Precision metagenomics is an approach that customizes the analysis of a metagenomic sample for the detection and classification of a specific pathogen. The development of culture-independent methods for the detection of foodborne pathogens can expedite source tracking and reduce prospective corrective measures during outbreak scenarios (Loman et al., 2013; Huang et al., 2017; Brown et al., 2019). For U.S. Food and Drug Administration (FDA)-designated zero-tolerance pathogens, such as Listeria monocytogenes and Salmonella spp., qPCR or metagenomic detection is sufficient to initiate microbiological isolate confirmation followed by regulatory action (Archer, 2018; CFSAN¹). However, the pervasive nature of Escherichia coli necessitates further classification. There are more than 400 serotypes of Shiga toxin-producing E. coli (STECs) that can range in potential pathogenicity, as determined by the presence of a combination of stx (shiga toxin), eae (intimin), and other putative virulence genes (Kaper et al., 2004; Garmendia et al., 2005; Gonzalez-Escalona and Kase, 2019; Gonzalez-Escalona et al., 2019a,b; National Advisory Committee On Microbiological Criteria For Foods, 2019). STECs are responsible for ~2,400 hospitalizations per year causing severe diseases, ranging from diarrhea to hemolytic uremic syndrome (HUS) (Tarr et al., 2005; Mellmann et al., 2008; Scallan et al., 2011; Beutin and Martin, 2012).

Produce-related outbreaks have reportedly increased from 6% of all foodborne transmission in the 1990s to ~18% in a survey from 2003 to 2014 (Sivapalasingam et al., 2004; Fischer et al., 2015). Agricultural water has been considered a potential contamination source due to adjacent land use, incomplete water sanitization, or wild animal activity (FDA;² Steele and Odumeru, 2004; Monaghan and Hutchison, 2012; Oliveira et al., 2012; Allende and Monaghan, 2015; Uyttendaele et al., 2015). Current STEC detection protocols as outlined by the FDA Bacteriological Analytical Manual (BAM) Chapter 4A (Feng et al., 2020) entail 24-h enrichment, qPCR detection of the stx1, stx2, and wzy genes of the O157 antigen, followed by several rounds of microbiological assays (TSAYE, TC-SMAC, and chromogenic agar plates), and single colony isolation used for whole-genome sequencing (WGS), which confirms the identity and pathogenic potential of the isolated STEC. This entire process can take ~2 weeks of analysis time, which may exceed the shelf life of produce and, in particular, leafy greens.

To expedite the analysis time, we have investigated culture-independent sequencing methods for the detection and classification of STECs directly in agricultural water. Long-read nanopore sequencing classified the highly abundant microbial community (>1% read abundance) but failed to accurately detect reads belonging to O157 in unenriched agricultural water due to low concentrations (Maguire et al., 2022). We have, therefore, suggested precision metagenomic analysis of enriched agricultural water instead, as the STEC concentration is increased to detectable levels by qPCR, according to the BAM Chapter 4A protocol. To investigate the use of long-read nanopore sequencing for precision metagenomics, it was necessary to establish the limits of detection and assembly of the technique (Maguire et al., 2021a). However, detection is not sufficient for the classification of the STEC serotype or virulotype. The limit of nanopore complete, fragmented assembly of a metagenome-assembled genome (MAG) was equivalent to the limit of qPCR detection at ~10⁵ CFU/ml. At STEC concentrations above ~10⁷ CFU/ml, a complete, closed O157:H7 MAG was obtained with the chromosome in single contig and the plasmid in the second contig (Maguire et al., 2021a).

Despite the benefits of nanopore sequencing, including affordability, portability, long reads, and real-time basecalling, the inherent error rate of nanopore sequencing (when using the fast-calling model) precludes the closed genomes from being used for phylogenic analysis. Short-read Illumina MiSeq sequencing technology, however, is highly accurate but creates a challenge in assembling highly repetitive regions, which can span several hundred base pairs (Bertrand et al., 2019; Gonzalez-Escalona et al., 2019a; Moss et al., 2020). Hybrid assembly using nanopore long-read sequencing as scaffolds and MiSeq for accuracy improves the overall outcome and generates closed or fragmented MAGs with enough quality to be used in phylogeny analysis (Gonzalez-Escalona and Sharma, 2020; Maguire et al., 2022). However, as determined for long-read sequencing, the limits of detection and assembly for the Illumina MiSeq technology have yet to be determined as well. We expected the limits of detection and assembly of Illumina MiSeq to be different from nanopore for several reasons: (1) the starting material is small (1–100 ng, depending on the DNA library preparation kit used vs. 1 μg for nanopore sequencing); (2) only a fraction of the library is used for sequencing, resulting in sampling bias (the entire material is used for nanopore sequencing instead); and (3) most MiSeq sequencing runs multiplex several samples (the exact level of detection or sequencing depth necessary for each sample has not been established). Therefore, we aimed to determine the limits of short-read Illumina MiSeq detection and MAG assembly using agricultural water artificially contaminated with STEC strain EDL933_2 (O157:H7), as previously determined for nanopore sequencing (Maguire et al., 2021a). We further aimed to test the use of hybrid assemblies for the recovery of a complete or fragmented O157:H7 MAG. For this last part, we have evaluated the performance of three hybrid assembly software (Unicycler, OPERA-MS, and SPAdes).

