This study included women treated at the Reproductive Medicine Center of the Second Hospital of Lanzhou University between May 2024 and May 2025. A total of 36 endometrial samples were collected, including 30 from patients diagnosed with PCOS undergoing in vitro fertilization–embryo transfer and 6 from healthy fertile controls.

Given the exploratory nature of this proteomic study, no a priori power calculation was performed. A post hoc power analysis based on the observed average effect size across six key differentially expressed proteins (Cohen’s d = 1.42; n₁ = 30, n₂ = 6; $\alpha$ = 0.05, two-tailed) indicated a statistical power of approximately 0.95, suggesting adequate sensitivity to detect large intergroup differences.

PCOS was diagnosed according to the Rotterdam criteria, requiring at least two of the following: (1) oligo- or anovulation; (2) clinical or biochemical hyperandrogenism; and (3) polycystic ovarian morphology ($\ge$12 follicles or ovarian volume $>$10 mL). Inclusion criteria were women aged 21–40 years with infertility lasting $\ge$1 year and partners with normal semen parameters. Exclusion criteria included tubal obstruction, endometriosis, uterine malformation or adhesions, diabetes, thyroid dysfunction, hepatic or renal impairment, and recent (within 3 months) use of medications affecting hormonal or metabolic function (e.g., contraceptives, insulin sensitizers, corticosteroids). Healthy controls were women with spontaneous conception and normal early pregnancy who had no history of metabolic or chronic inflammatory diseases and had never received assisted reproductive therapy.

All participants underwent endometrial biopsy under hysteroscopy during the midluteal phase. Biopsy samples were rinsed in saline, snap-frozen in liquid nitrogen, and stored at –80 °C for proteomic analysis.

This study was approved by the Ethics Committee of the Second Hospital of Lanzhou University (Approval No. 2025A-774). Written informed consent was obtained from all participants. All procedures were conducted in accordance with the ethical principles of the Declaration of Helsinki.

Baseline clinical data included maternal age, height, weight, and body mass index (BMI). In addition, age and BMI were further categorized into clinically relevant subgroups (25, 26–30, $>$30 years; 18.5, 18.5–24.9, $\ge$25 kg/m²) for descriptive comparison between groups [18, 19].

Venous blood was collected on the day of endometrial biopsy for biochemical and hormonal measurements. All biochemical and hormonal data were obtained from the hospital’s electronic medical record system, with assays performed by the Department of Clinical Laboratory under standardized procedures. Serum levels of anti-Müllerian hormone (AMH), luteinizing hormone (LH), follicle-stimulating hormone (FSH), testosterone (T), and fasting insulin (FINS) were measured using a chemiluminescence immunoassay on the Cobas 8000 analyzer (E602; Roche Diagnostics, Basel, Switzerland). Fasting blood glucose (FBG) and lipid parameters including total cholesterol (TC) and triglycerides (TGs) were measured enzymatically on the Cobas 8000 analyzer (C702; Roche Diagnostics). The homeostatic model assessment of IR (HOMA-IR) was calculated as follows: HOMA-IR = [FBG (mmol/L) $\times$ FINS (µU/mL)] / 22.5 [20, 21].

Clinical pregnancy was confirmed by both ultrasonography and laboratory testing. Specific criteria included observation of an intrauterine gestational sac and fetal cardiac activity via transvaginal ultrasound at $\ge$5 weeks. Pregnancy outcomes were categorized as live birth or adverse pregnancy. Adverse pregnancy outcomes included: biochemical pregnancy (defined as transient elevation of serum beta human chorionic gonadotropin without ultrasound evidence of a gestational sac), early miscarriage (defined as spontaneous pregnancy loss before 12 weeks of gestation), and late miscarriage (defined as fetal loss between 12 and 24 weeks of gestation). Pregnancy duration was recorded from the time of embryo transfer to final outcome for use in survival analyses and model construction.

Endometrial tissues were homogenized (FastPrep-24 5G; MP Biomedicals, Santa Ana, CA, USA) under liquid nitrogen and lysed in SDT lysis buffer (4% Sodium dodecyl sulfate (SDS), 100 mM dithiothreitol (DTT), 150 mM Tris-HCl, pH 8.0) (ED-8452; Guangzhou Saiguo Biotechnology, Guangzhou, China). Samples were shaken at 6.0 m/s for 60 s, followed by six cycles of intermittent sonication (5 s on, 5 s off). Lysates were heated at 100 °C for 15 min and centrifuged at 14,000 g for 40 min. Supernatants were collected for quantification using the BCA assay (G2026; Wuhan Servicebio, Hubei, China).

