Construction of prognostic risk prediction model of endometrial carcinoma based on bioinformatics analysis

Abstract

This study developed a prognostic risk prediction model for endometrial carcinoma (EC) by integrating data from The Cancer Genome Atlas and Gene Expression Omnibus for bioinformatics analysis. The relevant data of EC were downloaded from The Cancer Genome Atlas database and the GSE17025 dataset of the Gene Expression Omnibus database. Based on the R language, the differentially expressed genes (DEGs) and weighted gene co-expression network analysis were used to identify the gene modules with the strongest correlation with clinical features, and intersected with the DEGs of GSE17025 dataset. Subsequently, univariate and multivariate Cox regression analyses were conducted to construct and validate a prognostic risk prediction model for EC. Weighted gene co-expression network analysis identified 6 gene modules, with the turquoise module exhibiting the strongest correlation with EC prognosis and survival. By intersecting with DEGs from GSE17025 dataset, 65 candidate genes were identified. Univariate Cox regression revealed 19 genes significantly associated with overall survival, and multivariate Cox regression identified 5 prognostic genes. A 5-gene risk prediction model, including PDZ domain containing ring finger 3, KN motif and ankyrin repeat domains 4, prion protein, phosphoserine aminotransferase 1, and Annexin A1, was constructed. Kaplan–Meier survival curve analysis demonstrated that patients in the high-risk group had significantly lower overall survival compared to the low-risk group (P < .001). The ROC curve confirmed the model’s robust prognostic predictive performance. This study presents a 5-gene prognostic risk prediction model for EC, including PDZ domain containing ring finger 3, KN motif and ankyrin repeat domains 4, prion protein, phosphoserine aminotransferase 1, and Annexin A1, which can effectively predict patients’ prognosis and provide a reference for the clinical diagnosis and targeted therapy of EC.

1. Introduction

Endometrial carcinoma (EC) is a malignant tumor originating from the endometrial epithelium,^[1] and also one of the most common malignant tumors of the female reproductive system.^[2] It predominantly affects women aged 50 to 60, especially perimenopausal or postmenopausal women, and is mainly influenced by estrogen signaling.^[3] According to the 2022 global cancer statistics from the International Agency for Research on Cancer under the World Health Organization, there were approximately 4,20,200 new cases of EC worldwide, accounting for 2.1% of all new cancer cases. The number of global deaths due to EC was 97,800, representing 1.0% of all cancer-related deaths. The incidence and mortality rates of EC vary across different countries and regions and have shown an upward trend year by year, ranking among the top gynecological malignancies in developed countries.^[4]

EC is classified into 2 types based on clinic pathological features: type I, which is estrogen-dependent and the most common form, and type II, which is non-estrogen-dependent. Surgery is currently the primary treatment for EC, but approximately 15% to 20% of patients experience recurrence postoperatively.^[5] Early diagnosis of EC is challenging, and the prognosis for patients with advanced EC is poor. Therefore, there is an urgent need to identify potential targets and biomarkers for the prediction, early diagnosis, and prognosis assessment of EC to improve treatment outcomes and patient survival.

In recent years, the rapid development of microarray technology, biochip technology, and next-generation sequencing has been widely applied in clinical research. This study utilizes data from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO), combined with bioinformatics methods, to reanalyze and explore EC-related data, aiming to create a risk prediction model and provide a foundation for future EC research.

2. Material and methods

2.1. Data acquisition

RNA-seq data and related clinical information of EC were downloaded from TCGA database (https://portal.gdc.cancer.gov/), including 554 EC samples and 35 normal samples. Additionally, the gene expression dataset GSE17025 was obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/), which included 79 EC samples and 12 normal samples. The annotation platform was GPL570 [HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array.

