Cuffcompare/cuffdiff changes FPKM values of same BAM in sequential runs

[SEQond]

Dear all,

Question 1.

For an RNA Seq experiment I compare between 3 bam files (non paired-end). (Tophat- Cufflinks)

B vs A , C vs A (no replicate arrays)

In the first step : sequentially run Tophat-Cufflinks accross each of my files to create all the bam files.

In the second step: sequentially run comparisons with cuffcompare-cuffdiff.

For most XLOCs the FPKM of "setA" differs slightly from one comparison to the next as in the example below.
Why might that be? it is the same bam file in both cases. In certain rare cases the FPKM of "setA" differs considerably accross the comparisons.

Comparison B vs A

XLOC_002887 setB setA NOTEST 200,206 186,633
XLOC_003705 setB setA LOWDATA 152,669 223,705


Comparison C vs A

XLOC_002887 setC setA LOWDATA 201,762 185,595
XLOC_003705 setC setA LOWDATA 253,098 222,461


The basic workflow with options where applicable follows

Tophat -i 30 -I 20000 --segment-length 16 --segment-mismatches 1 (since my reads are 32bp long I have used half of the read length)
Cufflinks -N
Cuffcompare
Cuffdiff -N -L B,A --FDR 0.1



Question 2.

In a replay of this experiment as an exercise I used instead of "--segment-length 16" , "--segment-length 20" which is more than half the length of my reads. In this case , ALL FPMKs were similar to the 16 segment case but multiplied by 100

How can this happen?

Thanks for your input.

[zeam]

For question 1, did you use an annotation in cufflinks or cuffcompare? If so ,I thinks this might explain this:
Cuffdiff and Cufflinks now accept new options controlling whether all hits are counted towards the FPKM denominator, or only those compatible with some transcript in the reference annotation. Counting only compatible hits avoids certain types of bias that arise when one sample contains far more hits that aren't compatible with any transcript than the other sample does. For example, if one sample contains vastly more mapped ribosomal RNA hits, FPKM values will appear lower in that sample, potentially leading to false positive differential expression calls. Cuffdiff by default now uses only compatible hits. Cufflinks still uses total hits by default, as using compatible hit accounting requires a reference GTF.
So just set --compatible-hits-norm parameter to be identical when you do cufflinks and cuffdiff.

Also,I have one question, you said you have three bam files, A B and C.Did you use cuffdiff for each pair or just use cuffdiff once? If the later, did you get all cuffdiff test between all pairs of samples ? I have four tisssues, and command are as follows:
======================================
cuffdiff -o DEG_cuffdiff -p 20 cuffcmp.combined.gtf ./tophat_sd/accepted_hits.bam ./tophat_em/accepted_hits.bam tophat_en1/accepted_hits.bam ./tophat_en2/accepted_hits.bam
======================================
But I only got one cuffdiff test file between the fisrt two pairs? Follow the mannual, there should be 6 files for each pair of samples.
Thanks.

[SEQond]

Reply issue 1.
Yes I did use a GTF annotation file I had already created. I used the -G
-G/--GTF <reference_annotation.(gtf/gff)> Tells Cufflinks to use the supplied reference annotation (a GFF file) to estimate isoform expression. It will not assemble novel transcripts, and the program will ignore alignments not structurally compatible with any reference transcript.


I will check what --compatible-hits-norm returns coupled with the -G option and will get back to you.

Reply issue 2.
Related to the multiple comparisons in one--> What I did was to run cuffdiff separetely for each pair. I did try to do 4 comparisons at once at some point by putting the control as the first case in cuffdiff, but the software returned only the first comparison and not the rest 3.
I ll check again if there is a way around this.


New Question, issue 3, possibly related to issue 1.
In the way I am doing the analysis , I am getting a result of multiple NMs per XLOC

What I would prefer is a single NM per line