QIIME2 command-line workflow for 16S/ITS amplicon analysis. Alternative to DADA2/phyloseq R workflow with built-in provenance tracking...
Reference examples tested with: QIIME2 2026.1+ (amplicon distribution; framework now rachis), provenance-lib 2024.10+.
Before using code patterns, verify installed versions match. If versions differ:
qiime --version, qiime info, then qiime <plugin> <action> --help to confirm flagsIf code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.
The QIIME2 release tag (calendar-versioned YYYY.RELEASE, e.g. 2026.1) defines the plugin API AND the .qza artifact format; an sklearn taxonomy classifier .qza trained under one release may need retraining under another (the classifier is pinned to its scikit-learn version). The conda env name encodes the release and distribution (qiime2-amplicon-2026.1; renamed toward rachis-qiime2-<release> in 2026.4). Names and the install YAML URL are moving targets - verify the current release and distribution names against the live install page before pinning anything.
"Run my amplicon study through QIIME2" -> Move data through the framework as typed, provenance-carrying artifacts and defer every scientific choice to the owning skill - because a .qza is not a file, it is data plus its entire executable history, and that history is the deliverable.
qiime tools import, qiime <plugin> <action> --i-* --p-* --m-* --o-*, qiime tools peek/exportScope: the artifact/provenance/type machinery and the import/export/metadata/interface mechanics - this is the GLUE skill. Denoising params (trunc/trim/maxEE, DADA2 vs Deblur) -> amplicon-processing. Classifier/DB choice + training -> taxonomy-assignment. Diversity metric/sampling-depth/rarefaction + PERMANOVA-vs-dispersion -> diversity-analysis. DA tool choice/consensus -> differential-abundance. PICRUSt2 -> functional-prediction. Shotgun reads (moshpit distribution, Kraken2/MetaPhlAn/HUMAnN) -> metagenomics. This skill shows each scientific action and routes the decision out; it does not re-teach the method.
A .qza carries the data AND the complete computational graph that produced it. The semantic-type system plus the embedded provenance ARE the reproducibility guarantee - the whole reason to work inside the framework instead of passing loose BIOM/FASTA/Newick files. The cost is exact and unavoidable: there is no cat-ing the data. Three corollaries each common misuse violates:
qiime tools export writes native data and silently drops the QIIME2 wrapper AND the provenance. Export early and go ad-hoc, and the final figure has no history back to the raw reads - the framework overhead was paid and the deliverable thrown away. Export at the LAST step, or use qiime2R::qza_to_phyloseq so the chain survives as far as possible.core-metrics-phylogenetic refuses a FeatureData[Taxonomy] where a FeatureTable[Frequency] belongs, BEFORE running. A type error is the guard WORKING - fix the upstream action that made the wrong type, do not launder it by re-importing..qzv is terminal and a classifier is version-pinned. A Visualizer's output can never be another action's input (keep the .qza it was made from). An sklearn classifier .qza trained under 2024.x raises a version-mismatch under 2026.x - the training version is part of the method.Organize the analysis around protecting the provenance chain and the type contract, not around listing flags.
A .qza (QIIME Zipped Artifact) and .qzv (Visualization) are ZIP archives keyed at top level by a UUID. Every artifact carries four things:
FeatureTable[Frequency], SampleData[PairedEndSequencesWithQuality], Phylogeny[Rooted], FeatureData[Taxonomy], FeatureData[Sequence], DistanceMatrix, SampleData[AlphaDiversity]. Types can carry Properties (SampleData[AlphaDiversity] % Properties('phylogenetic')).BIOMV210DirFmt, a Newick file). Type is the meaning; format is the bytes.provenance/ subtree: for every upstream action the plugin/action name, every parameter value, input/output UUIDs, plugin + framework versions, execution environment, timestamp, and BibTeX citations. The references form a DAG of the whole analysis.qiime tools peek table.qza # UUID + Type + Format, without unzipping
qiime tools validate table.qza --level max # archive integrity + payload conforms to its format
qiime tools extract --input-path table.qza --output-path extracted/ # FULL archive incl provenance (read by hand)
qiime tools export --input-path table.qza --output-path exported/ # ONLY the native data - DROPS provenance
qiime tools extract keeps the QIIME2 structure (data + provenance/); qiime tools export is the one-way door out. A plain unzip table.qza works too (it is a standard ZIP).
