Multiple and Optimal Screening Subset: a method selecting global characteristic congeners for robust foodomics...

Area of Science:

• Metabolomics and Foodomics
• Analytical Chemistry
• Chemometrics

Background:

• Traditional feature selection methods in metabolomics and foodomics can lead to arbitrary and suboptimal model results.
• The complexity of food matrices requires advanced analytical techniques for comprehensive molecular analysis.

Purpose of the Study:

• To develop and validate a Multiple and Optimal Screening Subset (MOSS) approach for efficient and effective feature selection.
• To compare the performance of MOSS against traditional methods in statistical model setup for food analysis.
• To achieve a balance between a minimal number of predictors and high predictive accuracy.

Main Methods:

• Developed the Multiple and Optimal Screening Subset (MOSS) approach.
• Compared five statistical models: Student's t-test, ROC curve, PLS-DA, random forests, and SVM.
• Utilized cross-validation to identify feature candidates and determine optimal subset size.
• Analyzed 1406 mass spectral features from 122 bourbon samples.

Main Results:

• MOSS generated a feature subset for bourbon age prediction with 88% accuracy.
• MOSS improved the area under the curve for sweetness prediction to 0.898 using only four predictors.
• Compared to top-ranked features (AUC 0.681), MOSS demonstrated superior performance.

Conclusions:

• The MOSS approach offers an efficient and effective method for optimal feature selection in metabolomics and foodomics.
• MOSS enhances predictive accuracy while minimizing the number of predictors used in statistical models.
• This method provides a robust alternative to traditional, potentially arbitrary, feature selection techniques.