Machine Learning Modeling for Solubility Prediction in Pharmaceutical-Ionic Liquid Systems with Deep Feature Search

Drug solubility is a critical determinant of bioavailability and therapeutic efficacy in pharmaceutical development, with poorly soluble compounds presenting significant formulation challenges. Ionic liquids (ILs) have emerged as promising solubilizing agents due to their tunable physicochemical properties and ability to enhance drug dissolution. While machine learning (ML) has shown promise in predicting various IL properties, its application to pharmaceutical solubility prediction in IL media remains underexplored, largely due to the fragmented experimental data reported, the inherent complexity of binary (or multi-component) pharmaceutical−IL systems, and the susceptibility of conventional feature selection methods to overfitting on small datasets. This study addresses this gap by developing a data-driven quantitative structure−property relationship (QSPR) model for solubility prediction in pharmaceutical−IL systems. We curated a comprehensive experimental dataset comprising 234 solubility measurements spanning diverse pharmaceuticals and ILs (including binary IL mixtures, where IL1 and IL2 denote the two constituent ionic liquids) across varying temperatures. From SMILES representations of molecular structures, we computed 155 molecular descriptors using RDKit, followed by rigorous preprocessing including variance thresholding and Yeo-Johnson transformation. Two complementary feature selection strategies, deep feature selection (DFS) and SHapley Additive exPlanations (SHAP), were employed to identify the most predictive descriptors. Four regression models, including random forest (RF), extreme gradient boosting (XGB), decision tree (DT), and multilayer perceptron (MLP), were evaluated on the selected features. Under the adopted validation framework, DFS-selected features produced more consistent out-of-sample performance than SHAP-selected features, with XGB achieving the highest test performance (R2 = 0.84). While DFS prioritized structural descriptors (e.g., bicyclic rings and aromatic systems), SHAP analysis revealed that temperature and IL composition ratios exert the strongest influence on final predictions. This work establishes a reliable ML framework for solubility prediction in pharmaceutical−IL systems, providing valuable guidance for the rational design of IL-based drug delivery formulations.