WGCNA: Can we use high power for signed network?

Question

[Vinesh]

Hello,

I am performing WGCNA on RNA-seq data of 16 samples. I filtered the RNA-seq counts to obtain 18,841 genes and further performed batch correction and covariate adjustments on a vst transformed data using removebatcheffects().

Further using picksoftthreshold() to select the power above R2 0.8. I got the lowest power of 26 which was closest to 0.8 R2 (signed network).

Using this power I performed network and clustering

net = blockwiseModules(object, power = 26, maxBlockSize = 20000,  
                        deepSplit = 0, networkType = "signed", minModuleSize = 30,
                        reassignThreshold = 0, mergeCutHeight = 0.25, numericLabels = FALSE,
                        saveTOMs = FALSE, verbose = 5)

The cluster dendrogram I obtained was as follows:

I recently read a post which suggested to use the table provided in WGCNA faq. The table suggests me to use the power 18 (samples less than 20) for signed network. I tried that and the dendrogram is as follows

My question here is (1) Is 26 power too high or we can go ahead with it? Should we go with low power i.e. 18? (2) The cluster dendrogram we have seen in various papers looks different from what we got. So is the obtained cluster dendrogram fine?

Is there any way I can improve my results?

Thank you

Answer 1

[Crystal571]

1.I think your deepsplit and minmodulesize are too low. I had 18 samples with 18000 genes and I also used vst and removebatcheffects, but I set power=26, deepsplit=3 and minmodulesize=100. And this is my WGCNA below.

2.My teammates prefer to use TPM to do WGCNA because they don't need to care about covariates. Their power is lower and their WGCNA is more beautiful then mine. But I don't think we can use removebatcheffect to TPM, right?

1. Another possible problem is your design and covariates in removebatcheffct. I changed them and I got a better WGCNA.

[ATpoint]

18 is below what the authors recommend. TPM has nothing to do with covariantes, and yes you can regress batch from TPM.

[Crystal571]

I used to do that to my TPM, but I got some negative values in my matrix after removebatcheffect. Is that reasonable?

[Vinesh]

Hello,

Thank you so much for your response

I tried the suggestions mentioned by you but there was no change in the dendrogram. So I went with checking the covariates which might be causing this. I found "batch" to be the one responsible for my dendrogram to look like this otherwise adjusting for other variables I get a good dendrogram