AddaGCN: Spatial transcriptomics deconvolution using graph convolutional networks with adversarial discriminative...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics advances tissue analysis but often lacks single-cell resolution.
• Current methods struggle with gene expression profiles from mixed cell types.
• Batch effects between spatial and single-cell reference data pose a challenge.

Purpose of the Study:

• To develop AddaGCN, a robust deconvolution method for inferring cell type composition from spatial transcriptomic data.
• To address the limitations of existing methods in handling mixed cell types and batch effects.
• To enhance the understanding of spatial architecture and gene expression heterogeneity in tissues.

Main Methods:

• Utilized graph convolutional networks to integrate spatial information.
• Employed adversarial discriminative domain adaptation to mitigate batch effects.
• Validated performance on diverse real-world spatial transcriptomic datasets.

Main Results:

• AddaGCN demonstrated superior performance and robustness in cell-type deconvolution compared to existing methods.
• Successfully inferred cell type composition from spatial transcriptomic data across various technology platforms.
• Showcased potential for analyzing spatiotemporal changes and characterizing the tumor microenvironment.

Conclusions:

• AddaGCN offers a powerful and reliable solution for cell-type deconvolution in spatial transcriptomics.
• The method effectively handles technical challenges like batch effects.
• AddaGCN facilitates deeper insights into tissue biology, development, and disease contexts like cancer.