DiffDock

DiffDock frames molecular docking as a generative modeling problem by developing a diffusion generative model over the non-Euclidean manifold of ligand poses to predict the binding structure of small molecule ligands to proteins.

Key Features:

• Diffusion Generative Model: DiffDock maps the manifold of ligand poses to the product space of translational, rotational, and torsional degrees of freedom and develops an efficient diffusion process on this space.
• Confidence Estimates: DiffDock provides confidence estimates with high selective accuracy.
• Fast Inference: DiffDock achieves fast inference times compared to traditional search-based methods.
• Improved Version: DiffDock-L represents an improved version of the original DiffDock method.

Scientific Applications:

• Structure-Based Drug Design: DiffDock predicts binding structures of small molecule ligands to proteins with a 38% top-1 success rate (RMSD<2A) on PDBBind.
• Docking to Predicted Structures: DiffDock maintains significantly higher precision (21.7%) when docking to computationally folded structures, where previous methods achieve maximum accuracy of only 10.4%.

Methodology:

DiffDock frames molecular docking as a generative modeling problem using a diffusion generative model over the non-Euclidean manifold of ligand poses, mapping this manifold to the product space of translational, rotational, and torsional degrees of freedom involved in docking.