Identification and validation of a novel signature for prediction the prognosis and immunotherapy benefit in bladder cancer

Study design and data collection

Data were collected as previously described in Zhang et al. (2019). Specifically, the inclusion criteria of GEO datasets were as follows: (1) biospecimens were gained from patients with localized BC; (2) enrolling at least 5 pairs samples in each dataset; (3) only including transcriptomic data in each dataset. (4) containing both clinical features (clinical tumor stage (TNM) or molecular subtype) and survival outcomes (OS or PFS). The exclusion criteria were as follows: (1) duplicates of the previous eligible datasets; (2) not histo-pathologically confirmed urothelial carcinoma. Gene expression profile data (GSE37815, GSE13507, GSE121711, GSE40355, and GSE3167) were downloaded from the public Gene Expression Omnibus database (GEO, http://www.ncbi.nlm.nih.gov/geo/). Corresponding clinical information for GSE13507 was also obtained. Annotation information for the datasets and the platforms is shown in Table S1. The level 3 RNA-sequencing (RNA-seq) data (Fragments per kilobase million, FPKM) with the corresponding clinical information from 430 BLCA (Bladder Urothelial Carcinoma) samples (19 normal samples and 411 tumor samples) were downloaded from The Cancer Genome Atlas dataset (TCGA) (https://portal.gdc.cancer.gov/). The FPKM value was converted to the value per million transcripts (TPM) to make RNA-seq data more comparable with microarray data. The ENSEMBL IDs in RNA-seq data and the probes in microarray data were converted to gene symbol IDs using the annotation files from GENCODE (https://www.gencodegenes.org/). The IMvigor210 cohort was downloaded from http://research-pub.gene.com/IMvigor210CoreBiologies/#transcriptome-wide-gene-expression-data. It was a cohort evaluating the effect of atezolizumab (PD-L1 blockade) in patients with locally advanced or metastatic urothelial BC. A flowchart of the analysis performed in this study is shown in Fig. 1.

Data processing and screening for differentially expressed genes

To identify DEGs in the GEO datasets, the raw microarray data was converted to transcripts per million. The robust rank aggregation algorithm (RRA) in the “affy” package was used for background adjustment, log2 transformation and normalization. Then the RRA method was conducted to integrate the multiple-rank gene list of the five GEO datasets. The “edgeR” package was used to screen the DEGs from the TCGA dataset (|log2FC| > 1, adjusted P < 0.05). Subsequently, we obtained the intersection of the DEGs from the two datasets by using Venn diagram tool (http://genevenn.sourceforge.net/).

Identification and estimation of the prognostic multi-gene signature

After obtaining the DEGs in the previous step, univariate Cox regression analysis was conducted to determine which gene was significantly correlated with patients’ OS (P < 0.05). Lasso-penalized Cox regression analysis was employed to further minimize the numbers of DEGs. The role of Lasso is to add a constraint condition to the sum of the absolute values of the coefficients, in order to reduce data dimensionality and interference and obtain better fitting. For high-throughput gene expression data with high-dimensional latitude and strong correlations, the Lasso method is practical after compressing meaningless explanatory variables to zero (Gui & Li, 2005). The regression coefficients of eight prognosis genes were derived from the stepwise multivariate Cox regression model, and then were used to calculate risk scores. All samples in the training (TCGA-BLCA) sets were divided into high- or low-risk groups by the cutoff values calculated by X-tile software. The Kaplan-Meier survival curves were performed to compare the survival risk between the two groups. Receiver operating characteristic (ROC) curves were conducted to demonstrate the predictive accuracy of this signature.

To determine the predictive power of the prognostic model and other clinical parameters including age, gender, T-stage, N-stage, M-stage, molecular subtype and pathologic stage, univariate and multivariate Cox analyses were performed. The information of six molecular subtypes (luminal papillary, luminal non-specified, luminal unstable, stroma-rich, basal/squamous, and neuroendocrine-like) were obtained from the Supplemental Material of the article (Kamoun et al., 2020). Parameters with P < 0.05 in the univariate analysis were incorporated to conduct the multivariate analysis. Kaplan–Meier analysis was performed to determine whether the effect value was consistent among different subgroups.

Validation of prognostic multi‑gene signature

In order to validate the predictive capability of this prognostic model, the GSE13507 database was used for external validation. The risk scores were established using the prognostic gene signature. Patients were grouped by the median risk scores using the same method as above. Kaplan–Meier analysis and ROC analysis were generated to evaluate the power of the model. The protein expression levels of the eight genes in normal and BC tissues were validated in the Human Protein Atlas database (https://www.proteinatlas.org/).

