Year: 2026
Type: preprint
Abstract: Solvent response governs hydration free energy, yet high-fidelity calculations resolve it afresh for every new molecule. We ask whether information already encoded in disparate solvent datasets can instead be learned once and reused. SolvAI trains molecular surrogates on six calculated and empirical sources, predicts 15 complementary response descriptors from a SMILES string, and combines them with structural features in a separate predictor fitted to experimental hydration free energies. Inference requires no new solvent simulation. In a matched ExtraTrees comparison, the descriptors reduce mean absolute error from 0.303 to 0.202 kcal mol−1 on 85 ARROW solutes and from 1.532 to 1.153 kcal mol−1 on an endpoint-disjoint external cohort of 220 molecules. Shuffling descriptors among molecules removes the gain; it persists on the 97-molecule subset also absent from supervised response sources. The same descriptors often help other endpoint families, although the best complete predictor and the most useful source mixture change with the evaluation chemistry. Existing solvent calculations can therefore provide reusable supervision for molecular prediction, while source choice and transfer beyond small neutral solutes remain important tests.
Source: Research Square
Authors: Gal Oren, Boris Fain, Michael Levitt
Institution: Stanford University
Language: English
Subfield: Drug Discovery
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225758324