Precision Lasso: accounting for correlations and linear dependencies in high-dimensional genomic data

Area of Science:

• Bioinformatics and computational biology
• Genomic data analysis
• Statistical genetics

Background:

• Association studies are crucial for linking genetic markers to phenotypes in bioinformatics.
• Regularized regression methods like Lasso are popular but struggle with correlated and linearly dependent variables common in genomic data.
• This instability leads to under-performance in variable selection for classical methods.

Purpose of the Study:

• To develop a robust variable selection method that addresses the limitations of existing techniques in the presence of correlated and linearly dependent variables.
• To introduce the Precision Lasso, a novel variant of the Lasso regression designed for stable and consistent variable selection.

Main Methods:

• Proposed the Precision Lasso, a modified Lasso regression incorporating covariance and inverse covariance matrices of explanatory variables for regularization.
• Evaluated the method's performance using simulated datasets with highly correlated and linearly dependent variables.
• Applied the Precision Lasso to transcriptomic data from breast cancer patients to identify meaningful biological variables.

Main Results:

• The Precision Lasso demonstrated stable and consistent variable selection in simulations with challenging data structures.
• Empirical results showed the Precision Lasso's effectiveness in selecting meaningful variables from breast cancer transcriptomic profiles.
• Precision Lasso outperformed established methods like Lasso, Elastic Net, and MCP regression in settings with correlated and linearly dependent variables.

Conclusions:

• The Precision Lasso offers a significant improvement over existing methods for variable selection in bioinformatics, particularly for complex genomic datasets.
• The developed method provides a more reliable approach to identifying genetic associations and potential biomarkers.
• The software for Precision Lasso is publicly available for broader research application.