2.2. Data preprocessing and differential gene expression analysis of EC

Raw data preprocessing implemented in R 4.4.1 utilized the “Affy” package^[6] in R4.4.1 for normalization and standardization. Background correction was performed using the robust Multichip Average method, followed by log₂ transformation using the “Limma” package.^[7] Differentially expressed genes (DEGs)^[8,9] were identified based on a false discovery rate < .05 (Benjamini-Hochberg correction for multiple testing) and |log₂ FC|  ≥  1.5. Low-expression genes (|log₂ FC| < 1.5) were filtered out, and missing values were imputed using the k-nearest neighbors algorithm (k = 10).Volcano plots were generated using the “ggplot2” package to visualize DEGs.

2.3. Construction of weighted gene co-expression network

The “weighted gene co-expression network analysis (WGCNA)” package^[10] was used to construct the co-expression network of DEGs in endometrial cancer, and identify important modules related to EC prognosis and survival and other related WGCNA analysis.

Firstly, the hclust() function was used to perform sample clustering analysis and remove outlier samples; secondly, the pickSoftThreshold() function was used to screen the best soft threshold (β), and the softConnectivity() function was used to test the soft threshold screening results, so that the connections between genes in the gene network conformed to the distribution of the scale-free network; then analyze the Pearson correlation between genes, transform the correlation coefficient into the adjacency matrix, and construct the topological overlap matrix.

2.4. Identification of key gene modules

Hierarchical clustering of the topological overlap matrix was performed to create a gene dendrogram, and dynamic tree cutting was used to distinguish gene modules. The association between each module and clinical traits was then calculated. Furthermore, module eigengene, gene significance, and module significance were calculated, and module membership was assessed to identify key gene modules.

2.5. Construction of EC prognostic risk prediction model

The expression levels of intersecting genes between key modules and DEGs from the GSE17025 dataset, as well as relevant survival information, were merged in R4.4.1 software. A univariate Cox regression analysis was performed using the “survival” package in R4.4.1 software to identify potential genes highly correlated with overall survival (OS). In the univariate Cox regression analysis, genes with P < .05 were considered statistically significant. Significant genes were subsequently included in a multivariate Cox regression analysis to identify genes related to EC prognosis. To control for potential confounders, we employed a stepwise selection method (direction = “both”) in the multivariate Cox regression. The gene weights (β-values) are directly derived from the regression coefficients of the multivariate Cox regression, which quantify the independent contribution of each gene to survival risk. Specifically, the β-values are calculated using the maximum partial likelihood estimation method and are statistically assessed for significance using the Wald test (P-value threshold set at .05). Additionally, a forest plot was generated to visually represent the results. The final prognostic risk prediction model was constructed with the following formula:

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \underset{i=1}{{\displaystyle \sum ^n}}}\mathrm{e}\mathrm{x}{\mathrm{p}}_{\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{i}}\times {\mathrm{\beta }}_{\mathrm{g}\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{i}}}\end{array}}$$

where exp_{gene i} represents the normalized expression level of gene i, and β_{gene i} is the multivariate Cox regression coefficient.

2.6. Validation of EC prognostic risk prediction model

Each EC patient’s risk score was calculated, and patients were classified into high-risk groups (above the median value) and low-risk groups (equal to or below the median value). Kaplan–Meier survival analysis^[11,12] was performed, and differences between the 2 groups were compared using the log-rank test. Based on the median of gene expression, patients were divided into a high-expression group and a low-expression group, and Kaplan–Meier survival curves were plotted for each key gene respectively. Survival curves were plotted, and ROC analysis was conducted for 1-, 3-, and 5-year survival times using the “pROC” package. ROC curves^[13,14] were plotted, and the area under curve (AUC) was calculated to assess the performance of the risk prediction model.

2.7. External validation using independent datasets

To further validate the prognostic model, the gene expression dataset GSE120490 was downloaded from the GEO database, which included 145 EC samples with survival information. The annotation platform was GPL570 [HG-U133_Plus_2] Affymetrix Human Genome U133 Plus 2.0 Array. Raw data were preprocessed using the same pipeline as described in Section 2.1. Using the median risk score from Section 2.6 as the cutoff, patients in the GSE120490 dataset were classified into high- and low-risk groups. Kaplan–Meier survival analysis and ROC curves (1-, 3-, and 5-year AUC) were used to evaluate the model’s performance in external datasets.

