Ethical declaration

Adherence to ethical standards in animal research was paramount in the execution of this study. Scrutiny of the research protocol was exhaustive, culminating in its approval by the Animal Ethics Committee of China Medical University (approval number CMUXN2023100). All procedures involving animals were meticulously carried out in compliance with the committee's regulations, ensuring the ethical treatment and welfare of the animals involved, and attempts were undertaken to reduce unease and pain.

Data acquisition and preprocessing

From the Gene Expression Omnibus (GEO) database at http://www.ncbi.nlm.nih.gov/geo/, the PE datasets GSE10588, GSE24129, GSE25906, and GSE73374 were retrieved, and the ArrayExpress database (https://www.ebi.ac.uk/biostudies/arrayexpress) served as the source for dataset E-TABM-682. Basic information on these datasets is provided in Table S1. Platform annotation files were used to convert probe IDs in the expression matrices to gene symbols. The mean expression level was computed when multiple probes were associated with identical gene symbols. To normalize expression data from different datasets, application of the ComBat function sourced from the sva toolkit facilitated the eradication of batch discrepancies, and the assessment of batch effect eradication relied on principal component analysis (PCA) (Du et al. 2021). The GSE75010 dataset was used as a validation set.

Identification of ferroptosis-related genes (DE-FRG)

The process of selecting the DE-FRG involved the acquisition of FRG data from the FerrD database, accessible at http://www.zhounan.org/ferrdb/. Subsequently, the wilcoxTest method was employed to analyze the distinct FRG expression patterns, drawing comparisons between healthy individuals and those diagnosed with PE. DE-FRG was characterized as those genes exhibiting a statistical significance threshold with a p-value less than 0.05 (Su et al. 2023).

Gene ontology (GO) and kyoto encyclopedia of genes and genomes (KEGG) enrichment analysis

The analysis of GO functional enrichment and KEGG enrichment was implemented through the "clusterProfiler" package in R (Wang et al. 2021). Regarding statistical significance, outcomes meeting the predefined threshold for significance were those with a p-value less than 0.05 after correction.

NMF clustering for PE subtyping

DE-FRG was used for NMF clustering analysis to categorize 67 PE samples into subtypes. To determine the best cluster quantity (k), we monitored the trend of the cophenetic correlation coefficient as it diminished, revealing the stage where the ideal cluster amount was achieved. Following the determination of the optimal number of clusters, a subsequent step involved a comprehensive analysis of the differences between these subtypes. PCA was employed for this purpose, facilitating a deeper exploration of the underlying distinctions within the PE samples (Wang et al. 2022a, b, c).

Single sample gene set enrichment analysis (ssGSEA)

Enrichment score calculations for immune-related features in PE individuals were determined through the implementation of ssGSEA methodology. This analysis was executed using the GSVA (Gene Set Variation Analysis) package, a dynamic device for appraising the enrichment of gene collections in distinct specimens (Wang et al. 2021).

Weighted gene co-expression network analysis (WGCNA)

The "WGCNA" package within R software was deployed for the identification of co-expression modules. Initially, the subsequent WGCNA examination targeted the top 25% of genes demonstrating the most extensive diversity. This step aimed to ensure the quality and precision of the acquired findings. The appropriate soft threshold β was determined using the "pickSoftThreshold" function, enabling the transformation of the adjacency matrix. Subsequently, the calculation of the topological overlap matrix (TOM) was performed, leading to the construction of a hierarchical clustering dendrogram. This dendrogram facilitated the classification of genes exhibiting similar expression trends into distinct modules, each comprising no fewer than 40 genes. A threshold of 0.25 was established as the height cutoff for merging potentially similar modules. Lastly, analysis of module eigengenes (ME) facilitated the summarization of each module's expression characteristics and the assessment of their correlation with phenotypic traits. This correlation analysis enabled the selection of the most relevant modules for further analysis (Lai et al. 2022a, b).

Building and validating machine learning models

The objective of this investigation was to detect DE-FRG using machine learning models. The R package "caret" was utilized to construct these models based on the intersection genes of PE and its subtypes. Specifically, utilized in the project were four machine learning methodologies, namely Random Forest (RF), Support Vector Machine (SVM), Generalized Linear Model (GLM), and Extreme Gradient Boosting (XGB). To generate residual distribution results and identify essential features, the R package "DALEX" was employed to explain these models. Moreover, the "pROC" package in R was executed for the purpose of plotting the Receiver Operating Characteristic (ROC) curve and calculating the Area Under the Curve (AUC). Higher AUC values indicated higher accuracy in prediction. Ranking of the top 5 genes in each model was determined by the root mean square error (RMSE), suggesting their importance as predictive genes for PE.

