GeoDirDock (GDD): our new diffusion model for directed docking
We’re excited to introduce GeoDirDock (GDD), our new diffusion model for directed docking.
Diffusion models often overlook that the docking site on a protein is typically predetermined, along with specific molecular interactions at that site. GDD leverages this information, employing a directed diffusion process that guides docking to the desired pose through geodesic paths in implicit spaces.
Our approach not only surpasses other rigid diffusion models like DiffDock in PDBBind standard benchmarks but also generates more accurate molecular poses. These poses show fewer clashes and maintain better proximity to side chains and backbones, enhancing the plausibility and effectiveness of the docking predictions.
See in the examples below how GDD (green) it’s able to find the “right” pocket to bind (crystal ligand in red) while DiffDock (grey) explores a cavity that is not subject to the screening. Also, for large molecules, the imposed prior it’s able to fix most of the torsional angles and sets GDD to get the pose right. Even challenging for traditional docking methods!
Stay tuned for the upcoming release of GDD, which will be made accessible to all users soon.
This work has been accepted at the Generative and Experimental Perspectives for Biomolecular Design Workshop at ICLR. You can check out the paper here
