FPKM values differ for sample sample between comparisons?

[beeman]

Hi I have three biological samples (lets call them Sample_A-C)that I want to find DEGs between. Firstly, I'm performing pairwise comparisons between samples using cuffdiff using alignments generated using tophat.

The problem is when I compare the fpkm value of genes in Sample_A (in the gene_exp.diff file), the values differ depending on whether the comparison of Sample_A was against Sample_B or Sample_C.

How is this difference generated? I thought when running cuffdiff the fpkm value would be static and would originate from the mapping data in the bam/sam file used as input for cuffdiff.

Here is just the header and the first gene in the gene_exp.diff file.
Sample_A vs Sample_B
test_id gene_id gene locus sample_1 sample_2 status value_1 value_2 log2(fold_change) test_stat p_value q_value significant
ACYPI000001 ACYPI000001 acyp2eg0000191 chr1:2365250-2374181 q1 q2 OK 21.6621 25.2228 0.219556 0.389799 0.71015 0.991694 no
Sample_A vs Sample_C
test_id gene_id gene locus sample_1 sample_2 status value_1 value_2 log2(fold_change) test_stat p_value q_value significant
ACYPI000001 ACYPI000001 acyp2eg0000191 chr1:2365250-2374181 q1 q2 OK 26.0189 587.227 4.49629 4.16814 0.0011 0.422675 no

So as you can see, when compared against Sample_B, the fpkm value of ACYPI000001 in Sample_A is 21.6621 and in the second comparison it's 26.0189

There are even more pronounced differences in these datasets and I just don't get why these differences are arising.

Can someone please shed some light on what may be underlying this?

Thanks

EDIT: I'm using cufflinks version 2.1.1
cmd# cuffdiff -p 5 -o sample_A_vs_sample_B annotations.gff sample_A_accepted_hits.bam sample_B_accepted_hits.bam

[yueluo]

You might want to read about 'Library Normalization Methods' on the cufflinks manual:

http://cufflinks.cbcb.umd.edu/manual.html#library_norm_meth

FPKMs and fragment counts are scaled via the median of the geometric means of fragment counts across all libraries, as described in Anders and Huber (Genome Biology, 2010). This policy identical to the one used by DESeq. (default for Cuffdiff)

[beeman]

Great thanks for the info

[bhennuy]

Hello,
Did you try to compare the three conditions in the same process with cuffdiff ?
As explained, Cuffdiff should support that with "cuffdiff sample1 sample2 sample3". The 3 conditions should be compared instead of doing several pairwise comparison. In my case, it doesn't work ! Do you know if some options should be used to do that.
Thanks