Pangenome-based interpretable machine learning framework for predicting antimicrobial resistance in foodborne...

Antimicrobial resistance (AMR) in foodborne Escherichia coli (E. coli) remains a significant public health concern. Predicting AMR from whole-genome sequencing data has become a critical component of modern surveillance. However, conventional approaches relying solely on known resistance genes often fail to capture the broader genomic context associated with resistance. In this study, we developed interpretable machine learning models using pangenome-derived accessory genes from 655 foodborne E. coli isolates. These models achieved robust predictive performance across five antibiotics. SHAP analysis revealed that the models relied on genomic features beyond canonical resistance determinants, including markers of mobile genetic elements and stress response mechanisms. Furthermore, the disinfectant resistance gene qacEΔ1 emerged as a consistently important predictive feature across different antibiotic models. This suggests qacEΔ1 as a reliable positional marker for the physical linkage of resistance genes on integron platforms, representing a genomic context associated with potential co-selection pressures in food production environments. Overall, this study demonstrates the feasibility of the pangenome-based framework to provide interpretable predictions and capture potential genomic signals linked to the dissemination of resistance.