Vaginal microbiota and immune impacts of a Lactobacillus crispatus live biotherapeutic and predictors of colonization in a randomized controlled trial

Summary

Bacterial vaginosis (BV) affects >25% of women worldwide and often recurs after standard-of-care metronidazole (MTZ) treatment. LACTIN-V, a live biotherapeutic product (LBP) containing Lactobacillus crispatus strain CTV-05, significantly reduced recurrent BV in a Phase 2b clinical trial, but efficacy was incomplete. Here, we characterize microbiota and immune effects using multi-omics and define correlates of treatment success. By week 12, L. crispatus-dominant microbiota is achieved in 30% of LBP recipients compared to 9% of placebo (benefit ratio: 3.31; p<0.005). This is primarily due to CTV-05, but native L. crispatus strains are also present and increase over time. Inflammatory cytokines decrease in both arms after MTZ but return to baseline in placebo recipients. Successful L. crispatus colonization associates with pre-MTZ microbiota, baseline inflammatory profiles, post-MTZ bacterial load, and clinical and behavioral variables. These findings elucidate LBP microbiota effects and identify predictors of treatment success, informing improved intervention strategies to advance women’s health.

eTOC Blurb

Bloom, Symul et al. show a vaginal live biotherapeutic product (LBP) containing the healthassociated bacterium Lactobacillus crispatus prevents bacterial vaginosis recurrence and reduces vaginal inflammation. Colonization is usually driven by the administered strain, but endogenous strains sometimes replace it. Success varies depending on pre-treatment vaginal microbiota composition and host factors.

Graphical Abstract

Introduction

Bacterial vaginosis (BV), a syndrome characterized by Lactobacillus-deficient vaginal microbiota, affects 23–29% of reproductive age women globally, with estimated annual economic burden of $4.8 billion^1,2. BV symptoms include discharge, malodor, pain, and itching, significantly impairing quality of life, self-esteem, and sexual well-being^3,4. BV is also linked to mucosal inflammation and risk of adverse outcomes including HIV acquisition, sexually transmitted infections, preterm birth, human papillomavirus infection, and cervical cancer^5–10. While BV occurs globally, it disproportionately affects those with lower socioeconomic status and members of racial or ethnic minority groups across diverse settings^1,11–13. Effectively treating BV is therefore a key objective to improve women’s health¹⁴.

Current first-line BV therapy involves oral or intravaginal antibiotics such as metronidazole (MTZ), which target species in the diverse anaerobic bacterial communities characteristic of BV^2,14. In most cases, MTZ reduces abundance of BV-associated anaerobes leading to emergence of bacterial communities dominated by Lactobacillus species (which are intrinsically MTZ-resistant)¹⁵, but BV frequently recurs post-treatment at rates ranging from >50% within one year to >75% within 16 weeks^16,17. It is hypothesized MTZ’s incomplete efficacy results from causes including failure to eradicate BV-associated bacterial communities or from post-MTZ replacement by Lactobacillus species prone to reverting to BV^15,18. Specifically, Lactobacillus iners is associated with increased risk of transition to BV-like states^19–21 and higher rates of adverse health outcomes^6,22–25. MTZ frequently results in establishment of vaginal microbiota dominated by L. iners rather than L. crispatus^26–31, providing a strong rationale that therapies promoting L. crispatus over L. iners may improve outcomes and enhance vaginal health^15,32,33.

New strategies to promote L. crispatus colonization during BV treatment are currently in development, including vaginal microbiome transplants, adjunctive therapies that selectively inhibit L. iners or promote L. crispatus growth, and L. crispatus-containing live biotherapeutic products (LBPs)^32–39. LACTIN-V, a vaginal formulation containing L. crispatus strain CTV-05, is the only LBP evaluated in large-scale clinical trials to date^37,40. Cohen and colleagues reported clinical results of a phase 2b randomized, placebo-controlled, double-blinded study showing participants who received 11 weeks of LACTIN-V after intravaginal MTZ developed rBV at significantly lower rates than placebo³⁷. However, rBV remained common, occurring in 30% and 39% of LACTIN-V recipients by weeks 12 and 24, respectively (versus 45% and 54%, respectively, in placebo recipients). Notably, the study characterized rBV clinically, but comprehensive molecular analysis of microbiota and factors associated with L. crispatus colonization was not performed.

Here we analyze samples from that trial to characterize LACTIN-V’s effects on vaginal microbiota composition and identify correlates of successful L. crispatus colonization by assessing vaginal microbial communities, bacterial strain dynamics, cytokine trends, and demographic/behavioral parameters³⁷. Reduction in rBV among LBP recipients corresponded to >3-fold higher rates of achieving L. crispatus-dominant microbiota. However, only 30% achieved L. crispatus-dominance at week 12, consistent with observed incomplete clinical efficacy. The CTV-05 strain accounted for the majority of L. crispatus colonization among LBP recipients, but native strains increased over time and sometimes displaced CTV-05. We describe immune correlates of LBP treatment and show pre-MTZ microbiota, baseline inflammatory profiles, post-MTZ total bacterial load, and selected clinical/behavioral variables were associated with L. crispatus-dominance in LBP recipients. Our findings reveal how an L. crispatus LBP alters the vaginal microbiota to improve BV treatment efficacy and identify factors linked to LBP success that can guide development of therapies to improve women’s health.

Results

Study design and participant characteristics

Women aged 18 to 45 years with BV completed five days of intravaginal MTZ, then were randomized 2:1 to intravaginal LACTIN-V (Osel, Inc.) or placebo as previously described (see also Methods)³⁷. Enrollment occurred at four U.S. sites and ~50% of participants reported ≥5 prior lifetime BV episodes³⁷. LACTIN-V is a powder formulation containing live L. crispatus strain CTV-05^41–43, administered as 2×10⁹ colony-forming units per dose. BV was diagnosed at the baseline “pre-MTZ” visit based on at least three of four Amsel criteria and Nugent score ≥4^44,45. Participants received intravaginal MTZ for five days, then were randomized to LBP or placebo within 48 hours of antibiotic completion at a “post-MTZ” visit. The first LBP (or placebo) dose was clinician-administered at the post-MTZ visit, then self-administered daily for four consecutive days, then twice weekly for 10 weeks (Figure 1A). Vaginal swabs were collected at the pre-MTZ visit, post-MTZ (before product administration), and at weeks 4, 8, 12, and 24 post-randomization. Demographic, clinical, and behavioral data showed no notable between-arm differences³⁷. 228 participants (152 LBP, 76 placebo) were enrolled. Samples from 213 (142 LBP, 71 placebo) were available for our analysis, representing 1,156 unique visits.

Figure 1: LBP treatment increased the proportion of women with L. crispatus-dominant vaginal microbiota.

A. LACTIN-V trial design. See also Table S1 and Figure S1.

B. Sankey diagram showing transitions between microbiota categories: L. crispatus-dominance (≥50% L. crispatus relative abundance), Lactobacillus (non-crispatus)-dominance (≥50% Lactobacillus, <50% L. crispatus), or non-Lactobacillus-dominance (<50% Lactobacillus). See also Spreadsheet S1.

C. Week 12 microbiota composition for LBP and placebo recipients, grouped as by Figure 1B categories, with BV diagnosis at top. See also Table S2.

D. L. crispatus relative abundances by arm at Weeks 12 and 24. Black dots and whiskers show mean and 95% CI.

E. Total bacterial load (qPCR) at each scheduled visit. Boxplots (here and in subsequent figures) show median (middle horizontal line), 25th and 75th percentiles (lower and upper box boundaries, “IQR”), values within 1.5 times the IQR (whiskers), and individual values outside that range (dots).

LBP treatment significantly increased rates of L. crispatus-dominant vaginal microbiota

Microbiota composition was determined by bacterial 16S ribosomal RNA gene sequencing. Four samples were excluded due to technical failures. The remaining samples (Table S1, Figure S1A) had 31,177 median (IQR 23,774–41,251) analyzable reads. Microbiota composition did not differ between arms at pre-MTZ or post-MTZ visits (PERMANOVA p = 0.63 and p = 0.14, respectively; Figure S1B-C).

We pre-specified two microbiota outcome parameters: ≥50% L. crispatus relative abundance (“L. crispatus-dominant”; a threshold useful for distinguishing outcomes in other studies^6,21) and ≥50% summed relative abundance of all Lactobacillus species (“Lactobacillus-dominant”, including L. crispatus). The primary endpoint for our analysis was L. crispatus-dominance at week 12, the visit corresponding to the trial’s primary clinical endpoint. Secondary endpoints included L. crispatus-dominance at week 24 and total Lactobacillus-dominance at weeks 12 or 24. Almost all participants had <50% Lactobacillus abundance pre-MTZ, consistent with BV diagnosis (Figure 1B, Spreadsheet S1). Majorities in both arms transitioned to Lactobacillus-dominance at the post-MTZ visit, largely driven by L. iners (52% of participants L. iners-dominant, 8% L. jensenii/mulieris-dominant, 13% Lactobacillus-dominant with multiple species). Only two participants (1%, one per arm) had post-MTZ L. crispatus-dominance (Figure 1B). Microbiota composition diverged post-randomization, with 30% of LBP recipients (n=37) and only 9% of placebo (n=5) having L. crispatus-dominance at week 12, a benefit ratio of 3.4 (95% CI: 1.4–8.1; p <0.005; Table 1, Figure 1B-C). At week 24, 35% of LBP recipients (n=39) versus 8% of placebo (n=4) had L. crispatus-dominance (benefit ratio 4.5; 95% CI: 1.7–11.9). L. crispatus relative abundance at weeks 12 and 24 was bimodal and well-separated by the pre-specified 50% threshold, with <4% (week 12) and <7% (week 24) of participants exhibiting relative abundances between 33–67%, indicating results were robust to alternative thresholds (Figure 1D). LBP did not increase total Lactobacillus-dominance at week 12 (56% and 48% of LBP and placebo recipients, respectively; benefit ratio: 1.16; 95% CI: 0.85–1.59) but did at week 24 (58% and 33% for LBP and placebo, respectively; benefit ratio: 1.76; 95% CI: 1.15–2.68; Table 1, Figure 1B-C). To confirm replication of known relationships between Lactobacillus and BV absence⁴⁶, we compared microbiota category to concurrent BV status at each post-randomization visit (n = 729 visits; Table S2 and Figure 1C). rBV was present at no visits with ≥50% L. crispatus colonization (0 of 209) and only 2 of 194 visits with <50% L. crispatus but ≥50% total Lactobacillus. rBV was present in over half (164 of 326) of non-Lactobacillus-dominant visits, confirming Lactobacillus-dominance in our cohort was specific for BV absence, while non-Lactobacillus-dominance was sensitive but less specific for BV presence.

Table 1:

LBP treatment significantly increased rates of L. crispatus-dominant vaginal microbiota

Taxon	Week	Microbiota Endpoint	LACTIN-V	Placebo	Benefit Ratio (95% CI)	P Value
L. crispatus	Week 12	≥ 50% L.c.	37 (30%)	5 (9%)	3.37 (1.40 – 8.11)	< 0.005
		< 50% L.c.	86 (70%)	51 (91%)
	Week 24	≥ 50% L.c.	39 (35%)	4 (8%)	4.49 (1.69 – 11.90)	-
		< 50% L.c.	74 (65%)	48 (92%)
Lactobacillus	Week 12	≥ 50% Lacto	69 (56%)	27 (48%)	1.16 (0.85 – 1.59)	-
		< 50% Lacto	54 (44%)	29 (52%)
	Week 24	≥ 50% Lacto	65 (58%)	17 (33%)	1.76 (1.15 – 2.68)	-
		< 50% Lacto	48 (42%)	35 (67%)

Total vaginal bacterial load was assessed via quantitative PCR at post-MTZ and subsequent visits. Median load was significantly lower with wider range post-MTZ (median 10^2.69 copies/swab; IQR 10^0.8 - 10^6.44) than at subsequent visits (median 10^7.52 copies/swab; IQR 10^7.9 - 10^7.93), consistent with antibiotic-mediated microbiota biomass depletion followed by repopulation (Figure 1E). Post-MTZ microbiota composition and bacterial load did not significantly correlate as assessed using the RV coefficient, a multi-table multivariate generalization of the squared Pearson correlation^47,48 (RV coefficient = 0.02, p-value > 0.1).

