Machine learning-guided drug repurposing for EGFR inhibition using scaffold-split validation, docking, and molecular dynamics

Aberrant epidermal growth factor receptor (EGFR) signaling drives multiple cancers, but the clinical effectiveness of EGFR inhibitors is limited by relapse, toxicity, and mutation-associated resistance. This study applied an integrated computational drug-repurposing workflow combining machine learning-based potency prediction, structure-based docking, and molecular dynamics simulation to prioritize approved DrugBank compounds for mutant EGFR evaluation. EGFR bioactivity data from ChEMBL were curated, standardized, converted to pIC₅₀ values, and represented using SwissADME physicochemical descriptors and Morgan fingerprints. Among the evaluated regression models, ExtraTrees performed best and was selected for screening. Using ECFP plus descriptor features, ExtraTrees achieved R² = 0.71 ± 0.02 and RMSE = 0.74 ± 0.02 under random splits, and retained useful performance under Murcko scaffold splits with R² = 0.55 ± 0.01, RMSE = 0.90 ± 0.02, and Spearman ρ = 0.74 ± 0.02. The model was used to screen approved DrugBank compounds, prioritize 500 candidates, and guide docking against EGFR L858R/T790M/C797S (PDB: 6LUD). Docking identified Abemaciclib (-9.65 kcal/mol), Crizotinib (-8.29 kcal/mol), and Avapritinib (-8.10 kcal/mol) as favorable candidates. Abemaciclib scored slightly more favorably than Osimertinib (-9.44 kcal/mol) in this non-covalent docking setup. Subsequent 100 ns molecular dynamics simulations refined the ranking, with Crizotinib and Avapritinib showing more favorable dynamic profiles, while Abemaciclib showed greater ligand mobility. Oncology drugs were significantly enriched among the top 50 docked hits relative to the scored DrugBank background. These results support ML-guided docking and MD refinement as a practical strategy for prioritizing repurposing candidates for experimental EGFR validation.