Artificial intelligence approaches to resistome profiling and predictive antimicrobial resistance surveillance

Abstract Antimicrobial resistance (AMR) is a complex One Health challenge driven by the interconnected emergence, selection, persistence, and dissemination of resistance across clinical, animal, agricultural, food, and environmental microbiomes. Resistome profiling through whole-genome and metagenomic sequencing has transformed surveillance by enabling comprehensive characterization of antimicrobial resistance genes (ARGs); however, conventional homology-based approaches remain constrained by incomplete reference databases and may overlook divergent or previously uncharacterized resistance determinants. This review critically examines how artificial intelligence (AI), including machine learning, deep learning, transformer architectures, and biological language models, is advancing resistome profiling through ARG detection and classification, prioritization of putative novel resistance determinants, antimicrobial resistance phenotype prediction, and inference of microbial host and mobile genetic element associations. Particular emphasis is placed on the conceptual distinction between detecting resistance-associated sequence features and forecasting the future emergence of AMR. Although AI can identify sequence patterns associated with resistance and integrate genomic, metagenomic, multi-omic, clinical, epidemiological, and environmental data to estimate risk, these predictions do not independently establish that a candidate determinant will confer resistance in a new host, undergo successful horizontal transfer, persist within microbial populations, or become selectively advantageous under future ecological or clinical conditions. Resistance emergence is contingent on gene expression, genomic background, host compatibility, mobility, fitness effects, selection pressure, ecological persistence, and epidemiological connectivity. Accordingly, current AI outputs should generally be interpreted as detection, prioritization, phenotype prediction, or risk estimation rather than definitive forecasts of future resistance emergence. Major translational barriers include database bias, phylogenetic confounding, model opacity, limited genotype–phenotype data, poor external generalizability, inconsistent metadata, and insufficient prospective and experimental validation. Future progress will require biologically interpretable models, harmonized longitudinal One Health datasets, explicit evidence grading, and prospective validation against predefined emergence outcomes to establish whether AI-derived early-warning signals provide reliable and actionable lead time for AMR surveillance.