Identification of ribosomal stress related signature genes and immune microenvironment analysis in ovarian cancer based on multi-machine learning

Abstract

Background

Ovarian cancer (OC) is a highly aggressive malignancy with poor prognosis and limited response to immunotherapy. Ribosomal stress, a cellular response to disrupted ribosome biogenesis, has been increasingly implicated in tumorigenesis and immune regulation, yet its contribution to OC remains unclear.

Methods

We integrated four GEO transcriptomic datasets and identified ribosomal stress related signature genes (RSRGs) through differential expression and functional enrichment analyses. To construct a robust diagnostic model, three machine learning algorithms: LASSO regression, support vector machine recursive feature elimination (SVM-RFE), and random forest were combined. Immune infiltration patterns were evaluated using CIBERSORT, and interpretability analysis was performed using SHAP to determine feature importance. Functional validation of BMP6 was performed in ovarian cancer cell lines by RT-qPCR, Western blot, CCK-8, colony formation, and Transwell assays to evaluate its effects on proliferation, migration, and invasion.

Results

A total of 117 differentially expressed RSRGs were identified, mainly enriched in cytoskeletal regulation, lipid metabolism, proteoglycan signaling, and IL-17 mediated inflammatory pathways. The integrated machine learning approach identified six feature genes (SPP1, MAPK13, LCN2, JUP, DSP, and BMP6). SHAP analysis revealed that SPP1 and DSP had the greatest contributions to the predictive model. Immune profiling revealed increased macrophage M0/M2 and decreased CD8 + T cell infiltration in high-risk samples, with SPP1, MAPK13, and DSP positively correlated with macrophage abundance. Functional assays demonstrated that BMP6 was downregulated in ovarian cancer cells and that its overexpression significantly inhibited proliferation, migration, and invasion.

Conclusions

This study identifies a six-gene ribosomal stress signature linking tumor intrinsic pathways and immune remodeling in OC. BMP6 exerts tumor-suppressive effects, supporting the potential of targeting ribosomal stress and its immune axis as a therapeutic strategy in ovarian cancer.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05167-x.

Highlights

Machine learning, based integration of multi-dataset analysis identified six ribosomal stress related genes (SPP1, MAPK13, LCN2, JUP, DSP, and BMP6) as diagnostic biomarkers for ovarian cancer.

Functional validation of BMP6 demonstrates its tumor-suppressive role by inhibiting ovarian cancer cell proliferation, migration, and invasion.

Ribosomal stress pathways intersect with IL-17 signaling and metabolic remodeling, linking nucleolar dysfunction to immune evasion and tumor progression in ovarian cancer.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05167-x.

Introduction

Ovarian cancer (OC) is a malignant tumor that poses a serious threat to women’s health. In 2022, its age-standardized global incidence rate was 6.7 per 100,000, ranking only after breast, cervical, and endometrial cancers, while the rate in China was 5.7 per 100,000 [1]. OC is one of the most lethal malignancies of the female reproductive system, characterized by insidious early symptoms and a high tendency for recurrence and metastasis, which present major challenges in clinical management [2–4]. Its complex pathogenesis and pronounced intratumoral heterogeneity underscore the urgent need for novel molecular mechanisms and biomarkers to improve early diagnosis, prognostic prediction, and therapeutic stratification [5, 6].

In recent years, ribosomal stress has emerged as a crucial cellular event linking ribosome biogenesis dysfunction to tumorigenesis [7, 8]. Beyond its canonical role in protein synthesis, the nucleolus functions as a stress sensor that coordinates cellular responses to DNA damage, oncogene activation, and metabolic alterations [9]. Ribosomal stress can activate p53-dependent or independent signaling cascades that control cell cycle arrest, apoptosis, and senescence [10, 11]. Importantly, accumulating evidence indicates that ribosomal stress also exerts immune-modulatory effects, influencing the tumor immune microenvironment (TIME) by modulating immune evasion and inflammatory signaling [12–14]. These findings suggest that ribosomal stress could serve as a bridge between tumor metabolic reprogramming and immune regulation in OC. Concurrently, machine learning (ML) has revolutionized cancer genomics by enabling high-dimensional data integration and the discovery of key diagnostic and prognostic biomarkers [15, 16]. Integrating ML with ribosomal stress related gene profiling may therefore uncover novel OC-associated molecular features and provide new insights into tumor biology.

In addition, the tumor immune microenvironment (TIME) plays a pivotal role in OC progression and therapeutic response [17]. Immune infiltration patterns, including tumor-infiltrating lymphocytes (TILs), tumor-associated macrophages (TAMs), and regulatory T cells (Tregs), have been closely associated with clinical outcomes and sensitivity to immunotherapy [18, 19]. Dysregulation of immune checkpoint pathways and the immunosuppressive milieu are key contributors to poor responses to immune checkpoint inhibitors (ICIs) in OC [20]. Recent studies have further highlighted that the remodeling of immune cell composition, particularly the balance between pro-inflammatory and immunosuppressive cell populations is closely linked to tumor progression and therapeutic resistance in ovarian cancer [21–23]. Understanding how ribosomal stress related pathways interact with immune cell infiltration could provide a mechanistic basis for novel immunotherapeutic strategies.

Given these considerations, this study aims to identify ribosomal stress related signature genes in ovarian cancer using a multi-machine learning approach and to comprehensively characterize their association with the immune microenvironment. By integrating transcriptomic data from multiple public databases and applying complementary ML algorithms, we sought to construct a robust ribosomal stress related gene signature for diagnostic and prognostic prediction. Furthermore, immune infiltration analyses and immune checkpoint correlation studies were performed to elucidate the potential immunological mechanisms underlying ribosomal stress driven tumor biology. Our findings are expected to provide new insights into the molecular basis of ovarian cancer and contribute to the development of more effective diagnostic and therapeutic strategies.

Materials and methods

Data sources and preprocessing of expression profiles

Four ovarian cancer-related gene expression datasets(GSE6008, GSE4122, GSE12470, and GSE66957) were downloaded from the public Gene Expression Omnibus (GEO) database. These datasets included OC tissues and normal control samples obtained from different experimental platforms, aiming to improve the stability and generalizability of the results through integrated multi-dataset analysis. All raw expression matrix files were combined with corresponding clinical information files for sample classification and annotation (Supplementary Table S1).

To ensure cross-platform comparability, the raw expression matrices of each dataset were preprocessed. The avereps() function in the limma package was used to merge and average duplicate probe signals. The necessity of log₂ transformation was then determined: if the interquartile range exceeded a predefined threshold or if the 99th percentile was greater than 100, the data were considered untransformed, and a log2(x + 1) conversion was automatically applied. Background correction and intensity normalization were then performed using the normalizeBetweenArrays() function to eliminate technical bias. The resulting normalized gene expression profiles were used for subsequent integrated analyses.

