Spatial Transcriptomic Atlas

Unlike localized studies, a true atlas achieves whole-organ or whole-organism coverage by integrating thousands of serial sections or large tissue arrays. This requires computational spatial registration to align individual capture areas into a unified 3D coordinate system. The goal is to eliminate sampling bias and capture the full spectrum of tissue heterogeneity, from rare stem cell niches to broad parenchymal zones.

A canonical atlas co-registers spatial transcriptomics with complementary modalities from adjacent sections, such as spatial proteomics (e.g., CODEX, MIBI) and histology (H&E staining). This fusion anchors molecular profiles to classical anatomical landmarks. Computational pipelines for spatial multi-omics integration resolve the distinct resolutions of each modality, creating a single latent representation that defines tissue architecture by both gene and protein expression.

Atlases are organized hierarchically using unsupervised spatial domain detection algorithms. Graph neural networks or spatial hidden Markov models identify coarse anatomical regions (e.g., cortical layers) and fine-grained spatial niches (e.g., perivascular microenvironments). This structure allows users to query the atlas at multiple scales, from gross anatomy down to recurrent cellular communities defined by ligand-receptor co-localization patterns.

Every feature in an atlas is statistically validated. Spatially variable genes (SVGs) are identified using spatial autocorrelation metrics like Moran's I and Ripley's K function, distinguishing true biological gradients from random noise. Spatial permutation tests generate null distributions by shuffling spatial labels, ensuring that reported domain-specific expression patterns are statistically significant and reproducible across biological replicates.

Building an atlas often requires merging data from multiple laboratories and technologies (e.g., Visium, MERFISH, Slide-seq). Spatial batch correction algorithms remove technical variation while preserving true biological spatial heterogeneity. This harmonization ensures that a gene's expression gradient in one sample is directly comparable to the same gradient in another, creating a unified reference that transcends individual experimental platforms.

An atlas captures dynamic processes by applying spatial trajectory inference to ordered tissue locations. Rather than relying on pseudotime from dissociated cells, this approach reconstructs differentiation or migration paths along physical axes. The resulting vector fields map the flow of gene expression change across anatomical space, revealing how progenitor zones give rise to differentiated functional units in situ.

The foundational capture technology for many atlas-scale projects. Spatially indexed oligonucleotide arrays are placed on a slide, and polyadenylated mRNA from overlying tissue sections binds to them. Each spot's barcode encodes its unique spatial coordinates, allowing all captured transcripts to be mapped back to their tissue of origin after sequencing. This enables unbiased, genome-wide discovery without prior target selection.

A critical computational step for atlas construction when using multi-cellular capture spots. Each spot's expression profile is a mixed signal from multiple cell types. Deconvolution algorithms estimate the proportion of each constituent cell type using a reference single-cell RNA-seq signature matrix. Key methods include:

• RCTD (Robust Cell Type Decomposition): Models platform effects for accurate transfer.
• cell2location: A Bayesian hierarchical model that resolves fine-grained cell-type differences.
• Stereoscope: Uses negative binomial distributions for rate estimation.

The computational alignment of multiple tissue sections, or a tissue image with its spatial data, into a common coordinate system. For a whole-organ atlas, this is essential to integrate serial sections. Registration uses affine or elastic transformations based on landmark features, such as tissue edges or fluorescent fiducial markers, to warp images so that corresponding anatomical structures overlap perfectly, enabling cross-modality comparison.

The unsupervised identification of functionally coherent tissue regions that share similar gene expression profiles and histology. Unlike manual annotation, these methods are data-driven. Common approaches include:

• BayesSpace: Uses a spatial hidden Markov model to encourage neighborhood smoothness.
• SpaGCN: A graph convolutional network that integrates gene expression, spatial location, and histology.
• stLearn: Extracts morphological features from H&E images to guide clustering.

A computational analysis that infers cell-to-cell communication within the atlas. It identifies spatially proximal cell-type pairs where a ligand gene in one cell type and its cognate receptor gene in another are co-expressed. Tools like CellPhoneDB and NicheNet use curated databases of molecular interactions to predict active signaling pathways, revealing the tissue's functional wiring diagram.

A method for reconstructing dynamic biological processes directly in their spatial context. Instead of ordering cells purely by transcriptomic similarity, these algorithms constrain the inferred pseudotime by physical distance. This is crucial for understanding processes like tumor invasion or developmental patterning, where a cell's position is a direct readout of its history. Tools like SpaceFlow use deep graph networks to identify spatiotemporal patterns.