Most food safety laboratories are relying heavily on different methods of recovery and identification of foodborne pathogens to facilitate the fast identification and source tracking of specific strains during foodborne outbreaks. Among those foodborne outbreaks of high importance are those caused by STECs in produce. Due to the importance of agricultural water to produce and food safety, accurate detection and classification of STECs potentially present is of major importance, especially during a foodborne outbreak. Current methods for strain identification (usually by SNP or MLST analysis) during a foodborne outbreak require sequencing of pure strains by either short (usually Illumina) or long reads (usually PacBio or Oxford Nanopore Technologies). However, to obtain a single pure colony strain, there are several steps, such as initial detection [usually by real-time PCR (qPCR)] and extensive selective plating, before a single colony can be obtained. This is a time-consuming process that only provides confirmation of an isolate after almost 2 weeks of labor, when the associated contaminated product might be out of the market chain. By combining qPCR and short-read and long-read metagenomic analysis of the enrichment, we can definitively detect an STEC isolate and characterize its virulence potential in 3–4 days. While this will not replace eventual confirmation by microbiological methods, this reduces the time for a prospective corrective measure by a complete week.

A previous study has determined the limits of detection and assembly for STECs in enriched water using long reads (nanopore) (Maguire et al., 2021a). However, the same limits for Illumina MiSeq have not been determined yet. Our first goal from this study was to determine those. The importance of these determinations is paramount to evaluating the feasibility of the use of any technique to support outbreak investigations. Short reads have been used for culture-independent surveillance and have shown promising results in retrieving STEC, Salmonella, and L. monocytogenes genomes in as short as 24 h (Leonard et al., 2015, 2016; Ottesen et al., 2020; Saltykova et al., 2020; Townsend et al., 2020; Buytaers et al., 2021; Commichaux et al., 2021; Zhang et al., 2021; Vorimore et al., 2023). However, those genomes were in high concentration in the sample (~80%−90%), indicating that, when levels are ~10⁸ CFU/ml or higher, short reads will perform correctly (the target strain will be easily identified if only one strain is present). The main pitfall of the use of those short reads will be that the genome assembly will still be composed of many non-contiguous contigs, and many important traits or markers of the genome might be lost.

In this study, we started by taxonomically classifying the microorganisms present in the enriched water sample used as a negative control for the presence of STEC O157:H7 using short reads (Illumina MiSeq). In general, the proportions were similar to what was obtained previously for the same samples using long reads (nanopore) (Maguire et al., 2021a), except for two microorganisms (K. pneumoniae and E. cloacae). This was unexpected, and the explanations could be as follows: (1) Illumina reads are shorter, which might lead to misclassification and overrepresentation of certain taxa; (2) while both analyses use Centrifuge, the RefSeq databases used might include some differences in classification and could contribute to the different classifications; (3) MiSeq loads only part of the library; (4) the efficiency in library preparations can vary with the GC content of the microorganisms; and (5) since most taxonomy classifiers rely on read numbers matching to the organism, in the case of short reads, some bigger bacterial genomes (>5 Mb) might appear to be more represented in the sample than smaller genomes. There is a need for further experiments to test the accuracy of strain-level identification by Centrifuge in metagenomics samples between nanopore and MiSeq sequencing outputs.

