Early pregnancy is characterized by a significant decrease in the diversity of the oral microbiome and strong associations with lifestyle and conception method

Several reports have demonstrated that the oral microbiome undergoes distinct compositional changes throughout pregnancy, influenced by hormonal, immunologic, and metabolic shifts [ 1–3 ]. However, findings remain inconsistent across studies. While some studies report relatively stable overall microbial diversity, others observe a trend towards decreased microbial richness and shifts in community structure, particularly in the second (T2) and third (T3) trimesters [ 2 , 4 , 5 ]. Studies using 16S rRNA gene sequencing have shown significant increases in certain taxa in T2 compared to T1, such as the phylum Bacteroidota, and the genera Veillonella (notably Veillonella dispar and Veillonella atypica ) and Granulicatella , with a concurrent decrease in Streptococcus sanguinis and Selenomonas [ 2 , 5 , 6 ]. T3 is associated with a further reduction in microbial richness and increased dominance of specific genera, including Bacteroidota (notably Prevotella spp.), Veillonella spp. (especially V. dispar and V. atypica ), and S. mutans [ 2 ], suggesting a shift toward a more distinct, potentially pro-inflammatory community [ 7 , 8 ]. These changes occur in parallel with an increase in sex hormone levels [ 9 ]. The oral microbiome tends to revert to a pre-pregnancy state within weeks postpartum [ 4 ]. Importantly, while the oral microbiome is more stable than the vaginal or gut microbiome [ 10 ], pregnancy-associated changes have been associated with oral diseases and adverse pregnancy outcomes [ 11 ]. Although the evolution of the oral microbiome in T2 and T3 has been extensively studied, to our knowledge, there are currently no large-scale studies based on next generation sequencing that look at the evolution of the oral microbiome from T1 of pregnancy onwards, nor any clear evidence of a change in the oral microbiome during this period (see table 2 in [ 4 ] and [ 12 , 13 ]). Diet is a central determinant of oral microbial ecology, shaping both community structure and function [ 14 , 15 ]. Frequent intake of fermentable carbohydrates, particularly sugars, enriches acidogenic and aciduric taxa such as Streptococcus mutans and Lactobacillus , promoting dysbiosis and cariogenesis [ 16 , 17 ]. Conversely, diets rich in fiber, plant polyphenols, and micronutrients, such as vitamin D and calcium, are associated with greater oral microbial diversity, resilience, and anti-inflammatory properties [ 18 , 19 ]. In addition, nutritional deficiencies, including iron and folate, can also alter microbial colonization [ 20 ]. Thus, dietary patterns interact with host physiology to create either a protective or disease-prone oral environment. During pregnancy, nutritional demands increase to support maternal health and fetal development, and dietary habits often change due to nausea, cravings, or medical recommendations. These shifts can exacerbate pregnancy-related changes in the oral microbiome. For instance, increased snacking and carbohydrate intake during T1 may favor cariogenic bacteria, while micronutrient supplementation (e.g. folic acid, iron) may mitigate dysbiosis [ 21 , 22 ]. Restrictive diets such as gluten-free regimens, although necessary for individuals with celiac disease, can reduce dietary fiber and prebiotic intake, potentially leading to decreased microbial diversity and enrichment of opportunistic taxa in the oral cavity [ 23 ]. Together, these findings suggest that nutrition is a key, though underexplored, modifier of pregnancy-associated oral microbiome changes. Here, we characterized the oral microbiome at the end of T1, T2, and T3 in 346 Israeli women [ 24 ], aiming to elucidate how microbial composition and dynamics change across pregnancy stages in relation to maternal demographics, lifestyle, and dietary patterns. To assess the reproducibility of these findings, we validated our results in an independent cohort of 154 Russian women sampled during T2 and T3 [ 25 ].

