Abstract

   Preeclampsia (PE) is a serious pregnancy complication that contributes
   to maternal and perinatal morbidity and mortality. Understanding its
   pathogenesis and revealing predictive biomarkers are essential for
   guiding treatment decisions. In order to explore the global changes of
   serum metabolites in PE patients and identify potential predictive
   biomarkers for suspected PE patients (pregnant women who had already
   shown PE-related symptoms in the middle to late stages of pregnancy,
   but were not yet confirmatively diagnosed as PE.), a large-scale serum
   metabolomic analysis was conducted in this study with a prospective
   cohort of 328 suspected PE patients in the middle or late pregnancy
   stages, as well as a retrospective cohort of 30 healthy pregnant women
   and 30 PE patients. Using liquid chromatography mass spectrometry
   (LC − MS), serum metabolomic profiling revealed that the development of
   PE was closely associated with disturbed amino acid metabolism.
   Moreover, a panel of seven predictive biomarkers including
   2-methyl-3-hydroxy-5-formylpyridine-4-carboxylate,
   gamma-glutamyl-leucine, 2-hydroxyvaleric acid, LysoPC(16:1(9Z)/0:0),
   PC(DiMe(13,5)/MonoMe(13,5)), ADP-D-glycero-beta-D-manno-heptose and
   phenylalanyl-tryptophan were identified for PE development by
   performing multiple statistical analysis and LASSO regression analysis.
   The combination of these biomarkers showed promise in the prediction of
   PE development for suspected PE patients, with an AUC of 0.753 and
   0.885 for the discovery and validation cohorts, respectively. These
   findings highlight the potential of large-scale prospective metabolomic
   studies combined with machine learning algorithms in identifying key
   biomarkers for predicting PE development, while retrospective
   metabolomics studies provide insights into the pathogenesis of PE.

Supplementary Information

   The online version contains supplementary material available at
   10.1038/s41598-025-87905-9.

   Keywords: Preeclampsia, Serum metabolomic study, Predictive biomarkers,
   Altered metabolic pathways, Machine learning algorithms

   Subject terms: Diseases, Biomarkers, Predictive markers

Introduction

   Preeclampsia (PE) is a common and serious complication during pregnancy
   and postpartum, characterized by new-onset hypertension (> 140/90 mmHg)
   and proteinuria (> 300 mg/24 h), typically after 20 weeks of pregnancy.
   It can also be superimposed on preexisting hypertension or renal
   disease. Currently, PE has been recognized as public health concern due
   to several reasons: (1) it is diagnosed in 2–5% of pregnancies, with a
   higher incidence for other forms of pregnancy-related
   hypertension^[58]1; (2) there is an urgent need to explore more
   effective diagnostic biomarkers to assist in clinical decision-making
   for suspected PE patients, because the traditional clinical indicators
   have limited predictive value for PE or its adverse pregnancy
   outcomes^[59]2,[60]3; (3) PE is a complex multisystem disease and
   triggers problems in the liver, kidney, brain, and the clotting
   system^[61]4. Worsely, PE increases the risk of long-term
   complications, including metabolic disease and cardiovascular disease,
   in both mothers and offspring^[62]5. (4) there is currently no cure for
   PE, and treatments focus mainly on relieving the symptoms and
   minimizing complications^[63]6.

   Recent research has focused on revealing the development of PE and
   identifying biomarkers for prediction. The current strategy for
   predicting preeclampsia is based on a combination of baseline maternal
   factors, biophysical parameters, and placental-associated
   proteins^[64]7. Although this combination is effective in predicting PE
   in the early stage of pregnancy, it falls short in predicting PE
   development for those women have already shown PE-related
   symptoms^[65]8. Studies suggested that PE might stem from altered
   maternal pattern of circulating placentally derived proteins that
   regulating angiogenesis^[66]9, a specific sFlt-1:PlGF ratio cutoff of
   38 has shown promise in predicting PE development for women with
   clinical suspicion of the condition^[67]10. However, evidence for the
   diagnostic effectiveness of the ratio in screening women without
   clinical suspicion of the disease is poor^[68]11.