Construction and validation of predictive nomogram

The independent clinicopathological parameters identified in multivariate analysis in the previous step were selected to develop a nomogram to predict the 1-, 3-, and 5-year overall survival (OS) of BLCA patients. The AUC of the ROC curve and concordance index were constructed to evaluate the predictive power of this prognostic model, and calibration plot was conducted to compare the accuracy of nomogram-predicted probabilities with actual observation. Decision curve analysis (DCA) was also used to determine the clinical utility of the nomogram model.

Functional enrichment analyses

To reveal underlying functions of the prognostic gene signature in BC, GO and KEGG analyses were used with a false discovery rate (FDR) < 0.05 considered statistically significant. Patients from TCGA cohort were classed into two groups based on the risk scores as described previously. GSEA was performed to explore the potential molecular mechanisms enriched in the gene signature (FDR < 5%, nominal P < 1% and |NES| > 1).

Tumor immunity analyses

To systematically illustrate the relationship between immune infiltrating cell phenotype and BC survival, we applied the CIBERSORT algorithm and assessed the relative proportions of 22 distinct leukocyte subsets (Newman et al., 2015). Only the CIBERSORT samples with P < 0.05 were filtered and chosen for further analysis. After filtering the data, the relevant violin plot and correlation heatmap were displayed by R package. The correlation between immune infiltration level and expression of each gene in the gene signature was explored through TIMER2.0 (http://timer.comp-genomics.org/). Since the accuracy for estimating the proportion of cell components were different by applying for different algorithm methods, we used QUANTISEQ algorithm to estimate the changes in the proportion of CD8+ T cell and Tregs, and EPIC algorithm was used to estimate the changes in the proportion of CD4+ T cell, B cell, NK cell, macrophage, cancer associated fibroblast and endothelial cell (Sturm et al., 2019). To further explore the different infiltration degrees of immune cell types, immune-related functions, and immune-related pathways in two risk groups, single sample gene set enrichment analysis (ssGSEA) was applied by using the R package “GSVA” (Hänzelmann, Castelo & Guinney, 2013). The correlation between gene signature riskscore and immune related molecules was further investigated to better understand immune infiltration in BLCA.

Prediction and evaluation of immunotherapeutic response

Tumor Immune Dysfunction and Exclusion (TIDE) algorithm were performed to predict the potential response to immune checkpoint blockade. Moreover, an independent cohort (IMvigor210) that recorded expression data from patients who responded or did not respond to anti-PD-L1 immunotherapy was used for further verification. Kaplan–Meier analysis was generated to test the prognostic value of gene signature. CIBERSORT and ssGSEA were applied to explore the immune landscape between high- and low-risk group stratified by risk score.

Cell lines and culture conditions

The human BC cell lines (EJ, T24, UM-UC-3, SW780) and a normal human urinary tract epithelial cell line (SV-HUC-1) were purchased from Stem Cell Bank, Chinese Academy of Sciences (Shanghai, China). SV-HUC-1 was maintained in F-12K medium (Cat#11320033; Gibco, Grand Island, NY, USA). UM-UC-3, EJ and SW780 cell lines were routinely cultured in RPMI 1640 medium (Cat#11320033; Gibco, Grand Island, NY, USA), while T24 cell lines was cultivated in DMEM medium (Cat#11965092; Gibco, Grand Island, NY, USA). All the media were supplemented with 10% Fetal Bovine Serum (FBS, Gibco, Cat#10099141) and 1% penicillin/streptomycin (Cat#15070063; Gibco, Grand Island, NY, USA). All cells were incubated at 37 °C with in an atmosphere of 5% CO₂.

Validation of key genes by the quantitative real-time PCR (qRT-PCR) analysis

Total RNA was extracted from cultured cell lines using TRIzol reagent (Cat# 9108; TaKaRa, Dalian, China). The ratio of absorbance at 260 and 280 nm (the A260/280 ratio) was used to evaluate the purity of RNA. Subsequently, cDNAs were synthesized by using the PrimeScrip™ RT reagent Kit (Cat# RR047A; TaKaRa, Dalian, China) from 1 μg of total RNA in 20 μl of reaction volume. 2×TB^® Green qPCR Master Mix (Cat# RR820Q; TaKaRa, Dalian, China) was used for qRT-PCR with Roche Lightcycler 480 RT-PCR System. GAPDH fragment was used as an internal control for normalization of the data before calculation using the 2^−ΔΔCt method. Three independent replicates were conducted for each experiment. The primers used are listed in Table S2.

