A retrospective cohort dataset was obtained from 398 patients who visited a tertiary care private hospital in Bangkok for suspected UTI or routine monitoring between January 2018 and December 2020. The characteristics of the cohort are presented in Table 1; Supplementary Table S1. The median age of the overall cohort was 64 years, and the majority of patients were female (277, 69.6%). Samples were included for analysis if they contained more than 500 non-human reads after preprocessing and removal of human DNA sequences. A total of 368 samples satisfied these criteria and were included in the final analysis. The Bumrungrad International Hospital’s Institutional Review Board reviewed and approved the study, which used anonymized data (BI-IRB #312-02–23 CIEN-B Fub). No formal power calculation was performed for this retrospective analytical validation. Sample size (N = 398) was determined by available archived samples meeting inclusion criteria during the study period.

Urine was streaked onto blood agar and MacConkey agar without centrifugation. Plates were incubated aerobically at 35 °C for 24 hours. Pathogen identification used MALDI-TOF MS (VITEK MS v3.2, bioMérieux). We utilized a threshold of ≥10³ CFU/mL to define a positive urine culture. This threshold was selected to maximize sensitivity for the detection of potential uropathogens. Antimicrobial susceptibility testing employed VITEK 2 system (bioMérieux) with turbidimetric monitoring for up to 36 hours to determine minimum inhibitory concentrations.

Ten-milliliter urine aliquots were processed within 4 hours. Human cells were removed by centrifuge (300g×2min), then microbes pelleted (5000g×15min). Human DNA was depleted using Molysis Basic 5 Kit (Molzym D-301-100); microbial DNA was extracted using Maxwell RSC Viral kit (Promega AS1330). Following extraction, DNA was stored at -20 °C to facilitate sample batching. Samples ≥ 0.1ng/µL underwent library preparation using ONT Rapid PCR Barcoding Kit (SQK-RPB004, 14 cycles). Pooled libraries were sequenced on MinION flow cells (FLO-MIN106D R9.4.1) using MinKNOW v22.12 for 24h, with 6 multiplexed. All urine specimens were processed internally at the CAP-accredited Bumrungrad International Hospital laboratory by well-trained laboratory personnel, ensuring consistency and reliability in sample handling and analytical workflow (https://www.bumrungrad.com/laboratory/en).

The 500 non-human read threshold was established to ensure sufficient sequencing depth for reliable organism identification by the BIOTIA-DX classifier. There were 30 samples that failed this threshold, 8 of which were due to low DNA yields (<0.5 nq/uL). The other 22 samples exhibited normal DNA yields but may have failed at library preparation or sequencing stages.

After preprocessing steps removing low-quality reads and reads mapping to the human genome, the BIOTIA-DX (BDX) pipeline identifies microbes present in a metagenomics sample through a two-step process. First, the non-human reads are aligned to a large database of microbial genomes in a coarse classification step. Organisms passing a pre-determined threshold in coarse classification are then used in a fine classification step which uses machine learning (Couto-Rodriguez et al., 2022, 2024) to statistically quantify whether the organism is present or absent in the sample. Reads are aligned to curated pangenomes (Hyun et al., 2022) for each organism identified in the coarse classification step and summary statistics describing alignment quality and genome coverage are calculated for each organism. These statistics are then fed into the machine learning classifier which assigns a confidence score ranging from 0 to 1 for whether the organism was present (score=1) or absent (score=0). The BIOTIA-DX machine learning classifier was intentionally trained and designed to increase stringency and decrease false positive detection of urogenital commensals or opportunistic pathogens present at colonization levels. Organisms with a BIOTIA-DX confidence score greater than 0.5 were considered to be positive identifications. In addition, the pipeline is designed to provide identification as soon as sufficient reads are generated.

