a950fa_Science___22_October_2026_ebooksyard — 2026-10-24

AAAS

Science

AI IN MEDICINE

Guiding a clinical workup to reach a diagnosis p. 423

---

22 OCTOBER 2026

The FBI agent who sparked NIH's China investigations p. 378

Stopping an ice age in its tracks pp. 381 & 455

Science advocacy finds an ally in the arts p. 386

---

CASE WESTERN RESERVE UNIVERSITY

Learn more about our Faculty 100 Initiative on page 460 #

---

fastest-growing AAU research university

Top

in the nation

— TIME Magazine

CONTENTS

22 OCTOBER 2026 | VOLUME 394 | ISSUE 6822

---

381 & 455

The edge of the Barnes Ice Cap, Baffin Island, Canada, and its position of maximum extent around 1960

---

EDITORIAL

363 The AI debate is missing half the system

—P. Michelucchi

---

NEWS

364 Vaccine trial is a ray of hope in Congo's Ebola epidemic

Study will first test “mismatched” vaccine but Bundibugyo-specific shots may be added soon

—K. Kupferschmidt

366 Across fields, AI finds solutions hiding in plain sight "BootLoops" software enabled researchers to quickly find and attack problems in 22 disciplines —C. Zhao

367 How the domestication of a killer bacterium led to Botox

Scientists inadvertently turned C. botulinum into a tool for medicine and the beauty industry, study finds —J. Cohen

368 Microwave observatory peers into the infant universe In Chile, Simons Observatory begins hunt for galaxy clusters and signs of cosmic inflation —D. Clery

370 Deep-sea life is still reeling from the BP oil spill

A 13-year study shows

Gulf of Mexico ecosystems still haven't recovered

—E. Stokstad

371 Iron at Earth's heart may take a surprising form

X-ray experiments reveal cubic crystals that could complicate the picture of the inner core

—P. Voosen

372 Fired biologist wins \$6 million in defamation lawsuit

A jury ruled in favor of David Sabatini, dismissing accuser's claims of sexual harassment, retaliation —P. Jacobs

---

FEATURES

374 Eyes on the ball

A “robo ump” that calls balls and strikes in Major League Baseball makes its World Series debut this week.

It has to be near-perfect

—J. Cohen

378 The instigator

A single FBI agent spurred NIH's campaign against foreign intervention on U.S. campuses. Now, he's studying the extent of the problem —J. Mervis

---

COMMENTARY

PERSPECTIVES

381 Has the next ice age already come and gone?

The last remnant of a continental glacier shows that global warming may have forestalled the next glaciation —S. A. Marcott

RESEARCH ARTICLE p. 455

382 Putting glucose metabolism under electronic control

An electronic device controls the redox mechanisms that regulate glucose concentrations in mice —W. E. Bentley and G. F. Payne

RESEARCH SUMMARY p. 400

Two proteins that permeabilize mitochondria promote cell death through distinct mechanisms —C. Muñoz-Pinedo and S. J. Martin RESEARCH SUMMARY p. 401

---

2026's Top Employers: From AI to I do

Science Careers' annual survey spotlights the biotech and pharma leaders earning the highest praise from Science readers.

Read the article on p. 461 and explore these highly-rated employers online at ScienceCareers.org/TopEmployers

Science2026

TOP EMPLOYERS

---

385 The cognitive power of the avian brain

Carrion crows make decisions faster than monkeys and humans —M. Colombo and O. Poutrain-Sorel RESEARCH ARTICLE p. 445

---

BOOKS ET AL.

---

386 The microscope and the metaphor

Science shows us what to do to save the planet—poetry helps us understand why we should —P. Levin and D. Hassler PODCAST

387 The Monarch Bear

—L. Volski

---

LETTERS

388 Indigenous higher education needs support

—B. E. Soares et al.

389 India needs a tree conservation plan —A. Dhyani

389 AI biosecurity must include agriculture

—J. Feldman et al.

---

ANALYSIS

---

POLICY ARTICLE

390 Fiscal discipline without physical resilience

Reforms could shift fiscal discipline and resilience investment into reinforcing forces rather than competing ones

—A. Mishra et al.

---

RESEARCH

---

HIGHLIGHTS

394 From Science and other journals

---

RESEARCH SUMMARIES

---

397 Cancer

Intron retention regulates STAT2 function and associates with immunotherapy response in lung cancer —R. P. Englander et al.

---

398 Solar cells

Decoupled self-assembled monolayer and perovskite deposition for efficient flexible solar modules —Y. Wu et al.

---

399 Artificial intelligence

The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development —H. G. Zhang et al.

---

400 Metabolism

Wireless electrocatalytic closed-loop system for drug-free management of diabesity —D. Maity et al.

PERSPECTIVE p. 382

---

401 Cell biology

BAK and BAX form a multiprotein complex to drive mtDNA efflux and caspase-independent cell death —S. Yang et al.

PERSPECTIVE p. 383

---

RESEARCH ARTICLES

402 Molecular biology

Noncoding transcription controls the developmental dynamics of long-range gene regulation

—P. J. Batut et al.

---

408 Conservation biology

Safeguarding global terrestrial vertebrate species from future sea level rise —Z.-W. Jiang et al.

---

413 Natural products

Natural product biosynthesis by guanidine prenyltransferases in plants and bacteria

—D. Kolling et al.

---

419 Solid-state physics

Thermodynamic origin of the phonon Hall effect in a honeycomb antiferromagnet —Q. Meng et al.

---

423 Machine learning

Sequential diagnosis with language models —H. Nori et al.

---

429 Physics

Electronic bound states in the continuum in a two-dimensional metal —M. Thinel et al.

---

434 Membranes

Dynamic functionality-enabled fast and selective transport in nanochannels —H. Zhang et al.

---

440 2D materials

Visualizing symmetry-broken Chern insulators and their quantum melting —M. He et al.

---

445 Comparative cognition

Ultrarapid decision-making in an avian brain

—P. Schmidbauer and A. Nieder PERSPECTIVE p. 385; PODCAST

---

385 & 445

Crows perform random dot movement discrimination tasks with ultrarapid decision times.

---

CONTENTS

---

ON THE COVER

This image depicts a patient and a physician working together with artificial intelligence (AI) to reach a diagnosis. Researchers developed an AI-based approach to help physicians with real-life clinical workups, starting with the patient's chief complaint, systematically asking questions and deciding what laboratory and imaging tests to order, and then continuing the evaluation on the basis of the answers received. See page 423. Illustration: Jason Solo

---

450 2D materials

Imaging electron-hole asymmetry in the quantum melting of generalized Wigner crystals —E. Berger et al.

---

455 Glacial inception

Late Holocene reexpansion of the Laurentide Ice Sheet —N. E. Young et al.

PERSPECTIVE p. 381

---

WORKING LIFE

462 My hidden homeless past

—S. Raymond

362 Science Staff

460 Science Careers

Science NAAAS

EDITOR-IN-CHIEF Holden Thorp, hthorp@aaas.org

EXECUTIVE EDITOR Valda Vinson

EDITORS, RESEARCH Sacha Vignieri, Jake S. Yeston EDITOR, COMMENTARY Lisa D. Chong

DEPUTY EXECUTIVE EDITOR Lauren Kmec

DEPUTY EDITORS Stella M. Hurtley (UK), Phillip D. Szuromi SENIOR EDITORS Michael A. Funk, Angela Hessler, Di Jiang, Priscilla N. Kelly, Marc S. Lavine (Canada), Bianca Lopez, Mattia Maroso (UK), Yevgeniya Nusinovich, Ian S. Osborne (UK), Joana Osório (UK), Cheri Sirois, H. Jesse Smith, Keith T. Smith (UK), Jelena Stajic, Yury V. Suleymanov, Valerie B. Thompson, Brad Wible ASSOCIATE EDITORS Jack Huang, Sumin Jin, Sarah Ross (UK), Madeleine Seale (UK), Corinne Simonti, Unnati Sonawala SENIOR LETTERS EDITOR Jennifer Sills SENIOR NEWSLETTER EDITOR Christie Wilcox RESEARCH & DATA ANALYST Jessica L. Slater LEAD CONTENT PRODUCTION EDITORS Chris Filiatreau, Harry Jach SR. CONTENT PRODUCTION EDITORS Amelia Beyna, Julia Haber-Katris CONTENT PRODUCTION EDITORS Anne Abraham, Robert French, Nida Masiulis, Avery Shantha, Suzanne M. White SENIOR PROGRAM ASSOCIATE Maryrose Madrid EDITORIAL MANAGER Joi S. Granger EDITORIAL ASSOCIATES Aneera Dobbins, Lisa Johnson, Jerry Richardson, Anita Wynn SENIOR EDITORIAL COORDINATORS Clair Goodhead (UK), Alexander Kief, Kat Kirkman, Ronmel Navas, Isabel Schnaidt, Alice Whaley (UK), Brian White EDITORIAL COORDINATORS Samuel Bates, Daniel Young ADMINISTRATIVE COORDINATOR Karalee P. Rogers ASI DIRECTOR, OPERATIONS Janet Clements (UK) ASI OFFICE MANAGER Carly Hayward (UK) ASI SR. OFFICE ADMINISTRATORS Simon Brignell (UK), Jessica Waldock (UK) COMMUNICATIONS DIRECTOR Meagan Phelan DEPUTY DIRECTOR Matthew Wright SENIOR WRITERS Walter Beckwith, Joseph Cariz, Abigail Eisenstadt WRITER Mahathi Ramaswamy SENIOR COMMUNICATIONS ASSOCIATES Kiara Brooks, Zachary Graber, Mackenzie Williams, Sarah Woods

---

NEWS EDITOR Tim Appenzeller

NEWS MANAGING EDITOR John Travis GLOBAL HEALTH EDITOR Martin Enserink DEPUTY NEWS EDITORS Rachel Bernstein, David Grimm, Eric Hand, David Malakoff (Artificial Intelligence), Michael Price, Kelly Servick, Matthew Warren (International) SENIOR CORRESPONDENTS Daniel Clery (UK), Jon Cohen, Jocelyn Kaiser, Jeffrey Mervis ASSOCIATE EDITORS Michael Greshko, Katie Langin NEWS REPORTERS Jeffrey Brainard, Adrian Cho, Phie Jacobs, Sara Reardon, Jennie Erin Smith, Erik Stokstad, Paul Voosen, Celina Zhao CONSULTING EDITOR Elizabeth Culotta CONTRIBUTING CORRESPONDENTS Vaishnavi Chandrashekhar, Dan Charles, Warren Cornwall, Andrew Curry (Berlin), Ann Gibbons, Kai Kupferschmidt (Berlin), Andrew Lawler, Mitch Leslie, Virginia Morell, Dennis Normile (Tokyo), Catherine Offord, Cathleen O'Grady, Elisabeth Pain (Careers), Charles Piller, Hannah Richter, Zack Savitsky, Richard Stone (Senior Asia Correspondent), Gretchen Vogel (Berlin), Lizzie Wade (Mexico City) INTERNS Laura Martín Agudelo, Ailie McWhinnie, Mona Patterson, Julia Vaz COPY EDITORS Julia Cole (Senior Copy Editor), Hannah Knighton, Cyra Master (Copy Chief) ADMINISTRATIVE SUPPORT Meagan Weiland

---

CREATIVE DIRECTOR Beth Rakouskas

DESIGN MANAGING EDITOR Chrystal Smith GRAPHICS MANAGING EDITOR Chris Bickel PHOTOGRAPHY MANAGING EDITOR Emily Petersen MULTIMEDIA MANAGING PRODUCER Kevin McLean DIGITAL DIRECTOR Kara Estelle-Powers DESIGN EDITOR Marcy Atarod DESIGNER Noelle Jessup SENIOR SCIENTIFIC ILLUSTRATOR Noelle Burgess SCIENTIFIC ILLUSTRATORS Austin Fisher, Kellie Holoski, Ashley Mastin SENIOR GRAPHICS EDITOR Monica Hersher GRAPHICS EDITOR Veronica Penney SENIOR PHOTO EDITOR Charles Borst PHOTO EDITOR Elizabeth Billman SENIOR PODCAST PRODUCER Sarah Crespi SENIOR VIDEO PRODUCER Meagan Cantwell SOCIAL MEDIA STRATEGIST Jessica Hubbard SOCIAL MEDIA PRODUCER Sabrina Jenkins WEB DESIGNER Jennie Pajerowski

---

CHIEF EXECUTIVE OFFICER AND EXECUTIVE PUBLISHER

---

Sudip Parikh

PUBLISHER, SCIENCE FAMILY OF JOURNALS Bill Moran

DIRECTOR, BUSINESS OPERATIONS & ANALYSIS Eric Knott MANAGER, BUSINESS OPERATIONS Jessica Tierney SENIOR MANAGER, BUSINESS ANALYSIS Cory Lipman BUSINESS ANALYSTS Kurt Ennis, Maggie Clark, Isacco Fusi BUSINESS OPERATIONS ADMINISTRATOR Taylor Fisher DIGITAL SPECIALIST Marissa Zuckerman ASSOCIATE DIRECTOR OF PRODUCTION AND PLATFORM Jason Hillman SENIOR MANAGER, PUBLISHING AND CONTENT SYSTEMS Marcus Spiegler CONTENT OPERATIONS MANAGER Rebecca Doshi PUBLISHING PLATFORM MANAGER Jessica Loayza PUBLISHING SYSTEMS SPECIALIST, PROJECT COORDINATOR Jacob Hedrick PRODUCTION LEAD Kristin Wowk

MARKETING DIRECTOR Sharice Collins ASSOCIATE DIRECTOR, MARKETING Justin Sawyers GLOBAL MARKETING MANAGER Allison Pritchard

MARKETING MANAGER Ashley Evans SENIOR MARKETING ASSOCIATE Lorena Chirinos Rodriguez MARKETING COORDINATOR Jonathan Moskowitz SENIOR DESIGNER Kim Huynh

DIRECTOR AND SENIOR EDITOR, CUSTOM PUBLISHING Erika Gebel Berg ASSOCIATE DIRECTOR, CUSTOM PUBLISHING Michael Wheeler ADVERTISING PRODUCTION OPERATIONS MANAGER Deborah Tompkins DESIGNER, CUSTOM PUBLISHING Jeremy Huntsinger SENIOR TRAFFIC ASSOCIATE Christine Hall

DIRECTOR, PRODUCT MANAGEMENT Kris Bishop PRODUCT DEVELOPMENT MANAGER Scott Chernoff ASSOCIATE DIRECTOR, PARTNER MARKETING Shawana Arnold SPJ MARKETING MANAGER Ashley Hylton ASSOCIATE DIRECTOR, PUBLISHING INTELLIGENCE Rasmus Andersen PRODUCT MANAGERS Robert Koepke, Anne Mason PRODUCT ASSOCIATE Caroline Breul

ASSOCIATE DIRECTOR, INSTITUTIONAL LICENSING MARKETING Kess Knight ASSOCIATE DIRECTOR, INSTITUTIONAL LICENSING SALES

Ryan Rexroth INSTITUTIONAL LICENSING MANAGER Nazim Mohammedi, Claudia Paulsen-Young SENIOR MANAGER, INSTITUTIONAL LICENSING OPERATIONS Judy Lillibridge MANAGER, RENEWAL & RETENTION Lana Guz MARKETING ASSOCIATE Byron Campbell SYSTEMS & OPERATIONS ANALYST Ben Teincuff FULFILLMENT ANALYST Aminta Reyes

ASSOCIATE DIRECTOR, INTERNATIONAL Roger Goncalves ASSOCIATE DIRECTOR, US ADVERTISING Stephanie O'Connor US MID WEST, MID ATLANTIC AND SOUTH EAST SALES MANAGER Chris Hoag DIRECTOR, OUTREACH AND STRATEGIC PARTNERSHIPS, ASIA Shoupeng Liu ASI EMEA SALES ADVERTISING MANAGER Sarah Lelarge ADVERTISING SALES ASSISTANT, ROW Victoria Glasbey DIRECTOR OF GLOBAL COLLABORATION AND ACADEMIC PUBLISHING RELATIONS, ASIA Xiaoying Chu ASSOCIATE DIRECTOR, INTERNATIONAL COLLABORATION Grace Yao SALES MANAGER Danny Zhao MARKETING MANAGER Kilo Lan ASCA CORPORATION, JAPAN Rie Rambelli (Tokyo), Miyuki Tani (Osaka)

DIRECTOR, COPYRIGHT, LICENSING AND SPECIAL PROJECTS Emilie David RIGHTS AND PERMISSIONS ASSOCIATE Elizabeth Sandler LICENSING ASSOCIATE Virginia Warren RIGHTS AND LICENSING COORDINATOR Dana James

---

EDITORIAL

science\_editors@aaas.org

science\_news@aaas.org

INFORMATION FOR AUTHORS

science.org/authors/

science-information-authors

REPRINTS AND PERMISSIONS

science.org/help/

reprints-and-permissions

MULTIMEDIA CONTACTS

SciencePodcast@aaas.org

ScienceVideo@aaas.org

MEDIA CONTACTS

scipak@aaas.org

PRODUCT ADVERTISING & CUSTOM PUBLISHING

advertising.science.org

science\_advertising@aaas.org

CLASSIFIED ADVERTISING

advertising.science.org/

science-careers

advertise@sciencecareers.org

JOB POSTING CUSTOMER SERVICE

employers.sciencecareers.org

support@sciencecareers.org

MEMBERSHIP AND

INDIVIDUAL SUBSCRIPTIONS

science.org/subscriptions

---

MEMBER BENEFITS

MEMBER BENEFITS

aaas.org/membership/benefits

INSTITUTIONAL SALES AND SITE LICENSES science.org/librarian

AAAS BOARD OF DIRECTORS

CHAIR Joseph S. Francisco

IMMEDIATE PAST PRESIDENT

Theresa A. Maldonado

PRESIDENT

Marina Picciotto

PRESIDENT-ELECT

Eric D. Isaacs

TREASURER Mark F. Dingfield

COUNCIL CHAIR

Ichiro Nishimura

CHIEF EXECUTIVE OFFICER

Sudip Parikh

BOARD

Carolyn N. Ainslie

Morton Ann Gernsbacher

Kathleen Hall Jamieson

Jennifer Lodge

Babak Parviz

Gabriela Popescu

Juan S. Ramírez Lugo

Susan M. Rosenberg

Vassiliki Betty Smocovitis

Roger Wakimoto

Takuzo Aida, U. of Tokyo

---

BOARD OF REVIEWING EDITORS (Statistics board members indicated with S)

Leslie Aiello, Wenner-Gren Fdn.

Anastassia Alexandrova, UCLA

James Analysis, UC Berkeley

Amy Ando, The Ohio State U.

Paola Arlotta, Harvard U.

Madan Babu, St. Jude

Jennifer Balch, U. of Colorado

Nenad Ban, ETH Zürich

Carolina Barillas-Mury, NIH, NIAID

Christopher Barratt, U. of Dundee

François Barthelat,

U. of Colorado Boulder

Joseph Bass, Northwestern U.

Franz Bauer,

Universidad de Tarapacá

Andreas Baumler, UC Davis

Carlo Beenakker, Leiden U.

Camilla Bellone, Geneva U.

Sarah Bergbreiter,

Carnegie Mellon U.

Kiros T. Berhane, Columbia U.

Aude Bernheim, Inst. Pasteur

Joseph J. Berry, NREL

Aditi Bhargava, UCSF

Dominique Bonnet,

Francis Crick Inst.

Chris Bowler,

École Normale Supérieure

Ian Boyd, U. of St. Andrews

Malcolm Brenner,

Baylor College of Med.

Ron Brookmeyer, UCLA (\$) Christian Büchel, UKE Hamburg

Johannes Buchner, TUM

Dennis Burton, Scripps Res.

Carter Tribley Butts, UC Irvine

Erin Calipari, Vanderbilt U.

Annmarie Carlton, UC Irvine

Jane Carlton, Johns Hopkins U.

Pedro Carvalho, U. of Oxford

Simon Cauchemez, Inst. Pasteur

Ling-Ling Chen, SIBCB, CAS

Hilde Cheroutre, La Jolla Inst.

Wendy Cho, UIUC

Ib Chorkendorff, Denmark TU

Chunaram Choudhary,

Kabelene Hervis U.

Karlene Cimprich, Stanford U.

Laura Colgin, UT Austin

James J. Collins, MIT

Robert Cook-Deegan,

Carolyn Coyne, Duke U.

Roberta Croce, VU Amsterdam

Ismaila Dabo, Penn State U.

Nicolas Dauphas, U. of Chicago

Liset de la Prida,

Instituto Cajal CSIC

Orsola De Marco, Macquarie U.

Claude Desplan, NYU

Sandra Díaz, U. Nacional de Córdoba

Samuel Díaz-Muñoz, UC Davis

Ulrike Diebold, TU Wien

Stefanie Dimmeler,

Cootho U. Frankfurt

Hong Ding, Inst. of Physics, CAS

Dennis Discher, UPenn

Jennifer A. Doudna, UC Berkeley

Raissa M. D'Souza, UC Davis

Bruce Dunn, UCLA

Scott Edwards, Harvard 6.

Todd A. Ehlers, U. of Glasgow

Tobias Erb,

MPS, MPI Terrestrial Microbiology

Beate Escher, UFZ & U. of Tübingen

Vanessa Ezenwa, U. of Georgia

Fleur Ferguson, U. San Dieg

Toren Finkel, U. of Pitt. Med. Ctr. Natascha Förster Schreiber, MPI Extraterrestrial Phys.

Simon Greenhill, U. of Auckland

Gillian Griffiths, U. of Cambridge

Nicolas Gruber, ETH Zürich

Hua Guo, U. of New Mexico

Osvaldo Gutierrez, UCLA

Taekjip Ha, Johns Hopkins U.

Daniel Haber, Mass. General Hos..

Hamida Hammad, VIB IRC

Brian Hare, Duke U.

Kelley Harris, U. of Wash

Sheng-Yang He, Duke U.

Carl-Philipp Heisenberg,

IST Austria

U. of Basel & U. of Cambridge

Heather Hickman, NIAID, NIH

Brian Hie, Stanford U.

Janneke Hille Ris Lambers,

ETH Zürich

Pinshane Huang, UIUC

Christina Hulbe, U. of Otago,

New Zealand

Gwyneth Ingram, ENS Lyon

Darrell Irvine, Scripps Res.

Bharat Jalan, U. of Minnesota

Erich Jarvis, Rockefeller U.

Sheena Josselyn, U. of Toronto

Matt Kaeberlein, U. of Wash.

Daniel Kammen, UC Berkeley

Kisuk Kang, Seoul Nat. U.

V. Narry Kim, Seoul Nat. U.

Logan Kistler, Smithsonian Inst.

Nancy Knowlton, Smithsonian

Etienne Koechlin,

École Normale Supérieure

Peter Koo, Cold Spring Harbor Lab

Paul Kubes, U. of Calgary

Yulong Li, Peking U.

Ryan Lively, Georgia Tech

Luis Liz-Marzán, CIC biomaGUNE

Omar Lizardo, UCLA

Jeffery Molkentin, Cincinnati Children's Hospital Medical Center Alison Motsinger-Reif, NIEHS, NIH (S)

Rosa Moysés, U. of São Paulo

School of Medicine

Carey Nadell, Dartmouth College

Thi Hoang Duong Nguyen,

MRC LMB

Pilar Ossorio, U. of Wisconsin

Andrew Oswald, U. of Warwick

James Owen, Imperial College London

Pierre Paoletti, Ecole Normale

Supérieure, Paris

Philip Phillips, UIUC

Kathrin Plath, UCLA

Martin Plenio, Ulm U.

Katherine Pollard, UCSF

Julia Pongratz, Ludwig Maximilians U.

Philippe Poulin, CNRS

Suzie Pun, U. of Wash

Lei Stanley Qi, Stanford U.

Simona Radutoiu, Aarhus U.

Joeri Rogelj, Imperial College London

Sylvie Roke, École Polytechnique

Fédérale de Lausanne

Sarah Ross, U. of Pittsburgh

John Rubinstein, SickKids

Yvette Running Horse Collin,

Toulouse U.

Mike Ryan, UT Austin

Alberto Salleo, Stanford U.

Lawrence Berkeley Nat. Lab

Nitin Samarth, Penn State U.

Bhaven Sampat, Johns Hopkins U.

Erica Ollmann Saphire,

La Jolla Inst.

Vladimir Shalaev, Purdue U. Jie Shan, Cornell U.

Jay Shendure, U. of Wash.

Steve Sherwood,

U. of New South Wales

Robert Siliciano, JHU School of Med.

Emma Slack,

ETH Zürich & U. of Oxford

Richard Smith, UNC (S)

Carnegie Institution for Sci.

V. S. Subrahmanian,

Northwestern U.

Sandip Sukhtankar, U. of Virginia

Christopher Summerfield,

U. of Oxford

Tavneet Suri, MIT

Naomi Tague, UC Santa Barbara

A. Alec Talin, Sandia Natl. Labs

Patrick Tan, Duke-NUS Med. School

Fuchou Tang, BIOPIC Beijing

Dörthe Tetzlaff, Leibniz Institute of Freshwater Ecology and Inland Fisheries Amanda Thomas, UC Davis

Bozhi Tian, U. of Chicago

Rocio Titiunik, Princeton U.

Shubha Tole,

Tata Inst. of Fundamental Res.

Kimani Toussaint, Brown U.

Li-Huei Tsai, MIT

Jason Tylianakis, U. of Canterbury

Jo Van Ginderachter,

VIB, U. of Ghent

Matthew Vander Heiden, MIT

Ivo Vankelecom, KU Leuven

Henrique Veiga-Fernandes,

Champalimaud Fdn.

Elizabeth Villa, UC San Diego

Bert Vogelstein, Johns Hopkins U.

David Wallach, Weizmann Inst.

Jane-Ling Wang, UC Davis (\$) Jessica Ware,

Amer. Mus. of Natural Hist.

David Waxman, Fudan U.

Alex Webb, U. of Cambridge

Marjorie Weber, U. of Michigan

Chris Wikle, U. of Missouri (\$)

Ian A. Wilson, Scripps Res. (\$)

Sylvia Wirth, ISC Marc Jeannerod

Catherine Wolfram, MIT

Amir Yacoby, Harvard U.

Benjamin Youngblood, St. Jude

Kenneth Zaret, UPenn School of Med.

Magdalena Zernicka-Goetz, Caltech

Bing Zhai, Shenzhen Insts. of Advanced Technology

Lidong Zhao, Beihang U.

Bing Zhu, Inst. of Biophysics, CAS

Xiaowei Zhuang, Harvard U.

Maria Zuber, MIT

The AI debate is missing half the system

Pietro Michelucci

Last month, leading executives of artificial intelligence (AI) companies warned the United Nations Security Council about the risks of losing human control over this rapidly advancing technology. Their recommendations focused heavily on safeguards, evaluation, and governance of AI models and systems, largely overlooking the risks that arise from the interactions of AI with people and institutions. AI safety needs to be addressed at the level of the human-AI system, not the AI model alone.

The stakes are high. People are already embedded in systems that derive capabilities and behaviors from human interactions with AI. AI is influencing which medical treatments people pursue, where companies invest, how cities optimize civil infrastructure, and even how

optimize civil infrastructure, and even how countries approach national security. Humans are no longer thinking alone but in a continuous feedback loop with AI—they are not external to the system. As society becomes increasingly dependent upon AI systems, a kill switch cannot by itself serve as a viable or sustainable safety feature.

The real issue is the safety of the entire human-AI system. No single component, human or machine, can be made perfectly trustworthy. So the system itself must re-

main safe even when its components become unreliable. But how can self-correction and resilience be built into this complex system?

Although this might seem insurmountable, methods for building reliable systems from unreliable components have been under development in the field of human computation, which studies computing systems that rely on humans in a feedback loop. For example, one approach has been to implement a “trust score” modeled on the basic social principle that trust builds slowly but can be broken quickly. To accomplish this, a system randomly mixes test questions, for which the system already knows the correct answer, into work where the answers are unknown. As an individual gets a test question wrong or tries to cheat, their trust score drops. If the score dips below a certain threshold, the person can continue working, but their answers no longer affect the system output. The individual can slowly gain back trust by delivering correct answers again. Through dynamic monitoring, a participant’s contribution can remain valuable while their influence changes according to live fluctuations in their performance. The point is that trust, influence, and authority need not be fixed. They can be continuously adjusted as new evidence emerges.

Consider the recent Hugging Face incident, in which AI agents that were running cybersecurity evaluations bypassed restrictions of the testing environment, sought information about how they were being evaluated, and coordinated a cyberattack against an external open-source platform. A dynamic evaluation approach could monitor whether an AI system begins taking harmful actions while pursuing an assigned task, and rapidly reduce the system's access to tools or data or the ability to act autonomously. These capabilities could be restored only after sustained evidence of reliable behavior.

The same principle could be applied across the broader human-AI system. Reliable evidence for how well a human-AI system is performing a task could be used to adjust

various parameters, such as access to data, as conditions change. This shifts the discourse from whether an AI is safe enough to deploy to how trust, influence, and autonomy should be distributed across the human-AI system.

Many AI and social scientists are working to develop reliable and resilient human-AI systems. For example, the Assess Risks and Impacts of AI program at the US National Institute of Standards and Technology broadens AI evaluation

beyond the model to include interactions with human users in realistic settings. However, generating evidence for system performance can itself be an obstacle to evaluating whether it is safe. Performance cannot always be tested against known outcomes or independently audited decisions. Monitoring performance can itself create risks involving privacy, consent, and authority among the human participants. So any sustainable approach must address not only how trust and influence are adjusted, but what evidence justifies those changes, who can observe them, and who has the authority to act.

Instead of treating an AI model as a device that passes an initial test and is never checked again, regulators need to consider how AI performs in the real world alongside human users and how a system's permissions are continually monitored and adjusted on the basis of demonstrated reliability over time. This means developing safety protocols that shift from static guardrails to dynamic evaluation, and from AI models alone to the full human-AI system. □

Pietro Michelucci is executive director of the Human Computation Institute, Ithaca, NY, USA. pem@hcinst.org

INFECTIOUS DISEASES

Vaccine trial is a ray of hope in Congo's Ebola epidemic

Study will first test “mismatched” vaccine but Bundibugyo-specific shots may be added soon KAI KUPFERSCHMIDT

First responders in North Kivu province in the Democratic Republic of the Congo line up to receive an Ebola vaccine named Ervebo on a “compassionate” basis.

There are glimmers of hope amid the relentless bad news about the explosive epidemic of the Ebola Bundibugyo virus in the Democratic Republic of the Congo (DRC). After months of debates and preparations, a vaccine trial, named Tumaini—Swahili for “hope”—is expected to clear the final regulatory and ethical hurdles any time, and the first doses could be administered this week. The trial will first test Ervebo, a “mismatched” vaccine developed for another Ebolavirus species named Zaire. But two other vaccines, designed specifically to protect against Bundibugyo, may be added to the study in November.

Meanwhile, after more than 9500 reported cases and more than 4600 deaths, the epidemic appears to be slowing in Ituri, its epicenter and the hardest hit province. Both the number of newly reported cases and the percentage of tests that come back positive are going down, says Jean-Jacques Muyembe-Tamfun, who heads the DRC's National Institute of Biomedical Research: "It is not optimal, but it has improved."

And finally, a trial testing two potential Ebola therapeutics—an antibody cocktail named MBP134 and the antiviral remdesivir—has already enrolled more than 600 patients. There may soon be enough data to

see an effect, says Placide Mbala of the Africa Centres for Disease Control and Prevention, principal investigator on the trial. An effective treatment would be a big boost to the response, says epidemiologist Chikwe Ihekweazu, who heads the Health Emergencies Programme at the World Health Organization (WHO). “Every extra tool you have in the arsenal helps.”

Still, the challenges are enormous. The disease could still spread to more places. Earlier this month, Kenya announced its first Bundibugyo case, a 40-year-old Kenyan man who died in Nairobi after traveling from the DRC by way of Uganda. Four

family members and four health care workers have also tested positive and are in isolation. On 20 October, Nigeria reported a case as well, a woman who had traveled to Uganda, which had 20 cases earlier this year but declared its outbreak over in July.

In the DRC, two-thirds of the people who die of Ebola still do so in their communities instead of dedicated treatment centers, creating the risk of further spread. Contact tracing is still not working well; about 40% of newly identified cases aren't on the contact list of any known Ebola case. That's down from about 70% early in the outbreak, says Marie Roseline Belizaire, incident manager for the outbreak at WHO's Regional Office for Africa. "It's progress, but it's not enough."

And although cases in Ituri have gone down, they have trended upward in North Kivu, the second most affected province. WHO and other groups have scrambled to move more resources there. “I think it will be very important to see where the trajectory goes,” Ihekweazu says. With dozens of new cases reported each day, there will be no quick end to the outbreak, he warns. “The one game changer would be if we have a highly efficacious vaccine that we can deploy at scale.”

That's why the stakes in the vaccine trial are so high. Ervebo, developed more than 25 years ago, incorporates a surface protein of Ebola Zaire into a livestock virus that's innocuous to humans. It was first used during the devastating West African Ebola Zaire epidemic of 2013–16, where it showed complete protection in a trial in 2015. Animal studies suggest it may also work against Bundibugyo. The Tumaini trial will test that promise by giving the candidate vaccine to direct contacts of known Ebola cases, a “ring vaccination” design similar to the one used in the 2015 trial.

In that study, conducted in Guinea, entire rings of contacts were randomized to receive the Ebola vaccine either immediately or 3 weeks later. In the Tumaini trial, each person within a ring will be randomized. For now, half will receive the Ervebo vaccine on Day 1 and a typhoid vaccine 4 weeks later; for the other half, the sequence is reversed.

Randomizing individual people rather than entire rings should lead to efficacy data faster, says Ira

Longini, a biostatistician at the University of Florida who was involved in designing both trials. But the typhoid vaccine given to the control group could make it a little harder to detect any benefit from the Ebola vaccine because the typhoid shot might give the immune system a boost. That could protect some people in the control group from infection with Ebola and make any benefit of the vaccine seem a little smaller than it is.

The team started to explain the setup of the trial to people in the affected area 2 weeks ago, says Jean-Pierre Van geertruyden, an epidemiologist at the University of Antwerp and one of the trial's principal investigators. Complicating their outreach is the fact that Ervebo is already given on a “compassionate” basis to first responders in some parts of the country. It’s also used in health care workers in Ituri, in a separate trial called BRAVO that has no control group. Van geertruyden acknowledges that people might question why Ervebo is good enough to give to health care workers but needs testing on the rest of the population: “People might be a bit confused.”

Once Tumaini is up and running, the vaccine candidates specifically targeting the Ebola Bundibugyo virus will be added as they become available. Furthest along is a vaccine developed by researchers at the University of Oxford that contains a surface protein of Bundibugyo stitched into a chimpanzee adenovirus.

Data analyzed so far from ongoing phase 1 trials in the United Kingdom and Uganda suggest the vaccine can proceed to the next stage, says Oxford vaccinologist Teresa Lambe, who hopes it will “enter the trial as soon as possible.” A Moderna-made messenger RNA vaccine is closely behind. Van geertruyden believes both could be added to Tumaini in early November. It may take until February or March 2027 for efficacy data for the new vaccines to emerge, he says.

By then, the epicenter of the epidemic may have moved, and the trial would have to move with it. Organizers would have to shift their team, ensure they have housing and food, and keep the laboratory and refrigerators running in areas with poor infrastructure: “We have to follow the epidemic and that’s not easy,” Van geertruyden says. Perhaps by that point, the outbreak may be not just moving, but waning. □

---

IN OTHER NEWS

NO MORE LEAP SECOND The International Bureau of Weights and Measures announced last week it was bidding adieu to the leap second—an awkward cousin of the leap year, the extra day added to the calendar every 4 years to account for the fact that Earth orbits the Sun once every 365.24 days. The leap second compensated for unpredictable variations in Earth’s rotation and was added to Coordinated Universal Time (UTC)—which regulates clocks worldwide relative to the fixed stars—only when atomic clocks and observations indicated UTC was off by 0.9 seconds. Since 1972, a leap second had been added 27 times, but its idiosyncratic nature caused problems with computer networks and websites. The bureau will now add a leap hour whenever UTC wanders by that much, which means the next adjustment won’t be for thousands of years. —Adrian Cho

DOJ TO FUND GENDER RESEARCH The U.S. Department of Justice is moving to fund research on gender dysphoria, the feeling of distress when a person's gender identity doesn't align with their sex assigned at birth. The 9 October funding notice says it will grant up to \$1 million for studies of "gender dysphoria, its treatment, and the perpetration of violence," including research on the prevalence of violent crime among people with gender dysphoria. The initiative has raised concern among scientists and activists who worry it is driven by political bias that could cause harm to transgender people, who studies show are at increased risk of being victims rather than perpetrators of violent crime. —Julia Vaz

WEST NILE SURGES IN EUROPE This summer, West Nile virus (WNV) spread farther into Europe and sickened more people than ever before. In a report published last week, the European Centre for Disease Prevention and Control reported 1715 locally acquired human infections including 130 deaths since May. The hardest hit countries were Italy and Greece. But with Belgium reporting its first ever cases of the mosquito-borne illness, the agency said “the virus is now circulating further north, a trend that is expected to continue in the coming years.” Most WNV infections cause no symptoms, but about one in five people get sick and about one in 100 develop severe neurological disease, which can be fatal.

—Kai Kupferschmidt

ARTIFICIAL INTELLIGENCE

Across fields, AI finds solutions hiding in plain sight

"BootLoops" software enabled researchers to quickly find and attack problems in 22 disciplines CELINA ZHAO

A BootLoops-powered analysis of the Voynich manuscript, a 15th century book full of undecipherable script, concluded it does not hold a hidden message.

Borrowing an idea from one research discipline to solve a thorny problem in another is an old scientific trick. Charles Darwin's theory of natural selection, for example, inspired algorithms in computer science that "evolve" solutions to complex problems.

But such clever leaps are fleetingly rare. Usually, “A physics researcher wouldn’t spend 6 months reading linguistics research in order to figure out how to repurpose linguistics methods,” says Georgia Channing, artificial intelligence for science lead at Hugging Face, an AI tools platform.

For AI, however, hopping between fields can be a breeze. Researchers highlighted that ability earlier this month with the release of 36 preprints produced with a suite of AI tools called BootLoops. The so-called harness can be used with any large language model (LLM), such as Anthropic's Claude or OpenAI's ChatGPT. And in just 3 months, researchers used Claude outfitted with BootLoops to cross-pollinate at a massive scale, spotting—and solving—problems in seemingly unrelated fields that share the same mathematical structures. The resulting papers range widely, from figuring out the age of sunspots to attempting to decipher a medieval manuscript.

In essence, BootLoops “kept noticing the same pattern: many fields have problems that a technique from mathematics, physics, or computer science would solve outright if anyone knew it existed," Matthew Schwartz, a physicist at Harvard University and visiting researcher at Anthropic who developed the tools, wrote in a 1 October blog post unveiling the preprints.

The project also helps bring AI to scientific disciplines that have been slower to adopt it, Schwartz and other researchers say. “It’s phenomenal that the technology is being saturated through more fields,” Channing says.

Researchers caution that BootLoops still needs extensive human involvement to produce useful results. They also add that it suffers from limitations common in AI-powered science.

Schwartz began to explore AI's research capabilities in December 2025 when he took on Claude Opus 4.5 as his virtual “graduate student.” With heavy hand holding, it managed to spin up a reasonable physics research paper in just 2 weeks. Intrigued, Schwartz started to give the LLM bigger challenges, eventually as ambitious as reinventing quantum mechanics. But “It just wasn’t ready,” he says.

Then, he thought: Instead of forcing AI to perform like a human researcher, why not play to its strengths?

One obvious strength was coding. Schwartz tasked Claude Fable 5 with writing code to solve a type of integral used in theoretical analyses of high energy particle collisions. It combed physics and math papers for other examples of the integrals, then built software to solve 30 of them, including 15 that had never before been solved. The tool became BootLoops.

He then unleashed it on the broader scientific literature, where it soon identified many similar unsolved problems in other disciplines. From there, Schwartz says, “The project exploded,” and he began to recruit experts from other fields to collaborate.

One of the resulting papers, co-authored by evolutionary biologist Scott Edwards of Harvard, involved building evolutionary trees from DNA data. BootLoops calculated that two possible trees of the evolutionary relationships of humans, chimpanzees, gorillas, and orangutans were equally likely—something past estimation strategies could not establish. “This is my first experience with Claude-based science, and I can definitely see the potential,” Edwards says.

BootLoops was also able to find an optimal way to rank health insurance plans offered under the U.S. government's Medicare program. The current classification system, based on metrics such as patient experience, uses an approximate clustering algorithm to set the cutoff between five possible ratings. Those cutoffs then determine the bonus the government pays insurance companies. BootLoops, however, was able to calculate precise optimal bins. The accompanying preprint suggests implementing those suggestions could save the taxpayers roughly \$1 billion from 2025 to '27.

Soon, collaborations expanded beyond math problems to other work suited to LLMs. In yet another project, Schwartz and a trio of linguists used AI agents to scour the academic literature and amass the largest ever standardized catalog of languages' stress and tone patterns, using BootLoops to check the agents hadn't skipped any additional stress systems a language could presumably use. The resulting database, called ACCSTACK, is more than eight times larger than the largest previous data set of its kind, documenting the sound systems of more than 6000 languages. ACCSTACK is "of a scale that's really far, far beyond what we've been able to do as just humans doing it," says preprint co-author Joe Pater of the University of Massachusetts Amherst.

But as BootLoops wandered farther afield, Schwartz found that although the LLM could tell him what it could calculate, it couldn't identify what scientists would really care about. Put simply, it lacked research taste. For example, Claude saw a chance to test ecology's neutral biodiversity theory—which holds that many genetic differences between individuals and populations are due to random variations. The AI generated a counterexample, based on data from a research island in Panama, that showed tree diversity was changing faster than expected by chance alone.

When Schwartz asked ecologist James O'Dwyer of the University of Illinois Urbana-Champaign to examine the results, O'Dwyer found them “interesting” but not field changing. “I don’t think it’s going to replace ecologists of any kind just yet,” he says. “But I do know that everything moved much faster with this collaboration.”

Not all researchers embraced the tools. One preprint analyzes the Voynich manuscript, a 15th century book full of undecipherable script that has perplexed scholars for centuries. The paper claimed the manuscript may not encode a hidden linguistic message after all. But in an online forum dedicated to the manuscript, some contributors complained that the paper was so dense as to be AI “slop.”

Schwartz acknowledges the need for caution: On colleagues' advice, he withheld some results that were insufficiently validated. But he believes researchers in many areas will benefit from working with discipline-spanning AI tools. "Once everyone catches up," he says, "this human-AI collaborative era will be very exciting."

MICROBIOLOGY

How the domestication of a killer bacterium led to Botox

Scientists inadvertently turned C. botulinum into a tool for medicine and the beauty industry, study finds JON COHEN

A lavish 1919 dinner party that ended up killing five people. An attempt to save the canning industry in the 1920s. A World War II biological weapons program and a CIA plot against Fidel Castro. A cosmetic product, best known as Botox, that erases facial wrinkles and exploded into a multibillion-dollar business.

A new analysis of more than 1000 bacterial genomes connects these disparate episodes in U.S. history in a piece of microbiological detective work. The study, published on 15 October in the Proceedings of the National Academy of Sciences (PNAS), shows how scientists unwittingly triggered the evolution of a dangerous bacterium into one with major medical and commercial advantages.

That microbe is Clostridium botulinum, whose neurotoxin disrupts nerve communication, paralyzing muscles and causing victims to suffocate. The toxin, called BoNT in shorthand, is one of the most lethal natural substances known. Under oxygen-free conditions—in preserved food, for example—C. botulinum spores can “wake up,” developing into toxin-producing bacteria.

That's what happened with spores trapped in jarred Californian olives during the “Great Olive Poisoning of 1919,” which included a widely publicized dinner hosted by Detroit industrialist Murray Sales and his wife that killed two of the couple's children, a guest, and two maids. Nine other people in two other states died.

The tragedies posed a major threat to the burgeoning canned food industry and triggered a new wave of microbiological research. Among the researchers who joined the fight was Ivan Clifford Hall, a bacteriologist at the University of California who amassed a collection of C. botulinum isolates.

One, named Hall A, became a laboratory workhorse, propagated over and over. In the process, it became “domesticated”—reshaped by human-driven evolution, like countless other plants, animals, and microbes. Hall A

Ivan Clifford Hall collected many Clostridium botulinum strains, including a lab workhorse now known as Hall A.

lost its ability to form spores, which made it easier for scientists to purify BoNT, and churned out exceptionally high levels of the toxin. Those characteristics made Hall's strain a favorite of U.S. military researchers, who in 1942 started to use it to create biological weapons and a botulinum vaccine.

The military shared Hall A with others. In the early 1960s, CIA launched a plot—ultimately scuttled—to assassinate Cuban leader Fidel Castro by contaminating a box of his cigars with BoNT. Around the same time, a military researcher took a Hall A subculture with him when he moved to the University of Wisconsin and shared it with a Californian ophthalmologist, who began to conduct monkey experiments to see whether BoNT could relax eye muscles, in hopes of finding

treatments for misaligned eyes, a condition called strabismus. “This really is a swords into plowshares type of story,” says microbiologist Paul Keim of Northern Arizona University (NAU), the new study’s first author.

In the 1980s, U.K. microbiologist Andy Pickett and others, working with drug companies, found that tiny doses of BoNT—mere billionths of a gram—produced by descendants of Hall A could treat strabismus in humans. In 1989, the U.S. Food and Drug Administration approved the purified toxin as a treatment for that disease and eyelid spasms. The drug, now made by a few dozen companies, became widely known by the trade name Botox after it was found to temporarily paralyze facial muscles that create lines around the eyes, forehead, and nose.

Pickett has been hired as an expert witness in several legal battles around BoNT ownership and manufacturing secrets, as has Keim. To help unravel the differences between the strains in dispute, Keim suggested lawyers also recruit Richard Lenski, an evolutionary biologist at Michigan State University who in 1988 launched a now-famous experiment to follow 12 identical populations of Escherichia coli and document how they evolved over decades. (The study is still running, nearly 85,000 generations later.)

As part of their work on the case, the researchers constructed family trees of C. botulinum based on publicly available genome sequences and ones they sequenced themselves. They found that the Army Hall A strain and its commercial descendants were on an “unexpectedly long branch” of the family tree, says Lenski, a co-author on the PNAS paper, suggesting strains in that lineage mutated far more rapidly than their cousins.

“That brought to mind what we have seen in our experiments” with E. coli, Lenski explains. In those studies, some lab-propagated strains that started to mutate faster than normal had a defect in a gene involved with DNA repair. At Lenski’s suggestion, Keim and NAU microbial genomicist Jason Sahl looked for similar defects in the Hall A relatives. Sure enough, they found one, a mutation in a gene named mutS.

Keim says the mutation likely arose while Hall was working with the strain. In the wild, it might soon have disappeared, he says, but it thrived in the lab, in part because scientists favored C. botulinum cultures that displayed the mutation's effects. The mutS variant prevents C. botulinum from wasting energy making spores, allowing them to grow well and produce unusually large amounts of the toxin. The bacteria's metabolic pathways are "fine-tuned now for living in the laboratory as opposed to living out in nature," Keim says.

“This work really fills in a fantastic hole of knowledge,” says Pickett, who’s now retired and was not directly involved in the study. He is delighted that the study also spotlights Hall, whom he sees as an unsung hero. Although some of Hall’s “incredible” collection of anaerobic organisms made it to other labs, much of it no longer exists. “No one wanted it,” Pickett says. Unraveling the genetics of the Hall A strain, he says, is “a good tribute to someone who is basically an enigma.”

ASTRONOMY

Microwave observatory peers into the infant universe

In Chile, Simons Observatory begins hunt for galaxy clusters and signs of cosmic inflation DANIEL CLERY

From a perch high in Chile's Atacama Desert, a cluster of telescopes is taking aim at the oldest light in the

universe, the cosmic microwave background (CMB), emitted just 380,000 years after the Big Bang. With greater sensitivity and sharper resolution than earlier CMB surveys, the Simons Observatory will deepen cosmologists' understanding of the universe's makeup and look for signs of "cosmic inflation," a hypothesized burst of exponential expansion in the universe's first moments. "To a first approximation, it's going to do what those [earlier surveys] did, but better," says theoretical physicist Marc Kamionkowski of Johns Hopkins University.

But the observatory is much more than a simple CMB experiment, says co-director Mark Devlin of the University of Pennsylvania. By using the CMB as a backlight, the observatory will also study how the universe has evolved since the light was created. It will discover tens of thousands of galaxy clusters and observe the distribution of matter across cosmic history. A community of thousands of scientists is waiting for the data, Devlin says. “You have to have a broad appeal.”

Moisture in Earth's atmosphere blocks microwaves, so observers have mapped the CMB with satellites and balloons, or from the ground at high, dry spots. Prime locations are the South Pole and

The Simons Observatory's Large Aperture Telescope will map the oldest light in the universe, the cosmic microwave background.

the Atacama, where the \$120 million project, funded mostly by the Simons Foundation, has built three small-aperture telescopes, each 0.5 meters across, and a 6-meter Large Aperture Telescope that is set to begin its survey of the sky this month.

Simons is starting up as CMB researchers adjust to the loss of a much larger planned experiment. Last year, the U.S. National Science Foundation and Department of Energy ended support for a “CMB-S4” experiment, a next-generation project in Chile and at the South Pole that was designed to seek signals of cosmic inflation. “We’re not as sensitive as the original CMB-S4 was planned to be, but we’re covering all of the same science … and we’re doing it now,” says Simons Observatory spokesperson Jo Dunkley of Princeton University.

The CMB, the glow of the superhot infant universe, was stretched by cosmic expansion and now looks like a faint microwave haze from all directions. It is mottled with tiny temperature fluctuations that correspond to concentrations of matter, including unseen dark matter, in the infant universe—seeds that would grow

under gravity into the cosmic web of galaxies, filaments, and voids. By precisely measuring those fluctuations, cosmologists have reconstructed the basic recipe of the universe: roughly 5% ordinary matter, 26% dark matter, and 69% dark energy, a mysterious force that is accelerating the expansion of the universe.

Simons will not only refine those numbers, but also use its map of CMB fluctuations to address other questions, including the mass of the neutrino, a ghostly particle that streams through space at high speeds. Neutrinos don't clump like slow-moving matter and dark matter, so they suppress the growth of the cosmic web. By comparing the early-universe pattern encoded in the CMB with the later distribution of matter, Simons researchers can constrain the mass of these elusive particles. “We should be able to actually have a positive measurement of this mass,” says cosmologist Seshadri Nadathur of the University of Portsmouth.

The observatory will also probe an event thought to have preceded even the CMB: a brief spurt of exponential growth during a fraction of the universe's first second. In that instant of “cosmic inflation” the universe would have ballooned from less than one-trillionth the size of a proton to about as big as a beach ball. Inflation would have ironed most unevenness out of the newborn universe, explaining why today it is remarkably smooth and homogeneous on large scales yet contains a web of galaxies on smaller ones.

The growth spurt would have caused spacetime to reverberate with gravitational waves that might have left an imprint on the CMB: swirls in the polarization of the light. Researchers operating a CMB telescope at the South Pole called Background Imaging of Cosmic Extragalactic Polarization 2 declared they had found such swirls in 2014, only to retract the claim a year later after discovering that much of the signal came from dust in our Galaxy.

Simons hopes to do better. Its small-aperture telescopes are optimized to see fluctuations on scales wider than $1^{\circ}$ , where the primordial polarization signal is expected to be strongest. They also observe at several wavelengths, helping distinguish the swirls from polarized emission from

galactic dust and other signals from the nearer universe. The first two of the scopes began observations almost 2 years ago, with the third joining in more recently. “From here on out, we should be digging down,” Dunkley says. “After 5 years or so, we might be able to see [the swirls].” Some models suggest it might take longer.

While the small telescopes chase inflation's polarization signal, the 6-meter mirror of the Large Aperture Telescope will create fine-scale maps of a large chunk of the CMB. It will funnel light onto a receiver packed with 60,000 detectors, which are cooled to one-tenth of a degree above absolute zero to register the faint CMB signal.

The maps could hold clues to what triggered inflation, Nadathur says. The standard version of inflation is driven by a single field mediated by a particle called an inflaton. But “maybe there was more than one type of field that was interacting in the early universe,” Nadathur says. The two scenarios would result in different distributions of fluctuation sizes in the CMB.

Simons researchers will also use the CMB as a backlight. Hot electrons in the thin gas between galaxies scatter CMB photons, changing their energy in a process known as the Sunyaev-Zeldovich effect. The distortions act as signposts for galaxy clusters, even when the galaxies are too distant or faint to see. An earlier telescope at the Atacama site found thousands of galaxy clusters this way. With its wide field of view, Simons should detect as many as 40,000, including clusters in the early universe that raise questions about cosmic structure formation, Devlin says. “How did you form such a big thing that early on? And what are the physics of that?”

Researchers also hope to exploit a related phenomenon called the kinematic Sunyaev-Zeldovich effect in which moving bodies of gas shift the wavelengths of CMB photons streaming toward Earth. Such “redshifts” or “blueshifts” could reveal the presence of dark matter without the usual signpost of visible stars. “We always assume that dark matter and [normal matter] are distributed in the same way in the early universe. Now we’ll be able to test that,” Kamionkowski says.

Devlin is looking forward to using the facility for so much more than studying the CMB. “I feel like we’re serving a community now that we weren’t accessing before. And that also is a lot of fun.” □

ECOLOGY Deep-sea life is still reeling from the BP oil spill

A 13-year study shows Gulf of Mexico ecosystems still haven't recovered ERIK STOKSTAD

After a deadly blowout on the Deepwater Horizon rig in 2010 sent oil gushing into the Gulf of Mexico, an emergency response team gambled on an unprecedented tactic. It sent robotic submersibles to the wellhead, 1500 meters below the surface, to inject chemical dispersants to break up the oil into tiny droplets and dissolve it. The hope was that by diluting the oil, the approach would minimize harm to wildlife. The move likely spared many seabirds, turtles, and shallow-water fish from oil that would otherwise have pooled at the surface. But it may have taken an unexpected toll in the deep sea.

A 13-year survey, begun the year after the spill, has discovered a steep, sustained decline in fish, crustaceans, and cephalopods in the deep waters of the gulf. The most likely cause is the oil spill, the team concludes this week in the Proceedings of the National Academy of Sciences, although it doesn't directly blame the dispersants.

“They have witnessed a major change in an ecosystem that we thought was isolated and perhaps protected,” says marine ecologist Bruce Robison of the Monterey Bay Aquarium Research Institute, who was not involved in the study. “I just don’t know whether the system itself is resilient enough to recover," he says. "That's a big question mark."

The study began 10 months after the disaster, which killed 11 crew members onboard the BP oil rig and released 5 million barrels of oil. Tracey Sutton, a biological oceanographer at Nova Southeastern University, organized a survey of the deep-sea ecosystem as part of a federal investigation into damages from the spill. Starting in February 2011, the team collected data from a research vessel equipped with sonar, which allowed it to estimate the abundance of schools of fish and crustaceans. The researchers also put down a trawling net that snagged marine life at various depths at 46 locations around the gulf. Working in day and night shifts, they pulled up and counted about 1200 species to create a snapshot of the ecosystem in the aftermath of the spill.

When the team returned 4 years later, it hoped to see signs of recovery. Instead, researchers found further, drastic declines. The populations of three major groups of animals in the survey—fish, crustaceans, and cephalopods—were far below levels recorded in 2011. The picture was no better in subsequent surveys. The numbers remained stubbornly low through to 2023.

Deep-sea fish, such as this snaketooth fish, have had steep declines.

Fish populations registered the largest declines, dropping by 64% compared with 2011. Because the team has no baseline data from before the spill, the declines may have been even greater than it reported.

The team didn't measure how much oil is still present in the ecosystem. But it suspects it continued to have an effect via the food web long after the spill. Years after the disaster, study co-author Isabel Romero, a chemical oceanographer at the University of South Carolina, found polycyclic aromatic hydrocarbons—a particularly toxic component of oil—in the eggs and tissue of deep fish.

The study offers “a sobering example of the unintended consequences of human activity at sea,” says Cynthia Smith, a conservation veterinarian who leads the National Marine Mammal Foundation and was not part of the effort. Jeffrey Drazen, a marine ecologist at the University of Hawaii at Manoa, is particularly impressed by the scale of the survey, which involved collecting samples with a trawl 2000 times, often at depths of more than 1 kilometer. “This is one of the most comprehensive sampling efforts ever conducted in the mesopelagic realm.”

It also has implications for life closer to the surface. Lanternfish—small fish with large eyes and glowing patches of blue, yellow, and green—experienced some of the steepest declines. The fish spend their days in the ocean depths and migrate upward at night to feed on zooplankton, where they are in turn eaten by tuna, swordfish, and whales. A scarcity of lanternfish, with many of the survivors possibly contaminated with hydrocarbons, could help explain a recent decline in whales in the gulf. Sperm whales are now about 30% less common there than before the spill, for example.

Steven Murawski, a marine ecologist at the University of South Florida, is surprised by the lack of recovery. He thinks an influx of animals from nearby populations, for example in the western or southern gulf, should have bolstered the depleted region. “There must be some cascading effect from the footprint of the oil spill to the larger community in general in the gulf,” he says.

Smith says the new study reinforces the need for a broader look at when and whether to use dispersants in oil spills. “Those decisions should consider all animals within the oil spill footprint, from the surface to the sea floor.”

It is also a cautionary tale about other industrial activities in the deep sea such as seabed mining, the authors say. “There’s this idea that populations in deep-sea environments are abundant and can’t be harmed,” says co-author Heather Bracken-Grissom, an evolutionary biologist at Florida International University. But the study indicates that’s not the case. “The deep sea is a much more fragile system than we thought,” Sutton says. □

GEOPHYSICS

Iron at Earth's heart may take a surprising form

X-ray experiments reveal cubic crystals that could complicate the picture of the inner core PAUL VOOSEN

Four years ago, a dozen scientists packed into a control room at the European X-ray Free-Electron

Laser (XFEL) Facility in Germany. A sliver of iron thinner than a human hair was squeezed to deep-Earth pressures between the tips of diamond anvils, then blasted to extreme temperatures by the x-ray laser. A surprising picture emerged on the consoles, says Sébastien Merkel, a geophysicist at the University of Lille who leads the \$2.5 million project, called HotCores. The crystal structure of the iron was cubic—the first time the arrangement had been seen at the dynamo that generates Earth's magnetic field. "If the structure is different, then the properties will be different," says Shuai Zhang, a high-pressure researcher at the University of Rochester.

The finding was difficult to publish. It finally appeared in August in the journal Physical Review B, after several high-profile journals turned it down, Merkel says. The delays may reflect resistance to reopening a question many researchers thought was settled. Yet to Jon Eggert, a materials physicist at Lawrence Livermore National Laboratory, "The data is very clear. I'm convinced they see BCC."

---

It reopens a topic that I thought was closed.

Sébastien Merkel

University of Lille

such high pressures and temperatures. The result seemed impossible, Merkel says. “It’s really the last thing we expected,” he says. “It reopens a topic that I thought was closed.”

For decades, experiments and computer simulations of Earth's inner core, an alien world of solid iron some 5000 kilometers below the surface, suggested the iron atoms there form close-packed hexagons. But word of the new result has quietly roiled the small community of inner-core researchers. If the structure is instead what is commonly called body-centered cubic (BCC), it could change how heat is conducted, altering when in geological history the inner core began to solidify and how heat and elements were released into the molten outer core. Those processes affect how the outer core churns, powering

Experiments that try to reproduce the extreme pressures and temperatures deep inside Earth come with serious complications. Diamond-anvil cells generate deep-Earth pressures but can introduce impurities that alter iron's structure, whereas conventional laser experiments can boost samples to high temperatures, but only for a single vanishing moment—too little time for further measurements.

By combining the XFEL with a diamond cell, the HotCores team could blast its samples with repeated, 30-femtosecond x-ray pulses. The x-rays act as both “pump and probe,” Merkel says. Each pulse heats the sample, and the x-rays scatter from its atoms to produce a snapshot of their arrangement—a diffraction pattern.

At first, the snapshots showed the expected hexagonal structure. But once pressures and temperatures rose above 2.3 million atmospheres and some $5700^{\circ}\mathrm{C}$ , a clear cubic signal emerged, Merkel says. The finding isn't definitive. Attempts to reproduce it in subsequent samples faltered when the diamond cell could not reach the same pressure. And unpublished

NEWS

Researchers prepare a sliver of iron for heating and squeezing to corelike conditions.

experiments using traditional laser shock techniques have yielded only hexagonal iron.

Yet Maurizio Mattesini, a geophysicist at the Complutense University of Madrid, feels vindicated by the result. “I’ve been suffering for such a long time,” he says. Mattesini argues that only cubic iron can explain some features seen in the seismic waves from earthquakes that pass through the inner core, such as a surprisingly slow speed. BCC, with a more loosely packed, less rigid structure than hexagonal iron, would hamper the waves, like soft sand slowing a runner. But other modelers have found plausible ways for denser hexagonal iron to slow seismic waves, too.

A paper published in 2021 by Yang Sun, a mineral physicist at Xiamen University, and colleagues had proposed that iron solidifying in the inner core would first crystallize into a cubic structure before settling into its final hexagonal form. “The phase that forms first does not have to be the phase that eventually dominates,” Sun says. The model could explain why the HotCores team saw BCC when the laser experiments did not. The sample heats and cools with each laser pulse, giving the transient cubic phase time to emerge. In conventional laser shock experiments, the sample experiences a single violent heating event.

Because the hexagonal and cubic phases can form under such similar conditions, they might both exist in the inner core, says Sally June Tracy, a mineral physicist at the Carnegie Institution for Science. Small changes in the core—for example, a dash of extra nickel and other light elements—could shift the balance to cubic, she says. Indeed, Merkel's team has since moved to testing these alloys with the x-ray laser. Although the work is unpublished, he says his team has found that iron with nickel and silicon tends to stabilize in the cubic form.

There's an irony to HotCores's discovery of cubic iron, Merkel adds. The first few lines of the team's grant proposal stated plainly that iron in the core is hexagonal. But now, what lies at Earth's heart is uncertain once again. "What's happening in the core?" Merkel says. "I don't know."

SCIENTIFIC COMMUNITY

Fired biologist wins \$6 million in defamation lawsuit

A jury ruled in favor of David Sabatini, dismissing accuser's claims of sexual harassment, retaliation PHIE JACOBS

Biologist David Sabatini, whose prominent career unraveled 5 years ago after an investigation found he had violated sexual harassment policies,

scored a major victory last week when a Massachusetts jury awarded him \$6 million—plus interest—in damages in his defamation case against Kristin Knouse, a former colleague at the Whitehead Institute for Biomedical Research. Her claims he sexually harassed her led one research institution to fire Sabatini and precipitated his resignations from Whitehead and the Massachusetts Institute of Technology (MIT), where he was a tenured faculty member.

Sabatini tells Science the hours leading up to the 16 October verdict were “the most nerve-wracking of my life.” To him, the jury’s decision represents “a clear vindication” that has brought him “a massive sense of relief.” The jury also rejected Knouse’s counterclaims of sexual harassment and retaliation.

The outcome came as a surprise to some legal experts, in part because two judges had earlier dismissed claims Sabatini filed against Whitehead and its director, Ruth Lehmann, and even some he filed against Knouse. After investigating Sabatini's conduct, Whitehead in 2021 concluded he had breached the institute's consensual relationships policy and its antiharassment policy. He had met Knouse while he was associated director at Whitehead and she an M.D.-Ph.D. student at Harvard University and MIT, where he was one of her instructors and served on her thesis committee. The two began a sexual relationship just before Knouse joined Whitehead as a fellow and started her own lab.

Although Knouse says she was coerced into the relationship by Sabatini, who at the time served as director of the Whitehead Fellows Program, he maintained it was consensual. He contended Knouse only leveled harassment accusations after he broke up with her to pursue another woman. A Suffolk County jury has now ruled in his favor, finding that Knouse defamed Sabatini and interfered with his ability to work.

“That’s a pretty big finding,” says Ann Olivarius, a lawyer who specializes in cases of sexual harassment. “Obviously, the jury really went against her in a big way." Olivarius was also surprised the jury awarded Sabatini such a hefty sum in damages. "That's really rare. This is a standout case."

The civil lawsuit verdict has no direct impact on the conclusions of Whitehead's probe, although Sabatini and his supporters view it as an exoneration. And some, including billionaire Bill Ackman, who has used his large social media platform to defend Sabatini and bankrolled his recent research, called on the institute to reconsider its findings and demanded Lehmann resign.

Knouse's attorney, Ellen Zucker, tells Science her client intends to appeal the decision. “The verdict is deeply disappointing,” she says, noting that several other women testified in court about “concerning interactions” with Sabatini. Requiring more than \$6 million in damages from an early-career scientist, Zucker argues, “is proof positive that the jury was not well-equipped to make appropriate decisions in this case.”

Sabatini rose to prominence with research into a molecular signaling pathway that regulates cell growth and metabolism and malfunctions in diseases such as cancer. At Whitehead, he won a prestigious position as a Howard Hughes Medical Institute (HHMI) investigator. In 2021, after an employee survey identified concerns in Sabatini's lab, Whitehead hired an outside law firm to investigate.

In addition to finding that Sabatini breached the institute's sexual relationships policy with Knouse, the investigation also concluded that Sabatini cultivated a sexualized lab environment and tried to influence the probe by misleading investigators and attempting to shape interviewees' testimony. As his official employer, HHMI fired Sabatini, who also relinquished his position and lab at Whitehead. Sabatini later resigned from MIT after three senior officials recommended his tenure be revoked.

Sabatini's lawyer, Lisa Arrowood, says the once high-flying biologist “became a complete pariah in the scientific community.” Sabatini has since made efforts to resurrect his career. In 2023, Ackman and an anonymous donor pledged a total of \$25 million to restart his lab. Later that year, Sabatini took a position at the Institute of Organic Chemistry and Biochemistry of the Czech Academy

of Sciences. The institute recently opened a branch in the Boston area, where Sabatini now has lab space.

In 2024, the U.S. Department of Health and Human Services (HHS) announced it was seeking to debar him from receiving any federal research funding because of the accusations. Sabatini confirmed that in September 2025 the agency did ban him from receiving National Institutes of Health grants for a period of 1 year. But that debarment was terminated early, in April, according to an HHS email provided to Science by Sabatini.

Sabatini has argued there was no power imbalance between himself and Knouse. “She had her own lab, she had her own money, she had her own students,” Arrowood says of Knouse. But the idea that Sabatini did not hold a position of power over Knouse, a much younger colleague early in her career, strikes Olivarius as “ridiculous.” “Look who he is, and look who she is. ... That's not an equal relationship."

Yet Regina Barzilay, a computer scientist at MIT who describes herself as a friend and colleague of Sabatini, says the original Whitehead probe was “extremely one-sided” and she is “overjoyed” that Sabatini had won his case.

To Zucker, however, the verdict against her client is emblematic of larger cultural issues: “This trial and this jury’s verdict seems to me to be a sad marker of our times, where it is still hard for women to be believed and even harder to hold powerful men to account.”

Sabatini, his lawyer says, plans to appeal the earlier summary dismissals of his claims against Whitehead and Lehmann, which included defamation and, for accepting Knouse's version of events over his, gender discrimination. According to Olivarius, both he and Knouse likely face an uphill battle. "Appeals are hard to win, always." □

---

A lasting aftershock

The tsunami that hit Indonesia in 2004 following an undersea earthquake left 170,000 dead and forced more than half a million people from their homes. More than 20 years later, a new study has found, people who witnessed the devastation may have lasting health effects including metabolic imbalances. In a study posted last month on the preprint server bioRxiv, researchers tracked levels of thyroid hormone and other health markers in 3270 people in households from which they had baseline data dating before the disaster. The team found lower thyroid hormone levels among people from the areas worst hit by the tsunami, as well as more obesity, higher cholesterol, and other indicators of poor metabolic health. The researchers concluded that stress caused by the devastation—some of which is pictured here at a mosque in Banda Aceh 2 days after the tsunami—may have permanently altered the hormone's ability to regulate physiological functions. —Sara Reardon

EYES ON THE BALL

A “robo ump” that calls balls and strikes in Major League Baseball makes its World Series debut this week. It has to be near-perfect

JON COHEN, at Petco Park in San Diego

t 6 a.m. on 8 October, Kyle Wait and Tyler Danby towed Pelican cases and duffel bags filled with gear onto the home field of baseball's San Diego Padres. The night before, the team had lost a

heart-wrenching game here, silencing 47,712 roaring fans and eliminating the Padres from the playoffs.

Aside from 2 dozen pristine baseballs, known as pearls, the men brought no baseball gear into the still-dark ballpark. But their high-speed cameras, tripods, lasers, rulers, plumb lines, radars, and a custom-built, piston-powered cannon that fires baseballs had everything to do with a question key

Major League Baseball tested its Automated Ball-Strike Challenge

System during spring training in 2025.

to every game: whether a pitch is in the strike zone, crossing home plate at just the right height and depth, or misses the zone—a ball.

Wait and Danby, who both work at the Sports Science Laboratory run by Lloyd Smith at Washington State University (WSU), worried about

the darkness of the field, where most of the light towers were shut off, and the beer-flavored mist from the cleanup crew's power washing. At the invitation of Major League Baseball, they were about to do a “ground truth” of a technology introduced this year that represents one of the biggest changes ever in MLB's 123-year-old history: the Automated Ball-Strike (ABS) Challenge System, more commonly known as “robo ump.” And their cameras, which cost more than \$100,000 each, thrive on light and hate moisture.

After a flurry of emails and phone calls, the stadium lights started to blaze. The cleaners retreated. It was game on.

Over the next 3 hours, the two photographed 100 cannon-shot pearls crossing home plate at about 90 miles per hour, images they will analyze back at WSU to calibrate the park's camera-based ABS system. (Science style is to use metric measurements, but baseball decidedly does not.) “I love doing this,” said Wait, an erstwhile wildlife ecologist who once used cameras to track carnivores and has worked as a ground truth specialist checking MLB’s 30 stadiums for 7 years. “Perfecting our methods is just a constant battle,” Danby says.

The core confrontation in baseball is between a pitcher and a batter, but there's another important player in that drama—the home plate umpire, who decides in a split second whether an unswung pitch is a strike or a ball. Players and managers in the past could not, officially, challenge the ump's decision. But that didn't stop grown adults from swearing, wildly gesticulating, and acting like spoiled children when they disagreed with a call, risking ejection from the game. Given that three strikes end a batter's turn and four balls wins them a walk to first base, the stakes are high.

Even umps knew calls were regularly missed—19 in one 2022 game made especially memorable by a star slugger's temper tantrum over the incorrect calls. So MLB has turned to an ABS system made by Hawk-Eye Innovations that uses five cameras around the baseball field to check the umpire's calls. For an industry that grosses \$12 billion a year, with games, careers, and more billions in bets riding on the fate of pitches, the system has to be near-perfect. “People ask, ‘We know you use Hawk-Eye to grade the umpires, but who grades Hawk-Eye?’” says electrical engineer Clay Nunnally, a baseball scientist at MLB. “WSU grades Hawk-Eye.”

With MLB's ABS system, a pitcher, catcher, or batter can now challenge an umpire's call of a ball or strike by tapping the top of his helmet or cap within 2 seconds. A few seconds later, the pitch's trajectory appears in slow motion on the stadium's jumbotron, showing the precise location of the baseball when it crossed the halfway point of home plate (see graphic, p. 376). Teams start games with just two challenges, which are only retained if calls are overturned.

MLB's regular season this year saw 10,557 ball or strike challenges, and corrected calls changed games significantly. Three weeks ago, an extra-innings playoff contest ended on an ABS appeal, which turned a called ball into strike three. When a call gets overturned, fans sometimes go wild. But the robo ump calms down emotions on the field: Ejections over ball-strike disputes plummeted from 110 last season to 11, because the digital evidence leaves less room for contention.

Tracking a ball traveling not much faster than a car on a freeway for less than 60 feet might seem a simple assignment for today's technology. But the high stakes make fraction-of-an-inch precision crucial. Two ABS approaches, developed for different sports by a radar engineer and a computer scientist, competed for the major league job, with Hawk-Eye edging out its rival. And surprising variables can alter each system's accuracy. "It's not trivial to track the ball and figure out where it's at," Smith says. "There are a lot of things that come into play."

AT MLB'S START in 1903, the strike zone was a 3D concept. A baseball hit the zone if it crossed any part of home plate above the knees of a player and below his shoulders. For the Hawk-Eye ABS system this year, however, the strike zone is a 2D plane at the center of the plate that is adjusted for each player's height. MLB measures each player, without shoes on, at the beginning of the season. The measurement is done both by hand—using stadiometers or sliding wall rulers—and with lasers. It takes place between 10 a.m. and 12 p.m. to avoid the spinal compression that occurs over the course of a day.

ABS systems don't just capture the ball as it crosses home: They track its initial position, velocity, and acceleration throughout its half-second journey to the plate, calculating its trajectory toward the strike zone. “You gain overall precision in the measurement by using the full trajectory, rather than a single point,” says retired particle physicist Alan Nathan, a Smith collaborator who maintains a website called the Physics of Baseball.

