Mapping the topography of spatial gene expression with interpretable deep learning

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatially resolved transcriptomics offers high-throughput gene expression data but suffers from sparsity, hindering spatial pattern analysis.
• Analyzing complex spatial gene expression requires methods to interpret both continuous gradients and discontinuous variations.

Purpose of the Study:

• To develop an unsupervised deep learning algorithm for analyzing spatial transcriptomics data.
• To create a topographic map of tissue slices to identify spatial domains and gene expression changes.

Main Methods:

• Derivation of an 'isodepth' quantity to create topographic maps of tissue slices.
• Development of GASTON (gradient analysis of spatial transcriptomics organization with neural networks), a deep learning algorithm.
• GASTON simultaneously learns isodepth, spatial gradients, and piecewise linear expression functions.

Main Results:

• GASTON accurately identifies spatial domains and marker genes across multiple tissue types.
• The algorithm reveals gradients of neuronal differentiation and firing in the brain.
• GASTON elucidates gradients of metabolism and immune activity within the tumor microenvironment.

Conclusions:

• GASTON provides an interpretable deep learning framework for analyzing sparse spatial transcriptomics data.
• The method effectively models complex gene expression patterns, including continuous and discontinuous variations.
• GASTON enhances the understanding of tissue organization and cellular functions in various biological contexts.