Human longevity is influenced by many genetic variants: evidence from 75,000 UK Biobank participants

Genome Wide Association Study: smoking-related variants are associated with father's age at death

Of the 9,658,292 genetic variants included none were significantly associated (p<5×10⁻⁸) with “parent's age at death, combined” however one locus (36 variants in strong Linkage disequilibrium) on chromosome 15 was associated with “father's age at death,” and 1 variant (on chromosome 22) with “mother's age at death” (see Table Supplementary 2 for the top 1,000 results from each GWAS; see Figures Supplementary 2-5 for Manhattan and QQ plots). We focused on the locus on chromosome 15 associated with father's age at death, because the variant associated with mother's age at death was of low frequency (3%) and not in a typical “peak” expected of robust results (see Figures Supplementary 6-7 for LocusZoom plots [20]). rs1051730 is in this loci (beta between G allele and father's age at death=−0.0269, se=0.0049, p=3×10⁻⁸); the A allele of rs1051730 in the nicotinic acetylcholine receptor alpha 3 subunit CHRNA3 gene has been linked to smoking fewer cigarettes and lower risks of lung cancer, although this variant does not influence the chances of starting to smoke [21]. This analysis included fathers who died at ≥46 years of age; including only father's aged ≥66 (the short-lived cut-off empirically calculated – see methods) reduces the association but it remains significant (beta= −0.0207, p= 6×10⁻⁶; sample size reduced from 63,775 to 47,094).

We tested rs1051730 against smoking-status (in logistic regressions using binary current/never smoking status as the outcome) in the participants and found that per G allele there is an increased likelihood of current smoking (OR=1.063, 95% CIs=1.020 to 1.108, p=0.003). Participants’ smoking status is associated with father's age at death (per year OR=0.993, 95% CIs=0.991 to 0.996, p=2×10⁻⁶), however adjusting for smoking status (current/former/never) in the model of rs1051730 had little effect on the association with father's age at death (p=4.7×10⁻⁸).

The association between rs1051730 and mother's age at death (Beta=0.017, p=1.6×10-3) did not reach genome wide significance in our UK Biobank participants. We tested the association of rs1051730 with father's age at death in the Framingham Heart Study (FHS) generation 2, where 2033 participants were available (mean age at death of the fathers=77.4 years, SD=11.5, range=47.2 to 102.8). The association was non-significant, but was directionally consistent (per C allele coefficient= 0.008, p=0.98). However, the effects size of this SNP is modest and the power to detect the association in the FHS was only 1% (Quanto parameters: coefficient=0.0269, n=2033, MAF=0.17, alpha=0.05).

In the GWAS against the binary “extreme age” phenotype (at least one parent very long-lived) two variants (rs528161076 and rs75824829, on chromosome 7 and 9 respectively; see Figures Supplementary 8-9 for Locus Zoom plots) were found to be significant (p<5×10⁻⁸). rs75824829 may be anomalous as it is not in a “peak” of associations expected of robust results (Figure Supplementary 4; Figure Supplementary 9), however rs528161076 is in a distinct peak of variants in an intron of AP5Z1 (adaptor related protein complex 5, zeta 1 subunit, thought to be involved in homologous recombination DNA double-strand break repair). The association between rs528161076 and extreme longevity should still be interpreted with caution, as the variant is not associated with the continuous age at death variables (combined parent's age at death p=0.65, father's p=0.47 and mother's p=0.96).

In all four GWAS performed, no variants on the X or Y chromosomes were associated with the phenotypes (p>1×10⁻⁵). For mitochondrial variants, the smallest p-values for mother's age at death, was 9×10⁻³.

Heritability of parental longevity explained by genotyped common variants

We determined the variance in parental age at death explained by all the common genetic variation (minor allele prevalence >1%) robustly genotyped directly on the UK Biobank arrays (n=457,643; see methods). For combined parents age at death, the variance attributed to the measured genotypes was 8.47% (SD=1.06%); for mothers age at death 4.85% (SD=1.01%) and fathers age at death 5.35% (SD=1.04%). In a sensitivity analysis including all directly genotyped variants irrespective of minor allele frequency and missingness (n=845,997) the variance attributed for the combined age at death phenotypes was 10.24% (SD=1.26%); for mothers age at death 6.08% (SD=1.21%) and fathers age at death 5.79% (SD=1.23%).

