Virtual Biology Initiative

The Virtual Biology Initiative is a five-year research program launched on April 29, 2026 by Biohub, the science organization of the Chan Zuckerberg Initiative (CZI), with a founding commitment of $500 million.^[1][2][3] The initiative aims to generate the large, open biological datasets and computing infrastructure needed to build AI "virtual cell" models, predictive systems that simulate how human cells work in health and disease.^[6] Rather than funding a single model, the program is designed to galvanize a coordinated global effort across leading research institutions, with Biohub committing $100 million to external research and $400 million to internal technology development.^[1][2] At launch it brought together the Broad Institute of MIT and Harvard, the Allen Institute, the Arc Institute, the Wellcome Sanger Institute, the chipmaker NVIDIA as technology partner, and international data consortia including the Human Cell Atlas and the Human Protein Atlas.^[2][4][5] On October 7, 2026, Biohub, the U.S. Department of Energy (DOE) and the National Institutes of Health (NIH) announced an expansion that the partners valued at $1.8 billion in funding, data, computation and measurement technology.^[19][20] The total counts Biohub's founding $500 million, more than $500 million in planned DOE spending over five years, NIH datasets built through more than $500 million in prior federal investment, and $300 million invested collectively by Google DeepMind, Isomorphic Labs and Meta.^[19][23][27] The effort sits at the center of CZI's broader pivot toward AI for science and its founding goal of helping to cure, prevent, or manage all disease by the end of the century.^[7][8] In November 2025 CZI dropped the "managing disease" part of that goal and said it wanted to reach it sooner, without setting a new target date.^[7]

Background: CZI and the virtual cell vision

The Chan Zuckerberg Initiative was founded in 2015 by Meta chief executive Mark Zuckerberg and the pediatrician Priscilla Chan, who pledged 99 percent of their lifetime Meta shares to philanthropy. Its science arm launched the original Biohub in San Francisco in 2016, and CZI has since donated roughly $4 billion to basic science research.^[7] The "virtual cell" idea, an AI model that can simulate the behavior of a cell across molecular, spatial, and temporal scales, became a central organizing goal. In a 2023 essay in MIT Technology Review, Chan and Zuckerberg argued that bringing generative AI to biology at scale could let researchers query simulations of healthy and diseased cells instead of running every experiment at the bench.^[18]

A virtual cell model is broadly analogous to a large language model, but trained on biological measurements such as single-cell RNA sequencing, imaging, and protein data rather than text. The central obstacle has long been data. As Biohub Head of Science Alex Rives put it, building AI that can "accurately represent the full complexity of biology" requires "orders of magnitude more data than exists today."^[1] Existing biological datasets are often siloed, proprietary, and structured to answer one narrow question, which limits their usefulness for training general models. The Virtual Biology Initiative is CZI's bet that solving the data bottleneck is the prerequisite for building useful AI in healthcare.

In April 2025, CZI laid the groundwork by announcing four scientific grand challenges: building an AI virtual cell model; developing new imaging technologies to map biological systems; creating tools to sense and measure inflammation in tissues in real time; and harnessing the immune system for early disease detection and treatment.^[9] The Virtual Biology Initiative is the funding and data engine for the first of these challenges.

The initiative: funding and goals

Biohub committed $500 million over five years to anchor the program.^[1][2][3] The money is split into two parts. About $100 million funds external research to help nucleate a coordinated, worldwide data-generation effort, while roughly $400 million is invested internally at Biohub to generate data at scale and to develop next-generation technologies for measuring, imaging, and engineering biology.^[1][2] The figure and structure were reported consistently by Biohub's own announcement, a PR Newswire release, Axios, and trade press.^[1][2][3][4]

