Abstract: Cancer driver genes played a crucial regulatory role in the occurrence and development of cancer, and the identification was of great significance for revealing the molecular mechanisms of cancer and promoting precision diagnosis and treatment. Existing methods for cancer driver gene identification face challenges such as loss of semantic information in gene similarity matrices, single scale of protein-protein interaction (PPI) network topology modeling, and insufficient synergy of multi-view features, which limit the recognition performance of the models. Therefore, a cancer driver gene identification model based on cross-view collaborative networks (CVC-CDG) is proposed. First, a high-dimensional similarity semantic mapping strategy is used to treat the multi-source gene similarity matrix as a high-dimensional continuous semantic space to fully preserve the subtle functional differences between genes. Second, an embedding-diffusion coupled topological view is constructed, integrating local structural embeddings extracted by Node2Vec with globally weighted edge structures generated by personalized PageRank (PPR) diffusion to achieve collaborative modeling of local and global topological information of the PPI network. Finally, a cross-view collaborative interactive attention network was designed to achieve dynamic interaction and complementary enhancement of multi-view features before graph convolution aggregation, thereby further improving the recognition performance of cancer driver genes. Experimental results show that the CVC-CDG model outperforms the latest deepCDG model on three PPI networks: STRING, CPDB, and PathNet. Its AUC and AUPR reach 94.08% and 86.59% on the STRING network, 93.24% and 86.47% on the CPDB network, and 91.00% and 90.51% on the PathNet network, respectively, demonstrating good robustness and generalization ability.

摘要： 癌症驱动基因在癌症发生与发展过程中起着关键调控作用，其识别对于揭示癌症分子机制和推动精准诊疗具有重要意义。现有方法在癌症驱动基因识别任务中面临着基因相似度矩阵语义信息丢失、蛋白质相互作用(PPI)网络拓扑建模尺度单一以及多视图特征协同不足的问题，制约了模型的识别性能。因此，提出了一种基于跨视图协同网络的癌症驱动基因识别模型(CVC-CDG)。首先，通过高维相似度语义映射策略，将多源基因相似度矩阵作为高维连续语义空间，以充分保留基因间细微的功能关联差异；其次，构建嵌入-扩散耦合的拓扑结构视图，融合Node2Vec提取的局部结构嵌入与个性化网页排序(PPR)算法扩散生成的全局加权边结构，实现对PPI网络局部与全局拓扑信息的协同建模；最后，设计跨视图协同交互注意力网络，在图卷积聚合前实现多视图特征的动态交互与互补增强，从而进一步提升了癌症驱动基因的识别性能。实验结果表明，CVC-CDG模型在STRING、CPDB和PathNet三类PPI网络上均取得了优于最新deepCDG模型的性能，其AUC和AUPR分别在STRING网络上达到94.08%和86.59%，在CPDB网络上达到93.24%和86.47%，在PathNet网络上达到91.00%和90.51%，展现出良好的稳健性与泛化能力。