DGAT: a dual-graph attention network for inferring spatial protein landscapes from transcriptomics

Area of Science:

• Spatial biology
• Genomics
• Proteomics
• Deep learning

Background:

• Spatial transcriptomics (ST) offers genome-wide RNA profiles within tissue context.
• ST lacks direct protein measurements crucial for understanding cellular function and tissue organization.
• Bridging this gap is essential for comprehensive spatial biology research.

Purpose of the Study:

• To develop a deep learning framework for imputing spatial protein expression from ST data.
• To leverage RNA-protein relationships from integrated datasets.
• To enable protein-level interpretation of spatial transcriptomic data.

Main Methods:

• Developed DGAT (Dual-Graph Attention Network), a deep learning framework.
• Constructed heterogeneous graphs integrating transcriptomic, proteomic, and spatial information using graph attention networks.
• Employed task-specific decoders to reconstruct mRNA and predict protein abundance from a shared latent representation.

Main Results:

• DGAT demonstrated superior protein imputation accuracy compared to existing methods across multiple datasets.
• The framework successfully revealed spatially distinct cell states, immune phenotypes, and tissue architectures.
• DGAT enabled protein-level interpretation from transcriptomics-only spatial data, uncovering insights not apparent from transcriptomics alone.

Conclusions:

• DGAT accurately reconstructs spatial protein landscapes from spatial transcriptomic data.
• The framework enhances the understanding of tissue organization and cellular functions.
• DGAT provides a powerful tool for protein-level interpretation in transcriptomics-only spatial studies.