Hello,

Sorry for the simple question, but I could not find a straightforward answer to it.

I have RNA-seq data for three different cell lines. Each cell line is treated with three different conditions (Treatment A, Treatment B, and Control). The question I'm trying to address is "How is Treatment B different than Treatment A across the different cell lines?", and I'm using DESeq2 for the analysis.

In order to respond to this question, I believe I have to find the genes whose dysregulations are different between treatment A compared to control and treatment B compared to control. This way, genes that are dysregulated in the same direction with similar effect sizes should not be called differentially expressed and genes that are dysregulated in opposite directions or in the same direction but with differing effect sizes should be called differentially expressed. The primary problem is that I'm not sure about the appropriate design formulae and the contrast I should use for this.

I believe the design ~ Line + Treatment should be appropriate in this case, but how should I compare the condition effects of the Treatment A compared to control and Treatment B compared to control? Additionally, is it possible to use apeglm shrinkage for this comparison? or am I limited to ashr?

Any help is much Appreciated. Thanks.

Comparing each treatment to control and then comparing those differences is the same thing as just comparing treatment A vs treatment B, because control is the same set of samples. In other words, for cell line X you have treatmentA, treatmentB and control, right? And (treatmentA - control) - (treatmentB - control) reduces to treatmentA - treatmentB because control is the same set of samples for both of the original comparisons, and will therefore cancel out.

But what is different is the cell line, and the comparisons are nested within each cell line, and if you expect that there will be cell-line specific differences you would want to control for those differences because they are uninteresting in the context of the experiment, but may bias your results. This is a multi-level experiment, similar to section 3.5 in the edgeR user's guide, which you could emulate.

The usual answer would be that what you want is an interaction, and the design would be ~ Line + Treatment + Line:Treatment (same as ~ LIne * Treatment)

But I would worry about putting data from multiple cell lines together. The assumptions underlying library normalizations might be violated.

The kind of simple thing you could do is make three separate objects for each cell line, find the DE genes based on treatment for each cell line, then for each gene find the difference in fold changes between cell lines just by subtracting LFC values. If the fold change is different, for whatever reason, it means the cell line is responding differently to the treatment.