The Ecology of Barrett's Esophagus

Summary Barrett’s esophagus (BE) is the only known precursor to esophageal adenocarcinoma (EA), a highly lethal cancer, the incidence of which has been rapidly increasing in the US. Currently, we over-treat the vast majority of patients with BE who will never progress to EA, and under-diagnose those that will progress. However, the ecological constraints and pressures driving phenotypic changes in BE, and the evolution of EA, are currently poorly understood. Our goals are to interrogate the BE tissue ecosystem to map the cellular, microenvironmental, and phenotypic diversity, and to compare the ecological microenvironment in BE patients that progress to EA to those in which BE remains stable over long periods of time. In Aim 1 we will develop a computational method to test if the spatial structure of different cell types in BE biopsies predicts progression to cancer. Aim 2 develops computational methods to detect abnormal cell proliferation at the top of Barrett’s glands which predict progression to EA. Aim 3 develops deep learning methods and tests if they can detect DNA content abnormalities (increased 4N fractions and aneuploidy) which predict progression to EA. These aims are biologically interconnected, profiling the BE tissue ecosystem, yet are technologically independent. We will use state-of-the-art computer vision machine learning tools to automate quantitative analyses of the spatial distribution of key cell types in the BE ecology. Machine learning, both self-supervised and trained by expert pathologist annotations in conventional H&E images, will unambiguously identify multiple specific cell types and quantify spatial patterns of unique cell habitats on biopsy sections. We will test if these BE tissue ecology measures, in combination and separately from abnormal cell proliferation and genomic abnormalities, distinguish those who progress to cancer from those who remain cancer free during follow-up, to improve clinical care of individuals with Barrett’s esophagus. We will leverage our unique cohort of longitudinally collected BE tissue biopsies, analyzing 5,760 digital tissue section images (split evenly into training and test sets) from patients in the Seattle Cohort to develop spatio-temporal tissue maps at cellular resolution in BE as it progresses to EA. We will validate all our results in an independent cohort from UCSF. The Project brings together expertise in machine learning, computer vision, and digital pathology (Yuan), Barrett’s pathology (Stachler), statistical landscape ecology (Brown), Barrett’s esophagus (Grady, Paulson), BE clinical care (Grady), computational biology and the evolution of BE (Maley) to develop computational pathology systems to quantify the ecology, phenotypic heterogeneity, and genomic abnormality within the BE microenvironment. Our pipeline for calculating landscape ecology statistics to measure tumor microenvironments can be used on any tissue for which we have cell locations and classifications. Digital datasets and computational tools will be made publicly available, providing a rich resource for the early detection of cancer and cancer prevention communities, as well as clinical tools for BE patient care.