KOBT

KOBT performs model-free variable selection by integrating the knockoff filter framework with boosted tree models to control the false discovery rate (FDR) in complex datasets.

Key Features:

• Model-free variable selection: Selects informative variables without requiring a predefined model topology.
• Integration with boosted trees: Leverages boosted tree models to generate variable importance measures for selection in nonlinear and high-dimensional settings.
• Sparse covariance and principal component knockoffs: Implements two knockoff generation methods—sparse covariance knockoffs and principal component knockoffs—to create imitation variables that preserve correlation structure.
• Control of FDR and type I errors: Applies the knockoff filter framework to maintain false discovery rate control and limit type I errors while seeking high statistical power.
• Evaluation of importance statistics: Computes and compares importance test statistics derived from tree models to inform variable selection decisions.
• Comprehensive simulation testing: Assesses performance across simulated scenarios including main-effect, interaction, exponential, and second-order models.

Scientific Applications:

• Tumor purity estimation: Applied to tumor purity estimation tasks using genomic data.
• Tumor classification and discrimination: Applied to tumor classification using TCGA gene expression data and demonstrated improved discrimination between cancer types that are difficult to distinguish.

Methodology:

KOBT extends the knockoff method by integrating boosted tree models; implements sparse covariance and principal component knockoff generation methods; evaluates importance test statistics from tree models and compares combinations of knockoffs and statistics to optimize performance; performance is assessed via simulations of main-effect, interaction, exponential, and second-order models while controlling FDR and type I error and aiming to maximize power.