I have RNA-seq data for samples in 10 different groups and I would like to find differentially expressed genes. I have used Deseq2 package.

The differentially expressed genes have FDR between 0.03 and 0.05. I checked log2(FC) and found it is low value in most differentially expressed genes.

How I can explain this? Why genes with low log2(FC) are found to be differentially expressed? Also, I have checked the normalized count and I couldn't find any reason.


migrated from stackoverflow.com Oct 4 '16 at 7:49

This question came from our site for professional and enthusiast programmers.

  • 1
    $\begingroup$ Can you show us your model? When you say “10 different groups”, do you actually have 10 different conditions that you are testing against each other in one model? $\endgroup$ – Konrad Rudolph Sep 26 '16 at 9:32
  • $\begingroup$ Also, what’s a “low value”? An FDR > 0.03 isn’t amazing, either. I’d usually choose a lower threshold (though 0.05 is still convention but at least in my cases I can easily accept some false negatives in exchange for specificity). $\endgroup$ – Konrad Rudolph Sep 26 '16 at 9:36

It is perfectly possible that the genes are statistically differentially expressed but have low fold changes. The reason is that you are ultimately testing whether the means of different groups are different. And while they can be different statistically the difference doesn’t have to be biologicaly meaningful.

As a very simple example;

t.test(log2(c(18,19,19,20,20)) , log2( c(21,19,22,23,24)))

p-value = 0.04511

OK, at a 5% level this is just significant, but:

mean(log2(c(18,19,19,20,20)))  -  mean(log2( c(21,19,22,23,24)))


You can see that the means are different but the actual magnitude of change is very small. It is now up to you to decide if such a small change is meaningful. It usually isn’t because your experimental technique itself has some limitations (noise, sensitivity, etc). That is why most people use an FC cut-off, for example 2. This cut-off is in many cases very arbitrary, but is the best way to protect yourself from noise.

  • 2
    $\begingroup$ Well the whole point of DESeq2 and similar tools is that instead of using blunt fold change cutoffs they calculate power based on (pooled) variance estimates. I would generally trust a moderate but significant fold change more than an extreme, non-significant fold change: the latter probably has very low coverage and is thus unreliable, while the former almost certainly has substantial RNA-seq read coverage and is thus unlikely due to noise. $\endgroup$ – Konrad Rudolph Sep 26 '16 at 9:34
  • $\begingroup$ @KonradRudolph Good point! $\endgroup$ – USER_1 Sep 26 '16 at 10:02

Your Answer

By clicking "Post Your Answer", you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.