Multiscale Explainable Machine Learning Reveals Descriptor-Invariant Molecular Determinants of Small-Molecule PD-1/PD-L1 Inhibition.

Abstract

The PD-1/PD-L1 immune checkpoint pathway is a major target in cancer immunotherapy; however, small-molecule inhibitor development remains challenging due to the hydrophobic, structurally shallow PD-L1 interface. We developed an explainable artificial intelligence (XAI)-based QSAR framework to identify determinants governing PD-1/PD-L1 inhibition. A data set of 844 compounds, represented using MACCS, PubChem, and Mordred descriptors, was modeled using multiple machine learning algorithms with eXtreme Gradient Boosting (XGBoost) and Light Gradient Boosting Machine (LightGBM) models achieving the highest predictive performance. SHapley Additive exPlanations (SHAP) analysis and scaffold enrichment revealed convergence across descriptors, highlighting nitrogen-rich heteroaromatic systems, sulfur-containing motifs, and fused aromatic scaffolds as key determinants of activity. Active compounds occupied a distinct physicochemical space characterized by low molecular weight (MW) and topological polar surface area (TPSA) < 85 Å2. Docking identified conserved interactions within the PD-L1 dimer interface, and molecular dynamics confirmed their stability under-near physiological conditions. Molecular Mechanics/Poisson-Boltzmann Surface Area (MM/PBSA) calculations indicated that binding is predominantly driven by van der Waals interactions and hydrophobic stabilization, providing mechanistic insight and design guidance for next-generation PD-L1 inhibitors.