Plasma lipids are dysregulated preceding diagnosis of preeclampsia or delivery of a growth restricted infant

Summary

Background

Lipids serve as multifunctional metabolites that have important implications for the pregnant mother and developing fetus. Abnormalities in lipids have emerged as potential risk factors for pregnancy diseases, such as preeclampsia and fetal growth restriction. The aim of this study was to assess the potential of lipid metabolites for detection of late-onset preeclampsia and fetal growth restriction.

Methods

We used a case-cohort of 144 maternal plasma samples at 36 weeks’ gestation from patients before the diagnosis of late-onset preeclampsia (n = 22), delivery of a fetal growth restricted infant (n = 55, defined as <5th birthweight centile), gestation-matched controls (n = 72). We performed liquid chromatography-tandem mass spectrometry (LC-QQQ) -based targeted lipidomics to identify 421 lipids, and fitted logistic regression models for each lipid, correcting for maternal age, BMI, smoking, and gestational diabetes.

Findings

Phosphatidylinositol 32:1 (AUC = 0.81) and cholesterol ester 17:1 (AUC = 0.71) best predicted the risk of developing preeclampsia or delivering a fetal growth restricted infant, respectively. Five times repeated five-fold cross validation demonstrated the lipids alone did not out-perform existing protein biomarkers, soluble tyrosine kinase-1 (sFlt-1) and placental growth factor (PlGF) for the prediction of preeclampsia or fetal growth restriction. However, lipids combined with sFlt-1 and PlGF measurements improved disease prediction.

Interpretation

This study successfully identified 421 lipids in maternal plasma collected at 36 weeks’ gestation from participants who later developed preeclampsia or delivered a fetal growth restricted infant. Our results suggest the predictive capacity of lipid measurements for gestational disorders holds the potential to improve non-invasive assessment of maternal and fetal health.

Funding

This study was funded by a grant from National Health and Medical Research Council.

Research in context.

Evidence before this study

Previous lipidomics research has linked dysregulated plasma lipids with increased risk of pregnancy complications. However, none of these studies have used lipidomics to predict pregnancy disorders before they are diagnosed.

Added value of this study

Our study assesses the predictive potential of lipid metabolites using targeted lipidomics in a case cohort of 144 maternal plasma samples collected at 36 weeks’ gestation, prior to diagnosis of pregnancy complication. We strengthened these results by comparing our lipid biomarkers to current protein biomarkers sFlt-1 and PlGF. Further, we combined these protein markers with the best performing lipids to yield an improved predictive test.

Implications of all the available evidence

Our findings suggest preliminary evidence that lipidomics has the capacity to generate lipid profiles for the prediction of gestational disorders. Lipids hold significant potential as a non-invasive assessment of maternal and fetal health.

Introduction

Preeclampsia and fetal growth restriction are diseases of pregnancy that have severe consequences for mother and child, during and after pregnancy.¹ Many risk factors for pregnancy diseases, such as increased body mass index (BMI), advanced maternal age, smoking and diabetes are well-established.² Unfortunately, there are no biomarkers that can accurately predict if a woman is destined to develop preeclampsia or fetal growth restriction late in pregnancy.

Current clinical biomarkers such as soluble tyrosine kinase-1 (sFlt-1) and placental growth factor (PlGF) are proteins used to detect patients with preterm preeclampsia.³ Although these molecules perform well as a rule out test for preterm preeclampsia, a test to rule in patients who will develop preeclampsia or fetal growth restriction close to term is desperately needed.⁴ Indeed, preeclampsia occurring near term gestations is more common than preterm disease, and fetal growth restricted fetuses are at greatest risk of stillbirth.

Lipids serve as multifunctional metabolites that have important implications for a healthy pregnancy, with roles in cellular signalling, and membrane structure.^5,6 Abnormalities in lipid metabolites have emerged as potential risk factors for pregnancy diseases, like preeclampsia, where alterations in the lipidome are responsible for cellular adaptations in a disease state.⁷ Multiple studies have recognised a high abundance of dyslipidaemia in pregnancies complicated by preeclampsia,^8,9 however no studies have successfully identified lipids as biomarkers to predict preeclampsia or fetal growth restriction.

