Assembling RADseq reads - Clustering vs. de novo assembly

[dcard]

Being interested in non-model vertebrates, myself and colleagues typically work with organisms without a closely-related reference genome. A common way of understanding species relationships and population dynamics currently is using RADseq or similar methods, typically with 100bp SE or PE sequencing. If one cannot map their reads to a reference genome, I typically see them employ the Stacks pipeline, which uses ustacks to cluster reads together by similarity for SNP calling. Stacks performs admirably here, but the lack of standardized file formats in that pipeline is constricting, in my opinion, as those with the advantage of a reference genome would probably map their reads, produce BAM output, call SNPs, and then use many programs that are designed to work with VCF files (much greater flexibility and more resources available). A more recent alternative to Stacks is dDocent, which uses Rainbow and CD-hit to cluster reads and create "reference" contigs, which can be mapped back to in a standard mapping pipeline approach.

My question is whether a read clustering approach or a de novo assembly approach is best for creating these "reference" contigs from RADseq short reads. I've mentioned Rainbow, which seems to be the go-to way of assembling RADseq data, but I wondered if de novo assembly (and in particular, which de novo assemblers) could work also (or maybe even better). Perhaps someone has some experience with both or has the knowledge to tell me one way or the other. Some Google searching didn't yield a clear answer, so I figured I would pose the question on this forum. Thanks for the help.

[SNPsaurus]

I think Rainbow does a good job of creating the local assemblies, and uses steps that wouldn't be feasible on the full read set but are feasible for local assembly. It seems likely that it would beat any assembler built for whole-genome assembly.

These days, though, I would do a HiSeq 2x250 on a tight size selection of RAD fragments then just use a read overlap merger to combine the forward and reverse reads and end up with contigs about as long as you'd get with a good RAD-PE library.