 6, e24125 (2017).
RE FERENCES AND NOTES 20. I. De Smet et al., Development 134, 681–690 (2007).
water source (during the recovery phase)
1. B. Muller et al., Trends Plant Sci. 24, 810–825 (2019). 21. M. A. Moreno-Risueno et al., Science 329, 1306–1311 (2010).
rapidly reduces ABA levels in the outermost 22. K. Prado et al., Plant Cell 25, 1029–1039 (2013).
2. J. P. Lynch, Plant Physiol. 156, 1041–1049 (2011).
root tissues (>3 hours) and thus restores open 3. H. Motte, T. Beeckman, J. Exp. Bot. 70, 785–793 (2019). 23. Y. Benitez-Alfonso, Plant Cell Physiol. 60, 713–714 (2019).
PD and symplastic auxin transport. Once recon- 4. E. C. Morris et al., Curr. Biol. 27, R919–R930 (2017). 24. C. L. Thomas, E. M. Bayer, C. Ritzenthaler,
5. A. Zhan, H. Schneider, J. P. Lynch, Plant Physiol. 168, L. Fernandez-Calvino, A. J. Maule, PLOS Biol. 6, e7
nected with an external water source, the
1603–1615 (2015). (2008).
abundance and distribution of these signals 6. R. Karlova, D. Boer, S. Hayes, C. Testerink, Plant Physiol. 187, 25. B. Xie, X. Wang, M. Zhu, Z. Zhang, Z. Hong, Plant J. 65, 1–14
reset in root-tip tissues to pre–xerobranching- 1057–1070 (2021). (2011).

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 767

RES EARCH

26. S. Tylewicz et al., Science 360, 212–215 (2018). Horizon 2020 research and innovation program, Marie R.S., S.J.M., C.F., T.B., R.P.B., E.S., A.M.J., I.C.D., R.E.S., A.S., X.D.,
27. J. Petrášek, J. Friml, Development 136, 2675–2688 (2009). Skłodowska-Curie, XEROBRANCHING (891262 to P.M.); and M.J.B. Competing interests: The authors declare no
Biotechnology and Biological Sciences Research Council Discovery competing interests. Data and materials availability: No
ACKN OW LEDG MEN TS Fellowship (BB/V00557X/1 to B.K.P.); the Biotechnology and restrictions are placed on materials, such as materials transfer
We thank P. L. Rodriguez (CSIC-UPV, Spain) for sharing pyr/ Biological Sciences Research Council (BB/T001437/1 to M.J.B. and agreements. All data are available in the main text or the
pyl1124589, pyr/pyl1124578, and pyr/pyl112458 mutants; H. Fujii A.S.; BB/V003534/1 to M.J.B., A.S., and J.B.; BB/W008874/1 to supplementary materials. License information: Copyright © 2022
(University of Turku, Finland) for sharing snrk2.2/3/6, snrk2.1/4/ M.J.B. and C.J.S.; BB/W015080/1 to M.J.B.; and BB/P018572/1 the authors, some rights reserved; exclusive licensee American
5/7/8/9/10, and snrk2.1/3/4/5/6/7/8/9/10 mutants; and R. Miao to A.M.J.); European Research Council grant FUTUREROOTS Association for the Advancement of Science. No claim to original
(Fujian Agriculture and Forest University, China) for sharing (294729 to M.J.B.); European Research Council, SUMOrice (grant US government works. [Link]
hab1-1abi1-2pp2ca, hab1-1abi1-2abi2-2, and Qabi2-2 mutants. We no. 310235 to A.S.); Fonds de la Recherche Scientifique – FNRS licenses-journal-article-reuse
thank L. Kalmbach, Y. Helariutta (SLCU, UK), and N. Geldner (to V.C.); the Swedish Research Council (no. 2020-03522 to R.P.B.);
(University of Lausanne, Switzerland) for sharing pSUC2:GFP European Research Council, RobustHormoneTrans (757683 to E.S.); SUPPLEMENTARY MATERIALS
seeds. We also thank Z. Gong and Y. Wang (China Agricultural the NSF Plant Genome Research Program (IOS-1444448 to R.E.S.);
[Link]/doi/10.1126/science.add3771
University) for kindly sharing abi1-1 (Col-0) seeds. We acknowledge the Gatsby Charitable Foundation (to A.M.J. and J.R.); European
Materials and Methods
C. Luschnig (University of Natural Resources and Life Sciences – Research Council grant INTERCELLAR (725459 to C.F.); and
Figs. S1 to S33
BOKU, Vienna) and J. Friml (IST, Austria) for sharing eir1-4/PIN2: the Biotechnology and Biological Research Council Institute
Tables S1 and S2
PIN2-GFP and eir1-4/35S:PIN2-GFP lines. We thank F. Augstein and Strategic Programme, Plant Health (BBS/E/J/000PR9796 to C.F.).
References (28–59)
A. Carlsbecker (Uppsala University, Sweden) for providing UAS:abi1-1 Author contributions: Conceptualization: P.M. and M.J.B.
MDAR Reproducibility Checklist
and driver lines. We thank F. Chaumont (Louvain-la-Neuve, Methodology: P.M., B.K.P., V.C., J.R., and H.D.G. Investigation:
Belgium) and A. Schäffner (Helmholtz Zentrum München, P.M., B.K.P., D.M., J.B., N.L., V.C., J.R., M.A., H.D.G., E.M., C.J.S.,
Germany) for sharing the aquaporin mutant. Funding: This study and R.S. Funding acquisition: P.M. and M.J.B. Supervision: M.J.B.
was supported by a European Molecular Biology Organization Writing – original draft: M.J.B. and P.M. Writing – review & editing: Submitted 8 June 2022; accepted 24 October 2022
Long-Term Fellowship (ALTF 1140-2019 to P.M.); European Union’s P.M., B.K.P., D.M., J.B., N.L., V.C., J.R., M.A., H.D.G., E.M., C.J.S., 10.1126/science.add3771

◥ enhance printing speed and resolution (11, 12).

REPORTS Currently, almost all two-photon initiators are
organic molecules that do not enhance the
3D PRINTING properties and structural complexity of the
final printed products (13, 14). In addition,
Mechanical nanolattices printed using each organic photoinitiator is typically efficient
only for a single type of monomer. Composite
nanocluster-based photoresists nanostructures can be printed by adding metal
ions or inorganic particles to existing two-
Qi Li1*†, John Kulikowski1†, David Doan1†, Ottman A. Tertuliano1, Charles J. Zeman IV2, photon resists. This strategy is robust for simple
Melody M. Wang1, George C. Schatz2, X. Wendy Gu1* nanostructures (e.g., nanowires, 2D patterns)
(15) but ineffective for complex mechanical
Natural materials exhibit emergent mechanical properties as a result of their nanoarchitected, nanocomposite nanolattices, as performance suffers when small
structures with optimized hierarchy, anisotropy, and nanoporosity. Fabrication of such complex systems structural flaws are present (5, 16). Uncontrolla-
is currently challenging because high-quality three-dimensional (3D) nanoprinting is mostly limited ble metal ion reduction and particle growth
to simple, homogeneous materials. We report a strategy for the rapid nanoprinting of complex structural lead to structural flaws and reduced feature
nanocomposites using metal nanoclusters. These ultrasmall, quantum-confined nanoclusters function quality, and aggregated particles interfere with
as highly sensitive two-photon activators and simultaneously serve as precursors for mechanical light propagation (5, 16). Composite nano-
reinforcements and nanoscale porogens. Nanocomposites with complex 3D architectures are printed, lattices can also be fabricated by depositing
as well as structures with tunable, hierarchical, and anisotropic nanoporosity. Nanocluster-polymer an inorganic layer onto a printed polymeric
nanolattices exhibit high specific strength, energy absorption, deformability, and recoverability. This scaffold (17, 18). This approach results in im-
framework provides a generalizable, versatile approach for the use of photoactive nanomaterials in pressive mechanical performance but requires
additive manufacturing of complex systems with emergent mechanical properties. a time-consuming, multistep fabrication pro-
cess and is limited to core-shell structures.

A
In the past decade, ultrasmall metal nano-
dvances in three-dimensional (3D) print- (6, 7), stiffness (8), or energy absorption per clusters 1 to 2 nm in size have been synthe-
ing rely on the invention of new printing weight (9). Improved performance can be sized with precise atomic structures (19). These
techniques (1, 2) as well as the creation achieved by printing advanced materials with quantum-confined nanoclusters show non-
of compatible feedstock materials such complex internal nanofeatures, such as nano- metallic, molecule-like electronic structures and
as resins, filaments, powders, and photo- composites with optimized spacing and inter- optical properties. Metal nanoclusters consist
resists (3, 4). A current frontier is the printing actions between soft and hard components, of metal atoms with zero partial charge, which
of complex materials with arbitrary 3D nano- or the hierarchical graded porosity and struc- avoids the high energy requirements typical of
architectures. These structures are of partic- tural anisotropy of natural structural materials metal reduction. The ultrasmall size of the
ular interest as next-generation mechanical (10). Fabrication of these complex nanostruc- nanoclusters circumvents the light scattering
metamaterials (5) that combine nanoscale tures is currently limited by existing printing that would occur with larger metal particles.
material size effects with lightweight lattice techniques and lack of appropriate feedstock. The densities of many metal nanoclusters are
structures, which leads to optimized strength Two-photon (or multiphoton) lithography much lower than their bulk metal counterparts
can achieve polymeric nanostructures with (19). Therefore, ultrasmall nanoclusters are
1
Department of Mechanical Engineering, Stanford University, almost arbitrary geometry (1, 2). Two-photon prime candidates for replacement of tradi-
Stanford, CA 94305, USA. 2Department of Chemistry, resists generally use photoinitiators that ab- tional organic molecules as photoinitiators
Northwestern University, Evanston, IL 60208, USA. sorb photons to form reactive species (e.g., and substitutes for metal ions or particles as
*Corresponding author. Email: qilistan@[Link] (Q.L.);
xwgu@[Link] (X.W.G) radicals) that initiate polymerization. High- inorganic precursors. By doing so, fast print-
†These authors contributed equally to this work. performance two-photon initiators can greatly ing of nanocomposite nanolattices with high

768 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

A B radical polymerization cationic polymerization D 30

O O
O O O O

Laser Power (mW)

O O O O 25
O O
O O photoinitiator photoacid
HO 1 nm
HO generator O O 20
monomer nanocluster
O 15
photoresist O C O
substrate 3
O2 photosensitizer
O
O 10
C O OH 1
O2
OH O
objective oil O OH 5 Ag28Pt
O OH Au25
C O -sheet O 0
O 0 20 40 60 80 100 120 140 160
protein crosslinking
Scan Speed (mm/s)

C
Sn
S3 E
S2
singlet excited state (~us)
R1 singlet excited state (ps~ns)
S1 R2 S1
R3 T3
.... T2 ~400 nm
R
triplet excited state (~us)
slow decay

fast decay
T1

x
ground state ground state

nanocluster photoinitiator vs. organic photoinitiator

F G H I

Fig. 1. Photochemistry and printability of nanocluster-based photoresist. nanocluster, showing diverse fragmentation pathways (orange shading and
(A) Photoresist composition and printing setup. (B) Schematic of photo- gray arrows). The right inset shows a typical organic photoinitiator and
polymerization of different classes of monomers during two-photon lithography. its bond-dissociation position (orange X). (D) Fabrication window (laser power
The atomic structure of the nanocluster (Ag28Pt) is reproduced from (20). and scanning speed) for Ag28Pt and Au25 photoresists. (E) Open-faced table
(C) Comparison of photochemistry of nanocluster photoinitiators (left) with fabricated using Ag28Pt photoresist. (F and G) Octet lattices and (H and
typical organic photoinitiators (right). The left inset shows a snapshot from an I) Schwarz primitive (SP) lattices made with Au25 photoresist. Scale bars:
excited-state Born-Oppenheimer molecular dynamics simulation of an Ag28Pt (E) 2 μm, (F) 100 μm, (G) 10 μm, (H) 100 μm, (I) 10 μm.

structural quality, fidelity, and complexity can nanocluster (20) and the rod-shaped Au25 nano- is generally necessary for the generation of
be achieved. cluster (21) were selected as the two-photon reactive species (13, 14), which may reduce
Our nanocluster-based photoresist con- initiators in this study. The two-photon absorp- overall photoinitiation efficiency (Fig. 1C, right).
sists only of metal nanoclusters, mono- tion cross sections of Ag28Pt and Au25 were More importantly, compared with molecular
mers [e.g., an acrylate monomer such as found to be >30 Goeppert-Mayer units (GM) photoinitiators that usually have limited frag-
pentaerythritol triacrylate (PETA) or epoxy and >100 GM at sub–800-nm wavelengths, re- mentation pathways, nanoclusters provide for
monomer such as 3,4-epoxycyclohexylmethyl spectively, as determined by quadratic response a greater variety of bond scissions (Fig. 1C, left,
3,4-epoxycyclohexanecarboxylate (EEC)] time-dependent density functional theory cal- and movies S1 and S2), which lead to increased
(Fig. 1A), and a solvent (2-propanol), which culations (fig. S1). Furthermore, these two nano- reactivity with different reagents.
increases the solubility of nanoclusters in the clusters have stable, long-living S1 excited states The printability of nanocluster-acrylate pho-
photoresist. We find that nanoclusters can (lifetime ~3 ms) (21), which are beneficial for the toresists was systematically evaluated. Squares
serve as photoinitiators for radical polymer- generation of radical or other reactive species and lines were fabricated as test structures
ization, photoacid generators for cationic (Fig. 1C, left). This is different from organic to investigate the required scanning speed,
polymerization, and photosensitizers that pro- two-photon initiators in which the S1 state threshold laser power, and smallest achiev-
mote singlet oxygen formation to induce the is short lived (ps-ns). In organic two-photon able feature size of the nanocluster-acrylate
crosslinking of proteins (Fig. 1B). The Ag28Pt initiators, intersystem crossing to triplet states photoresist. For 5 wt % Ag28Pt photoresist

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 769

RES EARCH | R E P O R T S

A 350 B 10% Ag28Pt C

Ip-Dip (22)

Eng. Stress (MPa)

300 150

250
200 100
150
100 50
50
0 0
0 5 10 15 20 25 30 0 10 20 30 40 50
Eng. Strain (%) Eng. Strain (%)
50
D 0.19 E F G
Eng. Stress (MPa)

40
densification
30

0
0 10 20 30 40 50 60 70
Eng. Strain (%)
40
H 0.20 I J K
Eng. Stress (MPa)

30
densification
20

0
0 10 20 30 40 50 60 70
Eng. Strain (%)
Fig. 2. Mechanical behavior of nanocluster-polymer nanolattices. (A) Engineer- from (22). (D) Engineering stress–engineering strain curve for an octet lattice.
ing stress-strain curve of pillars under cyclic loading. Inset shows SEM image of (E to G) In situ SEM compression of an octet lattice. (H) Engineering stress–
a pillar before compression. (B) Engineering stress–engineering strain curves and engineering strain curve for SP lattice. (I to K) In situ SEM compression of an SP
(C) SEM image of honeycomb lattices. Stress-strain curve for Ip-Dip honeycomb is lattice. Scale bars: (A) 2 μm, (C) 8 μm, (E to G and I to K) 10 μm.

(5 mg Ag28Pt nanoclusters and 95 mg PETA), tion of epoxy monomers (EEC) with scanning The 8 wt % Au25 PETA photoresist was used to
square structures can be fabricated at laser speeds up to 100 mm/s (fig. S6). fabricate cylindrical pillars 2.5 mm in radius
power as low as 4 mW and scanning speeds up The Ag28Pt and Au25 nanoclusters show state- and 10 mm in height. Engineering stress-strain
to 100 mm/s (Fig. 1D and figs. S2 and S3). The of-the-art performance comparable to the most curves show a nonlinear loading curve with
fabricated structure shows photoluminescence efficient organic two-photon initiators. The significant hardening at strains above ~20%
with a maximum peak at 720 nm (fig. S4), sim- efficiency of different photoresists can be (fig. S8). This strain hardening is still apparent
ilar to Ag28Pt nanoclusters in solution, which compared using a dimensionless figure of merit when corrected for true stress and strain (fig.
suggests that the nanoclusters remain intact (FOM) (see fig. S7 for details). Kiefer et al. (13) S8). Sudden failure occurs at 1.0 ± 0.2 GPa and
in the printed structure. Lines were fabricated reported that only six types of organic photo- a strain of ~50%, although crack initiation was
with a lateral resolution of ~200 to 250 nm resists have a FOM ≥ 102 at velocities ≥ 10 mm/s observed at ~40% strain in some pillars (fig. S8
(fig. S5). Similar performance was shown by (fig. S7). Our nanocluster-PETA photoresists and movie S3). The elastic modulus was found
8 wt % Au25 photoresists, with threshold power have a FOM ≥ 102 at velocities up to 100 mm/s. to be 3.7 ± 0.2 GPa (fig. S9). Cyclic tests showed
down to 2.5 mW and scanning speed up to The minimum required concentration of the 70% recovery for samples that were loaded
150 mm/s (Fig. 1D). Open-faced tables, octet lat- nanocluster in resin is 0.47 to 2.75 mmol/g (de- to 30% strain (Fig. 2A, fig. S10, and movie
tices, and Schwarz primitive (SP) lattices made tails in SI), which is among the lowest compared S4). The high strength, stiffness, and strain
using Ag28Pt and Au25 photoresists are shown with hundreds of organic two-photon initiators hardening lead to an energy absorption of
in Fig. 1, E to I. Figure 1E shows that freestand- (13, 14). In addition, the nanocluster photo- 110 ± 50 MJ/m3 before initial crack formation.
ing 3D features can be as small as 400 nm. The resists can be applied to more than one class of This mechanical behavior is also apparent in
minimal strut thickness of the octet lattice is monomer, which provides an elegant strategy honeycomb structures fabricated using 10 wt %
1.27 mm (Fig. 1G). The minimal wall thickness for printing high-quality nanocomposites. Ag28Pt PETA photoresist (Fig. 2B). The honey-
of SP lattices is 850 nm (Fig. 1I). Both Au25 and The mechanical performance of the printed combs have a height of 8.2 ± 0.1 mm, cell side
Ag28Pt nanoclusters can also function as two- nanocluster-polymer structures is evaluated length of 2.6 ± 0.2 mm, and wall thickness of
photon acid generators for cationic polymeriza- with in situ SEM compression testing (Fig. 2). 800 ± 50 nm (Fig. 2C). This results in a density

770 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

A B C
1
100
lattices have
no recoverability

100

Specific Compressive Strength (MPa·m3/kg)

Energy Absorption (MJ/m3)

Compressive Stress (MPa)

10 0.1

1
0.01
1

0.1

0.001 0.1
0.1 1 0 25 50 75 0 20 40 60 80 100
Density (g/cm3) Compressive Strain (%) Recovery (%)

This work (pillar) MEA Coated Polymer (27) Polymeric Energy Absorbing Lattices (9) Ceramic Coated Polymer (26)
This work (lattice) HEA Coated Polymer (18) High Strength Ceramic Coated Polymer (17) NiB Coated Polymer (22)

Fig. 3. Comparison of mechanical performance. (A) Energy absorption versus density, (B) specific compressive strength versus compressive strain, and
(C) compressive stress versus recovery of the nanocluster-polymer nanolattices and pillars compared with other micro- and nanolattices (i.e., inorganic coatings
on polymer lattices). Data are from (9, 17, 18, 22, 26, 27).

of 0.56 g/cm3 and a relative density of ~48%. modulus of 24 GPa. This is reasonable given energy absorption capacities up to 7.6 MJ/m3
The compressive stress-strain response shows the ~5 nm average spacing of the nanoclusters. and 9.7 MJ/m3, respectively. In situ SEM com-
a linear region followed by a distinct yielding The interphase volume is roughly three times pression testing reveals that this best-of-class
event at 22.1 ± 2.1 MPa (Fig. 2B). After this, the larger than that of the nanoclusters. We hy- performance is due to the characteristic mate-
stress-strain curve is nonlinear, with signifi- pothesize that the nanocluster composite rial properties of the nanocluster-polymer com-
cant hardening at large strains until ultimate deformation mechanism is similar to that of posite (Fig. 2, D to K, and movies S5 and S6).
failure at ~25% strain. The honeycombs have thermoplastic copolymers (e.g., thermoplastic Low-density cellular lattices typically collapse
a nominal failure stress of 178 ± 52 MPa. In polyurethane) (23) or hybrid materials with in a layer-by-layer manner, which leads to re-
comparison, polymeric honeycombs (e.g., com- interpenetrating networks (24). These ther- gions of negative stiffness in the plateau portion
mercial Ip-Dip) with similar structural density moplastics have similar stress-strain behav- of the stress-strain curve (fig. S11) (18, 22, 25).
generally do not show such significant hard- ior, including strain hardening and hysteresis This results in reduced energy absorption ca-
ening beyond the yield point (22) (Fig. 2B). between loading and unloading. The average pacity. Layer-by-layer collapse and negative
Thus, our nanocluster-polymer structure re- spacing of the nanoclusters is on par with the stiffness are suppressed in the nanocluster-
sults in far better ultimate strength and energy spacing of hard domains in thermoplastic co- polymer nanolattices, which retain high recov-
absorption. polymers. The interactions of the nanoclusters erability of up to 80% even after compression
We examine the origin of the nanocluster may be similar to the alignment of the hard to ~60% strain. Tensegrity structures have
composite mechanical behavior. By using the domains in thermoplastic copolymers, which been constructed to promote uniform defor-
Halpin-Tsai model, we determine that the leads to high strength, stiffness, and significant mation in cellular polymeric lattices (25). We
nanocluster composites are not merely me- strain hardening despite the lack of chemical show that such complex lattice architectures
chanical reinforcements but perturb the bond- bonding. are not always necessary to achieve optimized
ing and physical properties of the polymer Octet and Schwarz primitive (SP) lattices behavior.
matrix in an interphase region (see supple- are fabricated from 8 wt % Au25 photoresist. The specific energy absorption of the
mentary materials for detailed calculations). These structures show high energy absorp- nanocluster-polymer nanolattices and pillars
This interphase region is estimated to have a tion before densification. Octet lattices with is better than polymer micro- and nanolattices
thickness of ~1 nm (0.25 of the volume fraction relative densities of 19 and 27% and SP lattices with inorganic coatings (9, 17, 18, 22, 26, 27), as
of the nanocluster composite) and an elastic with relative densities of 20 and 26% have well as conventional material systems (Fig. 3A).

