APOE Alleles and Extreme Human Longevity

Abstract

We assembled a collection of 28,297 participants from seven studies of longevity and healthy aging comprising New England Centenarian, Long Life Family, Longevity Gene Population, Southern Italian Centenarian, Japanese Centenarian, the Danish Longevity, and the Health and Retirement Studies to investigate the association between the APOE alleles ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ and extreme human longevity and age at death. By using three different genetic models and two definitions of extreme longevity based on either a threshold model or age at death, we show that ${\epsilon }_4$ is associated with a substantially decreased odds for extreme longevity, and increased risk for death that persists even beyond ages reached by less than 1% of the population. We also show that carrying the ${\epsilon }_2{\epsilon }_2$ or ${\epsilon }_2{\epsilon }_3$ genotype is associated with significantly increased odds to reach extreme longevity, with decreased risk for death compared with carrying the genotype ${\epsilon }_3{\epsilon }_3$ but with only a modest reduction in risk for death beyond an age reached by less than 1% of the population.

APOE, Extreme human longevity, Genetic association, Survival distribution

Consistent with Fries’ compression of morbidity hypothesis (1), we and others have observed a progressive compression of the time that centenarians experience both disability and morbidity, particularly for survival beyond 105 years, or the oldest 0.1 percentile of survival (2–6). Our work on heritability of extreme longevity (EL) has also demonstrated that while the vast majority of variation in survival around average life expectancy is explained by environmental factors and health-related behaviors (7,8), genetic variants are likely to be the major predisposing factors to EL, especially beyond ages 105+ years (9,10). Therefore, studying centenarians’ genomes should help discover those genetic factors that extend human lifespan and enable people to remain healthy as they age.

Apolipoprotein E (APOE) emerged as a candidate gene in longevity since Schachter et al. showed that French centenarians have a very low frequency of the allele ${\epsilon }_4$ which is also associated with increased risk for Alzheimer’s disease and vascular disease (11–,13). In addition, they found an increased frequency of the allele ${\epsilon }_2$ in centenarians (13). Three major APOE alleles in the population, known as ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ are defined by combinations of genotypes of the single nucleotide polymorphisms (SNPs) rs7412 and rs429358 (see reference (14) and Supplementary Figure 1). Corbo et al. (15) showed that the frequencies of these three alleles vary substantially with ethnicity, and that the prevalence of the ${\epsilon }_4$ allele—the ancestral allele—ranges between 41% in African Pygmies and 31% in European Lapps, 5% in Sardinians, and 10% in Japanese. The ${\epsilon }_3$ allele is the most common, with prevalence ranging between 90% in Southern Italians and Sardinians and 54% in African Pygmies. The ${\epsilon }_2$ allele is the least common; it is almost absent in Native Americans and the prevalence ranges between 3% and 12% in African populations and Europeans and reaches 14% in some Asian ethnicities.

The allele ${\epsilon }_4$ is a well-established genetic risk factor for Alzheimer’s disease (16) that notwithstanding its detrimental effects still persists in the population (17), whereas the role of the ${\epsilon }_2$ allele in longevity and healthy aging traits remains ambiguous (18). Underpowered studies have produced conflicting results claiming a negative association between the ${\epsilon }_2$ allele and longevity in Danish and Finnish centenarians (19,20), a neutral association in Chinese, Korean, and Finnish centenarians (21–23), and a positive association in Japanese, French, Spanish, northern and southern Italian centenarians, Ashkenazi Jews, and Greek nonagenarians (13,24–28). The meta-analysis in reference (29) pooled some of the results to estimate the odds ratio (OR) for longevity comparing carriers of the ε₂ε₂ genotype to carriers of ε₃ε₃ and reached the conclusion of no significant effect of ε₂ε₂ on human longevity. Our recent analysis of a subset of ethnically homogeneous studies showed a marginally significant association of ε₂ε₂ with human longevity (OR = 2.39, 95% confidence interval [CI]: 0.99, 5.76) (14). The common limitation to all these studies is the small sample size and, particularly, the small number of ε₂ε₂ genotypes; the largest collection of an ethnically homogenous sample thus far has been only 30 such cases in our meta-analysis (14).

