Bowtie2 or RSEM (with Bowtie2) alignment of PE reads with reference transcriptome?

Question

[deanpett]

Hello folks,

I'm working on an RNAseq experiment in a non-model species. I have a de novo transcriptome I'm using as my reference to align paired end illumina reads since we do not yet have a working genome.

I've done an initial alignment of my reads using both Bowtie2, but upon further consideration I thought that RSEM may be a better system due to its ability to more accurately give read counts for multiple-mapping reads. My alignment rate dropped significantly with the RSEM alignment, which I hypothesize is due to mainly to the fact that "Since currently RSEM does not handle indel, local and discordant alignments, the Bowtie2 parameters are set in a way to avoid those alignments".

I see a trade-off: higher alignment with Bowtie2, but randomly assign multireads (which could skew my analysis) or account for multireads with RSEM, but let my alignment suffer because indels and discordant alignments are not allowed.

QUESTIONS:

1. Is RSEM capable of properly handing a bowtie2 alignment (with "-a" option to report all alignments) as input, producing counts for each transcript which account for these multireads, while taking advantage of an excellent bowtie2 alignment? (if not)
2. Is it better to randomly assign multireads for an RNA-seq experiment using Bowtie2 alignment or more accurately quantify mutlireads with RSEM (w/ Bowtie2, excluding indels and discordant alignments)?

Thanks in advance.

Answer 1

[Devon Ryan]

1. Yes, in fact there's no point in using RSEM without giving it multiple alignments for multimappers. Note, however, that I wouldn't recommend using -a (-k is quicker) and that bowtie is more ideal than bowtie2 here, since you can additionally specify --best (you really only want the top stratum of scores). As an aside, you could instead feed the BAM file into Salmon and get similar results in a small fraction of the time. BTW, you might just use Salmon or Kallisto if this is a new project.
2. Either completely ignore multimappers or assign them with RSEM/eXpress/Salmon/etc. Note that if you map with Kallisto you can just use Sleuth, which is at least theoretically the most reasonable tool currently around for transcript-level DE.

[Rob]

Hi Devon - all great suggestions. Actually, it's now possible to use Sailfish with Sleuth (and Salmon support is imminent as well)!

[Devon Ryan]

Awesome, good to hear it!

Answer 2

[Ian]

This may not be terribly helpful, but is the question whether you trust multi-mapping reads or not. If you don't use STAR mapper. If you do then can't you use the Tophat/Bowtie route?

[deanpett]

My reference is a reference transcriptome designed to be robust for MANY spliceforms. I'm expecting a high rate of multireads due to the nature of this reference. As such, I trust multireads and I'm interested in maintaining them for differential expression analysis.