Genetic Risk Score associations

For each genetic risk score (GRS) we first verified the association with the best-fit phenotype corresponding to the reported association: for example the coronary artery disease (CAD) GRS was tested against prevalent coronary heart disease in the offspring (CHD= myocardial infarction or angina) (Table 2). All GRS were associated with their available phenotypes in UK Biobank in the expected direction, for example CAD GRS and CHD (per weighted allele increased OR=1.12, 95% CIs: 1.10 to 1.13, p=1×10⁻⁶⁴). The lipid GRS could not be tested directly, as serum lipid concentrations are not yet available, however all three separate risk scores are significantly associated with CHD. Although there were only 28 participants with dementia, a significant association was seen with the Alzheimer's disease GRS (per weighted allele OR=1.52, 95% CIs: 1.17 to 1.98, p=2×10⁻³). Telomere length was not assessed in the Biobank participants so validation was not possible.

We next tested associations between each GRS and parental age at death (combined mothers and fathers) in linear regression models plus ‘extreme longevity’ (at least 1 parent lived to the top 1% of the age at death distribution in UK Biobank) (Figure 1). These analyses included 76 statistical tests (19 GRS against 4 primary longevity phenotypes) and therefore we have ‘starred’ associations passing Benjamini-Hochberg correction only, although each test has a strong and distinct prior hypothesis (see full results in Table Supplementary 3). We observed associations in the expected directions (i.e. lower disease risk score or higher numbers of relatively protective alleles associated with older ages at death) between the combined parental age at death continuous trait and eight GRS’; for CAD (unadjusted p=1.5×10⁻¹¹), LDL cholesterol (p=1.8×10⁻⁷), CAD without lipid-associated alleles (p=1.8×10⁻⁷), triglycerides (p=1.0×10⁻²), systolic blood pressure (SBP, p=1.2×10⁻⁴), BMI (p=8.2×10⁻⁴), Inflammatory Bowel Disease (p=7.0×10⁻³), T1D (7.5×10⁻³), and Alzheimer's disease (p=1.4×10⁻²). Crohn's disease (p=2.3×10⁻²) and breast cancer (p=3.5×10⁻²) were associated at nominal significance (p<0.05) but not after multiple-testing correction.

A number of GRS had similar trends for the extreme longevity association but become statistically non-significant in the binary analyses of offspring of long-lived parents, which has reduced statistical power (binary analysis vs. continuous). The HDL cholesterol GRS was associated with extreme longevity (p=5.7×10⁻³) although this association did not reach statistical significance in the linear analysis of continuous age at death.

We observed consistent effect directions, but reduced effect size and significance in separate analyses of mother's age at death compared to father's (Table Supplementary 3); with the exception of Alzheimer's disease (AD) GRS which is only associated with mother's age at death but not father's (coefficient=−0.111, p=3.9×10⁻⁴; coefficient=−0.001, p=0.99, respectively) and Triglyceride level (TG) GRS was only associated with fathers not mothers age at death (father's coefficient=−1.366, p=9×10⁻⁴; mother's coefficient=−0.183, p=0.63, respectively). In linear regression models for AD and TG GRS there were no statistical interactions between mothers and fathers ages at death (p>0.05).

To illustrate combined effect sizes, we computed an estimate across the three GRS most robustly associated with parental longevity (unadjusted p<0.001): CAD, SBP and LDL. In total, 525 participants (of 45,627) were in the bottom quintile (20%) for all three GRS; these participants are the “lowest risk” group for these three cardiovascular traits. Correspondingly, 524 participants were categorized as having the “highest risk” (top 20% of genetic risk for all three traits). In adjusted logistic regression models, participants with the lowest genetic risk (highest number of protective alleles) had 3.25 times the likelihood of having at least 1 parent in the top 1% of ages at death, compared to those with the highest genetic risk (OR=3.25, 95%CI=1.3 to 8.1, p=0.012). Unadjusted prevalence: 19 participants (1.4% of 1,339) with at least one long-lived parent were in the bottom 20% for all three risk scores, where only 7 participants (0.5% of 1,339) were in the top 20% for all three risk scores.

