Problem getting soft threshold power using WGCNA for RNA-seq data

Hi jms – soft thresholding (and hard thresholding for that matter) are based on the assumption that use of such thresholds will cut out noise in correlation matrices, thereby “accentuating” the “true” networks in the data. I believe that dozens of empiric experiments substantiate this assumption. Nevertheless, there is still a lot of “art” here (though perhaps some science, too).

When I’ve faced the problem of threshold selection, I’ve used a method like this:

1) Establish a decision criterion for threshold selection first. Ideally, this decision criterion is unrelated to your question of interest.

2) Come up with a way to decide which threshold selection has performed the best according to your criterion.

3) Test several of the results you’ve generated against your decision criterion.

4) Based on 2., choose the threshold you think is performing the best.

Example, suppose a large amount of data from your and other labs suggest a certain battery of genes should be co-expressed. You could make a decision criterion around this observation, and then observe at what power that battery of genes is most clearly discernible. At a high enough threshold, the module disappear even if it is a true positive gene module in your data … that would be too high. At too low a threshold, it might be that each module contains what are likely to be many modules, or extraneous genes, etc.

You could further nuance the basic methodology above based on knowledge you have about the samples in your data. For instance, suppose you have reason to believe that i.e., gene set X should be coexpressed in WT, but not in KO. Now you can make a more focused decision rule, which should help you control type I and II error if properly applied.

Other authors have published other ways of selecting in their work. Can read papers from your target journal to see how they selected a threshold, too.

This link (or other posts here) may be of help, as well: support.bioconductor.org/p/87024/

Read more here: Source link