Deep learning tool Taxonova classifies microbiome species from short and long amplicon reads

Accurately naming the species present in a microbiome sample is harder than it sounds. Ribosomal amplicon sequencing, the workhorse of microbiome studies, often produces short reads that cannot resolve the subtle differences between closely related species, and similarity-based classifiers struggle when a sequence is poorly represented in reference databases. No existing method, the developers note, performs robustly across the full range of amplicon lengths generated by different sequencing platforms.

A new open tool called Taxonova aims to close that gap. Presented in PLOS Computational Biology by Kardokh Kakabra, Shriram Patel, Andriy Temko and Marcus J. Claesson, it is a deep learning framework designed to handle both short-read and long-read ribosomal amplicons, from 250 base pairs up to 4,900 base pairs, within a single workflow.

The framework was evaluated across physical and biome-simulated mock communities representing seven biomes, using an accompanying benchmarking tool called Taxobench. On species-level classification, Taxonova attained F-measures of 0.76 to 1.00, with a median of 0.94, for short-read amplicons, and 0.82 to 0.99, with a median of 0.97, for long-read amplicons.

The method also improved abundance estimates. Compared with other classifiers, Taxonova reduced the species-level abundance-estimation error, measured as the L1 norm, by a median of 0.52 to 1.50 depending on the amplicon region. Its species-level false-positive rate was a median of 16.2 × 10⁻⁴ lower than that of the other classifiers tested.

To test real-world performance, the team compared amplicon-derived taxonomic profiles against paired shotgun whole-genome sequencing data from gut, soil and vaginal microbiomes. Taxonova achieved the strongest species-level concordance, measured by Jaccard index, in all three studies, exceeding the next-best classifier by 0.009 to 0.022 across both high- and low-complexity communities.

The authors caution that species-level assignment is expected to be most reliable for well-characterised communities, such as human-associated microbiomes, where reference databases are more complete. Poorly described environments remain a harder problem for any classifier that depends on existing reference knowledge.

Speed is another selling point. Taxonova runs at roughly 100 sequences per second on a single CPU core, which the developers say makes it suitable for routine deployment in research settings without specialised hardware.

The benchmark results suggest it could improve species-level resolution across diverse biomes, though its performance ultimately depends on the quality and coverage of the reference data it was trained to interpret.

Subject of Research: Deep learning classification of microbiome species from short- and long-read amplicon sequences

Article Title: Taxonova: Deep learning enables fast and confidence-aware microbiome species-level amplicon classification

Article References: Kakabra, K., Patel, S., Temko, A., & Claesson, M. J. (2026). Taxonova: Deep learning enables fast and confidence-aware microbiome species-level amplicon classification. PLOS Computational Biology, 22(10), e1014912. https://doi.org/10.1371/journal.pcbi.1014912

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014912

Keywords: microbiome, deep learning, amplicon sequencing, taxonomy, species classification, Taxonova, ribosomal RNA, bioinformatics, long-read sequencing, metabarcoding, Deep, learning