Variable selection for generalized canonical correlation analysis.

Area of Science:

• Multivariate Statistics
• Bioinformatics
• Computational Biology

Background:

• Regularized Generalized Canonical Correlation Analysis (RGCCA) analyzes relationships across multiple variable sets.
• Identifying relevant variables within each block is crucial for RGCCA component quality and interpretability.
• Existing methods may not adequately address variable selection in multi-set analysis.

Purpose of the Study:

• To extend RGCCA for effective variable selection within blocks.
• To introduce Sparse Generalized Canonical Correlation Analysis (SGCCA) as a unified framework.
• To enhance the analysis of relationships between multiple sets of variables.

Main Methods:

• Proposed Sparse Generalized Canonical Correlation Analysis (SGCCA) by incorporating an L1-penalty into the RGCCA framework.
• Developed a flexible method where blocks are not necessarily fully connected.
• Utilized a simulated dataset and a real-world 3-block dataset (gene expression, CGH, phenotype) for validation.

Main Results:

• SGCCA effectively integrates RGCCA with variable selection, identifying significant variables across blocks.
• The method demonstrated flexibility in handling partially connected blocks.
• Successful application shown on both simulated and complex biological data from glioma patients.

Conclusions:

• SGCCA provides a powerful and flexible approach for multi-set variable relationship analysis.
• The method improves the interpretability and relevance of identified components.
• SGCCA is available as part of the RGCCA package on CRAN for practical applications.