Integrated Multiomics Analysis and Machine Learning Approaches in Bladder Cancer: Unveiling the Impact of Immunogenic Cell Death and Its Key Gene SLC2A3 on Prognosis and Personalized Treatment Strategies

Abstract

Recently, immunogenic cell death (ICD) has been identified as a regulatory cell death mechanism that induces an adaptive immune response, thereby improving enhancing the efficacy of immunotherapy and contributing to improved prognosis in bladder cancer (BLCA). This study established a risk signature based on ICD and identified ICD-related genes as diagnostic markers and therapeutic targets for BLCA. Thirty-two key ICD-risk genes (IRGs) were screened from correlation and univariate Cox regression analyses. Data obtained from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) databases regarding BLCA and normal samples were categorized based on IRGs. ICD-based prognostic signature was built based on IRGs, stratifying BLCA patients into high- and low-risk groups. SLC2A3 was identified as a hub gene by 101 combinatorial machine learning algorithms and 10 individual machine learning algorithms. Based on single-cell sequencing data, we identified the cancer-associated fibroblasts (CAFs), the key cell population exhibiting high SLC2A3 expression. Functional analyses were performed to explore the potential value of SLC2A3 as a target for precision therapy. A prognostic signature was constructed using IRGs, indicating significant differences in the tumor microenvironment (TME) and treatment response between different risk groups. We identified SLC2A3 as the most critical IRG exhibiting high expression in the fibroblast population of patients with BLCA, especially in CAFs, which play an important role in BLCA progression. We found that inhibiting SLC2A3 expression may enhance the effectiveness of immunotherapy and the identified potential drugs targeting SLC2A3. We demonstrated that the identified IRGs serve as risk factors for clinical prognosis in BLCA and successfully constructed an ICD-based prognostic signature. Additionally, SLC2A3 holds potential as a therapeutic target to advance precision and personalized treatment strategies for BLCA, in combination with immunotherapy.

Introduction

Bladder cancer (BLCA) is a leading cause of cancer-related mortality and one of the most common malignant neoplasms of the urinary system. In 2024, BLCA accounted for nearly 80,000 new cases and 17,000 deaths in the United States. Typically, patients with early stage BLCA are treated with intravesical immunoadjuvant therapy, which may involve chemotherapy. For advanced invasive cases, radical cystectomy remains the standard treatment. Owing to the recurrent nature of BLCA, adjuvant therapy can play a crucial role in improving patient prognosis. Cisplatin-based neoadjuvant chemotherapy before radical cystectomy has shown survival benefits in clinical trials; however, renal insufficiency renders approximately half of the patients with BLCA ineligible for this treatment. Recently, immunotherapy, particularly immune checkpoint inhibitors (ICIs), has shown significant survival advantages and reduced toxicity compared to chemotherapy. , Nonetheless, clinical studies have reported common immune-related adverse events (irAEs) may affecting severe neurological, hematological, and cardiac adverse events. − Clinical studies indicate a 30% recurrence rate at a median of 12 months postcystectomy, with an increased risk of more lethal metastases, highlighting the need for more precise biomarkers.

Immunogenic cell death (ICD) is a form of programmed cell death that elicits an immune response targeting antigens from dead cells, especially cancer cell-derived ones. The optimization of antigen supply to immune cells occurs through the activation of antigen-presenting cells (APCs), stimulation of CD8⁺ T cells, and induction of antitumor immune responses through ICD. The immunogenic properties of ICD are primarily mediated by damage-associated molecular patterns (DAMPs), which are released from dying cells in response to cellular stress or external stimuli. DAMPs are recognized by APCs, leading to the activation of immune responses, particularly cytotoxic T lymphocytes (CTLs) and tumor-specific immunity. This activation reverses the immunosuppressive TME and enhances tumor sensitivity to immunotherapy, providing new possibilities for developing personalized therapeutic strategies. − Although chemotherapies that induce ICD have improved tumor immunotherapy efficacy, the side effects such as renal toxicity highlight the need for precision therapy targets to maximize the benefits of ICD. ,

In recent years, RNA sequencing has become essential for understanding the origin and development of cancer. Conventional bulk RNA sequencing primarily analyzes the average gene expression levels of the entire cell population. However, it cannot distinguish gene expression differences between individual cells and may obscure subtle transcriptional differences or changes in gene expression over time. , Based on single-cell sequencing data, we can effectively reveal the heterogeneity of BLCA cells and identify key targets. , Comprehensive multiomics, including genomic, single-cell transcriptomic, and spatial transcriptomic approaches, have revealed in greater depth the molecular basis and potential mechanisms of signatures, defining tumor and immune cell states and revealing their interactions within specific disease contexts. − Machine learning has shown high reliability and stability in cancer diagnosis, prognosis, and treatment, improving the efficiency of tumor biomarker identification to locate key target genes. − Combining machine learning and multiomics analyses enables the evaluation of tumor development mechanisms and microenvironments, as well as the prediction of potential targets for precision therapy.

Materials and Methods

Data Collection and Processing

In our study, we obtained TPM-formatted data for 409 cancer samples and 19 normal samples of TCGA-BLCA from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov/). After removing duplicate genes and retaining the entry with the maximum expression value for each gene, only mRNA data were preserved for subsequent analysis.

A total of 2096 genes associated with immunogenic cell death (ICD) were retrieved from the Genecards database. After filtering for protein-coding genes and selecting those with correlation scores obtained through Spearman correlation analysis above the median, 919 high-confidence ICD-related genes were retained. Using the Wilcoxon rank-sum test in the TCGA-BLCA cohort, these 919 genes were analyzed for expression differences between tumor and normal samples. Then we identified 80 significantly upregulated and 111 downregulated ICD genes by applying thresholds of FDR <0.05 and |log2FoldChange| > 1. The 191 differentially expressed ICD genes (DEIGs) were analyzed using overall survival (OS) data from the TCGA-BLCA cohort through univariate Cox regression. With a P-value threshold of 0.01, 32 significant IRGs were identified eventually.

The 191 genes were analyzed in the STRING database with a minimum required interaction score of 0.7 (high confidence) to filter connected nodes and visualize the protein–protein interaction (PPI) network of DEIGs.

Analyses on ICD Subgroups

Unsupervised consensus clustering was performed using the ConsensusClusterPlus R package (v1.66.0) to classify the distinct molecular subtypes based on IRGs expression. The Kaplan–Meier (K–M) survival analysis was performed using the Survival (v3.5-7) and Survminer R packages (v0.4.9) to visually depict OS differences between the ICD-high and ICD-low groups. Differentially expressed genes (DEGs) between the two ICD subgroups were identified using the limma R package (v3.56.2), with criteria of |log2FC| > 1 and FDR < 0.05. 991 DEGs were identified, with 425 upregulated and 566 downregulated. Enrichment analyses for Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) were conducted utilizing the ClusterProfiler R package (v4.8.3). The p-values were adjusted using the Benjamini-Hochberg method. The top 10 significantly enriched pathways were displayed for each GO category: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC), while the top 20 pathways were presented for KEGG analysis. The ESTIMATE algorithm (v1.0.13) evaluated the immune and stromal scores for each BLCA patient. Additionally, the CIBERSORT algorithm (v0.1.0) was used to determine the proportion of immune cell subsets in each BLCA sample. The ssGSEA algorithm (GSVA R package, v1.48.3) was employed to assess human leukocyte antigen (HLA) gene levels, immune cell infiltration, and immune cell function in BLCA across various clusters.

