Biohub's ESMFold2 Beats AlphaFold3 on Antibodies With a 0.2B Open Model
Biohub's open-source ESMFold2 beats AlphaFold3 on antibody-antigen prediction, adds inference-time compute scaling, and predicts proteins with DNA, RNA, and ligands.
Takeaways
- Biohub released ESMFold2, an open MIT-licensed structure prediction model beating AlphaFold3 on antibody-antigen targets.
- Recurrent architecture with looped representations and stripped-down pair layer (no triangle attention) enables inference-time compute scaling.
- ESMFold2-Fast hits 65% pass rate after 1000 seeds on antibody-antigen prediction, beating AlphaFold3.
- Predicts all-atom complexes of proteins, DNA, RNA, ligands, and modified residues in one pass.
- Used to design binders against EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45 in cancer and immunology.
- Shipped with an atlas of 6.8 billion proteins and 1.1 billion predicted structures.
Biohub just dropped ESMFold2, a successor to the original ESMFold that overtakes AlphaFold3 on some of the hardest problems in structure prediction, including antibody-antigen complexes. It ships under an MIT license, with weights, inference code, and a companion atlas of more than a billion predicted structures all freely available.
The release is the structure-prediction head of a larger system. Biohub is making available an open discovery engine for protein structure prediction, design, and biological discovery built around three releases: ESMC, ESMFold2, and ESM Atlas. At its foundation is ESMC, a state-of-the-art language model that represents proteins, trained on approximately 2.8 billion sequences drawn from across all of life. ESMFold2 is the head that turns those sequence representations into atom-resolution 3D coordinates.
What is actually new
ESMFold2 predicts high-resolution, all-atom 3D structures directly from amino acid sequences, with optional MSA input for harder targets. The output bundle is a full structural report: backbone and side-chain coordinates, confidence metrics (pLDDT, pAE, pTM, iPTM), and optional distogram predictions.
The most notable upgrade over the original ESMFold is biomolecular scope. It can predict high-resolution, all-atom 3D structures of biomolecular complexes directly from protein, DNA, and RNA sequences, plus small-molecule ligands and modified residues. That puts it in the same problem class as AlphaFold3 rather than the original ESMFold, which only handled protein backbones.
The architecture trick: looping the network back on itself
ESMFold2 uses a recurrent architecture where representations from later stages are fed back into earlier stages, with a constraint applied to the recurrent update so activations do not blow up, and gradients are backpropagated through multiple loops. In practice, you control quality with a