Meta-analysis reveals the vaginal microbiome is a better predictor of earlier than later preterm birth

Abstract

Background: High-throughput sequencing measurements of the vaginal microbiome have yielded intriguing potential relationships between the vaginal microbiome and preterm birth (PTB; live birth prior to 37 weeks of gestation). However, results across studies have been inconsistent.

Results: Here, we perform an integrated analysis of previously published datasets from 12 cohorts of pregnant women whose vaginal microbiomes were measured by 16S rRNA gene sequencing. Of 2039 women included in our analysis, 586 went on to deliver prematurely. Substantial variation between these datasets existed in their definition of preterm birth, characteristics of the study populations, and sequencing methodology. Nevertheless, a small group of taxa comprised a vast majority of the measured microbiome in all cohorts. We trained machine learning (ML) models to predict PTB from the composition of the vaginal microbiome, finding low to modest predictive accuracy (0.28-0.79). Predictive accuracy was typically lower when ML models trained in one dataset predicted PTB in another dataset. Earlier preterm birth (< 32 weeks, < 34 weeks) was more predictable from the vaginal microbiome than late preterm birth (34-37 weeks), both within and across datasets. Integrated differential abundance analysis revealed a highly significant negative association between L. crispatus and PTB that was consistent across almost all studies. The presence of the majority (18 out of 25) of genera was associated with a higher risk of PTB, with L. iners, Prevotella, and Gardnerella showing particularly consistent and significant associations. Some example discrepancies between studies could be attributed to specific methodological differences but not most study-to-study variations in the relationship between the vaginal microbiome and preterm birth.

Conclusions: We believe future studies of the vaginal microbiome and PTB will benefit from a focus on earlier preterm births and improved reporting of specific patient metadata shown to influence the vaginal microbiome and/or birth outcomes.

Keywords: Machine learning; Meta-analysis; Preterm birth; Vaginal microbiome.

Fig. 1

The average proportion of all sequencing reads in each dataset derived from the set of common ASVs (found in every dataset), top ASVs (proportion larger than 0.1% in any dataset), and core genus-level taxonomic features (abundance > 0.1% and prevalence > 10% for at least 5 datasets). Note that common and top ASV features were determined within the V1V2 and V4 dataset groups independently, as non-overlapping ASVs are not directly comparable (Methods)

Fig. 2

A A schematic of different analytical strategies using machine learning. Each square represents a different dataset, and squares are colored by how they are used to train or test the ML model. B The prediction accuracy, as measured by the AUC, for random forest ML models trained on the vaginal microbiome profiles (genus-level proportion data) in one dataset (rows) and tested in the same or a different dataset (columns). “Ave.” indicates the average AUC of each row (same training dataset) or each column (same testing dataset)

Fig. 3

Assessment of prediction performance for different preterm birth groups using intra-dataset analysis (A), LODO analysis (B), and cross-dataset analysis (C) using CLR-transformed data. A resampling procedure is used to ensure each preterm birth group has the same sample size. The experiment is repeated 10 times and the average AUC and/or standard error are calculated. For each heatmap, diagonal AUC values are from intra-analysis, and off-diagonal values are from cross-analysis

Fig. 4

The feature importance ranking for genus-level taxonomic features (rows) in random forest ML models trained in different datasets (columns) using proportion data. Feature importance was quantified as the absolute SHAP value. Genera are ordered by their mean importance rank across all datasets

Fig. 5

Cross-dataset differential abundance analysis. A Point estimates and 95% confidence intervals of the log odds ratio of a genus being present in preterm births relative to term births using a generalized linear mixed model. Presence was defined as a relative abundance greater than 0.001. The model included all 12 datasets and no population characteristic covariates. Point estimates less than 0 are shown as blue points and greater than 0 as red points. Confidence intervals less than 0 are shown as blue bars and greater than 0 as red bars. B Posterior distribution of log odds ratio using pooling and set-specific methods for four selected genera/species