Graph theoretic and machine learning approaches in molecular property prediction of bladder cancer therapeutics

Abstract

This work introduces a hybrid computational approach in which degree-based topological descriptors are harnessed with the aid of advanced regression models and artificial neural networks (ANNs) to predict the crucial physicochemical properties of 17 drugs for the treatment of bladder cancer. Each molecule is assigned a molecular graph, from which a series of topological descriptors such as Zagreb indices, Randic index, Atom Bond Connectivity (ABC), and Symmetric Division Degree (SSD)are computed. These indices are used as input features by various regression models along with linear, cubic, and feedforward ANNs. The performance of the models is analyzed using metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination $$(R^2)$$ . ANNs showed the best predictive performance with the $$R^2$$ value achieving 0.99. Moreover, SHAP (SHapley Additive exPlanations) analysis was used to explain the contribution of each descriptor toward the models’ predictions. The findings validate the promise of the combination of graph-theoretic descriptors with the tools of machine learning to achieve solid and interpretable models of molecular property prediction, which hold the potential for drug discovery and optimization in oncologic applications.