Abstract PR015: Interpretable computational pathology reveals that vascular networks reflect evolutionary dynamics in kidney cancer

Abstract

Abstract Clear cell Renal Cell Carcinoma (ccRCC) is profoundly angiogenic, characterised by complex yet heterogenous vascular networks. Blood vessels are an important constituent part of the tumour microenvironment (TME) and, in addition to immune cells, are the target of drug therapies in advanced disease. The TME plays an important role in determining disease progression and response to therapy, acting as a selective pressure on the tumour cells thus influencing evolutionary trajectory. This selective pressure is sculpted by cross-talk between blood vessels, immune cells and the tumour cells themselves. Whilst the details of these carefully orchestrated cellular interactions is not understood their final read-out is reflected in tissue morphology, which can be assessed using an H&E-stained slide, a fundamental component of clinical diagnostic histopathological workflows. A computational pathology approach to assess vascular networks from digital H&E whole slide images (H&E WSIs) would present a powerful tool to understand disease biology. It would permit high-throughput analysis of large cohorts where routine multi-regional sampling captures disease heterogeneity. Such work would lay the foundations for developing a computational pathology biomarker to predict survival outcomes that could be easily implemented into existing clinical workflows. Intricacy of vascular network structures makes reproducible analysis challenging, which can be approached either using morphology, a qualitative evaluation of a shape, or using topography to quantify feature dimensions. Here we reconcile the two methods to develop an interpretable computational pathology solution to study the blood vessels in ccRCC. Further, we have built a deep-learning attention UNET model to segment blood vessels from H&E WSIs. By combining these tools we have developed a computational pathology pipeline able to robustly characterise vascular networks directly from H&E WSIs. We leverage 1064 tumour regions from 82 ccRCC tumours of the TRACERx Renal dataset where ex-vivo multi-regional sampling with closely linked specimens for histological and genomic analysis permits interrogation of the histo:genomic relationship contextualised within the evolutionary dynamics of each tumour. We demonstrate that vascular intratumoral heterogeneity is pervasive and we link different vascular topologies to genetic alterations associated with opposing evolutionary trajectories (PBRM1 and BAP1 mutations) and the acquisition of metastatic competence (loss of 9p). Finally, we show that progressive accumulation of genetic alterations alters vascular network structure, suggesting that vascular topology could be used to assess tumour evolution. Our pipeline is a powerful tool to study ccRCC vasculature in large cohorts with multi-regional sampling to capture intratumoral heterogeneity and ultimately could form the basis of a computational pathology biomarker to predict outcome to therapy. Citation Format: Charlotte E. Spencer, Graham Ross, Thomas Mead, Amy Strange, Anna Song, Katie Bentley, Samra Turajlic. Interpretable computational pathology reveals that vascular networks reflect evolutionary dynamics in kidney cancer [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Translating Cancer Evolution and Data Science: The Next Frontier; 2023 Dec 3-6; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(3 Suppl_2):Abstract nr PR015.