Transcriptome graph transformer: a graph transformer-based unsupervised model for transcriptome data analysis.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Transcriptomic datasets are rapidly growing, posing challenges for traditional analysis due to high dimensionality and complex gene relationships.
• Existing deep learning models struggle with fixed-length inputs and integrating biological network information.

Purpose of the Study:

• To develop an unsupervised graph Transformer framework for transcriptomic data analysis.
• To integrate biological network information with gene expression data for improved analytical performance.

Main Methods:

• Introduced Transcriptome Graph Transformer (TGT), an unsupervised graph Transformer.
• Constructed a heterogeneous gene-pathway graph using expression data, STRING interactions, and pathway annotations (GO/KEGG/Reactome).
• Pretrained TGT with masked-node prediction and fine-tuned for disease classification, biomarker discovery, and zero-shot clustering.

Main Results:

• TGT demonstrated superior performance on Alzheimer's disease, cancer, acute kidney injury, and COVID-19 datasets.
• The model generalized well across different platforms and outperformed state-of-the-art baselines.
• TGT provided biologically meaningful gene and pathway importance scores aligned with known disease mechanisms.

Conclusions:

• TGT offers an effective and generalizable approach for transcriptomic representation learning.
• Integrating biological network knowledge with graph Transformer architectures enhances analytical capabilities.
• TGT shows significant utility for broad transcriptomic applications and advancing precision medicine.