Batch effect correction and principal component analysis (PCA)

This study used the ComBat algorithm (based on the sva package) to perform batch effect correction. The specific procedure included the following steps: extracting the common genes from the four datasets as the intersection gene set, constructing an initial merged matrix containing all samples, and then using the ComBat() function to estimate and adjust parameters under known batch labels to generate the batch-corrected expression matrix. This process effectively improved the comparability of data from different sources. To further verify the effect of batch correction, Principal Component Analysis (PCA) was used to visualize the distribution pattern of samples in high-dimensional space. A customized function, bioPCA(), was applied to draw PCA plots before and after batch correction (“Before batch correction” and “After batch correction”). This function performed dimensionality reduction using the prcomp() function and plotted scatter diagrams with the ggplot2 and ggpubr packages, with colors and shapes representing different dataset sources. Ideally, after correction, samples should cluster according to biological status rather than dataset origin, indicating successful removal of batch interference.

Differential expression analysis

Based on the batch-corrected integrated expression matrix, differential expression analysis was performed. Sample names were parsed to identify the corresponding project, sample ID, and group type. A design matrix was constructed, and linear model fitting was conducted using the lmFit function. The contrast condition “Treatment vs Control” was defined with makeContrasts, and empirical Bayes moderation (eBayes) was applied to calculate the t-statistics and adjusted p-values (FDR). The screening criteria were set as |log₂FoldChange| > 1 and adj.P.Value < 0.05, indicating at least a twofold change in expression with statistical significance after multiple testing correction. The resulting significant differentially expressed gene list and corresponding expression subset were generated for subsequent analyses.

Intersection analysis of differentially expressed genes and ribosomal stress related genes

Given that this study focused on the role of ribosomal stress in ovarian cancer, a set of reported ribosomal stress-related genes (RSRGs) was obtained from the GeneCards database (https://www.genecards.org/) using the targeted search query “ribosomal stress”. The full gene set is provided in Supplementary Table S2. The previously identified significantly differentially expressed genes (DEGs) were intersected with this RSRG set to identify candidate genes that met both criteria-differential expression and involvement in ribosomal stress. These overlapping genes were defined as differentially expressed ribosomal stress-related genes (DE-RSRGs). The VennDiagram package was used to generate a Venn diagram to visually illustrate the overlap between the two gene sets. The intersecting gene list was then exported and used as the primary input features for subsequent machine learning modeling.

Functional enrichment analysis of DE-RSRGs

To elucidate the potential biological functions and pathway involvement of DE-RSRGs, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed. The intersecting gene symbols were first converted into Entrez IDs using the org.Hs.eg.db database. Subsequently, the enrichGO() and enrichKEGG() functions were applied to conduct hypergeometric tests. Terms with a q-value < 0.05 were considered significantly enriched. The enrichment results were visualized using bubble plots and customized circular plots, covering the three main GO categories: biological process (BP), cellular component (CC), and molecular function (MF), as well as key signaling pathways identified in the KEGG analysis.

Comparison of three machine learning models for feature gene selection

To further narrow the range of candidate genes and enhance the predictive performance of diagnostic biomarkers, three mainstream machine learning algorithms were applied for feature selection. The input features for all three models were the expression levels of the 45 DE-RSRGs identified from the intersection of DEGs and RSRGs, rather than the full transcriptomic profile. This pre-filtering step was applied prior to model training to reduce dimensionality and focus on biologically relevant candidates.

Lasso Regression (Least Absolute Shrinkage and Selection Operator): This method is suitable for high-dimensional datasets with relatively small sample sizes. By applying L1 regularization, it shrinks the coefficients of less important variables to zero, achieving sparse modeling. The glmnet package was used to construct a binary logistic regression model, and the optimal λvalue was determined through cross-validation. Genes with nonzero coefficients were extracted as candidate features.

Support Vector Machine-Recursive Feature Elimination (SVM-RFE): This approach combines the classification power of support vector machines with an iterative elimination strategy, sequentially removing the least important features to retain the most discriminative subset of genes. A custom R script (SHAP12.msvmRFE.R) was used to implement the algorithm, and the optimal number of features was determined based on the minimum cross-validation error criterion.

Random Forest (RF): As an ensemble learning method based on decision trees, random forest provides feature importance scores (Mean Decrease Accuracy or Gini Index). The randomForest package was used to train the classifier and generate a ranked list of gene importance, from which the top-ranked genes were selected as potential biomarkers.

The intersection of these three gene sets was identified to determine the feature genes of ovarian cancer, defined as genes simultaneously selected by all three machine learning models. The VennDiagram package was used to visualize the overlap among the gene sets, and the intersecting genes were exported for subsequent consistency analysis and comprehensive evaluation. A box plot illustrated the differential expression of the feature genes between cancer (Treat) and normal (Control) groups, while a circos plot displayed the chromosomal locations of these feature genes.

Interpretability analysis of machine learning models

Although machine learning models demonstrate high predictive performance, their “black-box” nature limits biological interpretability. To enhance model transparency, SHAP (SHapley Additive exPlanations) analysis was introduced in this study. To evaluate the discriminative performance of different algorithms, the dataset was randomly divided into a training set and a testing set at a ratio of 7:3 using the createDataPartition function to ensure balanced class distribution. A total of ten mainstream machine learning algorithms were compared, including Partial Least Squares (PLS), Random Forest (RF), Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Logistic Regression, K-Nearest Neighbors (KNN), XGBoost, Gradient Boosting Machine (GBM), Neural Network, and glmBoost. Model training and optimization were performed with fivefold repeated cross-validation using the trainControl function to improve generalization capability. Each model, once trained, predicted class probabilities on the independent testing set, and the pROC package was used to calculate the Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values. The model with the highest AUC was selected as the optimal classifier for subsequent interpretability analysis. To further elucidate the biological mechanisms underlying model decisions, SHAP value analysis was performed for both global and local interpretability. The kernelshap method (for high-dimensional features) or permshap method was applied to compute SHAP values for each gene in every training sample, followed by visualization using the shapviz package.

Immune cell infiltration analysis (CIBERSORT) for tumor microenvironment profiling

The composition and proportion of immune cells in the tumor microenvironment directly influence patient prognosis and therapeutic response. In this study, the CIBERSORT algorithm was used to systematically estimate the relative infiltration abundance of 22 human immune cell subsets in ovarian cancer samples. CIBERSORT was run via the online tool or a local R script to obtain the estimated percentages of each immune cell type in every sample. Results with p-value < 0.05 were retained, while unreliable fits were excluded. Stacked bar plots and box plots were generated to visualize the composition of immune cells and compare differences between the control (Control) and treatment (Treat) groups.