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 77 1

RES EARCH | R E P O R T S

A B C

surface

D E F
G

H I J

Fig. 4. Hierarchical, tunable, and anisotropic nanoporous structures of shell-solid core pillar. (E) Nanoporous pillar. (F and G) Nanoporous octet lattice. (H
glassy carbon and silk protein. (A) Nanoporous cube on supports. (B) Magnified to J) Silk fibroin honeycomb structures at different length scales. Scale bars: (A) 2 μm, (B)
view of top surface. (C) Cross-sectional image of nanoporous cube. (D) Nanoporous 1 μm, (C) 200 nm, (D) 5 μm (E) 1 μm, (F) 20 μm (G) 2 μm, (H) 10 μm, (I) 3 μm, (J) 500 nm.

The nanocluster composites also have high cube are ~56 ± 23 nm (Fig. 4B and fig. S13) form larger particles and eventually precipi-
specific compressive strength at large strains and ~50%, respectively. The nanopores extend tate out of the structure. The melting temper-
(Fig. 3B) and high recoverability at large ~500 to 700 nm into the structure (Fig. 4C) ature of 1- to 2-nm nanoclusters is estimated
compressive stresses (Fig. 3C). Glassy carbon and decrease in size and density further from to be 600 to 900°C (28), which is consistent
structures, although considerably strong and the surface. Thus, larger structures such as with the observation that nanoporous glassy
stiff (6), are not included in this comparison the ~8 mm diameter pillar shown in Fig. 4D carbon forms only above 550°C (fig. S14). This
because of their brittle and nonrecoverable have graded porosity that consists of a solid is also supported by the observation of ~100- to
behavior. By using nanocluster-based photo- core surrounded by a nanoporous shell. Small 500-nm metal particles on the surface of some
resists, these mechanical properties can be structures <2 mm in size are porous through- pyrolyzed structures and the nearby substrate
achieved in a single printing step, in contrast out the structure (Fig. 4E). Pyrolyzed octet (fig. S15). Energy-dispersive x-ray spectroscopy
to the multiple fabrication steps required for lattices have two levels of porosity: geomet- (EDS) confirmed the presence of Au or Ag in
core-shell composite lattices. rically designed spaces between lattice struts or on the pyrolyzed structures (fig. S16). Nano-
Nanocluster-based photoresists can be used and nanocluster-induced nanopores within porous glassy carbon pillars with a density of
to fabricate complex nanoporous structures lattice struts (Fig. 4, F and G). ~1.07 g/cm3 were found to have a compres-
(Fig. 4). Nanocluster-polymer composites were Printed structures containing fewer nano- sive failure strength of 1.3 ± 0.9 GPa (figs. S17
transformed into glassy carbon with complex clusters result in decreased porosity after pyrol- and S18).
nanoporous features through pyrolysis at 900°C ysis (fig. S14). Previous work involving pyrolysis It is challenging to fabricate hierarchical
under argon flow. The presence of glassy carbon of PETA-based photoresists resulted in glassy structures with anisotropic pores, although
was confirmed by Raman analysis (fig. S12). carbon without porosity (6). We propose that this is a common motif in natural structural
Figure 4A shows a nanoporous cube printed the nanopores result from the melting of nanomaterials (e.g., aligned collagen and pores
using 20 wt % Ag28Pt photoresist. The average clusters during pyrolysis. As the structure is in bone) (10). In addition, the two-photon
pore size and porosity on the surface of the heated to 900°C, the nanoclusters merge to lithography of protein structures is generally

772 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

slow, with reported speeds in the range of 1 to 17. J. Bauer, S. Hengsbach, I. Tesari, R. Schwaiger, O. Kraft, . Proc. Chemical Society Petroleum Research Doctoral New Investigator Grant.
1000 mm/s (29). Here, a nanocluster-protein Natl. Acad. Sci. U.S.A. 111, 2453–2458 (2014). D.D. acknowledges the National Science Foundation Graduate Research
18. X. Zhang et al., Nano Lett. 18, 4247–4256 (2018). Fellowship under grant 1656518. J.K. was supported by a Stanford
photoresist (detailed composition in the SM) 19. R. Jin, C. Zeng, M. Zhou, Y. Chen, Chem. Rev. 116, Graduate Fellowship. Part of this work was performed at the Stanford
is developed to address these issues. This pho- 10346–10413 (2016). Nano Shared Facilities (SNSF) and Stanford Nanofabrication Facility
toresist exploits the efficient singlet oxygen 20. X. Lin et al., Nano Res. 12, 309–314 (2019). (SNF), which are supported by the National Science Foundation under
21. Q. Li, C. J. Zeman4th, Z. Ma, G. C. Schatz, X. W. Gu, Small 17, award ECCS-1542152. Author contributions: Q.L. conceived the
generation and significant photothermal effect 2007992 (2021). initial idea. X.W.G. and Q.L. designed the research. J.K., D.D., and Q.L.
of metal nanoclusters under excitation (30). 22. M. Mieszala et al., Small 13, 1602514 (2017). performed 3D printing. J.K., D.D., and M.M.W. conducted mechanical
This induces photo-crosslinking of proteins 23. H. J. Qi, M. C. Boyce, Mech. Mater. 37, 817–839 (2005). tests. D.D., J.K., and O.A.T. performed structural characterization.
24. B. M. Novak, Adv. Mater. 5, 422–433 (1993). C.J.Z. and G.C.S. performed simulations. Q.L., J.K., D.D., O.A.T., C.J.Z.,
through oxidization of the tyrosine residues 25. J. Bauer, J. A. Kraus, C. Crook, J. J. Rimoli, L. Valdevit, Adv. and X.W.G. wrote the paper, and all authors commented on it.
(Fig. 1B) (31), as well as the formation of di- Mater. 33, 2005647 (2021). Competing interests: The authors declare no competing interests.
rectional β-sheet crystalline regions though 26. A. Schroer, J. M. Wheeler, R. Schwaiger, J. Mater. Res. 33, Data and materials availability: Data are available in the
274–289 (2018). manuscript or the supplementary material or are deposited in
local heating. A water-dispersible Au22 nano-
27. X. Feng et al., Mater. Today 42, 10–16 (2021). Dryad (33). License information: Copyright © 2022 the authors,
cluster (32) was synthesized and used in the 28. F. Gao, Z. Gu, in Handbook of Nanoparticles, M. Aliofkhazraei, some rights reserved; exclusive licensee American Association for
photoresist. Silk fibroin structures are printed Ed. (Springer, 2016), pp. 661–690. the Advancement of Science. No claim to original US government
at speeds up to 100 mm/s (laser power ~25 mW). 29. D. Serien, K. Sugioka, OEA 1, 180008–180018 (2018). works. [Link]
30. Y. Xie, W. Zheng, X. Jiang, ACS Appl. Mater. Interfaces 12, journal-article-reuse
The printed protein structures are composed of 9041–9049 (2020).
aligned bundles (Fig. 4, H to J), which indicates SUPPLEMENTARY MATERIALS
31. X. Mu, J. K. Sahoo, P. Cebe, D. L. Kaplan, Polymers 12, 2936 (2020).
that directional self-assembly occurred during 32. K. Pyo et al., J. Am. Chem. Soc. 137, 8244–8250 (2015). [Link]/doi/10.1126/science.abo6997
33. Q. Li et al., Mechanical nanolattices printed using nanocluster- Materials and Methods
fabrication. Raman spectroscopy confirms the Figs. S1 to S20
based photoresists, Dryad (2022); [Link]
formation of aligned crystalline domains made dryad.ksn02v77x. References (34–39)
of β-sheets (fig. S19). The anisotropic nanovoids Movies S1 to S6

(fig. S20) between these domains may be due to ACKN OWLED GMEN TS Submitted 19 February 2022; resubmitted 12 September 2022
This work was financially supported by the National Science Foundation Accepted 20 October 2022
the destruction of softer parts of the protein
(CMMI-2052251 and DMR-2002936/2002891) and the American 10.1126/science.abo6997
chains by the strong photothermal effect.
In summary, this work develops metal
nanocluster-based photoresists that are used
to fabricate nanocluster-polymer nanolattices, ORGANIC CHEMISTRY
as well as nanoporous glassy carbon and pro-
tein structures with unprecedented structural Halogenation of the 3-position of pyridines through
complexity. We show that nanoclusters are
versatile and highly efficient two-photon ac- Zincke imine intermediates
tivators that are applicable across different
classes of monomers. The nanocluster-polymer Benjamin T. Boyle†, Jeffrey N. Levy†, Louis de Lescure, Robert S. Paton*, Andrew McNally*
nanolattices show strain hardening behav-
ior, which leads to a combination of high Pyridine halogenation reactions are crucial for obtaining the vast array of derivatives required for drug
specific energy absorption, specific strength, and agrochemical development. However, despite more than a century of synthetic endeavors,
deformability, and recoverability. Based on halogenation processes that selectively functionalize the carbonÐhydrogen bond in the 3-position
this simple and general framework, there is a of a broad range of pyridine precursors remain largely elusive. We report a reaction sequence of pyridyl
broad opportunity to directly print additional ring opening, halogenation, and ring closing whereby the acyclic Zincke imine intermediates undergo
mechanical metamaterials by combining the highly regioselective halogenation reactions under mild conditions. Experimental and computational
hundreds of available metal nanoclusters with mechanistic studies indicate that the nature of the halogen electrophile can modify the selectivity-
different types of monomers and rationally determining step. Using this method, we produced a diverse set of 3-halopyridines and demonstrated
designed 3D topologies. late-stage halogenation of complex pharmaceuticals and agrochemicals.

P
RE FE RENCES AND N OT ES
1. S. K. Saha et al., Science 366, 105–109 (2019). yridines are widely found in biologi- (5–10). Halopyridines are also inherently valu-
2. S. Kawata, H.-B. Sun, T. Tanaka, K. Takada, Nature 412, cally relevant molecules, and their prev- able in bioactive compounds such as etoricoxib.
697–698 (2001). alence results from an interplay of the However, despite reports of pyridine halogen-
3. V. Hahn et al., Nat. Photonics 15, 932–938 (2021).
4. Z. C. Eckel et al., Science 351, 58–62 (2016).
heterocycle’s intrinsic properties and ation reactions dating back to the late 19th
5. J. Bauer et al., Adv. Mater. 29, 1701850 (2017). substituents (1–4). Therefore, regiose- century, there are still major limitations to this
6. X. Zhang, A. Vyatskikh, H. Gao, J. R. Greer, X. Li, Proc. Natl. lective pyridine C–H halogenation reactions valuable bond construction (11–13). In partic-
Acad. Sci. U.S.A. 116, 6665–6672 (2019). are vital because installing a carbon–halogen ular, processes selective for the 3-position tend
7. L. R. Meza, S. Das, J. R. Greer, Science 345, 1322–1326
(2014).
(C–Hal) bond enables numerous subsequent to be incompatible with the functionality of many
8. X. Zheng et al., Science 344, 1373–1377 (2014). bond-forming reactions (Fig. 1A). In pharma- substrates of interest for their bioactivity (14).
9. S. Yuan, C. K. Chua, K. Zhou, Adv. Mater. Technol. 4, 1800419 ceutical and agrochemical research, halopyr- Pyridine halogenation reactions using elec-
(2019).
10. O. A. Tertuliano, J. R. Greer, Nat. Mater. 15, 1195–1202 (2016).
idines are key intermediates to diversify candidate trophilic aromatic substitution are electroni-
11. W. Zhou et al., Science 296, 1106–1109 (2002). compounds for structure-activity relationship cally mismatched processes that require harsh
12. L. Li, R. R. Gattass, E. Gershgoren, H. Hwang, J. T. Fourkas, studies as well as target-orientated synthesis conditions (15, 16). Specifically, elemental halides
Science 324, 910–913 (2009).
13. P. Kiefer et al., Adv. Opt. Mater. 8, 2000895 (2020).
with strong Brønsted or Lewis acids are often
14. T. Wloka, M. Gottschaldt, U. S. Schubert, Chem. Eur. J. 28, used at elevated temperatures to compensate
e202104191 (2022). Department of Chemistry, Colorado State University, Fort for the poor p nucleophilicity of pyridine rings
15. M. Carlotti, V. Mattoli, Small 15, 1902687 (2019). Collins, CO 80523, USA. (5, 6). Substrate scope is limited, and despite
16. C. M. Spadaccini, in Three-Dimensional Microfabrication Using *Corresponding author. Email: [Link]@[Link] (A.M.);
Two-Photon Polymerization (Second Edition), T. Baldacchini, [Link]@[Link] (R.S.P.) being 3-selective, reactions often result in re-
Ed. (William Andrew Publishing, 2020), Chapter 13.4 †These authors contributed equally to this work. gioisomeric mixtures. Metalation-halogenation

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 773

RES EARCH | R E P O R T S

A Value of halogenated pyridines in pharmaceutical and agrochemical development

NHSO2n-Pr MeO2S
F
Hal C–C bond
Cl O Cl
Me C–N bond
N N
N F N N
F C–O bond
N N Me
Me H
C–S bond
Diversification of drug and agrochemical vemurafenib etoricoxib
candidates Bond formation via C–Hal Inherent function in bioactive molecules

B Pyridine C3-halogenation via Zincke imine intermediates C Zincke ring-opening study under mild reaction conditions

H Hal
TfN NR’2
R Hal+ R
N N Ph N Ph
1 2-6
Pyridines temporarily assume from HNR’2 :*
the reactivity of electron-rich
O
aromatics
i-Pr N i-Pr Ph N Ph
N H H
Hal H N N
TfN NR’2 Hal+ H H
TfN NR’2
2 3 4 5 6
R
R 40% 44% 62% 51% 84% (86% isolated yield)

Fig. 1. Importance of pyridine halogenation reactions and a distinct strategy nucleophiles. *Yields calculated from integrated 1H NMR spectra of the crude reaction
based on ring-opened intermediates. (A) Examples of pyridine halogenation mixture using triphenylmethane as an internal standard. R, general organic group;
products and derivatives in drug and agrochemical development. (B) A ring-opening, Me, methyl; Hal, halogen; Tf, trifluoromethylsulfonyl; collidine, 2,4,6-trimethylpyridine;
halogenation, ring-closing synthetic strategy. (C) Ring-opening study using amine Et, ethyl; Ac, acetate; Ph, phenyl; i-Pr, isopropyl; rt, room temperature.

sequences with strong bases are another ap- itation excludes many classes of substituted NTf-Zincke imine derived from pyridine and a
proach for pyridine halogenation, but most re- pyridines. To address these restrictions, we fo- pyridinium electrophile; however, there were
quire directing groups to access the 3-position cused on ring opening of NTf-pyridinium salts. no investigations into Zincke imines derived
reliably (17–19). Beyond these two approaches, These reactive intermediates are readily formed from substituted pyridines (30). When we sub-
there has been little progress over the past from the pyridine and Tf2O at low temperatures jected 6 to N-iodosuccinimide (NIS) at room
century, so the synthetic community turned and encompass most substitution patterns ex- temperature for 5 min, we observed 7-I in 92%
to other versatile functional groups in place of cept 2,6-disubstitution. Toscano et al. reported yield and >20:1 3- versus 5-selectivity (8-I)
halides. Iridium-catalyzed borylations and sily- a ring opening with Tf2O, but did not extend based on analysis of the crude 1H nuclear
lations are the most notable processes and are this process beyond pyridine, and observed magnetic resonance (NMR) spectrum. Using
3-selective for certain classes of substituted mixtures of ring-opened products (30). With N-bromosuccinimide (NBS), we observed a
pyridines or through control with ligand ar- 2-phenylpyridine (1), we surveyed a series of 4.4:1 ratio of 7-Br and 8-Br at room temper-
chitectures (20–26). Here, we present an alter- aliphatic amines as nucleophiles (Fig. 1C). Pyr- ature and >20:1 selectivity and 92% of 7-Br
native approach for pyridine halogenation using rolidine, piperidine, and morpholine gave mod- at –78°C in CH2Cl2. Chlorinating with N-
a ring-opening, halogenation, ring-closing strat- erate yields of ring-opened products, as did chlorosuccinimide (NCS) was lower yielding
egy (Fig. 1B). This “one-pot” protocol uses a diisobutylamine (2 to 5). However, dibenzyl- and less selective (7-Cl and 8-Cl), and it de-
modified version of the classic Zincke ring- amine proved optimal, with 6 formed in high composed 6. At this point, we tested condi-
opening reaction that converts pyridines into yield. Tables S1 and S2 show further studies of tions for ring closure and found that heating
azatriene intermediates, or “Zincke imines.” pyridine ring openings using this protocol. We 8-I in a EtOAc/EtOH solvent mixture at 60°C
This synthetic maneuver temporarily trans- anticipated that the ring closure of Zincke with 10 equivalents of ammonium acetate
forms pyridines from electron-deficient het- imines such as 6 would be straightforward smoothly formed 3-iodopyridine 9-I. Table S10
erocycles to a series of polarized alkenes that based on precedent in related Zincke systems shows further studies of this ring-closing step.
that undergo electrophilic substitution reac- (31). Examples typically use ammonium salts or Halogenating 3-substituted imines (Fig. 2B)
tions, much like electron-rich aromatics. The Brønsted acids to obtain pyridines, and these required a different procedure. Ring opening of
halogenation process is 3-selective across a reagents could potentially facilitate the one- 3-phenylpyridine produced Zincke imine 10 in
range of substituted pyridines and is func- pot sequence outlined in Fig. 1B (see below). 82% yield, but NIS did not react with 10 at room
tionally compatible with complex structures of For the next stage of reaction development, temperature and decomposed to unidentified
pharmaceutical interest. we tested Zincke imine 6 with halogen electro- by-products when heated to 50°C. However,
Current methods for Zincke ring-opening philes (Fig. 2A). At the outset, we did not when we included two equivalents of trifluoro-
chemistry cannot enable the general process expect that C–Hal bond formation would be acetic acid (TFA), iodopyridine 12-I formed in
for pyridine halogenation described in Fig. 1B selective because there was no clear electronic high yield (33). We attribute this in situ re-
(27–29). The pyridine N-activation step re- bias between the two d– sites at C3 and C5 (see aromatization to steric interactions between the
quires forcing conditions and often fails when below) in 6 (32). Toscano et al. did observe a 3- and 5-substituents in 11-I; TFA can promote
2-position substituents are present. This lim- 3-selective reaction between an unsubstituted the alkene isomerization-recyclization sequence,

774 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

A B
Hal Hal Ph Hal
3 5
TfN Ph Hal
TfN NBn2 TfN NBn2 NBn2
TfN NBn2
Ph
Ph Ph N
from 2-Ph from 3-Ph
7-I/Br/Cl 8-I/Br/Cl pyridine: pyridine: 11-I/Br/Cl 12-I/Br/Cl
6, 86% 10, 82%

