ABSTRACT
The outbreak of the COVID-19 pandemic caused catastrophic socioeconomic consequences and fundamentally reshaped the lives of billions across the globe. Our current understanding of the relationships between clinical variables (demographics, symptoms, follow-up symptoms, comorbidities, treatments, lab results, complications, and other clinical measurements) and COVID-19 outcomes remains obscure. Various computational approaches have been employed to elucidate the relationships between different COVID-19 clinical variables and their contributions to the disease outcomes. However, it is often challenging to capture the indirect relationships, as well as the direction of those relationships, with the conventional pairwise correlation methods. Graphical models (e.g., Bayesian networks) can address these limitations but are computationally expensive, which substantially limits their applications in reconstructing relationship networks ofumpteen clinical variables. In this study, we have developed a method named RAMEN, which employs Genetic Algorithm and random walks to infer the Bayesian relationship network between clinical variables. We applied RAMEN to a comprehensive COVID-19 dataset, Biobanque Quebecoise de la COVID-19 (BQC19). Most of the clinical variables in our reconstructed Bayesian network associated with COVID-19 severity, or long COVID, are supported by existing literature. We further computationally verified the effectiveness of the RAMEN method with statistical examinations of the multi-omics measurements (Clinical variables, RNA-seq, and Somascan) of the BQC19 data and simulations. The accurate inference of the relationships between clinical variables and disease outcomes powered by RAMEN will significantly advance the development of effective and early diagnostics of severe COVID-19 and long COVID, which can help save millions of lives.