The stated goal is to create the open data foundation for what Biohub calls "predictive models of life," AI systems that let scientists ask and answer biological questions digitally at a scale far beyond what is feasible in a physical laboratory.^[1][2] Biohub's launch announcement said that a high-accuracy predictive model of the cell would require "orders of magnitude more data than is currently available."^[1] Rives has since put the gap in terms of cell counts: existing datasets cover hundreds of millions of cells, while accurate virtual cell models may require data from billions and eventually trillions of cells.^[25] At launch, Biohub said it would make the data it generates open and freely available to the worldwide scientific community.^[1] Under the October 2026 expansion, data developed by the commercial partners is shared publicly after a one-year period of exclusive access for those partners.^[23]

The announcement followed a broader restructuring of CZI. Over 2024 and 2025 the organization wound down most non-science work, including grantmaking tied to diversity, education, and immigration advocacy, and unified its scientific efforts under the Biohub name.^[7][8] Reporting in late 2025 described CZI shifting the bulk of its philanthropy toward AI-powered biology, operating at roughly $1 billion per year and stating it was on track to double its cumulative $4 billion in science giving over the following decade.^[7][8] The Virtual Biology Initiative is the largest concrete expression of that pivot.

October 2026 expansion

On October 7, 2026, Biohub, the DOE and the NIH announced "a major expansion" of the initiative together with new funding partners.^[19][20] The organizations put the combined commitment at $1.8 billion in funding, data, computation and new measurement technology, which Biohub called "the largest coordinated commitment to generating AI-ready biological data to date." The headline of the release rounded the figure to "nearly $2 billion."^[19] The total mixes new money with existing resources: the NIH share consists of datasets and repositories built with earlier federal funding, and Biohub's portion is the $500 million it had already committed in April.^[23][25][27]

Contributions

Government partners

The DOE announcement describes a memorandum of understanding signed by the DOE Office of Science, the NIH and Biohub, and presents the partnership as delivering on the executive order that launched the Genesis Mission.^[21] DOE said its money covers fundamental cell research spanning data collection, analytics, laboratory measurement and imaging, modeling and computation. The resources it named include exascale supercomputers; user facilities such as the Joint Genome Institute and the Environmental Molecular Sciences Laboratory; X-ray and neutron scattering, cryo-electron microscopy and cryo-electron tomography; robotic and automated laboratories, including the Genesis Mission's Orchestrated Platforms for Autonomous Laboratories (OPAL) project; and the BRIDGE and LAMBDA biological data projects.^[21] The DOE and NIH releases describe the goal as "Super Intelligence (SI)" in biology, where Biohub's release says artificial intelligence. ExecutiveGov reported that a September 29 executive order directs federal agencies to use the term in place of artificial intelligence.^[21][22][26]

The NIH contribution runs through its Bio Genesis Mission, which will bring together existing biomedical datasets, national data infrastructure and research programs. The resources named include national biomedical repositories catalogued by the National Library of Medicine and the National Center for Biotechnology Information, and NIH Common Fund programs that are already developing coordinated biological atlases, shared data standards and AI-ready datasets.^[19][22] NIH said the effort aligns with the Predicting Living Systems National Science and Technology Challenge.^[22] Nicole Kleinstreuer, NIH Deputy Director for Program Coordination, Planning, and Strategic Initiatives, said combining resources could "accelerate the development of universal cell models with sufficient biological complexity to predict how any cell responds to an intervention."^[19][22]

Industry funders and data access

The three companies are collectively investing $300 million to create the technologies and multimodal datasets needed for predictive models of life.^[19] Max Jaderberg, president of Isomorphic Labs, said the company was joining the initiative "as a founding member" and was "helping build a massive, multimodal data foundation."^[19] Pushmeet Kohli, Google DeepMind's vice president for AI for Science and Google Cloud's chief scientist, said the investment "will help create an open, standardized data commons."^[19]

Axios reported that the commercial partners will have one year of exclusive access to the data they develop before it is shared publicly. "We have to have some incentive for commercial players to be a part of this, and the embargo period creates that," Rives told Axios. "But it's a one-year embargo. So what that means is that very rapidly, data becomes available broadly to scientific efforts."^[23] eWeek, citing Reuters, reported that government-funded work will not face the same restriction.^[25]

