FraGAT

FraGAT predicts molecular properties by representing molecules as fragment-oriented multi-scale graphs and applying a multi-scale graph attention network to capture hierarchical and functional-group-level structural information.

Key Features:

• Fragment-Oriented Representation: Defines molecule graph fragments that can include functional groups and serve as representation units linked to molecular properties.
• Multi-Scale Graph Attention Network: Employs a multi-scale attention mechanism that processes structural information at multiple levels of molecular hierarchy.
• Interpretability and Predictive Performance: Demonstrates state-of-the-art predictive performance on several benchmarks and improved interpretability when fragment-based representations containing functional groups are used.
• Case Studies and Validation: Validated by case studies that align model predictions with expected chemical behaviors and support interpretation of fragment contributions.

Scientific Applications:

• Chemistry and Pharmacy: Predicts molecular physio-chemical properties relevant to chemical research and pharmaceutical development.
• Binding Affinity and Property Prediction: Estimates binding affinities and other molecular properties that influence compound behavior.
• Drug Discovery and Development: Supports assessment of compound efficacy and safety to inform candidate selection in drug discovery workflows.

Methodology:

Defines molecule graph fragments that encapsulate functional groups; develops a multi-scale graph attention network that processes these fragments at various scales; and conducts experiments to validate model performance and interpretability.