Self-supervised graph representation learning integrates multiple molecular networks and decodes gene-disease...

Area of Science:

• Computational Biology
• Network Medicine
• Genomics

Background:

• Discovering disease-relevant modules in molecular networks is challenging due to data noise and context specificity.
• Integrating multiple interactomes requires advanced computational methods for robust gene functional analysis.

Purpose of the Study:

• Introduce Graphene, a two-step self-supervised representation learning framework for integrating molecular networks.
• Adapt Graphene for gene functional analysis and disease-related discoveries through downstream re-training.

Main Methods:

• Utilize graph neural network (GNN) pre-training for initial node embeddings.
• Employ a graph attention architecture for re-training Graphene.
• Integrate multiple molecular networks and omics features.

Main Results:

• Graphene achieves superior performance in pathway gene recovery, disease gene reprioritization, and comorbidity prediction.
• Successfully recapitulates tissue-specific gene expression and demonstrates shared heritability of common mental disorders.
• Demonstrates potential for refining genome-wide association study (GWAS) hits.

Conclusions:

• Graphene offers a powerful approach for deciphering gene function within network contexts.
• Provides mechanistic insights into diseases by integrating genomic, network, and phenotypic data.
• The framework is adaptable and can be updated with new biological data.