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

Accurate cancer subtype classification is crucial for guiding personalized clinical treatment. With the rapid advancement of high-throughput sequencing technologies, the massive multiomics data have provided a solid foundation for accurate cancer subtype classification. However, existing multiomics fusion methods typically rely on single-level fusion at either the feature level or the label level, making it difficult to simultaneously capture structured intraomics features and high-order associations among predictions from different omics, which limits the full exploitation of complementary information across omics. Therefore, we present a multiomics fusion framework with multilevel information integration, Multi2Fusion, which is designed to fully exploit deep dependencies both within and across modalities in multiomics data for cancer subtype classification. Specifically, Multi2Fusion constructs a sample-gene heterogeneity graph and combines it with heterogeneous graph convolution to extract structured features within the omics data. Concurrently, it designs a dual-level fusion mechanism at the feature and label levels, through adaptive fusion of omics features and joint optimization of multiview prediction results, to achieve the full integration of cross-omics information and improve classification performance. Extensive experiments on four publicly available cancer multiomics data sets have shown that Multi2Fusion outperforms existing methods in all classification metrics, validating its effectiveness in the task of cancer subtype classification.