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

Bacterial vaginosis (BV) is a common vaginal dysbiosis, and its diagnosis relies on traditional morphological assessment, which is subject to inter-observer variability in experience and subjectivity. We aimed to develop a targeted sequencing-based diagnostic classifier for BV using probe capture-based metagenomic sequencing. In this two-center case-control study, 220 women were enrolled, including an internal discovery cohort (80 BV and 59 Nugent-defined non-BV controls) and an independent external validation cohort (42 BV and 39 Nugent-defined non-BV controls). BV was defined by a Nugent score of ≥ 7, and controls by a score of 0–3. Vaginal secretion samples underwent probe capture-based metagenomic sequencing, and species-level count data were processed for compositional analysis and machine-learning modeling. Recursive feature elimination with cross-validation was used for feature selection in the internal cohort, and a random forest classifier was developed using the selected taxa. Model performance was evaluated by out-of-bag prediction internally and by direct application to the external cohort. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) and Gini importance. Reduced models were compared with the full model using DeLong’s test. BV and control samples showed significant group-level differences in both cohorts. A parsimonious five-taxon signature comprising Gardnerella vaginalis , Lactobacillus iners , Finegoldia magna , Streptococcus anginosus , and Enterobacter cloacae complex was selected for model construction. The final random forest model achieved an out-of-bag area under the receiver operating characteristic curve (AUC) of 0.939 (95% CI, 0.896–0.983) in the internal cohort and an external AUC of 0.843 (95% CI, 0.759–0.928). In the external cohort, sensitivity, specificity, and accuracy were 0.738, 0.821, and 0.778, respectively. G. vaginalis contributed most strongly to classification, but the full five-taxon model significantly outperformed reduced models based on G. vaginalis plus L. iners (AUC, 0.753; P = 0.011) or G. vaginalis alone (AUC, 0.710; P = 0.020). A compact five-taxon signature derived from probe capture-based metagenomic sequencing showed good diagnostic performance for BV and retained discrimination in an independent cohort. These findings support the potential utility of targeted sequencing-based microbial signatures for BV classification within the analytical boundaries of pathogen-oriented panels. Not applicable.