Hi everyone

I have an issue with my differential miRNA expression analysis project. I have two different comparisons treatment1 (T1) vs control and treatment2 (T2) vs control and I performed standard differential expression analysis with DESeq2 and the analysis resulted in 5 significantly differentialy expressed miRNA between T1 and Control and 0 significant miRNA between T2 and controls (padj<0.05). Thus I started to investage what could have gone wrong and thus I plotted the p-value histograms which looked as follows:

histogram1T1vC

histogramT2vC

Alright, thus i followed the instructions on following website https://www.huber.embl.de/users/klaus/Teaching/DESeq2Predoc2014.html where they use fdrtool to estimate the true null model variance. I did this for both datasets T1vC and T2v C according to the link above. output of the result for one dataset (the second one looks similar): https://paste.pics/4SXS5

So I redid the testing etc acording to the embl protocol from huber lab and now to my main question: new p-value histogram is more what one would expect https://paste.pics/4SXU4 However out of about 550 tested miRNAs I get 68 significantly differentially expressed miRNAs for 1 dataset T1vC and about 120 significant miRNAs for T2vC. And this number just seems way too high, no? I would never expect 1/5 of tested miRNAs to be significantly differentially expressed, or would you?

Which results should & can I trust more?

I appreciate every help, as I am not a statistican...

Best regards!

edit: Thanks for the input. Added 'fdrtool' as a tag

This is a good question for the fdrtool author Bernd Klaus. Can you add this package as a tag?

Hi,

Although I am not an expert in the use of fdrtool, normally I check I few things to understand the results of differential expression analysis.

Just to add some information to your question, How many replicates do you have in each group?

Did you plot the new genes are significant to know whether there is an outlier in your groups (an example explained in your link as the cause of the uphill pvalue distribution)? I would inspect a bunch of them, and see if you catch up some patterns.

Did you see whether the fold change are big or small for the significant genes?

As for the number of miRNA deregulated, it really depends on the biology involved. I had similar results for fibrosis study where the cells change a lot. So, I wouldn't discard that number because it is big. Many miRNA are involved in similar biological processes, and they are regulated in clusters. This doesn't mean that suddenly things make sense, but I think looking at the specific miRNA expression to catch up some obvious pattern, the fold change and the biological context should help to decide whether something is off or not.

Thanks for the question, it is very interesting for me.

Cheers

hi, i'd say that such a right-skewed distribution is often a product of some confounding factor that is not being adjusted. You can find a discussion on this in this old thread from the former Bioconductor mailing list.

there has been also a more recent discussion about hill-shaped histograms and the hint that Mike was referring to about fdrtool, in this other thread.

cheers, robert.