Metabolic Dynamics and Prediction of Gestational Age and Time to Delivery in Pregnant Women

Danish Pregnancy Cohort: A Study of Normal Pregnancy with High-Density Sampling

To capture the highly dynamic pregnancy process at high resolution, we established a multi-year single-center Danish normal pregnancy cohort and a design of high-density blood sampling. Consenting female participants submitted weekly blood draws beginning in week 5 of pregnancy and ending in the postpartum period. A total of 30 women with weekly blood samples were assigned to a discovery (N = 21) and a validation (test set 1, N = 9) cohort (Table 1; Figures 1A and S1A). The samples were analyzed in two separate years. In addition, another separate set of women (N = 8) was included as the secondary validation cohort. These samples were analyzed independently three years apart from the discovery cohort (test set 2) (Table 1).

Weekly Pregnancy Progression Is Precisely Ordered by Metabolites

We randomized the 784 samples from the first 30 subjects within each cohort (Discovery and test set 1), processed them by using a standardized protocol (Contrepois et al., 2015), and analyzed them by liquid chromatography-mass spectrometry (LC-MS) for untargeted metabolomics across two separate years. After quality control, data filtering, and normalization (see STAR Methods) (N = 30), we identified 9,651 metabolic features across the different samples. Of these, 4,995 features (51.7%) were altered during pregnancy and/or the postpartum period (false discovery rate [FDR] < 0.05), suggesting extensive metabolic changes occur during pregnancy. We examined the data globally with principal component analysis (PCA), in which the samples were distributed on the basis of the first two principal components according to their gestational stages (Figure 1B; Scree plot in Figure S1B and the partial least-squares discriminant analysis [PLS-DA] results in Figure S1C), regardless of individual variation and batches (Figures S1D and S1E). Interestingly, we found that metabolites with uni-directional behaviors dominated the features, and over half of them increased across pregnancy until reaching their peaks immediately before labor (Figures S1F and S1G).

To understand the potential function of pregnancy-related metabolites, we first annotated metabolic features by using an in-house library and a combined public spectral database (see details in STAR Methods). A total of 952 metabolic features were mapped to 687 compounds, which include plasma metabolites with important functions in humans. We then applied significance analysis for microarrays (SAM) to examine the correlation between the abundance of each compound and the reported gestational age of a woman at blood sampling. Among the 687 annotated compounds, 460 compounds were significantly associated with pregnancy (67.0%; FDR < 0.05, SAM). In addition, 264 compounds were identified with a metabolomics standards initiative (MSI) level 1 or 2 identification (Viant et al., 2017), among which 176 compounds (66.7%) were significantly associated with pregnancy, as determined by linear regression with gestational age (FDR < 0.05, SAM).

Our dense sampling revealed detailed temporal patterns of molecular changes. Among the top 68 metabolites (of the 176) that changed over 50% during whole pregnancy, those that increased (N = 30) included steroid hormones estriol-16-glucuronide, estrone 3-sulfate, and tetrahydrodeoxycorticosterone (THDOC). All three increased more rapidly than the well-known steroids progesterone and 17α-hydoxyprogesterone (FDR < 0.05, SAM) (Figures 1C and 1D). By contrast, the top metabolites that decreased during pregnancy (N = 38) were mostly lipids or lipid-like molecules, such as monoacylglycerides (MGs), lysophosphatidylcholines (LPC or lysoPC), and oleoylcarnitine (Figures 1C and 1E). Hierarchical clustering of the weekly samples on the basis of the top 68 altered metabolites revealed a week-order mostly consistent with the actual progression of gestational age (Table S1; Figures 1D and 1E). Intriguingly, most of these metabolite changes rapidly returned to baseline after childbirth (postpartum) (Figures 1B and 1D and 1E). Together, these results suggest a dramatic and programmed change of human blood metabolites at a system level during pregnancy.

Metabolite Groups Altered during Pregnancy

To detect the functional groups of metabolites that change during pregnancy, we performed correlation analysis on the temporal intensity profiles of the top 68 pregnancy-related compounds mentioned above. In Figure S2, metabolites that were significantly increased or decreased tended to cluster together. Using existing structural and biological information, we first categorized the top changing compounds in pregnancy into seven groups. Interestingly, compounds of the same groups tended to cluster together in the correlation matrix. On the basis of the correlation relationship, we constructed a regularized partial correlation network using all pregnancy-related compounds to explore the potential regulatory relationships (Figures 2A and S2B). The topology of the network indicates that different metabolite groups occupied different positions; dense interactions occurred between both inter- and intra- metabolite groups with the densest interactions between central steroid hormones (Figure 2A). These findings highlight that even though the amount of each compound dynamically changes during pregnancy, a highly coordinated metabolite regulatory network underlies the pregnancy process.

