Efficient link prediction in the protein-protein interaction network using topological information in a generative...

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Investigating protein interactions is crucial for understanding intracellular signaling but is experimentally expensive.
• In silico methods, such as network theory and link prediction, are used to identify potential protein-protein interactions (PPIs).
• Protein-protein interaction (PPI) networks represent proteins as nodes and interactions as edges, enabling link prediction for discovering new connections.

Purpose of the Study:

• To develop a novel machine learning approach for link prediction in PPI networks.
• To utilize a generative model for identifying potential protein-protein interactions.
• To offer an efficient in silico method for narrowing down candidate interactions for experimental validation.

Main Methods:

• A two-module tool was created, comprising a data processing framework and a machine learning model.
• The data processing module employed a modified breadth-first search algorithm to extract induced subgraphs.
• An image-to-image translation-inspired conditional generative adversarial network (cGAN) model was utilized for predicting unknown edges.

Main Results:

• The tool was evaluated on PPI networks from five species using metrics like AUROC, AUPRC, and NDCG.
• The machine learning model achieved an average AUROC of 0.915, AUPRC of 0.176, and NDCG of 0.763 across all species.
• The study demonstrated the effectiveness of a cGAN model conditioned on raw topological features for PPI prediction.

Conclusions:

• A software tool for machine learning-based link prediction in PPI networks was developed.
• This work is the first to show that a cGAN model can predict PPIs using only network topology, without needing molecular attributes.
• The developed scripts are publicly available for further research and application.