Biohub's Bet: Building Open AI World Models of Biology From Proteins Up to Cells

The founders recount that ten years ago Zuckerberg wanted to build an organization that could cure, prevent and manage all disease by the end of the century. Priscilla Chan recalls meetings in which famous scientists laughed at them, and says the team pressed them for why it was impossible. The answers were that scientists work in silos, publications lock information up, and tools die when a postdoc leaves the lab.

Zuckerberg stresses that Biohub does not expect to be the one curing diseases. The aim is to build tools that accelerate the whole scientific field. He once thought the end-of-century goal was a stretch, but he now thinks it is "too conservative."

The original Biohub model brought engineers and scientists together across universities for long-term tool development, and it worked well enough that the team kept investing. Biohub now spans San Francisco, New York and Chicago, and the unifying theme is a virtual biology initiative. Chan notes the work began as an early grant call on single-cell sequencing, grew into the Human Cell Atlas, and spawned the community-built annotation tool CELLxGENE.

Zuckerberg explains the data problem: unlike text, the biological data they want mostly does not exist. So Biohub builds imaging methods, cellular engineering in New York and inflammation-measuring devices in Chicago. As he puts it, there is no factory that will produce this data for you; you must invent new scientific approaches.

Chan says critics once called data gathering "stamp collecting" until large language models showed that vast data can be made sense of. For her the prize is turning biology from discovery-based into engineering-based science, where we can systematically understand why things go wrong. Alex Rives, who had been building his own AI-biology company, says what drew him was seeing the founders treat frontier AI and frontier biology as one integrated effort.

The host Elad Gil, a former biologist, asks whether modeling cells resembles modeling proteins. Zuckerberg says each layer will likely differ qualitatively, but you cannot skip layers, since cells cannot be understood without proteins, and immune-system simulation needs cells first. Rives adds that each level of biology is made up of lower ones, so models need the right basis at every level.

Sarah Guo asks about the mechanistic interpretability work in the release. Rives explains that the toolkit developed for language models can be applied to protein language models, which are trained only on protein sequences. Anything they learn about structure or function is emergent from predicting the next token.

Because the models train on billions of sequences, known and unknown biology alike, their internal representations capture parts of the reductive picture built over centuries. That lets researchers connect proteins about which almost nothing is known to proteins that are well understood. The hope is to open the black box and see the biology the model represents, potentially revealing new mechanisms.

Asked why Biohub is not a venture-backed company, Zuckerberg says he is not sure it could not be run as a business, but the team will have more impact by getting tools into more scientists' hands quickly as open source. He points to the 10-to-15-year horizon, the capital required, and the freedom from having to find a revenue model for each project. He also says data, not compute, is the bigger constraint, since these models are smaller than language models because the data is scarcer.

Chan adds that neutrality helps draw in more people, and that curing disease takes the whole academic and biotech world working in a unified way. Many diseases split into narrow subcategories, and rare ones often get orphaned. If tools are widely distributed, someone who cares deeply about spinal muscular atrophy can make progress, and understanding a niche disease can reveal how the human body works more generally.

Asked which diseases will benefit first, Chan says she does not think that way. She wants to understand a person's genetics, their risks, and the mechanistic chain from a gene variant to a protein to a disease process. Today, she says, patients with anything non-standard search PubMed and guess whether they resemble the people in a study.

Stages differ by disease. For some, simply knowing which variants cause disease is empowering, and for others the chain is known but the protein cannot be changed, so designing a protein that alters physiology could cure someone. On timing, Zuckerberg is optimistic it will be sooner than 100 years but notes biology is dynamic, so new problems will keep appearing.

Zuckerberg says Biohub prefers to study systems like inflammation and the immune system rather than individual diseases, leaving the last mile to biotech and academia. Chan thinks science will move fast but clinical research must change to shorten the distance from bench to patient. Biohub is exploring this through a partnership with a UCSF program.

Rives describes ESMFold2, announced a week earlier, as an open system for scientific discovery in protein biology: a language-model-based world model trained on billions of sequences. It predicts atomic-resolution structure very quickly and reaches state-of-the-art on structure benchmarks, especially for protein-protein and protein-antibody interactions. The team used it to fold over 1.1 billion proteins and to identify features connecting them through interpretability.

The model can also be searched to design new proteins, including single-chain antibodies. The team screened hundreds of thousands of designs digitally, synthesized 96, tested them on a plate, and found nanomolar binders. Rives stresses they built no antibody-specific model; protein design emerged from a general understanding of proteins, and results were checked with the Biohub structural biology center's cryo-EM microscopes.

Gil asks about the long, costly part of drug development. Chan hopes a cell-level model can predict off-target effects, such as a receptor unexpectedly present on kidney cells, before human trials, and says the choice of disease matters, citing the CHOP team's CRISPR treatment for baby KJ. Rives says the lower barrier to designing molecules could open a world of programmable, per-patient medicine.

Zuckerberg says a program called Rare As One illustrates how patient groups can organize to take experimental drugs, which changes the economics of trials where recruitment is a big share of cost. He notes that studying edge cases often reveals more about a system than hitting the common problems. Chan says rare-disease communities already build their own registries and biobanks, and one group advanced a gene therapy in perhaps three to five years rather than decades.

Zuckerberg adds that the general population will still want the vetting of the past, but allowing willing people to opt into frontier trials would help. Asked whether Biohub changes his view of AI at Meta, he says the philosophy is the same: put technology in individuals' hands. He does not believe in a small number of institutions, or one central superintelligence, advancing everything, and he flags biosafety as a consideration to balance.

On recruiting, Zuckerberg says AI researchers are in demand and could work at any main lab, but those labs lack a frontier biology effort attached, and he says no other organization does both. Rives calls the mission simple and says people are deeply motivated by it now that it looks achievable. Zuckerberg adds that you do not need hundreds of researchers, and that a strong group of a dozen or two dozen can make real progress.

Looking ahead, Rives is optimistic that hard problems like toxicity can be solved in an emergent way once accurate molecular models exist. He says people are already wiring ESMFold2 into agentic systems to automate design. The next big challenge is the virtual cell, a system spanning the protein, genetic and transcriptomic layers and connecting to phenotype, general enough to answer questions about interventions it has never seen.

Deciding what to do next is a normal constraint-management process, Zuckerberg says, since every lab feels compute-constrained and must balance protein work against cell work. Closing the gap needs an enormous effort to generate data. Five years out, he wants Biohub to have done meaningfully better, uniquely valuable work, and expects idea generation from users of the models to grow.

Asked for the biggest update of the past year, Zuckerberg says Biohub is now formally the main focus of the couple's philanthropy. He also stresses that Rives's team is talented, stable and used to working together, which compounds. He notes the change of leadership: earlier Biohub leaders were mostly biologists interested in technology, while Rives is primarily an AI researcher with a biology background.

On the broader industry, Zuckerberg says an exponential curve feels like it cannot possibly continue, yet it keeps accelerating, and AI remains on track, which validates making a big investment. Guo notes that the key is closing the loop with real biology, since code and research are closed-loop and fast to iterate while biology is open-loop. Chan says Biohub's teams now move arms linked toward a single directed goal.