Cuffdiff Very High/Low Ratios, Still Significant

Question

[GPR]

Dear all: I have a couple of questions about the ratios calculated by Cuffdiff, in particular those at the isoform level. I notice that a number of isoforms have very high fold change due to the fact that either "q1" or "q2" has been assigned a zero value. The ratio is therefore an extremely high number, which is also given in as the "test_stat" value. Some but not all of these fold changes are considered significant changes by CuffDiff, and when I inspect the ratios with CummeRbund, I see how a given isoform was not expressed at all (or to extremely low levels) in one sample, compare to the other. Can anybody comment on this?

A second related question: What is the standard FPKM cutoff when considering low abundant transcripts? Should one trust the fact that Cufflinks/CuffDiff, assigned a confidence value to a transcript, even if its read intensity is < 1 FPKM as an example? An example is: I have a gene composed by 3 isoforms. Two of these isoforms have a 1:1 ratio. A third one is overexpressed, with an extreme value, with one sample having no reads, like I mention above. This third isoform is a lot less abundant than the other 2, having an intensity lower than 5 FPKM. Cuffdiff tells me, this change is significant? should I consider it as such?

Many thanks in advance. G

Answer 1

[Istvan Albert]

Both of your questions are generic to situations where one looks and evaluates ratios. I remember them from the days when two color micro-arrays were popular.

In general the answer is that you should never trust values produced by 'black box' type of tools. Each problem is unique and needs to be evaluated in the right context.

I think both of your questions are valid but luckily you also seem to have the right answers as well.

Answer 2

[GPR]

Hello Istvan, thanks for your answer.

So just to make sure I understood you well. Is it therefore common practice to accept these extreme ratios tagged as significant?
This is: whether or not one goes about validating these ratios, would you show these in a publication figure?

I know one has to be cautious with low FPKM values, but given that CuffDiff assigns an OK status to some very low abundant transcripts, is it appropriate to show one of these in a publication figure? What is a standard FPKM cutoff?

Thanks again G.