Cross‐ancestry polygenic risk scores enhance Alzheimer's disease risk prediction in multiethnic cohorts

Abstract

INTRODUCTION

Genome‐wide association studies (GWAS) have identified 80+ genetic loci associated with Alzheimer's disease (AD), enabling the development of polygenic risk scores (PRS). However, the predictive accuracy of PRS in diverse populations remains low. Here, we evaluated the predictive accuracy of single‐, multi‐, and cross‐ancestry AD‐PRS models across multi‐ancestral populations.

METHODS

We used AD GWAS summary statistics from European, African, Admixed American, and East Asian populations to construct AD‐PRS for each target population. Model performance was assessed by estimating odds ratios, R², and area under the curve.

RESULTS

The cross‐ancestry Bayesian PRS model demonstrated the highest predictive performance in non‐European populations. It was significantly associated with poorer cognitive function, lower Aβ₄₂ cerebrospinal fluid levels, and the most severe category of Aβ and tau neuropathological burden.

DISCUSSION

Inclusive genetic datasets and cross‐ancestry PRS models are needed to enhance the transportability of AD‐PRS across multi‐ancestral populations.

Highlights

• Single‐ancestry polygenic risk score (PRS) is only predictive in participants of European populations.

• Cross‐ancestry PRS improves risk predictions in non‐European participants.

• Cross‐ancestry PRS is associated with abnormal Aβ and tau pathology and cognitive decline

• Cross‐ancestry PRS is associated with the Alzheimer's disease (AD) latent variable in a multi‐ancestral cohort.

1. BACKGROUND

The abundance of genome‐wide association studies (GWAS) and the growing number of disease‐associated genetic variants have enabled the development of polygenic risk scores (PRS). PRS calculate an individual's genetic liability for a given trait or a disease by aggregating all the variants weighted by their effect size estimates derived from GWAS summary statistics. ¹ In some diseases, at‐risk individuals identified through PRS exhibit up to 20‐fold higher than the carrier frequency of rare monogenic mutations conferring comparable risk. ² These findings highlight the potential of PRS to identify at‐risk individuals who may not be captured through family history or monogenic variants, creating opportunities to modify risk factors and personalize care before clinical symptoms emerge. ³

Alzheimer's disease (AD) GWAS have greatly advanced our understanding of AD genetic architecture, with heritability estimated to be 40‐60% ⁴ , ⁵ , ⁶ , ⁷ and facilitate the development of AD‐PRS. AD‐PRS show good predictive accuracy, with the best‐performing model having an area under the curve (AUC) of 0.74 while adjusting for apolipoprotein E (APOE) ε2 and ε4 status. ⁸ The predictive accuracy further increases to 81% when applied to neuropathologically confirmed cases and controls or when using extreme PRS cutoffs. ⁹ , ¹⁰ The high predictive accuracy of AD‐PRS demonstrates its potential utility for risk assessment. However, clinical translation has been slow, partly due to inconsistent performance across studies that raises concerns about generalizability. ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵

A major driver of the lack of generalizability of AD‐PRS stems from the lack of diversity in GWAS training datasets. Most AD‐PRS models have been derived from European ancestry GWAS, given the scarcity of large studies in non‐European or admixed populations. ¹⁶ , ¹⁷ As a result, AD‐PRSs derived from European datasets show markedly reduced predictive accuracy when applied to non‐European ancestry populations, with performance declining as genetic distance from the training population increases. ¹⁶ , ¹⁷ , ¹⁸ For instance, AD‐PRS derived from African American GWAS datasets outperformed those based on European or multi‐ancestry AD‐PRS in the African cohort, with some variability across different regions. ¹³ Similarly, including variants identified specifically in African populations has been shown to improve the accuracy of PRS across ancestrally diverse and admixed populations. ¹⁵ In some cases, AD‐PRS derived from European GWAS data remains associated with AD risk in non‐European populations by careful tuning of pruning and thresholding (P + T) parameters; ¹⁸ however, such optimizations may not be practical for routine clinical use. These trends highlight the limited portability of European‐derived PRS and the potential to exacerbate existing health disparities if such models are applied indiscriminately. ¹⁶

The need to address these disparities is particularly urgent given the disproportionate burden of ADRD among older Black and Latinx individuals compared to non‐Latinx White individuals. ¹⁹ Yet, relatively few studies have systematically evaluated the performance of different AD‐PRS models across ancestrally diverse populations, largely due to the limited availability of training and validation datasets. By increasing GWAS sample sizes of historically underrepresented populations and matching the statistical power of European ancestry GWAS, the accuracy of PRS can improve. ¹⁵ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ Although current non‐European AD GWAS are notably smaller than those of European ancestry, multi‐ancestry PRS derived from smaller ancestry‐matched GWAS have been shown to outperform PRS derived from larger European GWAS. ²⁵ Alternatively, cross‐ancestry PRS methods aim to jointly model GWAS summary statistics from multiple ancestry groups to improve PRS prediction within a specified target ancestry. ²² , ²³

The present study evaluates the predictive accuracy of AD‐PRS computed using single‐, multi‐, and cross‐ancestry PRS models in multi‐ancestral populations. We expect that AD‐PRS will exhibit reduced predictive accuracy in populations not represented in the training data. Beyond disease diagnosis, we also assessed the clinical validity of PRS by examining whether the best‐performing model was associated with AD endophenotypes, including biomarkers and cognitive domains within the A/T/N framework. ²⁶ , ²⁷ This evaluation provides insight into whether PRS can capture early disease‐related processes, complementing risk stratification for AD.

2. METHODS

2.1. Datasets and quality control

2.1.1. GWAS datasets

We utilized summary statistics from the largest publicly available ancestry‐specific GWAS datasets on AD and related dementias (ADRD) to date, including participants of European (Bellenguez et al. ⁵ : 39,106 cases, 46,828 proxy cases, 401,577 controls; FinnGen Release 6: 7329 cases, 131,102 controls), African (Kunkle et al. ²⁸ : 2748 cases, 5222 controls), East Asian (Shigemizu et al.: 3962 cases, 4074 controls ²⁹ ), and Caribbean Hispanic (Columbia University Study: 1088 cases, 1152 controls) descent. We also utilized multi‐ancestry meta‐analysis (MAMA) GWAS summary statistics, combining data from the five ancestry‐specific studies listed above (Lake et al.: 54,233 AD cases, 46,828 proxy cases, 543,127 controls ³⁰ ) to construct a multi‐ancestry PRS.

