Statistical construction of calibrated prediction intervals for polygenic score-based phenotype prediction

Area of Science:

• Genetics
• Biostatistics
• Clinical Prediction

Background:

• Accurate quantification of uncertainty in predicted phenotypes from polygenic scores (PGS) is vital for clinical interpretation and disease risk assessment.
• Existing methods may lack well-calibrated prediction intervals, hindering reliable decision-making in healthcare applications.

Purpose of the Study:

• To introduce PredInterval, a novel nonparametric method for constructing well-calibrated prediction intervals for PGS-based phenotype predictions.
• To enhance the clinical utility of PGS by providing reliable uncertainty quantification.

Main Methods:

• PredInterval is a nonparametric approach compatible with any PGS method, accepting individual-level data or summary statistics.
• It utilizes quantiles of phenotypic residuals derived from cross-validation to ensure well-calibrated coverage of true phenotypic values.
• The method is designed to be robust across diverse genetic architectures.

Main Results:

• PredInterval achieved well-calibrated prediction coverage across 17 diverse real-data trait applications, outperforming existing methods.
• The method offers a principled way to identify high-risk individuals, improving identification rates by 8.7-830.4% compared to current approaches.
• It demonstrated superior performance in maintaining coverage accuracy across various genetic architectures.

Conclusions:

• PredInterval is a robust and versatile tool for enhancing the clinical utility of polygenic scores.
• The method provides accurate and well-calibrated prediction intervals, crucial for reliable clinical interpretation and decision-making.
• PredInterval facilitates improved disease risk assessment and identification of at-risk individuals through enhanced prediction intervals.