Background Proteins often assemble into higher-order complexes to perform their biological functions. Such protein–protein interactions (PPI) are often experimentally measured for pairs of proteins and summarized in a weighted PPI network, to which community detection algorithms can be applied to define the various higher-order protein complexes. Current methods include unsupervised and supervised approaches, often assuming that protein complexes manifest only as dense subgraphs. Utilizing supervised approaches, the focus is not on how to find them in a network, but only on learning which subgraphs correspond to complexes, currently solved using heuristics. However, learning to walk trajectories on a network to identify protein complexes leads naturally to a reinforcement learning (RL) approach, a strategy not extensively explored for community detection. Here, we develop and evaluate a reinforcement learning pipeline for community detection on weighted protein–protein interaction networks to detect new protein complexes. The algorithm is trained to calculate the value of different subgraphs encountered while walking on the network to reconstruct known complexes. A distributed prediction algorithm then scales the RL pipeline to search for novel protein complexes on large PPI networks. Results The reinforcement learning pipeline is applied to a human PPI network consisting of 8k proteins and 60k PPI, which results in 1,157 protein complexes. The method demonstrated competitive accuracy with improved speed compared to previous algorithms. We highlight protein complexes such as C4orf19, C18orf21, and KIAA1522 which are currently minimally characterized. Additionally, the results suggest TMC04 be a putative additional subunit of the KICSTOR complex and confirm the involvement of C15orf41 in a higher-order complex with HIRA, CDAN1, ASF1A, and by 3D structural modeling. Conclusions Reinforcement learning offers several distinct advantages for community detection, including scalability and knowledge of the walk trajectories defining those communities. Applied to currently available human protein interaction networks, this method had comparable accuracy with other algorithms and notable savings in computational time, and in turn, led to clear predictions of protein function and interactions for several uncharacterized human proteins.

Many, if not most, proteins assemble into higher-order complexes to perform their biological functions. Such protein-protein interactions (PPI) are often experimentally measured for pairs of proteins and summarized in a weighted PPI network, to which community detection algorithms can be applied to define the various higher-order protein complexes. Current methods, which include both unsupervised and supervised approaches, often assume that protein complexes manifest only as dense subgraphs, and in the case of supervised approaches, focus only on learning which subgraphs correspond to complexes, not how to find them in a network, a task that is currently solved using heuristics. However, learning to walk trajectories on a network with the goal of finding protein complexes lends itself naturally to a reinforcement learning (RL) approach, a strategy that has not been extensively explored for community detection. Here, we evaluated the use of a reinforcement learning pipeline for community detection in weighted protein-protein interaction networks to detect new protein complexes. Using known complexes, the algorithm is trained to calculate the value of different possible subgraph densities in the process of walking on the network to find a protein complex. Then, a distributed prediction algorithm scales the RL pipeline to search for protein complexes on large PPI networks. The reinforcement learning pipeline applied to a human PPI network consisting of 8k proteins and 60k PPI results in 1,157 protein complexes and shows competitive accuracy with improved speed when compared to previous algorithms. We highlight protein complexes harboring minimally characterized proteins including C4orf19, C18orf21, and KIAA1522, suggest TMC04 to be a putative additional subunit of the KICSTOR complex, and confirm the participation of C15orf41 in a higher-order complex with CDAN1, ASF1A, and HIRA by 3D structural modeling. Reinforcement learning offers several distinct advantages for community detection, including scalability and knowledge of the walk trajectories defining those communities.