Prediction of antimicrobial resistance of Klebsiella pneumoniae through machine learning and revelation of significant role of insertion sequence elements

Klebsiella pneumoniae (K. pneumoniae) poses critical therapeutic challenges due to multidrug resistance (MDR). While antibiotic resistance genes (ARGs) are primary determinants, their expression and phenotypic impact are modulated by mobile genetic elements such as insertion sequence (IS) elements. However, the contribution of IS elements to antimicrobial resistance (AMR) prediction remains largely unexplored in K. pneumoniae. This study aimed to systematically evaluate whether integrating IS elements as genomic features improves machine learning-based AMR prediction in K. pneumoniae and to characterize IS-ARG co-occurrence patterns associated with MDR phenotypes. We retrieved genome sequences and corresponding antimicrobial susceptibility phenotypes for 2,732 K. pneumoniae isolates from the PATRIC database (2004–2024). We evaluated ARGs, IS elements, IS-ARG pairs, and curated key feature sets using six machine learning (ML) algorithms to predict resistance phenotypes for 15 antibiotics. Model performance was assessed using accuracy, recall, specificity, positive predictive value (PPV), F1-score, and the area under the receiver operating characteristic (ROC) curve (AUC). SHAP analysis was employed to interpret feature contributions, and enrichment of ARGs and IS elements in key feature sets was examined. Hierarchical clustering based on IS-ARG co-occurrence patterns was performed to explore associations with MDR profiles. The Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosting (GB) outperformed others, achieving mean accuracies of 88.11%, 87.89%, and 87.69% across all antibiotics. The key feature set yielded the highest mean accuracy of 89.59%, with GB attaining AUCs of 99.11% for ciprofloxacin and 98.23% for gentamicin. Models utilizing ARGs demonstrated superior overall performance over those employing IS elements, with a higher mean accuracy of 88.46% compared to 85.90%. RF with ARGs reached the highest per-antibiotic accuracy of 96.25% for ciprofloxacin, while IS elements achieved 93.42% for amikacin, exceeding ARGs. IS elements were significantly enriched in key feature sets of 13 drugs (fold change ≥ 2, P < 0.05), with RF and SHAP analysis identifying specific IS elements (e.g., IS1380, IS5, IS1182) as key predictors. IS-ARG co-occurrence analysis revealed 314 pairs, with IS6-blaSHV pairs accounting for 35.99%. Hierarchical clustering stratified isolates into seven clusters with distinct MDR profiles. Notably, IS21-blaKPC associations were linked to resistance against eight antimicrobials, while the absence of specific IS-ARG pairs correlated with resistance to ceftazidime, ciprofloxacin, and levofloxacin. This study established the potential to improve predictive accuracy in ML models for specific antibiotics when utilizing IS elements as genomic features in K. pneumoniae, with RF, SVM and GB achieving superior performance across multiple feature sets. IS elements served as critical predictors for specific resistance phenotypes beyond ARGs alone, and their co-occurrence with ARGs was associated with MDR phenotypes. The identification of IS21-blaKPC as a marker of resistance to eight antibiotics and the stratification of isolates into clinically relevant clusters provided candidate targets for genomic surveillance. These findings demonstrate that mobile genetic elements provided antibiotic-specific predictive improvements and enhanced biological interpretability by revealing the mobile genetic context of resistance gene dissemination.