Metagenomic identification of disease-causing Salmonella enterica serovars and antimicrobial resistance genes from paediatric faecal samples

Abstract

Background. Nontyphoidal Salmonella (NTS) is a common cause of enterocolitis and a major cause of death in children in low- and middle-income countries (LMICs). High antimicrobial resistance (AMR) prevalence in LMICs reduces treatment options for individuals at risk of severe infections.

Methods. We investigated the use of metagenomics to identify NTS and associated AMR genes in 28 faecal metagenomes from children with culture-confirmed salmonellosis in Vietnam, using accompanying NTS genomes from isolated serovars (one per metagenome). Read-based and assembly-based methods were utilised for NTS and AMR detection. Case metagenomes were compared to healthy control metagenomes (n=21) with respect to the microbiome composition, NTS relative abundances, number of unique AMR genes and antimicrobial classes to which the genes confer resistance, including classes used in Salmonella treatment.

Results. Salmonellosis cases displayed significantly higher relative abundances of Enterobacteriaceae than controls. Bracken and Centrifuge analysis facilitated the identification of Salmonella enterica sequences in case metagenomes at varying relative abundances (0.00259–27.7 % of total reads), which were significantly higher than controls. MetaPhlAn4 did not detect S. enterica in any control metagenomes, though 12 case metagenomes were also negative. The isolated serovars were identified in 78.6% of the associated case metagenomes with Centrifuge, suggesting this method is the most sensitive; however, the isolated genome serovar was the most abundant in only six case metagenomes, and serovar sequences were also identified in control metagenomes. Alignment to a Salmonella reference database, followed by local assembly and realignment, predicted the isolated serovar as the most likely serovar present in 35.7% of metagenomes, whereas Salmonella in silico typing resource classification of the local assembly was concordant with the isolate genome in 28.6% of cases. Metagenome-assembled genomes produced using two tools following de novo assembly identified the isolated serovar in 17.8–21.4% of cases. The percentage of NTS AMR genes identified in each case metagenome ranged between 0.00 and 100%. There was no significant difference in the number of unique AMR genes or antimicrobial classes between cases and controls, indicating comparable resistomes between cohorts.

Conclusions. This study highlights the potential of metagenomics for NTS identification in faecal samples, although overlap in S. enterica relative abundance between cohorts calls for further work to identify a diagnostic cutoff. Reliable characterisation of the organism to the serovar and AMR genotype level is affected by the complexity of the microbiome, sequencing and analysis approaches. Increased sequencing depth, for example through improved host DNA depletion, may facilitate enhanced characterisation. Detection of multiple serovars within individual samples with the Centrifuge suggests inaccurate classification or the presence of multiple serovars, making characterisation difficult.

Impact Statement.

Nontyphoidal Salmonella (NTS) is a common cause of diarrhoea. Paediatric infections are associated with higher rates of complications and severe disease, particularly in low- and middle-income countries like Vietnam. Clinical isolates of NTS in Vietnam often show high levels of antimicrobial resistance (AMR), which can limit treatment options in severe disease. Shotgun metagenomic sequencing allows examination of the whole microbial community within samples and could become the method of choice for diagnosis and epidemiologically relevant typing of infectious intestinal disease. It is important that the databases used to interpret these data are populated with DNA sequences representing different communities. The examination of the whole microbial community within stool samples from Vietnam is an important part of the process to potentiate direct identification and characterization of pathogens and microbiome patterns. This study assessed the use of metagenomics for the identification and characterization of NTS from faecal samples from children with confirmed salmonellosis (case metagenomes) and compared the results to healthy controls. The cases were marked with statistically significant differences in the microbiome compared to controls. The percentage of the sample classified as Salmonella enterica was also significantly higher in the case metagenomes compared to the controls. Characterisation of the serovar isolated from the case samples was possible in up to 78.6% of case metagenomes, but identification of low-abundance NTS serovar sequences in the control metagenomes may indicate low specificity of existing analysis approaches. NTS-associated AMR genes were identified in most case metagenomes, although many of these were likely carried by multiple organisms. There was no significant difference in the quantity of unique AMR genes or antimicrobial classes to which they confer resistance between the case and control metagenomes, indicating that although metagenomics can facilitate NTS identification in faecal samples, reliable characterisation for treatment guidance remains difficult.

Data Summary

The authors confirm that all supporting data have been provided within the article or through supplementary data files.

Introduction

Salmonella enterica are Gram-negative, facultative anaerobic bacilli in the Enterobacteriaceae family. Of the six subspecies described, S. enterica subspecies enterica is the most diverse, currently comprising over 1,500 serovars [1,2], many of which are clinically relevant. Contaminated food and water are the leading sources of nontyphoidal salmonellosis [3]. In comparison to the systemic disease associated with typhoidal S. enterica, nontyphoidal Salmonella (NTS) infections are generally characterised by mild to moderate enterocolitis [4]. However, paediatric NTS infections are associated with higher complication rates and subsequent invasive disease, particularly in low- and middle-income countries (LMICs), where diarrhoea is a leading cause of death in children [5].

Vietnam is an LMIC burdened with NTS infections in children [6]. Common clinically relevant serovars include Salmonella Typhimurium, Salmonella Weltevreden and Salmonella Stanley. S. Typhimurium, in particular, is often identified with high levels of antimicrobial resistance (AMR), resulting in nonsusceptibility to clinically important drug classes including beta-lactams, phenicols, fluoroquinolones and macrolides [7]. High antimicrobial consumption in Vietnam may have led to high levels of AMR [8]. Multidrug resistance (MDR) in NTS is particularly concerning, as resistance to three or more antimicrobial classes can result in reduced options for treatment of invasive disease and affect future therapeutic outcomes [9].

