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

Abstract

Predicting the binding structure of a small molecule ligand to a protein-a task known as molecular docking-is critical to drug design. Recent deep learning methods that treat docking as a regression problem have decreased runtime compared to traditional search-based methods but have yet to offer substantial improvements in accuracy. We instead frame molecular docking as a generative modeling problem and develop DIFFDOCK, a diffusion generative model over the non-Euclidean manifold of ligand poses. To do so, we map this manifold to the product space of the degrees of freedom (translational, rotational, and torsional) involved in docking and develop an efficient diffusion process on this space. Empirically, DIFFDOCK obtains a 38% top-1 success rate (RMSD<2 Å) on PDB-Bind, significantly outperforming the previous state-of-the-art of traditional docking (23%) and deep learning (20%) methods. Moreover, DIFFDOCK has fast inference times and provides confidence estimates with high selective accuracy.

BigBind: Learning from Nonstructural Data for Structure-Based Virtual Screening

“…In recent years there have been many advances in deep learning model architectures for analyzing molecules. [42][43][44][45][46] A deep learning model, however, is only as good as the dataset it is trained on. Previous attempts to utilize machine learning for protein-ligand binding affinity prediction often trained on PDBbind, but the small size and bias inherent to the dataset have hindered their utility.…”

Application of Computational Biology and Artificial Intelligence in Drug Design

“…Physics-based, empirical, knowledge-based, and machine learning-based scoring functions are available [ 202 ]. Additionally, new deep learning methods such as EquiBind, GNINA, DiffDock are developed to predict the binding mode between the ligand and a specific protein target [ 203 , 204 , 205 ]. Especially, Equibind and Diffdock have the potential to significantly change the VS landscape.…”

“…This method significantly speeds up with better quality compared to traditional docking methods [ 203 ]. DiffDock is a diffusion generative model over the non-Euclidean manifold of ligand poses, which has fast inference times and provides confidence estimates with high selective accuracy outperforming the previous traditional docking and deep learning methods [ 205 ]. GNINA utilizes an ensemble of convolutional neural networks (CNNs) as a scoring function and improves the quality of scoring and ranking binding poses for protein-ligand complexes [ 204 ].…”

“…The above-mentioned machine learning based target identification, binding site identification, docking prediction, develop ability predictions, affinity predictions, etc., are the whole or part of the work of drug discovery by screening the desired drug from existing compound data. EquiBind was developed to predict receptor binding location and ligand’s bound pose and orientation by applying geometric deep learning [ 203 ] ; additionally, DiffDock completed the molecular docking task by a diffusion generative approach [ 205 ]. In the following sections, we focus on reviewing the machine learning based frameworks which are relevant to de novo drug design.…”

Recent Developments in Structure-Based Virtual Screening Approaches

“…Both EquiBind and TANKBind are substantially faster than traditional docking programs. DiffDock, which is a diffusion generative model, was developed by the same research group as EquiBind, and their results indicate that it can be more accurate than any of the two previously mentioned blind docking methods (100). On the other hand it is considerably slower than EquiBind or TANKBind, however, it is still faster than most traditional docking programs.…”