Identification of regulated cell death related genes in polycystic ovary syndrome using machine learning

Abstract

Polycystic ovary syndrome (PCOS) is one of the most prevalent endocrine disorders affecting women during their reproductive years, with global prevalence estimates ranging from 5 to 15%, depending on the diagnostic criteria used. Emerging evidence suggests that various forms of regulated cell death (RCD) mechanisms play a significant role in the development and progression of PCOS. However, existing research has yet to systematically investigate how RCD processes interact with the molecular pathophysiology of PCOS. Mapping these complex interactions—including the associated regulatory networks and molecular cascades—could provide critical insights into disease mechanisms. This study aims to identify specific RCD-related genetic markers and signaling pathways, which could serve as potential therapeutic targets for PCOS management. Our team conducted computational bioinformatics analyses to find differentially expressed genes (DEGs) between healthy ovarian tissues and those affected by PCOS, revealing 389 genes linked to RCD. Through machine learning techniques—including Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and Support Vector Machine (SVM) algorithms—we identified five critical hub genes. To gauge their diagnostic potential, we performed receiver operating characteristic (ROC) curve evaluations and mapped out protein interaction networks (PPI) to uncover relationships among these key genes. We then delved deeper using Single-Sample Gene Set Enrichment Analysis (ssGSEA), Gene Ontology (GO) enrichment studies, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway assessments to shed light on biological processes tied to the hub genes. These findings were corroborated through additional Gene Set Enrichment Analysis (GSEA) validation. Leveraging the NetworkAnalyst and RegNetwork platforms, we predicted upstream regulators like microRNAs (miRNAs), transcription factors, and gene-associated compounds. Finally, interaction networks were visualized via Cytoscape to illustrate these complex relationships. Through comparative analysis of PCOS and control groups, DEGs were pinpointed and cross-referenced with genes linked to RCD mechanisms. Machine learning techniques highlighted five hub genes with significant biological relevance. Comprehensive bioinformatics profiling demonstrated that these key genes were significantly enriched in biological processes related to immune-inflammatory responses, metabolic regulation via adipocytokine signaling, reproductive hormone activity, and epigenetic regulation. Furthermore, we identified 25 therapeutic compounds, 42 regulatory miRNAs, and 30 transcription factors (TFs) with strong functional relationships to these critical genetic markers. We identified five RCD-related hub genes within the DEGs of PCOS and control samples and further analyzed upstream and downstream pathways, to elucidate potential pathogenic mechanisms.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-18867-1.

Introduction

Polycystic Ovary Syndrome (PCOS), often referred to as the most prevalent endocrine disorder among women of reproductive age, impacts an estimated 5–15% of this population, posing a significant challenge to global health systems^1–4. Based on the 2003 Rotterdam criteria, diagnosing PCOS involves meeting at least two of the following three conditions: hyperandrogenism (whether clinical or biochemical), irregular menstrual cycles, and the presence of polycystic ovary morphology. These criteria classify PCOS into four distinct phenotypes: A (characterized by ovulatory dysfunction, hyperandrogenism, and polycystic ovary morphology), B (ovulatory dysfunction and hyperandrogenism), C (ovulatory dysfunction and polycystic ovary morphology), and D (PCOS with hyperandrogenism)⁵. While the diagnostic framework is fairly straightforward, recent discussions among researchers suggest consolidating the current four phenotypes into just two broader categories. Despite these advancements in classification, the precise causes and underlying mechanisms of PCOS remain somewhat elusive, especially when compared to the clarity of its diagnostic guidelines.

Accumulating evidence indicates that PCOS is a multifactorial disorder characterized by the interplay of insulin resistance, hyperandrogenism, and chronic low-grade inflammation^6–9. Insulin resistance plays a central role, as lipid accumulation in hepatic and skeletal muscle tissues activates the diacylglycerol/protein kinase C pathway, impairing GLUT-4 expression and reducing glucose uptake, ultimately disrupting insulin signaling pathways^10,11. To compensate for reduced insulin sensitivity, hyperinsulinemia develops, which stimulates ovarian androgen production, enhances luteinizing hormone (LH) secretion, and further elevates circulating androgen levels^9,12. These hormonal changes are compounded by altered adrenal responses to adrenocorticotropic hormone (ACTH) and decreased hepatic production of sex hormone-binding globulin (SHBG), intensifying androgen excess.

Concurrently, hyperglycemia promotes monocyte activation and excessive release of pro-inflammatory cytokines, such as tumor necrosis factor-α (TNF-α), exacerbating metabolic disturbances. Women with PCOS also exhibit elevated levels of advanced glycation end products (AGEs) and their receptors, contributing to oxidative stress^13,14. Increased levels of plasminogen activator inhibitor-1 (PAI-1) and cytokines including TNF-α and interleukin-6 (IL-6) further indicate the persistence of chronic inflammation^15,16. This synergistic dysregulation of androgens, insulin signaling, and inflammation establishes a self-perpetuating pathological cycle. Moreover, testosterone disrupts adipocyte function by inhibiting protein kinase C (PKC) activity¹⁷ and interferes with insulin receptor signaling in skeletal muscle through insulin receptor substrate 1 (IRS-1) phosphorylation¹⁸. Collectively, these findings illustrate how the synergistic dysregulation of insulin signaling, androgen production, and inflammatory responses creates a vicious cycle that underpins the pathogenesis of PCOS.