NXS / conditions 7-I/Br/Cl (%)* 8-I/Br/Cl (%)* NXS / conditions 11-I/Br/Cl (%)* 12-l/Br/Cl (%)*

NIS (5 mins) 92 0 NIS 0 0

Hal
NBS 75 17 NIS + TFA (2.0 eq) 0 90
NBS, –78 ºC, CH2Cl2 92 1 NBS + TFA (2.0 eq) 0 76
Ph N
Hal = I, 9-I, 90%* NCS 6 21 NCS (1.2 eq) + HCl (4.0 eq), CH2Cl2 0 76

C One-pot protocols

Hal Ph Ph Hal
Hal = I, 9-I, 68% Hal = I, 12-I, 82%
Hal = Br, 9-Br, 58% N Hal = Br, 12-Br, 81%
Pyridines without 3-substituents Pyridines with 3-substituents Hal = Cl, 12-Cl, 79%
Ph N N
(2-Ph or 3-Ph)

D chlorination with NCS C3 C5 bromination with NBS iodination with NIS

G‡ = 3.4
G‡ = 0.4 G‡ = 2.6
G‡= 4.3
G‡ = 3.9 Int-I TS-II
TS-I 19.7 22.2
22.4 TS-I
19.3 Int-I TS-II
13.7 15.2
TS-II
Int-I
10.2
9.0
imine imine imine
0.0 0.0 0.0
- early halogenation TS product - late halogenation TS product - change in selectivity-determining step product
- lower regioselectivity -29.0 - highly C3 regioselective -22.8 - regio-determining TS-II -13.1

dNH 1.50
E selectivity-determining
Chlorination halogenation F selectivity-determining deprotonation dCH 1.25
O TS-I-C3-NBS TS-I-C5-NBS O
Fukui f – 0.24 0.25
N G‡ = 19.3 G‡ = 21.9 N
O O
TfN NBn2 Br H H Br O C3
TfN NMe2 TfN NMe2 N H R
Ph O TS-II-C3-NIS
Natural H I
-0.39 -0.44
charge Ph Ph G‡ = 22.2
G‡ = 2.6 TfN N R
Negligible difference in Fukui f-/ G‡ = 3.4
dNH 1.45
charge at C3 vs C5 Ph dCH 1.26

Unfavorable A1,3 strain develops

dNBr 2.25 dNBr 2.40
during C5-deprotonation in TS-II C5
dCBr 2.14 dCBr 2.07
dNCl 2.06
TS-II-C5-NIS
C3
dCCl 2.12 C5 G‡ = 25.6

Edist = 13.4 Edist = 18.0

G H/D H Hal H
TS-I-C3-NCS TfN NBn2
TfN NBn2
Edist = 9.8 (C3) vs. 9.4 (C5)
Ph Ph
Chlorination TS-I is early: smaller G‡ Bromination TS-I is later: larger G‡ reflects greater stability of the
NIS: [ 2H]-6/6 = 2.70
reflects substrate’s small electronic bias a, -unsaturated iminium for C3-halogenation 1 eq. 6, 1 eq [2H]-6
NBS: [ 2H]-6/6 = 1.06

Fig. 2. Reaction development and mechanistic investigation. (A) Halogena- for pyridines with 3-substituents: NIS (1.0 equiv) or NBS (1.0 equiv) and TFA
tion of 2-substituted Zincke imines and ring closure using ammonium salts. (2.0 equiv). NCS (1.0 equiv) and HCl (4.0 equiv) and ring opening were performed
*Yields calculated from 1H NMR of the crude reaction mixture using in CH2Cl2. Isolated yields are shown. (D) Computed Gibbs energy profiles (in
triphenylmethane as an internal standard. (B) Halogenation of 3-substituted kilocalories per mole) for imine halogenation by N-halosuccinimides. (E) Analysis
Zincke imines. (C) One-pot pyridine halogenation protocols. †Ring-opening of selectivity-determining chlorination and bromination transition structures
conditions: Tf2O (1.0 equiv), HNBn2 (1.2 equiv), collidine (1.0 equiv), EtOAc, –78°C, TS-I. (F) Analysis of selectivity-determining deprotonation transition structure
30 min then warm to rt. ‡Ring-opening and halogenation steps performed in TS-II for iodination. (G). Isotope competition experiment. Bn, benzyl group;
CH2Cl2 instead of EtOAc when using NBS. §Halogenation–ring-closure protocol NXS, N-halosuccinimides.

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 775

RES EARCH | R E P O R T S

R Hal
R Ph N Ph NBn2 R
H TfN
N N
NTf-Zincke imine (not isolated) Halogenated
Pyridine Dibenzylamine
pyridine

I I OCF3
Hal I n-Bu I F I CF3
I
O
N N
Ph N N F N N
O
X N
X = CF3, 13, 65% 15, Hal = I, 66% 20, two steps:
16, Hal = Br, 66% 17, 50% 18, 74% 19, 82% 21, 65%‡
X = BPin, 14, 68%‡ (52%), 82%‡

Bn O Ph O OMe O CF3
CN
Hal
Hal I I I Me I I
I OEt
N
N N N Me N Cl Me N Ph N
N
two steps: (80%)
22, Hal = I, 66% 24, 36% 25, 62% 26, 41% 27, 44% 30, 44% 31, 71%
23, Hal = Br, 52% 28, Hal = I, 83%
29, Hal = Br, 42%

Cl Br I Ph OEt
NO2 Br TsN
Hal OBn Hal
I Ph I F I F N Br Br

Me N N N
N N N N N
32, Hal = I, 67% 38, Hal = I, 53% 41, two steps:
33, Hal = Br, 69% 35, 57% 36, 39% 37, 25% 39, Hal = Br, 44% 42, 36% 43, 35%
(55%), 85%
34, Hal = Cl, 58% 40, Hal = Cl, 28%

I S O Cl
I OPh I Ts
O I N
I I I S
OMe I Ph
N
S N N N
N Me N N
N
44, 76% 45, 68% 46, 23% 47, 62% 48, 64% 49, 74% 50, 74% 51, 73%

from 7-I, (80%) from 7-Br, (80%) from 6, (86%)

I I I Br I Cl Br I Br Br Br Cl Cl Cl

Ph N Ph N Ph N Ph N Ph N Ph N Ph N

52, 85% 53, 85% 54, 80% 55, 76% 56, 63% 57, 71% 58, 45%

Fig. 3. One-pot pyridine halogenation: scope of building-block pyridines ring-closure protocol for pyridines with 3-substituents: NIS (1.0 equiv) or NBS
and dihalogenation reactions. Yields of isolated products are shown. Yields in (1.0 equiv) and TFA (2.0 equiv). NCS (1.0 equiv) and HCl (4.0 equiv), and ring
parentheses are those of isolated Zincke imines. *Ring-opening and halogenation opening is performed in CH2Cl2. ‡1H NMR yields using triphenylmethane as an
steps performed in CH2Cl2 instead of EtOAc when using NBS. †Halogenation– internal standard. Pin, pinacolato; n-Bu, normal butyl group; Ts, tosyl.

as well as the subsequent elimination of di- pot process (Fig. 2C). For pyridines without the one-pot process (12-I, 12-Br, and 12-Cl).
benzylamine and N–S bond cleavage of the 3-substituents, such as 2-phenylpyridine, ring These protocols do not require intermediate
resulting NTf-pyridinium salt to reform the opening using dibenzylamine and halogena- workup or purification steps and involve add-
pyridine ring system. Brominating with NBS tion with NIS or NBS (Fig. 2A) was appropri- ing reagents sequentially to the same reac-
and TFA (12-Br) resulted in the same outcome, ate. Adding NH4OAc and EtOH to the reaction tion vessel.
and a modified procedure using NCS with HCl mixture and heating to 60°C induced ring clo- We then studied the mechanism and re-
in CH2Cl2 chlorinated 10 effectively to form sure to reform pyridines 9-I and 9-Br. Adding gioselectivity of Zincke imine halogenation by
12-Cl. The results (Fig. 2, A and B) provide an equivalent of trimethoxybenzene after the N-halosuccinimides with quantum chemistry
two halogenation protocols to halogenate NTf- halogenation step proved helpful in quenching at the B3LYP-D3(BJ)/def2-TZVP//ωB97X-D/
Zincke imine intermediates depending on their any remaining N-halosuccinimide before cy- 6-31+G(d,p) level of theory with a “solvation
substitution pattern; further details of this opti- clization. When 3-substituents were present, model based on density (SMD)” description of
mization study are described in tables S3 to S8. such as in 3-phenylpyridine, ring opening fol- CH2Cl2. The most favorable pathway for halo-
Next, we combined the ring-opening, halo- lowed by addition of the N-halosuccinimides genating 6 (Fig. 2D), with the NBn2 portion
genation, and ring-closing steps into a one- and acids described in Fig. 2B comprised simplified to NMe2, consisted of electrophilic

776 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

R Hal
R Ph N Ph NBn2 R
H TfN
N †
N
NTf-Zincke imine (not isolated) Halogenated
Pyridine Dibenzylamine
pyridine

F
I Cl
I
MeO O
Boc N
Hal N N Me
OMe S O
N
F F Hal
N N
Me EtO2C
59, Hal = I, 70% 61, 51% 62, Hal = I, 79% N 64, 53%
60, Hal = Br, 68% 63, Hal = Br = 68%
Cl
O
S
O Me
O N N
S Hal N
N O
MeO N CF3 F
F F Hal
O Hal
Br N
Br
N Cl
N F
N
F3C N 67, Hal = I, 72% 69, Hal = I, 71% 71, Hal = I, 75% 1:1 d.r.
65, 36% 66, 53% 68, Hal = Br = 73% 70, Hal = Br = 68% 72, Hal = Br, 60% 1:1 d.r.

I AcO I Hal
O
H
O N I O
N Cl I
N O n-Bu N N
Cl Me
Cl N from loratadine N
zincke imine (52%)
from vismodegib from bisacodyl from nicoboxil N
76, Hal = I, 63% from nicotine
73, 43% 74, 56% 75, 59% 77, Hal = Br, 61%, 79, 51%
SO2Me OAc EtO2C 78, Hal = Cl, 43%

SO2Me F
I Cl
O Me H
Cl Cl O O N
I N
N N
I Hal
Boc N Cl
N Cl
N
N Me Cl N
F
from quinoxyfen SO2Me
from Boc-niaprazine from etoricoxib from vismodegib
82, Hal = I, 49%,
80, 34% 81, 43%
83, Hal = Br, 54% 84, two steps, (50%)‡, 74%

Fig. 4. One-pot halogenation of complex azines. Yields of isolated with 3-substituents: NIS (1.0 equiv) or NBS (1.0 equiv) and TFA (2.0 equiv).
products are shown. Yields in parentheses are those of isolated Zincke NCS (1.0 equiv) and HCl (4.0 equiv), and ring opening is performed in
imines. *Ring-opening and halogenation steps performed in CH2Cl2 instead of CH2Cl2. ‡Isolated yield of iodinated Zincke imine. Boc, tert-butoxycarbonyl
EtOAc when using NBS. †Halogenation–ring-closure protocol for pyridines group.

addition (TS-I) followed by deprotonation for 6 reacting with NCS, NBS, and NIS sug- The higher C3 selectivity for bromination over
(TS-II). Calculations showed that outer-sphere gest an irreversible overall reaction governed chlorination (DDG‡ of 2.5 and 0.4 kcal mol−1,
electron transfer processes were inaccessible by kinetically controlled regioselectivity in each respectively) can be understood on the basis of
(DG > 34 kcal mol−1); figs. S11 to S23 provide case, with DDG‡ values in quantitative agree- early versus late transition state arguments.
full computational details. ment with experiment. Breaking the extended Chlorination results in a more stable interme-
Regioselectivity in electrophilic halogena- conjugation in the Zincke imine to form posi- diate than bromination and, in accordance
tion is conventionally understood in terms of tively charged s-complex Int-I is endergonic: with the Hammond postulate, proceeds through
differences in frontier orbital coefficients, atomic Grel values of this high-energy intermediate an earlier TS [as shown by smaller TS distor-
charges, or nucleophilicity parameters (34). increase for C–Cl, C–Br, and C–I formation, tion energies in Fig. 2E and smaller pyramid-
However, the electronic environments at the respectively (9.0, 13.7, and 19.7 kcal mol−1, realization values at the reacting carbon atom
C3 and C5 positions of Zincke imine 6 are very spectively). These differences give rise to (fig. S19)]. Given the electronic similarity be-
similar in terms of Fukui f- coefficients (0.24 ver- two distinct selectivity-determining regimes: tween the C3 and C5 positions in the Zincke
sus 0.25), natural charges (–0.39 versus –0.44), an irreversible C–Hal bond-forming step (TS-I) imine and the absence of appreciable inter-
and HOMO coefficients (both 0.26), neces- determines regioselectivity for chlorination molecular noncovalent interactions (fig. S17),
sitating an alternative rationale for the high and bromination, whereas C–I bond formation there is little regioisomeric preference in the
levels of C3 selectivity in this transformation is reversible, and the second deprotonation early chlorination TS. Bromination proceeds
(Fig. 2E). Computed energy profiles (Fig. 2D) step (TS-II) dictates the regioselectivity (35). through a later TS that incurs larger substrate

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 777

RES EARCH | R E P O R T S

distortion and strongly favors addition at C3 to NBS then TFA resulted in 4-brominated heterocycle in each case (43). Pyridines 71
(DDEdist = 4.6 kcal mol−1). This difference in product 41. We observed a similar intermediate and 72 formed atropisomers during the one-
distortion energies reflects a more stable deve- in a one-pot process using 5-azaindole, resulting pot halogenation process, raising the possibility
loping a,b–unsaturated iminium ion in TS-I- in bromide 42. A C6-brominated product 43 of controlling axial chirality in further studies.
C3-NBS versus TS-I-C5-NBS (DDG‡ = 2.6 kcal formed from 4-ethoxyquinoline; it is less clear This protocol successfully halogenated a
mol−1). By contrast, the high regioselectivity what intermediate species operates in this case diverse collection of pharmaceuticals and agro-
of iodination (DDG‡ = 3.4 kcal mol−1) results because of broadening in the 1H NMR spectra chemicals, demonstrating its utility for late-stage
from a rate- and selectivity-determining depro- after ring opening. Figure S2 shows proposed functionalization. We obtained moderate yields
tonation (TS-II). Restoring planarity in TS-II- halogenation mechanisms and spectroscopic of iodinated vismodegib, bisacodyl, and nico-
C5-NIS results in increased A1,3 strain between evidence of the dearomatized adducts. boxil (73 to 75). Loratadine proved challenging
the enamine carbon substituent and the Next, we tested whether halogenating Zincke in the one-pot process, but we could synthe-
iodide (Fig. 2F). The bond and dihedral angles imine intermediates could outcompete neigh- size iodo-, bromo-, and chloro-analogs when we
in the C5-iodinated product clearly reflect this boring electron-rich aromatic rings and if isolated the intermediate Zincke imine (76 to
developing strain (fig. S21). dihalogenation processes were feasible. As 78). Nicotine, Boc-protected niaprazine, etor-
Intermolecular competition experiments vali- shown in Fig. 3, a pyridine is selectively iodi- icoxib, and quinoxyfen were competent sub-
date these computational predictions (Fig. 2G). nated over an anisole group (44), as evidenced strates (79 to 83). The dihalogenation sequence
We reacted one equivalent each of 6 and [2H]-6 by only traces of other compounds, aside from is also amenable to complex pyridines. Using
until this mixture consumed one equivalent of starting material, in the crude 1H NMR. As the basal cell carcinoma drug vismodegib, we
NIS. The [2H]-6/6 ratio of 2.70 measured at the shown in fig. S10, we applied standard arene obtained iodo-chloro analog 84 in reasonable
end of the reaction indicates that 6 is consumed iodination conditions (NIS and TFA in MeCN yield over two steps.
preferentially and supports the selectivity- at 50°C) and observed iodination on the ani- Given the lack of alternative methods for
determining deprotonation predicted by theory sole ring. Similarly, pyridines are iodinated over this transformation and the high value of the
(Fig. 2F). The same experiment using NBS did thiophenes, furans, and phenoxy groups (45 to halopyridine products, we anticipate that this
not show a significant bias between [2H]-6 48). Furthermore, iodination selectively occurs strategy will be useful for developing medicines
and 6 at the end of the reaction, consistent with on the pyridine in fused ring systems such as and agrochemicals. This approach for pyridine
selectivity-determining C–Br bond formation. furopyridines, protected azaindoles, and thie- functionalization is likely amenable to numer-
Next, we applied a series of azine building nopyridines that conventionally react at the ous other bond constructions as well.
block-type compounds to the one-pot halogena- five-membered heterocyclic portion (49 to 51)
REFERENCES AND NOTES
tion reaction (Fig. 3), selecting iodination for (36). C–Hal bond formation can also be applied
1. E. Vitaku, D. T. Smith, J. T. Njardarson, J. Med. Chem. 57,
most substrates because of its convenient room sequentially to form 3,5-dihalogenated pyridines, 10257–10274 (2014).
temperature C–I bond-forming step. The process a motif with two orthogonal vectors for further 2. M. Baumann, I. R. Baxendale, Beilstein J. Org. Chem. 9,
is compatible with pyridines with aryl, alkyl, synthetic transformations (37–39). Although a 2265–2319 (2013).
3. V. V. Zakharychev, A. V. Kuzenkov, A. M. Martsynkevich, Chem.
and acetal substituents at the 2-position (13 to one-pot procedure was unsuccessful, we were Heterocycl. Compd. 56, 1491–1516 (2020).
17), and the acid-mediated conditions were able to isolate Zincke imines 7-I and 7-Br and 4. M. Stolar, T. Baumgartner, Chem. Commun. 54, 3311–3322 (2018).
5. J. A. Joule, K. Mills, Heterocyclic Chemistry (Wiley, ed. 5, 2009).
effective for a series of pyridines (18 to 21). The subject them to a second round of halogenation, 6. M. R. Grimmett, Adv. Heterocycl. Chem. 58, 271–345 (1993).
one-pot iodination yield of 3-fluoropyridine resulting in various permutations of dihalopyr- 7. K. Murakami, S. Yamada, T. Kaneda, K. Itami, Chem. Rev. 117,
(20) was low, and we reverted to a two-step idines 52 to 57. We formed dichloropyridine 58 9302–9332 (2017).
8. M. L. Crawley, B. M. Trost, Applications of Transition Metal
process, isolating the intermediate Zincke imine. directly from Zincke imine 6 by modifying the Catalysis in Drug Discovery and Development: An Industrial
We recommend this alternative protocol to ob- acid-mediated halogenation-rearomatization Perspective (Wiley, 2012).
9. J. F. Bunnett, R. E. Zahler, Chem. Rev. 49, 273–412 (1951).
tain higher overall yields of halogenated pyr- protocol to include 2.05 equivalents of NCS. 10. W. F. Bailey, J. J. Patricia, J. Organomet. Chem. 352, 1–46 (1988).
idines in these instances. The final stage of this study focused on 11. A. W. Hoffmann, Ber. Dtsch. Chem. Ges. 12, 984–990 (1879).
Pyridines with benzyl, cyano, and ketone halogenating complex drug-like intermediates, 12. W. J. Sell, F. W. Dootson, J. Chem. Soc. Trans. 73, 442–445
(1898).
substituents at the 4-position are also effective pharmaceuticals, and agrochemicals (Fig. 13. W. J. Sell, F. W. Dootson, J. Chem. Soc. Trans. 73, 432–441 (1898).
substrates (22 to 25). We then halogenated a 4). These structures often contain multiple 14. P. S. Fier, J. F. Hartwig, Science 342, 956–960 (2013).
15. G. A. Olah, Acc. Chem. Res. 4, 240–248 (1971).
series of 2,3-, 2,4-, and 2,5-disubstituted pyri- (hetero)arenes, sites of reactivity, and Lewis 16. B. Galabov, D. Nalbantova, Pv. Schleyer, H. F. Schaefer3rd, Acc.
dines (26 to 34). Pyridine 27 is notable be- basic atoms, thus representing a significant Chem. Res. 49, 1191–1199 (2016).
cause of the sensitive 2-chloro substituent; synthetic challenge for the ring-opening, halo- 17. S. M. Manolikakes, N. M. Barl, C. Samann, P. Knochel,
Z. Naturforsch. B. J. Chem. Sci. 68, 411–422 (2013).
in this case, ring opening with dibenzylamine genation, ring-closing process. However, success 18. G. A. El-Hiti, K. Smith, A. S. Hegazy, M. B. Alshammari,
resulted in a complex mixture of products. in this endeavor would enable subsequent con- A. M. Masmali, ChemInform 46, 1 (2015).
19. G. A. El-Hiti, K. Smith, A. S. Hegazy, Heterocycles 91, 479–504
However, the reaction was successful using vergent coupling reactions, compound diver- (2015).
N-benzylaniline, and we attribute this out- sification, and late-stage functionalization efforts 20. M. A. Larsen, J. F. Hartwig, J. Am. Chem. Soc. 136, 4287–4299
come to its attenuated nucleophilicity. Halo- that are not possible with existing pyridine (2014).
21. S. A. Sadler et al., Org. Biomol. Chem. 12, 7318–7327 (2014).
pyridines 32 to 34 show that the Zincke imine halogenation methods (40–42). We began our 22. J. S. Wright, P. J. H. Scott, P. G. Steel, Angew. Chem. Int. Ed.
intermediate overrides the ortho-, para-directing study by halogenating five polycyclic struc- 60, 2796–2821 (2021).
23. L. Yang, N. Uemura, Y. Nakao, J. Am. Chem. Soc. 141,
benzyloxy group and maintains 3-selectivity. We tures representative of drug-like intermediates
7972–7979 (2019).
formed a set of di- and trihalogenated pyridines and obtained reasonable yields of halogenated 24. J. Trouvé, P. Zardi, S. Al-Shehimy, T. Roisnel, R. Gramage-Doria,
from 3,4-disubstituted systems (35 to 37), and products as single regioisomers in these cases Angew. Chem. Int. Ed. 60, 18006–18013 (2021).
25. C. Cheng, J. F. Hartwig, J. Am. Chem. Soc. 137, 592–595 (2015).
38 to 40 are preliminary examples showing (59 to 65). A substituted quinoline containing 26. X. Y. Zhou, M. Zhang, Z. Liu, J. H. He, X. C. Wang, J. Am. Chem.
that this process can halogenate pyrimidines. a cyclic sultam was again brominated at C6 Soc. 144, 14463–14470 (2022).
When using 5-nitroisoquinoline as a substrate, (66). Halopyridines 67 to 70 are notable be- 27. C. D. Vanderwal, J. Org. Chem. 76, 9555–9567 (2011).
28. W. C. Cheng, M. J. Kurth, Org. Prep. Proced. Int. 34, 585–608
we did not observe ring opening but instead cause C–Hal bond formation occurs on one (2002).
isolated a dearomatized adduct where diben- pyridine over another; we believe that this 29. S. Sowmiah, J. M. S. S. Esperanca, L. P. N. Rebelo,
C. A. M. Afonso, Org. Chem. Front. 5, 453–493 (2018).
zylamine adds to the C1 atom of the NTf- selectivity results from exclusive NTf-pyridinium 30. R. A. Toscano et al., Chem. Pharm. Bull. (Tokyo) 45, 957–961 (1997).
isoquinolinium salt. Subjecting this intermediate salt formation involving the less hindered 31. J. Becher, Synthesis 1980, 589–612 (1980).

