Deep Generative Model Integrates Multiscale Spatial Omics Across Diverse Tissues

Researchers have introduced Cross-scale spatial Omics Registration and Analysis via graph Learning, known as CORAL, a probabilistic deep generative model that leverages graph-attention aggregation mechanisms to achieve multiscale and multimodal integration of spatial omics data. Contemporary spatial omics technologies enable the detailed in situ characterization of complex tissue architecture through the direct measurement of distinct molecular modalities within intact biological samples. Nevertheless, integrating datasets across diverse experimental technologies has remained a substantial computational bottleneck, primarily due to pronounced differences in spatial resolution, difficulties in cross-modal alignment, and disparities in molecular coverage across platforms. To overcome these critical barriers, CORAL utilizes graph attention across spatial scales to jointly model high-resolution datasets possessing limited molecular coverage alongside low-resolution genome-wide datasets. Through this unified probabilistic formulation, the generative model synthesizes comprehensive representations that deliver both broad molecular coverage and high spatial resolution across spatial transcriptomic, proteomic, metabolomic, and epigenomic measurements.

Across extensive evaluations involving both synthetic benchmarks and experimental datasets derived from diverse tissues and instrumentation platforms, CORAL demonstrated consistent analytical accuracy across several core computational tasks. The deep generative model accurately identifies spatial domains, predicts high-resolution molecular profiles, detects spatially variable genes, and infers complex cell–cell interactions. By establishing robust multimodal integration for high-resolution, single-cell molecular mapping, CORAL effectively bridges the methodological trade-offs that typically exist between spatial granularity and molecular breadth.

The investigators applied the CORAL framework to several distinct anatomical and pathological tissue contexts to demonstrate its versatility across diverse biological systems and spatial measurement techniques.

In a direct clinical investigation, CORAL was applied to evaluate spatial transcriptomic and proteomic data collected from human hepatocellular carcinoma patients before and after immunotherapy treatment. By integrating single-cell resolution CODEX multiplexed imaging with Visium spatial transcriptomics, the deep generative model uncovered the intricate cellular organization of the hepatocellular carcinoma tumor microenvironment across different therapeutic stages. Through this cross-scale integration, CORAL successfully identified organized immune-cell subsets and resolved distinct molecular signaling pathways associated with therapeutic resistance to immunotherapy. The authors demonstrated that these empirical findings establish CORAL as a broadly applicable generative framework for resolving multicellular tissue organization and identifying clinically relevant intercellular interactions in complex clinical disease states.

Subject of Research: Multiscale spatial multi-omics data integration using deep generative modeling

Article Title: Multiscale multimodal integration of spatial omics through a deep generative model with CORAL

Article References: He, S., Bieniosek, M., Zhou, J., Song, D., Chidester, B., Wu, Z., Boen, J., Wang, G., Kaseb, A. O., Sharma, P., Trevino, A. E., & Zou, J. (2026). Multiscale multimodal integration of spatial omics through a deep generative model with CORAL. Nature Genetics. https://doi.org/10.1038/s41588-026-02821-6

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02821-6

Keywords: spatial omics, deep generative model, graph attention, spatial transcriptomics, proteomics, hepatocellular carcinoma, tumor microenvironment, multimodal integration, Multiscale, multimodal, integration, spatial