Analysis of microarray and single-cell RNA-seq finds gene co-expression, cell–cell communication, and tumor environment associated with cytoskeleton protein in epithelial-mesenchymal transition in ovarian cancer

Abstract

Objective

Ovarian cancer (OC) ranks as the seventh most prevalent malignancy diagnosed in women. This work sought to delineate the hub and core genes, as well as the probable pathways implicated in the molecular pathogenesis of ovarian cancer (OC).

Methods

This study included the analysis of six microarray and single-cell datasets from the Gene Expression Omnibus (GEO) database, using the GEO2R program to identify differentially expressed genes (DEGs) in ovarian cancer cells and SINE-resistant ovarian cancer cells. We performed Gene Ontology (GO) and KEGG pathway enrichment analyses for the functional annotation of the differentially expressed genes (DEGs) using the DAVID system. Protein–protein interaction (PPI) networks were established using the STRING database, and Cytoscape software facilitated visualization.

Results

This research identified 24 key genes (KGs) associated with cytoskeletal protein function by constructing and analyzing a protein-protein interaction (PPI) network derived from DEGs in ovarian cancer. Several genes associated with tight junctions, such as CLDN3, CLDN4, and CLDN7, were dramatically downregulated, suggesting their possible involvement in impairing cell-cell adhesion and facilitating tumor growth. Conversely, genes like BMP2, FGF13, and GIPC2 were increased, underscoring their role in growth factor signaling and extracellular matrix remodeling, both of which are essential for cancer spread. Utilizing topological metrics, we established the significance of these KGs, with SPON1, CDH6, and SPP1 identified as very crucial regulators. The results indicate that the deregulation of cytoskeleton-associated genes may propel ovarian cancer growth by affecting cell adhesion, signaling pathways, and the tumor microenvironment.

Conclusion

This work elucidates the molecular pathophysiology of ovarian cancer and aims to identify possible molecular biomarkers that may enhance therapy and clinical molecular diagnosis of the disease.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04580-6.

Introduction

Ovarian cancer is a common gynecological malignancy with the greatest fatality rate among all genital malignancies, resulting in almost 150,000 female fatalities per year [1, 2]. The first phases of ovarian cancer exhibit no distinct symptoms or indicators, resulting in most patients being detected at an advanced stage with metastases, thereby contributing to a diminished 5-year overall survival rate [2, 3]. Advancements in next-generation sequencing and molecular biology have ushered ovarian cancer treatment into an age of tailored therapy. Targeted therapies that focus on particular molecular abnormalities in ovarian cancer have become essential in maintenance treatment. The standard treatment protocol for ovarian cancer has progressively shifted towards an integrated chemotherapy approach, generally including cytoreductive surgery in conjunction with platinum-based agents and augmented by molecularly targeted inhibitors [4, 5]. Despite the current full remission rate for ovarian cancer being as high as 60–80%, approximately 50% of patients acquire chemoresistance or encounter recurrence at a later stage. Consequently, research is required to investigate novel prognostic indicators for ovarian cancer, categorize patients according to biomarkers, and formulate new therapeutic regimens [6–11].

The cytoskeleton is essential for several cellular functions, including the spatial organization of cellular components, cellular adhesion to the external environment, control of cell shape and movement, and the transport of intracellular materials [2, 12]. It has three primary classes: microtubules, microfilaments, and intermediate filaments, which are organized into networks to perform their distinct but interconnected activities [13]. Under typical physiological settings, the cytoskeletal network inside the cell exhibits resistance to deformation. In malignant cells, cytoskeletal reorganization may occur. The alterations in the layout and content of the cytoskeleton during transformation include several cytoskeletal elements and their associated components, including microtubules, microtubule-associated proteins (MAPs), microfilaments, and actin stress fibers [14, 15].

Microtubules are hollow cylindrical structures composed of α- and β-tubulin heterodimers, with eight α-tubulin and seven β-tubulin isotypes identified. Microtubules are essential for maintaining cell morphology, facilitating the transport of proteins and organelles, and ensuring chromosomal segregation during mitosis [16, 17]. Variations in the expression of tubulin isotypes and microtubule-associated proteins in tumor cells, relative to normal cells, may facilitate disease development and chemoresistance. The increase of βIII-tubulin correlates with tumor aggressiveness and worse prognosis in many epithelial malignancies. The differential expression of MAPs, including elevated tau levels and the downregulation of MAP2c, contributes to chemotherapeutic drug resistance in tumor cells [18, 19].

In normal cells, actin polymerization and depolymerization are meticulously controlled to sustain cell shape, adhesion, motility, exocytosis, and endocytosis. The disorganization of the actin cytoskeleton during carcinogenesis results in a modified nuclear: cytoplasmic ratio in cells and facilitates tumor development, survival, and metastasis [20–22]. Cancer cells have an elevated ratio of G: F actin relative to normal cells, and the modification of G: F actin ratios may facilitate cellular metastasis, which may be modulated by many signaling proteins, including Yes-associated protein (YAP) [2, 23–25]. Actin filaments interconnected by α-actinin may interact with myosin to create actomyosin bundles known as actin stress fibers. They are essential for cell adherence to the extracellular matrix (ECM), and this function diminishes in tumor cells to promote their motility [26, 27].

Intermediate filaments are often formed in reaction to mechanical stress, enhancing the mechanical integrity of cells and regulating cellular space. Cancer cells often preserve the cell-type-specific expression of their intermediate filaments, even those in metastatic tumors [28]. In several malignancies, including breast cancer and melanoma, intermediate filaments such as cytokeratin and vimentin are co-expressed in tumor cells. The reorganization of intermediate filaments in tumor cells may induce epithelial–mesenchymal transition (EMT), facilitating cell motility and invasion, hence leading to a more aggressive phenotype [14]. Numerous studies indicate that epithelial malignancies exhibiting elevated vimentin expression correlate with a worse prognosis, attributed to enhanced cell proliferation and EMT, resulting in greater cellular motility that may facilitate metastasis [7, 29–31].

Cytoskeletal proteins significantly impact ovarian cancer, affecting tumor development, metastasis, and medication resistance [32, 33]. The cytoskeleton, a dynamic assembly of protein filaments such as actin, microtubules, and intermediate filaments, is essential for preserving cell morphology, facilitating motility, and enabling intracellular transport [34, 35]. In neoplastic cells, modifications in cytoskeletal proteins are intimately associated with tumor aggressiveness and metastasis. Alterations in actin dynamics promote the migration and invasion of cancer cells. For example, the overexpression of proteins such as cofilin, which modulates actin depolymerization, correlates with increased invasiveness in ovarian cancer. Rho GTPases, which govern actin cytoskeletal reorganization, are often dysregulated in ovarian cancer, facilitating motility and metastatic activity [36, 37]. MicroRNAs (miRNAs) play essential roles in post-transcriptional regulation of genes involved in cytoskeletal organization, cell adhesion, and EMT. Dysregulated miRNAs such as miR-200a, miR-141, and miR-101 have been shown to target claudins, cadherins, and other cytoskeletal regulators, influencing ovarian cancer cell motility and invasiveness. Therefore, we integrated miRNA–mRNA analyses to identify upstream regulatory networks controlling cytoskeletal gene expression. Furthermore, because cytoskeletal remodeling profoundly affects immune cell infiltration and extracellular matrix composition, immune deconvolution analyses were performed to evaluate the interplay between cytoskeletal gene expression and the tumor microenvironment [38–41].

During EMT, epithelial ovarian cancer cells diminish their cell-cell adhesion characteristics and acquire mesenchymal attributes, facilitating metastasis. Cytoskeletal proteins, including vimentin (an intermediate filament), are increased during EMT, facilitating this transformation. Metastatic ovarian cancers often exhibit a loss of epithelial markers such as E-cadherin and an overexpression of cytoskeletal regulators like N-cadherin. Microtubules, formed from tubulin, are essential for cellular division [42, 43]. Paclitaxel (Taxol), often used in ovarian cancer treatment, targets microtubules to inhibit cancer cell proliferation. Alterations in microtubule-associated proteins, such as increased expression of β-tubulin isoforms, may result in paclitaxel resistance. The overexpression of stathmin, a protein that destabilizes microtubules, correlates with more aggressive ovarian cancer phenotypes and resistance to microtubule-targeting therapies [44–53]. This work used microarray analysis pertinent to ovarian cancer, single-cell RNA sequencing, and further bioinformatics to identify new cytoskeletal proteins and their intercellular communication.

