The nearest centroid analysis method as implemented in the WGCNA R library [63], was used to determine which combination of edges best characterized MDD subjects and differentiated them from normal controls. Because the subject pool was unbalanced with regard to group size (121 MDD versus 37 control subjects), the MDD subjects were divided into four datasets reflecting their original study source (groups of 31, 29, 25, and 36 subjects), such that each of the datasets contained roughly the same number of subjects and was comparable to the number of control subjects. The final nearest centroid categorization used the rankings from the four separate datasets and combined them with the metaanalysis method implemented in the rankPvalue function (pValueLowScale) of the WGCNA R library. Within each of the four datasets, we performed five-fold cross validation in which the data were split into five bins, with four of these used at any one time as a training set and the remaining bin used as a test set. Edge connectivity selection (based on the correlation test) in each of the training sets was performed
