Sylph overestimates percentage of unknown reads?

[pjoubert-git]

Hello, I gave sylph profile (v0.9.0) a try on samples generated from cultures that were expected to be pure, with the -u flag to estimate unknown sequence percentage. I used sylph against a database that contained the previously generated genomes of the bacteria of interest. For all samples, I got adjusted ANI scores of 100%, but for some of them I am getting sequence abundances as low as 75%. The genomes are quite complete (BUSCO completeness of 98%+). There is definitely a good number of unmapped reads in the samples (8%) but the number of unmapped reads according to bwa mem is a lot lower than the unknown reads estimated by sylph (8% vs 25%). A quick BLAST reveals that these unmapped reads most likely come from parts of the genome that weren't properly assembled; it is unlikely to be a contaminant (possible plasmids?). Is this a known issue? I understand that sylph wasn't designed for this application but I was still curious to hear the authors' thoughts on the issue and if there was any advice to improve this behavior. Thank you for your help.

[bluenote-1577]

Hi @pjoubert-git

Would it be possible to send me the reads and assemblies? I would like to take a look. Im a bit surprised at the discrepancy, but sylph should ultimately be less accurate than BWA for unclasified estimation.

[pjoubert-git]

Thanks a lot for your quick reply. I got permission to share the reads and assembly with you, I sent you an email to your gmail.

[bluenote-1577]

Hi @pjoubert-git

I'm looking into the genome and here's what I think.

1. For some reason, there are a few contigs in the genome with almost exactly 4x the coverage of the other contigs. For example: NZ_JACJMN020000003.1 is a 175kb contig with mean depth 1600x. But the majority of other contigs have 400x coverage. There's about 350kb of content that is exactly 4x multiplicity.

2. This messes up sylph's method, because it uses the median depth across all k-mers. So it does not handle heterogeneous/repetitive stuff as well.

I think this is hard to fix for sylph, since it uses k-mers and not reads to estimate unknown reads. This is a weird genome.... maybe you have a better idea why some parts of the genome are exactly duplicated four times.

I could probably add an option to tweak abundance calculation for isolate genomes to avoid this problem... but this is maybe a future feature if many people want it.

Thanks,
Jim

[pjoubert-git]

Hi Jim, thanks a lot for taking a look and for your helpful feedback. This makes a lot of sense and matches well with what I'm seeing on my side. It sounds like this isn't a common use case for sylph so I'm not sure how worth it it is to include it as a feature, but I'll keep an eye out in case you do decided to work on it. Thanks again!