Stool samples submitted to Douglass Hanly Moir Pathology (DHM) in Australia for conventional pathology testing were selected retrospectively based on positive identification of one or more targets in the mNGS assay (Figure 1A). These samples were from Australians with gastrointestinal symptoms referred by a treating doctor for standard infectious disease testing. Primary samples included archived material frozen at −80 °C for between 1 and 6 months, reference material, and samples undergoing contemporaneous testing. A secondary aliquot from each primary sample was taken with a Copan flocked swab (50 to 200 mg) and sent to Microba Laboratories (ISO15189 accredited) for mNGS assay testing (Figure 1B). DHM provided Microba with samples for 510 clinical specimens with results from one or more conventional diagnostic tests: PCR, MALDI-TOF, MCS (microscopy, culture, and sensitivity), and/or antigen test (Figure 1C; Supplementary Table S2). Additional testing was performed by Microba using Seegene PCR assays (Supplementary Table S2) to (i) supplement samples otherwise lacking conventional test results for study targets, (ii) evaluate the relationship between PCR cycle threshold (C_t) values and mNGS informative reads (Figure 1D), and (iii) resolve discrepant results. Biological replicates were also taken from 158 samples to assess mNGS assay reproducibility (Figure 1E).

Contrived samples used to determine analytical performance were created from whole organism isolates using a homologous faecal matrix pre-screened with mNGS and PCR assays to confirm the absence of target organisms (Figure 1F). Analytical targets were selected to span as many different target types (bacteria, virus, VF, and AMR) as possible, accommodating the availability of reference material. The base biomass of the matrix was 4.19 to 8.67 × 10¹⁰ equivalent organisms/g (equ. orgs/g; avg. = 6.23 × 10¹⁰) assuming an average genome size of 2.8 to 5.8 Mb (average = 3.9) for stool microorganisms (Nayfach and Pollard, 2015). This aligns with the expected range for faecal material (Sender et al., 2016; Fernández-Pato et al., 2024). Isolate cells were quantified using in-house methods, which included evaluation of cells per gram of faeces with the Zymobiomics Femto Bacterial and Fungal DNA Quantification Kit (Zymo Research, CA, USA) and confirmation by microscopy-based cell counting. Calculated equivalent organisms were added to the matrix with serial dilutions covering concentrations from 10⁸ to 10⁴ orgs/g. Each set of contrived samples was processed in replicates of 6, created by different technicians using varied reagent lots on different processing runs. Two independent manufacturing runs were conducted for selected targets to verify sample creation and assay performance.

Samples were processed at Microba Laboratories (ISO15189 accredited) using version-controlled protocols, analytical pipelines, and reference databases (Figure 1B) as previously described (Angel et al., 2024). Briefly, 1 ng of genomic DNA was used for library construction using the DNA Prep Kit with DNA UD Index sets A-D (Illumina, CA, USA) and sequenced on the NovaSeq 6000 (Illumina, CA, USA) in 2 × 150 bp format, generating a minimum of 16 million non-human high-quality read pairs per sample. Each processing run met minimum sequencing quality and control sample standards, including processing of sentinel negative controls for extraction and sequencing reagents and confirmation of absence of any mNGS assay targets in these controls (see Angel et al., 2024 for additional details). Sequencing reads were filtered to remove low-quality reads using Trimmomatic (Bolger et al., 2014), and reads with high-quality alignment to the GRCh38 human genome were discarded (2.8% of reads on average). Quality-controlled reads were mapped to genomes in Microba’s reference databases, and taxonomic profiling was performed using a proprietary bioinformatic pipeline utilising the Microba Community Profiler (Parks et al., 2021) and a database indicating genomic loci that are informative of a specific target. Informative loci were identified by comparing short genomic regions across reference genomes within a genus to identify loci that are nearly ubiquitous within a target species and absent from all other species in the genus. Reads mapping to informative loci were then used as evidence for the presence of a target, which allows for the robust identification of species at low relative abundance. Target AMR and VF genes were identified by mapping reads to Microba’s reference gene database. A gene target was considered present if reads (i) covered a sufficient portion of the gene and (ii) supported all single-nucleotide variants (SNVs) required for the target to be clinically relevant (e.g., SNVs required for a gene to confer antibiotic resistance). The clinical report classifies mNGS assay targets as “detected” or “not detected” based on target-specific criteria that include the number of sequencing reads assigned to informative loci within a pathogen target, or percent coverage and identified SNVs for a gene target.

