Multi-omics identification of a programmed cell death-related signature and potential target P4HB for bladder cancer based on a 101-combination machine learning and experimental validation

Abstract

Bladder cancer (BLCA) poses a significant clinical challenge due to its high mortality rates and the inadequacy of current prognostic biomarkers. Programmed cell death (PCD) is crucial in BLCA initiation, progression, and treatment, yet the interplay and specific roles of different PCD pathways in BLCA prognosis remain elusive. This study aimed to develop and validate predictive models by integrating 14 PCD patterns using comprehensive analyses of bulk RNA and single-cell RNA transcriptomic data from TCGA-BLCA and six GEO datasets. Through weighted gene co-expression network (WGCNA) analyses, 24 hub PCD-related genes (PCDGs) were identified in BLCA. Subsequently, we implemented a computational framework that integrated 10 machine learning algorithms along with 101 of their combined permutations. This framework was used to develop a programmed cell death-related signature (PCDRS). The final PCDRS consisted of 12 prognostic genes: P4HB, CHEK2, PTPN2, ATP13A2, CCT6A, TFRC, RRP12, TRAF7, POLR1B, B4GALT3, SIVA1, and TP73.The PCDRS was validated in training and external validation sets, with multivariate analysis confirming its independent prognostic value in BLCA. The PCDRS-integrated nomogram was also developed as a quantitative clinical tool. Furthermore, differences in reactive oxygen species (ROS) levels were observed in the tumor microenvironment between high- and low-risk groups based on PCDRS risk scores. Additionally, the elevated expression and tumorigenic role of P4HB in BLCA were validated through in vitro assays. In summary, P4HB may serve as a candidate gene with potential relevance to BLCA prognosis that could enhance personalized treatment strategies for patients with BLCA.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10238-026-02146-y.

Introduction

Bladder cancer (BLCA) ranked as the tenth most common malignant neoplasm in 2023, with higher incidence rates in males for whom it was the fourth commonest cancer [1]. Most cases of BLCA are urothelial carcinoma in subtype (> 90% of cases) [2]. The treatment of BLCA depends on whether it is muscle-invasive and the risk stratification, including transurethral resection of bladder tumor (TURBT) or radical cystectomy, while adjunctive therapies like cisplatin-based chemotherapy [3]. Some monoclonal antibodies targeting PD-L1, such as Pembrolizumab, Atezolizumab, Nivolumab, Durvalumab, and Avelumab, have also been introduced into clinical use, but overall response rates remain low [4]. Moreover, BLCA prognosis is intricate and diverse, marked by notable variations in survival outcomes based on patients’ disease stages and levels of differentiation [5]. Consequently, exploring robust biomarkers for prognostic prediction is crucial.

Programmed cell death (PCD) is a genetically orchestrated process entailing the intrinsic and sequential death of cells controlled by genes [6]. It is traditionally classified into apoptosis, autophagy, and programmed necrosis, and has recently been expanded to 14 types: disulfidptosis, apoptosis, entotic cell death, necroptosis, cuproptosis, oxeiptosis, ferroptosis, alkaliptosis, pyroptosis, parthanatos, autophagy-dependent cell death, lysosome-dependent cell death, netotic cell death, and immunogenic cell death (ICD) [7–9]. It has been reported that PCD plays a vital role in BLCA. For instance, dysregulation of autophagy impedes BLCA cell apoptosis and ferroptosis, thereby impacting tumor metastasis with epithelial-mesenchymal transition (EMT) mechanisms [10]. Ferroptosis, driven by iron overload and lipid peroxidation, was recently targeted in BLCA treatment using nanoparticle-based photodynamic immunotherapy to induce ferroptosis and immune stimulation [11]. Some studies have also explored several ICD inducers like Norcantharidin (NCTD), which are capable of triggering cell death in human BLCA cells and subsequently stimulating anti-tumor immune responses [12]. Nevertheless, the comprehensive association between these 14 PCD patterns and BLCA prognosis has yet to be fully elucidated.

To address this gap, our research characterized PCD activity profiles across both single-cell and bulk transcriptomes, and identified 24 hub PCD-related genes (PCDGs) in BLCA. Subsequently, we implemented an innovative machine-learning strategy that integrated 10 distinct algorithms along with 101 of their combined permutations. This framework was utilized to construct a unified PCD-related prognostic signature (PCDRS), designed to precisely forecast disease progression and survival in BLCA patients. Among these hub PCDGs, P4HB held the highest predictive ability of BLCA. Therefore, our study also delved into the function of P4HB in BLCA, and our study found that silencing P4HB in BLCA cells significantly inhibited cell proliferation and growth, suggesting a potential functional role for P4HB in BLCA, although the underlying mechanisms remain to be elucidated. This study provides deeper insights into the roles of PCD and P4HB in BLCA, paving the way for future research into targeted therapies.

Materials and methods

Data collection

Transcriptome sequencing data for relevant BLCA cases were acquired from The Cancer Genome Atlas (TCGA) and the Gene Expression Omnibus (GEO) databases. For the BLCA single-cell RNA sequencing data, GSE129845 (3 samples) and GSE135337 (8 samples) were included. Also, the comprehensive BLCA bulk RNA-Seq datasets were compiled, including GSE13507 (256 samples), GSE166716 (24 samples), GSE37815 (24 samples), and GSE154261 (99 samples). Figure 1 presented a workflow diagram of our study.

Fig. 1.

Flowchart of this study

Fourteen PCD types and their associated genes were manually collected from the published literature, including “Apoptosis”: 579, “Pyroptosis”: 52, “Ferroptosis”: 88, “Autophagy”: 367, “Necroptosis”: 101, “Cuproptosis”: 14, “Parthanatos”: 9, “Entotic cell death”: 15, “Netotic cell death”: 8, “Lysosome-dependent cell death”: 220, “Alkaliptosis”: 7, “Oxeiptosis”: 5, “Disulfidptosis”: 901, “Immunogenic cell death”: 57. After eliminating duplicate genes, a total of 2126 PCDGs were included for subsequent analysis (Table S1).

