Enhancing Multi-Label Protein Interaction Prediction via Hypergraph Modeling of Higher-Order Patterns.

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions and disease.
• Existing graph models predict pairwise PPIs but miss higher-order biological relationships.
• Capturing complex, multi-label PPIs remains a challenge.

Purpose of the Study:

• To develop HGNN-PPI, a novel framework integrating hypergraph neural networks (HGNNs) with graph neural networks (GNNs).
• To enhance multi-label PPI prediction by incorporating higher-order interaction motifs.
• To improve the accuracy of predicting rare and complex PPIs.

Main Methods:

• Developed HGNN-PPI, a hybrid framework combining GNNs and HGNNs.
• Modeled higher-order interaction motifs using hypergraphs from biological feedback and feedforward loops.
• Employed an asymmetric loss function to address class imbalance in multi-label PPI datasets.

Main Results:

• HGNN-PPI demonstrated superior performance over state-of-the-art methods on benchmark datasets (SHS27k, SHS148k).
• The framework effectively captured higher-order biological motifs and complex interaction patterns.
• Significant improvements were observed in predicting rare interaction types.

Conclusions:

• HGNN-PPI effectively leverages hypergraph structures to model higher-order biological relationships.
• The proposed method enhances the accuracy and scope of multi-label PPI prediction.
• Integrating higher-order motifs is key to advancing PPI prediction accuracy.