Development and Evaluation of a Core Genome Multilocus Sequencing Typing (cgMLST) Scheme for Serratia marcescens Molecular Surveillance and Outbreak Investigations

ABSTRACT

Serratia marcescens can cause a range of severe infections and contributes to nosocomial outbreaks. Although whole-genome sequencing (WGS)-based typing is the standard method for molecular surveillance and outbreak investigation, there is no standardized analytic scheme for S. marcescens core genome multilocus sequence typing (cgMLST). Here, the development and evaluation of a S. marcescens cgMLST scheme is reported with the goal of enabling a standardized methodology and typing nomenclature. Four hundred ninety-one high-quality S. marcescens WGS data sets were extracted from public databases and—using the genomic sequence of NCBI reference strain S. marcescens Db11 (NZ_HG326223.1) as a starting point—all Db11 genes present in ≥97% data sets used to create a cgMLST scheme. The novel scheme was evaluated using WGS data from 24 outbreak investigations (n = 175 isolates) distributed over three continents. Analysis of Db11 genes within the 491 data sets identified 2,692 target genes present in ≥97% of genomes (mean, 99.1%; median, 99.9%). These genes formed the novel cgMLST scheme, covering 47.8% of nucleotides in the Db11 genome. Analyzing 175 isolates from 24 outbreaks using the novel scheme gave comparable results to previous typing efforts for both general groupings and allelic distances within clusters. In summary, a novel cgMLST scheme for S. marcescens was developed and evaluated. The scheme and its associated nomenclature will improve standardization of typing efforts for molecular surveillance and outbreak investigation, allowing better understanding of S. marcescens genomic epidemiology and facilitating interlaboratory comparisons.

INTRODUCTION

Serratia marcescens is a Gram-negative bacterium of the order Enterobacterales. Considered to be nonpathogenic prior to 1960, S. marcescens has since been shown to cause a range of severe infections, including bacteremia, meningitis, urinary tract infection, pneumonia, and endocarditis (1–5).

As this bacterial species can easily be transmitted in health care settings, clusters of nosocomial colonizations and infections have been reported in hospital settings, often involving neonates and intensive care units (6–9). In outbreak settings, early and accurate identification of sources and routes of transmissions is crucial to implement infection control measures and to prevent further nosocomial spread of S. marcescens. Hence, a technique is needed to enable high-resolution typing of this species.

While in the past, pulsed-field gel electrophoresis (PFGE) was the gold standard for outbreak investigations (10), currently, whole-genome sequencing (WGS)-based typing is used for most bacterial species, including S. marcescens (9, 11). WGS provides higher discriminatory power and better interlaboratory comparability than PFGE (12, 13).

Whereas WGS has been well established as subtyping technology in several non–S. marcescens bacteria, criteria for determining relatedness of S. marcescens isolates using WGS are poorly defined. Two main approaches have been used to assess relatedness of WGS data: (i) construction of phylogenies based on the extraction of single-nucleotide polymorphisms (SNPs); and (ii) extended gene-by-gene comparison in analogy to classical multilocus sequence typing (MLST), or so-called core genome MSLT (cgMLST), comparing hundreds of predefined target genes. While SNP analysis is a suitable method, isolate dependency and choice of a representative reference genome limit standardization and affect resolution (14). In contrast, cgMLST uses loci that are readily maintained, and therefore results can be shared among laboratories, enabling standardized nomenclature (15).

Although WGS-based typing has been successfully applied in investigating S. marcescens outbreaks using SNP typing (16), and whole-genome MLST (17) and local ad hoc cgMLST schemes have been described (18–20), there is no public cgMLST scheme for S. marcescens. Therefore, the aim of this study was to define and evaluate a novel cgMLST scheme for WGS-based typing of S. marcescens.

MATERIALS AND METHODS

S. marcescens cgMLST definition.

Using the so-called “soft” core genome definition approach (21), all public available S. marcescens whole-genome sequence data sets (as of 2019-09-23) were downloaded with the Illumina paired-end filter activated from the NCBI genome database (n = 532 data sets) and the sequence read archive (SRA, n = 755 data sets) to obtain a comprehensive collection for target scheme definition. SRA data sets were assembled with SKESA (22) with default settings, followed by mapping of raw reads to FASTA assembly contigs using the BWA-MEM algorithm of the Burrows-Wheeler Aligner (23), as implemented in SeqSphere⁺ version 7.0 beta (24). To ensure a defined and quality-controlled set of genomes for cgMLST definition, a thorough manual filtering process was applied. Duplicates were removed, since prominent or culture collection strains are often overrepresented and could thus induce bias. Where duplications occurred between the NCBI genome database and the SRA, preference was given to the SRA data to optimize the typing scheme for draft genome data, leaving 167 NCBI genome and 675 SRA data sets. To verify that all data sets used were within species boundaries, in silico DNA-DNA hybridization was performed with fastANI (25). Only records with an identity of ≥95% to the S. marcescens type strain ATCC 13880 (CP041233.1, genome status complete) were retained, leaving 121 NCBI genome and 508 SRA data sets. Finally, since assembly quality depends on the genome coverage used, SRA data sets with less than 70-fold SKESA assembled coverage were excluded from further analysis to reduce the impact of potential sequencing errors (26). In total, 491 data sets (121 NCBI genome; 370 SRA data sets) were accepted as the genome collection for cgMLST definition (Table S1).

