The overall workflow of our study is shown in Figure 1 . The data sources and workflow details are shown in the Supplementary Material .

Based on several LLPS-related genes, the patients with BLCA were classified into different subtypes using non-negative matrix factorization (NMF). The patients were classified into three LLPS clusters: C1, C2, and C3. The biological and clinicopathological features of the three subtypes were investigated. The workflow is summarized in the Supplementary Material .

Based on previous studies, the individual LLPS patterns were identified. Stepwise multivariate Cox regression analysis was performed to create a scoring system called LLPS-related risk score (LLPSRS) (18). The LLPSRS formula is as follows:

The performance of the LLPSRS formula was further evaluated to predict the clinical outcomes of BLCA patients. The robustness and versatility of the LLPSRS formula were also validated. The details of the methodology are described in the Supplementary Material .

An artificial neural network (ANN) was established using a binary classifier to identify patients who might benefit from ICB. The formula for calculating classification score using the ANN model is as follows:

The details of the procedure are described in the Supplementary Material .

All statistical analyses were conducted using R (https://www.r-project.org/). The Wilcoxon test was used to compare two groups, and the Kruskal–Wallis test was used to compare more than two groups. The statistical details and experimental methods are summarized in the Supplementary Material .

Figure 1 shows a flow diagram that systematically describes our study. A total of 3,633 LLPS-related genes were identified from TCGA-BLCA cohort and extracted from the data resource of LLPS (DrLLPS) (19) of which a total of 586 prognostic genes were identified using univariate Cox regression analysis (p< 0.01). To determine the impact of these genes on BLCA, the Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed. These genes enriched processes associated with the extracellular matrix, immunoreaction, transcription, proliferation, and metabolism ( Figure 2A ). The patients from TCGA-BLCA cohort were categorized into three LLPS clusters based on the expression of 586 genes using NMF ( Supplementary Figures 1A, B , Figure 2B ). Principal component analysis (PCA) was further used to validate the differential expression of 586 genes in three clusters, and the clusters’ similar consistency could be distinguished ( Figure 2C ). The Kaplan–Meier (KM) survival curve revealed significant differences in the prognoses of patients among three clusters (log-rank test, p< 0.0001). The clinical outcomes of patients in C2 were significantly better compared to those in C1 and C3 ( Figure 2D , Supplementary Table 2 ). To determine the reproducibility of LLPS clusters, three external BLCA cohorts were integrated into a meta-BLCA cohort, and three distinct clusters were identified as anticipated ( Supplementary Figures 1C–E ). A significant difference was observed in the prognoses of patients among the three clusters (p< 0.001); the prognoses of patients in C2 were the best, thereby confirming that three robust LLPS clusters exist in BLCA ( Supplementary Figure 1F ).

Various studies have demonstrated LLPS’s correlation with dysregulation in the TME remodeling and signaling pathways. Hence, the characteristics of TME were analyzed in three LLPS clusters. A decrease in levels of most tumor-infiltrating immune cells (TIICs) like CD4⁺T, CD8⁺T, NK, dendritic cells, and macrophages was observed in cluster C2 ( Figure 2E ). In tracking tumor immunophenotype (TIP), a significant increase in antitumor immune responses was observed in C1 and C3 compared to C2; however, these results were inconsistent with survival outcomes ( Figure 2F ). Next, the representative hallmark gene sets of the three clusters were visualized. The hallmarks of C1 enriched oncogenic signaling pathways like the PI3K-AKT-mTOR, P53, mTORC1, and MYC signaling pathways. The hallmarks of C3 enriched the stromal signatures like epithelial–mesenchymal transition (EMT) and angiogenesis. However, the downregulation of hallmarks associated with immune responses like IL6-JAK-STAT3 or NF-κB-TNFα signaling pathways was observed in C2 ( Figure 2G ). A previous study has shown that TME can be divided into three immunophenotypes: inflamed, excluded, and desert (20). Despite the abundance of TIICs in excluded TME, the TIICs only surrounded the parenchyma. No TIICs were observed in the parenchyma, which indicates that the stromal barrier surrounding the tumor inhibited the cytotoxic effects of TIICs in the TME. Since interstitial activation could suppress the effect of TIICs, the differential enrichment of 12 BLCA signatures (21) was explored among three LLPS clusters ( Supplementary Figure 1G ). As expected, C1 was of basal subtype featured by heightened IFN response, mitochondrial dysfunction, etc., which indicates damage caused by inflammation and cancer development. C3 was featured by the presence of EMT, etc., indicating fibrosis and muscle invasion. Meanwhile, cluster C2 was of a luminal subtype with Ta stage and high differentiation grade; low infiltration of TIIC could likely be due to small tumor size. Further, the enrichment of processes associated with stromal activation, mismatch repair, and immune response-relevant (SA-MR-IR) signatures was determined in the three LLPS clusters ( Supplementary Figure 1H ). The processes enriched in patients in C1 were mismatch repair, including homologous recombination, base excision, and repair. C3 was characterized by stromal activation, including angiogenesis, Pan-F-TBRS, and EMT. Downregulation in processes related to an immune response, like CD8⁺T effector and immune checkpoint, was observed in cluster C2. These results suggested that the TME of three LLPS clusters had distinct immunophenotypes and enriched different oncogenic processes. Together, these results indicated that LLPS played an indispensable role in BLCA. Further, 19 oncogenic pathways’ enrichment was analyzed among LLPS clusters ( Supplementary Figure 1I ) (14). As expected, the pathways enriched in C1 were associated with the cell cycle, including activated MYC and PI3K signaling pathways. The WNT signaling pathway was enriched in C3, confirming the increased EMT, metastasis, and muscle invasion. Interestingly, the Hippo, NOTCH, and RAS pathways were inhibited in C2. Thus, the unique TME characteristics of three LLPS clusters were analyzed.

