GLMs vs pairwise testing

[jk1124]

I am looking for genes which are differentially expressed in any of 5 samples. Should I use a GLM or pairwise DE testing for this? I know that GLMs are included in DESeq and edgeR but am unsure how they compare to these programs' standard comparisons between two conditions. Can anyone explain this to me? Are the two equally reliable?

[Simon Anders]

You should explain a bit more about the biology of your experiment to get useful advice.

I say this because your question, "looking for genes which are differentially expressed in any of 5 samples", can be answered without any analysis whatsoever: I am very sure that the answer is "all your genes", because it is rather unlikely that a gene has exactly the same expression in all five samples. This should show that your question is ill-posed.

[jk1124]

Let me rephrase my question more generally. What are the advantages and disadvantages of using a GLM over an exact test? I noticed, for example, that DESeq2 has defaulted to using GLMs for all differential expression tests, even for pairwise testing. Is this merely to automatically encompass multifactor experiments, or is the exact test now considered less reliable?

[Simon Anders]

Th exact test is only suitable for simple two-group designs, while a GLM can also accommodate more complicated experimental design. The loss in precision of GLMs over the "exact" test is rather small, and hence we decided do drop the latter in DESeq2 in order to provide a more uniform user interface.