2.1. Study selection

Study selection was conducted using the free web-based software CADIMA (https://www.cadima.info/index.php). Search results from the three databases were imported into the platform, which automatically merged the records and removed duplicates. Two reviewers independently screened titles, abstracts, and full texts, applying the predefined inclusion criteria based on the PICO framework (detailed in S2 Appendix). The PICO criteria were defined as follows: Population (P): Pregnant women at risk of or diagnosed with PE; Intervention (I): Use of metabolomic profiling or analysis to identify biomarkers; Comparator (C): Healthy pregnant women; Outcome (O): Identification and/or validation of metabolomic biomarkers for the prediction or diagnosis of PE. Studies not meeting these criteria were excluded.

After completing the selection process, a standardized data extraction table was developed to capture key variables relevant to the research objectives. These included general study information (authors, year, country), characteristics of the study population (maternal age, gestational age, follow-up time), methodological details, type of biological matrix analyzed, and main findings. The final dataset, once verified and consolidated by two reviewers, was exported from CADIMA in Excel format for synthesis. All studies that employed metabolomic approaches to investigate PE were included, and for comparison purposes, studies were grouped according to the type of biological matrix analyzed.

To evaluate study quality and risk of bias, a critical appraisal was carried out within CADIMA (S1 Table), using predefined criteria adapted to the study designs and research context. These criteria focused on methodological rigor, sample size, and measurement validity. Three reviewers independently assessed each study, and any disagreements were resolved by consensus. The results of the quality assessment are graphically summarized in S1 Fig.

Due to substantial heterogeneity across studies—particularly in sample types, analytical platforms, metabolites assessed, and outcome definitions—a meta-analysis was not performed. Instead, a narrative synthesis was conducted, organizing findings by biological matrix to facilitate interpretation. Given the small number of included studies (n = 12) and the absence of a quantitative summary, subgroup and sensitivity analyses were not considered appropriate.

Study characteristics

The selected studies were published between 2010 and 2024 (Table 1). They were conducted in various countries: United States [22–24], United Kingdom [25–27], Norway [28–30], China [31], and Canada [32]. Additionally, one study was developed in multiple countries, including Australia, New Zealand, and Canada [33]. The study designs include cohort studies [24,28–30], case-cohort studies [27], case-control studies [22,23,32,33], prospective control studies [25], and the remaining are considered as translational studies [26,31].

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Comparison of the analytical techniques used in the selected studies.

https://doi.org/10.1371/journal.pone.0339292.t001

The most frequently used technique was mass spectrometry (MS) and also coupled with liquid chromatography (LC-MS) [24,26] and its derivatives, including liquid chromatography-tandem mass spectrometry (LC-MS/MS) [22,31], ultra-performance liquid chromatography-mass spectrometry (UPLC-MS) [33], ultra performance liquid chromatography – tandem mass spectrometry (UPLC-MS/MS) [27], and liquid chromatography-high-resolution mass spectrometry (LC-HRMS) [23]. Only one study used exclusively MS [32]. Nuclear magnetic resonance (NMR) spectroscopy was also widely employed [24,25,28] along with its derivative, high-resolution magic angle spinning magnetic resonance spectroscopy (HR-MAS MRS) [29,30]. The analysed samples in these studies were serum [22–25,28], plasma [26,27,32,33], urine [28], feces [31] and placental tissue [29,30].

Each metabolomic technique exhibits distinct analytical characteristics, which are summarized in Table 1 and have important implications for metabolite coverage, sensitivity, and confidence of identification.

Mass spectrometry (MS)-based approaches typically rely on a separation step, most commonly liquid chromatography (LC), and provide high sensitivity and broad metabolite coverage. However, they often require extensive sample preparation and can be affected by variable ionization efficiency, matrix effects, and limited reproducibility, which may compromise quantitative accuracy and metabolite identification [34,35]. Targeted LC–MS/MS strategies, including UPLC–MS/MS, improve analytical selectivity and quantitative robustness through multiple reaction monitoring, but require careful method optimization, specialized instrumentation, and often additional confirmatory analyses [36,37]. High-resolution mass spectrometry (LC–HRMS) offers high mass accuracy and broad untargeted coverage, as well as the possibility of retrospective data reprocessing, although it generates large datasets and may still suffer from ambiguity in isomeric metabolite assignment [38]. [39]. In contrast, NMR-based techniques provide inherently quantitative and highly reproducible data, together with direct structural information that facilitates confident metabolite identification. HR-MAS NMR allows the direct analysis of intact tissue, minimizing extraction-related variability, but is limited by lower sensitivity, reduced spectral resolution, and peak overlap [39]. Conventional solution-state NMR shares these advantages and limitations, offering robust quantification and structural insight at the expense of sensitivity compared with MS-based platforms [40].

