PC2P: parameter-free network-based prediction of protein complexes

Abstract

Motivation Prediction of protein complexes from protein–protein interaction (PPI) networks is an important problem in systems biology, as they control different cellular functions. The existing solutions employ algorithms for network community detection that identify dense subgraphs in PPI networks. However, gold standards in yeast and human indicate that protein complexes can also induce sparse subgraphs, introducing further challenges in protein complex prediction. Results To address this issue, we formalize protein complexes as biclique spanned subgraphs, which include both sparse and dense subgraphs. We then cast the problem of protein complex prediction as a network partitioning into biclique spanned subgraphs with removal of minimum number of edges, called coherent partition. Since finding a coherent partition is a computationally intractable problem, we devise a parameter-free greedy approximation algorithm, termed Protein Complexes from Coherent Partition (PC2P), based on key properties of biclique spanned subgraphs. Through comparison with nine contenders, we demonstrate that PC2P: (i) successfully identifies modular structure in networks, as a prerequisite for protein complex prediction, (ii) outperforms the existing solutions with respect to a composite score of five performance measures on 75% and 100% of the analyzed PPI networks and gold standards in yeast and human, respectively, and (iii,iv) does not compromise GO semantic similarity and enrichment score of the predicted protein complexes. Therefore, our study demonstrates that clustering of networks in terms of biclique spanned subgraphs is a promising framework for detection of complexes in PPI networks. Availability and implementation https://github.com/SaraOmranian/PC2P. Supplementary information Supplementary data are available at Bioinformatics online.

Efficient and accurate identification of protein complexes from protein-protein interaction networks based on the clustering coefficient

“…We compared the performance of the four versions of our greedy algorithm (GCC-v) with twelve state-of-the-art approaches, including: Markov Clustering (MCL) [24] , Molecular Complex Detection (MCODE) [20] , CFinder [21] , Affinity Propagation (AP) [23] , Clustering-based on Maximal Cliques (CMC) [22] , Clustering with Overlapping Neighbourhood Extension (ClusterOne) [18] , PEWCC [49] , Prorank + [50] , Discovering Protein Complexes based on Neighbor Affinity and Dynamic Protein Interaction Network (DPC-NADPIN) [51] , Core&Peel [19] , Inter Module Hub Removal Clustering (IMHRC) [52] , and Protein Complexes from Coherent Partition (PC2P) [38] . To facilitate fair comparison, we considered only approaches for which publicly available implementation exists and that do not rely on any additional knowledge (e.g.…”

“…The larger values for these scores are indicative of better performance. Moreover, to summarize these twelve performance measures, first we calculated a composite score that corresponds to the sum of four metrics, MMR, FMR, ACC, and F-measure [18] , [58] , [59] , [38] , [38] . Second, we calculated the MMR and F-measure + over predicted protein complexes with different overlap scores and employed their sum as suggested in [60] .…”

“…We have shown that partitioning a PPI network into biclique spanned subgraphs provides the best performing approach to identify protein complexes [38] . However, while this approach is parameter-free and allows to detect both dense and sparse protein complexes, it does not scale well.…”

“…Yet, these algorithms depend on multiple parameters, which render it difficult to gauge the performance in absence of optimal parameter values for different combinations of PPI networks and gold standards. It was recently shown that a parameter-free approach, that models protein complexes as biclique spanned subgraphs, outperforms the existing, seminal approaches [38] and allows for the identification of both dense and sparse clusters; however, only in unweighted networks of limited size.…”

A New Method for Recognizing Protein Complexes Based on Protein Interaction Networks and GO Terms

“…DPCMNE detects protein complexes via multilevel network embedding ( Meng et al, 2021 ). PC2P formalizes protein complexes as biclique spanned subgraphs and converts the problem of detecting protein complex to coherent partition ( Omranian et al, 2021 ). A semi-supervised model based on non-negative matrix tri-factorization is also used to detect protein complex ( Liu et al, 2021 ).…”