Robust Distance Correlation for Variable Screening

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• High-dimensional data analysis is crucial for scientific discovery.
• Traditional methods struggle with computational complexity and ultrahigh-dimensional data.
• Existing sure screening methods often fail to address heavy-tailed data characteristics.

Purpose of the Study:

• To develop a novel sure screening method for heavy-tailed, high-dimensional data.
• To enhance feature selection robustness in modern big data applications.
• To improve the identification of biologically relevant genes in cancer genomics.

Main Methods:

• Introduction of a new sure screening method utilizing robust distance correlation (RDC).
• Estimation of distance correlation robustly in the presence of heavy-tailed data.
• Development of a False Discovery Rate (FDR) control procedure using Reflection via Data Splitting (REDS).

Main Results:

• The proposed RDC-based screening method demonstrates superior performance over existing procedures in simulations with heavy-tailed data.
• The method successfully identified biologically meaningful genes predictive of MAPK1 protein expression in pancreatic cancer.
• Application to The Cancer Genome Atlas (TCGA) RNA-seq data confirmed the method's effectiveness.

Conclusions:

• The novel RDC screening method offers a robust solution for feature selection in high-dimensional, heavy-tailed datasets.
• This approach advances dimensionality reduction techniques for modern big data challenges.
• The method has significant implications for identifying key biomarkers in complex diseases like pancreatic cancer.