In a sensitivity analyses we removed variants mapped to the APOE locus from the risk scores (rs4420638 was included in the scores for CAD, LDL, HDL and AD) and assessed the associations with the parental age-at-death phenotypes. After exclusion the AD risk score is no longer associated with any of the parental age at death phenotypes. The effect sizes for CAD, LDL and HDL are attenuated but remain significant (see Figure Supplementary 10). Adjusting for presence of at least one ApoE-ε4 allele in the models against continuous age-at-death has the same effect on the associations: the CAD GRS association changes from (coef= −0.023, p= 1×10⁻¹⁰) to (coef= −0.022, p= 2×10⁻⁹); the LDL GRS association changes from (coef= −0.28, p= 6×10⁻⁸) to (coef= −0.22, p= 1×10⁻⁴); the HDL GRS association changes from (coef= 0.078, p=0.15) to (coef= 0.050, p=0.37); the AD GRS association changes from (coef= −0.015, p=0.008) to (coef= 0.002, p= 0.79).

Candidate genetic variants associated with longevity

We have highlighted the GWAS results for a number of SNPs a priori selected from the literature (see Table Supplementary 4 for the full details). We found significant (p<0.05, not ‘genome-wide’ significant p<5×10⁻⁸) associations between 4 of the 5 APOE/TOMM40 SNPs tested and parents age at death (rs2075650 p=1×10⁻⁴, rs429358 p=8×10⁻⁷, rs7412 p=0.02, rs4420638 p=3×10⁻⁵, but not rs405509 p=0.06). These variants are in moderate linkage disequilibrium (R² from 0.1 to 0.7). Of 12 FOXO3 variants assessed, none were significantly associated with a longevity phenotype (p>0.05). The 5q33 variant identified by Deelen at al. [13] is associated with combined parents age at death and father's, but not mother's age at death or extreme survival (p=0.005, p=0.01, p=0.1 and p=0.6, respectively). We observed modest associations with 2 of 7 variants at the CDKN2A locus (variants previously associated with coronary heart disease) with parent's age at death (rs1333049 G allele beta=0.015, p=0.0015; rs4977574 A allele beta=0.014, p=0.0018); the CDKN2A variants associated with type-2 diabetes were not associated with parent's longevity. Four variants were reported by Fortney to be associated with exceptional longevity; one is the APOE variant already discussed, and the other three are rs4977756 (CDKN2B/ANRIL), rs3184504 (SH2B3/ATXN2) and rs514659 (ABO). Only rs3184504 is associated with extreme longevity in this analysis (T allele beta=−0.0019, p=0.024), but rs4977756 was associated with continuous parent's age at death (G allele beta=0.017, p=0.025). The final SNP is not associated with any longevity phenotypes in our analysis.

Presence of at least one ApoE-ε4 allele (10,237 of 44,574 participants included in the analysis against continuous parents age at death, after excluding 1,053 ε2/ε4 participants) was associated with reduced age at parents death (coef= −0.088, 95% CI= −0.12 to −0.05, p=3×10⁻⁷) and reduced likelihood of having a parent in the top 1% of the age-at-death distribution (OR=0.78, 95% CI=0.66 to 0.91, p=0.002), compared to participants without an ε4 allele. Presence of at least one ApoE-ε2 allele (5,700 of 44,574) was associated with continuous parents age-at-death (coef= 0.052, 95% CI=0.01 to 0.09, p= 0.014) but not with having a parent in the top 1% of the age-at-death distribution (OR=1.09, 95% CI=0.91 to 1.31, p=0.37).

Comparison with previous work

There is consistent epidemiological evidence for a relationship between longevity and lower rates of type-2 diabetes [17], yet previous evidence from the Leiden study (n=2415) and the Life Long Family Study (n=1562 in generation 1; n=3102 generation 2) suggested that offspring of long-lived participants do not have lower burden of type-2 diabetes risk alleles compared to their partners [16,25]. We also did not observe an association between the type 2 diabetes GRS and longevity in the UK Biobank participants, although small effects cannot be excluded.

