Biohub, the nonprofit biomedical research organization co-founded by Mark Zuckerberg and Priscilla Chan, just dropped what might be the most ambitious open-source toolkit in the history of protein science. The suite includes three AI models designed to map, predict, and design proteins at a scale that was genuinely unthinkable a few years ago.
The centerpiece is the ESM Atlas, which covers 6.8 billion proteins. To put that in perspective, the human body contains somewhere around 20,000 protein-coding genes. This atlas is charting territory several orders of magnitude beyond what any individual lab could catalog in a lifetime.
What Biohub actually built
The release includes three distinct tools working in concert. ESMFold2 handles structure prediction and protein design, essentially letting researchers model how proteins fold and engineer new ones from scratch. ESMC is a protein language model trained on billions of sequences, treating amino acid chains the way GPT treats words. And the ESM Atlas ties it all together as a comprehensive database spanning those 6.8 billion proteins.
The practical payoff is already showing up in the lab. Biohub says the models can design functional binders with therapeutic-level affinity, meaning the AI-designed proteins actually stick to their targets well enough to work as potential drugs. These results have been validated through laboratory testing, not just computational prediction.
All three models are open-source, which means any researcher with the computational resources can download and use them. This is a deliberate choice. Biohub is positioning these tools as public infrastructure for the entire field of protein biology, not a proprietary advantage for one company.
The $500 million bet on virtual biology
This release is part of Biohub’s broader Virtual Biology Initiative, which was first announced on April 29, 2026. The total financial commitment behind the initiative stands at $500 million. Of that, $400 million is earmarked for internal investments, covering the development of models like the ones released today. The remaining $100 million goes toward external data-generation initiatives, funding the kind of wet-lab experiments that produce the training data AI models desperately need.
The nonprofit structure matters here too. Unlike pharmaceutical companies or venture-backed startups, Biohub doesn’t need to recoup its investment through drug sales or licensing fees. That freedom allows it to release everything as open-source without worrying about shareholder reaction.
What this means for investors and the broader market
The drug discovery market has been increasingly gravitating toward AI-driven approaches over the past several years. Open-source protein models of this caliber lower the barrier to entry for biotech startups that couldn’t previously afford to build their own foundation models from scratch. A small team with a good hypothesis and access to ESMFold2 can now compete with labs that have spent years and tens of millions building proprietary alternatives.
Biohub has also been explicit that no cryptocurrency elements or blockchain technologies are involved in this research, emphasizing the organization’s focus on scientific innovation rather than financial speculation.