The association between the pre-pregnancy vaginal microbiome and time-to-pregnancy: a Chinese pregnancy-planning cohort study

This study was divided into two phases. Phase I was conducted from 13 June 2018 to 31 October 2018; 106 women participated in this phase. This phase featured a nested case-control design. All participants were divided into pregnant or non-pregnant groups according to pregnancy outcomes after 1 year of participation. The potential biomarkers for bacteria that were identified in phase I were then detected in phase II, and their associations with TTP were verified by a cohort design. Phase II was conducted from 1 November 2018 to 30 May 2020. In total, 500 women were invited, and 495 women signed the informed consent; 23 women refused to provide vaginal swabs because of a menstrual period, and 51 women withdrew without the first visit. The final study included 89 women in phase I and 332 women in phase II. Further details are shown in Fig. 1.

Sample size estimations were performed when the research protocol was first designed (further details are provided in Additional file 1 [12]). All participants signed an informed consent form, and the study was approved by the Ethics Committee of Zhongda Hospital (Reference: 2018ZDSYLL116-P01).

At baseline, we performed a unified epidemiological survey for every female so that we could collate their age, the age difference within couples, educational level (high school and below/higher education and above), occupation (workers/office clerk/others), pregnancy history (yes/no), and menstrual status (regular or not). A regular menstrual cycle was defined as a cycle length of 21–35 days [13]. All data were acquired by one professional nurse to ensure that the information was credible.

All females would be contacted by the medical staff (by telephone) every 3 months. The main outcome was clinical pregnancy, as self-reported by the subjects; this needed to be confirmed by a pelvic ultrasound scan in the hospital. TTP was the interval between the dates of the last menstrual period (LMP) obtained at follow-up and before conception (pregnant within 1 year) or the last follow-up call (if not pregnant). TTP in months was calculated by TTP in days/30. The TTP in cycles was calculated by TTP in days/average length of the menstrual cycle. These indices were all round up to an integer.

Women were placed in a lithotomy position under standard operating procedures; then, gynecologists obtained two vaginal swabs with the aid of a sterile speculum. Swabs were rotated three times on the vaginal fornix to uniformly scrape any discharge and were transported to the laboratory within 4 h. Vaginal cleanliness was graded via microscopic examination of cervical smears. Grades I and II were regarded as normal while grades III and IV were regarded as disordered; these gradings were in accordance with the Chinese standards [14]. The second swab was stored in a dry tube at − 80 °C to await nucleic acid extraction. The detailed procedures for DNA extraction and 16S rRNA gene sequencing are described in Additional file 1. In brief, the swabs were eluted with PBS buffer and the TIANamp Bacterial DNA Kit (Tiangen Biochemical Technology, Beijing, China) as used to extract and purify nucleic acids. The V3–V4 region of the 16S rRNA gene was amplified and sequenced on an Illumina HiSeq 2500 platform (Beijing Biomarker Technologies Co. Ltd., Beijing, China). The raw sequencing data is stored in the figshare platform [15]. Then, sequencing data were processed using a standard procedure (Additional file 1). Denoised sequences were clustered using USEARCH (version 10.0), and tags with ≥ 97% similarity were regarded as an operational taxonomic unit (OTU). Representative sequences were annotated through the National Center for Biotechnology Information (NCBI) dataset using the QIIME software (https://​qiime2.​org). The numbers of reads for each sample were normalized according to the sample with the least sequence. All bioinformatics analyses were completed on the Biomarker BioCloud platform (www.​biocloud.​org).

We used the QIIME2 software (http://​qiime2.​org/​) to calculate α diversity for the vaginal microbiome, including Shannon, Simpson, Chao1, and ACE indices. These indices reflect the richness and diversity of the microbial community structure [16]. Based on the matrix of relative abundance of bacteria, we estimated the Jaccard Distance Index and then performed principal coordinate analysis (PCoA) to intuitively display different groups of the microbiome. Next, we performed a permutational multivariate analysis of variance (PERMANOVA) to test for statistical significance. Linear discriminant analysis (LDA) effect size (LEfSe) was used to identify potential biomarkers among different groups, which should meet P < 0.05, adjusted P < 0.01, and LDA > 4.0 [17]. The Benjamini-Hochberg (BH) method was used to adjust P values to minimize the false discovery rate when performing multiple comparisons.

