Evaluating the Effectiveness of Reference Solvent Solubility Calculations for Binary Mixtures Based on Pure Solvent Solubility: The Case of Phenolic Acids.

Predicting the solubility of active pharmaceutical ingredients (APIs) in binary solvent mixtures is a major challenge in formulation science, as physics-based models often fail to capture complex, non-additive mixing effects. This study presents a robust machine learning (ML) framework to overcome this limitation, enabling accurate predictions from pure solvent data alone and molecular descriptors derived from COSMO-RS (computed with COSMOtherm). Firstly, our experimental knowledge of binary solvent mixtures solubility was expanded through newly measured data of caffeic and ferulic acids in aqueous mixtures of DMF, DMSO, and 4-formylmorpholine (4-FM). These new data, combined with values in the literature, formed a comprehensive dataset of 1636 points for ten phenolic and benzoic acids. To build a predictive model, a systematic methodology was developed, with the acronym of DOO-IT (Dual-Objective Optimization with ITerative features pruning), which automates descriptor selection and hyperparameter optimization to yield a maximally parsimonious and generalizable model. An exhaustive, multi-run stability analysis identified a final 10-descriptor nuSVR model as the optimal solution. This model demonstrated outstanding predictive power, achieving an R