Transformed low-rank ANOVA models for high-dimensional variable selection.

Area of Science:

• Biostatistics
• Genomics
• Data Science

Background:

• High-dimensional data analysis is crucial in fields like biomedical research, especially with high-throughput omics data.
• Identifying predictive genetic variables for phenotypes is a key challenge, particularly when variables outnumber samples.
• Conventional regression models struggle with the "large p, small n" problem.

Purpose of the Study:

• To propose a general framework for analyzing high-dimensional data in exponential distribution families.
• To transform variable selection into a well-posed problem, overcoming the "large p, small n" issue.
• To develop and validate a model selection criterion for the proposed framework.

Main Methods:

• Utilizing ANOVA models to express the transformed mean of high-dimensional variables.
• Employing a low-rank interaction space to capture phenotype-variable associations.
• Developing a model selection criterion with proven consistency for diverging parameters.

Main Results:

• The proposed method effectively transforms the variable selection problem into a solvable format.
• Demonstrated superior performance in prediction and detection accuracy via simulations.
• Validated the approach through real-world data analyses, confirming its practical utility.

Conclusions:

• The novel framework offers a robust solution for high-dimensional variable selection in omics and related fields.
• The developed model selection criterion ensures reliable and consistent results.
• The method enhances accuracy in both prediction and detection for complex biological data.