In this study, we use genotype data of APOE alleles in 28,297 subjects from seven studies of healthy aging and EL that include 4,139 individuals that reached ages achieved by less than 1% of their birth year cohort, and 159 carriers of the ε₂ε₂ genotype to study the role of the APOE alleles in survival to extreme old age.

Results

We aggregated data from three centenarian studies of European ethnicity and a study of familial longevity (New England Centenarian Study (30), Long Life Family Study (31), Southern Italian Centenarian Study (32), and Longevity Gene Population Study (33)) that agreed to share individual-level data to generate a large discovery data set. We used three additional studies from the United States (the Health and Retirement Study [HRS] (34)), Denmark (the Danish Longevity Study (35)), and Japan (the Japanese Centenarian Study (6)) for replication. Characteristics of the 28,297 subjects assembled for this analysis are summarized in Table 1. The discovery set comprised 12,881 subjects from four different studies of longevity and approximately 60% of them were female. Ages at last contact of the participants ranged between 0 years for controls with DNA samples extracted from umbilical cords and 119 years for cases, with birth year between 1880 and 1975. The replication studies had similar demographic characteristics.

Table 1.

Demographic Characteristics of the 28,297 Study Participants

Study	Group	N	Birth Year	Sex (%M)	Age
Discovery	Cases	2,144	1904 (1880–1921)	42	102 (96–119)
	Controls	6,984	1953 (1904–1975)	44	53 (0–99)
	All	12,881	1911 (1896–1955)	43	72 (0–119)
JCS	Cases	790	1899 (1887–1907)	16	105 (100–115)
	Controls	967	//	61	36 (20–74)
HRS	Cases	85	1914 (1905–1918)	70	98 (96–103)
	Controls	10,305	1938 (1909–1981)	41	75 (33–99)
Danish	Cases	1,120	1910 (1895–1918)	45	101 (96–110)
	Controls	2,149	1941 (1902–1952)	43	75 (47–100)

Notes: Shown are median and range of birth year and age and percentages of male sex. Ages are either at last contact for alive participants or at death. The discovery set includes participants from Longevity Gene Population (N = 1,855), Long Life Family Study (N = 4,493), New England Centenarian (N = 1,977), Southern Italian Centenarian Study (N = 963), and 3593 additional controls. HRS = Health and Retirement Study; JCS = Japanese Centenarian Study.

Table 2 describes the distribution of the genotype data for APOE in the various studies. Notably, the frequency of ${\epsilon }_4{\epsilon }_4$ was substantially lower in cases than in controls. The frequency of ${\epsilon }_2{\epsilon }_2$ did not show substantial variation between cases and controls, whereas the ${\epsilon }_2{\epsilon }_3$ genotype appeared to be more common in long-lived individuals than controls.

Table 2.

Genotype and Allele Frequencies of APOE Variants

		Genotype Frequencies						Allele Frequencies
		${\epsilon }_2{\epsilon }_2$	${\epsilon }_2{\epsilon }_3$	${\epsilon }_2{\epsilon }_4$	${\epsilon }_3{\epsilon }_3$	${\epsilon }_3{\epsilon }_4$	${\epsilon }_4{\epsilon }_4$	${\epsilon }_2$	${\epsilon }_3$	${\epsilon }_4$
Discovery
	Cases (N = 2,144)	16 (0.01)	420 (0.19)	36 (0.02)	1,449 (0.68)	217 (0.10)	6 (0.003)	488 (0.11)	3535 (0.83)	265 (0.06)
	Controls (N = 6,984)	40 (0.01)	942 (0.13)	147 (0.02)	4,455 (0.64)	1,280 (0.19)	120 (0.02)	1,169 (0.08)	11,132 (0.80)	1,667 (0.12)
	All (N = 12,881)	83 (0.01)	1898 (0.15)	248 (0.02)	8389 (0.65)	2105 (0.16)	158 (0.01)	2,312 (0.09)	20,781 (0.81)	2,669 (0.10)
Replication
JCS	Cases (N =790)	1 (0.001)	109 (0.14)	2 (0.003)	611 (0.77)	63 (0.08)	4 (0.005)	113 (0.07)	1,394 (0.88)	73 (0.05)
	Controls (N = 967)	1 (0.001)	77 (0.08)	14 (0.01)	683 (0.71)	185 (0.19)	7 (0.007)	93 (0.05)	1,628 (0.84)	213 (0.11)
Danish	Cases (N =1,120)	6 (0.005)	195 (0.17)	34 (0.03)	708 (0.63)	170 (0.15)	7 (0.006)	241 (0.11)	1,781 (0.79)	218 (0.10)
	Controls (N = 2149)	12 (0.006)	290 (0.13)	62 (0.03)	1191 (0.55)	529 (0.25)	65 (0.03)	376 (0.09)	3,201 (0.74)	721 (0.17)
HRS	Cases (N = 85)	0 (0)	16 (0.19)	0 (0)	58 (0.68)	11 (0.13)	0 (0)	16 (0.09)	143 (0.84)	11 (0.06)
	Controls (N = 10,305)	56 (0.01)	1253 (0.12)	207 (0.02)	6381 (0.62)	2215 (0.21)	193 (0.02)	1,572 (0.08)	16,230 (0.79)	2,808 (0.14)

