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

Area of Science:

• Cheminformatics
• Computational Chemistry
• Machine Learning

Background:

• Quantitative prediction of aqueous solubility (logS) is vital for drug discovery and chemical development.
• High-accuracy modeling remains challenging due to molecular structural diversity and experimental data quality variations.

Purpose of the Study:

• To develop an improved model for accurate aqueous solubility prediction.
• To integrate engineered features with molecular graph representations for enhanced predictive power.

Main Methods:

• Proposed an XG graph neural network model (XGNN) combining XGBoost and a directed message passing neural network (D-MPNN).
• Integrated five public datasets for training under a unified protocol, using the Huuskonen dataset as an independent test set.
• Standardized, deduplicated, and filtered data, resulting in 20,030 training and 1,282 test samples.

Main Results:

• XGNN achieved superior performance on the independent test set compared to XGBoost, D-MPNN, Random Forest, LightGBM, GCN, and AttentiveFP.
• Achieved R² of 0.940, RMSE of 0.501, and MAE of 0.366, demonstrating a clear advantage.
• Feature ablation and applicability domain analyses confirmed model reliability and robustness.

Conclusions:

• XGNN offers a practical and robust approach for quantitative prediction of aqueous solubility (logS).
• The model's ability to capture both global physicochemical patterns and local structural information contributes to its high accuracy.
• This work advances chemometric property modeling through improved predictive accuracy.