SGTB: A graph representation learning model combining transformer and BERT for optimizing gene expression analysis in...

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Spatial transcriptomics (ST) offers insights into tissue organization and cellular interactions.
• Existing ST methods struggle with representing complex spatial dependencies and global features.
• Advanced computational frameworks are needed for deep mining of ST data.

Purpose of the Study:

• To propose an innovative spatial multi-scale graph convolutional network (SGTB) for optimizing spatial transcriptomics data representation.
• To address limitations in capturing complex spatial dependencies and global features in current ST analysis methods.
• To develop a novel computational framework for enhanced biological data mining.

Main Methods:

• Integration of Graph Convolutional Networks (GCN), Transformer, and BERT language models.
• GCN extracts hierarchical cell embeddings by capturing spatial dependencies and gene co-expression.
• Transformer layers model global relationships using attention mechanisms, followed by BERT for deep semantic feature extraction.

Main Results:

• Optimized representation of spatial transcriptomics data by fusing GCN, Transformer, and BERT features.
• Enhanced ability to capture complex spatial dependencies and global patterns within tissues.
• Improved feature learning for high-dimensional spatial transcriptomics data.

Conclusions:

• The proposed SGTB model offers a novel computational framework for spatial transcriptomics data analysis.
• SGTB significantly enhances the representation of spatial multi-scale biological data.
• This approach holds value for improving cell type classification and gene regulatory network construction accuracy.