TissueMosaic: Self-supervised learning of tissue representations enables differential spatial transcriptomics across...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics enables gene expression measurement in native tissue context.
• Existing computational methods struggle to link cell states with their microenvironment and compare these across samples.
• Limited tools exist for analyzing tissue architecture and its impact on cellular function.

Purpose of the Study:

• To introduce TissueMosaic, a self-supervised convolutional neural network for discovering and representing tissue architectural motifs.
• To link identified motifs to gene expression for studying structure-function relationships.
• To enhance spatial differential expression analysis and improve detection of structure-covarying genes.

Main Methods:

• Developed TissueMosaic, a self-supervised convolutional neural network.
• Applied TissueMosaic to multi-sample spatial transcriptomic datasets.
• Utilized a motif enrichment strategy to increase signal-to-noise ratio in differential expression analysis.

Main Results:

• TissueMosaic effectively discovers and represents tissue architectural motifs.
• The method successfully links tissue motifs to gene expression patterns.
• TissueMosaic outperforms baseline methods in downstream tasks, demonstrating superior representation learning.
• Improved signal-to-noise ratio in spatial differential expression analysis.

Conclusions:

• TissueMosaic provides a novel computational framework for spatial transcriptomics.
• Self-supervised learning holds significant potential for advancing discoveries in spatial transcriptomics.
• The method enables a deeper understanding of how tissue structure influences cell-intrinsic functions.