Spatial-MGCN: a novel multi-view graph convolutional network for identifying spatial domains with attention mechanism

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics technologies provide gene expression data with spatial context.
• Accurate spatial domain identification is vital for downstream analyses.
• Current computational methods struggle to adaptively learn complex gene expression-spatial relationships.

Purpose of the Study:

• To develop a novel deep learning method for enhanced spatial domain detection.
• To overcome the limitations of existing computational approaches in integrating gene expression and spatial information.

Main Methods:

• Proposed Spatial-MGCN, a Multi-view Graph Convolutional Network (GCN) with an attention mechanism.
• Constructed separate gene expression and spatial neighbor graphs.
• Employed a multi-view GCN encoder for feature and spatial graph embeddings.
• Utilized a zero-inflated negative binomial decoder for expression matrix reconstruction.
• Incorporated spatial regularization for end-to-end feature learning.

Main Results:

• Spatial-MGCN demonstrated superior performance compared to state-of-the-art methods.
• The method achieved consistent improvements in spatial clustering tasks.
• Effectiveness was also shown in trajectory inference applications.

Conclusions:

• Spatial-MGCN offers an effective deep learning framework for spatial domain identification.
• The method successfully integrates gene expression and spatial information for improved analysis.
• Spatial-MGCN advances the capabilities of computational tools in spatial transcriptomics.