stGNN: Spatially Informed Cell-Type Deconvolution Based on Deep Graph Learning and Statistical Modeling

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics (ST) revolutionizes tissue analysis but lacks single-cell resolution.
• Existing methods often ignore spatial context or fail to align reference data, limiting cell type mapping accuracy.

Purpose of the Study:

• To develop stGNN, a novel spatially-informed graph learning framework for accurate cell type deconvolution in ST data.
• To improve cell type resolution and mapping by integrating spatial context and statistical modeling.

Main Methods:

• Developed a dual encoding module using graph convolutional networks (GCN) and auto-encoders for spatial and non-spatial representations.
• Implemented an adaptive attention mechanism to integrate multi-scale spatial structures.
• Utilized a negative log-likelihood loss function to align ST and single-cell RNA sequencing (scRNA-seq) data distributions.

Main Results:

• stGNN consistently outperformed seven state-of-the-art methods across six diverse ST datasets (10x Visium, Slide-seqV2, Visium HD).
• Successfully resolved distinct cortical layers in mouse brain tissue at high resolution.
• Demonstrated effective performance across varying ST resolutions.

Conclusions:

• stGNN offers a powerful, accurate framework for analyzing cell type composition and spatial distribution in complex tissues.
• The method enhances deconvolution accuracy by effectively leveraging spatial information and aligning reference datasets.