Munawar Hraib

Background:

Prostate cancer (PCa) is a leading malignancy in men worldwide, with androgen receptor (AR) signaling central to disease progression and therapeutic resistance. Castration-resistant prostate cancer (CRPC) emerges in most patients receiving androgen deprivation therapy, underscoring the urgent need for novel AR-targeting compounds. Natural products represent an underexplored yet structurally diverse reservoir of bioactive molecules with potential AR-modulating activity.

Methods:

We developed AndroPredict, a machine-learning pipeline for prioritizing candidate AR modulators in natural-product chemical space. A curated dataset of 6,494 drug-like compounds with experimentally reported AR bioactivities was retrieved from ChEMBL. Molecular representations combined physicochemical descriptors with ECFP4 fingerprints. Random Forest and XGBoost classifiers and regressors were developed for binary activity classification and continuous pActivity prediction. The trained models were then applied to exploratory virtual screening of the NPASS natural-product database. Model interpretability was assessed using SHAP analysis.

Results:

Classification models achieved F1-scores of 0.8362 (Random Forest) and 0.8890 (XGBoost) on a held-out test set. Regression performance was moderate (R² ≈ 0.52–0.53; RMSE ≈ 0.81–0.82). NPASS screening produced a structurally diverse set of highest-ranked, near-threshold candidates spanning steroid-like, diterpenoid, coumarin-derived, and lipid-associated scaffolds. SHAP analysis indicated that predictions were driven predominantly by ECFP4-derived substructural features.

Conclusions:

AndroPredict provides a partially interpretable computational framework for narrowing large natural-product libraries to a tractable set of candidate AR modulators. The outputs should be interpreted as hypothesis-generating prioritization signals for downstream applicability-domain assessment, molecular docking, ADMET profiling, and experimental validation in prostate cancer research.