Single-cell RNA-sequencing analysis

We obtained single-cell RNA sequencing data from 11 BLCA samples in the GSE129845 (3 N: 3 Normal tissues) and GSE135337 (1N7T: 1 Normal and 7 Tumor tissues). During quality control (QC), cells were filtered by applying thresholds that included a mitochondrial gene percentage under 10%, and retaining only genes detected in at least three cells, with total gene counts per cell ranging from 200 to 7,000. Subsequently, a selection of 2000 highly variable genes were made for downstream analysis. After mitigating batch effects with the “Harmony” package, cells were clustered using the “FindNeighbors” and “FindClusters” packages, with visualization accomplished using the “t-SNE” method. Cell type identities were assigned based on established marker genes via a combination of manual annotation and the “singleR” R package. The quantification of gene set activity in each cell was conducted using the AddModuleScore algorithm.

Single-sample Gene Set Enrichment Analysis (ssGSEA) and Weighted Gene Co-Expression Network Analysis (WGCNA)

ssGSEA from the R package “GSVA” was used to calculated the PCD score for each BLCA sample. Univariate Cox regression analysis was conducted to determine the prognostic influence of individual PCD patterns. Additionally, the mutations across distinct PCD pathways and their implications on BLCA prognosis were performed with the “maftools” package. WGCNA was conducted employing the R package “WGCNA” on data from the TCGA-BLCA and GSE166716 datasets. For WGCNA, a soft-threshold power (β) was first chosen to achieve a scale-free network topology. And gene-gene correlation matrices were calculated, which served as the basis for constructing an adjacency matrix to represent inter-node connectivity. Gene modules were defined through the application of the dynamic tree-cutting algorithm, followed by hierarchical clustering, and a dendrogram was generated to visualize and define co-expression modules.

Machine learning approaches

To develop a prognostic signature through integrative machine learning techniques, differential analysis between normal and tumor samples was carried out for the bulk RNA-seq data from TCGA-BLCA, GSE13507, and GSE37815 using the “limma” R package (|logFC|>0.5 and p.adj < 0.05). The differentially expressed genes (DEGs) at the bulk RNA-seq level were intersected with the genes in the PCD-related module identified by WGCNA, denoted as hub PCDGs. To construct a robust prognostic model with high predictive accuracy, a series of methodical steps were followed:

1. Identification of prognostically relevant PCDGs: We conducted univariate Cox regression analysis to identify potential prognostic PCDGs in the TCGA-BLCA, GSE37815, and GSE13507 datasets.

2. For model development and evaluation, The TCGA-BLCA cohort was designated as the training set, while the GSE13507 and GSE37815 datasets served as independent external validation sets. We integrated ten distinct machine learning algorithms—including Lasso, Ridge, stepwise Cox, CoxBoost, RSF, Enet, plsRcox, SuperPC, GBM, and survival‑SVM—to construct a robust predictive framework. Each algorithm underwent 101 permutations to ensure stability, and a consensus signature was determined using a voting strategy, selecting genes frequently identified across algorithms.

3. Model assessment: The models generated were rigorously evaluated within both the internal and external validation sets. Each model’s performance was assessed using Harrell’s concordance index (C index), with the mean C index serving as the basis for ranking predictive efficacy, which was sorted from highest to lowest, and the results were visualized with a heatmap. The top-ranked model, demonstrating the highest predictive performance, was selected as the final predictive model. Subsequently, this final model served as a baseline for calculating the Net Reclassification Index (NRI) and the Integrated Discrimination Improvement (IDI) for other machine learning models, enabling an assessment of their predictive capabilities. Genes selected by the majority of algorithms with high selection frequency were included in the final signature, ensuring its robustness, reproducibility, and resistance to overfitting. Following meticulous evaluation, an amalgamation of algorithms demonstrating robust performance and clinical relevance was identified. This amalgamated approach culminated in the formulation of a definitive prognostic tool, denoted as the PCD-related signature (PCDRS), proficient in forecasting overall survival (OS) outcomes among BLCA patients.

Survival analysis and nomogram development

The TCGA-BLCA, GSE13507, and GSE378150 datasets were stratified into high- and low-risk cohorts based on the median PCDRS risk score. Subsequently, Kaplan-Meier (KM) curve analyses were conducted using the “survminer” R package to discern potential disparities in overall survival (OS) (log-rank test, p < 0.05). The receiver operating characteristic (ROC) curves were generated using the “timeROC” R package to evaluate the sensitivity and specificity of the PCDRS in predicting overall OS for BLCA patients. The association between PCDRS and diverse clinical attributes was further investigated through area under the curve (AUC) analyses. To determine whether the PCDRS serves as an independent prognostic factor for survival in BLCA patients, both univariate and multivariate Cox proportional hazards regression analyses were conducted. To heighten prognostic precision and predictive capacity, a nomogram integrating PCDRS and pertinent clinical features was formulated. This nomogram serves as a comprehensive tool for estimating the anticipated survival probabilities of individuals afflicted with BLCA.

Analysis of genomic variation between PCDRS risk subgroups

The Mutant Allele Tumor Heterogeneity (MATH) scores, a metric quantifying intratumoral genetic diversity, are typically computed from whole-exome sequencing data comparing tumor DNA with matched normal tissue samples. In this investigation, MATH scores were derived for every patient diagnosed with BLCA, in accordance with standard computational methodologies. Subsequently, survival analyses were conducted based on these MATH scores to probe potential correlations with patient outcomes. In an effort to delve into somatic mutations linked to PCD, we leveraged the “maftools” R package to generate waterfall plots.

Single-cell sequencing landscape of hub genes and pseudotime trajectory analysis

We identified 12 signature genes (P4HB, CHEK2, PTPN2, ATP13A2, CCT6A, TFRC, RRP12, TRAF7, POLR1B, B4GALT3, SIVA1, and TP73) as the hub genes for subsequent analyses. To investigate their activity patterns at single-cell resolution, we applied the “AddModuleScore” function to compute a hub_gene_score for each individual cell based on the combined expression of these 12 genes. This approach enabled the detection of distinct expression variations across different cell populations. Fifty hallmark gene sets were retrieved from The Molecular Signatures Database27. We summarized the involved genes of the ROSHALLMARK_REACTIVE_OXYGEN_SPECIES_PATHWAY from the GSEA-hallmark gene sets and applied the same approach to construct the ROS_genes_score.

