Statistical model choice including variable selection based on variable importance: A relevant way for biomarkers...

Area of Science:

• Computational statistics
• Bioinformatics
• Animal science

Background:

• Predicting numerical variables requires selecting appropriate regression models and relevant explanatory variables.
• Existing methods may not efficiently integrate model selection with variable selection.
• Identifying predictive biomarkers is crucial in fields like food science and animal breeding.

Purpose of the Study:

• To develop and validate a computational methodology for simultaneous regression model and variable selection.
• To implement this methodology in an R package for broad applicability.
• To identify protein biomarkers predictive of meat tenderness in cattle.

Main Methods:

• A computational approach combining parametric, semi-parametric, and non-parametric regression models (multiple linear regression, sliced inverse regression, random forests).
• Variable importance assessment using random perturbations to select covariates above a threshold.
• Learning/test sample approach to estimate Mean Square Error for model and variable selection accuracy.
• Development of the R package 'modvarsel' (MODel and VARiable SELection).

Main Results:

• The methodology demonstrated good performance on simulated data.
• The R package 'modvarsel' was successfully applied to a real-world dataset.
• Linear regression was identified as the best model for predicting meat tenderness.
• Heat shock proteins and metabolic proteins were confirmed as the predominant predictive biomarkers for meat tenderness.

Conclusions:

• The developed computational methodology provides an effective tool for integrated model and variable selection in regression analysis.
• The 'modvarsel' R package offers a practical implementation for diverse datasets.
• The study successfully identified key protein biomarkers related to meat tenderness, highlighting the importance of heat shock and metabolic proteins.