Biomarker-driven prospective clinical trial in renal cell carcinoma: Developing machine learning models to allocate patients to treatment arms using RNA sequencing.

person Anupama Reddy Vindhya Data Science, Durham, NC info_outline Anupama Reddy, Scott Mattox Haake, Katy Beckermann, Alex Nesta, Adnan Derti, Wendy Kimryn Rathmell, Nanguneri Nirmala, Brian I. Rini

Authors person Anupama Reddy Vindhya Data Science, Durham, NC info_outline Anupama Reddy, Scott Mattox Haake, Katy Beckermann, Alex Nesta, Adnan Derti, Wendy Kimryn Rathmell, Nanguneri Nirmala, Brian I. Rini Organizations Vindhya Data Science, Durham, NC, Vanderbilt University Medical Center, Nashville, TN, Vanderbilt University, Nashville, TN, Vindhya Data Science Inc., Morrisville, NC, Vindhya Data Science, Inc., West Roxbury, Department of Medicine, Vanderbilt University Medical Center, Nashville, TN Abstract Disclosures Research Funding Other Government Agency Department of Defense Background: RNA sequencing has shown promise in defining the biology of individual renal cell carcinoma (RCC) tumors. However, these biomarkers have not yet been translated to the clinic for prospectively assigning optimal treatments to patients. Challenges for RNA-seq biomarker development include translating classifiers across different assays/platforms, normalization of data collected from single patients in the clinic and establishing robust thresholds for assigning prediction groups. We have designed the prospective phase II OPTIC RCC clinical trial (clinicaltrials.gov NCT05361720) to test the utility of an RNA-seq based biomarker in predicting treatment based on biologic drivers relevant to angiogenesis (anti-angiogenesis; TKI) and immune microenvironment (immunotherapy; IO). Here, we will describe the development and optimization of a machine learning model for assigning individual patients to biologically driven clusters in real time to facilitate RNA-seq based biomarker trials. Methods: We have utilized RNA-seq data from the IMmotion 151 trial (ref) to develop a machine learning model. Clusters were grouped into three based on their association to tumor biology and treatment: (1) Cluster 1+2 (angiogenic signature → TKI+IO therapy), (2) Cluster 4+5 (immune/proliferative signature → dual IO therapy), (3) Cluster 3+6 (neither signature → exclude). Random forest classifier was used to train a multi-class model to predict the three groups. The model is evaluated using bootstrapped cross-validation. Results: Our machine learning classifier was built using 188 genes and has a cross-validation accuracy of 85% and sensitivity of >90% in predicting patients into one of the three biological clusters from our training data. Predictions of the classifier are significantly associated with progression-free survival across different treatments within each of the predicted groups. We also observed significant odds ratios when comparing responders (CR/PR/SD) to non-responders (PD) across the treatment groups. The model was then validated in two independent test sets treated with Angio and IO inhibitors: (1) 61 patient renal cell carcinoma cohort (2) 12 patient clear cell renal cell carcinoma cohort. We observed a significant enrichment of responders to Angio+IO treatment for the predicted Cluster_Angio patients compared to IO treatment (p=0.05), highlighting the importance of matching patients to their optimal treatment. Conclusions: We have developed an accurate machine learning model to assign individual patients to RNA-seq clusters in real time. This classifier will facilitate the prospective OPTIC RCC trial. If successful, our biomarker strategy will serve as a proof of concept for selecting optimal treatments for RCC patients.

Clinical