Quantitative real-time polymerase chain reaction (qPCR) was used to measure the absolute loads of specific vaginal bacteria, including L. crispatus, L. gasseri, L. iners, Gardnerella vaginalis (G. vaginalis), and Fannyhessea vaginae (F. vaginae, also called Atopobium vaginae). We used specific primers that had been verified by previous studies (further details are given in Additional file 2: Table S1 [18‐21]). We used the NCBI Blast database to predict the amplified products, and specific plasmids were synthesized by Sangon Biotech Company (Shanghai, China). The copy number concentration of the plasmid was calculated using the following formula: copies/mL = 6.02 × 10²³ × 10⁻⁶ × concentration (ng/μL)/(fragment length × 660). Then, 10-fold serial dilutions of the plasmid were prepared and subjected to qPCR to obtain a standard curve. In order to reduce variations in the total bacterial load from different swab samples, the copy number concentrations of the 16S V3–V4 region were also measured; these were then standardized to 1 × 10¹⁰ copies/mL for each sample. Under this condition, we measured the absolute abundance of another 5 species.

For each species, we calculated the absolute abundance z-score after the logarithmic transformation of the absolute loads. The clustering of microbial communities was explored with the k-means algorithm which minimizes the error inside the groups and maximizes the distance between clusters. We considered the Euclidean distance metric in our analysis and then tried to use the elbow method to determine the optimum number of clusters [22]. In this method, the slow-down point denotes the optimum number of clusters. Then, we compared the average z-score for specific species among different clusters using variance analysis.

All data were uploaded into the EpiData (Version 3.1) software by two independent researchers. The analyses followed a defined approach that was determined before running the models. Continuous variables are described by the mean and standard deviation (SD) (normal distribution) or median and quartile if not distributed normally. The t test or the Kruskal-Wallis test was used to test for the differences between the groups. Categorical variables are described by frequency and percentage; the chi-squared test or Fisher’s exact test was used to compare the distribution between the groups. Missing data were imputed by the multivariate imputation chained equations (MICE) package in the R software [23]. We set up five imputed datasets; the main analysis results were aggregated with Rubin’s rule after appropriate transformation [24]. We performed analyses using the completed case dataset as sensitivity analysis.

Spearman coefficients were calculated to determine the correlation between two relative abundances of bacteria. The Kaplan-Meier (KM) method was used to calculate the cumulative pregnancy rates in different types of microbiomes, and the log-rank test was used to test the differences. Cox models were used to estimate the fecundability ratios (FRs) and their 95% confidence intervals (CIs) for different types of microbiome after adjusting for potential confounding factors. FR reflects the ratio of pregnancies among females with certain characteristics compared with the reference groups; thus, an FR < 1.0 implies a lower fecundability or a longer TTP. All of these analyses were carried out using the R software (version 4.1.0), and two-sided probability values of < 0.05 were deemed to be statistically significant.

In phase I, the mean age of the participants was 28.66 ± 3.14 years old; most women did not have a history of pregnancy (79/89, 88.76%). In total, 59.6% (53/89) of women achieved pregnancy within 1 year. A comparison of the baseline characteristics between the two groups (pregnant or non-pregnant) revealed that there were no significant differences in terms of the age difference within couples, educational level, occupation, history of pregnancy, and the regularity of menstruation (P > 0.05, Table 1), although the mean age of the pregnancy group was significantly lower than that of the non-pregnancy group (27.98 vs 29.52, P = 0.036).

All nucleic acid samples from vaginal swabs were sequenced successfully, and the sequencing depths were sufficient (Additional file 2: Fig. S1). The sequencing quality of all samples was confirmed to be good (all Q20 indices were > 95%, Additional file 2: Table S2). At the genus level, the most common bacteria with the highest abundances were Lactobacilli (mean relative abundance 79.95%), Gardnerella (8.57%), Streptococcus (1.79%), and Atopobium (1.54%) (Additional file 2: Fig. S2). Comparisons of the Shannon, Simpson, Chao1, and ACE indices showed that pre-pregnancy vaginal bacterial diversities were not statistically different when compared between the pregnant and non-pregnant groups (P > 0.05, Additional file 2: Fig. S3). However, PCoA showed that the vaginal microbial community structures of these two groups were slightly different; 2.5% of variations were associated with pregnancy outcomes (PERMANOVA test, R² = 0.025, P = 0.049, Fig. 2A, B). To further identify the key species, we performed a Lefse analysis; the results showed that the abundance of the Lactobacillales order in the pregnant group was higher than that in the non-pregnant group (LDA > 4.0, average relative abundance 86.33% vs 75.63%). The abundance of the Actinobacteria phylum (8.64% vs 14.01%, LDA > 4.0) and the Gardnerella genus (6.34% vs 11.84%, LDA > 4.0) in the non-pregnant group was higher than those in the pregnant group (Fig. 2C). At the genus level, random forest model analysis identified potential biomarkers for distinguishing pregnancy or non-pregnancy, including Gardnerella, Lactobacillus, and Fannyhessea, with a relatively high Gini index (Fig. 2D). Based on these results, we further compared the relative abundance of this genus among the two groups (Fig. 3). The relative abundance of Gardnerella in the pregnant group was significantly lower than that in the non-pregnant group (P = 0.0029).

