Abstract
Learning aqueous solubility remains a key challenge in drug development for improving oral bioavailability. Traditional data-driven solubility estimations using standard supervised models, however, can often suppress the information embedded in a molecule’s chemical properties and the intricate connectivity of its atoms. In addition to the predictive analyses for solubility, it is also crucial to investigate the underlying scientific factors that influence the aqueous solubility of potential drug candidates. This paper attempts to learn such properties and further proposes a novel nonparametric model with an additive-type characterization that provides insights into the differential contribution of the chemical and graphical descriptors, that are important in learning a molecule’s solubility in an aqueous solvent. The dataset used in this work includes over 9000 experimental structural, chemical, and electronic properties. Using graph-theoretic attributes derived from molecular graphs, we characterize the structural traits of the molecules. Our analysis reveals several intriguing properties of solubility and its physicochemical relationship with both structural and chemical features of the molecules, specifically based on their weights and shapes.