Notes: Allele frequencies were computed by summing the number of alleles ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ in each genotype. Numbers in parentheses are frequencies. HRS = Health and Retirement Study; JCS = Japanese Centenarian Study.

Genotype Model

Logistic regression using the genotypic model showed significant negative association of the ${\epsilon }_3{\epsilon }_{4 }$ and ${\epsilon }_4{\epsilon }_4$ genotypes with EL relative to ${\epsilon }_3{\epsilon }_3$ in the discovery set (Supplementary Table 1 and Supplementary Figure 3). Compared to carriers of ${\epsilon }_3{\epsilon }_3$ carriers of ${\epsilon }_4{\epsilon }_4$ had an OR for EL of 0.17 (95% CI: 0.06, 0.35), whereas carriers of ${\epsilon }_3{\epsilon }_4$ had an OR for EL of 0.51 (95% CI: 0.44, 0.60). The negative association of ${\epsilon }_4{\epsilon }_4$ with EL was replicated in the Danish study (OR = 0.18, 95% CI: 0.08, 0.39) and was consistent in the Japanese study, but did not reach statistical significance (OR = 0.45, 95% CI: 0.12, 1.66) as shown by the 95% CI that includes 1. The HRS data could not be robustly analyzed with the genotype model because of the lack of cases carrying the ${\epsilon }_4{\epsilon }_4$${\epsilon }_2{\epsilon }_4$ or ${\epsilon }_2{\epsilon }_2$ genotype. Meta-analysis of the results from the three studies (discovery, Japanese, and Danish sets) estimated an 81% reduction in the odds for EL comparing carriers of ${\epsilon }_4{\epsilon }_4$ to ${\epsilon }_3{\epsilon }_3$ (OR = 0.19, 95% CI: 0.11, 0.33), and a 50% reduced odds for EL comparing carriers of ${\epsilon }_3{\epsilon }_4$ to ${\epsilon }_3{\epsilon }_3$ (OR = 0.50, 95% CI: 0.45, 0.56; Table 3 and Supplementary Figure 3). The association between ${\epsilon }_2{\epsilon }_4$ and EL was also negative although it did not reach statistical significance (OR = 0.76, 95% CI: 0.57, 1.01).

Table 3.

