by Christos Evangelou, MSc, PhD – Medical Writer and Editor

Spatial transcriptomics provides insights into how genes are expressed across tissues; however, current methods often struggle with data sparsity and fail to capture continuous variations in gene expression. A new algorithm called GASTON (Gradient Analysis of Spatial Transcriptomics Organization with Neural Networks) addresses these limitations by introducing the concept of isodepth, a topographic mapping approach that reveals both distinct spatial domains and continuous gradients in gene expression. 

The report was published in Nature Methods. 

Approach: Investigating Variation in Gene Expression Across Space

“Many existing clustering algorithms are based on the idea that there are distinct spatial domains in a tissue that can be distinguished by discrete changes in expression,” explained corresponding author Benjamin Raphael, PhD, in an interview with Pathology News. “The isodepth provides a continuous coordinate over a tissue slice that allows researchers to investigate gradual, or continuous, variation in expression across space.” 

This approach can model tissue organization similar to how topographic maps represent Earth’s surface, with isodepth contours functioning like elevation lines on a geographical map. Dr. Raphael noted that GASTON enhances signal detection in sparse spatial transcriptomics data, revealing previously hidden biological patterns by combining measurements from locations with similar isodepth values. 

GASTON Provides a Comprehensive View of Tissue Organization in the Brain

When the researchers applied GASTON to brain tissue samples, they were able to identify both known and novel marker genes in cerebellum layers, even from extremely sparse data. The algorithm successfully mapped the complex spatial organization of the mouse cerebellum, identifying distinct layers and the continuous gradients between them. 

“We were surprised that we could identify both known and novel marker genes of the layers of the cerebellum from extremely sparse spatial transcriptomics data,” said Dr. Raphael.  

He explained that the data they analyzed had extremely low coverage, with only approximately 500 unique RNA molecules measured per spatial location. However, by combining spots with similar values of isodepth, they identified individual genes that are expressed in specific layers of the cerebellum and even genes that vary continuously between the layers. 

GASTON outperformed six existing methods for identifying spatially varying genes, achieving higher accuracy in marker gene identification. According to the authors, the ability of the algorithm to detect both continuous gradients and sharp discontinuities in gene expression provides a more comprehensive view of tissue organization than previous approaches that focus primarily on discrete domains.

GASTON Reveals Distinct Spatial Expression Patterns in the Tumor Microenvironment

Using GASTON, the researchers were able to identify complex spatial patterns in tumor microenvironments that could inform diagnosis and treatment decisions. When applied to colorectal cancer samples, GASTON revealed three distinct spatial expression patterns: intrastromal variation, discontinuities at tumor-stroma boundaries, and intratumoral variation. 

Dr. Raphael explained that GASTON boosts the signal in spatial transcriptomics data by combining gene expression measurements from locations with similar values of the isodepth. This allowed the team to examine the expression of genes relative to specific anatomic features such as the tumor-normal boundary. 

“Suppose we want to examine how the expression of a gene varies from the boundary to the interior of the tumor. There are many spatial locations that comprise the boundary and the tumor interior,” he elaborated. “The isodepth answers the question of which locations should be analyzed to quantify this spatial variation, learning the coordinate that models the geometry of the tumor.” 

The algorithm revealed that genes associated with oxidative phosphorylation increased toward the tumor interior, whereas epithelial-mesenchymal transition genes showed increased expression near the tumor-stroma boundary. These patterns suggest a late-stage, vascularized primary tumor with a fully metastatic margin, a characterization consistent with the tumor’s clinical information. 

Current Limitations and Future Directions

Dr. Raphael acknowledged that GASTON has limitations that the research team is actively addressing. 

“GASTON currently models tissues with a single isodepth, which performs very well for tissues with layered structures, but not as well on tissues and tumors with more complicated geometries,” he said. “We are currently extending GASTON to model such cases.”  

One extension called GASTON-Mix uses a ‘mixture of experts’ AI model to partition the tissue into distinct domains and learn an isodepth within each domain, he revealed 

The team is also exploring applications beyond the tissues studied in the current paper.

“We are very interested in applying GASTON to developmental systems where it is well known that spatial gradients play important roles,” Dr. Raphael shared. “We are excited about the possibilities of identifying additional molecular gradients in developing tissues, and relating these to dedifferentiation processes that occur in cancer.” 

Future work may extend GASTON to other molecular modalities such as chromatin accessibility or protein abundance. The researchers also emphasize the need for further study of both the regulatory causes and downstream effects of the spatial gradients identified by GASTON. 

“We hope that GASTON could reveal gradients in cell types or genes that are biological and clinically relevant,” said Dr. Raphael. “Some examples we show in the paper are metabolic gradients, immune cell gradients, and signals of activated stroma. Some of these signals might serve as biomarkers to stratify patients for particular treatments.” 

The study was supported by the National Cancer Institute (NCI), Ludwig Cancer Research, National Science Foundation Graduate Research Fellowships, the Siebel Scholars program, the Schmidt DataX Fund at Princeton University, and the Damon Runyon Cancer Research Foundation. 

References

1. Chitra U, Arnold BJ, Sarkar H, et al. Mapping the topography of spatial gene expression with interpretable deep learning. Nat Methods. 2025;22(2):298-309. doi:10.1038/s41592-024-02503-3