We next examined the main clusters that were present in the correlation analysis. Three main clusters emerged from the hierarchical clustering of metabolites (Figure S2A), with a steroid cluster (e.g., antrostane-3,17-diol, estriol-16-glucuronide, progesterone, 17α-hydroxyprogesterone, and THDOC) sitting between the large clusters of lipids and non-lipid molecules. Compared with the other steroids in this cluster that slowly but steadily increased throughout pregnancy, estriol-16-glucuronide exhibited a rapid increase before week 24, (Figures 1D and 2B). Nearly all upregulated metabolites positively correlated with this cluster of steroids, whereas all downregulated metabolites negatively correlated with this cluster (Figure S2A). This result suggests that different steroid hormones might regulate global metabolome dynamics during pregnancy.

Within the lipid cluster, intra-correlation was relatively high. The largest cluster was composed of LysoPCs (Figures 2A and S2A), a class of phospholipids. LysoPCs gradually decreased during pregnancy and increased after childbirth in a pattern that highly correlates with the steroid dehydroepiandrosterone sulfate (DHEA-S) (Figure 2C). LysoPCs are bioactive pro-inflammatory lipids that have been linked with organismal oxidative stress and inflammation (Sevastou et al., 2013). The second-largest cluster of lipids included several free fatty acids that were highly correlated within the cluster (Figures 2A and S2A). Long-chain fatty acids showed intricate dynamics in their amounts revealed by the dense sampling. Hexadecadienoylcarnitine and tetracosahexaenoic acid (THA) decreased at the beginning of pregnancy, followed by waves of increased amounts similar to other fatty acids in the late second and third trimesters (Figure 2D). After childbirth, the amounts of most long-chain fatty acids decreased, except for hexadecadienoylcarnitine (Figure 2D).

Within the non-lipid cluster, one sub-cluster included five highly correlated metabolites belonging to the same caffeine metabolism pathway (Figures 2A and S2A). All five metabolites were consistently elevated during pregnancy, and caffeine reached a concentration three times higher at the end of pregnancy than at the beginning (Figure 2E). This elevation might be due to a slower caffeine metabolism in pregnant women rather than an increase in coffee intake (Knutti et al., 1981). Overall, among the 68 top-altered metabolites in pregnancy, functional metabolite groups (e.g., steroids, LysoPCs, fatty acids, and caffeine metabolites) were altered in an orchestrated manner during pregnancy, and individual compounds within each group showed inter-correlation to each other (Figure 2A).

Orchestrated Metabolome Reconfigurations Span Multiple Pathways during Pregnancy

Next, we longitudinally examined the global pathway changes of all 687 annotated compounds during normal pregnancy. Among the 48 mapped Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, 34 showed significant changes (70.8%, adjusted FDR < 0.05, global test) (Goeman et al., 2004, Xia and Wishart, 2010a) through metaboanalystR (Figure 3A) (Chong et al., 2018), suggesting large-scale pathway changes of metabolism in pregnancy. To quantify the pathway activities through gestational age, we calculated the average intensity of metabolites in the pathways (Figure 3B; see STAR Methods). Among the top altered pathways (Figure 3A), steroid hormone biosynthesis showed elevated activity precisely timed to gestation, peaking before the end of pregnancy and then declining sharply shortly after delivery (Figure 3B). Along with the essential roles of steroid hormones in maintaining pregnancy and later inducing parturition (Mendelson, 2009), we observed an orchestrated elevation of many components centered on progesterone, including some less well-characterized hormones (Figure S3A). Consistent with known sources of pregnancy metabolites (e.g., hormones) (Maltepe and Fisher, 2015), metabolite set enrichment analysis (MSEA) (Xia and Wishart, 2010b) revealed that the adrenal cortex, gonad, and placenta were among the top origins of pregnancy-related metabolites (Figure S3B). The ability to recognize many well-known and less-characterized steroid hormone changes across pregnancy validates our approach.

