Immune-related genes can accurately predict survival in bladder cancer: a retrospective study via two independent immunotherapy cohorts

Introduction

One of the most widespread urogenital tract tumors in the globe due to its aggressive nature, high recurrence rate, and rapidly rising incidence and prevalence is bladder cancer (BLCA) (1). Around 81,180 individuals were diagnosed with BLCA in 2022, and the disease was responsible for 17,100 fatalities in the United States (2). An epidemiological study demonstrates a 4:1 male-to-female ratio in both BLCA incidence and mortality rates, establishing male gender as a key demographic risk factor for this malignancy (3). BLCA classification is complex, according to whether the tumor invades the muscular layer of the bladder. BLCA is categorized into two types, that is, non-muscle invasive bladder cancer (NMIBC) and muscle invasive bladder cancer (MIBC), with about 75% NMIBC and about 25% MIBC in all newly diagnosed BLCA cases (4,5). Just 50% of patients with MIBC survive for five years after undergoing radical cystectomy and pelvic lymphadenectomy (6). About half of patients with MIBC treated with radical cystectomy develop recurrence or metastasis (7). BLCA continues to exhibit poor prognosis, with persistently high mortality and recurrence rates despite recent advancements in combined modality therapies including surgery and chemoradiotherapy (8). The fact that BLCA is highly diverse and varies widely from patient to patient, resulting in low responsiveness and multidrug resistance, is a significant aspect in this phenomenon. As a result, the anti-tumor effect is unsatisfactory.

Recent years have witnessed remarkable advances in systemic therapies for breast cancer, with particularly transformative progress in metastatic triple-negative breast cancer (TNBC) management. These breakthroughs have been propelled by the clinical introduction of several innovative therapeutic classes: immune checkpoint inhibitors (ICIs), poly(ADP-ribose) polymerase (PARP) inhibitors, and antibody-drug conjugates (ADCs) (9). Notably, these developments have fundamentally reshaped treatment paradigms for early-stage TNBC, where ICIs are increasingly incorporated into neoadjuvant strategies with demonstrated survival benefits (10). Parallel to these advances in breast oncology, immunotherapy has emerged as a first-line therapeutic option for BLCA, rapidly displacing conventional treatment regimens. The accelerated development of ICIs has been instrumental in this transition, highlighting the considerable potential of immunotherapy in BLCA management. Immunotherapy and chemotherapy for BLCA patients resulted in median total survivals of 10.3 months and 7.4 months, respectively (11). Even though immunotherapy is more effective than traditional platinum-based chemotherapy, it is anticipated that few individuals with solid tumors will benefit from it. Numerous studies suggest that the pathology of BLCA is influenced by genetics, the immune microenvironment, and a variety of confounding factors, including lifestyle, chemical exposure, and race. This results in the unpredictable overall effectiveness and prognosis of immunotherapy (12). Therefore, it is crucial to establish a precise predictive signature for BLCA patients in order to predict the effectiveness and prognosis of immunotherapy.

Even though deep learning is a relatively new branch of machine learning, it inherits widely used parts of the machine learning knowledge bases, such as basic statistics and probability, loss/cost functions, etc. Deep learning also has the capacity to build more intricate layers with greater predictive power (13). An autoencoder is a deep learning framework for unsupervised learning that uses a sequence of nonlinear transformations to recover its original input. Information about the input layer may be represented by hidden layers. Autoencoders are thought to match complex nonlinear relationships effectively when compared to other dimensionality reduction techniques (14). Meanwhile, genetic indicators at the messenger RNA (mRNA) level have enormous promise for forecasting patient outcomes (15). A strong BLCA risk assessment model based on transcriptome data has been constructed in this study, which has not been shown in earlier predictive models, as deep learning has not yet been employed to predict the BLCA immunotherapy’s outlook and effectiveness. In conclusion, this study developed a deep learning-based risk scoring model to systematically evaluate immunotherapy efficacy and prognostic outcomes in BLCA patients. We present this article in accordance with the TRIPOD reporting checklist (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-28/rc).

Methods

Data availability and analysis

Our analysis included two independent immunotherapeutic cohorts: the clinical data and gene expression profiles of the GSE176307 cohort were obtained from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo); the “IMvigor210CoreBiologies” R package’s original transcriptome and clinical data (https://www.r-project.org/) were collected for the IMvigor210 cohort, which consists specific immunotherapy data for BLCA. These two datasets included a total of 437 BLCA patients. After removing samples with insufficient data on the effectiveness and survival of immunotherapy, 298 samples were included in the IMvigor210 cohort as test dataset and 87 samples were included in the GSE176307 cohort as the validation dataset. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Sparse autoencoder for feature extraction

An autoencoder represents a category of artificial neural networks that operate in an unsupervised manner. This framework consists of two essential phases: encoding and decoding. In the encoding phase, the input information is converted into a compact representation, which can subsequently be reconstructed to its initial form in the decoding phase. Throughout both encoding and decoding processes, certain noise from the original data may be eliminated. This benefit allows the autoencoder to function as a tool for feature extraction. By incorporating a corruption operation and a sparsity constraint into the conventional autoencoder, the sparse autoencoder can derive more robust and beneficial features (16). This challenging learning procedure enables the autoencoder to acquire additional insights regarding the input data. A sparsity constraint is applied to the hidden layers, which restricts the neurons from being active for the majority of the time. Additionally, this aids the autoencoder in identifying more intriguing patterns within the data. In order to acquire immunotherapeutic-related genes associated with survival prognosis from transcriptomic data, we employed sparse autoencoder to realize the feature extraction (17). It consists a total of 1,000 and 100 nodes, respectively, were present in the encoding and decoding layers. 100 output nodes of the decoding layer were taken as new features, and then performed univariate Cox regression and Wilcox analysis for each feature. Finally, we acquired 4 prominent features.

