Representation Learning for the Clustering of Multi-Omics Data.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Science

Background:

• Multi-omics data integration is crucial for identifying complex disease subtypes.
• High-dimensional and heterogeneous multi-omics datasets require effective representation learning techniques.
• Existing methods for integrative clustering face challenges in balancing contributions from diverse data sources.

Purpose of the Study:

• To develop a unified framework for representation learning and clustering of multi-omics data.
• To compare statistical (group factor analysis) and deep learning (autoencoders) approaches for data integration.
• To identify biologically meaningful patient clusters by effectively leveraging information from multiple data sources.

Main Methods:

• Investigated representation learning techniques including Principal Component Analysis (PCA), Multiple Factor Analysis (MFA), and autoencoders.
• Proposed a novel disjointed deep autoencoder (DDAE) architecture with layer-wise reconstruction loss.
• Introduced a weighted internal clustering index for comparing statistical and deep learning models.

Main Results:

• The proposed DDAE architecture with layer-wise reconstruction loss demonstrated effective representation learning.
• The weighted internal clustering index provided a unified framework for model comparison.
• Applied methodology to TCGA Breast Cancer and TARGET Neuroblastoma datasets, yielding well-balanced clusters.

Conclusions:

• The developed framework offers a robust approach for multi-omics data integration and patient stratification.
• The novel DDAE architecture and weighted index facilitate improved identification of disease subtypes.
• The method successfully identified biologically meaningful clusters across different datasets, outperforming previous integrative clustering studies.