AI-Guided Drug Design: Reinforcement Learning and Docking for PPARγ Ligands

I have developed a machine learning model for designing drug-like compounds targeting the nuclear receptor peroxisome proliferator-activated receptor gamma (PPARγ), a key drug target for treating type II diabetes. The model employs reinforcement learning to generate novel compounds by assembling molecular fragments from curated databases organized by chemical reactivity. To ensure synthesizability, the model selects fragments that follow a predefined synthetic pathway. The generated compounds are evaluated based on their binding affinity to PPARγ, estimated by docking free energy (ΔG in kcal/mol) using AutoDock Vina. The docking score serves as the reward function, prioritizing compounds with strong binding affinity. To encourage exploration, a penalty is applied to previously evaluated compounds, preventing the model from repeatedly selecting the same structures. Model performance is monitored by tracking the average docking score across 100 docking batches, ensuring continuous optimization toward high-affinity binders. This approach aims to enhance target specificity and potentially reduce off-target effects and side effects.