Machine Learning Regression Models for Predicting Anti-Cancer Drug Properties: Insights from Topological Indices in QSPR Analysis

Source abstract

This study advances the prediction of anti-cancer drug properties by integrating machine learning regression techniques with topological indices derived from hydrogen-depleted molecular graphs. Focusing on distance-based metrics, we investigate how atomic spatial arrangements influence molecular characteristics, enhancing Quantitative Structure-Property Relationship (QSPR) frameworks. Our analysis reveals significant correlations between various topological indices and key molecular properties, including polarizability, molar refractivity, and boiling point. The regression models, particularly Linear and Ridge Regression, achieved high predictive performance, with R 2 values exceeding 0.80, low Root Mean Square Error (RMSE), significant F-test, and p-values below 0.05 for highly correlated properties. The importance of careful index selection is underscored, as models using indices with strong correlations show superior predictive accuracy. This approach offers a more robust framework for predicting physicochemical properties, enhancing the efficiency of screening processes and optimizing lead compounds in oncological research. By bridging theoretical modeling and practical drug discovery, this work has the potential to accelerate the development of more effective and targeted anti-cancer therapies.