Molecular docking is an important method in the early stages of the drug discovery and design pipeline as it helps to screen and predict the binding conformation and affinity of potential drugs to protein targets.
Conventional methods for docking typically utilize the computationally expensive search-and-score approach to sample the conformation space of ligands in the pocket of receptors.
DiffDock is a novel deep learning (DL)-based approach developed by researchers at MIT. While DiffDock is not the first DL method for docking, it is the first generative diffusion model for the task. Previous DL methods tackle docking as a regression problem limiting their speed, accuracy and diversity of poses generated. DiffDock, however, approaches the problem as a generative modeling task, building a diffusion process over the space of the ligand poses.
DiffDock comprises two models, the generative score model that produces a set of ligand poses bound to the target, and a confidence model that ranks the generated poses. Hence, DiffDock is able to provide a confidence estimate of its predictions, making it easier to identify high likelihood poses.
DiffDock shows improved performance over existing state-of-the-art physics-based and deep learning methods. In blind docking tests, DiffDock achieves 38.2 % accuracy for the top-1 pose (ligand RMSD below 2Å), compared to 21.8 % and 20.4 % by physics-based GLIDE and the next best DL TANKBind method. When using predicted protein structure as input, DiffDock generates poses with <2Å ligand RMSD in 21.7 % of test cases whereas the next best method achieves only 10.4 %. In terms of speed, DiffDock is up to 12 times faster than the best search-based method.
The most recent version of DiffDock, named DiffDock-L, has seen improved performance, including higher accuracy (from 38.2 to 43.0 %, for ligand RMSD <2Å), 2x increase in speed, and higher generalization ability to unseen targets.
References and Resources
DiffDock Paper: DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking
DiffDock-L Paper: Deep Confident Steps to New Pockets: Strategies for Docking Generalization
GitHub (now runs DiffDock-L by default): https://github.com/gcorso/DiffDock
Try DiffDock-L for free in the Genophore platform. Genophore is integrating molecular docking tools into discovery workflows, including DiffDock-L. Docking results can be visualized using our easy-to-use MolX molecular visualization tool.
