Reinforcement Learning with Real-time Docking of 3D Structures to Cover Chemical Space: Mining for Potent SARS-CoV-2 Main Protease Inhibitors

Li, Jie; Zhang, Oufan; Kearns, Fiona L.; Haghighatlari, Mojtaba; Parks, Conor; Guan, Xingyi; Leven, Itai; Amaro, Rommie E.; Head-Gordon, Teresa

We propose a novel framework that generates new inhibitor molecules for target proteins by combining deep reinforcement learning (RL) with real-time molecular docking on 3-dimensional structures. We illustrate the inhibitor mining (iMiner) approach on the main protease (MPro) protein of SARS-COV-2 that is further validated via consensus of two different docking software, as well as for druglikeness and ease of synthesis. Ultimately 54 molecules are proposed as potent Mpro inhibitors (7 of which have better synthetic accessibility), covering a much broader range than crowd-sourced projects like the COVID moonshot, and our generated molecules exhibit an optimized, precise, and energetically consistent fit within the catalytic binding pocket. Moreover, our approach only relies on the structure of the target protein, which means it can be easily adapted for future development of other inhibitors of any protein of disease origin.