A statistical boosting framework for polygenic risk scores based on large-scale genotype data.

Area of Science:

• Genetics
• Computational Biology
• Biostatistics

Background:

• Polygenic risk scores (PRS) assess genetic predisposition to traits and are vital for clinical risk stratification.
• Current PRS methods often rely on summary statistics from genome-wide association studies (GWAS).
• Fitting multivariable models directly on genetic data is computationally challenging for large, high-dimensional datasets.

Purpose of the Study:

• To develop an efficient algorithm for fitting multivariable models directly on genetic data for improved PRS prediction.
• To address the computational burden and memory demands of large-scale genetic datasets.
• To enhance the predictive performance of PRS beyond traditional univariate approaches.

Main Methods:

• An adapted component-wise L2-boosting algorithm (snpboost) was developed for continuous outcomes using linear base-learners.
• The method iteratively processes smaller batches of genetic variants, focusing on those most correlated with residuals.
• The framework was extended to predict binary traits, demonstrating modularity.

Main Results:

• Snpboost demonstrated competitive prediction accuracy and computational efficiency compared to existing methods like snpnet on UK Biobank data.
• The component-wise boosting approach significantly increases computational efficiency without compromising predictive performance.
• The method successfully predicted both continuous and binary traits, showcasing its versatility.

Conclusions:

• Snpboost offers an efficient and accurate computational framework for estimating polygenic risk scores from large genetic datasets.
• The algorithm's performance is comparable to established methods, providing a valuable tool for genetic risk stratification.
• The modular design allows for adaptation to various outcome types and effect estimations, broadening its applicability in precision medicine.