This cohort was independently collected and sequenced by Xiangya Hospital and included 57 bladder cancer (BLCA) samples and 13 normal bladder tissues. The samples were processed to generate mRNA sequencing profiles, and detailed descriptions can be found in our previous studies (Hu et al., 2021a; Sheng et al., 2021; Cai et al., 2023; Li et al., 2023). The Xiangya cohort has been deposited in the Gene Expression Omnibus (GEO) database under accession number GSE1887155 (Liu et al., 2021).

The TCGA-BLCA cohort consists of transcriptomic profiles and clinical annotations from 400 BLCA patients. The dataset was initially downloaded from the UCSC Xena platform. After primary data processing and normalization, three formats of RNA sequencing data were retained, including raw counts, fragments per kilobase per million reads (FPKM), and transcripts per million reads (TPM), all of which reflected protein-coding genes only. Somatic copy number variation (CNV) was analyzed using GISTIC2.0 (Mermel et al., 2011), and tumor mutational burden (TMB) was calculated using VarScan2.

IMvigor210 is a clinical immunotherapy trial that evaluated atezolizumab, a PD-L1–blocking antibody, in BLCA patients who were ineligible for cisplatin-based therapy (Balar et al., 2017). We obtained and curated the integrated RNA sequencing matrix with corresponding clinical annotations from the study by Mariathasan and colleagues (Mariathasan et al., 2018).

GSE48075 is a transcriptomic dataset of BLCA (Choi et al., 2014), whereas GSE32894 represents urinary tract carcinoma (Robertson et al., 2017). Both datasets were retrieved from the GEO database. To reduce batch effects, the two datasets were merged into a single meta-cohort using the “sva” package in R.

We comprehensively collected all potential genes related to the SWI/SNF chromatin-remodeling complex. The sources included REACTOME, gene set enrichment analysis (GSEA), Gene Ontology (GO), and PUBMED, and a raw gene list containing 1,653 genes was constructed, which we referred to as the Raw_Gene_Set. All included genes were derived exclusively from studies conducted in humans.

Next, univariate Cox regression was performed on the Raw_Gene_Set, yielding 32 prognostically relevant genes (p< 0.001). These genes were incorporated into the Cluster_Gene_Set and subjected to unsupervised clustering. After systematic evaluation, k = 2 was determined to provide the most stable clustering outcome, generating two SWI/SNF-associated molecular clusters, namely Cluster1 (n = 275) and Cluster2 (n = 125), which were referred to as SWI_SNF_Cluster.

The SWI_SNF_Score was derived from the SWI_SNF_Cluster. Using the “limma” package in R, we identified 919 differentially expressed genes (DEGs) between the two clusters. Univariate Cox regression was again applied to screen for DEGs with prognostic relevance, yielding 285 significant genes. Principal component analysis (PCA) was performed on these 285 genes. The loading vectors of the first principal component (PC1) were used as the coefficient matrix for score calculation. Specifically, the formula is written as: SWI_SNF_Scorei=∑k=1285×PC_1ik∗ZExpik.

SWI_SNF_Scorei represents the quantitative SWI_SNF_Score value for the  i−th BLCA patient. For example, k denotes the index of a gene, ranked from 1 to 285, representing a given gene Gene_k. PC_1ik is the coefficient of Gene_k on the first principal component. ZExpik is the Z-score–standardized expression level of Gene_k.

Molecular classification of BLCA has important diagnostic, prognostic, and therapeutic implications. Based on our previous studies (Hu et al., 2021a; Liu et al., 2021; Hu et al., 2022; Cai et al., 2023; Li et al., 2023; Hu et al., 2025a; Hu et al., 2025b), multiple classification frameworks have been widely used and validated (Sjödahl et al., 2012; Choi et al., 2014; Damrauer et al., 2014; Rebouissou et al., 2014; Robertson et al., 2017; Mo et al., 2018; Kamoun et al., 2020; Woerl et al., 2020; Zeng et al., 2020; Tate et al., 2021; He et al., 2022; Li et al., 2024). In this study, molecular subtyping was performed using RNA sequencing data from publicly available cohorts through the “ConsensusMIBC” and “BLCAsubtyping” packages in R. By integrating our SWI_SNF_Score system, we further evaluated its predictive potential for different molecular subtypes. Predictive accuracy was quantified using receiver operating characteristic (ROC) analysis and area under the curve (AUC).

