Combining single-cell and bulk RNA sequencing data to create a reliable prognostic model for predicting clear-cell renal cell carcinoma progression

Abstract

Background

Clear-cell renal cell carcinoma (ccRCC) is the most common pathological type of kidney cancer and is characterized by a low survival rate. Accurate prediction of the occurrence and progression of ccRCC is crucial for diagnosis and treatment. This study aimed to integrate multiple publicly available bulk sequencing and single-cell datasets on ccRCC to establish a novel prognostic model for reliable and precise predictions of ccRCC development.

Methods

We used data from five ccRCC samples from the Gene Expression Omnibus (GEO) database to identify 1,303 overlapping differentially expressed genes (DEGs). Through pseudotime analysis of single-cell ccRCC data sourced from the GEO database, we identified 4,002 genes that were highly associated with cancer progression. By using machine learning to screen reliable prognostic genes, we constructed a prognostic model for renal cancer with The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection (TCGA-KIRC) dataset and validated its effectiveness using a novel GEO renal cancer dataset. Finally, we explored the role of these prognostic genes in the progression of ccRCC through in vivo and in vitro experiments.

Results

The five ccRCC sequencing datasets exhibited significant heterogeneity. Therefore, we screened 211 DEGs that were highly associated with the development and prognosis of renal cancer. By exploring the biological functions of these genes, we found that they closely influenced the prognosis of patients with cancer. Upon screening, a reliable set of renal cancer-related DEGs was obtained from multiple samples. The prognostic model accurately identified renal cancer stages and predicted outcomes. In both in vivo and in vitro experimental results, we found that intervening in the expression of these prognostic genes can significantly slow down the progression of ccRCC.

Conclusions

We developed a valuable predictive tool for ccRCC progression, which can help estimate the survival period of patients with renal cancer and aid in the clinical diagnosis and targeted therapy of tumors.

Highlight box.

Key findings

• By integrating multiple datasets, a reliable prognostic model for clear-cell renal cell carcinoma (ccRCC) was constructed.

• The functional impact of prognostic genes on cancer progression was validated through in vivo and in vitro experiments.

What is known and what is new?

• ccRCC exhibits a highly variable prognosis, and existing clinical models often lack accuracy and molecular insight. While several genetic markers have been associated with ccRCC outcomes, a robust multi-gene model validated through functional experiments is still needed

• A clinically applicable prognostic model for ccRCC was developed. The molecular mechanisms underlying ccRCC progression were comprehensively investigated.

What is the implication, and what should change now?

• The predictive model effectively estimates survival outcomes in ccRCC patients.

• Provides a robust molecular foundation for future targeted therapies in ccRCC.

Introduction

Renal cell carcinoma (RCC) originates from the renal tubular epithelium and is one of the most common malignant tumors worldwide, accounting for approximately 3% of all malignant tumors in the human body and 90% of all malignant tumors in the kidneys (1). The global incidence of renal cancer is approximately 430,000 cases per year, with a male-to-female ratio of 2:1, and clear-cell type constitutes nearly 70% of these cases (2,3). For early-stage renal cancer, surgical resection often achieves satisfactory therapeutic outcomes; however, approximately 20–30% of patients who undergo surgery experience recurrence and develop metastatic renal cancer (4,5).Clear-cell renal cell carcinoma (ccRCC) is insensitive to traditional radiotherapy and chemotherapy, and once the tumor metastasizes, the 5-year average survival rate falls below 20%, with no effective treatment options currently available (6,7). Significant progress has been made in the development of personalized immunotherapies for patients with metastatic ccRCC. Immune checkpoint inhibitors combined with vascular endothelial growth factor receptor-targeted tyrosine kinase inhibitors are commonly used in clinical immunotherapy. However, the average survival time for patients receiving these immunotherapies is only two years, highlighting the persistent need for effective therapeutic strategies (5,8). For patients who are unresponsive to these immunotherapies, the use of mammalian target of rapamycin inhibitors should be considered (9). However, a recent large-scale phase III clinical trial of the mammalian target of rapamycin inhibitor everolimus indicated that the 5-year recurrence-free survival rate of patients with ccRCC treated with everolimus was not significantly different from that of the placebo group, and that patients receiving everolimus experienced a considerable number of complications, including diarrhea, nausea, infections, liver and kidney function impairment, and severe metabolic diseases (10). These results suggest that research barriers remain in the ongoing efforts to significantly improve the survival time and treatment outcomes of patients with metastatic RCC.

To improve the prognosis of patients with RCC, further investigations into the mechanisms underlying RCC development and progression are essential to identify reliable prognostic indicators and potential therapeutic targets that enable early diagnosis and optimal treatment. In recent years, with the universalization of sequencing technology, the amount of biological sequencing data associated with tumor tissues has significantly increased, leading to rapid developments in the field of bioinformatics. Several researchers have used sequencing data to screen for differentially expressed genes (DEGs) in tumor tissues, identify tumor markers, and establish prognostic models. However, traditional prognostic models established using biological molecular techniques have limitations, making it difficult to explore cell-specific expression differences within tumors. The emergence of single-cell RNA sequencing (RNA-seq) has allowed researchers to analyze the composition, distribution, function, and communication of different cell subgroups in tissues at the single-cell resolution, enabling more in-depth research on tumors (11). Currently, there are numerous types of bioinformatic predictive models for RCC, with previous research indicating that Kidney Renal Clear-Cell Carcinoma Collection (KIRC)-related molecular markers are primarily associated with proliferation (12), vascular endothelial growth factor (13), immunology (14), inflammation (15), somatic mutations (16), methylation (17), apoptosis/autophagy (18), and metabolic pathways (19). Prognostic models developed using these markers have partially elucidated the prognosis of RCC; however, they cannot overcome the tumor heterogeneity arising from single datasets, leading to poor efficacy (20,21). To address these limitations, we aimed to integrate multidimensional datasets of ccRCC from various sources to construct models capable of accurately predicting cancer progression and prognosis. This integrative strategy not only refines prognostic prediction but also offers mechanistic insights into the interplay between tumor metabolism and the immune microenvironment, thereby showing its potential for clinical translation and targeted therapy development. We present this article in accordance with the MDAR, ARRIVE and TRIPOD reporting checklists (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-442/rc).

