GatorST: A Versatile Contrastive Meta-Learning Framework for Spatial Transcriptomic Data Analysis.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) offers gene expression with spatial context, crucial for understanding cellular functions.
• Existing ST methods struggle to integrate local and global spatial information, and contrastive learning approaches can be unstable.
• Robust integration of spatial and transcriptomic data is essential for effective downstream analyses.

Purpose of the Study:

• To introduce GatorST, a novel framework for learning spatially informed representations of ST data.
• To address limitations in capturing local and global contexts and improve the stability of ST analysis.
• To generate biologically meaningful representations for advancing downstream ST analyses.

Main Methods:

• GatorST utilizes graph-based modeling and meta-learning strategies.
• It incorporates a spot-spot graph for local context and gene expression clustering for global context (weak supervision).
• An episodic training strategy enhances generalization to new spatial contexts.

Main Results:

• GatorST consistently outperforms fifteen state-of-the-art methods.
• Demonstrated superior performance in identifying spatial domains, imputing gene expression, removing batch effects, and inferring spatial trajectories.
• Provides robust integration of local spatial topology and global gene expression patterns.

Conclusions:

• GatorST effectively integrates local and global spatial information for ST data.
• The framework generates biologically meaningful representations that advance key downstream analyses.
• GatorST represents a significant advancement in spatial transcriptomics analysis.