61Predicting immunotherapy response in ccRCC using deep learning on histopathological slides

Abstract Background Immune checkpoint inhibitors (ICI) are a cornerstone of treatment for metastatic clear cell renal cell carcinoma (mccRCC), yet a significant proportion of patients do not achieve durable responses, while often experiencing immune-related adverse events. There are currently no robust, validated biomarkers to guide therapy selection in the clinic. Retrospective analysis of clinical trial data (eg, IMmotion150/151 and Javelin) suggest RNA based readouts of immune activity can predict response to ICI. While RNA-based signatures of immune activity have shown promise in research settings for predicting ICI response, their clinical adoption is hampered by high costs and standardization challenges. Furthermore, these bulk assays typically fail to capture crucial spatial information regarding immune cell infiltration and are susceptible to intra-tumor heterogeneity, which may contribute to their variable predictive performance across clinical trials. Methods To address these limitations, we developed a deep learning (DL) approach to quantify immune infiltration directly from routinely processed hematoxylin and eosin (H&E) stained histopathology slides, offering a cost-effective and spatially resolved alternative. We build upon an H&E based strategy previously validated by our group, where our prediction of an RNA angiogenesis score by identifying tumor vasculature matched the RNA score’s performance for Sunitinib response prediction in the IMmotion150 trial. Our method involves a DL model trained to classify individual nuclei as tumor, endothelial, immune (CD8+ as a key subset), or other, utilizing immunohistochemistry (PAX8 for tumor, ERG for endothelial, CD8 for cytotoxic T-cells) as ground truth for supervised learning. A separate segmentation model delineated tumor regions. We derived an ‘H&E DL Immune Score’ based on the proportion of immune cells within these tumor regions and validated its correlation with a previously established RNA T-effector signature score. Furthermore, we evaluated the prognostic significance of the H&E DL Immune Score for predicting time to next treatment (TNT) in a real-world cohort of mccRCC patients treated with first-line ipilimumab-nivolumab (Ipi-Nivo) at our institution. We also investigated whether incorporating information from multiple slides per patient and the spatial proximity of immune cells to tumor regions could enhance predictive accuracy. Results The H&E DL Immune Score showed a strong correlation (Spearman r = 0.77) with the RNA T-effector score in a validation cohort. The H&E DL Immune Score (derived from tumor areas) predicted TNT with a concordance index (C-index) of 0.60. Notably, the predictive performance improved to 0.75 upon incorporating immune cells within a 0.5 mm radius of the tumor regions for patients with multiple slides. Conclusions DL-based H&E analysis provides a cost-effective and readily implementable proxy for RNA-based immune signatures. The achieved C-index values (up to 0.75) are comparable to or exceed those reported for other investigational biomarkers for ICI response in mccRCC. Thus, our H&E DL Immune Score shows significant promise for predicting response to ICI therapy in mccRCC patients, with the potential to be developed into a clinically valuable biomarker to be evaluated prospectively in clinical trials.