Learning RMSD to Improve Protein-Ligand Scoring and Pose Selection

Docking algorithms are an essential part of the Structure Based Drug Design (SBDD) process as they aim to effectively identify the binding poses of chemical structures at the target site. These algorithms are reliant on scoring functions that evaluate the binding ability of a ligand conformation. Typically, scoring functions are designed to predict the binding affinity of various poses at the target site. In this work, we design a novel approach where the scoring function attempts to predict the Root Mean Square Deviation (RMSD) of a pose to the true binding pose. We show that a Convolutional Neural Network (CNN) can be trained to learn these RMSD values with high correlation between predicted and experimental values. Furthermore we show that this scoring function can improve pose selection performance when used in combination with orthogonal scoring functions like Autodock Vina.