There is evidence to suggest that long-lived individuals have reduced prevalence of Alzheimer's increasing alleles [16], which includes the TOMM40/APOE locus identified in multiple analyses of longevity [13,14,12,10]. We confirmed that the A allele of rs2075650 was positively associated with increased parental age at death (beta=0.039, p=1×10⁻⁴), however there was no significant association with “extreme longevity” (at least one parent in the top 1% of ages at death). The A allele, here associated with increased parents lifespan, is associated with decreased risk of dementia [26]. Presence of the ApoE-ε4 was associated with reduced likelihood of having a parent in the top 1% of survival (OR=0.78, p=0.002), consistent with previous reports on mortality risk [27]; however we do not see a significant increase with presence of an ε2 allele (OR=1.09, p=0.37). We find that a variant on chromosome 5q33.3 previously associated with survival to age 90 [13] is associated (p=0.003) with combined parental age at death in a consistent direction (T allele associated with increased age at death), but not with extreme survival (p=0.6) in this analysis.

A recent study found that longevity-associated loci are enriched for disease-associated loci, consistent with our findings in Biobank [28]. The authors also performed a disease-weighted GWAS analysis, reporting 4 loci associated with exceptional longevity (≥90 years) with replication. We replicate the association between rs3184504 (mapped to SH2B3/ATXN2), previously shown to be associated with celiac disease [29], in our extreme longevity phenotype, but not the other three variants (although they are associated with continuous parent's age at death p<0.05). rs3184504 has also been associated with blood pressure and cardiovascular disease [30]. This suggests rs3184504 is associated with survival to exceptional ages, but the others require further evidence.

Previous analyses have found that long-lived individuals have lower low-density lipoprotein genetic risk than young controls [31,32], which we also observe (offspring of long-lived parents have lower LDL-increasing genetic risk score). Our results extend this work by showing associations also with triglyceride and HDL genetic risk scores, plus BMI and systolic blood pressure, as well as associations with non-lipid cardiovascular disease genetic risk traits.

Limitations

This study is limited to white British UK Biobank participants of Caucasian genetic descent, thus the results may not be applicable to other populations. We plan to address this in future collaborations. Evidence from GWAS studies identifying novel markers is strongest when associations are shown to replicate in independent samples, but unfortunately no large-scale replication resources are currently available. We have therefore been very cautious in reporting the novel markers, which await validation. However, the results for previously proven SNPs (i.e. those with a high prior probability) are likely to be very robust, given our large sample size.

UK Biobank is a volunteer study that did not aim for population representativeness at baseline, although efforts were made to recruit a heterogeneous sample by varying geographic placement of examination sites, including in economically deprived areas; the final response rate was 5.47% [33]. It has been reported that with sufficient variation in the phenotypes being studied, results are generalizable to the wider UK population [34].

We are limited by the coverage of the genotyping microarray utilized; ∼800,000 genetic variants were directly measured, allowing imputation of 73 million (9 million of which were common enough and high quality enough for inclusion in this GWAS). Many variants may exist outside of the data available for this study, in particular on the X, Y and mitochondrial chromosomes, for which imputed data are not available.

We have studied the normal range of parental ages at death, as well as extreme longevity (top 1% of survival). Some would argue that normal age at death is necessarily influenced by disease and has little relevance to aging. However, age is the major risk factor for most causes of death in later life, and the geroscience view sees aging as a major contributing factor to these diseases and deaths. Our extreme survival group is older than those studied in many of the previous GWAS of longevity [13,14], but did suffer from a somewhat limited sample size of n=1339: analyses with bigger samples of those attaining extreme ages are needed. Our longevity phenotypes do involve some repeat testing of the same individuals, but these phenotypes are closely related rather than being independent, and therefore should not have produced additional multiple testing problems.

We excluded early parental deaths based on an empirical approach [17,35] to avoid using arbitrary cutoffs. Only 13.6% of the fathers were aged between 46 and 60 at death (only 2% 46 to 50), suggesting that the exact cut-offs for premature deaths would have a modest effect on findings, especially as the results were similar for the extreme longevity group comparisons.

The available information about the parents is very limited, with no dates of birth for those who had died and no cause of death information. We have excluded UK Biobank participants aged less than 55 to avoid adding large numbers of parents who are likely to be too young to have reached the longer lifespans which are our main focus in this analysis. The lack of cause of death data may have resulted in an underestimation of effect sizes on lifespan due to aging and related disease, as we have had to include e.g. accidental deaths unrelated to aging. It is unlikely that e.g. Second World War related parental deaths could have biased results, as few combatants were aged over 45, our earliest age of death for inclusion the analyses. The consistency of known allele results between normal and extreme ages at death also suggests that our findings are not driven by early deaths.

