XGNN: A chemometric dual-tower model for predicting aqueous solubility

Quantitative prediction of aqueous solubility (logS) is of great importance; however, achieving high-accuracy modelling remains challenging due to the structural diversity of molecules and variations in the quality of experimental data. To improve the prediction accuracy of aqueous solubility, this study proposes an XG graph neural network model (XGNN) that integrates engineered features with molecular graph representations. Specifically, XGNN combines XGBoost with a directed message passing neural network (D-MPNN), enabling simultaneous capture of global physicochemical patterns and local structural information. To ensure fair comparison and reproducibility, five public datasets were integrated as the training set under a unified data-cleaning and evaluation protocol. In contrast, the Huuskonen dataset was used as an independent test set. After standardisation, deduplication, and consistency filtering, 20,030 training samples and 1282 test samples were obtained. Compared with XGBoost, D-MPNN, and several widely used baseline models in recent years (e.g., Random Forest, LightGBM, GCN, and AttentiveFP), XGNN achieved the best performance on the independent test set, with an R² of 0.940, an RMSE of 0.501, and an MAE of 0.366, demonstrating a clear overall advantage over the competing methods. Feature ablation and applicability domain analyses further confirmed the complementarity of the two types of representations as well as the reliability and robustness of the proposed model. Collectively, XGNN provides a practical approach for logS prediction and related chemometric property modelling.