Abstract

Background

   Postpartum hemorrhage (PPH) is the leading cause of maternal mortality
   worldwide, with uterine atony accounting for approximately 70% of PPH
   cases. However, there is currently no effective prediction method to
   promote early management of PPH. In this study, we aimed to screen for
   potential predictive biomarkers for atonic PPH using combined omics
   approaches.

Methods

   Collection of cervicovaginal fluid (CVF) samples from 27 women with
   atonic PPH and 32 women with normal delivery was performed for
   metabolomic (LC-MS/MS) and proteomic (LC-MS/MS) detection and
   subsequent confirmation experiments in this nested case-control study.
   Mass spectrum and enzyme-linked immunosorbent assays (ELISA) were used
   to validate significantly different metabolites and proteins for
   screening potential biomarkers of atonic PPH. Furthermore, multivariate
   logistic regressions were performed for the prediction of PPH using the
   identified biomarkers mentioned above, and the area under the curve
   (AUC) was computed.

Results

   We identified 216 and 311 metabolites under positive and negative ion
   modes, respectively, as well as 1974 proteins. The PPH group had
   significant differences in metabolites and proteins belonging to the
   β-alanine metabolic pathway. Specifically, the PPH group had
   downregulation of critical metabolites, including histidine and protein
   dihydropyrimidine dehydrogenase (DPYD). Kyoto Encyclopedia of Genes and
   Genomes (KEGG) and Gene Ontology (GO) functional enrichment analysis of
   significantly differentially expressed proteins revealed that atonic
   PPH was associated with T cell- and macrophage-related immune
   inflammatory responses. Furthermore, we verified that concentrations of
   histidine (350.85 ± 207.87 vs. 648.33 ± 400.87) and DPYD (4.01 ± 2.56
   vs. 10.96 ± 10.71), and immune cell-related proteins such as CD163
   (0.29 ± 0.19 vs. 1.51 ± 0.83) and FGL2 (5.98 ± 4.23 vs. 11.37 ± 9.42)
   were significantly lower in the PPH group. Finally, the AUC for
   independent prediction of PPH using CD163, histidine, DPYD, and FGL2
   are 0.969 (0.897-1), 0.722 (0.536–0.874), 0.719 (0.528–0.864), and
   0.697 (0.492–0.844), respectively. A relatively high predictive
   efficiency was obtained when using joint histidine, DPYD, CD163, and
   FGL2, with AUC = 0. 964 (0.822-1).

Conclusions

   This study suggested that immune inflammation may play a role in the
   occurrence of PPH. The metabolite histidine and proteins of DPYD,
   CD163, and FGL2 in CVF were associated with uterine atony and could be
   used as predictive biomarkers for atonic PPH.

   Keywords: Proteomic, Metabonomic, Atonic postpartum hemorrhage,
   Maternal mortality, Prediction

Introduction

   Postpartum hemorrhage (PPH) is the leading cause of maternal mortality
   worldwide, accounting for approximately 18% [[38]1], with 60 -70% of
   cases being atonic PPH (caused by uterine atony) [[39]2]. Early
   prediction and identification of PPH are the basis for timely and
   effective rescue, but currently there is a lack of effective prediction
   models for atonic PPH.

   The mechanism of atonic PPH is still unclear, which may be the main
   reason for the lack of effective predictive tools. Previous studies
   have only suggested a correlation between immune inflammation, uterine
   atony, and PPH. “Postpartum acute myometritis (PAM)” has been proposed
   to support the theory that acute massive inflammatory reactions led to
   atony of myometrium [[40]3]. Complement system activation, mast cell
   degranulation, and massive infiltration of neutrophils and macrophages
   into the myometrium occured in patients with amniotic fluid embolism
   and refractory PPH without an infectious etiology, leading to PPH with
   uterine atony [[41]3]. Our previous study has found that elevated
   levels of immune-inflammatory factors (basic fibroblast growth factor
   (basic FGF), interleukin-1 (IL-1), and macrophage colony-stimulating
   factor (M-CSF)) were significantly associated with an increased risk of
   atonic PPH in late pregnancy [[42]2], which was supported by Gallo
   D.M.’s research [[43]4]. However, in general, research on the mechanism
   of atonic PPH is still extremely limited, and more predictive potential
   biomarkers still need to be developed.

