LARGE-SCALE MULTIVARIATE SPARSE REGRESSION WITH APPLICATIONS TO UK BIOBANK.

Area of Science:

• Genetics and Bioinformatics
• Statistical Modeling
• Computational Biology

Background:

• High-dimensional regression often involves identifying a small subset of relevant features.
• Reduced rank regression is effective for multitask learning with correlated outcomes.
• The UK Biobank presents challenges with ultrahigh-dimensional features and numerous phenotypes.

Purpose of the Study:

• To develop a scalable sparse reduced-rank regression model for large-scale, ultrahigh-dimensional data.
• To address practical issues like confounding variables and missing phenotype data.
• To provide a computational tool for analyzing complex human genetic studies.

Main Methods:

• An alternating optimization scheme combining sparse regression and reduced rank decomposition.
• A scalable iterative algorithm with adaptive screening for the sparse regression component.
• Reconstruction and validation of the full solution using an optimality condition.

Main Results:

• Demonstrated effectiveness of the proposed method on synthetic and UK Biobank data.
• Successful handling of ultrahigh-dimensional features and multiple correlated outcomes.
• Development of the multiSnpnet package for polygenic risk score generation.

Conclusions:

• The proposed sparse reduced-rank regression method is effective and scalable for large datasets.
• The multiSnpnet package offers a valuable tool for genetic research and polygenic risk score estimation.
• The approach successfully integrates confounding variable control and missing data imputation.