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

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer's global health impact is significant, with outcomes varying due to tumor heterogeneity.
• Accurate cancer subtype identification is crucial for prognosis and precision medicine.
• Multi-omics technologies offer deep cancer characterization but data integration remains challenging.

Purpose of the Study:

• To develop MoJKNet, a novel multi-omics integration framework for precise cancer subtype classification.
• To overcome limitations of existing methods, such as over-smoothing and inadequate feature representation in graph convolutional networks.
• To improve the accuracy of cancer subtype prediction by effectively integrating diverse molecular data.

Main Methods:

• MoJKNet utilizes a jumping knowledge network (JK-Net) to aggregate node representations and reduce over-smoothing.
• A multimodal autoencoder and similarity network fusion (SNF) capture cross-omics information.
• A graph attention network (GAT) assigns adaptive weights for accurate prediction.

Main Results:

• MoJKNet demonstrated superior performance over state-of-the-art methods on seven cancer types from The Cancer Genome Atlas (TCGA).
• The framework achieved nearly 10% improvement on the COADREAD dataset in precision, recall, and F1-score.
• Ablation studies validated the effectiveness of the jumping knowledge mechanism in representation learning.

Conclusions:

• MoJKNet offers an effective and generalizable solution for multi-omics data integration and cancer subtype classification.
• The framework shows potential for advancing biological interpretation and translational applications in oncology.
• Improved cancer subtype classification using MoJKNet can facilitate more personalized treatment strategies.