Oral Presentation 47th Lorne Genome Conference 2026

DOMINO: Diffusion-optimised graph contrastive learning framework for domain detection in spatial transcriptomic (133438)

Nora W Liu 1 2 , Pan Jia 3 , Zixu Ran 3 , Suzanne Maiolo 1 2 , Tianjun Zhang 1 2 , Monika Mohenska 1 2 , Xudong Guo 3 , Cong Wang 3 , Elly Walter 1 2 , Carmela Ricciardelli 4 , Noor A Lokman 4 , Rebecca Morrow 5 , Martin Oehler 4 6 , Jose M Polo 1 2 , Ning Liu 1 2 , Fuyi Li 3 7
  1. Cancer Epigenetics Program, The South Australian immunoGENomics Cancer Institute, The University Of Adelaide, Adelaide, SA, Australia
  2. Adelaide Centre for Epigenetics, School of Biomedicine, The University Of Adelaide, Adelaide, SA, Australia
  3. College of Information Engineering, Northwest A&F University, Yangling, Shanxi, China
  4. Robinson Research Institute, The University Of Adelaide, Adelaide, SA, Australia
  5. Gynaepathology, SA Pathology, Adelaide, SA, Australia
  6. Department of Gynaecological Oncology, Royal Adelaide Hospital, Adelaide, SA, Australia
  7. Computational Systems Oncology, The South Australian immunoGENomics Cancer Institute, The University of Adelaide, Adelaide, SA, Australia

Spatial transcriptomics enables molecular profiling in situ, and meaningful interpretation often relies on identifying spatial domains, defined as groups of cells that are molecularly coherent and spatially contiguous. These domains reflect microenvironmental niches and disease relevant tissue architecture. However, most existing domain calling methods do not scale to rapidly increasing data sizes, and focus only on local graph structure while missing the global view of the tissue.

We developed 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 global graph structure via contrastive learning. This integration yields biologically meaningful spatial domains with clearer boundaries, improves domain stability, and scales to millions of cells, outperforming leading domain detection methods across healthy and malignant benchmark datasets.

We applied DOMINO to an in house generated spatial transcriptomic data from four endometriosis associated ovarian cancers, an understudied group with distinct molecular features and poor treatment options. Across all tumours, DOMINO recovered a shared epithelial continuum along a proliferation associated axis spanning FOXJ1 high proliferative to CD55 high non proliferative domains, consistently coupled to coordinated stromal remodelling. Additionally, DOMINO resolved subtype specific spatial domains, including spatially restricted WNT beta catenin activation in endometriod and discrete NDRG1 high metabolic domains in clear cell ovarian cancers, which cannot be recovered by standard expression based clustering workflows.

Together, we present a novel deep learning framework that enables spatial domain detection that yields more stable and biologically meaningful domains and scales to dataset sizes that current approaches cannot compute on. Application of DOMINO to EAOCs reveals conserved and subtype specific spatial domains that provides a potential mechanistic basis for the coexistence of proliferative and non proliferative tumour states and offers new hypotheses for patient level heterogeneity in clinical outcome.