GAADE: identification spatially variable genes based on adaptive graph attention network.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) enables gene expression mapping with spatial coordinates.
• Existing methods for identifying spatially variable genes (SVGs) often overlook spatial domains, limiting accuracy.
• Predefined spot similarity in current domain-based SVG identification hinders adaptive learning and generalizability.

Purpose of the Study:

• To develop an advanced method for identifying SVGs that considers spatial domains.
• To improve the accuracy and generalizability of SVG detection in ST data.
• To address limitations of existing methods in capturing explicit spatial expression patterns.

Main Methods:

• Introduced GAADE, an unsupervised neural network using graph-structured data representation learning.
• GAADE employs encoder/decoder layers and a self-attention mechanism to capture spatial domain structures.
• SVGs are identified by differential expression analysis within spatial domains and their neighbors.

Main Results:

• GAADE effectively reconstructs spatial domain structures.
• The method successfully confines SVG identification within relevant spatial domains.
• Comparative evaluations show GAADE outperforms existing methods in detecting SVGs and their expression variation extent across diverse ST datasets.

Conclusions:

• GAADE offers a superior approach for identifying SVGs by integrating spatial domain information.
• The unsupervised neural network architecture enhances the analysis of spatial gene expression patterns.
• GAADE demonstrates robust performance and generalizability across different species, regions, and tissues.