Gene set enrichment analysis (GSEA) enrichment analysis

GSEA is a computational method that evaluates predefined gene sets for statistically significant enrichment patterns between distinct biological states. By analyzing the distribution trend of a preset gene set in the gene table ranked by phenotypic correlation, GSEA was applied to determine the contribution of genes to the phenotype. We performed GSEA analysis on the transcriptomic data of the above four prominent features, aiming to explore the biological pathways and immune activities (18). Following 1,000 substitutions, enriched gene sets with P<0.05 were gathered (19).

Consensus clustering

Auto-encoders are capable of dimensionality reduction (20). To assess the immune responses of various BLCA patients’ heterogeneity further, we conducted unsupervised cluster analysis of the four important features associated with survival and immunotherapy efficacy in the training set using the “ConsensusClusterPlus” R package (21), estimated the ideal cluster size using the cumulative distribution function (CDF), and selected stable clustering results to divide the samples into different clusters. Using Kaplan-Meier (K-M) analysis, differences in patient survival between various cluster subtypes were further analyzed. Subsequently, we compared the differences in immunotherapy responsiveness between different subtypes. “Pheatmap” R package was carried out to demonstrate the disparity expression of genes that four significant features associated to survival and immunotherapy efficacy in different cluster subtypes as well as in different clinicopathological parameters.

Biological pathway characteristics and immune infiltration of the different subtypes

To investigate the variations in biochemical mechanisms and processes that underlie phenotypic features. For single-sample gene set enrichment analysis (ssGSEA), the “GSVA” R package (22) was performed, and gene sets from the Kyoto Encyclopedia of Genes and Genomes (KEGG) were taken out of the msigdf R program and analyzed for enrichment using the “clusterProfiler” R tool.

Analysis of functional enrichment & development of risk-scoring model

To identify the genes most closely related to the effectiveness of immunotherapy, using the “limma” package, we first examined the differentially expressed genes (DEGs) between various groups. Using P<0.01 as the threshold, a total of 1,976 DEGs were identified. Gene Ontology (GO) and KEGG analyses were implemented by the “clusterProfiler” R package. Next, univariate Cox analysis of DEGs was performed to calculate survival significance, using P<0.01 as the threshold to extract DEGs related with survival, resulting in 461 genes left. The genes were then processed further using the Lasso algorithm, yielding 14 genes. Finally, multivariate Cox analysis was used for model building and the genes with P>0.05 were removed, and the risk score model was built using the remaining three genes (CXCL10, SYNGR4, UCP2) and their weights (23). The formula used for this analysis was as follows: risk score = (Expibi). (Example: model gene expression level; bi: model gene coefficient). To verify the reliability and applicability of this immunotherapy response-related prognostic risk model. Based on the median risk score, the training and validation cohort was divided into high- and low-risk groups. Both the K-M survival curve and receiver operating characteristic (ROC) curve (24) were used to assess the risk model’s capacity to predict patients’ 1- and 2-year survival. Finally, this risk score model formula was utilized to calculate the risk score for each of the training set’s cluster groups, and these cluster samples were further intersected with the risk score. Then, we drew a Venn diagram to compare the variations in risk score among the various cluster groupings.

Assessment of the relation between risk score and immune cell infiltration

To demonstrate the variations in immune cell infiltration abundance between various risk score groups. Using the ssGSEA and “Xcell” scripts, it was assessed what percentage of immune cells were invading. The relative abundance of the 28 immune cells in BLCA was calculated using the ssGSEA algorithm of the “GSVA” software and compared the difference in immune cells distribution between the low- and high-risk groups (25). ImmuneScores and StromaScores were analyzed via the “Xcell” package. The Pearson’s test was utilized to evaluate the relationship between the immune checkpoint gene and immune enhancement genes (Immunostimulator) profiles and the risk score in the two groups.

Risk score distribution for various clinical characteristics

In order to make clear how clinical traits and risk model ratings relate to one another, The Cancer Genome Atlas subtypes (TCGA subtypes), immune cell level (IC level), tumor cell level (TC level), immune phenotype, Baseline Eastern Cooperative Oncology Group Score (Baseline ECOG Score), and sex were used to build risk scores analysis.

Sensitivity testing for drugs within different risk cohorts

To identify prospective chemotherapeutic medications that might be applied in BLCA therapy and the connection between the treatment sensitivity and the risk level, using the “oncoPredict” program, the half maximum inhibitory concentration (IC50) values for 199 medications were determined for tumor patients in the IMvigor210 cohort. Drug sensitivity and risk score were correlated using the Spearman correlation analysis, with a threshold of |Rs| >0.25 and a significant correlation at adjusted P<0.01. We acquired information on drug responsiveness and drug targeting pathways from the Genomics of Drug Sensitivity in Cancer (GDSC) project. Moreover, we compared the risk score, neoantigen burden per Mb, Baseline ECOG Score, immunophenoscore (IPS) score, Tumor Immune Dysfunction and Exclusion (TIDE) score, and PDCD1 and CD274 expression differences in different immunotherapy efficacy groups to further analyze the risk model’s effectiveness.

Establishing and testing a nomogram for risk scoring

Using the “rms” R package, clinical characteristics and risk score were used to develop the nomogram model. Using calibration curves, the training group assessed the precision of predicting OS for BLCA patients at one and two years, then evaluated in the validation group. Its clinical dependability has been shown by a calibration curve analysis. Using decision curve analysis (DCA), the model’s net benefit value was evaluated for a range of criteria.

