Multi-omics and machine learning integration of diverse cell death pathways optimize risk stratification and inform drug therapy in Wilms tumor

Abstract

Introduction

Despite significant improvements in the overall survival of Wilms tumor (WT), a subset of patients still experiences poor outcomes. Programmed cell death (PCD) pathways are pivotal in cancer progression. A deeper understanding of their roles in WT is crucial for harnessing these mechanisms to optimize risk stratification.

Methods

Key tumor-associated genes were identified through limma differential analysis and WGCNA, and subsequently integrated with 12 distinct PCD patterns. TARGET-WT transcriptomic data was divided into training and validation sets to construct and validate a risk stratification model. It was subsequently integrated with clinical information to build a comprehensive prediction model. Immune infiltration and drug sensitivity analyses were performed. External validation was performed using scRNA-seq data from GSE200256.

Results

Key tumor-associated genes were enriched in multiple PCD pathways. The risk stratification model was constructed using 4 genes selected via the machine learning algorithm, stratifying the cohort into high- and low-risk groups. In the overall WT cohort, the high-risk group exhibited worse prognosis, with 1-, 3-, and 5-year AUC values of 0.820, 0.721, and 0.728, respectively. DCA demonstrated the superior predictive accuracy of the comprehensive prediction model. The high-risk group showed lower infiltration of TH17 cells and increased sensitivity to paclitaxel and sorafenib. Finally, the expression landscape of hub genes was validated in the single-cell dataset.

Conclusion

These results highlight a critical role for PCD genes in the progression and immune regulation of WT. Targeting these genes offers a promising avenue for improving clinical management of patients identified as high-risk.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-05326-0.

Introduction

Renal tumors account for 7% of all pediatric cancers, with WT representing the predominant subtype [1]. Advances in multimodal therapeutic strategies, including surgery, chemotherapy, and radiotherapy, have improved the overall survival rate to over 90% [2]. However, a subgroup of WT patients still exhibits poor prognosis, with 4-year survival rates ranging from only 30% to 85% [3]. These patients are typically diagnosed at advanced stages (III/IV), display unfavorable histological features such as diffuse anaplasia, or harbor specific molecular aberrations [4–6]. Unfortunately, existing pathological and molecular classification systems remain inadequate for the early and accurate identification of these high-risk patients and for guiding targeted therapeutic interventions [1, 6]. This limitation highlights the urgent need for a comprehensive investigation into the key molecular mechanisms driving WT metastasis and progression.

The application of machine learning methods serves as a driving force for improving cancer prognosis prediction, investigating the tumor immune microenvironment, and exploring personalized treatment strategies [7]. Simultaneously, the emergence of multi-omics technologies (particularly genomics and transcriptomics) has generated vast amounts of molecular data relevant to WT [8]. Furthermore, the integration of multi-omics data with machine learning techniques is emerging as a growing trend in renal tumor research [9]. The successful implementation of this integrated approach paves the way for refined risk stratification, more reliable prognostic evaluation, and ultimately, informed selection of the most effective treatment strategies [10, 11].

Over decades of investigation, pathways of programmed cell death (PCD) have been definitively characterized as fundamental mechanisms underlying pathogenesis and metastasis of malignant tumors. Acquired resistance to PCD constitutes a defining hallmark of cancer progression and therapeutic failure [12]. Consequently, multiple prognostic models incorporating genes related to various PCD modalities, such as pyroptosis, cuproptosis, and ferroptosis, have been successfully developed for WT [13–15]. However, most of the existing models are based on a single gene signature and fail to capture intricate crosstalk among multiple pathways. Given that tumor progression often involves complex interconnections across diverse molecular processes, such one-dimensional analyses may not fully elucidate the key drivers of disease evolution. Therefore, integrating multi-omics data with machine learning algorithms to comprehensively decode the PCD regulatory network is crucial for improving risk stratification, identifying novel therapeutic targets, and ultimately enhancing survival outcomes in children with WT.

Methods

Collection and processing of bulk transcriptomic data and acquisition of the programmed cell death (PCD) gene set

The RNA sequencing dataset for WT samples was obtained from the public TARGET database [16]. Samples fulfilling the following criteria were enrolled into this study: (a) complete availability of clinical information, and (b) exclusion of technical replicates where necessary. The final integrated dataset constituted 125 deduplicated tumor samples alongside 6 normal samples. Subsequently, protein-coding genes were selected. RNA-seq TPM data were subsequently processed through normalization and log2 transformation employing the “limma” R package, which generated the definitive expression profiles utilized in subsequent analyses [17].

