Four Genome-Wide Association Studies Identify New Extreme Longevity Variants

Abstract

The search for the genetic determinants of extreme human longevity has been challenged by the phenotype’s rarity and its nonspecific definition by investigators. To address these issues, we established a consortium of four studies of extreme longevity that contributed 2,070 individuals who survived to the oldest one percentile of survival for the 1900 U.S. birth year cohort. We conducted various analyses to discover longevity-associated variants (LAV) and characterized those LAVs that differentiate survival to extreme age at death (eSAVs) from those LAVs that become more frequent in centenarians because of mortality selection (eg, survival to younger years). The analyses identified new rare variants in chromosomes 4 and 7 associated with extreme survival and with reduced risk for cardiovascular disease and Alzheimer’s disease. The results confirm the importance of studying truly rare survival to discover those combinations of common and rare variants associated with extreme longevity and longer health span.

Consistent with Fries’ “Compression of Morbidity” hypothesis (1), we have observed a progressive compression of the time centenarians of the New England Centenarian Study (NECS) experience both disability and morbidity into shorter periods prior to death (2,3), particularly for those who survive beyond the age of 105 years (4). Because of this compression, people at these oldest ages are also phenotypically more homogeneous in terms of age of onset or absence of age-related diseases and disability, thus suggesting that they have underlying biological mechanisms in common that confer such a propensity for healthy aging (5). Accompanying this progressive compression of morbidity are findings supportive of an increasing genetic influence upon survival to older and older ages beyond approximately the oldest one percentile of survival (6). These observations are the rationale behind the hypothesis that studying centenarians should lead to discovering genetic factors that promote extreme health-span (lifespan spent in good health) in addition to extreme individual lifespan (7).

However, the search for the genetic determinants of extreme human longevity through genome-wide association studies (GWAS) has been challenged by the rarity of the phenotype and the need for substantial sample sizes to overcome the severe correction for multiple testing (8). In 2010, the prevalence of centenarians was 1.73 per 10,000 people and the total number of U.S. supercentenarians (ages 110+ years) was reported as about 300 (9). Caution was noted for over-reporting and the prevalence reported by the Gerontology Research Group (GRG) of about 60 to 70 supercentenarians (http://www.grg.org/adams/tables.htm) is more likely. In fact, it has taken the New England Centenarian Study 20 years to accumulate 150 supercentenarians in its sample.

Notwithstanding the challenge of recruiting extreme-old individuals, we and others have taken the approach of studying a more homogeneous and more heritable phenotype (eg, within the oldest one percentile) instead of younger and phenotypically more heterogeneous and therefore larger samples that have failed to identify significant longevity-associated variants (LAVs) (10–12). Demonstrating just how more select and different nonagenarians are from centenarians, in an analysis of U.S. life table data, Broer and van Duijn noted that 10% of women born in 1959 are projected to live to age 90 years while just 0.3% or 33 times fewer will live to 100 years (13).

To build a large enough sample to discover GWAS-significant variants associated with living to and beyond the oldest one percentile of survival of the U.S. 1900 birth cohort, we created a consortium of four studies, the New England Centenarian Study (NECS) (14), the Long Life Family Study (LLFS) (15), the Southern Italian Centenarian Study (SICS) (16), and the Longenity Gene Project (LGP) (17). We then conducted three levels of analyses: (i) a standard meta-analysis of the GWAS results from the four studies to discover LAV, (ii) an analysis of the aggregated data from the four studies (or mega-analysis) to identify those LAVs with significant joint effects on longevity (jLAV), and (iii) a series of survival analyses to characterize those significant LAVs that differentiate survival to extreme age at death (eSAVs) from those LAVs that are more frequent in centenarians because of mortality selection. Figure 1 provides a schematic of our analytic approach.

Figure 1.

Schematic of the analyses conducted to discover longevity-associated variants (LAV), LAV with significant joint effect on extreme longevity (jLAV), and LAV with effects on extreme survival age (eSAV). Extreme longevity is defined as survival to at least the oldest one percentile of the 1900 birth year cohort life table, regardless of the attained age at death. Extreme survival age considers specifically age at death, conditioning on survival beyond the oldest one percentile of the 1900 birth year cohort. The top panels include description of the four studies and additional controls from the Illumina repository.

Methods

Study Populations and Genetic Data

This consortium includes four studies of longevity with genome-wide genotype data (Supplementary Figure 1) that we used to design nested case–control studies of extreme longevity. The studies and the selection of additional controls are described in Supplement Information.

Definition of Extreme Longevity Phenotype

We define extreme longevity as living past the age at which less than 1% individuals from the 1900 birth year cohort survived (age 96 and older for males and 100 and older for females) (18). Controls were defined as either male study participants who died at age <90 years, or female study participants who died at age <95, or study controls.

Statistical Analysis (Case–Control Studies)

Each study was analyzed independently using Bayesian logistic regression, adjusted for sex, genome-wide principal components 1–4, and using SNP dosage with additive coding. Random effects were used to model the within family correlation in NECS, LLFS, and LGP. Details are in Supplement Information.

Statistical Analysis (Survival Analysis)

Analysis of age at death and of disease onset in the aggregated data used Cox-proportional hazard regression, stratified by sex, and adjusted by principal components 1–4. Sandwich estimators were used to adjust for relatedness. Details are in Supplement Information.

eQTL Analysis

Boxplots displaying the normalized expression by genotypes were generated on the GTEX portal (http://gtexportal.org/).