Lipidomics offers the potential to measure lipids in small sample volumes to correctly identify subtle changes in the maternal circulation.¹⁰ This technology can be conducted through an untargeted (analysis of all detectible lipids) and targeted (defined group of lipids) approach. Whilst both methods offer the potential for a broad range of lipids to be explored, the targeted lipidomics technique allows profiling and quantification of specific lipids of interest in a given population.¹¹ With targeted lipidomics technology, there is potential to discover biomarkers that can be used to predict pregnancy disorders before they are diagnosed. This study aims to assess the predictive potential of lipid metabolites using a targeted lipidomics approach for detection of pregnancy complications occurring near term gestations.

Methods

BUMPS cohort: plasma collection

The Biomarker and Ultrasound Measures for Preventable Stillbirth (BUMPS) study is a large prospective study conducted at the Mercy Hospital for Women, Melbourne, Australia.^12,13 This study involved the collection of blood at 36 weeks’ gestation (35⁺⁰–37⁺⁰) a time the pregnant women already undergo routine blood tests. At the time of the collection, the women were not clinically suspected of preeclampsia, and their disease developed, and was diagnosed after the blood test. Whole blood was collected in 9 mL ethylenediaminetetraacetic acid tubes, centrifuged at 1000 g for 10 min, and plasma was stored at −80 °C until required.

Ethics

Ethical approval was obtained from the Mercy Health and Human Research Ethics Committee (Approval number: 2019–012) and participants gave informed, written consent.

Participant characteristics

Preeclampsia was diagnosed according to the American College of Obstetricians and Gynaecologists (ACOG) guidelines.¹⁴ From the first 1000 BUMPS participants, a case cohort of 144 samples were selected for analysis, including 49 who delivered infants with fetal growth restriction (birthweight <5th centile according to the GROW software),¹⁵ 17 participants who developed preeclampsia, 6 participants who developed preeclampsia and delivered an infant with fetal growth restriction, and 72 randomly selected controls. English speaking women aged 18 years and over, with a singleton pregnancy and normal mid-trimester fetal morphology examination were eligible to participate. Women with multiple pregnancies and ultrasound greater than 42 weeks’ gestation were ineligible to participate in the study. Patient characteristics are shown in Table 1.

Table 1.

Maternal characteristics and pregnancy outcomes for the biomarker and ultrasound measures for preventable stillbirth (BUMPS) cohort.

	Controls (n = 72)	Preeclampsia (n = 23)	FGR (n = 55)	p value
Clinical characteristics
Gestation at collection (weeks) Median [IQR]	36.1 [35.7–36.4]	36.3 [35.6–36.6]	36.1 [35.7–36.4]	0.90
Maternal age (years) Median [IQR]	33.0 [31.0–35.0]	34.0 [32.0–37.0]	33.0 [30.5–36.0]	0.32
BMI (kg/m2) Median [IQR]	24.6 [22.2–27.7]	27.9 [24.1–30.7]	25.1 [22.7–29.4]	0.10
Birth outcomes
Birthweight (grams) Median [IQR]	3589 [3291–3825]	3150 [2550–3450]	2620 [2450–2805]	<0.0001
Crown-heel length (cm) Median [IQR]	51.0 [50.0–52.0]	51.0 [47.0–52.0]	48.0 [46.5–49.5]	<0.0001
Fetal sex No. (%)
Male	32 (44.4%)	11 (47.8%)	22 (44.9%)	0.96
Female	40 (55.6%)	12 (52.2%)	27 (55.1%)
Nulliparous No. (%)
Yes	37 (51.4%)	19 (82.6%)	32 (65.3%)	0.02
No	35 (48.6%)	4 (17.4%)	17 (34.7%)
Smoking No. (%)
Smoker	6 (8.3%)	2 (8.7%)	6 (12.2%)	0.76
Non-smoker	66 (91.7%)	21 (91.3%)	43 (87.8%)
GDM No. (%)
Yes	10 (13.9%)	6 (26.1%)	4 (8.2%)	0.12
No	62 (86.1%)	17 (73.9%)	45 (91.8%)

BMI, Body mass index; FGR, Fetal growth restriction. Data presented as median [25th–75th percentile] and as number (%) if categorical. Kruskal–Wallis tests were used for continuous data. Chi-square tests were used for categorical data. FGR is defined as an infant birthed <5th birthweight centile. BMI data missing for 1/72 control samples and 1/55 FGR samples.

The maternal age, BMI, smoking status, and gestational diabetes status for each sample was recorded. If BMI was not available, the population average of 26 for a pregnant woman of Australian European ethnicity was used. Smoking status was coded as positive for smoking at any point during pregnancy and negative otherwise. Gestational diabetes was coded as positive for any form of gestational diabetes (either diet or insulin controlled) and negative otherwise.