The PCD gene set, encompassing 12 distinct modalities, was obtained from the Molecular Signatures Database (MSigDB), the Kyoto Encyclopedia of Genes and Genomes (KEGG), the FerrDb database and a manually curated list [18–21] (Supplementary Table 1). The flowchart of the entire study is presented in Fig. 1.

Fig. 1.

The flowchart of this study

Identification of key tumor-associated genes

The “limma” R package was employed to perform differential expression analysis, aiming to identify differentially expressed genes (DEGs) between WT samples and healthy controls [17]. DEGs were filtered using the criteria of |log2 Fold Change (FC)| > 1 and an adjusted P-value (adj.P.Val) < 0.05.

The “WGCNA” R package was employed to deeply investigate key gene clusters determining WT phenotypes [22]. Following the construction of a weighted adjacency matrix (soft threshold power = 5) and its transformation into a topological overlap matrix (TOM), dynamic tree cutting was employed to pinpoint the module most strongly linked to WT characteristics. Genes from this key module were then intersected with DEGs. To elucidate biological functional heterogeneity among these genes, Gene Ontology (GO) enrichment analysis alongside KEGG pathway analysis were performed using the “clusterProfiler” R package [23–25]. Since pathways associated with PCD were significantly enriched, tumor-associated genes were delineated by the overlap between these genes and the PCD gene set.

Development and validation of the risk stratification model for WT

Based on the distribution of overall survival status, the cohort was partitioned into a training set (n = 63) and a validation set (n = 62). In the training set, multiple machine learning algorithms, including univariate Cox regression (UniCox), LASSO regression, and multivariate Cox regression (MultiCox), were applied to identify key tumor-associated genes and to construct an optimal risk stratification model (using the “survival” and “glmnet” R packages) [26]. Finally, the risk score for each patient was calculated using the following formula:

βi represents the regression coefficient for each gene, and Xi represents the log2-transformed TPM value of each gene.

Employing the median risk score, cohorts were divided into high- and low-risk subgroups. Kaplan-Meier analysis was then implemented using the “survival” and “survminer” R packages to ascertain prognostic value of the stratification. We used the “timeROC” R package to generate receiver operating characteristic (ROC) curves and calculate area under the curve (AUC), thereby quantifying accuracy and robustness of the risk stratification [27].

Development and validation of the comprehensive prediction model

UniCox and MultiCox were stepwise employed to analyze risk groups and clinical characteristics, aiming to pinpoint critical variables with significant bearing on clinical outcomes. In light of these results, a nomogram was constructed to estimate survival probabilities at 1, 3, and 5 years. Calibration curves and DCA, executed through “rmda” and “caret” R packages, were employed to appraise the model’s predictive capability [28].

Tumor immune infiltration and drug sensitivity analysis

The immune infiltration landscape of the cohort was comprehensively characterized using the “ImmuCellAI” algorithm, which leverages single-sample gene set enrichment analysis (ssGSEA) to calculate enrichment scores for immune cell-specific gene signatures. This enabled comparison of overall immune infiltration and the relative abundance of 24 immune cell subtypes between normal and tumor tissues, as well as between the high- and low-risk groups [29]. Furthermore, the “oncoPredict” R package, which employs a ridge regression model trained on the large-scale cell line screening data from the Genomics of Drug Sensitivity in Cancer project, was applied to assess the correlation between gene expression signatures and drug response phenotypes in the high- and low-risk groups [30]. Both analyses were based on in silico predictions.

Validation of the expression landscape of four core genes via single-cell RNA sequencing (scRNA-seq)

ScRNA-seq of WT was obtained from GSE200256 in the GEO database. Specifically, samples GSM6025609, GSM6025610, and GSM6025611 were classified as favorable histology WT samples (FWT), while samples GSM6025604, GSM6025605, and GSM6025606 were classified as anaplastic histology WT samples (AWT). The age at diagnosis was comparable among patients within each histological group. Data processing was performed using the “Seurat” (v5.1.0) R package [31]. Quality control was applied by retaining genes detected in at least 3 cells and cells expressing a minimum of 300 genes. Furthermore, an upper limit of 7500 genes detected per cell was set to filter out potential doublets. Cells with a mitochondrial gene expression ratio surpassing 10% were likewise removed to eliminate low-quality or dying cells. This process resulted in a final dataset comprising 29,962 cells for subsequent analysis. After batch effect correction using the “harmony” R package, dimensionality reduction and clustering were performed with the “RunUMAP” function. Finally, cell populations in the clustering results were annotated based on authoritatively reported marker genes [32–34].