Results

Participant Studies and Additional Controls

Characteristics of the study samples are in Supplementary Figure 1A. The studies differ in design, distribution of ages, and sample sizes. With the exception of the SICS, all studies enrolled probands, their siblings, and offspring. We designed a nested case–control study in each of the four cohorts with cases defined as participants who survived to an age older than the oldest one percentile of survival of the 1900 U.S. birth cohort (96 years or older for males and 100 years or older for females). This choice of extreme phenotype reduced the case set of each study substantially, and only the NECS provided more than 1,000 cases (Figure 1). Controls from each study were utilized in the analysis as described in Supplement Information. To increase statistical power of NECS and LLFS studies, we added controls selected from the Illumina repository. Genome wide principal components were used to select only white participants and characterize ethnicity of the participants of all four studies and the Illumina controls. This analysis confirmed that SICS and LGP were ethnically homogeneous studies, while LLFS and NECS included several European ethnicities (Supplementary Figure 1B and C).

Description of GWAS Results and Meta-Analysis

Approximately 6 million single nucleotide polymorphisms (SNPs) were selected for the analysis based on a conservative imputation quality score >0.9, minor allele frequency (MAF) >0.01, and no departure from Hardy–Weinberg equilibrium in controls (p > 10⁻⁶). GWASs were conducted in each study using mixed-effect logistic regression and the results were aggregated using meta-analysis. Both the Manhattan plot (Figure 2A) and QQ-plot (Supplementary Figure 1D) showed clean results and no apparent inflation of significant associations. Thirty-seven SNPs in chromosomes 7, 12, and 19 reached genome wide significance in the meta-analysis (p < 5E-8). The Manhattan plot in Figure 2A shows the three clusters of SNPs that reached genome-wide significance (LAVs) and additional clusters of 11 SNPs that reached 5E-7 level of significance. Table 1 lists the top 10 SNPs from the list of 48 with pairwise correlation <0.8. Several of the results were driven by the NECS data that was the largest study with also the oldest participants. Supplementary Table 1 includes the complete list of SNPs that reached 5E-7 level of significance. Genetic effects did not change substantially when Illumina controls were dropped from the analysis suggesting no bias. Of all the SNPs that we found to be genome-wide significant, only those within the APOE locus have been reported. The SNP rs2149954 that was described as a novel “longevity” variant in (10) did not replicate (p = .24). Established associations of SNPs in FOXO3A (11) did not reach 5E-7 level of significance. SNPs rs41266839, rs13217620 and rs10209741 in (12) did not replicate, while rs156033 reached nominal level of significance (p = .03).

Figure 2.

(A) Manhattan plot displaying the −log10(p-value) of SNP associations from the meta-analysis of the four GWAS of extreme longevity (y-axis). The x-axis displays the SNP positions by chromosomes that are in alternated colors. In addition to three loci that reached genome wide significance (horizontal bar in black, 5 × 10⁻⁸), 5 loci reached a 5 × 10⁻⁷ level of significance (horizontal bar in blue). (B–D) Regional plots of the locus on chromosome 7, 12, and 19. Data from the 1000 genomes were used to estimate linkage disequilibrium. Allele frequencies are summarized in Supplementary Figure 3C. (E–G) Functional annotation with expression data from GTEX. Rank-normalized expression by genotype. Plots were generated from the GTEX portal. Numbers on the x-axis are the genotype counts from GTEX.

Table 1.

SNPs Selected From the Meta-Analysis (p < 5E-7)