| Interface / tool | Role | When |
|---|---|---|
q2cli (qiime ...) |
the command-line interface; --i-* inputs, --p-* params, --m-* metadata, --o-*/--output-dir outputs |
default, most-documented, scriptable; what tutorials/forum answers use |
Artifact API (from qiime2 import Artifact, Metadata) |
the Python 3 interface; Artifact.load/.save/.view, actions importable as functions returning Results |
notebooks, embedding QIIME2 in a larger Python pipeline (no temp files) |
| view.qiime2.org | renders any .qzv viz AND the .qza/.qzv provenance DAG client-side, NO install |
sharing results and inspecting provenance without QIIME2 installed |
provenance-lib (qiime tools replay-provenance) |
parses an artifact's provenance DAG and regenerates executable code (Keefe 2023) | recovering the commands that made an artifact; reproducing a shared .qza |
Neither q2cli nor the Artifact API is "more reproducible" - provenance is identical; pick by host environment. An Action is a Method (Artifacts in -> Artifacts out), a Visualizer (-> exactly one terminal .qzv), or a Pipeline (-> many Artifacts and/or Visualizations, e.g. core-metrics-phylogenetic). The Method/Visualizer distinction is WHY a .qzv is a dead end. The legacy q2studio desktop GUI is dead (last release 2022.8); the no-CLI answers are Galaxy + view.qiime2.org.
| Scenario | Recommended | Why |
|---|---|---|
| Demultiplexed per-sample FASTQ, filenames are Casava 1.8 | --type SampleData[PairedEndSequencesWithQuality] --input-format CasavaOneEightSingleLanePerSampleDirFmt |
sample IDs parsed from filenames; no manifest needed |
| Demultiplexed FASTQ, arbitrary paths | V2 manifest (PairedEndFastqManifestPhred33V2) |
TSV of absolute paths; the most general/explicit on-ramp |
| Still multiplexed (one big FASTQ + barcodes) | import EMPPairedEndSequences, then qiime demux emp-paired |
demultiplexing is a QIIME2 step, not the import |
| A feature table built elsewhere | --input-format BIOMV210Format --type FeatureTable[Frequency] |
BIOM v2.1 (HDF5); attach metadata separately |
| Scripting a notebook / larger Python pipeline | Artifact API | returns Artifacts directly, no temp files; same provenance |
| Need to share a result with no-QIIME2 collaborators | upload .qzv to view.qiime2.org |
renders viz + provenance client-side |
Handed a single .qza, need the commands that made it |
qiime tools replay-provenance |
regenerates executable code from the provenance DAG |
| One-off custom R analysis, fighting the framework | qiime2R::qza_to_phyloseq / export, own the provenance loss |
the overhead is not worth it; be honest about the exit point |
| Shotgun / WGS reads | -> metagenomics (moshpit distribution) | different distribution and toolchain; cross-link, do not merge |
Goal: Turn raw demultiplexed reads into a typed, provenance-rooted artifact with the correct Phred offset.
Approach: Write a V2 manifest (TSV, absolute paths), declare the semantic type and the format whose name encodes the Phred offset, then immediately summarize to confirm the reads decoded sanely.
# manifest.tsv (TAB-separated, V2; absolute paths):
# sample-id<TAB>forward-absolute-filepath<TAB>reverse-absolute-filepath
qiime tools import \
--type 'SampleData[PairedEndSequencesWithQuality]' \
--input-path manifest.tsv \
--input-format PairedEndFastqManifestPhred33V2 \
--output-path demux.qza
# Phred offset is BAKED INTO the format name: Phred33V2 (modern Illumina) vs Phred64V2 (legacy).
# V1 was CSV with a `direction` column; V2 is TSV with separate forward/reverse columns - prefer V2.
qiime demux summarize --i-data demux.qza --o-visualization demux.qzv # per-base quality (drives trunc choices)
For EMP-multiplexed data: import --type 'EMPPairedEndSequences', then qiime demux emp-paired --i-seqs emp.qza --m-barcodes-file metadata.tsv --m-barcodes-column barcode-sequence --o-per-sample-sequences demux.qza --o-error-correction-details ec.qza. The per-base quality plot in demux.qzv is read by amplicon-processing to pick truncation - not here.