Furthermore, the R package "rms" was used to construct calibration plots and validate the predictive performance of the models. Calibration curves and Decision Curve Analysis (DCA) were employed for this purpose. Finally, the diagnostic value of the prediction model was validated in the GSE75010 dataset using the ROC curve (Su et al. 2023).

Prediction of upstream transcription factors for feature genes

Through the application of the "regulon" extension in Cytoscape, the upstream transcriptional factors (TFs) governing the feature genes were anticipated. The default parameter settings were used, which included a motif collection of 10 K consisting of 9713 position weight matrices (PWMs) and a dataset of 1120 ChIP-seq tracks sourced from ENCODE raw signals. The minimum NEScore was set to 3, and the rank threshold for visualization was set to 5000. The ROC threshold for calculating the AUC was 0.03 (%). Additionally, orthologous gene analysis required a minimum identity threshold of zero, and the motif similarity analysis adhered to a strict false discovery rate (FDR) threshold of 0.001. The motif rankings and track rankings databases encompassed a 20 kb region focused on the Transcription Start Site (TSS) for analysis, incorporating species from the 7-species motif rankings database and ChIP-seq-derived track rankings database, respectively. Subsequently, TFs that targeted a minimum of 3 genes were selected, and a regulatory network for TFs was constructed based on these criteria (Tang et al. 2022).

Cell cultivation and handling

The American Type Culture Collection (ATCC) in Manassas, VA supplied the HTR-8/SVneo cellular system (Ref: CRL-3271) and the BeWo cellular system (Ref: CCL-98). Cultivation took place in RPMI-1640 culture solution (Product code: 11875119, Thermo Fisher Scientific) comprising 10% fetal bovine serum (Product code: 10099141C, Thermo Fisher Scientific, Shanghai, China) and antibacterial medicines (100 units per milliliter of penicillin and 100 µg per milliliter of streptomycin). In a 75 cm2 vessel, the cells were nurtured at a temperature of 37 °C and in an environment with 5% CO2, and the fluid was altered every 48–72 h.

In addition, cell culturing for the establishment of an in vitro model was carried out using the AW200SG anaerobic workstation by Electrotek in the UK which operated at 37 °C and contained 1% O2, 5% CO2, and 94% N2. The cells were cultured under conditions of Erastin (2.5 μM, catalog number: M2679, AbMole) or RSL3 (0.1 μM, catalog number: M9060, AbMole) treatment and then supplemented with 0.1 μM ferrostatin-1 (Fer-1), a ferroptosis inhibitor (catalog number: M2698, AbMole) (Lai et al. 2022a, b).

Cell transfection

The HTR-8/SVneo and BeWo cell cultures were established in 6-well containers with 105 cells seeded in each well. si-ARID3A, si-EPHB3, si-PAPPA2, negative control siRNA (si-NC), ARID3A overexpression plasmid (oe-ARID3A), oe-EPHB3, oe-PAPPA2, and negative control overexpression plasmid (oe-NC) were procured from GenePharma (Shanghai, China). Below are the specified groupings: Nor group (cells under normal oxygen conditions), Hyp group (Control) (cells under hypoxic conditions), oe-NC group (hypoxic cells transfected with oe-NC), oe-ARID3A group (hypoxic cells transfected with oe-ARID3A), si-NC group (hypoxic cells transfected with si-NC), si-ARID3A group (hypoxic cells transfected with si-ARID3A), Erastin group (hypoxic cells exposed to Erastin), RSL3 group (hypoxic cells exposed to RSL3), Erastin + Fer-1 group (hypoxic cells exposed to both Erastin and Fer-1), RSL3 + Fer-1 group (hypoxic cells exposed to both RSL3 and Fer-1), oe-NC + Erastin group (hypoxic cells transfected with oe-NC and exposed to Erastin), oe-ARID3A + Erastin group (hypoxic cells transfected with oe-ARID3A and exposed to Erastin), oe-ARID3A + si-EPHB3 + Erastin group (hypoxic cells transfected with oe-ARID3A, si-EPHB3, and exposed to Erastin), oe-ARID3A + si-PAPPA2 + Erastin group (hypoxic cells transfected with oe-ARID3A, si-PAPPA2, and exposed to Erastin), si-NC + Erastin group (hypoxic cells transfected with si-NC and exposed to Erastin), si-ARID3A + Erastin group (hypoxic cells transfected with si-ARID3A and exposed to Erastin), si-ARID3A + oe-EPHB3 + Erastin group (hypoxic cells transfected with si-ARID3A, oe-EPHB3, and exposed to Erastin), and si-ARID3A + oe-PAPPA2 + Erastin group (hypoxic cells transfected with si-ARID3A, oe-PAPPA2, and exposed to Erastin). Per the guidelines provided by the producer, transfect 100 nM of siRNA or 1 mg of plasmid into HTR-8/SVneo and BeWo cells employing Lipofectamine 2000 (catalog number: 11668500, Thermo Fisher Scientific). Cells were harvested 36 h after transfection for further experiments (Zhang et al. 2020). Table S2 presents the siRNA sequences utilized during this investigation.

