Spatially Variable Genes (SVG)

An SVG is formally defined by a statistically significant departure from Complete Spatial Randomness (CSR) . This is not merely a difference in expression between arbitrary regions, but a quantifiable dependence on spatial coordinates. Detection relies on rejecting a null hypothesis where expression values are equally likely to be observed at any location, independent of their neighbors. Key statistical frameworks include:

• Spatial Autocorrelation Metrics: Statistics like Moran's I quantify the degree of clustering or dispersion.
• Gaussian Process Regression: Models expression as a continuous function over spatial coordinates to identify genes with non-zero spatial length scales.
• Permutation Tests: Randomly shuffling spatial labels to generate an empirical null distribution, against which the observed spatial pattern is tested.

SVGs are fundamentally distinct from Differentially Expressed Genes (DEGs) . A DEG is defined by a change in mean expression between pre-annotated groups (e.g., tumor vs. normal). An SVG is defined by its spatial pattern, which can manifest as a gradient, a patch, or a punctate distribution, often without requiring prior histological annotation. A gene can be an SVG without being a DEG, and vice versa. For example, a gene forming a sharp gradient across a histologically uniform tissue section is an SVG but not a DEG, as no discrete groups exist to compare. This makes SVG analysis a powerful tool for unsupervised tissue structure discovery.

SVGs are categorized by their spatial expression patterns, each with distinct biological implications:

• Clustered/Patchy Patterns: Indicate expression confined to specific cell types or anatomical microenvironments, such as a gene expressed only in a lymphoid follicle.
• Gradient Patterns: Reveal continuous morphogen-like signaling or developmental trajectories, such as a gene whose expression increases linearly from the outer cortex to the inner medulla of an organ.
• Punctate/Dispersed Patterns: Often signify expression in rare, scattered cells like infiltrating immune cells or specific neuronal subtypes.
• Laminar Patterns: Common in brain and retinal tissues, where expression aligns with defined anatomical layers.

A suite of computational tools has been developed to identify SVGs, each with different statistical assumptions:

• SpatialDE: Uses Gaussian process regression to decompose expression variance into spatial and non-spatial components.
• SPARK-X: A rapid, non-parametric method that directly tests the dependence of expression on spatial location using a generalized spatial covariance test, without assuming a specific expression distribution.
• trendsceek: Identifies SVGs by testing for significant spatial trends using marked point process statistics, such as Ripley's K function.
• Moran's I-based methods: Directly compute the spatial autocorrelation statistic for each gene, often using a spatial neighborhood graph to define proximity.

SVGs are the primary input for defining spatial domains and cellular niches. By clustering genes with similar spatial patterns, or by clustering spatial locations based on their SVG expression profiles, researchers can identify functionally coherent tissue regions de novo. This process, often implemented with Spatial Graph Neural Networks or Spatial Hidden Markov Models, reveals the tissue's functional architecture. For instance, co-expression of a set of immune-related SVGs can define an area of active inflammation, while a different set of SVGs defines a fibrotic capsule, all without prior knowledge of these structures.

The definition and detection of an SVG are intrinsically linked to the spatial resolution of the assay. A gene may appear as an SVG at single-cell resolution but not at a multi-cellular spot level where its signal is averaged out. Conversely, a gene with a broad, tissue-level gradient is best detected by lower-resolution, capture-based technologies like Visium. High-resolution techniques like MERFISH or Xenium are required to identify SVGs with subcellular or rare-cell punctate patterns. Therefore, an SVG is not an absolute property of a gene but a property of a gene measured at a specific spatial scale.

A statistical measure of the degree to which a variable's values at nearby locations in a tissue are more similar than expected by random chance. It is the foundational concept behind SVG detection.

• Positive autocorrelation: Similar expression levels cluster together.
• Negative autocorrelation: Dissimilar values are adjacent (rare in biology).
• Zero autocorrelation: A random spatial pattern with no spatial dependency.

SVG detection methods are essentially formal tests for the presence and strength of spatial autocorrelation in gene expression vectors.

A global spatial autocorrelation statistic that measures the overall clustering of a gene expression pattern across a tissue. It produces a single summary value ranging from -1 (dispersed) to +1 (highly clustered), with 0 indicating randomness.

• Compares the expression at each spot to the weighted average of its neighbors.
• Requires a pre-defined spatial weights matrix defining adjacency.
• Statistical significance is assessed via a permutation test.

Moran's I is one of the oldest and most interpretable methods for identifying SVGs, though it assumes stationarity across the tissue.

A non-parametric statistical test that randomly shuffles spatial labels to generate a null distribution for assessing the significance of observed spatial patterns.

• Null hypothesis: The observed spatial pattern is no different from random assignment.
• Procedure: Gene expression values are randomly reassigned to spatial locations thousands of times, and the test statistic is recalculated each time.
• The observed statistic is compared to this null distribution to compute an empirical p-value.

This approach makes no assumptions about the underlying data distribution and is the gold standard for significance testing in SVG detection.

A spatial point pattern analysis tool used to determine if cells or molecular events are clustered, dispersed, or randomly distributed across multiple distance scales simultaneously.

• Unlike Moran's I, it evaluates spatial dependency as a function of distance r.
• K(r) > expected: Clustering at distance r.
• K(r) < expected: Dispersion at distance r.
• The Besag L-function is a variance-stabilized transformation commonly used for visualization.

This multi-scale property makes it particularly valuable for identifying SVGs that exhibit distinct patterns at different spatial resolutions.

A data structure where each spatial location is a node, and edges connect neighboring locations based on a distance threshold or k-nearest neighbors. It is the computational backbone for most SVG detection algorithms.

• Delaunay triangulation: Connects spots based on spatial tessellation.
• k-Nearest Neighbors (kNN): Each spot connects to its k closest spots.
• Radius-based: All spots within a fixed distance d are connected.

The graph enables efficient computation of spatial lag vectors and is required input for spatial graph neural networks and Gaussian process models.

A probabilistic model that infers unobserved spatial domains by assuming that the observed gene expression at each location depends on a hidden state with spatial dependencies.

• Hidden states: Represent distinct tissue regions or functional zones.
• Transition probabilities: Govern the likelihood of moving between states across adjacent spots.
• Emission probabilities: Model the gene expression distribution given a state.

Spatial HMMs go beyond simple autocorrelation by segmenting tissue into coherent regions, making them powerful for identifying SVGs that define anatomical boundaries rather than smooth gradients.