Hi!

I am using package org.Mm.eg.db (org.Mm.egENSEMBL) to annotate the differential expressed genes (tophat(mm9)-featureCounts-RUVseq-edgeR). Although the results are nice some genes could not well being annotated and I guess it is is because I am using a GTF (NCBI37 mouse annotation) a bit different from that one used by org.Mm.egENSEMBL (both NCBI and ENSEMBL).  Do you have suggestions of which GTF should I better use? 

I am really interested in isoform transcript analysis, any tool suggestion? We have a batch effect, so I might be forced to use RUVseq tool or something similar before the analysis. I have tried cuffdiff but... not successful results to be honest. 

Thanks a lot for your help! 

Sergio

Hi Sergio,

I'll tell you what we do. We would use subjunc (a variant of subread) to realign to mm10, then use featureCounts to count reads by exon using the Rsubread package's built-in mouse annotation, then test for differential exon usage using edgeR's diffSpliceDGE() function or limma's diffSplice() function. Both diffsplice() and diffSpliceDGE() handle batch effects and complex linear models. diffSplice() is faster (very fast) but diffSpliceDGE() is more sensitive.

This pipeline is not the same as transcript level analysis of DE, but it does a good job of detecting differential splicing. It can all be done from the R prompt.

Hi Sergio,

I think your problem is that the org.Mm.eg.db is based on a newer version of the mouse annotation (mm10), while you're aligning to mm9. You have mainly two options: 1) map onto the mm10 genome, and use the latest Ensembl GTF (from here: ensembl.org/Mus_musculus/Info/Index) or 2) get an older annotation for your genes with the biomaRt package. See Section 5 of the biomaRt vignette (Using archived versions of Ensembl). If I remember correctly the latest Ensembl version that uses mm9 is 67.

As for the isoform analysis question: there are a few packages in Bioconductor that perform transcript-level differential expression, e.g., BitSeq and EBSeq. There might be others, but these are the two that come to mind.

I don't think there is a straightforward way to integrate the RUV factors in these models, so perhaps if your batch effects are strong, you will be better off using these methods just for transcript quantification and then run DE with limma on the log counts (perhaps using voom weights) and adding the RUV factors there.

Thanks a lot for your suggestions! I will try!!