1 Introduction

Sepsis, a life-threatening dysregulated host response to infections caused by viable microbes in the body, represents almost 20% of all global deaths [1]. There are several factors contributing to the high mortality rate. The long turnaround time for blood culture (1–5 days), which is the gold standard diagnostic tool for bloodstream infection, could be a contributing factor in certain cases where proper antibiotic treatment is not introduced with empiric therapy [2]. Antimicrobial treatment prior to blood culture sampling [3], insufficient sampling and low inoculum can often result in delays or even failure (up to 50% negative) in detecting microbes involved in bloodstream infections [4]. The use of multiplexed PCR in this situation is restricted by the limited range of pathogens targeted by the panels [5]. Although targeted sequenced-based methods, such as 16S rRNA and 18S rRNA, widely used in clinical diagnostics, are not limited to pre-set panels, they allow the identification of pathogens predictably only at the species level, and often only at the genus level [6, 7]. Shotgun metagenomics has the potential to overcome some of the limitations of routinely used microbial diagnostic techniques. It can provide a different level of information, from the detection and characterisation of microorganisms and viruses to the detection of virulence factors and the identification of antimicrobial resistance markers, in a single assay without a priori knowledge [8]. As the abundance of the infection-causing microbes can be less than 0.25 CFU/mL [9], the recommended volume is 40–60 mL for blood culture diagnostics [10]. Molecular tests use only 1–10 mL of blood [11], which in this context might hamper detection, but in certain cases, for example, in paediatric patients, only one blood draw of 1–7.5 mL is feasible anyway [12]. To date, attempts to use shotgun metagenomics to diagnose bloodstream infections have been performed using cell-free DNA from plasma [13-17] and blood samples [18]. Gyarmati et al. used shotgun metagenomics to detect bacterial, fungal and viral pathogens along with antimicrobial resistance genes in blood samples of acute leukaemia patients with suspected bloodstream infections at different time points of their antimicrobial treatment [18]. Despite the promises of diagnostic value in the specific clinical situations mentioned above, the application of shotgun metagenomics in routine settings is still limited by its analytical sensitivity owing to the low abundance of bacteria during septicaemia and the large amount of host nucleic acid (NA) in blood samples [9]. High-quality extraction of DNA from the disease-causing bacteria in the clinical specimen is thus of utmost importance for bloodstream infection diagnostics, for which there is currently no consensus regarding the sample matrix to be used, that is whole blood (WB), plasma or both. In addition, short-read sequencing is often used as the standard diagnostic technology for sequencing, but it has a long turnaround time compared to sequencing with Oxford Nanopore Technologies (ONT) [19, 20]. The primary aim of this study was therefore to develop a shotgun metagenomics protocol using contrived blood specimens spiked with bacteria and to evaluate the bacterial DNA extraction with human DNA depletion. A secondary aim was to enable same-day diagnostics to have a turnaround time short enough to be meaningful in a clinical context.

2 Materials and Methods

2.3 Preparation of Plasma Samples

One millilitre of plasma was obtained from 5 mL of spiked WB by centrifugation at 180 g (P-180 g) or 100 g (P-100 g) (Figure 1) for 10 min at room temperature. The bacteria were only spiked directly into WB to mimic the actual conditions of a clinical sample.

3 Results

3.1 Samples Spiked With Single Bacteria

3.1.1 ddPCR Analysis

The DNA extraction efficiencies of the protocols for spiked WB and plasma samples obtained after centrifugation at 100 g (P-100 g) and 180 g (P-180 g) were evaluated by quantifying the specific genes of each spiked bacterial species and the human-specific gene using ddPCR. All ddPCR analyses were performed in duplicate, and differences in the number of gene copies were compared as means of duplicates. The bacterial gene copy appeared to be higher in spiked WB samples compared to plasma samples (centrifuged at P-100 g or P-180 g). This trend seemed more pronounced in samples spiked with S. aureus and S. pneumoniae (1.2–3.3 and 3.4–7.5-fold differences, respectively) than in those spiked with E. coli (1.1–1.2-fold difference; Figure 2).

