MGN-Net: A multi-view graph normalizer for integrating heterogeneous biological network populations

Area of Science:

• Computational Biology
• Network Science
• Neuroscience

Background:

• Biological data increasingly involves complex, heterogeneous networks (graphs).
• Existing network science methods struggle to extract integral connectional fingerprints from multi-view graph populations.
• Disentangling typical from atypical variations across network samples remains a challenge.

Purpose of the Study:

• To introduce a data-driven method for normalizing and integrating multi-view biological networks.
• To create a single, representative, and topologically sound connectional template from a population of graphs.
• To identify connectional fingerprints and differentiate typical from atypical variations in biological networks.

Main Methods:

• Developed the multi-view graph normalizer network (MGN-Net), a graph neural network (GNN) based approach.
• Applied MGN-Net to normalize and integrate sets of multi-view biological networks into a unified template.
• Utilized MGN-Net to discover connectional fingerprints in healthy and neurologically disordered brain networks (Alzheimer's, Autism Spectrum Disorder).

Main Results:

• MGN-Net successfully generated centered, representative, and topologically sound connectional templates.
• The method effectively captured unique traits of healthy and disordered brain network populations.
• MGN-Net demonstrated superior performance compared to conventional network integration methods in extensive experiments.

Conclusions:

• MGN-Net provides a powerful and generic framework for normalizing and integrating multi-view biological networks.
• The approach enables the discovery of population-specific connectional fingerprints, aiding in the understanding of neurological disorders.
• MGN-Net is adaptable for various graph-based problems, including identifying relevant connections and data integration.