Abstract
Spatial omics face challenges in achieving high-parameter, multi-omics co-profiling. Serial-section profiling of complementary panels mitigates technical trade-offs but introduces the spatial diagonal integration problem. To address this, we present SpatialEx and its extension SpatialEx+, computational frameworks leveraging histology as a universal anchor to integrate spatial molecular data across tissue sections. SpatialEx combines a pre-trained H&E foundation model with hypergraph and contrastive learning to predict single-cell omics from histology, encoding multi-neighborhood spatial dependencies and global tissue context. SpatialEx+ further introduces an omics cycle module that encourages cross-omics consistency via slice-invariant mappings, enabling seamless integration without co-measured training data. Extensive validations show superior H&E-to-omics prediction, panel diagonal integration, and omics diagonal integration across various biological scenarios. The frameworks scale to datasets exceeding one million cells, maintain robustness with non-overlapping or heterogeneous sections, and support unlimited omics layers in principle. Our work makes multi-modal spatial profiling broadly accessible.
Introduction
In the rapidly evolving field of spatial biology, the ultimate goal is to achieve a comprehensive "full picture" of tissue architecture—mapping the precise location of every gene, protein, and metabolite. However, current technologies are often constrained by a fundamental trade-off: researchers can achieve high-resolution mapping of a few targets or low-resolution mapping of many, but rarely both simultaneously on a single tissue slice.
To circumvent these technical limitations, scientists frequently profile different molecular layers on serial (adjacent) tissue sections. While effective for data collection, this approach introduces the "spatial diagonal integration problem." Because adjacent slices are physically distinct, cells do not align perfectly, and there are often no shared molecular features to serve as a bridge for data integration. SpatialEx and SpatialEx+ address this by utilizing standard H&E (Hematoxylin and Eosin) histology—the clinical gold standard—as a "universal anchor" to stitch disparate omics layers together.
The SpatialEx Framework
The core of this approach is SpatialEx, a computational engine designed to predict molecular profiles directly from H&E images. The framework relies on three primary pillars:
- H&E Foundation Model: SpatialEx utilizes a pre-trained pathology foundation model (UNI) to extract high-dimensional morphological features from single-cell image patches. This allows the model to interpret the visual language of tissue at a granular level.
- Hypergraph Learning: Traditional graph neural networks often model simple cell-to-cell pairs. SpatialEx employs hypergraphs to model complex "neighborhoods" of cells. This captures the intricate microenvironmental interactions and spatial dependencies that define biological function.
- Contrastive Learning: By implementing a "local-global" contrastive strategy, the model ensures it understands both the specific morphology of an individual cell and the broader tissue context in which that cell resides.
SpatialEx+: Solving the Integration Problem
Building upon the core engine, SpatialEx+ introduces an Omics Cycle Module specifically designed for multi-slice integration. This module enforces a consistency constraint: the mathematical relationship between different omics types (e.g., the correlation between specific gene expressions and protein levels) should remain invariant across adjacent slices.
By leveraging these slice-invariant mappings, SpatialEx+ can "fill in the blanks." It can take the data learned from Slice A (e.g., transcriptomics) and project it onto Slice B (e.g., proteomics) using the shared H&E anchor, even when no co-measured training data is available. This enables the reconstruction of high-parameter multi-omics maps from standard serial sections.
Experimental Validation and Results
The researchers validated the frameworks across diverse biological contexts and high-profile datasets, demonstrating significant improvements over existing methods like DeepPT.
Superior Accuracy and Discovery
In benchmarks using 10x Xenium Human Breast Cancer data, SpatialEx significantly outperformed state-of-the-art models in Pearson Correlation Coefficient (PCC) and Structural Similarity (SSIM). Beyond metrics, the model demonstrated biological utility by identifying distinct immune and stromal boundaries in breast cancer samples that were not readily apparent to senior pathologists using H&E slides alone.
Multi-Modal Insights in Neurodegeneration
In a mouse brain model of Parkinson’s Disease, the framework integrated MALDI-MSI (metabolomics) and Visium (transcriptomics). The resulting integrated analysis successfully distinguished between "lesioned" and "intact" regions of the choroid plexus—a differentiation that neither individual omics modality could achieve in isolation.
Robustness and Scalability
A critical finding was the model's robustness to physical gaps. In simulations where tissue slices had zero physical overlap, SpatialEx+ maintained high accuracy. This suggests the model learns general biological rules mapping morphology to molecules rather than simply memorizing specific tissue coordinates. Furthermore, the framework successfully processed datasets exceeding one million cells, proving its readiness for large-scale clinical and research applications.
Discussion and Future Directions
While SpatialEx and SpatialEx+ represent a significant leap forward in democratizing multi-omics, several challenges remain. Currently, the model assumes similar spatial resolutions across different omics layers; future iterations will need to address "cross-resolution" integration (e.g., merging single-cell data with lower-resolution spot-based data). Additionally, addressing batch effects—variations in H&E staining across different laboratories—will be essential for clinical deployment.
The ultimate trajectory for this work is the extension from 2D pairs to 3D reconstruction, potentially allowing for the creation of fully integrated, high-parameter molecular maps of entire organs from serial histological sections.
Summary updated