Novel modelling approaches to elucidate the genetic architecture of resilience to Alzheimer’s disease

Abstract

Up to 30% of older adults meet pathological criteria for a diagnosis of Alzheimer’s disease at autopsy yet never show signs of cognitive impairment. Recent work has highlighted genetic drivers of this resilience, or better-than-expected cognitive performance given a level of neuropathology, that allow the aged brain to protect itself from the downstream consequences of amyloid and tau deposition. However, models of resilience have been constrained by reliance on measures of neuropathology, substantially limiting the number of participants available for analysis. We sought to determine whether new approaches using APOE allele status, age and other demographic variables as a proxy for neuropathology could still effectively quantify resilience and uncover novel genetic drivers associated with better-than-expected cognitive performance while vastly expanding sample size and statistical power.

Leveraging 20 513 participants from eight well-characterized cohort studies of ageing, we determined the effects of genetic variants on resilience metrics using mixed-effects regressions. The outcome of interest was residual cognitive resilience, quantified from residuals in three cognitive domains (memory, executive function and language) and built within two frameworks: ‘silver’ models, which obviate the requirement for neuropathological data (n = 17 241), and ‘gold’ models, which include post-mortem neuropathological assessments (n = 3272). We then performed cross-ancestry genome-wide association studies (European ancestry, n = 18 269; African ancestry, n = 2244), gene- and pathway-based tests and genetic correlation analyses. All analyses were conducted across all participants and repeated when restricted to those with unimpaired cognition at baseline.

Despite different modelling approaches, the silver and gold phenotypes were highly correlated (R = 0.77–0.88) and displayed comparable performance in quantifying better- or worse-than-expected cognition, enabling silver–gold meta-analyses. Genetic correlation analyses highlighted associations of resilience with multiple neuropsychiatric and cardiovascular traits [false discovery rate-corrected P (P_FDR) values < 5.0 × 10⁻²]. In pathway-level tests, we observed three significant associations with resilience: metabolism of amino acids and derivatives (P_FDR = 4.1 × 10⁻²), negative regulation of transforming growth factor beta (TGF-β) production (P_FDR = 1.9 × 10⁻²) and severe acute respiratory syndrome (P_FDR = 3.9 × 10⁻⁴). Finally, in single-variant analyses, we identified a locus on chromosome 17 approaching genome-wide significance among cognitively unimpaired participants (index single nucleotide polymorphism: rs757022, minor allele frequency = 0.18, β=0.08, P = 1.1 × 10⁻⁷). The top variant at this locus (rs757022) was significantly associated with expression of numerous ATP-binding cassette genes in brain.

Overall, through validating a novel modelling approach, we demonstrate the utility of silver models of resilience to increase statistical power and participant diversity.

Phillips et al. applied novel modelling approaches to quantify resilience to Alzheimer’s disease and performed a series of genetic analyses using these metrics. In doing so, they identified heritable traits, biological pathways, and genomic variants that may help preserve cognition in the presence of Alzheimer’s disease pathology.

Introduction

Alzheimer’s disease is a progressive, debilitating neurodegenerative disease characterized by the accumulation of extracellular amyloid-β plaques and intracellular neurofibrillary tangles followed by substantial brain atrophy and cognitive decline. However, post-mortem examinations of older adults with unimpaired cognition have discovered that ≤30% of such individuals meet the neuropathological criteria for Alzheimer’s disease yet never present with any cognitive impairment.¹ As such, there is a growing focus on the genetic drivers of resilience against Alzheimer’s disease, or better-than-expected cognitive performance given a level of neuropathology, that might confer protection against the downstream consequences of pathogenic plaques and tangles. Recent work from our group has highlighted numerous genetic variants associated with resilience that are distinct from Alzheimer’s disease risk loci, reinforcing the notion that investigating genetic contributions to resilience holds substantial promise for therapeutic development.^2,3

A common approach to resilience has relied on measures of neuropathology either from autopsy or from biomarkers of neuropathology (such as PET imaging measures of amyloidosis) to model better- or worse-than-expected cognitive performance for a given level of Alzheimer’s disease pathology.^2-5 These gold-standard models produce a continuous measure of resilience that demonstrates robust single nucleotide polymorphism (SNP)-based heritability and can be leveraged in genetic analyses to identify novel variants associated with resilience.^2,3 However, the reliance on measures of neuropathology when modelling resilience greatly limits the number and diversity of participants available for inclusion in large-scale genetic analyses.⁶ Recent strides in data harmonization have greatly enhanced the quality and accessibility of genetic and longitudinal cognitive phenotype data throughout the spectrum of preclinical and clinical Alzheimer’s disease.^7-9 This provides an exceptional opportunity to use these high-calibre data to construct quantitative endophenotypes of resilience.

To this end, we sought to determine whether innovative modelling approaches reliant on APOE allele status, race, age, sex and other high-order interaction terms as a proxy for neuropathology (referred to as the ‘silver’ model) could still effectively quantify resilience while greatly expanding sample size and participant diversity available for use in genome-wide association studies (GWAS). Such demographics-based variables are easier to measure and more accessible than neuropathological variables; furthermore, individuals of African ancestry are particularly under-represented in autopsy datasets,⁶ underscoring the need for novel modelling approaches. Approximation approaches such as GWAX (GWAS-by-proxy) have yielded a plethora of new insights into the genetic architecture of Alzheimer’s disease by greatly increasing statistical power.^10,11 Furthermore, APOE is a potent predictor of amyloid deposition in late life,^12-14 and thus might be sufficient to estimate resilience when incorporated into longitudinal models of Alzheimer’s disease-related cognitive decline. We harmonized cognitive data across eight cohorts of ageing, built models of resilience in silver and gold frameworks, and performed a series of genetic analyses to identify novel genetic loci implicated in resilience to Alzheimer’s disease. In doing so, we completed the largest genetic analysis of resilience to Alzheimer’s disease to date (n = 20 513) in a cross-ancestry approach. In addition, our analysis incorporates longitudinal data in a trajectory-based definition, providing key insight into genetic factors that slow the rate of cognitive decline despite presence of the APOE ɛ4 allele and Alzheimer’s disease neuropathology. Importantly, we included multiple layers of post-GWAS tests to explore the genetic architecture of resilience further, providing evidence for biological processes that might protect the brain from neurodegeneration amidst a multitude of risk factors.

