SuperAging is not the inverse of common-variant Alzheimer’s risk: evidence across genetic ancestries

Abstract

Background

As longevity increases and the population over age 65 expands, advancing age remains the most reliable predictor of cognitive decline, highlighting the need to identify biological mechanisms that support exceptional cognitive aging. We tested whether lower inherited risk of Alzheimer’s disease (AD) dementia predicts SuperAger status (adults ≥ 80 years with episodic memory at least as good as middle-age adults) using prospectively enrolled SuperAgers and Cognitively Average Controls (Controls) from the multisite SuperAging Research Initiative.

Methods

We studied 231 participants (SuperAgers n = 142; Controls n = 89). We confirmed that the genetic ancestry structure across groups was comparable. We evaluated whether APOE status (ε2, ε3, ε4) and three AD polygenic risk scores (PRS) derived from large contemporary Genome-Wide Association Studies (GWAS) (PRS_Lambert, PRS_Wightman, PRS_Bellenguez) predicted SuperAging status using logistic regression models adjusted for age, sex, and education, considering ancestry interactions.

Results

APOE allele and genotype distributions did not differ between groups, and neither APOE nor any of the three PRS predicted SuperAger status. Results were unchanged when accounting for global non-European or African ancestry or principal components. In this well-characterized cohort, neither APOE nor contemporary PRS explained SuperAger status.

Conclusions

These findings suggest that the exceptional late-life memory phenotype that is characteristic of SuperAging is not explained by common-variant AD genetic risk captured by APOE or contemporary AD PRS, motivating a deeper investigation of potential rare genetic variations and experiential factors contributing to exceptional cognitive aging.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13195-026-02124-2.

Background

Episodic memory typically declines with advancing age, paralleling age-related changes in hippocampal and distributed cortical systems that support memory function. Yet a small subset of older adults, hereafter referred to as SuperAgers, maintain episodic memory performance into their eighth decade and beyond at levels at least comparable to middle-age adults. Studying this exceptional phenotype offers a powerful opportunity to interrogate mechanisms that support memory function under conditions of highest age-related risk. By focusing on the extreme right tail of late-life cognition, SuperAging enhances spectrum-wide models of cognitive aging, resilience, and resistance by sharpening inference about biological and experiential factors that sustain memory despite advancing age and accumulating pathology. Critically, the SuperAging phenotype is defined using prospectively specified, age and performance-based cognitive criteria, enabling systematic identification and enrollment of individuals with exceptional late-life memory. The multisite SuperAging Research Initiative was established to harmonize deep phenotyping across SuperAgers and cognitively average older adults, providing a uniquely rigorous framework for isolating determinants of exceptional cognitive aging and distinguishing them from factors associated with typical aging and dementia. As such, SuperAging provides a tractable model for studying resilience and resistance mechanisms in late life.

Genetic factors represent a natural launch point for interrogating factors contributing to exceptional memory aging phenotypes. The APOE locus remains the strongest common-variant determinant of Alzheimer’s disease (AD) risk, with ε4 increasing risk and lowering age at onset and ε2 generally conferring protection. In contrast, contemporary polygenic risk scores (PRS) summarize thousands of additional variants into aggregate indices of inherited liability. However, how these liabilities operate in adults over age 80 remains uncertain. Our prior work using an early, prospectively enrolled SuperAging cohort, predominantly of European ancestry, found no difference in AD polygenic hazard between SuperAgers and age-similar controls, suggesting exceptional late-life memory is not simply explained by reduced common-variant AD risk and may instead reflect resilience or protective mechanisms [1, 2]. In contrast, large multi-cohort analyses report lower ε4 and higher ε2 frequencies among non-Hispanic White individuals with superior cognition, with evidence for similar patterns across ancestral groups [3]. Other cohorts demonstrate that superior memory confers reduced progression to mild cognitive impairment (MCI) or dementia despite comparable APOE ε4 or amyloid burden [4, 5]. Together, these findings raise unresolved questions about whether inherited AD risk distinguishes exceptional cognitive aging in advanced age, or whether genetic effects are modified by age, ancestry, or phenotype definition, questions that require rigorously defined, deeply phenotyped cohorts of the oldest-old.

