Introduction

Preeclampsia remains a leading cause of maternal death throughout the world and is responsible for considerable neonatal morbidity and mortality [1]. Management of preeclampsia risk throughout pregnancy therefore remains a key aspect of prenatal care worldwide. The accurate prediction of preeclampsia risk early in pregnancy has been flagged as a research priority in many clinical guidelines [2, 3] as it would enable risk-based prenatal care pathways, and hence more effective utilisation of prenatal care resources. The possible patient benefits of accurate preeclampsia risk prediction have recently been reaffirmed in the ASPRE trial, which reported that aspirin prophylaxis in women identified to be at risk of preeclampsia before 37 weeks of gestation, so-called preterm preeclampsia, reduced the incidence rate of preterm preeclampsia by 62% [4].

The National Institute for Health and Clinical Excellence (NICE) in England and the American College of Obstetricians and Gynecologists (ACOG) in the United States recommend the use of maternal demographic, obstetrical history and overt medical risk factors in simple dichotomous classifications to determine low or high risk status for the development of preeclampsia [5, 6]. However, the prediction accuracy of such approaches is limited. Using the NICE guidelines, sensitivity of 39% for preterm preeclampsia and 34% for term preeclampsia at 10.3% false positive rate (FPR) were reported; the corresponding sensitivity using the ACOG recommendations were 90% and 89%, but at 64.3% FPR [5, 7, 8]. This led clinical researchers to develop multivariable models which utilize these clinical risk factors to compute risk scores, as recently reviewed by Brunelli et al [9]. Several of these were subjected to external validation in two recent Dutch studies [10, 11], showing such models do have some moderate risk prediction ability, with some likely to outperform clinical guidelines [10]. However, these studies also re-iterated that the predictive performance decreased significantly in first-time pregnant women. Myers et al. further corroborated this finding by showing that in a population of low risk nulliparous women, detection rates for preterm preeclampsia were only 16.1%, 26.5% and 22.2% for respectively the NICE classifier [5], the maternal risk based competing risk model underpinning the classifier used in the Combined Multimarker Screening and Randomized Patient Treatment with Aspirin for Evidence-Based Preeclampsia Prevention (ASPRE) trial [12], and an alternative risk scoring method as proposed by Sovio et al [13, 14].

The need for novel biomarkers to improve preeclampsia prediction in nulliparous women without apparent risk factor is further emphasized by the fact that nulliparity is associated with increased risk (RR = 2.1 (1.9 to 2.4)) and accounts for the largest single population attributable fraction (32.3%) for preeclampsia [15].

To date, the most studied biomarker in preeclampsia is Placental Growth Factor (PlGF), the levels of which tend to drop in a sub-group of women destined to develop preeclampsia [16], and which has some demonstrated risk prediction potential [17]. Its use, in combination with maternal characteristics, medical history, obstetric history, blood pressure and, where available, uterine artery doppler measurements, is gaining acceptance for prediction of preterm preeclampsia risk in the general pregnancy population [18]. Yet additional biomarkers are still required to complement PlGF to improve preeclampsia risk prediction in nulliparous women without apparent risk factors.

The aim of this work was to inform the development of a novel biomarker-assisted test for the early and accurate prediction of preeclampsia risk in low-risk nulliparous women. With preeclampsia recognised as being a syndromic condition, accurate prediction of preeclampsia will require for several biomarkers to account for different preeclampsia subtypes and/or patient risk profiles [19–21]. Because liquid chromatography-triple quadrupole mass spectrometry (LC-QqQ-MS) is a well-established analytic technique used in clinical laboratories worldwide for multiplex metabolite analysis, such as neonatal screening for inborn errors of metabolism [22], we set metabolite “assay-ability” by LC-QqQ-MS as a stage-gate biomarker selection criterion. This will ensure that any biomarker panel identified will be easily portable to in vitro diagnostic workflows already operated in clinical laboratories.

We previously reported a metabolomic biomarker discovery study for early pregnancy prediction of later preeclampsia [23]. Building on these results (and those of others; S1 File & S1 Table), we set out to consolidate a panel of metabolite biomarkers which could be used in conjunction with PlGF to deliver accurate preeclampsia prediction in nulliparous women.

