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

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Cancer's heterogeneity necessitates precise subtyping for personalized treatment.
• Traditional single-omics methods struggle to capture cancer's complexity.
• Multi-omics integration offers a comprehensive view but faces challenges in interpretability and cross-omics correlation modeling.

Purpose of the Study:

• To develop a novel deep learning framework for multi-omics integration in cancer.
• To improve the accuracy and interpretability of cancer subtype classification.
• To enable scalable fusion of diverse omics data for precision oncology.

Main Methods:

• Developed MOFNet, a supervised deep learning framework utilizing similarity graph pooling (SGO) and a view correlation discovery network (VCDN).
• Processed mRNA expression, DNA methylation, and miRNA expression data using omics-specific graph learning and cross-omics label space fusion.
• Applied and evaluated MOFNet on three cancer types (BRCA, LGG, STAD) from The Cancer Genome Atlas (TCGA).

Main Results:

• MOFNet significantly outperformed baseline models across all tested cancer datasets (BRCA, LGG, STAD).
• Achieved high accuracy (e.g., 85.17% for BRCA) and improved F1 scores, with maximum gains up to 23.72%.
• Omics ablation studies confirmed the benefit of multi-omics integration, and identified key features linked to relevant biological pathways.

Conclusions:

• MOFNet provides a scalable and interpretable solution for multi-omics data fusion in cancer.
• The framework enhances predictive accuracy for cancer subtype classification while reducing feature complexity.
• MOFNet shows significant potential for applications in precision oncology and biomarker discovery.