Regulated cell death (RCD) refers to a controlled form of cell demise, distinguished from accidental cell death (ACD) by its regulated, orderly nature, influenced by genetic factors and specific molecular pathways. Unlike ACD, RCD can be modulated through pharmacological or genetic interventions. In 2018, the Nomenclature Committee on Cell Death introduced guidelines for the classification of cell death, identifying twelve distinct modes, including apoptosis, autophagy, and necroptosis¹⁹. New forms of cell death, such as oxeiptosis and mitotic cell death, are gaining attention for their role in cancer development^20–22. However, research on RCD in non-cancerous diseases, including PCOS, remains limited.

Growing evidence suggests a strong connection between RCD and the progression of PCOS. Several compounds have been identified as potential inhibitors of RCD pathways, offering promising therapeutic avenues for PCOS. For example, apoptosis in PCOS is triggered by the suppression of the phosphoinositide 3-kinase (PI3K)/protein kinase B (Akt) signaling pathway, while growth hormone has been shown to counteract this effect by reactivating the pathway²³. Recent findings have also highlighted the importance of autophagy in PCOS development²⁴particularly the role of proteins such as sirtuin 1 (SIRT1), which regulate both autophagy and circadian rhythms²⁵. Additionally, RCD contributes to inflammation through pathways like receptor-interacting serine/threonine-protein kinase 3 (RIPK3)/ mixed lineage kinase domain-like pseudokinase (MLKL) and NOD-, LRR- and pyrin domain-containing protein 3 (NLRP3)/caspase-1/ gasdermin D (GSDMD), further amplifying the inflammatory cycle in PCOS²⁶. Despite these insights, the intricate interplay between RCD and PCOS remains poorly understood, as research in this area is still emerging. Our study aims to identify key RCD-related targets and PCOS subtypes, while exploring potential therapeutic interventions. A deeper understanding of RCD mechanisms could pave the way for novel treatments for PCOS.

In this study, we conducted a comprehensive bioinformatics analysis to pinpoint differentially expressed genes (DEGs) in both normal ovarian tissues and those affected by PCOS. Through this process, we identified 389 genes linked to RCD. By employing advanced algorithms such as the Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and Support Vector Machine (SVM), we narrowed down five key hub genes. To evaluate their diagnostic potential, we utilized Receiver Operating Characteristic (ROC) analysis, while protein-protein interaction (PPI) networks were mapped out to visualize the relationships between these hub genes. Our findings revealed that these genes are significantly enriched in pathways related to immune-inflammatory responses, adipocytokine signaling, and sex hormone regulation. Notably, our research highlights the pivotal role of epigenetic mechanisms in the development of PCOS. To delve deeper, we performed single-sample gene set enrichment analysis (ssGSEA), Gene Ontology (GO) functional enrichment, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis to uncover the biological processes tied to these hub genes. Additionally, we validated pathway correlations using Gene Set Enrichment Analysis (GSEA) and predicted upstream regulatory elements, including gene-related compounds, microRNAs (miRNAs), and transcription factors (TFs), leveraging the NetworkAnalyst and RegNetwork databases. These interactions were visualized using Cytoscape software. A step-by-step overview of our methodology is illustrated in Fig. 1. In essence, our findings offer valuable insights that could enhance the diagnosis and therapeutic strategies for PCOS.

Fig. 1.

Study framework.

Materials and methods

Obtainment and preprocessing of PCOS and RCD datasets

Gene expression profiles for both PCOS patients and healthy controls were retrieved from the Gene Expression Omnibus (GEO) database, available at http://www.ncbi.nlm.nih.gov/geo. Three specific datasets—GSE34526, GSE168404, and GSE193123—were selected for further analysis. These datasets focus on the transcriptomic profiles of granulosa cells from both normal and PCOS-affected individuals and have been widely referenced in prior PCOS research. Detailed information regarding these datasets is provided in Table 1. Granulosa cells were collected from ovarian aspirates of healthy controls and PCOS patients undergoing in vitro fertilization (IVF), with all samples subjected to a gonadotropin-releasing hormone antagonist protocol. RCD-related genes were sourced from the literature²⁷ and are presented in Supplementary Information 1. The authors compiled a comprehensive collection of 18 signatures of RCD by utilizing various databases (MajiDB [http://www.xjlab.com.cn/MajiDB] and FerrDB [http://www.zhounan.org/ferrdb]) and published literature^28–35. The inclusion criteria required gene sets to be: (1) experimentally validated in peer-reviewed studies; (2) associated with specific RCD modalities defined by the Nomenclature Committee on Cell Death¹⁹; (3) functionally annotated in at least two independent datasets. Exclusion criteria removed: (1) non-RCD related apoptotic genes; (2) signatures derived solely from non-mammalian models; (3) computationally predicted sets without experimental validation. After demultiplexing, a total of 7,460 genes were identified.

Table 1.

GSE datasets cited in this research.

GSE dataset	Organism	Sample number	PMID	Platform
GSE34526	granulosa cell	Control:3 PCOS:7	22,904,171	GPL570-[HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array
GSE168404	granulosa cell	Control:5 PCOS:5	34,084,179	GPL16791-Illumina HiSeq 2500 (Homo sapiens)
GSE193123	granulosa cell	Control:3 PCOS:3	35,237,237	GPL24676-Illumina NovaSeq 6000 (Homo sapiens)

Identification of the differentially expressed genes

We pooled together the datasets (GSE34526, GSE168404, GSE193123) and removed batch effects by using the R packages “limma” and “sva”. To address cross-platform heterogeneity between microarray and RNA-seq datasets, we applied a standardized integration workflow, including log2(x + 1) transformation for RNA-seq data, quantile normalization for microarray data, harmonization of gene identifiers to retain only overlapping genes, and batch effect correction using the ComBat algorithm within an empirical Bayes framework. These steps were taken to cut down on experimental errors and enable the recombination of the datasets, making sure that the downstream analysis zeroes in on just the biological differences^36–39. We created a PCA plot with the R packages “FactoMineR” and “factoextra”, which gave us a visual way to see how similar the samples were. Moreover, we used the R package “preprocessCore” to standardize the dataset. Specifically, we applied the normalize. quantiles function with default parameters to achieve quantile normalization across all samples, ensuring comparability between datasets. We also made use of the “limma” package to pinpoint the differentially expressed genes (DEGs) between the PCOS and the control groups. Then, we generated a volcano plot to spotlight the differential expression of these DEGs. An adjusted P value < 0.05 and |logFC|> 0.5 were considered to be the cutoffs for DEGs. A heatmap was generated using the R package “pheatmap” based on the DEGs that were identified.