   Multi-omics approaches, such as the combined application of proteomics
   and metabolomics, have been widely used to reveal differentially
   regulated pathways in diseases and have contributed to understanding
   pathogenesis better and identifying biomarkers for disease prediction
   [[44]5, [45]6]. At present, the immune, metabolic or protein profile
   related to normal pregnancy has also been widely used in the
   exploration of pregnancy related diseases [[46]7–[47]9]. On this basis,
   multi-omics can effectively investigate multiple molecules (genes,
   proteins, and metabolites) simultaneously and obtain a large amount of
   data in a short period of time, thereby more effectively screening for
   the biomarkers with the best predictive efficiency [[48]8]. However,
   there is still no PPH related metabolic and protein profiles.

   Therefore, our study conducted a combined metabolomic and proteomic
   analysis of cervicovaginal fluid (CVF) in pregnant women to identify
   potential predictive biomarkers of atonic PPH, aiming to extend the
   discovery of biomarkers beyond cytokines.

Results

Clinical characteristics of participants

   Our study comprised 27 patients with atonic PPH (PPH group) and 32
   matched non-PPH cases, of which 11 pregnant women with PPH and 12
   pregnant women without PPH were in the discovery cohort and 16 pregnant
   women with PPH and 20 pregnant women without PPH were in the
   confirmation cohort (Fig. [49]1). The clinical characteristics of
   participants are presented in Table [50]1. There were no differences in
   high-risk factors for PPH between the two groups, including maternal
   age, BMI, HDP, GDM, fetal birth weight, induction of labor, and
   previous medical history (previous intrauterine surgery, PPH, and
   uterine diseases).

Fig. 1.

   [51]Fig. 1
   [52]Open in a new tab

   Flowchart of the study design, population, and experiment. A) Sample
   selection; B) Overview of proteomics and metabolomics data process,
   confirmation experiment and the construction of predictive models

Table 1.

   Clinical characteristics of pregnant women
   Characteristic PPH
   (N = 27) non-PPH
   (N = 32) P value
   Gestational age - w 39.77 ± 0.96 39.74 ± 0.80 0.893^a
   Maternal age - y 34.44 ± 3.09 34.47 ± 3.39 0.977^a
   Estimated blood loss - mL

   870.0

   (770.0, 1120.0)

   260.0

   (200.0, 307.5)
   < 0.001^b
   BMI before pregnancy - kg/m^2 26.83 ± 2.54 26.80 ± 3.34 0.972^a
   Way of conception 0.116^c
   Natural conception - No. (%) 15(55.6%) 24 (75.0%)
   Asisted reproduction - No. (%) 12 (44.4%) 8 (25.0%)
   HDP - No. (%) 2 (7.4%) 3 (9.4%) 0.073^c
   GDM - No. (%) 9 (33.3%) 9 (28.1%) 0.187^c
   Spontaneous labor or induction 0.224^c
   Spontaneous labor - No. (%) 11 (40.7%) 15 (46.9%)
   Induction of labor - No. (%) 16 (59.3%) 17 (53.1%)
   Primiparous - No. (%) 16 (59.3%) 18 (56.3%) 0.054^c
   Previous intrauterine surgery - No. (%) 6 (22.2%) 5 (15.6%) 0.420^c
   Previous PPH - No. (%) 1 (3.7%) 0 (0.0%) 0.272^c
   Previous uterine diseases - No. (%) 6 (22.2%) 6 (18.8%) 0.741^c
   Fetal birthweight - g

   3510.0

   (3340.0, 3680.0)

   3340.0

   (3095.0, 3565.0)
   0.089^b
   1 min Apgar Score 27 (100.0%) 32 (100.0%) 1.000^c
   [53]Open in a new tab

   ^a Student’s t-test, ^b Mann-Whitney U, ^c chi-quared test; BMI, body
   mass index; HDP, hypertension disorders of pregnancy; GDM, gestational
   diabetes mellitus; w, week; y, year

Results of differential metabolites and proteins analysis

   Compared with the non-PPH group, the differential metabolites under
   positive and negative ion modes in the PPH group are shown in
   Fig. [54]2A and B, with histidine being the most significantly
   differentiated metabolite. The OrthoPLSDA map results indicated that
   the samples were well-aggregated within the group, indicating that the
   data quality met the analysis requirements (Fig. [55]2C). The
   metabolites are listed in Table [56]2 and are mainly involved in the
   VEGF signaling pathway, platelet activation, vascular smooth muscle
   contraction, and the cAMP signaling pathway (Fig. [57]2D).

Fig. 2.