Statistical analysis

All data analyses were carried out using the R program (version 4.2.2). There are typically many features available when referring to medical data. By using the autoencoder algorithm, we determined the best features that significantly influenced the outcome of BLCA immunotherapy. Each analysis’s statistical cutoff was set at P<0.05.

Results

Results of sparse autoencoder feature extraction

In this study, we first used a sparse autoencoder to extract 100 features from the transcriptome. Then, the significant features related to survival and immunotherapy efficacy were screened by univariate Cox regression and Wilcox analysis and the significant features subjected to unsupervised clustering analysis. According to the results of unsupervised clustering, the training samples were divided into two groups and the genes between different clusters and related to prognosis were analyzed by differentially expressed analysis and univariate Cox analysis. Next, Lasso algorithm was used to screen the features. Ultimately, the resulting genes were model constructed using multivariate Cox analysis (Figure 1A). The GSEA analysis of the four significant features revealed 10 common enriched pathways: ECM receptor interaction, spliceosome, pyrimidine metabolism, proteasome, oocyte meiosis, mismatch repair, homologous recombination, endocytosis, cell cycle and DNA replication (Figure 1B). Compared to other features, feature 1 and feature 41 preferred to be enriched in signal transduction-related pathways. Feature 1 was enriched in calcium signaling pathway, Hedgehog signaling pathway, MAPK signaling pathway, phosphatidylinositol signaling system, VEGF signaling pathway and WNT signaling pathway (Figure 1C). Feature 41 was enriched in ERBB signaling pathway, Hedgehog signaling pathway, MAPK signaling pathway, VEGF signaling pathway and WNT signaling pathway (Figure 1D). Feature 46 was mostly enriched in the carbohydrate metabolism-related pathways, including amino sugar and nucleotide, fructose and Mannose Metabolism, galactose Metabolism and citrate cycle TCA cycle (Figure 1E). The feature 83 was mostly enriched in the lipid metabolism as well as carbohydrate metabolism-related pathways, including biosynthesis of unsaturated fatty acids, butanoate metabolism, citrate cycle TCA cycle, pentose phosphate pathway, primary bile acid biosynthesis and steroid biosynthesis (Figure 1F).

Figure 1 Analysis flow and the enrichment analysis results. (A) Analysis flow of this study. (B) Results of the GSEA analysis of the four significant features are obtained. (C) Pathways enriched by feature 1. (D) Pathways enriched by feature 41. (E) Pathways enriched by feature 46. (F) Pathways enriched by feature 83. AUC, area under curve; ECOG, Eastern Cooperative Oncology; GSEA, gene set enrichment analysis.

Results of unsupervised consensus clustering

Unsupervised cluster analysis of four significant features associated to survival and immunotherapy efficacy using the “ConsensusClusterPlus” package was carried out. When k=2, it can be clustered into two perfect clusters (Figure 2A). Based on the findings of the cluster analysis, the data were divided into two groups: cluster 1 (152 samples) and cluster 2 (146 samples). The analysis of survival differences between the two clusters showed statistically different results, and the cluster 2 had better OS (Figure 2B). Subsequently, the immunotherapy response of the two clusters was analyzed. The histogram revealed that the immunotherapy response was weaker in cluster 1 than in cluster 2 (Figure 2C). Then we analyzed the expression of the four significant features in the two clusters. The expressions of feature 1 and feature 41 were higher in cluster 1 and feature 46 and feature 83 showed higher expression in cluster 2 (Figure 2D). Next, the enrichment analysis was performed for clusters. The cluster 1 was mainly enriched in drug metabolism cytochrome P450, metabolism of xenobiotics by cytochrome P450, drug metabolism other enzymes, starch and sucrose metabolism, etc. While, the cluster 2 was mainly enriched in base excision repair, nucleotide excision repair, cell cycle, homologous recombination, etc. (Figure 2E). Meanwhile, we also performed relevant immune analysis of infiltration. In cluster 1, memory B cell had the higher degree of infiltration and activated CD8 T cell, DSC and activated CD4 T cell had the higher degree of infiltration in cluster 2 (Figure 2F).

Figure 2 Unsupervised cluster analysis results and other correlation analysis among clusters. (A) Results of a ConsensusClusterPlus clustering of 2. (B) Prognostic survival analysis of immunotherapy between the two clusters. (C) The response to immunotherapy between the two clusters. (D) The heatmaps of clusters and 4 significant features. (E) Differences in the enriched pathways between the two clusters. (F) Differences in the immune infiltration between the two clusters. ECOG, Eastern Cooperative Oncology Group; IC level, immune cell level; MDSC, myeloid-derived suppressor cell; TC level, tumor cell level; TCGA, The Cancer Genome Atlas.

