Three quick questions to the WGCNA experts (and perhaps Dr. Langfelder himself is out there),

1. I'm starting with RMA normalized Affy HGU133 arrays of human systemic lupus erythematosus and normals. Should I run WGCNA on the entire batch which includes both lupus and normals, or should I separate the two, run them separately, and then identify consensus modules? Could you please explain the basis of the best approach? Unfortunately I can only consider separating the two if I have clinical traits other than lupus or normal cohort designation, such as SLEDAI. The set I'm working on as we speak only has cohort designation, so it won't be a candidate for separation.

2. After running correlations against my clinical traits of interest, I see there are positively and negatively correlated modules. Should I only be considering positively correlated modules, or should I look at both positive and negative, thinking that upregulated genes are in  positively correlated modules, and downregulated are in the negatively correlated modules?

3. Speaking of correlations, how does WGCNA correlate a module to a binary trait like normal=0 or lupus=1? I thought a correlation would only make sense against a continuous variable?

Thanks deeply,

Robert Robl

You have mostly answered question 1 - just run WGCNA on the entire set and identify modules that relate to disease status (cohort). That's the most common approach and it identifies modules that group together genes that respond (transcriptionally) in a similar way to the perturbation(s). If you suspect that the disease affects a major transcriptional regulator, you could define modules in cases and controls separately, then study their preservation in the other set - a non-preserved module could indicate a difference in transcriptional (co-)regulation. A consensus analysis in your case does not answer any good questions that I could think of, so I wouldn't bother with it. 

You  should be interested in both positively (upregulated in cases vs. controls) and negatively (downregulated in cases vs. controls) modules.

You can calculate correlations of a binary (coded say as 0,1) and continuous variable. In fact, the correlation p-value in such a case is equivalent to Student t-test of the continuous against the binary variable with the assumption of equal variance.

HTH,

Peter