Examining response to immunotherapy in clear-cell renal cell carcinoma using baseline serum metabolic signatures.

person Leila Pirhaji Cambridge, MA info_outline Leila Pirhaji, Matthew Morris, Jonah Eaton, Min Zhang

Authors person Leila Pirhaji Cambridge, MA info_outline Leila Pirhaji, Matthew Morris, Jonah Eaton, Min Zhang Organizations Cambridge, MA, ReviveMed Inc., Cambridge, MA Abstract Disclosures Research Funding No funding sources reported Background: Clear cell renal cell carcinoma (ccRCC) accounts for 80% of kidney cancer cases. While immune checkpoint blockade (ICB) is a front-line treatment, most patients do not have durable clinical benefit, and current biomarkers fail to predict response. Previous studies reported metabolic alterations in response to ICB based on profiling of 202 targeted metabolites. To identify metabolic signatures predicting response to ICB at baseline, we have developed mzLearn, a data-driven algorithm that overcomes obstacles in liquid chromatography/mass spectrometry (LC/MS) measurements of untargeted metabolomics at scale. Methods: We used baseline and post-treatment raw LC/MS data (n=1,650) from ccRCC patients in Phase I (n=91) and Phase III (n=741) trials (Workbench IDs ST001236 and ST001237), where anti-PD-1 treatment yielded a significant increase in overall survival compared to mTOR inhibitor. mzLearn detected 4,018 signals present in >60% of all samples. The identified signals have 90% accuracy in detecting targeted metabolites with 10% false positives while effectively overcoming signal drifts and batch effects. We then performed unsupervised K-means clustering of baseline metabolomic features and supervised linear and Cox proportional hazard models to find minimal subsets of those features associated with clinical benefit after adjusting for clinical variables, e.g., age and prognostic scores. We then developed logistic regression models to predict clinical benefit (CB), i.e., tumor shrinkage with ≥6 months progression-free survival (PFS). Finally, we used PIUMet (1), a network-based tool for untargeted metabolomic analysis, to identify linked mechanisms. Results: We identified three main metabolomic subtypes with distinct PFS in response to treatments. We identified a cluster of anti-PD-1 responders (n=279) with significantly higher PFS (p=2e-5, HR=0.56) and a non-responder cluster (n=261) who showed no significant effect from anti-PD-1 relative to mTOR inhibition (p=0.874, HR=1.02). Importantly, one subtype (n=201) tended to have higher PFS after mTOR inhibition (p=0.11, HR=1.27). Next, using supervised methods, we identified 97 baseline metabolite features associated with clinical benefit in the ICB arm. A smaller subset of top-scoring features predicts CB with a cross-validated AUC of 78% and the AUC=71% when tested in the independent Phase I cohort. PIUMet analysis of these features predicted their association with sphingolipid metabolism and arginine/proline metabolism. Conclusions: We showed baseline metabolic stratification of ccRCC patients can predict treatment response and discovered novel blood-based metabolic signatures of predicting response to ICB at baseline. These results establish a solid foundation for predictive biomarker discovery and potentially therapeutic recommendations after further validation. 1. Pirhaji et al. Nat. Met. 2016.