We established a diagnostic platform for UTIs based on mNPS and subsequent data analysis. The workflow of our in-house mNPS platform includes DNA extraction, library preparation, sequencing, and pathogen identification (Figure 1).

In the first phase of our study, we assessed three methods of library preparation (rapid barcoding sequencing with RBK004, 16S sequencing with RAB204/16S024, and PCR barcoding sequencing with PBK004) using twenty culture-positive samples by comparing the consistency of mNPS results with culture-based results. Comparisons used the following criteria to evaluate library preparation methods: the sensitivity [referring to the positive criteria of sequencing established by Baldan et al. (2021) as follows: mono-species with largest number of reads, ≥ 10% difference from the next most abundant species] (Baldan et al., 2021), total reads number, the proportion of target reads and consistency in the detection of antibiotic resistance genes (ARGs) compared to clinical results. We then used the most comprehensive library preparation method to conduct mNPS testing on the remaining 56 samples.

We prospectively reviewed 76 cases of suspected acute or chronic infection between December 2020 and May 2021 at Dongfang Hospital in Beijing, China. Urine samples were collected via midstream clean catch for non-indwelled catheters patients and transurethral catheterization for indwelled patients, respectively. The urine remnants after standard-of-care clinical laboratory testing were then stored at −80°C before mNPS testing. The following criteria were used to determine inclusion in this study: no history of urological disorders and no prior clinical diagnosis of UTIs based on cell counts and cultures; with a few symptoms including urinary urgency, frequent urination, painful urination. A clinical diagnosis of UTI required to refer to the culture result and consider indicators including a white blood cell (WBC) count of > 10⁷/L, an Epithelial cell (EC) count of < 10⁷/L, fever, dysuria, frequency of urination, and urgency. WBC, EC, and red blood cell (RBC) counts were performed using phase-contrast microscopy. Urine cultures were performed on both MacConkey and blood agar plates, with incubation at 35°C for 16–18 h under aerobic conditions (Willner et al., 2014; Kumar et al., 2015). Demographic and baseline characteristics, clinical presentation, and laboratory findings of 76 patients were investigated for clinical diagnosis. Collection of surplus clinical samples was conducted under ethical approval by the Institutional Review Board (IRB) of the Beijing Dongfang Hospital (reference no. JDF-IRB-2020003101). All samples were obtained with the patient’s consent.

Clinical microbiological analysis of urine samples was performed according to standards formulated by the Clinical and Laboratory Standards Institute (CLSI). Briefly, 10 μl of urine were inoculated onto blood agar plates (bioMérieux) and incubated for 16–24 h at 35 ± 2°C in aerobic conditions. If cultures were negative for colony formation, plates were incubated continually until 48 h. Bacterial counts were determined from 10² to > 10⁵ CFU/ml. Bacterial identification and antibiotic susceptibility testing (AST) were performed separately with the VitekMS and VITECK2 Compact Systems (bioMérieux, SA), using the antimicrobial susceptibility test cards of the VITEK 2 AST-N334 and AST-GP67 Test Kits (bioMérieux, SA) for Gram-negative and Gram-positive organisms, respectively.

Urine samples were removed from storage at −80°C and incubated at room temperature until completely thawed. Samples were then gently shaken until well-mixed, and 1 ml of each sample was aliquoted into 1.5 ml Eppendorf tubes (Eppendorf) and centrifuged at 20,000 × g for 5 min to enrich for bacteria. The resulting bacterial pellet was resuspended in 180 μl phosphate-buffered saline (PBS) with brief, gentle vortexing. Then, 5 μl of lytic enzyme solution (Qiagen) and 10 μl of MetaPolyzyme (Sigma Aldrich; reconstituted in 750 μl PBS) were added to the samples and mixed by pipetting. Mixed samples were incubated for 1 h at 37°C to lyse bacterial cells. DNA was extracted from each post-lysed sample using an IndiSpin Pathogen Kit (Indical Bioscience). Sterile deionized water was extracted alongside the specimens as a negative control. DNA concentrations were assessed using a Qubit 4.0 fluorometer with the dsDNA HS Assay kit (Thermo Fisher Scientific).

