Beyond Prediction: Mechanistic Elucidation of Molecular Drivers for Aqueous Solubility Using Interpretable Machine Learning

Aqueous solubility is a critical property in drug discovery, yet its accurate prediction has remained elusive due to the complex balance between crystal lattice energy, solvation enthalpy, and entropic contributions. Conventional quantitative structure–property relationship (QSPR) models often provide limited accuracy and fail to yield mechanistic insights into the molecular origins of solubility. Here, we introduce an interpretable machine learning framework that integrates predictive modeling with explicit physicochemical reasoning, aiming not only to forecast solubility but also to uncover its molecular determinants. A high-quality dataset of 9,982 compounds was curated from multiple public sources, filtered for standardized conditions (25 ± 5 °C, pH 7 ± 1) and outlier removal. For each compound, 26 molecular descriptors were initially computed, spanning electronic, topological, and thermodynamic features. After redundancy reduction, 20 chemically meaningful descriptors were retained. Benchmarking of 18 algorithms identified tree-based ensemble methods as optimal, with LightGBM selected for its strong balance of accuracy and computational efficiency. The optimized model achieved R² = 0.80 and RMSE = 1 logS unit on a held-out test set, outperforming previously reported approaches. Interpretability analyses provided mechanistic clarity. SHAP (Shapley Additive exPlanations) identified the octanol–water partition coefficient (LogP) as the most influential descriptor, contributing nearly 40% of predictive variance. This directly reflects the hydrophobic effect, where transferring non-polar molecules into water incurs an entropic penalty from ordering solvent molecules. Topological polar surface area (TPSA) was the second most important factor, highlighting the enthalpic benefit of polar interactions and hydrogen bonding. Partial dependence plots revealed nonlinear threshold effects: solubility declines steeply when LogP exceeds ~3.5, aligning with Lipinski’s “rule of five,” while TPSA shows diminishing returns beyond ~100 Ų, where intramolecular hydrogen bonding may shield polar groups. Additional features such as molecular weight, number of rotatable bonds, and molecular complexity (BertzCT index) were identified as negative correlates of solubility. These findings are consistent with entropic losses upon dissolution and the unfavorable packing of bulky, branched structures in aqueous environments. Local SHAP explanations further elucidated molecule-specific predictions: for caffeine, solubility arises from balanced polarity and moderate hydrophobicity, whereas for fenbendazole, high LogP and large molecular size dominate its insoluble profile despite moderate TPSA. This study establishes a robust, interpretable framework that bridges machine learning with physical chemistry. By quantifying the roles of hydrophobicity, polarity, molecular size, flexibility, and complexity, our model not only predicts solubility with high accuracy but also recapitulates established thermodynamic principles. These insights provide actionable guidance for medicinal chemists, such as prioritizing LogP values below 3.5 and optimizing TPSA around 100 Ų for improved solubility. More broadly, this work illustrates how interpretable ML can move beyond correlation to causation, serving as a hypothesis-generating tool in chemical research.