10
$\begingroup$

I downloaded some publicly available RNA-seq data and want to compare those samples carrying a mutation (~4) against the rest (~800!). I ran both EdgeR and DESeq2, and the first results in an asymmetric volcano plot: skewed in one of the sides, meaning that it results in very large logFC that have large (NS) adjusted p-values. So next, in order to see if this asymmetry is persistant, I ran several random partitions of the data as groups (but keeping them highly unbalanced), obtaining the same behavior. Here an example of what I mean (I run it with 4, 20 and 100 to show how this effect decreases as I increase the small class size) upper panel: logFC vs logCPM, lower panel: logFC vs -log10(FDR)

*I tried EdgeR-robust and I still get the same effect.

I do not see this effect when running DESeq2, only very few DE genes. What method would you recommend using when comparing such different class sizes?

$\endgroup$
  • 1
    $\begingroup$ It'd be interesting to post this to the bioconductor site and see what Gordon Smyth says. $\endgroup$ – Devon Ryan Sep 28 '17 at 20:10
3
$\begingroup$

Random partitioning (as you have done) seems like a reasonable thing to do to work out what's going on. From the results you have suggested, it looks like DESeq2 is performing better than EdgeR, so use DESeq2. Can you show the volcano plots and the partition sizes?

Note: I'm a little bit biased in this, because DESeq2 is what I prefer to use (particularly where Salmon is not appropriate).

| improve this answer | |
$\endgroup$
  • $\begingroup$ I think the same, I usually use DESeq2 too, but I was trying to repeat some analyses done by someone else with EdgeR. I edited my original post to include an image of the results. $\endgroup$ – Kraken Sep 27 '17 at 22:57
3
$\begingroup$

In general, you would expect roughly the same distribution on both sides of your volcano plot. The first two plots are concerning since there's so much on the right and not on the left. With such a small number of samples, it is hard to conclude if this is an artifact of not having enough samples, a technical artifact from the way the data was processed, or true biology. I remember reading that DESeq and other packages like a high number of samples for estimating dispersion or something, so this could also be contributing. I think you should find a better dataset (more samples) to prove or disprove your hypothesis.

| improve this answer | |
$\endgroup$
  • $\begingroup$ I don't think this is a problem estimating dispersion: DEseq2 typically only requires a few replicates on either side (I've performed DE analysis on 3/3 with great results). Rather, the directionality could be due to the lack of power to detect significant differences in one direction vs the other. $\endgroup$ – Ben D. Sep 28 '17 at 17:29
  • $\begingroup$ @BenD. Wouldn't power be innately symmetric for cases like this? The power is finding a difference between two coefficient estimates (assumung a Wald test). Whether their difference is positive or negative would seem to be beside the point. $\endgroup$ – Devon Ryan Sep 28 '17 at 20:12

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.