PLS-Based and Regularization-Based Methods for the Selection of Relevant Variables in Non-targeted Metabolomics Data

Area of Science:

• Systems Biology
• Metabolomics
• Statistical Bioinformatics

Background:

• Non-targeted metabolomics generates high-dimensional data, necessitating effective variable selection for identifying disease biomarkers.
• Selecting relevant metabolites is critical for accurate classification and biomarker discovery in complex biological samples.

Purpose of the Study:

• To compare the performance of three statistical approaches for variable selection in untargeted metabolomics datasets.
• To evaluate Orthogonal Projections to Latent Structures-Discriminant Analysis (OPLS-DA) with and without multiple testing correction, and Least Absolute Shrinkage and Selection Operator (LASSO) with bootstrapping.
• To identify the most effective method for selecting robust biomarker candidates.

Main Methods:

• Utilized two metabolomics datasets (RH and PH studies).
• Applied OPLS-DA (with and without multiple testing correction) and LASSO with bootstrapping for variable selection.
• Compared variable selection performance based on criteria like Variable Importance in Projection (VIP) and reproducibility.

Main Results:

• OPLS-DA without correction selected a large number of variables (46-320).
• OPLS-DA with multiple testing correction significantly reduced selected variables (4-19).
• LASSO consistently selected a small number of highly reproducible variables (4-14), indicating robust feature selection.

Conclusions:

• LASSO offers a more robust and reproducible approach for variable selection in untargeted metabolomics compared to OPLS-DA.
• The study highlights LASSO's potential for identifying reliable disease biomarker candidates by promoting sparse solutions.
• This research is the first to apply LASSO penalized logistic regression in untargeted metabolomics studies.