Pitch-tracking technology first came to MLB ballparks 20 years ago, but originally just to help coaches and

players improve pitching and batting. It's been used as an ABS system in MLB's minor leagues for the past 5 years. "We test these things pretty significantly before we bring anything to the major league," says Joe Martinez, a former MLB pitcher who's now a vice president in the organization.

An early two-camera ABS system, PITCHf/x, used in the minor leagues between 2006 to '15, claimed an accuracy of half an inch and got TV viewers used to seeing whether a ball was truly in the strike zone. But then Fredrik Tuxen, an electrical engineer from

Denmark who knew nothing about baseball growing up, offered a more precise technology.

Tuxen had helped develop Doppler radars for the military to track missiles. Owners of golf driving ranges approached him and asked whether he could create a Doppler device to gauge the distance of their balls. A devoted golfer, Tuxen already knew it could be done: He had pulled out his clubs at the U.S. Army's Yuma Proving Ground 10 years earlier when he finished testing a missile-detecting radar. "After 2 hours, I managed to actually track a golf ball, but the problem was I had a \$2 million radar," he says.

Among the players who can challenge an umpire's strike or ball call, by tapping their heads, pitchers have the lowest success rate.

This led Tuxen in 2003 to co-found a company called Trackman, which developed a much cheaper radar that could fit into a large briefcase but still follow a golf ball. Sales took off. "People immediately said, 'Hey guys, you need to take this to baseball, they're crazy about data,'" he recalls.

The Trackman system sends out roughly 24,000 pulses of radio waves

The umpire's view of the strike differs from that of the cameras.

Catchers have the highest success rate with challenges.

In the zone

An ABS strike is defined as a pitch that crosses home plate's halfway point within a player-adjusted height range and inside its 17" width.

---

Umping the umps

The camera-driven

Automated

Ball-Strike (ABS)

Challenge System

introduced by Major

League Baseball

this season allows

the pitcher, catcher,

or batter to contest

pitch calls made

by umpires.

during a pitch's brief flight. Receivers capture the reflections from the ball, which reveal its speed, position, and—by tracking the ball's stitched seams—how fast it spins. The aerodynamic forces that spin creates, known as the Magnus effect, can deceive batters by causing a ball to deviate from a straight path. Many players and coaches didn't immediately see the value of the machine. “For years, players have looked with their eyes and talked about stuff—‘This was a nasty pitch’ and all that—and to me as an engineer, that didn’t make sense,” Tuxen says. “I need numbers.”

By 2010, Trackman delivered the numbers, putting science into the art of pitching. It showed, for example, that an MLB pitcher's average curve-ball spins at 2450 revolutions per minute (rpm) and revealed the degree at which it tumbles or twists. Pitchers learned for the first time that a curve spinning faster than 2600 rpm was more likely to fool a batter to swing and miss—a strike.

Trackman initially provided entertainment for MLB broadcasts, distinguishing pitch types—the eye can't easily tell a cutter from a slider—and trajectories, as well as pitch locations. To improve Trackman, Tuxen hired Smith's team to do ground truths.

---

Points of view

The ABS system designed by Hawk-Eye Innovations relies on five high-speed cameras around the ballpark. Because pitches can travel at more than 100 miles per hour, cameras might not capture the exact moment the baseball crosses midplate, so Hawk-Eye includes the trajectory of the ball in the equation that determines whether a pitch was a ball or a strike.

Then MLB decided to assess umpires with an ABS system. “That’s where I got involved,” Smith says. “If you’re going to use that to evaluate people, especially if there’s a union, you have to have some independent system that says, this actually works.”

Tuxen has been delighted to watch the robo ump gain traction in the sport. “It’s interesting when you have an old game like baseball, with all this culture around it, including having an umpire,” Tuxen says. “Now you have modern technology meeting it. How do you bridge that?”

MANY OF THE people who work on ABS systems played baseball themselves, but not Smith. “I would have been voted the least likely to be an athlete,” he says. He came to the game via his expertise in composite materials. Collaborating with Nathan, a lifelong amateur ball player, he began studies of the performance of different kinds of bats—wood, metal, and composite. The work involved firing baseballs at stationary bats, and Smith was hooked. “I kind of got into this by coincidence, but I found the research so interesting, and who wouldn’t enjoy a career building cannons?” he says.

MLB only allows wood bats, but colleges and other leagues permit other materials. That market is large, and the income from bat research allowed Smith to create the WSU sports lab, which has nearly 2 dozen staff. Together they have published more than 50 papers about baseball and softball, in journals ranging from Sports Engineering to Composite Structures. More than 2 dozen graduate students have come through with theses examining cricket bats, ice hockey sticks, shuttlecock drag, taekwondo, and the aerodynamics of baseball's oddest pitch, the knuckleball.

When Smith began to evaluate Trackman, his team soon discovered subtle factors can affect the radar results. The weight of a sellout crowd, for example, can cause a stadium to sag. “The radar orientation might change a 100th of a degree, but that translates into inches at the plate, and inches matter in baseball,” Tuxen says.

To improve accuracy, Trackman added a camera to its system, supplementing the radar with optical data. In 2015, MLB put Trackman in all 30 of its ballparks in conjunction with the rollout of its own data analytics platform, Statcast. Trackman data helped teams evaluate future players, and some professional pitchers bought units to sharpen their “stuff.” Professional baseball teams in South Korea and Japan began to use the radar system for analytics, too.

MLB didn't test Trackman as a robo ump until 2019, however, and the next year, it switched to Hawk-Eye, which relies entirely on cameras to track the ball. MLB won't comment directly on the decision, but Hawk-Eye was equally accurate, possibly cheaper, and provided more information. At the time, “it was much more sophisticated,” Tuxen acknowledges.

A U.K.-based company started in 2001 by computer scientist Paul Hawkins, Hawk-Eye first set out to adjudicate difficult calls in cricket, his first love. Purchased by Sony 10 years later, Hawk-Eye is now used in some 20 sports, including tennis, soccer, football, and rugby. For baseball, Hawk-Eye initially relied on 10 cameras, but added an 11th to detect the ball's spin and then another just because the computers could handle its data. Some of the cameras track player movements, but five focus on the ball, imaging its flight at 300 frames per second. "We're now at a point with optical tracking that we're measuring the actual orientation of the ball's seam," says Matthew Gaydon, a baseball tech manager at Hawk-Eye.

Even that fast frame rate may still miss capturing the ball at the moment it enters or misses the 2D strike zone now used by MLB. “The ball travels about 5 inches in between raw measurements and we’d have to be lucky to get an actual measurement right at the plane of the strike zone at midplate,” Nunnally says. So like Trackman, Hawk-Eye calculates a trajectory and extrapolates it across the strike zone.

By combing data from 3000 test pitches evaluated by WSU with both Trackman and Hawk-Eye, MLB has determined the median ABS “error” for both systems is now 0.17 inches. “Even though we call this ‘error,’ it is more a gauge of uncertainty,” Nunnally says.

MANY FACTORS can degrade Hawk-Eye's performance, and MLB had to pause its ABS system eight times this year during games. Hawk-Eye cameras can go out of alignment because of intense heat, the ground shifting under the ballpark, or having lenses cleaned during a game because of rain, fog, or sea spray. That's why WSU's team is here at Petco Park to evaluate how Hawk-Eye's five ABS cameras—four of them set on beams above the nosebleed

section, the highest tier of seats behind home plate, and one in the outfield—have held up.

During a ground truth test at Petco Park, Tyler Danby fires a fastball from the cannon to home plate.

The testers use two of their own cameras, which at 2500 frames per second operate much faster than the ABS ones. After using a level, laser, and a plumb at home plate to precisely calculate the height, width, and depth of a 3D strike zone, Wait and Danby record 100 pitches fired by the cannon on the mound, each accompanied by a boom and a puff of condensation as the compressed air expands and cools. At Hawk-Eye's control room in Atlanta, technicians simultaneously run Petco's ABS cameras. The technicians can remotely focus them if need be, combating changes in lighting from clouds or the Sun, which they did during each of the 24:30 games played in MLB this season.

The ground truthing goal is to test how accurately the system measures position, velocity, and spin. Danby

orients each pearl in the cannon's claw so the cameras can follow the seams and the writing printed on the ball, including the word “OFFICIAL” flanked by two stars. “To do spin rate, we track surface markers,” Wait explains.

The WSU team will calculate the “offsets” between its coordinates and the ones recorded by Hawk-Eye cameras, but the academic ball location evaluation is far slower than the ABS system’s in live play. “The crazy thing is, the pitch is thrown, and they know the [strike zone location] instantly,” Wait says. “We’ll do this, and I’ll tell you next week what the answer is.”

Hawk-Eye will then adjust its cameras to correct for imprecisions.

Gaydon thinks the automated ump revolution is just beginning. He expects Hawk-Eye technology will make its way down to high school games or even Little League contests, with smartphones in fixed locations providing the cameras. “I can’t wait for the day where the parent who’s invested everything into their 9-year-old starts arguing with the poor umpire, and he’s like, ‘I didn’t make the call. The camera way up there did,’” Gaydon says.

Besides making the umps' job less contentious, the ABS system has given them some heartening news. MLB's comparisons of human pitch calls with automated ones showed that this season, human umps were correct 93.5% of the time. Martinez, the former MLB pitcher, isn't surprised. "Everybody, even more so now than ever, appreciates the jobs the umpires do back there." □

Former FBI counterintelligence agent Tam Dao now works on research security at Rice University. 378

THE INSTIGATOR

A single FBI agent spurred NIH's campaign against foreign intervention on U.S. campuses. Now, he's studying the extent of the problem JEFFREY MERVIS

The National Institutes of Health (NIH) has spent the past decade compiling information on hundreds of U.S.-based scientists it believes tried to illegally

help their country of origin, most often China. But it's never been clear what triggered NIH's aggressive pursuit of these scientists, many of whom eventually lost their jobs or even went to jail, and why it was so much more extensive than efforts at other federal agencies.

Mystery solved. Science has learned that Tam Dao, an academic psychologist-turned-FBI special agent for counterintelligence, laid the groundwork for NIH's campaign. "I led the investigation that started all this," Dao says, acknowledging for the first time his pivotal role as the United States competes with China for global leadership in science.

Dao remains an important player in the debate over research security. Now head of campus safety and research security at Rice University, he will be tackling a question that arguably should have been asked before NIH plunged ahead: How often do foreign governments try to steal research at U.S. universities?

“Tam brought [the problem] to our attention,” agrees Michael Lauer, former head of NIH’s extramural research program, referring to several meetings starting in 2014 in which Dao told senior NIH officials he believed scientists at the MD Anderson Cancer Center had violated the confidentiality of the NIH peer-review process to benefit colleagues at Chinese institutions. “He had copies of their emails, text messages, and computers, which obviously we didn’t have access to,” Lauer recalls.

Dao's investigation of MD Anderson, which resulted in the firing of three scientists of Asian descent in 2019, was one of hundreds of counterintelligence cases Dao handled during his 11 years at FBI's Houston field office, from 2012 to '23. His efforts were part of a broader push by the U.S. government called the China Initiative, begun in 2018 during President Donald Trump's first term to root out espionage. And he convinced NIH there was a serious problem it needed to address.

Spurred on by what Dao had compiled and shared, Lauer assembled his own team to ferret out more information on questionable practices, including a failure to disclose foreign ties. That led to hundreds of U.S. universities getting letters from Lauer asking them to investigate one or more faculty members holding NIH grants and take appropriate action. The effort grew over the next few years, and as of June NIH says it has found 238 cases of undisclosed affiliations, including 197 instances of unreported grant support. As a result, 122 scientists have lost their jobs and 216 have been banned as peer reviewers.

While NIH's probes continued, the Department of Justice (DOJ), FBI's parent agency, ended the China Initiative in 2022 under then-President Joe Biden after acknowledging that many people believed it had unfairly targeted scientists of Chinese descent. And Dao's feelings about its methods and its impact are nuanced.

“I spent my whole [academic] career developing mental health inventories that don’t overdiagnose or are biased against Asian Americans,” he says. “So any accusation of targeting Asian Americans hits me at the core.”

He agrees with critics that the China Initiative damaged the reputations of many innocent scientists and “deterred good people from coming to the U.S.” But he also thinks it “has helped the country more than it has hurt the country by deterring bad people from coming here … and making the country safer.” And he strongly defends the government’s tactics.

“I’m pretty sure there are people out there who are racist and will target [those of Asian descent],” he says. “But the FBI looks at the behavior, not the ethnicity, of an individual, and that behavior has to provide what we call reasonable, articulable facts to support a case.”

DAO'S PERSONAL HISTORY makes him both an unlikely candidate—and a perfect fit—to influence government policy on research security. Born in what was then Saigon on 5 April 1975, he and his family were evacuated just days before the city fell to the North Vietnamese on 30 April. “We were told to go to a certain place at a certain time,” Dao recalls learning from his father, a translator for the U.S. government.

Housed temporarily in a re-location center at Fort Chaffee in Arkansas, Dao's family eventually ended up in Houston, where they struggled to make ends meet. His parents worked minimum-wage jobs, and during the summers he and his three brothers fended for themselves. "Today that would be considered child abuse," Dao says. But his parents "had no choice."

They also put a priority on education. Following in the footsteps of his two older brothers, Dao graduated from the University of Texas (UT) at Austin. After 2 years at a well-paid but unfulfilling job managing real estate, the desire to make a bigger contribution to society led him back to school and a Ph.D. in counseling psychology from Florida State University. In 2009, he joined the faculty of the University of Houston, where he specialized in personality assessments.

His department was willing to sponsor him for tenure after only 3 years, Dao says, but he felt it was time to show his appreciation for “everything this country has done for my family." He applied for an FBI job, even though he lacked training in law enforcement—"I didn't even know how to shoot a gun"—and was beyond the age limit of 37 for new agents. He received an age waiver, survived the grueling boot camp, and was assigned to the Houston FBI office in 2012. After a few years he moved into counterintelligence, investigating whether U.S.-based academic scientists might be helping foreign entities illegally acquire research funded by the U.S. government.

One of his first counterintelligence cases involved MD Anderson, part of the UT system. But when Dao presented his findings to then-NIH Director Francis Collins; his top deputy, Lawrence Tabak; and Lauer, “we were all rather skeptical that there was any problem,” Lauer recalls. Dao says: “They said I was not credible, and that NIH-funded scientists had been trained not to do those things.”

A major reason for that skepticism was what Lauer calls “a culture gap” between the two agencies. “The kind of people who work for the FBI … come from very different backgrounds and have different perspectives about how the world works,”

Any accusation of targeting Asian Americans hits me at the core.

Tam Dao, Rice University

he explains. “And we’re not used to dealing with them.”

Dao wanted to bridge that gap using the currency of science: evidence. He got permission to share with NIH the information he and other FBI agents had amassed—an unusual approach that Dao says required months of delicate negotiations with his bosses.

The research on mental health assessments that Dao had carried out also established some common ground with the NIH officials. “He had been applying various modern statistical techniques to problems that were similar to the statistical techniques I had been using when I was studying cardiovascular disease,” recalls Lauer, who worked at the Cleveland Clinic before coming to NIH in 2007. “So there was a connection. He’s also a really nice person.”

Once NIH leadership had come around to Dao's conclusion, Collins sent a letter to some 10,000 institutions that NIH funds. Dao helped craft that August 2018 bombshell, which warned that “NIH is aware that some foreign entities have mounted systematic programs to influence NIH researchers and peer reviewers and to take advantage of the long tradition of trust, fairness, and excellence of NIH-supported research activities.”

The subsequent flurry of letters from Lauer sent universities scrambling to investigate possible involvement by their faculty in what NIH labels foreign interference. The phrase encompasses violations of rules that govern peer review or the requirement of grantees to disclose ties to foreign institutions. But many university administrators complained they lacked access to what NIH—through Dao—already knew about their cases. That ignorance, Dao says, contributed to the perception that Asian scientists were being unfairly targeted. He decided once again that what was needed was greater transparency.

“Traditionally, the FBI would do its own investigation [without involving the institutions],” Dao explains. “But when you do that, you’re at risk of bias. I didn’t like that approach. So I started to give them the classified stuff. I would say, ‘Here it is. You decide whether you have a problem.’” Dao estimates he held such briefings for more than 300 universities and medical centers during his FBI years.

He believes sharing FBI information with institutions has improved research security on campus. One important ingredient, he says, is bringing in faculty with the expertise to decide whether a breach, if it had occurred, could pose a threat to national security. Dao says his FBI colleagues don't possess such expertise.

“Most agents are trying to protect something that they don’t even understand,” Dao explains. “For example, what aspects of quantum pose a national security risk? Most FBI agents would have no idea. Neither do I, for that matter.”

TO MANY POLITICIANS from both parties, the MD Anderson case was all the proof they needed that academic institutions weren't doing enough to combat the perceived threat. A steady stream of DOJ press releases touting the indictments of individuals or settlements with institutions in cases involving so-called “malign foreign influences” suggested the problem was huge—and growing.

But many of the cases involving academics were later dropped or ended in losses by the government.

Dao's current project, funded last month by the National Science Foundation (NSF), is the first rigorous attempt to quantify the problem. It's part of NSF's new Research on Research Security Program, mandated by a 2022 law giving the agency increased responsibility for improving research security at U.S. universities and academic medical centers.

Dao and his team hope to survey as many as 1600 faculty members, divided into STEM—science, technology, engineering, and math—or non-STEM fields and by whether their institution conducts more or less than \$50 million a year in research. Those distinctions could be important, Dao says. “What if the rate among non-STEM faculty is much higher?” he asks. “Or what if we learn that the problem is rare at most top-ranked institutions?”

The study will use a definition of foreign influence that covers four types of inappropriate behavior by foreign governments: offering incentives or rewards, demanding that any collaboration remain secret, coercing scientists to participate, and the outright theft of intellectual property. Dao says the survey methodology can collect accurate information about the type and frequency of any questionable behavior without linking the answers to an individual respondent. However, he acknowledges that the chilling effect of the China Initiative could make faculty less willing to report any contacts, much less any missteps.

Dao hopes the results of his survey, to be fielded in the coming months, will address what he sees as two flaws in the current political debate over foreign influence. One is an inherent bias among law enforcement agencies to inflate the problem because they see so many examples of possibly illegal behavior. The second is the reluctance of scientists to consider that their peers are capable of such misconduct.

Dao promises to avoid both tendencies. “People who know me know that I’m driven by data,” he says.

Lauer, for one, trusts Dao to be able to walk that fine line and establish a baseline for foreign influence on U.S. campuses. “Tam is an extraordinarily capable person, and he is quantitatively oriented,” Lauer says. “So I think he would be the obvious person to attempt to answer the question.”

COMMENTARY

Barnes Ice Cap in Baffin Island, Nunavut, Canada, which is the last remnant of a continental glacier, has been receding since 1960.

PERSPECTIVES

GLACIOLOGY

Has the next ice age already come and gone?

The last remnant of a continental glacier shows that global warming may have forestalled the next glaciation

Shaun A. Marcott

Like a metronome, Earth has experienced ice ages on predictable “beats” driven by its planetary orbit around the Sun over the past 2 million years (1). However, future anthropogenic greenhouse gas emissions could place a decisive end to this cycle, forestalling the next ice age for hundreds of thousands of years (2). On page 455 of this issue, Young et al. (3) report the history of Barnes Ice Cap in northeastern Canada in response to natural and anthropogenic climate changes over the past 4000 years. The ice cap is the last remnant of the Laurentide Ice Sheet, a massive continental glacier that once covered nearly 30% of North America during the peak glacial period. The study suggests the glacier began to grow 2000 years ago but started losing ice by 1960 because of preindustrial global warming, likely signaling the end of the next ice age.

The Laurentide Ice Sheet has been growing and shrinking in step with each ice age cycle (4). At the peak of the last glacial maximum [\~20,000 years ago (5)], the ice sheet covered most of Canada and parts of the northern United States. By contrast, during the present interglacial interval (a geologic epoch of warmer global average temperature), it has mostly disappeared. Young et al. used a combination of geologic dating techniques to determine when Barnes Ice Cap was advancing and retreating during the past 4000 years. Surface exposure dating, which is a commonly applied method in alpine terrains (6), determines the duration of time when a rock surface was exposed to the atmosphere and cosmic radiation. Uncovered rocks interact with the cosmic rays, creating a steady buildup of rare isotopes, such as beryllium-10 ( $^{10}$ Be), at known rates. Young et al. measured the ${}^{10}$ Be isotope in both bedrock surfaces that have been newly exposed by the retreating ice and the surface of boulders that lie atop an extensive network of moraines (distinct ridges of unconsolidated debris) left behind when the glacier retreated. Nearly half of the obtained ${}^{10}$ Be ages contained isotopes that were inherited from prior exposures to cosmic rays, which is a common complication in polar or high-latitude environments where the glaciers' base remains frozen to the bed and erodes very slowly (7). Young et al. used a second chronologic technique that measures radioactive carbon-14 ( ${}^{14}$ C) isotopes from previously ice-entombed vegetation that have been preserved below the low-eroding ice. Dead vegetation covered by the glacier no longer takes up ${}^{14}$ C, thus providing an exact time stamp of when the glacier advanced over the materials.

By combining the measured dates with prior work in the region [e.g., (8)], Young et al. found that Barnes Ice Cap achieved its minimum dimension \~3600 years ago and then started to regrow around 2000 years ago. The advancing ice demarcates the beginning of the next glaciation driven by regional cooling from a decline in the amount of incoming solar radiation in summer (9). By the late 1950s, Barnes Ice

Cap had reached its maximum historical position. However, the mass balance of the glacier tipped into the negative in 1960 (10), and the ice cap began its historical marginal retreat. This retreat marks the demise of the next glaciation, following the regular ice age cycle.

Young et al. offer an important benchmark into how an ice sheet responds to climate change, potentially providing information about the beginning of prior glacial periods (11). Identifying glacial inceptions is challenging because the initial ice bodies at the commencement of an ice age are often too small to be detected with traditional methods (12), and pinpointing the exact nucleation location of a massive ice sheet is difficult. Young et al.'s newly acquired chronological data provides a way to determine the precise timing and rate of glacial inception, demonstrating how quickly the glacier responds to shifts in summer temperature. The authors argue that a series of climate feedbacks, such as changes in sea ice, from progressive cooling temperatures likely helped accelerate the observed regrowth of the remnant Laurentide Ice Sheet 2000 years ago. This aligns with previous work that recognizes changes in the amount of incoming solar radiation in summer as a key initial driver of past glaciations (13). It also highlights that other climate effects, such as changing albedo (the amount of the solar radiation reflected by the land surface) from snow and ice cover, are equally important in initiating glacial inceptions (14).

Will another ice age ever transpire? The next glacial inception is generally considered unlikely to occur in the next predicted cycle, but ultimately how many skipped glacial cycles occur is left to choices made by humanity and tied to the future carbon emissions (2). Humans are now geologic agents capable of thwarting millions of years of natural glacial cycles and offsetting Earth's climate for hundreds of thousands of years to come. The absence of future glacial cycles will be profound in the context of the geologic history of the past 2 million years and will have ramifications that will extend to other biological and physical systems that co-evolved with the ice ages (15). The study of Young et al. provokes some pause for thought about what an ice-free world might mean as the pendulum of the ice ages comes to rest.

---

REFERENCES AND NOTES

1. J. D. Hays, J. Imbrie, N. J. Shackleton, Science 194, 1121 (1976).

2. D. Archer, A. Ganopolski, Geochem. Geophys. Geosyst. 6, Q05000 (2005).

3. N. E. Young et al., Science 394, 455 (2026).

4. A. S. Dyke et al., Quat. Sci. Rev. 21, 9 (2002).

5. P.U. Clark et al., Science 325, 710 (2009).

6. K. Nishiizumi et al., J. Geophys. Res. 94, 17907 (1989).

7. P.T. Davis, J. P. Briner, R. D. Coulthard, R. W. Finkel, G. H. Miller, Quat. Res. 65, 156 (2006).

8. G.H. Miller et al., Clim. Past 19, 2341 (2023).

9. B. S. Lecavalier et al., Proc. Natl. Acad. Sci. U.S.A. 114, 5952 (2017).

10. O.H. Loken, R. B. Sagar, Int. Assoc. Sci. Hydrol. Publ. 79, 282 (1967).

11. J. D. Ives, J. T. Andrews, R. G. Barry, Naturwissenschaften 62, 118 (1975).

12. A. C. Mix, W. F. Ruddiman, Quat. Res. 21, 1 (1984).

13. M. Milankovitch, Kanon Der Erdbestrahlung Und Seine Andwendung Auf Das Eiszeitenproblem (Royal Serbian Academy, 1941).

14. W.F.Budd, I.N.Smith, Geogr. Phys. Quat. 41, 279 (1987).

15. K. D. Burke et al., Proc. Natl. Acad. Sci. U.S.A. 115, 13288 (2018).

---

ACKNOWLEDGMENTS

The author acknowledges support from the US National Science Foundation and the University of Wisconsin–Madison.

10.1126/science.aem3295

METABOLISM

Putting glucose metabolism under electronic control

An electronic device controls the redox mechanisms that regulate glucose concentrations in mice

W. E. Bentley $^{1}$ and G. F. Payne $^{1}$

Electrodes have been important tools for biological discovery and disease treatment. Luigi Galvani and Alessandro Volta discovered “animal electricity” decades before the theory of bioelectricity was accepted, and Alan Hodgkin and Andrew Huxley studied how action potentials propagate electrical signals in cells before the structure of DNA and the genetic code were resolved. Today, electrical measurements are routinely used to assess cardiovascular health (e.g., electrocardiograms) and brain activity (e.g., electroencephalograms). Electrodes also enable minimally trained individuals to perform life-saving interventions using publicly available defibrillators. In all these examples, an electrode influences ion-based electrical processes in an animal’s body. On page 400 of this issue, Maity et al. (1) report an implantable electrode that interfaces with biology through an entirely different mechanism based on reduction-oxidation (redox) reactions. This approach, which they tested in mice, targets type 2 diabetes, a debilitating disease that affects millions of people worldwide.

Redox reactions are ubiquitous in biology. They can occur in isolation, such as those that generate reactive oxygen species, but they can also assemble into spatially organized pathways in the cell, such as the respiratory electron transport chain in mitochondria. Redox reactions are also embedded within metabolic networks that balance the production and degradation of complex molecules (anabolism and catabolism, respectively). In addition to metabolism, redox processes are involved in regulatory and protein interaction networks that transduce environmental, social, and lifestyle exposures into biological responses $(2-5)$ . For example, electrodes can provide the cues that act on redox-responsive transcription factors and thus affect gene expression $(6)$ , a concept known as “electrogenetics” $(7, 8)$ . Maity et al. used an electrode that interacts with the body’s biological processes at a metabolic level: It oxidizes glucose electrochemically. This approach could become a drug-free intervention to treat type 2 diabetes and obesity.

How an electrode influences biological redox mechanisms is fundamentally different from how it monitors or affects the movement of ions. The charged particles in a redox network are electrons that flow through electron transfer reactions. This poses several challenges for using electrodes to influence redox biology (9). Because they target localized reactions, redox-based interventions are usually more invasive than ion-based interventions that rely on electric fields generated from electrodes applied to the surface of the skin. In the study of Maity et al., the electrodes that enabled electrochemical glucose oxidation were implanted subcutaneously in mice. Redox reactions also often require a high amount of energy for activation, which biology overcomes using enzymes, proteins that trigger specific molecular reactions. To address this challenge, Maity et al. developed a nonbiological graphite felt matrix with platinum-decorated gold nanoclusters as an electrocatalyst that promotes the selective electrochemical oxidation of glucose. Implanted electrodes are also sensitive to damage, which reduces their active lifetime. To preserve the electrode's performance over several days and even weeks, Maity et al. encapsulated it in a protective three-layer structure composed of cellulose, alginate, and albumin. Because glucose oxidation occurs at the electrode's surface, the oxidation rate scales with the electrode's surface area, whereas the need to remove excess glucose from blood scales with the animal's volume. Maity et al. used a 0.5-cm $^{2}$ electrode to control blood glucose concentrations in their mice studies. To ensure that the electrochemical stimulation was sufficiently harmless while conferring benefits, the intervention used low voltages (<1 V). This was sufficient to generate gluconate (a product of glucose oxidation), which is cleared by the kidneys, and to promote a shift from glucose-based to fat-based metabolism.

Maity et al. integrated these electrochemical capabilities into a wireless electrocatalytic closed-loop health optimization device, which they coupled to a commercially available continuous glucose-monitoring system. This setup converts biological information (glucose concentrations) into an electrical signal. This signal is analyzed through standard electronic technologies to generate an electrical response that is used to guide the electrode's intervention (oxidizing glucose).

Efforts to scale up this approach for human intervention would need to consider how to accommodate a greater blood volume and ensure stability in a different immune and metabolic milieu. An implantable electronic device that automatically detects and corrects excessive glucose concentrations, just as implantable defibrillators detect and correct life-threatening arrhythmias, could have advantages over current devices that detect glucose levels and release insulin to treat diabetes. These advantages include the possibility of human-and drug-free (rather than hormone-mediated) operation. The electronic nature of such a device would also facilitate communication with a smartphone for health monitoring and human-independent learning, which would enable personalized diabetes management. Existing bioelectronics devices provide notable access to the nervous and neuromuscular systems. A redox-based version of bioelectronics could reach other physiological systems, such as the immune and endocrine systems and the gastrointestinal tract, which are not so heavily dependent on ion-based electrical regulation but are strongly influenced by redox networks. Therefore, future electronic intervention devices might help to manage diseases other than diabetes, modulate human-microbiome interactions, or maintain redox homeostasis in stress conditions, among many other applications.

---

REFERENCES AND NOTES

1. D. Maity, P. Guha Ray, P. Buchmann, G. Charpin-El-Hamri, M. Fussenegger, Science 394, eady8560 (2026).

2. D. P. Jones, H. Sies, Antioxid. Redox Signal. 23, 734 (2015).

3. H. Sies, R. J. Mailloux, U. Jakob, Nat. Rev. Mol. Cell Biol. 25, 701 (2024).

4. H. Sies, D. P. Jones, Nat. Rev. Mol. Cell Biol. 21, 363 (2020).

5. D. P. Jones, Free Radic. Biol. Med. 217, 179 (2024).

6. W. Weber et al., Nucleic Acids Res. 37, e33 (2009).

7. T. Tschirhart et al., Nat. Commun. 8, 14030 (2017).

8. J. M. Lawrence et al., Sci. Adv. 8, eabm5091 (2022).

9. E. Kim et al., Adv. Sci. 12, e10184 (2025).

---

ACKNOWLEDGMENTS

The authors acknowledge support by the Defense Threat Reduction Agency (grant HDTRA1–19–0021), the Gordon and Betty Moore Foundation (grant GBMF11395), and the National Science Foundation (grant 2607564).

10.1126/science.aem0762

CELL BIOLOGY

A cell death ballet in two acts

Two proteins that permeabilize mitochondria promote cell death through distinct mechanisms

Cristina Muñoz-Pinedo $^{1}$ and Seamus J. Martin $^{2}$

Complex multicellular organisms require ongoing replacement of cells for many reasons, including aging, infection, and injury. In many situations, cell death occurs in response to stress through a regulated process called apoptosis (1). A key pathway to apoptosis involves two related proteins, Bcl-2-associated X protein (BAX) and Bcl-2-homologous antagonist killer (BAK), that form pores in mitochondria (2, 3). This releases cytochrome c into the cytosol to trigger activation of caspase proteases that coordinate apoptosis (4). BAX-BAK pores can also lead to delayed release of mitochondrial DNA (mtDNA) into the cytosol and the production of type I interferons (IFNs), which is a classical antiviral immune response (5, 6). On page 401 of this issue, Yang et al. (7) report two distinct BAX-BAK complexes—one for cytochrome c release that promotes immunologically silent caspase-dependent apoptosis and another that promotes mtDNA release, caspase-independent cell death, and IFN production to stimulate inflammation.

An overarching objective of apoptosis is the surgical removal of

---

Dual acts at the mitochondria

In response to cell stress, distinct BAX-BAK complexes assemble to facilitate the release of factors that trigger distinct cell-death pathways. Complex I is immune silent, whereas complex II activates a robust immune response.

ANT1, adenine nucleotide translocator 1; BAK, Bcl-2–homologous antagonist killer; BAX, Bcl-2–associated X protein; cGAS, cyclic GMP-AMP synthase; IFN α/β, interferons α and β; PHB1,2, prohibitins 1 and 2; STING, stimulator of interferon genes; VDAC1,2,3, voltage-dependent channels 1, 2, and 3.

cells from tissues with minimal collateral damage to neighboring cells and the avoidance of inflammation (1, 8). This is largely facilitated by the swift recognition and removal of apoptotic cells by phagocytic cells before the loss of plasma membrane integrity, which precludes the release of proinflammatory molecules from dying cells (1, 9). For these reasons, apoptosis is typically immunologically silent. This is augmented by caspase-mediated inactivation of proteins that play key roles in promoting immune responses, such as cyclic GMP-AMP synthase (cGAS), interferon regulatory factor 3 (IRF3), and interleukin-33 (IL-33) (1, 10). Consequently, when apoptosis-associated caspase activity is suppressed—either pharmacologically or by pathogen-encoded caspase inhibitors—this can convert an otherwise immunologically silent mode of cell death event into a proinflammatory one (1, 5, 6, 8).

Diverse forms of cell stress, including DNA damage, protein misfolding, and viral infection, promote activation of BAX and BAK, which leads to mitochondrial cytochrome c release and formation of the apoptotic protease-activating factor 1 (APAF1)-caspase-9 apoptosome—a protein complex that activates “executioner” caspase-3 and caspase-7 downstream (4). This results in controlled cellular demolition through caspase-mediated proteolysis of hundreds of proteins, producing the classical apoptotic phenotype (1, 11). Because of the importance of BAX-BAK-mediated mitochondrial permeabilization for cell fate, the intricacies of this process have been intensively studied (2, 3, 12–14). Consequently, it is very well established that cell stress-initiated activation of one or more BH3-only proteins (those bearing a Bcl-2 homology domain 3) mobilizes BAX and/or BAK, resulting in their progressive oligomerization within the mitochondrial outer membrane to form pores or channels that permit cytochrome c efflux. BAX-BAK pores can subsequently coalesce to form macropores that have been implicated in mtDNA release, but how this occurs has been the subject of debate (12–14). BAX-BAK-dependent mtDNA release has important consequences because this can activate the cGAS-stimulator of interferon genes (STING) pathway to trigger the production of antiviral IFNs, but this is typically counteracted by caspase-mediated proteolysis of cGAS and IRF3 (10). However, in cells that lack APAF1 or where caspase activity is suppressed, mtDNA-driven IFN production is unleashed (5, 6).

Previous studies proposed that apoptosis-associated mtDNA release results from BAX-BAK macropore-driven herniation of the mitochondrial inner membrane through the outer membrane, leading to eventual rupture of the inner membrane and mtDNA release through such hernias (12, 13). However, Yang et al. provide evidence for the formation of two distinct BAX-BAK complexes in mitochondria: complex I, which has a low molecular weight (\~160 kDa) and facilitates cytochrome c release, and complex II, which has a high molecular weight (\~600 kDa)—comprising BAX, BAK, voltage-dependent anion-selective channel proteins 1 to 3 (VDAC1, VDAC2, and VDAC3), adenine nucleotide translocator 1 (ANT1), and prohibitins 1 and 2 (PHB1 and PHB2)—and is essential for mtDNA release and activation of the cGAS-STING pathway. In an unexpected twist, the study suggests that complex II is required for cell death (which the authors call baciptosis) under conditions where apoptosome-mediated caspase activation is suppressed (see the figure).

Yang et al. demonstrate that mammalian cells that are deficient in complex II (lacking VDAC1, VDAC2, or PHB2) can still release cytochrome c and undergo BAX-BAK-dependent mitochondrial herniation but exhibit greatly attenuated mtDNA release and IFN production. Precisely how mtDNA release is facilitated by complex II is not clear, although evidence suggests that BH3-only proteins can induce conformational remodeling of VDAC channels within this complex, leading to calcium influx and mtDNA efflux. Unexpectedly, the authors observed that mtDNA release is required for caspase-independent cell death. Cells deficient in VDAC1 or VDAC2 exhibited BAX-BAK-dependent cytochrome c efflux, but not mtDNA release, and survived for several days under conditions where downstream caspase activation was suppressed. Ablation of mtDNA also attenuated caspase-independent cell death under conditions where complex II was present. The study's findings challenge the prevailing view that caspase-independent cell death is a passive consequence of mitochondrial outer membrane permeabilization and cytochrome c release leading to mitochondrial dysfunction. However, how cytosolic mtDNA instigates caspase-independent cell death remains to be determined, although the cytosolic DNA sensors cGAS, Z-DNA binding protein 1 (ZBP1), and interferon-inducible protein 16 (IFI16) do not appear to be involved.

The study of Yang et al. suggests that BAX-BAK activation can result in two distinct modes of cell death, each guided by distinct BAX-BAK complexes that promote either cytochrome c-initiated caspase activation leading to apoptosis or mtDNA-initiated activation of an unknown cytosolic effector or effectors resulting in caspase-independent cell death. Notably, whereas the former mode of death is immunologically silent, the latter is accompanied by a robust antiviral immune response (5–7). The complex II-mediated route to cell death may have evolved as a backup to counter viruses and other pathogens that harbor caspase inhibitors to suppress the apoptosome route to cell death. Such an eventuality would warrant the initiation of a strong inflammatory response, which apoptosis typically mitigates against (1, 8), as well as a fail-safe route to cell death that operates independently of caspases. Thus, assembly of complex II to drive mtDNA release could be viewed as a form of effector-triggered immunity that is initiated in response to the detection of pathogen activity (i.e., caspase inhibition) as opposed to the detection of pathogen-associated molecular patterns, which is the predominant form of immunity in higher organisms (15). There are also instances where APAF1 expression is epigenetically silenced in certain cell types, such as neurons and cardiomyocytes, or in certain cancers, where complex II-initiated caspase-independent cell death may also play a role.

Questions that remain unresolved by the current study include the identity of the nucleic acid sensor that initiates cell death on detection of mtDNA in the cytosol. How complex II promotes mtDNA efflux and whether it is inhibited by caspases also remains to be clarified. The mechanism of mitochondrial permeabilization identified by Yang et al. sets an exciting stage for the next steps in the BAX-BAK cell death ballet. □

---

REFERENCES AND NOTES

1. S. J. Martin, C. M. Henry, S. P. Cullen, Mol. Cell 46, 387 (2012).

2. H. L. Glover, A. Schreiner, G. Dewson, S. W. G. Tait, Nat. Cell Biol. 26, 1434 (2024).

3. P. E. Czabotar, A. J. Garcia-Saez, Nat. Rev. Mol. Cell Biol. 24, 732 (2023).

4. X. Jiang, X. Wang, Annu. Rev. Biochem. 73, 87 (2004).

5. M.J. White et al., Cell 159, 1549 (2014).

6. A. Rongvaux et al., Cell 159, 1563 (2014).

7. S. Yang et al., Science 394, eaea2181 (2026).

8. M. Beltrán-Visiedo et al., Nat. Rev. Mol. Cell Biol. 26, 884 (2025).

9. I. K. H. Poon, K. S. Ravichandran, Annu. Rev. Pharmacol. Toxicol. 64, 339 (2024).

10. X. Ning et al., Mol. Cell 74, 19 (2019).

11. R. C. Taylor, S. P. Cullen, S. J. Martin, Nat. Rev. Mol. Cell Biol. 9, 231 (2008).

12. K. McArthur et al., Science 359, eaao6047 (2018).

13. J. S. Riley et al., EMBO J. 37, EMBJ201899238 (2018).

14. K. Cosentino et al., Mol. Cell 82, 933 (2022).

15. B. C. Remick, M. M. Gaidt, R. E. Vance, Annu. Rev. Immunol. 41, 453 (2023).

---

ACKNOWLEDGMENTS

S.J.M. acknowledges support from the European Research Council advanced grant program (DESTRESS, AdG 101020534).

10.1126/science.aem1247

$^{1}$ Institut d'Investigació Biomèdica de Bellvitge (IDIBELL), Barcelona, Spain. $^{2}$ Molecular Cell Biology Laboratory, Department of Genetics, Smurfit Institute of Genetics, Trinity College, Dublin, Ireland. Email: martinsj@tcd.ie

The cognitive power of the avian brain

Carrion crows make decisions faster than monkeys and humans Michael Colombo and Oriane Poutrain-Sorel

Once regarded primarily as creatures of instinct and simple learning, birds are revealing cognitive abilities that rival and sometimes surpass those of primates. Even the pigeon, an animal not typically associated with high cognitive skills, can distinguish between images of cancerous and healthy

cells, represent the concepts of same and different, perform numerical discriminations, and tell the difference between words and nonwords in ways that resemble aspects of human reading. Corvids (crows, ravens, scrub jays, and rooks) show even more impressive abilities, such as self-recognition, tool use, problem solving, episodic memory, theory of mind, and planning for future events, at levels that often exceed similar abilities in primates. On page 445 in this issue, Schmidbauer and Nieder (1) report that carrion crows can make ultrafast and accurate decisions.

These findings add to the growing appreciation of the abilities of the avian mind.

Having the same operating principles narrows the cognitive gap between birds and primates, but it does not fully explain the high levels of cognition and fast decision-making abilities of birds. Another explanation is that although small, the avian brain has a disproportionately large number of neurons. A rhesus monkey brain is eight times larger than that of a raven (another species of corvid with impressive cognitive abilities), but it only has three times the number of neurons (6.3 billion versus 2.2 billion). The source of this disparity is the size of neurons: Bird brain neurons are smaller than those in the primate brain, and therefore more neurons can be packed into the same volume of tissue (see the figure) (6).

However, if the rhesus monkey brain has three times the number of neurons as the raven brain, should that not still confer a cognitive

Schmidbauer and Nieder trained carrion crows to report whether random dots appearing on a computer screen moved either up or down, or left or right. Once the dots began to move, the crows reported their direction by pecking a target that appeared above or below the dots moving up and down, or on either side of the dots moving left and right. The crows took \~235 ms to make their decision and respond, a speed that is twice as fast as that of rhesus monkeys (2) and humans (3) performing a similar task. Furthermore, the accuracy of the crows' decisions remained very high. Schmidbauer and Nieder also measured the crows' neuronal activity in the nidopallium caudolaterale, an area of the brain that is functionally equivalent to the primate prefrontal cortex. They noted that direction-selective neurons activated \~70 ms after the onset of the moving dot stimuli and \~170 ms before the birds produced a response.

What factors account for the ultrafast decisions of crows and their impressive cognitive abilities? Researchers often attribute the cognitive performance of pri-

---

Punching above one's weight

Birds display high cognitive skills despite having much smaller brains than mammals. Birds do not have a layered brain organization but have smaller neurons that are arranged in clusters. A high neuronal density and short interneuron distances support fast information processing and decision-making.

Carrion crow

Corvus corone

Rhesus macaque

Macaca mulatta

mates to a large brain with billions of neurons and to a cortex that processes information through recurrent (repetitive) neural circuits that span six cortical layers. However, birds' brains are much smaller and have no cortex, which would suggest they have far less computational power. How then can their cognitive abilities and decision-making speed match or exceed those of primates? Part of the answer lies in their brain architecture. In contrast to the layered organization of the primate cortex, neurons in the avian pallium (which contains the nidopallium caudolaterale), an area broadly analogous to primate cortex, are arranged in clusters (4). Despite this difference, the repetitive neural circuitry of the primate brain is also present in the avian brain, with radially organized fibers that process sensory information connecting to tangentially running fibers that associate information across different brain regions (5). Functional abilities arise from this circuitry, regardless of the brain's overall architecture. Thus, the avian brain processes information in a manner similar to that of the primate brain.

advantage to the monkey? Not necessarily. Birds have proportionally more of their neurons in the pallium compared with the proportion of brain neurons in the primate cortex (6, 7). For example, monkeys have approximately 1.7 billion cortical neurons, whereas ravens have 1.2 billion pallial neurons (6). This would place ravens on nearly the same cognitive playing field as monkeys. The difference further diminishes when the comparison focuses on the number of neurons in the subregions of the cortex or pallium that are most involved in cognition, such as the prefrontal cortex of primates and the nidopallium and mesopallium of the avian brain. For example, rooks (a corvid species smaller than a raven) have the same number of neurons as chimpanzees—around 600 million—in these specialized regions involved in cognition (7).

The densely packed neuronal arrangement of birds' brains not only makes the brain compact, but it also makes it fast. Higher neuronal densities result in shorter distances between neurons, which leads to faster conduction times and a higher speed of information processing (6, 7). These effects are likely to explain the ultrafast decisions of carrion crows in the study of Schmidbauer and Nieder. Together with other emerging evidence on the remarkable cognitive abilities of birds, these findings prompt a rethink about the power that is packed into the small avian brain. □

---

REFERENCES AND NOTES

1. P. Schmidbauer, A. Nieder, Science 394, 445 (2026).

2. T. Hanks et al., eLife 3, e02260 (2014).

3. J. Palmer et al., J. Vis. 5, 376 (2005).

4. H.J. Karten, Ann. N. Y. Acad. Sci. 167, 164 (1969).

5. M. Stacho et al., Science 369, eabc5534 (2020).

6. S. Olkowicz et al., Proc. Natl. Acad. Sci. USA 113, 7255 (2016).

7. F. Ströckens et al., J. Comp. Neurol. 530, 1588 (2022).

10.1126/science.aem0341

BOOKS ET AL.

REFLECTION

The microscope and the metaphor

---

Science shows us what to do to save the planet—poetry helps us understand why we should

Phillip Levin $^{1}$ and David Hassler $^{2}$

In January 2017, as the incoming Trump administration removed climate change information from the White House website and imposed new restrictions on communications by federal scientists, poet Jane Hirshfield responded not with an essay or a protest speech but with a poem. Three months later, this poem, “On the Fifth Day,” appeared as a full-page advertisement in The Washington Post, a week before Hirshfield recited it during the March for Science on the National Mall (1).

The poem did not provide new evidence about climate change or biodiversity loss. Science had already done that. Instead, it reminded people why that evidence mattered. It gave voice to grief, responsibility, hope, and wonder at a moment when scientific facts alone seemed insufficient. As Hirshfield has often said, the microscope and the metaphor are both instruments of discovery.

---

POETRY ARTICULATES WHY EVIDENCE MATTERS

The relationship between poetry and science is often framed in terms of communication: Poetry and other art forms can help explain science to broader audiences. While this certainly is true, it understates the unique contributions each can bring to the conversation.

Science helps us understand what is happening, why it is happening, and which interventions are most likely to make a difference. Poetry can address a different challenge: helping people care enough to engage by activating emotion, imagination, identity, memory, and values—the qualities that often determine whether science becomes personally meaningful (2, 3).

Rather than just serving as a way to communicate science, poetry can help create the emotional and cultural conditions that make such knowledge meaningful. Put simply, science can help us ask “What should I do?” and “How will I know if it worked?” whereas poetry and other art forms can help us ask “Why should I care?”

---

FOSTERING CONNECTION FOR CONSERVATION

A growing body of research supports the complementary relationship between poetry and other art forms with conservation science in particular. People attending programs that intentionally combine art and science report learning more about biodiversity while expressing greater willingness to participate in conservation-related activities (4). Arts-based conservation education has also improved understanding while correcting misconceptions about threatened species (5). More broadly, studies suggest that artistic experiences can foster empathy, curiosity, and emotional connection in ways that make scientific information more memorable and relevant [e.g., (6)]. Reviews of environmental art likewise suggest that aesthetic experiences can strengthen connections to nature and increase intention to engage in proenvironmental behavior (7).

None of this suggests that poetry or other art forms can replace conservation science. Rather, the evidence points toward a productive complementarity: Art can strengthen motivation to engage, while science provides evidence that can help channel that motivation toward effective action. This relationship echoes a broader body of work arguing that conservation depends not only on ecological knowledge but also on human values, experiences, and relationships.

---

POETRY BROADENS WHAT COUNTS AS KNOWLEDGE

Conservation decisions are rarely based on ecological information alone. They are also shaped by history, culture, identity, memory, and lived experience. These forms of knowledge can be difficult to capture in maps, models, or journal articles, yet they can strongly influence whether conservation efforts succeed or fail. Artistic practices—including storytelling, poetry, photography, performance, and other creative methods—offer ways to make lived experience, cultural memory, and place-based knowledge visible without requiring them to conform to scientific conventions (8).

The boundary between scientific and artistic ways of knowing is permeable. Scientists increasingly use poetry and other creative forms to explore dimensions of their work that conventional scientific writing cannot capture: grief, wonder, uncertainty, attachment, and awe. Such work does not replace scientific inquiry; it reveals something about the human experience of conducting it.

To this end, as I (Levin) began working with 100s of colleagues on The Nature Record—the first-ever scientific assessment of nature in the US (9)—in 2024, my coauthor (Hassler), along with poets Luisa A. Igloria and Aileen Cassinetto, developed a companion anthology of poetry entitled The Nature of Our Times: Poems on America's Lands, Waters, Wildlife, and Other Natural Wonders (10). The anthology brings together voices from the arts, ecology, academia, and Indigenous communities, including marine ecologist Maya Zeff, whose poem “I translate the sea” lyrically captures the mutualisms that define marine ecosystems, and wildlife ecologist Lara Volski, whose poem “The Monarch Bear” incisively describes the reciprocal influence of people and animals on one another (see sidebar). Zeff and Volski’s work illustrates how scientific observation and poetic attention can emerge from—and contribute to—the same sustained engagement with the natural world.

The importance of multiple ways of knowing is particularly evident in work with Indigenous peoples and local communities. Arts-based approaches here have helped document cultural connections to the ocean, illuminate histories embedded in landscapes, and create space for multiple forms of knowledge to inform conservation planning (11).

In 2020, Anishinaabe poet, scholar, and environmental activist Kimberly Blaeser founded In-Na-Po—Indigenous Nations Poets—a national Indigenous poetry community committed to mentoring emerging writers, nurturing Indigenous poetic practices, and raising the visibility of Native writers (12). Blaeser and In-Na-Po have collaborated with other organizations that connect poetry, Indigenous knowledge, language, community, and relationships with the natural world. Their 2023 mentoring retreat in Door County, Wisconsin, produced “Poems for a Tattered Planet,” a series of poetic films created in collaboration with Overpass Light Brigade (13). Rather than competing with scientific evidence, such collaborations can enrich the knowledge available to conservation, contributing to a more complete understanding of the ecological and social systems conservation seeks to sustain.

---

SCIENCE HELPS DIRECT ACTION

Motivation alone is not enough. The history of conservation is full of well-intentioned actions that failed because they were based on incomplete understanding or ineffective interventions. Science provides an evidence-based foundation for deciding where to act, which actions are most likely to succeed, how to evaluate trade-offs, and whether desired outcomes are actually achieved. Whereas poetry can help generate the public will to conserve nature, science ensures that effort is directed toward actions that benefit biodiversity rather than simply making us feel that we have done something useful.

The relationship is therefore reciprocal rather than hierarchical. Art without science risks inspiring ineffective action; science without art often struggles to inspire action at all. Neither is sufficient by itself.

---

THE NEXT FRONTIER

Despite the growing interest in art-science collaborations, rigorous evidence about their long-term conservation effects remains limited. Much of the recent literature consists of conceptual frameworks, synthesis papers, project descriptions, and reflections on emerging practice rather than evaluations of conservation outcomes [e.g., (14)]. This work nevertheless identifies promising new roles for artists—not simply as communicators of scientific findings but as collaborators who can help shape research questions, reveal overlooked perspectives, and catalyze new forms of problem solving.

An important next step is to test the full pathway from artistic experience to motivation, from motivation to evidence-based decisions, and ultimately to lasting changes in human behavior and biodiversity. Comparative and longitudinal studies will be important for identifying which artistic approaches are most effective for different audiences and conservation contexts, and under what conditions art-science collaborations generate outcomes beyond those achieved through conventional conservation communication or engagement alone $(7, 14)$ .

It is within this emerging field that the Wick Poetry Center at Kent State University will convene the first Poets for Science Gathering in November 2026 (15). The gathering will bring together poets, scientists, educators, clinicians, students, environmental advocates, policymakers, artists, and the public to explore the meeting ground between science and poetry.

The meeting will celebrate art-science collaborations but also challenge them. When does a poem (or, in other contexts, performance art, visual art, or story) change not only what someone feels but what they do? Which collaborations produce lasting effects, for whom, and under what circumstances? And can those changes ultimately be linked to healthier ecosystems and more resilient communities?

Science reveals how the natural world works and how it is changing. Poetry reminds us why that world matters. Conservation needs the tools of both—the microscope and the metaphor. Understanding how they work together may become one of the field's most important questions.

---

REFERENCES AND NOTES

1. J. Hirshfield, "On the Fifth Day," The Washington Post, 14 April 2017.

2. D. J. Curtis, N. Reid, I. Reeve, S. A. P. I. EN. S. 7, 1655 (2014).

3. S. Dupont, J. Mar. Biol. Assoc. U.K. 97, 1211 (2017).

4. O. Opermanis, S. N. Kalnins, A. Aunins, J. Nat. Conserv. 28, 67 (2015).

5. M. Franquesa-Soler, L. Jorge-Sales, J. F. Aristizabal, P. Moreno-Casasola, J. C. Serio-Silva, PLOS ONE 15, e0228382 (2020).

6. P.-C. Wang, C.-Y. Yu, Sustainability 10, 1098 (2018).

7. A.-S. Tribot, D. Faget, T. Richard, T. Changeux, Biol. Conserv. 276, 109791 (2022).

8. R. E. Brennan, Ocean Coast. Manag. 162, 110 (2018).

9. The Nature Record; https://naturerecord.org/.

10. L.A. Igloria, A. Cassinetto, D. Hassler, Eds., The Nature of Our Times: Poems on America's Lands, Waters, Wildlife, and Other Natural Wonders (Paloma Press, 2025).

11. M. Strand, N. Rivers, B. Snow, Front. Mar. Sci. 9, 886632 (2022).

12. Indigenous Nations Poets, About In-Na-Po; https://www.indigenousnationspoets.org/about-us.

13. Indigenous Nations Poets, "Poems for a Tattered Planet" (2023); https://www.indigenous-nationspoets.org/poems-for-a-tattered-planet.

14. I. Jarić et al., Front. Ecol. Environ. 23, e70012 (2025).

15. Poets for Science Gathering, Wick Poetry Center, Kent State University, 12 to 14 November 2026; https://poetsforscience.org/gathering-2026/.

10.1126/science.aej4196

LETTERS

Edited by Jennifer Sills

Teams compete to raise tipis at the University of Regina in Saskatchewan, Canada. This annual event reflects the longstanding partnership between the University of Regina and the First Nations University of Canada.

---

Indigenous higher education needs support

Indigenous knowledge systems are central to addressing conservation challenges. Indigenous universities provide spaces where Indigenous languages, knowledge systems, and worldviews thrive. Such universities deserve reliable support from their countries' governments.

From stewardship of protected areas to environmental governance, Indigenous voices have reshaped scientific and policy discussions about humanity's responsibilities toward Earth (1, 2). Institutions such as the First Nations University of Canada, Tribal Colleges and Universities in the United States, and the Federal Indigenous University in Brazil help to revitalize and sustain languages, cultures, and knowledge systems that have guided human-environment relationships for generations (3–5). These Indigenous-led institutions also develop local expertise by training specialists who can bridge knowledge systems to address complex social-ecological interactions, thus strengthening research, governance, and local stewardship capacity (3).

Despite their important role, Indigenous universities remain highly dependent on unstable political support. In Brazil, the establishment of the Federal Indigenous University represents a historic investment in Indigenous-led higher education (6), but elsewhere, institutions continue to struggle. Amawtay Wasi University, the first intercultural university in Ecuador, has faced drastic budget and staff reductions, compromising its operations (7). In Mexico, intercultural universities have faced chronic resource limitations, as well as limited institutional autonomy owing to political interference (8, 9). The Indigenous Universities of Bolivia (UNIBOL) have also faced governance, infrastructure, and budgetary constraints (10), and the University of the Autonomous Regions of the Nicaraguan Caribbean Coast (URACCAN) in Nicaragua continues to operate under historical marginalization, ongoing challenges to Indigenous autonomy, and limited resources (11).

As Indigenous and allied scholars from institutions across the Americas, we applaud the development of new Indigenous universities and urge governments to move beyond symbolic commitments and invest in the long-term stability, autonomy, and growth of Indigenous-led higher education. In the United States, the Tribally Controlled Colleges and Universities Assistance Act of 1978 established a federal framework supporting Tribal Colleges and Universities, a network of Indigenous-led institutions that support language revitalization, cultural preservation, and training (5). Similarly, New Zealand strengthened long-term support for Māori-led wānanga by replacing annual grants with an ongoing statutory funding mechanism to provide greater funding certainty (12). Together, these examples demonstrate how long-term legislative and funding commitments can support the stability, autonomy, and growth of Indigenous-led higher-education institutions. Governments across the Americas should establish long-term funding mechanisms, strengthen institutional autonomy for intercultural and Indigenous-led organizations, and support Indigenous educational sovereignty.

Bruno Eleres Soares $^{1,2}$ , Thaïs Bernos $^{3}$ , Rodrigo F. R. Carmo $^{4}$ , Lucas Colares $^{5}$ , Jaime Grimm $^{6}$ , Alessandra Higa $^{7}$ , Anna-Leah King $^{8}$ , Andrew M. Miller $^{9}$ , Robert Newman $^{10}$ , Andrea M. D. Ortiz $^{11}$ , Andrea Sterzuk $^{8}$ , Vincent Ziffle $^{12}$

$^{1}$ Institute of Environmental Change & Society, University of Regina, Regina, SK, Canada. $^{2}$ Programa de Pós-Graduação em Ecologia, Universidade Federal do Rio de Janeiro, Rio de Janeiro, RJ, Brazil. $^{3}$ Rubenstein School of Environment and Natural Resources, University of Vermont, Burlington, VT, USA. $^{4}$ Departamento de Biologia, Universidade Federal Rural de Pernambuco, Recife, PE, Brazil. $^{5}$ Coordenação de Zoologia, Museu Paraense Emílio Goeldi, Belém, PA, Brazil. $^{6}$ Department of Biology, University of Victoria, Victoria, BC, Canada. $^{7}$ Department of Math, Science, and Technology, Oglala Lakota College, Kyle, SD, USA. $^{8}$ Faculty of Education, University of Regina, Regina, SK, Canada. $^{9}$ Department of Indigenous Studies, First Nations University of Canada, Regina, SK, Canada. $^{10}$ Department of Biology, University of North Dakota, Grand Forks, ND, USA. $^{11}$ Departamento de Geografia, Universidad de Concepción, Concepción, BI, Chile. $^{12}$ Department of Indigenous Knowledge and Science, First Nations University of Canada, Regina, SK, Canada. Email: soares.e.bruno@gmail.com

---

REFERENCES AND NOTES

1. S.J. Bruce, Conserv. Biol. 40, e70254 (2026).

2. E. S. Brondízio et al., Annu. Rev. Environ. Resour. 46, 481 (2021).

3. S. Brant, K. Williams, D. Longboat, Front. Educ. 11, 1702361 (2026).

4. K. Leonard, J. Comp. Int. Hig. Educ. 13, 90 (2021).

5. E. S. Huaman, B. Chiu, C. Billy, Educ. Pol. Anal. Archiv. 27, 101 (2019).

6. Presidência da República, Lei No. 15.418, de 28 de maio de 2026. Cria a Universidade Federal Indígena (Unind) (2026); https://www.planalto.gov.br/ccivil\_03/\_ato2023-2026/2026/lei/I15418.htm.

7. R. Revelo, "Consejo de Educación Superior aprueba intervenir la Universidad Amawtay Wasi, centro educativo habla de denuncia falsa," Primicias (13 July 2026); https://www.primicias.ec.

8. M. Lloyd, Intercult. Educ. 35, 402 (2024).

9. S. Schmelkes, Intercult. Educ. 20, 5 (2009).

10. M. M. Chundary, Rev. Educ. Sup. Soc. 36, 195.

11. J. Cupples, K. Glynn, Singap. J. Trop. Geo. 35, 56 (2014).

12. New Zealand Ministry of Education, Education Report: Providing Greater Funding Certainty to Wananga (24 April 2024); https://web-assets.education.govt.nz/s3fs-public/2026-02/MARKED\~1.PDF.

---

India needs a tree conservation plan

India holds 2603 known tree species, 650 of which are endemic, and 469 of which have already been assessed as threatened (1, 2). In October 2025, India's Ministry of Environment, Forest, and Climate Change launched the National Red List Assessment, a 5-year, nearly US\$10 million effort to evaluate roughly 11,000 species, including 7000 plants, against International Union for Conservation of Nature (IUCN) Red List criteria (3). However, identifying risk is not the same as acting on it. India should augment its substantial assessment investment with a conservation plan for the species found to be vulnerable.

Other countries have already built National Conservation Action Plans for their threatened trees. For example, Guinea, Gabon, Ghana, and Kenya, all of which have far fewer tree floras than India has (1), have each implemented and budgeted for time-bound frameworks that assign priority species to specific institutions for seed banking, habitat protection, or field research (4–8). Guinea's plan covers 76 threatened species. Colombia, the second-most tree-rich country in the world, also has a plan, which covers 701 threatened species (8). Every plan ties conservation action directly to Red List data and includes a formal review cycle.

India should build a similar system, timed to the 2030 deadline it has already set for its National Red Data Books. Existing programs such as the Green India Mission and Compensatory Afforestation Fund support tree planting at scale but do not designate funding or institutional responsibility for named threatened species (9). Every species assessed as threatened should be added automatically to a standing action list rather than awaiting separate political approval. The plan should set explicit conservation targets, such as seed-banked backup for a defined share of threatened species. In addition, the plan should include a 5-year review synchronized with the Red Data Book timeline, so that species assessment and conservation action remain on the same schedule rather than drifting apart.

---

Anurag Dhyani

Jawaharlal Nehru Tropical Botanic Garden and Research Institute, Thiruvananthapuram, Kerala, India. Email: anuragdhyani@gmail.com

---

REFERENCES AND NOTES

1. Botanic Gardens Conservation International (BGCI), "State of the world's trees" (BGCI, 2021).

2. M. Rivers, A. C. Newton, S. Oldfield, Plants People Planet 5, 466 (2023).

3. Ministry of Environment, Forest and Climate Change, Government of India, “National Red List Assessment of Indian flora and fauna: Vision 2025–2030” (2025).

4. C. Couch, Herbier National de Guinée, "Plan d'Action National de Conservation pour les Arbres Menacés de Guinée" (Royal Botanic Gardens, Kew, and Herbier National de Guinée, 2023); https://www.herbierguinee.org/uploads/2/6/3/0/26303479/panc\_arbres\_menaces\_guinea\_v2b\_2023.pdf.

5. GaboTREE Project, “Plans d’actions de conservation multi-espèces des espèces d’arbres les plus menacées du Gabon” (Fondation Franklinia, 2023); https://www.tropicos.org/projectimages/Gabon/GaboTREE/PACarbres.pdf.

6. BGCI, IUCN SSC Conservation Planning Specialist Group, CSIR-Forestry Research Institute of Ghana, "Planning conservation action for Ghana's threatened tree species" (BGCI, 2023); https://www.bgci.org/wp/wp-content/uploads/2023/03/PCAGhanaMedRes.pdf.

7. Y. Harvey-Brown, K. Shaw, Eds., "Planning Conservation Action for Kenya's Threatened Trees" (BGCI, 2020).

8. C. Lopez-Gallego et al., "Plan de acción para la conservación de los árboles de Colombia" (2024).

9. Ministry of Environment, Forest and Climate Change, Government of India, "Year-End Review 2025" (2025).

10.1126/science.ael9835

---

AI biosecurity must include agriculture

Advances in biological artificial intelligence (AI) have brought biosecurity to the center of public debate, and the US Congress is now considering the Biosecurity Modernization

and Innovation Act (1). Yet this debate, like biosecurity more broadly, remains focused largely on human health and overlooks the importance of pathogens that affect crops and livestock. The United States and governments worldwide should prioritize crop and livestock protection by minimizing the likelihood of deliberate pathogen attacks and making the targets less vulnerable to harm.

Agricultural pathogens can threaten food security, economies, and political stability. Southern corn leaf blight destroyed roughly 15% of the US corn crop in 1970, and Panama disease devastated global banana production and severely harmed the economies of several countries, fueling labor strikes and political upheaval across Central America (2–6). Genetic uniformity, dense production systems, and globally connected supply chains in modern agriculture amplify these risks.

Agricultural pathogens also present distinctive risks for misuse. Many important agricultural host-pathogen systems are well characterized, experimental materials are readily accessible, and host specificity can reduce risk to the developer (7–10). Unlike transmissible human pathogens, they may also allow harm to be concentrated in particular crops, livestock populations, or environments, potentially reducing blowback. AI could exacerbate agricultural products' vulnerability to targeted attacks.

Agricultural biosecurity therefore needs a continuous array of defenses spanning prevention through response. The US Department of Agriculture has already called for treating farms and food systems as national security assets (11); that effort deserves strong support and expansion. DNA synthesis screening should cover high-consequence agricultural pathogens as rigorously as human pathogens. AI evaluations should explicitly test capabilities relevant to agricultural misuse.

Biological surveillance should expand to encompass fields, herds, wildlife, vectors, and traded commodities. Governments should also strengthen agriculture-specific diagnostics, veterinary vaccines, resistant cultivars, quarantine capacity, and rapid-response pipelines. Layered defenses can preserve legitimate research while preventing deliberate misuse.

---

Jonathan Feldman $^{1,2}$ , Tal Feldman $^{3}$ , Pardis C. Sabeti $^{2,4,5,6}$

$^{1}$ College of Computing, Georgia Institute of Technology, Atlanta, GA, USA. $^{2}$ Broad Institute of MIT and Harvard, Cambridge, MA, USA. $^{3}$ Yale Law School, New Haven, CT, USA. $^{4}$ Department of Organismic and Evolutionary Biology, Harvard University, Cambridge, MA, USA. $^{5}$ Department of Immunology and Infectious Diseases, Harvard T. H. Chan School of Public Health, Harvard University, Boston, MA, USA. $^{6}$ Howard Hughes Medical Institute, Chevy Chase, MD, USA. Email: pardis@broadinstitute.org

---

REFERENCES AND NOTES

1. US Congress, Senate, Biosecurity Modernization and Innovation Act of 2026, S. 3741, 119th Congress, 2nd Session January 2026; introduced in the Senate on 29 January 2026.

2. S. Marquardt, Am. Hist. Rev. 106, 49 (2001).

3. L.A. Tatum, Science 171, 1113 (1971).

4. P. Balint-Kurti, J. Pataky, Phytopathology 114, 2007 (2024).

5. R. C. Ploetz, Crop Prot. 73, 7 (2015).

6. R. Harding et al., Plant Biotechnol. J. 23, 3628 (2025).

7. T. Nobori, Cell Host Microbe 33, 1227 (2025).

8. M. Urban et al., Nucleic Acids Res. 50, D837 (2022).

9. C. Verdonk, KK Gagalova, S. Raffaele, MC Derbyshire, Comput. Struct. Biotechnol. J. 27, 2881 (2025).

10. A. W. Woodward, B. Bartel, Genetics 208, 1337 (2018).

11. US Department of Agriculture, "National farm security action plan," technical report (US Department of Agriculture, July 2025).

---

COMPETING INTERESTS

P.C.S. is an inventor on intellectual property related to pathogen diagnostics and vaccines and a founder and shareholder at Delve Bio and Lyra Labs.

CLIMATE FINANCE

Fiscal discipline without physical resilience

Reforms could shift fiscal discipline and resilience investment into reinforcing forces rather than competing ones

Aayushi Mishra $^{1}$ , Advait Arun $^{2}$ , Evan Kodra $^{1}$ , Auroop R. Ganguly $^{1,3}$

As extreme weather intensifies, climate-vulnerable nations face two threats: debt default and physical exposure. The International Monetary Fund (IMF) and multilateral development banks manage the first threat well, rewarding fiscal discipline through credit ratings and lending terms, but

no equivalent mechanism rewards resilience investment. This gap can compound over time if a disaster strikes an underprepared economy, and resulting borrowing crowds out the next round of resilience investment—something prior work has modeled as a climate-debt doom loop. What hasn't yet been documented is how the loop unfolds in real time within a single country during a disaster. We trace this empirically using Hurricane Melissa in Jamaica in October 2025 and propose reforms—embedding resilience into debt sustainability analysis, scaling Resilience and Sustainability Trust (RST) funds, and adopting Climate Resilience Debt Clauses (CRDCs)—that could shift fiscal discipline and resilience investment into reinforcing forces rather than competing ones.

Between 2013 and 2024, Jamaica reduced its public debt (i.e., money owed both through loans from international lenders and bonds sold in financial markets) from roughly 144% of its gross domestic product (GDP) to 72% by 2023 (1), with a target of 60% by fiscal year (FY) 2027–2028. This debt reduction happened through years of disciplined spending restraint, primary surpluses, and rigid fiscal rules. Jamaica adopted these rules voluntarily, under its own Fiscal Responsibility Framework, rather than as an IMF condition. Maintaining them also kept borrowing costs low, giving the government further reason to stay the course. In September 2025, Jamaica's sovereign debt rating was upgraded by S&P Global to signal broad institutional confidence in its fiscal trajectory. Then Hurricane Melissa struck a few weeks later as a Category 5 storm, causing an estimated US\$12.2 billion in total damage, losses, and costs (roughly 56.7% of 2024 GDP) (2). Insurance payouts covered a small fraction of losses, with a large share of recovery financing being in the form of new sovereign borrowing. Within weeks, the hard-earned fiscal space that the debt reduction had created was once again under pressure.

---

CLIMATE-DEBT DOOM LOOP

There are many possible reinforcing mechanisms that can initiate a climate-debt doom loop, particularly for slower moving, chronic climate effects on growth and fiscal capacity (3). For the acute, disaster-driven version of the loop that is our focus here, the loop goes through three reinforcing mechanisms: institutional pressure to prioritize debt service over capital investment, a structural gap between disaster risk financing and actual recovery costs, and postdisaster borrowing that recreates the fiscal constraints that precluded investing in resilience in the first place. Here, we track each step of the loop in the weeks after Hurricane Melissa hit Jamaica in 2025, using Jamaica's public financial and disaster loss data as was reported (fig. S1). Two questions guide this tracing, drawing on disaster risk assessment and risk management research, and on mechanism design, respectively: first, how accurately disaster risk tools like insurance and catastrophe bonds capture real losses, and second, how institutional incentives could be redesigned to reward resilience building just as consistently as they reward debt repayment currently.

---

Underinvestment in physical resilience

In FY2024–2025, Jamaica allocated approximately US\$3.15 billion to debt service (\~15.8% of GDP), roughly six times its US\$510 million in total public capital expenditure (\~2.5% of GDP) [see supplementary materials (SM)]. This fiscal structure left coastal defenses underbuilt relative to physical risk, drainage infrastructure systems deferred, and critical systems underinvested in.

This imbalance reflects, in part, how sovereign creditworthiness is currently assessed. Although some credit rating agencies have begun to incorporate climate risk into their supplementary reports, these factors remain peripheral to their core decision-making. The fundamental framework of a sovereign credit rating is still based on a narrow and near-term liquidity logic: Rating agencies prioritize the size of the primary surplus, the pace of debt-to-GDP reduction, and how strictly a nation applies fiscal rules. This same incentive structure likely applies far beyond Jamaica, especially given evidence that climate vulnerability can materially affect sovereign creditworthiness (3, 4). Although dedicated resilience funds like the Green Climate Fund (GCF) exist, they are often narrow in scope and small relative to the broader physical resilience needs discussed here. They also do not address the debt service pressure or the rating agency incentives that exacerbate the doom loop. For example, Jamaica's first ever GCF investment was approved in 2026 for \$50 million for mostly farming resilience projects (see SM). Put differently, though there is resilience funding available, it may be both too small and disconnected from the mainstream systems that determine a government's fiscal flexibility.

---

A structural protection gap

Jamaica's postdisaster financing gap cannot be explained by underinvestment alone. Even countries with robust resilience infrastructure face residual losses due to extreme events, which disaster risk financing instruments like catastrophe bonds and parametric insurance are designed to address (5). Both instruments pay out on the basis of predefined physical measurements like wind speed, storm track, and central pressure exceedances, instead of assessed damages, though the money moves through different markets. With parametric insurance, Jamaica buys a policy directly from a regional insurer [such as the Caribbean Catastrophe Risk Insurance Facility (CCRIF)], which calculates

---

The protection gap

For hurricanes Beryl and Melissa, catastrophe (CAT) bonds and parametric insurance through the Caribbean Catastrophe Risk Insurance Facility (CCRIF) covered under 3% of losses. | Hurricane Beryl (Category 4, July 2024) | Hurricane Melissa (Category 5, October 2025) | Total losses | US$1 billion | US$12.2 billion | Losses (% of GDP) | ~5% | ~57% | CAT bond payout | US$0 million | US$150 million | CCRIF + other payout | ~US$27 million | ~US$92 million | Coverage ratio | ~3% | ~2% | a payout that scales with the storm's measured severity once losses cross a set threshold (see SM). A catastrophe bond works through capital markets instead: for example, the World Bank issues the bond to outside investors, who lose a share of their principal on a similar sliding scale if the storm crosses the same kind of threshold, with the money flowing to Jamaica. Either way, the payout is rapid and objective, but it is also decoupled from the disaster's actual cost. The gap between what a parametric instrument pays and the actual cost of the disaster is known as basis risk.

