Characterizing Molecular Property Variations Through Multivariate Machine Learning and Descriptor Analysis

Abstract

 The aqueous solubility of a molecule is one of the most important properties of a molecule that are determined by complex interactions among molecular and physicochemical characteristics. This study devised an interpretable multivariate machine-learning framework to characterize and predict solubility variation. AqSolDB enabled analysis of 9,881 compounds, following a quality screening and validation of structures. 208 two-dimensional descriptors were computed using RDKit, followed by variance, correlation and multicollinearity filtering to yield 58 descriptors. Application of Principal component analysis, correlation analysis, six regression models, locked-test validation and SHAP interpretation. XGBoost achieved the best performance, with cross-validated R² of 0.793 ± 0.011 and locked-test R² of 0.786, RMSE of 0.807, and MAE of 0.589. The variance of the descriptors was explained by the first five principal components (72.8%). MolLogP, MW, TPSA, aromaticity and complexity were the main drivers of solubility. The framework combines the ability of accurate prediction with chemically meaningful interpretation of the variation of molecular properties.