Culturing of faecal samples for pathogen identification and characterisation is common in clinical settings [10]. More recently, genomic approaches using sequence data from isolates have facilitated improved pathogen typing and prediction of AMR, reducing the dependency on phenotyping [11]. However, culturing and sequencing can be time-consuming, expensive and not always feasible [5]. Metagenomic sequencing, which involves the culture-free extraction and sequencing of total microbial DNA from samples, permits an examination of the whole microbial community within the sample and can give useful indications for potential causative agents of diarrhoea. Salmonella infections have been shown to be associated with a reduction in the gut microbial diversity and an increase in the abundance of Enterobacteriaceae, which in itself may generate a diagnostic signal, but further studies have been called for to better understand the interaction of Salmonella with the gastrointestinal microbiota [12]. Sequencing to sufficient depth can also facilitate the identification of specific causative agents of diarrhoea and the bacterial community resistome to guide treatment.

Here, we aimed to investigate the effectiveness of metagenomics for the identification of NTS, associated AMR genes and microbial community differences in faecal metagenomes collected from children with culture-confirmed salmonellosis in Vietnam. A variety of different informatics approaches were compared to attempt to identify the culture-confirmed Salmonella serovar from each metagenome: taxonomic profiling of metagenome reads, alignment of reads to a Salmonella reference database followed by local assembly and realignment, Salmonella in silico typing resource (SISTR) classification of local assemblies, and de novo assembly and metagenome-assembled genome (MAG) identification and classification. AMR genes were also identified from the reads and MAGs. The metagenomes from infected cases were compared to those from healthy controls with respect to the relative abundances of NTS, microbial composition and diversity and the resistome. We conclude that whilst S. enterica relative abundance was significantly higher in cases compared to controls, there is an overlap that makes determining a diagnostic cutoff difficult. Detection of serovars and Salmonella-specific AMR genes in the metagenomes is challenging, requiring higher sequencing depth to allow for more accurate diagnostics.

Methods

Sample collection and ethical approval

This study investigated 28 stool samples from paediatric culture-confirmed cases of salmonellosis. Samples were collected at Children’s Hospital No. 1 in Ho Chi Minh City, Vietnam, from children under 5 years of age presenting with symptoms of enterocolitis. Sample collection took place either on the day of admission (n=15; before any antibiotic treatment) or up to 3 days after admission (n=13) and thus possibly after antibiotic treatment was initiated. Ethical approval for this study was provided by the institutional review board of the hospital and the University of Oxford Tropical Research Ethics Committee (OxTREC no. 1045–13). The metadata associated with the samples can be accessed in Table S1 (available in the online Supplementary Material).

Salmonella isolation

All faecal samples were cultured on MacConkey agar (MC agar; Oxoid) and xylose-lysine-deoxycholate agar (Oxoid), as well as in selenite broth (Oxoid) at 37 °C for 18–24 h. Salmonella isolates (n=28; one per sample) were identified based on their distinctive appearance on xylose-lysine-deoxycholate and MC agar, and their identification was confirmed using MALDI-TOF MS (Bruker) and API20E (bioMérieux), following the manufacturer’s instructions.

DNA extraction, whole-genome sequencing and metagenome sequencing

The genomic DNA was extracted from all 28 isolates using the Wizard Genomic DNA Purification Kit (Promega), and whole-genome sequencing was performed at the Wellcome Sanger Institute. Extracted DNA of Salmonella isolates was sequenced on the Illumina HiSeq X Ten platform according to the manufacturer’s protocols to generate 150 bp paired-end reads.

Metagenome DNA from faecal samples was extracted using the FastDNA Spin Kit for Soil (MP Biomedicals), according to the manufacturer’s instructions. Metagenome sequencing was also performed at the Wellcome Sanger Institute, using the Illumina HiSeq 2500, generating 125 bp paired-end reads.

The sequence data are available in the National Center for Biotechnology Information (NCBI) Sequence Read Archive and European Nucleotide Archive (ENA) under project accession PRJEB21259.

Genome analysis

Analyses were performed for all 28 isolate genomes using Galaxy [13], a QIB cloud server (based on Cloud Infrastructure for Microbial Bioinformatics [14]) and the Norwich Research Park High Performance Computing cluster.

The Illumina paired reads were trimmed using fastp v0.23.2 [15]. Reads were assembled with Shovill v.1.1.0+galaxy0 (https://github.com/tseemann/shovill) using the SPAdes assembler [16] with an estimated genome size of 5 Mbp and minimum coverage to call part of a contig set to 0 (AUTO), and assembly quality was assessed with QUAST v5.0.2 [17] and CheckM v1.0.11 [18]. BWA-MEM Galaxy v0.7.17.1 [19,20] and CoverM v0.3.2 (https://github.com/wwood/CoverM) were used to estimate contig coverage. Assemblies that displayed >30× coverage, <500 contigs over 500 bp and <50 duplicate marker genes were accepted.

Assemblies were annotated with Prokka v1.14.5 [21], and Roary v3.13.0 [22] (95% identity threshold for blastp and 99% threshold for the percentage of total isolates a gene must be in to be considered a core gene) was used for core gene alignment. This was used to construct a maximum likelihood tree in IQ-TREE v1.6.11 [23] using a general time-reversible model on the web server (https://www.hiv.lanl.gov/content/sequence/IQTREE/iqtree.html) with ultrafast bootstrap approximation [24] and sH-like approximate likelihood ratio test (1,000 replicates) [25].

The species, subspecies, serovar and sequence types (STs) were determined with SISTR v1.0.2 [26] and MLST v2.16.1 (https://github.com/tseemann/mlst), respectively. AMR determinants were identified using KMA v1.4.3 [27] and the ResFinder database [28] with 90% query identity and template coverage thresholds. The aac(6′)-Iaa gene was considered to confer an aminoglycoside-resistant genotype, although this is a cryptic gene in Salmonella and may not typically confer phenotypic resistance [29]. StarAMR v0.4.0 was used to identify fluoroquinolone resistance mutations with the PointFinder database [30], using 90% blast identity and overlap thresholds.

Metagenome analysis

Dataset description and read processing

The Illumina paired reads of the ‘case’ faecal metagenomes (n=28) were each associated with one isolate genome from the ‘Genome analysis’ section (Table S2). The metagenome read files for individual runs were concatenated (merged) to obtain one pair of files per sample prior to analysis. ‘Control’ faecal metagenomes (n=21) from healthy children were obtained from a study by Pereira-Dias et al. [31] and downloaded from the ENA project PRJEB22032 (Table S3). These control samples were obtained from the Oxford University Clinical Research Unit at the Hospital for Tropical Diseases, also in Ho Chi Minh City, from children aged up to 5 years.