Kaplan–Meier survival analysis was performed using the “survival” package in R, and visualization was performed using “survminer.” For functional annotation, gene set enrichment at the single-sample level was estimated using the single-sample gene set enrichment analysis (ssGSEA) method. The gene set variation analysis (GSVA) approach was used to score molecular signatures across different cohorts. Gene set enrichment analysis (GSEA) was applied to evaluate the biological functions of DEGs, followed by GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) annotation to identify relevant pathways. Mutational profiling was visualized using the “maftools” package, and mutational differences were evaluated using chi-square tests.

Immune microenvironment analysis was conducted following methodologies from previous studies (Hu et al., 2021a; Hu et al., 2021b; Cai et al., 2023; Li et al., 2023; Yan et al., 2025). At the molecular level, we examined correlations between SWI_SNF_Score and immune-regulatory factors as well as representative immune cell markers. At the cellular level, tumor-infiltrating immune cells (TIICs) were estimated using MCP-COUNTER and TIMER, and the activation of the cancer immunity cycle was evaluated using the TIP platform. Enrichment analysis was subsequently performed to identify immune-related pathways associated with the SWI_SNF_Score.

We compared treatment-associated mutational features between high and low SWI_SNF_Score groups. Based on previous literature, chemotherapy-related genes such as TP53, RB1, and ERCC2 were selected (Motterle et al., 2020). For erdafitinib-targeted therapy, FGFR2 and FGFR3 mutations were specifically examined (Burki, 2019; Loriot et al., 2019; Siefker-Radtke et al., 2022; Guercio et al., 2023; Loriot et al., 2023). Additional analyses included radiotherapy-associated markers, targeted therapy signatures, and immunotherapy-relevant pathways.

In immune checkpoint blockade (ICB) analyses, particular attention was given to correlations between SWI_SNF_Score and markers of therapeutic response, including the T cell inflammation score (TIS), epithelial–mesenchymal transition (EMT) features, and TGF-β response signatures.

All statistical analyses and visualizations were performed in R software (version 4.3.3). A p-value below 0.05 was considered statistically significant. Patients with BLCA were dichotomized according to the median or the optimal cutoff of the SWI_SNF_Score.

For variables that met normal distribution assumptions, the Student’s t test was used for comparison between two groups. For skewed distributions, the Mann–Whitney U test was applied. Categorical variables were compared using chi-square tests. All statistical procedures adhered to biomedical analytical logic, and all hypothesis tests were conducted in a two-sided manner.

Figure 1A outlines the analytical workflow. As shown in Figure 1A and Supplementary Table S1, we identified 1,653 feature genes (Raw Gene Set) through systematic screening. Univariate Cox regression analysis prioritized 32 SWI/SNF-associated genes (COX Gene Set), whose functional relevance was corroborated by Gene Ontology (GO) enrichment analysis (Figure 1B). Using the 32-gene COX Gene Set, unsupervised consensus clustering of the TCGA-BLCA cohort identified two molecular subgroups: SWI/SNF Cluster 1 (n = 275) and SWI/SNF Cluster 2 (n = 125) (Figure 1C; Supplementary Figures S1A–H). Differential gene expression patterns between clusters were visualized via heatmap (Figure 1D). Kaplan–Meier survival analysis revealed a significant disparity in overall survival between clusters (p< 0.001), with Cluster 2 exhibiting poorer prognosis (Figure 1E).

Figure 2A illustrates the overall workflow for establishing the SWI_SNF_Score. We first identified 919 differentially DEGs between the two SWI/SNF clusters using the “limma” package (Supplementary Table S4). Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis and GO functional annotation were performed on these 285 prognostically significant DEGs, revealing their major biological functions (Figures 2B, C). These DEGs were predominantly enriched in tumor-associated pathways and extracellular matrix (ECM) regulatory processes. KEGG analysis indicated significant enrichment in the PI3K–Akt signaling pathway, TGF-beta signaling pathway, and ECM–receptor interaction.

At the biological process (BP) level, the DEGs were enriched in ECM organization, collagen metabolism, and cell–matrix adhesion. At the cellular component (CC) level, the genes were mainly located in ECM structures, collagen-associated regions, and basement membranes. At the molecular function (MF) level, the genes were enriched in collagen binding, integrin binding, and metallopeptidase activity. These results suggest that the DEGs may regulate ECM remodeling and the tumor microenvironment, thereby affecting BLCA development, molecular progression, and patient prognosis.