Functional annotation of the clusters

GO-biological processes (BP) results exhibited that the top three enrichment pathways were organelle fission, nuclear division and chromosome segregation; GO-cellular components (CC) results exhibited that the top three enrichment pathways were chromosomal region, spindle chromosome and centromeric region; GO-molecular function (MF) results exhibited that the top three enrichment pathways were tubulin binding, microtubule binding and catalytic activity acting on DNA. The top five enrichment pathways suggested by KEGG analysis were cell cycle, cellular senescence, human T-cell leukemia virus 1 infection, Fanconi anemia pathway and DNA replication (Figure 3A,3B). Patients with high-risk score exhibited significantly shorter survival times than those with low-risk score, according to the findings of the K-M analysis of prognostic scores from the IMvigor210 dataset (Figure 3C). The predictive properties of OS at 1 and 2 years were analyzed using ROC curves, and the corresponding area under curve (AUC) values were 0.7167 and 0.7272, respectively (Figure 3D). Similarly, we also performed the K-M analysis of the prognostic score based on GSE176307, which also suggested that patients with a high-risk score had worse survival rates than those with a low-risk score (Figure 3E). ROC curves analysis of the factors predicting OS at 1 and 2 years revealed AUC values of 0.6782 and 0.685 (Figure 3F). The two distinct clusters’ risk score significantly differ from one another, and cluster 1 had a risk score that is significantly greater than cluster 2 (Figure 3G). We compared different clusters and risk profiles and used Venn diagram to show common genes and unique genes. We can come to the realization that cluster 1 contained more high-risk genes, while cluster 2 included more low risk genes (Figure 3H).

Figure 3 Functional enrichment analysis. (A,B) Enrichment analysis of DEGs between the two clusters using GO and KEGG methods. (C,D) K-M analysis and AUC curves for prognostic scores from the IMvigor210 dataset. (E,F) K-M analysis and AUC curves for prognostic scores from the GSE176307 dataset. (G) Significant difference in risk score between two different clusters. (H) Venn diagram comparing common and unique genes under different clustering and risk characteristics. AUC, area under curve; BP, biological process; CC, cellular component; DEGs, differentially expressed genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; K-M, Kaplan-Meier; MF, molecular function.

Relationship between the risk model and immune cells

We compared the difference in immune infiltration between high and low risk score and indicated that cells such as CD56 bright natural killer cell, CD56 dim natural killer cell and type 17 T helper cell were more infiltrated in the high-risk group, while cells such as activated CD8 T cell, activated CD4 T cell and myeloid-derived suppressor cells (MDSC) were more infiltrated in the low-risk group (Figure 4A). We used the “Xcell” package to analyze the ImmuneScores and StromaScores, and the results showed that ImmuneScores of the low-risk group were considerably higher than those of the high-risk group (P<0.05, Figure 4B), but the there was no statistically significant difference in the StromaScores between the two risk groups (P>0.05, Figure 4C). The majority of the immune checkpoint genes were significantly expressed in the low-risk group, according to our comparison of immune checkpoint gene expression in the two groups (Figure 4D). Similarly, the Immunostimulator genes had the same results (Figure 4E).

Figure 4 The relationships between this model and the immune cells. (A) Differences in immune infiltration between high and low risk scores. (B,C) Differences in ImmuneScores and StromaScores between two risk groups. (D) Differential expression of immune checkpoint genes in high and low risk groups. (E) Differential expression of Immunostimulator gene in high and low risk groups. *, P<0.05; **, P<0.01; ***, P<0.001, ****, P<0.0001. MDSC, myeloid-derived suppressor cell.

Risk scores of the different clinical characteristics in patients with BLCA

To study the association between risk scores and clinical characteristics of BLCA patients, we investigated the distribution of risk score for TCGA subtype, IC level, TC Level, immunophenotype, Baseline ECOG Score and sex in the IMvigor210 cohort, as shown in Figure 5A-5F. We noted that TCGA subtype I had the highest risk score (P=1.3e−11, Figure 5A). The IC level was negatively correlated with the risk score (P=4.3e−10, Figure 5B). TC level 0 had the highest score (P=0.036, Figure 5C). Further analysis of risk score for the three-immune phenotype in the IMvigor210 cohort, including “immune inflamed”, “immune excluded” and “immune desert”, revealed that the risk score for the “immune inflamed” type was lower than the other two groups (P=9.5e−10, Figure 5D). In addition, our study showed that Baseline ECOG Score 1 had the highest risk score (P=0.016, Figure 5E) and there was no statistical difference between sex and risk score (P=0.64, Figure 5F).

Figure 5 Distribution of risk score among different clinical traits. (A-F) Risk score differences among TCGA subtypes, IC level, TC level, immunophenotype, Baseline ECOG Score and sex in the IMvigor210 cohort. ECOG, Eastern Cooperative Oncology Group; IC, immune cell; TC, tumor cell; TCGA, The Cancer Genome Atlas.

Potential therapeutic value of risk score and the predictable efficacy of immunotherapy risk score

To further understand the effect of risk score on drug response, the relationship between risk score and cancer cell line response to drugs was assessed. Carrying out Spearman correlation analysis, pairs of risk score significantly correlated with 35 drug sensitivities were identified in the GDSC database (Figure 6A). More research was done on the genes whose signaling pathways these medicines target. Among them, drugs related to PI3K/MTOR signaling, apoptosis regulation, protein stability and degradation, and WNT signaling pathways were positively correlated with risk score, and drugs of ERK MAPK signaling pathway were negatively correlated with risk score (Figure 6B). To analyze the relationship between immunotherapy and risk score, the immunotherapy efficacy was analyzed in terms of the percentage of high and low risk, and significantly better treatment outcomes were found in low-risk patients (Figure 6C). Further exploring the correlation between immunotherapy response and risk score, as demonstrated in Figure 6D, the immunotherapy responder group had a lower risk score. To assess the differences in neoantigen burden per Mb, Baseline ECOG Score, PDCD1 expression, CD274 expression, IPS score, we drew box plots. And the results indicated that TIDE score among different immunotherapy responses, neoantigen burden per Mb, Baseline ECOG Score, CD274 expression, and IPS score values were greater in the responder group than in the non-responder group (P<0.05, Figure 6E-6J).

