DiffDock

DiffDock (Corso et al., ICLR 2023) reframed molecular docking as generative modeling. Instead of regressing a single pose, it runs a diffusion process over the manifold of ligand poses, defined by translational, rotational, and torsional degrees of freedom, with an SE(3)-equivariant network. Given a protein structure and a ligand SMILES string, it samples diverse candidate poses and ranks them with a separate confidence model.

It does blind docking: no binding pocket needs to be specified, it searches the whole protein surface. The authors report 38 percent top-1 success (RMSD under 2 Å) on PDBBind, versus 23 percent for traditional docking and 20 percent for previous deep-learning methods, and notably it keeps 21.7 percent accuracy on computationally folded structures where baselines capped at 10.4 percent. The confidence score predicts pose quality, not binding affinity, a distinction the authors stress. The follow-up DiffDock-L (ICLR 2024) improved generalization and is now the repo default. Code is MIT-licensed.

Author-reported (Corso et al., ICLR 2023): 38% top-1 success (RMSD < 2 Å) on PDBBind vs 23% traditional and 20% deep-learning baselines; 21.7% on computationally folded structures (baselines max 10.4%); confidence model with high selective accuracy