Zuckerberg’s $500M Plan to Cure All Disease With AI Cells

What was actually announced

On April 29, Biohub published a public commitment to spend $500 million over five years on what it calls the Virtual Biology Initiative. The split is $400 million for Biohub’s internal effort to generate massive new datasets and develop next-generation tools for measuring, imaging, and engineering biology, and $100 million in grants to outside scientists and consortia who can help nucleate a coordinated data-generation effort worldwide.

The end goal is a working AI model of a human cell. Not a cartoon, not a static atlas — an actual predictive model that takes a cell’s state plus a perturbation (a drug, a virus, a mutation) and outputs what the cell will do next. If that sounds like the kind of thing biology has been chasing for decades, it is. The new ingredient is generative AI applied to biological data at a scale that wasn’t possible until very recently.

Where the money is going

The $400 million internal budget covers three things. First, generating new biological data — specifically the kinds of cell-level measurements that AI models need to learn from. Biohub’s leadership has said current public datasets cover roughly a billion cells, and they want at least ten times that. Second, building the AI infrastructure to train and serve cell-prediction models. Third, the wet-lab tooling needed to validate the predictions in real cells.

The $100 million in outside grants is structured to seed the broader ecosystem. The named partners (Allen Institute, Arc Institute, Broad Institute, Wellcome Sanger Institute, plus the Human Cell Atlas and Human Protein Atlas consortia) are some of the most respected basic-biology institutions in the world, and they bring datasets that no single lab could replicate.

The EvolutionaryScale acquisition

Folded into the same announcement: Biohub is acquiring EvolutionaryScale, the AI-for-biology lab co-founded by Alex Rives. EvolutionaryScale is best known for ESM3, a frontier protein-language model that the team open-sourced in 2024. Rives is joining Biohub as Head of Science, where he’ll lead the integrated research strategy across data, AI, and experimental biology. He will continue his affiliations with MIT and the Broad Institute.

That’s a meaningful signal. Rives is one of the people who showed protein-language models can predict 3D structure from sequence; bringing him in suggests Biohub plans to attempt the same trick at the cell level. (“Cell-language model” is the obvious next analogy, and it’s the analogy Biohub’s leadership has used in interviews.)

What a virtual cell actually means

Strip out the marketing. A virtual cell is software you can give a question to and get a useful answer from. “If I knock out this gene, what happens to the cell?” “If I expose this immune cell to this antigen, how does it respond?” “If I treat this tumor cell with this drug at this dose, does it die?” Today, those questions get answered with experiments that take weeks to years. A working virtual cell would answer some of them in minutes.

The catch is that a “useful answer” has to be more than plausible-sounding text. It has to be quantitatively right often enough that a drug team can use it to choose what to make next. That bar is high. Machine learning models in biology have been getting closer to that bar at the molecular and protein level for several years; the cell level is the next jump.

Who else is working on this

Biohub isn’t alone. Stanford bioengineering professor Emma Lundberg has been leading a parallel push for a virtual human cell, and described it in a Stanford piece this week as “the holy grail of biology.” Lundberg estimates a working version could take ten years or more, and was upfront that nobody yet knows whether the approach scales.

Other groups in the same race include Insitro, Recursion, Genentech’s gRED team, and DeepMind’s Isomorphic Labs (though Isomorphic is more drug-focused than cell-focused). The fact that multiple well-funded teams are now chasing the same target is itself useful information. It means the bet is mainstream enough that frontier labs and biotech companies are willing to commit nine-figure budgets, and competitive enough that progress is likely to come from several directions at once.

The skeptics’ view

Veteran immunologist Jeffrey Bluestone told Science magazine, “It ain’t going to work.” He added that the project could still help in other ways even if the headline goal isn’t reached — better datasets, better tooling, better trained scientists — but the core skeptical case is that human cells are heterogeneous, dynamic, and context-dependent in ways that protein-folding wasn’t, and the data needed to capture that is still being invented.

That’s a fair caveat. Protein folding worked because evolution had already done the brutal work of curating sequences that fold reliably; cell behavior doesn’t have the same evolutionary scaffolding the AI can lean on. Whether Biohub’s data-generation push closes that gap is the unanswered question. The Initiative is a five-year bet that it does.

What this could mean for patients

Three concrete things, in roughly increasing order of “wait and see.” First, faster drug-discovery cycles, especially for diseases like rare cancers and autoimmune conditions where current animal models are weak. Second, more personalized medicine — if the model can predict how your specific cell type responds to a drug, doctors can pick treatments with fewer rounds of trial and error. Third, in the longest term, the ability to test interventions on a virtual version of the patient before testing them on the patient. That last one is the kind of payoff that justifies a $500 million bet.

None of this is “cures coming this year.” Even on optimistic timelines, virtual-cell-driven drugs are a 5-to-15-year horizon. But the upstream work, the data, the tools, the trained scientists, has near-term value regardless of whether the headline goal hits.

The bottom line

$500 million is real money even by Silicon Valley philanthropy standards, and committing it to open data and open tools rather than a single proprietary model is the right move for biology. Whether it produces a working virtual cell by 2031 is genuinely unknown. What’s not unknown is that the AI-for-biology field just got a lot better resourced. Beginners in AI tracks the AI-meets-biology stories in plain English — subscribe to the daily newsletter to follow how this one develops over the next five years.

Sources and further reading

• Biohub: Virtual Biology Initiative announcement
• Alex Rives team page at Biohub
• PR Newswire: Biohub launches Virtual Biology Initiative
• Stanford: Emma Lundberg on building a virtual human cell
• What is machine learning? — Beginners in AI glossary