Large language model-assisted curation and sequence-informed modelling reveal membrane and genetic determinants of antibiotic resistance gene removal.

Water acts as a critical medium for the emergence and spread of antibiotic resistance genes (ARGs), which pose threats to public health. Efficient removal of extracellular ARGs (eARGs, hereafter ARGs) by functional membranes is critical for water security, yet their physicochemical heterogeneity poses significant challenges for membrane design. Herein, this study reports a large language model-driven machine learning framework to guide rational design of membranes with high selectivity for ARG removal. A large language model-based workflow was established to automatically extract standardized membrane and ARG datasets, followed by the development of a sequence-informed hierarchical filtration (SIHF) model for facilitating membrane design. By integrating sequence motif analysis of ARGs with membrane properties within a machine learning architecture, SIHF can identify the dominant factors governing ARG retention and quantify sequence-dependent incremental effects. SIHF analysis reveals that membrane pore size emerges as the primary design factor, with hydraulic flux modulating its impact. Independent filtration experiments further supported the predictive performance of SIHF. Overall, this work establishes an actionable and evidence-based design framework to advance the development of next generation membranes for controlling ARG dissemination and mitigating antimicrobial resistance.