Part of the hype on networks, co-expression networks are a fancy way to visualize correlation in gene expression. For smaller datasets, such correlation are more readily presented by the plain old heatmap with hierarchical clustering. For larger datasets, it might be helpful. iDEP uses the weighted correlation network analysis (WGCNA) package to identify and present co-expression networks. More information about WGCNA is available here.
Normalized gene expression data is used to select a subset of the most variable genes as input to WGCNA package. As WGCNA is time-consuming, at most 3000 genes is allowed. To reduce noise in correlation in the adjacency matrix, correlation coefficients are raised to a certain power according to the soft threshold. If the soft threshold is 5, then 0.9^5 = 0.59 and 0.3^5 = 0.0024. We should choose a soft threshold so that the network resembles a scale-free graph. Users should choose the smallest power that the scale-free topology fit index reaches 0.9. See here.
The co-expression network is partitioned into modules, which is visualized in the main panel and color-coded. Each module is a group of genes that are highly correlated within themselves. GO enrichment analyses are done on these modules. Modules can have hundreds of genes, making it difficult to visualize by iDEP. For each module, the top 10 genes are shown as a small network, just for visualization purpose. The entire modules can be downloaded an visualized by VisANT or Cytoscape.
By definition, weighted correlation network produced by WGCNA is a fully connected graph, even after the soft thresholding, which merely raises the correlation coefficients to the n-th power.