SpaVGMC: A Unified Representation Learning Framework via Structural and Semantic Alignment in Spatial Transcriptomics

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatial transcriptomics offers insights into tissue organization and disease by profiling gene expression in situ.
• Current methods struggle with accurate spatial domain identification due to oversmoothing and poor modeling of global structures.

Purpose of the Study:

• To develop SpaVGMC, a unified framework for robust spatial domain identification in spatial transcriptomics data.
• To jointly model structural dependencies and transcriptional semantics for improved biological interpretation.

Main Methods:

• SpaVGMC employs structured variational representation learning with context-aware message-passing.
• Structural information alignment preserves topological fidelity using mutual information.
• Semantic alignment organizes representations via distribution-aware contrastive learning.

Main Results:

• SpaVGMC consistently outperforms existing methods in spatial domain identification across diverse datasets.
• The framework demonstrates improved agreement with tissue structures and enhanced detection of subdomains.
• Learned embeddings are robust, discriminative, and biologically interpretable.

Conclusions:

• SpaVGMC provides a robust and scalable framework for spatial omics analysis.
• The joint modeling of structure and semantics significantly advances spatial domain identification.
• This approach enhances the understanding of tissue organization and disease mechanisms.