			Meta-Analysis		NECS			LGP			LLFS			SICS
SNP	Chr	LA/NLA	Beta	p	Beta	p	LAF Ce	Beta	p	LAF	Beta	p	LAF Ce/	Beta	p	LAF
			SE		SE		Co/	SE		Ce/	SE		Co/	SE		Ce/
			OR		OR		Ico	OR		Co	OR		Ico	OR		Co
			0.77		0.85		0.92/	0.58		0.91/	0.77		0.92/
rs6857	19	C/T	0.07	2.E-27	0.10	<1.E-30	0.84/	0.16	5.E-04		0.13	2.E-09	0.85/
			2.16		2.33		0.84	1.78		0.85	2.15		0.86
			0.87		0.95		0.95/	0.68		0.95/	1.04		0.96/	-0.02		0.96/
rs769449	19	G/A	0.09	1.E-23	0.12	5.E-15	0.88/	0.21	1.E-03		0.17	2.E-09	0.90/	0.40	1.E+00
			2.39		2.58		0.89	1.98		0.90	2.82		0.9	0.98		0.97
			0.43		0.52		0.87/	0.30		0.84/	0.43		0.86/	0.16		0.88/
rs59007384	19	G/T	0.05	5.E-15	0.08	3.E-10	0.80/	0.13	2.E-02		0.10	2.E-05	0.80/	0.22	5.E-01
			1.54		1.68		0.80	1.35		0.79	1.54		0.8	0.86		0.86
			0.5		0.75		0.14/	0.31		0.12/	0.24		0.09/	-0.19		0.09/
rs3764814	7	C/T	0.06	5.E-15	0.09	<1.E-30	0.11/	0.17	7.E-02		0.13	8.E-02	0.07/	0.24	4.E-01
			1.66		2.12		0.07	1.36		0.09	1.27		0.08	0.83		0.10
			0.25		0.18		0.69/	0.42		0.72/	0.36		0.72/	0.04		0.67/
rs7976168	12	A/G	0.04	4.E-09	0.06	3.E-03	0.65/	0.11	2.E-04		0.08	1.E-05	0.66/	0.15	8.E-01
			1.29		1.20		0.66	1.52		0.64	1.43		0.64	1.05		0.67
			0.2		0.20		0.58/	0.12		0.57/	0.28		0.58/	0.12		0.57/
rs405509	19	G/T	0.04	1.E-07	0.06	5.E-04	0.53/	0.10	2.E-01		0.07	1.E-04	0.53/	0.14	4.E-01
			1.23		1.22		0.52	1.13		0.54	1.32		0.49	1.12		0.54
			0.25		0.28		0.70/				0.20		0.68/	0.23		0.62/
rs7185374	16	A/C	0.05	2.E-07	0.06	1.E-05	0.64/				0.08	1.E-02	0.66/	0.15	1.E-01
			1.28		1.32		0.64				1.22		0.65	1.26		0.57
			0.73		0.57		0.02/	0.88		0.05/	0.6		0.02/	1.25		0.04/
rs28391193	4	A/G	0.14	2.E-07	0.23	1.E-02	0.01/	0.27	9.E-04		0.29	4.E-02	0.01/	0.47	7.E-03
			2.08		1.77		0.01	2.41		0.02	1.83		0.01	3.50		0.01
			0.22		0.23		0.47/				0.26		0.47/	0.06		0.43/
rs2008465	2	A/G	0.04	3.E-07	0.06	4.E-05	0.40/				0.08	6.E-04	0.43/	0.14	7.E-01
			1.25		1.25		0.42				1.30		0.41	1.06		0.42
			0.31		0.30		0.16/				0.40		0.17/	-0.02		0.10/
rs72834698	6	A/G	0.06	4.E-07	0.08	2.E-04	0.14/				0.10	1.E-04	0.12/	0.24	9.E-01
			1.36		1.35		0.13				1.49		0.13	0.98		0.10

Note: Beta = log-odds ratio for longevity for carriers of the longevity allele (LA) versus carrier of the nonlongevity allele (NLA). LAF: longevity allele frequency in centenarians (Ce), study controls (Co), and Illumina controls when used (Ico). “p Values” in the four genome-wide association studies were derived from the posterior distributions of the parameters. SNPs with poor quality score (<0.9) were not included in the analysis.

APOE Locus

Four of the most significant SNPs in Table 1 are in a locus on chromosome 19 that includes genes PVLR2, TOMM40, APOE, and APOC1 previously associated with extreme human longevity (11,19–23). The four SNPs demonstrated strong effects in NECS and LLFS (odds ratios [ORs] between 1.22 and 2.82), weaker effect in LGP and failed to reach statistical significance in the SICS. SNPs rs769449 and rs405509 have been recently characterized in the LGP using sequencing (24). Ninety-six percentage of carriers of the nonlongevity AA genotype of rs769449 in the LLFS cohort were homozygotes for the ε-4 allele of APOE that has high risk for Alzheimer’s disease, vascular disease and earlier mortality, thus suggesting that the increased frequency of the common allele of rs769449 in the long-lived individuals may be due to mortality selection. The pattern of linkage disequilibrium (LD) in the regional plot in Figure 2D suggests that the associations of the four SNPs may represent independent genetic effects. To test this hypothesis, a mixed-effect multivariable logistic regression model that included all four SNPs, in addition to sex and the first four principal components, was fitted to the data aggregated from the four studies. The analysis showed that only rs6857 and rs769449 remained strongly significant in the multivariable model (Supplementary Figure 3B and Supplementary Material). The SNPs were annotated by their potential role on gene transcription using the “Genotype Tissue Expression” (GTEX) portal (25) and gene regulation using the Regulome database from the ENCODE project (26). The SNP rs6857 is conserved among mammals and is a significant expression quantitative trait loci (eQTL) for APOE. The nonlongevity allele (T) is associated with decreased expression of the gene in whole blood (p = 0.005, Figure 2G). The SNP rs769449 is a significant eQTL for PVRL2 (p = 0.002), and the nonlongevity allele (A) is also associated with decreased gene expression in whole blood (Figure 2G). Additional annotation is in Supplementary Figure 2A and Supplementary Table 2.

Chromosome 7 Locus

The first new LAV discovered through the analysis was the SNP rs3764814: a synonymous variant in the 12th exome of “ubiquitin specific peptidase 42” (USP42). The longevity allele C was not common in controls of the NECS, LLFS, and LGP sample sets and the frequency matched the MAF = 0.07 in whites from dbSNP. The prevalence of the longevity allele increased by almost twofold in NECS and LGP centenarians (Table 1, Supplementary Figure 1G) but the effect was much smaller in the LLFS sample which was composed of less extreme ages and was not significant in SICS (also with younger ages than the NECS and LGP). This heterogeneity of effects was significant (p-value from heterogeneity Q test < 1E-4) and is highlighted in Supplementary Figure 1G that interestingly shows how the magnitude of ORs track ages at death of EL cases in the four studies. Figure 2B shows that the rs3764814 is not in strong LD with SNPs included in the analysis within a window of 20 Kb so that the isolated association was consistent with the data. This variant (or gene) has never been implicated with longevity, it has some evidence for binding (RegulomeDB score = 5), and is a significant eQTL for USP42 in various tissues (Figure 2E and Supplementary Figure 2B).

