Question: Contig kmer counts instead of mapping??

[shiraz-shah]

What are your thoughts on profiling contigs without Burrows-Wheeler style mapping? And is it possible VAMB could work with such alternative sources of counts? I ask because mapping is the main bottleneck currently.

If reads were simply split into long enough kmers, like the sizes typically used for assembly, they should yield enough specificity for unique mapping. Right? Or no? Although sequencing errors would interfere with string searches, they are rare, especially if reads are qc'd.

Default bwa mem is super unspecific for metagenomic profiling, and should be filtered with something like msamtools, which removes half of the read mappings, before generating e.g. an OTU table.

VAMB is magically able to see past bwa's noise and still define biologically meaningful clusters. Should it not be able to do the same with kmer counts?

[jakobnissen]

Possibly yes. Actually we don't really understand how co-abundance works in the details and I have lots of unanswered questions.
One recent experiment we did was to use strobealign --aemb to compute abundances per sample. This option is designed for binning ("aemb" is short for "abundance estimation for metagenomic binning"), and it works by skipping the "extend" part of the seed-chain-extend approach of modern mappers. That should be roughly equivalent to some kind of kmer counting, except it's more accurate (though less accurate that actual alignment) It's much faster than bwa mem.
Anyway, our experiments on the simulated CAMI dataset, and also experiments from other groups, have shown that strobealign --aemb gives just as good results as mapping for metagenomic binning.

I would give strobealign --aemb a try and see if you see a significant speed boost. If not, I would be interested in investigating kmer counting as a proxy for abundance.

[shiraz-shah]

That's a great idea, Jakob. I'm going to try strobealign then!