Cancer Subtype Classification Based on Graph Convolutional Network with Enhanced Structural Feature

Graph Convolutional Networks (GCNs) are biased toward local data information during training while neglecting global structural feature. Therefore, this paper investigates a cancer subtype classification method based on a GCN with enhanced structural feature. Aiming at genomics and transcrip-tomes datasets for stomach adenocarcinoma samples obtained from the Cancer Genome Atlas (TCGA), the proposed approach relies on a weighted autoencoder with a shared hidden layer to integrate multi-omics data feature. Subsequently, a Self-Organizing Mapping network is employed to learn the global structural feature of samples based on Geodesic distance. The integrated multi-omics feature and global structural feature serve as feature inputs for the GCN. At the same time, the input graph for the GCN is generated through Similarity Network Fusion algorithm. Finally, the classification of cancer samples is accomplished after training. Comparative experimental results demonstrate that the proposed model outperforms several popular algorithms in cancer subtype classification, achieving the highest accuracy. This is because enhancing global structural feature allows the model to capture the samples’ local and global structural feature, leading to more reasonable classification results.