Chromosome 12 Locus

The second novel LAV was the SNP rs7976168: an intronic variant of the gene “transmembrane and tetratricopeptide repeat containing 2” (TMTC2) in chromosome 12. The longevity allele was the common allele of the SNP, and the frequency increased from 64–67% in controls to 69–72% in long-lived individuals (Table 1). The association was consistent in the NECS, LLFS, and LGP but was not evident in SICS. The regional plot in Figure 2C shows a large cluster of SNPs in strong LD in a window of 100 Kb that support the association. The SNP is an eQTL for TMTC2 in brain hippocampus (p = .0081), and homozygosity for the longevity allele is associated with decreased expression (Figure 2F). SNPs in this gene have been associated with risk for glaucoma (27).

Additional “Borderline” Associations

An additional four SNPs in chromosomes 2, 4, 6, and 16 reached extreme level of significance (p < 5E-7) although failed to reach genome-wide significance. The regional plots in Figure 3 show that all these associations are supported by SNPs in LD.

Figure 3.

Regional plots of the four SNPs in chromosomes 2, 4, 6, and 16, with p-values between 5E-7 and 5E-8. Data from the 1000 genomes were used to estimate linkage disequilibrium. Allele frequencies are summarized in Supplementary Figure 3C.

The longevity allele of rs7185374 in chromosome 16 (A) was common in all control sets and became more common in the extreme old (Table 1). The SNP is an eQTL for the “seven in absentia homolog” gene (SIAH1) which was associated with certain forms of Parkinson’s disease (28). The nonlongevity allele of the SNP is associated with decreased gene expression in whole blood and thyroid (Supplementary Figure 3C) and the SNP is in a region with weak evidence of binding.

SNP rs2008465 is in a region of chromosome 2 between “grainyhead like transcription factor 1” (GRHL1) and “Kruppel like factor 11” (KLF11). The longevity allele was the uncommon allele in controls and became more common in the extreme old (MAF = 0.44 in whites from dbSNP and 0.47 in NECS and LLFS centenarians). The SNP is an eQTL for GRHL1 in lung and skin tissues and the longevity allele is associated with lower expression in skin and higher expression in lung tissues (Supplementary Figure 3D).

The SNP rs72834698 is in the histone cluster in chromosome 6 and is a relatively uncommon variant (MAF = 0.13 in whites from dbSNP) that was more prevalent in centenarians (MAF = 0.17 in LLFS cases, and 0.16 in NECS cases). The longevity allele is associated with down-regulation of HIST1H2BD in whole blood (Supplementary Figure 3E).

The SNP rs28391193 in chromosome 4 is a rare variant in whites (MAF = 0.007 in whites from dbSNP) that became more prevalent in the extreme old, although the prevalence was still very low (MAF = 0.02 in LLFS and NECS centenarians, 0.05 in LGP centenarians, and 0.04 in SICS centenarians). This SNP is an eQTL for ELOVL6 in whole blood (Supplementary Figure 3F), a fatty acid elongase in the 4q25 previously associated with extreme longevity in NECS (29,30).

Search for jLAV

To identify the set of variants that were jointly associated with extreme longevity, we built a multivariable mixed effect logistic regression model that included all 10 SNPs in Table 1 and iteratively simplified the model to remove nonsignificant SNPs. The SNPs rs59007384 and rs405509 in the APOE locus failed to reach statistical significance (p < .005). Supplementary Figure 3B shows that the joint effects of the two remaining SNPs in the APOE locus became much smaller than their individual effects while the effects of the other SNPs were virtually unchanged.

Search for eSAV

The next step of the analysis was to search for the LAVs that were also associated with differential survival at extreme ages. We conducted a survival analysis using the case-set from the GWAS (N = 1,922, restricted to individuals with complete SNP data) and examined the association of one SNP at a time with the hazard for mortality, conditional on having survived to the 99th percentile of survival. Only five of the eight jLAV were significantly associated with extreme differential survival in the oldest old (p < .005, Figure 4A). The five SNPs were in chromosomes 4, 7, and 19 and carriers of the longevity alleles had significantly lower risk for mortality (hazard ratios between 0.66 and 0.87). For comparison, we conducted a similar survival analysis in all individuals for whom we had age information (N = 10,194) and Figure 4B shows that six of the eight SNPs had a significantly lower hazard for mortality associated with the longevity alleles compared to the nonlongevity allele, while the hazard for mortality associated with the remaining two SNPs was not significant (Supplementary Figure 4B). The results did not change when the associations were also stratified by birth year cohort.

Figure 4.

Forest plot summarizing the association of each individual SNP. Panel A: analysis restricted to case only; Panel B: analysis extended to all study participants. HR: hazard for mortality for carrier of the longevity variant (LA). All models were stratified by sex, adjusted for four principal components and family relatedness. The associations of rs7976168 and rs7185374 were not significant in the case-only analysis, as shown by the 95% confidence intervals including 1. Full details are in Supplement Figure 4. All associations reached statistical significance in the analysis extended to all participants (see Supplementary Figure 4 for details).