Microbiota composition established early during treatment tended to persist

To characterize microbiota trajectories with greater resolution and reduce data dimensionality for subsequent analyses, we summarized microbiota composition using “topic” mixtures^49,50. Compared to clustering or categorizing samples into community state types (CSTs) ⁵¹, cervicotypes⁶, or sub-CSTs⁵², this method provides probabilistic weightings, enabling more accurate descriptions of taxonomic composition and longitudinal dynamics⁵⁰. In our cohort, reference CST assignment performed poorly, especially for Prevotella-dominated samples (median BC dissimilarity > 0.5), and with >40% of samples almost equally similar to ≥2 CSTs. To identify topics (which can be interpreted as bacterial subcommunities) and estimate topic proportions, we fitted a latent Dirichlet allocation (LDA) model (Bayesian “topic” model)⁵³, adapting a described approach⁵⁰ constraining topics to consist either of Lactobacillus or non-Lactobacillus species. We inferred four non-Lactobacillus topics (Figures 2A, S1D-E) and defined four Lactobacillus-dominated topics: three exclusively comprising a single species (L. crispatus, L. iners, or L. jensenii/mulieris) and one a mixture of remaining Lactobacillus species (Figure S1F). Figure 2B displays participant topic proportions throughout the trial.

Figure 2: Microbiota composition established early during treatment tended to persist.

A. Taxa proportions in each non-Lactobacillus topic for taxa comprising ≥1% of any topic; topic proportions sum to one. Topics are named by predominant taxon. “Ca. L. v. (BVAB1)”: Candidatus Lachnocurva vaginae (BVAB1). “P. bivia”: Prevotella bivia. “P. timonensis”: Prevotella timonensis. See also Figure S1.

B. Microbiota composition at each scheduled visit for participants with data from ≥3 visits, ordered by microbiota trajectories. See also Table S3.

C. Relative topic abundance at each visit. Lines connect individual participants’ values.

D. Bray-Curtis dissimilarity between each participant’s microbiota at indicated consecutive visits. See also Figure S1.

Three non-Lactobacillus topics predominated in both arms pre-MTZ, with Candidatus Lachnocurva vaginae (BVAB1), Gardnerella (ASV1), or Prevotella bivia as predominant taxa, respectively (Figure 2C and S1G). Composition shifted post-MTZ (Figure 2D), primarily due to increased L. iners, with the most abundant non-Lactobacillus topic being Gardnerella-predominant (Figure 2C and S1G). Another shift occurred between post-MTZ and week 4 visits, greater in the LBP arm (median pairwise BC dissimilarity: 0.81; IQR: 0.50–0.97) than placebo (median BC dissimilarity: 0.58; IQR 0.41–0.80) driven primarily by increased L. crispatus among LBP recipients (Figure 2C-D and S1G). Composition stabilized after week 4 with median BC dissimilarity <0.55 between all consecutive visits, comparable between arms (Figure 2D). LBP recipients who achieved L. crispatus-dominance had higher stability (median BC <0.26) than those who did not (median BC >0.48, Figure S1H).

Microbiota category analysis also showed stability after week 4. Among LBP recipients with available data, 71% with L. crispatus-dominance at week 12 or at week 24 were also L. crispatus-dominant at week 4 (Figure 2B, Table S3). 79% of those with L. crispatus-dominance at week 12 retained it at week 24. However, early L. crispatus-dominance did not guarantee persistence. Only 49% and 57% of LBP recipients with L. crispatus-dominance at week 4 retained it at weeks 12 and 24, respectively. By contrast, among 38 LBP recipients with week 4 non-Lactobacillus-dominance, 79% remained non-Lactobacillus-dominant at week 12 compared to just 10.5% each with week 4 L. crispatus-dominance or other Lactobacillus-dominance (Table S3). Thus, microbiota established early tended to persist through week 24, and early Lactobacillus-dominance – particularly L. crispatus-dominance – was strongly linked to ongoing Lactobacillus-dominance.

CTV-05 was the dominant L. crispatus strain in LBP recipients early, but native strains replaced it in a subset of recipients

We next examined what fraction of L. crispatus in LBP recipients consisted of CTV-05 versus non-LBP “native” strains, using shotgun metagenomic sequencing (metrics in Methods and Spreadsheet S2). L. crispatus strain genotypes and proportions were inferred from metagenomes using StrainFacts⁵⁴. StrainFacts defines strain genotypes based on single-nucleotide variant (SNV) profiles at defined biallelic sites in the species core genome. In a separate validation study, we found StrainFacts performs well inferring LBP (including CTV-05) and native strain genotypes and proportions in samples with ≥5% relative abundance of total L. crispatus⁵⁵, so samples were selected for strain analysis based on a 5% threshold. L. crispatus strain inference was successful in 313 samples. Based on validation results, a 10% strain fractional abundance was used to define presence in a sample⁵⁵.

We inferred presence of 24 distinct L. crispatus strains detected in ≥1 sample within the cohort. To determine which strain represented CTV-05, we generated a closed CTV-05 genome, then determined its StrainFacts genotype. CTV-05’s genotype near-perfectly matched a strain widely prevalent in metagenomes (Jaccard similarity >0.999), identifying that strain as CTV-05. CTV-05 was inferred in 217 of 256 (84%) of post-randomization LBP recipient samples with ascertainable strain composition but only 5 of 57 (8.7%) samples from placebo recipients or pre-treatment LBP recipients (Figure 3A), suggesting CTV-05 detection was largely accurate but occasionally misclassified native strains as CTV-05. Most participants with an inferred native strain had only one native strain (occasionally two or three) throughout the study, and participants often had the same native strain(s) at multiple visits (Figure 3B). Since strain inference was blinded to participant identity^54,55, these results supported analysis validity.

Figure 3: CTV-05 was the dominant L. crispatus strain in LBP recipients early, but native strains replaced it in a subset of recipients.

A. CTV-05 and native L. crispatus strain dynamics inferred from metagenomes. Samples were categorized by proportional strain abundance: >90% CTV-05 (“High CTV-05”), 10–90% CTV-05 (“Mixed”), and <10% CTV-05 (“Low CTV-05”). Green: samples with technical failures in strain inference, see also Spreadsheet S2.

B. Number (percentage) of participants in each arm who had the indicated number of native strains detected at least once during the trial.

C. Metagenomically inferred L. crispatus strain composition of LBP recipients selected for isolations (n=6). Color indicates unique inferred strains; gray shows summed abundance of non-L. crispatus taxa. Symbols indicate samples used for L. crispatus isolation. See also Figure S2.

D. Phylogenetic tree of CTV-05 (arrow) and L. crispatus isolates from 3C. Symbols indicate isolate source; colors show each isolate’s closest matching inferred strain (see 3C). Inset: typical L. crispatus colony morphology. See also Figure S2 and Spreadsheet S3.

E. Transition plot for LBP recipients showing consecutively scheduled pairs of post-randomization visits with L. crispatus strain data or with insufficient L. crispatus for strain inference. Starting (“Visit t”, top row) and subsequent (“Visit t+1”, bottom row) L. crispatus relative abundance. Second row: study week at Visit t. Third and fourth rows: CTV-05 and summed native strain proportions at Visit t (third row) and Visit t+1 (fourth row). Visit pairs are grouped by Visit t strain category and ordered by Visit t+1 CTV-05 proportion.

F. Pie charts for Visit t strain categories (Figure 3E), summarizing strain category frequencies at Visit t+1. See also Figure S2.

To experimentally test strain-tracing accuracy, we performed isolations in 12 distinct samples from 6 participants metagenomically predicted to contain no (<0.5%) L. crispatus (n=2 samples) or abundant (>40%) L. crispatus (n=10). This included samples predicted to contain solely CTV-05, solely native strains, or mixed CTV-05 and native strains (Figures 3C & S2A). We recovered L. crispatus from all samples predicted to contain it (106 total isolates from 10 samples), presumptively identifying isolates by colony morphology confirmed via genome sequencing (Figures 3D and S2A and Spreadsheet S3). A phylogenetic tree of L. crispatus isolate genomes revealed 7 clades, each containing multiple nearidentical genomes. The largest clade, which clustered with CTV-05’s genome (Figure 3D, red arrow), included isolates from 7 of 8 samples and all 5 participants with predicted CTV-05 (Figure 3C-D and S2A). Other clades each contained isolates from a single participant. Samples predicted to have two native strains each produced two distinct non-CTV-05 clades and samples predicted to have a single native strain each produced a single clade. The only exception was a sample (participant STI.00625, week 24) predicted to contain CTV-05 plus a native strain that produced no CTV-05 isolates, but instead produced two non-CTV-05 clades. Among participants with multiple cultured timepoints, participant STI.00629 produced CTV-05 isolates at week 12 and a native strain at week 24 (consistent with predicted complete strain replacement; Figures 3C-D, S3A), participant STI.00356 produced CTV-05 isolates at week 12 and both CTV-05 and a native strain at week 24 (consistent with predicted partial strain replacement), and participant STI.00625 produced CTV-05 at week 8, a mix of CTV-05 and a native strain at week 12, and a mix of the same native strain plus a new native strain at week 24 (consistent with predictions except as noted above for week 24). Jaccard similarity between each isolate’s StrainFacts genotype and all inferred strains showed each isolate’s most closely matching inferred strain was present in the sample from which the isolate was derived (Figures 3C-D, S3B). All members of each clade matched the same inferred strain. In participant STI.00625’s week 24 sample, the closest inferred strain for one clade’s genomes (Clade 5) was CTV-05, but at lower similarity than the true CTV-05 clade (Figure S3B). Thus, isolations showed strain inference was accurate but occasionally misclassified native strains as CTV-05.

To analyze L. crispatus strain dynamics, we classified samples into three categories: >90% inferred CTV-05 fractional strain abundance (“High-CTV-05”), 10–90% CTV-05 (“mixed strains”), and <10% CTV-05 (“Low-CTV-05” or high native strain(s); Figure 3A). From week 4 onwards, most LBP recipients with analyzable L. crispatus had high-CTV-05 strain abundance, but the number and proportion with high native strains progressively increased while numbers and proportions with high-CTV-05 or mixed strains decreased (Figure 3A). At week 4, 96.1% (n=74) of LBP recipients with ascertainable strains had high-CTV-05 or mixed strains, while just 3.9% (n=3) had high native strains. However, by week 24, 63.2% (n=31) had high-CTV-05 or mixed strains, while 36.8% (n=18) had high native strains. Participant-level analysis showed among 144 visits with high-CTV-05, two main outcomes were observed at the next visit: 68% (n=98) maintained high-CTV-05 and 26% (n=38) fell below the strain detection threshold (Figure 3E-F). Among 18 visits with high native strains, 61% (n=11) retained high native strains and none transitioned to high-CTV-05. Among 23 visits with mixed strains, just one (4%) transitioned to high-CTV-05, while 39% (n=9) remained mixed and 35% (n=8) transitioned to high native strains, including participants still receiving LACTIN-V treatment (Figure S2C). Thus, CTV-05 was frequently the primary L. crispatus strain in LBP recipients, but when native strains dominated or co-occurred with CTV-05, they often subsequently replaced it whereas CTV-05 almost never replaced native strains.

