fail to pick soft threshold in WGCNA

Question

[RC]

Hi there！

I'm trying to find the co-methylation sites using WGCNA. Methylation status were assessed using the Illumina EPIC microarray. When I run WGCNA's pickSoftThreshold() on 810135 sites (after some common preprocessing steps, such as filter out probes with bad detection P value), server killed this process and reported "task 1 failed - "cannot allocate vector of size 30.2Gb" ".

the code for picksoftthreshold is

sft1=pickSoftThreshold(resids2,blockSize=5000,powerVector =c(seq(4,10,by=1),seq(10,20,by=2)),networkType="signed",verbose=5)

memory of my server: 220GiB

i have tested what if i use a smaller block size, such as 2000, in this step. It result in 12min for each block, so maybe 4800min(more than 3 days ) for the whole data.As this step takes too much time and eats too much memory, how do you guys deal with it? maybe some basic filters? or can i pick the soft threshold on data subsets? one study noted that they get the power from the calculation by the scale-free topology criterion on data subsets.However,they didn't give any details about how they select the subsets(title:Mosaic Epigenetic Dysregulation of Ectodermal Cells in Autism Spectrum Disorder).

Any suggestion?

[WouterDeCoster]

I added code markup to your post for increased readability. You can do this by selecting the text and clicking the 101010 button. When you compose or edit a post that button is in your toolbar, see image below:

Answer 1

[WouterDeCoster]

on 810135 sites

That's quite a lot. I have used WGCNA for RNA-seq analysis, which is with far less "sites". It doesn't surprise me that it takes such a long time for a dataset like that. I usually filter out sites with a low variance.

Answer 2

[Kenny Westerman]

In my (limited) experience, the time taken scales roughly with the size of the overall dataset, both genes/CpGs and samples. So, as you mention it can be helpful to remove low-variance probes as well as perform at least the threshold selection step in a subset of the total samples. When selecting the subset, I would probably attempt to preserve some level of balance across any heterogeneity in the samples, e.g. case/control, sex, etc.