Spatial Differential Expression

The core mechanism involves comparing gene expression between pre-annotated spatial regions within a tissue section. Unlike global differential expression, which compares entire samples, this framework tests contrasts such as tumor core vs. invasive margin or cortical layer 2/3 vs. layer 5. The analysis accounts for the spatial covariance structure of the data, ensuring that observed differences reflect true biological variation rather than positional artifacts. Statistical models typically incorporate spatial random effects or Gaussian process kernels to model the correlation between neighboring spots.

Standard t-tests and Wilcoxon rank-sum tests assume independent observations, an assumption violated by spatially autocorrelated transcriptomic data. Spatial differential expression frameworks employ specialized models:

• Spatial Generalized Linear Mixed Models (GLMMs): Incorporate spatial random effects to account for location-based correlation
• Gaussian Process Regression: Uses kernel functions to model smooth expression gradients across tissue coordinates
• Moran's I-adjusted tests: Correct test statistics by estimating the effective sample size after accounting for spatial autocorrelation
• Spatial permutation tests: Shuffle region labels while preserving spatial structure to generate valid null distributions

When testing thousands of genes across spatial regions, false discovery rate (FDR) control becomes critical. However, the spatial dependency between tests violates the independence assumptions of standard Benjamini-Hochberg correction. Advanced approaches include:

• Spatial FDR: Estimates the proportion of null hypotheses while modeling the spatial correlation structure
• Cluster-based permutation: Identifies contiguous spatial clusters of significant expression and assesses their significance as a whole
• Bayesian hierarchical models: Share information across spatially proximal genes to improve power while controlling false positives
• Spatial knockoff filters: Generate synthetic null variables that preserve the spatial correlation structure for rigorous FDR control

Beyond discrete region comparisons, spatial differential expression can identify genes with continuous expression gradients radiating from a reference structure. For example, a gene may show decreasing expression with increasing distance from a blood vessel or increasing expression toward the necrotic core of a tumor. This is modeled using:

• Distance-based regression: Expression ~ f(distance_to_landmark) with spline or polynomial basis functions
• Spatial trend tests: Non-parametric tests for monotonic expression changes along a spatial trajectory
• Boundary detection algorithms: Identify genes whose expression sharply changes at the interface between two tissue compartments

Complex experimental designs extend spatial differential expression beyond simple two-region comparisons:

• Multi-region ANOVA: Identifies genes differentially expressed across three or more anatomical zones simultaneously
• Spatial interaction effects: Tests whether the difference between two regions depends on a covariate, such as treatment vs. control × tumor vs. stroma
• Paired spatial designs: Compares matched regions within the same tissue section, controlling for inter-sample variability
• Spatial case-control contrasts: Compares equivalent anatomical regions across disease and healthy tissues, requiring robust spatial registration before differential testing

Modern spatial differential expression frameworks jointly model gene expression and tissue morphology. By incorporating features extracted from paired H&E or immunofluorescence images—such as cell density, nuclear atypia, or extracellular matrix alignment—the analysis can identify genes whose spatial expression patterns correlate with histological transitions. This multimodal approach distinguishes expression changes driven by cellular composition shifts from those reflecting true transcriptional regulation within a cell type, providing mechanistic insight beyond simple differential lists.

Genes whose expression exhibits non-random spatial patterning across tissue. These are the primary output of spatial differential expression analysis.

• Detection methods: Gaussian process regression, spatial autocorrelation statistics, or marked point processes
• Biological significance: Often mark tissue boundaries, signaling centers, or disease-specific microenvironments
• Example: A gene highly expressed only at the invasive front of a tumor but not in the core

A non-parametric significance test that generates an empirical null distribution by randomly shuffling spatial coordinates while preserving gene expression values.

• Process: Permute labels 1,000–10,000 times, recalculate the test statistic each iteration
• Advantage: No assumption of normality in the underlying expression data
• Output: An empirical p-value indicating how extreme the observed spatial pattern is compared to random chance

The computational partitioning of tissue images into anatomically or functionally distinct regions that define the spatial domains for differential comparison.

• Input: H&E stained images, immunofluorescence, or spatial transcriptomics data
• Methods: Deep learning architectures like U-Net, graph-based clustering, or manual annotation
• Role in SDE: Defines the categorical spatial regions between which differential expression is tested

A graph data structure where nodes represent spatial measurement locations and edges encode proximity relationships.

• Construction: k-nearest neighbors (k-NN) or distance-thresholded Delaunay triangulation
• Utility: Enables graph neural networks and spatial smoothing operations
• Edge weights: Often incorporate physical distance, enabling distance-weighted statistical models

Statistical adjustment required when testing thousands of genes simultaneously across spatial regions to control false discovery rates.

• Benjamini-Hochberg: Controls the expected proportion of false positives among all discoveries
• Bonferroni: More conservative, controls family-wise error rate
• Spatial FDR: Specialized methods that account for spatial dependency between tests, reducing overcorrection