Methods

Data collection

We collected five publicly available human ccRCC bulk RNA-seq gene datasets from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/), GSE117890, GSE15641, GSE76351, GSE97327, and GSE168845, which were used to remove heterogeneity between the datasets. Single-cell data for ccRCC were obtained from GSE224630 in the GEO database and used to screen for DEGs related to the occurrence and development of ccRCC. The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/), TCGA-KIRC, and GEO database GSE29609 were used to screen for DEGs associated with cancer survival. Finally, the co-expressed genes filtered through the above steps were combined with TCGA-KIRC data to establish a prediction model for RCC, which was validated using a uniquedataset, GSE167573. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Data processing from different sources and nomenclature

The “limma” (22) package was employed to perform differential analysis of tumor and benign tissue expression in the five human ccRCC datasets from GEO database. DEGs were statistically identified using the criteria P<0.05 and log fold change (FC) >1, and the intersection of these results was referred to as Bulk-seq-DEGs.

Single-cell data for ccRCC were obtained from the GSE224630 dataset, and five samples (GSM7028034–7028038) were selected for the single-cell analysis. The “Seurat” (23) package in R software was used to merge, reduce dimensionality, normalize, cluster cells, and analyze differences in single-cell data. Quality control standards were applied to single-cell data based on the mitochondrial ratio and gene median count, retaining cells with mitochondrial gene content below 25% and gene expression counts between 100 and 5,000. Finally, 2,000 highly variable genes were used to remove batch effects from the data. The “FindClusters” function was then used to construct cell clusters, and uniform manifold approximation and projection (UMAP) was used to visualize the results. Next, the “SingleR” package was used for the automatic annotation of cell clusters, effectively distinguishing the clustering results through marker genes for each cell type, and cell types with significant expression differences were screened. Pseudotime analysis was conducted using the R package “Monocle2” (24) and cells related to the occurrence of ccRCC were identified via visualization results of the pseudotime analysis. From these data, DEGs were extracted, summarized, and subsequently referred to as single-cell (sc) RNA-seq-DEGs.

Weighted gene co-expression network analysis (WGCNA)

Bulk sequencing (bulk-seq) data with prognostic information for ccRCC were obtained from the GSE29609 dataset of the GEO database. After normalizing the whole-genome expression data of the 39 ccRCC tissue samples from GSE29609, the data were merged with the TCGA-KIRC dataset, and samples with the same gene names in both datasets were identified. Duplicates and samples lacking prognostic information were removed, and the “Combat” function was used to eliminate batch effects, resulting in the identification of duplicate genes. To identify the most critical ccRCC gene expression modules related to clinical prognostic information, we used the curated overlapping genes and clinical phenotype data of tumor samples from TCGA-KIRC to perform WGCNA. The correlation coefficients between the modules were determined to be the most reliable at a threshold of 5. Subsequently, we distinguished the gene expression clusters using the WGCNA dendrogram (25) and set a minimum of 100 genes per gene module to reduce the number of clusters. In this manner, renal cancer genes that were strongly correlated with the clinical information for each expressed gene between the modules were identified. Finally, we performed a prognostic-related Cox univariate regression analysis of these DEGs using clinical phenotype information from TCGA-KIRC. Genes significantly associated with prognosis (P<0.05), defined as Prognosis-related-DEGs, were used to construct a prognostic model.

Biological function evaluation of co-expressed DEGs

The ccRCC-DEGs from bulk-seq-DEGs, scRNA-seq-DEGs, and prognosis-related-DEGs were merged, and the intersection yielded a total of 211 co-expressed DEGs. Kyoto Encyclopedia of Genes and Genomes (KEGG) (26,27) enrichment analysis was performed on the 211 genes, and the top five enriched genes from each pathway were selected, resulting in 25 genes, designated as top-co-DEGs, for subsequent biological function analysis.

Single-sample gene set enrichment analysis (ssGSEA) was conducted on the combined gene expression data from TCGA-KIRC using the “CIBERSORT” (28) function to explore the effects of top-co-DEGs on tumor immune cells and to further assess the roles of the identified genes in the RCC tumor microenvironment. STRING (https://string-db.org/) was employed to construct a protein-protein interaction (PPI) network for the top-co-DEGs, which was visualized using Cytoscape software (version 3.9.1) (29).

The Human Protein Atlas (30) comprises proteomics, transcriptomics, and systems biology data, including protein expression in tumors and normal tissues. This database was used to validate the expression of top-co-DEGs.

