SpaMask: Dual masking graph autoencoder with contrastive learning for spatial transcriptomics.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Spatial resolved transcriptomics (SRT) is vital for understanding tissue organization and cellular diversity.
• Spatial domain characterization is a critical initial step in SRT data analysis.
• Graph neural networks (GNNs) are commonly used for SRT analysis but struggle with data sparsity.

Purpose of the Study:

• To develop a robust GNN-based method for spatial domain characterization in SRT data.
• To enhance the performance and stability of GNNs in the presence of sparse SRT data.
• To improve clustering accuracy and batch correction in SRT analysis.

Main Methods:

• Proposed SpaMask, a dual masking graph autoencoder with contrastive learning for SRT.
• Implemented node masking in Masked Graph Autoencoders (MGAE) to leverage spatial neighbors.
• Applied edge masking in Masked Graph Contrastive Learning (MGCL) to tighten embeddings of spatially proximate nodes.

Main Results:

• SpaMask demonstrated superior clustering accuracy compared to existing methods across eight diverse datasets.
• The dual masking approach enhanced model robustness against the inherent sparsity of SRT data.
• SpaMask achieved effective batch correction, improving data integration and comparability.

Conclusions:

• SpaMask offers a robust and accurate solution for spatial domain characterization in SRT.
• The proposed masking strategies effectively mitigate challenges posed by sparse SRT data.
• SpaMask advances the analysis of spatial transcriptomics, enabling deeper biological insights.