A spatially informed matrix normal model for gene co-expression analysis in spatial transcriptomics studies

Chichun Tan, Ying MaNucleic 2025 · DOI

TL;DR

This paper introduces spMOCA, a novel statistical framework that accurately infers gene co-expression networks from spatial transcriptomics data by simultaneously modeling gene-gene interactions and spatial dependencies. This approach provides more biologically relevant insights into gene function within tissue microenvironments compared to existing methods.

What Problem Does This Paper Address?

Spatial transcriptomics (SRT) technologies allow us to measure gene expression while preserving the spatial context of tissues. Analyzing gene co-expression in SRT data is crucial for understanding how genes work together within this spatial context. However, current methods struggle to effectively capture the interplay between gene-gene relationships and spatial organization. This limitation can lead to biased gene co-expression estimates, where apparent correlations are driven by spatial proximity rather than true biological interactions. The authors aim to bridge this gap by developing a method that explicitly models both gene-gene and spatial dependencies.

Methods

The core of spMOCA is a matrix-normal model. This statistical framework allows the authors to jointly model the gene expression data as a matrix where rows represent genes and columns represent spatial locations. The model explicitly accounts for:

  • Gene covariance: How genes' expression levels relate to each other.
  • Spatial covariance: How gene expression varies across different locations in the tissue.
  • Gene-spatial interaction: How gene-gene relationships are influenced by spatial proximity.

This is achieved by using a Kronecker product structure for the covariance matrices, which elegantly disentangles intrinsic co-expression from spatially induced effects. The method uses maximum likelihood estimation and efficient matrix calculations, making it scalable to large SRT datasets. For modeling spatial dependency, they employ a Gaussian kernel with Silverman's bandwidth selection rule, though they note that alternative kernels could be explored.

Datasets

The authors evaluated spMOCA using a combination of simulated data and nine publicly available spatial transcriptomics datasets:

  • Simulated Data: Two simulation frameworks were used:
  • Spatial-gene interactive dependency: Mimics spMOCA's core assumption where gene and spatial effects jointly influence expression.
  • Spatial-gene additive dependency: Violates spMOCA's core assumption, where gene and spatial effects are independent. These simulations were conducted under varying levels of spatial dependency (weak, moderate, strong) and in scenarios with no true gene co-expression to assess robustness.

  • Real Datasets:

    • Four 10X Visium SRT datasets from different cancer types: Breast Ductal Carcinoma (BRCA), Colorectal Cancer (CRC), Lung Squamous Cell Carcinoma (LUSC), and Ovarian Carcinoma (OVCA).
  • Three 10X Visium SRT datasets from mouse whole brain tissue across different ages (6, 18, and 21 months).
  • Two MERFISH datasets from human and mouse cortex tissues.

SpaceX was excluded from some real data analyses due to computational convergence issues.

Key Insights and Ideas

  • Spatial context is not just noise: The paper strongly argues that spatial organization isn't merely a nuisance to be smoothed over but a fundamental driver of gene expression patterns that needs to be explicitly modeled.
  • Disentangling true vs. spatial correlations: The matrix-normal model is key to separating genuine gene-gene regulatory relationships from correlations that arise simply because genes are located near each other.
  • Beyond additive models: Existing methods often treat spatial and gene effects additively. spMOCA's interactive modeling captures a more complex and realistic biological scenario.
  • Cross-species conservation: The method can identify conserved gene modules and cell-type specific pathways across different species, hinting at fundamental biological processes.

Key Results

  • Simulation Performance: spMOCA consistently outperformed existing methods (Pearson's correlation, SpaceX, Giotto, CSCORE) in simulations across various metrics (PCC, RV, RMSE, MAD, ARI) for both interactive and additive dependency scenarios, especially when spatial dependencies were strong. It showed particularly high accuracy in recovering true gene modules.
  • Cancer Data Analysis:
    • spMOCA identified a higher proportion of experimentally validated transcription factor (TF) target genes compared to other methods.
  • It uncovered gene modules associated with tumorigenesis and immune pathways, revealing potential prognostic markers.
  • Module scores from spMOCA showed strong correlations with specific tissue regions (e.g., invasive carcinoma vs. immune cells) and cell types.
  • Aging Mouse Brain Data: spMOCA captured dynamic changes in gene co-expression relationships associated with aging and neurodegeneration, identifying key gene markers for these processes.
  • Cross-Species Brain Data: spMOCA identified conserved gene modules and cell type-specific pathways in human and mouse cortex, demonstrating its utility for comparative analyses.

Limitations and Future Work

  • Spatial Kernel Choice: While spMOCA is robust to bandwidth selection, the authors suggest exploring alternative kernels (e.g., periodic, cosine) to capture diverse spatial correlation patterns.
  • Unified Co-expression Assumption: Currently, spMOCA assumes a single co-expression relationship across the entire tissue. The authors propose extending spMOCA to a mixture model for unsupervised inference of region-specific GCNs, which would not require pre-defined spatial domains.
  • Spatial Domain Detection: The paper hints at the potential of spMOCA-derived GCNs to improve spatial domain detection, suggesting this as a future research direction.
  • Computational Cost: While spMOCA is designed to be scalable, the paper notes that SpaceX, a competing method, faced computational challenges. Further optimization for extremely large datasets might be beneficial.

Summary updated

Abstract

The rapid advancement of spatially resolved transcriptomics (SRT) technology enables gene expression profiling across tissue locations while preserving spatial context. Gene co-expression analysis in SRT data provides critical insights into how genes function together within the tissue microenvironment. However, existing methods fail to effectively capture the joint influence of gene-gene interactions and spatial dependencies, limiting their biological interpretability. Here, we introduce spMOCA (SPatially informed Matrix-nOrmal model for gene Co-expression Analysis), a statistical framework for inferring gene co-expression networks while explicitly modeling spatial dependencies. By leveraging a matrix-normal model, spMOCA jointly accounts for gene-gene and spatial covariance, disentangling intrinsic co-expression relationships from spatially induced effects. Through extensive simulations, we show that spMOCA provides more accurate and unbiased estimates of gene-gene correlations than existing approaches across a range of spatial dependency levels. In applications to nine SRT datasets spanning diverse technologies, tissues, and species, spMOCA consistently identifies more experimentally validated transcription factor target genes than alternative methods. In tumors, it uncovers gene modules linked to tumorigenesis and immune pathways, revealing prognostic markers. In aging mouse brains, it captures dynamic co-expression changes associated with neurodegeneration. In cross-species analyses, it detects conserved gene modules and cell type-specific pathways in the mouse and human cortex.
BibTeX
@inproceedings{tan2025spatially,
  author = {Chichun Tan and Ying Ma},
  title = {{A spatially informed matrix normal model for gene co-expression analysis in spatial transcriptomics studies}},
  booktitle = {Nucleic},
  year = {2025},
  doi = {10.1093/nar/gkaf1264},
}
Figure