LARGE: A large language model-based tool that unveils high-resolution temporal dynamics of the antibiotic resistome during human aging

The gut microbiome undergoes profound changes during aging, including shifts in the microbial antibiotic resistome: the collective repertoire of antibiotic resistance genes (ARGs). We developed a language model-based ARG explorer (LARGE), which used frozen ESM-2 and FGeneBERT encoders to embed protein sequences and three independent multilayer perceptron classifiers to predict resistance type, mechanism, and gene name. LARGE was trained on 34,008 ARG and 30,309 non-ARG sequences from the NCRD95 and UniProt databases and benchmarked against DeepARG, RGI, PLM-ARG, and ARGNet using F1-scores on multiple independent test sets. We applied LARGE to a longitudinal cohort (GNHS, n = 1078, mean baseline age 64.6 years), and validated key findings in independent cohorts (CHNS, n = 356; ZMSC, n = 1361). Linear mixed-effects models were used to assess age-related ARG trajectories, and logistic regression examined associations with 13 chronic diseases. In benchmark test, LARGE substantially outperformed all four tools, achieving superior F1-scores. In longitudinal cohorts, LARGE revealed declining ARG burdens in dominant resistance drug classes and identified 31 species with significant ARG changes. The ARGs of these 31 species are closely associated with multiple age-related chronic diseases, such as chronic kidney disease and coronary heart disease; a specific chronic kidney disease ARG score was consistently associated with disease risk in both discovery and validation cohorts. The findings highlight the pivotal role of aging in resistome evolution and potentially provide microbiome-targeted interventions to mitigate antibiotic resistance and promote healthy aging.