This study was approved by the Ethics Committee of Northwestern University (Xi’an, China) and the Ethics Committee of the First Affiliated Hospital of Xi’an Medical College (Xi’an, China) and was conducted in compliance with the ethical guidelines of the Declaration of Helsinki of the World Medical Association. We randomly collected vaginal secretion samples from 4865 subjects from the First Affiliated Hospital of Xi’an Medical College from January 2022 to April 2022. The inclusion criteria were at least 18 years old, not pregnant, no current use of hormonal or barrier contraceptive products, vaginal douching, tobacco or alcohol abuse, no hospitalization or systemic use of medication for chronic diseases or antibiotics/probiotics within the 6 months before sample collection, and no intercourse in the day before sampling. And then, smear microscopy, Gram staining and drying chemoenzymatic method (Bacterial vaginosis Test kit, BioPerfectus Co., Ltd. China) were used to determine the following parameters in accordance with the provided instructions: pH, flora density and diversity, Donders’ score, Nugent score (Nugent et al., 1991), Lactobacillus, Trichomonas, blastospore, catalase, cleanliness, leukocyte esterase, sialidase, and other relevant indicators. According to the test results, the samples were divided into the following types: (1) Normal flora: the density of vaginal flora was grade II-III, the diversity of vaginal flora was grade II-III, the dominant bacteria was Lactobacillus spp., no other pathogens were detected and the vaginal microbial function was normal. (2) Aerobic vaginitis (AV): Donders’ score ≥ 3 points, Hydrogen peroxide, leukocyte esterase and β-glucuronidase/coagulase positive can be used for auxiliary diagnosis. (3) Vulvovaginal candidiasis (VVC): Gram positive ovoid blastospores or tubular pseudohyphae were observed under the oil immersion objective. (4) Trichomonas vaginitis (TV): a large number of white blood cells with Trichomonas spp. were observed. (5) BV: Nugent score ≥ 7 points, Clue cells positive, PH ≥ 4.4 and Whiff test positive. (6) Intermediate Microbiota (IBV): the Nugent score was 4 to 6 points. After excluding 3397 cases with other reproductive tract infections (such as VVC, TV, etc.), the samples were divided into two groups according to Nugent score: the 7–10 group (376 cases) and the 0–3 group (1092 cases). In the 7–10 group, 131 pregnant women, 16 patients with malignant tumors, 88 patients with other non-reproductive tract infections, 43 patients who had received treatment, and 23 patients with missing information or without symptoms were excluded, and 74 symptomatic BV patients were retained as the BV group in the modeling cohort. The healthy group in the modeling cohort consisted of 25 non-pregnant, asymptomatic healthy women who underwent pre-pregnancy physical examination in the 0–3 group. Then, 25 healthy individuals (NO group) and 74 BV samples (BV group) in the modeling cohort were analyzed to screen candidate Lactobacilli and BVPs, and then the BV diagnostic model was constructed and internally validated. Finally, another vaginal secretion samples from 10607 subjects were randomly collected from the Affiliated Hospital of Xi’an Medical College from May 2022 to March 2023. And 83 healthy samples, 76 BV samples and 30 intermediate flora samples were obtained for external data validation according to the sample screening method described above. Clinical information and the data of the Laboratory for all participants were recorded in detail ( Supplementary Table S2 ).

Two vaginal swabs per patient were sampled simultaneously. Sample secretions were collected from the lowest one third of the vagina by rotating the swabs (HCY Transsystem CY93050T swab, HCY technology, China) three times ensuring uniform distribution on both swabs. One sample was used for non-molecular tests, which performed at the laboratory of the First Affiliated Hospital of Xi’an Medical College. The second swab was used for the extraction and purification of microbial nucleic acids by The Upure Microbial DNA Kit^® (BioKeystone, Chengdu, China). Pure DNA was evaluated by the ratio of 260/280 with using NanoDrop 2000 (Thermo Scientific, Wilmington, DE, United States) and stored at −80°C for further use.

Our qPCR methods have not been described elsewhere and were developed in-house. The whole genomic sequences of 11 BVPs (Gardnerella vaginalis [G. vaginalis], Atopobium vaginae [A. vaginae], Bacteroides fragilis [B. fragilis], BVAB2, Prevotella bivia [P. bivia], Megasphaera type 1 [Megasphaera 1], Megasphaera type 2 [Megasphaera 2], Mobiluncus mulieris [M. mulieris], Mobiluncus curtisii [M.curtisii], Mycoplasma hominis [M. hominis], and Ureaplasma urealyticum [U. urealyticum]) and 4 lactobacilli (L. crispatus, L. gasseri, L. jensenii, and L. iners) were downloaded from the NCBI database. MAFFT software was used for global sequence alignment to screen-conserved sequences. Primers were designed using the Primer-BLAST (https://www.ncbi.nlm.nih.gov/tools/primer-blast/) online tool according to the primer design principles and commissioned to be synthesized by Bioengineering (Shanghai, China) Co., Ltd. The PCR reaction employed universal primers for the host gene (β-actin) to provide the internal positive control. The specificity of primers and probes were verified by sanger sequencing of the PCR products.