Results of Meta-analysis of Genetic Associations

Genotype	OR (95% CI)	HR (95% CI)	HR (95% CI) in ES
	Genotypic Model
${\epsilon }_2{\epsilon }_2$	0.99 (0.58, 1.71)	0.95 (0.67, 1.35)	0.76 (0.48, 1.19)
${\epsilon }_2{\epsilon }_3$	1.33 (1.19, 1.48)	0.94 (0.87, 1.01)	1.00 (0.91, 1.10)
${\epsilon }_2{\epsilon }_4$	0.76 (0.57, 1.01)	1.27 (1.05; 1.53)	1.43 (0.96, 2.13)
${\epsilon }_3{\epsilon }_3$	1	1	1
${\epsilon }_3{\epsilon }_4$	0.50 (0.45, 0.56)	1.19 (1.05, 1.28)	1.19 (1.05, 1.33)
${\epsilon }_4{\epsilon }_4$	0.19 (0.11, 0.33)	1.86 (1.48, 2.32)	1.89 (1.24, 2.86)
	Additive Model
${\epsilon }_3{\epsilon }_3$	1	1	1
${\epsilon }_2$	1.28 (1.16, 1.41)	0.94 (0.88, 1.01)	0.98 (0.9, 1.06)
	Genotype Group Model
$E_2$	1.32 (1.19, 1.47)	0.94 (0.87, 1.01)	0.98 (0.90, 1.07)
$E_3$	1	1	1
$E_4$	0.50 (0.45, 0.56)	1.21 (1.12, 1.30)	1.21 (1.10, 1.35)

Notes: Genotypic model: The first column reports the odds ratio (OR) for extreme longevity (EL) and 95% confidence interval (CI) for carriers of the genotype in the row relative to the referent genotype ${\epsilon }_3{\epsilon }_3$ The second column reports the hazard ratio (HR) for mortality and 95% CI comparing the same groups. The last column reports HR for mortality and 95% CI in the extreme survivors (ES) defined as the subset of study participants who survived beyond the age reached by less than 1% of the population. Additive model: OR, HRs, and 95% CI for carriers of one copy of the ${\epsilon }_2$ allele. The analysis was restricted to carriers of ${\epsilon }_2{\epsilon }_2,{\epsilon }_2{\epsilon }_3, {\epsilon }_3{\epsilon }_3$Genotype Group model: OR, HRs, and 95% CI for carriers of either $E_2={\epsilon }_2{\epsilon }_2,{\epsilon }_2{\epsilon }_3$ or $E_4={\epsilon }_2{\epsilon }_4,{\epsilon }_3{\epsilon }_4,{\epsilon }_4{\epsilon }_4$ versus $E_3={\epsilon }_3{\epsilon }_3$ Bold-face denotes statistically significant results as shown by 95% CI not including 1.

Carriers of ${\epsilon }_2{\epsilon }_{3 }$ had increased odds for EL compared with carriers of ${\epsilon }_3{\epsilon }_3$ in the discovery set, the Japanese and Danish studies, and in the meta-analysis of the results the genotype ${\epsilon }_2{\epsilon }_3$ was associated with significantly increased odds for longevity (OR = 1.33, 95% CI: 1.19, 1.48). The association of ${\epsilon }_2{\epsilon }_{2 }$ was inconsistent across studies, and the meta-analysis produced a nonsignificant result (OR = 0.99, 95% CI: 0.58, 1.71). To increase statistical power, we added to this meta-analysis the results of the associations of ${\epsilon }_2{\epsilon }_{2 }, {\epsilon }_2{\epsilon }_{3 }$ and ${\epsilon }_3{\epsilon }_{3 }$ with EL from an additional meta-analysis of six studies of centenarians described in references (14,24–26,29,36). The data included 1,001 centenarians and 2,772 controls and showed a twofold increase in the odds for EL comparing ${\epsilon }_2{\epsilon }_{2 }$ and ${\epsilon }_2{\epsilon }_{3 }$ carriers (Supplementary Figures 4 and 5). Figure 1 summarizes the results of the extended meta-analysis and shows increased odds for EL comparing ${\epsilon }_2{\epsilon }_{2 }$ to ${\epsilon }_3{\epsilon }_{3 }$ although the association did not reach statistical significance (OR = 1.26, 95% CI: 0.80, 1.99), and statistically significant increased odds for EL in carriers of ${\epsilon }_2{\epsilon }_{3 }$ compared with ${\epsilon }_3{\epsilon }_{3 }$ (OR = 1.34, 95% CI: 1.21, 1.47). Interestingly, the OR for EL associated with ${\epsilon }_2{\epsilon }_{2 }$ and ${\epsilon }_2{\epsilon }_{3 }$ in this larger meta-analysis were similar, and the meta-analysis of the results comparing the odds for EL between ${\epsilon }_2{\epsilon }_{2 }$ and ${\epsilon }_2{\epsilon }_{3 }$ showed no significant difference (Supplementary Figure 5). Note that the sample size of the meta-analysis including all these different studies provides 72% power to detect a significant OR = 1.25, assuming a prevalence of EL of about 1% in carriers of ${\epsilon }_2{\epsilon }_{2 }.$ Therefore, the lack of statistical significance is not attributable to low statistical power, assuming that there is no significant heterogeneity of study-specific effects. We formally tested for heterogeneity of effects among the studies using Cochran’s Q statistic and none of the tests was statistically significant (Figure 1 and Supplementary Figure 3).

