MUSSEL: Enhanced Bayesian Polygenic Risk Prediction Leveraging Information across Multiple Ancestry Groups

Area of Science:

• Population Genetics
• Statistical Genomics
• Machine Learning in Bioinformatics

Background:

• Polygenic risk scores (PRS) show promise for predicting complex traits and diseases.
• A significant performance gap exists in PRS prediction accuracy across different ancestral populations.
• Existing methods often fail to adequately account for population-specific genetic architectures.

Approach:

• Introduced MUSSEL, a novel method for ancestry-specific polygenic prediction.
• MUSSEL utilizes Bayesian hierarchical modeling with a MUltivariate Spike-and-Slab (MUSS) model for effect-size distributions.
• Incorporates an Ensemble Learning (EL) step using super learner to combine information across tuning parameters and ancestries.

Key Points:

• MUSSEL demonstrated superior performance in simulation studies and analyses of 16 traits across diverse populations (5.7 million participants).
• Achieved substantial gains in prediction R-squared (e.g., 40.2% and 49.3% over PRS-CSx and CT-SLEB in African Ancestry populations).
• Optimal method performance is contingent on GWAS sample size, target ancestry, trait architecture, and LD reference panel choice.

Conclusions:

• MUSSEL offers a significant advancement in developing more equitable and accurate polygenic risk scores.
• The method effectively leverages cross-ancestry GWAS summary statistics to improve prediction in diverse groups.
• A combination of methods may be necessary for the most robust PRS generation across all populations.