Biohub

Biohub, founded in 2016 as the Chan Zuckerberg Biohub, is a nonprofit biomedical research organization created by Priscilla Chan and Mark Zuckerberg to build tools and datasets for biology rather than to develop drugs. It began as a San Francisco institute shared by three Bay Area universities, grew into a network of sites in California, Chicago and New York, and in November 2025 became the couple's primary philanthropic effort with a stated focus on AI models of how cells work. As of October 2026 it anchors what its partners call the largest coordinated commitment yet to generating data for Virtual cell models, on the premise that data is what such models lack.

§Overview

Biohub's stated mission is to "cure or prevent all disease", a goal the founders originally framed as a century-long project. Its method has been consistent as its scale changed: fund measurement technology, cell atlases, open software and grants to academic scientists, and publish the results openly. It is not a drug company and has no clinical pipeline, which distinguishes it from technology-funded biology ventures such as Calico Life Sciences and Altos Labs and from AI drug discovery start-ups; in structure it is closer to an independent research institute such as the Buck Institute for Research on Aging.

Chan Zuckerberg Biohub, Inc. is a 501(c)(3) medical research organization; strategic, funding and operational support comes from the Chan Zuckerberg Initiative, LLC. As of October 2026 the Initiative lists Alex Rives as head of science, with site heads in San Francisco (Scott Fraser), Chicago (Shana Kelley) and New York (Andrea Califano).

§History

Chan and Zuckerberg announced the Biohub in September 2016 as the first science investment of a plan to spend $3 billion on medical research over ten years. It received $600 million over a decade, a building beside UCSF's Mission Bay campus, and two co-directors, the UCSF biochemist Joseph DeRisi and the Stanford bioengineer Stephen Quake. Jennifer Doudna and Robert Tjian of UC Berkeley took advisory roles.¹ An Investigator Program gave no-strings funding to roughly a hundred faculty at the partner universities in its first five years.

In December 2021 the Chan Zuckerberg Initiative announced up to $3.4 billion for science over ten to fifteen years, including $800 million to $1 billion more for the Biohub and $1 billion for a new Biohub Network.² Quake moved to lead the network in 2022 and DeRisi became sole president of the San Francisco institute. A Chicago Biohub, co-led by Northwestern, the University of Chicago and the University of Illinois Urbana-Champaign to measure inflammation inside human tissue, was chosen from 58 applicant teams in March 2023. A New York Biohub followed in October 2023 with Columbia, Rockefeller and Yale, to engineer immune cells that detect Cancer and neurodegenerative disease early. An imaging institute in Redwood City completed the network of four.

1. 2016
   Founded
   The Chan Zuckerberg Biohub opens in San Francisco with $600 million over ten years and partners at UCSF, Stanford and UC Berkeley.
2. 2020
   COVID-19 testing
   Biohub, UCSF and the Chan Zuckerberg Initiative build a clinical testing laboratory in eight days and offer free testing to every California county.
3. 2021
   Network funded
   The Chan Zuckerberg Initiative commits $1 billion to a Biohub Network of new institutes beyond the Bay Area.
4. 2023
   Chicago and New York
   Biohubs are announced in Chicago, focused on tissue inflammation, and New York, focused on engineered immune cells.
5. 2025
   AI becomes the centre
   The EvolutionaryScale team joins, Alex Rives becomes head of science, and Biohub becomes the founders' primary philanthropic effort.
6. 2026
   Virtual Biology Initiative
   A $500 million data programme launches in April and is expanded in October with federal agencies and three AI companies.

The decisive change came on 6 November 2025. Zuckerberg and Chan made Biohub their primary philanthropic effort, with the Chan Zuckerberg Initiative as its infrastructure and support; they reported having given $4 billion to basic science since 2016, excluding operating costs, and said they were on track to double that over the next decade.³ The same day Biohub announced that the team of EvolutionaryScale, a public-benefit AI company that builds protein language models, was joining, with its co-founder and chief scientist Alex Rives as head of science, and that it would expand computing capacity tenfold to 10,000 GPUs by 2028.⁴ Financial terms were not disclosed.

§Research programme

§Cell atlases and data platforms

Biohub's best-known outputs are reference datasets and software. It supported Tabula Sapiens, a 2022 single-cell atlas of about 500,000 cells from 24 human tissues whose coordinators included Quake, and the Chan Zuckerberg Initiative's CZ CELLxGENE Discover platform hosts a standardized corpus of more than 93 million community-contributed cells.⁵ These collections became training data for machine-learning models of cells, including the Initiative's own TranscriptFormer, a generative model trained on 112 million cells from 12 species and described in a 2025 preprint.

§AI models

The arrival of the EvolutionaryScale team brought the ESM family of protein language models, which Rives had started at Meta, under Biohub's name. In May 2026 Biohub released a new generation, including the ESMFold2 structure predictor, as fully open models, working in the same area of AI protein design as Google DeepMind and Isomorphic Labs. Its 2025 announcement also listed in-house models including rBio,⁴ a language model trained to reason about cell biology using virtual-cell simulators as a source of reward; like TranscriptFormer, it was described in a preprint rather than a peer-reviewed paper.

