Enhancing predictive models for solubility in multicomponent solvent systems using semi-supervised graph neural networks

Abstract

Solubility plays a critical role in guiding molecular design, reaction optimization, and product formulation across diverse chemical applications. Despite its importance, current approaches for measuring solubility face significant challenges, including time- and resource-intensive experiments and limited applicability to novel compounds. Computational prediction strategies, ranging from theoretical models to machine learning (ML) based methods, offer promising pathways to address these challenges. However, such methodologies need further improvement to achieve accurate predictions of solubilities in multicomponent solvent systems, as expanding the modeling approaches to multicomponent mixtures enables broader practical applications in chemistry. This study focuses on modeling solubility in multicomponent solvent systems, where data scarcity and model generalizability remain key hurdles. We curated a comprehensive experimental solubility dataset (MixSolDB) and examined two graph neural network (GNN) architectures – concatenation and subgraph – for improved predictive performance. By further integrating computationally derived COSMO-RS data via a teacher–student semi-supervised distillation (SSD) framework, we significantly expanded the chemical space and corrected previously high error margins. These results illustrate the feasibility of unifying experimental and computational data in a robust, flexible GNN-SSD pipeline, enabling greater coverage, improved accuracy, and enhanced applicability of solubility models for complex multicomponent solvent systems.