The stakes of getting this right are considerable. Bacterial vaginosis is associated with a range of adverse obstetric and gynecological outcomes, including an increased risk of acquiring sexually transmitted infections, heightened susceptibility to HIV, elevated rates of cervical cancer, and pre-term birth. Because the syndrome is defined by a shift in relative abundance between Lactobacilli and anaerobes rather than by the presence of a single pathogen, diagnosis has traditionally relied on broad criteria that may not capture the full complexity of the microbial community. Understanding precisely which species drive the transition, and at what abundances, could sharpen both diagnostics and research into how the condition develops.

To address this, the research team assembled 16S rRNA sequencing data from patients presenting with bacterial vaginosis and subjected it to a rigorous, systematic comparison of machine learning architectures and feature selection methods. Rather than committing to a single algorithm, the authors evaluated multiple classifiers to determine which combination of model and feature selection strategy could best predict BV status from the microbial profiles. This kind of head-to-head benchmarking is essential in microbiome studies, where results can vary dramatically depending on the analytical pipeline chosen, and where overfitting to small or heterogeneous datasets is a persistent concern.

The comparison produced a clear verdict. Support vector machine and random forest models, when paired with appropriate feature selection, predicted BV diagnosis with the most balanced accuracy of all the approaches tested. Balanced accuracy is a particularly informative metric in this context because it accounts for performance on both BV-positive and BV-negative cases, avoiding the trap of models that appear accurate simply because they favor the majority class. The finding aligns with a broader pattern in microbiome research, in which tree-based ensembles and margin-based classifiers often outperform more flexible deep learning architectures when the number of samples is modest relative to the number of microbial features.

With the best-performing models in hand, the researchers turned to explainable artificial intelligence methods to interrogate why the classifiers made the predictions they did. Explainable AI techniques allow researchers to move beyond a single accuracy score and identify which input features — in this case, which bacterial taxa — carry the most weight in the model’s decisions. Applied to the vaginal microbiome data, this analysis pinpointed ten species as the most important contributors to BV diagnosis: four species of Lactobacilli and six anaerobic bacteria. This shortlist represents a data-driven distillation of the microbial community down to the organisms that most reliably signal whether a patient’s ecosystem has tipped toward disease.

The identification of these ten key species has practical implications for both diagnostics and pathogenesis research. Several of the anaerobes implicated in BV have received varying degrees of scientific attention over the years, and the authors note that the determination of key bacteria can inform research into species that have previously eluded scientific focus. In other words, the machine learning pipeline did not merely confirm what was already known; it surfaced specific organisms whose role in the syndrome may have been underappreciated, providing a prioritized list of targets for future experimental and clinical investigation.

Perhaps the most visually and conceptually striking contribution of the study, however, lies in its use of Voronoi-based decision boundaries. A Voronoi diagram partitions a space into regions based on proximity to a set of reference points, and in this application the technique was adapted to visualize how a classifier divides the space of possible microbial abundance combinations into BV-positive and BV-negative territories. By plotting the relative abundances of pairs of key bacteria, the researchers produced maps in which every point corresponds to a possible microbial state, and the boundaries between regions mark the tipping points at which the model’s diagnosis flips from healthy to BV-positive.

These decision boundary plots offer something that standard classification metrics cannot: an intuitive, interpretable point of reference for how the relative abundances of key vaginal flora translate into diagnostic outcomes. A clinician or researcher looking at such a plot can immediately see, for example, how a decline in protective Lactobacilli combined with a rise in a particular anaerobe moves a sample across the boundary into BV territory. Because the boundaries are built from the highest-performing models rather than arbitrary thresholds, they reflect patterns learned directly from patient data, making them a candidate tool for guiding diagnostic interpretation and for generating hypotheses about the ecological dynamics of the syndrome.

The study also speaks to a larger methodological conversation in microbiome science. Population-specific microbial interactions — the ways in which bacterial species relate to one another within a given community — are notoriously difficult to quantify, and findings from one population do not always generalize to another. By combining careful model benchmarking, feature selection, and explainability techniques, the authors demonstrate a framework that could be applied beyond BV to other conditions defined by community-level shifts rather than single pathogens. The approach treats the microbiome as a system of interacting players whose collective balance, not any one organism alone, determines health or disease.

For patients and clinicians, the immediate promise of this work is a more precise, species-level understanding of what bacterial vaginosis actually is at the microbial level. For researchers, the ten key species and the Voronoi boundary maps provide concrete, testable targets: experiments can now be designed to probe how these organisms interact, how their abundance thresholds vary across populations, and whether manipulating them can prevent or reverse the transition to BV. As machine learning continues to move from the margins of microbiology into its methodological mainstream, studies like this one illustrate how predictive models, when paired with interpretability tools, can do more than classify — they can illuminate the hidden structure of the ecosystems within us.