Construction of a prediction model related to renal cancer

We extracted the expression data of 211 co-expressed DEGs and their corresponding prognostic information from the GSE167573 and TCGA-KIRC datasets. We then used the “coxph” function from the survival package to systematically build univariate Cox models to analyze the prognostic capability of the 211 co-expressed DEGs in the GSE167573 and TCGA-KIRC datasets. Based on the results of the univariate Cox models from the two datasets, we screened for genes with a hazard ratio (HR) >1 and P<0.05, and obtained a core set of 27 hub-DEGs through the dataset intersection. We combined these 27 hub-DEGs with survival data from TCGA-KIRC to establish a training set for the multivariate Cox regression model, which is displayed in a nomogram. The predictive performance of the model was evaluated using decision curve analysis (DCA) and calibration curves. The prognostic efficacy of the model was visualized using receiver operating characteristic (ROC) and Kaplan-Meier curves. To validate our findings, we performed random sampling of TCGA-KIRC cases, allocating 70% of the samples to establish an internal validation cohort. Additionally, we utilized the prognostic data from GSE167573 to construct an independent external validation set for further verification.

Machine learning algorithms

To further screen for potential biomarkers for ccRCC, the expression matrix of TCGA-KIRC was combined, and genes with diagnostic efficacy were identified using three machine learning algorithms: least absolute shrinkage and selection operator (LASSO), random forest (RF), and support vector machine recursive elimination (SVM-RE). From the results of these three algorithms, overlapping genes were identified and considered the most valuable for diagnostic purposes.

Cell culture, quantitative polymerase chain reaction (qPCR), and in vivo experiments

HK-2, Caki-2, 786-O, and OS-RC-2 cells were acquired from Wuhan Pricella Biotechnology Co., Ltd. (Wuhan, China). The ccRCC cell lines 786-O and Caki-2 were derived from primary tumor tissues, whereas the control human cell line HK-2 was derived from benign human renal epithelial cells. The cells were cultured in a CO₂ incubator (Thermo Fisher Scientific, Waltham, MA, USA) at 37 with 5% CO₂, and maintained in RPMI-1640 medium (350-030-CL, Wisent, Saint-Jean-Baptiste, QC, Canada), MEM medium (PM150410, Pricella, Wuhan, China) and McCoy’s 5A medium (PM150710, Pricella) supplemented with 10% fetal bovine serum (085-450, Wisent) and 1% penicillin-streptomycin (450-201-EL, Wisent). Following DEG-specific experiments (fully detailed in Appendix 1), the cells were prepared for reverse transcription quantitative PCR (RT-qPCR). An RNA extraction kit (74004, Qiagen, Hilden, Germany) was used to isolate total RNA from 786-O, Caki-2, OS-RC-2 and HK-2 cells according to the manufacturer’s protocol. After measuring the total RNA concentration, cDNA was generated using a T100 Thermal Cycler (Bio-Rad, Hercules, CA, USA). The LightCycler^® 96 Instrument (Roche, Basel, Switzerland) was used for real-time qPCR fluorescence analysis.

The detailed operational steps for the in vivo experiments on DEGs are provided in Appendix 1.

In vivo tumorigenesis assay

For the in vivo tumorigenesis assay, all experimental animals were maintained under specific pathogen-free (SPF) conditions with a 12/12-h light/dark cycle and were provided free access to food and water. Female BALB/c nude mice (6-week-old) were purchased from SPF (Beijing) Biotechnology Co., Ltd. (Beijing, China) (No. 110324251101350632). For xenograft establishment, the BALB/c nude mice were randomly divided into three groups (sample size: 5 mice per group) and inoculated with cells as follows: sh-BCAT1 stable transfected 786-O Cell [3×10⁶ cells, suspended in 100 µL phosphate-buffered saline (PBS)]; sh-IFNGR2 stable transfected 786-O Cell (3×10⁶ cells, suspended in 100 µL PBS) and 786-O NC group cells (3×10⁶ cells, suspended in 100 µL PBS). Mice were euthanized 5 weeks after injection (cervical dislocation following 1.25% tribromoethanol), Tumor dimensions were measured at designated intervals using digital calipers. A protocol was prepared before the study without registration. All animal experiments were performed under a project license (No. 202503003) granted by the ethics committee of Guangxi Medical University, in compliance with Chinese national guidelines and institutional guidelines for the care and use of animals.

Statistical analysis

All analyses were conducted using R language (Version 4.1.2, www.r-project.org) and GraphPad Prism 8.0 (GraphPad, USA). One-way analysis of variance was used to evaluate the differences between groups, and the Wilcoxon rank-sum test was used to analyze continuous variables. Correlations were analyzed using Pearson’s (for normally distributed variables) and Spearman’s (for non-normally distributed variables) analyses. Kaplan-Meier analysis was used to evaluate the survival differences between the groups.

Results

ccRCC dataset analysis and merging

The roadmap of this study is shown in Figure 1A. We collected and analyzed five bulk-seq ccRCC datasets from the GEO database (tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). We analyzed the differences between the tumor and benign tissue data and identified 1,303 overlapping DEGs in the five datasets (Figure 1B). When the filtering criteria of P<0.05 and logFC >1 were applied, only three overlapping DEGs were identified (Figure 1C). This result supports our previous notion that individual ccRCC datasets exhibit high heterogeneity and that models generated from a single dataset may not accurately capture the complex transcriptomic landscape of the tumor tissue. Accordingly, this approach was employed to identify additional prognostic genes that are universally applicable to renal cancer. To broaden the range of DEGs for the prognostic model, we defined the aforementioned 1,303 genes as Bulk-seq-DEGs for subsequent analysis (table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX).

Figure 1.

Roadmap of the study and screening scheme for Bulk-seq-DEGs. (A) Technical roadmap. The process of establishing a prognostic model involves data screening, model establishment, and evaluation of the accuracy and effectiveness. (B,C) Venn diagrams show poor heterogeneity and accuracy in the existing ccRCC dataset. (C) It is screened based on the criteria P<0.05 and logFC >1. ccRCC, clear-cell renal cell carcinoma; DEGs, differentially expressed genes; FC, fold change; scRNA-seq, single-cell RNA sequencing; TCGA-KIRC, The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection; WGCNA, weighted gene co-expression network analysis.

