Fast and scalable ensemble learning method for versatile polygenic risk prediction.

Area of Science:

• Genetics
• Biostatistics
• Computational Biology

Background:

• Polygenic risk scores (PRS) are crucial for risk stratification and personalized medicine.
• Existing PRS methods struggle with computational efficiency, accuracy, and diverse genetic architectures.

Purpose of the Study:

• To introduce Aggregated L0Learn using Summary-level data (ALL-Sum), an efficient and scalable method for computing PRS.
• To address limitations of current PRS methods in terms of speed, accuracy, and adaptability.

Main Methods:

• Developed ALL-Sum, an ensemble learning method utilizing L0L2 penalized regression on genome-wide association study (GWAS) summary statistics.
• Employed ensemble learning across tuning parameters to model various genetic architectures.
• Validated using large-scale simulations and real-world data for 11 complex traits.

Main Results:

• ALL-Sum outperformed existing methods in simulations by 10% in accuracy, 20-fold in speed, and threefold in memory efficiency.
• Real-world data analysis showed ALL-Sum achieved 25% higher PRS accuracy, 15x faster computation, and 50% less memory usage.
• Demonstrated robustness across diverse genetic architectures and stability with varying linkage disequilibrium data.

Conclusions:

• ALL-Sum provides a fast, scalable, and accurate solution for PRS computation, advancing personalized medicine.
• The method is robust across various genetic architectures and data sources.
• ALL-Sum is available as an R package for accessible use in genetic research.