Cao, S.; Yuan, Y.

Here we present stFormer, a foundation model which incorporates ligand genes within spatial niches into Transformer encoders of single-cell transcriptomics. We demonstrate that despite a moderate pretraining data size, the spatially informed gene representations generated by stFormer more consistently cluster the cells, more accurately encode hierarchy and membership within receptor-dependent gene networks, remarkably boost identification of ligand-receptor interaction pairs, and could simulate perturbation effects of ligand-receptor interactions on downstream targets.