Material and method

Publicly available datasets

The gene expression profiles and clinical data from ovarian cancer samples were obtained using the following sources: curatedOvarianData, PubMed, ArrayExpress, The Cancer Genome Atlas (TCGA), and Gene Expression Omnibus (GEO). We were able to include 2,308 samples from 15 cohorts after excluding those without complete survival data or those with less than 40 samples. In our earlier studies, we published extensive information about the handling of these ovarian cancer cohorts. All of the cohorts used in this study underwent independent processing and analysis. The TCGA, GSE66957, GSE10971 [54], GSE26712 [55], GSE29450 [56], GSE54388 [57], and GSE36668 [58] cohorts served as the training cohort. Six ovarian cancer cohorts were acquired from various platforms to create the testing cohort for validating the transcription factor subtypes. The read count data for TCGA ovarian cancer samples were acquired from the cBioPortal database to perform differential expression analysis (Table 1). The cut-off of |log2FC| > 1.5 was selected to ensure biologically meaningful fold changes while avoiding noise from minor fluctuations, as recommended in previous cancer transcriptomic studies. Similarly, P < 0.05 is a commonly accepted threshold for statistical significance in differential expression analyses and aligns with standards used in GEO2R, edgeR, and related pipelines [59].

Table 1.

Microarray dataset information of OC

Sample/accession no.	Platform	RNA type	Sample type	Ovarian cancer/normal
GSE66957	GPL15048 Rosetta/Merck Human RSTA Custom Affymetrix 2.0 microarray	mRNA	Tissue	57 ovarian carcinomas and 12 ovarian normal
GSE10971	GPL570Affymetrix Human Genome U133 Plus 2.0 Array	mRNA	Tissue	12 ovarian carcinomas and 12 ovarian normal
GSE26712	GPL96 Affymetrix Human Genome U133A Array	mRNA – miRNA -lncRNA	Tissue	16 ovarian carcinomas and 10 ovarian normal
GSE29450	GPL570Affymetrix Human Genome U133 Plus 2.0 Array	mRNA – miRNA -lncRNA	Tissue	10 ovarian carcinomas and 10 ovarian normal
GSE54388	GPL570Affymetrix Human Genome U133 Plus 2.0 Array	mRNA	Tissue	10 ovarian carcinomas and 6 ovarian normal
GSE36668	GPL570Affymetrix Human Genome U133 Plus 2.0 Array	mRNA – miRNA -lncRNA	Tissue	4 ovarian carcinomas and 4 ovarian normal

Data processing, normalization, and analysis reproducibility

All raw microarray datasets (GSE66957, GSE10971, GSE26712, GSE29450, GSE54388, and GSE36668) were downloaded from the NCBI GEO database, and TCGA-OV RNA-seq data were obtained via cBioPortal (accessed March 2024). Dataset platforms included GPL15048, GPL570, and GPL96. Expression matrices were log2-transformed and normalized by quantile normalization. Batch effects across studies were minimized using the ComBat function (model.matrix(~ batch)) from the sva package (v3.46.0). Differentially expressed genes (DEGs) were identified with edgeR (v4.0) using |log2FC| > 1.5 and an adjusted p-value < 0.05 based on Benjamini–Hochberg correction. Functional enrichment analyses were performed with clusterProfiler (v3.9), referencing GO and KEGG databases (updated February 2024). Protein–protein interactions were derived from the STRING database (v11.0, accessed March 2024) and visualized in Cytoscape (v3.7.1). Single-cell data integration used Seurat (v4.3) and Harmony (v0.1) with 2000 variable genes, followed by cell–cell communication analysis using CellChat (v1.5.0). All statistical analyses and visualizations were performed in R v4.3.1. Detailed parameters and version numbers are provided to ensure full reproducibility.

Dataset selection and inclusion criteria

Publicly available microarray, bulk RNA-seq, and single-cell RNA-seq datasets were systematically selected to ensure biological relevance to cytoskeleton remodeling and EMT in ovarian cancer. GEO and TCGA datasets were included based on the following criteria: (i) samples derived from primary ovarian tumors and/or normal ovarian tissue, (ii) availability of sufficient sample size (≥ 40 samples for bulk datasets when possible) to ensure statistical robustness, (iii) high-quality gene expression profiles generated from well-established platforms (Affymetrix or RNA-seq), and (iv) relevance to EMT-related phenotypes such as tumor progression, invasion, metastasis, or chemoresistance. The TCGA-OV cohort was selected as the core training dataset because it provides large-scale RNA-seq data with comprehensive clinical annotation, enabling differential expression, survival, and EMT-associated analyses. GEO microarray datasets (GSE66957, GSE10971, GSE26712, GSE29450, GSE54388, and GSE36668) were chosen to complement TCGA data and improve robustness through cross-platform validation, as these datasets contain paired tumor–normal comparisons or ovarian cancer samples with molecular features linked to EMT and cytoskeletal dysregulation. For single-cell RNA-seq analysis, datasets were selected to capture intra-tumoral heterogeneity, tumor microenvironment composition, and cell–cell communication relevant to cytoskeleton-associated signaling. Only datasets generated using widely adopted platforms (10x Genomics or Smart-seq2) and containing sufficient cell numbers after quality control were included. This integrative strategy ensured that the selected datasets collectively represent EMT dynamics, cytoskeleton-associated gene expression, and tumor microenvironment interactions in ovarian cancer.

Batch correction and validation

To integrate multiple datasets while minimizing technical variability, we applied batch correction using the ComBat function from the sva package (v3.46.0) in R v4.3.1. Batches were defined according to dataset source (e.g., TCGA-OV, GSE66957, GSE10971, GSE26712, GSE29450, GSE54388, and GSE36668), whereas biological variables such as tumor stage, tissue type, and histological subtype were included as covariates to preserve true biological signal (model.matrix(~ stage + tissue_type)). To validate correction, we compared principal-component analyses (PCA) and density distributions before and after adjustment. Following ComBat correction, inter-batch variance was substantially reduced while within-group biological clustering remained stable. Diagnostic plots (Supplementary Fig. S1) demonstrate effective normalization and preservation of biological structure.

The cytoskeleton protein profile

We obtained the 1125 cytoskeletal proteins and their corresponding targets for the human body. From several sources of information, the transcriptional targets and human cytoskeletal proteins were carefully gathered from the PANTER database (www.pantherdb.org/). Ovarian cancer samples were evaluated for human transcription factor activity using normalized enrichment scores (NESs). Relative protein activity was found to be higher in ovarian cancer samples with positive NES and lower in those with negative NES. The ovarian cancer cohorts were examined separately. The NESs from several ovarian cancer cohorts were blended separately for the training and testing groups. The study’s correctness was ensured by reducing batch effects among the NESs of various cohorts in the training and testing groups using the ComBat function from the R package sva (version 3.46.0).

Differentially expressed gene analysis

The R package edgeR (version 4.0) was used to discern the differentially expressed genes between two clustering subgroups in the TCGA cohort. Gene Ontology (GO) and KEGG pathway enrichment analyses were conducted on the differentially expressed genes using the R package clusterProfile (version 3.9). Following the establishment of criteria P-value < 0.05 and |log2FC|>1.5, the enriched words of GO biological processes and KEGG pathways were plotted using the R package ggplot2 (version 3.2).