Here, we leverage the multisite SuperAging Research Initiative to test the role of inherited AD genetic liability in participants with exceptional late-life memory using a prospectively defined, performance-based cognitive phenotype. By studying a large, well-characterized cohort of SuperAgers and similar-age cognitively average controls, and incorporating ancestry-aware analytic approaches, this study provides a rigorous evaluation of whether common-variant AD genetic risk, captured by APOE and contemporary polygenic risk scores, distinguishes individuals with exceptional memory performance in advanced age. This design enables assessment of genetic effects across ancestries and clarifies whether exceptional cognitive aging reflects reduced inherited AD risk or mechanisms of resilience beyond common-variant liability.

Methods

Participants

Participants included prospectively enrolled SuperAgers and cognitively average 80 + -year-old Controls from the original Chicagoland cohort and through the multisite SuperAging Research Initiative through 2024. Established in 2021, the SuperAging Research Initiative is a National Institute on Aging and McKnight Brain Research Foundation-funded consortium enrolling participants across five regions in North America (Ann Arbor/Detroit, MI; Atlanta, GA; Chicago, IL; London, ON, Canada; and Madison, WI). The Initiative builds upon and extends the original Chicagoland SuperAging cohort established and operationalized by Rogalski and colleagues by expanding geographic and demographic representation and enabling harmonized, multidomain phenotyping across sites [6–8].

Eligibility criteria at enrollment for the SuperAging Research Initiative included age ≥ 80 years, English-speaking adults with the absence of neurological disorders known to affect cognition (e.g., Parkinson’s disease, multiple sclerosis), and no significant uncontrolled medical conditions at enrollment. Participants were classified as SuperAgers or cognitively average controls based on prospectively specified, a priori cognitive performance criteria consistent with our prior SuperAging studies [6, 7].

Briefly, SuperAgers were required to demonstrate exceptional episodic memory performance, defined as scoring at or above the mean for 56–64-year-old adults on the delayed free recall trial of the Rey Auditory Verbal Learning Test (RAVLT); this corresponds to a standard score ≥ 10 based on Mayo’s Older Americans Normative Studies (MOANS) norms and is equivalent to recalling ≥ 9 words at delay [9, 10]. In addition, SuperAgers were required to perform within or above age-appropriate normative expectations on non-memory cognitive measures, including naming (Boston Naming Test, BNT, or Multilingual Naming Test, MiNT), Category Fluency Test (Animals), and executive functioning (Trail Making Test Part B), using age-, sex-, education-, and race-adjusted normative values as appropriate [11–15].

Controls were required to perform within the average normative range for their age on the RAVLT delayed free recall and memory measure and within or above normative expectations for the same non-episodic memory cognitive measures used to classify SuperAgers. All SuperAgers and Controls had a Clinical Dementia Rating (CDR) equal to zero, confirming the absence of clinical impairment in both groups.

All study procedures were approved by the appropriate Institutional Review Boards, including Advarra, Clinical Trials Ontario, Northwestern University, and the University of Chicago. Written informed consent was obtained from all participants in accordance with the Declaration of Helsinki.

This expanded, multisite cohort provides a rigorous framework for evaluating common genetic influences on exceptional late-life memory using larger, more diverse samples and updated polygenic risk scoring approaches.

Genotyping

DNA from SuperAgers and Controls participants was extracted from whole blood using the QIAamp DNA Blood Midi Kits (Qiagen, USA) and analyzed via whole exome sequencing (WES). Samples were collected and sequenced between 2013 and 2025 using multiple library preparation kits (TruSeq 62 Mb Exome Enrichment Kit or the Nextera 62 Mb Rapid Capture Expanded Exome Kit (Illumina, Inc., San Diego, CA, USA)), following the manufacturers’ protocols. Sequencing was performed on a HiSeq 2500 System using 100-bp paired-end reads (Illumina, Inc., San Diego, CA, USA). FASTQ files from SuperAgers and Controls were aligned to the human reference genome GRCh38 using the Parabricks pipeline v4.0.1. Joint genotyping was performed using GLnexus v1.4.1. Variants were then filtered to exclude those with: quality score (QUAL) < 300, depth (DP) < 10, genotype quality (GQ) < 20, single nucleotide polymorphisms (SNPs) located within 10 bp of insertion/deletions (INDELS), and allelic balance for heterozygous genotypes outside 0.2–0.8. Only biallelic SNPs were retained, and INDELS were excluded. The resulting filtered variant call format (VCF) files were converted to PLINK format [16] for further SNP and sample filtering. Specifically, we removed SNPs with a genotyping rate < 95%, those failing Hardy–Weinberg Equilibrium (HWE) with P < 1.0 × 10^–06, and with minor allele frequency (MAF) < 1%. Samples with genotyping call rates below 20% were excluded. A lower threshold than the standard 95–99% was applied to minimize sample loss in this limited dataset, as subsequent imputation was expected to recover missing genotypes with high accuracy. Duplicates were identified using the kinship-based inference for Genome-Wide Association Studies (KING) robust estimator [17], with duplicate pairs defined as those with KING-kinship ≥ 0.354. Heterozygosity was estimated using the –het command from PLINK2, then removing samples below or above ± 3 standard deviations from the sample average. In total, we removed four samples as follows: low genotyping rate (n = 1), duplicates (n = 2), and sex mismatches (n = 1), resulting in a total of 142 SuperAgers and 89 Controls for imputation. Imputation was conducted using the TOPMED server [18] with the vr3 reference panel, retaining only imputed variants with r² ≥ 0.3. Another round of SNPs quality controls was conducted on the imputed data (HWE and MAF, using the same cutoffs as above), after which we obtained a total of 2,812,549 SNPs with MAF ≥ 1%. APOE genotype was determined using SNPs rs429358 and rs7412 [19].

