ConSpaS: a contrastive learning framework for identifying spatial domains by integrating local and global...

Area of Science:

• Spatial transcriptomics
• Computational biology
• Genomics

Background:

• Spatial transcriptomics enables understanding of tissue organization and biological processes at high resolution.
• Current methods often overlook global semantic structures, leading to inaccurate spatial domain identification.
• Integrating spatial information with gene expression profiles is crucial for biological insights.

Purpose of the Study:

• To develop a novel framework, ConSpaS, for precise spatial domain identification in spatial transcriptomics data.
• To integrate local and global similarities for improved characterization of spatial domains.
• To enhance the biological interpretability of spatial transcriptomics analyses.

Main Methods:

• Developed ConSpaS, a node representation learning framework utilizing graph autoencoder (GAE) and contrastive learning (CL).
• GAE integrates local spatial information and gene expression profiles for spatially continuous cluster assignment.
• CL enhances global semantic information using an augmentation-free mechanism and semi-easy negative sampling.

Main Results:

• ConSpaS demonstrated improved identification accuracy of spatial domains across multiple tissue types and platforms.
• The method identified biologically meaningful spatial patterns and denoised gene expression data.
• ConSpaS effectively depicted spatial trajectories by integrating local and global similarities.

Conclusions:

• ConSpaS offers a robust approach for accurate and biologically relevant spatial domain identification in spatial transcriptomics.
• The integration of local and global similarities provides a more comprehensive understanding of tissue architecture.
• This framework advances the analysis of spatial transcriptomics data for biological discovery.