Genome-scale metabolic modelling identifies vaginal microbiome members as potential probiotics

Abstract

Probiotic supplements are marketed for diverse health benefits, yet species inclusion often lacks functional rationale. Our survey of 352 over-the-counter probiotic products available in the USA revealed 36 unique microbial species. However, there is no clear link between species inclusion and the intended health benefit. Here, to address this gap, we developed HaPaPro, a collection of 1,012 genome-scale metabolic models spanning pathogenic, probiotic and host-associated bacteria, constructed from publicly available genome sequences. Flux balance analysis revealed that probiotic species fail to capture the metabolic diversity of host-associated microbes. Focusing on vaginal health, we computationally identified vaginal microbes with metabolic profiles overlapping Gardnerella vaginalis. In vitro spent media assays using 11 vaginal isolates showed variable inhibition of G. vaginalis, primarily driven by d-lactic acid production, which was also produced by non-Lactobacillus species. These findings highlight the need for function-based probiotic design and demonstrate a scalable framework integrating metabolic modelling with experimental validation.

Metabolic modelling and experimental validation reveal that current probiotics lack the functional diversity of native microbes, identifying vaginal species that inhibit Gardnerella vaginalis through the dual mechanisms of resource competition and d-lactic acid production.

Main

Over the past two decades, the therapeutic landscape of probiotics¹ has expanded exponentially (Extended Data Fig. 1), with microbial supplementation linked to a broad spectrum of improved health outcomes ranging from obesity prevention to immune modulation^2–5. However, this proliferation of potential clinical application has largely outpaced our fundamental understanding of metabolic mechanisms that govern probiotic efficacy within the host niche. While the historical use of fermented microbes provides a dietary precedent⁶, the translation to clinical therapeutics remains limited; currently, only two microbial products (VOWST and REBYOTA^7,8) have achieved US Food and Drug Administration (FDA) approval as therapeutics specifically for recurrent Clostridioides difficile infection. The vast remainder of the market is dominated by supplements targeting gut, vaginal or neurological health, which often lack the rigorous in vitro or clinical validation required to support their functional claims^9,10. This disconnect highlights a critical need to transition from empirical species selection towards a standardized functional rationale, enabling the rational design of next-generation live biotherapeutics for targeted dysbiosis.

Extended Data Fig. 1. Number of ‘Probiotics’ PubMed articles published each year since before 1950.

The number of articles published on PubMed on ‘Probiotics’ from before 1950 to present. Data was retrieved from PubMed in 2024.

Women’s health represents a primary frontier for probiotic development, with the vaginal microbiome serving as a critical niche for ecological intervention^11–14. In reproductive-age women, homeostasis is typically maintained by Lactobacillus dominance and low pH (pH <4.5), which provides a robust barrier against pathogen colonization¹⁵. Conversely, the dysbiotic state of bacterial vaginosis (BV) is characterized by a rise in pH (>4.5), an increase in taxonomic diversity and the proliferation of anaerobic pathobionts such as Gardnerella species^16–18. While antibiotics remain the clinical standard, high recurrence rates^19,20 and the rise of antimicrobial resistance underscore the limitations of these broad-spectrum approaches²¹.

Critically, recent evidence has identified healthy, asymptomatic populations with naturally diverse microbiomes that exist in a dynamic equilibrium with both Gardnerella and Lactobacillus²². This heterogeneity suggests that health may be governed by functional niche stability rather than simple species exclusion, challenging the efficacy of ‘one-size-fits-all’ treatments. There is a pressing need for a patient-specific framework that accounts for these demographic variations utilizing rationally designed probiotics to stabilize the metabolic landscape and prevent pathogen establishment across diverse host environments^16,19,20,22.

Microbial restoration through oral or vaginal probiotic administration offers a compelling strategy for re-establishing vaginal homeostasis^23,24. To contextualize the efficacy of current interventions, we surveyed 32 independent research studies treating recurrent BV with probiotics, antibiotics or combination therapies^25–55. Across these studies, there was no clear link between treatment modality and durable protection against BV recurrence. However, current clinical interventions have yielded highly variable results, with existing formulations frequently failing to provide durable protection against BV recurrence compared with standard antibiotic regimens^37,41,50. This inconsistency probably reflects the underlying ecological and demographic heterogeneity of the vaginal niche, suggesting that a ‘one-size-fits-all’ approach to strain selection is inadequate. These observations highlight a critical opportunity for the development of personalized, rationally designed biotherapeutics. By shifting focus towards the specific metabolic mechanisms required to inhibit Gardnerella colonization, development of tailored probiotic consortia that provide robust and enduring support for vaginal health across diverse patient populations is possible.

In this study, we integrate market analysis, systems modelling and in vitro experimentation to establish a functional rationale for probiotic design. We first characterize the current market landscape through an extensive survey of products from the three largest US pharmacies. To evaluate the metabolic capacity of these species, we utilize genome-scale metabolic network reconstructions (GENREs) and constraint-based reconstruction and analysis (COBRA)⁵⁶ methods to simulate the strain-specific phenotypes of 1,012 probiotic, pathogenic and host-associated bacteria. This computational framework, termed the HaPaPro collection, enables the identification of unique metabolic signatures and significant functional gaps in current probiotic formulations.

Focusing on the vaginal microbiome as a clinical case study, we then combine these in silico simulations with in vitro spent media and co-culture assays to identify native species that effectively inhibit Gardnerella vaginalis. By elucidating the shared metabolic traits and inhibitory mechanisms of these successful competitors, we provide a mechanistic blueprint for the rational development of targeted biotherapeutics capable of preventing vaginal dysbiosis.

Results

Probiotic supplement market survey

To gain a deeper understanding of the landscape of over-the-counter probiotic supplements and the diversity of bacterial species they contain, we performed a survey of all over-the-counter probiotic supplements available at the top three pharmacies in the USA⁵⁷. Across these pharmacies, there were 352 unique probiotic supplement products available for purchase, 36 unique probiotic species across these products and 70 distinct brand names.

The ten most prevalent species of bacteria found across the 352 probiotic supplements are presented in Table 1. The top two species Lactobacillus rhamnosus and L. acidophilus, both lactic acid-producing bacteria, were present in 46% of all probiotic supplements. In addition, Bacillus coagulans was present in 22% of probiotic supplements. We noted that over half of the probiotic supplements contain one probiotic species (Fig. 1a). Despite multistrain probiotics being more effective at displacing pathogens owing to synergy across strains⁵⁸, there are less than five supplements that contain 17 unique species, perhaps reflecting industrial and manufacturing constraints rather than biological efficacy.

Table 1.

Survey of over-the-counter probiotic supplements

Top ten most prevalent species	Number of probiotic supplements	Percentage of all probiotic products surveyed
L. rhamnosus	162	46%
L. acidophilus	162	46%
B. longum	103	29%
L. plantarum	101	26%
B. lactis	100	26%
L. paracasei	90	26%
B. bifidum	82	23%
B. coagulans	79	22%
L. casei	60	17%
B. breve	59	17%

Top ten most prevalent probiotic species across surveyed supplements.

