Optimization of binding affinities in chemical space with generative pre-trained transformer and deep reinforcement learning

Background: The key challenge in drug discovery is to discover novel compounds with desirable properties. Among the properties, binding affinity a target one of prerequisites and usually evaluated by molecular docking or quantitative structure activity relationship (QSAR) models. Methods: In this study, we developed Simplified input line entry system Generative Pre-trained Transformer Reinforcement Learning (SGPT-RL), which uses transformer decoder as policy network reinforcement learning agent optimize target. SGPT-RL was on Moses distribution benchmark two goal-directed generation tasks, Dopamine Receptor D2 (DRD2) Angiotensin-Converting Enzyme 2 (ACE2) targets. Both QSAR model were implemented optimization goals tasks. The popular Reinvent method used baseline for comparison. Results: results showed that learned good property distributions generated molecules high validity novelty. On both able generate valid improved scores. achieved better than ACE2 task, where goal. Further analysis shows conserved scaffold patterns during exploration. Conclusions: superior performance task indicates it can be applied virtual screening process widely criteria. Besides, exploration assist chemists design lead candidates.