The BDX pipeline covers general detection of some of the most prevalent and relevant gene markers in urine, including beta-lactamases (blaOXA, blaNDM, blaCTX-M, blaTEM, blaKPC, blaCMY, blaADC, blaSHV, blaVIM, blaPDC, blaZ, cfxA), mec gene classes (mecA-D), aminoglycoside, sulfonamide (sul1-4) and trimethoprim (dfr) resistance markers. A comprehensive database was constructed from the AMRFinder and CARD databases (Feldgarden et al., 2021; Alcock et al., 2022). As a first step, sequencing reads are aligned to AMR gene clusters, and alignment statistics are extracted. After filtering alignments with low-quality metrics, gene sequences are reconstructed and translated into amino acid sequences. Tripeptide fragments of the protein sequences are then fed into a shallow neural network for gene identification. Additionally, fluoroquinolone resistance in E. coli is monitored using a protein variant-based machine learning classifier trained on isolates from the BV-BRC database (Couto-Rodriguez et al., 2024). The model encompasses mutations in gyrA, gyrB, parC, and parE proteins.

For comparison of organisms identified by culture and by BIOTIA-DX, we considered cases where culture and BIOTIA-DX identified the same organism as a true positive. Where culture and BIOTIA-DX were both negative, we considered the case a true negative. Where culture identified an organism that was not identified by BIOTIA-DX with prediction probability > 0.5, we labeled the case false negative. In cases where a culture-identified organism did not pass the coarse classification stage of BIOTIA-DX and thus was not assigned a prediction probability during the fine classification stage, we excluded the organism from the analysis. As we cannot definitively determine whether organisms identified by BIOTIA-DX but not by culture are false positive identifications by BIOTIA-DX or false negative detections by culture, we simply counted these organisms as ‘additional positives.’ Thus, we were able to calculate the sensitivity of BIOTIA-DX compared to culture (true positives/true positives + false negatives), but we were not able to calculate a value for specificity.

We also assessed the performance of BIOTIA-DX at the overall sample level. If at least one microbe was identified by culture and at least one microbe was detected by BIOTIA-DX, we considered the sample to be in agreement between the two methods. The sample was considered to be a complete agreement if all organisms detected by culture were also detected by BIOTIA-DX. In the case that some, but not all microbes identified by culture were identified by BIOTIA-DX (or vice versa), the sample was considered in partial agreement. This classification was applied when both platforms identified members of the same established taxonomic complex or genus such as the Acinetobacter baumannii (ACB) or Klebsiella pneumoniae (KpSC) complexes where high proteomic similarity often limits the discriminative power of MALDI-TOF MS. Thus, partial agreement reflects concordance at the genus/complex level while acknowledging the superior species-level resolution of mNGS. The sample was also considered a partial agreement if the two methods both found the sample to be positive but identified the species differently. Finally, the methods were considered in disagreement when culture identified one or more microbes, but BIOTIA-DX did not identify any.

For comparison of AMR calls, we considered BIOTIA-DX and AST to be in agreement if either of the following were true: the isolate identified by culture was marked as intermediate or resistant to at least one tested drug in a drug class and BIOTIA-DX identified at least one antimicrobial resistance gene (ARG) to that drug class; OR the isolate identified by culture was susceptible to all tested drugs in a drug class and BIOTIA-DX did not identify any ARGs to that drug class. In any case where BIOTIA-DX did not identify resistance genes in a phenotypically resistant or intermediate isolate, we marked the sample as a disagreement. When BIOTIA-DX identified ARGs in a sample where the isolate was susceptible by AST, we also marked the sample as a disagreement; however, many of these samples contained multiple organisms, so we noted that it was possible that the discrepant ARGs belonged to a different organism in the same sample. For isolates found to be resistant to drugs that were only tested in combination (e.g., sulfamethoxazole/trimethoprim), we required ARGs from both drug classes to be present to mark the sample as an agreement, thereby minimizing the potential for over-interpretation of the genotypic data.

Samples were classified as culture-positive or culture-negative. Culture-positive means any bacterial organism identified by standard culture methods, whereas culture-negative means no bacterial growth on standard urine culture. Fungal organisms were excluded from analysis because the DNA extraction method used (Molysis Basic 5 Kit) does not efficiently lyse fungal cells, making BIOTIA-DX unable to detect fungi regardless of their presence. Additionally, Corynebacterium imitans and Lactobacillus spp. were excluded as they likely represent urogenital commensals rather than infection. Culture identifications lacking any genomic signal in BIOTIA-DX (no reads aligned in coarse classification, n=24) were excluded from organism-level concordance calculations as likely culture contaminations, while maintaining conservative sample-level classifications.

