Five bacterial species flag bacterial vaginosis in sequencing-based diagnostic test

Bacterial vaginosis, one of the most common vaginal conditions affecting women of reproductive age, is currently diagnosed largely through microscopy-based scoring, a method that depends heavily on the experience and judgment of the observer. A new two-center case-control study now reports that a machine-learning model built on just five bacterial species, detected through a targeted metagenomic sequencing approach, can classify the condition with high accuracy in one group of women and retain useful discrimination in an independent group from another center.

The study, published in BMC Microbiology, enrolled 220 women across two centers. Participants were divided into an internal discovery cohort of 80 women with bacterial vaginosis and 59 controls, and an external validation cohort of 42 cases and 39 controls. Cases were defined by a Nugent score of 7 or higher, the standard morphological criterion, while controls had scores between 0 and 3. Vaginal secretion samples from all participants underwent probe capture-based metagenomic sequencing, a technique that uses synthetic probes to enrich DNA from targeted microbial species before sequencing, allowing precise species-level quantification of the vaginal microbiome.

Using the sequencing data, the researchers applied compositional analysis and machine-learning methods to species-level count data. Feature selection was carried out in the internal cohort with recursive feature elimination combined with cross-validation, a procedure that repeatedly trains and tests a model while progressively removing the least informative features. This process converged on a compact five-taxon signature: Gardnerella vaginalis, Lactobacillus iners, Finegoldia magna, Streptococcus anginosus, and the Enterobacter cloacae complex. A random forest classifier, an ensemble method that aggregates many decision trees to improve robustness, was then built using these five taxa.

Performance was assessed in two ways. Within the internal cohort, the model was evaluated by out-of-bag prediction, a technique in which each decision tree is tested on the samples it did not use during training. The out-of-bag area under the receiver operating characteristic curve, or AUC, reached 0.939, with a 95 percent confidence interval of 0.896 to 0.983. An AUC of this magnitude indicates that the model distinguished cases from controls with high accuracy in the population where it was developed.

The more demanding test came from the external cohort, where the model was applied directly without retraining. There, the AUC was 0.843, with a 95 percent confidence interval of 0.759 to 0.928. Sensitivity, the proportion of true cases correctly identified, was 0.738, while specificity, the proportion of controls correctly classified, was 0.821. Overall accuracy in the external cohort was 0.778. These figures show that the signature retained meaningful discrimination in an independent population, though performance was modestly lower than in the discovery setting, a pattern commonly observed when diagnostic models are transported between centers.

Interpretability analyses using SHapley Additive exPlanations, a game-theoretic method for quantifying each feature’s contribution to individual predictions, together with Gini importance, identified Gardnerella vaginalis as the strongest single contributor to classification. This species has long been associated with bacterial vaginosis, and its prominence in the model is consistent with that association. However, the study found that relying on this organism alone would sacrifice considerable accuracy.

When the researchers compared the full five-taxon model with reduced versions using DeLong’s test, a statistical method for comparing correlated receiver operating characteristic curves, the complete signature performed significantly better. A two-taxon model based on Gardnerella vaginalis plus Lactobacillus iners achieved an AUC of only 0.753, a difference from the full model that was statistically significant at P equal to 0.011. A model using Gardnerella vaginalis alone reached an AUC of 0.710, also significantly inferior at P equal to 0.020. These comparisons indicate that the less abundant organisms, including Finegoldia magna, Streptococcus anginosus, and the Enterobacter cloacae complex, carry genuine diagnostic information beyond what the dominant species provide.

The authors note that bacterial vaginosis diagnosis currently relies on traditional morphological assessment, which is subject to inter-observer variability in experience and subjectivity. By contrast, the sequencing-based approach produces quantitative, species-level molecular data that can be processed through standardized computational pipelines. The study’s design, with cases and controls defined by the Nugent criterion at both centers, means the model was trained and validated against this existing morphological standard rather than against clinical outcomes, and the findings apply to the analytical boundaries of pathogen-oriented sequencing panels.

As a case-control study, the research establishes diagnostic classification performance in women already stratified by Nugent score rather than demonstrating how the test would perform in routine screening of unselected patients. The external validation cohort provides evidence that the signature is not merely an artifact of a single center’s population, but broader prospective evaluation would be needed before any clinical deployment. The reported metrics describe model performance against the Nugent reference standard in the enrolled populations and should not be generalized beyond that scope.

The study’s conclusion is that a compact five-taxon signature derived from probe capture-based metagenomic sequencing showed good diagnostic performance for bacterial vaginosis and retained discrimination in an independent cohort. The work adds to a growing body of research applying targeted sequencing and machine learning to vaginal microbiome data, suggesting that molecular microbial signatures could complement or eventually refine existing morphology-based diagnostic criteria for this common condition.

Subject of Research: A machine-learning diagnostic model using five bacterial taxa detected by targeted metagenomic sequencing for classifying bacterial vaginosis.

Article Title: A five-taxon diagnostic model for bacterial vaginosis using probe capture-based metagenomic sequencing: a two-center case-control study

Article References: Li, Y., Wang, B., Pan, M., Liu, Z., Liu, Y., Lv, P., Zhao, J., & Zhao, Z. (2026). A five-taxon diagnostic model for bacterial vaginosis using probe capture-based metagenomic sequencing: a two-center case-control study. BMC Microbiology. https://doi.org/10.1186/s12866-026-05827-6

Image Credits: AI Generated

DOI: 10.1186/s12866-026-05827-6

Keywords: bacterial vaginosis, metagenomic sequencing, random forest, Nugent score, Gardnerella vaginalis, diagnostic classifier, vaginal microbiome, machine learning, external validation, probe capture, AUC, case-control study