Lipid extraction

Lipid extraction and targeted lipidomics analyses were performed by Metabolomics Australia (Melbourne, Australia). First, 20uL of plasma was mixed with 380uL of extraction solvent containing chloroform:methanol (2:1) containing a mixture of internal standards (PC 19:0 19:0, LPC 17:1 and TG-d5). Samples were then vortexed for 30 s and placed in a thermomix at room temperature, for 5mins. Samples were centrifuged at a speed of 15,000 rpm for 10 min. The supernatant was evaporated under SpeedVac to complete dryness (100 μl per cycle). The extracts were reconstituted in 90 μL methanol:water-saturated butanol (1:9) and transferred into compatible vials with glass inserts for subsequent analysis. Samples were deidentified and randomized prior to analysis and pooled biological quality control samples (pbQCs) were prepared by pooling 10 μl of all samples lysate and run every 5 samples to monitor instrument reproducibility.

Lipid measurement

Analyses of plasma lipids was performed on the Agilent 6490 LC-QQQ mass spectrometer interfaced with an Agilent 1290 series HPLC system. Separation of lipids were conducted on a ZORBAX eclipse plus C18 column (2.1 × 100 mm 1.8 mm, Agilent technologies) with the temperature of 60 °C with solvent A and B made up of water, acetonitrile, and isopropanol (50:30:20 and 1:9:90 respectively) with 10 mM ammonium formate.

Targeted mass spectrometry analysis was conducted in ESI positive ion mode with dynamic scheduled multiple reaction monitoring. Mass spectrometry settings and multiple reaction monitoring transitions for each lipid subclass and individual species was run as previously described.^16,17

Mass spectrometer running conditions were as follows: gas temperature 150 °C, gas flow rate 17 l/min, nebulizer gas pressure 20 psi, sheath gas temperature 200 °C, capillary voltage 3500 V and sheath gas flow 10 l/min. Isolation widths for Q1 and Q3 were set to unit resolution (0.7 amu) for both Q1 and Q3. QC samples were analysed with batch samples to monitor sample extraction efficiency. Data processing of each lipid species was performed using Agilent's Mass Hunter Quantitative Analysis (QQQ) software (Agilent Technologies Australia) to determine the area of the resultant chromatograms for the lipid species.

Measurement of plasma protein levels

Plasma placental growth factor (PlGF) and soluble FMS-like tyrosine kinase-1 (sFlt-1) were measured using commercial electrochemiluminescence immunoassay platform (cobas e 411 analyser, Roche Diagnostics). sFlt-1 and PlGF measurements for one fetal growth restriction sample could not be obtained.

Statistics

In 144 samples collected, a total of 468 lipids were quantified using mass spectrometry. In total, 4 samples contained missing or negative measurements for a single lipid. Missing/negative measurements were imputed using sample-wise k-nearest neighbours.¹⁸ The relative standard deviation (RSD) of each lipid was calculated and the lipids with the lowest 10% RSD were removed.¹⁹ The remaining 421 lipids were log transformed, mean centred and normalised by the standard deviation to produce a set of pre-processed lipid measurements for further analysis.

Heatmaps coupled with dendrograms were produced to visualise sample clusters illustrating the similarity or dissimilarity. The top 25 differentially expressed lipids were presented on the heatmap. Principal component analysis (PCA) was performed to visualise distinct groupings within the data.

To determine significantly dysregulated lipids, a set of logistic regression models were fitted where the independent variables included the pre-processed lipid measurements, maternal age, BMI, smoking status and gestational diabetes status and the dependent variable was either preeclampsia or fetal growth restriction. For each model, the p value for the coefficient representing the lipid term was calculated. Benjamini-Hochberg correction was used to compute q values from the p value stack, with significantly dysregulated lipids reported using a false discovery threshold of q < 0.1.

Univariate receiver operating characteristic (ROC) curves were constructed to determine the diagnostic power of significantly dysregulated lipids. ROC curves were also generated for combinations of lipids established in univariate analysis. Model selection for the optimal predictive model between single and combined lipid models were performed using the Akaike information criterion (AIC). Regression model for preeclampsia prediction included case controls and patients who developed preeclampsia. Regression model for fetal growth restriction included case controls and patients who birthed a growth restricted infant.

