A First-Trimester Serum Proteomic Signature for Early Prediction of Preeclampsia: Integrated Untargeted and Targeted Mass Spectrometry with Machine Learning

1. Introduction

Preeclampsia (PE) is a leading cause of maternal and perinatal morbidity and mortality globally, particularly when it develops early and leads to indicated preterm delivery [1]. Consequently, the reliable first-trimester identification of high-risk women is a paramount objective in modern obstetrics. This remains challenging, especially for nulliparous patients, in whom traditional risk-factor-based screening performs suboptimally [2,3]. Current guidelines rely primarily on maternal demographic and historical factors, an approach that limits detection rates and leaves a substantial proportion of cases undetected [3,4,5,6,7,8,9,10].

While more sophisticated first-trimester algorithms, such as the Fetal Medicine Foundation (FMF) model—which integrates maternal factors with uterine artery Doppler and biochemical markers—can detect 75–90% of early or preterm preeclampsia at a 10–16% false-positive rate, their performance is imperfect and varies across populations and settings [3,4,11,12,13]. Furthermore, the implementation of these multivariable tools requires standardized measurements, specialized software, and trained personnel [7,14,15]. Critically, they may not fully capture the complex systemic pathophysiological shifts that precede clinical disease [16,17]. Maternal serum offers a compelling matrix for biomarker discovery due to its minimally invasive collection and its reflection of systemic processes underpinning preeclampsia, including placental dysfunction, endothelial activation, and inflammation [16].

Modern mass-spectrometry-based proteomics enables the specific, reproducible, and multiplexed quantification of hundreds of serum proteins, often surpassing conventional immunoassays in analytical performance [16,18,19,20,21,22]. Data-independent acquisition (DIA) provides unbiased, broad-coverage discovery of the serum proteome, while targeted multiple reaction monitoring (MRM) mass spectrometry permits precise validation of candidate proteins across cohorts, establishing a powerful complementary pipeline [16,19,20,21,22,23,24]. The integration of these advanced proteomic approaches in early pregnancy holds promise for uncovering molecular signatures of subsequent preeclampsia, refining individual risk stratification beyond existing algorithms, and informing more effective preventive strategies [17,25,26].

This study employs an integrated DIA-PASEF-MS and targeted MRM-MS strategy on first-trimester serum samples from women who later developed preeclampsia and matched controls. The objective is to identify and validate robust protein biomarkers and to construct predictive models with the potential to augment current FMF-based screening in routine clinical practice.

2. Materials and Methods

2.1. Study Design

This prospective cohort study was conducted at the I. Kulakov National Medical Research Center for Obstetrics, Gynecology, and Perinatology (Moscow, Russia) from January to December 2022. The analytical cohort was selected from an initial screening population of 1869 women aged 18–45 years who underwent routine first-trimester FMF-based screening between 11⁺² and 14⁺² weeks of gestation. The final cohort comprised 64 women with singleton pregnancies, including 32 who subsequently developed PE and 32 age-matched controls with uncomplicated pregnancies. Exclusion criteria were pre-existing diabetes mellitus, autoimmune disease, a history of solid-organ transplantation, known malignancy, or fetal chromosomal abnormalities. Preeclampsia was defined according to contemporary guidelines as de novo hypertension after 20 weeks of gestation accompanied by proteinuria, maternal organ dysfunction (renal, hepatic, hematologic, or neurologic), or uteroplacental dysfunction including fetal growth restriction [14]. The control group consisted of age-matched (p>0.05) women with uncomplicated singleton pregnancies who conceived spontaneously and did not develop any hypertensive disorder, gestational diabetes, preterm delivery before 37 weeks, or intrauterine growth restriction (estimated fetal weight <10th percentile), and did not have high FMF-based first trimester risk of PE.

