SCTGE infers transformer-based graph embeddings to improve cell-cell interaction identification and cell identity...

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Single-cell transcriptomic data analysis presents challenges in spatial embedding interpretation due to high dimensionality and noise.
• Existing graph-based models show promise but face limitations in scalability and computational efficiency.

Purpose of the Study:

• To introduce SCTGE (Single-Cell Transformer-based Graph Embeddings), a novel architecture for spatially aware single-cell embeddings.
• To address scalability and noise challenges in spatial transcriptomics data analysis.

Main Methods:

• Integration of transformer-based self-attention with graph neural networks.
• Implementation of adaptive multi-scale attention and graph-specific encodings.
• Utilizing context-aware adaptive multi-head attention and APPNP-enhanced feature propagation.

Main Results:

• SCTGE demonstrated superior performance in cell-cell interaction prediction compared to state-of-the-art models.
• SCTGE generally outperformed or showed improvements over seven scRNA-seq tools for cell identity annotation across six datasets.

Conclusions:

• SCTGE offers a scalable and interpretable solution for spatial embedding extraction in single-cell analysis.
• The framework advances intercellular communication analysis and cell identity annotation in spatial transcriptomics.