DockingRL - A Reinforcement Learning Guided Method to Generate Moleculesthat Dock Well to Drug Targets
Drug discovery is an important problem that is quite time consuming and requires extensively combing through existing datasets to identify possible hits. In this study we propose a computational strategy for de novo design of drug molecules that dock strongly to target proteins. This approach combines a generative model along with the binding affinity computed through docking which finds the orientation of the molecule when it binds to a target. The generative model uses a stack-augmented recurrent neural network (RNN) which is first trained separately to generate valid drug like molecules in the form of simplified molecular-input line-entry system (SMILES) strings and is then biased using reinforcement learning during which it is rewarded for molecules that bind strongly to the given target and penalized for drugs that do not. We present the application of this approach on generating drug molecules that bind to SARS-CoV-2 main protease Mpro an important target for COVID-19. This approach can potentially be generalized potentially to generate focused drug libraries for any target.
Manan Goel
Shampa Raghunathan
Siddhartha Laghavarapu
Deva Priyakumar U.
Drug Design.
Algorithm.
Technology designed and implemented.
Find potential drugs for novel targets that can be used as candidates in drug trials.