Written informed consent was obtained from all participants at enrollment. A standardized first-trimester assessment included measurement of maternal weight, height, and blood pressure [14], along with collection of detailed obstetric and medical history. Ultrasonographic evaluation involved transabdominal color Doppler assessment of the uterine arteries with calculation of the pulsatility index (UtA-PI) [15]. Venous blood was collected into Serum Z/9 tubes (Monovette, Sarstedt, Germany) and processed within 2 hours. After centrifugation at 300×g for 20 minutes at room temperature, the serum supernatant was aliquoted and stored. Serum concentrations of placental growth factor (PlGF) and pregnancy-associated plasma protein-A (PAPP-A) were measured using the Delfia Xpress system (PerkinElmer, USA) per the manufacturer’s instructions. Residual serum was aliquoted and stored at −80°C for subsequent untargeted DIA-MS and targeted MRM-MS proteomic analyses of first-trimester biomarkers [23]. The study protocol was approved by the Institutional Review Board of the I. Kulakov Center (protocol No. 2, 9 March 2017). All procedures followed the principles of the Declaration of Helsinki and Good Clinical Practice guidelines.

2.2. Serum Preparation

Serum samples and a surrogate matrix of bovine serum albumin (BSA, 10 mg/mL in PBS) were subjected to tryptic digestion following standard procedures [21,23,24,27]. Briefly, 10 µL of serum were diluted in denaturation buffer containing 7.2 M urea, 16 mM dithiothreitol, and 240 mM Tris-HCl (pH 8.0) and incubated for 30 minutes at 37°C. Al-kylation was performed with iodoacetamide at a final concentration of 40 mM for 30 minutes at room temperature in the dark. Trypsin (Trypsin Gold, Promega, USA) was added at a 1:25 enzyme-to-protein ratio, and digestion proceeded overnight at 37°C. Total protein concentration was determined using a bicinchoninic acid (BCA) assay kit (Thermo Fisher Scientific, USA). Digestion was terminated by acidification with 1.0% formic acid to pH ≤ 2. The resulting peptide mixture was adjusted to 1 µg/µL, kept on ice, and analyzed by mass spectrometry on the same day.

For targeted MRM-MS analysis of 139 serum proteins, an in-house panel based on the BAK125/270 MRM kit (Canada) was employed. This panel, previously developed by Prof. Borchers’ group, enables multiplexed quantification of potential plasma protein bi-omarkers across various diseases including neurodegenerative, autoimmune, cardiovas-cular, and renal disorders, as well as diabetes [21,22,23]. Stable isotope-labeled internal standard (SIS) peptides and corresponding native (NAT) peptides were synthesized and characterized at the Omics Laboratory of Skolkovo Institute of Science and Technology (Skoltech) [21,22]. The SIS peptide panel featured isotopically labeled C-terminal lysine (+8 Da) or arginine (+10 Da) residues.

The SIS peptide mixture was dissolved and diluted to a working solution at 10× the lower limit of quantification (LLOQ). For each sample, 40 µL of serum digest were com-bined with 10 µL of the SIS peptide mixture. Peptide purification was performed via sol-id-phase extraction using Oasis HLB 96-well plates. After conditioning and sample load-ing, wells were washed and peptides were eluted with 70% acetonitrile/0.1% formic acid. Eluates were dried in a vacuum concentrator and stored at −80°C until analysis.

2.3. Targeted HPLC-MS Analysis

For HPLC-MRM-MS analysis, dried peptide samples were reconstituted in 0.1% for-mic acid to a final concentration of 1 µg/µL. Aliquots of 10 µL from each reconstituted se-rum digest, quality-control (QC) sample, and calibration standard were injected onto a Zorbax Eclipse Plus reversed-phase UHPLC column (Agilent Technologies, USA) con-nected to an ExionLC™ system (Thermo Fisher Scientific, USA). Peptides were separated at a flow rate of 0.4 mL/min over a 60-minute multistep gradient [23].