Fig. 1. Survey of over-the-counter probiotic supplements.

a, Histogram of the number of probiotic supplements containing a certain number of probiotic species. b, Box-and-whisker plot of the number of species contained in a probiotic supplement by brand name. Only brand names containing three or more unique products are displayed on this plot. Brand 1: n = 43, min/max of 1/15, median of 4, Q1–Q3 of 1.5–8, whiskers of 1–15, P10/P25/P50/P75/P90 of 1/1.5/4/8/10. Brand 2: n = 16, min/max of 1/3, median of 1, Q1–Q3 of 1–1, whiskers of 1–1, P10/P25/P50/P75/P90 of 1/1/1/1/2. Brand 3: n = 17, min/max of 1/15, median of 12, Q1–Q3 of 2–14, whiskers of 1–15, P10/P25/P50/P75/P90 of 1.6/2/12/14/15. Brand 4: n = 25, min/max of 1/4, median of 1, Q1–Q3 of 1–2, whiskers of 1–2, P10/P25/P50/P75/P90 of 1/1/1/2/2. Brand 5: n = 9, min/max of 1/2, median of 2, Q1–Q3 of 1–2, whiskers of 1–2, P10/P25/P50/P75/P90 of 1/1/1/2/2. Brand 6: n = 9, min/max of 1/1, median of 1, Q1–Q3 of 1–1, whiskers of 1–1, P10/P25/P50/P75/P90 of 1/1/1/1/1. Brand 7: n = 36, min/max of 1/9, median of 3, Q1–Q3 of 1–4, whiskers of 1–8, P10/P25/P50/P75/P90 of 1/1/3/4/5. Brand 8: n = 18, min/max of 1/10, median of 1.5, Q1–Q3 of 1–8.5, whiskers of 1–10, P10/P25/P50/P75/P90 of 1/1/1.5/8.5/10. Brand 9: n = 9, min/max of 1/11, median of 9, Q1–Q3 of 2–9, whiskers of 1–11, P10/P25/P50/P75/P90 of 1.8/2/9/9/10.2. Brand 10: n = 9, min/max of 1/15, median of 4, Q1–Q3 of 2–8, whiskers of 1–15, P10/P25/P50/P75/P90 of 1.8/2/4/10/11. c, PCA plot of the species profiles of each probiotic supplement. Colours of points represent the marketed use for each probiotic supplement.

As the variation in number of unique probiotic species included in a supplement, we explored if the number of unique species was linked with brand name (Fig. 1b). Evaluation of formulation complexity by brand revealed significant inter- and intrabrand variation, with species counts per product ranging from 1 to 17. While some brands maintain consistent species numbers across their product lines, other exhibit a broad distribution of in-formulation diversity.

We identified the targeted use for a supplement along with the profile of species either present or absent in each supplement. We observe no apparent clustering in this principal component analysis (PCA) plot on the basis of marketed use for the probiotic (Fig. 1c). This result suggests that across all over-the-counter probiotic supplements, there is no specific combination of species that is consistently used to specifically support gut health, vaginal health and so on. This result highlights the gap in understanding the functional capabilities of individual probiotic species.

Metabolic modelling of diverse bacterial species

We generated GENREs representing 1,012 bacterial species using the automated GENRE construction tool, Reconstructor⁵⁹ (Fig. 2a; sequence selection outlined in the Methods) to assess the functional capabilities of probiotic, pathogenic and other host-associated bacterial species—the HaPaPro collection. To systematically categorize this functional space, we defined probiotics as species identified in our commercial market survey, pathogens as species included in the established PATHGENN database and host-associated microbes as any remaining human-associated bacterial taxa in the BV-BRC database not captured by the prior two categories. No manual curation steps were performed with this collection owing to the high quality of the draft reconstructions created with Reconstructor and the number of reconstructions in the collection. With these GENREs, we explored reaction content across probiotic, pathogenic and other host-associated species.

Fig. 2. Description of genome-scale metabolic network models of probiotic, pathogen and host-associated bacterial metabolism.

a, Workflow for generating the HaPaPro GENRE collection. b, Number of unique reactions across the GENREs of each category. c, Number of shared reactions between probiotic, pathogen and host-associated organisms. d–f, Metabolic subsystems of reactions that are unique to host-associated pathogen (d), pathogen ((e) and probiotic (f) GENREs. g, PCA of metabolic phenotypes of probiotic GENREs. Colours represent distinct taxonomic families. Triangles represent cluster centroids; ellipses are plotted 2 standard deviations from the centroid. Kruskall–Wallis test for median distances between centroids, P = 1.16 × 10⁻¹¹². Cluster standard deviations are as follows: Lactobacillus, 995.19; Streptococcaceae, 533.53; Bacillaceae, 635.82; Bifidobacteriaceae, 609.72; Akkermansiaceae, 248.95; Enterococcaceae, 488.13. h, PCA of metabolic phenotypes across probiotic, pathogen and host-associated GENREs. Triangles represent cluster centroids; ellipses are plotted 2 standard deviations from the centroid. Kruskall–Wallis test for median distances between centroids, P < 0.0001. H-A, host-associated. Schematic in a created in BioRender; Glass, E. https://biorender.com/np18ap2 (2026).

Host-associated bacterial species have the greatest number of unique reactions across GENREs (4,296 reactions), probably owing to category size and diversity of bacterial species it contains (Fig. 2b). Probiotics have the fewest unique reactions (2,457 reactions), probably owing to the homogeneity of species the probiotic category contains (Lactobacillus, Bacillus and Bifidobacteria). There is significant overlap in reaction content between the three groups, with pathogens and non-pathogens sharing the greatest number of reactions (3,588; Fig. 2c).

Across unique reaction sets, there were differences in reaction metabolic subsystems. Unique reaction sets were partitioned by group: host-associated pathogens were enriched in amino acid metabolism (Fig. 2d); host-associated bacteria are essential for the synthesis and absorption of certain amino acids in the human body^60,61. Pathogens were enriched in terpenoid and polyketide biosynthesis (Fig. 2e), which are responsible for the production of virulence factors in pathogens^62,63. Probiotics were enriched in carbohydrate metabolism (Fig. 2f), a result consistent with previous studies showing that carbohydrate metabolism in probiotic bacteria aids in digestion and promotes short-chain fatty acid synthesis^64–67. This metabolic heterogeneity confirms that host-associated, pathogen and probiotic categories occupy distinct functional niches⁶⁸.

