In Silico Identification of Natural-Product Aromatase (CYP19A1) Inhibitors via Integrated Molecular Docking and QSAR...

Aromatase (CYP19A1) is the rate-limiting enzyme in estrogen biosynthesis and the principal therapeutic target for estrogen receptor-positive (ER+) breast cancer. Despite the clinical success of third-generation aromatase inhibitors (AIs) such as letrozole and anastrozole, acquired resistance and systemic adverse effects, including musculoskeletal pain, bone density loss, and cardiovascular complications, which necessitate the identification of novel scaffolds. This study employs an integrated computational pipeline combining ligand-based quantitative structure-activity relationship (QSAR) modelling and structure-based molecular docking to screen the AfroDB and EANPDB natural product libraries for putative aromatase inhibitors. A curated dataset of 3,068 unique compounds derived from ChEMBL (v33) was used to train Random Forest models, with the binary classification model achieving an area under the receiver operating characteristic curve (AUC) of 0.95 on the independent holdout set, an accuracy of 0.90, and a Matthews correlation coefficient (MCC) of 0.76. The regression model for pIC₅₀ prediction attained an R² of 0.65 and an RMSE of 0.81 pIC₅₀ units on the test set. The trained classifiers were applied to screen 1,871 AfroDB compounds; 361 were predicted active, of which 71 resided within the defined Applicability Domain. Molecular docking into the human aromatase crystal structure (PDB ID: 3S79) revealed that the top candidate, a prenylated dihydroflavone (5,7-dihydroxy-4'-methoxy-3'-(3-hydroxy-3-methylbut-1-enyl)-5'-(3-methylbut-2-enyl)flavanone), achieved a binding affinity of -9.598 kcal/mol, outperforming both letrozole (-7.11 kcal/mol) and anastrozole (-7.617 kcal/mol) under identical docking conditions. The docking protocol was validated by redocking of the co-crystallised ligand (RMSD = 1.35 Å). Comprehensive interaction analysis identified contacts with key active-site residues Arg115, Phe221, Thr310, Val370, Met374, and Phe430. ADMET profiling indicated that 65% of prioritised hits satisfy Lipinski's Rule of Five, while 95% meet Veber bioavailability criteria. This study provides a reproducible computational framework for next-generation AI discovery, adhering to the TRIPOD+AI reporting guidelines for machine learning in biomedical research.