2.8. Single-cell RNA sequencing data analysis

To investigate the impact of tumor heterogeneity on prognosis, single-cell RNA sequencing (scRNA-seq) dataset GSE225689 for EC was downloaded from the GEO database. Data preprocessing was performed using the “Seurat” package, including quality control, data normalization (LogNormalize method), identification of highly variable genes, and batch effect correction using the Harmony algorithm. Subsequent analysis focused on malignant epithelial cells. High-dimensional weighted gene co-expression network analysis (hdWGCNA) was conducted using the “hdWGCNA” package to identify potential biomarkers. The optimal soft threshold was determined to construct gene modules, and modules significantly correlated with clinical traits were identified. Genes from the key module were extracted and intersected with the 5-gene prognostic model to validate their biological relevance.^{[8,9,11–14]}

3. Results

3.1. Results of differentially expressed genes (DEGs) screening

Using the “limma” package with R4.4.1 software, the results showed that 869 significant DEGs were identified in the GSE17025 dataset, including 539 upregulated genes and 330 downregulated genes, with a false discovery rate < .05 and |log₂FC|  ≥  1.5 as the filtering conditions. In the TCGA RNA-seq data for EC, 1059 significant DEGs were screened, including 420 upregulated genes and 639 downregulated genes. And the differential gene volcano maps were plotted respectively (Fig. 1A, B).

Figure 1.

(A) GSE17025 differential gene volcano map: red dots are upregulated genes, blue dots are downregulated genes. (B) TCGA-EC differential gene volcano plot: red dots are upregulated genes, blue dots are downregulated genes. TCGA-EC = The Cancer Genome Atlas-Endometrial Carcinoma.

3.2. Results of WGCNA analysis

In this study, the “WGCNA” package in Rstudio4.4.1 was used to construct a gene co-expression network for EC. Hierarchical clustering of the samples was performed, setting the cut height to 90,000 to remove outlier samples (Fig. 2A). The optimal soft threshold (β) was chosen as 5, with R² = 0.869, successfully constructing a scale-free network (Fig. 2B).

Figure 2.

(A) TCGA-EC samples clustering tree note: The red line represents the cut height. (B) Best soft-threshold (β) Screening note: The y-axis of the left plot represents the scale-free topology fit (R²), while the y-axis of the right plot indicates the mean connectivity. The x-axis in both plots shows the soft-threshold power (β). (C) Gene clustering tree note: Each color represents a gene module. (D) Correlation matrix between gene modules and clinical traits. (E) Scatter plot of genes in the turquoise module correlated with clinical traits. TCGA-EC = The Cancer Genome Atlas-Endometrial Carcinoma.

Pearson correlation analysis was used to calculate the correlation coefficients between each gene module and clinical traits (such as survival time), and heatmaps were generated to visualize the relationships (Fig. 2C, D). The gene module with the highest correlation coefficient (r) was module eigengene turquoise (r = −0.14, P = .001). Therefore, the turquoise module was selected for further analysis as a significant module (Fig. 2E).

3.3. Construction of EC prognostic risk prediction model

Using the “venn” package, an intersection analysis was performed between the turquoise module genes and the DEGs from the GSE17025 dataset, resulting in a total of 65 intersecting genes (Fig. 3A).

Figure 3.

(A) Venn diagram of the intersection between turquoise module and GSE17025 differentially expressed genes. (B) Forest plot of univariate Cox regression analysis results. (C) Forest plot of multivariate Cox regression analysis results. DEGs = differentially expressed genes, GEO = Gene Expression Omnibus.

