RSEM: directly compare the gene expression ratio in the same sample?

[mrfox]

Hi all,
I wonder if I could use RSEM value to compare the expression level of 2 genes directly? For instance, gene A's RSEM value is 100, gene B's RSEM value is 200. Then can I say the expression ratio is 2 (gene B vs gene A)?

I have been using FPKM and raw counts but did not have a chance to read the detailed RSEM algorithm.

Thanks!

[dpryan]

You certainly can't use raw or even expected counts for this, though FPKM could work. Things that need to be kept in mind when comparing two genes are: (1) gene length differences need to be accounted for (FPKM should handle this), (2) GC content differences need to be accounted for, (3) differences in 5' or 3' bias need to be accounted for. You might be interested in using Salmon rather than RSEM, since it's mentioned in the preprint that it can account for at least some (possibly all, I'd have to check) of these issues.

[mrfox]

Thanks dpryan for the reply. I agree with you that the gene/transcript length, GC content, 3'UTR bias should be considered. I will definitely try Salmon when having a chance.

[lpachter]

You need to be careful in comparing genes within a single sample, especially if you intend to argue for a "statistically significant" difference. In the absence of replicates you will be unable to determine the extent to which differences are meaningful.

Having said that, if you use the bootstraps in kallisto you will at least be able to assess part of the technical variance (due to the random sequencing of the library and uncertainty introduced in quantification). sleuth uses that information in differential analysis but it will require biological replicates. In other words, with a set of biological replicates in a single condition you can use sleuth to compare two different genes. Otherwise you will be embarking on an ill-advised fishing expedition.

[mrfox]

Hi lpachter,

Thanks a lot for the advice. I agree that lacking biological replicates would not lead to meaningful p-values. I just checked the relative expression ratios, according to my collaborators' request, but did not perform any test.

May I ask you an un-relevant question: for a group of 10 co-expressed genes, may I use the average of the log2(RSEM+1) of these 10 genes to roughly represent the general expression profile? If so, it will be convenient to correlate this single average value to other features.

Thank you again!

You need to be careful in comparing genes within a single sample, especially if you intend to argue for a "statistically significant" difference. In the absence of replicates you will be unable to determine the extent to which differences are meaningful.

Having said that, if you use the bootstraps in kallisto you will at least be able to assess part of the technical variance (due to the random sequencing of the library and uncertainty introduced in quantification). sleuth uses that information in differential analysis but it will require biological replicates. In other words, with a set of biological replicates in a single condition you can use sleuth to compare two different genes. Otherwise you will be embarking on an ill-advised fishing expedition.