Jamaica's two recent storms illustrate the consequences. Hurricane Beryl (July 2024), caused an estimated US\$1 billion in damages. Yet, Jamaica's World Bank catastrophe bond paid nothing because its measured intensity fell short of the bond's minimum payout threshold, and the CCRIF covered only US\$26.6 million (2.7% of losses). By contrast, Melissa met every trigger, resulting in a full US\$150 million

---

Decomposition of the post-Melissa increase

Relative to the pre-Melissa debt-to-GDP ratio, the estimated post-Melissa increase is driven by a numerator effect from recovery borrowing and a denominator effect from lost output, following standard debt-ratio accounting. | Stage | Public Debt(% of GDP) | Pre-Melissa baseline | 60% | Recovery borrowing | +6.4 pp | GDP shortfall (lost output) | +2.1 pp | Post-Melissa peak | 68.5% | pp, percentage points.

payout; when combined with other aid through the CCRIF, the total coverage amounted to US\$242 million (2% of losses) (see the first table).

In both cases, the instruments performed exactly as specified, yet for sovereign nations, this results in a gap between contractual performance and fiscal adequacy: Jamaica's disaster risk financing strategy worked, instruments triggered and paid as designed, yet most of the losses were left uncovered. Current institutional architecture has no mechanism to reconcile these two frameworks at present (6).

Expanding parametric coverage carries recurring costs of annual premiums, paid regardless of whether a storm strikes. When concessional borrowing (more generous terms than market lending) is available after a disaster, the long-run costs might be comparable to insurance premium costs. However, parametric instruments offer two key advantages over postdisaster borrowing. Payouts arrived within 3 to 14 days of landfall, and the World Bank catastrophe bond's payout was confirmed within 10 days, compared with the roughly 3 months Jamaica waited for its IMF disbursement to be approved (see SM). Parametric payouts also do not add to the sovereign debt stock. The second advantage is crucial, as every dollar borrowed after a disaster compounds through interest and crowds out future capital spending, whereas an insurance payout does not.

---

Debt-financed recovery

With parametric instruments covering under 2% of the losses, the government's recovery financing came overwhelmingly from new borrowing, roughly US\$3.6 billion (\~18% of pre-Melissa GDP). This single-year borrowing already exceeds the \~15% of GDP flow-financing threshold that the IMF's debt sustainability framework treats as a risk marker for emerging-market sovereigns and approaches the 20% threshold that is applied to advanced economies (see SM). The new post-Melissa debt is spread across the IMF, World Bank, and three regional development banks at mostly undisclosed terms (see SM).

We modeled total nominal obligations (principal plus interest) under terms ranging from concessional to near-market pricing: US\$4.9 billion to US\$5.7 billion (24 to 29% of pre-Melissa GDP). At concessional rates, their net present value falls below the principal, so lending is a real fiscal concession. What binds Jamaica, though, is the annual cash payment, not its discounted value, because those payouts are what compete with the capital budget each year. At peak repayment, the new debt service alone would approach, or exceed, Jamaica's entire annual capital budget (see SM).

This closes the doom loop: Postdisaster borrowing crowds out precisely the predisaster investment that would have reduced losses. Combined, this recovery borrowing and the disaster's dampening effect on output are projected to push Jamaica's debt-to-GDP ratio to a peak of $\sim 68.5\%$ , an 8.5 percentage-point increase over the $\sim 60\%$ trajectory Jamaica was on track for absent the storm (see the figure and the second table).

---

The scale of the investment gap

Using benefit-to-cost ratios from disaster risk literature, we estimate that annual predisaster resilience investment of between US\$360 million and US\$900 million (1.8% to 4.5% of GDP) could have offset that new borrowing entirely, depending on the ratio assumed (see SM). These figures cannot say which investments would have worked best, nor capture implementation lags; still, they show that this estimated gap between the resilience investment Jamaica would have needed and what it has actually been able to invest is large and grounded in established benefit-cost-ratio research. The IMF's 2025 Article IV consultation, a health check of a nation's economy, found that Jamaica could spend only 0.2

to 0.4% of GDP per year (US\$40 million to US\$80 million) on adaptation without straining the debt path, which is 4 to 22 times below what offsetting the postdisaster borrowing would require.

This gap narrows considerably once financing structure is considered: Public financing can cover roughly one-third of total adaptation cost before returns become negative, with private capital expected to fund the remainder (7). If we apply that share to our \$360 million to \$900 million range, it implies a government-financed portion of roughly US\$120 million to \$300 million annually. This is still 1.5 to 7.5 times above what the IMF's Article IV consultation treats as fiscally sustainable, but it is a substantially smaller shortfall than the 4 to 22 times gap mentioned earlier. What Jamaica can sustainably afford under current rules falls short of what would have changed its post-Melissa trajectory, because those rules presently leave too little fiscal room for resilience spending to begin with (see SM).

---

The architecture perpetuates the cycle

The loop persists because each actor responds rationally to the incentives it faces, yet the combined effect leaves both borrowing countries and their lenders worse off than a coordinated approach would. Governments prioritize debt service because creditworthiness signals carry real consequences for market access and borrowing costs. Rating agencies have limited incentives to change this unilaterally: Physical resilience metrics are not yet standardized, their links to sovereign default probabilities are underresearched, and incorporating them would require restructuring entire rating frameworks (4, 8). Reforming sovereign risk frameworks is more workable, given that agencies treat them as credible inputs.

Credit rating agencies reward what their frameworks are designed to measure. In Jamaica's case post-Melissa, activating disaster risk financing instruments, using disciplined concessional lending, and temporarily pausing its own budget rules through its fiscal escape clause (triggered once a disaster's fiscal impact exceeds 1.5% of GDP; Melissa's was \~3.9%) (see SM) were all real improvements in financial risk management. But what current frameworks do not capture is any change in the physical exposure that produced the crisis in the first place. IMF programs like the Debt Sustainability Analysis (DSA) are built to address sovereign debt distress. The August 2024 supplemental note introduced climate risks into DSA for some programs, yet implementation is still not part of the core methodology (8, 9).

Disaster risk financing instruments are real innovations beyond traditional aid mechanisms, with CCRIF designed for rapid postdisaster liquidity, but the limitation is that of scale and design fit: Instruments designed for investor liquidity are not necessarily built for national recovery. The current architecture produces a coordination failure, where every institution operates rationally within its own mandate, yet the collective outcome is chronic underinvestment in resilience for the world's most vulnerable nations, and the same conditions recur wherever debt constraints and physical risk coincide. Here, sovereign fiscal efforts alone cannot close the gap, as adaptation pays for itself only when governments cover the reactive share of adaptation spending (i.e., disaster relief and recovery), whereas the private sector funds proactive measures like resilient infrastructure and early warning systems. That government share averages about one-third of total adaptation costs across the countries (7). If governments instead fund adaptation entirely themselves, the costs outweigh the savings, and even this best-case split would not fully restore debt sustainability for a high-exposure, high-debt country like Jamaica (7).

---

DISRUPTING THE DOOM LOOP

Multilayered resilience frameworks—combining fiscal buffers (reserves or prearranged credit lines that are set aside before a disaster), parametric insurance, and contingent financing like debt-pause clauses—have worked in comparable Caribbean economies $(10)$ . What is missing, however, is the institutional incentive to build these three layers. The interventions below target that gap, so that fiscal discipline and resilience investment reinforce each other instead of competing, benefiting both borrowers and lenders. All are aimed at multilateral institutions, because the IMF, multilateral development banks (MDBs), and sovereign debt markets are more open to coordinated reform (compared to rating agencies, for example) that could shift market signals at scale. Each reform below strengthens one of these layers: Embedding resilience into IMF DSAs builds the incentive structure that ties them together, scaling the RST builds fiscal buffers, recalibrating parametric triggers strengthens the insurance layer, and CRDCs build contingent financing.

---

Embed physical resilience into IMF DSAs

The IMF's existing DSA framework, given its mandate for sovereign stability, relationships with all major borrowing countries, and an analytical capacity, is the most viable path to standardizing how physical resilience is measured and rewarded in sovereign risk assessment. The foundational case for integrating climate risks into stochastic DSA (9) links integrated assessment models with debt dynamics to capture how global warming affects growth, fiscal costs, and sovereign wealth. Jamaica's 2025 Article IV consultation represents a step in this direction, with a climate module projecting debt paths under higher adaptation spending (see SM), yet the analysis asked only how much adaptation the country could afford without straining its debt path. It did not, critically, model the reverse: What happens to debt sustainability when adaptation falls short, and a catastrophe follows? Jamaica's post-Melissa trajectory represents that scenario.

Climate warming may suppress growth, and slower growth erodes debt capacity $(9)$ . In addition to this “slow” channel in a DSA, an “acute” channel could see underinvestment in resilience increase the probability and fiscal severity of a disaster, requiring a complementary approach, in which predisaster investment enters the DSA as a variable that shifts the probability distribution of catastrophic-loss scenarios. As noted earlier, financing reactive adaptation can break even and lower the primary balance needed to stabilize debt $(7)$ . Cutting this spending to make room for debt service does not necessarily save money, but instead gives up a return the government would otherwise capture. This share should be protected from crowding out, and embedding it into the core DSA methodology would do it.

---

Scale up RST funds

IMF's Resilience and Sustainability Trust, established in 2022 to provide concessional financing, should be more explicitly used to expand the fiscal horizons of highly indebted and climate vulnerable nations.

---

Jamaica's public debt-to-GDP trajectory

The historical path shows debt declining from $\sim 144\%$ in 2013 to $\sim 65\%$ by the end of 2024. Following Hurricane Melissa, the debt ratio is projected to peak at $68.5\%$ , an 8.5 percentage point increase over the no-storm counterfactual ( $\sim 60\%$ ).

Early Caribbean pilots have been promising, like Barbados using RST resources to retire high-cost debt and divert those savings toward essential water and food security (11). The instrument is presently underscaled relative to the magnitude of the fiscal constraints, despite a strong economic case for scaling up: The fiscal cost of climate finance at scale to advanced-country donors is modest as a share of GDP, and backing climate-vulnerable economies is in the collective economic interest of financier countries beyond just development obligations (12). The IMF should scale this approach and clearly link RST disbursements to measurable increases in public capital expenditure. In Jamaica's case, this would directly address the crowding-out mechanism due to debt repayment obligations.

Scaling this approach across the roughly 35 to 40 highly indebted, climate-vulnerable small island developing states (SIDS) with active IMF relationships would require cumulative RST disbursements on the order of US\$20 billion to US\$30 billion, values that fall within current RST capacity if climate-vulnerable SIDS are explicitly prioritized (see SM). This requires targeted deployment for countries where RST's concessional terms and long maturities fit the problem, and not for the IMF to serve as the primary creditor for every climate-exposed sovereign. When combined with the fiscal buffers, parametric instruments, and contingent financing, RST-linked commitments carry less of the burden on IMF resources alone (10).

---

CRDCs

CRDCs let a borrower defer principal and interest for up to 2 years after a disaster, which redirects existing capital toward emergency response without new debt. Seventy-three countries and major MDBs like the World Bank and the Inter-American Development Bank have committed to implementation (13). In April 2026, Fitch Ratings revised its Sovereign Rating Criteria to clarify that bounded, rules-based payment deferrals in well-designed pause clauses need not be considered a default, provided triggers and protections are clearly specified (14); Moody's Ratings has engaged but has not issued an equivalent revision. Bondholders, initially resistant, are increasingly viewing well-structured CRDCs as credit enhancing, because payment flexibility lowers the odds of disorderly default.

Contingent debt instruments have been proposed for decades but rarely adopted. GDP-linked bonds, the best-known attempt, had three problems: Triggers were tied to GDP figures that a government could influence, difficult pricing (see SM), and a history of issuance mainly through distressed restructurings. CRDCs address these problems: Storm triggers cannot be manipulated, and standardized development bank templates give markets a common reference that earlier instruments did not. Countries like Barbados have adopted them as a precaution, from a position of fiscal strength rather than in a crisis $13$ . CRDCs complement parametric instruments rather than replacing them: Catastrophe bonds supply liquidity when a storm meets the trigger, and a CRDC defers payments on a declared emergency regardless of storm track.

Coordination remains the main barrier, because the price itself is modest. The World Bank charges a fee of just 0.05% of the disbursed balance for its CRDC option, and the clauses' transparency and disclosure benefits can offset even that $(13)$ . The deeper challenge is how CRDCs have been used, given that most have been added during debt restructurings rather than written into new borrowing upfront, which ties the whole idea to distress and puts off the very countries that would benefit most. Had Jamaica's debt included standard CRDC provisions before Melissa, its debt service payments would have paused automatically, but the benefit only holds if the pause applies across all of a country's debt at once, which means standardizing the clause language across every source, including bilateral lenders, development banks, and private bond markets.

---

Recalibrate parametric trigger design and coverage limits

Parametric instruments need recalibration on two fronts. First, trigger design should move toward multiparameter indices, making it a hazard-modeling problem as much as an insurance one: incorporating rainfall accumulation, storm surge, and wind speed, and defining triggers over smaller, more localized grid cells rather than large regional ones. This will help reduce the basis risk that left Jamaica with zero payout despite Beryl's billion-dollar damages (5). Hybrid parametric-indemnity structures, where a parametric trigger initiates immediate liquidity while a claims adjustment process determines a supplementary indemnity component (i.e., a payout based on actual verified losses), can be a direct path to closing the gap between contractual performance and actual recovery. This is important because traditional loss-verified insurance is not really available to sovereigns on its own, because insuring a whole country's losses this way requires assessors on the ground and takes months (which is exactly why parametric instruments were built in the first place). A hybrid structure would provide both speed from the parametric trigger and accuracy from the verification.

Second, coverage limits should be expanded by having MDBs back these instruments with additional risk capital or guarantees, so the risk pool itself can safely support larger payouts. This works by growing how large a loss the pool can cover, and not by changing how any individual country's premium is set. Premiums can still scale with a country's actual risk profile, even as the pool's total capacity grows. This distinction is important because if premiums stopped reflecting a country's actual exposure, there would be little reason to continue investing in the resilience measures that lower it. MDB capitalization avoids this by expanding the pool's capacity to pay out while leaving each country's premium tied to its own risk.

The cost trade-off remains a crucial point: Whereas multiparameter triggers improve actuarial efficiency, hybrid structures cost more per policy than pure parametric ones, and that added cost needs to be weighed against the borrowing it can help avoid after a disaster. For the most constrained sovereigns, concessional premium financing through climate funds like United Nations' Loss and Damage Fund can absorb it. These reforms would allow disaster risk financing to meet national recovery needs instead of focusing only on supplying market liquidity.

Jamaica's trajectory from over a decade of hard-won fiscal discipline to a sharp reversal of that progress in a matter of weeks is an urgent signal that institutional frameworks are not keeping pace with a changing climate. The institutional conditions that led to Jamaica's outcome are structural and will return wherever creditworthiness measures fiscal discipline without accounting for the physical resilience that helps sustain it. Reforming how sovereign resilience is measured, financed, and rewarded is a prerequisite for making fiscal discipline meaningful as physical risk intensifies. $\square$

---

REFERENCES AND NOTES

1. S. Arslanalp, B. Eichengreen, P. B. Henry, Brookings Pap. Econ. Act. 55, 133 (2024).

2. Planning Institute of Jamaica, Assessment of the socioeconomic and environmental impact of Hurricane Melissa (damage and loss assessment prepared with UN ECLAC and the Government of Jamaica, Planning Institute of Jamaica, 2026); https://jis.gov.jm/total-loss-and-damage-from-hurricane-melissa-estimated-at-1-952-trillion/.

3. S.A. Zenios, Curr. Opin. Environ. Sustain. 67, 101414 (2024).

4. P. Klusak, M. Agarwala, M. Burke, M. Kraemer, K. Mohaddes, Manage. Sci. 69, 7468 (2023).

5. Financial Stability Institute, International Association of Insurance Supervisors, "Uncertain waters: Can parametric insurance help bridge NatCat protection gaps?" (FSI Insights on Policy Implementation No. 62, Bank for International Settlements, 2024); https://www.iais.org/uploads/2024/12/FSI-IAIS-Insights-on-parametric-insurance.pdf.

6. D. J. Clarke, S. Dercon, Dull Disasters? How Planning Ahead Will Make a Difference (Oxford Univ. Press, 2016).

7. M. Calcaterra, A. Consiglio, V. Martorana, M. Tavoni, S. A. Zenios, "Sovereigns on thinning ice: Debt sustainability, climate impacts and adaptation," Working paper 06/2025, Bruegel, 2025; https://www.bruegel.org/working-paper/sovereigns-thinning-ice-debt-sustainability-climate-impacts-and-adaptation.

8. International Monetary Fund, "Staff guidance note on the debt sustainability framework for low-income countries: Supplemental guidance on climate change" (IMF, 2024); https://www.imf.org/en/publications/policy-papers/issues/2024/08/05/supplement-to-2018-guidance-note-on-the-bank-fund-debt-sustainability-framework-for-low-553151.

9. S.A. Zenios, Clim. Change 172, 30 (2022).

10. S. Chen, S. Siegel, C. Tovar Mora, "Fiscal sustainability and natural disaster risks in the ECCU" (Selected Issues Paper 2025/068, International Monetary Fund, 2025); https://doi.org/10.5089/9798229011778.018.

11. J. Rambarran, S. J. Ahmed, “Strengthening the International Monetary Fund’s Resilience and Sustainability Trust to promote a just, global climate transition: Early evidence from Barbados and Jamaica” (Task Force Working Paper 015, Boston University Global Development Policy Center, 2025); https://www.bu.edu/gdp/files/2025/04/TF-WP-015-FIN.pdf.

12. P. Bolton, A. M. Kleinnijenhuis, J. Zettelmeyer, "The economic case for climate finance at scale" (Policy Brief 09/2024, Bruegel, 2024); https://www.bruegel.org/policy-brief/economic-case-climate-finance-scale.

13. Sustainability-linked Sovereign Debt Hub, "Climate resilient debt clauses: State of play" (SSDH, 2025); https://www.ssdh.net/climate-resilient-debt-clauses.

14. Fitch Ratings, "Sovereign Rating Criteria" (Fitch Ratings, 27 April 2026).

---

ACKNOWLEDGMENTS

The authors thank K. Mackenzie and S. Evans for their expert review and feedback on the manuscript. Funding was provided by the Lizzy Warner Fellowship, Northeastern University (A.M.), and the US Department of Defense's Strategic Environmental Research and Development Program (RC20-1183 to A.R.G.). E.K. is also employed by Intercontinental Exchange Data Services but contributes here solely in his capacity as adjunct faculty at Northeastern University. The views expressed here are those of the authors alone and do not reflect the views of Intercontinental Exchange Data Services.

---

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.aeh3608

10.1126/science.aeh3608

$^{1}$ Sustainability and Data Sciences Laboratory, Northeastern University, Boston, MA, USA. $^{2}$ Center for Public Enterprise, Washington, DC, USA. $^{3}$ Artificial Intelligence for Climate & Sustainability, The Institute for Experiential Artificial Intelligence, Northeastern University, Boston, MA, USA. Email: mishra.aay@northeastern.edu; a.ganguly@northeastern.edu

RESEARCH

IN SCIENCE JOURNALS

Edited by

Michael Funk

---

NEUROSCIENCE An epigenetic brake for motor control

Cross-talk between blood vessels and neurons supports central nervous system function. Lu et al. identified a vascular-specific mechanism that was critical for developing neurocircuitry in the motor coordination center of the brain (see also the Focus by Bonanomi). The methyltransferase PRMT5 epigenetically repressed the expression of the gene encoding the vascular remodeling factor ANGPTL4. Mice deficient in brain endothelial cell PRMT5 had high amounts of secreted ANGPTL4, which increased the inhibitory synaptic input to the cerebellum and caused progressive motor deficits. These findings broaden the role of neurovascular coupling in development. —Leslie Ferrarelli Sci. Signal. (2026) 10.1126/scisignal.aed6605; 10.1126/scisignal.aek9091

Blood vessels in the brain visualized in a magnetic resonance imaging-based reconstruction

---

MOLECULAR BIOLOGY

---

ncRNAs tune gene expression dynamics

Most of the genome does not encode proteins but instead produces noncoding RNAs (ncRNAs), the functions of which remain largely mysterious. Using live imaging in developing fly embryos, Batut and Levine found that these ncRNAs help to control when and how strongly protein-coding genes are turned on. For the Hox gene Scr, ncRNAs produced by a distant enhancer delay activation, whereas ncRNAs from a “tethering” element boost expression. Disrupting the enhancer RNA strengthened long-range DNA contacts, leading to premature

activation of Scr. These findings suggest that ncRNAs help to shape genome organization and long-range communication across the genome, ensuring precise timing and levels of gene expression during development.

—Di Jiang

Science p. 402, 10.1126/science.aek5696

---

MACHINE LEARNING Stepwise diagnosis with AI

AI applications for clinical medicine continue to advance but often rely on written vignettes. Unfortunately, these are not a realistic representation of clinical medicine, where patients do not come in with a tidy write-up of key symptoms and pertinent laboratory findings. To address this issue, Nori et al. made their AI system start from the patient's chief complaint and decide what follow-up questions to ask and what testing to order to collect the information needed for a diagnosis. In multiple test settings, this system achieved higher accuracy than human responders and at a lower total cost, demonstrating its potential as a tool to assist physicians.—Yevgeniya Nusinovich

Science p. 423, 10.1126/science.aeh7930

---

SOLID-STATE PHYSICS Elucidating the phonon

The thermal Hall effect (THE), the appearance of a transverse temperature gradient in response to a longitudinal heat current and in the presence of a perpendicular magnetic field, has been observed in many materials. Depending on whether the material is a metal or an insulator and if it exhibits magnetic ordering, THE has been variously ascribed to electrons, magnons, and, more recently, phonons. But just how phonons cause THE remains controversial. Meng et al. studied this effect in the honeycomb antiferromagnet NiPS $_{3}$ . By measuring five different samples, the researchers found that the amplitude of THE was correlated with the degree of magnetic anisotropy. Along with calculations, which indicated the presence of magnon-phonon band crossings in this material, these results are consistent with an intrinsic phonon THE that may be driven by magnon-phonon hybridization. —Jelena Stajic

Science p. 419, 10.1126/science.adt1992

---

MEMBRANES

---

Spinning gates to accelerate flow

Ionic species in the channels of porous membranes can be used to enhance the selectivity of a membrane, but this often comes at the expense of flow rate because of a narrowing of the channels. Zhang et al. incorporated ammonium ions into layered MXene membranes. The self-rotating ammonium functional groups endowed the otherwise-static subnanometer channels with dynamic responsiveness, which enabled rapid and selective transport.

—Marc S. Lavine

Science p. 434, 10.1126/science.aed8319

---

2D MATERIALS Visualizing strongly correlated phases

Moiré superlattices in stacked van der Waals materials host a variety of phases featuring strong electronic correlations. Two groups used scanning tunneling microscopy to visualize such phases and tuned experimental conditions to induce their melting. Berger et al. focused on twisted bilayers of molybdenum diselenide, the triangular moiré superlattice of which hosts crystalline electronic states. As they varied the carrier density, the researchers observed the electronic crystal melt, which became either a uniform liquid-like state when electrons were added or a disordered state when the system was doped with holes. He et al. worked with Bernal-stacked bilayer graphene fully aligned with hexagonal boron nitride. Placing the sample in an external magnetic field reshaped the electronic energy levels, leading to the formation of strongly correlated topological

states. Direct imaging of these states revealed their intricate structure and elucidated the process of their quantum melting. —Jelena Stajic

Science p. 450, 10.1126/science.aef0940,

p. 440, 10.1126/science.aee6967

---

CONSERVATION BIOLOGY Water is coming

Global warming is causing sea level to rise, inundating and shifting coastal ecosystems. Sea-level rise has already led to population declines of coastal species and even an extinction. Jiang et al. estimated the extent of projected future sea-level rise's effects on terrestrial vertebrate diversity at the global scale. They found that one-fifth to two-thirds of species will be exposed to sea-level rise in some part of their habitats, depending on the taxonomic group and projection, with especially high diversity threatened by sea-level rise in Southeast Asia. Considering refugia from sea-level rise in conservation priorities may be important for protecting species in the future. —Bianca Lopez

Science p. 408, 10.1126/science.aed5772

---

NATURAL PRODUCTS Constructing nitrogen-rich compounds

The guanidine moiety is found in many plant- and microbe-derived bioactive small molecules, but installing this functional group is chemically challenging. Kolling et al. identified a group of prenyl transferases involved in the production of guanidine-containing natural products and found that they bind and activate free guanidine as a nucleophile to attack prenyl diphosphates. Crystal structures and biochemical experiments established the mechanism of this transformation. The researchers tested a number of enzymes from plant and bacteria that had similar activity, thus establishing a genetic basis for the production of an important group of natural products. —Michael A. Funk

Science p. 413, 10.1126/science.aeg1821

---

IN OTHER JOURNALS

Edited by Corinne Simonti and Jesse Smith

---

PLANT SENSORS

---

A fitness sensor for plants

Guttation is the release of fluid through small pores on leaf margins called hydathodes, and it occurs when root pressure exceeds transpiration. Guttation provides an accessible readout of a plant's nutritional, metabolic, and stress status, but its small volume and rapid evaporation make continuous sampling difficult. Drawing on noninvasive sweat sensors for humans, Xue et al. designed a microfluidic electrochemical monitoring platform called Plantlonics. Its wedge-shaped microchannel, inspired by the pitcher plant's rim, provides a self-driven, unidirectional flow. A multichannel sensing array measures sodium, potassium, and glucose alongside on-leaf temperature and humidity. This allows noninvasive monitoring of stress-induced changes in these biomarkers over several days across multiple plant species, representing a step toward precision agriculture. —Unnati Sonawala

Nat. Commun. (2026)

10.1038/s41467-026-77955-6

---

CELL BIOLOGY Understanding implantation failure

Female fertility declines sharply after 35 years of age. Cavanaugh et al. investigated age-related infertility in aged mouse models compared with in vitro fertilization (IVF) clinical datasets. Overly strong contractility in mouse embryos from older females impaired embryonic cell-fate specification and implantation. Mouse models replicated age-related defects seen in humans in relation to

pregnancy outcomes and in vitro implantation assays. A peri-implantation age-related defect involved increased surface tension, substrate adhesion, and viscosity of the embryo, which restricted spreading during implantation. Noninvasive imaging signatures predicted implantation success for mouse embryos from young and aged females. Thus, embryo mechanics can regulate reproductive longevity, providing a potential route for screening embryos to improve IVF outcomes.

—Stella M. Hurtley

Nat. Cell Biol. (2026)

10.1038/s41556-026-02052-1

---

2D MATERIALS Cling to the surface

Atomically thin crystals with metallic elements, such as intermetallic compounds and high-entropy metal alloys, can exhibit distinct physical and chemical properties desirable for electronic devices. However, synthesizing two-dimensional (2D) films of these materials has been challenging because of thermodynamic instability and kinetic limitations. Zhang et al. report an adsorption-conversion method that can synthesize a wide range of high-quality intermetallic and high-entropy metal alloy thin films using metal chloride salts. Under a hydrogen environment, adsorbed metal chloride compounds were reduced on the surface, where the released metal atoms lowered the surface energy. Adsorbed chlorine atoms further suppressed vertical 3D growth. The approach produced ultrathin films a few tens of micrometers in size on practical substrates. —Sumin Jin

Nat. Mater. (2026)

10.1038/s41563-026-02752-x

PUBLIC HEALTH

---

Cooled by close, curvy green spaces

Urban green space is associated with cooling and reduced heat-related illness. How that space is organized seems to matter, too, although this has not been well resolved at standard spatial resolutions. Wang et al. overlaid statewide health records for 2016 to 2023 onto 1-meter aerial imagery of more 2000 neighborhoods in Arizona. As expected, more green space was linked to reduced heat-related illness. As important, however, were green space complexity and connectedness. Their findings suggest that urban green-cooling initiatives should consider the shape and linkages of spaces as well as their extent. —Angela Hessler

GeoHealth (2026) 10.1029/2026GH001978

The benefits of urban green space, like Rymill Park in Adelaide, Australia, are correlated with its shape and connectivity.

---

ACTINIDE CHEMISTRY Carbon and uranium triple up

In a Fischer carbyne, a carbon atom forms a triple bond to a metal and a single bond to an oxygen or nitrogen substituent, with some potential for electron delocalization along the axis. A variety of examples have been characterized with transition metals, and Seed et al. now extend the bonding paradigm to uranium. Specifically, reaction of a triamide uranium complex with a diazophosphorus ylide produced a dinuclear complex with the

uranium centers linked by a CNN bridge. Crystallography and spectroscopy as well as theory support a triple-bond character between the carbon and uranium. —Jake S. Yeston

Nat. Chem. (2026)

10.1038/s41557-026-02260-0

TOXICOLOGY

---

Formaldehyde's impact on the brain

Formaldehyde is one of the simplest toxicants, and, if inhaled at high levels, it can cause both acute irritation and serious long-term consequences such as cancer. Associations with depression have also been claimed, but the molecular mechanism of such effects remain poorly studied. Xu and Shen tackled this question by analyzing brain activity of mice exposed to formaldehyde using cyclic voltammetry. They found that this treatment reduced stimulated release of dopamine in the nucleus accumbens. The authors propose a potential association with changes in the gut microbiome, but further mechanistic experiments will be necessary to fully understand and validate these effects. —Michael Funk

ACS Chem. Neurosci. (2026) 10.1021/acschemneuro.6c00121

PLANT BIOLOGY

---

Trees straighten themselves by feel

Trees keep their trunks upright by responding to gravity and light, but it is unclear whether they can also sense and correct their own curvature. Caulus et al. investigated whether poplar trees actively straighten bent stems using a sense of their own shape, called proprioception, rather than springing back passively. By tilting young poplars and then spinning them to cancel out gravity sensing, the authors found that the trees still straightened themselves, growing a muscle-like “tension wood” that shifted from one side of the stem to the other to pull it back into shape. This may mean that trees actively monitor and correct their posture, which could help to improve wood quality and the breeding of crops more resistant to storms. —Di Jiang

New Phytol. (2026) 10.1111/nph.71238

CANCER

---

STAT2 intron retention and immunotherapy

Alternative RNA splicing is a regulatory process that allows a gene to encode multiple messenger RNA (mRNA) isoforms that can produce proteins with different functions. Using long-read RNA sequencing, Englander et al. mapped alternatively spliced mRNA variants in lung adenocarcinoma, a cancer type that is only partially responsive to current therapies. Thousands of mRNA isoforms were identified in immune-related genes, and intron retention events in the Signal Transducer and Activator of Transcription 2 (STAT2) gene led to the production of altered protein isoforms that regulated immune signaling and interferon response. The molecular changes correlated with the response of lung cancer patients to anti-PD-L1 immunotherapy, suggesting that regulated intron retention in the STAT2 gene may contribute to treatment outcome. —Priscilla N. Kelly

—Priscilla N. Kelly

Science p. 397, 10.1126/science.aea7156

SOLAR CELLS

---

Anchors for flexible perovskite modules

Decoupling the deposition of self-assembled monolayers (SAMs) and perovskites enables the fabrication of flexible perovskite modules (PSMs) with high power conversion efficiency (PCE). The stability and PCE of flexible devices fall short of those of their rigid devices because surface hydroxyls desorbing on flexible substrates disrupts SAM assembly. Wu et al. accelerated SAM anchoring with a passivator and reduced their interference with perovskite film formation. A 25-square-centimeter flexible device retained 92.5% of its

Edited by Michael Funk

initial PCE of 22% after 1044 hours of maximum power point tracking at 85°C and 65% relative humidity. —Phil Szuromi

Science p. 398, 10.1126/science.aef9745

---

ARTIFICIAL INTELLIGENCE Tag-teaming AI agents

Conventional drug discovery is a labor-intensive and expensive process, requiring separate teams of people for different steps. The Virtual Biotech approach was designed by Zhang et al. to streamline early drug discovery, with an integrated workflow made up of teams of specialized AI agents, all directed by an agent acting as a chief scientific officer. As with any AI-driven biological research, the agents' recommendations will need to be translated into the real world by synthesizing the candidate drugs and testing them. This approach can only work for diseases for which there are enough background data for the AI to interpret, but it may help to identify therapeutic opportunities that would otherwise be missed.

—Yevgeniya Nusinovich

Science p. 399, 10.1126/science.aeg6779

---

METABOLISM Goodbye to excess glucose

Diabetes and obesity are becoming increasingly common around the world, and they frequently coexist and reinforce each other. A variety of pharmacological and biological interventions are already available, but each has its advantages and disadvantages and many patients are still in need of better tolerated, more effective, and/or more affordable treatment options. An implantable device developed by Maity et al. helps to fill this gap by providing a controllable way to adjust blood glucose concentration

without medications or insulin (see the Perspective by Bentley and Payne). The wirelessly powered device can be linked to a clinically available continuous glucose monitor to catalyze the conversion of glucose to gluconate, which is then renally excreted. This normalized blood glucose and simultaneously removed some of the excess caloric load when tested in mice.

—Yevgeniya Nusinovich

Science p. 400, 10.1126/science.ady8560; see also p.382, 10.1126/science.aem0762

---

CELL BIOLOGY Mitochondria dictate death mechanism

Control of programmed cell death is central to diverse processes from development to cancer therapies. A key event leading to apoptotic cell death is permeabilization of the mitochondrial outer membrane. This normally leads to release of cytochrome c and activation of caspases and the apoptosome. If this apoptotic death process is not completed, then mitochondrial DNA can leak out as well, causing caspase-independent cell death and inflammation. Yang et al. showed that release of cytochrome c and release of mitochondrial DNA are separate biochemical processes (see the Perspective by Muñoz-Pinedo and Martin). They are mediated, respectively, by complexes formed of the proteins BAK and BAX alone or by the incorporation of BAK and BAX in a larger complex with other proteins. The latter form of cell death, referred to as baciptosis by the authors, may be of particular importance in apoptosome-deficient human cancers. —L. Bryan Ray

Science p. 401, 10.1126/science.aea2181; see also p. 383, 10.1126/science.aem1247

---

COMPARATIVE COGNITION Fast thinking on the wing

Bird brains were once maligned, but it is increasingly clear that, although structurally different from mammals, they provide high-level analogous cognitive functions. In terms of being able to make rapid decisions based on visual inputs, birds are under some of the strongest selection pressure given their dependence on rapid movement during flight. Schmidbauer and Neider tested decision speed in carrion crows and found that it was roughly twice as fast as that in primates (see the Perspective by Colombo and Poutrain-Sorel). Furthermore, recordings from neurons in the nidopallium caudolaterale, the functional analog of the primate prefrontal cortex, exhibited ramping choice signals similar to those seen within this region during decision-making in primates.

—Sacha Vignieri

Science p. 445, 10.1126/science.aed6459; see also p. 385, 10.1126/science.aem0341

---

GLACIAL INCEPTION Anthropogenic reversal of glaciation

Has anthropogenic warming prevented the next glacial cycle? Young et al. showed that it may have done just that (see the Perspective by Marcott). They looked at the evolution of the Barnes Ice Cap on Baffin Island, Canada, the last remaining piece of the Laurentide Ice Sheet (LIS), which once covered much of North America and was the largest of the Pleistocene ice sheets. After shrinking throughout much of the early- and mid-Holocene, the LIS began to grow again around 2000 years ago but then began shrinking once more around 1960. The authors attribute

this reversal to anthropogenic greenhouse forcing, evidence that humans caused the termination of the next glaciation.

—Jesse Smith

Science p. 455, 10.1126/science.aed3379; see also p. 381, 10.1126/science.aem3295

PHYSICS

---

Stranded in a metallic sea

Bound states in the continuum (BIC), localized states with energies lying within the continuous spectrum of spatially extended states, appear in systems described by a wave equation. They have been observed largely in photonic and acoustic systems because visualizing electronic BIC in metals is trickier. Thinel et al. used scanning tunneling spectroscopy to measure the local density of states of $Pd_{5}AlI_{2}$ , a van der Waals metal with a decorated checkerboard lattice. On the sites identified as palladium defects, the authors observed sharp, isolated peaks at energies inside the metallic continuum but electronically decoupled from the continuum states. Calculations suggest that these electronic BIC states stem from flat bands associated with frustrated hopping, a mechanism that is generalizable to several types of crystal lattices. —Jelena Stajic

Science p. 429, 10.1126/science.adt9825

T CELLS

---

Sugarcoating the fate of γδ T cells

γδ T cells play an important role in maintaining tissue homeostasis but become dysregulated in inflammatory conditions. Santos-Pereira et al. revealed that the γδ T cell glycome plays an important role in γδ T17 cell-fate commitment, promoting a hyperactivated proinflammatory state that exacerbates immune-mediated inflammatory diseases. Analysis of human transcriptomic data and murine models revealed that the absence of β1,6 N-acetylglucosamine–branched N-glycans promoted type-3 interleukin-17A–producing γδ T cells and loss of tolerance. Loss of N-glycan branching in glycoengineered mouse models increased susceptibility to psoriasis, lupus, and inflammatory bowel disease. Supplementing γδ T cells with N-acetylglucosamine reduced disease pathology and γδ T17 cell activation, revealing the importance of glycobiology in γδ T cell behavior and strategies for disease intervention.

—Leoma Bere

Sci. Immunol. (2026)

10.1126/sciimmunol.aec9553

Intron retention regulates STAT2 function and associates with immunotherapy response in lung cancer

Full article and list of author affiliations: https://doi.org/10.1126/science.aea7156

Ryan P. Englander, Mattia Brugiolo, Maëva Devoucoux, Te-Chia Wu, Mitch Kostich, Nathan K. Leclair, Isha Walawalkar, Hyeon Gu Kang, SungHee Park, Florentina Marches, Mark E. Peeples, Phylip Chen, Jacques Banchereau, Peter Yu, Andrew Salner, Romain Banchereau, Karolina Palucka†, Olga Anczuków†\*

INTRODUCTION: Immune checkpoint blockade has transformed the treatment of lung cancer, yet most patients still fail to respond and the molecular determinants of therapy response remain incompletely understood. Alternative RNA splicing, the process by which cells generate multiple messenger RNA (mRNA) isoforms from a single gene by selectively including exons or introns, is frequently dysregulated in cancers, including lung cancer. However, its contribution to antitumor immunity and immunotherapy responses has not been well characterized.

RATIONALE: Splicing alterations in tumors often generate isoforms with altered or disrupted function relative to their canonical counterparts, potentially influencing tumor growth and therapeutic response, and yet a large fraction of these isoforms remains functionally uncharacterized. Long-read RNA sequencing, which captures full-length transcripts rather than short fragments, has substantially improved the ability to identify and functionally interpret these isoforms in health and disease. This approach has previously been applied to catalog isoforms in lung cancer cell lines, but not in primary human lung tumors.

RESULTS: We applied long-read RNA sequencing to primary human lung adenocarcinoma samples to map the alternative splicing landscape of the most common subtype of lung cancer. We identified more than 149,000 full-length mRNA isoforms, more than half of which were unannotated, with many occurring in immune-related genes. Isoforms with immune relevance were enriched for genes involved in interferon responses, a pathway central to both immune activation and suppression in tumors. Among these, we identified two retained introns in STAT2, which encodes a transcription factor required for type I interferon signaling; each of these introns was up-regulated in tumors and predicted to disrupt normal protein function. We functionally characterized both isoforms: Retention of either intron depleted full-length STAT2 protein, but retention of intron 12 specifically produced a truncated isoform that actively suppressed interferon signaling rather than merely failing to support it. This dominant-negative isoform is expressed in lung adenocarcinoma cell lines upon suppression of nonsense-mediated decay, a pathway often attenuated in tumors. Finally, we analyzed pretreatment RNA sequencing data from clinical trials in lung and urothelial cancer and found that the relative abundance of these two STAT2 intron-retention isoforms was associated with checkpoint blockade response, showing stronger retrospective association than existing biomarkers.

CONCLUSION: Alternative RNA splicing is a pervasive and functionally important layer of gene regulation, and yet its role in cancer biology remains incompletely defined. Here, we show that alternative splicing regulates the function of genes implicated in immune response in lung cancer, including in STAT2. Intron retention in STAT2 disrupts type I interferon signaling through

STAT2 intron retention disrupts interferon signaling and correlates with checkpoint blockade response. Long-read RNA sequencing of primary lung adenocarcinomas identified 149,673 full-length isoforms, more than half unannotated. Retained intron 6 depletes full-length STAT2, lowering interferon-stimulated gene (ISG) transcription, whereas retained intron 12 generates a dominant-negative isoform that actively inhibits signaling. The difference between intron 6 and intron 12 retention is associated with anti-programmed death-ligand 1 (PD-L1) response across patient cohorts. FL, full length; ISRE, interferon-stimulated response element; ri12, retained intron 12.

two distinct mechanisms: depletion of functional protein and generation of a dominant-negative isoform under conditions of nonsense-mediated decay suppression in tumors. The relative levels of these intron-retained isoforms represent a candidate biomarker associated with checkpoint blockade response, which warrants prospective validation. More broadly, the scale of unannotated isoform diversity uncovered by this approach suggests that long-read RNA sequencing of primary tumors may reveal additional functionally relevant isoforms of clinical utility beyond STAT2. These findings identify a previously unrecognized splicing-based mechanism of immunotherapy response, with potential implications for future patient stratification strategies pending further clinical evaluation.

CANCER

Intron retention regulates STAT2 function and associates with immunotherapy response in lung cancer

Ryan P. Englander $^{1,2}$ , Mattia Brugiolo $^{1}$ , Maëva Devoucoux $^{1}$ , Te-Chia Wu $^{1}$ , Mitch Kostich $^{1}$ , Nathan K. Leclair $^{1}$ , Isha Walawalkar $^{1}$ , Hyeon Gu Kang $^{1}$ , SungHee Park $^{1}$ , Florentina Marches $^{1}$ , Mark E. Peeples $^{3,4,5}$ , Phylip Chen $^{3}$ , Jacques Banchereau $^{1,6}$ , Peter Yu $^{7}$ , Andrew Salner $^{7}$ , Romain Banchereau $^{8}$ , Karolina Palucka $^{1,9\dagger}$ , Olga Anczuków $^{1,9,10\dagger*}$

Immunotherapy with checkpoint inhibitors has revolutionized cancer therapy and enabled considerable therapeutic advances for treating various diseases, including lung cancer. However, the molecular determinants of variable patient outcomes remain unclear. Using long-read RNA sequencing, we mapped the landscape of alternative RNA splicing in human primary lung adenocarcinomas. We identified more than 149,000 full-length messenger RNA (mRNA) isoforms, more than half of which were unannotated and many that occurred in immune-related genes, particularly within the type I interferon response pathway. We functionally characterized retained introns in the STAT2 gene that produced altered protein isoforms that regulated immune signaling and interferon responses and correlated with patient survival after anti-programmed death-ligand 1 (PD-L1) immunotherapy. These findings position alternative splicing as a regulator of tumor immune responses and reveal intron retention in STAT2 as an interferon regulatory axis that correlates with PD-L1 checkpoint inhibitor response.

Lung cancer is a leading cause of death globally, with lung adenocarcinoma (LUAD) being the most prevalent subtype $(1)$ . Checkpoint inhibitor immunotherapies such as anti-programmed cell death protein 1 (PD-1) antibodies have substantially improved overall survival of patients with LUAD $(2, 3)$ . These therapies promote antitumor immune responses by blocking immunosuppressive pathways that reduce the effector functions of endogenous tumor-reactive CD8 $^{+}$ T cells; however, many patients do not respond to this treatment, and existing biomarkers fail to predict responses adequately $(4)$ . A better understanding of the molecular pathways that regulate immune function in LUAD is vital for improving immunotherapy.

Alternative RNA splicing, a ubiquitous posttranscriptional process that generates multiple mRNAs from a single gene (5), may be key to a more precise understanding of immune responses in tumors. Although almost all multiexonic genes are alternatively spliced in humans, most isoforms remain uncharacterized (6, 7). Splicing is often altered in cancer to favor oncogenic and immune-evasive isoforms (8–11), but the mechanisms underlying these changes in LUAD are poorly understood.

Long-read RNA sequencing (RNA-seq) allows sequencing of full-length transcripts and thus enables deeper functional analyses than would be possible using conventional short-read RNA-seq technology $12-14$ . However, to date, long-read RNA-seq has only been applied to studies of alternative splicing in LUAD cell lines and not primary tumors $15-17$ . In primary tumors from other cancer types, long-read RNA-seq has revealed thousands of previously uncharacterized isoforms $18-21$ , highlighting its potential to uncover biologically and clinically relevant isoforms in LUAD.

Using long-read RNA-seq and a multilevel analytical platform, we mapped full-length mRNA isoforms in LUAD tumors. Many long-read-derived isoforms affected genes in immune pathways, notably interferon (IFN) signaling. IFNs, including type I (IFN-I) and type III (IFN-III), play dual roles in antitumor immunity and immune evasion $22–27$ . We identified and characterized two retained introns using long-read RNA-seq in the STAT2 gene, an essential transcription factor necessary for transducing IFN-I and IFN-III responses. These events impair IFN responsiveness, are tumor-enriched, are targetable by RNA-based methods, and retrospectively associate with patient survival after programmed death-ligand 1 (PD-L1) inhibition. Our findings underscore the utility of long-read RNA-seq to discover immune-regulatory mechanisms in cancer.

---

Long-read RNA-seq reveals thousands of unannotated alternatively spliced isoforms in human lung tumors

To interrogate the alternative splicing landscape of human lung cancer, we performed paired long-read and short-read RNA-seq on RNA from 12 LUAD tumors (table S1, A to C) using PacBio single-molecule real-time (SMRT) circular consensus sequencing and Illumina sequencing, respectively (Fig. 1A). Full-length isoforms were assembled using the isoform sequencing (Iso-seq) pipeline, aligned, refined, and merged into a unified transcriptome (28, 29). Unannotated splice junctions detected by long reads were validated using paired short-read RNA-seq from the same RNA sample. We further validated transcript 5' and 3' boundaries using orthogonal datasets, including existing reference transcriptome annotations and previously mapped polyadenylation and transcription start sites (30–34). We determined empirically optimized cutoffs by varying the stringency of these parameters systematically (fig. S1A and table S1, D to H). To assess which filtering parameters were most important for isoform classification, we trained gradient-boosted machine-learning classifiers on our dataset to assess the importance of each parameter for differentiating true isoforms from artifacts (fig. S1, D to I). High-quality isoforms were further filtered for ≥5 long reads in ≥1 sample or ≥3 long reads in ≥2 samples, and isoforms were annotated using SQANTI3 structural categories (35). After stringent filtering, we identified 149,673 full-length mRNA isoforms from 18,300 loci (Fig. 1B and fig. S1B), with 51% of isoforms unannotated compared with the reference transcriptome (Fig. 1B and table S2A). Isoform quantification across tumors showed consistent structural category proportions between patients (fig. S1C and table S2B). Notably, 56% of isoforms (n = 33,267) with unannotated splice junctions [novel-in-catalog (NIC) and novel-not-in-catalog (NNC) categories] were found in ≥50% of patients, suggesting that some unannotated long-read-derived isoforms are recurrent components of LUAD transcriptomes (Fig. 1C and table S2B). Additionally, isoforms across structural categories exhibited shorter 3' untranslated regions (UTRs) than the closest matching reference (fig. S1, J to K, and table S3A). This result is consistent with 3' UTR shortening previously described in cancer (36) but may also reflect a contribution from 3' truncated artifacts arising from degraded mRNA. Of the isoforms, 2% (2940) arose from intergenic regions, with 39% of these being multiexonic (fig. S1M), potentially representing unannotated or reexpressed genes.

We predicted isoform protein-coding sequences for long-read-derived isoforms and compared them with human reference proteins from UniProtKB/Swiss-Prot (37). Most reference-matching isoforms

A

B

C

D

E

F

G

H

|

J

Fig. 1. Long-read RNA-seq detects unannotated alternatively spliced isoforms in LUAD primary tumors that are enriched in genes associated with IFN response. (A) Pipeline for the discovery of full-length spliced mRNA isoforms in human primary LUAD tumors ( $n = 12$ patients). RNA from each tumor was sequenced with long-read (LR) RNA-seq to identify unannotated isoforms, which underwent extensive quality control and filtering, including short-read (SR) RNA-seq data from the same sample to validate unannotated spliced junctions, known polyA sites, and known transcription start sites. Isoforms with $\geq 5$ reads in $\geq 1$ sample or $\geq 3$ reads in $\geq 2$ samples were retained. See Materials and methods for filtering details. Full-length isoforms were annotated using SQANTI3 isoform structural categories: FSM, RM, ISM, NIC, and NNC. ORFs were predicted and annotated per isoform. (B) Long-read-derived full-length mRNA isoforms detected in LUAD were classified using SQANTI3 isoform structural categories. (C) Number of patient samples positive for each isoform as quantified by FLAIR, shown by SQANTI3 isoform structural category. (D) The impact of long-read-derived full-length mRNA isoforms on protein-coding potential was predicted by comparing the in silico translated mRNA sequence to the reference protein sequence from UniProt and is shown for "Known" and "Unannotated" isoforms. See Materials and methods for more details. (E) Pathway enrichment analysis of genes with more unannotated than known isoforms using immune-related Hallmark gene sets from MSigDB. See table S4A for gene names and isoform numbers (significance was assessed using a one-sided hypergeometric test; $^{*}P < 0.05$ , and ns is not significant). (F) The impact of long-read-derived mRNA isoforms on protein-coding potential was predicted for unannotated long-read-derived

isoforms affecting genes associated with IFN- $\alpha$ and IFN- $\gamma$ response pathways (n = 1079) versus all unannotated long-read-derived isoforms (n = 52,010) (significance was assessed using the hypergeometric test; $*P < 0.05$ , $P < 0.01$ , $*P < 0.001$ , and ns is not significant). (G) Intron retention events are quantified as PSI using short-read RNA-seq data from TCGA-LUAD tumor (n = 437 patients) and normal lung (n = 46 patients) samples. Intron retention events are filtered for those supported by long-read RNA-seq and skipped in the canonical reference isoform, ranked by FDR and the difference in median PSI between tumor and normal lung samples (median PSI is indicated by a bar, and FDR-corrected values are from rMATS; $***P < 0.001$ ). (H) Full-length spliced mRNA isoforms of STAT2 detected by long-read RNA-seq in LUAD tumor samples, grouped by intron retained (middle) and intron skipped (bottom) isoforms. The full gene structure is shown on top, with the retained introns 6 and 12 highlighted. SQANTI3 structural categories are listed next to each isoform. (I and J) Intron 6 (I) and 12 (J) retention was assessed by RT-PCR with primers located in flanking exons in LUAD tumors (n = 10 patients) used for long-read RNA-seq (left) and normal bronchial epithelial cells (NLE; n = 9 patients) from healthy donors (middle) and quantified as PSI (right; boxplots show median and interquartile range with whiskers extending to 1.5× interquartile; significance was assessed using Student's t test).

had exact protein matches [79% reference match (RM), 75% full-splice match (FSM)], whereas novel structural categories [incomplete-splice match (ISM), NIC, and NNC] frequently encoded divergent proteins (fig. S1L and table S3B). Thus, a substantial proportion of the protein-coding transcriptome in LUAD remains uncharacterized.

To determine how unannotated long-read-derived isoforms might contribute to the LUAD proteome, we classified the predicted open reading frames (ORFs) of isoforms from known genes as either known, unannotated, or truncated relative to the reference (table S3C). We further classified isoforms predicted to encode truncated ORFs into those with or without intron retention events, which often lead to ORF truncation by shifting the reading frame or introducing in-frame premature termination codons (PTCs) $(38)$ . Introns in long-read-derived isoforms were enriched in the terminal 10% of isoforms and were otherwise evenly distributed across gene bodies (fig. S1N). Notably, out of the 52,010 unannotated long-read-derived isoforms with ORF predictions, 35% were predicted to encode truncated proteins, 34% unannotated proteins, and 31% known ORFs (Fig. 1D and table S3C). These data show that unannotated long-read-derived isoforms that may encode uncharacterized and truncated proteins are frequently detected in LUAD.

---

Unannotated isoforms are enriched in IFN-related genes

Despite their ubiquity, most alternatively spliced isoforms are functionally uncharacterized $(39)$ . We chose to focus on the role of unannotated long-read-derived isoforms in immune-related pathways, which are critical for tumor immunity and the target of immunemodulating therapies $(40-42)$ . Intriguingly, pathway analysis of genes enriched in unannotated isoforms revealed significant enrichment in IFN- $\alpha$ and IFN- $\gamma$ responses (Fig. 1E). We found unannotated long-read-derived isoforms in 169 of 217 IFN-related genes, including STAT1, STAT2, and HLA-A (table S4B), with 27% $(n = 288)$ of unannotated long-read-derived isoforms in IFN-related genes predicted to encode truncated ORFs owing to intron retention, which is higher than the overall unannotated long-read-derived isoform average (Fig. 1F). Using short-read RNA-seq data from the same tumor samples $(n = 12)$ , we found that $\geq 90\%$ of unannotated long-read-derived isoforms were detected by short-read-based expression quantification of full-length isoforms, providing orthogonal cross-platform validation of these IFN-related unannotated isoforms across all coding categories (fig. S2A). Because intron retention can reduce protein expression by PTC introduction or inhibition of mRNA export, these results suggest that unannotated isoforms containing retained introns may disrupt IFN signaling in LUAD.

To further explore tumor-up-regulated retained introns in IFN-related genes in a larger patient cohort, we quantified splicing using short-read RNA-seq from tumor $n = 437$ and normal lung $n = 46$ samples from The Cancer Genome Altas Lung Adenocarcinoma (TCGA-LUAD) database (fig. S2B). We found 3038 differentially spliced events where the percent spliced-in (PSI) values significantly differed between tumor and normal samples $\left|\left|\Delta\mathrm{PSI}\right| > 5\%\right|$ , false discovery rate (FDR) < 0.05] across 1656 genes, including 69 events in 32 IFN-related genes (fig. S2, C and D, and table S4C). Differential splicing events included comparable overall proportions of intron retention across all genes and IFN-related genes (fig. S2, C and D), suggesting that intron retention is not specific to IFN genes in LUAD. Filtering for intron retention identified 10 long-read–validated events in seven IFN-related genes (table S4D). These intron retention events include (i) two events in STAT2, the essential IFN-I and IFN-III transcription factor (43, 44); (ii) three events in UBA7, an enzyme that covalently modifies other proteins with the polypeptide ISG15 to promote IFN responses and tumor immunity (45); and (iii) events in DHX58, OAS1, and MOV10, all of which encode antiviral effectors with complex roles in tumor immunity (46–48) (Fig. 1G and table S4, C and D). To further validate that these events are not being detected because of differential group sizes, we confirmed that all events except for retained intron 5 of OAS1 were robust to repeat subsampling of our LUAD samples (table S4E).

---

STAT2 isoforms with retained introns are expressed in tumor cells

We focused on unraveling the function and regulation of intron retention in STAT2, as retained introns 6 (ri6) and 12 (ri12) were among the most up-regulated IFN-related events in TCGA, and STAT2 is an essential, nonredundant mediator of IFN-I and IFN-III signaling (27, 43). Long-read RNA-seq identified 22 STAT2 isoforms, of which 15 were unannotated, including 12 retaining ri6, ri12, or both (Fig. 1H). All long-read-derived ri6 and/or ri12 isoforms in LUAD tumors were structurally distinct from reference transcriptome annotations (fig. S2E and table S4F).

As an orthogonal validation, we quantified STAT2 intron 6 and intron 12 retention by reverse transcription polymerase chain reaction (RT-PCR) with primers that amplify both intron-containing and fully spliced transcripts in the primary LUAD tumors analyzed by long-read RNA-seq. As a normal reference, we used primary bronchial epithelial cells from healthy donors cultured at the air-liquid interface, a system that recapitulates bronchoalveolar epithelia, including alveolar cells thought to give rise to LUAD (49–51) (Fig. 1, I and J). Consistent with the TCGA data, both introns were retained at significantly higher levels in tumors compared with normal bronchial epithelial cells (medianPSI $_{ri6}$ = 32% tumors versus 15% normal; medianPSI $_{ri12}$ = 25% tumors versus 15% normal) (Fig. 1, I and J). RNA fluorescence in situ hybridization (FISH) on LUAD tumor sections confirmed STAT2 ri6 and STAT2 ri12 signals in KRT8 $^{+}$ tumor cells (fig. S2, F and G), indicating that tumor cells themselves express retained-intron STAT2 isoforms. Consistently, both introns were detectable by RT-PCR in six LUAD cancer cell lines (fig. S2, H and I), although at lower levels (medianPSI $_{ri6}$ = 15%; medianPSI $_{ri12}$ = 12%) than in primary tumors, possibly reflecting differences in baseline intron retention levels between cells in culture and primary tumors. These data indicate that STAT2 ri6 and ri12 are up-regulated in tumors compared with normal lung and are expressed within tumor cells, while not excluding their expression in other cell types.

---

STAT2 intron retention regulates STAT2 protein expression and IFN signaling

To assess how STAT2 intron retention affects IFN signaling, we designed splice-switching antisense oligonucleotides (ASOs) targeting the splice sites of STAT2 ri6 (ri6-ASO1) or ri12 (ri12-ASO1 and ri12-ASO2) as molecular tools to promote intron retention (Fig. 2, A and B). In A549 lung cancer cells, these ASOs increased retention of either ri6, ri12, or both, with largely consistent effects at baseline and after IFN-I stimulation (Fig. 2, C to H). The ri6-targeting ASO increased intron 12 retention without a significant effect on intron 6 retention, leading to a >10-fold decrease in STAT2 protein expression (Fig. 2, C, E, and G). Both ri12-targeting

B

F

H

J

Fig. 2. ASOs targeting STAT2 retained introns permit dynamic modulation of STAT2 protein expression and IFN response. (A and B) Schematic representation of the position of ASOs targeting the STAT2 splice sites of introns 6 (A) and 12 (B). (C to F) Endogenous STAT2 intron 6 [(C) and (D)] or intron 12 [(E) and (F)] retention in A549 cells transfected with ASOs targeting intron 6 [(C) and (E)] or intron 12 [(D) and (F)], or nontargeting control (NTC), and treated with 100 U/ml IFN-I for 0, 6, or 24 hours before collection, was assessed by semiquantitative RT-PCR with primers that amplify both the skipped and retained-intron isoforms. Representative images along with PSI quantification are shown ( $n = 3$ independent biological replicates; data are mean ± SD; significance was assessed using Student's t test; \*P < 0.05, \*\*P < 0.01, \*\*\*P < 0.001, and ns is not significant; all tests are relative to NTC-ASO for the same IFN-I time point). (G and H) Endogenous STAT2 protein expression in A549 cells transfected with ASOs targeting intron 6 (G) or intron 12 (H), or nontargeting control (NTC), and treated with 100 U/ml IFN-I for 0, 6, or 24 hours before collection was quantified with Western blot with an anti-STAT2 antibody and GAPDH as loading control. Representative images along with quantification normalized to NTC-ASO for the same IFN-I time point are shown ( $n = 3$ independent biological replicates; data are mean ± SD; significance was assessed using Student's t test; \*\*P < 0.01, and \*\*\*P < 0.001). (I and J) Expression of ISGs in A549 cells transfected with ASOs targeting intron 6 (I) or intron 12 (J), or nontargeting control (NTC), and treated with 100 U/ml IFN-I for 6 or 24 hours was quantified by quantitative real-time PCR. Expression is normalized to NTC-ASO for the same IFN-I time point ( $n = 3$ independent biological replicates). FC, fold change.

ASOs increased intron 6 retention under all conditions and increased intron 12 retention with IFN-I, depleting full-length STAT2 by up to \~100-fold (Fig. 2, D, F, and H). Extended RT-PCR with primers flanking introns 6 and 12 confirmed that these effects reflect bona fide intron retention changes rather than exon skipping (fig. S3, A to E) and additionally revealed an isoform retaining both introns 12 and 13, which was also detected in LUAD primary tumors by long reads (Fig. 1H and fig. S3, D and E). ASO effects on intron retention and protein expression were largely consistent across three additional LUAD cell lines (fig. S4, A to C). Computational mapping of known RNA-binding protein (RBP) binding motifs from in vitro and in cellulose binding studies (52) revealed overlaps between ASO target sequences and predicted binding sites for heterogeneous nuclear ribonucleoprotein (hnRNP), serine/arginine-rich (SR), and RNA-binding motif (RBM) protein families (fig. S3, F and G), suggesting that ASO-mediated steric inhibition of regulatory RBP binding contributes to the increase in intron retention. We note, however, that the RBP binding patterns observed at ASO-targeted splice sites are not fully recapitulated by existing computational prediction models, underscoring the known limitations of such tools in capturing the complexity of cotranscriptional RBP dynamics. Finally, we investigated how ASO treatment affects IFN-I signaling using the expression of known IFN-stimulated genes (ISGs) MX1 and RSAD2 as a readout of IFN-I activity. All ASOs promoting intron retention suppressed ISG induction after IFN-I stimulation (Fig. 2, I and J). Together, these results demonstrate that STAT2 splice-switching ASOs can increase intron retention, thereby reducing full-length STAT2 expression and IFN-I activity.

---

STAT2 retained-intron isoforms can produce truncated proteins when nonsense-mediated decay is inefficient

Although intron retention frequently introduces PTCs that trigger nonsense-mediated decay (NMD), some PTC-containing mRNAs evade degradation and generate truncated proteins $53, 54$ . Because NMD is often less efficient in tumors than in normal tissues, including in LUAD $55–57$ , we reasoned that tumors with reduced NMD efficiency may express higher levels of truncated STAT2-ri6 or STAT2-ri12 proteins.

To test this, we first scored NMD efficiency in 437 TCGA-LUAD samples by quantifying the inclusion of 12 well-established and experimentally validated NMD-sensitive exons (58–60), where a lower score indicates lower NMD efficiency (fig. S5A). This approach was validated against published RNA-seq data from short interfering RNA-mediated silencing of the essential NMD factor UPF1 (61), in which NMD-deficient cells showed significantly reduced NMD scores (fig. S5B). Tumors with lower NMD scores displayed higher STAT2 ri6 and ri12 retention (fig. S5, C and D, and table S5F), suggesting that intertumoral variation in NMD efficiency contributes to differential STAT2 intron retention. Notably, tumors had lower NMD scores than A549 cells, suggesting higher NMD efficiency in cell lines than in LUAD tumors in vivo. This difference was not attributable to tumor purity, as NMD score did not correlate with purity in TCGA-LUAD (62) (fig. S5, E and F).

To directly test whether truncated STAT2 proteins are expressed endogenously, we inserted a 3xFLAG tag into the 5' region of the endogenous STAT2 locus in A549 and H1975 cells by CRISPR (Fig. 3A), which we confirmed by detection of full-length FLAG-STAT2 with both FLAG and C-terminal STAT2 antibodies (fig. S5G). In both cell lines, endogenous truncated FLAG-STAT2 proteins were detected at molecular weights matching those of cDNA-expressed STAT2-ri6 (\~30 kDa) and STAT2-ri12 (\~50 kDa) controls (Fig. 3, B and C). Treatment with SMGIi, a selective inhibitor of the NMD-essential kinase SMG1 (63), increased STAT2-ri12 protein \~2- to 3-fold alongside elevated ri12- and ri6-containing transcripts (Fig. 3, B and C, and fig. S5, H and I), confirming that NMD actively limits truncated STAT2 protein accumulation.

Finally, we examined truncated-protein isoform expression upon forced intron retention using STAT2 splice-switching ASOs under conditions of NMD inhibition. We treated endogenously tagged cells with the intron retention—boosting ri6-ASO1 or ri12-ASO1 combined with

B

C

E

D

Fig. 3. STAT2 retained intron 6– and intron 12–containing isoforms can produce truncated proteins regulated by NMD. (A) Strategy for CRISPR-mediated FLAG epitope tagging of endogenous STAT2. Cells were nucleofected with STAT2 start codon–targeting sgRNA, Cas9 ribonucleoproteins, and donor oligonucleotide encoding the FLAG tag. Single-cell clones were expanded and genotyped, and FLAG-STAT2 protein was characterized by Western blot. (B and C) FLAG-tagged STAT2 protein isoforms in endogenous FLAG-STAT2 A549 (B) or H1975 (C) cells treated with SMG1i (500 ng/ml for 24 hours) or DMSO as negative control were detected by Western blot with anti-FLAG antibody. Transiently transfected FLAG-STAT2-cDNA with ri6 or ri12 were used as molecular weight controls. Representative images are shown along with quantification of STAT2-ri6 and STAT2-ri12 with GAPDH as loading control, normalized to DMSO for each clone (n = 3 independent biological replicates; data are mean ± SD; significance was assessed using Student's t test; \*P < 0.05 and, ns is not significant). (D and E) FLAG-tagged STAT2 protein isoforms in endogenously tagged FLAG-STAT2 A549 (D) and H1975 (E) clones transfected with STAT2 intron 6– or intron 12–targeting ASOs (ri6-ASO1 and ri12-ASO1) or nontargeting control (NTC) and treated with SMG1i (500 ng/ml for 24 hours) or DMSO as negative control were detected by Western blot with anti-FLAG antibody. Representative images are shown along with quantification of STAT2-ri6 and STAT2-ri12 with GAPDH as loading control, normalized to DMSO for each ASO (n = 3 independent biological replicates; data are mean ± SD; significance was assessed using Student's t test; \*P < 0.05, \*\*P < 0.01, and ns is not significant; all tests are relative to DMSO for the same ASO transfection).

NMD inhibition. ri6-ASO1 with SMG1i increased FLAG-STAT2-ri6 protein levels by \~3- to 5-fold, consistent with its increased ri6 retention at the RNA level (PSI $_{ri6}$ = 50 to 80% versus \~20% in control) (Fig. 3, D and E, and fig. S5, J and K). ri12-ASO1 with SMG1i increased FLAG-STAT2-ri12 protein levels by \~5- to 6-fold and FLAG-STAT2-ri6 by \~3-fold (Fig. 3, D and E). Cell line-specific differences in ri6 retention upon ri12-ASO1 treatment and NMD inhibition likely reflect competition among multiple NMD-sensitive STAT2 isoforms, with STAT2-ri12–encoding isoforms increasing preferentially under these conditions. Notably, whereas full-length STAT2 (STAT2-FL) remained the most abundant isoform across conditions, combined ri12-ASO1 and SMG1i brought STAT2-ri6 and STAT2-ri12 to approximately equal abundance as STAT2-FL in A549 cells (fig. S5L), demonstrating that high intron retention alongside inefficient NMD can substantially shift the balance of STAT2 protein isoforms. Taken together, these results demonstrate that truncated STAT2-ri6 and STAT2-ri12 proteins are produced from the endogenous locus, are actively suppressed by NMD, and accumulate when NMD is inhibited or intron retention is elevated.

---

STAT2-ri12 inhibits the IFN-I response

Having established that STAT2 intron retention produces truncated STAT2 proteins, we investigated their functional consequences. AlphaFold 3 structural predictions revealed that the STAT2-ri6 coiled-coil domain (CCD) is disrupted by a PTC that replaces the native C-terminal CCD sequence with 20 amino acids encoded by intron 6, predicted to extend outward as a naked $\alpha$ helix that may impair protein-protein interactions (Fig. 4A and fig. S6A). By contrast, STAT2-ri12 undergoes a clean truncation at glycine-372 with no distinctive amino acids, leaving its N-terminal domain (NTD) and CCD structurally intact and potentially preserving canonical interaction surfaces (Fig. 4A and fig. S6A). Both truncated isoforms lack the Src homology 2 (SH2) and transcriptional activation (TAD) domains required for transcriptional up-regulation of STAT2 targets in response to IFN (Fig. 4A). To confirm that neither isoform can transduce IFN responses, we reexpressed FLAG-tagged STAT2-FL, STAT2-ri6, STAT2-ri12, or empty vector (EV) in two independent $\mathrm{STAT2}^{\mathrm{KO}}$ A549 clones in which endogenous STAT2 has been inactivated with CRISPR (fig. S6, B to F). Only STAT2-FL rescued IFN-I-stimulated MX1 and RSAD2 expression, confirming that neither truncated isoform can transduce IFN-I responses (fig. S6G).

We next asked whether truncated isoforms affect IFN-I signaling in cells also expressing full-length STAT2 using doxycycline (DOX)-inducible A549 lines expressing FLAG-tagged STAT2-ri6, STAT2-ri12, or STAT2-FL cDNAs that are truncated after their respective stop codons (Fig. 4, B and C). Upon IFN-I stimulation, STAT2-ri12 expression reduced ISG expression at all IFN-I doses tested, whereas STAT2-ri6 had no effect and STAT2-FL increased ISG expression as expected (Fig. 4D). To define the mechanism, we performed coimmunoprecipitation assays assessing interactions with IRF9 and STAT1, the partners that together with STAT2 form the ISGF3

transcription factor–mediating IFN-I responses (64). STAT2-ri12 retained IRF9 binding but failed to associate with STAT1, consistent with loss of the SH2 domain required for STAT1:STAT2 interaction (64). STAT2-ri6 interacted with neither partner, consistent with its disrupted CCD, which is required for IRF9 binding (65). STAT2-FL interacted with both partners as expected (Fig. 4, A and E). Both truncated isoforms localized predominantly to the cytoplasm regardless of IFN-I stimulation (fig. S7A), suggesting that these interactions occur in the cytosol. Because IRF9 binding and nuclear translocation are required for productive STAT2 signaling (66), STAT2-ri12 likely acts as a dominant-negative inhibitor by sequestering IRF9 and preventing ISGF3 assembly.

C

B

D

E

Fig. 4. STAT2 retained intron 6– and intron 12–containing isoforms encode proteins that fail to respond to IFN-I and differ in their inhibitory capacity. (A) Protein structures of STAT2-FL, STAT2-ri6, and STAT2-ri12 protein isoforms predicted by AlphaFold 3. Amino acid residues are colored by protein domain. DBD, DNA binding domain. (B) Strategy for generating DOX-inducible truncated STAT2 protein isoform–expressing A549 lines. Parental A549 cells expressing wild-type STAT2 were infected with lentivirus constructs encoding STAT2-FL, STAT2-ri6.TR, or STAT2-ri12.TR. DOX binding activates the reverse tetracycline-controlled transactivator (rtTA) leading to activation of the tetracycline-responsive element (TRE) synthetic promoter that drives STAT2 isoform expression. iSTAT2-ri6.TR and iSTAT2-ri12.TR are truncated after their respective PTC. (C) Expression of FLAG-STAT2 protein isoforms in A549 cells harboring DOX-inducible STAT2-FL, STAT2-ri6.TR, or STAT2-ri12.TR treated with DOX or DMSO for 48 hours was assessed by Western blot with anti-FLAG antibody, with GAPDH as loading control. Representative images are shown, and the expected positions of STAT2-FL, STAT2-ri6, and STAT2-ri12 are indicated. FL, full length. (D) Expression of ISGs at 6 hours after IFN-I stimulation was assessed by quantitative real-time PCR in A549 cells expressing DOX-inducible empty vector, STAT2-FL, STAT2-ri6.TR, or STAT2-ri12.TR treated with DOX or DMSO for 48 hours, expressed as fold change to the respective DMSO control (n = 3 independent biological replicates; data are mean ± SD; significance was assessed using Student's t test; \*P < 0.05, \*\*P < 0.01, \*\*\*P < 0.001, and ns is not significant). (E) Interaction between STAT2 and STAT1 or IRF9 is assessed in A549 cells harboring DOX-inducible FLAG-tagged STAT2-FL, STAT2-ri12, or STAT2-ri6 treated with DOX for 48 hours to induce expression, followed by 1000 U/ml IFN-I stimulation for 1.5 hours before lysis. FLAG-tagged proteins were immunoprecipitated (IP) using anti-FLAG antibody, eluted with FLAG peptide, and analyzed by Western blot with the indicated antibodies.

These data reveal a functionally important asymmetry between the two isoforms. STAT2-ri12 actively suppresses IFN-I signaling, whereas STAT2-ri6 is functionally inert, neither transducing nor inhibiting the response. This asymmetry has an important implication: Because ri6 lies upstream of ri12, isoforms retaining both introns will encode the STAT2-ri6 protein rather than STAT2-ri12, as the ri6 PTC terminates translation before the ri12-encoded sequence is reached.

---

STAT2 retained introns associate with response to PD-L1 inhibition

IFN signaling plays a complex role in cancer immunotherapy: Although essential for leukocyte activation, differentiation, and tumor rejection (27, 67), chronic IFN signaling can drive immunosuppression and contribute to treatment failure (23, 25, 26). To assess how STAT2 intron retention interacts with this dual function, we examined STAT2 intron retention in nonsquamous non-small cell lung cancer patients from the OAK clinical trial, which compared the PD-L1 inhibitor atezolizumab to docetaxel chemotherapy (2, 68). We mapped OAK short-read RNA-seq data to our long-read-enriched LUAD transcriptome and quantified splicing events using rMATS (Fig. 5A and table S5A). We performed multivariate Cox proportional hazards analysis on atezolizumab-treated patients with complete metadata to assess the association of overall survival with STAT2 ri6 and ri12 retention, overall STAT2 expression, intratumoral PD-L1 expression by immunohistochemistry (IHC), and tumor mutational burden (TMB). Cox proportional hazards models estimate the independent effect on survival of each variable measured while controlling for the other variables included in the model. For continuous variables, the hazard ratio (HR) indicates the independent effect of increasing each variable by one increment on the risk of death. For discrete variables, the HR indicates the independent effect of that variable on the risk of death compared with baseline. STAT2 ri12 retention predicted improved survival in atezolizumab-treated patients (HR for death: 0.5, P = 0.005), whereas STAT2 ri6 predicted worsened survival (HR for death: 2.02, P = 0.002), independently of TMB, PD-L1 IHC, and STAT2 expression (fig. S8A). High TMB predicted improved survival (HR for death: 0.47, P = 0.035), which was expected, as TMB is a US Food and Drug Administration (FDA)-approved immunotherapy biomarker. By contrast, PD-L1 staining did not predict overall survival (P = 0.811), consistent with published OAK data (2) (fig. S8A). These associations persisted after adjusting for age, sex, smoking, and transcriptomic markers of IFN signaling IFNG.GS and ISG.RS (69) (fig. S8B).

