SpaFoundation: a visual foundation model for spatial transcriptomics

Spatial transcriptomics has revolutionized biological research but is limited by trade-offs between spatial resolution and gene coverage. Here we present SpaFoundation, a versatile visual foundational model with 80 million trainable parameters, pre-trained on 1.84 million histological image patches to learn general-purpose imaging representations. Using histology images alone, SpaFoundation accurately infers spatial gene expression and demonstrates strong robustness and transferability across tasks such as super-resolution inference, tumor detection, and spatial clustering.