Scientific partners and outlook

The October release named the Allen Institute, the Broad Institute, the Gladstone Institutes, the Human Cell Atlas, the Human Protein Atlas and the Wellcome Sanger Institute as institutions and consortia helping to organize the scientific community around the effort. The Gladstone Institutes had not been among the institutions listed at the April launch.^[1][19] NVIDIA continues to support the initiative with accelerated computing infrastructure, domain-specific software and technical expertise, and Renaissance Philanthropy is helping to expand funding for data generation.^[19] Biohub said it would build a layer that lets partners' datasets work together, with "shared standards, common identifiers, and a single point of access," drawing on its experience with projects such as Tabula Sapiens, OpenCell and Zebrahub and with data infrastructure such as CELLxGENE and the CryoET Data Portal.^[19]

Rives told Axios that within a year of having the first large-scale dataset, researchers should be able to train models, measure their capabilities and work out which kinds of additional biological data make them better.^[23] According to eWeek, he told Reuters that he expects a first dataset in about a year and accurate predictive models within five years.^[25] Interesting Engineering noted that the goal remains prospective: the initiative has not yet produced a universal virtual cell able to predict arbitrary biological responses.^[24]

Partners and infrastructure

The initiative is structured as a consortium rather than a grant to a single lab. Biohub coordinates participating institutions and consortia, each contributing data-generation capacity, expertise, or computing resources. It builds on Biohub's prior support for large-scale atlases including the Human Cell Atlas, the Billion Cells Project (which coordinates 17 projects across institutions such as MIT, Stanford, UC San Francisco, Columbia, the University of Washington, ETH Zurich, and the Genome Institute of Singapore), and the Tabula Sapiens multi-organ cell atlas.

NVIDIA's role centers on compute. CZI operates one of the largest computing systems dedicated to nonprofit life-science research. Announced in September 2023, the cluster is built around more than 1,000 NVIDIA H100 GPUs; trade reporting described it as 1,024 H100s in an NVIDIA DGX SuperPOD configuration paired with VAST data storage, intended to come online in 2024.^[12][13] The cluster lets CZI train large biological models in-house, an unusual capability for a nonprofit. In October 2025, CZI and NVIDIA expanded their collaboration to develop open virtual cell models, GPU-accelerated tools for harmonizing petabyte-scale biological data spanning billions of cellular observations, and a shared evaluation suite called cz-benchmarks, all delivered through CZI's open Virtual Cells Platform.^[10][11]

AI virtual-cell models

The Virtual Biology Initiative is deliberately distinct from any specific model: it funds the data and infrastructure, while the models are built on top. CZI and its collaborators have already released several virtual-cell models that the initiative is designed to feed.

TranscriptFormer, posted as a bioRxiv preprint on April 25, 2025, is a generative, transformer-based model trained on single-cell transcriptomics from more than 110 million cells across 12 species, spanning roughly 1.5 billion years of evolution.^[14] Stephen Quake, then CZI's head of science and a co-corresponding author, said CZI viewed it "as a way to do the equivalent of genome assembly for all the cell atlas data," and it supports tasks such as disease-state identification and gene-interaction prediction.^[14] Like a foundation model for cells, it is meant to generalize across many biological problems.

GREmLN (a gene-regulatory, embedding-based neural model), released in mid-2025 and developed with researchers at Columbia University, takes a more specialized, graph-aware approach.^[15] Instead of treating each gene as an isolated token, it integrates gene-regulatory network structure into its attention mechanism, learning embeddings that capture each gene's role in the broader network. It was initially trained on about 11 million single-cell RNA-seq profiles spanning 162 cell types across tissues including brain, lung, kidney, and blood, and is aimed at applications such as pinpointing cancer-associated cell states.^[15] These transformer- and graph-based systems are extensions of modern deep learning applied to molecular biology, and they rely on the same kind of neural network architectures used in language and vision.

