Efficient Implementation of Penalized Regression for Genetic Risk Prediction

Area of Science:

• Genetics and Bioinformatics
• Computational Biology
• Statistical Genetics

Background:

• Polygenic Risk Scores (PRS) are crucial for identifying genetic predisposition to diseases.
• The common Clumping+Thresholding (C+T) method for PRS derivation uses univariate GWAS summary statistics.
• Jointly estimating SNP effects may significantly enhance PRS predictive performance compared to C+T.

Purpose of the Study:

• To develop and evaluate an efficient penalized logistic regression (PLR) method for computing PRS using individual-level data.
• To implement penalized linear regression for quantitative traits.
• To compare the predictive performance and scalability of PLR against C+T and random forests.

Main Methods:

• Developed an efficient penalized logistic regression (PLR) implementation for joint SNP effect estimation on large individual-level datasets.
• Included automatic hyper-parameter selection within the PLR implementation.
• Applied penalized linear regression for quantitative traits and compared PLR, C+T, and random forests on simulated and real data.

Main Results:

• PLR achieved equal or superior predictive performance compared to C+T across various scenarios, demonstrating scalability to biobank-scale data.
• In simulations, PLR improved AUC from 83% to 92.5%; in celiac disease, AUC increased from 82.5% to 89%.
• PLR predicted height (correlation ~65% vs ~55%) and breast cancer risk more accurately than C+T in large UK Biobank cohorts.

Conclusions:

• Penalized regression is a feasible and relevant method for PRS computation with large individual-level datasets.
• The efficient R package bigstatsr enables practical application of PLR for enhanced genetic risk prediction.
• PLR offers significant improvements in predictive power for both disease and quantitative traits.