Simulation of metabolic niche differences

We used the GENREs of pathogenic, probiotic and host-associated bacterial metabolism from the HaPaPro collection to probe each species’ genotype–phenotype relationship using COBRA. To characterize metabolic phenotypes across bacterial species in the HaPaPro collection, we performed flux balance analysis and flux space sampling⁶⁹. We observed clustering of metabolic phenotypes in probiotics by family (P < 0.001) using PCA (Fig. 2g). This functional landscape identified opportunities for rational consortium design; for instance, combining Lactobacillacea and Bifidobacteriacea species maximizes metabolic coverage. However, the close proximity of Akkermansiacea and Bifidobacteriaceae species suggests metabolic redundancy; combining them may not offer any added metabolic functional diversity compared with administering Bifidobacteriacea strains alone. Such mapping permits the intentional selection of strains that either fill distinct metabolic niches or provide ecological benefits through niche partitioning.

Gaining a deeper understanding of how pathogenic, probiotic and other host-associated bacteria’s metabolic phenotypes are similar and different is essential for intentional probiotic strain design. Currently, many pathogenic and host-associated metabolic phenotypes are not captured by probiotic species (Fig. 2h), highlighting a significant functional void in the supplement landscape. Importantly, filling this void does not mean that a probiotic should broadly mimic the metabolic range of all host-associated species, as many native microbes are associated with dysbiosis. Rather, this metabolic gap presents an opportunity for targeted niche occupation. By intentionally expanding a probiotic’s metabolic repertoire to specifically overlap with the limiting resources required by a target pathogen, we can engineer direct resource competition. We expanded on this analysis to demonstrate this concept of targeted competitive exclusion in a specific use case—women’s health and the vaginal microbiome.

Metabolic landscape of vaginal microbes

We identified 23 vaginal health supplements using 22 unique probiotic bacterial species, 68% of which were Lactobacillus sp., a dominant member of the healthy vaginal microbiome¹⁵. However, certain populations do not exhibit Lactobacillus-dominant community state types (CSTs) nor symptoms consistent with vaginal dysbiosis²², suggesting that homeostatic functions are not restricted to Lactobacillius. To systematically characterize these protective metabolic functions and identify the specific mechanisms preventing pathogen expansion, we performed a combination of computational and in vitro analyses.

Through dimensionality reduction and visualization of the metabolic phenotypes, we observe that there are significant functional differences in metabolism between species used as vaginal probiotics in commercially available supplements and the native flora of the vaginal microbiome, with non-probiotic species covering a wider range of metabolic functions (non-probiotic standard deviation, 897.7; vaginal probiotic standard deviation, 863.35; Kruskal–Wallis test for median distances between centroids, P = 2.2 × 10⁻⁷) (Fig. 3a). Consistent with our global analysis (Fig. 2h), endogenous species occupy a significantly broader metabolic space, possessing functional traits not captured by commercial probiotic species. This result presents an opportunity to leverage the expanded metabolic repertoire of native vaginal microbes to identify protective functional traits missing in current formulations.

Fig. 3. Functional analysis of vaginal probiotic species and other vaginal bacteria.

a, PCA plot of metabolic phenotypes of vaginal bacteria and probiotic species. Triangles represent cluster centroids; ellipses are plotted 2 standard deviations from the centroid. Kruskal–Wallis test for median distances between centroids, P = 2.2 × 10⁻⁷. Vaginal bacteria standard deviation, 897.7; vaginal probiotic standard deviation, 863.35. b, Hierarchical clustering of metabolic phenotypes across vaginal probiotic species, two G. vaginalis strains and other culturable vaginal bacterial species. Metabolic phenotype vectors are calculated as the per-reaction median value across 500 flux samples. DSMZ, German Collection of Microorganisms and Cell Cultures.

Prevention of Gardnerella spp. growth, frequently associated with BV, is one specific target to consider when designing next-generation probiotics for the vaginal microbiome⁷⁰. To identify potential competitors of Gardnerella spp., we identified the median metabolic signature of two G. vaginalis strains against native vaginal isolates and probiotic species. Importantly, ‘vaginal probiotics’ are species that were included in commercially available probiotic supplements that claim vaginal health benefits. Hierarchical clustering was used to map functional proximity across groups focusing on species with poorly characterized roles in the vaginal niche. For example, bifidobacteria are widely used as probiotics but their efficacy as vaginal probiotics has not been well studied (CST IV-C3, Bifidobacterium dominated, uncertain clinical risk). This analysis allowed us to observe the similarities and differences in metabolic signatures across strains (Fig. 3b and Extended Data Fig. 2).

Extended Data Fig. 2. In silico metabolic phenotype comparison between vaginal strains and G. vaginalis strains.

Hierarchical clustering of metabolic phenotypes across two G. vaginalis strains, and culturable vaginal commensal species used in the in vitro spent media experiment, rows are colored by G. vaginalis growth dynamic category. Metabolic phenotype vectors are calculated as the per-reaction median value across 500 flux samples.

G. vaginalis strains exhibited high functional similarity to bifidobacteria species currently used in vaginal probiotics yet diverged significantly from the Lactobacillus and Bacillus vaginal probiotic species. This lack of metabolic overlap between the pathobiont and the traditional protective species implies that overlap in function may not be the only driver of resource competition or Lactobacillus-mediated competitive exclusion. Lactobacilli are essential for maintaining vaginal health through the secretion of lactic acid to maintain a low-pH environment. Our result could suggest that when considering the rational design of probiotics, preventing G. vaginalis colonization may be a more nuanced problem that considers environmental factors in addition to competitive exclusion. To resolve this discrepancy and identify the role of environmental factors, we performed in vitro spent media assays to model competitive interactions between G. vaginalis and other vaginal microbiome isolates.

Inhibition of G. vaginalis by vaginal isolates

Vaginal microbiome isolates that compete with G. vaginalis could be key candidates for a probiotic consortium to prevent G. vaginalis colonization, if they themselves are not associated with high-clinical-risk CSTs. To further understand differences in metabolic functionality across vaginal microbes and identify which isolates effectively inhibit G. vaginalis growth, we designed an in vitro spent media assay (Methods) as a proxy for assessing competition between vaginal microbes and G. vaginalis. This assay allows us to determine if a stable population of each vaginal isolate produces metabolic byproducts that would inhibit the growth of G. vaginalis. We generated 11 spent media conditions, one for each of the culturable vaginal isolates shown in Fig. 3b, that were able to be cultured in vitro. Then, an overnight culture of G. vaginalis was inoculated into the cell-free supernatants of each isolate to assess the inhibitory potential of secreted metabolic byproducts (Fig. 4a; Methods).

