A Sparse Structured Shrinkage Estimator for Nonparametric Varying-Coefficient Model with an Application in Genomics.

Area of Science:

• Genomics
• Statistical modeling
• Bioinformatics

Background:

• Genomic data analysis often involves high-dimensional variable selection.
• Genomic covariates can exhibit graph structures, and biological processes are dynamic.
• Selecting variables in nonparametric models with structured, high-dimensional covariates is a significant challenge.

Purpose of the Study:

• To address the challenge of variable selection in high-dimensional nonparametric varying-coefficient models.
• To develop a method for regression-based motif discovery.
• To introduce a sparse structured shrinkage (SSS) estimator.

Main Methods:

• Utilized basis function expansions for nonparametric modeling.
• Developed a novel smoothed penalty function for the sparse structured shrinkage (SSS) estimator.
• Designed an efficient algorithm for computing the SSS estimator.

Main Results:

• Derived theoretical results on model selection consistency and estimation bounds.
• Demonstrated improved finite-sample performance through simulations, highlighting the impact of high-dimensionality and covariate structure.
• Successfully applied the method to motif discovery using yeast cell-cycle gene expression data.

Conclusions:

• The proposed SSS method enhances variable selection and prediction for high-dimensional regression with nonparametric models and structured covariates.
• The approach is effective for biological applications like motif finding.
• This work advances statistical methods for analyzing complex genomic data.