PAMS Seminar: "DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking" by Hannes Stärk & Gabriele Corso

Hannes Stärk & Gabriele Corso
Massachusetts Institute of Technology
Computer Science & Artificial Intelligence Laboratory

Abstract:

Molecular docking, the process of predicting how small molecule ligands bind to proteins, is pivotal in drug design. Our work introduces a novel perspective by conceptualizing molecular docking as a generative modeling challenge. We present DiffDock, a diffusion generative model that operates on the complex, non-Euclidean manifold of ligand poses, efficiently navigating through the translational, rotational, and torsional degrees of freedom essential for accurate docking. Our empirical results showcase DiffDock's superiority, achieving a 38% top-1 success rate (RMSD<2Å) on the PDBBind dataset, markedly surpassing both traditional (23%) and other deep learning methodologies (20%). Notably, DiffDock demonstrates remarkable performance on computationally folded structures, achieving more than double the accuracy of existing methods. This talk will delve into the intuition and methodology behind DiffDock, and its strengths and limitations.

Hannes and Gabriele are PhD students at MIT in the CS and AI Laboratory (CSAIL).

This is a Zoom seminar: 915 2072 1543

Free