A total of 213 leftover body fluids, along with their corresponding culture results, were collected from the clinical microbiology laboratories of four public hospitals in Hong Kong: Pamela Youde Nethersole Eastern Hospital, Prince of Wales Hospital, Princess Margaret Hospital, and Tuen Mun Hospital. Detailed information regarding the specimen types and culture results are shown in Supplementary Table 1. Upon receival, DNA extraction was performed using QIAamp BiOstic Bacteremia DNA Kit.

Library preparation was performed using 16S Barcoding Kit 1 – 24 (SQK-16S024) from Oxford Nanopore Technologies (ONT) according to the manufacturer’s protocol with some modifications. The suggested input is 10 ng of genomic DNA in the protocol. Since it is impossible to quantify only bacterial DNA in clinical samples with high human DNA background, the maximum input volume of 15 μl was added instead of 10 ng. The PCR cycle number was also increased from 25 to 35 cycles in order to increase the sensitivity of the assay. A total of 24 barcoded libraries were pooled in equal concentration and sequenced for up to 24 h using the flow cell FLO-MIN106 R9.4.1 with the sequencer GridION on the MinKNOW platform, with super-accuracy basecalling model. To reduce index misassignment, “mid-read barcode filtering” and “barcode both ends” were adopted, and the minimum barcoding score was set to be 85.

The sequencing reads were uploaded to Epi2me, the cloud-based analysis platform developed by ONT, for real-time analysis. The FASTQ 16S workflow (v2021.09.09) in Epi2me was adopted with a minimum QSCORE of 10, which is the default minimum QSCORE of super-accuracy basecalling model. The reported average read accuracy of R9.4.1 flow cell ranges from 92.3 to 96.52% in recent studies (Ni et al., 2023; Zhao et al., 2023), which is lower than the traditional 97% similarity cutoff to differentiate different species. Considering the relatively lower read accuracy of Nanopore sequencing, a minimum coverage and a minimum identity of 90% were adopted and only reads between 1,000 to 2,000 bps were included for the analysis. The sequencing reads were further analyzed with two additional pipelines: Emu (Curry et al., 2022) and NanoCLUST (Rodriguez-Perez et al., 2021), with default parameters.

The concordance between culture and the Nanopore 16S, coupled with the three analysis pipelines in culture-positive samples were calculated by (number of cultured species detected by Nanopore 16S)/(total number of cultured species). For the culture-negative samples, if clinically important species were obtained in Nanopore 16S, the results would be correlated with the clinical manifestations and medical history of the patients. All the statistical analysis was performed using GraphPad Prism (v9.5.0), a p-value less than 0.05 was regarded as statistically significant. The statistical difference between two populations of sequencing reads and classified species were calculated using Mann–Whitney U test.

Relative abundance of a species was calculated by (number of reads of a species in a sample)/(total number of classified reads of a sample). To determine the threshold of relative abundance (T_RA) for detecting potential pathogens in body fluids using Nanopore 16S, a receiver operating characteristic (ROC) curve was illustrated for each analysis pipeline based on the relative abundance of true positives (detected by both culture and Nanopore 16S) and false positives (detected only by Nanopore 16S) using GraphPad Prism (v9.5.0). ROC analyses of monomicrobial and polymicrobial samples were performed separately. The optimal point with the maximum Youden’s index in the ROC curve was considered the T_RA for detecting pathogens in each analysis pipeline, and the average value of the three pipelines was considered the T_RA in Nanopore 16S.

To determine the LOD of Nanopore 16S, simulated bacteremic blood samples were prepared by spiking two reference strains (Staphylococcus aureus BAA-3114 and Klebsiella pneumoniae BAA-3079) in EDTA-blood provided by a healthy individual with a final concentration of 150, 100, 50, and 10 CFU/ml. The simulated samples were extracted and sequenced as authentic samples and analyzed using Epi2me. The experiment was repeated twice to eliminate random errors. The average relative abundances of each species in Nanopore 16S were plotted against the bacterial concentrations. A trend line was used to determine the minimal bacterial concentration that meets the T_RA.

