DOMINO: diffusion-optimised graph learning identifies domain structures with enhanced accuracy and scalability

Abstract Spatial transcriptomics enables in situ molecular profiling, allowing to measure the cellular transcriptional output within the tissue. As the tissue architecture is conserved, spatial domains with specific transcriptional patterns can be identified, facilitating the discovery and understanding of functional tissue compartments. Thus, several methods to uncover and identify these spatial domains have been developed. However, most of these existing methods do not scale to rapidly increasing data sizes and focus only on local structure while missing the global view of the tissue. Here, we present DOMINO, a diffusion-optimised contrastive learning framework for spatial domain detection. DOMINO utilises graph diffusion convolution to propagate information beyond immediate neighbours and jointly optimises local and, importantly, global graph structure via contrastive learning. This novel framework yields biologically interpretable domains with clearer boundaries and scales to large datasets, outperforming state-of-the-art methods across healthy and malignant benchmark datasets. We apply DOMINO to a newly generated spatial transcriptomic dataset of endometriosis-associated ovarian cancers, which could not be processed by existing domain detection methods owing to its size. We uncovered conserved proliferative and non-proliferative tumour states that recurred across these tumours and were independently validated in an external clear cell ovarian cancer spatial transcriptomic dataset. Proliferative domains were characterised by elevated expression of EIF4A1 and HSPA8, increased cell cycle activity, reduced mast cell abundance, and coordinated stromal remodelling, including altered fibroblast states and spatial organisation. In parallel, integrative analysis across tumours revealed subtype-specific multicellular ecosystems associated with either endometrioid or clear cell ovarian carcinomas, together with a tumour-excluded stromal domain that could only be resolved through the integration of spatial and transcriptional information. These findings demonstrate how well DOMINO scales up and that it uncovers biologically meaningful spatial programs spanning tumour intrinsic states, tumour microenvironment interactions, and subtype-specific tissue architecture that are not recovered by conventional expression-based clustering approaches. Competing Interest Statement The authors have declared no competing interest. Footnotes Figures 3 and 4 are revised. Descriptions are updated to clarify the innovation of our method