Machine learning prediction of human antibiotic resistance risk using 16S rRNA profiles.
Antimicrobial resistance poses a growing global health threat, yet large-scale surveillance and risk evaluation remain constrained by the cost and accessibility of metagenomic sequencing. Here, we demonstrate that antibiotic resistance risk, integrating gene mobility, human accessibility, clinical relevance, and host pathogenicity, can be quantitatively inferred from microbial taxonomic composition through its ecological coupling with microbial hosts. By integrating 177,134 metagenome-assembled genomes, 3,058 metagenomes, and 31,216 16S rRNA profiles, we defined a comprehensive ARG host catalogue and conserved core taxa across sequencing platforms. A machine learning model built on this framework achieved high predictive accuracy in held-out test data (R2 > 0.96) and retained strong performance in an independent dataset with paired 16S rRNA and metagenomic profiles (Pearson r = 0.74; Lin's CCC = 0.73), supporting its robustness and cross-platform transferability. Applying this tool on a global scale, we demonstrate that resistance risk exhibits consistent structure across populations, with host-associated ecological factors explaining more variation than socioeconomic conditions, supporting the feasibility of translating taxonomic profiles into quantitative estimates of functional risk. This work establishes a scalable framework for inferring antibiotic resistance risk from 16S data, enabling equitable, large-scale surveillance of antimicrobial resistance while positioning microbiome composition as a predictive basis for functional risk and advancing a general paradigm for inferring microbial traits from community structure.