MTZ reduced mucosal inflammation associated with low Lactobacillus abundance, which reverted by week 24 in placebo but not LBP recipients

To assess treatment effects on vaginal inflammation, we measured cytokines and chemokines in vaginal swab eluates (Figure S3A)^6,50. Analyte concentrations were strongly positively correlated (Figure S3B), suggesting a technical “size-effect”⁵⁶ from sampling variation for which we adjusted using PC1-subtraction. Participant-level immune dynamics were assessed by comparing each participant’s adjusted pre-MTZ concentration for each analyte with concentrations at subsequent visits. IL-1β decreased post-MTZ whereas IP-10 rose, consistent with reported microbiota and BV associations^5,6,57,58 (Figure 4A-B). Shifts persisted in both arms through week 12, but placebo recipients returned to pre-MTZ levels by week 24, producing differences of −0.39 (95% CI: −0.66 - −0.14) between LBP and placebo for IL-1β and 0.51 (95% CI: 0.21–0.79) for IP-10. By contrast, IL-6 – which has been reported not to differ in BV⁵⁷ – remained unchanged throughout (Figure 4C).

Figure 4: MTZ reduced mucosal inflammation associated with low Lactobacillus abundance, which reverted by week 24 in placebo but not LBP recipients.

A. Differences in log10-transformed, adjusted IL-1β concentrations between participants’ pre-MTZ and week 12 (left) or week 24 (right) visits. Significance of week 24 differences are shown (Wilcoxon signed ranked test with Benjamini-Hochberg correction, **: <0.01, *: <0.05, ns: >0.05). See also Figure S3.

B. Same as Figure 4A for IP-10.

C. Same as Figure 4A-B for IL-6.

D. DiSTATIS screeplot: eigenvalues (a.u.) of DiSTATIS compromise for first 10 latent components.

E. DiSTATIS correlations between 1st and 2nd compromise latent dimensions and microbiota topic proportions or cytokine transformed concentrations.

F. Same as Figure 4E for 3rd and 4th latent components. See also Figure S3.

Week 24 results for IL-1β and IP-10 suggested associations with total Lactobacillus relative abundance, which differed between arms at week 24 (Table 1). Microbiota category analysis confirmed lower IL-1β and higher IP-10 were associated with Lactobacillus-dominance (Figure 4A-B). Overall microbiota composition (topic proportions) significantly correlated with concentrations of the 13 analyzed cytokines/chemokines (RV coefficient 0.18; p-value <0.005). Analyzing visits individually produced similar results (Figure S3C) except for lower correlation at the post-MTZ visit (where bacterial load varied widely). To characterize microbiota-immune relationships in greater detail, we employed DiSTATIS^59,60, which infers a consensus representation of the two datasets from between-sample dissimilarities. Projecting microbiota topics and cytokines/chemokines revealed the first latent dimension discriminated Lactobacillus-dominated from non-Lactobacillus-dominated samples (Figure 4D-F). MIG, IP-10, and ITAC were positively correlated (covaried with Lactobacillus) while IL-1β, TNF-α, and IL-1α were negatively correlated (covaried with non-Lactobacillus topics, Figure 4E). The second dimension discriminated L. crispatus from L. iners but most cytokines and chemokines showed little correlation, indicating minimal immune association with individual Lactobacillus species. Most cytokines and chemokines had high correlation with the 3rd or 4th latent components but microbiota topics did not, indicating a degree of microbiota-independent immune variation (Figure 4F). In these dimensions, IFNγ and IL-17 correlated but were anti-correlated with IL-6, MIP-3α, MIP-1β, and MIP-1α. Analysis distinguishing CTV-05 from other L. crispatus strains showed no clear differences in microbiota-immune correlations (Figure S3D-G).

LBP treatment benefits differed for women with different pre-MTZ microbiota

We performed exploratory post-hoc analyses assessing whether participants with different pre-MTZ microbiota composition responded differently to the LBP (treatment effect heterogeneity). Stratifying participants by their most abundant pre-MTZ taxon (genus-level) identified four groups: Lactobacillus-predominant, Gardnerella-predominant, Prevotella-predominant, and Ca. Lachnocurva vaginae (BVAB1)-predominant. LBP benefit was evaluated for two separate outcomes: L. crispatus-dominance (at week 12 or 24) and BV recurrence (by week 12 or 24) (Figure 5A-B). Analysis of L. crispatus-dominance suggested some groups differentially benefited from LBP treatment (adjusted p-values = 0.11 and 0.03 at weeks 12 and 24, respectively; Figure 5C-D). LBP recipients with pre-MTZ Lactobacillus-, Gardnerella-, or Prevotella-predominance attained higher rates of subsequent L. crispatus-dominance than their placebo counterparts, with modestly larger differences at week 24 than week 12 (Figure 5D). In contrast, LBP treatment was not associated with higher rates of L. crispatus-dominance for participants with pre-MTZ Ca. Lachnocurva vaginae (BVAB1)-predominance. LBP benefit in preventing rBV also differed by pre-MTZ microbiota (adjusted p-values <0.05 and <0.01 at weeks 12 and 24; Figure 5E-F). BV recurred in almost all placebo recipients with pre-MTZ Prevotella-predominance by week 24 versus just over half of LBP recipients. By contrast, LBP recipients with pre-MTZ Ca. Lachnocurva vaginae (BVAB1)-predominance had slightly higher rBV rates than placebo (Figure 5E-F). Sensitivity analyses grouping participants by pre-MTZ CST⁵² (Figure S4) or using a model-based approach relying on topic relative abundances provided largely consistent results. Collectively, these exploratory analyses suggested lack of LBP benefit for participants with pre-MTZ Ca. Lachnocurva vaginae (BVAB1)-predominance and variable degrees of benefit for all others.

Figure 5: LBP treatment benefits differed for women with different pre-MTZ microbiota.

A. Week 12 and 24 outcomes including L. crispatus-dominance (dark orange), <50% L. crispatus (light orange), rBV (dark blue), or no rBV (light blue), grouped horizontally by predominant pre-MTZ genus (“L.”: Lactobacillus; “BVAB1”: Ca. Lachnocurva vaginae), showing groups with ≥2 participants per arm. (White: no data.)

B. Pre-MTZ microbiota composition organized as in 5A.

C. Rates and 95% CI of L. crispatus-dominance at week 12 (left) or 24 (right) by pre-MTZ microbiota group (5A).

D. Between-arm rate differences (LBP - placebo) and 95% CI based on Figure 5C. Color shows degree of benefit in achieving L. crispatus-dominance (blue: benefit with LBP). Significance determined by analysis of deviance test, with Benjamini-Hochberg adjustment for p-values across 5D and 5F.

E. Same as 5C, but for rBV by week 12 or 24.

F. Same as 5D, but for rBV by week 12 or 24 (blue: LBP benefit in reducing rBV). See also Figure S4.

Microbiota, vaginal, immune, and sociodemographic factors were associated with L. crispatus dominance in LBP recipients

To identify factors associated with achieving L. crispatus-dominance in LBP recipients, we performed multiblock partial least square discriminant analyses (MB-PLS-DA) on LACTIN-V arm data (Figure 6A-C, S5), leveraging the biological data described above and sociodemographic, clinical, and behavioral data collected during the clinical trial (variables and their unadjusted relationships to rBV and microbiota outcomes in Spreadsheet S4). MB-PLS-DA simultaneously examines how variables contribute to a model’s ability to explain outcomes both individually (“variable importance”) and in thematic groups of variables (“block importance”)^61,62. A block could thus contribute to model performance either through modest contributions from each of its variables or strong contribution from a few variables. We defined three models corresponding to different trial phases: “initial phase” (post-MTZ to week 4), “continuation phase” (week 4 to week 12), and “follow-up phase” (week 12 to week 24, Figure S5A). Explanatory variables were grouped into 13 total blocks characterizing participant demographics, baseline (pre-MTZ) vaginal ecosystem, vaginal ecosystem at the preceding visit, and clinical/behavioral parameters (Figure 6A and Spreadsheet S4). For each phase, we evaluated whether models could discriminate among three microbiota categories: ≥50% L. crispatus, ≥ 50% Lactobacillus but <50 % L. crispatus, or <50% Lactobacillus. We used cross-validation to select latent component number and avoid over-fitting, and nested models to further assess block contributions (Figure S5A-C). Results showed initial-phase colonization was more difficult to explain and explained by more factors than colonization at later phases (see Methods, Figure 6A and S5B-C).

Figure 6: Microbiota, vaginal, immune, and sociodemographic factors were associated with L. crispatus dominance in LBP recipients.

A. Relative cumulative block importances from MB-PLS-DA analysis of variables predictive of microbiota categories defined in 1B. Point-estimates and 95% CI are shown for each variable block for initial-phase (left), continuation-phase (middle), and follow-up phase models (right) for the LBP arm. Black vertical lines show the (null) hypothesis that all blocks have equal importance. See also Figure S5, Spreadsheet S4.

B. Bi-plot of initial-phase model (LBP arm). Dots (colors: week 4 microbiota category) represent participant scores; arrows (colors: blocks as in 6A) show loadings of the most important variables in the first 2 latent variables. Perfect separation of microbiota categories would indicate perfect model performance. See also Figure S5.

C. Scree-plot of first 15 eigenvalues of model from 6B. Dark gray: first two latent variables.

D. Concentrations of cytokines (MIG and IL-1b) whose pre-MTZ levels (left) predicted week 4 category in the initial-phase model (LBP arm). Post-MTZ (middle) and week 4 (left) concentrations are also shown. Data are grouped and colored by week 4 microbiota category in all panels. Left panel shows t-tests comparing pre-MTZ concentrations for participants achieving L. crispatus-dominance at Week 4 to all others (**: <0.01, *: <0.05). See also Figure S6-7.

E. Rates and 95% CI of L. crispatus-dominance at week 12 (left) or 24 (right) by self-declared race.

F. Between-arm rate differences (LBP - placebo) and 95% CI for achieving L. crispatus-dominance based on 6E.

In the initial-phase model for the LBP arm, the most important block was previous (i.e., post-MTZ) vaginal environment (Figure 6A, left panel), which included post-MTZ total bacterial load, α-diversity, and pH. Lower values were associated with higher chance of week 4 L. crispatus-dominance and these were among the 5 most important variables (Figure 6B-C and S5D). The next most important blocks were demographics and blocks characterizing pre-MTZ vaginal environment, cytokine concentrations, and microbiota composition (Figure 6A and S5D). Specifically, high pre-MTZ α-diversity and pH were negatively associated with week 4 L. crispatus-dominance (Figure 6B). Interestingly, pre-MTZ IL-1β levels were positively associated with week 4 L. crispatus-dominance while pre-MTZ IP-10 and MIG levels showed opposite association (Figure 6B, S5D, 6D). Self-identified race and education were the most important demographic variables and slightly contributed to model performance, with participants identifying as more highly educated and white achieving higher rates of week 4 L. crispatus-dominance (Figure S5D). However, self-identified race and education were highly correlated with each other (χ² p-value < 0.01; Figure S5E) and with certain pre-MTZ vaginal characteristics such as microbiota α-diversity as well as study site (Figure S5F-G), but adding study site as a demographic block variable did not improve model performance. Thus self-identified race – itself a non-biological, socially defined category^63,64 – was confounded with measured and unmeasured socioeconomic, geographic, and clinical factors in this cohort such that individual relationships with L. crispatus colonization could not be identified. Notably, women self-identifying as white had higher rates of L. crispatus-dominance in both the placebo and LBP arms, but LBP treatment increased L. crispatus-dominance relative to placebo at a similar degree for all self-declared racial groups at week 12 (Figures 6E-F). There was a modest trend toward greater benefit in women identifying as white at week 24, which was not statistically significant (adjusted p-values = 0.57 at both weeks). Remaining blocks were not important for initial-phase predictions, although within these blocks, factors with higher variability across participants like menstrual bleeding and sexual activity showed greater variable importance (Figure 6A-B, Figure S5D).

