An Ensemble Model for Molecular Solubility Prediction Across Aqueous and Organic Solvents

Solubility underpins decisions in drug discovery, separations, crystallization, formulation, and materials synthesis, yet accurate prediction across diverse solute–solvent pairs and temperatures remains challenging. We present a supervised learning framework that unifies two high-quality resources—AqSolDB for aqueous solubility and BigSolDB v2.0 for solubility in 213 solvents—into a standardized schema with RDKit-derived 2D descriptors for both solute and solvent, plus temperature. The modeling stack couples gradient-boosted decision trees (XGBoost) and a lightweight one-dimensional convolutional neural network trained on tabular descriptors. A validation-driven weight optimization determines ensemble contributions. The approach is data-efficient, reproducible, and deployable: it requires only SMILES of solute and solvent and temperature to estimate log10 S (in mol/L). On held-out validation and test splits, the ensemble achieves R² = 0.945 and RMSE = 0.341 log units, consistently outperforming single learners. We provide a complete, reproducible workflow with explicit data lineage and source code.