Cufflinks RABT assembly

[swaraj]

Dear All,

I have run cufflinks with "-g" (RABT assembly) and "--no-faux-reads" option on a single end, un-stranded sequencing data. As reference I have provided a pooled gtf file comprising of refseq, ensembl and mRNA transcripts. I observed that cufflinks builds many transcripts on the basis of the reference which do not have any read coverage or very less read coverage.

I would like to avoid this situation. I thought using the "--no-faux-reads" solves my problem but it does not. I would like to use the RABT assembly with a reference file but want to disallow cufflinks from building transcripts for which there are no mapped reads.

[arvid]

You could simply remove the transcripts without read coverage in the end, instead.

[swaraj]

Yes thats what I did, but I still have many transcripts built from reference gtf which have like 0.1-0.5 coverage on their exons. I am not sure how to treat them as cufflinks has build them because they were present in the reference gtf.

I am looking towards identifying novel intergenic non-coding transcripts, many of which are build by cufflinks using reference intergenic unannotated mRNA from the gtf. Hence I am not sure whether to take such transcripts at all.

[arvid]

So if you look for novel transcripts, I guess you would like to filter out all known features, wouldn't you? (with e.g. BEDTools "intersectBed -v") Regardless of their coverage? I'm still a bit confused about what your actual problem is then...

[swaraj]

Yes I would like to filter out known features, but keep those mRNA features which do not overlap a protein coding gene. In this regard I get many reported transcripts which are build from reference mRNAs but do not overlap protein coding genes. For me they are putative stage specific intergenic transcripts. What I am not able to resolve is whether I should consider their presence even after a very low coverage. I suspect most of them have been built by cufflinks simply because they were present in the reference gtf file.

[arvid]

I guess that depends on what you would like to do downstream with the candidates. I'd simply remove all really-low-coverage features (something like less than 5 reads aligned, or so, or where less than 25 % is covered at all), if there are tons of them, and focus on what's left. Then I'd label them according to "previously described" or "completely novel", and carry on according to the interest of the study (typically differential expression analysis in my case).

[swaraj]

Thanks a lot. This was quite helpful.