New AI framework separates chemical & structural factors in drug solubility prediction

A new additive MLP-GNN framework from researchers Bhattacharya and Roy (arXiv, July 2026) tackles a core problem in drug discovery: understanding whether a solubility prediction comes from global chemical properties or molecular structure. Most existing models merge these two information sources into a single black-box representation, making interpretation difficult. This framework keeps them separate throughout training—a multilayer perceptron (MLP) encodes physicochemical descriptors (the 'chemical branch') while a graph neural network (GNN) captures molecular graph topology (the 'structural branch'). The outputs are combined only at the final prediction stage through an additive model, with an optional multiplicative interaction.

The approach dramatically improves explainability. By pretraining on the larger AqSolDB dataset and fine-tuning on the smaller BigSolDB2 dataset, the model reduces run-to-run variation and boosts accuracy. The researchers interpret the model using best linear projections of branch outputs, molecule-level embedding summaries across solubility classes, and atom-level GNNExplainer masks aggregated over functional groups. Results show the chemical branch aligns with familiar physicochemical descriptors (like logP and hydrogen bond counts), while the structural branch captures graph-topological and functional-group patterns linked to solubility. This transparency could help medicinal chemists prioritize compounds for synthesis based on which factors dominate poor solubility.