Construction of a novel mRNA-signature prediction model for prognosis of bladder cancer based on a statistical analysis

Our study applied public datasets to conduct analysis based on the Cancer Genome Atlas (TCGA, https://​portal.​gdc.​cancer.​gov/​) and Gene Expression Omnibus (GEO, https://​www.​ncbi.​nlm.​nih.​gov/​geo/​). Gene expression data and corresponding clinical data were obtained from TCGA website via “gdc-client” tool. GEO datasets included GSE13507 [18] and GSE133624 [19].

The gene expression data of GSE13507 was conducted the DEGs analysis via R package “limma” [20]. For gene expression data of GSE133624 and TCGA-BLCA, the DEGs analysis was conducted via R package “DESeq2” [21]. After these analyses, the downregulated or upregulated gene was defined with adjust P value < 0.05, the |log2 fold change| > 1 [12, 22]. The shared DEGs among three datasets were showed in Venn diagram by R package “VennDiagram” [23].

In this part, we excluded the samples without corresponding survival data and clinical data. Then we divided bladder cancer data of TCGA into training set and testing set randomly. The information about training set and testing set is shown in Table 1. The univariate Cox analysis was utilized to select prognostic genes via R package “survival” in training set [24], which were obtained with the threshold of P < 0.05. The overlapping candidate genes (OCGs) were obtained by intersection analysis between prognostic genes and shared DEGs.

We utilized LASSO regression with tenfold cross-validation to narrow down OCGs by R package “glmnet” [25]. A gene-based prognostic signature was constructed via stepwise multivariate Cox regression. Risk score based on gene prognostic signature was calculated for each TCGA-BLCA patient via gene expression multiplied by the regression coefficient in stepwise multivariate Cox regression.

The testing set (n = 162) and the whole set (n = 405) were utilized to assess the predictive validity of the multi-gene prognostic signature. In the validation set, the risk score of each patient was calculated via the coefficient of the candidate genes obtained above. Then the patients were stratified into high-risk and low-risk groups based on the median risk score as the cutoff. The Kaplan-Meier (KM) survival analysis with log-rank test and time-dependent receiver operating characteristic (ROC) analysis was applied to validate the gene-based prognostic signature. Furthermore, the mutation type of selected genes was explored in cBioPortal (https://​www.​cbioportal.​org/​) [26, 27].

Based on risk score and some clinical parameters, a nomogram was established to predict the probability of 1-year, 3-year, and 5-year OS using R package “rms” [28]. The score of the prediction of nomograms for each patient was calculated via R package “nomogramFormular” [29]. With the source code provided on the MSKCC website(https://​www.​mskcc.​org/​departments/​epidemiology-biostatistics/​biostatistics/​decision-curve-analysis), we performed the DCA analysis of survival outcome [6]. The calibration curve analysis was conducted via “calibrate” function of “rms” R package [28]. The time-dependent ROC analysis for nomogram score was performed via R package “timeROC” [30].

Gene ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways analysis of candidate genes were performed by R package “clusterProfiler” [31, 32]. The threshold for analysis was set P-value < 0.05, indicating significantly enriched functional annotations. Bioinformatic methods “guilt by association” (GBA) [17] and GSEA were applied to conduct potential functional analysis. GSEA was conducted in R with R package “clusterProfiler” [32]. GBA was performed with Spearman method.

The samples in TCGA were randomly divided into training set and testing set with “sample” function of R. Heatmap of DEGs obtained in three datasets were plotted with R package “pheatmap” [33]. Two groups of boxplots were analyzed with Wilcoxon-test. The comparison of clinical parameters between training set and testing set was conducted with χ² test or exact Fisher test. As for KM survival analysis, P-value and hazard ratio (HR) was generated via log-rank tests and univariate Cox proportional hazards regression. All analysis above and R packages were performed in R software version 3.6.3 (The R Foundation for Statistical Computing, 2020). All statistical tests were two-sided. P < 0.05 was regarded as statically.

The flowchart of this study is shown in Fig. 1. According to the differential gene selection criteria for differential analysis, 2606 up-regulated genes and 2046 down-regulated differential genes were screened in the TCGA-BLCA (Fig. 2 A, D). 293 up-regulated genes and 697 down-regulated genes were screened in GSE13507 (Fig. 2 B, E). 1984 up-regulated genes and 546 down-regulated genes were screened in GSE133624 (Fig. 2 C, F). Taking the intersection of the up-regulated and down-regulated genes in the three data sets, 151 up-regulated genes and 143 down-regulated genes were obtained (Fig. 2 G, H).

Based on the univariate Cox analysis, 808 prognosis-related genes (HR > 1) and 879 prognosis-related genes (HR < 1) were screened in training set. Then the prognostic genes were intersected with 294 DEGs. Finally, 20 shared genes were obtained, which includes 7 DEGs with HR < 1 and 13 DEGs with HR > 1 (Fig. 2 I).

