Knockoff boosted tree for model-free variable selection.

Area of Science:

• Computational biology
• Statistical genetics
• Machine learning

Background:

• The knockoff filter framework controls false discovery rate (FDR) for variable selection.
• Current knockoff methods are limited to linear regression models.
• Tree-based models, like boosted trees, are widely used in machine learning but lack robust variable selection with FDR control.

Purpose of the Study:

• Extend the knockoff filter framework to boosted tree models for model-free variable selection.
• Develop and evaluate novel methods for generating knockoffs suitable for tree-based models.
• Assess the performance of the proposed methods in controlling Type I errors and improving statistical power.

Main Methods:

• Developed a novel strategy integrating the knockoff method with boosted tree models for model-free variable selection.
• Proposed and evaluated two new knockoff generation methods: sparse covariance and principal component knockoffs.
• Compared the proposed methods against the original knockoff method using simulation tests across various model scenarios (main-effect, interaction, exponential, second-order).

Main Results:

• The extended knockoff method with boosted trees effectively controls Type I errors and enhances power in variable selection without prior model knowledge.
• The novel sparse covariance and principal component knockoff methods demonstrate competitive performance.
• Application to Cancer Genome Atlas (TCGA) gene expression data for tumor purity estimation and classification showed improved discrimination between cancer types.

Conclusions:

• The proposed approach successfully extends the knockoff filter to boosted tree models, enabling FDR-controlled, model-free variable selection.
• The new knockoff generation strategies offer valuable alternatives for enhancing variable selection performance.
• The method shows promise for applications in cancer genomics, improving classification accuracy.