Background. Heritability estimates suggest that 60–80% of Alzheimer’s disease (AD) risk is genetic, with >75 risk loci identified through genome-wide association studies (GWAS). Polygenic risk scores (PRS) aggregate and weigh the effects of those common variants and have been shown to modestly improve AD risk prediction in European populations. However, their added value beyond APOE diplotyping and their applicability to non-European populations remain unclear. Furthermore, the impact of methodological choices—such as GWAS summary statistics, PRS algorithms, and hyperparameter tuning—has not been systematically assessed using the latest AD Sequencing Project (ADSP) release 5 dataset.

Methods. Using the GenoPred pipeline, which standardizes PRS to the 1000 Genomes and the Human Genome Diversity Projects using HapMap3-restricted variants, we evaluated nine PRS algorithms and 584 hyperparameter configurations across three sources of GWAS summary statistics. In the ADSP release 5 dataset, 11,200 AD cases and 19,321 controls (age ≥65) were operationalized using a combination of autopsy and clinical data. Analyses were stratified by genetic ancestry, and PRS correlations across methods were examined. PRS metrics were stratified by APOE diplotype and calibrated to ensure broad applicability.

Results. PRS generally had little predictive value of AD risk until they exceeded a certain risk threshold, which varied depending on algorithmic parameters. PRS results also varied considerably across methodological choices (Spearman’s ρ: –0.487 to 1.0). Top-performing PRS algorithms and thresholds significantly improve AD risk prediction for populations with African (OR=1.52; P= 2.0428´10^-26), Admixed American (OR=1.47; P= 4.5652´10^-21), East Asian (OR=1.84; P= 6.2125´10^-40), and European (OR=1.8; P= 7.8925´10^-187) ancestries after stratifying for APOE diplotypes.

Conclusions. With optimal parameters, PRS significantly improve AD risk stratification beyond APOE diplotyping for individuals within the ADSP release 5 from African, Admixed American, East Asian, and European ancestries. Differing results across PRS methods underscore the need for algorithmic fine-tuning to ensure accuracy and consistency.