Among the 128 monomicrobial body fluids, the most frequently identified pathogen in culture was coagulase-negative Staphylococci (21/128), followed by Staphylococcus aureus (20/128), Escherichia coli (12/128), and Enterococcus faecalis (12/128). Among the 65 polymicrobial body fluids, the most common isolated pathogen in culture was Escherichia coli (42/65), followed by Klebsiella pneumoniae (16/65) and Bacteroides fragilis (14/65). For Nanopore 16S, the twelve most abundant species and their respective relative abundances of each sample classified by the three analysis pipelines were summarized in Supplementary Table 2.

After 24-h sequencing, the monomicrobial samples yielded an average of 76,551 reads per sample (s.d. ± 91,867) with a median of 32,699 reads. The polymicrobial samples had comparatively more reads per sample than the monomicrobial samples (p < 0.0001), with an average of 160,172 reads (s.d. ± 168,470) and a median of 127,662 reads per sample. The culture-negative samples had significantly lower number of reads compared to those culture-positive samples (P < 0.0001), with an average of 23,442 reads (s.d. ± 54,352) and a median of 1,373 reads per sample.

An overview of the concordance between culture and Nanopore 16S results was illustrated in Figure 1, while Supplementary Tables 3, 4 record the concordance of each monomicrobial and polymicrobial sample, respectively. In both monomicrobial and polymicrobial samples, Nanopore 16S coupled with Emu demonstrated the highest concordance. Among the 128 monomicrobial samples, Emu correctly identified the taxon of 125 (97.7%) samples, compared to 109 (85.2%) samples and 102 (79.7%) samples identified by Epi2me and NanoCLUST, respectively. Among the 65 polymicrobial samples, Emu correctly identified all the cultured species in 35 (53.8%) samples, compared to 19 (29.2%) samples and 9 (13.8%) samples in Epi2me and NanoCLUST, respectively.

Historically, a threshold of 97% sequence similarity in the 16S rRNA gene was used to differentiate two species (Janda and Abbott, 2007; Beye et al., 2018; Johnson et al., 2019). However, in some closely related species, the sequence similarity can be 98% or higher (Kawamura et al., 1995; Beye et al., 2018; Devanga Ragupathi et al., 2018). Supplementary Table 5 provides a list of closely related species (sequence similarity ≥ 98%). If those closely related species were considered concordant classifications, the number of correctly detected samples increased for Epi2me. In this case, the number of correctly detected samples for Epi2me, Emu, and NanoCLUST in monomicrobial samples increased to 127 (99.2%), 126 (98.4%), and 119 (93.0%), respectively. For the 65 polymicrobial samples, Epi2me successfully detected all the cultured species in 50 samples (76.9%), compared to 37 samples (56.9%) and 30 samples (46.2%) in Emu and NanoCLUST, respectively. Partially concordant samples were defined as those that failed to detect all but at least one of the cultured species by Nanopore 16S. In Epi2me, Emu, and NanoCLUST, a total of 15 samples (23.1%), 28 samples (43.1%), and 33 samples (50.8%) of polymicrobial samples were partially concordant with the culture results, respectively.

A total of 230 species were cultured from 65 polymicrobial body fluids. When considering all the cultured species in polymicrobial body fluids together, Nanopore 16S coupled with Emu showed the highest concordance (81.7%) if closely related species were not considered, compared to 75.7% for Epi2me and 54.3% for NanoCLUST (Figure 2). If closely related species were considered concordantly classified, Epi2me showed the highest concordance (91.7%), followed by Emu (83.0%), and NanoCLUST (72.6%).

The overall concordance of Nanopore 16S with culture in the 193 culture-positive body fluids, totaling 358 cultured species, was summarized in Figure 3. The overall concordance of Epi2me, Emu, and NanoCLUST was 79.1, 87.4, and 63.4%, respectively. If closely related species were considered concordantly classified, the overall concordance of Epi2me, Emu, and NanoCLUST would be 94.4, 88.6, and 79.9%, respectively.

In both monomicrobial and polymicrobial samples, there were cases where Nanopore 16S failed to detect the expected species based on the culture reference or their closely related species. In monomicrobial samples, the number of discordant samples was 1 (0.8%) for Epi2me, 2 (1.6%) for Emu, and 6 (4.7%) for NanoCLUST. Notably, three monomicrobial samples could not be analyzed by NanoCLUST, possibly due to insufficient reads (<120 reads) for cluster generation. For polymicrobial samples, two completely discordant samples were found solely in NanoCLUST, as it failed to detect any of the cultured species in these samples. In contrast, Epi2me and Emu were able to detect at least one of the cultured species in a sample.

