Spatial Omics and Tissue Models
Spatial omics adds tissue position to molecular measurement. The modelling problem changes because cell state, neighbourhood, morphology, and tissue architecture now influence interpretation. Platforms differ in resolution and sensitivity: Visium is spot-level, MERFISH and Xenium are imaging-based, CosMx is subcellular sequencing, and Stereo-seq is high-resolution sequencing. SpatialData provides a data-frame layer for storing spatial assays without erasing platform-specific geometry, while Nicheformer and Novae show how foundation-model approaches are moving from cell-state embeddings toward spatial-context embeddings (Marconato et al., 2024; Tejada-Lapuerta et al., 2025; Blampey et al., 2025). Spatial AI is most useful when it links molecular signals to tissue structure with clear resolution limits and explicit platform-aware comparisons. Assigning cell-type labels to spots is necessary but not sufficient.
Use this chapter to:
- Place molecular measurements back into tissue architecture so cell state, neighborhood, and spatial organization can be studied together.
- Platform physics matters: spot size, capture efficiency, segmentation, imaging modality, tissue processing, and coordinate systems shape every downstream claim.
Prerequisites: Single-Cell Foundation Models for the dissociated-cell modelling context; Cell Painting and Image-Based Phenotyping for image-based phenotyping background.
Summary: Place molecular measurements back into tissue architecture so cell state, neighborhood, and spatial organization can be studied together. Spatial AI is advancing through shared data structures and tissue-aware models, but arbitrary tissue-response simulation remains early.
Key point: Platform physics matters: spot size, capture efficiency, segmentation, imaging modality, tissue processing, and coordinate systems shape every downstream claim. Open question: whether tissue-aware models generalize across platforms, disease states, and spatial scales.
Bottom line: Spatial omics connects single-cell biology to histopathology, organoids, systems biology, tumor biology, neuroscience, and biomarker translation.
What is this field trying to solve? Place molecular measurements back into tissue architecture so cell state, neighborhood, and spatial organization can be studied together.
What is the core idea? Platform physics matters: spot size, capture efficiency, segmentation, imaging modality, tissue processing, and coordinate systems shape every downstream claim.
What is the current state of the field? Spatial AI is advancing through shared data structures and tissue-aware models, but arbitrary tissue-response simulation remains early.
What do we know, and what remains open? Known reference points include Visium, MERFISH, CosMx, Xenium, Stereo-seq, SpatialData, Nicheformer, Novae, spatial deconvolution methods, and histology-linked datasets. What remains open is whether tissue-aware models generalize across platforms, disease states, and spatial scales.
Why does this matter? Spatial omics connects single-cell biology to histopathology, organoids, systems biology, tumor biology, neuroscience, and biomarker translation.
Introduction
Spatial transcriptomics and single-cell technologies have produced datasets that connect cell identity and tissue architecture. A Nature Reviews Molecular Cell Biology review describes advances and challenges in characterising cell states and multicellular neighbourhoods, including deep learning methods (Gulati et al., 2025).
Spatial omics is not one assay. It spans barcoded arrays that preserve tissue coordinates (Ståhl et al., 2016), single-molecule imaging approaches such as MERFISH (Chen et al., 2015), high-plex spatial molecular imaging of RNA and protein (He et al., 2022), and high-resolution sequencing-based systems such as Stereo-seq (Chen et al., 2022). The biological question determines which tradeoff is acceptable: transcriptome breadth, cell boundary confidence, subcellular localisation, tissue area, protein co-measurement, or compatibility with archival tissue.
Platform physics sets the ceiling
Spatial AI begins with platform physics. Spot-based methods trade molecular breadth and tissue area against cell-level precision. Imaging-based methods can reach subcellular localization for targeted panels, but the panel limits discovery. Sequencing-based high-resolution methods improve density, but sample processing and segmentation remain decisive. A model cannot recover information the assay never measured.
This is why platform-aware reporting belongs in the first paragraph of any spatial result: tissue type, preservation method, platform, resolution, panel or transcriptome breadth, image channel, segmentation method, coordinate system, and normalization procedure. Without those details, “spatial model” is too broad to evaluate.
Cell-cell communication is an inference, not a measurement
Spatial data are often used to infer ligand-receptor interactions, niches, and tissue neighborhoods. These are valuable hypotheses, but most are not direct measurements of signaling. Co-location of a ligand-expressing cell and receptor-expressing cell does not prove contact, ligand availability, receptor activation, downstream signaling, or functional consequence.
The strongest spatial communication claims therefore combine spatial proximity, expression, protein or phospho-protein evidence when available, perturbation or natural experiment, and biological plausibility. Static tissue snapshots are weaker when the claim is dynamic response.
