In this study six public transcriptomic datasets GSE148241 (12), GSE44711 (13), GSE74341 (14), GSE114691 (15), GSE10588 (16) and GSE75010 (17) ( Supplementary Table S1 ) were downloaded from the Gene Expression Omnibus database (GEO) (18) (https://www.ncbi.nlm.nih.gov/geo/) based on: 1) Sample size ≥ 5 per group (EOPE vs. control); 2) Clinically confirmed EOPE diagnosis per ACOG criteria; 3) Exclusive use of placental tissue (no cell lines or other tissues). Platform-specific normalization pipelines were implemented: RNA-seq datasets (GSE114691) underwent median-of-ratios normalization using DESeq2, while microarray datasets (GSE148241, GSE10588, GSE44711, GSE74341, GSE75010) were processed with limma quantile normalization.

The ferroptosis-related gene sets were acquired from FerrDb (http://www.zhounan.org/ferrdb/index.html) (19). FerrDb contained 258 experimentally validated FRGs at time of access (2022). Database expansions since then include indirect regulators not analyzed here. The full list of FRGs was shown in Supplementary Material .

Placental specimens were obtained from six normal normotensive controls and six EOPE cases (determined through power analysis and clinical constraints) after obtaining informed consent ( Supplementary Table S2 ). EOPE diagnosis followed the American College of Obstetricians and Gynecologists (ACOG) criteria, requiring new-onset hypertension (systolic BP ≥140 mmHg/diastolic ≥90 mmHg) with proteinuria (≥300 mg/24h) or end-organ dysfunction, occurring after 20 weeks and before 34 weeks of gestation. Exclusion criteria include: pregnant women diagnosed with cardiovascular disease, metabolic syndrome, endocrine disorders, liver, or kidney diseases, as well as cases involving fetal malformations or chromosomal abnormalities. The samples were then cleansed with balanced salt solution to eliminate any blood, flash-frozen in liquid N₂ within 20 min of delivery, and stored at -80°C for ≤6 months before analysis for western blot analyses. The Ethics Committee of the First Affiliated Hospital of Bengbu Medical College granted approval for this study under protocol number ([2024]KY023).

HTR-8/SVneo cells were cultured in RPMI-1640 complete medium (RPMI-1640 medium containing 10% fetal bovine serum (FBS, 164210, Procell Life Science & Technology, Wuhan, Hubei, China), 100 U/ml penicillin, and 100 μg/ml streptomycin) at 37 °C with 5% CO2. Cells were seeded in the RPMI-1640 complete medium for 24 hours and then transfected with siRNAs or negative control siRNA using Lipo3000 (Thermo Fisher Scientific, USA) for 48 hours and the knockdown efficiency was confirmed by real-time quantitative PCR and western blot. The siRNA sequences are shown as Supplementary Table S3 .

Total RNA from cells was extracted using TRIzol (Invitrogen, Carlsbad, CA). The concentration and purity of the extracted RNA were evaluated with a NanoDrop™ One/OneC (Thermo Fisher Scientific). cDNA was generated using a Novo-Scrip^® Plus All-in-one 1st Strand cDNA Synthesis Kit (E042-01B, Novoprotein, China) at 42°C for 5 minutes, 50°C for 15 minutes, and 75°C for 5 minutes. The qRT-PCR was performed using NovoStart^® SYBR qPCR SuperMix Plus (E167-01A, Novoprotein, China) and the LightCycler 480 (Roche, Basel, Switzerland) was used for amplifications and quantifications. Primers were validated through melt curve analysis (single-peak confirmation) and the sequences were shown as follows (5′–3′):

Cultured cells or pre-homogenized tissues were lysed with ​ice-cold RIPA buffer (Beyotime, P1005, China) containing 1× protease inhibitor cocktail​(Beyotime, P0013B, China) with gentle agitation at 4°C for 30 min, followed by​centrifugation at 12,000×g for 15 min at 4°C to collect clarified lysates. The total protein was quantified using the Total Protein Assay Kit (Beyotime, P0012, China) and subsequently denatured. Proteins (50 μg per lane) were separated on SDS–polyacrylamide gels and then transferred onto polyvinylidene fluoride (PVDF) membranes. The membranes were incubated overnight at 4°C with the following primary antibodies: anti-HILPDA (sc-376704, 1:100, Santa Cruz) and anti-β-actin (LF201, 1:5000, Epizyme, Shanghai, China). After being washed with TBST, the membranes were incubated for 2 hours at room temperature with HRP-labeled Goat Anti-Mouse IgG secondary antibodies (LF101, 1:5000, Epizyme, Shanghai, China). Protein bands were detected using a ChemiDoc Imaging System (UVP ChemStudio, Analytik Jena, Germany) and quantified with ImageJ software.