Functional enrichment analysis

To evaluate the biological roles of the 854 differentially expressed genes (DEGs), we conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) analyses using the “clusterProfiler” R package. The GO analysis focused on identifying enriched terms across three categories: biological processes (BP), molecular functions (MF), and cellular components (CC). The findings were organized by GeneRatio, with circle size reflecting the count of enriched genes and color intensity representing the significance of the P value. Additionally, the top five pathways and their associated genes from the KEGG analysis were highlighted and visualized for further interpretation.

Identification of differentially expressed rcd‑related genes and functional enrichment analysis

RCD-related genes that were differentially expressed in the DEGs from the datasets (GSE34526, GSE168404, GSE193123) were identified using the “limma” package with P < 0.05 and |logFC| > 0.5. GO (including BP, MF, and CC) and KEGG analyses were conducted on the identified RCD-related genes, and the top five pathways and genes from the KEGG analysis were visualized.

Screening for hub genes by machine-learning algorithms

Our investigation leveraged three machine-learning approaches—random forests (RF), LASSO regression, and support vector machine-recursive feature elimination (SVM-RFE)—to pinpoint Hub RCD-associated genes that play a key role in PCOS. Using R package “randomForest”, we implemented RF analysis, while LASSO modeling was conducted via “glmnet” with optimal parameters determined by minimizing lambda. To ensure robust parameter selection, ten-fold cross-validation was applied, confirming alignment with partial likelihood deviation thresholds. Feature ranking was achieved through recursive elimination within the RF framework, prioritizing RCD-linked DEGs by importance. SVM-RFE outperformed linear discriminant analysis and mean squared error methods in streamlining feature selection, efficiently filtering noise while retaining critical variables—a process validated through repeated cross-validation. Diagnostic accuracy was evaluated using ROC curves, with AUC values computed via the “pROC” package to quantify predictive performance.

Human ovarian granulosa cell collection and ethics statement

To validate the expression of identified hub genes (WFIKKN2, LPAR3, TNFRSF1A, ACO1, ZNF521), granulosa cells were collected from an cohort of 18 women (9 PCOS patients and 9 healthy controls) undergoing in vitro fertilization (IVF) at the Affiliated Hospital of Nanjing University of Chinese Medicine. All participants provided written informed consent, and the study was approved by the Institutional Ethics Committee (No.2021NL-019-03). Patients in the PCOS group were selected in accordance with the Rotterdam diagnostic criteria⁵. The standards used to choose control women were as follows: (a) Menstrual cycles occurring at regular intervals of 25–35 days, (b) Absence of any anomalies in the endocrine system, and (c) Normal appearance of the ovaries as confirmed by ultrasound imaging.

RNA isolation and quantitative real‑time PCR (qRT‑PCR)

Total RNA was isolated using TRIzol^® Reagent (Thermo Fisher, 15596026) following the manufacturer’s protocol, including DNase I treatment (RNase-Free DNase Set, Qiagen, 79254) to eliminate genomic DNA contamination. RNA concentration/purity Measured by NanoDrop™ (Thermo Fisher). All samples had A260/280 ratios of 1.98–2.05 and A260/230 > 1.8. Integrity evaluated by Agilent 2100 Bioanalyzer with RNA Nano Chips (Agilent Technologies, 5067 − 1511). Only samples with RNA Integrity Number (RIN) ≥ 7.0 were used. Reverse Transcription:1 µg total RNA was reverse-transcribed using PrimeScript™ RT Kit (Takara, RR047A) with random hexamers and oligo(dT) primers in a 20 µL reaction. cDNA was diluted 1:5 with nuclease-free water and stored at − 20 °C. Gene-specific primers were designed using Primer-BLAST (NCBI) with the following criteria: Amplicon size: 80–150 bp. Primer efficiency: 90–110% (validated by standard curve, R² >0.99). The relative expression level was calculated by 2^−∆∆Ct method and three replicate experiments were involved. All specific primers were shown in Table 2.

Table 2.

Primers used for real-time PCR.

Gene	Sequence (5’→3’)
GAPDH	Forward: GAAGGTGAAGGTCGGAGTC
	Reverse: GAAGATGGTGATGGGATTTC
TNFRSF1A	Forward: CTCTGCCGCTTCATCATCAC
	Reverse: GATGTCGTAGCCGGTGTAGA
WFIKKN2	Forward: CAGGTGCTGGACTTCATGGT
	Reverse: TGGTAGCCACAGTGAGGACA
LPAR3	Forward: TGGGCTACATCCTGGTCATC
	Reverse: CAGGGTCCAGGTAGAGGTCA
ACO1	Forward: AGCCTGGAGAAAGTGCTGAC
	Reverse: TCCAGGTCCAGGTAGAGGTG
ZNF521	Forward: GCTGGAGACCTGAAGCTGAT
	Reverse: TGCTGTTGCTGAGGTAGAGG

Protein–protein interaction (PPI) network analysis

PPI networks were generated by inputting gene datasets into genemania database (https://genemania.org/), with automatically selected weighting method being used for network construction. Organism is Homo sapiens (human), while the Application version is 3.6.0. Network nodes are proteins, and lines represent predicted relationships.

