Predicting Cancer Driver Genes Using Contrastive Graph Diffusion and Dynamic Weighting Adjustment

Cancer development is often driven by gene mutations. Accurately identifying driver genes containing mutations is crucial for understanding cancer biology and developing targeted therapies. Many methods for identifying cancer driver genes use Protein-Protein Interaction (PPI) networks, but these networks usually contain noise hindering accuracy. To address this, we design a Diffusion Graph Contrastive Learning method called DGCL, which leverages graph diffusion techniques and contrastive learning to extract robust gene representations from the PPI network for cancer driver gene identification. DGCL employs personalized PageRank graph diffusion to create an enhanced view of the PPI network, preserving the original structure while revealing hidden biological connections. Secondly, we leverage Chebyshev graph convolution to extract features from both networks. By applying neighborhood contrastive learning, we harmonize gene representations across different contexts, mitigating noise in the PPI network. Then, we refine network-specific features using Chebyshev graph convolutions and constrain them with node classification and link prediction tasks. Additionally, we extend DGCL by introducing a dynamic weight adjustment (DWA) strategy to balance task-specific losses in the model training. The extended method is DGCL_DWA. Finally, logistic regression is then applied to predict cancer driver genes using the gene features. Compared with the state-of-the-art methods, experimental results show that our DGCL and DGCL_DWA models perform excellently in identifying driver genes for both pan-cancer and specific cancer types. Moreover, the ablation experiments prove that the diffusion graph network, contrastive learning, and the dynamic weight adjustment strategy positively enhance predictive performance.