Figure 6 The association between risk index, drug sensitivity, and immunotherapy effectiveness. (A) The association between risk score and drug sensitivity estimated by the Spearman analysis. (B) Signaling pathways targeted by drugs that are resistant or sensitive to risk score. (C) Proportion of patients responding to immunotherapy in high and low risk groups. (D) Risk score for different immunotherapy response groups. (E-J) Distribution of neoantigen burden per Mb, Baseline ECOG Score, PDCD1 expression, CD274 expression, IPS score, and TIDE score in different immunotherapy efficacy. ECOG, Eastern Cooperative Oncology Group; IPS, Immunophenoscore; TIDE, Tumor Immune Dysfunction and Exclusion.

Construction and evaluation of the nomogram model

To see if the risk score may function as an independent prognostic factor, multivariate Cox regression method was performed. Among them, risk score, neoantigen burden per Mb, and Baseline ECOG Score were found to be a robust and independent prognostic marker for assessing patients’ prognosis (Figure 7A). Therefore, we selected these three indicators to construct a nomogram model for BLCA patients (Figure 7B). Moreover, we assessed the calibration of the nomogram model and its impact on clinical decision making. The calibration curves indicated a high accuracy of the nomogram (Figure 7C). The DCA curve was utilized to explore whether the decisions of the nomogram model were clinically applicable, and we were able to successfully confirm the agreement of the probabilities of actual and predicted survival by DCA (Figure 7D,7E). We also further investigated the nomogram model for BLCA using K-M analysis, showing that patients with lower risk score had better OS (P<0.001, Figure 7F). Finally, we utilized ROC analysis to estimate the prognostic accuracy of the nomogram, risk score, neoantigen burden per Mb, and Baseline ECOG Score, and obtained AUC area corresponding values of 0.8002, 0.7167, 0.6591, and 0.6354, respectively (Figure 7G).

Figure 7 Construction of a prognostic nomogram for BLCA patients and assessment of the prognostic value of the nomogram model. (A) Multivariate Cox analysis of risk score and risk ratios for clinical traits. (B) Construction of nomogram model. (C) Calibration plot of the nomogram for probabilistic forecasts of 1- and 2-year overall survival of patients. (D,E) The DCA evaluated the accuracy of predicting 1-, 2-year prognosis. (F) K-M analysis with the nomogram model. (G) ROC curve of the nomogram model and individual factors to estimate the prognostic accuracy. **, P<0.01; ***, P<0.001. AUC, area under curve; BLCA, bladder cancer; CI, confidence interval; DCA, decision curve analysis; ECOG, Eastern Cooperative Oncology Group; HR, hazard ratio; IC, immune cell; K-M, Kaplan-Meier; ROC, receiver operating characteristic; TC, tumor cell; TCGA, The Cancer Genome Atlas.

Discussion

Approximately 30% of BLCA patients present with muscle-invasive disease, which carries a high risk of metastatic progression and cancer-related mortality, while the remaining 70% exhibit non-muscle-invasive tumors characterized by frequent recurrence but generally favorable survival outcomes (26-28). The poor outcomes that accompany these patients are typically correlated with deteriorated disease or relapse after radical cystectomy. First-line therapy for metastatic urothelial carcinoma (mUC) has historically been cisplatin-based combinations (28,29). However, almost all patients will eventually deteriorate and pass away, despite the cisplatin-based combinations’ initial response. Recent advances in biological and bioinformatic technologies have significantly enhanced our understanding of the molecular pathogenesis underlying BLCA initiation and progression.

Cancer therapy approaches that modify the immune status play a crucial part in oncology over the past few years. Typically, immunotherapy was used in conjunction with traditional cancer treatments such as surgery, radiation, and chemotherapy. Immunotherapies are considered as first-line therapy for many cancers (30). Recently, immune checkpoint blockers, such as T-cell transfer treatment, monoclonal antibodies, cancer vaccines, and immune system regulators are only a few examples of the various immunotherapy approaches utilized to treat cancer.

Innovative deep learning techniques based on unsupervised learning have been effectively used in visual computing applications. Moreover, deep learning approaches could be used as unsupervised feature learning approaches that speed up training and testing for supervised machine learning, enhancing performance. Here, 100 features were extracted from the transcriptome using sparse autoencoder, and then significant features associated with survival and immunotherapy efficacy were screened and unsupervised clustering analysis was performed on the significant features. Ultimately, we acquired four prominent features and two clusters. To predict accurately the effectiveness and prognosis of BLCA immunotherapy, we then created a prognostic risk signature by employing univariate Cox analysis, Lasso regression analysis and multivariate Cox analysis. A three-gene risk score signature (CXCL10, SYNGR4, UCP2) was then developed. To further comprehend the functional role of the four significant features, GSEA enrichment analyses were performed. K-M analysis indicated that cluster 2 fared better in terms of survival than cluster 1 did. Furthermore, to validate the prognostic benefit, we selected a different immunotherapy GEO dataset (GSE176307), and the outcomes agreed with the earlier analysis.

The tumor microenvironment (TME), characterized by dynamic interactions between malignant cells and surrounding stromal components through epigenetic reprogramming and oncogenic signaling, has emerged as a pivotal research focus in contemporary oncology (31). A small number of non-cancerous cells in the TME effect cancer cell growth, which in turn has a big impact on tumor spread and the effectiveness of pharmacological treatments (32). The TME comprises heterogeneous cellular components and their secreted factors that collectively influence tumor progression. They work together to create a persistently inflammatory, immunosuppressive, and pro-tumor environment, which is crucial for the growth and spread of tumors (33). In our research, we found that there existed differences between the two clusters in immune infiltration, memory B cells were more infiltrated in cluster 1. Activated CD8 T cells, MDSC and activated CD4 T cells were more infiltrated in cluster 2. Activated CD8 T cells and activated CD4 T cells could inhibit tumor development, it can be one of the factors contributing to cluster 2’s superior prognosis. Therefore, we could better understand the potential mechanism of the TME-related immune response and improve the effectiveness of current immunotherapy methods for BLCA patients by identifying the involvement of unique TME immune cell infiltration.

