Triangular Clique Based Multilevel Approaches to Identify Protein Functional Modules

Abstract

Identifying functional modules is believed to reveal most cellular processes. There have been many computational approaches to investigate the underlying biological structures[1,4,9,13]. A spectral clustering method plays a critical role identifying functional modules in a yeast protein-protein network in [9]. One of major obstacles clustering algorithms face and deal with is the limited information on how close two proteins with or without interactions are. We present an unweighted-graph version of a multilevel spectral algorithm which identifies more protein complexes with less computational time [8]. Existing multilevel approaches are hampered with no preliminary knowledge how many levels should be used to expect the best or near best results. While existing matching based multilevel algorithms try to merge pairs of nodes, we here present a new multilevel algorithms which merges groups of three nodes in triangular cliques. These new algorithms produce as good clustering results as previously best known matching based coarsening algorithms. Moreover, our algorithms use only one or two levels of coarsening, so we can avoid a major weakness of matching based algorithms.

Topic: Computing in Biosciences, Data Processing, Numerical Methods.

This work was supported in part by NSF ITR grant DMS-0213305.