Uncovering multidrug-resistance genes and potential biomarkers in Klebsiella pneumoniae using pan-genomic machine-learning feature-selection approach

Background. Antimicrobial resistance is one of the biggest threats to modern medicine, yet our understanding of its genetic basis remains incomplete. Current surveillance strategies rely on established resistance genes catalogued in curated databases; however, bacterial populations continuously evolve resistance mechanisms that may not be recognized or characterized in those existing catalogues. Here, we combined pan-genomic analysis with machine-learning feature selection to identify genes significantly associated with antibiotic resistance in Klebsiella pneumoniae and to explore potential resistance determinants that have not been recognized before. Results. A total of 8,132 K . pneumoniae genomes were collected to construct a comprehensive pan-genome comprising >180,000 gene families. Using AdaBoost with an incremental feature-selection approach, we identified parsimonious gene sets (14–27 genes per antibiotic) that achieved high predictive accuracy (mean area under the receiver operating characteristic curve: 0.96; F 1 -score: 0.93). Notably, about 15.8% of these genes were previously annotated as antimicrobial resistance determinants, whereas hypothetical proteins (HPs) and mobile genetic elements accounted for 32.5 and 22.3%, respectively. Among known resistance genes, we found that the ble gene, which encodes the bleomycin resistance protein, was widely associated with resistance to 8 of 12 included antibiotics. The functional re-annotation process also identified 15 HPs with diverse functions that are highly relevant to resistance, including transposases, components of the plasmid mobilization machinery, DNA manipulation proteins and regulatory factors. These findings underscore the importance of selecting and analysing highly associated features (or genes) in bacterial resistance. Conclusion. The study shows that pan-genomic analysis, coupled with feature selection, can uncover resistance-associated genes that conventional approaches miss. The findings also support enhancing surveillance strategies that incorporate emerging resistance mechanisms through large-scale genomic analysis, rather than relying exclusively on catalogued repertoires.