Statistical analysis and data visualization

All statistical analyses were performed using R (version 4.4.1) and GraphPad Prism. Comparative analysis between groups was performed using the Student’s t-test, the Mann-Whitney U test, the Wilcoxon test, or the log-rank test, as appropriate. The correlation analysis was performed using Spearman’s correlation coefficient. Data visualization was conducted employing the ggplot2 package in R and GraphPad Prism.

Results

Identification and analysis of the key tumor-associated genes

In the TARGET-WT cohort, 5,174 DEGs were detected, comprising 2,537 upregulated and 2,637 downregulated transcripts in WT samples (Fig. 2A). In Supplementary Table 2 is presented a comprehensive catalog of the DEGs. The expression spectrum of the top 30 genes exhibiting the greatest upregulation and downregulation across tumor samples is shown (Fig. 2B). To identify hub gene module most relevant to WT phenotypes, WGCNA was performed on the entire cohort. A scale-free topology emerged in the network once the soft-thresholding power β was assigned a value of 5 (Fig. 2C). From the TOM-based correlation analysis linking tumor attributes with gene modules, it was turquoise module that demonstrated the most pronounced relationship with WT phenotypes (R = 0.81, p-value < 0.001) (Fig. 2D, E). The notably strong correlation between turquoise module membership and gene significance for WT traits (R = 0.89, p < 0.001) (Fig. 2F) further indicated that genes within this module were closely related to the molecular mechanisms underlying tumor pathogenesis. A total of 1873 genes were obtained from the turquoise module. GO and KEGG enrichment analyses performed on the 1,236 overlapping genes between the DEGs and the turquoise module demonstrated significant enrichment in multiple PCD-related pathways, including “fatty acid metabolic process”, “Apoptosis”, “cell cycle checkpoint signaling”, “TNF signaling pathway”, and “p53 signaling pathway” (Fig. 2G). Taken together, these results highlight pivotal role of PCD-related genes in WT pathogenesis. Consequently, overlap among DEGs, turquoise module genes, and the PCD gene set revealed 106 key tumor-associated genes (Fig. 2H).

Fig. 2.

Identification of the key tumor-associated genes. A, B Volcano plot and heatmap of the TARGET-WT cohort; C Selection of a soft-thresholding power utilizing the scale-free topology model; D Associations between gene co-expression modules and clinical traits; E Hierarchical clustering dendrogram; F Correlation between gene significance for the tumor traits and module membership in the turquoise module.; G GO and KEGG enrichment analyses of DEGs and turquoise module genes; H Venn diagram showing the overlap among DEGs, turquoise module genes, and PCD-related gene sets

The risk stratification model based on the key tumor-associated genes predicts outcome events in the WT cohort

Within the training cohort, UniCox analysis pinpointed 11 genes that were markedly linked to survival outcomes (P < 0.05) (Fig. 3A). It was through LASSO regression and MultiCox that four independent biomarkers emerged as the strongest predictors of WT prognosis (Fig. 3B–D). Based on these results, a risk stratification model was constructed according to the following formula:

Fig. 3.

Construction and validation of the risk stratification model. A UniCox of the key tumor-associated genes; B, C LASSO regression analysis; D MultiCox; E–G Training set: risk score distribution, survival status plot, and heatmap of gene expression; Kaplan-Meier survival curve between high- and low-risk groups; ROC curves at 1, 3, and 5 years; H–J Validation set: risk score distribution, survival status plot; Kaplan-Meier survival curve between high- and low-risk groups; ROC curves at 1, 3, and 5 years

Using the median risk score as the cutoff value, the training samples were stratified into high- and low-risk groups. HMGB1 and IDUA were more abundant in the high-risk group, while CALCOCO2 and DNASE2 showed increased expression in the low-risk group (Fig. 3E). Analysis of the risk score and survival distribution plots revealed a marked increase in death events with higher risk scores, indicating a strong association between the risk score and patient outcomes (Fig. 3E). To assess the prognostic discrimination of the risk stratification model, Kaplan-Meier survival analysis demonstrated that the high-risk group had a significantly poorer survival rate compared to the low-risk group (P = 0.0016) (Fig. 3F). Furthermore, ROC curve revealed the model performed robustly, yielding AUC values of 0.877, 0.809, and 0.806 at 1, 3, and 5 years (Fig. 3G).

