Abstract
Maternal morbidity and mortality continue to rise, and pre-eclampsia is a major driver of this burden1. Yet the ability to assess underlying pathophysiology before clinical presentation to enable identification of pregnancies at risk remains elusive. Here we demonstrate the ability of plasma cell-free RNA (cfRNA) to reveal patterns of normal pregnancy progression and determine the risk of developing pre-eclampsia months before clinical presentation. Our results centre on comprehensive transcriptome data from eight independent prospectively collected cohorts comprising 1,840 racially diverse pregnancies and retrospective analysis of 2,539 banked plasma samples. The pre-eclampsia data include 524 samples (72 cases and 452 non-cases) from two diverse independent cohorts collected 14.5 weeks (s.d., 4.5 weeks) before delivery. We show that cfRNA signatures from a single blood draw can track pregnancy progression at the placental, maternal and fetal levels and can robustly predict pre-eclampsia, with a sensitivity of 75% and a positive predictive value of 32.3% (s.d., 3%), which is superior to the state-of-the-art method2. cfRNA signatures of normal pregnancy progression and pre-eclampsia are independent of clinical factors, such as maternal age, body mass index and race, which cumulatively account for less than 1% of model variance. Further, the cfRNA signature for pre-eclampsia contains gene features linked to biological processes implicated in the underlying pathophysiology of pre-eclampsia.
Main
The period from conception to delivery represents the most rapid growth and development in an individual’s life. The ability to support this development requires dramatic and poorly understood alterations in maternal physiology. Research into human pregnancy has clear ethical constraints, and the unique character of human gestation has limited deeper understanding of the physiology and pathophysiology of pregnancy3. Haemochorial placentation is found among many mammalian species; however, in humans, it involves a unique degree of trophoblastic invasion4,5, and because pre-eclampsia occurs predominantly in humans, conventional animal models are of limited value6,7. Pre-eclampsia, a condition marked by maternal endothelial dysfunction and associated new-onset maternal hypertension, complicates up to 1 in 12 pregnancies and is a significant cause of maternal morbidity and higher lifetime risk of cardiovascular disease1.
Here we demonstrate the ability of cfRNA transcripts to establish the normative responses of both maternal and fetal tissues characteristic of normal pregnancy progression. By implication, deviation from normative cfRNA expression patterns should allow the prediction of impending pathology before its presentation. We demonstrate the use of cfRNA to characterize women at risk of pre-eclampsia months before diagnosis. Notably, the cfRNA profiles identify risk solely through molecular mechanisms common to pre-eclampsia and are therefore exclusive of clinical variables such as race, body mass index (BMI), maternal comorbidities and/or obstetrical history.
In this study, we gather the largest and most diverse dataset of maternal transcriptomes to date. Samples were drawn from eight prospectively collected cohorts that provided n = 2,539 plasma samples from n = 1,840 pregnancies for women of multiple ethnicities, nationalities, geographic locations and socioeconomic contexts, while covering a range of gestational ages (Fig. 1a). The broad sociodemographic spectrum of our data (Table 1 and Supplementary Table 1) enabled us to test the applicability of maternal transcriptomes at one gestational time point. A detailed description of each cohort and the methodology is available in the Supplementary Information.
RNA signal independent of clinical factors
Ultrasound-based gestational age has long been used as a surrogate measure of pregnancy progression. Here, we show that a cfRNA signature is as accurate a measure of gestational age while also providing insights into the biology of pregnancy progression. As a first step to develop a machine learning model, we divided our data from all full-term pregnancies without complications into a training set (n = 1,908 samples) and a test set (n = 474 samples), stratified by gestational age so that all age strata were represented proportionally. Before modelling, we standardized the means of gene counts across all cohorts (Methods and Extended Data Fig. 5). A Lasso linear model was fitted to predict gestational age in the training set, with a test set performance of a mean absolute error of 14.7 days (Fig. 1b, Extended Data Fig. 6 and Supplementary Data 1), referencing to first-trimester fetal ultrasound biometry. Overall, the error of our model is equivalent to that of second-trimester ultrasound and superior to that with third-trimester ultrasound8, and could provide an alternative dating procedure for women who start prenatal care later in pregnancy.
Next, we explored whether inclusion of clinical variables altered model performance. By analysis of variance (ANOVA), we showed that the model was driven almost entirely by information from the cfRNA transcripts, with BMI, maternal age and race accounting for less than 1% of variance (Fig. 1c). Rebuilding the gestational age model including maternal race, BMI and age provided no improvement in accuracy (0.07 days, not significant by bootstrap test).