The single-cell ccRCC dataset GSE224630 from the GEO database contained 23,248 cells (Figure S1A-S1C). Following cell annotation and UMAP visualization, the ccRCC cells were clustered into 21 clusters, roughly classified into five cell subtypes: epithelial, endothelial, tissue, T, and stem cells (Figure 2A, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). We then performed pseudotime analysis of the single-cell data from different perspectives, including cell clustering, cell annotation, and renal cancer progression (Figure 2B,2C). We found that endothelial, epithelial, and tissue stem cells corresponded to the three stages of ccRCC progression. “Cell-Chat” analysis showed that these three cell types had the strongest interactions with other cells, indicating their prominent roles in RCC progression (Figure 2D). Therefore, we focused on analyzing the differential genes of these three cell subtypes and explored the cell signaling pathways involved in the three stages of ccRCC through KEGG enrichment analysis (Figure 2E-2G, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). The most prominent pathway identified in the enrichment analysis was the metabolic pathway. After intersecting the DEGs from the three cell types (endothelial, epithelial, and tissue stem cells), 4,002 genes were identified, named scRNA-Seq-DEGs, and used for subsequent analysis (table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX).

Figure 2.

Identification of scRNA-DEGs most closely related to the development of cancer cells through ccRCC scRNA-seq. (A) UMAP of 23,248 single cells from five ccRCC tissues in the GSM7028034–GSM7028038 datasets from the GEO database. Left, clustering results; right, cell type. (B,C) Results of pseudotime analysis of single-cell datasets. (C) Left, according to cell type; middle, according to cell clustering; right, according to the stage of cancer development. (D) Cell-Chat intensity of each cell type. KEGG enrichment analysis of DEGs in three pseudo temporal analysis methods: cell clustering (E), cell annotation (F), and the progression of renal cancer (G). ccRCC, clear-cell renal cell carcinoma; DEGs, differentially expressed genes; KEGG, Kyoto Encyclopedia of Genes and Genomes; scRNA-seq, single-cell RNA sequencing; UMAP, uniform manifold approximation and projection.

WGCNA and result identification

We merged the normalized bulk RNA data from TCGA-KIRC and GSE29609 (tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX), resulting in 15,294 duplicate genes. These genes were used in conjunction with survival data from TCGA-KIRC for WGCNA. Scatter plots were used to present the organization and clustering analysis results of the TCGA-KIRC samples (Figure S1D,S1E). Based on these results, we constructed a WGCNA co-expression network and found that most genes were associated with co-expression modules (Figure 3A). The expression values for each sample in each module were calculated, and the clustering results of the modules were visualized based on our analysis of the association between each module and the phenotype information of TCGA-KIRC (Figure 3B,3C, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). The expression matrix in the red module had a strong positive correlation with the prognostic information of ccRCC, whereas that of the blue module exhibited a strong negative correlation. We then performed differential analysis on the genes in each module using the criteria Log FC >1 and P<0.05 (tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX), and in conjunction with the clinical information of TCGA-KIRC, we conducted a univariate Cox regression analysis. The genes significantly associated with cancer prognosis were identified and subsequently named prognosis-related DEGs (table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX).

Figure 3.

Identification of prognosis-related-DEGs using WGCNA in the TCGA-KIRC and GSE29609 datasets. (A) Construction of gene expression modules and merging of cluster dendrograms. (B) The heatmap shows the correlation between gene modules and clinical information of cancer (OS, cancer grade, and TNM stage). (C) Cluster dendrograms of the gene modules. (D) Venn diagram showing the merging of DEGs from the three sources (Bulk-seq-DEGs, scRNA-seq-DEGs, and prognosis-related-DEGs), resulting in the screening of 211 co-expressed DEGs. (E) KEGG pathway enrichment results for co-expressed DEGs (database source: www.kegg.jp/kegg/kegg1.html). DEGs, differentially expressed genes; KEGG, Kyoto Encyclopedia of Genes and Genomes; OS, overall survival; scRNA-seq, single-cell RNA sequencing; TCGA-KIRC, The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection; TNM, tumor-node-metastasis; WGCNA, weighted gene co-expression network analysis.

Biological functions of co-expressed DEGs

The intersection of the bulk-seq-DEGs, scRNA-seq-DEGs, and prognosis-related-DEGs yielded 211 co-expressed DEGs related to ccRCC with low heterogeneity from multiple samples (Figure 3D, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). To further validate the rationale for these DEGs, we explored the biological functions and expression profiles of co-expressed DEGs in renal cancer.

KEGG enrichment analysis of the 211 co-expressed DEGs revealed that these genes were predominantly enriched in five pathways closely related to renal cancer: amino acid metabolism, lipid metabolism, immune system, resistance to anticancer drugs, and signal transduction (Figure 3E, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). The top 25 co-expressed genes that were significantly enriched in these pathways were: ABCC3, ACAA2, ACADM, ACAT1, ARPC5, AUH, BCAT1, CPT1A, EIF4EBP1, ENO2, FCGR2B, FOLR1, GAPDH, HADHB, HMGCL, IFNGR2, IGF1R, LYN, MARCKS, RAC2, SHMT1, SLC2A1, TYMS, VASP, and WAS. After combining these genes with prognostic information from TCGA-KIRC, we plotted Kaplan-Meier curves to examine whether these genes influenced the survival length and status of patients with RCC (Figure S2, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). The results showed that the identified prognostic genes independently predicted the survival time of patients with renal cancer. The expression profiles of these top-co-DEGs in the single-cell sequencing results were investigated, revealing that most genes exhibited significant expression in the single-cell renal cancer data (Figure 4A). Using the STRING database, we established a PPI network for these genes and found that they could be roughly divided into metabolism-related and immune-related genes, with glyceraldehyde-3-phosphate dehydrogenase (GAPDH) serving as a hub for crosstalk between the two (Figure 4B).