ChIP assay

The detection of ChIP was accomplished through the application of the BeyoChIP™ ChIP Assay Kit (P2080S, Beyotime) as instructed by the manufacturer. Initially, lysis of the HTR-8/SVneo cellular model was performed with SDS Lysis Buffer that included a protease inhibitor. Subsequently, sonication was performed to obtain chromatin fragments. The lysate was subjected to immunoprecipitation using Protein A/G conjugated with ARID3A antibody (1:100, ab227274, Abcam) or IgG antibody (1:30, ab313801, Abcam). The enriched DNA fragments for the target genes, EPHB3 and PAPPA2, were then detected within the precipitated proteins (Zhang et al. 2020). Refer to Table S3 for the exact ChIP primer sequences.

Dual luciferase reporter gene experiment

Execution of the dual luciferase reporter gene analysis adhered to the manufacturer's provided protocol. HEK-293 cells (catalog number: CRL-3271, ATCC) were transfected with a pGL3-based vector containing the wild-type (WT) or mutant (MUT) EPHB3 and PAPPA2 promoters, an ARID3A overexpression plasmid, and a SEAP reporter plasmid using Lipofectamine 2000. After 48 h of transfection, the activity of the luciferase enzyme was measured using the Glomaxmutil detection system (Promega) (Zhang et al. 2020). The trial took place on three separate occasions.

Evaluation of cellular viability

Cell viability was assessed following the guidelines provided by the manufacturer using the MTT assay kit (referenced as M6494, Thermo Fisher Scientific Inc.). Seeding of cells took place in a 96-unit microtiter grid with 5000 cell quantities per well. Subsequently, 10 µl of MTT solution (5 mg/mL) were introduced into the respective wells. A microplate reader (Synergy HT, Bio-Tek, USA) was employed to determine the absorbance at 450 nm (Xie et al. 2019).

Intracellular Fe2+ level detection

The digested cell trypsin was transferred to a 100μL confocal well with a cell density of 1 × 104/mL, following the manufacturer's specified protocols. The cells were treated with a solution of FerroOrange at a concentration of 1 μmol/L (catalog number: F374, Shanghai Golden Harvest Biotech Co., Ltd.) for a duration of 30 min. Later on, the cells were viewed through a confocal fluorescence microscope (Yang et al. 2023).

Identification of intracellular ROS

Following the provided instructions from the manufacturer, the seeding of cells was performed in a 96-well plate, black in color with a clear bottom, at an approximate density of 1 × 104 cells/well. The cells were cultured for half an hour at 37 °C in the presence of a 5 μM MitoSOX™ mitochondrial superoxide indicator (catalog number: M36009, Thermo Fisher Scientific). Subsequent to 2 whole days, a triple PBS wash was performed on the cellular specimens. Quantification of the fluorescence emission at 510/580 nm was conducted through the utilization of a specialized microplate reader for fluorescence detection (Yang et al. 2023).

Lipid peroxidation detection

Following the instructions provided by the manufacturer, approximately 1 × 104 cells were placed in each well of a 6-well plate during the cell seeding process. The cell cultures were then subjected to low light conditions and administered with a 50 μM C11-BODIPY™ 581/591 probe (D3861 code, Thermo Fisher Scientific) for an hour. After the incubation, the cells underwent two rounds of PBS washing to eliminate surplus probes. A minor portion of serum-lacking medium for cultivation was supplied to enclose the cellular populace within the cultivation receptacle. Visuals of fluorescence were captured with the use of a confocal laser scanning microscope (Yang et al. 2022).