In the LBP arm continuation-phase (week 4 to 12), microbiota category at the previous visit was by far the most important block (Figure 6A). While no other blocks significantly improved model performance (Figure S5C), individual variables including sexual activity, bleeding, douching, and non-hormonal IUDs were negatively associated with L. crispatus-dominance (Figure S5D). In the follow-up phase (week 12 to 24), microbiota category at the previous visit again had high importance and sexual behavior and antibiotic use mildly contributed to explaining outcomes (Figure 6A, right panel, Figure S5D).

Analogous placebo arm models showed no significant predictive value for the initial-phase or follow-up phase in cross-validation (Figure S6-7). The placebo continuation-phase model had modest predictive value, with previous visit microbiota category serving as the best predictor (Figure S6-7).

Discussion

Vaginal LBPs offer significant promise to improve health, but mechanisms and strain dynamics of LBP colonization and efficacy are incompletely understood^18,46. We analyzed samples and data from a randomized, placebo-controlled trial of LACTIN-V – a single-strain L. crispatus LBP to prevent BV recurrence³⁷ – employing microbiome sequencing, immunologic characterization, and multi-omic analyses to investigate LBP effects and correlates of success. LBP treatment produced L. crispatus-dominant vaginal microbiota in 30% and 35% of recipients at weeks 12 and 24, respectively, substantially superior to placebo. L. crispatus colonization was primarily due to the LACTIN-V strain CTV-05, although CTV-05 diminished and native L. crispatus strains increased over time. Lactobacillus-dominance ‒ particularly L. crispatus-dominance ‒ at early timepoints was linked to persistent Lactobacillus-dominance, while early non-Lactobacillus-dominance also tended to persist. Exploratory analyses showed LBP benefits differed by baseline microbiota composition and identified microbial, immune, demographic, and behavioral factors associated with L. crispatus-dominance among LBP recipients.

Investigation of L. crispatus strain dynamics showed CTV-05 accounted for most L. crispatus colonization in LBP recipients, but the number and proportion with high native strains progressively increased over time from 3.9% at week 4 to 36.8% at week 24. When CTV-05 co-occurred with native strains, the native strains frequently replaced CTV-05 at subsequent visits. By contrast, native strain replacement by CTV-05 was very rare. Further research is needed to determine if native strains that outcompete CTV-05 have functional characteristics providing selective advantages or if initial CTV-05 colonization establishes a permissive environment for native strains. It is also unknown if strain dynamics observed in this US cohort will be replicated in non-US populations treated with LACTIN-V or in women receiving LBPs containing L. crispatus from different clinical or geographic sources.

Vaginal mucosal cytokine and chemokine levels changed significantly with microbiota shifts from BV to Lactobacillus-dominance, and reverted if microbiota returned to non-Lactobacillus-dominance. Specifically, cytokines including IL-1α, IL-1ꞵ, and TNFα increased with decreasing Lactobacillus, whereas IP-10, MIG, and ITAC showed opposite patterns, consistent with prior reports^5,6,57,58. A prior study of a small subset of trial participants reported lower IP-10 levels at study end in women with CTV-05 compared to other L. crispatus⁶⁵, but we saw no similar association in the larger cohort. Favorable cytokine changes persisted in LBP recipients at week 24 compared to placebo, largely due to higher rates of overall Lactobacillus-dominance rather than CTV-05-specific effects.

LACTIN-V efficacy varied with differences in pre-MTZ microbiota composition. Compared to placebo, LBP treatment particularly reduced rBV among participants with pre-MTZ Prevotella-predominance, whereas participants with pre-MTZ Ca. Lachnocurva vaginae (BVAB1)-predominance uniquely trended toward more rBV with the LBP and little or no benefit in achieving L. crispatus-dominance. Reasons for these patterns are unclear, but may reflect greater competitive ability or MTZ resistance in Ca. Lachnocurva vaginae or its co-occurring species^66,67. Possible differences in Ca. Lachnocurva vaginae response to oral versus intravaginal MTZ have been reported⁶⁸. Ca. Lachnocurva vaginae remains uncultured^69,70, so our results emphasize that cultivating and phenotypically characterizing it are key research priorities. Importantly, these exploratory post-hoc analysis results should be considered hypothesis-generating observations requiring validation in future studies.

Integrated multiblock analysis of LBP recipients showed low bacterial load, vaginal pH, and α-diversity at the post-MTZ visit were associated with subsequent L. crispatus-dominance, elucidating the microbiota dynamics underlying a post-hoc analysis that found participants clinically cured of BV at the post-MTZ visit had less subsequent rBV¹⁸. Low pre-MTZ MIG and IP-10 levels and high pre-MTZ IL-1β were also associated with early L. crispatus-dominance – an unexpected finding given their opposite correlations with Lactobacillus at concurrent visits. Reasons are unclear but may involve more vigorous pre-MTZ immune response helping clear BV-associated bacteria, presence of inflammatory but MTZ-responsive bacteria, or other mechanisms. After week 4, the strongest predictor of microbiota category was the prior visit’s microbiota, indicating early post-LBP microbiota best predicts subsequent outcomes.

Self-declared race and education were also linked to microbiota composition at week 4 and beyond in LBP recipients, but LBP treatment increased L. crispatus-dominance compared to placebo for all racial groups. Prior US studies have linked BV and non-Lactobacillus-dominance to self-identified non-white race and/or Hispanic ethnicity^12,51. However, education and race correlated closely in our cohort, differed by study site, and thus likely correlated with unmeasured factors relevant to microbiota composition including nutrition and health service access, environmental exposures, and psychosocial stressors that elevate levels of hormones like cortisol which can alter sex hormone regulation, menstrual bleeding, and cervicovaginal epithelial function^63,64,71,72. BV’s complex relationship with socially defined categories like race/ethnicity is highlighted by the fact that BV disparities are observed in ancestrally unlinked, socioeconomically disadvantaged populations globally including nomadic women in Iran⁷³, aboriginal women in Canada⁷⁴, and Tibetan women in China⁷⁵, while lower educational and socioeconomic status correlated with lower L. crispatus-dominance in a homogeneous cohort of Caucasian Finnish women⁷⁶. A recent Chinese study reported associations between vaginal microbiota and environmental factors including ambient temperature and pollutant exposure, neither of which were addressable in our study⁷⁷. Thus, our results should not be interpreted as showing education or race are mechanistic causes of BV risk or treatment outcome and we believe additional research on determinants of these associations is needed. Compared to placebo, LBP treatment increased L. crispatus-dominance for women identifying as Black to a similar degree as other groups, even though their absolute rates of L. crispatus-dominance were lower in both placebo and LBP arms. Thus, further improving LBPs or other BV therapies has potential to offer particular benefit for these women.

Despite evidence sexual activity influences BV risk and vaginal microbiota composition⁷⁸, self-reported sexual behavior was not strongly predictive of microbiota composition in LBP recipients. This may be due to consistent sexual behavior throughout the trial such that effects of sex occurred early and were obscured by other factors associated with early L. crispatus-dominance. Since self-report is often an unreliable measure of true sexual activity^79–82, inaccurate self-reporting may also have decreased power to detect effects. We stress that these results are exploratory and require future validation.

In summary, our analysis of microbiota and strain dynamics underlying clinical effects of a vaginal L. crispatus LBP identifies key microbiota and host factors associated with treatment outcomes. If validated, these findings may help identify patients likely to benefit from treatment and guide discovery of novel targets to enhance LBP efficacy and improve women’s health globally.

Limitations

This study has several limitations. First, samples from a few participants were unavailable (10 of 152 LBP and 5 of 76 placebo recipients) and technical sequencing failures affected four samples, so our cohort was slightly smaller than the original trial. Second, the study was powered to detect differences in rBV, so its limited size reduced power to detect heterogeneity in treatment effects or correlates of L. crispatus-dominance. Third, the sampling schedule did not permit high-resolution temporal assessment of microbiota dynamics. Fourth, participants were followed for 24 weeks post-randomization (including 13 weeks post-treatment), precluding assessment of longer-term microbiota patterns. Finally, detailed sexual, contraceptive, hygiene, and clinical data were available, but only a narrow range of environmental and sociodemographic variables were collected, limiting analysis of factors that might help explain associations with microbiota composition.

Resource availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Douglas S. Kwon (dkwon@mgh.harvard.edu).

Materials availability

• There are restrictions to the availability of L. crispatus strain CTV-05, including CTV-05 strains reisolated from trial samples as part of this study, because of intellectual property and commercial interests of Osel, Inc., the manufacturer of LACTIN-V, an investigational drug under Investigational New Drug (IND) status with the US Food and Drug Administration. Requests to use CTV-05 for academic research purposes under a materials transfer agreement should be directed to Osel, Inc., at tparks@oselinc.com or info@oselinc.com.

• All other unique resources generated in this study are available from the lead contact with a completed materials transfer agreement.

Data and code availability

• Sequence data and genome assemblies have been deposited under NCBI BioProject PRJNA1303956 and are publicly available as of the date of publication.

• Clinical and processed and transformed omics data are available on Zenodo at DOI 10.5281/zenodo.17755173

• Original code is available on Zenodo at DOI 10.5281/zenodo.18524760.

• Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Study design and sample collection

Samples analyzed in this study were obtained as part of a previously reported phase 2b, randomized, double-blind, placebo-controlled trial of the L. crispatus LBP LACTIN-V (Osel, Inc., Mountain View, CA) for prevention of rBV³⁷. Use of samples was approved by the Mass General Brigham Institutional Review Board (IRB Protocol #2020P002237) as well as the UCSF Institutional Review Board (IRB Protocol #19–28337). LACTIN-V is a single-strain LBP formulated as a powder containing a preservation matrix and 2×10⁹ colony-forming units per dose of the L. crispatus strain CTV-05, which was isolated in 1993 from a vaginal sample of a healthy US woman without BV or sexually transmitted infections in Seattle, Washington^37,41–43. CTV-05 is administered using a pre-filled, single use vaginal applicator and was compared in the trial to a placebo formulation consisting of the preservation matrix without CTV-05. The trial was conducted at four centers within the USA and enrolled premenopausal, non-pregnant women aged 18–45 years. Eligibility criteria were previously described³⁷. The original clinical trial enrolled 152 LBP and 76 placebo recipients, of whom 142 LBP and 71 placebo recipients (213 total) had samples available for our analysis. Participants’ self-declared race and educational level are reported in Figure S5E, and 20% of participants identified as being of Hispanic or Latino ethnicity.

Briefly, potential participants attended a screening (pre-MTZ) visit at which they were determined to be eligible for the study if testing revealed presence of BV as determined by both a Nugent score of 4–10 on Gram stain of a vaginal smear⁴⁵ and presence of at least three of four Amsel criteria (characteristic vaginal discharge, >20% clue cells on microscopy of a vaginal wet prep, vaginal fluid pH >4.5, and presence of a fishy odor upon addition of 10% potassium hydroxide to a vaginal specimen)⁴⁴, as well as negative testing for HIV, syphilis, gonorrhea, chlamydia, trichomonas, and urinary tract infection. Women found to be eligible based on this evaluation completed 5 days of intravaginal MTZ therapy within 30 days of their pre-MTZ visit (Figure 1A, S1A). They then returned to the trial clinic within 48 hours of completing antibiotics (post-MTZ visit) and were randomized in a 2:1 ratio to receive either LBP or placebo after providing written informed consent. The first dose of LBP or placebo was clinician-administered at the randomization (post-MTZ) visit, then doses were vaginally self-delivered daily for the next four days, then twice weekly for ten additional weeks. In-person study visits were scheduled 4, 8, 12, and 24 weeks after randomization, at which vaginal swabs were collected and stored (details below) and clinical report forms (CRFs) were completed. In addition, two phone visits were planned at week 16 and 20 during which a subset of the clinical report forms were filled to capture information on adverse events, menstruation, concomitant medication use, and sexual behavior protected or unprotected by condoms. Participants who desired additional in-person visits were invited to present to the clinics. Swabs and CRFs were collected at these additional visits.