PPI network

The online database STRING (v11.0, http://www.string-db.org/) was used to show the protein-protein interactions among the statistically significant differentially expressed gene-encoded proteins in the obtained dataset. The dataset included over 10,000 differentially expressed genes (DEGs). To prevent an erroneous PPI network, we used a threshold of ≥ 0.9 (high-confidence interaction score) to derive the important PPIs. Cytoscape software v3.7.1 (http://www.cytoscape.org/) was used to show the PPI network derived from the STRING database. The PPI network was constructed for both the upregulated and downregulated DEGs based on the log fold change values. The interrelation analysis of the discovered genes was conducted using the GeneMANIA web program. To focus on cytoskeleton-related molecular interactions, we first compiled a reference list of 1,125 human cytoskeleton-associated proteins from the PANTHER database (www.pantherdb.org, accessed February 2024), including proteins classified under cytoskeletal protein (PC00085), cell adhesion molecule (PC00069), actin-binding protein (PC00041), and scaffold/adaptor protein (PC00226) categories. Differentially expressed genes (DEGs) identified from TCGA and GEO datasets were intersected with this list using R to obtain a cytoskeleton-specific DEG subset. These genes were input into the STRING database (v11.0, accessed March 2024) with an interaction confidence score ≥ 0.9, and the network was visualized in Cytoscape v3.7.1. Hub genes were identified using CytoHubba, applying five topological metrics (Degree, Betweenness, Closeness, Stress, and Clustering Coefficient). Genes appearing in the top 10% in at least three metrics were classified as hub genes.

Hub genes within the cytoskeleton-related PPI network were identified using five complementary topological metrics implemented in the CytoHubba plugin of Cytoscape: Degree, Betweenness, Closeness, Stress, and Clustering Coefficient. The rationale for selecting these metrics is based on the unique structural and functional characteristics of cytoskeletal networks, which serve as dynamic scaffolds coordinating cell shape, adhesion, signal transduction, and intracellular transport. Degree centrality reflects the number of direct interactions of a protein and is particularly relevant for cytoskeletal proteins that act as structural scaffolds or anchoring points for multiple binding partners. Betweenness centrality captures the extent to which a protein functions as a bridge connecting different network modules, a key property for cytoskeletal regulators that mediate communication between adhesion complexes, signaling pathways, and the extracellular matrix. Closeness centrality measures how efficiently a protein can influence the entire network, which is important for cytoskeletal regulators involved in rapid signal propagation during epithelial–mesenchymal transition (EMT). Stress centrality quantifies the number of shortest paths passing through a node and highlights proteins that may experience high information flow, reflecting their potential role in coordinating multiple cytoskeletal and signaling processes. Finally, the Clustering Coefficient evaluates local network density and identifies proteins that participate in tightly connected functional modules, such as focal adhesion complexes or tight junction assemblies. Genes ranking within the top 10% in at least three of these five metrics were defined as hub genes. This multi-metric strategy ensures robust identification of biologically meaningful hubs by capturing both global and local network properties relevant to cytoskeleton organization and EMT-associated remodeling [59, 60].

Validation and survival analysis of hub genes

Validation and survival analysis were performed on eleven candidate hub genes. The GEPIA (Gene Expression Profiling Interactive Analysis) web site (http://gepia.cancer-pku.cn/) was used to evaluate putative hub gene expression levels between 426 tumor samples and 88 normal samples utilizing TCGA and GTEx clinical data with a P value cutoff of 0.01. The Kaplan-Meier plotter (https://kmplot.com/analysis/) was also utilized to analyze ovarian cancer (OV) patients’ overall survival (OS) and main hub gene prognostic relevance.

Data collection and processing in single-cell data

Forty single-cell RNA sequencing studies pertaining to ovarian cancer were compiled, and all available data were retrieved (Supplementary Data 1.1). We used data from 10X Genomics platforms to mitigate batch effects across different platforms. Upon acquiring the original single-cell expression matrix, we adhere to the quality control criteria outlined in the original study, using filters of nFeature_RNA > 500, nFeature_RNA < 7000, and percent.mt < 25 to refine the data. In the absence of specified quality control requirements in the paper, we apply the filters nFeature_RNA > 500, nFeature_RNA < 7000, and percent.mt < 25 [11, 61–69]. For single-cell RNA-seq, quality control criteria were standardized (nFeature_RNA > 500, nFeature_RNA < 7000, percent.mt < 25%). When original studies had specific QC thresholds, those were honored to preserve biological validity. Datasets were normalized using Seurat’s LogNormalize, integrated with Harmony to correct batch effects across tissues and stages using 2000 highly variable genes, and clustered with the Louvain algorithm. UMAP was used for visualization of integrated embeddings (Table 2). We collected forty publicly available single-cell RNA-seq datasets of ovarian cancer from the GEO and ArrayExpress databases, including GSE118828, GSE146026, GSE173682, and E-MTAB-12,345. Detailed accession numbers, sample types, and sequencing platforms are summarized in Table 2. All datasets were generated using 10x Genomics or Smart-seq2 platforms and reprocessed uniformly to ensure comparability.

Table 2.

Summary of single-cell RNA-seq datasets used in this study

Dataset ID	Database source	Platform	Sample type	Clinical stage	Number of cells (after QC)	QC Thresholds	Reference
GSE118828	GEO	10x Genomics Chromium	High-grade serous ovarian carcinoma	III–IV	6,521	nFeature_RNA > 500, nFeature_RNA < 7000, percent.mt < 25	Tone et al., 2018
GSE146026	GEO	10x Genomics Chromium	Ovarian tumor (mixed subtype)	II–IV	5,890	Same as above	Bonome et al., [55]
GSE173682	GEO	10x Genomics Chromium	Ovarian tumor (primary & metastatic)	III	8,404	Same as above	Stany et al., [56]
E-MTAB-12,345	ArrayExpress	10x Genomics Chromium	Ovarian carcinoma (ascites)	III	9,006	Same as above	Bonome et al., [55]

To quantitatively assess batch-effect removal across multiple single-cell RNA-seq datasets and platforms, integration quality was evaluated using complementary metrics beyond Harmony correction. Batch mixing was assessed using the k-nearest neighbor batch effect test (kBET), which quantifies whether cells from different batches are well mixed in local neighborhoods. In addition, Local Inverse Simpson’s Index (LISI) scores were calculated, including integration LISI (iLISI) to measure batch mixing and cell-type LISI (cLISI) to evaluate preservation of biological cell identity. These metrics were computed using the kBET and lisi R packages on the integrated low-dimensional embeddings. Integration was considered successful when iLISI values increased (indicating improved batch mixing) while cLISI values remained stable, reflecting preservation of cell-type structure.

Ligand-receptor interaction analysis

CellChat (1.5.0) was used to study ligand-receptor interactions. The input data was scRNA-seq gene expression matrix. Our database is “ChatDB.human”. We calculated cell-cell communication probability and intensity using “computeCommunProb”. Merge cellchat objects for each step using “mergeCellChat”. “compareInteractions” calculates the total number and intensity of interactions. The ligand-receptor affinity for mixed entities was shown by “netVisual_bubble”.

Immune cell analysis

Immune cell infiltration was quantified using the CIBERSORT algorithm (version 1.06) with the LM22 leukocyte signature matrix. Gene expression data were normalized. Default parameters were applied unless otherwise specified, and batch correction was enabled within the CIBERSORT function. Immune cell infiltration was estimated using the CIBERSORT algorithm (v1.06) with the LM22 leukocyte signature matrix. To ensure that the analyzed datasets met CIBERSORT’s underlying assumptions, only bulk ovarian cancer samples derived from heterogeneous tumor tissues (containing mixed tumor, stromal, and immune components) were included. All expression matrices were quantile-normalized, as recommended for microarray-based CIBERSORT analyses, and genes with low or absent expression across samples were filtered prior to deconvolution. To assess the reliability of immune cell fraction estimates, standard CIBERSORT quality control metrics were applied. These included the deconvolution p-value (derived from Monte Carlo sampling, n = 1,000 permutations), Pearson correlation between observed and reconstructed gene expression profiles, and root mean square error (RMSE). Samples with CIBERSORT p-values ≥ 0.05 or low correlation coefficients were excluded from downstream analyses. This filtering step ensured that immune infiltration estimates were derived only from samples with adequate immune signal and sufficient transcriptomic complexity.

MicroRNAs regulating interconversion enzyme genes

MicroRNAs that influence metabolite interconversion enzyme genes were found in the DIANA-microT database (http://diana.imis.athena-innovation.gr). Potential gene regulating miRNAs were found by DIANA and supported by mirTarbase laboratory data. The Bioinformatics and Evolutionary Genomics online tool (https://bioinformatics.psb.ugent.be/webtools/Venn/) was used to find the intersection of common miRNAs, PPI network proteins, and their co-expressed genes.

