Histopathology based AI model predicts anti-angiogenic therapy response in renal cancer clinical trial
Jay Jasti, Hua Zhong, Vandana Panwar, Vipul Jarmale, Jeffrey Miyata, Deyssy Carrillo, Alana Christie, Dinesh Rakheja, Zora Modrusan, Edward E. Kadel, Niha Beig, Mahrukh Huseni, James Brugarolas, Payal Kapur, Satwik Rajaram
Abstract
Anti-angiogenic (AA) therapy is a cornerstone of metastatic clear cell renal cell carcinoma (ccRCC) treatment, but not everyone responds, and predictive biomarkers are lacking. CD31, a marker of vasculature, is insufficient, and the Angioscore, an RNA-based angiogenesis quantification method, is costly, associated with delays, difficult to standardize, and does not account for tumor heterogeneity. Here, we developed an interpretable deep learning (DL) model that predicts the Angioscore directly from ubiquitous histopathology slides yielding a visual vascular network (H&E DL Angio). H&E DL Angio achieves a strong correlation with the Angioscore across multiple cohorts (spearman correlations of 0.77 and 0.73). Using this approach, we found that angiogenesis inversely correlates with grade and stage and is associated with driver mutation status. Importantly, DL Angio expediently predicts AA response in both a real-world and IMmotion150 trial cohorts, out-performing CD31, and closely approximating the Angioscore (c-index 0.66 vs 0.67) at a fraction of the cost.