Clinician-collected vaginal swab samples were obtained via speculum exam at in-person study visits. Two types of swabs were collected in parallel for analysis³⁷. One set of swabs was collected at all in-person visits including the pre-MTZ (screening) visit using the Starplex^™ Scientific Multitrans^™ Collection and Transportation System (Starplex^™ Scientific S1600), which comprises a plastic-shaft Dacron^™-tipped swab stored in a glass bead-containing transport medium consisting gelatin (5.0 g/L), sucrose (68.46 g/L), glutamic acid (0.70 g/L), HEPES sodium salt (3.4 g/L), modified Hank’s balanced salts (9.8 g/L), sodium bicarbonate (0.35 g/L), bovine serum albumin (10 g/L) and the antimicrobial agents vancomycin (0.1 g/L), amphotericin B (2.5 mg/L), and colistin (0.015 g/L) at a pH of 7.2–7.8. The other set of swabs was collected at the post-MTZ (randomization) visit and all subsequent visits in the ESwab^Ⓡ Liquid Based Collection and Transport System (Copan ESwab 480C^Ⓡ). Samples were stored at room temperature for 1–4 hours after collection, then frozen at −80°C.

Microbiota sequencing, bacterial isolation, and cytokine analysis was performed on samples stored using the Starplex^™ system. Samples were thawed on ice, vortexed at maximum speed for 5 seconds, then the swabs were removed from the transport media and media was divided into aliquots and re-frozen at −80°C. Subsequent processing was performed as described below. Measurement of bacterial load via quantitative PCR (qPCR) was performed using swabs collected via the Copan ESwab^Ⓡ system.

Bacterial isolations and cultivation

Bacterial isolation and cultivation were performed under anaerobic conditions at 37°C in an AS-580 anaerobic chamber (Anaerobe Systems) with an atmosphere of 5% carbon dioxide, 5% hydrogen, and 90% nitrogen (Airgas, Inc.). All culture media was pre-reduced prior to use by being placed in the anaerobic chamber overnight. Bacteria were isolated and cultivated on solid media including Lactobacillus MRS agar (Hardy Diagnostics, #G117), Columbia Blood Agar (“CBA”, Hardy Diagnostics, #A16), or CDC Anaerobe Laked Sheep Blood Agar with Kanamycin and Vancomycin (“LKV”, BD BBL^™ Prepared Plated Media, #221846). Known L. crispatus strains and not-yet-identified isolates obtained on Lactobacillus MRS agar were expanded by culture in liquid media consisting of Lactobacillus MRS broth (BD #288130) prepared according to manufacturer instructions. Isolates obtained on CBA agar or LKV agar were expanded in liquid media consisting of either Wilkins-Chalgren Anaerobe Broth (Thermo Scientific^™ Oxoid^™, #CM0643B; prepared according to manufacturer instructions) or of NYCIII broth (American Type Culture Collection (ATCC) medium 1685), whichever produced better growth. NYCIII broth was prepared using a slightly modified version of the standard ATCC protocol³⁵. Pre-media consisted of 4 g/L HEPES (Fisher Scientific, #BP310–500), 15 g/L proteose peptone no. 3 (BD Biosciences, #BD 211693), and 5 g/L sodium chloride in 875 ml distilled water, which was pH-adjusted to 7.3 and autoclaved on liquid protocol at 121°C for 15 minutes, then cooled and stored at 4°C. One day before use, complete NYCIII broth was prepared from the autoclaved, cooled pre-media by adding dextrose (from a stock of 3 g per 45 ml; Fisher Chemical^™, #D16–500) at 7.5% v/v, yeast extract solution (Gibco, #18180–059) at 2.5% v/v, and heat-inactivated horse serum (Gibco, #26050070) at 10% v/v, then sterilized by passage through a 0.22 μm vacuum filter.

To establish a complete genome sequence for the CTV-05 strain, a cryopreserved pure culture of CTV-05 was obtained from Osel, Inc. The strain was streaked for isolation on Lactobacillus MRS agar and cultured for 48 hours, then a single colony was picked into Lactobacillus MRS broth and incubated for 20 hours. The broth culture was harvested by centrifugation to obtain bacterial pellets for genomic DNA extraction and sequencing.

Bacterial isolations were performed from twelve selected trial samples (see Figure 3C, S2A) using a modification of previously described methods³⁵. Since samples had been collected into Starplex^™ transport medium containing antibiotics (see above), the samples were thawed on ice, immediately diluted into pre-reduced Dulbecco’s Phosphate Buffered Saline (“PBS”, Millipore Sigma, #D8537) at 1:12 v/v, centrifuged for 10 minutes at 10,000 rcf, then supernatant was removed. Pellets were re-suspended in PBS and re-centrifuged with removal of supernatant two more times, then resuspended in PBS, diluted in serial 10-fold dilutions, and 100 𝜇l aliquots from each dilution were plated evenly on MRS, LKV, and CBA agar in parallel. Plates were incubated for 7 days and multiple examples of each distinct colony morphology from each sample were picked and subcultured onto solid media of the same type as the source media, with an emphasis on colonies from MRS agar with characteristic L. crispatus morphology (e.g., Figure 3D). After sub-culture for 3–7 days, colonies of the sub-cultured bacteria were picked into MRS broth (for colonies isolated on MRS agar) or into both NYCIII broth and Wilkens-Chalgren broth in parallel and incubated for 1–4 days, depending on growth rate. The resulting liquid cultures were then cryopreserved, with aliquots of each culture centrifuged and pellets saved for genomic DNA extraction and sequencing (see below).

METHOD DETAILS

DNA extraction for short-read sequencing

Total nucleic acids (TNA) extraction from cervicovaginal swabs for microbiota profiling was performed via a phenol-chloroform method, which includes a previously described bead beating process to disrupt bacteria⁷⁰ and modified for processing in 96-well plate format (Phenol:Chloroform:IAA, 25:24:1, pH 6.6, Invitrogen, #AM9730, which has since been discontinued; Sodium Dodecyl Sulfate 20% Solution, Fisher Scientific, #BP1311–200; EDTA, Invitrogen, #AM9260G; 2-Propanol, Sigma, #I9516–500ML; 3M Sodium Acetate, pH 5.5, Life Technologies, #AM9740). Aliquoted samples were thawed on ice, then TNA extraction was performed using 200uL of well-mixed Star media. The extracted TNA sample was eluted into 80uL of TE buffer (Promega, #V6321).

To extract bacterial genomic DNA (gDNA) for genome sequencing of cultured bacterial strains, each isolated strain was streaked on the indicated solid media and a single, clonal colony was picked into broth culture and incubated in static culture under anaerobic conditions for between 18 and 120 hours (depending on strain growth kinetics). Cultures were centrifuged and gDNA was extracted from the pellets using a plate-based protocol including a bead beating process and combining phenol-chloroform isolation⁸³ with QIAamp 96 DNA QIAcube HT kit (Qiagen, #51331) procedures.

Bacterial 16S rRNA gene sequencing

Bacterial microbiota taxonomic composition in cervicovaginal samples was determined by sequencing the V4 region of the bacterial 16S ribosomal RNA (rRNA) gene. The V4 region was amplified via polymerase chain reaction (PCR) using the primer set 515F/806R at 200 pM each (515F primer sequence 5’-AATGATACGGCGACCACCGAGACGTACGTACGGTGTGCCAGCMGCCGCGGTAA-3’ and barcoded 806R primer sequence 5’-CAAGCAGAAGACGGCATACGAGATxrefxrefAGTCAGTCAGCCGGACTACHVGGGTWTCTAAT-3’, in which the underlined sequences in each primer represent the regions of complementarity to 5’ and 3’ ends of the V4 region of the bacterial 16S rRNA gene, respectively, and the barcode positions in the 806R primer are indicated by X; IDT), with the 806R primers barcoded for multiplexing^35,84. PCR was performed in 25 µl reactions containing 1X Q5 reaction buffer (NEB, #B9027), 0.2 mM of dNTPs (NEB, #N0447), 0.2 μM of each primer, 0.5 unit of Q5 high-fidelity DNA polymerase (NEB, #M0491), and 2 µl of the TNA sample. PCR was performed in triplicate for each sample in the following program: 98°C for 30 seconds, followed by 30 cycles of 98°C for 10 seconds, 60°C for 30 seconds, and 72°C for 20 seconds, followed by a final extension at 72°C for 2 minutes. The triplicate PCR reactions for each sample were combined and amplicon production and size were confirmed on an agarose gel. Negative controls with PCR-quality water (Invitrogen, #10977) as a template were amplified in parallel for each primer barcode mix and assessed in parallel by gel electrophoresis to confirm absence of contamination and non-specific amplification. PCR products were pooled, with the amount for each sample semi-quantitatively adjusted based on its gel band intensity, then purified with a QIAquick PCR purification kit (Qiagen, #28104) and quality controlled with the Qubit^™ 4 Fluorometer (Invitrogen #Q33226), and TapeStation (Agilent Technologies, 4200 TapeStation). Libraries were mixed with 10% PhiX and single-end sequenced on an Illumina MiSeq using a 300-cycle v2 kit (Illumina, #MS-102–2002) employing the custom Earth Microbiome Project sequencing primers (Read 1 sequencing primer sequence: 5’-ACGTACGTACGGTGTGCCAGCMGCCGCGGTAA-3’; read 2 sequencing primer sequence: 5’-ACGTACGTACCCGGACTACHVGGGTWTCTAAT-3’; index sequencing primer sequence: 5’-ATTAGAWACCCBDGTAGTCCGGCTGACTGACT-3’; IDT)⁸⁴. Negative controls for TNA extractions and PCRs were included in each sequencing library. Study samples were sequenced in a total of six libraries. Samples with read counts < 10,000 in initial libraries were re-amplified and re-pooled into subsequent libraries and data from the run producing the highest number of reads for each sample (if the sample was sequenced multiple times) were selected for subsequent analysis. Analyzable 16S rRNA gene data was generated from a total of 1152 out of 1156 available trial samples, with the remaining 4 samples failing due to technical challenges with extraction, amplification, or sequencing (<10³ processed reads per sample).

Bacterial short-read genome sequencing

Shotgun metagenomic and genomic libraries were prepared following a modified protocol of Baym et. al⁸⁵, using the Nextera DNA Library Preparation Kit (Illumina, #20034211) and KAPA HiFi Library Amplification Kit (Kapa Biosystems, #KK2602). In brief, DNA from each sample was standardized to a concentration of 1ng/mL after quantification with SYBR Green (Invitrogen, #S7653), followed by simultaneous fragmentation and sequencing adaptor incorporation by mixing 1ng of DNA (1 mL) with 1.25 ml TD buffer and 0.25 mL TDE1 provided in the Nextera kit and incubating for 9min at 55°C. Tagmented DNA fragments were amplified in PCR using the KAPA high fidelity library amplification reagents, with Illumina adaptor sequences and sample barcodes incorporated in primers. PCR products were pooled, purified with magnetic beads (MagBio Genomics #AC-60050) and paired-end sequenced on Illumina NovaSeq X with a 300-cycle kit (Psomagen, Inc.).