Ancestry

Genetic ancestry analysis was estimated using ADMIXTURE [20], incorporating reference data from 3,202 individuals from the 1000 Genomes Project Phase 3 [21], representing five major population groups: American Mixed (AMR), African (AFR), European (EUR), East Asian (EAS), and South Asian (SAS). ADMIXTURE analyses were conducted with K = 5, corresponding to the five population groups, using fivefold cross-validation to evaluate model performance. Continuous ancestry proportions were derived for each participant and used as covariates and interaction terms in logistic regression models evaluating associations between APOE, AD polygenic risk scores (PRS), and SuperAger status. Specifically, AFR and non-European (non-EUR) ancestry proportions were examined as potential effect modifiers. Principal component analysis (PCA) was performed to capture major axes of genetic variation and account for population structure. Ancestry-based PCA was conducted using SNPs with MAF ≥ 5% following linkage disequilibrium (LD) pruning in PLINK2 (command –indep-pairwise 50 5 0.2). This command performs LD pruning in sliding windows of 50 SNPs, advancing the window by 5 SNPs each time, and removes one SNP from any pair with r² ≥ 0.2. The resulting principal components capture the major axes of genetic variation and were included as covariates in genetic association models to account for population structure.

Polygenic risk score

The AD polygenic risk scores (AD-PRS) were estimated using PRsice-2 [22]. Three AD-PRSs were generated based on summary statistics from widely used genome-wide association study (GWAS) meta-analyses representing successive stages of discovery: PRS_Lambert [23], PRS_Wightman [24] and PRS_Bellenguez [25]. These GWAS were selected to provide complementary coverage of the common-variant genetic architecture of AD across increasing sample sizes and analytic frameworks spanning a decade of discovery.

Lambert et al., 2013 (International Genomics of Alzheimer’s Project (IGAP) stage-I/II) was the first large consortium meta-analysis and established many of the canonical late-onset AD loci used in early PRS benchmarking; using this score enables continuity with a vast prior literature and enables historical comparability. The phase-1 GWAS from Lambert et al. includes 17,008 AD cases and 37,154 controls and a dataset of 7,055,881 SNPs. Wightman et al., 2021 expanded the sample size to > 1.1 million individuals (using clinical and proxy phenotypes), adding new loci and updated effect sizes; a PRS from this study tests whether newer signals discovered with greater power improve discrimination beyond the prior published IGAP-based scores. The meta-analysis from Wightman et al. includes 90,338 AD cases (46,613 proxy) and 1,036,225 (318,246 proxy) controls assessed across a total of 12,206,525 SNPs. Bellenguez et al., 2022 further increased power (64,498 clinically diagnosed or ‘proxy’ AD cases and 677,663 controls), covering 21,101,115 SNPs and reported ~ 75 risk loci (42 novel at the time), refining pathway inferences and variant weights; evaluating a PRS from this meta-analysis provides a state-of-the-art benchmark anchored in the current consensus locus set.

Summary statistics genomics coordinates from Lambert et al. and Wightman et al. were converted using genome liftover with the MungeSumstats R-package [26] (GRCh37 to GRCh38). SNPs names were reformatted to the structure chromosome:bp:a1:a2, where a1 and a2 represent the allelic variants in alphabetical order. Palindromic SNPs were removed from the dataset. Target datasets were filtered to retain variants with minor allele frequency ≥ 1%, SNP and sample call rates ≥ 95%, and HWE P ≥ 1 × 10⁻⁵. SNP names were harmonized to the summary statistics format using the command –set-all-var-ids '@:#:$1:$2' from PLINK2 [16].