Fig. 4. In vitro spent media assay and lactic acid quantification.

a, Growth curves of G. vaginalis 14018 in 12 media conditions (11 spent media and PGY-mod control). Three growth dynamics are specified to the right of the growth curves. N, non-inhibitory (n = 4); M, moderate (n = 4); I, inhibitory (n = 4). b, Quantification of the AUC for each of the three growth dynamic categories. Pairwise t-tests, one tailed: N versus M, P = 5.68 × 10⁻⁴; N versus I, P = 4.19 × 10⁻⁷; M versus I, P = 6.76 × 10⁻⁶. c, pH of the spent media in each growth dynamic category. Pairwise t-tests, one tailed: I versus M, P = 6.46 × 10⁻⁴; I versus N, P = 6.32 × 10⁻⁴; M versus N, P = 0.4039. d, l-Lactic acid concentration in the spent media of each growth dynamic category. Pairwise t-tests, one tailed: I versus M, P = 0.133; I versus N, P = 0.0532; M versus N, P = 0.1605. e, d-Lactic acid concentration in the spent media of each growth dynamic category. Pairwise t-test, one tailed: I versus M, P = 0.0292; I versus N, P = 0.02418; M versus N, P = 0.1632. f, Relationship between d-lactic acid and G. vaginalis AUC. R² = 0.7026, exponential fit equation, y = 122.504 × 10^−0.564x. ***P < 0.001, ** P < 0.01, *P < 0.1. All data are presented as mean values ± s.d. n.s., not significant.

We observed some variation in the G. vaginalis growth profiles on the 12 media conditions (Fig. 4a). More specifically, we see three distinct response groups; non-inhibitory (N), moderate (M) and inhibitory (I) (P < 0.01 for all pairwise comparisons; Fig. 4b). G. vaginalis grown on non-inhibitory media (Mobiliuncus curtisii, Veillonellaceae bacterium and Ezakiella massiliensis spent media) had the same growth dynamics as G. vaginalis grown on the Peptone–Glucose–Yeast (modified) (PGY-mod) media control. In the moderate group, we see a significant change in the area under the growth curve (AUC) of G. vaginalis grown in Peptoniphilus vaginalis, Anaerococcus marseille, Criibacterium bergeronii and Aerococcus christensenii spent media compared with the uninhibited group (P < 0.01). In the inhibited group, we see a statistically significant difference in the AUC of G. vaginalis grown in Fannyhessea vaginae spent media, Lactobacillus jensenii, Anaerococcus lactolyticus and Anaerococcus tetradius when compared with the uninhibited (P < 0.001) and moderate (P < 0.001) conditions. This potent inhibition suggests diverse mechanisms of suppression, prompting an investigation into whether resource competition or environmental modifiers, such as media acidification, drove the observed exclusion.

d-Lactic acid drives G. vaginalis inhibition

It is well known that the Lactobacillus dominance in the vaginal microbiome is responsible for protecting against pathogen invasion by maintaining an acidic environment through lactic acid production¹⁵. We observed significant differences in pH across the spent media of the three growth dynamics groups (N, M and I) (Fig. 4c). The spent media that inhibited G. vaginalis growth had a significantly lower pH than the spent media of the moderate and uninhibited groups (P < 0.01). However, there was no significant difference between the pH in the uninhibited and moderate groups (P > 0.8). This result suggests that the pH of the spent media plays a major role in inhibiting G. vaginalis growth. Consequently, we quantified D- and L-lactic acid concentration to determine the metabolic basis of this acidification (Methods).

The mmol per litre concentrations of l-lactic acid and d-lactic acid determined by the assays are presented in Table 2. We observed no significant difference in the concentration of l-lactic acid between the inhibitory, non-inhibitory and moderate spent media (Fig. 4d). We observed significant differences in d-lactic acid concentration in the inhibitory spent media condition compared with the moderate and non-inhibitory spent media conditions (Fig. 4e). In addition, we observed a strong inverse correlation between d-lactic acid concentration and the G. vaginalis AUC (Fig. 4f). These results suggest that d-lactic acid production is a key metabolic driver for G. vaginalis growth inhibition in the spent media of L. jensenii, F. vaginae, A. tetradius and A. lactolyticus.

Table 2.

Concentration of l-lactic acid, d-lactic acid, G. vaginalis AUC and growth dynamic category of each spent media condition

Media condition	G. vaginalis growth dynamic category	G. vaginalis AUC	l-Lactic acid concentration (mmol l−1)	d-Lactic acid concentration (mmol l−1)
PGY-mod	Uninhibitory	6.812	7.176	1.678
M. curtisii	Uninhibitory	7.074	7.528	1.483
E. massiliensis	Uninhibitory	7.116	7.566	1.64
V. bacterium	Uninhibitory	7.055	1.453	1.052
A. christensenii	Moderate	5.696	11.79	17.643
C. bergeronii	Moderate	5.393	7.159	1.793
A. marseille	Moderate	5.886	7.128	1.476
P. vaginalis	Moderate	6.315	6.144	3.156
F. vaginae	Inhibitory	2.442	7.025	53.472
A. tetradius	Inhibitory	1.667	15.138	21.192
L. jensenii	Inhibitory	1.315	7.689	78.059a
A. lactolyticus	Inhibitory	1.999	16.263	25.016

^aThis concentration was beyond the standard curve and was extrapolated.

These results are consistent with the general knowledge that Lactobacillus species are typically present in a healthy vagina owing to their ability to produce lactic acid and lower vaginal pH. The high d-lactic acid production by F. vaginae, A. tetradius and A. lactolyticus demonstrates that pathogen-inhibitory acidification is not exclusive to the Lactobacillus genus. These findings suggest that non-canonical d-lactic acid producers may offer functional equivalence in maintaining a protective vaginal environment. To determine if this inhibitory potential translates to direct competitive exclusion in a shared environment, we performed co-culture assays.

Characterizing competitive interactions

We performed co-culture assays with G. vaginalis and three high d-lactate-producing microbes L. jensenii, A. lactolyticus and A. tetradius. While the data presented in Extended Data Fig. 5 are preliminary given n = 1, our results show that the three high d-lactate-producing microbes effectively suppressed the growth of G. vaginalis in co-culture compared with monoculture controls, confirming that its metabolic advantage drives competitive exclusion. Similarly, both A. lactolyticus and A. tetradius exhibited strong suppressive effects, particularly at equal or higher starting abundances. These results demonstrate that d-lactate production, regardless of whether it originates from a health-associated Lactobacillus or a dysbiosis-associated Anaerococcus, is a potent driver of G. vaginalis inhibition in this system. In addition, we assessed the biofilm-forming capacity of the vaginal isolates. Furthermore, while G. vaginalis exhibited robust biofilm formation relative to F. vaginae, A. tetradius and A. lactolyticus (Extended Data Table 1), the sustained inhibition observed in portioned co-culture confirms this suppression is mediated by secreted soluble factors rather than physical displacement.

Extended Data Fig. 5. G. vaginalis grown in co-culture with high D-lactate producing bacteria.

Growth curve results of G. vaginalis grown in co-culture with L. jensenii (A), A. lactolyticus (B), and A. tetradius (C) over a 16 hour time course. All growth curve data presented here are from one replicate (n = 1). 1:10 ratio represents a ratio of OD 0.005:0.05 where 0.005 is a 1:10 dilution from a culture with an OD of 0.05.

Extended Data Table 1.

