A novel age-informative polygenic score improves predictive ability for kidney function and kidney function decline
Preprint
Abstract
Polygenic scores (PGSs) are widely used to summarize the joint genetic effects for disease-related traits. However, while age-dependent genetic effects are increasingly recognized, their integration into PGSs remains underexplored. Kidney function, assessed by estimated glomerular filtration rate (eGFR), has strong age-related genetic effects, and prediction of kidney function decline is an unmet need. We develop an age-informative PGS for quantitative traits by generating age-specific weights via main and interaction effects and compare its performance to the age-agnostic PGS in theory and real data of eGFR. We test PGSs across 282 kidney function SNPs in cross-sectional and longitudinal data from UK Biobank (n=348,275, m=1,520,382) and independent population-based individuals aged 25 to 98 years (KORA&AugUR; n=9,057, m=16,804). In theory and real data, we illustrate that ignoring age mis-specifies genetic effects. The age-informative PGS has better performance than the age-agnostic PGS in young and old individuals (KORA&AugUR: 6.3% versus 5.9% of eGFR variance in 25-to 45-year-old, 2.3% versus 1.8% in 75-to 98-year-old). The PGS based on interaction effects explains more of the eGFR-decline variability than the age-agnostic PGS. The highest versus lowest PGS quintile predict eGFR-decline of −0.88 (95%-CI=[−0.93;−0.83]) versus −0.75 ml/min/1.73m² (95%-CI=[−0.79;−0.71]) on the population-level, similar to strata by acquired risks like diabetes, obesity or albuminuria. Prediction of eGFR-decline on the individual level remains challenging by both, genetic or acquired risks. Overall, we provide a simple approach to an age-informative PGS for quantitative disease traits and illustrate its chances and challenges for predicting kidney function and kidney function decline.
What the paper shows and why it matters
(AI-generated)
A polygenic score condenses the effects of many genetic variants into one number for predicting a trait. Each variant is weighted by its effect from a genome-wide association study, averaged over all ages in the sample. Genetic effects on some traits change with age, and a 2024 approach brings gene-by-environment interactions in general into scores by using two separate ones, for main and interaction effects (Jayasinghe et al. 2024). That is a reasonable way to include interactions. What remains open is how to build a single age-specific score for a quantitative trait, and how much it gains over age-averaged weights, which treat a variant as equally strong at every age.
The paper argues that a polygenic score should weight each variant by its effect at the person’s age, combining an age-independent main effect and an age-dependent interaction effect, rather than by an age-averaged effect. In a constructed ten-variant example, an age-averaged score explained 3.8% of variance at every age, an age-specific score 7.7% at 40 and 5.2% at 80. The paper tests this on kidney function, where some variants’ effects strengthen with age, using 282 variants weighted with UK Biobank data (Wiegrebe et al. 2024) and independent pooled data from three German cohorts aged 25 to 98 years. In the UK Biobank estimates, age-averaged effects overstated genetic effects in the young and understated them in the old. The age-specific score explained more variance overall, more in the young (in the German cohorts, 6.3% versus 5.9%) and very old (2.3% versus 1.8%), and matched the two-score model of Jayasinghe et al. with one score. Its interaction part, used as a score for kidney function decline, explained more of the variation in decline than the conventional score in the German cohort (1.0% versus 0.7%, both small).
The overall gain was small, but as the score is simple to build, the paper sees no good reason to keep age-averaged weights where effects depend on age. It presents the approach as applicable to other quantitative traits and factors, and as complementary to the two-score model. For kidney function, the genetic score separated people by their average annual decline about as much as groups split by diabetes, obesity or albuminuria, yet individual decline remains hard to predict because prediction intervals were wide and overlapping. The study is a preprint, not peer reviewed.
References
- Jayasinghe et al. (2024). Mitigating type 1 error inflation and power loss in GxE PRS: Genotype–environment interaction in polygenic risk score models. Genetic Epidemiology 48, 85–100. doi:10.1002/gepi.22546
- Wiegrebe et al. (2024). Analyzing longitudinal trait trajectories using GWAS identifies genetic variants for kidney function decline. Nature Communications 15, 10061. doi:10.1038/s41467-024-54483-9