To investigate the differentiation trajectories and stemness of tumor cells, CytoTRACE2 and Monocle2 were employed. CytoTRACE2 was used to assess the stemness of tumor cells, quantifying their differentiation potential based on the diversity of gene expression, where higher scores indicated less differentiated states. Monocle2 was utilized to reconstruct differentiation trajectories. Cells were subsequently ordered along a reconstructed pseudotime trajectory and projected into a low-dimensional space for visualization. The starting root of this trajectory was assigned according to predefined cell subcluster identities. Genes with dynamic expression patterns were visualized by the “plot_genes_in_pseudotime” function. To further resolve the branching architecture, we additionally performed trajectory inference using the Slingshot algorithm. The root clusters were specified as cluster4 and cluster2 based on CytoTRACE2 stemness scores, and lineages were constructed using default parameters.

Cell culture, lentiviral transduction, quantitative Real-Time Polymerase Chain Reaction (qRT-PCR), and western blot

Six BLCA cell lines (T24, RT4, UMUC3, TCC, J82, and BIU87), along with a human ureteral epithelial immortalized cell line (SV-HUC-1) were acquired from the Chinese Academy of Sciences (Shanghai, China). All cell lines were cultured in appropriate media supplemented with 10% fetal bovine serum. The lentivirus constructs for P4HB knockdown were obtained from OBIO (Shanghai, China). For lentiviral transduction, BIU87 and UMUC3 cells were seeded in 6-well plates at 50% confluence and infected with P4HB knockdown lentivirus (referred to as shP4HB), or a scramble control (referred to as shNC), respectively.

Total RNA was extracted from cultured cells or tissues employing a cellular RNA extraction kit (Vazyme, Nanjing, China). For qRT-PCR, the purified RNA was then reverse-transcribed into complementary DNA (cDNA), which served as the template for quantitative amplification performed on a StepOne real-time PCR system. For protein analysis, cellular proteins were subjected to separation by 10% SDS-PAGE. The resolved proteins were subsequently transferred onto a polyvinylidene fluoride (PVDF) membrane (Millipore, USA). After blocked, the membranes were then incubated overnight at 4 °C with primary antibodies targeting P4HB (dilution 1:1000, Cat. #3501, CST, USA). After probed with the corresponding secondary antibody, and protein signals were visualized using a chemiluminescence system (Bio-Rad Laboratories, USA).

Cell proliferation, colony formation, and flow cytometry assays

UMUC3 and BIU87 cells were seeded into 96-well plates at densities of 1,000 and 2,000 cells per well, respectively. Cell proliferation was monitored using the Cell Counting Kit-8 (CCK-8), with measurements taken at 450 nm absorbance after 1 h of incubation at 37 °C on days 0, 1, 2, 3, and 4. For colony formation assays, the same BLCA cell lines were seeded in 6-well plates and allowed to grow for approximately 10 days. The resulting colonies were then fixed with 4% paraformaldehyde and stained with crystal violet. For flow cytometry assays, cells were digested and stained with Annexin V-APC and PI for 30 min at room temperature. Subsequently, flow cytometry (BD, USA) was employed to evaluate cells apoptotic.

Statistical analysis

All statistical analyses were performed using R software (R 4.3.0). Comparisons between two groups for normally distributed variables were conducted using the Student’s T-test, whereas the Wilcoxon test was applied for non-normally distributed data. In cases involving multiple groups, the Analysis of Variance (ANOVA) analysis was employed. A p-value of less than 0.05 was considered statistically significant.

Results

Characteristics of PCD in single-cell RNA sequencing

Single-cell RNA sequencing was performed on samples derived from 11 BLCA patients, yielding a total of 55,401 individual cells. After removing batch effects and integrating the 11 samples, the cells were aggregated into 21 clusters (Fig. S1A). Then, we annotated the cells into seven major clusters, Macrophage: 12,118 cells; Cancer cell: 28,499 cells; Epithelial cell: 5763 cells; Mesenchymal cell: 6966 cells; T cell: 877 cells; B cell: 576 cells; Endothelial cell: 430 cells; unknown: 172 cells (Fig. 2A). Figure 2B displays the expression patterns of marker genes across individual cells. To assess PCD activity across various cell types, we employed the “AddModuleScore” function to evaluate the expression levels of a 2126-gene set associated with PCD in all cells (Fig. 2C). Subsequently, the ssGSEA algorithm was leveraged to assign a quantitative PCD activity score to each BLCA sample. Among the seven cell types analyzed, notably elevated PCD activity was observed in B cells and Cancer cells (Fig. 2D).

Fig. 2.

Molecular characteristics of programmed cell death at the single-cell level. (A) UMAP plots for 55,401 cells in 11 BLCA samples from 2 single-cell datasets (GSE129845: 3 N; GSE135337: 1N7T). (B) The expression patterns of marker genes within each individual cell. (C) The distribution of median PCD-scores. (D) B cells and cancer cells exhibited heightened PCD activity

Identification of PCD‑Associated Hub Modules and Genes from Bulk RNA‑Sequencing Data

To uncover PCD modules linked with BLCA, we conducted WGCNA on the TCGA-BLCA dataset. Initially, all 2126 PCDGS were utilized to establish a co-expression network following the removal of outlier samples (Fig. 3A). The optimal soft threshold was identified as a power of 5 (R² = 0.9), which yielded a scale-free topological network (Fig. S1B). With the minimum module gene count set to 50 and MEDissThres set to 0.25, a total of five distinct PCD modules were delineated (Fig. 3B). Notably, our results highlighted a positive correlation between the MEbrown module and BLCA (cor = 0.38, p = 1e − 24, Fig. 3C). Furthermore, the scatter plot illustrating gene significance (GS) versus module membership (MM) for the MEbrown module exhibited a significant association with BLCA (cor = 0.51, p = 3.6e − 14, Fig. 3D).

Fig. 3.