The human gene RPP30 copy number was highest in P-100 g plasma samples and lowest in WB samples (the RPP30 gene was undetectable in S. aureus-spiked WB [Figure 2a]). In the WB sample spiked with 10⁵ CFU/mL S. pneumoniae (Figure 2b), there was an unexpectedly high RPP30 gene copy number compared to all the other samples, indicating a technical issue in the human depletion step in this sample. Overall, the P-100 g protocol yielded a higher RPP30 gene copy number and a lower copy number of bacterial genes than the WB samples. The negative controls showed very low levels of 16S DNA in the ddPCR and was thus not sequenced.

3.1.2 ONT Sequencing and Data Analysis

Next, pooled DNA of duplicates for each spiked sample was sequenced. The proportion of reads classified as bacteria and host (human) was compared between WB and the different plasma preparations (Figure 3). The proportion of reads classified as bacteria was highest in WB (Figure 3b,c), except for S. aureus, for which bacterial-classified reads were highest in the P-180 g plasma samples (Figure 3a). The lowest proportion of bacterial reads was detected in P-100 g samples. Moreover, more than 95% of the reads (min 95.2%–max 99.49%) were classified as human in all samples, with the highest in P-100 g samples (Table S2). The depth of genome sequencing for bacteria was low overall, with a maximum of 4.8x coverage in S. aureus P-180 g samples, which also had the highest proportion of bacterial reads (2.3%; Table S2).

In general, the P-100 g protocol yielded the highest proportion and gene copy number of human DNA and the lowest proportion and gene copy number of bacterial DNA. Therefore, the P-100 g protocol was excluded from the ATCC-MSA-2002 mock community spiking experiments.

3.2 Polymicrobial (Mock Microbial Community) Spiked Samples

3.2.1 ddPCR Analysis

In the ATCC-MSA-2002 mock community spiked samples, gene copies/μL for nuc, uidA and 16S rRNA genes as determined by ddPCR were at least two times higher in WB samples than in plasma (P-180 g) at three different concentrations: 4 × 10², 4 × 10³ and 4 × 10⁴ cells/mL (Table S1 and Figure S1). The RPP30 gene copy number was lowest in 4 × 10⁴ cells/mL spiked WB but unexpectedly highest in 4 × 10³ and 4 × 10² cells/mL spiked WB. Although the RPP30 gene copy numbers among all samples were more consistent for P-180 g samples (n = 11; median 2.0 copies/μL; IQR 1.45–2.1), they varied substantially for WB samples (n = 11; median 3.0 copies/μL; IQR 0.31–10.45). The variation in gene copies among the replicates was also observed for nuc, uidA and 16S rRNA genes in WB samples, most likely due to extraction variability (Table S1 and Figure S1). The negative controls showed very low levels of 16S DNA in the ddPCR and were thus not sequenced.

3.2.2 ONT Sequencing and Data Analysis

The replicates of WB and P-180 g samples spiked with the ATCC-MSA-2002 mock community were sequenced. In the first dilution (4 × 10⁴ cells/mL), 48% and 10% of the reads were classified as bacteria in WB and P-180 g samples, respectively (Figure 4). The proportion of reads for total bacteria was also higher in WB samples than plasma samples at dilution 2 (4 × 10³ cells/mL). However, 99% of the reads were classified as human in both WB and plasma samples at dilution 3 (4 × 10² cells/mL). The mean read length obtained from plasma samples was shorter than from WB samples (Table S3).

The DNA extraction efficiency for each bacterium in the mock community was investigated by looking at the respective proportion of reads (Figure 5) and the number of reads (Figure S2) classified in WB and plasma. Using Kraken2, the highest relative abundance was seen for Enterococcus faecalis and S. aureus, followed by Acinetobacter baumannii in both WB and plasma samples (Figures 5 and S2). Most of the other species were barely detected. Pseudomonas aeruginosa and Cutibacterium acnes were not detected in either WB or plasma samples. To further investigate this result, the mock community bacterial suspensions in 1 x PBS (not spiked into WB) at three different concentrations were sequenced. P. aeruginosa, which could not be detected either in WB or plasma (Figure 5a,b), was recovered from the bacterial suspension in PBS (Figure S3). Furthermore, including a pre-enzyme treatment step into the protocol increased the DNA extraction yield for Gram (+) bacteria but decreased it for Gram (−) bacteria (Figure S3).

When studying the abundances of S. aureus and E. coli in the mock community spiked samples, the detection limit was estimated to be 20 genomic copies/mL WB +/−1 log₁₀ for ONT sequencing (Figure 5a).

