Training classifier after trimming primers

[colinbrislawn (Colin J Brislawn)]

Hello @xchromosome,

I think the central question is about trimming primers, so let's start there.

@xchromosome

I have come across several suggestions that it's good practice to trim the primer regions from the sequences, either at the Deblur/DADA2 step, or using cutadapt.

It's essential to remove barcodes, otherwise the same feature from different samples would be split into different ASVs due to the barcode.

It's not essential to remove the region targeted by the primer, as it's just a conserved region around the variable region, so it should not matter much

It can be unnescessary to remove the primer it you use a sequencing chemistry that starts the sequencing-by-synthesis using the same primer used to amplify the amplicon, instead of the generic Illumina adapter. In this case, R1 and R2 start after the primer so there's no primer region there to remove.

You can align your ASVs against a full length 16S gene to see if the region being primed it showing up in your reads, or not!

@xchromosome

I'm now at the taxonomy stage and need to train, or pick, a classifier. I see that weighted classifiers are considered superior so I want to use that. I see there is an animal gut one available on Zenodo. Cool. My question is two-fold.

Cool

@xchromosome

• If I train my own classifier, will it matter that I've already removed the sequence-specific primer regions from my reads?

It shouldn't. I don't think those primer regions are included in the sklearn pre-trained databases, and it won't matter to q2-vsearch due to how vsearch calculates alignment scores.

@xchromosome

1. If I use a pre-trained one, does it matter whether the pre-trained classifier has been trained using the older Caporaso primer set while I'm using a very slightly different one?

Well... it shouldn't, as the old and new primers target the same region. However, I wonder if primer biases get embedded in the skbio classifier and subtly skew results.

If you do want to build your own classifier, try using RESCRIPt, the same plugin used to build the official pre-trained databases!