Abstract Tissues are organized through the assembly of diverse cell types into multicellular structures that exhibit hierarchical spatial organization. We present HRCHY-CytoCommunity, a graph neural network framework for identifying multi-level tissue structures directly from cell-type annotated spatial maps. It integrates differentiable graph pooling, adaptive edge pruning, and consistency and balance regularization in an end-to-end model, simultaneously inferring robust structures across multiple scales while preserving complete cellular coverage and fully nested relationships. The framework also supports cross-sample hierarchy alignment via cell-type enrichment-based clustering. Benchmarking on diverse spatial omics datasets, HRCHY-CytoCommunity outperforms existing hierarchical and non-hierarchical methods in identifying both coarse-grained tissue compartments and fine-grained cellular neighborhoods. Applied to a breast cancer cohort with clinical outcomes, the framework enables hierarchical prognostic stratification of patients and reveals survival-associated spatial patterns. HRCHY-CytoCommunity represents a general and scalable tool for deciphering tissue organization from single cells to multicellular modules, and ultimately to intact tissues and organs.
Xie et al. (Sat,) studied this question.