Figure 1.

Forest plots of odds ratio (OR) for extreme longevity estimated using the genotypic model. Meta_2017 denotes results from the meta-analysis published in (14). ${\chi }_n^2$ denotes Cochran’s Q test of heterogeneity with n degrees of freedom and p is the p value to test the null hypothesis of no heterogeneity across studies (a nonsignificant result for this test suggests that meta-analysis is appropriate). CI = confidence interval.

The meta-analysis of the association between APOE genotypes and age at death did not include the Japanese study, in which age at last contact of controls was not available. This analysis showed that any genotype including ${\epsilon }_4$ was associated with an increased risk for death (Table 3). The hazard ratio (HR) for mortality for carriers of ${\epsilon }_4{\epsilon }_4$ compared with ${\epsilon }_3{\epsilon }_{3 }$ was significantly greater than 1 (HR = 1.86, 95% CI: 1.48, 2.32), and similarly the HRs for mortality of carriers of ${\epsilon }_3{\epsilon }_4$ and of ${\epsilon }_2{\epsilon }_4$ compared with ${\epsilon }_3{\epsilon }_{3 }$ were significantly greater than 1 ( ${\epsilon }_3{\epsilon }_4$ vs ${\epsilon }_3{\epsilon }_{3 }:$ HR = 1.19, 95% CI: 1.05, 1.28, and ${\epsilon }_2{\epsilon }_4$ vs ${\epsilon }_3{\epsilon }_{3 }:$ HR = 1.27, 95% CI: 1.05, 1.53). In this meta-analysis, carriers of ${\epsilon }_2{\epsilon }_{2 }$ or ${\epsilon }_2{\epsilon }_{3 }$ had a similar, slightly reduced risk for mortality compared with carriers of ${\epsilon }_3{\epsilon }_{3 }$ but the associations did not reach statistical significance.

The deleterious effect of ${\epsilon }_4{\epsilon }_4$${\epsilon }_3{\epsilon }_4$ and ${\epsilon }_2{\epsilon }_4$ persisted in the analysis of survivors past the 1% survival age, and only the effect of ${\epsilon }_2{\epsilon }_4$ did not reach statistical significance ${\epsilon }_4{\epsilon }_4$ vs ${\epsilon }_3{\epsilon }_{3 }:$ HR = 1.89, 95% CI: 1.24, 2.86; ${\epsilon }_3{\epsilon }_4$ vs ${\epsilon }_3{\epsilon }_{3 }:$ HR = 1.19, 95% CI: 1.05, 1.33; ${\epsilon }_2{\epsilon }_4$ vs ${\epsilon }_3{\epsilon }_{3 }:$ HR = 1.43, 95% CI: 0.96, 2.13), whereas the associations of ${\epsilon }_2{\epsilon }_2$ and ${\epsilon }_2{\epsilon }_3$ with risk for mortality did not reach statistical significance (Table 3). Note that the sample size of this analysis provided more than 80% power to detect HR equal to 0.75 so that the lack of statistical significance of the effect of ${\epsilon }_2{\epsilon }_{2 }$ relative to ${\epsilon }_3{\epsilon }_{3 }$ should not be due to the lack of statistical power. We also tested for heterogeneity of the effects among the studies using Cochran’s Q statistic, and none of the tests was statistically significant (Supplementary Figure 3).

We subsequently ran two additional sets of analyses, one based on the additive genetic model and one based on the genotype group model for larger statistical power.