Reads were trimmed using fastp v0.23.4 and host reads removed with Hostile v0.1.0 [32] using the T2T-CHM13 v2.0+IPD-IMGT/HLA v3.51 human reference genome. Further analysis was performed on the trimmed, host-depleted reads.

Classification of reads for family-, genus- and species-level metagenome profiling, identification of S. enterica and S. enterica serovars

The reads were classified with Kraken2 v2.1.1+galaxy0 [33] (k2_nt_20230502 with 0.1 confidence level) and the abundance of species, genera and families estimated with Bracken v2.8 [34] (100mer k-mer distribution, read threshold to include as part of Bracken report set to 10). The 20 most abundant species and genera and 10 most abundant families in the metagenomes were determined using the Bracken outputs, by dividing the new_est_reads in Bracken reports by the number of overall metagenome reads obtained from Kraken reports (sum of unclassified and root), and multiplying by 100 to obtain percentages. S. enterica relative abundances were calculated in the same way.

Reads were also classified with Centrifuge v1.0.3 (nt_2018_3_3 database, with -k set to 1) [35]. The Centrifuge outputs were converted to Kraken-style reports using Centrifuge k-report to quantify S. enterica reads. The percentage of the overall faecal microbiome represented by S. enterica was calculated as described above.

MetaPhlAn4 v4.0.3 (mpa_vJan21_CHOCOPhlAnSGB_202103) [36] was also used for classification, as an alternative marker gene-based method. MetaPhlAn outputs were combined into one report using the merge_metaphlan_results.py script, and the S. enterica relative abundances were reported directly.

Centrifuge results were also used to identify reads representing the isolated S. enterica serovar for the case metagenomes, and the most abundant S. enterica serovar for the case and control metagenomes. For the detection of monophasic Salmonella Paratyphi B var. Java, the serovar was considered to be present regardless of monophasic or biphasic S. Java identification. However, for the monophasic variant of S. Typhimurium (I 4,[5],12:i:-), the metagenomes were specifically screened for S. 4,[5],12:i:-, as this variant was present in the databases used. The relative abundance of S. enterica serovars was calculated in the same way as species from the Centrifuge k-reports.

Alpha and beta diversity analysis

The alpha and beta diversities of the case and control metagenomes were assessed using the genus-level Bracken reports as input to make an operational taxonomic unit (OTU) table using R v4.2.3 [37]. Bracken by default only retains genera with at least 10 reads; thus, only those genera were included. NA values representing either absent taxa or taxa with fewer than 10 reads were changed to 0 for the analysis.

For alpha diversity, the metagenomes were rarefied to the smallest library size (28,644) with ‘rrarefy’ from vegan v2.6–4 [38]. Observed richness was calculated manually by counting the number of genera with non-zero values in each metagenome. Shannon and Simpson diversity indices were calculated using vegan.

Beta diversity analysis was also performed using vegan. The OTU table was transformed such that each genus abundance was divided by the sum for the sample to obtain relative abundance. This was used to calculate Bray–Curtis distance, which was used for non-metric multidimensional scaling (NMDS) in two dimensions (trymax=100).

Alignment of reads to the Salmonella database

NCBI (https://www.ncbi.nlm.nih.gov/) was searched for Salmonella reference genomes representing different serovars; 117 were identified and used to form a database after removing plasmid contigs. BBsplit v38.75 (https://github.com/BioInfoTools/BBMap/blob/master/sh/bbsplit.sh) was used to align the reads of each metagenome against this database and extract the Salmonella reads, which were assembled using SPAdes v3.14.1. The assembled contigs were aligned back to the database with Nucmer v3.1 [39] using 95% identity and coverage thresholds. The proportion of each chromosome in the database to which the contigs aligned was calculated. The serovar displaying the highest coverage was defined as the most likely serovar present. The assemblies were also analysed with SISTR core genome multilocus sequence typing (cgMLST) to determine the most likely serovars present.

Assembly of contigs and MAGs

Reads were assembled into contigs using MEGAHIT v1.1.2 [40] and coverage estimated with Bowtie2 v2.3.4.1 [41] and CoverM. MAGs were assembled using Metabat2 v2.14 [42] and Maxbin2 v2.2.4 [43]. MAG quality was assessed with QUAST [17] and CheckM [18]. BAT v5.0.3 [44] was used to identify Salmonella MAGs; the serovar of these was then classified with Kmerfinder v3.0.2+galaxy0 (https://cge.food.dtu.dk/services/KmerFinder/) using the bacteria database (compiled 17 October 2021) [27,45, 46]. Salmonella MAGs displaying less than 10% completeness were discarded.

Detection of AMR genes in metagenome reads and MAGs

KMA was used for AMR gene identification, using ≥90% query identity and ≥60% template coverage thresholds. The isolate genome and metagenome KMA results were compared to determine whether or not S. enterica AMR genes were identified in the associated metagenomes. AMR gene variant names were collapsed at the root, such that multiple variants were considered as one unique gene and a gene was considered to be present in the associated metagenome regardless of the specific variant. gyrA mutations were excluded when calculating the percentage of isolate genome AMR determinants identified in the associated metagenome, as these could not be identified in the metagenomes. The relative abundance of each AMR gene in the metagenome was calculated by dividing the number of bases associated with the gene (calculated by multiplying the KMA template length and the depth) by the total number of bases associated with AMR genes in that sample and then displayed as a percentage. The metagenomes were then screened for the AMR genes identified in the Salmonella isolate genomes. Relative abundances of these genes were summed up to determine the relative abundance of S. enterica AMR genes in each metagenome. For genes with multiple variants, the proportions of all variants were added up. The classes of antimicrobials to which the genes confer resistance were reported according to the Comprehensive Antibiotic Resistance Database (CARD) [47] (Table S4) to identify AMR gene classes, except for beta-lactamase genes, for which the antimicrobial class was simplified to ‘beta-lactams’. The drug class for mcr genes was changed to polymyxins instead of the peptide classification given by CARD, as this is a more distinct representation of the resistance class [48]. For genes that were not present in the CARD database, the AMR gene classes were obtained from published literature. The gene classes were summarised as ‘multiple’ for genes conferring resistance to more than one antimicrobial class. The class summary was used for plotting or analysis unless stated otherwise.

