SpatialFormer: universal spatial representation learning from subcellular molecular to multicellular landscapes.

SpatialFormer achieves accurate spatial gene expression modeling by integrating multimodal data, enabling comprehensive analysis across single-cell and multicellular systems.

• Why it matters: Understanding spatial gene expression and cellular organization is crucial for disease diagnosis and biological research, but current models lack effective integration of diverse spatial data.
• What they did: SpatialFormer combines convolutional networks and transformers, trained on 700 million cell pairs from 17 million spatially resolved cells across 71 Xenium slides, to learn multiscale spatial information.
• The result: The model improves tasks like batch correction, cell-type annotation, and co-localization detection, revealing key gene interactions in immune communication, tissue organization, and cancer progression.