In the univariate Cox regression analysis, 19 genes were found to be significantly associated with the OS, including: PDZ domain containing ring finger 3 (PDZRN3), SOSTDC1, CENPF, KN motif and ankyrin repeat domains 4 (KANK4), MAGEH1, AMOTL2, ECT2, PIK3R3, LPCAT1, SLC38A1, ID4, TOP2A, prion protein (PRNP), NDRG1, CTSV, phosphoserine aminotransferase 1 (PSAT1), MYBL2, ID3, Annexin A1 (ANXA1). The above 19 genes were included in the subsequent analysis, as shown in Fig. 3B.

In the multivariate Cox regression analysis, 5 genes were identified as being significantly related to the prognostic of EC patients, including: PDZRN3, KANK4, PRNP, PSAT1, and ANXA1, in which hazard ratio (HR) < 1 was the protective genes for EC prognosis, and HR > 1 was the risk genes for EC prognosis, as shown in Fig. 3C. The final EC prognostic risk prediction model was constructed as follows:

Risk Score = −0.297 × expPDZRN3 + 0.236 × expKANK4 + 0.253 × expPRNP + 0.173 × exp PSAT1–0.202 × expANXA1

For example, if the expression level of PDZRN3 in a patient is 1.5 (normalized value), and the corresponding β value is −0.297, then the contribution of this gene is 1.5 × (−0.297) = −0.446. The final risk score, calculated as the algebraic sum of contributions from all 5 genes, was employed to stratify patients into high- and low-risk groups using the median value as the stratification cutoff.

3.4. Validation of the EC prognostic risk prediction model

The number of patients in the high-risk and low-risk groups were 270 and 271, respectively. Kaplan–Meier survival analysis showed that the risk score had a significant impact on the prognosis of EC patients (P < .001), with the survival rate of patients in the high-risk group being lower than that of the low-risk group (Fig. 4A–D). ROC analysis results showed that the AUC values for the EC prognostic risk prediction model at 1, 3, and 5 years were 0.692, 0.720, and 0.721, respectively, all >0.6, indicating that the model had good survival prediction performance (Fig. 4E). Patients in the low-expression groups of PDZRN3 and ANXA1 exhibited significantly poorer survival rates compared to their high-expression counterparts (P < .001), consistent with their designation as “protective genes” in the prognostic model. Conversely, patients with high expression levels of KANK4, PRNP, and PSAT1 demonstrated significantly poorer survival rates than those in the low-expression groups (P < .001), supporting their classification as “risk genes” with adverse prognostic implications (Fig. 4F).

Figure 4.

(A) Distribution of risk scores in high-risk and low-risk groups of EC. (B) Distribution of survival status in high-risk and low-risk groups of EC. (C) Heatmap of prognostic gene expression in high-risk and low-risk groups of EC. (D) Kaplan–Meier survival analysis results. (E) ROC curve of the prognostic risk prediction model. (F) Kaplan–Meier survival analysis of key genes. ANXA1 = Annexin A1, AUC = area under curve, EC = endometrial carcinoma, KANK4 = KN motif and ankyrin repeat domains 4, PDZRN3 = PDZ domain containing ring finger 3, PRNP = prion protein, PSAT1 = phosphoserine aminotransferase 1, ROC = receiver operating characteristic.

3.5. External validation in the GSE120490 dataset

The prognostic model demonstrated consistent performance in the independent GSE120490 dataset. Kaplan–Meier analysis revealed significantly poorer OS in the high-risk group compared to the low-risk group (log-rank P = .005; Fig. 5A). The ROC analysis yielded AUC values of 0.609 (1-year), 0.640 (3-year), and 0.627 (5-year), indicating robust predictive accuracy (Fig. 5B).

Figure 5.

(A) Kaplan–Meier survival analysis results of GSE120490 dataset. (B) ROC curve of GSE120490 dataset. AUC = area under curve, ROC = xxx.

3.6. Single-cell RNA sequencing reveals tumor heterogeneity

Five gene modules were identified by hdWGCNA analysis of malignant epithelial cells (soft threshold = 6). Among these, module “M1” exhibited the strongest correlation with clinical features, suggesting its potential involvement in EC progression. Intersection analysis between genes in module “M1” and the 5-gene prognostic model revealed co-expression of only KANK4 and PSAT1 within the “M1” module (Fig. 6).

