FuncNet: A Machine Learning Framework Capable of Enhancing Neural Cell Type Classification via Functional Subtype Clustering

Cell type identification in single cell RNA sequencing (scRNA-seq) remains an essential yet challenging task. Existing algorithms assign cell type nomenclature typically by (i) subgrouping cells by computing proximity of cellular gene expression patterns at the single cell level, (ii) identifying marker genes in each cluster, and (iii) performing a lookup of these genes in the literature and databases to assign cell subtype labels. Unfortunately, this approach falls short when classifying cells based on functions, because biological function(s) of a group of cells cannot be determined based on a limited set of marker genes. Here we propose a new modeling formalism, namely FuncNet, which is a form of neural network-based method capable of discerning the functional profiles of individual cells. The key idea of FuncNet is to incorporate "a secondary learning arm" into the otherwise classical multi-layer neural network in order to bring in "prior knowledge" into the overall learning process. The prior knowledge that is brought into the FuncNet learning is Gene Ontology Biological Processes that should have been captured in the cells being studied. We applied FuncNet to three publicly available single cell transcriptomics datasets from human and mouse brains. The results demonstrate that FuncNet can successfully reproduce the functional nomenclatures associated with brain cells in the literature. FuncNet suggests a new way of subtyping cells, specifically for the functional subtyping of single cells—an area which has not been well established in the current single cell data analysis practices.