R4ST: a reference-guided graph-generative model for robust reconstruction of spatial transcriptomic profiles

The trade-off between spatial granularity and transcriptome coverage in current spatial transcriptomics (ST) technologies results in sparse and incomplete expression profiles. Meanwhile, the rich local and global spatial topology inherent in spatial data are crucial for accurate biological interpretation but remain underutilized by existing methods.

Here, we propose R4ST, an end-to-end framework designed to complete ST data. R4ST leverages scRNA-seq data as a reference and employs dual learning channels based on graph inductive and transductive modeling to capture complementary spatial topology information in ST data, enabling accurate reconstruction of missing gene expression. Extensive evaluations across multiple datasets from different platforms demonstrate that R4ST enables accurate recovery of large-scale gene expression profiles from a small subset of measured genes, uncovers novel spatial patterns associated with rare cell types, and substantially enhances the biological interpretability of ST data.

https://github.com/zpliulab/R4ST.