Graph contrastive learning of subcellular-resolution spatial transcriptomics improves cell type annotation and...

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• Imaging-based spatial transcriptomics (iST) quantifies gene expression in space at single-molecule resolution, revealing subcellular RNA distribution.
• Subcellular RNA localization is critical for gene regulation and cell identity.
• Current iST cell type annotation methods overlook RNA's intracellular spatial patterns.

Purpose of the Study:

• To introduce Focus, a novel semi-supervised graph contrastive learning method for iST.
• To explicitly model RNA's subcellular distribution and community for improved cell type annotation.
• To leverage subcellular spatial gene patterns for enhanced biological interpretation.

Main Methods:

• Developed Focus, a semi-supervised graph contrastive learning framework.
• Applied Focus to various iST platforms, including MERFISH, CosMx SMI, and Xenium.
• Integrated gene expression data with subcellular RNA spatial distribution information.

Main Results:

• Focus significantly improves cell type annotation accuracy (up to 27.8%) and F1-score (up to 51.9%) over existing methods.
• The method effectively captures intricate, cell type-specific subcellular spatial gene expression patterns.
• Focus provides interpretable gene importance scores, highlighting genes relevant to cell type-specific pathways.

Conclusions:

• Focus represents a breakthrough in iST cell type annotation by incorporating subcellular RNA distribution.
• The method enhances cell characterization and offers insights into unique subcellular regulatory mechanisms.
• Focus has potential for discovering novel regulatory programs in diverse biological systems.