Firstly we developed a translational research workflow based on LC-QqQ-MS based metabolite quantification; only candidate biomarkers amenable to this workflow were considered for further analysis in a case-control study. Biomarkers with univariable prediction merits were pre-selected for the development of risk predictor models. Models were evaluated in function of two pre-defined clinical use scenarios: (1) identify women at risk of developing preterm preeclampsia (primary objective), and (2) identify women at risk of developing preeclampsia at any stage of the pregnancy (secondary objective). We found that combining dilinoleoyl-glycerol (DLG) with PlGF effectively predicted increased preterm preeclampsia risk at ca. 15 weeks of gestation. The further addition of heptadecanoyl-2-hydroxy-sn-glycero-3-phosphocholine (1-HGP) expanded the capacity to also identify pregnant women at decreased risk of developing any form of preeclampsia. PlGF, DLG and 1-HGP were therefore proposed for progression to the final development stage of a risk prediction test for (preterm) preeclampsia and subsequent validation in the IMproved Pregnancy Outcomes by Early Detection (IMPROvED) study [24].

Materials and methods

Results

Study population

A nested case-control study (n = 432) was performed within the 2364 women recruited into the European SCOPE cohort, for whom pregnancy outcomes were known. Twenty-three women developed preterm preeclampsia (1%) and seventy-four (4%) developed term preeclampsia. Background characteristics and outcomes of pregnancy in women who did and did not develop preeclampsia are shown in Table 1. Mean age, and gestational age at sampling did not differ significantly among the groups. Significant differences are observed for BMI and blood pressure at sampling. Gestational age at delivery was lower in preeclampsia cases compared with controls (no preeclampsia). Likewise, the babies born to preeclamptic women were significantly lighter than the babies in the control group.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of the study population.

https://doi.org/10.1371/journal.pone.0244369.t001

Predictor selection

Biomarkers and biometric variables selected based on their stand-alone prediction performance of preterm, term or any preeclampsia are shown in Table 2. Mean Arterial Pressure (MAP) showed significant discriminative performance for term and all preeclampsia (AUROC = 0.67 and 0.69 respectively). BMI delivered similar discrimination for preterm, term and all preeclampsia (AUROC = 0.63 and 0.65 respectively). DLG and PlGF stand-alone prediction performances were higher for preterm preeclampsia (AUROC = 0.70 and 0.73 respectively). 1-HGP showed similar discriminative performance for both preterm and term preeclampsia (AUC = 0.61) but this performance was only significantly superior to 0.50 for term preeclampsia (p<0.05). With high FDR associated with the predictor selections for preterm preeclampsia and term preeclampsia, only the predictors as selected for preeclampsia were considered as variables for modelling.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Selection of predictors for multivariable modelling.

https://doi.org/10.1371/journal.pone.0244369.t002

Modelling

Model development was targeted to answer two pre-set outcome based clinical use scenarios (preterm preeclampsia and any preeclampsia), and three associated minimal performance requirements, expressed as Positive and Negative Predictive Value thresholds (PPV and NPV, as detailed in Methods–Clinical use cases). The multivariable models developed within this framework are shown in Table 3; the comparator models based on the classic predictors MAP, BMI and PlGF are available in S7 Table. For the primary objective, i.e., the detection of women at increased risk for preterm preeclampsia at a pre-set PPV cut-off of ≥ 0.05, the observed sensitivity using only PlGF was 48% (PPV = 0.05), and up to 61% for the comparator models combining PlGF with MAP and/or BMI. Combining PlGF with DLG by means of recursive partitioning increased the sensitivity significantly to 74% (PPV = 0.06). The resulting partitioning rule-in model (PlGF || DLG) is shown in Fig 1. Predictor cut-off values in each partition step were 0.00526 ng/mL and 2.26 relative concentration (RC) for PlGF and DLG respectively. By combining PlGF with a two-predictor partial least squares discriminant analysis (PLS-DA) model constituting DLG and 1-HGP (PlGF || (DLG+1-HGP)), the sensitivity could be increased further to 78% (PPV = 0.05; S1 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Partitioning rule-in model (PlGF || DLG).

(A) Partition 1: PlGF concentration (ng/mL) versus clinical outcome PlGF cut-off value = 0.00526 (dotted line). (B) Partition 2: DLG relative concentration versus clinical outcome for patients ruled-out in partition 1. DLG cut-off value = 2.26 (dotted line). (C) Partitions 1 and 2: DLG relative concentration versus PlGF concentration (red stars: preterm preeclampsia, orange triangles: term preeclampsia, blue circles: no preeclampsia). Dotted line: DLG and PlGF cut-off values. (D) Estimated predictive performance of the model for 10,000 patients.

https://doi.org/10.1371/journal.pone.0244369.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Prediction models for the pre-set clinical use scenarios.

https://doi.org/10.1371/journal.pone.0244369.t003

With regards to predicting increased risk for any type of preeclampsia at a pre-set PPV cut-off of ≥ 0.15 (secondary objective), comprehensive analysis of the generated prediction models revealed that no classifier compliant with the PPV≥0.15 requirement could be identified.

