PDL1-inhi.predictor: a two-tiered machine learning framework for predicting and prioritizing PDL1 inhibitors

Small-molecule inhibition of PD-1/PDL1 via PDL1 dimerization presents a promising yet challenging therapeutic approach for cancer treatment. To accelerate the discovery of effective PDL1 inhibitors, we developed machine learning (ML)-based models optimized for PDL1-specific predictions. By systematically constructing and evaluating multiple classification and regression models (30 classifiers and 30 regressors), we identified Stacking_standard as the best-performing classifier and SVM_standard as the optimal regressor. Our approach combines classifier-based activity prediction with regressor-based affinity estimation, enabling a two-tiered screening strategy that enhances accuracy. Screening a 12-million-compound library resulted in 6704 potential PDL1 inhibitors, which were further refined through molecular docking. Seven top candidates exhibited strong binding affinity (binding energy ≤ -11.5 kcal/mol), and HTRF-based assays confirmed potent PDL1 inhibition for two compounds, with IC₅₀ values of 17.49 nM and 63.17 nM. Subsequent 300-ns molecular dynamics simulations revealed stable binding conformations for the inhibitors. Beyond these validated hits, our analysis uncovered 15 critical substructural motifs that provide valuable insights for rational inhibitor design. Additionally, SwissADME profiling indicated favorable drug-likeness, particularly for Hit 1, supporting their potential as lead compounds. The entire ML screening pipeline has been made accessible through the PDL1-inhi.predictor web server (http://www.pdl1predictor.com/chemicals), offering researchers a powerful tool for PDL1 inhibitor discovery.