The number of classified species per sample for each pipeline was summarized in Figure 4. In monomicrobial samples, despite single species being cultured, Nanopore 16S often classified more than one species (Figure 4A), regardless of the analysis pipelines. Additional organisms, apart from the cultured species, were detected in 128 (100%) samples, 122 (95.3%) samples, and 106 (82.8%) samples by Epi2me, Emu, and NanoCLUST, respectively. The average number of classified species per sample in monomicrobial samples was 83.9, 15.1, and 7.7 for Epi2me, Emu, and NanoCLUST, respectively. Interestingly, the cultured species were not necessarily the most abundant species in the correctly detected samples. In Epi2me, out of the 109 correctly detected samples, the cultured species in 27 samples were not the classified species with the highest relative abundance, including 3 samples with extremely low relative abundances (<0.1%). Similarly, in Emu, 37 out of 125 correctly detected samples were not the most abundant species among the classified species, and one sample had an extremely low relative abundance (<0.1%). For NanoCLUST, 25 out of 102 correctly detected samples were not the most abundantly classified species.

In the case of polymicrobial samples, the number of classified species per sample was significantly higher (p < 0.0001) compared to monomicrobial samples, regardless of the taxonomic classifiers used (Figure 4B). Similarly, Epi2me exhibited the highest number of classified species per sample, with an average of 153 species classified per sample, compared to 62 and 30 classified species in Emu and NanoCLUST, respectively. Additional organisms, beyond the expected ones, were detected in 65 (100%), 64 (98.5%), and 62 (95.4%) samples in Epi2me, Emu, and NanoCLUST, respectively. Notably, in Epi2me, 32 out of 50 (64.0%) concordant samples included species with extremely low relative abundances (<0.1%). In contrast, in Emu and NanoCLUST, 12 out of 37 (32.4%) and 1 out of 30 (3.33%) concordant samples included species with extremely low relative abundances, respectively. Therefore, Epi2me demonstrated a higher sensitivity for detecting extremely low abundance species.

Among the three taxonomical classifiers, Emu showed the highest concordance with the reference culture method as it correctly differentiated E. coli from Shigella and other Escherichia species. On the other hand, both Epi2me and NanoCLUST misidentified E. coli as other closely related species in 12 monomicrobial samples. In polymicrobial samples, Epi2me misidentified E. coli in 37 samples, while NanoCLUST misidentified E. coli in 35 samples. It should be noted that in Epi2me, most of the reads of E. coli were classified at the family level (Enterobacteriaceae), resulting in a lower relative abundance of E. coli or related species at the species level compared to the results of Emu and NanoCLUST.

E. coli and Shigella are known to be closely related species with highly similar 16S rRNA gene sequences (Devanga Ragupathi et al., 2018). Although Emu outperformed the other two pipelines in identifying E. coli, their performance in identifying Shigella species was unknown due to the absence of Shigella species in the collected body fluids. To assess their ability to differentiate E. coli and Shigella, Nanopore 16S was performed using isolates of E. coli, S. sonnei, and S. flexneri, and the sequencing reads were analyzed using the three pipelines.

As shown in Table 1, both Epi2me and NanoCLUST could not differentiate E. coli and the two Shigella species, as they resulted in the same classified species, E. fergusonii. Furthermore, the relative abundance of E. fergusonii was much lower in Epi2me as most of the reads were classified only at the family level. In contrast, Emu correctly differentiated the three species.

Table 2 presents the average number of classified species obtained from Nanopore 16S coupled with Epi2me, Emu, and NanoCLUST, respectively, for various types of body fluids. Overall, ascitic fluid and peritoneal fluid exhibited a relatively higher average number of classified species compared to other body fluids, with most of these species belonging to the gut microbiota. It is worth noting that most of the polymicrobial body fluids observed in this study were ascitic fluid and peritoneal fluid. However, these gut microbes are often overlooked in culture-based methods since it is impractical to identify numerous gut microbes through culture.