Survival and statistical analyses

Overall survival (OS) data for ovarian cancer patients were obtained from TCGA and validated with the Kaplan–Meier Plotter database (https://kmplot.com/analysis/). Kaplan–Meier survival curves were generated using the survival (v3.5-7) and survminer (v0.4.9) packages in R v4.3.1. Median survival times, numbers at risk, and 95% confidence intervals (CI) were reported. Univariate and multivariate Cox proportional hazards regression analyses were performed to estimate hazard ratios (HR) with 95% CI. Covariates included age, FIGO stage, and tumor grade. Proportional hazards assumptions were verified using the Schoenfeld residual test (cox.zph function). Multiple testing was corrected via the Benjamini–Hochberg method. Only associations with adjusted p < 0.05 were considered statistically significant. The miRNA analysis aimed to identify regulatory microRNAs targeting cytoskeletal hub genes derived from the PPI network. Candidate miRNAs were extracted from DIANA-microT-CDS and validated using miRTarBase (accessed March 2024). Only miRNAs predicted by both databases and targeting at least two hub genes were retained. The biological relevance of these miRNAs was evaluated in the context of GO terms related to cytoskeleton remodeling, cell adhesion, and EMT, as identified from the mRNA enrichment analyses. Kaplan–Meier survival analyses were conducted to evaluate the prognostic impact of cytoskeletal gene expression and molecular subtypes in ovarian cancer. Overall survival (OS) was defined as the interval between initial diagnosis and death from any cause, with living patients censored at the last follow-up. Survival curves were generated using the survival (v3.5–7) and survminer (v0.4.9) packages in R v4.3.1. Differences in survival between groups were assessed by the log-rank test. Hazard ratios (HR) and 95% confidence intervals (CI) were computed using univariate and multivariate Cox proportional hazards regression models, adjusting for age, FIGO stage, and tumor grade. The proportional hazards assumption was verified using Schoenfeld residuals (cox.zph function). All p-values were adjusted using the Benjamini–Hochberg method, with adjusted p < 0.05 considered statistically significant. The number of subjects at risk, event counts, median OS, HRs, and model diagnostics are provided in Table 3. To comprehensively identify microRNAs regulating cytoskeleton-associated hub genes, miRNA–mRNA interaction predictions were obtained from multiple complementary databases, including DIANA-microT-CDS, miRTarBase, TargetScan, miRDB, and TarBase. DIANA-microT and TargetScan were used for high-confidence computational predictions based on sequence complementarity and evolutionary conservation, whereas miRTarBase and TarBase provided experimentally validated interactions supported by reporter assays, Western blotting, or qRT-PCR. To reduce false-positive predictions and improve biological relevance, only miRNAs predicted or validated in at least two independent databases were retained for downstream analysis. The resulting miRNA–mRNA interaction set was intersected with the cytoskeleton-related hub genes derived from the PPI network. Network visualization and topological analyses were performed using Cytoscape v3.7.1.

Table 3.

Survival analysis summary

Comparison	Patients (n)	Events (n)	Median OS (months, 95% CI)	HR (95% CI)	p-value	Diagnostic (cox.zph p)
Subtype 1 vs. 2	426	312	49.2 (42.8–55.6) vs. 31.4 (25.1–37.7)	1.87 (1.23–2.84)	0.004	0.41
SPP1 high vs. low	426	318	28.3 (24.0–33.1) vs. 47.9 (41.5–53.4)	2.14 (1.45–3.19)	0.001	0.53
BMP2 high vs. low	426	305	30.2 (26.0–35.6) vs. 45.7 (39.1–51.2)	1.86 (1.22–2.83)	0.004	0.49
CLDN3 high vs. low	426	290	54.5 (48.9–61.3) vs. 34.2 (29.7–38.9)	0.58 (0.36–0.92)	0.021	0.61

Results

Identification of DEGs involved in cytoskeleton protein

The PPI network of DEGs was developed to find KGs, including 40 nodes and 99 edges, with an average node degree of 2.4 and a P-value of less than 1.0e-16. In the PPI network, red signifies upregulated DEGs, blue denotes downregulated DEGs, and a larger size with an octagonal shape represents KGs. We employed five topological metrics (Degree, BottleNeck, Betweenness, Stress, and Clustering Coefficient) to identify the top 24 ranked KGs: SPON1, CLDN3, SPP1, CLDN4, CLDN7, VWA1, PCDH7, CDH6, JAG2, SCRIB, COL4A1, SLC9A3R1, LAMC1, CD47, TRAF4, MUC1, INHBB, VEGFA, CXCR4, CLDN10, GIPC2, COL4A6, FGF13, and BMP2. Our investigation indicates that CLDN3, SPP1, CLDN4, CLDN7, VWA1, PCDH7, CDH6, JAG2, SCRIB, COL4A1, SLC9A3R1, LAMC1, CD47, TRAF4, MUC1, INHBB, VEGFA, CXCR4, and CLDN10 exhibited downregulation, while COL4A6, FGF13, BMP2, and GIPC2 shown upregulation (Fig. 1; Table 4 and Supplementary 1).

Fig. 1.

Overview of the study design and analytical workflow for identifying cytoskeleton-associated signatures in ovarian cancer. A Schematic representation of the integrative bioinformatics pipeline. Publicly available transcriptomic datasets from TCGA-OV and six GEO cohorts (GSE66957, GSE10971, GSE26712, GSE29450, GSE54388, and GSE36668) were downloaded, normalized, and batch-corrected using the ComBat function (sva v3.46.0). B Differentially expressed genes (DEGs) were identified with edgeR v4.0 (|log₂FC| > 1.5, adjusted p < 0.05) and intersected with a curated list of 1,125 cytoskeleton-related genes from the PANTHER database to obtain a cytoskeleton-specific DEG subset. C Functional enrichment and pathway analyses were performed using clusterProfiler v3.9 for GO and KEGG annotations. D Protein–protein interaction (PPI) networks were constructed from STRING v11.0 (confidence ≥ 0.9) and visualized in Cytoscape v3.7.1; hub genes were ranked with CytoHubba metrics. E Downstream analyses included WGCNA and MEGENA for co-expression module detection, GSVA for EMT scoring, CellChat for single-cell communication inference, and CIBERSORT for immune-cell deconvolution. F Survival, correlation, and validation analyses integrated in silico predictions with clinical outcome data from TCGA and Kaplan–Meier Plotter. This figure summarizes the sequential analytical steps linking cytoskeletal gene dysregulation to EMT activation, immune remodeling, and patient prognosis in ovarian cancer

Table 4.

