Evaluation of polygenic prediction methodology within a reference-standardized framework.

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Polygenic scores (PGS) are increasingly used for predicting complex traits.
• Numerous PGS methods exist, but their comparative performance is not well-established.
• A standardized framework is needed to evaluate these methods objectively.

Purpose of the Study:

• To evaluate and compare the predictive utility of eight popular polygenic score methods.
• To assess different strategies for parameter optimization, including cross-validation and pseudovalidation.
• To identify the most effective methods for polygenic prediction within a reference-standardized framework.

Main Methods:

• Eight PGS methods (pT+clump, SBLUP, lassosum, LDpred1, LDpred2, PRScs, DBSLMM, SBayesR) were applied.
• Performance was evaluated using UK Biobank and Twins Early Development Study (TEDS) datasets.
• Parameter optimization strategies included 10-fold cross-validation, pseudovalidation, and elastic net models.

Main Results:

• LDpred2, lassosum, and PRScs showed strong performance with 10-fold cross-validation, improving prediction by 16-18% over pT+clump.
• PRScs, DBSLMM, and SBayesR performed best under pseudovalidation.
• Ensemble models combining multiple PGS parameters consistently enhanced predictive accuracy.

Conclusions:

• LDpred2, lassosum, and PRScs are recommended for polygenic prediction when using 10-fold cross-validation.
• Pseudovalidation highlights PRScs, DBSLMM, and SBayesR as strong contenders.
• Modeling multiple polygenic scores using optimized parameters offers the best predictive power.