Genomic predictors of response to PD-1 axis inhibitors in metastatic urothelial cancer (mUC) patients using machine learning analysis of tissue comprehensive genomic profiling (CGP).

Abstract

4542 Background: Immune check point inhibitors (ICI), specifically PD-1 axis inhibitors, are an established treatment for mUC patients (pts), though molecular markers of response are not well-established. Probabilistic Graphical Models (PGMs) are artificial intelligence (AI) algorithms that capture multivariate, multi-level dependencies in complex patterns in large datasets while retaining human interpretability. We hypothesize that machine learning analysis can reveal biomarkers of response to ICI beyond Tumor Mutational Burden (TMB) and PD-L1 expression. Methods: Pts with mUC receiving systemic therapy and available tumor CGP were included in the analysis. CGP was performed by a CLIA validated NGS panel (Foundation Medicine). 17 clinically relevant variables were included in the analyses. Multilevel molecular and clinical interdependencies between ICI response were assessed using a Bayesian network (BN) machine learning approach. To account for high computational cost of the BN dependence structure discovery, variables were selected based on primary and secondary dependencies to ICI Objective Response Rate (ORR) via Hill Climbing Algorithm. Genes selected for analysis included those present in > 5% of the cohort . All variants of unknown significance were removed. Kaplan-Meyer (KM) survival analysis was also performed. Results: 174 pts were eligible and included: median age was 67 (33 – 86), 54.6% were smokers, 27.6% male, 14.4% had upper tract disease at diagnosis and 77 pts were treated with an approved ICI. Results from BN analysis revealed strong positive interdependencies between ICI ORR and TMB ≥ 10 or mutated PIK3CA gene. KM analysis was in agreement with BN results, a low TMB ( < 10 muts/MB) and wild-type PIK3CA were predictive of poor PFS (Table). PGMs will be presented at the meeting. Conclusions: BN and KM survival analysis supported TMB ≥ 10 as a biomarker of improved outcomes to ICI. Moreover, these results reveal mutated PIK3CA as a novel biomarker of improved outcomes to ICI. Study has limitations as expected in a retrospective analysis. These hypothesis-generating data require external validation.[Table: see text]