Abstract
Spatial cellular context is crucial in shaping intratumor heterogeneity. However, understanding how each tumor establishes its unique spatial landscape and what factors drive the landscape for tumor fitness remains significantly challenging. Here, we analyze over 2 million cells from 50 tumor biospecimens using spatial single-cell imaging and single-cell RNA sequencing. We develop a deep learning-based strategy to spatially map tumor cell states and their surrounding environmental architecture, and find that different tumor cell states can be organized into distinct clusters, or "villages," each supported by unique microenvironments. Notably, tumor cell villages exhibit village-specific molecular co-dependencies between tumor cells and their microenvironment and are associated with patient outcomes. Perturbation of molecular co-dependencies via random spatial shuffling of the microenvironment results in destabilization of the corresponding villages. We validate our findings using single-cell, spatial, and bulk transcriptome data from 740 liver cancer patients. This study provides insights into understanding tumor spatial landscape and its impact on tumor aggressiveness.