To unravel the potential biological behavior of three LLPS clusters, a total of 470 differentially expressed genes (DEGs) were identified and annotated ( Supplementary Figure 2A ). These DEGs significantly enriched the pathways associated with metabolism reprogramming, dyssecretosis, increase in cell-autonomous proliferation, alteration biosynthesis, extracellular matrix, and immunoediting ( Figures 3A–D ). To explore a more accurate classification of BLCA subtypes and uncover underlying mechanisms, TCGA-BLCA cohort was classified into three clusters based on differential expression of 197 prognosis-related genes (p< 0.01) using NMF ( Supplementary Figures 2B, C , Figure 3E ). The patients were classified into three clusters: DEG-C1, C2, and C3. A total of 99 patients with BLCA were classified in C1, 175 patients with BLCA were grouped in C2, and 108 patients with BLCA were classified in C3 ( Figure 3E ). PCA was used to validate the expression patterns of 197 DEGs in three DEG clusters, and similar consistency was distinguished ( Figure 3F ). Additionally, the KM curve showed significant differences in the prognoses of patients among three DEG clusters (log-rank test, p< 0.0001). The OS of patients in DEG-C2 was significantly better compared to that of patients in DEG-C1 and C3 ( Figure 3G ). Most patients in LLPS-C2 were classified in DEG-C2 (166/181 = 91.72%), 81.30% of patients in LLPS-C1 (100/123) were reassigned to DEG-C3, and 87.18% of patients in LLPS-C3 (68/78) were included in DEG-C1. Eventually, the TME characteristics of three DEG clusters were analyzed, and the results were similar to our previous results ( Figures 3H–J , Supplementary Figures 2D–F , Supplementary Table 3 ).

Given the heterogeneity and complexity of LLPS, the LLPSRS was calculated to quantify LLPS-related clusters and predict patients’ prognoses. In the training cohort, 424 prognosis-related genes were identified. Univariate and least absolute shrinkage and selection operator Cox regression analyses identified 60 genes as candidate genes ( Supplementary Figures 3A, B ). Next, stepwise multivariate Cox proportional regression analysis was used to screen for 29 robust genes to calculate LLPSRS ( Figure 4A , Supplementary Table 4 ). In the training cohort, patients were classified into the high- (n = 135) and low-risk (n = 135) subgroups based on median LLPSRS as a cutoff value. The patients in the validation cohort were also divided based on these criteria. The difference in the distribution of subtypes, risk, and OS was calculated, and the results revealed significant differences in LLPSRS among LLPS or DEG clusters (p< 0.0001, Figures 4B–D ). The LLPSRS of patients in C2 was lower compared to that of patients in C1 and C3, thus suggesting that LLPSRS may be useful in predicting BLCA subtypes.