For calibration standards and QC samples, 40 µL of surrogate BSA digest were mixed with 10 µL of SIS mixture and 10 µL of native peptide mixture. A lyophilized mixture of native peptides, pre-balanced at the LLOQ for each analyte, was dissolved and serially diluted to generate eight calibration levels: 100×, 40×, 16×, 4×, 2×, 0.5×, 0.25× and 0.1× LLOQ. QC samples at three concentration levels (0.35× LLOQ, QC-A; 3.5× LLOQ, QC-B; and 35× LLOQ, QC-C) were analyzed in triplicate.

MRM data were processed using Skyline software [28]. Quantification adhered to ICH guidelines for Bioanalytical Method Validation [29], employing weighted (1/x²) linear regression of SIS-to-native peak area ratios. Calibration curve performance was assessed in Skyline, with accuracy considered acceptable when measured concentrations for at least 6 of 9 calibration points and at least 1 of 3 QC replicates at each level were within ±20% of nominal values.

2.4. Non-Targeted DIA-PASEF-MS Analysis

Tryptic peptide fractions were analyzed on a Dionex Ultimate 3000 nano-LC system (Thermo Fisher Scientific, USA) coupled with a timsTOF Pro mass spectrometer (Bruker Daltonics, Billerica, MA, USA). 1 µL of each peptide sample was injected onto a packed emitter C18 column (25 cm × 75 µm, 1.6 µm; Ion Optics, Parkville, Australia). LC separa-tion was carried out at a flow rate of 400 nl/min using a 90-minute linear gradient from 2% to 37% solvent B (0.1% FA in ACN), followed by a column wash step (10-minute iso-cratic elution with 90% solvent B) and equilibration (15 minutes, isocratic elution with 2% solvent B).

MS data were acquired using the DIA-PASEF method. The electrospray ionization (ESI) source parameters were set as follows: capillary voltage 1400 V, dry gas flow — 3.0 l/min at 180 °C. MS and MS/MS spectra were recorded in the range of 100 to 1700 m/z and in the ion mobility range from 0.6 to 1.6 V*s/cm². The scan time (ramp) was set to 100 ms. Collision energy changed linearly depending on mobility: from 59 eV at 1/K0 = 1.6 V·s/cm² to 20 eV at 1/K0 = 0.6 V·s/cm².

The LC-MS/MS spectra were analyzed using the DIA NN software (Data-Independent Acquisition by Neural Networks, version 1.8.1) in library-free mode with the following parameters: mass accuracy – 20 ppm; MS1 accuracy – 20 ppm; peptide length range – from 7 to 30 amino acids [7,8,9]. The search was performed against the SwissProt Human database with carbamidomethylation (C) and oxidation (M) specified as variable modifi-cations. The false discovery rate threshold was set at 0.1%. Data filtering and calculation of LFQ values were performed using the R package DIAgui [30].

2.5. Statistical Analysis

Continuous clinical variables were summarized as median (first quartile; third quartile) and categorical variables as counts (percentage). Between-group differences were assessed using the Mann–Whitney U test for continuous variables and Pearson’s chi-square test for categorical variables, with statistical significance defined as p < 0.05.

For untargeted DIA-MS data, only proteins detected in ≥70% of samples were analyzed. Missing values were imputed in Perseus (MaxQuant environment) using a normal distribution-based procedure [31].

For targeted MRM-MS data to correct for batch effects and analytical variability, data normalization was performed using RobNorm methods [32]. For each protein, the mean CV was calculated across standard and quality control samples between batches; proteins with a CV exceeding 50% were excluded from further analysis.

Significant proteins from both targeted and untargeted datasets were identified using the Mann–Whitney U test with Benjamini–Hochberg correction (false discovery rate (FDR)-adjusted p-value < 0.01). Orthogonal projections to latent structures discriminant analysis (OPLS-DA) was performed on each proteomic profile, and features with variable importance in projection (VIP) > 1.5 and FDR < 0.01 were retained as protein markers [33]. For DIA-MS data, additional machine learning classifiers were trained, including support vector machines (SVM) with linear, polynomial, radial and sigmoid kernels, as well as random forest models [34,35]. Hyperparameters of SVM with polynomial, radial and sigmoid kernels were optimized using particle swarm optimization (PSO) [36]. Model performance was evaluated by 10-fold cross-validation, calculating sensitivity, specificity, accuracy, area under the receiver-operating characteristic curve (AUC), positive predictive value (PPV), negative predictive value (NPV) and F-score.