Correlation heatmap of immune cells to reveal synergistic or antagonistic relationships

To further understand the interaction network within the immune microenvironment, we calculated the Spearman’s rank correlation coefficients between various immune cell types in the experimental group samples based on the immune cell abundance data output by CIBERSORT. The corrplot package was used to draw the correlation heatmap, where the color gradient indicates the strength of positive and negative correlations. Significant correlations (|r| > 0.5 and p < 0.05) were marked with asterisks. This analysis helps identify potential immune cell co-aggregation patterns or inhibitory interactions, providing clues for understanding the mechanism of immune escape.

Correlation analysis between feature genes and immune cells

Given that ribosomal stress may influence the tumor immune microenvironment, we further investigated the associations between the previously machine learning–selected feature genes and immune cell infiltration levels. Expression data for the feature genes and the corresponding immune infiltration profiles were first extracted, ensuring proper sample matching. Spearman correlation tests were then performed for each “gene-immune cell” pair. The results were visualized using a correlation heatmap generated with the ggplot2 package.

Validation of OC transcriptome using TCGA database

We downloaded RNA-Seq data and comprehensive clinical information from 434 OC tissue samples from the Cancer Genome Atlas (TCGA) database (https://tcga-data.nci.nih.gov/tcga). Initially, we isolated mRNA and lncRNA expression matrices from the TCGA-OV whole transcriptome expression matrix and extracted the expression levels of six characteristic genes from the mRNA matrix. Subsequently, we performed correlation analysis on these matrices to identify feature gene-related lncRNAs (FG-lncRNAs) and obtained the correlation coefficients and P-values from the correlation tests. A correlation coefficient greater than 0 indicates a positive regulatory relationship, whereas a value less than 0 indicates a negative regulatory relationship. Correlation coefficients exceeding 0.3 with P-values less than 0.05 were considered statistically significant.

Construction of an OC risk model based on FG-lncRNAs

First, we extracted the survival time and status of OC patients from the clinical information file, excluding any empty or unknown entries. We then merged the TC-lncRNAs expression matrix with the clinical information data using the R package limma, retaining only the intersecting data. We employed univariate Cox analysis to identify prognostic FG-lncRNAs (PFG-lncRNAs) and performed LASSO regression analysis on these PFG-lncRNAs to construct a regression model. The optimal values for each PFG-lncRNA were determined through cross-validation. Subsequently, we constructed a Cox model using the selected PFG-lncRNAs and optimized it to develop the OC risk model. The risk score formula of the model was derived to differentiate between high-risk and low-risk patients.

Survival and risk analyses of high-risk and low-risk groups

To assess the predictive value of the prognostic risk model, we conducted survival analyses, including overall survival (OS) and progression-free survival (PFS) analyses. We compared the survival differences between the high-risk and low-risk groups and visualized these differences using survival curves. Additionally, to explore the correlations between patient survival time, survival status, PFG-lncRNA expression, and risk score, we performed risk analyses on both high-risk and low-risk groups. The results were visualized using risk curves, survival scatter plots, and risk gene heatmaps.

Independent prognostic analysis of the OC risk model

We conducted an independent prognostic analysis of the OC risk model to determine whether the risk score could serve as an independent prognostic factor for OC patients. We performed univariate and multivariate independent prognostic analyses by integrating risk data (including survival time, survival status, risk score, and PFG-lncRNA expression matrix) and clinical feature lists (including patient age, tumor staging, etc.) from all patients. The results were visualized using a forest plot.

ROC curve and C-index curve analyses of the OC risk model

The Receiver Operating Characteristic (ROC) curve is a comprehensive indicator reflecting the relationship between sensitivity and specificity, two continuous variables, typically represented by a curve. A larger Area under the ROC curve (AUC) indicates higher diagnostic value for the test. We used an R package to draw ROC curves to evaluate the predictive performance of the OC risk model. We integrated survival time, survival status, PFG-lncRNA expression, risk score, and clinical features (patient age, tumor stage) and plotted ROC curves for 1-year, 3-year, and 5-year survival rates, as well as for clinical features. Additionally, we used the concordance index (C-index) to assess the predictive ability of the risk model, calculating C-index values for risk score, survival time, survival status, and various clinical features, and visualizing them through curves.

Analysis of the tumor microenvironment in OC samples

The tumor microenvironment (TME) refers to the intricate relationship between tumor occurrence, growth, metastasis, and the internal and external environments of tumor cells. Using the R packages limma and estimate, we calculated the StromalScore, ImmuneScore, and ESTIMATEScore for high-risk and low-risk samples in the TCGA-OV cohort and visualized them using violin plots.

Sample collection and cell lines and culture

To validate the bioinformatic findings, ovarian cancer tissues and paired normal ovarian epithelial tissues were collected from 45 patients who underwent primary surgical resection at the Department of Obstetrics and Gynecology, People’s Hospital of Ningxia Hui Autonomous Region. All 45 ovarian cancer tissue samples were histopathologically confirmed as high-grade serous ovarian carcinoma (HGSOC), the most prevalent subtype of epithelial ovarian cancer, which is consistent with the predominant tumor subtype represented in the GEO datasets (GSE6008, GSE4122, GSE12470, and GSE66957) used in the bioinformatic analysis, thereby ensuring comparability between the clinical validation cohort and the computational discovery datasets. All patients were pathologically diagnosed with ovarian cancer, and none had received chemotherapy, radiotherapy, or immunotherapy prior to surgery. Tumor tissues were obtained directly from the resected lesions, while the corresponding normal ovarian epithelial tissues were collected from sites located 5 cm away from the edge of the tumor from the tumor margin during the same surgical procedure. All tissue samples were immediately snap-frozen in liquid nitrogen after excision and stored at -80 °C until further molecular analyses. This study was approved by the Institutional Review Board of People’s Hospital of Ningxia Hui Autonomous Region, and written informed consent was obtained from all participants prior to surgery, in accordance with the Declaration of Helsinki.

Human ovarian cancer cell lines A2780(#H-C020, BHcell, China), SKOV3(#ZQ0074, Zhongqiaoxinzhou, China), CAOV-3 (#ZQ0484, Zhongqiaoxinzhou, China), OVCAR3 (#XY-H923, XYBIO, China) and normal ovarian epithelial cells (IOSE80, #GDC-49229, ChemicalBook, China) were obtained from the indicated domestic suppliers and cultured in DMEM (Gibco, USA) supplemented with 10% fetal bovine serum (FBS; Gibco, USA), 100 U/mL penicillin, and 100 µg/mL streptomycin. All cells were maintained at 37 °C in a humidified incubator with 5% CO₂.