In models developed to meet the “rule-out” performance criteria (Table 3), specificity at the target NPV≥0.99 for preeclampsia improved from 0.02 (PlGF only) to 0.37 when combining PlGF, DLG and 1-HGP predictors using recursive partitioning. The resulting partitioning rule-out model (PlGF || DLG || 1-HGP) and the estimated predictive performance are shown in Fig 2. Predictor cut-off values in each partition step were 0.00702 ng/mL, 1.12 RC and 0.458 RC for PlGF, DLG and 1-HGP respectively. This model compares favourably to the comparator models combining PlGF with MAP and/or BMI; the best comparator model, i.e., the recursive partitioning model PlGF || BMI || MAP delivers a rule-out specificity of 0.32.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Partitioning rule-out model (PlGF || DLG || 1-HGP).

(A) Partition 1: PlGF (ng/mL) versus clinical outcome. PlGF cut-off value = 0.00702 (dotted line). (B) Partition 2: DLG relative concentration vs clinical outcome for patients ruled-out in partition 1. DLG cut-off value = 1.12 (dotted line). (C) Partition 3: 1-HGP relative concentration versus clinical outcome for patients ruled-out in partitions 1 and 2. 1-HGP cut-off value = 0.458 (dotted line). (D) Estimated predictive performance of the model for 10,000 patients.

https://doi.org/10.1371/journal.pone.0244369.g002

Discussion

In this study we re-evaluated the relevance of a set of putative metabolite biomarkers to further inform the development of a test to predict preeclampsia in nulliparous women without apparent risk factors. Upon application of both technical assay-ability and data quality assurance criteria, forty-three metabolites, PlGF and the biometric variables MAP and BMI, well-known to associate with preeclampsia, were retained as predictors for consideration.

In line with the findings that placental pathology is more common in preterm preeclampsia [35], we adopted the hypothesis that within the group of women who developed preterm preeclampsia, there would be a greater prevalence of preeclampsia associated with placental insufficiency, whereas the group of women developing term preeclampsia would be more reflective of preeclampsia of maternal origin [36]. Hence, in addition to the outcome preeclampsia, we considered the outcomes preterm and term preeclampsia as two disease subtypes in the process of pre-selecting variables.

This led to an initial selection of twelve predictors, comprising the clinical risk factors relevant in the low-risk population studied (e.g., MAP and BMI), PlGF and nine metabolites from the original set of forty-three quantified metabolites. For all predictors but BMI, the single variables showed discriminative performance which aligned with the prediction of either preterm preeclampsia or term preeclampsia. This observation supports the assumption that in this study population the two disease subtypes considered do indeed align well with distinct underlying physiological pathways. Interestingly, we note that in this study population, MAP is only a stand-alone predictor for term preeclampsia and not for preterm preeclampsia, whereas typically MAP is reported to be a strong stand-alone predictor for preterm preeclampsia [37]. With high FDR associated with the predictors selected for preterm and term preeclampsia, we restricted the final predictor selection for modelling to these predictors which showed performances when assessed for all preeclampsia prediction, i.e., the biometric predictors MAP and BMI and the biomarkers PlGF, DLG and 1-HGP.

Consultation with a panel of clinicians (IMPROvED Consortium meeting, London 2018; see also Acknowledgements) established two relevant clinical use scenarios for stratifying nulliparous pregnant women without overt risk factors to their preeclampsia risk. The primary clinical use scenario was a rule-in classification to identify women at high risk of developing preterm preeclampsia. The secondary clinical use scenario was delivering the same preeclampsia risk information to nulliparous pregnant women without overt risk factors, as the risk information available to parous women based on their obstetrics history [38].

In risk stratification, the clinical usefulness of a test is best expressed in terms of positive and negative predictive values [38]. Hence, to establish robust performance targets for the prognostic test(s) for the identified clinical use scenarios, we derived these risk probabilities from published large clinical studies reporting risk probabilities for well-powered population studies. The ASPRE study, involving a cohort of 25,797 pregnancies with mixed prior risk profiles and parity who were screened for preterm preeclampsia, provided a rule-in classification threshold of PPV ≥ 0.05 [39]. The population-based study of Hernandez-Diaz included data from 504,849 parous pregnancies. These authors reported a preeclampsia risk of 14.7% in the second pregnancy for women with a history of preeclampsia, whereas the risk of preeclampsia for parous women without a history of preeclampsia was 1.1% [40].