Figure 4.

Exploring the biological functions of top-co-DEGs. (A) Expression of top-co-DEGs in the single-cell dataset GSE224630. (B) PPI network of the top-co-DEGs. (C) Heatmap showing the distribution of 22 immune-cell subtypes in TCGA-KIRC. (D) Box plot displaying the differences in the distribution of each type of immune-cell. (E,F) Spearman’s correlation heatmap of top-co-DEGs with immune-cell subtypes using the “Mantle test” function. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001; ns, not significant (P>0.05). DEGs, differentially expressed genes; PPI, protein-protein interaction; TCGA-KIRC, The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection.

Next, we conducted ssGSEA to investigate the effects of top-co-DEGs on immune cells in renal cancer. Figure 4C,4D visually present the expression of immune cells in patients with tumors or controls from the TCGA-KIRC. Compared to the control group, the differences in immune subtypes in patients with tumors were primarily concentrated in T cells and macrophages, and the functions of these two cell types were highly related to the top-co-DEGs selected in this study (Figure 4E,4F).

Finally, to confirm the differences in the expression of the selected top-co-DEGs in ccRCC, we validated the expression of the aforementioned genes in human 786-O and Caki2 cells and obtained immunohistochemical results for top-co-DEGs from the Human Protein Atlas database in normal and renal cancer tissues (Figure 5, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX). Molecular biology results revealed significant differences in the expression of top-co-DEGs between the tumor and control groups. In summary, co-expressed DEGs can influence the occurrence and development of renal cancer through different mechanisms, further highlighting their biological significance.

Figure 5.

Validation of top-co-DEG expression in ccRCC based on molecular biology methods. (A,B) Immunohistochemical results of top-co-DEGs in renal cancer and normal tissues from the HPA database (see table available at https://cdn.amegroups.cn/static/public/tau-2025-442-1.XLSX for details). The close-up view shows a ×5 magnification of the original image. (C) Expression of top-co-DEGs in human HK2 renal tubular and 786-O renal cancer cells. (D) Expression of top-co-DEGs in human HK2 and Caki2 cells. ****, P<0.0001. ccRCC, clear-cell renal cell carcinoma; DEGs, differentially expressed genes; HPA, Human Protein Atlas.

Constructing a prognostic model using co-expressed DEGs

Expression data for 211 co-expressed DEGs and prognostic information from the samples were extracted from TCGA-KIRC and GSE167573 datasets. Univariate Cox regression analyses with the criteria HR >1 and P<0.05 resulted in 27 hub-DEGs (Figure 6A,6B, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-2.XLSX). We used the TCGA-KIRC dataset as the training set and performed multivariate Cox regression analyses using the expression levels of hub-DEGs and the prognostic information from TCGA-KIRC samples to establish a prognostic model for ccRCC (Figure S3, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-2.XLSX). To comprehensively evaluate the prognostic performance of our model, we employed a dual-validation approach: 70% of TCGA-KIRC samples served as the internal validation cohort, while the GSE167573 dataset was utilized as an independent external validation set (tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-2.XLSX). The training set model was visualized using nomograms (Figure 6C), and both DCA and calibration curves demonstrated that the model exhibited good efficacy in predicting 1-, 3-, and 5-year survival of patients with RCC (Figure 6D,6E). Subsequently, Kaplan-Meier and 1-, 3-, and 5-year ROC curve analyses were performed in both the training and validation sets. Notably, the model effectively predicted the prognostic risk among patients with renal cancer at different time points (Figure 7A-7D, Figures S4-S6). Finally, by integrating clinical information from TCGA-KIRC, we found that the prognostic model could predict clinical parameters such as TNM staging and clinical grading for renal cancer (Figure 7E-7H). In summary, the established prognostic model significantly distinguished high- and low-risk survival in patients and effectively predicted survival outcomes at different time points.

Figure 6.

Establishing a predictive model for ccRCC through co-expressed DEGs. (A) Following univariable Cox regression analysis using the criteria HR >1 and P<0.05, we identified the hub-DEGs of co-expressed DEGs that showed significant correlation in both the TCGA-KIRC and GSE167573 datasets. (B) Univariate regression analysis results of hub-DEGs in TCGA-KIRC. (C) Combined hub-DEGs with survival data from TCGA-KIRC to create a nomogram for multivariate regression analysis. (D) DCA plot at different follow-up times. (E) Prognostic calibration curves for different follow-up times. *, P<0.05; **, P<0.01. ccRCC, clear-cell renal cell carcinoma; CI, confidence interval; DCA, decision curve analysis; DEGs, differentially expressed genes; HR, hazard ratio; OS, overall survival; TCGA-KIRC, The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection.

Figure 7.

