STAID: A Self‐Refining Deep Learning Framework for Spatial Cell‐Type Deconvolution with Biologically Informed Modeling

ABSTRACT Spatial transcriptomics provides high‐throughput measurement of gene expression while retaining spatial context; however, inferring accurate cell‐type compositions within individual spots remains a major challenge. Here, we present STAID, a unified framework that effectively integrates pseudo‐spot generation with deep learning training through iterative pseudo‐spot refinement and leverages graph signal processing to capture higher‐order gene‐wise relationships. By creating a self‐reinforcing cycle, STAID enables accurate spot‐level deconvolution of cell‐type compositions for spatial transcriptomics data. Comprehensive benchmarking demonstrates that STAID outperforms existing methods, accurately reconstructs cell‐type spatial distributions, and effectively resolves the cellular colocalization. In clinical breast cancer sections, STAID precisely infers tumor epithelial distributions and reveals their spatial associations with immune cells. In human embryonic limb datasets, STAID captures the ordered spatial distributions of key progenitor populations, reflecting hierarchical tissue organization and demonstrating that incorporating cell‐type composition information can enhance tissue segmentation. STAID also resolves the spatial cellular organization in Crohn's disease and reveals the characteristics of TLS‐like immune niches. Collectively, by delivering high‐resolution cell‐type distributions, STAID provides deeper insights into tissue organization and cellular heterogeneity.