LaGAT

LaGAT predicts drug-drug interactions (DDI) from large-scale biological knowledge graphs using a link-aware graph attention mechanism to improve prediction accuracy and interpretability.

Key Features:

• Link-aware graph attention: Generates link-specific attention pathways by using one drug's embedding as a query to compute attention weights for each drug pair.
• Knowledge-graph integration: Operates on large-scale biological knowledge graphs compiling extensive drug information from biomedical databases.
• Embedding-based query mechanism: Uses the embedding representation of one drug to query and weight neighbor nodes relevant to the other drug in the pair.
• Topological neighbor selection: Selects topological neighbor nodes according to attention weights to capture semantically relevant information.
• Noise mitigation: Prioritizes semantically relevant nodes per link to reduce the impact of noisy edges and entities in knowledge graphs.
• Classification support: Applied to binary and multi-class DDI classification tasks and reported improved performance relative to classical and state-of-the-art models.
• Attention visualization: Produces visualizable attention pathways that reveal how the model prioritizes nodes for specific drug pairs.

Scientific Applications:

• Drug-drug interaction prediction: Predicts and classifies pharmacological interactions between drug pairs using knowledge-graph-derived features.
• Pharmacology and clinical research: Supports research in pharmacology and related clinical applications by improving DDI prediction accuracy and interpretability.
• Model interpretability analysis: Enables investigation of semantic relationships in biomedical knowledge graphs through attention-pathway visualization.

Methodology:

LaGAT embeds drugs from large-scale biological knowledge graphs and applies a link-aware graph attention mechanism that uses one drug's embedding as a query to compute attention weights, selects topological neighbor nodes based on those weights to form link-specific attention pathways, and provides attention-pathway visualizations.