Additive Association

In this analysis, we excluded all carriers of the ${\epsilon }_4$ allele ${\epsilon }_2{\epsilon }_4$${\epsilon }_3{\epsilon }_4$ and ${\epsilon }_4{\epsilon }_4)$ and we tested a linear dosage effect of the ${\epsilon }_2$ allele in the ${\epsilon }_2{\epsilon }_3$ and ${\epsilon }_2{\epsilon }_2$ genotypes on the log odds for EL or log hazard for death, relative to ${\epsilon }_3{\epsilon }_3$ Logistic regression showed increased odds for EL for one copy of ${\epsilon }_2$ in all four studies (Figure 2), and the meta-analysis given in Table 3 produced a significant positive association between a linear dosage effect of the ${\epsilon }_2$ allele and the odds for EL (OR = 1.28, 95% CI: 1.16, 1.41). Survival analysis in both the discovery and HRS sets showed reduced risk for death for increasing copies of ${\epsilon }_2$ in this dosage model (HR = 0.93, 95% CI: 0.85, 1.02 in the discovery set, and HR = 0.81, 95% CI: 0.68, 0.97 in HRS), whereas the Danish study produced a null result (HR = 0.94, 95% CI: 0.85, 1.09). Meta-analysis showed that carrying one copy of the ${\epsilon }_2$ allele in this dosage model decreases the risk for mortality by 6% compared with ${\epsilon }_3{\epsilon }_3$ but the association was only borderline significant (95% CI: 0.88, 1.01). No significant heterogeneity of study-specific effects was detected in this meta-analysis (Figure 2). The survival analysis restricted to extreme survivors showed a nonsignificant association between ${\epsilon }_2$ and risk for death at the most extreme ages (HR = 0.98, 95% CI: 0.90, 1.06) and also no significant heterogeneity of study-specific effects (Figure 2).

Figure 2.

Forest plots of odds ratio (OR) and hazard ratios, relative risk (RR) estimated using the additive genetic models. The analysis was restricted to carriers of the ${\epsilon }_3{\epsilon }_3$${\epsilon }_2{\epsilon }_3$ and ${\epsilon }_2{\epsilon }_2$ genotypes. The Japanese set was not included in the survival analysis of all study participants because ages of controls are unknown. ES = extreme survivors. ${\chi }_n^2$ denotes Cochran’s Q test of heterogeneity with n degrees of freedom and p is the p value to test the null hypothesis of no heterogeneity across studies. CI = confidence interval; HRS = Health and Retirement Study.

Genotype Group Association

We created three genotype groups defined as $E2={\epsilon }_2{\epsilon }_2, {\epsilon }_2{\epsilon }_3$$E3={\epsilon }_3{\epsilon }_3$$E4={\epsilon }_2{\epsilon }_4,{\epsilon }_3{\epsilon }_4,{\epsilon }_4{\epsilon }_4$ and analyzed their association with EL. The study-specific analyses (Supplementary Table 3) showed that carriers of $E2$ have increased odds for EL compared with carriers of $E3$ with effects ranging from 1.62 in the Japanese study (95% CI: 1.14, 2.31), to 1.38 in the discovery set (95% CI: 1.20, 1.58), to 1.37 in the HRS (95% CI: 0.78, 2.40), and to 1.12 in the Danish sample (95% CI: 0.92, 1.37). The negative association of $E4$ with achieving EL was strong, with the OR for EL ranging from 0.36 in the Japanese set (95% CI: 0.26, 0.50) to 0.51 in the discovery set and HRS (95% CI: 0.43, 0.59), 0.50 in the HRS (95% CI: 0.45, 0.55), and 0.54 in the Danish set (95% CI: 0.45, 0.64). Meta-analysis produced very significant results (Table 3; OR for EL comparing $E2 vs E3$ = 1.32, 95% CI: 1.19, 1.47; OR for EL comparing $E4 vs E3$ = 0.50, 95% CI: 0.45, 0.56). Survival analysis showed consistent results: all studies estimated a decreased risk for death in carriers of $E2$ and an increased risk for death in carriers of $E4$ relative to $E3$ and the meta-analysis (Figure 3) showed that carriers of $E4$ have a 21% increased risk of death compared with carriers of $E3$ (95% CI: 1.12, 1.30), whereas carriers of $E2$ have a 6% borderline significant decreased risk for death compared with carriers of $E3$ (HR = 0.94, 95% CI: 0.87, 1.01). The analyses restricted to extreme survivors (those who survived to the oldest 1 percentile) showed a persistent effect of $E4$ even in this subset of individuals and the $E4$ genotype group was associated with a relative risk for death of 1.21 (95% CI: 1.10, 1.35) compared with $E3$ genotype group. As in the analyses of other genetic models, the protective effect of $E2$ relative to $E3$ in the extreme survivors became smaller (HR = 0.98, 95% CI: 0.90, 1.07). Note that this analysis restricted to extreme survivors has 80% power to detect an HR of at least 0.83 comparing $E2$ to $E3$ and no significant heterogeneity of study-specific effects (Figure 3).