Materials and methods

Participants

Data from participants were sourced from an array of cohort studies, including the Adult Changes in Thought (ACT) Study, the Alzheimer’s Disease Neuroimaging Initiative (ADNI), Biomarkers for Older Controls at Risk for Dementia (BIOCARD), the Baltimore Longitudinal Study of Aging (BLSA), the National Alzheimer’s Coordinating Center (NACC), the Religious Orders Study (ROS), the Rush Memory and Aging Project (MAP), the Minority Aging Research Study (MARS), the Washington University Memory & Aging Project (WASHU) and the Wisconsin Registry for Alzheimer’s Prevention (WRAP). Beginning its enrolment in 1994 in Seattle, ACT selected a random cohort of participants 65 years and older who were cognitively unimpaired.¹⁵ Since its inception in 2003, ADNI has encompassed >1800 individuals between 55 and 90 years of age, through four study phases, with the principal objective of validating biomarkers for Alzheimer’s disease clinical trial applications (http://adni.loni.usc.edu/). Starting in 1995, the ongoing BIOCARD study initially enrolled cognitively normal individuals, predominantly in the middle-aged group (https://reporter.nih.gov/project-details/8072622). The BLSA, initiated by the National Institute on Aging in 1958, stands as one of the world’s most enduring longitudinal ageing studies, including >3200 participants (https://www.nia.nih.gov/research/labs/blsa). Established by the NIA in 1999, NACC amalgamates participant data from Alzheimer’s Disease Centers funded by the NIA, with records dating back to 1984, forming the most extensive database of its kind in the USA.¹⁶ ROS, initiated in 1994, engaged Catholic nuns, priests and brothers across the nation, and MAP, starting in 1997, focused on cognitively intact older adults from the Chicago area.¹⁷ Beginning in 2004, MARS specifically targeted the enrolment of older African-Americans displaying no cognitive impairment.¹⁸ WASHU, active since 1979, has consistently enrolled participants, either cognitively unimpaired or with mild dementia, at their initial visit. WRAP has been evaluating participants since 2001, now boasting a cohort exceeding 1700 individuals, particularly those with a family history of probable Alzheimer’s disease.¹⁹

All participants provided informed consent, and the studies were carried out in accordance with Institutional Review Board-approved protocols. The Vanderbilt University Medical Center Institutional Review Board authorized secondary analyses of the data. Supplementary Tables 1 and 2 provide an overview of participant demographics in each analysis by cohort. Data were accessed and harmonized as part of the Alzheimer’s Disease Sequencing Project Phenotype Harmonization Consortium.

Post-mortem assessment of neuropathology

Post-mortem assessments were conducted for participants in the ACT, NACC and ROS/MAP cohorts. Data were harmonized to align with the uniform dataset neuropathology form from NACC. For our analyses, we leveraged a well-established neuropathological staging of amyloid plaque burden, the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) neuritic plaque staging scores.²⁰ Braak staging of neurofibrillary tangle distribution and severity was also included.²¹ CERAD and Braak scores were standardized between ACT, NACC and ROS/MAP, such that higher CERAD and Braak scores were representative of higher individual amyloid and tau burden in all cohorts, respectively. For more extensive details regarding post-mortem assessment of neuropathology, see our recent paper.²

Cognitive harmonization

Cognitive data were harmonized across all cohorts using published, modern psychometric techniques.²² Briefly, qualified neuropsychologists/behavioural neurologists categorized test items into memory, language or executive function domains (or neither). Test items that overlapped across cohorts were set as anchor items. Scaled anchor items allowed non-overlapping test items to be estimated freely. Subsequently, test items were entered as indicators in a confirmatory factor analysis. All three cognitive outcomes were successfully harmonized across all eight cohorts. The harmonized memory, language and executive function composite scores were extracted and leveraged in resilience phenotype modelling. For more extensive details regarding cognitive harmonization, see our recent paper.²

Building ‘silver’ and ‘gold’ resilience models

Cognitive resilience models were created using a derivation of previously published protocols (Fig. 1).^2,5 A dataset composed of eight cohorts without neuropathology measures and a subset composed of three cohorts with neuropathology measures were used to build latent variable models of resilience. Both models were restricted to participants with at least two cognitive visits. All linear mixed-effects models covaried for the fixed effects of age at baseline, sex, race, APOE ɛ2 allele positivity, APOE ɛ4 allele count, and an age-at-baseline-squared term to represent the non-linear effects of ageing. Race was categorized as non-Hispanic White, non-Hispanic Black or Other (any other or mixed race). Time was modelled as years from baseline and was included as both a fixed and a random effect. Both models included all three-way interactions of covariates with age and years from baseline, along with all lower-order terms.

Figure 1.

Resilience metric quantification. Mixed-effects regression models were built, in which a cognitive score was regressed on age at baseline, sex, race, APOE ɛ2 allele positivity, APOE ɛ4 allele count, and an age-at-baseline term to represent the non-linear effects of ageing. Time was modelled as years from baseline and was included as both a fixed and a random effect. Both models included all three-way interactions of covariates with either age at baseline or age at baseline squared and years from baseline, along with all lower order terms. The gold model also included covariates for amyloid and tau pathology. Residuals were extracted and entered as indicator variables in a partial least squares path model using established procedures. Educational attainment was included as an additional indicator variable to yield residual cognitive resilience. APOE ɛ4 = apolipoprotein E ɛ4; CERAD = Consortium to Establish a Registry for Alzheimer's Disease; EXF = harmonized executive function score; LAN = harmonized language score; MEM = harmonized memory score.

The gold model also included covariates for amyloid and tau pathology. Amyloid pathology was coded as a categorical variable (from one to four) corresponding to CERAD stages ‘none’, ‘sparse’, ‘moderate’ or ‘frequent’. Tau pathology was coded as a categorical variable corresponding to Braak stages 0–6. In the gold models that incorporated neuropathology at autopsy, time between last site visit and participant death was restricted to 5 years or fewer. Additionally in the gold model, CERAD and Braak measures of amyloid and tau pathology, respectively, were treated as factor variables. Outcomes were longitudinal harmonized scores for memory, executive function and language. Residuals were generated from models including all participants and in a subset including only participants who were cognitively unimpaired (CU) at baseline. Diagnosis across cohorts was categorized as normal cognition, mild cognitive impairment (MCI), and dementia based on clinical consensus that varied by cohort. The CU status was characterized as participants who had a diagnosis of normal cognition. Cognitively impaired status was characterized as participants who had a diagnosis of MCI or dementia.

