Machine Learning–Based QSAR Screening of Natural Product and ZINC Libraries Uncovers Stable BACE-1 Inhibitors Validated by Gaussian Accelerated Molecular Dynamics (GaMD) Simulations

Abstract

Beta-site amyloid precursor protein cleaving enzyme 1 (BACE-1) plays a central role in amyloid-β production and remains an important therapeutic target for Alzheimer’s disease. This study developed an integrated quantitative structure–activity relationship and molecular dynamics (QSAR–MD) framework to identify promising BACE-1 inhibitor candidates. From 15,642 initially retrieved bioactivity records, 5,245 remained after preprocessing, and 5,235 compounds with complete fingerprint data constituted the final QSAR dataset. Random Forest models based on twelve molecular fingerprint classes were benchmarked against AdaBoost, Gradient Boosting, and XGBoost using identical training and external-validation sets. CDK and CDK Extended fingerprints produced the highest validation performance (Q²CV = 0.78 ± 0.03; Q²EXT = 0.79 ± 0.01), while PubChem, MACCS, Klekota–Roth, and Substructure Count fingerprints yielded Q²EXT values of approximately 0.76–0.78. Model reliability was assessed using RMSE, MAE, Y-randomization, applicability-domain analysis, and SHAP interpretation. The selected CDK Extended RF model was applied to the ZINC and NPASS libraries, resulting in five prioritized candidates. Molecular docking indicated favorable interactions with key BACE-1 residues, including the catalytic dyad Asp32 and Asp228 and active-site residues Thr72 and Thr231. Analysis of 200 ns GaMD trajectories using RMSD, RMSF, radius of gyration, SASA, and PCA supported the overall structural stability of the complexes. MM/GBSA analysis identified NPC475168 as the most favorable candidate, with an estimated binding energy of − 64.84 kcal/mol. NPC313966, NPC252056, and NPC471858 also demonstrated favorable dynamic and energetic profiles relative to CNP520. Overall, NPC475168 emerged as the leading computational candidate, although experimental validation, selectivity profiling, and ADMET assessment remain necessary.