CellFeature: Cell and Feature Co-Embedding from Single-Cell Multi-Omics with Heterogeneous Graph Model

Most current single-cell multi-omics analysis methods are limited to the co-embedding of different omics cells and lack the ability to directly analyze the relationship between different omics cells and different types of features. Here, we describe a multi-omics cells and heterogeneous features co-embedding algorithm, named CellFeature, by eliminating the heterogeneity between different types of nodes based on heterogeneous graph representation learning framework. Using different single-cell multimodal datasets, we compared multi-omics cell embeddings with current state-of-the-art methods and achieved comparable results. By leveraging co-embedding of cells and features, we demonstrate that CellFeature can identify cell-type-specific features with better performance than existing methods. Based on the co-embeddings, CellFeature can also perform cell subtypes discovery and enable trajectory-specific genes identification.