stFormer integrates spatial ligand signaling into a foundation model for spatial transcriptomics

stFormer incorporates ligand genes within the spatial niche into transformer encoder of single-cell transcriptomics, and outputs gene embeddings specific to the intracellular context and spatial niche. These gene representations can serve as input of various downstream applications, including cell clustering, cell type prediction, gene function prediction, and in silico perturbation analysis of ligand-receptor interaction.

The model architecture is designed for ST data resolved at the single-cell level. We propose a biased cross-attention method to enable the model to do learning with single-cell resolution on low-resolution, whole-transcriptome Visium data, which is a widely available spatial resource.

We assembled a pretraining corpus comprising ~4.1 million spatial samples from public human Visium datasets, spanning diverse tissues, development stages, and disease states. After pretraining, stFormer is compatible with both single-cell and spot resolution ST data.