   Metabolomics, as the final manifestation of integrated upstream
   biological information flow, determines the eventual phenotype.
   Compared to targeted metabolomics focusing on well-defined metabolites,
   the high-throughput untargeted metabolomics aims at monitoring all
   low-molecular-weight metabolites in a biological fluid and has been
   widely used to discover specific metabolic patterns of diseases^[69]12.
   Besides, recent work has highlighted the potential of machine learning
   (ML) algorithms for processing a large amount of data and screening the
   candidate biomarkers effectively. The combination of untargeted
   metabolomics and ML provides comprehensive insights into metabolic
   alterations, leading to improved predictions^[70]13–[71]15. Despite
   advancements in metabolomics for PE, there are still challenges for
   clinical application. The heterogeneity of metabolites contributed to
   poor specificity and low positive predictive values (PPV,
   8-33%)^[72]16, leading to unnecessary tests and interventions for
   false-positive patients. Additionally, previous PE predictive models
   were established between healthy controls (HC) and PE patients, the
   differences between suspected-PE^+ (patients with PE-related symptoms
   who are diagnosed with PE) and suspected-PE^− (patients with PE-related
   symptoms but did not develop to PE) were not revealed. Therefore, more
   diverse populations should be recruited to validate the biomarkers.

   In this work, an integrated work-flow for metabolome profiling of
   maternal serum was carried out to explore the global changes of serum
   metabolites in PE and identify potential predictive biomarkers for PE
   development. Firstly, HC, PE patients and suspected PE participants
   were recruited in independent experiments for untargeted serum
   metabolomics study. Subsequently, multivariate statistical analysis was
   employed to reveal the metabolic differences between HC vs. PE and
   suspected-PE^−vs. suspected-PE^+. Thirdly, predictive biomarkers for
   suspected PE patients were identified and validated in two cohorts. To
   our knowledge, this work is the first to reveal the candidate
   predictive biomarkers for suspected PE patients based on large-scale
   prospective serum metabolomics, we envision that the integrated
   researches could deepen our understandings of PE.

Results

Characteristics of serum metabolomic profiles in all groups based on
untargeted metabolomics

   Studies on determinants of PE in pregnant women suggested parity, age,
   pre-pregnancy BMI and gestational week at sampling as potential
   covariates of PE during pregnancy^[73]16–[74]18. As shown in
   Fig. [75]1, retrospective and prospective cohorts were constructed and
   the baseline characteristics including these potential covariates of
   the study population were summarized, no statistically significant
   differences were observed (Table [76]1). Following appropriate
   pre-treatment, serum samples were injected into UPLC − QE-MS for
   untargeted metabolomic analysis. A total of 8033 features were
   discovered in the positive mode, and 6291 features were detected in the
   negative mode (Figure [77]S1). After signal de-noising and dataset
   normalization, 1569 metabolites in the positive and 825 metabolites in
   the negative mode were annotated in the public database. A total of 187
   compounds were identified in both positive and negative ion modes
   (Fig. [78]2A). Among these annotated metabolites, lipids and lipid-like
   molecules were the most abundant, accounting for 41.31% of the total
   metabolites. They were followed by organic acids and derivatives, organ
   heterocyclic compounds, benzenoids, organic oxygen compounds,
   phenylpropanoids and polyketides and organic nitrogen compounds, among
   others (Fig. [79]2B).

Fig. 1.

   [80]Fig. 1
   [81]Open in a new tab

   The schematic diagram represents the strategy for discovery and
   validation of the serum candidate biomarkers for PE prediction.

Table 1.

   The baseline characteristics of the study population.
   Age, mean ± SD Pre-pregnancy BMI Sampling week Diagnosis week
   Retrospective cohort HC (n = 30) 32.83 ± 4.39 23.84 ± 3.17 28.73 ± 4.47
   PE (n = 30) 33.33 ± 4.86 25.35 ± 4.02 29.23 ± 3.79 35.90 ± 2.95
   p^a value 0.677 0.112 0.642
   Discovery cohort Suspected-PE^−(n = 215) 32.75 ± 4.63 24.18 ± 4.29
   30.45 ± 4.19
   Suspected-PE^+ (n = 47) 33.91 ± 4.47 25.11 ± 4.77 28.93 ± 5.27
   36.17 ± 2.91
   p^a value 0.117 0.095 0.067
   Validation cohort Suspected-PE^−(n = 56) 34.00 ± 4.88 23.92 ± 3.81
   28.41 ± 4. 52
   suspected-PE^+ (n = 10) 31.90 ± 4.04 24.43 ± 4.81 31.40 ± 3.53
   35.40 ± 5.19
   p^a value 0.204 0.705 0.052
   p^b value 0.413 0.307 0.193 0.779
   [82]Open in a new tab

   ^aCompared inter-groups; ^b: compared intra-groups.

