Abstract
The rapid development of multi-omics data has provided new opportunities for cancer subtype classification. Due to the ability to model gene associations by constructing graph among genes, multi-omics cancer classification based on graphs has attracted considerable attention from researchers. However, due to the high-dimensional nature of multi-omics data, existing graph construction methods based on cosine similarity may lead to noise and low-quality edges. Moreover, considering the complexity of associations within and across omics, existing methods either focus solely on multi-omics fusion at the graph structure level or only consider multi-omics fusion at the representation level, making it challenging to comprehensively model this complex relationship. To this end, we propose a novel multi-omics integration method that combines graph-level structure fusion and locus-level feature fusion to enhance the performance of cancer subtype classification (GL-Fusion). In the graph-level structure fusion module, we integrate multi-omics gene-gene graphs using similarity network fusion method and optimize the graph structure with structural entropy and the protein-protein interaction network. In the locus-level feature fusion module, we employ a locus-level graph convolutional network to integrate multi-omics features, learn gene-level fused representations, and predict cancer subtypes. Empirical validation across four publicly accessible datasets (BRCA, HNSC, LGG, THCA) indicated that our method achieved superior performance compared to 12 representative multi-omics cancer subtyping methods. Ablation experiments, experiments with different omics combinations, and evaluation experiments on modeling associations within and across omics were used to further validate the effectiveness of the method.