Genomic Features Associated with the Degree of Phenotypic Resistance to Carbapenems in Carbapenem-Resistant Klebsiella pneumoniae

Abstract

Carbapenem-resistant Klebsiella pneumoniae strains cause severe infections that are difficult to treat. The production of carbapenemases such as the K. pneumoniae carbapenemase (KPC) is a common mechanism by which these strains resist killing by the carbapenems. However, the degree of phenotypic carbapenem resistance (MIC) may differ markedly between isolates with similar carbapenemase genes, suggesting that our understanding of the underlying mechanisms of carbapenem resistance remains incomplete. To address this problem, we determined the whole-genome sequences of 166 K. pneumoniae clinical isolates resistant to meropenem, imipenem, or ertapenem. Multiple linear regression analysis of this collection of largely bla_KPC-3-containing sequence type 258 (ST258) isolates indicated that bla_KPC copy number and some outer membrane porin gene mutations were associated with higher MICs to carbapenems. A trend toward higher MICs was also observed with those bla_KPC genes carried by the d isoform of Tn4401. In contrast, ompK37 mutations were associated with lower carbapenem MICs, and extended spectrum β-lactamase genes were not associated with higher or lower MICs in carbapenem-resistant K. pneumoniae. A machine learning approach based on the whole-genome sequences of these isolates did not result in a substantial improvement in prediction of isolates with high or low MICs. These results build upon previous findings suggesting that multiple factors influence the overall carbapenem resistance levels in carbapenem-resistant K. pneumoniae isolates. IMPORTANCE Klebsiella pneumoniae can cause severe infections in the blood, urinary tract, and lungs. Resistance to carbapenems in K. pneumoniae is an urgent public health threat, since it can make these isolates difficult to treat. While individual contributors to carbapenem resistance in K. pneumoniae have been studied, few reports explore their combined effects in clinical isolates. We sequenced 166 clinical carbapenem-resistant K. pneumoniae isolates to evaluate the contribution of known genes to carbapenem MICs and to try to identify novel genes associated with higher carbapenem MICs. The bla_KPC copy number and some outer membrane porin gene mutations were associated with higher carbapenem MICs. In contrast, mutations in one specific porin, ompK37, were associated with lower carbapenem MICs. Machine learning did not result in a substantial improvement in the prediction of carbapenem resistance nor did it identify novel genes associated with carbapenem resistance. These findings enhance our understanding of the many contributors to carbapenem resistance in K. pneumoniae.

Keywords: Klebsiella pneumoniae; antibiotic resistance; carbapenem; machine learning; whole-genome sequencing.

FIG 1

Core genome phylogenetic tree of the 166 CR-Kp isolates included in the final analyses and corresponding genotypic/phenotypic isolate information. Sequence type and capsule locus type (if predicted) are listed next to each isolate. Carbapenem MICs to meropenem, ertapenem, and imipenem are indicated by the gradient from green (low MIC) to red (high MIC). The presence of carbapenemase or other β-lactamase genes is depicted by blue and orange boxes, respectively. The bla_KPC copy number is presented with a gradient from yellow (low) to red (high). The status of the ompK35, ompK36, and ompK37 outer membrane porin channel genes are designated by the dark purple (wild-type allele), light purple (allele with mutations of unclear significance), and gray (absent or truncated allele) boxes. Transposon isoforms are indicated in blue (Tn4401a), yellow (Tn4401b), green (Tn44401d), and orange (Tn4401e). The ertapenem breakpoints for susceptibility and resistance are ≤0.5 mg/liter and ≥2 mg/liter, respectively. The imipenem and meropenem breakpoints for susceptibility and resistance are ≤1 mg/liter and ≥4 mg/liter, respectively.

FIG 2

Distribution of MICs obtained from Etest for ertapenem (A), imipenem (B), and meropenem (C) for CR-Kp isolates that contain carbapenemases (blue) and those without carbapenemases (green). (D) The distribution of phenotypic MICs of all three carbapenems in each CR-Kp isolate. Circles represent isolates with the corresponding ertapenem, imipenem, and meropenem MICs; the color of the circle corresponds to the meropenem MIC, and the size of the circle corresponds to the number of isolates with that phenotype (e.g., the large brown circle in the upper right corner represents 11 CR-Kp isolates that have ertapenem, imipenem, and meropenem MICs of >32, >32, and >32 mg/liter, respectively).

FIG 3

Ertapenem (purple), imipenem (orange), and meropenem (green) MICs for each CR-Kp isolate stratified by the predicted function of OmpK35 (A), OmpK36 (B), and OmpK37 (C) outer membrane proteins. Black lines represent the median MIC for each group. MIC values of >32 mg/liter were converted to 64 mg/liter for analysis.

FIG 4

Ertapenem (purple), imipenem (orange), and meropenem (green) MICs for each CR-Kp isolate stratified by the type of transposon on which the bla_KPC gene was located. Black lines represent the median MIC for each group. MIC values of >32 mg/liter were converted to 64 mg/liter for analysis.

FIG 5

Relationship between the relative bla_KPC copy number in each CR-Kp isolate and its ertapenem (A), imipenem (B), and meropenem (C) MIC, where “r” represents the Spearman’s correlation coefficient. (D) Comparison of carbapenem MICs for isolates with fewer than 4 copies of the bla_KPC gene and those with ≥4 copies of the bla_KPC gene. Black lines represent the median MIC for each group. MIC values of >32 mg/liter were converted to 64 mg/liter for analysis.

FIG 6

Nested cross-validation performance of the support vector classifier machine learning approach to predict high versus low carbapenem resistance among ST258 K. pneumoniae isolates (n = 123). Levels of meropenem and imipenem resistance were predicted using models trained on curated (A and C, respectively) and whole-genome sequence features (B and D, respectively). Accuracy, sensitivity, specificity, positive predictive value (PPV), area under the receiver operating characteristic curve (AUC), and F1 scores were determined for models built in each cross-validation fold, with mean and 95% confidence intervals displayed in red.