Fast and accurate Bayesian polygenic risk modeling with variational inference

Area of Science:

• Genetics
• Statistical genetics
• Bioinformatics

Background:

• Genome-wide association studies (GWAS) generate large datasets for genetic prediction.
• Polygenic risk score (PRS) methods, often using multiple linear regression, infer genetic variant effects.
• Bayesian PRS methods using GWAS summary statistics show promise but face computational challenges with Markov chain Monte Carlo (MCMC) algorithms.

Purpose of the Study:

• To introduce Variational Inference of Polygenic Risk Scores (VIPRS), a novel Bayesian PRS method.
• To address the computational inefficiency of existing MCMC-based Bayesian PRS methods.
• To evaluate VIPRS's prediction accuracy, speed, and transferability across diverse populations and genetic architectures.

Main Methods:

• Developed VIPRS, a Bayesian PRS method utilizing variational inference for posterior approximation of effect sizes.
• Compared VIPRS against state-of-the-art PRS methods using 36 simulation configurations.
• Validated VIPRS on 12 real phenotypes from the UK Biobank dataset, including large-scale genomic data.

Main Results:

• VIPRS demonstrated competitive prediction accuracy compared to existing methods.
• VIPRS was over twice as fast as popular MCMC-based Bayesian approaches.
• VIPRS showed robust performance across various genetic architectures, heritabilities, and GWAS cohorts, with improved transferability to non-European ancestries.

Conclusions:

• VIPRS offers a computationally efficient and accurate alternative for genetic prediction from GWAS summary statistics.
• The method's speed and accuracy advantages are robust and applicable to large genomic datasets.
• VIPRS shows potential for improved PRS transferability across diverse ethnic groups, enhancing its clinical utility.