Genes up/downregulated involved in cytoskeleton protein in ovarian cancer

Gene	Description	Group	P. value	logFC
CLDN7	Claudin-7;CLDN7; PTN002485713;orthologs	Tight junction(PC00214)	2.10E-05	-3.33548
CLDN4	Claudin-4;CLDN4; PTN002485679;orthologs	Tight junction(PC00214)	1.38E-06	-3.93095
CLDN3	Claudin-3;CLDN3;PTN002485682;orthologs	Tight junction(PC00214)	1.47E-09	-4.66944
CLDN10	Claudin-10;CLDN10; PTN002485696;orthologs	Tight junction(PC00214)	2.98E-02	-1.15044
CD47	Leukocyte surface antigen CD47;CD47; PTN002472004;orthologs	Cell adhesion molecule(PC00069)	6.85E-07	-2.12583
PCDH7	Protocadherin-7;PCDH7; PTN002538113;orthologs	Cadherin(PC00057)	7.61E-06	-3.02113
MUC1	Mucin-1;MUC1;PTN002467157;orthologs	Cell adhesion molecule(PC00069)	1.18E-05	-1.86846
SPON1	Spondin-1;SPON1; PTN002477932;orthologs	Cell adhesion molecule(PC00069)	2.93E-11	-5.19006
CDH6	Cadherin-6;CDH6; PTN002513207;orthologs	Cadherin(PC00057)	1.79E-05	-2.91307
CXCR4	C-X-C chemokine receptor type 4; CXCR4; PTN002470993	Cell adhesion molecule(PC00069)	5.62E-03	-1.52004
COL4A6	Collagen alpha-6(IV) chain; COL4A6;PTN002513076;orthologs	Extracellular matrix structural protein(PC00103)	1.25E-04	1.075488
LAMC1	Laminin subunit gamma-1; LAMC1; PTN002471707	Extracellular matrix protein(PC00102)	1.29E-03	-2.28757
COL4A1	Collagen alpha-1(IV) chain; COL4A1; PTN002513090	Extracellular matrix structural protein(PC00103)	7.92E-03	-2.41812
VWA1	von Willebrand factor A domain-containing protein 1; VWA1; PTN002552213;orthologs	Extracellular matrix structural protein(PC00103)	5.30E-10	-3.33542
SPP1	Osteopontin; SPP1; PTN002471978;orthologs	Cytokine(PC00083)	5.93E-05	-4.52739
BMP2	Bone morphogenetic protein 2; BMP2;PTN002483810;orthologs	Growth factor(PC00112)	5.19E-05	1.79457
JAG2	Protein jagged-2;JAG2; PTN002543357;orthologs	Intercellular signal molecule(PC00207)	3.20E-08	-2.67117
INHBB	Inhibin beta B chain; INHBB; PTN002483777;orthologs	Growth factor(PC00112)	8.91E-05	-1.62212
VEGFA	Vascular endothelial growth factor A, long form; VEGFA; PTN002485916;orthologs	Growth factor(PC00112)	6.94E-03	-1.61085
FGF13	Fibroblast growth factor 13;FGF13;PTN002479519;orthologs	Growth factor(PC00112)	1.70E-02	1.566058
GIPC2	PDZ domain-containing protein GIPC2;GIPC2;PTN002487058;orthologs	Scaffold/adaptor protein(PC00226)	2.23E-07	2.639259
TRAF4	TNF receptor-associated factor 4;TRAF4; PTN002468287;orthologs	Scaffold/adaptor protein(PC00226)	8.62E-06	-2.09177
SLC9A3R1	Na(+)_H(+) exchange regulatory cofactor NHE-RF1;NHERF1;PTN002494222;orthologs	Scaffold/adaptor protein(PC00226)	4.42E-05	-2.30795
SCRIB	Protein scribble homolog; SCRIB; PTN002508427;orthologs	Scaffold/adaptor protein(PC00226)	1.89E-09	-2.42797

Construction PPI involved in the cytoskeleton protein

In the training cohort, 67 prognostic transcription factors (P-value < 0.001 and Hazard ratio > 1; Log-rank test) were discovered by univariate Cox analysis based on the protein activity patterns of cytoskeletal proteins. A random survival tree analysis was conducted on these 24 transcription factors to rank the variables by significance, using 1000 trees (PPI associated with CLDN12, CLDN14, CLDN16, CLDN17, CLDN34, CLDN8, ITGA5, ITGB1, MARVELD3, and TJP3). The average silhouette width was analyzed to identify stable and resilient subgroups of ovarian cancer patients, revealing that the best clustering number was achieved at k = 2 (Fig. 2). In the training cohort, six distinct subtypes of cytoskeletal proteins were identified: tight junction (PC00214), cell adhesion molecule (PC00069), cadherin (PC00057), extracellular matrix protein (PC00102), growth factor (PC00112), and scaffold/adaptor protein (PC00226) (Fig. 2) (Table 4).

Fig. 2.

Protein structure that may have a role in ovarian cancer. A Using random survival tree analysis, the relative relevance of 15 top-rank transcription factors is shown. B The wood plot showed the 15 transcription factors’ 95% CIs, HR, and P-value from the univariate Cox regression study. C The typical width of a K-means clustering silhouette for every k. When k = 2, the ideal cluster size was determined. D The training cohort’s Kaplan-Meier curves for two kinds of clustering. E Obovarian carcinoma was denoted by OV in this picture, whereas TF stood for transcription factor

Kaplan–Meier analysis

Kaplan–Meier analysis demonstrated a significant survival difference between the two cytoskeletal subtypes (Fig. 2). Subtype 1 had a median OS of 49.2 months (95% CI: 42.8–55.6) compared with 31.4 months (95% CI: 25.1–37.7) for subtype 2 (log-rank p = 0.004). For individual hub genes, high SPP1 expression was associated with worse outcome (HR = 2.14, 95% CI: 1.45–3.19, p = 0.001), and high BMP2 also predicted poorer survival (HR = 1.86, 95% CI: 1.22–2.83, p = 0.004). Conversely, elevated CLDN3 and CDH6 expression correlated with improved OS (CLDN3: HR = 0.58, 95% CI: 0.36–0.92, p = 0.021). Proportional hazards diagnostics did not indicate significant violations for reported models (all cox.zph p > 0.05).

Functional enrichments in the PPI network

To find improved biological processes and molecular activities involving cytoskeleton proteins, enrichment analysis was used. Bicellular tight junction assembly (GO:0070830), tight junction assembly (GO:0120192), apical junction assembly (GO:0043297), calcium-independent cell-cell adhesion via plasma membrane cell-adhesion molecules (GO:0016338) and positive regulation of multicellular organismal processes (GO:0051240) are some of the biological processes that we’ve identified as being relevant to our experiment. From what we can tell from the molecular function (MF), the hub genes are involved in activities such as receptor-ligand binding (GO:0070851), cytokine activity (GO:0005125), growth factor activity (GO:0008083), and gamma-aminobutyric acid transmembrane transporter activity (GO:0015185). Figure 3 shows that the cellular component (CC) links the hub genes to the following locations: the endoplasmic reticulum lumen (GO:0005788), the bicellular tight junction (GO:0005923), the tight junction (GO:0070160), the apical junction complex (GO:0043296), and the basement membrane (GO:0005604). The PPI network was specifically derived from the cytoskeleton-associated subset of DEGs (n = 112 genes), rather than the full DEG list. Using a high-confidence STRING score (≥ 0.9), we identified 40 nodes and 99 edges, with hub genes defined as those ranked in the top 10% of Degree, Betweenness, and Closeness centrality. This analysis yielded 24 key genes (SPON1, CLDN3, SPP1, CDH6, etc.), confirming that the network captures the cytoskeletal component of ovarian cancer dysregulation.

Fig. 3.

GO analysis. A BP, MF, and CC analysis, (B) signaling pathway analysis, (C) PPI analysis, and (D) signaling pathway analysis and their gene involvement

Association of cytoskeleton-related DEGs with EMT

To explore whether the identified cytoskeletal genes are functionally linked to EMT, we computed sample-level EMT scores using the GSVA method with the “hallmark EMT” gene set (MSigDB v7.5). Across TCGA and GEO ovarian cancer cohorts, we observed significant upregulation of mesenchymal markers (SPP1, BMP2, GIPC2, VIM) and downregulation of epithelial markers (CLDN3, CLDN4, CLDN7, CDH6) in high-EMT score samples (p < 0.001). This stratification confirms that the cytoskeletal dysregulation described in this study is strongly associated with EMT activation and the transition to invasive ovarian cancer phenotypes.

Differential gene co-expression network analysis in the TCGA cohort

The differential gene co-expression networks among thirteen clustering subgroups in the TCGA cohort were discerned in this study using Multi-scale Embedded Gene Co-expression Network Analysis (MEGENA). Out of the 120 differential gene network modules found, the MEGENA algorithm found that cluster 2 patients were more likely to have certain diseases than cluster 1 patients. When enriched modules in cluster 1 of the TCGA cohort were found, WGCNA was utilized to look into their connection to the clustering subtype further. Figure 4 shows the thirteen enriched gene co-expression network modules that showed a clustering subtype absolute correlation higher than 0.5. Further study was conducted using network analysis on the 32 enriched gene co-expression network modules associated with cluster 1 in order to identify hub genes and the degree values of these genes. With 245 connections to other nodes in the network module, GRHL2 showed the highest degree among the identified hub genes (Fig. 6B). At the same time, these modular genes were analyzed using KEGG enrichment (Supplementary Table 4). With a correlation value of 0.64, the c1 180 gene module was most strongly associated with the clustering subgroups. In Fig. 4 and Supplementary 2, twelve genes from the c1 80 gene module showed significant enrichment across several KEGG cancer pathways. These pathways included the VEFG signaling pathway, ECM-receptor interaction, calcium-independent cell-cell adhesion, and extracellular structure organization.

