A machine learning-based basement membrane gene signature model for predicting ovarian cancer survival

Introduction

Ovarian cancer is one of the most common malignancies in women and is characterized by its high invasiveness and a lack of early symptoms, resulting in a high mortality rate (1). Despite significant advances in oncological research achieved over the past few decades, the prognostic prediction for patients with ovarian cancer remains unsatisfactory.

Currently, clinical prognostic evaluation for ovarian cancer relies on two core methods with critical limitations. First, traditional clinicopathological factors [age, Fédération Internationale de Gynécologie et d’Obstétrique stage (FIGO) stage, grade, cancer antigen 125 (CA125)] are easily accessible but lack predictive accuracy due to high tumor heterogeneity and fail to reflect molecular traits (2-4). Second, BRCA1/2 and homologous recombination deficiency (HRD) testing only applies to a small patient subset, cannot fully characterize tumors, and lacks dynamic monitoring capacity (3,4). Thus, a novel, accurate molecular signature-based prognostic model is urgently needed, which is the core focus of this study.

The basement membrane, situated between epithelial and stromal tissues, serves important functions in cellular localization, structural support, and regulation (5). It is also vital for cell adhesion and positioning and for the maintenance of tissue structure. The basement membrane interacts with cell surface receptors, such as integrins, to influence cellular function, morphology, and polarity. These interactions with intracellular signaling molecules regulate critical processes such as cell proliferation, differentiation, and survival (6). Previous studies have shown that the basement membrane plays a significant role in tumor development and invasion. Under normal circumstances, the basement membrane acts as a barrier to restrict cell movement and invasion, limiting tumor spread. However, in malignant tumors, tumor cells alter the structure and composition of the basement membrane, disrupting its integrity and promoting tumor cell infiltration, invasion, and metastasis (7). Moreover, the basement membrane plays a crucial role in shaping the tumor microenvironment (TME), as it provides a supportive structure for tumor growth, facilitates metastasis, and contributes to the evasion of immune surveillance (8). Basement membrane genes regulate the basement membrane through various mechanisms. They control the transcription and expression of basement membrane proteins and participate in posttranslational modifications, such as glycosylation and phosphorylation. Furthermore, basement membrane genes interact with other cellular and extracellular matrix components, contributing to the organization and structural integrity of the basement membrane (9). Aberrant expression of basement membrane genes in ovarian cancer is closely associated with tumor cell infiltration, invasion, and metastasis (10). Furthermore, abnormal basement membrane gene expression have been reported to be associated with aggressive cases and poor prognosis (11,12).

Prognosis prediction and treatment planning for patients with ovarian cancer are complicated by disease heterogeneity and the variability in treatment responses (2). Traditional ovarian cancer prognostic factors, such as age, tumor stage, and grade, have been found to provide limited accuracy in predicting patient outcomes (13). Reliable prognostic models based on molecular markers, identified through comprehensive analysis of genomic and transcriptomic data, can enhance prognostic accuracy (14). Machine learning has proven to be a powerful tool in the development of prognostic models for various types of cancer, accurately predicting patient survival outcomes in breast (15) and colorectal cancer (16). Prognostic models for prostate (17) and lung (18) cancer have also benefited from the integration of clinical, genetic, and imaging data via machine learning techniques. The models for ovarian cancer have incorporated immune-related genes (19), DNA damage response and mitochondrial assessment (20), and metabolism-related signatures (21) to evaluate patient survival. Several studies were published in 2025 that discussed the optimization of prognostic modeling for patients with ovarian cancer. For instance, a four-gene model derived from least absolute shrinkage and selection operator (LASSO) and Cox analysis demonstrated predictive value for survival outcomes in patients with ovarian cancer; however, it has not been subjected to multicenter validation (22). In a 2025 study, Yu et al. built a Surveillance, Epidemiology, and End Results (SEER)-based nomogram to predict 1-, 2-, 3-year overall survival (OS) in epithelial ovarian cancer. It integrated pathological grade, FIGO stage, surgery, radiotherapy, and diagnosis-to-treatment interval—key factors for risk stratification and personalized treatment, offering a practical tool to improve prognosis assessment (23). Other research built a model integrating glycolysis, lipid, choline, and sphingolipid metabolism, serving as a promising biomarker for predicting prognosis, guiding targeted therapy, and analyzing immune landscape in ovarian cancer. It also identified C1QC-positive tumor-associated macrophages (C1QC⁺ TAMs) and FCN1-positive resident tissue macrophages (FCN1⁺ RTMs) as ovarian cancer markers and candesartan/PD-123319 as potential drugs (24). Additionally, the multimodal follicle maturation (FoMu) model was reported to provide superior OS and progression-free survival prediction for patients with high-grade serous ovarian carcinoma (HGSOC) (25). The Cervical Squamous cell carcinoma and Ovarian Adenocarcinoma Risk Grading (CSOARG) model, incorporating cellular senescence-associated and ovarian aging-associated gene sets, was found to be capable of predicting ovarian cancer prognosis and therapeutic responses (26). In another work, it was found that the Artificial Intelligence-Driven Predictive Index (AIDPI), via ensemble machine learning algorithms, outperformed other models (27), and an another study reported that a parthanatos-related gene index could independently predict outcomes and reflect tumor immune features (28). However, to the best of our knowledge, no study has constructed a prognostic model for ovarian cancer based on a series of basal membrane gene expression patterns.

In our study, the objective was to develop a prognostic model for patients with ovarian cancer based on basement membrane gene expression using a machine learning framework. By integrating RNA-sequencing (RNA-seq) data, conducting differential expression analysis, employing feature selection algorithms, applying machine learning techniques, and validating the model with clinical variables and biological features, we sought to generate valuable insights into disease progression, discover novel prognostic biomarkers, and contribute to improved personalized treatment and patient outcomes. Ultimately, a novel marker for the evaluation the survival among patients with ovarian cancer was developed through the creation of a gene signature incorporating metabolism-related genes. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0227/rc).

