Discovery of Novel PD‐L1 Inhibitors Using Machine Learning and Molecular Simulations

Abstract Programmed death ligand 1 (PD‐L1) plays a key role in tumor immune escape and is an important target for current tumor immunotherapy. In order to break through the limitations of long cycle time and high cost of traditional screening methods, this study constructed a machine learning regression model based on 2044 known PD‐L1 inhibitors to predict their pIC 50 values. Among them, the classification boosting method (CatBoost) performed optimally, achieving an R 2 of 0.93 on the training set and 0.83 on the test set. The model was used to screen 3,883,857 class lead molecules in the ZINC database to obtain four candidate compounds with the highest activity and to analyze their binding modes to PD‐L1 by molecular docking. Further, ADMET evaluation was performed to screen out 2 candidate molecules with good pharmacokinetic properties and low toxicity. Molecular dynamics simulations showed that both formed complexes with PD‐L1 with high conformational stability, supporting their feasibility as potential inhibitors. This study established an efficient virtual screening process and identified novel PD‐L1 small molecule inhibitors with development potential, providing new ideas for targeted immunotherapy.