CUBCO: Prediction of Protein Complexes Based on Min-cut Network Partitioning into Biclique Spanned Subgraphs

Abstract

Proteins are essential components of all living organisms and participate in almost every biological process. However, most proteins do not function as a single entity; instead, they often interact with other proteins to form large macromolecules, i.e. protein complexes, that are involved in different cellular functions. Identifying protein complexes allows assigning functions to proteins of yet unknown roles by using the known function of their interacting partners, following the principle of guilt-by-association (Tian et al. 2008). Moreover, due to the protein structures, proteins are often involved in more than one complex in different subcellular compartments and biological processes. Therefore, studying protein complexes is important to understand the functional principles of the cell system, from signaling to metabolism (Pawson and Nash 2000;Maslov and Sneppen 2002;Reyes-Turcu et al. 2009;Sweetlove and Fernie 2018), and provide a better understanding hierarchy of intra-and inter-cellular activities (Bauer and Kuster 2003).

Computational identification of protein complexes from network interactions: Present state, challenges, and the way forward

“…The last three approaches in this category, namely: Protein Complexes from Coherent Partition (PC2P) [58] , Greedy Clustering Coefficient and its Variants (GCC-v) [57] , and minimum CUt to detect Biclique spanned subgraphs as protein COmplexes (CUBCO) [59] , formalize the protein complexes as biclique spanned subgraphs to include both sparse and dense complexes. As a result, they resolve the issues with community density and size observed in existing approaches.…”

“…Unlike the two previous approaches based on local graph properties, CUBCO [59] utilizes global properties to partition the network into biclique spanned subgraphs. The local algorithms utilize the local node properties, such as their first and second neighborhoods.…”

“…To facilitate a fair comparison, the approaches are selected based on two criteria, (i) the public availability of executable code and implementations and (ii) that the method does not rely on any additional biological knowledge or data. Therefore, MCL [22] ), MCODE [7] , CFinder [1] , AP [23] , CMC [45] , PEWCC [99] , Prorank+ [30] , and DPC-NADPIN [73] are selected from the node affinity-based category; ClusterOne [55] ), Core&Peel [63] , IMHRC [48] , PC2P [58] , CC [57] , WCC [57] , OCC [57] , OWCC [57] , and CUBCO [59] from the cluster quality-based category; and DPCMNE [53] from network embedding-based category. For the parameter(s)-dependent approaches (see Table 3 ), we used the default parameter values as suggested in the corresponding original studies.…”

“…To summarize the twelve performance measures and to enable its visualizations, two composite scores are defined. The first composite score is given by the sum over MMR, FRM, SEP, ACC, F-measure [55] , [13] , [87] , [57] , [57] , [59] , and the second one is the sum over MMR and F-measure + across different threshold values [49] , [57] , [57] , [59] .…”

“…The interactions with low confidence values may, in turn, be discarded in subsequent analyses. On the other hand, link prediction algorithms [10] , [41] and different local and global network topological metrics can be employed to score false-negative interactions and insert the high-scored ones to the original PPI networks as a result [59] .…”