REPEL - Random Embedding Perturbation for Enhanced Learning of Protein Function.

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Protein function prediction is vital for understanding biological systems.
• Current embedding methods for protein-protein association networks struggle with spurious proximity, limiting accuracy.
• Heterogeneous network structures complicate distinguishing truly dissimilar proteins.

Purpose of the Study:

• To introduce REPEL, a novel tool for protein function prediction.
• To address the challenge of spurious protein proximity in network embeddings.
• To improve the robustness and accuracy of protein function prediction.

Main Methods:

• Developed REPEL, a function prediction tool utilizing random graph augmentation.
• Applied a uniform weak repelling force to push network nodes apart.
• Assessed the method on simulated and real multiplex protein association networks (yeast, E.coli).

Main Results:

• REPEL consistently improved protein function prediction accuracy compared to Mashup, deepNF, and BIONIC.
• The random repelling augmentation effectively denoises learning by distancing spurious proximities.
• The method demonstrated increased robustness in function prediction.

Conclusions:

• REPEL offers a significant advancement in protein function prediction accuracy and robustness.
• Random graph augmentation acts as a denoising mechanism for network-based learning.
• The proposed graph augmentation principle holds potential for broader applications in graph-based algorithms.