Artificial intelligence-driven consensus gene signatures for improving bladder cancer clinical outcomes identified by multi-center integration analysis

Abstract

To accurately predict the prognosis and further improve the clinical outcomes of bladder cancer (BLCA), we leveraged large-scale data to develop and validate a robust signature consisting of small gene sets. Ten machine-learning algorithms were enrolled and subsequently transformed into 76 combinations, which were further performed on eight independent cohorts (n = 1218). We ultimately determined a consensus artificial intelligence-derived gene signature (AIGS) with the best performance among 76 model types. In this model, patients with high AIGS showed a higher risk of mortality, recurrence, and disease progression. AIGS is not only independent of traditional clinical traits [(e.g., American Joint Committee on Cancer (AJCC) stage)] and molecular features (e.g., TP53 mutation) but also demonstrated superior performance to these variables. Comparisons with 58 published signatures also indicated that AIGS possessed the best performance. Additionally, the combination of AIGS and AJCC stage could achieve better performance. Patients with low AIGS scores were sensitive to immunotherapy, whereas patients with high AIGS scores might benefit from seven potential therapeutics: BRD-K45681478, 1S,3R-RSL-3, RITA, U-0126, temsirolimus, MRS-1220, and LY2784544. Additionally, some mutations (TP53 and RB1), copy number variations (7p11.2), and a methylation-driven target were characterized by AIGS-related multi-omics alterations. Overall, AIGS provides an attractive platform to optimize decision-making and surveillance protocol for individual BLCA patients.

Keywords: biomarker; bladder cancer; immunotherapy; multi-omics; prognosis.

Fig. 1

The flow chart of this study. CNV, copy number variation; CTRP, cancer therapeutics response portal datasets; OS, overall survival; ssGSEA, single‐sample gene set enrichment analysis.

Fig. 2

Generation of the artificial intelligence‐derived gene signature. (A) Discovery of robust OS‐related genes (RORGs). (B) 30 robust OS‐related genes were identified in eight cohorts (n = 1218). (C) The C‐indices of 76 machine‐learning algorithms in seven validation cohorts (n = 818).

Fig. 3

Survival analysis and performance evaluation of artificial intelligence‐derived gene signature (AIGS). (A–H) Kaplan–Meier survival analysis between the high and low AIGS groups across eight OS cohorts. (I) Kaplan–Meier survival analysis between the high and low AIGS groups in GSE31684 (n = 93). (J) Time‐dependent ROC analysis for predicting OS at 1, 3, and 5 years in TCGA‐BLCA (n = 400), GSE13507 (n = 165), GSE19423 (n = 48), GSE31684 (n = 93), GSE37815 (n = 18), GSE48075 (n = 73), and GSE48276 (n = 73). (K) The C‐indexes of AIGS in TCGA‐BLCA, GSE13507, GSE19423, GSE31684, GSE37815, GSE48075, GSE48276, and IMvigor210 (n = 348); the error bars indicate 95% confidence interval (CI).

Fig. 4

The performance of AIGS was compared with common clinical and molecular variables in predicting prognosis across all training and validation cohorts. The error bars indicate 95% confidence interval (CI). TCGA‐BLCA (n = 400), GSE13507 (n = 165), GSE19423 (n = 48), GSE31684 (n = 93), GSE37815 (n = 18), GSE48075 (n = 73), GSE48276 (n = 73), and IMvigor210 (n = 348). Z‐score test: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. cM, Clinical M stage; cN, Clinical N stage; cT, Clinical T stage; M, M stage; N, N stage; T, T stage.

Fig. 5

Comparisons between AIGS and gene expression signatures. (A) Univariate cox regression analysis of AIGS and 58 published signatures. (B) C‐indices of AIGS and 58 published signatures in TCGA‐BLCA, GSE13507, GSE19423, GSE31684, GSE37815, GSE48075, GSE48276, and IMvigor210. Z‐score test: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. The error bars indicate 95% confidence interval (CI).

Fig. 6

Nomogram construction (A) and performance evaluation (B, C). OS, Overall survival. TCGA‐BLCA (n = 400).

Fig. 7

Immunotherapy prediction and immune landscape with regard to AIGS. (A, B) The relationship between AIGS and immunotherapy response in IMvigor210 (n = 298) and GSE91061 (n = 39); t‐test. (C) The relationship between AIGS and chemotherapy response in GSE52219 (N = 23); t‐test. (D–F) The performance of AIGS in TIDE (D, fisher test), IPS (E, t‐test), and SubMap (F) algorithms. (G) The correlation analysis between AIGS and 24 immune cell infiltration abundance; t‐test. (H) Boxplot of 27 immune checkpoints profiles between high and low AIGS patients; t‐test. The error bars indicate 95% confidence interval (CI). T‐test or Wilcoxon rank‐sum test: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. aDC, Activated dendritic cells; AIGS, Artificial intelligence‐derived gene signature; DC, Dendritic cells; iDC, Immature dendritic cells; IPS, Immunophenoscore; pDC, Plasmacytoid dendritic cells; SubMap, Subclass mapping; Tcm, Central memory T cell; Tem, effector memory T cell; TFH, Follicular helper T cell; tgd, T γ δ cells; TIDE, tumour immune dysfunction and exclusion; Treg, regulatory T cells.

Fig. 8

Multi‐omics analysis based on mutation, copy number variations (CNVs) and methylation. (A) The mutational landscape of the top 30 frequently mutated genes (FMGs). (B) The mutation frequency of 30 frequently mutated genes (FMGs) between the high and low AIGS groups; Chisq test. (C) Identification of independent AIGS‐associated mutations by univariate and multivariate logistic regression analysis. (D) The CNV landscape of the top 15 AMP and Homdel chromosome fragments between high and low AIGS patients. (E) Determination of independent AIGS‐related CNV chromosome segment through univariate and multivariate logistic regression analysis. (F, G) Methylation level and expression difference of RTP4 between high and low AIGS groups; t‐test. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. The error bars indicate 95% confidence interval (CI). TCGA‐BLCA (n = 400).

Fig. 9

Identification of candidate agents with higher drug sensitivity in high‐AIGS score patients. (A) Schematic outlining the strategy to identify agents with higher drug sensitivity in high AIGS score patients. (B) Comparison of estimated cisplatin's sensitivity between high and low BRCA1 expression groups; t‐test. (C) Potential therapeutic compounds for high‐AIGS BLCA patients; Wilcox test. The error bars indicate 95% confidence interval (CI). *P < 0.05, ***P < 0.001. AIGS, Artificial intelligence‐derived gene signature.