Single-sample gene set enrichment analysis

Single-sample gene set enrichment analysis (ssGSEA) is an advanced method of analyzing gene set enrichment that can be utilized for analyzing individual samples. This algorithm was employed to investigate pathway enrichment in samples from patients with PCOS compared to control groups (GSE34526, GSE168404, GSE193123), using the KEGG pathway dataset (c2.cp.kegg.v2023.1.Hs.symbols.gmt). The R package “ggplot2” facilitated the quantification of enrichment for pathways associated with thrombosis, glycolipid metabolism, inflammatory responses, oxidative stress, and hormonal variations. To evaluate the relationship between the expression of key genes and the scores of KEGG pathways, Spearman’s correlation analysis was conducted. A significance level of P < 0.05 was established to determine statistical relevance.

Correlation analysis and gene set enrichment analysis

We assessed how the five RCD-linked DEGs co-expressed across all specimens and within PCOS cases specifically. Pearson’s correlation tests were run with the cor.test tool from R package “stats”, and heatmaps generated via “pheatmap” illustrated these relationships. To better highlight expression trends in key biological pathways, GSEA was implemented in R. Leveraging the “clusterProfiler” and “enrichplot” packages, we plotted the 20 most statistically prominent Reactome pathways identified through GSEA. An adjusted P value threshold of < 0.05 determined significance.

Predication of drug‑gene interaction and miRNA-TF regulatory network

NetworkAnalyst (https://www.networkanalyst.ca/) is a sophisticated web-based tool designed for in-depth gene expression analysis, leveraging network-driven visual analytics. It also incorporates drug-gene interaction data from multiple sources, making it a valuable resource for identifying existing drugs or small-molecule compounds associated with key hub genes. To improve the reliability of predictions, both NetworkAnalyst and the RegNetwork database (https://regnetworkweb.org/) were utilized to forecast miRNAs and transcription factors (TFs) linked to the hub genes, applying a p-value cutoff of 0.05 for filtering. The resulting miRNA-TF regulatory networks were then mapped out and visualized using Cytoscape (v3.10.1; https://cytoscape.org).

Statistical analysis

Statistical analyses were conducted in R version 4.2.2. Group comparisons were assessed using either the Wilcoxon rank-sum test or Student’s t-test, depending on data distribution. Relationships between variables were evaluated via Pearson’s correlation coefficient or Spearman’s rank-order correlation, chosen based on the nature of the data. Two-tailed hypothesis testing was applied throughout the study, with results achieving p-values below 0.05 considered statistically significant.

Results

Screening of DEGs in PCOS and functional analysis

We integrated data from three datasets—GSE34526, GSE168404, and GSE193123—yielding a combined total of 26 samples, encompassing both PCOS and control groups. Principal Component Analysis (PCA) plots were employed to visually highlight the similarities among these samples (Fig. 2A, B). To mitigate potential bias introduced by individual sample variations, we normalized the dataset using the R package “preprocessCore.” The comparative results of the data before and after normalization are illustrated in Fig. 2C, D. Utilizing the “limma” package in R, we identified 246 up-regulated and 608 down-regulated genes, applying thresholds of an adjusted P value < 0.05 and |logFC| > 0.5 to define DEGs. These findings were visually represented through volcano plots and heatmaps (Fig. 2E, F). To explore the biological implications of these DEGs, GO and KEGG analyses were performed using the “clusterProfiler” package. The analytical results and the key biological implications of these DEGs are illustrated in Fig. 3 and summarized in Table 3.

Fig. 2.

Gene expression data of PCOS patients and normal women from GSE34526, GSE168404, GSE193123. (A) PCA plot shows the data before merging. (B) PCA plot shows the data after merging, total of 26 samples. (C) Data before homogenization. (D) Data after homogenization using R package “preprocessCore”. (E) A volcano plot of the 854 DEGs in PCOS and control. Upregulated: red, downregulated: blue. (F) An expression heatmap corresponding to the DEGs.

Fig. 3.

Functional annotation and pathway enrichment analysis of DEGs. (A) GO enrichment analysis showing top BP and CC. (B) KEGG pathway enrichment results ranked by GeneRatio. (C) Chord diagram linking the top five KEGG pathways with their associated genes^40–42. Circle size represents gene counts; color gradients indicate adjusted P-values.

Table 3.

Functional enrichment analysis of DEGs.