Pathway enrichment analysis was conducted using STRING (FDR < 0.01) [37]. Associations between DIA-MS and MRM-MS protein levels were assessed using Pearson correlation, while correlations between DIA-MS protein levels and clinical parameters were evaluated using Spearman’s rank correlation.

Statistical analysis was performed by scripts based on R version 4.3.2 [38] the RStudio environment (version 2023.09.1) [39]. A comprehensive suite of R packages was employed to support both statistical modeling and data visualization. For modeling and machine learning tasks, the packages ropls 1.34.0 [40], effsize 0.8.1 [41], pwr 1.3-0 [42], e1071 1.7-16 [43], caret 7.0-1[44], dplyr 1.1.4 [45], RandomForest [35] were utilized. Visualization of the results was accomplished using ggplot2 3.5.2 [46], reshape2 1.4.4 [47], forcats 1.0.0 [48], ggrepel 0.9.6 [49], pheatmap 1.013 [50] and pROC 1.18.5[51].

3. Results

3.1. Clinical Characteristics

The final cohort comprised 32 women who subsequently developed PE and 32 age-matched controls with uncomplicated pregnancies. The clinical characteristics of both groups are summarized in Table 1.

Maternal age, body mass index (BMI), gestational age at sample collection, mean arterial pressure (MAP) multiples of median (MoM), and the use of assisted reproductive technologies (IVF) did not differ significantly between the groups (all p > 0.05). However, women who later developed PE were more frequently nulliparous (72% vs. 34%, p < 0.001), had a higher prevalence of previous PE (22% vs. 0%, p = 0.02), habitual miscarriage (38% vs. 0%, p < 0.001), and previous preterm delivery (25% vs. 0%, p = 0.01) compared to controls.

First-trimester FMF-based screening identified a high risk for PE in 53% of women who subsequently developed the disease, whereas none of the controls were classified as high-risk (p < 0.001). First-trimester placental growth factor (PlGF) levels (MoM) were significantly lower in the PE group compared to controls (median 0.55 vs. 0.74, p = 0.002).

Regarding pregnancy outcomes, women in the PE group delivered earlier than controls (median 37.2 vs. 39.4 weeks, p < 0.001) and had higher maximum systolic and diastolic blood pressures (135 vs. 115 mmHg and 89 vs. 70 mmHg, respectively, both p < 0.001). The PE group also demonstrated higher umbilical artery pulsatility index (0.95 vs. 0.79, p = 0.002) and lower cerebro-placental ratio (1.44 vs. 1.89, p < 0.001).

Markers of disease severity were significantly elevated in the PE group, including 24-hour proteinuria (1.1 vs. 0 g/L, p < 0.001) and serum creatinine (80.8 vs. 66.6 µM/L, p < 0.001). Intrapartum blood loss was higher in the PE group (700 vs. 300 ml, p < 0.001), and emergency cesarean section was more frequent (59% vs. 3%, p < 0.001). Neonatal outcomes were less favorable in the PE group, with lower birthweight (2745 vs. 3400 g, p < 0.001) and lower Apgar scores at both 1 minute (p = 0.002) and 5 minutes (p = 0.001) compared to controls.

3.2. Non-Targeted DIA-MS Proteome Profile

Non-targeted high-performance liquid chromatography HPLC-DIA-MS analysis of first-trimester serum identified over 455 protein groups. Statistical analysis was per-formed on 274 proteins consistently detected in at least 70% of samples (Table S1). Principal component analysis (PCA) revealed clear separation between samples from women who later developed PE and controls. This separation was further supported by an OPLS-DA model demonstrating high explanatory and predictive capability (R2Y = 0.96, Q2Y = 0.78) (Figure 1a,b).