AD-PRS were computed using PRSice-2 [22] across a range of GWAS P-value thresholds (Benjamini and Hochberg adjusted P: 5 × 10⁻⁸ to 0.05) with LD clumping (r² = 0.1; 250 kb window). SNP weights were derived from GWAS β coefficients. Optimal P-value thresholds for SNP inclusion were determined using logistic regression in an independent, neuropathologically confirmed AD cohort (TGEN2; 1,019 AD cases and 591 cognitively normal controls), modeling AD status as the outcome and standardized PRS as the predictor, adjusting for age, sex, and the first three ancestry principal components. Details of genotyping and imputation in the TGEN2 cohort have been described previously [27]. The optimized PRS thresholds identified in TGEN2 were subsequently applied to the SuperAging and Control datasets for primary analyses.

Statistical analysis

Differences in demographic characteristics and raw neuropsychological test scores between the SuperAgers and Controls were examined using t-tests, Pearson’s χ² test or Mann–Whitney U-tests as appropriate. Allele and genotype frequencies were compared between the SuperAgers and Controls using a Pearson’s χ² test, quantifying the effect sizes using Cramer’s V statistics.

Associations between APOE and SuperAger status were evaluated using logistic regression models. Binary indicator variables were created to denote the presence of at least one APOE ε2, ε3, and ε4 allele; mixed-risk genotypes ε2/ε4 (n = 7) and ε3/ε4 (n = 35) were excluded from allele-specific analyses. The comparison between SuperAgers and Controls on APOE genotype (APOE ε2, ε3, and ε4 allele) was carried out using separate logistic regression models including terms for age, sex, years of education, and either PRS or APOE. Age and years of education were mean-centered by subtracting the sample mean. Interaction models with APOE also included self-reported race (Black/White) and genetic ancestry proportions for AFR and non-EUR, estimated using the admixture method and centered by subtracting the sample mean.

Polygenic risk score (PRS) associations with SuperAger status were examined using logistic regression models adjusted for age, sex, and education. To evaluate potential effect modification, interaction models additionally included self-reported race (Black/White) and continuous genetic ancestry proportions for African (AFR) and non-European (non-EUR) ancestry, estimated using the admixture method and mean-centered.

Sensitivity analysis was conducted by removing the APOE region from the summary statistics, as APOE is the strongest genetic factor associated with AD and the large effect size of this region can dominate polygenic risk score estimations and obscure the contribution of other variants. The SNPs removed were defined based on the genomic interval showing significant association in the original AD GWAS (Benjamini and Hochberg adjusted P < 0.05) and ranged from chr19:43,349,456 bp to chr19:49,932,605 bp.

Genetic association analysis was performed using all overlapping SNPs included in the PRS derived from the three GWAS, excluding low frequency variants (MAF < 1%). The same logistic regressions were fitted but now with a term for each SNP instead of the PRS, and using PLINK (–glm command). As a conservative threshold of significance, we used 5.0 × 10^–08, which represents the standard genome-wide significance level derived from the total number of independent common variants in the human genome [28]. For exploratory purposes, we also considered a less stringent and more tailored threshold based on the number of independent SNPs in our dataset after including only the PRS SNPs. The number of independent variants was determined using PLINK LD pruning as for the PCA. P-values from the association tests were then adjusted with the Bonferroni method, accounting for the number of independent SNPs after pruning (n = 43) and the number of models evaluated (n = 5), yielding an adjusted significance threshold of (P < 2.3 × 10^–04). SNPs annotation was conducted using ANNOVAR [29].

Results

The study design is summarized in Fig. 1A. After SNP and sample quality controls (see Methods for details), we included a total of 142 SuperAger and 89 Control participant samples. Mean age was 83.7 (SD = 4.06) in SuperAger and 84.7 (SD = 4.38) in Control (Table 1). Global ancestry proportions showed substantial overlap between SuperAger and Control groups, indicating comparable ancestry structures across the two groups (Fig. 1B–C).

Fig. 1.