Fig. 2.

   [83]Fig. 2
   [84]Open in a new tab

   Veen diagram of metabolites detected in the positive ion mode and
   negative ion mode (A); Types and proportions of metabolites detected in
   untargeted metabolomics (B); Score scattering plot yielded from partial
   least-squares discrimination analysis (PLS-DA) on the quantitative
   metabolome datafile of serum samples from HC, PE, suspected-PE^− and
   suspected-PE^+ groups (C).

   The metabolites collected in the positive and negative ion modes were
   combined for multivariate statistical analysis. Unfortunately, the
   unsupervised pattern recognition method was not sufficient to
   differentiate the four groups (Figure [85]S2). Subsequent partial least
   squares discriminant analysis (PLS-DA) analysis revealed that HC
   samples mainly located in the first quadrant, with mild variations
   observed for suspected-PE^− samples. On the other hand, significant
   intra-group variations were observed for suspected-PE^+ samples,
   particularly for PE samples. It is noteworthy that during the
   development of PE, there is a gradual deviation from normal
   physiological patterns (Fig. [86]2C). This suggests that certain
   candidate biomarkers could be utilized to reflect the
   pathophysiological progression of PE. To validate the reliability of
   the developed PLS-DA model, a permutation test was carried out and no
   overfitting was observed (Figure S3). Additionally, an unsupervised
   principal component analysis (PCA) was performed to evaluate the
   variance of QC samples (Figure S4). The results indicated that the QC
   samples clustered closely within a range of no more than two standard
   deviations, demonstrating the stability of the analysis system and the
   high quality of the data.

Differential metabolites and dysregulated metabolic pathways between healthy
pregnant women and PE patients

   In order to enhance the diversity among groups in metabolomics analysis
   and identify key differential metabolites, supervised orthogonal
   partial least squares discriminant analysis (OPLS-DA) was performed.
   The scattering plot showed a clear separation between HC and PE
   (Fig. [87]3A). To avoid overfitting, a 200 times permutation test was
   carried out and the results suggested that the supervised model was
   able to provide an objective comparison between HC and PE (Figure S5).
   Metabolites with VIP value above 1.5 were dispersed from the origin in
   the loading plot (Fig. [88]3B). Additionally, through univariate
   nonparametric Wilcoxon’s analysis, 98 differential metabolites with
   VIP > 1.5, p < 0.05 and fold change (FC) > 1.2 / FC < 0.8 were
   recognized as the primary contributors to group classification
   (Fig. [89]3C).

Fig. 3.

   [90]Fig. 3
   [91]Open in a new tab

   Orthogonal partial least-squares discrimination analysis (OPLS-DA) of
   serum samples from HC and PE groups: score scattering plots (A) and
   S-plot loading plots (B); Volcano plot analysis of significantly
   altered metabolites in the serum samples of HC versus PE (C); ROC
   analysis of the candidate biomarkers for PE prediction in the
   retrospective cohort (D).

   Among the 98 metabolites, 73 were up-regulated and 25 were
   down-regulated in the PE group. These significantly regulated
   metabolites fell into diverse categories of structural identities,
   including organic compounds, bile acids, steroid hormones, amino acids,
   nucleotides, and purine metabolites. The top 5 metabolites with the
   highest FC in both upregulation and downregulation were highlighted in
   Fig. [92]3C. Notably, citric acid, taurocholic acid,
   3-dehydrosphinganine, glycocholic acid and 3-phosphonooxypyruvate were
   significantly elevated in PE, while cortisol, estrone sulfate,
   dehydroepiandrosterone sulfate, L-phenylalanine and L-cysteine were
   significantly reduced compared to HC. Afterwards, a genetic
   algorithm-based optimization evolutionary method was employed to
   identify discriminant metabolites, with the top metabolites selected
   from 200 repeated 10-fold cross-validations as potential biomarkers to
   enhance the reliability and clinical applicability of the prediction
   models. And a LASSO machine learning model was utilized for variable
   selection to yield concise outcomes (Figure S6). Ultimately, a total of
   10 metabolites, including dehydroepiandrosterone sulfate, cortisol,
   valproic acid glucuronide, L-proline, 3-hydroxybutyric acid,
   L-histidine, 2-hydroxybutyric acid, L-glutamic acid, citric acid and
   3-phosphonooxypyruvate were identified (Table [93]S1). As depicted in
   Fig. [94]3D, these selected biomarkers exhibited an area under the
   curve (AUC) ranging from 0.635 to 0.910 in the retrospective cohort.
   Unexpectedly, the combination of biomarkers showed outstanding accuracy
   (AUC = 1).

