HiGNN

HiGNN enhances molecular property prediction for drug discovery by integrating hierarchical information from chemically synthesizable breaking of retrosynthetically interesting chemical substructure (BRICS) fragments and a feature-wise attention mechanism into graph neural networks.

Key Features:

• Hierarchical Information Utilization: HiGNN implements a corepresentation learning approach on molecular graphs using BRICS fragments to capture inherent hierarchical molecular structure.
• Feature-Wise Attention Mechanism: A feature-wise attention block adaptively recalibrates atomic features following the message passing phase to emphasize relevant feature channels for prediction.
• Predictive Performance on Benchmark Datasets: HiGNN demonstrates superior predictive performance across drug discovery-associated benchmark datasets.
• Molecule-Fragment Similarity Mechanism: A molecule-fragment similarity mechanism links model predictions to BRICS subgraphs to enable subgraph-level interpretability.

Scientific Applications:

• Molecular Druggability and Bioactivity Prediction: HiGNN predicts molecular druggability and bioactivities from graph-based molecular representations.
• Lead Discovery and Compound Design: HiGNN aids identification of crucial molecular components and supports design of novel compounds with targeted therapeutic properties.
• Substructure-Level Interpretation: The molecule-fragment similarity mechanism enables investigation of model predictions at the BRICS subgraph level.
• Optimization of Molecular Structures: Interpretability features support rational decision-making during optimization of molecular structures.

Methodology:

HiGNN constructs hierarchical graph representations from BRICS fragments, applies corepresentation learning and message passing, uses a feature-wise attention block to recalibrate atomic features, and computes molecule-fragment similarity for interpretability.