Methods

Data sources

We obtained RNA-sequencing data from The Cancer Genome Atlas (TCGA) and Genotype-Tissue Expression (GTEx) databases, quantified as fragments per kilobase of transcript per million mapped reads (FPKM) using the expectation-maximization (RSEM) method. The data were downloaded from the UCSC Xena platform (https://toil.xenahubs.net) and transformed as log₂(FPKM + 0.001) for subsequent analyses (29). This dataset integrates transcriptomic expression data from all tumor samples without matched normal samples from the TCGA database (30) and normal control samples in the GTEx database (31). For this analysis, after tumor samples lacking clinical data were excluded, the data corresponding to the ovary tissue, which included 427 tumor samples and 88 normal control samples, were included for subsequent analyses, and the corresponding clinical data for these samples were obtained. Additionally, we downloaded the human gene annotation file (GRCh37.p13) from the GENCODE database (https://www.gencodegenes.org/human/)(32) and extracted gene symbols using this annotation file. We further retrieved 222 basement membrane genes from the basement membraneBASE database (33) and compiled their expression data from the previously obtained dataset and name it “TCGA-OV” for further analysis. Furthermore, to validate the reliability of subsequent models, the GSE30161 dataset was downloaded from the Gene Expression Omnibus (GEO) database (34) (http://www.ncbi.nlm.nih.gov/geo) of the National Center for Biotechnology Information (NCBI). The data related to 58 tumor tissue samples with prognostic information were extracted from this dataset for model validation. The corresponding platform annotation files were used to convert probe IDs into to gene symbols. For different probes corresponding to the same gene symbol, the average value was taken as the expression value for that gene.

Basement membrane gene differential analysis

Genes with zero expression in over 50% of the samples were removed. The package limma (version 3.56.2; http://www.bioconductor.org/packages/2.9/bioc/html/limma.html) in R (The R Foundation for Statistical Computing, Vienna, Austria) was then used to analyze the differential expression between cancer and normal samples via the TCGA-OV dataset. Significant differentially expressed genes were selected based on statistical significance (P value <0.05) and fold change (FC) (|log₂ FC| >0.585).

Identification of tumor-related basement membrane genes

Based on the differential analysis of basement membrane genes and clinical information from tumor samples, the Boruta feature selection algorithm (version 8.0.0; https://cran.r-project.org/web/packages/Boruta/index.html) was used to identify tumor-related basement membrane genes—specifically, to retain only those genes confirmed to be significantly associated with tumor status, thereby distinguishing them from tentative or rejected candidates. With the tumor status serving as the outcome feature, tumor association analysis was performed on all differential genes. Genes with “tentative” or “rejected” status were removed, while genes with “confirmed” status were retained as tumor-associated basement membrane genes. Importance scoring and visualization were then conducted for these genes.

Construction of the basement membrane gene signature (BMGS)

Based on seven machine learning algorithms [extreme gradient boosting (XGBoost), LASSO, ridge, elastic net (Enet), Cox proportional hazards model with likelihood-based boosting (CoxBoost), supervised principal components (SuperPC), and partial least squares regression for Cox models (plsRcox)], the identified genes were combined with tumor sample prognostic survival information to construct a survival model. Benchmark testing was performed, and 10-fold cross-validation was used to select the optimal model parameters in order to identify a hyperparameter tuning model with the best accuracy and lowered risk of overfitting. During this process, the R packages Hmisc (version 5.1-0; https://cran.r-project.org/web/packages/Hmisc/index.html) and survival (version 3.5-5; https://cran.r-project.org/web/packages/survival/index.html) were used to calculate the concordance index (C-index), integrated Brier score (IBS), and area under the curve (AUC) values at 1, 3, 5, and 10 years for each survival model. These indices were used to select the best model, and the data from all ovarian cancer tumor samples and prognostic survival information were integrated to construct the BMGS prognostic model. The gene features of the BMGS model were also assigned importance scores to assess the contribution of each gene to the BMGS model.

Evaluation of the BMGS

To validate the accuracy and reliability of the BMGS model, risk scores for all tumor samples were calculated with the BMGS model. The samples were then divided into high- and low-risk groups based on the median risk score. The association between the grouping of high-risk and low-risk and actual survival prognosis information was assessed via the Kaplan-Meier curve method from the survival package in R. The log-rank test was used to evaluate the significance of survival prognosis between the two groups. Additionally, the AUC values of the receiver operating characteristic (ROC) curves at 1, 3, and 5 years were used to assess the prognostic value of the BMGS model in the TCGA-OV training set and the GSE30161 validation set.

To assess whether the BMGS model and other clinical variables act as independent prognostic factors for ovarian cancer, we first performed univariate Cox regression analysis to examine the associations of risk grouping (BMGS), age, American Joint Committee on Cancer (AJCC) stage, and tumor grade with OS. Variables showing a significant association (P<0.05) were then selected for subsequent multivariate Cox regression analysis, which determined their independence as prognostic predictors. Subsequently, these significantly correlated features, along with the risk grouping, were used to construct a multivariate Cox regression model. Features with a significance level of P<0.05 were considered independent risk factors. Forest plots were generated to visually represent the results of the two rounds of feature selection.

Association between the BMGS and clinical variables

To ascertain the correlation between different clinical features and risk scores, the risk scores were integrated with the clinical data in the training set TCGA-OV samples. The Wilcoxon rank-sum test was then used to calculate the differential P values between different clinical features and risk scores. The clinical variables included age (<60 and ≥60 years), AJCC stage, and tumor grade. Box plots were generated to visualize the differences in risk scores between these clinical variables.

