Sparse support vector machines with L₀ approximation for ultra-high dimensional omics data.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Omics data present challenges with ultra-high dimensionality (p) and small sample sizes (n).
• Existing L1-based Support Vector Machines (SVMs) for feature selection are computationally intensive and may exhibit bias.
• Current methods often solve primal formulations, limiting scalability with increasing feature numbers.

Purpose of the Study:

• To develop an efficient sparse SVM method for high-dimensional omics data.
• To address the computational limitations and inconsistencies of L1-based SVMs in feature selection.
• To enable both feature and sample selection through L0 norm approximation.

Main Methods:

• Developed a novel sparse SVM approach using L0 norm approximation.
• Approximated L0 minimization by solving a series of L2 optimization problems formulated with dual variables.
• Focused on estimating n dual variables to find solutions for p primal variables, enhancing efficiency for small sample sizes.

Main Results:

• The L0 approximation method achieves sparsity in both dual and primal variables.
• The proposed method demonstrates efficiency and effectiveness in feature selection for omics data.
• Simulations show comparable performance to existing methods with a reduced number of selected features.

Conclusions:

• The L0 approximated SVM is computationally efficient for high-dimensional omics data.
• This method facilitates effective feature selection and can be applied to metagenomic and gene expression data.
• The approach successfully identifies biologically relevant genes and taxa, offering a scalable solution for omics analysis.