Proteins were alkylated with 100 mM iodoacetamide for 30 min in the dark and filtered using 10 kDa ultrafiltration units. After washing with urea-alkylating (UA) buffer and 25 mM NH₄HCO₃, proteins were digested overnight with trypsin (1:50, w/w) at 37 °C. Peptides were desalted using C18 cartridges (567270U; Sigma, St. Louis, MO, USA), vacuum dried, and reconstituted in 0.1% formic acid. The peptide concentration was measured at 280 nm, after which samples were fractionated into 10 fractions using a high pH reversed-phase kit (84868; Thermo Fisher Scientific, Waltham, MA, USA) and then stored in 0.1% formic acid.

Peptide fractions were analyzed by Shanghai Applied Protein Technology Co., Ltd. (Shanghai, China) using the Evosep One LC system coupled to a timsTOF Pro mass spectrometer (Bruker, Billerica, MA, USA).

Data-dependent acquisition (DDA) was used to generate a spectral library, followed by data-independent acquisition (DIA) for quantification. DDA settings were as follows: precursor scan range 300–1800 m/z; resolution 60,000; automatic gain control (AGC) target 3 $\times$ 10⁶; max injection time (IT) 25 ms; and dynamic exclusion 30 s. Each full MS1 scan was followed by 20 MS2 scans (resolution 15,000, AGC target 5 $\times$ 10⁶, normalized collision energy 30 eV). DIA acquisition used a full scan range of 300–1800 m/z with MS1 resolution of 120,000, AGC target of 3 $\times$ 10⁶, and maximum IT of 50 ms. The LC gradient was 90 min at a flow rate of 250 nL/min, using 0.1% formic acid in 80% acetonitrile as mobile phase B. Quality control samples—pooled from all samples—were analyzed after every six runs to ensure data stability.

Raw MS data were processed with Spectronaut (version 14.4.200727.47784; Biognosys AG, Schlieren, Switzerland) using the UniProt human protein database. Search parameters were as follows: trypsin digestion (maximum of 2 missed cleavages); one fixed modification (carbamidomethylation); and two variable modifications (oxidation on methionine residues and acetylation on the protein’s N-terminus. Precursor and fragment tolerances were 10 ppm and 0.02 Da, respectively, with a false discovery rate (FDR) 1%, calculated as FDR = N(decoy) $\times$ 2 / [N(decoy) + N(target)], ensuring a 99% confidence level for all reported identifications. DIA data were log₂-transformed and normalized; proteins with $>$50% missing values were excluded. Differential expression was determined by the Student’s t-test with Benjamini–Hochberg correction (fold change $>$1.5, FDR 0.05).

DEPs were identified using the “limma” package (version 3.66.0; Bioconductor, https://bioconductor.org/packages/limma/) in R (v4.5.1; R Foundation for Statistical Computing, Vienna, Austria) with thresholds $|$log₂FC$|$ $>$0.58 and FDR 0.05. Subcellular localization was predicted using CELLO v2.5. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were conducted using the “clusterProfiler” package (version 4.18.2; Bioconductor, https://bioconductor.org/packages/clusterProfiler/) to elucidate potential biological processes, cellular components, molecular functions, and pathways.

Necroptosis-related gene sets were retrieved from GeneCards (https://www.genecards.org/), KEGG (https://www.kegg.jp/), and National Center for Biotechnology Information (NCBI) (https://www.ncbi.nlm.nih.gov/) databases. These genes were intersected with the DEPs to identify necroptosis-related DEPs. Machine learning algorithms were applied using R packages “e1071”, “kernlab”, and “caret” to screen for core necroptosis-associated proteins.

Protein–protein interaction (PPI) networks were constructed using the STRING database (v11.5, https://string-db.org/) and visualized in Cytoscape. Correlation networks among necroptosis proteins were plotted using “corrplot”, “reshape2”, and “igraph”.

Univariate Cox regression (R package “survival”) was used to identify necroptosis-related DEPs significantly associated with pregnancy outcomes (p 0.05). Then the Least Absolute Shrinkage and Selection Operator (LASSO)–Cox regression (R package “glmnet”, seed = 2000) was applied as a penalized method to minimize overfitting, with internal 10-fold cross-validation to enhance model stability. The necroptosis risk (NecroSig) score was calculated as: NecroSig Score = ${\sum }_{i=1}^n\beta i\times Expri$, where $\beta$₁–$\beta$_n are LASSO coefficients and Expr₁–Expr_n are standardized expression values [log₂(TPM + 1)].

Patients were divided into high- and low-risk groups based on the median risk score. Principal component analysis was conducted using “Rtsne” and “ggplot2”. Kaplan–Meier survival analysis was performed to compare pregnancy outcomes between groups. Time-dependent receiver operating characteristic (ROC) curves (“timeROC”) were used to assess predictive accuracy at 6, 28, and 37 weeks of gestation. Decision curve analysis (DCA) (“ggDCA”) compared the net clinical benefit of the NecroSig model with conventional predictors (age, BMI, HOMA-IR, lipids). A nomogram and calibration curve were generated to evaluate consistency between predicted and observed live birth rates.

