Genomic correlates of response to PD-1 axis inhibitors (IO) in metastatic urothelial cancer (mUC) using explainable artificial intelligence (AI).

Abstract

570 Background: AI is increasingly being used in clinical cancer genomics research. Probabilistic Graphical Models (PGMs) are AI algorithms that capture multivariate, mutli-level dependencies in complex patterns in large datasets while retaining human interpretability. We hypothesize that PGMs can identify clinical and genomic features that correlate with IO response in patients (pts) with mUC. Methods: In this retrospective study eligibility criteria were: diagnosis of mUC, receipt of IO for mUC, comprehensive genomic profiling data available from CLIA certified labs. The Bayesian Network (BN, PGM based AI) was used to discover clinical characteristics and selected genomic alterations relevant to IO response by RECIST 1.1 (investigator assessed). Results: Overall, 95 pts (73 men) with mUC were evaluated. 45 (47%) were ever smokers.The presented BN correctly captured the clinical landscape of mUC explaining significant relationship between included variables (p<0.0001). Ever smokers and pts with de novo metastasis had higher TMB and better response to IO. Inactivating MLL2 alterations were more prevalent in non-smokers, and negatively correlated with response to IO. FGFR3 alterations did not predict response to IO. Significant associations are presented in Table. Conclusions: These hypothesis-generating data (by a novel approach, i.e. PGM based AI) showed that smoking and high-TMB were associated with improved response to IO; in contrast, inactivating MLL2 alternations and visceral metastasis predicted inferior response. FGFR3 alterations did not correlate with response. This model validated previous findings and found new hypothesis-generating relationship, such as altered MLL2 gene; external validation is needed.[Table: see text]