To analyze the development of the oral microbiome during pregnancy, we recruited an Israeli cohort of pregnant women sampled during T1, T2, and T3. A total of 346 Israeli women were recruited, and 467 oral microbiome samples were processed: 235 from T1 (11–14 gestational weeks), 144 from T2 (24–28 gestational weeks), and 88 from T3 (32–38 gestational weeks; Table 1 ). It is important to note that a subset of participants was diagnosed with gestational diabetes mellitus in T2 (GDM; 79 cases and 267 controls). Full cohort details are provided in Pinto et al. [ 24 ]. Distribution of Israeli cohort participants across trimesters and their overlap. A second cohort of pregnant women, a Russian cohort, was recruited to validate our findings; In total, 154 pregnant women were recruited, of whom 138 were sampled in T2 (16-28 gestational weeks) and 106 in T3 (34–37 gestational weeks), including 90 women who were sampled at both T2 and T3, yielding a total of 244 oral microbiome samples. The majority of participants were diagnosed with GDM (139 cases and 15 controls); full cohort details are provided in Popova et al. [ 25 ]. At recruitment, the height and current weight of the study participants were recorded, and participants were interviewed by a dietitian for stress level, working hours, sleeping hours, smoking, education, and a 24-hour recall for food intake. Participants then provided a saliva sample, according to Human Microbiome Project protocols [ 26 ], collected in 1.5 ml tubes and stored at −80 °C until processing. Similar procedures were repeated in the study population in T2 (24–28 gestational weeks) and T3 (32–38 gestational weeks) of pregnancy. At recruitment (T2; 16–28 weeks of gestation), the participants’ pre-pregnancy height and weight were recorded, and participants were asked to fill out questionnaires regarding smoking habits and nutrition [ 25 ]. Participants then provided saliva, as described above. Samples were collected again in T3 (34–37 gestational weeks). It is important to note that the questionnaires used in this cohort differed from those administered in the Israeli cohort, and thus, for validation analyses, only the variables common to both cohorts were used. DNA was extracted from all collected samples using the PowerSoil DNA Isolation Kit (MO BIO, Carlsbad, CA, USA) according to the manufacturer’s instructions and following a 2-min bead beating step (BioSpec, Bartlesville, OK, USA). The variable V4 region of the 16S rRNA gene was PCR-amplified using the 515 F barcoded and 806 R primers following the Earth Microbiome Project protocol [ 27 ]. Each PCR reaction contained 25 μ l with ~40 ng/ μ l of DNA, 2 μ l 515 F (forward, 10 μ M) primer, 2 μ l 806 R (reverse, 10 μ M) primer, and 25 μ l PrimeSTAR Max PCR Readymix (Takara, Mountain View, CA, USA). PCR conditions were as follows: 30 cycles of denaturation at 98 °C for 10 s, annealing at 55 °C for 5 s, and extension at 72 °C for 20 s, followed by a final elongation at 72 °C for 1 min. Amplicons were purified using AMPure magnetic beads (Beckman Coulter, Indianapolis, IN, USA) and quantified using the Picogreen dsDNA quantitation kit (ThermoFisher, Waltham, MA, USA). Equimolar amounts of DNA from individual samples were pooled and sequenced using the Illumina MiSeq platform at the Azrieli Faculty of Medicine’s Genome Center. Sequencing data was run through the YaMAS pipeline ( https://github.com/louzounlab/YaMAS ) [ 28 ], which uses QIIME2 [ 29 ] for quality control and sequence processing. Quality control measures include filtering out low-quality reads, removing samples with fewer than 2,000 total reads, and trimming sequences. The filtered sequences were then processed to identify Amplicon Sequence Variants (ASVs), which were used to generate a table summarizing their abundance. Taxonomy was assigned with a Naive Bayes classifier trained on the GreenGenes 13_8 99% ASVs reference database [ 30 ]. The resulting ASV tables for the Israeli and Russian cohorts were then preprocessed separately using the MIPMLP pipeline [ 31 ], consisting of four steps: (1) merging similar features based on taxonomic classifications, (2) scaling the distribution, (3) standardizing features to z-scores, and (4) applying dimensionality reduction. Taxa were collapsed to the most resolved taxonomic level using the mean. All analyses were conducted in Python (v3.9). Analyses were performed separately for the Israeli and Russian cohorts to ensure independent validation. Alpha diversity metrics (Shannon diversity index and observed richness) were computed from relative abundance tables using scikit-bio. Global differences in alpha diversity across trimesters were assessed using the Kruskal-Wallis test, followed by pairwise two-tailed Mann-Whitney U tests with FDR correction (T1 vs. T2, T2 vs. T3, and T1 vs. T3). Since the Shannon index inherently includes a logarithmic component, all computations were performed on relative (non-log-transformed) abundance data. To quantify within- and between-subject variability in microbial composition over time, we computed multiple distance metrics on abundance profiles: cosine distance, Bray-Curtis dissimilarity, Jaccard dissimilarity, and Euclidean distance. For each metric, abundance profiles were first log-transformed (to reduce the disproportionate influence of highly abundant taxa) and then z-score normalized (to standardize scale across samples). Pairwise beta-diversity distances were calculated between all available trimester combinations (T1-T2, T2-T3, T1-T3) for each subject using each metric. We compared three types of distance distributions: (1) intra-individual consecutive (same woman across consecutive trimesters: T1-T2 or T2-T3), (2) intra-individual non-consecutive (same