All the samples included in this study were processed and sequenced regardless of DNA concentrations to provide an accurate representation of the data that would likely be obtained from metagenomic analysis of samples in clinical settings.

Library preparation for the MinION platform was performed according to the manufacturer’s protocols (see Supplementary Methods) for (1) 16S sequencing using a 16S Barcoding Kit (SQK-RAB204/SQK-16S024), (2) rapid sequencing with a Rapid Barcoding Kit (SQK-RBK004), and (3) PCR sequencing by PCR Barcoding Kit (SQK-PBK004), with the following minor alterations. The 16S PCR reaction was performed as per the manufacturer’s instructions. However, the number of PCR cycles was increased to 30. For PBK method, we used a 2 min extension time and 15 cycles. We selected 20 samples that were positive for clinical microbial culture and prepared libraries for each sample using all three of the above methods. The other samples were only sequenced using the PBK method.

MinION sequencing was performed using R9.4.1 flow cells (FLO-MIN106). A total of 75 μl of library DNA was loaded into the flow cell according to the manufacturer’s instructions. The MinION instrument was run for approximately 2 h. ONT MinKNOW GUI software (version 4.2.8) was used to collect raw sequencing data.

Figure 2 provides an overview of the bioinformatic workflow and the specific role of each tool in the process of MinION data analysis.

In cases of discrepancy between the results of mNPS and urine cultures, we sent 1 ml aliquots of the samples for Illumina 16S rDNA sequencing through an independent third-party company (LC Biotech, Hangzhou, China). Sequencing of the V3–V4 region of the 16S rDNA provided genus-based species distribution. The detailed methods are provided in Supplementary Methods.

To evaluate the limit of detection for our mNPS testing, we chose Escherichia coli and Enterococcus faecium as the representative species spiked into healthy donor negative urine samples with diluted concentrations of 10³, 10⁴, and 10⁵CFU/ml. Then we subjected these samples to our described mNPS testing by PBK method.

Statistical analysis to assess test accuracy used a two-tailed Mann-Whitney U-test at a significance threshold of P = 0.05 to evaluate if there were systematic differences in total reads, host reads, and bacterial reads between culture-positive and culture-negative samples. A P-value < 0.05 indicated significant differences. The sensitivity and specificity of tests were calculated based on the mNPS criteria for pathogen detection. All the above statistical analyses were performed with Graphpad 8.

We used urine culture as a clinical gold standard, and Illumina 16S rDNA sequencing as the confirmatory testing. The specific scoring algorithm is outlined as follows (see Supplementary Table 1): based on the clinical gold standard, for the culture-positive samples, true positives or false negatives were scored for each microorganism detected or not detected by mNPS positive criteria, respectively; for the culture-negative samples, true negatives were scored if no microorganism was detected by mNPS, otherwise, a false positive was scored. When multiple microorganisms were detected as positive by mNPS in one culture-negative sample, we counted all these false-positive microorganisms as one false positive result overall.

Demographic features of the 76 patients in this study are provided in Table 1. For culture-based identification, among the 76 samples, 45 (59.2%) were culture-positive, including 17 Escherichia coli (37.8%), 12 Enterococcus spp. (26.7%), 4 Klebsiella pneumoniae (8.9%), 3 Corynebacterium spp. (6.7%), 2 Proteus mirabilis (4.4%), 2 Staphylococcus epidermidis (4.4%), 2 Staphylococcus haemolyticus (4.4%), 1 Pseudomonas aeruginosa (2.2%), 1 Enterobacter hormaechei (2.2%), and 1 Morganella morganii (2.2%). The remaining 31 (40.8%) were culture-negative. We performed AST analysis of the 43 culture-positive samples (expect 2 Corynebacterium striatum samples have no AST result) and found that 31 (72.1%) were resistant to fluoroquinolone, 24 (55.8%) to penicillin, 16 (37.2%) to sulfonamide, 14 (32.6%) to cephalosporin, aminoglycoside and macrolide, 10 (23.3%) to nitrofurantoin, 9 (20.9%) to monobactam, 5 (11.6%) to tetracycline, 2 (4.7%) to carbapenem, cephamycin and nitroimidazole, and 1 (2.3%) to lincosamide, streptogramin, and glycopeptide. Detailed results of AST are shown in Supplementary Table 2.

