Dynamics of cognitive variability with age and its genetic underpinning in NIHR BioResource Genes and Cognition cohort participants

Dementia affects a growing global population, projected to reach approximately 139 million by 2050. Despite recent therapeutic advances, many candidate treatments fail in late-stage clinical trials, potentially because they are tested too late in the disease course. There is an urgent need to understand mechanisms during preclinical and prodromal stages and to evaluate interventions earlier. Large, recallable cohorts are required to enable early studies and careful stratification. The NIHR BioResource Genes and Cognition (G&C) cohort was established to profile cognition at scale across adulthood, to quantify demographic and environmental determinants, and to investigate genetic influences—including APOE genotype and Alzheimer’s disease polygenic risk—on cognitive performance and trajectories. The study further aimed to identify genetic loci underlying general cognitive ability and to assess whether cognition-related genetics overlap with Alzheimer’s disease risk biology.

Prior work highlights frequent translational failures in Alzheimer’s disease clinical trials, attributed to limited pathophysiological understanding, imperfect animal models, and evaluation too late in the disease course. Small studies have reported inconsistent associations between APOE genotype and cognition in healthy individuals, with some age-specific effects noted for processing speed and visual memory. Vocabulary performance (a proxy for crystallized intelligence) is generally reported to increase through midlife and decline later, though findings vary regarding the age of decline onset. Polygenic risk scores for Alzheimer’s disease have been proposed for risk stratification, yet their relation to cognition in healthy populations is uncertain. Previous GWAS have identified many loci for general cognitive function, fluid intelligence, and educational attainment, with moderate to high heritability and substantial polygenicity. The extent to which normal cognition genetics overlaps with Alzheimer’s disease risk has remained unclear.

Study design and population: The G&C study is a prospective open cohort nested within the NIHR BioResource (England), recruiting from the general population and NHS organizations. Participants consented to recall for future studies. Two phases occurred: pilot (June–August 2020) and main (November 2020–November 2021). A total of 21,052 participants formed the study base; after exclusions (withdrawn consent or missing key data), 21,051 were available. Participants were considered cognitively healthy at recruitment. Ethical approvals were obtained (REC REF: 19/NS/0118; NIHR BioResource ethics REC REF: 13/EE/0325 and 17/EE/0025). Data collection: Participants provided blood (or saliva) for DNA and completed questionnaires at NIHR BioResource recruitment on demographics and lifestyle (age, gender, ethnicity, BMI, smoking, alcohol) and diagnoses (for example, diabetes, stroke, mental health). Before testing, a brief questionnaire captured additional items (e.g., first language, color blindness, learning disability). Cognitive testing: Participants completed online cognitive tests using the Cognitive Test app (versions 4.4.7–5.6.7) on compatible devices (iOS, Android, Windows). Eleven tests were administered: Reaction Time (RT), Stroop Box (SB), Stroop Ink (SI), Symbol Digit (SD), Trail Making Numeric (TMN), Trail Making Alphanumeric (TMA), Matrices (MX), Working Memory (WM), Quiz (QZ), Vocabulary (VY), and Pairing 7 (PR). Total time was approximately 30 minutes. Scores were transformed so that higher values indicate poorer performance across tests. Two data-driven summary measures of general cognitive ability were derived using principal component analysis: G6 (PC1 of RT, SB, SI, SD, TMN, TMA; explained variance 66.5%) and G4 (PC1 of MX, QZ, VY, WM; explained variance 46.6%). Device type effects were assessed; except for WM, device-related differences persisted after adjustment and were included as covariates in analyses. Genotyping and QC: DNA was genotyped on Affymetrix Axiom UKB arrays (v1.0 hg37 and v2.1 hg38; lifted to hg37; 708,654 shared variants pre-imputation). QC included minor allele frequency <0.01, marker and individual missingness >0.01, Hardy–Weinberg equilibrium P<1e-6, extreme heterozygosity exclusion, removal of monomorphic variants and allele mismatches to HRC. After QC, 518,164 autosomal markers were imputed to the HRC reference on the Michigan server. Genetic sex and ancestry were inferred; only genetically inferred European ancestry samples were analyzed. Twenty genetic principal components were created post-imputation. APOE alleles were determined from imputed rs429358 and rs7412; ambiguous genotypes (n=279) were excluded, leaving e2/e2 (n=69), e2/e3 (n=1,238), e3/e3 (n=5,931), e3/e4 (n=2,304), e4/e4 (n=218). Polygenic risk scores: AD-PRS were based on the PGS Catalog PGS002289 (Lambert et al.). Two scores were computed with PRSice-2: AD-PRSAPOE (20 SNPs including APOE) and AD-PRS without APOE (18 SNPs). Participants were categorized as high-risk (>95th percentile) vs low-risk (≤95th percentile). Statistical analyses: Descriptive statistics summarized demographics and cognitive scores for the full sample and genetic subset. Individuals self-reporting diagnoses likely to affect cognition (n=123) were excluded from analyses. Phenotypic correlations used Pearson’s r. Device associations used linear regression with age and gender covariates. Age and gender effects were modeled with stepwise linear regression including centered age, age^2, age-by-gender, age^2-by-gender, and device (except WM), with deprivation and ethnicity tested in sensitivity analyses. Education and multiple deprivation associations were tested with linear regression and linear trend, adjusting for base covariates. Associations with self-reported diagnoses used linear regression adjusted for age terms, gender, and device. Multiple testing used Bonferroni–Holm across 13 phenotypes. APOE and PRS analyses: Cognitive trajectories were plotted against age stratified by genotype or PRS group. Candidate phenotypes were tested using linear mixed-effects models including age terms, genotype/PRS group, and interactions, adjusted for sex (genetically inferred), device, genotyping batch (random effect), array, and first five genetic PCs; multiple testing was Bonferroni–Holm corrected. Heritability and genetic correlation: SNP heritability for 13 phenotypes was estimated with BOLT-REML; bivariate GREML (GCTA) estimated genetic correlations, using phenotype residuals adjusted for covariates and including batch, array, and PCs. LD Score Regression (LDSR) estimated genetic correlations of G4 and G6 with childhood and adult IQ, educational attainment, and Alzheimer’s disease using HapMap3/1000 Genomes European LD scores. GWAS and functional analyses: GWAS of G4 and G6 used BOLT-LMM with additive SNP effects and covariates (age, age^2, sex, batch, array, PCs; per Supplementary Table 17). Filters: MAF ≥0.05, INFO ≥0.50, HWE P≥1e-6; genome-wide significance at P<5e-8. Loci were mapped and annotated with FUMA (MAGMA gene-based tests, positional, eQTL, and chromatin interaction mapping; ANNOVAR annotations with CADD and RegulomeDB). Colocalization with microglia eQTLs used COLOC; SMR/HEIDI tested putative causal gene expression effects in 12 GTEx v8 brain regions. Fine-mapping employed GCTA-COJO and FINEMAP allowing up to five causal variants. Replication of top SNPs was assessed in prior intelligence meta-analyses (Savage et al.; Sniekers et al.), harmonizing directions to the G&C phenotype definitions. Data access: GWAS summary statistics for G4 and G6 were deposited in Zenodo (10.5281/zenodo.10836380).

