Continual representation learning for evolving biomedical bipartite networks.

Area of Science:

• Biomedical Informatics
• Network Science
• Machine Learning

Background:

• Biomedical interactions (gene-disease, drug-target) are often bipartite networks.
• Network Representation Learning (NRL) aims to create low-dimensional vector embeddings for these networks.
• Existing NRL methods struggle with the unique topology of bipartite networks and static data, limiting their application to dynamic biological systems.

Purpose of the Study:

• To develop a novel NRL approach for biomedical bipartite networks.
• To accurately preserve the intricate bipartite structure during representation learning.
• To enable efficient, online updates of node representations for evolving networks.

Main Methods:

• A customized autoencoder designed to capture proximity within bipartite bicliques (2x2 subgraphs).
• Integration of structure-preserving techniques with continual machine learning principles for incremental learning.
• Development of an online strategy to update node representations dynamically.

Main Results:

• The proposed approach produces high-fidelity and computationally efficient node representations.
• It effectively preserves both global and local structures of the bipartite networks.
• Experiments on biomedical bipartite networks demonstrate the approach's effectiveness and validity.

Conclusions:

• The novel NRL method successfully addresses limitations of existing approaches for biomedical bipartite networks.
• It offers a robust solution for learning representations from dynamic, evolving biological network data.
• The approach provides meaningful embeddings with high accuracy and efficiency.