Further, we showed the potential drugs sensitivity of BLCA patients in IMvigor210 cohort to 199 drugs. Risk score was related to drugs resistance targeting ERK MAPK signaling and associated with sensitivity to drugs targeting PI3K/MTOR signaling pathway, apoptosis regulation, protein breakdown and stability, WNT signaling pathways. The clinical result of cancer patients has considerably improved because of immunotherapy. The observed heterogeneity in immunotherapy response rates likely stems from the current lack of robust biomarkers capable of reliably distinguishing responders from non-responders (34,35). Among them, neoantigen burden, which is typically stated as the number of tumors neoantigens per Mb of tumor genomic area, is a measure of the overall number of neoantigens in cancerous cells. Cancer patients with high neoantigen burden levels represent a higher number of tumor neoplastic antigens on the surface of their cancerous cells, and their immune cells can produce more effective killing effects on tumor cells, indicating that tumor patients with high neoantigen burden levels can have a better therapeutic response to ICI drugs (36). Before cancer treatment, doctors will evaluate the general health status of the patient, and the performance status (PS) is an important indicator to evaluate the general health status. ECOG is a standard that classifies the patient’s activity status into six levels from zero to five, each level corresponds to the corresponding physical status (37). It has been discovered that the proteins programmed cell death-1 (PDCD1/PD-1) and programmed cell death 1 ligand 1 (CD274/PD-L1) decrease anti-tumor T cell-mediated immune responses (38). An immune response score based on machine learning is called the IPS from The Cancer Immunome Atlas. A better immune response is related with a higher IPS score (39). It is a computational framework exploited to assess the likelihood of tumor immune evasion based on gene expression profiles of cancerous samples. The TIDE score serves as a computational biomarker for predicting ICIs treatment efficacy in individual tumor samples (40). Herein, we found the distribution of neoantigen burden per Mb, Baseline ECOG Score, PDCD1 expression, CD274 expression, IPS score, and TIDE score in different immunotherapy efficacy, and the findings exhibited that our model has better prognostic scores than the currently used immunotherapy.

Notably, our findings opened up new opportunities for enhancing the effectiveness of immunotherapy for BLCA by detecting various immune phenotypes of tumors and enabling tailored cancer immunotherapy.

Additionally, CXCL10 (chemokine interferon-γinducible protein 10 kDa) is classified functionally into a Th1-chemokine. It binds to the CXCR3 receptor and controls immunological responses by attracting and activating leukocytes like T cells, monocytes, and eosinophils (41). CXCL10 is highly expressed in various human disorders. It has been shown to contribute to the pathological emergence of three major human diseases, including inflammatory diseases, autoimmune diseases and infectious diseases and the cancer (42). SYNGR4’s (synaptogyrin-4) functions are unknown; it may also affect how synaptic-like microvesicles (SLMVs) behave in motor neurons (MNs), and its overexpression at the onset of symptoms may negatively affect motor neuronal function through this pathway (43,44). An ever-increasing number of studies emphasize the role of UCP2 (uncoupling protein 2) in a broad range of physiological and pathological processes. A previous study has proved that it is sufficient for cancer cells to overexpress the mitochondrial membrane transport protein UCP2 in order to shift the balance back toward oxidative phosphorylation and suppress malignant characteristics (45).

Notably, our risk score model demonstrates substantial potential to bridge the gap between computational prediction and clinical application in BLCA immunotherapy. By synergizing deep learning algorithms with transcriptomic profiling, we have developed a predictive tool that not only forecasts survival outcomes but also effectively identifies patients most likely to benefit from immune checkpoint inhibition. Clinically, this model could guide therapeutic decision-making by stratifying high-risk patients for combinatorial treatment strategies while identifying low-risk candidates suitable for de-escalation approaches. However, several limitations merit careful consideration. First, the retrospective nature of our study and the moderate sample sizes in both training and validation cohorts may introduce selection bias. Multi-center prospective validation studies will be essential to confirm the model’s robustness. Such as through collaboration with the tertiary hospitals, we will validate the model in diverse ethnic populations and treatment settings. Second, while we identified CXCL10, SYNGR4, and UCP2 as key prognostic biomarkers, their mechanistic roles in BLCA immune evasion require further elucidation. Functional studies are needed to determine whether these genes directly interact with immune checkpoint pathways. Third, the current model’s reliance on transcriptomic data alone could be augmented by incorporating genomic alterations and proteomic profiles to enhance predictive accuracy. Finally, clinical implementation faces practical challenges including standardization of RNA-seq protocols and development of cost-effective biomarker testing platforms. Looking forward, emerging technologies promise to refine our approach. The advent of single-cell sequencing and spatial transcriptomics will enable unprecedented resolution of tumor-immune interactions, potentially allowing spatial refinement of our risk score. Furthermore, next-generation artificial intelligence platforms integrating multi-omics data may eventually surpass conventional gene-expression-based models. Our current work establishes an important foundation for these advanced approaches by demonstrating the clinical feasibility of deep learning in BLCA prognostication. Addressing these priorities will be crucial for translating our findings into meaningful clinical applications.

