The transcripts.gtf file is one of the inputs to cuffdiff. This file could be either:

• the reference transcriptome from, say, iGenomes
• the merged predicted gtf file resulting from a cuffmerge (which in turn takes input from cufflinks)

The first approach can avoid using cufflinks/cuffmerge entirely. I understand that both of these are valid options?

My question is, when following a reference-based pipeline (as opposed to reference-guided or de novo), how does one decide whether it is appropriate to follow a tophat->cuffdiff approach rather than a tophat->cufflinks->cuffmerge->cuffdiff approach? What are the advantages of the former method (besides being less time-consuming), and in particular, what are the disadvantages?