MVGCN

MVGCN predicts unobserved links in biomedical bipartite networks to elucidate biomolecular interactions underlying human diseases.

Key Features:

• Multi-View Heterogeneous Network (MVHN) Construction: Constructs an MVHN by integrating similarity networks with the original biomedical bipartite network to represent bioentities across multiple views.
• Self-Supervised Learning Strategy: Derives initial node embeddings from the bipartite network via a self-supervised learning approach.
• Neighborhood Information Aggregation (NIA) Layer: Uses a Neighborhood Information Aggregation (NIA) layer that iteratively updates node embeddings by aggregating inter- and intra-domain neighbor information across each view.
• Integration of Multiple Views: Combines embeddings from multiple NIA layers within each view and integrates embeddings across views to produce final node representations.
• Discriminator-Based Link Prediction: Applies a discriminator model on final node embeddings to predict the existence of links between bioentities.

Scientific Applications:

• Link prediction in biomedical bipartite networks: Predicts unobserved associations among biomolecular entities to inform studies of molecular mechanisms and disease.
• Protein–protein interaction networks: Identifies potential protein-protein interactions.
• Gene–disease associations: Detects candidate gene-disease associations.
• Generalization and benchmarking: Generalizes across datasets and tasks and demonstrated superior performance on six benchmark datasets covering three biomedical tasks.

Methodology:

Constructs an MVHN integrating similarity networks and the original bipartite network, derives initial embeddings via self-supervised learning, applies iterative NIA layers aggregating inter- and intra-domain neighbor information across views, combines embeddings across layers and views to obtain final node embeddings, and uses a discriminator on the final embeddings for link prediction.