Development and Validation of an ICD Prognostic Signature

Training and validation cohorts were created to construct a reliable ICD prognostic signature. The training cohort included BLCA samples from the TCGA data set, while the validation cohort consisted of BLCA samples from the GEO data set. The glmnet R (v4.1-8) package was utilized to conduct LASSO Cox regression analysis for assessing gene selection. We configured the LASSO regression parameters as family = “cox” and nfolds = 10 for 10-fold cross-validation, identifying 9 genes (ADAMTSL1, CALR, ELN, NES, OLR1, PDGFRA, SETBP1, SLC2A3, VHL). Subsequently, multivariate Cox regression was performed with the parameter direction = “both” (bidirectional stepwise selection combining forward inclusion and backward elimination) and default settings for other parameters. Finally, 6 genes (CALR, NES, OLR1, SETBP1, SLC2A3, VHL) were selected from the initial 9 to construct a prognostic risk score signature. A formula was used to compute the risk score: risk score = ∑(Expi × Coefi). BLCA patients were divided into high- and low-risk groups based on their median risk scores. The Kaplan–Meier analysis evaluated the survival outcomes of the two risk score groups. A risk curve was developed to evaluate prognostic predictions in BLCA patients. Cox regression analyses, both univariate and multivariate, were performed to evaluate the risk score as an independent prognostic factor.

The rms R package (v6.3-0) was utilized to develop a nomogram model incorporating clinicopathological features and risk score. Cox regression analysis was employed to identify factors and predict patient survival probabilities at 1, 2, and 3 year intervals. The nomogram’s precision was evaluated through a calibration graph and the consistency index (C-index). The C-index served as a measure of nomogram accuracy, indicating a positive correlation. The risk model’s predictive capability was validated using time-dependent ROC analysis via the timeROC R package (v0.4).

Immune Cell Infiltration

The relationship between immune cell infiltration and ICD-related signatures was examined using various algorithms, including XCELL, TIMER, QUANTISEQ, MCPCOUNTER, EPIC, CIBERSORT-ABS, and CIBERSORT, each represented in different colors. Correlation analyses across multiple algorithms were conducted using Spearman’s correlation method. The ssGSEA method was employed to calculate infiltration scores for 22 immune cell types, 13 immune function-related pathways, and 7 stromal-related pathways. Three TME-related scoresimmune, stromal, and estimatewere calculated using the IOBR package (v0.99.9). Immune checkpoint and HLA gene expression values were obtained from groups with varying ICD levels. Submap analysis was performed via the GenePattern platform (https://cloud.genepattern.org/gp) with Bonferroni correction to adjust p-values. Tumor immune dysfunction and exclusion (TIDE) analyses were performed through the Web site http://tide.dfci.harvard.edu/ with statistical significance assessed using Fisher’s exact test. Immunotherapy cohort data were obtained from the TIGER database (http://tiger.canceromics.org/#/). Data set GSE78220 contained mRNA expression data in 28 pretreatment melanomas treated with anti-PD-1 checkpoint inhibition, including 15 responded to treatment and 13 did not. Data set GSE91061 was obtained from 109 RNASeq samples (58 on-treatment and 51 pretreatment) from 65 patients with melanoma receiving anti-PD-1 checkpoint inhibition therapy, including 20 responded and 78 did not.

Identification of the Key Gene Based on Machine Learning

The TCGA-BLCA data set served as the training cohort, while the GSE13507 (256 BLCA samples) and GSE31684 (93 BLCA patients) data sets were assigned as independent validation cohorts. The study utilized 12 machine-learning algorithms: Efficient Neural Network (Enet), Random Forest (RF), NaiveBayes, Lasso, Support Vector Machine (SVM), Stepglm, glmBoost, Ridge, XGBoost, GBM, LDA, and plsRglm. Using a 10-fold cross-validation framework, we evaluated 101 combinations of 12 algorithms on the TCGA-BLCA training data set for variable selection and model construction. Despite the 101 ensemble machine learning combinations failing to identify the most critical genes, we supplemented the analysis with ten individual machine learning algorithms. Remarkably, SLC2A3 consistently ranked within the top three across all individual models, confirming its role as one of the most pivotal IRGs. Immunohistochemical and fluorescence analyses were conducted to examine SLC2A3 expression distribution. GESA was used to identify the enriched functions of SLC2A3.

Single-Cell Analysis

We collected 16 bladder cancer samples and 8 normal bladder samples from six single-cell sequencing data sets (GSE135337, GSE222315, GSE129845, GSE172433, GSE192575, and GSE225190) for single-cell analysis. Using the Seurat R package (v4.4.0), we integrated the samples and filtered scRNA-seq data to identify core cells. The harmony algorithm (v1.2.3) was employed to remove batch effects from single-cell RNA-seq data, excluding those with fewer than 300 genes, more than 5000 genes, over 10% mitochondrial genes, or more than 3% hemoglobin content. The data were normalized using the NormalizeData function and scaled via the ScaleData function in the Seurat package. Subsequently, the FindVariableFeatures function with the variance-stabilizing transformation (vst) algorithm was applied to identify the top 3000 highly variable genes. Single-cell samples underwent Principal component analysis (PCA), and the leading principal components were chosen for further analysis. The cells were clustered by the FindNeighbors and FindClusters functions in scRNA-seq data. DEGs of a specific cell type were identified using the “FindAllMarkers” function. The Uniform Manifold Approximation and Projection (UMAP) algorithm was then applied for dimensionality reduction and cell clustering. Marker genes for manual cluster annotation were identified using the singleR R package and the CellMarker database.

Pseudotime analysis using the “Monocle” algorithm (v2.28.2) was conducted to elucidate the molecular mechanisms underlying BLCA progression and to compare differentiation patterns between tumor and normal samples. The CellChat package (v2.1.2) was employed to examine cell communication patterns, identifying key senders, receivers, mediators, and influencers within intercellular communication networks. The UMAP analysis was used to identify subgroups of fibroblasts and then pseudotime analysis was performed on each groups using the “Monocle” algorithm, too.

Somatic large-scale chromosomal copy number variation (CNV) scores for individual cells were computed using the R package “inferCNV” (v1.16.0). Input filesincluding the raw counts matrix, cell annotation file, and gene-chromosome positional referencewere prepared in accordance with the package’s data preparation guidelines (https://github.com/broadinstitute/inferCNV). Cells derived from normal tissues were designated as the reference normal cell population. Analyses were executed with default parameters (cutoff = 0, denoise = 0.1).

Immunotherapy and Drug Sensitivity Analysis

We utilized the TIDE tool to predict immunotherapy responses in BLCA patients, aiming to further assess the tumor-immune microenvironment across various subgroups. TME scores and expression values of immune checkpoint-relevant genes were extracted in high-SLC2A3 and low-SLC2A3 groups. To ensure reliable immune score evaluation, we utilized the “immuneeconv” R software package.

