Accurate identification of cancer driver genes is critical for understanding cancer mechanisms and advancing precision oncology. We propose a graph neural network framework that integrates diverse biological features and employs two-stage pretraining to improve cancer driver gene prediction. Our model incorporates multi-omics signals and pathway-level information to represent genes and captures biologically meaningful relationships through graph-based learning. Experiments on 12 types of cancer demonstrate that our method consistently outperforms existing deep learning approaches. Ablation studies confirm the importance of key components, including pathway features, expression-based connectivity, and pretraining. A case study on breast cancer further highlights the ability of the model to prioritize novel candidate driver genes. These results underscore the value of integrating heterogeneous biological knowledge into graph learning for gene discovery in cancer.