778 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

32. U. Stämpfli, M. Neuenschwander, Helv. Chim. Acta 66, ACKN OWLED GMEN TS absolute thermochemical data for all stationary points in
1427–1435 (1983). computational studies are available in data S1. License
Funding: This work was supported by Colorado State University
33. T. M. Nguyen, M. del Rayo Sanchez-Salvatori, J. C. Wypych, information: Copyright © 2022 the authors, some rights reserved;
with funds from the National Institutes of Health (NIH grant
C. Marazano, J. Org. Chem. 72, 5916–5919 (2007). exclusive licensee American Association for the Advancement of
34. A. Tomberg, M. J. Johansson, P. O. Norrby, J. Org. Chem. 84, R01 GM144591-01). R.S.P. acknowledges support from the
Science. No claim to original US government works. [Link]
4695–4703 (2019). National Science Foundation (NSF grant CHE-1955876); use
[Link]/about/science-licenses-journal-article-reuse
35. M. Liljenberg, J. H. Stenlid, T. Brinck, J. Phys. Chem. A 122, of computational resources from the RMACC Summit
3270–3279 (2018). Supercomputer, which is funded by NSF grants ACI-1532235 and
36. L. H. Klemm, J. N. Louris, J. Heterocycl. Chem. 21, 785–789 (1984). ACI-1532236, the University of Colorado Boulder, and Colorado
State University; and support from XSEDE through allocation SUPPLEMENTARY MATERIALS
37. E. K. Reeves, E. D. Entz, S. R. Neufeldt, Chemistry 27,
6161–6177 (2021). TG-CHE180056. Author contributions: B.T.B. and J.N.L. [Link]/doi/10.1126/science.add8980
38. M. H. Keylor, Z. L. Niemeyer, M. S. Sigman, K. L. Tan, J. Am. performed the experimental work. A.M., B.T.B., and J.N.L. Materials and Methods
Chem. Soc. 139, 10613–10616 (2017). conceptualized the work. R.S.P. and L.d.L. performed the Figs. S1 to S23
39. J. Ji, T. Li, W. H. Bunnelle, Org. Lett. 5, 4611–4614 (2003). computational studies. All authors contributed to the design of Tables S1 to S12
40. T. Cernak, K. D. Dykstra, S. Tyagarajan, P. Vachal, S. W. Krska, the experimental and computational work, performed data NMR Spectra
Chem. Soc. Rev. 45, 546–576 (2016). analysis, discussed the results, and commented on the References (44Ð74)
41. L. Zhang, T. Ritter, J. Am. Chem. Soc. 144, 2399–2414 (2022). manuscript. A.M. and R.S.P. wrote the manuscript. Competing Data S1
42. D. C. Blakemore et al., Nat. Chem. 10, 383–394 (2018). interests: The authors declare no competing interests. Data
43. R. D. Dolewski, P. J. Fricke, A. McNally, J. Am. Chem. Soc. 140, and materials availability: All data are available in the main text Submitted 12 July 2022; accepted 7 October 2022
8020–8026 (2018). or the supplementary materials. Cartesian coordinates and 10.1126/science.add8980

ORGANIC CHEMISTRY transformed into activated electron-rich inter-

Radical and ionic meta-CÐH functionalization mediates. The ensuing electrophilic reactions
and subsequent rearomatization provide ex-

of pyridines, quinolines, and isoquinolines clusively meta-substituted heteroarenes (Fig.

1C) (26–31). Several strategies for the reduc-
tive dearomatization of pyridines have been
Hui Cao†, Qiang Cheng†, Armido Studer* developed along those lines (32, 33). However,
most of them generate either stable enamines
Carbon-hydrogen (C−H) functionalization of pyridines is a powerful tool for the rapid construction and so that rearomatization becomes energetically
derivatization of many agrochemicals, pharmaceuticals, and materials. Because of the inherent unfavorable, or alternatively very unstable in-
electronic properties of pyridines, selective meta-C−H functionalization is challenging. Here, we present termediate N-silyl or N-boryl species that con-
a protocol for highly regioselective meta-C−H trifluoromethylation, perfluoroalkylation, chlorination, strain applicable functionalization reactions.
bromination, iodination, nitration, sulfanylation, and selenylation of pyridines through a redox-neutral For example, the interrupted reductive dearo-
dearomatization-rearomatization process. The introduced dearomative activation mode provides a matization of pyridiniums allowed for sub-
diversification platform for meta-selective reactions on pyridines and other azaarenes through radical sequent selective C3-functionalization, but
as well as ionic pathways. The broad scope and high selectivity of these catalyst-free reactions render tetrahydropyridines instead of the aromat-
these processes applicable for late-stage functionalization of drugs. ized heteroarenes were formed as the final
products (34). meta-Selective silylation, alkyl-

P
ation, and trifluoromethylthiolation have been
yridines and their derivatives are among documented examples (12). In particular, meta- developed through 1,4-reduction of pyridines
the most frequently occurring hetero- selective processes proceeding in both radical by silanes or boranes followed by in situ func-
arenes in pharmaceuticals and agro- and ionic reactivity modes are unavailable. tionalization and oxidative rearomatization
chemicals that contain N-heterocycles Although electrophilic aromatic substitu- (28–30). Nevertheless, introduction of the tri-
(Fig. 1A) (1, 2). They also play prominent tions have been applied to meta-selective ha- fluoromethyl and halogen groups to these re-
roles in ligands as well as functional materials logenation and nitration of pyridines, the harsh ductively dearomatized intermediates remains
(3, 4). Selective modification of pyridines with- acidic conditions limit their usage and fre- unachieved, most likely because of the sen-
out preinstalled transformable functional groups quently lead to regioisomeric products (13, 14). sitivity of these intermediates to oxidation in
can substantially increase step economy for Milder meta-functionalization reactions have the presence of electrophilic reagents (35). To
the preparation of related, more elaborate com- been developed in recent decades, mainly through date, none of the reported methods provide
pounds (5). The known methods for carbon- transition-metal catalysis (Fig. 1B) (15). On access to stable dearomatized intermediates of
hydrogen (C–H) functionalization of pyridines the basis of steric control, iridium-catalyzed pyridines that possess enough redox stability
mainly rely on the electronically biased reac- C–H borylation and silylation reactions oc- to undergo radical addition and electrophilic
tivity of the ortho- and para-positions owing cur regioselectively at the meta position (16–21). halogenation instead of direct oxidative rear-
to the electronic deficiency of the p-system Palladium-catalyzed meta-C–H olefination and omatization yet still rearomatize under spe-
as well as the efficient s-donor property of the arylation reactions were pioneered by the Yu cific conditions, such as acid treatment. We
nitrogen atom (5, 6). Current strategies for di- group by using designed ligands for regiose- present a protocol for versatile and practical
rect functionalization of pyridines include di- lectivity control (22–25). Despite the extensive meta-C–H functionalization of pyridines, quino-
rected metalation (7, 8), Minisci-type radical development and application of those tran- lines, and isoquinolines through a redox-neutral
reactions (9), and nucleophilic addition to sition metal–catalyzed reactions, the reaction dearomatization-rearomatization process (Fig.
N-activated pyridines (10, 11). However, for type and scope are still limited; for example, 1D). Pyridines can react with inexpensive and
unbiased pyridines, direct meta-functional- 3-substituted pyridines are required for regio- commercially available acetylenedicarboxylates
ization is far more challenging, with limited selectivity control in most cases (15). to form Huisgen 1,4-dipoles, which readily un-
An alternative strategy for the meta- dergo high-yielding dearomative cycloaddition
Organisch-Chemisches Institut, Westfälische Wilhelms-Universität, functionalization of pyridines is the tempo- reactions with carbonyl dipolarophiles (36, 37).
Münster, Germany.
*Corresponding author. Email: studer@[Link] rary dearomatization approach, so that the We found that the formed oxazino pyridines are
†These authors contributed equally to this work. initially electron-deficient heteroarenes are bench-stable intermediates, which could engage

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 779

RES EARCH | R E P O R T S

Fig. 1. meta-Functionalization of pyridines. (A) Various biologically pyridine. (D) Developed strategy through redox-neutral activation of
active compounds containing a meta-functionalized pyridine moiety. pyridines and subsequent regioselective ionic and radical C−H function-
(B) Established routes for the direct meta-C−H functionalization of alization. FG, functional group; DMAD, dimethyl acetylenedicarboxylate;
pyridines. (C) meta-Functionalization through reduction to the dihydro- MP, methyl pyruvate.

in regioselective radical and ionic functional- electrophilic trapping reagents. Separation the organic base DBU (1,8-diazabicyclo[5.4.0]
izations by use of a variety of radical precursors of the isomers was not required because they undec-7-ene) effectively promote trifluorome-
and electrophilic reagents. Acid-promoted re- converged to the same product after func- thylation of the oxazino pyridine intermedi-
aromatization then provides the corresponding tionalization and rearomatization. We found ates under blue light irradiation. We propose
meta-functionalized pyridines. With the redox- that rearomatization of the regioselectively that the generated electrophilic trifluoromethyl
neutral activation method of N-heteroarenes, C−H functionalized oxazino pyridine inter- radical (figs. S12 to S14) should electronically
we could achieve exclusively meta-selective C–H mediates is readily achieved in aqueous acid match with the nucleophilic dearomatized in-
functionalization to form trifluoromethyl, per- at 60°C in quantitative yields (table S4). termediates (40), which should also be respon-
fluoroalkyl, chloro, bromo, iodo, nitro, thio, and Considering the radical meta-pyridine func- sible for the high regioselectivity. Blue light
seleno N-heteroarenes under mild conditions. tionalization, C−H trifluoromethylation of the was required for chain initiation through carbon-
We evaluated first the reaction conditions oxazino pyridine intermediates merits atten- iodine (C−I) bond homolysis, and we assumed
for redox-neutral dearomatization and rearo- tion. Although radical trifluoromethylation is that the radical generated by the addition of
matization of pyridines (Fig. 2). A broad range a practical method for installing the pharma- the trifluoromethyl radical to the oxazino
of pyridines engaged in three-component cou- ceutically highly relevant trifluoromethyl group pyridine intermediate further reacts through
pling with dimethyl acetylenedicarboxylate on arenes (38), poor regioselectivity has been iodine atom abstraction from CF3I, sustaining
(DMAD) and methyl pyruvate (MP) in aceto- observed for unbiased pyridines owing to polarity- the chain (fig. S14). DBU-mediated HI elimi-
nitrile at room temperature under air (tables mismatch (39). A general strategy to directly nation leads to the C−H trifluoromethylated
S1 to S3), with an excellent average yield of 89% install a trifluoromethyl or other perfluoroalkyl oxazino pyridine, driving the endothermic I
(fig. S5). The generated oxazino pyridine inter- groups selectively at the meta-position of pyr- atom transfer step (41).
mediates, used as mixtures of diastereoisomers idines under mild conditions is highly desirable. Both electron-donating phenyl (2), phenoxy
[in some cases also as a regioisomeric mixture In this work, we realized formal direct meta- (3), alkyl groups (7), and electron-withdrawing
(fig. S5)], were stable toward air, water, and selective trifluoromethylation of pyridines by halo substituents (9, 10, 11, and 16) on pyridines
silica-gel-column chromatography (fig. S7), merging the dearomatization-rearomatization at different positions were compatible with
and these activated pyridines could then be strategy with trifluoromethyl radical chemis- the reaction. For substrates with two meta-C−H
used as efficient carbon-radical acceptors as try. Trifluoroiodomethane, one of the cheap- positions available (1 to 6), including unsub-
well as nucleophiles in reactions with various est trifluoromethyl sources (38), together with stituted pyridine (1), only mono-functionalized

780 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

Fig. 2. Scope of radical meta-fluoroalkylation and ionic meta-halogenation. H2SO4 was used instead of HCl. Paragraph symbol (¶) indicates 60% conversion
Compounds 1 to 23 are accessible through radical C−H functionalization by based on the analysis of 1H NMR spectra of crude product by using
using condition A. Ionic functionalization leads to halogenated heteroarenes 24 dibromomethane as an internal standard. Pound symbol (#) indicates mono-
to 42 by applying condition B. All yields of pyridine products correspond to and di-chlorination at 3- and 5-positions, with a ratio of 5:1 (mono:di). Two
isolated yields on the basis of the isolated dearomatized intermediates. Asterisk asterisks (**) indicate trimethyl silyl chloride (2.0 equiv) as additive, heating at
indicates 1.3 equiv alkyl iodide. Dagger symbol (†) indicates that yield was 40°C. Two dagger symbols (††) indicate trimethyl silyl chloride (2.0 equiv) as
determined by the analysis of 1H nuclear magnetic resonance (NMR) spectra of additive, heating at 60°C. DMAD, dimethyl acetylenedicarboxylate; MP, methyl
crude product by using dibromomethane as an internal standard. Double-dagger pyruvate; MeCN, acetonitrile; DBU, 1,8-diazabicyclo[5.4.0]undec-7-ene; NCS,
symbol (‡) indicates 10 equiv alkyl iodide. Section symbol (§) indicates that N-chlorosuccinimide; NBS, N-bromosuccinimide; NIS, N-iodosuccinimide.

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 781

RES EARCH | R E P O R T S

Fig. 3. meta-Functionalization of drugs and drug derivatives. Radical yields on the basis of the isolated dearomatized intermediates. Experimental
trifluoromethylation of various heteroarene containing biologically active details are provided in the supplementary materials. Asterisk indicates from the
molecules is achieved. The ionic C−H functionalization can be used for meta-C−H corresponding acetate. Dagger symbol (†) indicates from the corresponding
halogenation, nitration, sulfanylation, selenylation, and deuteration on more acetamide. Double-dagger symbol (‡) indicates from the corresponding tert-
complex heteroarenes. All yields of pyridine products correspond to isolated butyl carbamate.

products were obtained. The dienamine entity compounds, were accessible in moderate yields. activity patterns. Along these lines, our strat-
of the oxazino pyridine intermediates was reac- The reaction was also compatible with other egy enables selective trifluoromethylation
tive at the b- and d-position, and 2-aryl pyri- heteroarenes, such as thiophenes (5), pyrazoles of pyridines in the presence of electron-rich
dines gave a mixture of 3- and 5-functionalized (13), and oxazoles (14), with activation (de- arenes (3 and 5). Multisubstituted pyridines
products (5 and 6). The reverse selectivity aromatization) and accordingly C−H function- (15, 16, and 20 to 23) were compatible sub-
for products 5 and 6 indicates the inherent alization chemoselectively occurring on the strates. Quinolines (17), isoquinolines (18),
steric bias of the radical addition process. pyridine moiety. In general, electron-rich arenes and naphthyridines (19) also engaged in the
Chloro-, bromo-, and iodo-trifluoromethylated such as thiophenes are better trifluoromethyl transformation on application of the same ac-
pyridines (9, 10, and 11, respectively), which radical acceptors than pyridines (39), showing tivation strategy to provide the correspond-
can serve as versatile precursors to drug-like that our activation mode overrides innate re- ing trifluoromethylated heteroarenes with

782 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

Fig. 4. Synthetic applications. (A) meta-Trifluoromethylation of loratadine and meta-chlorination of (–)-cotinine as one-pot processes on a larger scale. (B) Sequential
radical-ionic or ionic-ionic meta,meta′-difunctionalization of pyridines. (C) Chemoselective meta-C−H functionalization of molecules containing two pyridine rings.

complete meta-selectivity. Along with the mainly due to low conversions (figs. S10 and strategy, selective electrophilic halogenation
trifluoromethyl group, other polyfluoroalkyl S11). Varying the oxazino pyridine moiety by at the meta-position of pyridines possessing
groups (4, 6, 8, 12 to 15, and 20 to 22) could using different dipolarophiles for pyridine both electron-rich and electron-poor substitu-
also be introduced to the meta position of dearomatization did not provide better re- ents was readily achieved with commercially
pyridines, quinolines, and isoquinolines un- sults (figs. S8 and S9). With iodoacetonitrile, available N-halosuccinimides (24 to 42). For
der the same conditions by simply varying the meta-alkylation of pyridines could be achieved, 2-aryl pyridines with two meta-positions avail-
radical precursors. As compared with the tri- albeit in low yield (23). able, mono-chlorination and mono-bromination
fluoromethylation, perfluoroalkylation with Direct meta-halogenation of pyridines un- were achieved with opposite regioselectivity
the corresponding commercial iodides worked der mild conditions has been a longstanding (24 to 28) as chlorination occurred exclu-
with higher efficiency. The relatively low yields challenge in synthetic chemistry (42). Using sively at the 3-position, whereas bromination
of the trifluoromethylation reactions were the redox-neutral dearomatization activation proceeded at the 5-position, addressing the