To help clarify whether the different results between the case-only analysis and all-subjects analysis were due to the different statistical powers of the two datasets, or due to different genetic effects, we generated Kaplan–Meier curves of the age at death, censored at last contact, stratified by the genotypes of the eight SNPS (extreme survivors only, Figure 5A, and all subjects, Figure 5B). The earlier decline in the survival curve for carriers of the nonlongevity allele of rs7976168 in chromosome 12, and rs7185374 in chromosome 16 suggested that those associations with extreme longevity in the original meta-GWAS were determined by higher mortality selection at earlier ages, so that the longevity allele became more prevalent in the oldest olds because carriers of the nonlongevity variants died at earlier ages. The longevity allele of SNP rs72834698 in chromosome 6 appeared to provide a survival advantage at earlier ages, but not in the extreme survivors group where the survival curves became undistinguishable (Supplementary Figure 5A). The effect of SNPs in chromosome 19 was unique: homozygosity for the nonlongevity allele of rs769449 had a dramatic effect on earlier mortality but carriers of one copy of the nonlongevity allele were also at higher risk for mortality in the extreme survivors group relative to noncarriers (Figure 5). The effect of rs6857 was less dramatic but persistent at any age. Only rs28391193 in chromosomes 4 (ELOVL6) and rs3764814 in 7 (USP42) provided a clear survival advantage at any age, even at the extreme ages, with a substantial separation of the survival curves (Figures 5). The plots also suggest that the C allele of rs3764814 had a recessive effect, while the A allele of rs28391193 has a dominant effect.

Figure 5.

Kaplan–Meier curves of survival to extreme old ages stratified by genotype. Panel A: Analysis restricted to case only; Panel B: Analysis extended to all study participants. In each plot, the least common genotype is colored in green/light grey, and the most common genotype in black. The longevity allele is the most common allele of rs6857, rs769449, rs7976168, and rs7185374, and the least common allele of rs3764814 and rs28391193. Similar plots for rs72834698 and rs2008465 are in Supplementary Figure 5.

We next generated a survival model that included all six significant SNPs using data from the cases (N = 1,922) and simplified the model by removing the SNPs that were not significantly associated with the hazard for mortality using a backward selection strategy. Only the three SNPs rs6857, rs3764814, and rs28391193 (Figure 6A) remained simultaneously significantly associated with differential survival at extreme ages. We tabulated the genetic profiles based on the genotypes of these three SNPs and the genetic models inferred from the one-SNP-at-a-time survival analysis, and displayed the Kaplan–Meier curves stratified by the most prevalent genetic profiles (Figure 6B). The most common genetic profiles carried by 1523 extreme survivors was characterized by the combination of all the common alleles of the three SNPs and included the longevity allele of rs6857 (genotype CC), and the common genotypes of rs3764814 (TT/TC) and rs28391193 (GG) that were not longevity alleles. Carriers of this genetic profile were characterized by the survival curve in black/solid line and represent the referent group in the oldest old. The second most common profile was carried by 280 extreme survivors and differed from the common genetic profile only in the genotype of rs6857 (CT/TT). Carriers of this genetic profile had higher risk for earlier mortality at extreme old age compared to the referent group (survival curve in green/dotted line). Two additional, rare genetic profiles were associated with a survival advantage even at extreme old ages (survival curves in blue/dashed-dotted line and red/dashed line) and were characterized by either carrying the longevity allele of rs3764814 (CC, only 33 subjects, survival curve in blue/dashed-dotted line), or carrying the longevity allele of rs28391193 (AG or AA, survival curve in red/dashed line). The complete list of genetic profiles based on the three SNPs genotype is in Supplementary Figure 6.

Figure 6.

Result of the joint survival analysis in case only (N = 1922). Panel A: forest plot displaying the three SNPs that remained significantly associated with the hazard for mortality when analyzed jointly. HR: hazard for mortality for carrier of the longevity variant (LA). The model was stratified by sex, and adjusted for four principal components and family relatedness. Full details are in Supplementary Figure 4. Panel B: Kaplan–Meier curves of survival to extreme old ages stratified by genetic profiles described in the legend. Longevity alleles: rs6857: C; rs769449: G; rs3764814: C; rs28391193: A; rs7185374: A. Line thickness represents the prevalence of the associated genetic profile.

The same analysis was conducted in the complete data set. Five SNPs remained significantly associated with the hazard for mortality in the joint model (Figure 7A). We tabulated the genetic profiles based on combination of genotypes of these five SNPs and the genetic models inferred earlier, and displayed the Kaplan–Meier curves for six genetic profiles (Figure 7B). The analysis in the larger data set confirmed the survival advantage of carriers of the longevity alleles of rs3764814 and rs28391193, and the damaging effects of the uncommon alleles of the other SNPs that increase the risk for earlier mortality. Interestingly, this analysis suggested a potential biological interaction between rs7185374 in chromosome 16 and SNPs in the APOE locus. Carriers of the genotypes CT/TT of rs6857 and GG of rs769449 had substantially different survival if they also carried the longevity allele A (genotype AA/AC) of rs7185374 (survival curve in green/dotted line) or the CC genotype (survival curve in cyan) indicating a buffering effect.