This methodological heterogeneity across metabolomic technologies has important consequences for the interpretation and comparison of reported metabolite profiles. Differences between studies may reflect not only underlying biological variation but also platform-specific constraints related to metabolite coverage, sensitivity, and confidence of identification. MS-based approaches may detect a wider range of metabolites but are more susceptible to ambiguous annotation of isomeric compounds or analytical artefacts, whereas NMR-based techniques tend to identify fewer, more abundant metabolites with higher structural confidence. Consequently, metabolites consistently detected across multiple studies and analytical platforms are more likely to represent high-confidence biological signals, while technique-specific findings should be interpreted with caution, as they may partly arise from methodological or analytical limitations rather than true biological differences.

Participant characteristics

The participants in the selected studies range in age from 18 to 43 years (Table 2). Regarding preeclampsia classification, four studies combined EO-PE and LO-PE cases [22,26,30,33], three focused on LO-PE [25] (with a single EO-PE case included in two of them [23,28]), and five studies did not specify the PE subtype [24,27,29,32]. The majority of the studies were carried out on the first trimester of pregnancy [22–25,27–29], while the others took place in the second [33] and third [26] trimester or after delivery [30,32]. The remaining studies [31] did not specify the gestational age. This variability in case definition and gestational age represents a source of clinical heterogeneity across the included studies. Some of the patients were followed after delivery [24,29,30], or until delivery [22,33], while others were monitored only during the first trimester [23,25,28]. The remaining studies did not report follow-up duration [26,27,31,32].

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Studies and participants characteristics.

https://doi.org/10.1371/journal.pone.0339292.t002

3.1. PE metabolites in serum

The study of Odibo, A.O. et al. [22] conducted a nested case-control study that revealed four key metabolites: hydroxyhexanoylcarnitine, alanine, phenylalanine, and glutamate significantly elevated in maternal blood samples from women who later developed PE compared to pregnant controls (PC; women without HDP) during the first trimester of pregnancy (Table 3). They found that the predictive ability of these metabolites was robust, with an area under the curve (AUC) that ranged from 0.77 to 0.80 when analysed individually. Notably, when these metabolites were combined into a single predictive model, the AUC improved to 0.82 for all cases of PE and even reached 0.85 for EO-PE, indicating a strong predictive capacity. Furthermore, the detection rates of the combined model showed significant practical application, utilizing the combination of significant metabolites, were 50% at a 10% false positive rate (FPR) and 60% at a 20% FPR. For EO-PE, the combined model identified 50% of cases at a 10% FPR, and this detection rate raised to 70% when the FPR was increased to 20%.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Significant metabolites found in different matrixes.

https://doi.org/10.1371/journal.pone.0339292.t003

Austdal, M. et al [28] utilized Partial Least Squares Discriminant Analysis (PLS-DA) to predict PE, GH, and their combination in pregnant women during the first trimester. The results showed varying prediction accuracy at a 10% FPR: 15% for PE, 33% for GH, and 30% for both combined. Key metabolites linked to these hypertensive disorders included elevated lipid levels, particularly triglycerides, and decreased levels of phosphatidylcholines (especially those associated with high-density lipoproteins or HDL). Additionally, lower glucose, lactate and alanine levels were significant in distinguishing hypertensive conditions (Table 3). The relevance of the variability in these metabolites was quantified with variable importance in projection (VIP) scores. Despite the results indicating a correlation between lipid metabolism alterations and hypertension development, the overall prediction accuracy in serum samples was lower compared to the urine analyses also performed in this study. This lower sensitivity was attributed to the challenges in detecting small molecular weight metabolites in serum due to factors like viscosity and signal overlap.

