Weighted gene co-expression network analysis (WGCNA)

We constructed gene co-expression networks with the WGCNA R package (Langfelder and Horvath, 2008). We used the counts per million normalizations with measured values in at least 60% of all the samples analyzed (n = 17,588 and n = 17,646 for the high and low response groups, respectively). We chose the cutoff to eliminate noise from very rarely detected genes or whose expression was difficult to correlate with other genes due to the few samples measured. An initial hierarchical clustering with the expression data confirmed that there were no outliers. We used a scale-free topology criterion to choose the soft threshold power β. The β, six, resulted in approximate scale-free topology using a free topology fitting index. A signed hybrid network was constructed using the adjacency matrix of the biweight mid-correlation of all pairwise comparisons of gene expression values, that is, the matrix of connection strengths by using the soft threshold power. Modules were identified using the cutreeDynamic function, setting the branch height cutoff to 0.99, the minimum module size to 50, and the merging cut height, that is, the dissimilarity threshold below which separate modules would be merged, to 0.35.

We calculated eigengene-based connectivity, also known as module membership kME measures for a particular gene within a given non-preserved module to identify highly connected or hub genes representative of the module’s overall function with a high likelihood to be critical components within the module. We determined module membership by correlating the gene and the module eigengene value (the first principal component of each module output value). This value quantified how close a gene is within a given module, and we applied this measure to detect the hub genes. Genes with absolute kME membership ≥0.9 were considered hub genes to the respective module (Langfelder and Horvath, 2008).