ON BLOCKWISE AND REFERENCE PANEL-BASED ESTIMATORS FOR GENETIC DATA PREDICTION IN HIGH DIMENSIONS.

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Genetic prediction aims to translate genetic discoveries into medical advances.
• High-dimensional genetic data often exhibits a block-diagonal covariance structure due to linkage disequilibrium (LD).
• Current methods often estimate variant dependence within LD blocks using external reference panels due to privacy concerns.

Purpose of the Study:

• To provide a unified analysis of blockwise and reference panel-based estimators in high-dimensional genetic prediction.
• To investigate the accuracy of different estimation methods under block-diagonal covariance structures without sparsity assumptions.
• To compare the performance of methods using original training data versus external reference panels.

Main Methods:

• Developed a unified theoretical framework for analyzing blockwise and reference panel-based estimators.
• Applied novel results from random matrix theory to high-dimensional block-diagonal covariance matrices.
• Conducted extensive numerical evaluations using simulations and real-world data from the UK Biobank.

Main Results:

• Surprisingly, blockwise estimation methods can be less accurate than whole-covariance matrix methods, even with clear LD block structures.
• Estimation methods relying on external reference panels may exhibit varying performance compared to those using original training data in high dimensions.
• The performance differences highlight potential costs associated with using only summary-level data from training sets.

Conclusions:

• Rethinking the reliance on strictly blockwise estimation in genetic prediction is warranted.
• The choice between using original training data and external reference panels significantly impacts prediction accuracy.
• Further research into robust high-dimensional covariance estimation is crucial for advancing genetic prediction.