Inspection of the risk score and survival distribution profiles in the validation cohort likewise mirrored the patterns observed in the training cohort (Fig. 3H). Stratification of the validation cohort into high- and low-risk groups was likewise performed according to the median risk score. Kaplan-Meier analysis confirmed a statistically significant difference between the two groups (P = 0.031) (Fig. 3I), accompanied by AUC values of 0.753, 0.626, and 0.649 for 1, 3, and 5 years (Fig. 3J).

The development of the comprehensive prediction model

Moreover, accuracy of the risk stratification was corroborated across the entire TARGET-WT cohort (Fig. 4A, B). Subsequently, entire cohort was utilized for the construction of comprehensive prediction model. Four clinical characteristics: tumor stage, gender, histological type, and age at diagnosis, were included as potential prognostic factors. When coupled with the risk groups defined by median risk score, UniCox indicated that only gender, tumor stage, and the risk classifications showed significant relevance to prognosis in children with WT (P < 0.05). Notably, risk groups exhibited a higher hazard ratio compared to other clinical features (Fig. 4C). To control for potential confounding factors, variables with prognostic significance in the UniCox were incorporated into the MultiCox, which further confirmed that risk groups served as an independent prognostic factor for WT patients (Fig. 4D). Thereafter, a nomogram was formulated using tumor stage, risk groups, and gender to predict 1-, 3-, and 5-year survival probabilities in pediatric WT patients (Fig. 4E). Calibration curve, in turn, attested to dependability of forecasts generated by the comprehensive prediction model (Fig. 4F). DCA likewise verified that this model yielded greater clinical net benefit than other prognostic metrics across the spectrum of threshold probabilities (Fig. 4G).

Fig. 4.

Construction of comprehensive predictive model. A, B Kaplan-Meier survival analysis between high- and low-risk groups in the overall WT cohort; ROC curves at 1, 3, and 5 years; C, D UniCox and MultiCox incorporating the risk groups and clinical characteristics; E Nomogram integrating the risk groups, tumor stage, and gender; F, G Calibration curve and DCA of the comprehensive predictive model

Immune infiltration variations between high- and low-risk groups

To delineate immune infiltration landscape, 24 distinct immune cell populations were profiled across the entire TARGET-WT cohort. Marked distinctions became evident when contrasting immune cell abundance across tumor and normal samples (Fig. 5A). Moreover, a significantly elevated overall immune infiltration score was detected in normal samples (Fig. 5B). Among the 125 tumor samples, discernible variation in immune cell infiltration likewise emerged when contrasting individuals assigned to the high- and low-risk groups (Fig. 5C). Consistently, the high-risk group exhibited a lower overall immune infiltration score (Fig. 5D). Furthermore, a significant negative correlation was observed between the risk score and the immune infiltration score of each sample. Notably, the infiltration proportion of TH17 cells was higher in normal samples than in tumor samples, and similarly elevated in the low-risk group compared with the high-risk group. Consistent with this trend, the risk score was negatively correlated with the TH17 cell infiltration proportion across all samples. Collectively, these findings suggest that the four key hub genes in our prognostic model may influence patient outcomes by modulating TH17 cell abundance, thereby reshaping the immune microenvironment within high-risk tumor tissues (Fig. 5E).

Fig. 5.

Immune infiltration analysis. A–D Differences in the infiltration levels of 24 immune cell types and the overall immune infiltration score between normal and tumor samples, as well as between the high- and low-risk groups. E Correlation between the expression levels of the four hub genes/risk score and the infiltration proportions of 24 immune cell types. Line thickness indicates the P value, and line color intensity represents the strength of the association. (P* < 0.05, P** < 0.01, P*** < 0.001, P**** < 0.0001)

Variations in drug sensitivity between high- and low-risk groups

To forecast prospective therapeutic agents for pediatric patients harboring high-risk WT, we calculated the half-maximal inhibitory concentration (IC50) of multiple compounds for each specimen, pinpointing those exhibiting notable disparities between the high- and low-risk cohorts. In the high-risk cohort, IC50 values were notably lower, indicating increased sensitivity to chemotherapeutic agents recommended in WT treatment protocols, including Dactinomycin and Vincristine (Fig. 6A, B) [35, 36]. Additionally, Paclitaxel and Sorafenib were highlighted as promising agents for children classified as high-risk WT (Fig. 6C, D).