We continued by determining empirically the detection and assembly limits of the STEC strain EDL933_2 using Illumina MiSeq reads. The main difference with the experiment in our previous study, using long reads (nanopore), was that we multiplexed six samples per MiSeq run instead of a single sample per run. This allows us to perform sequencing of the same sample per duplicate per run. The fact that we used six samples per MiSeq run was because the limit of detection or assembly from metagenomic samples using that equipment was currently unknown. MiSeq detection limit using all reads was determined to be 10⁵ CFU/ml for STECs (when using six samples in a V3 cartridge), and the assembly limit was ~10⁷ CFU/ml to obtain a fragmented MAG. Even after binning the reads matching to E. coli, the assembly limit remained the same, although with a lower number of contigs (Tables 1, 3). When using a single sample per run, the estimated detection and assembly limits should therefore be six times higher than what we observed. The cost for a nanopore run is approximately USD$ 1,100, while an Illumina MiSeq run without multiplexing is approximately USD$ 1,500. All these variables (price, limits of detection, and assembly) should be taken into consideration when planning metagenomic studies using a MiSeq platform, specifically if multiplexing samples are used in a single run.

Assembling STEC genomes is a very complicated matter, and if the STECs are in mixed culture, it is even more difficult (Buytaers et al., 2021; Maguire et al., 2021a; Jaudou et al., 2022). We have previously shown that a completely closed STEC O157:H7 MAG can be recovered from enriched samples by sequencing using long reads (nanopore) at 10⁷ CFU/ml and higher levels (Maguire et al., 2021a). However, this completely closed MAG was not of high quality to be used for high-precision SNP analysis that could reveal the correct placement of the genome in a phylogenetic tree (results not shown). Therefore, we intended to test whether, by using a hybrid assembly approach, we could obtain similar results (completely closed MAG from a metagenomic sample) and with enough quality to be used for SNP analysis for source tracking as commonly used in outbreak investigations (Hoffmann et al., 2016; Crowe et al., 2017; Brown et al., 2019; Saltykova et al., 2020; Haendiges et al., 2021; Vorimore et al., 2023). For the hybrid assembly, we selected three known software (SPAdes) and pipelines (Unicycler and OPERA-MS). As mentioned in the Results section, neither of the assemblers performed adequately to meet our needs (recovery of the completely closed or fragmented O157:H7 STEC MAG) when using the complete datasets using default parameters. That is not to say that they performed inadequately for MAG hybrid assemblies overall for the entire metagenomic sample. That was not part of our tests. Considering that, the binned E. coli reads extracted from both long and short-read datasets were used instead for testing the performance of the hybrid assemblers. They performed with varying degrees of effectivity with SPAdes producing the better outcome (Table 5). They all showed the same limit of assembly (10⁷ CFU/ml) but failed to recover a completely closed MAG for any of the spiking levels. This was contrary to what was observed when using long reads alone (Maguire et al., 2021a), where a completely closed O157:H7 STEC MAG was recovered for spiking levels above 10⁷ CFU/ml. A similar observation when using hybrid assemblies for metagenomic sequences was observed for L. monocytogenes in enrichments from ice cream (Commichaux et al., 2021), where the hybrid assemblers using short reads as the first step of the assembly resulted in more accurate assemblies but were more fragmented and failed to recover a completely closed genome.

Another important consideration is that OPERA-MS requires 4 Gb of output data for samples with low complexity and 10 Gb for samples with high complexity. In our case, we had ~2 Gb per sample, which might explain why our OPERA-MS assemblies were the worst performers of the three tested ones, suggesting that a single sample should be run in a MiSeq system if a better hybrid assembly is necessary to obtain better or more complete MAG assemblies. However, a different system with higher output per sample could be used (i.e., NextSeq 2000 or another short-read sequencer) to fulfill the software pipeline requirements, albeit with higher expenses per run. If a simple estimation of taxonomic diversity is desired, the MiSeq approach might suffice if lower depth per sample is not a problem.

One of the reasons why we picked OPERA-MS for the hybrid assembly was that the authors claimed that their pipeline was able to detect strains of the same organism in the metagenomic sample (Bertrand et al., 2019). This is a very important and attractive feature that is desired when performing metagenomic analysis of enriched food or water samples. When food samples or water samples are analyzed, single colonies are picked and then tested for specific targets and sequenced (Walters et al., 2013). This process might miss some other co-occurring strains present in the sample, and those strains might be of lower prevalence (Commichaux et al., 2021). However, Unicycler was not designed for the assembly of metagenomic samples (Wick et al., 2017), but since we extracted the E. coli reads from both long- and short-read sequencing data for each individual sample, we expected it to perform successfully when used for a single organism.