Mutation analysis of BMGS-high and BMGS-low groups in TCGA-OV

Considering that differences in omics data may impact patient prognosis, we downloaded ovarian cancer somatic mutation data from TCGA. For the somatic mutation data, we performed tumor mutation burden (TMB) analysis using the maftools package (35) (version 2.17.0; https://bioconductor.org/packages/release/bioc/html/maftools.html) in R. We calculated the mutation frequency of the top 20 genes and determined the TMB for all ovarian cancer tumor samples. Additionally, we used Genomic Identification of Significant Targets in Cancer version 2.0 (GISTIC 2.0) (36) (https://www.genepattern.org/modules/docs/GISTIC_2.0) to detect copy number alterations between the high-risk and low-risk groups. The aneuploidy score and tumor neoantigen burden (TNB) data were obtained from the study by Thorsson et al. (37).

Relationship between the BMGS and hallmark gene sets

We downloaded 50 tumor hallmark gene sets from the Molecular Signatures Database (MSigDB) (http://www.gsea-msigdb.org/gsea/index.jsp) using the R package IOBR (version 0.99.9; https://iobr.github.io/book/, method=’ssgsea’) and calculated the tumor hallmark enrichment score based on single-set gene set enrichment analysis (ssGSEA). Differential analysis was performed between the high- and low-risk groups based on sample enrichment scores to evaluate the hallmark difference between the two groups delineated by the BMGS.

The relationship between BMGS and the immune microenvironment

To further clarify the relationship between BMGS and the TME, we used the CIBERSORT (38) algorithm for immune microenvironment analysis and TCGA-OV expression data to calculate the immune infiltration scores of the tumor samples. We performed Wilcoxon rank-sum tests to calculate P values of the immune infiltration scores between the high- and low-risk groups. Additionally, we generated box plots of immune cell scores for each sample based on different groupings, aiming to demonstrate the differences in immune infiltration levels between the high- and low-risk groups.

Identification of potential targets and drugs for patients in the BMGS-high group

To identify the potential targets and drugs for high-risk patients, we first obtained a list of 2,249 druggable genes from the study by Huang et al. (39). We then compiled the expression data of these genes in TCGA-OV and performed a correlation analysis between the BMGS risk score and the druggable gene data using a significance threshold of correlation >0.3 and a P <0.05. Additionally, we used the Dependency Map (DepMap) portal to perform CERES scoring on the significantly correlated genes identified from the correlation analysis. Moreover, based on the TCGA-OV data, we conducted differential expression analysis between the BMGS-high and BMGS-low groups with the following thresholds: P<0.05 and |log₂ FC| >1. We selected the top 150 upregulated genes and the top 150 downregulated genes as the BMGS genes. Using the Connectivity Map (CMap) database (20) (https://clue.io/), we predicted CMap scores for each compound based on the BMGS genes. We employed the screening criteria used by Li et al. (40) (CMap score <−90 or >90) to identify potential effective compounds and their corresponding drugs.

Statistical analyses

Continuous variables were compared with the Mann-Whitney test, and categorical variables were compared with the Chi-squared test. Survival was estimated via Kaplan-Meier curves, with P values determined by a log-rank test. Hazard ratios (HRs) were determined through univariable and multivariable Cox regression. Variables with P<0.10 in the univariable regression were included in the multivariable Cox regression. The false-discovery rate (FDR) was used to estimate the significance of differences between the messenger RNA (mRNA) expression levels. All reported P values were two-tailed, and P<0.05 or FDR <0.05 was considered statistically significant. All analyses were performed with SPSS 25.0 (IBM Corp., Armonk, NY, USA) and R version 3.5.2. Graphs were drawn with GraphPad Prism 8 (Dotmatics, Boston, MA, USA) and R version 3.5.2.

Ethics

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Results

Construction of the BMGS prognostic model

The RNA-seq RSEM FPKM dataset of TCGA, and GTEx was downloaded from UCSC Xena, and the transcriptome expression data for ovarian tissue (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) were extracted. Moreover, the clinical data of TCGA-OV (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) were downloaded. Gene symbols were extracted using the GENCODE annotation file, and tumor samples lacking clinical data were excluded. This resulted in gene expression data for 466 samples. Subsequently, 222 basement membrane genes (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) were obtained from the basement membraneBASE database, and the expression data for these genes (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) were extracted. Finally, a dataset, TCGA-OV, was created for further analysis, consisting of expression data for 221 basement membrane genes and 466 samples. Meanwhile, the GSE30161 dataset was downloaded from the GEO database, and samples with clinical information were extracted, resulting in expression data for 58 samples (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip).

Differential expression analysis of basement membrane genes in TCGA-OV datasets between cancer and normal tissues (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) identified 181 differentially expressed genes, including 68 upregulated and 113 downregulated ones. A volcano plot (Figure 1A) was generated based on the differential analysis results. The Boruta algorithm was applied to further identify tumor-related genes from the 181 differentially expressed basement membrane genes, resulting in a final set of 77 genes (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Additionally, gene importance scores were calculated with the Boruta algorithm (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), and an importance box plot (Figure 1B) was created for the tumor-related genes.