We leveraged a residuals-based approach to model better- or worse-than-expected cognitive performance to represent higher versus lower resilience. Standardized residuals from all mixed-effects models were extracted and entered as indicators into latent variable models in Mplus (version 7.31)²³ to summarize the extent to which a participant performed better or worse than predicted given their unique demographic characteristics, APOE allele status and burden of neuropathology (gold model participants only). One resilience model was built: residual cognitive resilience, with the standardized residuals from each of the three cognitive domains and educational attainment measured in years as indicators (Fig. 1). Inclusion criteria for the models required participants to have residuals for at least two of the three cognitive domains. Models showed good fit [root mean square error of approximation (RMSEA) = 0.013–0.034, comparative fit index (CFI) = 0.996–0.999]. Factor scores were extracted from all models. In total, four resilience composite scores were generated: two corresponding to the silver models and two corresponding to the gold models, further stratified into composites built from models including all participants and separately in CU-at-baseline participants. These scores were then used as endophenotypes in all GWAS and post-GWAS analyses. For average resilience scores by cohort, see Supplementary Tables 1 and 2, and for more extensive details regarding the latent variable modelling procedure, see our recent paper.²

Genotyping, quality control and imputation

Participants included in this study were genotyped using DNA extracted from either brain or whole blood. Genotyping chips for each cohort are as follows. For ACT, genetic data were collected with two arrays (Illumina Human660W-Quad Array and Infinium Global Screening Array-24). For ADNI, genetic data were collected with four arrays (Illumina Human610-Quad, Illumina HumanOmniExpress, Illumina Omni 2.5 M, and Illumina Global Screening Array v.2). For BIOCARD, genetic data were collected with the Illumina HumanOmniExpress array. For BLSA, genetic data were collected with two arrays (Illumina HumanOmni2.5 and Illumina HumanOmniExpress). NACC is a consortium of 37+ Alzheimer’s disease research centres, and all genotyping information is outlined on the NACC website (https://naccdata.org/nacc-collaborations/partnerships). For ROSMAPMARS, genetic data were collected with three arrays (Global Screening Array-24 v.3.0, Affymetrix GeneChip 6.0 and Illumina HumanOmniExpress). For WASHU, genetic data were collected with two arrays (Illumina Human610 and Illumina HumanOmniExpress). For WRAP, genetic data were collected using the Illumina HumanOmniExpress array.

All genetic data were processed with a standardized quality control and imputation pipeline.³ Initially, variants that had a low genotype rate (<95%), low minor allele frequency (<1%) or were outside of Hardy–Weinberg equilibrium (P < 1 × 10⁻⁶) were removed. Participants were excluded if the reported and genotypic sex differed, if there was poor genotyping efficiency (missing >1% of variants) or if cryptic relatedness was present (PIHAT > 0.25). Imputation was performed on the University of Michigan Imputation Server using the TOPMed reference panel (hg38) with SHAPEIT phasing.^24-26 Following imputation, datasets were filtered to exclude variants with low imputation quality (R² < 0.8 and used called genotypes from the imputation), duplicated/multi-allelic variants and minor allele frequency of <1%. Within each self-identified racial group (non-Hispanic White and non-Hispanic Black), principal component analysis was conducted, and genetic ancestry outliers relative to 1000 Genomes reference populations [e.g. Utah residents with Northern and Western European Ancestry (CEU) and Yoruba in Ibadan, Nigeria (YRI)] were excluded.

Statistical analysis

Prior to performing GWAS, cryptic relatedness across genetic datasets was assessed, removing 1607 related individuals across all eight datasets. Related individuals present in multiple datasets were preferentially removed from the highest-powered dataset.

Resilience model assessment

Silver models were evaluated by building the model in all participants, then comparing the resilience phenotypes in all participants and the subset with available neuropathology data. Models were considered valid if the predictors in the model were statistically significant and explained a substantial amount of the variance in longitudinal cognitive performance based on Cohen’s benchmarks for correlation (R² ≥ 0.25)²⁷ and if the resilience scores performed better in neuropathology-positive participants compared with neuropathology-negative participants. We used several approaches to visualize and assess the silver resilience models. Initially, we compared the silver and gold models in overlapping participants using Pearson’s correlation. Next, we visually assessed the ability of the silver model preferentially to differentiate slower-than-expected longitudinal cognitive decline among those who were amyloid and tau positive. Finally, we excluded participants with neuropathology data from the silver model to ensure zero sample overlap with the gold model and compared its genetic correlation with the gold model to determine whether they shared a common genetic architecture. We completed additional sensitivity analyses in all participants by incorporating last visit clinical diagnosis into each mixed-effects model to compare with our base models.

GWAS and genome-wide meta-analyses

A workflow outlining GWAS, meta-analysis and post-GWAS analyses is presented in Fig. 2. GWAS were performed with PLINK linear association models (v.1.90b5.2; https://www.cog-genomics.org/plink/1.9).²⁸ All GWAS were run in cohorts individually for all resilience phenotypes and separately in European ancestry (EUR) and African ancestry (AFR) participants. GWAS covariates included age, sex, APOE ɛ4 allele count, APOE ɛ2 allele positivity and the first five genetic ancestry principal components. GWAS results were then meta-analysed across cohorts using a fixed-effects model with beta and standard error input (GWAMA v.2.2.2).²⁹ Models were also run identically in the sample restricted to individuals who were CU at baseline, with the fixed-effects meta-analyses implementing the minor allele frequencies calculated based on these individuals only. Following the individual meta-analyses of EUR and AFR outcomes, these results were integrated to produce a cross-ancestry analysis with increased statistical power. Importantly, GWAS were included in ancestry-specific meta-analyses only if ≥30 participants were present in cohort-specific GWAS.

Figure 2.

Analytical workflow. AFR = African; CU = cognitively unimpaired; EUR = European; GWAS = genome-wide association study.