The identification of hub modules in bulk RNA sequences of TCGA-BLCA based on WGCNA. (A) Dendrogram showing the hierarchical clustering of TCGA-BLCA samples. (B) Cluster dendrogram of the WGCNA analysis. (C) A positive significant correlation (cor = 0.38) between the MEbrown module and BLCA. (D) A highly significant correlation observed in the scatter plot of gene significance (GS) and module membership (MM) within the MEbrown module (cor = 0.51, p = 3.6e − 14). (E) Venn diagram shows the modular genes intersection of TCGA-BLCA and GSE166716. (F) Venn diagram shows the intersection of module genes from WGCNA, TCGA-BLCA and GSE166716 datasets. (G) 24 PCDGs were summarized to distinct PCD patterns. (H) GO enrichment analysis of these 24 hub genes

Similarly, we conducted WGCNA on the GSE166716 dataset to pinpoint the PCD modules significantly linked with BLCA. Specifically, following the exclusion of outlier samples, the PCDGs were used to construct a gene co-expression network (Fig. S2A). For the establishment of a scale-free topological network, a power of 12 (R² = 0.9) was deemed as the optimal soft threshold (Fig. S2B). With the minimum module size set to 50 genes and the MEDissThres set to 0.25, a total of seven distinct modules were identified (Fig. S2C). Our findings indicated strong correlations between BLCA and the MEturquoise module (cor = 0.74, p = 3e-05), the MEbrown module (cor = 0.47, p = 0.02), and the MEgreen module (cor = 0.44, p = 0.03, Fig. S2D). Moreover, highly significant correlations Moreover, highly significant correlations were observed in the scatter plot of GS and MM within the MEbrown module (cor = 0.31, p = 3e − 7, Fig. S2E), MEgreen module (cor = 0.21, p = 0.015, Fig. S2F), and MEturquoise module (cor = 0.72, p = 1.8e − 82, Fig. S2G).

Next, we picked the modular PCDGs associated with BLCA via intersection of TCGA-BLCA and GSE166716 with Venn diagram (Fig. 3E). Then, we intersected the module PCDGs from WGCNA with the DEGs from the TCGA-BLCA and GSE166716 datasets, finally identifying a total of 24 hub PCDGs (Fig. 3F, Table S2). These PCDGs, which were considered to be involved in BLCA at both TCGA-BLCA and GSE166716, and were summarized to distinct PCD patterns (Fig. 3G). GO and KEGG enrichment analysis of these 24 PCDGs indicated significant enrichment across multiple biological processes (BP), including response to manganese ion, and chaperone−mediated autophagy, as well as in cellular components (CC) such as lysosomal lumen, and in molecular functions (MF) including disulfide oxidoreductase activity (Fig. 3H, Table S3).

Prognostic risk profiles of the hub PCDGs

Subsequently, we performed univariate Cox regression analysis on 24 PCDGs using TCGA-BLCA cohort, identifying 12 significant genes affecting the OS of BLCA, including ATP13A2, CCT6A, CHEK2, HSP90AA1, P4HB, POLR1B, PTPN2, RRP12, TFRC, TMX2, TRAF7, and TTF2 (Fig. 4A). The prognostic curves of the 12 genes in the OS of BLCA were presented in Fig. S3A. Among the 24 hub PCDGs, only 21 PCDGs expressed in GSE13507 and GSE37815 (Table S4). Similarly, we performed univariate Cox regression analysis on the 24 PCDGs in GSE13507, identifying 10 significant genes affecting the OS of BLCA, including B4GALT3, CCNE1, CDT1, CENPO, HSP90AA1, P4HB, RPN1, TFRC, TP73, and TTF2 (Fig. 4B). The prognostic curves of the 10 genes in the OS of BLCA were presented in Fig. S3B. Also, in GSE37815, we identified 7 significant genes affecting the OS of BLCA, including ATP13A2, CCT6A, CDT1, CENPO, HSP90AA1, PTPN2, and RRP12 (Fig. 4C). The prognostic curves of the 10 genes in the OS of BLCA were presented in Fig. S3C.

Fig. 4.

Prognostic risk profile of hub PCDGs. (A) Univariate risk regression analysis was conducted on these 24 PCDGs and an association network was constructed (TCGA-BLCA). (B) Univariate risk regression analysis was conducted on these 21 PCDGs and an association network was constructed (GSE13507). (C) Univariate risk regression analysis was conducted on these 21 PCDGs and an association network was constructed (GSE37815)

Moreover, we elevated the molecular features and the mutations of these prognostic genes in TCGA-BLCA. Fig. S4A-C presented the mutation proportion and mutation types of prognostic genes within the TCGA-BLCA cohort, highlighting the genetic landscape associated with BLCA prognosis. The detailed mutational information for these prognostic genes was provided in Fig. S4D-G, offering insights into specific alterations that may influence patient outcomes. Then, we illustrated the expression trajectories of these genes across different stages of BLCA (Fig. S4H), while the expression distribution of the GSVA scores derived from these prognostic genes (Fig. S4I). Additionally, we explored the relationship between the GSVA scores and various survival characteristics, revealing their prognostic significance (Fig. S4J). Meanwhile, we examined the association between GSVA scores and tumor stage, further linking gene expression patterns with disease progression (Fig. S4K). Finally, we assessed the correlation between GSVA scores and key biological pathways, suggesting potential mechanisms through which these genes may impact BLCA progression and patient survival (Fig. S4L).

Multiple machine learning algorithms were integrated to screen BLCA signature genes

