From Descriptor Learning to Binding Stability: An Explainable Machine Learning Pipeline for EGFR Double-Mutant Inhibitor Discovery

Abstract

Drug resistance arising during cancer development and progression remains a major challenge in the treatment of epidermal growth factor receptor (EGFR)-driven tumors, particularly those harboring the clinically relevant T790M/L858R double mutation. In this study, we developed an integrated computational workflow combining explainable machine learning, virtual screening, molecular dynamics simulations, and binding free-energy calculations to identify novel inhibitors of this drug-resistant EGFR variant. An XGBoost regression model was trained using scaffold-aware cross-validation, Bayesian hyperparameter optimization, and sequential feature selection, resulting in a compact model based on 16 molecular descriptors. The model demonstrated robust predictive performance on external validation data, while SHAP analysis identified descriptors related to the local electronic environment, fragment distribution, and molecular topology as the primary contributors to activity prediction. The optimized model was subsequently applied to screen compounds from the Enamine REAL database. Top-ranked candidates were evaluated using explicit-solvent molecular dynamics simulations and MM/GBSA binding free-energy calculations. Several compounds formed stable protein–ligand complexes and maintained key interactions with residues known to be important for EGFR inhibition, including Lys745, Met790, and Leu718. These results demonstrate that the proposed workflow can efficiently prioritize computational candidates of drug-resistant EGFR mutants and may support the development of new therapeutic strategies for overcoming resistance in EGFR-driven cancers.