Pan-genome-based Insight Extraction for Antimicrobial Resistance Prediction in Klebsiella pneumoniae

BACKGROUND: Antimicrobial resistance (AMR) is one of the most critical global health challenges we face today, driven by the rise of drug-resistant pathogens and a diminishing pipeline of novel antibiotics. Klebsiella pneumoniae is a common cause of hospital-acquired infections worldwide among these pathogens. It is increasingly resistant to antibiotics, including the last resort carbapenems. Therefore, understanding the underlying mechanisms contributing to this resistance is crucial for preventing infections and potential outbreaks in the future. Most current studies on predicting resistant strains focus primarily on analyzing known AMR genes and whole-genome sequences. However, most of these conventional approaches, which entirely rely on existing resistance genes, often fail to capture the full genetic diversity underlying resistance mechanisms. Thus, this study aims to improve predictive performance and accurately identify the genetic determinants of antimicrobial resistance in K. pneumoniae.
METHOD: Whole genome sequences of Klebsiella pneumoniae were garnered and integrated with their corresponding antimicrobial resistance phenotypes to generate the drug-associated pan-genome matrices. These matrices documented gene families' presence/absence patterns (alternatively referred to as gene clusters) across the analyzed genomes. Subsequently, the pan-genome matrices were utilized as input for extracting critical genes using machine learning and feature selection algorithms. The gene subsets identified from the feature selection process, which demonstrate strong correlations with resistance profiles of K. pneumoniae, will undergo comprehensive functional characterization and analysis to elucidate their potential resistance mechanisms.
RESULTS: Our analysis utilized 8,132 high-quality K. pneumoniae genomes with annotated AMR phenotypes, enabling the construction of 12 drug-specific pan-genome matrices for subsequent machine learning implementation. Through the feature selection process, we identified highly predictive gene subsets comprising approximately 20 genes for each antibiotic. These selected gene subsets achieved a remarkable mean Area Under the Curve (AUC) of 0.96, statistically comparable to both the complete pan-genome with tens of thousands of genes and the established repertoire of hundreds of known AMR genes (mean AUC = 0.97 and 0.96, respectively; all comparisons yielded p > 0.05, Wilcoxon rank-sum test). Notably, only 18.1% of our selected genomic features corresponded to previously characterized antibiotic-resistance determinants, while approximately 27% encoded hypothetical proteins with undetermined functional roles. In addition to this, genes associated with mobile genetic elements constituted about 23.1% of the identified features. These findings underscore the significant contribution of uncharacterized genetic determinants and mobile genetic elements to antimicrobial resistance phenotypes, reflecting the dynamic nature of genetic exchange in bacterial evolution.
CONCLUSIONS: Our results demonstrate that pan-genome-wide feature selection represents a robust methodological framework for developing high-performance AMR prediction models that transcend the limitations of approaches relying on known AMR determinants. Furthermore, this strategy facilitates the identification of previously unrecognized genetic factors contributing to antimicrobial resistance, thereby expanding our understanding of resistance mechanisms in K. pneumoniae. The novel genetic determinants identified through this approach may reveal unexplored biological pathways involved in antimicrobial resistance, potentially guiding the development of innovative diagnostic techniques and therapeutic interventions to address the growing challenge of antimicrobial resistance.