When comparing the effect sizes between different GRS, it is important to consider that each GRS explains a different proportion of its associated phenotype: this will in turn affect the interpretation of the association with longevity. For instance, only approximately 1% of the variance in mean length of the telomeres is explained by the 7 reported SNPs [36], whereas more than 10% of the variance in the blood lipids is accounted for by associated variants [37]; therefore the strength of association between these two phenotypes and longevity may not be proportionately represented by the effect sizes of the GRS longevity associations.

Parental age at death and longevity phenotypes

Participants were asked the age at which their parents had died (or their current age if still alive). Analyses were performed separately on mother's age at death and father's age at death, and also on a combined phenotype. To reduce the effect of higher ages at death of mothers (compared to the fathers) we first z-transformed the mothers and fathers age at deaths before combining the z-scores into a single summed phenotype. Offspring of parents who died prematurely were excluded because the cause of death of the participant's parents was not asked, so we could not exclude accidental deaths explicitly. To determine the premature age at death cut-offs we used previously described methods to define the normal range of age at death for mothers and fathers separately, and excluded participants below these values (methods described here [17], description of method applied to UK Biobank here (Atkins et al. in review 2016)). In brief, non-linear least square regression models are used to fit a normal curve to the right-side of the distribution of parents age at death; the left half of the curve is then fitted and cut-offs determined (see [35] and [17]). The analysis identified the following cut-points for short, intermediate, and long-lived parents ages at death: for mothers (57 to 72 years, 73 to 92 years, and ≥93 years, respectively) and for fathers (46 to 65 years, 66 to 89 years, and ≥90 years, respectively). Parents below the “short” definition are excluded as premature deaths. We also defined a binary trait for an “extreme longevity” phenotype by determining the top 1% of age at death for mothers (≥98 years) and fathers (≥95 years) separately. We were able to increase the number of “long lived” parents in all the binary phenotypes by including participants who had responded to the question regarding the age of their parent if still alive. Participants with one long-lived and one short-lived parent were excluded from this analysis.

Finally, we defined two further binary traits for sensitivity analyses testing the consistency of results using a non-linear definition: offspring of at least one long-lived parent vs. offspring of two intermediate-lived parents, and offspring of at least one long-lived parent vs. offspring of at least one short-lived parent (very low numbers of participants with either “both long” or “both short” meant we used the aforementioned “at least one…” definitions). Figure Supplementary 1 contains a flow-chart showing how the phenotypes are derived and the numbers included in each analysis.

UK Biobank genetics data

We used genetic data available (autumn 2015) from 120,286 participants identified as ‘white British’ through self-report and verified through principal components analysis based on genotypes. Kinship coefficients were estimated and related individuals (3rd degree or higher) were removed to provide the maximal unrelated set of individuals. The central UK Biobank analysis team performed these analyses. Details of principal component analyses and kinship analyses can be found in the official UK Biobank genotyping document (http://biobank.ctsu.ox.ac.uk/crystal/docs/genotyping_qc.pdf; accessed 1st December 2015).

We used imputed genotypes available from the UK Biobank for association analyses. Briefly, phasing of individuals was carried out using SHAPEIT-2. Imputation was performed using IMPUTE2 and a combined 1000 Genomes / UK10K reference panel. Full details can be found in the official UK Biobank imputation document (http://biobank.ctsu.ox.ac.uk/crystal/docs/impute_ukb_v1.pdf; accessed 1st December 2015). After filtering for variants with MAF ≥1%, missingness <1.5%, imputation quality >0.4 and with Hardy-Weinberg equilibrium (HWE) P>1×10⁻⁶ within the white British participants, 9,658,292 imputed autosomal variants were eligible for the analyses.

We also utilized data directly from the microarrays for variants on the X (n=19,381) and Y (n=284) chromosomes, and on the mitochondrial genome (n=135), which were unavailable in the imputed dataset.

Within “white British” principal components

We selected 95,535 independent SNPs (pairwise r2<0.1) directly genotyped with a minor allele frequency (MAF) ≥ 2.5% and missingness <1.5% across all UK Biobank participants with genetic data available at the time of this study (n=152,732), and with HWE P>1×10-6 within the white British participants. Principal components were subsequently generated using FlashPCA and the first five adjusted for in all analyses [39].

