Predicting Clinical Phenotypes in Crohn's Disease Using Machine Learning and Single-Cell 'omics

PROJECT SUMMARY/ABSTRACT Pediatric Crohn's disease presents as a chronic, relapsing inflammatory condition of the gastrointestinal tract, leading to malabsorption, anemia, and psychosocial decline. The incidence rate of Crohn's disease has been growing in the 10- to 18-year age group. Crohn’s disease exists on a spectrum of clinical severity, ranging from mild disease responsive to standard anti-TNFɑ therapy to severe, treatment-resistant disease with stricturing (B2) or penetrating (B3) complications often requiring surgical intervention. Distinguishing which patients will progress to more severe disease from patients who will require minimal intervention at the time of diagnosis is an urgent unmet need. Accurate and automated prediction of disease outcomes will significantly improve patient health by informing personalized interventions for individual patients. Previous attempts at generating predictive models of Crohn’s disease relying solely on clinical features of the disease and patient biodata have demonstrated promising, yet inadequate accuracies for clinical practice applications. This proposal addresses these limitations by leveraging large cohorts of archival and prospective patient clinical metadata, ‘omics, and machine learning derived tissue features to build and test machine learning models for predicting specific Crohn's disease outcomes. In Aim 1, we will build, test, and validate predictive models of Crohn’s disease using computational image analysis of gold-standard biopsy histopathology slides. We will use saliency maps and gene correlations analysis to validate our models by visualizing the tissue features of importance to our predictive models and identify specific transcriptomic changes associated with these features. In Aim 2, we will generate a clinically-relevant predictive model of Crohn’s disease by integrating the deep features extracted from histology image analysis with other patient metadata collected as part of standard clincal care. Additionally, we will collate a thorough list of published predictive models of Crohn's disease to benchmark the performance of our proposed and future predictive models. Lastly, in Aim 3 we will use cutting edge single-cell RNA sequencing and spatial transcriptomics approaches to elucidate a transcriptomic signature of Crohn's disease and characterize specific genetic profiles associated with the hallmark morphological changes in diseased tissue. These data will provide a framework for studying the subtypes and clinical outcomes of Crohn’s disease and other gastrointestinal diseases, thus driving the clinical adaptation of personalized therapy and precision medicine. This proposed research will increase the resolution of both diagnostic and prognostic information to better manage Crohn’s disease in patients and significantly shft clinical management to an individualized treatment paradigm.