woman across T1-T3), and (3) inter-individual (different women, within or across trimesters). Differences between distributions were evaluated using one-sided Mann-Whitney U tests to determine whether samples from the same woman were more similar to each other than samples from different women, indicating subject-specific temporal stability. To assess overall differences in microbial community structure, we applied PERMANOVA (permutational multivariate analysis of variance, 9,999 permutations) as implemented in scikit-bio [ 32 ]. PERMANOVA was applied to Bray-Curtis and Jaccard dissimilarity matrices computed from the log-transformed and z-score normalized abundance profiles described above. For the GDM analysis, a two-way PERMANOVA was used: (1) the effect of GDM status (GDM vs. Control), (2) the effect of trimester (T1, T2, T3), and (3) the Group × Time interaction. To determine which specific trimesters differed in microbial composition, post-hoc pairwise PERMANOVA comparisons were conducted between all timepoint pairs (T1 vs T2, T1 vs T3, T2 vs T3), with p -values adjusted for multiple testing using the Benjamini-Hochberg false discovery rate (FDR) correction. Stratified analyses by trimester were also performed to assess GDM effects at individual timepoints. Effect sizes were reported as η 2 (proportion of variance explained). To identify compositional changes at the phylum level, relative abundances were log-transformed to reduce the influence of highly abundant taxa, and differences across trimesters were assessed using two-sided Mann-Whitney U tests with FDR correction. For finer taxonomic resolution, two complementary frameworks were used: (1) GIMIC [ 33 ], which identifies groups of taxa exhibiting coordinated changes between consecutive trimesters (T1 vs. T2 and T2 vs. T3), and (2) miMic [ 34 ], which detects taxa with significant differences between trimesters (T1 vs. T2, T1 vs. T3, and T2 vs. T3) using Mann-Whitney U tests followed by FDR correction. This hierarchical approach allowed consistent evaluation of both collapsed (at class or phylum levels) and specific (species-level) changes over time. Note that species classifications were based an amplicon sequencing and should be interpreted with caution. Associations between microbial taxa and maternal metadata were analyzed within each trimester (T1, T2, T3) and longitudinally between consecutive trimesters (T1-T2, T2-T3). For longitudinal analyses, we computed the difference in microbial abundance between consecutive trimesters for each woman and correlated these differences with baseline metadata values. Metadata included nutritional, demographic, and clinical features collected via standardized questionnaires (Table S1). Categorical variables were one-hot encoded to generate binary indicator variables. To avoid spurious correlations, we retained only features (including binary encoded categories) present in at least five participants and excluded features with no variance across the cohort. Continuous numerical variables were retained without additional filtering. For association testing, microbial relative abundances were log-transformed and z-score normalization (centered-log-ratio transformation) to ensure comparability across taxa. Spearman’s rank correlations were used for continuous metadata features, and point-biserial correlations for binary features. Multiple testing corrections were performed separately for each maternal feature across all taxa using the FDR method (Benjamini-Hochberg), with significance defined as adjusted p < 0.05 . To assess whether the gluten-free diet association reflected an independent dietary signal rather than correlated lifestyle or clinical characteristics, we fitted a linear mixed model for each taxon of the form: (1) log ( Abundance )   ∼   GlutenFree + Smoking + residential   greenness + Migraine + Conception + ( 1   |   Woman ) where Woman was included as a random effect to account for repeated measures across trimesters. The model was fitted using the mixedlm function from the statsmodels Python package (v0.14). Only taxa with prevalence ≥ 10 % across samples were included. Multiple testing corrections were applied using the Benjamini-Hochberg FDR method, with significance defined as adjusted p < 0.05 . Results were compared to a model without adjusting for covariates , which contained only the gluten-free term to identify taxa whose associations were robust to confounder adjustment versus those potentially attributable to correlated lifestyle or clinical characteristics. To assess whether maternal features with similar microbiome associations also correlate with each other, we performed an additional correlation analysis. For each maternal feature from the taxon-feature association analysis described above, we extracted its vector of correlation coefficients across all microbial taxa. We then computed the Spearman correlations between these microbial correlation profiles across all pairs of maternal features, creating a feature-feature similarity matrix. This quantified whether features showing similar patterns of association with the microbiome also tend to co-vary in the metadata. Statistical significance was defined as p < 0.05 after FDR correction (Benjamini-Hochberg method). To validate the reproducibility of our findings, all analyses, including diversity metrics, PERMANOVA tests, GIMIC and miMic analyses, distance-based intra-individual variability, and taxon-feature correlations, were independently repeated in the Russian cohort using identical preprocessing, normalization, and statistical parameters. The association with external parameters was limited to compatible measures. Because T1 samples were unavailable, analyses were limited to T2 and T3.