Statistical analysis

All statistical analyses were conducted using the R software (version 3.6.2; https://www.r-project.org/). Cox regression model with Lasso based on the R package “glmnet” were performed to generate optimal prognostic signature for BC. The Risk score was calculated using this formula: The risk score = ${\sum }_{i=1}^n(Coefi\ast Expi)$, where Expi represents the expression level of gene, i and coefi represents the regression coefficient of gene i in the signature. Kaplan–Meier plots and log-rank tests were used to compare the difference in the survival status between the high- and low- risk groups. The time-dependent receiver operating characteristic curve (ROC) and calculated the area under the curve (AUC) for 1-year, 3-year, and 5-year OS were conducted to validated the prediction ability of the risk signature. Mann–Whitney-Wilcoxon Test or Student’s t test was executed to compare the difference between defined groups for continuous variables. Categorical clinical variables were assessed using Chi-square or Fisher’s exact tests. Differences among BC cell lines and urinary tract epithelial cell line were analyzed by one‑way ANOVA, and the Dunnett test was used as the post hoc test. P < 0.05 was set as the cutoffs for statistical significance.

Identification of robust DEGs

The annotation information for the datasets and platforms was exhibited in Table S1. We identified 306, 780, 4,035, 452, and 2,038 DEGs between normal and tumor tissues in the GEO datasets GSE3167, GSE37815, GSE40355, GSE13507, and GSE121711, respectively. Volcano plots displayed the distribution of DEGs in each dataset (Figs. 2A–2E). Based on the results of robust rank aggregation method 403 DEGs were screened out, including 132 upregulated and 271 downregulated genes. The top 20 up-regulated and down-regulated DEGs were listed in Fig. 2F. A total of 2,568 DEGs were selected from the TCGA-BLCA dataset, including 1,415 up-regulated and 1,153 down-regulated genes. By using Venn diagram web tool we chose a total of 302 DEGs for subsequent analysis in the two cohorts.

Identification and validation of eight-gene prognostic signature

Univariate Cox regression modeling revealed 74-OS related genes in BLCA patients (P < 0.05). In the training set (TCGA-BLCA), Lasso-penalized Cox analysis was applied to further analyze the mRNAs data, and 19 genes were identified. The aim of backward stepwise regression was to construct a minimum set of independent variables by including all variables and deleting one variable at a time, to test which was least statistically significant. Model with the highest determination coefficient was built on the remaining variables (Kuhn & Johnson, 2013). 8-prognostic-gene model was finally selected by stepwise regression analysis: CNKSR, COPZ2, CXorf57, FASN, PCOLCE2, RGS1, SPINT1 and TPST1.

The risk scores were calculated for all patient according to the following formula: (0.0289 × exp_COPZ2) – (0.0573 × exp_CNKSR1) – (0.0602 × exp_CXorf57) + (0.00678 × exp_FASN) + (0.0297 × exp_PCOLCE2) – (0.0353 × exp_RGS1) + (0.0032 × exp_SPINT1) + (0.04 × exp_TPST1).

In the training dataset, 394 patients were separated into high- and low-risk groups using the median risk score calculated by X-tile software. The low-risk group showed significant better overall survival compared with the high-risk group (P < 0.001; Fig. 3A). The AUC values of the ROC for 1-, 3-, and 5-year OS were 0.771, 0.735, and 0.718, respectively (Fig. 3C). Subgroup analyses stratified by age, gender, AJCC stage and molecular subtype were conducted to evaluate the prognostic values of the eight-gene signature in different subtypes. The patients with high riskscores had worse OS than the patients with low riskscores in age < 65 (P < 0.0024), age > 65 (P < 0.0001), female (P = 0.00011), male (P < 0.0001), stage III + IV (P < 0.0001), molecular subtype of basal/squamous (P = 0.0019), molecular subtype of luminal papillary (P = 0.00017) and molecular subtype of luminal non-specified (P = 0.043) (Fig. S1).

Eight-gene signature is a prognostic factor independent of other clinicopathological parameters

In line with the results in the training dataset, the eight-gene signature could successfully stratify the samples in the validation dataset into different risk groups (Fig. 3B). The AUC values of the ROC for 1-, 3-, and 5-year OS were 0.677, 0.694, 0.676, respectively (Fig. 3D). Cox regression analyses were processed to access whether the eight-gene signature was an independent variable of BC patient survival. Univariate Cox regression revealed that 8-gene risk score, age, pathologic stage, T stage, and N stage were correlated with OS in BC (Fig. 3E). After the multivariate Cox analysis, risk score, age, and pathologic stage remained as independent prognostic parameters (Fig. 3F). The expression of CNKSR1, CXorf57 and FASN at the protein level were significantly elevated in BC tissues compared with non-cancerous tissues, while the level of COPZ2, PCOLCE2, TPST1 and RGS1 were notably decreased in BC tissues than in normal tissues. No difference was found for SPINT1 protein expression (Fig. S2). In addition, high risk group was correlated with a higher histological grade, T stage, N stage, M stage and clinical stage as shown in Table 1.

