ESMFold2 Launches: 6.8B Protein Atlas, Edges AlphaFold3
Chan Zuckerberg Biohub and EvolutionaryScale released ESMFold2, ESMC-6B, and ESM Atlas as fully open models and datasets on HuggingScience. The atlas contains 6.8 billion protein sequences and 1.1 billion predicted structures — 800 million more entries than AlphaFold's database — and reportedly edges AlphaFold3 on protein complexes including antibody-antigen binding. The team demonstrated validated miniprotein binders and single-chain antibodies across five therapeutic targets in cancer and immunology, with affinities at levels consistent with therapeutic activity.
Why It Matters
Fully open protein world models at this scale lower the barrier to therapeutic design by orders of magnitude, making biohub's release directly competitive with AlphaFold infrastructure that previously required proprietary access or significant compute budgets.