   Subsequently, the up-regulated and down-regulated metabolites were
   analyzed using MetaboAnalyst 5.0 to identify enriched pathways,
   respectively. We found that the down-regulated metabolites were mainly
   enriched in steroid hormone biosynthesis (p < 0.05, FDR < 0.25)
   (Fig. [95]4A, Table [96]S2). On the other hand, the up-regulated
   metabolites were associated with pathways such as phenylalanine,
   tyrosine and tryptophan biosynthesis, nitrogen metabolism,
   phenylalanine metabolism, glyoxylate and dicarboxylate metabolism,
   glycine, serine and threonine metabolism, biosynthesis of unsaturated
   fatty acids, arginine biosynthesis, butanoate metabolism and histidine
   metabolism (Fig. [97]4B, Table S3). However, some metabolites can
   participate in multiple metabolic pathways. To explore the potential
   functional relationships between these key metabolites, a
   metabolite-metabolite interaction network was built. As shown in
   Fig. [98]4C, key metabolites such as L-glutamic acid, citric acid,
   L-phenylalanine, L-cysteine and L-proline, which had high degrees in
   the network, appeared to play critical roles in connecting different
   pathways, indicating the dysregulated amino acid metabolism pathway is
   the premise of various dysregulated metabolic pathways in PE.

Fig. 4.

   [99]Fig. 4
   [100]Open in a new tab

   Enrichment pathway analysis of the metabolites that were differentially
   downregulated (A) or upregulated (B) in PE versus HC;
   metabolite–metabolite interaction network analysis of the differential
   metabolites between PE and HC (C).

Biomarkers discovery between suspected-PE^− and suspected-PE^+ groups

   Predicting the development of PE in the patients with PE-related
   symptoms is crucial in clinical practice. Therefore, a prospective
   study was conducted with a large sample size of suspected-PE pregnant
   women (n = 336), which were randomly divided into discovery cohort and
   validation cohort. The discovery cohort aimed to reveal the metabolic
   differences between suspected-PE^− and suspected-PE^+ groups,
   nevertheless, initial analysis using PCA did not show clear
   distinctions between the groups (Figure S7). Subsequently, a supervised
   OPLS-DA model was employed to highlight the differences more
   effectively. The model demonstrated a slight separation in metabolite
   profiles between the two groups (Fig. [101]5A), supported by
   satisfactory permutation results between two similar groups (Figure
   S8). Suspected-PE^− samples were predominantly distributed on the
   positive half axis of the X-axis, while suspected-PE^+ samples were
   mainly on the negative half axis of the X-axis. According to the
   corresponding loading plots, 117 variables with VIP > 1.5 were retained
   and presented in the Fig. [102]5B. Moreover, volcano plots were used to
   visualize statistical significance and fold change values, leading to
   the identification of 19 differential metabolites meeting specific
   criteria (VIP > 1.5, p < 0.05, and FC > 1.2 / FC < 0.8) (Fig. [103]5C,
   Table S4). According to the significantly altered metabolites,
   metabolite set enrichment analysis (MSEA) and KEGG pathway analysis
   were used to determine the altered metabolic pathways in
   suspected-PE^+. As a result, the enriched pathways mainly involved in
   vitamin B6 metabolism, glutathione metabolism, arginine and proline
   metabolism and warburg effect (Fig. [104]5D).

Fig. 5.

   [105]Fig. 5
   [106]Open in a new tab

   Orthogonal partial least-squares discrimination analysis (OPLS-DA) of
   serum samples from suspected-PE^− and suspected-PE^+ groups: score
   scattering plots (A) and S-plot loading plots (B); Volcano plot
   analysis of metabolites significantly altered in the serum samples of
   suspected-PE^− versus suspected-PE^+ (C), Enrichment pathway analysis
   of the metabolites that were differentially altered in suspected-PE^−
   versus suspected-PE^+ (D).