We analyzed 467 saliva samples collected from 346 Israeli women across all three trimesters of pregnancy ( Table 1 , Figure 1A ). In addition to 16S rRNA gene sequencing, each participant completed dietary and health questionnaires and underwent clinical tests (Table S1 for questionnaire details; Figure S1 for feature distributions). We observed significant differences in community composition across pregnancy trimesters (PERMANOVA on Bray-Curtis distances: pseudo-F = 2.2, p = 0.001 , η 2  = 0.9%; Jaccard: pseudo-F = 4.4, p < 0.001 , η 2  = 1.6%; Figure S2, S3), indicating substantial temporal shifts in the oral microbiome during gestation. Post-hoc pairwise comparisons with FDR correction revealed that the first trimester differed significantly from both the second (adjusted p = 0.006 ) and third trimesters (adjusted p = 0.003 ), whereas the second and third trimesters were compositionally similar (adjusted p = 0.052 ). To further contextualize the biological significance of this global effect size, we performed differential abundance analysis at the genus level. Despite the modest PERMANOVA η 2  = 0.9%, 51 genera showed statistically significant abundance shifts between T1 and T3 (adjusted p   0.05), independently corroborating the PERMANOVA finding that compositional change is concentrated between the first and later trimesters and demonstrating that taxon-specific shifts are biologically meaningful despite modest global variance explained. To ensure that these temporal patterns were not confounded by GDM status, we performed a two-way PERMANOVA testing for GDM effects and Group × Time interactions. GDM status explained minimal variance compared to trimester (Bray-Curtis: η 2  = 0.3%, p = 0.094 ; Jaccard: η 2  = 0.5%, p = 0.016 ). While Group × Time interactions were detected in both metrics ( p ≤ 0.002 ), stratified analyses revealed metric-dependent patterns: Bray-Curtis showed a modest T2-specific difference ( p = 0.041 ), whereas with Jaccard distances, no individual time point reached significance (all p > 0.25 ). Importantly, both groups exhibited qualitatively similar temporal progressions. Therefore, subsequent analyses focused on characterizing temporal dynamics across pregnancy. To identify the taxa contributing to temporal community shifts, we first examined changes at the phylum level. We observed a significant increase in Synergistota from T1 to T2 and a significant decrease in Verrucomicrobiota over the same interval (adjusted p = 0.003 and adjusted p = 0.006 , respectively) ( Figure 1B ; Table S2 for mean and median phylum-level abundances). Alpha diversity for all Israeli women, as quantified by the Shannon diversity index, exhibited significant differences across pregnancy trimesters (Kruskal-Wallis test, p = 0.0023 ). Pairwise Mann-Whitney tests indicated a significant decrease from T1 to T2 (adjusted p = 0.031 ) and from T1 to T3 (adjusted p = 0.010 ; Figure 1C ). Similar patterns were observed for species richness (Figure S6). Shannon diversity did not differ significantly between T2 and T3 (adjusted p = 0.264 ); median values were 3.261 (T1), 3.173 (T2), and 3.109 (T3) ( Figure 1C ). To further characterize the microbial shifts driving these community-level changes, we examined differential taxonomic composition across trimesters and across taxonomic levels using GIMIC [ 33 ]. Several taxa within the classes Alphaproteobacteria, Verrucomicrobiota, and Erysipelotrichia consistently decreased from T1 to T2, while Methanobacteria and Deferribacteres decreased from T2 to T3. The direction and magnitude of these shifts were stable across taxonomic levels, with Deferribacteres showing a consistent decline throughout pregnancy ( Figure 1D ). To statistically test the difference between microbial frequencies, we used miMic (the statistical test matching of GIMIC) [ 34 ] to detect specific taxa differing between trimesters (T1 vs. T2 and T1 vs. T3; Figure 1E and Figure S7). A. muciniphila , belonging to the Verrucomicrobiota phylum, which consistently decreased from T1 to T2 in the GIMIC analysis, also showed a significant species-level decrease ( Figure 1E ). Other notable shifts included an increase in Gammaproteobacteria and a decrease in Erysipelotrichia in T2. We hypothesized that the modest contribution of the trimesters to the total variability is the result of a conserved personal microbiome. To test this, we computed the variability in microbial composition over time within and between women. Specifically, we calculated cosine distances between samples. The distances between microbiomes of the same women across consecutive trimesters were significantly lower than those between different women, indicating subject-specific temporal stability in community composition (adjusted p < 0.001 ). However, when comparing non-consecutive trimesters, comparisons between the same women from T1 and T3 revealed significantly greater dissimilarity than T1-T2 or T2-T3 alone, consistent with cumulative compositional divergence that accumulates progressively