An ensemble penalized regression method for multi-ancestry polygenic risk prediction.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Advanced polygenic risk scores (PRS) aim to predict complex traits and diseases.
• Current PRS methods often lack transferability due to training predominantly on European ancestry data.
• This limits their utility in diverse global populations.

Purpose of the Study:

• To introduce PROSPER (Polygenic Risk scOres based on enSemble of PEnalized Regression models), a novel method for generating multi-ancestry PRS.
• To enhance the predictive power of PRS for minority populations by leveraging diverse genomic data.
• To provide a computationally scalable solution for large-scale, multi-ancestry PRS analysis.

Main Methods:

• PROSPER integrates genome-wide association studies (GWAS) summary statistics from multiple ancestries.
• It employs a combination of LASSO and RIDGE penalized regression models.
• An ensemble approach combines PRS derived from different penalty parameters for optimal performance.

Main Results:

• PROSPER demonstrated substantial improvements in multi-ancestry polygenic prediction across various genetic architectures.
• In African ancestry populations, PROSPER increased prediction R² by an average of 70% compared to PRS-CSx for continuous traits.
• The method showed significant gains in out-of-sample predictive accuracy in real-world datasets.

Conclusions:

• PROSPER offers a significant advancement in developing accurate and transferable polygenic risk scores across diverse ancestries.
• The method effectively addresses the limitations of existing PRS in minority populations.
• PROSPER is computationally efficient and scalable for large genomic datasets and multiple populations.