cuffdiff output different FPKM if the comparing samples are different

[rossini]

Hi,

Hope someone can kindly help me. I am confused about why the FPKM of a gene is different if different comparing samples are used!

I have four samples (say A, B, C, and D) and I run two cuffdiff processes:

cuffdiff I: cuffdiff annotation.gtf -p 8 sampleA sampleB sampleC
cuffdiff II: cuffdiff annotation.gtf -p 8 sampleA sampleB sampleC sampleD

(1)The FPKM of gene X in sampleA (and also in sampleB and sampleC) should be the same in both cuffdiff I and cuffdiff II. However, they are different.

(2) Even though for gene Y, cuffdiff I and cuffdiff II have similar log2(sampleA(Y)/sampleB(Y)), the t-statistic from cuffdiff I and cuffdiff II are quite different. cuffdiff II seem to have smaller value of the t-statistic even if the log ratio of sampleA/sampleB are similar in cuffdiff I and cuffdiff II.

The weird thing is that all the parameters are the same between cuffdiff I and cuffdiff II. The only difference is I add one more sample (sampleD).

Does this mean that the detection of differential expression largely dependent on the number of samples you provide? It seems that both the FPKM and the way they calculate t-statistic (between two samples) are influenced by the number of samples you provide, although some of the samples are not even compared (in my case, sampleC and sampleD). This would cause a big issue especially when dealing with time course data.

Thanks for your kind help.

[chenyao]

what's the version of cufflinks?

[rossini]

I used the latest version: 1.0.3 release - 6/1/2011

[wangxj]

I have the same problem.I found that different expression gene from I is more than II,and include II.
My cuffdiff version is 2.0.2.

[Krish_143]

Acoording to me:
cufdiff results will vary based on how many samples your providing, It always calculate based on the first sample. It will produce better results if one have more no of replicates. else cufdiff is not a best option when you do not have any replicates and trying to find differentally expressed genes across many samples.. that might not be good idea.

Try cufdiff only when replicates are available then it gives better result.

[Isabel Gerhardt]

Hi Rossini,
I observed the same as you. I have 3 tissues, 3 genotypes and 2 biological replicates. When I run cuffdiff with all samples and compare the number of differentially expressed genes between sample A and sample B, I have twice the number of differentially expressed genes than when I run sample A and sample B.
Did you figure out why?