Study design, ancestry comparability, and genetic risk measures in SuperAgers (SA) and Cognitively Average Controls (CTL). A Cohort overview illustrating prospective cognitive classification of adults ≥ 80 years and construction of Alzheimer’s disease polygenic risk scores (AD-PRS). B Individual-level global ancestry proportions estimated using ADMIXTURE (AFR, African; EUR, European; others). C Principal component analysis (PC1 vs PC2) of genetic ancestry demonstrating substantial overlap between SuperAgers and Controls after quality control. D APOE allele (top) and genotype (bottom) frequencies by group; numeric counts are shown above bars. E Distribution of standardized AD-PRS in SuperAgers and Controls

Table 1.

Demographic characteristics of the analyzed sample after quality controls

	SuperAgers	Controls
	(N = 142)	(N = 89)
Age, years, mean (SD)	83.7 (4.06)	84.7 (4.38)
Women, N (%)	105 (73.9%)	44 (49.4%)
Black race, N (%)	17 (12.0%)	12 (13.5%)
Latino(a), N (%)	3 (2.1%)	0 (0.0%)
Education, years, mean (SD)	16.8 (2.23)	16.9 (2.43)
Neuropsychological measures, mean (SD)
RAVLT, delayed recall	11.0 (1.9)	5.38 (1.5)
*Multilingual Naming Test	30.1 (1.9)	30.0 (2.2)
*30-item Boston Naming Test	27.9 (2.9)	26.2 (3.1)
Category fluency (animals)	21.4 (5.1)	19.2 (4.5)
Trail Making Test (Part B)	87.7 (40.6)	111.8 (56.1)

Demographics variables are self-reported. Neuropsychological data are presented as raw score means and standard deviations (SD). Episodic memory was assessed using delayed free recall from the Rey Auditory Verbal Learning Test (RAVLT; possible range 0–15). Executive function was assessed using Trail Making Test Part B, a timed measure reported in seconds, with testing discontinued at 300 s. Semantic fluency was assessed using category fluency (Animals), reported as the number of items generated in 60 s. Object naming was assessed using the Boston Naming Test, 30-item version (BNT-30; possible range 0–30) and the Multilingual Naming Test (MiNT; possible range 0–32)

*Participants completed either the BNT or MiNT depending on testing protocol. Therefore, sample is mutually exclusive with respective sample of n = 168 (SuperAgers n = 90, Control n = 78) and n = 61 (SuperAgers n = 50, Control n = 11)

The distribution of APOE allele and genotype frequency is reported in Table S1 and Fig. 1D. Groupwise comparisons of APOE status did not differ significantly between SuperAgers and Controls. Using allele- and genotype-level χ² tests, neither overall allele frequencies nor genotype distributions showed evidence of group differences (alleles: χ² = 2.81, Cramer’s V = 0.078, P = 0.245; genotypes: χ² = 4.566, Cramer’s V = 0.141, P = 0.335; Table S2). In models adjusted for age, sex, and years of education, APOE ε2, ε3, and ε4 allele status were not associated with odds of SuperAger classification (ε2: β = 0.094, SE = 9.434, P = 0.828; ε3: β = 1.226, SE = 0.735, P = 0.095; ε4: β = − 0.252, SE = 0.380, P = 0.507; Table 3). Descriptively, the proportion with at least one ε2 allele was similar across groups (SuperAger 12.7%, Control 13.1%), as was the proportion carrying at least one ε4 allele (SuperAger 15.7%, Control 19.0%) (Table S1).

Table 3.

Main and ancestry-interaction effects of APOE genotype and Alzheimer’s disease polygenic risk scores on SuperAger classification

	Effect	Beta	SE	P
Main term
	APOE ε2	0.094	9.434	0.828
	APOE ε3	1.226	0.735	0.095
	APOE ε4	−0.252	0.380	0.507
	AD-PRSLambert	1.565	23.108	0.946
	AD-PRSWightman	4.743	13.132	0.718
	AD-PRSBellenguez	2.238	70.728	0.975
Interaction with Non-European ancestry*
	APOE ε2	0.816	2.311	0.724
	APOE ε3	8.843	8.862	0.318
	APOE ε4	−1.664	1.659	0.316
	AD-PRSLambert	−206.9	127.2	0.104
	AD-PRSWightman	−34.449	67.962	0.612
	AD-PRSBellenguez	−282.295	348.116	0.417
Interaction with African ancestry†
	APOE ε2	0.599	2.017	0.766
	APOE ε3	15.630	24.990	0.531
	APOE ε4	−1.635	1.499	0.275
	AD-PRSLambert	−172.577	133.828	0.129
	AD-PRSWightman	−21.396	60.893	0.752
	AD-PRSBellenguez	−220.748	317.132	0.486
Interaction with first principal component of ancestry ‡
	APOE ε2	2.552	7.020	0.716
	APOE ε3	52.868	70.553	0.453
	APOE ε4	−5.983	5.162	0.246
	AD-PRSLambert	−625.623	391.730	0.110
	AD-PRSWightman	−100.388	209.473	0.631
	AD-PRSBellenguez	−824.939	1081.778	0.445
Interaction with self-reported race **
	APOE ε2	0.666	1.305	0.609
	APOE ε3	15.528	791.011	0.984
	APOE ε4	−1.246	1.027	0.225
	AD-PRSLambert	−113.723	77.039	0.139
	AD-PRSWightman	−13.287	39.539	0.736
	AD-PRSBellenguez	−146.472	211.627	0.489