Conclusions

In conclusion, we have developed and validated a risk score signature that robustly predicts the immunotherapy efficacy, drug sensitivity, and prognostic outcomes in BLCA. This model represents an independent prognostic factor for overall survival and may inform clinical decision-making in immunotherapy administration.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-28/rc

Peer Review File: Available at https://tau.amegroups.com/article/view/10.21037/tau-2025-28/prf

Funding: This work was supported in part by the Natural Science Research Key Project of Anhui Province (No. KJ2021A1156).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tau.amegroups.com/article/view/10.21037/tau-2025-28/coif). The authors have no conflicts of interest to declare.

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

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

References

1. Zhang Y, Lin Y, Lv D, et al. Identification and validation of a novel signature for prediction the prognosis and immunotherapy benefit in bladder cancer. PeerJ 2022;10:e12843. [Crossref] [PubMed]
2. Siegel RL, Miller KD, Fuchs HE, et al. Cancer statistics, 2022. CA Cancer J Clin 2022;72:7-33. [Crossref] [PubMed]
3. Zhang Z, Yu Y, Li P, et al. Identification and validation of an immune signature associated with EMT and metabolic reprogramming for predicting prognosis and drug response in bladder cancer. Front Immunol 2022;13:954616. [Crossref] [PubMed]
4. Wullweber A, Strick R, Lange F, et al. Bladder Tumor Subtype Commitment Occurs in Carcinoma In Situ Driven by Key Signaling Pathways Including ECM Remodeling. Cancer Res 2021;81:1552-66. [Crossref] [PubMed]
5. Zhang Q, Tan Y, Zhang J, et al. Pyroptosis-Related Signature Predicts Prognosis and Immunotherapy Efficacy in Muscle-Invasive Bladder Cancer. Front Immunol 2022;13:782982. [Crossref] [PubMed]
6. Patel VG, Oh WK, Galsky MD. Treatment of muscle-invasive and advanced bladder cancer in 2020. CA Cancer J Clin 2020;70:404-23. [Crossref] [PubMed]
7. Zhao H, Chen Z, Fang Y, et al. Prediction of Prognosis and Recurrence of Bladder Cancer by ECM-Related Genes. J Immunol Res 2022;2022:1793005. [Crossref] [PubMed]
8. Dyrskjøt L, Hansel DE, Efstathiou JA, et al. Bladder cancer. Nat Rev Dis Primers 2023;9:58. [Crossref] [PubMed]
9. Schipilliti FM, Drittone D, Mazzuca F, et al. Datopotamab deruxtecan: A novel antibody drug conjugate for triple-negative breast cancer. Heliyon 2024;10:e28385. [Crossref] [PubMed]
10. Rizzo A, Cusmai A, Acquafredda S, et al. KEYNOTE-522, IMpassion031 and GeparNUEVO: changing the paradigm of neoadjuvant immune checkpoint inhibitors in early triple-negative breast cancer. Future Oncol 2022;18:2301-9. [Crossref] [PubMed]
11. Bellmunt J, de Wit R, Vaughn DJ, et al. Pembrolizumab as Second-Line Therapy for Advanced Urothelial Carcinoma. N Engl J Med 2017;376:1015-26. [Crossref] [PubMed]
12. Deshpande RP, Sharma S, Watabe K. The Confounders of Cancer Immunotherapy: Roles of Lifestyle, Metabolic Disorders and Sociological Factors. Cancers (Basel) 2020;12:2983. [Crossref] [PubMed]
13. Zhang L, Zhou L, Wang Y, et al. Deep learning-based transcriptome model predicts survival of T-cell acute lymphoblastic leukemia. Front Oncol 2022;12:1057153. [Crossref] [PubMed]
14. Chen S, Zang Y, Xu B, et al. An Unsupervised Deep Learning-Based Model Using Multiomics Data to Predict Prognosis of Patients with Stomach Adenocarcinoma. Comput Math Methods Med 2022;2022:5844846. [Crossref] [PubMed]
15. Kamoun A, de Reyniès A, Allory Y, et al. A Consensus Molecular Classification of Muscle-invasive Bladder Cancer. Eur Urol 2020;77:420-33. [Crossref] [PubMed]
16. Franco EF, Rana P, Cruz A, et al. Performance Comparison of Deep Learning Autoencoders for Cancer Subtype Detection Using Multi-Omics Data. Cancers (Basel) 2021;13:2013. [Crossref] [PubMed]
17. Gao W, Shen J, Lin Y, et al. Sequential sparse autoencoder for dynamic heading representation in ventral intraparietal area. Comput Biol Med 2023;163:107114. [Crossref] [PubMed]
18. Dezem FS, Marção M, Ben-Cheikh B, et al. A machine learning one-class logistic regression model to predict stemness for single cell transcriptomics and spatial omics. BMC Genomics 2023;24:717. [Crossref] [PubMed]
19. Chen S, Gu J, Zhang Q, et al. Development of Biomarker Signatures Associated with Anoikis to Predict Prognosis in Endometrial Carcinoma Patients. J Oncol 2021;2021:3375297. [Crossref] [PubMed]
20. van de Leur RR, Hassink RJ, van Es R. Variational auto-encoders improve explainability over currently employed heatmap methods for deep learning-based interpretation of the electrocardiogram. Eur Heart J Digit Health 2022;3:502-4. [Crossref] [PubMed]