Figure 1 Gene-Signature Prognostic Model for ovarian cancer. (A) Volcano plot of differentially expressed genes in TCGA-OV, where blue represents downregulated genes and red represents upregulated genes. Black or grey dots represent genes that do not meet the screening thresholds for both fold change and statistical significance. (B) Box plot showing the gene importance scores obtained from the Boruta algorithm. (C) Box plot displaying the C-index values and integrated Brier score for seven machine learning algorithms. (D) Bar chart illustrating the importance scores of major genes in the fitted prognostic model. (E) Bar chart illustrating the importance scores of the major genes in the BMGS model. AUC, area under the curve; BMGS, basement membrane gene signature; CoxBoost, Cox proportional hazards model with likelihood-based boosting; Enet, elastic net; LASSO, least absolute shrinkage and selection operator; plsRcox, partial least squares regression for Cox models; SuperPC, supervised principal components; TCGA, The Cancer Genome Atlas; XGBoost, extreme gradient boosting.

A benchmark test was conducted on the 77 selected signature genes via seven machine learning algorithms. The performance of the models was evaluated and compared with metrics including the C-index, IBS, and AUC values for 1-, 3-, 5-, and 10-year predictions (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). The results are visualized in Figure 1C,1D. It was observed that the XGBoost model achieved the highest C-index score and the lowest IBS score. It also demonstrated excellent performance in the 1-, 3-, and 5-year survival prediction tasks. Meanwhile, several other models showed poor performance in the 10-year AUC prediction, indicating that there might be a significant impact of various potential risk factors on longer-term survival, affecting the data analysis. Finally, the optimal parameters for the XGBoost model were chosen. The model was then fitted with the tumor sample expression data and survival information from TCGA-OV. Additionally, the contribution values of each gene feature to the model were calculated. Genes with a contribution value (importance) below 0.01 were removed, resulting in a set of 39 major signature genes (Figure 1E).

Evaluation of the BMGS prognostic model

We reconstructed the BMGS prognostic model integrating the TCGA-OV dataset and the 39 major signature genes. Subsequently, the BMGS model was used to calculate the risk scores for both the TCGA-OV tumor samples and the validation set samples from GSE30161 (tables available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Considering the poor performance of the models in predicting 10-year survival, we removed all samples with a survival period greater than 5 years (tables available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip).

Based on the median risk score of the samples, the TCGA-OV training set samples and GSE30161 validation set samples were divided into high- and low-risk groups. The association between the high- and low-risk groups and the actual patient outcomes was evaluated via the Kaplan-Meier method. The results showed a significant decrease in OS in the high-risk group compared to the low-risk group in both the training set [HR =6.02; 95% confidence interval (CI): 4.21–8.62; P<0.001] and the validation set (HR =2.53; 95% CI: 1.15–5.57; P=0.02) (Figure 2A). Furthermore, according to the survival time, survival status, and risk score values of the samples in the TCGA training set and GEO validation set, ROC curves were plotted to predict 1-, 3-, and 5-year survival. In the training set, the AUC values for 1-, 3-, and 5-year predictions were all above 0.9, indicating good predictive accuracy of the constructed model. In the validation set, the AUC values for all three time points were distributed between 0.6 and 0.8, also indicating good predictive value (Figure 2B).

Figure 2 Evaluation of BMGS prognostic model. (A) Prognosis-related K-M curve chart based on the risk score prediction model (left: TCGA training set; right: GEO validation set). (B) The ROC curve of the model for predicting 1-, 3-, 5-year survival in the sample (left: TCGA training set; right: GEO validation set). (C) Forest plot for the single-factor and multifactor analysis for the TCGA training set and GEO validation (top panels: multivariate analysis results; bottom panels: univariate analysis results). AUC, area under the curve; BMGS, basement membrane gene signature; CI, confidence interval; GEO, Gene Expression Omnibus; HR, hazard ratio; K-M, Kaplan-Meier; OV, ovarian cancer; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas.

We further integrated clinical features from both the TCGA training set and the GEO validation set, including age, AJCC stage, and grade. We performed univariate and multivariate Cox regression analyses separately (tables available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Variables with a P value <0.05 were selected (tables available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), and a forest plot was generated. It was observed that in both the TCGA training set and the GEO validation set, the BMGS model was an independent risk factor, demonstrating the prognostic accuracy of the BMGS model (Figure 2C). We also conducted statistical integration of the distribution of clinical information in TCGA-OV tumor samples (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip) and generated box plots to clarify the levels of risk scores under different clinical feature grouping conditions. We calculated the correlation using the rank-sum test. The results showed that age was a relevant factor in risk scoring. Patients with an age >60 years had significantly higher overall risk scores than did those with an age <60 years (Figure S1).

Analysis of genetic variation in the high- and low-risk groups

An analysis of somatic mutations between the two groups was performed, and the differences in mutation frequencies for each gene were calculated via the Chi-squared test. Significant differences between the two groups were found in the mutations of TTN, VCAN, and DGKH (Figure 3A). The somatic copy number variations (CNVs) between the high-risk and low-risk groups were detected via GISTIC 2.0. The gene amplification and deletion patterns were separately analyzed for both groups (Figure S2), and it was found that the high-risk group exhibited higher frequencies of amplifications and deletions than did the low-risk group (Figure 3B). These findings have important implications for the research and personalized treatment of ovarian cancer. The TMB values were calculated using maftools R package, and the tumor aneuploidy scores and TNB scores were obtained from the referenced literature (41) (tables available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). The results revealed no significant differences in TMB scores or TNB scores between the risk groups (Figure 3B).

Figure 3 Mutation analysis of high- and low-risk groups in TCGA-OV. (A) Waterfall plot of common somatic mutation frequencies between the BMGS high- and low-risk groups. *, 0.01≤P≤0.05; **, 0.001≤P<0.01; ‘.’ indicates a P value between 0.05 and 0.1 (0.05<P≤0.1); “-” indicates a P value greater than 0.1 (P>0.1). (B) Violin plot of differences in genomic mutation characteristics between the BMGS high- and low-risk groups. ***, P<0.001; -, not significant. BMGS, basement membrane gene signature; TCGA, The Cancer Genome Atlas; TMB, tumor mutation burden; TNB, tumor neoantigen burden.