HTR-8/SVneo cells transfected with siHILPDA or controls (scrambled siRNA) were scratched using 200-μl pipette tips when cells were cultured to approximately 90% confluency. After gently washing the cells twice with phosphate-buffered saline (PBS), the wells were replenished with serum-free RPMI 1640 medium (Thermo Fisher Scientific, USA) to eliminate proliferation bias. Microscopic images of the wound areas were captured at 0 h (baseline) and 24 h post-wounding using an inverted phase-contrast microscope (Motic, Xiamen, China). Quantitative analysis of cell migration was performed by measuring the residual scratch area at both time points with ImageJ software. The migration rate was calculated using the following formula: [(Initial area - Final area)/Initial area] × 100%.

The limma and DESeq2 packages in R were used to identify differentially expressed genes (DEGs). For RNA-seq count data, genes were filtered by requiring ≥10 raw counts in a minimum of k samples, with k defined as the sample size of the smallest experimental group (EOPE or CON). DEGs were defined by p-value <0.05 and |log₂FC| > 1. In GSE114691 datasets, all patients were stratified into HILPDA-high and HILPDA-low groups based on median HILPDA expression, with DEGs identified using p-value <0.05 and |FC| > 1. The VennDiagram package in R was used to identify common genes among DEGs in the GSE148241, DEGs in GSE44711 and ferroptosis-related gene sets. Receiver operating characteristic (ROC) curves and the area under the ROC curve (AUC) were calculated to evaluate the diagnostic potential of the candidate genes.

Three machine learning models (LASSO, SVM-RFE, Random Forest) were trained on z-score normalized expression of the 171 WGCNA-selected genes. 10-fold cross-validation with 5 repeats was used for hyperparameter tuning and performance evaluation. The least absolute shrinkage and selection operator (LASSO) regression, random forest and support vector machine-recursive feature elimination (SVM-RFE) were performed on GSE114691(n=61: 21 EOPE/40 controls). Model discrimination was quantified by mean cross-validated AUC. External validation was performed on GSE75010.

The LASSO regression, implemented via the R package glmnet (v4.1-8), employs L1 regularization to achieve high-dimensional data compression and feature selection. In this analysis, the algorithm effectively distilled 171 candidate genomic features into 5 non-zero biomarkers through an optimized λ parameter (lambda.1se = 0.186) determined via 10-fold cross-validation.

The random forest algorithm was implemented using the randomForest (v4.7-1.2) package in R. Initial model training employed 1,000 decision trees (ntree = 1000) to capture feature interactions, with subsequent optimization using out-of-bag (OOB) error analysis. The optimal mtry (number of features per split) was determined via iterative testing (1–20 features), selecting the value minimizing OOB error (mtry = 14).

The SVM-RFE algorithm integrates L2-norm regularization with backward feature elimination to identify optimal feature subsets for predictive modeling. Implemented via the e1071 (v1.7-16) and caret (v7.0-1) R packages. A 10-fold cross-validation (nfold = 10) was applied to partition samples into training/validation subsets.

To identify the co-expression gene module associated with EOPE and the screened candidate gene HILPDA, WGCNA was performed in GSE114691 datasets using the WGCNA package in R. Specifically, hierarchical clustering analysis was first performed using hclust function, and then the suitable soft threshold power was determined using the pickSoftThreshold function and an adjacency cutoff value of 0.9. We selected a soft-thresholding power of 19 based on rigorous evaluation of the scale-free topology criterion. This power represented the minimum value achieving satisfactory scale-free model fit (R² = 0.92) while maintaining biological interpretability. Finally, co-expression gene modules were constructed by set the minimum module size of 30.

To perform functional enrichment analysis of screened genes associated with EOPE and HILPDA expression, Functional enrichment was performed using the clusterProfiler package (v4.10.0). Gene ontology (GO) terms (20) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (21) analyses were conducted with enrichGO and enrichKEGG package respectively. The GO analysis encompassed biological processes (BPs), cellular components (CCs), molecular functions (MFs), and pathway analyses. The Benjamini-Hochberg method was employed to adjust the P values, with adjusted P-values < 0.05 set as the threshold for significance.

Immune cell infiltration analysis was performed using ​CIBERSORT (v1.03) (22).​​ The LM22 gene signature​ (downloaded from CIBERSORT portal) containing 547 genes defining 22 human immune cell types. Gene expression matrices were formatted with unique gene symbols​ in row names (duplicates resolved by ​mean expression). We ensured ​no missing values​ by excluding genes with missing data and enabled ​quantile normalization (QN=TRUE)​​ to minimize batch effects across samples. To investigate immunomodulatory relationships, Spearman’s rank correlation analysis was systematically performed between HILPDA expression and immune subsets.