   [58]Fig. 2
   [59]Open in a new tab

   Significant differences in metabolites and proteins and analysis
   results. Analysis of significant differences in metabolite expression
   in positive ion mode (A) and negative ion mode (B) by multiple
   differences: The horizontal axis represents the log2 FC value of
   differential metabolites, which is the logarithmic value of the
   difference multiple of differential metabolites based on 2. The
   vertical axis represents significant differential metabolites. Red
   indicates upregulation of differential metabolites, while green
   indicates downregulation of differential metabolites; OrthoPLSDA map
   showing metabolomics results (C): T score [[60]1] represents principal
   component 1, Orthogonal T score [[61]1] represents principal component
   2, and the dots of the same color represent various biological
   replicates within a group, and the distribution status of the dots
   reflects the degree of difference between and within groups; Enrichment
   pathways for all differential metabolites (D); Histogram of
   significantly differentially expressed proteins (E), where blue
   represents the number of downregulated proteins and red represents the
   number of upregulated proteins; Cluster analysis of differentially
   expressed proteins (F), represented by a tree heat map; Biological
   process related GO enrichment pathway map (G) and KEGG enrichment
   pathway map (H) of significantly different proteins: (G) Each row
   represents a differential protein and each column represents a group of
   samples. Red represents significant upregulation, blue represents
   significant downregulation, and color depth indicates the degree of
   upregulation and downregulation. Proteins with similar expression
   patterns are clustered in the same cluster on the left. (H) The x-axis
   represents the p value of Fisher’s exact test (logarithmic to 10), and
   the y-axis represents the path name. The pathways involved in
   upregulation and downregulation of proteins are represented by the
   right and left bars respectively

Table 2.

   Statistics of POS and NEG metabolite between PPH and non-PPH
   ID Metabolite name KEGG
   ID SubClass PPH vs. non-PPH
   P value POS or NEG
   M263T821_1 Ifosfamide [62]C07047 Isofamides 0.015 POS
   M122T115 Benzamide [63]C09815 Benzoic acids and derivatives 0.016 POS
   M163T51_1 Coniferyl alcohol [64]C00590 Methoxyphenols 0.026 POS
   M156T521 Histidine [65]C00135 Amino acids, peptides, and analogues
   0.040 POS
   M138T488 4-nitrophenol [66]C00870 Nitrophenols 0.008 NEG
   M141T66 2-mercaptoethanesulfonic acid [67]C03576 Organosulfonic acids
   and derivatives 0.014 NEG
   M351T240 Prostaglandin i2 [68]C01312 Eicosanoids 0.024 NEG
   M383T50 Artesunate Sesquiterpenoids 0.037 NEG
   M251T924

   2,3-quinoxalinedione,

   1,4-dihydro-6,7-dinitro-
   Benzodiazines 0.047 NEG
   [69]Open in a new tab

   POS indicate positive ion mode; NEG indicate negative ion mode

   We identified 1974 proteins and compared them with differentially
   expressed proteins between the two groups (Fig. [70]2E and F). Compared
   to the non-PPH group, 29 significantly different proteins were
   expressed in the PPH group, including 27 downregulated and two
   upregulated proteins (Fig. [71]2E). GO and KEGG functional enrichment
   analyses of significantly differentially expressed proteins revealed
   that the differentially expressed proteins were associated with T cell-
   and macrophage-related immune inflammatory responses, lysosomal
   function, and energy metabolism (Fig. [72]2G and H). Table [73]3 lists
   two significantly upregulated proteins and ten significantly
   downregulated ones in the PPH group.

Table 3.

   Statistics of significantly differential proteins between PPH and
   non-PPH group
   Uniprot ID Protein name Gene name PPH vs. non-PPH
   P value Difference multiple Regulation
   [74]Q66K66 Transmembrane protein 198 TMEM198 0.005 2.080 Up
   [75]P25398 40 S ribosomal protein S12 RPS12 0.023 2.037 Up
   [76]Q9NZK5 Adenosine deaminase 2 ADA2 0.030 0.118 Down
   [77]P01889 HLA class I histocompatibility antigen, B alpha chain HLA-B
   0.034 0.132 Down
   [78]P20933 N(4)-(beta-N-acetylglucosaminyl)-L-asparaginase AGA 0.039
   0.176 Down
   [79]P50452 Serpin B8 SERPINB8 0.046 0.240 Down
   [80]Q96C86 m7GpppX diphosphatase DCPS 0.048 0.243 Down
   [81]P17050 Alpha-N-acetylgalactosaminidase NAGA 0.035 0.268 Down
   [82]Q12882 Dihydropyrimidine dehydrogenase [NADP (+)] DPYD 0.017 0.273
   Down
   [83]O95498

   Vascular

   non-inflammatory molecule2
   VNN2 0.030 0.297 Down
   [84]Q9Y3D6 Mitochondrial fission 1 protein FIS1 0.011 0.298 Down
   [85]P28065 Proteasome subunit beta type 9 PSMB9 0.011 0.303 Down
   [86]Open in a new tab