We used four independent reactions to achieve simultaneous detection of 15 microbial targets, with four targets per reaction. Probes were labeled with different fluorescent molecules.The mPCR amplification was performed using ABI 7500 platform (Applied Biosystems, Wilmington, DE, United States) in 20 μL of reaction mixture solution containing 5 μL of nucleic acid template, 4 μL of 5×buffer (FAPON Biotech Inc., Guangzhou, China), 2 μL of 10×solution I (FAPON Biotech Inc., Guangzhou, China), 1 μL of dNTP mix (FAPON Biotech Inc., Guangzhou, China), and 0.275 μL of Hotstart Hitaq polymerase mix (FAPON Biotech Inc., Guangzhou, China), and sterilized distilled water. The reaction conditions were as follows: 95°C for 5 min, followed by 40 cycles of 95°C for 50 s, and 60°C for 40 s.

Univariate analysis was used to compare the characteristics of 15 microorganisms between the healthy and the BV groups, and to screen indicators that could be used to construct diagnostic models. The differential factors or high-resolution indicators (sensitivity or specificity higher than 80%) were used as candidate features for model construction. Three machine learning algorithms, support vector machine (SVM), decision tree (DT), and random forest (RF) algorithm, were used for the construction of mixed diagnostic models of BV. The modeling cohort were randomly divided into a training dataset and a validation dataset in a ratio of 2:1 in the modeling process. We then internally validated our whole model development strategy using the validation dataset of modeling cohort. In addition, an independent validation dataset of 159 samples was used to verify the performance of diagnostic models. The predictive performance of diagnostic models was evaluated by a series of evaluation metrics, including area under the curve (AUC), sensitivity and specificity.

Quantification cycle (Ct) values obtained by the mPCR were utilized for subsequent data analysis, with a uniform assignment of 40 to microorganisms that were either undetected or had Ct values exceeding the maximum quantification cycle. A lower Ct value indicated a higher abundance of this organism in the sample. Data preprocessing, model constuction and statistical analysis were performed using perl (www.perl.org) and R (https://www.r-project.org/) software. The Kolmogorov–Smirnov and Shapiro–Wilk tests were used to test the normality of the data distribution. ANOVA, Pearson’s chi-squared test, and Fisher’s exact test were used to compare continuous and categorical data. Feature selection and diagnostic performance evaluation of the models were performed by differential analysis and ROC analysis, respectively. Statistical analysis were only performed for data with a sample size of more than 20 cases. A p value less than 0.05 was considered statistically significant.

Species-specific primers designed manually from variable regions are summarized in Supplementary Table S1 . In order to assess the sensitivity of the BVLaB assay, the detection limit of the mPCR was performed with different dilution of the template. As shown in Figure 2A and Supplementary Table S1 , the newly developed method was able to amplify all targets in a range of 10³–10⁸ copies μL⁻¹ with a strong linear relationship, and all 15 targets were detectable even as DNA template amounts was as low as 15 copies μL⁻¹. In addition, non-specific amplification of other organisms or non-template agents was undetected by the BVLaB assay. These results indicate that the new assay was proved to be highly sensitive for amplification of low quantities of DNA and thus could be applied to detect pathogenic microorganisms.

A total of 288 samples were collected for differential analysis and model construction in this study. Information regarding the clinical samples is summarized in the Table 1 . There were significant differences in the detection rates of 15 microorganisms in the samples. Overall, the detection rate of each microorganism ranged from 9.09% to 94.95%. Among BVPs, the detection rate of G. vaginalis was the highest (94.95%), followed by A. vaginae (77.78%), and that of Megasphaera 1 was the lowest (9.09%). Among the four species of Lactobacillus spp., L. iners could be detected in the largest number of samples, with the detection rate of 85.86% and 75.29% in the Lactobacilli cohort and modeling cohort, respectively, followed by L. crispatus (64.7% and 41.41% in the Lactobacilli and modeling cohort, respectively).

To determine whether the distribution of 11 BVPs were associated with healthy women and BV patients, 99 samples (NO = 25 and BV = 74) were independently tested by the assay in the modeling cohort. As shown in Figures 2B–E , the heatmap plot displayed the alteration of BVPs’ distribution between the two groups. And the PCA results indicated a significant difference in the distribution of the first two principal components between the two groups of samples. P. bivia (76.0%) and G. vaginalis (84.0%) were detected in more than half of the samples in the healthy group, while seven microbes (G. vaginalis [98.6%], P. bivia [81.1%], A. vaginae [97.3%], BVAB2 [77.0%], M. hominis [67.6%], Megasphaera 2 [63.5%], and U. urealyticum [51.4%]) were detected in the BV group. Five BVPs were detected more frequently in the BV group than in the healthy group: A. vaginae (p < 0.0001), BVAB2 (p < 0.0001), Megasphaera 2 (p = 0.001), M. hominis (p < 0.0001), G. vaginalis (p = 0.0138). As expected, these six BVPs were present in higher abundance in the BV group than in the healthy group: A. vaginae (p ≤ 0.0001), BVAB2 (p ≤ 0.0001), Megasphaera 2 (p ≤ 0.0001), M. hominis (p ≤ 0.0001), G. vaginalis (p = 0.0138), P. bivia (p = 0.0039).

