Multi2Fusion: A Multiomics Fusion Framework with Multilevel Information Integration for Cancer Subtype Classification

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate cancer subtype classification is vital for personalized medicine.
• High-throughput sequencing generates massive multiomics data for cancer research.
• Existing fusion methods have limitations in capturing comprehensive cross-omics information.

Purpose of the Study:

• To develop a multiomics fusion framework, Multi2Fusion, for improved cancer subtype classification.
• To fully exploit deep dependencies within and across different omics data modalities.
• To enhance the integration of complementary information from multiomics data.

Main Methods:

• Constructed a sample-gene heterogeneity graph combined with heterogeneous graph convolution.
• Implemented a dual-level fusion mechanism integrating feature and label levels.
• Employed adaptive fusion of omics features and joint optimization of multiview predictions.

Main Results:

• Multi2Fusion demonstrated superior performance across all classification metrics compared to existing methods.
• The framework effectively captured structured intraomics features and high-order cross-omics associations.
• Validated effectiveness on four public cancer multiomics datasets.

Conclusions:

• Multi2Fusion offers a powerful approach for multiomics data integration in cancer research.
• The multilevel information integration framework significantly improves cancer subtype classification accuracy.
• This method holds promise for advancing personalized cancer treatment strategies.