Trans-Driver: A Deep Learning Approach for Cancer Driver Gene Discovery With Multi-Omics Data

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Cancer driver genes are critical for tumor growth and therapeutic targeting.
• Identifying driver genes is challenging due to the complexity and heterogeneity of multi-omics data.

Purpose of the Study:

• To develop a novel deep learning method, Transformer-Driver (Trans-Driver), for accurate cancer driver gene identification.
• To integrate diverse multi-omics data for enhanced driver gene discovery.

Main Methods:

• Proposed Trans-Driver, a deep supervised learning method utilizing a novel transformer architecture.
• Implemented a kernel-based multi-head self-attention mechanism with gated residual connections and Dynamic Tanh (DyT) normalization for heterogeneous feature integration.
• Evaluated performance on TCGA, CGC, and PCAWG datasets.

Main Results:

• Trans-Driver demonstrated superior performance compared to existing state-of-the-art methods.
• Identified 269 candidate driver genes from ~20,000 protein-coding genes, with 132 (49.1%) validated against the CGC dataset.
• Feature contribution analysis confirmed the benefit of integrating multi-omics data over using somatic mutations alone.

Conclusions:

• Trans-Driver effectively integrates multi-omics data for robust cancer driver gene identification.
• The identified candidate driver genes hold significant clinical relevance, highlighting the method's practical value in cancer research and therapy development.