A new study suggests that genomic language models (gLMs) can improve the detection of antibiotic resistance genes, especially those that traditional tools often miss.
Overview
A study published in npj Antimicrobials and Resistance introduced resLens, a family of genomic language models designed to detect antibiotic resistance genes (ARGs). The models aim to identify ARGs that conventional alignment-based tools may overlook.
Methods
Researchers obtained ARG sequences from the NCBI Pathogen Detection RefGene and ResFinder databases. After preprocessing, the dataset comprised over 7,600 ARGs across 12 antibiotic classes, combined with an equal number of non-resistance genes from GenBank. They fine-tuned four models: two for long-read (LR) data and two for short-read (SR) data. One model performed binary classification (ARG vs. non-ARG), while the other classified predicted ARGs into specific antibiotic classes. The models were evaluated against five alignment-based tools (AMR++, KARGA, ResFinder, Meta-MARC, RGI) and two deep learning models (DeepARG, ARGNet).
Results
On the LR dataset, resLens outperformed other models, though the difference with KARGA and RGI was modest. On the SR dataset, KARGA and RGI outperformed resLens. The model closely replicated the class distribution in the LR test set. In tests on novel ARGs (families ANT and blaADC), resLens showed variable performance and generalized beyond close database matches, while ResFinder failed to identify any blaADC genes. In a stricter clustered-split analysis, performance declined, indicating limitations under strong distribution shifts. On whole-genome sequencing data from 79 genomes, resLens and RGI identified at least one gene corresponding to labeled phenotypes more often than ResFinder.
"Genomic language models can improve ARG detection, especially for poorly represented ARGs."
Limitations
The authors noted that the whole-genome analysis was exploratory due to limited sample size and lack of gene-level annotation. Manual validation revealed false positives and ambiguous classifications, emphasizing the need for careful validation.
Conclusion
The study suggests that genomic language models can improve ARG detection, but warns that these tools should be used for screening and hypothesis generation, not final conclusions.