The opposing effects of ri6 and ri12 prompted us to examine their difference as a composite biomarker. We found that $PSI_{ri12-ri6}$ —the difference between ri12 and ri6 retention—strongly predicted improved survival after PD-L1 blockade (HR for death: 0.48, P = 0.0008) (fig. S8C), an effect that was absent in docetaxel-treated patients, indicating an immune-specific association. We dichotomized patients at a cutoff optimized in biomarker-negative patients ( $PSI_{ri12-ri6} \geq 5\%$ ; see Materials and methods) and found that High- $PSI_{ri12-ri6}$ patients had markedly improved survival compared with Low- $PSI_{ri12-ri6}$ patients, with a stronger effect than High-TMB (Fig. 5, B and C). Notably, only High- $PSI_{ri12-ri6}$ patients derived a survival benefit from atezolizumab over docetaxel (HR for death: 0.49, P = 0.001), with no such benefit observed in Low- $PSI_{ri12-ri6}$ patients (Fig. 5D). Patients with High- $PSI_{ri12-ri6}$ had minimal overlap with High-TMB and PD-L1 positivity (Fig. 5E and fig. S8E). Incorporating $PSI_{ri12-ri6}$ alongside TMB nearly doubled the number of biomarker-positive responders from 9 to 16 (Fig. 5F), positioning $PSI_{ri12-ri6}$ as a possible independent and complementary immunotherapy biomarker pending prospective confirmation.

In total, 304 genes were differentially expressed between High-PSI $_{ri12-ri6}$ and Low-PSI $_{ri12-ri6}$ patients in OAK (table S5B). Gene set enrichment analysis revealed depletion of IFN- $\alpha$ and IFN- $\gamma$ signaling in High-PSI $_{ri12-ri6}$ tumors (Fig. 5G, fig. S8, G and I, and table S5C), suggesting that High-PSI $_{ri12-ri6}$ tumors may have reduced chronic interferon signaling that facilitates responses to PD-L1 inhibition. This result was reproduced in TCGA (fig. S8, H, J, and K and table S5, D and E), showing that the inverse association between PSI $_{ri12-ri6}$ and IFN signaling is reproducible across cohorts. Consistent with a possible role for STAT2-ri12 protein in suppressing IFN-I, High-PSI $_{ri12-ri6}$ patients had a lower NMD score than Low-PSI $_{ri12-ri6}$ patients in both OAK and TCGA (Fig. 5H, fig. S8F, and table S5, F and G), which suggests that their tumors may more efficiently produce STAT2-ri12 protein. Notably, NMD score did not independently predict survival (fig. S8D), indicating that $PSI_{ri12-ri6}$ captures biology beyond NMD efficiency.

A

B

C

D

E

F

G

H

Fig. 5. STAT2 retained introns associate with response to atezolizumab in a phase 3 clinical trial. (A) RNA-seq data from pretreatment nonsquamous non-small cell lung cancer tumors ( $n = 503$ patients) from the OAK open label phase 3 clinical trial were used to quantify alternative splicing events using rMATS, focusing on STAT2 retained- intron events and their relationship to overall survival in patients treated with PD-L1 inhibitor atezolizumab or chemotherapy docetaxel. CA, cassette exon; IR, intron retention; A5SS, alternative $5'$ splice site; A3SS, alternative $3'$ splice site; MXE, mutually exclusive exon. (B and C) Kaplan-Meier plots of overall survival against patients with High-PSI $_{ri12-ri6}$ ( $\geq 5\%$ ; $n = 43$ patients) versus Low-PSI $_{ri12-ri6}$ ( $< 5\%$ ; $n = 212$ patients) (B) or High-TMB ( $\geq 16$ mutations/Mb; $n = 37$ patients) versus Low-TMB ( $< 16$ mutations/Mb; $n = 113$ patients) (C) in OAK patients from the atezolizumab treatment arm. Significance was assessed using the log-rank test. An optimal dichotomized cutoff for PSI $_{ri12-ri6}$ was derived using the function surv\_cutpoint from the R package survminer (see Materials and methods). (D) Cox proportional hazards analysis of overall survival among nonsquamous non-small cell lung cancer patients from OAK in the atezolizumab treatment arm who are High-PSI $_{ri12-ri6}$ ( $n = 43$ patients) or Low-PSI $_{ri12-ri6}$ ( $n = 212$ patients) compared with patients treated with docetaxel ( $n = 248$ patients). Data indicate HR $\pm 95\%$ confidence interval (CI), reflecting the relative rate of death in each group compared with docetaxel-treated patients. Significance was assessed using the Wald test. (E) Overlap of High-PSI $_{ri12-ri6}$ ( $\geq 5\%$ ) and High-TMB ( $\geq 16$ mutations/Mb) patients in the atezolizumab treatment arm of OAK. (F) Atezolizumab responders who achieved a complete or partial response per

Finally, to assess cross-tumor applicability, we computed STAT2 intron retention and $PSI_{ri12-ri6}$ from pretreatment short-read RNA-seq data from IMvigor210, a single-arm phase 2 trial of atezolizumab in metastatic urothelial carcinoma (70). $PSI_{ri12-ri6}$ again correlated with improved survival after atezolizumab treatment, both as a singular variable and as a covariate, with High- $PSI_{ri12-ri6}$ patients showing significantly better outcomes at an independently optimized cutoff (fig. S8, L and M, and table S5, H and I).

---

Discussion

Here, we used long-read RNA-seq, which is capable of resolving full-length isoform structures that short-read sequencing cannot distinguish, to comprehensively profile alternatively spliced isoforms in LUAD, revealing more than 149,000 mRNA isoforms, including in genes implicated in IFN responses. Focusing on STAT2, we show that retained introns 6 and 12 impair full-length STAT2 expression, can encode dominant-negative truncated proteins, and may predict patient responses to immunotherapy. These findings may open new avenues for therapeutic targeting of RNA splicing in LUAD.

Existing immunotherapy biomarkers are imperfect (4, 71, 72), and there is a pressing need for more and better tools to guide treatment decisions. We show that patients with higher STAT2 ri12 inclusion relative to ri6 (High-PSI $_{ri12-ri6}$ ) have significantly improved survival after atezolizumab treatment, an association observed in two independent clinical cohorts across distinct tumor types. Although splicing signatures have prognostic value in LUAD (73), to our knowledge, this is the first study to link specific alternative splicing events directly to immunotherapy outcomes. PSI $_{ri12-ri6}$ identifies a patient subset with minimal overlap with TMB or PD-L1 IHC, and incorporating it alongside TMB nearly doubles the number of biomarker-positive responders. This positions PSI $_{ri12-ri6}$ as a potentially clinically actionable transcriptomic biomarker capable of capturing immunotherapy benefits that existing tools miss, warranting prospective validation in future studies. The mechanistic basis for this association may relate to the ability of STAT2-ri12, but not STAT2-ri6, to act as a dominant-negative inhibitor of IFN-I signaling: Given that chronic IFN-I impairs immunotherapy efficacy (25, 26, 69), tumors expressing STAT2-ri12 may be better primed to respond to anti-PD-L1 immunotherapy. Because ri6 is upstream of ri12, ri6 retention can block STAT2-ri12 expression, potentially explaining the association between ri6 and poorer survival. Alternatively, High-PSI $_{ri12-ri6}$ tumors may harbor reduced NMD efficiency, leading to higher levels of PTC-encoding neoantigens that enhance immune recognition. Additional research to investigate whether STAT2 intron retention plays a direct mechanistic role in driving antitumor immune responses in vivo represents a clear next step for future research.

If STAT2 intron retention directly influences therapy response, manipulating it may represent an innovative therapeutic strategy. Using an FDA-approved ASO chemistry, we demonstrate that STAT2 splicing

trial criteria stratified by biomarker positivity. (G) Significantly enriched pathways in High-PSI $_{ri12-ri6}$ or Low-PSI $_{ri12-ri6}$ patients from OAK (FDR < 0.05) using MSigDB Hallmark 2020 Gene Sets (94, 95). Genes were ranked by their Wald test statistic from differential gene expression analysis comparing High-PSI $_{ri12-ri6}$ to Low-PSI $_{ri12-ri6}$ patients with DESeq2, and gene set enrichment analysis was conducted on the ranked gene list using gseapy. A positive normalized enrichment score (NES) indicates pathway enrichment in High-PSI $_{ri12-ri6}$ patients, whereas a negative NES indicates pathway enrichment in Low-PSI $_{ri12-ri6}$ patients. Immune-related pathways are highlighted in purple. (H) NMD score in OAK patients who are High-PSI $_{ri12-ri6}$ ( $\geq 5\%$ ; n = 70 patients) versus Low-PSI $_{ri12-ri6}$ (<5%; n = 433 patients) (boxplots show median and interquartile range, with whiskers extending to 1.5× interquartile; significance was assessed using Student's t test). A lower NMD score reflects less efficient NMD (see Materials and methods).

can be pharmacologically modulated to regulate STAT2 protein expression and IFN-I signaling strength. ASOs targeting STAT2 splicing could enable clinicians to precisely inhibit STAT2 activity as well as IFN-I and IFN-III signaling in cancer as necessary, providing a more refined degree of control over inflammation than is presently possible. Additional studies to assess the impact of ASO-mediated STAT2 splicing manipulation alongside immunotherapy in vivo are therefore warranted.

Though differences in STAT2 ri6 and ri12 retention between tumor and normal tissue appear modest, small splicing changes can have outsized functional consequences: NMD can mask increases in the production of NMD-sensitive isoforms, as described for other targets (74, 75), and simultaneous shifts across multiple isoforms produce nonlinear changes in isoform ratios that are often amplified by complex interactions between transcription, mRNA stability, translation efficiency, and protein turnover (76). Notably, LUAD tumors with reduced NMD efficiency display substantially higher STAT2 intron retention, and pharmacological NMD inhibition elevates truncated STAT2 protein 2- to 3-fold in cell lines, which, taken together, suggests that tumors with low NMD activity plausibly express truncated STAT2 proteins at functionally relevant levels. Key limitations of our study include the inability to directly measure STAT2 protein isoforms in primary tumors owing to the absence of isoform-specific antibodies and the fact that isoforms identified by long-read RNA-seq are not necessarily cancer specific or cancer enriched, as they have not been compared with long-read data from normal lung tissue.

Although RNA splicing has been studied for decades, the full diversity of spliced isoforms, especially in immune-related genes, has only recently begun to be fully appreciated $12, 13, 15, 18, 21, 77$ . The extensive catalog of previously unannotated isoforms uncovered here in well-studied cancer genes underscores how much of the tumor transcriptome, and its functional consequences, remains invisible to short-read approaches alone. Some of these isoforms, including those of STAT2, may be expressed across multiple cell types within the tumor microenvironment, and single-cell and spatial long-read sequencing will be essential to resolve their cell-type specificity in future studies.

---

Materials and methods

---

Clinical samples

Studies involving human tumor samples were conducted according to the protocol approved by the Institutional Review Board of The Jackson Laboratory for Genomic Medicine (IRB number 2019-RELY-002-HHC and 2019-RELY-002). Postsurgical tumor samples from patients with invasive LUAD were collected with appropriate informed consent at Hartford Hospital. Human bronchial epithelial cells and derived cultures used in this study were obtained from deceased donors through the National Disease Research Interchange (NDRI) and from deidentified deceased donors procured through (Ohio) donor recovery programs. The research was reviewed by The Jackson Laboratory Institutional Review Board (2019-NHSR-048; 2022-NHSR-004) and determined to qualify as non-human subjects research because all specimens were obtained from deceased donors and provided without direct identifiers. All research activities were conducted in accordance with institutional policies and ethical standards.

---

Human cell lines

Cell lines were obtained from the American Type Culture Collection (ATCC). A549 lines were cultured in Ham's F-12 Nutrient Mixture (ThermoFisher Scientific no. 11765-054) supplemented with $10\%$ fetal bovine serum (FBS) (GeminiBio no. 100-500), 25 mM HEPES (ThermoFisher no. 15630-080), 2 mM L-glutamine (ThermoFisher no. 25030-081), and 1% penicillin/streptomycin mixture (ThermoFisher no. 15140-122). NCI-H1373, NCI-H1573, NCI-H1693, NCI-H1755, and NCI-H1975 lines were cultured in RPMI-1640 medium (ThermoFisher no. 11875-093) supplemented with 10% FBS (Denville Scientific no. C788U23) and 1% diluted penicillin/streptomycin mixture (ThermoFisher no. 15140-122). All cell lines were grown at 37°C with 5% CO₂. For routine subculturing, cells were lifted by incubating cell cultures with 0.05% trypsin-EDTA (ThermoFisher no. 25300-054) at 37°C until cells were detached. Cells were routinely tested negative for mycoplasma using the MycoAlert Mycoplasma Detection Kit (Lonza no. LT07-318).

---

Human primary bronchial epithelial cells

To generate air-liquid-interface (ALI) cultures from lung epithelial progenitors, primary lung epithelial cells were plated on inserts with a pore size of 0.4 $\mu$ m (Corning no. 3470) coated with 0.03 mg/ml Collagen I Rat Tail Protein (Gibco no. A1048301). After seeding, wells were maintained for 8 to 10 days in Pneumacult-Ex Plus Medium (STEMCELL no. 05040) supplemented with 10 $\mu$ M ROCK inhibitor Y-27632 (Abmole no. Y-27632) until cells reached 100% confluence, with media changes every 2 to 3 days. Cells were then differentiated for 5 to 7 days in Pneumacult-ALI Medium (STEMCELL no. 05001) supplemented with 10 $\mu$ M ROCK inhibitor Y-27632 (Abmole no. Y-27632) with media changes every 2 to 3 days, alongside measurement of Trans-Epithelial/Endothelial Electrical Resistance (TEER) every 2 to 3 days until cells were fully differentiated, as described previously (78, 79). Finally, cells were differentiated in Pneumacult-ALI Medium (STEMCELL no. 05001) without ROCK inhibitor for \~28 days with media changes on the basal side of the insert every 2 to 3 days. Once TEER values reached >500 ohms/cm $^{2}$ , media was removed from the apical layer and checked under the microscope every 2 to 3 days for cilia beating and mucus production. After differentiation, the air-liquid interface membrane was harvested and lysed directly in RLT for subsequent RNA extraction.

---

Plasmids

All plasmids were sequence-confirmed using whole-plasmid Nanopore sequencing or standard Sanger sequencing (Quintara Biosciences). All primer sequences used for cloning are listed in table S6H.

pVAX1 transient transfection plasmids: Full-length STAT2 isoform sequences were derived via fusion PCR from A549 cDNA using Phusion Hot Start II DNA Polymerase (ThermoFisher Scientific no. F-549L). All constructs were obtained by sequential fusion PCRs of partial elements of STAT2 obtained from A549 cDNA or gDNA, finally NheI and EcoRI sites were added. All cloning oligos used are listed in table S6H. The resulting amplicons were digested using NheI-HF (New England Biolabs no. R3131) and EcoRI-HF (NEB no. R3101) and ligated into the linearized pVAX1 (Addgene no. 196624) backbone using T4 DNA Ligase (NEB no. M0202). Empty vector pVAX1 (pVAX1-EV) was generated by restriction digest using NheI-HF (NEB no. R3131) and XbaI (NEB no. R0145) and ligation of the backbone to itself using T4 DNA Ligase (NEB no. M0202).

pLJM1 lentiviral insert plasmids: Third-generation lentivirus encoding 3xFLAG-tagged STAT2 isoforms were generated by subcloning each isoform from pVAX1 into pLJM1-Polylinker (Addgene no. 164143) using T7\_F and BGH\_R to amplify the insert followed by restriction digest by NheI-HF and EcoRI-HF.

pINDUCER20 inducible insert plasmids: All inserts were amplified from pVAX1 using primers AgeI\_T7 and PacI\_BGH and subcloned into pCR8-ccdb-u1 (gift from A. Cheng) using AgeI-HF (NEB no. R3552) and PacI (NEB no. R0547) sites. Inserts were transferred into pINDUCER20 (Addgene no. 44012) using Gateway cloning (Invitrogen no. 11791020).

---

Generation of STAT2 knockout cell lines

STAT2-targeting gRNA was designed using the CRISPR Design Tool (https://design.synthego.com/#/, Synthego) using the Ensembl GRCh38. p10 human genome build targeting STAT2 (table S6D). Single guide RNA (sgRNA) with standard chemical modifications (2'-O-methylated nucleotides and phosphorothiorate linkages in the first and last three bases) targeting the relevant sequences were synthesized (Synthego). Then, 16 $\mu$ g of sgRNA was incubated with 20 $\mu$ g of Alt-R HiFi Cas9 Nuclease V3 (IDT no. 1081060) for 30 min at room temperature. A549 cells were harvested and nucleofected with the sgRNA:Cas9 ribonucleoprotein mixture using the 4D-Nucleofector X Unit (Lonza) with nucleofection solutions and programs as listed in table S6E. Cells were incubated in complete growth media supplemented with 5 $\mu$ M Alt-R HDR Enhancer V2 (IDT no. 10007921). Twenty-four hours after nucleofection, media was removed and replaced with complete growth media without HDR Enhancer. Nucleofected cells were harvested and serially diluted to a concentration of 5 cells/ml and plated into each well of a 96-well plate. Single cell clones were expanded and gDNA was extracted using the QIAamp DNA Mini Kit (Qiagen no. 51306). The edited region was amplified using primers listed in table S6F, and PCR products were confirmed by Sanger sequencing. Chromatogram files from single cell clones were deconvoluted into constituent alleles with DECODR (https://decodr.org/) using parental A549 cells as control. STAT2 knockout (deletion) was confirmed via Western blot and IFN stimulation followed by RT-qPCR for known STAT2-dependent ISGs MX1 and RSAD2 (80).

---

CRISPR-mediated endogenous STAT2 FLAG epitope tagging

A sequence encoding a 3xFLAG epitope tag was inserted immediately downstream of the STAT2 start codon in A549 and H1975 cells by CRISPR-Cas9-mediated homology-directed repair (HDR). An sgRNA targeting the STAT2 start codon region (table S6D) was synthesized and complexed with Cas9 as described above. Then, 200 pmol of a single-stranded HDR donor oligonucleotide encoding the 3xFLAG tag flanked by STAT2 homology arms (IDT; sequence listed in table S6G) was added to the nucleofection reaction. Cells were nucleofected with the sgRNA:Cas9 ribonucleoprotein and donor oligonucleotide using the 4D-Nucleofector X Unit with nucleofection solutions and programs as listed in table S6E. Immediately after nucleofection, cells were resuspended in complete growth media supplemented with Alt-R HDR Enhancer V2 (IDT no. 10007921) at 5 $\mu$ M (A549) or 3 $\mu$ M (H1975). After 24 hours, media was replaced with complete growth media without HDR Enhancer. After 3 days, cells were harvested and serially diluted to 5 cells/ml and plated into 96-well plates for single-cell cloning. The edited region was amplified using primers listed in table S6F, and PCR products were confirmed by Sanger sequencing. Successful in-frame FLAG tag insertion was confirmed by Western blot using an anti-FLAG antibody.

---

Generating stable cell line expressing STAT2 isoforms

Ten micrograms of lentiviral plasmids were transfected alongside 5 $\mu$ g psPAX2 (Addgene no. 12260) and 2.5 $\mu$ g pMD2.G (Addgene no. 12259) into HEK293T cells using Lipofectamine 3000 (Invitrogen no. L3000015) according to manufacturer instructions. Lentiviral supernatants were collected at 48 and 72 hours and concentrated as previously described (81). Concentrated lentivirus was then resuspended in Opti-MEM (Invitrogen no. 31985) and stored at $-80^{\circ}$ C until further use.

Cells were plated at $4 \times 10^{4}$ cells/cm $^{2}$ and transduced after 2 hours by supplementing media with polybrene (Sigma no. TR-1003-G) to a final concentration of 8 $\mu$ g/ml and adding the desired volume of concentrated lentivirus dropwise. Cells were selected in either 1 $\mu$ g/ml puromycin (Gibco no. A1113803) for lines generated with pLJM1-derived virus or 4 mg/ml G418 (Gibco no. 10131035) for lines generated with pINDUCER20-derived virus.

---

IFN stimulation

Cells were stimulated with Universal Type I Human Interferon (PBL Assay Science no. 11200) at indicated doses and durations using complete growth medium as the diluent.

---

ASO transfections

LUAD cell lines were reverse transfected in a 24-well plate with 2'-O-methoxyethyl-phosphorothioated ASO (IDT) at a final concentration of 100 nM using Lipofectamine RNAiMAX (Invitrogen no. 13778150) according to manufacturer instructions. Target ASO sequences are listed in table S6C. Seventy-two hours after transfection, cell lysate for RNA or Western blot was collected by direct lysis in the plate.

---

NMD inhibition

To inhibit NMD, cells were treated with complete growth media supplemented with 500 ng/ml SMG1i (Sigma-Aldrich no. SML3864) or dimethyl sulfoxide (DMSO) as negative control for 24 hours before collection for RNA or Western blot. For combined ASO transfection and NMD inhibition, cells were transfected with ASO as described above for 72 hours and treated with SMG1i for the last 24 hours before sample collection.

---

RNA extraction from tumors and cell lines

High-quality RNA was extracted from samples using the RNeasy Mini Kit (Qiagen no. 74106). For primary tumor samples, three 30- $\mu$ m tissue sections were cut using a Cryostat from each OCT-embedded tumor and placed directly into 350 $\mu$ l of RLT Lysis Buffer supplemented with 1% $\beta$ -mercaptoethanol. For cell lines, cells were lysed by direct addition of 350 $\mu$ l of RLT Lysis Buffer supplemented with 1% $\beta$ -mercaptoethanol to the wells directly. Samples were treated with DNase I (Qiagen no. 79256) for 1 hour at room temperature following manufacturer instructions. For RNA extracted from primary tumor samples, RNA quality was assessed with the RNA 6000 Pico Kit (Agilent no. 5067-1513) on a 2100 Bioanalyzer (Agilent). Only samples with an RNA integrity number above 7.0 were selected for sequencing.

---

cDNA synthesis and RT-PCR

RNA was extracted as described above, and 200 to 1000 ng of RNA was reverse-transcribed using SuperScript III Reverse Transcriptase (Invitrogen no. 18080093). Semiquantitative RT-PCR was used to amplify 20 ng of cDNA with optimized melting temperatures and cycle numbers using Q5 High-Fidelity Polymerase with High-GC Buffer (NEB no. M0491S). Primer sequences are listed in table S6A. PCR products were separated on a 2% agarose gel with SYBR Safe (Invitrogen no. S33102) and imaged on a ChemiDoc MP Imaging System (Bio-Rad). Band intensities were quantified with ImageLab 6.1.0 (Bio-Rad). The PSI was calculated as the intensity of the included band of the target event divided by the sum of the included and skipped bands of the target event. For each target event, bands of interest were extracted with a Gel Extraction Kit (Qiagen no. 28704) and Sanger sequenced (Quintara Biosciences) to confirm their identity. Only verified bands were included in the calculation of PSI.

---

Quantitative RT-PCR

cDNA was synthesized as described above. Quantitative PCR (qPCR) was used to amplify 10 ng of cDNA with iTaq Universal SYBR Green Supermix (Bio-Rad no. 1725120) in 384-well plates (Applied Biosystems no. 4309849) using a QuantStudio 7 Flex machine (Applied Biosystems). Primers are listed in table S6B. Data were analyzed using QuantStudio Real-Time PCR software. All gene expression was normalized to the housekeeping gene GAPDH (glyceraldehyde 3-phosphate dehydrogenase).

---

Western blot analysis

Cells were harvested by direct lysis in the cell-culture dish in 2x Laemmli buffer (Bio-Rad no. 1610737) with 5% (v/v) β-mercaptoethanol preheated to 95°C. Samples were boiled at 95°C for 20 min, with brief vortexing during incubation. Samples were allowed to cool and either frozen at -80°C or run on a 4 to 15% gradient TGX Stain-Free gel (Bio-Rad no. 4568086). Gels were transferred onto 0.2-μm nitrocellulose membranes (Bio-Rad no. 1620168) and blocked in 5% (w/v) milk in Tween 20-TBST (50 mM Tris pH 7.5, 150 mM NaCl, 0.05% (v/v) Tween 20). Blots were incubated with primary antibody overnight at 4°C and with secondary antibody for 1 hour at room temperature. A ChemiDoc MP Imaging System (Bio-Rad) was used for infrared or chemiluminescent detection. Blots probed with an horseradish peroxidase (HRP)-conjugated secondary were incubated in Clarity Western ECL Substrate (Bio-Rad no. 1705060) for 1 min before image acquisition. Protein expression was quantified using ImageLab 6.1.0 software (Bio-Rad) and normalized first to GAPDH as a loading control and then expressed as fold change to experimental controls. All antibodies used and their respective dilutions are listed in table S6I.

---

Immunofluorescence

Cells were cultured in black 96-well imaging plates (Revvity no. 6055300) for 48 hours and fixed by incubating wells with 4% (w/v) paraformaldehyde in phosphate-buffered saline (PBS). Cells were permeabilized with 0.2% (v/v) Triton X-100 in PBS and blocked with 10% (v/v) goat serum (Sigma-Aldrich no. G6767 in 0.2% (v/v) Triton X-100 in PBS. Wells were incubated in primary antibody diluted in blocking solution overnight at 4°C and detected using Alexa-fluor-488 anti-rabbit or Alexa-fluor-568 anti-mouse conjugated secondary antibodies (Invitrogen) used at 1:500 in blocking solution. Cells were costained with Hoechst 33342 (ThermoFisher no. H3570, 1:2000) and phalloidin-AlexaFluor 647 conjugate (ThermoFisher no. A22287, 1:1000) in PBS for 15 min at room temperature. Wells were imaged on an Opera Phenix High-Content Screening machine (Revvity). All antibodies used and their respective dilutions are listed in table S6J.

---

ViewRNA FISH

RNA transcripts were visualized in optimal cutting temperature (OCT)-embedded lung tumor sections using the ViewRNA Tissue Assay Fluorescence Kits (ThermoFisher no. QVT0700). Human STAT2 ViewRNA type 4 probes, human KRT8 ViewRNA type 6 probes, and human STAT2\_ri6 or \_ri12 ViewRNA type 1 probes were obtained from ThermoFisher (table S6K). The assay was performed according to the tissue-based ViewRNA assay protocol with formaldehyde fixation and a 20-min protease treatment. ViewRNA probes were detected with Leica TCS SP8 confocal microscope at $40\times$ magnification.

---

Whole-cell extraction and FLAG immunoprecipitation

Doxycycline-inducible A549 cell lines expressing 3×FLAG-tagged STAT2-FL, STAT2-ri12, or STAT2-ri6 were induced with DOX (100 ng/ml) for 48 hours and subsequently stimulated with IFN-I (1000 U/ml) for 1.5 hours before harvest. Whole-cell extracts (WCEs) were prepared using a high-salt lysis protocol. Briefly, cells were washed with ice-cold PBS and lysed in buffer containing 50 mM Tris-HCl (pH 8.0), 450 mM NaCl, 1% Triton X-100, 2 mM MgCl₂, 0.1 mM ZnCl₂, 2 mM EDTA, 10% glycerol, 1 mM DTT, protease inhibitor cocktail, 1 mM Na₃VO₄, and 50 mM NaF for 30 min at 4°C. An equal volume of NaCl-free buffer was added to adjust the final NaCl concentration to 225 mM. Lysates were incubated on ice for 5 min, clarified by centrifugation (14,000 rpm, 30 min, 4°C), and the supernatant collected as WCE.

For FLAG immunoprecipitation, anti-FLAG M2 agarose beads (Sigma no. A2220) preequilibrated in WCE buffer (225 mM NaCl) were incubated with extracts overnight at 4°C with rotation. Beads were washed four times with WCE buffer and FLAG-tagged proteins were eluted sequentially using 3×FLAG peptide (200 μg/ml) in elution buffer (20 mM HEPES pH 7.5, 100 mM KCl, 0.1% Triton X-100, 20% glycerol, 100 $\mu$ M ZnCl $_{2}$ , 1 mM DTT). Eluates were collected after two successive incubations at 4°C before downstream analyses.

---

Long-read RNA-seq of tumor samples

Iso-seq libraries were prepared according to Iso-seq Express Template Preparation (Pacific Biosciences) using 300 ng of RNA. Full length cDNAs were generated using NEBNext Single Cell/Low Input cDNA synthesis and Amplification Module (NEB) in combination with Iso-Seq Express Oligo Kit (Pacific Biosciences), with 12 cycles of PCR amplification. Amplified cDNAs were purified using ProNex beads. All samples yielded more than 160 ng of cDNA and were used for library preparation. SMRTbell library preparation including DNA damage repair, end repair and A-tailing and ligation with SMRTbell Adaptors (Pacific Biosciences) were performed following manufacturer recommendations. Loading protocols for each sample were made using PacBio SMRTLink software following manufacturer recommendations. Libraries were sequenced on either a PacBio Sequel II or PacBio Sequel IIe instrument (Pacific Biosciences). Each sample was individually sequenced on a single SMRT cell.

---

Short-read RNA-seq of tumor samples and cell lines

Short-read RNA-seq libraries were prepared with a KAPA mRNA Hyperprep Kit (Roche) according to manufacturer instructions. First, polyA RNA was isolated from 300 ng total RNA using oligo-dT magnetic beads. Purified RNA was then fragmented at 85°C for 6 min, targeting 250–to 300-base pair (bp) fragments. Fragmented RNA was reverse transcribed with random primers with an incubation of 25°C for 10 min, 42°C for 15 min, and an inactivation step at 70°C for 15 min. This was followed by second strand synthesis and A-tailing at 16°C for 30 min and 62°C for 10 min. A-tailed, double stranded cDNA fragments were ligated with Illumina unique adaptors (Illumina). Adaptor-ligated DNA was purified using KAPA Pure beads. This was followed by 10 cycles of PCR amplification. The final library was cleaned up using KAPA Pure Beads (Roche). Quantification of libraries was performed using real-time qPCR (ThermoFisher). Sequencing was performed on an Illumina NovaSeq 6000 instrument generating paired end reads of 150 bp, targeting 100 million to 200 million reads for each sample.

---

Long-read RNA data processing

Long-read RNA-seq files were processed according to the Iso-Seq pipeline provided by Pacific Biosciences at https://github.com/PacificBiosciences/pbbioconda. Briefly, HiFi CCS BAM files from sequencing runs were trimmed and demultiplexed using lima (v2.7.1) with the-ccs and-peak-guess parameters. The resultant reads were quality controlled to remove polyA tails and artifactual concatemers using isoseq refine (v4.0.0) with the-require-polya parameter. Reads were assembled into distinctive isoform clusters using isoseq cluster2 (v4.0.0) and mapped to the hg38 genome with pbmm2 (v1.13.0). Finally, mapped isoform clusters were collapsed into nonredundant full-length isoforms using isoseq collapse (v4.0.0).

---

Long-read isoform filtering and quality control

Isoforms identified by long-read RNA-seq were filtered to remove low-quality isoforms and artifacts with the SQANTI3 (v5.2) pipeline $(35)$ , found at https://github.com/ConesaLab/SQANTI3. To validate unannotated splicing junctions with SQANTI3, samples selected for long-read RNA-seq were also sequenced with short-read RNA-seq. Briefly, short-read RNA-seq FASTQ files were trimmed using fastp (v0.20.0) and aligned to the hg38 genome using STAR $(82)$ (v2.7.10b) using multi-sample two-pass mapping to maximize sensitivity for detection of unannotated splice junctions. Long-read isoforms were classified using the sqanti3\_qc.py script from the SQANTI3 distribution using the GENCODE v44 comprehensive gene annotation on reference chromosomes $(30)$ as the reference transcriptome, CAGE peaks from the refTSS database (31) (v3.1), the human polyA motif list from SQANTI3, and splice junction output files obtained from STAR. Isoforms were filtered using the rules mode of sqanti3\_filter.py from the SQANTI3 distribution to remove artifacts; for a complete description of filtering parameters used, please refer to the rulesFilter.json file (Data S1). We used per-sample isoform counts from FLAIR (see section "Characterizing long-read isoforms" below) to further filter our isoforms; only isoforms with $\geq5$ long-read counts in $\geq1$ sample or $\geq3$ long-read counts in $\geq2$ samples were retained. The resultant filtered GTF was annotated and merged with well-supported isoforms from the reference GTF GENCODE v44 using a custom Python script.

---

Characterizing long-read isoforms

Per-sample isoform quantification from long-read RNA-seq: We used FLAIR (83) (v3.0.0) to quantify isoform counts in long-read RNA-seq samples. Briefly, our long-read GTF was converted to BED12 format using the gtfToGenePred and genePredToBed utilities from the UCSC Genome Browser (84) (https://hgdownload.soe.ucsc.edu/admin/exe/) with default options. We converted HiFi CCS FASTQ files from sequencing runs into FASTA format from the command line and ran flair quantify using the long-read BED12 file using the options -tpm, -stringent, and -check\_splice. We queried the raw counts file across each patient sample and considered an isoform detected in a sample if the raw count was $\geq 1$ using a custom Python script.

Calculating UTR length differences of full-splice match isoforms detected in LUAD: Isoforms were filtered to include only isoforms categorized as full-splice match or reference match by SQANTI3. The diff\_to\_TSS and diff\_to\_TTS parameters from the SQANTI3 classification step were multiplied by -1 to obtain values for the difference in isoform 5' and 3' UTR lengths relative to the closest matching reference isoform, such that a positive number indicates a longer UTR in that isoform relative to the closest matching reference isoform and a negative number indicates a shorter UTR in that isoform relative to the closest matching reference isoform.

Isoform ORF prediction and comparison to reference sequences: We ran the isoform open reading frame prediction tool CPAT (85) (v3.0.2) on our long-read transcriptome FASTA file using options -top-orf=10 and -min-orf=25. The highest scoring predicted ORF for each isoform was collected and filtered to keep all ORFs with a coding probability score $\geq0.364$ (the optimal value for human ORFs). These putative ORFs were compared to human reference protein sequences from the UniProtKB/Swiss-Prot database (37) using blastp (86) (v2.14.0+). We calculated ORF identity to the UniProt reference as follows: $(2\times\text{matching residues})/(\text{query length}+\text{reference length})\times100\%$ . For each ORF, we took the best ORF identity value and rounded with the floor function. Results were grouped by isoform structural category from SQANTI3 and binned as indicated.

Prediction of isoform coding potential: To identify isoforms with likely frameshifts or in-frame premature termination codons, we used a custom Python script and human reference protein sequences from the UniProtKB/Swiss-Prot database $(37)$ as the reference. For each isoform corresponding to a gene present in the reference, the script translated the transcript in all three reading frames and compared the resulting peptide sequences with the annotated reference protein. We first identified isoforms whose translated products contained N- and C-terminal regions matching those of the reference protein. Isoforms in which these termini were split across different reading frames or were separated by a stop codon within the same frame and whose N-terminal ORFs had a stop codon more than 55 nucleotides upstream of the last splice junction $(87)$ were classified as containing a frameshift or an in-frame PTC, respectively. If the stop codon was no more than 55 nucleotides upstream of the last splice junction, the isoform was classified as encoding a previously uncharacterized C terminus. Isoforms whose translations matched the reference protein's N- and C-termini within a single uninterrupted ORF were instead grouped as (i) exact matches (identical to the reference) or (ii) inexact matches (containing internal mismatches or gaps relative to the reference). Isoforms lacking either the N- or C-terminal match were flagged accordingly. Isoforms that were not able to be classified using these definitions were classified as unknown and not considered in further analysis.

Identification of isoforms with retained introns: To identify isoforms with retained introns, we used a custom Python script to collect all possible retained introns from our long-read GTF and remove those introns that are retained in the canonical reference isoform from the GENCODE v44 comprehensive gene annotation $(30)$ .

Classification of isoforms: Isoforms were grouped into the following categories based on their coding potential and intron retention status: "Exact match to reference protein" if the isoform contained an ORF that exactly matched a reference protein; "Novel protein sequence" if the closest matching ORF contained a novel amino acid sequence either internally or at the C terminus; "Intron retention leading to protein truncation" if the isoform contained a frameshift or PTC-containing ORF and contained a retained intron; or "Protein truncation without intron retention" if the isoform contained a frameshift or PTC-containing ORF and did not contain a retained intron.

Machine-learning-based feature importance analysis: Each isoform in the SQANTI3 classification table was labeled “Isoform” or “Artifact” according to its filter\_result. Per isoform structural category, a gradient-boosted classifier (XGBoost v3.2.0; 300 trees, max\_depth = 5, learning rate = 0.05) was trained to recover the Isoform/Artifact decision from the SQANTI3 feature set, with mean area under the ROC curve (AUC) estimated by stratified five-fold cross-validation (scikit-learn v1.7.2, sklearn.model\_selection.StratifiedKFold, sklearn.metrics.roc\_curve and auc). Per-feature importance was then quantified as the mean reduction in AUC across 30 independent random permutations of that feature's values across isoforms, using sklearn.inspection.permutation\_importance with scoring="roc\_auc."

---

Differential splicing and gene expression analysis

We used a custom in-house pipeline to conduct differential splicing analysis, available at https://github.com/TheJacksonLaboratory/splicing-pipelines-nf. Briefly, the pipeline was used to trim FASTQ files with trimmomatic (88) (v0.39) and align them to the human reference genome with the merged long-read and reference GTF using STAR (82) (v2.7.9a). A PSI for each splicing event in each sample was calculated using rMATS (89) (v4.1.2) using both splice junction read counts as well as exon and intron body read counts to infer PSI. When applicable, a $\Delta$ PSI was calculated wherein $\Delta$ PSI = (mean PSI) $_{case}$ – (mean PSI) $_{control}$ . Gene and isoform count estimates were made using StringTie (90) (v2.1.7), with STAR-aligned BAM files as input and compiled into a counts matrix using a custom Python script. Differential gene analysis was conducted using DESeq2 (91) (v1.44.0) using a Wald test to calculate P values and the Benjamini-Hochberg method to derive FDR-corrected P values.

---

TCGA splicing analysis, gene expression, and tumor purity estimates

BAM files corresponding to genome-mapped RNA-seq files from LUAD patients (TCGA-LUAD) were downloaded from the Genomic Data Commons (accession no. phs000178.v11.p8) using the gdc-client executable (https://github.com/NCI-GDC/gdc-client) with a manifest derived by checking options for Data Type:Aligned Reads, Experimental Strategy:RNA-Seq, Workflow:STAR 2-Pass Transcriptome, and Data Format:bam. BAM files were converted back to raw FASTQ format using samtools (92) (v1.4.1) to sort BAM files by name (option -n) and the bamtofastq utility from bedtools (93) to convert the sorted BAM file into paired-end FASTQ files. FASTQC (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/, v0.11.9) was used to calculate quality metrics for the resultant FASTQ files. Only samples for which both FASTQ files passed default metrics were considered for further analysis. Splicing analysis and gene expression analysis were conducted using the in-house pipeline described above with a read length of 50 bp and strandedness set to “false” to indicate that libraries were unstranded. Tumor purity estimates from TCGA-LUAD (62) were downloaded from https://raw.githubusercontent.com/BioinformaticsFMRP/TCGAbiolinks/master/data/Tumor.purity.rda.

---

Splicing analysis and gene expression from OAK and IMvigor210

FASTQ files from pretreatment tumor samples from nonsquamous non-small cell lung cancer patients from OAK, metastatic urothelial carcinoma patients from IMvigor210, and associated metadata were downloaded from the European Genome-phenome Archive (accession numbers EGAD00001008549, EGAD00001008550, EGAD00001007703, EGAD00001006630, and EGAD00001006631). Splicing analysis and gene expression analysis were conducted using the in-house pipeline described above with a read length of 51 bp and strandedness set to "first-strand" to indicate that libraries were reverse stranded.

---

NMD score

To score NMD efficiency, we used PSI values from our rMATS-based pipeline for 12 NMD-sensitive poison exons in SR family splicing factors (58–60). Splicing event PSI values were Z-scored and individual event Z-scores were summed in each sample, then multiplied by -1 to derive the NMD score. Splicing event IDs for the included events are listed in table S5, F and G. FASTQ files from siRNA-mediated UPF1 silencing experiments were downloaded from the Gene Expression Omnibus (accession number GSE176197), and splicing analysis and gene expression analysis were conducted using the in-house pipeline described above with a read length of 150 bp and strandedness set to “first-strand” to indicate libraries were reverse stranded.

---

IFN signaling scores

IFNG.GS and ISG.RS scores were calculated using gene lists as previously described $(69)$ . Briefly, VST-normalized counts from DESeq2 of genes from each list (excluding STAT2) were summed within each sample and Z-scored across each group.

---

Gene set enrichment analysis

To conduct gene set enrichment analysis on differentially expressed genes, genes were ranked by their Wald test statistics as calculated by DESeq2 (91) (v1.44.0). GSEApy (94) (v1.1.3) was used to conduct enrichment analysis on the ranked gene list using the "Prerank" module with the Molecular Signatures Database (MSigDB) Hallmark 2020 Gene Set (95).

---

Prediction of RBP motifs

We used the tool RBPmap at http://rbpmap.technion.ac.il/ (52) to identify predicted RNA-binding protein motifs that overlapped the target sites of STAT2 ri6- and ri12-targeting ASOs. Overlapping motifs were visualized using the UCSC genome browser.

---

AlphaFold 3 protein structure predictions

To predict protein structures, polypeptide sequences were processed by AlphaFold 3 (96) using the online server at: https://alphafoldserver.com. Resultant protein structure files were visualized using ChimeraX-1.9 (97).

---

Statistical analysis

For the proportion of novel isoforms stratified by coding sequence impact, significance was assessed with a hypergeometric test taking the P value for over- or underenrichment as appropriate. For splicing from short-read RNA-seq, significance was assessed using the rMATS statistical model $(89)$ using the FDR-corrected P value. For subsampling analysis, an equal number of tumor and normal samples were selected using the function random.choice from numpy (v1.26.4), and significance was assessed using a two-sided Mann-Whitney U test and was repeated 100 times. All gene, pathway, and cell-type enrichment analysis results were filtered for FDR-adjusted P < 0.05. For RT-PCR, Western blots, and quantitative RT-PCR, data are shown as mean ± SD, and significance was assessed using an unpaired two-tailed Student's t test unless otherwise specified. Survival analysis was conducted using right-censored survival data from OAK and IMvigor210, and significance was assessed using the Wald test without multiple comparisons adjustments for Cox proportional hazards models using the R package survivalAnalysis (v0.3.0) or the log-rank test for Kaplan-Meier curves using the R package survival (v3.8.3) (98). We used the function surv\_cutpoint from the R package survminer (v0.5.0) on biomarker-negative (TMB < 16 mutations/Mb or TMB not specified) patients from OAK and IMvigor210 to identify optimal PSI $_{ri12-ri6}$ cutoffs for Kaplan-Meier survival analysis.

---

Generative AI for writing code

We used Claude Code (Claude Sonnet 4.6; Anthropic, PBC) for assistance with writing code and as an editorial aide during the editing process. All artificial intelligence (AI)-generated code was produced in specific subdirectories without access to parent data or the ability to write outside of those folders. Prompts for script generation included strict instructions to not modify any data in parent directories, and all AI-generated or AI-edited scripts were manually checked before use.

---

Graphs and figures

Data were processed using Python or R, and plots were generated with R using the ggplot2 library or Microsoft Excel. Figures were made with Adobe Illustrator 2023 for Mac in compliance with Science/AAAS guidelines concerning image integrity.

---

REFERENCES AND NOTES

1. A. Leiter, R. R. Veluswamy, J. P. Wisnivesky, The global burden of lung cancer: current status and future trends. Nat. Rev. Clin. Oncol. 20, 624–639 (2023). doi: 10.1038/s41571-023-00798-3; pmid: 37479810

2. A. Rittmeyer et al., Atezolizumab versus docetaxel in patients with previously treated non-small-cell lung cancer (OAK): A phase 3, open-label, multicentre randomised controlled trial. Lancet 389, 255–265 (2017). doi: 10.1016/S0140-6736(16)32517-X; pmid: 27979383

3. M. Reck et al., Pembrolizumab versus chemotherapy for PD-L1-positive non-small-cell lung cancer. N. Engl. J. Med. 375, 1823–1833 (2016). doi: 10.1056/NEJMoa1606774; pmid: 27718847

4. M. Mino-Kenudson et al., Predictive biomarkers for immunotherapy in lung cancer: Perspective from the International Association for the Study of Lung Cancer Pathology Committee. J. Thorac. Oncol. 17, 1335–1354 (2022). doi: 10.1016/j.jtho.2022.09.109; pmid: 36184066

5. K. S. Manning, T. A. Cooper, The roles of RNA processing in translating genotype to phenotype. Nat. Rev. Mol. Cell Biol. 18, 102–114 (2017). doi: 10.1038/nrm.2016.139; pmid: 27847391

6. Q. Pan, O. Shai, L. J. Lee, B. J. Frey, B. J. Blencowe, Deep surveying of alternative splicing complexity in the human transcriptome by high-throughput sequencing. Nat. Genet. 40, 1413–1415 (2008). doi: 10.1038/ng.259; pmid: 18978789

7. E. T. Wang et al., Alternative isoform regulation in human tissue transcriptomes. Nature 456, 470–476 (2008). doi: 10.1038/nature07509; pmid: 18978772

8. L. M. Urbanski, N. Leclair, O. Anczuków, Alternative-splicing defects in cancer: Splicing regulators and their downstream targets, guiding the way to novel cancer therapeutics. Wiley Interdiscip. Rev. RNA 9, e1476 (2018). doi: 10.1002/wrna.1476; pmid: 29693319

9. A. Kahles et al., Comprehensive analysis of alternative splicing across tumors from 8,705 patients. Cancer Cell 34, 211–224.e6 (2018). doi: 10.1016/j.ccell.2018.07.001; pmid: 30078747

10. R. K. Bradley, O. Anczuków, RNA splicing dysregulation and the hallmarks of cancer. Nat. Rev. Cancer 23, 135–155 (2023). doi: 10.1038/s41568-022-00541-7; pmid: 36627445

11. O. Anczukow et al., Steering research on mRNA splicing in cancer towards clinical translation. Nat. Rev. Cancer 24, 887–905 (2024). doi: 10.1038/s41568-024-00750-2; pmid: 39384951

12. D. Sharon, H. Tilgner, F. Grubert, M. Snyder, A single-molecule long-read survey of the human transcriptome. Nat. Biotechnol. 31, 1009–1014 (2013). doi: 10.1038/nbt.2705; pmid: 24108091

13. L. Ermini, P. Driguez, The application of long-read sequencing to cancer. Cancers 16, 1275 (2024). doi: 10.3390/cancers16071275; pmid: 38610953

14. A. Rhoads, K. F. Au, PacBio sequencing and its applications. Genomics Proteomics Bioinformatics 13, 278–289 (2015). doi: 10.1016/j.gpb.2015.08.002; pmid: 26542840

15. Q. Shi, X. Li, Y. Liu, Z. Chen, X. He, FLIBase: A comprehensive repository of full-length isoforms across human cancers and tissues. Nucleic Acids Res. 52, D124–D133 (2024). doi: 10.1093/nar/gkad745; pmid: 37697439

16. M. Oka et al., Aberrant splicing isoforms detected by full-length transcriptome sequencing as transcripts of potential neoantigens in non-small cell lung cancer. Genome Biol. 22, 9 (2021). doi: 10.1186/s13059-020-02240-8; pmid: 33397462

17. Y. Chen et al., A systematic benchmark of Nanopore long-read RNA sequencing for transcript-level analysis in human cell lines. Nat. Methods 22, 801–812 (2025). doi: 10.1038/s41592-025-02623-4; pmid: 40082608

18. D. F. T. Veiga et al., A comprehensive long-read isoform analysis platform and sequencing resource for breast cancer. Sci. Adv. 8, eabg6711 (2022). doi: 10.1126/sciadv.abg6711; pmid: 35044822

19. Z. Li et al., An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing. Cell Genom. 4, 100641 (2024). doi: 10.1016/j.xgen.2024.100641; pmid: 39216476

20. C.-K. Shiau et al., High throughput single cell long-read sequencing analyses of same-cell genotypes and phenotypes in human tumors. Nat. Commun. 14, 4124 (2023). doi: 10.1038/s41467-023-39813-7; pmid: 37433798

21. Q. Sun et al., Long-read sequencing reveals the landscape of aberrant alternative splicing and novel therapeutic target in colorectal cancer. Genome Med. 15, 76 (2023). doi: 10.1186/s13073-023-01226-y; pmid: 37735421

22. G. P. Dunn, C. M. Koebel, R. D. Schreiber, Interferons, immunity and cancer immunoediting. Nat. Rev. Immunol. 6, 836–848 (2006). doi: 10.1038/nri1961; pmid: 17063185

23. N. Jacquelot et al., Sustained type I interferon signaling as a mechanism of resistance to PD-1 blockade. Cell Res. 29, 846–861 (2019). doi: 10.1038/s41422-019-0224-x; pmid: 31481761

24. D. Memon et al., Clinical and molecular features of acquired resistance to immunotherapy in non-small cell lung cancer. Cancer Cell 42, 209–224.e9 (2024). doi: 10.1016/j.ccell.2023.12.013; pmid: 38215748

25. J. Qiu et al., Cancer cells resistant to immune checkpoint blockade acquire interferon-associated epigenetic memory to sustain T cell dysfunction. Nat. Cancer 4, 43–61 (2023). doi: 10.1038/s43018-022-00490-y; pmid: 36646856

26. J. L. Benci et al., Tumor interferon signaling regulates a multigenic resistance program to immune checkpoint blockade. Cell 167, 1540–1554.e12 (2016). doi: 10.1016/j.cell.2016.11.022; pmid: 27912061

27. G. P. Dunn et al., A critical function for type I interferons in cancer immunoediting. Nat. Immunol. 6, 722–729 (2005). doi: 10.1038/ni1213; pmid: 15951814

28. A. M. Wenger et al., Accurate circular consensus long-read sequencing improves variant detection and assembly of a human genome. Nat. Biotechnol. 37, 1155–1162 (2019). doi: 10.1038/s41587-019-0217-9; pmid: 31406327

29. S. Guizard et al., nf-core/isoseq: Simple gene and isoform annotation with PacBio Iso-Seq long-read sequencing. Bioinformatics 39, btad150 (2023). doi: 10.1093/bioinformatics/btad150; pmid: 36961337

30. A. Frankish et al., GENCODE reference annotation for the human and mouse genomes. Nucleic Acids Res. 47, D766–D773 (2019). doi: 10.1093/nar/gky955; pmid: 30357393

31. I. Abugessaisa et al., refTSS: A reference data set for human and mouse transcription start sites. J. Mol. Biol. 431, 2407–2422 (2019). doi: 10.1016/j.jmb.2019.04.045; pmid: 31075273

32. P. Carninci et al., Genome-wide analysis of mammalian promoter architecture and evolution. Nat. Genet. 38, 626–635 (2006). doi: 10.1038/ng1789; pmid: 16645617

33. T. Shiraki et al., Cap analysis gene expression for high-throughput analysis of transcriptional starting point and identification of promoter usage. Proc. Natl. Acad. Sci. U.S.A. 100, 15776–15781 (2003). doi: 10.1073/pnas.2136655100; pmid: 14663149

34. C. J. Herrmann et al., PolyASite 2.0: A consolidated atlas of polyadenylation sites from 3' end sequencing. Nucleic Acids Res. 48, D174–D179 (2019). doi: 10.1093/nar/gkz918; pmid: 31617559

35. F. J. Pardo-Palacios et al., SQANTI3: Curation of long-read transcriptomes for accurate identification of known and novel isoforms. Nat. Methods 21, 793–797 (2024). doi: 10.1038/s41592-024-02229-2; pmid: 38509328

36. C. Mayr, D. P. Bartel, Widespread shortening of 3'UTRs by alternative cleavage and polyadenylation activates oncogenes in cancer cells. Cell 138, 673–684 (2009). doi: 10.1016/j.cell.2009.06.016; pmid: 19703394

37. A. Bateman et al., UniProt: The Universal Protein Knowledgebase in 2023. Nucleic Acids Res. 51, D523–D531 (2023). doi: 10.1093/nar/gkac1052; pmid: 36408920

38. D. F. Grabski et al., Intron retention and its impact on gene expression and protein diversity: A review and a practical guide. Wiley Interdiscip. Rev. RNA 12, e1631 (2021). doi: 10.1002/wrna.1631; pmid: 33073477

39. C. J. Wright, C. W. J. Smith, C. D. Jiggins, Alternative splicing as a source of phenotypic diversity. Nat. Rev. Genet. 23, 697–710 (2022). doi: 10.1038/s41576-022-00514-4; pmid: 35821097

40. A. Lahiri et al., Lung cancer immunotherapy: Progress, pitfalls, and promises. Mol. Cancer 22, 40 (2023). doi: 10.1186/s12943-023-01740-y; pmid: 36810079

41. M. Reck, J. Remon, M. D. Hellmann, First-line immunotherapy for non-small-cell lung cancer. J. Clin. Oncol. 40, 586–597 (2022). doi: 10.1200/JCO.21.01497; pmid: 34985920

42. M. Spella, G. T. Stathopoulos, Immune resistance in lung adenocarcinoma. Cancers 13, 384 (2021). doi: 10.3390/cancers13030384; pmid: 33494181

43. C. Yue et al., Host STAT2/type I interferon axis controls tumor growth. Int. J. Cancer 136, 117–126 (2015). doi: 10.1002/ijc.29004; pmid: 24895110

44. S. Leung, S. A. Qureshi, I. M. Kerr, J. E. Darnell Jr., G. R. Stark, Role of STAT2 in the alpha interferon signaling pathway. Mol. Cell. Biol. 15, 1312–1317 (1995). doi: 10.1128/MCB.15.3.1312; pmid: 7532278

45. J. B. Fan et al., Type I interferon regulates a coordinated gene network to enhance cytotoxic T cell-mediated tumor killing. Cancer Discov. 10, 382–393 (2020). doi: 10.1158/2159-8290.CD-19-0608; pmid: 31974171

46. L. Yang, A. Shen, R. Wang, Z. Zheng, S100A16 stabilizes the ITGA3-mediated ECM-receptor interaction pathway to drive the malignant properties of lung adenocarcinoma cells via binding MOV10. Mol. Med. Rep. 31, 11 (2025). doi: 10.3892/mmr.2024.13376; pmid: 39450567

47. R. Yang et al., Multi-omics analysis reveals interferon-stimulated gene OAS1 as a prognostic and immunological biomarker in pan-cancer. Front. Immunol. 14, 1249731 (2023). doi: 10.3389/fimmu.2023.1249731; pmid: 37928544

48. Z. Xiao et al., MATR3 promotes liver cancer progression by suppressing DHX58-mediated type I interferon response. Cancer Lett. 604, 217231 (2024). doi: 10.1016/j.canlet.2024.217231; pmid: 39276912

49. D. C. Castaneda et al., Spatiotemporally organized immunomodulatory response to SARS-CoV-2 virus in primary human broncho-alveolar epithelia. iScience 26, 107374 (2023). doi: 10.1016/j.isci.2023.107374; pmid: 37520727

50. J. Sainz de Aja, A. F. M. Dost, C. F. Kim, Alveolar progenitor cells and the origin of lung cancer. J. Intern. Med. 289, 629–635 (2021). doi: 10.1111/joim.13201; pmid: 33340175

51. Z. Wang et al., Deciphering cell lineage specification of human lung adenocarcinoma with single-cell RNA sequencing. Nat. Commun. 12, 6500 (2021). doi: 10.1038/s41467-021-26770-2; pmid: 34764257

52. I. Paz, I. Kosti, M. Ares Jr., M. Cline, Y. Mandel-Gutfreund, RBPmap: A web server for mapping binding sites of RNA-binding proteins. Nucleic Acids Res. 42, W361–W367 (2014). doi: 10.1093/nar/gku406; pmid: 24829458

53. A. E. Campbell et al., Compromised nonsense-mediated RNA decay results in truncated RNA-binding protein production upon DUX4 expression. Cell Rep. 42, 112642 (2023). doi: 10.1016/j.celrep.2023.112642; pmid: 37314931

54. H. Sato, R. H. Singer, Cellular variability of nonsense-mediated mRNA decay. Nat. Commun. 12, 7203 (2021). doi: 10.1038/s41467-021-27423-0; pmid: 34893608

55. K. Litchfield et al., Escape from nonsense-mediated decay associates with anti-tumor immunogenicity. Nat. Commun. 11, 3800 (2020). doi: 10.1038/s41467-020-17526-5; pmid: 32733040

56. R. G. H. Lindeboom, M. Vermeulen, B. Lehner, F. Supek, The impact of nonsense-mediated mRNA decay on genetic disease, gene editing and cancer immunotherapy. Nat. Genet. 51, 1645–1651 (2019). doi: 10.1038/s41588-019-0517-5; pmid: 31659324

57. G. Palou-Márquez, F. Supek, Variable efficiency of nonsense-mediated mRNA decay across human tissues, tumors and individuals. Genome Biol. 26, 316 (2025). doi: 10.1186/s13059-025-03727-y; pmid: 41023738

58. L. F. Lareau, S. E. Brenner, Regulation of splicing factors by alternative splicing and NMD is conserved between kingdoms yet evolutionarily flexible. Mol. Biol. Evol. 32, 1072–1079 (2015). doi: 10.1093/molbev/msv002; pmid: 25576366

59. N. K. Leclair et al., Poison exon splicing regulates a coordinated network of SR protein expression during differentiation and tumorigenesis. Mol. Cell 80, 648–665.e9 (2020). doi: 10.1016/j.molcel.2020.10.019; pmid: 33176162

60. L. F. Lareau, M. Inada, R. E. Green, J. C. Wengrod, S. E. Brenner, Unproductive splicing of SR genes associated with highly conserved and ultraconserved DNA elements. Nature 446, 926–929 (2007). doi: 10.1038/nature05676; pmid: 17361132

61. S. E. Fritz, S. Ranganathan, C. D. Wang, J. R. Hogg, An alternative UPF1 isoform drives conditional remodeling of nonsense-mediated mRNA decay. EMBO J. 41, e108898 (2022). doi: 10.15252/embj.2021108898; pmid: 35403729

62. D. Aran, M. Sirota, A. J. Butte, Systematic pan-cancer analysis of tumour purity. Nat. Commun. 6, 8971 (2015). doi: 10.1038/ncomms9971; pmid: 26634437

63. A. Gopalsamy et al., Identification of pyrimidine derivatives as hSMG-1 inhibitors. Bioorg. Med. Chem. Lett. 22, 6636–6641 (2012). doi: 10.1016/j.bmcl.2012.08.107; pmid: 23021994

64. S. Gupta et al., The SH2 domains of Stat1 and Stat2 mediate multiple interactions in the transduction of IFN- $\alpha$ signals. EMBO J. 15, 1075–1084 (1996). doi: 10.1002/j.1460-2075.1996.tb00445.x; pmid: 8605877

65. S. Rengachari et al., Structural basis of STAT2 recognition by IRF9 reveals molecular insights into ISGF3 function. Proc. Natl. Acad. Sci. U.S.A. 115, E601–E609 (2018). doi: 10.1073/pnas.1718426115; pmid: 29317535

66. M. Martinez-Moczygemba, M. J. Gutch, D. L. French, N. C. Reich, Distinct STAT structure promotes interaction of STAT2 with the p48 subunit of the interferon- $\alpha$ -stimulated

transcription factor ISGF3. J. Biol. Chem. 272, 20070–20076 (1997). doi: 10.1074/jbc.272.32.20070; pmid: 9242679

67. M. B. Fuertes, S. R. Woo, B. Burnett, Y. X. Fu, T. F. Gajewski, Type I interferon response and innate immune sensing of cancer. Trends Immunol. 34, 67–73 (2013). doi: 10.1016/j.it.2012.10.004; pmid: 23122052

68. N. S. Patil et al., Intratumoral plasma cells predict outcomes to PD-L1 blockade in non-small cell lung cancer. Cancer Cell 40, 289–300.e4 (2022). doi: 10.1016/j.ccell.2022.02.002; pmid: 35216676

69. J. L. Benci et al., Opposing functions of interferon coordinate adaptive and innate immune responses to cancer immune checkpoint blockade. Cell 178, 933–948.e14 (2019). doi: 10.1016/j.cell.2019.07.019; pmid: 31398344

70. J. E. Rosenberg et al., Atezolizumab monotherapy for metastatic urothelial carcinoma: Final analysis from the phase II IMvigor210 trial. ESMO Open 9, 103972 (2024). doi: 10.1016/j.esmoop.2024.103972; pmid: 39642637

71. D. L. Jardim, A. Goodman, D. de Melo Gagliato, R. Kurzrock, The challenges of tumor mutational burden as an immunotherapy biomarker. Cancer Cell 39, 154–173 (2021). doi: 10.1016/j.ccell.2020.10.001; pmid: 33125859

72. D. R. Camidge, R. C. Doebele, K. M. Kerr, Comparing and contrasting predictive biomarkers for immunotherapy and targeted therapy of NSCLC. Nat. Rev. Clin. Oncol. 16, 341–355 (2019). doi: 10.1038/s41571-019-0173-9; pmid: 30718843

73. Y. Li et al., Prognostic alternative mRNA splicing signature in non-small cell lung cancer. Cancer Lett. 393, 40–51 (2017). doi: 10.1016/j.canlet.2017.02.016; pmid: 28223168

74. B. Fair et al., Global impact of unproductive splicing on human gene expression. Nat. Genet. 56, 1851–1861 (2024). doi: 10.1038/s41588-024-01872-x; pmid: 39223315

75. D. Inoue et al., Spliceosomal disruption of the non-canonical BAF complex in cancer. Nature 574, 432–436 (2019). doi: 10.1038/s41586-019-1646-9; pmid: 31597964

76. B. Schwanhäusser et al., Global quantification of mammalian gene expression control. Nature 473, 337–342 (2011). doi: 10.1038/nature10098; pmid: 21593866

77. D. A. Glinos et al., Transcriptome variation in human tissues revealed by long-read sequencing. Nature 608, 353–359 (2022). doi: 10.1038/s41586-022-05035-y; pmid: 35922509

78. X. Cao et al., Invited review: Human air-liquid-interface organotypic airway tissue models derived from primary tracheobronchial epithelial cells-overview and perspectives. In Vitro Cell. Dev. Biol. Anim. 57, 104–132 (2021). doi: 10.1007/s11626-020-00517-7; pmid: 33175307

79. R. E. Rayner, P. Makena, G. L. Prasad, E. Cormet-Boyaka, Optimization of normal human bronchial epithelial (NHBE) cell 3D cultures for in vitro lung model studies. Sci. Rep. 9, 500 (2019). doi: 10.1038/s41598-018-36735-z; pmid: 30679531

80. J. W. Schoggins, C. M. Rice, Interferon-stimulated genes and their antiviral effector functions. Curr. Opin. Virol. 1, 519–525 (2011). doi: 10.1016/j.coviro.2011.10.008; pmid: 22328912

81. N. K. Leclair et al., Antisense oligonucleotide-mediated TRA2β poison exon inclusion induces the expression of a lncRNA with anti-tumor effects. Nat. Commun. 16, 1670 (2025). doi: 10.1038/s41467-025-56913-8; pmid: 39955311

82. A. Dobin et al., STAR: Ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013). doi: 10.1093/bioinformatics/bts635; pmid: 23104886

83. A. D. Tang et al., Full-length transcript characterization of SF3B1 mutation in chronic lymphocytic leukemia reveals downregulation of retained introns. Nat. Commun. 11, 1438 (2020). doi: 10.1038/s41467-020-15171-6; pmid: 32188845

84. G. Perez et al., The UCSC Genome Browser database: 2025 update. Nucleic Acids Res. 53, D1243–D1249 (2025). doi: 10.1093/nar/gkae974; pmid: 39460617

85. L. Wang et al., CPAT: Coding-Potential Assessment Tool using an alignment-free logistic regression model. Nucleic Acids Res. 41, e74 (2013). doi: 10.1093/nar/gkt006; prmid: 23335781

86. C. Camacho et al., BLAST+: Architecture and applications. BMC Bioinformatics 10, 421 (2009). doi: 10.1186/1471-2105-10-421; pmid: 20003500

87. L. E. Maquat, Nonsense-mediated mRNA decay: Splicing, translation and mRNP dynamics. Nat. Rev. Mol. Cell Biol. 5, 89–99 (2004). doi: 10.1038/nrm1310; pmid: 15040442

88. A. M. Bolger, M. Lohse, B. Usadel, Trimmomatic: A flexible trimmer for Illumina sequence data. Bioinformatics 30, 2114–2120 (2014). doi: 10.1093/bioinformatics/btu170; pmid: 24695404

89. S. Shen et al., rMATS: Robust and flexible detection of differential alternative splicing from replicate RNA-Seq data. Proc. Natl. Acad. Sci. U.S.A. 111, E5593–E5601 (2014). doi: 10.1073/pnas.141916111; pmid: 25480548

90. M. Pertea et al., StringTie enables improved reconstruction of a transcriptome from RNA-seq reads. Nat. Biotechnol. 33, 290–295 (2015). doi: 10.1038/nbt.3122; pmid: 25690850

91. M. I. Love, W. Huber, S. Anders, Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014). doi: 10.1186/s13059-014-0550-8; pmid: 25516281

92. P. Danecek et al., Twelve years of SAMtools and BCFtools. Gigascience 10, giab008 (2021). doi: 10.1093/gigascience/giab008; pmid: 33590861

93. A. R. Quinlan, I. M. Hall, BEDTools: A flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842 (2010). doi: 10.1093/bioinformatics/btq033; pmid: 20110278

94. Z. Fang, X. Liu, G. Peltz, GSEApy: A comprehensive package for performing gene set enrichment analysis in Python. Bioinformatics 39, btac757 (2023). doi: 10.1093/bioinformatics/btac757; pmid: 36426870

95. A. Liberzon et al., The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst. 1, 417–425 (2015). doi: 10.1016/j.cels.2015.12.004; pmid: 26771021

96. J. Abramson et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024). doi: 10.1038/s41586-024-07487-w; pmid: 38718835

97. E. C. Meng et al., UCSF ChimeraX: Tools for structure building and analysis. Protein Sci. 32, e4792 (2023). doi: 10.1002/pro.4792; pmid: 37774136

98. D. R. Cox, Regression models and life-tables. J. R. Stat. Soc. Series B Stat. Methodol. 34, 187–202 (1972). doi: 10.1111/j.2517-6161.1972.tb00899.x

99. L. Urbanski et al., MYC regulates a pan-cancer network of co-expressed oncogenic splicing factors. Cell Rep. 41, 111704 (2022). doi: 10.1016/j.celrep.2022.111704; pmid: 36417849

---

ACKNOWLEDGMENTS

We thank all members of the Anczuków and Palucka labs for helpful discussions and comments on the manuscript and A. Patel from ThermoFisher for assistance with ViewRNA probe design. We acknowledge assistance from the Microscopy, Genome Editing, and Genome Technologies Sequencing Services at The Jackson Laboratory, which are supported in part by National Cancer Institute (NCI) grant P30CA034196. We used Claude Code (Claude Sonnet 4.6; Anthropic, PBC) as an assistive editing tool to optimize code design and condense language during the editing process. We performed molecular graphics and analyses with UCSF ChimeraX, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from National Institutes of Health grant R01-GM129325 and the Office of Cyber Infrastructure and Computational Biology, National Institute of Allergy and Infectious Diseases. This research makes use of data generated by Genentech. Funding: We acknowledge funding from National Institutes of Health grants U19AI142733 (K.P., O.A.), R01CA248317 (O.A.), R01GM138541 (O.A.), F30CA278494 (R.P.E.), and P30CA034196 (K.P., O.A., R.P.E.) and funding from Sanofi S.A. SANOFI-SRA-FY19-JFB (J.B.). Author contributions: Conceptualization: R.P.E., O.A., J.B., K.P.; Methodology: R.P.E., O.A., K.P., M.K., N.K.L., M.B., T.-C.W., S.P., I.W., F.M., M.E.P., P.C.; Investigation: R.P.E., M.B., A.S., P.Y., R.B., M.D., H.G.K., I.W.; Visualization: R.P.E., O.A., M.B.; Funding acquisition: O.A., K.P., J.B., R.P.E.; Project administration: R.P.E., O.A., K.P.; Supervision: O.A., K.P., J.B.; Writing – original draft: R.P.E., O.A., M.B.; Writing – review & editing: R.P.E., O.A., K.P., R.B., J.B., N.K.L., M.B., S.P. Competing interests: R.P.E., O.A., N.K.L., J.B., and K.P. are listed on a patent application (D25-008-USPROV1, US provisional application no. 63/852,095) submitted by The Jackson Laboratory that covers the use of splice-modulating ASOs targeting STAT2 as well as STAT2 intron retention events as immunotherapy biomarkers. In the early stage of the study, J.B. was a member of the board of directors and scientific advisory board of Neovacs and Ascend Biopharma and a scientific advisory board member of Cue Biopharma. Now, J.B. is the founder of Immunoledge LLC, an entity designed to advise biotech start-ups. In this role, J.B. serves as chief information officer of Georgiamune, scientific advisory board member of Metis Therapeutics, and adviser to The Jackson Laboratory. During the course of this study, K.P. was a member of the advisory board for Cue BioPharma and a founding member of Guardian Bio. OA is a member of the scientific advisory board of NeoSplice, Akari Scientific, and EnteloBio. The authors declare no other competing interests. Data, code, and materials availability: Our splicing analysis pipeline v2.0 has previously been described (99) and is freely available on Github (https://github.com/The.JacksonLaboratory/splicing-pipelines-nf). All other data are available in the paper and supplementary materials or are available from the corresponding author upon request. Tables S1 to S6 may also be accessed at https://github.com/ryanpe13002/Supplementary-Tables-2/. Sequencing data generated as part of this study are available on dbGaP (the Database of Genotypes and Phenotypes) under accession no. phs004211.v1.p1 and ImmPort under study accession no. SDY3373. License information: Copyright © 2026 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licenses-journal-article-reuse

---

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.aea7156

Figs. S1 to S8; Tables S1 to S6; MDAR Reproducibility Checklist; Data S1

Submitted 24 July 2025; resubmitted 1 April 2026; accepted 26 August 2026

10.1126/science.aea7156

Decoupled self-assembled monolayer and perovskite deposition for efficient flexible solar modules

Yeyong Wu, Xikun Zhang, Tiankai Zhang, Guiying Xu\*, Yunxiu Shen, Chengxin Guo, Qing Miao, Junyuan Ding, Yue Yin, Qinrong Cheng, Wen Liu, Xiaoxiao Wu, Kaiqiang Chen, Wenwen Kou, Zhiyun Cao, Haiyang Chen, Weijie Chen, Yujin Chen, Feng Gao, Yaowen Li\*, Yongfang Li

Full article and list of author affiliations: https://doi.org/10.1126/science.aef9745

INTRODUCTION: Flexible perovskite solar cells (PSCs) enable efficient solar energy harvesting in curved and mobile scenarios such as buildings, transportation, wearables, and portable devices, substantially expanding the clean energy supply boundary. Although small-area flexible PSCs have reached a power conversion efficiency (PCE) of 26.22%, approaching that of rigid PSCs, device upscaling triggers pronounced declines in both PCE and operational stability, creating a notable performance gap between large-area flexible perovskite solar modules (PSMs) and their rigid counterparts. This considerable efficiency and stability deficit remains a critical barrier to the commercialization of both single-junction and tandem flexible perovskite photovoltaics.

RATIONALE: The inferior performance of large-area flexible PSMs is closely associated with the distinct surface properties of flexible substrates. Their ITO surfaces typically have a lower density and less uniform distribution of hydroxyl groups, compromising self-assembled monolayer (SAM) anchoring and wetting of polar perovskite precursors. Surface treatment can temporarily increase hydroxyl density, whereas SAM-perovskite precursor codeposition can overcome the wetting limitation. However, these hydroxyl groups are rapidly lost in the polar precursor solution, creating a kinetic mismatch between hydroxyl loss and SAM anchoring. Specifically, interactions between perovskite colloids and the phosphonic acid groups of SAM molecules retard SAM adsorption, causing hydroxyl groups to be depleted before efficient SAM anchoring and ultimately compromising SAM coverage and perovskite crystallization. Accordingly, we introduced 3,5-bis(methylsulfonyl)phenylammonium iodide (BSAI) as a decoupling agent into the codeposition precursor. We hypothesized that BSAI preferentially interacts with perovskite precursor species, weakening their interaction with Me-4PACz and accelerating SAM anchoring.

