Molecular complex detection in protein interaction networks through reinforcement learning.

Area of Science:

• Computational Biology
• Network Science
• Systems Biology

Background:

• Protein complexes are crucial for biological functions and are often studied using protein-protein interaction (PPI) networks.
• Existing community detection algorithms for PPI networks often assume dense subgraphs and may lack efficiency in identifying complex structures.
• Reinforcement learning (RL) presents a novel strategy for community detection in biological networks, an area not extensively explored.

Purpose of the Study:

• To develop and evaluate a reinforcement learning pipeline for detecting protein complexes in weighted PPI networks.
• To leverage RL's ability to learn network walk trajectories for identifying higher-order protein structures.
• To scale the RL approach for discovering novel protein complexes in large-scale PPI networks.

Main Methods:

• Developed a reinforcement learning pipeline trained to evaluate network subgraphs for complex identification.
• Employed a distributed prediction algorithm to scale the RL pipeline for large PPI networks.
• Applied the RL method to a human PPI network comprising 8,000 proteins and 60,000 interactions.

Main Results:

• Identified 1,157 protein complexes in the human PPI network.
• Achieved competitive accuracy with improved speed compared to existing algorithms.
• Highlighted uncharacterized protein complexes (C4orf19, C18orf21, KIAA1522) and suggested new subunits for known complexes (TMC04 in KICSTOR, C15orf41 involvement).

Conclusions:

• Reinforcement learning provides scalable and efficient community detection in biological networks by utilizing walk trajectory knowledge.
• The RL pipeline demonstrates comparable accuracy to other methods while significantly reducing computational time.
• The approach facilitates predictions of protein function and interactions, aiding in the characterization of unstudied proteins.