Reducing work-load of pathotype and gene detection in Klebsiella pneumoniae by leveraging machine learning

BackgroundThe growing prevalence of carbapenem resistance has caused an increasing number of bacterial isolates with multi-drug resistance features, especially in the Enterobacteriaceae family. Klebsiella pneumoniae, as one of the important members of the Enterobacteriaceae family, causes serious infections, which has attracted the attention of scientists due to the emergence of hypervirulent pathotypes with increasing antibiotic resistance and has been raised as a major concern worldwide. Early detection of this new super bacterium and its antibiotic resistance is of great help in reducing mortality and costs. The lack of new antibiotic options underscores the need to optimize current diagnostics. Therefore, this study was designed to leverage machine-learning approach for optimized selection of crucial antibiotics to reduce the experiments needed for the detection of pathotypes and genes presence in two classical and hypervirulent K. pneumoniae pathotypes. Methods341 non-duplicate clinical isolates of K. pneumoniae were collected from five university hospitals in Tehran and Qazvin, Iran. Pathotype differentiation of classical (cKp) and hypervirulent K.pneumoniae (hvKp) was done by PCR method by two molecular biomarkers including iuc and iut. After identifying the phenotypic antibiotic resistance, the presence of antibiotic resistance genes was detected by PCR method. Then, the relevance of resistance/susceptibility of the antibiotics and presence of pathotypes, aerobactin, and beta-lactamase genes was investigated and analyzed using five supervised machine learning algorithms by selecting crucial antibiotics through feature selection methods. ResultsAmong the 341 K.pneumoniae isolates, 102 and 239 isolates were hvKp and cKp respectively. The highest rate of antibiotic resistance after ampicillin (100%) was related to cefotaxime (76.2%) and the lowest rate of resistance was found in meropenem (24.3%). Imipenem, Meropenem, Aztreonam, Ceftazidime, Ceftriaxone, and Gentamicin are crucial antibiotics for detection of the pathotypes and the aerobactin genes. Moreover, Cefotaxime, Ciprofloxacin, Cefepime, Meropenem, and Imipenem are essential for detection of the beta-lactamase genes. ConclusionImplementing a machine learning approach including various feature selection methods and algorithms, results in less-required experiments on more limited antibiotics to detect genes and pathotypes. Our findings reveal that using machine learning in the prediction of the presence of genes and pathotypes of clinical isolates was a suitable method in terms of rapidity and cost-efficiency on top of accuracy.