Differential abundance analysis identified 69 proteins with significant unadjusted p-values (p<0.01), of which 48 remained significant following FDR correction (FDR<0.01). From this set, 33 proteins with a VIP score exceeding 1.5 were designated as candidate PE markers (Figure 1c, Table S1). The majority of these proteins were down-regulated in serum from women who later developed PE (CFHR4, PRSS1, IGHV1-46, C1QB, KRT1, NOTUM, AHSG, GPX3, LBP, PRG2, SAA4, ORM2, FBLN1, SERPINC1, SERPINA7, C1S, C9, SERPINA3, C3, TF, GC, ALB, APOA1 and SERPINA1), while a smaller subset was up-regulated (CFD, CPB2, BCHE, IL1RAP, LTBP1, PROC, INHBC, SELENOP and VNN1).

When these 33 DIA-derived markers were used as input features for classification models, SVM with linear and polynomial kernels and a random forest classifiers achieved the highest quality metrics (PPV and NPV of 97% and 95%, respectively) during 10-fold cross-validation (Table 2, Figure 1d). Pathway enrichment analysis demonstrated that this protein set was strongly associated with complement activation pathways and also mapped to biological processes including post-translational protein phosphorylation, regulation of insulin-like growth factor (IGF) transport and uptake, platelet degranulation, and selenium-related micronutrient networks (Figure 2, Table S2).

3.3. Targeted MRM-MS Validation

Targeted HPLC-MRM-MS quantification was successfully performed for 104 plasma proteins across all serum samples (n=64) (Table S4). Unsupervised PCA again resolved two distinct clusters corresponding to future clinical outcomes. A subsequent OPLS-DA model based on the targeted panel demonstrated robust discriminatory power (R2Y = 0.94, Q2Y = 0.74) (Figure 3a, b).

Among the quantified proteins, 12 exhibited significant unadjusted p-values (p < 0.01), with 9 proteins remaining significant after FDR correction (FDR<0.01) and possessing high VIP scores (VIP>1.5) in the OPLS-DA model (Table S3, Figure 3c). Notably, four of these nine validated proteins (AFM, AHSG, C8A, IGHG1) overlapped with candidates identified in the discovery phase, providing cross-platform confirmation.

Correlation analysis between DIA-MS and MRM-MS measurements demonstrated significant concordance for the majority of overlapping proteins: 71% showed Pearson correlation coefficients greater than 0.5 with p < 0.001, indicating good reproducibility be-tween the semi-quantitative and quantitative approaches (Table S4). A correlation heatmap highlighted four coherent clusters of markers across platforms, including a tight subcluster comprising AHSG, C8A and F10 from the MRM panel and AHSG, C1QB and PRSS1 from the DIA panel, all with correlation coefficients above 0.6. AHSG levels measured by DIA-MS and MRM-MS were particularly consistent, with a correlation coefficient of 0.82 (p < 0.001), supporting its role as a robust early marker candidate (Figure 4a).

3.4. Associations Between Protein Markers and Clinical Parameters

Integration of DIA-MS protein markers with clinical data revealed two main clusters of clinical variables on the correlation heatmap (Figure 4b). Gestational age at delivery and birthweight grouped together and showed correlation patterns opposite to those of diastolic and systolic blood pressure before delivery, duration of post-partum hospitalization and proteinuria level, which formed a second cluster.

Several candidate protein markers (GPX3, PRSS1, TF, AHSG, C3, KRT1, GC, C1QB and IGHV1-46) were inversely associated with clinical parameters indicative of disease severity, including umbilical artery pulsatility indices, intrapartum blood loss, and pre-delivery creatinine levels. Conversely, these same proteins demonstrated positive correlations with factors such as parity, first-trimester placental growth factor levels, Apgar scores and the cerebral-placental ratio before delivery. The strongest individual associations were observed for PRSS1, which correlated negatively with proteinuria (r = −0.74), and for IGHV1-46, which correlated positively with gestational age at delivery (r = 0.72), linking the proteomic signature to clinically meaningful manifestations of PE.

