Learning representation for multiple biological networks via a robust graph regularized integration approach.

Area of Science:

• Bioinformatics
• Network Science
• Computational Biology

Background:

• Learning node representations is crucial for analyzing complex biological networks.
• Existing methods struggle with noise and integrating information from multiple, heterogeneous networks.

Purpose of the Study:

• To develop a robust method for learning node representations from multiple biological networks.
• To address challenges posed by network noise and the integration of common and specific information across networks.

Main Methods:

• Utilized denoised diffusion to handle noise and spurious edges in biological networks.
• Introduced a graph regularized integration model to combine refined networks and learn common representation features.
• Employed regularized decomposition to preserve common properties while accommodating specific network information.

Main Results:

• Demonstrated superior performance on noisy networks through simulation studies.
• Achieved robust performance in real-world tasks: drug-target interaction prediction, gene function identification, and species categorization.
• Showcased effectiveness across biological networks of varying scales and sparsity levels.

Conclusions:

• The proposed method effectively eliminates noise and integrates information from multiple biological networks.
• Learned representations are consistent and useful for various downstream biological inference tasks.
• Offers a robust solution for multi-network biological data analysis.