Development of a non-muscle invasive bladder cancer (NMIBC) genomic score (NGS) to predict early BCG failure.

Abstract

850 Background: Although BCG is the standard of care for high-grade (HG)-NMIBC, half of patients recur and 25% progress to muscle invasive disease or require cystectomy. Given emerging intravesical therapies showing promising efficacy, better biomarkers are urgently needed to identify patients at risk for BCG failure and guide personalized therapy. Methods: We included HG-NMIBC patients treated at Memorial Sloan Kettering (MSKCC) who developed early BCG failure [HG recurrence within 6 months of complete induction (n = 24)] or long-term BCG response [no recurrence 5+ years after at least complete induction (n = 40)]. All patients were profiled using MSK-IMPACT – a targeted exome panel. Genes selected for analysis balanced effect size ( > 10% frequency difference) and biological relevance. Three machine learning algorithms (logistic regression, random forest and support vector machine) were employed for feature selection using a stratified 60:40 train-test split to maximize recurrence cases in model evaluation. Models were trained via 5-fold cross validation with AUC-ROC optimization. Random forest achieved the highest test AUC and was selected for feature identification. A normalized NMIBC Genomic Score (NGS) from 0 (low risk) to 100 (high risk) was developed using variable importance weights, with patients stratified into risk tertiles. Kaplan-Meier analysis and multivariable Cox regression adjusting for AUA risk, demographics and CIS were performed. Time-dependent ROC analysis evaluated discriminatory performance. Results: The groups had similar demographic and clinicopathologic characteristics. Random forest variable importance weights were: Mutation Count (100.0), TP53 (38.6), PIK3CA (28.7), TERT (23.2), KDM6A (17.9), FAT1 (17.4) and ERCC2 (15.1). All variables except TP53 were associated with long-term response. Patients were stratified into Intermediate/High NGS (n = 42) and Low NGS (n = 22) groups. Low NGS patients demonstrated superior high-grade recurrence-free survival (HG-RFS) compared to Intermediate/High NGS patients (median not reached, p = 0.025). In a multivariable Cox regression, Low NGS remained independently protective against early BCG failure (HR 0.32, 95% CI 0.11–0.96, p = 0.04). Time dependent ROC analysis at 12 months demonstrated that adding genomic variables to AUA risk improved discriminatory accuracy by 22% (AUA Risk AUC: 0.545, Combined: 0.666). Conclusions: This preliminary evaluation of the novel NGS, derived using data from the clinically deployed MSK-IMPACT assay, demonstrates an independent association with early BCG failure in HG-NMIBC, suggesting a role in risk stratification beyond clinical parameters. Prospective validation in a larger cohort is ongoing and will determine whether this biomarker could identify patients who may benefit from alternative therapies or early radical cystectomy.