Multiscale multimodal integration of spatial omics through a deep generative model with CORAL
Abstract
Spatial omics enable the in situ characterization of tissue architecture by measuring molecular modalities. However, integrating these datasets remains challenging due to differences in spatial resolution, cross-modal alignment and molecular coverage. Here we show that Cross-scale spatial Omics Registration and Analysis via graph Learning (CORAL), a probabilistic deep generative model, leverages graph attention to integrate diverse spatial transcriptomic, proteomic, metabolomic and epigenomic data across spatial scales. CORAL jointly models high-resolution datasets with limited molecular coverage and low-resolution genome-wide datasets, generating unified representations with broad molecular coverage and high spatial resolution. Across synthetic and experimental datasets from diverse tissues and platforms, CORAL accurately identifies spatial domains, predicts high-resolution molecular profiles, detects spatially variable genes and infers cell–cell interactions. Applied to spatial transcriptomic and proteomic data from hepatocellular carcinoma patients before and after immunotherapy, CORAL identified organized immune-cell subsets and signaling pathways associated with therapeutic resistance. These results demonstrate CORAL’s broad applicability for resolving tissue organization and clinically relevant multicellular interactions. Cross-scale spatial Omics Registration and Analysis via graph Learning is a deep generative model that integrates multiscale spatial multi-omics data using graph-attention aggregation mechanisms, enabling a comprehensive spatial analysis.