Virulence prediction of intrahospital multi-drug resistant bacteria by machine learning

Klebsiella pneumoniae is a globally significant pathogen responsible for a wide range of infections, often complicated by multidrug resistance (MDR) and hypervirulent (hv) traits. Understanding the genomic architecture and distribution of virulence factors (VFs) is crucial to improve clinical management and surveillance. This study presents an integrative genomic analysis of 1038 K. pneumoniae isolates, focusing on the hierarchical organization of the virulome and its association with clinically relevant lineages. The VF panel was constructed by combining K. pneumoniae-specific genes from the Virulence Factor Database (VFDB) with curated entries from the literature, including both established and recently discovered factors. To capture potential horizontal gene transfers, non-K. pneumoniae VFs from other Enterobacteriaceae were also included. Sequence similarity searches were performed using MMseqs2 to compare each genome against the reference VF panel. Identity and coverage thresholds were applied to generate a binary matrix representing VF presence or absence per isolate. This matrix was used for unsupervised clustering (K-means and hierarchical) and resulting groups were post-labeled with clinical metadata and MLST to explore correspondence between virulome profiles and pathogenic lineages. Clustering revealed distinct virulome-based groupings, or virutypes, some of which corresponded closely with known clinical lineages, including hv ST23 and MDR-associated clones like ST147. The eight known MDR subgroups showed heterogeneous but partially conserved virulence profiles, suggesting subgroup-specific adaptation. Co-occurrence analysis identified functionally coherent VF modules, including mobile-associated clusters like Peg344, PagO, aerobactin complex and RmpA, and chromosomal modules like Kpi and FhaBC. Detection of Pla in ST490 highlights the role of horizontally acquired genes in shaping pathogenicity. While a supervised predictive model could not be completed, post hoc labeling showed rigid classification into hv, MDR or non-pathogenic groups oversimplifies the diversity. This study supports virulome profiling as a complementary epidemiological tool for refined surveillance and rapid diagnostics.