Curvature-enhanced graph convolutional network for biomolecular interaction prediction.

Area of Science:

• Computational biology
• Network science
• Machine learning

Background:

• Geometric deep learning excels at analyzing non-Euclidean data.
• Integrating geometric insights into learning architectures is crucial for performance.
• Biomolecular interaction prediction is a key challenge in computational biology.

Purpose of the Study:

• To propose a novel curvature-enhanced graph convolutional network (CGCN) for improved biomolecular interaction prediction.
• To leverage Ollivier-Ricci curvature (ORC) to characterize local geometric properties of networks.
• To enhance the learning capability of graph convolutional networks (GCNs) by incorporating geometric features.

Main Methods:

• Developed a CGCN model that utilizes Ollivier-Ricci curvature (ORC) to assess local network topology.
• Incorporated ORCs into the weight function for feature aggregation during the message-passing process.
• Validated the CGCN model on fourteen real-world biomolecular interaction networks and simulated data.

Main Results:

• The CGCN model achieved state-of-the-art performance, outperforming existing models on thirteen out of fourteen datasets.
• Demonstrated superior performance compared to traditional GCN models across various network densities, sizes, and curvature ratios in simulated data.
• The model's effectiveness was robust across different network characteristics.

Conclusions:

• The proposed CGCN model significantly enhances biomolecular interaction prediction by incorporating network geometry.
• Ollivier-Ricci curvature is an effective feature for improving GCN performance in biological networks.
• CGCN represents a promising advancement in applying geometric deep learning to biological network analysis.