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

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Drug Discovery

Background:

• Aqueous solubility is critical for drug pharmacokinetics, efficacy, and bioavailability.
• Experimental solubility measurement is time-consuming and expensive.
• Computational methods are needed for rapid and reliable solubility estimation.

Purpose of the Study:

• To develop a novel computational method for predicting aqueous drug solubility.
• To create an accurate and interpretable tool for drug discovery.

Main Methods:

• Utilized an edge-conditioned message passing mechanism (NNCov) with dynamic filter weight generation.
• Employed gated recurrent units (GRUs) for sequential dependencies in molecular graphs.
• Applied residual gated graph convolutional layers and Set2Set pooling for robust graph representations.

Main Results:

• The model consistently outperformed existing methods on five benchmark datasets (R² and RMSE).
• Residual plot analysis confirmed reliable error distribution.
• Graph explainability identified key molecular features influencing solubility.

Conclusions:

• The developed model is an accurate and interpretable computational tool for aqueous drug solubility prediction.
• This method has the potential to significantly advance drug discovery processes.