Hello all,

I have 16 patient rna samples who belong to 8 patients, taken before and after their treatment.

My samples table is that: 

no sample condition replicate reponse 1 1 102_screen screen P1 N 2 2 102_treated treatment P1 N 3 3 105_treated treatment P2 N 4 4 105 screen P2 N 5 5 107_treated treatment P3 N 6 6 107 screen P3 N 7 7 108_treated treatment P4 N 8 8 108 screen P4 N 9 9 110_treated treatment P5 Y 10 10 110 screen P5 Y 11 11 111_treated treatment P6 N 12 12 111 screen P6 N 13 13 115_treated treatment P7 N 14 14 115 screen P7 N 15 15 116_treated treatment P8 N 16 16 116 screen P8 N

One of them responded to the treatment and I would like to check the significance of the gene count difference between the responder and the others, but taking into consideration that the samples are paired.

I am not sure if it is correct to use this design, because I think that it considers all the samples as indipendant.

 dds <- DESeqDataSetFromTximport(txi, coldata, ~reponse)

Is there anyone who can help me sort this out please?

Firstly, you should appreciate that this is a really under-powered analysis, because it all hinges on the after vs before treatment for a single patient. So unless that patient is very strongly different in after vs before compared to the others, you won't find anything. And then, if that patient is very different, you don't really know if it's just an outlier.

That said, it is technically possible to do the analysis you're suggesting using the interaction model described in the DESeq2 vignette.

In your case, the group is response (yes vs no), and condition is after vs before treatment. See the vignette on how to specify the reference level of factors. The design will be group + group:individual + group:condition, and you are interested in contrasting "groupY.conditionAfter" with "groupN.conditionAfter". You can do this using the list() argument of results() after running DESeq(). It would look like contrast=list("groupY.conditionAfter","groupN.conditionAfter").

The individual variables should be a factor taking levels 1-7 for the non-responders and 1 for the responder. Then you will have to remove a number of columns of the model matrix for which there are all zeros (because there are no patients 2-7 in the non-responder group).

mm <- model.matrix(~ group + group:individual + group:condition, colData(dds))
mm2 <- mm[,!apply(mm,2,function(x) all(x==0))]

then

dds <- DESeq(dds, full=mm2, betaPrior=FALSE)