Detecting Interactions in High-Dimensional Data Using Cross Leverage Scores.

Area of Science:

• Genetics
• Statistical genetics
• Bioinformatics

Background:

• Investigating gene interactions (e.g., single-nucleotide polymorphisms or SNPs) is crucial for understanding complex health outcomes.
• Analyzing interactions in large genetic datasets is computationally challenging due to the high dimensionality.

Purpose of the Study:

• To develop a computationally efficient variable selection method for detecting interactions in large-scale regression models.
• To introduce and evaluate cross leverage scores (CLSs) for identifying important variable interactions while maintaining interpretability.

Main Methods:

• Developed a variable selection method based on cross leverage scores (CLSs) for interaction detection.
• Implemented data batching and windowing techniques to scale computations for large datasets.
• Utilized sketching-based approximations to further enhance computational efficiency.

Main Results:

• Cross leverage scores (CLSs) were shown to be directly correlated with the importance of variables in interaction effects.
• Approximation methods using sketching were found to be effective for large-scale data analysis, preserving the interaction detection capabilities of CLSs.
• The methods demonstrated scalability for genome-wide data analysis.

Conclusions:

• The developed CLS method and its approximations offer a scalable solution for identifying gene-gene interactions in large genetic datasets.
• These methods facilitate efficient analysis of complex genetic architectures influencing health outcomes.
• The approach is validated through simulations and application to real-world genetic data (HapMap project).