We allocated the TCGA-BLCA dataset to function as the training set and reserved the GSE13507 and GSE37815 datasets for external validation. As we mentioned above, the external validation sets GSE13507 and GSE37815 lacked the expression signature of IFI27L1, GGCT, and TMX2, so we excluded these three genes. Finally, 21 PCDGs were carried forward into the subsequent phase of the analysis. Using ten-fold cross-validation within the TCGA-BLCA training set, we constructed 101 predictive models and evaluated their performance by computing the C-index on both the training and validation sets. (Fig. 5A and Table S5). Based on the average C-index, the top four performing models were all constructed using the Random Survival Forest (RSF) algorithm. However, while these RSF-based models demonstrated strong performance within the internal training set, they generalized poorly to the GEO datasets, as evidenced by C-index values consistently below 0.6. Consequently, these overfitted models were excluded from further analysis. We then focused on the CoxBoost + GBM model, which exhibited good predictive ability across the training set (C-index = 0.76) and external validation (GSE13507: C-index = 0.63, GSE37815: C-index = 0.62). For performance comparison, CoxBoost + GBM was used as the baseline model. The other machine learning models were evaluated against this baseline using the NRI and the IDI. The results showed that only three models outperformed CoxBoost + GBM; however, their predictive performance was inconsistent across different datasets. Detailed NRI and IDI metrics are provided in Table S6. The signature included all 12 genes in the CoxBoost + GBM model: P4HB, CHEK2, PTPN2, ATP13A2, CCT6A, TFRC, RRP12, TRAF7, POLR1B, B4GALT3, SIVA1, and TP73, as depicted in Fig. 5B. Among them, P4HB showed the greatest predictive power (Fig. 5C-D). The cutoff value of -0.1308913 was applied to dichotomize TCGA-BLCA patients into two prognostic risk categories: high-risk and low-risk (Table S7). The heat map revealed a significant upregulation of P4HB, ATP13A2, TFRC, TRAF7, and RRP12 in the high-risk group samples, while the remaining genes exhibited down-regulated expression in the low-risk group samples (Fig. 5E). The KM-curve suggested a significantly poorer OS for high-risk BLCA patients compared to their low-risk counterparts in the TCGA cohort (Fig. 5F). The ROC curve showed excellent predictive accuracy of the risk score at 1, 3, and 5 years in the TCGA-BLCA cohort, yielding area under the curve (AUC) values of 0.796, 0.827, and 0.841, respectively (Fig. 5G). These results demonstrate the strong discriminatory power of the PCDRS.

Fig. 5.

Multiple machine learning algorithms were integrated to screen BLCA signature genes. (A) 10 different machine learning algorithms were incorporated to identify the genetic markers based on these 21 PCDGs. (B-D) 12 signature genes were identified by the CoxBoost algorithm and GBM algorithm. (E) The heat map revealed a significant upregulation of P4HB, ATP13A2, TFRC, TRAF7, and RRP12 in the high-risk group sample. (F) Patients with BLCA in the high-risk group had significantly lower OS compared to those in the low-risk group. (G) ROC curve showed excellent predictive accuracy for 1-, 3-, and 5-year interval risk scores in the TCGA-BLCA cohort

Construction and evaluation of prognostic model based on the PCDRS

We present the distribution of risk scores alongside the OS status for BLCA patients in the TCGA cohort in Fig. 6A-B. The results showed that the incidence of mortality exhibited an upward trend corresponding to higher risk scores (Fig. 6A) and the high-risk group demonstrated significantly poorer OS compared to the low-risk group (Fig. 6B). Then, significant differences were observed between the high- and low-risk groups across several key clinical parameters, including age, tumor grade, tumor stage, and Fustat (p < 0.05, Fig. 6C). Then, univariate Cox regression analysis identified age, tumor stage, and the PCDRS risk score as factors significantly associated with the prognosis of BLCA (Fig. 6D). Furthermore, age, tumor stage, and the PCDRS risk score were validated as independent prognostic determinants by multivariate Cox regression analysis (Fig. 6E). To facilitate the clinical translation of the PCDRS, we constructed a nomogram that integrates the PCDRS with clinical characteristics (Fig. 6F). The predictive accuracy of this nomogram was subsequently validated by calibration curves, which showed a high degree of concordance between its predicted probabilities and the observed survival outcomes (Fig. 6G). Furthermore, the C-index confirmed the nomogram’s stable and robust predictive performance, surpassing other clinical characteristics in forecasting OS from 1 to 10 years (Fig. 6H). The results indicate that the PCDRS-based nomogram serves as a dependable and precise tool for personalized prognosis prediction in BLCA patients.

Fig. 6.

Construction of prognostic model based on the 12 signature genes. (A) The incidence of mortality exhibited an upward trend corresponding to higher risk scores. (B) The high-risk group demonstrated significantly poorer overall survival compared to the low-risk group. (C) Correlation between the low- and high-risk groups and clinical characteristics. (D) Univariate Cox regression analysis demonstrated significant correlations between age, overall disease stage, and risk-Score with the prognosis of BLCA. (E) Multivariate Cox regression analysis confirmed that age, overall disease stage, and risk-Score independently serve as prognostic factors. (F) A prognostic predictive nomogram based on the PCDRS and clinical characteristics. (G) The calibration curve of the nomogram for 1, 3, and 5-year OS. (H) The C-index between the nomogram and other clinical characteristics

Consistent with the results in TCGA-BLCA, we also constructed risk scores using 12 signature genes to evaluate their prognostic value in the GSE13507 dataset (Table S8). Our findings indicated that higher risk scores were associated with an increased incidence of mortality (Fig. S5A), and OS was significantly inferior among patients assigned to the high-risk group relative to their low-risk counterparts. (Fig. S5B). A heat map analysis revealed significant upregulation of P4HB, ATP13A2, B4GALT3, TFRC, CCT6A, and RRP12 in the high-risk group, while the remaining genes were downregulated in the low-risk group (Fig. S5C). The KM curve showed that the high-risk group also demonstrated significantly lower OS (Fig. S5D). The ROC curve analysis demonstrated good predictive accuracy for 1-, 3-, and 5-year intervals, with AUC values of 0.657, 0.633, and 0.595, respectively (Fig. S5E). Univariate Cox regression identified age, tumor stage, tumor grade, and the risk score as variables significantly associated with BLCA prognosis (Fig. S5F). Subsequently, a subsequent multivariate Cox regression analysis confirmed that age, tumor stage, and the risk score each retained independent prognostic significance (Fig. S5G).