RNA extraction and quantitative real-time PCR (RT-qPCR)

Total RNA was extracted from cultured cells and tissue samples (A2780, SKOV3, CAOV-3, OVCAR3, and IOSE80) using TRIzol reagent (Invitrogen, USA) strictly following the manufacturer’s protocol. RNA concentration and purity were assessed using a NanoDrop spectrophotometer (Thermo Fisher Scientific, USA). Complementary DNA (cDNA) was synthesized using a PrimeScript RT Reagent Kit (Takara, Japan). RT-qPCR was performed using SYBR Green Master Mix (Applied Biosystems, USA) on an ABI 7500 Real-Time PCR System. DSP forward primer: 5′-TGAAAACCTGCTGAAAGCGTC-3′; reverse primer: 5′- GCCTCCTGTTTCTGAGCGAT-3′; JUP forward primer: 5′- GGGTGTTGACTGCTGCCCAC-3′; reverse primer: 5′- GAGTTCGGTGGATTGGACTCTT-3′; LCN2 forward primer: 5′- CCCCATCTCTGCTCACTGTC-3′; reverse primer: 5′- TTTTTCTGGACCGCATTG-3′; MAPK13 forward primer: 5′- GATGACTGGCTACGTGGTGA-3′; reverse primer: 5′- AGATGTGCTGCTTCCATTCA-3′; SPP1 forward primer: 5′- AGACCTGACATCCAGTACCCTG-3′; reverse primer: 5′- CCAAGCCCTCCCAGAATTTAA-3′; BMP6 forward primer: 5′- TGGGATGGCAGGACTGGAT-3′; reverse primer: 5′- CAGCATGGTTTGGGGACGTA-3′; GAPDH forward primer: 5′- CGGAGTCAACGGATTTGGTCGTAT-3′; reverse primer: 5′- AGCCTTCTCCATGGTGGTGAAGAC-3′. Relative mRNA expression levels were calculated using the 2^−ΔΔCt method. All reactions were performed in triplicate.

Western blot analysis

Total protein was extracted from cells using RIPA buffer (Beyotime, China) supplemented with protease inhibitors. Protein concentrations were determined by BCA assay (Pierce, USA). Equal amounts of protein were separated by SDS-PAGE and transferred onto PVDF membranes (Millipore, USA). Membranes were blocked with 5% non-fat milk for 1 h and incubated overnight at 4 °C with primary antibodies against BMP6 (1:1000, #AF5196, Affinity) and GAPDH (1:10000, #AB0037, Abways) (as loading control). After washing, membranes were incubated with HRP-conjugated secondary antibodies, goat anti-rabbit IgG(H + L) HRP (1:5000, #S0001, Affinity), for 1 hours at room temperature. Protein bands were visualized using an enhanced chemiluminescence (ECL) detection system (Bio-Rad, USA) and quantified using ImageJ software.

Plasmid construction and cell transfection

The human BMP6 coding sequence was amplified from the cDNA of IOSE80 cells and cloned into the pcDNA3.1 (+) vector. After verification by restriction enzyme digestion and sequencing, the overexpression plasmid (pcDNA3.1-BMP6) and empty vector control were obtained. OVCAR3 and SKOV3 cells were seeded in 6-well plates at 2 × 10⁵ cells/well. When reaching 70–80% confluence, the cells were transfected with 4 µg of plasmids using Lipofectamine 3000. The complete medium was replaced after 6 h. Forty-eight hours post-transfection, the expression levels of BMP6 mRNA and protein were verified by RT-qPCR and Western blot. Cells with expression levels ≥ 3-fold higher than the control were considered successfully transfected and used for subsequent experiments.

Cell proliferation assays and colony formation assays

Cell viability was assessed using the Cell Counting Kit-8 (CCK-8) assay (Dojindo, Japan). Briefly, transfected OVCAR3 and SKOV3 cells were seeded in 96-well plates (3 × 10³ cells/well) and incubated for 24, 48, 72, and 96 h. CCK-8 solution was added to each well and incubated for 2 h at 37 °C. Absorbance at 450 nm was measured using a microplate reader (Bio-Rad, USA). Colony formation assays were performed by seeding 500 transfected cells in 6-well plates and culturing for 10–14 days. Colonies were fixed with 4% paraformaldehyde (Beyotime, China), stained with 0.1% crystal violet (Sigma-Aldrich, USA), and counted manually.

Cell migration and invasion assays

Transwell migration assays were performed using 5 × 10⁴ transfected cells suspended in serum-free medium and seeded into the upper chamber of a Transwell insert (8 μm pore size, Corning, USA). Medium containing 10% FBS was added to the lower chamber as a chemoattractant. After 24 hours of incubation, cells on the upper surface were removed, and migrated cells on the lower surface were fixed, stained with crystal violet, and counted in five random fields under a microscope.

Transwell invasion assays were performed using the procedure similar to that of the migration assay, except that the upper chamber was pre-coated with Matrigel (BD Biosciences, USA) to assess invasive capacity. After 48 h, invaded cells were fixed, stained, and quantified as described above.

Statistical analysis

All analyses were performed in R (v4.4.1) and GraphPad Prism 7 (GraphPad Software, La Jolla, CA). For comparisons between two groups, the Student’s t-test was applied when data followed a normal distribution, the Wilcoxon rank-sum test was used. For multi-group comparisons, one-way ANOVA was applied for normally distributed data, or the Kruskal-Wallis test for non-parametric data. Categorical variables were analyzed by the chi-square test. P < 0.05 was considered statistically significant.

Results

Data preprocessing and normalization achieved high quality

The raw expression matrices from the four GEO datasets were individually log-transformed and normalized the influence of extreme values and background noise was effectively eliminated. The results showed that after correction, the expression value distributions of all datasets became consistent, with smooth density curves concentrated within reasonable ranges, indicating high data quality suitable for subsequent integrative analysis. PCA revealed that, before correction, samples were clearly clustered according to their dataset origins, suggesting a strong batch effect (Fig. 1A). After applying the ComBat algorithm, however, samples clustered mainly according to disease status (Control vs. Treat), and data from different sources were well integrated. This demonstrated effective batch effect correction, ensuring the reliability of the downstream differential expression analysis (Fig. 1B).

Fig. 1.

PCA analysis before and after batch effect correction. A Before correction; B After ComBat correction

Identification of DEGs

Differential expression analysis identified a total of 456 DEGs (adj.P.Value < 0.05, |logFC| > 1), including 267 upregulated and 189 downregulated genes (Fig. 2A). A heatmap of the top 100 most significant DEGs (50 upregulated and 50 downregulated) further demonstrated strong sample discrimination, suggesting their potential biological relevance (Fig. 2B).

Fig. 2.

Differential expression analysis of genes. A Volcano plot; B Heatmap displaying expression patterns of the top 100 DEGs

Identification of DE-RSRGs and enriched in key functional categories and signaling pathways

By intersecting the identified DEGs with RSRGs obtained from the GeneCards database, a total of 45 overlapping genes were identified, defined as DE-RSRGs. The Venn diagram clearly illustrated the overlap between the two gene sets. Although the intersecting region was relatively small, it was highly specific, indicating that these genes are not only aberrantly expressed in ovarian cancer but also directly involved in ribosomal function regulation (Fig. 3A).

