Boltz2: State of the Art Structure and Binding Affinity Prediction

Introducing Boltz-2, an open-source model from MIT and Recursion that predicts both protein structures and binding affinities. It reaches free-energy-perturbation (FEP) accuracy at ~1,000× the speed.

Use it on Tamarind Bio today: https://app.tamarind.bio/boltz

Binding affinity determines whether a drug is potent enough; FEP is precise but slow, while docking is fast but noisy. Deep-learning alternatives have lagged—until now.

Performance:

• Matches OpenFE on the standard FEP+ benchmark while running 1,000× faster.

• Tops the CASP16 affinity challenge (140 complexes).

• In MF-PCBA hit-discovery screens, doubles average precision over ML and docking.

Coupled with SynFlowNet, Boltz-2 powered a prospective TYK2 screen; ABFE predicted binding for all top-10 generated compounds.

For structure prediction, Boltz-2 Matches or exceeds Boltz-1, with major improvements on DNA–protein, RNA, and antibody–antigen complexes.

Specify contact constraints, templates, or experimental methods directly in the model.

Architecture upgrades: Builds on Boltz-1 with a new affinity head, finer controllability, GPU optimizations, and extensive synthetic + MD training data.

Boltz-2 brings FEP-level accuracy to deep learning, at a fraction of the compute cost, and slots cleanly into generative screening pipelines. Excited to see how Boltz-2 will be used for faster small molecule discovery! 

Performance Benchmarks

In addition to protein-ligand interactions, the authors claim meaningful improvements in docking quality for protein-protein and antibody-antigen interactions.

Code & model: https://github.com/jwohlwend/boltz

Paper: https://cdn.prod.website-files.com/68404fd075dba49e58331ad9/6842ee1285b9af247ac5a122_boltz2.pdf