In addition to the steroid pathway, we observed a dynamic pattern of metabolite changes with pregnancy in other pathways, such as the arachidonic acid metabolism pathway (Figures 3A, 3B, and S3C). We observed 20-HETE amounts increased until week 34; 20-HETE is potentially linked to the regulation of blood pressure and renal function during pregnancy (Wang et al., 2002, Wu et al., 2014) (Figures S3C and S3D). By contrast, 5-HETE amounts generally decreased during pregnancy, potentially associated with its regulation of the uterus (Figures S3C and S3D) (Edwin et al., 1997, Pearson et al., 2010). Thus, beyond energy metabolism and hormones, a system-wide reconfiguration of the metabolome occurs as the mother adapts to pregnancy. In addition, based on MSEA analysis, many pregnancy-related metabolites are implicated in human disease states, including obesity and prepartum depression (Figure 3C) (Xia and Wishart, 2010b).

The Metabolic Clock of Normal Pregnancy Identified by Machine Learning

We next determined whether we can build a metabolic clock based on the high-resolution profile to predict gestational age for individual plasma samples. In the discovery cohort (samples n = 507, subjects N = 21), we applied feature selection (lasso [least absolute shrinkage and selection operator]) with all 9,651 features to build the linear regression model that shows optimal cross-validation performance for predicting a given phenotype in this cohort. We then ran the validation cohort data (test set 1, samples n = 245, subjects N = 9) through the model established in the discovery cohort to measure the independent performance of our model (Figure 4A; see STAR Methods).

We first tested whether the metabolome change can quantitatively determine the gestational age in normal pregnant women. Feature selection in the discovery cohort yielded a linear model that included 42 metabolic features (Figure S4A; Table S2). In the cross-validation test of 507 samples in the discovery cohort, the metabolic model predicted gestational age in weeks (GA_metabolic) that correlated with gestational age estimated by the first-trimester ultrasound (GA_ultrasound, in compliance with the clinical standard of care) with a Pearson correlation coefficient (R) of 0.96 (R² = 0.93, p < 1 X 10⁻¹⁰⁰, root mean squared error [RMSE]= 2.49) (Figure S4B). In the independent-validation cohort, the model yielded a similar R of 0.95 (R² = 0.91, p < 1 X 10⁻¹⁰⁰, RMSE = 2.76, test set 1) (Figure S4C). This indicates metabolic features can accurately predict the gestational age on the basis of a blood sample from a pregnant woman.

For potential clinical use, we next tested whether we can use the annotated compounds in blood to predict the gestational age in pregnant women. We performed feature selection in discovery cohort by using the 264 level 1 and level 2 compounds identified in the Human Metabolome Database (HMDB) in the discovery cohort (Table S3). This yielded a linear model including five compounds (Figures S4D and 4D) that together are highly predictive. We first evaluated the performance of the model in a 10-fold cross-validation (CV) test in the discovery cohort, in which samples were distributed into folds by subject instead of by sample to prevent person-specific information cross-over between the training folds and the test fold. In the CV test, the metabolic-clock model produced a result (GA_metabolic) that correlated with the gestational age estimated by the first-trimester ultrasound (GA_ultrasound) with a Pearson correlation coefficient (R) of 0.92 (R² = 0.85, p = 8e−222, RMSE = 3.67) (Figure 4B). To avoid the hyperparametric selection bias, we further evaluated the performance of our model in two independent validation cohorts (test set 1 and test set 2). In test set 1, the model yielded an R of 0.89 (R² = 0.80, p = 8e−93, RMSE = 4.11) (Figure 4C). The model, including four steroids and one lipid (Figure 4D), was further verified in a second independent-validation cohort of eight individuals with R of 0.91 (R² = 0.83, RMSE = 3.05, samples n = 32, test set 2) (Table 1; Figure 4E). The compound identifications were confirmed by chemical standards (Figures 4F–4H, S4E, and S4F; Table S4; see STAR Methods). We noted that four of these five compounds are among the central steroid cluster forming a dense correlation network with one another (Figure 2A).

