Biohub, the Chan Zuckerberg-funded biomedical research institute, released an open three-layer AI system for protein biology on May 28, putting a foundation model, a structure-prediction engine, and a navigable protein database into the hands of any researcher with an internet connection. The announcement positions Biohub as a direct challenger to DeepMind’s AlphaFold lineage and the prior Meta ESM family, from which Biohub’s lead scientist Alex Rives originally emerged.
The three components work as a stack. ESMC is the foundation model, trained on roughly 2.8 billion protein sequences drawn from across all of life. The training objective is minimal: predict which amino acids evolution selects. Because evolution preserves proteins that function, the patterns encoded across billions of years of data implicitly encode the physical rules that govern how proteins fold and interact. ESMC produces sequence representations that downstream components use as inputs.
ESMFold2 takes those representations and produces atomically-resolved 3D structures of proteins and biomolecular complexes. Biohub’s release announcement states that ESMFold2 outperforms AlphaFold 3 on antibody-antigen binding pose prediction when using ESMC representations alone, and leads both general protein-protein and antibody-antigen benchmarks when both models are given the same evolutionary information. The release was accompanied by a preprint describing binder design against five cancer and immunology targets, including PD-L1 and CTLA-4, with lab-validated hit rates of 36 to 88 percent for compact minibinders.
ESM Atlas is the navigable database layer: 6.8 billion protein sequences and 1.1 billion predicted structures organized by relationships the model has learned, surfacing evolutionary connections that existing curated databases have not captured. Biohub describes it as the largest application of AI to protein biology to date.
The open-weight release is the structural differentiator against AlphaFold. DeepMind open-sourced AlphaFold’s code and released a large prediction database, but the full AlphaFold 3 model weights were not made freely available at launch, and running the system at scale required compute most academic labs cannot easily provision. Biohub is publishing all three layers as freely accessible tools through the Biohub Platform, with no stated licensing restrictions for the scientific community.
Skepticism is warranted on the benchmark claims. Biohub’s announcement states ESMFold2 leads across standard protein folding benchmarks for protein-protein and antibody-antigen prediction, and explicitly claims it surpasses AlphaFold 3 on antibody-antigen complex prediction. The benchmarks cited are from Biohub’s own preprint, not independent evaluation. RoseTTAFold All-Atom, Chai-1, and Boltz-1 are not addressed directly in the release. Independent replication on held-out targets by groups without a stake in the result will be necessary before the state-of-the-art label settles. The preprint has not yet passed peer review.
The binder design results are more concrete. Designing functional protein binders against five therapeutic targets in days rather than years, with confirmed lab binding and T cell signaling restoration for PD-L1, is a meaningful demonstration if the results replicate. A single preclinical binder candidate typically takes three to four years to develop under conventional screening approaches, according to Biohub’s own release.
For biotech and pharma teams, the immediate evaluation question is whether ESMFold2 plus ESM Atlas can compress the early-stage binder discovery cycle enough to change R&D economics. The PD-L1 result is the headline proof point: a known checkpoint target with validated biology, where the model produced binders that worked in cell assays. Teams running antibody or minibinder discovery programs should run the Biohub stack against their current hit rates before the next platform contract renewal. For closed-stack structure prediction vendors, an open, openly accessible system that claims benchmark parity with AlphaFold 3 on therapeutic targets narrows the moat they can charge for.
Published by Biohub on 2026-05-28.
