Alzheimer's polygenic risk scores vary widely across models
Across 1,752 models, correlations ranged from -0.49 to 1, with more than 95% of individuals placing in both the top and bottom risk deciles.
medRxiv
In a preprint analyzing genomic data from 11,200 clinical or autopsy-confirmed Alzheimer's disease cases and 19,321 controls aged 65 and older, researchers found that polygenic risk scores (PRS) yield starkly inconsistent results depending on the modeling method. Using the standardized GenoPred pipeline, the authors evaluated 1,752 PRS models across nine algorithms, 584 configurations, and three genome-wide association studies, stratifying subjects by genetic ancestry and APOE diplotype. Model predictions showed correlations ranging from Spearman's ρ of -0.49 to 1. More than 95% of individuals were classified into both the top and bottom risk deciles depending on the model used. Despite these contradictions, top-performing models effectively stratified disease risk across African, Admixed American, East Asian, and European ancestries.
Why it matters
Polygenic risk scores are increasingly studied to identify age-related neurodegenerative risk before clinical symptoms manifest. These findings show that unoptimized models could assign contradictory risk levels to the same individual, highlighting the need for standardized, ancestry-specific tuning.
Caveats
The findings are reported in a preprint and have not yet been peer-reviewed. The study is computational and evaluated specific algorithms and datasets from the Alzheimer's Disease Sequencing Project.
The paper
Alzheimer's Polygenic Risk Scores Are Not Interchangeable: Evidence from 1,752 Models
University of Kentucky
medRxiv · 7 Sep 2026 · Preprint, not peer-reviewed