Bacterial 16S rRNA gene sequence annotation

Demultiplexing of Illumina MiSeq bacterial 16S rRNA gene sequence data was performed using QIIME 1 version 1.9.188⁸⁶. Mapping files created in QIIME 1 format were validated using validate_mapping_file.py, then sequences were demultiplexed with split_libraries_fastq.py using parameter store_demultiplexed_fastq and no quality filtering or trimming, and demultiplexed sequences were organized into individual fastq files using split_sequence_file_on_sample_ids.py. Sequence reads were trimmed and filtered using dada2 version 1.6.0⁸⁷, trimming at positions 10 (left) and 230 (right) using the filterAndTrim function with truncQ = 11, MaxEE = 2, and MaxN = 0. Sequences were then inferred, then initial taxonomy assigned using the dada2 assignTaxonomy function with the RDP training database rdp_train_set_16.fa.gz (https://www.mothur.org/wiki/RDP_reference_files). Amplicon sequence variant (ASV) taxonomic assignments were refined via extensive manual review. Final ASV taxonomic assignments are included in Spreadsheet S1 for ease of review and comparison to other literature. The resulting annotated sequences were analyzed in R using phyloseq version 1.30.0⁸⁸ and custom R scripts (see STAR Methods). Sequence processing and taxonomy assignment was performed blinded to information about participants’ and samples’ clinical and demographic characteristics, treatments, and trial outcomes.

Total bacterial load via qPCR

Quantification of total bacterial load via qPCR was performed and reported as part of the initial LACTIN-V clinical trial³⁶. In brief, DNA extracted from samples stored in the Copan ESwab^Ⓡ system were amplified using bacterial 16S rRNA gene primers targeting total bacteria (16S ribosomal DNA, AGAGTTTGATCCTGGCTCAG, GCTGCCTCCCGTAGGAGT, 312bp). Bacterial concentration was calculated using a standard curve based on serial dilutions of the CTV-05 strain as previously described³⁷.

Assessing baseline balance between arms

Since pre-intervention (i.e., pre-MTZ or post-MTZ) microbiota composition may be associated with differential microbiota composition post-intervention, imbalances between arms could lead to biases in our primary outcome benefit ratio estimates. To assess whether significant imbalances in microbiota composition existed in this cohort, we performed a PERMANOVA analysis, as implemented in the vegan R package⁸⁹, to test whether the intervention arm (LBP vs placebo) explained significant variability in microbiota β-diversity, computed using the Bray-Curtis dissimilarity on ASV relative abundances.

Calculating benefit ratios

Benefit ratios, and associated confidence intervals and p-value were computed using Wald’s method as implemented in the epitools R package⁹⁰.

Identification of microbiota topics

Microbiota topics were identified using a modification of a previously described approach⁵⁰. One main difference is that here, Lactobacillus topics (i.e., topics composed exclusively of Lactobacillus species) were defined independently from non-Lactobacillus topics (i.e., topics composed exclusively of non-Lactobacillus species), which were identified by fitting a Latent Dirichlet Allocation model⁵³ to the non-Lactobacillus ASV counts aggregated at the species level. This was done to facilitate the interpretation of topic composition and, specifically, to allow for “pure” topics for the most prevalent Lactobacillus species in this cohort (L. crispatus, L. iners, and L. jensenii).

To determine K, the optimal number of non-Lactobacillus topics, we relied on a method called “topic alignment”⁹¹ which examines the robustness of topics across resolutions (increasing values of K) and provides diagnostics scores that facilitate the identification of spurious topics. We selected K to minimize the number of spurious topics (flagged by low coherence scores) and such that the number of paths (collection of similar topics across resolution) in the alignment presented a plateau⁹¹ (Figure S1D-E). The estimated proportions (relative abundances) of non-Lactobacillus topics in each sample (𝑝̂_ki where k is the topic and i is the sample) were computed by multiplying the proportions estimated by the model on non-Lactobacillus counts (𝜋̂_ki) by the total non-Lactobacillus proportions in each samples (𝛱̂_i = ∑_v∈V 𝑝_vi where 𝑝_vi is the observed proportion of ASV 𝑣 in sample 𝑖 and 𝑉 is the set of non-Lactobacillus ASVs, such that 𝑝̂_ki = 𝜋̂_ki 𝛱̂_i).

Lactobacillus topics were defined as follows: Lactobacillus species which reached 50% of a microbiota composition in at least 10 samples made up their own topic, while the remaining Lactobacillus species were grouped into a single topic (“Other L.”) as their total prevalence and abundance was overall small (Figure S1F). The composition of this topic was estimated from the species average prevalences in this cohort. The estimation of proportions of Lactobacillus topics in each sample was straightforward: they were computed from the proportions of the corresponding species in each sample.

Isolate genome assemblies from short-reads

Paired-end short-read genomic sequencing data from bacterial isolates were processed through a quality control and assembly pipeline. Raw reads were first trimmed for adapter sequences using Cutadapt v5.2⁹² with the Nextera adapter sequence (CTGTCTCTTAT). Quality filtering was then performed using Sickle-trim v1.33⁹³ with a quality threshold of Q20 and minimum read length of 50 bp after trimming. Read quality was assessed using FastQC⁹⁴. De novo genome assembly was performed using Unicycler v0.5.0 in standard mode with default parameters⁹⁵. Unicycler produces high-quality assemblies from short reads alone by optimizing SPAdes assembly, followed by graph simplification. The resulting assemblies were annotated using Bakta v1.9.4⁹⁶ with the Bakta database v5⁹⁷, which provides comprehensive bacterial genome annotation including coding sequences, rRNAs, tRNAs, and other genomic features. Assembly completeness and quality were evaluated using BUSCO v5.5⁹⁸ with the Bacteroidales lineage dataset (bacteroidales_odb10), providing a measure of genome completeness based on conserved single-copy orthologs. Isolate genome-assembly statistics and NCBI BioSample numbers are in Spreadsheet S3; the median N50 was >23,000, median completeness >99.8%, and median contamination <0.36%.

CTV-05 sequencing and genome assembly

For long-read (Oxford Nanopore Technologies) genome sequencing of the CTV-05 strain, a bacterial pellet was prepared from a clonal colony cultured in Lactobacillus MRS broth media as described above. gDNA was extracted using a non-bead-beating protocol to minimize shearing of DNA fragments, with extractions performed and sequenced using five parallel aliquots of the same original broth culture to maximize yield and consistency. Since L. crispatus is known to have a thick cell wall containing surface-layer (S-layer) proteins, cell pellets were resuspended with 5M lithium chloride (Molecular Dimensions #MD2–100-43) and then washed with PBS to begin to degrade the cell wall. Using the manufacturer’s protocol for the MasterPureTM Gram Positive DNA Purification Kit (Biosearch Technologies #MGP04100), the bacterial cells were then lysed, proteins and RNA were digested, and genomic DNA (gDNA) was extracted. Extracted gDNA concentration was diluted 1:5 and measured using the Qubit^™ 4 Fluorometer, then DNA libraries were prepared using Oxford Nanopore Sequencing’s Rapid Barcoding Kit (Oxford Nanopore Technologies #SQK-RBK004; note that this product has been discontinued as of March 2024). In brief, 400 ng of gDNA from five replicate extractions was barcoded, pooled, purified, and then loaded onto the MinION Flow Cell using R9.4.1 chemistry (Oxford Nanopore Technologies #FLO-MIN106D; note that this product has been discontinued as of July 2024) in the MinION Sequencing Device (Oxford Nanopore Technologies #MIN-101B). To produce the maximum number of reads possible using the Rapid Barcoding Kit, the sequencing run was continued for a full 72 hours.

Demultiplexed genomic long-reads from MinKNOW v2.2 were concatenated, and reads were separately filtered for quality and length using Filtlong v0.2.1 (https://github.com/rrwick/Filtlong). Filtered reads were subsampled to make 12 different read sets using Trycycler v0.5.3⁹⁹ and duplicate reads from each read set were removed using BBMap v39.00¹⁰⁰. The read sets were assembled using Flye v2.9.1¹⁰¹, Minipolish v0.1.3¹⁰², Raven v1.8.1¹⁰³, and Canu v2.2¹⁰⁴. Contigs from different assembly methods were clustered based on similarity to generate consensus sequences and improve assembly quality using Trycycler. Clusters were manually chosen and Trycycler was used to reconcile the contigs within each cluster. Multiple sequence alignment and read partitioning were performed on the reconciled contigs using Trycycler to understand sequence variation within each cluster and assign each read to the cluster it best aligned with respectively. Trycycler was then used to generate a consensus contig sequence for each cluster based on the multiple sequence alignment and read partitioning steps. Medaka v1.7.2 (https://github.com/nanoporetech/medaka), a tool for creating consensus sequences specifically from nanopore sequence data, was used to polish the Trycycler consensus sequences. The consensus sequences were then combined, after which Polypolish v0.5.0¹⁰⁵ was used to polish the long-read sequence with Illumina short-read data generated from an aliquot of the same clonal culture of CTV-05 (extracted and sequenced as described above). The complete CTV-05 genome was annotated using Bakta v1.7.0⁹⁶ and published online as part of this manuscript under NCBI BioSample reference number SAMN55039759, which includes the unassembled short- and long-reads described above as well as the complete CTV-05 genome assembly. BioSample reference number SAMN55039759 is included under the study’s overall BioProject reference number: PRJNA1303956.

L. crispatus SNV database for strain analysis

A GT-Pro database of biallelic core sites was constructed using 236 L. crispatus genomes using k-mers with thresholds for SNV prevalence of 0.95 and minor allele frequencies of 0.02 with MAAST^55,106,107. Then k-mers for samples were counted to generate a metagenotype, or a matrix of the number of reads at biallelic SNV sites for a single species⁵⁴. This matrix was then used by StrainFacts to infer suitable allele and relative abundance matrices that represent strains across all given samples⁵⁵.

L. crispatus strain inference in metagenomes

Raw metagenomic paired-end reads from vaginal swab samples were quality-trimmed and adapter-filtered using Cutadapt v3.9⁹². Human reads were depleted by aligning reads to the T2T-CHM13 reference genome using HISAT2 v2.2.1¹⁰⁸. Samples were sequenced to a median depth of 22,329,925 (IQR: 14,928,888 – 33,699,201) reads, yielding a median of 2,748,850 (IQR: 1,186,960 – 7,950,264) analyzable microbial reads after host and quality filtering for a median of 12.7% (IQR: 4.53 – 25.9%) of reads retained after processing (sample-level read depths pre- and post-processing, as well as NCBI BioSample numbers, are provided in Spreadsheet S2). Processed reads were profiled using GT-Pro (v1.0.1) against our custom SNV database consisting of ~26,000 biallelic SNV sites¹⁰⁷. Resulting metagenotype profiles were analyzed with StrainFacts (v0.6.0) to infer strain genotypes and abundances for L. crispatus, excluding samples with <25% coverage of the biallelic SNVs⁵⁴. The fit was then run through a cleanup step to remove low abundance strains and strains with extremely similar genotypes using sfacts cleanup_fit --abundance 0.01 --dissimilarity 0.01 --discretized. Analysis was performed on samples with ≥5% overall relative abundance of L. crispatus as determined by 16S rRNA gene sequencing based on results of a separate validation study evaluating parameters for accurate StrainFacts LBP strain inference⁵⁵. Since most samples had L. crispatus relative abundance either substantially higher or lower than 5%, with only 31 (<3%) exhibiting relative abundance between 2.5–5% (e.g., Figure 1D), decreasing this threshold would likely have increased error while negligibly increasing the number of samples selected for analysis. L. crispatus strain inference was successful in 313 of 343 samples in which species relative abundance exceeded the analysis threshold, while strain analysis was unsuccessful in the remaining 30 samples due to failed or inadequately deep shotgun metagenomic sequencing. A 10% strain proportional abundance was used as a threshold for confident strain identification based on prior simulations and experimental validation⁵⁵. Most samples had L. crispatus fractional strain abundances either substantially higher or lower than 10%, with only 12 of 313 samples (3.8%) containing inferred strains with fractional abundance between 7.5% and 12.5%, showing strain detection was robust to threshold choice.