Fig. 6.

Drug sensitivity analysis. A–D Differences in the IC50 of various drugs between the high- and low-risk groups, and Spearman correlation coefficients between the risk score and drug IC50 values (P* < 0.05, P** < 0.01, P*** < 0.001)

ScRNA-seq validation of the four hub gene signatures

We analysed the scRNA-seq data from patients with FWT and AWT. All samples were clustered into four major clusters: Wilms Tumor Cell exhibiting high expression of “WT1, NCAM1, PAX2”; Cancer Associated Fibroblast with high expression of “ACTA2, COL1A1, PDGFRB”; Immune Cell characterized by high expression of “PTPRC”; and Endothelial cell showing high expression of “CLDN5, ENG, VWF” (Fig. 7A, B). Cell composition analysis revealed a higher proportion of Wilms Tumor Cell in AWT, whereas Endothelial cell and Immune Cell were more abundant in FWT samples (Fig. 7C). We next examined the expression patterns of the four model genes across clusters in both tumor types. CALCOCO2 and DNASE2 were predominantly localized within the Endothelial cell of FWT samples. In contrast, IDUA was mainly expressed in the Endothelial cell of AWT samples, and HMGB1 was primarily expressed in the Wilms Tumor Cell (Fig. 7D–G).

Fig. 7.

Validation of hub genes expression through scRNA-seq. A Cluster annotation; B UMAP plot; C Bar plot comparing the cellular composition between FWT and AWT samples; D–G Expression patterns of individual hub genes in FWT versus AWT samples

Discussion

In recent years, the integrative application of diverse bioinformatics algorithms and machine learning models, combined with multi-omics and multi-batch data, has provided valuable insights for elucidating disease mechanisms and developing therapeutic strategies [37, 38]. In our study, integration of WGCNA with limma differential analysis allowed the identification of key tumor-associated genes. A four-gene based risk stratification model was subsequently developed in the WT training cohort through machine learning approaches. Patients were categorized into high- and low-risk groups by this model, which displayed robust prognostic accuracy. The model’s predictive performance was then validated in the WT validation cohort. Considering the limited availability of multi-center WT cohorts with high-quality sequencing data, a random sampling-based internal validation strategy was employed to minimize the risk of model overfitting. Ultimately, a comprehensive prediction model incorporating key clinical variables was established, achieving accurate and robust prediction of survival outcomes in pediatric WT patients. Furthermore, differences in tumor immune microenvironment and drug sensitivity between high- and low-risk groups were systematically investigated. The expression landscape of the key PCD-related hub genes was also validated using an external single-cell RNA sequencing dataset.

Various PCD pathways exert diverse and critical roles in the development and progression of malignant tumors [39, 40]. In our previous work, we specifically investigated the role of pyroptosis, a distinct form of PCD, in WT [14]. In contrast, the present study focuses on the crosstalk among multiple PCD pathways and provides a comprehensive evaluation of their biological relevance in WT progression.

Ultimately, we identified four risk classification genes: DNASE2 and IDUA from the lysosome-dependent cell death pathway, CALCOCO2 from the autophagy pathway, and HMGB1 which is associated with multiple pathways including cuproptosis, necroptosis, autophagy and pyroptosis.

