ShadowVIMP: permutation-based multiple testing-controlled variable selection.

Area of Science:

• Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Identifying biomarkers is crucial for precision medicine, especially with high-dimensional data.
• Random Forests (RFs) are effective for high-dimensional data but face challenges in variable selection due to complex VIMP score distributions.
• Standard statistical testing and multiple testing adjustments for RF variable importance (VIMP) are difficult.

Purpose of the Study:

• To introduce shadowVIMP, a novel method for multiple testing-controlled variable selection in RF analysis.
• To address limitations of existing RF variable selection methods, particularly concerning correlated and categorical variables.
• To provide a robust approach for biomarker discovery in high-dimensional datasets.

Main Methods:

• Propose shadowVIMP, a method inspired by permutation testing for variable selection.
• shadowVIMP generates permuted variable counterparts to calculate adjusted p-values for VIMP scores.
• The method preserves the correlation structure between variables, mitigating selection bias.

Main Results:

• shadowVIMP demonstrates improved sensitivity and provides multiple testing-adjusted results in high-dimensional settings.
• The method shows robustness against VIMP biases caused by correlated and categorical variables.
• shadowVIMP can visually annotate VIMP plots for selected variable sets.

Conclusions:

• shadowVIMP offers a promising approach for reliable variable selection in RF analysis.
• It effectively addresses known biases in permutation-based VIMP measures.
• The shadowVIMP R package is available on CRAN for practical application.