Figure 7.

Result of the joint survival analysis in all subjects in the dataset (N = 10,194). Panel A: forest plot displaying the five SNPs that remained significantly associated with the hazard for mortality when analyzed jointly. HR: hazard for mortality for carrier of the longevity variant (LA). As in (A), the model was stratified by sex, and adjusted for four principal components and family relatedness. Full details are in Supplementary Figure 4. Panel B: Kaplan–Meier curves of survival to extreme old ages stratified by genetic profiles described in the legend. Longevity alleles: rs6857: C; rs769449: G; rs3764814: C; rs28391193: A; rs7185374: A. Line thickness represents the prevalence of the associated genetic profile.

Genetic Variants Associated With Disease-Free Aging

We examined the associations of the 10 LAVs with delay of aging-related diseases in participants enrolled in NECS and LLFS for whom we had information about age of onset of cancer, cardio-vascular disease, type 2 diabetes, hypertension, and stroke. The data were collected using similar instruments in both studies (2) so that aggregated data could be analyzed. The analysis included on average 6,070 individuals comprising both proband and offspring generations in LLFS and NECS (Supplementary Figure 1A). The SNPs in the APOE locus were associated with risk for dementia and Alzheimer’s and not having the disease-associated variant was associated with a risk reduction as high as 63%. The longevity alleles of rs3764814, rs7185374, and rs2008465 were significantly associated with a reduced risk for CVD and hypertension. SNP rs28391193 was associated with reduced risk for CVD although the association failed to remain significant after correction for multiple comparisons.

Discussion

Summary of the Findings

We assembled a consortium of three centenarian studies and a familial longevity study that contributed more than 2,000 genome-wide genotyped samples of individuals who survived to ages within the oldest one percentile of survival for the 1900 U.S. birth cohort (age 95+ years for men and 100+ years for women). With this unique data set and a large set of controls, we conducted a variety of analyses to discover LAVs using a meta-analysis of study specific GWASs. We characterized the subset of these variants (8 of the 10 variants, with 2 of the 4 APOE variants dropping out) that have a joint effect on longevity and then we determined which of these variants have an effect on survival to extreme old ages, and which of these variants are more prevalent in centenarians because of mortality of carriers of the non-longevity variants. The analyses identified extreme longevity-promoting variants on chromosomes 4 (with almost genome-wide significance) and 7 (with genome-wide significance). We confirmed the association between established SNPs in the APOE locus and extreme longevity but showed that these associations are the result of mortality selection of the carriers of the nonlongevity variants. In the NECS and LLFS where we had age of onset data for age-related diseases, we found that the longevity alleles of rs3764814 (chromosome 7) and rs7185374 (chromosome 16) were significantly associated with reduced risks for CVD and hypertension and rs2008465 (chromosome 2) was associated with reduced risk for hypertension. We used publicly available functional data to show that several of these variants are eQTLs for known genes that could be new targets for healthy aging therapeutics.

The Role of the Apolipoprotein E Gene in Extreme Longevity

Our analysis confirmed the strong association of SNPs in the APOE locus (chromosome 19) with extreme longevity. However, the data provided clear evidence that the associations of SNPs in this locus with extreme longevity are the results of selection for earlier mortality for Alzheimer’s and vascular disease which are both associated with the ε-4 allele, and confirmed our earlier findings on the role of APOE and longevity (21). Analysis of the extreme survivors also showed that individuals who survived to old ages with only one copy of the nonlongevity allele (presumably the ε-4 allele) were still at risk for earlier mortality compared to those without this allele. Several publications have provided suggestive evidence supporting a role of the ε-2 allele of APOE in longevity (24), and the LLFS published data showing an association between the APOE ε-2 allele and human longevity (31). The common genotypes of rs6857 and rs769449 were associated with the combined ε-2 and ε-3 alleles in the LLFS set, but more conclusive data are needed to dissect the signal of the ε-2 allele from the other genotypes.

New Variants Associated With Extreme Longevity

With a large sample and extremely old subjects, our meta-analysis identified two new LAVs on chromosomes 7 and 12 with genome-wide significance and 4 new LAVs with nearly genome-wide significance. The association of rs7976168 (TMTC2) in chromosome 12 was marginally significant in NECS, LGP, and LLFS and only through a meta-analysis, genome-wide significance was achieved. This association was very robust and supported by a cluster of SNPs in linkage disequilibrium. The survival analyses showed that individuals who carried the nonlongevity allele had higher risk for Alzheimer’s disease (Table 2) and for early death (Figure 5). Higher mortality selection in carriers of the nonlongevity allele might have produced an over-representation of the “longevity allele” in centenarians that would not be causal, and the lack of an association between this SNP and survival to extreme ages in the case-only analysis confirmed that this LAV was not an eSAV. This characterization is important because it suggests that rs7976168 would not be a useful target for healthy-aging therapeutics. The association of rs2008465 (chromosome 2) led to a similar interpretation: the genotype GG of this SNP appeared to promote earlier mortality, possibly through cardiovascular disease (Table 2) but this variant was not an eSAV.

Table 2.

