DeFusion: a denoised network regularization framework for multi-omics integration.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Diverse omics data offer comprehensive biological insights but present integration challenges.
• Existing multi-omics integration methods often neglect noise effects and data-specific patterns.
• Uncovering consistent patterns and learning compact representations from heterogeneous omics data remains difficult.

Purpose of the Study:

• To develop a multi-omics integration method that explicitly models noise and data-specific patterns.
• To improve the discovery of consistent patterns and the learning of compact representations from multi-omics data.
• To enhance disease understanding and survival prediction using integrated omics data.

Main Methods:

• A novel multi-omics integration approach is presented, explicitly modeling error terms and noise effects.
• Denoised network regularization is utilized to suppress noise and data-specific patterns.
• An inexact alternating minimization algorithm is employed to solve the optimization problem.

Main Results:

• Comparative simulations demonstrate superior performance in discovering common patterns across data types at varying noise levels.
• Integration of transcriptomics and epigenomics data from The Cancer Genome Atlas reveals survival predictive power.
• The method achieves an average Harrell's C-index of 0.78 for survival prediction in liver hepatocellular carcinoma, outperforming existing methods.
• An aggressive subtype in liver hepatocellular carcinoma was identified and validated using an external dataset.

Conclusions:

• The proposed method effectively integrates multi-omics data by addressing noise and data-specific patterns.
• The learned integrative representation holds significant potential for unsupervised survival prediction and subtype discovery.
• DeFusion demonstrates broad applicability for integrating various omics data types.