MoJKNet: a jumping knowledge graph framework for multi-omics cancer subtype prediction

Cancer remains one of the leading causes of morbidity and mortality worldwide and poses a major threat to global public health. Despite substantial advances in early diagnosis and therapeutic strategies, patient outcomes vary widely due to the pronounced molecular and clinical heterogeneity of tumors. Accurate identification of cancer subtypes is therefore essential for elucidating tumor heterogeneity, improving prognostic assessment, and enabling precision medicine. In recent years, multi-omics technologies have provided unprecedented opportunities to characterize cancer at multiple molecular layers, including genomic, epigenomic, transcriptomic, and proteomic levels. However, effectively integrating high-dimensional and heterogeneous multi-omics data remains a major challenge. Moreover, many existing graph convolutional network-based integration methods suffer from over-smoothing and limited utilization of deep feature representations, which restrict their ability to capture complex multi-scale relationships inherent in cancer biology. To address these challenges, we propose MoJKNet, a novel multi-omics integration framework for cancer subtype classification. MoJKNet incorporates a jumping knowledge network (JK-Net) to adaptively aggregate node representations across multiple propagation depths, thereby alleviating over-smoothing and enhancing feature extraction within each omics modality. Subsequently, a multimodal autoencoder combined with similarity network fusion (SNF) is employed to capture complementary information across different omics layers. Finally, a graph attention network (GAT) assigns adaptive feature weights to enable accurate cancer subtype prediction. We evaluated MoJKNet on seven cancer types from The Cancer Genome Atlas (TCGA). Experimental results demonstrate that MoJKNet consistently outperforms state-of-the-art methods, including MOGCAN, MOGONET, and MoGCN, in terms of precision, recall, and F1-score, achieving nearly a 10% performance improvement on the COADREAD dataset. Ablation studies further confirm the critical contribution of the jumping knowledge mechanism to improved representation learning. Overall, MoJKNet provides an effective and generalizable solution for multi-omics data integration and cancer subtype classification, with strong potential for downstream biological interpretation and translational applications.