Heterogeneous graph neural networks for link prediction in biomedical networks

Area of Science:

• Biomedical informatics
• Network science
• Machine learning

Background:

• Heterogeneous graph neural networks (HGNNs) are effective for analyzing complex networks.
• Current HGNN applications are limited in the biomedical field.
• Biomedical link prediction is crucial for understanding biological systems.

Purpose of the Study:

• To evaluate the effectiveness of generic HGNNs for biomedical link prediction.
• To benchmark generic HGNNs against specialized biomedical methods.
• To provide guidelines for hyperparameter optimization in HGNNs.

Main Methods:

• Conducted a comprehensive benchmarking study.
• Evaluated nine generic HGNNs and 42 techniques.
• Utilized eight diverse biomedical datasets and multiple evaluation metrics.

Main Results:

• Generic HGNNs achieve comparable or superior performance to specialized methods.
• The Simple-HGN model demonstrated top performance on four out of eight datasets.
• Results indicate the broad applicability of generic HGNNs in biomedicine.

Conclusions:

• Generic HGNNs are powerful and accessible tools for biomedical link prediction.
• Readily available HGNNs offer a competitive alternative to domain-specific approaches.
• This study provides practical insights and resources for researchers.