Description	P.adjust	geneID(Top5)	Count
A GO-BP enrichment results
Phospholipid metabolic process	0.001049938	Agpat1/dgka/socs3/pnpla6/pitpnm1	39
Phosphatidic acid biosynthetic process	0.001049938	Agpat1/dgka/dgkz/nr1h4/pnpla3	10
Secondary alcohol metabolic process	0.001049938	Cyp27a1/stard3/nr1h4/vldlr/apoc1	21
Phospholipid biosynthetic process	0.001049938	AGPAT1/DGKA/SOCS3/PITPNM1/DGKZ	29
Cholesterol metabolic process	0.001049938	CYP27A1/STARD3/NR1H4/VLDLR/APOC1	20
B GO-CC enrichment results
Apical part of cell	0.024679333	DCHS1/CASR/SPTBN5/C2CD2L/SLC12A2	36
Apical junction complex	0.025136855	CYTH2/ARHGEF2/CYTH1/VASP/JUP	17
Cell projection membrane	0.025136855	ARHGEF2/INPP5K/EPB41L3/VASP/FSCN1	29
Basement membrane	0.025136855	COL4A3/EGFLAM/CASK/VWA1/COL8A2	13
Apical plasma membrane	0.059262797	CASR/C2CD2L/SLC12A2/CLDN1/PODXL	29
C KEGG pathway enrichment results.
Cortisol synthesis and secretion	0.022727883	NR4A1/ITPR1/PDE8A/MRAP/STAR	11
Steroid biosynthesis	0.022727883	EBP/LSS/NSDHL/DHCR24/SQLE	6
Phospholipase D signaling pathway	0.022727883	CYTH2/AGPAT1/DGKA/GAB2/CYTH1	17
Glycerolipid metabolism	0.022727883	AGPAT1/DGKA/DGKZ/PNPLA3/PNLIP	10
PPAR signaling pathway	0.028068991	CYP27A1/OLR1/FABP3/SCD/FABP2	11

Identification of RCD-related genes in PCOS and functional analysis

A total of 7,460 genes associated with RCD were identified from existing literature²⁷which has has already been shown in the previous section. These RCD-related genes were cross-referenced with the DEGs, as shown in Fig. 4. We identified 147 up-regulated and 242 down-regulated RCD-related DEGs (Fig. 4A, B). GO analysis was performed on these RCD-related DEGs across biological process (BP), cellular component (CC), and molecular function (MF) categories (Fig. 4C-E). KEGG analysis was also performed on the RCD-related DEGs (Fig. 4F), and the top 5 pathways were visualized as shown in Fig. G. The key biological implications of RCD-related DEGs are summarized in Table 4.

Fig. 4.

Identification and functional enrichment of RCD-related genes. (A) Overlap of upregulated DEGs with RCD-related genes from published datasets²⁷. (B) Overlap of downregulated DEGs with RCD-related genes. (C–E) GO enrichment analysis showing top BP, CC, and MF. (F) KEGG pathway enrichment results. (G) Chord diagram linking the top five KEGG pathways to the associated genes, with color codes denoting pathway categories^40–42.

Table 4.

Functional enrichment analysis of RCD-related DEGs.

Description	P.adjust	geneID(Top5)	Count
A GO-BP enrichment results
Regulation of cellular component size	0.000109952	SEMA4A/PPP1R15A/SRF/LIMK1/SPTBN5	26
Response to mechanical stimulus	0.002577331	TNFRSF1A/BTG2/CNN2/IRF1/RELA	17
Response to oxidative stress	0.007136885	GPX3/HMOX1/IL10/KLF2/PLK3	24
Regulation of extent of cell growth	0.007136885	SEMA4A/SRF/LIMK1/ANAPC2/ULK1	11
Extrinsic apoptotic signaling pathway	0.007136885	TNFRSF1A/HMOX1/ARHGEF2/RELA/ICAM1	16
B GO-CC enrichment results
Cell projection membrane	0.00614065	ARHGEF2/INPP5K/VASP/SYTL1/SLC12A2	19
Mitochondrial outer membrane	0.00614065	HMOX1/PPP1R15A/ULK1/CYB5A/BNIP3	14
Organelle outer membrane	0.00614065	HMOX1/PPP1R15A/ULK1/CYB5A/BNIP3	15
Outer membrane	0.00614065	HMOX1/PPP1R15A/ULK1/CYB5A/BNIP3	15
Endoplasmic reticulum lumen	0.032032232	C3/VCAN/IL23A/THBS1/SERPINA1	16
C GO-MF enrichment results
Cytokine binding	0.161131228	TNFRSF1A/ZFP36/THBS1/WFIKKN2/TRIM16	10
Calmodulin binding	0.161131228	CNN2/DAPK1/MYH11/CASK/MAP2	12
Iron ion transmembrane transporter activity	0.161131228	MCOLN1/TTYH1/SLC39A14	3
G-protein beta/gamma-subunit complex binding	0.161131228	GNAI2/CETN2/CETN3/GNAZ	4
Phosphoric ester hydrolase activity	0.161131228	HMOX1/CTDSP1/INPP5K/PPP1CA/CTDP1	17
D KEGG pathway enrichment results.
C-type lectin receptor signaling pathway	0.009559102	IL10/IRF1/PLK3/RELA/IL23A	11
Phospholipase D signaling pathway	0.009559102	CYTH2/AGPAT1/DGKA/GAB2/DGKZ	13
Pertussis	0.009559102	C3/IL10/IRF1/RELA/IL23A	9
TNF signaling pathway	0.026073796	TNFRSF1A/IRF1/RELA/SOCS3/ICAM1	10
Legionellosis	0.026073796	C3/RELA/NFKB2/CD14/CXCL2	7

Identification of hub genes through machine learning and testing of diagnostic efficacy