Construction and validation of the predictive nomogram

In order to predict the prognosis of BLCA patients, we integrated the independent prognostic parameters, including 8-gene risk score, age and AJCC stage, to develop a nomogram model (Fig. 4A). The AUCs for the 1-, 3-, and 5-year OS were 0.769, 0.751, and 0.764, respectively (Fig. 4B). The C-index of the nomogram model was 0.703 (95% CI [0.547–0.872]), whereas that for AJCC stage was 0.634 (95% CI [0.451–0.818]), with 1,000 cycles of bootstrapping. Calibration plots showed that the results of predicted OS were consistent with the actual observations (Fig. 4C). DCA was used to evaluate the clinical utility of the nomogram model. The nomogram showed the greatest net benefit when compared with AJCC stage, T stage and risk score alone (Fig. 4D).

GO/KEGG/GSEA

In biological processes, the gene signature was significantly enriched in focal adhesion, microRNAs in cancer, proteoglycans in cancer, cell cycle, and extracellular matrix receptor interaction; these processes have close connections with tumor proliferation and metastasis (Fig. 5A). The KEGG pathway analysis gave similar results (Fig. 5B).

To explore the significantly enriched pathways of the eight prognostic genes GSEA was performed. The results showed the high-risk group was enriched in more T cell suppressive pathways, such as the transforming growth factor β (TGF-β) signaling pathway, Wnt signaling pathway, and calcium signaling pathway (Fig. 5C).

Correlation between tumor immunity and eight-gene signature

The association between risk score and the distribution of tumor-infiltrating immune cells was further investigated in two risk groups. Results showed that significantly higher proportions of M0 macrophages (P = 0.013), M2 macrophage (P = 0.031) and resting mast cells (P = 0.04) were exhibited in the high-risk group while the proportions of CD8⁺ T cells (P < 0.001), CD4 memory-activated T cells (P = 0.001), regulatory T cells (Tregs) (P = 0.003) and follicular helper T cells (P = 0.034) were dramatically lower in the high-risk group (Figs. 6A and 6B). It’s worth noting that the expression levels of eight genes were positively correlated with macrophage and cancer associated fibroblast infiltration (Fig. S3). BLCA samples were successfully divided into two clusters by applying “GSVA” algorithm. The ESTIMATE Score, Immune Score and Stromal Score were significantly higher in low-risk group than that of high-risk group. Also, low risk group had higher infiltration degrees of immune-related functions as well as immune-related pathways compared with high- risk group (Fig. 6C). Interestingly, the risk score was strongly positively correlated with the expression of immune checkpoints such as PD-1, PD-L1, CTLA-4, TIM3, LAG3 and TIGIT (Fig. 6D).

Verification of the correlation between the risk signature and immunotherapeutic response

The CTLA-4 and PD-1/PD-L1 blockade has emerged as a critical weapon in treating cancer (Postow, Callahan & Wolchok, 2015). Thus, we evaluated the eight-gene signature in predicting the sensitivity to ICBs in the two groups through TIDE algorithm and subclass mapping. The results showed that BLCA patients in low-risk group were more likely to response to PD-1 blockade (Bonferroni corrected P = 0.014), and high-risk groups were more likely to response to CTLA-4 blockade (Bonferroni corrected P = 0.041) (Fig. 7A). We investigated the prognostic value of the predictive model in IMvigor210 cohort, which included patients with metastatic urothelial cancer treated with anti-PD-L1 antibody. The results showed low-risk group stratified by the eight gene signature has a significant survival benefit compared with high-risk group (Fig. 7B). The validation cohort also showed a similar immune pattern with training cohort, and the risk score was found to be significantly correlated with immune-related molecules (Figs. 7C–7F).

Distribution of risk score in different BC molecular subtypes

According to different transcriptomic and genomic profiling BC can be classified into six molecular subtypes, each with different clinical prognosis and therapeutic responses to chemotherapy and immunotherapy. We evaluated the distribution of risk score in six BC molecular subtypes, and the results showed the risk score was significantly different in different molecular subtypes (Fig. 8). Ba/Sq tumor with poor prognosis had the highest risk score, but the best prognostic outcome LumP and LumU had the lowest risk score.

Validation of the expression levels of the eight genes in cell lines

The expression signatures at the mRNA level of the 8 genes were then investigated in four human BC cell lines (i.e. EJ, T24, UM-UC-3 and SW780) and demonstrated that the mRNA levels of CNKSR1 were relatively higher in most of the BC cells when compared with normal human urinary tract epithelial cell line (SV-HUC-1), however CXorf57 and FASN were dramatically decreased in most of the BC cells (Figs. S4A–S4H).