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 783

RES EARCH | R E P O R T S

b- or d-reactivity of the dienamine intermediate, (63; 92%) and 5-brominated (64; 69%) pro- pyridines, quinolines, and isoquinolines through
respectively. A plausible explanation is that ducts with exclusive regiocontrol. For the drug a redox-neutral dearomatization-rearomatization
chlorination occurs irreversibly at the more abametapir with a bipyridine structure, the strategy. The high selectivity, robust reaction
nucleophilic b-position, delivering the kinetic mono-dearomatized intermediate gave meta, operations, broad scope of transformations,
product, whereas because of the weaker carbon- meta′-dichlorination (65; 70%) and meta,meta′- and late-stage application all contribute to the
bromine (C−Br) bond, bromination is reversible dibromination (66; 51%) products, probably practicality of the method. Therefore, this redox-
and leads to the less sterically hindered and because of high reactivity of the corresponding neutral activation approach for meta-selective
thermodynamically more stable d-bromo di- dearomatized intermediate. To illustrate the functionalization of heteroarenes should be
enamine intermediates (42, 43). This analysis potential use of this redox-neutral activation widely applicable in the pharmaceutical and
is in line with the iodination that also occurs method for other previously inaccessible meta- agrochemical research arenas.
exclusively at the 5-position (26). We observed functionalization reactions, we achieved se-
good selectivity of mono-chlorination for lective deuteration of fasudil (68; 86%, >99%
4-aryl pyridines (29). Activation of the chlori- deuterium) by treating the dearomatized in- REFERENCES AND NOTES

nating reagent, N-chlorosuccinimide, with tri- termediate with deuterated trifluoroacetic acid 1. M. Baumann, I. R. Baxendale, Beilstein J. Org. Chem. 9,
2265–2319 (2013).
methylsilyl chloride (44) further extended the and deuterium oxide (D2O). 2. E. Vitaku, D. T. Smith, J. T. Njardarson, J. Med. Chem. 57,
scope to more electron-deficient pyridines (31 The meta-functionalizations of pyridines de- 10257–10274 (2014).
to 34, 36, and 37) that are virtually unreac- scribed above were carried out by means of 3. J. K. Kallitsis, M. Geormezi, S. G. Neophytides, Polym. Int. 58,
1226–1233 (2009).
tive under conventional electrophilic aromatic two-pot processes, in which the dearomatized 4. M. N. Zafar et al., Russ. J. Coord. Chem. 42, 1–18 (2016).
halogenation conditions (13). Moreover, bipyr- intermediates were isolated before the ensu- 5. K. Murakami, S. Yamada, T. Kaneda, K. Itami, Chem. Rev. 117,
idine and terpyridine ligands, such as 37 and ing C−H functionalization step to guarantee 9302–9332 (2017).
6. F.-Y. Zhou, L. Jiao, Synlett 32, 159–178 (2021).
38, could be selectively mono-dearomatized facile isolation of products and limit side re- 7. M. Balkenhohl, P. Knochel, SynOpen 2, 0078–0095 (2018).
at one of the pyridine rings, delivering after actions. This allowed for ready diversification 8. Y. Nakao, Synthesis 2011, 3209–3219 (2011).
rearomatization monochlorinated products. because various radical precursors and elec- 9. R. S. J. Proctor, R. J. Phipps, Angew. Chem. Int. Ed. 58,
13666–13699 (2019).
Isoquinolines (39) and 1-azaphenanthrenes trophiles engage in the C−H functionalization, 10. J. A. Bull, J. J. Mousseau, G. Pelletier, A. B. Charette, Chem.
(40 and 41) can also be selectively halogenated as we have documented. However, targeting Rev. 112, 2642–2713 (2012).
at the meta position of the pyridine core. The a single meta-functionalized compound, we 11. R. D. Dolewski, M. C. Hilton, A. McNally, Synlett 29, 08–14
(2018).
potential application of this redox-neutral activa- found that most reactions can be readily con-
12. D. E. Stephens, O. V. Larionov, Tetrahedron 71, 8683–8716
tion strategy toward other N-heteroarenes could ducted as one-pot processes. To demonstrate the (2015).
be demonstrated by the successful 4-selective practicality and robustness of the method in 13. G. A. Olah, Acc. Chem. Res. 4, 240–248 (1971).
chlorination of a thiazole compound (42). both medicinal and process chemistry, we dis- 14. B. Galabov, D. Nalbantova, Pv. Schleyer, H. F. Schaefer 3rd,
Acc. Chem. Res. 49, 1191–1199 (2016).
Late-stage functionalization enables rapid played one-pot reactions at a larger scale, start- 15. R. Das, M. Kapur, Asian J. Org. Chem. 7, 1217–1235 (2018).
modification of drugs, drug candidates, and ing from loratadine and (–)-cotinine (Fig. 4A and 16. M. A. Larsen, J. F. Hartwig, J. Am. Chem. Soc. 136, 4287–4299
materials without the need for de novo syn- figs. S3 and S4). meta-Trifluoromethylation of (2014).
17. J. S. Wright, P. J. H. Scott, P. G. Steel, Angew. Chem. Int. Ed.
thesis (45, 46). We demonstrated that the loratadine on a 1-mmol scale gave 47% isolated 60, 2796–2821 (2021).
developed redox-neutral dearomatization- yield of 60 without any solvent change. Gram- 18. S. A. Sadler et al., Org. Biomol. Chem. 12, 7318–7327 (2014).
functionalization-rearomatization strategy is a scale synthesis of meta-chlorinated (–)-cotinine 19. L. Yang, N. Uemura, Y. Nakao, J. Am. Chem. Soc. 141,
7972–7979 (2019).
practical method for the direct late-stage mod- 45 was achieved under air without the need for 20. C. Cheng, J. F. Hartwig, J. Am. Chem. Soc. 137, 592–595 (2015).
ification of drugs that contain pyridine moieties column chromatography. Through sequential 21. B.-J. Li, Z.-J. Shi, Chem. Sci. 2, 488–493 (2011).
(Fig. 3). Twelve drugs and drug derivatives with ionic chlorination and radical trifluorometh- 22. M. Ye, G.-L. Gao, J.-Q. Yu, J. Am. Chem. Soc. 133, 6964–6967
(2011).
varied substitution patterns underwent diverse ylation, regioselective consecutive function- 23. M. Ye et al., J. Am. Chem. Soc. 133, 19090–19093 (2011).
meta-selective functionalizations in generally alization of a ligand was realized to generate 24. T. Zhang et al., Nat. Chem. 13, 1207–1213 (2021).
good yields (30 to 92%; 43 to 68). As an il- 3,5-difunctionalized pyridine 71 (66% yield) 25. Z. Fan et al., Nature 610, 87–93 (2022).
26. C. S. Giam, S. D. Abbott, J. Am. Chem. Soc. 93, 1294–1296 (1971).
lustration, various functionalities—including without isolation of the mono-functionalized
27. O. Tsuge, S. Kanemasa, T. Naritomi, J. Tanaka, Chem. Lett. 13,
trifluoromethyl (43), pentafluoroethyl (44), intermediates, leveraging the potential of the 1255–1258 (1984).
chloro (45), bromo (46), iodo (47), nitro (48), b- and d-reactivity of the dienamine interme- 28. S. Wübbolt, M. Oestreich, Angew. Chem. Int. Ed. 54,
sulfanyl (49), and selenyl (50) groups—could diates (Fig. 4B). Likewise, sequential ionic-ionic 15876–15879 (2015).
29. Z. Liu et al., J. Am. Chem. Soc. 144, 4810–4818 (2022).
be selectively introduced into the meta position difunctionalization of the atazanavir precursor 30. X.-Y. Zhou, M. Zhang, Z. Liu, J.-H. He, X.-C. Wang, J. Am. Chem.
of the alkaloid (–)-cotinine through radical and was accomplished (74; 74%). When two differ- Soc. 144, 14463–14470 (2022).
ionic pathways. Although some drugs and pre- ently substituted pyridines were presented 31. A. Grozavu, H. B. Hepburn, E. P. Bailey, P. J. Lindsay-Scott,
T. J. Donohoe, Chem. Sci. 11, 8595–8599 (2020).
cursors required protection of the hydroxyl or in the same molecule, the redox-neutral dearo- 32. S. Park, S. Chang, Angew. Chem. Int. Ed. 56, 7720–7738 (2017).
amino groups before dearomatization (53, 56, matization showed high selectivity toward 33. G. Bertuzzi, L. Bernardi, M. Fochi, Catalysts 8, 632 (2018).
67, and 68), acid-promoted rearomatization more electron-rich and less sterically hindered 34. A. Grozavu et al., Nat. Chem. 11, 242–247 (2019).
35. M. W. Gribble Jr., S. Guo, S. L. Buchwald, J. Am. Chem. Soc.
expediently furnished the free alcohol or amine pyridines, allowing for chemoselective mono- 140, 5057–5060 (2018).
products. The relatively electron-rich pyridine meta-functionalization of polypyridine com- 36. R. Huisgen, M. Morikawa, K. Herbig, E. Brunn, Chem. Ber. 100,
ring in metyrapone was selectively trifluoro- pounds [54 and 55 (Fig. 3) and 75 to 79 (Fig. 1094–1106 (1967).
37. V. Nair, A. Deepthi, D. Ashok, A. E. Raveendran, R. R. Paul,
methylated (54) and chlorinated (55) in 47 4C)]. Even for compounds that contain pyri- Tetrahedron 70, 3085–3105 (2014).
and 68% yield, respectively. The pyrimidine dines in the same chemical environment, such 38. H. Baguia, G. Evano, Chem. Eur. J. 28, e202200975 (2022).
moiety in the imatinib precursor was well tol- as the linker molecule used to construct metal 39. F. O’Hara, D. G. Blackmond, P. S. Baran, J. Am. Chem. Soc.
135, 12122–12134 (2013).
erated (56; 76%). Tropicamide with a 4-alkyl organic frameworks (78 and 79) (47), controlled
40. A. Studer, Angew. Chem. Int. Ed. 51, 8950–8958 (2012).
pyridine moiety was mono-trifluoromethylated mono- and di-dearomatization can lead to 41. F. Juliá, T. Constantin, D. Leonori, Chem. Rev. 122, 2292–2352
in moderate yield (59), whereas milrinone with a the corresponding mono- and di-chlorination (2022).
4-aryl pyridine moiety delivered an isolable mix- products, respectively, in high yields. 42. B. T. Boyle, J. N. Levy, L. de Lescure, R. S. Paton, A. McNally,
ChemRxiv 88802 [Preprint] (2022); doi:10.26434/chemrxiv-
ture of mono- and di-trifluoromethylated pro- We have developed a general method for rad- 2022-88802-v3.
ducts (58). Vismodegib furnished 3-chlorinated ical and ionic meta-C–H functionalization of 43. L. A. Paquette, W. C. Farley, J. Org. Chem. 32, 2725–2731 (1967).

784 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

44. T. Maibunkaew, C. Thongsornkleeb, J. Tummatorn, A. Bunrit, and A.S. designed and analyzed the experiments. H.C. and Q.C. SUPPLEMENTARY MATERIALS
S. Ruchirawat, Synlett 25, 1769–1775 (2014). conducted the experiments. H.C., Q.C., and A.S. wrote the [Link]/doi/10.1126/science.ade6029
45. T. Cernak, K. D. Dykstra, S. Tyagarajan, P. Vachal, S. W. Krska, manuscript. A.S. directed the project. All authors contributed to Materials and Methods
Chem. Soc. Rev. 45, 546–576 (2016). discussions. Competing interests: The authors declare that they Supplementary Text
46. J. Börgel, T. Ritter, Chem 6, 1877–1887 (2020). have no competing interests. Data and materials availability: Figs. S1 to S14
47. F. A. Almeida Paz, J. Klinowski, Inorg. Chem. 43, 3882–3893 All data are available in the main text or the supplementary Tables S1 to S4
(2004). materials. License information: Copyright © 2022 the authors, NMR Spectra
some rights reserved; exclusive licensee American Association References (48–56)
ACKN OW LEDG MEN TS for the Advancement of Science. No claim to original US government
Funding: We thank the Westfälische Wilhelms–Universität Münster works. [Link] Submitted 26 August 2022; accepted 18 October 2022
for supporting this work. Author contributions: H.C., Q.C., article-reuse 10.1126/science.ade6029

QUANTUM PHYSICS alizing the Kitaev model, for example, in nano-

wires with spin-orbit interactions placed in
Noise-resilient edge modes on a chain of the proximity of a superconductor (9–16). Here,
the underlying ℤ2 symmetry cannot be broken
superconducting qubits by local perturbations within a closed system.
Nevertheless, theoretical results have widely
X. Mi1, M. Sonner2, M. Y. Niu1, K. W. Lee1, B. Foxen1, R. Acharya1, I. Aleiner1, T. I. Andersen1, F. Arute1, suggested that MEMs remain susceptible to
K. Arya1, A. Asfaw1, J. Atalaya1, J. C. Bardin1,3, J. Basso1, A. Bengtsson1, G. Bortoli1, A. Bourassa1, a variety of decoherence effects from their
L. Brill1, M. Broughton1, B. B. Buckley1, D. A. Buell1, B. Burkett1, N. Bushnell1, Z. Chen1, B. Chiaro1, open solid-state environment (17–20). Exper-
R. Collins1, P. Conner1, W. Courtney1, A. L. Crook1, D. M. Debroy1, S. Demura1, A. Dunsworth1, imental results have also established that the
D. Eppens1, C. Erickson1, L. Faoro1, E. Farhi1, R. Fatemi1, L. Flores1, E. Forati1, A. G. Fowler1, W. Giang1, density of subgap quasiparticles is often orders
C. Gidney1, D. Gilboa1, M. Giustina1, A. G. Dau1, J. A. Gross1, S. Habegger1, M. P. Harrigan1, M. Hoffmann1, of magnitude higher than predictions from
S. Hong1, T. Huang1, A. Huff1, W. J. Huggins1, L. B. Ioffe1, S. V. Isakov1, J. Iveland1, E. Jeffrey1, Z. Jiang1, simple thermal population arguments (21–24).
C. Jones1, D. Kafri1, K. Kechedzhi1, T. Khattar1, S. Kim1, A. Y. Kitaev1,4, P. V. Klimov1, A. R. Klots1, The incoherent processes involving these quasi-
A. N. Korotkov1,5, F. Kostritsa1, J. M. Kreikebaum1, D. Landhuis1, P. Laptev1, K.-M. Lau1, J. Lee1, L. Laws1, particles can change the parity of the ground state
W. Liu1, A. Locharla1, O. Martin1, J. R. McClean1, M. McEwen1,6, B. Meurer Costa1, K. C. Miao1, and, consequently, destroy the topological pro-
M. Mohseni1, S. Montazeri1, A. Morvan1, E. Mount1, W. Mruczkiewicz1, O. Naaman1, M. Neeley1, tection. These results highlight the importance
C. Neill1, M. Newman1, T. E. OÕBrien1, A. Opremcak1, A. Petukhov1, R. Potter1, C. Quintana1, N. C. Rubin1, of characterizing the extent of symmetry pro-
N. Saei1, D. Sank1, K. Sankaragomathi1, K. J. Satzinger1, C. Schuster1, M. J. Shearn1, V. Shvarts1, D. Strain1, tection in realistic open-system environments.
Y. Su1, M. Szalay1, G. Vidal1, B. Villalonga1, C. Vollgraff-Heidweiller1, T. White1, Z. Yao1, P. Yeh1, J. Yoo1, The advent of high-fidelity quantum proces-
A. Zalcman1, Y. Zhang1, N. Zhu1, H. Neven1, D. Bacon1, J. Hilton1, E. Lucero1, R. Babbush1, S. Boixo1, sors and simulators suggests an alternative
A. Megrant1, Y. Chen1, J. Kelly1, V. Smelyanskiy1, D. A. Abanin1,2*, P. Roushan1* approach to examining the realistic extent of
protection for a given symmetry (25–27). In
Inherent symmetry of a quantum system may protect its otherwise fragile states. Leveraging such this study, we use the Jordan-Wigner trans-
protection requires testing its robustness against uncontrolled environmental interactions. Using formation (JWT) to map the Kitaev model to a
47 superconducting qubits, we implement the one-dimensional kicked Ising model, which exhibits nonlocal transverse Ising spin model (28), which is more
Majorana edge modes (MEMs) with ℤ2 parity symmetry. We find that any multiqubit Pauli operator compatible with a chain of qubits (29–32). The
overlapping with the MEMs exhibits a uniform late-time decay rate comparable to single-qubit relaxation JWT also maps each MEM, commonly repre-
rates, irrespective of its size or composition. This characteristic allows us to accurately reconstruct the sented by a sum of local Majorana operators
exponentially localized spatial profiles of the MEMs. Furthermore, the MEMs are found to be resilient against in the fermionic chain, to a sum of Pauli spin
certain symmetry-breaking noise owing to a prethermalization mechanism. Our work elucidates the complex operators that can be individually character-
interplay between noise and symmetry-protected edge modes in a solid-state environment. ized on a quantum processor. Given the non-
local nature of the JWT, the MEMs in the Pauli

T
basis are prone to local symmetry-breaking
he symmetry of a quantum system can for topological quantum computing (5, 6). An noise even within a closed system, which dis-
give rise to topologically distinct degen- example model supporting symmetry-protected tinguishes them from MEMs in fermionic sys-
erate ground states. A quantum super- topological states is the Kitaev model of spinless tems. Despite this disadvantage, we find that
position of such states is, in principle, fermions in a one-dimensional (1D) wire (7). The the interplay between ℤ2 parity symmetry
immune to dephasing; additionally, an ℤ2 parity symmetry of the model leads to a pair and a prethermalization mechanism endows
energy gap separates the ground states from the of degenerate ground states. The distinct parities the MEMs with a strong resilience toward
excited states and further protects the ground of the two ground states protect them against both closed-system thermalization and open-
states from energy decay. As such, symmetry- local parity-preserving noise, such as potential system perturbations such as low-frequency
protected ground states may form decoherence- fluctuations (8). The topological property of noise. Furthermore, we discover a method
free subspaces (1–4) and are promising candidates these degenerate ground states is commonly for accurately reconstructing the Pauli expan-
1
Google Research, Mountain View, CA, USA. 2Department of
described by a pair of localized Majorana sion of MEMs in the presence of decoher-
Theoretical Physics, University of Geneva, Geneva, edge modes (MEMs) at the ends of the wire. ence, which may be extended to study other
Switzerland. 3Department of Electrical and Computer Whereas the degree of symmetry protection integrals of motion in many-body quantum
Engineering, University of Massachusetts, Amherst, MA, USA. in a closed quantum system is often under- systems.
4
Institute for Quantum Information and Matter, California
Institute of Technology, Pasadena, CA, USA. 5Department of stood, experimental quantum systems are in- The experiment is conducted on an open-
Electrical and Computer Engineering, University of California, variably subject to physical noise sources that ended chain of L = 47 superconducting qubits
Riverside, CA, USA. 6Department of Physics, University of do not necessarily respect the underlying sym- [see (33) for device details]. The qubit chain is
California, Santa Barbara, CA, USA.
*Corresponding author. Email: banin@[Link] (D.A.A.); metry. In the context of MEMs, notable efforts periodically driven by a quantum circuit cor-
pedramr@[Link] (P.R.) have been directed toward experimentally re- responding to a kicked Ising model (Fig. 1B),

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 785

RES EARCH | R E P O R T S

Fig. 1. Observation of long-lived edge modes in a kicked Ising model. shown on a unit circle according to their quasienergies. (C) Z^ j ðtÞ as a function
(A) Schematic illustration of the Jordan-Wigner transformation between a 1D of t and qubit location Qj. The initial state is a random product state, j0101001:::i.

fermionic Kitaev chain and a qubit chain. In the fermionic (qubit) chain, the (Inset) Z^ j ðtÞ
for the
three leftmost and rightmost qubits, between t = 50 and
sizes (widths) of the colored spheres (bars) denote the relative weights of the t = 200. (D) Z^ j ðtÞ for the two edge qubits j = 1 and 47 (left panel) and two
edge modes in the Majorana fermion (Pauli) basis. The right edge mode in qubits within the bulk j = 16 and 32 (right panel). Top axis for each plot indicates
the Pauli basis is dominated by long Pauli operators spanning the entire chain. real time, calculated on the basis of the time needed to execute U ^ F (93 ns).
(B) (Left) Quantum circuit implementation of a kicked Ising model. An identical Locations for the qubits shown in this panel are also indicated by colored
unitary U^ F is repeated a total of t times. (Right) Eigenstates of U
^ F (g > 0.5), arrows in (C).

with the following unitary applied in each is a set of local z-fields that break integrability on a unit circle, as illustrated in the right panel
cycle of the model. Compared to a digitized imple- of Fig. 1B.
mentation of the transverse Ising Hamiltonian Given that there is no ground state in a
L
X L
X1 L
i ipJ ipg X (see fig. S11 for experimental data of this ap- Floquet system, the twofold degeneracy of the
2 hj Z^ j 2 Z^ jZ^ jþ1 2
^j
X
^F ¼e
U j¼1
e j¼1
e j¼1
ð1Þ proach), the periodic (i.e., Floquet) evolution ground state in the Kitaev model instead be-
here generates faster dynamics in real time comes a pairing of eigenstates across the entire
where X ^ j and Z^ j denote Pauli operators and is advantageous given the finite coherence spectrum. In our work, we fix J = 1/2, wherein
acting on a given qubit Qj. Here J and g de- times of the qubits. Under such a drive, the the Floquet system, in the integrable limit hj = 0,
note the strengths of the Ising interaction Hilbert space of the system may be described has ℤ2 spin-flip symmetry and exhibits two
and the transverse fields, respectively; and hj by the eigenstates of U ^ F whose eigenvalues lie phases with distinct spectral pairings. For

786 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

Fig. 2. Quasienergy spectroscopy. (A) (Top) Z^ 1 ðtÞ measured in the spectra, Z^ 1 ðtÞ is measured up to t = 300, 200, and 150 cycles for L = 6, 12, and
integrable limit hj = 0 and system size L =
12. (Bottom)
Frequency-dependent 18, respectively. (C) (Top) n(w) for different L, showing the quasienergy peaks
amplitude n(w) of the Fourier transform of Z^ 1 ðtÞ in the top panel. The arrows of the hybridized MEMs with a splitting 2D/p. Data are offset for clarity. (Bottom)
indicate the single-particle quasienergy peaks for the bulk and edge fermionic D/p measured as a function of L at different values of g. Solid lines represent
modes. a.u., arbitrary units. (B) n as a function of both frequency w/p and g, exact numerical results from diagonalizing U ^ F in the fermionic basis (33).
measured for three different values of L. In all cases, hj = 0. To obtain the Random product states are used as initial states in all measurements.

