Predictive modeling of physicochemical properties of antihypertensive drugs using degree-based topological indices and machine learning algorithm

Abstract

Quantitative prediction of physicochemical properties through molecular graph theory has become an important focus in cheminformatics. This study introduces a set of degree-based topological indices—ABC, ABS, MMR, SDD, SI, SO, SO${}_3$, and SO${}_4$—to model 23 antihypertensive drugs. A QSPR framework is developed using both classical linear regression and ensemble-based machine learning algorithms (Random Forest and XGBoost). Model performance is evaluated using standard error metrics (MAE, MSE, RMSE, $R^2$), and feature importance is analyzed through Gini, permutation, and Shapley Additive exPlanations (SHAP). The proposed indices show strong correlations with boiling point, melting point, critical volume, LogP, molar refractivity, and CLogP. Among the tested models, XGBoost performs best, achieving $R^2>0.99$ across all properties. Beyond predictive accuracy, the findings show that degree-based indices capture structural features of drug molecules while offering interpretable insights into lipophilicity, stability, and thermodynamic behavior. These results demonstrate the potential of graph-theoretical descriptors as cost-effective alternatives to experimental assays, thereby accelerating rational drug design and screening workflows. Overall, this study establishes a generalizable modeling framework that bridges mathematical chemistry and pharmaceutical applications, providing valuable directions for high-throughput drug discovery.