AJCC staging is insufficient for precision treatment. Bladder cancer is the fourth most commonly diagnosed malignancy globally, with over 430,000 new cases in 2020. Despite the widespread use of AJCC staging as the standard risk classification tool, patients with the same stage can have dramatically different clinical outcomes, suggesting that staging alone misses important biological information that drives prognosis and treatment response.
Immunotherapy with checkpoint inhibitors targeting PD-1, PD-L1, and CTLA-4 has transformed bladder cancer treatment, with atezolizumab and pembrolizumab approved as first-line therapy for platinum-ineligible patients. However, only a minority of patients respond to immunotherapy, and current predictive biomarkers including PD-L1 expression and tumor mutation burden are imperfect predictors of who will benefit.
Existing multigene signatures lack rigorous validation. A large number of gene expression-based prognostic signatures have been published for bladder cancer, but most suffer from underutilization of available data, inappropriate machine learning methods, and lack of validation across multiple independent cohorts and clinical trial datasets, severely limiting their clinical applicability. Many signatures perform well in the cohort they were trained on but fail to generalize.
The need for a robust, generalizable signature that is independently validated across large multi-center datasets, compared against existing published models, and linked to treatment response prediction motivates this study's large-scale systematic approach using 10 machine learning algorithms and 11 independent patient cohorts.
Two-step gene identification and signature construction. The study first identified robust overall-survival-related genes (RORGs) by applying univariate Cox regression across 14,843 intersection genes from 8 cohorts with complete survival data, retaining genes with consistent hazard ratio directions in at least 6 of 8 cohorts. This cross-cohort consistency requirement reduced the gene pool from nearly 15,000 to 30 robust candidates.
Ten machine learning algorithms were then applied to the 30 RORGs: Random Survival Forest (RSF), Elastic Net (Enet), Lasso, Ridge, Stepwise Cox, CoxBoost, Partial Least Squares Regression for Cox (plsRcox), Supervised Principal Components (SuperPC), Generalized Boosted Regression Modelling (GBM), and Survival Support Vector Machine (survival-SVM). These were combined into 76 algorithm combinations using feature-selecting algorithms (Lasso, Stepwise Cox, CoxBoost, RSF) as upstream selectors paired with all other algorithms as downstream model builders.
10-fold cross-validation and multi-cohort ranking. All 76 models were trained in TCGA-BLCA with 10-fold cross-validation, then applied to seven independent validation cohorts (818 patients). The consensus AI-derived Gene Signature (AIGS) was defined as the model achieving the highest average C-index across all validation cohorts, selecting the Ridge regression-based model as optimal (average C-index 0.709).
The final dataset for model development and validation included 11 independent cohorts totaling 1,351 bladder cancer patients from GEO, TCGA, and the IMvigor210 immunotherapy trial cohort. Two immunotherapy cohorts and one chemotherapy cohort enabled assessment of AIGS for treatment response prediction beyond its primary prognostic use.
Comprehensive multi-omics analyses supplemented the signature evaluation, including somatic mutation profiling of the top 30 frequently mutated genes, copy number variation analysis, methylation-driven gene identification via MethylMix, and drug sensitivity prediction using CTRP and PRISM datasets covering over 1,500 compounds in hundreds of cancer cell lines.
Significant survival separation in all 8 OS cohorts. Kaplan-Meier analysis demonstrated that high AIGS scores were significantly associated with worse overall survival in the TCGA-BLCA training cohort (n=400, p below 0.0001) and in all seven independent validation cohorts including GSE13507, GSE19423, GSE31684, GSE37815, GSE48075, GSE48276 (all p below 0.001), and the IMvigor210 immunotherapy cohort (n=348, p=0.0008). AIGS also predicted recurrence-free and progression-free survival in separate cohorts.
Time-dependent AUC analysis showed robust and consistent predictive accuracy: TCGA-BLCA achieved AUCs of 0.758, 0.738, and 0.745 at 1, 3, and 5 years, with comparable performance across validation cohorts including GSE19423 (0.941, 0.817, 0.871) and GSE37815 (0.824, 0.959, 0.938). C-indices across the 8 cohorts ranged from 0.577 to 0.825, confirming stable predictive ability across diverse datasets.
Independent of clinical and molecular features. Multivariate Cox analysis adjusting for age, gender, tumor stage, T/N/M classification, grade, treatment history (intravesical therapy, systemic chemotherapy, BCG, platinum, neoadjuvant chemotherapy), and multiple molecular mutations (FGFR3, TP53, RAS, RB1) confirmed that AIGS remains a statistically significant independent predictor of OS, RFS, and PFS across all evaluated cohorts.
AIGS demonstrated superior predictive accuracy compared to all individual clinical variables and molecular features tested, including AJCC stage, T stage, N stage, grade, and multiple mutation statuses. This positions AIGS as a more informative prognostic tool than any single currently used clinical or molecular classifier for bladder cancer.
Head-to-head comparison with published models. A systematic comparison was conducted against 58 previously published mRNA and lncRNA prognostic signatures for bladder cancer, all based on various machine learning algorithms including Lasso, GBM, and Ridge. AIGS demonstrated the highest C-index in the majority of cohorts and the best overall average performance across all eight evaluation cohorts.
While several signatures trained on TCGA data showed higher TCGA-specific C-indices than AIGS, these same signatures performed poorly in external validation cohorts, with C-indices sometimes falling below 0.6, indicating severe overfitting. AIGS maintained consistent performance across all cohorts regardless of their size, origin, or imaging protocols, demonstrating superior generalizability.
