Model-free feature screening for categorical outcomes: Nonlinear effect detection and false discovery rate control

Area of Science:

• Genomics
• Biostatistics
• Bioinformatics

Background:

• High-dimensional genomic data analysis requires effective feature screening to reduce dimensionality and remove redundant features.
• Existing methods often assume linear effects and feature independence, which may not hold true in real-world biological data.
• Selecting continuous features for categorical outcomes in high-dimensional settings remains a challenge.

Purpose of the Study:

• To propose a novel statistical procedure for selecting continuous features associated with a categorical outcome in high-dimensional genomic data.
• To address the limitations of existing methods by accommodating nonlinear effects and feature dependence.
• To identify genes associated with specific cancers using real genomic datasets.

Main Methods:

• A two-step statistical procedure combining a nonparametric significance test based on edge count and a multiple testing procedure.
• The edge-count test is designed to be sensitive to nonlinear effects by directly assessing distributional differences between groups.
• An adapted Efron's procedure is employed to adjust for dependence between features, controlling the false discovery rate.

Main Results:

• The proposed procedure demonstrates strong performance in terms of statistical power and false discovery rate control, as shown through simulations.
• The method successfully identified genes associated with colon, cervical, and prostate cancers in three distinct genomic datasets.
• The edge-count test proved effective in detecting nonlinear relationships, a significant improvement over traditional methods.

Conclusions:

• The developed statistical procedure offers a powerful and flexible approach for feature screening in high-dimensional genomic data.
• This method overcomes key limitations of existing techniques by handling nonlinear effects and feature dependencies.
• The successful application to cancer genomics datasets highlights its utility in identifying disease-associated genes.