Dear reader

I am using edgeR to do differential gene expression analysis on my mRNA seq data.

I just have two group, each of 5 replicates, that I want to compare.

Out of curiosity I performed both the classic EdgeR (exact test) and the GLM (glmQlFtest) approaches.

Surprisingly, I got really different results, the exact test gave me 192 upregulated genes but the GLM approach only 82.

Is it normal to get such big differences between the two methods?

Which test is preferred for a simple design like this?

Thanks, Jurgen

You should just use the quasi-likelihood pipeline, which controls type I error better than the exact test or likelihood ratio test.

Yes, what you see is what is expected. The whole purpose of the edgeR-quasi pipeline is to give more rigorous control of the FDR so it will naturally be more conservative than the older pipelines. The exact test was a gold-standard when it was first published (16 years ago) so it is still not a bad test in an absolute sense.

By the way, the difference that you see is considered a small difference in the context of differential expression analyses. If you have 15,000 genes in your analysis (say), then the two tests apparently give different results for only 110/15000 = 0.7% of all genes. To compare the two tests it is more relevant to compare the two gene rankings rather than just the number of genes passing a particular FDR theshold.