PRANA: A Deep Learning Method for Adapting Polygenic Risk Scores to Diverse Ethnic Groups

Area of Science:

• Genomics
• Computational Biology
• Precision Medicine

Background:

• Polygenic risk scores (PRSs) assess inherited disease susceptibility but show reduced accuracy in non-European populations due to genetic architecture differences.
• Genome-wide association studies (GWAS) predominantly feature European cohorts, leading to inequities in PRS deployment across diverse global populations.

Purpose of the Study:

• To introduce PRANA (Polygenic Risk Adaptation via Neural-network Architecture), a deep learning framework designed to adapt existing PRSs to new ancestries.
• To enhance the predictive accuracy and equitable application of PRSs in underrepresented populations.

Main Methods:

• PRANA utilizes a deep learning framework to adapt pre-trained PRSs from European cohorts.
• The adaptation process requires modestly sized cohorts from the target population, avoiding the need for large-scale GWAS in those groups.
• The framework was evaluated on seven complex traits across South Asian, East Asian, and Ashkenazi Jewish populations, including challenging small East Asian subpopulations.

Main Results:

• PRANA demonstrated improvements in predictive performance, typically increasing effect size (β) and Nagelkerke's R² by 5%-20% compared to baseline PRS models.
• In most evaluations, PRANA outperformed existing cross-ancestry multi-PRS methods.
• The framework proved effective even in subpopulations with limited training data.

Conclusions:

• PRANA offers a scalable and practical solution to enhance PRS accuracy in diverse populations.
• The framework effectively reduces disparities in genomic risk prediction, advancing equitable PRS application.
• PRANA represents a significant step towards more inclusive precision medicine through improved polygenic risk assessment.