Tophat-Cufflinks-Cuffdiff: different libraries used as replicates

[Erinyes]

Hi all,

I was wondering whether you can offer me some advice. I have 2 sets of RNA-seq libraries (plus and minus conditions per set - activation of a signalling pathway); one performed and sequenced 2 years ago (36bp SE, 7-8mil reads), and one performed and sequenced a few months back (51bp SE, 12-17mil reads). I wanted to re-analyze them in mm10 and check the biological variation between the two sets. I run the newest versions of TopHat, Cufflinks, Cuffmerge and Cuffdiff, with Cuffdiff being run in individual libraries as well as in both sets as replicates. The oldest set gives me 60 DE genes, whereas the newest gives me none (when run as individual sets). The log2FC values between the two sets for the 60 DE genes are very similar, and the FPKM values are very similar as well. Nonetheless, they only appear significant only for the oldest library. When I run Cuffdiff for both sets as replicates, only 35 of them were recognized as significant. Both sets of libraries were quality-checked, trimmed (if appropriate), had adapters and primers removed (if appropriate), and had similar duplication levels (60% and 55% for the old and the new set respectively).

It seems counter-intuitive the smallest libraries to give DE genes, whereas the bigger ones to give none. Do you have any idea what might be the reason?

I could apply less stringent parameters on the newest set, but I am more interested in understanding the difference first. Also, after reading the Cuffdiff manual and papers a few times, I am no closer in understanding how exactly they calculate p and q values for no-replicate comparisons.

Any thought would be greatly appreciated. And even more than that.
Cheers


Parameters used:

TopHat (v 2.0.10)

tophat --read-realign-edit-dist 0 -g 2 --b2-very-sensitive -G mm10.gtf -o TopHat-set1 -M --transcriptome-index=transcriptome_data/known mm10-genome Set01_minus.fastq

Cufflinks (v 2.1.1)

cufflinks -p 4 -b mm10.fa -o Cufflinks-set1 -g mm10.gtf -u Set01_minus.sam

Cuffmerge

cuffmerge -o Set1-samples -g mm10.gtf -s mm10-genome/ assembly_list.txt

Cuffdiff

cuffdiff -p 4 -b mm10.fa -o Both-sets -L minus,plus -u -c 0 --min-reps-for-js-test 1 merged.gtf Set01_minus.sam,Set02_minus.sam Set01_plus.sam,Set02_plus.sam

or

cuffdiff -p 4 --dispersion-method=blind -b mm10.fa -o Set1 -L minus,plus -u -c 0 --min-reps-for-js-test 1 merged.gtf Set01_minus.sam Set01_plus.sam

[dpryan]

The p and q values are nearly meaningless without replicates. Since you're using mice as model organism, I'd encourage you to simply redo the entire experiment with more samples, since no conceivable method will ever give you meaningful results with your current data.

[Erinyes]

I understand that, and repeats are underway. What I do not understand, however, is how I get significant DE genes in one of the sets and not in the other with very similar log2FC and FPKM values. And the only reason I am asking is to see whether I need to alter any experimental conditions (i.e. library prep method, stimulation protocol, depth of sequencing etc). The RNA-seq library of the older set for example was generated by a 'manual' method, while for the newest set an Illumina kit was used.

Thanks for the reply though