Hello, I've reviewed A: WGCNA Hub Gene Selection Method. I'm calculating module membership (MM) using geneModuleMembership = as.data.frame(cor(datExpr, MEs, use = "p"))

Because kME is a more sophisticated approach as it includes error checking and other features, I also used signedKME( datExpr, MEs, outputColumnName="KME", corFnc="bicor")

For this I'm using reads from the affy hgu133plus2 platform. As a test, I examined the three probes that target a certain gene, and found that the most significant MM correlation was to the tan module in three cases , whereas the kME significance was tan, tan, and dark orange.

However, the dendrogram tree cutting places these three probes in green, gray, and tan. Which of these calculations should I use as the final call for which module a probe most significantly resides within?

Also, kIM only calculates the connectivity of a probe within its home "dendrogram" module, so I'm not easily able to compare kIM values within all the modules? Dr. Langfelder writes kME and kIM are usually very similar, but I have no way to establish this?

Thanks much,

Robert Robl

It often happens that one gene (probeset) has a high kME to several modules. Since modules are based on TOM, not on the adjacency (correlations), they provide a slightly different picture of the data which can change the apparent module membership, especially if the modules have strongly correlated eigengenes. TOM also often seems to "prefer" larger modules over small ones. Which one should you settle on? I personally prefer the kME as measured by bicor, but I often settle on simple binary module membership as determined by tree cutting. The point is that the most important genes are those that are intramodular hub genes, and the module to which the hubs belong is usually pretty clear.

The difference in MM vs. kME is most likely due to using bicor for one and Pearson correlation for the other, otherwise they should be exactly the same.

kIM and kME values are "similar" in the sense that for each module, the values of kME and kIM for genes in the module tend to be highly correlated. The ability to quickly calculate kME for all genes (not just those within a module) is one of the advantages of kME. Another advantage is that kME values can be directly compared. Since kIM is a sum of connectivities and the number of genes in each module differs, the kIM values across different modules cannot be easily compared.

Hope this helps.

At low sample sizes, I'm finding that kME can prioritize some genes that are only moderately or weakly connected to other genes in it's assigned module, but just happen to correlate well with the module eigengene anyways. Since hub implies high connectivity and kIM directly measures connectivity, whereas kME is more of a proxy for kIM with nicer properties, I'm finding myself preferring kIM when possible.

One downside as you've identified is that kIM is less comparable across modules. You can address this somewhat by normalizing kIM across modules. You can use the scaleByMax = TRUE parameter of intramodularConnectivity to normalize by max kIM of genes in the module. Alternatively, you can scale by maximum possible kIM using kIM/(n-1) where n is the module size.

Another potential issue is that since kIM is sum of adjacencies, it's somewhat sensitive to the chosen power. Higher powers will emphasize high correlation to a small number of genes, and lower powers will emphasize moderate correlation to a large number of genes.

You can compare kIM and kME within module assignments. The plot below is for my network (I used Pearson correlation and signed network).

Some genes have higher kME with a different module. For example, in the heat map below (colored by kME), all genes were assigned to module "blue".