BoCluSt: bootstrap clustering stability algorithm for community detection in networks

The identification of modules or communities of related variables is a key step in the analysis and modelling of biological systems. Many module identification procedures are available, but few of these can determine the module partitions best fitting a given dataset in the absence of previous information, in an unsupervised way, and when the links between variables have different weights. Here I propose such a procedure, which uses the stability under bootstrap resampling of different alternative module structures as a criterion to identify the structure best fitting to a set of variables. In its present implementation, the procedure uses linear correlations as link weights. Computer simulations show that the procedure is useful for problems involving moderate numbers of variables, such as those commonly found in gene regulation cascades or metabolic pathways, and also that it can detect hierarchical network structures, in which modules are composed of smaller sub modules. The procedure becomes less practical as the number of variables increases, due to increases in processing time.The proposed procedure may be a valuable and robust network analysis tool. Because it is based on comparing the amount of evidence for different module partitions structures, this procedure may detect the existence of hierarchical network structures.