stFormer

Recent foundation models for single-cell transcriptomics data generate informative, context-aware gene representations. The spatially resolved transcriptomics data offer extra positional insights, yet corresponding gene representation methods that integrate both intracellular and spatial context are still lacking. Here, we introduce a gene representation framework tailored for spatial transcriptomics data. It incorporates ligand genes within the spatial niche into the transformer encoder of single-cell transcriptomics. We further propose a biased cross-attention method to enable the framework to do learning with single-cell resolution on low-resolution, whole-transcriptome Visium data. We implemented our framework on a hybrid Visium dataset derived from two human tissue types with distinct developmental and disease states, and tested on various downstream applications. Compared with the latest foundation model for single-cell transcriptomics, our spatially informed gene representations could identify cell groups and gene functions more accurately, and could predict the perturbation effects of cell-cell ligand-receptor interactions on downstream targets.