Chemotherapeutic responses in BLCA patients were evaluated using the Cancer Immunome Database (https://tcia.at/home) and the OncoPredict R package (v1.2). The correlation between SLC2A3 expression and antitumor drug sensitivity scores was analyzed using point diagrams in the Ggplot2 R package. The CellMiner database (https://discover.nci.nih.gov/cellminer/home.do) was utilized to investigate the interaction between drugs and SLC2A3 by examining the model genes. The ggplot2 R package was utilized for visualizing the results.

Potential bioactive drug components and the three-dimensional structure of SLC2A3 were downloaded from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) and the Protein Data Bank (PDB, https://www.rcsb.org/), respectively. Structural preprocessingincluding the removal of water molecules and addition of hydrogen atoms to the receptor proteinwas performed using the AutoDock Tools package (The Scripps Research Institute Molecular Graphics Laboratory, Version 1.5.6, released September 17, 2014). Molecular docking between SLC2A3 and the small-molecule compounds was then executed with the Lamarckian genetic algorithm (LGA), retaining default parameters for all genetic algorithm settings. Finally, the docking results were visualized using PyMOL.

Molecular dynamics (MD) simulations of the complexes were performed for 100 ns using GROMACS 2022. The CHARMM36 force field parameters were applied to the protein, while ligand topologies were generated using the General AMBER Force Field 2 (GAFF2). Periodic boundary conditions (PBC) were implemented by embedding the protein–ligand complex in a cubic box. The system was solvated with TIP3P water molecules, maintaining a periodic buffer distance of 1.2 nm. Electrostatic interactions were calculated using the Particle Mesh Ewald (PME) method, and the Verlet cutoff scheme was employed for neighbor searching. Then 100,000 steps of isothermal isovolumetric ensemble equilibrium and isothermal isobaric ensemble equilibrium were simulated with a coupling constant of 0.1 ps and a duration of 100 ps. Both van der Waals and Coulomb interactions are calculated using 1.0 nm cutoff values. Finally, the system was simulated using Gromacs 2022 at constant temperature (310 K) and constant pressure (1 bar) for a total of 100 ns.

RT-qPCR Analysis

Total RNA from tumor and normal control tissues was extracted using TRIzol reagent (Invitrogen, USA). The sample size for each group was n = 5. First-strand cDNA was synthesized using a Superscript II First-Strand cDNA Synthesis Kit (TaKaRa, Japan). RT-qPCR was performed to quantify the expression levels of SLC2A3. GAPDH served as the internal control. The primer sequences are listed in Table .

1. Primers Used for RT-qPCR.

	forward: 5′ to 3′	reverse: 3′ to 5′
SLC2A3	CACGCTTGCGTATGTCTGTT	CTGCCTTACTGCCAACCTAC
GAPDH	AATGGGCAGCCGTTAGGAAA	GCGCCCAATACGACCAAATC

Statistical Analysis

Statistical analyses were performed using R software version 4.2.2. Chi-square tests were employed to assess differences in clinical characteristics between the training and internal validation cohorts. The Wilcoxon rank-sum test, a nonparametric method, was used to evaluate differences between two non-normally distributed variables. For DEG analysis, FDR-adjusted p-values were applied to determine significant DEGs. Kaplan–Meier survival analysis with log-rank tests compared OS across subgroups, while univariate and multivariate Cox regression analyses identified independent prognostic factors. Enrichment analysis p-values were adjusted using the Benjamini-Hochberg correction method. To evaluate model performance, time-dependent ROC curve analysis and AUC calculations were performed with the timeROC R package. Spearman’s correlation analysis explored associations between risk scores and immune cell infiltration. Student’s t tests were applied to analyze RT-qPCR results. Statistical significance was set at p < 0.05 unless otherwise specified.

Results

Transcriptome and Single-Cell Analyses Confirmed the Importance of ICD in BLCA Progression

We screened 2097 ICD-related genes from GeneCards and identified 191 DEIGs based on protein-coding genes and correlation scores greater than the median. Subsequently, we performed unifactorial Cox analysis of the DEIGs and identified 32 key IRGs (P < 0.01). Additionally, we analyzed the expression levels of these IRGs to explore their association with ICD and BLCA, revealing significant differences between the tumor and normal groups for most IRGs (Figure A). Kaplan–Meier survival analysis indicated that IRGs significantly impacted the OS with BLCA (Figure B), implying their abnormal expression in BLCA tissues and substantial effect on prognosis.

1.

Importance of ICD for BLCA. (A) Expression of IRGs in tumorous and normal tissues. (B) K–M curve results show the influence of each IRG on BLCA prognosis. (C) The heatmap shows statistically significant differences in gene expression between ICD-high and -low groups. (D) K–M survival analysis shows that IRGs were favorable prognostic factors for BLCA. (E) The UMAP algorithm was used to classify different cell types and shows the distribution of the types in normal and tumor tissues.

To advance our understanding of the specificity of ICD across various cell types, we analyzed single-cell RNA sequencing (scRNA-seq) data. During the quality control process, we excluded low-quality cells with high mitochondrial gene expression in both the tumor and normal groups. Subsequently, we used the screened core cells for further analyses (Supporting Information Figure S1). Using the UMAP algorithm, the core cells were classified into four distinct clusters: immune cells, fibroblasts, epithelial cells, and endothelial cells, with their distribution differing significantly between the tumor and normal groups. We categorized the groups into high- and low-IRG expression groups based on the IRG scores. Notable differences in gene expression levels were observed between the high- and low-IRG expression groups (Figure C). K–M survival analysis was conducted to evaluate the impact of varying IRG expression patterns on BLCA prognosis across the two clusters. The findings revealed that outcomes were significantly poorer in the high IRG cluster than in the low IRG cluster (Figure D). Compared to normal tissues, BLCA tumor tissues exhibited a decreased percentage of immune cells and an increased percentage of endothelial and epithelial cells with statistically differences (Figure E). We subjected all the 32 IRGs to UMAP projection analysis and observed high expression levels in BLCA tissues, especially in fibroblasts (Figure S2). These findings suggest the possible involvement of IRGs in the complex regulatory network underlying the pathophysiology of BLCA, as well as in its initiation and development. Nevertheless, further research on IRGs is crucial to elucidate the mechanisms underlying BLCA and to improve clinical outcomes.

We performed differential analysis to explore differences in IRG expression levels between the normal and tumor samples. We identified 191 DEIGs and categorized them into 80 upregulated and 111 downregulated genes (Figure A,B). GO and KEGG analyses revealed that DEIGs were mainly enriched in xenobiotic response, the external plasma membrane, and signal transduction functions, including the PI3K-Akt signaling pathway (Figure C). The PPI network identified the expression products of multiple DEIGs and selected hub genes for further enrichment analyses (Figure D). Furthermore, consensus clustering analysis of BLCA based on IRGs provided valuable insights into tumor heterogeneity, thereby improving the customization of treatments for different patient groups. The consensus heat map identified two clusters (c1 and c2) with well-defined boundaries (Figure E), and the results of consensus clustering analysis indicated a significant difference when k = 2, as evidenced by the curve with a gentle slope (Figure F–H). PCA revealed no distinct separation between the two clusters, which can be attributed to their shared tissue origin and underlying biological similarity (Figure I).

