Antibiotic resistance is one of the most urgent threats to global health. The human gut harbors a vast community of bacteria carrying antibiotic resistance genes, collectively known as the antibiotic resistome. These genes can spread among bacteria, complicate infection treatment, and influence health throughout life. Yet most detection tools rely on databases of known resistance sequences, so researchers have struggled to track resistance genes at the species level, recognize novel genes, and follow how the resistome changes with age.
Now, in a study published in hLife , researchers at Westlake University present LARGE, a large language model-based Antibiotic Resistance Gene Explorer. Trained on more than 64,000 experimentally curated sequences, LARGE classifies resistance type, mechanism, and gene name with a micro-F1 score of 0.987, outperforming established tools such as DeepARG, RGI, PLM-ARG, and ARGNet, and identifies the most novel resistance genes among all tools tested. Applying LARGE to 2,795 gut metagenomes from three independent Chinese cohorts, including 1,078 participants in the longitudinal Guangzhou Nutrition and Health Study, the researchers found that resistance genes carried by 31 bacterial species changed significantly during aging: 15 species gained resistance genes and 16 lost them.
“Our results show that the antibiotic resistome is not a fixed feature of the gut—it shifts dynamically as we age, and those shifts are tied to the risk of age-related chronic disease,” said Professor Ju-Sheng Zheng of Westlake University, one of the corresponding authors of the study. “LARGE is fast and runs on an ordinary graphics card, so we hope it will allow researchers around the world to map resistance genes in their own data and deepen our understanding of how gut bacteria influence human health.”
LARGE uses two frozen pre-trained language models, ESM-2 and FGeneBERT, which learn rich representations of protein and gene sequences, together with three lightweight classifiers that predict resistance type, mechanism, and gene name. This helps the model recognize previously unseen resistance genes that similarity-based tools would miss. LARGE is also remarkably efficient: it was trained on a single NVIDIA GeForce 4090 GPU in about 50 minutes and can screen 100,000 protein sequences in 38 minutes. The age-related trends from a longitudinal cohort were reproduced in two additional cohorts, and resistance gene changes in 31 species were linked to 13 age-related chronic diseases, including chronic kidney disease and coronary heart disease. Seven bacterial species from these 31 species formed a CKD-specific resistance score consistently associated with CKD risk.
These findings suggest that the gut antibiotic resistome may be a modifiable factor in healthy aging and that species-level resistance dynamics could serve as early indicators of chronic disease risk. LARGE is freely available as an online tool and as open-source software, so researchers worldwide can apply it to their own metagenomic data without specialized hardware. The team plans to extend LARGE to larger, more diverse populations, and to test whether diet or probiotics can reshape the resistome to benefit long-term health.
This work was conducted by researchers at Westlake University and Sun Yat-sen University, in collaboration with other institutions in China.
About Author:
Professor Ju-Sheng Zheng is an associate professor at Westlake University in Hangzhou, China. His research focuses on precision nutrition, multi-omics, and gut microbiome epidemiology, aiming to understand how diet and gut bacteria shape aging and chronic disease. He is a corresponding author of this study. For more information, please visit his research homepage: https://www.westlake.edu.cn/faculty/jusheng-zheng.html
Professor Stan Z Li is a Chair Professor of Artificial Intelligence at Westlake University in Hangzhou, China. He is a Fellow of IEEE and IAPR. His research focuses on AI fundamentals and AI for Science. For more information, please visit his research homepage: https://www.westlake.edu.cn/faculty/stan-zq-li.html
D OI Link:
https://doi.org/10.1016/j.hlife.2026.08.002
hLife
LARGE: A large language model-based tool that unveils high-resolution temporal dynamics of the antibiotic resistome during human aging
28-Aug-2026