Development and validation of the diagnostic model for PE in suspected
patients

   Systematic metabolomic investigations revealed metabolic alteration in
   serum between suspected-PE^− and suspected-PE^+ individuals.
   Differential metabolites were analyzed using logistic regression models
   with constrained parameters as in LASSO to establish the prediction
   models. A panel of seven potential biomarkers including
   2-Methyl-3-hydroxy-5-formylpyridine-4-carboxylate,
   gamma-glutamyl-leucine, 2-hydroxyvaleric acid, LysoPC(16:1(9Z)/0:0),
   PC(DiMe(13,5)/MonoMe(13,5)), ADP-D-glycero-beta-D-manno-heptose and
   phenylalanyl-tryptophan, was selected for predicting PE diagnosis
   (Fig. [107]6A). These differentially altered metabolites fell into
   diverse categories of structural identities, including
   pyridinecarboxylic acids and derivatives, amino acids, peptides, and
   analogues, fatty acids and conjugates, glycerophosphocholines, and
   purine nucleotide sugars. The corresponding intercept and coefficients
   in the LASSO model were summarized in Table [108]2. As illustrated in
   Fig. [109]6B, levels of
   2-Methyl-3-hydroxy-5-formylpyridine-4-carboxylate was significantly
   down-regulated from suspected-PE^− to suspected-PE^+, while the levels
   of gamma-glutamyl-leucine, 2-hydroxyvaleric acid, LysoPC(16:1(9Z)/0:0),
   PC(DiMe(13,5)/MonoMe(13,5)), ADP-D-glycero-beta-D-manno-heptose and
   phenylalanyl-tryptophan were significantly elevated in suspected-PE^+.

Fig. 6.

   [110]Fig. 6
   [111]Open in a new tab

   The minimum penalty coefficient model constructed using the LASSO
   regression model (A); Comparison between suspected-PE^− (n = 215) and
   suspected-PE^+ (n = 47) groups towards relative abundance of candidate
   biomarkers.

Table 2.

   The metabolites with the highest potential prognostic significance were
   identified by LASSO regression analysis.
   Metabolites Coefficient HMDB ID Supper class Sub class
   (Intercept) 4.41
   2-methyl-3-hydroxy-5-formylpyridine-4-carboxylate 2.4e^− 7 HMDB0006954
   Organoheterocyclic compounds Pyridinecarboxylic acids and derivatives
   gamma-glutamyl leucine − 7.2e^− 8 HMDB0011171 Organic acids and
   derivatives Amino acids, peptides, and analogues
   2-hydroxyvaleric acid − 3.5e^− 9 HMDB0001863 Lipids and lipid-like
   molecules Fatty acids and conjugates
   LysoPC(16:1(9Z)/0:0) − 1.6e^− 9 HMDB0010383 Lipids and lipid-like
   molecules Glycerophosphocholines
   PC(DiMe(13,5)/MonoMe(13,5)) − 9.1e^− 8 HMDB0061455 Lipids and
   lipid-like molecules Glycerophosphocholines
   ADP-D-glycero-beta-D-manno-heptose − 2.4e^− 7 HMDB0029952 Nucleosides,
   nucleotides, and analogues Purine nucleotide sugars
   Phenylalanyl-tryptophan − 8.6e^− 8 HMDB0029006 Organic acids and
   derivatives Amino acids, peptides, and analogues
   [112]Open in a new tab

