I don't think that controlling the FDR (or any other "statistically founded") cut-off makes sense or is helpful. You should anyways make the whole data available, so there is no need to focus on a subest of gene you deem "differentially expressed" vs. those you don't deem so. If you pick promising candidates, then I would simply look down the sorted list (sorted e.g. by p-value, or just by LFC) until the first gene (or the first k genes) that looks sufficiently interesting to be further investigated.

Note that these high-throuput methods are not to test hypotheses. They are rather to generate hypotheses. It already does not make much sense philosophically to control error rates of tests about hypothese that had not been planned and that lack an a priori model/idea.

I am happy to see this kind of opinion about gene selection, I get your point here. To be clear, I am not so much interested in control of the error rate, rather concerned about the repeatability of the procedure. I have used FDR as cut-off, I wanted to check the sensitivity/specificity of the method. My results, along with your explanation help me understand that FDR, alone, as a cut-off is a arguable strategy.

Thanks.

Hi François,

This seems to be more a comment on how DESeq2 works, rather than the FDR adjustment. The FDR p-values are a function of the number of and size of the un- adjusted values and the algorithm has no concept of number of samples.

The R help for p.adjust will give you details and a set of references for other common adjustment methods other than BH. This is the function DESeq uses internally, either directly or via the filtered_p function (Bioconductor genefilter) to adjust p values.

Of course, the results of the DESeq2 analysis are very much dependent on the experimental design.

Greg