Gene copy-number features generalize better than SNPs for antimicrobial resistance prediction in Staphylococcus aureus

Genome-based prediction of antimicrobial resistance in Staphylococcus aureus is substantially improved by using pan-genome gene copy-number features rather than traditional SNP-based models. Across 4,255 isolates and six antibiotics, machine-learning models trained on gene content achieved higher accuracy and far better generalization to previously unseen lineages, while SNP-based models performed well only when closely related strains were included in training. The predictive signal was spread across many gene families rather than a few resistance markers, explaining the stronger robustness across lineages. Because gene copy-number data can be reliably derived from low-coverage sequencing, this approach offers a more generalizable and clinically practical method for rapid AMR prediction.