Results of the Associations of the 10 SNPs in Terms of Hazard Ratio (HR) With Onset of Aging-Related Diseases

	Chr	LA/NLA	HR Cancer	p	HR CVD	p	HR AD	p	HR T2D	p	HR HTN	p	HR CVA	p
rs6857	19	C/T	0.98	.700	0.99	.877	0.53	1.0E-06	0.91	.346	0.91	.023	0.96	.773
rs769449	19	G/A	1.05	.569	0.97	.666	0.37	0.0E+00	0.95	.672	0.9	.078	0.92	.596
rs59007384	19	G/T	1.07	.271	1.02	.682	0.69	6.0E-04	0.97	.683	0.95	.188	0.92	.446
rs3764814	7	C/T	0.96	.529	0.82	.001	0.83	6.2E-02	0.84	.159	0.84	.001	1.02	.866
rs7976168	12	A/G	0.97	.520	0.98	.558	0.83	2.4E-02	1.1	.223	1.03	.315	1	.954
rs405509	19	G/T	1	.965	0.99	.862	0.85	2.6E-02	1.05	.500	0.99	.637	1	.971
rs7185374	16	A/C	0.96	.388	0.9	.005	0.98	8.2E-01	0.94	.405	0.91	.003	0.9	.173
rs28391193	4	A/G	1.14	.409	0.65	.015	0.46	4.7E-02	1.33	.278	0.99	.953	1.45	.161
rs2008465	2	A/G	0.91	.032	0.91	.009	1.09	2.2E-01	0.98	.736	0.91	.004	0.98	.817
rs72834698	6	A/G	0.96	.444	0.98	.675	1.13	2.3E-01	0.97	.699	0.95	.222	0.95	.649

Note: Cancer: all type excluding skin cancer; CVD: angina, congestive heart failure, atrial fibrillation, myocardial infarction; AD: Alzheimer’s disease; T2D: type 2 diabetes; HTN: hypertension; CVA: Stroke or TIA. Highlighted in boldface are the associations that remain significant after correction for multiple comparisons. All associations were estimated using Cox proportional hazard regression, stratified by sex and adjusted for the four principal components.

More promising to the field of longevity are the associations of the uncommon SNPs rs3764814 (chromosome 7), rs28391193 (chromosome 4), and rs72834698 (chromosome 6) that appear to be specific longevity promoting variants. The association of rs3764814 with extreme longevity was driven by the dominant effect of the C allele in NECS, and was probably missed in our earlier analysis that used a much smaller number of cases (801 rather than 1088), and was less enriched for extremely old individuals than this sample, and of Ashkenazi Jewish descents (21). The association was also likely missed in other studies of longevity that did not meet our more select definition of extreme survival (10,11,19). The gene harboring this variant (USP42) is a possible transcription regulator, it can target p53 and contribute to stabilization of p53 in response to stress (32). While the HaploReg software search showed no evidence of regulatory features for this particular SNP, the data from GTEX showed that rs3764814 is an eQTL for USP42 in a variety of tissues. The longevity genotype CC is not common in the general population and more data are needed to better understand the mechanism of action linking the genotype to the phenotype.

Although the association of rs28391193 in chromosome 4 did not reach genome-wide significance, the consistent effects observed in all four centenarian studies made the association a likely true positive finding and worth further follow-up studies. This variant is an eQTL for ELOVL6, a fatty acid elongase in 4q25, a region associated with extreme human longevity by linkage analysis (29). We replicated this linkage peak using a larger set of sib pairs (33). The longevity allele of rs28391193 is associated with lower expression of ELOVL6. Reduced expression of this gene in mice was linked to high levels of palmitoleic acid determining insulin sensitivity despite the animals being subjected to a high-fat diet (34). The important role of ELOVL6 in insulin sensitivity, cardiovascular disease and healthy aging was recently confirmed in humans (35). Furthermore, palmitoleic acid was shown to be higher in centenarians (36) and in long-living worms (37).

The associations of rs72834698 in chromosome 6 was only supported by data in NECS and LLFS samples and was consistent with our earlier findings of the genetics of extreme human longevity through GWAS and sequencing that identified a cluster of significant SNPs in the MHC complex (21,38).

We also noted that the less common genotype CC of SNP rs7185374 in chromosome 16 appeared to be associated with exceptional longevity but the survival analysis showed that this was likely due to its association with decreased mortality at earlier ages. The joint analysis showed that this SNP remained associated with survival even when all other variants were included in the model. Most intriguingly, the analysis of genetic profiles in Figure 7B suggests that the longevity genotype AA and AC of this SNP might reduce the damaging effect of the deleterious variants of APOE.

The Important Distinction Between LAVs and eSAVs

The traditional method of analysis based on aggregating the GWAS results of the individual studies through a meta-analysis was sufficiently powered to detect new LAVs with genome-wide significance, and new LAVs with suggestive levels of significance. However, the additional analysis of age at death in the aggregated data was critical to differentiate between LAVs that appeared to be more frequent in the oldest old because of mortality selection at earlier ages from those LAVs that appeared to confer a real survival advantage. This distinction is important to select the best therapeutic targets for future work, but it is also important to better understand the genetic model underlying extreme longevity.