Subsequently, the ability of LLPSRS to predict the prognosis was determined. In training, TCGA-BLCA, validation, and several external cohorts, the prognoses of patients with high LLPSRS were poor ( Figures 4E–H , Supplementary Figure 3F–I ). The area under the receiver operating characteristic (ROC) curve (AUC) was used to validate the performance of LLPSRS to predict the OS of patients with BLCA ( Figures 4I, J ). In TCGA-BLCA cohort and training and validation cohorts, a decrease in the expression of 14 genes that conferred protection was observed, while 15 risk-associated genes were upregulated as LLPSRS increased ( Supplementary Figures 3C–E ). In summary, these results determined the utility and robustness of LLPSRS in predicting the clinical outcomes of patients with BLCA. In other publicly available cohorts, most patients with BLCA were Caucasians or Africans, whereas patients with BLCA in GSE188715 were Chinese; therefore, LLPSRS could be used to predict patients’ prognoses from different ethnicities. Moreover, compared to the performance of the previously published six prognostic models (14, 18, 22–25), the performance of our model was better on several appraisal algorithms. In TCGA-BLCA cohort, our model showed the highest net benefit, concordance index, and AUC, thus confirming that the adaptability of our model was better compared to that of the previously published six models ( Figures 4K–M , Supplementary Figure 3J ). Since our retrospective study was currently restricted to retrospective studies, the differential expression of seven core genes from 29 LLPSRS-related genes (RSGs) was verified in vitro ( Figure 5A , Supplementary Figure 4A ). qRT-PCR was used to study the differential expression of these genes between one bladder and two BLCA cell lines. The results revealed a significant increase in the expression of five genes in BLCA cell lines, consistent with results via TCGA-BLCA cohort; however, no difference in the expression of the other two genes was observed.

Since LLPSRS was related to the advanced stage of cancer, the performance of LLPSRS was compared with clinicopathological features in predicting patients’ OS. The univariate and multivariate Cox regression analyses were performed on TCGA-BLCA cohort, and the results showed that the hazard ratios of LLPSRS were 1.70 and 1.69 (p< 0.0001), respectively. This indicates that the performance of LLPSRS to predict prognosis was robust ( Figure 5B , Supplementary Figure 4B ). Moreover, the AUC and decision curve analysis (DCA) results confirmed the highest efficacy of LLPSRS over other clinical parameters ( Figure 5C , Supplementary Figure 4C–G ). Furthermore, to compensate for bias caused by differences in clinicopathological features, univariate Cox analysis was performed on subgroups with different clinicopathological features. It demonstrated that LLPSRS was an independent prognostic factor after adjustment ( Figure 5D ). The molecular subtypes aid in designing personalized treatment (26). Further, an association between several BLCA subtypes (21) and LLPSRS exists. Hence, the differences in the clinical landscape between the two subgroups in TCGA-BLCA cohort were explored ( Figure 5E ). The patients in the high-risk subgroup had a basal subtype characterized by high malignancy, whereas the patients in the low-risk subgroup had a more differentiated luminal subtype, and their clinical outcomes were better (chi-squared test, p< 0.05). Furthermore, the clinicopathological features of patients in the low-risk subgroup were similar to those in C2. Therefore, these results indicated that the LLPSRS could be used as a biomarker for predicting heterogeneity and designing personalized therapy for patients with BLCA.

To determine the underlying genomic alterations caused by LLPSRS, the genomic variations between the two subgroups were explored. The copy number variations (CNVs) in RSGs on chromosomes were analyzed and visualized. The results showed significant amplifications in 16 RSGs, whereas five RSGs harbored deletions ( Figure 5F , Supplementary Figure 4H ). Furthermore, the mutation landscape of the top 20 mutated RSGs is shown in Figures 5G, H . The patients in the high-risk subgroup had distinct mutation patterns, the top three mutated genes were observed in patients from the same subgroup, and there was a difference in the abundance of other mutated genes in the patients from the two subgroups. Moreover, as potential indices for ICB response and neoantigens epitopes, tumor mutational burden (TMB) and purity data were obtained. A negative correlation was observed between TMB, tumor purity, and the LLPSRS ( Supplementary Figures 4I, J , p< 0.05). In addition, the interconnection and mutation landscape of RSGs were visualized ( Supplementary Figures 4K, L ). In addition to genomic mutations, epigenetic aberrations play an important role in oncogenesis. Changes in DNA methylation pattern, which plays an important role in pre-transcriptional modification, were next explored. Previous studies have shown that the prognoses of patients with DNA hypomethylation were poor, whereas patients with DNA hypermethylation may experience cachexia (27, 28). A total of 26,583 differentially methylated CpG islands (DMCGs) were identified ( Figure 5I ). The methylation levels in patients in the high-risk subgroup were higher, which confirms our hypothesis. Furthermore, 240 genes were identified as DNA methylation driver genes (MET-DGs). These genes enriched pathways associated with biosynthesis disorders, cancer, and cell proliferation pathways, which indicates that these genes are involved in activating various oncogenic pathways ( Supplementary Figure 5A ). The top 35 MET-DGs were visualized, and the results revealed that the LLPSRS was characterized by the hypermethylation of tumor suppressor genes ( Supplementary Figure 5B ).

