Detailed results of Application 5: Preservation of female mouse liver modules in male mice

Several studies have explored how co-expression module networks change between mouse tissues [1] and/or genders [2]. Here we re-analyze mouse liver samples of an F2 mouse intercross [3,4] to study whether “female” modules (i.e., modules found in a network based on female mice) are preserved in the corresponding male network. The data sets consist of 135 female and 124 male samples. The modules were identified in the female data as follows [3]: the original over 20000 probes were reduced to 8000 most variant, and of those the 3600 most connected were selected for network analysis. Further, probes mapping to the same gene were merged, resulting in 3421 measured gene expressions. An unsigned network was constructed using the soft thresholding power β = 6, genes were clustered using average linkage hierarchical clustering with dissimilarity equal to one minus the topological overlap matrix [5], and the Dynamic Tree Cut [6] was used to identify modules in the clustering tree. The authors of [3] manually merged several modules whose eigengenes were highly correlated. This analysis identified 12 modules (labeled by colors black, blue, brown cyan, green, greenyellow, lightcyan, lightyellow, midnightblue, purple, red, and salmon) with sizes ranging from 34 genes to 772 genes. The biological significance of the modules was discussed in [3]; one of the modules (blue) was found to relate significantly to body weight. In [3], the male mice were omitted from the analysis since the male network appeared significantly different from the female network. Here we use our module preservation statistics to make this statement more quantitative. Since module identification has not been performed in the male data, we do not use the cross-tabulation methods here and rely instead fully on the network preservation statistics. We re-created the gene clustering dendrogram in the female data using the same network construction techniques as were used in [3] (Figure 1A). Additionally, we constructed an analogous clustering dendrogram in the male data, but did not identify modules. Instead, in Figure 1B we show the male dendrogram with the corresponding female module colors. This figure provides a visual clue to module preservation. Solid blocks of a single color (such as, for example, brown) indicate that the female module is strongly preserved in the male network. This figure indicates that most modules are likely to be strongly preserved since they can be observed as blocks of solid color under the male dendrogram. The only possible exceptions are the salmon and lightyellow modules that are difficult to locate under the male dendrogram. The composite statistics Zsummary and medianRank tell a similar though more precise story. In Figures 1C and D we show Zsummary and medianRank as a function of module size. The composite Zsummary indicates that all modules except the lightyellow and salmon are strongly preserved. For the lightyellow module we find very weak evidence of preservation (Zsummary = 2.0), while for the salmon module we find moderate evidence for preservation (Zsummary = 7.6). The salmon and lightyellow modules also exhibit the worst median rank (13) of observed statistics. Detailed results including values of all preservation statistics are presented below and in Supplementary Table S6. What is the meaning of the weak preservation scores of the lightyellow and salmon modules? Since the network construction is sensitive to outliers, one may suspect that the correlations underlying the lightyellow and salmon modules are caused by outlier samples. Such samples may be visually identified in module expression heatmaps and the corresponding eigengene plots. In Figure 1E we show the heatmap and eigengene plot for the salmon module. It clearly shows an outlier sample in which genes of the salmon module have abnormally high