In the GSE114691 datasets, gene set enrichment analysis (GSEA) (23) was performed between low- and high-HILPDA expression (dichotomized at median expression) to identify the biological process (GO term) and pathways (KEGG) associated with HILPDA expression. Gene set permutations to compute log2 fold change. GSEA analysis was performed using the gseGO and gseKEGG methods of the clusterprofiler package in R.

Continuous variables between two groups were compared using unpaired Student’s t-tests​, implemented in GraphPad Prism 9. Bioinformatics analyses were conducted using R software (v4.3.1). The chi-squared tests were used for the analysis of count data between two groups. All data are presented as mean ± standard deviation (SD). A P-value of less than 0.05 was considered statistically significant.

To screen for the differentially expressed FRGs in EOPE, differential expression analysis between the placenta samples from EOPE and control patients was performed in GSE148241(32 controls, 9 EOPE) and GSE44711 (8 controls, 8 EOPE) datasets. The study workflow was shown in Figure 1 . Volcano plots identified 763 DEGs (432 upregulated) in GSE148241 placentas ( Figure 2A ) and 1,014 DEGs (597 upregulated) in GSE44711 ( Figure 2B ) based on p value < 0.05 and |logFC| > 1. Two differentially expressed FRGs (HILPDA and SCD) were identified in EOPE after intersection with DEGs in GSE148241 and GSE44711 datasets ( Figure 2C ). To verify the expression of the two screened genes in the placenta of PE subtypes, we further analyzed further the expression of HILPDA and SCD in GSE74341. The GSE74341 dataset was prioritized for validation due to its ​comprehensive stratification of PE subtypes: early-onset PE (n=7, <34 weeks), late-onset PE (n=8, >36 weeks), preterm controls (n=5, <34 weeks), and term controls (n=5, >36 weeks). This design enables rigorous comparison of HILPDA expression across gestation-matched pathological and physiological contexts. The HILPDA expression in EOPE placentas was significantly increased compared to the Preterm, Term and LOPE groups ( Figure 2D ). However, the expression of SCD in the EOPE placentas showed no significant difference in either the GSE74341 dataset (vs. preterm and term placentas; Supplementary Figure 1A ) or the GSE114691 dataset (vs. normotensive controls; Supplementary Figure 1B ). Additionally, ROC curve analysis demonstrated that HILPDA exhibited high diagnostic accuracy for EOPE (AUC = 0.812 in GSE114691 and AUC = 0.871 in GSE10588; Figures 2E, F ), while the AUC of SCD was 0.655 in GSE10588 ( Supplementary Figure 1C ).

Given the elevated HILPDA expression in EOPE placentas and its strong diagnostic performance, additional genes correlated with HILPDA expression and associated with EOPE were identified in GSE114691 datasets. Stratification of GSE114691 samples (n=61) by median HILPDA expression yielded 33 HILPDA-high and 28 HILPDA-low placentas. As shown in Figure 3A , in the GSE114691 dataset, HILPDA expression was significantly upregulated in EOPE placentas, consistent with results from the GSE148241 and GSE44711 datasets. Compared to the HILPDA-low expression group, 298 DEGs were identified in the HILPDA-high samples (logFC > 1 and p value < 0.05), comprising 241 upregulated and 57 downregulated genes ( Figure 3B ). The 30 most significantly upregulated and 30 downregulated genes (ranked by log2FoldChange) were displayed in a heatmap ( Figure 3C ). These findings suggested HILPDA may take important roles in regulating placenta transcriptome.

Functional enrichment analyses of all 298 DEGs between HILPDA-high and HILPDA-low groups were systematically conducted to elucidate the biological significance of HILPDA-associated genes. GO enrichment plots for BP, CC, and MF terms are shown in Figures 4A–C , which revealed significant enrichment of differentially expressed genes (DEGs) between HILPDA-high and -low group across all three GO domains. The biological process (BP) category revealed prominent involvement in leukocyte migration and bacterial molecular pattern response, while cellular component (CC) analysis identified enrichment in collagen-rich extracellular matrix and MHC protein complexes ( Figures 4A, B , Supplementary Figure 2 ). Molecular function (MF) characterization further showed significant enrichment in immune receptor binding activity and humoral immune mediators ( Figure 4C , Supplementary Figure 2 ). Furthermore, the KEGG pathway analysis revealed that DEGs between HILPDA-high and -low group were significantly involved in pathways including cell adhesion molecules and HIF−1 signaling pathway ( Figure 4D ). The results from GSEA showed that biological processes leukocyte migration and mononuclear cell migration were significantly enriched in HILPDA-low samples, while cell adhesion molecules, HIF-1 signaling pathway, and response to hypoxia pathway significantly enriched in HILPDA-high samples ( Figure 4E ). The enrichment of these pathways (immune dysregulation, matrix dysmorphia, and hypoxia adaptation) suggested HILPDA’s role in EOPE pathogenesis.

