LLM-powered construction of antibiotic resistance datasets and knowledge graphs for accelerated antibiotic resistance gene discovery

Antibiotic resistance is a mounting global public health and development threat, with the misuse and overuse of antibiotics driving the rapid dissemination of multidrug-resistant pathogens. Developing next-generation small-molecule antibiotics requires in-depth investigation of resistance mechanisms and the identification of novel targets to overcome the limitations of current treatments. However, current strategies mainly rely on sequence-based homology search or models trained on manually curated datasets, which often suffer from outdated annotations, limited coverage, and high false-positive rates. These limitations hinder the identification of novel resistance targets. To address these limitations, we developed a Large Language Model (LLM)-powered automated pipeline that systematically mines resistance-related knowledge from full-text publications and supplementary materials, thereby supporting scalable and iterative data integration. Building on this framework, we constructed AResDB , a curated resource that integrates data from over 8570 published papers covering bacterial species, compounds, genes, mutations, phenotypes, and quantitative measurements. On this foundation, we constructed AResKG , a large-scale, multi-species knowledge graph of antibiotic resistance integrating 14 entity types, 27 relation types, and over 52.1 million triples. Experimental results based on AResKG demonstrate its effectiveness in improving the performance of commonly used machine learning methods for predicting gene-antibiotic resistance, compared with existing knowledge graphs. This work establishes a scalable, LLM-powered framework for antibiotic resistance research that advances the understanding of resistance mechanisms and accelerates the discovery of next-generation therapeutics.