Building a Classifier

Building a classifier

You need a QIIME2 classifier to run this pipeline. If you update your QIIME2 version, you will generally need to recompile the classifier.

From UNITE

Preferred method: the rescript plugin, included in the QIIME2 distribution, downloads UNITE directly from its PlutoF REST API -- no manual download needed, and it's the way UNITE itself now recommends fetching this data for QIIME2 (https://unite.ut.ee/repository.php now gates its files behind DOI landing pages with no stable direct-download URL, which is why the URL-based auto-download this pipeline used before generally no longer works):

conda activate rachis-qiime2-2026.7

qiime rescript get-unite-data \
    --p-version 2025-02-19 \
    --p-taxon-group fungi \
    --p-cluster-id 99 \
    --p-no-singletons \
    --o-sequences unite-sequences.qza \
    --o-taxonomy unite-taxonomy.qza

qiime feature-classifier fit-classifier-naive-bayes \
    --i-reference-reads unite-sequences.qza \
    --i-reference-taxonomy unite-taxonomy.qza \
    --o-classifier unite-classifier.qza

Run qiime rescript get-unite-data --help for the current list of available --p-version values.

Fallback: if you already have a UNITE QIIME-release archive (.tgz) from another source (a manually downloaded copy, a colleague, an institutional mirror), qiime2-its-train-unite can extract, import, and train from it directly:

qiime2-its-train-unite \
    -u ~/Downloads/sh_qiime_release_19.02.2025.tgz \
    -o /db/UNITE \
    -q rachis-qiime2-2026.7

-u also accepts a URL, but only if it resolves directly to the archive file -- most current UNITE download links do not.

From Eukaryome

Eukaryome is a broader alternative/complement to UNITE, also fetchable directly via rescript -- but unlike UNITE it isn't fungi-only: it covers ITS across all eukaryotes (animals, plants, protists included), which can actually help ITS metabarcoding by letting a classifier flag non-fungal amplification instead of forcing everything into a fungal lineage. Confirmed by a real download (Eukaryome 2.0's --p-rrna-gene ITS): ~1.6 million sequences, ~200 MB -- an order of magnitude more than UNITE, so training takes correspondingly longer and more RAM.

qiime rescript get-eukaryome-data \
    --p-rrna-gene ITS \
    --output-dir eukaryome_its

qiime feature-classifier fit-classifier-naive-bayes \
    --i-reference-reads eukaryome_its/eukaryome_sequences/ITS_seqs.qza \
    --i-reference-taxonomy eukaryome_its/eukaryome_taxonomy/ITS_taxa.qza \
    --o-classifier eukaryome-classifier.qza

--p-rrna-gene is a Collection output, hence --output-dir rather than individual --o-... flags -- it writes <dir>/eukaryome_sequences/ITS_seqs.qza and <dir>/eukaryome_taxonomy/ITS_taxa.qza (naming matches whatever gene(s) you requested). If you only want fungi, filter to it before training:

qiime taxa filter-seqs \
    --i-sequences eukaryome_its/eukaryome_sequences/ITS_seqs.qza \
    --i-taxonomy eukaryome_its/eukaryome_taxonomy/ITS_taxa.qza \
    --p-include k__Fungi \
    --o-filtered-sequences eukaryome-its-fungi-seqs.qza

Then train against eukaryome-its-fungi-seqs.qza and the original (unfiltered) taxonomy artifact -- fit-classifier-naive-bayes ignores taxonomy entries that don't correspond to any reference sequence, so there's no need to filter the taxonomy artifact too.

From GenBank (NCBI query or accession list)

qiime2-its-train-ncbi downloads sequences, downloads/parses taxonomy, and trains the classifier. It accepts either an NCBI query string or a text file with one accession number per line. Test your query on the NCBI website first to make sure it returns the sequences you expect.

# Using a query string
qiime2-its-train-ncbi \
    -q "txid4762[Organism:exp] AND (\"internal transcribed spacer\"[Title]) NOT uncultured[Title]" \
    -t 48 \
    -o /oomycetes_DB \
    -e your_email@provider.org \
    -a your_ncbi_api_key_optional_but_recommended

# Using a text file of accessions
qiime2-its-train-ncbi \
    -q ~/Downloads/accession.list \
    -t 48 \
    -o /oomycetes_DB

--acc2taxid/--dead-acc2taxid let you point at already-downloaded copies of NCBI's nucl_gb.accession2taxid.gz/dead_nucl.accession2taxid.gz (each several GB) instead of downloading them again.

From your own fasta file

qiime2-its-train-fasta builds a classifier from a fasta file and a matching "accession to taxid" table (two tab-separated columns; accessions must match exactly between the fasta headers and the table, up to the first whitespace):

qiime2-its-train-fasta \
    -q my_sequences.fasta \
    -i acc2taxid_table.tsv \
    -o /path/to/output_folder/

Every sequence ID in the fasta file must be in the table; this is checked before anything is downloaded or imported, and the missing IDs are listed. Several sequences can share a taxid (the usual case: many reference sequences per species) -- each one gets its own taxonomy line.

Other databases (not for ITS)

rescript and the wider QIIME2 ecosystem also integrate several other reference databases: SILVA (qiime rescript get-silva-data), GTDB (get-gtdb-data), PR2 (get-pr2-data), and a few others. None of these cover the ITS region -- confirmed against each action's real --help output: SILVA's --p-target only offers SSURef/LSURef (16S/18S and 23S/28S rRNA), GTDB is bacterial/archaeal SSU only, PR2 is protist SSU only. This isn't a QIIME2 limitation to work around -- the ITS region is the non-conserved spacer between rRNA genes, which is exactly why these curated rRNA databases exclude it, and why UNITE/Eukaryome (both amplicon-database projects, not rRNA-alignment projects) exist as separate resources for it in the first place.

SILVA also publishes ready-to-use, pretrained .qza classifiers for common SSU/LSU regions/primer sets -- no training step at all, just download and point --classifier at it. arb-silva.de/current-release/QIIME2 is a real browsable directory tree (confirmed, not just a curated links page) laid out as <QIIME2 version>/<SSU or LSU>/<region or primer set>/<uniform or weighted>/, where "uniform" is the standard classifier and "weighted" is an experimental, habitat-adjusted variant. Confirmed with a real download, e.g. the V3-V4 (341F/806R) 16S classifier:

https://www.arb-silva.de/fileadmin/silva_databases/current/QIIME2/2026.7/SSU/V3V4-341f-806r/uniform/SILVA_144_SSURef_NR99_uniform_classifier_V3V4-341f-806r.qza

(112 MB, 200 OK). The path encodes the QIIME2 version it was built for (2026.7 here) -- match it to your own installed QIIME2 version, since a classifier's scikit-learn version has to match the one your QIIME2 environment ships (classify-sklearn will warn, and may give unreliable results, across a scikit-learn version mismatch).

That said, this pipeline isn't hardcoded to ITS: --extract-its1/--extract-its2 are optional (they just run ITSxpress before DADA2). Omit both and supply a matching classifier -- e.g. one of SILVA's pretrained SSU/LSU classifiers above -- and the rest of the pipeline (DADA2, phylogeny, diversity, taxonomy, the advanced-stats/report features) works the same way for a 16S/18S/23S/28S rRNA amplicon dataset as it does for ITS.

Next: Pipeline usage.