Considering these “target” inputs led to the establishment of a set of hierarchical goals for the development of predictive models for preeclampsia in nulliparous women without apparent risk factors. The primary goal was defined as the identification of pregnant women at increased risk for preterm preeclampsia with a post-test risk ≥ 0.05 (or PPV ≥ 0.05). Secondary to this, the establishment of generic risk for developing preeclampsia at any stage of pregnancy. For the latter, both the identification of women at high risk, defined as a post-test risk ≥ 15% (PPV ≥ 0.15), and the identification of women at low risk, defined as a post-test risk ≤ 1% (NPV ≥ 0.99), were considered of interest to clinical practice.

Among all preeclampsia biomarkers studied, PlGF is the best understood. Low PlGF levels throughout pregnancy are associated with a disease phenotype which has placental insufficiency as a key characteristic [16, 41–44]. Indeed, it has been shown that low PlGF levels early in pregnancy has some preeclampsia risk prediction merits, especially for preterm preeclampsia wherein placental compromise is more common [16].

Given the multi-factorial nature of preeclampsia, a Bayesian modelling approach may deliver more accurate predictive models than PLS-DA [12, 45]. However, the available number of patients prevented calibration and therefore the use of probabilistic methods. For this reason, recursive partitioning was considered as an alternative as it does not necessitate calibration. In the absence of calibration, generalisability of the models developed cannot be gauged in this study; this will be assessed in a readily available independent cohort [24].

We indeed found that recursive partitioning delivered better prediction than PLS-DA models. By establishing PlGF as the first independent predictor, the recursive partitioning modelling allowed us to align the model development with the pre-defined classification objectives. We found that complementing PlGF with a single metabolite biomarker (DLG) increased the sensitivity of the test from 48% (at PPV = 0.05) to 74% (at PPV = 0.06) for predicting preterm preeclampsia risk in nulliparous women without overt risk factors. Preliminary results show combining 1-HGP with DLG in the second partition may further improve the prediction of preterm preeclampsia, increasing the sensitivity to 78%.

Developing recursive partitioning models with three splits suggested that 1-HGP may also be a third independent predictor. Adding 1-HGP enabled identification of significant fraction of women at low risk of developing preeclampsia (specificity = 36%) in accordance with the pre-set rule-out target of NPV = 0.99.

In this data set we were unable to establish a prediction model which also complied with the other secondary goal: the identification of a sizeable group of women with a post-test risk ≥ 15% of developing any type of preeclampsia in their pregnancy. This may be explained in the context of the competing risk model concept as established for preeclampsia by Wright et al. In this model, pregnant women at risk of developing preeclampsia may experience an uncomplicated pregnancy, as their pregnancy (naturally) concluded prior to their preeclampsia risk coming to expression [46].

It is however noted that the list of putative biomarkers considered here was primarily informed by our earlier work [23] and from findings other groups around the same time (S1 Table). Additional putative metabolite biomarkers are still being identified in de-novo metabolomics biomarker studies within other pregnancy demographics [47–50]. Evaluation of these metabolites within the technical and clinical use framework established here, may lead to further prediction improvements and achieving the secondary goal in full.

A review of the literature reveals that DLG and 1-HGP may map onto different but complementary pathways relative to PlGF. Genetic studies have provided new insights into the mechanisms that link intracellular diacylglycerol concentrations to insulin resistance [51, 52], which associates with preeclampsia risk [53]. Erion & Shulman reported that increases in intracellular diacylglycerol content could lead to activation of new protein kinase C which in turn, inhibits insulin action in liver and skeletal muscle [54]. Diacylglycerols in plasma were reported as the basis of a biomarker profile of metabolic syndrome onset in rhesus monkeys; diacylglycerol-derived linoleate levels were markedly upregulated in insulin-resistant metabolic syndrome animals [55]. The physiological metabolic alterations occurring in normal pregnancy: transient excursion into insulin resistance, hyperlipidemia, and increased coagulation are essentially the hallmarks of metabolic syndrome, and are exacerbated in preeclampsia [56, 57]. With the metabolic syndrome manifesting in women with preeclampsia typically associated with insulin resistance and endothelial dysfunction [58], it is therefore plausible that DLG levels reflect risk pathway for the development of preeclampsia that is associated with a pregnancy-specific disposition for metabolic syndrome [56]. With reference to 1-HGP, previous studies have shown that lysophosphatidic acid (LPA), the hydrolysed form of 1-HGP, is a potent mediator of the immune response and could contribute to improper immune activation upregulating the production of inflammatory cytokines [59, 60]. Studies in humans and animal models have indicated that the development of hypertension in preeclampsia has a strong pro-inflammatory component and that placental ischemia is the stimulus for this immune activation [61–64]. In this context, 1-HGP may be related to vascular inflammation and dysfunction. Recently, Fujii et al. reported that LPA signalling may be directly involved in placental homeostasis. They found that dysregulation of LPA signalling as mediated by LPA specific G-protein-coupled receptors (LPAR), and more specifically LPAR3, may be associated with placental dysfunction in preeclampsia [65]. Circulating 1-HGP levels may therefore be indicate LPA signalling issues which in turn could be directly or indirectly be associated with the risk of developing preeclampsia.

