I have an experiment with 45 samples. These 45 samples consist of 3 different surgeries, two treatments and two time periods for a total of 10 different experimental conditions with 4-5 experimental replicates for each. 

I'm using DESeq2 to analyze for DE of genes from RNAcounts.

There are really multiple comparisons we are interested in

• Surgeries versus Sham at two time points
• Surgery 1 over sham vs. Surgery2 vs. Sham
• Treatment vs control (treatment is different from surgery)
• Surgery differences in comparison to treatment vs. control

My question revolves around best practices. I originally ran DESeq2 on (treatment) which is a concatenation of two treatments, 3 surgeries and two time periods (this is 10 groups of 4-5 replicates each, because time period 2 is missing one surgery).  design ~ treatment

Then I have been using contrast statements to examine individual comparisons such as the following:

res.d1.ozone <-results(dds,contrast=c("treatment","Day1OzoneSHAM","Day1AirSHAM"),parallel = TRUE)

This compares time period one, sham surgery, Ozone to Air. etc...

My question: Should I be running a separate dds (and normalization etc...) followed by: 

dds <- DESeq(dds)
res <- results(dds)

on subsets of data (in this case 4reps of each) or am I better off normalizing all day in a single dds as I have done on all 45 samples and using contrast statements as above to evaluate?

There is a FAQ in the vignette about this question:

vignette("DESeq2")

Let me know if that helps answer your question.

In general, the best approach is to include all the samples in a single model and then pick out specific comparisons from the full model using contrasts. The only reason you might consider fitting multiple separate models on subsets of the data would be if you believed that there were substantial differences in the dispersions of each treatment group.. Although in that case I would rather recommend using voomWithQualityWeights from the limma package, which can accommodate samples/groups with different variances in a single model.