Based on the findings from phase I, we further focused on the Gardnerella, Fannyhessea, and Lactobacillus genera. In consideration of the leading role of Lactobacillus in the vaginal microbiome, and the potentially differential effects of different Lactobacillus species, except for the G. vaginalis and F. vaginae, we also detected the three most common Lactobacillus species, including L. cripatus, L. iners, and L. gassri. In total, 332 women were included in the phase II analysis. The mean age was 29.50 ± 3.95 years,, and the mean age difference within couples was 1.35 years. The baseline characteristics of women who were excluded for various reasons were comparable with those included (Additional file 2: Table S3). The absolute loads of L. crispatus, L. gasseri, L. iners, G. vaginalis, and F. vaginae in the vaginal swabs taken at baseline were detected using standard curves (Additional file 2: Fig. S4). The correlation analysis shows that L. crispatus was positively associated with L. gasseri (ρ = 0.56, P < 0.001) and negatively associated with L. iners (ρ = 0.18, P = 0.004). L. iners was negatively associated with L. gasseri (ρ = − 0.14, P = 0.009). G. vaginalis was positively associated with F. vaginae (ρ = 0.31, P < 0.001). The associations between the remaining species were not statistically significant (P > 0.05), the details were shown in Additional file 2: Table S4.

The absolute loads of these species at baseline were then divided into four groups (Q1–Q4) based on the interquartile range. Cox models were then used to estimate the association between these four groups and female fecundability (Additional file 2: Table S5). Data showed that female fecundability increased with higher absolute loads of L. gasseri (Q4 vs Q1, FR = 1.71, 95%CI 1.02–2.87) but decreased with higher absolute loads of F. vaginae (Q4 vs Q1, FR = 0.62, 95%CI 0.38–1.00). Other species were not statistically associated with female fecundability (P > 0.05).

Based on the absolute loads of five bacterial species, we found that the vaginal microbiome clustered into five types (A–E). Figure 4A shows that type A (24.7%, 82/332) featured three high abundant Lactobacillus species, with a low abundance of G. vaginalis and F. vaginae. Type B (13.0%, 43/332) was characterized by a high abundance of G. vaginalis and a low abundance of the three Lactobacillus species. Type C (20.2%, 67/332) was characterized by a high abundance of F. vaginae and a modest abundance of the three Lactobacillus species. Type D (22.3%, 74/332) was characterized by a high abundance of L. iners abundance and low abundances of the other four species. Type E (19.9%, 66/332) was characterized by high abundances of L. crispatus and L. gasseri and low abundances of the other three species. The z-scores of the absolute abundance of specific species grouped by different types are shown in Additional file 2: Table S6. Figure 4B showed that women with different types of vaginal microbiome had different levels of fecundability (log-rank test, P = 0.014). Women with the type A vaginal microbiome had the highest cumulative pregnancy rate (12th month, 54.7%, 95%CI 41.2–65.1%) while women with the type D vaginal microbiome had the lowest cumulative pregnancy rate (12th month, 28.2%, 95%CI 16.8–38.1%). Types B, C, and E had similar cumulative pregnancy rates (12th month, 44.5% vs 45.0% vs 45.2%).

After adjusting for potential confounding factors, including female age, educational level, occupation, pregnancy history, vaginal cleanliness grading, and the age difference between couples, we found that compared to women with a type A vaginal microbiome, women with a type D microbiome showed a 55% reduction in fecundability (model A, FR = 0.45, 95%CI 0.26–0.76). This association was robust irrespective of whether the TTP was determined by month or menstrual cycle (model B, FR = 0.45, 95%CI 0.27–0.77). Women with vaginal microbiome types B, C, and E had lower tendencies of fecundability compared with type A, although these differences were not statistically significant (Table 2). Sensitivity analysis based on the dataset of completed cases was consistent with these primary results (Additional file 2: Table S7).