The mean patient age was 70 years and 52.6% were female. Among the 76 patients, 71 (93.4%) were hospitalized and the median duration of hospitalization was 20 days. Information on the WBC counts, C-reactive protein, and procalcitonin inflammation indicators of all patients is also listed in Table 1.

In order to select the most suitable library preparation method for our mNPS platform, we compared the detection results of the three methods, using clinical urine culture as the gold standard. We found that 16S sequencing results were concordant with culture-based identification in 16 (80%) of the 20 samples. Notably, the rapid barcode sequencing results were concordant with only 11 (55%) of the 20 of the culture assays, while the remaining 9 samples showed disagreement in their identification or were negative for pathogenic taxa. Sequencing results from the PBK method were concordant with the culture results of 16 (80%) samples, although one could only identify bacteria to the genus level (Corynebacterium spp.). The detailed results are listed in Supplementary Table 3. These results indicate that library preparation by the PBK method could provide higher consistency with culture-based assays than that of the RBK method. In addition, the PBK method enabled deeper exploration of the genomic information relevant to drug resistance genes compared to 16S method. Taken together, these findings indicate that library preparation by PBK could provide the most reliable and informative results for mNPS-based pathogen detection in the full set of culture-positive samples.

Based on our above results, we then used the PBK method to generate libraries and performed sequencing for the remaining set of 56 clinical urine samples. We compared the total reads, percentage of host reads, and bacterial reads in the culture-positive and culture-negative samples, respectively (see Table 2 for medians, IQR, and range, see Supplementary Table 4 for raw data). We found that the total read counts showed no significant difference between the culture-positive and culture-negative samples (P = 0.7928), whereas the proportion of host reads was significantly lower (P < 0.0001) and the proportion of bacterial reads was significantly higher (P < 0.0001) in culture-positive samples compared with those of culture-negative samples. We examined the distribution of host reads and the top 3 most abundant bacterial reads (Figure 3) and found that 73.3% of samples had one obviously dominant bacterial taxon (e.g., P1 and P2, Figure 3A), while 11.1% of samples had two or more apparently predominant bacterial strains (e.g., P8 and P9, Figure 3B). There were also 8.9% of samples with no obvious dominant bacteria and relatively few bacterial reads (e.g., P33, Figure 3C).

Evaluation of test accuracy focused on the performance of mNPS testing in the identification of causal UTI pathogens relative to that of culture-based diagnostics. After evaluation of library preparation methods, we divided all the 76 samples into a training set (n = 35, including 20 culture-positive samples and 15 culture-negative samples) and a validation set (n = 41, including 25 culture-positive samples and 16 culture-negative samples) randomly. ROC curve was generated for the training set based on the results of clinical culture using the RPM metric (Figure 4).

