Statistical knockoffs improve biomarker discovery from transcriptomic data

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• High-throughput sequencing generates vast biological data, enabling gene-phenotype association studies.
• Variable selection in high-dimensional, correlated data (p >> n) presents significant challenges for identifying true associations.

Purpose of the Study:

• To evaluate the knockoff (KO) procedure for variable selection in high-dimensional transcriptomic classification.
• To assess the KO framework's performance against established variable selection models using simulated and real data.

Main Methods:

• The study applied the knockoff (KO) variable selection procedure, designed to control the false discovery rate while accounting for variable correlations.
• Extensive simulations were conducted using real transcriptomic data to test the KO framework in a classification context.
• KO aggregation was explored to enhance the stability of KO results.

Main Results:

• The KO framework demonstrated superior performance compared to widely used variable selection models in high-dimensional classification.
• KO aggregation improved the stability of findings without compromising statistical power.
• Application to real transcriptomic datasets revealed that the KO framework is conservative, making few discoveries.

Conclusions:

• The knockoff (KO) procedure is a robust method for variable selection in high-dimensional transcriptomic classification.
• The KO framework's conservative nature suggests it is reliable in identifying truly relevant features, unlike methods that may overestimate feature importance.
• KO aggregation offers a strategy to improve the stability of variable selection results.