Comparison of Penalty Functions for Sparse Canonical Correlation Analysis

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• Canonical correlation analysis (CCA) is standard for assessing associations between variable sets.
• Traditional CCA is unsuitable for high-dimensional genomic data where variables outnumber subjects.
• High variable correlation destabilizes covariance matrices in traditional CCA.

Purpose of the Study:

• To compare four penalty functions (Lasso, Elastic-net, SCAD, Hard-threshold) for sparse canonical correlation analysis (SCCA).
• To evaluate the effectiveness of the Bayesian Information Criterion (BIC) filtering step in SCCA.
• To determine the optimal SCCA method for analyzing genomic data.

Main Methods:

• Implemented and compared Lasso, Elastic-net, SCAD, and Hard-threshold penalties for SCCA.
• Applied SCCA with and without the Bayesian Information Criterion (BIC) filtering step.
• Utilized both simulated and real genotypic and mRNA expression data for validation.

Main Results:

• SCAD penalty demonstrated superior performance in SCCA for genomic datasets.
• The addition of the BIC filtering step further enhanced the selection of relevant features.
• The combination of SCAD penalty and BIC filtering provided robust and interpretable results.

Conclusions:

• The SCAD penalty combined with BIC filtering is the preferred method for sparse canonical correlation analysis in genomics.
• This approach effectively addresses challenges posed by high-dimensional and correlated variables in genomic studies.
• The findings offer valuable guidance for researchers analyzing complex genomic datasets.