Finally, we compared the predictive performance of the optimal lipid models for fetal growth restriction and preeclampsia prediction against known angiogenic biomarkers, sFlt-1 and PlGF. Five times repeated five-fold cross validation of each logistic regression model was used to estimate the performance of the best lipids model compared to sFlt-1 and PlGF. Five times repeated five-fold cross validation was performed using Matlab R2020b.

MetaboAnalyst 4.0, Matlab R2020b and GraphPad Prism 9.4.1 were used for data pre-processing, statistical analysis, and visualisation.

Role of funders

This research was exclusively funded by the National Health and Medical Research Council. The funding body did not play any role in study design, data collection, data analyses, interpretation, or writing of report.

Results

Clinical characteristics

Plasma samples from the biomarker and ultrasound measures for preventable stillbirth cohort were collected at 36 weeks’ gestation from pregnant patients with no signs of pregnancy complications. Patient characteristics are shown in Table 1. There was a significant difference in parity (p = 0.02), birthweight (p < 0.0001), and crown-heel length (p < 0.0001) between groups (Kruskal–Wallis test). No significant differences in gestation at collection, maternal age, BMI, fetal sex, smoking status, or gestational diabetes status were identified between groups.

Lipidomics to predict fetal growth restriction or preeclampsia

This study successfully identified 421 lipids for analysis after pre-processing. Initially, heatmaps were used to identify lipid changes between fetal growth restriction, fetal growth restriction/preeclampsia, and preeclampsia groups, compared to control. Fig. 1a indicates there was not clear groupings between controls and different disease groups within the samples. Likewise, Fig. 1b illustrates an absence of clustering within groups after PCA, indicating considerable overlap between patient samples collected at 36 weeks’ gestation. Fig. 1c presents the top 25 lipid metabolite averages within each group. Overall, patients that develop fetal growth restriction had a higher abundance of circulating cholesterol esters (CE), compared to control samples. Furthermore, patients that later developed preeclampsia had a higher abundance of Diacylglycerol (DG), Triacylglycerol (TG), compared to control samples.

Fig. 1.

Heatmap, dendrogram and principal component analysis (PCA) plots of targeted lipidomic analysis of plasma collected at 36 weeks' gestation from 55 who were fetal growth restricted (Including 6 with both FGR/PE; FGR = coral); 23 who developed preeclampsia (Including 6 with both FGR/PE; PE = blue) compared to 72 who did not (control; green). Samples are denoted in the top x-axis of both (a) and (c), and in the circles of (b). Heatmap scales presented red for higher, and blue for lower abundance of lipids. Scale is used for each of the top 25 lipids listed on the y-axis. Ward clustering algorithm is used for dendrograms. (a) Heatmap of each individual sample and top 25 lipids in targeted analysis; (b) PCA of samples illustrating PC1 (dimension 1) and PC2 (dimension 2); (c) heatmap of group averages.PCA, principal component analysis, PC1, principal component 1, PC2, principal component 2.

We assessed circulating lipids in patients before they were diagnosed with preeclampsia or fetal growth restriction. The following lipid classes were dysregulated (q < 0.1) in the circulation of 55 patients who birthed a fetal growth restriction infant: CE and ceramide (Cer). Details of dysregulated lipids were presented in Table 2.

Table 2.

Significantly dysregulated lipids in 55 patients who birthed a fetal growth restricted infant compared to 72 who did not.

Lipid	Odds ratio (95% CI)	p value	q value
CE 15:0	2.03 (1.35–3.19)	1.1E-3	0.08
CE 16:1	2.05 (1.36–3.23)	1.1E-3	0.08
CE 17:1	2.08 (1.37–3.33)	1.1E-3	0.08
CE 22:4	1.95 (1.33–2.97)	9.7E-4	0.08
CE 24:6	2.23 (1.46–3.59)	4.4E-4	0.08
Cer(d20:1/24:1)	0.47 (0.30–0.72)	8.0E-4	0.08

Odds ratios are given for a one standard deviation increase in the log-transformed lipid measurement. Lipid measurements were corrected for maternal age, BMI, smoking status, and gestational diabetes status. CE, Cholesterol ester; Cer, Ceramide.