Gene- and pathway-based tests

Gene- and pathway-level tests were performed with Multi-marker Analysis of GenoMic Annotation (MAGMA v.1.09)³⁰ on all meta-analysis results. Initially, permutation gene tests were performed to determine whether a higher number of significant variant-level P-values existed in a predefined gene window than expected by chance. This process was conducted across the entire genome. All gene-level results were then entered into permutation pathway tests to determine whether there were more significant gene test P-values associated with known biological pathways than expected by chance. We leveraged two sets of curated pathway annotations from the Molecular Signatures Database (MSigDB) v.7.0 (downloaded on 5 February 2020), the curated gene set (C2) and the ontology gene set (C5).³¹ We conducted MAGMA using GWAS summary statistics for each of the silver and gold phenotypes and for silver–gold meta-analyses. In total, we tested 17 922 genes and 12 173 biological pathways. All gene and pathway tests were adjusted for multiple comparisons using the false discovery rate (FDR) procedure, and an a priori significance threshold was set at P_FDR < 0.05.

Genetic correlation and heritability analyses

We conducted genetic correlation and heritability analyses using the Genetic Covariance Analyzer (GNOVA).³² Genetic covariances were computed within GNOVA by comparing resilience meta-analysis summary statistics with GWAS summary statistics for 65 complex traits, deriving z-scores for each variant-level association from the GWAS summary datasets. We also determined linkage disequilibrium scores from a reference panel aligned with the ancestry of our sample (e.g. 1000 Genomes Project’s European reference panel). Subsequently, we calculated genetic covariances between pairs of traits using the aforementioned z-scores. Adjustments for inflation caused by the linkage disequilibrium structure were made by incorporating the ancestry-matched genome-wide linkage disequilibrium scores. Additionally, we corrected genetic covariances for overlaps in sample sizes across different GWAS. Throughout all genetic correlation analyses we used most basic model in GNOVA, which does not include annotations. Upon completing these analyses, we adjusted the genetic covariances for the risk of multiple comparisons through the FDR procedure, setting an a priori significance cut-off at FDR < 0.05. We leveraged the EUR within-ancestry meta-analyses, because all prior complex traits focused on EUR ancestry.

Expression quantitative trait locus analysis

All genetic variants surpassing the suggestive significance threshold were explored in the following expression quantitative trait loci (eQTL) databases: BRAINEAC (http://www.braineac.org), eQTLGen Consortium (whole blood; https://www.eqtlgen.org), Brain xQTLServe³³ and BrainSeq (dorsolateral prefrontal cortex and hippocampus; http://eqtl.brainseq.org). The eQTL significance threshold was set a priori at P < 0.05. For each eQTL, P-values were determined by the given P-values in each database for the given tissue(s).

Results

Gold and silver resilience phenotypes displayed strong, positive correlations (R = 0.77–0.88) across comparable models in participants with neuropathological data available (Supplementary Fig. 1). Indeed, the silver model adequately identified better-than-predicted cognitive performance among those carrying an APOE-ɛ4 allele as expected, but also in the subset of participants with available neuropathological data who were amyloid and tau positive (see Fig. 3). Individuals with higher resilience scores displayed a similar rate of memory decline despite the presence of the ɛ4 allele or neuropathology at autopsy, validating our resilience metric as quantifying better-than-expected cognitive performance. Although memory decline is depicted in Fig. 3, we observed similar trends for executive function and language (see Supplementary Fig. 2A–D), further underscoring the translatability of our approach across cognitive domains. Furthermore, in our silver standard models, we were able to explain 40%–58% of variance, in comparison to 39%–57% in the gold standard models (Supplementary Table 3). Notably, when we calculate the silver standard model in participants with neuropathological data available (to compare the variance explained by gold and silver models in an identical set of participants), the gold standard models do explain 2.5%–8% more variance in cognitive performance than the silver model, as expected (Supplementary Table 3). Across all silver and gold models, we observed strong, expected, positive correlations between resilience scores and annual rates of decline in each cognitive domain (Supplementary Fig. 3). As expected, the correlation with cognition decreases as the variance explained in the initial resilience model increases. In sensitivity analyses incorporating last visit diagnosis as a predictor in each mixed-effects regression for all participants, we observed strong correlations with the base models (silver, R = 0.84; gold, R = 0.88), and the number of significant interactions with ɛ4, amyloid and tau positivity was comparable across models with and without clinical diagnosis (Supplementary Figs 2A and B and 4A and B).

Figure 3.

Silver and gold phenotype comparison. Scatterplots of residual cognitive resilience scores for the silver and gold all-participants models and annual rate of memory decline, coloured according to APOE ɛ4 allele status, amyloid positivity or tau positivity. Amyloid positivity is defined as CERAD stages 3 and 4, whereas tau positivity is defined as Braak stages 3, 4, 5 and 6. Plots of silver resilience scores coloured according to neuropathology are in the subset of participants with available neuropathology measures at autopsy. CERAD = Consortium to Establish a Registry for Alzheimer's Disease.

Across the eight cohorts, a maximum of 20 513 individuals had both genome-wide genotype and resilience phenotype data, making this the largest genetic analysis of resilience to date and the first to incorporate participants of diverse genetic ancestries. Participant characteristics for the silver–gold cross-ancestry meta-analyses in all participants and in participants who were CU at baseline are presented in Table 1. Full participant characteristics by cohort for all models are presented in Supplementary Tables 1 and 2.

Table 1.

Participant demographics

	Silver		Gold		Combined
Variables	CU	All	CU	All	CU	All
Sample size	12 234	17 241	1974	3272	14 208	20 513
AD diagnosis at baseline	0 (0)	2249 (13)	0 (0)	665 (20)	0 (0)	2914 (14)
APOE ɛ2 carrier	1878 (15)	2315 (13)	292 (15)	429 (13)	2170 (15)	2744 (13)
APOE ɛ4 carrier	3652 (30)	6448 (37)	474 (24)	1101 (34)	4126 (29)	7549 (37)
Female	7726 (63)	10 287 (60)	1218 (62)	1824 (56)	8944 (63)	12 111 (59)
European ancestry	10 621 (87)	14 997 (87)	1974 (100)	3272 (100)	12 595 (89)	18 269 (89)
Age at baseline, years	70.54 ± 9.23	71.78 ± 9.02	79.19 ± 7.2	79.28 ± 7.66	71.74 ± 9.46	72.98 ± 9.23
Education, years	15.94 ± 2.91	15.78 ± 2.97	15.85 ± 3.33	15.82 ± 3.19	15.93 ± 2.97	15.78 ± 3.01
Number of study visits	6.29 ± 3.76	5.81 ± 3.6	8.87 ± 5.59	7.04 ± 4.48	6.55 ± 3.97	6.01 ± 3.78
Follow-up time, years	7.89 ± 6.1	6.6 ± 5.76	8.17 ± 4.76	7.11 ± 5.34	8.02 ± 6.04	6.68 ± 5.69
Cognitive resilience score	0.23 ± 0.91	0.05 ± 1.07	0.00 ± 0.99	0.01 ± 1	0.04 ± 1.01	0.04 ± 1.06

Baseline is defined as the first site visit with cognitive scores for two or more domains available. Values shown are the mean ± standard deviation or n (%). AD = Alzheimer’s disease; CU = cognitively unimpaired at baseline.

