Why Alzheimer’s DNA tests don’t treat all populations equally

By Hugo Francisco de SouzaReviewed by Susha Cheriyedath, M.Sc.Jun 19 2025

A major global study shows that European-derived Alzheimer’s risk scores can predict disease in many ancestries, but fall short in genetically distinct groups, spotlighting the need for equitable genomic tools.

Study: Transferability of European-derived Alzheimer’s disease polygenic risk scores across multiancestry populations. Image Credit: Azurhino / Shutterstock

In a recent study published in the journal Nature Genetics, a multi-institutional team of researchers evaluated whether polygenic risk scores (PGS) for Alzheimer’s Disease (AD), derived from predominantly European genetic data, can be applied to global populations from diverse ancestries.

The study found that European-derived genome-wide association study (GWAS) data can be used to predict biomarker concentrations and age at onset of Alzheimer's disease across various ancestry groups, including Asian, African, Hispanic, and others, with significant accuracy. However, the predictive power of PGS was observed to attenuate in many non-European groups and was particularly weak in individuals of African ancestry. Encouragingly, when the key Alzheimer’s risk gene, APOE, was included, a cross-ancestry model that integrated multi-ancestry data improved PGS risk estimation in non-European populations. These findings underscore the current clinical utility of PGS and highlight the need to develop more equitable, ancestry-inclusive genetic tools for Alzheimer’s disease.

Background

Polygenic risk scores (PGSs) are a combined metric that estimates the collective impact of many genetic variants, thereby computing an individual’s risk of developing complex diseases, especially those with high heritability. Alzheimer’s Disease (AD), a progressive neurodegenerative disorder whose heritability ranges from 60-80%, is a prime candidate for PGS application.

A growing critique of PGS’s global application is that most PGSs to date have been derived from genome-wide association studies (GWAS), which are heavily biased toward individuals of European ancestry. Critiques claim that this skew limits the accuracy of the PSG model in other populations, raising concerns about equity in genetic risk assessment. Prior small-scale tests using PSG models in Korean and Black cohorts demonstrated reduced predictive performance but suggested promising similarity in outcomes.

Unfortunately, given the lack of large-scale investigations specifically focusing on how well European-derived Alzheimer’s PGSs transfer across multi-ancestry populations, critical questions remain unanswered: Can predictive tools built for one ethnic group diagnose everyone relatively, or do we risk widening preexisting health disparities?

About the study

The present study assembled PSG scores using a large European GWAS meta-analysis, from which major datasets like the UK Biobank were specifically excluded to ensure statistical independence. Researchers used this data to generate a novel PGS termed “PGSALZ” that focuses on 83 AD-associated sentinel single-nucleotide polymorphisms (SNPs), excluding the APOE (apolipoprotein E) locus.

The novel PGSALZ model scores were applied to multiple target populations, including European, East Asian, African, Hispanic, and others, totaling hundreds of thousands of participants. These ethnically diverse GWAS datasets, admittedly limited and varied in data collection and summary generation protocols, were obtained from sources including NIAGADS, Japan’s National Bioscience Database Center (NBDC), and various US-based and international studies.

To investigate whether trans-ancestry PGS models demonstrated improved predictive accuracy, preexisting European datasets were supplemented with Japanese, Indian, African, US, and other non-European GWAS data. Notably, European-only PGSALZ model scores were compared against trans-ancestry model versions to assess performance changes in non-European populations directly. All models were adjusted for potential confounders, including APOE status, age, sex, and population structure.

Statistical analyses evaluated the PSG’s performance across populations, measuring how well scores tracked actual Alzheimer’s cases, age of onset, and biomarker (e.g., amyloid-beta) levels in cerebrospinal fluid. The predictive accuracy of the PSG model was evaluated using odds ratios (ORs), predictive values such as Nagelkerke R², and other standard metrics. Sensitivity tests, precise meta-analysis, and replication across multiple cohorts strengthened researchers’ conclusions.

Study findings

The present ‘mega-analysis’ produced several significant findings. Most notably, the European-derived PGSALZ model was significantly associated with Alzheimer’s risk in many non-European ethnic groups, including Asian, Hispanic, and North African populations, with statistically significant, though often weaker, associations compared to European subjects. Odds ratios for model-predicted risk and disease onset remained significant, and PGS correlated with CSF biomarkers across ancestries.

However, some populations, especially those in African regions, demonstrated notable reductions in model predictive performance. Sub-Saharan groups are notable examples, possibly due to their linkage disequilibrium (LD) and allele frequency patterns that differ most significantly from those of Europeans.

Crucially, the benefit of incorporating diverse data was nuanced. A more complex cross-ancestry risk score generally did not outperform the simple European-derived score when the APOE genetic region was excluded. However, the cross-ancestry model was most effective and showed a clear improvement in risk prediction for non-European populations when the APOE region was included, suggesting that the APOE locus itself holds key genetic information that varies across ancestries and is critical for improving risk prediction in diverse groups.

Furthermore, the study validated the specificity of the genetic scores. Their association with disease risk was strongest for diagnosed Alzheimer’s and weakened when the diagnosis was broadened to all-cause dementia, confirming the scores capture AD-specific genetic information.

Direct comparisons between European-only and trans-ancestry models confirmed these results, suggesting that European-derived PGS models capture a substantial portion of the shared Alzheimer’s risk architecture across ethnically diverse cohorts but lose precision in genetically distant populations.

Conclusions

The present landmark study confirms that European-derived Alzheimer’s polygenic scores have predictive value across multiple ancestries, cautioning that they underperform in genetically distant populations such as those in sub-Saharan Africa. It highlights how integrating even limited non-European genetic data into current European-derived PGS models can improve predictive accuracy, particularly by better characterizing the effects of the APOE gene region across diverse groups.

The findings underscore a path forward: expanding GWAS diversity isn’t merely beneficial, it is necessary to build fair, generalizable, and clinically useful genetic tools. As the field moves toward genetic-based prevention, early intervention, and personalized interventions for AD, equity demands that we ensure risk assessments work for everyone, not just those of European descent.