MicrobeRT: A Framework and Suite of Fine-tuned Language Models for Genomic Sequences

MicrobeRT fine-tuned genomic language models classify metagenomic reads for species, antibiotic resistance genes, and pathogenicity, achieving strong performance while reducing computational overhead across simulated and empirical datasets.

• Why it matters: Metagenomic surveillance needs accurate, reproducible detection, but targeted assays miss atypical or novel infections and shotgun workflows are complex, costly, and tied to mutable reference databases.
• What they did: The suite fine-tunes three pre-trained models—Nucleotide Transformer, DNABERT, and HyenaDNA—on curated datasets spanning more than 500,000 genomes and 250,000 antibiotic resistance gene sequences, then benchmarks against SOTA tools.
• The result: Results on both simulated and empirical datasets show strong task performance with lower compute, and MicrobeRT is released open-source with drop-in integration into TaxTriage for scalable, reproducible diagnostics.

AI summary of the abstract · may contain errors · not medical advice · read the paper.