Knockoff statistics-driven interpretable deep learning models for uncovering potential biomarkers for COVID-19 risk prediction

Canada's healthcare systems have been greatly affected by the coronavirus disease 2019 (COVID-19). Some patients show more severe symptoms of the disease, and thus, need more healthcare resources, while others show mild or even no symptoms. Previous studies have identified several clinical factors, such as age and sex, that are related to this difference. There are also studies examining genetic factors that play important roles in leading to the severity difference. However, these COVID-19 genetic studies did not use enough Canadian patients' information, and their sample sizes are not big enough, which limited the discovery of significant genetic factors for the Canadian population. In this study, we will develop a powerful artificial intelligence (AI) tool to examine which genetic factors are related to the difference of COVID-19 severity using whole-genome sequencing data from a large Canadian COVID-19 cohort. Although AI is a successful technique because of its high accuracy in prediction, it is limited in real clinical practice due to its low interpretability. We will integrate a recently developed statistic method that can handle high-dimensional data and control false discoveries to address the low interpretability issue in the AI tool. This statistic method can provide information about the AI's learning process, thus making the AI model easier to be understood by humans. We will perform a sex-specific genome-wide association (GWA) analysis to explore the potential genetic risk factor for males and females. Finally, we will also build a predictive AI model to predict an individual's risk of developing severe COVID-19 using both the genetic risk score and other known clinical risk factors such as age and gender. The results of this study will help us understand how genetic factors can influence COVID-19 severity and may help us develop better plans for individual COVID-19 treatment, and eventually allocate healthcare resources more efficiently.