CAGNet: a structure-aware clustering-alternated graph network for cell-cell interaction inference in spatial...

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Understanding cell-cell interactions (CCIs) is vital for spatial transcriptomics.
• Current graph-based models struggle with dynamic cellular relationships due to static clustering.
• Limitations exist in capturing dynamic cellular relationships in tissue organization.

Purpose of the Study:

• To propose CAGNet, a novel two-stage framework for robust CCI inference.
• To overcome limitations of static models in spatial transcriptomics analysis.
• To improve the understanding of spatial organization and functional heterogeneity in tissues.

Main Methods:

• CAGNet employs a two-stage framework for CCI inference.
• Stage 1: Graph Attention Network encoder for structure-aware node embeddings.
• Stage 2: Alternating optimization for iterative refinement of cluster centers and node embeddings.

Main Results:

• CAGNet consistently outperforms six CCI inference baselines on three 10x Genomics Visium datasets.
• Achieved the highest Adjusted Rand Index for spatial domain identification, indicating biologically relevant embeddings.
• Information-theoretic analysis confirms superior mutual information retention in learned embeddings.

Conclusions:

• CAGNet provides a powerful new approach for cell-cell interaction inference.
• The framework effectively captures dynamic cellular relationships and spatial organization.
• CAGNet enhances the analysis of spatial transcriptomics data for biological discovery.