A powerful representation learning method for enhanced analysis of incomplete multi-omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Integrative multi-omics analysis offers deeper biological insights but faces challenges like high dimensionality and missing data.
• Current methods for incomplete multi-omics data are limited, failing to fully utilize shared and specific information, leading to biased representations.

Purpose of the Study:

• To introduce JASMINE, a novel self-supervised representation learning framework for handling incomplete multi-omics data.
• To develop a method that preserves both modality-specific and joint information while improving sample similarity structure.

Main Methods:

• JASMINE employs a self-supervised learning approach to generate embeddings from incomplete multi-omics datasets.
• The method is designed to capture both unique features of each data modality and shared biological signals across modalities.

Main Results:

• JASMINE demonstrated superior performance across various downstream tasks on two distinct incomplete multi-omics datasets.
• The learned representations effectively preserved modality-specific and joint information, enhancing sample similarity.
• The method requires only a single training phase per dataset, indicating computational efficiency.

Conclusions:

• JASMINE provides a robust solution for analyzing incomplete multi-omics data, overcoming limitations of existing methods.
• The framework generates high-quality, task-unbiased embeddings, advancing biological data integration and interpretation.