Dissecting and Predicting Lethal Prostate Cancer using Biologically Informed Artificial Intelligence

点击翻译按钮获取中文摘要

PROJECT SUMMARY – PROJECT THREE Treatment strategies for intermediate and high-risk localized prostate cancer (PCa) include surgery or radiation with or without hormonal therapy. Multiple molecular factors, including germline and somatic alterations in DNA repair genes and tissue-based transcriptional biomarkers, have biological and prognostic relevance in these clinical settings yet are rarely used today to guide treatment decisions. Determination of the interacting and co- occurring molecular features that jointly drive indolent or aggressive clinical outcomes in this setting is urgently needed to enable molecularly guided therapeutic strategies and biologically grounded predictive models for clinical use. Furthermore, complex molecular states may converge on histopathological patterns to augment these predictions, but these properties are difficult to quantify, integrate, and generalize across diverse patient populations. The advent of large and diverse patient cohorts with clinically embedded molecular characterization, digital histopathology techniques, and key outcome measures, along with innovations in computation and deep learning to analyze and interpret these data, has created an opportunity to profoundly expand the discovery and translational potential of molecular, pathologic, and phenotypic data for patients with localized PCa. Our overarching hypothesis is that interacting molecular, pathologic, and phenotypic features define prognostic outcomes in intermediate and high-risk localized PCa after surgery, and that biologically guided interpretable deep learning, paired with harmonized cohorts representative of PCa diversity, will transform our understanding of indolent versus potentially lethal localized PCa and deliver on the promise of precision cancer medicine. Toward that end, the specific aims of this proposal are: 1) Dissect the interacting germline and somatic properties that mediate localized PCa using biologically guided neural networks; 2) Determine the convergent spatial histopathologic properties of molecularly and clinically distinct forms of PCa; 3) Develop and validate a clinical grade molecular prognostic model guided by biological networks in real-world and clinical trial settings. For these aims, we will build on our team’s extensive expertise in PCa genomics, computer science, and medical and urologic oncology. Critically, we will embed our approaches in the context of harmonized and representative PCa cohorts. The ability to understand why some intermediate and high-risk localized prostate cancers are phenotypically aggressive, and therefore predict which PCa will progress following curative-intent treatment in this manner, would significantly advance basic PCa research and clinical translation. Broadly, this project will strive to transform precision cancer medicine for prostate cancer and serve as a model for the creation, development, and application of these emerging methodologies across cancer types and contexts.