Of these 29 RSGs, NSUN5 is an RNA methyltransferase responsible for 5-methylcytidine (m⁵C) modification, whereas DHX16 and HNRNPH3 regulate alternative splicing (AS) during pre-mRNA splicing. Knockdown of NSUN5 expression reduces cell proliferation (29), DHX16 is a biomarker for immune-related adverse events (irAEs) (30), and HNRNPH3 directly alters the mRNA splicing of proto-oncogene MST1R (31). Since RNA editing and AS are key regulators of carcinogenesis, the correlation between the LLPSRS and post-transcriptional regulation was explored. The expression of 95 RNA-editing regulators in different subgroups was analyzed, and the results revealed differential expression in most RNA-editing regulators between two subgroups ( Supplementary Figures 5C–E ). Next, a correlation was observed between 786 out of 1,416 DEGs and 71 RNA-editing regulators. These genes were annotated as hormone secretion, cell division, metabolic rewiring, and immunosuppressive TME (|R| > 0.3, p< 0.0001, Supplementary Figures 5F, G ). A network was constructed using 21 RNA-editing regulators and LLPSRS to demonstrate their correlation and prognostic roles. Interestingly, a correlation was observed between LLPSRS and most regulators ( Figure 5J , univariate Cox regression analysis, p< 0.05). In addition to the correlation between LLPSRS and m⁵C regulators, a correlation between LLPSRS and other regulators was also observed. This indicates that diverse crosstalk between LLPSRS and RNA-editing regulators plays a crucial role at the epitranscriptomic level in BLCA. Moreover, 483 AS events were observed in 161 DEGs. These genes enriched pathways associated with carcinogenesis and immunogenicity, thus implying that LLPSRS may regulate these pathways by AS of DEG ( Figure 5K , Supplementary Figure 5H ). Correlation networks were constructed between 37 AS events and 22 splicing factors ( Supplementary Figure 5I ). Taken together, our results demonstrated that the LLPS-mediated epigenetic alterations played an important role in the progression of BLCA.

The GO and KEGG pathway enrichment analyses were performed using DEGs to decipher LLPSRS-relevant biological mechanisms using KOBAS-i (32). The correlation between the LLPSRS, SA-MR-IR, and oncogene signatures was further explored. The LLPSRS was associated with metabolic rewiring, cell-autonomous hyperproliferation, immunoediting, and biosynthesis derangements, thus indicating that LLPSRS played an important role in carcinogenesis ( Figures 6A–E ). Higher LLPSRS reveals persistent stromal activation and transcriptional dysregulation, thus implying tumor cell-autonomous proliferation and excluded immunophenotype ( Figure 6F ). Meanwhile, a positive correlation was observed between LLPSRS and a decrease in the expression of tumor suppressor genes, as well as an increase in oncogenes’ expression like PI3K, RAS, and TGF-β, thereby suggesting a higher degree of malignancy. As expected, the gene set enrichment analysis results revealed that hallmarks of cancers like angiogenesis, EMT, and the TGF-β and WNT signaling pathways were upregulated in the patients in the high-risk subgroup. This indicates the frequent occurrence of cachexia-relevant signatures and an increase in the activation of oncogenes ( Figure 6G ).

