Association of machine learning (ML)–derived histological features with transcriptomic molecular subtypes in advanced renal cell carcinoma (RCC).
The results suggest that clinically relevant RCC subtypes may be extracted directly from H&E-stained WSI and may complement gene expression based patient stratification and selection strategies.
Abstract
4519 Background: Metastatic RCC (mRCC) is a molecularly heterogeneous disease. Transcriptomic analysis in the Phase 3 IMmotion 151 (Im151) trial identified 7 molecular subtypes that showed differential outcomes to Atezolizumab+Bevacizumab (A+B) vs Sunitinib (S) treatment (Motzer, Cancer Cell 2020). Here, we present histological correlates of these subtypes as identified in whole slide images (WSI) of hematoxylin and eosin (H&E) stained tumors. Methods: ML models identified 922 H&E derived, human interpretable histological features (ML HIFs) in RCC associated with tumor and stromal (including vessels, immune cells, fibroblasts) cell and tissue morphologies, and nucleus shape. These ML HIFs were then extracted from WSI in 2 mRCC trials – Im151 (n=797) and IMmotion150 (Im150, n=203). Previously described 7 molecular subtypes were combined into 4 subgroups (Angiogenic [comprised of Angiogenic/Stromal and Angiogenic], Complement/OmegaOxidation, T-effector, and Proliferative [comprised of Proliferative and Stromal Proliferative]) for computational power. Due to low prevalence, snoRNA subset was excluded. Univariate analysis with FDR correction was applied to identify positively associated ML HIFs in each of the 4 subgroups in the Im151 WSI and then validated in Im150 subgroups. Representative ML HIFs that showed uniquely higher abundance in each molecular subgroup in both studies were dichotomized by tertiles as ‘high’ or ‘low/intermediate’ and associated with progression free survival (PFS) to fit Cox proportional hazard models in Im151 study. Results: 169 ML HIFs were differentially enriched across 3 molecular subgroups in both Im151 and Im150 data sets. Angiogenic subgroup had higher prevalence of 40 ML HIFs associated with density of endothelial cells in cancer epithelium. T-effector subtype showed higher abundance of 64 ML HIFs associated with immune cell presence in stroma. Proliferative subgroup showed higher prevalence of 40 ML HIFs associated with nuclear morphologies. No ML HIFs were uniquely associated with the Complement/OmegaOxidation subgroup. Consistent with transcriptional findings in Im151, ML HIFs that were enriched in T-effector and Proliferative subgroups showed improved PFS benefit to A+B vs S (Table). Conclusions: We identified unique histological features of RCC tumors that correlate with previously defined molecular subtypes. Our results suggest that clinically relevant RCC subtypes may be extracted directly from H&E-stained WSI and may complement gene expression based patient stratification and selection strategies. [Table: see text]