2.

Key genes were screened and BLCA samples were divided into two clusters. (A) Volcano plot of DEGs between BLCA and normal samples in TCGA. P < 0.05 and |log2FoldChange| > 1 were identified as significant DEIGs. Red dots represent upregulated genes and blue dots represent downregulated genes. (B) Heatmap of DEIGs. (C) Bubble plots of the BP, CC, and MF pathways of DEIGs. (D) PPI network shows the interaction between DEIG products. (E) Consensus heatmap defining the two clusters (k = 2). (F) Relative change in area under the CDF curve ranging k 2–9. (G) Consensus CDF in consistent clustering (k = 2–9). (H) Tracking plot of BLCA samples (k = 2–9). (I) PCA further supports classification of subgroups. (J) The heatmap shows large differences in gene expression between clusters c1 and c2. Rows correspond to genes, and columns represent samples. (K) K–M analysis of clusters c1 and c2. (L) GSEA identified top gene sets in clusters c1 and c2. (M) GO and KEGG results across TCGA-BLCA including BP, CC, and MF analyses, showing the top related pathways. (N) TMB in clusters c1 and c2 was predicted separately. TME-related scores between clusters c1 and c2. (O) Immune infiltration analysis of clusters c1 and c2. (P,Q) Different expression levels of HIRGs and HLA genes between clusters c1 and c2.

The heatmap revealed significant gene expression disparities between clusters c1 and c2 (Figure J). K–M analysis indicated that patients in cluster c1 exhibited a significantly improved prognosis compared to those in cluster c2 (Figure K). Functional enrichment analyses were performed to elucidate the molecular mechanisms distinguishing c1 and c2. The findings indicated that cluster c1 pathways were primarily linked to immune responses, including TLRs and STING signaling, while cluster c2 pathways were mainly related to ribosomal functions (Figure L). BP analysis identified the cytokine-mediated signaling pathway, axonogenesis, and leukocyte migration as the top three enriched functions. Moreover, CC analysis identified the external side of the plasma membrane and the collagen-containing extracellular matrix as the two most enriched functions. Molecular function analysis identified signaling receptor activator and receptor ligand activities as the most enriched functions. KEGG analysis identified cytokine–cytokine receptor interaction, neuroactive ligand–receptor interaction, and the PI3K-Akt signaling pathway as the top three enriched pathways (Figure M). Figure N shows the tumor mutational burdens (TMB) of clusters c1 and c2, revealing a higher TMB in cluster c1 than in c2. These findings indicate that cluster c1 may exhibit a more favorable response to immunotherapy. Using the IOBR package, we observed that cluster c1 exhibited greater immune cell infiltration than cluster c2, indicating a more favorable prognostic outcome (Figure O). Regarding humoral immunity-related genes (HIRGs) and HLA genes, most HIRGs were significantly differentially expressed between clusters c1 and c2, and all HLA genes showed significant differences (Figure P,Q). In summary, the ICD demonstrated substantial potential in predicting clinical outcomes for patients with BLCA.

Construction of an ICD-Based Prognostic Signature and Exploration of BLCA Immune Cell Infiltration

An accurate ICD-related prognostic signature was developed by applying the LASSO algorithm to refine the IRGs, resulting in the selection of six IRGs with optimal λ values (Figure A). A risk formula was developed using six characteristic genes. Patients were divided into low- and high-risk groups based on the median risk score. PCA showed that the two groups had clear edges (Figure B). A risk curve analysis revealed that patients classified as high-risk exhibited higher mortality rates and increased expression levels of SLC2A3, SETBP1, OLR1, NES, and CALR; and decreased expression level of VHL decreased (Figure C). To confirm the effectiveness of our prognostic signature, we plotted the receiver operating characteristic (ROC) curve for OS. Although the area under the curve (AUC) values for the three subsets at 1, 3, and 5 years did not exceed 0.7, our prognostic signature retained modest predictive efficacy (Figure D). Kaplan–Meier survival analysis revealed that patients with high-risk scores exhibited significantly reduced OS compared to those with low-risk scores (Figure E). Additionally, K–M survival analysis demonstrated a significantly improved prognosis for low-risk BLCA compared to high-risk BLCA across various clinical subgroups, including age <65 years, age ≥65 years, male, M0, N0, stage III–IV, T1-2, and T3-4 (Supporting Information Figure S3A–L). When clinical characteristics were compared, we observed a significantly better prognosis for age <65 years than age ≥65 years, M0 than M1, N0-1 than N2-3 and T1-2 than T3-4 (Supporting Information Figure S3M–R). Univariate Cox regression analysis assessed the influence of candidate genes on BLCA, while multivariate Cox regression analysis evaluated their independence as influencing factors alongside age, stage III, stage IV, and risk score (Supporting Information Figure S3S,T). A nomogram incorporating age, stage, and risk score was developed (Supporting Information Figure S4A). The 1, 2, and 3 year survival rates of each patient were predicted using the final nomogram scores consisting of the three items. Calibration curves demonstrated the nomogram’s high accuracy in predicting survival (Supporting Information Figure S4B–D). ROC analysis demonstrated that the nomograms exhibited a high AUC, indicating their superior predictive ability and enhanced quantification capability compared to a single nomogram (Supporting Information Figure S4E).

3.

Construction and validation of senescence-related signature and immune cell infiltration pattern in low- and high-risk BLCA. (A) Coefficients of LASSO analysis and ICD-related signature obtained six prognostic genes with a minimum lambda value. (B) PCA showing significant separation of low- and high-risk BLCA. (C) Survival curve, survival status, and heatmaps showing the expression of signature genes of patients with BLCA. (D) ROC analysis showing the stable prediction ability of the ICD-related signature. (E) K–M survival analysis showing a significant survival difference between low- and high-risk BLCA. (F) Box plot visualizing the expression levels of 13 immune function-related pathways between low- and high-risk groups. (G) Correlation analysis of risk score and diverse immune cells using the XCELL, TIMER, QUANTISEQ, MCPCOUNTER, EPIC, CIBERSORT-ABS, and CIBERSORT algorithms. (H) Violin plot showing infiltration of immune cells in different states. (I) Box plot showing ssGSEA scores of 23 immune cells between high- and low-risk groups. (J) Correlation analysis between different immune cells. (K) Different expression levels of HLA genes between high- and low-risk groups. (L) Different expression levels of HIRGs between high- and low-risk groups. (M) Prediction of the response to immunotherapy in high- and low-risk groups. (N) The TIDE database was used to predict response to immune checkpoint inhibitor therapy in high- and low-risk groups. (O) K–M analysis verified the reliability of the prognosis signature.

