Solubility is governed by intermolecular interactions between solutes and solvents, and accurate prediction is essential in drug discovery and chemical process design. However, real-world applications often involve novel combinations that fall outside the training distribution, making out-of-distribution (OOD) generalization a critical challenge.
A new arXiv preprint introduces PRISM (Prior Relational Information for Self-supervised Modeling), a method designed to enhance solubility prediction under OOD conditions. By leveraging prior relational information within a self-supervised framework, PRISM aims to better capture the physical interactions that determine solubility, rather than relying solely on memorized patterns from the training set.
The authors position PRISM as a step toward more reliable predictive models for solubility, with direct relevance to pharmaceutical and chemical engineering workflows. The abstract does not provide quantitative results, so the reported gains are not detailed in the source.