Predictive performance of the 6 significantly dysregulated lipids for fetal growth restriction were assessed using the area under the receiver operating characteristic curve (AUC). Lipid metabolites CE 15:0 (Fig. 2a, AUC = 0.69); CE 16:1 (Fig. 2b, AUC = 0.69); CE 17:1 (Fig. 2c, AUC = 0.71); CE 22:4 (Fig. 2d, AUC = 0.70); CE 24:6 (Fig. 2e, AUC = 0.70); Cer(d20:1/24:1) (Fig. 2f, AUC = 0.70) demonstrated reasonable performance to predict fetal growth restriction pregnancies. AUC values remained unchanged when patients with both preeclampsia and fetal growth restriction were excluded (Fig. S1). Further analysis revealed no significant associations between lipids (CE24:1, Cer(d20:1/24:1), CE17:1, CE16:1, CE22:4, PI32:1, and sFlt/PlGF and gestation (days) (Fig. S2).

Fig. 2.

Circulating lipids were measured at 36 weeks' gestation in 55 maternal plasma samples prior to delivery of a fetal growth restricted infant, compared to 72 control samples. The discriminatory power of each lipid after correcting for maternal age, BMI, smoking status, and gestational diabetes status is shown as a ROC curve with the AUC annotated. (a) CE 15:0 (0.59–0.78 95% CI), (b) CE 16:1 (0.60–0.78 95% CI), (c) CE 17:1 (0.61–0.80 95% CI), (d) CE 22:1 (0.61–0.79 95% CI), (e) CE 24:6 (0.61–0.79 95% CI), (f) Cer(d20:1/24:1) (0.60–0.79 95% CI). AUC, area under the curve; ROC, receiver operator characteristic; CE, cholesterol ester; Cer, ceramide and CI, confidence interval.

The following lipid classes were dysregulated (q < 0.1, Benjamini-Hochberg correction) in circulation of 23 participants who were later diagnosed with preeclampsia, compared to 72 who were not: Diacylglycerol (DG), Phosphatidylcholine (PC), Phosphatidylinositol (PI), Triacylglycerol (TG), Phosphatidylethanolamine (PE) and Phosphatidylserine (PS). Details of dysregulated lipids were presented in Table 3. AUC values remained unchanged when patients with both preeclampsia and fetal growth restriction were excluded (Fig. S3). Further analysis revealed no significant associations between lipids (PI32:1, PC32:1, Hex1Cer (d18:1/16:0), TG 16:1 16:1 18:1, DG 34:2-(16:1), TG 14:1 16:1 18:0, TG 16:0 16:1 18:1, PI 34:1, DG 32:1-(16:1) and sFlt/PlGF) and gestation (days) (Fig. S4).

Table 3.

Significantly dysregulated lipids in preeclampsia pregnancies against controls.

Lipid	Odds Ratio (95% CI)	p value	q value
DG 34:2 –(16:1)	2.88 (1.63–5.61)	7.2E-4	0.05
PC 32:1	3.34 (1.82–6.88)	3.2E-4	0.05
PI 32:1	3.76 (1.97–8.10)	2.0E-4	0.05
PI 34:1	3.35 (1.75–7.21)	6.9E-4	0.05
TG 14:1 16:1 18:0	3.15 (1.71–6.63)	7.7E-4	0.05
TG 16:1 16:1 18:1	3.46 (1.80–7.77)	7.7E-4	0.05
DG 32:1 –(16:1)	3.05 (1.65–6.33)	1.0E-3	0.06
PC 36:3 (a∖b∖c)	3.05 (1.57–6.53)	1.9E-3	0.08
PE 32:1	2.46 (1.44–4.55)	1.9E-3	0.08
PS 36:1	0.34 (0.16–0.64)	2.0E-3	0.08

Odds ratios are given for a one standard deviation increase in the log-transformed lipid measurement. Lipid measurements were corrected for maternal age, BMI, smoking status, and gestational diabetes status. BMI, Body mass index; DG, Diacylglycerol; PC, Phosphatidylcholine; PI, Phosphatidylinositol; TG, Triacylglycerol; PC, Phosphatidylcholine; PE, Phosphatidylethanolamine; PS, Phosphatidylserine.

Predictive performance of the 10 significantly dysregulated lipids for PE were assessed using the area under the receiver operating characteristic curve (AUC). Lipid metabolites, CE 15:0 (Fig. 3a, AUC = 0.80); PC 32:1 (Fig. 3b, AUC = 0.81); PI 32:1 (Fig. 3c, AUC = 0.81); PI 34:1 (Fig. 3d, AUC = 0.79); TG 14:1 16:1 18:0 (Fig. 3e, AUC = 0.79); TG 16:1 16:1 18:1 (Fig. 3f, AUC = 0.80); DG 32:1 –(16:1) (Fig. 3g, AUC = 0.79); PC 36:3 (a/b/c) (Fig. 3h, AUC = 0.78); PE 32:1 (Fig. 3i, AUC = 0.79); and PS 36:1 (Fig. 3j, AUC = 0.79) were among the best lipids to predict those who developed preeclampsia later in pregnancy.