Overall, 24 different bacterial species and 4 fungal species were identified through standard urine culture in this cohort (Table 2). Escherichia coli was the dominant species (n=157), followed by Klebsiella pneumoniae (n=32) and Enterococcus faecalis (n=30), all of which are well-known uropathogens. Staphylococcus species were the next most common organism detected by culture, identified as Staphylococcus aureus (n=5), Staphylococcus epidermidis (n=3), Staphylococcus haemolyticus (n=2), and Staphylococcus saprophyticus (n=1). Other common organisms included Pseudomonas aeruginosa (n=10), Streptococcus agalactiae (n=6), Proteus mirabilis (n=5), Enterobacter cloacae (n=5), Citrobacter koseri (n=5), Enterococcus faecium (n=4), and Acinetobacter baumannii (n=4). The remaining 11 bacterial species identified by culture were each detected two or fewer times: Actinomyces urogenitalis, Aeromonas caviae, Burkholderia cepacia, Citrobacter freundii, Corynebacterium imitans, Enterococcus raffinosus, Klebsiella aerogenes, Lactobacillus species, Morganella morganii, Serratia marcescens, and Stenotrophomonas maltophilia. Fungal species were also common, with Candida albicans (n=7) being the most prevalent, followed by Candida glabrata (n=3), Candida tropicalis (n=2), and Candida krusei (n=1).

Samples were considered suitable for BIOTIA-DX analysis if greater than 500 reads were present in the sequencing data after preprocessing and human DNA removal. Of the 398 samples in this study, 30 samples either did not exceed this threshold, were contaminated, or were not sequenced and were excluded, resulting in 368 samples suitable for analysis. Of these 368 samples, 260 were positive by culture and 108 were negative or inconclusive. There were 24 organisms across the culture positive samples where BIOTIA-DX did not find any genomic signal of the organism in the coarse classification stage (Supplementary Table S2A). Because lack of a genomic signal is extremely unlikely in samples where an organism is present at the clinical threshold (≥10³ CFU/mL), we categorized these 24 cases as culture-positive with no detected genomic signal (a present organism missed by BIOTIA-DX will almost always pass coarse classification but have a prediction probability < 0.5), these identifications may represent false positive identifications by culture and were removed from the analysis. These discrepancies where a culture-positive organism exhibited no detected genomic signal may be partly due to DNA degradation or other preprocessing factors, reflecting the complexities of real-world clinical settings (El Bali et al., 2014; Bundgaard-Nielsen et al., 2020). In 15 of 24 cases, the removal of the culture-positive organism with no detected genomic signal left no culture-positive organisms in the sample, and these samples were reclassified as culture-negative (for bacteria). The remaining 9 cases had at least one additional organism called by culture and remained classified as culture positive.

Additionally, due to limitations in the extraction method, we were unable to analyze fungal targets using BIOTIA-DX; therefore, 10 samples positive only for a Candida species were also considered culture-negative. The remaining 2 samples with Candida identifications were culture-positive for an additional bacterial species. Finally, 4 organisms identified by culture are not considered typical causative agents of UTIs and were also excluded from BIOTIA-DX analysis (Corynebacterium imitans and Lactobacillus species). These re-classifications resulted in 231 culture-positive samples and 137 culture-negative samples (Figure 1).

We next then compared microbes identified by BIOTIA-DX to those identified by urine culture at both the sample and organism level. A complete list of all BIOTIA-DX identifications along with prediction probabilities is available in Supplementary Table S2A.

At the sample level, the comparison revealed a high degree of concordance for culture-positive samples. Of the 231 culture-positive samples, BIOTIA-DX identified at least one microbe in 228 samples (98.7%) and identified no microbes in 3 of the samples (1.3%) (Figure 2A). In samples where microbes were identified, 173 samples (74.9%) were in complete agreement between culture and BIOTIA-DX. In 36 cases (15.6%), BIOTIA-DX identified the culture-positive organism but also found additional organisms not identified by culture. In 13 cases (5.6%), at least one microbe identified by BIOTIA-DX was different than the one identified by culture. Finally, in 6 cases, BIOTIA-DX did not identify one of the organisms present in a co- or polymicrobial infection that was identified by culture (2.6%).

