Hi there,

I was looking through the vignette for tximport, and it has recommendations for how to import data for downstream analysis in edgeR, DESeq2 and limma-voom, but it does not mention the lesser-used limma-trend. The edgeR method stores the length corrections in y$offset, but the voom() function does not use the y$offset so tximport recommends importing either "scaledTPM" or "lengthScaledTPM". The limmaUsersGuide() suggests doing logCPM <- cpm(y, log = TRUE, prior.count = 3). I thought that since cpm() is an edgeR function it would use the y$offset, but looking at the code of cpm.DGEList, it doesn't use y$offset either. So am I correct in assuming that I should use the tximport method for limma-voom, but then should use cpm() instead of voom()?

Thanks,

Jenny

You can force cpm to use the offset matrix by passing exp(offset) as the library size. For cpm, I don't think this should make any difference relative to using lengthScaledTPM without offsets, so the only advantage is being able to use the same tximport run for edgeR, DESeq2, and limma.

For voom, I've written a custom version that uses an offset matrix in place of the normalized library sizes: https://github.com/DarwinAwardWinner/CD4-csaw/blob/master/scripts/utilities.R#L254-L390 (Although now that I think about it, it should be possible to modify voom to accept a matrix-like lib.size argument just like cpm, instead of having a separate function for it.) I don't know that it's optimal or handles every edge case, but it has been working for me.

One way of computing log-CPMs with offsets is to do something like this:

cpm(y$counts, lib.size=exp(y$offset), log=TRUE, prior.count=3)

... assuming your offsets are on a scale that is interpretable as the log-library size. This is what edgeR assumes the offsets to be, check out ?scaleOffset.

Also, I assume that the length corrections occur between samples, rather than between genes. edgeR will mostly ignore systematic differences in the sizes of the offsets between genes.

Note in the tximport vignette that instead of count + offset as input to limma-voom, the hand off is to use countsFromAbundance="lengthScaledTPM" to get around the offset.