2.1.2. Alzheimer's Disease Sequencing Project

Data from the Alzheimer's Disease Sequencing Project (ADSP) Release 4 whole genome sequencing dataset were accessed from the National Institute on Aging Genetics of Alzheimer's Disease Data Storage Site (NIAGADS) (https://dss.niagads.org/). The participants in the ADSP are recruited from large, existing family‐based and case/control studies that provide comprehensive phenotypic and genetic data. The ADSP has made a concerted effort to include participants from multiracial and multiethnic groups, as many earlier genetic studies have predominantly involved individuals of European descent. AD diagnosis was harmonized across the cohorts by the ADSP Phenotype Harmonization Consortium (ADSP‐PHC). ³¹ The ADSP‐PHC also harmonized AD endophenotype measurements, including cognitive function (language, memory, executive functions; total N = 11,046), cerebrospinal fluid (CSF) biomarkers (Tau, pTau₁₈₁, Aβ₄₂; total N = 815), and neuropathology (Consortium to Establish a Registry for Alzheimer's Disease [CERAD], N = 2727; Thal Phase, N = 1186; BRAAK staging, N = 2725). ³² , ³³ For exclusion, A total of 515 participants diagnosed with mild cognitive impairment (MCI) were excluded from the analysis. An additional 117 participants from a family‐based study that overlapped with case/control studies were removed. Additionally, 488 samples were removed due to missing genotype data, and 5763 participants were removed due to missing diagnosis information.

RESEARCH IN CONTEXT

1. Systematic review: Using diverse genome‐wide association studies (GWAS) datasets to construct Alzheimer's disease (AD) polygenic risk score (PRS) is a promising yet underexplored approach to improve risk prediction accuracy across different populations. Integrating diverse base GWAS datasets and evaluating various PRS models can enhance model performance, but such approaches have not yet been widely applied to multi‐ancestral cohorts to measure AD risks or abnormalities in biomarkers.

2. Interpretation: Incorporation of ancestrally diverse base GWAS datasets enhanced the association between PRS and AD risk across multiple populations. Leveraging both these diverse discovery datasets and a Bayesian framework markedly improved model performance, extending its potential clinical applicability beyond AD case‐control classification to the prediction of biomarker abnormalities.

3. Future directions: Future research should prioritize the validation of cross‐ancestry PRS models in larger and more heterogeneous populations, alongside systematic benchmarking against an expanding repertoire of PRS methodologies. Clinical implementation of AD‐PRS will require rigorous validation in large, diverse as well as community‐based populations to ensure reproducibility and generalizability, thereby enhancing its translational relevance.

A standardized quality control QC pipeline was applied to the whole genome sequencing (WGS) datasets using default settings in GenoTools. ³⁴ Variants failing to meet predefined quality thresholds were excluded, including those with a minimum depth of coverage below 10, genotype quality below 20, and minor allele frequency (MAF) below 0.01. Variants with missing rsIDs were assigned a chromosome position ID. To further assess quality, sample‐level and variant‐level QC were performed using GenoTools to undergo a series of QC checks. Sample‐level QC includes call rate (≥ 0.95), sex concordance (male: F ≥ 0.99, female: F ≥ 0.03), and relatedness between samples (cutoff = 0.0884). Variant‐level QC was performed to examine non‐random missingness by haplotypes, Hardy–Weinberg equilibrium (≤ 1 x 10⁻⁴), and genotype missingness (< 0.05). 2743 participants were removed due to sex discrepancy, 359 participants were removed due to heterozygosity pruning, and 3164 participants were removed due to the relatedness check. The African GWAS summary statistics were derived from participants in the Alzheimer's Disease Genetic Consortium (ADGC), which includes sample overlaps with those from the ADSP. To retain sample independence between the base and the target datasets and avoid PRS inflation caused by sample overlaps, ³⁵ relationship inference was performed using software, KING. ³⁶ to identify and exclude samples based on relatedness. Pairwise samples with kinship coefficients ≥ 0.354, ³⁶ 2776 participants, were removed from the ADSP target datasets. After the main GenoTools filters and excluding participants with missing data, 19,398 participants (7111 cases and 12,287 controls) remained for the downstream analysis. Additionally, WHICAP participants (N = 442) were excluded from the ADSP to remove potential overlapping participants with the Caribbean Hispanic GWAS. The Stage I Bellenguez et al. GWAS includes European Alzheimer & Dementia Biobank, European Association of Development Research and Training Institutes, GERAD, Bonn, RS1&2, Genome Research at Fundació Ace, DemGene, the CCHS study, NxC, UK Biobank, which do not overlap with the ADSP cohorts.

Genetic ancestry was determined using a principal component analysis (PCA) with the 1000 Genome Project (1KG) + Human Genome Diversity Project (HGDP) as the reference panel, followed by classification via a random forest algorithm implemented in pgsc_calc. ³⁷ Out of 19,398 participants after initial exclusion (eMethods), 8043 (41.5%) participants were classified as 1KG + HGDP‐EUR‐like (EUR; European), 1945 (10.0%) as 1KG + HGDP‐AFR‐like (AFR; African), 6901 (35.6%) as 1KG + HGDP‐AMR‐like (AMR; Admixed American), 2365 (12.2%) as 1KG + HGDP‐CSA‐like (CSA; Central/South Asian), 72 (0.37%) as 1KG + HGDP‐EAS‐like (EAS; East Asian), and 72 (0.37%) participants were assigned as 1KG + HGDP‐MID‐like (MID; Middle Eastern) ancestry. Participants of EAS and MID were removed due to small sample size, while participants of CSA were excluded due to imbalanced case‐control ratios (18 cases: 2347 controls). Global genetic admixture for each sample was estimated using ADMIXTURE version 1.3. ³⁸ Briefly, we identified a representative reference panel composed of samples distributed by the 1KG. The reference panel and target ADSP dataset were subset to include only overlapping single‐nucleotide polymorphisms (SNPs). We performed an unsupervised admixture analysis of the reference dataset, setting K = 5. The target samples were then projected onto the population structure learned from the reference panel.