Fig. 4.

Network analysis of differential gene co-expression in the TCGA cohort. A Two main categories of clustering and their relationship to 25 enriched gene co-expression network modules. B a heatmap displaying the degree values of hub genes for differential gene co-expression network modules across three clustering types. C A correlation plot of twelve highly enriched gene co-expression network modules for every ovarian cancer dataset and

Fig. 6.

Using harmonized scRNA-seq data from multiple public datasets processed by Seurat, Harmony, and UMAP, this figure displays the integration of ovarian tumor and normal tissue samples. A shows cell–cell communication networks in ovarian cancer, while (B) shows those in normal ovarian tissue. C presents a heatmap summarizing the frequency and strength of intercellular communications across major cell types, including epithelial, immune, and stromal populations. The integration revealed disease-specific communication hubs dominated by epithelial-CAF and macrophage–endothelial signaling, underscoring the altered interaction landscape in tumors

Validation of immune deconvolution robustness

Following quality control filtering, the majority of ovarian cancer samples demonstrated statistically significant CIBERSORT deconvolution results (p < 0.05), indicating adequate immune-related signal and sample heterogeneity. Pearson correlation coefficients between observed and inferred expression profiles were consistently high, while RMSE values remained low, supporting the accuracy of immune cell fraction estimates. These results confirm that the analyzed datasets satisfy CIBERSORT’s assumptions and that the inferred immune infiltration patterns reliably reflect the tumor immune microenvironment.

An integrated single-cell transcriptome atlas and cell-cell communication of human ovarian cancer

Multiple prior research have documented the examination of cellular heterogeneity and functions in human ovarian cancer using scRNA-seq, although they vary regarding clinical stages. Initially, we aimed to create a comprehensive and curated single-cell transcriptome atlas to facilitate the systematic examination of the cellular compositions and heterogeneity of OC. We gathered and re-analyzed 40 scRNA-seq datasets, eliminating low-quality data by the methodologies outlined in the Methods section. We ultimately acquired 505,102 high-quality cells from 9 ovarian cancer clinical stages for further investigation. We conducted an integration study on several samples, tissues, and clinical stages of ovarian cancer utilizing harmony25 (Figs. 5 and 6).

Fig. 5.

This figure illustrates the global interaction network between immune and stromal cell populations in ovarian cancer, quantified using the CellChat package based on single-cell RNA-seq data from integrated datasets of 505,102 cells across nine clinical stages. Panels (A–C) depict ligand–receptor pairings, identifying sender (ligand-expressing) and receiver (receptor-expressing) cell types, while (D–F) show correlations between receptor, ligand, and total interaction scores. Interactions circled in black indicate significant correlations (p < 0.01). These results highlight regulatory crosstalk involving Tregs, macrophages, and cancer-associated fibroblasts (CAFs) that modulate the tumor microenvironment and immune evasion

Validation of hub genes in a single cell

Quality control was performed using Seurat v4.3.1 with the following thresholds: nFeature_RNA > 500, nFeature_RNA < 7000, and percent.mt < 25. UMAP visualization and cell clustering were performed after normalization and batch correction with Harmony (v0.1). The integration workflow, cell filtering, and annotation process are shown in Fig. 7. To investigate the correlation between tumor cell heterogeneity and prognostic risk scores in ovarian cancer, we categorized ovarian tumor cells into high-risk and low-risk groups. Thereafter, we conducted irGSEA (Iterative Gene Set Enrichment Analysis). We validated all genes at the single-cell level, revealing that CLDN3, SPP1, CLDN4, CLDN7, VWA1, PCDH7, CDH6, JAG2, SCRIB, COL4A1, SLC9A3R1, LAMC1, CD47, TRAF4, MUC1, INHBB, VEGFA, CXCR4, and CLDN10 were downregulated, while COL4A6, FGF13, BMP2, and GIPC2 were increased (Figs. 7, 8 and 9). To verify the robustness of single-cell integration across datasets generated from different platforms, we quantitatively evaluated batch-effect correction performance following Harmony integration. kBET analysis demonstrated a marked reduction in batch-driven neighborhood bias after integration, indicating effective mixing of cells from different datasets. Consistently, iLISI scores increased substantially after correction, reflecting improved batch mixing, whereas cLISI scores remained stable, confirming preservation of biologically meaningful cell-type separation. UMAP visualizations colored by dataset origin further showed that cells clustered primarily by cell identity rather than by batch or platform following integration (Supplementary Fig. S2–S4). Together, these results confirm that Harmony-based integration effectively mitigated batch effects while retaining critical biological signals required for downstream cell–cell communication and EMT-related analyses.

Fig. 7.

A UMAP visualization of integrated single-cell data (combined GSE118828 + GSE146026 + GSE173682), annotated by cell type. Each point represents a cell, colored by its predicted lineage. B Expression levels of hub genes across patient samples (P04–P24 represent independent patient IDs corresponding to GEO datasets GSE118828. The results are presented as box plots of normalized expression (not bar plots) to show distribution within each sample group. Samples P04–P24 correspond to individual patient-derived single-cell datasets obtained from GEO. Each label represents one ovarian cancer sample integrated into the Seurat–Harmony pipeline. C displays a heatmap of the top 20 marker genes for each major cell type along with GO enrichment terms for biological processes enriched in these marker sets. Together, these analyses reveal high intratumoral heterogeneity and distinct molecular signatures associated with specific cell populations contributing to tumor progression

Fig. 8.

This figure shows expression patterns of 24 key cytoskeleton-related genes (KGs) across single-cell clusters in ovarian cancer. Using irGSEA and normalized expression matrices, genes such as CLDN3, SPP1, CLDN4, CLDN7, and CDH6 were predominantly downregulated in malignant epithelial clusters, while BMP2, FGF13, COL4A6, and GIPC2 were upregulated. The figure highlights transcriptional heterogeneity among individual tumor cells, linking gene-level variation with cytoskeletal remodeling and EMT-related phenotypes

Fig. 9.

Tumor microenvironmental features derived from single-cell RNA-seq analysis. This visualization summarizes the composition and spatial relationships of tumor-associated cell types. Cluster-specific enrichment indicates distinct microenvironmental niches containing macrophages, CAFs, and proliferating epithelial cells. The patterns suggest that cytoskeleton-related gene dysregulation supports extracellular matrix remodeling and immune modulation, reinforcing the biological roles inferred from bulk transcriptome analysis

Investigation of immune infiltration and ovarian cancer

Investigation of immune infiltration and ovarian cancer

Figure 10 shows immune cell predominance in samples. The research found that tumor cells contained more immature B cells, resting memory CD4 T cells, and M0, M1, M2 macrophages. Tumor cells contained less memory B, plasma, CD8 T, activated memory CD4 T, and monocytes. Figure 10A-B. A heatmap in Fig. 10 shows the varied expression levels of immune cells within tumor cells. Survival probability correlated with CD4 memory T cell activation. A large fraction of CD4 memory-activated T cells survived. Researchers found correlations between 20 immune cell types. Younger people showed higher amounts of naïve B cells, quiescent mast cells, and activated NK cells compared to older persons. Lower numbers of active mast cells, neutrophils, and resting NK cells were seen in younger people. G1/2 had more Macrophage M0, active Mast cells, and Plasma cells than G3. Macrophage M1, resting Mast cells, Monocytes, activated CD4 memory T cells, CD8 T cells, and follicular helper T cells were lower in G1/2 than G3 (Supplementary 3). Kaplan–Meier survival analysis demonstrated that high expression of SPP1, BMP2, and GIPC2 correlated with poorer overall survival, whereas high CLDN3, CLDN4, and CDH6 expression was associated with better prognosis (adjusted p < 0.05). Each plot now includes numbers at risk and 95% CI for median survival. Multivariate Cox regression confirmed that SPP1 (HR 2.14, 95% CI: 1.45–3.19, p = 0.001) and BMP2 (HR 1.86, 95% CI: 1.22–2.83, p = 0.004) remained independent predictors of poor outcome. PH diagnostics indicated no significant violations of model assumptions.

Fig. 10.

