Prediction of antimicrobial resistance in clinical Campylobacter jejuni isolates from whole-genome sequencing data

Campylobacter is a foodborne bacterial pathogen, which is able to cause gastroenteritis in humans and is found in the environment where it is widely distributed in the gastrointestinal tract of most warm-blooded animals [1]. Foodborne transmission is believed to be the main route of human infection [2, 3]. In 2018, Campylobacter continued to be the most common foodborne pathogen causing infections in Denmark with more than 4500 registered cases; most of these were caused by C. jejuni (85–95%) [4]. Clinical symptoms during infection often include diarrhoea, fever, abdominal pain, nausea, headache, and/or vomiting [5, 6] and in some cases result in Guillain-Barré syndrome (GBS) or reactive arthritis [7, 8]. Most patients recover without any antimicrobial treatment, whereas more severe and immunocompromised cases may need antibiotic treatment. Due to the extensive use of antimicrobials worldwide, within the production of food animals and human medicine, antimicrobial resistance (AMR) against clinically relevant antibiotics is emerging in Campylobacter spp. presenting a serious health concern [2, 9, 10]. In Denmark, AMR surveillance in Campylobacter jejuni has been in place since 1996 based on phenotypic susceptibility testing against ciprofloxacin (CIP), nalidixic acid (NAL), erythromycin (ERY), gentamicin (GEN), streptomycin (STR), and tetracycline (TET). The data on clinical isolates is reported as part of The Danish Integrated Antimicrobial Resistance Monitoring and Research Programme (DANMAP) (https://​www.​danmap.​org/​).

Campylobacter has developed several mechanisms for antibiotic resistance including point mutations, acquisition of resistance genes, and efflux systems [11‐13]. Resistance towards fluoroquinolones (FQs) in Campylobacter is mainly caused by point mutations in the DNA gyrase A (GyrA) region at multiple positions [10, 14]. The mechanisms involved in erythromycin resistance include ribosomal target modification and efflux of the antibiotic from the cell and this resistance is often associated with point mutation in domain V of the 23S rRNA, which is the target of macrolides, or the ribosomal proteins L4 and L22 [10, 15]. Campylobacter harbour three copies of the 23S rRNA gene, but only two copies need to be affected by point mutations to result in erythromycin resistance [16, 17]. Streptomycin resistance in Campylobacter is associated with substitution within the ribosomal RpsL protein or by expression of an aminoglycoside-modifying enzyme (ANT(6)-I) [18, 19]. Resistance towards gentamicin and tetracycline is in general conferred by the acquisition of resistance-associated genes. Resistance to tetracycline in Campylobacter is mainly conferred by the ribosomal protection protein tet(O), a plasmid-encoded gene, whereas the resistance gene aph(2″)-If often is associated with gentamicin resistance [20‐22].

The whole-genome sequencing (WGS) technology provides complete genomic DNA sequence of a bacterium, and WGS has proved to be an effective surveillance tool for foodborne pathogens [23, 24]. This method provides a large amount of information generated from a single assay [25] and allows multiple tests to be performed in silico simultaneously, whereas phenotypic susceptibility tests can be time-consuming, labour intensive, and are highly dependent on standardised laboratory procedures. Several studies have demonstrated that WGS can be used to identify resistant genotypes for different foodborne pathogens [23, 24, 26]. The aim of this study was to evaluate the concordance between WGS-based AMR prediction and the phenotypes identified from conventional antimicrobial susceptibility in a set of clinical C. jejuni isolates. The focus of the evaluation is to allow the replacement of phenotypic testing with a WGS-based surveillance of AMR in Campylobacter.

Resistance against clinically important antibiotics in Campylobacter continues to be a public health concern worldwide. Campylobacter is exposed to antibiotics in both animal production and human medicine making development and transmission of AMR possible. To assess the ability of using WGS to identify and predict AMR in C. jejuni, a dataset of 516 clinical isolates were analysed by comparing their resistance genotypes to their respective phenotypic resistance profile. The resistance profiles were found based on susceptibility to six antimicrobials all included in the Danish antimicrobial surveillance programme, DANMAP [2]. In this study, we evaluated and optimised the analysis of WGS data to identify—with high confidence—the relevant genetic markers that enable the prediction of the phenotypic resistance against specific antimicrobials.

For the genome-based prediction of AMR, we analysed WGS data by the use of mapping- and assembly-based bioinformatics methods to detect known AMR genes listed in a curated database, ResFinder. For detection of point mutations, we constructed a database containing reference genes and associated point mutations known from the literature. For each of the six antimicrobials tested, there was a concordance between 92 and 100% between the presence of known resistance genes or mutations and MICs above the ECOFF breakpoint. To obtain this high concordance, a number of modifications were made to the raw output from the ResFinder database search as well as the detected point mutations validated before inclusion in the final output presented here.