In our previous analysis, we utilized three distinct algorithms to identify hub genes from the DEGs. For the LASSO algorithm, we determined the optimal lambda value via a ten-fold cross-validation method. Consequently, we chose the minimal lambda value to build the LASSO classifier, as it demonstrated superior accuracy compared to other options, ultimately revealing seven key genes: TNFRSF1A, WFIKKN2, LPAR3, ACO1, GABBR2, ZNF521, and DECR1 (Fig. 5A). The random forest algorithm ranked 50 significant genes according to their importance, with the top 15 highlighted in Fig. 5B. These included genes like OSBPL10, HS2ST1, SLC12A2, and others. Meanwhile, in the SVM-RFE algorithm, we minimized error by selecting the optimal number of features, which included genes such as TRIM16, SLC12A2, RCAN1, and more (Fig. 5C). After comparing the results, we identified five hub genes that were consistently recognized across the LASSO, random forest, and SVM-RFE analyses: WFIKKN2, LPAR3, TNFRSF1A, ACO1, and ZNF521 (Fig. 5D). To further explore the interactions between these hub genes, we built a protein-protein interaction (PPI) network with GeneMANIA (https://genemania.org/), which revealed 20 nodes (refer to Supplementary Information 2). Notable expression differences were observed for the five hub genes when comparing PCOS patients with the control group. Specifically, TNFRSF1A was found to be up-regulated, while WFIKKN2, LPAR3, ACO1, and ZNF521 were down-regulated in the PCOS cohort (Fig. 6A). Besides, qPCR analysis confirmed significant dysregulation of all five hub genes in PCOS granulosa cells compared to controls: TNFRSF1A was up-regulated (P < 0.001) while WFIKKN2, LPAR3, ACO1, and ZNF521 were down-regulated (P < 0.01) (Fig. 6B). These results align with our bioinformatic predictions and underscore the relevance of RCD-related genes in PCOS pathogenesis. These findings suggest that these genes may play a significant role in the development of PCOS. We further assessed the diagnostic efficacy of each hub gene in PCOS diagnosis across the merged cohorts of GSE34526, GSE168404, and GSE193123. The area under the curve (AUC) values from the ROC analyses for the hub genes were as follows: 0.958 for WFIKKN2, 0.933 for LPAR3, 0.897 for TNFRSF1A, 0.933 for ACO1, and 0.897 for ZNF521, suggesting that these genes are promising indicators for predicting the progression of PCOS (Fig. 5E).

Fig. 5.

Machine learning–based identification of hub genes from DEGs. (A) Seven genes selected by LASSO regression. (B) Top 15 genes ranked by importance from RF analysis. (C) Genes selected by SVM-RFE. (D) Five hub genes identified by intersecting the three algorithms. (E) ROC curves of the five hub genes for predicting disease occurrence, with the x-axis representing 1–specificity and the y-axis representing sensitivity.

Fig. 6.

Differences in the expression of five hub genes between PCOS and controls. (A) Expression level of ACO1, LPAR3, TNFRSF1A, WFIKKN2 and ZNF521 in the datasets (GSE34526, GSE168404, GSE193123). ***P < 0.001. (B) Relative mRNA expression of the hub genes by qRT-PCR. P values were shown as: **P < 0.01; ***P < 0.001, n = 3.

SsGSEA analysis and correlation between hub genes and pathway scores

Utilizing the KEGG datasets, we applied the ssGSEA algorithm to investigate the pathway enrichment of DEGs in both PCOS and control groups. To effectively visualize and assess the enrichment of these genes across various metabolic pathways, we employed the ggplot2 package in R. Our analysis revealed that the DEGs in the PCOS group were particularly enriched in metabolism-related pathways, with significant results (P < 0.001) observed in the degradation of valine, leucine, and isoleucine, as well as expression related to peroxisome function, propanoate metabolism, and the biosynthesis of unsaturated fatty acids (Fig. 7). Moreover, we conducted a Spearman correlation analysis to explore the relationship between the expression levels of hub genes and the KEGG pathway scores. As illustrated in Fig. 8, this analysis further corroborated the association between the core genes and the metabolism-related pathways.

Fig. 7.

Box plot showing differences in KEGG pathway scores assessed by ssGSEA algorithm between PCOS and control groups. **: P < 0.01, ***: P < 0.001.

Fig. 8.

Correlation between 5 hub genes and pathway scores (only P < 0.05 is shown). Circle size indicates correlation coefficient size and color indicates P-value.

Correlation analysis and biological significance underlying hub genes

We employed GSEA method to delve into the functional significance of key hub genes. To kick off, we conducted a correlation analysis between the hub genes and the entire gene set (PCOS and control), and then illustrated the expression patterns of the top 50 genes that showed a positive correlation through a heatmap (Fig. 9). Following this, we implemented GSEA to assess the signaling pathways linked to these hub genes (Fig. 10). The results from the GSEA highlighted that the elevated expression levels of ACO1 and ZNF521 are chiefly involved in pathways related to lipid metabolism, such as cholesterol biosynthesis and fatty acid metabolism. Furthermore, the ZNF521 gene showed a notable enrichment in the metabolic pathway of sex hormones, which also showed a connection to pathways associated with lipid metabolism. We also observed significant enrichment in immune and inflammatory pathways, including antigen processing and cross-presentation, the interferon-gamma signaling pathway, signaling by interleukins, cilium assembly, and the TLR7/8 cascade, linked to genes like LPAR3, TNFRSF1A, and WFIKKN2. These observations imply that women with PCOS may undergo biological changes tied to inflammation, immune dysfunction, and lipid metabolism, as suggested by the behavior of these pivotal hub genes.

Fig. 9.

Correlation analysis of 5 hub genes with entire gene set (PCOS and control), heatmap shows positive correlation of top50 gene expression.

Fig. 10.

Top pathways identified by GSEA. The twenty most significantly enriched pathways are displayed, with enrichment scores indicating the direction of association: positive values denote a direct relationship between gene activity and biological processes, while negative values indicate an opposing trend.