The box plan showed the link between BLCA and immune function-related pathways, which were significantly expressed in the high-risk group (Figure F). The seven immune algorithms revealed a negative correlation between the ICD-related signature and CD4 T cells, plasma B cells, central memory CD4 T cells, and activated myeloid dendritic cells, while showing a positive correlation with M1 and M2 macrophages, monocytes, and CAFs (Figure G). The CIBERSORT tool indicated increased infiltration of activated memory and naïve CD4 T cells, monocytes, M0, M1, and M2 macrophages, as well as both activated and resting dendritic cells in the high-risk group (Figure H). The ssGSEA analysis revealed statistically significant differences in the infiltration levels of 21 immune cell species, except for CD56 bright NK cells and Th17 cells, which showed no distinct variations (Figure I). Pearson correlation results showed a correlation between each cell type (Figure J). Differential expression levels of all HIRGs and HLA were observed in both low- and high-risk groups (Figure K,L). The high-risk group showed a greater likelihood of responding to PD-1 checkpoint immunotherapy (Figure M). A low TIDE score and low-risk score were associated with a high response to ICIs, with a significantly greater proportion of responders in the low-risk group than in the high-risk group (Figure N). The K–M survival analysis showed that the low-risk group had a better outcome (Figure O). In this section, we established a relatively stable prognostic signature for BLCA to detect tumor progression.

Combining Machine Learning and Experimental Methods to Screen the Key Gene SLC2A3

The low-risk group showed a better response to immunotherapy, highlighting the need for further exploration of novel therapeutic targets for high-risk patients. Consequently, we identified the most critical genes for this purpose. We applied 101 combinatory machine-learning techniques to the TCGA-BLCA data set, partitioning it into training and internal validation sets at a 7:3 ratio. Figure A presents the C-index calculations for each model. We screened the five IRGs with the highest association with BLCA: CALR, NES, SETBP1, SLC2A3, and VHL (Figure B). To extract the most critical genes, we combined six separate machine-learning methods and selected only the top three genes, leading to the identification of SLC2A3 (Figure C). In the GSE31189 data set, SLC2A3 expression was notably higher in the tumor group than in the normal group (Figure D). We analyzed SLC2A3 expression across different T stages of the TNM classification and found that its expression level increased with the advancement of the T stage (Figure E). We also examined variations in SLC2A3 expression across various clinical subgroups (Supporting Information Figure S5). RT-qPCR analysis revealed that the relative gene expression levels were elevated in the tumor group and progressively increased with the advancement of the T stage (Figure F). The expression levels of SLC2A3 in the tissue samples were examined using immunohistochemistry. As expected, SLC2A3 was highly expressed in BLCA samples (Figure G). In summary, these results confirmed that SLC2A3 is a key gene involved in BLCA progression.

4.

Various machine-learning algorithms and single-cell data were combined to screen the key gene SLC2A3. (A,B) Multiple machine-learning algorithms showing the most critical ICD-related genes in BLCA. (C) Ten machine-learning methods were combined to extract information regarding the key gene, SLC2A3. (D) Expression difference of SLC2A3 in tumor and normal cells. (E) SLC2A3 expression differences in T staging across different data sets. (F) RT-qPCR verifying the expression difference of SLC2A3 in tumorous and normal cells and in T staging. The sample size for each group was n = 5. (G) Expression of SLC2A3 in normal and BLCA tissues. Scale bar = 200 μm.

Functional Analysis of SLC2A3 Based on scRNA-seq

K–M survival analysis identified SLC2A3 as a prognostic risk factor for BLCA (Figure A). Our analyses consistently yielded the same conclusion regarding survival rates among patients with varying SLC2A3 expression levels across different clinical characteristics (Supporting Information Figure S6). High SLC2A3 expression served as a risk factor in patients with the following clinical characteristics: White race, presence of lymphovascular invasion, no prior radiotherapy, higher histologic grade, advanced tumor stage (Stage III/IV, T3/T4), nodal involvement (N1/N2/N3), and nonmetastatic status (M0). We performed several analyses on SLC2A3 to elucidate the specific mechanism through which it contributes to BLCA progression. Immunofluorescence analysis revealed the subcellular localization of SLC2A3 to the plasma membrane and aggresomes (Figure B–D). The single-gene coexpression heatmap indicated positive correlations between SLC2A3 expression and other genes, with the upregulation of additional genes associated with active SLC2A3 expression (Figure E–G). GSEA was performed to elucidate the distinct molecular mechanisms between the high and low SLC2A3 groups. The analysis indicated that the high-SLC2A3 group was enriched in pathways related to the extracellular matrix, including structural constituents and collagen trimers. In contrast, the low-SLC2A3 group was enriched in pathways associated with P450 functions, such as aromatase and arachidonic acid monooxygenase activities (Figure H). Furthermore, GO analysis indicated that SLC2A3 significantly influenced extracellular matrix functions related to fibroblasts, while KEGG analysis identified the PI3K-Akt signaling pathway and cytokine–cytokine receptor interaction as the most enriched pathways (Figure I).

5.

Functional analysis of SLC2A3. (A) K–M analysis showing that SLC2A3 exhibited an adverse prognostic effect on BLCA. (B–D) Subcellular localization of SLC2A3 in CACO-2/Hep-G2/U2OS cells from HPA data sets. Scale bar = 20 μm. Green, target protein; blue, nucleus; red, microtubules. (E) Coexpression heat map of SLC2A3. (F–G) The volcano map and heatmap show the expression of genes regulated with the expression of SLC2A3. (H) GSEA was performed to identify molecular mechanisms of the high- and low-SLC2A3 groups. (I) GO and KEGG analysis show the pathways affected by SLC2A3.

The distribution of these clusters in the tumor group differed significantly from that in the normal group (Figure A). Cell type annotation revealed that the collagen family was a key characteristic of fibroblasts. Moreover, the bubble diagram showed that SLC2A3 expression in fibroblasts was significantly different between the tumor and normal groups (Figure B). Spatial transcriptomic analysis confirmed that SLC2A3 exhibits broadly elevated expression in tissue regions enriched with fibroblasts (Figure C). Intercellular communication networks were developed to predict cellular interactions based on specific pathways and ligand–receptor dynamics. These interactions were significantly more numerous and stronger in the tumor state than in the normal state (Figure D,I). Collagen was identified as a key signaling pathway in both normal and tumor tissues, primarily affecting endothelial and epithelial cells through fibroblasts (Figure E,J). The bubble diagram revealed that the APP-CD74 ligand–receptor pair was the most crucial interaction between fibroblasts and other cells in normal tissues. However, in tumor tissues, the CD99–CD99 interaction emerged as the most significant. Nonetheless, the collagen family still exhibited the highest number of significant receptor–ligand pairs and remained the most important pathway (Figure F,K). Regardless of the tissue type, fibroblasts were more likely to act as secreting cells rather than target cells. The detailed communication patterns are shown in the Sankey diagram (Figure G,H,L,M). These findings indicate the significant role of fibroblasts in BLCA progression and their close association with SLC2A3 expression. Nevertheless, this association necessitates further investigation.

6.

