phylotypy

A Naive Bayesian Classifier for 16S rRNA gene sequences, inspired by the phylotypr R package by Riffomonas. Designed for classifying amplicon sequence variants (ASVs) from DADA2, QIIME2, or raw FASTA files against a reference database of 16S rRNA sequences. The RDP training data is provided here in the data directory located at the github repository. But Silva and others can be used.

Thanks to Riffomonas for the inspiration — check out the videos on his YouTube channel.

Performance

The full RDP reference database takes ~7.5 seconds and the full Silva reference database (genus level) takes ~19 seconds on a 2020 Apple Intel MacBook Pro with 16Gb of RAM. Newer systems should see a substantial increase in performance.

How to Install

Using pip:

pip install phylotypy

Using uv (recommended — how to install uv):

uv pip install phylotypy

Note: Intel Mac (x86_64) users are limited to numba 0.62.1, which is pinned in this package. Apple Silicon (M-series) users are not affected.

Training Data

Download the RDP reference training set and an example dataset before classifying:

"Kingdom", "Phylum", "Class", "Order", "Family", "Genus", "Species"

The taxon string in the fasta description should follow the semicolon-separated format like this:

>Bacteria;Pseudomonadota;Gammaproteobacteria;Enterobacterales;Enterobacteriaceae;Citrobacter
TAGAGTTTGATCCATGGCTCAGATTGAACGCTGGCGGCAGGCCTAACAC.....

Reference Database: Taxonomic Levels

Every reference sequence needs a taxonomy string in its FASTA header (see the format above), and all headers in the file must have the same number of ;-delimited levels, e.g. Kingdom;Phylum;Class;Order;Family;Genus is 6 levels. This matters because make_classifier() and classify_sequences() compare taxonomy strings assuming a fixed depth; a reference set with mixed depths will fail, either right away or partway through classification.

1. Headers already share the same depth

If your reference fasta is already uniform (true of the RDP training set used above, at 6 levels/genus), just load and build normally:

from phylotypy import read_fasta, classifier

rdp = read_fasta.read_taxa_fasta("rdp_16S_v19.dada2.fasta", is_ref=True)
database = classifier.make_classifier(rdp, verbose=True)

Pass is_ref=True to read_taxa_fasta() to check this up front: it counts the semicolon-delimited levels in every header, and raises a ValueError naming the file and the depths it found if they're not all the same. So a bad reference file is caught immediately, not partway through building the classifier or later during classification. Leave is_ref at its default (False) when reading sequences you're going to classify, since those don't carry a fixed-depth taxonomy string.

2. Fixing ragged headers

"Ragged" headers means the taxonomy strings in the reference fasta don't all reach the same depth. Some sequences are annotated down to Genus, others only down to Class or Order:

>Bacteria;Firmicutes;Bacilli;Lactobacillales;Lactobacillaceae;Lactobacillus
TAGAGTTTGATCCTGGCTCAG...
>Bacteria;Firmicutes;Bacilli
TAGAGTTTGATCCTGGCTCAG...

This is common with broader databases like Silva. Trying to build a classifier from ragged reference data (or reading it with is_ref=True) raises:

ValueError: ref_db contains taxonomy strings with inconsistent numbers of taxonomic levels:
id
3      1204
6     88213
All 'id' entries must have the same number of ';'-delimited levels before building a
classifier, or classify_sequences() will fail later on.
Options:
  - Fix/pad the taxonomy strings in ref_db so every entry has the same depth.
  - Call make_classifier(ref_db, filter_db=True) to filter ref_db down to a single
    consistent depth (see training_data.filter_train_set).
  - Pass n_levels=<int> together with filter_db=True to control which depth is kept.