To determine the gene sets closely associated with EOPE, WGCNA was performed on the top 5,000 most variably expressed genes (selected by median absolute deviation to capture biologically relevant signals) across GSE114691 samples (n=61). As shown in Figures 5A, B , a soft threshold of 11(Minimum value achieving scale-free topology and lower powers compromise network biological validity) was employed to achieve an adjacency cutoff value of 0.85. After setting the minimum gene module size to 30, a total of ten modules were identified ( Figure 5C ), with the blue module (enriched in immune regulation genes) showing the strongest positive correlation (cor = 0.85, P = 5.1×10^-192) with EOPE status​. The overlap between genes in the blue module and DEGs in the HILPDA-low/high groups included 171 genes, potentially identified as EOPE-correlated genes associated with HILPDA expression ( Figure 5F ).

Three machine learning models were trained on the 171 HILPDA-associated genes using GSE114691 samples with stratified 10-fold cross-validation (repeated 5 times) to maintain class balance. LASSO regression (glmnet v4.1-8) analysis (lambda.1se = 0.186) identified 5 candidate diagnostic biomarkers (PART1 NRAD1, HTRA1, LOC107985906, HTRA4) ( Figure 6A ). Subsequent application of SVM-RFE with 10-fold cross-validation revealed 4 high-confidence genes (LOC105369957, PART1, MYO7B, CLDN9) demonstrating optimal classification performance ( Figure 6B ). Finally, parallel random forest analysis (1000 decision trees, Mean Decrease Accuracy (ranked exclusively by Mean Decrease Accuracy) and Gini impurity criterion) independently prioritized 21 candidate genes through variable importance scoring ( Figure 6C ). Intersection analysis of the three method-specific candidate pools (LASSO:5, SVM-RFE:4, RF:21) identified one consensus biomarker PART1 ( Figure 6D ). PART1 is a long noncoding RNA that has been reported to has important role in regulating cell proliferation and invasion (24, 25). We validated the diagnostic performance of the identified biomarker PART1 using the independently selected external dataset GSE75010. Within this dataset, PE patients with a gestational age of less than 34 weeks were designated as the EOPE group. The ROC curve for PART1 was generated using raw expression values on a GSE75010 (49 EOPE, 77 controls) and present a good diagnostic performance for EOPE (AUC=0.889; 95% CI: 0.831-0.948) ( Figure 6E ).

Utilizing the CIBERSORT deconvolution algorithm, we quantitatively assessed immune cell infiltration patterns in normal and EOPE placentas ( Supplementary Figure 4 ). Comparative analysis revealed distinct immunological profiles between groups ( Figure 7A ). Specifically, EOPE placentas exhibited significantly elevated proportions of plasma cells and activated NK cells compared to normal counterparts. Conversely, normal placental tissues demonstrated higher infiltration levels of resting NK cells, monocyte, and resting mast cells ( Figure 7A ). Furthermore, Spearman’s rank correlation analysis uncovered significant associations between HILPDA expression and specific immune subsets ( Figures 7B–E , Supplementary Figure 5 , Supplementary Table S4 ). Notably, HILPDA demonstrated a strong positive correlation with activated dendritic cells and neutrophils, while showing an inverse relationship with resting mast cells.

To evaluate the differential expression of HILPDA in placental pathophysiology, western blot analysis of 6 biological replicates per group was performed in placental tissues from EOPE and normotensive pregnancy. As shown in Figures 8A, B , western blot analyses showed that the levels of HILPDA (band intensities normalized to β-actin) were markedly elevated in EOPE placentas compared to those from normal pregnancies. Furthermore, to investigate the effect of HILPDA on trophoblast cell migration, siRNAs were used to silence HILPDA expression in HTR-8/SVneo cells and the result showed a good silencing efficiency of HILPDA at both the mRNA and protein levels ( Figures 8C–E ). Wound healing assays quantified at 24 hours post-scratch demonstrated significantly enhanced migration in HILPDA-knockdown cells versus scrambled siRNA controls ( Figures 8F, G ). These findings suggested that HILPDA critically suppresses trophoblast migration capacity, and its pathological overexpression in EOPE placentas may directly contribute to impaired spiral artery remodeling, a hallmark defect in EOPE pathogenesis.