No main or interaction effects survived multiple-testing correction

^*In separate logistic regression analysis with terms for sex, age and years of education

^**In separate logistic regression analysis with terms for sex, age, years of education and self-reported race (Black/White)

^†In separate logistic regression analysis with terms for sex, age, years of education and African Ancestry (proportion)

^‡In separate logistic regression analysis with terms for sex, age, years of education and the first principal component for ancestry

Mean AD-PRS values derived from the three discovery GWAS (Fig S1 to S3) were broadly comparable between SuperAgers and Controls (PRS_Lambert: SA 1.30 × 10⁻³ vs CTL 1.15 × 10⁻³; PRS_Wightman: SA 1.28 × 10⁻³ vs CTL 0.91 × 10⁻³; PRS_Bellenguez: SA 2.57 × 10⁻³ vs CTL 2.47 × 10⁻³; Table 2, Fig. 1E). In adjusted logistic regression models (age, sex, and years of education), none of the PRS were associated with odds of SuperAger classification (PRS_Lambert: β = 1.565, SE = 23.108, P = 0.946; PRS_Wightman: β = 4.743, SE = 13.132, P = 0.718; PRS_Bellenguez: β = 2.238, SE = 70.728, P = 0.975; Table 3). Next, we tested if there was a differential association between genetic predictors (PRS, APOE and PRS without APOE ε4) and race/ancestry, represented as (i) a non-EUR global ancestry proportion, (ii) AFR ancestry proportion, (iii) the first genetic principal component and (iv) self-reported race (indicator of Black race 1, White 0). There was no evidence of a significant interaction (Table 3).

Table 2.

Distribution of APOE allele counts and Alzheimer’s disease Polygenic Risk Score (AD-PRS) in SuperAgers and Controls

	SuperAgers N = 142	Controls N = 89
APOE Allele Counts—n/n total; %
APOE ε2	20/284; 7.0%	16/178; 9.0%
APOE ε3	239/284; 84.2%	139/178; 78.1%
APOE ε4	25/284; 8.8%	23/178; 12.9%
AD—Polygenic Risk Score (PRS)—mean (STD), 10–3
PRSLambert	0.186 (5.97)	-0.0123 (6.37)
PRSWightman	-20.0 (10.4)	-20.7 (11.4)
PRSBellenguez	-3.26 (1.91)	-3.27 (2.16)

The three PRS scores were computed using summary statistics from three different GWAS [23, 24, 30]. The table shows the mean and standard deviation of the three AD-PRS in SuperAgers and Controls

We conducted exploratory SNP-level association analysis using the 768 shared SNP variants across the three AD-PRS (Table S4). No individual SNP was significantly associated with SuperAger status after correction for multiple testing (adjusted P < 2.3 × 10⁻⁴). A heatmap showing the top 5 nominal associations for each model is reported in Fig. 2 and all associations are summarized in Table S4. In the primary model (adjusted for age, sex, and years of education), the top association was in the intergenic region between PICALM and EED (OR = 0.463; P = 3.1 × 10^–04; adj-P = 0.068), followed by other variants within PICALM (OR = 0.560; P = 4.3 × 10^–03; adj-P = 0.925); none survived multiple-testing correction. Interaction analysis (SNP x self-reported race), similarly yielded nominal associations clustered in the MS4A gene cluster, including SNPs between MS4A4E and MS4A4A (OR = 6.066; P = 0.017; adj-P = 1.000) and for a second MS4A4E variant (OR = 5.785; P = 0.017; adj-P = 1.000), but all interaction effects were non-significant after correction (Table S4).