Further analyses of the biological role of SLC2A3. (A) The UMAP algorithm was used to classify different cell types and showed the distribution of the types in normal and tumor tissues. (B) Expression of the key gene SLC2A3 in different cell clusters. (C) Spatial transcriptomic analysis. (D) Number and strength of interactions between each cell type in normal tissues. (E) Outgoing and incoming signaling patterns of different cell types in normal tissues. (F) Relative contribution of each ligand–receptor pair to the primarily signaling pathway between different cell clusters in normal tissues. (G,H) Interaction patterns of different cell types in normal tissues. (I–M) Similar to C–G, although the object of description were tumor tissues.

The UMAP analysis categorized fibroblasts into four subclasses: fibroblasts, CAFs, myofibroblasts, and others. The proportion of CAFs was significantly higher in tumor tissues than in normal tissues, indicating their potential role in BLCA progression (Figure A). SLC2A3 expression significantly increased in the CAF subgroup (Figure B). Additionally, pseudotime analysis was performed on all annotated cells to explore the differentiation trajectories of various fibroblast subtypes during BLCA development. In normal tissues, SLC2A3 was predominantly expressed in fibroblasts and played crucial roles in fibroblast development and differentiation. In BLCA tissues, SLC2A3 played an important role in the development and differentiation of fibroblasts, myofibroblasts, CAFs, and other cell types. Notably, SLC2A3 expression reached its peak in CAFs, whereas the differentiation of CAFs in BLCA tissue increased significantly, indicating that SLC2A3 may primarily influence BLCA progression by modulating CAFs (Figure C,I). Additionally, intercellular communication networks were constructed to predict cell interactions through specific pathways and ligand–receptor pairs. CellChat analysis indicated the importance of fibroblasts and myofibroblasts in normal tissue networks; in tumor tissues, interactions involving CAFs and other cell types were prominent (Figure D,J). This suggests that increased SLC2A3 expression in CAFs may significantly influence the development of BLCA. Moreover, analysis of receptor–ligand interactions between tumor and normal groups revealed significant changes in collagen family molecule expression, suggesting their role as key signaling molecules in tumor progression (Figure E,K). In normal tissues, myofibroblasts were the most important cell type within the cell communication network (Figure F). Although collagen was more prominent in the outgoing communication patterns (Figure G), myofibroblasts were the cell population with the highest collagen expression activity (Figure H). However, in tumor tissues, CAFs replaced myofibroblasts, as the most active cell population in intercellular communication and the most highly expressed collagen cell population (Figure L–N).

7.

Further study of fibroblast subgroups. (A) The UMAP algorithm was used to classify different types of fibroblasts and showed the distribution of the types in normal and tumor tissues. (B) Expression of the key gene SLC2A3 in different fibroblast clusters. (C) Pseudotime analysis revealing three subsets of normal tissue cells with distinct differentiation patterns and showing the relative expression of SLC2A3 in different fibroblast types in normal tissues. (D) Number and strength of interactions between each fibroblast type in normal tissues. (E) Relative contribution of each ligand–receptor pair to the primarily signaling pathway between different fibroblast clusters in normal tissues. (F) Outgoing and incoming signaling patterns of different fibroblast types in normal tissues. (G) Interaction patterns of different fibroblast types in normal tissues. (H) The inferred collagen signaling pathway networks and heatmap showing the relative importance of each fibroblast group based on the computed four network centrality measures of the collagen signaling pathway. The relative contribution of each fibroblast type to the collagen signaling pathway is shown. (I–N) Similar to C–H, although the object of description were tumor tissues.

We analyzed the fibroblast subpopulation with high SLC2A3 expression to investigate its relationship with CAFs (Figure A). Copy number variation (CNV) analysis revealed a significant increase in the CNV level of SLC2A3-positive fibroblasts (SLC2A3 + Fs) in the tumor group compared to that in the normal group, indicating that SLC2A3+Fs are key malignant cells in BLCA (Figure B). Spatial transcriptomic analysis confirmed significant colocalization between SLC2A3 and the key cellular subpopulation of CAFs (Figure C). Pseudotime analysis showed that SLC2A3+Fs primarily existed in tumor tissues, and their numbers gradually increased with cell differentiation; they were almost nonexistent in normal tissues (Figure D,J). Additionally, intercellular communication networks were developed to predict interactions through specific pathways and ligand–receptor pairs. CellChat analysis revealed that fibroblasts and myofibroblasts played key roles in intercellular communication within normal tissue networks, while interactions between CAFs and SLC2A3+Fs were prominent in tumor tissues (Figure E,K). This finding suggests that increased levels of SLC2A3 + Fs may significantly influence the development of BLCA. Comparison of receptor–ligand performance between tumor and normal groups revealed significant changes in the expression of collagen family molecules. SLC2A3 + Fs primarily interacted with CAFs via the collagen pathway, suggesting the role of SLC2A3 in CAF transformation (Figure F,L). In normal tissues, myofibroblasts were the most important cell type in the cell communication network (Figure G). Compared to normal tissues, collagen was more frequently observed in both incoming and outgoing communication patterns (Figure H,N), with myofibroblasts being the most actively expressed cell population in the collagen pathway (Figure I). However, in tumor tissues, CAFs and SLC2A3 + Fs replaced myofibroblasts, becoming the most active groups for intercellular communication and the most highly expressed cell populations in the collagen pathway (Figure M,O).

8.

Further study of SLC2A3 high-expression fibroblast (SLC2A3 + F) population defined from fibroblast cell population. Subgroups. (A) Expression of the key gene SLC2A3 in different fibroblast clusters. (B) CNV analysis was performed to measure the malignancy of the cell population. (C) Spatial transcriptomic analysis on CAFs and SLC2A3. (D) Pseudotime analysis showing three subsets of normal tissue cells with distinct differentiation patterns and the relative expression of SLC2A3 in different fibroblast types in normal tissues. (E) Number and strength of interactions between each fibroblast type in normal tissues. (F) Relative contribution of each ligand–receptor pair to the primarily signaling pathway between different fibroblast clusters in normal tissues. (G) Outgoing and incoming signaling patterns of different fibroblast types in normal tissues. (H) Interaction patterns of different fibroblast types in normal tissues. (I) The inferred collagen signaling pathway networks and the heatmap showing the relative importance of each fibroblast group based on the computed four network centrality measures of the collagen signaling pathway. The relative contribution of each fibroblast type to the collagen signaling pathway is shown. (J–O) Similar to C–H, although the object of description were tumor tissues.

We explored the expression of PFKFB3, a gene related to SLC2A3, across different cell populations. The MEbrown module was identified as the key module based on the correlation coefficient and P value (Supporting Information Figure S7A). Genes with high connectivity in the MEbrown module were selected to build a connectivity network, and key genes within this network were analyzed (Supporting Information Figure S7B,C). We confirmed that PFKFB3 was significantly correlated with SLC2A3 and that high expression of PFKFB3 resulted in a lower survival rate (Supporting Information Figure S7D,E). Pan-cancer analysis revealed a significant upregulation of PFKFB3 expression in various cancers, such as BLCA, relative to that in normal tissues, indicating its potential role as a cancer-promoting gene (Supporting Information Figure S8A). Using the UMAP algorithm for further analysis, we found that PFKFB3 expression was found to be elevated in fibroblasts and endothelial cells within tumor tissues compared to normal tissues (Supporting Information Figure S8B). Our analysis of the fibroblast subgroup revealed a significantly higher expression levels of PFKFB3 in tumor CAFs than in normal tissues (Supporting Information Figure S8C). Furthermore, pseudotime analysis revealed a progressive increase in PFKFB3 expression in CAFs as the tumor advanced, implying the potential activation of the SLC2A3 signaling axis in these cells (Supporting Information Figure S8D). This study identified a potential developmental pathway for BLCA involving SLC2A3, CAFs, and the collagen family, with SLC2A3 and PFKFB3 genes potentially playing key roles.

