Accelerating Drugs Discovery with Deep Reinforcement Learning: An Early Approach

Clinical research can be remarkably enhanced making use of virtual screening techniques that predicts how ligands interact with pharmacological receptors. This accelerates the long and expensive process of finding new drugs. Current methods often involve computer clusters that require high computational cost and several programming models. As an alternative to alleviate these problems, we developed DQN-Docking, a novel approach that takes advantage of the latest breakthroughs achieved in deep reinforcement learning by applying a Deep Q-Network to the context of protein-ligand docking prediction. The goal is to design a system able to "teach" the agent--the ligand--how to successfully couple with the receptor by an iterative trial-and-error process. A Deep Q-Network estimates the Q-value function to guide the agent to move in the environment, while METADOCK, a parallel metaheuristic schema for virtual screening methods, provides the resulting state of those movements and its corresponding docking score. Preliminary results seem promising but more research needs to be conducted to refine DQN-Docking and make it an effective alternative to solve the molecular Docking problem.