Interpretable Hierarchical Bayesian Modeling of Cell-Type Distributions in COVID-19 Disease

High-throughput sequencing of ribonucleic acid molecules is used increasingly to understand gene expression in organs, tissues, and therapies, at a single-cell level. To facilitate the discovery of the heterogeneity and cell-specific factors of the COVID-19 disease, we use an interpretable computational approach that derives cell mixtures from peripheral blood mononuclear cells of healthy donors, and influenza, asymptomatic, mild and severe COVID-19 patients. Cell mixtures are generated using hierarchical Bayesian modeling and are subsequently used as features in the gradient boosting tree classifier. Balanced accuracy of five-fold cross-validation was 68%, significantly higher than expected by random chance. Moreover, 11 out of 19 donors' samples were classified accurately. The main advantage of the mixture-based approach compared to the traditional feature-based classification, is its ability to capture associations between genes as well as between cells.