Immunotherapy and Drug Sensitivity Analysis of SLC2A3

TIDE analysis was employed to assess the impact of SLC2A3 on BLCA prognosis by predicting tumor immune escape likelihood in patients. The results showed that T cells in tumors with high SLC2A3 expression were more dysfunctional and exclusionary than those in tumors with low SLC2A3 expression. Additionally, T cells in tumors with high SLC2A3 expression exhibited a lower MSI expression signature, leading to a higher TIDE score (Figure A,B). Thus, SLC2A3 expression may cause an adverse immunotherapeutic response. TME analysis showed that tumors with lower SLC2A3 expression had higher tumor purity (Figure C). The analysis of humoral immunity-related gene expression revealed that genes such as IDO1 and CD44, which are indicators of poor prognosis, were significantly upregulated in the high SLC2A3 expression group, reinforcing the association between elevated SLC2A3 levels and adverse outcomes in BLCA (Figure D). The IPS analysis of the four subgroups revealed that patients with BLCA having low SLC2A3 expression exhibited higher IPS in the CTLA4-PD1- and CTLA4⁺ PD1- subgroups, but lower IPS in the CTLA4⁺ PD1⁺ subgroup, indicating that low-risk patients have greater immunogenicity and a more favorable response to immunotherapy (Figure E). Regarding PFKFB3, T cells in tumors with high PFKFB3 expression levels were more exclusionary than those in tumors with low PFKFB3 expression levels, leading to a higher TIDE score (Supporting Information Figure S8E). TME analysis indicated that tumors with low PFKFB3 expression had high tumor purity, similar to tumors with low SLC2A3 expression (Supporting Information Figure S8F). IPS comparison across the four subgroups showed no significant differences between patients with BLCA having low and high PFKFB3 expression (Supporting Information Figure S8G). BP analysis revealed a link between immune function pathways and SLC2A3 expression, indicating that pathways such as APC coinhibition, APC costimulation, CCR, checkpoint, cytolytic activity, HLA, inflammation-promoting, parainflammation, T-cell coinhibition, T-cell costimulation, and type I IFN response were significantly upregulated in the high-SLC2A3 expression group (Figure F). In the immunotherapy cohort, ICL analysis was conducted on both high- and low-SLC2A3 expression groups, highlighting that low SLC2A3 expression may enhance immunotherapy outcomes.

9.

Immunotherapy and drug susceptibility analyses results centering on SLC2A3. (A,B) The TIDE analysis was used to predict the likelihood of tumor immune escape in high- and low-SLC2A3 groups. (C) TME-related scores between high- and low-SLC2A3 groups. (D) Different expression levels of HIRGs between the high- and low-SLC2A3 groups. (E) IPS score of the low- and high-SLC2A3 expression in the CTLA4- PD1-, CTLA4- PD1⁺, CTLA⁺ PD1-, and CTLA⁺ PD1⁺ subgroups. (F) The expression levels of 13 immune function-related pathways between low- and high-SLC2A3 groups. (G) Drug sensitivity analysis was performed for low- and high-SLC2A3 groups (H) Drugs highly associated with SLC2A3 were screened based on GDSC and CTRP database. (I) Molecular docking simulating the binding of the drug to the target SLC2A3. (J) Molecular dynamics simulation of the protein–ligand complex. RMSD, Rg, SASA, RMSF and HBonds of the protein–ligand complex over time were presented.

A drug sensitivity analysis was performed on the low- and high-SLC2A3 expression groups to predict their sensitivity to common cancer drugs. Sensitivity differences between the two groups were assessed for staurosporine, alpelisib, talazoparib, and niraparib, along with their approved clinical usage or trials (Figure G,H). Molecular docking analysis revealed that the binding energies required for these drugs to bind to the SLC2A3 protein were −10.0, −10.2, −10.1, and −9.5 kcal/mol, respectively, indicating strong affinities between SLC2A3 and these drugs (Figure I). As shown in Figure J, the smaller root-mean-square deviation and root-mean-square fluctuation values indicated that the four drugs exhibited high stability when bound to the SLC2A3 target protein. The radius of gyration and solvent-accessible surface area of the SLC2A3–alpelisib, SLC2A3–niraparib, SLC2A3–talazoparib, and SLC2A3–staurosporine complexes showed minor fluctuations during the simulation, suggesting conformational changes in the protein–ligand complexes during dynamic motions. Hydrogen bond analysis further confirmed robust hydrogen bonding interactions in all four complexes. These results collectively demonstrated that alpelisib, niraparib, talazoparib, and staurosporine bound effectively to SLC2A3, highlighting their potential as targeted therapeutic agents.

Discussion

Previous studies have shown that BLCA progression is closely linked to tumor advancement and an inhibitory tumor immune microenvironment, which promotes T-cell exclusion and immune escape, ultimately reducing patients’ response to immunotherapy. , ICD is recognized as a beneficial factor against tumors. When combined with immunotherapy, particularly ICIs, ICD can reverse the immunosuppressive environment, enhance cytotoxic T lymphocyte (CTL) activity, amplify antitumor effects, and inhibit metastasis. , ICD is currently being used to develop prognostic and immune response signatures in BLCA, with the ICD-related gene HSP90AA1 identified as a potential therapeutic target. , Our study differs from previous ones by examining the role of specific cell populations in BLCA development and the localization of key genes within these cell subsets, providing valuable insights for advancing precision therapies. Additionally, our study revealed that expression levels of IRGs were higher in BLCA than in normal tissues and were negatively correlated with patient prognosis, consistent with previous research. Recognizing the heterogeneity of BLCA, we performed CC analysis on IRGs and categorized TCGA data set patients into subgroups c1 and c2 based on variations in IRG expression. Cluster c1 exhibited increased expression of immune core components, including HIRGs, HLA, and immune checkpoint genes. We also found that the gene set associated with cluster c1 was significantly enriched in the immune activity–signaling pathway. Therefore, cluster c1 was more sensitive to immunotherapy response than c2, which helped us gain a better prognosis in patients with BLCA.

