A whitening approach to probabilistic canonical correlation analysis for omics data integration

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Canonical Correlation Analysis (CCA) is a statistical method for analyzing relationships between two sets of variables.
• CCA has broad applications across various scientific disciplines, including biology, medicine, and social sciences.
• Recent advancements focus on developing probabilistic models for CCA to enhance its applicability in large-scale data analysis.

Purpose of the Study:

• To propose a new probabilistic model for Canonical Correlation Analysis (CCA) based on statistical whitening.
• To enhance the interpretability and flexibility of CCA by addressing limitations of existing methods.
• To develop computationally efficient methods for applying probabilistic CCA to high-dimensional data.

Main Methods:

• Developed a two-layer latent variable generative model for probabilistic CCA.
• Utilized statistical whitening of random variables as a core component of the proposed model.
• Employed regularized inference for computationally efficient estimation in high-dimensional settings.

Main Results:

• The proposed probabilistic CCA model offers non-ambiguous latent variables and accommodates negative canonical correlations.
• The model allows for non-normal generative variables and provides enhanced interpretability.
• Demonstrated efficient estimation in high-dimensional data and applied the method to omics datasets integrating gene expression, lipids, and methylation data.

Conclusions:

• The whitening approach provides a unifying framework for CCA, connecting sphering, multivariate regression, and probabilistic models.
• An R package named "whitening" is available for implementing the proposed CCA method.
• This work advances CCA by offering a more robust, interpretable, and computationally efficient probabilistic framework.