Construction of the miRNA-TF regulatory network and predication of drug‑gene interaction

The hub genes emerged as promising therapeutic targets for addressing PCOS. By leveraging the NetworkAnalyst database (https://www.networkanalyst.ca/), a drug–gene interaction analysis pinpointed potential drugs and compounds that could be effective in treating PCOS. The top 25 candidates are illustrated in Fig. 11A. Furthermore, the same database was employed to identify upstream miRNAs and transcription factors (TFs) associated with the hub genes, as depicted in Fig. 11B. To validate these findings, the RegNetwork database (https://regnetworkweb.org/) was utilized to predict miRNAs and TFs for the hub genes, and the resulting regulatory network was mapped using Cytoscape (v3.10.1; https://cytoscape.org). This comprehensive network included 42 miRNAs, 30 transcription factors, and 5 hub genes, corroborating earlier predictions (Fig. 12). Of particular interest, USF1 and VSX2 were found to interact with both LPAR3 and WFIKKN2, while PPARG showed connections to ZNF521 and WFIKKN2.

Fig. 11.

A Predicting drug–gene interaction results (NetworkAnalyst database). B Predicting miRNA and TF upstream of hub genes (NetworkAnalyst database).

Fig. 12.

Predicting miRNA and TF upstream of hub genes (regnetwork database).

Discussion

Regulated cell death (RCD) refers to a highly coordinated process enabling cells to undergo self-destruction in a controlled manner¹⁹which is essential for various physiological functions, including development, tissue homeostasis, and immune regulation. RCD manifests in several forms, such as apoptosis, necrosis, autophagy, and pyroptosis. Increasing evidence suggests a link between PCOS and various RCD pathways^24,43,44although a comprehensive exploration of RCD’s role in PCOS remains limited. This study illuminates the potential relationship between RCD and PCOS by identifying five key biomarkers associated with RCD in PCOS using advanced machine learning techniques. Moreover, we mapped the relevant pathways and miRNA-TF regulatory networks tied to these central genes, contributing to a more nuanced understanding of the underlying mechanisms of PCOS.

We initially performed Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and single-sample Gene Set Enrichment Analysis (ssGSEA) on the DEGs identified from the GSE dataset. These analyses suggest that the DEGs are involved in lipid processing and steroid hormone production, which aligns with findings from previous studies. Disruptions in fat metabolism are a well-established feature of PCOS, which exacerbates the condition’s metabolic imbalances. Following the National Cholesterol Education Program (NCEP) guidelines, approximately 70% of PCOS patients exhibit abnormal blood lipid profiles⁴⁵. Cholesterol, an essential molecule for cell growth and steroid hormone synthesis, plays a critical role in ovarian hormone activity and follicle maturation⁴⁶. Additionally, lipid metabolism disruptions are closely linked to insulin resistance⁴⁵a hallmark symptom of PCOS. Notably, recent studies have identified PCSK9, a protein associated with both lipid regulation and ovarian dysfunction, as a promising therapeutic target. Furthermore, drugs such as Alirocumab have shown promise in improving fat metabolism, rebalancing reproductive hormones, and reversing ovarian changes in PCOS mouse models⁴⁶. Recent work also connects messed-up phospholipid metabolism to PCOS⁴⁷though the “how” remains murky. Our analysis points to potential involvement of the Phospholipase D signaling pathway.

Hormonal dysregulation, particularly elevated DHEA levels, is a common feature of PCOS, often indicative of excess androgens. Elevated testosterone in PCOS patients may arise from overactive aromatase enzymes, which convert androgens into estrogens⁴⁸. Additionally, the hypothalamic-pituitary-adrenal (HPA) axis may modulate adrenal hormone production, contributing to the development of PCOS⁴⁹. Interestingly, cortisol-blocking treatments have shown more significant improvements than standard hormone therapies, particularly in adolescents with PCOS⁵⁰.

In contrast to traditional screening methods, the integration of biological analysis with machine learning approaches enabled us to pinpoint five core differentially expressed genes (DEGs)—WFIKKN2, LPAR3, TNFRSF1A, ACO1, and ZNF521—that differentiate PCOS patients from healthy controls. These hub genes were identified from an extensive pool of 18 RCD-related genes implicated in various cell death mechanisms such as alkaliptosis, anoikis, apoptosis, and pyroptosis. This finding underscores the growing complexity and scale of biological data available for analysis⁵¹. Our in-depth functional analysis revealed that the five hub genes (WFIKKN2, LPAR3, TNFRSF1A, ACO1, ZNF521) are associated with immune signaling, inflammation, and metabolic pathways, including glycolipid and steroid metabolism. One crucial player in these metabolic processes is the SREBP-1 pathway, which regulates lipid and glucose homeostasis^52–54. Hyperactivation of SREBP-1 not only contributes to metabolic disturbances in PCOS but may also increase the risk of endometrial cancer by fueling aberrant fat metabolism in uterine tissue^55,56. This positions SREBP-1 as a central figure in the pathogenesis of PCOS, potentially exceeding its previously understood role in metabolic regulation.

While inflammation’s role in PCOS is well-documented^7,9our study reveals a deeper connection. For example, alpha-linolenic acid (ALA), a precursor to omega-3 fatty acids, exacerbates inflammation in ovarian granulosa cells by activating the nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) pathway. This activation leads to phosphorylation of RelA/p65 (p65) and inhibitor of kappa B alpha (IκBα), facilitating the nuclear translocation of the NF-κB p50-p65 complex, which subsequently drives the transcription of inflammatory genes^57,58. Similarly, the interferon-gamma (IFN-γ) pathway, mediated by overactive toll-like receptor 8 (TLR8) in granulosa cells, promotes the production of pro-inflammatory cytokines such as IFN-γ, tumor necrosis factor-alpha (TNF-α), and interleukin-12 (IL-12). The interaction between microRNA-21 (miR-21) and TLR8 forms a feedback loop that modulates inflammation and influences granulosa cell survival in PCOS⁵⁹.