^ 1 ðtÞ
Fig. 3. Low-frequency noise resilience of MEMs and comparison with realizations

hj/p ∈ [−0.05,0.05].
with (Bottom panels) Disorder-averaged X
unprotected edge modes. (A) (Left) Quantum circuit corresponding to the XY ^ ^ ^ ^
Z 1 ð0ÞZ 1 ðtÞ for the U XY U F edge mode, shown over four different disorder
^
model,
wherean identical
cycle unitary U XY is applied t times. (Right) Top panel strengths d. Eighty disorder instances hj/p ∈ [−d/p, d/p] are used for averaging
^ ^
shows X 1 ðtÞ and Y 1 ðtÞ measured at Q1, with the control parameter z/p = 1.0 in each case, and the initial states are additionally randomized between instances
^ F. (D) Red lines: Fourier spectra n(w) obtained from the disorder instances in

and no disorder
hj/p = 0. Bottom panel shows the Fourier spectrum n(w) of for U
^ 1 ðtÞ þ i Y^ 1 ðtÞ . (B) n(w) as a function of w/p and z for the U
X ^ XY model, where the upper panels of (C). Black lines: n(w) for the disorder-averaged observables

^
X

1 ðt Þ ^
Y ðtÞ t ^ 1 ðtÞ
X
and 1 are measured
up to = 100. (C) (Top panels) (d = 0.05) in the lower panels of (C). (E) Maximum Fourier amplitude nmax as a
^ ^ ^ ^
Z 1 ð0ÞZ 1 ðtÞ for the U XY U F edge modes, measured for four different disorder function of d. Data are normalized by nmax at d = 0.

the phase characterized by g > 0.5, which is the energy q + p (34). The transition between any c ^F ¼
^LU ^ F^
U cL; c ^F ¼
^R U ^ Fc
U ^R ð2Þ
focus of the main text, the quasienergy levels paired eigenstates is enabled by an application
have a p-pairing (Fig. 1B, right panel): Every of the so-called p-MEMs, c ^ L and ^
c R (35, 36). At g < 0.5, the eigenspectrum of U ^ F has a
many-body eigenstate of the U ^ F with quasi- The p-MEMs anticommute with U ^ F in the double degeneracy, that is, each eigenstate
energy q has a “partner” state with quasi- large L limit has a partner state with the same quasienergy.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 787

RES EARCH | R E P O R T S

Fig. 4. Reconstructing the Pauli operator expansion of MEMs. (A) (Top panels) experimentally reconstructed Pauli operator expansion of the MEMs c ^ L;R ;

aZ;n and aY;n correspond to the coefficients of the Pauli operators shown
^ ðtÞ for g = 0.8 and g = 0.6, with the compositions of C
Correlators Z^ 1 ð0ÞC ^ ðtÞ in
the legends. The bottom two panels show experimental values of aZ;n
shown in the legend. Here, the g = 0.8 ( g = 0.6) data are averaged over 10 (12)
(points) and theoretical predictions (solid lines). Error bars correspond to
disorder realizations and initial random product states. Bottom panels show the
statistical uncertainty stemming from single-shot measurements [see
^ ðtÞ. (B) The top eight panels show
absolute values of the correlators, Z^ 1 ð0ÞC methods (33)].

Here, the transition between paired eigen- for the edge qubits. In addition, Z^ j ðt Þ for each first to a metastable state before decaying to
states is described by two so-called 0-MEMs edge qubit shows a subharmonic oscillation ergodic states. A common mechanism for pre-
that commute with U ^ F . Experimental data at a period twice that of the drive U ^ F, because thermalization is the existence of spectral
for this regime, which is analogous to the each application of U ^ F changes the sign of gaps, which make relaxation processes driven
ferromagnetic phase of the transverse Ising ^ L;R owing to their anticommutation (Eq. 2).
c by integrability-breaking perturbations off-
model, are shown in fig. S10. At the critical The bulk-edge difference is further illus- resonant, thereby preventing energy absorption.
point g = 0.5, the eigenstates are distributed trated in Fig. 1D, where data for four qubits To experimentally establish prethermalization
uniformly on the unit circle with a gap of are shown. The lifetimes of the edge modes, in our system, we characterize the excitation
p/L, which vanishes in the limit L = ∞. which include contributions from both exter- spectrum in the integrable limit, hj = 0. Here, the
In the presence of finite local fields hj ≠ 0, nal decoherence effects and internal non- many-body spectrum of U ^ F may be constructed
U^ F is no longer integrable and the ℤ2 sym- integrable dynamics, are

extracted
by fitting from a total of 2L energy quanta, corresponding
metry is also broken. We begin by searching the envelope of Z^ 1 ðt Þ Z^ 47 ðt Þ to an expo- to the quasienergies of noninteracting Bogoliubov
for signatures of stable edge modes in this nential (fig. S8) and found to be 19.5 ms (17.2 ms) fermionic quasiparticles in the fermionic repre-
regime, focusing on the Z^ operators that, in for Q1 (Q47). These values are close to the typ- sentation of U ^ F . These quasienergies can be
the JWT, have large overlap with MEMs on ical single-qubit relaxation time T1 = 22.2 ms obtained through a Fourier analysis of time-
the edge (Fig. 1A). Figure
1C shows
experi- on the device—a preliminary indication that domain signals (42, 43) [see supplementary
mental measurements of Z^ j ðt Þ for all qubits the MEMs are resilient toward integrability- text sections III and
V (33)].
Figure 2A shows
in the chain, where we have chosen hj/p from a and symmetry-breaking fields as well as de- measurements of Z^ 1 ðt Þ ðhj ¼ 0Þ for
a short

random uniform distribution [−1,1] to maxi- phasing effects such as low-frequency noise. chain L = 12. The time evolution for Z^ 1 ðt Þ is
mize the effect of integrability breaking. We Recent theoretical works have suggested now seemingly featureless, which results from
observe a stark contrast in the behavior of the that the resilience of the edge modes toward interference between different eigenmodes of
edge qubits, Q1 and Q47, and U^ F . To obtain the quasienergies, a Fourier trans-

qubits
within the nonintegrable dynamics is a result of prether-
chain, Q2 to Q46. Whereas Z^ j ðt Þ decays rap- malization (37–41). Unlike thermalizing sys- form of the time-domain data is then performed

∼20 cycles (∼2 ms) for any qubit
idly to 0 after tems, which monotonically decay to ergodic (Fig. 2A, bottom). The Fourier spectrum n(w)
in the bulk, Z^ j ðt Þ decays much more slowly states over time, a prethermal system relaxes reveals a total of 2L distinct peaks at values of

788 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

w corresponding to the quasienergies of the 2L To probe one of the edge modes for U ^ XY, we U^ F remains stable at w = p, irrespective of
noninteracting fermionic modes in the system. prepare the system in a superposition state disorder realizations. Lastly, we measure the
The two dominant peaks close to w = p in p1ffiffi ðj0000:::i þ j1000:::iÞ to maximize the ini- disorder-averaged quasienergy peak height,
2
^þ ^
the spectrum of Fig. 2A are associated with tial value s 1 ðt ¼ 0Þ ¼ 1. We then apply U XY nmax = Max[n(w)], and show the results in Fig.
the MEMs, which are split in quasienergy be- t times before measuring the time-dependent 3E. For the U ^ XY model, we observe that nmax
^þ ^ ^
cause of their hybridization in this short chain. observable s 1 ðt Þ ¼ X 1 ðt Þ þ i Y 1 ðt Þ , which decays exponentially as a function of d irre-
To confirm this interpretation, we change the precesses at a frequency corresponding to the spective of z. For the U ^ F model, nmax is com-
localization length x of the MEMs by tuning g quasienergy of the edge mode. Example ex- pletely insensitive to d for sufficiently localized
and measure the spectra over three different periment data and the corresponding Fourier MEMs (g = 0.8) and remains insensitive for
system sizes. The results, shown in Fig. 2B, re- spectrum n(w) for z/p = 1.0 are both shown in small d < 0.05 even in the more delocalized
veal two important features: First, we observe Fig. 3A, demonstrating a slowly decaying sub- regime g = 0.6. These results highlight the crit-
that the quasienergy splitting 2D of the two harmonic response and a quasienergy peak at ical role of symmetry in stabilizing the quasi-
MEMs decreases as g increases. This is caused w = p that are similar to those of the MEMs of energies of MEMs and protecting their lifetimes
by a reduced x that leads to weaker hybrid- the U ^ F model. Figure 3B shows experimentally against low-frequency noise.
ization between ^ c L and c ^ R . Second, we ob- measured n as a function of both z and w. At Finally, using the full L = 47 qubit chain, we
serve a finite quasienergy gap X between the 0.25 ≲ z/p ≲ 1.75, we observe a dominant quasi- demonstrate an error-mitigation strategy for ac-
MEMs and the other bulk fermionic modes, energy peak that corresponds to an edge mode curately reconstructing the Pauli operator ex-
which increases at larger g. This quasienergy and is separated from the L − 2 bulk modes, pansion of c^ L;R in the presence of noise. Figure
gap, which crucially remains open as L in- visible as smaller peaks outside the range 0.5 < 4A shows the late-time evolution of eight multi-
creases, suppresses transitions between bulk wp < 1.5, by spectral gaps akin to the bulk gap X qubit Pauli operators C^ entering the JWT of ^c L;R,
and edge states and is the key to protecting the of the U ^ F model. Despite these apparent simi- experimentally obtained by rotating each qubit
MEMs against integrability-breaking fields. larities, a crucial distinction exists between the 1into the appropriate basis followed by multi-
two models: The quasienergy of the U ^ XY edge
qubit readout. We observe that each
Further discussion of this prethermalization ^ ðt Þ
Z^ ð0ÞC 1
mechanism is presented in supplementary mode is first-order sensitive to z at all values of
text section IV (33). z, whereas the quasienergy of the U ^ F edge mode exhibits a similar subharmonic response, with
Although the bulk gap protects the MEMs from asymptotically approaches p as g increases. This an amplitude that decreases when C ^ incorpo-
internal thermalization, the finite quasienergy distinction stems from the lack of ℤ2 symmetry rates more qubits and has less overlap with
difference 2D between the two MEMs is sen- in U ^ XY and leads to drastically different robust- c L;R (Fig. 1A). Operators ending with Y^ also
^
sitive to disorder fluctuations. Such a sensi- ness of the two models toward low-frequency show smaller amplitudes than those ending
tivity may lead to dephasing of the MEMs noise in hj, which we explore next. with Z^ because they have no overlap with c ^ L;R
through low-frequency noise (20), as shown The upper panels of Fig. 3C show X ^ 1 ðt Þ in the time-independent transverse Ising mod-
by fig. S5. This effect may be suppressed by Z^ 1 ð0ÞZ^ 1 ðt Þ of the U ^ XY U
^ F model, mea- el and only arise as corrections to the JWT of
reducing the hybridization between ^ c L and sured for four different realizations of hj/p ^
c L;R as a consequence of the time-dependent,
^ R , which is achievable through increasing
c that are uniformly chosen from [−d, d]. Here, periodic dynamics. Notably, as shown also in
either g (Fig. 2B) or L. The dependence of D on the autocorrelator Z^ 1 ð0ÞZ^ 1 ðt Þ differs from Fig. 4A, the absolute values (i.e., magnitudes) of

L is mapped out in detail by the experimental Z^ 1 ðt Þ only by a random ± sign given by the ^ ðt Þ , exhibit nearly iden-
these operators, Z^ 1 ð0ÞC
measurements shown in Fig. 2C. We observe initial state of Q1, and d = 0.05 is a disorder tical decay rates despite their different lengths
that for g > 0.6, D is exponentially suppressed strength chosen to be comparable to the low- and compositions.
by larger L, in agreement with theory (33). For frequency fluctuation of the quantum device. The observation in Fig. 4A is contrary to
g < 0.6, the suppression is no longer exponen- We observe that X ^ 1 ðt Þ exhibits beating patterns naïve expectations, wherein the decay rate of
tial, given the proximity to the phase transi- that depend sensitively on the disorder realiza- a quantum operator is expected to scale with
tion point g = 0.5, where the bulk gap closes. tion. This is a result of the first-order sensitivity the number of qubits it incorporates. The result
We also find excellent agreement between toward control parameters demonstrated in may be qualitatively understood by the fact that
exact numerical results and experimental mea- Fig. 3B. On the other hand, Z^ 1 ð0ÞZ^ 1 ðt Þ is vir- ^ ^ R anticommute with U
c L and c ^ F (Eq. 2) and are
surements even for D/p ≈ 0.01, which is a result tually unchanged between different disorder conserved under the periodic dynamics. Even
of accurate gate calibrations described in (33). realizations. The impact of low-frequency noise though external decoherence and integrability-
We next perform a systematic study on the on each edge mode realization is then emulated breaking fields violate this commutation, c ^ L;R
low-frequency noise resilience of the MEMs by averaging the corresponding observable remains a slowly decaying mode [see supple-
for a moderately long chain, L = 20. To ex- over an ensemble of disorder realizations, which mentary text section VII (33)]. As a result, any
amine the role of symmetry in noise protec- mimics the process of dephasing. The disorder- multiqubit operator having a finite overlap with

tion, we have also experimentally realized edge averaged X ^ 1 ðt Þ in the U ^ XY model, shown in ^
c L;R will exhibit a slow-decaying expectation
modes in a different periodic circuit with a the lower panel of Fig. 3C, decays significantly value in its late-time dynamics, with an ampli-
cycle unitary U ^ XY that does not have ℤ2 sym- faster as the disorder strength d increases. On tude proportional to the overlap.

metry. As illustrated in Fig. 3A, U ^ XY consists the other hand, Z^ 1 ð0ÞZ^ 1 ðt Þ in the U ^ F model The uniform decay rates of Z^ 1 ð0ÞC ^ ðt Þ in-
of two layers of two-qubit gates applied pffiffiffiffiffiffiffiffiffiffiffiffiffibe- remains unchanged over d. form an experimental strategy for reconstruct-
tween all nearest-neighbor qubits, iSPðzÞ ¼ The sensitivity of the two edge mode real- ing the expansion of the MEMs in the Pauli
e i4 ðZ j Z jþ1 Þ e i4 ðs^ j s^jþ1 þ s^j s^jþ1Þ e i4 ðZ j Z jþ1 Þ, where
z ^ ^ p þ þ z ^ ^
izations toward low-frequency noise is further operator basis
þ;
^
s denotes Pauli raising and lowering opera- elucidated by inspecting the Fourier spectrum " ! ! #
tors. In the single-excitation subspace, U ^ XY has n(w) of each disorder realization, shown in Fig. X
L Y
N Y
N
^
c L;R ¼ aZ;n X j Z^ n þ aY;n
^ X j Y^ n
^
L eigenmodes, including two localized edge 3D. Here, we observe that the quasienergy peak n¼1 j¼M j¼M
modes [see appendix E of (44)] for control for the U ^ XY edge mode is different for each dis-
ð3Þ
parameter z=p ∈½0:25; 1:75. The leading order order realization, resulting in a broadened spec-
terms in the Pauli operator expansion of the trum with a lower peak height upon averaging. where the products over j have limits N = n − 1
edge modes are s ^1þ and s ^Lþ , respectively. On the other hand, the quasienergy peak for ^ L , and N = L and M = n + 1 for ^
and M = 1 for c cR

SCIENCE [Link] 18 NOVEMBER 2022 ¥ VOL 378 ISSUE 6621 789

RES EARCH | R E P O R T S

(45). The coefficients a and a normalize

X Z;n 2 Y;n 2
12. L. P. Rokhinson, X. Liu, J. K. Furdyna, . Nat. Phys. 8, 795–799 44. C. Neill et al., Nature 594, 508–512 (2021).
Y
to unity for L = ∞: n aZ;n þaY;n ¼ 1. To (2012). 45. In Eq. 3, product terms Nj¼M X ^ j with N < M should be treated

estimate their values, we measure different

13. A. Das et al., Nat. Phys. 8, 887–895 (2012). as identity ^I. Additionally, the JWT of c ^ R should, in principle,
14. S. Nadj-Perge et al., Science 346, 602–607 (2014).
^ ðt Þ at 10 late-time cycles. The average
Z^ 1 ð0ÞC 15. S. M. Albrecht et al., Nature 531, 206–209 (2016).
have higher weights
Y in longer Pauli strings, with the dominant
L 1^ ^
string being ^ L up to
X j Y L . This is equivalent to Z
16. P. Yu et al., Nat. Phys. 17, 482–488 (2021). j¼1
value of each operator and the normalization Y
the parity transformation Lj¼1 X ^ F . In
^ j , which commutes with U
17. G. Goldstein, C. Chamon, Phys. Rev. B Condens. Matter
condition allow us to determine the ideal val- Mater. Phys. 84, 205109 (2011). ^ R is expressed in this alternative form and has a
Eq. 3, c
ues of aZ;n and aY;n [see methods and supple- 18. M. Cheng, R. M. Lutchyn, S. Das Sarma, Phys. Rev. B Condens. leading term aZ;L Z^ L . Such a basis requires measurements of
Matter Mater. Phys. 85, 165124 (2012). shorter Pauli strings.
mentary text section VI (33) for details].
19. J. C. Budich, S. Walter, B. Trauzettel, Phys. Rev. B Condens. 46. Google Quantum AI, Data and Simulation Code for “Noise Resilient
The experimentally measured coefficients, Matter Mater. Phys. 85, 121405 (2012). Edge Modes on a Chain of Superconducting Qubits,” version 1,
shown in Fig. 4B for four values of g, both 20. C. Knapp, T. Karzig, R. M. Lutchyn, C. Nayak, Phys. Rev. B 97, Zenodo (2022); [Link]
oscillate in sign and decay exponentially as 125404 (2018).
AC KNOWLED GME NTS
n moves away from the edge. The decay rate 21. E. M. Levenson-Falk, F. Kos, R. Vijay, L. Glazman, I. Siddiqi,
Phys. Rev. Lett. 112, 047002 (2014). We have benefited from discussions with M. H. Devoret, L. G. Dias,
is also observed to decrease as g approaches 22. B. Feldman et al., Nat. Phys. 13, 286–291 (2016). I. K. Drozdov, P. Ghaemi, and A. Rahmani. Funding: D.B. is a CIFAR
the critical value g = 0.5. This is a result of the 23. K. Serniak et al., Phys. Rev. Lett. 121, 157701 (2018). Associate Fellow in the Quantum Information Science Program.
fact that the decay constant for the coeffi- 24. M. Hays et al., Phys. Rev. Lett. 121, 047001 (2018). Author contributions: D.A.A. and [Link]. conceived of the project.
25. R. Blatt, C. F. Roos, Nat. Phys. 8, 277–284 (2012). X.M., D.A.A., and [Link]. designed the experiment. X.M. executed
cients is the localization length x of the MEMs, 26. C. Gross, I. Bloch, Science 357, 995–1001 (2017). the experiment. P.R., X.M., D.A.A., [Link]., and M.Y.N. performed
which diverges at g = 0.5 [see supplementary 27. I. Carusotto et al., Nat. Phys. 16, 268–279 (2020). analysis of the experimental results. K.W.L. and B.F. contributed to
text section III (33)]. A comparison between 28. E. Lieb, T. Schultz, D. Mattis, Ann. Phys. 16, 407–466 (1961). measurements in the supplementary materials. X.M., P.R., and D.A.A.

