RNA-seq expression measurements and tests

[Rainbird]

Got few dumb questions, might already been asked before.

1, Are the sum of RPKM for all transcripts in a genome supposed to be identical across tissues or samples ?

2, For detecting differential expression, which method works better: parameter method using negative binomial or other distribution, or the nonparameter method like fisher test? I personally prefer fisher, but I am not sure.

3, It seems to me the parameter method needs replicates to do the differential expression test. If I have only one RPKM value for each transcript, it means I have no choice but the fisher method, correct?

Thanks,
Rainbird

[dpryan]

2/3. No one will believe results from a Fisher's exact test these days for measuring differential expression. The best you can do with an experiment lacking replicates is to make a ranked list of candidates to verify with qPCR. It'd prove more economical (and less a waste of your time) to run some biological replicates, if that's possible for whatever experiment you're doing.

1. No, that will vary by tissue and sample. You can see why the sum will vary if you imagine two scenarios. (1) There are only two genes: gene A is 1Kb and has 500,000 reads and gene B is 10Kb and also has 500,000 reads. (2) There are, again, two genes: gene A is 1Kb and has 750,000 reads and gene B is 10Kb and has 250,000 reads. These are, of course, highly contrived examples, but they demonstrate the principal.

[Rainbird]

Hi dpryan, thanks for your relies.

2/3, why pepole no longer trust the fisher's results? For me, at least the idea/hypothesis is solid. Parameter methods would be more problematic since the distribution assumption is often invalid.

1, is this because that the gene length is considered in the RPKM normalization? When trying to measure the expression fold change from RPKM values, should we use: (target gene RPKM in sample A/sum of all gene RPKM in sample A)/(target gene RPKM in sample B/sum of all gene RPKM in sample B) ?

Thanks,

[dpryan]

2/3. You have no idea what the biological variance is, so any statistical test will be meaningless. That's why you're best off just ranking things by fold change. You can also take a rough stab at things using DESeq with the "blind" method, which is about the closest you can come to worthwhile statistics (n.b., read the vignette, it's very informative).

1. Yes, because transcript length is included in RPKM calculations, the sum of them will not be constant across samples or tissues. I'm not really sure what normalizing by the RPKM sum will get you. I would think that more likely to just screw everything up. You can search this forum for other discussions on how to take a stab at creating a ranked list.

Make your life easier and just sequence some biological replicates.