STELLAR: A flexible ensemble learning framework integrating rare variants to enhance polygenic risk prediction

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Whole-exome and whole-genome sequencing identify rare genetic variants linked to diseases.
• Current methods struggle to integrate rare and common variants for genetic risk prediction models.

Purpose of the Study:

• To introduce STELLAR, an ensemble learning method for rare variant polygenic risk scores (PRS).
• To enhance common variant PRS by incorporating rare variants using association summary statistics.

Main Methods:

• STELLAR combines burden-based and penalty-based rare variant analyses.
• It uses functional annotation to prioritize variants in prediction models.
• The approach leverages association summary statistics for PRS computation.

Main Results:

• STELLAR demonstrated superior prediction accuracy in simulations compared to common variant-only or rare variant burden models.
• In UK Biobank data, STELLAR significantly improved prediction for eight continuous and five binary traits.
• The method refined genetic risk stratification and identified biologically relevant genes.

Conclusions:

• STELLAR offers a scalable strategy to integrate rare variants into PRS alongside common variants.
• This advances precision risk prediction for complex diseases.
• It enables a more comprehensive assessment of genetic contributions to health and disease.