We further developed an ICD-related prognostic signature to assess BLCA prognosis based on IRGs, categorizing patients into high- and low-risk groups based on risk scores. Immune infiltration and TMB analyses revealed that low-risk patients exhibited greater immune cell infiltration and a lower TMB, indicating a potentially better response to immunotherapy. Although an AUC value above 0.7 typically indicates strong discrimination between high- and low-risk patients, our model, despite not reaching this threshold, showed improvement compared to a previous study (AUC < 0.65), highlighting its potential in identifying IRGs. To overcome the limited predictive power of single-factor prognostic signatures, nomograms improve accuracy by integrating multiple risk factors into a comprehensive predictive framework. Our nomogram, which integrates age, tumor stage, and our risk score, demonstrated high predictive accuracy (AUC > 0.7) for OS in patients with BLCA, confirming that our prognostic model serves as a clinically actionable tool for outcome assessment. Another study found that integrating multimodal features, such as pathomics, radiomics, and immune scores, into prognostic models can significantly enhance the accuracy of OS prediction in cancer patients (AUC > 0.85), highlighting a promising direction for refining prognostic models. Through the strategies outlined above, ensemble machine learning can also significantly enhance the robustness and clinical utility of prognostic models.

To further explore the biological mechanism through which ICD affects BLCA progression, we used machine learning to screen for the most critical gene, SLC2A3. We found that SLC2A3 was highly expressed in BLCA tissues, especially in CAFs, and was associated with a poor prognosis. SLC2A3 is a member of the solute carrier 2A (SLC2A) family (also known as the glucose transporter or GLUT family), which plays a crucial role in mediating glucose transport across the plasma membrane and is regarded as a central regulator of cellular energetics. Dysregulation of glucose metabolism is a primary characteristic of BLCA, with tumor cells shifting toward aerobic glycolysis as their primary energy source to support proliferation. GLUT3 promotes tumor growth by enhancing glucose uptake and nucleotide synthesis. Therefore, inhibiting SLC2A3 expression can hinder tumor proliferation and progression in BLCA. This study identified SCL2A3 as both a prognostic marker and a potential therapeutic target for BLCA. By analyzing the scRNA-seq and spatial transcriptome data, we found that SLC2A3 was predominantly expressed in fibroblasts. Through intercellular communication analysis, we found different communication patterns between normal and tumor tissues. In normal tissues, endothelial cells primarily functioned as signal recipients, while fibroblasts predominantly served as signal transmitters. In tumor tissues, epithelial cells primarily functioned as signal recipients, and fibroblasts mainly served as signal transmitters. In both normal and BLCA tissues, various cell types primarily interacted via the collagen pathway. As the predominant protein in the TME, collagen plays a crucial role in maintaining TME structure and significantly influences tumor progression. This pathway can serve as both a prognostic marker and a therapeutic target in BLCA.

Analysis of the fibroblast subgroup revealed a significant increase in the number of CAFs in tumor tissues compared to that in normal tissues. As the tumor progressed, peak SLC2A3 expression shifted from fibroblasts and myofibroblasts to CAFs, indicating that SLC2A3 plays a crucial role in promoting CAF proliferation and differentiation. The central cells of the collagen pathway also shifted from myofibroblasts to CAFs, suggesting a potential association between SLC2A3 and this pathway. Therefore, we defined SLC2A3 + F population from the fibroblast cell population. We observed that SLC2A3 + Fs gradually underwent differentiation and proliferation, exhibiting close intercellular communication with CAFs primarily through the collagen pathway, which potentially serves as a mechanism by which SLC2A3 promotes BLCA progression. In addition, PFKFB3 was identified as a relevant gene of SLC2A3. We found that PFKFB3 was abnormally elevated in the fibroblast group of tumor tissues, particularly in the CAF subgroup, confirming that SLC2A3 is a key gene involved in BLCA progression. CAFs, often originating from fibroblasts, are key elements of the TME that influence tumor cell behavior and promote processes such as proliferation and invasion. Targeting CAFs or their mediators presents a promising strategy for cancer therapy. Studies have highlighted the crucial role of CAFs in BLCA progression, particularly in remodeling the extracellular matrix, which affects immune surveillance. , Thus, inhibiting SLC2A3 can modify CAF behavior and the TME, thereby improving patient prognosis and broadening treatment options. Currently, various innovative therapies aim to enhance the TME and induce ICD by targeting CAFs, which enhances the effectiveness of immunotherapy. , Our findings further support the development of new treatment strategies for BLCA by targeting the SLC2A3–CAF axis. Combining SLC2A3 inhibitors with conventional immunotherapeutic options, such as ICIs, may enhance immunotherapy response rates and improve the prognosis for patients with BLCA.

Using drug sensitivity analysis, we further identified potential SLC2A3 inhibitors with clinical applicability. Among the candidates, staurosporine, alpelisib, talazoparib, and tiraparib showed significant inhibitory activity and strong binding affinity to SLC2A3 in molecular docking. Staurosporine, a highly potent natural kinase inhibitor, is a leading compound in drug development across various therapeutic areas, such as cancer therapy, owing to its pro-apoptotic properties. Alpelisib, a novel breast cancer drug, shows significant potential in immunotherapy applications. It blocks the proliferation of cancer cells by inhibiting PI3K, which is closely related to the GLUT family. Talazoparib is a targeted drug used to treat breast and prostate cancers. , It primarily inhibits the proliferation and repair of cancer cells by blocking poly­(ADP-ribose) polymerase (PARP), leading to cell death and reduced growth. , Similarly, niraparib is another PARP inhibitor that targets prostate and ovarian cancers, enhancing tumor sensitivity to PD-L1 inhibitors. , Alpelisib, talazoparib, and niraparib have been approved by the FDA for clinical use as anticancer drugs, whereas staurosporine, although not directly used in clinical settings, highlights the potential value of its derivatives and analogues. These drugs may provide accurate and personalized treatment strategies for addressing tumor heterogeneity in patients with abnormal SLC2A3 expression.

The limitations of this study include unvalidated collagen–SLC2A3 linkage, undefined mechanisms of SLC2A3-mediated CAF differentiation, and CAF heterogeneity effects, which require further validation through multicenter cohorts, preclinical models, and mechanistic investigations. Notably, candidate drugs exhibit toxicity profiles. Combining these drugs with emerging nanotargeting technologies may enable precise targeting of CAFs with high SLC2A3 expression, leading to improved treatment specificity and reduced side effects.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c01496.

• Figures (Figures S1–S8), detailed descriptions of the specific software tools and data sets utilized in this study (PDF)

§.

M.S., Q.Z. and Z.H. contributed equally. M.S. & Y.Z; Data curation: Q.Z., L.P., J.Z., H.M. & T.Q; Formal analysis: Q.Z. & Z.H; Methodology: J.Z., H.M. & T.Q.; Visualization: M.S., Q.Z., Z.H. & L.P.; Writingoriginal draft: M.S., Q.Z., Z.H. & Y.L.; Writingreview and editing: M.S., Z.H., Y.L., Y.Z. & Y.T.; Funding acquisition: M.S. Y.Z. & Y.T.; Supervision: M.S. & Y.Z. All authors have read and agreed to the published version of the manuscript.

The Human Research Ethics Committee of the First Affiliated Hospital of Wenzhou Medical University reviewed and approved the research involving human participants. This study collected samples from 10 to 20 BLCA patients during 2022–2024. All procedures were conducted according to the Helsinki Declaration’s outlined guidelines. The ethics number was KY2022-069 (approval date: 2022-05-19). Clinical trial number: not applicable. The participants gave their written informed consent to take part in this study.

The authors declare no competing financial interest.