Sparse Canonical Correlation Analysis for Multiple Measurements With Latent Trajectories

Area of Science:

• Multivariate Statistics
• Bioinformatics
• Computational Biology

Background:

• Canonical Correlation Analysis (CCA) integrates high-dimensional omics datasets by identifying correlations between observed features.
• Standard CCA requires independent observations, limiting its use with repeated or longitudinal measurements.
• Existing CCA extensions for repeated measures are suboptimal for high-dimensional data and longitudinal analysis.

Purpose of the Study:

• To develop a novel extension of sparse CCA that incorporates time dynamics for analyzing high-dimensional longitudinal data.
• To address the limitations of standard CCA in handling correlated repeated measurements.
• To improve the interpretability and computational efficiency of CCA in omics research.

Main Methods:

• Proposed a novel sparse CCA extension incorporating time dynamics at the latent variable level using longitudinal models.
• Implemented an $\ell _0$ penalty for fixed sparsity levels, enhancing interpretability and computational efficiency.
• Estimated longitudinal trajectories by fitting models to low-dimensional latent variables, leveraging clustered data structures.

Main Results:

• The novel CCA method effectively handles repeated measurements and incorporates time dynamics in high-dimensional datasets.
• The approach provides interpretable longitudinal trajectories, revealing shared latent mechanisms.
• Demonstrated substantial reduction in computational time compared to existing methods for high-dimensional analyses.

Conclusions:

• The proposed CCA method offers an efficient and interpretable solution for analyzing high-dimensional longitudinal omics data with repeated measurements.
• This method enables the estimation of canonical correlations across measurements for clustered data, capturing temporal dynamics.
• The approach is applicable to sparsely and irregularly observed data, as shown with Human Microbiome Project data.