Beyond single-cell models, Biohub released ESMFold 2 on May 27, 2026, an open protein-biology engine from the team led by Alex Rives, whose EvolutionaryScale team joined Biohub in November 2025.^[16][17][33] ESMFold 2 builds on the ESM family of protein language models and shipped with an atlas of more than one billion predicted protein structures, positioning it as an open alternative to structure-prediction systems like AlphaFold.^[16] CZI's portfolio also includes rBio and the RNA model CodonFM. The initiative is not the only effort of its kind; the Arc Institute's State model and a 2025 Arc Virtual Cell Challenge reflect a wider race to build predictive cell models.

Significance

The Virtual Biology Initiative is notable for its scale, its open-data philosophy, and what it signals about the direction of one of the world's largest science philanthropies. By committing $500 million primarily to generating shared, open datasets rather than to a proprietary model, Biohub is treating data scarcity, not algorithms, as the binding constraint on AI for biology. That framing echoes broader debates in machine learning about whether progress is limited more by compute and data than by model design.

For CZI, the initiative consolidates a decade of work into a single coordinated push behind the virtual cell, aligning its in-house GPU cluster, its model portfolio, and a network of elite research institutions. If successful, the resulting datasets and models could let researchers simulate how cells transition from healthy to diseased states and test interventions computationally before bringing them to the bench. The October 2026 expansion added federal agencies and three technology companies to the effort. eWeek pointed out that the $1.8 billion headline combines new commitments with existing scientific resources rather than entirely new cash, and that faster drug development remains a goal rather than a demonstrated benefit until laboratory tests validate model predictions.^[25] Data generated with commercial funding will also reach the wider research community only after the one-year embargo.^[23] The program's significance will ultimately depend on whether the open data it generates proves sufficient to train models that make reliable, testable biological predictions.

Reception

Coverage of the April launch and the October expansion mostly relayed the funding figures, the partner list and Biohub's account of the plan. The outside commentary that did appear concentrated on how the $1.8 billion total was counted,^[25] on the terms of access to commercially funded data, and on how well current virtual cell models perform.

Openness and commercial access

Reuters described the embargo period for commercial funders as the arrangement by which "Biohub is drawing private money into a project it describes as open science." Rives told the agency that the government-funded work would carry no such restrictions and that Biohub planned to approach pharmaceutical companies and philanthropies next.^[28] Reuters also set the initiative alongside other AI groups' biology data efforts, noting that Anthropic had expanded its biology work with a wet lab and that the OpenAI Foundation had started a grant program of more than $125 million for biological and medical datasets.^[28]

Performance of current virtual cell models

Skepticism about the underlying science predates the initiative. In an October 2025 Science feature, six months before the launch, several researchers questioned how well existing AI cell models performed. "Despite the hype, [the models] are underperforming," said Alex Lu of Microsoft Research. Computational biologist Qin Ma of Ohio State University called the number of competing models a bad sign: "One model should be powerful enough that we wouldn't need to see so many of them."^[29] Other scientists in the same article described the problem the initiative was later built to address. Jonathan Karr of the Mount Sinai School of Medicine said, "The data limitation is a massive barrier to cell models," and argued that the field needed a central repository comparable to the Protein Data Bank, which had underpinned AlphaFold's success. Unlike protein structures, he noted, cell data come with no deposit requirement and cannot be characterized with a single technology.^[29] Kasia Kedzierska of the Allen Institute, one of the initiative's partners, had co-authored a study in which simpler methods beat two foundation models in zero-shot tasks such as cell classification, but she said she remained optimistic: "I wouldn't give up on these models. Where they fail gives us a lot of information."^[29]

