An online tool for evaluating diagnostic and prognostic gene expression biomarkers in bladder cancer

We performed a systematic search of the Gene Expression Omnibus (GEO) [ 13] and Array Express [ 9] and identified all entries with the keyword “bladder cancer” whose data was obtained by “transcription profiling by array”. Cohorts were excluded if all patients received neoadjuvant or adjuvant chemotherapy. Of the patient cohorts identified, four cohorts (GSE88, GSE89, GSE7476, GSE27448) were then excluded due to an insufficient sample size for both diagnostic (<10 per group) and prognostic (<10 events) biomarker evaluation (see below). Eleven patient cohorts from GEO and ArrayExpress met our inclusion criteria. Two additional cohorts were identified following a literature search and downloaded from supplemental material to publication [ 14, 15]. In these cohorts, individual patients were excluded from the survival analyses if they were treated by neoadjuvant chemotherapy and excluded completely if their tumors did not have urothelial histology. Individual patients treated with intravesical therapy or adjuvant chemotherapy can be optionally excluded from the prognostic biomarker evaluation analysis (see Prognostic biomarker evaluation and survival analysis). In all, BC-BET contains 13 patient cohorts ( N = 1454) profiled on 10 distinct microarray platforms. Cohorts are named according to the institution or individual responsible for collecting and uploading the dataset, and are summarized in Table  1. All patient samples were collected with approval from an appropriate Institutional Review Board, as documented in the original publications (see Table  1).

For each cohort, the processed gene expression data was downloaded. Three cohorts (Blaveri, AUH-2, and MDA-1) had missing values. Microarray probes with missing values in >20 % of samples were removed and expression values imputed using the impute package ( impute.knn function) in R with default parameters. Probes are matched to genes based on current microarray (e.g., Affymetrix) annotation. When multiple probes exist for a gene, the probe with the highest mean expression is used [ 16]. In cohorts with replicate samples (AUH-2, MSKCC, Stransky-1, and Stransky-2), replicate samples were averaged to produce a single gene expression profile for each patient. In the MDA-1 cohort, because only two individuals had NMI tumors, the two individuals with NMI tumors are excluded from the database. When the same patient was profiled in multiple cohorts, and both cohorts are analyzed (i.e., the gene of interest is profiled in both cohorts), then duplicate patients are removed from one of the cohorts in order to prevent biasing of the results. For details, and for more information about the patient cohorts, sample processing, and patient exclusion, see Additional file 1: Table S1.

For a specified gene, BC-BET evaluates whether the candidate gene is differentially expressed between the following groups which we denote as group A and group B, respectively: tumor and normal samples; high-grade (HG) and low-grade (LG) tumors; and muscle invasive (MI) and non-muscle invasive (NMI) tumors. LG and HG tumors correspond to G1-G2 and G3, respectively, or are classified according to the low- vs. high-grade classification system [ 17, 18]. NMI and MI tumors correspond to Ta-T1 or T2-T4 tumors, respectively [ 19]. Group differences are quantified by fold-change (FC) or by the area under the receiver operating characteristics curve (AUC). FC is calculated using the formula

The AUC quantifies the ability of gene expression to discriminate between two groups and is equivalent to the probability that a randomly selected sample from group A has higher expression than a randomly selected sample from group B. We note that AUC is a better measure of a biomarker’s predictive value, while FC may have a more intuitive biological interpretation. Statistical significance is assessed by either the two-sample t-test (which tests the hypothesis that FC = 1) or the non-parametric Wilcoxon-Rank sum test (which tests the hypothesis that AUC = 0.50).

Endpoints include disease-specific survival (DSS), overall survival (OS), or recurrence-free survival (RFS). The events for these endpoints are death from disease, death from any cause, and disease recurrence for DSS, OS, and RFS, respectively. Because there is no single endpoint consistently available across all cohorts with survival information, a user can opt to use the Best Available endpoint, which is taken to be the first available endpoint in the following order: DSS, OS, and RFS. The endpoints available for each cohort are listed in Table  1. Although selection of the Best Available endpoint results in the selection of different endpoints for different cohorts, patients with MI tumors have significantly ( P < 0.05) poorer outcomes than patients with NMI tumors using the best available endpoint in 5/6 cohorts that include both NMI and MI tumors (Additional file 1: Table S2), which is consistent with RFS and OS analyses in a long-term study of over one thousand patients [ 8]. Patients with HG tumors also have significantly ( P < 0.05) poorer outcomes than patients with LG tumors in all cohorts with more than 10 NMI tumors (Additional file 1: Table S2). Therefore, stage and grade are consistently associated with outcome in these diverse cohorts when the best available endpoint is used.

