resLens

Antibiotic resistance is a mounting global health threat, and surveillance depends on accurately identifying antibiotic resistance genes (ARGs) in genomic and metagenomic sequence data. The dominant detection strategy is alignment-based: query sequences are matched against curated ARG reference databases. This works well for genes that resemble known references but degrades sharply for novel or divergent ARGs — exactly the variants of greatest concern for emerging resistance — because alignment cannot recognize what is not already cataloged.

resLens, developed by Mollerus and colleagues at The George Washington University and posted to bioRxiv (with an updated version in February 2026), addresses this gap by reframing ARG detection as a learned classification problem over genomic language model representations. The approach fine-tunes a pretrained DNA language model on curated ARG datasets, producing a family of models that classify resistance genes from latent sequence representations rather than explicit alignment. Because the underlying language model captures generalizable sequence features, resLens is designed to recognize ARGs even when their sequences and resistance mechanisms differ substantially from those in reference databases.

The authors report that resLens achieves competitive or superior performance to alignment-based baselines across multiple evaluation scenarios, including the challenging case of ARGs dissimilar to reference data.

#Key Features

• Genomic language model backbone: resLens fine-tunes a pretrained DNA language model, leveraging latent genomic representations instead of relying solely on alignment to reference databases.
• Generalization beyond references: The method is designed to detect ARGs whose sequences and resistance mechanisms diverge from cataloged references, where alignment-based tools tend to fail.
• Family of models: resLens is presented as a family of models for ARG detection and classification across multiple evaluation settings.
• Competitive accuracy: The authors report competitive or superior classification performance relative to alignment-based methods across the evaluated scenarios.

#Technical Details

resLens is built by fine-tuning a pretrained DNA language model on curated antibiotic resistance gene datasets, casting ARG detection and classification as a supervised task over the model's learned sequence representations. By operating on latent genomic features rather than direct alignment to reference sequences, the models aim to generalize to divergent ARGs. The authors evaluate across multiple scenarios — including cases where test ARGs are dissimilar in sequence and mechanism to the reference data — and report competitive or superior performance to alignment-based baselines. The preprint does not yet provide a public code or weights link, and specific details such as the base DNA language model, parameter count, and training-set composition should be confirmed against the full text as the work advances through review.

#Applications

resLens is intended for microbiologists, genomic epidemiologists, and surveillance programs that screen bacterial genomes and metagenomes for resistance determinants. Its advantage is most pronounced when monitoring for novel or rapidly evolving ARGs that existing reference databases do not yet contain — a recurring blind spot in alignment-based pipelines. Such a tool can complement standard ARG-annotation workflows in clinical microbiology, environmental resistance monitoring, and One Health surveillance, where catching divergent resistance genes early has direct public-health value.

#Impact

resLens illustrates how DNA language models can move antibiotic resistance gene detection beyond the inherent ceiling of reference-database alignment, contributing to a broader shift toward representation-learning approaches in genomic surveillance. By targeting the hardest case — ARGs unlike anything in current references — it speaks directly to the needs of resistance monitoring. As a bioRxiv preprint without a released code or weights link at the time of writing, its reported gains await peer review, independent reproduction, and public tooling before it can be integrated into routine surveillance pipelines.