which package should be selected for differential gene expression analysis

Question

[lkianmehr]

Hello,

I need your help to choose proper software for doing differential gene expression analysis of these data. they are transcript quantification from Salmon, then import them to tximport to got gene and transcript counts from RNA-seqs. Actually, I have different groups and I want to do DGE for each one. I would describe conditions and samples.

In the first group, I have 8 samples named: D1 , D3, R1, and R3. each one of them has two technical replicates. for example D1 (D1_L001 and D1_L002); D3 (D3_L001 and D3_L002), R1 (R1_L001 and R1_L002) and R3 (R3_L001 and R3_L002). on the other hand, D1 and D3 are biological replicates as well as R1 and R3. I'd like to make a comparison between total of them I mean all D vs R, and D1 vs R1, and D3 vs R3.

what is your suggestion for these samples of first group to do DGE?

I did DGE by DESeq2, but its developer would not recommend DESeq2 for such conditions that compare individually and combining samples. please help me to choose best statistical software to do differential gene expression analysis for such conditions?

Thanks in advance

[ATpoint]

I did DGE by DESeq2, but its developer would not recommend DESeq2 for such conditions

Do you have a reference for that, and (given that you asked this question yourself) didn't Michael Love give a recommendation for your setup?

[lkianmehr]

he said that This isn’t possible with our software. It requires replicates to run DESeq(). You can compute vst() and look at the fold changes though. you want to make comparisons of individual samples, sometimes by combining other samples. I would not recommend this. It sounds arbitrary and you’re likely to trick yourself into a false positive result. now I got mixed up, I can't trust the results!

Answer 1

[Charles Warden]

In general, I would recommend that you test a few different methods for every project (such as DESeq2 / limma-voom / edgeR), and see what seems to work best with your data (and/or if you get similar results).

As for the 1-versus-1 comparison, I would strongly recommending having biological replicates as much as possible. However, I think I am confused about your experimental design. Are you trying to compare D versus R, considering pairing of 3 pairs of samples?

For all the 3 packages above, you can perform a multivariate test (comparing D versus R status, adjusting for variability between pairs). From what I understand, I think that is what you want to do (which would make use of all your samples with a single comparison).

[lkianmehr]

thank you for comment, yes I tried to compare D ( plus D1 and D3) versus R ( plus R1 and R3).

Answer 2

[Devon Ryan]

There exist no packages that perform useful 1 vs. 1 comparisons because it's then statistically impossible to estimate any biological variance. You must rethink the question you want to ask (in short, group-level differences can be used for predictions, individual-level differences generally can't).