Conventional tests can provide less taxonomic resolution than mNGS assays. Consequently, the specific species that produces a positive conventional test result may not be resolved. Notably, 5 of the 11 bacterial targets evaluated using Seegene PCR assays specify a target genus (e.g., Aeromonas spp., Campylobacter spp.; Supplementary Table S3). To evaluate the mNGS assay, targets corresponding to each reference test were established using available vendor specifications and expanded based on preliminary assessment of the mNGS assay (Supplementary Table S3).

The consensus of the conventional test results was compared to the mNGS assay test result, and discrepant findings were resolved by additional testing (Supplementary Figure S1). Discrepancy testing was performed in-house using Seegene PCR assays for which Microba Laboratories is a certified pathology service provider. Samples were processed on the Seegene CFX96™ Real-time PCR System using the following Allplex™ assays: GI-Bacteria (I) (Cat. No. G19801Y), GI-Bacteria (II) (Cat. No. G19702Y), GI-Helminth (I) (Cat. No. GI10189Z), GI-Parasite (Cat. No. G19703Y), GI-Virus (Cat. No. G19701X), and H. pylori & ClariR (Cat. No. HC10389Z). Validation outcomes were obtainable for 61 to 497 samples depending on the target (Supplementary Table S4).

Cycle threshold (C_t) values obtained by Microba Laboratories with Seegene PCR assays were compared to the number of informative reads identified by the mNGS assay, that is, reads mapping to informative loci in target species or to any portion of a target gene. Targets were considered absent for C_t values above 40, 43, 45, or 50 as specified by the manufacturer (Supplementary Table S3). The average C_t value was used for samples with multiple PCR test results. For reference targets with multiple mNGS assay targets (Supplementary Table S3), the highest number of informative reads was used.

In silico samples were created for all pathogen and gene targets in the mNGS assay (Figure 1G). Read pairs were generated at four sampling depths, corresponding to a fixed number of read pairs for pathogens and a specific coverage for genes (Supplementary Table S5). At the standard mNGS assay depth of 16 million read pairs, pathogen targets comprised 0.0001% (ultra-low) to 0.1% (high) of read pairs, while gene targets averaged 0.0000116% (ultra-low) to 0.000116% (high) abundance. Read pairs were generated from three independent reference genomes for each target using InSilicoSeq v1.6.0 with its NovaSeq error model (Gourlé et al., 2019), with genomes not in Microba’s reference genome databases used when available to simulate true biological variability. For targets with fewer than three genomes, independent read sets were generated using the available genomes to ensure three read sets per target. In-silico target reads were added to three faecal samples pre-screened to confirm they contained few mNGS assay targets, with any targets present being excluded when establishing the mNGS assay performance (Supplementary Table S6). Targets were divided into 22 sets covering all 176 mNGS targets to manage sample volume (Supplementary Table S7). A total of 792 samples were required to generate mocks at four depths, with three read sets per target added to three faecal samples.

Results of the mNGS assay were compared to conventional diagnostic test results for clinical samples (Supplementary Figure S1), or known results for contrived and in silico samples to establish if a sample represented a true positive (TP), true negative (TN), false positive (FP), or false negative (FN). Performance of the mNGS assay was then assessed using sensitivity [TP/(TP + FN)], specificity [TN/(TN + FP)], positive predictive value [PPV = TP/(TP + FP)], and negative predictive value [NPV = TN/(TN + FN)]. The Clopper-Pearson exact method (Clopper and Pearson, 1934) was used to determine 95% confidence intervals (CIs) for these performance measures.

