Correcting for multiple-testing - interaction model with large differences in variance between groups
Entering edit mode
4.1 years ago
npharmston • 0


I am attempting to find genes whose response to treatment is different depending on the whether gene X is KO'ed out or not.

So I have an experiment where I have condition + treatment + condition:treatment .... where condition is KO or WT and treatment is Ctrl, M1, M2 (i.e. grown on different media).

So I do the standard

dds = nbinomLRT(dds, full= ~ condition + treatment + condition:treatment, reduced = ~condition + treatment)

 results.interaction  = results(dds)

 table(results.interaction$padj < 0.1)


16183     5


hist of p-values

Now from this I see this hill-shaped distribution of p-values which comes from - as I understand - large differences within groups - which I do - if I look at the level of individual genes its pretty obvious that the expression of genes in the KO condition are much more variable than in the WT.

So the question what is the best way of proceeding with this / fixing it. Are there more diagnostic tests I should do first before even considering this?

I have found a lot of talk/presentations/posts using fdrtool to reestimate the p-values and go from there. Typically by taking the stat from a DESeq2 result and using this to re-estimate p-values. However I am really after all those genes with a significant interaction not those that I can identify from pairwise comparisons (i.e. extracting stat from conditionKO.treatmentM1 or conditionKO.treatmentM2), so ....

corrected <- fdrtool(results(dds)$stat,  statistic= "normal", plot = T)
table(corrected$qval < 0.1)

12971  3219
  1. Is using FDRtool to recompute p-values the best option in this situation? If not, are there any other suggestions?

  2. Are there diagnostics I should perform to actually say "yes, lets do this" rather than my p-value distribution looks funny?

  3. Is there anyway to implicitly tell DESeq to consider things differently - to account for large variances in the groups ?

  4. Is there something else going on here?

Thanks in advance,

deseq2 fdrtool fdr multifactorial design • 1.1k views

Login before adding your answer.

Traffic: 2184 users visited in the last hour
Help About
Access RSS

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6