Cuffdiff - 1vs1 or multiple testing?

[DonDolowy]

Dear all,

I am analyzing my RNA-seq data from mice that are isogenic, but yet exhibit different phenotypes. I am interesting in getting DEG and have therefore, in parallel with DESeq, analyzed my data with Cuffdiff (no gene/transcript discovery mode).

In essence, I have a control wildtype sample (WT) and then I have samples from two mice that are isogenic but different phenotypes (Pheno01 and Pheno02).

Firstly, I am interested in knowing what genes are DE between:
WT vs Pheno01
WT vs Pheno02

but more interestingly, what genes are DE between the two isogenic mice:
Pheno01 vs Pheno02

Question
My question then goes on if it would be most correct to run Cuffdiff three times, or just run Cuffdiff one time but then include Wt, Pheno01 and Pheno02 and then do the pairwise comparison using cummeRbund.
The reason for asking is that it gives completely different results regarding the amount of DEG.

The same holds true for a colleagues experiment where he has progenitor cells which he has differentiated in to five different lineages, where he wants to compare them with each other to see how different/similar they are.

Thanks a lot! I appreciate your help.

[zinky]

I encountered the same proplem, for multicondition's comparision, I first use cuffdiff to do reads counting and normalizition (with fpkm output) by --nondiff option, then extracted those expression information and test them manually(test method : anova or GLM). hope it helps!

[dietmar13]

you need biological replicates!

my opinion is that you can't (by no ways) get a meaningful DE list without biological replicates.

you never know if your DE genes are (really) differentially expressed (from your conditions) or are only fluctuating...

if you would have, let's say, at least three replicates from each type you could use programs which can deal with your design, like edgeR (w/ limma) or limma, using contrasts.

cuffdiff is always the worst choice...

[zinky]

well , I mean , I just employed cuffdiff to do reads counting as well as normalizition(fpkm , the most classic method), and of course there are replicates on my hand. DonDolowy's question is not about wether the DElist is meanful but on how to avoid three times cufdiff running , actually, you can put all your data into one command (with " " delimed for each condition ,ps: you must have replicates), cuffdiff can do comparision of each two condition for you.