RESULTS: Systematic investigations reveal BSAI exerted a dual function through its ammonium and dual sulfonyl groups bound strongly to perovskite colloids. First, it disrupted coordination between Pb-I colloids and the phosphonate groups of Me-4PACz to facilitate the quick adsorption to the substrate before surface hydroxyl groups desorption, enabling uniform SAM anchoring on large-area substrates. Meanwhile, it also retarded perovskite nucleation to promote oriented crystal growth, yielding optimized film morphology over large areas. The compact Me-4PACz anchoring and high perovskite crystallinity in decoupled codeposition strategy contributed to a certified PCE of 26.9% for small-area flexible PSCs, while certified PCEs of 22.3, 20.6, and 20.2% are achieved on flexible PSMs with substrate areas of 25, 100, and 1050 cm $^{2}$ (device areas of 14.55, 80, and 750 cm $^{2}$ ), respectively. The modules retained more than 92% of their PCE after \~1000 hours of maximum power point tracking under the ISOS-L-3 protocol (85°C, 65% relative humidity) and passed all IEC TS 63163:2021 test sequences, as certified by TÜV SÜD. The strategy also shows broad applicability to rigid substrates and different SAM molecules.

CONCLUSION: We identified the key factors by which substrate surface properties constrain the performance of flexible PSMs, and developed decoupled codeposition strategy to address the long-standing bottleneck of performance loss during flexible PSMs upscaling. The decoupling agent BSAI reduced perovskite-SAM strong interactions to facilitate compact SAM anchoring and high perovskite crystallinity, which enables high-efficiency, stable large-area flexible photovoltaic devices. This work advanced the commercial viability of flexible perovskite photovoltaics and demonstrated their practical potential in portable and mobile power supply scenarios. □

Compact and uniform Me-4PACz and perovskite film on flexible substrate through Decoupled SAM-perovskite codeposition strategy. The decoupling agent BSAI preferentially binds to the Pb-I framework, thereby reducing the Me-4PACz-perovskite interaction. This effect facilitates the quick anchoring of Me-4PACz to surface hydroxyl groups on PET/ITO substrates and ameliorates perovskite crystallization kinetics, ultimately yielding stable tridentate-anchored SAMs and uniform perovskite films.

Me-4PACz:BSAI:perovskite solution

Decoupled self-assembled monolayer and perovskite deposition for efficient flexible solar modules

Yeyong Wu $^{1\dagger}$ , Xikun Zhang $^{1\dagger}$ , Tiankai Zhang $^{2\dagger}$ , Guiying Xu $^{1*}$ , Yunxiu Shen $^{1}$ , Chengxin Guo $^{3}$ , Qing Miao $^{1}$ , Junyuan Ding $^{1}$ , Yue Yin $^{1}$ , Qinrong Cheng $^{1,3}$ , Wen Liu $^{1}$ , Xiaoxiao Wu $^{1}$ , Kaiqiang Chen $^{1}$ , Wenwen Kou $^{1}$ , Zhiyun Cao $^{1}$ , Haiyang Chen $^{1}$ , Weijie Chen $^{1}$ , Yujin Chen $^{3}$ , Feng Gao $^{4}$ , Yaowen Li $^{1,5*}$ , Yongfang Li $^{1,3,6}$

The power conversion efficiency (PCE) and stability of flexible perovskite solar modules (PSMs) are less than those of their rigid devices because surface hydroxyls desorption on flexible substrates, which disrupts self-assembled monolayer (SAM) anchoring and perovskite crystallization. We developed a decoupled SAM-perovskite codeposition strategy that accelerates SAM anchoring by reducing the SAM-perovskite interaction and ameliorates perovskite crystallization. The certified PCE with phenethylammonium iodide as the passivation molecule was 20.2% on PSMs with device areas of 750 cm $^{2}$ . With 1,3-diaminopropane dihydroiodide as the passivator, the PCE of 22.0% for a device area of 14.55 cm $^{2}$ decreased by 8% after 1044 hours under maximum power point tracking at 85°C and 65% relative humidity. The devices pass all sequences of an international commercialization assessment standard for flexible portable PSMs.

Cost-effective, highly efficient flexible perovskite solar cells (PSCs) could expand the application of solar energy to curved buildings, vehicles, airplanes and wearable electronics (1-3). Although small-area flexible PSCs have already achieved power conversion efficiencies (PCE) of $26.22\%$ (4-6), appreciable PCE and stability drops occur upon device upscaling (fig. S1 and tables S1 and S2), which results in a performance gap between large-area flexible perovskite solar modules (PSMs) and their rigid counterparts (7-13). The substantial efficiency and stability deficit poses challenges for the commercial viability of both single junction and tandem flexible PSMs.

The substantial differences in substrate properties are the primary cause of the inferior performance of large-area PSMs (14–16). Flexible plastic substrates only permit low-temperature indium tin oxide (ITO) deposition, which forms ITO layers with low crystallinity that have surface hydroxyl groups that readily desorb in polar solvents. These properties increased the difficulty of uniformly depositing large-area self-assembled monolayers (SAMs) and perovskite films compared with rigid glass substrates (17–19). Even the SAM-perovskite precursor codeposition method, which circumvents the issue of poor large-area wettability of perovskite on the SAM layer and realizes state-of-the-art efficiency on rigid perovskite modules, still fails when being applied on flexible substrates (20, 21).

The key reason for the ineffectiveness of the codeposition method on the flexible substrate lies in the dynamic competition between rapid surface hydroxyl groups desorption and delayed SAM anchoring. The surface hydroxyl groups on the flexible substrate quickly desorb, whereas the strong interaction between perovskite and phosphonic acid groups in the SAM molecules impedes the SAM anchoring process (22, 23), yielding nonuniform SAM deposition and compromised perovskite crystallinity (24–27). To address this dynamic competition, we developed a decoupled codeposition strategy by incorporating a decoupling agent, 3,5-bis(methylsulfonyl) phenylammonium iodide (BSAI), into the perovskite-SAM ([4-(3,6-dimethyl-9H-carbazol-9-yl) butyl] phosphonic acid, Me-4PACz) codeposition process. The ammonium group and two sulfonyl groups of BSAI cause it to bind strongly to perovskite colloids and disrupt the coordination between Pb ions and the phosphonic acid groups in Me-4PACz.

This dual function not only enabled effective and uniform SAM layer deposition on large-area flexible substrates with minimal interference, but also retarded nucleation to promote oriented perovskite growth, which realized optimized morphologies over areas up to $1050 \, cm^2$ . This decoupled codeposition strategy enables a record PCE of 26.9% (26.9% certified) on small-area ( $0.0628 \, cm^2$ ) flexible PSCs. Moreover, it delivered scalable flexible PSMs with record PCE of 23.1% (22.3% certified), 21.0% (20.6% certified), and 20.6% (20.2% certified) on 25, 100, and $1050 \, cm^2$ substrates (device areas of 14.55, 80, and $750 \, cm^2$ , respectively), mitigating the constraint of PCE loss as the module device size increased. Moreover, the flexible PSMs retained >92% of their initial PCEs after \~1000 hours of continuous operation under maximum power point tracking (MPPT) under ISOS-L-3 protocol [ $85^\circ C$ , 65% relative humidity (RH)] and successfully passed all test sequences defined by IEC TS 63163:2021, an international commercialization assessment standard for flexible portable photovoltaic modules, as certified by TÜV SÜD. We also applied this decoupled codeposition strategy to rigid modules, delivering rigid PSMs ( $5 \, cm \times 5 \, cm$ ) with a certified PCE of 23.8% and $T_{97} >2500$ hours aging under the same protocol. The decoupled codeposition strategy deployed high-efficiency flexible PSMs, allowing stable power supply while being integrated into portable chargers and radio-controlled unmanned aerial vehicles (RC-UAVs), validating their practical applicability and commercial viability.

---

Desorption-prone nature of surface hydroxyl oxygen on flexible substrates

Initially, we compared the performance of rigid and flexible PSMs (with a size of $5 \, cm \times 5 \, cm$ ) by applying two different common fabrication methods: sequential deposition and conventional codeposition method (detailed fabrication procedures are provided in the Experimental section). On glass/ITO substrates, PSMs fabricated with the codeposition method achieved a slightly higher average PCE than those prepared by Me-4PACz-perovskite sequential deposition (22.0% versus 21.7%), highlighting its superiority in the fabrication of large-area modules. However, on polyethylene terephthalate (PET)/ITO substrates (the electrical and optical properties are similar to the glass/ITO substrates, fig. S2), the PSMs fabricated by the conventional codeposition method and sequential deposition method both delivered a much lower average PCE (14.6 and 15.8%, respectively) compared with rigid devices fabricated with the same deposition method (22.0 and 21.7%), which highlights the intrinsic limitations of the previously reported fabrication methods for high-performance flexible PSMs (Fig. 1A, fig. S3, and table S3).

Altered surface chemical properties, especially the stability of surface hydroxyl groups, have presented a major obstacle for applying common fabrication methods on flexible substrates. X-ray photoelectron spectroscopy (XPS) revealed that the ratio of hydroxyl oxygen to lattice oxygen on PET/ITO (15%) was only half of that on glass/ITO (29%) (Fig. 1, B and C, fig. S4, and table S4) (28–30). Low-density surface hydroxyl coverage results in poor wettability (fig. S5) (31, 32). Specific surface treatment, such as plasma treatments, can temporarily increase hydroxyl group density on the surface of PET/ITO and improve the wettability at the initial stage (fig. S5). However, these surface hydroxyl groups generated by plasma treatment still quickly desorb, especially under external perturbations (28, 33).

Fig. 1. Desorption-prone nature of surface hydroxyl oxygen on flexible substrates. (A) The distribution of PCEs for PSMs on rigid (glass/ITO) and flexible (PET/ITO) substrates using sequential deposition and codeposition. (B and C) XPS O 1s spectra of glass/ITO (B) and PET/ITO (C) substrates at the initial state, after plasma treatment, after EtOH washing, and after DMF washing. (D) XPS P 2p spectra of glass/ITO and PET/ITO surfaces through sequential deposition and codeposition. (E) Schematics of hydroxyl group density, SAM anchoring, and perovskite film quality on PET/ITO through sequential deposition and codeposition.

After exposure to polar solvents, including ethanol (EtOH) and dimethylformamide (DMF)—which are commonly used in processing SAM and perovskite layers, respectively—the hydroxyl-to-lattice oxygen ratio on glass/ITO decreased only slightly (from 58 to 48% after EtOH washing and from 58 to 49% after DMF washing). By contrast, the corresponding hydroxyl-to-lattice oxygen ratio on PET/ITO decreased sharply (from 53 to 19% after EtOH washing and from 53 to 21% after DMF washing) (Fig. 1, B and C, fig. S4, and table S4) (28, 34), implying the rapid desorption of surface hydroxyl groups. We attributed the desorption-prone nature of hydroxyl groups on PET/ITO to the lower crystallinity of the ITO layer, which resulted in a higher concentration of oxygen vacancies and more electron-deficient In/Sn sites (fig. S6 and supplementary text section 1) (35). These electron-deficient In/Sn sites readily interacted with Lewis basic solvents, such as DMF and EtOH, through Lewis acid-base interactions, thereby perturbing surface hydroxyl binding and promoting hydroxyl loss (fig. S7 and supplementary text section 2) (36–38).

The desorption-prone nature of surface hydroxyl groups on PET/ITO largely affected SAM deposition in the sequential process. The P 2p signal intensity in XPS from Me-4PACz on PET/ITO was approximately half of that on glass/ITO (\~6800 versus \~3300) (Fig. 1D), indicating less efficient anchoring of Me-4PACz on the flexible substrate. In addition, the subsequent perovskite crystallization was adversely affected by the intrinsically nonpolar Me-4PACz terminal group (figs. S8 to S11), which led to a broad distribution of PCEs and reduced reproducibility of flexible PSMs fabricated by sequential deposition (Fig. 1A and fig. S3). These findings indicated that sequential deposition was not well-suited for large-area flexible PSMs (schematically illustrated in Fig. 1E).

The codeposition method mitigated the large-area wettability limitations of perovskite precursors on SAM-covered surfaces (fig. S12). However, the easy removal of surface hydroxyl groups by DMF also severely disrupted the Me-4PACz anchoring process in the codeposition method on flexible substrates. After rinsing codeposited Me-4PACz-perovskite layers with DMF (fig. S13), the P 2p signal intensity on PET/ITO was again approximately half of that on glass/ITO (\~2800 versus \~4500) (Fig. 1D), confirming reduced anchoring efficiency on flexible substrates even with codeposition. Given the strong interactions between Me-4PACz and the perovskite precursors in solution (figs. S14 to S17 and supplementary text section 3), the limited Me-4PACz anchoring would impair hole extraction and the residual Me-4PACz in precursors would perturb perovskite film growth, as evidenced by the poor film morphology observed on flexible substrates (figs. S18 to S21 and supplementary text section 4; schematically illustrated in Fig. 1E).

---

Decoupled codeposition strategy and the selection of decoupling agent

Given that surface hydroxyl groups on the PET/ITO substrates are prone to desorption, merely increasing their population cannot achieve effective SAM anchoring. Ideal SAM-perovskite codeposition requires accelerating the anchoring kinetics between SAM molecules and hydroxyl groups. We introduced a decoupling agent with a much stronger affinity to Pb-I framework in conventional codeposition method to develop a “decoupled codeposition” strategy that avoided the poor wettability issue of depositing perovskite film on SAM, especially over a large-scale. The decoupling agent preferentially bound to Pb-I framework, thereby reducing the Me-4PACz-perovskite interaction and enabling Me-4PACz quick anchoring onto PET/ITO. Once anchored, Me-4PACz could be stabilized on the ITO surface.

We designed three molecules containing Lewis base or ammonium groups that could interact with Pb-I framework through coordination and ionic bonding: 3,5-bis(methylsulfonyl)phenylamine (BSA, two -S=O groups), phenylammonium iodide (PAI, ammonium group), and 3,5-bis(methylsulfonyl)phenylammonium iodide (BSAI, integrating both functionalities) [fig. S22; detailed synthesis procedures of BSAI are in the materials and methods in the supplementary materials (SM), as candidates for decoupling agents] (39). To evaluate the decoupling efficacy of these additives, we systematically tracked the interactions of Me-4PACz in the full precursor solution upon the introduction of BSA, PAI, and BSAI, respectively. Fourier-transform infrared (FTIR) and nuclear magnetic resonance (NMR) spectroscopies (Fig. 2, A and B, fig. S23, and supplementary text section 5) jointly revealed that the bifunctional BSAI triggered the greatest release of Me-4PACz from the Pb-I framework colloids, surpassing those of BSA and PAI, and led to the most effective decoupling of Me-4PACz and perovskite precursors. Density functional theory (DFT) calculations also showed that BSAI exhibited strong binding to the precursor Pb-I framework (-1.43 eV), substantially exceeding that of Me-4PACz (-0.60 eV), through cooperative S=O…Pb coordination and N-H…I ionic interactions (fig. S24). Interaction region indicator analysis further confirmed strong attractive interactions for both binding motifs (fig. S25). Such synergistic dual strong interactions enabled BSAI to outcompete Me-4PACz for binding to the Pb-I colloids, thereby effectively producing the strongest decoupling effect.

We incorporated BSAI as the decoupling agent in the codeposition method to reduce the interaction between Me-4PACz-perovskite and promote Me-4PACz release to anchor to ITO. The adsorption and anchoring kinetics of Me-4PACz with and without the BSAI were tracked from the in situ FTIR spectra (Fig. 2, C to E, and figs. S26 and S27). At the pre stage (0 min), the as-deposited film without BSAI showed only slight shifts of the P=O and P-O-(H) peaks (1260.17 and 1107.05 cm $^{-1}$ , respectively), relative to the precursor solution (1260.74 and 1106.46 cm $^{-1}$ respectively), indicating that most Me-4PACz was associated with Pb-I framework colloids, thereby suppressing its initial adsorption onto PET/ITO (fig. S26). By contrast, the film with BSAI exhibited pronounced peak shifts from 1266.04 and 1103.08 cm $^{-1}$ to 1259.50 and 1105.60 cm $^{-1}$ , respectively, implying substantial coordination of Me-4PACz onto PET/ITO with electron-deficient In/Sn sites.

During subsequent processing, both samples exhibited three progressive stages (stages I through III). For the sample without BSAI, we observed a negligible P–O–(H) peak shift (from 1107.05 to 1106.66 cm $^{-1}$ ) within the first 9 min (stage I), which then shifted to 1104.83 cm $^{-1}$ (at a rate of $\sim0.17~cm^{-1}~min^{-1}$ ) in the next 11 min, which indicated that only a limited number of P–O–In/Sn bonds formed during this period (stage II). The peak position stabilized at 1104.64 cm $^{-1}$ after 20 min, implying that the anchoring process was nearly complete (stage III). By contrast, the sample with BSAI exhibited distinctly accelerated anchoring kinetics. The rapid shift occurred much sooner (after 5 min) and proceeded at a higher rate of $\sim0.64~cm^{-1}~min^{-1}$ , indicating more pronounced formation of P–O–In/Sn bonds (stage II). The P–O–(H) vibration signal stabilized at 1097.98 cm $^{-1}$ after 16 min, indicating that anchoring was near saturation (stage III). This greater peak shift rate revealed that the incorporation of BSAI effectively disrupted the interaction between Me-4PACz and the Pb-I framework and facilitated the rapid release and more effective anchoring of Me-4PACz on the PET/ITO surface. In addition, DFT calculations revealed a tridentate anchoring configuration, characterized by one P=O $\cdots$ In/Sn coordination interaction and two covalent P–O–In/Sn bonds, with a high binding energy ( $E_{b}$ ) of -4.650 eV (fig. S28). Overall, the combined in situ FTIR and DFT results indicated that Me-4PACz adopted a tridentate anchoring configuration on PET/ITO.

We further checked the effectiveness of Me-4PACz anchoring with the BSAI incorporation through time-resolved ex situ measurements, including XPS (Fig. 2, F to H, fig. S29, and table S5), atomic force microscopy-infrared spectroscopy (AFM-IR, fig. S30), and scanning electrochemical cell microscopy-thin-layer cyclic voltammetry (SECCM-TLCV, fig. S31) (40). All of the measurement results showed evidence of a three-stage Me-4PACz anchoring process (supplementary text section 6).

Based on these results, the anchoring process of Me-4PACz on PET/ITO during conventional and decoupled codeposition can be summarized as follows (Figs. 2, I to P): During the pre stage, the conventional codeposition process (without BSAI) exhibited strong interactions between Me-4PACz and Pb-I framework colloids and retained most Me-4PACz in the precursor solution and suppressed its adsorption onto PET/ITO, resulting in limited $\mathrm{P = O\dots In / Sn}$ coordination. The facile coordination of DMF with electron-deficient In/Sn sites perturbed the local electronic environment, weakening hydroxyl-In/Sn interactions, and promoting interactions between neighboring hydroxyl groups (Fig. 2I). By contrast, with BSAI, Me-4PACz was effectively released from the Pb-I framework colloids and rapidly adsorbed onto PET/ITO through stronger coordination interactions. Surface hydroxyl groups then interacted with the phosphonic acid moieties of adsorbed Me-4PACz, stabilizing hydroxyl-rich interfacial sites against DMF-induced perturbation (Fig. 2M).

During stage I, at relatively low temperatures, dehydration-condensation slowly occurred between P-OH groups of adsorbed Me-4PACz and surface hydroxyl groups, gradually forming In/Sn-O-P bonds in both systems. In the conventional process, a competing DMF-induced dehydration-condensation simultaneously consumed surface hydroxyl species and reformed metal-oxygen-metal bonds (Fig. 2, J and N). During stage II, at elevated temperatures, anchoring reactions become more favorable. However, in the conventional process, hydroxyl depletion suppressed efficient anchoring despite partial migration of Me-4PACz toward the ITO interface induced by perovskite crystallization (Fig. 2K). By contrast, the BSAI-containing system showed accelerated anchoring caused by effective initial adsorption (Fig. 2O). During stage III, as anchoring sites approach saturation and surface hydroxyl density decreases, the anchoring rate slows markedly in both systems (Fig. 2, L and P).

Fig. 2. Decoupled codeposition strategy and anchoring kinetics of Me-4PACz. (A) FTIR spectra of pure Me-4PACz, Me-4PACz:DMF:FAI:PbI $_{2}$ solution, Me-4PACz:DMF:FAI:PbI $_{2}$ :BSA solution, Me-4PACz:DMF:FAI:PbI $_{2}$ :PAI solution, and Me-4PACz:DMF:FAI:PbI $_{2}$ :BSAI solution. (B) ${}^{1}$ H NMR spectra of pure Me-4PACz, Me-4PACz:FAI:PbI $_{2}$ , Me-4PACz:FAI:PbI $_{2}$ :BSA, Me-4PACz:FAI:PbI $_{2}$ :PAI, and Me-4PACz:FAI:PbI $_{2}$ :BSAI samples in DMF-d $_{7}$ . (C to E) In situ FTIR spectra of samples without BSAI (C) and with BSAI (D), together with the corresponding evolution of the P–O–(H) vibrational peak position (E) on PET/ITO substrates during the annealing process. (F to H) Time-resolved ex situ XPS tracking of the anchoring kinetics. High-resolution P 2p spectra of the samples processed without BSAI (F) and with BSAI (G). Evolution of the P/In signal intensity ratio plotted as a function of annealing time (H). (I to P) Schematic illustration of the anchoring process of Me-4PACz on PET/ITO during both the conventional (I to L) and decoupled (M to P) codeposition processes.

---

SAM morphology and perovskite crystallization modulation by the decoupled codeposition

The accelerated anchoring kinetics of Me-4PACz in the decoupled code-position process enabled a uniform SAM morphology on the PET/ITO surface. Time-of-flight secondary ion mass spectrometry (ToF-SIMS) analysis revealed a more compact and vertically concentrated distribution of Me-4PACz, represented by the $\mathrm{PO}_2^-$ signal, near the substrate region in the BSAI incorporated decoupled codeposited (target) film compared with the conventionally codeposited (control) film (Fig. 3, A and B). In addition, we lifted off the perovskite layer from the substrate to directly monitor Me-4PACz signals using different codeposition methods. A more pronounced Me-4PACz signal was observed from the buried interface of perovskite in the control sample, whereas a stronger Me-4PACz signal was tracked from the ITO surface in the target sample, both of which reflect the ideal morphology of SAM in the decoupled codeposition strategy (fig. S32).

The lateral distribution of Me-4PACz was also more compact and homogeneous in the target sample. In the cross-sectional images under high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) (Fig. 3, C and D), Me-4PACz appeared as a faint, brighter layer between the darker ITO and molybdenum oxide $(\mathrm{MoO_x})$ layers caused by atomic number contrast. Compared with irregular thickness and sparse coverage (with some regions measuring $< 1\mathrm{nm}$ ) in the control samples, Me-4PACz exhibited a continuous coverage and uniform thickness of 1 to $2\mathrm{nm}$ across the whole monitored region in the target samples (41). The lateral uniform anchoring of Me-4PACz on substrate was further corroborated by AFM-IR, tracking from the uniform distribution of characteristic P–C stretching vibration signal of Me-4PACz at 14:18 cm $^{-1}$ (Fig. 3, E and F). The homogeneous Me-4PACz anchoring enhanced hole extraction and suppressed nonradiative recombination (figs. S33 to S37, table S6, and supplementary text section 7). Moreover, uniform Me-4PACz anchoring was achieved across large-area flexible substrates up to 30 cm × 35 cm, demonstrating the scalability of the decoupled codeposition strategy for large-area flexible modules (figs. S38 to S44, table S7, and supplementary text section 8).

A

C

B

D

F

G

H

|

J

Fig. 3. Perovskite crystallization and SAM release by decoupled codeposition. (A and B) ToF-SIMS depth profiles of control (A) and target (B) perovskite films. The spatial distribution of $PbI^{-}$ , $In^{-}$ , $PO_{2}^{-}$ , and $SO_{2}^{-}$ ions correspond to the distributions of perovskite, ITO, Me-4PACz, and BSAI, respectively. (C and D) Cross-sectional HAADF-STEM images of control (C) and target (D) SAMs between $MoO_{x}$ and PET/ITO. Scale bars, 10 nm. (E and F) AFM-IR mapping at $1418\ cm^{-1}$ (P–C stretching of Me-4PACz) of control-ITO (E) and target-ITO (F). Scale bars, 2 $\mu$ m. (G and H) In situ PL spectra of control (G) and target (H) perovskite films during $N_{2}$ -knife treatment and subsequent annealing. (I and J) In situ synchrotron GIWAXS of control (I) and target (J) perovskite films during annealing.

In addition to assisting robust and compact anchoring of Me-4PACz, the decoupling agent BSAI could also modulate the crystallization kinetics in the target film compared with the control film. From in situ photoluminescence (PL) spectroscopy during slot-die coating, the perovskite growth underwent three stages: First was the prenucleation stage (stage I) dominated by the $Pb_{3}I_{8}\cdot2DMF\cdot2FA$ solvent complex (42), which had a low PL intensity. Second was the intermediate phase nucleation stage (stage II) which started with a steep PL surge at $\sim780$ nm. Upon annealing, the film entered the crystal growth stage (stage III), accompanied by a rapid PL intensity rise and a slight red shift of the peak. Compared with the conventional codeposition, the PL intensity near 700 nm during stage I in the BSAI incorporated decoupled codeposition was clearly reduced, which was evidence of effective suppression of solvent complex formation, likely resulting from the strong interaction of BSAI with Pb–I frameworks. This strong interaction retarded the formation of intermediate phase, thereby delaying the onset of nucleation from 22.1 s in the control codeposition to 30.5 s (Fig. 3, G and H). In addition, the control film exhibited poorer stability under ambient conditions, undergoing gradual decomposition and a corresponding decline in PL intensity (fig. S45) compared with the target films. The retarded perovskite crystallization in the decoupled codeposition was further confirmed by the slower emergence of the absorption edge in the in situ UV–vis absorption spectra (fig. S46 and supplementary text section 9).

The retarded crystallization kinetics ameliorated the perovskite crystallinity in the decoupled codeposition strategy. From the in situ synchrotron-based grazing incidence wide-angle x-ray scattering (GIWAXS) results, the control films contained not only $\delta$ -phase but also $\alpha$ -phase crystallites with varied (001), (011), and (111) orientations during early growth, resulting from the heterogeneous nucleation. After annealing, the $\delta$ -phase and disordered $\alpha$ -phase perovskite domains remained (Fig. 3I). By contrast, the target film showed effective suppression of both the $\delta$ -phase and the $\alpha$ -phase components with (011) and (111) orientations. Instead, a pronounced and gradual growth of the $\alpha$ -phase (001) signal was observed (Fig. 3J and fig. S47), confirming a facet-oriented crystallization in the decoupled codeposition. The more detailed incident-angle dependent GIWAXS measurements further revealed that the BSAI incorporation in the decoupled codeposition facilitated consistently high crystallinity and eliminated impurities across the whole perovskite film (figs. S48 to S52 and supplementary text section 10).

With the ameliorated perovskite crystallinity, we assessed the large-scale horizontal uniformity of the target films. Perovskite layers with the size of $5 \times 5 \, cm^{2}$ , $10 \times 10 \, cm^{2}$ , and $30 \times 35 \, cm^{2}$ were slot-die coated and nine subregions (labeled I through IX) were examined through optical and morphological characterizations (figs. S8, S40, and S53). The PL intensity and absorbance of different subregions vary considerably across the control $25 \, cm^{2}$ film, and the variation became more pronounced in the larger samples ( $100 \, cm^{2}$ and $1050 \, cm^{2}$ ) (Fig. 4A and figs. S54 to S57). By contrast, the target films displayed negligible variations in PL intensity and absorbance across all areas even for the $1050 \, cm^2$ sample (Fig. 4B and figs. S54 to S57). In addition, substantial voids were observed in scanning electron microscopy (SEM) images (fig. S58) at the bottom interface of the control films caused by local nonwettability, whereas the target films showed only dense, pinhole-free microstructures even across the entire $1050 \, cm^2$ area (fig. S59). We observed similar retarded crystallization process and oriented lattice growth in perovskite with only BSAI addition, which implies that the optimization of the perovskite layer mainly came from strong interactions between the decouple agent BSAI and Pb–I framework colloids (fig. S60 and supplementary text section 11).

---

Efficiency and stability improvement with the decoupled codeposition strategy

We fabricated flexible PSCs through the BSAI incorporated decoupled codeposition strategy [with the configuration of PET/ITO/perovskite: Me-4PACz/phenethylammonium iodide (PEAI)/C $_{60}$ /2,9-dimethyl-4,7-diphenyl-1,10-phenanthroline (BCP)/copper (Cu)] that demonstrated fast transient photocurrent decay, a prolonged charge-carrier lifetime, a reduced trap density, and minimized nonradiative recombination (fig. S61 and supplementary text section 12). For a small-area (0.0628 cm $^{2}$ ) flexible device fabricated by spin-coating, a record PCE of 26.9% [26.9% certified by TÜV SÜD, 26.9% by Verband Deutscher Elektrotechniker Testing and Certification Institute (VDE), and 26.8% by China Photovoltaic Product Quality Inspection and Testing Center (CPVT)] was achieved with the decoupled codeposition strategy with 0.5 mg mL $^{-1}$

A

B

C

D

F

G

H

|

Fig. 4. Efficiency of flexible PSMs enabled by decoupled codeposition strategy. (A and B) PL intensity mapping of nine regions from control (A) and target (B) perovskite films $(30 \times 35 \, \text{cm}^{2})$ . (C) J-V curves of $0.0628 \, cm^{2}$ flexible PSCs under AM 1.5G illumination $(100 \, \text{mW cm}^{-2})$ . (D) PCE evolution versus year chart of the reported flexible PSCs. The solid and open symbols represent the certified and in-laboratory-measured values, respectively. (E to G) I-V curves of flexible modules with aperture area of $14.55 \, cm^{2}$ (E), $80 \, cm^{2}$ (F), and active area of $750 \, cm^{2}$ (G); insets show module photographs. (H) PCE of flexible PSCs as a function of device area. (I) Certified record PCE statistics of flexible modules in different device areas.

BSAI (Fig. 4, C and D, figs. S62 to S68, tables S8 to S11, supplementary text sections 13 and 14), which outperformed the PCE of 24.5% using the conventional codeposition method.

The high homogeneity of both SAM anchoring and perovskite crystallization also contributed to efficient flexible PSMs. The detailed fabrication process of the flexible PSMs is presented in figs. S69 and S70. The light beam-induced current (LBIC) mapping demonstrated that the target modules with the decoupled codeposition exhibited more homogeneous and stronger external quantum efficiency response across the whole region (fig. S71). Thus, the target mini-module $(5 \times 5 \, \text{cm}^2, \text{aperture area: } 14.55 \, \text{cm}^2)$ using the decoupled codeposition strategy achieved a PCE of 23.1% and stabilized output of 22.6% (certified PCE under reverse scan was 22.3% and certified steady-state PCE is 22.2%, Fig. 4E, figs. S72 and S73, and table S12), which was substantially higher than that of the control flexible submodules (15.3%). The PCEs of the Target flexible PSMs also exhibited superior reproducibility, as evidenced by the notably narrower device-to-device variation distribution (fig. S74). With further upscaling, we fabricated flexible PSMs with an area of $10 \times 10 \, cm^2$ (aperture area: $80 \, cm^2$ ) and $30 \times 35 \, cm^2$ (active area: $750 \, cm^2$ ) based on the decoupled codeposition strategy, and achieved PCEs of 21.0% (20.6% certified) and 20.6% (20.2% certified), respectively (Fig. 4, F and G, figs. S75 and S79, and table S12), whereas the control modules exhibited lower PCEs of only 11.5 and 7.95%. The decoupled codeposition strategy exhibited negligible PCE “scale-up efficiency loss” (Fig. 4H) and delivered record PCEs of commercial-size flexible PSMs (Fig. 4I and table S13).

We further demonstrated the general applicability of the decoupled codeposition strategy to different substrates and SAM molecules. Rigid PSMs with different areas fabricated by the decoupled codeposition achieved PCEs of 24.2% (23.8% certified, 14.55 cm $^{2}$ ), 22.1% (80 cm $^{2}$ ), and 21.6% (750 cm $^{2}$ ), respectively (figs. S80 and S81 and table S14). In addition, flexible PSMs incorporating different SAM molecules, namely [2-(3,6-Dimethoxy-9H-carbazol-9-yl)ethyl]phosphonic acid (MeO-2PACz), [4-(7H-Dibenzo[c,g]carbazol-7-yl)butyl]phosphonic acid (4PADCB), and (4-(2,7-dibromo-9,9-dimethylacridin-10(9H)-yl)butyl)phosphonic acid (DMAcPA), into the perovskite precursor delivered similar PCEs (22.5, 22.6, and 22.8%), underscoring the broad applicability of this approach with different SAMs (fig. S82 and table S15).

To assess the commercial viability of flexible PSMs, we systematically evaluated their operational stability following the standardized International Summit on Organic Solar Cells Stability (ISOS) protocols, including photothermal aging, damp heat exposure and thermal cycling. To minimize the disturbance from passivation for long-term stability evaluation of decoupled codeposition strategy, we replaced PEAI with a more robust passivation molecule 1,3-diaminopropane dihydroiodide (figs. S83 and S84, table S16). As shown in Fig. 5A and table S17, under ISOS-L-3 testing conditions (1-sun illumination at $85^{\circ}$ C with MPPT), the encapsulated target flexible PSMs (5 cm × 5 cm) retained 92.5% of its initial PCE of 22.0% after 1044 hours, whereas the control device maintains only 42.1%.

The decoupled codeposition strategy first realized stable operation of flexible PSM at $85^{\circ}$ C, yet previous stability test results were limited to $<70^{\circ}$ C. (Fig. 5E and table S18). A similar trend was observed for unencapsulated devices under continuous 1-sun illumination at $85^{\circ}$ C in a $N_{2}$ atmosphere (fig. S85). The target device retained 91.8% of its initial efficiency after 900 hours, compared with only 43.0% for the control device, which confirms that the decoupled codeposition strategy significantly improves the intrinsic operational stability of flexible modules. During the damp heat test ( $85^{\circ}$ C and 85% RH, ISOS-D-3), the target modules maintained 94.6% of their initial PCE after 1050 hours, whereas the control modules degraded to 62.9% (Fig. 5B, fig. S86A, and table S17). In the thermal cycling test between $-40^{\circ}$ C and $+85^{\circ}$ C (ISOST-3), the target modules retained 95.1% of their initial performance after 200 cycles, whereas the control modules substantially degraded, with PCE decreasing to 51.5% of the initial value (Fig. 5C, fig. S86B, and table S17).

We attributed enhanced stability to two main factors. First, the decoupling agent inhibited interaction between perovskite and Me-4PACz and promoted complete anchoring of Me-4PACz to substrate, which suppressed deleterious reactions between phosphonic acid groups and perovskites under high humidity, elevated temperature, and illumination (figs. S87 to S94, supplementary text sections 15 and 16) (43). Second, the presence of electron-deficient In/Sn surface sites on PET/ITO facilitated Me-4PACz anchoring in strong tridentate binding mode, thereby further enhancing the anchoring stability (figs. S95 to S98, table S19, and supplementary text section 17).

The mechanical durability of the flexible modules was also evaluated by cyclic bending at a radius of 10 mm (1% strain, axis $\perp$ current flow) (figs. S99 and S100, table S20, and supplementary text section 18). After 10,000 bending cycles, the target modules retained 92.1% of their initial PCE, whereas the control modules preserved only 75.5%. We attributed this enhanced mechanical durability to the strengthened SAM/perovskite interface that formed from the dense and pinhole-free bottom surface of the perovskite film, which minimized internal free surfaces and thereby suppressing crack initiation and propagation along the buried interface under cyclic shear stress (figs. S101 to S105, table S21, and supplementary text section 19).

To evaluate the operational robustness under combined complex aging conditions, our PSMs were submitted to the Technical Inspection Association TÜV SÜD, where they successfully completed all required tests in accordance with IEC TS 63163:2021 (an international commercialization assessment standard for flexible portable photovoltaic modules, figs. S106 and S107, table S17, and supplementary text section 20). Additionally, the rigid perovskite modules fabricated with the decoupled codeposition strategy retained \~97.4% PCE after 2520 hours of aging according to ISOS-L-3 protocol after encapsulation, whereas the control device PCE decayed to \~63.3% (Fig. 5D and table S17). The operational lifetime of rigid modules further validated the ability to enhance stability of decoupled codeposition strategy (Fig. 5F and table S22).

We demonstrate real-world applications of high-efficiency flexible perovskite modules in two critical scenarios: portable chargers and RC-UAVs, showcasing technology readiness near commercial maturity. The flexible PSMs used in these demonstrations weighed only 20.31 g (equivalent to four sheets of A4 paper; see movie S1). The portable PSMs generated 20.5 W of electricity under clear weather (with irradiance of $1058 \, W m^{-2}$ ) and maintained an output of 12.1 W on cloudy late afternoons (with irradiance of $635 \, W m^{-2}$ ), fully meeting charging demands of personal electronics (Fig. 5G and movie S2). Laminating flexible PSMs onto the aerodynamically designed curved wings of an RC UAV increased the total weight by $\sim8\%$ . However, it extended the flight endurance from 8 to 19 min and expanded the flight range from $\sim4$ to $\sim9.5 \, km$ . This enhancement substantially increased the flying scope of the RC UAV without compromising its flight altitude or balance (Fig. 5H, figs. S108 and S109, movie S3, and supplementary text section 21).

---

Discussion

We developed a decoupled codeposition strategy by incorporating BSAI as a tailored modulator in the conventional SAM-perovskite codeposition precursor solution. The strong interaction between BSAI and perovskite precursor impedes interaction between Me-4PACz and perovskite, facilitating rapid and efficient Me-4PACz anchoring on the PET/ITO electrode, as well as high-quality perovskite film growth, which addressed both the challenges of achieving homogeneous and compact SAM/perovskite stack layers over large areas on flexible substrates. This decoupled codeposition strategy contributed to a record PCE of 26.9% (26.9% certified by TÜV SÜD, 26.9% by VDE, and 26.8% by CPVT) for small-area flexible PSCs, while PCEs of 23.1% (22.3% certified), 21.0% (20.6% certified), and 20.6% (20.2% certified) are achieved on flexible PSMs with substrate areas of 25, 100, and 1050 cm $^{2}$ (active areas of 14.55, 80, and 750 cm $^{2}$ ), respectively. The resulting flexible PSMs (device area 14.55 cm $^{2}$ ) retained >92% of their initial efficiencies (22.0%) after \~1000 hours of MPPT at 85°C. Moreover, the commercial-scale flexible PSMs (device area 900 cm $^{2}$ ; PCE 19.9%) successfully passed all sequences of the IEC TS 63163:2021 reliability tests, as certified by TÜV SÜD. The rigid PSMs also remained >97% PCE after \~2500 hours aging under the same conditions, underscoring the universality of the decoupled codeposition strategy. Our work provides a mechanistic understanding at the molecular level of the competitive interplay between SAM anchoring and perovskite crystallization and offers an effective and feasible regulatory strategy for efficient commercial-scale flexible PSMs.

C

B

E

D

G

Fig. 5. Stability and real-world applications of flexible perovskite modules. (A) PCE evolution of encapsulated flexible PSMs (aperture area of $14.55 \, cm^2$ ) when operating under 1-sun full spectrum illumination and MPPT at $85^\circ C$ and $\sim65\%$ RH (ISOS-L-3). (B) PCE evolution of encapsulated flexible PSMs (aperture area of $14.55 \, cm^2$ ) when heated under $85^\circ C$ and $85\%$ RH (ISOS-D-3). (C) Thermal cycling aging from $-40^\circ C$ to $+85^\circ C$ under dark (ISOS-T-3). (D) PCE evolution of encapsulated rigid PSMs (aperture area of $14.55 \, cm^2$ ) when operating under 1-sun full spectrum illumination and MPPT at $85^\circ C$ and $\sim65\%$ RH (ISOS-L-3). (E) MPPT stability ( $T_{90}$ ) versus test temperature chart of the reported flexible PSCs and PSMs. (F) MPPT stability (PCE maintained at $85^\circ C$ ) versus lifetime chart of the reported rigid PSMs. (G) Portable charging performance and real-time power output of flexible PSMs from 18 to 24 November 2025; deep blue dots indicate irradiance. (H) Power output of RC-UAVs integrated with flexible PSMs from 22 to 28 November 2025; dark blue dots indicate irradiance.

The SM presents additional information supporting the conclusions of this work, including materials and methods, 21 sections of supplementary text, 109 supplementary figures, 22 supplementary tables, four dynamic movies, and additional cited references. Collectively, these contents corroborate the experimental observations and validate the mechanistic framework proposed in the main text.

---

REFERENCES AND NOTES

1. A. Cui et al., Chemical bath deposition of $SnO_{2}$ films on PEN/ITO substrates for efficient flexible perovskite solar cells. Nanotechnology 35, 375401 (2024). doi: 10.1088/1361-6528/ad568f

2. W. Cai et al., Interlayer reinforcement for improved mechanical reliability for wearable perovskite solar cells. Energy Environ. Sci. 17, 8162–8173 (2024). doi: 10.1039/D4EE03503H

3. X. Li et al., Visible light-triggered self-welding perovskite solar cells and modules. Adv. Mater. 37, e2410338 (2025). doi: 10.1002/adma.202410338; pmid: 39568249

4. S. Shi et al., Mechanically interlocked polymer scaffolds enable high-efficiency printed flexible perovskite photovoltaics. Nat. Synth. 5, 209–220 (2025). doi: 10.1038/s44160-025-00904-6

5. X. Tang et al., Enhancing the efficiency and stability of perovskite solar cells via a polymer heterointerface bridge. Nat. Photonics 19, 701–708 (2025). doi: 10.1038/s41566-025-01676-3

6. Z. Chu et al., Synergistic macroscopic–microscopic regulation: Dual constraints of the island effect and coffee-ring effect in printing efficient flexible perovskite photovoltaics. Adv. Funct. Mater. 35, 2424191 (2025). doi: 10.1002/adfm.202424191

7. Z. Chu et al., Laser annealing enables rapid, degradation-free ambient processing of perovskite solar modules. Science 390, 905–910 (2025). doi: 10.1126/science.adx9650; pmid: 41308129

8. B. Yan et al., 3D laminar flow-assisted crystallization of perovskites for square meter-sized solar modules. Science 388, eadt5001 (2025). doi: 10.1126/science.adt5001; pmid: 40403065

9. C. Gong et al., An equalized flow velocity strategy for perovskite colloidal particles in flexible perovskite solar cells. Adv. Mater. 36, e2405572 (2024). doi: 10.1002/adma.202405572; pmid: 38809575

10. H. Wang et al., Impurity-healing interface engineering for efficient perovskite submodules. Nature 634, 1091–1095 (2024). doi: 10.1038/s41586-024-08073-w; pmid: 39326517

11. D. S. Lee et al., Overcoming stability limitations of efficient, flexible perovskite solar modules. Joule 8, 1380–1393 (2024). doi: 10.1016/j.joule.2024.02.008

12. C. Wang et al., $SnO_{2}$ ink engineering for printing efficient flexible perovskite solar modules. Sci. Adv. 11, eadu1116 (2025). doi: 10.1126/sciadv.adu1116; pmid: 41134905

13. Y. Ding et al., Cation reactivity inhibits perovskite degradation in efficient and stable solar modules. Science 386, 531–538 (2024). doi: 10.1126/science.ado6619; pmid: 39480944

14. J. H. Heo et al., Chemical oxidization of PTAA enables stable slot-die-coated perovskite solar modules. Joule 9, 101850 (2025). doi: 10.1016/j.joule.2025.101850

15. S. H. Huang et al., Controlling nuclei growth and phase transition of CsFA-Based perovskite during slot-die coating via solvent engineering for high-performance solar cells and modules. Adv. Funct. Mater. 35, 2418777 (2025). doi: 10.1002/adfm.202418777

16. S. Yuan et al., Scalable preparation of perovskite films with homogeneous structure via immobilizing strategy for high-performance solar modules. Nat. Commun. 16, 2052 (2025). doi: 10.1038/s41467-025-57303-w; pmid: 40021660

17. X. Tong et al., Large orientation angle buried substrate enables efficient flexible perovskite solar cells and modules. Adv. Mater. 36, e2407032 (2024). doi: 10.1002/adma.202407032; pmid: 39049807

18. W. Xu et al., Multifunctional entinostat enhances the mechanical robustness and efficiency of flexible perovskite solar cells and minimodules. Nat. Photonics 18, 379–387 (2024). doi: 10.1038/s41566-023-01373-z

19. Y. Xu et al., Uniform coverage functional layers enable high-efficient flexible perovskite solar modules with an outstanding fill factor. Sol. RRL 7, 2300238 (2023). doi: 10.1002/solr.202300283

20. X. Zheng et al., Codeposition of hole-selective contact and absorber for improving the processability of perovskite solar cells. Nat. Energy 8, 462–472 (2023). doi: 10.1038/s41560-023-01227-6

21. D. P. Gao et al., High-efficiency perovskite solar cells enabled by suppressing intermolecular aggregation in hole-selective contacts. Nat. Photonics 19, 1070–1077 (2025). doi: 10.1038/s41566-025-01725-x

22. Y. Huang et al., High-efficiency inverted perovskite solar cells via in situ passivation directed crystallization. Adv. Mater. 36, e2408101 (2024). doi: 10.1002/adma.202408101; pmid: 39140642

23. Q. Tan et al., Inverted perovskite solar cells using dimethylacridine-based dopants. Nature 620, 545–551 (2023). doi: 10.1038/s41586-023-06207-0; pmid: 37224876

24. R. Tian et al., Nucleation regulation and mesoscopic dielectric screening in $\alpha$ -FAPbI $_{3}$ . Adv. Mater. 36, 2309998 (2024). doi: 10.1002/adma.202309998

25. M. Li et al., Orientated crystallization of FA-based perovskite via hydrogen-bonded polymer network for efficient and stable solar cells. Nat. Commun. 14, 573 (2023). doi: 10.1038/s41467-023-36224-6; pmid: 36732540

26. P. Shi et al., Oriented nucleation in formamidinium perovskite for photovoltaics. Nature 620, 323–327 (2023). doi: 10.1038/s41586-023-06208-z; pmid: 37344595

27. X. Jiang et al., Top-down induced crystallization orientation toward highly efficient p-i-n perovskite solar cells. Adv. Mater. 36, e2313524 (2024). doi: 10.1002/adma.202313524; pmid: 38453665

28. C. Luo et al., Engineering bonding sites enables uniform and robust self-assembled monolayer for stable perovskite solar cells. Nat. Mater. 24, 1265–1272 (2025). doi: 10.1038/s41563-025-02275-x; pmid: 40555860

29. E. Aydin et al., Enhanced optoelectronic coupling for perovskite/silicon tandem solar cells. Nature 623, 732–738 (2023). doi: 10.1038/s41586-023-06667-4; pmid: 37769785

30. H. Tang et al., Reinforcing self-assembly of hole transport molecules for stable inverted perovskite solar cells. Science 383, 1236–1240 (2024). doi: 10.1126/science.adj9602; pmid: 38484063

31. M. Iqbal et al., Controlled surface wettability by plasma polymer surface modification. Surfaces 2, 349–371 (2019). doi: 10.3390/surfaces2020026

32. X. Zhu et al., Uniform-saturation modification for hydrophilicity improvement of large-scale PET by plasma-electrified treatment. Eur. Polym. J. 181, 111656 (2022). doi: 10.1016/j.eurpolymj.2022.111656

33. X. Zhu, F. Li, F. Xia, Z. Zhao, W. Xue, Properties of indium tin oxide films deposited on unheated polymer substrates by ion beam assisted deposition. Thin Solid Films 517, 5395–5398 (2009). doi: 10.1016/j.tsf.2008.12.057

34. M. Wu et al., Reconstruction of the indium tin oxide surface enhances the adsorption of high-density self-assembled monolayer for perovskite/silicon tandem solar cells. Adv. Funct. Mater. 33, 2304708 (2023). doi: 10.1002/adfm.202304708

35. M. Sohail, W. Lv, Z. Mei, Recent progress in Ruthenium-based electrocatalysts for water oxidation under acidic condition. ACS Sustain. Chem.& Eng. 11, 17564–17594 (2023). doi: 10.1021/acssuschemeng.3c05415

36. J. M. Vohs, Site requirements for the adsorption and reaction of oxygenates on metal oxide surfaces. Chem. Rev. 113, 4136–4163 (2013). doi: 10.1021/cr300328u; pmid: 23181433

37. M. J. Gilkey, B. Xu, Heterogeneous catalytic transfer hydrogenation as an effective pathway in biomass upgrading. ACS Catal. 6, 1420–1436 (2016). doi: 10.1021/acscatal.5b02171

38. Z.-J. Zuo, P.-D. Han, J.-S. Hu, W. Huang, Effect of surface hydroxyls on DME and methanol adsorption over $\gamma$ -Al(2)O(3) (hkl) surfaces and solvent effects: A density functional theory study. J. Mol. Model. 18, 5107–5111 (2012). doi: 10.1007/s00894-012-1495-z; pmid: 22763911

39. H. Chen et al., Improved charge extraction in inverted perovskite solar cells with dual-site-binding ligands. Science 384, 189–193 (2024). doi: 10.1126/science.adm9474; pmid: 38603485

40. W. Wu et al., Stable and uniform self-assembled organic diradical molecules for perovskite photovoltaics. Science 389, 195–199 (2025). doi: 10.1126/science.adv4551; pmid: 40570056

41. S. M. Park et al., Low-loss contacts on textured substrates for inverted perovskite solar cells. Nature 624, 289–294 (2023). doi: 10.1038/s41586-023-06745-7; pmid: 37871614

42. M. Liu et al., Efficient and stable p-i-n perovskite solar cells enabled by insitu functional group conversion. J. Am. Chem. Soc. 146, 32105–32116 (2024). doi: 10.1021/jacs.4c13248; pmid: 39512138

43. C. Fei et al., Limiting phosphonic acid interlayer-perovskite reactivity to stabilize perovskite solar modules. Science 391, eadz7969 (2026). doi: 10.1126/science.adz7969; pmid: 41477887

44. Y. Wu et al., Source data for: Decoupled self-assembled monolayer and perovskite deposition for efficient flexible solar modules, Zenodo (2026); https://doi.org/10.5281/zenodo.22030944.

---

ACKNOWLEDGMENTS

This work was supported by the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), Collaborative Innovation Center of Suzhou Nano Science and Technology, and the Key Laboratory of Polymeric Materials Design and Synthesis for Biomedical Function, Soochow University, and Laboratory of Flexible Electronics Technology, Tsinghua University. The authors are grateful for the technical support for Nano-X from Suzhou Institute of Nano-Tech and Nano-Bionics, Chinese Academy of Sciences (SINANO). We gratefully acknowledge the support from Shanghai Synchrotron Radiation Facility (SSRF) for use of the BL14B beamline. Funding: This work was supported by the following: National Natural Science Foundation of China (grants 52325307 and 52273188 to Y.W.L.); National Natural Science Foundation of China (grant 52573207 to G.Y.X.); National Key Research and Development Program (2024YFF1401103 to Y.W.L.); Suzhou Key Core Technologies project (SYG2024146 to Y.W.L.); the Key R&D and Achievement Transformation Plan Project of Inner Mongolia Autonomous Region (2025YFHH0021 to Y.W.L.); the Science and Technology Major Project of Jiangsu Province (BG2025024 to Y.W.L.); Natural Science Foundation of the Jiangsu Higher Education Institutions of China (25KJB480012 to G.Y.X.) Author contributions: Conceptualization: Y.W.L. Methodology: Y.W.L., G.Y.X., Y.J.C., C.X.G., W.J.C., H.Y.C., and Q.R.C. Investigation: Y.W.L., G.Y.X., Y.Y.W., X.K.Z., Y.X.S., Q.M., J.Y.D., Y.Y., W.L., X.X.W., Z.Y.C., K.Q.C., and W.W.K. Visualization: Y.W.L., G.Y.X., and Y.Y.W. Funding acquisition: Y.W.L. Project administration: Y.W.L. Supervision: Y.W.L. and G.Y.X. Writing - original draft: Y.W.L., G.Y.X., Y.Y.W., and T.K.Z. Writing - review & editing: Y.W.L., G.Y.X., Y.Y.W., T.K.Z., F.G., and Y.F.L. Competing interests: Y.W.L., G.Y.X., Y.Y.W., and Y.F.L. are inventors on a patent application related to this work filed by Soochow University. The remaining authors declare no competing interests. Data, code, and materials availability: All data needed to evaluate the conclusions in the study are available in the main text or the SM. Source data and key quantitative analyses are available at Zenodo (44). License information: Copyright © 2026 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/content/page/science-licenses-journal-article-reuse

---

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.aef9745

Materials and Methods; Supplementary Text; Figs. S1 to S109; Tables S1 to S22; References (45–110); Movies S1 to S4

Submitted 30 January 2026; resubmitted 31 May 2026; accepted 28 August 2026

10.1126/science.aef9745

LET YOUR KNOWLEDGE BLOSSOM

Visit the Science Custom Publishing sites today and grow your knowledge with a wide selection of booklets, podcasts, posters, sponsored features, and webinars!

Booklets

Podcasts

Posters

Sponsored Features

Webinars

Brought to you by the Science/AAAS Custom Publishing Office.

Science

AAAS

Scan the code and start exploring the latest advances in science and technology innovation!

Science.org/custom-publishing

The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development

Harrison G. Zhang\*, Peter Eckmann, Jiacheng Miao, Andrew B. Mahon, James Zou\*

Full article and list of author affiliations: https://doi.org/10.1126/science.aeg6779

INTRODUCTION: Drug discovery requires integrating evidence across biological scales and data modalities to make a sequence of complex decisions, from target prioritization to clinical development. The relevant data, tools, and expertise remain fragmented across organizations and siloed workflows. Overcoming this fragmentation requires not just individual artificial intelligence (AI) models, but holistic systems that can organize specialized expertise, tools, and data into coordinated therapeutic decision-making.

RATIONALE: AI agents are language models equipped with tools and databases that can autonomously perform multistep analyses. We introduce the Virtual Biotech, a multi-agent organization modeled on a drug-development company, with agentic divisions spanning target discovery, safety assessment, modality selection, and clinical development. A virtual chief scientific officer interprets scientific questions, delegates analyses to domain-specialized scientist agents, and synthesizes their findings. The system provides access to more than 100 specialized tools covering 78,726 targets and 39,530 diseases, enabling in-depth and nuanced primary data analysis. A scientific reviewer critiques analyses before evidence synthesis, and analysis artifacts are retained for transparency and reproducibility. We applied the Virtual Biotech to three early-development decision points: target prioritization, modality selection, and analysis of clinical translation failures.

RESULTS: First, the Virtual Biotech autonomously annotated and analyzed clinical trial outcomes to identify transcriptomic features of drug targets associated with trial success. More than 37,000 agents curated structured trial outcomes. Across 55,984 trials, the Virtual Biotech found that drugs targeting cell type-specific genes were $40\%$ more likely to progress from phase 1 to phase 2 and $48\%$ more likely

A multi-agent AI system for therapeutic research. The Virtual Biotech is coordinated by a virtual chief scientific officer (center) that directs domain-specialized AI agents to perform analyses across early drug discovery. In three case studies, the Virtual Biotech conducted multistep investigations spanning target prioritization, target validation and modality selection, and analysis of clinical translation failures.

to reach market, while exhibiting 32% lower adverse event rates. These associations persisted after accounting for genetic evidence, suggesting that single-cell features provide information complementary to genetics for target prioritization. Second, the Virtual Biotech evaluated B7-H3 as a lung cancer target, integrating statistical genetics, single-cell, spatial, and clinicogenomic data. The agents localized B7-H3 overexpression to cancer-associated fibroblasts, linked B7-H3-high regions to immune exclusion and worse survival, and proposed an antibody-drug conjugate strategy. Third, the system analyzed a terminated ulcerative colitis trial targeting oncostatin M receptor (OSMR). Longitudinal single-cell analyses suggested that OSMR was not the dominant driver of the pathway the antibody was intended to block because related receptors converge on the same signaling pathway. Across all case studies, the Virtual Biotech outperformed three state-of-the-art agentic systems, producing more accurate trial annotations and more biologically meaningful conclusions from diverse data modalities. Human experts reimplemented core analyses, confirming the agent conclusions and code correctness.

CONCLUSION: The Virtual Biotech demonstrates that an organized system of AI agents can carry out integrative and multimodal reasoning for early therapeutic development at a scale difficult for human-only teams. By systematically bringing diverse evidence to bear on therapeutic decisions, while keeping human scientists in the loop, it can help identify opportunities and support more informed development choices. □

\*Corresponding author. Email: hgzhang@stanford.edu (H.G.Z.); jamesz@stanford.edu (J.Z.) Cite this article as H. G. Zhang et al., Science 394, eaeg6779 (2026). DOI: 10.1126/science.aeg6779

Target Prioritization

Target Validation & Modality Selection

Clinical Translation

Failure Analysis

ARTIFICIAL INTELLIGENCE

The Virtual Biotech: A multi-agent AI framework for therapeutic discovery and development

Harrison G. Zhang $^{1*}$ , Peter Eckmann $^{2}$ , Jiacheng Miao $^{1,3}$ , Andrew B. Mahon $^{4}$ , James Zou $^{1,2*}$

Drug development requires evidence integration across biological scales and modalities, but relevant tools are fragmented. We introduce the Virtual Biotech, an organization of artificial intelligence (AI) agents modeled on a drug-development company, with agentic divisions spanning target discovery, safety assessment, modality selection, and clinical development. We demonstrate its utility at three drug-development decision points. First, more than 37,000 agents annotated outcomes from 55,984 trials and found that drugs targeting cell type-specific genes were 48% more likely to reach market with 32% fewer adverse events. Second, it integrated multimodal evidence to propose a therapeutic strategy in lung cancer. Third, it analyzed a terminated ulcerative colitis trial and inferred potential mechanisms of failure. These results demonstrate that human-guided multi-agent systems can conduct transparent, multiscale analyses to inform therapeutic-development decisions.

Drug discovery is one of the most consequential and challenging endeavors in biomedicine. Bringing a therapeutic from concept to clinic often requires more than a decade of development and costs on the order of billions of dollars (1-3). Despite these investments, \~90% of drug candidates entering phase 1 trials never reach approval, with failures most often driven by insufficient efficacy and safety liabilities (4-7). To reduce this attrition, modern drug discovery increasingly draws on evidence from diverse sources to validate targets, including but not limited to human genetics, functional genomics, single-cell and spatial profiling, structural biology, medicinal chemistry, and clinical medicine (8).

Organizational and scientific bottlenecks have limited the proper use of this evidence in practice. The expertise to collect and analyze it is fragmented across disciplines, formats, and biological scales, producing siloed workflows (3, 9, 10). Additionally, weighing disparate and sometimes conflicting evidence is difficult, particularly when organizational dynamics or differing prior beliefs shape decisions (3, 9). Fragmented teams also document their workflows differently, making findings hard to audit and decisions hard to reconstruct (3, 11, 12). Further, the recent explosion in the scale of biological data renders manual integration across modalities more difficult (3, 9). Thus, despite advances in data-driven approaches, early-stage drug development remains slow, resource intensive, and prone to failure (1-7).

Artificial intelligence (AI) agents have been proposed to accelerate scientific tasks that involve extensive data analysis, as they are language models equipped with tools and databases that enable them to autonomously carry out complex multistep tasks. Specifying a role and providing domain-specific tools, such as bioinformatics software, specializes an agent for a class of tasks. Recent work also explores multi-agent systems, where multiple specialized agents collaborate to solve complex problems (13).

Here, we introduce the Virtual Biotech, a multi-agent AI research platform that retrieves, analyzes, and integrates multimodal evidence to answer scientific questions in early-stage human therapeutic discovery. A virtual chief scientific officer (CSO) orchestrator breaks down queries into manageable subtasks, directs them to specialized AI scientist agents, and reasons about their results. This structured collaboration encourages deep data-driven analyses of primary evidence and allows for weighing of disparate sources of evidence using a unified reasoning layer to arrive at nuanced conclusions.

Agentic systems have already shown promise in scientific domains such as chemistry and AI research automation (14–17). In biomedicine, agentic methods tailored to specific workflows have emerged, with applications such as gene editing, spatial and single-cell omics, rare disease diagnosis, and nanobody design (18–22). Recent work expanded the domain of agentic reasoning by incorporating a wider range of tools, databases, and code execution workflows for general-purpose data analysis (23–28). Nonetheless, these frameworks largely concentrate on biological data analysis and are typically governed by a single agent. In comparison, the Virtual Biotech is a multi-agent system explicitly designed for translational reasoning across the entire human therapeutic R&D pipeline. By coordinating analysis over the large scale and complexity of drug discovery data, the system can answer scientific questions with holistic, data-driven reasoning.

We first describe the Virtual Biotech tool environment and multi-agent orchestration. We then present three translational case studies that demonstrate how the Virtual Biotech can generate evidence for key decision points in early therapeutic discovery and development. The first case study focuses on large-scale target prioritization by connecting target-level molecular characteristics to clinical-trial outcomes. The second examines drug target B7-H3 in lung cancer from the angle of target validation and modality selection. The third investigates a discontinued oncostatin M receptor (OSMR) ulcerative colitis trial to infer possible reasons for clinical translation failure and to outline biomarker-driven strategies for clinical development.

---

Results

---

Overview of the Virtual Biotech

The Virtual Biotech emulates a cross-functional therapeutic research organization. It consists of four divisions of scientist agents that address key steps of the drug discovery pipeline (Fig. 1A and table S1). Domain-specialized scientist agents in each division have complementary expertise and domain-specific tools that allow them to query, analyze, and integrate biomedical information across the translational spectrum for evidence-based assessments of drug discovery questions (see Materials and methods for details).

We created domain-specific model context protocol (MCP) servers to provide a unified programmatic interface to major biomedical databases and knowledge sources, abstracting the complexity of heterogeneous data formats and query APIs (application programming interfaces) into standardized tools (Fig. 1B and supplementary text I) (29–31). Each scientist agent accesses curated, expertise-aligned MCP tools to gather evidence from relevant sources, transforming static databases into interactive knowledge layers that agents dynamically analyze in response to user questions.

We developed specialized MCPs that provide Virtual Biotech scientist agents with >100 analytical tools to interrogate large-scale, heterogeneous biomedical data spanning 78,726 targets, 39,530 diseases, 14.5 million protein–protein interactions, >3 million credible genetic sets, and >100 million single-cell profiles (32, 33). The platform further integrates chemoinformatics and pharmacology resources covering 18,119 drugs, 6332 distinct mechanisms of action, 114,270 US Food and Drug Administration (FDA) adverse event reports,

A

Fig. 1. Overview of the Virtual Biotech AI agents, tools, and orchestration procedure. (A) The multi-agent AI system organization and scientific divisions. (B) Overview of model context protocol (MCP) servers, grouped by domain area, which provide customized tools to query and analyze diverse biological and clinical data sources. (C) The workflow of the Virtual Biotech. In response to a user query, the virtual CSO first asks any relevant follow-up questions to clarify user intent before doing expensive analyses. In parallel, the chief of staff agent prepares a briefing on the topic with recent developments, data availability, and general field context. The virtual CSO then coordinates divisions of specialized scientist agents to perform analyses to answer the user's query. Next, the scientific reviewer agent evaluates the methodology and claims from scientist agents, and upon reviewing the feedback, the CSO either coordinates revised analyses with scientist agents or synthesizes the evidence into a report.

32,783 pharmacogenomic variants, and drug perturbation expression measurements from Tahoe-100M (34). In addition, ClinicalTrials.gov records are accessible through our MCP (35). Scientist agents can further query biomedical literature through PubMed and retrieve additional targeted datasets via web-based tools. These resources are largely human-focused and thus support target-centered queries in human diseases represented by the data.

We developed a user interface for the Virtual Biotech that displays the agents' reasoning process, indicates which tools are currently in use, and allows users to download the corresponding data, reproducible code, and reports generated during each session (fig. S1).

---

Virtual Biotech orchestration

Upon receiving a user's query and before initiating scientist work, the virtual CSO initiates its own strategic orientation (Fig. 1C and supplementary text II). Its chief of staff agent first generates a briefing on general field awareness, data landscape, and recent developments, to help the virtual CSO frame the analysis, set proper feasibility expectations for the user, and identify key questions. In parallel, the virtual CSO interviews the user to clarify the intent of the user and increase the alignment of human-AI collaboration (36, 37). Once briefed, the virtual CSO routes tasks to relevant scientist agents (Fig. 1C).

After the scientist agents complete their analyses, the scientific reviewer agent evaluates their outputs to understand how well they address the user's original question, the strength of the evidence supporting their conclusions, and the overall thoroughness of the analysis (Fig. 1C and supplementary text III). If gaps or unsubstantiated claims are identified by the scientific reviewer agent, the CSO redelegates to relevant scientist agents for iterative refinement.

---

Target prioritization: Virtual Biotech connects preclinical biology with clinical outcomes to identify drug target features associated with trial success

Drug discovery programs face substantial attrition, with lack of efficacy and safety concerns accounting for \~79% of clinical trial failures $(38)$ . Prior target prioritization work established that targets supported by human genetic evidence are much more likely to progress through clinical development $(39-41)$ . To expand on these results for better target prioritization, we provided Virtual Biotech with this earlier work and instructed the system to critically evaluate its own infrastructure and recommend directions to expand this analysis $(40)$ . The virtual CSO used its clinical trialist agent to enrich the 55,984-trial Open Targets clinical trial dataset with additional outcome annotations. It also proposed using its single-cell agent to design genomic features for association testing with trial outcomes, following Razuvayevskaya et al. $(33, 40, 42)$ . Agent-based outcome annotation was necessary because the original Open Targets dataset lacked endpoint data and ClinicalTrials.gov records often provide incomplete, inconsistent, or free-text-only outcomes $(43, 44)$ . Agent curation was therefore needed to standardize outcomes and recover missing or conflicting data from sources such as publications and press releases.

Massively parallel, source-tracked extraction of trial outcomes: The clinical trialist agent proposed key outcome variables for large-scale extraction, including trial-phase progression, primary and secondary endpoint results, and adverse event rates. It developed a systematic protocol to obtain trial results from sources such as publications and press releases. The workflow began with a human user instructing the virtual CSO to curate clinical trial outcomes for phase 2 and 3 trials in the Open Targets dataset. The virtual CSO translated this goal into a distributed extraction task and delegated protocol design to a clinical trialist agent, which specified the required outcome fields, evidence hierarchy, and source-tracking requirements. The virtual CSO then dispatched 37,075 clinical trialist agents in parallel, with one agent assigned to each National Clinical Trial identifier (NCT ID) (Fig. 2A). At this scale, comprehensive curation is impractical without parallelization, necessitating a multi-agent architecture capable of executing data curation in a distributed manner.

Each clinical trialist agent followed a three-tier evidence cascade for its assigned NCT ID: It first systematically queried the ClinicalTrials.gov database, then searched PubMed for publications containing trial results when primary data were incomplete, and finally consulted online press releases and regulatory announcements (Fig. 2A and supplementary text IV) (35). Agents recorded the source of information for each annotation. The median per-trial cost was \$0.23 in Anthropic API fees. For phase 1 trials, success was algorithmically defined as progression to phase 2, because phase 1 trials assess safety and dosing rather than efficacy.

To assess annotation quality, manual review of 100 randomly selected clinical trials was performed and agreement with agent annotations was 88.4% for primary endpoints (76 out of 86 applicable trials), 88.4% for secondary endpoints (76 out of 86), and 92.4% for adverse event rates (61 out of 66 reporting exact statistics). Furthermore, agreement was 85.6% for primary endpoints across 7666 trials overlapping with the manually curated Therapeutic Data Commons (TDC) Trial Outcome Prediction dataset (see Materials and methods for more details) $(45)$ . These agreement metrics demonstrated that the agent annotated trials reliably and that the resulting labels were suitable for analyses.

Translational feature engineering with single-cell atlases: To derive single-cell expression features of drug targets, the single-cell agent proposed and implemented a comprehensive feature extraction framework for Tabula Sapiens, a healthy human reference atlas of 27 tissues (46). The agent designed features organized into two categories: cell-type specificity metrics and expression heterogeneity measures. For specificity, the agent proposed the tau cell-type specificity index, where values approaching zero indicate ubiquitous expression whereas values approaching one indicate expression restricted to a single cell type (47). To capture expression heterogeneity within cell populations, the agent proposed a bimodality coefficient calculated from the skewness and kurtosis of expression distributions among expressing cells, where higher values may indicate “on/off” expression patterns within a tissue (48, 49). Various measures of bimodality have been proposed for gene expression analysis, but this specific bimodality coefficient, originally developed in psychology and social sciences, is an underexplored and interpretable metric for gene expression bimodality (50–55). These two features were moderately correlated (Pearson $\rho = 0.54$ ), suggesting that they provide some complementary information.

Virtual Biotech–driven analysis reveals genomic features associated with trial success: The final integrated dataset comprised 55,984 clinical trials spanning phase 1 (n = 14,237), phase 2 (n = 22,164), phase 3 (n = 14,911), and phase 4 (n = 4672) (Fig. 2B). Trial-level outcome annotations were linked to target-level molecular features extracted by the Virtual Biotech and to genetic evidence scores from Open Targets.

The clinical trialist agent first successfully replicated the genetic evidence results reported by Razuvayevskaya et al., finding that genetic support is associated with trial progression and also protective of early trial stoppage (fig. S2A) (40). Using the additional trial outcomes curated by the Virtual Biotech, it also found that trials that achieved primary endpoints and showed lower adverse events were enriched with genetic evidence (fig. S2, A and B). Following the approach from the analogous study on genetic evidence, the clinical trialist agent then evaluated the association between single-cell transcriptomic features of drug targets and trial outcomes (39, 40).

A

B

Fig. 2. Large-scale clinical trial data extraction using the Virtual Biotech. (A) Overview of the data extraction and curation process. For each National Clinical Trial (NCT) ID, the clinical trialist agent uses tool calls to ClinicalTrials.gov, PubMed, and the web to gather information about the trial. After reasoning about its findings, the agent enters information into a structured JSON ontology for downstream analysis. (B) Sankey plot of the curated clinical trial dataset (n = 55,984) showing trial phase, trial status, and outcomes.

Both tau cell-type specificity and expression bimodality score were associated with trial progression through clinical development phases (Fig. 3A). Following binarization of targets based on their tau metric into cell type-specific versus broadly expressed classes (see Materials and methods for details), the clinical trialist agent found that drugs acting on cell type-specific targets were 48% more likely to ever reach phase 4.

Targets with higher cell-type specificity were more likely to meet defined primary endpoints [odds ratio (OR) = 1.12; 95% confidence interval (CI): 1.09–1.14], secondary endpoints (OR = 1.12; 95% CI: 1.09–1.14), and successful phase 1 to phase 2 progression (OR = 1.27; 95% CI: 1.22–1.33) (Fig. 3A). Drugs with cell type-specific targets were 40% more likely to progress from phase 1 to phase 2. Similarly, targets with a higher expression bimodality score were more likely to achieve a positive primary endpoint, a positive secondary endpoint, and successful phase 1 to phase 2 progression (Fig. 3A).

The Virtual Biotech further found that trials with targets that have a higher cell-type specificity and expression bimodality score reported lower adverse event rates across many organ systems (Fig. 3B). Trials testing cell type-specific targets had on average $32\%$ lower adverse event rates. This suggested that targets with more heterogeneous and cell type-restricted expression patterns may offer improved safety profiles.

Next, we demonstrated that these findings are robust to confounders. To assess whether the observed associations could arise by chance or through confounding in the complex trial-to-target mapping, we performed permutation testing >1000 iterations, confirming that all findings remained statistically significant (permutation P < 0.05) against the empirical permuted null (fig. S3, A and B). To control for potential confounding between disease indications and between drug modalities, we further estimated ORs using mixed-effects models adjusted for trial phase and enrollment year, with drug modality and disease indication area as random effects. The adjusted odds ratios for cell-type specificity and expression bimodality score remained statistically significant across mixed-effects models (fig. S3, C and D).

Another key question is whether single-cell transcriptomic features of targets provide independent value beyond established human genetic evidence linking targets to disease. Therefore, we replicated our findings adjusting for genetic evidence. After adjustment, the associations between trial outcomes and each of the two single-cell features—cell-type specificity and expression bimodality score—remained statistically significant with negligible reductions in effect sizes (fig. S4, A and B). We further estimated ORs for single-cell features among trials with no genetic evidence (75% of trials) and found that single-cell features had similar or even stronger effect sizes in this subset (fig. S4, C and D). Together, these results indicated that single-cell information captures additional biological signal associated with clinical trial outcomes beyond genetic evidence alone. A recent report by Dann et al. similarly found that cell type-restricted targets were associated with phase progression but did not study association with endpoints or adverse events $(56)$ .

