MicroDUET: a dual-headed genomic foundation model for pathogen surveillance

MicroDUET, a dual-headed ProkBERT-based genomic foundation model, identifies pathogen species, gene families, and resistance-linked function with up to 97.3% accuracy and AUROC 0.985.

• Why it matters: Surveillance needs fast, scalable comparisons from raw sequence data while jointly capturing evolutionary ancestry and phenotypic capability, which may not co-vary. Existing embeddings are compact but often lack reliable phylogeny-function separation for retrieval and triage.
• What they did: MicroDUET uses a 26.6M ProkBERT backbone with two 3.5M adapter heads: a contrastive phylogeny head on genome windows and a function head trained to identify genes across taxa. It was evaluated on held-out retrieval benchmarks and independent antimicrobial resistance-class benchmarks, plus lineage and drug-susceptibility-free tests.
• The result: The phylogeny head returned correct species for 80.3% (cellular) and 66.6% (viral) queries, while the function head achieved 97.3% and 97.2% gene-family accuracy and 74.4% on an antimicrobial resistance-class benchmark. It also reached 0.985 AUROC for rifampicin resistance without lineage adjustment and 0.974 AUROC for detecting an emerging mpox lineage as anomalous, while running 10^3–10^4× faster than single-threaded alignments.

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