The four approaches to identify the presence of acquired resistance genes (modified KMA, ABRicate, ARIBA, ResFinder Batch Upload) worked well and with an overall good compliance between the results obtained, however with a few important differences that made us chose a modified KMA as the preferred method. A systematic error was attained when using three of the four methods for the detection of the mosaic gene tet(O/32/O) conferring resistance to tetracycline. Although tet(O/32/O) is present in the ResFinder database, only the mapping-based KMA method was able to detect this gene in 27 isolates, whereas neither of the assembly-based methods (ARIBA and ResFinder Batch Upload) detected tet(O/32/O) with a gene length > 90% in any of the genomes. This would potentially make an under-estimation of the occurrence of TET resistance as this gene was present in 23% of the TET-resistant genomes and it was widely distributed in our C. jejuni population. For all approaches, another systematic error was found to be related to frameshift mutations resulting in non-functional genes. We discovered 11 non-functional tet(O) genes resulting in a disagreement between phenotype and genotype. For ten of the isolates, the frameshift was due to a nucleotide deletion, whereas for the last isolate, the tet gene harboured an extra nucleotide compared with the reference. As a result, 3% of the TET sensitive isolates were wrongly determined as resistant. Such frameshift-related errors can be avoided by an automated frameshift check of the matched open reading frames or by a tblastx search.

Verification of markers responsible for resistance is essential for building a reliable pipeline. In this study, it was not possible to verify the significance of some point mutations described in the literature. This was the case for the ribosomal protein L22 (rplV) and the substitution P104S in gyrA associated with resistance against ERY and CIP/NAL, respectively. One mutation in L22 (A103V) was equally common in resistant and sensitive isolates, whereas another mutation in L22 (D72N) was only detected in a sensitive isolate. Thus, our findings suggest that D72N and A103V in L22 do not contribute to erythromycin resistance although these substitutions were reported to be the responsible marker for ERY resistance in one and two isolates, respectively [37, 42]. The P104S substitution in gyrA has been found previously in clinical isolates [43, 44], but our data could not support the association to fluoroquinolone resistance, as P104S was found either together with T86I in gyrA in resistant isolates or in a phenotypically sensitive isolate. These findings, summarised in Table 1, highlight the importance of being cautious when including suggested mechanisms of resistance in a routine analysis of predicting phenotypic resistance.

The most frequent resistance profile observed among the isolates was CIP/NAL resistance. Quinolone resistance in Campylobacter is most often due to mutations in the DNA gyrase A gene, gyrA, and especially substitutions at Thr86 are associated with high-level resistance to quinolones [32]. The substitution T86I was present in 97% of the quinolone resistant isolates and none of the sensitive isolates in our study. Two isolates carrying the T86I gyrA substitution showed high-level resistance to CIP but were fully susceptible to NAL. This phenomenon is described previously [45] and shows that the exact phenotype may not be possible to predict with the current knowledge. Another mutation, the more atypical T86A gyrA substitution, was observed in one isolate that — in accordance with previous studies [45‐47] — displayed high-level resistance towards NAL, but was susceptible to CIP. Double point mutations resulting in two amino acid substitutions in gyrA were also seen. In addition to the common T86I, 13 isolates had the P104S substitution and one isolate had the D90N substitution. The D90N substitution is less common and associated with moderate resistance [31‐33]; however, we detected D90N as the only marker of quinolone resistance in one of the CIP/NAL-resistant isolates.

Resistance to erythromycin is commonly mediated by mutations in 23S rRNA, especially A2075G [36, 48]. C. coli is more often resistant to macrolides than C. jejuni, and the ermB gene has been shown to mediate resistance to erythromycin in C. coli isolates of animal origin [49, 50], but recently ermB was also identified in a clinical C. jejuni isolate [51]. In this study, none of the isolates had the ermB gene. All isolates exhibiting phenotypic resistance against ERY had a mutation in the 23S rRNA genes at position 2074 or 2075, resulting in a 100% concordance between the phenotype and genotype. None of the isolates had any of the described point mutations in the ribosomal protein L4 (rplD).

When disregarding detection of the mentioned non-verified point mutations and the non-functional tet(O) genes, discrepancies between phenotype and genotypic prediction were seen for eight isolates, four for CIP/NAL, one for STR, and three for TET. In one of the TET-resistant isolates, a tet gene was clearly present as determined by the alternative methods. Otherwise, it is possible that the discrepancies were caused by unknown resistance mechanisms, sequence gaps in the genome preventing detection of known resistance genes or point mutations, or due to mixed cultures in the sample.

Based on our findings, we will establish a routine WGS-based analysis pipeline for AMR detection that we believe has a sufficiently high level of performance for replacing the phenotypic surveillance of AMR in clinical isolates of Campylobacter. Phenotypic susceptibility testing may still be optimal for clinical purposes as the WGS approach cannot deliver an MIC value. However, a sequence-based approach predicting MIC values from markers influencing the MICs may be clinically relevant and could be implemented in the database in the future. In our selected C. jejuni population, we only detected a limited number of genes and mutations that seem to be relevant for predicting the AMR profile. Thus, future routine analysis will only include the verified mutations and the acquired genes in the curated public databases with focus on avoiding the systematic errors encountered in this study, e.g. non-functional genes and detection of the mosaic tet gene. The pipeline should be maintained continuously. We were only able to detect a minority of the many reported point mutations, and therefore, it is relevant to include all potentially important mutations in an additional analysis and evaluate these by phenotypic testing of the isolates. Likewise, new genes or variants of genes may be added to the curated databases, and the performance of our detection pipeline should be evaluated for these. In addition, a subset of the isolates should be tested phenotypically to be able to detect new mechanisms, non-functional genes, and generally perform quality assurance on the analysis pipeline. This further evaluation and development can be done on small selected strain collections at regular intervals.