Enhancing aqueous solubility prediction with residual gated graph convolutions and sequential modeling

Aqueous solubility is one of the most crucial physicochemical characteristics affecting a compound's pharmacokinetics, efficacy, and bioavailability. Traditional experimental methods for measuring solubility are costly and time-consuming, motivating the development of computational approaches that can rapidly and reliably estimate drug solubility. In this work, we introduce a novel method for aqueous drug solubility prediction. Our model employs an edge-conditioned message passing mechanism (NNCov) that dynamically generates filter weights from edge features, combined with gated recurrent units (GRUs) to capture sequential dependencies within molecular graphs. Additionally, residual gated graph convolutional layers are applied to handle graphs with variable sizes and topologies, preserving original node features via residual connections to prevent over-smoothing. Finally, a Set2Set pooling layer produces robust, fixed-size graph-level representations for solubility prediction. We evaluated our model on five different benchmark datasets, where it consistently outperforms existing methods in terms of R² and RMSE. Residual plot analysis confirms the reliability of the error distribution, and graph explainability techniques highlight the molecular features most strongly influencing solubility. These results demonstrate the potential of our model as both an accurate and interpretable computational tool for advancing drug discovery.