ESMFold2 and ESM-C: A World Model of Protein Biology, Now on Tamarind
Today Biohub released ESMFold2 and the ESM-C family of models, a fully open-source, MIT-licensed release that represents the next generation of evolutionary-scale protein modeling. Tamarind Bio is a day-1 launch partner. Both models are live on our web interface, API, and AI agent right now.
The Biohub team frames this release as a "world model of protein biology." Based on the lab-validated results released today, the framing is earned.
ESMFold2: state-of-the-art structure prediction
ESMFold2 achieves state-of-the-art accuracy on Foldbench, with particularly strong performance on antibody-antigen complex prediction. It correctly predicts 55% of antibody-antigen complexes, 71% of protein-protein complexes from a single sequence, and 77% with MSAs.
A lighter variant, ESMFold2-Fast, predicts a 1024-residue structure in 9.4 seconds while still outperforming prior models on antibody-antigen folding. That speed is what makes the design results below feasible.
De novo binder design at therapeutic-level affinity
The most striking result in the release is what ESMFold2 can do when used as the backbone for de novo binder design.
The Biohub team designed protein binders against five disease targets in cancer and immunology (EGFR, PDGFRβ, PD-L1, CTLA-4, and CD45) in two formats: minibinders and scFvs. ESMFold2 was not trained or fine-tuned for antibodies. The binders emerged from a simple search through the model's latent space, with candidate generation taking roughly two days and scoring under one.
Hit rates at higher inference compute reached 70% for minibinders and 21% for scFvs.
The PD-L1 result is the standout. An ESMFold2-designed, de novo scFv bound PD-L1 with a measured affinity of 4.3 nM, and in cell-based assays relieved PD-L1-mediated suppression of T-cell signaling with potency comparable to approved checkpoint inhibitor therapies. Designed computationally, in days, with no antibody-specific training.
ESM-C: the next-generation protein language model
ESM-C (Cambrian) is trained on roughly 2.8 billion protein sequences. For teams already using ESM2 embeddings, it is a natural drop-in upgrade for representation-driven workflows: property prediction, similarity search, clustering, variant effect prediction, and downstream generative design.
Try ESMFold2 and ESM-C on Tamarind.
Link
ESMFold2, ESMFold2-Fast, and ESM-C are live on Tamarind today. You can run structure prediction through our web app, API, or AI agent, pipeline ESMFold2 with our 300+ other molecular design workflows (RFAntibody, BoltzGen, and others), use ESM-C embeddings in downstream ML pipelines, and fine-tune on proprietary data using our managed infrastructure.
A structure prediction model producing antibody-format binders at therapeutic-level affinity, with a simple search procedure, under an MIT license, is a meaningful moment for the field.