• Cohort: 21,051 participants (ages 17–85; mean 50.48, SD 14.81); 63.2% female; 96.8% White; device usage: iOS 46%, Android 31%, Windows 23%. After excluding 123 individuals with conditions affecting cognition, 20,928 were analyzed for primary results; genotyping and specific analyses included subsets (e.g., APOE n≈9,691). - Device effects: Significant device-related differences in test scores (except WM) remained after adjusting for age and gender; device was included as a covariate in analyses. - Age and cognition: Performance worsened with age across all tests except Vocabulary (VY), which increased with age; no decline in VY beyond 60 years was evident in this cohort (Bonferroni–Holm P<0.05). - Gender: Small differences by gender across some phenotypes; overall, gender explained only 0.1–1.33% of variance. Significant age-by-gender interactions were observed for SD, VY, and G4; minimal contributions of age and gender to WM (1.09%), QZ (1.16%), and G4 (2.53%). - Socioeconomic factors: Lower education associated with worse performance across all phenotypes with a significant linear trend; higher multiple deprivation correlated with worse performance for all phenotypes except PR (Bonferroni–Holm P<0.05). - APOE genotype: ε4 carriers showed subtle increases (worse performance) in several measures emerging in mid-life (45–64 years), but associations did not withstand covariate adjustment or correction for multiple testing. No compelling evidence that APOE genotype influenced performance across the 11 cognitive tests in the 9,691 individuals with unambiguous APOE genotypes. - AD polygenic risk scores: Comparing top 5% vs bottom 95%, small deviations in some scores emerged only in later life. In adjusted models, age-by-AD-PRSAPOE interaction was detected for SI and G6, but after multiple testing only SI remained significant (adjusted P=0.039). No significant interactions for PRS without APOE. Overall, AD-PRS had minimal impact on cognitive performance and were inferior to APOE genotype for early change detection. - Heritability: SNP heritability for cognitive phenotypes ranged from 0.06 to 0.28, with genetic correlations between phenotypes stronger than phenotypic correlations. - GWAS of general cognitive ability: Distinct loci underlie G4 and G6 with no evidence of population stratification confounding (LDSR intercepts near 1). • G4: Strongest association on chromosome 16; independent SNP rs62034351 (intronic, within CCDC101/SGF29 region) reached genome-wide significance and explained 0.37% of G4 variance (ANOVA P=1.38×10^-8), 185-fold more than APOE (0.002%, P=0.93). Four additional loci were suggestive (P<1e-6). The rs62034351 signal replicated directionally in two intelligence meta-analyses. Functional annotation implicated microglial immune pathways (e.g., interferon gamma response) with colocalized microglia eQTLs for TUFM, SULT1A1, and SULT1A2; SMR analyses supported brain region-specific expression effects. Fine-mapping highlighted rs3743963, rs11074904, rs62031607, and rs2411453 as plausible causal variants. • G6: Strongest locus on chromosome 3 near GBE1; rs11705789 explained 0.11% of G6 variance (ANOVA P=2.52×10^-5), 5.5-fold more than APOE (0.02%, P=0.21). Three additional loci were suggestive. GWGBA identified GBE1; rs11705789 was an eQTL for GBE1. Replication in prior intelligence meta-analyses was not observed, possibly due to differences in phenotype composition. Fine-mapping nominated rs12635671, rs820270, and rs2691073 as likely causal. - Genetic correlations with external traits: G4 and G6 showed high genetic correlation with childhood and adult IQ, with higher estimates for G4; G4 had approximately 2.4-fold higher genetic correlation with educational attainment than G6. Genetic correlation between cognition (G4, G6) and Alzheimer’s disease was very weak, indicating largely distinct genetic architectures.

