Abstract PR-02: Machine learning models to quantify lineage plasticity and neuroendocrine differentiation in high-grade prostate cancer

Abstract

Neuroendocrine differentiation (NED) is a type of lineage plasticity that occurs in multiple cancer types either de-novo, or as a result of treatment-induced selection pressure. NED is identified by different, partially overlapping clinical, morphological and genomic features which include small cell neuroendocrine histomorphology, gene expression signatures of Rb1 loss and lethal tumor behavior. Through the eyes of pathologists, the most extreme form of NED is small cell, neuroendocrine carcinoma (SmCC), a rare cancer type that can arise in many organs. Adenocarcinomas, the most common cancer type in the adult population can harbor clinical and genomic characteristics of SmCC without recognizable morphologic features. Therefore, we determined whether machine learning algorithms can be trained to identify salient morphologic features of NED in adenocarcinoma. Using a cohort of prostate needle biopsies, we trained complementary algorithms to generate NED scores from high-grade prostate cancer regions and tested NED scores for correlation with gene expression and pathway activation from RNA sequencing of the exact same tumor regions. Using this approach as a validation strategy, we observed correlations between NED scores and pathway activation known to be increased in SmCC. In addition, we discovered novel morphologic features through functional interpretations of pathways correlated with NED scores. Finally, we used multi-variate regression analysis to demonstrate that computer generated NED scores predict the risk of lethal prostate cancer in diagnostic prostate needle biopsies. Altogether, we demonstrate the feasibility to quantify NED - lineage plasticity in regular diagnostic HE 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PR-02.