Identifying Cancer Driver Genes Based on Vision Transformer and Hierarchical Feature Fusion Module

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cancer remains a significant global health challenge, necessitating improved methods for identifying cancer driver genes.
• Understanding driver genes is crucial for elucidating cancer development mechanisms.
• Current deep learning methods for driver gene identification often focus on local structures, limiting comprehensive analysis.

Purpose of the Study:

• To develop an advanced computational method, VGDriver, for more accurate prediction of cancer driver genes.
• To enhance the global learning capacity in cancer driver gene identification by integrating multi-omics data.

Main Methods:

• Constructed a gene property feature matrix using gene expression, copy number aberrations, and single nucleotide variations.
• Designed a hierarchical feature fusion module combining graph convolutional networks and multilayer perceptrons for multi-source feature integration.
• Utilized Vision Transformer (ViT) to process the multi-source feature matrix for driver gene prediction.

Main Results:

• VGDriver demonstrated superior performance compared to seven existing methods across five pan-cancer and eight single-cancer datasets.
• The method achieved high evaluation metrics, including AUC and AUPR, indicating robust predictive accuracy.
• The hierarchical feature fusion and ViT integration effectively captured complex genomic relationships.

Conclusions:

• VGDriver offers a powerful new approach for identifying cancer driver genes, outperforming current state-of-the-art methods.
• The integration of multi-omics data through hierarchical fusion and ViT is effective for understanding cancer genomics.
• This method holds promise for advancing cancer research and therapeutic target discovery.