2.2. Health and Brain Aging Study—Health Disparities

2.2.1. Participants

The Health and Aging Brain Study—Health Disparities (HABS‐HD) cohort comprises participants from Black, Mexican American, and non‐Hispanic White populations recruited at the University of North Texas Health Science Center, Fort Worth, Texas, USA. ³⁹ , ⁴⁰ The study collected information on demographics, AD biomarkers, neuroimaging, clinical history, and genomics, providing a comprehensive resource for AD and related dementia research. Eligibility to participate in the study was restricted to generally healthy individuals without serious mental or medical conditions, who were willing to undergo sample collections and neuroimaging procedures. We utilized participant data from the HABS‐HD release 5 baseline visit 1. We further applied the exclusion criteria to remove participants younger than 55 years old, those without genetic data, and those with plasma biomarker values identified as outliers. Outliers were detected following the published methodology using the interquartile range approach. ⁴¹ All HABS‐HD participants (and/or their legal guardians) signed written informed consent to participate in the study.

2.2.2. AD endophenotypes

HABS‐HD participants have undergone extensive phenotyping for AD endophenotypes, including plasma A/T/N biomarkers, brain morphometry, amyloid and tau positron emission tomography (PET), and neuropsychological testing, as previously described. ³⁹ , ⁴⁰ For downstream analyses of each endophenotype, we applied the following approaches:

Plasma biomarkers: Fasting blood samples were collected, and the levels of plasma biomarkers (i.e., Aβ₄₂, Aβ₄₀, pTau₂₁₇, total Tau, NfL) were quantified using commercially available kits, Quanterix, for all the participants of the HABS‐HD. Aβ₄₀, Aβ₄₂, pTau₂₁₇, Tau, and NfL values were natural log‐transformed to reduce skewness, and then standardized to z‐scores (mean = 0, SD = 1) across the study samples.

Brain morphometry: Cortical thickness was defined as the surface area–weighted average of cortical thickness in the right and left entorhinal cortex, fusiform, and inferior and middle temporal cortex. Hippocampal volumes were determined by taking the mean volume of the right and left hippocampal volumes determined on a T1‐weighted volume scan.

Neuropsychological testing: Composite scores for each cognitive domain (memory, language, and executive function) were constructed by averaging z‐score‐normalized test scores from the neuropsychological battery. The memory domain was assessed using immediate and delayed recall from the Wechsler Memory Scale (WMS‐III) Logical Memory and the Spanish English Verbal Learning Test (SEVLT). The language domain was assessed using Letter Fluency and Animal Naming tests. The executive function domain was assessed using the WMS‐III Digit Span and the Trail Making Test, Parts A and B. In addition, the Mini‐Mental State Examination (MMSE) and Clinical Dementia Rating (CDR) scale were administered to all participants as part of the neuropsychological assessment.

Amyloid PET: For amyloid PET scan variables, the global standardized uptake value ratio (SUVR) was derived by normalizing tracer uptake to the whole cerebellum. Amyloid PET positivity was defined using an SUVR threshold of 1.08.

2.2.3. Genome‐wide genotyping and QC

Participants were genotyped on the Illumina Global Screening Array (GSA), and genotype data underwent stringent QC checks using an in‐house Snakemake Pipeline. ⁴² Variants were excluded if the call rate was < 0.95 and not in Hardy–Weinberg equilibrium (p < 1 × 10⁻⁶), and samples were excluded if the call rate was < 0.95; discordant sex was reported based on X chromosome heterozygosity, excessive/insufficient heterozygosity, and cryptic relatedness. Related individuals were determined within and across cohorts by identity‐by‐descent (IBD) using KING. ³⁶ with individuals excluded based on a proportion of IBD < 0.1875, corresponding to less than halfway between second‐ and third‐degree relatives. Single‐nucleotide polymorphisms that were not directly genotyped were imputed on the TOPMed Imputation Server. ⁴³ Ancestry groups were imputed separately using all ethnicities of the TOPMed reference panel (Version R3) to allow ancestry‐specific imputation quality (R²) estimation, with Eagle used for phasing and Minimac3 used for imputation. Rather than imputing the joint dataset, ancestry‐specific imputation was performed to preserve accuracy and reduce poorly imputed variants in individual ancestry groups. ⁴⁴ Following imputation, poorly imputed (R² < 0.3) or rare (MAF < 0.01) variants were removed, and the ancestry groups were merged for joint analysis. Following this merger, variants with low call rate due to differential imputation (< 95%) were removed, and then samples with low call rates (< 95%) were removed. To note, different genetic QC software were used for the ADSP and the HABS‐HD to accommodate differences in genotyping platforms and data types; accordingly, platform‐specific and widely accepted QC thresholds were applied for each cohort. These differences reflect standard best practices for WGS versus array‐based data and are not expected to materially affect downstream analyses. The final dataset contained 2559 samples with ∼22 M high‐quality SNPs per individual.

Of the total of 2559 participants that were included in the downstream statistical analysis, 1040 were of European ancestry, 891 were of Admixed American ancestry, and 574 were of African ancestry based on genetic ancestry inference (Refer to ancestry assignment in 2.3.2).

2.3. PRS construction:

2.3.1. PRS scorefile generation

PRSice was used to implement P + T to construct four AD‐PRS, stratified by EUR ancestry and MAMA ancestry, at two p‐value thresholds. P + T generates scorefiles for SNPs that exceed a given p‐value threshold (P_T ) and its corresponding effect size estimated from an independent GWAS. Scorefiles for the single‐ (EUR) and multi‐ancestry (MAMA) PRS models were constructed using SNPs with a p‐value < 0.1 (P _T0.1) and genome‐wide significance (GWS) (P _TGWS). SNPs from APOE regions ± 500 kb (GRCh38, chr19:bp44405791‐45409393) were excluded from all the base datasets. A p‐value threshold of 0.1 in the construction of PRS has been previously shown to have the best predictive accuracy and will be tested in this study. ⁸ Linkage disequilibrium (LD) clumping using 1KG + HGDP LD reference panels, with an r² threshold of 0.1 and clumping distance of 250 kb, was used to select independent SNPs. For the single‐ancestry PRS model, the 1KG + HGDP reference panel was filtered for European‐only samples, and the whole 1KG + HGDP reference panel was used for the multi‐ancestry PRS model.