Correlation analysis between TTN mutation status and TMB values was performed in the TCGA-OV cohort, and the results showed a significant positive correlation between the two indicators (Figure S3A, table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Based on CIBERSORT immune infiltration analysis, we further compared immune cell infiltration differences between TTN-mutant and TTN-wild-type ovarian cancer samples. The findings revealed that TTN mutations were significantly associated with reduced infiltration of monocytes and increased infiltration of M1 macrophages (Figure S3B, table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), confirming that TTN mutations directly regulate the ovarian cancer immune microenvironment in a TMB-correlated manner, rather than independently of TMB levels.

Analysis of pathways and immune microenvironment in the high- and low-risk groups

ssGSEA analysis was performed on the tumor hallmarks of the high- and low-risk groups in the BMGS (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). The heatmap illustrates the distribution of enrichment scores for tumor hallmarks between the two risk groups. It can be observed that several hallmark processes, such as myogenesis, apical surface, hedgehog signaling, and apical-junction, exhibit significant differences between the two risk groups (Figure 4A). Pathway analysis revealed that the BMGS-high group of ovarian cancer exhibited lower cellular differentiation, potentially leading to uncontrolled proliferation and invasive behavior. Loss of cellular polarity and abnormal activation of the hedgehog signaling pathway play important roles in the metastasis of ovarian cancer (42). Disruption of cell adhesion may also promote tumor progression and metastasis. Aberrant regulation of cell differentiation and cell adhesion is closely associated with the prognosis of patients with ovarian cancer (43). The CIBERSORT algorithm was used to estimate the relative infiltration abundance of immune cell types in each sample (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Subsequently, the Wilcoxon test was employed to compare and calculate the significance of immune cell differences between the two risk groups. It was found that three immune cell types exhibited significant differences (P value <0.05) between the two groups. Significantly higher infiltration of T-helper lymphocytes and activated dendritic cells was observed in the low-risk group, while an increase in neutrophil infiltration was observed in the high-risk group (Figure 4B).

Figure 4 Analysis of pathways and immune microenvironment in the high- and low-risk groups. (A) Heatmap of tumor hallmark enrichment scores in the high- and low-risk groups. Red indicates an upregulation trend, and blue indicates a downregulation trend. (B) Box plot for the relative abundance of distribution of immune cells estimated by the CIBERSORT algorithm. *, 0.01≤P≤0.05; **, 0.001≤P<0.01; ***, P<0.001; -, not significant.

Potential targets and drug identification for patients in the BMGS-high group

A total of 2,249 druggable targets were identified (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), of which 2,198 were expressed in TCGA-OV samples. Spearman correlation analysis was performed between the BMGS risk scores and the expression of these 2,198 druggable targets (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), resulting in the identification of 6 BMGS-related targets (correlation coefficient >0.3; P<0.05). The scatter plot in Figure 5A illustrates the distribution of these correlations. The six genes were further assessed for dependency based on CERES scores (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip). Among them, ITGA5 and NUAK1 had the smallest CERES scores among all samples, indicating that these two are the most likely potential therapeutic targets for patients at high risk according to the BMGS. Inhibiting the function of these two genes may achieve good therapeutic effects. The MMP17 target was found to be less prominent in driving ovarian cancer than were the other five targets (Figure 5B).

Figure 5 Potential targets and drug identification for patients in the BMGS-high group. (A) Scatter plot of Spearman correlation analysis between BMGS risk score and druggable gene expression. (B) Box plot of CERES score distribution of BMGS-related target genes. (C) Rose plot of CMap analysis of the differential genes in the high-risk group corresponding to drugs. BMGS, basement membrane gene signature; CAR, constitutive androstane receptor; CMap, connectivity Map; EGFR, epidermal growth factor receptor; MDM, mouse double minute.

The differential expression patterns of these two genes between ovarian cancer tissues and normal ovarian tissues were validated using the TCGA-OV and GTEx databases, with corresponding box plot results supplemented (Figure S4A,S4B). Subsequently, univariate and multivariate Cox regression analyses were performed, and the results confirmed that high expression levels of ITGA5 and NUAK1 are independent adverse prognostic factors for ovarian cancer patients (Figure S4C).

Additionally, a differential analysis was conducted between the high- and low-risk groups in TCGA-OV tumor samples via the R limma package (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), resulting in the identification of 252 upregulated genes and 149 downregulated genes. The top 150 upregulated genes and the 149 downregulated genes were selected as the BMGS genes. The CMap database was used to determine the CMap scores for each compound. The criteria for selecting effective compounds was a CMap score <−90 or a CMap score >90. Consequently, 18 potentially effective compounds were identified (table available at https://cdn.amegroups.cn/static/public/tcr-2026-1-0227-1.zip), corresponding to a total of 16 predicted candidate drugs. The top 10 candidates included ATPase inhibitors, histone lysine methyltransferase inhibitors, epidermal growth factor receptor (EGFR) inhibitors, constitutive androstane receptor (CAR) agonists, vitamin K antagonists, progestins, dipeptidyl peptidase inhibitors, aromatase inhibitors, and mouse double minute (MDM) inhibitors (Figure 5C).

