Spatial Omics Technologies

• Spatial omics technologies are a suite of experimental and computational approaches that profile diverse molecular states while preserving spatial context in tissues.
• They combine imaging-based and barcoding methods to achieve subcellular to spot-level resolution and whole-transcriptome or proteome coverage.
• Advanced computational frameworks, including graph neural networks and deep learning, enable spatial domain segmentation, cell-type deconvolution, and integration of multi-omic data.

Spatial omics technologies comprise a suite of experimental and computational approaches that simultaneously profile molecular states—transcriptomic, proteomic, epigenomic, and metabolic—while explicitly preserving the spatial coordinates of individual cells or tissue regions. These modalities enable direct inference of cellular organization, microenvironmental heterogeneity, and spatially encoded regulatory programs that underlie physiological and pathological processes. Spatial omics bridges the gap between dissociated single-cell measurements and intact tissue context, allowing for quantitative dissection of cellular architectures, communication networks, and evolutionary dynamics across biological scales.

1. Experimental Modalities and Technological Principles

Spatial omics platforms fall into four principal experimental classes, each with distinct trade-offs in spatial resolution, molecular coverage, and throughput:

1. Imaging-based in situ hybridization and sequencing: Single-molecule FISH (smFISH), seqFISH+, MERFISH, ExSeq, and STARmap directly hybridize fluorescently labeled probes or amplify nucleic acids in fixed tissue sections, with multiplexing enabled via iterative barcoding and imaging. These methods achieve subcellular (70–200 nm) resolution, enabling mapping of up to tens of thousands of transcripts per cell, and in some protocols, 3D spatial reconstruction. Representative throughput ranges from 100 (classic MERFISH) to over 10,000 genes (seqFISH+) (Jena et al., 2023, Fan et al., 16 Sep 2025).
2. Spatial barcoding and sequencing-based transcriptomics: Approaches such as 10x Visium, Slide-seqV2, HDST, Seq-Scope, and Stereo-seq employ patterned arrays of DNA-barcoded spots, beads, or nanoballs to capture bulk- or single-cell-level mRNA. RT and next-generation sequencing yield spatially indexed transcriptomes with spot sizes ranging from ∼50 μm (Visium) down to submicron (Seq-Scope, Stereo-seq), typically affording whole-transcriptome coverage (>20,000 genes). Principal strengths are high throughput (mm²–cm² field-of-view) and unbiased transcriptome-wide profiling; limitations include spatial averaging and technical noise due to capture/registration (Liu et al., 2021, Fan et al., 16 Sep 2025).
3. Spatial proteomics: Multiplexed fluorescence/immunohistochemistry (CODEX, IBEX, CycIF, CosMx SMI) and imaging mass cytometry (IMC, MIBI, MALDI-IMS) cyclically label tissues with panels of antibodies or tags, achieving single-cell to subcellular resolutions over 40–100 protein markers. Label-free mass spectrometry methods (Deep Visual Proteomics) afford single-cell proteome coverage, but with coarser spatial mapping (10–100 μm). Panel design and detection noise are key considerations (Fan et al., 16 Sep 2025, Isik et al., 18 Jan 2026).
4. Spatial epigenomics and metabolomics: Chromatin accessibility (spatial ATAC-seq, CUT&Tag) and DNA methylation can be mapped at ∼20 μm–cellular resolution via combinatorial barcoding or microfluidics, while spatial metabolome imaging (DESI, MALDI-FTICR, nano-DESI) detects hundreds to thousands of metabolites per voxel, with 10–100 μm spatial granularity. Integrated multi-omic chemistries (e.g., DBiT-seq, sciMAP-ATAC) are emerging for same-section profiling (Jena et al., 2023, Fan et al., 16 Sep 2025, Isik et al., 18 Jan 2026).

These technologies provide the basis for comprehensive spatial mapping of molecular phenotypes across biological systems.

2. Computational Frameworks and Statistical Principles

Spatial omics data require specialized computational frameworks to extract biologically meaningful patterns from high-dimensional, spatially indexed measurements:

• Spatial domain segmentation and clustering: Methods for spatial clustering, such as SpaGCN, STAGATE, SemanticST, and BNPMFA, build spot/cell graphs using spatial coordinates and molecular features, integrating deep graph neural networks (GNNs) or Markov random field (MRF) priors to enforce spatial contiguity and discover tissue domains. Probabilistic and Bayesian approaches (e.g., BNPMFA) directly infer spatial domains and cluster number without ad hoc dimension reduction (Zhu et al., 2024, Zahedi et al., 13 Jun 2025).
• Detection of spatially variable genes: SpatialDE, SPARK, and Moran’s I compute global or local measures of spatial autocorrelation in gene expression, leveraging Gaussian process or generalized linear spatial models. Cluster-based and deep learning models (SEDR, SpaGCN) integrate spatial smoothness constraints into latent representations for subsequent gene variability assessment (Liu et al., 2021, Nasab et al., 2022).
• Cell-type deconvolution and mapping: For lower-resolution platforms, methods such as DOT, Stereoscope, Cell2location, and SpaRTaCo align spatial transcriptomics with single-cell references to estimate cell-type composition in each spot by optimizing multivariate transfer objectives or co-clustering blocks. Optimal transport (Tangram, DOT) aligns cells and spatial locations subject to spatial and expression constraints (Rahimi et al., 2023, Sottosanti et al., 2021).
• Integration of multi-omic and imaging data: Multimodal data fusion is enabled by joint nonnegative matrix factorization (LIGER, SPOTlight), variational autoencoders (MOFA+, MEFISTO), and deep learning frameworks (PRAGA, MUSE, CellScape), which learn shared low-dimensional representations while denoising technical artifacts and reconciling scale/resolution differences. Cross-modal graph adaptation and dynamic prototype learning (PRAGA) address cell type and annotation sparsity (Huang et al., 2024, Isik et al., 18 Jan 2026, Yan et al., 13 Feb 2026).
• Inference of cell–cell interactions and spatial communication: Statistical methods decompose expression variance into intrinsic and context-dependent components (SVCA), while spatial graphs and permutation testing identify ligand–receptor signaling and spatially coordinated gene programs. Deep graph models directly infer interaction networks by learning task-specific edge weights (Jena et al., 2023, Nasab et al., 2022, Yan et al., 13 Feb 2026).
• Visualization and spatial summary statistics: Lattice-based and point-pattern statistics (Ripley’s K, Moran’s I, variograms, marked point process models) implemented via flexible software frameworks (e.g., pasta) quantify spatial clustering, domain structure, and spatial heterogeneity across data modalities (Emons et al., 2024).

These computational pipelines are increasingly modular, scale-aware, and optimized for contemporary high-resolution spatial omics datasets.

3. Machine Learning and Deep Learning Architectures

Recent advances leverage deep learning for robust modeling and interpretation of spatial omics data:

• Graph Neural Networks (GNNs) dominate spatial domain segmentation in both unimodal and multi-modal settings, learning node embeddings via kNN or adaptive graphs that encode proximity, expression similarity, and domain structure. SemanticST, CellScape, and PRAGA extend classic GNNs with multi-semantic graphs, disentanglement objectives, and prototype-aware contrastive learning, enabling recovery of biologically meaningful tissue architectures at massive scale (up to 500,000 cells) (Zahedi et al., 13 Jun 2025, Yan et al., 13 Feb 2026, Huang et al., 2024).
• Autoencoders and Variational Autoencoders (AEs/VAEs) are widely used for dimensionality reduction and integration of diverse omic and imaging modalities (SEDR, MUSE, MOFA+). VAEs with negative-binomial or zero-inflated decoders explicitly address overdispersion and dropout in gene expression data (Nasab et al., 2022, Isik et al., 18 Jan 2026).
• Attention Mechanisms and Transformers: Self-attention modules (e.g., in HisToGene, SpaFoundation, Nicheformer) capture long-range spatial, molecular, and morphological dependencies, supporting domain recovery, imputation, and cell-type annotation. Vision transformer (ViT) backbones, contrastive multimodal pretraining (CLIP style), and masked token prediction losses are increasingly central in emergent foundation models for spatial omics (Hao et al., 12 Jan 2026, Noorbakhsh et al., 30 Jun 2025).
• Deep generative and contrastive paradigms are applied for cross-modality mapping (GANs for virtual proteomics), imputation, and robust transfer learning. Hybrid models (CellScape) fuse CNN-based encoders for intrinsic genomics with GAT-based encoders for spatial context, calibrated by multi-task and contrastive objectives (Jiang et al., 5 Feb 2026, Yan et al., 13 Feb 2026).

These architectures simultaneously address scale, noise, unknown cluster structure, and modality heterogeneity across spatial omic datasets.