PRS‐CSx was used to construct a scorefile for the cross‐ancestry AD‐PRS model, a method that improves cross‐population polygenic prediction by jointly modeling GWAS summary statistics from multiple populations. ²³ The global shrinkage parameter (phi) was set to “auto”, which allows for automatic estimation of the effect size from the datasets. Additionally, the “meta” parameter was set to “true” to combine the SNP effect sizes across populations using an inverse‐variance‐weighted meta‐analysis. For PRS‐CSx, SNPs were restricted to the HapMap3 panel as the software was designed for HapMap3 variants, whereas PRSice obtains independent SNPs using a P + T approach. Both software used the 1KG reference panel to maintain the LD estimation consistent between the two software. The summary statistics for ancestry‐specific AD GWAS were used to construct the scorefile. SNPs from APOE regions ± 500 kb (GRCh37, chr19:bp44909053‐45912650) were excluded from the final scorefile prior to PRS calculation. Liftover was automatically applied during PRS calculation to convert scorefile build to GRCh38. The same methodology was applied to the HABS‐HD cohort with the same parameters and base GWAS datasets.

2.3.2. PRS calculation and ancestry normalization

To mitigate ancestry‐related confounding due to differences in PRS distribution and to enable more accurate PRS comparisons across ancestrally diverse populations, ⁴⁵ we used pgsc_calc ³⁷ to construct PRS and normalize for ancestry differences in the ADSP and the HABS‐HD cohort using the PCA method. ⁴⁶ PCA was performed on the 1KG reference panel to derive principal components (PCs) representing global genetic variation. Target individuals were then projected onto these PCs to determine their position in genetic ancestry space. A RandomForest classifier, trained on the reference panel's PCA loadings (default: 10 PCs), was used to predict the reference population to which each target sample was most similar. ⁴⁷

Following PC projection and ancestry assignment, PRS were normalized by ancestry. First, a regression model predicted PRS from PC loadings, and the residuals were used to center the PRS distribution at zero across ancestry groups (znorm1). A second regression on the squared residuals adjusted for variance, standardizing the PRS spread to approximately one (znorm2). All five PRS models underwent ancestry normalization, and znorm2 scores were used for the downstream analysis. Importantly, phenotypic information (e.g., case–control status) was not used in the normalization process. This approach removes PRS variation attributable to population stratification driven by allele frequency and LD shifts while retaining the GWAS‐derived SNP weights and rankings of individual risks within ancestry groups. This normalization step scales PRS to enable inter‐ancestry comparisons, while still capturing disease‐relevant genetic risks.

This study was conducted and reported in accordance with the PRS reporting guidelines; a completed checklist is provided in supplementary files.

2.4. Statistical analysis

2.4.1. AD risk

Logistic regression models were used to evaluate the associations of EUR, MAMA, and PRS‐CSx PRS models with AD, adjusting for age at baseline, sex, APOE genotypes, and 10 PCs. The APOE status was encoded as a categorical variable, categorized as APOE ε2+ carriers (ε2/ε2, ε2/ε3) and ε4+ carriers (ε2/ε4, ε3/ε4, or ε4/ε4), with the ε3/ε3 allele serving as the reference group. Model performance was evaluated via area under the receiver operating characteristics curves (AUCROC) to assess model discrimination. R2redux ⁴⁸ was used to calculate and statistically compare the differences in R² for the AD outcome between models. Calibration of AD‐PRS was assessed to evaluate the estimated probability of AD against the observed probability using PRS as a predictor.

2.4.2. Endophenotypes

In the ADSP, multiple linear regression models were fit to assess the relationship between PRS‐CSx and (i) cerebrospinal fluid biomarkers (i.e., pTau₁₈₁, total tau, and Aβ₄₂), and (ii) cognitive domain scores (i.e., language, memory, and executive function). Models were adjusted for age, sex, APOE genotype, PC1‐10, and study cohort. Models for cognitive outcomes were additionally adjusted for years of education. Ordinal logistic regression models, performed using the polr function in the MASS R package, were fitted for Thal phase, CERAD score, and Braak staging as ordinal outcomes, adjusting for age, sex, APOE genotype, PC1‐10, and study cohort. All p‐values were adjusted for multiple testing at 5% false discovery rate (FDR). Genetic ancestry stratification was performed for cognitive outcomes. CSF biomarkers and other cognitive measures had insufficient sample sizes for ancestry‐stratified analyses.

Similarly, in the HABS‐HD, multiple linear regression models were fit to assess the associations between PRS‐CSx and (i) plasma biomarkers (i.e., Aβ₄₀, Aβ₄₂, pTau₂₁₇, Tau, and NfL), (ii) neuroimaging variables (i.e., cortical thickness and hippocampal volume), (iii) composite scores for each domain (i.e., language, memory, and executive function), and (iv) global cognition tests (i.e., CDR and MMSE). Models were adjusted for age, sex, APOE genotype, and PC1‐4. Additionally, body mass index (BMI) and estimated glomerular filtration rate (eGFR) were accounted for in the plasma biomarkers. Intracranial volume was accounted for in the neuroimaging variables. Years of education were accounted for in the neuropsychological test battery variables. Both raw and FDR‐adjusted p‐values (5%) were reported.

Continuous and binary outcomes were modeled using linear and logistic regressions, respectively. PRS effect estimates are interpreted as the change in the outcome per one SD increase in PRS. For ordinal outcomes, ordinal logistic regression was performed, and predicted probabilities represent the estimated likelihood of being in each severity category based on PRS values.

3. RESULTS

3.1. Patient characteristics

Of 16,461 participants in the ADSP, 9645 were classified as cognitively normal, and 6816 were diagnosed with AD (Table 1). The cohort had a mean age of 68 ± 10 years and consisted of 60% females. EUR ancestry was the most common (49%), followed by AMR (40%) and AFR participants (11%), respectively. The majority of people with AD were APOE ε4 allele carriers (54%). A similar profile is presented in the HABS‐HD, with 61% female, with a mean age of 69 ± 8 years. The cohort primarily consists of EUR (41%), AMR (35%), and AFR participants (22%) (Table 1, Table S1, Table S2).