Based on the data generated in this study, DLG and 1-HGP are selected as the key markers for the development of a dedicated, rapid, quantitative LC-MS/MS assay prototype.

Conclusions

In this study, we adopted an end-user focused translational research methodology for the prioritization of putative metabolite biomarkers for further development into a clinical test for preeclampsia risk prediction. Three key determinants were considered in the selection process: compatibility with LC-MS/MS technology already available in clinical laboratories, ability to improve the prediction of the established preeclampsia biomarker PlGF, and ability to meet patient stratification requirements as identified by clinical practitioners.

Using this translational research framework to re-evaluate fifty-eight metabolite biomarker candidates, we were able to prioritise two metabolite biomarkers from an initial pre-selection of nine metabolites, for development of a clinical assay. We found that the metabolites DLG and 1-HGP effectively complemented PlGF in a prediction model for (preterm) preeclampsia in nulliparous women without over risk factors. A review of the literature supports that PlGF, DLG and 1-HGP may represent separate pathways for preeclampsia risk in the population studied.

A strength of this study lies in the well-defined pregnancy risk profile within the SCOPE study, allowing us to focus this translation biomarker research to the pregnancy population most in need for biomarker-assisted risk prediction. In addition, a rational approach considering both technical usability and clinical utility criteria for refining the selection of previously identified putative biomarkers of interest for further test development, was adopted to increase the chance of successfully porting novel biomarkers into clinical practice.

A limitation of this work lies within the ethnic make-up of the SCOPE-Europe study; most of the women were of white ethnicity. In line with the hypothesis that different pregnancy populations may constitute different patient risk profiles, there may be a need for additional biomarkers, representing additional preeclampsia risk pathways, to improve prediction for other ethnicities. Another limitation of this study is the gestational age considered; risk prediction involved samples taken early in the second trimester of pregnancy (15 +/- 1 week of gestation), whereas there is trend to consolidate prenatal risk screening at the end of the first trimester. However, we note that none of the verified metabolite biomarkers show any dependency on gestational age of sampling (S2 Fig).

The next steps in our test development work will involve developing a multiplex clinical assay for the selected markers, as well as further assessment of their generalisability to complement PlGF for the prediction in preeclampsia in other pregnancy populations. In a first instance, the markers will be assessed in IMPROvED; a pan-European multi-centre phase 2A clinical study which recruited ca 4000 low risk primiparous women for the purposes of assessing and refining innovative biomarker-assisted prototype predictive tests [24].

Supporting information

Acknowledgments

The Authors Louise C. Kenny (Louise.Kenny@liverpool.ac.uk), Lucilla Poston, Jenny E. Myers, Nigel A.B. Simpson, Fergus P. McCarthy and Philip N. Baker contributed to this work on behalf of the Screening for Pregnancy Endpoints (SCOPE) Consortium.

The investigators thank the pregnant women who participated in the SCOPE study for their indispensable contribution. Clinicians from the IMPROvED (Improved Pregnancy Outcomes by Early Detection) study who attended the IMPROvED meeting in London (April 2018) are acknowledged for their contribution: Dr. F. McCarthy (Cork University Maternity Hospital, Ireland), Pr. Z. Alfirevic (University of Liverpool, UK), Dr. K. Navaratnam (University of Liverpool, UK), Dr. J.J. Duvekot (Erasmus MC, Netherlands), Pr. S. O’Brien (Keele University, UK) and Dr. P. Wu (Keele University, UK). Paloma Muñiz-Ortea PhD and Caroline Nolan (Metabolomic Diagnostics) are acknowledged for the assistance in the preparation of the study samples; Paloma Muñiz-Ortea for an independent review of LC-MS/MS data. Some writing assistance for the preparation of this manuscript was provided under the direction of the authors by Carmen Romera PhD (Metabolomic Diagnostics).