Biohub

Biohub is a nonprofit biomedical research organization that combines artificial intelligence, experimental biology, and scientific instrumentation to study cells and disease. Founded by Priscilla Chan and Mark Zuckerberg, it develops models, laboratory technologies, and shared research resources. Its stated mission is to help scientists cure or prevent disease.^[1][2]

Organization and history

Biohub grew out of the Chan Zuckerberg Biohub network. The San Francisco institute began in 2016; the network subsequently included institutes in Chicago and New York. In March 2025, the Chan Zuckerberg Initiative announced that the San Francisco Biohub and its imaging institute would combine their teams, with a new Redwood City science campus planned for 2027.^[3]

On November 6, 2025, the organization announced that its scientific teams would operate together under the Biohub name. The EvolutionaryScale team would join the organization, and its co-founder Alex Rives would become head of science. Biohub also announced a plan to expand its computing capacity to 10,000 GPUs by 2028.^[4][5]

Biohub's organizational structure includes Chan Zuckerberg Biohub, Inc., a 501(c)(3) medical research organization. Chan Zuckerberg Initiative, LLC provides strategic, funding, and operational support; the structure also includes the Chan Zuckerberg Initiative Foundation and a donor-advised fund. Its investigator program supports research and collaborations with university scientists.^[2]

Research programs

Biohub's research agenda connects biological measurements with models that can predict cellular behavior. Its four scientific grand challenges define the principal areas of work:^[6]

The Virtual Immune System project, announced in November 2025, is a research roadmap for predictive immunology. It emphasizes intervention-based data and models suited to different biological scales. The proposed workflow links experimental data generation, modeling, predictions, and follow-up experiments, rather than treating a static cell atlas as a complete simulation of immunity.^[7]

AI models

Protein models and ESM Atlas

Biohub's protein-model platform uses protein language models to learn representations from amino acid sequences. Its protein world-model release includes the following components:^[8][9]

The accompanying June 2026 preprint reports laboratory tests of designed miniproteins and single-chain antibodies with nanomolar binding affinities. These experiments concern molecular binding and protein design. The release page provides platform access and links to model weights and code, and states that the models have an MIT license.^[8][9]

Cellular and biological reasoning models

Biohub's November 2025 announcement introduced VariantFormer, CryoLens, and scLDM, alongside models already available through its virtual-cell platform:^[4]

Open biological datasets

Biohub's data resources include measurements of gene expression, protein localization, and molecular interactions. These datasets describe different aspects of biology rather than interchangeable measurements of the same phenomenon.^[10][11]

The 2022 OpenCell paper combines genome engineering, live-cell confocal imaging, mass spectrometry, and data analysis. Its unsupervised learning approach groups proteins by spatial and interaction patterns to investigate cellular organization.^[12]

Examples of published research

Virtual Lab

Biohub San Francisco and Stanford researchers developed Virtual Lab, a system in which a principal-investigator AI agent directs specialist agents while a human scientist gives high-level feedback. A 2025 Nature paper describes a computational pipeline combining ESM, AlphaFold-Multimer, and Rosetta to design 92 nanobodies. Laboratory testing identified two with improved binding to the SARS-CoV-2 JN.1 or KP.3 variants while retaining binding to the ancestral spike protein. The study presents them as candidates for further investigation.^[13]

AI-guided screening for psoriasis targets

A 2026 Nature Communications study by Biohub Chicago researchers and collaborators combined a genome-wide CRISPR screen with VirtualCRISPR, a language-model framework trained on functional-genomics data. The work identified ALOX5 and OXTR as regulators of the IL-17 receptor in human keratinocytes. Topical inhibitors of these targets reduced experimentally induced psoriasis-like dermatitis in mice. These results are preclinical, rather than evidence of treatment effectiveness in human patients.^[14]

Virtual Biology Initiative

On April 29, 2026, Biohub announced a five-year Virtual Biology Initiative to expand open data for predictive biological models. Its $500 million commitment comprises $400 million for internal technologies and data generation and $100 million for external research. The initiative coordinates with research institutes and biological atlas consortia, with NVIDIA as a technology partner.^[15]

October 2026 expansion

On October 7, 2026, Biohub, the U.S. Department of Energy (DOE) and the National Institutes of Health (NIH) announced an expansion of the Virtual Biology Initiative with new funding partners. The organizations described a combined $1.8 billion in funding, data, computation, and measurement technology, which Biohub called the largest coordinated commitment to generating AI-ready biological data to date. The headline of Biohub's release rounds the figure to "nearly $2 billion".^[16] The total combines new commitments with existing resources: it counts Biohub's April commitment and NIH datasets built with earlier federal funding, so it is not $1.8 billion of newly raised cash.^[16][22]