All statistical analyses were performed using SPSS 26.0 (IBM Corp., Armonk, NY, USA) and R 4.5.1. Normality was tested with the Kolmogorov–Smirnov test. Continuous variables with non-normal distribution are expressed as the median (P25, P75) and were compared using the Mann–Whitney U test; categorical variables were compared using the chi-square or Fisher’s exact test. p 0.05 was considered statistically significant. Spearman’s rank correlation analysis was performed to evaluate the associations of necroptosis-related proteins with both clinical metabolic indicators and endometrial receptivity markers within the PCOS cohort. Correlations with p 0.05 were considered statistically significant. Missing values for clinical and proteomic variables were examined prior to analysis. Proteins with $>$50% missing values were excluded, and remaining missing data were handled using pairwise deletion for correlation analysis and complete-case analysis for regression models. The proportional hazards assumption of the Cox regression model was verified using Schoenfeld residuals, and no significant violation was detected (p $>$ 0.05).

The clinical and pregnancy data of the two groups are summarized in Table 1. Women with PCOS showed significant elevations in FINS, HOMA-IR, AMH, LH, LH/FSH ratio, and T levels, along with a marked decrease in endometrial thickness (ET). No significant differences were observed in age, height, weight, BMI, FSH, or gestational duration of live birth between the two groups. In the control group, no adverse pregnancy outcomes were observed. By contrast, among the 30 women with PCOS, 14 experienced adverse pregnancy outcomes including 10 early miscarriages, 3 biochemical pregnancies, and 1 late miscarriage. No cases of preterm delivery or stillbirth were recorded.

Quantitative proteomic analysis of all endometrial samples was performed to identify PCOS-associated proteins. A total of 4425 proteins were commonly detected in both PCOS and control samples. Among them, 611 proteins were identified as differentially expressed, including 132 upregulated and 479 downregulated proteins in the PCOS group (Fig. 1A).

Fig. 1.

Identification and enrichment analysis of differentially expressed proteins. (A) Volcano plot showing DEPs between PCOS and control endometria. (B) Functional enrichment analysis of DEPs. DEPs, differentially expressed proteins; KEGG, Kyoto Encyclopedia of Genes and Genomes; BP, Biological Process; CC, Cellular Component; MF, Molecular Function.

To further elucidate the biological functions of the DEPs, GO and KEGG pathway enrichment analyses were performed (Fig. 1B). GO analysis revealed that the DEPs were significantly enriched in biological processes such as ribonucleoprotein complex biogenesis, regulation of viral process, and translational control in response to stress. For the cellular component category, DEPs were mainly associated with the small nuclear ribonucleoprotein complex, trans-Golgi network transport vesicle membrane, and clathrin-coated vesicle. Regarding molecular function, enrichment was observed in double-stranded RNA binding, histone methyltransferase binding, and ribonucleoprotein complex binding. KEGG pathway analysis identified influenza A infection and necroptosis as the two most significantly enriched pathways, indicating a potential inflammatory association underlying endometrial dysfunction in PCOS, rather than confirming a direct causal mechanism.

To determine necroptosis-related DEPs associated with PCOS prognosis, 628 necroptosis-related genes were retrieved from public databases (GeneCards, KEGG, and NCBI) and intersected with the PCOS DEPs, yielding 15 overlapping proteins.

Univariate Cox regression identified several necroptosis-related proteins significantly associated with pregnancy outcomes. Subsequent LASSO regression further reduced the candidates to six prognostically relevant necroptosis-related proteins: dynamin 1 like (DNM1L), thioredoxin (TXN), eukaryotic translation initiation factor 2 alpha kinase 2 (EIF2AK2), adenosine deaminase acting on RNA (ADAR), calmodulin-dependent protein kinase II gamma (CAMK2G), and calmodulin-dependent protein kinase II delta (CAMK2D) (Fig. 2A–C).

Fig. 2.

Identification and analysis of prognostic necroptosis-related DEPs. (A) Forest plot of univariate Cox regression for necroptosis-related DEPs. (B) Partial likelihood deviance plot from LASSO regression. (C) LASSO coefficient profiles of the selected proteins. (D) PPI network of prognostic necroptosis-related proteins. (E) Correlation heatmap of the six necroptosis-related DEPs. DNM1L, dynamin 1 like; TXN, thioredoxin; EIF2AK2, eukaryotic translation initiation factor 2 alpha kinase 2; ADAR, adenosine deaminase acting on RNA; CAMK2D, calmodulin-dependent protein kinase II delta; CAMK2G, calmodulin-dependent protein kinase II gamma; LASSO, Least Absolute Shrinkage and Selection Operator; PPI, protein–protein interaction.

