We used DESeq2 a lot at our research projects. Recently I was asked about specific experimental design for DESeq2.

We have 10 subjects that we will perform surgery on and measure mRNAs before surgery, during surgery, and 3 follow up time points.

All 10 subjects were collected at the pre, during, and the first follow up.

Only 7 were collect at the 2nd follow up and 6 were collected at the 3rd follow up.

This is a question of paired or unpaired analyses. We can perform a paired analysis for the “during” and “1st” follow up to the “pre”, but we don’t have a full paired set for the 2nd and 3rd follow up time points.

What do you recommendation for comparing 2nd follow up vs 1st follow up? Should we use all 10 samples in 1st follow up and 7 samples in 2nd follow up and set the patient ids as covariance even there are three patients existed at 1st follow up only? Or we have to use the 7 paired samples only?

There was another question posted here:

A: Samples without pair in paired design experiment

There, Gordon Smith recommends to use duplicateCorrelation() with limma-voom to make use of all the samples. This makes sense to me. You can’t include all the samples with DESeq2, only the paired ones.

I wrote a comment on that post where I try to get at, when do I think it’s not necessary to include the unpaired samples: basically if the pairs are very different from each other and there are few unpaired samples, I don’t think much is gained by adding the unpaired samples. For example, if in the PCA plot, the paired samples are always very close to each other and far from others, then I think the pairing is what’s most important and the unpaired samples can’t add much. The other argument is that duplicateCorrelation and limma voom works either way so you can just use that.