Results of correlation analysis

   Differentially expressed proteins and metabolites between the PPH and
   non-PPH groups were involved in 60 and 15 metabolic pathways,
   respectively (Fig. [87]3B). The two parts jointly participated in just
   one pathway, the β-alanine metabolism pathway (Fig. [88]3B).
   Hierarchical heatmaps represented the correlation coefficients of
   differential proteins and metabolites (Fig. [89]3A), and it was noted
   that metabolites involved in histidine metabolism and β-alanine
   metabolism were closely related to differential proteins. In the
   PPH/non-PPH-related network, four metabolites within the nodes were
   closely related to differentially expressed proteins (Fig. [90]3C),
   where histidine was associated with the highest number of
   differentially expressed proteins. In metabolomics, histidine was most
   closely related to β-alanine metabolism, which plays a vital role in
   oxidative stress and immune inflammation. In addition, the
   significantly different histidine-related proteins, including DPYD,
   CD163, FGL2, ADA2, GPX3, FIS1, NAGA, and so on, which are involved in
   immune cell and lysosomal functions and oxidative stress, are presented
   in Fig. [91]3D; Table [92]4.

Fig. 3.

   [93]Fig. 3
   [94]Open in a new tab

   Joint analysis results of significant differences in metabolites and
   proteins. The correlation coefficients of differential proteins and
   metabolites showed by layered heat maps (A); The Venn diagram (B) of
   differentially expressed proteins and metabolites involved in the
   pathway, where blue represents the proteome, yellow represents the
   metabolome, and the cross region represents metabolic pathways involved
   in both omics; A metabolite-protein interaction network centered on
   histidine (C), showing the correlation analysis network between
   significantly different proteins and significantly different
   metabolites; Boxplot showing significant different proteins related to
   histidine (D): Data were presented as the median and interquartile
   range (25-75%); * indicate statistical differences between two groups

Table 4.

   Significant differential proteins and metabolites which closely related
   to histidine between PPH and non-PPH
   Metabolite Protein Coeffcient P value Protein function
   Histidine DPYD 0.684 0.005 Pantothenate and CoA biosynthesis
   ADA2 0.780 0.003 Adenosine related metabolism, ATP biosynthesis
   AGA 0.881 0.000 Lysosome function (lysosomal acid hydrolases), glycan
   degradation
   NAGA 0.881 0.000 Lysosome function (lysosomal acid hydrolases), glycan
   degradation
   MANBA 0.949 0.000 Lysosome function (lysosomal acid hydrolases)
   SORT1 0.574 0.024 Lysosome function (lysosomal membrane peoteins)
   CD163 0.817 0.000 Macrophage function
   FGL2 0.614 0.011 Lymphocyte function
   PSMB9 0.653 0.025 Formation of immunoproteasomes (IFN-γ)
   [95]Open in a new tab

Molecular validation of significant differential expression between PPH group
and non-PPH group

   The confirmation cohort was used for the validation of proteins with
   significant differences noted in the aforementioned omics data, and the
   results demonstrated that the concentrations of histidine
   (350.85 ± 207.87 vs. 648.33 ± 400.87, P = 0.021) and DPYD (4.01 ± 2.56
   vs. 10.96 ± 10.71, P = 0.016) in the atonic PPH group were
   significantly lower than those in the non-PPH group (Fig. [96]4A and
   B). The proteins related to immune function and oxidative stress that
   were listed in Fig. [97]3 have also been validated. The concentrations
   of CD163 (0.29 ± 0.19 vs. 1.51 ± 0.83, P = 0.000) and FGL2 (5.98 ± 4.23
   vs. 11.37 ± 9.42, P = 0.047) were significantly lower in the atonic PPH
   group than in the control group (Fig. [98]4C and D). The concentrations
   of ADA (1.02 ± 0.68 vs. 1.35 ± 0.94, P = 0.252) and FIS1 (1.31 ± 1.47
   vs. 1.50 ± 1.52, P = 0.711) in the atonic PPH group were lower than
   those in the control group, while GPX3 (23.81 ± 12.03 vs. 20.10 ± 8.11,
   P = 0.278) had a higher concentration in the atonic PPH group, although
   there was no statistically significant difference between the two
   groups.

Fig. 4.