Fig. 3.

Circulating lipids were measured at 36 weeks' gestation in 23 participants preceding diagnosis of preeclampsia, compared to 72 who did not. The discriminatory power of each lipid metabolite was presented after correcting for maternal age, BMI, smoking status, and gestational diabetes status. Graphs are shown as ROC curves with the AUC annotated. (a) DG 34:2-(16:1) (0.70–0.89 95% CI), (b) PC 32:1 (0.72–0.90 95% CI), (c) PI 32:1 (0.72–0.91 95% CI), (d) PI 34:1 (0.68–0.89 95% CI), (e) TG 14:1 16:1 18:0 (0.70–0.89 95% CI), (f) TG 16:1 16:1 18:1 (0.70–0.90 95% CI), (g) DG 32:1 –(16:1) (0.70–0.88 95% CI), (h) PC 36:3 (a/b/c) (0.68–0.87 95% CI), (i) PE 32:1 (0.69–0.88 95% CI), (j) PS 36:1 (0.69–0.88 95% CI). BMI, Body mass index; ROC, receiver operator characteristic; AUC, area under the curve; CE, cholesterol ester; PC, Phosphatidylcholine; PI, Phosphatidylinositol; TG, Triacylglycerol, DG, Diacylglycerol; PE, Phosphatidylethanolamine; PS, Phosphatidylserine; and CI, confidence interval.

Lipid combinations improve prediction of pregnancy pathology

Next, we combined the best performing single upregulated and downregulated lipid for each pregnancy pathology to determine if combining lipids will improve predictive performance. In patients who later developed fetal growth restriction, combining CE 17:1 (AUC = 0.71) and Cer(d20:1/24:1) (AUC = 0.70) resulted in an improved AUC when combined (Table 4, AUC = 0.75); based on the Akaike Information Criterion (AIC).

Table 4.

Best performing upregulated and downregulated lipids for predicting 55 fetal growth restriction, and 23 preeclamptic pregnancies, as lone and combination biomarkers.

Pregnancy pathology	Lipid(s)	AUC (95% CI)	AIC
Fetal growth restriction	CE 17:1	0.71 (0.62–0.80)	171.2
	Cer(d20:1/24:1)	0.70 (0.60–0.79)	170.5
	CE 17:1 + Cer(d20:1/24:1)	0.75 (0.66–0.83)	163.5
Preeclampsia	PI 32:1	0.81 (0.72–0.91)	94.5
	PS 36:1	0.79 (0.69–0.88)	101.1
	PI 32:1 + PS 36:1	0.88 (0.81–0.95)	86.0

Lipids were log-transformed prior to analysis and included as a single biomarker, or in combination. Logistic regression was performed with each lipid as an independent variable corrected for maternal age, BMI, smoking status, and gestational diabetes status. BMI, Body mass index; AUC, Area under the curve; AIC, Akaike information criterion; CE, cholesterol ester; Cer, Ceramide; PI, Phosphatidylinositol; PS, Phosphatidylserine.

Using the odds ratio values, PI 32:1 (AUC = 0.81) and PS 36:1 (AUC = 0.79) were the best lipids to discriminate between 23 patients who later developed preeclampsia, and 72 who did not. In combination, PI 32:1 and PS 36:1 were able to improve predictive performance (Table 4, AUC = 0.88); based on the AIC.

Comparison of lipidomics and sFlt/PlGF to assess biomarker potential

Finally, we compared the top lipid combinations with sFlt-1 and PlGF; two existing biomarkers for predicting pregnancies at risk of fetal growth restriction and preeclampsia. To minimise overfitting, we performed five times repeated five-fold cross validation of the following logistic regression models for fetal growth restriction and preeclampsia individually: 1) best performing lipids alone, 2) a ratio of sFlt-1 and PlGF, and 3) best performing lipids combined with sFlt-1 and PlGF. For each model, the AUC and sensitivity at 90% and 80% were averaged across folds and repeats. Patient characteristics, maternal age, BMI, smoking status, and gestational diabetes were not included in the cross-validated models as these variables are not used in clinical tests utilising the sFlt-1 and PlGF biomarkers.