Fig. 3.

Identification and functional enrichment of DE-RSRGs. A Venn diagram showing the intersection between DEGs and RSRGs; B Circular plots; C Bubble plots of GO (BP/CC/MF) pathways enriched by DE-RSRGs (q < 0.05); D Bubble plots of KEGG pathways enriched by DE-RSRGs (q < 0.05)

GO enrichment analysis revealed that DE-RSRGs were significantly associated with multiple core biological processes, including response to mechanical stimulus, regulation of cardiac muscle contraction, and cardiomyocyte action potential in BP; intercalated disc, cell-cell contact zone, and myofibril membrane in CC; as well as enzyme inhibitor activity, structural constituent of cytoskeleton, and integrin binding in MF (Fig. 3B and C). KEGG pathway analysis indicated that these genes were involved in classical pathways such as cytoskeleton organization in muscle cells, lipid metabolism and atherosclerosis, osteoclast differentiation, proteoglycans in cancer, and the IL-17 signaling pathway (Fig. 3D).

Machine learning models identify a robust set of feature genes

The three machine learning algorithms independently identified the most discriminative feature genes from the DE-RSRGs expression matrix: Lasso regression selected 18 genes, with the cross-validation error curve reaching its minimum at λ_min, indicating good model generalizability (Fig. 4A and B). mSVM-RFE identified 25 key genes through recursive feature elimination. The cross-validation accuracy initially increased and then decreased as the number of features was reduced, with the peak indicating the optimal feature set (Fig. 4C and D). Random forest ranked genes based on out-of-bag error, yielding the top 15 most important genes (Fig. 4E and F). By comparing the results of the three approaches, six ovarian cancer feature genes were ultimately defined: BMP6, DSP, JUP, LCN2, MAPK13, and SPP1 (Fig. 5A). These genes were not only significantly differentially expressed but also consistently identified as core discriminative factors across multiple algorithms, highlighting their potential as novel biomarkers for ovarian cancer. A box plot illustrated the differential expression of these feature genes between cancer (Treat) and normal (Control) groups: DSP, JUP, LCN2, MAPK13, and SPP1 were upregulated in the Treat group, whereas BMP6 was downregulated compared to controls (Fig. 5B). A circos plot displayed the chromosomal locations of the six feature genes (Fig. 5C).

Fig. 4.

Feature selection using machine learning models. A, B Lasso regression cross-validation error curve and coefficient shrinkage path; C, D SVM-RFE accuracy curve with varying feature numbers and the optimal gene subset; E, F Random forest out-of-bag (OOB) error and gene importance ranking

Fig. 5.

Validation of feature genes. A Venn diagram showing the six feature genes shared by all three machine learning models; B Box plot illustrating differential expression of the feature genes between cancer and normal groups, ***p < 0.001; C Circos plot indicating the chromosomal locations of the feature genes

SHAP analysis reveals the directional contributions of key feature genes

SHAP analysis was employed to investigate the contribution mechanisms of key feature genes in the predictive model. To select the optimal classifier, all 10 previously described machine learning algorithms were evaluated. The resulting ROC curves summarized the discriminative performance of each model (Fig. 6A). The PLS model achieved the highest AUC of 0.968, followed by glmBoost (0.963) and Logistic Regression (0.960). XGBoost and SVM also performed well with AUCs of 0.944 and 0.946, respectively, whereas the DTS model showed the lowest performance (AUC = 0.854). Consequently, all subsequent SHAP interpretability analyses were based on the PLS model. To identify the genes with the greatest impact on predictions, SHAP values were calculated for each sample in the training set and visualized using multiple plots. Bar plots and bee swarm plots indicated that the top six key genes by average SHAP values were: SPP1 (0.0933), DSP (0.0714), MAPK13 (0.0498), BMP6 (0.0352), LCN2 (0.0320), and JUP (0.0181) (Figs. 6B, C). This suggests that SPP1 and DSP are the most influential drivers, potentially playing central roles in ribosomal stress regulation in ovarian cancer.

Fig. 6.

SHAP-based interpretability analysis. A ROC curves comparing the 10 machine learning models; B, C Bar plot and bee swarm plot showing the ranking and directional contributions of gene SHAP values; D Dependence plot illustrating the nonlinear relationship between feature gene expression and SHAP values; E, F Force plot and waterfall plot for a single sample, visualizing the stepwise decision pathway of the prediction

Dependence plots illustrated SHAP values exhibited positive trends for SPP1, DSP, MAPK13, LCN2, JUP consistent with pro-tumor activity, while BMP6 showed negative contributions (Fig. 6D). For individualized interpretation, the first test sample was selected to generate a force plot (Fig. 6E) and a waterfall plot (Fig. 6F), visually displaying the decision pathway for predicting this sample as a “case.” The force plot showed the baseline prediction (Expected Value: E[f(x)] = 0.8) being gradually adjusted to the actual output (f(x) = 0.245), with the largest positive contributions from SPP1 = 4.99 (-0.166) and DSP = 5.5 (-0.161), while other genes such as MAPK13 and LCN2 exhibited minor negative adjustments. The waterfall plot dynamically illustrated the stepwise transition from E[f(x)] = 0.8 to f(x) = 0.245.

Overall immune cell infiltration patterns and group differences in ovarian cancer tissues

CIBERSORT analysis was performed on 172 ovarian cancer samples (including Control and Treat groups) to quantify the relative infiltration levels of 22 immune cell types in the tumor microenvironment. Stacked bar plots indicated that the major immune components included T cells (CD4 + memory T cells, regulatory T cells, γδ T cells), monocytes, macrophages (M0/M1/M2), dendritic cells (DCs), and natural killer (NK) cells. Considerable heterogeneity in immune composition was observed across samples, suggesting substantial inter-individual variability in immune responses (Fig. 7A). Comparisons between the Control and Treat groups revealed statistically significant differences in multiple immune cell subsets (Fig. 7B). Specifically, naive B cells, memory B cells, CD4 naive T cells, CD4 memory resting T cells, T follicular helper cells, monocytes, macrophages M0, macrophages M1, and resting mast cells differed significantly between groups (*p < 0.05, **p < 0.01, ***p < 0.001), with CD4 memory resting T cells and macrophages M1 showing the most pronounced differences. Other subsets, such as regulatory T cells, NK cells, and dendritic cells, showed no significant differences. These findings suggest that ovarian cancer may modulate specific immune cell subsets to influence anti-tumor immune responses.

Fig. 7.

Immune cell infiltration analysis. A Stacked bar plot showing the distribution of 22 immune cell types across samples. B Box plot comparing immune cell subset differences between Control and Treat groups, *p < 0.05

Correlation networks among immune cells and between feature genes and immune infiltrates in ovarian cancer

Spearman correlation analysis was used to construct a correlation heatmap among immune cell types in the Treat group (Fig. 8A). Results indicated that Macrophages M1 were positively correlated with activated NK cells (r = 0.38), suggesting potential co-participation in pro-inflammatory responses, whereas resting mast cells were negatively correlated with activated mast cells, indicating a relatively suppressed state. Clustering analysis grouped immune cells into functional modules, facilitating understanding of the dynamic regulation of the ovarian cancer immune network.