The study conducted by Bahado-Singh, R.O. et al. [25] focused on the validation of metabolomic models to predict EO- PE using serum samples collected in the first trimester. Participants were divided into two groups: a discovery group, comprising 30 EO-PE cases and 65 controls, which was utilized for developing the metabolomic biomarker models, and a validation group, containing 20 EO-PE cases and 43 PC, to test the predictive accuracy of the developed algorithms independently. The study employed NMR spectroscopy with an internal spectral database that permitted to analyse around 50 metabolites in each serum sample. Key findings indicated that the metabolite-only model achieved an AUC of 0.835, with a sensitivity of 75% and specificity of 74.4%. In this model, the significant biomarkers predictors were 2-hydroxybutyrate, 3-hydroxyisovalerate, acetone, citrate and glycerol (Table 3). When combined with UtAPI, the performance improved significantly, resulting in an AUC of 0.916, with sensitivity and specificity rates increasing to 90% and 88.4%, respectively. The validation of the models showed reproducibility, as the diagnostic accuracy in the validation group mirrored that of the discovery group, confirming the robustness of the algorithms developed.

Ferranti, E.P. et al. [23] identified specific metabolites in serum samples obtained during the first trimester as significantly differentially regulated between women who developed PE and those who developed GH. A total of 470 significant features were identified between PE and PC, and 388 were identified between GH and controls. A PLS-DA identified 169 significant features based on both p-value < 0.05, using generalized linear regression methods, and a VIP score > 2, between PE and GH. Only metabolites verified by Schymanski Level 1 criteria showing significant differences between PE and GH were reported in detail. They found increased levels of retinoate and kynurenine, and decreased levels of sn-glycerol-3-phosphocholine and 2’4’-dihydroxyacetophenone in PE compared to GH (Table 3). In addition to these metabolites, they also identified significantly enriched metabolic pathways compared to PC. These pathways included alterations in porphyrin metabolism, steroid hormone biosynthesis, retinol metabolism, and arachidonic acid metabolism in PE, as well as changes in fructose and mannose metabolism and various amino acid metabolism pathways in GH.

Five groups of populations were analysed by Harville, E.W. et al. [24] related to pregnancy complications in the first trimester: women with PE, GH, HDP and preterm birth (PTB) or spontaneous preterm birth (sPTB). Advanced techniques such as UPLC-HR-MS and NMR were used to identify metabolites from serum samples associated with these disorders. In the case of PE, several predictive metabolites, such as bolasterone, cerasinone and unidentified signals, significantly improved the model’s predictive ability with an AUC of 0.95. Additionally, the NMR analysis revealed that asparagine, N,N-dimethylglycine and trimethylamine were linked to lower probability of PE, with an AUC of 0.177. For GH, predictive signals were found using UPLC-HR-MS, such as 2,6-di-tert-butyl-4-hydroxymethylphenol, which showed an odds ratio (OR) of 2.57, with an AUC of 0.945, significantly improving the model. However, no predictive metabolites were identified with NMR. In HDP, which include both PE and GH, UPLC-HR-MS identified predictive signals such as glycochenodeoxycholic acid 3-glucuronide, which showed an OR of 5.43, and unidentified signals associated with lower odds of HDP, with an AUC of 0.954. NMR also contributed with signals associated with lower odds of HDP, such as the combination of asparagine/albumin, with an AUC of 0.0412. Finally, for PTB and sPTB, predictive metabolites were identified using UPLC-HR-MS, such as {3-[(2E)-3-phenylprop-2-enoyl]phenyl} oxidanesulfonic acid in PTB and anisoxide in sPTB, among others, which improved the AUC of the models, reaching 0.821 for PTB and 0.865 for sPTB (Table 3). NMR found metabolites like threonine and urea, associated with a reduced risk of PTB and sPTB, though they did not significantly contribute to the predictive value of the model.

3.2. PE metabolites in plasma

Kenny, L.C. et al. [33] conducted a two-phase metabolic profiling study using UPLC-MS to identify plasma metabolic biomarkers for predicting PE in the second trimester of pregnancy. In the first phase, the discovery phase, they identified 14 metabolites that were discriminant for PE, including fatty acids, acylcarnitines, and amino acid–related compounds, among others (Table 3). The multivariate model developed to predict these metabolites showed an odds ratio of 36 (95% Confidence Interval, CI: 12–108) for developing PE, with an AUC of 0.94. These findings were further validated in an independent case-control study, where the same 14 metabolites produced an odds ratio of 23 (95% CI: 7–73) and an AUC of 0.92. They performed a PLS-DA model using the data from both studies demonstrating that a combination of 14 metabolites that represents systemic interactions in the metabolome is sufficient to generate a robust predictive model with an AUC greater than 0.9. This model showed promising detection performance, with subsequent PE detection rates of 77% and 73% in the discovery and validation datasets at a 10% false-positive rate, and false-positive rates of 21% and 24% at a 90% detection rate.

