Prediction of Aqueous and Non-Aqueous Solubility Using Machine Learning

Solubility, the capacity of a solute to dissolve in a solvent, forming a solution, is a crucial design parameter across various materials and life science applications. Due to the high cost of experimental measurements, we have developed quantitative structure-property relationship (QSPR) models to rapidly and accurately predict aqueous solubility in water and non-aqueous solubility in organic solvents. For this purpose, we gathered 14,485 room temperature aqueous solubility data points and 45,313 temperature-dependent non-aqueous solubility data points from literature and open-source databases. Additionally, we incorporated advanced cheminformatics-based, graph-based, and physics-based descriptors computed through classical molecular dynamics to optimize machine learning performance. These models can significantly streamline molecular discovery by providing rapid, accurate solubility predictions, reducing the need for costly experiments, and accelerating the identification and optimization of promising candidates.