Taxonova models
Pretrained checkpoints for Taxonova, a classifier that assigns 16S rRNA and rRNA operon amplicon sequences to a lineage from domain to species, with a confidence at each rank.
This repository is a mirror of the files attached to the
GitHub releases. You do not
need to download anything by hand: the taxonova package fetches a model the
first time it is used, from GitHub or, if that fails, from here, and checks its
sha256.
pip install taxonova
taxonova classify reads.fasta --model v4
Or classify in the browser with the web app.
Files
| model id | amplicon | reference | file | sha256 |
|---|---|---|---|---|
v4 |
16S V4 | Greengenes2 2024.09 | taxonova-V4-greengenes2-26.10.pt |
c52aaafe2b9b79092171c11f67de469ac9d64f3e8fadac74e63e756792c0a3a3 |
v3v4 |
16S V3-V4 | Greengenes2 2024.09 | taxonova-V3V4-greengenes2-26.10.pt |
005677c10327b981847b9ec5c5c529828716d0d628706fa65e2f77222309e0aa |
v5v7 |
16S V5-V7 | Greengenes2 2024.09 | taxonova-V5V7-greengenes2-26.10.pt |
77f7a75bf38c4923422dadf88e971764a3ccf604b23c1c96b46c0b325a670385 |
16s |
whole 16S gene (V1-V9) | Greengenes2 2024.09 | taxonova-16S-greengenes2-26.10.pt |
d2ec6d014cf771525887c136fa0591b09d5284db6035022bb60f7f96d2e453ca |
operon |
rRNA operon (16S-ITS-23S) | GROND (GTDB R207 build) | taxonova-operon-grond-1.0.pt |
9ac36fea212edd569198027c9e6fb182199bb316acacad188d758929c9408289 |
Species names follow GTDB. The number at the end of a file name is the model version (when the checkpoint was made), not the package version. A released file is never changed; a new checkpoint gets a new file name.
Each file is a PyTorch checkpoint that holds the network weights together with
the label encoders and the taxonomy, as written by taxonova training. Load
the files with the taxonova package rather than with torch.load directly.
Loading a checkpoint runs Python's pickle, so use files from this repository
or the GitHub releases, and check the sha256.
How each model was made, and the earlier file names of the same files, are described in docs/models.md. The method and its evaluation are in the paper.
Licence
MIT, as the Taxonova software.
Citation
Kakabra K, Patel S, Temko A, Claesson MJ. Taxonova: deep learning enables fast and confidence-aware microbiome species-level amplicon classification. PLoS Comput Biol. 2026;22(10):e1014912. doi:10.1371/journal.pcbi.1014912
@article{kakabra2026taxonova,
author = {Kakabra, Kardokh and Patel, Shriram and Temko, Andriy and Claesson, Marcus J.},
title = {Taxonova: Deep learning enables fast and confidence-aware microbiome species-level amplicon classification},
journal = {PLOS Computational Biology},
year = {2026},
volume = {22},
number = {10},
pages = {e1014912},
doi = {10.1371/journal.pcbi.1014912}
}
Developed in the Claesson Lab, University College Cork and APC Microbiome Ireland. Questions and problems: GitHub issues.