Probabilistic-graph-based spatial context-aware framework for interpretable spatial omics denoising and augmentation.

Spatially resolved omics technologies offer unprecedented insight into tissue organization, yet current analytical methods face challenges in effectively handling technical noise while preserving biological heterogeneity. We present CadaST, an interpretable and unified computational framework that integrates spatially aware feature selection and an adaptive imputation strategy to address this limitation. By inferring the spatial molecular pattern for each feature and conducting pattern-guided aggregation, CadaST effectively denoises and augments spatial omics data while preserving sharp biological boundaries. This gene-centric approach robustly denoises data without the oversmoothing common to other methods. CadaST is versatile and highly effective, outperforming existing methods across diverse spatial technologies. It accurately resolves delicate anatomical layers in the brain, characterizes complex tumor microenvironments, and scales efficiently to large-scale developmental atlases. By providing a more accurate, interpretable, and scalable solution, CadaST represents a significant methodological advance for elucidating the principles of tissue architecture in health and disease.