Fig. 8.

Correlation networks. A Spearman correlation heatmap among immune cell types, color gradient: red means positive, blue means negative; B Correlation heatmap between feature genes and immune cell infiltration

We further analyzed correlations between the expression of previously identified feature genes (SPP1, MAPK13, LCN2, JUP, DSP, BMP6) and immune cell infiltration levels (Fig. 8B). Key findings include: SPP1 showed significant positive correlations with memory B cells, Macrophages M0/M2, activated mast cells, and neutrophils (***p < 0.001, *p < 0.05), and negative correlations with naive B cells, resting mast cells, and CD8 + T cells (**p < 0.01, *p < 0.05). MAPK13 was positively correlated with Macrophages M0 and negatively correlated with Tregs and resting mast cells (**p < 0.01, *p < 0.05). LCN2 was positively associated with neutrophils and Tregs (*p < 0.05) and negatively with CD8 + T cells (*p < 0.05). JUP showed positive correlations with memory B cells, Macrophages M0, and Tregs (*p < 0.05). DSP correlated positively with memory B cells and Macrophages M0 (*p < 0.05). BMP6 was strongly positively correlated with resting dendritic cells (***p < 0.001) and negatively with follicular helper T cells (*p < 0.05). Overall, Macrophages M0 emerged as the immune cell type most significantly associated with multiple feature genes (SPP1, MAPK13, JUP, DSP), while SPP1 showed the strongest positive correlation with neutrophils and BMP6 with resting dendritic cells (***p < 0.001). Notably, SPP1 demonstrated significant associations across the widest range of immune cell populations. These gene immune correlations provide insights into potential interactions between specific signaling pathways and the tumor immune microenvironment in ovarian cancer.

Identification of FG-lncRNAs

An expression matrix comprising 60,660 RNA transcripts was downloaded from the TCGA database, including 16,902 lncRNAs and 19,969 mRNAs. A total of 89 FG-lncRNAs were identified through co-expression correlation analysis with six FGs. The expression matrix and interaction relationships of FG-lncRNAs are detailed in the supplementary documents (Supplementary Table S3 and Table S4 ). Finally, the co-expression relationships between FGs and FG-lncRNAs were visualized using Sankey diagrams (Fig. 9).

Fig. 9.

Co-expression relationship between FGs and FG-lncRNAs. The top bar represents FGs, the bottom FG-lncRNAs, with lines indicating co-expression

Prognostic risk model for OC

Univariate Cox analysis was conducted on the 89 FG-lncRNAs, identifying six PFG-lncRNAs. Among them, PTGES2-AS1 was identified as a high-risk lncRNA (Hazard ratio > 1 and p-value < 0.05), while AC004540.1, LNC-LBCS, AL031058.1, AC010636.1, and AC242842.1 were identified as low-risk lncRNAs (Hazard ratio < 1 and p-value < 0.05) (Fig. 10A, Table S5). A Cox model was constructed using LASSO regression (Fig. 10B, C), and the predictive values of the risk model were evaluated and optimized through cross-validation. Ultimately, the prognostic risk model consisted of three PFG-lncRNAs (LNC-LBCS, PTGES2-AS1, and AC242842.1).

Fig. 10.

Construction of OC prognostic risk model. A Forest plot of univariate Cox regression on six PFG-lncRNAs in OC patients. Green box: low HR; red box: high HR; blue line: 95% CI; B, C Optimal parameters selected via LASSO and cross-validation

Survival and risk analyses of the OC prognostic risk model

Survival curves were plotted by comparing the survival differences between all high-risk and low-risk patients. As depicted in Fig. 11A and B, the overall survival rate of patients in the high-risk OC group was significantly lower than that in the low-risk group (P = 0.004), and the progression-free survival rate of patients in the high-risk group was significantly lower than that in the low-risk group (P = 0.026). Furthermore, a risk analysis was performed to evaluate the prognostic risk of patients in the high-risk and low-risk groups. The risk score of the high-risk group was significantly higher than that of the low-risk group (P < 0.05) (Fig. 11C). As the risk score increased, the patient survival rate decreased, and the number of deaths increased (Fig. 11D). As shown in Fig. 11E, LNC-LBCS and AC242842.1 are low-risk PFG-lncRNAs, while PTGES2-AS1 is a high-risk PFG-lncRNA. These results indicate that the prognostic risk model can accurately predict the survival and risk outcomes of the two patient groups.

Fig. 11.

Survival analysis and risk analysis of OC prognostic risk model. A, B Overall and progression-free survival curves for high- (red) and low-risk (blue) OC patients. X-axis: survival time (years); Y-axis: survival rate. Table under curve shows annual survivors. C Risk curve for OC patients. X-axis: risk score (low to high); Y-axis: score. Red/blue dots: high/low-risk. D Risk scatter plot. Y-axis: survival time. Red/blue dots: deceased/living. E Risk heatmap of 3 PFG-lncRNAs. Red/blue squares: high/low expression

Evaluation and validation of the prognostic predictive ability of the risk model

Independent prognostic analysis, ROC curve analysis, and C-index curve analysis were performed to evaluate and validate the prognostic predictive ability of the risk model. As shown in Fig. 12A and B, in both univariate and multivariate logistic analyses, the p-value for the risk score was less than 0.001, and the HR value was greater than 1.8, suggesting that the risk score in the prognostic risk model may be a reliable independent prognostic factor for OC. As shown in Fig. 12C, the AUC values for 1-year, 3-year, and 5-year survival rates were 0.685, 0.586, and 0.620, respectively. Compared with other clinical features, the AUC value of the risk score based on the prognostic risk model was the highest (0.685) (Fig. 12D), and the C-index value was also the highest (Fig. 12E). These results indicate that the prognostic model can accurately and independently predict the prognosis of OC patients.

Fig. 12.

Evaluate and validate the prognostic ability of risk model. A, B Forest plots of univariate and multivariate logistic analyses on OC clinical features. Green/red box: HR; blue line: 95% CI. C–E ROC and C-index curves for risk model. Curves show survival rates (1, 3, 5 years) or clinical features. AUC: area under ROC

12 TME differences between high-risk and low-risk groups

Through an analysis of the differences in the tumor microenvironment (TME) between the high-risk and low-risk groups, we found that the StromalScore, ImmuneScore, and ESTIMATEScore of the high-risk group were significantly higher than those of the low-risk group (P < 0.001) (Fig. 13). These results suggest that the high-risk group is characterized by a more complex tumor microenvironment with greater stromal and immune infiltration, which may reflect a more immunosuppressive yet reactive milieu that contributes to poorer clinical outcomes.