Fix it with training_data.filter_train_set(), which drops sequences that don't match a target depth (n_levels) and strips out common noise terms (Eukaryota, metagenome, Candidatus, etc.. See training_data.DEFAULT_NOISE_TERMS):

from phylotypy import read_fasta, training_data, classifier

# is_ref left at False here — the file is ragged, so the check would raise
silva = read_fasta.read_taxa_fasta("silva_nr99.fasta.gz")

filtered = training_data.filter_train_set(silva, n_levels=6, verbose=True)
database = classifier.make_classifier(filtered, verbose=True)

With verbose=True, filter_train_set() reports how many sequences it dropped:

Removed 12,458 sequences with taxonomic depth != 6 (kept 158,970)

Or skip the manual filter step and let make_classifier() do it for you:

database = classifier.make_classifier(silva, filter_db=True, verbose=True)

This runs filter_train_set() internally, auto-detecting n_levels from the majority depth in your data if you don't pass one, and prints the same before/after counts when verbose=True.

Command line interface

Classifying sequences can be done on the command line using:

phylotypy classify --input dna_moving_pictures.fasta \
                   --db rdp_16S_v19.dada2.fasta \
                   --out classied_seqs.tsv \
                   --verbose

You can save the classifier (a pickle file) and reuse it later by specificying --save-db:

phylotypy classify --input dna_moving_pictures.fasta \
                   --db rdp_16S_v19.dada2.fasta \
                   --save-db rdp_classifer.pickle \ # set the path
                   --out classied_seqs.tsv \
                   --verbose
                   
# resuse on another run saves time since the db is already built
phylotypy classify --input my_sequences.fasta \
                   --db rdp_classifer.pickle \
                   --out classied_my_seqs.tsv \
                   --verbose

Help menu:

phylotypy --help

usage: phylotypy [-h] [--version] {build,classify} ...

Naive Bayes classifier for 16S rRNA sequence data.

positional arguments:
  {build,classify}
    build           build a classifier database from a reference fasta and save it to disk
    classify        classify sequences against a database

options:
  -h, --help        show this help message and exit
  --version         show program's version number and exit

--

phylotypy classify --help
usage: phylotypy classify [-h] -i INPUT -d DB -o OUT [--save-db SAVE_DB] [--res-extended] [--kmer-size KMER_SIZE] [--num-bootstrap NUM_BOOTSTRAP]
                          [--min-consensus MIN_CONSENSUS] [--n-levels N_LEVELS] [--threads THREADS] [--force] [-v]

options:
  -h, --help            show this help message and exit
  -i, --input INPUT     fasta of representative/dereplicated sequences to classify (e.g. ASVs or OTU centroids -- not raw reads)
  -d, --db DB           reference database: a prebuilt classifier (.pkl/.pickle, from `phylotypy build`) or a raw reference fasta, which will be built on the fly
  -o, --out OUT         output path for classification results (.tsv or .csv; default tab-separated)
  --save-db SAVE_DB     if --db is a raw fasta, add a file path to save the built classifier here for reuse (avoids rebuilding it on the next run)
  --res-extended        Created an extended results report including qiime formatted lineage, lineages split into taxonomic levels
  --kmer-size KMER_SIZE
                        k-mer size; must match the database's k-mer size (default: 8)
  --num-bootstrap NUM_BOOTSTRAP
                        number of bootstrap replicates for the confidence estimate (default: 100)
  --min-consensus MIN_CONSENSUS
                        bootstrap confidence threshold 0-100; ranks below this are reported as '_unclassified' (default: 80)
  --n-levels N_LEVELS   number of taxonomic levels in the output (default: auto-detected from the reference database)
  --threads THREADS     number of CPU threads to use for classification (default: all cores)
  --force               skip the memory-safety check before bootstrap sampling
  -v, --verbose         print progress messages

