DISSOLVR: An Interpretable and Fast Framework for Aqueous and Organic Solubility Prediction

High-fidelity solubility prediction is fundamental to pharmaceutical development and environmental partitioning, where accurate modeling must couple molecular structure with thermodynamic behavior across diverse chemical environments. However, recent advancements have been dominated by deep learning architectures that often sacrifice physical interpretability for predictive power. We challenge this trend by showing that state-of-the-art performance does not require such non-transparent architectures. To address this, we introduce DISSOLVR, a transparent framework for molecular solubility prediction. In addition, we perform a comprehensive literature review and a benchmarking study against various methods. We show that Dissolvr approaches the aleatoric limit of experimental uncertainty and achieves OOD generalization through structural invariance derived by mapping molecules to physically-grounded descriptors. Then, we present an LLM-assisted post-hoc explanation pipeline that bridges the gap between symbolic model artifacts and chemically grounded narratives. Finally, a comparative benchmark of a survey involving 22 expert chemists reveals that expert evaluators provide deep insights.

Knowing how well a drug dissolves in water or other solvents is one of the most important problems in pharmaceutical development. Nearly 40% of marketed drugs suffer from poor solubility, yet measuring it experimentally is slow and error-prone: two labs measuring the same compound can disagree by almost a full order of magnitude. Computational models help, but the field has drifted toward deep architectures that trade interpretability for marginal gains in accuracy—a bad idea when chemists need to understand why a molecule behaves the way it does, not just get a number. We introduce DISSOLVR, a framework that challenges this trend. Instead of learning hidden representations, DISSOLVR maps molecules onto 176 physically grounded descriptors—counts of functional groups, topological motifs, Abraham solvation parameters, and thermodynamic estimates—and feeds them into a gradient-boosted decision tree (CatBoost). In multi-solvent settings, a compact cross-attention layer produces an explicit $24 \times 24$ matrix that quantifies how each solute property couples to each solvent property, directly reflecting the "like dissolves like" principle. Temperature constraints are enforced as hard monotonic rules derived from the van't Hoff relation, guaranteeing thermodynamic consistency without soft penalties or extra compute. On every benchmark tested—ESOL, AqSolDB, BigSolDB, and the Leeds solute-holdout set—DISSOLVR matches or outperforms deep learning baselines while using $81\times$ fewer parameters and running in under a second on a laptop CPU. Crucially, the model's symbolic artifacts (SHAP values, decision paths, interaction matrices) are passed through a six-stage LLM pipeline that produces plain chemical narratives grounded strictly in model evidence. In a blind survey of 22 expert chemists, these explanations received a mean quality rating of 3.75/5 and produced a statistically significant increase in expert agreement with the model's predictions. This demonstrates that interpretability and accuracy are not in conflict, and that approaching the aleatoric limit of experimental noise does not require sacrificing scientific transparency.