Deep learning to link normal variation in aortic valve hemodynamics to pathogenesis of aortic stenosis

PROJECT SUMMARY Calcific aortic stenosis, an age-related calcification and narrowing of the aortic valve, is the most common heart valve disease and is responsible for over 10,000 deaths and 90,000 heart surgeries each year in the United States. Risk for aortic stenosis is known to have an inherited component, and genetic syndromes such as familial hypercholesterolemia predispose to early disease onset. Nevertheless, trials of cholesterol- lowering medications did not slow disease progression, and at present there are no effective medical therapies for aortic stenosis. There is a critical need to identify risk factors for the development of aortic stenosis and to identify new therapeutic targets. We submit that studying normal population variation in aortic valve measurements (including valve area, mean gradient, and peak velocity) has the potential to address these challenges. Our preliminary data from over 40,000 people provide early evidence that there are shared factors—such as genetic variants near the gene encoding lipoprotein(a)—that are common both to normal variation in aortic valve measurements and to clinically identified aortic stenosis. The specific objective of this proposal is to quantify aortic valve measurements in 100,000 people and to deeply characterize their epidemiological and genetic bases. The overall hypothesis for the proposed work is that factors linked to normal variation in aortic valve hemodynamics in the population underlie the risk for the pathogenesis of aortic stenosis. In Aim #1 of this proposal, we will measure the aortic valve in 100,000 UK Biobank participants using our deep learning models; validate these measurements externally in clinical data from UCSF; and develop deep learning models to identify bicuspid aortic valves. In Aim #2, we will comprehensively characterize the epidemiologic, metabolomic, proteomic properties of aortic valve measurements as well as aortic stenosis as identified in the electronic health record. We will also develop deep learning models to estimate the aortic valve measurements from electrocardiographic data, and externally validate these models in data from tens of thousands of UCSF patients. In Aim #3, we will perform comprehensive analyses of the common genetic basis for variation in 100,000 people with aortic valve measurements and nearly 1.5 million people in a case/control analysis for aortic stenosis. We will also study contributions from rare pathogenic variants and clonal blood cells. The large-scale study of normal variation in aortic valve hemodynamics is an innovative approach for understanding the pathogenesis for aortic stenosis. We believe that our aims are consistent with the NHLBI’s mission of understanding the causes of disease and enabling translation of basic discoveries into clinical practice.