Early prediction and longitudinal modeling of preeclampsia from multiomics

Abstract

Preeclampsia is a complex disease of pregnancy whose physiopathology remains unclear. We developed machine-learning models for early prediction of preeclampsia (first 16 weeks of pregnancy) and over gestation by analyzing six omics datasets from a longitudinal cohort of pregnant women. For early pregnancy, a prediction model using nine urine metabolites had the highest accuracy and was validated on an independent cohort (area under the receiver-operating characteristic curve [AUC] = 0.88, 95% confidence interval [CI] [0.76, 0.99] cross-validated; AUC = 0.83, 95% CI [0.62,1] validated). Univariate analysis demonstrated statistical significance of identified metabolites. An integrated multiomics model further improved accuracy (AUC = 0.94). Several biological pathways were identified including tryptophan, caffeine, and arachidonic acid metabolisms. Integration with immune cytometry data suggested novel associations between immune and proteomic dynamics. While further validation in a larger population is necessary, these encouraging results can serve as a basis for a simple, early diagnostic test for preeclampsia.

Keywords: biomarkers; machine learning; multiomics; predictive modeling; preeclampsia.

Figure 1

Overview of the study (A) Two independent cohorts were analyzed using six different omics. (B) Sample collection timeline for plasma in our discovery and validation cohorts. Circles indicate pre-delivery sample collection times, and inverted triangles indicate delivery dates for individual women (one per horizontal line).

Figure 2

Prediction models in early pregnancy Samples obtained in the first 16 weeks of pregnancy were used. (A) Performance comparison of EN models derived from different omics in terms of the AUC. The integrated (stacked) model utilizing stacked regression exhibited the highest accuracy (AUC = 0.94, 95% CI [0.86, 1]). Among omics sets the urine metabolomic model (AUC = 0.88, 95% CI [0.76, 0.99]) and plasma proteome (AUC = 0.87, 95% CI of [0.75, 0.99]) performed best. (B) Heatmap of ranked values of features identified by EN, perfectly distinguishing preeclamptic from normotensive women.

Figure 3

Urine metabolome prediction model using nine metabolites sampled early in gestation validates on the validation cohort Samples obtained in the first 16 weeks of pregnancy are used. (A) AUC = 0.83, 95% CI [0.62, 1] and prediction values (scores) obtained by EN for preeclamptic (PE) and normotensive women. (B) Metabolites identified by EN as biomarkers of preeclampsia. y axis shows the value in early pregnancy stratified by normotensive (gray) and preeclamptic (light blue) pregnancies. p values obtained using Wilcoxon signed-rank univariate analysis show statistical significance of each protein (∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001).

Figure 4

An integrated multiomics machine model outperforms single omics models for preeclampsia (A) Cross-validated performance of machine-learning models in terms of the AUC is shown on the y axis. Each model was obtained using all available samples over gestation. The integrated (stacked) model utilizing stacked regression exhibited the highest accuracy (AUC = 0.91, 95% CI [0.85, 0.97]). Both proteome and metabolome (urine) had high prediction performance (AUC = 0.89, 95% CI [0.83, 0.95] proteome; AUC = 0.87, 95% CI [0.80, 0.94] urine metabolome). (B) Urine metabolome prediction model using ten metabolites sampled over gestation validates on the validation cohort. AUC = 0.874, 95% CI [0.76, 0.99], and prediction values (scores) obtained by EN for normotensive and preeclamptic (PE) women. (C) Metabolites identified by EN as biomarkers of preeclampsia over gestation. y axis shows values stratified by normotensive (gray) and preeclamptic (light blue) pregnancies. p values obtained using linear mixed-effects univariate analysis show statistical significance of each metabolite (p < 0.05). (D) Proteins identified by EN as biomarkers of preeclampsia over gestation. y axis shows a protein value stratified by normotensive (gray) and preeclamptic (light blue) pregnancies. p values obtained using linear mixed-effects univariate analysis show statistical significance of each metabolite (p < 0.05).

Figure 5

Visualization of predictive features of the transcriptome (yellow), proteome (orange), urine metabolome (dark blue), and plasma metabolome (light blue) Features obtained using all available samples over gestation. Vertices represent features selected by EN laid out using t-SNE. Edges are drawn between features with Spearman correlation >0.55 clearly illustrating high correlations between different omics sets. Size of each node is proportional to the frequency at which it was chosen in prediction models during cross-validation. High frequency of occurrence indicates that a feature is relevant for all or majority of patients, resulting in a more stable model.

Figure 6

Correlation between predictive immune features and multiomics model (A) Previously identified predictive immune features strongly correlate with the multiomics predictive model. Visualization shows features most correlated with the prediction of the stacked model. Features shown in orange are the seven most predictive immunome features that also highly correlate with the multiomics predictive model. Size of each node is proportional to the −log₁₀(p value) of Spearman correlation. (B) Comparison of p value of correlation for the top immune and top proteome features. Each node is a pair comprising an immune and a proteome feature.

Figure 7

Identified enriched pathways from urine metabolome urine over gestation and in early pregnancy (A) Pathway enrichment analysis over gestation using metabolites from urine that were significant (FDR < 0.05, Wilcoxon signed-rank test with Benjamini-Hochberg procedure). Pathways shown above the dotted line were significant (p < 0.05). (B) Pathway enrichment analysis for early pregnancy using metabolites from urine that were significant (FDR < 0.05, linear mixed-effects model with Benjamini-Hochberg procedure). The color and the size of a circle are proportional to the −log(p) and pathway impact value, respectively, where p denotes a p value.