Large-scale, device-enabled cognitive phenotyping across adulthood confirmed known determinants of cognition (age, education, socioeconomic status) and showed negligible differences between males and females, with minimal explanatory power for gender overall. Contrary to smaller prior studies, neither APOE genotype nor Alzheimer’s disease PRS meaningfully influenced cognitive performance in healthy individuals across most of adulthood; small effects appeared only in later life and did not robustly survive multiple corrections, suggesting limited utility for early stratification by AD genetic risk in the general population. In contrast, genome-wide analyses of general cognitive ability revealed distinct genetic architectures for fluid/crystallized domains (G4) versus processing speed/executive function (G6). For G4, convergent functional evidence implicated microglia-mediated immune pathways, including interferon gamma responses, and prioritized genes such as TUFM, SULT1A1, and SULT1A2 with brain region-specific expression effects. For G6, the lead locus near GBE1 highlighted a plausible role for glycogen metabolism in attention and executive function, consistent with clinical observations in adult polyglucosan body disease and emerging literature on brain glycogen. The weak genetic correlation between cognition measures and Alzheimer’s disease risk suggests that the molecular basis of normal cognitive variation is largely distinct from dementia risk biology, pointing to new opportunities to target age-related cognitive decline independently of Alzheimer’s disease susceptibility pathways. The NIHR BioResource G&C cohort’s consented recall-by-genotype/phenotype enables well-controlled natural history and interventional studies, with matching on key confounders and targeted inclusion of mechanistically relevant subgroups.

This study establishes a large, recallable resource linking cognitive phenotypes and genetics across adulthood, demonstrating that common dementia risk variants, including APOE and AD-PRS, have minimal impact on cognition in healthy individuals until late life. GWAS of general cognitive ability identified distinct biology for different cognitive domains: microglial immune processes for fluid/crystallized abilities (G4) and glycogen metabolism via GBE1 for processing speed/executive function (G6). These findings indicate that mechanisms underlying normal cognitive variability differ from Alzheimer’s disease risk loci and suggest novel targets to prevent or slow age-related cognitive decline. Future work includes longitudinal retesting to define trajectories, expansion to more diverse ancestries, incorporation of long-read genome sequencing, and recall for deep phenotyping (including disease-specific biomarkers and neuroimaging) and early-phase interventional trials.

• Cross-sectional cognitive assessments to date may dilute associations due to measurement error; longitudinal follow-up is pending. - Device dependence of most tests (except WM) could confound results; although adjusted analytically, device effects must be considered in recall studies. - The cognitive battery does not cover all domains (for example, verbal episodic memory, visuospatial skills). - Genetic analyses focused on participants of European ancestry and a cohort enriched for higher education, limiting generalizability across populations. - Lack of disease-specific biomarkers (e.g., imaging, fluid biomarkers) means recalled participants may not differ from the background population with respect to neurodegeneration markers. - Some APOE and PRS effects appeared only in later life and were small; multiple testing corrections further limited detectable associations.