Fig. 2.

Exploratory Single Nucleotide Polymorphism (SNP)-level association results for variants derived from Alzheimer’s disease polygenic risk scores in SuperAging. Heatmap displays the top five nominally associated SNPs per gene region across logistic regression models with SuperAger status as the outcome (adjusting for age and sex). Color intensity reflects − log10(P value), and color hue indicates direction of effect (blue: OR < 1; red: OR > 1). Five models were evaluated, each adjusted for age, sex, and years of education: SNP: main SNP effect only; SNP × AFR: interaction with African ancestry proportion; SNP × NEUR: interaction with non-European ancestry proportion; SNP × PC1: interaction with the first ancestry principal component; SNP × Race: interaction with self-reported race (Black vs White). No SNPs remained significant after correction for multiple testing. All association results are provided in Table S4

Finally, we screened our cohort for the presence of AD rare protective variants, including the APP Icelandic variant (rs63750847; chr21:25,897,620), the PLCG2 P552R variant (rs72824905; chr16:81,908,423), and APOE Christchurch variant (rs121918393; chr19:44,908,756). We identified a single heterozygous SuperAger carrying the PLCG2 P522R variant, but no carriers of the other mutations were identified in SuperAgers or Controls. Given our SuperAgers sample size of N = 142, this corresponds to a minor allele frequency (MAF) of 0.0035 for PLCG2 P552R, consistent with the frequencies reported in gnomAD for European non-Finnish (0.0083) and African/African American (0.0012). Current analyses do not suggest enrichment of AD-protective variants across our SuperAging sample.

Discussion

In this multisite, harmonized cohort of prospectively enrolled adults aged 80 years and older, we tested whether lower inherited Alzheimer’s disease risk differentiates the SuperAger phenotype, defined by youthful episodic memory, from a cognitively average phenotype. There was no association between SuperAger status and APOE (ε2/ε3/ε4) or any of the three AD polygenic risk scores (AD-PRS_Lambert, AD-PRS_Wightman, AD-PRS_Bellenguez) in ancestry-adjusted logistic models. There was also no interaction between genetic ancestry (non-European or African ancestry proportions or by ancestry principal components) and APOE or the AD polygenic risk scores. Taken together, these findings indicate that common-variant AD risk, as captured by APOE and contemporary AD-PRS, does not account for exceptional late-life memory, but does not preclude contributions from other biological or experiential mechanisms.

The three PRS were drawn from successive, large GWAS that together encompass canonical and newly discovered loci, providing state-of-the-art coverage of common-variant liability; yet none predicted SuperAger status in adjusted models. This divergence implies that genetic architectures linked to case–control AD risk may be only weakly informative for the extreme right tail of late-life cognition, reinforcing the need to study SuperAgers with designs tailored to their phenotype rather than extrapolating from AD genetics alone. The expanded, ancestry-diverse SuperAging Research Initiative specifically enables such inquiry by re-examining common genetic influences with larger, more representative samples and updated scoring approaches.

Importantly, while APOE allele frequencies and AD polygenic risk scores did not distinguish SuperAgers from cognitively average older adults, both groups differed meaningfully from individuals with AD, in whom these common genetic risk factors are strongly enriched. This pattern indicates that lower inherited AD risk at the group level is a shared feature of successful cognitive aging into advanced age, but it is not sufficient to explain the exceptional memory performance that defines the SuperAging phenotype. Thus, common AD genetic risk appears permissive rather than determinative; that is, it may enable survival into late life without dementia, but it does not fully account for the preservation of youthful episodic memory among SuperAgers. This is not unexpected, as AD polygenic risk scores are optimized for case–control discrimination and may have reduced sensitivity for resolving variance at the upper extremes of cognitive performance among non-demented older adults. However, the present findings empirically demonstrate that approaches beyond AD-risk are required to explain exceptional late-life memory. Consistent with this interpretation, screening for rare AD-protective variants also did not suggest enrichment in the SuperAging cohort.

Together, these findings place an upper bound on the contribution of common-variant Alzheimer’s disease genetic risk to the SuperAging phenotype. Variants of the magnitude captured by APOE and contemporary AD polygenic risk scores do not fully account for exceptional late-life memory relative to cognitively average older adults. While lower inherited AD risk appears to be a shared feature of dementia-free cognitive aging into advanced age, it is not sufficient to explain the preservation of youthful episodic memory that defines SuperAging. Thus, common AD genetic risk may be permissive, enabling survival into late life without dementia, but not determinative of exceptional cognitive performance. In sum, SuperAging is not simply the inverse of AD risk as indexed by common variants.