DNASE2, which functions in the degradation of DNA derived from dead cells [41, 42], was identified as a protective factor in our classification model. Interestingly, a pan-cancer study reported high DNASE2 expression in glioblastoma multiforme cells [43]. Such observations could result from the complex crosstalk between DNASE2 and tumor microenvironment. IDUA, a lysosomal hydrolase, was identified as a risk factor in our model. Consistently, knockdown of IDUA has been shown to suppress the proliferation and migration of renal cell carcinoma cells [44, 45]. CALCOCO2, a regulator of xenophagosome maturation [46], is transcriptionally modulated by the nuclear factor NFE2L2 [47]. Targeting the CALCOCO2-associated pathway plays a critical role in a range of cancer-related processes, including cancer metabolism, chemoprevention, and cancer immunotherapy [48, 49]. HMGB1 can antagonize immune cell activation triggered by PCD [50, 51]. In the present study, HMGB1 was identified as a promoter of WT progression. Similarly, circulating HMGB1 mRNA levels in bladder cancer are closely associated with tumor prognosis [52]. In liver cancer, simultaneous induction of ferroptosis and neutralization of HMGB1 markedly enhances tumor cell sensitivity to immune checkpoint inhibitors [53, 54]. Therefore, a therapeutic strategy that combines induction of PCD with HMGB1 inhibition may offer a promising adjuvant option for patients with high-risk WT.

The tumor immune microenvironment is widely recognized as a pivotal factor influencing tumor initiation, progression, and patient prognosis [55]. The composition of specific immune cell types determines the prevailing orientation of immune responses toward either pro-tumor or anti-tumor activity [56]. With the rapid accumulation of multi-omics datasets, current bioinformatics frameworks have become essential for characterizing the immune landscape of diseases [57]. Our findings highlight a significant link between diminished Th17 cell infiltration and heightened tumor aggressiveness within WT immune microenvironment. Th17 cells, a distinct CD4⁺ T cell subset, can exert anti-tumor effects either by recruiting other immune effector cells or by transdifferentiating into Th1-like cells that produce interferon-gamma (IFN-γ) [58]. Consistently, a large-scale study characterizing the evolutionary modes of clear cell renal cell carcinoma demonstrated that patients with extended overall survival harbored elevated Th17 cell infiltration [59].

The clinical management of children with high-risk WT remains a major challenge [60]. This study further substantiated the pivotal roles of vincristine and dactinomycin in the therapeutic framework for high-risk WT [35, 36]. However, current therapeutic regimens that primarily rely on conventional cytotoxic agents are reaching the tolerance limits of pediatric patients, with some children requiring dose reductions due to unacceptable toxicity [6, 61, 62]. Consequently, there is an urgent need to explore agents with novel mechanisms of action that can enhance survival outcomes in high-risk patients while minimizing treatment-related adverse effects. Case reports have suggested that paclitaxel and the multi-kinase inhibitor sorafenib may offer therapeutic benefit in the palliative treatment for children with high-risk relapsed WT [63, 64]. Our analysis supports these findings, yet the efficacy of these agents awaits validation in prospective clinical cohorts.

Although the proposed model exhibits clear advantages in risk stratification for pediatric WT, several limitations should be noted. First, long-term follow-up data from larger, multi-center pediatric cohorts are necessary to achieve robust clinical validation. Second, given that the present study primarily relies on retrospective data and in silico analyses, the biological roles of the model-associated genes warrant further experimental verification both in vivo and in vitro. Third, the sequencing data for children with WT are limited in scale; therefore, in constructing the model, the entire cohort was split into training and validation sets in a single step, which may introduce a risk of model overfitting.

Conclusions

In summary, our risk stratification model enhances the molecular understanding of high-risk WT, demonstrating robust prognostic performance and predictive value for responses to immunotherapy and chemotherapy. Our findings elucidate the crucial role of PCD pathways in shaping WT prognosis and lay a foundation for improving risk assessment, guiding targeted therapy, and developing personalized treatment strategies for this pediatric malignancy.

Author contributions

Zj L and Qj S contributed equally to this study. Zj L and Qj S designed the study and wrote the manuscript. Jj G, Sy F, and Sy W performed the statistical analysis and created the figures. Jy P reviewed and edited the manuscript. Zz L acquired the funding. All authors reviewed the manuscript.

Funding

This work was supported by the Key Research and Development Program of Heilongjiang Province[2022ZX06C08].

Data availability

All data used in this study are publicly available. scRNA-seq data were obtained from the Gene Expression Omnibus database under accession number GSE200256 and are available at the following URL: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi? acc=GSE200256. TARGET-WT cohort were obtained from the Genomic Data Commons (GDC) and are available at the following URL: https://portal.gdc.cancer.gov/.

Declarations

Ethics approval and consent to participate

Ethical approval was not required for this study, as all data were obtained from publicly available databases. Patient consent was not required for this study, as all data were obtained from publicly available databases.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.