As discussed previously, LLPSRS-related mechanisms play an important role in regulating immune responses; hence, the correlation between LLPSRS and intrinsic as well as extrinsic immunoediting was analyzed. A negative correlation was observed between LLPSRS and MHC receptors. A positive correlation was observed between LLPSRS and immunosuppressive checkpoints, thus indicating an increase in immune evasion by tumor cells and a decrease in immunogenicity. Interestingly, a correlation was observed between low LLPSRS and some ICB-related genes like (CTLA4, LAG3, PDCD1, etc.), thus indicating higher sensitivity of patients to ICB ( Figure 6H ). Apart from the chemokines and their receptors, a positive correlation was observed between the protooncogenes and LLPSRS. A negative correlation was observed between other tumor suppressors and LLPSRS. On the contrary, a negative correlation was observed between several immune inhibitors, immune stimulators, and LLPSRS ( Figure 6I ). Considering the multiple complex roles of immunomodulators, the association between immunomodulators and LLPSRS was insufficient to elucidate overall immunological features. The extrinsic immunoediting indicated by TIICs was equally important since they revealed a functional repertoire of antitumor immunity. A correlation between the LLPSRS and TIICs was observed. Increased levels of antitumor TIICs, like CD4⁺T and CD8⁺T, were observed, whereas the levels of pro-tumor TIICs like Treg and Th2 were low in patients in the low-risk subgroup ( Figure 6J ). As expected, a positive correlation was observed between myeloid cells, stromal cells, and LLPSRS, thereby indicating an increase in damage caused by inflammation and interstitial activation ( Figure 6K ). Together, these results suggested that LLPSRS played an important role in immunoediting, which indicated that patients with lower LLPSRS had higher immunogenicity and sensitivity to ICB.

The correlation between LLPSRS and anticancer drug regimens was next determined, which could aid in designing precision medicine at the pharmacogenomics level. First, the immunophenoscore of patients in the low-risk group was high, regardless of the status of indicators, which suggests better efficacy of ICB ( Figure 7A ). As expected, the Tumor Immune Dysfunction and Exclusion analysis revealed that the patients who responded to ICB treatment had higher LLPSRS. This indicates that LLPSRS could predict the efficacy of ICB treatment ( Figure 7B ). Furthermore, the association between LLPSRS and gene signatures of adjuvant treatments was explored ( Figure 7C ). A negative correlation was observed between LLPSRS and oncogenic pathways, and a positive correlation was observed between the predicted EGFR pathway, radiotherapy, and irAEs. Therefore, patients with higher LLPSRS were sensitive to EGFR-targeted therapies and radiotherapy, whereas the patients with low LLPSRS were sensitive to oncogenes’ blockades. Additionally, an inverse correlation was observed between LLPSRS and Ta stage, luminal, and urothelial differentiation. The patients in the high-risk subgroup were characterized by immune evasion and dysfunction, and neuronal and basal differentiation, thereby indicating that patients in the high-risk group could benefit from neoadjuvant chemotherapy, whereas ICB might induce hyperprogression.

Although adjuvant treatment was a major breakthrough in cancer therapeutics, chemotherapy is still an indispensable part of BLCA treatment. To determine the use of LLPSRS in determining BLCA treatment, data on experimentally or clinically used drugs in BLCA were retrieved from the Genomics of Drug Sensitivity in Cancer database, and their efficacy was determined. Chemotherapeutic drugs like dasatinib and IWP-2 were more suitable for the patients in the high-risk group ( Figure 7D–G ), whereas the other eight chemotherapeutic drugs were more suitable for the patients in the low-risk group ( Supplementary Figures 6A–H ). To provide more avenues for LLPSRS-based therapies, the Cancer Therapeutics Response Portal (CTRP) and Profiling Relative Inhibition Simultaneously in Mixtures (PRISM) were used to analyze LLPSRS’s correlation with chemotherapeutic agents. Five CTRP (e.g., dasatinib and fluvastatin) and five PRISM drugs (e.g., epothilone-b and tosedostat) were more suitable for the treatment of patients in the high-risk group ( Figure 7H ), whereas four CTRP (e.g., apicidin and brefeldin A) and nine PRISM drugs (e.g., poziotinib and RITA; Supplementary Figure 6I ) were more suitable for the treatment of patients in the low-risk group. Further, the data obtained from the CellMiner (33) were analyzed, and 20 negative and 12 positive LLPSRS-related drugs were identified. Therefore, all patients with different LLPSRSs may respond to different chemotherapeutic drugs ( Figure 7I , Supplementary Figure 6J ).

