Integrating multi-omics data through deep learning for accurate cancer prognosis prediction

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Individual omics data types are limited by noise and bias for cancer prognosis.
• Integrating multi-omics data is challenging due to high dimensionality and small sample sizes.
• Existing deep learning models for multi-omics integration can be fragile to data noise.

Purpose of the Study:

• To develop a robust framework for integrating multi-omics data for cancer prognosis prediction.
• To improve the accuracy of cancer risk estimation using integrated omics data.
• To identify novel prognostic markers associated with cancer survival.

Main Methods:

• Employed a denoising Autoencoder to extract robust features from multi-omics data.
• Utilized learned features for patient risk estimation.
• Trained XGBoost models using mRNA data for prognosis prediction.
• Validated the model on 15 cancers from The Cancer Genome Atlas (TCGA) and external datasets.

Main Results:

• Achieved an average 6.5% improvement in C-index values over previous methods across 15 cancers.
• XGBoost models trained on mRNA data achieved an average C-index of 0.627.
• Breast cancer model demonstrated significant separation of high-risk from low-risk patients on independent datasets (C-index > 0.6, p < 0.05).
• Identified nine prognostic markers for breast cancer, with seven validated by literature.

Conclusions:

• The developed framework provides an accurate and robust method for multi-omics data integration in cancer prognosis.
• The approach is effective for discovering novel genes associated with cancer prognosis.
• The method shows promise for clinical application in personalized cancer treatment.