Genetic correlation and heritability

The gold and silver resilience scores showed no evidence of genetic correlation in all participants (P_FDR = 0.65) and a modest correlation of 0.13 (P_FDR = 3.4 × 10⁻³) in participants who were CU at baseline, perhaps owing to the clinical heterogeneity when including participants with Alzheimer’s disease diagnoses at baseline, similar to our previous report.²

Compared to our previous cross-sectional models of resilience, we observed lower heritability for the silver model [h² = 0.037 (all) − 0.089 (CU at baseline)] and gold model [h² = −0.113 (all) − 0.105 (CU at baseline)]. However, our previous analyses leveraging neuropathology showed lower heritability than our biomarker-based approaches,² hence this might be attributable, in part, to the additional measurement error that is included in our residual when incorporating neuropathology data that are not as proximal to the cognitive assessment as in the biomarker datasets.

Despite the lower levels of heritability, we did note several interesting genetic correlations. Leveraging summary statistics from the EUR ancestry silver–gold meta-analysis, we conducted comprehensive genetic correlation analyses of resilience phenotypes and 65 complex traits to determine the extent of their shared genetic architecture. Pairwise genetic covariances from all-participants and CU models are depicted in Fig. 4. Pairwise genetic covariances for comparable silver and gold models are depicted in Supplementary Figs 5 and 6, and all results are presented in Supplementary Table 4. We observed robust and expected positive correlations with cognitive performance and educational attainment in the CU-at-baseline model (all P_FDR < 1.9 × 10⁻³; Fig. 4), validating our resilience metrics and mirroring previous results from our group. Furthermore, we observed novel correlations between resilience and numerous cardiovascular and neuropsychiatric traits (all P_FDR < 5.0 × 10⁻²). Resilience in the model incorporating all participants was positively associated with multiple resting heart rate parameters (all P_FDR < 5.0 × 10⁻²). and negatively associated with Tourette’s syndrome (P_FDR = 7.0 × 10⁻⁴) and schizophrenia (P_FDR = 1.7 × 10⁻²). There was also a strong negative correlation of resilience with multiple sclerosis in both the all-participants and CU-at-baseline models (P_FDR < 5.8 × 10⁻⁶). Notably, we observed significant negative correlations with Alzheimer’s disease, indicating shared genetic architecture between Alzheimer’s disease and resilience. This stands in contrast to our previous cross-sectional models, suggesting that the silver models might not be removing the contributions of Alzheimer’s disease neuropathology to the same degree as our previous work. As shown in Supplementary Figs 5 and 6, the genetic architecture of resilience phenotypes from the silver and gold models largely overlapped. Traits where silver and gold estimates deviated from one another were largely immune related, potentially related to the effect of neuropathological burden being regressed out of the gold model.

Figure 4.

Genetic covariance results. Genetic covariance estimates with 95% confidence intervals. The top 25 traits ordered on corrected rho are shown. ADHD = attention deficit hyperactivity disorder; BMI = body mass index; CU = cognitively unimpaired; FDR = false discovery rate; HDL = high-density lipoprotein; ICV = intracranial volume; RMSSD = root mean square of successive differences, pvRSAHF = peak–valley sinus arrhythmia, high-frequency power; SDNN = standard deviation of the NN interval (NN interval = the interval between two heart beats).

Gene- and pathway-level results

Gene and pathway test results are presented in Supplementary Tables 5–7. We observed three significant pathway associations with resilience: metabolism of amino acids and derivatives in the silver all-participants meta-analysis (P_FDR = 4.1 × 10⁻²), negative regulation of TGF-β production in the silver CU-at-baseline meta-analysis (P_FDR = 1.9 × 10⁻²), and the severe acute respiratory syndrome (SARS) pathway in the gold CU-at-baseline meta-analysis (P_FDR = 3.9 × 10⁻⁴). We did not observe any significant gene associations with any of the resilience phenotypes.

Single-variant results

Variant-level results are presented in Supplementary Tables 8 and 9, and QQ plots for each meta-analysis are depicted in Supplementary Fig. 7. Results for the silver–gold cross-ancestry meta-analysis in CU individuals at baseline are presented in Fig. 5. Although we did not observe any genome-wide significant signals in any of our single-variant analyses, we observed multiple loci approaching genome-wide significance. The minor allele of the top SNP in the silver–gold cross-ancestry CU analysis, located on chromosome 17 (rs757022*G, minor allele frequency = 0.18), was associated with higher levels of residual cognitive resilience (β = 0.08, P = 1.1 × 10⁻⁷; Fig. 5). The chromosome 17 locus had similar directions of effect across cohorts in the silver and gold models (Fig. 5). In addition, carriage of the minor allele exerted a protective effect on memory and language slopes in the presence of the APOE ɛ4 allele, and there was a significant interaction between ɛ4 allele count and rs757022*G allele dosage on memory slopes (Supplementary Fig. 8). Furthermore, we found evidence for positive associations of rs757022*G with numerous brain volume-related traits using the PheWeb database (https://open.win.ox.ac.uk/ukbiobank/big40/pheweb33k/), including greater cortical thickness in the superior temporal gyrus (P = 1.1 × 10⁻⁴). The minor allele of the top SNP in the all-participants silver–gold meta-analysis analysis, located on chromosome 11 (rs7935771*C, minor allele frequency = 0.40), was associated with lower levels of residual cognitive resilience (β = −0.05, P = 3.4 × 10⁻⁷; Supplementary Fig. 9). rs7935771*C was also associated with multiple brain volume and cortical thickness metrics in the PheWeb database, including higher superior temporal gyrus thickness that could contribute to neuroprotection in late life. Additionally, we found evidence that each SNP acts as an eQTL in multiple databases. rs757022*G was most strongly associated with KCNJ2, ABCA5, ABCA8 and ABCA10 expression (all P < 5.0 × 10⁻²; Supplementary Table 10). Specifically, rs757022*G was associated with decreased ABCA5 expression in hippocampus (β = −0.19, P = 4.5 × 10⁻⁴) and increased ABCA10 expression in dorsolateral prefrontal cortex (β = 0.065, P = 6.4 × 10⁻³; Supplementary Table 10). rs7935771*C was most strongly associated with PAX6 expression in temporal cortex (P = 2.9 × 10⁻⁵) and lower expression of DNAJC24 expression in hippocampus (β = −0.17, P = 2.5.4 × 10⁻⁴; Supplementary Table 11). We did not find any eQTL evidence for rs7935771 in the Brain xQTLServe database nor either variant in eQTLGEN. Finally, we examined our meta-analysis results with respect to rs2571244, a genome-wide significant SNP identified in past resilience work by our group,² and did not find evidence of replication in any of the present analyses.

