Improving polygenic prediction from summary data by learning patterns of effect sharing across multiple phenotypes.

Area of Science:

• Human genetics
• Statistical genetics
• Bioinformatics

Background:

• Polygenic prediction of complex traits is crucial for precision medicine.
• The mr.mash method improves prediction by modeling multiple phenotypes jointly but requires individual-level data.
• Individual-level genetic data is often unavailable or not publicly shared.

Purpose of the Study:

• Introduce mr.mash-rss, an extension of mr.mash that uses summary statistics.
• Enhance the applicability and scalability of mr.mash for large biobank datasets.
• Improve polygenic prediction accuracy using publicly available GWAS summary statistics.

Main Methods:

• Developed mr.mash-rss, a method leveraging summary statistics and linkage disequilibrium (LD) estimates.
• Evaluated mr.mash-rss performance through extensive simulations across various genetic architectures.
• Applied mr.mash-rss to a real-world dataset of 16 blood cell phenotypes from the UK Biobank.

Main Results:

• mr.mash-rss demonstrates competitive or superior performance compared to state-of-the-art methods in polygenic prediction.
• The method shows high accuracy across diverse scenarios, including varying effect sharing, phenotype numbers, and heritability.
• mr.mash-rss achieved higher prediction accuracy for most blood cell traits in the UK Biobank analysis, particularly with smaller sample sizes.

Conclusions:

• mr.mash-rss effectively extends the utility of mr.mash to summary statistics, broadening its application.
• The method offers a scalable and powerful tool for multi-phenotype polygenic prediction using readily available GWAS data.
• mr.mash-rss represents a significant advancement for precision medicine by enabling more accessible and accurate genetic predictions.