A ROC analysis was conducted to establish the threshold of relative abundance for detecting potential pathogens using Nanopore 16S (Figure 5). In monomicrobial samples, the determined T_RA for Epi2me, Emu, and NanoCLUST were 0.043, 0.066, and 0.066, respectively, with a sensitivity and specificity equal to or greater than 90%. The average threshold across the three analysis pipelines was considered the Nanopore 16S threshold, which was 0.058. Applying this threshold correctly identified 107 (83.6%), 117 (91.4%), and 114 (91.2%) monomicrobial samples for Epi2me, Emu, and NanoCLUST, respectively. Notably, this included 20 (15.6%), 29 (22.7%), and 23 (18.0%) samples that were not identified as the most abundant species in Epi2me, Emu, and NanoCLUST, respectively.

However, T_RA could not be determined in the ROC analysis of polymicrobial samples. Nanopore 16S appeared to function as a random classifier in polymicrobial samples as numerous species exclusively detected by Nanopore 16S were considered “false positives” when compared to the culture results. Nevertheless, these “false positives” may actually be present in the body fluids but overlooked in culture-based methods. If the threshold of 0.058 was applied to the polymicrobial samples, the total number of cultured species detected decreased to 46 (20.0%), 59 (25.7%), and 64 (27.8%) for Epi2me, Emu, and NanoCLUST, respectively. This suggests in polymicrobial samples, most species had a relative abundance below the threshold of 0.058, which could be attributed to the presence of dominant species or the high complexity of the bacterial population in the sample.

When applying the T_RA of 0.058, 13 out of the 20 culture-negative body fluids revealed the presence of clinically important species, including some opportunistic pathogens. The sequencing results were summarized in Supplementary Table 6. The clinical background of the patients was examined for these thirteen Nanopore 16S-positive samples (Table 3). Particularly, the inferred species from the Nanopore 16S results corresponded to the patients’ medical history in two samples (19MB090618 and 20MB068970). In the case of 19MB090618, Streptococcus pneumoniae was inferred by Nanopore 16S, which matched the detection of pneumococcal antigen in urine samples. Similarly, Prevotella detected by Nanopore 16S in 20MB068970 correlated with the isolation of Prevotella from a previous sample. The main reason for the inability to detect the Nanopore 16S inferred species in culture was the prescription of empirical antibiotics. Out of the thirteen patients, eleven had been prescribed antibiotics prior to sample collection. However, for the two samples without antibiotic treatment (21M2019392 and 21M2019576), the positive Nanopore 16S results were likely due to contamination from skin flora. In the case of 21M2019576, the detection of E. coli was also expected considering the presence of a duodenal ulcer.

To determine the limit of detection (LOD) of Nanopore 16S, one Gram-positive (S. aureus) and one Gram-negative (K. pneumoniae) pathogen, commonly found in body fluids, were spiked in EDTA-blood at various dilutions (Figure 6). During the LOD test, Nanopore 16S showed sensitivity to environmental contamination when the bacterial load in the sample was extremely low. Consequently, some environmental bacteria such as Pelomonas saccharophila, Cutibacterium acnes, and Stenotrophomonas rhizophila were detected. Based on the T_RA calculated from the ROC analysis (0.058), the LOD of Nanopore 16S for S. aureus and K. pneumoniae was determined to be 89.32 CFU/ml and 14.47 CFU/ml, respectively. Therefore, the overall LOD was approximately 90 CFU/ml.

Figure 7 illustrates the workflow of Nanopore 16S compared to conventional culture-based identification. The culture-based identification typically requires at least 24 h to obtain results, with even longer timeframes for fastidious bacteria. In contrast, Nanopore 16S coupled with Epi2me provides preliminary results within just 6 h. Among the 127 concordant monomicrobial samples (including closely related species) analyzed using Nanopore 16S coupled with Epi2me, all targeted species or closely related species were detected within the first hour of sequencing (Supplementary Table 7). Among the 50 concordant polymicrobial samples (including closely related species), 31 samples (62%) showed detection of all targeted species within 1 h of sequencing using Nanopore 16S coupled with Epi2me. Since the bacterial load varies greatly among the samples, the sequencing time was set to 24 h in this study to increase the sequencing depth of samples with extremely low DNA concentration. However, the sequencing time can be significantly shortened since the targeted species could be detected within the first few hours in most of the samples.