The accepted Metabat2 and Maxbin2 MAGs were analysed with ABRicate v0.9.7 (https://github.com/tseemann/abricate) using the ResFinder database and 90% identity and 60% coverage thresholds. MAG contigs containing AMR genes were mapped against publicly available sequences with blastn v2.15.0+ (megablast) [49] using the nucleotide collection (nr/nt, updated 12 January 2024) [50] on the online server (https://blast.ncbi.nlm.nih.gov/Blast.cgi) with default parameters. Contigs containing aac(6′)-Iaa genes were not mapped as these genes are considered cryptic (not conferring phenotypic resistance) in Salmonella [29]. From the first 100 chromosome and plasmid hits, those with ≥90% identity were reported.

Statistical analysis

Statistical analysis was performed in R. The relative abundance of S. enterica and Enterobacteriaceae in the case and control metagenomes, the percentage of metagenome reads aligned to the custom Salmonella database and the percentage of reference chromosome covered following assembly and realignment (top result per metagenome), as well as the number of total AMR variants, unique AMR genes and number of antimicrobial classes to which the genes confer resistance, were compared using Mann–Whitney U tests. The alpha diversity indices of case and control metagenomes were also compared using Mann–Whitney U tests. Fisher’s exact tests were used to compare the proportions of metagenomes containing genes conferring resistance to AMR classes commonly used in Salmonella treatment; this included genes conferring resistance to multiple antimicrobial classes where relevant.

Kendall’s correlation tests were performed to investigate whether or not the ability to detect S. enterica or the isolated serovar and its AMR determinants in the associated metagenome was associated with sequencing depth.

It was assumed that higher relative abundances of S. enterica reads in the metagenome would increase the likelihood of detection of the isolated S. enterica AMR determinants. The relative abundance of S. enterica in the metagenomes and the percentage of S. enterica AMR genes identified in the case metagenomes were compared using Kendall’s correlation to test this formally.

For beta diversity analysis, centroids were compared using PERMANOVA (adonis2), and heterogeneity of dispersions was assessed with betadisper and permutest from vegan.

A significance level threshold of 0.05 was used for all statistical tests.

Results

Salmonella enterica isolated from diarrhoeic stool samples

The 28 genomes representing isolates collected from salmonellosis patients were classified as S. enterica subsp. enterica. A total of 11 different serovars and 13 associated STs were identified (Table 1).

Table 1. The serovars and STs identified amongst the 28 cultured S. enterica isolate genomes.

Serogroup	Serovar	ST	No. of isolate genomes
B	Typhimurium	34	2
		19	2
		36	1
	Monophasic Typhimurium (I 4,[5],12:i:-)	34	11
	Stanley	29	1
		2615	1
	Paratyphi B var. Java monophasic	423	1
	Saintpaul	50	2
C2–C3	Newport	4157	1
	Albany	292	1
D1	Enteritidis	11	2
E1	Weltevreden	365	1
E4	Meleagridis	463	1
I	Hvittingfoss	446	1

Thirty-five unique AMR determinants, including genes and resistance-associated mutations, were identified amongst the 28 Salmonella isolate genomes with KMA and starAMR (Table S5). The AMR determinants identified conferred resistance to 12 different antimicrobial classes (Fig. 1). A high proportion of isolate genomes (82.1%) exhibited a MDR genotype (predicted resistance to three or more antimicrobial classes). The number of unique AMR determinants per genome ranged between 1-18 (median=9).

Fig. 1. Distribution of the percentage of 28 case-associated S. enterica isolate genomes containing genes conferring resistance to specific antimicrobial classes.

Microbial communities within case and control metagenomes

The total number of reads obtained for the metagenomes after trimming and host read depletion ranged between 30,576 and 94,890,832 for the case metagenomes and 10,821,636 and 15,177,561 for the control metagenomes.

The ten most abundant families in the faecal samples were compared and represented 64.6–99.6% (median 97.7%) and 57.9–92.5% (median 78.2%) of the overall case and control metagenomes, respectively. The most dominant microbial family amongst the case metagenomes was Enterobacteriaceae, representing the most abundant family in 14 samples (Table S6). The relative abundance of Enterobacteriaceae was significantly higher (P=7.66×10⁻⁵) in the case metagenomes (0.219–96.6%, median=40.2%) compared to controls (0.0246–19.4%, median=1.47%). Within this, Escherichia was the most prevalent genus in case metagenomes (Fig. S1), along with Klebsiella, and Escherichia coli was the most dominant species (Figs S2 and S3, Table S7). A smaller number of case metagenomes were predominated by Bacteroidaceae (n=5), Bifidobacteriaceae (n=4), and to a lesser extent other microbial families, whereas control metagenomes displayed more diverse and uniform microbial profiles, with Lachnospiraceae being the most abundant family in eight samples, followed by Bacteroidaceae, Bifidobacteriaceae and Enterobacteriaceae. At the genus level, the control metagenomes were predominated by Bifidobacterium and Bacteroides. Taken together, this suggests less diversity in the case metagenome group compared to controls, with the top ten most abundant families representing a larger proportion of the overall population (Fig. 2).

Fig. 2. The ten most abundant families in each case and control metagenome based on Bracken results, with remaining classifications grouped into the ‘other’ category in each sample.

To explore microbiome differences between sample groups further, alpha and beta diversities of case and control metagenomes were investigated. There was a significant difference in the observed richness (P=2.31×10⁸), Shannon (P=2.31×10¹⁰) and Simpson (P=4.09×10⁹) diversity indices between the sample types (Fig. 3a), with the control group displaying higher alpha diversity compared to case metagenomes. There was also a significant difference in the beta diversity between case and control metagenomes (adonis2 P=0.001). Analysis of the heterogeneity of dispersions revealed that the variances of the sample groups also differed significantly (P=0.001), which was reflected in the NMDS plot (Fig. 3b). The control samples clustered closely together, whereas the case metagenomes were more dispersed. Overall, the case metagenomes displayed less diversity and more heterogeneity with a predominance of select microbial families compared to controls.