Then, at the optimal Youden’s index derived from the training set ROC curve, the RPM threshold was set to 0.3065 with an AUC value of 0.8867. After meeting the precondition of ≥ 100 total pathogen reads, a minimum threshold of 0.3065 RPM was designated for reporting the positive or negative detection of a bacterial taxon (RPM ≥ 0.3065 or < 0.3065 for positive or negative, respectively). We then investigated the mNPS result of spiked-in samples based on positive criteria and found that all the species of ≥ 10⁴CFU/ml were successfully detected as positives. However, the remaining two samples with 10³CFU/ml spiked-in species were only detected as positive in one, with an RPM = 1.4 (see Supplementary Table 5). Next, we used the results of clinical culture to establish the sensitivity and specificity of bacterial detection by mNPS testing for the validation set and all the sample set, respectively. The results showed that mNPS-based diagnostic detection had 84.0% (95% CI, 64.7–94.2%) sensitivity and 100% (95% CI, 77.3–100%) specificity for the validation set and 86.7% (95% CI, 73.5–94.1%) sensitivity and 96.8% (95% CI, 82.4–99.9%) specificity for all the sample set compared to the standard clinical culture method.

Based on criteria for positive detection, sequencing results were in agreement with culture-based detection in 39 of 45 culture-positive samples (86.7%). The six remaining mNPS-negative samples were therefore considered false negatives (see Supplementary Table 1). In the case of P26, both clinical culture and mNPS found the presence of bacterial pathogen(s), but disagreed on the pathogen identification, with mNPS indicating the presence of dual-dominant pathogens K. pneumoniae and Lactobacillus delbrueckii and culture-based identification indicating the presence of P. mirabilis. To resolve this discrepancy, we conducted Illumina 16S rDNA sequencing and found that sample P26 was indeed positive for the two strains identified by mNPS, which supported the credibility of the mNPS identification method. In the case of sample P33, insufficient reads were obtained by mNPS to meet the criteria for a positive detection, whereas culture methods indicated the presence of E. faecium. Illumina 16S rDNA sequencing revealed the pathogenic agent as Enterobacter spp., which was also inconsistent with culture results (see Supplementary Table 1). In the case of sample P13 and P43, the culture results suggested that these two samples have bacterial counts of 10³CFU/ml and with our mNPS method it is still somewhat difficult to detect such low biomass samples, which is also in line with the result of spiked specimens (see Supplementary Table 5).

Among the 31 culture-negative samples, only one (N14) met the threshold for an mNPS-positive ID of E. coli, which was undetected by culture-based testing and therefore designated as a false positive. Among the other culture-negative samples, none met the mNPS threshold criteria for positive detection and were therefore considered true negatives (see Supplementary Table 1).

Following pathogen detection, we next explored the sequencing data to screen for antibiotic resistance genes (ARGs) in the 39 mNPS-positive samples. This analysis revealed a total of 73 ARGs across 31 samples, with 8 samples showing no mNPS reads that mapped to known ARGs. In these 31 samples, the largest number of ARGs per sample mapped to E. coli (Figure 5). Among the 73 identified ARGs, most were efflux components or β-lactamase genes, several of which conferred multidrug resistance. The abundance of ARGs in each sample and the corresponding types of antibacterial agents are shown in Figure 5. Multidrug resistance-related genes accounted for the largest proportion. Notably, the number of ARGs in output data was positively correlated with sequencing depth, since limited reads in some samples consequently also limited the number of detectable ARGs (e.g., P1 and P24). The AST results indicated that β-lactam, fluoroquinolone, and cotrimoxazole resistance appeared at the highest frequencies. We then compared the AGRs output with clinical AST results to investigate the concordance between the both (excluded 2 samples without AST result). The two methods produced fully concordant results in 6 (20.7%) samples and partially consistent results in the remaining 23 (79.3%) samples, and nitrofurantoin was the most frequently missed in ARG screening of the mNPS data. In addition, we found that all ESBL phenotypes conferred resistance to third generation cephalosporins due to the presence of blaCTX-M genes and the majority of fluoroquinolone resistance was due to CRP, emrH, and/or gadX genes. We also detected the vanA gene which is associated with vancomycin resistance in sample P18 and AST results indicated that it was indeed a vancomycin resistant strain. Other ARGs conferred resistance to untested and non-clinically relevant antibiotics (e.g., acridine_dye and rhodamine).