Core Genome Multilocus Sequence Typing and Prediction of Antimicrobial Susceptibility Using Whole-Genome Sequences of Escherichia coli Bloodstream Infection Isolates

ABSTRACT

In total, 50 Escherichia coli bloodstream isolates from the clinical laboratory and 12 E. coli isolates referred for pulsed-field gel electrophoresis (PFGE) were sequenced, assessed for clonality using core genome multilocus sequence typing (cgMLST), and evaluated for genomic susceptibility predictions using ARESdb. Results of sequence typing using whole-genome sequencing (WGS)-based MLST and sequence type (ST)-specific PCR were identical. Overall categorical agreement between genotypic (ARESdb) and phenotypic susceptibility testing for 62 isolates and 11 antimicrobial agents was 91%. Among the referred isolates, high major error rates were found for ceftazidime, cefepime, and piperacillin-tazobactam.

INTRODUCTION

The prevalence of antimicrobial-resistant (AMR) bacteria has increased worldwide, resulting in substantial morbidity, mortality, and health care costs. AMR Gram-negative bacilli are public health threats because they include rapidly expanding sequence types (STs) that harbor multiple resistance mechanisms and often have few effective therapeutic options.

Clinicians currently rely on phenotypic antibacterial susceptibility testing (AST) to guide treatment of culture-proven bacterial infections. However, phenotypic results using conventional methods are not typically available until 48 to 72 h after a culture turns positive. During this time, patients may be over- or undertreated for their infections, resulting in drug-associated toxicities or untreated infection and associated consequences. Molecular methods that rapidly detect antimicrobial susceptibility and resistance can enable more timely administration of effective, targeted treatment.

Our team has used whole-genome sequencing (WGS) and commercial analytics to assess clonality for outbreak investigations and, for Staphylococcus aureus and Klebsiella pneumoniae isolates, prediction of phenotypic AST results (1–5). We have incorporated into our routine clinical practice a WGS workflow that uses core genome multilocus sequence typing (cgMLST) (1, 2, 4–6). If WGS data could be generated rapidly (from isolates or, conceivably, directly from specimens) and results quickly translated to prediction of drug susceptibility, genomic analysis might be able to rapidly identify appropriate antimicrobial options for each patient on an individualized basis.

Here, we evaluated use of WGS to predict phenotypic resistance of the most common Gram-negative pathogen, Escherichia coli, using a commercial bioinformatics platform. We compared AST predictions from WGS with reference to phenotypic breakpoint agar dilution (AD) AST. We also compared the clonalities of isolates determined using WGS-based cgMLST and MLST determined by ST-specific PCR or pulsed-field gel electrophoresis (PFGE).

Analyses were conducted using 50 E. coli bloodstream isolates from the Antibacterial Resistance Leadership Group (ARLG), collected during a previous Mayo Clinic clinical trial evaluating a rapid blood culture diagnostic test (7), and 12 E. coli isolates referred to Mayo Clinic from a single institution for PFGE. As described in greater detail below, isolates underwent phenotypic AST, ST-specific PCR (ARLG isolates), WGS-based cgMLST (2,513 alleles), and MLST (7-allele Warrick scheme) using SeqSphere+ version 6 (Ridom Muenster, Germany).

Phenotypic AST was performed using AD and interpreted according to CLSI M100 guidelines (8). For the ARLG isolates, established multiplex PCR-based assays had been previously performed to identify E. coli phylogroups and STs (9–11).

For WGS, DNA was extracted and prepared using the Quick-DNA Fungal Bacterial kit (Zymo Research, Irvine, CA). WGS libraries were prepared with Nextera XT and sequenced on a MiSeq instrument utilizing v2 500-cycle chemistry (Illumina, San Diego, CA) with a targeted coverage of 200×. Read files were demultiplexed and underwent index and adapter removal using onboard MiSeq Reporter software. Read files were fed into the SeqSphere+ E. coli assembly and typing pipeline for cgMLST and MLST analysis.

FASTA assembly files were exported and uploaded into the ARESdb platform, release 2020-04 (Ares Genetics GmbH, Vienna, Austria), for genomic prediction of antimicrobial resistance. In brief, the platform used susceptibility/resistance (S/R) stacked classification models trained per species-compound pair on the AMR reference database ARESdb (3). The database contained genotype (WGS) to phenotype (AST) associations for more than 3,500 E. coli isolates. The model stacks consisted of extreme gradient boosting, elastic net regularized logistic regression (ENLR), and set covering machine models. Unlike conventional gene-based resistance prediction models, ARESdb models were trained on informative sequence motifs, i.e., feature-selected count matrices of overlapping k-mers, and combined by ENLR as described by Lüftinger et al. (12). A total of 11 compound-specific stacked classification models were used for genomic prediction of antimicrobial resistance on uploaded assembly files. Very major error (VME), major error (ME), and minor error (mE) rates were defined following CLSI document M52 (13). Intermediate phenotypes were treated as minor errors because ARESdb did not predict the intermediate interpretive category.