SuperAging may represent an active preservation of memory systems through mechanisms that are orthogonal to those captured by case–control AD genetics. Importantly, SuperAging is defined phenotypically, and likely reflects convergence of multiple biological and experiential pathways rather than a single genetic determinant. These mechanisms could include enhanced neural maintenance, compensatory network dynamics, or life-course experiential factors (e.g., vascular health, physical activity, and other modifiable exposures) that buffer age-related cognitive decline even in the presence of typical genetic risk. These factors are actively being investigated through the multisite SuperAging Research Initiative.

At the individual level, we observed a small number of SuperAgers with elevated AD genetic risk (e.g., APOE ε4 homozygosity or high PRS) who nonetheless maintained youthful cognition. Although not the focus of the present analyses, such individuals may be particularly informative targets for future in-depth study for identifying resilience mechanisms, analogous to cognitively intact “escapees” observed in autosomal Alzheimer’s disease [31–33].

The study has several strengths: consistent, a priori cognitive criteria for SuperAging and Controls; centralized variant processing and imputation; explicit modeling of genetic ancestry using both global admixture and principal components; and external optimization of PRS thresholds in a neuropathologically confirmed case–control resource before application to the SuperAging cohort. Limitations include modest power for small genetic effects, limited power for ancestry-stratified analyses within subgroups, potential constraints on PRS portability across ancestries, and the inability of cross-sectional analyses to establish temporal precedence. Additionally, the present analyses did not incorporate vascular, lifestyle, or behavioral factors, which may contribute to the SuperAging phenotype and represent an important direction for future work.

By demonstrating that common-variant AD genetic risk does not explain exceptional memory in advanced age, this work refines the biological scope of SuperAging and underscores the need for discovery frameworks that move beyond disease risk toward mechanisms of cognitive preservation, resilience, and resistance across the lifespan. Future work would benefit from expansion of multi-ancestry samples, integration of rare-variant analyses and gene-environment measures, lifestyle factors assessment, and measurements of the presence of AD pathology through biomarkers and/or at autopsy. The assessment of underlying AD pathology (e.g., circulating pTau) would allow us to investigate whether preserved cognition in SuperAgers reflects AD pathological burden, resilience, or both, an important direction currently underway within the SuperAging Research Initiative. Finally, linking genetic profiles to longitudinal cognitive trajectories will allow us to identify mechanisms that support exceptional memory in very late life.

Abbreviations

AD

Alzheimer’s disease

PRS

Polygenic risk score

MCI

Mild Cognitive Impairment

RAVLT

Rey Auditory Verbal Learning Test

MOANS

Mayo’s Older Americans Normative Studies

BNT

Boston Naming Test

MiNT

Multilingual Naming Test

QUAL

Quality score

DP

Depth

GQ

Genotype quality

SNPs

Single nucleotide polymorphisms (SNPs)

INDELS

Insertion/deletions

VCF

Variant call format

HWE

Hardy-Weinberg Equilibrium

MAF

Minimum allele frequency

KING

Kinship-based inference for Genome-Wide Association Studies

AMR

American Mixed

AFR

African

EUR

European

EAS

East Asian

SAS

South Asian

PCA

Principal component analysis

LD

Linkage Disequilibrium

GWAS

Genome-wide association study

IGAP

International Genomics of Alzheimer’s Project

kb

Kilobases

Authors’ contributions

MJH, ISP, and ER contributed to the conceptualization of the study. All authors contributed to data collection and curation. ISP and AWC conducted the formal analyses. ER acquired funding for the study. ISP, AWC, MJH, and ER contributed to analysis planning and methodological design. ISP, AWC, MJH, and ER drafted the original manuscript. All authors contributed to manuscript review and editing and approved the final version.

Funding

This project is supported in part by the following NIH Research Grants U19AG073153, R01AG045571, R56AG045571, 5R01AG067781, P30AG13854, P30AG072977, and by the McKnight Brain Research Foundation (MBRF). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Data availability

The datasets used analyzed in the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

All study procedures were approved by the appropriate Institutional Review Boards, including Advarra, Clinical Trials Ontario, Northwestern University, and the University of Chicago. Written informed consent was obtained from all participants prior to study participation in accordance with the Declaration of Helsinki.

Competing interests

The authors declare no competing interests.