cuffdiff vs edgeR

[Balat]

Hi,
I have used the new version of cuffdiff with biological replicates. After identifying differentially expressed genes, I did gene ontology based enrichment trests. I did not find any significant enriched categories. Then I used Tophat with all my libraries to get one bam file and used this in cufflinks to get one gtf file. I did this as we don't have a good annotations for the organism that I am working on. Then I used Tophat separately with each library and got bam files for each library. I used the bam file and the gtf file with bedtools and got read counts for each of library separately. I used these read count in edgeR to test differential gene expression. Gene ontology tests with differentially expressed gene showed more than 100 significantly cotegories and the up and downregulated gene categories are all relevant to the treatment I used in my experiment.

To directly compare the results from cuffdiff, I obtained read counts based on the coordinates from gene exp file from cuffdiff. I used these read counts in edgeR and used the DE genes in gene ontology enrichemnt tests. I obtained several significant gene categories and all of the enriched categories are relevant for the treatment of my experiment.

I am just wondering has anyone seen such discrepancy between cuffdiff and other read count based gene expression tests?

[Cole Trapnell]

Can you be a bit more detailed? How many genes were reported significantly different by cuffdiff? How did you run cuffdiff?

[Balat]

Total 20700 genes were were tested with cuffdiff and 5000 genes were found to be significant between two treatments. I ran cuffdiff using gtf file obtained from cuffcompare. I had 3 replicates for two of my treatments. I used bam files from 3 replicates from each treatment together as comma separated files in cuffdiff. I extracted read counts from the gene exp file using gene coordinates with bedtools. Total 15500 genes were used for extracting read counts. As cuffdiff estimates several transcripts from same gene coordinates, the number of genes with read counts were less when I used gene coordinates for extracting gene counts. I used these read counts with edgeR.

I used top most 2000 significant genes from cuffdiff in testing for gene ontology enrichment. Similarly I used top 2000 significant genes from edgeR for testing gene category enrichment.