SumGNN: multi-typed drug interaction prediction via efficient knowledge graph summarization

Area of Science:

• Biomedical Informatics
• Machine Learning
• Pharmacology

Background:

• Accurate drug-drug interaction (DDI) detection is crucial for patient safety.
• Large biomedical knowledge graphs (KGs) offer rich information but are challenging to utilize effectively due to size and noise.
• Existing methods often ignore KGs or struggle to integrate them with other data for DDI prediction, particularly for multi-typed interactions.

Purpose of the Study:

• To develop a novel method, SumGNN (knowledge summarization graph neural network), for enhanced multi-typed DDI prediction.
• To effectively leverage large and noisy biomedical KGs by extracting relevant subgraphs and summarizing reasoning paths.
• To improve the accuracy and interpretability of DDI predictions, especially in data-scarce scenarios.

Main Methods:

• SumGNN employs a subgraph extraction module to identify relevant information within large KGs.
• A self-attention mechanism is used for subgraph summarization, generating interpretable reasoning paths.
• A multi-channel module integrates KG knowledge with other data for improved DDI prediction.

Main Results:

• SumGNN significantly outperforms existing methods in multi-typed DDI prediction, achieving up to a 5.54% performance gain.
• The performance improvement is particularly notable for low-frequency DDI types.
• The model provides interpretable predictions through generated reasoning paths.

Conclusions:

• SumGNN offers an effective approach to utilizing large biomedical KGs for accurate and interpretable DDI prediction.
• The method addresses limitations of previous approaches by successfully integrating noisy KG data.
• SumGNN advances the field of DDI prediction, especially for complex, multi-typed interactions.