Reflection knockoffs via householder reflection: applications in proteomics and genetic fine mapping

Area of Science:

• Statistics
• Genetics
• Proteomics

Background:

• High-dimensional datasets often contain highly correlated features, complicating accurate statistical inference.
• Existing knockoff methods struggle with feature selection when feature correlation is substantial.

Purpose of the Study:

• To introduce a novel knockoff construction method, "reflection knockoffs," designed for datasets with high feature correlation.
• To evaluate the performance of reflection knockoffs in identifying proteomic signatures of age and in genetic fine mapping.
• To propose a method for aggregating knockoff statistics to enhance filter consistency.

Main Methods:

• Constructing reflection knockoffs using Householder reflections of original features.
• Applying knockoff filters with reflection knockoffs and aggregation to proteomic and genetic datasets.
• Comparing reflection knockoffs against Model-X knockoffs and a state-of-the-art genetic fine mapping method.

Main Results:

• Reflection knockoffs significantly outperform Model-X knockoffs in feature selection accuracy, especially with highly correlated features.
• In proteomic analysis, reflection knockoffs revealed that many initially identified age associations were non-drivers (hitchhikers).
• In genetic fine mapping, reflection knockoffs with aggregation surpassed existing state-of-the-art methods.

Conclusions:

• Reflection knockoffs provide a superior approach for feature selection in correlated data.
• The method offers improved accuracy in biological applications like identifying true drivers of age-related proteomic changes and genetic associations.
• Reflection knockoffs may enable secure sharing of genetic data by mitigating privacy concerns.