Use skani for genome clustering

[SilasK]

Skani

The tool Skani claims to be better and faster than the combination of mash + FastANI as used by dRep
I implemented the skin for species clustering.
We now do the species clustering in the atlas run binning step.
So you get information about the number of dereplicated species in the binning report. This allows you to run different binners before choosing the one to use for the genome annotation.
Also, the file storage was improved all important files are in Binning/{binner}/

My custom species clustering does the following steps:

1. Pre-cluster genomes with single-linkage at 92.5 ANI.
2. Re-calibrate checkm2 results.

• If a minority of genomes from a pre-cluster use a different translation table they are removed
• If some genomes of a pre-cluster don't use the specialed completeness model we re-calibrate completeness to the minimum value.
  This ensures that not a bad genome evaluated on the general model is preferred over a better genome evaluated on the specific model.
  See also https://silask.github.io/post/better_genomes/ Section 2.
• Drop genomes that don't correspond to the filter criteria after re-calibration

3. Cluster genomes with ANI threshold default 95%
4. Select the best genome as representative based on the Quality score Completeness - 5x Contamination

New Contributors

• @jotech made their first contribution in Improvement of offline cluster mode #667

Full Changelog: v2.16.3...v2.17.0

---

This discussion was created from the release Use skani for genome clustering.

[SilasK]

Copy of discussion from #672

Do you have some recommendations about the alignment fraction?
I think FastANI takes the average alignment fraction.

My intuition was to use the max alignment fraction as a small genome fragment aligned to a large genome with high ANI would indicate that the two should be clustered together. Anyway, the larger genome would be chosen as the cluster representative.

Maybe even better would be to remove the smaller genome if the difference between the two genomes is too large...

FastANI appears to output only the alignment fraction for the query when doing something like fastANI -q gn1.fa -r gn2.fa and doesn't do any averaging (see output format in https://github.com/ParBLiSS/FastANI#an-example-run).

I remember reading about this alignment fraction problem in dRep: https://drep.readthedocs.io/en/latest/choosing_parameters.html#minimum-alignment-coverage where they just set the ANI to 0 if the aligned fraction is < 10% (by default, I think). But because there are 2 aligned fractions, I'm not sure which one they use.

I'm not sure your exact use case, but the maximum aligned fraction may become an issue if it's a mobile element, where this mobile element can have high ANI and high alignment fraction to two distant genomes, but these two distant genomes should not be clustered together. Something to think about...