The benchmark results behind that skepticism are specific. A 2025 Nature Methods study by Constantin Ahlmann-Eltze, Wolfgang Huber and Simon Anders compared five single-cell foundation models and two other deep learning models with deliberately simple baselines at predicting transcriptome changes after single or double gene perturbations, and found that none of the deep learning models outperformed the baselines.^[30] "I was very surprised. Based on the publications, they all looked very good," Ahlmann-Eltze told Science.^[29] The first Arc Institute Virtual Cell Challenge, whose results were announced in December 2025, drew more than 5,000 registrants from 114 countries and submissions from more than 1,200 teams. Its organizers concluded that perturbation prediction models "are not yet consistently outperforming naive baselines across all metrics," that almost all entries did worse than the baseline on mean absolute error, and that hybrid models combining deep learning with statistical features beat pure neural networks.^[31] The baseline findings have also been contested. An October 2026 Nature Biotechnology paper by Henry Miller, Bo Wang and colleagues argued that common benchmark metrics such as mean squared error and Pearson delta are frequently miscalibrated. Across 14 datasets and 18 metrics, the authors reported that deep learning perturbation models can outperform uninformative baselines when the metrics are calibrated.^[32]

None of these evaluations tested models trained on the initiative's own data. Rives told Reuters at the October announcement that the first dataset should be ready in about a year.^[28] Future Human Wiki's article on virtual cell models sets the initiative's premise against these benchmarks and covers the open questions about donor consent and data access.