The pipeline shape, with every method choice routed to its owning skill:
# Denoise -> ASV table + rep-seqs. PARAM CHOICE (trunc/trim/maxEE, DADA2 vs Deblur) -> amplicon-processing
qiime dada2 denoise-paired --i-demultiplexed-seqs demux.qza \
--p-trunc-len-f 0 --p-trunc-len-r 0 \
--o-table table.qza --o-representative-sequences rep-seqs.qza --o-denoising-stats stats.qza
qiime tools peek table.qza # confirm Type is FeatureTable[Frequency] before wiring downstream
# Taxonomy. CLASSIFIER + DB choice and training -> taxonomy-assignment
# Use a classifier .qza trained for THIS release (data.qiime2.org/<release>/common/...); old ones break.
qiime feature-classifier classify-sklearn \
--i-classifier classifier.qza --i-reads rep-seqs.qza --o-classification taxonomy.qza
# Phylogeny (Pipeline) -> rooted tree for UniFrac/Faith PD
qiime phylogeny align-to-tree-mafft-fasttree --i-sequences rep-seqs.qza \
--o-alignment aln.qza --o-masked-alignment masked-aln.qza \
--o-tree unrooted-tree.qza --o-rooted-tree rooted-tree.qza
# Diversity (Pipeline). SAMPLING DEPTH + metric + rarefy-or-not -> diversity-analysis (pick depth from alpha-rarefaction)
qiime diversity core-metrics-phylogenetic --i-phylogeny rooted-tree.qza --i-table table.qza \
--p-sampling-depth 10000 --m-metadata-file metadata.tsv --output-dir core-metrics/
# PERMANOVA via diversity beta-group-significance; the location-vs-dispersion (betadisper) confound -> diversity-analysis
# Differential abundance - MODERN q2-composition (NOT add-pseudocount+ancom). Tool choice/consensus -> differential-abundance
qiime composition ancombc --i-table table.qza --m-metadata-file metadata.tsv \
--p-formula 'group' --o-differentials ancombc.qza
qiime composition da-barplot --i-data ancombc.qza --o-visualization ancombc-barplot.qzv
core-metrics-phylogenetic and align-to-tree-mafft-fasttree are Pipelines (one call, a directory of artifacts + Emperor .qzvs out). --p-formula takes column names from the Metadata; annotate integer ID/batch columns categorical (below) or they enter the model as continuous covariates.
The Metadata TSV is the spine - the same --m-metadata-file drives demux barcodes, group-significance, taxa barplots, ANCOM-BC grouping, and Emperor coloring. First column header is the ID column (sample-id, id, #SampleID, ...). An optional second row #q2:types overrides type inference per column (categorical / numeric):
sample-id subject group
#q2:types categorical categorical
s1 101 treatment
s2 102 control
Without the #q2:types row, a column of only integers is inferred numeric - so a subject/batch/timepoint ID silently becomes a continuous covariate. Annotate ID-like integer columns categorical. Validate the sheet with Keemei (Rideout 2016 GigaScience 5:27) before running - a malformed metadata file is a top cause of cryptic action failures. qiime metadata tabulate --m-input-file metadata.tsv --o-visualization metadata.qzv renders any metadata (including an artifact viewed as metadata, e.g. taxonomy or denoising stats) as a table.
Goal: Recover the executable commands that produced an artifact, from the artifact alone.
Approach: Parse the embedded provenance DAG and regenerate a q2cli (or Artifact-API) script plus a citations BibTeX.
qiime tools replay-provenance --in-fp core-metrics/ --out-fp replay.sh --usage-driver cli
qiime tools replay-citations --in-fp core-metrics/ --out-fp citations.bib
# --usage-driver selects cli vs python3/artifact-api output. Verify flag spelling with
# `qiime tools replay-provenance --help` on the installed build (the interface is still maturing).
Replay recovers the commands; it is not a guaranteed bit-identical rerun across very different releases (plugin versions are part of the record). The aggregated DAG citations are how a methods section's references come straight from provenance, also via the .qzv Citations tab on view.qiime2.org.
Goal: Hand the data to R/Python when the analysis is no longer expressible in QIIME2 - while losing as little provenance as possible.
Approach: Stay in artifacts as long as the work is QIIME2-expressible; export (or read into phyloseq) only at the last step, and keep the upstream .qzas so the chain survives up to the exit.
qiime tools export --input-path table.qza --output-path exported/ # -> exported/feature-table.biom
biom convert -i exported/feature-table.biom -o feature-table.tsv --to-tsv
# FeatureData[Sequence] -> dna-sequences.fasta; FeatureData[Taxonomy] -> taxonomy.tsv; Phylogeny[Rooted] -> tree.nwk
Export DROPS the QIIME2 wrapper and the provenance - the exported TSV has no history back to the reads. For R, prefer qiime2R::qza_to_phyloseq('table.qza', 'taxonomy.qza', 'rooted-tree.qza', 'metadata.tsv') (Bisanz), which reads artifacts directly and assembles a phyloseq object without manual export. Record where the chain ends.