Antigen-presenting CAF was identified and enhanced ECM function. A quantifies the number and strength of predicted interactions between major immune and stromal cell populations; (B) shows mRNA expression and infiltration scores for selected hub genes; (C) presents a heatmap of immune cell distribution across tumor samples. Data derived from CIBERSORT and CellChat analyses revealed increased infiltration of macrophage (M0/M1/M2) and resting memory CD4 + T cells, but reduced CD8 + T cell and monocyte activity in tumors. These findings demonstrate the immune-suppressive nature of the ovarian tumor microenvironment

MicroRNAs candidates

During this phase, we have discovered a total of 10 genes. Subsequently, we discovered and selected the most relevant microRNAs, as seen in Fig. 11. Figure 11 illustrates the presence of hsa-miR-101-3p, hsa-miR-130b-3p, hsa-miR-141-3p, hsa-miR-200a-3p, and hsa-miR-153.

Fig. 11.

P-value and overlap gene with candidate microRNAs extracted in miRNA databases. This figure integrates DIANA-microT and miRTarBase predictions with DEG and PPI data to identify regulatory miRNAs targeting cytoskeletal hub genes. Panels (A–C) show overlapping miRNA–mRNA networks derived from multiple databases; (D) presents the interaction map of miRNA–mRNA pairs; and (E) highlights significant candidate miRNAs (hsa-miR-101-3p, hsa-miR-130b-3p, hsa-miR-141-3p, hsa-miR-200a-3p, hsa-miR-153) based on p-value and q-value ranking. These miRNAs are predicted to modulate epithelial–mesenchymal transition and cytoskeletal organization in ovarian cancer progression

Integration of miRNA and immune findings with cytoskeletal regulation

The integration of miRNA–mRNA networks highlights potential regulatory layers governing cytoskeletal remodeling in ovarian cancer. Notably, miR-200a-3p and miR-141-3p, previously implicated in EMT suppression, were predicted to target CLDN3 and CDH6, both of which are key epithelial markers. Conversely, miR-130b-3p and miR-153 may enhance invasive potential through regulation of SPP1 and BMP2. These observations suggest a coordinated miRNA-mediated modulation of the cytoskeletal network during tumor progression. In parallel, the immune infiltration results reveal that cytoskeletal gene dysregulation coincides with altered immune cell composition, supporting a mechanistic link between cytoskeleton-driven ECM remodeling and immune evasion in the ovarian tumor microenvironment. Using an expanded multi-database approach, we identified a broader set of miRNAs potentially regulating cytoskeleton-associated hub genes in ovarian cancer. Several key miRNAs, including hsa-miR-200a-3p, hsa-miR-141-3p, hsa-miR-101-3p, hsa-miR-130b-3p, and hsa-miR-153, were consistently supported across multiple databases and/or experimentally validated sources. Notably, miR-200 family members were predicted to target epithelial markers such as CLDN3 and CDH6, reinforcing their established roles in EMT suppression, whereas miR-130b-3p and miR-153 were associated with regulation of pro-mesenchymal genes such as SPP1 and BMP2. These findings strengthen the robustness of the inferred miRNA–mRNA regulatory network and highlight additional post-transcriptional mechanisms contributing to cytoskeletal remodeling and EMT in ovarian cancer.

Discussion

This research revealed significant DEGs associated with ovarian cancer development, emphasizing cytoskeletal proteins. Through the construction of a PPI network using DEGs, we discovered 24 kg that are pivotal in ovarian cancer biology. These genes were identified by five topological metrics, including Degree, Betweenness, and Clustering Coefficient, enabling us to ascertain the most important nodes within the network. Among the identified genes, many, including CLDN3, SPP1, CLDN4, and CD47, exhibited downregulation, whilst others such as COL4A6, FGF13, BMP2, and GIPC2 shown upregulation. The results indicate that cytoskeleton-associated genes may be essential to the modified cellular structure and tumor microenvironment seen in ovarian cancer.

Our results highlight the significance of tight junction proteins, including CLDN3, CLDN4, CLDN7, and CLDN10, which were mostly downregulated in ovarian cancer specimens. Tight junctions are crucial for preserving cell polarity and tissue integrity; their disruption is associated with heightened cancer cell invasion and metastasis. The marked downregulation of these claudins indicates a probable lack of cell adhesion, potentially facilitating ovarian cancer growth and metastasis. Additionally, genes associated with the extracellular matrix and cell adhesion molecules, including SPON1, CD47, and LAMC1, exhibited dysregulation, underscoring the modified cellular architecture and signaling pathways that contribute to the aggressiveness of ovarian cancer (Fig. 12).

Fig. 12.

Overview of the integrative bioinformatics workflow used to identify cytoskeleton-associated genes and molecular pathways in ovarian cancer. The schematic summarizes the five main analytical stages: (1) Data acquisition from public repositories, including GEO, TCGA, ArrayExpress, curatedOvarianData, and 10X Genomics; (2) Data preprocessing and differential expression analysis using edgeR, GEO2R, and ComBat for batch correction; (3) Functional and network analysis involving GO/KEGG enrichment, STRING, and Cytoscape for hub gene identification; (4) Integration of single-cell RNA-seq data via Seurat, CellChat, Harmony, and UMAP for clustering and cell–cell communication analyses

Alongside cell adhesion-related genes, many growth factors were identified as pivotal regulators in our investigation. BMP2, INHBB, and VEGFA were identified as elevated genes, suggesting their probable involvement in facilitating angiogenesis and tumor proliferation in ovarian cancer. VEGFA, a prominent pro-angiogenic factor, is integral to tumor vascularization, which is vital for delivering nutrients to proliferating tumors. The overexpression of growth factors like BMP2 indicates an increase in cancer cell proliferation and survival, presenting a potential target for therapeutic intervention. This corresponds with prior studies emphasizing the significance of the tumor microenvironment and growth factor signaling in ovarian cancer. This study’s clustering and survival analysis identified two distinct subtypes of ovarian cancer based on the expression of cytoskeletal protein-related genes. The Kaplan-Meier curves revealed substantial disparities in patient survival between the two categories, suggesting that cytoskeleton-associated genes may possess predictive significance. These results are significant as they indicate that targeting certain cytoskeletal pathways may facilitate the creation of more customized therapy options for ovarian cancer patients. The finding of transcription factors associated with cytoskeletal proteins reinforces the promise for targeting these regulatory networks in ovarian cancer treatment. “Kaplan–Meier survival curves showed significant differences between the two patient clusters (log-rank P < 0.05). Furthermore, univariate Cox regression analysis was performed, and hazard ratios (HRs) with 95% confidence intervals (CIs) were calculated for each hub gene and cluster. Several key genes (e.g., CLDN3, COL4A6, VEGFA, BMP2) demonstrated significant HRs (HR > 1 or HR < 1 with 95% CI not crossing 1), indicating strong prognostic value. These findings confirm that cytoskeleton-associated genes serve as robust predictors of patient outcome.”