Fig. 13.

The tumor microenvironment (TME) score differences. ImmuneScore, StromalScore, and ESTIMATEScore differences between high- and low-risk groups. Significance: P < 0.05, P < 0.01, P < 0.001

BMP6 is downregulated in ovarian cancer cells

To validate the bioinformatic findings, we examined the expression ofpreviously identified feature genes (SPP1, MAPK13, LCN2, JUP, DSP, BMP6) in normal ovarian epithelial tissues (n = 45) and ovarian cancer tumor tissues (n = 45). RT-qPCR analysis revealed that these target genes (SPP1, MAPK13, LCN2, JUP, DSP) were significantly overexpressed in tumor tissues, whereas BMP6 was significantly underexpressed in tumor tissues compared to normal ovarian epithelial tissues (Fig. 14A). RT-qPCR analysis showed that BMP6 was significantly downregulated in all ovarian cancer cell lines compared to IOSE80 cells (Fig. 14B). Consistently, Western blot analysis confirmed lower BMP6 protein expression in ovarian cancer cells relative to normal ovarian epithelial cells (Fig. 14C). These results collectively indicate that BMP6 is markedly suppressed in ovarian cancer, supporting its potential role as a tumor suppressor.

Fig. 14.

BMP6 is downregulated in ovarian cancer tissues and cells. A RT-qPCR analysis of the expression of 6 feature genes (DSP, JUP, LCN2, MAPK13, SPP1, and BMP6) in normal ovarian tissues and ovarian tumor tissues. B RT-qPCR analysis of BMP6 mRNA levels in normal ovarian epithelial cells (IOSE80) and ovarian cancer cell lines (A2780, SKOV3, CAOV-3, and OVCAR3). C Western blot analysis of BMP6 protein levels in IOSE80 and ovarian cancer cell lines. *p < 0.05, **p < 0.01, ***p < 0.001 versus Normal (A) and IOSE80 groups (B and C)

BMP6 overexpression inhibits ovarian cancer cell proliferation, migration, and invasion

To investigate the functional role of BMP6 in ovarian cancer, BMP6 was ectopically overexpressed in OVCAR3 and SKOV3 cells. RT-qPCR and Western blot assays confirmed efficient BMP6 overexpression at both mRNA and protein levels (Fig. 15A and B). Functional assays demonstrated that BMP6 overexpression significantly suppressed ovarian cancer cell proliferation. CCK-8 assays revealed a marked reduction in cell viability over time compared with Vector groups (Fig. 15C), and colony formation assays further confirmed the decreased clonogenic potential of BMP6-overexpressing cells (Fig. 15D). Moreover, BMP6 overexpression impaired the migratory and invasive capabilities of ovarian cancer cells. Transwell migration assays showed a significant reduction in the number of migrating cells upon BMP6 overexpression (Fig. 15E), while Matrigel-coated Transwell invasion assays demonstrated a similar inhibitory effect on cell invasion (Fig. 15F). Collectively, these findings suggest that BMP6 acts as a tumor suppressor in ovarian cancer by inhibiting cell proliferation, migration, and invasion, consistent with its low expression in ovarian cancer tissues and cell lines.

Fig. 15.

BMP6 overexpression suppresses ovarian cancer cell proliferation, migration, and invasion. A RT-qPCR and B Western blot analyses were performed to verify the overexpression efficiency of BMP6 in ovarian cancer cells (OVCAR3 and CAOV-3). C CCK-8 assays and D colony formation assays were used to assess the effects of BMP6 overexpression on the proliferative capacity of ovarian cancer cells. E Transwell migration and F Matrigel invasion assays were conducted to evaluate the impact of BMP6 overexpression on the migratory and invasive abilities of ovarian cancer cells. ***p < 0.001 versus Vector groups

Discussion

In this study, we systematically investigated the molecular characteristics of RSRGs in OC and their relationship with the tumor immune microenvironment using integrated multi-dataset analysis and multi-machine learning approaches. Through rigorous preprocessing, batch correction, and differential expression analysis across four GEO datasets, we identified 45 DE-RSRGs. Functional enrichment revealed their involvement in cytoskeletal organization, proteoglycan signaling, lipid metabolism, and the IL-17 pathway biological processes that are closely linked to both tumor progression and stress signaling. Subsequent application of three complementary machine learning algorithms (LASSO, SVM-RFE, and Random Forest) robustly identified six feature genes (SPP1, MAPK13, LCN2, JUP, DSP, and BMP6) as the most critical ribosomal stress–related biomarkers in OC. SHAP analysis provided further interpretability of the predictive models, confirming SPP1 and DSP as the most influential determinants in model decisions. Moreover, immune infiltration analysis revealed significant alterations in macrophages, T cells, and mast cells between tumor and normal tissues, with strong correlations between several feature genes (especially SPP1 and MAPK13) and specific immune cell subsets. Collectively, these findings delineate a potential mechanistic link between ribosomal stress dysregulation and the remodeling of the ovarian cancer immune microenvironment.

Ribosomal stress represents a key node connecting nucleolar function to oncogenic transformation [24, 25]. Disruption of ribosome biogenesis activates signaling cascades that engage p53 and other tumor suppressors to maintain genomic stability [26]. Recent studies have demonstrated that ribosomal stress influences cell proliferation, DNA repair, and metabolic homeostasis in multiple cancers [27]. Our data suggest that ribosomal stress related pathways are aberrantly regulated in OC, with DE-RSRGs enriched in cytoskeletal and integrin-binding functions, which are essential for cell adhesion, invasion, and metastasis. The enrichment of IL-17 and lipid metabolism pathways also implies potential crosstalk between inflammatory signaling and metabolic reprogramming, consistent with evidence that IL-17 promotes pro-inflammatory microenvironments and enhances tumor growth in OC [28]. Furthermore, inhibition of the of the IL-17 A/IL-6 axis has been shown to alleviate ovarian pathology in preclinical models [29], reinforcing the significance of IL-17-mediated signaling in OC. These results highlight ribosomal stress not merely as a bystander response but as an active participant in the molecular pathogenesis of OC.