4. Applications and Biological Insights

Spatial omics technologies yield novel insights across key biological systems:

• Tissue architecture and developmental patterning: High-resolution spatial transcriptomic and epigenomic maps delineate canonical cortical layers, hippocampal domains, liver zonation, and embryonic patterning, revealing transcriptomic and chromatin gradients underlying morphogenesis. CellScape and SemanticST recover laminar organization and region-specific marker gene localization (Zahedi et al., 13 Jun 2025, Yan et al., 13 Feb 2026).
• Cancer biology and tumor microenvironment (TME): Spatial omics uncovers niche structures (e.g., triple-receptor-positive microdomains in breast cancer), invasion fronts, and immune cell spatial networks. AI-driven approaches dissect TME communication channels, response to therapy, and rare subpopulations underlying clinical outcomes (Noorbakhsh et al., 30 Jun 2025, Fan et al., 16 Sep 2025). For example, SemanticST identifies non-canonical EMT programs and spatially distinct transition zones in breast tumor tissues (Zahedi et al., 13 Jun 2025).
• Cell-type discovery and domain segmentation: Adaptive graph-based methods (PRAGA, BNPMFA) and STProtein reveal novel cellular clusters (e.g., new macrophage subsets in the spleen) and fine-grained domain boundaries, including domains invisible to unimodal or nonspatial analyses (Huang et al., 2024, Jiang et al., 5 Feb 2026, Zhu et al., 2024).
• Spatial multi-omic mapping and integration: Joint profiling of transcriptomes, proteomes, and chromatin accessibility enables inference of signaling circuits, identification of spatially regulated protein “dark matter,” and causal testing of clonal dynamics or rare cellular states (Isik et al., 18 Jan 2026, Jiang et al., 5 Feb 2026, Jena et al., 2023).

These insights are transforming understanding of multicellular systems, pathogenesis, and therapeutic response.

5. Practical Trade-offs, Challenges, and Limitations

Spatial omics analyses involve key technical and conceptual challenges:

• Resolution vs. molecular coverage: Imaging-based methods achieve subcellular precision at the cost of panel size, whereas sequencing-based barcoding yields genome-wide data at coarser (spot-level) resolution. Multi-omic and true single-cell co-profiling remain limited in throughput and coverage (Fan et al., 16 Sep 2025, Isik et al., 18 Jan 2026).
• Scalability and computational complexity: Deep graph models, Gaussian processes, and MCMC inference scale superlinearly with either cell/spot count or gene dimensionality. Methodological advances (mini-batch GNNs in SemanticST, scalable inference in BNPMFA, efficient Frank–Wolfe optimization in DOT) address large-scale tissue mapping, but terabyte-scale data and 3D/4D atlases pose persistent barriers (Zahedi et al., 13 Jun 2025, Rahimi et al., 2023, Zhu et al., 2024).
• Batch effects, technical variance, and data normalization: Spatial autocorrelation complicates batch effect removal and QC; normalization methods must accommodate overdispersion, zero inflation, and differing detection characteristics across omic layers (Nasab et al., 2022, Camacho et al., 2024).
• Interpretability, annotation, and benchmark standards: Many deep learning architectures are “black box”; robust benchmarking, standardized pipelines (SpatialData, pasta), and unsupervised model selection remain open areas of method development (Emons et al., 2024, Zhu et al., 2024).
• Integration of imaging, omics, and 3D structure: Co-registration, normalization, and analysis of serial tissue sections, histology, and multi-modal omics require unified data models, flexible graph representations, and scalable analytics (Hao et al., 12 Jan 2026, Isik et al., 18 Jan 2026).

Future progress will depend on harmonized data standards, shared benchmarks, and development of theory-guided, multiscale, and interpretable learning approaches.

6. Emerging Directions and Foundation Models

Cutting-edge work is extending spatial omics in several directions:

• Foundation models and transfer learning: Large-scale, cross-tissue pretraining of vision transformers and multimodal contrastive models (Nicheformer, OmiCLIP, PAST) provide transferable representations for downstream spatial inference, imputation, and domain adaptation (Hao et al., 12 Jan 2026, Noorbakhsh et al., 30 Jun 2025).
• Integrated and universal multi-omic prediction: Novel frameworks (STProtein, PRAGA) leverage multi-task, prototype-aware, and cross-modal learning to transduce spatial transcriptomic data into proteomic or epigenomic predictions, bridging experimental bottlenecks (Jiang et al., 5 Feb 2026, Huang et al., 2024).
• Hybrid data-driven and mechanistic models: Development is ongoing in integrating biophysical, information-theoretic, and AI paradigms to enable causal mapping, spatial modeling of tissue mechanics, and dynamic (3D/4D) cell fate inference (Noorbakhsh et al., 30 Jun 2025, Jena et al., 2023).
• Software and statistical toolkit maturation: Adoption of comprehensive spatial statistics packages (pasta, Squidpy) and structured workflow frameworks (SpatialExperiment, SingleCellExperiment) is standardizing analysis and reporting, while still enabling scale- and task-appropriate metric selection (Emons et al., 2024).

Continued innovation in algorithmic frameworks, data integration, and mechanistic interpretation will expand the impact of spatial omics technologies across cell biology, oncology, neurodevelopment, and precision medicine.