Test results from the mNGS assay were compared with conventional pathology testing for eight bacterial, four protozoa, three invertebrate, one viral, and six VF targets across 510 faecal clinical samples (Figure 1C; Supplementary Table S8). The clinical samples were obtained from children (28.8% of samples <18 years; 58.5% male) and adults (71.2% of samples; 52.3% female) presenting with gastrointestinal symptoms. Conventional testing of these faecal samples was not uniform due to technical limitations and individual clinician choice of test (Supplementary Table S4), and the prevalence of targets in these samples varied substantially (Supplementary Table S9). Using conventional pathology testing, 268 (52.5%) of the 510 clinical samples tested positive for one target, 171 (33.5%) were positive for more than one target, and 71 (13.9%) were negative for all study targets.

The mNGS assay exhibited strong performance on the majority of evaluated targets. Specificity was ≥99% for all targets except Adenovirus F (serotype 40/41) with 96.8% specificity, and the assay achieved high positive predictive value (PPV: average 98.1%; median 100.0%) and negative predictive value (NPV: average 95.7%; median 99.1%) across all targets (Table 1). Average sensitivity across the 16 target pathogens was 80.7% (median 91.0%) and increased to 89.0% (median 95.6%) with the removal of Helicobacter pylori and Entamoeba histolytica. The sensitivity of VFs (average 56.3%; median 58.7%) was generally lower, which can be attributed to the smaller genomic regions of these targets and, consequently, the lower likelihood of obtaining DNA sequencing reads from these regions. The broad coverage of targets in the mNGS assay resulted in 256 of 510 (50.2%) samples having 1 or more additional pathogens or VFs identified, including 83 samples with E. coli pathotypes and a confirmed Tropheryma whipplei case, compared to current standard-of-care pathology (Supplementary Table S10). An additional 181 samples had one or more AMR genes detected, including nine samples with carbapenemase bla_OXA-23 or bla_OXA-48.

PCR is the most comparable conventional diagnostic method to the mNGS assay, as both are nucleic acid tests, in contrast to MCS, MALDI-TOF, and antigen-based tests, which evaluate the viability or expression of targets. Comparing mNGS assay results to PCR resulted in sensitivity increasing to 93.9% for Salmonella spp. (16.4% increase), 85.0% for Vibrio spp. (8.8% increase), and 72.7% for Aeromonas spp. (6% increase) with sensitivity of Cryptosporidium spp. decreasing to 93.5% (6.5% decrease; Supplementary Table S11). Specificity, PPV, and NPV all remained the same or improved when considering just PCR as a reference diagnostic test, with the exception of NPV, which decreased slightly for Cryptosporidium spp. (100% to 99.3%), Entamoeba histolytica (99.4% to 98.3%), and Giardia intestinalis (98.9% to 97.1%).

We explored the relationship between the Seegene PCR cycle threshold (C_t) values obtained by Microba Laboratories and the number of informative sequencing reads identified by the mNGS assay (Figure 1D). The number of mNGS informative reads (log₂) was found to be inversely proportional to the C_t value obtained by Seegene PCR assays (Figure 2). As expected, a substantial number of false negative (FN) results between the mNGS and PCR assays occur at high C_t values, as PCR test results are less reproducible at higher C_t values (Caraguel et al., 2011; Rhoads et al., 2021), and samples with sufficiently low pathogen load may not reach the limit of detection of the mNGS assay. Consequently, the sensitivity of the mNGS assay was found to increase steadily when compared to PCR test results on samples with a C_t value less than a specified threshold (Figure 3).

