SpatialEx

Spatial omics technologies profile molecules in situ, but each assay is limited in how many genes or proteins it can measure and in which omics layers it captures. Combining complementary molecular panels or different omics across serial tissue sections is difficult because the datasets often share no common molecular features, a challenge the authors term the spatial diagonal integration problem. SpatialEx tackles this by using hematoxylin and eosin (H&E) histology, which is universally available, as a shared anchor across sections.

Developed by Yonghao Liu, Chuyao Wang, and colleagues at Jilin University with collaborators including Monash University, SpatialEx and its extension SpatialEx+ were published in Nature Methods in 2025 following a February 2025 bioRxiv preprint. The framework predicts single-cell omics directly from histology by combining a pretrained H&E foundation model with a cellular hypergraph and contrastive learning, encoding both local multi-neighborhood spatial dependencies and global tissue context. Because histology serves as the anchor, complementary measurements made on different sections can be projected into a unified multi-omics representation.

#Key Features

• Histology-anchored integration: H&E images act as a universal anchor, letting complementary panels or distinct omics layers measured on separate serial sections be reconciled into one multi-omics state.
• H&E foundation model embeddings: For each cell, an image patch is passed through a pretrained histology foundation model to produce a low-dimensional embedding of local tissue appearance.
• Cellular hypergraph encoding: A hypergraph built from spatial cell locations, refined by a hypergraph encoder with contrastive learning, captures multi-neighborhood dependencies and global tissue context.
• Spatial diagonal integration: SpatialEx+ adds an omics cycle module that enforces cross-omics consistency through slice-invariant mappings, enabling integration without co-measured training data.
• Scales to large tissues: The framework handles datasets exceeding one million cells and remains robust with non-overlapping or heterogeneous sections, supporting many omics layers in principle.

#Technical Details

SpatialEx crops an image patch around each cell and feeds it to a pretrained H&E foundation model to generate a cellular embedding. A cellular hypergraph is constructed from spatial coordinates, and a hypergraph encoder refines these features through contrastive learning before a decoder predicts the cell's omics profile. SpatialEx+ extends this with an omics cycle module that learns slice-invariant mappings between omics layers, encouraging cross-omics consistency so that panels or modalities measured on different sections can be integrated even when they share no features. The method was validated on Xenium human breast cancer data, Spatial Multimodal Analysis (SMA) datasets, and immunofluorescence and mouse brain data, and scales to more than one million cells. Code is released under an MIT license.

#Applications

SpatialEx serves researchers in spatial biology and computational pathology who want to expand molecular coverage beyond what a single assay measures. It can predict single-cell transcriptomic or proteomic profiles from routine H&E images, stitch complementary gene panels measured on adjacent sections into a wider effective panel, and integrate transcriptomics with proteomics across slices lacking shared features. These capabilities support tissue characterization in cancer and neuroscience studies where sample material and assay throughput constrain direct multi-omics measurement.

#Impact

SpatialEx formalizes and addresses spatial diagonal integration, using ubiquitous histology as the bridge between molecular measurements that cannot otherwise be aligned. Its peer-reviewed publication in Nature Methods, open MIT-licensed code, and demonstrated scaling to million-cell datasets make it a practical tool for extending the reach of spatial omics experiments. Because predictions are inferred from histology rather than directly measured, downstream conclusions depend on the fidelity of the learned image-to-omics mappings, and predicted profiles are best treated as hypotheses for targeted validation.