A new artificial intelligence framework is giving researchers a sharper way to read the spatial language of cells, connecting molecular activity inside individual cells with the neighborhoods they form in tissues. Called SpatialFormer, the hybrid model combines convolutional neural networks and transformer architectures to learn how gene expression is organized across multiple biological scales, from subcellular locations to multicellular landscapes. The study, published in Nature Computational Science, presents the system as a general-purpose representation-learning platform for spatial biology, a rapidly expanding field that aims to understand not only which genes are active, but also where that activity occurs and how neighboring cells influence one another.
Traditional gene-expression analysis often treats cells as isolated collections of molecular measurements. Spatial technologies such as Xenium add a crucial layer of information by recording the physical positions of RNA molecules within tissue sections. Yet this richer data creates a computational challenge: a cell may contain thousands of transcripts distributed across different compartments, while its behavior can also depend on nearby immune, epithelial, stromal or malignant cells. SpatialFormer was designed to integrate these signals instead of analyzing them separately. Its objective is to produce a compact numerical representation of each cell that reflects both its molecular identity and its location within a tissue niche.
The model combines two complementary forms of machine learning. Convolutional networks are well suited to recognizing local patterns, such as the arrangement of transcripts within a cell or the immediate organization of neighboring cells. Transformers, by contrast, use attention mechanisms to evaluate relationships across a broader context. In SpatialFormer, these capabilities are used to capture information at both subcellular and multicellular scales. The result is intended to preserve fine-grained spatial gene-expression patterns while also representing the larger cellular environment in which those patterns acquire biological meaning.
To train the system, the researchers used a pairwise strategy based on relationships between cells. Rather than learning from a single cell in isolation, the model was exposed to cell pairs and their associated spatial and expression information. This approach allows it to distinguish cellular features that are intrinsic to a cell from signals that emerge through proximity or shared niche context. During pretraining, SpatialFormer processed approximately 700 million cell pairs drawn from 17 million spatially resolved single cells across 71 Xenium slides. The scale of this dataset is central to the project: exposure to diverse tissues and cellular arrangements can help the model learn reusable biological patterns rather than memorizing one experimental sample.
The training data came from Xenium spatial transcriptomics, a technology that detects selected RNA molecules while retaining their coordinates inside tissue. In practical terms, this means that researchers can ask whether a gene is concentrated near the nucleus, enriched toward a cell boundary, or distributed in a pattern associated with a particular cellular state. They can also examine whether cells expressing complementary or competing molecular programs are positioned next to one another. SpatialFormer uses these measurements as more than a list of gene counts, treating them as structured spatial signals that can be translated into representations for downstream analysis.
The authors report that these learned representations supported several important tasks in single-cell and spatial biology. In batch correction, the model helped reduce technical differences between experiments while retaining biological variation, an essential step when datasets are generated on different slides, in different laboratories or under different conditions. In cell-type annotation, the representation provided molecular and contextual information that can assist in assigning identities to cells. The framework was also used for co-localization detection, identifying cell populations that repeatedly occupy the same tissue regions or appear to participate in shared microenvironments.
Co-localization is particularly significant because physical proximity can provide clues about communication, competition or coordinated function. A neighboring relationship alone does not prove that two cell types directly interact, but it can identify tissue regions where signaling is biologically plausible. By incorporating gene expression and spatial arrangement together, SpatialFormer may help researchers move beyond maps that simply label cells toward models that explain why particular cell types gather in specific locations. This could be valuable in tissues where disease progression depends on complex local interactions rather than on the behavior of one cell population alone.
The study also used perturbation analysis to investigate which molecular signals may be especially important in disease-associated cellular relationships. In pulmonary fibrosis, the analysis identified gene pairs linked to immune cell–cell communication. In breast cancer, it highlighted signals associated with epithelial–myoepithelial co-localization and tumor transition states. These findings suggest that the model can be used to computationally test the importance of individual genes or gene pairs within a spatial network. Such perturbation analyses do not automatically establish clinical causation, but they can prioritize molecular interactions for laboratory experiments and potentially reveal mechanisms that would be difficult to detect from expression levels alone.
SpatialFormer’s broader promise lies in its attempt to create a common language for spatial biology. Researchers currently work with datasets that differ in tissue type, measurement technology, resolution and experimental design. A representation learned from millions of spatially resolved cells could make it easier to compare these datasets and transfer knowledge between studies. If successful across additional tissues and platforms, systems of this kind could support the discovery of disease-specific niches, improve the interpretation of tissue biopsies and help identify cellular interactions that influence treatment response. The framework does not replace biological validation, but it offers a high-dimensional map of cellular relationships that may guide the next generation of experiments in immunology, fibrosis and cancer research.
Subject of Research: Spatial representation learning for single-cell multimodal and multiscale gene expression, cellular niches and spatial cell–cell interactions.
Article Title: SpatialFormer: universal spatial representation learning from subcellular molecular to multicellular landscapes
Article References: Wang, J., Huang, Y. & Winther, O. SpatialFormer: universal spatial representation learning from subcellular molecular to multicellular landscapes. Nat Comput Sci (2026). https://doi.org/10.1038/s43588-026-01016-7
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s43588-026-01016-7
Keywords: SpatialFormer, spatial transcriptomics, single-cell biology, gene expression, cellular niches, transformers, convolutional neural networks, Xenium, cell–cell communication, pulmonary fibrosis, breast cancer, tumor biology
