I have RNA-seq data from two targeted panels. So, for each sample/mouse, I have one counts matrix for one set of genes, and another matrix for another set of genes. The major problem here is that the library size for a given sample is different across the two panels, so just slapping the two matrices together and proceeding as normal is right out (one of the samples has like 100 times fewer reads than the rest, but, surprisingly, other quality metrics look great). I've used
limma+voom to test differential expression of individual genes, applying these to each panel separately.
I now wish to ask, "is this set of genes expressed higher in this group than in that?" The problem is that the gene sets I'm looking at combine genes from both panels.
cameraPR using the
t statistics obtained in from each panel separately using the Broad's canoncial pathways. I'm not sure if this is hinky statistically, but it seems ok? If so, it also makes sense scientifically as a way of asking whether those gene sets are changing in specific contrasts across groups.
However, the null hypothesis tested by CAMERA is not "the level of expression for these genes is the same across groups." That's (roughly) the null for ROAST, and I want to test against that null for some specific gene signatures. Is there any hope for me?
Worst case, I'll use something like log-CPM and do parametric testing.
(Searching for combining panels mostly produces results about combining samples that have the same rows, but that's not the problem facing me. Any help would be most appreciated.)