We then performed PCA on these DEGs to construct the SWI_SNF_Score (Figure 2D). Genes upregulated in the high-score group were primarily stromal and ECM-associated markers, including CTHRC1, POSTN, FAP, COL6A2, and ADAMTS2, indicating stromal activation and ECM remodeling (Figure 2E). Overall, the SWI_SNF_Score distinguished stromal differentiation characteristics, suggesting potential clinical implications for patient stratification and prognosis assessment.

Functional enrichment analyses revealed pronounced differences between high and low SWI_SNF_Score groups. Genes in the high-score group were enriched in pathways related to cell adhesion, immune regulation, and ECM organization, whereas genes in the low-score group were enriched in metabolic programs such as fatty acid metabolism, redox processes, and xenobiotic metabolism (Figures 2F, G). Kaplan–Meier survival analysis confirmed a significant difference in patient prognosis, with the high-score group showing poorer survival outcomes (Figure 2H). A Sankey diagram demonstrated the correspondence between SWI_SNF_Score, molecular clusters, gender, and age (Figure 2I).

The SWI_SNF_Score demonstrated clear stratification across multiple molecular classification systems. Samples with high scores were predominantly enriched in basal differentiation subtypes and exhibited EMT and immune activation features. In contrast, samples with low scores showed luminal differentiation, urothelial differentiation, and Ta pathway characteristics (Figure 3A). ROC analysis confirmed the discriminative performance of the SWI_SNF_Score across Baylor, UNC, MDA, and TCGA systems, AUC values greater than 0.8, indicating strong prediction accuracy for molecular subtypes (Figure 3B). Consistent trends were observed in pathway enrichment heatmaps. The high-score group showed elevated expression in basal differentiation, EMT differentiation, and immune differentiation, while the low-score group was characterized by luminal differentiation, urothelial differentiation, and Ta pathway signatures (Figure 3C). Previous studies on molecular stratification have shown similar patterns, where the high-score group corresponds to a more aggressive tumor phenotype and the low-score group aligns with favorable prognostic tendencies (Dyrskjøt et al., 2023).

In therapeutic response analyses, the two SWI_SNF_Score groups displayed distinct patterns (Figure 3D). The high-score group exhibited greater activity in EGFR signaling and immune-related signatures, suggesting potential sensitivity to EGFR-targeted therapy, immune checkpoint blockade, and radiotherapy. Conversely, the low-score group showed higher enrichment in Sunitinib–CSF1R and Trastuzumab–ERBB2 signatures, indicating possible suitability for CSF1R-targeting or HER2-targeting therapeutic strategies.

Mutational landscape analysis revealed a high mutational burden in both groups, with an overall mutation rate of 93.97 percent. TP53 was frequently mutated in both score groups. However, the accompanying mutation patterns differed substantially. The high-score group exhibited more frequent RB1 co-mutations, which matched basal and aggressive phenotypes. The low-score group was enriched with FGFR3 and KDM6A mutations, which were biologically consistent with luminal molecular features (Figures 3E, F). These findings demonstrate that the SWI_SNF_Score not only differentiates molecular subtypes, but also reflects differences in treatment susceptibility and mutational architecture, thereby offering potential implications for individualized treatment in BLCA.

The high SWI_SNF_Score group and the SWI_SNF_Cluster2 group exhibited elevated expression of immune-regulatory factors (Figure 4A). Consistent with this observation, the high SWI_SNF_Score group and the SWI_SNF_Cluster2 group showed significantly higher levels of immune cell infiltration (Figure 4B). As expected, most steps of the antitumor immune cycle were more activated in the high SWI_SNF_Score group and in the SWI_SNF_Cluster2 group (Figure 4B). The immune score was significantly higher in the high SWI_SNF_Score group compared with the low-score group (Figure 4C). All effector signatures showed strong correlations with the SWI_SNF_Score (Figure 4D), indicating enhanced immune infiltration in the high-score group.

Compared with the low-score group, the high SWI_SNF_Score group exhibited a stronger T cell inflammation score (TIS), indicating more pronounced immune cell infiltration. A linear relationship between the SWI_SNF_Score and TIS was confirmed, with a slope of 0.47, highlighting a strong correlation between the two variables (p< 0.001, Figure 5B). The SWI_SNF_Score was significantly associated with most key genes used to construct the TIS, except PSMB10, which showed no significant correlation (p > 0.05, Figure 5A). The SWI_SNF_Score also showed positive correlations with the majority of immune checkpoint molecules, although LGALS3 and VTCN1 did not exhibit statistically significant associations (p > 0.05, Figure 5A).

