Benchmarking machine learning algorithms for solubility prediction of pharmaceuticals in ethanol

In this study, the experimental mole fraction solubility data for drugs in ethanol at various temperatures are collected from the literature and refined based on low solubility values, inconsistencies in similar systems, prediction errors, and enthalpic analysis. A curated dataset comprising 784 experimental mole-fraction solubility measurements for 77 pharmaceutical compounds in ethanol over a range of temperatures is assembled to develop predictive mathematical models.

Herein, the van’t Hoff equation is integrated with Abraham-Hansen solubility parameters and RDkit parameters to establish conventional linear correlation models. Moreover, six regression algorithms including multiple linear regression, support vector regression , random forest , LightGBM, CatBoost, and an artificial neural network are evaluated using three descriptor representations: Abraham-Hansen solvation descriptors, RDKit molecular descriptors, and a combined descriptor set obtained through feature selection.