Power calculations

Quanto software version 1.2.4 was utilized for power calculations with parameters: continuous, additive model, mean/SD of standardized outcome=0/1 [40].

Genome Wide Association Study

We used BOLT-LMM to model the associations between imputed variants (dosages) and each phenotype [41], which uses a linear mixed effects model approach. We looked at the results for variants with imputation quality >0.4, HWE p-values >1×10⁻⁶ and minor allele frequencies >0.1% in the white/British subset used for all analyses. For variants on the X, Y and mitochondrial chromosomes only in the directly-genotyped data we used Plink (v1.9) [42] in linear (additive) or binary (fisher) models, as appropriate, adjusted for the same covariates as above including the first 5 principal components from FlashPCA.

Independent testing of GWAS results

We obtained estimates of our genome wide significant results for fathers age at death GWAS in the Framingham Heart Study (FHS) generation 2 [43], with data on n=2033 offspring available. The inclusion criteria and model specification were the same as described for Biobank, with the exception that family structure was taken into account (using R package ‘pedigreem’).

Genetic heritability estimation

To estimate the variance in mean parental age at death (and age at death of mothers/fathers considered separately) explained by common genetic variants we utilized the BOLT-REML package [44]. This package uses Restricted Maximum Likelihood (REML) methods for variance component estimation, in this case on the genetics data. We used SNPs that met the following criteria in the 120,286 individuals; variants were excluded if not in Hardy-Weinberg Equilibrium (HWE) (P<1×10-6), or had a minor allele frequency <1%, or an overall missing rate >1.5% in any individual batch, or were on the Y chromosome. This resulted in 457,643 directly genotyped variants for inclusion. We also performed an additional sensitivity analysis relaxing the exclusion criteria to include all 845,997 variants. BOLT-REML determines the “heritability” of phenotypes based on the variance components of the genetics data provided. Prior to analyses, phenotypes are adjusted (by taking the residuals from a linear regression analysis) for confounding factors (age, sex, array type, assessment center, principal components 1-5).

Genetic Risk Score (GRS) creation

For each trait we identified the most recent GWAS meta-analysis and downloaded the results tables. We selected the SNPs identified at genome wide significance (p<5×10⁻⁸) and extracted the corresponding genotype information from the imputed data in the UK Biobank data. For all SNPs included in a GRS we checked for high imputation quality (>0.9), no significant deviation from HWE (p>1×10⁻⁶) and low missingness (<5% sample missing). We used the imputed genotype (dosage) information and effect size (from the meta-analysis) with Plink (v1.9) function `scores` to generate the weighted GRS for each trait in each participant [42]. The Plink function uses genotype coding for each participant as 0, 1 or 2 trait-increasing alleles, which is multiplied by the effect size (coefficient or odds ratio) from the published study. The resulting “weighted allele score” is summed for all variants associated with a particular trait. See the accompanying Supplementary Methods document for the definition of each genetic risk score, and Table Supplementary 1 for the variants and effect sizes used to create the scores.

Statistical analysis of Genetic Risk Scores

Generalized linear regression models were used throughout (R v3.2.0), with adjustment for age, sex, array type (‘axiom’ or ‘bileve’), assessment center (22 possible categories), and the first 5 genetic principal components. The regression linker functions `logit` or ‘Gaussian’ were used respectively for logistic and linear outcomes. Forest plots comparing the associations between different GRS’ and each outcome were generated using the R package ‘rmeta’ (v2.16). For these analyses the GRS was first z-transformed so that the statistics reported could be compared between each GRS on the forest plot. In the rest of the manuscript, we refer to the unstandardized, “per weighted allele” effects. In all respects except for the z-transformation the models were identical. Benjamini-Hochberg p-values are calculated to correct for potential false-positive associations (19 GRS against 4 primary longevity phenotypes).

ApoE haplotype definition

The ApoE haplotype is defined using two genetic variants (rs429358 and rs7412). We created a binary phenotype “any ε4 alleles vs. rest” where ε3/ε4 and ε4/ε4 participants were grouped together (ε2/ε4 participants were excluded), and the ε2/ε2, ε2/ε3 and ε3/ε3 alleles were the control participants. A second phenotype “any ε2 alleles vs. rest” is defined correspondingly.