The separation of metabolites into polar and apolar fractions allowed Sander, K.N. et al. [26] to identify 35 metabolites associated with PE at third trimester of pregnancy, using LC-MS, although not all could be fully characterized. A total of 4379 peaks corresponding to polar compounds and 3153 peaks for apolar compounds were initially detected. From the full dataset, 37 metabolites were selected through both univariate and multivariate analysis. After excluding two compounds identified as drugs, 35 remained as potential biomarkers. These showed clear group separation in Principal Component Analysis (PCA) (and OPLS-DA, with a highly significant model (R²X(cum) = 0.629, Q²(cum) = 0.648, p = 5.5 × 10^-12), achieving 91% prediction accuracy and 100% specificity and an AUC value of 0.964. Among them, nine compounds were tentatively identified using metabolomics databases. In the polar fraction, taurodeoxycholic acid isomer, methionine sulfoxide, 3-hydroxyanthranilic acid, N1-methyl-pyridone-carboxamide isomer, and urocanic acid were identified. The apolar fraction contained hydroxyhexacosanoic acid isomer, two types of diacylglycerols, and glycerophosphoinositol (Table 3).

Sovio, U. et al. [27] identified 4-hydroxyglutamate as a novel biochemical predictor of PE, demonstrating a significantly stronger association compared to existing biomarkers like PlGF and PAPP-A. They analised serum samples of women with PE (delivering at term and pre-term) and women without pre-term PE with 12, 20, 28 and 36 weeks of gestational age (wkGA). Statistical analyses revealed that among the 1193 untargeted metabolites screened, 4-hydroxyglutamate showed promising predictive capabilities (Table 3). Specifically, the predictive ability of 4-hydroxyglutamate at 12 wkGA had an AUC of 0.673. At 20 wkGA, it remained strong, and by 28 wkGA, its predictive power was further enhanced, resulting in an AUC of 0.848 when combined with C-glycosyltryptophan. Although C-glycosyltryptophan also demonstrated some potential as a predictor, its strong association with pre-term PE was primarily evident at 28 weeks, with less consistent results at earlier gestation periods. The addition of C-glycosyltryptophan to predictive models improved their overall capability, but its clinical utility is considered limited compared to that of 4-hydroxyglutamate. The predictive accuracy of these metabolites was validated in an external cohort of samples from Born, in Bradford, highlighting their robustness across diverse populations. This evidence suggested that 4-hydroxyglutamate could serve as a valuable biomarker for early identification of women at high risk for PE.

The metabolomics analysis conducted in the study of de Almeida, L.G.N. et al. [32] highlighted significant alterations in the metabolic profiles of plasma samples from participants in the third trimester of pregnancy in various groups: non-pregnant (NP), PC, GH and PE. A total of 52 metabolites were identified during the analysis. Among the metabolites examined, three specific metabolites (allantonin, glutamine and histidine) exhibited markedly elevated levels in the PE group compared to PC (Table 3). Specifically, allantoin showed significant increase, with higher intensities in PE (p < 0.01) when compared to PC, while glutamine and histidine also demonstrated significant increases in PE (p < 0.05). The distinct metabolic profiles associated with PE were further supported by hierarchical clustering analysis, which effectively separated the NP group from the other three groups (PC, GH, and PE).

3.3. PE metabolites in urine

Austdal, M. et al. [28] demonstrated that first-trimester urinary metabolomic profiles had the potential to predict both PE and GH. Specifically, multivariate analyses revealed that at a 10% false positive rate, PE could be predicted with a sensitivity of 51.3%, while GH was predicted with a sensitivity of 40%, and both conditions combined at 37%. Hippurate was found to be significantly decreased in women who subsequently developed either PE or GH. In women who later developed PE, the urinary levels of creatinine, glycine, 4-deoxythreonic acid, α-hydroxyisobutyrate, histidine and dimethylamine were increased, while hippurate, lactate, and proline betaine were decreased (Table 3). For those who developed GH, there was an additional decrease in urinary citrate excretion alongside the changes observed in PE. The PCA of the ¹H NMR spectra of urine samples showed a distinct clustering of women who were set to develop PE or GH, indicating a clear difference in their urinary metabolic profiles compared to PC. Notably, combining the urinary hippurate/creatinine ratio with maternal mean arterial pressure (MAP) and a variable for maternal age improved the prediction rates for PE in a logistic regression model (AUC of 0.778) compared to using UtAPI combined with MAP and age (AUC of 0.738).

