Deep Learning-Driven Coronary Artery Disease Diagnostics leveraging Cell-Free DNA Methylation

Abstract

Coronary artery disease (CAD), caused by atherosclerotic plaque buildup within coronary arteries, usually remains undiagnosed until severe symptoms occur, such as heart attacks (1-4). CAD affects approximately 200 million people globally, a leading cause of death in every population (1-3). The current ‘gold standard’ for CAD diagnostics is cardiac catheterization and angiogram, a costly and invasive procedure (2). Collectively coronary heart diseases (CHDs) cost approximately 11% of the EU healthcare budget over 77 billion euros annually. Due to the high prevalence and costs associated with CAD, novel diagnostic and predictive tests are ever more necessary. Here we propose the introduction of cell-free DNA (cfDNA) diagnostics, already leveraged in cancer and prenatal diagnostics, within the cardiac diagnostic field (5,6,8,9,13-15). CfDNA is released into blood from dying cells, it contains sequence and methylation information from their tissue of origin throughout the body. While varying cfDNA compositions have been described for cardiovascular diseases, reliable biomarkers and diagnostic tests have yet to be determined (10-12). Therefore, we propose the incorporation of natural language processing to create novel deep learning approaches using cfDNA for a classification model to differentiate CAD between patients. Using the novel human methylation atlas as a pre-training data corpus and creating novel tokenizers we aim to pre-train and fine tune a novel methylation language model capable of incorporating DNA methylation within the model pre-training (17). Furthermore, leveraging language model’s unique capabilities to track model attention mechanisms we can calculate which parts of the sequence or sequences help differentiate CAD patients from healthy, allowing us to further speculate alternative biomarkers (16). Given the success of such a model we can significantly impact the clinical setting of CAD diagnostics as well as create a novel language model which can be fine-tuned for various DNA methylation tasks. Furthermore, this approach could be leveraged to identify additional CHDs in the future.