Strain analysis of isolate genomes

Isolate genomes were taxonomically identified using GTDB-Tk v2¹⁰⁹. Isolate taxonomic details are in Spreadsheet S3. Gene-calling was performed for L. crispatus isolate genomes, including the completed CTV-05 genome (see above) using Prodigal¹¹⁰. Core ribosomal proteins were identified using HMMER¹¹¹, aligned using mafft^112,113, then concatenated and a phylogenetic tree was constructed using RAxML-NG¹¹⁴, which was used to compute branch lengths to identify isolates representing distinct strains and determine which isolates represented CTV-05 or endogenous strains. The tree was rooted to L. jensenii. GT-Pro¹⁰⁷ was then used to calculate genotypes from representative isolate reads for comparison to metagenomically inferred strain genotypes, calculated based on Jaccard similarity of the StrainFacts genotypes⁵⁵.

Mucosal cytokine and chemokine measurement

Cytokines and chemokines were measured using a previously described custom 20-plex High Sensitivity Luminex Assay Kit from EMD Millipore that measures interferon gamma-induced protein (IP-10), interleukin 8 (IL-8), interleukin 6 (IL-6), monokine induced by interferon gamma (MIG), interferon-inducible T cell alpha chemoattractant (ITAC), interleukin 1 alpha (IL-1α), interleukin 1 beta (IL-1β), macrophage inflammatory protein-1 alpha (MIP-1α), macrophage inflammatory protein-1 beta (MIP-1β), macrophage inflammatory protein-3 alpha (MIP-3α), tumour necrosis factor alpha (TNFα), interleukin 21 (IL-21), interleukin 17 (IL-17), interferon gamma (IFNγ), interleukin 23 (IL-23), interleukin 12 (IL-12 p70), interleukin 13 (IL-13), interleukin 10 (IL-10), interleukin 4 (IL-4), and interleukin 5 (IL-5)⁶. Reagents were used as supplied by the manufacturer, except the Mixed Beads solution, detection antibody mixture, and Streptavidin-phycoerythrin (Strep-PE) were diluted 3-fold (1:2 volume:volume ratio) for use. The Mixed Beads were diluted in Bead Diluent, while the detection antibodies and Strep-PE were diluted individually in Assay Buffer. All reagents were from a single manufacturer lot to minimize batch effects.

Samples were first pre-processed to remove interfering mucus and debris. In brief, 200 𝜇L aliquots of each vaginal swab supernatant were thawed on ice, vortexed for 10 seconds to resuspend, then centrifuged at 1000 rcf for 15 minutes at 4°C to pellet mucous and cells. The maximum amount of supernatant recoverable from each sample without disturbing the resulting pellet was then transferred to a well of a 0.22 m PVDF filter plate (EMD Millipore #MSGVS2210), filtered by centrifugation at 2,451 rcf for 1 hour at 4°C, then transferred to freezer-safe tubes and stored at −80°C until assayed. Samples were assayed in 96-well plates as per manufacturer protocol with the addition of periodic sonication steps to minimize bead clumping. Each plate assay included one blank background control, 7 standard serial dilutions assayed in duplicate, two manufacturer-supplied Quality Control (QC) samples assayed in duplicate, 72 experimental samples, and 4 biological control samples used as internal quality controls on all plates. Batches of standards and QC samples were prepared from the kit on the day of the assay according to manufacturer protocol. For each sample or standard, 25 𝜇L each of Assay Buffer, Mixed Beads, and sample were aliquoted into each plate well. After the addition of Mixed Beads, all incubation steps were performed while the plate was protected from light. Plates were sonicated for 30 seconds at room temperature in a bath sonicator, then incubated for 16–18 hours on a horizontal plate shaker at 600 rpm in 4°C. An additional 30 second sonication at room temperature was then performed. A handheld plate magnet was applied to the base of the plate to retain the magnetic beads and excess sample and reagents were removed via three consecutive washes using manufacturer-supplied 1X Wash Buffer. Detection antibodies were added as per manufacturer protocol and plates were incubated for 1 hour at room temperature on a horizontal plate shaker at 600 rpm. Next, Strep-PE reagent was added, followed by a 30 minute incubation at room temperature on a horizontal plate shaker at 600 rpm. Plates were washed 3 times as above using the magnet. After the final wash, 150 𝜇L of sheath fluid was added to all wells and a final 30 second sonication at room temperature was performed. Analyte concentrations were measured using a Luminex^® FLEXMAP 3D instrument with Luminex xPONENT^® software (Diasorin).

Cytokine concentration transformations

Cytokine and chemokine concentrations were log-transformed (log) to stabilize the mean-variance relationship. Compounds with concentration values below the lower limit of quantification (LLOQ) were imputed at half the LLOQ; those with concentration values above the upper limit of quantification (ULOQ) were imputed at the ULOQ (Figure S3A).

Measurement of soluble molecules in swab samples can be complicated by a “size effect” phenomenon whereby a component of between-swab variation in concentrations is due to technical differences in the amount of material collected on each swab. In such situations, the first principal component (PC1), which reflects the “size” of observations ⁵⁶, can primarily be driven by the amount of material on the swab rather than biological variation. Consistent with this size effect phenomenon, we observed that per-sample cytokine/chemokine concentrations were highly collinear, with PC1 of the standardized log-transformed cytokine concentrations accounting for >50% of the total variance (Figure S3B) but showed no substantial correlation with clinical and biological factors such as proportions of total Lactobacillus or of Lactobacillus crispatus, or participants’ contraceptives. Such size effects can be addressed by subtracting PC1 from the data. We therefore performed PC1 subtraction by re-assigning scores corresponding to the 1st PC to a value of 0, then transforming the data back to its original variable space using the transposed rotation matrix. Sensitivity analyses employing unadjusted concentrations revealed qualitatively similar findings to results obtained using the PC1-subtracted data, but with reduced magnitude. Finally, due to the large uncertainty regarding their distributions, cytokines with values below the LLOQ in ≥60% of the samples or with values above the ULOQ in ≥30% of the samples were excluded from the analysis (Figure S3A).

Association of microbiota and cytokines

To quantify the overall association between microbiota composition and cytokine profiles, we computed the RV coefficient between the microbiota composition as expressed as topic proportions and cytokine transformed log₁₀-concentrations. Associated p-value was computed with a permutation test^115,116. To confirm that correlation between these tables was not driven by potential subject (longitudinal) effects, we also computed correlation between the two tables at each visit independently. If correlations were found significant, further analyses were carried out to characterize the relationship between tables. Specifically, DISTATIS^59,117, as implemented in the distatisR R package, was used. This method presents the advantage of estimating compromise scores (and partial residual scores) directly from dissimilarity matrices such that ecological measures of (dis)similarity, such as the Bray-Curtis dissimilarity, can be used for representing microbiota composition. Associations with the original variables can then be assessed using correlations and displayed in correlation circles as typically done with PCA, PCoA, or NMDS results. The minimum number of components used for display was chosen based on the presence of an elbow in the DiSTATIS scree plot.

Estimating effect heterogeneity

Binary outcomes (Y: ≥50% colonization by L. crispatus at week 12 or 24 or absence of rBV by week 12 or 24) were predicted using logistic regression (logit link function) with input variables representing the intervention arm (A: LBP or placebo), the baseline (pre-MTZ) microbiota (V: stratified by most prevalent genus, CST, or topic proportions; Figure 5C-F and S4) or self-declared racial group (Figure 6E-F), and their interactions (A:V). In stratification analyses, strata with less than two participants per arm were excluded from the analysis and visualizations. Since Lactobacillus species other than L. iners had low pre-MTZ relative abundance, prior to determining fits relying on topic proportions, Lactobacillus-dominated topics were agglomerated such that the proportion of total Lactobacillus was considered, which improved fit convergence. The statistical significance of quantitative effect heterogeneity was tested by comparing the null model (which only includes the intervention arm as predictor) with the full model where pre-MTZ microbiota (or self-declared racial group) and interaction effects were included (analysis of deviance F test). P-values were adjusted to account for multiple hypotheses testing using the Benjamini-Hochberg procedure to control the false discovery rate¹¹⁸. For the stratified analyses, CI were computed using Wilson scores due to the small sample size in each stratum. When pre-MTZ microbiota composition or self-declared racial group was included as a quantitative multivariate variable, counterfactual probabilities of success and associated 95% confidence intervals were predicted from the logistic regression fitted model by setting the intervention arm to LBP or placebo. The predicted participant-level odd ratios were computed as the ratio between predicted odds of L. crispatus colonization at week 12 and 24 or rBV by week 12 or 24 had participants been receiving LBP or the placebo.

Factors predictive of microbiota category

To identify demographic, clinical, behavioral, or microbiologic factors associated with successful L. crispatus colonization (≥ 50%) in the intervention arm, we relied on a multiblock partial least square discriminant analysis (MB-PLS-DA)⁶² where the 3-category response variable indicated whether (1) relative abundance of L. crispatus was ≥50%, (2) relative abundance of total Lactobacillus (any species) was ≥50% but relative abundance of L. crispatus was <50%, or (3) the relative abundance of total Lactobacillus was <50%. MB-PLS-DA relies on the same principles as PLS-DA but allows for explanatory variables to be grouped into several thematic blocks, which enables the computation of block importance indices and covariances with the response block^61,62. Our analysis grouped explanatory variables into 13 blocks (Figure 6A, Spreadsheet S4). The first block characterized participant demographics, blocks 2–4 described participant baseline vaginal ecosystem (i.e., pre-MTZ microbiota composition and diversity, pH, cytokines), blocks 5–8 quantified the vaginal ecosystem at the previous visit (e.g., at the post-MTZ visit for the initial phase), while blocks 9–13 characterized sexual behavior, douching/bleeding, antibiotic use, and product adherence (Spreadsheet S4).

Several explanatory variables were correlated. For example, the abundances of some cytokines correlated with microbiota composition (Figure 4), and the baseline α-diversity was lower in White participants (Figure S5F, p-value < 0.05). We addressed these existing correlations between explanatory blocks and/or variables in two different ways depending on our assumptions on the underlying correlation source or cause. We either relied on nested models to evaluate the additive predictive power of specific blocks (e.g., demographics) that had variables correlated with microbiological blocks or used one variable or one block to predict the values of another one and included the residuals instead of the observed values in the model. This was indicated by the symbol (r) in the variable or block names. We used this approach for the cytokine blocks whose residuals were computed using their PLS-predicted abundances based on microbiota composition. Similarly, we computed residual microbiota composition, α-diversity, and pH at the previous visit such that topic proportions, α-diversity, or pH were relative to those expected based on the participants’ colonization status at the same visit.

Categorical variables such as self-declared race or birth control were included using one-hot encoding, and variables were standardized. To ensure convergence of the fits in cross-validation or using the bootstrap, we added Gaussian noise with very small variance (𝜎_i² = 10 ⁻³𝑆_i² where 𝜎_i² is the variance of the Gaussian noise for variable 𝑖, and 𝑆_i² the empirical variance of the 𝑖^𝑡ℎ variable) to one-hot encoded variables that had few participants in some categories.

