Prediction of non-muscle invasive bladder cancer outcomes assessed by innovative multimarker prognostic models

Here we present a high dimensional model considering the time-to-event nature of the information and censored data enabling to accommodate a large number of variables in a relatively small number of individuals. To our knowledge, this is the first time that such a model is applied in the clinical and genetic epidemiology fields. More specifically, we have applied it to study the predictive ability of prognostic models for NMIBC patients.

The major goal in managing NMIBC patients is to prevent tumor relapse, this including both the high number of recurrences and the progression to MIBC. To this end, treatment needs to be tailored according to the aggressiveness of the disease. Therefore, accurate prognostic models are crucial. Currently, there are no validated prognostic molecular biomarkers to guide the clinical management of patients [22, 23] and the therapeutic decisions are still based on risk tables only including clinico-pathological prognosticators [3]. Here we have investigated the potential clinical utility of inherited genetic markers (SNP profiles) based on their robustness and precise measurements as well as on their time-independent nature in comparison to serological and histological markers. To this end we have assessed the ability to improve TFR and TP risk stratification in NMIBC patients of genome-wide common SNPs profiles. We have also evaluated the performance of well-known clinico-pathological prognosticators and how much the whole genome approach improved their performance to better classify patients.

Regarding the classification performance of clinico-pathological prognosticators alone, our sequential threshold models for both TFR and TP got similar estimates to those obtained previously by us with a Cox proportional hazard regression analysis [11]. Discrimination of patients according to the risk of TFR using clinico-pathological prognosticators was poorer than previously reported by Hernandez et al [24] (0.62 vs. 0.75), although better than that reported by Vedder et al [25] in a large cohort including ours. Nevertheless, it is worth noting that the definition of the outcome differs (recurrence vs. first recurrence) between our and these studies [24, 25]. Regarding TP outcome, our clinico-pathological prognosticators model classified the patients better than in Hernandez et al [24] (0.76 vs. 0.54) and than in a Danish cohort using both EORTC (0.76 vs. 0.72) and CUETO (0.76 vs. 0.74) scores [25]. However, it performed worse than in a Dutch cohort using the same classifiers: EORTC (0.76 vs. 0.81 and 0.77) and CUETO scores (0.76 vs. 0.82 and 0.81) [25].

The prediction ability of clinico-pathological prognosticators depends on the outcome. They clearly perform better in predicting TP than TFR, both in terms of classification (AUC, 0.76 vs. 0.62) and proportion of the explained variance (R _probit ² , 5.4 % vs. 3.1 %). Their lower performance when predicting TFR could be due to the dependence of factors other than biological explanations such as the potential incomplete resection of the tumor during the TURB and the tumour cell reimplantation on first tumour recurrence [23], factors that are difficult to be assessed and therefore are not accounted for in the model. When the patients were stratified according to their risk status, clinico-pathological prognosticators explained a larger proportion of the phenotypic variance (~15 %) in the HiR group than in the LR NMIBC, probably because these factors were specifically selected to identify patients with HiR tumors with a high potential of progression. However, the overall classification performance of HiR NMIBC patients was poorer (AUC = 0.57) than in the whole cohort. While the discriminatory ability of clinical-pathological parameters for both NMIBC outcomes is valuable, there is room for improvement. More accurate discriminatory models would better select patients for aggressive treatment as well as would avoid unnecessary treatments towards a better patient management. This justifies the search of further prognostic factors, among them tumour molecular alteration and inherited variation markers [3, 26, 27].

Our results showed that common genome-wide SNPs similarly, though poorly, classified patients regarding both TFR and TP in the whole series and in the HiR and LR subcohorts, AUCs ranging from 0.55 to 0.58. Adding SNP to the models did not improve the classification performance of clinico-pathological prognosticators although improvements of R _probit ² were achieved for TFR (3–4) and TP in the HiR cohort (15.1 - 15.5 %). Surprisingly, adding SNP to clinico-pathological prognosticators worsened the percentage of phenotypic variance (R _probit ² ) explained by the model with clinico-pathological prognosticators only by 7 and 25 % when predicting TP in the whole and the LR-NMIBC cohorts, respectively. The little improvement or even deterioration in terms of R _probit ² could be explained by a correlation between the prediction of clinico-pathological prognosticators and that of SNPs. To confirm this, we calculated the R _probit ² of a model with \( \mathbf{X}\widehat{\boldsymbol{\upbeta}} \)obtained from clinico-pathological prognosticators only as dependent variable and the SNPs as independent variables (see Tables 2 and Additional file 1: Table S6). The proportion of the clinico-pathological prognosticators prediction variances of TFR and TP explained by SNPs was larger than that of the TFR and TP phenotypic variances. The calculation of R _probit ² allowed us to report the first ĥ ² for TFR and TP in the whole series and in the HiR and LR subcohorts. The largest ĥ ² corresponded to TFR (1 %) and to TP of patients at HiR (1 %), although they may be underestimated because of the sample size and the limitation on the number of SNPs included in the model [28]. All the above explains the small or nil contribution of the SNPs to the predictive ability of clinico-pathological prognosticators of the phenotypes of interest. The poor predictive ability of common SNPs in NMIBC prognosis is in line with a previous study reporting low GWAS risk predictive values for UBC [19], as well as with those obtained in studies predicting risk for other neoplasms, such as breast cancer [29, 30]. The different results obtained with AUC and R _probit ² can be explained by the different scales in which the predictions are expressed (observable for AUC and liability for R _probit ² ), their non-monotonic relationship, and the lower number of events, especially when the individuals were stratified.

While this is one of the largest and well-characterized NMIBC cohort worldwide, the restricted sample size in the subgroup analyses is one of the limitations we face here because the small number of events limits the prediction accuracy of the genomic profile achieved with the SNPs. This is even clearer when patients were further stratified as LR-NMIBC. Although increasing sample size of the study would be desirable, heterogeneity across studies regarding patient recruitment, pathological classifications applied, and treatment or patient management would increase random misclassification and, therefore, would dilute estimates. While we conducted a genome-wide exploration, the models did not include all genotyped SNPs (1 million) but a subset that were filtered by a restrict LD. When we applied a less restrictive LD threshold (r ²  < 0.8) and considered a larger number of common SNPs neither the classification performance nor the percentage of the phenotypic variance explained improved (results not shown). Including in the models both rare and structural variants may help in further characterizing and increase the precision of the predictive estimates. Application of other statistical modeling approaches could indeed yield improvements in the predictive power, for example by considering non-additive models that include epistatic interactions between SNPs or adding functional information in the model. Exploring the integration of other –omics data such as microRNAs, as well as considering possible interactions between treatment and variants could also help in this regard.

This study also presents several strengths as its population-based nature, detailed medical information, long follow-up, and centralized pathological review decreasing heterogeneity of the covariates stage and grade. The use of state-of-the art methodology applied here allowed to handle a highly dimensional problem and time-to-event data, as well as censoring. The application of such methodology allowed us to provide the first estimates of heritability for UBC outcomes.