As the author of limma suggested, the Log-transformed RSEM expected count could be reversal of the log-transformation and then feed to voom without round (How to use RSEM value to find differentially expressed genes). But is there necessary to do normalization when using voom function? Besides, the RSEM Expected count data was downloaded from GTEX database.
v <- voom(dge,design) #'none' default
v <- voom(dge,design, normalize.method = 'quantile')
If you have expected count data from RSEM, you just use those data directly. That's assuming you have gene count data. You would then use calcNormFactors on your DGEList prior to running voom, which will do the usual normalization (which is not 'none' as you state). You could hypothetically also use an additional normalization method like 'quantile', but needing to do that is a rare case indeed.
Thanks for your quick and detailed reply sir!
Regarding "just use those data directly", shouldn't I reversal of the log-transformed expected count data (gene level)?
My understanding of GTEx data is that you can get gene or isoform counts as well as TPM values. I have never seen or heard of them providing log counts, which wouldn't make any sense since nobody uses log counts for anything. But I suppose anything is possible, so if you have log counts then yes, you have to convert back to actual counts. There are no methods that I know of that use log counts.
If you have TPM values (way more likely IMO), then you are left with limma-trend to do any analyses. Simply exponentiating won't undo that transformation.
Thanks for your quick and detailed reply sir! Regarding "just use those data directly", shouldn't I reversal of the log-transformed expected count data (gene level)?
My understanding of GTEx data is that you can get gene or isoform counts as well as TPM values. I have never seen or heard of them providing log counts, which wouldn't make any sense since nobody uses log counts for anything. But I suppose anything is possible, so if you have log counts then yes, you have to convert back to actual counts. There are no methods that I know of that use log counts.
If you have TPM values (way more likely IMO), then you are left with limma-trend to do any analyses. Simply exponentiating won't undo that transformation.
Thanks so much sir!