DGSIST: Clustering spatial transcriptome data based on deep graph structure Infomax.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Spatial transcriptomics offers insights into tissue gene expression and structure but often underutilizes spatial data.
• Graph neural networks present an opportunity to integrate spatial information with gene expression data.
• Existing methods may not fully exploit the rich spatial context available in transcriptomic datasets.

Purpose of the Study:

• To develop an unsupervised model, DGSI (Deep Graph Structure Infomax), for processing graph data from spatial transcriptomics.
• To introduce the DGSIST framework, integrating DGSI with dimensionality reduction and clustering for accurate cell type identification.
• To enhance the analysis of spatial transcriptomics data, improving cell clustering and spatial domain identification.

Main Methods:

• Developed the DGSI model using graph convolutional neural networks and an unsupervised learning approach to maximize mutual information between graph and node representations.
• Integrated DGSI with Singular Value Decomposition (SVD) and k-means++ for the DGSIST unsupervised cell clustering framework.
• Applied DGSIST to various spatial transcriptomics datasets across different tissue types and technologies.

Main Results:

• DGSIST accurately identifies cell types and spatial domains, outperforming existing methods.
• The framework effectively eliminates batch effects without explicit correction.
• Demonstrated robust performance across diverse tissue types and technological platforms.

Conclusions:

• DGSIST is a powerful unsupervised framework for cell clustering and spatial analysis using spatial transcriptomics data.
• The model effectively captures local spatial information, leading to improved accuracy in identifying cellular structures.
• DGSIST has significant potential for advancing the understanding of spatial organization in diseases like cancer.