Graph-based deep learning for integrating single-cell and bulk transcriptomic data to identify clinical cancer...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) and bulk transcriptomic data are crucial for understanding cancer heterogeneity and identifying subtypes.
• Challenges in scRNA-seq data (high dimensionality, sparsity, noise) limit clinical translation.
• Integrating multimodal transcriptomic data is essential for robust cancer subtype identification and outcome prediction.

Purpose of the Study:

• To introduce a novel graph deep learning method, scBGDL, for synergistic integration of scRNA-seq and bulk transcriptomic data.
• To accurately identify cancer subtypes and predict clinical outcomes using integrated transcriptomic data.
• To overcome limitations of scRNA-seq data for clinical applications.

Main Methods:

• Developed scBGDL, a graph-based deep learning framework utilizing Graph Attention Networks, MinCutPool, and Transformer modules.
• Constructed sample-specific gene graphs to model gene-gene interactions and cellular relationships.
• Validated scBGDL across 16 The Cancer Genome Atlas (TCGA) cancer types and three independent multicenter cohorts.

Main Results:

• scBGDL significantly outperformed existing methods in prognostic accuracy across 16 TCGA cancer types (mean C-index: 0.7060).
• Demonstrated robust risk stratification and clinical utility in lung adenocarcinoma, ovarian cancer, and melanoma cohorts.
• Identified key driver edges and uncovered clinically relevant biological interpretations.

Conclusions:

• scBGDL effectively integrates multimodal transcriptomic data for precise cancer subtype identification and prognosis prediction.
• The method offers robust, generalizable performance across diverse cancer types and therapeutic contexts.
• scBGDL advances precision oncology by enabling interpretable biological insights for therapy optimization and biomarker discovery.