The NIH contribution consists of existing datasets and infrastructure, not a newly announced $500 million cash grant.^[16][20] DOE's announcement describes a memorandum of understanding signed by its Office of Science, NIH, and Biohub, and says the partnership delivers on the executive order launching the Genesis Mission. The government releases describe the goal as data for "Super Intelligence (SI)" models of biology, while Biohub's release uses the term AI. DOE resources named in its announcement include exascale supercomputing; user facilities such as the Joint Genome Institute and the Environmental Molecular Sciences Laboratory; X-ray and neutron scattering; cryo-electron microscopy and tomography; and automated laboratories, including the Genesis Mission's Orchestrated Platforms for Autonomous Laboratories (OPAL) project.^[19] NIH said its part runs through its Bio Genesis Mission. It draws on repositories catalogued by the National Library of Medicine and the National Center for Biotechnology Information, and on NIH Common Fund programs that build biological atlases and shared data standards. NIH also tied the effort to the Predicting Living Systems National Science and Technology Challenge.^[20]

Several research organizations joined the effort to help coordinate the scientific community: the Allen Institute, the Broad Institute, the Gladstone Institutes, the Human Cell Atlas, the Human Protein Atlas, and the Wellcome Sanger Institute. Biohub said NVIDIA will support the initiative with accelerated computing infrastructure, domain-specific software, and technical expertise, and Renaissance Philanthropy is helping to expand funding for data generation. Isomorphic Labs president Max Jaderberg described the company as a founding member of the initiative.^[16][27] Biohub said it would build shared standards, common identifiers, and a single point of access so that partners' datasets can be used together. It presented this as a continuation of its earlier data projects, including Tabula Sapiens, OpenCell, Zebrahub, CELLxGENE, and the CryoET Data Portal.^[16]

Biohub describes the resulting datasets as an open resource for the research community.^[16] The Next Web reported that, according to interviews Rives gave to Axios and Reuters, the commercial funders will receive one year of exclusive access before the data become public, while government-funded work will carry no such restriction. Rives called the embargo an incentive for commercial participation. According to the same report, Rives told Reuters that current cell datasets contain hundreds of millions of cells, while an accurate model will need billions and eventually trillions. The partners aim to release a first dataset in about a year and to build accurate predictive models within five years.^[21] Interesting Engineering noted that the goal remains prospective: the initiative has not yet produced a universal virtual cell model, and the commitment is meant to build the data, measurement technologies, and computing infrastructure needed to develop and test such models.^[22]

Scientific scope and evaluation

A 2024 Cell perspective co-authored by researchers at the Chan Zuckerberg Initiative and Biohub institutes describes an AI virtual cell as a model connecting molecular, cellular, and multicellular measurements. Proposed uses include predicting previously untested interventions and selecting informative follow-up experiments. The authors recommend evaluating generalization across biological contexts, calibrating prediction uncertainty, and assessing biologically meaningful capabilities rather than relying on narrow benchmarks. The paper presents a research agenda, not a completed universal-cell system.^[26]

The October 2026 expansion addresses the data requirements for this work. NIH describes measurements of intervention responses across cell types and conditions as a training-data requirement, and says predictive models could help prioritize ideas for laboratory and clinical testing. These are prospective uses of models under development, not demonstrated treatments.^[20]

Validation and responsible use

Biohub's platform limitations state that ESM models and the atlas support research and hypothesis generation. Generated structures, annotations, similarity scores, and cluster assignments are computational predictions that require independent validation before drawing biological or clinical conclusions. Reliability can vary by target and sequence context, and the atlas's AI search agent can return inaccurate or incomplete answers.^[17]

Biohub's ethics policy addresses scientific misconduct, credit, grant funding, and respectful treatment of participants. It also requires compliance with institutional and national standards for research involving human participants or animals.^[18]

Reception

When CZI folded its science work into Biohub in November 2025, Science reported both praise and criticism. Robert Tjian, a University of California, Berkeley biochemist who once led the Howard Hughes Medical Institute and has advised CZI, called the founders' plan to give their fortune to science "fantastic" and said "there are a lot of other very, very wealthy individuals that are not doing it." The wind-down of CZI's other programs had drawn "sharp criticism from some former CZI employees and jilted grantees," the magazine reported. Matthew Porteus, a Stanford University pediatrician and former CZI grantee, questioned Zuckerberg's support for President Donald Trump, who wanted to cut the NIH budget by 37%: "There's an inconsistency between what the CZI is doing and what is going on at the federal government level." Cori Bargmann, CZI's head of science for its first six years, said she knew Zuckerberg had to wear two hats: "It's a lot easier to do philanthropy when you're not trying to run a major business at the same time." Immunologist Jeffrey Bluestone, head of Sonoma Biotherapeutics, doubted that the planned virtual immune system would be the key to curing and preventing all diseases ("It ain't going to work"), though he said the project could still be "transformational" for immunology.^[23]

Outside scientists interviewed by Nature about the May 2026 protein release were largely positive, particularly about ESMFold2 being fully open source. Gemma Atkinson of Lund University called the ESM Atlas "an extraordinary resource for biology." Christine Orengo of University College London said the predictions would first need evaluating but could help uncover new protein folds and functions. Martin Steinegger of Seoul National University said his main question was how well ESMFold2 predicts proteins that differ greatly from known ones; his group had found the first ESMFold weak on unusual structures, especially those from metagenomic data. Sergey Ovchinnikov of the Massachusetts Institute of Technology described the atlas as a supplement to the AlphaFold database rather than a replacement, and called the model's predictions of interacting proteins impressive but not all that surprising, since Isomorphic Labs had reported substantial gains with a proprietary model earlier in the year. He expected the absence of restrictions on commercial use to bring wide adoption: "I expect many people will be excited to try ESMFold2."^[24]

