Molecular Generation Strategy and Optimization Based on PPO Reinforcement Learning in De Novo Drug Design

The drug discovery process tends to be grueling, lengthy and expensive, with cost estimates approximating $2.6 billion, consuming over 10 years to complete. Such drawbacks have set many eyes locked onto reducing the costs and accelerating the development. The emergence of Deep Reinforced Learning (DRL) within cheminformatics and bioinformatics has broadens the horizons of de novo drug design. Realizing the full potential of DRL in molecular generation requires selecting the appropriate reinforcement learning (RL) algorithm. In this work, we address these problems by utilizing Proximal Policy Optimization (PPO) algorithm within the DRL framework for molecular generation. We proposed a new method by utilizing PPO algorithm within the DRL framework, termed PSQ, to enable the generation of new chemical compounds with desired properties. This methodology has demonstrated significant potential in exploring and generating specific molecules by optimizing for targeted characteristics. The PPO algorithm's superior performance in exploring the chemical space and generating compounds with diverse pharmacophore features, functional groups, and biological activities underscores its potential in drug discovery and chemical synthesis. By systematically comparing the outputs of PPO and REINFORCE, we highlighted the robustness and efficiency of PPO in optimizing molecular properties for targeted therapeutic applications.