Multimodality spatial analysis in prostate cancer to improve prognostic estimation and cast light into the black box of pathology artificial intelligence algorithms

PROJECT ABSTRACT: Prostate cancer is the most common cancer diagnosis in men, with approximately 300,000 new cases per year, marked by exceptionally heterogeneous prognosis across patients. We are now a decade into the era of clinically available genomic biomarker tests to improve clinical prognostic estimates and help guide decision-making after diagnosis of prostate cancer. Three RNA-based tests from prostate tissues are currently included in NCCN guidelines, and they have recently been joined by the first predictive biomarker derived from pathology artificial intelligence (PAI) analysis of standard H&E-stained cancer tissue. However, current genomic tests offer limited insights into tumor heterogeneity and biological variability, and little is understood of how PAI systems reflect the underlying biology, limiting their potential in advancing cancer research across patient sub-groups. The goal of this proposal is to apply two cutting-edge spatial proteogenomic technologies together with PAI to identify aspects of previously obscure biology which drive AI outcomes, and to explore and elucidate prostate cancer heterogeneity and prognosis at unprecedented depth. We will identify and validate, through micron-scale subcellular spatial analysis, novel genomic and proteomic markers of outcomes after prostate cancer treatment that can both explain and extend emerging pathology artificial intelligence (PAI) algorithms. We propose to 1) understand competitive and/or additive relationships between established standard-of-care genomic expression and PAI scores in predicting outcomes after radical prostatectomy, 2) employ spatial proteomics and transcriptomics to describe prostate cancer heterogeneity, local evolution, and cellular diversity at unprecedented detail; and to identify novel markers that can improve on both existing genomic scores and PAI, and 3) illuminate the black box of PAI—and build a better box by comparing subcellular proteogenomic and PAI convolutional features, and incorporating all these streams into next-generation AI algorithms. These platforms in concert will improve on our current ability to predict outcomes based on clinical and imaging parameters alone, and will yield novel insights into prostate cancer’s heterogeneity within patients, between individuals, and across diverse population groups. Overall, this study will greatly enhance our understanding of prostate cancer biology and heterogeneity within and between patients, and across critical sub-populations. We expect to generate next- generation artificial intelligence-based tools to drive a new level of personalized treatment for patients across the disease spectrum.