With the exception of CNUH and DFCI, definitive treatment of all patients was either transurethral resection of the bladder or radical cystectomy. In CNUH, 56 patients with NMI tumors received intravesical Bacillus Calmette-Guerin therapy and 26 patients with MI tumors received cisplatin-based adjuvant chemotherapy (along with one patient with an NMI tumor). Because neither of these treatments are associated with outcome in this cohort (Additional file 2: Figure S1A-B), these patients are included by default in the survival analysis, but optionally can be excluded. In DFCI, there is a confounding between TNM staging and treatment with chemotherapy. In particular, 32 patients with MI tumors receive adjuvant chemotherapy while 1 patient with an NMI tumor receives adjuvant chemotherapy. Of the treated patients having MI tumors, 84 % have nodal involvement or distant metastases, and treated patients have a higher risk of recurrence than non-treated patients (Additional file 2: Figure S1C). The user can optionally remove these treated patients from the survival analysis if desired.

Survival analysis is carried out using the coxph function in R to obtain a hazard ratio (HR) and log-rank p-value based on the cutpoint selected by the user. If the median cutpoint is selected, patients are separated into two groups according to those with high expression (≥ the median) and low expression (< the median), and the HR corresponds to the hazard rate of the high expressers relative to the hazard rate of the low expressers. Alternatively, the continuous expression value of the gene may be used, in which case the HR is relative to a two-fold increase in gene expression.

In order to demonstrate the functionality of BC-BET, we will evaluate the diagnostic and prognostic value of fibroblast growth factor receptor 3 ( FGFR3). Genomic studies in bladder cancer suggest that MI and NMI tumors arise through divergent genomic pathways [ 20] and may arise from distinct progenitor cell types [ 21]. Specifically, activating HRAS and TP53/ RB1 mutations are more prevalent in MI tumors while FGFR3 and KDM6A mutations are more prevalent in NMI and LG tumors [ 22– 26]. Furthermore, FGFR3 mutation status is positively correlated with FGFR3 expression [ 25]. We therefore expect high FGFR3 mRNA expression to be associated with NMI and LG tumors in BC-BET.

A screenshot of the BC-BET homepage is provided in Fig.  1. Here the user selects the gene symbol of interest from a drop down menu of available genes, and sets the statistical parameters for the analysis. For class comparison, we will use FC to measure differential expression across tumor status, grade, and stage, and calculate p-values using the Wilcoxon Rank-Sum test. For survival analysis, we will look at continuous FGFR3 expression using the Best Available endpoint. The user also has the option of including or excluding treated patients from the CNUH and DFCI cohorts. Clicking on the Patient Analysis button carries out the desired evaluation.

Figure  2 contains a screenshot summarizing the diagnostic (Fig.  2a) and prognostic (Fig.  2b) value of FGFR3. The diagnostic value of the gene is summarized and illustrated according to whether the gene is up-regulated (FC > 1) or down-regulated (FC < 1) in group A (i.e., tumor samples, HG tumors, or MI tumors) relative to group B (i.e., normal samples, LG tumors, or NMI tumors, respectively), and whether or not the result is statistically significant ( P < 0.05). The prognostic value of the gene is summarized according to whether the gene is negatively (HR > 1) or positively (HR < 1) associated with survival, and whether or not the result is statistically significant ( P < 0.05). A legend is included on the results page displayed by BC-BET (Fig.  2c).

Clicking on the appropriate View Results radio button provides detailed numeric results indicating the diagnostic value (FC and p-value) and prognostic value (HR and logrank p-value) for each cohort for the chosen analysis. By default, the “Tumor vs. Normal” and “Poor vs. Good Prognosis (NMI + MI patients)” detailed results are shown. The full set of results can be downloaded as an Excel spreadsheet by clicking the Download Results link at the top of the page (Additional file 3: Table S3).

From these results we see that FGFR3 expression is significantly ( P < 0.05) up-regulated in tumor samples (i.e., is higher in tumor samples than in normal samples) in 4/5 cohorts with this information, indicating its value as a diagnostic biomarker. FGFR3 expression is also significantly ( P < 0.05) down-regulated in HG tumors (i.e., is higher in LG tumors than HG tumors) in 8/10 cohorts, and significantly down-regulated in MI tumors in 11/12 cohorts with stage information. However, as a prognostic biomarker, FGFR3 expression is only significantly ( P < 0.05) associated with outcome in 2/6 cohorts, although in all cohorts it is positively associated with survival (HR < 1). In addition, FGFR3 does not have any prognostic value independent of stage and grade, as it is not significantly associated with outcome in patients with LG, NMI tumors or consistently associated with outcome in patients with HG, MI tumors. The above results are consistent with previous studies showing that FGFR3 protein expression is associated with stage and grade in bladder cancer [ 25], and supports investigation into targeting FGFR3 in patients with NMI tumors [ 20]. However, FGFR3 expression has little prognostic value based on the available gene expression cohorts.

BC-BET will be updated periodically with additional gene expression datasets as they become available. In addition, new features including the ability to query the database with a list of genes or by probe, and the option to generate Kaplan-Meier curves are expected to be implemented in future versions of BC-BET.