Question about de novo multi-species Nocardiopsis database

[ERBringHorvath]

I was hoping to ask a question about a draft Nocardiopsis spp. model I've built using PopPUNK; I work in the Natural Products space, primarily with phylum Actinomycetota—a group of bacteria that can be incredibly genetically diverse even at the species level. I am working with a novel Nocardiopsis species and was using PopPUNK to try to determine its taxonomic position relative to Nocardiopsis genomes available via NCBI. Because my strain is a new species (confirmed chemotaxonomically), I need a tool that would be appropriate for mixed-species taxonomic characterization. I started with PopPUNK because the 2019 publication highlighted its utility in classifying multi-species cohorts of bacterial pathogens, and also because I'd like to use core + accessory genomes to build a phylogeny. My question is this—is PopPUNK appropriate for a diverse genus like Nocardiopsis, given the following results?

I've attached all files generated from the database creation here as a .tar.gz file.

Reference Nocardiopsis genomes were downloaded from NCBI using Datasets. 157 genomes (either complete or draft) were used.

The following workflow was used to generate this first database:
Sketch:
poppunk --create-db --output nocardiopsis_ref --r-files ref.txt --min-k 13 --max-k 29

QC database:
poppunk --qc-db --ref-db nocardiopsis_ref --qc-keep

This resulted in 128 files failing QC for various reasons, mostly for failing distance.

Fit model:
poppunk --fit-model bgmm --ref-db nocardiopsis_ref --output nocardiopsis_ref

Fit summary:
Avg. entropy of assignment 0.0160
Number of components used 2

Scaled component means:
[0.69171908 0.07866528]
[0.36643846 0.3914161 ]

Network summary:
Components 1
Density 0.1921
Transitivity 0.5733
Mean betweenness 0.1924
Weighted-mean betweenness 0.1924
Score 0.4631
Score (w/ betweenness) 0.3740
Score (w/ weighted-betweenness) 0.3740
Removing 136 sequences

Thanks!

nocardiopsis_ref.tar.gz

[johnlees]

Some thoughts:

• Generally PopPUNK works with within-species datasets, rather than across a genus. Some of the defaults may not work well at this larger range.
• The QC seems to be removing a lot of your sequences, and I would review that the criteria being applied are appropriate here too.
• The distances are the most important thing for you to look at. Presently, they don't look like what you'd expect.
• I'd suggest reviewing the k-mer range, and looking at the k-mer regression plots to ensure you are getting a good fit. See https://poppunk.bacpop.org/sketching.html#choosing-the-right-k-mer-lengths
• An alternative starting point could be to use sketchlib.rust to find ANI between your samples and try clustering at 95%.

[ERBringHorvath]

Thanks for the advice, I'll looking into modifying both my QC criteria and investigating a more optimal kmer range. I did run Kchooser4 from the kSNP4 package on my cohort, which reported an optimal kmer length of 23; however, even using that parameter for the kSNP4 pipeline resulted in a tree with low confidence. As far as ANI is concerned, I did run fastANI on my dataset, however there are no strains that meet the ANI threshold when compared with my strain–all other strains of importance exhibit ANI values <90%, which encapsulates my core problem, I think.