---

Target validation and modality selection: Virtual Biotech research builds preclinical rationale for B7-H3 antibody-drug conjugate development in lung cancers

A

We next investigated a common drug program-level decision point: whether multimodal evidence can support a coherent therapeutic hypothesis and modality choice for a nominated target in a given disease context. To demonstrate the Virtual Biotech's capacity to rapidly generate rigorous evidence to support this task, we prompted the system to assess the therapeutic potential of B7-H3 (also known as CD276) in lung adenocarcinoma (LUAD) and small cell lung cancer (SCLC). B7-H3 is a transmembrane immune checkpoint protein that is highly expressed in solid tumors, including lung cancers $57$ . Its receptor is unknown, and its role in cancer remains unclear, though it likely has both immunologic and nonimmunologic functions $58$ . The user's initial query was: "Evaluate the therapeutic candidacy of B7-H3 as a target in lung cancer." The virtual CSO interpreted this as a therapeutic target evaluation task and decomposed it into a set of evidence-generation modules spanning human genetics, target biology, single-cell transcriptomics, spatial transcriptomics, biomarker-stratified survival analyses, and prior clinical trial data.

The statistical genetics agent first evaluated germline genetic support for B7-H3 as a lung cancer target. No significant associations or credible sets were identified at the B7-H3 locus for lung cancer, and the agent concluded that germline genetic evidence for B7-H3 as a lung cancer target was weak (Fig. 4A). Despite this, the virtual CSO correctly interpreted the absence of genome-wide association study (GWAS) signal as nondisqualifying, noting that for potential immune checkpoint targets, therapeutic rationale can derive from somatic tumor overexpression rather than germline risk variants. This reasoning distinguished B7-H3 from targets where genetic evidence is essential for

B

Fig. 3. Preclinical single-cell atlas features of drug targets significantly associated with key clinical trial outcomes. (A) Associations between single cell-derived features of the drug target and phase progression, trial stoppage, and efficacy outcomes. Plot shows the standardized odds ratio (all features normalized to have a standard deviation of 1) with 95% CIs. (B) Associations between single cell-derived features and serious adverse event rates grouped by organ system. Forest plot shows beta regression coefficients with 95% CIs.

A

---

"Evaluating B7-H3 (CD276) therapeutic potential in lung cancer...

B

\- No B7-H3 variants or credible sets identified across four population-level lung cancer GWAS.

\- No machine learning locus-to-gene causal predictions identified.

\- Genetic constraint showed predicted LoF observed/expected ratio of 0.67, indicating moderate LoF tolerance.

\- ENCODE E2G identified high-confidence regulatory regions at the B7-H3 locus in normal lung, supporting strong regulatory activity.

\- Limited germline evidence does NOT invalidate therapeutic potential. Rationale may derive from somatic tumor overexpression rather than inherited risk.

Proceeding to expression-based analyses."

C

"Tumor-enriched B7-H3 expression for both LUAD and SCLC is strongest in cancer-associated fibroblasts (CAFs), not malignant cells.

A next step is understanding how cell-cell communication is disrupted and what signals B7-H3+ CAFs send to immune cells. This bridges the gap between knowing which cells express B7-H3 and understanding their function."

E

"Analysis of Visium LUAD spatial transcriptomics suggests that B7-H3 expression contributes to a immune-excluded niche. Mixed-effects modeling showed reduced immune cell abundance around high B7-H3 expression spots relative to low B7-H3 spots. This provides spatial validation for the immunosuppressive ligand-receptor interactions identified.

I want to understand whether this immunosuppressive phenotype translates into clinically meaningful survival differences."

validation, prompting the virtual CSO to instruct its single-cell agent to perform analyses for more clarity.

The single-cell agent queried and downloaded single-cell atlases from CELLxGENE Census, comprising 62,341 cells from 9 donors in the SCLC atlas, 337,002 cells from 69 donors in the LUAD atlas, and 86,478 cells from lung tissue of 26 healthy donors (Fig. 4B) (59). The agent conducted cell type-resolved differential expression analysis with pseudobulk aggregation, revealing that B7-H3 up-regulation was not uniform across all cell types but concentrated in fibroblasts in both SCLC $\left[\log_{2}\text{ fold change (log2FC)} = 2.13, \text{ false discovery rate (FDR) adjusted P value} = 3.5\ 10^{-9}\right)$ and LUAD ( $\log_{2}FC = 1.79, FDR P value = 2.0 \times 10^{-8}$ ) (Fig. 4C).

F

"Querying CELLxGENE Census for LUAD, SCLC, and healthy lung atlases to characterize B7-H3 expression at single cell resolution. Performing QC, correcting for batch effects with Harmony, visualizing atlases with UMAP."

"B7-H3+ fibroblasts show a immunosuppressive program.

Let's analyze spatial transcriptomics next to validate whether these predicted immunosuppressive interactions translate into a physical immune exclusion phenotype."

"B7-H3 seems to be a potential prognostic biomarker in NSCLC. With the spatial evidence of immune exclusion, these results support the hypothesis that B7-H3-mediated immunosuppression contributes to adverse clinical outcomes—providing a mechanistic and clinical rationale for therapeutic targeting of B7-H3 in lung cancer"

Fig. 4. Virtual Biotech evaluates B7-H3 (CD276) as a therapeutic target in lung cancer. (A) Summary from statistical genetics agent of germline genetic findings for B7-H3 in lung cancer. (B) UMAP of lung adenocarcinoma (LUAD) and small cell lung cancer (SCLC) single-cell atlases prepared by the single-cell atlas agent. (C) Differential expression analysis of B7-H3 by cell type identified significantly increased expression in fibroblasts. (D) Ensemble cell-cell communication inference by the single-cell agent revealed distinct signaling pathways found in B7-H3+ cancer-associated fibroblasts in SCLC (shown here) and LUAD (fig. S5). (E) Spatial transcriptomic analyses of LUAD primary-site tissue identifies immune-excluded niches in the top quartile of B7-H3-expressing spots relative to the bottom quartile of B7-H3-expressing spots. Bar plot quantifies the percent change (95% bootstrap CI over 10,000 resamples) in neighbor immune abundance between top and bottom quartile of B7-H3-expressing spots across 12 LUAD tumor samples. Statistical significance was assessed using a linear mixed-effects model adjusted for focal-spot total UMI count, fibroblast abundance, malignant epithelial abundance, and endothelial abundance as fixed effects with nested random effects for patient and sample within patient ( $P < 0.01$ , $*P < 0.001$ ). (F) Survival analysis of the lung adenocarcinoma cohort from The Cancer Genome Atlas project identifies B7-H3 as a potential prognostic biomarker of disease-free survival.

On the basis of these findings, the single-cell agent hypothesized that B7-H3-high fibroblast cells might actively shape an immunosuppressive tumor microenvironment (TME). It then performed a ligand-receptor interaction analysis to identify differences in signaling between fibroblasts whose B7-H3 expression fell within the highest versus lowest quartiles of the cell population (60). This identified 180 significant interactions in SCLC and 226 interactions in LUAD (see Materials and methods for details) that were specific to B7-H3-high fibroblasts communicating with immune cells (Fig. 4D and fig. S5A). The single-cell agent identified five main signaling groups: sheddases that release activating ligands from immune cell surfaces, myeloid-reprogramming mediators that polarize macrophages toward immunosuppressive states, proangiogenic factors that remodel tumor vasculature, immunosuppressive cytokines that engage inhibitory receptors, and chemokine pathways that retain immune cells in the fibroblast niche. These results suggest that B7-H3-high stromal cells may use additional immunomodulatory mechanisms beyond B7-H3's canonical checkpoint function.

The scientific reviewer agent was then dispatched and noted that the ligand-receptor analysis inferred cell-cell communication from dissociated single-cell data that lost spatial tissue context. Thus, the virtual CSO reasoned that spatial profiling was needed to verify whether these predicted immunosuppressive interactions corresponded to a physical immune exclusion phenotype. It then dispatched the single-cell agent to analyze a Visium dataset comprising 12 lung tissue samples from patients diagnosed with LUAD, totaling over 25,000 spatially resolved spots (61). To infer spot-level cell-type abundances, the single-cell agent applied Cell2Location using the prior LUAD single-cell atlas as reference (fig. S5B) (62). It then implemented a mixed-effects regression model to assess whether spots expressing B7-H3 in the top quartile showed altered immune cell compositions within their immediate spatial neighborhood relative to the bottom quartile. B7-H3-high spots exhibited lower abundance of neighboring immune cells such as T cells, macrophages, monocytes, and dendritic cells (Fig. 4E). Notably, these effects showed spatial attenuation as depletion was strongest in the immediate neighborhood (first hexagonal ring around Visium spot) and progressively weakened at the second and third rings, consistent with a local exclusion mechanism. This demonstrated that B7-H3 expression spatially correlated with an immune-excluded TME, supporting the hypothesis of an immunosuppressive phenotype.

Next, the virtual CSO wanted to understand whether this immunosuppressive phenotype corresponded to differences in patient outcomes. The virtual CSO tasked the clinical trialist agent with performing a biomarker-stratified survival analysis on the TCGA PanCancer Atlas LUAD cohort, which included 477 patients with gene expression measurements and detailed survival outcomes (63, 64). To maximize separation between biologically distinct groups, the clinical trialist agent stratified patients by B7-H3 expression quartiles, comparing patients in the top and bottom quartiles on the basis of their B7-H3 expression (Fig. 4F). In a Cox proportional hazard model adjusted for age, cancer stage, and sex, elevated B7-H3 expression was significantly associated with worse overall survival [hazard ratio (HR) = 1.62; 95% CI: 1.05–2.49; P = 0.028] and worse disease-free survival (HR = 2.06; 95% CI: 1.08–3.90; P = 0.027). Together with spatial evidence of immune exclusion, the virtual CSO suggested that B7-H3-mediated immunosuppression could drive poor clinical outcomes, providing mechanistic and clinical rationale for targeting B7-H3 in lung cancer.

Finally, the virtual CSO dispatched the target biologist agent to determine the optimal modality. The agent retrieved Human Protein Atlas data confirming B7-H3 cell surface expression with subcellular vesicle localization (65). Together with the single-cell finding that

B7-H3 is overexpressed predominantly in TME fibroblasts rather than only on malignant cells, this led the virtual CSO to nominate an antibody-drug conjugate (ADC). The CSO reasoned that an ADC could act through three mechanisms consistent with these findings: cytotoxic payload delivery to B7-H3-expressing cancer cells, TME alteration through inhibition of B7-H3-mediated immunosuppressive checkpoint interactions, and bystander killing of adjacent B7-H3-negative cancer cells. Small-molecule and proteolysis targeting chimera (PROTAC) approaches were deprioritized after the pharmacologist agent's tractability analysis identified no existing chemical binders.

This case study was conducted without web search to prevent information leakage, and the underlying language model had a knowledge cutoff of January 2025. Of note, existing literature emphasized B7-H3's expression on cancer cells and canonical checkpoint-associated functions, motivating antibody- and ADC-based programs largely through a cancer cell-centric lens (57, 58). By contrast, the Virtual Biotech integrated evidence to nominate a less explored but mechanistically coherent hypothesis: that B7-H3 is more strongly overexpressed in cancer-associated fibroblasts, spatially associates with immune exclusion, and stratifies poor clinical outcomes, an interpretation that is not directly identified from any single dataset or publication. In August 2025, the FDA granted Breakthrough Therapy Designation to ifinatamab deruxtecan, a first-in-class B7-H3-targeted ADC (66–68). As the Virtual Biotech did not have access to this information, the regulatory announcement provided external support for the system's mechanism-driven therapeutic design.

This systematic and data-driven analysis was completed for a total cost of \$50.00 in Anthropic API credits. It showed how the Virtual Biotech can accelerate target validation and modality selection, a typically manual process performed by cross-functional drug discovery teams. The full conversation transcript is in supplementary text V.

---

Clinical translation failure analysis: Virtual Biotech identifies possible failure mechanisms for terminated phase 2 trial testing first-in-class monoclonal antibody (mAb) against OSMR in ulcerative colitis

We subsequently used the Virtual Biotech to analyze cases of late-stage clinical development failure. We focused on the phase 2 MOONGLOW study (NCT06137183), a trial in ulcerative colitis (UC) evaluating the oncostatin M receptor subunit OSMRβ as a therapeutic target (69). OSMRβ is a cytokine receptor subunit that heterodimerizes with gp130 to mediate oncostatin M signaling pathways implicated in inflammation and fibrosis such as JAK-STAT (70, 71). Prior research demonstrated that elevated mucosal OSMR expression appears to mark a treatment-refractory UC inflammatory state, and mucosal OSMR expression has been independently associated with anti-TNF (tumor necrosis factor) nonresponse (71–73).

Vixarelimab is a first-in-class mAb engineered to block OSMRβ-mediated signaling (69). In MOONGLOW, adults with moderate-to-severe UC were randomized to receive vixarelimab or placebo. However, the trial was terminated in June 2025 because an interim futility analysis indicated that the primary endpoint was unlikely to be achieved (69). Accordingly, the initial query to the virtual CSO was: "We want to understand why trial NCT06137183 was terminated. This trial tested a first-in-class monoclonal antibody against OSMR in patients with UC. The results have been posted, and the trial was terminated due to interim futility analysis, which showed lack of efficacy. This termination was not related to safety concerns." The virtual CSO interpreted this as a clinical translation failure analysis and decomposed the query into subtasks focused on trial protocol and eligibility review, mechanistic assessment of OSMRβ signaling, pathway redundancy analysis in treatment-refractory gut tissue, and cross-disease exploration of OSMR-associated stromal activation programs.

The clinical trialist agent first analyzed trial design using the ClinicalTrials.gov record, which included the full study protocol (Protocol GA44839, Version 2 PDF) (69). It focused on the inclusion criteria, noting that the trial selected for a treatment-refractory study population of patients with UC who showed poor response to conventional or advanced therapies (Fig. 5A) (69). Because the drug target OSMRβ signals through the JAK-STAT pathway, the agent reasoned that it was important to understand whether OSMRβ serves as a primary driver of JAK-STAT signaling in treatment-refractory UC biology or if it instead represented a redundant point of entry into the same signaling cascade.

To answer this question, the single-cell agent then independently identified the TAURUS dataset (435,857 cells from 20 patients with UC) to investigate MOONGLOW failure mechanisms because it is a single-cell transcriptomic atlas of UC gut biopsies collected before and after anti-TNF therapy. The agent reasoned that this dataset, rather than a treatment-naïve UC atlas, enabled investigation of OSMR cell states from biologic nonresponders that resembled the specific patient population of MOONGLOW (74).

Accessing supplementary materials of the original study, the agent first confirmed that OSMR expression increased in three fibroblast cell states exclusively in nonresponders after treatment, consistent with the trial's rationale for targeting this receptor and with existing studies $(71, 72)$ . The agent then designed computational experiments not found in the original study. To investigate how this OSMR up-regulation affects downstream signaling, the agent implemented a genome-wide screen of transcription factor (TF) activity using the CollecTRI regulon database and applied mixed-effects modeling to identify differentially active TFs between nonresponders and responders in posttreatment tissue $(75)$ . This screen revealed that STAT1 activity was broadly elevated in nonresponder stroma where OSMR is expressed, reaching significance in 9 of 10 stromal cell types (mixed-effects P < 0.05 after FDR correction). This established that the JAK-STAT signaling axis downstream of OSMRβ is persistently active in the treatment-refractory tissue.

However, the agent noted a key structural aspect of the target biology: OSMRβ forms a heterodimer with gp130 (IL6ST), the common signal-transducing subunit used by the interleukin-6 (IL-6) cytokine family that includes IL-6, LIF, IL-11, CNTF, and CT-1. This observation brought into question the extent to which the prolonged STAT activation was genuinely attributable to OSMRβ itself versus being mediated by other gp130-family receptors that converge on the same pathway. To quantify the contribution of each receptor on inferred STAT1 activity, the agent designed a Shapley-based variance decomposition analysis across gp130-family receptor genes within each stromal cell state of posttreatment tissue from nonresponders. This analysis found that despite being the drug target, OSMR expression accounted for less variation in STAT1 activity than other co-receptors in most stromal cell types where OSMR is expressed (Fig. 5B). These findings raised the possibility that anti-OSMRβ therapy constituted a partial blockade in the context of treatment-refractory UC biology. This concept is consistent with a recent study that reported that treatment-refractory UC can reflect inflammatory pathways persisting through parallel receptor complexes that converge on the same downstream pathways (76).

The agent then hypothesized that, if redundancy within the gp130 axis contributes to treatment-refractory UC biology, a composite score capturing baseline gp130 axis expression could better predict nonresponse to therapy than OSMR alone. Thus, the clinical trialist agent then designed a gp130-axis risk score (see Materials and methods for details) and compared its performance to both OSMR expression alone and a previously reported five-gene nonresponse signature across four infliximab-treated UC cohort datasets (77–80). The gp130-axis score consistently outperformed OSMR alone across all four cohorts (Fig. 5C). The gp130-axis score also performed comparably to the Arijs signature across all cohorts, despite being derived entirely from the agent's prior single-cell mechanistic insight rather than from supervised gene selection. These findings supported the hypothesis that the gp130 signaling axis, rather than OSMR $\beta$ alone, better captures treatment-refractory disease in UC.

Finally, the Virtual Biotech investigated whether OSMR up-regulation is specific to UC or reflects a conserved stromal activation program across diseases. It conducted a systematic cross-disease expression survey by querying OSMR-pathway gene expression across 112 diseases represented in the CELLxGENE Census single-cell atlas, applying linear mixed models to test for disease-specific OSMR up-regulation relative to healthy tissue while accounting for donor and dataset batch effects (Fig. 5D) (59). The survey analysis recovered established OSMR biology in the stromal compartments of UC, colorectal, and gastric cancers, and in cardiomyocytes in myocarditis (74, 81–83). Beyond this, the Virtual Biotech highlighted roles for OSMR in acute kidney injury, pilocytic astrocytoma, periodontitis, and Alzheimer's disease and related dementias, with increased expression compartmentalized predominantly in fibroblast, endothelial, or neuronal cell types. The enrichment of OSMR expression across anatomically and etiologically diverse diseases suggested a conserved disease program associated with chronic inflammation. The cross-disease signals raised the possibility that OSMR-directed therapeutic strategies may have relevance beyond UC.

Taken together, this retrospective assessment demonstrated that the Virtual Biotech can start from a terminated clinical study and generate mechanistic insight. The agents combined trial protocol review, treatment-refractory study population target biology, single-cell signaling interrogation, biomarker performance comparison, and cross-disease expression mining to arrive at a unified rationale for the absence of efficacy: OSMRβ may label a treatment-refractory stromal inflammatory phenotype in UC but may not be the predominant driver of downstream JAK-STAT signaling in relevant OSMR-expressing cell states. Instead, these findings pointed to broader redundancy along the gp130 axis as a more accurate representation of treatment-refractory UC biology, enabling the derivation of a composite biomarker that surpassed OSMR alone in predicting nonresponse in independent cohorts. This case study was completed for a total cost of \$59.00 in Anthropic API credits, and the full conversation transcript is in supplementary text VI.

---

Correctness and reproducibility of Virtual Biotech analyses

To evaluate the correctness and reproducibility of the case study analyses performed by the Virtual Biotech, we conducted human expert review. We split each of the two case studies into four stand-alone analyses (eight total). For B7-H3, these analyses correspond to Fig. 4, panels C to F. For OSMR, they correspond to Fig. 5, panels B to D. We recruited four human reviewers, and each was assigned to two analyses conducted by the Virtual Biotech. For each analysis, reviewers were asked to perform analyses following the same instructions from the virtual CSO to AI scientists by writing their own code. They were then asked, "After implementing the analysis yourself, do your conclusions generally match that of the Virtual Biotech? Please add notes describing your findings and thought process."

Experts answered “Yes” for all eight analyses, confirming that their analyses matched the agent’s findings. Then, experts reviewed the corresponding agent-written code for the analysis and were asked “Is the code written by the Virtual Biotech correct and free from mistakes that significantly impact the conclusions of the analysis? Please add notes describing your findings and thought process.” The reviewers answered “Yes” for all eight of the analyses, identifying minor stylistic design choices that the reviewers had made differently in their own analyses. These differences did not affect findings and interpretations. See supplementary text VII for example notes from the experts and figs. S6 and S7 for side-by-side comparisons of the Virtual Biotech and the expert-generated findings. These results demonstrated that the Virtual Biotech’s implementations are accurate.

---

Comparison of the Virtual Biotech with other agentic systems

To compare the quality of Virtual Biotech responses to that of similar agentic systems, we queried Biomni, Kosmos, and PantheonOS (23–25), three state-of-the-art general purpose agentic systems, with our case studies: clinical trial annotations, B7-H3, and OSMR. For clinical trial annotation, we found that all three systems produced less accurate

A

"Analyzing NCT06137183 trial protocol and posted results. I'll delegate this task to the clinical trialist agent..."

"Accessing the ClinicalTrials.gov record. I see that this is a terminated Phase II trial testing vixarelimab in participants with moderate to severe ulcerative colitis (UC).

Analyzing study population..."

---

Terminated

Based on a fatality analysis, which suggested that the Moonglow study was unlikely to meet its primary endpoint.

A Study to Evaluate the Efficacy, Safety, and Pharmacokinetics of Vixarelimab in Participants With Moderate to Severe Ulcerative Colitis (UC)

ClinicelTrials.gov ID NCT06187183

"I found TAURUS, a single cell atlas of adalimumab treatment in UC. Original study authors found increased OSMR expression in 3 fibroblast cell states exclusively in non-responders (NRs) after adalimumab."

I'm now scoring cell-type resolved TF activity followed by mixed-effects screening...

I see elevated STAT1 activity in NR post-treatment tissue in 9/10 stromal cell types. But, OSMR shares the gp130 signaling chain with 6 other receptors. Performing Shapley dominance decomposition to quantify redundancy...

Fascinating: despite being the drug target, OSMR explains less STAT1 variance than gp130, LIFR, or IL27RA in the vast majority of stromal cell types — JAK-STAT signaling driven through other receptors."

D

"I wonder if OSMR upregulation itself specific to UC? Launching OSMR expression analysis across 112 diseases with single cell data in CELLxGENE Census.

The survey analysis recovered established OSMR biology in UC fibroblasts and cardiac muscle. Interestingly, it suggests previously underexplored roles of OSMR in acute kidney injury, pilocytic astrocytoma, periodontitis, and Alzheimer's — predominantly in fibroblast, endothelial, or neuronal cell compartments.

Significantly increased OSMR expression spans many diseases, suggesting a conserved inflammatory program with potential therapeutic relevance for vixarelimab beyond UC.

---

Eligibility Criteria

---

Description

Inclusion Criteria:

\+ Diagnosis of UC for at least 3 months + Moderately to severely active UC, assessed by mMS + Inadequate response, loss of response to, or intolerance to conventional or advanced therapies for UC

---

Exclusion Criteria:

\+ Diagnosis of Crohn's disease or indeterminate colitis

+ Suspicion of ischemic, radiation, microscopic, or infectious colitis

• Prior colectomy

B

\- Inadequate response or loss of response to previous treatment of UC with tofacitinib, upadacitinib, or other systemic janus kinase (JAK) inhibitor

"NCT06137183 enrolled a treatment-refractory UC cohort: patients showed prior poor response to conventional therapies and/or advanced therapies (biologics and targeted small molecules). They also did not enroll JAK-inhibitor non-responders."

I'd like to better understand the treatment refractory biology in this patient population..."

"If signaling redundancy potentially explains why anti-OSMR mAb failed, I hypothesize that a gp130-axis score should outperform OSMR alone in predicting non-response.

I'm constructing a gp130-axis score by averaging pre-treatment expression of gp130-axis genes. Testing across four infliximab UC trial datasets as validation...

Fig. 5. Retrospective evaluation of terminated phase 2 trial protocol by the Virtual Biotech leads to refined development strategy for an OSMRβ-directed therapy in ulcerative colitis (UC). (A) Agent summary of the phase 2 UC trial (NCT06137183) investigating vixarelimab. (B) Single-cell signaling analysis of the TAURUS UC atlas. Shapley dominance analysis decomposes STAT1 activity across gp130-family α receptors, showing that OSMR is not the leading receptor driving persistent STAT1 activity in the majority of nonresponder stromal cell types. Right panels show OSMR detection fraction among each cell type, STAT1 activity enrichment in nonresponder versus

The gp130-axis score consistently outperforms OSMR alone. Virtual Biotech mechanistic insight from single-cell decomposition validates as a stronger clinical biomarker in bulk tissue."

responder posttreatment tissue, and overall cell-state abundances. (C) External validation in four cohorts of UC patients with bulk RNA-seq data collected of a 10-gene gp130-axis score to predict responders from nonresponders. Performance was compared against OSMR alone and a published five-gene inflammatory signature. AUC annotated per score within each trial. (D) Cross-disease OSMR expression analysis. Per-donor pseudobulk OSMR expression queried across 112 diseases in CELLxGENE Census and tested via linear mixed-effects models (disease versus matched normal tissue). Heatmap shows disease settings and cell types with significantly increased OSMR expression; black dots indicate FDR P < 0.05.

annotations on both the manual review set and the TDC labels as compared to the Virtual Biotech (table S2). For B7-H3 and OSMR, we queried each system with the same initial prompt used for the Virtual Biotech and had them perform open-ended analysis (tables S3 and S4). For a fair comparison, we provided similar levels of human steering to each system as was done for the Virtual Biotech in our case studies: Ask the system for a number of possible next steps after each analysis and manually select the most promising. We selected the choice that most closely aligned with analyses that were done in the case studies for comparability. Overall, these other agentic systems used fewer data modalities, performed less rigorous analyses, and produced fewer biologically meaningful results compared to the Virtual Biotech for both case studies (see supplementary text VIII for more details).

---

Discussion

The Virtual Biotech is a multi-agent platform for early-stage drug discovery. At its core, a virtual CSO plans and coordinates multiple domain-specific scientist agents that analyze and reason over primary data sources. The domain-specialized scientist agents mirror core R&D functions and have access to a wide variety of tools and data sources. This collaborative setup allows for unified decision-making based on a collective and nuanced understanding of many primary data sources.

Our experiments showed that the Virtual Biotech is capable of performing large-scale analysis of clinical trials and identified several transcriptomic features of drug targets associated with clinical trial success. Additionally, two case studies on real-world drug development scenarios showed that the Virtual Biotech can yield insightful hypotheses that complement the evidence and conclusions reached by modern biopharmaceutical companies. This work contributes to a broader emergence of agentic AI systems organized around expert biomedical workflows $(21)$ . Beyond these scientific findings, the Virtual Biotech enables integrated cross-domain reasoning, more auditable evidence-based decision-making, and scalable analyses (see supplementary text IX for extended discussion).

The single-cell transcriptomic features of drug targets identified by the Virtual Biotech are themselves of independent interest as predictors of clinical trial success. Genetic evidence is recognized as one of the strongest predictors of clinical success and plays a central role in modern target prioritization $(39)$ . Prior work has also suggested that single cell-derived signals contain meaningful clinical information, linking these features to clinical phase progression $(56)$ . Building on this, we demonstrated that single cell-based features are not only associated with phase advancement but also with successful endpoint outcomes and reduced adverse event rates. Furthermore, we showed that single-cell evidence is orthogonal to genetic evidence and that associations between single-cell features and trial outcomes remained statistically significant within trials that had no supporting genetic evidence. Together, these findings suggest that incorporating single cell-derived evidence alongside genetic support may provide important guidance in drug development programs.

This study has limitations. The Virtual Biotech is a decision-support system for early-stage human therapeutic research, and it answers queries comprehensively on target-centered questions in well-characterized human diseases with deep coverage across genetics, transcriptomics, functional genomics, chemical biology, and clinical-trial resources. Conclusions are correspondingly weaker for rare or understudied diseases, poorly annotated targets, underrepresented tissues or populations, and questions lacking informative human genetic or single-cell evidence. The present configuration is human focused rather than veterinary, and pathogen-focused infectious-disease analyses are further constrained both by model-provider safety safeguards and by the shortage of dedicated pathogen and antimicrobial resources.

Of note, these are boundaries of the current data and tool coverage rather than of the architecture: Because data access are mediated through modular MCP interfaces, public, proprietary, or user-specific resources can be connected without redesigning the multi-agent framework, though each extension requires its own curation, specialist tools, and evaluation (supplementary text IX for extended discussion).

Additionally, our trial analyses are observational and should be interpreted as such. Possible confounders include trial targets, indications, endpoint definitions, and sponsor-related influences. In this vein, we adjusted for a trial's phase, year of enrollment, therapeutic area, and drug modality to address their potential confounding, observing consistent results. The analysis was also limited to trials registered in the ClinicalTrials.gov database and drugs that have mapped molecular targets. For the case studies, the Virtual Biotech generated data-driven hypotheses, which must be tested and further validated. There is a risk that the language model training can bias the agent toward existing hypotheses. To mitigate this, we intentionally selected case studies with trial readouts that occurred months after the language model knowledge cutoff date of January 2025.

Finally, the Virtual Biotech as presently implemented captures the computational analysis, evidence integration, and prioritization components of early therapeutic development rather than the full operations of a biotechnology organization. Experimental testing remains outside its scope. Its modality assessments reflect the data resources available to the system, so the absence of known binders or established tractability should be interpreted as an evidence gap and a source of development uncertainty rather than as proof that a modality is infeasible. More broadly, the increased throughput and scope of agentic systems raise considerations of hypothesis prioritization, provenance, independent validation, biosecurity, and scientific training (see supplementary text IX for extended discussion).

In the future, we envision expanding the Virtual Biotech to encompass even more of the drug discovery pipeline, such as molecule design, virtual screening, toxicity prediction, and clinical decision-making agents, bringing the benefits of interdisciplinary collaboration, speed, and objectivity to the entire end-to-end process. The Virtual Biotech illustrates a shift from isolated AI tools toward coordinated systems that reason across biological scales and stages of translation. As agentic platforms mature, they may enable a model of scientific organization in which hypotheses are continuously generated, stress-tested against heterogeneous evidence, and refined in partnership with human experts. Rather than replacing scientists, agentic systems could expand the scope and speed of therapeutic hypothesis exploration while making the reasoning process more transparent and reproducible.

---

Materials and methods

---

Virtual Biotech system architecture

Virtual Biotech is a multi-agent AI system that is organized to mimic the structure of a biotechnology research and development organization to accelerate and support therapeutic development and strategy. The system uses a hierarchical orchestration model comprising a chief scientific officer (CSO) as the global strategic orchestrator who oversees domain-specialized AI scientist agents and dedicated research support agents.

The virtual CSO agent supports multi-turn conversations with users, interprets queries, and dispatches analytical tasks to appropriate scientist agents. The CSO functions exclusively as an orchestrator: It routes queries, synthesizes findings, and provides strategic recommendations, but never directly accesses data or performs analyses. This separation of roles ensures that domain expertise remains encapsulated within scientist agents while the CSO maintains a global view of the research process.

---

Human interaction with the Virtual Biotech

Virtual Biotech is designed for lightweight human interaction, in which the primary role of the user is to set intent and steer the system through high-level scientific questions or instructions. After a human user submits a scientific query, the CSO agents conducts a brief clarification interview—posing follow-up questions to align on scope feasibility—before initiating computational analyses (Fig. 1C). After objectives are clarified, subsequent actions are autonomous: The CSO decomposes the task and routes subtasks to scientist agents, scientist agents perform analyses and reason over findings, and the CSO synthesizes cross-division results into recommendations. Consistent with this design, in the case studies presented in this paper, human involvement was limited to (i) responding to CSO clarification prompts and (ii) asking follow-up questions after each iteration (e.g., to probe assumptions, request sensitivity checks, or trigger additional analyses). Experimental ideation, design, and execution remained agent-driven. Users can independently verify key results by inspecting and executing the generated code, enabling human oversight.

---

AI scientist agents

The Virtual Biotech consists of a virtual CSO agent, eight distinct scientist agents distributed across four research divisions, plus two additional agents in the CSO's office that oversee the scientific direction and quality of the work (Fig. 1A). Each agent is configured with domain-specific system prompts and MCP access.

1) Target Identification and Prioritization Division integrates genetic causality, functional dependency, and cellular context to validate targets. The statistical genetics agent evaluates human genetic evidence by querying GWAS, locus-to-gene (L2G) machine learning predictions for causal gene assignment, fine-mapped credible sets, rare variant burden meta-analyses, and quantitative trait locus (QTL) colocalization to determine whether genetic variation affecting a gene influences disease risk. The functional genomics and perturbation agent assesses target essentiality and pharmacological tractability using CRISPR knockout screens and drug perturbation transcriptomic profiles from the Tahoe-100M dataset to evaluate whether a target can be modulated. The single-cell atlas agent queries CELLxGENE Census, which contains 143 million harmonized single-cell profiles, to learn cell type-specific expression patterns, differential expression between disease and healthy states, and cell-cell communication patterns

2) Target Safety Division integrates pathway reasoning, interaction-based effects, cell type-specific expression, and historical FDA regulatory safety signals to assess potential on- and off-target safety concerns. The biopathways and protein-protein interaction agent reasons through the target's biological context through pathway and molecular interaction maps to identify involvement in critical cellular processes that may pose safety concerns upon modulation. The single-cell atlas agent complements this network-level assessment by examining cell type-specific expression patterns across cell types found in 27 tissues of the Tabula Sapiens v2 atlas to identify potential off-target toxicity liabilities in critical cell types (46). The FDA safety officer agent complements these mechanistic assessments with empirical safety data from OpenFDA adverse event reports for drugs with similar mechanisms, drug safety warnings and black box labels from drug labels, and mouse knockout phenotypes to reason through potential target-level safety liabilities.

3) Modality Selection Division determines the optimal therapeutic modality for validated targets using structural biology, druggability assessment, and clinical precedence. The target biologist agent evaluates target properties that constrain modality choice, including protein family classification, subcellular localization, and structural tractability predictions that assess the likelihood of developing small-molecule binders, therapeutic antibodies, or other modality types on the basis of protein structure and binding site accessibility. The pharmacologist integrates this biological evaluation with existing clinical precedence, querying ChEMBL-derived drug databases to find approved therapies, pipeline compounds, and chemical probes that share the same target or mechanism. In addition, they analyze homologous proteins to determine family-level tractability and assess development practicality for each modality, including projected timelines, manufacturing demands, and likely hurdles.

4) Clinical Officers Division bridges preclinical assessments with clinical precedence by analyzing regulatory safety data and extracting insights from prior clinical development efforts. Working in tandem with the FDA safety officer agent, the clinical trialist agent enhances regulatory assessments with clinical precedence analysis. It locates prior trials aimed at the same gene or pathway; extracts information on trial phase, enrollment, endpoints, and results to determine the level of clinical validation; and examines discontinued studies to identify reasons for failure, including futility, safety concerns, or business-related decisions. For oncology targets, the clinical trialist agent can access clinicogenomics data to determine prognostic biomarkers across tumor types and survival outcomes from cBioPortal.

5) Office of CSO Agents: The chief of staff provides rapid strategic intelligence briefings on research fields and data landscape by inventorying MCP tools and performing web searches. The scientific reviewer performs quality assurance on AI scientist agent outputs, identifying gaps, unsupported conclusions, or misalignment with user queries (supplementary text III).

The system prompts for each agent are in supplementary text X. All agents as presented operated using Claude Sonnet 4.5, except for the chief of staff and scientific reviewer, which used Claude Haiku 4.5 for efficiency and speed. Scientist agents have access to standard computational tools including file operations and code execution, plus domain-specific data access tools in the form of MCP servers (described below). Scientist agents are instructed to maintain critical scientific thinking by questioning assumptions, validating findings across complementary datasets, and classifying evidence strength as weak or strong based on the convergence of independent lines of evidence.

---

Implementation and language model configuration

Agents and orchestration were implemented using the Claude Agent SDK, an Anthropic agent development framework to build AI platforms $(84)$ . The SDK provides several capabilities for multi-agent scientific workflows: hierarchical agent orchestration enabling a parent agent to spawn and coordinate child agents with independent execution contexts, native integration with MCP servers for standardized tool access, and built-in support for multi-turn conversations with persistent conversation state. These capabilities allow the Virtual Biotech to implement complex delegation patterns where the CSO orchestrator manages scientist agents that operate autonomously while maintaining coordinated access to shared data resources. Agent behavior is controlled through structured system prompts that define role identity and expertise boundaries, available tools and data access permissions, analysis workflows and decision frameworks, and output formatting requirements. We summarize the Virtual Biotech workflow algorithms in supplementary text II. Specialized scientific workflows, such as single-cell quality-control and analysis, were implemented in the form of agent Skills with progressive disclosure of workflow guidance.

---

Data and Virtual Biotech tool infrastructure

The Virtual Biotech integrates data from multiple public biomedical databases to enable a comprehensive evaluation of drug targets across the translational scientific spectrum and multiple biological scales. The Open Targets Platform (version 25.09) serves as the primary data backbone (33, 42). This is an extensive database that enables systematic discovery and evaluation of candidate therapeutic drug targets. By combining publicly available datasets, including those produced by the Open Targets experimental and informatics research programs, Open Targets curates data that support the development of therapeutic hypotheses. Each component contributes specific evidence types and detailed, high-resolution information on drug targets, diseases, genetic variants, QTL colocalization, mechanisms of action of therapeutic compounds, scientific publications, drug labels, FDA-reported safety signals, clinical trials, and other related data (33, 42).

We further integrated additional data resources into the Virtual Biotech that are key evidence sources in drug discovery. Single-cell atlas data are accessed and retrieved through MCP tools that leverage the CELLxGENE Census API, which aggregates harmonized single-cell RNA-sequencing (RNA-seq) profiles from over a thousand published studies, offering standardized cell-type annotations across diverse tissues and disease conditions $(59)$ . In addition, we provided the Virtual Biotech with a local copy of the Tabula Sapiens v2 healthy reference atlas $(46)$ . Pseudobulked differential expression profiles were procured from Tahoe-100M, a gigascale single-cell perturbation atlas of 50 cancer lines exposed to 1100 small-molecule perturbations $(34)$ . Agent access to clinicogenomic and outcome data in different cancer cohorts was implemented using MCP tools powered by the cBioPortal REST API $(85)$ . Finally, clinical trial information was made available to agents via tools that leveraged the ClinicalTrials.gov API $(86)$ .

Data access to each agent is mediated through 10 MCP servers, each implemented as a FastMCP Python instance that exposes domain-specific query functions as callable tools (Fig. 1B and supplementary text I). MCP provides a standardized interface between language models and external data sources, enabling structured queries with typed parameters and formatted responses $(29, 30)$ . Each MCP tool is implemented as a Python function with comprehensive docstrings and typed parameter signatures, which FastMCP automatically introspects to generate structured tool schemas exposed to the language model at runtime, injecting required context that enables agents to discover available tools and understand their purpose, required inputs, and expected outputs without manual specification $(87)$ . Tools provide not only raw outputs but also interpretable summaries and preview results, allowing agents to rapidly gauge relevance and determine whether more in-depth analysis is warranted. This facilitates efficient multistep reasoning and iterative tool invocation across diverse data sources. Each MCP server offers specialized tools tailored to frequent analytical tasks. These tools accept structured inputs—such as gene symbols, disease identifiers, and various filter parameters—and return well-formatted outputs optimized for language model consumption. The Virtual Biotech MCP tools are intentionally composable, enabling agents to chain multiple tool invocations together to construct rich, comprehensive evidence profiles.

---

Benefits of multi-agent architecture

The multi-agent architecture of the Virtual Biotech offers several key advantages when compared to single-agent approaches. The breadth of data sources and analytical tools incorporated into the Virtual Biotech, on the scale of hundreds of tools and terabytes of data spanning $\sim$ 80,000 targets, 40 million papers, and over 100 million cell profiles, makes in-depth analyses of these sources by a single agent infeasible owing to context, time, and memory constraints (33, 34, 88). By contrast, a multi-agent framework allows these analyses to be decomposed and distributed across parallel agents with isolated contexts, enabling each agent to conduct a thorough, focused analysis without overwhelming the shared context. High-level summaries from these agents are then integrated by a coordinating agent to support final decision-making.

Second, the multi-agent architecture enables parallel workflows at scale. For example, in extracting clinical trial outcomes, each extraction followed a multistep workflow: Given an input NCT ID, the clinical trialist agent first calls the ClinicalTrials.gov API to retrieve trial metadata, then evaluates whether all required fields are sufficiently populated before moving through a three-tier evidence cascade. This cascade involves searching PubMed with NCT ID confirmation and then expanding to additional web sources. Throughout this process, the agent reviews numerous information sources to understand trial designs, interpret complex clinical outcomes, determine endpoint results, map termination reasons to a standardized ontology, and consolidate all evidence into a validated JSON output that adheres to a Pydantic schema. Consequently, we assigned a dedicated agent to each trial, allowing it to use its full context window on that single study—gaining deep familiarity with its design, results, and literature—without the distraction or context dilution of managing multiple trials in one conversation. Additionally, this design enables our multi-agent parallelized workflow where 37,075 phase 2 and phase 3 trials were systematically processed in 6 hours. By contrast, processing the same number of trials with a single agent would require \~1839 hours or 76.6 days of runtime. The parallelized multi-agent architecture therefore achieves an approximately 184-fold speedup, reducing a months-long extraction to an overnight run.

---

Analysis of single-cell features and clinical trial outcomes

Clinical trial manual validation: We sampled 100 trials from the Open Targets clinical trial dataset that was included in the agent-driven curation process (50 randomly sampled from all phase 2 trials, and 50 from all phase 3 trials). Two human annotators assessed agreement with the agent's annotations for primary outcomes, secondary outcomes, and adverse event (AE) rates. For both primary and secondary endpoints, trials were classified into two groups on the basis of whether or not there was sufficient evidence supporting a positive trial endpoint. Not all outcomes were applicable to each trial; for example, primary endpoints were not counted for trials stopped early. For each trial, the annotators first reviewed the ClinicalTrials.gov record for posted results and any linked publications. If none were available, they searched Google by NCT identifier and trial name to identify relevant abstracts or reports. The annotators then compared these sources against the agent's extracted outcomes, performing additional review when discrepancies arose and documenting the reasons for disagreement. The two human annotators discussed ambiguous cases to reach a consensus. Phase 1 trials were not reviewed because their success was defined algorithmically as progression to phase 2. Phase 1 trials were classified as having progressed to phase 2 if there existed at least one phase 2 trial that tested the same set of drugs for the same indications, with a subsequent or concurrent trial start date.

Among the trials where the agent's annotations differed from manual review or TDC data, the differences were confined to intrinsically ambiguous cases: studies with multiple experimental arms in which some arms achieved the primary endpoint whereas others did not; trials where the primary endpoint failed to reach conventional statistical significance (for example, $P = 0.06$ ) despite supportive secondary or longer-term outcomes; and single-arm or dosage-based study designs that did not cleanly map onto a simple success/failure dichotomous outcome. These edge cases indicated intrinsic ambiguity in reducing complex, multi-endpoint results to discrete classes, rather than systematic mistakes by either annotation source.

Single-cell feature extraction from Tabula Sapiens: Single-cell expression features were derived from the Tabula Sapiens atlas, comprising 27 human tissues (bladder, blood, bone marrow, eye, fat, heart, kidney, large intestine, liver, lung, lymph node, mammary, muscle, pancreas, prostate, salivary gland, skin, small intestine, spleen, stomach, testis, thymus, tongue, trachea, uterus, vasculature, and ear). Cell-type annotations were based on the Cell Ontology class assignments provided in the atlas. For each target gene, features were first computed per tissue and then aggregated across tissues to obtain a single global value per gene.

Cell-type specificity: Tau index: Cell-type specificity was quantified using the Tau index ( $\tau$ ), which measures the degree to which gene expression is restricted to specific cell types (47). For a gene with mean expression $\overline{x}_{j}$ in cell type j across n cell types within a given tissue:

$$

\tau_ {\mathrm{tissue}} = \frac {\sum_ {j = 1} ^ {n} \left(1 - \frac {\overline {{x}} _ {j}}{\overline {{x}} _ {\max}}\right)}{n - 1}

$$

where $\overline{x}_{\max} = \max_{j} (\overline{x}_{j})$ is the maximum mean expression across all cell types within that tissue. The Tau index ranges from 0 (ubiquitous expression, equal across all cell types) to 1 (perfectly specific expression, expressed in only one cell type). Cell types with fewer than 20 cells were excluded from Tau computation, and mean expression was computed from log-normalized counts. For a given gene target, the Virtual Biotech then aggregated Tau metrics across all tissues as the arithmetic mean of per-tissue Tau values across tissues in which the gene was expressed.

Expression heterogeneity: Bimodality coefficient: Within-population expression heterogeneity was assessed using the bimodality coefficient (BC), which detects the presence of distinct expression subpopulations (48, 49). For a gene's expression vector across n expressing cells (cells with expression > 0) within a given tissue:

$$

\mathrm{BC} _ {\mathrm{tissue}} = \frac {m _ {3} ^ {2} + 1}{m _ {4} + \frac {3 (n - 1) ^ {2}}{(n - 2) (n - 3)}}

$$

where $m_{3}$ is the sample skewness and $m_{4}$ is the Pearson kurtosis. The bimodality coefficient ranges from 0 to 1. Only expressing cells were included in the computation to focus on heterogeneity among cells that transcribe the gene. The global bimodality score for each gene was computed as the arithmetic mean of per-tissue bimodality coefficients.

Statistical analysis of single-cell atlas features with trial outcomes: Our statistical analyses of single-cell features followed the same approach in analogous studies of genetic evidence $(39, 40)$ . Clinical trial outcomes (phase progression, trial termination, endpoint success) were analyzed with univariate logistic regression to estimate the association between single-cell expression features and outcome probability. Features were z-score standardized prior to model fitting to enable comparison of effect sizes across continuous features with different scales. Model parameters were estimated via maximum likelihood using generalized linear models with a binomial family and logit link (statsmodels.GLM,Binomial family). For trials assessing drugs that act on multiple molecular targets, and thus yield several tau and bimodality features per trial, the minimum value across all targets for each feature type was used as the single-cell agent hypothesized that the least specific target likely drives the overall safety profile. For continuous percentage outcomes such adverse event rates by organ system calculated at the patient-level, we used beta regression as implemented in R's betareg package $(89)$ . The resulting regression coefficient corresponds to the change in the logit of the mean adverse event proportion associated with a one-standard-deviation increase in the predictor. These univariate analyses are the main results we present, consistent with the approaches of Razuvayevskaya et al. $(40)$ and Minikel et al. $(39)$ .

In addition to these results, we conducted robustness analyses to further support our findings. First, to assess whether observed associations exceeded chance, we conducted permutation tests with 1000 iterations per feature-outcome combination. In each iteration, the outcome vector was randomly shuffled while retaining the original feature values, and the regression was refit to obtain a null coefficient. Empirical two-sided P values were computed as the proportion of null coefficients with absolute value greater than or equal to the observed coefficient. Permutation tests were run for both the logistic (binary outcome) and beta regression (adverse event rate) analyses using model specifications identical to those of the primary analyses.

Next, to account for potential confounding by drug modality, therapeutic area, trial start year, and clinical phase, we fit generalized mixed-effects models (GMMs). For binary outcomes, we used logistic GMMs via lme4::glmer in R with crossed random intercepts for drug-type combination and therapeutic area combination. For adverse event rate outcomes, we used beta mixed-effects regression via glmmTMB::glmmTMB with a beta family and the same crossed random-effects structure.

To evaluate whether single-cell expression features provided additional information independent of genetic evidence, we fit bivariate logistic regression models adjusting for genetic evidence. For adverse event outcomes, the analogous bivariate beta regression was fit using R's betareg. As a complementary analysis, we repeated the univariate analyses restricting to the subset of trials (75%) that have no genetic evidence. Following the approach of Razuvayevskaya et al., genetic evidence was defined as a binary indicator of whether any target-disease indication pair in a trial had a direct genetic association reported in Open Targets, which aggregated evidence from GWAS, phenomewide association studies, gene burden analyses, and clinically annotated variants (ClinVar/EVA) (33, 40).

To adjust for multiple hypothesis testing, we used the Benjamini-Hochberg procedure to control for a FDR of 5% (90).

To enable interpretable quantification, we also binarized the continuous tau cell-type specificity index into either “cell-type-specific” or “broadly expressed” categories using K-means clustering $(k=2)$ on the trial-level distribution, setting the threshold at the midpoint of the two cluster centers $(\tau=0.69)$ .

---

Single-cell and spatial transcriptomics analyses

Quality control and data preprocessing: Raw single-cell or single-nucleus RNA-sequencing count matrices were subjected to quality control filtering. Cells were retained if they expressed between 300 and 9000 genes and had $\leq15\%$ mitochondrial reads. To mitigate compositional biases from uneven sampling, donor-level downsampling was performed (maximum 10,000 cells per donor), stratified by cell type to preserve relative proportions. Gene expression was normalized using library-size normalization and log1p transformation via Scanpy (v1.9+) (91). Highly variable genes (3000) were selected for downstream dimensionality reduction.

Cell-type annotations across datasets were harmonized to a common vocabulary using the Cell Ontology $(92)$ . Fine-grained annotations were mapped to standardized Level 3 ontology categories through hierarchical ontology traversal. Cell types were filtered to those eligible for downstream differential expression analysis, requiring representation by $\geq3$ donors per condition and $\geq20$ cells per donor.

Dimensionality reduction was performed using principal component analysis (40 components). Batch effects were corrected using Harmony $(93)$ to preserve biological variation while mitigating technical differences across datasets. Uniform manifold approximation and projection (UMAP) embeddings were computed on the Harmony-corrected principal components for visualization. Batch mixing was assessed by comparing corrected and uncorrected embeddings.

Pseudobulk differential expression: Differential expression analysis was performed using a pseudobulk approach. Raw counts were aggregated by donor and cell type, and differential expression was tested using PyDESeq2 (94). Genes were defined as differentially expressed at FDR-adjusted P value < 0.05 and $\left|\log_{2}FC\right| > 1.0$ .

Cell-cell communication analysis: Cell-cell communication inference was performed using LIANA (60), which computes a consensus score across multiple ligand-receptor inference methods. LIANA's rank\_aggregate method was run with 1000 permutations within two groups. In SCLC, the first group was fibroblasts in the top quartile ranked by

B7-H3 expression, and the second group was fibroblasts in the bottom quartile. In LUAD, the top quartile cutoff for B7-H3 expression was not sufficiently high to isolate B7-H3-expressing fibroblasts, so the first group was therefore B7-H3-expressing fibroblasts and the second group was nonexpressing fibroblasts. Interactions were retained if they ranked in the top $10\%$ of interactions by coexpression magnitude, CellPhoneDB permutation test $P < 0.01$ , and were supported by $\geq 3$ constituent methods. Group-specific interactions were identified by comparing interaction sets that met the aforementioned criteria between conditions.

Cell--type deconvolution of spatial transcriptomics data: Cell-type deconvolution of 10x Visium spatial transcriptomics data was performed using Cell2Location on a NVIDIA H100 GPU (62). Briefly, this published workflow first trained a negative binomial regression model on the LUAD single-cell atlas to estimate gene expression signatures for each cell type. The trained reference signatures were then applied to spatial data to infer cell-type abundances per spot, with an expected cell density of eight cells per spot. We excluded the gene of interest (B7-H3) from this workflow to avoid circularity in the downstream analyses. The reference model was trained for 250 epochs and the spatial mapping model for 10,000 epochs. Cell-type proportions were computed from the top fifth percentile posterior estimates of cell abundances.

Spatial immune neighborhood analysis: To assess immune cell composition in the local microenvironment of gene-expressing spots, a k-nearest neighbor graph $(k = 6)$ was constructed for each tissue section using spatial coordinates (95). Within each sample, expressing spots were stratified into gene-high (top quartile) and gene-low (bottom quartile) groups on the basis of per-sample thresholds. The middle 50% of spots were excluded. Samples with fewer than 25 expressing spots were omitted. For each spot, the mean Cell2Location-estimated cell-type abundance across its k nearest spatial neighbors was computed. To test for immune depletion while controlling for confounders of spatial tissue architecture, mixed-effects models were fit separately for each immune cell type:

$$

\begin{array}{c} Y _ {i j k} \sim \beta_ {0} + \beta_ {1} \cdot \text {Gene} _ {\text {high}} + \beta_ {2} \cdot \text {UMI} _ {z} + \beta_ {3} \cdot \text {Fibroblast} _ {z} + \beta_ {4} \cdot \\ \text {Epithelial} _ {z} + \beta_ {5} \cdot \text {Endothelial} _ {z} + u _ {j} + v _ {j k} \end{array}

$$

where $Y_{ijk}$ is the mean neighbor immune abundance for spot i in sample k nested within patient j; total UMI serves as a proxy for sequencing quality, fibroblast, epithelial, and endothelial cell abundances (all measured at the focal spot) control for spatial compartment identity; and $u_{j} \sim \mathcal{N}(0, \sigma_{u}^{2})$ and $v_{jk} \sim \mathcal{N}(0, \sigma_{v}^{2})$ are patient- and sample-level random intercepts, respectively, to account for pseudoreplication and slide-to-slide variation. Spatial attenuation was assessed by repeating the analysis at increasing neighbor rings (neighbors 7–15, 16–30). Models were estimated using restricted maximum likelihood (REML) via lmerTest (96).

Biomarker-stratified survival analysis: To assess the prognostic significance of B7-H3 expression in LUAD, survival analysis was performed on the TCGA PanCancer Atlas LUAD cohort accessed via cBioPortal (63, 64). Gene expression profiles and clinical annotations—including age at diagnosis, AJCC pathologic tumor stage, sex, and survival endpoints—were retrieved for 566 LUAD patients. Analysis was restricted to patients with complete covariate data for age, stage, and sex. A total of 477 patients were ultimately used for the analysis. Patients were stratified into expression quartiles on the basis of B7-H3 (CD276) expression, comparing the top quartile (high expression) against the bottom quartile (low expression). Multivariable Cox proportional hazards models were fit for overall survival (OS), progression-free survival (PFS), disease-specific survival (DSS), and disease-free survival (DFS), adjusting for age (continuous), cancer stage [advanced (III/IV) versus early (I/II)], and sex. Survival analyses were performed using the lifelines Python package.

Transcription factor activity inference: To identify transcription factors with persistently altered activity in treatment-refractory UC tissue, TF activity scores were inferred from the TAURUS longitudinal single-cell atlas of anti-TNF (adalimumab) treatment in inflammatory bowel disease $74$ . The analysis focused on UC patients $n = 20; 14$ nonresponders, 6 responders) with paired pre- and posttreatment colonic biopsies, after excluding patients without outcome annotation. Raw counts were normalized to 10,000 counts per cell and log1p-transformed. TF activity was scored using the CollecTRI regulon collection $75, 97$ via decoupler using the multivariate linear model (MLM), which jointly scores TFs sharing target genes to correctly partition shared regulatory signals $75$ .

To compare TF activity between nonresponders and responders, linear mixed-effects models were fit for each TF and cell type on post-treatment cells:

$$

\mathrm{TFactivity} _ {i j} \sim \beta_ {0} + \beta_ {1} \cdot \text { Remission\_status } _ {j} + u _ {j}

$$

where $u_{j} \sim \mathcal{N}(0, \sigma_{u}^{2})$ is a sample-level random intercept. Models were fit using lmerTest (96) via rpy2, with Satterthwaite degrees of freedom for P-value estimation. P values were corrected using the Benjamini-Hochberg procedure within each cell type (90). This screen was applied across 766 TFs in 10 OSMR-expressing stromal cell types.

Shapley variance decomposition analysis across gp130-family receptors: To quantify the relative contribution of each gp130-family $\alpha$ -receptor to downstream transcription factor activity, the agents performed variance decomposition using the Lindeman–Merenda–Gold (LMG) method (98), which computes Shapley value-based relative importance shares by averaging marginal $R^{2}$ contributions over all possible predictor orderings. For each of the 10 OSMR-expressing stromal cell types in non-responder posttreatment tissue (defined as $\geq5\%$ OSMR detection), STAT1 activity was scored via univariate linear model (ULM) using a cleaned CollecTRI regulon from which any predictor receptor gene appearing as a regulon target (IL6ST) had been removed to prevent circularity (75, 97). Pseudobulk mean expression of each receptor and mean STAT1 ULM activity were computed per sample and cell type (minimum 10 cells per pseudobulk group). The outcome (z-scored STAT1 activity) was regressed onto seven gp130-family $\alpha$ -receptors (OSMR, IL6ST, LIFR, IL6R, IL11RA, IL27RA, CNTFR) with patient-level random intercepts:

$$

\mathrm {STAT1\_activity} _ {z} \sim \sum_ {r} \beta_ {r} \cdot \mathrm{Receptor} _ {r, z} + (1 | \text { Patient })

$$

where all predictors were z-scored within each cell type. LMG shares were computed by exhaustive enumeration of all $2^{7} = 128$ receptor subsets (one mixed-effects submodel per subset per cell type), extracting the marginal $R^{2}$ from each via MuMIn using r.squaredGLMM (99). Ninety-five percent confidence intervals were estimated via cluster bootstrap (2000 resamples at the patient level), using ordinary least squares for computational tractability.

gp130-axis biomarker score validation in bulk cohorts: To validate the gp130-redundancy hypothesis in independent patient cohorts, bulk gene expression data from four infliximab-treated UC clinical trial cohorts were obtained from the Gene Expression Omnibus (GSE12251, GSE16879, GSE23597, GSE73661). When multiple microarray probes mapped to the same gene, expression was summarized as the mean across probes. A gp130-axis composite score was defined as the simple mean of within-cohort z-scored baseline expression across 10 genes (OSMR, IL6ST, LIFR, IL6R, IL11RA, IL6, IL11, OSM, LIF, STAT1), selected by the agents based on the single-cell variance decomposition results with no empirical model fitting. Predictive performance for treatment nonresponse was assessed by area under the receiver operating characteristic curve (AUC) and compared to OSMR expression alone at baseline and to the established five-gene anti-TNF response signature of Arijs et al.

(IL13RA2, TNFRSF11B, STC1, PTGS2, IL11), scored using the same within-cohort z-score averaging procedure (77).

Cross-disease OSMR expression survey: To determine whether stromal OSMR up-regulation is specific to UC or reflects a conserved inflammatory response, we queried the CZ CELLxGENE Census (59) across 112 diseases that passed quality filters. For each disease, per-donor pseudo-bulk expression was calculated as $\log_{2}(\mathrm{CP10K}+1)$ for the OSMR gene, including only donors with at least 25 cells per cell type. Cell types were harmonized to Level 1 categories using the Cell Ontology (92). Disease and normal tissue-matched control arms were constructed, with normal donors prioritized by study and assay overlap. Differential expression was tested using weighted linear mixed-effects models:

$$

\text { mean\_log2\_OSMR\_cp10k } \sim \beta_ {0} + \beta_ {1} \cdot \text { condition } +

$$

$$

\beta_ {2} \cdot \log \_ d e p t h [ + \beta_ {3} \cdot a s s a y \_ g r o u p ] + (1 | d a t a s e t \_ i d)

$$

where condition indicates disease versus normal, log depth controls for residual sequencing depth differences, assay group (included when more than one assay type was present) accounts for platform differences, and the dataset-level random intercept accounts for study heterogeneity. Observations were weighted by the number of cells contributing to each pseudobulk estimate. Models were fit using lmerTest with Satterthwaite degrees of freedom; when fewer than two studies were available, an ordinary weighted linear model was used instead $96$ . P values were corrected globally across all disease–cell type–gene tests controlling for a FDR of 5% $90$ .

---

REFERENCES AND NOTES

1. J. A. DiMasi, H. G. Grabowski, R. W. Hansen, Innovation in the pharmaceutical industry: New estimates of R&D costs. J. Health Econ. 47, 20–33 (2016). doi: 10.1016/j.jhealeco.2016.01.012; pmid: 26928437

2. O. J. Wouters, M. McKee, J. Luyten, Estimated research and development investment needed to bring a new medicine to market, 2009-2018. JAMA 323, 844–853 (2020). doi: 10.1001/jama.2020.1166; pmid: 32125404

3. D. Sun, W. Gao, H. Hu, S. Zhou, Why 90% of clinical drug development fails and how to improve it? Acta Pharm. Sin. B 12, 3049–3062 (2022). doi: 10.1016/j.apsb.2022.02.002; pmid: 35865092

4. M. Hay, D. W. Thomas, J. L. Craighead, C. Economides, J. Rosenthal, Clinical development success rates for investigational drugs. Nat. Biotechnol. 32, 40–51 (2014). doi: 10.1038/nbt.2786; pmid: 24406927

5. Y. Zhou et al., Dynamic clinical trial success rates for drugs in the 21st century. Nat. Commun. 16, 9537 (2025). doi: 10.1038/s41467-025-64552-2; pmid: 41162353

6. C. H. Wong, K. W. Siah, A. W. Lo, Estimation of clinical trial success rates and related parameters. Biostatistics 20, 273–286 (2019). doi: 10.1093/biostatistics/kxx069; pmid: 29394327

7. D. B. Fogel, Factors associated with clinical trials that fail and opportunities for improving the likelihood of success: A review. Contemp. Clin. Trials Commun. 11, 156–164 (2018). doi: 10.1016/j.conctc.2018.08.001; pmid: 30112460

8. E. M. McDonagh et al., Human genetics and genomics for drug target identification and prioritization: Open targets' perspective. Annu. Rev. Biomed. Data Sci. 7, 59–81 (2024). doi: 10.1146/annurev-biodatasci-102523-103838; pmid: 38608311

9. D. B. Searls, Data integration: Challenges for drug discovery. Nat. Rev. Drug Discov. 4, 45–58 (2005). doi: 10.1038/nrd1608; pmid: 15688072

10. N. Denton et al., Data silos are undermining drug development and failing rare disease patients. Orphanet J. Rare Dis. 16, 161 (2021). doi: 10.1186/s13023-021-01806-4; pmid: 33827602

11. M. F. Jarvis, M. Williams, Irreproducibility in preclinical biomedical research: Perceptions, uncertainties, and knowledge gaps. Trends Pharmacol. Sci. 37, 290–302 (2016). doi: 10.1016/j.tips.2015.12.001; pmid: 26776451

12. K. Gierend et al., Provenance information for biomedical data and workflows: Scoping review. J. Med. Internet Res. 26, e51297 (2024). doi: 10.2196/51297; pmid: 39178413

13. S. Han, Q. Zhang, Y. Yao, W. Jin, Z. Xu, LLM multi-agent systems: Challenges and open problems. arXiv:2402.03578 [cs.MA] (2024).

14. D. A. Boiko, R. MacKnight, B. Kline, G. Gomes, Autonomous chemical research with large language models. Nature 624, 570–578 (2023). doi: 10.1038/s41586-023-06792-0; pmid: 38123806

15. A. M Bran et al., Augmenting large language models with chemistry tools. Nat. Mach. Intell. 6, 525–535 (2024). doi: 10.1038/s42256-024-00832-8; pmid: 38799228

16. Ch. Lu et al., The AI scientist: Towards fully automated open-ended scientific discovery. arXiv:2408.06292 [cs.AI] (2024).

17. J. Miao, J. R. Davis, Y. Zhang, J. K. Pritchard, J. Zou, Paper2Agent: Reimagining research papers as interactive and reliable AI agents. arXiv:2509.06917 [cs.AI] (2025).

18. Y. Qu et al., CRISPR-GPT for agentic automation of gene-editing experiments. Nat. Biomed. Eng. 10, 245–258 (2026). doi: 10.1038/s41551-025-01463-z; pmid: 40738974

19. H. Wang et al., SpatialAgent: An autonomous AI agent for spatial biology. bioRxiv 2025.04.03.646459 [Preprint] (2025); https://doi.org/10.1101/2025.04.03.646459.

20. S. Alber et al., CellVoyager: AI CompBio agent generates new insights by autonomously analyzing biological data. Nat. Methods 23, 749–759 (2026). doi: 10.1038/s41592-026-03029-6; pmid: 41845065

21. W. Zhao et al., An agentic system for rare disease diagnosis with traceable reasoning. Nature 651, 775–784 (2026). doi: 10.1038/s41586-025-10097-9; pmid: 41708847

22. K. Swanson, W. Wu, N. L. Bulaong, J. E. Pak, J. Zou, The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies. Nature 646, 716–723 (2025). doi: 10.1038/s41586-025-09442-9; pmid: 40730228

23. K. Huang et al., Biomni: A general-purpose biomedical AI agent. bioRxiv 2025.05.30.656746 [Preprint] (2025); https://doi.org/10.1101/2025.05.30.656746.

24. L. Mitchener et al., Kosmos: An AI scientist for autonomous discovery. arXiv:2511.02824 [cs.AI] (2025).

25. W. Xu et al., PantheonOS: An evolvable multi-agent framework for automatic genomics discovery. bioRxiv 2026.02.26.707870 [Preprint] (2026); https://doi.org/10.64898/2026.02.26.707870.

26. R. Jin et al., STELLA: Towards a biomedical world model with self-evolving multimodal agents. bioRxiv 2025.07.01.662467 [Preprint] (2025); https://doi.org/10.1101/2025.07.01.662467.

27. B. Gao et al., PharmAgents: Building a virtual pharma with large language model agents. arXiv:2503.22164 [q-bio.BM] (2025).

28. P. Sui et al., Medea: An omics AI agent for therapeutic discovery. bioRxiv 2026.01.16.696667 [Preprint] (2026); https://doi.org/10.64898/2026.01.16.696667.

29. Anthropic, Introducing the Model Context Protocol (2024). https://www.anthropic.com/news/model-context-protocol.

30. What is the Model Context Protocol (MCP)? (2026); https://modelcontextprotocol.io/docs/getting-started/intro.

31. X. Hou, Y. Zhao, S. Wang, H. Wang, Model context protocol (MCP): Landscape, security threats, and future research directions. arXiv:2503.23278 [cs.CR] (2025).

32. Open Targets, 25.09 Platform release now live! (2025).

33. A. Buniello et al., Open Targets Platform: Facilitating therapeutic hypotheses building in drug discovery. Nucleic Acids Res. 53, D1467–D1475 (2025). doi: 10.1093/nar/gkae1128; pmid: 39657122

34. J. Zhang et al., Tahoe-100M: A giga-scale single-cell perturbation atlas for context-dependent gene function and cellular modeling. bioRxiv, 2025.02.20.639398 [Preprint] (2025); https://doi.org/10.1101/2025.02.20.639398.

35. A. T. McCray, N. C. Ide, Design and implementation of a national clinical trials registry. J. Am. Med. Inform. Assoc. 7, 313–323 (2000). doi: 10.1136/jamia.2000.0070313; pmid: 10833169

36. S. Wu et al., J. GaoCollabLLM: From passive responders to active collaborators. arXiv:2502.00640 [cs.AI] (2025).

37. L. Kuhn, Y. Gal, S. Farquhar, CLAM: Selective clarification for ambiguous questions with generative language models. arXiv:2212.07769 [cs.CL] (2022).

38. H. Dowden, J. Munro, Trends in clinical success rates and therapeutic focus. Nat. Rev. Drug Discov. 18, 495–496 (2019). doi: 10.1038/d41573-019-00074-z; pmid: 31267067

39. E. V. Minikel, J. L. Painter, C. C. Dong, M. R. Nelson, Refining the impact of genetic evidence on clinical success. Nature 629, 624–629 (2024). doi: 10.1038/s41586-024-07316-0; pmid: 38632401

40. O. Razuvayevskaya, I. Lopez, I. Dunham, D. Ochoa, Genetic factors associated with reasons for clinical trial stoppage. Nat. Genet. 56, 1862–1867 (2024). doi: 10.1038/s41588-024-01854-z; pmid: 39075208

41. D. Ochoa et al., Human genetics evidence supports two-thirds of the 2021 FDA-approved drugs. Nat. Rev. Drug Discov. 21, 551 (2022). doi: 10.1038/d41573-022-00120-3; pmid: 35804044

42. G. Koscielny et al., Open Targets: A platform for therapeutic target identification and validation. Nucleic Acids Res. 45, D985–D994 (2017). doi: 10.1093/nar/gkw1055; pmid: 27899665

43. T. Tse, K. M. Fain, D. A. Zarin, How to avoid common problems when using ClinicalTrials.gov in research: 10 issues to consider. BMJ 361, k1452 (2018). doi: 10.1136/bmj.k1452; pmid: 29802130

44. J. E. Becker, J. S. Ross, Reporting discrepancies between the clinicaltrials.gov results database and peer-reviewed publications. Ann. Intern. Med. 161, 760 (2014). doi: 10.7326/L14-5022; pmid: 25402518

45. T. Fu, K. Huang, C. Xiao, L. M. Glass, J. Sun, HINT: Hierarchical interaction network for clinical-trial-outcome predictions. Patterns (N. Y.) 3, 100445 (2022). doi: 10.1016/j.patter.2022.100445; pmid: 35465223

46. S. R. Quake, The Tabula Sapiens Consortium, Tabula sapiens reveals transcription factor expression, senescence effects, and sex-specific features in cell types from 28 human organs and tissues. bioRxiv 2024.12.03.626516 [Preprint] (2025); https://doi.org/10.1101/2024.12.03.626516.

47. N. Kryuchkova-Mostacci, M. Robinson-Rechavi, A benchmark of gene expression tissue-specificity metrics. Brief. Bioinform. 18, 205–214 (2017). pmid: 26891983

48. J. A. Hartigan, P. M. Hartigan, The dip test of unimodality. Ann. Stat. 13, 70–84 (1985). doi: 10.1214/aos/1176346577

49. R. Pfister, K. A. Schwarz, M. Janczyk, R. Dale, J. B. Freeman, Good things peak in pairs: A note on the bimodality coefficient. Front. Psychol. 4, 700 (2013). doi: 10.3389/fpsyg.2013.00700; pmid: 24109465

50. P. Tong, Y. Chen, X. Su, K. R. Coombes, SIBER: Systematic identification of bimodally expressed genes using RNAseq data. Bioinformatics 29, 605–613 (2013). doi: 10.1093/bioinformatics/bts713; pmid: 23303507

51. L. Moody, S. Mantha, H. Chen, Y.-X. Pan, Computational methods to identify bimodal gene expression and facilitate personalized treatment in cancer patients. J. Biomed. Inform. 100S, 100001 (2019). doi: 10.1016/j.yjbinx.2018.100001; pmid: 34384574

52. K. D. Korthauer et al., A statistical approach for identifying differential distributions in single-cell RNA-seq experiments. Genome Biol. 17, 222 (2016). doi: 10.1186/s13059-016-1077-y; pmid: 27782827

53. A. Bessi et al., Users polarization on facebook and YouTube. PLOS ONE 11, e0159641 (2016). doi: 10.1371/journal.pone.0159641; pmid: 27551783

54. Y. Lelkes, Mass polarization: Manifestations and measurements. Public Opin. Q. 80 (S1), 392–410 (2016). doi: 10.1093/poq/nfw005

55. J. B. Freeman, R. Dale, Assessing bimodality to detect the presence of a dual cognitive process. Behav. Res. Methods 45, 83–97 (2013). doi: 10.3758/s13428-012-0225-x; pmid: 22806703

56. E. Dann. et al., Estimating the impact of single-cell RNA sequencing of human tissues on drug target validation. medRxiv 2024.04.04.24305313 [Preprint] (2024); https://doi.org/10.1101/2024.04.04.24305313.

57. W.-T. Zhou, W.-L. Jin, B7-H3/CD276: An emerging cancer immunotherapy. Front. Immunol. 12, 701006 (2021). doi: 10.3389/fimmu.2021.701006; pmid: 34349762

58. F. Kontos et al., B7-H3: An attractive target for antibody-based immunotherapy. Clin. Cancer Res. 27, 1227–1235 (2021). doi: 10.1158/1078-0432.CCR-20-2584; pmid: 33051306

59. S. Abdulla et al., CZ CELLxGENE Discover: A single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Res. 53 (D1), D886–D900 (2025). doi: 10.1093/nar/gkae1142; pmid: 39607691

60. D. Dimitrov et al., LIANA+ provides an all-in-one framework for cell-cell communication inference. Nat. Cell Biol. 26, 1613–1622 (2024). doi: 10.1038/s41556-024-01469-w; pmid: 39223377

61. M. De Zuani et al., Single-cell and spatial transcriptomics analysis of non-small cell lung cancer. Nat. Commun. 15, 4388 (2024). doi: 10.1038/s41467-024-48700-8; prmid: 38782901

62. V. Kleshchevnikov et al., Cell2location maps fine-grained cell types in spatial transcriptomics. Nat. Biotechnol. 40, 661–671 (2022). doi: 10.1038/s41587-021-01139-4; pmid: 35027729

63. J. Gao et al., Integrative analysis of complex cancer genomics and clinical profiles using the cBioPortal. Sci. Signal. 6, pl1–pl1 (2013). doi: 10.1126/scisignal.2004088; pmid: 23550210

64. J. N. Weinstein et al., The cancer genome atlas pan-cancer analysis project. Nat. Genet. 45, 1113–1120 (2013). doi: 10.1038/ng.2764; pmid: 24071849

65. M. Uhlen et al., Towards a knowledge-based human protein atlas. Nat. Biotechnol. 28, 1248–1250 (2010). doi: 10.1038/nbt1210-1248; pmid: 21139605

66. Merck, Ifinatamab deruxtecan granted breakthrough therapy designation by U.S. FDA for patients with pretreated extensive-stage small cell lung cancer (2025); https://www.merck.com/news/ifinatamab-deruxtecan-granted-breakthrough-therapy-designation-by-u-s-fda-for-patients-with-pretreated-extensive-stage-small-cell-lung-cancer/.

67. Merck, Ifinatamab deruxtecan demonstrated clinically meaningful response rates in patients with extensive-stage small cell lung cancer in IDeate-Lung01 phase 2 trial (2025); https://www.merck.com/news/ifinatamab-deruxtecan-demonstrated-clinically-meaningful-response-rates-in-patients-with-extensive-stage-small-cell-lung-cancer-in-ideate-lung01-phase-2-trial/.