   ROC analysis and ROC AUC (area under curve) were utilized to assess the
   diagnostic performance of all parameters. In the ROC analysis of
   discovery cohort, AUCs of seven biomarker candidates were presented in
   descending order: phenylalanyl-tryptophan, gamma-glutamyl-leucine,
   2-methyl-3-hydroxy-5-formylpyridine-4-carboxylate,
   PC(DiMe(13,5)/MonoMe(13,5)), lysoPC(16:1(9Z)/0:0), 2-hydroxyvaleric
   acid and ADP-D-glycero-beta-D-manno-heptose, with the best AUC of 0.673
   (95% CI, 0.597–0.750). Multivariate ROC curve analysis may yield a more
   effective approach for creating and evaluating predictive biomarker
   models compared to univariate ROC curve analysis, it yielded an AUC of
   0.753 (95% CI, 0.683–0.824) and achieved a sensitivity of 67.91%, a
   specificity of 70.21%, a PPV of 80.86% and an NPV of 65.44% at the
   cutoff determined by the Youden’s index (Fig. [113]7A; Table [114]3).
   In the independent validation cohort, the panel of biomarker candidates
   showed the best predictive power with an AUC of 0.885 (95% CI,
   0.789–0.982). Using the cutoff determined by the Youden’s index, the
   biomarker panel showed a sensitivity of 76.79%, a specificity of
   80.00%, a PPV of 93.26% and an NPV of 75.77% (Fig. [115]7B;
   Table [116]3).

Fig. 7.

   [117]Fig. 7
   [118]Open in a new tab

   The ROC analysis of the candidate biomarkers for PE prediction in the
   discovery (A) (suspected-PE^+n = 47, suspected-PE^−n = 215) and
   validation (B) (suspected-PE^+n = 10, suspected-PE^−n = 56) cohorts.

Table 3.

   ROC analysis of the potential biomarkers in PE prediction.
   Gamma-glutamyl- leucine 2-hydroxyvaleric acid lysoPC(16:1(9Z)/0:0)
   PC(DiMe(13,5)/MonoMe(13,5)) Phenylalanyl-tryptophan
   2-Methyl-3-hydroxy-5-formylpyridine-4-carboxylate
   ADP-D-glycero-beta-D-manno-heptose Combination
   Discovery cohort (n = 262)
    AUC (95%CI) 0.6639 (0.5771–0.7508) 0.6113 (0.5181-0.7045) 0.6363
   (0.5453–0.7273) 0.6391 (0.5530-0.7252) 0.6737 (0.5972–0.7503) 0.6582
   (0.5715-0.7448) 0.6038 (0.5172–0.6903) 0.7539 (0.6833–0.8245)
    Sensitivity,% 66.05 57.21 60.47 61.86 63.26 68.69 56.28 67.91
    Specificity,% 63.83 57.45 61.70 61.70 63.83 61.70 57.45 70.21
    PPV,% 74.83 59.47 82.77 83.40 90.69 76.31 56.94 80.86
    NPV,% 56.65 55.50 52.16 52.29 55.60 54.09 56.78 65.44
   Validation cohort (n = 66)
    AUC (95%CI) 0.7482 (0.5925–0.9039) 0.6286 (0.4035-0.8536) 0.7929
   (0.6272–0.9585) 0.7339 (0.5825-0.8853) 0.6214 (0.4518–0.7910) 0.7268
   (0.4518-0.7910) 0.6393 (0.4353-0.8433) 0.8857 (0.7894–0.9820)
    Sensitivity,% 85.71 91.07 75.00 66.07 64.29 87.50 53.57 76.79
    Specificity,% 60.00 50.00 70.00 60.00 50.00 60.00 60.00 80.00
    PPV,% 81.70 64.55 80.42 81.7 79.69 68.63 76.27 93.26
    NPV,% 61.91 84.83 60.43 61.91 61.17 82.76 57.01 75.77
   [119]Open in a new tab

   PPV, positive predictive value; NPV, negative predictive value.

Discussion

   Hypertension and proteinuria are critical clinical symptoms of PE, as
   well as major objective diagnostic indicators for this condition.
   However, the presence of proteinuria or hypertension does not always
   lead to the development of PE, and the metabolic alterations in
   suspected-PE^− and suspected-PE^+ remain poorly defined. In this
   untargeted UHPLC − MS metabolomics study, both retrospective and
   prospective cohorts were included and strict inclusion criterias were
   used to recruit healthy controls, PE and suspected-PE patients. After
   metabolomic profiling and systematic comparisons, significantly
   different serum metabolomic patterns were observed between HC and PE
   groups, while the metabolic patterns were similar between
   suspected-PE^− and suspected-PE^+. Pathway enrichment analysis combined
   with metabolite-metabolite interaction network analysis revealed that
   altered amino acids bridged the various dysregulated metabolic
   metabolites during PE, and the endpoint of interaction network mainly
   manifested as significant downregulation of steroid hormone metabolic
   pathways. Additionally, we constructed the predictive models based on
   differential metabolites between suspected-PE^− and suspected-PE^+
   using machine learning algorithms, and then developed a consensus model
   with satisfactory predictive ability.