Subject of Research: Machine learning analysis of 16S rRNA data to identify key bacterial interactions and diagnostic decision boundaries in bacterial vaginosis

Article Title: Machine learning and Voronoi-based decision boundaries for Bacterial vaginosis to determine population- specific microbial interactions

Article References: Celeste, C. G., Sokolik, C. C., Gachunga, W., & Parker, I. K. (2026). Machine learning and Voronoi-based decision boundaries for Bacterial vaginosis to determine population- specific microbial interactions. PLOS Computational Biology, 22(10), e1014767. https://doi.org/10.1371/journal.pcbi.1014767

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014767

Keywords: bacterial vaginosis, machine learning, microbiome, 16S rRNA, Lactobacillus, anaerobic bacteria, support vector machine, random forest, Voronoi diagram, explainable AI, feature selection, vaginal health

The rationale for targeting L. iners rather than simply eliminating all bacteria rests on a growing body of ecological evidence. Vaginal communities are classified into five major community state types, with CST I dominated by L. crispatus and CST III dominated by L. iners. Although CST III also occurs in asymptomatic women, L. iners-dominated communities have been associated with greater Candida colonization, reduced clearance of high-risk human papillomavirus infection and, in specific cohorts, poorer in vitro fertilization outcomes. Reduced L. crispatus abundance, meanwhile, has been linked to spontaneous preterm birth. Because bacterial vaginosis recurrence remains common after metronidazole therapy and post-treatment communities frequently remain or become dominated by L. iners, the post-antibiotic interval represents a strategic window in which ecological steering toward L. crispatus dominance might be achievable.

The researchers began with metabolic profiling. Using API 50 CH carbohydrate strips across five strains of each Lactobacillus species, they found that L. crispatus displayed broader, strain-dependent carbohydrate utilization, while all five L. iners strains fermented only glucose and maltose and none utilized galactose, xylose or raffinose. This raised the possibility of prebiotic selectivity: substrates that L. crispatus could exploit but L. iners could not. However, when galactose, xylose and raffinose were tested in monocultures, all five L. iners strains reached comparable viable counts, meaning the carbohydrates provided insufficient selectivity to favour L. crispatus under the nutrient-rich conditions tested. The authors report this negative result candidly, noting that alternative excipients such as alginate, sucrose and trehalose, which have shown promise in other studies, warrant further comparative evaluation.

A striking incidental finding emerged from the culture work: when defibrinated horse blood was omitted from the growth medium, none of the five L. iners strains could replicate. This observation, consistent with the species’ known cysteine dependence and predicted heme-associated functions, suggested that L. iners relies on blood-derived nutrients, including heme-associated iron. The team therefore turned to nutritional restriction as a selective strategy, selecting Vitis vinifera-derived oligomeric proanthocyanidins (OPCs), which can sequester iron, amino acids and peptides, and ethylenediaminetetraacetic acid (EDTA), a metal chelator with established antibiofilm credentials. In simulated vaginal fluid supplemented with 3% boric acid, 0.25% OPCs, 0.05% EDTA and 0.5% raffinose, L. iners viability fell by approximately 3 log10 CFU/mL over 48 hours, while most L. crispatus strains were affected far less. Subsequent compatibility screening identified 0.25% OPCs plus 0.05% EDTA as the combination least inhibitory to L. crispatus.

The decisive selectivity, however, came from fatty acids. Testing oleic, linoleic and palmitoleic acids by broth microdilution, the researchers found that oleic acid offered the widest differential between the two lactobacilli. For L. iners DSM 13335, the minimum inhibitory and bactericidal concentrations were 0.4 and 0.8 mM respectively, whereas the corresponding endpoints for three L. crispatus strains ranged from 1.6 mM to beyond 6.4 mM, the highest concentration tested. Linoleic and palmitoleic acids inhibited both species within a narrower window. The mechanistic explanation, drawn from prior work, is elegant: L. crispatus possesses the oleate hydratase OhyA and a putative fatty-acid efflux protein, FarE, which together support fatty-acid utilization and tolerance, while L. iners lacks these functions. Oleic acid thus acts as an ecological selector rather than a blunt antimicrobial.

When the full three-component mixture of 0.25% OPCs, 0.05% EDTA and 6.4 mM oleic acid was challenged against vaginal microorganisms for 24 hours, the results were pronounced. L. iners counts fell by approximately 4.8 log10 CFU/mL relative to controls, while the three L. crispatus strains showed only 0.3 to 0.5 log10 CFU/mL reductions that were not statistically significant. All four clinically relevant Candida species tested, C. albicans, C. glabrata, C. krusei and C. parapsilosis, showed significant reductions of roughly 1.4 to 2.5 log10 CFU/mL, even though individual fatty acids alone had shown no fungicidal activity at the concentrations tested, suggesting that component interactions, antiadhesive effects or antibiofilm mechanisms may underlie the antifungal activity. Cytotoxicity testing in HeLa cells showed no significant increase in lactate dehydrogenase release at dilutions of 1.56 to 12.5% of the undiluted preparation, but significant, concentration-dependent membrane damage appeared at 25 to 100%, highlighting that an overlapping active and cytocompatible exposure window has yet to be established.