Fig. 3. Alpha and beta diversity analyses of the case and control faecal metagenomes. Observed richness, Shannon and Simpson diversity indices calculated from Bracken results at the genus level for the case and control metagenomes, after rarefying to the smallest sample size (a); NMDS plot displaying dissimilarity between case and control metagenomes based on Bray–Curtis distance calculated from genus relative abundances from Bracken (b).

Detection of S. enterica and isolated serovars in the associated faecal metagenomes

To assess the utility of metagenomics as a tool to identify diarrhoea-causative Salmonella directly from culture-confirmed stool samples, the percentage of reads represented by S. enterica in the case metagenomes was determined and compared to the percentage of reads represented by S. enterica in the healthy control stool metagenomes.

S. enterica was identified in all 28 faecal case metagenomes with Centrifuge and Bracken. However, the percentage of the metagenome reads represented by S. enterica was variable (Fig. 4, Table S8). There was no significant association (P=0.150–0.298) between the percentage of S. enterica and the total number of metagenome reads. S. enterica was identified in all 21 of the control metagenomes with Centrifuge and 17 (81.0%) control metagenomes with Bracken (Table S9). The percentage of S. enterica reads in the case and control metagenomes varied significantly for both the Centrifuge (P=1.59×10⁻⁷) and Bracken (P=1.17×10⁻⁶) methods (Fig. 5). However, the distribution of S. enterica relative abundance in the case and control metagenomes was not entirely distinct, with some overlap at lower case percentages. Relative abundances of at least 0.0581% and 0.598% with Centrifuge and Bracken, respectively, were associated exclusively with case metagenomes, although there was a difference in the relative abundances for individual samples determined following analysis with the two tools (Tables S8 and S9). Sixteen case metagenomes were at or above the Bracken threshold, and 21 were at or above the Centrifuge threshold. All of the metagenomes above the Bracken threshold were also above the Centrifuge threshold. As an alternative approach, we also classified reads with MetaPhlAn4 to see whether or not this improved the discrimination between case and control metagenomes. With this approach, S. enterica was not detected in any of the control metagenomes (Table S9). However, the species was also not detected in 12 of 28 case metagenomes (Table S8). The lowest relative abundance of S. enterica detected with MetaPhlAn4 was 0.00577%, indicating that the potential diagnostic threshold varies by classification method.

Fig. 4. Proportion of case and control metagenome reads represented by S. enterica with Centrifuge (a), Bracken (b) and MetaPhlAn (c), with the Mann-Whitney U P-value displayed above the boxplots, and the lowest proportion associated with only case metagenomes highlighted by the red, dashed threshold line.

Fig. 5. The proportion of metagenome reads aligning to the custom Salmonella serovar database (a) and the highest proportion of reference chromosome covered after assembly of the aligned reads and realignment to the custom Salmonella serovar database (b), with the Mann-Whitney U P-value displayed above the boxplots, and the lowest proportion associated with only case metagenomes highlighted by the red, dashed threshold line.

The specific serovar of the isolated S. enterica was identifiable in 22 (78.6%) of the associated case metagenomes with Centrifuge (0.00000208–1.82 % of total reads, median=0.0000745%). There was no significant correlation between the percentage of the isolate genome serovar in the metagenome and the total number of metagenome reads (P=0.953). Despite being identifiable in a large proportion of metagenomes, the isolate genome serovar was the most abundant in only six (21.4%) case metagenomes based on Centrifuge results. S. enterica serovar sequences were also identified in all control metagenomes.

The metagenome reads were also mapped to a custom Salmonella database consisting of available reference genomes from NCBI, to facilitate serovar detection. The percentage of metagenome reads aligning to this database ranged between 0.200 and 49.1% for cases and 0.107 and 9.53% for controls (Fig. 5); these percentages were significantly different (P=8.51×10⁻⁵). Alignment percentages of at least 9.86% were associated exclusively with case metagenomes; this is considerably higher than the relative abundance thresholds for species relative abundance with taxonomic profiling. These reads were assembled and then aligned back to the genomes in the reference database. The serovar reference displaying the highest percentage of chromosome covered through alignment to a Salmonella reference database, assembly and realignment ranged between 0.243% and 92.8% (median=3.89%) in cases, though high initial read alignment percentage did not always result in a high percentage of reference chromosome covered (Fig. 5, Table S8). Nonetheless, the distribution of the highest percentage of chromosome covered was significantly different between cohorts (P=0.00969), with alignment proportions of at least 7.47% associated exclusively with case metagenomes. The isolated serovar was predicted as the most likely serovar present in 10 of the 28 case metagenomes with the Nucmer alignment method. SISTR analysis on the assemblies resulted in the correct prediction of the associated isolate genome serovar in eight case metagenomes (Table S8); however, it is important to note that the majority of assignments had low confidence as denoted by a QC warning or fail due to missing cgMLST loci, which was expected as this tool was not designed for use with fragmented assemblies from metagenomes. A SISTR cgMLST serovar prediction was also given for all controls.

Salmonella MAGs with >10% completeness were assembled in six case metagenomes only. Metabat2 assembled 1–3 S. enterica MAGs in five samples (completeness, 12.5–75.9 %; contamination, 0.00–1.75 %), whereas Maxbin2 identified one S. enterica MAG in six samples (completeness, 14.7–99.2%; contamination, 0.00–3.03%). Accepted Salmonella MAGs were identified in samples with high S. enterica relative abundance or high sequencing depth (Table S8). Kmerfinder classifications of the MAGs revealed the associated isolate genome serovar in four and three metagenomes for Metabat2 and Maxbin2, respectively. Overall, alignment and assembly methods identified the serovar of the associated isolate genome in 12 samples (Fig. 6).