Figure 5.

Variant-level resilience GWAS results. (A) Manhattan plot showing variant associations with residual cognitive resilience in the CU-at-baseline cross-ancestry meta-analysis. (B) LocusZoom plot displaying the genomic region around rs757022 on chromosome 17. (C) Forest plot of rs757022*G in the silver and gold CU-at-baseline models, including fixed-effects meta-analysis estimates. CU = cognitively unimpaired; GWAS = genome-wide association study.

Discussion

We performed a series of genetic analyses on cognitive resilience phenotypes to identify genetic variants, biological pathways and complex traits associated with resilience to cognitive decline. Initially, we demonstrated that a silver standard approach leveraging longitudinal modelling and APOE with higher-order interaction terms approximates resilience sufficiently to facilitate genomic analysis. However, we also noted some limitations with this approach, including the lower level of SNP heritability and higher genetic correlation with Alzheimer’s disease, suggesting that residual contributions of Alzheimer’s disease neuropathology might remain in our silver standard models. Yet, with this tool in hand we were able to characterize the genetic architecture of resilience further. We observed novel genetic correlations with resilience and numerous behavioural, cardiovascular and neurological traits. We also identified multiple biological pathways associated with resilience, including metabolism of amino acids and derivatives, negative regulation of TGF-β production, and the SARS pathway. Finally, at the single-variant level, we identified multiple candidate loci with biological relevance approaching genome-wide significance. Together, these results highlight the utility of silver models of resilience and complementary approaches to well-validated gold standard techniques to increase sample size and diversity in analyses of genetic resilience to Alzheimer’s disease neuropathology.

The utility of silver standard models of resilience

Our primary results highlight the potential of combining silver standard approximations of resilience with gold standard models to maximize sample size and measurement precision. In our silver standard models, we were able to explain 40%–58% of the variance in cognitive decline and observed high phenotypic correlation with gold standard models that explained a comparable amount of variance in decline (Supplementary Fig. 1 and Supplementary Table 3). Moreover, when visually assessing the resilience phenotypes, higher resilience scores from the silver model were associated with a slower rate of cognitive decline, particularly among those who were neuropathology positive (Supplementary Fig. 2A and B). That said, the reliance on APOE is a limitation because the effect of APOE differs across ancestries, age bands and biological sex,^8,34 suggesting that the model accuracy will also vary along these dimensions. Moreover, the present modelling approach relies on a linear longitudinal model that is likely to underestimate the non-linear changes that occur, particularly in datasets with very long follow-up periods. Despite these limitations, the modelling approach allowed us to increase our sample size more than 5-fold while still retaining information about neuropathology in the participants who have it available. Similar to the GWAX approach that leverages family history as a proxy for disease,^10,11,35 we think a proxy approach to resilience might ultimately produce the sample sizes needed to identify common and rare genetic effects that protect from the downstream consequences of Alzheimer’s disease neuropathology.

It is also notable that the narrow-sense heritability estimates that we see for resilience phenotypes built with cross-sectional compared with longitudinal data, built with silver standard versus gold standard approaches, and built with autopsy versus biomarker measures of neuropathology all vary substantially. In part, these differences are likely to be attributable to the small sample sizes at which the heritability estimates are negative or unstable. It is also possible that key differences in the study populations based on age, sex and concomitant pathways of injury in these various designs also contribute to the discrepancies. Furthermore, our previous analyses leveraging neuropathology showed lower heritability than our biomarker-based approaches,² hence this might be attributable, in part, to the additional measurement error that is included in our residual when incorporating neuropathology data that are not as proximal to the cognitive assessment as in the biomarker datasets. We were also unable to replicate the association of rs2571244 with resilience as identified previously by our group.² There are several differences between the models that might explain the discrepancy. The study by Dumitrescu et al.² used a cross-sectional design as opposed to the longitudinal design presented in the present work. Furthermore, the study by Dumitrescu et al.² used both PET imaging and autopsy cohorts to build resilience models, whereas the present analysis features a larger sample size owing to the incorporation of the largely APOE allele status-dependent silver model. Ultimately, it will be crucial to develop multiple complementary approaches to estimating better-than-expected cognitive performance in the face of neuropathology to ensure that residual measures reflect protection and not simply other elements that are sure to contribute to the residual in such models (e.g. measurement error, other neuropathology or disease processes). Despite these limitations, our results highlight how a silver standard model can be incorporated into genomic analyses to increase sample size, improve statistical power and identify novel genes and pathways that contribute to neuroprotection to carry forwards into mechanistic studies.