The second module of the study addressed direct fungal suppression through statin chemistry. Monacolin K is chemically identical to lovastatin, and its β-hydroxy acid form, monacolin KA, directly inhibits HMG-CoA reductase in the fungal mevalonate pathway, restricting ergosterol synthesis. In overnight assays, purified monacolin KA showed the broadest antifungal activity, reducing C. albicans counts by approximately 4.3 log10 CFU/mL at 50 µg/mL, with C. albicans and C. glabrata particularly responsive and C. krusei less susceptible, a pattern consistent with known species-dependent statin responses. Monacolin K showed a narrower spectrum, active against C. albicans, C. glabrata and C. parapsilosis. Critically, both compounds preserved high L. crispatus viability at the tested concentrations, with M247, SS and DC reaching counts of 8.34 to 8.88 log10 CFU/mL at 25 to 50 µg/mL. By testing purified compounds rather than complex red-yeast-rice preparations, the authors avoided confounding by other fermentation metabolites and identified monacolin KA as the priority antifungal candidate.

Translational support came from an exploratory longitudinal observation of a 32-year-old woman undergoing assisted reproductive treatment, reported under the CARE guidelines. Her vaginal microbiota initially showed a CST I profile with L. crispatus at roughly 95% relative abundance, but symptomatic bacterial vaginosis in April 2024 shifted the community to CST IV. After intravaginal and oral metronidazole, the profile became CST III, with L. iners accounting for 99.087% of the community. Following recurrent Gardnerella dominance, an intravaginal prototype containing 0.25% OPCs and 0.05% EDTA in an extra-virgin olive-oil vehicle was introduced alongside continued oral L. crispatus M247 supplementation and a short boric acid course. Over the following months, L. crispatus rose from 0.020% to 5.695%, then 25.341%, 68.338% and finally 96.862% by April 2026, at which point the community had returned to CST I, Gardnerella was undetectable, L. iners accounted for only 0.337% and the patient was asymptomatic. No treatment-related adverse events were reported.

The authors are careful to emphasize the limitations of this clinical signal. The prototype differed from the laboratory mixture containing free oleic acid, and the concomitant treatments preclude any attribution of the microbiota transition to the prototype alone. The evidence base remains predominantly in vitro, no animal studies were performed, and only a single clinical observation is available. Short-term monocultures cannot reproduce epithelial interactions, hormonal variation, multispecies biofilms, local retention or repeated exposure, and effects on the CST IV-associated bacteria that drive bacterial vaginosis require direct evaluation. Cytocompatibility also needs refinement, with further assessment planned in non-transformed vaginal epithelial cells or reconstructed vaginal tissue, including repeated exposures and measurements of barrier integrity.

Nevertheless, the study lays out a coherent blueprint for a new generation of intravaginal formulations that work with vaginal ecology rather than against it. The bacterial-selective module could serve as an adjunctive or post-antibiotic therapy for bacterial vaginosis, targeting the frequent emergence of L. iners during recovery, and the association between CST III and reduced HPV clearance provides a rationale for future studies of whether microbiota modulation affects viral persistence. The monacolin module targets settings where Candida colonization or recurrent vulvovaginal candidiasis predominates, a condition with a substantial global burden. Ongoing work is exploring a standardized, fractionated Pistacia lentiscus berry oil rich in oleic acid as a natural alternative to purified oleic acid, potentially extending activity to Streptococcus agalactiae, and reassessing boric acid inclusion. The next steps, standardized manufacture, stability studies, polymicrobial vaginal models and controlled clinical comparisons, will determine whether this laboratory selectivity can translate into sustained L. crispatus dominance in patients.

Subject of Research: Selective modulation of the vaginal microbiota to suppress Lactobacillus iners and Candida while preserving Lactobacillus crispatus

Article Title: Vaginal formulations selectively target Lactobacillus iners and Candida spp. while preserving Lactobacillus crispatus

Article References: Di Pierro, F., Sagheddu, V., Galletti, S., Casaroli, A., Soldi, S., Bertuccioli, A., Palazzi, C. M., Senatori, R., Carugno, A., Antipov, M. O., Briko, N. I., & Zerbinati, N. (2026). Vaginal formulations selectively target Lactobacillus iners and Candida spp. while preserving Lactobacillus crispatus. International Microbiology. https://doi.org/10.1007/s10123-026-00906-0

Image Credits: AI Generated

DOI: 10.1007/s10123-026-00906-0

Keywords: vaginal microbiota, Lactobacillus crispatus, Lactobacillus iners, Candida, oleic acid, oligomeric proanthocyanidins, EDTA, monacolin KA, bacterial vaginosis, community state types, probiotics, antifungal