We frequently analyze our gene expression data using WGCNA to identify modules of gene co-expression profiles. However, in generating our original dendrograms we obtain blocks that have unique characteristics.

Some blocks have higher level branches (see block 1, below):

Whereas other blocks from the same data set are more deeply branched (see block 2, below):

In performing subsequent deep splitting, and tree re-cutting, a detectCutHeight is selected (usually 0.99) that is utilized across all blocks. My question is: would it be prudent to select different detectCutHeights for individual blocks in order to isolate apparently unique branches?

Below is a modified example of block 2, wherein if a detectCutHeight of 0.9 was selected (blue dashed line), rather than the standard 0.99, two modules would be identified (red ovals), as opposed to a single larger turquoise module.

We understand the inherent problems that may arise from this approach. We've discovered that some modules identified in the cutting process, have the same color names spanning multiple blocks. Also, that selecting deeper cut heights would result in loss of some other branch information. Any guidance or experience in this matter is greatly appreciated.

I don't recommend setting the cutting arguments differently for each block (it is also not an option in standard WGCNA blockwiseModules). If you are concerned about apparently distinct branches being lumped into a single module, you can (1) decrease mergeCutHeight (this should also get rid of modules with genes in different blocks) and (2) increase deepSplit. Both adjustments may be necessary since they both merge less distinct branches. You may also want to increase minimum module size if you start getting too many small modules. If at all possible, I also recommend running the analysis in a single block; an article about the blockwise analysis is posted at http://www.peterlangfelder.com/blockwise-network-analysis-of-large-data/, here's a quote:

"I emphasize that the blockwise analysis creates an approximation to the network that would result from a single block analysis. The approximation is often very good but the modules are not quite the same. If possible, I recommend running the analysis in a single block; if not, use the largest blocks your computer can handle. "

HTH,

Peter