Biased sampling driven by bacterial population structure confounds machine learning prediction of antimicrobial resistance

Machine-learning models are increasingly used to predict antimicrobial resistance from bacterial genome data, but their performance is strongly undermined by hidden biases in how data are collected. Using more than 24,000 genomes from five different pathogens, this study shows that bacterial population structure and over-representation of human disease isolates can falsely link resistance to lineage rather than biology. As a result, many ML models perform poorly, even when trained on very large datasets. Model accuracy varies by species and antibiotic, highlighting that one-size-fits-all approaches do not work. Overall, the findings demonstrate that current ML methods for AMR prediction often fail under realistic conditions and must explicitly account for population structure and sampling bias, supported by more diverse and representative genomic datasets.