--

phylotypy build --help
usage: phylotypy build [-h] -i INPUT -o OUT [--kmer-size KMER_SIZE] [--threads THREADS] [--filter-db] [--max-per-genus MAX_PER_GENUS] [--n-levels N_LEVELS] [-v]

options:
  -h, --help            show this help message and exit
  -i, --input INPUT     reference fasta with taxonomy strings in the sequence headers
  -o, --out OUT         output path for the pickled classifier database
  --kmer-size KMER_SIZE
                        k-mer size used to build the database (default: 8)
  --threads THREADS     number of CPU processes to use while building (default: 4)
  --filter-db           filter the reference database to a single consistent taxonomic depth and drop noisy entries before building
  --max-per-genus MAX_PER_GENUS
                        down-sample to at most this many sequences per genus (requires --filter-db)
  --n-levels N_LEVELS   taxonomic depth to keep when --filter-db is set (default: auto-detected)
  -v, --verbose         print progress messages

Using the API

1. Load training data and sequences to classify

from phylotypy import classifier, results, read_fasta

rdp = read_fasta.read_taxa_fasta("rdp_16S_v19.dada2.fasta")
moving_pics = read_fasta.read_taxa_fasta("dna_moving_pictures.fasta")

2. Train the classifier

# Accepts fasta or csv/tsv with 'id' and 'sequence' column names
database = classifier.make_classifier(rdp, verbose=True)

# If the reference has ragged (inconsistent) taxonomic levels, use filter_db=True
# to remove the mismatched records — see "Reference Database: Taxonomic Levels" above
database = classifier.make_classifier(rdp, filter_db=True, verbose=True)

3. Classify sequences

classified = classifier.classify_sequences(moving_pics, database)

4. Format and export results

classified = results.summarize_predictions(classified)
print(classified.columns)

Output:

Index(['id', 'sequence', 'classification', 'Kingdom', 'Phylum', 'Class',
       'Order', 'Family', 'Genus', 'observed', 'lineage'],
      dtype='object')

classified.to_csv("classified_results.csv")

Complete Code Block

from phylotypy import classifier, results, read_fasta

rdp = read_fasta.read_taxa_fasta("rdp_16S_v19.dada2.fasta")
moving_pics = read_fasta.read_taxa_fasta("dna_moving_pictures.fasta")

database = classifier.make_classifier(rdp)

classified = classifier.classify_sequences(moving_pics, database)
classified = results.summarize_predictions(classified)
print(classified.head())

classified.to_csv("classified_results.csv")

Example Classification Output

Taxonomic levels (Domain → Genus) are semicolon-separated. Numbers in parentheses represent bootstrap confidence scores. The default confidence threshold is 80%.

Bacteria(100);Pseudomonadota(99);Alphaproteobacteria(99);Rhodospirillales(99);Acetobacteraceae(99);Roseomonas(83)

Bacteria(99);Bacteroidota(97);Bacteroidia(93);Bacteroidales(93);Bacteroidales_unclassified(93);Bacteroidales_unclassified(93)

Bacteria(100);Bacteroidota(100);Bacteroidia(100);Bacteroidales(100);Bacteroidaceae(100);Bacteroides(100)

Working with Your Own Data

phylotypy works with FASTA files from DADA2, QIIME2, or any standard pipeline. See read_fasta.py for utilities to load and convert sequence data into the required format.

A complete walkthrough is available in vignette.py.

Requirements

Dependencies are installed automatically via pip. See pyproject.toml for the full list.

Citation

If you use phylotypy in your research, please cite:

• Wang, Q., Garrity, G.M., Tiedje, J.M., Cole, J.R. (2007) Naive Bayesian Classifier for Rapid Assignment of rRNA Sequences into the New Bacterial Taxonomy. Applied and Environmental Microbiology, 73(16), 5261–5267.
• Schloss PD.2025.phylotypr: an R package for classifying DNA sequences. Microbiol Resour Announc14:e01144-24.https://doi.org/10.1128/mra.01144-24
• Saltikov, C. (2024) phylotypy: Python implementation of a Naive Bayesian 16S rRNA classifier. https://github.com/csaltikov/phylotypy