Among the 50 E. coli bloodstream isolates, ST-specific PCR identified eight STs in 33 isolates, as follows: ST131, 12 (24%); ST95, 8 (16%); ST127, 2 (4%); ST73, 4 (8%); ST81/550, 1 (2%); ST141, 2 (4%); ST69/CGA, 3 (6%); and ST405, 3 (6%). WGS-based Warrick MLST by SeqSphere+ corresponded perfectly with ST-specific PCR. Additionally, WGS-based MLST characterized the remaining 17 isolates as belonging to unique STs that were not included in the ST-specific PCR panel.

Both isolate collections were also analyzed using cgMLST. Among the ARLG isolates, there was no evidence of high-level relatedness; all isolates exhibited at least 17 allelic differences (Fig. 1, left). Among the 12 referred isolates, 9 isolates (all of which were indistinguishable by PFGE) differed by 0 or 1 allele. The 3 remaining isolates (which differed from all other isolates by ≥3 PFGE bands) differed from other referred isolates by ≥31, ≥43, and ≥2,389 alleles (Fig. 1, right) (14). Overall, these results show that SeqSphere+ can accurately determine the ST of E. coli isolates, and that SeqSphere+ cgMLST analysis provides similar results to those provided by PFGE, as shown previously with Staphylococcus aureus, Klebsiella pneumoniae, and Clostridioides difficile (1, 2, 4–6). Analysis of WGS data has become standard for assessment of clonality in clinical microbiology laboratories. If cgMLST could be performed rapidly and in real time, it might be useful for prompt recognition of outbreaks. At present, however, the turnaround time, as well as the cost, of WGS is a challenge to its widespread, routine use.

FIG 1.

Minimum spanning tree showing phylogenetic relationships (per core genome multilocus sequence typing [cgMLST]) among 50 Antibacterial Resistance Leadership Group (ARLG) Escherichia coli isolates (left) and 12 referred isolates (right). Each isolate is designated with a unique number. The lines between the circles show the numbers of allelic differences.

Among the 50 ARLG bloodstream isolates, phenotypic antimicrobial resistance prevalence, as determined by AD, was 52% to ampicillin, 30% to ciprofloxacin, 14% to ceftriaxone, and 28% to 3 or more drug classes (Table 1). In contrast, all 12 referred isolates were multidrug resistant (MDR), with resistance to 3 or more drug classes, including to ampicillin, ceftriaxone, and ciprofloxacin and levofloxacin (Table 1). Among the ARLG isolates, agreement between genotypic and phenotypic resistance determinations was high, ranging from 90% to 100% for all drugs except ampicillin-sulbactam (68%), which exhibited a high mE rate (20%) (Table 2). Overall agreement was 94%, with VMEs in 3 (1%), MEs in 16 (3%), and mEs in 15 (3%) (Table 2). In contrast, for the 12 referred isolates, agreement between genotypic AMR prediction and phenotypic AST was high for all drugs except ceftazidime, cefepime, and piperacillin-tazobactam. These three antibiotics had high associated ME rates, with genotypic predictions overcalling phenotypic resistance to ceftazidime for 75% of isolates and that to cefepime for all isolates (Table 2). Among these 12 isolates, overall agreement was 80%, and VME, ME, and mE rates were 0%, 17%, and 2%, respectively. Nine of these isolates were found to contain a β-lactamase gene in the OXA-1 family, detected by querying the WGS data in the PATRIC AMR database (15). OXA-30, a member of the OXA 1 family, has previously been reported to possess enhanced activity against cefepime (16). Furthermore, 2/12 possessed TEM family alleles, and all 12 possessed a multitude of AMR-associated efflux pumps. The clonality of this collection limits the generalizability of this finding.

TABLE 1.

Antimicrobial resistance by agar dilution among two collections of Escherichia coli isolates

Antibiotic	No. of resistant isolatesa (column %)
	ARLG (n = 50)	Referred (n = 12)
Ampicillin	26 (52)	12 (100)
Ampicillin-sulbactam	21 (42)	11 (92)
Cefepime	5 (10)	1 (8)b
Ceftazidime	5 (10)	3 (25)
Ceftriaxone	7 (14)	12 (100)
Ciprofloxacin	15 (30)	12 (100)
Gentamicin	7 (14)	0 (0)
Levofloxacin	16 (32)	12 (100)
Meropenem	0 (0)	0 (0)
Piperacillin-tazobactam	3 (6)	1 (8)
Trimethoprim-sulfamethoxazole	11 (22)	10 (83)
≥3 classesc	14 (28)	12 (100)

^aIntermediate isolates were considered resistant.

