Identification of cancer driver genes is crucial for understanding the molecular mechanisms of cancer. To address the limitations of graph convolutional networks-based cancer driver gene identification methods, including biased prediction results caused by convolutional layer structures that focus more on either the structural characteristics (e.g., degree) or biological features (e.g., mutation frequency) of nodes in the network, as well as sparse supervisory information, we propose a method called Multi-Task Graph Contrastive Learning (MTGCL) for the identification of cancer driver genes. MTGCL designs a new graph convolutional layer structure which can improve the performance of cancer driver gene identification by effectively integrating graph structure topology information and node features information, while a semi-supervised graph contrastive learning task is presented as a regularizer within a multitask learning paradigm to enhance the performance of the main task of driver gene identification by utilizing a small portion of labeled nodes and a large amount of unlabeled nodes information. The experimental results on pan-cancer and some specific cancers demonstrate the effectiveness of MTGCL. In addition, we also find the features of different mutation types derived from somatic mutation data can effectively improve the performance of identifying driver genes for some specific cancer types.