Biofilm quantification of G. vaginalis and high D-lactate producers grown in co-culture

Ratio of G. vaginalis : L. jensenii	Ratio of Biomass/Biofilm
	G. vaginalis	L. jensenii
01:01	0.557	0.282
01:10	0.321	0.238
10:01	1.214	0.295
	Ratio of Biomass/Biofilm
Ratio of G. vaginalis : A. lactolyticus	G. vaginalis	A. lactolyticus
01:01	0.948	0.153
01:10	0.368	0.521
10:01	0.665	0.206
	Ratio of Biomass/Biofilm
Ration of G. vaginalis : A. tetradius	G. vaginalis	A. tetradius
01:01	0.516	0.273
01:10	0.563	0.081
10:01	0.635	0.231
	Ratio of Biomass/Biofilm
1:1 G. vaginalis : G. vaginalis	A	B
1:1 L. jensenii : L. jensenii	0.516	0.273
1:1 A. lactolyticus : A. lactolyticus	0.563	0.081
1:1 A. tetradius : A. tetradius	0.635	0.231

Ratio of biomass/biofilm for G. vaginalis and L. jensenii (1), G. vaginalis and A. lactolyticus (2), and G. vaginalis and A. tetradius (3) grown in co-culture in three different ratios 1:1, 1:10, and 10:1. 4) Ratio of biomass/biofilm for G. vaginalis, L. jensenii, A. lactolyticus, and A. tetradius grown in co-culture with itself at a 1:1 ratio.

To mechanistically elucidate these in vitro interactions, we utilized microbial community-scale modelling simulations (MICOM)⁷¹ to simulate pairwise in silico co-cultures of G. vaginalis with high d-lactate producers: L. jensenii, A. lactolyticus, A. tetradius and F. vaginae. The models predicted significant competitive pressure on G. vaginalis driven purely by resource competition. When modelled at a 1:1 ratio, L. jensenii reduced the predicted growth rate of G. vaginalis by ~87% relative to the monoculture, a magnitude of suppression that is comparable to that of the BV-associated Anaerococcus species (~89%) (Extended Data Table 2). Increasing the partner’s relative abundance (1:10 ratio of G. vaginalis to partner) yielded a near-total suppression (>98%) of pathogen growth across all pairs, aligning with our in vitro co-culture dynamics (Extended Data Fig. 5). Notably, while the model predicted robust suppression when the pathogen was dominant (10:1), our experimental data revealed incomplete inhibition at this ratio. This discrepancy demonstrates that while resource competition is a fundamental baseline driver of suppression, successful in vivo exclusion probably depends on reaching a critical threshold of accumulated inhibitory metabolites, such as d-lactate.

Extended Data Table 2.

MICOM Simulation Results

Pair	Partner Name	Abundance Ratio	G. vaginalis Growth	Partner Growth	G. vaginalis Growth Alone	G. vaginalis growth supression (%)
G. vaginalis vs L. jensenii	Lactobacillus jensenii	1.0:1.0	46.2232683	46.2232843	358.422939	87.103708
G. vaginalis vs L. jensenii	Lactobacillus jensenii	1.0:10.0	6.1470105	61.4621445	358.422939	98.284984
G. vaginalis vs L. jensenii	Lactobacillus jensenii	10.0:1.0	39.2219324	3.92218262	358.422939	89.057081
G. vaginalis vs A. lactolyticus	Anaerococcus lactolyticus	1.0:1.0	36.8846366	36.8851553	358.422939	89.709186
G. vaginalis vs A. lactolyticus	Anaerococcus lactolyticus	1.0:10.0	4.29884095	42.9718659	358.422939	98.800623
G. vaginalis vs A. lactolyticus	Anaerococcus lactolyticus	10.0:1.0	37.3729003	3.73729167	358.422939	89.572961
G. vaginalis vs A. tetradius	Anaerococcus tetradius	1.0:1.0	40.6771234	40.6771291	358.422939	88.651083
G. vaginalis vs A. tetradius	Anaerococcus tetradius	1.0:10.0	5.04783607	50.478537	358.422939	98.591654
G. vaginalis vs A. tetradius	Anaerococcus tetradius	10.0:1.0	38.1235632	3.81235427	358.422939	89.363526
G. vaginalis vs F. vaginae	Fannyhessea vaginae	1.0:1.0	46.2231304	46.2232453	358.422939	87.103747
G. vaginalis vs F. vaginae	Fannyhessea vaginae	1.0:10.0	6.14738535	61.4621725	358.422939	98.28488
G. vaginalis vs F. vaginae	Fannyhessea vaginae	10.0:1.0	39.2219701	3.92219827	358.422939	89.05707

G. vaginalis simulation with high D-lactate producing bacteria in three different abundance ratios 1:1, 10:1, 1:10. Growth supression is reported as percentage of total G. vaginalis growth with partner compared to without partner.

Discussion

Despite the proliferation of over-the-counter probiotic products^2–5, a functional rationale for species selection is often absent. Our survey of 352 products identified only 36 unique species, most of which are marketed without a clear mechanistic link to their intended health benefit. To address this knowledge gap, we developed the HaPaPro collection utilizing 1,012 genome-scale metabolic models to evaluate the functional diversity of current probiotics relative to native microbes. Our analysis revealed that current probiotic species capture only a small subset of the metabolic functions found across pathogenic and host-associated bacteria. This functional gap suggests a significant opportunity to move beyond traditional strain selection towards a design strategy that expands the metabolic niches occupied by probiotic communities.

In our vaginal case study, we identified specific metabolic functions required to inhibit the pathobiont G. vaginalis. While the protective role of lactic acid is well documented^72–75, our work distinguishes between the isomers, clarifying that d-lactic acid is a key driver of G. vaginalis-growth inhibition. Through spent media assays, we demonstrated that not only Lactobacillius species but also certain non-lactic acid bacteria, including A. lactolyticus and A. tetradius, can produce sufficient d-lactate to suppress pathogen growth. This finding shifts the focus from simple acidification to the specific accumulation of therapeutic metabolic byproducts.

Integrating metabolic modelling with in vitro co-cultures (Extended Data Fig. 5) revealed that pathogen suppression is driven by dual mechanisms: resource competition and metabolic inhibition, consistent with previous work⁷⁶. While our experimental assays highlighted d-lactic acid, MICOM simulations demonstrated that nutrient sequestration is a concurrent force (Extended Data Table 2). Even in the absence of d-lactate production, our models predicted that L. jensenii could suppress G. vaginalis growth by over 85% simply by outcompeting the pathogen for limiting substrates. This result suggests that the most effective probiotic candidates are those that provide protection through both acidification and efficient niche occupation, preventing pathogen establishment even when metabolic byproduct concentrations are transiently low.

