A masked generative graph representation learning framework empowering precise spatial domain identification

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics (ST) provides gene expression data with spatial context.
• ST data sparsity hinders effective use of gene expression and spatial information.
• This leads to poorly represented embeddings, challenging downstream analyses.

Purpose of the Study:

• To develop a novel framework for self-supervised representation learning in ST data.
• To enhance the utilization of gene expression and spatial information from ST data.
• To improve spatial domain identification and downstream analysis of ST data.

Main Methods:

• Introduced GSG, a generative self-supervised representation learning framework.
• Employed a masking mechanism within GSG to learn informative representations.
• Applied GSG to ST data, including an in-house human fetal heart dataset.

Main Results:

• GSG consistently outperformed state-of-the-art methods in spatial domain identification across datasets and platforms.
• GSG revealed anatomically coherent spatial domains in a human fetal heart dataset.
• Identified APCDD1 as a novel endocardial-specific marker potentially linked to congenital heart disease.

Conclusions:

• GSG demonstrates superior performance in ST data analysis.
• The framework effectively leverages gene expression and spatial information for improved representation learning.
• GSG offers valuable contributions to advancing spatial transcriptomics analysis and disease marker discovery.