Molecular solubility directly impacts key development aspects such as reaction feasibility, formulation performance, separation efficiency, and solvent selection. Experimental measurement across solutes, solvents, and temperatures is costly and sparsely sampled. Existing computational models often rely on fixed‑solvent assumptions, deterministic formulations, or simplified solute‑solvent interaction representations, limiting their ability to capture complex molecular interactions, continuous temperature effects, and experimental uncertainty.
EnSol introduces an environment‑aware probabilistic framework. It first represents the solute and solvent as separate molecular graphs and learns their embeddings with distinct graph neural networks. A cross‑attention mechanism then fuses the two embeddings to capture fine‑grained solute‑solvent interactions. Temperature is incorporated directly into the solvent environment via feature‑wise modulation, enabling continuous temperature influence on the embeddings. Finally, a mixture density network outputs the full solubility distribution, jointly modeling temperature‑dependent behavior and experimental measurement uncertainty.
On the independent SolProp and Leeds benchmark datasets, EnSol achieved Spearman correlations of $0.876$ and $0.601$, respectively, outperforming state‑of‑the‑art solubility prediction models. Experimental validation across chemically diverse solute‑solvent pairs confirmed strong predictive performance and reliable solvent ranking, with a Spearman correlation of $0.715$. These results demonstrate that EnSol can provide trustworthy solubility predictions and solvent selection across diverse chemical systems while naturally quantifying predictive uncertainty.
Review: By integrating graph neural networks, cross‑attention, and a mixture density network, EnSol offers a nuanced treatment of environmental factors, delivering higher accuracy and reliability for molecular solubility prediction.