MOFNet: a deep learning framework for multi-omics data fusion in cancer subtype classification

Abstract

Background: cancer exhibits high molecular and clinical heterogeneity, making accurate subtyping essential for personalized treatment. Traditional single-omics approaches often fail to capture this complexity. Multi-omics integration offers a more holistic understanding, but many existing methods either lack interpretability or fail to model cross-omics correlations effectively. Methods: we developed MOFNet, a novel supervised deep learning framework for multi-omics integration, incorporating a similarity graph pooling (SGO) module and a view correlation discovery network (VCDN). MOFNet processes omics data—including mRNA expression, DNA methylation, and miRNA expression—via omics-specific graph learning and cross-omics label space fusion. Three cancer types—breast cancer (BRCA), low-grade glioma (LGG), and stomach adenocarcinoma (STAD)—were analyzed using datasets from the cancer genome atlas (TCGA). Statistical evaluation was performed using accuracy, weighted F1 score, and macro F1 score across stratified training/testing splits. Results: MOFNet achieved superior performance across all datasets. For BRCA, it obtained an accuracy of 85.17%, F1_weighted of 85.36%, and macro F1 of 80.93%, outperforming all baseline models by up to 18.25%. In LGG and STAD, MOFNet also showed robust gains, with maximum improvements of 23.72% and 21.56%, respectively. Omics ablation studies demonstrated enhanced performance with multi-omics integration. Functional enrichment analysis revealed that MOFNet-identified key features were involved in biologically relevant pathways such as cell cycle regulation, synaptic signaling, and ion transport. Conclusions: MOFNet enables scalable and interpretable multi-omics data fusion for cancer subtype classification, significantly improving predictive accuracy while retaining only 25% of input features. The integration of SGO and VCDN modules offers both biological interpretability and computational efficiency. These results suggest MOFNet's promising application in precision oncology and biomarker discovery.