Calibrated prediction intervals for polygenic scores across diverse contexts

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• Polygenic scores (PGS) are widely used for genomic prediction.
• Existing PGS methods exhibit performance variability across different populations and contexts.
• Factors like age, sex, income, and genetic ancestry significantly influence PGS accuracy.

Purpose of the Study:

• To introduce CalPred, a novel approach for joint modeling of contextual factors in genomic prediction.
• To develop context-specific prediction intervals for improved calibration of PGS.
• To address the miscalibration of current PGS methods in diverse populations.

Main Methods:

• Developed CalPred, an approach that jointly models multiple contextual variables.
• Applied CalPred to 72 traits using data from the All of Us and UK Biobank.
• Evaluated prediction interval calibration across various demographic and socioeconomic contexts.

Main Results:

• PGS performance varies substantially across different biobanks and contexts.
• Contextual factors like age, sex, and income impact PGS accuracy similarly to genetic ancestry.
• CalPred produces calibrated prediction intervals that vary by context, unlike existing miscalibrated methods.
• Quantitative trait prediction intervals required up to 80% adjustment.
• Disease trait predictions showed miscalibration across socioeconomic contexts, such as income levels.

Conclusions:

• Accounting for context is crucial for accurate and calibrated polygenic score predictions.
• CalPred offers a more reliable approach to genomic prediction in diverse populations.
• The findings underscore the need to incorporate socioeconomic and demographic data into PGS models to ensure equity and precision.