The finished and publicly available genome of S. marcescens Db11 listed as the representative genome on the NCBI genome homepage (version1, GenBank accession no. NZ_HG326223.1, 2017-03-24) was selected as a seed genome—serving as a well-curated and -annotated genome sequence— as starting point for extraction of potential cgMLST targets. Using the cgMLST Target Definer tool (v. 1.5 with default parameters) from Ridom SeqSphere⁺ software (24), a preliminary cgMLST target list was defined that contained all genes from the seed genome that were not homologous, did not contain internal stop codons, and did not overlap other genes. Furthermore, genes that had a match with all available S. marcescens plasmid sequences (19 NCBI entries, as of 2019-10-21) were excluded (Table S2). All data sets of 491 genomes chosen for soft cgMLST definition were then analyzed with the preliminary cgMLST target list. Targets from the preliminary target list present in ≥97% of data sets were used for the final cgMLST typing scheme; the remaining targets were moved to a list of accessory targets.

Evaluation of the S. marcescens cgMLST scheme.

Recently published data from S. marcescens outbreaks in Germany, India, and the Netherlands (17, 19, 27), as well as unpublished cluster investigation data from four different U.S. cities (Table S3) were used to evaluate the novel cgMLST scheme. In all cases (in total 175 isolates were included), previous epidemiological investigations combined with various typing efforts (e.g., whole-genome MLST (wgMLST), single nucleotide polymorphism (SNP), PFGE, or matrix assisted laser desorption ionization-time of flight mass spectrometry (MALDI-TOF MS) based typing) were used to evaluate cgMLST typing results. In addition, we also extracted all SNPs within the cgMLST targets and used this information to further compare cgMLST- and SNP-based grouping of genotypes.

Raw reads from the published outbreaks were downloaded, and SKESA used for assembly, followed by BWA-MEM mapping implemented in SeqSphere⁺. For the remaining S. marcescens isolates (Table S3), DNA was extracted from cultured isolates using the Zymo Research Quick-DNA. Fungal/Bacterial Miniprep kit (Zymo Research Corporation, Irvine, CA) according to the manufacturer’s instructions. Subsequently, paired-end sequencing libraries were prepared using a NEBnext Ultra II Library Prep kit (New England Biolabs) with a targeted fragmentation size of 500 bp. Next generation sequencing was performed on an Illumina HiSeq 2500 (Illumina, Inc., San Diego, CA) using Rapid SBS V2 2 × 250-bp chemistry and targeting 200× depth of coverage.

For comparison, PFGE was performed on a subset of strains (Table S3). Cells in an overnight brain heart infusion broth culture were normalized to a turbidity of 0.45 to 0.55 using a turbidometer (Dade Behring, Deerfield, IL) and pelleted by centrifugation; the culture broth was removed and replaced with 500 μL of EET (100 mM EDTA, 10 mM EGTA, 10 mM TRIS, pH 8.0). Five hundred μL of sample agarose (SeaPlaque, Lonza, Basal, Switzerland) was added to create sample plugs. Encased cells were lysed in EET-LS (EET, 200 mg/μL lysozyme, 0.05% sarkosyl) for 4 h at 30°C. Restriction fragmentation was carried out using XbaI endonuclease (New England Biolabs, Ipswich, MA). PFGE was performed on a CHEF Mapper XA system (Bio-Rad, Hercules, CA). Ethidium bromide-stained gels were visualized on a Gel Doc XR system (Bio-Rad). Determination of clonal groups was made by comparison of banding patterns, with isolates with ≥1 band difference considered different in accordance with the Mayo Clinic laboratory method (28).

cgMLST-based analysis.

The quality filtered genome collection, the published outbreak data, as well as the newly sequenced outbreak isolates, were analyzed using the final cgMLST scheme. The deployed SeqSphere⁺ software called and automatically assigned alleles for each target, ensuring a unique nomenclature. The combination of all alleles in each isolate formed an allelic profile that was used to generate minimum spanning trees (MST) using the parameter “pairwise ignore missing values” during distance calculation. Links between MST nodes represent the allelic distance between the two connected genotypes. Distances were used to determine if outbreak isolates could be attributed to the same cluster and clearly separated from other clusters. The cluster cutoff value was defined as the maximum pairwise distance between epidemiologically linked isolates within well-defined outbreaks.

