HI,

I have recently used the WGCNA for a whole brain imaging experiment and it seems to have worked well, but I have a question. First some background: The experiment consisted of imaging whole brain samples(n=20) which were stained with c-fos (a marker of neural activity). So I have a 20x300 matrix with each column representing the number of activated cells in said brain region. I can achieve a scale free topology with a power of 7 or higher and I get a decent number of modules (4-6 depending on the minmodule size) which make sense biologically and correlate (as high as .75) with my trait of interest. However, I did not filter out any regions with the exception of regions with mostly 0 cells activated. My questions is should I filter out regions and if so, how? I've seen some remove the mean/median for each gene (in my case brain region) but I have also seen filtering out genes with low variance. If I use one of these techniques should I do it before or after I log transform my data?

Thanks!

I'd try to filter on the counts, not on variance, i.e., filter out regions for which the proportion of measurements with number of activated cells bigger than a threshold (say 5) is smaller than a minimum acceptable proportion (e.g., 25%). If you're just log-transforming the data (without other transformation such as normalization), it doesn't matter when you do the filtering, but do adjust the threshold (e.g., if your log transformation is log(x+1), turn the threshold 5 into log(5+1)).

I would worry about normalization. In this case, you may want to work not with the number of activated cells but with the proportion, since the sections presumably don't cover the exact same volume or total number of cells. The question is whether your staining also shows non-activated cells and whether you can count them.