Biohub's founders say their original goal of curing all disease by century's end now looks too conservative, though the organization does not expect to cure diseases itself.
Zuckerberg argues AI alone is not enough because biology lacks internet-scale data, so Biohub must invent new experimental methods to generate it.
Biohub's Alex Rives says the new ESM Fold, announced a week earlier, folded over 1.1bn proteins and designed binders that reached nanomolar potency in lab tests.
Biohub bets open-source AI models can accelerate all of biology
Biohub is a nonprofit betting that biology advances fastest when AI models are trained on new data it generates itself and then released openly. Biology lacks the internet-scale data behind language models, so AI and wet lab teams work as one, building from proteins to cells to whole systems. Open tools and a long time horizon let it speed up the entire field rather than chase cures alone.
Takeaways
AI needs new biological data to work. Biology lacks the internet-scale data that trained language models, so Biohub must invent experimental methods to generate it. Its AI and wet lab efforts run as one, starting with proteins and moving to cells and whole systems, built up hierarchically.
Biohub aims to accelerate the whole field. The founders set out a decade ago to cure, prevent and manage all disease by century's end. Zuckerberg now calls that too conservative, but says Biohub will not cure diseases itself. It builds tools and studies systems like inflammation so others can build therapies.
Nonprofit and open source maximize impact. Zuckerberg says open source puts tools in more scientists' hands faster. The work needs a ten to fifteen year horizon and more capital for compute clusters than a typical startup. The founders committed $500m to the virtual biology initiative.
ESM Fold release shows protein models working. Announced a week before the episode, the new ESM Fold folded over 1.1bn proteins. Of 96 designed proteins synthesized, lab tests reached nanomolar binders. Rives says protein language models learn structure and function without being taught directly, and interpretability could reveal how.
Virtual models could cut drug development cost. Gil cites about fifteen years and $1.5bn per clinical trial. Zuckerberg hopes a biological model can predict off-target effects, such as kidney toxicity, before human trials. Guo suggests companies would then try many more programs. Gil says an FDA-accepted virtual trial would be exciting.
The virtual cell is the next big challenge. Rives names it as the next frontier and says agentic systems may automate design. Predictions must generalize, which will take an enormous data generation effort. Zuckerberg describes choosing between protein and cell work as constrained optimization.
Treatments should be designed for individual patients. Chan wants to understand each patient's genetics and disease mechanisms. She cites the CHOP CRISPR therapy for baby KJ, a liver-targeted edit chosen because the liver was reachable. Biohub also partners with Jennifer Downey on the CRISPR Cures program at UCSF.
Tap a name to go there · tap a line to see why they are linked