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

   Differentially expressed proteins validation, and their correlation
   with blood loss, gestational age, and maternal age. Boxplot showing
   concentration of histidine (A), DPYD (B), CD163 (C) and FGL2 (D)
   between the PPH and non-PPH group: Data were presented as the median
   and interquartile range (25-75%); Scatter plot of correlation
   coefficients showing the correlation between histidine (A), CD163 (C),
   FGL2 (D) with blood loss, gestational age, and maternal age between the
   PPH and non-PPH group, respectively. DPYD, dihydropyrimidine
   dehydrogenase; CD163, scavenger receptor cysteine-rich type 1 protein
   M130; FGL2, fibroleukin; PPH, postpartum hemorrhage; * indicate
   statistical differences between two groups

   Moreover, we analyzed the correlation between histidine, DPYD, CD163,
   FGL2, ADA, GPX3, and FIS1 with blood loss, gestational age, and
   maternal age, respectively, as shown in Fig. [101]4. Histidine
   (R=-0.420, P = 0.019) and CD163 (R=-0.740, P = 0.000) exhibited a
   significant correlation with blood loss, while there was no
   statistically significant difference in the correlation between
   histidine, DPYD, CD163, and FGL2 with maternal age and gestational age
   (Fig. [102]4A-D).

   Finally, we analyzed the correlation between the proteins mentioned
   above and noted that CD163 was significantly positively correlated with
   histidine (R = 0.420, P = 0.020), GPX3 was significantly negatively
   correlated with DPYD (R=-0.410, P = 0.014). There was no significant
   correlation between other proteins (Fig. [103]5).

Fig. 5.

   Fig. 5
   [104]Open in a new tab

   Correlation analysis between significantly different proteins. Scatter
   plot of correlation coefficients showing the correlation between CD163
   (A), FGL2 (B), ADA (C), GPX3 (D) and FIS1 (E) with histidine and DPYD
   between the PPH and non-PPH group, respectively. CD163, scavenger
   receptor cysteine-rich type 1 protein M130; FGL2, fibroleukin; ADA,
   adenosine deaminase; GPX3, glutathione peroxidase 3; FIS1,
   mitochondrial fission 1 protein; DPYD, dihydropyrimidine dehydrogenase;
   PPH, postpartum hemorrhage

The prediction efficiency of histidine, DPYD, CD163 and FGL2 for atonic PPH

   Then, we constructed prediction models using significantly different
   metabolites and proteins and further compared the efficiency of four
   integrated models (Fig. [105]6). The AUC for independent prediction of
   PPH using CD163, histidine, DPYD, and FGL2 are 0.969 (0.897-1), 0.722
   (0.536–0.874), 0.719 (0.528–0.864), and 0.697 (0.492–0.844),
   respectively (Fig. [106]6A). In addition, we established a prediction
   model combining histidine and DPYD, which obtained a relatively high
   AUC of 0.887 (Fig. [107]6B). Then T-statistic as a feature ranking
   method was used to construct prediction models randomly combining
   histidine, DPYD, CD163, and FGL2 (Fig. [108]6C). The highest predictive
   efficiency was obtained when using histidine and three proteins
   (AUC = 0.964 [0.822-1]). The combination of histidine and CD163
   obtained the highest efficacy among every two biomarker combinations
   (AUC = 0.926 [0.766-1]), while a prediction model combining histidine,
   DPYD and CD163 obtained a relatively high AUC of 0.949 (0.783-1).

Fig. 6.

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

   Prediction models for atonic PPH using histidine, DPYD, CD163 and FGL2.
   ROC curves of histidine, DPYD, CD163 and FGL2 in the prediction of
   atonic PPH (A); ROC curves of histidine combined with DPYD (B) in the
   prediction of atonic PPH; ROC curves of histidine combined with CD163,
   histidine combined with DPYD and CD163, histidine combined with DPYD,
   CD163 and FGL2 (C) in the prediction of atonic PPH. AUC, area under the
   curve; DPYD, dihydropyrimidine dehydrogenase; PPH, postpartum
   hemorrhage; ROC, receiver operating characteristic

Discussion

   Our previous study has demonstrated a correlation between immune
   inflammation and the occurrence of atonic PPH [[111]2]. This
   prospective cohort study continued to explore the predictive biomarkers
   of atonic PPH using a joint omics approach. The biomarkers we
   identified exhibited good predictive efficiency. A relatively high
   predictive efficiency was obtained when using joint histidine, DPYD,
   CD163, and FGL2.