Coverage of the October 2026 expansion of the Virtual Biology Initiative paid particular attention to its access terms. Reuters wrote that embargo periods for commercial funders, among them Meta, were how "Biohub is drawing private money into a project it describes as open science."^[25]

See also

• Foundation Models
• AI Drug Discovery
• EvolutionaryScale

References

1. ^Biohub. Biohub research mission and programs. Accessed October 8, 2026.
2. ^1 ^2Biohub. Team. Organizational structure, founders, and investigator program. Accessed October 8, 2026.
3. ^Biohub. CZI announces Biohub to develop breakthrough imaging technologies to observe cells in action. March 26, 2025.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Biohub. AI-biology launch announcement. November 6, 2025.
5. ^Biohub. Launching the first large-scale scientific initiative combining frontier AI with frontier biology to solve disease. November 6, 2025.
6. ^1 ^2 ^3 ^4 ^5Biohub. 4 scientific grand challenges to transform human health at the intersection of AI and biology. April 16, 2025.
7. ^Aly Khan. The Virtual Immune System: A roadmap for predictive human immunology. Biohub, November 6, 2025.
8. ^1 ^2 ^3 ^4 ^5Biohub. A World Model of Protein Biology. Model descriptions, experimental results, and availability. Accessed October 8, 2026.
9. ^1 ^2 ^3Salvatore Candido, Thomas Hayes, et al. Language Modeling Materializes a World Model of Protein Biology. bioRxiv preprint, posted June 4, 2026. DOI: 10.64898/2026.06.03.729735.
10. ^1 ^2Tabula Sapiens Consortium. Tabula Sapiens data portal. Combined release description and cell counts. Accessed October 8, 2026.
11. ^1 ^2OpenCell. OpenCell protein data portal. Dataset counts and search interface. Accessed October 8, 2026.
12. ^Nathan H. Cho, Keith C. Cheveralls, Andreas-David Brunner, et al. OpenCell: Endogenous tagging for the cartography of human cellular organization. Science 375(6585), eabi6983, March 11, 2022. DOI: 10.1126/science.abi6983.
13. ^Kyle Swanson, Wesley Wu, Nash L. Bulaong, John E. Pak, and James Zou. The Virtual Lab of AI agents designs new SARS-CoV-2 nanobodies. Nature 646, 716-723, published July 29, 2025. DOI: 10.1038/s41586-025-09442-9.
14. ^Chenlin Zhao, Mushaine Shih, Sharif Ahmed, et al. AI-guided CRISPR screening reveals therapeutic targets in psoriasis. Nature Communications 17, 8346, published July 6, 2026. DOI: 10.1038/s41467-026-75249-5.
15. ^Biohub. Biohub Launches the Virtual Biology Initiative to Galvanize a Global Effort to Create the Open Data Foundation for AI-Accelerated Biology. April 29, 2026.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Biohub. International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease. October 7, 2026.
17. ^Biohub. Limitations. Updated May 27, 2026. Accessed October 8, 2026.
18. ^Biohub. Ethics. Accessed October 8, 2026.
19. ^1 ^2U.S. Department of Energy, Office of Science. DOE, NIH, and Biohub Partner to Build Foundational Data for Predictive Biological Super Intelligence Models. October 7, 2026.
20. ^1 ^2 ^3 ^4National Institutes of Health. NIH joins effort to build SI-ready data for predictive models of human biology. October 7, 2026.
21. ^Ana Maria Constantin. Biohub, Meta, Google DeepMind and the US pool $1.8bn for AI biology data. The Next Web, October 7, 2026.
22. ^1 ^2Neetika Walter. Google, Meta, and US build biology datasets for 'virtual cells'. Interesting Engineering, October 7, 2026.
23. ^Jon Cohen. AI drives dramatic expansion of Chan Zuckerberg Initiative's funding to end all diseases. Science, November 6, 2025.
24. ^Ewen Callaway and Miryam Naddaf. Move over, AlphaFold: open-source model predicts shape of 1 billion proteins. Nature 654, 13-14, May 27, 2026. DOI: 10.1038/d41586-026-01686-3.
25. ^Krystal Hu. US government, Google join Zuckerberg-backed Biohub in $1.8 billion push for AI biology data. Reuters, October 7, 2026; republished by CNBC.
26. ^Charlotte Bunne, Yusuf Roohani, Yanay Rosen, et al. How to build the virtual cell with artificial intelligence: Priorities and opportunities. Cell 187(25), 7045-7063, December 12, 2024. DOI: 10.1016/j.cell.2024.11.015. Perspective.
27. ^Isomorphic Labs. Isomorphic Labs joins the Virtual Biology Initiative to build foundational data for AI models to predict and treat disease. October 7, 2026.