Many investigators have conjectured that escaping disease through a depletion of “bad” variants is what makes people survive long (39). To date, the only published example that supports this hypothesis is the ε-4 allele of APOE and our analysis showed that surviving to old age with the ε₄ allele of APOE is indeed very rare. The prevalence of the ε-4 allele of APOE is only about 5% in whites, therefore the majority of individuals do not carry this variant. So, common sense suggests that depletion of this disease-associated and rare genotype plays a nonessential role in extreme longevity. In general, earlier mortality of the carriers of the disease allele of a SNP will lead to an over-representation of the nondisease allele in the survivors, and the nondisease allele will appear to be associated with EL in a case–control study design even when it is neutral to longevity. Our analysis, which distinguishes eSAVs from LAVs, attempts to address this problem by searching for LAVs that are associated with varying age at death at the extremes of human lifespan. In addition, we and others have shown that centenarians appear to carry many disease-associated variants (21,38,40–42) and argued that a possible explanation of this phenomenon is the presence of protective variants that buffer the effect of deleterious variants (43). We showed that combinations of the genotypes of 281 SNPs produced genetic signatures of extreme longevity that correlated with different health-spans and lifespans (21). About 50% of the SNPs were replicated in a meta-analysis that includes three of the four studies included here (44). The limitation of that work was the smaller sample size compared to the sample in this meta-analysis, thus likely explaining why we did not find in that study more SNPs beyond those in the APOE locus that reached genome-wide significance levels.

With a much larger sample size, the current analyses produced results with genome-wide levels of significance. By focusing on the set of variants that are jointly associated with extreme longevity defined as either a dichotomous trait or age at death, we could build an initial set of genetic profiles that were easy to interpret and visualize. We also showed that some rare profiles made by combinations of common and rare genotypes are associated with more extreme lifespan. The data and survival analysis provide support for the hypothesis that the genetic makeup of extreme longevity is based on a combination of common and rare variants, with common variants that create the background to survive to relatively common old ages, and specific combinations of uncommon and rare variants that add an additional survival advantage to even older ages. Our analysis showed that LAVs discovered through a case–control study are not necessarily the variants that make someone live to extreme old age, and additional survival analysis is needed to characterize and distinguish eSAV from the set of LAVs. This analysis establishes a framework to distinguish between longevity variants that may confer an advantage to survive to extreme old ages, and those longevity variants that are more common in centenarians because of earlier mortality selection.

Rare Variants

Although the yield of discovery is more substantial than any GWAS of extreme longevity published so far, it is still disappointing. The two most significant longevity genotypes (genotype CC of rs3764814 in USP42, chromosome 7, and genotype AA/AG of rs28391193, ELOVL6, chromosome 4) are carried by a very small proportion of the cases included in the analysis and much of the variability around extreme lifespan remains unexplained. In this analysis, we decided to be extremely conservative and included only SNPs with imputation quality score of 0.9 and higher. This selection led to ignoring uncommon and rare variants that are challenging to impute with good quality. We expect that many more uncommon variants remain to be discovered through sequencing of centenarian samples. Larger sample sizes are needed to detect association of rare variants such as rs28391193 (45) and therefore promising associations that miss the threshold for genome-wide significance are important to discuss (12).

Meta- Versus Mega-Analysis

In this article, we used meta-analysis to aggregate the results from four GWAS of EL. Unlike a meta-analysis that combines study results, a mega-analysis combines data from each study in one dataset. This approach has been shown to increase the power to detect associations of uncommon/rare variants by aggregating small genotype counts from the different studies (46–48), and provides a means to discover associations of variants with ethnic specific effects that can be lost in a meta-analyses of studies with different ethnicities (48,49). We limited analysis of data aggregated from the four longevity studies to the most significant associations detected with the meta-analysis. In future work, we will extend this approach to a genome-wide analysis.

Limitations

A limitation of the current analysis was the choice of controls. Our study controls were not matched by birth year cohort to cases, and most of them are still alive and may include individuals who might eventually become centenarians although we expect this number to be small given the low prevalence of centenarians in the population. While possible differences in the genetic background were adjusted for by using genome-wide principal components, the different birth year cohorts of cases and controls may introduce nongenetic confounders. The survival analysis restricted to case only did not suffer from this limitation but additional nonmeasured confounders could influence life-span even at the extreme ages, including for example use of medications. Our analyses were not adjusted for medications because in two of the four studies and in all additional controls this information is not available. More informative data about medications and other nongenetic risk factors are needed to completely characterize the mechanisms that link genotypes to EL.

We were able to annotate the potential role of common variants on gene transcription and regulation using the GTEX portal (25) and the Regulome database from the ENCODE project (26) but the analyses showed the limitation of these data for characterizing rare variants associated with extreme human longevity. Additional functional data corresponding to rare variants that are more prevalent in centenarians will be needed to translate the findings from genetic studies into more actionable targets. In addition, it will be important to move from associations between longevity SNPs and gene expression to association between genetic profiles of extreme longevity and molecular signatures based on gene expression profiles in order to characterize the biological mechanisms that help people remain healthy as they age.

Supplementary Material

Supplementary data is available at The Journals of Gerontology, Series A: Biological Sciences and Medical Sciences online.

Funding

This work was supported by the National Institute on Aging (NIA cooperative agreements U01-AG023755, U19-AG023122, P30AG038072, R01AG046949), the National Heart Lung Blood Institute (R21HL114237), and the William Wood Foundation.