Histology and molecular data answer different questions
Histology captures architecture, morphology, and pathologist-recognizable tissue patterns. Spatial transcriptomics captures molecular state in coordinate space. Combining the two is useful because tissue context and molecular state constrain each other, but the labels should stay distinct. A histology-derived tumor region, a transcriptomic cluster, and an inferred neighborhood are different evidence objects.
What is demonstrated?
Demonstrated capability includes cell-state mapping, neighbourhood analysis, tissue segmentation support, spatially aware representation learning, and multimodal alignment between histology and gene expression in bounded settings. The spatial transcriptomics review documents the increasing role of deep learning in single-cell and spatial data analysis (Gulati et al., 2025). SpatialData demonstrates the infrastructure side: spatial omics needs a data framework that preserves images, shapes, labels, molecular tables, and coordinate systems together (Marconato et al., 2024). Nicheformer demonstrates that a foundation model trained on SpatialCorpus-110M can encode spatial context for downstream niche, region, and composition tasks (Tejada-Lapuerta et al., 2025). Novae adds a graph-based spatial foundation model with zero-shot spatial-domain inference and spatial-context embeddings (Blampey et al., 2025).
Platform-Aware Measurement
The platform determines the claim. Visium-style spot data can support tissue-region and cell-mixture inference, but each spot can contain multiple cells. MERFISH and spatial molecular imaging support subcellular localisation for targeted panels, but panel design constrains discovery. Stereo-seq and related sequencing-based methods increase spatial density and tissue coverage, but still require platform-specific normalisation and segmentation logic. A result that is valid for one platform should not be promoted as a general tissue-model result without replication across acquisition and processing conditions.
Single-Cell Mapping and Deconvolution
Most practical spatial AI work maps dissociated single-cell references into tissue coordinates or decomposes mixed measurements. Tangram aligns single-cell and spatial transcriptomes to impute and map cell states (Biancalani et al., 2021). cell2location estimates spatial cell-type abundance using single-cell references and spot-level spatial data (Kleshchevnikov et al., 2022). These methods support spatial annotation and neighbourhood analysis, but they inherit the reference atlas, tissue-processing, and resolution limits of the input data.
| Evidence Anchor | What It Supports | Practical Constraint |
|---|---|---|
| Spatial transcriptomics review (Gulati 2024) | Cell identity and tissue architecture as linked data | Resolution, sensitivity, and platform differences constrain inference |
| Primary platform papers | Measurement modality and resolution claims | Platform-specific chemistry and panels shape the biological signal |
| Tangram / cell2location | Mapping and deconvolution from single-cell references | Reference atlas quality limits spatial inference |
| SpatialData | Platform-aware spatial data storage and interoperability | Data structure does not solve biological normalization |
| Nicheformer | Spatially aware representation learning | Spatial pretraining helps spatial tasks; it does not make static data dynamic |
| Novae | Graph-based spatial-domain representation | Benchmark cautiously; domain inference is not tissue simulation |
| Cell Painting | Image-based phenotyping as cellular morphology data | Morphology is a proxy requiring biological validation |
| Single-cell foundation models | Cell-state representation learning | Dissociated cells lose native spatial context |
What is theoretical?
Theoretical capability includes tissue foundation models that represent histology, spatial transcriptomics, proteomics, and perturbations together. This is plausible, but dataset harmonisation, resolution mismatch, and tissue processing artefacts remain hard.
What is beyond current capability?
Beyond current capabilities includes reliable simulation of tissue response to arbitrary perturbations from static spatial data alone. Dynamic experiments and perturbational measurements are still required.
What would make this more promising?
The claim would strengthen if spatial models showed reproducible improvement across platforms, tissues, and laboratories while preserving platform geometry and resolution limits. Stronger evidence would include paired histology, spatial transcriptomics, protein or imaging validation, and perturbation or longitudinal data when claims involve signaling or response.
The claim would weaken if domain labels track platform, segmentation, section quality, or coordinate registration more than biology. Static-to-dynamic predictions should be downgraded unless matched perturbational evidence is available.
What should researchers, biotech teams, funders, and program leaders do with this?
- State the spatial resolution and whether measurements are spot-level, cell-level, or subcellular.
- Keep histology-derived labels separate from molecular labels.
- Validate inferred neighbourhoods with biological markers.
- Avoid comparing platforms without platform-aware normalisation.
- Track gene-panel size and sensitivity per platform when reporting findings.
- For tissue foundation model claims, expect independent reproduction before adopting in a research program.