I have an experimental design with 3 main factor effects. (data = mapped rnaseq counts (raw))

1. Days of treatment (levels = 1 and 5)
2. Diet (levels of BD, HFD, and MCD)
3. Exposure (levels of Vehicle and PERC)

I created a full model where condition.group= concatenation of each of these:

dds<-DESeqDataSetFromHTSeqCount(sampleTable = sampleTable, directory = input_dir, design=~condition.Group)

Question 1) We are not super interested in interaction effects between all 3 main effects. I went ahead and ran a full model as listed above instead of splitting day1 and day2.  My understanding from a prior post is that this is the recommendation unless one expects the variances to be significantly different from each other on days of treatment =1 vs. days of treatment =2.  Is this still accurate?

Question 2) Main effect contrasts are pretty straight-forward. For example, I used the following to compare the difference between HFD and BD when exposure = Vehicle on Day1  :

d1.hfd.veh <-as.data.frame(
  results(dds, contrast=c("condition.Group","HFD+Vehicle+1","BD+Vehicle+1"),
​          parallel = TRUE)
)

 My question regards a more complex contrast.  I want to compare effect of (PERC vs. Vehicle on HFD on Day1) to (PERC vs. Vehicle on BD on Day1).  This is the contrast I wrote after reading this post DESeq2: with multiple factors and interaction terms won't show all effects and I wanted to make sure my understanding is correct.

d1.perc.hfd.vs.bd <-as.data.frame(
results(dds, contrast=
  list(c("condition.GroupHFD.PERC.1","condition.GroupHFD.Vehicle.1"),
  c("condition.GroupBD.PERC.1","condition.GroupBD.Vehicle.1")),
  listValues=c(1/2,-1/2),
  parallel = TRUE)
);

Question 3:) In the past, to examine PCA plots, etc..I have take normalized counts from the dds object with normcounts<-counts(dds,normalized=TRUE) and then log transformed them to use in PCA plots. Is using plotPCA on rlog(dds, blind=FALSE) better, similar, or different?

Much appreciated.

john

The best way to set up the model is to ask yourself if it makes sense that there would be interactions between these variables. It sounds like you may want to have an interaction between Exposure and Diet, to test for differences of differences as you've stated in Question 2. And then you could have Days in the model as an added factor, but not interacting with the other two. Does this sound right?

Question 3: If you want a simple log plus pseudocount transformation, you can use normTransform() which produces a DESeqTransform object that you can run plotPCA() on. The idea behind rlog() and vst() are statistical approaches to transforming the data such that the variance is relatively stable across the mean as opposed to what can often be arbitrary decisions about cutoffs and pseudocounts. You can read the DESeq2 paper which talks about this, or look in the DESeq2 vignette section on transformations, and in particular the plots of the SD over the mean for the various transformations.