^bSusceptible dose-dependent isolates were considered susceptible (n = 4).

^cCiprofloxacin and levofloxacin were considered a single drug class.

TABLE 2.

Agreement between phenotypic antimicrobial susceptibility and genotypic predictions among 62 Escherichia coli isolates

Isolate collection and antibiotic	Phenotype-genotype comparison (no. of isolates [row %])
	Agreement	Minor error	Major error	Very major error
ARLG isolates (N = 50)
Ampicillin	45 (90)	1 (2)	3 (6)	1 (2)
Ampicillin-sulbactam	34 (68)	10 (20)	5 (10)	1 (2)
Ceftriaxone	50 (100)
Trimethoprim-sulfamethoxazole	49 (98)		1 (2)
Ciprofloxacin	47 (94)		3 (6)
Ceftazidime	49 (98)	1 (2)
Gentamicin	50 (100)
Levofloxacin	47 (94)		2 (4)	1 (2)
Cefepime	48 (96)a	1 (2)	1 (2)
Meropenem	50 (100)
Piperacillin-tazobactam	47 (94)	2 (4)	1 (2)
Referred isolates (N = 12)
Ampicillin	12 (100)
Ampicillin-sulbactam	11 (92)	1 (8)
Ceftriaxone	11 (92)	1 (8)
Trimethoprim-sulfamethoxazole	12 (100)
Ciprofloxacin	12 (100)
Ceftazidime	3 (25)		9 (75)
Gentamicin	12 (100)
Levofloxacin	12 (100)
Cefepime	1 (8)		11 (92)
Meropenem	12 (100)
Piperacillin-tazobactam	8 (67)	1 (8)	3 (25)

^aOne isolate was phenotypically susceptible dose-dependent by agar dilution testing and had a susceptible genotype; this was counted as agreement.

Compared to gold-standard phenotypic AST using AD, WGS prediction of susceptibility was accurate for drug-susceptible E. coli isolates but had unacceptably high ME rates for ampicillin-sulbactam among the ARLG bloodstream isolates and for piperacillin-tazobactam, ceftazidime, and cefepime among the referred MDR isolates. High ME rates may result from a failure to predict inhibitor activity for the β-lactam–β-lactamase inhibitor combinations. Notably, ceftazidime and cefepime showed ME rates of 0% and 2%, respectively, in the ARLG bloodstream isolates, even though 90% of isolates were susceptible. Differences in ME rates between ARLG isolates and referred isolates may be due to the considerable cgMLST relatedness of 9 of the 12 referred isolates (Fig. 1, right, group 1).

Possible reasons for discrepancies between WGS prediction and phenotypic AST include variability in phenotypic AST results, lack of expression of resistance genes, loss of resistance genes during subculturing, and limitations of the bioinformatic pipeline. Phenotypic resistance can vary with the level of resistance gene expression, which can differ because of alterations in promoters, repressors, or other regulatory mechanisms and may be hard to predict using existing databases. This may be especially relevant for MDR isolates and those categorized as susceptible dose-dependent (i.e., to cefepime). A recent study of isolates from respiratory specimens also found similar overall agreement of 89% across 16 species (95% for E. coli specifically) using ARESdb to predict AST, with high ME and VME rates for β-lactam–β-lactamase inhibitor combinations and cephalosporins. Findings from the current study similarly show that the ARESdb system needs to be improved for genotypic resistance prediction for select drugs in E. coli. Here, ARESdb was evaluated based on release 2020-04; since the system’s content and algorithms are updated regularly, comparisons to other studies that used this analytic must account for the version used.

This study has several limitations, including the relatively small sample size; clonality of most of the referred MDR isolates; the absence of carbapenem-resistant isolates; the scarcity of isolates resistant to ceftazidime, gentamicin, or cefepime; the use of AD as the sole phenotypic AST method; and the use of a single resistance prediction application. Despite these limitations, this study demonstrates that although WGS-based technologies are promising, they cannot yet replace phenotypic AST of E. coli, especially for predicting resistance in highly resistant strains that harbor a multiplicity of complex resistance mechanisms and variable gene expression patterns. Furthermore, WGS remains costly in comparison to conventional AST methods, and total testing time is roughly equivalent. Time to receive results is laboratory/resource dependent.

In conclusion, SeqSphere+ was an accurate method for high-resolution clonality assessment based on WGS data. The ARESdb WGS AMR prediction platform showed 91% overall agreement in E. coli isolates (with agreement of 100% for some antibiotics) but requires enhancement for accurate prediction of resistance across all antibiotics. Development of more sophisticated resistance prediction models and of less expensive and faster methods for WGS will help advance the clinical use of WGS for bioinformatic prediction of AST.

Data availability.

Whole-genome sequencing read data and isolate antibiogram data were deposited in the National Center for Biotechnology Information archives under BioProject identifier PRJNA736958.