Paper Review

Gene expression varies across tissues due to both the spatial arrangement of cell types and localized changes in cell states. Some genes show sharp boundaries between spatial domains, such as cortical layers in the brain, while others form gradients that drive key biological processes, including neuronal heterogeneity and tumor microenvironment interactions. Spatial transcriptomics (SRT) is a cutting-edge field that measures gene expression in tissue slices, helping scientists understand how genes work in specific locations. However, SRT data is often sparse, making it hard to analyze spatial patterns accurately. This challenge has limited our ability to study complex biological systems like the brain and tumors.

A recent paper introduces GASTON, an unsupervised and interpretable deep learning algorithm. GASTON creates a "topographic map" using a concept called "isodepth," which identifies areas in tissues with different cell types. It can model both smooth (continuous) and abrupt (discontinuous) changes in gene expression, addressing a key limitation of previous methods.

GASTON was tested on various tissues, including the mouse cerebellum, human colorectal cancer (CRC) tumors, and mouse olfactory bulb. It revealed detailed patterns, such as neuronal differentiation in the brain and metabolic gradients in tumors. For CRC tumors, it identified changes near tumor boundaries that might indicate aggressiveness, offering new insights into cancer biology. The code for GASTON is publicly accessible at https://github.com/raphael-group/GASTON and via Zenodo at https://doi.org/10.5281/zenodo.12702592

The study used data from multiple sources, including Stereo-seq data from the mouse olfactory bulb, which includes the expression of 27,106 transcripts within 9,825 cells.

Read more about the research:  https://doi.org/10.1038/s41592-024-02503-3