Hi,

I am using DEseq2 to analyse RNAsequencing data in a timecourse setup. I am however unsure about setting up the correct experimental design. In particular, I struggle to decide when to combine factors and use the group to extract results and when to use interaction terms. I have tried to find answers by reading extensively on experimental design, however it only made me less confident that I have the correct experimental design for my analyses. I would be grateful for advice on how to properly setup the experimental design in the following case:

I am interested in differences of a particular gene knockout in comparison to wild type controls at different timepoints after injury. For this I have generated RNA sequencing data in biological triplicates. At the moment, I am combining factors running design = ~ condition such as I would have for the gene knockout: KO_timepoint1, KO_timepoint_2, KO_timepoint_3 and equivalently for the wild type control: wt_timepoint1, wt_timepoint_2, wt_timepoint_3

With this I think I can get all kinds of needed results by using contrasts e.g.

res=results(dds,contrast = c("condition","KO_timepoint1","wt_timepoint1")))

or

res=results(dds,contrast = c("condition","KO_timepoint2","KO_timepoint1")))

and so on. However, I have read in posts and vignette that an experimental design such as ~ genotype+time+genotype:time is used. Sometimes also using LRT. I am now unsure if I need to change my experimental design.

I would be grateful for advice on which strategy is most suitable in my case and how the two different designs would influence results. Thanks a lot in advance!

Best, Katja

The one thing I've learned that's not obvious when you're first learning about design matrices is that two different design matrices can represent the same design. When using the likelihood ratio test and similar tests, combining genotype and time into a single "group" variable is equivalent to using ~genotype + time + genotype:time.  Both designs will give you the exact same (un-shrunken) log-fold changes and exact same p-values for the same tests, so it doesn't matter which one you use as long as you form the correct contrasts in each case. We call these different ways of parametrizing the same design. You can also think of them as the same solution space but with 2 different coordinate systems.

However, when log-fold change shrinkage comes in to the picture, these designs may no longer be equivalent, because the direction of the shrinkage can depend on the choice of parametrization. I don't have much experience with the implementation of log-fold change shrinkage in DESeq2, so you should wait for Mike Love's comments on that subject.