Attention-Based Graph Neural Network for Molecular Solubility Prediction.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in pharmacology

Background:

• Solubility is a critical physicochemical property for active pharmaceutical ingredients (APIs) in drug discovery.
• Accurate aqueous solubility (AS) prediction is essential for efficient API characterization and reduces experimental costs.
• Machine learning and deep learning techniques show promise for in silico solubility prediction.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting the aqueous solubility of diverse molecules.
• To utilize the largest available solubility dataset for training and testing predictive models.
• To identify the most effective deep learning architecture for solubility prediction.

Main Methods:

• Utilized Simplified Molecular Input Line Entry System (SMILES) strings for molecular representation.
• Developed and compared deep learning models including simple graph convolution, graph isomorphism network, graph attention network, and AttentiveFP.
• Trained and tested models on a dataset of 9943 compounds.

Main Results:

• The AttentiveFP-based network model demonstrated superior performance in predicting aqueous solubility.
• The selected model achieved a Pearson correlation coefficient (R²) of 0.52 and a root-mean-square error of 0.61 on 62 anticancer compounds.
• The study highlights the potential of graph-based deep learning for solubility prediction.

Conclusions:

• Deep learning models, particularly AttentiveFP, can accurately predict molecular aqueous solubility.
• Improved graph algorithms or the inclusion of additional molecular properties can further enhance solubility prediction accuracy.
• This approach offers a valuable tool for accelerating drug discovery by reducing experimental efforts.