Performance of the prognostic model for predicting OS of patients with ccRCC. AUC values of the ROC curve in the training sets (A) and validation set (B) are shown. All AUC values are greater than 0.5, indicating that the risk factors are significantly associated with the occurrence of ccRCC. Kaplan-Meier analysis demonstrates the effectiveness of the prognostic model in distinguishing survival differences between the high- and low-risk groups in both the training (C) and validation sets (D). By integrating the prognostic model with the clinical documents of TCGA-KIRC, we discovered that our prognostic model is capable of accurately predicting a range of clinical information pertinent to renal cancer, including pathological stage (E), N (F), M (G), and T (H) staging. *, P<0.05; **, P<0.01; ***, P<0.001; ns, not significant (P>0.05). AUC, area under the curve; ccRCC, clear-cell renal cell carcinoma; NA, not available; OS, overall survival; ROC, receiver operating characteristic; TCGA-KIRC, The Cancer Genome Atlas-Kidney Renal Clear-Cell Carcinoma Collection; TNM, tumor-node-metastasis.

Exploring the biological functions of hub-DEGs

We conducted KEGG enrichment analysis of hub-DEGs and found that these genes were mainly enriched in metabolic pathways, such as amino acid and fatty acid metabolism (Figure S7A, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). After constructing the PPI network of hub-DEGs, we found that GAPDH was an important hub protein (Figure S7B). After intersecting the top-co-DEGs and hub-DEGs, four genes appeared in both screening modes: LYN, BCAT1, GAPDH, and ENO2 (Figure S7C). Subsequently, we explored the biological functions of these genes and, in conjunction with ssGSEA, investigated their roles in the tumor microenvironment. We found that BCAT1, LYN, and GAPDH were significantly associated with the functions of different subtypes of T cells and macrophages in the tumor environment (Figure S8A-S8C). Furthermore, these three genes were significantly correlated with the immune, stromal, and tumor purity scores in renal cancer (Figure S8D-S8O, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). Finally, the expression of hub-DEGs was validated in human (HK2) and renal cancer cells (786-O and Caki2; Figure S9, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX).

Exploring the effect of DEGs on the proliferation function of ccRCC

The top-co-DEGs were merged with 27 hub-DEGs, combined with the gene expression file of TCGA-KIRC, and subjected to three machine learning methods (LASSO, RF, and RVM-RE) to screen for the most reliable diagnostic DEGs (Figure S10A-S10E, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). We found that the GAPDH and IFNGR2 genes not only co-occurred in the results of the three machine learning methods, but were also identified as predictive genes for ccRCC in previous cancer prognosis analyses (Figure S10F). After combining the results of the immune infiltration analysis, we sought to further investigate the functions of GAPDH, IFNGR2, and BCAT1 in ccRCC tumor cells.

We knocked down GAPDH, IFNGR2, and BCAT1 in both 786-O and OS-RC-2 cells (Figure S11A,S11B, tables available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). After selecting the most efficacious siRNA sequences, we performed cell scratch assays in both cell lines. Post-scratch analysis revealed that silencing BCAT1 and IFNGR2 significantly inhibited the proliferation of ccRCC tumor cells (Figure 8A, Figure S11C). Transwell experiments demonstrated that silencing BCAT1 and IFNGR2 significantly inhibited the invasive ability of ccRCC cells (Figure 8B,8C).

Figure 8.

Validation of the effects of hub genes on ccRCC proliferation in vivo and in vitro. (A) Cell scratch assay results of 786-O cells after intervention of BCAT1 and IFNGR2 genes (no staining; magnification ×100). (B,C) Transwell experiment results of BCAT1 and IFNGR2 gene expression interference in 786-O (B) and OS-RC-2 (C) cells (crystal violet staining; magnification ×200). (D,E) Verification of successful knockout of BCAT1 (D) and IFNGR2 (E) in 786-O cells by quantitative PCR. (F) Experimental results of subcutaneous tumors in nude mice. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. ccRCC, clear-cell renal cell carcinoma; NC, negative control; PCR, polymerase chain reaction.

Subsequently, we sought to validate these molecular findings using preliminary in vivo experiments. The qPCR results demonstrated successful and stable BCAT1 and GAPDH knockout in the 786-O cells (Figure 8D,8E, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). After 786-O cells were subcutaneously injected into mice for 5 weeks, the xenografted tumor was harvested, and the tumor volume and weight was calculated (Figure 8F, table available at https://cdn.amegroups.cn/static/public/tau-2025-442-3.XLSX). The results showed that compared to the sh-NC group, sh-BCAT1 and sh-IFNGR2 significantly inhibited the proliferation of ccRCC cells in vivo.

In this study, the GAPDH gene was repeatedly implicated as an important molecule in the development of ccRCC. Therefore, we explored the biological function of GAPDH in the proliferation of ccRCC using cell scratch and Transwell assays. Cellular assays showed that interference with GAPDH expression significantly inhibited the proliferation and invasion of ccRCC cells (Figure S11D-S11G).

