Supervised Deep Learning with Gene Annotation for Cell Classification

Gene-by-gene differential expression analysis is a widely used supervised approach for interpreting single-cell RNA-sequencing (scRNA-seq) data. However, modern scRNA-seq datasets often contain large numbers of cells, which can produce numerous differentially expressed genes with exceedingly small p-values but minimal effect sizes, and thus making biological interpretation difficult. To overcome this challenge, we developed Supervised Deep learning with gene ANnotation (SDAN), a method that integrates gene-annotation information with gene-expression profiles using a graph neural network. SDAN identifies functionally coherent gene sets that best classify cells, and the resulting cell-level classification scores can be aggregated to make individual-level predictions. We evaluated SDAN and two representative existing methods in three real-data applications to identify gene sets associated with severe COVID-19, dementia, and immunotherapy response in cancer. SDAN consistently outperformed alternative approaches by achieving two key objectives simultaneously: accurate classification of outcomes and unambiguous assignment of genes to gene sets of functionally related genes.