over the course of pregnancy (T1-T2 vs. T1-T3: one-sided Mann-Whitney U = 1052, adjusted p = 0.033 ; T2-T3 vs. T1-T3: one-sided Mann-Whitney U = 596, adjusted p = 0.005 ; Figure 1F ). Similar patterns were observed for alternative distance metrics (Figure S8). We next examined maternal factors associated with the oral microbiome development. The correlation between normalized microbial frequencies and 54 maternal variables collected pre-pregnancy or during specific trimesters, encompassing dietary habits ( n  = 12), demographic factors ( n  = 8), medical history ( n  = 15), clinical measurements ( n  = 14), and pregnancy outcomes ( n  = 8) was computed (see Figure 1A and Table S1 for complete feature descriptions). Of these 54 features, three variables demonstrated consistent and significant correlations with the oral microbiome across trimesters after FDR correction, indicating that the majority of maternal features had little association with the oral community composition ( Figure 2A ). The most robust associations were observed for adherence to a gluten-free diet, which correlated with oral microbiome structure in all three trimesters as well as with inter-trimester changes (T1-T2 and T2-T3). Specifically, adherence to a gluten-free diet was associated with an increase in the relative abundance of the associated taxa. Tree-based association mapping ( Figure 2B ) revealed that a gluten-free diet in T2 was linked to multiple taxa, with the strongest effects involving members of the families Lachnospiraceae and Ruminococcaceae. Associations between maternal characteristics and the oral microbiome. (A) Bar plot showing the number of significant correlations (adjusted p  < 0.05) between maternal features and microbial taxa in T1 (blue), T2 (purple), and T3 (orange), as well as within-woman differences between consecutive trimesters (T2-T1, light blue; T3-T2, red). Only the top 20 features (ranked by total number of significant associations across all comparisons) are shown. (B, C) Circular taxonomic tree mapping microbes significantly associated with gluten-free diet, and smoking history, respectively, in oral microbiome T2. Following the same format as the miMic tree in Figure 1E , Dot colors indicate directionality (blue for increased abundance, red for decreased abundance in women following a gluten-free diet, or that smoked in the past), dot size reflects the absolute Mann-Whitney test statistic multiplied by 30, and background shading groups taxa by family. (B, C) share the same family colors. The trees were filtered to show only the significant genera and species (See Figure S9-S18 for additional full circular taxonomic trees for other features and trimesters, formatted similarly to the miMic tree). To assess whether the gluten-free association reflected an independent dietary signal rather than correlated lifestyle or clinical characteristics, we fitted a linear mixed model for each taxon, adjusting for smoking history, residential greenness [ 35 ], migraine comorbidity, and conception method, with the woman as a random effect to account for repeated measures across trimesters. Of 48 taxa significantly associated with gluten-free diet in the unadjusted model, 32 (66.7%) retained significance after full confounder adjustment, indicating that the majority of associations are robust and not attributable to the tested confounders. The 16 taxa that lost significance potentially reflect effects of lifestyle or clinical characteristics correlated with dietary choice rather than diet itself (Figure S29). A history of smoking was also significantly associated with oral microbial composition. Most associations were in taxa within Clostridiales and Bacteroidales classes ( Figure 2A ; with T2 associations shown in Figure 2C ). These associations were present in all trimesters and with inter-trimester changes. Conception aided by clominphene citrate (Ikaclomin) was also associated with oral microbial composition, sharing associations with taxa from the orders Clostridiales and Bacteroidales to Actinomycetales, in T2 ( Figure 2A , Figure 3 ). Associations between maternal metadata and microbial taxa. Plot of significant correlations (adjusted p  < 0.01) between oral microbiome taxa and maternal metadata across pregnancy trimesters, excluding smoking and gluten-free diet (see Figure 2 ). The left y-axis lists microbial taxa, and the right y-axis indicates their phylogeny. The plot is divided into four panels: correlations within T1, within T2, differences between trimesters (T2-T1), and within T3. Differences between T3 and T2 were also analyzed but showed no significant associations after applying FDR correction. Columns represent maternal features (nutrition, demographics, and medical history). Each dot denotes a significant association, with color indicating directionality (blue for positive, red for negative). Dot size is scaled linearly to the absolute correlation coefficient, such that stronger correlations are displayed with larger dots. Spearman’s rank