taxonomy classfier : taxnomic weights

[Nicholas_Bokulich]

Hi @Lu_Zhang,

@Lu_Zhang

I am reading your current protocol paper for the static version 2019.10.

It's not really clear what you are referring to here. If this is important for answering your questions, could you please link to the source?

@Lu_Zhang

In my case, I have samples from shell or sand. It’s kind of different source. In that case, I couldn’t find an exactly matched pre-calculated one in readytowear/inventory.tsv at master · BenKaehler/readytowear · GitHub

You could follow this tutorial if you would like to train your own:

Using q2-clawback to assemble taxonomic weights

Note: This guide assumes you have installed QIIME 2 using one of the procedures in the install documents. This tutorial extends the q2-feature-classifier tutorial to show how you can improve taxonomic classification accuracy by assembling ecologically informed taxonomic weights for your samples using q2-clawback. This approach is described in this article: https://doi.org/10.1038/s41467-019-12669-6 We will retrain the naive Bayes classifier from the q2-feature-classifier to classify the ski…

Or determine the most appropriate EMPO-3 type from the earth microbiome project (sand would fit soil or sediment, but not sure about shell, maybe animal surface?).

@Lu_Zhang

I have noticed that all cases are tested by gg13_8, but not the SLIVA

I am not really sure what you are referring to here. As you can see in the readytowear repository you linked to, there are pre-assembled taxonomic weights compatible with SILVA, greengenes, and GTDB taxonomies. Taxonomic weights can be assembled based on any taxonomy of your choosing, following the tutorial above.

@Lu_Zhang

And in the paper, you also used the 99% but not the 97%, is it better to use 99%?

Again it is not clear what paper you are referring to here, but I recommend searching this forum for a detailed answer to this question. Yes, in general it is better to use more specific taxonomic databases, and a database clustered at 97% similarity will lose some resolution compared to 99%.

I hope that helps!