TABLE 1.

Characteristics of the ADSP and the HABS‐HD cohort, including age, sex, genetic ancestry, APOE genotype, and endophenotypes

Cohort	ADSP		HABS‐HD
Characteristic	Cognitively normal	AD	Cognitively normal	MCI + dementia
Sample size a	9645	6816	1855	704
Age, years	68 (11)	73 (9)	67 (7.3)	68 (8.5)
Sex
Female	6088 (63%)	3986 (58%)	1194 (64%)	371 (53%)
Male	3557 (37%)	2830 (42%)	661 (36%)	333 (47%)
Ancestry
European	3637 (30%)	4406 (62%)	845 (47%)	195 (28%)
African	1377 (11%)	470 (6.8%)	330 (18%)	244 (35%)
Admixed American	4631 (38%)	1941 (28%)	637 (35%)	254 (37%)
APOE genotype
e3/e3	6195 (64%)	2843 (42%)	1019 (65%)	317 (56%)
e2+	932 (9.7%)	282 (4.1%)	165 (11%)	65 (11%)
e4+	2518 (26%)	3691 (54%)	390 (25%)	189 (33%)
Unknown	–	–	281	133
CSF Biomarkers
Aβ42	0.32 (0.97)	−0.64 (0.87)	–	–
pTau	−0.25 (0.86)	0.62 (0.95)	–	–
Tau	−0.25 (0.90)	0.57 (0.91)	–	–
Plasma biomarkers
Aβ42/Aβ40	–	–	0.02 (0.99)	−0.02 (1.02)
pTau	–	–	−0.10 (0.90)	0.26 (1.19)
Tau	–	–	−0.02 (0.97)	0.07 (1.08)
Neuroimaging
Cortical thickness	–	–	2.75 (0.13)	2.69 (0.17)
Hippocampal volume	–	–	0.12 (0.92)	−0.32 (1.13)
PET b
Aβ SUVR	–	–	1.04 (0.15)	1.37 (0.24)
Aβ positivity	–	–	195 (19%)	18 (90%)
Cognitive domains
Memory	0.67 (0.53)	−0.66 (0.82)	0.33 (0.67)	−0.85 (0.78)
Language	0.43 (0.70)	−0.32 (0.80)	0.24 (0.75)	−0.62 (0.84)
Executive function	0.45 (0.82)	−0.41 (0.88)	−0.05 (0.48)	0.12 (0.72)
Global cognition
MMSE	–	–	28.20 (2.11)	25.37 (4.33)
CDR	–	–	0.00 (0.00)	2.07 (2.17)
Neuropathology c
Thal phase
None	46 (29%)	21 (3%)	–	–
Phase 1	33 (21%)	10 (1%)	–	–
Phase 2	15 (9.4%)	16 (2%)	–	–
Phase 3	32 (20%)	54 (8%)	–	–
Phase 4	24 (15%)	121 (17%)	–	–
Phase 5	9 (6%)	475 (68%)	–	–
CERAD score
None	377 (58%)	47 (3%)	–	–
Sparse/possible	167 (26%)	65 (4%)	–	–
Moderate/probable	71 (11%)	320 (21%)	–	–
Definite/frequent	30 (5%)	1086 (72%)	–	–
BRAAK stage
None	31 (5%)	8 (0.5%)	–	–
Stage I	117 (18%)	17 (1%)	–	–
Stage II	179 (28%)	37 (2%)	–	–
Stage III	182 (28%)	80 (5%)	–	–
Stage IV	110 (17%)	188(12%)	–	–
Stage V	22 (3%)	458 (30%)	–	–
Stage VI	3 (0.5%)	732 (48%)	–	–

Abbreviations: AD, Alzheimer's disease; ADSP, Alzheimer's Disease Sequencing Project; APOE, apolipoprotein E; CDR, Clinical Dementia Rating; CERAD, Consortium to Establish a Registry for Alzheimer's Disease; HABS‐HD, Health and Aging Brain Study—Health Disparities; MMSE, Mini‐Mental State Examination; PET, positron emission tomography; SUVR, standardized uptake value ratio.

^a Mean (SD); n (%).

^b N = 1029; Aβ PET positivity from available PET data; % by diagnosis.

^c Available sample sizes; CERAD: N = 2713, Thal Phase: N = 1175, BRAAK staging: N = 2711.

The PCA projection of the ADSP revealed a high admixture profile among AMR participants. Most of these participants clustered into two distinct patterns, separated from the EUR group, while some overlapped with the AFR group (Figure 1(B), Figure S1). Consistent with this, the admixture analysis showed that AMR participants exhibited genetic heterogeneity, with contributions from EUR, Amerindian, and AFR ancestries (Figure 1(C)). To account for ancestry‐related differences in genetic architecture, we examined the distribution of all the PRS models across ancestral groups. Following ancestry normalization, the distributions were well‐aligned and approximately normally distributed across groups, confirming that ancestral confounding was mitigated (Figure S2).

FIGURE 1.

Genetic ancestry and global admixture analysis in the ADSP cohort. (A) PCA of genetic ancestry using the 1000 Genomes (1KG) reference panel. ADSP participants (black dots) are overlaid onto 1KG reference populations (square dots). (B) PCA of ADSP participants only. (C) Global ancestry proportions estimated using admixture analysis, illustrating varying degrees of admixture across ancestry groups. Same ancestry labels were used for all analyses. ADSP, Alzheimer's Disease Sequencing Project; PCA, principal component analysis.