correlation was used for continuous metadata features, and biserial correlation was applied for binary features. See Figure S19 for the complete heatmap of associations. To assess whether relationships among maternal features were stable across pregnancy, we computed pairwise correlations among all metadata variables for each trimester (Figures S26-28). Overall, a similar network structure of inter-feature dependencies was observed across trimesters, with several features repeatedly forming co-varying clusters. In T1, lifestyle and metabolic variables, such as gluten-free diet, migraine comorbidity, and past smoking were strongly intercorrelated ( | r |  = 0.48-0.54, all p < 0.01 ), while preterm delivery showed negative associations with both gluten-free diet ( r = − 0.40 , p < 0.001 ) and very high GCT ( r = − 0.31 , p = 0.046 ). By T2, this pattern became more pronounced: Gluten-free diet remained correlated with migraine comorbidity, with a negative correlation coefficient ( r = − 0.52 , p < 0.001 ), and a cluster involving smoking (past and current), conception type, and preterm delivery emerged ( | r |  = 0.45-0.77, all p < 0.01 ). These associations suggest increasing interdependence between lifestyle, conception-related, and clinical variables as pregnancy progresses. In T3, inter-feature correlations became sparser, but several consistent relationships persisted. A gluten-free diet remained correlated with past smoking ( r = 0.52 , p = 0.008 ), while migraine comorbidity and antibiotic use formed a minor positive cluster ( r = 0.31 , p = 0.007 ). To validate the reproducibility of the findings obtained for the Israeli cohort in an independent population, we analyzed the oral microbiome of 154 Russian women sampled during T2 and T3 of pregnancy. PERMANOVA revealed significant differences in beta diversity between T2 and T3, both for Bray-Curtis dissimilarity (pseudo-F = 2.93, p = 0.002 ) and for the Jaccard index (pseudo-F = 7.10, p < 0.001 ), indicating that temporal compositional shifts were detectable in this independent population (Figures S20, S21). At the phylum level, significant decreases in T3 were observed in Euryarchaeota (adjusted p = 0.0002 ), Actinobacteria (adjusted p = 0.0487 ), and Verrucomicrobiota (adjusted p = 0.0002 ), while no phyla exhibited significant increases between trimesters (Figure S22). Alpha diversity did not change from T2 to T3 ( p = 0.22 ). Consistent with the Israeli cohort, GIMIC analysis revealed a pronounced and consistent decrease in Verrucomicrobiota (particularly A. muciniphila ) and in Methanobacteria across taxonomic levels (see FigureS23). miMic analysis identified A. muciniphila as significantly reduced from T2 to T3 (adjusted p = 0.0005 ), and within the genus Veillonella , V. parvula and V. dispar showed a significant increases in T3 (adjusted p = 0.004 for both). As detected in the phylum-level analysis, Collinsella aerofaciens (phylum Actinobacteria) showed a consistent decrease in T3 (adjusted p = 0.03 ), while P. copri exhibited an even stronger decrease (adjusted p < 0.001 ). Methanobrevibacter likewise showed a consistent decrease across all taxonomic levels (adjusted p < 0.001 ) (Figure S24). Analysis of intra-individual distances further demonstrated that, despite this population-level shift, samples from the same woman remained significantly more similar to each other (T2-T3) than to samples from different women ( p < 0.001 ), supporting subject-specific microbial stability over short time intervals (Figure S25). Furthermore, in the Russian cohort, Dialister (Veillonellaceae family) in T2 was significantly associated with past smoking history (adjusted p = 0.01 ), aligning with smoking-related associations observed in the Israeli cohort. The gluten-free dietary pattern identified in the Israeli cohort could not be directly validated in the Russian cohort, as dietary gluten was captured as a continuous intake variable rather than a binary dietary pattern, precluding a direct comparison. However, when gluten intake was operationalized using a binary cutoff of 0.5 g/day [ 36 ], Enterobacteriaceae showed a significant association in the Russian cohort at T3 (adjusted p = 0.044 ), consistent with its association in the Israeli cohort at T2 (adjusted p  = 0.010), though the T3 association in the Israeli cohort did not remain significant following FDR correction (adjusted p = 0.135 ). This partial replication suggests that certain taxon-level associations with gluten intake may be detectable across populations when a comparable operationalization of the variable is used, while the broader pattern of gluten-free associations observed in the Israeli cohort remains cohort-specific, likely reflecting cohort-specific factors, such as differences in dietary assessment methodology and population background. These results confirm the reproducibility of key microbial shifts and maternal factor associations across both cohorts, highlighting the robustness of the observed oral microbiome dynamics during pregnancy.

