DeepHostSeverity: A Knockoff Statistics-Driven and Interpretable Deep Learning Model for Identifying Biomarkers Associated with COVID-19 Severity

Coronavirus Disease 2019 (COVID-19) remains a public health concern, with some patients experiencing mild or asymptomatic cases, while others require hospitalization and mechanical ventilation. Current approaches for investigating genetic differences among COVID-19 patients have identified genetic variants linked to disease severity. However, these variants often account for a small fraction of variation in COVID-19 severity and have limited predictive value. Translating these discoveries into practical knowledge is difficult as current methods fall short of identifying causal genetic variants and fail to account for SNP-SNP interactions. Although artificial intelligence (AI) methods can model complex relationships between genetic variants and have high prediction accuracy, these methods have limited clinical use due to their low interpretability. In our study, we will develop an interpretable AI tool to predict COVID-19 severity using genomic and demographic data. Using genomic data from a Canadian COVID-19 cohort, we will develop an AI model to identify genetic variants associated with COVID-19 severity. We will explore various model architectures and rigorously tune our model to achieve optimal performance. We will incorporate statistical methods to control the false discovery rate and give insight into the model's learning and decision-making process. Finally, we will build a predictive AI model to assess an individual's risk of developing severe COVID-19 using their genetic risk score and demographic risk factors. The results of our work will shed light on genetic factors and biological mechanisms underlying COVID-19 severity and may improve vaccine development, individual patient outcomes, and healthcare decision-making.