A Contrastive-Learning-Based Deep Neural Network for Cancer Subtyping by Integrating Multi-Omics Data.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Accurate cancer subtyping is vital for prognosis and personalized treatment.
• Multi-omics data offers insights but faces challenges like high dimensionality and small sample sizes, leading to ambiguous subtypes.
• Existing clustering methods struggle with overlapping cancer subtypes.

Purpose of the Study:

• To develop a novel contrastive-learning-based deep learning approach for cancer subtyping.
• To address challenges of high dimensionality and small sample sizes in multi-omics data for patient clustering.
• To improve the accuracy and biological significance of identified cancer subtypes.

Main Methods:

• Proposed an end-to-end deep learning method utilizing contrastive learning and self-supervised learning.
• Applied the method to nine public cancer datasets for patient clustering.
• Developed an XGBoost classification model to evaluate the importance of different omics data types.

Main Results:

• The proposed method demonstrated superior performance in separating patients with different survival outcomes (p < 0.05) compared to existing methods.
• mRNA was identified as the most important omics feature for cancer survival, followed by DNA methylation and miRNA.
• A case study successfully clustered subtypes and identified 14 cancer-related genes, with 12 (85.7%) validated by literature.

Conclusions:

• The developed method effectively identifies statistically and biologically significant cancer subtypes.
• This approach enhances patient stratification and personalized medicine strategies.
• The study provides a valuable tool for cancer molecular subtyping using multi-omics data.