Animal models for PE

Procured from Beijing Vitonlihua Experimental Animal Technology Co., Ltd., the Sprague–Dawley rats, 9 weeks old and weighing 200–220 g, belonged to Strain 101. Rats received a conventional diet in the laboratory setting and maintained a light/dark cycle of 12 h each. Once a vaginal plug was observed, that day was designated as Gestational Day 1 (GD1). 7 divisions were established for the pregnant SD rat cohort through random classification: the standard control group (n = 8), the PE group (n = 8), the Ad-ARID3A group (where PE rats were intravenously injected with Ad-ARID3A, n = 8), the Ad-CTL + Fer-1 group (where PE rats were intravenously injected with Ad-CTL and Fer-1, n = 8), the Ad-ARID3A + Fer-1 group (where PE rats were intravenously injected with Ad-ARID3A and Fer-1, n = 8), the Ad-ARID3A + sh-EPHB3 group (where PE rats were intravenously injected with Ad-ARID3A and sh-EPHB3, n = 8), and the Ad-ARID3A + sh-PAPPA2 group (where PE rats were intravenously injected with Ad-ARID3A and sh-PAPPA2, n = 8). Surgery was performed on GD14 in order to induce a PE rat model, which involved reducing uterine perfusion pressure (RUPP). An occlusive silver clip with dimensions of 0.203 mm was placed on the aorta, as well as on the ovarian vessels (using a 14.1 mm clip), above the iliac bifurcation (Zhang et al. 2020). Similar surgery was carried out on rats that mimicked PE rats, but without occlusion.

Short hairpin RNA (shRNA) targeting EPHB3 and PAPPA2, recombinant adenovirus overexpressing ARID3A (Ad-ARID3A), and Ad-CTL were purchased from OBiO Technology Co., Ltd. (Shanghai, China). After modeling, the adenovirus was introduced via the rat's caudal vein, delivering a quantity of roughly 1 × 109 pfu per rat. The sequences of shRNA utilized in this research are specified in Table S4. After adenovirus injection, rats were intraperitoneally injected with either physiological saline (control) or Fer-1 (2 μmol/kg) every other day. SBP and MBP were measured using a non-invasive blood pressure monitor (MRBP system, IITC Life Science, Woodland Hills, CA, USA) in GD14, GD16, and GD19. GD19 pregnant rats were sacrificed to collect urine, fetal, and placental tissues. The Animal Care and Use Committee at our institution has authorized all animal experiments, in accordance with the guidelines provided in the National Institutes of Health Guide for the Care and Use of Laboratory Animals (Zhang et al. 2020).

Immunohistochemistry

Following the initial step of cleansing paraffin-embedded placental tissue with PBS, it was meticulously sectioned into slices, each measuring 4 μm in thickness. Subsequently, these sections were subjected to an overnight fixation process at room temperature, employing a 4% paraformaldehyde solution. The following day, another round of fixation was carried out.

To prepare the sliced sections for further analysis, a series of steps including dewaxing and rehydration were executed using a graded alcohol sequence. The retrieval of antigens was achieved through microwave treatment, involving exposure to a 10 mM sodium citrate solution at pH 6.0 for a duration of 15 min. To inhibit endogenous peroxidase activity, the sections underwent treatment with a 15% H2O2 solution at room temperature, with an incubation period lasting 3 min. Subsequently, the sections were exposed to primary antibodies, including anti-ARID3A (1:50, DF12558, Affinity Biosciences, Jiangsu, China), anti-EPHB3 (1:50, sc-100299, Santa Cruz Biotechnology, Santa Cruz, CA, USA), anti-PAPPA2 (1:50, DF13413, Affinity Biosciences), anti-4-HNE (1:100, catalog number: 6313R, Shanghai Yaji Biotechnology Co., Ltd., Shanghai, China), and anti-β-actin (1:200, sc-8432, Santa Cruz Biotechnology). Overnight at 4 °C, primary antibodies were subjected to an extended incubation. Afterward, the sections were exposed to a horseradish peroxidase compound for a time span of 30 min. Finally, diaminobenzidine was utilized to visualize the immune complexes. Imaging of the stained samples was accomplished using an EVOS microscope (Life Technologies, Carlsbad, CA, USA), and quantitative evaluation of positive staining in each sample was performed with the assistance of analysis software, specifically Image J (Version 1.49v, National Institutes of Health, Bethesda, MD, USA) (Yang et al. 2022).