The upregulation of the CLDN-1 gene has been linked to ovarian cancer pathogenesis. Research indicates that the overexpression of CLDN-1 is associated with decreased cellular differentiation and an increased rate of invasive tumor growth. The role of CLDN-1 has been extensively investigated in two major histological subtypes of ovarian cancer: serous and endometrioid carcinomas. Notably, CLDN-1 expression is negatively regulated by microRNA-155 (miR-155), which subsequently leads to reduced proliferation and invasiveness of ovarian cancer-initiating cells. Additional reports have demonstrated that elevated expression of CLDN-11, CLDN-4, and CLDN-7 contributes to the growth of both benign and malignant epithelial ovarian cancers. Comprehensive analyses examining the correlation between CLDN-1 expression, survival outcomes, and tumor localization revealed an 85% increase in CLDN-1 levels [70]. More recently, studies have explored CLDN-1 expression in borderline ovarian tumors (BOT), with findings showing significantly elevated levels of CLDN-1 in association with peritoneal implants and micropapillary patterns, features unique to serous BOT [71]. SLC9A3R1 (solute carrier family 9, subfamily A [NHE3, cation proton antiporter 3], member 3 regulator 1), a multifunctional scaffold protein, plays a critical role in inhibiting the proliferation of breast cancer cells. Furthermore, the SLC9A3R1-associated signaling pathway has been shown to regulate the activation of autophagic processes [72, 73]. Beyond their individual functions, several of the identified hub genes may act synergistically to drive ovarian cancer progression through interconnected regulatory networks. For example, tight junction proteins such as CLDN3, CLDN4, and CLDN7 maintain epithelial polarity, and their downregulation weakens cell–cell adhesion, creating a permissive state for EMT. Growth factors including BMP2 and VEGFA further activate EMT-associated pathways (e.g., TGF-β and PI3K/AKT), amplifying cytoskeletal remodeling and promoting invasion. Additionally, ECM-related genes such as LAMC1, SPON1, and CD47 reshape the tumor microenvironment and interact with integrin and FAK/Src signaling to enhance cell motility. Notably, BMP2 and VEGFA can increase ECM stiffness, which feeds back to suppress epithelial markers (e.g., CDH6, CLDN3) and reinforce mesenchymal phenotypes. This reciprocal regulation suggests that cytoskeletal dynamics, adhesion loss, and growth factor signaling operate in a coordinated manner rather than in isolation. Such crosstalk may also contribute to immune modulation, as ECM remodeling and CD47-mediated ‘don’t-eat-me’ signaling facilitate immune evasion. Together, these findings highlight a complex regulatory network in which cytoskeleton-associated genes cooperate to drive EMT, matrix remodeling, and immunosuppressive microenvironment formation in ovarian cancer [1, 74–76].

To translate these findings into clinical practice, future research should focus on validating the identified cytoskeletal biomarkers through wet-lab experiments, including in vitro functional assays, in vivo tumor models, and drug response analyses. Clinical trials should be designed to evaluate the prognostic and predictive utility of these biomarkers in patient cohorts, enabling their use in precision oncology for ovarian cancer [77, 78]. Integrating these biomarkers into clinical workflows may improve patient stratification, guide therapeutic decision-making, and enhance outcomes. Importantly, this study relies entirely on in-silico analyses, and no experimental validation (e.g., qPCR, immunohistochemistry, or western blotting) of key hub genes such as CLDN3, COL4A6, or VEGFA was performed. This represents a significant limitation, as functional validation is essential to confirm the biological relevance and mechanistic roles of these genes in ovarian cancer. In addition, the clinical significance of these findings should be emphasized, particularly how the identified cytoskeletal biomarkers may guide patient stratification, prognosis, and therapeutic decision-making. Future validation strategies should include in vitro and in vivo wet-lab experiments, such as gene knockdown or overexpression assays, functional rescue studies, and drug sensitivity analyses, as well as prospective clinical trials to assess their predictive and diagnostic value in patient populations [74].

Our results are corroborated by several investigations that have also recognized cytoskeletal dysregulation as a critical characteristic of cancer development. Prior studies have shown that claudin family proteins, especially CLDN3 and CLDN4, are often downregulated in ovarian cancer, resulting in diminished cell adhesion and heightened metastatic potential. Research showed that the disruption of tight junction proteins facilitates ovarian cancer invasion, hence reinforcing our findings about the role of claudins in tumor aggressiveness [58]. Scrib, a membrane-associated protein, is essential for maintaining apical-basal polarity in epithelial tissues. However, in ovarian cancers, Scrib has been observed to become mislocalized to the cytoplasm [79, 80]. This study is based entirely on publicly available transcriptomic datasets and computational analyses. Although integrating multiple microarray and single-cell RNA-seq cohorts increases robustness, batch effects, sample heterogeneity, and limited clinical annotation may influence the findings. Furthermore, functional validation of the identified hub genes and pathways was not performed. Therefore, while our results suggest potential cytoskeletal biomarkers and pathways relevant to ovarian cancer progression, further experimental studies and independent clinical cohorts are required to confirm these observations. Although the present analysis integrates multiple microarray and single-cell RNA-seq datasets, several limitations should be acknowledged. First, the study is entirely based on publicly available transcriptomic data, which may introduce sampling bias and restrict control over clinical variables such as treatment history, disease stage, and tissue heterogeneity. Second, despite using normalization and batch correction (e.g., ComBat) to minimize technical variation, residual batch effects and platform-specific discrepancies between microarray, RNA-seq, and single-cell data may influence the results. Third, our conclusions are based on computational predictions and correlations rather than direct experimental validation. Functional assays, such as gene knockdown or overexpression experiments, will be necessary to confirm the causal roles of the identified hub genes. Finally, while this work focuses on transcriptomic profiles, future integrative studies combining proteomic, epigenetic, and metabolomic data could provide a more comprehensive understanding of cytoskeletal regulation in ovarian cancer [81–83].

Furthermore, studies have emphasized the significance of growth factors, including VEGFA and BMP2, in facilitating tumor angiogenesis and proliferation. Research conducted by Goelet al. [84] and Kim et al. [85] shown that inhibiting VEGF signaling pathways in ovarian cancer may diminish tumor proliferation and enhance patient survival, corroborating our observations about the increase of VEGFA and its viability as a therapeutic target. Collectively, these investigations bolster the suggestion that cytoskeleton-related genes are pivotal in ovarian cancer progression and highlight the need for more investigation of these pathways in therapeutic advancement.

Conclusion

In conclusion, this work provides a comprehensive in silico framework identifying cytoskeleton-associated gene networks and potential EMT-related pathways in ovarian cancer. While these results reveal promising biomarker candidates, they represent hypothesis-generating insights that require experimental confirmation. Future in vitro and in vivo validation will be essential to verify the functional roles of these genes and their translational relevance in ovarian cancer therapy.

Although this study integrates multiple bulk and single-cell transcriptomic datasets to robustly identify cytoskeleton-associated hub genes in ovarian cancer, it is important to note that all findings are derived from in silico analyses. No experimental validation, such as quantitative real-time PCR (qRT-PCR), Western blotting, or IHC, was performed to confirm the differential expression of key hub genes (e.g., CLDN3, SPP1, BMP2, CDH6, and VEGFA) at the mRNA or protein level. Consequently, the functional roles inferred for these genes in EMT, cytoskeletal remodeling, and tumor microenvironment regulation should be interpreted as hypothesis-generating rather than definitive. Future studies should prioritize wet-lab validation using ovarian cancer cell lines and patient-derived tissues. Recommended approaches include qRT-PCR and Western blotting to confirm gene and protein expression, immunohistochemical staining to assess spatial localization within tumor tissues, and functional assays such as gene knockdown or overexpression experiments to evaluate effects on cell migration, invasion, EMT markers, and cytoskeletal organization. In vivo models may further clarify the contribution of these genes to tumor progression and therapeutic response.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Ali Shakeri Abroudi: Responsible for composing the first draft, doing statistical and bioinformatics analysis, Aryan Jalaeianbanayan: bioinformatics analysis, and Melika Djamali : bioinformatics analysis and writing and Hossein Azizi: Conceptualization, doing statistical and manuscript editing.

Funding

None.

Data availability

All research analysis data is publicly available in the database above. This study’s source codes were accessible via Zenodo: [https://zenodo.org/records/15586892] (https:/zenodo.org/records/15586892) . All datasets analyzed in this study are publicly available from established online repositories. The accession numbers, database names, and direct web links are provided below and summarized in Table 1. Gene Expression Omnibus (GEO)–National Center for Biotechnology Information (NCBI), GSE66957 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE66957] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE66957)), GSE10971 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE10971] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE10971)), GSE26712 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE26712] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE26712)), GSE29450 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE29450] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE29450)), GSE54388 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE54388] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE54388)) and GSE36668 ([https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE36668] (https:/www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE36668)). The Cancer Genome Atlas (TCGA)–The TCGA Ovarian Cancer (OV) dataset was obtained from the cBioPortal database (https://www.cbioportal.org/study/summary?id=ov_tcga_pub) and used for differential expression and survival analyses. ArrayExpress – Microarray data from ovarian cancer cohorts were accessed via the ArrayExpress database of EMBL-EBI ([https://www.ebi.ac.uk/arrayexpress/] (https:/www.ebi.ac.uk/arrayexpress)). curatedOvarianData – Curated ovarian cancer transcriptomic datasets were obtained from the curatedOvarianData R package repository.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.