Resilience is genetically correlated with cardiovascular and neurological traits

We observed significant associations between resilience and numerous cardiovascular and neurological traits, notably with multiple heart rate-related outcomes, in which the genetic architecture of resilience covaried positively with these traits (Fig. 4). Conversely, resilience was negatively correlated with the genetic architecture of Tourette’s syndrome, multiple sclerosis and schizophrenia (Fig. 4). The negative association with multiple sclerosis was consistent across the all-participants and CU-at-baseline models, indicating that genetic risk for multiple sclerosis might be a strong predictor of faster-than-expected cognitive decline. When comparing the silver and gold models, genetic covariance estimates largely overlapped, whereas select traits, notably immune-related phenotypes, deviated (Supplementary Figs 5 and 6). These differences might be explained by the inclusion of neuropathological measures in the gold model and the lack thereof in the silver model. Genetic risk for dysregulation of immune pathways involved in certain inflammatory conditions, such as Crohn’s disease and inflammatory bowel disease, might be negatively correlated with genetic liability for resilience when measures of amyloid and tau are included in the model, which themselves contribute to brain inflammatory states.³⁶ Without measures of amyloid and tau in the silver model, it might be capturing a genetic architecture that is distinct from the gold definition of resilience regarding overlap with immune-related traits. This is underscored by the lack of significant correlation in either direction in the silver longitudinal models (Supplementary Figs 5 and 6).

Furthermore, we observed consistent negative correlations with genetic risk for Alzheimer’s disease, which deviates from the results reported previously by our group.² Reasons for this observation might stem from differences in the way resilience was modelled; past work incorporated both autopsy and PET-derived measures of amyloidosis, whereas the present work incorporates only autopsy measures of both amyloid and tau. As such, it is expected that the genetic architecture of resilience would differ based on the way the resilience metrics themselves were generated.

Pathway tests implicate genetic variation in amino acid metabolism, TGF-β and SARS-related pathways in resilience

MAGMA pathway tests identified three biological pathways significantly associated with resilience: the Reactome metabolism of amino acids and derivatives pathway in the silver all-participants analysis (P_FDR = 4.1 × 10⁻²), the Gene Ontology negative regulation of the TGF-β pathway in the silver CU-at-baseline-participants analysis (P_FDR = 1.9 × 10⁻²), and the Biocarta SARS pathway (P_FDR = 3.9 × 10⁻⁴) in the gold CU-at-baseline-participants analysis. Although the role of amino acid metabolism in Alzheimer’s disease remains unclear, multiple studies have proposed a connection. Notably, mouse models of Alzheimer’s disease display higher plasma branched chain amino acid levels compared with controls,³⁷ and deficits in branched chain amino acid metabolism can contribute to Alzheimer’s disease neuropathology via mTOR signalling.³⁸ Several amino acids serve directly or as precursors to neurotransmitters, importantly glutamate, tryptophan and tyrosine, and changes in amino acid metabolism might affect neurotransmitter levels known to be altered in Alzheimer’s disease.^39,40 Furthermore, alterations in branched chain amino acid metabolism have been observed in insulin resistance, type 2 diabetes and obesity, all of which are metabolic conditions that increase risk of Alzheimer’s disease.^41,42 It is worth noting that our group observed nominal associations with amino acid metabolic pathways and resilience in past work, underscoring the potential connection between these two phenotypes.²

TGF-β is a ubiquitously expressed cytokine, and deficiency of TGF-β signalling has been shown to increase both amyloid-β accumulation and amyloid-β-induced neurodegeneration in models of Alzheimer’s disease.⁴³ TGF-β is upregulated in astrocytes and microglia in Alzheimer’s disease⁴⁴ and exerts a neuroprotective effect, blocking amyloid-β-mediated synapse degradation in an Alzheimer’s mouse model.⁴⁵ Although TGF-β activity modulates expression of a plethora of downstream effectors, it is becoming increasingly relevant to Alzheimer’s disease therapeutics given its role in glial responses to injury and attenuation of neuroinflammation.⁴³

The viral hypothesis of Alzheimer’s disease remains controversial, yet substantial neurological manifestations have been described in a portion of SARS coronavirus 2 cases that present as memory impairment and difficulty in concentrating.^46,47 Furthermore, recent evidence has implicated infection with the Epstein–Barr virus as a leading cause of multiple sclerosis, a demyelinating disorder characterized by chronic inflammation, pain, difficulty with motor coordination, and cognitive impairment.^48,49 Past research indicates that certain viruses can reside latently within the nervous system and potentially cause chronic inflammation or directly damage neural tissue upon reactivation.^50-52 As such, inflammatory responses to viral infection might represent a common pathway between Alzheimer’s disease pathogenesis and multiple sclerosis, with our results indicating that genetic variants in the SARS response pathway might confer protection from virally mediated pro-inflammatory responses that contribute to neurodegeneration and cognitive decline. This connection is particularly intriguing when considering the negative genetic correlation between resilience and multiple sclerosis reported here and in previous work by our group.³ It is worth noting that the SARS-related pathway was identified in the gold model and that the association was not present in the higher-powered silver model, potentially owing to the effects of neuropathology being regressed out of the gold model, suggesting that additional validation will be needed in the future.

Top variant-level associations highlight ABC family transporters, PAX6

We observed multiple loci approaching genome-wide significance across our analyses and chose to focus on the top signals from the CU-at-baseline and all-participants meta-analyses. The top SNP from the CU-at-baseline analysis, rs757022, is categorized as an intron variant in LINC01483 and positioned upstream of MAP2K6 and numerous ATP-binding cassette (ABC) family genes on chromosome 17 (Fig. 5). The minor allele at this SNP, rs757022*G, was associated with higher levels of residual cognitive resilience, representing a putative resilience locus. Interestingly, we also found evidence that this SNP acts as an eQTL in brain, notably for multiple ABC genes, including ABCA5, ABCA6, ABCA8, ABCA9 and ABCA10 (Supplementary Table 10). ABCA7, located on chromosome 19, has been identified as a high-confidence gene candidate for late-onset Alzheimer’s disease risk, whereas certain variants in ABCA7 exert a protective effect.⁵³ Consequently, the ABC superfamily is shown to play crucial roles in maintaining proper cholesterol efflux in the brain, with family members displaying preferential protein expression on distinct cell types.⁵⁴ This includes ABCA5 and ABCA10 expression in neurons and ABCA8 expression in oligodendrocytes, whereas ABCA9 expression is less thoroughly characterized.^55,56  ABCA5 expression is elevated in the hippocampus of Alzheimer’s disease brains,⁵⁷ which is particularly interesting given the observed positive association of rs757022*G with resilience and its negative association with ABCA5 expression in hippocampus. Likewise, ABCA5 expression in brain is upregulated in Parkinson’s disease and strongly correlated with sphingomyelin levels.⁵⁵ Genetic variation in ABCA8 has previously been linked to increased risk of multiple sclerosis, underscored by the negative association we observed between the genetic architectures of resilience and multiple sclerosis (Fig. 4).⁵⁸ In parallel, the identification of genes involved in cholesterol efflux potentially implicated in resilience is especially intriguing given the negative association we observed between the genetic architectures of resilience and high-density lipoprotein cholesterol (Fig. 4). As such, understanding mechanisms by which ABC transporter family members affect cholesterol efflux, amyloid-β production and clearance, and myelination could provide valuable context as to how genetic variation at this locus relates to resilience to Alzheimer’s disease.