Single-Cell Transcriptomic Profiling of Hub Genes in BLCA

At the single-cell level, the “AddModuleScore” algorithm was utilized to assess the activity of 12 identified hub genes (hub_genes_Score) in 11 BLCA samples from 2 single-cell datasets (GSE129845: 3 N; GSE135337: 7T1N, Fig. 7A). The distribution analysis revealed higher median hub_genes_Scores in Cancer cells, Epithelial cells, and B cells compared to other cell types (Fig. 7B-C). Notably, tumor tissues exhibited significantly elevated hub_genes_Scores relative to normal tissues (p < 2.22e-16, Fig. 7D). Furthermore, Hallmark gene set enrichment analysis indicated the most significant enrichment of the ROS pathway (HALLMARK_REACTIVE_OXYGEN_SPECIES_PATHWAY) in Cancer cells with elevated hub_genes_Scores (Fig. 7E, Table S9). Next, the “AddModuleScore” algorithm was also utilized to evaluate the activity of ROS-related genes (ROS_genes_Score), showing a higher ROS score in the high hub_genes_Score group, indicating their positive correlations (Fig. 7F). The UMAP visualization demonstrated the expression distribution of these 12 hub genes across cell types, respectively (Fig. 7G). Additionally, the summary of all 12 PCDRS genes and their known functions or literature links to BLCA was displayed in Table S10. These findings highlight the distinct molecular characteristics of the hub PCDGs in BLCA, particularly their association with the ROS pathway, which may contribute to the differential behavior of cancer cells at the single-cell level.

Fig. 7.

Molecular characteristics of hub-genes in BLCA at single-cell sequencing level. (A) The AddModuleScore algorithm was employed to evaluate the activity of the 12 genes (hub_genes_Score) at the single-cell sequencing level in 11 BLCA samples from 2 single-cell datasets (GSE129845: 3 N; GSE135337: 1N7T). (B) Distribution of median hub_genes_Score. (C) Higher hub_genes_Scores were observed in Cancer cells, Epithelial cells and B cells compared to other cell populations. (D) Higher hub_genes_Score was observed in tumor tissues compared to normal tissues. (E) Hallmark gene set enrichment analysis in each cell type (high-hub_genes_Score vs. low-hub_genes_Score) revealed that ROS pathway were significantly enriched in Cancer cells with higher hub_genes_Score. (F) The AddModuleScore algorithm was employed to evaluate the activity of the ROS-related genes (ROS_genes_Score) at the single-cell sequencing level. (G) UMAP of the expression distribution of the 12 hub genes

Pseudotime trajectory analysis of tumor cells

To visualize scRNA-seq data, we applied the UMAP algorithm, identifying 14 distinct tumor cell clusters (Fig. 8A). The expression of P4HB across tumor cell clusters was shown in Fig. 8B. To explore the differentiation and developmental relationships among the 14 tumor cell subgroups, CytoTRACE2 was used to analyze their differentiation status (Fig. 8C). Notably, cluster4 and cluster2 exhibit the highest stemness, highlighting its primitive position within the differentiation hierarchy (Fig. 8D). Consistently, Monocle2 pseudotime analysis showed that the pseudo-ordering of tumor cells is organized into two main branches (Fig. 8E). Most cluster4 and cluster2 subgroup cells were located at the beginning of the pseudotime series (Fig. 8F). The pseudotime scatter plots displayed that P4HB expression is elevated at the terminal stages of differentiation along the pseudotime axis (Fig. 8G), offering valuable insights into the biological relevance of P4HB expression in BLCA. The Slingshot trajectory inference analysis revealed three distinct developmental lineages originating from the stem-like clusters (cluster4 and cluster2) (Fig. 8H). Examination of P4HB expression dynamics along each lineage showed lineage-specific patterns. P4HB expression progressively increased along Lineage 1 and Lineage 3, peaking at later pseudotime stages, while Lineage 2 exhibited a relatively moderate and stable expression pattern (Fig. 8I). These findings suggest that P4HB is differentially regulated during distinct differentiation paths, further supporting its role in tumor progression and heterogeneity.

Fig. 8.

Pseudotime trajectory analysis of P4HB in tumor cells using CytoTRACE2 and Monocle2. (A) Tumor cells were extracted and clustered into 14 distinct subgroups. (B) Expression of P4HB across the 14 tumor subgroups. (C) The differentiation status of tumor cells was visualized using CytoTRACE2, with colors representing the degree of differentiation. (D) The box plot shows the predicted ordering of tumor cell subpopulations by CytoTRACE2. (E-F) Pseudotime trajectory of tumor cells colored by pseudotime (E) and distribution of tumor cells (F). (G) Expression of P4HB across pseudotime. (H) Slingshot trajectory inference showing three distinct lineages originating from stem-like clusters. (I) Dynamic expression patterns of P4HB along each inferred lineage

Association of P4HB with Tumorigenesis and the ROS Pathway in BLCA

A violin plot illustrated the differential expression of 12 hub genes when comparing high- and low-risk score groups, underscoring the potential role of these genes in disease progression (Fig. 9A). Further supporting these findings, boxplots demonstrate that these 12 hub genes were all elevated in tumor tissues compared with and normal tissues in the TCGA-BLCA cohort (Fig. 9B), which is further validated through paired difference analysis within the same cohort, emphasizing the higher expression of these hub genes (Fig. 9C). Moreover, we found that P4HB was up-expression across various cancers including BLCA with a radar map (Fig. 9D). IHC staining data sourced from the HPA database further confirmed P4HB high expression in tumor tissues of BLCA patients (Fig. 9E). In addition, risk regression analyses underscored the significance of P4HB in predicting poor prognosis outcomes of BLCA across TCGA, GSE13507, and GSE154261 (p < 0.05, Fig. 9F), with significant increased expression of P4HB in the progression of BLCA across GSE13507 (p = 0.017), and GSE37815 (p = 0.041, Fig. 9G). Specifically, the expression of P4HB is notably elevated in advanced tumor stages of BLCA in GSE13507 (p = 0.0055), suggesting its involvement in the aggressiveness of BLCA (Fig. 9H). Notably, single-cell transcriptomic analysis reveals that P4HB expression is highly consistent with the distribution of ROS pathway scores in cancer cells, highlighting its potential role in oxidative stress regulation within the tumor microenvironment (Fig. 9I). These results have demonstrated a significant association between P4HB and tumorigenesis, particularly within the ROS pathway in BLCA.

Fig. 9.

