Metabolomic profiling of preterm birth in pregnant women living with HIV

Abstract

Background: Preterm birth is a leading cause of death in children under the age of five. The risk of preterm birth is increased by maternal HIV infection as well as by certain antiretroviral regimens, leading to a disproportionate burden on low- and medium-income settings where HIV is most prevalent. Despite decades of research, the mechanisms underlying spontaneous preterm birth, particularly in resource limited areas with high HIV infection rates, are still poorly understood and accurate prediction and therapeutic intervention remain elusive.

Objectives: Metabolomics was utilized to identify profiles of preterm birth among pregnant women living with HIV on two different antiretroviral therapy (ART) regimens.

Methods: This pilot study comprised 100 mother-infant dyads prior to antiretroviral initiation, on zidovudine monotherapy or on protease inhibitor-based antiretroviral therapy. Pregnancies that resulted in preterm births were matched 1:1 with controls by gestational age at time of sample collection. Maternal plasma and blood spots at 23-35 weeks gestation and infant dried blood spots at birth, were assayed using an untargeted metabolomics method. Linear regression and random forests classification models were used to identify shared and treatment-specific markers of preterm birth.

Results: Classification models for preterm birth achieved accuracies of 95.5%, 95.7%, and 80.7% in the untreated, zidovudine monotherapy, and protease inhibitor-based treatment groups, respectively. Urate, methionine sulfone, cortisone, and 17α-hydroxypregnanolone glucuronide were identified as shared markers of preterm birth. Other compounds including hippurate and N-acetyl-1-methylhistidine were found to be significantly altered in a treatment-specific context.

Conclusion: This study identified previously known as well as novel metabolomic features of preterm birth in pregnant women living with HIV. Validation of these models in a larger, independent cohort is necessary to ascertain whether they can be utilized to predict preterm birth during a stage of gestation that allows for therapeutic intervention or more effective resource allocation.

Keywords: Dried blood spots; Metabolomics; Plasma; Preterm birth; Women living with HIV; Zidovudine.

Fig. 1

Study design. Schematic showing exclusion criteria and the number of samples used for analysis of maternal plasma, maternal DBS, and infant DBS samples

Fig. 2

Maternal metabolites in preterm birth. a Principal components analysis of maternal metabolite profiles using Euclidean distances. Ellipses show 95% confidence areas for the treatment regimens as marked. Numbers in brackets denote percent of overall variation explained by each component. b Coefficients from linear regression analysis of maternal metabolites stratified by treatment regimen. Only metabolites that were significant in any single analysis are shown. Metabolites with positive estimates are increased in women who deliver preterm and metabolites with negative estimates are increased in women who deliver at term. Error bars denote 95% confidence intervals. Values in red are statistically significant with FDR-adjusted p < 0.05

Fig. 3

Maternal signatures of preterm birth. Random forests (RF) model for preterm birth in a untreated women, b women on zidovudine monotherapy, and c women on PI-ART. Features shown represent the sparse set selected by cross-validation and are ordered by decreasing importance in the RF model as indicated by shaded boxes on the left. Points and error bars show the coefficients and 95% confidence intervals from linear regression analysis of the same metabolite. Metabolites with positive estimates are increased in women who deliver preterm and metabolites with negative estimates are increased in women who deliver at term. Values in red were statistically significant in the linear regression analysis

Fig. 4

Selected metabolite abundances. Boxplots of normalized metabolite abundance values for selected features. Bold lines indicate medians, whiskers indicate 1.5*IQR (interquartile range) from first and third quartiles, and points indicate individual sample values. Statistically significant comparisons are marked with * FDR-adjusted p < 0.05, **FDR-adjusted p < 0.01 for comparison of PTB versus term delivery by treatment group (small brackets) or all PTB versus term delivery (large bracket)

Fig. 5

Infant signatures of preterm birth. Random forests (RF) model for preterm birth using infant DBS-derived metabolite profiles. a Receiver-operator characteristic curves for classification models of birth status for each treatment regimen as indicated. Numbers in parentheses indicate the area under the ROC curve (AUC). b–c RF models for preterm birth in infants exposed to (b) zidovudine monotherapy and c PI-ART in utero. Features are ordered by decreasing importance in the RF model as shown by shaded boxes on the left. Points and error bars show the coefficients and 95% confidence intervals from linear regression analysis of the same metabolite. Metabolites with positive estimates are increased in preterm infants and metabolites with negative estimates are increased in term infants