Reinforcement Learning Strategies for De Novo Drug Design: an Overview
Yang Xuelan, Jasni Mohamad Zain, Tao Hai · 2024
Recently, reinforcement learning (RL) has led to a great advance of drug design, especially such in virtual screening (VS) and high-throughput screening (HTS) areas. This paper aims at introducing RL technique in an end to end de novo drug designing scheme and demonstrate how a potential RL based model could be adopted to navigate the gigantic chemical space to find novel drug molecules.Traditional methods, while effective, are limited in their scope and efficiency, often constrained to predefined chemical libraries. By contrast, Reinforcement Learning (RL) represents a dynamic design strategy to iteratively improve both molecular structures and screening routes from real-time feedback. This process leads to the generation of compounds with desirable biological activities and physicochemical properties. In this review, several RL frameworks including FREED, MolDQN, etc., are discussed and results from a number of studies help to highlight important aspects in the applications of these algorithms to VS, and HTS as well as how it could potentially improve the process of drug-like molecule discovery and optimization. Additionally, the use of RL in bias control and chemical reaction optimization is discussed, showcasing the potential of RL to revolutionize experimental design in drug discovery. These results support the value of RL-informed generative models for broadening the space of molecular design and advancing new therapeutic agents.