theoretical and experimental values of aZ;n
29. L. S. Levitov, T. P. Orlando, J. B. Majer, J. E. Mooij, wrote the manuscript. [Link]. and P.R. led and coordinated the project.
arXiv:cond-mat/0108266 [[Link]-hall] (2001). Infrastructure support was provided by Google Quantum AI.
is shown in the bottom panels of Fig. 4B, where 30. J. Q. You, Z. D. Wang, W. Zhang, F. Nori, Sci. Rep. 4, 5535 (2014). All authors contributed to revising the manuscript and the
good agreement is found over a span of nearly 31. J.-S. Xu et al., Sci. Adv. 4, eaat6533 (2018). supplementary materials. Competing interests: The authors
32. M. I. Dykman, Phys. Rev. A 100, 042101 (2019). declare no competing interests. Data and materials availability:
three orders of magnitude. 33. Materials and methods are available as supplementary materials. Experimental data shown in the main text and supplementary
In this study, we simulate MEMs using a 34. V. Khemani, A. Lazarides, R. Moessner, S. L. Sondhi, Phys. Rev. Lett. materials, as well as simulation software code, are available in Zenodo
system of driven transmon qubits and com- 116, 250401 (2016). (46). License information: Copyright © 2022 the authors, some
35. M. Thakurathi, A. A. Patel, D. Sen, A. Dutta, Phys. Rev. B rights reserved; exclusive licensee American Association for the
prehensively study their symmetry protection Advancement of Science. No claim to original US government works.
Condens. Matter Mater. Phys. 88, 155133 (2013).
against noise in their solid-state environment. 36. D. J. Yates, F. H. L. Essler, A. Mitra, Phys. Rev. B 99, 205419 (2019). [Link]
We find that the degree of protection sensi- 37. P. Fendley, J. Phys. A Math. Theor. 49, 30LT01 (2016).
38. D. Abanin, W. De Roeck, W. W. Ho, F. Huveneers, SUPPLEMENTARY MATERIALS
tively depends on the physical characteristic
Commun. Math. Phys. 354, 809–827 (2017). [Link]/doi/10.1126/science.abq5769
of the noise and generally does not extend to 39. T. Mori, T. Kuwahara, K. Saito, Phys. Rev. Lett. 116, 120401 (2016). Materials and Methods
noise that breaks the underlying symmetry, 40. D. V. Else, B. Bauer, C. Nayak, Phys. Rev. X 7, 011026 (2017). Supplementary Text
such as T1 decay of the transmon qubits. We 41. D. V. Else, P. Fendley, J. Kemp, C. Nayak, Phys. Rev. X 7, Figs. S1 to S15
041062 (2017). References (47–60)
also find that, because of a prethermalization 42. P. Roushan et al., Science 358, 1175–1179 (2017).
mechanism, the MEMs in our system are pro- 43. N. Harle, O. Shtanko, R. Movassagh, arXiv:2203.15083 Submitted 17 April 2022; accepted 21 October 2022
tected against certain noise that seemingly [quant-ph] (2022). 10.1126/science.abq5769
violates ℤ2 symmetry, for example, local Z^ fluc-
tuations. These results highlight the complex
interplay between physical noise and protection CHEMICAL PHYSICS
and indicate the crucial importance of testing
symmetry against open-system dynamics in any Cavity-enabled enhancement of ultrafast
experimental platform. Furthermore, we find
that even in the presence of decoherence, the intramolecular vibrational redistribution
Pauli expansion of conserved quantities such as
MEMs can be accurately determined by mea- over pseudorotation
suring and renormalizing late-time expectation
values of Pauli operators. This error-mitigation Teng-Teng Chen1†, Matthew Du1†, Zimo Yang2†, Joel Yuen-Zhou1*, Wei Xiong1,2,3*
strategy may be applied to study integrals of mo-
tion in physical models that are more difficult to Vibrational strong coupling (VSC) between molecular vibrations and microcavity photons yields a few
compute classically. Preliminary results on non- polaritons (light-matter modes) and many dark modes (with negligible photonic character). Although VSC is
integrable dynamics are shown in fig. S7. reported to alter thermally activated chemical reactions, its mechanisms remain opaque. To elucidate this
problem, we followed ultrafast dynamics of a simple unimolecular vibrational energy exchange in iron
RE FE RENCES AND N OT ES pentacarbonyl [Fe(CO)5] under VSC, which showed two competing channels: pseudorotation and
1. P. Zanardi, M. Rasetti, Phys. Rev. Lett. 79, 3306–3309 (1997). intramolecular vibrational-energy redistribution (IVR). We found that under polariton excitation, energy
2. D. A. Lidar, I. L. Chuang, K. B. Whaley, Phys. Rev. Lett. 81,
2594–2597 (1998).
exchange was overall accelerated, with IVR becoming faster and pseudorotation being slowed down. However,
3. D. Bacon, J. Kempe, D. A. Lidar, K. B. Whaley, Phys. Rev. Lett. dark-mode excitation revealed unchanged dynamics compared with those outside of the cavity, with
85, 1758–1761 (2000). pseudorotation dominating. Thus, despite controversies around thermally activated VSC modified chemistry,
4. A. Y. Kitaev, Ann. Phys. 303, 2–30 (2003).
5. C. Nayak, S. H. Simon, A. Stern, M. Freedman, S. Das Sarma,
our work shows that VSC can indeed alter chemistry through a nonequilibrium preparation of polaritons.
Rev. Mod. Phys. 80, 1083–1159 (2008).

V
6. A. G. Fowler, M. Mariantoni, J. M. Martinis, A. N. Cleland,
Phys. Rev. A 86, 032324 (2012).
ibrational strong coupling (VSC) gives rise nipulating chemical reactions in condensed
7. A. Y. Kitaev, Phys. Uspekhi 44, 131–136 (2001).
8. N. Read, D. Green, Phys. Rev. B Condens. Matter 61, 10267–10297 to delocalized superpositions of molecular phases (3–5). Extensive experimental evidence
(2000). vibrations and electromagnetic modes has shown that without photoexcitation, re-
9. R. M. Lutchyn, J. D. Sau, S. Das Sarma, Phys. Rev. Lett. 105, (cavity modes), known as molecular action rates can be either accelerated or dece-
077001 (2010).
10. [Link], G. Refael, F. von Oppen, Phys. Rev. Lett. 105, 177002 (2010). vibrational polaritons (1, 2). Recently, lerated by VSC and reaction selectivity can
11. V. Mourik et al., Science 336, 1003–1007 (2012). VSC has become a promising method for ma- even be altered (3–5). Although much effort

790 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

has been devoted to providing a sound expla- celerate the overall 2D IR cross-peak dynamics, ilar to the dynamics of molecules outside the
nation for VSC-modified chemistry, consensus and more notably, the excited polaritons make cavity. Thus, the fundamental concept of VSC-
between theory and experiments is still mis- IVR faster than pseudorotation (bottom of modified chemistry—that polaritons can change
sing (6–11). Though it is clear that polaritons Fig. 1A). By contrast, the dynamics triggered reactions—holds. However, because dark modes
are different from bare molecular states, and by exciting the dark reservoir modes were sim- are statistically dominant, the overall influence
thereby have the potential to modify chemistry,
dark modes, which greatly outnumber polar-
itons, have the same excitation energies as the
uncoupled vibrational modes (12). Hence, some A B
theoretical work predicts that reactivity in the
collective VSC regime is similar to that outside

2050
Pseudorotation
an optical cavity (6, 8, 9, 13), which is con- UP
sistent with a few recent experimental results (2045)

Wavenumber (cm-1)
IVR
a2”
that show no modification of reactions from
VSC (10, 14). Resolution of the discrepancies is Cavity MP (2022)
(2012) (2014)

2000
hindered by the following factors: Most reac- e’
tions studied so far are quite complex, that is, (1999)
involve multiple steps or are diffusion limited; LP
and reactive events involving dark modes and (1976)
polaritons simultaneously are probed together

1950
(3–5). To delineate the effect of VSC, it is there- Pseudorotation
fore critical to study elementary reactions and
use a technique that can differentiate the con- IVR
tributions from polaritons and dark modes.
Here, we used ultrafast two-dimensional
infrared (2D IR) spectroscopy to follow how
the polaritons and dark modes evolve in both C D
pseudorotation and intramolecular vibrational- LP UP Cross Peaks Fitting Result
energy redistribution (IVR) of Fe(CO)5 (15). -0.7
a2” Excitation

Normalized Intensity
We did so in a state-resolved manner, thereby -0.75 Outside Cavity
MP a2” Excitation
meeting the desired criteria listed earlier. Fe -0.8
Inside Cavity
(CO)5 features two IR-active vibrational bands, 500 -0.85 Excitation
a doubly degenerate e′ mode at 1999 cm−1 that 2060 [ ,
UP LP
][ UP
, MP
] UP
Inside Cavity
-0.9
involves three equatorial CO groups, and an
0 -0.95
a2″ mode at 2022 cm−1 that involves the axial 2040
-1
(cm-1)

CO groups (fig. S1). Harris and co-workers

applied 2D IR spectroscopy and showed that 2020 -500 2 4 6 8 10 12
Fe(CO)5 can rearrange from its D3h equilib- time (ps)
1

rium geometry to a C4v transition state and 2000 t2 = 30 ps

-1000
back to D3h, during which the equatorial and Diagonal Peaks Fitting Result
-0.15
axial CO ligands interconvert, leading to vibra- 1980
Normalized Intensity
tional energy exchange between the a2″ and e′ -1500 -0.2
modes (16). This process, referred to as Berry’s 1960
pseudorotation, is a single barrier crossing and a2” Excitation
-0.25
Intensity

0
Outside Cavity
thus represents the essence of elementary re-
-2000 a2” Excitation
actions, although the product is indistinguishable -0.3
Inside Cavity
-4000
from the reactant (17). Given that Berry’s Excitation
UP
pseudorotation competes with IVR between a2″ Cut at 1
= UP
-6000 -0.35
Inside Cavity
and e′ modes whose transition dipoles are -8000
perpendicular to each other (Fig. 1A, top), Fe 1960 1980 2000 2020 2040 2060 2 4 6 8 10 12

(CO)5 is an ideal testbed to understand how (cm-1) time (ps)

3
VSC affects single barrier-crossing events and
the branching ratio between various dynami- Fig. 1. Influence of VSC on Fe(CO)5 energy-exchange dynamics. (A) Schematic drawing showing that
cal processes. when Fe(CO)5 is outside of the cavity, pseudorotation is the dominating channel (top); when the molecule
Using 2D IR spectroscopy, we found that is placed in an optical cavity, IVR becomes the dominant energy-exchange process and pseudorotation
when polaritons were pumped, they could ac- is suppressed (bottom). (B) Strong coupling diagram and IR spectrum of Fe(CO)5 inside the cavity.
(C) Normalized 2D IR spectrum using the linear spectrum of strongly coupled Fe(CO)5 at waiting time
(t2) = 30 ps in dodecane (blue and red boxes represent [wUP, wLP] and [wUP, wMP] peaks, respectively),
1
Department of Chemistry and Biochemistry, University of
California, San Diego, La Jolla, CA, USA. 2Materials Science
along with the corresponding linear spectrum (top) and normalized narrowband pump probe spectrum at
and Engineering Program, University of California, San Diego,
La Jolla, CA, USA. 3Department of Electrical and Computer w1 = wUP (bottom). (D) Experimental dynamics of cross-peaks (top) and diagonal peaks (bottom) for Fe(CO)5
Engineering, University of California, San Diego, La Jolla, outside the cavity upon pumping of the a2″ modes (red dots) and inside the cavity upon pumping of
CA, USA. the UP (blue dots) and the a2″ dark modes (gray dots). The black dashed, dotted, and solid lines are the
*Corresponding author. Email: joelyuen@[Link] (J.Y.-Z.);
w2xiong@[Link] (W.X.) corresponding fits. Energy is exchanged at a faster rate when pumping the UP, whereas pumping the a2″
†These authors contributed equally to this work. dark modes leads to a rate similar to that outside the cavity.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 791

RES EARCH | R E P O R T S

of VSC on the dynamics of Fe(CO)5, when itated by pseudorotation and IVR. At a first vided a good fit to the [ωUP, ωMP] and [ωUP,
measured without differentiating polariton and glance, the dynamics of energy transfer under ωLP] dynamics (Fig. 2B). Furthermore, the
dark modes (e.g., without optical pumping at VSC (reported by [ωUP, ωMP] and [ωUP, ωLP] measured dynamics could be separated into
room temperature), should be negligible. peak dynamics) were a bit faster than those three components: polariton relaxation to dark
The VSC condition was achieved by placing of outside the cavity (reported by the [2022, modes at short times (cavity leakage is implic-
a solution of Fe(CO)5 in n-dodecane into a 2010] diagonal and [2022, 1986] cross-peak itly accounted for; see SM section 2.3.2), energy
Fabry-Pérot microcavity. Unless specifically dynamics) (Fig. 1D). exchange at intermediate times, and vibra-
mentioned, we set the Fe(CO)5 concentration To quantify the energy-exchange dynamics tional decay at long times. From the fitted re-
to ~180 mM and the cavity longitudinal thick- upon pumping the UP, we used the kinetic sults, kex under VSC was 0.113 ± 0.009 ps−1 at
ness to ~12.5 mm. The e′ and a2″ vibrational model shown in Fig. 2A. First, the population room temperature, 30% faster than that out-
modes of Fe(CO)5 strongly couple to a fifth- of the UP relaxed to both the a2″ and e′ dark side the cavity.
order cavity mode. The IR spectrum (Fig. 1B) modes within the polariton lifetime. Then, the Using 2D IR and the same analysis, we
shows the transitions of upper, middle, and a 2″ and e′ dark modes exchanged energy, found when exciting the a2″ dark reservoir
lower polaritons (UP, MP, and LP) at a reso- through pseudorotation and IVR, and at the modes directly, the energy-exchange dynamics
nant frequency (ω) of ωUP = 2045 cm−1, ωMP = same time dissipated energy to their environ- had similar trends as those outside the cavity
2014 cm−1, and ωLP = 1976 cm−1, respectively. ment. The solution of this kinetic model pro- (Fig. 2C) and kex was 0.090 ± 0.006 ps−1.
By fitting to a coupled oscillator model [see Clearly, VSC was only modifying the dynamics
supplementary materials (SM) section 2.1], we when the polaritons were pumped, whereas
determined that the cavity mode is 2013 cm−1 A pumping the a2″ dark reservoir modes caused
and it interacts with the e′ and a 2″ modes Vibrational
the system to evolve similarly to the molecules
with amplitudes gcav-e′ = 26 cm−1 and gcav-a2″ = decay kR Polariton Vibrational decay kR
outside the cavity, agreeing with the reservoir’s
19 cm−1, respectively. Because the full widths UP relaxation kpa a ” purely molecular character. Similar findings
2
at half maximum of the e′ and a2″ modes are MP Energy have been predicted by a recent theoretical
e’
8 and 5 cm−1, respectively, and that of the cavity LP Polariton exchange kex
work (24). The contrast of dynamics between
mode is 11 cm−1, the samples satisfy the criteria relaxation kpe
Vibrational decay kR
pumping the UP and dark a2″ modes suggests
Pump
for VSC (18). 0 that the energy-exchange rates depend on the
We used 2D IR to monitor the cross-peak B initial populated states.
dynamics both outside and inside the cavity 1
Expt. data
Although we have shown that VSC leads to
[ , ] peak fitting
faster energy exchange between the a2″ and
Normalized Intensity

(19–22). Through pseudorotation and IVR, UP LP Fitted data

Energy transfer
the a2″ and e′ modes can exchange energy. By 0
Vibrational decay
e′ modes, this acceleration could be due to
measuring the dynamics of the [2022, 1986] Polariton relaxation
an enhancement of either pseudorotation or
cross and [2022, 2010] diagonal peaks of the IVR. To qualitatively distinguish between the
1→2 transitions (fig. S6) and fitting them to 0.2 [ UP
, MP
] peak fitting two processes, we measured the vibrational
a kinetic model (see SM section 2.3.1), we de- 0 anisotropy (25) associated with the cross-peaks.
termined the energy-exchange rate constant -0.2
IVR involved energy transfer between e′ and a2″
kex to be 0.084 ± 0.002 ps−1 at 25°C (Fig. 1D). modes that were perpendicular to each other
-0.4
Here, unless specifically mentioned, all mea- 5 10 15 20 25 30 (Fig. 3A), and pseudorotation caused energy
surements were done under magic angle con- time (ps) exchange between e′ and a2″ modes that were
ditions to remove contributions from rotational C parallel to each other (Fig. 3B). Thus, the ani-
-0.7
dynamics. sotropy should start from −0.2 and 0.4 for the
Normalized Intensity

Similarly, for Fe(CO)5 under VSC (Fig. 1C), -0.8

[2022, ] peak fitting former and latter (25, 26), respectively. In gen-
we followed the dynamics of the [ωUP, ωMP]
LP
-0.9 eral, both processes occurred concurrently, and
(red box in Fig. 1C) and [ωUP, ωLP] (blue box the initial anisotropy was between these values.
in Fig. 1C) peaks of the 2D IR spectrum (or Outside the cavity, the initial value of ani-
-0.02
the corresponding narrowband pump probe sotropy was ~0.06 for exciting the a2″ modes
spectra; see SM section 1.3 for details). Here, -0.03 (Fig. 3C), suggesting that pseudorotation
we specifically focused on the dynamics in- [2022, MP
] peak fitting dominated over IVR. However, under VSC, the
-0.04
volving pumping the UP to avoid complications opposite trend was observed: The anisotropy
5 10 15 20 25 30
of hot (i.e., highly excited) vibrational states time (ps) started at about −0.08 for exciting the UP (Fig.
when exciting the LP mode (23). The inter- 3D). This contrast indicates that under VSC,
pretation of these peaks was discussed in our Fig. 2. Energy-exchange dynamics between a2″ IVR dominated over pseudorotation. Not sur-
previous work (20). Basically, the polariton and e′ modes. (A) Schematic drawing of the prisingly, when pumping the a2″ dark modes
transitions at ωMP and ωLP overlap with kinetic model for Fe(CO)5 under VSC. See SM under VSC, the anisotropy was ~0.06 (fig. S18),
the 1→2 transition of the a2″ and e′ modes, section 2.3.2 for details of the kinetic model. kpe similar to the cavity-free case.
respectively [this assignment is further con- (kpa), the rate constant for the polariton relaxation To determine the rate constants of pseudo-
firmed by spectral simulations (SM section from UP to the e′ (a2″) modes; kex, the energy rotation and IVR, a more-detailed kinetic model
2.6) and input-output theory (SM section 4)]. transfer rate constant between the e′ and a2″ was developed (see SM section 2.4) (25, 26).
Upon exciting the UP, when the waiting time modes; kR, the rate constant for the vibrational The anisotropy can be calculated based on the
was beyond the polariton lifetime, the [ωUP, decay. (B) Experimental data (blue dots) and fits energy-exchange dynamics simulated from the
ωMP] and [ωUP, ωLP] peaks corresponded, (black dotted lines) including each component for kinetic model. By fitting the measured anisot-
respectively, to the excited state population of [wUP, wLP] (top) and [wUP, wMP] (bottom) peaks. ropy dynamics to the kinetic model, we deter-
the a2″ and e′ modes. Therefore, the dynamics (C) Experimental data (red dots) and fits (black dotted mined the rate constants for pseudorotation
of these peaks reported the energy transfer lines) for the [2022, wLP] (top) and [2022, wMP] (kps) and IVR (kIVR) to be 0.035 ± 0.001 and
between the a2″ and e′ modes that was facil- (bottom) peaks when the a2″ dark modes are pumped. 0.024 ± 0.001 ps−1, respectively, outside the

