ICoN: integration using co-attention across biological networks.

Area of Science:

• Bioinformatics
• Computational Biology
• Network Science

Background:

• Molecular interaction networks are crucial for understanding cellular functions.
• Integrating diverse networks improves tasks like gene module detection and protein function prediction.
• Extracting meaningful protein features from noisy, sparse heterogeneous networks is challenging.

Purpose of the Study:

• To propose ICoN, an unsupervised graph neural network model for integrating multiple protein-protein association networks.
• To generate robust protein feature representations by integrating topological information across networks.
• To enhance downstream task performance through effective network integration.

Main Methods:

• ICoN utilizes an unsupervised graph neural network architecture.
• A co-attention mechanism facilitates cross-network communication during training.
• Denoising training techniques improve model robustness by reconstructing original networks from corrupted versions.

Main Results:

• ICoN successfully integrates multiple protein-protein association networks.
• The model generates improved protein feature representations.
• ICoN surpasses individual networks and existing integration models in gene module detection, gene coannotation, and protein function prediction.
• ICoN demonstrates enhanced robustness against network noise.

Conclusions:

• ICoN offers a promising approach for integrating diverse protein-protein association networks.
• The model achieves biologically meaningful protein representations.
• ICoN improves performance and robustness in key bioinformatics tasks.