4. Discussion

The FMF algorithm detects 80.6% (95% CI 64.0–91.8) of preterm PE cases (<37 weeks) and 31.8% (95% CI 18.6–47.6) of term PE cases (≥37 weeks) and has a long-standing clinical track record [52]. Nevertheless, false-negative results persist in real-world practice. Some women without pronounced anamnestic risk factors who are classified as low-risk subsequently develop severe PE, leading to adverse outcomes [3,4,12,13,52,53,54]. This may be attributed to operator dependence in UtA-PI measurement, insufficient sphygmomanometer quality control, and lack of regular calibration. In the present cohort, the FMF model identified high risk in only 53% of women who later developed PE, underscoring the potential added value of a proteomic approach.

In contrast, integrated untargeted DIA-MS and targeted MRM-MS profiling (n = 64) of first-trimester maternal serum reliably identified women at risk for subsequent PE. In a discovery study, the analysis identified 33 DIA-MS markers (FDR < 0.01, VIP > 1.5) and achieved excellent predictive performance (AUC 0.95, sensitivity 95%, specificity 97%) using SVM and random forest classifiers. Targeted MRM-MS validated nine proteins, of which four — AFM, AHSG, C8A, and IGHG1 — confirmed the discovery findings. High cross-platform concordance (71% of overlapping proteins with r > 0.5, p < 0.001) underscores technical reproducibility. These findings, together with previous first-trimester proteomic studies [16,24,55,56] demonstrate the ability of this approach to detect early molecular changes preceding clinical disease. This capability is particularly important given the limited effectiveness of existing prevention strategies: low-dose aspirin reduces PE risk by only 18–48% [53,54]. This modest efficacy likely reflects biological heterogeneity, encompassing distinct molecular subclasses such as placental, metabolic, maternal anti-fetal, and extracellular matrix-related subtypes [17,57,58].

The biological coherence of the discovered markers strengthens their clinical plausibility. Pathway enrichment analysis prominently featured complement cascade activation, a central mechanism in PE pathogenesis [17,57,59,60]. Altered concentrations of complement-related proteins, including C1QB, C3, CFHR4, and other components of the classical pathway, may reflect their consumption due to early complement activation. Several complement-related proteins, including C1QB, C1S, C3, C8A, and C9, were significantly altered in women who later developed PE, consistent with a chronic low-grade inflammatory and complement-dysregulated state established in the first trimester [16,59,60,61,62]. Interestingly, a prospective study by He YD et al. (2020) demonstrated that by the second and third trimesters, blood differ from those in women with uncomplicated pregnancies [59]. Furthermore, Matsuyama T. et al. (2021) showed in vitro that an imbalance of pro-angiogenic and anti-angiogenic factors—observed as early as the first trimester in pregnancies subsequently complicated by PE—inhibits the synthesis of complement factor H by placental endothelial cells, leading to complement activation and endothelial dysfunction [60].

Similarly, several liver-derived and transport proteins—such as albumin (ALB), alpha-1-antitrypsin (SERPINA1), alpha-1-antichymotrypsin (SERPINA3), serotransferrin (TF), and vitamin D-binding protein (GC)—may reflect early changes in hepatic synthetic function, systemic inflammation, or redistribution processes previously associated with adverse pregnancy outcomes [16,24]. Reduced first-trimester ALB levels have demonstrated moderate prognostic value for PE; however, its diagnostic performance improves substantially when incorporated into combined predictive models [16,24]. SERPINA1 has been proposed as a potential component of proteomic panels for early PE prediction, demonstrating high diagnostic accuracy within multi-protein models [24]. Experimental evidence indicates its involvement in regulating trophoblast invasion through endoplasmic reticulum stress pathways [63]. Elevated levels of non-tryptic SERPINA1 peptides in urine are associated with the clinical course of preeclampsia and may represent a promising non-invasive marker of disease severity [61,64,65]. By contrast, SERPINA3—belonging to the same family of serine protease inhibitors—has not shown a significant association between genetic variants and PE risk [66], and its role in disease pathogenesis remains incompletely understood.