792 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

RES EARCH | R E P O R T S

cavity (Fig. 3C) and 0.022 ± 0.005 and 0.043 ± the reaction. However, the coexistence of IVR altered by VSC (27). This correspondence sup-
0.002 ps−1 under VSC (Fig. 3D). The energy- enhancement and pseudorotation suppression ports that the insights obtained here should be
exchange rate is kex = kps + 2*kIVR because suggests otherwise. By quickly going through relevant to understanding the previous exper-
the pseudorotation adiabatically morphs the IVR, molecules may lose their driving force for iments (28).
a2″ mode into one of the e′ modes but the IVR pseudorotation, leading to the pseudorotation By using 2D IR to resolve ultrafast chem-
causes relaxation of a2″ into both e″ modes slowdown. It is also possible that the pseudo- ical dynamics with specific initial states, we
(see SM, section 2.4.1e). The former results rotation motion is hindered by the phonons quantified the energy-exchange dynamics in
qualitatively agree with previous work show- that are excited by the transition from the UP Fe(CO)5 under VSC. We showed that when the
ing that pseudorotation dominated the dy- to the dark modes. The temperature-dependent UP was excited under VSC, IVR was promoted
namics outside a cavity (16), except that now measurements further showed that VSC shifts and pseudorotation was suppressed compared
we quantified the relative contribution of the thermodynamic parameters of activa- with the bare molecular system. However, pump-
IVR. The quantitative results agreed with the tion in the same direction (see SM section 2.4), ing the dark reservoir modes under VSC led to
qualitative analysis above, indicating that which has also been observed—and rather little change in the dynamics compared with
VSC shifted the balance between pseudorota- consistently—in reports of reaction kinetics those outside the cavity. Because the population
tion and other energy-exchange channels: Out-
side the cavity, exciting the a2″ mode yielded
dynamics where pseudorotation dominated A e’ e’ a2”
over IVR, and under VSC, exciting the UP
promoted IVR and suppressed pseudorota-
tion; yet, under VSC, exciting the a2″ dark
modes did not change the dynamics relative to
molecules outside the cavity. We note that this
effect is VSC-exclusive, because weak coupling
to the cavity did not lead to the modification of
the dynamics (see SM section 3.4); further, the
acceleration of energy transfer and promotion
of IVR through VSC was robust against dif-
ferent solvent environments (see results for B
1-octanol, SM section 3.9). Pseudorotation
The sharp contrast between the VSC dynam-
ics starting in the UP and a2″ dark reservoir D3h C4v D3h
modes is interesting because, even when the
UP was excited, the population relaxed to the
dark modes on a much shorter time scale than
pseudorotation and IVR. The difference then
lies in the relaxation processes available to the
initial states. Several mechanisms could ex-
plain the faster IVR upon pumping the UP. For
example, the decay from the UP to the a2″ dark 120° 104° 90°
modes was accompanied by the excitation of
low-frequency vibrations (i.e., phonons), and
some of these phonons could be further ex-
cited during the energetically downhill IVR
from a2″ to e′ modes. It follows that IVR would C D
be accelerated by the first scattering process,
and this enhancement would not occur if the Anisotropy of Cross Peak outside Cavity Anisotropy of Cross Peak inside Cavity
system were initialized in the a2″ dark modes. 0.06
0
A limitation of this hypothesis is that at room -0.01
0.05 Expt. data Expt. data
Anisotropy

Anisotropy

temperature, the phonons should have a high -0.02

0.04 Fitted data Fitted data
occupation number (≈10), which should not -0.03
substantially change when the UP relax to dark 0.03 kps = -0.04 kps =
modes (through one- or few-phonon excita- 0.02 0.035± 0.001 ps-1 -0.05 0.022± 0.005 ps-1
tion). Another possibility is that the VSC- kIVR = -0.06 kIVR =
0.01
induced speedup in the dynamics reflects a 0.024± 0.001 ps-1 -0.07 0.043± 0.002 ps-1
polariton-induced intermolecular vibrational 0
10 20 30 40 50 60 10 20 30 40 50 60
energy transfer. In this case, though, the ob- time (ps) time (ps)
served anisotropy dynamics should practically
be zero, or have a fast decay (20), instead of Fig. 3. Cross-peak anisotropy dynamics of IVR and pseudorotation. (A) Depiction of the eigenvectors
starting from a negative value, owing to the for the a2″ and doubly degenerate e′ vibrational modes of Fe(CO)5. IVR leads to energy transfer between
lack of orientational correlation between donor modes that are perpendicular to each other. (B) Pseudorotation leads to energy transfer between a2″
and acceptor molecules. Conversely, the mild and e′ modes that are parallel to each other. (C and D) Experimental anisotropy (red line) and corresponding
suppression of pseudorotation is surprising, fits (black dotted line) for the cross-peak of Fe(CO)5 (C) outside the cavity ([2022, 1986]) and (D) under
because it is conventionally thought that these VSC ([wUP, wMP]). Listed are the rate constants, extracted from the fitting, of IVR (kIVR) and pseudorotation
high-frequency vibrational modes do not drive (kps). The rate constants indicate that VSC accelerated IVR and suppressed pseudorotation.

SCIENCE [Link] 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 793

RES EARCH | R E P O R T S

at thermal equilibrium resides predominantly 9. M. Du, J. A. Campos-Gonzalez-Angulo, J. Yuen-Zhou, J. Chem. AC KNOWLED GME NTS
in the dark modes, the overall influence of VSC Phys. 154, 084108 (2021). We thank H. H. Bhakta for assisting with the Mathematica code.
10. M. V. Imperatore, J. B. Asbury, N. C. Giebink, J. Chem. Phys. Funding: T.-T.C. is supported by National Science Foundation
on Fe(CO)5 should be negligible without exter- 154, 191103 (2021). (NSF) grant DMR-1848215. Z.Y. is supported by Air Force Office of
nal (e.g., laser) pumping. Yet, the present re- 11. X. Li, A. Mandal, P. Huo, Nat. Commun. 12, 1315 (2021). Scientific Research grant FA9550-21-1-0369. W.X. acknowledges
sults show an important insight that unifies 12. J. A. Campos-Gonzalez-Angulo, J. Yuen-Zhou, J. Chem. Phys. general support for summer salaries from Alfred P. Sloan
156, 194308 (2022). Foundation grant FG-2020-12845 and NSF grant CHE-2101988.
the works that report VSC-modified reactions 13. J. A. Campos-Gonzalez-Angulo, J. Yuen-Zhou, J. Chem. Phys. M.D. is supported through the American Chemical Society
and the ones that report or predict the oppo- 152, 161101 (2020). Petroleum Research Fund grant 60968-ND6. J.Y.-Z. was supported
site: Regardless of how reactions behave under 14. G. D. Wiesehan, W. Xiong, J. Chem. Phys. 155, 241103 (2021). by the US Department of Energy, Office of Science, Basic Energy
15. R. S. Berry, J. Chem. Phys. 32, 933–938 (1960). Sciences, Condensed Phase and Interfacial Molecular Science
thermally activated conditions, the basic concept 16. J. F. Cahoon, K. R. Sawyer, J. P. Schlegel, C. B. Harris, Science (CPIMS) program under Early Career Research Program Award
of VSC-modified chemistry works—populated 319, 1820–1823 (2008). DE-SC0019188. Author contributions: W.X. conceived the original
polaritons can influence chemical dynamics. 17. F. H. Westheimer, Acc. Chem. Res. 1, 70–78 (1968). idea and supervised the overall research. T.-T.C., Z.Y., and W.X.
18. P. Törmä, W. L. Barnes, Rep. Prog. Phys. 78, 013901 (2014). designed the experiments. T.-T.C. and Z.Y. conducted the
These findings suggest that the future of VSC-
19. B. Xiang et al., Proc. Natl. Acad. Sci. U.S.A. 115, 4845–4850 experimental work. T.-T.C., M.D., J.Y.-Z., and W.X. analyzed data.
modified thermal chemistry lies in controlling (2018). M.D. and W.X. developed the anisotropy model. M.D. and J.Y.-Z.
the dark modes, either by reducing the num- 20. B. Xiang et al., Science 368, 665–667 (2020). proposed mechanisms and developed the spectral theory in
ber of dark reservoir modes (29) or even going 21. B. Xiang, J. Wang, Z. Yang, W. Xiong, Sci. Adv. 7, eabf6397 SM section 4. T.-T.C., M.D., J.Y.-Z., and W.X. interpreted the
(2021). results and wrote the manuscript. Competing interests: The
to the single-molecule regimes (30), for ex- 22. B. Xiang, W. Xiong, J. Chem. Phys. 155, 050901 (2021). authors declare no competing interests. Data and materials
ample, through cavity miniaturization or by 23. B. Xiang et al., J. Phys. Chem. A 123, 5918–5927 (2019). availability: All data needed to support the conclusions of the
24. D. Wellnitz, G. Pupillo, J. Schachenmayer, Commun. Phys. 5, main text and supplementary materials have been uploaded to
making dark modes more delocalized through
120 (2022). Zenodo (33). License information: Copyright © 2022 the
heterogeneity (31, 32). 25. R. M. Hochstrasser, Chem. Phys. 266, 273–284 (2001). authors, some rights reserved; exclusive licensee American
26. A. Tokmakoff et al., J. Chem. Phys. 102, 3919–3931 (1995). Association for the Advancement of Science. No claim to original
27. K. Hirai, J. A. Hutchison, H. Uji-I, ChemPlusChem 85, US government works. [Link]
RE FE RENCES AND N OT ES
1981–1988 (2020). licenses-journal-article-reuse
1. J. P. Long, B. S. Simpkins, ACS Photonics 2, 130–136 28. A. D. Dunkelberger, B. S. Simpkins, I. Vurgaftman,
(2015). J. C. Owrutsky, Annu. Rev. Phys. Chem. 73, 429–451 (2022).
2. A. Shalabney et al., Nat. Commun. 6, 5981 (2015). 29. T. E. Li, A. Nitzan, J. E. Subotnik, J. Chem. Phys. 154, 094124 SUPPLEMENTARY MATERIALS
3. A. Thomas et al., Angew. Chem. Int. Ed. 55, 11462–11466 (2016). (2021). [Link]/doi/10.1126/science.add0276
4. A. Thomas et al., Science 363, 615–619 (2019). 30. C. C. Schäfer, J. Flick, E. Ronca, P. Narang, A. Rubio, Materials and Methods
5. F. J. Garcia-Vidal, C. Ciuti, T. W. Ebbesen, Science 373, arXiv:2104.12429 [quant-ph] (2021).
Supplementary Text
eabd0336 (2021). 31. G. D. Scholes, Proc. R. Soc. London Ser. A 476, 20200278
Figs. S1 to S45
6. J. Galego, C. Climent, F. J. Garcia-Vidal, J. Feist, Phys. Rev. X 9, (2020). Tables S1 to S7
021057 (2019). 32. M. Du, J. Yuen-Zhou, Phys. Rev. Lett. 128, 096001 (2022).
References (34–42)
7. T. E. Li, A. Nitzan, J. E. Subotnik, J. Chem. Phys. 152, 234107 33. T.-T. Chen, M. Du, Z. Yang, J. Yuen-Zhou, W. Xiong, Cavity-
(2020). enabled enhancement of ultrafast intramolecular vibrational Submitted 16 May 2022; resubmitted 14 August 2022
8. I. Vurgaftman, B. S. Simpkins, A. D. Dunkelberger, redistribution over pseudorotation, Zenodo (2022); Accepted 26 September 2022
J. C. Owrutsky, J. Phys. Chem. Lett. 11, 3557–3562 (2020). [Link] 10.1126/science.add0276

794 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

Pushing the Boundaries of Knowledge
As AAAS’s first multidisciplinary, open access journal, Science Advances publishes
research that reflects the selectivity of high impact, innovative research you expect
from the Science family of journals, published in an open access format to serve
a vast and growing global audience. Check out the latest findings or learn how to
submit your research: [Link]/journal/sciadv

G O L D O P E N AC C E S S , D I G I TA L , A N D F R E E T O A L L R E A D E R S

[Link] 795 11/10/22 8:21 AM

online @[Link]

Science Careers helps you advance

your career. Learn how !
 Register for a free online account on  Download our career booklets, including
[Link]. Career Basics, Careers Beyond the Bench,
 Search hundreds of job postings and and Developing Your Skills.
find your perfect job.  Complete an interactive, personalized career
 Sign up to receive e-mail alerts about job plan at “my IDP.”
postings that match your criteria.  Visit our Employer Profiles to learn more
 Upload your resume into our database and about prospective employers.
connect with employers.  Read relevant career advice articles from our
 Watch one of our many webinars on library of thousands.
different career topics such as job searching,
networking, and more.

Visit [Link]
today — all resources are free

SCI E N CECA RE E RS.O RG

1118Recruitment_RC.indd 796 11/14/22 12:32 PM

online @[Link]
POSTDOCTORAL RESEARCH ASSOCIATES POSTDOCTORAL RESEARCH FELLOWS
NRC RESEARCH ASSOCIATESHIP PROGRAMS AIR FORCE SCIENCE & TECHNOLOGY
The National Academies of Sciences, Engineering, and Medicine administers FELLOWSHIP PROGRAM
postdoctoral and senior research awards at participating federal laboratories and
affiliated institutions at locations throughout the U.S and abroad. The National Academies of Sciences, Engineering, and Medicine administers
postdoctoral and senior fellowship awards at the U.S. Air Force Research
We are seeking highly qualified candidates who hold, or anticipate earning, a Laboratory (AFRL), the U.S. Air Force Institute of Technology (AFIT), and the
doctorate in a variety of fields of science or engineering. Degrees from foreign U.S. Air Force Academy (USAFA) under the Air Force Science & Technology
universities should be equivalent in training and research experience to a Fellowship Program (AF STFP).
doctoral degree from a U.S. institution. Citizenship eligibility varies among
the sponsoring laboratories. We are seeking highly qualified candidates who are U.S. citizens and hold, or
anticipate earning, a doctorate in a variety of fields of science or engineering.
Application deadline dates (four annual review cycles):
• February 1 Application deadline dates (four annual review cycles):
• May 1 • February 1
• August 1 • May 1
• November 1 • August 1
• November 1
Awardees have the opportunity to:
• Conduct independent research in an area compatible with the interests of the Awardees have the opportunity to:
sponsoring laboratory • Conduct independent research in an area compatible with the interests of the
• Devote full-time effort to research and publication Air Force laboratories
• Access the excellent and often unique facilities of the federal • Devote full-time effort to research and publication
research enterprise • Access the excellent and often unique Air Force research facilities
• Collaborate with leading scientists and engineers at the • Collaborate with leading scientists and engineers
sponsoring laboratories
Awardee benefits:
Awardee benefits include:
• Base stipend starting at $76,542; may be higher based on experience
• Stipends ranging from $45,000 to $97,490; may be higher based
• Health insurance (including dental and vision), relocation benefits, and
on experience
• Health insurance (including dental and vision), relocation benefits, and a professional travel allowance
professional travel allowance
For detailed program information, to search for AFRL, AFIT, and USAFA
For detailed program information, to search Research Opportunities, and to Research Opportunities, and to contact prospective Research Adviser(s), visit
contact prospective Research Adviser(s) visit [Link]/rap. [Link]/afstfp.

ARE YOU THE KIND

WHO WORKS TO HELP The University of Missouri (MU) Department of Biochemistry invites
applications for a tenure-track faculty position in plant radiochemistry.

ALL HUMANKIND? We seek candidates working at the interface of plant metabolism and
radiochemistry that leverage the generation of radio-isotopes at the University
of Missouri Research Reactor (MURR). The faculty member is expected to
develop a robust and externally funded research program exploiting the
resources and-, capabilities; of MURR to address fundamental questions in
plant biology. Examples include, studies of plant carbon or nitrogen dynamics
utilizing radioactive 11C or 13N, micronutrient assimilation and/or exchange
between plants, soil, and microorganisms, as well as metabolic ﬂux and
regulation using isotope ratio analyses. Applications at the Assistant Professor
level are encouraged. Candidates at a higher level will be considered if there
is a record of extramural research funding, publications, teaching and service
expected of tenured faculty at a research-intensive university.

The department has an undergraduate and graduate teaching mission in the

College of Agriculture and Natural Resources (CAFNR). The successful
candidate will be a member of the Interdisciplinary Plant Group (IPG), a
world-renowned group of researchers from seven departments/units with
interests in plant biology. He/she will also be a core faculty member of MURR,
which operates the highest power (10 MW) university research reactor in the US
and is planning a further expansion. For additional information, visit https://
[Link] and [Link]

APPLICATION: To apply, visit [Link]

positions/, click on Open Positions and submit: (1) a cover letter;
(2) curriculum vitae; (3) a one-page summary of research accomplishments
with two pages of future plans; (4) a one-page narrative of teaching philosophy;
Find your next job at [Link] (5) a description of current and future efforts to expand diversity & inclusion in
science; and (6) names and contact information of four references. Review of
applications will begin on December 15, 2022 and will continue until the
position is ﬁlled. For information, contact biochemchair@[Link].

The University of Missouri is an Equal Access, Equal Opportunity, and

Afﬁrmative Action Employer. We are fully committed to achieving the goal
of a diverse and inclusive academic community of faculty, staff and students.
We seek individuals who are committed to this goal and our core campus
values of respect, responsibility, discovery and excellence.

1118Recruitment_RC.indd 797 11/14/22 12:32 PM

WORKING LIFE
By Walter P. Suza

Silent no longer

F
ear kept me silent. When I was a child in Tanzania, kids from a different tribe would wait for me at
the end of the school day to beat me up. The adrenaline rush of fear may have sometimes helped
me outrun my nemeses. Yet it also kept me silent. I feared that if I told anyone at home, my older
siblings would try to intervene, which would only make future beatings worse. That was many
years ago. The United States is my country now. But I am still afraid. To America, I am not just a
man; I am a Black man. And for years, the baggage of that label kept me afraid and silent.

Since coming to the United States as the Confederate flag being flown.
a graduate student, I have been sub- Still, I remained silent, unable to
ject to microaggressions and worse. overcome my fear. Instead of speak-
On the first day of my biochemis- ing out, I dove into diversity and in-
try class, a white professor asked, clusion work. I sought professional
“Are you sure you are in the right development to learn how to make
class?” as I was walking in the door. my classroom more inclusive, hosted
Outside a restaurant and sports bar a student from an internship pro-
off-campus, a white man called me gram intended in part to increase
the N-word. During my postdoctoral diversity, and served on department
training, a white colleague giving a committees to strengthen diversity,
lab tour to a white family of a pro- equity, inclusion, and belonging. I
spective student asked them, “Have also read about the history of the
you seen the Black guy?” I assume he U.S. civil rights struggle.
did not see me sitting in the corner. Over time, these efforts helped
Despite the grave reality that me gain the courage to write about
such dehumanizing language can
pave the way for death at the hands
“Despite the threats, racial and social justice for the stu-
dent newspaper. Several students
of the police and vigilantes, I was I continue to write.” and colleagues offered encouraging,
afraid to speak about it in public. supportive responses, which helped
Professionally, I was afraid if I spoke, potential employers reduce my fear of career repercussions. But I was still afraid
might view me as a troublemaker and I would face greater to speak out to the public at large.
barriers to pursuing my goal of becoming a tenured profes- The tipping point came when Ahmaud Arbery was shot
sor. Personally, I feared it might jeopardize my chances of dead by white men and George Floyd was choked to death by
being awarded permanent residency and eventually becom- a white police officer. My rage and despair were nearly over-
ing a U.S. citizen. So, I kept my head down and focused on whelming. I was still afraid. But by that time I had become a
my work and my family. U.S. citizen, and I felt I could no longer stay silent. I owed it to
The stress from not speaking out felt constant and myself, and to those who were more vulnerable, to speak out.
inescapable. I felt trapped inside my own skin. So, I started to write opinion columns for local newspapers.
I was afraid to speak about being followed in grocery Some readers encourage me to keep writing. Some tell me
stores. I was afraid to speak about a white woman at the to “get out of my state.” Some make phone calls to my work,
park ordering her child not to play with my daughter. I was calling me a racist. Some send disturbing threats to my
afraid to speak about being seated in the back of the restau- home and work. Despite the threats, I continue to write.
rant when we were the only family there. I am still a Black man in America, with all the resulting
I remained silent after a white vigilante shot and killed challenges and fears. I am still afraid. But I have learned
Trayvon Martin. I remained silent after a white police of- that fear need not mean silence. Instead, I can channel the

ILLUSTRATION: ROBERT NEUBECKER

ficer shot and killed Michael Brown. I remained silent after adrenaline—not to outrun and escape my fear, but to face
another white police officer shot and killed Tamir Rice. it head on, in the hopes that someday I and others like me
I moved on to a non–tenure-track faculty position, pleased will no longer have cause to feel afraid. j
to be progressing in my career. But I could not outrun rac-
ism. On campus, incidents of white supremacist symbols and Walter P. Suza is an adjunct associate professor at Iowa State University.
racist graffiti were frequently reported. Off campus, I saw Send your career story to SciCareerEditor@[Link].

798 18 NOVEMBER 2022 • VOL 378 ISSUE 6621 [Link] SCIENCE

1118WL_16180857.indd 798 11/11/22 4:49 PM

QUALITY CONTENT FOR THE GLOBAL SCIENTIFIC COMMUNITY
Multiple ways to stay informed on issues related to your research

Scan the code and start exploring

the latest advances in science and
technology innovation!
[Link]/custom-publishing

Brought to you by the Science/AAAS Custom Publishing Office.

Posters Podcasts Sponsored Advertorials Webinars

Collection Booklets

Apply by 24 January 2023 at [Link]/noster

[Link] 800 11/10/22 8:21 AM

Common questions