Detecting Genetic Susceptibility for Fatty Liver Disease in Lean vs. Non-lean Individuals: Use of a Deep Learning Algorithm PROJECT SUMMARY/ABSTRACT Fatty liver disease (FLD) is the most common liver disease in the world, affecting nearly 1 billion people. Although it commonly occurs among overweight and obese (non-lean) individuals, a quarter of those affected are lean. Paradoxically, although lean individuals with FLD are less likely to have cardiometabolic risk factors (e.g., diabetes, hypertension, and dyslipidemia), they have a two-fold increased risk of cardiovascular and liver-related mortality. Metabolic dysfunction and alcohol use are the most common causes of FLD. However, they have overlapping genetic risk factors and cannot be distinguished by medical imaging or histopathology. To date, it remains unknown if there are any lean-specific genetic variants for FLD. Current thresholds of alcohol consumption associated with FLD are arbitrary, and it is also unclear if alcohol consumption accounts for the difference in outcomes observed between lean and non-lean individuals with FLD. With the unique clinical and imaging data available within the Veterans’ Affairs (VA) system, combined with large-scale genetic data available in the Million Veteran Program (MVP), the goal of this project is to identify convergent and divergent features of lean and non-lean FLD by comparing associated clinical and genetic risk factors in those with and without substantial alcohol exposure. We have established a cohort of over 81,000 veterans in the MVP with genetic data who underwent non-contrast abdominal computed tomography (CT) imaging for any indication in the context of clinical care between 2011 to 2023. Our preliminary study of 45 lean and 72 non- lean individuals shows that Automatic Liver Attenuation Region-of-interest-based Measurement (ALARM), a deep learning tool was accurate in identifying hepatic steatosis in both lean and non-lean individuals compared to blinded expert radiology read. The central hypothesis to be tested is that application of the ALARM tool to non-contrast abdominal CT scans will accurately classify lean FLD and improve the identification of new genetic variants for non-alcoholic and alcohol-related FLD in lean individuals. This hypothesis will be tested through three specific aims: (1) We will assess the accuracy of ALARM in phenotyping lean FLD; (2) We will characterize lean and non-lean FLD by level of alcohol consumption in the MVP cohort; and (3) Identify genetic variants associated with lean FLD. The proposed project will assemble the largest ever multi-ancestry cohort of persons with FLD based on the application of artificial intelligence methods to clinically obtained non-contrast abdominal CT scans. The proposed research will improve our understanding of the clinical and genetic underpinnings of FLD in lean individuals and how alcohol consumption impacts these clinical and genetic risk factors. We expect to identify dist...