scDGG: Dynamic gene graphs for enhancing clustering analysis of single-cell RNA sequencing data via spatiotemporal representations.

For high-throughput single-cell RNA sequencing (scRNA-seq) data, spatial features have emerged as a powerful representations for downstream analysis. These spatial features contain but not limited to gene graphs and cell graphs. Specifically, gene graphs have been inferred to capture functional interactions between transcriptional factors and marker genes, which are associated with abnormal expression patterns and molecular heterogeneity. Furthermore, incorporation of spatial features is useful to enhance the accuracy of single-cell clustering. However, static gene graphs encode limited cellular information in conveying dynamic regulatory mechanisms that govern cell fates as well as disease progression. To alleviate this drawback, this work extracts and employs dynamic gene graphs, which contribute to a more comprehensive observation of regulatory mechanisms. This study proposes an multi-view graph learning architecture named scDGG to compress dynamic gene graphs from various signaling pathways, with each graph representing a specific biological context. Experimental results about benchmark scRNA-seq datasets have demonstrated the effectiveness and advantages of the scDGG method over SOTA single-cell clustering approaches that take deep learning architecture. It seems dynamic gene graphs could be regarded as high-quality graph representations that outperform static spatial features in single-cell clustering.