Integrative factorization of bidimensionally linked matrices.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Molecular omics technologies generate high-content biomedical data, necessitating advanced integration methods.
• Existing methods often limit integration to single dimensions (vertical or horizontal), restricting analysis of bidimensionally linked data.
• Bidimensionally linked data, involving multiple cohorts and platforms, are increasingly prevalent in large-scale biomedical studies.

Purpose of the Study:

• To introduce bidimensional integrative factorization (BIDIFAC), a novel statistical method for integrative dimension reduction and signal approximation.
• To enable the analysis of bidimensionally linked data matrices, capturing both shared and unique patterns of variability.
• To provide a flexible framework for integrating diverse omics data across multiple sample cohorts.

Main Methods:

• BIDIFAC factorizes data into globally shared, row-shared, column-shared, and single-matrix components.
• A penalized objective function, extending nuclear norm penalization, is used for estimation.
• Random matrix theory guides the selection of tuning parameters, circumventing complex rank selection.

Main Results:

• The method was applied to integrate messenger RNA and microRNA expression data from tumor and normal breast cancer tissue samples.
• BIDIFAC successfully decomposed the bidimensionally linked data into distinct structural components.
• The approach facilitates the investigation of complex interrelationships within and between omics platforms and sample cohorts.

Conclusions:

• BIDIFAC offers a powerful approach for integrating bidimensionally linked high-content biomedical data.
• The method enhances the ability to discover shared and unique biological signals across multiple omics and cohorts.
• The provided R code supports the application and further development of BIDIFAC in biomedical research.