DNA Polymerase Inhibitor Discovery Using Machine Learning-Enhanced QSAR Modeling

Cisplatin resistance is a major obstacle in cancer therapy, often driven by translesion DNA synthesis (TLS) mechanisms utilizing specialized polymerases like human DNA polymerase η (hpol η). While small-molecule inhibitors such as PNR-7-02 have shown potential in disrupting hpol η activity, existing compounds lack the necessary potency and specificity to fully address chemoresistance. Traditional drug discovery methods are limited by the vast chemical space, highlighting the need for advanced computational strategies like machine learning (ML)-enhanced Quantitative Structure-Activity Relationship (QSAR) modeling. In this study, we used a curated library of 85 indole thio-barbituric acid (ITBA) analogs with validated hpol η inhibition data, excluding outliers to ensure integrity. Molecular descriptors (1D–4D) were computed, resulting in 220 features. Seventeen ML algorithms, including Random Forests, XGBoost, and Neural Networks, were trained on an 80% training set and evaluated across 14 performance metrics. Hyperparameter optimization and 5-fold cross-validation ensured robustness. Ensemble methods outperformed others, with Random Forest achieving near-perfect accuracy (training MSE = 0.0002, R² = 0.9999; testing MSE = 0.0003, R² = 0.9998). SHAP analysis identified electronic properties, lipophilicity, and topological atomic distances as top predictors of inhibition. Linear models showed higher errors, underscoring the non-linear nature of the relationship between molecular descriptors and hpol η inhibition by ITBA analogs. Integrating ML with QSAR modeling offers a robust framework for optimizing hpol η inhibition, combining high predictive accuracy with biochemical interpretability. This approach accelerates the discovery of potent, selective inhibitors, providing a promising strategy to overcome cisplatin resistance and enhance precision oncology.