Figure 6.

Identification of key modules related to endometrial carcinoma epithelial cells by hdWGCNA analysis. (A) Best soft-threshold (β) screening. (B) Epithelial cells hdWGCNA dendrogram. (C) Five gene modules were obtained and the top 10 hub genes were presented according to the hdWGCNA. (D) UMAP plots illustrating the distribution of each module. (E) Correlation analysis between different module. hdWGCNA = high-dimensional weighted gene co-expression network analysis.

4. Discussion

EC is a common gynecological malignancy that poses a significant threat to women’s health worldwide. It is difficult to diagnose early, and the prognosis for patients in the advanced stages is poor. Therefore, this study aimed to perform bioinformatics analysis to identify potential prognostic genes in EC patients and construct a prognostic risk prediction model, providing a theoretical basis for future research on clinical diagnosis, treatment, and prognosis evaluation.

A total of 5 genes were included in the prognostic risk prediction model of EC constructed in this study, including PDZRN3, KANK4, PRNP, PSAT1 and ANXA1.

The external validation using the GSE120490 dataset reinforced the generalizability of our 5 genes prognostic model. Despite differences in datasets and patients, the model retained predictive power, underscoring its robustness against technical and biological heterogeneity. However, the slightly lower AUC values in the external dataset suggest that future iterations of the model may benefit from integrating additional clinical variables to enhance precision. These results collectively affirm the clinical translatability of our model while guiding refinements for broader applicability.

This study is the first to integrate scRNA-seq data to elucidate the biological basis of the 5-gene prognostic model. While the original model was constructed using bulk RNA-seq data, scRNA-seq analysis demonstrated that only KANK4 and PSAT1 were co-expressed in a malignant epithelial-specific module. This discrepancy may arise from tumor heterogeneity: bulk sequencing captures signals from mixed cell populations, potentially diluting cell type-specific expression patterns. The restricted overlap highlights the necessity of incorporating single-cell approaches to refine prognostic biomarkers and account for cellular diversity in EC.

PDZRN3 is widely expressed in ovarian, endometrial, and 20 other tissues, encoding a member of the LNX family of RING-type ubiquitin E3 ligases. It may play an important role in angiogenesis and is potentially targeted for degradation by human papillomavirus E6 protein. Studies have shown that PDZRN3 plays a role in the metastasis and proliferation of EC, with its expression significantly lower in EC tissues compared to normal endometrium. Lower expression is also correlated with certain clinicopathological features, such as patient age, tumor grade, and tumor subtype, indicating that PDZRN3 could be a potential therapeutic target for EC.^[15] In this study, PDZRN3 was identified as a protective prognostic gene for EC, with its reduced expression associated with poor prognosis. The gene exhibited lower expression in the high-risk EC group and higher expression in the low-risk group.

KANK4 encodes a protein from the KANK family, containing multiple ankyrin repeat domains. To date, no studies have reported the functional mechanism of KANK4 in EC. In this study, KANK4 was identified as a risk gene for EC prognosis, with higher expression potentially leading to poor prognosis. It was highly expressed in the high-risk group and expressed at lower levels in the low-risk group. Research has also found that KANK4 is closely related to the OS and disease-free survival of cervical cancer patients in the TCGA database. Cholesterol metabolism mediated by DHCR7 upregulates KANK4 expression, activating the PI3K/AKT signaling pathway to promote the migration and invasion of cervical cancer cells.^[16]

PRNP encodes a membrane-anchored glycoprotein. The relationship between PRNP and EC is still under investigation, with no conclusive evidence yet. However, recent studies suggest that PRNP may play a role in tumor development, with high expression indicating poor prognosis in cancer patients.^[17] PRNP promotes cancer cell proliferation, metastasis, and invasion, and targeting PRNP could be a novel therapeutic strategy.^[18,19] In this study, PRNP was identified as a risk gene for EC prognosis, with increased expression potentially leading to poor outcomes. The gene was highly expressed in the high-risk group and lowly expressed in the low-risk group.

