PandoraRL: DQN and Graph Convolution based ligand pose learning for SARS-COV1 Mprotease

The ability to predict the correct ligand binding pose for protein-ligand complex is vital for drug design. Recently several machine learning methods have suggested knowledge based scoring functions for binding energy prediction. In this study, we propose a reinforcement learning (RL) based model, PandoraRL, where the RL agent helps the ligand traverse to the optimal binding pose. The underlying representation of molecules utilizes generalized graph convolution to represent the protein ligand complex with various atomic and spatial features. The representation consists of edges formed on the basis of inter molecular interactions such as hydrogen bonds, hydrophobic interactions, etc, and nodes representing atomic features. This study presents our initial model which can train on a protein-ligand pair and predict optimal binding pose for a different ligand with the same protein. To the best of our knowledge, this is the first time an RL based approach has been put forward for predicting optimized ligand pose.