Abstract LB-65: A gene expression predictor of pathological node stage of urothelial bladder cancer

Abstract

Abstract The use of systemic neoadjuvant cisplatin-based chemotherapy prior to radical cystectomy for clinically muscle invasive but clinically node negative urothelial carcinoma (UC) has been shown in a recent meta-analysis of randomized trials to confer a 5-7% survival benefit (1). This relatively modest effect, therapy toxicity, and treatment delay in non-responsive patients has diminished enthusiasm and use of this approach. In contrast, defining patients that would benefit most would provide a compelling reason for use of such therapy. Despite negative clinical staging, up to 30% of patients have lymph node metastasis at cystectomy, the most important predictor for subsequent disease relapse. Here we develop a gene expression-based predictor of node positive disease at cystectomy in clinically node negative patients that could be applied before definitive resection using tissue from the patient's endoscopic diagnosis. Importantly, we develop a predictor that can be applied on formalin fixed paraffin embedding (FFPE) tumor samples, for facile deployment of this clinical tool. We first performed an analysis of correlation of gene expression between fresh frozen (FF) and FFPE across 32 paired tissues samples preserved both ways, using bootstrapping to select only array probes that maintain robust correlation irrespective of tissue preservation method. Using the “robust” gene set, we identified a subset of genes that were differentially expressed between node positive and node negative disease found at cystectomy in a cohort of 90 FFPE tumor specimens we profiled by oligonucleotide microarrays. A model based on these genes was then optimized for prediction of node positive disease on an independent set of 91 FF specimens profiled on the same microarray platform (2), using a weighted kNN algorithm. A final, 108-feature (probeset) predictor was then tested on a third cohort of 185 FFPE samples from a prospective randomized clinical trial (AUO-AB 05/95) (3). On the independent dataset, the predictor exhibited favorable discriminant characteristics (AUC=0.70, 95%CI 0.62 to 0.78, P<0.0001, ROC Analysis) for prediction of node metastasis at cystectomy in clinically node negative patients. Based on the performance on this prospectively collected dataset, we have developed a dual cutoff system, as has been used for other molecular classifiers (4), providing a threshold yielding a 50% PPV for a priori identification of high risk therapy candidates and a threshold yielding an 85% NPV for identification of low risk candidates to exclude from treatment. Multivariate logistic regression demonstrated this predictor's performance was independently associated with nodal status when adjusting for clinicopathologic factors (P<0.001). We suggest that this strategy may aid in the decision of whether to use neoadjuvant chemotherapy and may help to close the gap between evidence in practice in this regard.