Multi-fusion strategy network-guided cancer subtypes discovering based on multi-omics data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Next-generation sequencing and The Cancer Genome Atlas (TCGA) data enable advanced cancer subtype discovery.
• Identifying molecular heterogeneity is crucial for understanding cancer.
• Accurate identification of cancer subtypes aids in personalized medicine.

Purpose of the Study:

• To propose the Self-supervised Multi-fusion Strategy Network (SMMSN) for effective cancer subtype discovery.
• To integrate multi-level and multi-omics data for robust subtype identification.
• To address the challenge of limited labeled data in multi-omics analysis.

Main Methods:

• The SMMSN model fuses single-omics data using Graph Convolutional Networks (GCN) and Stacked Autoencoder Networks (SAE).
• It integrates multi-omics data through various fusion strategies.
• A dual self-supervised method is employed for clustering cancer subtypes from integrated data.

Main Results:

• Experiments on labeled and unlabeled multi-omics datasets demonstrated SMMSN's ability to distinguish potential cancer subtypes.
• Kaplan Meier survival analysis confirmed significant differences among the identified subtypes.
• Case studies on Glioblastoma Multiforme (GBM) and Breast Invasive Carcinoma (BIC) revealed clinically meaningful subtypes.

Conclusions:

• SMMSN successfully discovers clinically relevant cancer subtypes from integrated multi-omics data.
• The model's self-supervised approach effectively handles unlabeled data.
• Findings support SMMSN as a valuable tool for cancer research and precision oncology.