3.4. PE metabolites in feces

Liu, X. et al. [31] revealed a significant alteration in the fecal metabolome in women with severe PE compared to PC. Through untargeted metabolomic analysis using LC-MS/MS, a total of 686 fecal metabolites were identified after normalization. Further analysis pointed 31 metabolite features exhibiting distinct enrichment patterns between the two groups. Notably, guanidoacetic acid and valine were found to be significantly more abundant in PC, while 5-deoxy-D-glucuronate, phenylpropanoate, agmatine, and N-acetylputrescine were enriched in the severe PE group (Table 3). KEGG pathway analysis indicated that metabolites with higher concentrations in PC were significantly enriched in arginine and proline metabolism and glycine, serine and threonine metabolism. Conversely, the inositol phosphate metabolism and pentose and glucuronate interconversions pathways were enriched by metabolites more abundant in severe PE. Furthermore, the study identified Limosilactobacillus fermentum as a key bacterial biomarker associated with severe PE. Notably, its abundance showed a significant negative correlation with the metabolites phenylpropanoate (p < 0.001) and agmatine (p < 0.05), suggesting a potential metabolic interaction. In addition, phenylpropanoate exhibited excellent diagnostic performance, with an AUC of 98.57%, in distinguishing women with severe PE from PC. Co-occurrence network analysis between gut bacteria and fecal metabolites showed strong inter-relationships, with metabolites more abundant in PC generally clustering separately from those elevated in severe PE. Limosilactobacillus fermentum showed a significant positive correlation with bacterial species and metabolites enriched in PC.

3.5. PE metabolites in placental tissue

The study of Austdal, M. et al.[29] identified 25 metabolites in a HR-MAS MRS analysis of placental biopsies. A univariate analysis revealed that 12 metabolites showed significant differences between placentas from pregnancies with PE and PC after correction for multiple testing. Specifically, in preeclamptic placentas, a decrease in the levels of ethanolamine, glutamine, glutamate, glycine, taurine, dihydroxyacetone, 3-hydroxybutyrate and ascorbate was observed, while the levels of glycerophosphocholine were found to be increased. A decrease in valine, threonine, and lysine was also observed in preeclamptic placentas. PCA indicated that an increase in aspartate, phosphocholine, and glycerophosphocholine, and a decrease in glutamate, taurine, ascorbate and glutamine corresponded with the preeclamptic phenotype. PLS-DA models showed that the levels of glycerophosphocholine, phosphocholine, and aspartate were increased, while the levels of ethanolamine, taurine, glutamate, ascorbate, and glycine were decreased in PE compared to PC (classification accuracy 92.7%, p < 0.001). In severe PE, an increase in choline, lysine, alanine, glucose, and myo-inositol, along with a decrease in 3-hydroxybutyrate, was observed compared to non-severe PE (accuracy 83.2%, p = 0.003) (Table 3).

In 2019, Austdal, M. et al. [30] studied three distinct metabolic placental groups through hierarchical clustering of metabolic profiles obtained by HR-MAS MRS: the normal placenta group, the moderate placental dysfunction group, which contained almost all LO-PE pregnancies and a substantial number of PC; and the severe placental dysfunction group, that included more women with EO-PE, fetal growth restriction (FGR) and preterm delivery, and lower placenta weight. The normal placenta group was metabolically distinct, showing significant differences in 21 metabolites involved in pathways like BCAAs degradation, glycine, serine, and threonine metabolism, and phospholipid biosynthesis. Compared to the normal placenta group, in the moderate placental dysfunction group, the levels of glycerophosphocholine, phosphocholine, aspartate, and myoinositol were significantly increased, while ethanolamine, dihydroxyacetone, taurine, acetate, isoleucine, lysine, leucine, valine, alanine, glucose, threonine, glycine, and glutamate were significantly decreased. In the severe placental dysfunction group, the levels of phosphocholine, aspartate, and creatine were significantly increased, while ethanolamine, glycerol, dihydroxyacetone, taurine, acetate, lysine, valine, glucose, 3-hydroxybutyrate, ascorbate, threonine, glycine, and glutamate were significantly decreased (Table 3).