Predicting Biomedical Interactions With Higher-Order Graph Convolutional Networks.

Area of Science:

• Biomedical informatics
• Network biology
• Machine learning

Background:

• Biomedical interaction networks are crucial for predicting biological interactions, identifying disease biomarkers, and discovering drug targets.
• Graph neural networks (GNNs) have advanced biomedical interaction prediction but typically only use immediate neighbor information.
• Existing GNNs struggle to capture complex relationships by not integrating features from neighbors at varying distances.

Purpose of the Study:

• To introduce a novel higher-order graph convolutional network (HOGCN) for enhanced biomedical interaction prediction.
• To enable the aggregation of information from higher-order neighborhoods in biomedical networks.
• To improve the learning of informative entity representations by considering multi-distance neighbor features.

Main Methods:

• Developed a higher-order graph convolutional network (HOGCN) model.
• Implemented a mechanism to collect and linearly mix feature representations from neighbors at various distances.
• Applied HOGCN to four distinct biomedical interaction networks: protein-protein, drug-drug, drug-target, and gene-disease interactions.

Main Results:

• HOGCN demonstrated more accurate and calibrated predictions across all tested biomedical interaction networks.
• The model exhibited robust performance on noisy and sparse interaction networks.
• Consideration of features from various neighbor distances significantly improved prediction accuracy.

Conclusions:

• Higher-order graph convolutional networks (HOGCN) offer a superior approach for biomedical interaction prediction compared to standard GNNs.
• HOGCN effectively integrates information from diverse neighborhood levels, leading to improved biological insights.
• The method's success on challenging datasets suggests its potential for broader applications in bioinformatics and drug discovery.