MAEST: accurately spatial domain detection in spatial transcriptomics with graph masked autoencoder.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) provides gene expression with spatial context, crucial for understanding tissue architecture.
• Spatial domain identification in ST is vital but challenged by existing methods' limited use of spatial information.
• Suboptimal clustering accuracy and representational capacity hinder current ST analysis.

Purpose of the Study:

• To introduce MAEST, a novel graph neural network model for enhanced spatial domain identification in ST data.
• To improve the accuracy and robustness of spatial domain detection by leveraging comprehensive spatial relationships.
• To enable accurate integration of multi-slice ST data for joint domain identification.

Main Methods:

• MAEST utilizes graph masked autoencoders for denoising and refining ST data representations.
• Graph contrastive learning is incorporated to prevent feature collapse and enhance model robustness.
• Integration of one-hop and multi-hop representations captures local and global spatial relationships.

Main Results:

• MAEST significantly outperforms seven state-of-the-art methods in spatial domain identification across diverse datasets (human brain, mouse hippocampus, etc.).
• The model demonstrates high accuracy in identifying joint spatial domains across multiple horizontal tissue sections.
• MAEST effectively refines representations and improves clustering precision by capturing intricate spatial relationships.

Conclusions:

• MAEST offers a versatile and effective approach for unraveling the spatial organization of complex tissues using ST data.
• The model's ability to integrate multi-slice data opens new avenues for cross-tissue spatial analysis.
• MAEST represents a significant advancement in computational methods for spatial transcriptomics analysis.