Looking for advice on generating a transcriptome from mapped .bam or .sam files

[ovon]

Hello all. I'm starting off here with something fairly complex, I guess.

I'm looking for some advice because I'm starting a new project and I haven't done this sort of thing before.

I've got a few samples worth of RNAseq reads and I'd like to generate expression information with them. They are human, but likely to contain sequences that are not in an existing reference transcriptomes. The read depth is also not very high, so I'm hesitant to just use the genome as a reference when generating expression data.

My plan right now is to map them to the current version of the human genome using STAR, then to use the alignments generated to produce a fasta file that has consensus sequences for everything aligned.

From there I can use Sailfish or Salmon to get read counts, RPKM, etc, and compare my samples using some kind of differential expression analysis in R. This part I'm solid on, it's the middle step that I'm not so sure about.

Does anyone think that generating a reference transcriptome this way is inadvisable (and why)?

If this sounds reasonable, what do you think is the best way to do so? I see that Trinity has a genome guided transcriptome generation option. I'd like to try that out. If not, I also see that there are ways to get just the aligned portions of a .bam file in .fasta format. Seems like a relatively convoluted process, though. I'd prefer to keep things relatively simple where possible.

One last thing, I'm not sure about preserving annotations for either of these options. So I'm open to advice on how to do that no matter what route I end up going.

Thanks in advance!

[Brian Bushnell]

BBMap is, probably, substantially more sensitive for finding novel isoforms/splice-sites from low-coverage data (the only non-default flag you would need is "maxindel=200k"). I have not directly compared it to STAR, though.

As for generating a reference transcriptome, you might consider making a union of the official transcriptome and your experimental transcriptome prior to quantification. That way, transcripts that are already annotated, and present in your samples but in levels too low to assemble, will still become valid mapping targets.

I think the simplest approach would probably be to combine all your data, assemble with Trinity, then deduplicate your assembled transcriptome in conjunction with an official transcriptome using Dedupe (part of the BBMap package), which also removes contained subsequences. At that point you will have all official and novel transcripts for optimal quantification. This solution is not perfect, but eukaryotic RNA-seq is always messy.

[ovon]

Thanks Brian, I wasn't aware of bbmap's capabilities wrt this analysis. That is all very helpful and I think I will go ahead with your suggestions. Definitely sounds easier than what I was planning.