Among the six feature genes, SPP1 emerged as the top-ranked gene by SHAP importance [30]. As a multifunctional glycoprotein involved in extracellular matrix remodeling and immune regulation [31], SPP1 positively correlated with macrophages M0/M2 and neutrophils but negatively with CD8 + T cells, suggesting a central role in immune evasion. From a translational perspective, elevated SPP1 may predict resistance to immune checkpoint inhibitors (ICIs), as M2-enriched immunosuppressive microenvironments are well-established predictors of ICI failure [32, 33]. Given SPP1 as reported association with platinum resistance, risk stratification based on this six-gene signature may further help identify patients requiring alternative therapeutic strategies [34]. Similarly, elevated LCN2, linked to Treg infiltration and neutrophil recruitment, may mark a subgroup responsive to dual innate/adaptive immune-targeting approaches [35]. These hypotheses warrant prospective validation in cohorts with matched treatment data. Notably, BMP6, identified as a downregulated feature gene in OC, was further validated experimentally. RT-qPCR and Western blot confirmed its low expression in multiple ovarian cancer cell lines, and functional assays demonstrated that BMP6 overexpression significantly inhibited proliferation, migration, and invasion of OC cells (OVCAR3 and CAOV-3). These results support BMP6 as a tumor-suppressive RSRG that may counteract pro-tumorigenic signaling and influence the tumor immune microenvironment.

BMP6, as a member of the TGF-β superfamily, signals primarily through the phosphorylation of SMAD1/5/8 and has been reported to exert anti-tumor effects in several cancers [36]. The inhibitory effects of BMP6 on ovarian cancer cell migration and invasion observed in the present study are consistent with a role in suppressing epithelial-to-mesenchymal transition (EMT) [37]. DSP (desmoplakin) and JUP (junction plakoglobin) are key components of desmosomal and adherens junctions, respectively [38]. Dysregulation of these cytoskeletal proteins contributes to altered cell adhesion, mechanical stress responses, and metastatic potential [39]. Both DSP and JUP were significantly upregulated in OC samples and showed strong SHAP contributions, implying potential roles in tumor cell cohesion and mechanical stress adaptation. The enrichment of cytoskeleton-related GO terms further supports this interpretation. Although high expression of SPP1 and DSP is generally associated with tumor progression, in this sample, the combined inhibitory effects of multiple genes outweighed individual pro-tumor signals, leading the model to classify it closer to the “normal” state. This finding suggests that the cancer phenotype may result from multi-pathway interactions rather than being dominated by a single gene.

The use of multi-machine learning approaches combining LASSO, SVM-RFE, and Random Forest, enabled robust feature selection from a high-dimensional dataset, effectively minimizing algorithm-specific bias [40]. The convergence of results across three independent methods underscores the reliability of the identified six-gene signature [41, 42]. Furthermore, the application of SHAP analysis, an emerging framework for model interpretability, provided mechanistic transparency by quantifying the individual contribution of each gene to model predictions. This integration of predictive modeling and biological interpretation represents an important advancement in bioinformatics-based cancer research, bridging computational accuracy with biological plausibility.

The immune infiltration analysis revealed that macrophages, CD4 + memory T cells, and mast cells exhibited significant differences between tumor and normal tissues, indicating extensive immune remodeling in OC. Macrophages, particularly M0 and M2 subtypes, play pivotal roles in tumor progression by secreting cytokines, promoting angiogenesis, and suppressing cytotoxic immune responses [43, 44]. The consistent correlation of multiple feature genes SPP1, DSP, JUP, and MAPK13, with M0 macrophages suggests a potential ribosomal stress driven signaling axis influencing macrophage recruitment or polarization [45, 46]. Moreover, the negative association between SPP1/LCN2 and CD8 + T cells indicates that ribosomal stress related signaling may impair anti-tumor immunity. Based on these observations, we propose a hypothetical working model: ribosomal stress in ovarian cancer cells leads to upregulation of SPP1, DSP, JUP, LCN2, and MAPK13, alongside downregulation of BMP6. Elevated SPP1 promotes M2 macrophage polarization and neutrophil recruitment while suppressing CD8 + T-cell infiltration, thereby fostering an immunosuppressive microenvironment. Simultaneously, BMP6 downregulation may relieve its inhibitory effects on pro-tumorigenic signaling, potentially through derepression of TGF-β/SMAD2/3-mediated EMT and PI3K/AKT pathway activation further facilitating immune evasion and tumor progression. Together, these molecular events collectively reshape the tumor immune microenvironment to promote ovarian cancer progression. Interestingly, IL-17 pathway enrichment among DE-RSRGs and the observed changes in T-cell subsets further support the hypothesis that ribosomal stress may trigger a pro-inflammatory yet immunosuppressive environment, analogous to chronic stress induced inflammation observed in other malignancies [28]. The experimental validation of BMP6 further strengthens this concept, suggesting that restoring tumor-suppressive RSRGs could modulate both tumor cell behavior and immune composition, offering potential therapeutic avenues.

Several limitations of this study should be acknowledged. First, although bioinformatic findings were partially validated through in vitro experiments, the functional characterization was focused on BMP6, and the roles of the remaining five feature genes (SPP1, MAPK13, LCN2, JUP, and DSP) at the protein and cellular levels remain to be experimentally confirmed. Second, although multi-machine learning integration improves robustness, the sample size remains relatively limited, and external validation in independent clinical cohorts is necessary. Third, although significant correlations were observed between feature gene expression and immune cell infiltration abundance, these associations are based on statistical co-expression analyses and do not establish causal relationships. Future studies employing co-culture systems or in vivo models are needed to validate how these feature genes mechanistically regulate immune cell recruitment and polarization in the ovarian cancer microenvironment. And transcriptomic changes of ribosomal stress related genes may not fully reflect functional ribosomal stress in tumor cells. Future studies should use direct indicators, such as nucleolar morphology, rDNA transcription activity, ribosomal protein stoichiometry, or p53 stabilization, to validate these findings.

Conclusion

This study identified six key ribosomal stress related genes in ovarian cancer, including SPP1, MAPK13, LCN2, and BMP6, linking ribosomal stress to immune dysregulation. A hypothetical mechanistic model is proposed in which ribosomal stress-driven upregulation of SPP1 and downregulation of BMP6 collectively promote an immunosuppressive microenvironment characterized by M2 macrophage polarization and CD8 + T-cell exclusion, facilitating ovarian cancer progression. Functional assays confirmed that BMP6 suppresses proliferation, migration, and invasion of ovarian cancer cells. Our integrative framework provides a strategy for exploring stress immune interactions and potential precision immunotherapies in OC.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Xuechuan Han conceived and designed research. Xuechuan Han, Yan Yu conducted experiments. Yang Fan, Miao Zhang analyzed data. Xuechuan Han wrote the manuscript. All authors read and approved the manuscript.

Funding

Study on the Mechanism of the PROM1/Wnt/β-Catenin Signaling Pathway Regulating the Sensitivity of High-Grade Serous Ovarian Cancer to Cisplatin No.2026AAC030670

Data availability

The public gene expression datasets analyzed in this study (GSE6008, GSE4122, GSE12470, GSE66957) can be obtained from the Gene Expression Omnibus (GEO) database. The list of ribosomal stress-related genes used in this study was derived from the GeneCards database (https://www.genecards.org/). The new data generated in this study can be obtained from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of People’s Hospital of Ningxia Hui Autonomous Region, and written informed consent was obtained from all participants prior to surgery, in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to surgery.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.