Fig. 6. Summary of the detection (yes/no) of the isolated S. enterica serovar in the case metagenomes by each detection method used; for Centrifuge, the serovar was marked as detected if it was the most abundant serovar in the sample and for Nucmer if it displayed the highest proportion of the reference chromosome covered.

This indicates that the efficacy of detection of the isolated serovar also varies by method, though overall the identification of isolated serovars was less common than the pathogen species. S. enterica serovar Meleagridis was not identified in the associated case metagenome through any method. Identification was not possible with the serovar alignment method as a reference genome was not available in the database. Another metagenome (Metagenome_6916498) did not produce any MAGs.

The resistome of case and control metagenomes and detection of AMR genes identified in the S. enterica isolate genomes through metagenomics

The resistome of the case and control metagenomes was explored. There was a significant difference in the number of total AMR variants between cases and control metagenomes (P=0.0338), though no significant difference between the number of unique AMR genes (P=0.132) nor the number of antimicrobial classes to which these genes were predicted to confer resistance (P=0.697, Fig. 7). Most of the resistance classes were observed in both metagenome groups, except fusidane resistance genes were only observed in the case group, whereas aminocoumarin and nitroimidazole resistance genes were only observed in the control group. The percentage of metagenomes with genes conferring resistance to diaminopyrimidines, glycopeptides, lincosamides, macrolides and rifamycins varied considerably between groups. However, the level of resistance to classes including macrolides and lincosamides may be partly represented by genes conferring resistance to multiple antimicrobial classes, which were observed in 96.4% of case and 100% control metagenomes. When comparing the proportion of samples with resistance genotypes for antimicrobial classes used in salmonellosis treatment, there was no significant difference between the proportion of case and control metagenomes with genes conferring resistance to beta-lactams (Fisher’s exact test, P=1), macrolides (P=0.5) or fluoroquinolones (P=1). Overall, this indicates similar resistomes between cohorts.

Fig. 7. Comparison of the distribution of the number of individual AMR genes (a), unique genes (b) and antimicrobial classes represented by the AMR genes (c) identified in the case and control metagenomes with the median indicated in red, as well as the percentage of metagenomes containing genes conferring resistance to specific antimicrobial classes (d).

Characterisation of S. enterica AMR genotypes directly in stool metagenomes may be beneficial for guiding treatment in severe cases and invasive disease. Therefore, the metagenomes from the cases were screened for AMR genes identified in the associated S. enterica isolate genomes. The percentage of AMR genes in the isolate genomes identified in the associated metagenomes ranged between 0.00 and 100% (Fig. 8, Tables S5 and S10). All the unique AMR genes identified in the isolate genomes were also identified in seven (25.0%) of the associated metagenomes. There was a significant association between the percentage of unique isolate genome AMR genes identified in the associated metagenomes and the percentage of S. enterica reads in the metagenomes based on Centrifuge (P=0.00531) and Kraken2 and Bracken (P=0.000177) classification. However, the total number of metagenome reads did not significantly correlate with the percentage of AMR genes identified (P=0.113). Many genes commonly identified in the isolated S. enterica and the case metagenomes from this study were also identified in the control metagenomes (Figure S4, Table S11), except for aac(6′)-Iaa, aadA22, bla_CTX-M-55, lnu(F), mcr-3.1 and mcr-3.20. The aac(6′)-Iaa gene, a cryptic AMR gene found in all of the S. enterica isolate genomes investigated, was not detected in any of the control metagenomes. However, this gene was only identified in eight (28.6%) of the case metagenomes.

Fig. 8. Maximum likelihood tree based on core gene alignment of the 28 S. enterica isolate genomes, with the associated metagenomes labelled, as well as the serovars, STs and the AMR determinants of the S. enterica isolate genomes identified with KMA indicated; the AMR determinant matrix also indicates whether or not the gene was identified in the associated faecal metagenome. (A) The percentage of S. enterica determined based on Centrifuge results, (B) overall number of reads in the metagenome, (C) serovar of the S. enterica isolate genome, (D) ST of the S. enterica isolate genome.

The proportion of the sample resistome represented by S. enterica genes (AMR genes identified in the associated isolate genome) compared to other AMR genes present was investigated. The number of unique Salmonella-associated AMR genes identified in the metagenome ranged between 0 and 11 (Fig. 9a). The relative abundances of Salmonella AMR genes were calculated and compared to the overall AMR gene abundance in each sample (Fig. 9b). Salmonella AMR genes represented 0.00–67.1 % of the overall sample resistome, suggesting that although the number of genes commonly found in Salmonella is generally low compared to other AMR genes in the samples, the AMR genes associated with Salmonella can be highly abundant in the metagenome, which could potentially facilitate treatment guidance for salmonellosis cases. However, as most of the case metagenomes also contained high relative abundances of other Enterobacteriaceae, it is likely that the abundance of many of the Salmonella-associated genes was due to the genes also being present in related bacteria.

Fig. 9. Comparison of the number of unique S. enterica-associated AMR genes and other genes present in the associated faecal metagenomes according to KMA (a) and the abundance of unique S. enterica AMR genes compared to the overall AMR gene abundance in the metagenomes (b).

The MAGs obtained with Metabat2 and Maxbin2 were also screened for AMR genes, which resulted in the detection of 0–1 genes within the Metabat2 MAGs and 2–10 genes in the Maxbin2 MAGs (Table S12). The AMR gene identified in the Metabat2 MAGs (samples metagenome_6916509 and metagenome_6916514) was also identified in the Salmonella isolate genomes from those samples. However, the Maxbin2 MAGs from these samples contained erm genes that were not identified in the associated isolate genomes. This was also the case with other genes identified in the other Maxbin2 MAGs. The contigs of the Maxbin2 MAGs containing AMR genes were screened with blastn to investigate if sequences from other bacteria were incorporated into the MAGs. This revealed that the contigs containing mecA and erm genes were mostly associated with Gram-positive bacteria. Other genes (dfrA12, tet(M), qnrB4, ant(3″)-Ia, bla_DHA-1, bla_CTX-M-15, bla_CTX-M-27, aph(3′)-Ia, lnu(F), mef(B) and bla_CTX-M-55) were often associated with plasmids or chromosomes of bacteria in the Enterobacterales order, including Salmonella. The AMR genes identified in the isolated S. enterica were always associated with Enterobacterales plasmids or chromosomes or other Gram-negative bacteria. This highlights that the MAGs from this study may not accurately reflect isolated Salmonella AMR genes.