In agreement with previous studies (Boers et al., 2020), a substantial number of FN results between the mNGS and PCR assay test results were observed to occur at C_t values ≥35. Specifically, the sensitivity of the mNGS assay on samples with C_t ≥ 35 was 20.2% compared to 79.6% for samples with C_t < 35 (Supplementary Table S12). This improvement is consistent across all target groups with the average sensitivity of bacterial targets increasing from 42.3% to 97.5%, protozoan targets from 13.3% to 73.7%, Adenovirus F (serotype 40/41) from 0% to 29.6%, and VF targets from 2.0% to 63.8% (no invertebrate targets had a C_t ≥ 35). Targets with >20% improvements in sensitivity when considering only test results with a C_t < 35 as opposed to all test results include Aeromonas spp. (54.2% to 91.7%), H. pylori (6.25% to 100%), Salmonella spp. (68.2% to 96.8%), and the ETEC heat-labile/stable toxins (53.3% to 76.2%). Notably, Yersinia enterocolitica and Campylobacter spp. appeared to be relatively more sensitive on samples with a C_t ≥ 35, with 17 of 20 (85.0%) and 11 of 15 (73.3%) samples being detected, respectively; though sensitivity still improved for these targets when considering samples with C_t < 35 (Y. enterocolitica = 100%; Campylobacter spp. = 96.8%).

Biological triplicates were taken for 158 clinical specimens to evaluate the reproducibility of the mNGS assay (Figure 1E). Reproducibility was evaluated by determining if the results of the mNGS assay were identical for a target across all three biological replicates. The assay evaluates 176 targets per specimen, resulting in a total of 27,808 (i.e., 158 specimens × 176 targets) test results of which 27,656 (99.5%) were concordant across all replicates (Supplementary Table S13). The majority of targets (114/176; 64.7%) were 100% or `triple/absolute’ concordant across all replicates. Of the targets with discordant results, only 2/176 (1.1%) exhibited concordance in <95% of specimens (Campylobacter concisus: 91.1% and rmtD: 89.9%).

Contrived samples were created for three bacterial, one viral, 14 AMR, and one VF target by adding isolate cells into the faecal matrix at equivalent organisms/g faeces decreasing from 10⁸ to 10⁴, except for Adenovirus F (serotype 40/41), where equ. orgs/g decreased from 2 × 10¹⁰ to 2 × 10⁶, to establish the limit of detection for a range of mNGS targets (Figure 1F). Contrived samples were created in replicates of 6, with select samples independently created twice, denoted by repeat (Table 2; see Methods). Sensitivity of target pathogens attained 100% at equ. orgs/g ≥ 10⁶ with intermittent identification of Listeria monocytogenes and Pseudomonas aeruginosa at 10⁵ equ. orgs/g. AMR and VF genes generally required 10⁸ equ. orgs/g to be robustly identified with bla_GES, bla_lMP, and qnrA proving challenging to identify even at this concentration. Similarly, Adenovirus F (serotype 40/41) attained 100% sensitivity at 2 × 10⁸ equ. orgs/g.

In silico faecal samples were generated by adding in silico reads for all 176 mNGS targets at four read depths into three independent faecal samples in triplicate (Figure 1G; Supplementary Table S5), with ultra-low to low read depth expected to correspond to the limit of detection established using contrived samples (Supplementary Figure S2, Supplementary Note S1). The 792 in silico faecal samples were processed with the mNGS assay and collated to determine the performance of the assay on each target (Supplementary Table S14).

The performance of the mNGS assay was summarised for increasing read depths for each taxonomic group of pathogens and for the AMR and VF targets (Figure 4). Sensitivity of pathogen targets increased with read depth with all bacterial and eukaryotic targets being identified at ultra-low depth except for Aeromonas veronii, A. hydrophila, Salmonella bongori, S. enterica, and Anncaliia algerae (Table 3; Supplementary Table S14). In contrast, identification of viral targets required at least low read depth (average sensitivity = 72.7%), with a few targets requiring medium read depth before they were identified and sensitivity reached 100%. Mean sensitivity was >97.8% for all gene target groups at low to high read depth (Table 4), with the only gene targets resulting in FN predictions above ultra-low read depth being CTX-M-G1, bla_KPC, bla_SIM, and EHEC O157:H7 VF genes (Supplementary Table 14). The mNGS assay had ≥95% mean specificity, PPV, and NPV for all pathogen and gene groups at all read depths, with these performance statistics often exceeding ≥99%. Notably, FP predictions only occurred for nine of 176 (5.1%) targets (Supplementary Note S2, Supplementary Table S15).