RT-qPCR

Pursuant to the guidelines provided by the producer, frozen placental tissue and cell cultures were processed for total RNA extraction utilizing TRIzol reagent (code: 15596026) provided by Thermo Fisher Scientific, and TaKaRa Bio (Beijing) Co., Ltd. supplied the PrimeScript™ RT reagent Kit with gDNA Eraser (RR047A) for performing reverse transcription of the RNA to cDNA. BaoRui Biotechnology (Beijing) Co., Ltd. provided the TB Green® Premix Ex Taq™ (Tli RNaseH Plus) (Catalog No. RR420A) for the RT-qPCR investigation, and gene-specific primers with a 0.3 nM concentration. Refer to Table S5 for information on primers.

The comparative quantitative approach (2−ΔΔCT technique) was applied for determining the relative gene transcription level of the specific target, utilizing β-actin as the internal reference. The fold change in gene expression between the experimental and control groups is denoted by 2−ΔΔCt. Presented below is the calculation formula: ΔΔCT = ΔCt experimental group—ΔCt control group, where ΔCt = Ct target gene—Ct reference gene. Ct represents the amplification cycle number at which the real-time fluorescence intensity reaches the set threshold, indicating logarithmic growth of amplification at this point. Every sample underwent the setting of 3 repeated apertures, with the experiment being redone thrice (Zhang et al. 2020; Liao et al. 2022).

Western blot

The lysis of tissue and cell specimens was conducted with RIPA lysis buffer obtained from Thermo Fisher Scientific (catalog number: 89901). Determination of protein concentrations in individual samples was performed utilizing the BCA Protein Quantitation Kit provided by Thermo Fisher Scientific (catalog number: 23225). Subsequently, the total protein underwent separation applying 10% SDS-PAGE and was then conveyed onto a polyvinylidene fluoride (PVDF) membrane from Thermo Fisher Scientific (catalog number: 88518). Incubation of primary antibodies was carried out overnight at 4 °C following the blocking of the membrane with 5% skim milk. The primary antibodies employed were ARID3A (1:500, Santa Cruz Biotechnology, sc-398367), EPHB3 (1:200, Santa Cruz Biotechnology, sc-100299), PAPPA2 (1:500, Affinity Biosciences, DF13413), and β-actin (1:200, Santa Cruz Biotechnology, sc-8432). Post membrane cleansing, a secondary antibody conjugated with horseradish peroxidase was utilized in an hour-long incubation. Submerging the membrane in Beyotime's ECL solution (Code: P0018AS) and leaving it at ambient temperature for 60 s. Measurement of the protein's relative expression was achieved through a contrast in the intensity of grayscale between the specific band and the β-actin band, which functioned as the internal reference point (Zhang et al. 2020). Each trial underwent three rounds of testing.

Analysis of glutathione content

Tissues and cell samples were lysed using a 5% solution of 5-sulfosalicylic acid. According to the instructions provided by Abcam (Shanghai) Biotechnology Co., Ltd., glutathione levels were assessed by means of the Glutathione Assay Kit (Catalog Number: abs580006) (Zhang et al. 2020).

MDA (Methylenebis(4-phenylisocyanate))

Subsequently, the resuspended solution was blended with a 0.6% thiobarbituric acid solution (2 mL) in a 10 mL test tube, employing lysed tissue and cellular extracts. Then, the solution underwent heating in a boiling water container for 15 min, followed by measuring the solution's optical density at 532 nm (Zhang et al. 2020).

Analysis of statistics

R software v4.2.1 (R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism 8 (version 8.0.2.263, GraphPad Software, USA) were employed for statistical analysis in this study. Mean values with standard deviation were showcased, and group distinctions were appraised through the wilcoxTest approach. The correlation analysis was performed utilizing Spearman's methodological approach. Diverse comparisons among groups were conducted via one-way ANOVA, with subsequent post hoc assessments done using Tukey's methodology. The criterion for statistical significance was P < 0.05.