Sparse logistic regression with Lp penalty for biomarker identification.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Sparse logistic regression is crucial for analyzing high-dimensional biological data.
• Existing methods often struggle with non-convex regularization penalties.
• Accurate biomarker identification requires efficient and robust algorithms.

Purpose of the Study:

• To propose novel algorithms for sparse logistic regression with non-convex Lp regularization (p < 1).
• To develop fast and efficient methods applicable to high-dimensional datasets like gene expression.
• To introduce the first algorithms for sparse logistic regression utilizing Lp and elastic net (Le) penalties.

Main Methods:

• Utilizing smooth approximation techniques for algorithm development.
• Implementing non-convex Lp regularization for enhanced sparsity.
• Employing area under the ROC curve (AUC) maximization for regularization parameter selection.

Main Results:

• Demonstrated accuracy, sparsity, and efficiency on methylation and microarray data.
• Identified biomarkers comparable to existing literature findings.
• Lp logistic regression (p < 1) showed superior performance over L1 logistic regression and SCAD SVM.

Conclusions:

• The proposed Lp logistic regression method offers a powerful tool for high-dimensional data analysis.
• The developed algorithms provide an efficient and accurate approach for biomarker discovery.
• This work represents a significant advancement in sparse logistic regression with non-convex penalties.