At the organism level, a total of 242 organisms were identified by urine culture across the 231 culture-positive samples (Figure 2B). BIOTIA-DX correctly identified 229 of these targets with confidence scores > 0.5, resulting in an overall organism-level sensitivity of 94.6% (Supplementary Table S2A). Notably, 5 of these targets were identified as a closely related species in the same genus by BIOTIA-DX (Table 3). Because of the likelihood that these were simply misclassified by culture, we considered these identifications to be true positives and in agreement with the culture result.

There were 13 targets identified by culture but not by BIOTIA-DX due to prediction probability scores < 0.5 (Table 4). Of these 13 targets, more than half (n=8) had borderline scores between 0.4 and 0.5. Capturing these borderline calls would have increased the organism-level sensitivity from 94.6% to 97.9%. As BIOTIA-DX is trained on short-read sequencing data, these results indicate that the BDX assay is also highly effective at detecting the same pathogens identified by standard urine culture using long-read sequencing, which has a higher error rate than short-read technologies.

For culture-negative samples (n=137), a much greater degree of discordance was observed between culture and BIOTIA-DX. Only 39 of 137 (28.5%) culture-negative samples were found to be completely negative by BIOTIA-DX (Supplementary Table S2A). In the remaining 98 culture-negative samples, BIOTIA-DX identified one urogenital pathogen in 54 samples (39.4%), two co-infecting pathogens in 23 samples (16.8%), and polymicrobial infections of 3 or more pathogens in 21 samples (15.3%) (Figure 2C), demonstrating that mNGS identifies microbial signatures in a substantial number of cases deemed negative by the current standard. However, the clinical significance of these detections (infection vs. colonization vs. contamination) requires prospective validation with clinical correlation.

In total, across all 368 samples, BIOTIA-DX identified 264 organisms that were not identified by standard urine culture. The frequencies of organisms identified by BIOTIA-DX that were not identified by culture are shown in the second column of Supplementary Table S2. Eighty-four of the 264 additional identifications were identified in culture-positive samples known to contain at least one other organism, and 180 organisms were found across the 137 culture-negative samples. In general, the organisms identified by BIOTIA-DX that were not identified by standard urine culture were either anaerobic or fastidious organisms, with 204 out of the 264 detections (77.3%) falling into one of these two categories (Figure 2D).

The most commonly identified pathogen found by BIOTA-DX but not by culture was Gardnerella vaginalis (n=56, 15.2% of all samples) (Table 5; Supplementary Table S2), an anaerobic organism often associated with bacterial vaginosis (BV) that also plays a significant role in UTIs. Other anaerobic taxa that can cause UTIs frequently identified by BIOTIA-DX but not by culture included Fannyhessea vaginae (n=22, 6.0%), Peptoniphilus species (n=19, 5.2%), and Anaerococcus species (n=24, 6.6%). In total, 150 anaerobic organisms were identified in 80 different samples (21.7% of all samples, 56.8% of the 264 additional BIOTIA-DX identifications). While it is unclear which of these organisms may represent opportunistic infection or synergistic growth alongside true causative anaerobic organisms, anaerobic organisms nonetheless constitute a large burden of undiagnosed infections in UTI samples. Anaerobic urinary tract infections can be clinically serious, especially in the setting of complicated UTIs or immunocompromised patients, and are frequently missed by standard urine culture. No anaerobic organisms were detected by standard urine culture in the entire cohort, demonstrating a major diagnostic advantage of BIOTIA-DX over culture.