Discussion

Ovarian cancer is a common malignant tumor, and studies have shown that basement membrane genes play an important role in its development and progression (44-46). We constructed a prognostic model for ovarian cancer based on key basement membrane signature genes, termed the BMGS model, which can effectively evaluate the prognosis of patients with ovarian cancer. By performing risk scoring on TCGA-OV data and the validation set GSE30161, we divided the samples into high- and low-risk groups and found a correlation between the high-risk group and poor prognosis. In the correlation analysis with other factors, it was found that patients with ovarian cancer and over the age of 60 years have higher BMGS risk scores. Moreover, the BMGS risk score was independent of tumor stage and grade. The BMGS model is primarily based on gene expression data, reflecting the molecular characteristics and gene expression patterns of the tumor. The association with age may be due to the influence of age on tumor gene expression and molecular features (47). In contrast, tumor stage and grade are commonly used clinical indicators and based on pathological features such as tumor size, depth of invasion, and lymph node metastasis. Overall, in the univariable and multivariable Cox regression analyses, the BMGS model was considered an independent risk factor for ovarian cancer prognosis, further confirming its reliability in the prognostic assessment of ovarian cancer.

Previous studies have reported that aneuploidy is an adverse prognostic factor in cancers such as breast cancer and colorectal cancer (CRC). One study found that aneuploidy G2 breast cancer was associated with worse disease-free survival and OS (P=0.001 and P<0.001, respectively) and that aneuploidy G1 cancer was associated with worse OS (P=0.01) (48). Meanwhile, in a meta-analysis involving 7,072 patients, it was found that late-stage CRC had a higher frequency of aneuploidy compared to early-stage CRC (odds ratio =1.51; 95% CI: 1.37–1.67; P<0.001). Regardless of tumor stage, the overall range of aneuploidy was 39–81% (median 58%), and 21 (54.1%) studies reported a significant prognostic impact of aneuploidy on OS, disease-specific survival, and recurrence-free survival (49). In our study, we also found that chromosomal-level variations, and not point mutations, may be a contributing factor to the poorer prognosis in patients with ovarian cancer. It was found that in the BMGS-high group, the aneuploidy score was significantly higher, but there was no significant difference in TMB or TNB score between the two groups. TMB and TNB scores primarily reflect the burden of gene mutations, while the aneuploidy score is more directly associated with chromosomal-level variations. Chromosome instability can promote tumor genesis by increasing genetic heterogeneity and promoting tumor evolution (50). Research has shown that tumor aneuploidy provides independent prognostic value among patients with lower TMB tumors treated with immunotherapy, with a higher aneuploidy score being associated with a poor prognosis (51).

Through analysis of somatic mutation, it was found that the mutation rates of the TTN genes were significantly higher in the BMGS-high group. The TTN encodes titin, which may have an impact on the proliferative, migratory, and invasive abilities of tumor cells. Research has shown that TTN mutation can serve as an independent prognostic factor in patients with lung squamous cell carcinoma (HR =0.64, 95% CI: 0.48–0.85; P=0.001) and is significantly associated with the enrichment of M1 macrophages (P<0.05) (52). Similarly, we observed higher infiltration of T helper cells and activated dendritic cells in the low-risk group. T helper cells play a crucial role in antitumor immune response by interacting with B lymphocytes to promote antibody production and immune memory formation (53). Meanwhile, activated dendritic cells are important antigen-presenting cells that can recognize and present antigens, thereby activating other immune cells (54). The higher levels of infiltration of these two cells in the low-risk group indicate enhanced antitumor immune activity and are associated with a better prognosis. Conversely, an increase in neutrophils was observed in the high-risk group. In the TME, neutrophils excessively accumulate around the tumor and produce anti-inflammatory factors and immunosuppressive molecules, thereby inhibiting the activity of immune cells, promoting immune tolerance, and providing support for tumor growth and metastasis (55).

As a high-frequency passenger gene in tumors, the prognostic mechanism of TTN mutation in ovarian cancer has not been fully clarified. Our findings confirm that TTN mutation is not correlated with TMB, but directly reshapes the tumor immune microenvironment by regulating immune cell infiltration, which provides a novel mechanistic clue for interpreting the poor prognosis of TTN-mutant patients in the high-BMGS risk group.

The BMGS model, while promising, has limitations that need to be considered. Its generalizability across different populations and cancer subtypes may vary, and it may not fully capture the complex interactions and dynamic nature of tumor biology. Moreover, the model’s reliance on a predefined set of genes might overlook other relevant molecular markers, and its applicability in the context of clinical heterogeneity remains to be more thoroughly evaluated. Ongoing research efforts should focus on refining the model by incorporating additional molecular features, validating it in diverse populations, and considering the dynamic nature of cancer biology to enhance its predictive accuracy and clinical utility.

Conclusions

The BMGS model constructed in this study can effectively evaluate the prognosis of patients’ ovarian cancer. Differences in aneuploidy score, TTN gene mutation frequency, hallmark enrichment, and immune cell infiltration may reveal the underlying reasons for the BMGS risk score’s association with poor prognosis. Additionally, we found several predicted candidate drugs screened via database correlation analysis that may have clinical significance through database analysis, but the findings need to be verified in future research. ITGA5 and NUAK1 are predicted as potential therapeutic targets solely based on bioinformatics analyses, and all candidate targets and drugs require further in vitro/in vivo experimental verification and clinical trial validation before clinical application.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0227/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0227/prf