   Previous studies have suggested that atonic PPH was associated with
   abnormal activation of immune inflammation [[112]2–[113]4]. Amino acid
   metabolism might play an important role in immune cell activation by
   maintaining the redox state to promote the production of inflammatory
   cytokines [[114]10, [115]11]. Then redox and immune-inflammation
   balance could influence uterine muscle contraction [[116]12–[117]14].
   Our combined omics results highlighted the β-alanine metabolism, in
   which histidine and DPYD were the differential metabolite and protein
   related to β-alanine, respectively. Histidine is a component of
   β-alanine metabolism, while DPYD participates as a rate limiting enzyme
   in β-alanine synthesis for decomposing uracil and thymine into
   β-alanine [[118]15]. Excessive activation of pro-inflammatory cytokines
   caused by abnormal metabolism of β-alanine and histidine, that may lead
   to uterine atony. On one hand, β-alanine and histidine can inhibit
   pro-inflammatory factors. The decrease in histidine was negatively
   associated with proinflammatory cytokines such as TNF-α and IL-6
   [[119]16, [120]17]. β-alanine has also been linked to inflammatory
   cytokines such as IL-1β, IL-6, and iNOS [[121]18, [122]19]. On the
   other hand, β-alanine and histidine affect immune cell activation by
   regulating ROS levels [[123]11, [124]20–[125]24]. In our study, the
   expression of histidine and DPYD was decreased in the PPH group, which
   may promote the occurrence of atonic PPH through oxidative stress and a
   pro-inflammatory status. Moreover, Cierny et al. [[126]25] identified a
   relation between the level of pro-inflammatory cytokines and longer
   labor, which might be explained by our study that a decrease in
   metabolite histidine could promote an increase in pro-inflammatory
   cytokines, which in turn affected the contractility of uterine smooth
   muscle and led to prolonged labor. Therefore, we used the key
   metabolite (histidine) and key protein (DPYD) in β-alanine metabolism
   to predict atonic PPH and obtained a relatively high predictive
   efficiency.

   In addition, there were also differences in proteins related to immune
   function and oxidative stress, such as CD163, FGL2, ADA, FIS1, and
   GPX3, between the two groups. CD163, as a specific biomarker for
   macrophages, is involved in various immunomodulatory functions such as
   the production of anti-inflammatory and anti-oxidative substances
   (IL-10, ferritin, bilirubin, and CO) [[127]26] and promotion of a
   TH2-type immune response [[128]27, [129]28]. Previous studies have
   demonstrated that uterine myometrial contraction required aseptic
   inflammatory activation [[130]29]. Our study revealed that CD163 was
   significantly reduced in atonic PPH, which may affect uterine
   contractions by affecting immune inflammation. Soluble CD163 (sCD163)
   has been identified as a potential inflammation biomarker that was used
   in predicting the incidence rate of diseases, such as acute-on-chronic
   liver failure (ACLF) [[131]30]. Our study revealed that the predictive
   efficiency of the prediction model constructed solely using CD163 could
   reach as high as 0.969, indicating that CD163 might play an important
   role in atonic PPH and had excellent predictive ability. In addition,
   FGL2 is expressed in the human myometrium and maintains myometrial
   function during labor [[132]31]. Soluble fibrinogen-like protein 2
   (sFGL2) could also selectively induce pro-repair macrophage
   polarization [[133]32]. When the FGL2 level was abnormal, oxidative
   stress could be affected by regulating ROS production, thereby
   affecting the inflammatory response [[134]33]. Our study also confirmed
   a significant decrease in the expression of FGL2 in atonic PPH,
   suggesting that the reduction of FGL2 had a potential impact on uterine
   myometrial contraction and promotes the occurrence of atonic PPH.
   However, the predictive performance of FGL2 is inferior to that of
   CD163 (AUC = 0. 697).

   Previous predictive models for PPH were mainly constructed based on
   clinical information. Most of them had poor predictive ability (AUC
   ranged from 0.59 to 0.8) [[135]34]. Some PPH models constructed using
   coagulation indicators such as fibrinogen, D-dimer, and platelets did
   not exhibit a significant correlation [[136]35]. Currently, there is
   still a lack of molecular predictive models for atonic PPH. Only a few
   studies used immune inflammation-related biomarkers to predict the
   occurrence of PPH. For example, Jiang et al. constructed predictive
   models with efficacy of 0.88 and 0.84 using 12 and 3 cytokines,
   respectively [[137]2], while Gallo DM et al. constructed predictive
   models with efficacy of 0.71–0.76 using cytokines and high-risk factors
   [[138]4]. Compared with the above models, our predictive model was the
   first to use joint omics to screen a wider range of biomarkers and
   achieve higher predictive efficiency (AUC = 0.964).

   The key strengths of this study are that the CVF can effectively
   reflect the local immune-inflammatory-metabolic environment of the
   maternal-fetal interface. Due to the close relationship between PPH and
   labor process, this study aimed to make early predictions, therefore
   samples from the early stages of delivery were required. This is the
   reason why we chose cervical dilation < 3 cm as research object.
   Compared to blood samples, extracting secretions is non-invasive,
   making it easier to achieve clinical transformation. In addition, we
   selected samples during labor because they could effectively reflect
   the state of the maternal-fetal interface, including the
   immune–inflammatory system activated during labor initiation. However,
   this study has several limitations. First, the sample size of this
   study was small and limited to a single center. Further experiments are
   required to expand the sample size for multicenter joint studies.
   Second, follow-up research is needed to conduct in-depth verification
   of critical molecule expression and essential pathway functions.