Using the repeated cross validation of the model, sFlt-1/PlGF ratio (AUC = 0.77) outperformed the model with the best performing lipids CE 17:1 + Cer(d20:1/24:1) (AUC = 0.73) to predict fetal growth restriction (Table 5). However, in combination, CE 17:1 + Cer(d20:1/24:1) + sFlt-1/PlGF ratio outperformed all other models (AUC = 0.82). Circulating sFlt-1/PlGF ratio (AUC = 0.87) outperformed the best performing lipids PI 32:1 + PS 36:1 (AUC = 0.79) to predict preeclampsia. However, the model with the best performing lipids PI 32:1 + PS 36:1 combined with sFlt-1 and PlGF outperformed all other models (AUC = 0.89).

Table 5.

Averaged performance of logistic regression models after five times repeated five-fold cross validation utilizing the best performing lipids combinations alongside sFlt-1 and PlGF in predicting fetal growth restriction and preeclamptic pregnancies.

Pregnancy pathology	Model	Cross-validated sensitivity @ 90% specificity	Cross-validated @ensitivity @ 80% specificity	Cross-validated AUC
Fetal growth restriction	CE 17:1 + Cer(d20: 1/24:1)	42.8%	53.3%	0.73
	sFlt-1 + PlGF	47.4%	53.6%	0.77
	CE 17:1 + Cer (d20:1/24:1) + sFlt-1 + PlGF	56.6%	62.0%	0.82
Preeclampsia	PI 32:1 + PS 36:1	42.8%	62.4%	0.79
	sFlt-1 + PlGF	71.4%	75.2%	0.87
	PI 32:1 + PS 36:1 + sFlt-1 + PlGF	74.2%	82.8%	0.89

AUC, Area under the curve; CE, Cholesterol.ester; Cer, Ceramide; PI, Phosphatidylinositol; PS, Phosphatidylserine. sFlt-1 and PlGF measurements could not be obtained for one FGR sample which was excluded from this analysis.

Discussion

This study performed targeted lipidomics in maternal plasma collected at 36 weeks’ gestation to identify biomarkers to predict term preeclampsia, or fetal growth restriction. We successfully identified 421 lipids in patients destined to develop preeclampsia, and 55 in those who delivered a <5th centile baby (before correcting for maternal age, BMI, smoking and gestational diabetes). In this study, we used a case-cohort study design which is an advantage over nested case–control study as it allows for the randomly selected controls to be used in both disease outcomes.²⁰

Cholesterol esters were commonly altered in the circulation of patients who delivered a <5th centile baby at term. As a single marker, CE 17:1 provided the best predictive power for fetal growth restriction pregnancies. Cholesterol esters are long-chain fatty acid cholesterol molecules with a linked hydroxyl group and function as a transport form of cholesterol in blood plasma, and lipid droplets in cells.²¹ They are a protective form of cholesterol for storage in cells and transport in plasma. Cholesterol esters are required for fetal development, and can be transported to the fetus via the trophoblast and endothelial cells of the fetoplacental vasculature.²² Whilst cholesterol esters are important for transport around the circulation, cholesterol can accumulate within the cells. In excess, cholesterol metabolism produces a high volume of reactive oxygen species often associated with lipid peroxidation and inflammation in the placenta. Inflammation is an important contributor to pregnancy pathologies, like fetal growth restriction.²³ Future studies should measure circulating free cholesterol, and enzymes involved in cholesterol metabolism to understand the molecular mechanisms involved.

Phosphatidylinositol, phosphatidylserine, diacylglycerol, and triacylglycerol were among the top dysregulated lipids in circulation of participants who developed preeclampsia. As a single molecule, we found phosphatidylinositol, specifically PI 32:1 as the best molecule to predict patients who developed preeclampsia. Phosphatidylinositols consist of a phosphatidic acid backbone linked via the phosphate group to inositol (hexahydroxycyclohexane). They are important for metabolism cell signalling and form the cell membrane structure.²⁴ Transport of nutrients, enzymes and receptors are greatly influenced by the state of the lipid bilayer within the cells. Although in excess, these lipids like cholesterol and phospholipids are associated with enhanced formation of lipid peroxides. These are compounds which give rise to endothelial dysfunction, often observed in preeclampsia.²⁵ Further, studies by Huang et al. found high levels of phospholipids in placental tissue complicated by preeclampsia which compliments the elevated phospholipids we observed in our cohort.²⁶ Further studies would benefit from investigating the relationship between lipid transport and preeclampsia.