4 Discussion

In this study, a shotgun metagenomic protocol using ONT sequencing, with the potential of same-day diagnostics of bloodstream infections, was investigated. WB and plasma were evaluated as clinical specimens to reach an optimum shotgun metagenomics protocol with a similar NA extraction efficiency for different bacterial species despite their varying cellular structures. Considering the challenges of the conventional diagnostic approaches mentioned previously, shotgun metagenomics has the advantage of potentially providing information on all microbial pathogens and associated antimicrobial resistance in a single assay directly from the sample. The fast sequencing technology of Oxford Nanopore was used to get test results within 1–2 days by processing 10 mL of WB and 1 mL of plasma (obtained from 5 mL of WB). In general, more bacteria, with the highest bacterial gene copy number detected by ddPCR and the highest proportion of reads classified as bacteria through sequencing, were recovered using 10 mL WB. One exception was S. aureus, for which the number of bacterial reads was highest in the P-180 g plasma samples. The WB and P-180 g samples spiked with the ATCC-MSA-2002 mock community (4 × 10⁴ cells/mL) showed proportions of reads classified as bacteria at, on average, 48% and 10%, respectively. In turn, the bacterial DNA extraction was least efficient in plasma P-180 g samples, giving the lowest proportion of reads classified as bacteria and with the highest level of human DNA. The mean read length obtained from plasma samples was also shorter than from WB samples (Table S3). However, reproducibility was higher in plasma compared to WB, shown by IQRs of 1.34–2.1 and 0.31–10.45 copies/μL, respectively. Together with the fact that the Molzym kit resulted in a consistently higher DNA extraction efficiency for Gram (+) bacteria in all preparations, these results make it hard to definitely conclude on the impact of the choice of sample matrix. Although the spiked mock community theoretically contains an even whole cell mixture of 20 different bacterial species, Gram (+) bacteria, particularly E. faecalis, were detected in high proportions in both WB and plasma samples. To study the efficiency of the DNA extraction without the effects of a sample matrix but also without the effects of the human depletion step, possibly affecting the extraction of DNA from Gram (−) bacteria, DNA was also extracted directly from the mock community suspension in PBS in three different concentrations. However, another DNA extraction kit had to be used, magDEA Dx SV, since the Molzym kit is only designed for DNA extraction from clinical samples that include human material. We found the P. aeruginosa, which could not be detected either in WB or plasma (Figure 5a,b), was recovered from the bacterial suspension in PBS (Figure S3). Other Gram (−) bacteria, such as E. coli, Neisseria meningitidis and A. baumannii, were also detected in higher proportions in the protocol without the human depletion step. We additionally evaluated how enzymatic pre-treatment affected bacterial DNA isolation and found it to improve extraction efficiency for Gram (+) bacteria but decrease it for Gram (−) (Figure S3). This finding should be considered when trying to determine an optimal extraction procedure for clinical samples prior to shotgun metagenomics.

We observed that there was no strict linear relationship in the dilution series in the detection of bacteria. This result might be explained by the low concentrations of bacterial DNA in the higher dilutions, as human depletion is likely more important in samples containing less bacterial DNA to reach effective sequencing. A disagreement was also observed in the amount of host DNA between ddPCR and sequencing. Although the ddPCR copy number of the RPP30 gene was lower than the bacterial gene copy number in WB and P-180 g samples, most of the reads were classified as human with ONT sequencing in samples spiked with single bacteria (> 95% human reads) and the ATCC-MSA-2002 mock community (> 50% human reads). This difference is most likely explained by the modified library preparation protocol, which includes an increased amount of PCR cycles, which may not be optimal. Since the library preparation includes this PCR step, in which all material amplifies exponentially, considering the size of the human genome vs. the bacterial genome, the included PCR step increases the human DNA copy number to a much greater extent. Theoretically, in 10 mL WB with a concentration of E. coli of 10⁴ CFU/mL, the ratio of leucocyte (10 × 10⁶ cells/mL) human DNA (~6.4 billion bp) to E. coli DNA (~5 Mb) should be 320,000,000: 500 [10 × (5 × 10⁶) × 6.4: 10 × 10⁴ × 0.005]. The difference in amplification also explains why the detected amount of human DNA is more significant with sequencing than ddPCR, since the ddPCR only targets one single gene from the whole human genome (RPP30) and sequencing detects the whole human genome within the sample. Yet, the increased amplification enables the detection of low-abundant pathogens present in a sample, which can be very important in a clinical context.