PSAT1 encodes a protein involved in the biosynthesis of phosphoserine. Research has indicated that PSAT1 plays a role in EC cell growth, immune regulation, and the cell cycle. PSAT1 expression is positively correlated with Th2 cells and negatively correlated with Th17 cells. High PSAT1 expression is associated with poor prognosis in EC patients, and it has been identified as a potential target for EC diagnosis and immunotherapy.^[20] Yuan Fan et al,^[21] also found that CA3, HNMT, PHGDH, CD38, PSAT1, and GPI could be potential prognostic biomarkers for EC. In this study, PSAT1 showed a HR of 1.312 (95% confidence interval: 1.107–1.555, P = .002) in univariate Cox regression, but it was not an independent prognostic factor in multivariate Cox analysis (HR: 1.188, 95% confidence interval: 0.996–1.418, P = .056), suggesting that PSAT1 may indirectly affect survival in EC patients. Increased expression of PSAT1 was observed in the high-risk group.

ANXA1 encodes a calcium-dependent phospholipid-binding protein that inhibits phospholipase A2 activity and has anti-inflammatory properties. ANXA1 shows functional loss or low expression in multiple tumors.^[22] Research has shown that the expression of Annexin-1 in endometrial adenocarcinoma is significantly lower than that in normal and atypical endometrial tissues. Annexin-1 may play a crucial role in the early diagnosis of endometrial adenocarcinoma and could aid in assessing the risk of progression from precancerous lesions to cancer.^[22,23] In this study, ANXA1 was identified as a protective prognostic gene, with reduced expression associated with poor outcomes. It was expressed at lower levels in the high-risk group and at higher levels in the low-risk group.

The 5-gene prognostic risk prediction model (PDZRN3, KANK4, PRNP, PSAT1, ANXA1) proposed in this study holds significant potential for clinical application in EC management. Below, we elaborate on its implications for patient stratification, treatment response prediction, and targeted therapy guidance:

4.1. Patient stratification

The risk score derived from the model effectively categorizes EC patients into high- and low-risk groups, as demonstrated by Kaplan–Meier survival analysis (P < .001). This stratification could aid clinicians in tailoring surveillance protocols and therapeutic intensity. For example, high-risk patients, characterized by lower survival rates, may benefit from intensified follow-up schedules and adjuvant therapies such as chemotherapy or radiation, even at earlier stages. Conversely, low-risk patients might be candidates for less aggressive interventions, reducing overtreatment and associated side effects. Additionally, integrating this model with existing clinicopathological staging systems could enhance prognostic accuracy and refine risk-adapted clinical pathways.

4.2. Predicting treatment response

The 5 genes in the model are mechanistically linked to pathways influencing tumor behavior and therapy resistance. For example, high expression of certain genes may indicate sensitivity or resistance to chemotherapy or radiation therapy, thus guiding treatment selection.

4.3. Guiding targeted therapies

Based on the involvement of the identified genes in specific signaling pathways or molecular mechanisms within our study model, targeted therapeutic agents directed against these pathways can be proposed. Additionally, it will be discussed how to select appropriate targeted therapies by detecting the expression of these genes.

In summary, based on TCGA and GEO databases, this study analyzed gene expression and clinical data of endometrial cancer by bioinformatics methods, and constructed a prognostic risk prediction model of EC containing 5 genes, which is of positive significance for the prediction of prognostic survival and gene-targeted therapy of EC. However, the specific mechanism of action needs further research validation.

Author contributions

Conceptualization: Gongwei Zhou, Yu Zhang.

Data curation: Yu Zhang.

Formal analysis: Gongwei Zhou, Yu Zhang.

Methodology: Yu Zhang.

Project administration: Gongwei Zhou.

Resources: Yu Zhang.

Software: Yu Zhang.

Supervision: Gongwei Zhou.

Writing – original draft: Yu Zhang.

Writing – review & editing: Gongwei Zhou.