Interestingly, our study highlights the immune system’s significant role in PCOS, surpassing even inflammation in importance. Moreover, bioinformatics analyses revealed significant enrichment of genes related to neutrophil activation and degranulation in PCOS, although experimental validation is still lacking⁶⁰. The interplay between immune cells, particularly macrophages and neutron-phils, and adipose tissue may perpetuate chronic low-grade inflammation, further compounding the condition⁶¹. Despite these advances, further research is essential to fully unravel the complex interplay between PCOS, inflammation, and immune dysregulation.

Recent genetic studies have revealed how DNA diversity contributes to the development of PCOS, categorizing the condition into distinct reproductive and metabolic subtypes. Studies in animal models suggest that prenatal environmental exposures may induce epigenetic modifications, potentially transmitting PCOS traits across generations⁶². These findings indicate that the origins of PCOS may lie in fetal development, where disruptions in egg precursor cell formation may predispose individuals to the condition later in life. Moreover, DNA methylation plays a crucial role in regulating genes associated with PCOS, influencing gene expression without altering the DNA sequence. In mouse models of PCOS, abnormalities in the transcriptome, neuroendocrine system, and metabolism were reversed using S-adenosylmethionine (SAM), a methyl donor, highlighting the potential for epigenetic therapies⁶³. Maternal androgen exposure has also been identified as a key driver of transgenerational androgen effects, suggesting that both intrauterine conditions and germ cells contribute to the inheritance and progression of PCOS⁶⁴. Bioinformatics analyses have further implicated several biological processes, including glycosylphosphatidylinositol (GPI) anchor biosynthesis, basal transcription factors, mRNA splicing, and eukaryotic translation elongation, in the mechanisms underlying PCOS. However, the validity of these findings hinges on more extensive, long-term studies to fully understand the transgenerational impact of PCOS.

The clinical value of the identified biomarkers and their combined predictive model was assessed using ROC curve analysis and nomogram modeling, both of which suggest that the five core genes hold substantial clinical promise for diagnosing PCOS. Notably, WFIKKN2 emerged as the most accurate biomarker for diagnostic purposes.

Additionally, we explored drug-gene interactions and identified 25 promising drugs or compounds with potential applications in treating PCOS. Among these, tretinoin, valproic acid, and benzophenone stood out due to their interactions with multiple key genes, suggesting their therapeutic potential. The miRNA-TF regulatory network, which includes 42 miRNAs, 30 transcription factors, and the five hub genes, was constructed using Cytoscape. One key regulatory element, peroxisome proliferator-activated receptor γ (PPARγ), encoded by the PPARG gene, plays a pivotal role in regulating adipocyte differentiation, glycolipid metabolism, and inflammation. The rs1801282 C > G polymorphism in the PPARG gene, particularly the Pro12Ala variant, has been linked to altered PPARG activity and may lower the risk of PCOS, especially among Caucasian populations⁶⁵. Epigenetic regulation, including DNA methylation and histone modifications, further influences PPARG gene expression, highlighting the intricate interplay between genetics and epigenetics in PCOS development⁶⁶.

Together, our findings further underscore the potential involvement of several key genes in the pathophysiological processes underlying PCOS. TNFRSF1A, a receptor mediating TNF-α signaling, has been implicated in NF-κB–driven inflammatory responses⁶⁷which aligns with the well-documented chronic low-grade inflammation observed in PCOS. ZNF521 and ACO1 may linked to lipid metabolism regulation via the SREBP-1 pathway^68,69providing a plausible molecular basis for the dyslipidemia and metabolic disturbances frequently reported in PCOS patients. In addition, LPAR3, a lysophosphatidic acid receptor, has been shown to contribute to ovarian fibrosis⁷⁰a pathological feature that may exacerbate ovarian dysfunction in PCOS. These observations suggest that inflammatory signaling, altered lipid metabolism, and tissue remodeling may converge through these genes to drive PCOS progression, offering novel targets for future mechanistic studies and therapeutic interventions.

This study is limited by the use of a relatively small sample derived from publicly available datasets, and the validation measures employed were somewhat restricted. To validate our findings, an independent, prospective cohort study with a larger sample and comprehensive clinical data is necessary. Moreover, future studies should incorporate in vitro and in vivo approaches to further elucidate the complex mechanisms underpinning PCOS.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Rongkui Hu spearheaded the conceptualization and design of the study, meticulously revised the manuscript, and granted final approval for its publication. Ronghuang Li and Qianyu Chen took charge of data analysis and played a pivotal role in drafting the manuscript. Yuehua Yan and Yang Yang contributed by analyzing the data and interpreting the findings. Every author thoroughly reviewed and endorsed the final version of the paper.

Funding

This work is supported by Jiangsu Province 333 Project Project ([2022]3-25-049), Key program of Administration of Traditional Chinese Medicine of Jiangsu Province, China (ZX202102).

Data availability

The RNA expression dataset is publicly accessible through the NCBI GEO repository at https://www.ncbi.nlm.nih.gov/geo/. All findings from this research have been incorporated into the manuscript and its supplementary materials. Additional data can be obtained by contacting the corresponding author with a formal request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

This study was carefully reviewed and approved by the Ethics Committee of Affiliated Hospital of Nanjing University of Chinese Medicine (No.2021NL-019-03). All procedures performed in studies were in accordance with the ethical standards of Ethics Committee of Affiliated Hospital of Nanjing University of Chinese Medicine and the Declaration of Helsinki.

Consent to publish

Ethical approval not required. Publicly accessible anonymized data from a secondary source repository was utilized. No direct interaction with individuals occurred.