Identification of other diarrhoeagenic pathogens

In 12 case metagenomes, other potential diarrhoeagenic bacteria (Shigella flexneri, Shigella boydii, Shigella dysenteriae, Shigella sonnei, Staphylococcus aureus, Clostridium perfringens and Clostridioides difficile) were identified amongst the 20 most abundant species based on Bracken analysis. Shigella flexneri was identified in 10 case metagenomes with the relative abundance ranging between 0.0599 and 5.16 %, Shigella boydii in 8 (0.0198–0.694 %), Shigella dysenteriae in 5 (0.0251–0.485 %), Shigella sonnei in 4 (0.0295–0.132 %), Staphylococcus aureus in 4 (0.701–19.7 %), Clostridium perfringens in 1 (0.0720%) and Clostridioides difficile in 2 (0.193–0.265 %). Clostridioides difficile was also identified amongst the 20 most abundant species in 1 control (0.687%) metagenome. In 6 case metagenomes, at least one of the Shigella species identified amongst the 20 most abundant bacteria was more abundant than S. enterica, and in one case metagenome Staphylococcus aureus was more abundant than S. enterica. This highlights potential co-infection cases amongst the case metagenomes.

Discussion

In this study, we aimed to determine if metagenomics can be used as a potential diagnostic tool to identify S. enterica, associated AMR genes for treatment guidance and microbial community changes associated with infection, using faecal metagenomes from culture-confirmed paediatric salmonellosis cases.

We were able to specifically identify S. enterica using metagenomics in all 28 case metagenomes with both Bracken and Centrifuge. Although S. enterica reads were also detectable in 81% and 100% of 21 control metagenomes obtained from individuals without diarrhoeal symptoms for at least 6 months prior to sampling with Bracken and Centrifuge, respectively [31], the relative abundance was significantly lower compared to cases. An overlap between cases and controls could indicate asymptomatic carriage in some controls, which has been previously reported in Vietnam [51]. The control samples were screened with the Luminex xTag Gastrointestinal Pathogen Panel Kit to determine the absence of pathogens [31], which is a screening kit for nucleic acids of gastrointestinal pathogens or their toxins. However, this assay has previously shown 83.3–92.3 % sensitivity in the detection of Salmonella spp. from clinical samples [52,54], thus suggesting the possibility of false negative results.

Low S. enterica relative abundances were also observed in a number of case metagenomes, though such observations in symptomatic cases are not unusual and could be related to different stages of infection, the progression of which can be affected by the innate gut microbiome [12,55]. Coinfections could also explain lower NTS abundance and have been previously observed in paediatric gastroenteritis cases in Southeast Asia [53]. Indeed, a number of other bacterial pathogens were identified amongst the most abundant taxa in the case samples. However, definitive metagenomic identification of certain pathogens, including Shigella, is challenging as Shigella is difficult to distinguish from Escherichia [56]; therefore, these results should be interpreted with caution. Similarly, the presence of bacterial toxins associated with gastroenteritis was not evaluated; thus, identification of organisms like S. aureus and C. difficile in the case metagenomes provides insufficient evidence of co-infection with these organisms. Taken together, these findings highlight that verification of a diagnostic threshold for S. enterica abundance using metagenomics may require further investigation in a larger study that accounts for these potential confounders.

A combination of methods was used to identify the serovar of the S. enterica isolated from the case metagenomes and pathogen-associated AMR genes for treatment guidance, which may be particularly important for paediatric cases that are at increased risk of progressive disease [57]. Although reads representing the isolated serovar could be identified in up to 78.6% of case metagenomes, the isolated serovar was the most abundant in only six (21.4%) case metagenomes according to Centrifuge, and Salmonella serovars were also detectable in all control metagenomes. The prediction capability of the alignment and local assembly, as well as SISTR serovar prediction, may be related to the S. enterica relative abundance in the case samples, although higher S. enterica relative abundance did not always result in the prediction of the isolated serovar. We observed significantly higher percentages of reads aligned to the custom Salmonella database in cases compared to controls, though after assembly of the aligned reads and realignment back to the reference database, the highest percentage of chromosome covered was not necessarily as high; this could be due to initial alignment of non-specific regions, which can be an issue in samples containing high abundances of other Enterobacteriaceae. The reliability of serovar characterisation with these methods is, therefore, limited in the absence of culture and genomic data. Use of methods that take into account coverage depth and breadth of alignment could help to resolve this in future studies [58].

The mapping approach used for detecting AMR genes showed similar indications. The percentage of the Salmonella-associated AMR genes identified was correlated with the S. enterica relative abundance and the unique genes identified in the Salmonella isolate genomes represented up to 67% of the sample resistome. However, many of the genes identified were likely carried by multiple organisms present in cases and also controls, as the resistomes were overall not significantly different between the cohorts. The identification of S. enterica-specific genes, such as aac(6′)-Iaa, was possible in only 28.6% of case metagenomes, indicating that mapping approaches may be insufficient to accurately characterise the resistome of individual pathogens within samples [59]. MAGs could provide this capability, but they can be contaminated with sequences from other organisms [60], which could help to explain why many of the AMR genes found in the MAGs were not identified in the Salmonella isolate genomes associated with the samples. This was particularly apparent for the Maxbin2 MAGs, the contigs of which often mapped to distantly related bacteria or plasmids of Gram-negative bacteria, including Enterobacterales, but not exclusively Salmonella. However, it is important to note that the threshold for Salmonella MAG completeness in this study was relatively low (10%) compared to the recommended threshold for medium- and high-quality MAGs [61], and a contamination cutoff was not applied, which could partly explain the lack of consistency between MAGs and isolate genomes. Higher sequencing depth and the use of long-read sequencing could improve MAG assembly and thus detection of NTS serovars and associated AMR genes in the metagenomes with increased confidence.

