Spatially Variable Genes

Pipeline question: Which genes exhibit expression patterns that are spatially organized rather than randomly distributed across the tissue?

Overview

Spatially variable gene (SVG) detection identifies genes whose expression varies in a structured way across tissue space. These genes mark spatial domains, gradients, and local microenvironments that would be invisible in dissociated scRNA-seq data. SVG detection typically serves as an exploratory step — surfacing candidate genes for spatial domain identification, ligand-receptor analysis, or biological interpretation.

Key Methods

SpatialDE / SpatialDE2

• Paper: Nature Methods, 2018 / bioRxiv, 2021
• Code: github.com/PMBio/SpatialDE
• Key innovation: Gaussian process regression to model spatial gene expression, with automatic relevance determination for spatial versus non-spatial variance.
• Strengths:
  • Principled statistical framework with well-calibrated p-values
  • SpatialDE2 adds automatic expression histology for grouping genes by spatial pattern
• Limitations:
  • Computationally expensive — O(n^3) scaling limits applicability to large datasets
  • SpatialDE2 improves speed but remains slower than non-GP methods
• Technology compatibility: Visium, Slide-seq, ST, any spot-level data

SPARK / SPARK-X

• Paper: Nature Methods, 2020 / Genome Biology, 2021
• Code: github.com/xzhoulab/SPARK
• Key innovation: SPARK uses a spatial generalized linear mixed model; SPARK-X uses non-parametric covariance testing for scalability to millions of cells.
• Strengths:
  • SPARK-X is among the fastest SVG methods available
  • Non-parametric approach avoids distributional assumptions
  • Scales to Stereo-seq and other high-resolution platforms
• Limitations:
  • SPARK-X trades some statistical power for speed
  • Less sensitive to subtle spatial patterns than GP-based methods
• Technology compatibility: Visium, Slide-seq, Stereo-seq, Xenium, MERFISH

nnSVG

• Paper: Nature Communications, 2023
• Code: github.com/lmweber/nnSVG
• Key innovation: Nearest-neighbor Gaussian processes that approximate full GP inference in O(n) time, enabling scalable and accurate SVG detection.
• Strengths:
  • Best accuracy in independent benchmarks
  • Scalable through nearest-neighbor approximation
  • Provides effect size estimates (proportion of spatial variance)
• Limitations:
  • Slower than SPARK-X despite approximations
  • R/Bioconductor only
• Technology compatibility: Visium, Visium HD, Slide-seq, Xenium, MERFISH, Stereo-seq

SOMDE

• Paper: Bioinformatics, 2021
• Code: github.com/WhittakerLab/SOMDE
• Key innovation: Self-organizing map dimensionality reduction of spatial coordinates before Gaussian process testing, dramatically reducing computation time.
• Strengths:
  • 5-10x faster than SpatialDE with comparable accuracy
  • Simple to use with minimal parameters
• Limitations:
  • SOM-based approximation may miss fine-grained spatial patterns
  • Less well-maintained than alternatives
• Technology compatibility: Visium, Slide-seq, ST

Hotspot

• Paper: Cell Systems, 2021
• Code: github.com/yoseflab/hotspot
• Key innovation: Identifies informative genes and gene modules using local autocorrelation statistics on any cell-cell similarity graph, not just spatial coordinates.
• Strengths:
  • Flexible — works with spatial, kNN, or other similarity graphs
  • Identifies gene modules (co-varying gene sets), not just individual SVGs
  • Fast computation
• Limitations:
  • Graph-based approach conflates spatial and transcriptomic neighborhoods if not careful
  • Less interpretable spatial pattern decomposition than SpatialDE
• Technology compatibility: Any spatial platform, also applicable to scRNA-seq

SpatialCorr

• Paper: bioRxiv, 2023
• Code: github.com/mbernste/SpatialCorr
• Key innovation: Tests for spatially varying gene-gene correlations, detecting genes whose co-expression structure changes across tissue space.
• Strengths:
  • Unique focus on spatial co-expression, not just spatial expression
  • Identifies genes involved in spatially-localized regulatory programs
• Limitations:
  • Computationally intensive for large gene panels
  • Conceptually more complex — harder to interpret than standard SVGs
• Technology compatibility: Visium, Slide-seq, any spot-level data

trendsceek

• Paper: Nature Methods, 2018
• Code: github.com/edsgard/trendsceek
• Key innovation: Uses marked point process statistics to detect spatial expression trends without assuming specific pattern types.
• Strengths:
  • Non-parametric approach detects diverse spatial patterns
  • Among the earliest SVG methods with solid statistical foundations
• Limitations:
  • Very slow — impractical for genome-wide testing on large datasets
  • Largely superseded by faster methods
• Technology compatibility: Visium, ST, Slide-seq

PROST

• Paper: Nature Communications, 2024
• Code: github.com/WenMi1/PROST
• Key innovation: Quantifies spatial expression patterns using a Positive-and-Negative-index (PI) score that captures both the strength and type of spatial autocorrelation.
• Strengths:
  • Fast and scalable
  • PI score provides interpretable pattern characterization
  • Also performs spatial domain detection
• Limitations:
  • Dual-purpose design (SVG + domains) may not optimize either fully
  • Less independent validation than nnSVG or SPARK-X
• Technology compatibility: Visium, Slide-seq, Stereo-seq

Benchmark Summary

Independent benchmarks consistently rank nnSVG as the most accurate SVG detection method, with well-calibrated p-values and the best true-positive recovery rate. SPARK-X offers the best speed-accuracy tradeoff, making it the practical choice for large datasets (>50,000 spots/cells) where nnSVG becomes slow. Notably, the classical Moran's I statistic remains a competitive baseline — often matching or exceeding more complex methods in power, especially for strong spatial patterns.

Start simple

Moran's I (available in squidpy and scanpy) is a reasonable first pass for SVG detection. Move to nnSVG for rigorous analysis or SPARK-X when dataset size demands speed.

When to Use What

Your data	Your goal	Recommended	Why
Visium (<10K spots)	Rigorous SVG detection	nnSVG	Best accuracy, manageable runtime at this scale
Large dataset (>50K cells)	Fast SVG screen	SPARK-X	Scalable non-parametric testing
Quick exploration	Initial SVG candidates	Moran's I (squidpy)	Fast, no additional installation, competitive accuracy
Any data	Gene module discovery	Hotspot	Identifies co-varying gene groups, not just individual SVGs
Any data	Spatially varying co-expression	SpatialCorr	Unique capability to detect changing gene-gene correlations

Technology Compatibility

Method	Visium	Visium HD	Xenium	MERFISH	CosMx	CODEX	Stereo-seq
SpatialDE/2	Yes	-	-	-	-	-	-
SPARK-X	Yes	Yes	Yes	Yes	-	-	Yes
nnSVG	Yes	Yes	Yes	Yes	-	-	Yes
SOMDE	Yes	-	-	-	-	-	-
Hotspot	Yes	Yes	Yes	Yes	Yes	-	Yes
SpatialCorr	Yes	-	-	-	-	-	-
trendsceek	Yes	-	-	-	-	-	-
PROST	Yes	-	-	-	-	-	Yes