Fetal signatures in maternal circulation
As the cfRNA signatures for gestational age demonstrated a dynamic change in transcripts as pregnancy progresses, we then explored whether transcripts found in the maternal circulation during pregnancy could be linked to their tissue of origin. Specifically, we sought to ascertain whether the molecular status of the placenta, fetal organs and/or maternal tissues (cervix and/or uterus) could be assessed by examining cfRNA profiles. While fetal cells are known to pass into the maternal circulation9,10, individual transcripts from the fetus or fetal cell types are relatively rare in maternal plasma; thus, we investigated these signals by analysing gene sets from Gene Ontology11 or the Molecular Signatures Database12,13. Using longitudinal data from cohort H covering 93 women sampled four times during pregnancy (Supplementary Information), we first confirmed that we could identify pregnancy-related sets such as those for gonadotropin and oestrogen pathways (Extended Data Fig. 1) and that the signal from the gestational age model increased with gestational age as did signal from the placenta (Fig. 2a, b and Methods). We show that hundreds of independently identified gene sets in maternal blood mirror the maternal and fetal physiological changes expected during pregnancy. Specifically, using single-cell RNA-seq data from adult and fetal organs (Supplementary Table 2), we were able to confirm changes in fetal gene sets, including those involved in fetal heart development, in maternal blood (Fig. 2c). Furthermore, the cfRNA profiles reflect expected changes in maternal tissues, such as the uterus and cervix, with progressively increasing expression of collagen and extracellular matrix gene sets14 (Fig. 2d). Extended Data Fig. 2 shows additional examples of fetal gene sets, including those of nephron progenitor cells for which expression become less abundant with gestational age in accordance with a decrease in the nephrogenic zone width15,16 and those in the gastrointestinal tract, where the oesophagus develops early with associated gene expression decreasing later versus small intestine where associated gene expression shows a steady increase17.
To test whether the identified gene sets were uniquely associated with pregnancy progression, we next compared the observed gestational age collection time labels to a set of randomly permuted collection time labels. This comparison verified that all selected gene sets were associated with pregnancy progression (Extended Data Fig. 3). The directional signals could be confirmed in three independent cohorts (n = 351 women) for which longitudinal data were available (Fig. 2e–h). In all cases, the slopes for the gestational age coefficients were distinct from 0 at a 0.05 confidence level. In total, we tested 793 gene sets from single-cell analyses12,13, comprising 384 gene sets from adult and 409 gene sets from fetal tissues. Of these, 129 gene sets (55 fetal) were significantly correlated with gestational age, of which 99 gene sets (40 fetal) showed increased signal and 30 gene sets (15 fetal) showed decreased signal as a function of gestational age at collection in cohort H, and were confirmed in at least two other cohorts with longitudinally sampled individuals (Supplementary Data 2). As changes in these predefined gene sets were only significant in the context of gestational age across at least three cohorts with longitudinal information, we present here a non-invasive window into maternal–fetal development from a maternal blood sample.
Early prediction of pre-eclampsia
Having established that cfRNA profiles can reveal and characterize molecular changes in the maternal–placental–fetal unit over gestation, it is likely that disruption of these pathways might identify women at risk for adverse pregnancy outcomes such as pre-eclampsia.
We evaluated the ability of cfRNA signatures in maternal blood, during the second trimester (16–27 weeks), to predict the development of pre-eclampsia. Maternal blood draws occurred, on average, 14.5 weeks (s.d., 4.5 weeks) before delivery (Fig. 3a); in contrast to work by Munchel et al.18 where plasma was collected at the time of diagnosis, the gestational age time points in our analysis correspond to timepoints where women are asymptomatic. A case–control study with 72 cases of pre-eclampsia and 452 non-cases selected from two independent cohorts (cohorts A and E) was performed (Supplementary Information). Cohort E included 31 controls with chronic hypertension and 19 controls with gestational hypertension and both cohorts included spontaneous preterm birth samples along with the normotensive term controls. Pre-eclampsia was defined by criteria consistent with those from the 2013 Task Force on Hypertension in Pregnancy (ACOG 2013), and each case was adjudicated by two board-certified physicians. As before, a cohort correction was applied before modelling.
Two-sided Spearman correlation tests identified signatures that separated the cases and controls; in each round of cross-validation, we retained features with an adjusted P value below 0.05 (Methods) and consistently identified seven genes: CLDN7, PAPPA2, SNORD14A, PLEKHH1, MAGEA10, TLE6 and FABP1 (Fig. 3b).
