Training Silva 132 classifier for qiime2-amplicon-2023.9

[SoilRotifer (Mike Robeson)]

@Mudit_Bhatia

I have tried using RESCRIPT to train SILVA 132 using the pre-clustered reads both with and without the primer trimming and have had no luck in getting good alignment.

Based on what I've been able to extract from the reference database, I'm not surprised.

@Mudit_Bhatia

Full-length classifiers which are pre-trained and are available on qiime2 website work well and show good alignment with the found ASVs with less than 0.1% ASVs being unassigned or assigned to bacteria/archaea if at all.

To make sure I understand correctly, are you saying that you are able to identify most of your 18S sequences as Eukaryotes using the full length classifier? It's just the amplicon specific one is not working well?

@Mudit_Bhatia

above protocol (with or without RESCRIPT) should ideally provide results which are coherent with each other. However, that did not look like the case.

Most of the time the differences are not substantial. But generally speaking, most report improved classification when training a classifier based on the amplicon region. However, it appears that a substantial portion of the region you need to construct an amplicon specific classifier is not completely present in the database, thus reducing the effectiveness the classifier. In these cases, we suggest using this approach, as I mentioned earlier. But it is understandable why the full-length classifier might be better in this case, as your sequences are hitting the partial region of the full length sequence, that is removed from the PCR-primer based extraction method.

@Mudit_Bhatia

So I am wondering if this is the same method being used to train the full length classifier or that takes some different steps?

Currently, I think we use the pipeline as presented, for the full length and ampicon region. Again, we provide those pre-made classifiers as a convenience, and might further optimize those classifiers in the future (some of us devs have been discussing some ideas on this). But generally, you can curate the reference database as you like. In fact, you can simply make use of the full length SSURef_NR99 as not perform any curation at all. Though I'd recommend at least dereplicating the database, to shrink the file and memory sizes for the classifier.

@Mudit_Bhatia

Also, is there a method using which a classifier ".qza" file could be converted into a ".qzv" file and could be compared?

Not currently. But you should be able to use several of the evaluate functions in RESCRIPt, to compare the taxonomy files (the ones used as input to the classifier). See the SILVA tutorial for examples.