Our RNA-Seq design includes two variables: time (three points) and treatment (treatment vs. control). There are a total of 6 samples per group which have been balanced and randomized in two different sequencing runs. We want to compare the treatment groups vs. the control group for each time point (a total of 3 comparisons). We have considered all the samples for the TMM normalization and then we have performed the comparisons in which we are interested in.

Would it be correct to consider in the normalization step only the samples that we want to compare? I mean, to do 3 independent normalization, one per comparison? 

Thanks a lot!

We almost always do the normalization on all the samples at once. There's nothing particular to be gained by doing normalizations on subsets.

The same goes for the DE analysis. We would usually analyse all the samples together in one big linear model.