Most of the current use of WGS in bacterial-associated diseases is focused on the sequencing of isolated microorganisms from selective culture plates (Hoffmann et al., 2016; Gobin et al., 2018). Most clinical and diagnostic laboratories are moving toward the use of culture-independent diagnostic tests (CIDTs) to quickly identify within minutes the etiological agent causing the illness. However, contrary to the culture method, which takes longer and requires more effort, a physical isolate is not produced at the end of the diagnosis. The lack of a physical isolate dramatically affects how public health agencies can properly identify potential outbreak clusters in outbreak investigations. This, in turn, has a negative impact on public health and increases the time to resolve ongoing or potential outbreaks (Carleton et al., 2019). A comparable situation occurs when trying to isolate foodborne bacteria from foods for surveillance or outbreak investigations, such as STECs, which take ~2 weeks to obtain a single isolate before doing WGS on that isolate and complete the outbreak investigation (Maguire et al., 2021a). Consequently, there is a renewed effort to move to culture-independent subtyping and outbreak investigations for pathogen infections in both clinical and food investigations (Carleton et al., 2019; Peña-Gonzalez et al., 2019; Maguire et al., 2021a). These methods or approaches need to be rapid, cheap (95–300 USD), and highly accurate (Carleton et al., 2019). In a previous publication, we were able to obtain completely closed STEC genomes from enriched samples (using a metagenomic approach); however, the technology employed (ONT) produced genomes that were not suitable for source tracking using an SNP approach. In this study, by adding Illumina reads and assembling the data using both Illumina and ONT data using a hybrid approach, we were able to obtain fragmented STEC O157:H7 MAGs with enough quality and accuracy to be able to be placed in the correct cluster when analyzed against 4,200 O157:H7 genomes available at NCBI (Figure 2). Both replicates for each spike level (1e⁸ and 7e⁷ CFU/ml) clustered together with the genome generated for the completely circular closed genome (EDL933_2 FDA Unicycler) spiked strain and differed among them by a maximum of three SNPs.

Current developments in ONT sequencing are showing promising results, with newer sequencing kits producing data of higher quality (Q20+) (preliminary unpublished data by authors). However, more studies and advances are still needed to make it suitable to be used as an alternative method to culture-based approaches, as well as the creation of improved assemblers for that kind of data.

Overall, we tested the limits of detection and assembly for EDL933 O157:H7 in enriched irrigation water using a shotgun short-read sequencing approach (MiSeq Illumina). In our previous study, the nanopore sequencing detection and assembly limits were determined to be 10³ and 10⁵ CFU/ml, respectively. In this study, we showed that, for MiSeq sequencing, the detection and assembly limits when using six samples per MiSeq V3 kit were two 10-fold lower at 10⁵ and 10⁷ CFU/ml, respectively, when compared to nanopore sequencing. Furthermore, short-read sequencing (MiSeq Illumina) of the highest level of EDL933_2 spiking of 7 × 10⁸ CFU/ml did not result in a completely closed genome as observed with the nanopore sequencing in our previous study. Nevertheless, the completely fragmented genome or fragmented MAG obtained with levels above 10⁷ CFU/ml with short reads was enough to make a complete characterization of the STEC strain, including serotype and virulotype. In this study, we ran six samples per single MiSeq reagent cartridge, and the analysis of the results provided vital information regarding how many samples could be run successfully in a single cartridge and their coverage depth, answering several unknowns about the limit of detection and assemblies for samples with varying degrees of abundance in the sample. For example, if your target of interest in your sample is low, then we suggest running a single sample per run or sequencing them in other instruments (i.e., NextSeq 2000) that can produce higher output data. Still, the cost of your sampling/analysis ratio must be taken into consideration as well. Contrary to what we were expecting, a hybrid assembly approach did not produce a completely closed genome even at the highest concentration (7 × 10⁸ CFU/ml), highlighting the complexity of STEC genome assemblies. Nonetheless, these fragmented MAG hybrid genomes clustered with the correct samples in an SNP tree of more than 4200 STEC O157:H7 genomes. Further studies using nanopore sequencing kits that show promising results in read quality and accuracy could help in solving these issues of closing high-quality STEC genomes from metagenomic samples to be used for outbreak investigations while the single purified strain is isolated for regulatory reasons. The impact of this single in silico purified strain would be remarkable, allowing to speed up the process of pulling potentially contaminated products from the shell earlier, reducing the number of illnesses associated with that product, and helping in root case analysis of the possible sources of contamination of such products. We must stress though that additional advances in the technology (more samples and output per run) and price reduction per run will make this approach more accessible for use in routine food testing at most laboratories.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://www.ncbi.nlm.nih.gov/genbank/, CP120944-CP120946 and PRJNA639799.

NG-E and MMag conceived and designed the experiments and wrote the manuscript. NG-E, MMag, and MMam analyzed the data. NG-E, MMag, MMam, MA, EB, and SM contributed to reagents, materials, and analysis tools. MMag, MA, ST, EB, and NG-E revised and drafted the study.