We also observed a separate locus approaching genome-wide significance in the all-participants meta-analysis, in which the minor allele, rs7935771*C, was associated with lower levels of residual cognitive resilience. The strongest eQTL evidence for this SNP was for PAX6, a transcription factor important for normal ocular and neural development (Supplementary Table 11).⁵⁹ Moreover, PAX6 signalling has been implicated in amyloid toxicity-mediated tau phosphorylation through direct upregulation of GSK-3β, one of the dominant tau kinases.^60,61 These findings are underscored by recent work linking elevated PAX6 to inhibition of isocitrate dehydrogenase 3β, a crucial citric acid cycle enzyme, leading to tau hyperphosphorylation and synapse impairment.⁶² Collectively, the loci identified in single-variant analyses could modulate resilience to cognitive decline through a plethora of biologically plausible mechanisms, reinforcing the utility of our approach.

It is important to highlight that in sensitivity analyses, the P-value of the top variant in our silver–gold all-participants meta-analysis, rs7935771, was attenuated in models incorporating last visit clinical diagnosis (β = −0.043, P = 1.2 × 10⁻⁴). There are a few possible explanations for the attenuated P-value; first, clinical diagnosis at last visit is closely related to the resilience phenotype, because both are influenced by the underlying neuropathological burden and cognitive performance. Including clinical diagnosis as a covariate in the model might account for some of the variance that was previously attributed to the genetic variant, thereby weakening the apparent association between the variant and resilience. Furthermore, the genetic variant might influence resilience through mechanisms that are also captured by the clinical diagnosis. For example, if the variant contributes to cognitive performance independent of neuropathology, and clinical diagnosis is partly based on cognitive outcomes, the inclusion of clinical diagnosis could weaken the direct association between the variant and resilience. Finally, resilience is conceptually distinct from clinical diagnosis, focusing on better-than-expected cognitive performance given a level of Alzheimer's disease neuropathology. Including clinical diagnosis, which might be influenced by other factors (e.g. comorbidities, diagnostic thresholds), could introduce noise or confounding, leading to an attenuated P-value. We chose to keep our primary models agnostic to clinical diagnosis owing to the evolving theoretical discussion surrounding this topic and plan to investigate further in future work.

Strengths and limitations

Our study has numerous strengths. We harmonized data across eight deeply characterized cohorts of cognitive ageing and incorporated a cross-ancestry analytical approach, completing the largest and most diverse genetic analysis of resilience to date. Our longitudinal modelling approach, with an average follow-up time of 6–8 years, represents a methodological advancement for the field and will benefit from rapidly expanding sample sizes as more participant data become available. Generation of resilience composites from residuals in three cognitive domains yields a robust anchoring of resilience metrics across cohorts. Gold models incorporated measures of both amyloid and tau at autopsy, providing a thorough view of an individual’s neuropathological landscape. Moreover, comprehensive post-GWAS tests bring valuable insight to the genetic architecture of resilience and biological pathways potentially implicated in the brain’s shield against neurodegeneration.

In addition to its strengths, our study also has some limitations. Although the cross-ancestry approach is novel in genetic analyses of resilience, participants were still predominantly of European descent (89%), limiting the generalizability of our results to more diverse populations. Furthermore, reliance on APOE as a proxy for neuropathology is a limitation, because the effect of APOE varies based on genetic ancestry, age and sex,^8,34 and it is likely that the model accuracy would change in accordance with these dimensions. Additional loci beyond APOE contributing to Alzheimer’s disease risk could be characterized through genetic risk scores and incorporated in our modelling approach, but it will be crucial to address how to analyse the regions around those loci properly in diverse genetic ancestries to enable this approach in the future. Additionally, we considered self-reported race/ethnicity to be synonymous with ancestry. Newer tools that consider population structure at the SNP level will allow for more rigorous admixed GWAS. Although we did not formally integrate diagnostic information into our base resilience models as regressors, we did note in sensitivity analyses that these models performed well, suggesting that future work might benefit from harmonizing and integrating longitudinal diagnostic information as a strong alternative approach to build resilience metrics that are more sensitive to premorbid and post-dementia cognitive decline. The present analysis also focused on common variants with small effect sizes, and it is likely that rare variants with large effect sizes could also affect resilience. As such, leveraging whole genome and exome sequencing to identify rare variant associations with resilience remains an important focus for future work. Finally, our linear mixed-effects approach is likely to underestimate the non-linear changes that occur in ageing, and future work should focus on incorporating these effects in models of resilience to Alzheimer’s disease neuropathology.

Conclusion

The findings of our cross-ancestry genetic analysis of resilience to cognitive decline validate the utility of silver models of resilience that forego the need for measures of neuropathology, greatly expanding statistical power and participant diversity. Our findings suggest that such modelling approaches complement gold-standard approaches in the identification of protective variants, genes and biological pathways that protect the brain from the downstream consequences of neuropathology.

Data availability

The present analyses used phenotype and genetic data that can be accessed on NIAGADS (https://dss.niagads.org/). Additionally, the ADSP-PHC provides other phenotype meta-data accessible through a data curation tool hosted at Vanderbilt (https://vmacdata.org/adsp-phc). The findings presented in this publication also rely, in part, on data acquired from the Accelerating Medicines Partnerships—Alzheimer’s Disease Target Discovery and Preclinical Validation Project (AMP-AD) (https://adknowledgeportal.synapse.org/). ROSMAPMARS data are available at www.radc.rush.edu.

Funding

This work was supported by the National Institutes of Health under award numbers F31 AG085980, U24 AG074855, R01 AG059716 and R01 AG073439.

Competing interests

T.J.H. sits on the scientific advisory board for Vivid Genomics. The remaining authors report no competing interests.

Supplementary material

Supplementary material is available at Brain online.