Funding: This work was supported by the Central Government Guidance Fund for Local Science and Technology Development (No. 202407AB110013).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0227/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Liberto JM, Chen SY, Shih IM, et al. Current and Emerging Methods for Ovarian Cancer Screening and Diagnostics: A Comprehensive Review. Cancers (Basel) 2022;14:2885. [Crossref] [PubMed]
2. Veneziani AC, Gonzalez-Ochoa E, Alqaisi H, et al. Heterogeneity and treatment landscape of ovarian carcinoma. Nat Rev Clin Oncol 2023;20:820-42. [Crossref] [PubMed]
3. Ittner E, Swenson H, Werner L, et al. Diagnostic and prognostic biomarkers associated with histotype in advanced epithelial ovarian cancer. Sci Rep 2025;15:37171. [Crossref] [PubMed]
4. Quesada S, Penault-Llorca F, Matias-Guiu X, et al. Homologous recombination deficiency in ovarian cancer: Global expert consensus on testing and a comparison of companion diagnostics. Eur J Cancer 2025;215:115169. [Crossref] [PubMed]
5. Yurchenco PD. Basement membranes: cell scaffoldings and signaling platforms. Cold Spring Harb Perspect Biol 2011;3:a004911. [Crossref] [PubMed]
6. Pozzi A, Yurchenco PD, Iozzo RV. The nature and biology of basement membranes. Matrix Biol 2017;57-58:1-11. [Crossref] [PubMed]
7. Chang J, Chaudhuri O. Beyond proteases: Basement membrane mechanics and cancer invasion. J Cell Biol 2019;218:2456-69. [Crossref] [PubMed]
8. Khalaf K, Hana D, Chou JT, et al. Aspects of the Tumor Microenvironment Involved in Immune Resistance and Drug Resistance. Front Immunol 2021;12:656364. [Crossref] [PubMed]
9. Miner JH. Basement membranes. In: The extracellular matrix: An overview. Berlin: Springer; 2010:117-45.
10. Zong S, Xu PP, Xu YH, et al. A bioinformatics analysis: ZFHX4 is associated with metastasis and poor survival in ovarian cancer. J Ovarian Res 2022;15:90. [Crossref] [PubMed]
11. Patel V, Aldridge K, Ensley JF, et al. Laminin-gamma2 overexpression in head-and-neck squamous cell carcinoma. Int J Cancer 2002;99:583-8. [Crossref] [PubMed]
12. Yamamoto H, Itoh F, Iku S, et al. Expression of the gamma(2) chain of laminin-5 at the invasive front is associated with recurrence and poor prognosis in human esophageal squamous cell carcinoma. Clin Cancer Res 2001;7:896-900.
13. Bookman MA. Can we predict who lives long with ovarian cancer? Cancer 2019;125:4578-81. [Crossref] [PubMed]
14. Dienstmann R, Villacampa G, Sveen A, et al. Relative contribution of clinicopathological variables, genomic markers, transcriptomic subtyping and microenvironment features for outcome prediction in stage II/III colorectal cancer. Ann Oncol 2019;30:1622-9. [Crossref] [PubMed]
15. Ferroni P, Zanzotto FM, Riondino S, et al. Breast Cancer Prognosis Using a Machine Learning Approach. Cancers (Basel) 2019;11:328. [Crossref] [PubMed]
16. Zhang Z, Huang L, Li J, et al. Bioinformatics analysis reveals immune prognostic markers for overall survival of colorectal cancer patients: a novel machine learning survival predictive system. BMC Bioinformatics 2022;23:124. [Crossref] [PubMed]
17. Tătaru OS, Vartolomei MD, Rassweiler JJ, et al. Artificial Intelligence and Machine Learning in Prostate Cancer Patient Management-Current Trends and Future Perspectives. Diagnostics (Basel) 2021;11:354. [Crossref] [PubMed]
18. Li Y, Wu X, Yang P, et al. Machine Learning for Lung Cancer Diagnosis, Treatment, and Prognosis. Genomics Proteomics Bioinformatics 2022;20:850-66. [Crossref] [PubMed]
19. Dong L, Qian YP, Li SX, et al. Development of a machine learning-based signature utilizing inflammatory response genes for predicting prognosis and immune microenvironment in ovarian cancer. Open Med (Wars) 2023;18:20230734. [Crossref] [PubMed]
20. Walker TDJ, Faraahi ZF, Price MJ, et al. The DNA damage response in advanced ovarian cancer: functional analysis combined with machine learning identifies signatures that correlate with chemotherapy sensitivity and patient outcome. Br J Cancer 2023;128:1765-76. [Crossref] [PubMed]
21. Fu Y, Huang Z, Huang J, et al. Metabolism-related gene vaccines and immune infiltration in ovarian cancer: A novel risk score model of machine learning. J Gene Med 2024;26:e3568. [Crossref] [PubMed]
22. Hu J, Su M, Qin Z, et al. DNA methylation and transcription factor-driven immune subtypes in ovarian cancer. Discov Oncol 2025;16:1646. [Crossref] [PubMed]
23. Yu X, Fan Y, Liu Y. Establishment of a predictive survival prognosis model for epithelial ovarian cancer: a study based on the SEER database. Eur J Gynaecol Oncol 2025;46:70-9.
24. Ling L, Li B, Ke B, et al. Metabolism-associated marker gene-based predictive model for prognosis, targeted therapy, and immune landscape in ovarian cancer: an integrative analysis of single-cell and bulk RNA sequencing with spatial transcriptomics. BMC Womens Health 2025;25:233. [Crossref] [PubMed]
25. Bi Q, Ai C, Qu L, et al. Foundation model-driven multimodal prognostic prediction in patients undergoing primary surgery for high-grade serous ovarian cancer. NPJ Precis Oncol 2025;9:114. [Crossref] [PubMed]
26. Chen J, Guan B, Zhang J, et al. Development of CSOARG: a single-cell and multi-omics-based machine learning model for ovarian cancer prognosis and drug response prediction. Front Oncol 2025;15:1592426. [Crossref] [PubMed]
