Histology-aware Inference of Cell-cell Communication in Spatial Transcriptomics

Cell-cell communication (CCC) is fundamental to tissue organization and function, yet inferring communication from spatial transcriptomics (ST) remains challenging because molecular and spatial information does not fully capture the anatomical context of signaling. Paired hematoxylin and eosin (H&E)-stained images provide complementary insights into tissue architecture and cellular morphology, which remain underutilized as the explicit contextual constraint of CCC inference. Here we present HARMONIC, a histology-aware graph learning framework for anatomically contextualized CCC inference from ST and H&E images. HARMONIC uses conditional mutual information to model dependencies among tissue architecture, ligand-receptor-mediated signaling, gene expression and morphology, and integrates these representations through self-supervised graph learning. Across synthetic and biological datasets spanning platforms, species and tissues, HARMONIC suppresses less anatomically plausible interactions while recovering context-supported signaling across tumor-immune interfaces, necrotic transitions, anatomical compartments and microvascular structures. These findings establish tissue architecture as the explicit contextual constraint of spatial CCC inference.