HYMET: a hybrid metagenomic pipeline for accurate and efficient taxonomic classification

Reliable taxonomic classification of metagenomic sequences remains constrained by high mutation rates, fragmented assemblies, and large heterogeneous reference databases. HYMET (Hybrid Metagenomic Tool) was developed to overcome these challenges through a 2-stage hybrid design combining adaptive Mash-based screening with Minimap2 alignment and a coverage-weighted Lowest Common Ancestor classifier. Its sample-adaptive thresholds and on-the-fly reference database construction enable efficient, domain-agnostic classification while maintaining accuracy across divergent genomes.

Across 7 CAMI assembly datasets in contig mode, HYMET achieved a mean F1 of 83.89%, with genus-level F1 of 76.75% and species-level F1 of 60.18%, while averaging 115.93 s runtime and a mean peak memory of 6.24 GB. Performance remained stable under mutation rates up to 30% for most domains (F1 $\ge$ 0.8), with viral sequences showing the expected decline (F1 $\approx$ 0.5 at 30%). Read and contig inputs produced nearly identical results when sharing reference caches, and real-world datasets confirmed robustness with the human gut metagenome, which reproduced typical anaerobic profiles, while in the ZymoBIOMICS mock community, HYMET recovered all bacterial members; a further ground-truth evaluation on the ZymoBIOMICS Gut Microbiome Standard (D6331) yielded near-perfect genus-level concordance (Pearson $r = 0.998$, Bray-Curtis $= 0.04$) across bacteria, fungi, and archaea.

HYMET achieves a practical balance of accuracy, efficiency, and scalability for metagenomic classification. Its adaptive candidate selection, alignment-anchored taxonomy, and reproducible reference caching collectively enhance performance across domains. HYMET source code is fully available at https://github.com/ieeta-pt/HYMET.