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Biohub. "Virtual Biology Initiative." biohub.org. biohub.org/...virtual-biology-initiative
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7PR Newswire. "Biohub Launches the Virtual Biology Initiative to Galvanize a Global Effort to Create the Open Data Foundation for AI-Accelerated Biology." April 29, 2026. prnewswire.com/...ai-accelerated-biology-302756743
3. ^1 ^2 ^3Axios. "Mark Zuckerberg and Priscilla Chan commit $500 million to AI biology." April 29, 2026. axios.com/...g-chan-biohub-philanthropy-ai-disease
4. ^1 ^2EdTech Innovation Hub. "Biohub puts $500 million behind AI biology push with MIT, Harvard and NVIDIA." edtechinnovationhub.com/...-mit-harvard-and-nvidia
5. ^ion genomics. "Biohub Leads $500M Push to Generate Biological Data for AI Models." iongenomics.bio/...hub-leads-500m-push-to-generate
6. ^Time. "If AI Can Model Cells, Science Can Deliver Cures." April 29, 2026. time.com/...-model-cells-science-can-deliver-cures
7. ^1 ^2 ^3 ^4 ^5Science (AAAS). "AI drives dramatic expansion of Chan Zuckerberg Initiative's funding to end all diseases." science.org/...itiative-s-funding-end-all-diseases
8. ^1 ^2 ^3Axios. "Zuckerberg, Chan pivot philanthropy to AI-powered biology." November 6, 2025. axios.com/...zuckerberg-chan-biohub-ai-disease
9. ^Chan Zuckerberg Initiative. "CZI Sets Four Scientific Grand Challenges to Transform Human Health at the Intersection of AI and Biology." chanzuckerberg.com/...-grand-scientific-challenges
10. ^Chan Zuckerberg Initiative. "CZI and NVIDIA Accelerate Virtual Cell Model Development." October 28, 2025. chanzuckerberg.com/...rtnership-virtual-cell-model
11. ^PR Newswire. "CZI and NVIDIA Accelerate Virtual Cell Model Development for Scientific Discovery." prnewswire.com/...cell-model-development-302595996
12. ^Data Center Dynamics. "Chan Zuckerberg Initiative to build 1,000+ GPU cluster for biomedical AI research." datacenterdynamics.com/...r-biomedical-ai-research
13. ^The Register. "Chan Zuckerberg Initiative to build giant AI H100 cluster." September 20, 2023. theregister.com/...n_zuckerberg_initiative_to_spin
14. ^1 ^2GEN (Genetic Engineering & Biotechnology News). "The Cross-Species Cell Atlas: Chan Zuckerberg Initiative Releases TranscriptFormer AI Model." genengnews.com/...leases-transcriptformer-ai-model
15. ^1 ^2Chan Zuckerberg Initiative. "GREmLN: A New AI Model Trained in Molecular Logic." chanzuckerberg.com/...ng-ai-think-like-cell-gremln
16. ^1 ^2Nature. "Move over, AlphaFold: open-source model predicts shape of 1 billion proteins." 2026. nature.com/...d41586-026-01686-3
17. ^Axios. "Zuckerberg's Biohub unveils AI 'world model' of proteins." May 27, 2026. axios.com/...ub-unveils-ai-world-model-of-proteins
18. ^MIT Technology Review. "Priscilla Chan & Mark Zuckerberg: How AI can help us understand how cells work and help cure diseases." September 19, 2023. technologyreview.com/...czi-ai-cell-disease
19. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21Biohub. "International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease." October 7, 2026. biohub.org/...virtual-biology-initiative-expansion
20. ^1 ^2PR Newswire. "International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease." October 7, 2026. prnewswire.com/...dict-and-treat-disease-302901172
21. ^1 ^2 ^3 ^4 ^5U.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. energy.gov/...nal-data-predictive-biological-super
22. ^1 ^2 ^3 ^4 ^5 ^6Mirage News. "NIH Joins Effort to Build SI-Ready Data for Human Biology Models." Republication of a National Institutes of Health news release, October 7, 2026. miragenews.com/...-build-si-ready-data-for-1756288
23. ^1 ^2 ^3 ^4 ^5 ^6Ina Fried. "Zuckerberg teams with Google, U.S. in push to map cells." Axios, October 7, 2026. axios.com/...zuckerberg-biohub-ai-biology-data
24. ^Neetika Walter. "Google, Meta, and US build biology datasets for 'virtual cells'." Interesting Engineering, October 7, 2026. interestingengineering.com/...-dataset-google-meta
25. ^1 ^2 ^3 ^4 ^5 ^6Kezia Jungco. "Google, Meta Join $1.8B AI 'Virtual Cell' Push." eWeek, October 9, 2026. eweek.com/...google-meta-virtual-cell-ai-biology
26. ^Mendel Cohen. "DOE, NIH and Biohub Sign Agreement to Advance Super Intelligence in Biology." ExecutiveGov, October 8, 2026. executivegov.com/...intelligence-biology-agreement
27. ^1 ^2Emma Thompson. "Biohub expands $1.8 billion virtual cell AI research effort." EdTech Innovation Hub, October 12, 2026. edtechinnovationhub.com/...ls-for-disease-research
28. ^1 ^2 ^3Krystal Hu. "US government, Google join Zuckerberg-backed Biohub in $1.8 billion push for AI biology data." Reuters, October 7, 2026. Republished by CNBC. cnbc.com/...ckerberg-backed-biohub-ai-biology-data
29. ^1 ^2 ^3 ^4Mitch Leslie. "Can AI capture the mind-boggling complexity of a human cell?" Science, October 30, 2025. science.org/...mind-boggling-complexity-human-cell
30. ^Constantin Ahlmann-Eltze, Wolfgang Huber and Simon Anders. "Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines." Nature Methods 22, 1657-1661, August 4, 2025. DOI: 10.1038/s41592-025-02772-6. nature.com/...s41592-025-02772-6
31. ^Arc Institute. "Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections." December 6, 2025. arcinstitute.org/...al-cell-challenge-2025-wrap-up
32. ^Henry E. Miller, Gabriel M. Mejia, Francis J. A. Leblanc, Brendan Swain, Bo Wang and Lucas Paulo de Lima Camillo. "Deep learning perturbation models can outperform baselines on calibrated metrics." Nature Biotechnology, October 1, 2026. nature.com/...s41587-026-03307-w
33. ^Biohub. "Biohub launches first large-scale scientific initiative combining frontier AI with frontier biology to cure or prevent disease." November 6, 2025. biohub.org/...ai-biology-cure-disease