Furthermore, our study highlights a distinction between metabolic capability and colonization potential⁷⁷. While we identified strong inhibitors, many of these candidates lack the robust biofilm-forming capacity of G. vaginalis. As BV is characterized by resilient polymicrobial biofilms^78,79, the therapeutic potential of these candidates probably relies on the secretion of diffusible metabolites, such as d-lactic acid, that can penetrate the biofilm matrix without requiring the probiotic to physically integrate into the established structure. This distinction is critical for designing probiotics that can effectively displace pathobionts from protected niches.

Clinical context remains essential for translating these mechanistic findings. Although Anaerococcus spp. and F. vaginae exhibited potent inhibition in vitro, they are clinically associated with CST IV and BV⁸⁰. Their inhibitory capacity probably reflects competitive dynamics with a dysbiotic niche rather than a health-promoting mechanism^81,82. Conversely, L. jensenii is a hallmark of the healthy CST V. The convergence of our mechanistic data with clinical health associations identifies L. jensenii as the primary actionable candidate for vaginal health interventions.

Looking forwards, this systems biology framework provides a scalable foundation for predictive, rationally designed biotherapeutics. While d-lactate drove inhibition in this baseline model, adapting our in silico pipeline to simulate distinct physiological constraints, such as those dictated by specific demographic or clinical parameters, will allow us to uncover alternative, context-dependent metabolic targets. For instance, our analysis of health records revealed that 22% of vaginitis cases (International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM): N76.0 and N76.1) occur in postmenopausal women, a population with distinct metabolic constraints driven by lower glycogen and oestrogen compared with premenopausal women⁸⁰ (Extended Data Table 4; Methods). Ultimately, the approach presented here informs the rational selection and metabolic engineering of targeted consortia, moving the field beyond empirical strain supplementation towards functional, evidence-based ecological interventions for diverse patient populations.

Extended Data Table 4.

Patients presenting with vaginitis associated symptoms in the UVA health system

	All Vaginitis Patients	Vaginitis 55+ Years Old
Total patients	35040	7580
Minimum Age	0	55
Maximum Age	90	90
Mean Age	41	67
Standard Deviation	18	10
Ethnicity
Not Hispanic or Latino	83.82%	93.40%
Hispanic or Latino	14.04%	4.88%
Unknown Ethnicity	2.17%	1.85%
Race
Whtie	60.73%	71.11%
Black or African American	22.60%	21.24%
Other Race	12.70%	4.88%
Asian	2.23%	1.85%
Unknown Race	1.57%	1.19%
American Indian or Alaska Native	0.17%	0.13%
Native Hawaiian or Other Pacific Islander	0.09%	0.13%

Methods

Survey of commercial probiotics

We performed a survey of probiotics available for purchase on the CVS, Walgreens and Walmart websites, the top three pharmacies in the USA by number of locations (9,554, 9,398 and 6,860 stores respectively)⁵⁷. In this survey, we collected the name of each probiotic supplement product, specified uses (for example, gut, vaginal or urinary) and all probiotic strains included in each supplement. In addition, we determined the total number of strains and the total number of species present in each probiotic supplement. We then created a binary species presence dataframe (row: probiotic supplement product, column: probiotic species, 1: species present, 0: species absent; Extended Data Table 3). We then used this dataframe for dimensionality reduction and visualization with PCA (sklearn (version 1.7), matplotlib (version 3.10.7), numpy (version 2.3.5) and pandas (version 2.3.3)) to observe clustering patterns⁸³.

Extended Data Table 3.

List of probiotic genome sequences used to create the probiotic GENREs

Name	BV-BRC Accession Number
Pediococcus acidilactici strain PMC65	1254.353
Streptococcus thermophilus strain STH_CIRM_65	1308.1128
Enterococcus faecium SRR24	1352.14893
Lactococcus lactis strain LAC460	1358.1413
Lactobacillus brevis strain NPS-QW-145	1580.66
Lactobacillus helveticus SCB643	1587.750
Lactiplantibacillus plantarum BGI-N6	1590.4735
Lactobacillus gasseri strain BIO6369	1596.253
Lactobacillus reuteri	1598.864
Limosilactobacillus reuteri strain PNG008_24h	1598.1606
Lactobacillus fermentum strain HFD1	1613.511
Ligilactobacillus salivarius LPM01	1624.699
Bifidobacterium adolescentis strain 1–11	1680.104
Bifidobacterium longum subsp. Infantis BB-02	1682.258
Bifidobacterium breve strain BR3	1685.10
Bifidobacterium animalis A1-hRTP31088sv1_210526	28025.291
Lactobacillus johnsonii strain GHZ10a	33959.508
Lactobacillus rhamnosus strain 1.0320	47715.587
Lacticaseibacillus rhamnosus TCI366	47715.1833
Lactobacillus crispatus strain DC21.1	47770.632
Bacillus clausii strain 7520-2	79880.20
Lactobacillus jensenii strain SNUV360	109790.33
Lactobacillus casei subsp. Casei ATCC 393	219334.4
Bacillus subtilis subsp. Subtilis str 168	224308.43
Akkermansia muciniphila strain JCM 30893	239935.2189
Pediococcus pentosaceus ATCC 25745	278197.12
Bifidobacterium longum subsp longum JCM 1217	565042.3
Lactobacillus rhamnosus GG ATCC 53103	568703.30
Lactobacillus delbrueckii subsp. Bulgaricus PB2003/044-T3-4	784613.4
Bifidobacterium animalis subsp. Lactis BLC1	1075106.4
Bacillus coagulans DSM 1 = ATCC 7050	1121088.10
Lactobacillus acidophilus La-14	1314884.3
Lactococcus plantarum NBRC 100936	1348632.3
Bacillus infantis NRRL B-14911	1367477.3

Metabolic network reconstruction of pathogenic, probiotic and other host-associated species

We generated 1,007 GENREs of pathogenic, probiotic and other host-associated bacterial species. To do this, we began by selecting genome sequences from the BV-BRC (version 3.52.11)⁸⁴. We selected one genome sequence for each of the 35 bacterial probiotic species identified in our survey of commercially available probiotics from the BV-BRC (Extended Data Table 4). The criteria for selecting probiotic genome sequences were as follows: (1) select the representative or reference sequence of a given species if it exits; and (2) if there is no representative or reference sequence, randomly select a sequence that is considered good quality and complete. Of the 35 probiotic sequences selected, 26 were considered reference or representative sequences and 9 sequences were considered good and complete without the reference or representative designation. We selected pathogen sequences in a similar manner. Using the previously published database of metabolic network models as a guide, we selected all reference and representative sequences of the species included in the PATHGENN database⁶⁸, resulting in 197 pathogenic sequences selected. Finally, we considered other host-associated bacteria to be any species in the BV-BRC database that was not considered pathogenic or probiotic. We selected all reference and representative host-associated sequences that met these criteria, resulting in 775 host-associated sequences.

