using multiple taxonomy classifiers, for comparison or in filtering workflows

Sometimes it's helpful to compare the output of two taxonomy classifiers - for example to see what the different classifiers would call the same ASV, or because they provide complementary information. Here I show how to run two different classifiers (GTDB and Silva), generate a summary visualization that allows for comparison of both, and finally uses the Silva annotations to filter organelle sequences (which are not annotated in GTDB as of release 232) from the feature table.

Data and classifiers

Step 1: Generate a summary of the ASV table

First, generate the standard ASV table summary. We'll incorporate one of the outputs in our tabular summary of the sequences.

qiime feature-table summarize \
  --i-table asv-table.qza \
  --output-dir asv-table-summary

Step 2: Classify the ASV sequences with each classifier

The classifiers can be obtained from their respective sources, which are linked from the QIIME 2 Library's Resources page.

qiime feature-classifier classify-sklearn \
  --i-classifier gtdb-r232.0-2026.7.0-classifier.qza \
  --i-reads asv-seqs.qza \
  --o-classification asv-taxonomy-gtdb-r232.qza

qiime feature-classifier classify-sklearn \
  --i-classifier SILVA_144_SSURef_NR99_uniform_classifier_V4-515f-806r.qza \
  --i-reads asv-seqs.qza \
  --o-classification asv-taxonomy-silva-144.qza

Step 3: Call tabulate-seqs[1] with the two taxonomies for a visual comparison

qiime feature-table tabulate-seqs \
  --i-data asv-seqs.qza \
  --i-taxonomy gtdb_r232:asv-taxonomy-gtdb-r232.qza silva_144:asv-taxonomy-silva-144.qza \
  --m-metadata-file asv-table-summary/feature_frequencies.qza \
  --p-merge-method intersect \
  --o-visualization asv-seqs-gtdb-silva.qzv

This is a useful tabular summary of ASV sequences. Each column starts with the ASV ID, and then provides information including the taxonomy called by each of the two classifiers as well as the total number of times each feature was observed, and the total number of samples that each feature was observed in. The latter two values come from the feature_frequencies.qza artifact that was output from the qiime feature-table summarize we ran above.

Note that the file paths for the taxonomy inputs are preceded by labels (gtdb_r232 and silva_144). When you're providing a Collection as input, as we are here, you can optional label the items in the collection by putting them before the filepath followed by a :. These labels show up in the column headers in the resulting visualization.

See the result here[1:1].

In the resulting visualization, you can for example search the Taxon: silva_144 column for f__Mitochondria or o__Chloroplast to see how the organelle sequences are labeled by the GTDB classifier.

Step 4. Filter putative organelle ASVs using the SILVA assignments

Finally, we can apply a taxonomy-based filter on the feature table to remove the organelle sequences[2]. Then, we can use that filtered table to filter the ASV sequences, which can be helpful if you want to build a phylogenetic tree from these sequences in a later step.

qiime taxa filter-table \
  --i-table asv-table.qza \
  --i-taxonomy asv-taxonomy-silva-144.qza \
  --p-exclude f__Mitochondria --p-exclude o__Chloroplast \
  --o-filtered-table asv-table-no-organelles.qza

qiime taxa filter-seqs \
  --i-sequences asv-seqs.qza \
  --i-taxonomy asv-taxonomy-silva-144.qza \
  --p-exclude f__Mitochondria --p-exclude o__Chloroplast \
  --o-filtered-sequences asv-seqs-no-organelles.qza

Downstream analysis can subsequently proceed with either taxonomic annotations, GTDB or Silva - nothing in the workflow will care which.


Thanks to @SoilRotifer and @Nicholas_Bokulich for a review and discussion of this post as I was preparing it!


  1. This was built with 2026.7.dev, as is a preview of the updated qiime feature-table tabulate-seqs visualization that will be available in QIIME 2 2026.10. ↩︎ ↩︎

  2. In SILVA 144 mitochondria are the family f__Mitochondria and chloroplasts the order o__Chloroplast.
    Keep the rank prefixes: --p-exclude chloroplast would also match the eukaryotic k__Chloroplastida (green plant 18S).
    Here this removes 454 ASVs and 165,708 reads (0.78% of the table). ↩︎

4 Likes