To develop a potential predictive test for fetal growth restriction, and preeclampsia, we combined the top two lipids along with patient characteristics including maternal age, BMI, smoking status, and gestational diabetes status. Different biomarker models were compared using Akaike's information criteria to evaluate and improve model fitness.²⁷ As the aim of this study was to discover new biomarkers to predict fetal growth restriction and preeclampsia, we also measured sFlt-1 and PlGF to compare the performance of lipid metabolites. Interestingly, while sFlt-1/PlGF ratio outperformed single lipids, in combination, lipids + sFlt/PlGF ratio improved AUC and sensitivity to predict fetal growth restriction and preeclampsia. This suggest that adding lipids to an algorithm may improve early prediction of fetal growth restriction and preeclampsia.

There have been recent studies using an omics approach to generate new and accurate biomarkers. Bahado-Singh et al. used artificial intelligence and multiomics approaches in amniotic fluid to better predict perinatal outcomes.²⁸ Additionally, Sovio et al. have measured metabolites and developed a metabolite ratio to predict fetal growth restriction at term.²⁹ Others have begun to explore metabolic biomarkers in the field of severe preeclampsia development.30, 31, 32 Future research could focus on combining multiomics metabolites found in circulation to improve detection of pregnancy pathologies late in pregnancy.

Like most studies, there are limitations that should be discussed. At the beginning of this study, we used mass spectrometry to measure 468 lipids in a case-cohort of 144 samples. The women were not fasted prior to collection of these samples, and dietary consumption was not recorded. Thus, this study is reflective of a ‘real world’ scenario where pregnant women attend the clinic around 36 weeks' gestation without fasting. Future studies could assess more stringent sampling considerations. We also note that this study was a case-cohort design and thus the data should be representative of the data had we run the entire cohort population of 1000. We acknowledge that omics research often results in over-optimistic performance estimates. To overcome this challenge in our study, we performed a five times repeated five-fold cross validation which is a preferred cross-validation technique for machine learning models.³³ The results of this study should still be further validated in other cohorts.

This study utilised plasma for discovery of lipid biomarkers. Ishikawa et al. and Sovio et al. have compared lipids in plasma and serum and have concluded plasma is the best for lipid metabolite measurements.^29,34 This is partly due to the clotting process which occurs before serum is collected. As clotting occurs, by-products like DG and inositol 1,4,5-phosphate are released via the degradation of phosphatidylinositol 4,5-bisphosphate and the activation of phospholipase C.²⁷ This cascade of events elicits a biological response in serum that may not be reflective of normal biological effects.³⁵

Additionally, discovery of lipids as potential biomarkers are often simple to study through mass spectrometry.³⁶ However, due to the complexity of lipid science, one challenge is measuring lipid subspecies without the use of mass spectrometry for further clinical use.³⁷ Currently there are no tests as sensitive as a mass spectrometer and therefore lipidomic measurement for clinical performance requires more efforts to solve these potential challenges before proceeding.³⁸

This study used lipidomics to predict pregnancies that are at risk of term preeclampsia, or fetal growth restriction. We successfully quantified 421 lipids in plasma from 144 samples collected at 36 weeks’ gestation. Although lipids alone did not outperform the current protein biomarkers, sFlt-1 and PlGF ratio, we found that by adding lipids to protein measurements, the AUC and sensitivity of each test was slightly improved. The predictive capacity of generating lipid profiles for gestational disorders holds significant potential as a non-invasive assessment of maternal and fetal health. More studies are required to further investigate lipid metabolism in pregnancy and validate these changes in external cohorts.

Contributors

LAB: Conceptulization, Formal analysis, Writing—Original draft, Writing—Review & Editing, Visualization, EK: Conceptulization, Formal analysis, Writing—Original draft, Writing—Review & Editing, SPW: Resources, Project administration, Funding acquisition, Writing—Review & Editing, TMM: Resources, Project administration, Funding acquisition, Writing—Review & Editing, BN: Formal analysis, ST: Conceptualization, resources, data curation, Writing—Review & Editing, Supervision, Project administration, Funding acquisition, TKL: Conceptualization, resources, data curation, Writing—Review & Editing, Supervision, Project administration, Funding acquisition. LAB and EK had direct access and verified data from the lipidomics measurements. All authors read and approved the final version of the manuscript.

Data sharing statement

The data that support the findings of this study are available from the corresponding author, LAB, upon reasonable request.

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.