After selecting the sequences, we generated an annotated protein sequence through an automated pipeline. We used this annotated protein sequence as an input to Reconstructor (version 1.1.2), a tool used for automated GENRE creation⁵⁹. We used Reconstructor to generate GENREs of all 1,007 sequences. All GENREs were created in the context of the same rich media.

Reaction and subsystem analysis

We identified reactions that were unique to pathogen, probiotic and other host-associated bacterial species through model analysis and visualized these data using an upset plot (Fig. 2b,c). We then took the list of unique metabolic reactions from each group (pathogenic, probiotic and other host-associated bacteria) and identified the metabolic subsystem to which they belong. We did this by querying the Kyoto Encyclopedia of Genes and Genomes (KEGG; version 115.0) Application Programming Interface (API) to identify the general subsystem to which each of these unique reactions belongs⁸⁵. We then displayed the number of unique reactions in each group that were a part of each metabolic subsystem.

Flux sampling and metabolic phenotype analysis

We simulated all reconstructions on a rich in silico medium containing the following compounds with corresponding modelseed compound IDs: cpd00001 (water), cpd00035 (l-alanine), cpd00041(l-aspartate), cpd00023 (l-glutamate), cpd00119 (l-histidine), cpd00107 (l-leucine), cpd00060 (l-methionine), cpd00161 (l-threonine), cpd00069 (l-tyrosine), cpd00084 (l-cysteine), cpd00033 (glycine), cpd00322 (l-isoleucine acid), cpd00066 (l-phenylalanine), cpd00054 (l-serine), cpd00065 (l-tryptophan), cpd00156 (l-valine), cpd00220 (riboflavin), cpd00644 (pantothenate), cpd00393 (folate), cpd00133 (nicotinamide), cpd00263 (pyridoxal), cpd00104 (biotin), cpd00149 (cobalt), cpd00971 (sodium), cpd00099 (chloride), cpd00205 (potassium), cpd00009 (phosphate), cpd00063 (calcium), cpd00254 (magnesium), cpd10515 (iron), cpd00030 (manganese), cpd00242 (bicarbonate), cpd00226 (hypoxanthine), cpd01242 (thyminose), cpd00307 (cytosine), cpd00092 (uracil), cpd00117 (d-alanine), cpd00067 (hydrogen), cpd00567 (d-proline), cpd00132 (l-asparagine), cpd00210 (taurine), cpd00320 (d-aspartate), cpd03279 (deoxyinosine), cpd00246 (inosine), cpd00311 (guanosine), cpd00367 (cytidine), cpd00277 (deoxyguanosine), cpd00182 (adenosine), cpd00654 (deoxycytidine), cpd00412 (deoxyuridine), cpd00438 (deoxyadenosine), cpd00274 (citrulline), cpd00186 (d-glutamate), cpd00637 (d-methionine), cpd00105 (d-ribose), cpd00305 (thiamin), cpd00309 (xanthine), cpd00098 (choline), cpd00207 (guanine), cpd00082 (d-fructose) and cpd00129 (l-proline).

We used the Gapsplit algorithm for sampling flux distributions⁸⁶. We generated 500 flux distributions for each GENRE to most accurately capture the range of all metabolic functions (metabolic phenotypes). When comparing flux sampling outputs between taxa, we used the union of all reactions across reconstructions; missing reaction fluxes were imputed with zeros for comparison. To visualize similarities and differences in flux distributions across non-pathogen, pathogen and probiotic species, we used PCA through the sklearn (version 1.7) Python (version 3.9.12) package⁸³. We reduced each flux distribution to a two-dimensional space (two principal components) to plot. We plotted host-associated, pathogenic and probiotic (Fig. 2d–f) bacterial species to best visualize the metabolic phenotype coverage of each group. We used the same method of flux sampling and dimensionality reduction in Figs. 2g,h and 3a.

Heat map and hierarchical clustering of metabolic flux through G. vaginalis, other vaginal microbiome species and vaginal probiotic bacterial species

We generated 500 flux distributions per model of interest using Gapsplit (version 1.1.1)⁸⁶. Then, we generated a median flux vector across all 500 samples per each model. We then combined all median flux vectors across models of interest into one dataframe, removed low-variance reactions (variance threshold of 0.1) and dropped highly correlated reactions (correlation threshold of 0.9). After these preprocessing steps, we generated a cluster map using seaborn (version 0.13.2)⁸⁷, clustering on both rows (species-specific models) and columns (reactions) using the Canberra distance. This method was applied to the heat maps in Fig. 3b and Extended Data Fig. 2.

In vitro spent media assays

We obtained 16 vaginal microbiome isolates from the DSMZ: L. jensenii strain DSM (German Collection of Microorganisms) 20557 (DSM 20557), A. christensenii strain CCUG28831 (DSM 15819), C. bergeronii CCRI-22567 (C. bergeronii CCRI-22567), Ezakiella sp. Marseille-P2951 (DSM 103122), Peptoniphilus sp. KHD2 strain kHD2 (DSM 101742), Anaerococcus sp. Marseille-P2765 (DSM 103343), Corynebacterium fournierii strain Marseille-P2948 (DSM 103271), Corynebacterium accolens DSM 44278 (DSM 44278), Mobiluncus curtsii subsp. Holmesii American Type Culture Collection (ATCC) 35242 (DSM 21655), Veillonellaceae bacterium KA00182 (DSM 111202), Corynebacterium aurimucosum ATCC 700975 (DSM 44827), A. tetradius strain DSM 2951 (DSM 2951), Megasphaera sp. UPII 199-6 (DSM 111201), A. lactolyticus DSM 7456 (DSM 7456), Dermabacter vaginalis strain AD1-86 (DSM 100050) and Atopobium vaginae strain DSM 15829 (DSM 15829). None of these isolates are found on the list of known misidentified cell lines maintained by the International Cell Line Authenication Committee. We selected these isolates from the larger collection of identified vaginal microbes for several reasons: (1) the isolates were easily obtainable from one reputable source and (2) we selected strains that were specifically isolated from the human vagina. There were strains of some isolates that were available but were not isolated from the vagina; we did not select these.

Ultimately, we were able to successfully grow 11 of these species robustly in anaerobic conditions. After successfully growing these bacterial species in their specified media, we determined a media condition that would successfully grow these 11 species as well as G. vaginalis. We determined the media that was the most successful at robust growth was PGY media (DSMZ) modified with HEPES (6.3 ml) and 10% fetal bovine serum (FBS). We also tried unmodified PGY, NYCIII and NYCIII media modified with vitamin K (100 µl) and hemin (5 ml) (Extended Data Fig. 3).

Extended Data Fig. 3. Comparison of culturable vaginal commensal species growth in four media conditions.

Growth over 48 and 72 hour periods in PGY, PGY (Modified), NYCIII, and NYCIII (Modified). N = 1.

After determining a universal media condition (PGY-mod) that would lead to maximal growth across isolates, we next needed to collect spent media. We inoculated each isolate into 10 ml PGY-mod. For each isolate, we repeated this process five times. Each 10-ml aliquot of spent media was filtered sterilized with a 20-μm filter and was pooled to generate 50 ml spent media per isolate. Spent media were not normalized by culture density.