A PPI network constructed using the STRING database illustrated their putative molecular interconnections (Fig. 2D), whereas the correlation heatmap demonstrated their expression relationships (Fig. 2E).

Using the six prognostic necroptosis-related proteins, a necroptosis risk model (NecroSig) was established based on 30 patients with PCOS with pregnancy outcome data. The risk score was calculated as follows:

NecroSig (PCOS) = 0.000280179 $\times$ DNM1L – 0.000601845 $\times$ TXN + 0.003339755 $\times$ EIF2AK2 + 0.004178775 $\times$ ADAR – 0.001023549 $\times$ CAMK2G – 0.000960423 $\times$ CAMK2D

Among these, DNM1L, EIF2AK2, and ADAR had positive coefficients, indicating potential risk factors for adverse pregnancy outcomes; whereas TXN, CAMK2G, and CAMK2D showed negative coefficients, suggesting protective roles. Based on the median NecroSig score, patients with PCOS were stratified into high- and low-risk groups (Fig. 3A). The high-risk group exhibited significantly lower live birth rates than the low-risk group (Fig. 3B). Among the six prognostic proteins, EIF2AK2 and ADAR showed significant expression differences between risk groups (Fig. 3C). Kaplan-Meier survival analysis confirmed that the low-risk group consistently demonstrated superior pregnancy outcomes (Fig. 3D).

Fig. 3.

Construction of the NecroSig model. (A) Distribution of risk scores. (B) Survival status plot showing the relationship between risk score and pregnancy outcome. (C) Boxplots of protein expression levels between risk groups. (D) Kaplan–Meier survival curves comparing pregnancy outcomes between high- and low-risk groups. ** p 0.01; *** p 0.001; ns, not significant.

A nomogram integrating the six prognostic proteins was constructed to predict the probability of adverse pregnancy outcomes at 6 weeks (42 days), 28 weeks (196 days), and 37 weeks (259 days) (Fig. 4A). Time-dependent ROC analysis demonstrated strong predictive performance, with area under the curve (AUC) values of 0.828, 0.878, and 0.903 at 6, 28, and 37 weeks, respectively (Fig. 4B). At 37 weeks, the NecroSig model showed higher predictive accuracy than individual clinical parameters such as age (AUC = 0.551), BMI (0.712), AMH (0.592), insulin (0.684), HOMA-IR (0.673), TC (0.633), and TGs (0.633) (Fig. 4C1). A combined clinical model integrating these indicators yielded an AUC of 0.847, still lower than the NecroSig model (AUC = 0.903). When both proteomic and clinical indicators were integrated, the multimodal model demonstrated the highest predictive performance (AUC = 0.944) (Fig. 4C2). The multimodal model showed the highest discriminative ability, highlighting the complementary value of proteomic and clinical markers. DCA further confirmed that the NecroSig model provided greater net clinical benefit compared with conventional indicators (Fig. 4D). Heatmaps revealed distinct clinical feature distributions between high- and low-risk groups (Fig. 4E). Correlation analysis showed significant associations between EIF2AK2 and both fasting insulin and HOMA-IR, whereas TXN was negatively correlated with BMI (Fig. 4F).

Fig. 4.

Evaluation of model performance and correlation with clinical data. (A) Nomogram predicting adverse pregnancy risk at 6, 28, and 37 weeks. (B) Time-dependent ROC curves for live birth prediction. (C1) ROC curves of individual clinical indicators for predicting live birth at 37 weeks. (C2) Comparison between the NecroSig model, the combined clinical model, and the integrated multimodal model. (D) DCA for 37-week outcome prediction. (E) Heatmap showing distribution of clinical features across risk groups. (F) Correlation heatmap between clinical variables and prognostic proteins. AUC, area under the curve; TC, total cholesterol; TG, triglyceride; ROC, receiver operating characteristic; DCA, decision curve analysis; FPR, false positive rate; TPR, true positive rate.

Within the PCOS cohort, exploratory correlation analyses were conducted between the six necroptosis-related proteins and classical markers of endometrial receptivity, including homeobox A11 (HOXA11) and integrin subunit beta-1 (ITGB1), ITGB3, and ITGB5 [22]. Three significant associations were identified: DNM1L–ITGB1 (r = 0.38, p = 0.049), CAMK2G–ITGB3 (r = –0.41, p = 0.030), and CAMK2D–HOXA11 (r = –0.57, p = 0.013). These findings support a potential molecular crosstalk between necroptosis and integrin-mediated receptivity pathways (Supplementary Fig. 1; Supplementary Tables 1,2).