3.2. EUR single‐ancestry PRS

The European‐only, single‐ancestry PRS model showed robust associations with AD risk in EUR participants at both P _TGWS (odds ratio [OR] [95% confidence interval {CI}] = 1.4[1.37–1.52], p = 5.53 × 10⁻⁴²) and P _T0.1 (OR[95%CI] = 1.38[1.31–1.47], p = 1.69 × 10⁻²⁷). R² values also declined as the genetic distance for AFR and AMR participants increased from the EUR participants (R² = 0.013, 0.018, 0.026, respectively) (Table S3). Association of single‐ancestry PRS remained statistically significant in the AMR participants at both P_TGWS (OR[95%CI] = 1.17[1.10–1.24], p = 5.59 × 10⁻⁷) and P _T0.1 (OR[95%CI] = 1.14[1.06–1.21], p = 1.69 × 10⁻⁴) with similar effect sizes. It showed no association with AD risk in the AFR participants, both at P _TGWS (OR[95%CI] = 1.08[0.96–1.21], p = 0.22) and P _T0.1 (OR[95%CI] = 0.93[0.81–1.04], p = 0.22) thresholds (Figure 2; Table S4).

FIGURE 2.

(A) Odds ratio of ancestry‐normalized single‐ (EUR), multi‐ (MAMA), and cross‐ancestry (PRS‐CSx) PRS models (right x‐axis) stratified by genetic ancestry (top y‐axis). p‑Values were adjusted for multiple testing using the Benjamini–Hochberg FDR procedure, controlling the FDR at 5%. (B) Comparison of model performance across ancestry groups. (Left) R² for EUR PRS (GWS), MAMA PRS (GWS), and PRS‐CSx PRS models in the participants of EUR, AMR, and AFR ancestry groups. Error bars represent 95% confidence intervals. (Right) Differences in R² between EUR versus MAMA PRS, EUR versus PRS‐CSx PRS, and MAMA versus PRS‐CSx PRS models. Error bars represent 95% confidence intervals of the difference. AFR, African; AMR, Admixed American; EUR, European; FDR, false discovery rate; PRS, polygenic risk scores.

3.3. MAMA multi‐ancestry PRS

When leveraging the MAMA GWAS summary statistics, multi‐ancestry PRS performance improved substantially in the AFR and AMR participants while remaining comparable in the EUR participants. In the EUR participants, the multi‐ancestry PRS showed similar associations as EUR PRS with AD risk for P _TGWS (OR[95%CI] = 1.44[1.37–1.52], p = 5.53 × 10⁻⁴²) and P _T0.1 (OR[95%CI] = 1.38[1.31–1.47], p = 1.69 × 10⁻²⁷). P_T0.1 model showed a stronger association with AD risk than P _TGWS model in the AMR participants (P _TGWS (OR[95%CI] = 1.17[1.10–1.24], p = 1.27 × 10⁻⁶) and P _T0.1 (OR[95%CI] = 1.34[1.26–1.41], p = 1.71 × 10⁻²³). The effect was even larger in the AFR participants with P _T0.1 model, demonstrating 80% increase in the odds of developing AD, while P _TGWS model remained nonsignificant (P _TGWS (OR[95%CI] = 1.13[1.00–1.27], p = 0.053) and P _T0.1 (OR[95%CI] = 1.80[1.62–2.01], p = 7.30 × 10⁻²⁷) (Figure 2; Table S4).

3.4. PRS‐CSx cross‐ancestry PRS

The PRS‐CSx cross‐ancestry model showed similar trends as P_T0,1 MAMA PRS models in regards to the association with AD. It was associated with the highest AD risk in the AFR participants (OR[95%CI] = 1.71[1.52–1.93], p = 1.45 × 10⁻¹⁸), while remaining significant in the AMR and EUR participants (OR[95%CI] = 1.29[1.21–1.37], p = 1.76 × 10⁻¹⁵ and OR[95%CI] = 1.31[1.24–1.38], p = 7.86 × 10⁻²², respectively) (Figure 2). The model also achieved the highest R² values in the AFR and AMR participants (R² = 0.047, 0.010, respectively) compared to other models (Figure 2; Table S3). Differences in R² were also significant between PRS‐CSx and EUR or MAMA PRS models in the AFR participants (ΔR² p = 4.6 × 10⁻⁷ and 1.3 × 10⁻⁶, respectively), showcasing PRS‐CSx strong predictive performance. A less prominent yet similar trend was seen in the AMR participants (ΔR² p = 0.016, 0.023, EUR–PRS‐CSx, MAMA–PRS‐CSx) (Table S3).

Calibration analyses further illustrate that while all models calibrate best in the EUR participants, and maintain high accuracy in the AMR participants, predictive accuracy decays progressively with increasing genetic distance from Europeans, with the greatest calibration gains in AFR participants arising from ancestry‐normalized MAMA and PRS‐CSx PRS models (Figures S3, S4).

3.5. AD endophenotypes: ADSP

To measure PRS performance beyond their established utility in predicting AD risk, we further evaluated the association of the best‐predictive PRS model, PRS‐CSx, with cognitive domains (memory, language, and executive function) and CSF biomarkers (Tau, pTau₁₈₁, and Aβ₄₂). PRS‐CSx demonstrated strong associations with poorer cognitive performance across multiple domains (β[95% CI]: Memory = 0.94[0.93–0.95], p = 6.69 × 10⁻¹⁸, Language = 0.95[0.94–0.97], p = 1.10 × 10⁻¹², Executive function = 0.95[0.94–0.97], p = 1.18 × 10⁻⁹) in the total cohort (Table S5).

PRS‐CSx demonstrated significant associations with poorer cognitive functions in participants of European ancestry (Memory (β[95% CI] = ‐0.08[−0.099, −0.062], p = 3.71 × 10⁻¹⁷); Language (β[95% CI] = −0.059[−0.077, −0.042], p = 2.87 × 10⁻¹⁰); and Executive function (β[95% CI] = ‐0.056[−0.075, −0.037], p = 1.22 × 10⁻⁸). In AFR and AMR participants, PRS‐CSx was associated with memory language, respectively. Incremental R² values were marginal, ranging from 1.8 × 10⁻⁴ to 8.7 × 10⁻³ in AFR and 3.3 × 10⁻⁴ to 1.6 × 10⁻³ in AMR participants across the domains (Table S5).