Four of the genes selected for modelling have functions relevant to pre-eclampsia or placental development. PAPPA2, encoding pregnancy-associated plasma protein 2, is expressed in the placenta19, specifically in trophoblast cells. It has previously been linked to the development of pre-eclampsia and has been associated with inhibition of trophoblast migration, invasion and tube formation20,21. Claudin 7 (CLDN7) is involved in tight cell junction formation and blastocyst implantation; in healthy pregnancies, expression of CLDN7 is reduced in response to oestrogen at the time of implantation22,23. Similarly, TLE6 has also been linked to preimplantation and early embryonic lethality24. Fatty acid-binding protein 1 (FABP1) was first purified from human cytotrophoblasts and is known to be highly expressed in the fetal liver; it is critical for fatty acid uptake and transport25 and is upregulated threefold when cytotrophoblasts differentiate to syncytiotrophoblasts at implantation26. The other three genes that make up the pre-eclampsia cfRNA signature (SNORD14A, PLEKHH1 and MAGEA10) have been associated with pre-eclampsia through bioinformatic analyses, although their function is less well understood27,28. Two of the identified genes, PAPPA2 and FABP1, were also identified in the gestational age model and highlight the imbalance in cfRNA signatures between pregnancy progression and pathology.
On the basis of these identified gene features, a logistic regression model in a leave-one-out cross-validation set-up was used to estimate the probability of pre-eclampsia. This model framework was chosen on the basis of learning curve analyses (Methods and Extended Data Fig. 7). At a sensitivity of 75%, our cfRNA model achieved a positive predictive value (PPV) of 32.3% (s.d., 3%) given a prevalence of pre-eclampsia of 13.7% in our study, superior to PPVs reported from current clinical state-of-the-art models, which are driven largely by maternal factors2 ; the area under the curve (AUC) for the model was 0.82 (95% confidence interval, ±0.06; Fig. 3c). Consistent with our findings with the gestational age model, inclusion of clinical variables (maternal BMI, age and race) had no effect on performance, as the classifier assigns zero weight to these clinical variables and they explain <1% of the variance based on ANOVA analyses. The lack of contribution to cfRNA profiles from clinical factors highlights the generalizability of these profiles to diverse populations.
When comparing gestational age at delivery between test-positive and test-negative individuals, a significant shift was found in the timing of delivery, with the test-positive population delivering earlier during gestation (P < 2 × 10–7; Fig. 3d). A positive test correctly identified 73% of individuals destined to have a medically indicated preterm birth over 3 months in advance of the onset of clinical symptoms or delivery.
To further understand molecular signature changes and how they might reflect the pathophysiology driving pre-eclampsia, we performed pathway analysis. The top upregulated pathways were dominated by structural cell functions, including placental blood vessel development, artery morphogenesis and embryonic placental development (Extended Data Fig. 4a), while the majority of downregulated pathways were related to immune pathways (Extended Data Fig. 4b). Both the upregulated and downregulated gene sets aligned with the accepted mechanism of pathogenesis for pre-eclampsia29.
In cohort E, the non-case group contained both normotensive women (n = 263) and women with chronic (n = 31) or gestational (n = 19) hypertension. Genes identified through comparison of the groups with chronic or gestational hypertension with the normotensive group showed no overlap with genes significant for pre-eclampsia (two-sided Spearman correlation test, P < 0.05). Additionally, no genes were differentially expressed in the chronic or gestational hypertensive groups when compared with the normotensive group. While others have published studies designed to determine the effect of hypertension more generally on gene expression (e.g., Zeller et al.30), here, we demonstrate that the signal for pre-eclampsia is specific to hypertension driven by a placental disorder and the signature is independent of signals associated with chronic hypertension. Clinically, it can be quite challenging to differentiate superimposed pre-eclampsia in women with pre-existing hypertension from exacerbation of baseline chronic hypertension. This difference is important, as one requires delivery for cure while the other usually does not.
As pre-eclampsia and spontaneous preterm birth are theorized to have some overlapping molecular pathways31,32, we tested whether excluding non-case samples with deliveries before gestational week 37 (n = 85) would affect test prediction. Removal of spontaneous preterm delivery samples did not alter the performance of the model (AUC = 0.79; 95% confidence interval, ±0.06), suggesting that inclusion of spontaneous preterm birth samples in the non-case group does not affect the pre-eclampsia classifier.
We report a standalone molecular predictor that has the potential to be an early detector of pre-eclampsia with a PPV of 32% that is based entirely on transcripts and is exclusive of clinical variables. This predictor contrasts with state-of-the-art methods, which are dependent on clinical factors and achieve a PPV of 4.4%2….