Unfortunately, we could not conclude how the different matrices affected the bacterial DNA isolation efficiency using the same Molzym kit. However, extraction of bacterial suspension using the magDEA Dx SV kit also showed a difference in the proportion of reads classified for each bacterium despite theoretically even cell mixtures in suspension. A substantial variation in Molzym kit extractions among the replicates was also observed, especially in WB samples. For example, one of the replicates spiked with 4 × 10⁴ cells/mL did not yield any quantifiable DNA. The failure during automatic extraction could be due to a high amount of total NA released from the bacteria and the human cells that made the magnetic beads malfunction. Nucleic acid-based detection methods might also be hampered by inhibitory components such as haemoglobin, immunoglobulin and heparin in WB [25]. These factors affect reproducibility when using WB as a sample matrix and should be considered when introducing this into a clinical context. The presence of inhibitors could also explain the source of variation observed in the replicates within the same dilution, shown in general within this study. A previous study comparing different DNA extraction kits from spiked WB also showed similar inconsistencies in DNA concentrations among different concentrations of bacteria using the MolYsis Complete5 kit of Molzym [26].

Overall, the detection limit was estimated to be 20 genomic copies/mL WB +/−1 log for ONT sequencing, and DNA isolation was generally more efficient for Gram (+) than Gram (−) bacteria. To assess this more accurately, it would have been ideal to check the gene copy number of each of the 20 bacteria included in the mock community by ddPCR. Nevertheless, one should consider the differences in genome size, which might favour the number of reads from larger bacterial genomes. However, we did not see any association between the proportion of reads and the genome size of the bacteria in the mock community.

One limitation of this study is the low coverage since a higher coverage is required for reliable data. At the time of these experiments, the newer Q20 technology and analytic tools from Nanopore were not available, and therefore all experiments were run using the old technology, known to have a higher error rate and therefore the species-level call had lower reliability [27, 28]. Hence, the pathogen content of the samples was here known in advance, and the species-level call could be trusted. This finding highlights the importance of further improving the method using the newest Nanopore technology with higher accuracy to make it more sensitive for low levels of bacteria. Improvements would facilitate reaching down to species level to a greater extent than seen in this study, in which several reads were classified at the genus level. In addition, the study was limited by the volume of samples available and the cost to process all biological and technical replicates of samples in the ddPCR and Nanopore sequencing experiments. Another limitation is also the fact that the blood used for spiking was contrived from healthy donors. WB is not a passive medium, and in patients presenting with bloodstream infections, with a storm of events occurring, blood cells are exceedingly active, which may influence the result. Further studies are needed with specimens from patients with bloodstream infections. Careful consideration should also be given to the choice of human depletion method; saponin and DNAse can also be used to reduce human DNA in similar samples and should be compared with the choice of method in this study. In addition, sequencing technologies are constantly evolving and selective sequencing using Nanopore could be a useful approach to try out in the future [29].

5 Conclusion

In this study, we used nanopore-based shotgun metagenomics for spiked WB and plasma samples. The study revealed several challenges and critical steps, the first of which is achieving optimal extraction, including human DNA depletion. The proportion of bacteria detected from 10 mL of WB appeared to be higher, but it also seemed to have a lower reproducibility compared to plasma. Second, the DNA extraction efficiency was higher for Gram (+) bacteria, regardless of the sample matrix. These two factors prevent us from drawing definite conclusions as to whether WB or plasma is the best sample material. Third, the low genome coverage also indicates that further optimisation is necessary to obtain more detailed information beyond pathogen detection. Several challenges remain with shotgun metagenomics-based diagnostics of bloodstream infections. This study contributes by highlighting some of the critical steps for future development of shotgun metagenomics-based diagnostics to bring it closer to a method with general clinical applicability.

Conflicts of Interest

At the time of the study John Rossen was consulting for IDbyDNA (now Illumina). This did not have any influence on the interpretation of reviewed data and conclusions drawn nor on the drafting of the manuscript, and no support was obtained from them. Remaining authors declared no conflict of interest.