Thus, early disturbances in hepatic synthetic and metabolic function may play an important role in the pathogenesis of preeclampsia as early as the first trimester. Elevated first-trimester hepatic steatosis index (HSI) values were recently associated with an increased risk of gestational hypertension and PE in a study by Zhang et al (2025) [67], confirming the role of early structural liver changes closely linked to metabolic and functional dysfunction. The down-regulation of AHSG—a liver-synthesized protein involved in metabolic regulation—observed in both DIA-MS and MRM-MS platforms (Pearson r = 0.82, p < 0.001) aligns with previous reports linking low AHSG to preterm PE and to metabolic dysfunction-associated steatotic liver disease (MASLD)[68], which itself associates with adverse pregnancy outcomes [69,70]. A longitudinal study by Chaemsaithong P. et al. (2015) found that AHSG levels increased from the first trimester to 26 weeks of gestation and then declined, being significantly lower in the group that developed early-onset PE [71]. However, the direction of AHSG changes varies across studies: both elevated levels in manifest PE and reduced levels in early-onset disease and longitudinal observations have been reported, reflecting disease heterogeneity, differences in gestational timing, and population characteristics [71,72] Reduced levels of several proteins, including AHSG, may be attributable to multiple mechanisms: redistribution and local accumulation in placental tissue (as shown for SERPINA1) [61,73], consumption under conditions of immune cascade activation [16,24,62], as well as differences in disease stage and phenotypic heterogeneity [17,57].

Accumulating evidence points to a possible role of hemostasis imbalance in the development of PE. Pei-Pei Jin et al. (2023) conducted a prospective study consistent with the present findings, evaluating first-trimester antithrombin III (SERPINC1) levels in 853 pregnant women (322 who subsequently developed PE and 531 controls). That study demonstrated that reduced SERPINC1 levels were significantly associated with disease development after adjustment for age and BMI [74]. The role of other coagulation proteins, such as PROC and CPB2, in early PE prediction remains limited; nevertheless, hemostatic disturbances confirm their involvement in disease pathogenesis [62]. Furthermore, the down-regulation of GPX3, SERPINC1, and SERPINA1 points to impaired antioxidant capacity and coagulation regulation, both implicated in PE progression [63,66,72,73,74].

Proteins involved in extracellular matrix (ECM) organization and vascular remodeling, including fibulin-1 (FBLN1), may reflect impaired vascular adaptation and defective placentation. This aligns with evidence implicating ECM dysregulation in PE pathogenesis [75]. First-trimester proteomic studies have demonstrated the potential of ECM-related proteins as early disease biomarkers, and experimental data indicate that alterations in collagen composition may disrupt trophoblast function and contribute to defective placentation [24,75].

An important translational insight emerges from the correlation analysis between protein markers and clinical parameters (Figure 4b). PRSS1 showed a strong negative correlation with proteinuria (r = −0.74), and IGHV1-46 correlated positively with gestational age at delivery (r = 0.72), directly linking the proteomic signature to clinically meaningful endophenotypes of disease severity. Moreover, the inverse associations of PRSS1, IGHV1-46, AHSG, GPX3, C1QB and C3 with umbilical artery pulsatility indices and pre-delivery creatinine levels suggest that these markers capture not only the risk of developing PE but also the anticipated trajectory of placental and maternal organ dysfunction. This dual prognostic capacity—predicting both the occurrence and the likely severity of PE—represents an advance over many existing single-analyte or purely demographic models [3,60,72].