Conclusions

   Histidine, DPYD, CD163, and FGL2 in CVF were associated with atonic
   PPH. The predictive models based on above differential expressed
   metabolite and proteins had an excellent predictive efficiency. These
   biomarkers could be utilized to identify patients at risk of PPH and
   enable prompt action to prevent PPH-related maternal mortality.

Materials and methods

Study design

   This study was a nested case-control study based on a prospective
   cohort. To obtain sufficient PPH samples and matched control group
   samples, we carefully designed a prospective cohort consisting of 958
   pregnant women with complete clinical data on pregnancy outcomes at
   Peking University Third Hospital from March 28, 2022, to July 29, 2022,
   118 of whom were diagnosed with PPH. The cohort was divided into two
   parts, including the discovery cohort and the confirmation cohort.
   First, a discovery cohort was used for proteomic and metabolomic
   detection to screen for significantly differentially expressed proteins
   and metabolites between the PPH group and non-PPH group. Then, a
   confirmation cohort was used to verify the differential expression
   stability of significantly changing molecules. Finally, proteins and
   metabolites identified in the discovery cohort and stably
   differentially expressed in the confirmation cohort were included to
   construct a predictive model.

   We included pregnant women who met the following inclusion criteria:
   single gestational women aged over 18 years old who were full-term and
   had a vaginal delivery, and whose cumulative bleeding volume measured
   by the weighing method exceeded 500 ml within 24 h after vaginal
   delivery. The following exclusion criteria were used: refusal to
   participate, PPH caused by reasons other than atony (placental accreta,
   soft birth canal lacerations, underlying coagulation disorders, etc.),
   autoimmune diseases, preeclampsia, reproductive tract infection, fetal
   appendage infection, suffering from gestational diabetes mellitus (GDM)
   requiring drug control, multiple pregnancies, macrosomia, and
   precipitous deliveries. Based on the above criteria, we ultimately
   included 27 pregnant women with PPH. The control (non-PPH) group was
   matched with baseline information (including maternal age, gestational
   age, and pregnancy complications, including hypertensive disorders in
   pregnancy (HDP) and GDM) without PPH. Each pregnant woman was enrolled
   from early pregnancy, followed up after labor. All procedures were
   approved by the Medical Science Research Ethics Committee of the Peking
   University Third Hospital (M2021685).

Collection of cervicovaginal fluid (CVF) collection

   CVF samples were collected from pregnant women using a vaginal swab
   when the cervical dilation was < 3 cm during labor. Samples were
   centrifuged at 4 °C 1650 rcf (g) in physiological saline for 10 min,
   and the supernatant was extracted and stored at -80 °C. Laboratory
   personnel were blinded to the clinical diagnosis of samples.

Omics detection

Proteomic and metabonomic sample processing and analysis

   Proteomic samples were lysed, and proteins were extracted using SDT
   buffer (4% SDS, 100 mM Tris-HCl, pH 7.6). The amount of protein was
   quantified using the BCA Protein Assay Kit (Bio-Rad, USA). The digested
   peptides were desalted, concentrated, and reconstituted in 40 µl of
   0.1% (v/v) formic acid. LC-MS/MS analysis was performed on a timeTOF
   Pro mass spectrometer (Bruker) coupled to a nanoelute (Bruker) and
   operated in the positive ion mode. The precursors and fragments were
   analyzed using a TOF detector (m/z 100–1700). The timsTOF Pro was
   operated in parallel accumulation serial fragmentation (PASEF) mode.
   MaxQuant 1.6.14 software combined and searched the raw MS data for
   identification and quantitation analysis.

   After metabonomic samples were vortexed, incubated, and dried, and were
   suspended in 100µL acetonitrile solution (1:1, v/v) and transferred to
   LC vials. A quadrupole time-of-flight mass spectrometer (Sciex
   TripleTOF 6600) was used to analyze the samples. Solvents A (25 mM
   ammonium acetate and 25 mM ammonium hydroxide in water) and B
   (acetonitrile) were used for LC separation. The mass spectrometer was
   operated in the negative ion and positive ionization modes. The
   instrument was set to collect over the m/z ranges of 60-1000 Da and
   25-1000 Da, respectively, in MS and auto-MS/MS acquisition. The
   information-dependent acquisition was used to acquire product ion
   scans. The original data were converted into MzXML format using
   ProteoWizard to perform peak pairing, retention time correction, and
   peak area extraction. The metabolites were identified by MS/MS using an
   in-house database established using authentic standards. An unpaired
   Student’s t-test was used (variable importance in projection (VIP)
   value > 1, P < 0.05) for statistical analyses.

