Multi-order graph attention network for water solubility prediction and interpretation.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Molecular property prediction

Background:

• Molecular water solubility is crucial in chemical and medical research.
• Machine learning (ML) models reduce computational costs for property prediction.
• Existing ML methods lack interpretability for predicted molecular properties.

Purpose of the Study:

• To propose a novel multi-order graph attention network (MoGAT) for enhanced water solubility prediction.
• To improve the interpretability of ML-based molecular property predictions.
• To achieve superior predictive performance compared to existing methods.

Main Methods:

• Utilized a multi-order graph attention network (MoGAT) architecture.
• Extracted graph embeddings from multiple node embedding layers.
• Integrated diverse neighboring order information using an attention mechanism.
• Generated atomic-specific importance scores for chemical interpretability.

Main Results:

• MoGAT demonstrated superior performance over state-of-the-art methods in water solubility prediction.
• The model provides chemically interpretable results by identifying influential atoms.
• Incorporating multi-order graph representations enhanced predictive accuracy.

Conclusions:

• MoGAT offers improved predictive performance and chemical interpretability for water solubility.
• The method aligns with established chemical knowledge, validating its predictions.
• This approach advances the application of ML in molecular property analysis.