Immunotherapy was regarded as an epoch-making breakthrough. LLPSRS’s correlation with the immunosuppressive milieu and TMB was observed, so LLPSRS could likely play a role in predicting patients’ response to ICB. We analyzed the association between LLPSRS and sensitivity to ICB in the IMvigor210 cohort. The patients in the low-risk subgroup had a longer life span and demonstrated prolonged survival as compared to those in the high-risk subgroup (log-rank test, p< 0.0001, Figures 8A, B ). We also discovered that the progressive disease subgroup, in which TIICs exerted a faint effect, had the highest LLPSRS (Kruskal–Wallis test, p< 0.01, Figure 8C ). Together, these results show a positive correlation between LLPSRS and irAEs, thus suggesting that patients with lower LLPSRS may respond better to ICB.

PD-L1 expression is an important biomarker for ICB responsiveness. Hence, we analyzed the correlation between TC (PD-L1 located on tumor cells), IC (PD-L1 located on immune cells), immunophenotype (desert, inflamed, and excluded), and LLPSRS. In the patients in the low-risk subgroup, TC2+ was the most abundant TC, thus indicating that the patients had higher sensitivity to ICB treatment (chi-squared test, p< 0.05, Figure 8D ). Next, the patients in the IC2 subgroup or immune-inflamed type had the lowest LLPSRS (Kruskal–Wallis test, p< 0.01, p< 0.05, respectively, Figures 8E, F ). A significant difference was observed in TMB between the high- and low-risk subgroups (Wilcoxon test, p< 0.001, Figure 8G ). Further, the predictive ability of the combined TMB and LLPSRS was better as compared to LLPSRS and TMB alone ( Supplementary Figure 7A ). Therefore, for those patients treated with ICB, LLPSRS could effectively predict individual responses and immunosuppressive properties.

The outcomes of patients treated with ICB were better; however, not all patients experienced durable response to ICB treatment. Hence, there is an urgent need to identify eligible patients. Further, a diagnostic model was created based on LLPSRS and cytotoxicity using ANN. Weighted Gene Co-expression Network Analysis (WGCNA) was used to predict the patient’s response to ICB and determine its relevance with LLPSRS. WGCNA was used to extract eigengenes from 1,758 DEGs between the complete response/partial response (CR/PR) and stable disease/progressive disease (SD/PD) subgroups, the optimal fitting degree was R² = 0.85, and soft-thresholding was β = 3 ( Supplementary Figures 7B–F ). After the modules were merged with a disparity coefficient< 0.45 and overall gene counts< 30, the DEGs were divided into six modules ( Figure 8H ). Given that the correlation between clinical characteristics and module eigengenes (MEs) exists, a correlation was observed between MEbrown, (the main module) and LLPSRS, binary response, OS time, OS, and cytotoxicity ( Figure 8I , Spearman’s rank test, p< 0.001). Eventually, in the MEbrown module, gene significance’s associations with module membership were analyzed, and the results of LLPSRS (rho = 0.79, p< 0.0001) and cytotoxicity (rho = 0.86, p< 0.0001) were significant ( Supplementary Figures 7G,H ); thus, 122 genes were regarded as eigengenes. The eigengenes enriched immune responses such as cytotoxicity, immunogenicity, PD-L1/PD-1 checkpoints, chemotaxis, T-cell differentiation, and interferon response, thereby confirming the successful extraction of eigengenes ( Figure 8J ).

To build a classifier for identifying ICB-sensitive patients, a combined prediction model was constructed using random forest (RF) and ANN algorithms based on eigengenes. Since the number of patients with BLCA treated with ICB was few, all samples from the IMvigor210 cohort were randomly split. In the training cohort, eigengenes were incorporated into the RF classifier, and 180 trees were selected as the parameter ( Figure 8K ). Next, the variable importance was measured, and the top 30 genes were selected ( Supplementary Figure 7I ). Based on these genes, an ANN model was created using the training cohort, which consisted of three layers: input (expression of 30 genes), hidden (scores and weights of genes), and output layer (SD/PD or CR/PR results) ( Figure 8L ). Finally, the ROC–AUC values for the ANN model in training, validation, and IMvigor210 cohorts were 1.000, 0.839, and 0.834, respectively. The accuracy of the model in predicting the response to ICB treatment in the training, validation, and IMvigor210 cohorts was 1.000, 0.750, and 0.720, respectively. Together, these results show the robustness and utility in predicting response to ICB treatment in patients with BLCA ( Figure 8M ).