Data availability.

Raw reads generated were submitted to the European Nucleotide Archive (http://www.ebi.ac.uk/ena/) under accession no. PRJEB55016. The list of target genes and all allelic sequences are freely available under www.cgMLST.org.

RESULTS AND DISCUSSION

Development of the S. marcescens cgMLST scheme.

After removal of genes that were homologous, contained internal stop codons, or overlapped with other genes, a preliminary target list comprised 4,428 genes of the S. marcescens strain Db11 reference genome. Analysis of the presence of these genes within the 491 data sets of the quality-controlled and filtered genome collection identified 2,692 target genes present in ≥97% of all genomes. These targets formed the final cgMLST scheme, covering 47.8% nucleotides of the seed genome. The remaining 1,736 targets were present in <97% of the genomes and were thus moved to the list of accessory targets (Table S4). In addition, all targets and allelic sequences are also freely available at www.cgMLST.org. To ensure representativeness of the novel cgMLST scheme among the available S. marcescens population, the 491 data sets were reanalyzed with the reduced target set. This resulted in a mean and median percentage of detected targets of 99.1% and 99.9%, respectively.

To ensure a core genome-based typing scheme suitable for all strains of a species, it is important to consider genetic variability within the species. Here, the so-called “soft defined core genome” approach was applied (21) to infer the cgMLST target gene set. This approach is one way to define the “core genome” for typing, useful for pragmatic and reproducibility reasons, in contrast to traditional approaches defined on nucleotide and not protein levels; it also uses rather stringent search criteria. Therefore, the thus determined core genome is usually smaller than by traditional approaches. During the “soft” core genome definition, a very large collection of genomes containing hundreds to thousands of sequences is usually used and all genes present in ≥95% of analyzed genomes included in the cgMLST scheme (21). This approach is implemented in the BIGSdb Oxford and EnteroBase databases (29, 30) and is particularly useful when widely used standardized typing methods, such as MLST or serotyping, which usually reflect the diversity within a certain species, are not available, as is the case for S. marcescens. This contrasts with our previous efforts to develop cgMLST schemes (15, 31–34), where a “hard defined core genome” approach was used to define cgMLST schemes. Using this alternative approach, a limited number of genomes are included to define the cgMLST target set, and final target genes must be present in all genomes. As these genomes must therefore reflect whole-species genomic variability to achieve an optimal typeability, extensive preknowledge about the species under investigation to choose suitable representatives of the population, such as MLST data and other, frequently phenotypic data like serotypes is a basic requirement. This “hard defined core genome” approach is applied within the BIGSdb Pasteur database (https://bigsdb.pasteur.fr/) and the cgMLST Target Definer tool of the SeqSphere⁺ software.

Evaluation of the S. marcescens cgMLST scheme.

To evaluate the novel cgMLST scheme for detection of clusters during outbreak investigations, various outbreak data sets from both literature and new outbreak investigations were analyzed. The first group of cluster investigations originated from three European cities, where different outbreak scenarios (n = 14) were previously investigated using wgMLST (17). Applying the novel cgMLST scheme, results were comparable for both the general grouping within MSTs as well as maximum allelic distances within clusters to the central nodes and pairwise comparisons (Fig. 1). This is in line with other findings, where cgMLST and wgMLST also led to similar results (35). Interestingly, although the investigation of Rossen et al. targeted 9,377 genes, only roughly half of the genes were analyzed during the single outbreak investigations, questioning the use of wgMLST for longitudinal outbreak surveillance. Moreover, the various denominator of genes compared further complicates the interpretation of allelic distances among isolates. A recent comparison of SNP typing, wgMLST, and cgMLST for Pseudomonas aeruginosa (14) corroborated these findings with the inclusion of only approximately 40% of all possible wgMLST targets, resulting in nearly identical typing results between cgMLST and wgMLST among outbreak isolates. Blanc et al. additionally determined the benefit of a fixed in comparison to a variable wgMLST scheme during recombination events. In comparison to SNP typing, cgMLST exhibited higher concordance than wgMLST, explained by the fact that the wgMLST scheme included accessory genes and that their presence/absence might have an impact on the resulting phylogeny (14). Another benefit of cgMLST in comparison to wgMLST is the ability to form a stable nomenclature, which then can be used for the development for a hierarchical clustering (30).

FIG 1.