27. Wu Y, Wang K, Song Y, et al. Enhancing ovarian cancer prognosis with an artificial intelligence-derived model: Multi-omics integration and therapeutic implications. Transl Oncol 2025;59:102439. [Crossref] [PubMed]
28. Qi X, Jiang C, Wang N, et al. Identification of parthanatos-related molecular subtypes and development of prognostic risk models in ovarian cancer based on multi-omics analysis. J Ovarian Res 2025;18:137. [Crossref] [PubMed]
29. Goldman MJ, Craft B, Hastie M, et al. Visualizing and interpreting cancer genomics data via the Xena platform. Nat Biotechnol 2020;38:675-8. [Crossref] [PubMed]
30. Comprehensive genomic characterization defines human glioblastoma genes and core pathways. Nature 2008;455:1061-8.
31. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 2020;369:1318-30.
32. Frankish A, Diekhans M, Jungreis I, et al. GENCODE 2021. Nucleic Acids Res 2021;49:D916-23. [Crossref] [PubMed]
33. Jayadev R, Morais MRPT, Ellingford JM, et al. A basement membrane discovery pipeline uncovers network complexity, regulators, and human disease associations. Sci Adv 2022;8:eabn2265. [Crossref] [PubMed]
34. Barrett T, Wilhite SE, Ledoux P, et al. NCBI GEO: archive for functional genomics data sets--update. Nucleic Acids Res 2013;41:D991-5. [Crossref] [PubMed]
35. Mayakonda A, Lin DC, Assenov Y, et al. Maftools: efficient and comprehensive analysis of somatic variants in cancer. Genome Res 2018;28:1747-56. [Crossref] [PubMed]
36. Reich M, Liefeld T, Gould J, et al. GenePattern 2.0. Nat Genet 2006;38:500-1. [Crossref] [PubMed]
37. Thorsson V, Gibbs DL, Brown SD, et al. The Immune Landscape of Cancer. Immunity 2018;48:812-830.e14. [Crossref] [PubMed]
38. Newman AM, Liu CL, Green MR, et al. Robust enumeration of cell subsets from tissue expression profiles. Nat Methods 2015;12:453-7. [Crossref] [PubMed]
39. Huang RH, Hong YK, Du H, et al. A machine learning framework develops a DNA replication stress model for predicting clinical outcomes and therapeutic vulnerability in primary prostate cancer. J Transl Med 2023;21:20. [Crossref] [PubMed]
40. Li C, Liu Z. Bioinformatic Analysis for Potential Biomarkers and Therapeutic Targets of T2DM-related MI. Int J Gen Med 2021;14:4337-47. [Crossref] [PubMed]
41. Yu L, Zhang Y, Wang D, et al. Harmonizing tumor mutational burden analysis: Insights from a multicenter study using in silico reference data sets in clinical whole-exome sequencing (WES). Am J Clin Pathol 2024;162:408-19. [Crossref] [PubMed]
42. Zhang H, Wang Y, Chen T, et al. Aberrant Activation Of Hedgehog Signalling Promotes Cell Migration And Invasion Via Matrix Metalloproteinase-7 In Ovarian Cancer Cells. J Cancer 2019;10:990-1003. [Crossref] [PubMed]
43. Lee JG, Ahn JH, Jin Kim T, et al. Mutant p53 promotes ovarian cancer cell adhesion to mesothelial cells via integrin β4 and Akt signals. Sci Rep 2015;5:12642. [Crossref] [PubMed]
44. Roland IH, Yang WL, Yang DH, et al. Loss of surface and cyst epithelial basement membranes and preneoplastic morphologic changes in prophylactic oophorectomies. Cancer 2003;98:2607-23. [Crossref] [PubMed]
45. Diao B, Yang P. Comprehensive Analysis of the Expression and Prognosis for Laminin Genes in Ovarian Cancer. Pathol Oncol Res 2021;27:1609855. [Crossref] [PubMed]
46. Yang WL, Godwin AK, Xu XX. Tumor necrosis factor-alpha-induced matrix proteolytic enzyme production and basement membrane remodeling by human ovarian surface epithelial cells: molecular basis linking ovulation and cancer risk. Cancer Res 2004;64:1534-40. [Crossref] [PubMed]
47. Li CH, Haider S, Boutros PC. Age influences on the molecular presentation of tumours. Nat Commun 2022;13:208. [Crossref] [PubMed]
48. Pinto AE, Pereira T, Silva GL, et al. Aneuploidy identifies subsets of patients with poor clinical outcome in grade 1 and grade 2 breast cancer. Breast 2015;24:449-55. [Crossref] [PubMed]
49. Laubert T, Freitag-Wolf S, Linnebacher M, et al. Stage-specific frequency and prognostic significance of aneuploidy in patients with sporadic colorectal cancer--a meta-analysis and current overview. Int J Colorectal Dis 2015;30:1015-28. [Crossref] [PubMed]
50. Ben-David U, Amon A. Context is everything: aneuploidy in cancer. Nat Rev Genet 2020;21:44-62. [Crossref] [PubMed]
51. Spurr LF, Weichselbaum RR, Pitroda SP. Tumor aneuploidy predicts survival following immunotherapy across multiple cancers. Nat Genet 2022;54:1782-5. [Crossref] [PubMed]
52. Zou S, Ye J, Hu S, et al. Mutations in the TTN Gene are a Prognostic Factor for Patients with Lung Squamous Cell Carcinomas. Int J Gen Med 2022;15:19-31. [Crossref] [PubMed]
53. Kennedy R, Celis E. Multiple roles for CD4+ T cells in anti-tumor immune responses. Immunol Rev 2008;222:129-44. [Crossref] [PubMed]
54. Schmidt SV, Nino-Castro AC, Schultze JL. Regulatory dendritic cells: there is more than just immune activation. Front Immunol 2012;3:274. [Crossref] [PubMed]
55. Li L, Yu R, Cai T, et al. Effects of immune cells and cytokines on inflammation and immunosuppression in the tumor microenvironment. Int Immunopharmacol 2020;88:106939. [Crossref] [PubMed]

(English Language Editor: J. Gray)