Quick clarification on feature classifiers

[ctekellogg (Colleen Kellogg)]

Hi there,
I have a (hopefully) quick question regarding the current feature classifiers and how they compare to the blast from QIIME v 1.9, for example. I am turning back to an old, partially analyzed 16S/18S V6-V8 dataset and want to update the classifications to use the newest version of Silva. Previously, I classified my OTUs using Blast in QIIME 1.9. I thought I'd use the new classifiers from QIIME2 when updating the classifications and noticed that each of them (vsearch, consensus-blast+ and the naive-bayes classifier) yield lower taxonomic resolution than when I classified in QIIME1. I read this post: Feature-classifier with Blast reference dataset and it leads me to believe that since they are consensus based, rather than top hit based (as the old QIIME1 version of blast was...correct?), that the newer classifiers yield a more conservative ID, thus the lower taxonomic resolution. And then, I cross-checked a few where the resolution was really low compared to my old ID by running them through the Silva classifier online (Alignment, Classification and Tree Service) and that classified them to quite a bit higher taxonomic resolution as well; so is this again a case of "top hit" vs. consensus. Just wondering what I should trust or go with for my final IDs! If anyone has any opinions or advice, I'd greatly appreciate it! Maybe I just need to adjust the settings in the new classifiers a bit to approach the IDs that I got during my last go-around in QIIME1.9. Or maybe I should just stick with this conservative IDs, as that is better....

Thank you in advance for your input!
Colleen

[Nicholas_Bokulich]

Hi @ctekellogg,
Good questions!

@ctekellogg

it leads me to believe that since they are consensus based, rather than top hit based (as the old QIIME1 version of blast was…correct?), that the newer classifiers yield a more conservative ID, thus the lower taxonomic resolution.

correct

@ctekellogg

I cross-checked a few where the resolution was really low compared to my old ID by running them through the Silva classifier online (Alignment, Classification and Tree Service) and that classified them to quite a bit higher taxonomic resolution as well; so is this again a case of “top hit” vs. consensus.

yes (to my knowledge)

@ctekellogg

Just wondering what I should trust

trust us!

@ctekellogg

Maybe I just need to adjust the settings in the new classifiers a bit to approach the IDs that I got during my last go-around in QIIME1.9.

you could set maxaccepts=1 to do that... but then it is top hit and totally inaccurate!

@ctekellogg

Or maybe I should just stick with this conservative IDs, as that is better…

if it is too conservative we can help... check out the paper for recommendations. But overall you should use the consensus classifiers to avoid dire misclassifications.

[ctekellogg (Colleen Kellogg)]

Hi @Nicholas_Bokulich - Thank you so much for your response and clarifications. I'll look back over the paper you've provided a link to as well and let you know if I have any further questions.

Colleen