Similarly, PRS‐CSx was associated with lower Aβ₄₂ CSF levels (β[95% CI] = 0.91[0.85–0.98], p = 0.015) (Figure 3), but was not associated with Tau or pTau₁₈₁. Lower CSF levels of Aβ₄₂ indicate a worse endophenotype for AD. Furthermore, higher PRS‐CSx values were associated with an increased probability of being in the most severe category of CERAD, Thal, and BRAAK classification (e.g., “Frequent/Definite” for CERAD, “Phase 5” for Thal, and “Stage VI” for Braak), suggesting a strong relationship between PRS‐CSx and advanced AD pathology (Figure 3; Table S5).

FIGURE 3.

Associations of the best‐predictive PRS model, cross‐ancestry PRS‐CSx, with (A) Cognitive function (N = 10,903; AFR: 409, AMR: 2,888, EUR: 7,606) and (B) CSF biomarkers (N = 808) in the ADSP. Darker colors indicate statistical significance after correction at a 5% FDR. Pooled analyses were performed for CSF biomarkers due to limited sample sizes precluding ancestry‐stratified analyses. (C) Predicted probabilities of neuropathological staging outcomes by PRS‐CSx. CERAD: N = 2713, Thal Phase: N = 1175, BRAAK staging: N = 2711. The sample sizes reported in Table 1 vary, as certain participants contributing to the endophenotype analyses lack an AD diagnosis. AFR, African; AMR, Admixed American; CERAD, Consortium to Establish a Registry for Alzheimer's Disease; CSF, cerebrospinal fluid; EUR, European; FDR, false discovery rate; PRS, polygenic risk scores.

3.6. AD endophenotypes: HABS‐HD

We further validated the association of the cross‐ancestry PRS in HABS‐HD using AD endophenotypes. Higher PRS‐CSx was significantly associated with lower levels of plasma levels of Aβ₄₂, higher levels of pTau₂₁₇, and higher Aβ PET global SUVR in the total cohort. The association with pTau₂₁₇ and Aβ global SUVR was significant in the participants of European ancestry, retaining significance after multiple testing correction. There were no other ancestry‐specific effects in the AFR and AMR participants (Table S6).

4. DISCUSSION

We evaluated the predictive accuracy of single‐, multi‐, and cross‐ancestry AD‐PRS models in participants of African, Admixed American, and European ancestry within the ADSP and the HABS‐HD cohort to assess PRS transportability. Single‐ancestry PRS showed the strongest performance in participants of European ancestry but declined notably in participants of Admixed American and African ancestry, especially among individuals of African ancestry, consistent with previous reports highlighting ancestry‐specific limitations in PRS‐based risk stratification. ⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ This highlights reduced accuracy and portability of a single‐ancestry PRS model across ancestries, likely driven by differences in LD structure, allele frequency, and genetic architecture. ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ In contrast, multi‐ and cross‐ancestry PRS models demonstrated improved predictive accuracy in African and Admixed American ancestry groups. When models incorporated a large number of SNPs from multi‐ancestry GWAS (e.g., MAMA P_T0.1, PRS‐CSx), both associations with AD risk and explained variance (R²) improved, indicating that including more ancestry‐matched SNPs captures better risk profiles in African and Admixed populations. The absence of this trend for single‐ancestry EUR PRS likely reflects the limited portability of European‐derived variants to non‐European populations with more variable genetic architectures. This supports previous findings that PRS constructed from smaller, ancestry‐matched GWAS can outperform those derived from larger European datasets. ²⁵

Although the non‐European GWAS datasets used in this study were substantially smaller than the European GWAS dataset, incorporating base datasets that match the ancestry of the target populations improved overall PRS performance. In contrast, one study reported similar performance across single‐, multi‐, and cross‐ancestry AD‐PRS models using the MVP data. ⁵⁷ This discrepancy may be due to differences in SNP selections. For example, their single‐ancestry PRS included all GWS SNPs, whereas our approach applied LD pruning to select independent SNPs. As a result, incorporating additional SNPs from multi‐ancestry GWAS in their framework may have contributed only marginal effects. Differences in cohort characteristics may also play a role, as AD cases and controls in MVP were defined using International Classification of Diseases (ICD) codes rather than clinical or biomarker‐confirmed diagnoses.

Beyond AD diagnosis, we evaluated the clinical validity of PRS in predicting abnormalities in AD endophenotypes for Aβ, Tau, and cognitive function. The best‐performing model, PRS‐CSx, was associated with CSF Aβ₄₂ levels and cognitive domains of memory, executive function, and language. Notably, PRS‐CSx was not associated with CSF pTau₁₈₁ or total tau in the ADSP cohort, which may stem from sampling variability, data harmonization differences, and/or small sample sizes. Higher levels of plasma pTau₂₁₇ were significant in the HABSHD cohort, in addition to associations with lowered levels of plasma Aβ₄₂. Significant associations with key plasma biomarkers of AD pathology suggest that polygenic risk prediction may capture upstream pathogenic processes, including Aβ and tau pathways. Higher PRS‐CSx was also associated with the most severe neuropathology burdens across CERAD, Thal phase, and Braak stage in the ADSP cohort. In contrast, associations with lower or intermediate stages were weak or absent, suggesting that PRS‐CSx predicts advanced pathology but may be less sensitive to earlier or moderate changes. Although we observed significant associations between AD‐PRS and endophenotypes, the effect sizes of PRS still remain modest, explaining only a small fraction of variance in cognitive, neuroimaging, and biomarker outcomes, especially when looking at incremental R². This is consistent with previous studies that, while PRS captures cumulative genetic effects, other factors such as age, sex, APOE genotype often explain more variance in AD phenotypes than PRS alone. ⁸ , ⁵⁸ However, individuals in the extreme tail of the PRS distribution exhibit substantially elevated disease risk, reportedly as high as a 30‐fold increase among those in the top decile. ² , ⁵⁷ By defining thresholds that identify this high‐risk group, PRS can still be clinically meaningful for identifying individuals who may benefit from further testing and more regular screenings.