Metabolomic and proteomic data analysis

   Cluster 3.0 and Java Treeview software performed the hierarchical
   clustering analysis. The NCBI BLAST + client software, InterProScan was
   used in protein analysis. Gene ontology (GO) terms were mapped, and
   sequences were annotated using the software program Blast2GO. The GO
   annotation results were plotted using R scripts. The Kyoto Encyclopedia
   of Genes and Genomes (KEGG) database ([139]http://geneontology.org/)
   was used to identify the KEGG pathways. Enrichment analysis was
   performed using Fisher’s exact test. The Benjamini-Hochberg correction
   was applied to adjust the derived P values (P < 0.05).

Integrative analysis

   Protein-protein interaction (PPI) information was retrieved from the
   IntAct database ([140]http://www.ebi.ac.uk/intact/). The results were
   imported into Cytoscape software ([141]http://www.cytoscape.org/,
   version 3.2.1) in XGMML format to derive a functional protein-protein
   interaction network and evaluate the importance of proteins.

Verification of biomarkers discovered in omics

   The mass spectrum was used for metabolite validation. 100µL of CVF was
   mixed with 300µL methanol, and 310µL of the supernatant was transferred
   and supplemented with 50µL 85% methanol. Samples were then centrifuged
   for 5 min (15000 g, 4 °C). The supernatant was transferred for LC-MS
   analysis, and UHPLC-QqQ MS was used for quantitative analysis of
   metabolites. The concentration of each analyte in the standard curve
   was plotted on the horizontal axis, and the peak area ratio of the
   analyte to the internal standard was plotted on the vertical axis.
   Linear regression was performed using the weighted (1/X^2) least
   squares method, and the concentration of the analyte in each sample was
   calculated based on the peak area ratio of the analyte to the internal
   standard.

   Enzyme-linked immunosorbent assay (ELISA) was used for protein
   validation. DPYD (ELK Biotechnology, ELK2382), Scavenger receptor
   cysteine-rich type 1 protein M130 (CD163) (ELK Biotechnology, ELK5116),
   Fibroleukin (FGL2) (ELK Biotechnology, ELK1956), Mitochondrial fission
   1 protein (FIS1) (ELK Biotechnology, ELK4182), Glutathione peroxidase 3
   (GPX3) (ELK Biotechnology, ELK3444), and Adenosine deaminase (ADA) (ELK
   Biotechnology, ELK1970) ELISA kits were used to determine if these
   proteins were differentially expressed in accordance with the
   manufacturers’ instructions.

Construction of predictive models

   Firstly, histidine, DPYD, CD163, and FGL2 that were identified by omics
   and had significant differences in the validation cohort were included
   in the construction of the prediction model. Secondly, we conducted a
   correlation analysis between histidine, DPYD, CD163, and FGL2 and blood
   loss by using Spearman’s rank correlation coefficient. Then, using the
   above biomarkers, constructed a multi factor analysis for predicting
   PPH. Multivariate logistic regressions were performed for significantly
   different proteins and metabolites; the curve of ROC was drawn, and the
   area under the curve (AUC) was computed. We have constructed
   independent prediction models for histidine, DPYD, CD163, and FGL2, as
   well as a joint prediction model for histidine and DPYD, a joint
   prediction model for histidine and CD163, a joint prediction model for
   histidine, DPYD and CD163, and a joint prediction model for histidine,
   DPYD, CD163, and FGL2. In these models, indicators such as maternal age
   and gestational age were not included, firstly because these indicators
   were not related to blood loss, and secondly, to make the model more
   concise. [142]https://www.metaboanalyst.ca/ was used to plot OrthoPLSDA
   map and histograms for metabolite and protein validation and ROC
   curves.

Statistical analysis

   Continuous variables were tested for normality. Maternal age,
   gestational age, and BMI were described as mean ± standard, while
   estimated blood loss and fetal birthweight were described as median
   (25th to 75th percentile). Differences between groups were assessed for
   significance using Student’s t-test (normality test) or Mann-Whitney U
   (non-normality test) for continuous variables. Categorical variables
   are presented as numbers (percentages) and were compared using the
   chi-squared test. The method of multivariate logistic regressions can
   be found in Sect. 2.5. Statistical analyses were performed using SPSS
   (version 26.0; SPSS, Inc., Chicago, IL). Statistical significance was
   set at P < 0.05.

Acknowledgements