In July 2026 Biohub researchers reported a genome-wide CRISPR knockout screen of primary adult human skin cells that pointed to two existing drugs as candidate psoriasis treatments; topical gels of both reduced disease severity in a mouse model, a result that has not been tested in people.

§Imaging

The Redwood City institute develops imaging systems meant to show the dynamic architecture of living systems, from single proteins to whole organisms, and Biohub runs an open portal for cryo-electron tomography data. In June 2026 Biohub and UC Berkeley reported a laser phase plate that improves contrast in electron microscopy.

§Infectious disease

IDseq, later renamed CZ ID, is a free cloud pipeline that identifies pathogens in metagenomic sequencing data, designed for laboratories with limited computing resources.⁶ A 2021 Biohub study sequenced 148 Californian mosquitoes individually and found 70 viral taxa, 46 of them previously undescribed. In March 2020 the San Francisco team converted UCSF space into a certified COVID-19 testing laboratory in eight days, offered free testing to all 58 California counties, and had processed more than 150,000 tests by October of that year.

§The Virtual Biology Initiative

Biohub launched the Virtual Biology Initiative on 29 April 2026 with $500 million over five years: $400 million for its own data generation and measurement technology and $100 million for outside research. It promised to make the data openly and freely available.⁷

On 7 October 2026 Biohub, the US Department of Energy and the National Institutes of Health announced an expansion that the partners valued at $1.8 billion in funding, data, computation and measurement technology, and described as the largest coordinated commitment to generating AI-ready biological data to date. DOE plans more than $500 million over five years for laboratory measurement, modelling and computation, delivered through the Genesis Mission, the AI-for-science programme created by executive order in November 2025. NIH contributes no new cash: through what it calls its Bio Genesis Mission it will coordinate existing datasets and repositories built with more than $500 million of prior federal investment, which Biohub will help standardize for model training. Google DeepMind, Isomorphic Labs and Meta are collectively investing $300 million, with no breakdown given; NVIDIA supplies computing infrastructure and Renaissance Philanthropy is helping raise further funds.⁸

Reuters reported terms the releases omit: data paid for by the corporate funders is embargoed so that they can work on it before public release, while the government-funded work carries no such restriction; Rives said he expected a first dataset in about a year and accurate predictive models within five years.⁹ Widely shared summaries presented it as a new $1.8 billion launch; in fact it expanded the April programme, Biohub's share was the April commitment, and NIH's $500 million is past federal spending on existing data.

§Funding

Biohub is funded by its founders through the Chan Zuckerberg Initiative and, since 2026, by programme partners. Every figure above is a commitment rather than audited spending, over differing periods. The arrangement has drawn scrutiny since 2015, when the founders placed their giving in a limited liability company rather than a foundation; commentators noted that an LLC faces no lobbying limits or annual payout requirement, and Zuckerberg replied that the couple received no tax benefit from it. Biohub itself, as a 501(c)(3), files the public returns the LLC does not.

§Reception

Biohub's open datasets and software are community resources across single-cell biology, and its open-release practice is the least disputed part of its record. The 2025 consolidation drew more mixed comment. The Associated Press noted that it followed the Chan Zuckerberg Initiative's decision earlier that year to end grants for diversity, equity and inclusion and immigration advocacy, and that it came as the federal government was cutting billions from scientific research; the founders said the shift toward science predated the 2024 election.³

The October 2026 arrangement raises questions of concentration: one of the corporate funders with early access to embargoed data is Meta, whose chief executive co-founded Biohub, and the companies paying for the largest datasets can train on them first. The scientific premise is also contested. Independent benchmarks have been unflattering to current models: a 2025 comparison found that none of seven deep-learning systems, five of them single-cell foundation models, beat deliberately simple baselines at predicting the effect of a genetic perturbation.¹⁰ Biohub's answer, that the models are starved of data, is the hypothesis its programme exists to test, not a finding.

§Outlook

The test Biohub has set itself is whether a model trained on its data can predict a cell's response to a perturbation it has never seen, and whether the prediction holds in the laboratory. The five-year horizon for accurate models is a forecast by the party funding the work. Unanswered questions include the length of the corporate embargo, how consent and access are governed for donor-derived data, whether predictions about a person's cells could inform decisions by insurers or employers, and whether unrepresentative donors carry the equity problem of biomedical datasets into the models. If data alone proves insufficient, the programme would still leave a large open collection of measurements of human cells; whether it yields a working model of a cell, or a step toward patient-level simulation, is what the next five years are meant to show.

§See also

• Virtual cell models
• Isomorphic Labs
• AI protein design
• AI drug discovery
• Human digital twins
• Calico Life Sciences
• Altos Labs
• Buck Institute for Research on Aging

§References