Joint identification of spatially variable genes via a network-assisted Bayesian regularization approach

Mingcong Wu, Yang Li, Shuangge Ma, Mengyun Wu

TL;DR

Researchers have developed a new Bayesian framework that identifies spatially variable (SV) genes by looking at them "jointly" rather than one by one. By incorporating known biological networks and correcting for cell-type mixtures, this method significantly improves the accuracy of spatial transcriptomics analysis, especially for genes with weak signals.

What Problem Does This Paper Address?

Spatial Transcriptomics (ST) is a revolutionary technology that allows scientists to see not just which genes are active, but where they are active within a tissue. A primary goal in ST is identifying Spatially Variable (SV) genes—those whose expression levels change based on their location (e.g., genes active only in a tumor core vs. the periphery).

However, current methods face two major hurdles:

  1. The "Lone Wolf" Problem: Most tools analyze genes individually (marginal analysis). In reality, genes work in complex networks; if one gene shows a spatial pattern, its biological "neighbors" likely do too.
  2. The "Mixed Bag" Problem: Many ST technologies (like 10X Visium) measure "spots" that contain a mixture of different cell types. If a gene looks spatially variable, is it because of a true spatial program, or just because the cell-type composition of the spots is changing?

Methods

The authors propose a network-assisted Bayesian regularization approach. Here’s the breakdown of their "secret sauce":

  • ZINB Modeling: They use a Zero-Inflated Negative Binomial (ZINB) distribution. This is crucial because ST data is "noisy"—it’s full of zeros (dropouts), count-based, and has high variance (overdispersion).
  • Thresholded Graph Laplacian Regularization: This is the core innovation. They take a known biological network (like a protein-protein interaction map) and use it to "nudge" the model. If two genes are connected in the network, the model assumes they should have similar spatial patterns. The "thresholded" part allows the model to be selective, identifying true SV genes while ignoring noise.
  • Confounding Correction: The model includes cell-type proportions (derived from deconvolution) as covariates. This ensures that the spatial patterns identified are truly "spatial" and not just artifacts of varying cell densities.
  • Mean-Value-Based Strategy: Instead of complex Gaussian processes, they model the log-mean expression as a linear combination of spatial coordinates using specific "kernels" (linear, exponential, and periodic).

Datasets

The researchers validated their method using:

  • Extensive Simulations: 12 different scenarios with varying dropout rates, spatial patterns, and network structures (Star and Scale-free networks).
  • Visium Triple-Negative Breast Cancer (TNBC): 1,109 spots and over 36,000 genes.
  • Visium Primary Liver Cancer (PLC): 3,181 spots.
  • Xenium Pancreatic Cancer (PAC): A high-resolution dataset involving 164,274 single cells.

Key Insights and Ideas

  • Strength in Numbers: By analyzing genes jointly, the model can "rescue" genes with weak spatial signals. If a gene has a faint spatial pattern but is connected to a "hub" gene with a very strong spatial pattern, the network prior helps the model identify it correctly.
  • Interpretability over Complexity: While many methods use Gaussian Processes (which are powerful but "black box" and computationally expensive), this mean-value-based approach is more intuitive. It tells you exactly how a gene varies—whether it's a gradient (linear) or a localized cluster (exponential).
  • Biological "Communities": The method doesn't just find a list of genes; it finds connected communities of genes that share spatial functions, such as RNA splicing in breast cancer or carbohydrate metabolism in liver cancer.

Key Results

  • Superior Accuracy: In simulations, the proposed method consistently achieved the highest F1 scores (a balance of precision and recall) compared to nine other popular tools like SpatialDE, SPARK-X, and nnSVG.
  • Robustness to Noise: Even when the provided biological network was "noisy" (containing incorrect connections), the method maintained a better balance of recall and precision than its competitors.
  • Real-World Discovery: In the TNBC dataset, the method identified 1,084 SV genes. Notably, 782 of these were connected in a known protein interaction network, forming distinct functional "communities" related to cancer progression that other methods missed.

Limitations and Future Work

  • Pre-defined Patterns: The current model relies on pre-specified kernels (linear, exponential, periodic). If a gene follows a highly irregular or "wavy" pattern that doesn't fit these shapes, the model might struggle.
  • Rotation Sensitivity: The authors note that their coordinate-based approach isn't "rotation invariant"—if you rotate the tissue sample, the axis-specific results might change.
  • Network Dependence: The method's performance is tied to the quality of the input biological network. While it handles some noise well, a completely irrelevant network would hinder the results.
  • Future Directions: The authors suggest moving toward non-parametric methods to capture unknown spatial patterns and developing rotation-invariant strategies using relative distances between spots.

Summary updated

Abstract

Identifying genes that display spatial patterns is critical to investigating expression interactions within a spatial context and further dissecting biological understanding of complex mechanistic functionality. Despite the increase in statistical methods designed to identify spatially variable genes, they are mostly based on marginal analysis and share the limitation that the dependence (network) structures among genes are not well accommodated, where a biological process usually involves changes in multiple genes that interact in a complex network. In addition, the latent cellular composition within the spots can introduce confounding variations, negatively affecting the accuracy of the identification. In this study we develop a novel Bayesian regularization approach for spatial transcriptomic data, with confounding variations induced by varying cellular distributions effectively corrected. Significantly advancing from existing studies, a thresholded graph Laplacian regularization is proposed to simultaneously identify spatially variable genes and accommodate the network structure among genes. The proposed method is based on a zero-inflated negative binomial distribution, effectively accommodating the count nature, zero inflation, and overdispersion of spatial transcriptomic data. Extensive simulations and applications to real data demonstrate the competitive performance of the proposed method.
BibTeX
@article{wu2025joint,
  author = {Mingcong Wu and Yang Li and Shuangge Ma and Mengyun Wu},
  title = {{Joint identification of spatially variable genes via a network-assisted Bayesian regularization approach}},
  journal = {The Annals of Applied Statistics},
  year = {2025},
  doi = {10.1214/25-aoas2097},
}

Loading related content...

Figure