The second most commonly identified pathogen found by BIOTIA-DX, but not by culture, was Streptococcus anginosus (n=24, 6.6% of all samples) (Table 5; Supplementary Table S2). Streptococcus species are facultatively anaerobic and considered to be fastidious, making them susceptible to missed detection via standard urine culture. Only S. agalactiae was identified by culture in 6 samples, and BIOTIA-DX identified it in an additional 5 samples, demonstrating it was potentially missed by culture in nearly 50% of cases. Following S. anginosus and S. agalactiae, viridans group streptococci were the third most commonly identified Streptococcus species (n=7, 1.9% of all samples), including S. mitis, S. cristatus, S. parasanguinis, and S. oralis. In total, BIOTIA-DX identified 40 cases (10.9% of all samples) of Streptococcus species that were not detected by culture. While Streptococcus species were the most common fastidious organism identified by BIOTIA-DX, other fastidious species identified by BIOTIA-DX included Actinotignum species (n=6), Oligella urethralis (n=6), and Winkia neuii (n=2). In total, 54 non-anaerobic fastidious organisms were identified (20.5% of the 264 additional BIOTIA-DX identifications). Together, these results demonstrate the additional significant advantage of BIOTIA-DX over urine culture in the identification of non-anaerobic fastidious organisms.

While anaerobic and fastidious organisms constituted the majority of BIOTIA-DX organisms not identified by culture, 60 of the 264 additional BIOTIA-DX detections were non-anaerobic and non-fastidious organisms identified across 51 unique samples (Supplementary Table S2A). For 47 of these 60 organisms (17.8% of the 264 additional BIOTIA-DX identifications), the infection was a co- or polymicrobial infection (n=38 samples). In 21 samples, a BIOTIA-DX organism was identified in addition to at least one co-infecting organism also identified by culture, and in 17 samples, none of the co- or polymicrobial infection organisms identified by BIOTIA-DX were detected by culture. In only 13 samples (3.5% of all samples, 4.9% of the 264 additional BIOTIA-DX identifications) was a single, non-anaerobic and non-fastidious organism identified by BIOTIA-DX but not by culture. In total, 284 detections of non-anaerobic non-fastidious organisms were made by either culture or BIOTIA-DX in this cohort, and with 60 identifications made only by BIOTIA-DX and not by culture, potentially 21.1% of all such infections may have been missed by culture. These results demonstrate that, even for non-anaerobic and non-fastidious organisms, BIOTIA-DX still has substantial utility over standard urine culture, particularly in the identification of co- and poly-microbial infections. The significance of these findings is amplified in hospitalized, catheterized, or immunocompromised patients, where co-infections are clinically meaningful and support more precise, risk-stratified treatment decisions.

The estimated sample-to-result timeline for the BDX platform is 11–13 hours. It supports the potential for same-day clinical reporting (< 24 hours), compared to the 24–48 hours typically required for standard urine culture. This includes a median of 5 hours for wet-lab processing (extraction and library preparation) followed by 6–8 hours of sequencing and automated analysis.

A total of 137 samples were considered to be culture-negative for BIOTIA-DX analysis. Within these samples, BIOTIA-DX predicted the presence of at least one ARG in 96 samples. Beta-lactam ARGs were found in 63/137 (46.0%) culture-negative samples, with 51 (37.2%) occurring in samples where BIOTIA-DX made at least one positive taxa call. Among beta-lactam ARGs, blaZ (n=27), mecA (n=18), and their operon elements were commonly detected. In 11/18 detections of mecA, a Staphylococcus species was also identified as additional taxa call by BIOTIA-DX. Other markers including blaTEM (n=32), cfxA (n=8), blaCTX-M (n=3), blaSHV (n=1), and blaOXA (n=3) were also identified. Either sul (n=12) or dfr (n=17) were detected in 24 (17.5%) culture-negative samples, 19 (13.9%) of which were also taxa-positive by BIOTIA-DX. Of these, 5 showed the presence of both sul and dfr genes, four of which were also taxa-positive. Aminoglycoside ARGs were detected in 73 (53.3%) culture-negative samples, 54 (39.4%) of which were taxa-positive by BIOTIA-DX. For aminoglycoside markers, aph(3’)-Ia (n=42), aph(3’)-IIIa (n=29), and ant(6)-Ia (n=28) were the most prevalent ARGs identified. Tetracycline resistance markers were the most prevalent in the culture-negative sample set (n=98, 71.5%) most commonly including tet(M) (n=80), tet(L) (n=34), and tet(W) (n=26).