Then, to determine if G. vaginalis would grow on the spent media of each isolate, we performed a spent media assay. This assay involved inoculating G. vaginalis into 5 ml PGY-mod and allowing it to grow for 24 h. After 24 h, the culture was spun down at 6,500 rpm for 5 min, media were aspirated and culture was resuspended in fresh PGY-mod media. Then, G. vaginalis was inoculated into the spent media in a 12-well plate at an optical density (OD) of 0.1 and allowed to grow until stationary phase with continuous growth monitoring using a Cerrillo stratus plate reader. We repeated this assay with two different strains of G. vaginalis, GV14018 (results shown in Fig. 4a) and JCP7672 (Extended Data Fig. 4), to ensure the trends we observed were consistent across strains. Notably, these two G. vaginalis isolates are from the same clade. We observed inhibition of G. vaginalis JCP767 in A. tetradius, A. lactolyticus, L. jensenii and F. vaginae spent media as well. We observed growth of JCP7672 in the moderate and non-inhibitory spent media that were mostly consistent with GV14018, but there was some variation (Extended Data Fig. 4).

Extended Data Fig. 4. Growth of G. vaginalis JCP7672.

This alternative G. vaginalis strain was grown on 11 spent media conditions and PGY (modified) control. N = 3 per condition.

AUC calculations and identifying differences in growth dynamics

We calculated the AUC in Fig. 4b using the simps integration method from the scipy package (version 1.16.0)⁸⁸. We used the mean value across replicates for this growth-curve calculation. To determine the statistical difference between the three observed groups of growth curves, we used a one-way analysis of variance test, with subsequent pairwise t-tests between groups, with a significance value of 0.05 (n = 3 per condition).

Spent media pH measurements and differences between groups

We used a pH metre (Accumet basic AB15) to determine the pH of each media condition used in the spent media assay, with accuracy to 0.01. Then, to determine statistically significant differences in pH between the three growth dynamics groups, we used pairwise Welch’s t-tests, with a significance value of 0.05.

l-Lactic acid and d-lactic acid assays

We quantified concentrations of d-lactic acid and l-lactic acid using the Novus Biologicals d-Lactic Acid/Lactate Assay Kit (Colorimetric) NBP3-25788 and l-Lactic Acid Assay Kit (Colorimetric) NBP3-25875, respectively (n = 2 per spent media condition). After completing the specified kit protocol, the OD was read at 530 nm using a TECAN plate reader. The assay was rerun (standards and samples) at dilution for samples that read above the standard curve range. d-Lactic acid and l-lactic acid concentrations were computed according to the standard curve.

Co-culture assay

Cultures that produced the most lactic acid were grown under anaerobic conditions at 37 °C in modified PGY (described above), absorbance at 600 nm (A₆₀₀) was measured and cultures were adjusted to a starting density of 0.05–0.005 in fresh modified PGY. Co-culture conditions included strains at a 1:1 or 10:1 ratio, where each side of a Duet co-culture plate (Cerillo) contained 450 µl, with one strain grown on each side. Plates were sealed using a Breathe-Easy membrane (USA Scientific) and growth was monitored at A₆₀₀ with shaking (Alto plate reader; Cerillo) under anaerobic conditions (Extended Data Fig. 5).

Biofilm assay

Duets used in the co-culture assay described above were used to measure biofilm with the crystal violet assay. In brief, cultures were gently aspirated off and plates allowed to dry in a biosafety cabinet (5 min). Biofilms were fixed with 75% ethanol (10 min), stained with 0.5% Crystal Violet (5 min), rinsed with water and biofilm-associated crystal violet solubilized with 95% ethanol and the A₅₇₀ was measured (Extended Data Table 1).

Community modelling

Pairwise community modelling was performed using the MICOM (micom version 0.31.1)⁷¹ framework to simulate interactions between G. vaginalis and potential inhibitory partners (Extended Data Table 2). For each pairwise comparison, community models were constructed by combining genome-scale metabolic models of G. vaginalis with high lactate-producing vaginal microbes (L. jensenii, A. lactolyticus, A. tetradius, and F. vaginae). Community simulations were performed using the cooperative trade-off optimization algorithm (fraction of 0.5), which balances individual organism growth objectives with community-level fitness through quadratic programming optimization (solved using Operator Splitting Quadratic Program. Each pairwise interaction was simulated at three different initial relative abundance rations (1:1, 1:10 and 10:1, representing G. vaginalis:partner) to assess how competitive dynamics vary with numerical dominance. From each simulation, we calculated the growth suppression as the fractional reduction in G. vaginalis growth rate relative to its growth in monoculture to identify key metabolic features associated with effective G. vaginalis suppression and to inform engineering targets for probiotic strain development.

TriNetX cohort exploration

To explore the number of vaginitis cases presented in the University of Virginia (UVA) health system, we used TriNetX. The data used in this study were collected on 5 June 2025 from the TriNetX UVA Network, which provided access to electronic medical records (diagnoses, procedures, medications, laboratory values and genomic information) from 730,801 patients in the healthcare organization. This retrospective study is exempt from the requirement for informed consent. The data reviewed are a secondary analysis of existing data, do not involve intervention or interaction with human individuals and are de-identified per the de-identification standard defined in Section §164.514(a) of the HIPAA Privacy Rule. The process by which the data are de-identified is attested to through a formal determination by a qualified expert as defined in section §164.514(b)(1) of the HIPAA Privacy Rule. This formal determination by a qualified expert refreshed in December 2020.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

E.M.G. wrote the initial paper draft. E.M.G. performed computational studies. E.M.G. and G.L.K. performed experimental analyses. J.A.P. and G.L.K. supervised the work. E.M.G., G.L.K. and J.A.P. edited and approved the final paper.

Peer review

Peer review information

Nature Microbiology thanks Nick Bohmann, Caroline Mitchell and Jo-Ann Passmore for their contribution to the peer review of this work. Peer reviewer reports are available.

Data availability

The datasets generated and analysed during the current study are available via Zenodo at 10.5281/zenodo.19338012 (ref. ⁸⁹). Please note that the clinical cohort data obtained from the TriNetX University of Virginia Network are not publicly available owing to privacy restrictions and database-use agreements. Researchers wishing to request access to this specific clinical dataset may contact the UVA TriNetX account manager at glyons@virginia.edu. You can expect a response to data access requests within 1–2 weeks.

Code availability

The code generated during the current study are available via Zenodo at 10.5281/zenodo.19338012 (ref. ⁸⁹).

Competing interests

J.A.P. has financial stake in Cerillo, the manufacturer of the plate reader used in some experimental analyses. All other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41564-026-02380-w.

Supplementary information

The online version contains supplementary material available at 10.1038/s41564-026-02380-w.