Minimum spanning trees of 110 S. marcescens isolates from Rossen et al. (17) using the novel cgMLST scheme. Distances are based on the allelic profiles of up to 2,692 target genes, pairwise ignoring missing targets. The values on the connecting lines indicate the number of allelic differences between the connected isolates. Circle sizes are proportional to the numbers of isolates per genotype (i.e., the allelic profile). Related genotypes (≤12 alleles distance) are shaded in gray. Each tree represents the typing results of the different given cities (A) Groningen, (B) Freiburg, and (C) Cologne, and the genotypes are colored according to the cluster colors from the original publication and named identically for comparability.

The second set of six outbreak investigations originated from different U.S. cities where PFGE was done to determine clonal relationships of epidemiological clusters (Table S3). All six PFGE based clusters were confirmed by grouping isolates into clusters with a maximum allelic distance of three alleles within a PFGE cluster (Fig. 2).

FIG 2.

Minimum spanning tree of 35 S. marcescens isolates using the novel cgMLST scheme from suspected clusters previously characterized by PFGE. Distances are based on the allelic profiles of up to 2,692 target genes, pairwise ignoring missing targets. Values on the connecting lines indicate the number of allelic differences between the connected isolates. Circle sizes are proportional to the numbers of isolates per genotype (i.e., the allelic profile). Related genotypes (≤12 alleles distance) are shaded in gray. The genotypes are colored according to the PGFE groups. Isolates with related PFGE types were grouped.

The third data set was from an outbreak of S. marcescens (19) initially analyzed by SNP typing (8). Again, the novel cgMLST scheme produced identical clustering results (Fig. 3) with seven of nine isolates clustered with at a maximum two alleles difference. The remaining two isolates (“patient-3” and “razor-1,” respectively) were clearly separated as had been demonstrated by SNP typing (see Fig. 1 in [8]).

FIG 3.

Minimum spanning tree of nine S. marcescens isolates using the novel cgMLST scheme from a suspected outbreak previously characterized by single nucleotide polymorphisms (SNPs) (19). Distances are based on the allelic profiles of up to 2,692 target genes, pairwise ignoring missing targets. The values on the connecting lines indicate the number of allelic differences between the connected isolates. Circle sizes are proportional to the numbers of isolates per genotype (i.e., the allelic profile). Related genotypes (≤12 alleles distance) are shaded in gray. Genotypes are named identically from the original publication for comparability.

The last outbreak investigations were initially characterized by MALDI-TOF MS and an ad hoc cgMLST scheme (27). MALDI-TOF MS is well established to determine bacterial species; however, the use of protein mass spectra for subtyping is debatable because of its discriminatory abilities may be lower than genotyping methods (36–42). Therefore, it was interesting to see that the novel cgMLST scheme was not only able to reproduce the previous findings but also that—with the exception of a single isolate (A6)—all MALDI-TOF MS-based clusters could be confirmed with at maximum four differing alleles within the three clusters (Fig. 4).

FIG 4.

Minimum spanning tree of 21 S. marcescens isolates using the novel cgMLST scheme previously characterized by MALDI-TOF MS clustering from suspected clusters (27). Distances are based on the allelic profiles of up to 2,692 target genes, pairwise ignoring missing targets. The values on the connecting lines indicate the number of allelic differences between the connected isolates. Circle sizes are proportional to the numbers of isolates per genotype (i.e., the allelic profile). Related genotypes (≤12 alleles distance) are shaded in gray. Genotypes are colored according to the MALDI-TOF MS pattern/cluster. Isolates with related MALDI-TOF MS types are grouped.

Finally, we extracted all SNPs present in the cgMLST targets to compare cMLST- and SNP-based grouping of genotypes. Here, both methods resulted in identical groupings (Fig. S1-S4), which is in concordance with previous investigations of other species such as P. aeruginosa, Salmonella Enteritidis, Listeria monocytogenes, Staphylococcus aureus, and Enterococcus faecium (14, 35, 43–45).

In summary, analysis of 24 cgMLST clusters from various outbreak investigations demonstrated the applicability of the novel cgMLST scheme to discern outbreak from non-outbreak isolates with sufficient discriminatory power. The number of differing alleles within a cluster, which reflects the extent of intraoutbreak microevolution, relates, at least in part, to duration of the outbreak (33). The numbers varied between 0 and 12 alleles between two isolates within a cluster, and between 0 and 8 allelic distances within a cluster to the central node (Fig. 1B). Although detailed epidemiological information was not available for all samples and patients investigated, it is reasonable and helpful to recommend a 12-allele cluster cutoff value to support data interpretation, i.e., isolate pairs that differ by ≤12 alleles should be investigated in-depth for a possible epidemiological relationship. This cluster cutoff value is shown in the figures as gray shadings.

Conclusion.

In conclusion, a novel cgMLST scheme for S. marcescens was developed and evaluated. This scheme will improve standardization of typing efforts for molecular surveillance and outbreak investigations to better understand the genomic epidemiology of S. marcescens, facilitating interlaboratory comparisons.