As pregnancy progresses toward term, clinical classifications and decisions often need to be made based on the timing of pregnancy (e.g., < 37 weeks for preterm birth). Babies born before 37 weeks are considered preterm, those born before 20 weeks are considered a miscarriage, and those born before 24 weeks have low survival. Therefore, for clinical action it is important to accurately classify the gestational age by clinical cutoff points at weeks 20–37. As a proof-of-principle, we tested the potential use of the metabolome data to classify the normal pregnancy samples as before or after 20, 24, 28, 32, and 37 gestational weeks (Figure S5A). First, using only samples from the third-trimester (> 28 weeks of gestation), the time window where women were more susceptible to preterm delivery, we determined whether the identified maternal blood metabolites can distinguish the sample gestational age as before or after 37 weeks. Both the discovery and the validation prediction yielded an area under the receiver operating characteristics (AUROC) over or close to 0.90 (Figure S5B; see STAR Methods). Remarkably, the prediction model contained only three metabolites, and the abundance range of each individual metabolite separated the > 37 week samples from the < 37 week samples for all but one to two validation subjects (Figures S5C–S5F). Similarly, using samples across the whole pregnancy, we found that metabolites can also accurately distinguish pregnancy samples before or after other important gestational age cutoffs, such as 20, 24, 28, and 32 gestational weeks (Figures S5A and S5G–S5J).

Personal Metabolic Clock of Pregnancy Linked with Timing of Delivery and Fetal Growth

Next, we examined the metabolic clock prediction performance in individuals. First, we noted that for most individuals, our model produced predictions consistently aligned with the gestational age estimated by the first-trimester ultrasound (Figures 5A and 5B). In both cross validations for Discovery and test set 1, the prediction deviation (measured by RMSE) in individuals centered around 3 weeks (Figures S6A and S6B). However, in each dataset, there is a small population of individuals with higher prediction deviation. When we examined these individuals (e.g., subjects 1, 2, and 4), we found the predictions were not more randomly scattered than other individuals. Rather, in the majority of them, predictions shifted away from the actual gestational ages in a portion of the pregnancy duration (Figures 5A and 5B), suggesting effects from non-random causes.

We hypothesized that some of these large prediction deviations might arise from biological causes, particularly from the maternal-fetal interaction. It is reported that the fetoplacental unit secretes hormones in conjunction with fetal growth and development (Murphy et al., 2006). Indeed, we noted that the average prediction deviation strongly correlated with adjusted infant birth weight (Figure 5C, adjusted for gestational length; see Figure S6C and STAR Methods). Thus, the overall metabolic clock tends to outpace the gestational age estimation determined by first-trimester ultrasound in mothers with a heavier fetus while being delayed in the mothers with a lighter fetus. The finding suggests that fetal growth appears to be one of the inputs read by the metabolic clock.

In addition, within the 18 women with natural labor onset (i.e., excluding women with induction before labor onset and scheduled cesarean-section), we found that the women whose overall metabolic clock of pregnancy outpaced ultrasound evaluation tended to deliver earlier, whereas a delay in metabolic clock correlated with a delayed time to child delivery compared to ultrasound estimated due date (Figure 5D). Interestingly, five out of six women (83%) with a metabolic-clock-predicted gestational age more than one week later than the ultrasound-estimated gestational age had natural labor onset after their due date (estimated by ultrasound, marked in red in Figure 5D), and four out of five women (80%) with metabolic-clock-predicted gestational age more than one week earlier than the ultrasound-estimated gestational age had natural labor onset before their due date (marked in blue in Figure 5D). These results suggest the metabolic clock of pregnancy with maternal metabolites contains information on the timing of delivery in normal pregnancy.

Prediction for Timing of Delivery

We then tested whether the maternal blood metabolites can also predict the timing of a normal delivery event within a defined period (2, 4, and 8 weeks from delivery) approaching the labor events (in the third trimester). We first examined whether metabolites can predict a delivery within 2 weeks (weeks to delivery [WD] < 2w). To predict delivery triggered naturally without outside procedures (such as scheduled cesarean-section), we only included delivery events naturally triggered (subjects N = 18, samples n = 193). With just three metabolites, the metabolome accurately predicted an upcoming delivery event within 2 weeks in both discovery and validation cohorts with AUROC close to 0.9 (Figures 5E–5H, S6D, and S6E; see STAR Methods). Similarly, identified metabolites can also be used to predict the timing of a normal delivery event within 4 and 8 weeks (Figures 5E and S6F–S6I). Intriguingly, the panels of metabolites partially overlapped between the models, whereas the individual metabolites contributed differently to the models (Figure 5E). All of the metabolite markers were identified as steroids, except for phospholipid PE(P-16:0e/0:0), and most of them (three out of five in total) also appeared in the aforementioned metabolic clock for gestational age (Figure 5E; Table S4). These results demonstrate that we can precisely categorize critical pregnancy stages in normal subjects by using a small number of maternal blood metabolites, which can be further validated in larger and independent cohorts.

Key Resources Table

Resource Availability

Experimental Model and Subject Details

Method Details

Quantification and Statistical Analysis