LASSO regression with tenfold cross-validation was conducted to get the optimal lambda value from the minimum partial likelihood deviance (λ_min = 0.03522) (Fig. 3 A, B) [12]. Prognostic DEGs were narrowed down to an eight-gene signature. The correlation analysis showed that the expression of these eight genes was not significant (Fig. 3 C), which means the signature based on these genes was not overfitting. Six candidate genes (SORBS2, GPC2, SETBP1, FGF11, APOL1, H1–2) were selected via stepwise multivariate Cox regression (Fig. 3 D, E). Then six-gene-based signature (Risk score = 0.11061*Exp _(SORBS2) - 0.18866*Exp _(GPC2) + 0.24538*Exp _(SETBP1) + 0.38858*Exp _(FGF11) – 0.16433*Exp _(APOL1) -0.23161*Exp _(H1–2)) was constructed.

The patients were divided into the high-risk group and low-risk group based on the median risk score as cutoff (Fig. 4 A). Figure 4 B showed the survival status of the patients. And Fig. 4 C showed the heatmap of six prognosis-related genes. The KM survival analysis for the training set showed that the high-risk group had a worse OS compared with the low-risk group (Fig. 4 D). The AUC of 1-year, 3-year, 5-year for training set were 0.635, 0.732, 0.737, respectively (Fig. 4 E).

The expression of six genes in tumor and normal tissue were shown in Fig. 5 A-C, which indicated that GPC2, FGF11, APOL1, and H1–2 were highly expressed in tumor tissue, while SORBS2 and SETBP1 were highly expressed in normal tissue. The expression of six genes in TCGA-BLCA database indicated that SORBS2, SETBP1, APOL1 were differently expressed in different stages. SORBS2 and SETBP1 were significantly up-regulated in stage IV (Fig. 5 D, E). APOL1 was significantly up-regulated in stage I/II (Fig. 5 F). Therefore, these three genes may be associated with the pathological stage of BC. The genetic alteration type of six genes was analyzed in the cBioPortal database (Fig. 5 G). KM survival analysis for each prognostic gene showed that the expression of GPC2, SETBP1, FGF11, APOL1, H1–2, PDGFD were significantly correlated with OS of patients with BC (Fig. 6). Moreover, the immunohistochemistry from Human Protein Atlas database (Supplementary Fig. S1) showed that APOL1, GPC2, and H1–2 had a higher protein level in urothelial cancer, while SETBP1 had a higher protein level in the urinary bladder. This result was consistent with the mRNA analysis in TCGA database.

The survival analysis in different subgroups showed that the risk score had a satisfactory performance. The group of age (Fig. 7A-B), gender (Fig. 7C-D), race (Fig. 7E-F), AJCC-N (Fig. 7 L-M), AJCC-M (Fig. 7N-O) indicated that the patients with high-risk score had significantly worse OS. In the group of AJCC-stage, the patients with the high-risk score in the early stage did not have significantly worse OS (Fig. 7G), while in stage III and stage IV, the patients with the high-risk score have significantly worse OS (Fig. 7 H-I). In the group of AJCC-T, there was no significantly different OS between high- and low-risk score in the T0/1/2 group (Fig. 7J), while in the T3/4 group, the patients with the high-risk score have worse OS (Fig. 7K). Then the risk score of each patient in testing set and entire set was calculated via the coefficient and gene expression of six genes. Then the KM survival analysis was conducted in the entire set, testing set, and external dataset GSE13507 (Fig. 8 A, B, C). Time-dependent ROC analysis was performed to evaluate risk score in the entire set, testing set, and external dataset GSE13507 (Fig. 8 C, D, E). In summary, the six-gene prognostic signature is a reasonably adequate OS predictor for BC patients.

The six-gene prognosis-related signature with other clinical parameters, such as age, gender, AJCC pathological stage, was performed to construct a nomogram to predict 1-year, 3-year, 5-year OS of patients with BC (Fig. 9 A). Considering the accuracy of Cox proportional hazard model, the age and the pathological stage were set polytomous variables in the construction of nomogram [34]. The calibration plot for patient survival prediction suggested that the predicted outcome of the six-gene prognostic nomogram showed consistency with the actual outcome (Fig. 9 B-D). DCA indicated that utilizing the nomogram gained more benefit than utilizing the stage model and TNM model when the threshold probabilities were set more than 0.25 (Fig. 9 E-G). The time-dependent ROC curves indicated that the nomogram had higher predictive accuracy than the stage model and risk score model. The AUC of nomogram time-dependent ROC was 0.763, 0.805, and 0.806 for 1-year, 3-year, and 5-year, respectively (Fig. 9 H-J).

GO and KEGG enrichment analysis was utilized to explore the biological function of genes correlated with candidate genes. In GO biological analysis, these genes were enriched in extracellular structure organization, extracellular matrix organization, second messenger mediated signaling et al. (Fig. 10 A, Additional file: Table S1). In KEGG pathway analysis, PI3K-Akt signaling pathway, Calcium signaling pathway, cGMP-PKG signaling pathway, ECM-receptor interaction, et al. were identified for genes correlated with candidate genes (Fig. 10 B, Additional file: Table S2).

GSEA was performed to identify the potential biological process of 6 prognosis-related genes. Results suggested that the samples with high expression of SORBS2, SETBP1 were enriched in epithelial-mesenchymal transition (Fig. 10 C, E). The samples with low expression of GPC2, H1–2 were enriched in interferon-alpha response (Fig. 10 D, I). While the samples with high expression of FGF11 were enriched in PI3K-AKT-MTOR signaling and hypoxia (Fig. 10 F). And the samples with high expression of APOL1 enriched in P53 pathway, interferon-alpha response (Fig. 10 G).