Lactobacillus spp. is closely associated with the maintenance of a stable microenvironment in the FLGT. As presented in Figure 3A , There were statistically significant differences in the distribution of L. crispatus and L. jensenii in the NO and BV groups. The PCA results demonstrated that the distribution features of these four species of Lactobacillus spp. can be used to distinguish whether the subjects are BV positive ( Figure 3B ). The detection rate of L. iners was higher in the BV group than in the NO group, whereas that of L. crispatus and L. jensenii showed an opposite trend ( Figure 3C ). The abundance of L. crispatus and L. jensenii in the BV group was significantly lower than that in the NO group (Both p ≤ 0.0001) ( Figure 3D ). In addition, coexistence analysis ( Figure 3E ) showed that L. iners was the most predominant type in the BV group, mainly in samples with one and two species of Lactobacillus detected. There is a significant correlation between L. jensenii and L. gasseri, which may be related to the low detection rate of these two lactobacilli ( Figure 3E ).

The similarity in the distribution of microorganisms in different infection states might be used to predict the interactions between microorganisms. The detection rates and abundance levels of L. iners and all BVPs were higher in the BV group than in the NO group ( Figures 4A, B ). As expected from Figure 4C , we found positive correlations between the abundance of L. iners and some BVPs in both groups, but the microbial species of BVPs associated with L. iners were not consistent between the two groups. For example, in the NO group, the abundance of P. bivia and U. urealyticum showed significant positive correlations with L. iners, while L. iners was found to have a significant positive correlation with BVAB2 and A. vaginae in the BV group, which may indicate that L. iners may be an independent factor in the development of BV. In addition, we also found that there were correlations between various BVPs in the BV group, such as a significant positive correlation ermerged between the abundance of A. vaginae and P. bivia, G. vaginalis was also significantly positively correlated with M.curtisii, suggesting the synergistic effect of these BVPs on the occurrence of BV.

Given the observed differences in the distribution of microorganisms between healthy individuals and BV patients, the performance of the developed method for detecting and measuring abundance as a diagnostic indicator for each microorganism was evaluated ( Table 2 ). When the presence (Ct < 40) or absence (Ct ≥ 40) of microorganisms was used as the BV diagnostic criteria, A. vaginae showed the best diagnostic performance (sensitivity 91%, specificity 94%), Considering the influence of age on the microbial community of the vaginal microenvironment, we further divided the samples into different age subgroups for analysis, and the results showed that BVAB2 and A. vaginae had the best diagnostic performance in the age group of 25–34 (sensitivity 91%, specificity 100%) and 35–45 (sensitivity 75%, specificity 95%), respectively. Moreover, the ROC analysis was then performed based on the abundance of each microorganism (Ct value) and the cut-off value (CoV) with the maximum Youden’ s index was used as the BV diagnostic criteria, L. crispatus showed the best diagnosis performance (CoV: 21.24, sensitivity 97%, specificity 94%) in the modeling cohort and the age group of 25–34 (CoV: 21.55, sensitivity 100%, specificity 94%), while A. vaginae performed best in the age group of 35–44 (CoV: 29.68, sensitivity 95%, specificity 100%). These results indicated that the presence and abundance of microorganisms can be used as candidate markers for the diagnosing BV, but the influence of age should still be considered in the diagnosis.

Given our findings on microbial distribution patterns in BV patients, three lactobacilli (L. crispatus, L. jensenii, and L. iners), seven BVPs (G. vaginalis, A. vaginae, Megasphaera 2, P. bivia, M. mulieris and M. hominis) and age of the subjects were used for modeling. As shown in Figure 5 , the diagnostic AUC of the DT model was 0.736 (95% CI: 0.594–0.932) in the modeling cohort, with the sensitivity and specificity of 82.6% and 70.0%, respectively. The diagnostic AUC, sensitivity, and specificity of the SVM and RF models were all 100%. In the external validation dataset, the diagnostic AUC of the DT was 0.745 (95% CI: 0.678–0.813), with the sensitivity and specificity of 81.9% and 67.1%, respectively. The diagnostic AUC of the RF model was 0.830 (95% CI: 0.775–0.885), with the sensitivity and specificity of 97.6% and 68.4%, respectively. The SVM model had a diagnostic AUC of 0.969 (95% CI: 0.945–0.99), with the sensitivity and specificity of 90.4% and 96.1%, respectively. Overall, the SVM algorithm resulted in the highest performance in terms of AUC value in the internal validation and external validation datasets among the other two models, proving that the SVM model has more potential to diagnose BV.

Further investigation of the diagnostic consistency with intermediate scores using the SVM model is presented in Figure 5D . The heatmap results showed that 93.3% (28/30) of the IBV samples were also identified as BV positive by the new method, but the diagnosis of the other 6.7% (2/30) of the intermediate BV samples were different between the two methods.