Evaluation of multiple variate selection methods from a biological perspective: a nutrigenomics case study.

Area of Science:

• Genomics and Bioinformatics
• Nutrigenomics
• Statistical Modeling

Background:

• Genomics technologies generate vast datasets requiring robust interpretation methods.
• Multivariate regression and variate selection are crucial for identifying phenotype-related genetic factors.
• Biological interpretability of these models often remains a challenge.

Purpose of the Study:

• To compare the performance, utility, and biological interpretability of various multivariate regression and variate selection methods.
• To apply these methods to a nutrigenomics dataset from a high-fat diet study in ApoE3Leiden mice.
• To evaluate gene ranking approaches for pathway analysis.

Main Methods:

• Analysis of hepatic transcriptome and plasma protein data (Leptin, TIMP-1).
• Comparison of Partial Least Squares (PLS), Genetic Algorithm-based Multiple Linear Regression (GA-MLR), LASSO, ELASTIC NET, and Covariance-based PLS (CovProc).
• Evaluation of gene ranking via correlation with protein data versus PLS regression coefficient stability for Gene Set Enrichment Analysis (GSEA).

Main Results:

• Regression methods showed similar statistical performance; CovProc and GA-MLR performed best and worst, respectively.
• CovProc, LASSO, and ELASTIC NET yielded parsimonious models and consistent identification of key variates.
• Correlation-based gene ranking showed high agreement with PLS-based ranking but identified more significant gene sets.

Conclusions:

• Covariance-based variate selection (CovProc) is recommended for selecting important genes in nutrigenomics studies.
• Combining CovProc with univariate methods enhances biological data interpretation.
• Correlation-based ranking is suitable for GSEA-like pathway analyses.