Two related transcriptomes: merging but avoiding fake fusion transcripts

[danwiththeplan]

Hi. I'd like to discuss a situation that has been partially discussed in this thread:

http://seqanswers.com/forums/showthread.php?t=33917

I have RNAseq data from two cultivars of a plant which is polyploid. I've taken the approach of doing a de novo transcriptome assembly separately for each cultivar.

Examining the transcriptomes reveals that about 1/3 of the transcripts are unique to a cultivar, and 2/3 of the transcripts have a very close or identical BLAST hit in the other transcriptome.. but these hits are rarely full-length. There are actually few transcripts with a 100% full-length match in the other cultivar/genotype.

Doing differential expression analysis for different conditions in the same cultivar, no problem, I can do that with standard approaches. However, I'm not entirely sure how to do differential expression analyses across cultivars/genotype.

I know that I somehow need to produce a "combined" reference transcriptome, by one of two approaches:

(1) Simply throw all RNAseq reads from both cultivars into a new de novo assembly (which I'm doing now)

(2) Combine the two existing de novo assemblies into a new assembly using an OLC-based method like CAP3 or MIRA

To me, the main thing to avoid is assemblies containing "fake" transcripts that are half from one cultivar and half from another, and I can see approach (1) doing that a lot, because the de novo assembly breaks everything into kmers and you lose information about which full-length transcripts come from which cultivar/genotype. I am thinking that approach (2) is better for avoiding "fake" fusion transcripts since it starts from the point of long transcripts that are known to come from just one cultivar/genotype.

PS. Did I mention it's a horrible polyploid and there's no genome?

Does anyone have an opinion or similar experience?

[dongilbert]

You don't say how you did your single cultivar assemblies that were short, but if it was Trinity, then add Velvet/Oases and SoapTrans and/or TransAbyss, all of which give you more complete assemblies if your input is paired end reads. Use multi-kmers up to size of reads, as that gives more complete assembly of the high expressed genes.

See here for software that picks your best gene subset of several transcript assemblies of the same data:

http://arthropods.eugenes.org/EvidentialGene/

see about/EvidentialGene_trassembly_pipe.html for the software.

This paper is an independently done comparison of methods, with essentially same conclusions, that combining best of several assembliers, using CDS-size metrics, gives you the most complete genes:

http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0091776

The artifact gene-joins (fake fusions) are exacerbated using post-assembly mergers such as CAP and velvet/o -merge (maybe also mira, i've not tested that tho). In general the post-assembly mergers don't use all the read pair info and make more mistakes by joining things that don't belong.

You can use your cultivar mixed read set for another assembly, if above re-assembly with other assemblers doesn't help enough. The CDS-selection pipeline I've built throws out those mistakes as the CDS never spans gene joins (too many stop codons).

There are several tips here that work to improve mRNA assemblies

http://arthropods.eugenes.org/EvidentialGene/evigene/docs/perfect-mrna-assembly-2013jan.txt

For matching your 2 cultivars I suggest matching CDS also/instead as much of the assembly differences (artifacts, shortness) will be in UTRs. You may also want to measure expression differences only on CDS (or CDS +100bp)
to avoid those assembly artifacts.