Knowledge Graph Neural Network With Spatial-Aware Capsule for Drug-Drug Interaction Prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Artificial intelligence in medicine

Background:

• Predicting drug-drug interactions (DDIs) is crucial for drug development and patient safety.
• Graph neural networks (GNNs) show promise for DDI prediction but struggle with spatial relationships.
• Existing methods lack effective capture of higher-level features from neighboring nodes.

Purpose of the Study:

• To introduce KGCNN, a novel model for comprehensive DDI prediction.
• To address limitations in capturing spatial relationships and higher-level features in GNN-based DDI prediction.
• To leverage biomedical knowledge graphs (BKGs) for improved DDI prediction.

Main Methods:

• Developed KGCNN, a message-passing GNN framework with propagation and aggregation.
• Utilized a biomedical knowledge graph (BKG) to govern information propagation based on semantic relationships.
• Introduced a spatial-aware capsule aggregator to capture spatial relationships and higher-level features.

Main Results:

• KGCNN demonstrated superior performance in DDI prediction tasks.
• Experimental results on two datasets validated the model's effectiveness.
• The model achieved quantified performance improvements in predicting DDIs.

Conclusions:

• KGCNN effectively captures spatial relationships within BKGs for DDI prediction.
• The spatial-aware capsule aggregator enhances the representation of molecular interactions.
• KGCNN offers a superior approach for computational DDI prediction.