Are full length 16S bespoke taxonomy references ok to use on specific regions?

I’m just wondering if full-length weighted taxonomy references made using clawback (as found in the readytowear collections ex. are ok to use with targeted regions like V3-V4. I know with non-weighted classifiers they performs fine enough, perhaps a little bit worse than if they were trained on the specific region, but I’m wondering if this has been tested with bespoke classifiers? Safe to use?

Additionally, would there be any downstream issues if the reference databases used to train the bespoke classifier and insertion trees don’t match? For example, using full length Silva 132 to train a bespoke classifier but using gg_13_8 to build an insertion tree.

Preemptively tagging @BenKaehler, @Nicholas_Bokulich.



Good questions, @Mehrbod_Estaki!

Yes. I am still benchmarking this but, anecdotally speaking I would say these are safe to use and in my hands I get very reasonable results.

Really the quality of these weights will be related to some degree to how well full-length sequences perform for uniform classification (since that is how these weights are being built off of the source data) and, as you say:

Even though the different amplicon targets will have their own biases (and hence true species frequencies may be slightly different), these weights will almost certainly still outperform uniform weights, because we find that even weights from related sample types outperform uniform and the accuracy improvement yielded by bespoke is correlated with the fitness of the weights (see the clawback paper for more details).

So we actually added the full-length 16S weights recently for this very use case (implying our approval of this methodology): weights may be assembled from taxonomic frequencies observed using one subunit of 16S, and then used to classify another 16S domain. Even with primer bias factored in, the weights should still be quite close to the true taxonomic frequencies of the target region, and much better than assuming uniform weights!

As always, though, take a look at your results and decide whether the classifications you get make sense. Bespoke classification quality will be tied to the quality of the source data, so the method will improve over time but will be limited by contemporary limitations in the reference databases (e.g., misannotated sequences) and source data (e.g., misannotated samples, contaminants). Don’t like the result? Build better weights! Then share them on readytowear so others can try them on.

Not at all, to my knowledge. Feature classifier is just providing more feature metadata (no matter what reference, weights, classification method, etc you use) and is not touching the feature IDs. Fragment insertion does not do anything with that feature metadata, and is also operating on the sequences themselves. As long as neither alters the feature IDs (they don’t) then they should not interfere with each other.


Just a minor addendum to @Nicholas_Bokulich’s excellent response.

Currently, if you wish to use the full length weights, you have to train your classifier using the full length reference sequences and taxonomy.

We are considering changing this in a future release (to make it possible to train a classifier on a targeted region using full length weights), but the current system does not allow that.


Thank you both for the thorough responses and this awesome plugin!


A post was split to a new topic: training classifiers: performance of full-length vs. extract-reads