68. C. M. Rudin et al., Ifinatamab deruxtecan in patients with extensive-stage small cell lung cancer: Primary analysis of the phase II IDeate-Lung01 trial. J. Clin. Oncol. 44, 261–273 (2026). doi: 10.1200/JCO-25-02142; pmid: 41086386

69. ClinicalTrials.gov, A study to evaluate the efficacy, safety, and pharmacokinetics of vixarelimab in participants with moderate to severe ulcerative colitis (UC) (2026); https://clinicaltrials.gov/study/NCT06137183.

70. C. L. Wolf, C. Pruett, D. Lighter, C. L. Jorcyk, The clinical relevance of OSM in inflammatory diseases: A comprehensive review. Front. Immunol. 14, 1239732 (2023). doi: 10.3389/fimmu.2023.1239732; pmid: 37841259

71. N. R. West et al., Oncostatin M drives intestinal inflammation and predicts response to tumor necrosis factor-neutralizing therapy in patients with inflammatory bowel disease. Nat. Med. 23, 579–589 (2017). doi: 10.1038/nm.4307; pmid: 28368383

72. S. Verstockt et al., Oncostatin m is a biomarker of diagnosis, worse disease prognosis, and therapeutic nonresponse in inflammatory bowel disease. Inflamm. Bowel Dis. 27, 1564–1575 (2021). doi: 10.1093/ibd/izab032; pmid: 33624092

73. K. Nishioka et al., Mucosal IL23A expression predicts the response to Ustekinumab in inflammatory bowel disease. J. Gastroenterol. 56, 976–987 (2021). doi: 10.1007/s00535-021-01819-7; pmid: 34448069

74. T. Thomas et al., A longitudinal single-cell atlas of anti-tumour necrosis factor treatment in inflammatory bowel disease. Nat. Immunol. 25, 2152–2165 (2024). doi: 10.1038/s41590-024-01994-8; pmid: 39438660

75. P. Badia-I-Mompel et al., decoupleR: Ensemble of computational methods to infer biological activities from omics data. Bioinform. Adv. 2, vbac016 (2022). doi: 10.1093/bioadv/vbac016; pmid: 36699385

76. J. van Oostrom et al., Anti-TNF nonresponse in ulcerative colitis: Correcting for mucosal drug exposure reveals distinct cytokine profiles. J. Crohns Colitis 19, jjae200 (2025). doi: 10.1093/ecco-jcc/jjae200; pmid: 39745888

77. I. Arijs et al., Mucosal gene signatures to predict response to infliximab in patients with ulcerative colitis. Gut 58, 1612–1619 (2009). doi: 10.1136/gut.2009.178665; pmid: 19700435

78. I. Arijs et al., Mucosal gene expression of antimicrobial peptides in inflammatory bowel disease before and after first infliximab treatment. PLOS ONE 4, e7984 (2009). doi: 10.1371/journal.pone.0007984; pmid: 19956723

79. G. Toedter et al., Gene expression profiling and response signatures associated with differential responses to infliximab treatment in ulcerative colitis. Am. J. Gastroenterol. 106, 1272–1280 (2011). doi: 10.1038/ajg.2011.83; pmid: 21448149

80. I. Arijs et al., Effect of vedolizumab (anti- $\alpha4\beta7$ -integrin) therapy on histological healing and mucosal gene expression in patients with UC. Gut 67, 43–52 (2018). doi: 10.1136/gutjnl-2016-312293; pmid: 27802155

81. T. Kubin et al., The role of oncostatin m and its receptor complexes in cardiomyocyte protection, regeneration, and failure. Int. J. Mol. Sci. 23, 1811 (2022). doi: 10.3390/ijms23031811; pmid: 35163735

82. I. Avraham-Davidi et al., Spatially defined multicellular functional units in colorectal cancer revealed from single cell and spatial transcriptomics. eLife 14, RP104815 (2025). doi: 10.7554/eLife.104815.3; pmid: 41384492

83. K. Cheng, J. Xu, J. Yang, Z. Zhang, Spatial atlas of gastric cancer reveals macrophage-driven stromal activation via ICAM1 $^{+}$ smooth muscle cells. BMC Cancer 26, 466 (2026). doi: 10.1186/s12885-026-15647-6; pmid: 41761123

84. Anthropic, Agent SDK overview (2026); https://platform.claude.com/docs/en/agent-sdk/overview.

85. M. Valerio, A. Inno, S. Gori, pyBioPortal: A Python package for simplifying cBioPortal data access in cancer research. JAMIA Open 8, ooae146 (2024). doi: 10.1093/jamiaopen/ooae146; pmid: 39735786

86. ClinicalTrials.gov, API migration guide (info API endpoints) (2024); https://clinicaltrials.gov/data-api/about-api/api-migration#info-endpoints.

87. Prefect, Welcome to FastMCP (2026); https://gofastmcp.com/getting-started/welcome.

88. National Center for Biotechnology Information, About – PubMed (2025); https://pubmed.ncbi.nlm.nih.gov/about/.

89. F. Cribari-Neto, A. Zeileis, Beta regression in r. J. Stat. Softw. 34, 1–24 (2010). doi: 10.18637/jss.v034.i02

90. Y. Benjamini, Y. Hochberg, Controlling the false discovery rate: A practical and powerful approach to multiple testing. J. R. Stat. Soc. Series B Stat. Methodol. 57, 289–300 (1995). doi: 10.1111/j.2517-6161.1995.tb02031.x

91. F. A. Wolf, P. Angerer, F. J. Theis, SCANPY: Large-scale single-cell gene expression data analysis. Genome Biol. 19, 15 (2018). doi: 10.1186/s13059-017-1382-0; pmid: 29409532

92. A. D. Diehl et al., The Cell Ontology 2016: Enhanced content, modularization, and ontology interoperability. J. Biomed. Semantics 7, 44 (2016). doi: 10.1186/s13326-016-0088-7; pmid: 27377652

93. I. Korsunsky et al., Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods 16, 1289–1296 (2019). doi: 10.1038/s41592-019-0619-0; pmid: 31740819

94. B. Muzellec, M. Teleńczuk, V. Cabeli, M. Andreux, PyDESeq2: A python package for bulk RNA-seq differential expression analysis. Bioinformatics 39, btad547 (2023). doi: 10.1093/bioinformatics/btad547; pmid: 37669147

95. P. Virtanen et al., SciPy 1.0: Fundamental algorithms for scientific computing in Python. Nat. Methods 17, 261–272 (2020). doi: 10.1038/s41592-019-0686-2; pmid: 32015543

96. A. Kuznetsova, P. B. Brockhoff, R. H. Christensen, ImerTest package: Tests in linear mixed effects models. J. Stat. Softw. 82, 1–26 (2017). doi: 10.18637/jss.v082.i13

97. S. Müller-Dott et al., Expanding the coverage of regulons from high-confidence prior knowledge for accurate estimation of transcription factor activities. Nucleic Acids Res. 51, 10934–10949 (2023). doi: 10.1093/nar/gkad841; pmid: 37843125

98. R. H. Lindeman, P. F. Merenda, R. Z. Gold, Introduction to Bivariate and Multivariate Analysis, vol. 4 (Scott, 1980).

99. K. Barton, MuMIn: Multi-Model Inference (R package version 1.18, 2015).

100. H. G. Zhang, P. Eckmann, J. Miao, A. B. Mahon, J. Zou, The Virtual Biotech: Case Study Documentation, Zenodo (2026); https://doi.org/10.5281/zenodo.22259123.

---

ACKNOWLEDGMENTS

The authors thank S. He, S. Alber, W. Deng, and other members of the Zou Lab at Stanford University for their valuable input and useful discussions on this work. Funding: National

Institutes of Health (T32-GM145402 to H.G.Z.); Knight-Hennessy Scholarship (to H.G.Z.) National Science Foundation Graduate Research Fellowship Program (DGE-2146755 to P.E.); Stanford Graduate Fellowship (to P.E.). Author contributions: H.G.Z. and J.Z. conceived and supervised the study. H.G.Z. developed, implemented, and optimized the framework. H.G.Z. oversaw analyses and visualization of the case studies. A.B.M. provided input and guidance on the case studies. P.E. designed the human expert evaluation and performed benchmarking of the framework. H.G.Z. and J.M. designed the user interface. H.G.Z., P.E., and J.Z. drafted the manuscript. All authors discussed, reviewed, and edited the final manuscript. Competing interests: A.B.M. is an employee of PHD Biosciences. All other authors declare that they have no competing interests. AI use declaration: An AI coding tool (Claude Code) was used to assist with code generation and debugging. All AI-generated code were reviewed and verified by the authors. Data, code, and materials availability: Data from the Open Targets Platform can be directly accessed at https://platform.opentargets.org. CELLxGENE data can be accessed at https://cellxgene.cziscience.com. The Tabula Sapiens dataset can be accessed at https://tabula-sapiens.sf.czbiohub.org/. The Visium lung adenocarcinoma dataset can be accessed at https://pubmed.ncbi.nlm.nih.gov/38782901/. The TCGA PanCancer Atlas lung adenocarcinoma cohort was accessed through cBioPortal (study ID luad\_tcga\_pan\_can\_atlas\_2018; https://www.cbioportal.org/study/summary?id=luad\_tcga\_pan\_can\_atlas\_2018). The microarray data for the four UC trials analyzed can be accessed with GEO

accession IDs GSE12251, GSE16879, GSE23597, and GSE73661. The TAURUS dataset can be accessed at https://pubmed.ncbi.nlm.nih.gov/39438660/. The full case study documentation including the codebase, intermediate results, agent-generated reports, figures, Virtual Biotech clinical trials dataset, and benchmarking outputs are available at Zenodo (100). The code for the Virtual Biotech, including a runnable command line interface and the complete set of system prompts and MCP servers, can be found at: https://github.com/harrisongzhang/TheVirtualBiotech. No new materials were generated in this study. License information: Copyright © 2026 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licenses-journal-article-reuse

---

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.aeg6779

Supplementary Text; Figs. S1 to S7; Tables S1 to S4; MDAR Reproducibility Checklist

Submitted 23 February 2026; resubmitted 25 May 2026; accepted 2 September 2026; published online 17 September 2026

10.1126/science.aeg6779

Wireless electrocatalytic closed-loop system for drug-free management of diabesity

Debasis Maity, Preetam Guha Ray, Peter Buchmann, Ghislaine Charpin-El Hamri, Martin Fussenegger\*

Full article and list of author affiliations: https://doi.org/10.1126/science.ady8560

INTRODUCTION: Type 2 diabetes and obesity frequently coexist (as “diabesity”) and affect billions of people worldwide. Current treatments rely largely on drugs that lower blood glucose or body weight, but these therapies can produce side effects, require lifelong administration, and may lose effectiveness over time. A subcutaneous implant containing a solid electrocatalyst that directly removes excess glucose from the bloodstream at an electrically adjustable rate could provide a dose-controllable, drug-free therapeutic alternative.

RATIONALE: Electrocatalysts use electricity to accelerate chemical reactions. We developed an implantable device containing platinum (Pt)-decorated gold (Au) nanoclusters embedded in a porous graphite matrix. When wirelessly powered, the ECHO (electrocatalytic closed-loop health optimization) device selectively oxidized glucose into gluconate, a lower-energy metabolite that is readily cleared through the kidneys. Because electrical stimulation can be precisely controlled, the rate of glucose removal could also be precisely regulated. We designed the implant to restore glucose homeostasis, reshape energy metabolism, and enable autonomous closed-loop control when coupled to a commercial continuous glucose monitor.

RESULTS: The electrocatalyst selectively converted glucose into gluconate under physiologically relevant conditions and retained its activity during prolonged operation. After implantation in mice with diabesity, wireless activation rapidly restored elevated blood glucose to the normal range, with repeated treatment maintaining glucose control for 6 weeks without detectable hypoglycemia.

Lowering glucose shifted whole-body metabolism toward the use of stored fat. Treated mice exhibited reduced circulating triglyceride concentrations and body weight, together with increased concentrations of free fatty acids and ketone bodies. The implanted device remained functional throughout the 6-week treatment period, and treatment produced no detectable adverse physiological, inflammatory, or behavioral effects.

Coupling the implant to a commercial continuous glucose monitor enabled autonomous closed-loop operation. The system activated the electrocatalyst only when glucose exceeded a predefined threshold, thereby maintaining glucose within a desired range.

CONCLUSION: A wirelessly powered electrocatalytic implant directly removed excess blood glucose while promoting the use of stored fat, all without administering drugs. Coupling the implant to continuous glucose monitoring enabled autonomous closed-loop metabolic control and demonstrated the potential of electrocatalysis as a drug-free approach to managing diabetes and obesity. □

---

Wireless electrocatalytic control

of diabesity. The ECHO system combines continuous glucose monitoring with a wirelessly powered electrocatalytic implant that oxidizes excess glucose to gluconate for renal clearance. In mice with diabesity, autonomous closed-loop control restored glucose homeostasis, reduced body weight, and promoted fat utilization without glucose-lowering drugs. [Figure created with BioRender.com. Maity, D. (2026) https://BioRender.com/r5kl19g]

Wireless electrocatalytic closed-loop system for drug-free management of diabesity

Debasis Maity $^{1}$ , Preetam Guha Ray $^{1}$ , Peter Buchmann $^{1}$ , Ghislaine Charpin-El Hamri $^{2}$ , Martin Fussenegger $^{1,3*}$

Lifestyle-related type 2 diabetes and associated obesity (diabesity) affect >2.5 billion people worldwide and impose nearly \$3 trillion in annual health care costs. We developed an implantable, wirelessly powered electrocatalytic closed-loop health optimization (ECHO) device containing platinum-decorated gold nanoclusters embedded in graphite felt that enabled selective, electrically tunable oxidation of blood glucose to renally cleared gluconate. In type 2 diabetic mice, wireless, low-power activation (0.75 volts, 10 milliamperes) restored glucose homeostasis within 1 hour, shifted metabolism from glucose to lipid utilization, reduced body weight and triglycerides, and increased circulating free fatty acids and ketone bodies without detectable hypoglycemia or adverse physiological, inflammatory, or behavioral effects during 6 weeks of treatment. Coupling this ECHO device to a commercial continuous glucose monitoring system enabled autonomous, smartphone-controlled, drug-free closed-loop management of experimental diabesity.

Diabetes and obesity are two genetic predisposition-enhanced interlinked chronic physiological lifestyle conditions associated with deregulation of metabolic energy homeostasis (1–5), the combination of which is known as diabesity. The physiological relationship between obesity and diabetes is bidirectional and self-reinforcing, involving a complex interplay of metabolic, hormonal, and inflammatory changes (1, 5–7). In particular, carbohydrate-triggered fat storage is a central driver for the development and reinforcement of diabetes and obesity, thereby creating a vicious cycle fueling the diabesity pandemic, which affects nearly 30% of the global population and incurs health care costs reaching >1% of the global gross domestic product (GDP) (1–3, 5, 8–10). As diabesity-promoting lifestyle changes correlate with welfare and increasing GDP, the diabesity pandemic continues to drastically increase globally (2, 3, 5, 10, 11).

With no cure available and low compliance with recommended lifestyle changes, current diabesity strategies focus on lifelong drug-based management of diabetes and obesity $(5, 11-13)$ . Diabetes management remains particularly challenging given the limited precision of manual open-loop interventions and the lack of real-time dynamic dosing, resulting in chronically deregulated glucose metabolism and persistent insulin administration, which may lead to acute emergencies and numerous devastating comorbidities $(14, 15)$ . Continuous glucose-monitoring devices and their closed-loop integration with drug-delivery systems such as insulin pumps have greatly improved real-time blood glucose profiling, precision insulin dosing, and prompt delivery, but technical and mechanical issues, such as insufficient sensor accuracy, sensor biofouling and signal drift, poor reliability, lack of personalization, and limited flexibility, as well as high cost and limited availability, remain major limitations $(16-20)$ .

These challenges have spurred the development of advanced cell-based proof-of-concept therapies using designer cells (21) that secrete blood glucose-modulating proteins (22–26) in response to molecular (27–30) or physical cues (31–34). These cells can be coupled with real-time in situ molecular blood glucose monitoring for closed-loop control (35, 36), a concept that has also been explored in obesity management (31, 37–40). For instance, cells engineered to sense food intake (37) or blood lipid levels (39) can regulate the production of biopharmaceuticals or satiety hormones, thereby reducing food consumption, restoring metabolic homeostasis, and maintaining a healthy body weight (37–40). Despite the promise of these engineered cells, multiple hurdles—including tumorigenicity, immune rejection, and manufacturing costs—currently impede clinical implementation (41–44). In parallel, glucagon-like peptide 1 (GLP-1) and glucose-dependent insulinotropic polypeptide (GIP) receptor agonists [semaglutide (45–47) and tirzepatide (22, 23), respectively] have emerged as a recent advance in treating diabesity by simultaneously addressing diabetes and obesity (48–52). However, high cost, frequent adverse gastrointestinal effects, and ongoing questions about long-term safety, including risks that may not yet be apparent, leave the population-level impact of these widely used therapies on the diabesity burden uncertain. Despite their early success, these drugs illustrate key challenges associated with modern medication: cost, dosing, and side effects (47, 48, 52–58). Enzyme-catalyzed glucose oxidation, exemplified by glucose oxidase-based biosensors, is among the most successful biomedical applications of catalysis. However, enzyme-based systems can suffer from limited stability and operational lifetime, and implantation introduces additional challenges that restrict long-term in vivo use. These limitations motivated the development of durable nonenzymatic electrocatalytic approaches for sustained, externally tunable, closed-loop glucose control.

In particular, electrocatalysis—accelerating chemical reactions at electrode surfaces—has recently come into the limelight in a variety of fields, such as sustainable energy storage (hydrogen production) $(59, 60)$ and carbon re- and upcycling $\left(\mathrm{CO}_{2}\right.$ reduction $(61, 62)$ . Electrocatalysts such as Pt-Ag nanoclusters $(63)$ , Au-doped $\mathrm{Co(OH)}_{2}$ nanoparticles $(64)$ , and Pt anchored on defective $TiO_{2}$ $(65)$ enable oxidation of glucose to high-value glucaric acid, which is found in fruits, honey, and wine. Also, nanoparticular electrode structures containing noble metals such as Pt and Au closely mimic the activity of glucose oxidase and can capture electrical energy, leading to applications in sensors and biofuel cells $(35, 66–72)$ . These advances underscore the promise of nonenzymatic electrocatalysts, which can offer improved robustness and tunability for sustained glucose modulation. Prior glucose biofuel cell studies established that catalytic materials can oxidize glucose in vivo $(35)$ , and classic metabolic work showed that gluconate is largely cleared or metabolized after systemic exposure $(43)$ . However, existing nonenzymatic glucose biofuel cells oxidize glucose continuously and do not enable externally tunable, on-demand glucose clearance $(66–72)$ .

Motivated by the need for controllable glucose clearance and by the limitations of chronic pharmacotherapy, including tolerability, adherence, and uncertain long-term safety (5, 14, 15, 56, 73–80), we developed an implantable, wirelessly powered electrocatalytic closed-loop health optimization (ECHO) device containing a composite electrocatalyst comprising platinum (Pt)-decorated gold (Au) nanoclusters integrated into a graphite felt (GF) matrix (GF\_Au@Pt). This ECHO device enables specific, electrically tunable oxidation of blood glucose to renally excreted gluconate at low input power. In a proof-of-concept study in obese mice with type 2 diabetes, the ECHO implant, linked through a smartphone to a commercial continuous glucose monitoring (CGM) device, automatically restored blood glucose homeostasis and shifted metabolic energy balance without drugs.

Figure 1 summarizes the operating principle and metabolic consequences of the ECHO system. The device operates in a closed loop to detect hyperglycemia and trigger on-demand electrocatalysis. When CGM readings exceed a preset threshold, the smartphone initiates wireless power transfer to the subcutaneous implant, where the GF\_Au@Pt catalyst oxidizes glucose to gluconate. This reduces circulating glucose and creates a controlled caloric sink, shifting substrate utilization toward lipid catabolism. Consequently, adipose lipolysis increases, free fatty acid concentrations rise, and hepatic ketogenesis is promoted. By maintaining glucose within the normoglycemic range without exogenous insulin, this closed-loop strategy couples glycemic control to a metabolic shift that can counteract diabetes-associated obesity.

Fig. 1. Closed-loop ECHO-CGM system for electrochemical management of diabesity. (A) Schematic of glucose regulation by the implanted ECHO system. Dietary glucose absorbed in the intestine enters the circulation and is monitored by a CGM. Under hyperglycemia, the subcutaneous ECHO implant catalyzes oxidation of glucose to gluconate, lowering circulating glucose. Reduced glucose availability promotes adipose lipolysis, converting triglycerides to glycerol and free fatty acids, which are further metabolized to ketone bodies and cleared via the kidneys. (B) Operating principle of the closed-loop ECHO-CGM system incorporating the GF\_Au@Pt electrocatalyst. When glucose is elevated (20 mM; ECHO ON), CGM feedback activates wireless power transfer and drives electrocatalytic conversion of glucose to gluconate. Under normoglycemia (10 mM; ECHO OFF), the wireless controller deactivates stimulation to maintain homeostatic glucose levels. (C) Conceptual therapeutic outcome. Subject A (pretreatment) exhibits elevated glucose and triglycerides, whereas subject B (posttreatment) shows restored metabolic balance after ECHO-assisted therapy. [Figure created with BioRender.com. Maity, D. (2026) https://BioRender.com/to7go11]

---

Results

---

Development and characterization of the GF\_Au@Pt electrocatalyst

Capitalizing on the intrinsic ability of Pt and Au to break down glucose, we synthesized a GF\_Au@Pt heterogeneous electrocatalyst by first assembling Au nanoflakes (around 150 nm in size) into spherical Au nanostructures (around 600 nm in size) on conductive high-porosity GF, then decorating these Au nanostructures with Pt dots. We constructed this structure by chemical deposition from Au and Pt salts using reducing $(\mathrm{NaBH}_{4})$ and complexing (ascorbic acid) agents in a hydrothermal reaction (24 hours, 150°C; and 1 hour, 80°C) (Fig. 2A).

We extensively characterized GF\_Au@Pt using scanning electron microscopy (SEM) (Fig. 2, B to D), low-magnification energy-dispersive spectroscopy (EDS) mapping (Fig. 2, E to H), high-resolution transmission electron microscopy (HR-TEM) (Fig. 2I), and high-magnification EDS mapping (Fig. 2, J to L), which confirmed the surface morphology of the Pt-decorated Au nanoclusters (Fig. 2, B to D). The Au nanoflakes (100 to $200\mathrm{nm}$ ) formed spherical structures (500 to $1000\mathrm{nm}$ ) on the porous nodule framework of the GF, providing ample internal space and a large surface area for catalytic activity. At higher magnification, Pt dots (<10 nm) are visible on the spherical Au clusters (Fig. 2I). EDS line-scanning spectra indicate homogeneous distribution of carbon (C), Au, and Pt, suggesting successful formation of the Au@Pt catalyst on the GF (Fig. 2, E to H). HR-TEM visualization and EDS elemental mapping of single Au flakes (around 100 nm in diameter) confirmed their decoration with Pt dots (<10 nm) (Fig. 2, I to L). Au flakes provide a stable, conductive, high-surface-area scaffold that supports Pt deposition and facilitates electron transport. Decorating the Au flakes with nanoscale Pt dots concentrates highly active surface motifs while minimizing precious-metal loading. Whereas Au alone exhibits limited glucose oxidation activity, Pt provides the dominant catalytic sites, and together, this Au flake/Pt dot architecture shows improved catalytic efficiency, selectivity, and stability relative to Au alone.

The high-resolution C 1s x-ray photoelectron spectroscopy (XPS) spectrum of pristine GF reveals characteristic peaks corresponding to $\mathfrak{sp}^2$ -hybridized carbon (284.49 eV), C-O bonds (285.97 eV), and O=C-O groups (289.56 eV), indicating the presence of oxygen-containing functional groups on the graphene framework (fig. S1A). After the incorporation of Au@Pt nanoparticles, the C 1s spectrum of GF\_Au@Pt shows dominant $\mathfrak{sp}^2$ carbon (284.42 eV) with a reduced intensity of oxygen-containing species (C-O

at 285.32 eV), suggesting partial reduction or surface modification during the synthesis process (fig. S1B). The Au 4f spectrum of GF\_Au@Pt shows well-defined doublets at 83.81 eV (Au $^{0}$ 4f $_{7/2}$ ) and 87.52 eV (Au $^{0}$ 4f $_{5/2}$ ), confirming the presence of metallic gold (Au $^{0}$ ) in the composite structure (Fig. 2M). The Pt 4f spectrum reveals the coexistence of metallic platinum (Pt $^{0}$ ) and oxidized platinum (Pt $^{2+}$ ) species, with peaks assigned to Pt $^{0}$ 4f $_{7/2}$ (71.13 eV) and Pt $^{0}$ 4f $_{5/2}$ (74.52 eV), along with Pt $^{2+}$ 4f $_{7/2}$ (71.72 eV) and Pt $^{2+}$ 4f $_{5/2}$ (74.90 eV) (Fig. 2N). The presence of both Pt oxidation states suggests partial surface oxidation of Pt nanoparticles during or after synthesis.

Electrochemical analysis of glucose oxidation by GF\_Au and GF\_Au@Pt

We first assessed the electrocatalytic glucose oxidation reaction (GOR) in a standard three-electrode setup under alkaline (1 M KOH) conditions at 21°C. We recorded linear sweep voltammetry profiles at a scan rate of 100 mV s $^{-1}$ for GF, Au-anchored GF (GF\_Au), and Pt-decorated Au-anchored GF (GF\_Au@Pt) in the presence and absence of 20 mM glucose (Fig. 3A). The profiles confirm the occurrence of the oxygen evolution reaction (OER) in the presence of GF, GF\_Au, and GF\_Au@Pt at a potential of 1.26 V versus RHE (reversible hydrogen electrode) delivering a current density of 30 mA cm $^{-2}$ . Upon addition of 20 mM glucose, GOR of GF\_Au and GF\_Au@Pt could be recorded at 0.7 and 0.6 V versus RHE knee voltage, respectively (no knee voltage for pure GF), indicating that GOR is kinetically more favorable than OER (Fig. 3A). Notably, there is no GOR in the absence of electrical power and below the knee voltage of 0.6 V. GOR's catalytic current densities of 25 mA cm $^{-2}$ (GF\_Au) and

Fig. 2. Synthesis and analysis of gold-platinum nanoclusters on graphite felt substrate. (A) Schematic illustration of the synthesis of platinum-decorated gold nanoclusters integrated into a graphite felt matrix, serving as electrocatalyst for electro-controlled glucose oxidation. $C_{6}H_{8}O_{6}$ , ascorbic acid; GF, graphite felt; GF\_Au, GF-integrated gold nanocluster; GF\_Au@Pt, GF-integrated platinum-decorated gold nanocluster; HAuCl $_{4}$ , chloroauric acid; H $_{2}$ PtCl $_{4}$ , chloroplatinic acid; NaBH $_{4}$ , sodium borohydride; NaOH, sodium hydroxide. [Panel created with BioRender.com. Maity, D. (2026) https://BioRender.com/jwq27ew] (B to D) Scanning electron micrographs of GF\_Au@Pt at different magnifications: 2360× (B), 20,070× (C), and 76,730× (D). (E to H) EDX spectroscopy mapping of GF\_Au@Pt (E) showing elemental distribution of carbon (F), gold (G), and platinum (H). (I to L) Scanning transmission electron micrographs of a single Au@Pt nanoflake at 600,000× (I) and 200,000× (J) magnification with corresponding EDX spectroscopy mapping of Au (K) and Pt (L). (M and N) XPS spectra of GF\_Au@Pt showing Au 4f (M) and Pt 4f (N) regions, confirming surface composition and chemical states. a.u., arbitrary units.

50 mA cm $^{-2}$ (GF\_Au@Pt) were reached at 0.75 V. GF\_Au@Pt's catalytic activity outperformed that of GF\_Au, as it reached higher current density (50 mA cm $^{-2}$ ) at lower voltage (0.75 and 1.25 V versus RHE, respectively), confirming both the synergistic effect of Pt decoration on the Au flakes and efficient glucose oxidation by GF\_Au@Pt (Fig. 3A). Nyquist plots confirmed that GF\_Au@Pt shows lower charge-transfer resistance owing to higher glucose adsorption, consistent with faster glucose oxidation compared with GF\_Au (Fig. 3B). Double-layer capacitance ( $C_{dl}$ ) values of GF, GF\_Au, and GF\_Au@Pt in the non-Faradaic region were 0.14, 0.27, and 0.30 mF cm $^{-2}$ , respectively (fig. S1, C to F), confirming the higher active surface area of GF\_Au@Pt and GF\_Au compared with GF. Overall, the GF\_Au@Pt catalyst outperformed GF\_Au in all respects and showed a direct correlation between glucose level and glucose oxidation over the entire physiological glucose concentration range without saturation of catalytic activity (Fig. 3C). GF\_Au@Pt exhibits maximal GOR activity in an alkaline electrolyte, while retaining substantial activity under near-physiological conditions (fig. S2).

---

Performance profiling and specificity of GF\_Au@Pt-based glucose oxidation

The voltage decrease associated with electrified GF\_Au@Pt indicates a shift from OER to GOR, and we identified gluconate as the glucose oxidation product. Liquid chromatography–mass spectrometry (LC-MS) and ${}^{1}H-{}^{13}C$ heteronuclear multiple bond correlation (HMBC) analyses confirm selective electrochemical oxidation of glucose to gluconate (glucose → gluconate), supporting high reaction selectivity under ECHO operating conditions (figs. S3 and S4). This confirms that GF\_Au@Pt-based glucose oxidation proceeds via the following two-electron transfer process: $C_{6}H_{12}O_{6}$ (glucose) + 2 OH $^{-}$ → $C_{6}H_{12}O_{7}$ (gluconate) + H $_{2}$ O + 2e $^{-}$ . When GF\_Au@Pt, immersed in 30 mM glucose solution, was powered at 0.75 V and 10 mA, the glucose concentration started to decrease within 30 min and reached 7 mM after 2 hours. Concomitantly, the gluconate concentration increased to 12 mM (Fig. 3D). While glucose oxidation reached a steady state under these in vitro conditions, GF\_Au@Pt activation continuously converted glucose to gluconate, demonstrating sustained electrochemical glucose clearance. In vivo, \~90% of gluconate is renally cleared, whereas a smaller fraction enters intermediary metabolism. We therefore hypothesized that ECHO-mediated glucose oxidation would lower circulating glucose while removing part of its caloric load (35, 43, 81, 82).

Comprehensive performance analyses showed that electro-triggered GF\_Au@Pt-mediated glucose oxidation is robust across conditions and exhibits little dependence on glucose concentration over the tested range (Fig. 3E). In contrast, the glucose oxidation rate increases with

A

B

C

D

E

F

G

H

|

Fig. 3. GF\_Au@Pt-mediated electrocatalytic glucose oxidation. (A) Linear sweep voltammetry of GF, GF\_Au, and GF\_Au@Pt-mediated glucose oxidation (20 mM glucose, 1 M KOH, scanning rate 100 mV/s). (B) Electrochemical impedance spectroscopy of GF, GF\_Au, and GF\_Au@Pt during glucose (20 mM) oxidation recorded at 0.75 V against a RHE. (C) Changes in current density of GF\_Au@Pt catalyst with glucose concentration in the range of 1 to 50 mM. (D to G) Dependence of GF\_Au@Pt-mediated electrocatalytic glucose oxidation [Dulbecco's phosphate-buffered saline (DPBS), 10 ml, 10 mA, 0.75 V] and gluconate production on (D) electrostimulation time, (E) glucose concentration (DPBS, 10 ml, 10 mA, 0.75 V), (F) GF\_Au@Pt electrode surface area (DPBS, 20 mM glucose, 10 mA, 10 ml, 1 hour, 0.75 V), and (G) sample volume (DPBS, 10 mA, 2 hours, 0.75 V). (H) Reversibility of GF\_Au@Pt-mediated electrocatalytic glucose oxidation (30 mM glucose, 10 ml, 2 hours, 0.75 V, 10 mA, 1 cm by 1 cm surface area). (I) Chronoamperometric analysis of GF\_Au@Pt-mediated electrocatalytic glucose specificity in the presence of physiological concentrations of the indicated metabolites (0.75 V, 10 ml, PBS). Data are represented as mean ± SEM ( $n = 5$ biological replicates). Statistical significance was determined using two-way analysis of variance (ANOVA) with Dunnett's multiple comparisons test.

GF\_Au@Pt working area (Fig. 3F) and decreases with increasing reaction volume (Fig. 3G). Furthermore, performance is reversible and stable during prolonged operation (Fig. 3H), consistent with the material integrity of GF\_Au@Pt (fig. S5).

Under all conditions examined, electro-triggered GF\_Au@Pt activation exclusively oxidized glucose without targeting other sugars or representative electroactive molecules (Fig. 3I). Linear sweep voltammetry and amperometric analyses showed that GF\_Au@Pt was electrochemically specific for glucose under the tested conditions, as metabolites, proteins, hormones, vitamins, and drugs tested at 5- to 10-fold physiological concentrations produced negligible current responses and did not measurably interfere with glucose oxidation (figs. S6 and S7).

---

Electrotunability of GF\_Au@Pt-mediated glucose oxidation

Restoring blood glucose homeostasis in type 2 diabetes and resetting the metabolic energy balance for the treatment of obesity require precise and dynamic control of electrocatalytic GF\_Au@Pt activity. To comprehensively analyze electrotunability of GF\_Au@Pt-mediated electrocatalytic glucose oxidation, we immersed GF\_Au@Pt in a physiological saline solution containing $30~\mathrm{mM}$ glucose and varied the voltage at a fixed current while profiling glucose-clearance kinetics. The electrochemical onset (knee) voltage for GF\_Au@Pt-mediated glucose oxidation was approximately 0.6 V, while 0.75 V was used as the practical operating voltage to achieve efficient glucose oxidation (Fig. 3A). Next, with the activation threshold voltage set to 0.75 V, we recorded the glucose oxidation rate while varying the current (Fig. 3D). The results confirmed that the glucose concentration could be precisely adjusted to desired levels by tuning the input power. GF\_Au@Pt displays a concentration-dependent increase in glucose oxidation rate over the tested range (fig. S8). Overall, these findings demonstrate that GF\_Au@Pt-mediated GOR can be fine-tuned and scaled as required by adjusting (i) the surface area of GF\_Au@Pt, (ii) the operating time, and (iii) the electrical power, suggesting that the electrocatalyst can be flexibly adapted to different systems and conditions. Chlorine and reactive oxygen species (ROS) assays showed no detectable oxidant generation at the ECHO operating potential of 0.75 V, supporting a nonoxidative operating window and good biocompatibility in accordance with the results of cytotoxicity testing (figs. S9 and S10).

---

Electrocatalytic closed-loop control of GF\_Au@Pt-mediated glucose oxidation

Having established that GF\_Au@Pt-mediated glucose oxidation can be electrically tuned, we next asked whether this reaction could be coupled to real-time glucose sensing to create closed-loop control over glucose clearance. In preparation for using GF\_Au@Pt-mediated GOR for dynamic closed-loop control of blood glucose, we first simulated prandial glucose intake by addition of glucose at regular intervals during electrocatalyst operation. Then, to simulate closed-loop control, we connected electrotunable GF\_Au@Pt activation to glucose-level input data from a commercial CGM device. In this system, the CGM detected simulated postprandial glucose spikes in real time and repeatedly activated GF\_Au@Pt electrocatalysis. This setup maintained glucose near the preset 5 mM target by dynamically switching electrocatalytic glucose oxidation on and off (fig. S11).

---

Construction of wirelessly powered ECHO implants for tunable glucose oxidation

To enable drug-free real-time wireless-powered electrotunable blood glucose control, we assembled the GF\_Au@Pt electrocatalyst and associated switchboard electronics in a sealed IPX7-certified waterproof (IEC 60529) 3D-printed US Food and Drug Administration (FDA)-licensed polyamide casing [26 mm by 14 mm by 5 mm, length × width × height (LWH)] (fig. S12). The ECHO module features a compact, polyamide-encapsulated printed circuit board that provides biocompatible electrical insulation (fig. S12). A dual-layer radiofrequency regulator stabilizes wireless power delivery (fig. S13), and a 13.56 MHz transmitter enables efficient power transfer and synchronized activation for closed-loop electrocatalysis (figs. S14 and S15). The switchboard also contains a receiver coil synchronizing power transfer from the emitter coil of the extracorporeal power generator at an industrial, scientific, and medical (ISM) band frequency of 13.56 MHz. For increased biocompatibility, we encapsulated the bioelectronic GF\_Au@Pt-containing implant in a three-layer structure consisting of a BioGlue-sealed polycellulose bag and alginate coating, with EDC-NHS [1-ethyl-3-(3-dimethylaminopropyl)carbodiimide-N-hydroxysuccinimide] surface functionalization of the alginate layer. The ECHO-CGM system integrates continuous glucose sensing with wireless electrocatalytic feedback to enable autonomous, biocompatible glucose regulation through closed-loop electrochemical control (fig. S16). The final GF\_Au@Pt-containing ECHO implants retained their electrochemical glucose-oxidation activity after 3 months of subcutaneous implantation in obese type 2 diabetic (db/db) mice (fig. S17).

---

Electrocatalytic blood glucose control managing type 2 diabetes and associated obesity

To validate drug-free electrocatalytic control of blood glucose concentrations and diabetes-associated obesity, we implanted the ECHO device in db/db mice (fig. S18) and measured the blood glucose responses across applied voltages from 0 to $0.9\mathrm{V}$ (Fig. 4A) and currents from 0 to $10\mathrm{mA}$ (Fig. 4B). We found that operating at $0.75\mathrm{V}$ and $10\mathrm{mA}$ was sufficient to drive efficient blood glucose oxidation. Upon activation of the ECHO implant at $0.75\mathrm{V}$ and $10\mathrm{mA}$ , normoglycemia is achieved within 60 to $90\mathrm{min}$ (Fig. 4C), which is comparable to the postprandial blood glucose profile of wild-type (WT) mice (Fig. 4D) as also observed in glucose tolerance test (GTT) (fig. 4E). These results show that electrocatalytic glucose clearance can reduce hyperglycemia on a timescale comparable to physiological postprandial glucose normalization in WT mice. Hyperglycemia gradually returned after switching off electrocatalysis (Fig. 4D). Repeated activation of ECHO-controlled electrocatalytic glucose oxidation for different time periods confirmed rapid and reversible blood glucose control, suggesting that ECHO can provide robust, reproducible, and precise dynamic control of blood glucose in experimental type 2 diabetes (Fig. 4F). With optimal stimulation power $(0.75\mathrm{V}, 10\mathrm{mA})$ and activation time $(60\mathrm{min/day})$ , hyperglycemic excursions could be consistently attenuated to reach normoglycemia and manage experimental type 2 diabetes for at least a 6-week period (Fig. 4, F and G). Blood pH remained stable and never decreased below 7.35, indicating that the bicarbonate buffering system (37) was unaffected by gluconate produced by ECHO-triggered glucose oxidation, and suggesting the absence of ketoacidosis (fig. S19). Inductively coupled plasma mass spectrometry (ICP-MS) and SEM-EDX (energy-dispersive x-ray) analyses support the chemical and structural stability of the GF\_Au@Pt electrocatalyst, with no Au or Pt detectable in serum or organ tissues under the tested conditions (table S1 and fig. S20). These findings indicate that the alginate-polycellulose encapsulation effectively limits systemic metal exposure, and they also support in vivo biostability of the ECHO system.

In contrast to classical blood glucose management, which focuses on insulin-triggered glucose uptake and metabolism by individual cells, ECHO-based electrocatalytic conversion of glucose into renally cleared gluconate results in a decrease of overall systemic metabolic energy, which is expected to favor a shift from glucose- to fat-based energy metabolism and may therefore reduce diabetes-associated obesity. Indeed, db/db mice treated for a 6-week period with ECHO-based electrocatalytic glucose oxidation not only had their normoglycemia restored but also lost up to 10% of their body weight (Fig. 5A). Reduced insulin concentrations in ECHO-treated type 2 diabetic (T2D) mice are consistent with improved glycemic control and may reflect enhanced insulin sensitivity (Fig. 5B). Furthermore, the animals showed lower concentrations of blood triglycerides (Fig. 5C) as well as higher amounts of free fatty acids (Fig. 5D) and ketone bodies (Fig. 5, E and F) compared with untreated T2D and WT animals, confirming the electrocatalytically mediated transition from glucose- to fat-based metabolism. Notably, ECHO's electrocatalyst remained fully intact, and explanted devices exhibited native performance (fig. S17). Electrocatalytic glucose oxidation may therefore provide a drug-free route to modulate both hyperglycemia and obesity-associated metabolic features in experimental diabesity. During the clamp procedure, the ECHO system maintained glucose within the euglycemic range, supporting effective control without evidence of hypoglycemia under the tested conditions.

---

Closed-loop electrocatalytic blood glucose oxidation for treatment of diabesity

As with any open-loop therapeutic intervention, hypoglycemic excursions and excessive metabolic energy depletion associated with open-loop drug-free electrocatalytic glucose oxidation cannot be completely mitigated. We therefore constructed a dynamic, closed-loop system by interfacing the ECHO implant with a commercial CGM device (fig. S18). This CGM device profiles interstitial glucose concentration in real time and wirelessly transfers the data to a compatible reader or smartphone, which was programmed to wirelessly activate the ECHO implant whenever glucose reached hyperglycemic levels (Fig. 6A and movie S1). To optimize the closed-loop control dynamics and prevent hypoglycemic excursions, we profiled the closed-loop ECHO-CGM communication and control system at various threshold glucose concentrations (5 to 15 mM glucose; Fig. 6B) and ECHO activation cycles (5 to 30 min; Fig. 6C). This characterization revealed that the closed-loop ECHO-CGM system could be set to any glucose threshold in the above range and that short ECHO activation cycles of 5 min were sufficient to provide stable glucose homeostasis and prevent hypoglycemic excursions (Fig. 6C and movie S2). During the licensed lifespan of the CGM (12 days) (83–86), the closed-loop ECHO-CGM system provided automatic, robust, and stable glucose homeostasis in implanted T2D mice (Fig. 6D and movie S2). Cytokine profiling together with histological and immunohistochemical analyses revealed no marked differences in systemic inflammatory cytokines or local inflammatory cell infiltration between stimulated and nonstimulated T2D cohorts, supporting preserved tissue integrity and controlled remodeling at the device-tissue interface over 30 days in vivo (figs. S21 to S23).

The loss of metabolic energy resulting from the glucose drain driven by the electrocatalytic glucose oxidation also induced a shift to fat-based energy metabolism, characterized by lower blood triglyceride concentrations (Fig. 6E), higher blood fatty acids (Fig. 6F), and higher ketone body concentrations (Fig. 6, G and H), indicating that ECHO stimulation induced a shift from glucose-based to fat-based metabolism. Consistent with this, hyperinsulinemic-euglycemic clamp studies showed that ECHO-treated diabetic mice required significantly higher glucose infusion rates than unstimulated diabetic controls under controlled glycemic conditions (fig. S24; P < 0.0001). Under the tested dosing regimens, ECHO electrocatalysis produced rapid glucose lowering with stable CGM profiles and no hypoglycemic overshoot compared with the pharmacological comparator groups (fig. S25). The platform further showed scalable and stable glucose oxidation in a 5-liter-scale circulation setup, with larger GF\_Au@Pt electrodes increasing catalytic capacity and enabling near-complete oxidation. These results support further preclinical development and demonstrate scalable glucose oxidation under human volume-relevant perfusion conditions (fig. S26).

A

B

C

D

E

F

G

Fig. 4. ECHO-controlled blood-glucose levels in experimental type 2 diabetes. (A and B) Blood glucose profiles of ECHO-implanted T2D mice stimulated with different voltages (A) and currents (B). (C) Blood glucose profiles of ECHO-implanted T2D mice [T2D-ECHO (On)] stimulated with 0.75 V and 10 mA compared with WT, untreated T2D, unstimulated ECHO-implanted [WT+ECHO (Off)], and unstimulated ECHO-implanted [T2D+ECHO (Off)] animals. (D) Real-time ECHO-mediated blood glucose control of T2D mice repeatedly stimulated for different periods of time with 0.75 V and 10 mA. (E) GTTs of WT, T2D, WT+ECHO (Off), and T2D+ECHO (Off) animals, and ECHO-implanted stimulated (0.5 cm $^{2}$ , 0.75 V and 10 mA) T2D mice. (F and G) Long-term blood glucose profiling of WT, T2D, and ECHO-implanted T2D mice stimulated once daily with 0.75 V and 10 mA. Blood glucose levels were monitored daily for 7 days (F) and weekly for 6 weeks (G). Error bars represent mean ± SEM ( $n = 8$ biological replicates). Statistical significance was determined using two-way ANOVA with Tukey's multiple comparisons test.

---

Discussion

Glucose is the major metabolic energy carrier in mammals, providing short-term power (87, 88). It is supplied by dietary intake and internally produced by gluconeogenesis in the liver, and blood glucose homeostasis is maintained by insulin-triggered cellular glucose uptake and consumption as well as by conversion into fatty acids and triglycerides for reversible longer-term energy storage (1, 88). Imbalanced and chronically elevated blood glucose is the root cause of diabesity, with interlinked pathophysiology and comorbidities of diabetes and obesity (5, 14, 15). The range of drug-based therapies for diabesity, designed to directly or indirectly address chronically elevated blood glucose concentrations, includes treatments such as gluconeogenesis blockers (79, 89), insulin and insulinogenic peptide hormones [GLP-1 (90), GIP (91), and adiponectin (92)], and drugs such as semaglutide (23, 45, 46) and tirzepatide (22, 23). Additionally, advanced cell therapies are currently undergoing proof-of-concept studies or are in preclinical development (87). However, like all drug-, gene-, or cell-based treatment strategies, current diabesity therapies are limited by challenging control dynamics, side effects, poor patient adherence, high cost per dose, and the need for lifelong administration (93, 94). Therefore, for next-generation medicine, the quest for control modalities providing reversible, adjustable, and

A

B

C

D

E

F

Fig. 5. Impact of ECHO-controlled blood glucose homeostasis on interlinked obesity. Body weight (A) and concentrations of insulin (B), triglycerides (C), free fatty acids (D), $\beta$ -hydroxybutyrate (E), and acetoacetate (F) in WT, T2D, and ECHO-implanted T2D mice stimulated once daily with $0.75\mathrm{V}$ and $10\mathrm{mA}$ for at least 6 weeks. Error bars represent mean $\pm$ SEM ( $n = 8$ biological replicates). Statistical significance was assessed using two-way ANOVA with Tukey's multiple comparisons test.

dynamic control of therapeutic interventions is ongoing (24, 25, 95, 96). Early efforts centered on open-loop therapeutic transgene control using small-molecule compounds enabling oral (21, 24, 25, 29, 30, 40, 96–106), percutaneous (100, 107), and inhaled administration (27–30, 105). Subsequently, pioneering closed-loop systems that dynamically coordinate biomarker input to control implanted designer cells (21) have been successfully used for the treatment of diverse conditions (35, 36, 103, 108). A range of gene switches and therapeutic gene networks responsive to nonmolecular physical cues such as light (31–34), temperature (109, 110), mechanical pressure (111), and sound (112, 113) have similarly advanced genetic precision control, and the emerging science of electrogenetics, which interfaces wearable electronics with metabolism using electrical fields (114), direct current (115), and electromagnetic fields (116), has enabled remote control of therapeutic gene expression (31, 32, 117, 118). Recent advances in electrogenetics have also incorporated material science to expand the reaction space by using catalytic nanocomposite materials. A notable example includes a metabolic fuel cell, in which a cuprous oxide-containing nanocomposite material continuously catalyzes blood glucose oxidation, producing a glycemia-correlated direct current (35, 108). An electronic power-management circuit draws direct current only during hyperglycemia and electrostimulates coimplanted designer cells either optogenetically or directly to produce and release insulin, restoring normoglycemia in a closed-loop manner $(35, 108)$ . Wireless control and information exchange are mediated by wearable devices and smartphones. However, as the glucose-to-direct current conversion is irreversible, the nanocomposite material of the metabolic fuel cell cannot operate in the reverse mode: fine-tuning blood glucose concentrations by adjusting direct current input.

Building on prior work showing that nanocomposite materials can interface glucose oxidation with bioelectronics $(35)$ and on advances in noble metal electrocatalyst design $(35, 119–121)$ , we designed GF\_Au@Pt for a distinct purpose: externally powered, reversible, and tunable glucose oxidation rather than continuous glucose-to-current conversion. The resulting electrocatalyst consists of high-porosity conductive graphite felt with Au nanoflakes assembled into spherical nanostructures decorated with Pt dots. This structure facilitates efficient electron transfer and enhances glucose accessibility owing to its high surface area. The flower-shaped Pt dot-decorated Au framework promotes glucose adsorption and improves oxidation kinetics, resulting in superior overall reaction performance compared with conventional noble metal catalysts. This hierarchical structure showed high morphological stability over extended operation times, rapid electron-transfer kinetics, and high specificity for glucose, and it enabled reversible and

C

D

E

F

G

H

Fig. 6. Smartphone-enabled closed-loop control of glucose homeostasis using a commercial continuous glucose monitoring device integrated with the ECHO system. (A) Schematic representation of the smartphone-enabled wireless closed-loop CGM-ECHO integration to control glucose homeostasis in T2D mice. The commercial CGM continuously profiles glucose levels in real time and initiates wireless stimulation of the ECHO implant when glucose levels rise above a preset threshold concentration. [Panel created with BioRender.com. Maity, D. (2026) https://BioRender.com/Obj48wp] (B) Glucose homeostasis kinetics in response to different preset glucose threshold concentrations. (C) Impact of different ECHO stimulation times (0.5 cm $^{2}$ , 0.75 V, 10 mA) and glucose homeostasis kinetics in ECHO-implanted T2D mice. (D) Autonomous maintenance of glucose homeostasis of T2D mice controlled by the integrated closed-loop CGM-ECHO device during the licensed CGM lifespan of 12 days. (E to H) Impact of CGM-ECHO treatment on obesity. Concentrations of triglycerides (E), free fatty acids (F), acetoacetate (G), and β-hydroxybutyrate (H). Animal groups included WT mice and T2D mice, both with CGM; WT mice and T2D mice equipped with the closed-loop CGM-ECHO system but not stimulated (Off); and T2D mice equipped with the closed-loop CGM-ECHO system and stimulated (On). Data are represented as mean ± SEM ( $n = 8$ biological replicates). Statistical significance was assessed using two-way ANOVA with Dunnett's multiple comparisons test.

precisely tunable glucose oxidation in the physiological range using low electrical power (less than or equal to the power required by pacemakers). Most importantly, the GF\_Au@Pt electrocatalyst exhibited rapid electrotunability, efficiently restoring postprandial blood glucose concentrations with kinetics comparable to those observed in WT or insulin-treated T2D mice, without occurrence of hypoglycemia or associated behaviors such as shivering. In contrast to native physiology or insulin-based therapies where the body's overall metabolic energy level remains constant after hormone-triggered cellular glucose uptake, electrocatalytic glucose oxidation produces gluconate, 90% of which is rapidly cleared by the kidneys, resulting in a net decrease in overall systemic metabolic energy (35, 43, 81, 82). Notably, we did not observe gluconate-associated acidosis during this process, indicating that the body's bicarbonate buffering system remains effectively regulated (37). These results suggest that the electrocatalytic gluconate-based approach to reducing net metabolic energy (43, 81) can reset the body's overall metabolic load by promoting fat mobilization while preventing further fat accumulation. Electrocatalytic regulation of blood glucose homeostasis also resulted in lower blood triglyceride concentrations, higher circulating free fatty acids, increased ketone body production, and reduced body weight—all achieved with no detectable change in food intake. These effects align closely with outcomes previously observed using cell-based appetite suppression strategies tailored for managing diet-induced obesity (39).

Throughout prolonged periods of diabesity management, the electrocatalytic implant was stable, maintaining its native performance without apparent local inflammation or elevation in systemic inflammatory biomarkers. Moreover, ECHO operation did not alter feeding habits or overall behavior of the animals, further supporting tolerability under the tested in vivo conditions. In particular, the absence of altered eating behavior and hypoglycemic episodes highlights a potential advantage of drug-free electrocatalytic glucose modulation, which avoids systemic exposure to glucose-lowering drugs in this experimental setting. This drug-free mode of action is relevant because current pharmacological therapies for diabesity can be limited by adverse effects, tolerability, adherence, cost, and unresolved long-term safety questions (122).

The electrocatalytic ECHO device can seamlessly integrate with modern electronics and can in principle connect to a variety of wearable devices, smartphones, and cloud-based Internet of Things networks. This enables real-time monitoring, online diagnostics, and therapeutic interventions, by either the patient or medical professionals, or through artificial intelligence-driven algorithms that anticipate daily routines and lifestyle changes. As a proof of concept, we wirelessly linked the ECHO device implanted in T2D obese mice to a commercial CGM system. The CGM system tracked glucose concentrations and transmitted the data to a smartphone, which processed the data and sent instructions back to the ECHO device, which in turn rapidly adjusted glucose concentrations to maintain homeostasis in response—achieving this without hypoglycemic episodes or the need for small-molecule drugs or biopharmaceuticals. Standard practice for CGM-licensed applications is to replace the CGM every 12 days (16, 83–86). Because the electrocatalyst was not consumed or degraded over this timescale, and because the ECHO implant showed preserved local tissue compatibility, longer-term drug-free modulation of glucose and metabolic energy balance may be feasible with routine CGM replacement, provided that implant performance and tissue compatibility remain stable over extended periods. We believe that closed-loop electrocatalytic devices could establish a platform concept for regulating disease-relevant metabolites, offering a potential route toward drug-sparing, dynamically controlled metabolic interventions.

---

Materials and methods Chemicals

Chloroplatinic acid (H $_2$ PtCl $_6$ ), chloroauric acid (HAuCl $_4$ ), sodium borohydride (NaBH $_4$ ), ascorbic acid (C $_6$ H $_8$ O $_6$ ), glucose (C $_6$ H $_{12}$ O $_6$ ), hydroxylamine hydrochloride (NH $_2$ OH·HCl), $N$ -hydroxysuccinimide (NHS, C $_4$ H $_5$ NO $_3$ ), phosphate-buffered saline (PBS), $N,N'$ -bis(acryloyl)cystamine (BAC), polycellulose, sodium dodecyl sulfate (SDS), 1-ethyl-3-(3-dimethylaminopropyl) carbodiimide (EDC), 2,2'-azobis(2-methylpropionamidine)dihydrochloride (AAPH), and bovine serum albumin (BSA) were purchased from Sigma-Aldrich. Glucose-to-gluconate conversion was quantified using a colorimetric D-Gluconate Assay Kit (Sigma-Aldrich, MAK279). Alginate was purchased from Buechi AG (cat. no. 11061528). Acetone, carbon black, ethanol, potassium hydroxide (KOH), and sodium hydroxide (NaOH) were sourced from ThermoFisher Scientific. Platinum (Pt) round wire (0.50 mm diameter, density 20.10 g/cm $^3$ ) was bought from Cooksongold, silver ink (cat. no. 186-3600) from RS Components, graphite felt (GF, cat. no. G100) from Fuel Cell Store, and BioGlue (cat. no. BG3510-5-US) from CryoLife Inc. Integrated circuits (LT3092ETS8), ceramic capacitors (MC0402), diodes (Diodes Inc., 1N4148; ST Microelectronics, BAT54J), and other small electronic components such as resistors were obtained from Digi-Key Inc.

---

Electrocatalyst fabrication

Graphite felt preparation: GF was cut into 5 cm by 5 cm squares, which were washed once with 70% (v/v) ethanol, dried overnight at 21°C, autoclaved in double-distilled water (ddH₂O) at 120°C for 24 hours to remove loose nanofibers and activate the GF surface, washed in ddH₂O, and vacuum-dried at 40°C overnight.

Synthesis of GF\_Au electrocatalyst: Au nanoclusters were synthesized by mixing 10 ml of HAuCl $_4$ solution (10 mM) with 10 ml of NaOH solution (0.1 M). The mixture was stirred for 30 min and sonicated for 30 min at 21°C (82% amplitude, 20 kHz, 250 W, 5 s on/off cycles; Sonopuls

HD4100, Bandelin Electronic GmbH), then the activated GF was added, followed by 1 ml of $NaBH_{4}$ (0.001 M) and 1 ml of ascorbic acid (0.1 mM). After mixing for 5 min, the immersed GF was collected and incubated in an airtight hydrothermal autoclave reactor for 24 hours at 150°C to complete the flake-based Au cluster formation on the GF. The hydrothermal autoclave reactor was then rapidly cooled in a water bath set to 21°C—this is crucial to maintain the flake-based Au clusters anchored on the GF (GF\_Au) (Fig. 2C). The GF\_Au was collected, washed five times with 70% (v/v) ethanol, and vacuum-dried at 50°C overnight.

Synthesis of GF\_Au@Pt electrocatalyst: To decorate the flake-based Au clusters anchored on the GF with platinum, we immersed GF\_Au in 30 ml of $H_{2}PtCl_{6}$ (1 mM) and centrifuged it at 8000g for 20 min. The GF\_Au was transferred to a mixture of 40 ml of ascorbic acid (0.1 mM) and 5 ml of $NaBH_{4}$ (0.001 M), incubated for 1 hour at 80°C under stirring at 100 rpm, washed six times in an acetone- $ddH_{2}O$ mixture (1:10), and vacuum-dried at 45°C for 48 hours to afford Pt-decorated flake-based Au cluster-anchored GF, the final GF\_Au@Pt electrocatalyst, which was ready to be placed in the wireless-powered implant.

---

ECHO device

The ECHO device was assembled sequentially, including the wireless-powered electronics, the GF\_Au@Pt plug-in and the encapsulation of the biocompatible implant. Electronics: (i) Field generator. We constructed a field generator to wirelessly power and control the bioelectronic ECHO implant. The 35.5 cm by 19.5 cm by 10 cm (LWH) field generator incorporates components such as a frequency generator using an oscillating circuit and impedance converter amplifying the transistor's output signal, which drives the resonance unit (13.56 MHz) of the transmitter coil and wirelessly transfers electromagnetic power to the receiver coil of the ECHO implant. (ii) Wireless-powered bioelectronic ECHO implant. The wireless-powered bioelectronic ECHO implant casing (26 mm by 14 mm by 5 mm, LWH), fabricated from 3D-printed (Formiga P110 3D printer; EOS GmbH, Krailling, Germany) FDA-licensed polyamide (Fabb-It 3D Druckservice, Loerrach, Germany), contains the receiver coil that collects the electromagnetic output of the power generator at the resonance frequency and rectifies it to direct current through diodes. The direct current power activates ECHO's custom-printed switchboard (22 mm by 8 mm by 0.7 mm) containing a programmable two-terminal current source (IC LT3092) that regulates device operation, capacitors (C4, C5) to adjust the resonance frequency, resistance (R1) to control the output current, and diodes (D3, D4) to control the output voltage at the anode containing the plugged-in GF\_Au@Pt electrocatalyst (0.5 cm $^{2}$ ), fixed with conductive silver ink. To prevent any short circuit, the anode and catalyst were insulated with BioGlue. The cathode consisted of a platinum wire (diameter 0.5 mm, length 10 mm). The circuit diagram and a list of electronic components are provided in figs. S12 to S16 and table S2. Overall, the ECHO device was rated IPX7 waterproof according to the International Protection Marking standards (IEC 60529). (iii) ECHO implant encapsulation. To improve its biocompatibility, we encapsulated the ECHO implant in a three-layer structure consisting of polycellulose, alginate, and EDC-NHS for surface functionalization. We first packed the ECHO device in a polycellulose membrane bag sealed with BioGlue and then covered the bag with alginate by immersing it in 0.2 (wt %) alginate solution for 6 hours at 21°C and dipping it for 30 min in 0.1 M CaCl $_{2}$ to allow polymerization. The alginate-covered ECHO bag was then washed three times with ddH $_{2}$ O and functionalized on its surface with EDC-NHS by immersing it for 2 hours in a 1:1 solution containing 0.2 M EDC and 0.05 M NHS, then incubating it for 2 hours in 1 M Na $_{2}$ HPO $_{4}$ , followed by washing four times in ddH $_{2}$ O. This process was repeated three times. The EDC-NHS-alginate ECHO-bag was then immersed for 10 min in 0.1 M BSA, washed once in PBS, and UV-sterilized for 30 min. The final ECHO implant (30 mm by 17.5 mm by 10 mm, LWH) was stored at 4°C until required.

---

GF\_Au@Pt analysis

Scanning electron microscopy (SEM, FEI Versa 3D DualBeam) and scanning transmission electron microscopy (STEM, JEOL JEM-F200 cFEG) were used to analyze the structural properties of GF\_Au@Pt. X-ray photoelectron spectroscopy (XPS) was performed using a VG ESCALAB 210 system (VG Scientific Ltd., East Grinstead, UK). Infrared thermal imaging was performed using a FLIR E8 Pro (Teledyne FLIR LLC, USA). Elemental mapping was conducted using energy-dispersive x-ray spectroscopy (EDS, Zeiss GeminiSEM 450). Selected-area electron diffraction (SAED, JEM-F200(HRP)) was used for crystallographic characterization. Electrochemical analyses such as cyclic voltammetry (-0.2 to 0.2 V, 20 to 100 mV/s), linear sweep voltammetry (0 to 1.4 V, 100 mV/s), and electrical impedance spectroscopy (100 kHz to 0.1 Hz, 100 mV amplitude) were performed on a CH Instruments electrochemical workstation (CHI660) using GF\_Au@Pt, RHE as a reference and platinum as counter electrodes immersed in PBS (pH 7.4, 37°C and 1 M KOH, 21°C). Electrocatalysis dynamics of GF\_Au@Pt were recorded using GF\_Au@Pt as the anode and platinum as the cathode. The RND LAB DC Power Supply (Model: RND 320-KD3005D, China) was used to provide a stable DC power source during the electrocatalysis experiments.

---

Measurement of chlorine gas

Dissolved chlorine was measured using DPD-1 test strips (cat. no. 486637) with an eXact EZ photometer (cat. no. 486205) (ITS Europe).

---

ROS quantification

Electrostimulated DCINS cells were washed with PBS (300 $\mu$ l; cat. no. 14190-094, Thermo Fisher Scientific) and incubated for 45 min in serum-free Dulbecco's modified Eagle's medium (DMEM; 500 $\mu$ l) containing 25 $\mu$ mol liter $^{-1}$ 2',7'-dichlorofluorescein diacetate (cat. no. D6883, Sigma-Aldrich) (115). Cells were rinsed with PBS (300 $\mu$ l), and ROS levels were quantified fluorometrically (excitation/emission, 485/535 nm) using a Tecan Infinite 200 PRO plate reader.

---

Infrared thermal imaging

Infrared thermal imaging (FLIR E8 Pro, Teledyne FLIR LLC, USA) detected no measurable temperature increase ( $\pm0.1^{\circ}C$ ) after 2 hours of operation at 0.75 V and 10 mA in 10 ml DMEM (fig. S27).

---

Analytical characterization

LC-MS analysis: LC-MS measurements were performed on a Bruker Maxis 4G ESI-Q-TOF mass spectrometer coupled to a high-performance liquid chromatography system. Separations used a Reprospher 100 C18-Aqua column (2.0 mm by 125 mm, 5 $\mu$ m) maintained at 40°C. The mobile phase consisted of (i) H $_{2}$ O containing 0.1% formic acid and (ii) acetonitrile containing 0.1% formic acid. The gradient program was: 0 to 2 min, 10% B; 2 to 12 min, linear increase to 95% B; 12 to 18 min, 95% B; 18 to 22 min, return to 10% B. The flow rate was 0.5 ml min $^{-1}$ , and the injection volume was 1 $\mu$ l. The MS was operated in positive electrospray ionization (ESI $^{+}$ ) mode with the following settings: capillary voltage, 4500 V; nebulizer pressure, 2.0 bar; dry gas, 8 L min $^{-1}$ ; dry gas temperature, 200°C; scan range, mass/charge ratio 75 to 700. Data were processed using Bruker Compass DataAnalysis.

Nuclear magnetic resonance (NMR) spectroscopy: NMR spectra were acquired on a Varian 600 MHz spectrometer ( $^{1}$ H, 600 MHz; $^{13}$ C, 151 MHz) using a gradient-selected echo-antiecho HMBC pulse sequence. D-Glucose was electro-oxidized in an ECHO cell for 2 hours under aqueous conditions, and the reaction mixture was then redissolved in D $_{2}$ O for NMR analysis at 298 K. Spectra were collected over spectral widths of 10 ppm ( $^{1}$ H) and 220 ppm ( $^{13}$ C), using 2048 × 512 data points and 32 scans per increment. Chemical shifts were referenced to residual HDO ( $\delta$ H = 4.79 ppm). Data were processed in MestReNova (v14.2) with sine-bell apodization and zero-filling before Fourier transformation.

ICP-MS analysis: Gold ( $^{197}$ Au) and platinum ( $^{195}$ Pt) were quantified using an ICP-MS NexION 350D (PerkinElmer) operated in kinetic energy discrimination (KED) mode with helium as the collision gas. Yttrium [2 parts per billion (ppb)] was used as an internal standard. External calibration employed standards from 0.05 to 50 ppb. Limits of detection were 0.005 ppb for Au and 0.010 ppb for Pt. Serum samples (0.75 ml) were digested in concentrated HNO $_{3}$ (69%) using a DigiPREP Jr. heating block at 100°C for 55 min. Digests were diluted to 5 ml with Milli-Q water, then further diluted 20-fold in 2% HNO $_{3}$ before ICP-MS measurement.

---

Animal experiments

Mouse models: This study used two mouse models: type 2 diabetic (T2D), male BKS(D)- $Lepr^{db}$ /JOrlRj mice (6 to 12 weeks old, average weight $\sim$ 35 to 45 g, purchased from Janvier Labs), male and female BKS.Cg-Dock7m+/+Lepr $^{db}$ J (db/db) (6 to 10 weeks old; Charles River Laboratories, Germany) and WT, male and female C57BL/6JRJ WT mice (6 to 12 weeks old, 18 to 28 g, purchased from Janvier Labs). Experiments were performed predominantly in male mice to reduce variability associated with sex-dependent hormonal effects. For the experiments presented in fig. S28, blood glucose profiles during electrical stimulation and glucose tolerance tests (GTTs) were evaluated in female T2D mice. The glucose-lowering effects were comparable to those observed in male T2D mice (Fig. 4, C and E), and no significant sex-dependent differences were detected. The mice were housed in controlled conditions with a stable temperature ( $21^{\circ} \pm 2^{\circ}\mathrm{C}$ ), ambient environment and controlled humidity ( $55 \pm 10\%$ ). They were maintained on a 12-hour reverse light-dark cycle and provided unrestricted access to standard chow and water. Each experimental group consisted of randomly assigned mice ( $n = 8$ ).

Animal experimentation license: All experimental procedures were conducted following ethical guidelines established by the Swiss animal welfare legislation and approved by the Veterinary Office of the Canton Basel-Stadt, Switzerland (license number: 2996\_34477 and 3341\_38467) as well as the French Republic (project no. DR2018-40v5 and APAFIS no. 16753). Experiments were conducted by licensed researchers: P.G.R (license number: LTK 5507) and G. Charpin-El Hamri (license number: 69266309) at the ETH Zurich and/or at the University of Lyon, Institut Universitaire de Technologie.

Experimental procedures: To prepare for surgical interventions, the dorsal region of the mice was shaved, and the surgical area was disinfected with mild ethanol and betadine solution. To provide preoperative analgesia, a mixture of bupivacaine (0.25%) and lidocaine (0.5%) was administered locally at the implantation site before the procedure. Meloxicam (5 mg kg $^{-1}$ , subcutaneous) was additionally administered 30 min before surgery. General anesthesia was induced using 4% isoflurane to ensure the mice were fully anesthetized and maintained using 2% isoflurane during the surgical procedures to keep the animals in a stable, unconscious state while minimizing the potential for anesthetic-related complications. The ECHO device was surgically implanted in a subcutaneous pocket and the incision was closed using clinically approved sutures (110). After surgery, the animals were stabilized for 24 hours before conducting ECHO stimulation and blood glucose profiling. The wireless-powered ECHO device was stimulated at 0.75 V and 10 mA to confirm the performance of the catalyst compared with control groups. Then, voltage (0 to 0.9 V) and current (0 to 10 mA) parameters of the implanted ECHO device were standardized for blood-glucose oxidation by continuous stimulation for 2 hours. Subsequently, the optimized voltage (0.75 V) and current (10 mA) were used to standardize the stimulation time. To enable closed-loop control, the ECHO device was wirelessly connected to a commercial CGM system (FreeStyle Libre 3) (figs. S16 and S18 and movie S1). The electronic circuit interconnecting CGM and ECHO to achieve a bioelectronic closed-loop system is shown in fig. S16. To equip animals with this bioelectronic closed-loop system, the ECHO device was implanted on the dorsal side, and this was followed by a 24-hour recovery period before the dorsoventral area of the mice was shaved and sterilized prior to placement of the CGM device. A compatible CGM reader was used for continuous data acquisition and monitoring (FreeStyle Libre 3 reader). To establish closed-loop control between the CGM and ECHO devices, we calibrated and programmed the CGM reader to vibrate at high-normal levels of 10 mM or more interstitial glucose. These vibrations were captured by a vibration sensor (SW-420 Vibration Sensor Module; BerryBase.ch) and relayed via an Arduino Controller (Arduino Nano A000005 by ATmega328; DigiKey.ch) to switch on the aforementioned power generator and activate the ECHO implant with step-sized stimulation periods of 5 min to safely lower glucose levels without the risk of hypoglycemic excursions. After each control interval, the radio frequency generator was reactivated on the basis of feedback from the vibrational sensor connected to the CGM reader's vibration unit. The wireless vibration-communication interface was established because CGM communication is proprietary and not accessible for open-source modifications.

Glucose tolerance test: Mice were fasted for 6 hours before blood glucose measurements. For the GTT, mice received an intraperitoneal injection of glucose (1.5 g/kg). The mice were then stimulated with optimized electrical parameters (0.75 V, 10 mA), and the glycemic profiles were recorded at regular intervals using ContourNext test strips and the ContourNext ONE reader (Ascensia Diabetes Care; cat. nos. 84191451 and 85659367). For long-term analyses, electrical power (0.75V, 10 mA, 60 min) was used to stimulate the ECHO device for the indicated intervals. After stimulation, blood glucose was measured, and blood samples were collected for further analysis.

Blood collection and analysis: Blood samples were collected in microtainer serum separator tubes (Becton Dickinson, cat. no. 365967), centrifuged at 6000g for 10 min at 4°C, and stored at -80°C for further analysis. Blood metabolites were profiled using triglyceride assay (Abcam, cat. no. b65336), free fatty acid quantitation (Sigma, cat. no. MAK044), and ketone body assay kits (Sigma, cat. no. MAK134). Glucagon (Mercodia, cat. no. 10-1271-01) and insulin (Mercodia, cat. no. 10-1249-01) in the blood serum were measured using the corresponding ELISA kit. In vivo biocompatibility of the ECHO device was evaluated by quantifying the inflammatory cytokines interferon- $\gamma$ (IFN- $\gamma$ ), interleukin-6 (IL-6), and tumor necrosis factor- $\alpha$ (TNF- $\alpha$ ) levels using corresponding ELISA kits [IFN- $\gamma$ (cat. no. BMS606-2), IL-6 (cat. no. BMS603HS), and TNF- $\alpha$ (cat. no. BMS607HS), all from Thermo Fisher Scientific]. In open-loop experiments, glucose was measured from tail-vein blood samples. In closed-loop studies, glucose was monitored continuously using a CGM positioned near the implant, with periodic tail-vein sampling for validation. In contrast to insulin-based regulation, ECHO modulates glucose through direct electrocatalytic oxidation at the electrode surface, governed by transport processes, including blood flow and molecular diffusion.

Resin embedding and section preparation: Samples were processed by Morphisto GmbH (Offenbach am Main, Germany) according to protocol AF-086388. Specimens were rinsed in distilled water for 1 hour to remove residual fixative and sectioned using an EXAKT diamond band saw (Model 312). Samples were dehydrated at room temperature through graded ethanol (30 to 99%), followed by isopropanol treatment and xylene clearing. Preinfiltration was performed at 4°C using a 1:1 mixture of xylene and stabilized Technovit 9100 base solution (Kulzer/Morphisto GmbH), followed by infiltration in Technovit 9100 supplemented with Hardener I and final infiltration in Technovit 9100 containing Hardener I and PMMA powder at 4°C. Samples were embedded in freshly prepared Technovit 9100, degassed (\~200 mbar, 10 min), hermetically sealed to exclude oxygen, and polymerized at $-4^{\circ}$ to $-8^{\circ}\mathrm{C}$ for 48 hours. Polymerized blocks were off-gassed overnight. Ground sections were prepared using EXAKT cutting, grinding, and thin-section technology (Donath method) with an Exakt Micro Grinder 400 CS, Diamond Band Saw 300 CP, and 402 Precision Adhesive Press. Sections were subsequently deplasticized with xylene and 2-methoxyethyl acetate, rinsed in acetone, and rehydrated.