Discussion

Traditional single-sample sequencing models often lack accuracy owing to tumor heterogeneity and are unable to cluster the most critical pathways involved in cancer development. In this study, we developed an effective prognostic model for ccRCC by integrating multiple datasets and identified 211 prognostic genes involved in the cancer metabolism pathways. The top 25 genes enriched in these pathways were significantly associated with survival prognosis, with the majority of genes enriched in metabolic pathways, downregulated in ccRCC, and acting as protective factors, including ACADM, HMGCL, AUH, HADHB, GAPDH, ACAA2, CPT1A, HMGCL, SHMT1, and SLC2A1. However, genes enriched in immune pathways, including BCAT1, LYN, WAS, ARC5, MARKS, VASP, FCGR2B, and RAC2, were upregulated in cancer and negatively correlated with the prognosis of patients with kidney cancer, identifying them as oncogenes in kidney cancers. Alternatively, the DEGs selected from large-scale datasets enhance accuracy. The pathways enriched by these genes more faithfully represent the key pathways underlying ccRCC tumorigenesis. Consequently, the ccRCC prognostic model derived from these genes exhibits superior generalizability. Furthermore, the model presented in this study demonstrates significantly enhanced predictive performance compared to models derived from either a single dataset or a single pathway gene set (31-33). Furthermore, PPI analysis indicated that GAPDH may serve as a pivotal hub gene in these immune and metabolic pathways. These observations highlight the importance of future research focusing on the relationship between tumor metabolic reprogramming and immunity in kidney cancer. Subsequently, we screened the most critical DEGs using machine learning and identified three cancer-related genes in ccRCC: BCAT1, GAPDH, and IFNGR2. Finally, the expression of these three genes in ccRCC cells was genetically reduced, and experiments demonstrated that interfering with their expression significantly inhibited the invasion and development of ccRCC. Critically, these top co-hub genes function not in isolation but as coordinated networks. Taking ‘tumor metabolic pathways’ as an exemplar, KEGG analysis of the top 25 genes reveals multifaceted dysregulation in the tumor microenvironment—including glucose/lipid metabolism and immunometabolic crosstalk. Genes enriched in immune pathways (e.g., BCAT1, LYN, WAS) demonstrated experimentally validated anti-tumor effects. Crucially, PPI analysis confirmed their interactions with metabolic genes and direct connections to GAPDH (Figure 4B). These findings demonstrate that tumorigenesis involves dynamic interplay between immune-metabolic circuits and cross-pathway metabolic adaptations. Studies targeting isolated metabolic gene sets often fail to address the plasticity of tumor microenvironments. Thus, elucidating GAPDH-centered metabolic-immune networks represents a critical priority for future renal cancer research.

Metabolic reprogramming is a current focus of cancer research, and fatty acid metabolism pathways play key roles in energy production, membrane synthesis, and signal transduction, all of which are essential for tumor cell proliferation (34,35). Studies have linked alterations in fatty acid metabolism to several cancers, including prostate (36), breast (37), colorectal (38), and lung cancers (39). In ccRCC, mutations in oncogenes disrupt amino acid, fatty acid, and energy metabolism (40). Approximately 90% of metabolic alterations in ccRCC are associated with VHL mutations and hypermethylation of its promoter region (41). This mechanism likely involves the loss of the VHL tumor-suppressor gene due to chromosome 3p deletions, resulting in increased VHL protein levels (42). VHL, as part of the E3 ubiquitin ligase complex, activates the HIF1 signaling pathway, which is involved in various oncogenic processes beyond metabolism, including cell survival, energy production, and angiogenesis. Furthermore, the HIF1 signaling pathway was the most significantly enriched pathway among the DEGs identified in this study (43,44).

Interventions targeting genes involved in fatty acid metabolism can directly inhibit the development of various tumors. This treatment approach is currently being evaluated in clinical trials (45). For example, targeting the interactions between fatty acid synthase and estrogen receptors can inhibit breast cancer progression (46). The modulation of metabolism to maintain a balance in fatty acid oxidation in cancer cells has become an important therapeutic strategy (47). The co-expressed DEGs identified in this study play a pivotal role in the metabolic reprogramming of tumors. For instance, GAPDH and BCAT1, which were identified as both top-co-DEGs and hub-DEGs, are recognized as hub genes in cancer metabolic pathways. BCAT1 (branched chain amino acid transaminase 1) is a key enzyme in amino acid metabolism that is overexpressed in multiple cancers, including breast, cholangiocarcinoma, esophageal, and hepatocellular carcinomas, and contributes to tumor growth, metastasis, cell cycle regulation, apoptosis, necroptosis, and angiogenesis in ccRCC (48,49). BCAT1 can transfer the amino groups on branched chain amino acids to alpha ketoglutarate, producing glutamate and alpha ketoacid, which are ultimately oxidized to provide energy for cells (50). This process is accelerated in tumors that overexpress BCAT1. For example, BCAT1 interacts with immune cells, accelerating amino acid metabolism, such as leucine catabolism in CD4⁺ T cells, to enhance energy availability in cancer cells (51). Recent studies have shown that BCAT1 overexpression in macrophages and RCC cells promotes RCC progression, increases immune-cell infiltration, and impedes immune regulation (52,53), which is consistent with our ssGSEA results.

Similarly, aerobic glycolysis, a hallmark of cancer progression, is driven by HIF1α suppression of mitochondrial function, with GAPDH being a key regulator. GAPDH is a key enzyme in glycolysis that plays a central role in basal metabolism. GAPDH not only regulates lactate levels but also promotes angiogenesis and programmed cell death in cancer, influencing the transcriptional and post-transcriptional processes in cancer cells (54). This enzyme is widely distributed in different cells of the body. A pan-cancer study showed that GAPDH is generally highly expressed in most tumor cells, and its expression level is strongly and negatively correlated with the survival prognosis of patients with cancer (55). In addition, highly expressed GAPDH is a risk factor for disease-free interval (DFI) and progression-free interval (PFI) in numerous cancers, similar to the results of this study. In several recent studies, scientists have attempted to interfere with the target of this gene in the glycolysis pathway using various methods to regulate tumor metabolism and inhibit tumor development (56,57). For example, by reducing the concentration of extracellular pyruvate in tumor cells, the Warbug effect mediated by HIF-1α can be inhibited, and by increasing the mitochondrial NADH/NAD⁺ ratio, the activity of the glycolytic enzyme GAPDH can be inhibited, thereby preventing the proliferation and development of various tumor cells, including ccRCC (58). Other studies have shown that 5-HT can directly induce serotoninization of GAPDH in CD8⁺ T cells, increasing glucose metabolism and anti-tumor immune activity (59). Network pharmacology has predicted GAPDH as a potential core target in RCC, as it inhibits the PI3K/Akt pathway and suppresses cancer cell growth (60). Our PPI results similarly showed that GAPDH is a crucial link between the immune and metabolic pathways. These findings suggest new directions for immune modulation in RCC. Other fatty acid metabolism genes, such as SLC2A1 (61), IGF2BP1 (62), CAPT1A (63), and ACADM (64), are also strongly associated with ccRCC development, underscoring their critical roles in metabolic pathways such as fatty acid metabolism, amino acid metabolism, and glycolysis.

