Tan, X.; Chen, X.; Ren, Z.; Tian, R.; Jiang, M.; Yang, D.; Guo, X.

Cancer heterogeneity is recognized as a major barrier to precise molecular subtype classification. Conventional approaches inadequately leverage the complementary nature of multimodal data, resulting in overfitting when handling high-dimensional omics data and constraining the comprehensive characterization of cancer subtype heterogeneity. To address this, Triad-LMF, a novel multi-omics integration framework, is introduced, leveraging a low-rank multimodal fusion mechanism to improve the Accuracy of cancer subtype classification. Triad-LMF consolidates multi-omics datasets, mitigating feature dimension disparities through optimized preprocessing. To enable efficient multimodal information integration, Triad-LMF implements a two-stage hierarchical fusion strategy, where in Local Pairwise Fusion and Global Triadic Fusion are combined via the Two-Feature and Three-way LMF modules, facilitating a progressive transition from local modality interactions to global feature integration. Experimental results demonstrate that Triad-LMF outperforms traditional machine learning methods and previous published methods in classification performance. UMAP verifies that Global Triadic Fusion significantly enhances the ability to discriminate against subtypes in implicit space representation. We further illustrate the effectiveness of Triad-LMF in making full use of multimodal interactive information. Additionally, SHAP feature importance analysis was adopted for extracting the important features. Moreover, across independent datasets, Triad-LMF demonstrates superior generalization capabilities. Triad-LMF offers an efficient and robust framework for multi-omics-driven cancer subtype classification.