Pathways associated with immune checkpoint inhibitor (ICI) response were significantly upregulated in the high SWI_SNF_Score group (Figure 5C). Interestingly, CDKN2A and CDKN2B are genes negatively associated with immune hyper progression (Giusti et al., 2019), and their expression is also significantly enhanced in the high SWI-SNF-Score group (Figures 5D, E). Taken together, the high SWI_SNF_Score group is associated with enhanced immune infiltration, indicating the potential benefits of immunotherapy, but it is also associated with potential risks of immune hyper progression and adverse outcomes—highlighting its significant role in predicting the clinical efficacy of ICI.

In the Xiangya cohort, the SWI_SNF_Score exhibited significant associations with the molecular characteristics and clinical outcomes of BLCA. Kaplan–Meier survival analysis showed that patients with high SWI_SNF_Score values had poorer prognosis (p = 0.032, Figure 6A). The SWI_SNF_Score demonstrated strong discriminative performance for predicting classical molecular subtypes, with the area under the ROC curve ranging from 0.84 to 0.98 (Figure 6B).

Molecular subtype analysis further revealed that the low-score group was enriched for luminal differentiation, urothelial differentiation, and Ta pathway characteristics. In contrast, the high-score group presented stromal-associated features, including fibroblast activation and EMT differentiation (Figures 6C, D).

Immune-related analyses indicated that the high SWI_SNF_Score was consistent with an inflammatory tumor microenvironment. The score was significantly correlated with multiple steps of the cancer immunity cycle, including T cell recruitment and natural killer cell activation. Additionally, the high SWI_SNF_Score was closely associated with pathways involving DNA damage repair, cell cycle regulation, and virus-associated carcinogenesis (Figure 6E).

Importantly, the SWI_SNF_Score showed strong positive correlations with multiple immune checkpoint molecules, including PDCD1 (PD-1), CD274 (PD-L1), CTLA4, and TIGIT. The score was also highly consistent with the expression of a wide range of immune-related genes (Supplementary Figure S2A). Collectively, these results demonstrate that the SWI_SNF_Score not only reflects molecular subtype and prognosis in BLCA, but also reveals immune microenvironment characteristics and pathway features, providing potential value for guiding future clinical decision-making.

In the immunotherapy cohort (IMvigor210), the SWI_SNF_Score underwent additional validation. In the overall patient population, there was no significant survival difference between high and low SWI_SNF_Score groups (Figure 7A). However, more nuanced findings emerged when patients were further stratified based on treatment response. In the complete response and partial response (CR/PR) subgroup, patients with low SWI_SNF_Score values exhibited more favorable outcomes (Figure 7B). In contrast, in the stable disease and progressive disease (SD/PD) subgroup, patients with high SWI_SNF_Score values showed superior outcomes (Figure 7C).

Considering the reshaping effect of immunotherapy within the tumor microenvironment, patients in the IMvigor210 cohort were categorized into three immune phenotypes, namely the desert phenotype, the excluded phenotype, and the inflamed phenotype. Regression analysis showed a positive linear correlation between the SWI_SNF_Score and TIS in both the desert and excluded phenotypes, indicating potential predictive utility of the SWI_SNF_Score in these immune contexts (Figures 7J–L).

In the excluded phenotype and inflamed phenotype, patients with high SWI_SNF_Score values consistently exhibited poorer prognosis (Figures 7D, E). In contrast, in the desert phenotype, patients with high SWI_SNF_Score values unexpectedly demonstrated better outcomes (Figure 7F).

We presented the expression profiles of tumor-infiltrating lymphocyte (TIL) effectors, immune checkpoints (IC) markers, and ICI, and observed that TIL and IC markers were upregulated in the high SWI_SNF_Score group across all immune phenotypes, while ICI showed a more significant upregulation in the low SWI_SNF_Score group across all immune phenotypes (Figure 7M). Both TIL and ICI signatures demonstrated strong correlations with the SWI_SNF_Score (Supplementary Figure S3A). Significant differences in SWI_SNF_Score were also observed across immune cell subsets, immune phenotypes, and binary response patterns (Figures 7G–I).

By integrating the SWI_SNF_Score with the three immune phenotypes, clinicians may more effectively select treatment strategies and assess prognosis for patients receiving immunotherapy.