21. Hu X, Ni S, Zhao K, et al. Bioinformatics-Led Discovery of Osteoarthritis Biomarkers and Inflammatory Infiltrates. Front Immunol 2022;13:871008. [Crossref] [PubMed]
22. Hänzelmann S, Castelo R, Guinney J. GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics 2013;14:7. [Crossref] [PubMed]
23. Tay JK, Narasimhan B, Hastie T. Elastic Net Regularization Paths for All Generalized Linear Models. J Stat Softw 2023;106:1. [Crossref] [PubMed]
24. Zheng H, Zhang G, Zhang L, et al. Comprehensive Review of Web Servers and Bioinformatics Tools for Cancer Prognosis Analysis. Front Oncol 2020;10:68. [Crossref] [PubMed]
25. Tan J, Wang HL, Yang J, et al. JMJD2B-induced amino acid alterations enhance the survival of colorectal cancer cells under glucose-deprivation via autophagy. Theranostics 2020;10:5763-77. [Crossref] [PubMed]
26. van Ginkel N, Hermans TJN, Meijer D, et al. Survival outcomes of patients with muscle-invasive bladder cancer according to pathological response at radical cystectomy with or without neo-adjuvant chemotherapy: a case-control matching study. Int Urol Nephrol 2022;54:3145-52. [Crossref] [PubMed]
27. Cronise KE, Das S, Hernandez BG, et al. Characterizing the molecular and immune landscape of canine bladder cancer. Vet Comp Oncol 2022;20:69-81. [Crossref] [PubMed]
28. Lavoie JM, Sridhar SS, Ong M, et al. The Rapidly Evolving Landscape of First-Line Targeted Therapy in Metastatic Urothelial Cancer: A Systematic Review. Oncologist 2021;26:e1381-94. [Crossref] [PubMed]
29. Hanna KS, Larson S, Nguyen J, et al. The role of enfortumab vedotin and sacituzumab govitecan in treatment of advanced bladder cancer. Am J Health Syst Pharm 2022;79:629-35. [Crossref] [PubMed]
30. Jagodinsky JC, Harari PM, Morris ZS. The Promise of Combining Radiation Therapy With Immunotherapy. Int J Radiat Oncol Biol Phys 2020;108:6-16. [Crossref] [PubMed]
31. Yao H, Jiang X, Fu H, et al. Exploration of the Immune-Related Long Noncoding RNA Prognostic Signature and Inflammatory Microenvironment for Cervical Cancer. Front Pharmacol 2022;13:870221. [Crossref] [PubMed]
32. Xiao Y, Yu D. Tumor microenvironment as a therapeutic target in cancer. Pharmacol Ther 2021;221:107753. [Crossref] [PubMed]
33. Gaggero S, Witt K, Carlsten M, et al. Cytokines Orchestrating the Natural Killer-Myeloid Cell Crosstalk in the Tumor Microenvironment: Implications for Natural Killer Cell-Based Cancer Immunotherapy. Front Immunol 2020;11:621225. [Crossref] [PubMed]
34. Walter NS, Bhattacharyya S. Mining parasites for their potential as novel therapeutic agents against cancer. Med Oncol 2024;41:211. [Crossref] [PubMed]
35. Wu W, Pu Y, Shi J. Nanomedicine-enabled chemotherapy-based synergetic cancer treatments. J Nanobiotechnology 2022;20:4. [Crossref] [PubMed]
36. Wang P, Chen Y, Wang C. Beyond Tumor Mutation Burden: Tumor Neoantigen Burden as a Biomarker for Immunotherapy and Other Types of Therapy. Front Oncol 2021;11:672677. [Crossref] [PubMed]
37. Strang P. Palliative oncology and palliative care. Mol Oncol 2022;16:3399-409. [Crossref] [PubMed]
38. Wang S, Yuan B, Wang Y, et al. Clinicopathological and prognostic significance of PD-L1 expression in colorectal cancer: a meta-analysis. Int J Colorectal Dis 2021;36:117-30. [Crossref] [PubMed]
39. Zou R, Jiang Q, Jin T, et al. Pan-cancer analyses and molecular subtypes based on the cancer-associated fibroblast landscape and tumor microenvironment infiltration characterization reveal clinical outcome and immunotherapy response in epithelial ovarian cancer. Front Immunol 2022;13:956224. [Crossref] [PubMed]
40. Huang X, Liu Y, Qian C, et al. CHSY3 promotes proliferation and migration in gastric cancer and is associated with immune infiltration. J Transl Med 2023;21:474. [Crossref] [PubMed]
41. Li X, Lu M, Yuan M, et al. CXCL10-armed oncolytic adenovirus promotes tumor-infiltrating T-cell chemotaxis to enhance anti-PD-1 therapy. Oncoimmunology 2022;11:2118210. [Crossref] [PubMed]
42. Kong YF, Sha WL, Wu XB, et al. CXCL10/CXCR3 Signaling in the DRG Exacerbates Neuropathic Pain in Mice. Neurosci Bull 2021;37:339-52. [Crossref] [PubMed]
43. Marques RF, Engler JB, Küchler K, et al. Motor neuron translatome reveals deregulation of SYNGR4 and PLEKHB1 in mutant TDP-43 amyotrophic lateral sclerosis models. Hum Mol Genet 2020;29:2647-61. [Crossref] [PubMed]
44. Marques RF, Duncan KE. SYNGR4 and PLEKHB1 deregulation in motor neurons of amyotrophic lateral sclerosis models: potential contributions to pathobiology. Neural Regen Res 2022;17:266-70. [Crossref] [PubMed]
45. Luo JY, Cheng CK, He L, et al. Endothelial UCP2 Is a Mechanosensitive Suppressor of Atherosclerosis. Circ Res 2022;131:424-41. [Crossref] [PubMed]