I am currently conducting an RNA-Seq experiment in which i have several samples (1 normal sample, 2 cancer samples and 2 cancer + virus samples). I have used cuffdiff to produce datasets for each of these samples vs the normal sample and use these results to check for differences in gene expression between the cancer and the cancer + virus samples. In order to ensure that normalization would make comparisons of FPKM between these different runs valid would i need to run cuffdiff with a cuffmerge file that has merged the assemblies from each different sample (i.e cuffmerge the normal, cancer and cancer + virus assemblies) or just the samples in the run (i.e cancer sample vs normal sample). As the geometric normalization method which cuffdiff uses scales based on library size i am guessing that this scaling would be different when all the assemblies are used in the same cuffdiff run, resulting in FPKM values which are not comparable between these runs.
I hope this makes some sense.