From a methodological perspective, the combination of DIA-PASEF-MS for broad discovery and MRM-MS for multiplexed, absolute quantification proved highly synergistic. The DIA-MS platform enabled unbiased interrogation of >450 protein groups in blood serum, while the targeted MRM panel—originally designed for other disease contexts but applied here to PE—permitted rapid, quantitative validation of candidate markers without requiring de novo assay development [19,21,22,23,24,25,27]. The high cross-platform correlation for more than 70% of overlapping proteins indicates that semi-quantitative DIA-MS data can be reliably used for initial screening, with MRM-MS serving as a confirmatory or clinical-grade quantification method. Of particular interest is the identification of coherent protein clusters demonstrating similar patterns of change across both methods. Specifically, a tightly connected subcluster was identified, including AHSG, C8A, and F10 (by MRM-MS) as well as AHSG, C1QB, and PRSS1 (by DIA-MS), all of which exhibited high correlation coefficients. This indicates coordinated changes in these proteins and likely reflects their involvement in common pathophysiological processes, such as complement activation, inflammatory response, and hemostatic disturbances. The highest cross-platform concordance was observed for AHSG, confirming its robustness to analytical variability and positioning it as one of the most reliable candidate early biomarkers of preeclampsia.

Several limitations warrant consideration. First, the sample size (n = 64), while adequately powered for discovery and for training machine learning models with cross-validation, requires external validation in larger, multi-center cohorts before clinical implementation. Second, the study was conducted at a single tertiary referral center with a predominantly nulliparous and Caucasian population; generalizability to other ethnic groups, multiparous women, and lower-resource settings remains to be established. Third, the MRM panel, though multiplexed, was not specifically optimized for PE; a purpose-built PE panel might yield even higher performance. Fourth, proteomic analysis was performed at a single time point in the first trimester, which does not allow assessment of the dynamics of molecular changes throughout pregnancy. Finally, while association with clinical severity parameters was assessed, the study was not designed to distinguish early- vs. late-onset PE or preterm vs. term PE; these molecular subtypes may have distinct proteomic profiles [16,17,57,58].

Despite these limitations, the present findings align with and extend a growing body of literature implicating complement, coagulation, and metabolic regulation pathways in early PE pathogenesis [16,17,55]. The four cross-platform validated markers (AFM, AHSG, C8A, IGHG1) represent particularly robust candidates for further development. Importantly, the ability to predict PE from first-trimester serum using quantitative MS—without requiring specialized ultrasonographic equipment or operator training—could democratize risk assessment, particularly in settings where Doppler expertise is limited. The high NPV (95%) of the SVM and random forest models also suggests that such an approach could safely reassure a large proportion of nulliparous women, reducing unnecessary surveillance and anxiety.

5. Conclusions

Integrated untargeted DIA-MS and targeted MRM-MS profiling of first-trimester maternal serum reliably identifies women at risk for subsequent preeclampsia. In a discovery-validation cohort (n = 64), the analysis identified 33 DIA-MS markers (FDR < 0.01, VIP > 1.5), validated four by MRM-MS (AFM, AHSG, C8A, IGHG1), and achieved excellent predictive performance (AUC 0.95, sensitivity 95%, specificity 97%) using SVM and random forest classifiers. High cross-platform concordance (71% of overlapping proteins with r > 0.5, p < 0.001) underscores technical reproducibility.

Pathway enrichment featured complement activation, IGF transport dysregulation, and platelet degranulation. PRSS1 correlated negatively with proteinuria (r = −0.74), and IGHV1-46 correlated positively with gestational age at delivery (r = 0.72), directly linking the proteomic signature to clinical severity.

The four cross-platform validated markers represent robust candidates for further development. The ability to predict PE from first-trimester serum using quantitative mass spectrometry—without specialized ultrasound equipment or operator training—could democratize risk assessment, particularly in settings where Doppler expertise is limited. The high negative predictive value (95%) of the SVM and random forest models suggests that such an approach could safely reassure a large proportion of nulliparous women, reducing unnecessary surveillance and anxiety.

Future directions include external validation in diverse populations, direct comparison with the FMF model, development of a clinically validated targeted panel, and investigation of whether the proteomic signature can guide aspirin prophylaxis. Integrating quantitative proteomics into first-trimester screening may improve early identification of women at risk for PE, enabling timely surveillance and preventive therapy.