Module representatives for refining gene co-expression modules

This paper concerns the identification of gene co-expression modules in transcriptomics data, i.e. collections of genes which are highly co-expressed and potentially linked to a biological mechanism. Weighted gene co-expression network analysis (WGCNA) is a widely used method for module detection based on the computation of eigengenes, the weights of the first principal component for the module gene expression matrix. This eigengene has been used as a centroid in ak-means algorithm to improve module memberships. In this paper, we present four new module representatives: the eigengene subspace, flag mean, flag median and module expression vector. The eigengene subspace, flag mean and flag median are subspace module representatives which capture more variance of the gene expression within a module. The module expression vector is a weighted centroid of the module which leverages the structure of the module gene co-expression network. We use these module representatives in Linde-Buzo-Gray clustering algorithms to refine WGCNA module membership. We evaluate these methodologies on two transcriptomics data sets. We find that most of our module refinement techniques improve upon the WGCNA modules by two statistics: (1) module classification between phenotype and (2) module biological significance according to Gene Ontology terms.