   In the perturbed metabolic network of PE patients, amino acids played a
   central role and connected the entire network. Among them, L-glutamic
   acid, a key regulator of glutathione metabolism, exhibited the highest
   degrees in the metabolic network. According to previous
   studies^[120]19,[121]20, aberrant glutathione homeostasis also
   contributes to complications in PE. Therefore, glutathione metabolism
   is one of the key pathways to target in PE. Except for amino acids,
   endogenous molecules including arachidonate and linoleate also served
   as important communication points. Over-accumulation of linoleate and
   arachidonate in PE indicated excessive biosynthesis of unsaturated
   fatty acids, leading to endothelial dysfunction due to oxidative stress
   and inflammation^[122]21. Moreover, metabolites like taurocholic acid,
   glycocholic acid, estrone sulfate and dehydroepiandrosterone sulfate,
   which are related to cholesterol metabolism, showed the highest FC
   between HC and PE. Interestingly, bile acids and steroid hormones
   exhibited opposite trends of variations, warranting further
   investigation.

   Metabolomics combined with ML algorithm has been increasingly
   implemented for developing diagnostic models for various human diseases
   in recent years^[123]22–[124]24. Systematic metabolomic investigations
   revealed metabolic alteration in plasma between HC and PE, as well as
   suspected-PE^− and suspected-PE^+ individuals. Diagnostic models based
   on differential metabolites were established for PE using machine
   learning algorithms. For the comparison between HC and PE, the selected
   biomarkers showed promising results with an AUC ranging from 0.635 to
   0.910 in the retrospective cohort, but further validation with a larger
   population from multiple-center is necessary for clinical application.

   Currently, the sFlt-1/PlGF ratio cutoff of 38 has been employed for the
   short-term prediction of PE in Asian women with suspected PE with good
   performance, which has also been validated in our previous research
   (training cohort AUC = 0.637; validation cohort AUC = 0.733). Despite
   the similar metabolic profiles between suspected-PE^− and
   suspected-PE^+ were found in this research, a panel of seven features
   were screened as predictive biomarkers with higher diagnostic
   efficiency (training cohort AUC = 0.753; validation cohort
   AUC = 0.885). Interestingly, although these 7 biomarkers have not been
   previously reported in metabolomics studies on PE, they are all linked
   to the pathogenesis of PE. Inflammation, particularly in the
   preeclamptic placenta, has been highlighted as a key process. Molecules
   such as ADP-d-glycero-β-d-manno-heptose (ADP-heptose) and certain
   glycerophospholipids including lysoPC(16:1(9Z)/0:0) and
   PC(DiMe(13,5)/MonoMe(13,5)) have been implicated in modulating
   inflammation^[125]25–[126]27. Meanwhile, lipid over-accumulation in
   maternal serum contributes to endothelial dysfunction secondary to
   oxidative stress, and abnormal glucose metabolism and lipid metabolism
   often occur in parallel^[127]28. Moreover, increased 2-hydroxyvaleric
   acid, gamma-glutamyl-leucine and phenylalanyl-tryptophan were detected
   in suspected-PE^+ samples. 2-hydroxyvaleric acid, a known organic acid
   in human fluids, has shown potential as a predictor of type 1
   diabetes^[128]29,[129]30. Additionally, gamma-glutamyl-leucine and
   phenylalanyl-tryptophan are bioactive peptides associated with obesity
   and type 2 diabetes^[130]31–[131]33. As an intermediate metabolite in
   vitamin B6 production,
   2-methyl-3-hydroxy-5-formylpyridine-4-carboxylate plays a role in
   regulating cellular homocysteine levels in the transsulfuration pathway
   and acetylcholine-induced endothelium-dependent relaxation, thereby
   helping to prevent insulin resistance and vascular dysfunction^[132]34.
   Lower levels of 2-methyl-3-hydroxy-5-formylpyridine-4-carboxylate in
   suspected-PE^+ groups indicated an increase in insulin resistance.

   This study has several inevitable limitations that warrant
   acknowledgment. Firstly, inadequate clinical information was available