Multiple analyses indicating a significant association between P4HB and the tumorigenesis as well as ROS pathway in BLCA. (A) Violin plot shows the differential expression of 12 hub genes between high- and low-risk score groups. (B) Boxplots show the overall expression levels of the 12 hub genes between cancer and normal tissues in the TCGA-BLCA cohort. (C) Paired difference analysis of 12 hub genes in the TCGA-BLCA cohort. (D) Radar map of P4HB expression levels in pan-cancer. (E) Immunohistochemical staining results from HPA database showed the expression of P4HB. (F) Risk regression analysis of P4HB in multiple bladder cancer cohorts (G) Differential expression analysis of P4HB in multiple progressive bladder cancer cohorts. (H) Differential expression analysis of P4HB in bladder cancer samples of different stages. (I) Expression density of P4HB at the single-cell level and cellular distribution of ROS pathway score

Additionally, we conducted an analysis of P4HB’s interacting partners using publicly available protein-protein interaction databases, GeneMANIA (Fig. S6A) and STRING (Fig. S6B), which revealed its interactions with proteins involved in Protein processing in endoplasmic reticulum, Thyroid hormone synthesis, and Arginine and proline metabolism and metabolic processes (Fig. S6C-D), thereby offering deeper insights into the functional mechanisms of P4HB in BLCA pathogenesis.

Validation of the expression and function of P4HB in BLCA

Consistent with results from public databases, our RT-qPCR analysis indicated that P4HB mRNA levels were increased in BLCA tissues from 25 clinical samples at our center (Fig. 10A). Additionally, in BLCA cell lines (J82, UMUC3, TCC, BIU87, T24, and RT4), P4HB protein and mRNA expressions were up-regulated compared to SVHUC-1 cells, as confirmed by RT-qPCR (Fig. 10B) and western blot (Fig. 10C), respectively. Additionally, we validated the oncogenic role of P4HB in BLCA. As depicted in Fig. 10D-E, successful transfection of P4HB shRNA into BIU87 and UMUC3 cells was confirmed through qRT-PCR and western blot analyses. Suppression of P4HB led to a significant inhibition in the proliferation and growth of BLCA cells, as evidenced by CCK-8 (Fig. 10F) and colony formation assays (Fig. 10G). The flow cytometry assay determined that blocking P4HB could promote BLCA cells apoptosis (Fig. 10H). These findings establish P4HB as a pivotal driver of BLCA tumorigenesis, and highlight its promise as a novel therapeutic target for combating this malignancy.

Fig. 10.

P4HB expression and function in BLCA. (A) RT-qPCR results of P4HB mRNA expression in BLCA tissues from 25 paired adjacent normal and tumor samples. (B-C) P4HB mRNA (B) and protein (C) expression in BLCA cell lines (mean ± SD, n = 3, ***p < 0.001). (D-E) The interference efficiency of P4HB shRNA in BIU87 (D) and UMUC3 (E) was evaluated by RT-qPCR and western blot (mean ± SD, n = 3). (F-H) CCK-8 (F), colony assays (G), and flow cytometry (H) determined that blocking P4HB inhibited BLCA cells’ proliferation and growth, but promoted cells apoptosis (mean ± SD, n = 3, ***p < 0.001)

Discussion

BLCA is a significant public health issue because of its high incidence and mortality rates, which have remained largely unaltered over time [13]. Due to tumor location and complexity, conventional surgical interventions and chemotherapy may not be suitable for some BLCA patients [14]. Some clinical studies have validated the BLCA biomarkers DHCR24 and RRM2, which are associated with apoptosis [15, 16]. Moreover, recent studies using bioinformatics methods have identified numerous prognostic indicators of BLCA, including ERCC2, EPB41L2, POLE2 and MMP14 [17–20]. However, to date, these biomarkers lack sufficient sensitivity when examined in isolation. In this study, we identified hub PCDGs associated with BLCA, and constructed a powerful PCDRS at the bulk transcriptome, and single-cell transcriptome levels with 101-combination machine learning algorithms, demonstrating a strong association between PCD and BLCA prognosis, as well as ROS status, which may improve the management of BLCA, as well as the care and prognosis for BLCA patients.

Previous studies have shown that the single form of PCD plays a crucial role in BLCA, including disulfidptosis, anoikis, autophagy, and ICD [21–24]. However, these earlier investigations primarily focused on isolated aspects of PCD without a comprehensive examination of PCD interconnected dynamics in the context of BLCA. To dada, a comprehensive analysis examining PCDGs in BLCA has not yet been reported. As a result, we comprehensively analyzed PCDGs from 14 PCD types combining single-cell sequencing data with bulk transcriptome data in this study. Subsequently, a novel computational framework, PCDRS, has been developed to identify stable and reliable prognostic markers for BLCA, surpassing the studies mentioned above in terms of predictive accuracy and clinical translational application. ROC curve analysis further validates the robustness of the PCDRS, with AUCs of 0.796, 0.827, and 0.841 for 1-, 3-, and 5-year risk scores in the TCGA-BLCA cohort, respectively. Previously established prognostic models based on 12 types of PCD with LASSO regression analyses only had AUC values of 0.751, 0.753, and 0.763 in the TCGA-BLCA cohort [25], indicating that our 14 types of PCD prognostic model exhibited superior advantages.

Meanwhile, various prognostic models for BLCA have been developed with only one or several machine learning algorithms [26–28]. However, given the inherent molecular complexity of BLCA, prognostic models based on a single biomarker or analytical criterion often lack the robustness required for accurate patient outcome prediction. Therefore, we developed 101 predictive models using a tenfold cross-validation framework, leading to the identification of the optimal CoxBoost + GBM model with the highest C-index. The CoxBoost algorithm implements a Cox proportional hazards model through gradient boosting trees, specifically designed for the analysis and prediction of time-to-event data [29]. GBM algorithm is a widely used and powerful machine learning model, but exists a heightened risk of overfitting [30]. To address these issues, we used CoxBoost primarily for dimensionality reduction and variable screening, which helps to reduce the risk of overfitting. As a result, this combined approach significantly enhanced accuracy and outperformed other models in our study. We also established a nomogram consisting of gender, age, stage, grade, and risk score, which achieved a significantly higher C-index than a previously established PCD prognostic model only using WGCAN analysis [31, 32]. Our multivariate Cox regression analyses demonstrated that clinical variables such as age, stage, and risk score are significant independent prognostic factors for patient survival. Notably, the risk score exhibited a strong association with prognosis, with higher scores correlating with poorer outcomes. Furthermore, the time-dependent analysis indicated that the risk score maintained a consistent impact on prognosis over the follow-up period. These results showcase the superior stability and precision of our model across both training and external validation datasets, affirming its considerable edge and promising applicability in real-world clinical scenarios.

