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

Accurate species-level taxonomic classification of microbial amplicon sequences remains challenging due to the difficulty of resolving subtle variations within conserved ribosomal regions. Traditional similarity-based methods struggle with short amplicons and sequences with poor database representation. No existing method performs robustly across the full spectrum of amplicon lengths, limiting applicability across diverse sequencing platforms and variable-region targets. We present Taxonova, a deep learning framework for both short- and long-read ribosomal amplicons. Evaluated across physical and biome-simulated mock communities representing seven biomes (via the accompanying benchmark tool Taxobench), Taxonova attains species-level F-measures of 0.76–1.00 (median 0.94) for short-read and 0.82–0.99 (median 0.97) for long-read amplicons spanning 250 bp to 4,900 bp, reduces the species-level abundance-estimation error (L1 norm) by a median of 0.52 to 1.50 depending on amplicon region, and has a species-level false-positive rate that is a median of 16.2 × 10 ⁻⁴ lower than that of the other classifiers. Species-level assignment is expected to be most reliable for well-characterised communities, such as human-associated microbiomes. To demonstrate real-world applicability, we also compared amplicon-derived taxonomic profiles to paired shotgun whole-genome sequencing from gut, soil and vaginal microbiomes; Taxonova achieves the strongest species-level concordance (Jaccard) in all three studies, with a species-level Jaccard index 0.009 to 0.022 higher than that of the next-best classifier across both high- and low-complexity communities. Taxonova runs at a throughput of roughly 100 sequences per second on a single CPU core, making it suitable for routine deployment in research settings.