scGraphETM: Graph-Based Deep Learning Approach for Unraveling Cell Type-Specific Gene Regulatory Networks from Single-Cell Multi-Omics Data

In the forefront of single-cell multi-omics research, the challenge of elucidating intricate gene regulatory networks (GRNs) at a cellular level is paramount. This study introduces the Single Cell Graph Network Embedded Topic Model (scGraphETM), a novel computational approach aimed at unraveling the complexities of cell-specific GRNs from multiomics single-cell sequencing data. Central to our investigation is the integration of single-cell RNA sequencing and single-cell ATAC sequencing data, leveraging the strengths of both to uncover the underpinnings of cellular regulation. The scGraphETM model innovatively combines a variational autoencoder framework with a graph neural network. By conceptualizing transcription factors (TFs), genes, and regulatory elements (RE) as nodes, and their regulatory interactions as edges, the model adeptly captures the dynamic regulatory interplay within cells. It uniquely incorporates both universal and cell-specific features, enabling the model to generalize across cell populations while also identifying unique regulatory dynamics within individual cells. Our results reveal that scGraphETM surpasses existing methodologies in accurately modeling cell-type clustering, cross-modality imputation and cell-type specific TF-RE relationships.