Nomogram combining AIGS and AJCC stage. A prognostic nomogram integrating AIGS and AJCC clinical stage achieved AUCs of 0.774, 0.752, and 0.778 at 1, 3, and 5 years in TCGA-BLCA, outperforming either predictor alone. Calibration plots confirmed the nomogram's superior individual prognosis prediction capability, suggesting the two variables provide complementary information for risk assessment.
Low AIGS scores associated with better treatment response. In two independent immunotherapy cohorts (IMvigor210, n=298 and GSE91061, n=39), patients with low AIGS scores had significantly better immune checkpoint inhibitor response (p below 0.01 in both). The same pattern was observed in the chemotherapy cohort GSE52219 (p below 0.05), with low AIGS score patients responding better to chemotherapy as well.
Multiple independent immunotherapy response prediction algorithms confirmed this relationship: low AIGS patients showed higher immune response rates by TIDE analysis, higher immunophenoscore (IPS) values reflecting greater immunogenicity, and superior predicted efficacy for both anti-PD-1 and anti-CTLA-4 immunotherapy by SubMap analysis.
Immune landscape differences explain the pattern. Immune infiltration analysis in the IMvigor210 cohort revealed that low AIGS patients had significantly higher CD8+ T cell infiltration and dendritic cell abundance, both of which are associated with effective anti-tumor immune responses. Conversely, high AIGS patients showed greater macrophage and T helper 2 cell infiltration, cell types associated with immune suppression and poor immunotherapy response.
Immune checkpoint expression analysis identified CTLA-4, CD274 (PD-L1), TMIGD2, CD40, and SIGLEC15 as significantly upregulated in low AIGS patients, suggesting these checkpoints as potential immunotherapy targets, while VTCN1, CD70, and TNFRSF18 were downregulated in the low AIGS group.
AIGS correlates with specific mutation patterns. Among the top 30 frequently mutated genes in bladder cancer, seven showed significantly different mutation frequencies between high and low AIGS groups: TP53, TTN, RB1, FGFR3, ELF3, SPTAN1, and NEB. Multivariate logistic regression confirmed that AIGS remained independently associated with TP53, RB1, FGFR3, and ELF3 mutations after adjusting for age, gender, stage, and tumor mutation burden.
Copy number variation analysis identified specific chromosomal amplifications and deletions differentially distributed between AIGS risk groups. Key alterations included amplification of 3p25.2, 3q26.33, 7p21.1, and 7p11.2, and deletion of 22q13.32, 4q34.2, and 9q22.33. AIGS independently predicted amplification of 3p25.2 and 7p11.2 and deletions of 22q13.32, 4q34.2, and 9q22.33 after multivariate adjustment.
RTP4 identified as methylation-driven AIGS target. Among 55 methylation-driven genes identified (where methylation negatively correlated with expression), RTP4 showed the strongest and most significant relationship with AIGS: high AIGS score patients exhibited higher RTP4 promoter methylation and correspondingly lower RTP4 gene expression, identifying RTP4 as a potential epigenetic mediator of the high-risk AIGS biology.
GSEA and GSVA analyses revealed that AIGS-associated genes are enriched in cancer-relevant pathways including DNA replication, chromosome segregation, cell cycle regulation, mismatch repair, ECM receptor interaction, and epidermal cell differentiation, providing mechanistic context for the signature's prognostic power.
Drug sensitivity predicted from gene expression. Using gene expression profiles from cancer cell lines in the CTRP (266 compounds) and PRISM (1,285 compounds) databases, drug sensitivity was predicted for bladder cancer patients by correlation of cell line expression profiles with patient tumor expression. The prediction approach was validated by confirming that low BRCA1 expression (a known cisplatin sensitivity marker) predicted lower cisplatin AUC (greater sensitivity) as expected from clinical data.
Seven compounds were identified that showed both significantly lower predicted AUC values (greater sensitivity) and negative correlation between AUC and AIGS score, indicating they may be preferentially effective in high AIGS score patients who respond poorly to standard immunotherapy. These candidates are: BRD-K45681478, 1S,3R-RSL-3, RITA, U-0126, temsirolimus, MRS-1220, and LY2784544.
Therapeutic targeting of the high-risk group. The identification of seven potential therapeutic agents specific to high AIGS patients addresses a critical clinical gap: patients who are unlikely to respond to immunotherapy need alternative treatment strategies. These seven compounds represent testable hypotheses for clinical investigation, spanning diverse mechanisms including ferroptosis induction (RSL-3), MDM2 inhibition (RITA), MEK inhibition (U-0126), and mTOR inhibition (temsirolimus).
A comprehensive platform for bladder cancer risk stratification. AIGS provides an integrated platform that simultaneously predicts overall survival, recurrence-free survival, progression-free survival, immunotherapy response, and chemotherapy response in bladder cancer patients, outperforming all 58 compared published signatures and all individual clinical and molecular variables across multiple independent cohorts.
The combination of AIGS with AJCC staging in a nomogram improves upon either alone, suggesting that AI-driven molecular signatures and traditional clinical staging provide complementary information that together enable more precise individual risk prediction than either currently available tool.
Toward actionable clinical implementation. AIGS divides patients into two biologically and clinically distinct groups: low-risk patients who are likely to respond to immunotherapy and standard chemotherapy, and high-risk patients who need alternative treatment strategies, for whom seven potential therapeutic agents have been identified. Multi-omics characterization of the two groups provides mechanistic insights that could guide future drug development and clinical trial design.
Future prospective validation studies and functional characterization of the AIGS genes, particularly the methylation-driven RTP4 target, are needed before clinical implementation. However, the rigorous multi-cohort validation framework and comprehensive benchmark against published literature establish AIGS as the most robustly validated bladder cancer prognostic signature to date.