The results of the ssGSEA indicated a significant correlation between metabolic pathway genes and the tumor immune microenvironment. This provides further evidence for the hypothesis that metabolic reprogramming and immune-cell infiltration play roles in cancer development and emphasizes the importance of metabolic pathways in immune-cell infiltration. The maintenance of metabolic homeostasis within cells not only provides the body with a robust innate immune defense but also facilitates muscle regeneration, tissue damage repair, and resistance to the effects of aging (65,66). The interaction between metabolic homeostasis and the tumor immune microenvironment may represent a novel avenue for elucidating the mechanisms underlying tumor formation and progression. Augmented fatty acid metabolism can result in immune-cell-mediated tumor evasion, thereby facilitating the invasive growth of multiform tumors (67). For example, CPT1A, a rate-limiting enzyme in fatty acid oxidation, functions in conjunction with L-carnitine derived from tumor-associated macrophages to inactivate CD8⁺ T cells in lung cancer, thereby promoting tumor infiltration and development (68). In this study, GAPDH emerges as a core gene regulating multiple metabolic pathways (e.g., glucose and lipid metabolism), underscoring its pivotal role in ccRCC metabolic reprogramming. Current research increasingly focuses on crosstalk among distinct metabolic pathways within the tumor microenvironment. Rather than targeting isolated pathways, identifying common therapeutic nodes across interconnected metabolic networks represents a promising anticancer strategy. Within this framework, GAPDH constitutes a compelling therapeutic target. Notably, a hepatocellular carcinoma study developed a novel glucose-lipid metabolism signature incorporating GAPDH, which demonstrated enhanced prognostic accuracy for tumor progression (69). Compelling evidence further indicates that short-chain fatty acids can upregulate GAPDH expression, enhance aerobic glycolysis, and alter TH1 cell polarization (70). This mechanistically reinforces the role of GAPDH as a pivotal hub integrating multiple metabolic pathways. GAPDH was identified as a key hub gene between the immune and metabolic pathways in this study. The precise mechanism by which it alters tumor and immune-cell subtypes via changes in intracellular metabolism remains unclear. However, recent studies have indicated that GAPDH can influence macrophage polarity by regulating glycolysis (71,72). In response, Jürgens et al. (73) discovered that the binding of GAPDH to siRNA can facilitate the transformation of a subset of M2 tumor-associated macrophages into tumor-suppressor M1 macrophages. Another study demonstrated that GAPDH can induce the polarization of M2 macrophages by binding to tumor-derived exosomes, thereby creating an immunosuppressive tumor microenvironment and improving the prognosis of patients with liver cancer (74), which is in accordance with the ssGSEA results of this study. Furthermore, macrophages secrete colony-stimulating factor-1, a crucial cytokine for tumor progression. In ovarian cancer, the downregulation of GAPDH promotes the stability of colony-stimulating factor-1 mRNA, which may contribute to tumor progression (75). GAPDH has been found to be activated and ubiquitinated in hepatocellular carcinoma, thereby increasing the classical immune signaling pathway NF-κB and HIF-1α transcription, ultimately enhancing the invasiveness of tumor cells (76). This mechanism has been well studied in hematologic cancers (77,78), but less so in solid tumors. This may be a new direction for future exploration of ccRCC. Collectively, this evidence suggests that research on kidney cancer should not exclusively focus on changes in the tumor immune microenvironment. Rather, it should be combined with multiple metabolic pathways, such as fatty acid metabolism and glycolysis, to explore the macroscopic mechanisms underlying tumor development. This approach may facilitate the identification of effective markers for predicting cancer and identifying targets for cancer treatment.

Our study presents an effective prognostic model for ccRCC driven by metabolic reprogramming genes, which could aid in predicting patient outcomes. However, there are some limitations in this study, such as its reliance on public databases, which may introduce individual biases despite integrating diverse sequencing data. The prognostic model validation set, GSE167573, referenced in this study was derived from RCC rather than ccRCC. Further clinical validation and experimental studies, including in vivo and in vitro models, are required to validate these findings and explore the specific molecular mechanisms underlying ccRCC progression.

Conclusions

In conclusion, we developed a reliable prognostic model for ccRCC by integrating multiple public datasets and robustly validated its accuracy using several approaches. These findings suggest that studying changes in cancer cell metabolic pathways may be a new area to improve the understanding of the molecular mechanisms of ccRCC and to seek targeted therapies. This model provides a precise tool for identifying ccRCC and establishes a solid foundation for future clinical applications.

Supplementary

The article’s supplementary files as

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. All animal experiments were performed under a project license (No. 202503003) granted by the ethics committee of Guangxi Medical University, in compliance with Chinese national guidelines and institutional guidelines for the care and use of animals.

Data Sharing Statement

Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-442/dss

tau-14-10-2805-dss.pdf ^{(80KB, pdf)}

DOI: 10.21037/tau-2025-442