We have presented here the first large-scale longitudinal investigation of oral microbiome changes from T1 through T3, addressing a critical gap in understanding early gestational microbial dynamics. In the Israeli cohort, significant changes in oral microbial community structure were observed between T1 and T3. Pairwise comparisons indicated that the largest compositional shifts occurred between T1 and T2, with comparatively smaller changes between T2 and T3. In contrast, the Russian cohort, which included sampling in T2 and T3 only demonstrated significant T2-T3 differences, indicating that temporal shifts persist beyond mid-gestation. Although earlier studies have documented oral microbiome alterations, primarily in T2 and T3 [ 2 , 5 , 6 ], our results in the Israeli cohort highlight the importance of early gestational transitions. The overall pattern suggests a progressive restructuring of the oral microbial community across pregnancy rather than abrupt late-gestation changes. The significant decrease in Verrucomicrobiota, particularly A. muciniphila , in the Israeli cohort, and the concurrent increase in Synergistota from T1 to T2 provides important insights into the microbial dynamics of early pregnancy. A. muciniphila , a mucin-degrading bacterium extensively studied in the gut microbiome for its roles in maintaining intestinal barrier integrity and metabolic health [ 37 , 38 ], has also been detected in oral samples, though its prevalence and functional significance in the oral cavity remain less well characterized. There is some evidence that this taxon is protective against periodontitis, mirroring its beneficial role in the gut [ 37–39 ]. The reduction of A. muciniphila in the oral cavity during early pregnancy may reflect systemic changes in host-microbe interactions driven by hormonal fluctuations [ 3 , 40 ]. Synergistota are typically present in low abundance in the oral microbiome, and have been consistently reported in subgingival plaque, periodontal pockets, and endodontic infections, often co-occurring with keystone periodontal pathogens such as Porphyromonas gingivalis [ 1,40–42 ]. Their increased abundance has also been associated with proteolytic metabolism, persistence in inflamed niches, and contribution to chronic biofilm dysbiosis [ 43 ]. Thus, the relative increase of Synergistota and the relative decrease in A. muciniphila with pregnancy progression may reflect ecological remodeling of the oral microbiome during pregnancy, shifting from a potentially protective profile toward one enriched in taxa linked to tissue inflammation [ 41 ]. The increase in Gammaproteobacteria and the decrease in Erysipelotrichia classes further support the concept of a directional shift in the structure of the microbial community between T1 and T2. Gammaproteobacteria are known members of the oral cavity that are often increased under inflammatory or metabolically altered conditions [ 44 ]. Erysipelotrichia, in contrast, are more commonly associated with the gut microbiome but have also been detected in saliva [ 45 ]. This strengthens the above hypothesis that the oral microbiota in later pregnancy is pro-inflammatory [ 46 , 47 ]. Despite these alterations, intra-individual profiles remained relatively conserved within individuals compared to between individuals. This suggests that while pregnancy exerts directional pressure on microbial community structure, individual microbial ‘signatures’ persist. The strong association between a gluten-free diet and oral microbiome during pregnancy was among the most intriguing findings in our Israeli cohort. These predominantly positive correlations with taxa within Clostridiales and Bacteroidales, relative to those not adhering to a gluten-free diet, persisted throughout pregnancy. We suggest that eliminating gluten-rich foods may reduce probiotic substrates, enabling expansion of taxa that exploit broader carbohydrate niches. Smoking history was also significantly associated with oral community composition across trimesters, with consistent divergence between women with and without a history of smoking observed in all trimesters. This finding aligns with prior cohort and review evidence showing enrichment of periodontal-associated genera (e.g. Prevotella , Porphyromonas ) and depletion of commensal Proteobacteria such as Neisseria among smokers, likely reflecting sustained effects on the oral mucosal environment, including altered oxygen tension, inflammation, and immune modulation [ 48–52 ]. We also identified an association between Ikaclomin-aided conception and T2 oral microbiome, particularly involving Actinomycetales. Interestingly, this signal appeared to attenuate by T3, suggesting assisted reproductive technologies’ (ART) effects may be strongest in early pregnancy. Possible mechanisms include hormonal stimulation during ART [ 53 ] and links between oral dysbiosis and infertility-related conditions like polycystic