DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking

Code: https://github.com/gcorso/DiffDock
DiffDock (https://arxiv.org/abs/2210.01776) is a deep-learning diffusion generative model that predicts the 3D structure in which a small molecule binds to a protein structure without prior knowledge of the binding pocket. This blind docking task is often evaluated with holo-protein structures, which means that the methods receive the bound protein structure as input - the coordinates it will take on when the small molecule is already bound to it. In most use cases, this structure is not available.

To evaluate how able docking methods are to dock to computationally generated protein structures, we provide this dataset. It contains the time-split based test set used in DiffDock and prior work such as EquiBind (https://arxiv.org/abs/2202.05146), but with the protein structures generated by ESMFold: https://github.com/facebookresearch/esm.

The generated proteins have their pockets RMSD/Kabsch aligned to the original protein structure as described in our paper. Code is provided in the GitHub repository. The RMSDs after alignment are provided as well.