Mapping the topography of spatial gene expression with interpretable deep learning.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) offers high-throughput gene expression data but suffers from sparsity, complicating the analysis of spatial patterns like gene expression gradients.
• Existing methods struggle to effectively model both continuous gradients and discontinuous variations in gene expression across tissue samples.

Approach:

• We introduce GASTON, an unsupervised and interpretable deep learning algorithm designed to analyze sparse SRT data.
• GASTON learns a "topographic map" of the tissue using a novel "isodepth" quantity, where contours define spatial domains and isodepth gradients indicate directions of maximal gene expression change.
• The algorithm models gene expression using piecewise linear functions of the isodepth, capturing both continuous and discontinuous spatial variations.

Key Points:

• GASTON accurately identifies spatial domains and marker genes in diverse biological systems.
• The method reveals gradients in neuronal differentiation and firing in brain SRT data.
• GASTON infers metabolic activity and epithelial-mesenchymal transition (EMT) gradients in tumor microenvironments.

Conclusions:

• GASTON provides a powerful new tool for dissecting complex spatial gene expression patterns in SRT data.
• This approach enhances our understanding of tissue organization, developmental processes, and disease mechanisms by revealing hidden spatial relationships in gene activity.