Scalable Condition-relevant Cell Niche Analysis of Spatial Omics Data with Taichi

Tissues are composed of heterogeneous cell niches, which can be investigated using spatial omics technologies. Large consortia have accumulated vast amounts of spatially resolved data, which typically assign slice-level condition labels without considering intra-slice heterogeneity, particularly differential cell niches that respond to certain perturbations. Here, we present Taichi, an efficient and scalable method for condition-relevant cell niche analysis that does not rely on pre-defined discrete spatial clustering. Taichi utilizes a scalable spatial co-embedding approach that effectively accounts for batch effects, incorporating advanced label refinement and graph heat diffusion techniques to explore condition-relevant cell niches across extensive multi-slice and multi-condition spatial omics datasets. Comprehensive benchmarks demonstrate Taichis ability to precisely identify condition-relevant niches under various levels of perturbations. We showcase Taichis effectiveness in accurately delineating major shifts in cell niches in a mouse model of diabetic kidney disease compared to a normal group, revealing disease-specific cell-cell interactions and spatial gene expression patterns. Furthermore, Taichi can identify key subtype-relevant niches between colorectal cancer patient groups with significantly different survival outcomes. Moreover, we demonstrate that Taichi can help discover more fine-grained clinical properties within the originally coarse-defined patient groups in large-scale tumor spatial atlases, reflecting intra-group heterogeneity obscured previously. Additionally, we combine Taichi and tensor decomposition to discover higher-order biomarkers relevant to the immunotherapy response of triple-negative breast cancer. Finally, we highlight Taichis speed and scalability by confirming its unique applicability in large-scale scenarios containing up to 16 million cells in [~] 12 minutes. Taichi provides a powerful tool for mining disease-relevant spatially resolved insights in the era of big data in spatial biology.