---

Histological staining and image acquisition

ECHO device sections: Rehydrated resin-embedded sections were stained with hematoxylin and eosin using an automated Leica Multistainer ST5020 system or manually with Paragon stain at $38^{\circ}$ C. Slides were dehydrated, cleared in xylene, and coverslipped with xylene-based CV Mount (Leica). After drying at $40^{\circ}$ C, slides were digitized using a ZEISS Axio Scan.Z1 slide scanner equipped with a $20\times$ objective under standardized acquisition settings.

Surrounding tissue sections: Peri-implant tissue samples embedded in paraffin blocks were sectioned using a Leica HistoCore AUTOCUT (5 $\mu$ m thick) (Leica Biosystems, Germany). Sections were stained with hematoxylin and eosin (Roth; cat. no. 9194.2) according to the manufacturers' protocols. Micrographs of the stained tissue were acquired using a Nikon Eclipse Ti2 microscope equipped with a color camera.

---

Immunofluorescence analysis

Peri-implant tissue surrounding the ECHO device was harvested at the designated endpoints, fixed in 4% paraformaldehyde, paraffin-embedded, and sectioned at 5 $\mu$ m using a motorized rotary microtome (Leica HistoCore AUTOCUT, Leica Biosystems, Germany). Sections were mounted on glass slides, dried, deparaffinized in xylene, and rehydrated through graded ethanol. Heat-mediated (95°C) antigen retrieval was performed in 10 mM citrate buffer (pH 6.0) using the EZ Retriever IR antigen retrieval system (Biogenex; cat. no. MW015-IR). After cooling, sections were permeabilized with 0.1% Triton X-100 and blocked with 5% normal goat serum for 1 hour at room temperature. Sections were incubated overnight at 4°C with rabbit primary antibodies diluted in blocking buffer: anti-CD68 recombinant monoclonal antibody (clone 230504G10; Cat# 83014-5-RR; Thermo Fisher Scientific; 1:500), anti-TNF- $\alpha$ polyclonal antibody (Cat# BS-0078R; Thermo Fisher Scientific; 1:500), and anti-TGF- $\beta$ 1 recombinant monoclonal antibody (clone PD00-17; Cat# MA5-44667; Thermo Fisher Scientific; 1:500). After washing with PBS, sections were incubated for 1 hour at room temperature with Alexa Fluor 488-conjugated goat anti-rabbit IgG (H+L), cross-adsorbed secondary antibody (Cat# A-11008; Thermo Fisher Scientific; 1:250), following the manufacturer's instructions. Stained samples were imaged using a Nikon Eclipse Ti2 microscope equipped with a color camera.

---

Jugular vein catheterization and hyperinsulinemic-euglycemic clamp assay

Hyperinsulinemic-euglycemic clamp assays were performed to assess device efficacy. For catheter implantation, mice received meloxicam (5 mg kg $^{-1}$ , subcutaneous) before isoflurane anesthesia. The right jugular vein was exposed and cannulated with a PE-10 catheter (C20PU-MJV2814; Instech Laboratories, Inc., Plymouth Meeting, PA, USA) filled with heparinized saline (50 U ml $^{-1}$ ). The catheter was advanced toward the right atrium, secured with silk ligatures, and tunneled subcutaneously to exit dorsally, and the incision was closed. Postoperative analgesia was maintained for 48 hours. For clamp experiments, mice were fasted for 5 to 6 hours with free access to water, and baseline blood glucose was measured from the tail vein. A priming insulin bolus (80 mU kg $^{-1}$ ) was administered through the jugular catheter, followed by a continuous glucose infusion using a precision syringe pump (Pump 11 Plus, Harvard Apparatus, Holliston, MA, USA). Blood glucose was measured every 10 min, and the glucose infusion rate (GIR) required to maintain stable glycemia was recorded. At the end of the experiment, infusions were stopped, catheters were flushed with heparinized saline, and animals were either euthanized or allowed to recover under warming conditions.

---

Statistics

Unless specified otherwise, all data presented are representative of two independent experiments done in triplicate. Details of data presentation, sample sizes of biological replicates $(n)$ , statistical analyses, and significance levels are provided in the figures or their respective legends. Statistical evaluations were conducted to assess the significance of differences between datasets two-way ANOVA either with Tukey's multiple comparisons test or Dunnett's multiple comparisons test (as mentioned in figure legends). Data processing and analysis were performed using Microsoft Excel (version 16.51) and GraphPad Prism (version 9.2.0).

---

REFERENCES AND NOTES

1. B. Chen et al., The interactions between energy homeostasis and neurovascular plasticity. Nat. Rev. Endocrinol. 20, 749–759 (2024). doi: 10.1038/s41574-024-01021-8; pmid: 39054359

2. GBD 2021 US Obesity Forecasting Collaborators, National-level and state-level prevalence of overweight and obesity among children, adolescents, and adults in the USA, 1990-2021, and forecasts up to 2050. Lancet 404, 2278–2298 (2024). doi: 10.1016/S0140-6736(24)01548-4; pmid: 39551059

3. NCD Risk Factor Collaboration (NCD-RisC), Worldwide trends in diabetes prevalence and treatment from 1990 to 2022: A pooled analysis of 1108 population-representative studies with 141 million participants. Lancet 404, 2077–2093 (2024). doi: 10.1016/S0140-6736(24)02317-1; pmid: 39549716

4. I. J. Neeland et al., Metabolic syndrome. Nat. Rev. Dis. Primers 10, 77 (2024). doi: 10.1038/s41572-024-00563-5; pmid: 39420195

5. A. C. T. Ng, V. Delgado, B. A. Borlaug, J. J. Bax, Diabesity: The combined burden of obesity and diabetes on heart disease and the role of imaging. Nat. Rev. Cardiol. 18, 291–304 (2021). doi: 10.1038/s41569-020-00465-5; pmid: 33188304

6. D. E. Coral et al., A phenome-wide comparative analysis of genetic discordance between obesity and type 2 diabetes. Nat. Metab. 5, 237–247 (2023). doi: 10.1038/s42255-022-00731-5; pmid: 36703017

7. P. V. Piazza, D. Cota, G. Marsicano, The CB1 receptor as the cornerstone of exostasis. Neuron 93, 1252–1274 (2017). doi: 10.1016/j.neuron.2017.02.002; pmid: 28334603

8. D. S. Ludwig et al., Competing paradigms of obesity pathogenesis: Energy balance versus carbohydrate-insulin models. Eur. J. Clin. Nutr. 76, 1209–1221 (2022). doi: 10.1038/s41430-022-01179-2; pmid: 35896818

9. J. R. Speakman, K. D. Hall, Carbohydrates, insulin, and obesity. Science 372, 577–578 (2021). doi: 10.1126/science.aav0448; pmid: 33958466

10. D. Yach, D. Stuckler, K. D. Brownell, Epidemiologic and economic consequences of the global epidemics of obesity and diabetes. Nat. Med. 12, 62–66 (2006). doi: 10.1038/nm0106-62; pmid: 16397571

11. V. S. Malik, W. C. Willett, F. B. Hu, Global obesity: Trends, risk factors and policy implications. Nat. Rev. Endocrinol. 9, 13–27 (2013). doi: 10.1038/nrendo.2012.199; pmid: 23165161

12. T. D. Müller, M. Blüher, M. H. Tschöp, R. D. DiMarchi, Anti-obesity drug discovery: Advances and challenges. Nat. Rev. Drug Discov. 21, 201–223 (2022). doi: 10.1038/s41573-021-00337-8; pmid: 34815532

13. C. Oggioni, J. Lara, J. C. Wells, K. Soroka, M. Siervo, Shifts in population dietary patterns and physical inactivity as determinants of global trends in the prevalence of diabetes: An ecological analysis. Nutr. Metab. Cardiovasc. Dis. 24, 1105–1111 (2014). doi: 10.1016/j.numecd.2014.05.005; pmid: 24954422

14. E. W. Gregg et al., The burden of diabetes-associated multiple long-term conditions on years of life spent and lost. Nat. Med. 30, 2830–2837 (2024). doi: 10.1038/s41591-024-03123-2; pmid: 39090411

15. D. Tomic, J. E. Shaw, D. J. Magliano, The burden and risks of emerging complications of diabetes mellitus. Nat. Rev. Endocrinol. 18, 525–539 (2022). doi: 10.1038/s41574-022-00690-7; pmid: 35668219

16. R. A. Ajjan et al., Continuous glucose monitoring for the routine care of type 2 diabetes mellitus. Nat. Rev. Endocrinol. 20, 426–440 (2024). doi: 10.1038/s41574-024-00973-1; pmid: 38589493

17. A. B. Daly et al., Fully automated closed-loop insulin delivery in adults with type 2 diabetes: An open-label, single-center, randomized crossover trial. Nat. Med. 29, 203–208 (2023). doi: 10.1038/s41591-022-02144-z; pmid: 36631592

18. R. Hovorka, Closed-loop insulin delivery: From bench to clinical practice. Nat. Rev. Endocrinol. 7, 385–395 (2011). doi: 10.1038/nrendo.2011.32; pmid: 21343892

19. Y. Liu, L. Yang, Y. Cui, A wearable, rapidly manufacturable, stability-enhancing microneedle patch for closed-loop diabetes management. Microsyst. Nanoeng. 10, 112 (2024). doi: 10.1038/s41378-024-00663-y; pmid: 39166137

20. E. Vargas, P. Nandhakumar, S. Ding, T. Saha, J. Wang, Insulin detection in diabetes mellitus: Challenges and new prospects. Nat. Rev. Endocrinol. 19, 487–495 (2023). doi: 10.1038/s41574-023-00842-3; pmid: 37217746

21. T. Kitada, B. DiAndreth, B. Teague, R. Weiss, Programming gene and engineered-cell therapies with synthetic biology. Science 359, eaad1067 (2018). doi: 10.1126/science.aad1067; pmid: 29439214

22. T. Battelino et al., Efficacy of once-weekly tirzepatide versus once-daily insulin degludec on glycaemic control measured by continuous glucose monitoring in adults with type 2 diabetes (SURPASS-3 CGM): A substudy of the randomised, open-label, parallel-group, phase 3 SURPASS-3 trial. Lancet Diabetes Endocrinol. 10, 407–417 (2022). doi: 10.1016/S2213-8587(22)00077-8; pmid: 35468321

23. P. H. Reis-Barbosa, I. Marcondes-de-Castro, T. S. Marinho, M. B. Aguila, C. A. Mandarim-de-Lacerda, The dual glucose-dependent insulinotropic polypeptide/glucagon-like peptide-1 receptor agonist tirzepatide effects on body weight evolution, adiponectin, insulin and leptin levels in the combination of obesity, type 2 diabetes and menopause in mice. Diabetes Obes. Metab. 26, 4613–4621 (2024). doi: 10.1111/dom.15820; pmid: 39113264

24. A. P. Teixeira, M. Fussenegger, Synthetic gene circuits for regulation of next-generation cell-based therapeutics. Adv. Sci. 11, e2309088 (2024). doi: 10.1002/advs.202309088; pmid: 38126677

25. A. P. Teixeira, M. Fussenegger, Synthetic macromolecular switches for precision control of therapeutic cell functions. Nat. Rev. Bioeng. 2, 1005–1022 (2024). doi: 10.1038/s44222-024-00235-9

26. M. P. Valdecantos et al., The dual GLP-1/glucagon receptor agonist G49 mimics bariatric surgery effects by inducing metabolic rewiring and inter-organ crosstalk. Nat. Commun. 15, 10342 (2024). doi: 10.1038/s41467-024-54080-w; pmid: 39609390

27. A. Bertschi et al., Controlling therapeutic protein expression via inhalation of a butter flavor molecule. Nucleic Acids Res. 51, e28 (2023). doi: 10.1093/nar/gkac1256; pmid: 36625292

28. M. Müller et al., Designed cell consortia as fragrance-programmable analog-to-digital converters. Nat. Chem. Biol. 13, 309–316 (2017). doi: 10.1038/nchembio.2281; pmid: 28092361

29. H. Wang, M. Xie, G. Charpin-El Hamri, H. Ye, M. Fussenegger, Treatment of chronic pain by designer cells controlled by spearmint aromatherapy. Nat. Biomed. Eng. 2, 114–123 (2018). doi: 10.1038/s41551-018-0192-3; pmid: 31015627

30. W. Weber et al., Gas-inducible transgene expression in mammalian cells and mice. Nat. Biotechnol. 22, 1440–1444 (2004). doi: 10.1038/nbt1021; pmid: 15502819

31. M. Mansouri et al., Smart-watch-programmed green-light-operated percutaneous control of therapeutic transgenes. Nat. Commun. 12, 3388 (2021). doi: 10.1038/s41467-021-23572-4; pmid: 34099676

32. M. Mansouri et al., Smartphone-flashlight-mediated remote control of rapid insulin secretion restores glucose homeostasis in experimental type-1 diabetes. Small 17, e2101939 (2021). doi: 10.1002/smll.202101939; pmid: 34227232

33. H. Ye, M. Daoud-El Baba, R. W. Peng, M. Fussenegger, A synthetic optogenetic transcription device enhances blood-glucose homeostasis in mice. Science 332, 1565–1568 (2011). doi: 10.1126/science.1203535; pmid: 21700876

34. Y. Zhou et al., A small and highly sensitive red/far-red optogenetic switch for applications in mammals. Nat. Biotechnol. 40, 262–272 (2022). doi: 10.1038/s41587-021-01036-w; pmid: 34608325

35. D. Maity, P. Guha Ray, P. Buchmann, M. Mansouri, M. Fussenegger, Blood-glucose-powered metabolic fuel cell for self-sufficient bioelectronics. Adv. Mater. 35, e2300890 (2023). doi: 10.1002/adma.202300890; pmid: 36893359

36. M. Xie et al., $\beta$ -cell–mimetic designer cells provide closed-loop glycemic control. Science 354, 1296–1301 (2016). doi: 10.1126/science.aaf4006; pmid: 27940875

37. D. Ausländer et al., A synthetic multifunctional mammalian pH sensor and $CO_{2}$ transgene-control device. Mol. Cell 55, 397–408 (2014). doi: 10.1016/j.molcel.2014.06.007; pmid: 25018017

38. Y. Liu, G. Charpin-El Hamri, H. Ye, M. Fussenegger, A synthetic free fatty acid-regulated transgene switch in mammalian cells and mice. Nucleic Acids Res. 46, 9864–9874 (2018). doi: 10.1093/nar/gky805; pmid: 30219861

39. K. Rössger, G. Charpin-El Hamri, M. Fussenegger, A closed-loop synthetic gene circuit for the treatment of diet-induced obesity in mice. Nat. Commun. 4, 2825 (2013). doi: 10.1038/ncomms3825; pmid: 24281397

40. H. Ye et al., Pharmaceutically controlled designer circuit for the treatment of the metabolic syndrome. Proc. Natl. Acad. Sci. U.S.A. 110, 141–146 (2013). doi: 10.1073/pnas.1216801110; pmid: 23248313

41. C. J. Bashor, I. B. Hilton, H. Bandukwala, D. M. Smith, O. Veiseh, Engineering the next generation of cell-based therapeutics. Nat. Rev. Drug Discov. 21, 655–675 (2022). doi: 10.1038/s41573-022-00476-6; pmid: 35637318

42. J. E. Eyles et al., Cell therapy products: Focus on issues with manufacturing and quality control of chimeric antigen receptor T-cell therapies. J. Chem. Technol. Biotechnol. 94, 1008–1016 (2019). doi: 10.1002/jctb.5829

43. M. R. Stetten, D. Stetten Jr., The metabolism of gluconic acid. J. Biol. Chem. 187, 241–252 (1950). doi: 10.1016/S0021-9258(19)50948-6; pmid: 14794708

44. P. Zhang, G. Zhang, X. Wan, Challenges and new technologies in adoptive cell therapy. J. Hematol. Oncol. 16, 97 (2023). doi: 10.1186/s13045-023-01492-8; pmid: 37596653

45. B. A. Borlaug et al., Semaglutide in HFpEF across obesity class and by body weight reduction: A prespecified analysis of the STEP-HFpEF trial. Nat. Med. 29, 2358–2365 (2023). doi: 10.1038/s41591-023-02526-x; pmid: 37635157

46. A. M. Lincoff et al., Semaglutide and cardiovascular outcomes in obesity without diabetes. N. Engl. J. Med. 389, 2221–2232 (2023). doi: 10.1056/NEJMoa2307563; pmid: 37952131

47. D. H. Ryan et al., Long-term weight loss effects of semaglutide in obesity without diabetes in the SELECT trial. Nat. Med. 30, 2049–2057 (2024). doi: 10.1038/s41591-024-02996-7; pmid: 38740993

48. M. Ala, R. Mohammad Jafari, A. R. Dehpour, M. Poursalehian, Tirzepatide outcompetes long-acting insulin in managing type 2 diabetes: A meta-analysis of three phase 3 randomized controlled trials. Int. J. Obes. 48, 1684–1695 (2024). doi: 10.1038/s41366-024-01621-4; pmid: 39210008

49. L. Gao et al., Tirzepatide versus insulin glargine as second-line or third-line therapy in type 2 diabetes in the Asia-Pacific region: The SURPASS-AP-Combo trial. Nat. Med. 29, 1500–1510 (2023). doi: 10.1038/s41591-023-02344-1; pmid: 37231074

50. C. M. Kusminski, P. E. Bickel, P. E. Scherer, Targeting adipose tissue in the treatment of obesity-associated diabetes. Nat. Rev. Drug Discov. 15, 639–660 (2016). doi: 10.1038/nrd.2016.75; pmid: 27256476

51. A. N. Webster et al., Molecular connectomics reveals a glucagon-like peptide 1-sensitive neural circuit for satiety. Nat. Metab. 6, 2354–2373 (2024). doi: 10.1038/s42255-024-01168-8; pmid: 39627618

52. F. Zhao et al., Structural insights into multiplexed pharmacological actions of tirzepatide and peptide 20 at the GIP, GLP-1 or glucagon receptors. Nat. Commun. 13, 1057 (2022). doi: 10.1038/s41467-022-28683-0; pmid: 35217653

53. E. M. Apperloo et al., Semaglutide in patients with overweight or obesity and chronic kidney disease without diabetes: A randomized double-blind placebo-controlled clinical trial. Nat. Med. 31, 278–285 (2025). doi: 10.1038/s41591-024-03327-6; pmid: 39455729

54. B. A. Borlaug et al., Effects of tirzepatide on circulatory overload and end-organ damage in heart failure with preserved ejection fraction and obesity: A secondary analysis of the SUMMIT trial. Nat. Med. 31, 544–551 (2025). doi: 10.1038/s41591-024-03374-z; pmid: 39551891

55. H. M. Colhoun et al., Long-term kidney outcomes of semaglutide in obesity and cardiovascular disease in the SELECT trial. Nat. Med. 30, 2058–2066 (2024). doi: 10.1038/s41591-024-03015-5; pmid: 38796653

56. M. G. Savelieff et al., The global and regional burden of diabetic peripheral neuropathy. Nat. Rev. Neurol. 21, 17–31 (2025). doi: 10.1038/s41582-024-01041-y; pmid: 39639140

57. W. J. Valentine, M. Hoog, R. Mody, M. Belger, R. Pollock, Long-term cost-effectiveness analysis of tirzepatide versus semaglutide 1.0 mg for the management of type 2 diabetes in the United States. Diabetes Obes. Metab. 25, 1292–1300 (2023). doi: 10.1111/dom.14979; pmid: 36655340

58. W. Wang et al., Association of semaglutide with risk of suicidal ideation in a real-world cohort. Nat. Med. 30, 168–176 (2024). doi: 10.1038/s41591-023-02672-2; pmid: 38182782

59. R. Wang et al., Crystal OH mediating pathway for hydrogen peroxide production via two-electron water oxidation in non-carbonate electrolytes. Nat. Commun. 15, 10456 (2024). doi: 10.1038/s41467-024-54593-4; pmid: 39622859

60. M. Yang et al., A self-reactivated PdCu catalyst for aldehyde electro-oxidation with anodic hydrogen production. Nat. Commun. 15, 9852 (2024). doi: 10.1038/s41467-024-54286-y; pmid: 39543188

61. N. Wang et al., Contact-electro-catalytic $CO_{2}$ reduction from ambient air. Nat. Commun. 15, 5913 (2024). doi: 10.1038/s41467-024-50118-1; pmid: 39003260

62. K. Ye et al., Molecular level insights on the pulsed electrochemical $CO_{2}$ reduction. Nat. Commun. 15, 9781 (2024). doi: 10.1038/s41467-024-54122-3; pmid: 39532852

63. L. Chen et al., Structural analysis and intrinsic enzyme mimicking activities of ligand-free PtAg nanoalloys. Small 19, e2206772 (2023). doi: 10.1002/smll.202206772; pmid: 36755199

64. M. A. Sayeed, A. P. O'Mullane, A multifunctional gold doped Co(OH) $_{2}$ electrocatalyst tailored for water oxidation, oxygen reduction, hydrogen evolution and glucose detection. J. Mater. Chem. A 5, 23776–23784 (2017). doi: 10.1039/C7TA08928G

65. Z. Tian et al., Selective photoelectrochemical oxidation of glucose to glucaric acid by single atom Pt decorated defective $TiO_{2}$ . Nat. Commun. 14, 142 (2023). doi: 10.1038/s41467-023-35875-9; pmid: 36627303

66. J. Chen et al., Glucose-oxidase like catalytic mechanism of noble metal nanozymes. Nat. Commun. 12, 3375 (2021). doi: 10.1038/s41467-021-23737-1; pmid: 34099730

67. C. Gyan-Barimah et al., High vacancy formation energy boosts the stability of structurally ordered PtMg in hydrogen fuel cells. Nat. Commun. 15, 7034 (2024). doi: 10.1038/s41467-024-51280-2; pmid: 39147744

68. W. Luo et al., Self-catalyzed, self-limiting growth of glucose oxidase-mimicking gold nanoparticles. ACS Nano 4, 7451–7458 (2010). doi: 10.1021/nn102592h; pmid: 21128689

69. X. Shen et al., Mechanisms of oxidase and superoxide dismutation-like activities of gold, silver, platinum, and palladium, and their alloys: A general way to the activation of molecular oxygen. J. Am. Chem. Soc. 137, 15882–15891 (2015). doi: 10.1021/jacs.5b10346; pmid: 26642084

70. Z. Wang et al., Natural oxidase-mimicking copper-organic frameworks for targeted identification of ascorbate in sensitive sweat sensing. Nat. Commun. 14, 69 (2023). doi: 10.1038/s41467-022-35721-4; pmid: 36604444

71. J. Wu et al., Nanomaterials with enzyme-like characteristics (nanozymes): Next-generation artificial enzymes (II). Chem. Soc. Rev. 48, 1004–1076 (2019). doi: 10.1039/C8CS00457A; pmid: 30534770

72. H. Zhang, L. Wang, L. Lu, N. Toshima, Preparation and catalytic activity for aerobic glucose oxidation of crown jewel structured Pt/Au bimetallic nanoclusters. Sci. Rep. 6, 30752 (2016). doi: 10.1038/srep30752; pmid: 27476577

73. D. A. Antonetti, P. S. Silva, A. W. Stitt, Current understanding of the molecular and cellular pathology of diabetic retinopathy. Nat. Rev. Endocrinol. 17, 195–206 (2021). doi: 10.1038/s41574-020-00451-4; pmid: 33469209

74. F. G. De Felice, R. A. Gonçalves, S. T. Ferreira, Impaired insulin signalling and allostatic load in Alzheimer disease. Nat. Rev. Neurosci. 23, 215–230 (2022). doi: 10.1038/s41583-022-00558-9; pmid: 35228741

75. R. I. Holt, M. de Groot, S. H. Golden, Diabetes and depression. Curr. Diab. Rep. 14, 491 (2014). doi: 10.1007/s11892-014-0491-3; pmid: 24743941

76. N. B. Lister et al., Child and adolescent obesity. Nat. Rev. Dis. Primers 9, 24 (2023). doi: 10.1038/s41572-023-00435-4; pmid: 37202378

77. D. M. Muoio, C. B. Newgard, Molecular and metabolic mechanisms of insulin resistance and $\beta$ -cell failure in type 2 diabetes. Nat. Rev. Mol. Cell Biol. 9, 193–205 (2008). doi: 10.1038/nrm2327; pmid: 18200017

78. J. Pearson-Stuttard et al., Type 2 diabetes and cancer: An umbrella review of observational and Mendelian randomization studies. Cancer Epidemiol. Biomarkers Prev. 30, 1218–1228 (2021). doi: 10.1158/1055-9965.EPI-20-1245; pmid: 33737302

79. M. Sun, W. M. Chen, S. Y. Wu, J. Zhang, Metformin in elderly type 2 diabetes mellitus: Dose-dependent dementia risk reduction. Brain 147, 1474–1482 (2024). doi: 10.1093/brain/awad366; pmid: 37878862

80. D. H. van Raalte et al., Combination therapy for kidney disease in people with diabetes mellitus. Nat. Rev. Nephrol. 20, 433–446 (2024). doi: 10.1038/s41581-024-00827-z; pmid: 38570632

81. M. R. Stetten, Y. J. Topper, Pathways from gluconic acid to glucose in vivo. J. Biol. Chem. 203, 653–664 (1953). doi: 10.1016/S0021-9258(19)52335-3; pmid: 13084635

82. A. P. Teixeira, S. Xue, J. Huang, M. Fussenegger, Evolution of molecular switches for regulation of transgene expression by clinically licensed gluconate. Nucleic Acids Res. 51, e85 (2023). doi: 10.1093/nar/gkad600; pmid: 37497781

83. J. Fuchs, R. Hovorka, Closed-loop control in insulin pumps for type-1 diabetes mellitus: Safety and efficacy. Expert Rev. Med. Devices 17, 707–720 (2020). doi: 10.1080/17434440.2020.1784724; pmid: 32569476

84. J. R. Gruber et al., Impact of blood glucose on cognitive function in insulin resistance: Novel insights from ambulatory assessment. Nutr. Diabetes 14, 74 (2024). doi: 10.1038/s41387-024-00331-0; pmid: 39261457

85. D. Kleiman, M. Littor, M. Nawas, R. Ben-Haroush Schyr, D. Ben-Zvi, Simple continuous glucose monitoring in freely moving mice. J. Vis. Exp. 192, e64743 (2023). doi:10.3791/64743; prmid: 36912552

86. L. Morales-Dopico, S. A. MacLeish, Expanding the horizon of continuous glucose monitoring into the future of pediatric medicine. Pediatr. Res. 96, 1464–1474 (2024). doi: 10.1038/s41390-024-03573-x; pmid: 39306610

87. A. Grattoni et al., Harnessing cellular therapeutics for type 1 diabetes mellitus: Progress, challenges, and the road ahead. Nat. Rev. Endocrinol. 21, 14–30 (2025). doi: 10.1038/s41574-024-01029-0; pmid: 39227741

88. F. Mauvais-Jarvis, Sex differences in energy metabolism: Natural selection, mechanisms and consequences. Nat. Rev. Nephrol. 20, 56–69 (2024). doi: 10.1038/s41581-023-00781-2; pmid: 37923858

89. T. Mason et al., Metformin in gestational diabetes: Physiological actions and clinical applications. Nat. Rev. Endocrinol. 21, 77–91 (2025). doi: 10.1038/s41574-024-01049-w; pmid: 39455749

90. B. Bouvard, G. Mabilleau, Gut hormones and bone homeostasis: Potential therapeutic implications. Nat. Rev. Endocrinol. 20, 553–564 (2024). doi: 10.1038/s41574-024-01000-z; pmid: 38858581

91. A. Liskiewicz et al., Glucose-dependent insulinotropic polypeptide regulates body weight and food intake via GABAergic neurons in mice. Nat. Metab. 5, 2075–2085 (2023). doi: 10.1038/s42255-023-00931-7; pmid: 37946085

92. H. Tilg, G. Ianiro, A. Gasbarrini, T. E. Adolph, Adipokines: Masterminds of metabolic inflammation. Nat. Rev. Immunol. 25, 250–265 (2025). doi: 10.1038/s41577-024-01103-8; pmid: 39511425

93. M. Chakhtoura et al., Pharmacotherapy of obesity: An update on the available medications and drugs under investigation. EClinicalMedicine 58, 101882 (2023). doi: 10.1016/j.eclinm.2023.101882; pmid: 36992862

94. The Lancet, Treating obesity and diabetes: Drugs alone are not enough. Lancet 403, 1 (2024). doi: 10.1016/S0140-6736(24)00003-5; pmid: 38184328

95. W. Weber, M. Fussenegger, Emerging biomedical applications of synthetic biology. Nat. Rev. Genet. 13, 21–35 (2011). doi: 10.1038/nrg3094; pmid: 22124480

96. M. Xie, M. Fussenegger, Designing cell function: Assembly of synthetic gene circuits for cell biology applications. Nat. Rev. Mol. Cell Biol. 19, 507–525 (2018). doi: 10.1038/s41580-018-0024-z; pmid: 29858606

97. D. Bojar, L. Scheller, G. C. Hamri, M. Xie, M. Fussenegger, Caffeine-inducible gene switches controlling experimental diabetes. Nat. Commun. 9, 2318 (2018). doi: 10.1038/s41467-018-04744-1; pmid: 29921872

98. J. Dowarah, V. P. Singh, Anti-diabetic drugs recent approaches and advancements. Bioorg. Med. Chem. 28, 115263 (2020). doi: 10.1016/j.bmc.2019.115263; pmid: 32008883

99. M. Fussenegger et al., Streptogramin-based gene regulation systems for mammalian cells. Nat. Biotechnol. 18, 1203–1208 (2000). doi: 10.1038/81208; pmid: 11062442

100. M. Gitzinger, C. Kemmer, M. D. El-Baba, W. Weber, M. Fussenegger, Controlling transgene expression in subcutaneous implants using a skin lotion containing the apple metabolite phloretin. Proc. Natl. Acad. Sci. U.S.A. 106, 10638–10643 (2009). doi: 10.1073/pnas.0901501106; pmid: 19549857

101. M. Gitzinger et al., The food additive vanillic acid controls transgene expression in mammalian cells and mice. Nucleic Acids Res. 40, e37 (2012). doi: 10.1093/nar/gkr1251; pmid: 22187155

102. J. Huang et al., Aspirin-responsive gene switch regulating therapeutic protein expression. Nat. Commun. 16, 2028 (2025). doi: 10.1038/s41467-025-57275-x; pmid: 40016240

103. M. Mahameed, S. Xue, B. A. Stefanov, G. C. Hamri, M. Fussenegger, Engineering a rapid insulin release system controlled by oral drug administration. Adv. Sci. 9, e2105619 (2022). doi: 10.1002/advs.202105619; pmid: 35048556

104. W. Weber et al., A synthetic time-delay circuit in mammalian cells and mice. Proc. Natl. Acad. Sci. U.S.A. 104, 2643–2648 (2007). doi: 10.1073/pnas.0606398104; pmid: 17296937

105. X. Wu et al., AAV-delivered muscone-induced transgene system for treating chronic diseases in mice via inhalation. Nat. Commun. 15, 1122 (2024). doi: 10.1038/s41467-024-45383-z; pmid: 38321056

106. T. Yang et al., Ligand-switchable nanoparticles resembling viral surface for sequential drug delivery and improved oral insulin therapy. Nat. Commun. 13, 6649 (2022). doi: 10.1038/s41467-022-34357-8; pmid: 36333321

107. M. Mahameed et al., Nitroglycerin-responsive gene switch for the on-demand production of therapeutic proteins. Nat. Biomed. Eng. 9, 1129–1143 (2025). doi: 10.1038/s41551-025-01350-7; pmid: 39953326

108. D. Maity, P. Guha Ray, M. Fussenegger, Glucose-Operated Widget (GLOW) for closed-loop optogenetic glycemic control. Adv. Mater. 36, e2408537 (2024). doi: 10.1002/adma.202408537; pmid: 39210629

109. P. Bai et al., A fully human transgene switch to regulate therapeutic protein production by cooling sensation. Nat. Med. 25, 1266–1273 (2019). doi: 10.1038/s41591-019-0501-8; pmid: 31285633

110. B. A. Stefanov et al., Genetically encoded protein thermometer enables precise electrothermal control of transgene expression. Adv. Sci. 8, e2101813 (2021). doi: 10.1002/advs.202101813; pmid: 34496151

111. H. Zhao, S. Xue, M. D. Hussherr, A. P. Teixeira, M. Fussenegger, Autonomous push button-controlled rapid insulin release from a piezoelectrically activated subcutaneous cell implant. Sci. Adv. 8, eabm4389 (2022). doi: 10.1126/sciadv.abm4389; pmid: 35704573

112. F. Ivanovski et al., Ultrasound-mediated spatial and temporal control of engineered cells in vivo. Nat. Commun. 15, 7369 (2024). doi: 10.1038/s41467-024-51620-2; pmid: 39191796

113. H. Zhao et al., Tuning of cellular insulin release by music for real-time diabetes control. Lancet Diabetes Endocrinol. 11, 637–640 (2023). doi: 10.1016/S2213-8587(23)00153-5; pmid: 37620062

114. K. Krawczyk et al., Electrogenetic cellular insulin release for real-time glycemic control in type 1 diabetic mice. Science 368, 993–1001 (2020). doi: 10.1126/science.aau7187; pmid: 32467389

115. J. Huang, S. Xue, P. Buchmann, A. P. Teixeira, M. Fussenegger, An electrogenetic interface to program mammalian gene expression by direct current. Nat. Metab. 5, 1395–1407 (2023). doi: 10.1038/s42255-023-00850-7; pmid: 37524785

116. Z. Lin, P. Guha Ray, J. Huang, P. Buchmann, M. Fussenegger, Electromagnetic wireless remote control of mammalian transgene expression. Nat. Nanotechnol. 20, 1071–1078 (2025). doi: 10.1038/s41565-025-01929-w; pmid: 40325210

117. M. Folcher et al., Mind-controlled transgene expression by a wireless-powered optogenetic designer cell implant. Nat. Commun. 5, 5392 (2014). doi: 10.1038/ncomms6392; pmid: 25386727

118. J. Shao et al., Smartphone-controlled optogenetically engineered cells enable semiautomatic glucose homeostasis in diabetic mice. Sci. Transl. Med. 9, eaal2298 (2017). doi: 10.1126/scitranslmed.aal2298; pmid: 28446682

119. X. Li et al., Advances in heterogeneous single-cluster catalysis. Nat. Rev. Chem. 7, 754–767 (2023). doi: 10.1038/s41570-023-00540-8; pmid: 37814032

120. R. Luque, Z. A. ALOthman, A. M. Balu, L. Voskressensky, Heterogeneous catalysis to drive the waste-to-pharma concept: From furanics to active pharmaceutical ingredients. Molecules 26, 6738 (2021). doi: 10.3390/molecules26216738; pmid: 34771146

121. Y. Zhang et al., Nanozymes for nanohealthcare. Nat. Rev. Methods Primers 4, 36 (2024). doi: 10.1038/s43586-024-00315-5

122. D. S. Popovic et al., Tirzepatide use and the risk of cancer among individuals with type 2 diabetes mellitus: A meta-analysis of randomized controlled trials. Diabetes Res. Clin. Pract. 213, 111758 (2024). doi: 10.1016/j.diabres.2024.111758; pmid: 38925294

---

ACKNOWLEDGMENTS

We thank H. Zulewski, A. P. Teixeira, and A. Hierholzer for generous advice. We thank M. Pfeffer (Department of Chemistry, University of Basel) for LC-MS analyses, I. Cornu (University of Basel) for assistance with NMR measurements, and N. Gasilova (EPFL, ISIC-GE-VS, Sion) for performing ICP-MS analyses. We thank M.-D. Hussherr (D-BSSE Animal Facility, ETH Zurich) for expertise in performing jugular vein catheterization and surgical procedures. Funding: This work was financially supported through a European Research Council advanced grant (ElectroGene, no. 785800; M.F.) and in part by the Swiss National Centre of Competence in Research (NCCR) for Molecular Systems Engineering (M.F.). Author contributions: D.M. and M.F. designed the project. D.M. and M.F. analyzed the results and wrote the manuscript. D.M. conducted the in vitro work, and P.G.R. and G.C.-E.H. performed the animal experiments. D.M. and P.B. designed ECHO and fabricated the wireless power circuit. Competing interests: The authors declare that they have no competing interests. Data, code, and materials availability: All data supporting the findings of this study are available in the paper and its supplementary materials. Original data are provided in data S1. The materials generated during this study, including the GF\_Au@Pt electrocatalyst and the ECHO implant components, are available from the corresponding author upon reasonable request. License information: Copyright © 2026 the authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original US government works. https://www.science.org/about/science-licenses-journal-article-reuse. This research was funded in whole or in part by the European Research Council (785800); as required, the author will make the Author Accepted Manuscript (AAM) version available under a CC BY public copyright license.

---

SUPPLEMENTARY MATERIALS

science.org/doi/10.1126/science.ady8560

Figs. S1 to S28; Tables S1 and S2; MDAR Reproducibility Checklist; Movies S1 and S2; Data S1 Submitted 8 May 2025; resubmitted 30 May 2026; accepted 31 August 2026

10.1126/science.ady8560

BAK and BAX form a multiprotein complex to drive mtDNA efflux and caspase-independent cell death

Full article and list of author affiliations: https://doi.org/10.1126/science.aea2181

Shaoyuan Yang†, Hyungjin Kim†, Guanghong Liao, Na Ta, Ho-Chou Tu, Shriram Ramani, Chih-Yu Chen, Zih-Yu Li, Nobuhiro Tanno, James J. Hsieh, Emily H. Cheng\*

INTRODUCTION: Mitochondria have a pivotal role in mammalian cell death. In response to apoptotic signals, activator BH3-only proteins activate BAX and BAK, leading to mitochondrial outer membrane permeabilization (MOMP), cytochrome c release, apoptosome assembly, and caspase-mediated apoptosis. In cells lacking apoptosome function, activation of BAX and BAK instead leads to sustained release of mitochondrial DNA (mtDNA) into the cytosol, activation of cGAS-STING-type I interferon signaling, and caspase-independent cell death. The apoptotic caspase cascade normally suppresses this innate immune pathway, rendering apoptosis immunologically silent. Given that impaired apoptosome function is common in human cancers, BAX-BAK activation-induced mtDNA efflux and immunogenic cell death may provide a new avenue for cancer therapy. However, the molecular and biochemical basis of mtDNA efflux was not understood.

RATIONALE: BAX-BAK activation-induced mtDNA efflux has been attributed to MOMP-associated mitochondrial inner membrane (MIM) herniation. After cytochrome c efflux, BAX and BAK coalesce to form macropores in the mitochondrial outer membrane (MOM), through which the MIM herniates and subsequently loses membrane integrity, resulting in mtDNA efflux. However, only a minority of herniated MIM structures lose membrane integrity, suggesting that mtDNA efflux may also involve an active biochemical process. We therefore used biochemical and genetic studies to determine whether BAK and BAX engage distinct protein complexes that differentially regulate cytochrome c and mtDNA efflux. We further examined this process in human cancer cells with impaired apoptosome function.

RESULTS: BAK and BAX, once activated by activator BH3-only proteins, assembled into two functionally distinct mitochondrial complexes. The machinery responsible for cytochrome c release probably consisted of a BAK or BAX hexamer. By contrast, we purified a preexisting 600-kD complex comprising BAK, VDAC1, VDAC2, PHB1, and PHB2, which may localize at contact sites between the MOM and MIM. Activation by BH3-only proteins appears to remodel this complex and promote recruitment of cytosolic BAX to form mtDNA-releasing machinery, which directly released mtDNA from mitoplasts. This remodeling was associated with VDAC channel closure and increased mitochondrial calcium influx. Loss of VDAC1, VDAC2, or prohibitins impaired mtDNA efflux and caspase-independent cell death without affecting MOMP or MIM herniation. Depletion of mtDNA also protected cells from this form of death, supporting a contribution of cytosolic mtDNA to cellular demise. We designated this BAX-BAK activation-induced caspase-independent cell death “baciptosis.” Baciptosis displayed necrotic morphology, lacked widespread genomic DNA fragmentation, and was accompanied by a progressive decline in

Graphical model illustrating distinct BAK-BAX complexes mediating cytochrome c and mtDNA efflux. Activator BH3-only proteins activate BAK and BAX to form a hexameric complex that mediates cytochrome c release. In parallel, they remodel a preexisting 600-kD complex containing BAK, VDAC1, VDAC2, PHB1, and PHB2, and recruit cytosolic BAX to generate mtDNA-releasing machinery. The model highlights the mechanistic separation between cytochrome c release and mtDNA efflux.

ATP concentrations before cellular demise. These findings were recapitulated in apoptosome-deficient human kidney and lung cancer cells.

CONCLUSION: Our findings identify mtDNA-releasing machinery that is distinct from the cytochrome c-releasing MOMP machinery. Thus, mtDNA efflux is not simply a downstream consequence of MOMP or MIM herniation but instead involves a distinct, active biochemical process. mtDNA efflux contributes to bacitosis, a form of BAX-BAK activation-induced programmed necrosis that occurs when apoptosome function is impaired. This work provides a mechanistic basis for exploring BAX-BAK activation-induced bacitosis as a potential therapeutic avenue in apoptosome-deficient cancers. □

CELL BIOLOGY

BAK and BAX form a multiprotein complex to drive mtDNA efflux and caspase-independent cell death

Shaoyuan Yang $^{1\dagger}$ , Hyungjin Kim $^{2\dagger\ddagger}$ , Guanghong Liao $^{1}$ , Na Ta $^{1}$ , Ho-Chou Tu $^{1}$ , Shriram Ramani $^{1}$ , Chih-Yu Chen $^{1}$ , Zih-Yu Li $^{1}$ , Nobuhiro Tanno $^{1}$ , James J. Hsieh $^{2}$ , Emily H. Cheng $^{1,3,4*}$

The apoptotic regulators BAX and BAK, once activated, form homo-oligomers to trigger mitochondrial outer membrane permeabilization (MOMP), activating the apoptosome and caspases. Activation of BAX-BAK without apoptosome activation triggers sustained efflux of mitochondrial DNA (mtDNA) and caspase-independent cell death. However, the molecular basis of mtDNA efflux remains unclear. We purified a macromolecular complex comprising BAK, BAX, activator BH3-only molecules, voltage-dependent anion channels 1 and 2 (VDAC1 and VDAC2), and prohibitins that released mtDNA from mitoplasts. BAX-BAK activation resulted in VDAC closure, mitochondrial calcium influx, and mtDNA efflux. Loss of VDAC1, VDAC2, or prohibitins prevented BAX-BAK activation-induced mtDNA efflux and caspase-independent cell death without affecting MOMP or mitochondrial inner membrane herniation. These findings support differential regulation of cytochrome c and mtDNA efflux by distinct BAX-BAK complexes.

Mitochondria have a pivotal role in mammalian cell death. Diverse apoptotic signals trigger cytochrome c release from the mitochondrial intermembrane space, enabling apoptotic protease activating factor 1 (APAF1) to assemble the heptameric apoptosome and activate caspases $(1, 2)$ . The multidomain proapoptotic proteins BAX and BAK permeabilize the mitochondrial outer membrane (MOM), whereas antiapoptotic BCL-2, BCL-X $_{L}$ , and MCL-1 preserve mitochondrial integrity $(3–5)$ . BH3-only molecules (BH3s) integrate upstream apoptotic signals to promote apoptosis by either directly activating BAX and BAK or inactivating BCL-2, BCL-X $_{L}$ , or MCL-1 $(4–13)$ . In response to apoptotic signals, activator BH3s bind directly to BAX or BAK to induce their homooligomerization, leading to MOM permeabilization (MOMP) $(4–13)$ . BCL-2 family proteins regulate apoptosis in normal physiological and disease processes and are targets for anti-cancer drugs that disrupt protein-protein interactions $(14–16)$ .

Activation of BAX and BAK not only triggers APAF1-mediated caspase activation but also initiates caspase-independent mitochondrial dysfunction and cellular demise (4). Activation of BAX and BAK in cells lacking APAF1, caspase-9, or caspase-3 and caspase-7 leads to sustained release of mitochondrial DNA (mtDNA) into the cytosol, activating the cGAS-STING pathway and type I interferon (IFN) response (17, 18). Imaging analyses revealed that cytochrome $c$ exits from BAX or BAK pores that are too small to be resolved by super-resolution microscopy (19). After cytochrome $c$ efflux, BAX and BAK coalesce to form macropores in the MOM through which the mitochondrial inner membrane (MIM) herniates and then loses membrane integrity, resulting in efflux of mtDNA $(19)$ . Only a minority of herniated MIM loses membrane integrity and most transcription factor A, mitochondrial (TFAM)-bound mtDNA remains inside the mitochondria during this process $(19)$ . During caspase-mediated apoptosis, mtDNA efflux occurs transiently on the verge of cell death because of rapid caspase-mediated killing $(19)$ . Even if a small amount of mtDNA is released during apoptosis, cGAS-STING-type I IFN signaling is abrogated by caspase-3-mediated cleavage of cGAS and IRF3, rendering apoptosis immunologically silent $(17, 18, 20, 21)$ .

The discovery of mtDNA efflux induced by activation of BAX or BAK and subsequent activation of immunogenic cell death opens new avenues for cancer cell death research, particularly given that impaired apoptosome function is common in human cancers $(22)$ . However, the molecular and biochemical basis of mtDNA efflux is not understood. Specifically, it remains unclear whether mtDNA efflux is simply a consequence of MOMP and MOMP-associated MIM herniation or if it involves yet-to-be-identified active biochemical processes. Furthermore, it is unknown whether mtDNA release contributes to caspase-independent cell death. Given that neurons and myocardium express little APAF1 $(23, 24)$ , mtDNA efflux may contribute to the pathogenesis of diseases associated with neuronal and myocardial cell death. We used biochemical and genetic studies to identify and characterize a macromolecular complex associated with BAK and BAX, which regulated mtDNA efflux and caspase-independent cell death in various cell types, including apoptosome-deficient human cancer cells.

---

Results

---

Identification of a 600-kD BAK-associated macromolecular complex

To determine whether BAK engages distinct protein complexes to regulate the efflux of cytochrome c and mtDNA, $Bax^{-/-}Bak^{-/-}$ mouse embryonic fibroblasts (MEFs) were reconstituted with N-terminal protein C-tagged BAK to facilitate the affinity purification of BAK-associated complexes. We focused on BAK because it is an integral MOM protein, whereas BAX translocates from the cytosol to mitochondria upon apoptosis (8, 25). To enhance the extraction of detergent-resistant BAK and associated integral membrane proteins, mitochondria were sequentially extracted with 1% and 2% 3-[(3-cholamidopropyl)dimethylammonio]-1-propanesulfonate (CHAPS) detergent. The 1% CHAPS-insoluble, 2% CHAPS-soluble subfraction was subjected to anti-protein C affinity chromatography, followed by silver staining and tandem mass spectrometry. The voltage-dependent anion channel 1-3 (VDAC1-3) isoforms, adenine nucleotide translocator 1 (ANT1), and prohibitin 2 (PHB2) were copurified with BAK at nearly 1:1 stoichiometry (Fig. 1A and table S1). The identity of these proteins was further verified by immunoblotting (Fig. 1B). The mitochondrial matrix protein cyclophilin D, a critical regulator of the mitochondrial permeability transition pore complex (mPTP) (26), and the MIM protein ATP synthase β served as negative controls. The association of BAK with proteins localized to both the MOM (VDAC1-3) and the MIM (ANT1 and PHB2) indicates that these components may assemble into a complex at contact sites between the MOM and MIM. Indeed, fractionation of submitochondrial vesicles on a discontinuous sucrose gradient revealed overlapping localization of BAK, VDAC1, VDAC2, ANT1, and PHB2 at MOM-MIM contact sites (Fig. 1C).

To further characterize this complex, we used gel filtration chromatography of the 2% CHAPS-soluble mitochondrial subfraction. BAK, VDACs, ANT1, and PHB2 eluted in high-molecular-weight fractions around 600 kD, consistent with these proteins forming a 600-kD complex. Cyclophilin D, cytochrome c oxidase subunit IV (COX IV), and ATP synthase β were not found in these fractions, supporting the specificity of this complex (Fig. 1D). To further confirm that these proteins form a 600-kD complex, fractions of high molecular size were collected from gel filtration chromatography and subjected to immunoprecipitation with antibody to protein C. The resulting silver-stained gel showed patterns similar to those obtained without size fractionation shown in Fig. 1A (Fig. 1E), and their identity was further confirmed by immunoblotting (Fig. 1F). This 600-kD complex appears to exist in healthy cells, distinguishing it from oligomerized BAK, which is only detected upon activation by BH3s (8, 10).

A

D

c

B

G

E

F

H

|

J

Fig. 1. BAK exists in three protein complexes with distinct death-regulatory activities. (A) Mitochondria from $Bax^{-/-}Bak^{-/-}$ MEFs ± protein C (PrC)-tagged BAK were sequentially extracted with 1% and 2% CHAPS. The 2% CHAPS-soluble subfraction was purified by anti-PrC affinity chromatography and silver stained. Bands were excised for tryptic digestion and liquid chromatography-tandem mass spectrometry. (B) $Bax^{-/-}Bak^{-/-}$ MEFs ± PrC-tagged BAK were subjected to anti-PrC immunoprecipitation and immunoblotting. (C) Mitochondria from WT MEFs were subjected to hypotonic swelling and sonication, fractionated on a discontinuous sucrose gradient into MOM, MIM, and contact site (CS) fractions, and analyzed by immunoblotting. FIS1 marks the MOM. (D) The 2% CHAPS-soluble mitochondrial subfraction from WT MEFs was analyzed by Superdex 200 (HR 10/30) gel filtration and immunoblotting. Protein standards indicate apparent molecular size in kilodaltons. (E and F) The 2% CHAPS-soluble mitochondrial subfraction from $Bax^{-/-}Bak^{-/-}$ MEFs ± PrC-tagged BAK was subjected to gel filtration. Fractions around 600 kD were purified by anti-PrC affinity chromatography and analyzed by silver staining (E) or immunoblotting (F). (G) Mitochondria from WT MEFs ± recombinant tBID were sequentially extracted with 1% and 2% CHAPS and analyzed by gel filtration chromatography and immunoblotting. The control anti-BAK immunoblot (2% CHAPS-soluble) is identical to that in (D). (H) The 1% CHAPS-soluble mitochondrial subfraction from $Bax^{-/-}Bak^{-/-}$ MEFs ± PrC-tagged BAK was analyzed by gel filtration. Fractions around 160 kD were purified by anti-PrC affinity chromatography and silver stained. (I and J) The 2% CHAPS-soluble mitochondrial subfraction from WT MEFs was analyzed by gel filtration. The 60-kD fraction before tBID and the 160- and 600-kD fractions after tBID were subjected to BMH cross-linking (I) or incubated with $Bax^{-/-}Bak^{-/-}$ mitochondria to assess cytochrome c release by ELISA (J). Data are shown as mean ± SD ( $n = 3$ independent experiments). \*\*\*P < 0.001 (Student's t test).

---

BAK exists in three protein complexes with distinct death-regulatory activities

To investigate the existence of distinct BAK-associated complexes, gel filtration chromatography (Superdex 200 HR 10/30) was performed on 1% or 2% CHAPS-soluble subfractions obtained by sequential extraction of wild-type (WT) MEF mitochondria before or after incubation with recombinant truncated BID protein (tBID) to activate BAK in vitro. BAK eluted in three distinct complexes (Fig. 1G). In both the 1% and 2% CHAPS-soluble subfractions, BAK eluted at 60 kD before tBID treatment and shifted to 160 kD after tBID treatment (Fig. 1G). We reasoned that the 160-kD complex might represent oligomerized BAK induced by tBID to mediate cytochrome c release, whereas the 60-kD complex could contain inactive BAK. BAK eluted around 600 kD from the 2% CHAPS-soluble subfraction both before and after tBID treatment (Fig. 1G). The 600-kD complex was less abundant in the 1% CHAPS-soluble subfraction, consistent with its lower solubility in 1% CHAPS. Recombinant tBID protein predominantly eluted with the 600-kD complex (Fig. 1G).

To further assess the 160-kD BAK complex, the 160-kD gel filtration fractions were subjected to anti-protein C affinity chromatography. No distinct proteins were detected in association with BAK (Fig. 1H). Therefore, the 160-kD complex probably represents a BAK hexamer composed of six 25-kD BAK monomers. To assess the functional significance of distinct BAK complexes, protein cross-linking and cytochrome c release assays were done with the 60-, 160-, and 600-kD fractions. Only BAK present in the 160-kD fraction was cross-linked into dimers and trimers by the bismaleimidohexane (BMH) crosslinker (Fig. 1I) and released cytochrome c from $Bax^{-/-}Bak^{-/-}$ mitochondria (Fig. 1J). These data indicate that the 600-kD BAK complex is functionally distinct from the 160-kD BAK complex that promotes cytochrome c release (Fig. 1J).

We next investigated whether BAK formed the same complexes after doxycycline-induced tBID expression in $Apaf1^{-/-}$ MEFs. This system enabled analysis of tBID-induced mtDNA efflux. It also allowed us to determine whether BAX formed distinct complexes and was recruited to the 600-kD complex after tBID expression, because most BAX is absent from mitochondria in healthy cells (Fig. 1G). The 2% CHAPSSoluble mitochondrial subfraction from tBID-inducible $Apaf1^{-/-}$ MEFs before or after 6 hours of tBID induction was analyzed by gel filtration chromatography. Using a different Superdex 200 (10/300 GL) column, BAK similarly shifted from 60- to 160-kD fractions upon tBID induction and consistently eluted with VDAC1, VDAC2, ANT1, and PHB2 at an apparent size of \~600 kD regardless of tBID expression (Fig. 2A). Consistent with BAX translocation from the cytosol to mitochondria during apoptosis, mitochondrial BAX increased and shifted from primarily 40 kD to 140 kD and 600 kD after tBID expression (Fig. 2A). These findings indicate that BAX, once activated by tBID, is recruited to the 600-kD complex. BAX still formed the 140- and 600-kD complexes after tBID accumulation in BAK-deficient cells (fig. S1A), indicating that these BAX-associated complexes assemble independently of BAK. We reasoned that the 40-kD BAX complex contains inactive BAX, whereas the 140-kD BAX complex is composed of a BAX hexamer analogous to the 160-kD BAK hexameric complex, representing the MOMP machinery. tBID was also recruited to the 600-kD complex, although the elution profile of endogenously overexpressed tBID differed from that of exogenously added recombinant tBID (Figs. 2A and 1G). Coimmunoprecipitation of BAK, BAX, tBID, and PHB2 with hemagglutinin (HA)-VDAC1 was confirmed in tBID-inducible $Apaf1^{-/-}$ MEFs in which endogenous VDAC1 was replaced with HA-tagged VDAC1 (fig. S1B).

---

Loss of VDAC1 or VDAC2 inhibits BAX-BAK-dependent caspase-independent cell death

To assess the function of the 600-kD BAK-BAX complex, we targeted each Vdac gene with Cas9 and single-guide RNAs (sgRNAs) in tBID-inducible $Apaf1^{-/-}$ MEFs. Targeting Vdac1 or Vdac2, but not Vdac3, protected $Apaf1^{-/-}$ cells from cell death triggered by tBID or by combined inhibition of BCL-2, BCL- $\mathrm{X_L}$ , and MCL-1 using ABT-263 (navitoclax, a dual inhibitor of BCL-2 and BCL- $\mathrm{X_L}$ ) and S63845 (an MCL-1 inhibitor) (Fig. 2B) (27, 28). Simultaneous targeting of both Vdac1 and Vdac2 conferred greater protection. These findings were further validated in tBID-inducible $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ MEFs (Fig. 2C). By contrast, targeting both Vdac1 and Vdac2 failed to prevent and even slightly increased apoptosis in WT MEFs (Fig. 2D). However, WT MEFs lacking Vdac1 and Vdac2 were protected from tBID-induced cell death when caspase activity was inhibited by the pan-caspase inhibitor Q-VD-OPh (fig. S1C). As reported previously (4), BAX-BAK activation killed $Apaf1^{-/-}$ MEFs at a much slower rate than it did WT MEFs (fig. S1D). ABT-263 plus S63845 (AS) failed to activate BAX or BAK in the absence of activator BH3s (fig. S1E), consistent with their displacement of endogenous activator BH3s to indirectly activate BAX-BAK (10, 12, 29). Moreover, retroviral transduction of tBID, BIM, or PUMA triggered cell death in $Apaf1^{-/-}$ MEFs, which was suppressed by the loss of VDAC1-VDAC2 (fig. S1F). Genetic studies indicate that VDACs are dispensable for MOMP (30), and deficiency of Vdac2 alone or of all three Vdac isoforms even accelerates caspase-dependent apoptosis (30–32).

Consistent with the viability data, loss of VDAC1-VDAC2 did not inhibit tBID-induced cytochrome c release, although it appeared to reduce baseline mitochondrial cytochrome c levels (Fig. 2E). tBID-induced oligomerization of BAX or BAK was still detected by cross-linking in $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ cells (Fig. 2F). VDAC2 not only keeps BAK in an inactive monomeric form but also stabilizes its mitochondrial localization (31–34). Accordingly, BAK became BAX-like in the absence of VDAC2: Once activated by death signals, it translocated from the cytosol to mitochondria (Fig. 2E) and formed homo-oligomers, although BAK was unstable and less abundant (Fig. 2F). Consistent with VDAC2-mediated BAK stabilization, loss of VDAC2, but not VDAC1 or VDAC3, led to reduced abundance of BAK (fig. S1G). A small molecule reportedly binds VDAC2 and enhances its inhibition of BAK-driven apoptosis (35). Collectively, VDAC2 appears to negatively regulate BAK-dependent apoptosis while promoting BAK-mediated caspase-independent cell death, likely through association with distinct BAK complexes (60 versus 600 kD).

Our findings that loss of VDAC1-VDAC2 inhibits tBID-induced death in $Apaf1^{-/-}$ cells without blocking MOMP (Fig. 2E) indicate that BAX-BAK-dependent, caspase-independent cell death is not merely a consequence of MOMP. Furthermore, MEFs survived and proliferated in the absence of cytochrome $c$ , and its deficiency did not affect tBID-induced killing of $Apaf1^{-/-}$ cells (fig. S1H). We tested specific inhibitors of other known cell death pathways in BAX-BAK activation-induced caspase-independent cell death that we call baciptosis (fig. S1I). None of these inhibitors prevented baciptosis, thereby excluding the involvement of pyroptosis (inhibited by Q-VD-OPh), necroptosis (inhibited by the protein kinase RIPK1 inhibitor necrostatin-1 and the RIPK3 inhibitor GSK'872), ferroptosis (inhibited by ferrostatin-1 and deferoxamine), and mPTP (inhibited by cyclosporin A). Conversely, CRISPR-Cas9-mediated targeting of mitochondrial DNA polymerase gamma (Polg) depleted mtDNA and protected $Apaf1^{-/-}$ cells from tBID-induced cell death (Fig. 2G). Similar protection was observed in cells treated with ethidium bromide to deplete mtDNA (fig. S1J), supporting the notion that cytosolic release of mtDNA contributes to baciptosis.

We characterized the morphological and biochemical features of bac-iptosis. Time-lapse live-cell imaging showed that WT MEFs expressing tBID exhibited classical apoptotic features, including membrane blebbing, cell shrinkage, and fragmentation (movie S1), whereas $Apaf1^{-/-}$ MEFs expressing tBID displayed prominent balloon-like swelling before cell death, consistent with necrotic morphology (movie S2). Electron microscopy revealed that tBID expression in WT MEFs produced characteristic apoptotic features, including chromatin condensation along the nuclear periphery and plasma membrane blebbing (fig. S2A). By contrast, tBID expression in $Apaf1^{-/-}$ MEFs resulted in necrotic features, including plasma membrane and organelle disruption, induction of autophagy and mitophagy, and lysosomal accumulation (fig. S2, B and C). Consistently, developmental cell death in the interdigital webs of $Apaf1^{-/-}$ mice has necrotic rather than apoptotic morphology (36). These observations support the classification of baciptosis as a distinct form of programmed necrosis. To exclude additional death modalities, we assessed lysosomal membrane permeabilization and detected no evidence of lysosomal disruption in either $Apaf1^{-/-}$ or $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ MEFs (fig. S2D). We also evaluated genomic DNA integrity by terminal deoxynucleotidyl transferase-mediated deoxyuridine triphosphate nick end labeling (TUNEL) staining. Apoptotic cells showed TUNEL staining, whereas baciptotic cells did not (fig. S2E). Baciptosis was accompanied by a progressive decline in ATP concentrations preceding cellular demise, whereas loss of VDAC1-VDAC2 preserved ATP concentrations (fig. S2F). Thus, baciptotic cell death appears to cause catastrophic bioenergetic failure downstream of mtDNA release.

A

B

C

F

D

E

G

Fig. 2. Loss of VDAC1 or VDAC2 inhibits BAX-BAK-dependent caspase-independent but not caspase-dependent cell death. (A) The 2% CHAPS-soluble mitochondrial subfraction from tBID-inducible $Apaf1^{-/-}$ MEFs ± doxycycline (1 μg/ml) for 6 hours was analyzed by Superdex 200 (Increase 10/300 GL) gel filtration chromatography, followed by immunoblotting. (B) tBID-inducible $Apaf1^{-/-}$ MEFs transduced with Cas9 and the indicated sgRNAs were analyzed by immunoblotting. Cells were treated with doxycycline or ABT-263 (2 μM) plus S63845 (10 μM), and cell death was quantified by PI staining and flow cytometry at 48 hours. (C) tBID-inducible $Apaf1^{-/-}$ or $Apaf1^{-/-}$ Vdac1 $^{-/-}$ -Vdac2 $^{-/-}$ MEFs were treated as in (B), and cell death was quantified by PI staining at 48 hours. (D) tBID-inducible WT MEFs transduced with Cas9 and the indicated sgRNAs were treated as in (B), and cell death was quantified by annexin V staining at 6 hours. (E) tBID-inducible $Apaf1^{-/-}$ or $Apaf1^{-/-}$ -Vdac1 $^{-/-}$ -Vdac2 $^{-/-}$ MEFs ± doxycycline for the indicated times were subjected to subcellular fractionation and analyzed by immunoblotting. (F) tBID-inducible $Apaf1^{-/-}$ or $Apaf1^{-/-}$ -Vdac1 $^{-/-}$ -Vdac2 $^{-/-}$ MEFs ± doxycycline (6 hours) were treated with BMH and analyzed by immunoblotting. An asterisk denotes an intramolecularly cross-linked BAK monomer. (G) tBID-inducible $Apaf1^{-/-}$ MEFs transduced with Cas9 and the indicated sgRNAs were subjected to qPCR-based mtDNA quantification (left). Cell death was quantified by PI staining at 48 hours after doxycycline treatment (right). Data in (B) to (D) and (G) are shown as mean ± SD ( $n = 3$ independent experiments). \* $P < 0.05$ ; \*\*\* $P < 0.001$ ; \*\*\*\* $P < 0.0001$ ; ns, not significant (Student's t test).

---

Loss of VDAC1, VDAC2, or prohibitins impairs BAX-BAK-dependent mtDNA efflux

Because mtDNA efflux contributes to baciptosis, which is inhibited by VDAC1 and VDAC2 loss, we tested whether the 600-kD BAK-BAX complex promotes mtDNA release. tBID-induced mtDNA efflux depended on BAK and BAX (Fig. 3A). As in apoptosis, BAK and BAX functioned redundantly to drive baciptosis (fig. S3A). Loss of both VDAC1 and VDAC2 reduced mtDNA efflux induced by tBID or AS in $Apaf1^{-/-}$ cells (Fig. 3, B and C). Retroviral transduction of VDAC1 or VDAC2, but not VDAC3, restored tBID-induced mtDNA efflux and baciptosis in $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ cells (Fig. 3D). Consistent with the role of VDAC2 in stabilizing BAK, reexpression of VDAC2, but not VDAC1 or VDAC3, in $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ cells restored BAK abundance (Fig. 3D). Thus, VDAC1 and VDAC2, but not VDAC3, have critical roles in BAX-BAK-dependent mtDNA efflux and baciptosis.

To determine whether loss of VDAC1-VDAC2 affects the formation of distinct BAK-BAX complexes, we performed gel filtration chromatography on the 2% CHAPS-soluble mitochondrial subfraction from $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ cells before and after tBID expression. Consistent with BAK becoming BAX-like in the absence of VDAC2, both BAK and BAX were present exclusively in small-molecular-size fractions (40 to 60 kD) before tBID expression and shifted to 140- to 160-kD and 600-kD fractions after tBID expression (Fig. 3E). Thus, loss of VDAC1-VDAC2 did not impair the formation of the MOMP machinery, as supported by gel filtration chromatography (Fig. 3E) and protein cross-linking (Fig. 2F). It also did not alter the elution profiles of tBID, ANT1, or PHB2 (Fig. 3E), indicating that deficiency of VDAC1-VDAC2 does not grossly disrupt the formation of the 600-kD BAK-BAX complex. Similarly, loss of BAX-BAK did not affect the localization of VDAC1, VDAC2, prohibitins, and tBID to the 600-kD complex (fig. S3B). These data indicate that components of the 600-kD complex may be independently targeted to contact sites between the MIM and MOM, where the specialized lipid and protein microenvironment may facilitate their assembly.

Because mtDNA efflux requires passage across both mitochondrial membranes, PHB2, ANT1, or both (MIM proteins associated with the 600-kD BAK-BAX complex) may contribute to this process. We could not generate stable Phb2-deficient cell lines due to its essentiality $(37)$ . Nonetheless, transient targeting of Phb2 significantly attenuated tBID-induced mtDNA efflux and bacitosis in $Apaf1^{-/-}$ MEFs (Fig. 3, F to H). By contrast, depletion of Phb2 did not affect tBID-induced apoptosis in WT MEFs (fig. S3C). Consistent with the interdependence of PHB1 and PHB2 within a heterodimeric MIM complex (37, 38), targeting Phb2 reduced the abundance of both PHB2 and PHB1 but slightly increased the abundance of BAK (Fig. 3F). PHB1 also eluted with the 600-kD BAK-BAX complex and coimmunoprecipitated with BAK (fig. S3, D and E). Loss of PHB1-PHB2 did not impair recruitment of BAX, BAK, VDAC1, VDAC2, or tBID to the 600-kD complex (fig. S3F). Thus, prohibitins appear to be required for mtDNA release but not complex assembly. Loss of VDAC1-VDAC2 or prohibitins significantly attenuated accumulation of IFN- $\beta$ , an indicator of mtDNA-mediated activation of the cGAS-STING-type I IFN pathway, in tBID-expressing $Apaf1^{-/-}$ MEFs (fig. S3, G and H).

Because ANT1 interacts with ANT2, which also eluted with the 600-kD complex (fig. S3D), we evaluated both isoforms. Lentiviral transduction of Cas9 and sgRNAs targeting both Ant1 and Ant2 had minimal effects on tBID-induced mtDNA efflux and baciptosis (Fig. 3, I and J). Thus, despite their presence in the 600-kD complex, ANTs and VDAC3 are not key mediators of BAX-BAK-dependent mtDNA efflux and baciptosis and may instead have structural or auxiliary roles. Although VDAC-mediated mitochondrial calcium influx could potentially activate the mPTP, depletion of cyclophilin D had little effect on tBID-induced mtDNA efflux and baciptosis (fig. S3I), arguing against mPTP involvement. Similar results were obtained with cyclosporin A treatment (fig. S1I). The roles of VDAC1, VDAC2, and PHB2 in BAX-BAK activation-induced mtDNA efflux and baciptosis were further validated in B16-F10 mouse melanoma cells (fig. S4).

---

Characterization of BAX-BAK activation-induced mtDNA efflux and cell death in apoptosome-deficient human cancers

Given the multiple mechanisms of apoptosome inactivation in human cancers (22), we investigated whether baciptosis occurs under these conditions and whether it requires VDAC1, VDAC2, and prohibitins. We screened commonly used human cancer cell lines for minimal effector caspase activity after BAX-BAK activation with AS. AS effectively triggered MOMP, as evidenced by cytochrome c release, in A498 kidney cancer, PC9 lung cancer, and MDA-MB-231 breast cancer cells; however, minimal effector caspase activity was detected in A498 and PC9 cells, likely due to their low APAF1 expression (Fig. 4A and fig. S5, A to C). Consistent with caspase activation, MDA-MB-231 cells died faster than A498 and PC9 cells, and only cell death in MDA-MB-231 cells was blocked by Q-VD-OPh (Fig. 4B and fig. S5D). Depletion of both BAX and BAK protected A498 and PC9 cells from AS-induced cell death (fig. S5E). Thus, activation of BAX-BAK by AS induces baciptosis in A498 and PC9 cells.

To determine whether human BAK and BAX form the aforementioned complexes upon activation, we isolated mitochondria from A498 cells before and after AS treatment and analyzed the 2% CHAPSSoluble mitochondrial subfraction by gel filtration chromatography (Fig. 4C). Similar to mouse BAK, apparent size of human BAK shifted from 60 to 160 kD after AS treatment and eluted with VDAC1, VDAC2, PHB1, and PHB2 at \~600 kD both before and after treatment (Fig. 4C). Consistent with mouse BAX, human BAX eluted exclusively with apparent size of 40 kD before treatment and shifted to 140 and 600 kD after treatment (Fig. 4C). BIM predominantly eluted in fractions with apparent size of \~50 kD before treatment but shifted to 600 kD after AS treatment (Fig. 4C). These findings indicate that AS displaces BIM from antiapoptotic BCL-2 family members to activate BAX-BAK present within the 40- to 60-kD and 600-kD complexes, one for the formation of the MOMP machinery and the other for the mtDNA-releasing machinery, respectively.

We investigated the roles of VDACs and prohibitins in AS-induced mtDNA efflux and bacitosis in A498 and PC9 cells. Targeting $VDAC1$ or $VDAC2$ , but not $VDAC3$ , significantly attenuated both responses (Fig. 4, D and E, and fig. S5, F and G). Simultaneous depletion of VDAC1 and VDAC2 could not be assessed because these cells did not tolerate loss of both proteins. As in MEFs, loss of VDAC2, but not VDAC1 or VDAC3, reduced BAK abundance (fig. S5, F and G). Because prohibitins are essential and human cells did not tolerate CRISPR-Cas9-mediated PHB2 depletion, we transiently depleted PHB2 with lentiviral short hairpin RNAs (shRNAs). Depletion of PHB2 reduced the abundance of both PHB2 and PHB1 and significantly attenuated AS-induced mtDNA efflux (Fig. 4, F and G, and fig. S5, F and G). However, we were unable to assess the effect of PHB1-PHB2 loss on bacitosis, because cells deficient in both proteins died before the onset of AS-induced bacitosis. By contrast, loss of VDAC1 or VDAC2 failed to inhibit AS-induced apoptosis in MDA-MB-231 cells or in APAF1-reconstituted A498 cells (Fig. 4, H and I, and fig. S5, H to J). Loss of VDAC1, VDAC2, or prohibitins also attenuated AS-induced IFN- $\beta$ expression in A498 and PC9 cells (fig. S5, K and L), consistent with their participation in BAX-BAK-dependent mtDNA efflux. Thus, human cancer cells recapitulated the mouse phenotypes, confirming that apoptosome-deficient cancers can undergo bacitosis after BAX-BAK activation, a process regulated by VDAC1 and VDAC2.

A

C

B

D

E

F

G

H

|

J

Fig. 3. Loss of VDAC1, VDAC2, or prohibitins impairs BAX-BAK-dependent mtDNA efflux and caspase-independent cell death. (A) tBID-inducible $Apaf1^{-/-}$ MEFs transduced with Cas9 and the indicated sgRNAs ± doxycycline or AS were subjected to cell death quantification at 48 hours and immunoblotting at 6 hours. Cytosolic mtDNA was assessed by qPCR at 6 hours after doxycycline treatment and 4 hours after AS treatment. (B and C) tBID-inducible $Apaf1^{-/-}$ or $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ MEFs were untreated or treated with doxycycline (B) or AS (C). Cytosolic mtDNA was assessed by qPCR at 6 and 4 hours, respectively. (D) tBID-inducible $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ MEFs transduced with control retrovirus or retroviruses expressing VDAC1, VDAC2, or VDAC3 ± doxycycline (6 hours) were analyzed by immunoblotting. Cell death was quantified at 48 hours, and cytosolic mtDNA was assessed at 6 hours. (E) The 2% CHAPS-soluble mitochondrial subfraction from tBID-inducible $Apaf1^{-/-}Vdac1^{-/-}Vdac2^{-/-}$ MEFs ± doxycycline (6 hours) was analyzed by Superdex 200 (Increase 10/300 GL) gel filtration chromatography, followed by immunoblotting. (F to J) tBID-inducible $Apaf1^{-/-}$ MEFs transduced with Cas9 and the indicated sgRNAs ± doxycycline or AS were subjected to cell death quantification at 48 hours and immunoblotting at 6 hours [(F) and (I)]. Cytosolic mtDNA was assessed by qPCR at 6 hours (G and J) or 4 hours (H). Data in (A) to (D) and (F) to (J) are shown as mean ± SD ( $n = 3$ or 4 independent experiments). \*P < 0.05; \*\*P < 0.01; \*\*\*P < 0.001; \*\*\*\*P < 0.0001; ns, not significant (Student's t test).

A

B

C