Routine susceptibility panels predict reserve-antibiotic resistance across ESKAPE pathogens: an ontology-anchored machine learning analysis with cross-species transfer

Abstract

Resistance to reserve antibiotics such as carbapenems, glycopeptides, and oxazolidinones is the most consequential and least measured part of antimicrobial resistance. We asked whether resistance to these last-line agents can be predicted from the routine susceptibility panel that laboratories already produce, whether the predictive signal reflects shared resistance mechanisms, and whether it generalises across species. Using per-isolate susceptibility data for seven high-priority bacterial pathogens from a public bacterial bioinformatics resource (about 24,000 isolates), we built gradient-boosted models that predict resistance to each antibiotic from the rest of the panel, with every agent annotated for drug class and mechanism from a curated resistance ontology. Of 220 candidate pathogen-antibiotic pairs, 73 met a pre-specified eligibility rule and were evaluated by five-fold cross-validation. Resistance was broadly predictable (mean area under the curve 0.80 across the 73 models; 0.82 for reserve agents), reaching 0.90 for carbapenem resistance in Klebsiella pneumoniae and 0.97 for glycopeptide and oxazolidinone resistance in Enterococcus faecium. Performance was largely retained after same-class predictors were removed, indicating genuine cross-class co-resistance, and carbapenem resistance transferred across Gram-negative species. Routine panels therefore carry a strong, mechanism-aware signal for reserve-drug resistance.