In the current study, the detection of S. enterica and S. enterica AMR genes in the associated metagenomes was largely reliant on the genomes of S. enterica from isolates recovered from the faecal samples. It is possible that the faecal samples contained more than one serovar or strain of S. enterica that was not detected with culture and genomics. This could explain why the Centrifuge results indicated that a different serovar was more abundant than the recovered serovar in multiple metagenomes. Detection of more than one Salmonella serotype from faecal samples has been previously documented [62,63]; similarly, multiple strains displaying differing gene repertoires within the same serovar can be recovered from one faecal sample [64]. However, as the practice of selecting multiple isolates per sample is generally rare, it is not possible to accurately determine the frequency of such occurrences.

The salmonellosis cases were associated with a marked disturbance in the faecal microbiome compared to controls, as indicated by significant differences in alpha and beta diversity. The case metagenomes were marked with significantly fewer genera than controls and the controls clustered closer together than case metagenomes, suggesting reduced diversity in the gut microbiome and a shift away from a more stable microbial community in the NTS cases. Reduced microbial alpha diversity has been previously observed in diarrhoea cases [65], and the microbiomes of infants with diarrhoea have been shown to display wider variation than healthy controls [66].

We specifically observed a significant increase in the relative abundance of Enterobacteriaceae in case metagenomes, mostly associated with E. coli and Klebsiella species. These species can replicate in the conditions created through inflammation initiated by S. enterica [67], though such changes have been reported across a range of different bacterial aetiologies of diarrhoea [68,69]. Increases in the abundance of other facultative anaerobes, such as Enterococcus [70], can also occur under these conditions, particularly in dysenteric diarrhoea, as observed in some samples in the current study. We hypothesise that the differences in the microbiome profiles between salmonellosis cases could be used to characterise different stages of infection, with increased relative abundances of facultative anaerobes indicative of later infection stages in which the conditions allow proliferation of specific taxa. Conversely, metagenomes with less observed microbiome disturbance could represent earlier stages of infection, as more normal microbiome profiles were observed in a minority of cases here and reported elsewhere [65], or, alternatively, later stages in which the microbiome begins to return to a more stable state [65]. This was also reflected in the beta diversity NMDS plot, which suggested some overlap between cases and controls. However, such inferences are complicated by the observation of increased levels of Enterobacteriaceae in some control samples from healthy individuals, though this again could indicate asymptomatic S. enterica carriage. This highlights the limitations associated with microbiome characterisation in infectious disease, but also reinforces the need for further evaluation of potential diagnostic thresholds and additional studies to understand the microbiome dynamics of symptomatic and asymptomatic individuals. Alternative methods based on symptom history may be currently more feasible for infection staging.

This work supports conclusions from previous studies that the choice of tools and databases used can have a large effect on findings [59,71], as evidenced in the differing relative abundances of S. enterica in individual samples between methods and the serovar and AMR prediction results. Differences in S. enterica relative abundances inferred with Centrifuge compared to Kraken2 and Bracken can be a result of differences in tool algorithms, parameters and databases used; in particular, Bracken analysis involves abundance re-estimation at a given taxonomic level, which was not done in the Centrifuge approach. It is also important to reiterate that taxonomic classifiers used to infer the relative abundances of taxa in the current study have been associated with false positive results due to low specificity [71], and thus, low-level detection of S. enterica reads could indicate false positive results, again highlighting that further work is necessary to determine the diagnostic potential of metagenomics. Whilst not so much of an issue for the case samples, which were cultured for the pathogen, the identification of S. enterica reads in the controls does not necessarily mean all of the healthy individuals are carriers. Use of more conservative approaches, for example, those using marker genes to aid classification, could potentially help to distinguish cases and controls and thus facilitate the establishment of diagnostic thresholds. Additional classification with MetaPhlAn4 indeed indicated the absence of S. enterica in the healthy control cohort, though the pathogen was also not identified in 12 of the culture-confirmed case samples; therefore, the use of more conservative methods may control potential false positive detection, at the risk of instead increasing false negative results. Additionally, S. enterica serovar Meleagridis was not identified in the sample from which this serovar was isolated through alignment and local assembly, as a high-quality reference genome was not available in our database.

Another important limitation noted by other authors is microbial coverage reduction due to host DNA presence, which can be particularly high in gastroenteritis cases, especially if these feature bloody diarrhoea [69,72]. There was no correlation between the relative abundance of NTS or NTS serovars and total metagenome reads in the current study, indicating that the pathogen detection signal was not linked to the overall sequencing depth. However, there was a large range in the total size of case metagenomes after host-depletion compared to the controls, which highlights the need for effective host DNA depletion methods prior to sequencing. This would allow increased sequencing depth of bacterial reads and facilitate more robust detection, for example, through the assembly of high-quality MAGs.

In conclusion, this study demonstrates the potential of metagenomics for direct pathogen identification from stool metagenomes, although the stage of infection and possible coinfections can affect diagnosis in the absence of the isolated pathogens. Reliable characterisation of the causative pathogen with metagenomics at the serovar and AMR genotype level may not be possible with the current methods and tools, especially when the resolution is reduced by high proportions of host reads in disease cases. This limits the suitability for epidemiological studies and guiding antimicrobial treatment. Improvements to the method could allow more precise pathogen characterisation for clinical use. This could include implementing long-read and adaptive sequencing [73] or hybrid-capture target enrichment [74] to increase resolution of the pathogen of interest and improvements in host depletion prior to sequencing. Improvements in analysis pipelines, for example using comprehensive databases with pathogen-specific markers [69], or tools able to discriminate between closely related organisms to reduce false positive hits, could also be beneficial.

Abbreviations

AMR

antimicrobial resistance

CARD

Comprehensive Antibiotic Resistance Database

cgMLST

core genome multilocus sequence typing

ENA

European Nucleotide Archive

LMICs

low- and middle-income countries

MAG

metagenome-assembled genome

MDR

multidrug resistance

NCBI

National Center for Biotechnology Information

NMDS

non-metric multidimensional scaling

NTS

nontyphoidal Salmonella

OTU

operational taxonomic unit

SISTR

Salmonella in silico typing resource

STs

sequence types