SpatialFormer, a new artificial intelligence framework, enables universal spatial learning across molecular and multicellular landscapes by combining convolutional neural networks and transformer architectures. Published in Nature Computational Science, the hybrid model learns how gene expression is organized across biological scales, from subcellular locations to multicellular tissues, integrating molecular activity with cellular neighborhoods. Spatial technologies such as Xenium record the physical positions of RNA molecules within tissue sections, but a cell may contain thousands of transcripts, and its behavior can depend on nearby immune, epithelial, stromal or malignant cells. SpatialFormer integrates these signals to produce a compact numerical representation of each cell reflecting both its molecular identity and location within a tissue niche. Convolutional networks recognize local patterns, while transformers use attention mechanisms to evaluate broader relationships, capturing information at both subcellular and multicellular scales. The researchers trained the system using a pairwise strategy based on relationships between cells, exposing the model to cell pairs with associated spatial and expression information to distinguish intrinsic cellular features from signals emerging through proximity or shared niche context. During pretraining, SpatialFormer processed approximately 700 million cell pairs drawn from 17 million spatially resolved single cells across 71 Xenium slides. This large dataset helps the model learn reusable biological patterns rather than memorizing one experimental sample.