Given that our categorical response had three categories, two latent components for the MB-PLS-DA models were a natural choice to avoid overfitting. Further, we performed cross-validation analyses (40 random 75–25% split into calibration and validation sets) and found that two latent components maximized the mean average F1 score on the validation sets for all models except for the initial phase of the placebo in which no variables were found to predict our response, leading to poor and unstable performances in cross-validation (Figure S5B and S6A). This was consistent with the presence of an elbow in the screeplots and robust to other choices of metrics such as accuracy or root mean square error (RMSE). Overall, results on calibration and validation sets indicated initial-phase colonization was more difficult to explain (average validation F1 score: 0.46, 95%CI 0.42–0.49) and explained by more factors (best validation F1 score obtained for the full model) than colonization at later visits (best average validation F1 obtained for the minimal models of 0.64; 95%CI 0.62–0.66 for continuation phase and 0.47; 95%CI 0.46–0.49 follow-up phase, Figure 6A, S5BC). Directions of associations between explanatory variables and outcomes were consistent across nested models. Relative cumulative block importance indices were computed by dividing cumulative block importance, the BIPC as in Brandolini-Bunlon et al.⁶², by the total inertia of each block, which corresponds to the expected BIPC if all variables had a similar importance; 95% CIs were calculated using a bootstrap approach.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical Analysis

Statistical analyses are described in figure legends as appropriate and details for individual analyses are provided in the corresponding sections or the STAR methods describing the relevant bioinformatic and experimental Method Details.

Software

R packages used in the analyses included SummarizedExperiment¹¹⁹, MultiAssayExperiment¹²⁰, tidyverse¹²¹, janitor¹²², gt¹²³, alto¹²⁴, topicmodels¹²⁵, gtsummary¹²⁶, ade4¹¹⁵, factoextra¹²⁷, DistatisR¹²⁸, PropCIs¹²⁹, pls¹³⁰, packMBPLSDA¹³¹ (see also Key Resource Table and publicly available code).

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Bacterial and virus strains
Lactobacillus crispatus strain CTV-05	Osel, Inc.	CTV-05
Vaginal bacterial primary isolates	This study	See Spreadsheet S3
Biological samples
Human vaginal swab samples	Cohen et al.37	N/A
Chemicals, peptides, and recombinant proteins
Lactobacillus MRS agar	Hardy Diagnostics	#G117
Columbia Blood Agar (“CBA”)	Hardy Diagnostics	#A16
CDC Anaerobe Laked Sheep Blood Agar with Kanamycin and Vancomycin (“LKV”)	BD BBL™ Prepared Plated Media	#221846
Lactobacillus MRS broth	BD	#288130
Wilkins-Chalgren Anaerobe Broth	Thermo Scientific™ Oxoid™	#CM0643B
HEPES	Fisher Scientific	#BP310–500
Proteose peptone no. 3	BD Biosciences	#BD 211693
Dextrose	Fisher Chemical™	#D16–500
Yeast extract solution	Gibco	#18180–059
Heat-inactivated horse serum	Gibco	#26050070
Dulbecco’s Phosphate Buffered Saline (“PBS”)	Millipore Sigma	#D8537
Phenol:Chloroform:IAA, 25:24:1, pH 6.6	Invitrogen	#AM9730 (since discontinued)
Sodium Dodecyl Sulfate 20% Solution	Fisher Scientific	#BP1311–200
EDTA	Invitrogen	#AM9260G
2-Propanol	Sigma	#I9516–500ML
Sodium Acetate, pH 5.5	Life Technologies	#AM9740
TE Buffer	Promega	#V6321
QIAamp 96 DNA QIAcube HT kit	Qiagen	#51331
Q5 reaction buffer	NEB	#B9027
Ultrapure dNTP mix	NEB	#N0447
Q5 high-fidelity DNA polymerase	NEB	#M0491
Invitrogen UltraPure Water - 0.1-μm Membrane-Filtered Water	Invitrogen	#10977
SYBR™ Green I Nucleic Acid Gel Stain, 10,000X concentrate in DMSO	Invitrogen	#S7653
HighPrep PCR magnetic beads	MagBio Genomics	#AC-60050
5M Lithium Chloride	Molecular Dimensions	#MD2–100-43
Multiscreen® 96 well Plate, hydrophilic PVDF membrane (pore size 0.22 μm, sterile)	EMD Millipore	#MSGVS2210
Critical commercial assays
Starplex™ Scientific Multitrans™ Collection and Transportation System	Starplex™ Scientific	S1600
ESwabⓇ Liquid Based Collection and Transport System	Copan	ESwab 480CⓇ
QIAquick PCR purification kit	Qiagen	#28104
MiSeq Reagent Kit v2, 300 Cycles (PE)	Illumina	#MS-102–2002
Nextera DNA Library Preparation Kit	Illumina	#20034211
KAPA HiFi Library Amplification Kit	Kapa Biosystems	#KK2602
MasterPure Gram Positive DNA Purification Kit	Biosearch Technologies	#MGP04100
Oxford Nanopore Sequencing Rapid Barcoding Kit	Oxford Nanopore Technologies	#SQK-RBK004 (since discontinued)
MinION Flow Cell using R9.4.1 chemistry	Oxford Nanopore Technologies	#FLO-MIN106D (since discontinued)
MinION Sequencing Device	Oxford Nanopore Technologies	#MIN-101B
Custom 20-plex High Sensitivity Luminex Assay Kit	EMD Millipore	Gosmann et al., 2017
Deposited data
Metagenomic sequence data, bacterial 16S rRNA gene sequence data, isolate shotgun genomic read sequence data, and isolate genome assemblies.	NCBI	BioProject PRJNA1303956
CTV-05 closed genome and the Illumina short-read shotgun genomic sequences and Oxford Nanopore long-read shotgun genomic sequences from which it was assembled	NCBI	BioSample SAMN55039759 (part of BioProject PRJNA1303956, above)
Clinical data and processed and transformed omics data	Zenodo	DOI 10.5281/zenodo.17755173
Analysis code	Zenodo	DOI 10.5281/zenodo.18524760
Experimental models: Cell lines
Experimental models: Organisms/strains
Oligonucleotides
Bacterial 16S rRNA gene 515F primer AATGATACGGCGACCACCGAGACGTACGTACGGTGTGCCAGCMGCCGCGGTAA (underlined sequence represents the region of complementarity to the bacterial gene)	Caporaso et al.84; primer manufactured by IDT	N/A
Bacterial 16S rRNA gene 806R primer CAAGCAGAAGACGGCATACGAGATxrefxrefAGTCAGTCAGCCGGACTACHVGGGTWTCTAAT (underlined sequence represents the region of complementarity to the bacterial gene; barcode positions in the primer are indicated by X)	Caporaso et al.84; primer manufactured by IDT	N/A
Custom Earth Microbiome Project Read 1 sequencing primer: ACGTACGTACGGTGTGCCAGCMGCCGCGGTAA	Caporaso et al.84; primer manufactured by IDT	N/A
Custom Earth Microbiome Project Read 2 sequencing primer: ACGTACGTACCCGGACTACHVGGGTWTCTAAT	Caporaso et al.84; primer manufactured by IDT	N/A
Custom Earth Microbiome Project index sequencing primer: ATTAGAWACCCBDGTAGTCCGGCTGA CTGACT	Caporaso et al.84; primer manufactured by IDT	N/A
Total bacterial 16S qPCR primer, forward: AGAGTTTGATCCTGGCTCAG	Cohen et al.37	N/A
Total bacterial 16S qPCR primer, reverse: GCTGCCTCCCGTAGGAGT	Cohen et al.37	N/A
Recombinant DNA
Software and algorithms
QIIME 1 version 1.9.188	Caporaso et al.86	N/A
dada2 version 1.6.0 (R package)	Callahan et al.87	N/A
Ribosomal Database Project (RDP) training database rdp_train_set_16.fa.gz	https://www.mothur.org/wiki/RDP_reference_files	N/A
phyloseq (R package)	McMurdie and Holmes88	N/A
vegan (R package)	https://vegandevs.github.io/vegan/	N/A
epitools (R package)	Aragon90	N/A
GT-Pro v1.0.1	Shi et al.107	N/A
MAAST	Shi et al.106	N/A
StrainFacts	Smith et al.54	N/A
Bakta v1.9.4	Schwengers et al.96	N/A
MinKNOW v2.2	Oxford Nanopore Technologies	N/A
Luminex xPONENT software	Diasorin	N/A
Filtlong v0.2.1	https://github.com/rrwick/Filtlong	N/A
Trycycler v0.5.3	Wick et al.99	N/A
BBMap v39.00	Bushnell100	N/A
Flye v2.9.1	Kolmogorov et al.101	N/A
Minipolish v0.1.3	Wick and Holt102	N/A
Raven v1.8.1	Vaser and Šikić103	N/A
Canu v2.2	Koren et al.104	N/A
Medaka v1.7.2	https://github.com/nanoporetech/medaka	N/A
Polypolish v0.5.0	Wick and Holt105	N/A
Cutadapt v5.2	Martin92	N/A
HISAT2 v2.2.1	Kim et al.108	N/A
T2T-CHM13	National Center for Biotechnology Information (NCBI)	BioProject PRJNA559484; RefSeq assembly GCF_009914755.1
GTDB-Tk v2	Chaumeil et al.109	N/A
Prodigal v3.5.0	Hyatt et al.110	N/A
HMMER v3.3.2	Eddy111	N/A
mafft v7.490	Katoh et al.112; Kuraku et al.113	N/A
RAxML-ng v1.1.0	Kozlov et al.114
SummarizedExperiment (R Bioconductor package)	Morgan et al.119	N/A
MultiAssayExperiment (R Bioconductor package)	Ramos et al.120	N/A
tidyverse (R package suite)	Wickham et al.121	N/A
janitor (R package)	Firke122	N/A
gt (R package)	Iannone123	N/A
alto (R package)	Fukuyama et al., 2024124 https://github.com/lasy/alto	N/A
topicmodels (R package)	Grün et al125	N/A
gtsummary (R package)	Sjoberg et al.126	N/A
ade4 (R package)	Dray et al.115	N/A
factoextra (R package)	Kassambara et al.127	N/A
DistatisR (R package)	Beaton et al.128	N/A
PropCIs (R package)	Scherer et al.129	N/A
pls (R package)	Liland et al.130	N/A
packMBPLSDA (R package)	Brandolini-Bunlon et al.131	N/A
Sickle v1.33	Joshi and Fass93	N/A
FastQC v0.12.1	Andrews94	N/A
Unicycler v0.5.1	Wick et al.95	N/A
Bakta database v5.1	Schwengers97	N/A
BUSCO v5.5	Manni98	N/A
Other

Supplementary Material

1

Document S1. Figures S1-S7, Tables S1-S3, and supplemental references

2

Spreadsheet S1. Bacterial 16S ribosomal RNA (rRNA) gene v4 region amplicon sequence variant taxonomic assignments used in this study, related to Figure 1

3

Spreadsheet S2. Vaginal shotgun metagenomic whole genome sequencing (mWGS) sample NCBI BioSample accession numbers, read depths, and quality statistics before and after filtering, related to Figure 3

4

Spreadsheet S3. Vaginal bacterial isolate genome NCBI BioSample accession numbers, assembly quality metrics, and taxonomy as assigned by GTDB-tk, related to Figure 3

5

Spreadsheet S4. Unadjusted associations between explanatory variables by thematic block and rBV by weeks 12 and 24 or L. crispatus dominance at weeks 12 or 24, related to Figure 6

Highlights.

• LACTIN-V recipients achieve higher rates of Lactobacillus crispatus dominance

• The LBP strain predominates, but native strains increasingly replace it over time

• LBP efficacy differs in women with different pre-treatment vaginal microbiota

• L. crispatus dominance in LBP arm varies by biological and sociodemographic factors