The machine learning integration identified 12 prognostic PCDGs to construct the PCDRS, including ATP13A2, B4GALT3, CCT6A, CHEK2, P4HB, PTPN2, POLR1B, RRP12, SIVA1, TFRC, TRAF7, and TP73. Among them, CCT6A serves as a molecular chaperone that assists in ATP-dependent protein folding, involved in key signaling pathways, including the cell cycle, p53, and apoptosis [33]. Knockdown of CCT6A has been shown to induce apoptosis and affect cell cycle progression at the G0/G1 phases in osteosarcoma [33, 34]. PTPN2 is a member of the protein tyrosine phosphatase (PTP) family and functions in T cell anti-tumor immunity [35–37]. The knockdown of PTPN2 can induce cell cycle changes and apoptosis in pancreatic adenocarcinoma cells via JAK-STAT signaling pathway [38, 39]. The latest research has identified PTPN2 as a target that enhances the sensitivity of cancer cells to immune attack and T cell activation [40]. TFRC is a ferroptosis modulator that causes arrest of ferroptosis progression to facilitate tumor development by affecting cellular iron uptake [41, 42]. Research indicates that TFRC is an effective antigen for the development of anti-BLCA mRNA vaccines [43]. TRAF7 acts as a signal transducer for the TNF receptor superfamily and suppresses apoptosis, and promotes tumorigenesis by targeting P53 for ubiquitin-mediated proteasomal degradation in hepatocellular carcinoma [44]. It was reported that knockdown of POLR1B (RNA Polymerase I Subunit B) could induce massive p53-dependent apoptosis and suppress lung cancer cell proliferation [45, 46]. Existing research indicates that B4GALT3 is associated with clinical prognosis in BLCA by suppressing cancer immunity through synthesizing the glycan structure of molecules on the CD8 T cell surface [47]. Besides, our study indicated a correlation between BLCA and the PCDGs (CHEK2, ATP13A2, RRP12, and TP73), which has not been accessed in previous studies.

Among these 12 hub PCDGs, P4HB hold the highest AUC value, suggesting its potential importance in BLCA, which requires further functional validation. The P4HB gene is part of the protein disulfide isomerase (PDI) family. It encodes the beta subunit of prolyl 4-hydroxylase and promotes thiol-disulfide exchange and disulfide bond formation [48]. In various cancers, downregulation of P4HB can induce apoptosis, partly by increasing ROS levels and inactivating the STAT3 pathway (e.g., in colon and prostate cancers) [30, 49]. In addition, P4HB plays an important role by modulating chemoresistance in liver cancer, lung cancer, malignant glioma, and BLCA [50–53]. Currently, there are no studies on the use of P4HB inhibitors for cancer therapy. However, it is well established that P4HB inhibitors like PACMA31 can directly induce ferroptosis, or indirectly promote apoptosis via PDI activity inhibition [54]. This suggests promising potential for P4HB inhibitors as a therapeutic approach in BLCA in the future.

Consistently, our single-cell transcriptomic analysis revealed that P4HB expression closely aligns with the distribution of ROS pathway scores in cancer cells, highlighting its potential role in regulating oxidative stress within the tumor microenvironment. Similarly, Wang et al. reported that inhibition of P4HB could increase ROS contents and regulate the cell cycle in BLCA [51]. Meanwhile, several studies have reported the promoting role of P4HB in BLCA, but they only explored the role of P4HB in T24 and 5637 BLCA cell lines [55–57]. Our study extends these findings by confirming its anti-apoptotic effects in additional cell models, specifically UMUC3 and BIU87. Taken together, these studies further suggest that targeting P4HB could modulate PCD in BLCA, offering significant therapeutic potential target for BLCA.

In summary, our prognostic model based on 14 forms of PCD outperforms those relying on individual PCD types such as apoptosis, autophagy, or ferroptosis. Moreover, the 101-combination machine learning algorithm shows superior performance compared to single algorithms. In addition, our study identifies P4HB as a potentially valuable biomarker for BLCA with in vitro knockdown experiments, paving the way for future research into prognostic biomarkers and personalized treatment strategies. Nevertheless, there are still some limitations that must be acknowledged. Firstly, due to the still incomplete understanding of PCD, additional PCDGs beyond the 2423 included in this study may be discovered in the future, as well as new PCD panels. Secondly, the accuracy of the PCDRS was tested using only retrospective external datasets from public databases, and validation through prospective clinical trials is necessary to confirm the current conclusions. Thirdly, this study’s complexity may limit clarity and claim validation, requiring a more focused scope and stronger mechanistic support. Also, P4HB exhibited the highest predictive importance in our bioinformatic model, suggesting it may serve as a potential biomarker. Preliminary in vitro knockdown experiments further support its role in regulating tumor cell proliferation and apoptosis. However, whether P4HB directly modulates PCD or ROS-related pathways requires further mechanistic investigation. Fourthly, there is no unified or standardized gene set panel for PCD subtypes, and the manually curated PCD gene sets may introduce inherent limitations, including variable gene set sizes, potential selection bias, and gene overlap across functionally related pathways.

Conclusion

In conclusion, this study provides a novel and comprehensive PCDRS for BLCA prognosis, developed through an innovative multi-algorithm machine learning framework. Additionally, our findings suggest a possible involvement of P4HB in BLCA progression. Nonetheless, additional research is required to address the outlined limitations, thereby enhancing the robustness and applicability of our results.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

CY and HYH generated the idea, and designed the study. LC and HYB conducted the experiments and analysis. LZY double checked the analysis. WTT and SQY interpreted the results with the help of HYB. CY drafted the paper, HYH and LZY critically revised the paper. All authors reviewed and approved the final version.

Funding

No funding.

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This study was approved by the ethics committee of the First Affiliated Hospital of Soochow University.

Consent for publication

Not applicable.