AucPR: an AUC-based approach using penalized regression for disease prediction with high-dimensional omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Optimal marker combination is crucial for disease classification and prediction.
• Existing Area Under the Curve (AUC)-based methods for high-dimensional data rely on non-parametric approximations.
• There is a need for parametric AUC-based approaches in high-dimensional settings.

Purpose of the Study:

• To propose a parametric, AUC-based penalized regression approach (AucPR) for high-dimensional data.
• To develop a method for obtaining an optimal linear combination of markers that maximizes AUC.
• To address limitations of non-parametric AUC methods in high-dimensional contexts.

Main Methods:

• AucPR transforms a classical parametric AUC maximizer into a regression framework.
• Penalization regression techniques, specifically lasso and elastic net, are applied.
• The method is evaluated using real microarray and synthetic datasets.

Main Results:

• AucPR demonstrates superior performance compared to penalized logistic regression and non-parametric AUC methods.
• The approach achieves better AUC and sensitivity for a given specificity, especially with correlated genes.
• AucPR effectively performs gene selection and classification.

Conclusions:

• AucPR is a powerful, easily-implementable parametric linear classifier for high-dimensional data.
• The method offers excellent prediction performance for gene selection and disease prediction.
• AucPR is applicable to various high-dimensional omics data, including gene expression, miRNA, and protein data.