Robust prioritization of genomic features with stability selection.

Area of Science:

• Genomics
• Biostatistics
• Computational Biology

Background:

• Complex diseases like cancer exhibit heavy-tailed trait distributions, making non-robust variable selection methods susceptible to data contamination and outliers.
• Existing methods that use pseudo-features for error control can exacerbate the curse of dimensionality in high-dimensional genomic data.

Purpose of the Study:

• To develop a robust variable selection framework that enhances feature prioritization in the presence of data contamination and outliers.
• To address the limitations of existing methods by improving robustness and mitigating the curse of dimensionality.

Main Methods:

• A novel robust variable selection framework incorporating stability selection.
• Utilizes least absolute deviation (LAD) LASSO for robustness against outliers and heavy-tailed errors.
• Avoids augmenting the genotype matrix with pseudo-features, thus mitigating the curse of dimensionality.

Main Results:

• The proposed method demonstrates superior performance in simulation studies compared to competing variable selection techniques.
• Application to The Cancer Genome Atlas (TCGA) Skin Cutaneous Melanoma (SKCM) and eQTL datasets shows enhanced identification of reproducible genomic features.
• Achieves double robustness by combining LAD LASSO with the avoidance of pseudo-features.

Conclusions:

• The developed robust framework offers a more reliable approach for genomic feature selection in complex diseases with heterogeneous traits.
• The method provides superior performance and reproducibility, particularly in high-dimensional genomic datasets prone to contamination.