Unsupervised Ensemble Learning for Efficient Integration of Pre-trained Polygenic Risk Scores

Area of Science:

• Genetics
• Bioinformatics
• Machine Learning

Background:

• Pre-trained polygenic risk score (PRS) models are increasingly available for real-world use.
• Challenges include PRS model transferability, data heterogeneity, and lack of phenotype data in target populations.
• Existing ensemble methods often require target population data or genome-wide association studies (GWAS), limiting real-time application.

Purpose of the Study:

• To develop an unsupervised ensemble learning framework for combining pre-trained PRS models.
• To enable accurate genetic risk prediction without requiring phenotype data or summaries from the target population.
• To facilitate the integration of PRS into real-world clinical and research settings.

Main Methods:

• Introduced UNSemblePRS, an unsupervised ensemble learning framework.
• Aggregated pre-trained PRS models based on prediction concordance across a subset of candidate models.
• Evaluated performance using continuous and binary traits in the All of Us database.

Main Results:

• UNSemblePRS demonstrated scalability and robust performance across diverse populations.
• The framework successfully combined PRS models without relying on target population phenotype data.
• Achieved accurate genetic risk prediction in real-world settings.

Conclusions:

• UNSemblePRS provides an accessible solution for integrating diverse PRS models.
• The unsupervised approach overcomes limitations of existing methods, enhancing PRS utility.
• This framework supports broader application of PRS as model availability expands.