Trigger: qiime tools export to TSV at step three, then everything else in a notebook. Mechanism: export writes native data only and drops the provenance/ subtree. Symptom: the final figure has no provenance back to the raw reads - the framework overhead bought nothing. Fix: export at the LAST step; save upstream .qzas; or use qiime2R::qza_to_phyloseq so the chain survives to the exit.
Trigger: feeding a Phylogeny[Unrooted] or a FeatureData[Taxonomy] where a FeatureTable[Frequency] is required. Mechanism: the type system refuses incompatible inputs at the interface boundary before running. Symptom: "expected an artifact of type ..." error. Fix: this is the guard WORKING; qiime tools peek to read the actual Type, then fix the UPSTREAM action that produced the wrong type - do not re-import to coerce it.
Trigger: a silva-138-99-nb-classifier.qza from 2024.x used under 2026.x. Mechanism: the sklearn naive-Bayes classifier is pinned to its scikit-learn version; provenance replay assumes recorded plugin versions. Symptom: scikit-learn version-mismatch warning/error, or refusal to load. Fix: download/train the classifier for YOUR release (the data.qiime2.org/<release>/common/... URLs are release-namespaced); retrain or pin the whole env if reusing an old one.
Trigger: Phred64V2 on modern Illumina, V1-vs-V2 manifest confusion, relative paths, or SampleData[...] for still-multiplexed EMP data. Mechanism: the Phred offset is baked into the format name and is applied without checking. Symptom: silently mis-decoded quality scores, or an import that "works" but demux summarize shows garbage qualities. Fix: modern Illumina = Phred33; V2 TSV manifests with absolute paths; qiime demux summarize immediately after import; EMP data needs EMPPairedEndSequences + qiime demux.
Trigger: trying to feed a .qzv into the next action. Mechanism: a Visualizer's output is terminal by the framework's type contract. Symptom: the action will not accept it as an input. Fix: keep and feed the .qza the Visualizer was MADE from; a .qzv is for viewing only (browser or view.qiime2.org).
Trigger: an integer subject/batch/timepoint column with no #q2:types row. Mechanism: inference casts an all-integer column to numeric. Symptom: an ID enters a model as a continuous covariate; nonsensical group results. Fix: add a #q2:types row annotating ID-like columns categorical; validate with Keemei.
Trigger: expecting amplicon plugins in moshpit, or shotgun assembly in amplicon. Mechanism: distributions are curated, partially-disjoint plugin sets. Symptom: "plugin not found." Fix: amplicon/marker-gene -> amplicon distribution (renamed qiime2 in 2026.4); shotgun -> moshpit (and -> metagenomics); pin both distribution and release.
| Threshold | Source | Rationale |
|---|---|---|
--sampling-depth (rarefaction depth) |
-> diversity-analysis | required by core-metrics; pick from alpha-rarefaction, not a default - the 10000 in examples is a placeholder |
--p-formula integer columns annotated categorical |
use.qiime2.org metadata reference | otherwise inferred numeric and used as a continuous covariate |
| Phred offset = 33 (modern Illumina) | Illumina format history | Phred64 only for pre-2011 pipelines; wrong choice silently mis-decodes quality |
| Classifier release-match | Bokulich 2018 Microbiome 6:90 | the classifier is pinned to its scikit-learn version; cross-release reuse breaks |
| denoise / taxonomy / DA tuning | -> the owning sibling skill | this skill owns no scientific thresholds by design |
Most scientific magic numbers live in the five sibling skills, not here - this skill owns the machinery.
| Error / symptom | Cause | Solution |
|---|---|---|
| "The scikit-learn version ... could not be found" / classifier won't load | classifier .qza trained under a different release |
use the release-namespaced classifier or retrain under the current release |
| "Argument ... is not a subtype of ..." / type error | wrong semantic type wired into an action | qiime tools peek; fix the upstream action, do not re-import |
demux summarize shows nonsense quality scores |
wrong Phred offset in the import format name | re-import with ...Phred33V2; modern Illumina is Phred33 |
| Import fails on the manifest | V1/V2 confusion, relative paths, wrong delimiter | V2 TSV, absolute paths, tab-separated header sample-id |
A .qzv rejected as an action input |
Visualizations are terminal | feed the .qza it was made from |
| Action treats an ID column as continuous | no #q2:types row |
annotate the column categorical; validate with Keemei |
| Plugin not found | wrong distribution installed | install the amplicon (a.k.a. qiime2 in 2026.4) distribution |
