A 20-gene model for molecular nodal staging of bladder cancer: development and prospective assessment

Abstract

Background: Neoadjuvant chemotherapy before cystectomy confers a survival benefit in bladder cancer, but it has not been widely adopted since most patients do not benefit and we are at present unable to predict those that do. Since the most important predictor of recurrence after cystectomy is pathologically positive nodes, our aim was to assess techniques that define this stage for the selection of patients for neoadjuvant chemotherapy.

Methods: We developed a gene expression model (GEM) to predict the pathological node status in primary tumour tissue from three independent cohorts of patients who were clinically node negative. From a subset of transcripts detected faithfully by microarrays from both paired frozen and formalin-fixed tissues (32 pairs), we developed both the GEM and cutoffs that identified patient strata with raised risk of nodal involvement by use of two separate training cohorts (90 and 66 patients). We then assessed the GEM and cutoffs to predict node-positive disease in tissues from a phase 3 trial cohort (AUO-AB-05/95; 185 patients).

Findings: We developed a 20-gene GEM with an area under the curve of 0·67 (95% CI 0·60-0·75) for prediction of nodal disease at cystectomy in AUO-AB-05/95. The cutoff system identified patients with high relative risk (1·74, 95% CI 1·03-2·93) and low relative risk (0·70, 95% CI 0·51-0·96) of node-positive disease. Multivariate logistic regression showed the GEM predictor was independent of age, sex, pathological stage, and lymphovascular space invasion (coefficient 9·81, 95% CI 1·64-18·00; p=0·019).

Interpretation: Selecting patients for neoadjuvant chemotherapy on the basis of risk of node-positive disease has the potential to benefit high-risk patients while sparing other patients toxic effects and delay to cystectomy.

Funding: US National Cancer Institute (R01CA143971).

Figure 1. Development of a GEM predictor of pathological nodal status at cystectomy

A. High fidelity transcript discovery: For ease of implementation, we first used Affymetrix HG-U133 plus 2·0 microrarray data for a cohort of 32 paired tissues that had been preserved by formalin fixation, paraffin embedding (FFPE) and fresh freezing (FF) to develop a set of probesets detected with high-fidelity by either means of tissue preservation. After 1000-fold bootstrapping the correlation of probes across the paired tissues, we selected probes maintaining positive correlation at the 2·5^th percentile. 12,402 of the high-fidelity probes were common to the U133 plus 2·0 platform and the U133A platforms, and after ensuring these genes were not expressed differently in TUR specimens, these were used for GEM development. B. Model development using training cohorts: The GEM was developed by linear forward searching of probes most significantly correlated with nodal status in the Laval Cohort and predicting nodal status on the MSKCC Cohort, using a weighted nearest neighbor (WNN) predictor based on correlation of samples. The maximal AUC for discrimination was attained at 21 probes, representing 20 transcripts, (AUC=0·72, [95%CI 0·58–0·85]), and this model was selected as the final model based on training data alone. C. The 20-gene model: Univariate differences for each probe’s expression across nodal status are shown. The normalized expression in node positive samples (pN1-3) in the Laval Cohort is plotted in terms of standard deviations of difference compared to the node negative (pN0) samples of the Laval Cohort (error bars represent 95% confidence intervals).

Figure 2. A. Risk cutoff development for 20-gene model using training cohorts

Using a cross-validation approach applied to both Laval and MSKCC Cohorts, cutoffs were developed from the training cohorts to identify patient risk strata with significantly high or low relative risk of nodal positive disease. B. Comparison of high fidelity vs. all transcript use in GEM based prediction of node positive disease: Box-whisker (boxes, median and interquartile range; whiskers 5th and 95th percentiles) plot of distributions of Area Under the ROC Curve performance of models based on the top 5 through 150 FFPE to FF high-fidelity node-associated genes compared to models based on the top 5 through 150 of all genes derived from the Laval cohort (FFPE). Predictions were made from the Laval Cohort (FFPE) on the MSKCC Cohort (FF). The distributions show a highly significant trend of superiority of the models based on high-fidelity genes, supporting the usefulness of this methodology. The top performing models are also plotted (solid triangles and AUCs for these shown).

Figure 3. Performance of the 20-gene model on a prospectively collected independent test cohort

A. Receiver Operating Characteristic analysis of the predictions outputted by the 20-gene model showed highly significant ability to discriminate between nodal negative and positive tumors in the AUO Cohort. B. Testing of the performance of the relative risk cutoffs on the independent test cohort found that the cutoffs developed from training data again identified groups within the AUO Cohort with significantly increased or decreased risk of nodal positive disease.