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

Rapid prediction of antimicrobial resistance (AMR) from genome sequences is essential for timely therapy, yet models based on curated marker panels or core-genome Single Nucleotide Polymorphisms (SNPs) often fail to generalize to novel bacterial lineages. We evaluate AMR prediction in Staphylococcus aureus using pan-genome features that encode homologous gene copy number (including absence) and compare them to SNP-based models across six antibiotics and 4255 isolates. Gradient-boosted decision tree ensembles (XGBoost) trained on gene copy number achieve macro-averaged F1-scores of 0.925-0.988, surpassing SNP-based models (0.838-0.935). Under lineage-held-out evaluation, which withholds entire clades to mimic previously unseen lineages, gene-content models retain markedly higher performance (F1 = 0.875 and 0.904 across two split schemes), whereas SNP-based models degrade substantially (F1 = 0.557 and 0.638). Feature ablation indicates that predictive signal is distributed across many homologous gene families rather than dominated by a few markers, a structure consistent with stronger cross-lineage generalization. Because gene-content features can be robustly obtained even from low-coverage sequencing, this approach extends genome-based AMR prediction to real-world clinical and epidemiological datasets. Together, these results show that copy-number-based pan-genome representations provide a robust alternative to SNP-only approaches, particularly when models must generalize to lineages not represented in training data.