Sparse generalized linear model with L₀ approximation for feature selection and prediction with big omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Data Mining

Background:

• Feature selection and prediction are crucial for big data mining.
• Existing methods like L1, SCAD, and MC+ do not directly optimize L0 regularization.
• L0 regularization directly penalizes the number of non-zero features, offering a distinct advantage.

Purpose of the Study:

• To develop a novel sparse generalized linear model (GLM) using L0 approximation for feature selection and prediction in big omics data.
• To address the non-convexity of L0 optimization through sequential convex approximations.
• To introduce adaptive ridge algorithms (L0ADRIDGE) for L0-penalized GLM in ultra-high dimensional settings.

Main Methods:

• Developed a novel sparse generalized linear model (GLM) with L0 approximation.
• Approximated the non-convex L0 optimization problem using sequential convex optimizations.
• Introduced adaptive ridge algorithms (L0ADRIDGE) for efficient computation with ultra-high dimensional data.

Main Results:

• The proposed L0ADRIDGE method demonstrates superior performance compared to SCAD and MC+ in simulations.
• Applied to TCGA ovarian cancer data, it identified multilevel gene signatures associated with suboptimal debulking.
• The identified genes showed biological significance and potential clinical importance.

Conclusions:

• The L0ADRIDGE algorithm provides an effective approach for feature selection and prediction in big omics data.
• The method successfully identified clinically relevant gene signatures in ovarian cancer.
• MATLAB software L0ADRIDGE is publicly available for research use.