Should zero expression ever be tested for significance?

Question

[ivan.o.kryukov]

I am trying to figure out when the comparison between FPKM is performed in cuffdiff. The NOTEST clause is invoked when both values are close. If one of the values is zero, there should be no testing for significance (or so it seems). However, using version 2.2.1, I get several examples where -inf log fold change results in a significant q-value (always 0.047).

I have checked, and for all of such cases, I get no reads for the transcripts in question.

#regular case - no objections here
OK      2.7692  0       -inf    -nan    0.0011  0.226273

#weird case - not sure why this is "significant"
OK      2.71508 0       -inf    -nan    5e-05   0.0470179       yes

Is there a general case where 0 expression is tested?

[Martombo]

imagine that the most highly expressed gene in condition A is switched off completely in condition B and has therefore 0 counts. wouldn't you think that's a very significant difference? that said, I don't know exactly the steps in cuffdiff

[ivan.o.kryukov]

This makes sense biologically, yes. But I am not sure what statistical basis there could be for comparing 0s (other than pseudocounts, or Bayesian priors). In the previous cases I've seen - those are not compared at all. Otherwise, they will automatically _always_ given you a significant difference.