Valence Labs just released Nesso-1, an open-source coarse-grained protein–ligand co-folding model for binding affinity prediction.

Nesso-1 replaces full atomistic diffusion with residue-level protein representations, ligand heavy atoms, ESM-2 embeddings, and a Pairformer-style trunk. It also uses NVIDIA cuEquivariance, which provides optimized GPU kernels for expensive geometric operations such as triangle-based pair updates.

The result is better reported affinity ranking than Boltz-2 across several benchmarks with more than a 10x inference speedup.

A strong release for practical virtual screening and a good example of how model design and GPU-level optimization can expand the scale of chemical space we can screen.