Ancestry normalization emerged as a critical step in improving prediction accuracy for underrepresented groups by accounting for population‐specific genetic structures that differ in allele frequency and LD. Although several PRS methods have been developed to include ancestry stratification or ancestry‐specific SNP effects to address bias in cross‐population prediction, ²² , ²³ , ³⁰ , ⁵⁹ , ⁶⁰ , ⁶¹ an explicit normalization of the PRS itself based on global ancestry is still not widely implemented for AD‐PRS. Shifts in PRS distributions that arise from differences in MAF and LD complicate the assessment of risks across populations and studies. By employing a reproducible framework such as the pgsc_calc Nextflow pipeline, genetic ancestry estimation and ancestry normalization can be performed effectively and seamlessly. Integrating ancestry normalization into PRS construction enables more equitable genetic prediction tools that better reflect the genetic architecture of ancestrally diverse populations. Given the high degree of genetic heterogeneity within a single superpopulation ⁶² (e.g., African and admixed populations), ancestry normalization is a critical step to account for differential risk profiles that arise from ancestral differences rather than true genetic risk.

Several limitations should be acknowledged in our study. First, a lack of representation of people of South and East Asian ancestry limited our ability to assess the cross‐ancestry transportability of the PRS models. The dataset is also limited to participants in the United States, and how multi‐ and cross‐ancestry perform in non‐European populations from other continents is unknown. Furthermore, the currently available GWAS summary statistics from minority populations still don't reflect the full genetic diversity and substructures that exist within one superpopulation. This calls for the necessities to continue expanding and improving the diversity and sample sizes of future GWAS to more fully represent global and intra‐ancestry genetic variation. Second, we did not incorporate genotype–environment (GxE) interactions, a framework that explains the interplay of environmental, socioeconomic, and lifestyle factors that can modify genetic risk. These factors are known to contribute additively or multiplicatively to AD susceptibility across populations. ⁶³ , ⁶⁴ , ⁶⁵ Third, future iterations of such studies should prioritize larger, more representative cohorts and integrate extensive phenotypic data from biobanks (e.g., All of Us, UK biobanks) to enhance both the calibration and clinical adaptability of risk predictions. Finally, the application of other advanced data‐driven models, such as LDpred2, ⁶⁶ PolyPred+, ⁶⁹ and BridgPRS, ⁶⁷ and the incorporation of SNP heritability estimates could further refine the predictive power and cross‐population transferability of AD‐PRS. ⁶⁸ With the rapid proliferation of PRS methods, ⁶⁹ future studies may focus on benchmarking across all the methods in large, diverse cohorts as well as community‐based populations.

Despite these limitations, our study demonstrates several key strengths. We systematically evaluated single‐, multi‐, and cross‐ancestry PRS models across participants of African, Admixed American, and European ancestries, providing a comprehensive assessment of PRS transferability in diverse cohorts. By incorporating ancestry‐specific normalization and leveraging multi‐ancestry GWAS summary statistics, we improved predictive accuracy in underrepresented populations and highlighted the importance of including ancestry‐matched SNPs in risk models. Furthermore, by extending the evaluation beyond AD diagnosis to biomarkers and cognitive endophenotypes within the A/T/N framework, we provide evidence for the clinical validity of PRS in capturing biologically relevant AD pathology. The broader implications of our findings align with the objectives of the eMERGE Consortium's Genome Informed Risk Assessment (GIRA) initiative, which emphasizes the integration of PRS into clinical care while ensuring analytical validity and health equity across diverse populations. ⁷⁰ By leveraging diverse datasets and applying ancestry‐specific normalization to PRS calculations, our methodology enhances risk stratification accuracy, supporting the growing consensus that such approaches are crucial for equitable and effective genomic medicine. ¹⁶ , ²⁰ , ⁶⁰

In summary, this study highlights methodological innovations of leveraging multi‐ancestry base GWAS summary statistics, applying ancestry normalization, and extending evaluation beyond diagnosis to AD endophenotypes that can enhance both the accuracy and clinical relevance of AD‐PRS. Importantly, our results call for equitable genomic medicine through increasing genetic representation and moving beyond Eurocentric risk models to reflect the full spectrum of human genetic diversity. By demonstrating practical strategies to improve prediction in underrepresented populations, this work provides a foundation for more inclusive risk assessment tools that can be translated into meaningful clinical applications.

CONFLICT OF INTEREST STATEMENT

C.J.’s participation in this project was part of a competitive contract awarded to DataTecnica LLC by the National Institutes of Health to support open science research. J.S.Y. serves on the scientific advisory board for the Epstein Family Alzheimer's Research Collaboration and the Charleston Conference on Alzheimer's Disease and is the editor‐in‐chief of npj Dementia. Other co‐authors report no conflicts of interest. Author disclosures are available in the supporting information. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants in the ADSP provided written informed consent, and institutional review board (IRB) approvals were obtained by each contributing cohort. All aspects of the HABS‐HD study protocol are managed by the North Texas Regional IRB. All HABS‐HD partners (and or legal guardian) undergo informed consent and provide written informed authorization to engage in the research study.

Supporting information

Supporting Information: alz71529‐sup‐0001‐figuresS1‐S4.docx

Supporting Information: alz71529‐sup‐0002‐tablesS1‐S6.xlsx

Supporting Information: alz71529‐sup‐0003‐SuppMat.docx

Supporting Information: alz71529‐sup‐0004‐SuppMat.pdf

DATA AVAILABILITY STATEMENT

European summary statistics from Bellenguez et al. 2022 ⁵ were accessed through the National Human Genome Research Institute‐European Bioinformatics Institute using accession number GCST90027158. Finnish European summary statistics from FinnGen Release 6 were accessed through https://www.finngen.fi/en/access_results. African summary Statistics from Kunkle et al. 2021 ²⁸ were accessed through NIAGADS (https://www.niagads.org/) using accession number NG00100. East Asian summary statistics from Shigemizu et al. 2021 ²⁹ were accessed through the National Bioscience Database Center (NBDC) at the Japan Science and Technology Agency (JST) at https://humandbs.biosciencedbc.jp/en/ through accession number hum0237.v1.gwas.v1. Caribbean Hispanics summary statistics from the Columbia University Study of Caribbean Hispanics and Late‐Onset Alzheimer's disease were accessed via application to dbGaP accession number phs000496.v1.p1. All codes developed and used are available at https://github.com/AndrewsLabUCSF/AFRPRS. Information regarding all the software and reference datasets used in the analysis can also be found in the repository.