ovary syndrome or endometriosis [ 54 ]. Although conception method was associated with microbial variation, we did not observe evidence of persistent or long-term effects of ART on maternal microbial composition. These findings were validated in an independent cohort of 154 pregnant women from Russia, underscoring their generalizability across populations and geographic regions. In this cohort, we observed a decline in A. muciniphila from T2 to T3, reinforcing its role as a key taxon. In contrast, several taxa that showed differential abundance during mid-to-late gestation (T2-T3) in the Russian cohort were not duplicated in the Israeli cohort. This may arise from dietary and lifestyle differences between the cohorts. Notably, a history of past smoking was again associated with a reduction in Dialister abundance at T2, echoing the smoking-related associations found in the Israeli dataset. These results confirm that oral microbiome community dynamics during pregnancy are partially shared across distinct populations, supporting their relevance to maternal-fetal health. Several limitations should be considered. First, the observational design precludes causal inference, and unmeasured factors like oral health status (e.g. gingivitis, caries, salivary pH) were not systematically assessed. Second, self-reported dietary data raises the possibility of recall bias, while sampling at fixed gestational weeks may miss continuous microbial dynamics. Furthermore, variations in cohort composition (e.g. trimesters, GDM prevalence) may complicate cross-cohort comparisons. Finally, our use of V4 16S rRNA gene amplicon sequencing limits functional inference and taxonomic resolution compared to metagenomics or metabolomics. Inherent primer biases and database constraints mean that species-level assignments remain putative; therefore, species-specific inferences should be interpreted cautiously and ideally validated with orthogonal methods, such as species-specific qPCR, in future studies. To assess whether missing samples introduced systematic bias, we compared microbiota profiles of women at a given time point based on whether they were subsequently sampled. Specifically, women who had a T1 sample and were later sampled at T2 ( n  = 62) were compared to women who had a T1 sample but were not sampled at T2 ( n  = 173), using only their T1 data. The same approach was applied at T2: women subsequently sampled at T3 ( n  = 41) were compared to those who were not ( n  = 103), using only their T2 data. No significant differences were found in any clinical or demographic variable after FDR correction (all adjusted p > 0.05 ). Observed richness at T1 was modestly lower in women subsequently sampled at T2 (median 84 vs 92, adjusted p = 0.014 ), though Shannon diversity did not differ (adjusted p = 0.088 ), and no alpha diversity differences were observed at T2 (all adjusted p > 0.87 ). Although PERMANOVA detected significant compositional differences at T1 (Pseudo-F = 2.99, p = 0.001 ), PERMDISP analysis indicated this was driven by differences in within-group dispersion rather than centroid location (F = 9.02, p  = 0.006), and sequencing depth was comparable between groups ( p = 0.758 ). Together, these analyses indicate that missing samples are unlikely to have introduced systematic bias. To formally assess whether cohort differences affect generalizability, we fitted a linear model for each genus: log ( Abundance ) ∼ Timepoint + Cohort + Timepoint × Cohort , where the interaction term directly tests whether temporal dynamics differ between cohorts. Of 101 common taxa tested, 95 (94.1%) showed no significant Taxon × Cohort interaction after FDR correction, indicating that temporal abundance changes from T2 to T3 were statistically consistent across populations. Notably, Akkermansia , the genus associated with the Verrucomicrobiota reduction highlighted in our findings, showed concordant depletion in both cohorts (interaction adjusted p-value = 0.163). Only five genera showed significant cohort-specific effects: Prevotella , Blautia , Oscillospira , Acidaminococcus , and Coprococcus . We attribute these differences to population-specific dietary and clinical factors.

Longitudinal profiling across all trimesters of pregnancy demonstrates pronounced remodeling of the maternal oral microbiota, most evident from T1 to T2, shaped by maternal factors such as smoking history, diet, and mode of conception. Like with the gut microbiota, these shifts in oral microbiota are consistent with a transition toward a composition associated with inflammation. Collectively, we conclude that early-mid pregnancy is a particularly dynamic window in which the oral microbiota may be a relevant therapeutic target. By mapping trimester-specific trajectories tied to modifiable exposures, our work lays the groundwork for microbiota-informed screening and timed intervention in routine obstetrics.

supp_mat_oral.pdf