Abstract PR02: Profiling intratumoral heterogeneity of bladder cancer subtypes at the single-cell level using machine-learning assisted histopathology

Purpose: Machine-learning assisted histopathology using markers of basal and luminal differentiation was employed to profile the intratumoral heterogeneity of bladder cancer from cystectomy patients and predict disease-free survival in this high-risk patient population. Methods: Urothelial carcinomas are biologically heterogeneous and vary greatly in clinical progression as well as treatment response. Delineation of molecular subtypes by gene expression analysis of luminal and basal markers has indicated differential outcomes associated with basal and luminal subtypes. However, histologic validation of this classification using protein markers (basal = KRT5/6, P63; luminal = KRT20/GATA3) has been challenging. While using multiplex-immunofluorescence to subtype a retrospective cystectomy cohort (a TMA of 380 patients), we determined that nearly 50% of tumors did not exhibit cytokeratin markers. Subtyping was further confounded by frequent loss of basal-to-luminal stratification and the emergence of intratumoral spatial heterogeneity with the basal and luminal subtypes being completely intermixed throughout the tumor. These observations caused us to hypothesize that previously undefined but clinically relevant subtypes might exist. To address this challenge we developed a single-cell image analysis pipeline that leveraged machine learning to classify molecular subtype and spatial heterogeneity within each tumor. Using the informatics software KNIME we achieved single-cell segmentation and extracted 285 features for 5 protein markers (P63, GATA3, collagen, nuclear stain, and pan-cytokeratin) from each ~20,000 cells contained in 2 cores of tumor and adjacent benign for each patient. Under guidance from a pathologist, definitive urothelial cells (luminal, intermediate, and basal cells) as well as stromal cells were selected from 25 cores normal urothelium to form the ground truth for XGboost-based machine-learning. Summary Findings: Single-cell profiling with machine learning on transcription factors could classify basal and luminal subtypes with greater than 97% accuracy according to validation in normal urothelium using keratin markers. While we were able to recapitulate differential survival associated with a pure basal subtype, it was the intratumoral heterogeneity of basal and luminal cells that was the predominant driver of disease-free survival. Conclusion: A newly identified bladder cancer subtype defined by intratumoral heterogeneity is a clinically relevant driver of disease-free survival. This abstract is also being presented as Poster B22. Citation Format: Adel Eskaros, Tatiana Novitskaya, Andries Zijlstra. Profiling intratumoral heterogeneity of bladder cancer subtypes at the single-cell level using machine-learning assisted histopathology [abstract]. In: Proceedings of the AACR Special Conference on Bladder Cancer: Transforming the Field; 2019 May 18-21; Denver, CO. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(15_Suppl):Abstract nr PR02.