Spatial Omics Integration

The computational process of registering spatially barcoded gene expression spots onto corresponding high-resolution histology images. This alignment enables the direct correlation of transcriptomic profiles with morphological features such as tumor boundaries, necrotic zones, and immune infiltrates. Algorithms solve a non-rigid co-registration problem by identifying fiducial markers and tissue landmarks visible in both modalities, often using iterative closest point (ICP) or deep learning-based feature matching to achieve sub-cellular precision.

The integration of heterogeneous data types—spatial transcriptomics, multiplex immunofluorescence, and H&E morphology—into a unified analytical framework. This fusion leverages canonical correlation analysis (CCA) and multi-modal autoencoders to learn a shared latent representation that captures both molecular and morphological tissue states. The resulting joint embedding space enables cross-modality prediction, such as inferring gene expression from histology images alone using models like HistoGene or SEQUOIA.

The quantitative characterization of cellular microenvironments by analyzing the composition and spatial arrangement of cell types within defined tissue regions. Techniques include:

• Spatial autocorrelation metrics (Moran's I, Geary's C) to detect gene expression gradients
• Cell-cell interaction networks built via Delaunay triangulation or k-nearest neighbor graphs
• Niche identification using graph neural networks to discover recurrent multicellular structures predictive of immunotherapy response

Statistical methods that identify genes with location-dependent expression changes rather than global condition-specific differences. Unlike bulk differential expression, these approaches model spatial covariates using Gaussian process regression or spatialDE frameworks to distinguish genuine spatial patterns from technical noise. This reveals gradient genes that vary smoothly across tissue zones and hotspot genes concentrated in specific anatomical compartments, critical for understanding tumor-immune interfaces.

Algorithms that estimate cell-type proportions within each spatial capture spot, which typically contains multiple cells. Methods like RCTD (Robust Cell Type Decomposition) and SPOTlight integrate single-cell RNA-seq reference data with spatial transcriptomics to infer the cellular composition at each tissue coordinate. This transforms spot-level data into cell-type-resolved spatial maps, enabling the study of colocalization patterns between tumor clones and specific immune subsets.

The volumetric assembly of serial tissue sections with aligned spatial omics data to reconstruct three-dimensional molecular architectures. This requires inter-section registration algorithms that compensate for tissue deformation, tearing, and staining variability across consecutive slices. The resulting 3D spatial atlas reveals molecular gradients along anatomical axes—such as perivascular immune niches or depth-dependent hypoxia signatures—that are invisible in single 2D sections.

A family of molecular profiling methods that quantify gene expression within intact tissue sections, preserving the two-dimensional spatial coordinates of each transcript. Unlike single-cell RNA sequencing, which loses spatial context during dissociation, spatial transcriptomics maps expression patterns directly onto histology.

• Visium (10x Genomics): Captures polyadenylated mRNA on spatially barcoded slides with 55 μm spot resolution
• MERFISH: Multiplexed error-robust fluorescence in situ hybridization achieving single-molecule resolution
• Slide-seq: Transfers RNA onto a puck of DNA-barcoded beads for near-cellular resolution
• Stereo-seq: Nanoscale-resolution capture using randomly barcoded DNA nanoball arrays

The computational alignment of spatially-resolved molecular data with corresponding histology images, enabling direct correlation between gene expression patterns and morphological tissue features. This process is foundational to spatial omics integration.

• Landmark-based registration: Uses manually or automatically detected fiducial points to compute affine or elastic transformations
• Intensity-based registration: Optimizes mutual information or cross-correlation between image modalities
• Non-rigid deformable registration: Accounts for tissue distortion during molecular processing using B-spline or diffeomorphic models
• Multi-modal fusion: Aligns H&E, IHC, and spatial transcriptomics into a common coordinate frame

The deep learning task of identifying individual cell boundaries within tissue images, a critical preprocessing step for assigning spatial molecular profiles to specific cells. Modern approaches combine nuclear detection with membrane staining or computational expansion.

• Hover-Net: Simultaneously segments and classifies nuclei using horizontal and vertical gradient predictions to separate touching instances
• Cellpose: A generalist segmentation model using spatial gradients and a style-aware architecture adaptable to diverse cell morphologies
• StarDist: Predicts star-convex polygon representations for each nucleus, guaranteeing non-overlapping segmentations
• Mesmer: A multiplexed image segmentation pipeline that combines nuclear and membrane channels for whole-cell delineation

Computational methods that characterize the local cellular microenvironment by analyzing the composition and organization of cells within defined spatial windows around each cell or region of interest.

• Cellular neighborhoods: Clusters of recurrent multi-cellular spatial patterns identified using k-means or graph-based community detection
• Spatial autocorrelation: Moran's I or Geary's C statistics quantify whether gene expression or cell types are clustered, dispersed, or random
• Niche identification: Graph neural networks learn latent representations of tissue microenvironments by aggregating features from neighboring cells
• Distance-based enrichment: Permutation testing determines if specific cell-type pairs co-localize more frequently than expected by chance

The computational inference of cell-cell communication by identifying co-expression of ligand and receptor gene pairs between spatially adjacent cells. This transforms spatial omics data into functional interaction networks.

• CellChat: Infers biologically meaningful communication by modeling ligand-receptor interactions with cofactors and signaling pathway databases
• NicheNet: Predicts which sender cells drive gene expression changes in receiver cells using prior knowledge of intracellular signaling
• Spatial-constrained analysis: Restricts inferred interactions to cells within a plausible diffusion distance, reducing false positives from dissociated data
• Commot: Uses optimal transport to model spatial signaling gradients and cumulative ligand effects across tissue domains

Computational methods that resolve the cellular composition of spatial transcriptomics spots, which often capture multiple cells. Deconvolution estimates the proportion of each cell type contributing to the mixed expression signal at each spatial location.

• RCTD (Robust Cell Type Decomposition): Uses a reference single-cell RNA-seq signature matrix and models platform-specific technical effects
• Stereoscope: A probabilistic model that jointly estimates cell-type proportions and spatial gene expression profiles
• SPOTlight: Seeded non-negative matrix factorization initialized with cell-type marker genes from reference data
• Tangram: Deep learning framework that maps single-cell transcriptomes onto spatial coordinates by optimizing gene expression similarity