Figure 3.

Forest plots of odds ratio and hazard ratios estimated using the genotype group models. $E2={\epsilon }_2{\epsilon }_2, {\epsilon }_2{\epsilon }_3$$E3={\epsilon }_3{\epsilon }_3$ and $E4={\epsilon }_2{\epsilon }_4,{\epsilon }_3{\epsilon }_4,{\epsilon }_4{\epsilon }_4$ The Japanese set was not included in the survival analysis of all study participants because ages of controls are unknown. ES = extreme survivors. ${\chi }_n^2$ denotes Cochran’s Q test of heterogeneity with n degrees of freedom and p is the p value to test the null hypothesis of no heterogeneity across studies. CI = confidence interval; HRS = Health and Retirement Study; RR = relative risk.

Haplotype Analysis of SNPs in the APOE Locus

To investigate whether the APOE alleles explain the associations with EL of SNPs in the PRVL2, TOMM40, APOE, and APOC1 gene cluster that were detected in genome-wide association studies (37), we analyzed their joint effect using haplotype analysis. We estimated haplotypes comprising rs6857, rs2075650, rs769449 (Supplementary Figure 6A), and rs429358 and rs7412 that define the APOE alleles ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ using an iterative procedure based on the EM algorithm (38). The analysis identified 20 haplotypes, but only 11 listed as shown in Supplementary Figure 6B had frequencies more than 0.1%. The analysis also showed that only three haplotypes were individually associated with EL (Supplementary Figure 6C). The haplotype 16 (GAGTT) that includes ${\epsilon }_2$ was associated with a 34% increase odds for EL (p = 7.8 E−07) compared with the most common haplotype (15: GAGTC) that includes ${\epsilon }_3$ whereas haplotype 5 (AGACC) that includes ${\epsilon }_4$ was associated with a 50% reduction of the odds for EL (p < 1.0 E−08) compared with haplotype 15. In addition to these two associations, the uncommon haplotype 10 (AGGTC) that includes ${\epsilon }_3$ was associated with a significant 20% reduction of the odds for EL compared with haplotype 15 (p = .04).

We next analyzed the association between alleles of SNPs rs6857, rs2075650, and rs769449 and EL within haplotypes that include the APOE alleles, using a haplotype-based conditional analysis. Within haplotypes that include APOE = ${\epsilon }_2$ we essentially did not observe variability of the other three SNPs and almost all 1,600 haplotypes that include ${\epsilon }_2$ also include the common variants of rs6857, rs2075650, and rs769449 (Supplementary Figure 6D). Within haplotypes that include APOE = ${\epsilon }_4$ we did not detect any significant association between the three SNPs and EL (Supplementary Figure 6E). Within haplotypes that include APOE = ${\epsilon }_3$ the G allele of rs2075650 was associated with a 4% reduction of the odds for EL (p = .02), and the A allele of rs6857 was associated with a 3% reduction of the odds for EL (p = .06). This analysis shows that carriers of the haplotype that includes ${\epsilon }_3$ and the G allele of rs2075650 have reduced odds for EL, compared with carriers of the haplotype that includes ${\epsilon }_3$ and the common A allele of rs2075650, and suggests that there are SNPs in this locus with a negative effect on longevity, in addition to APOE ${\epsilon }_4.$

Discussion

We conducted a thorough analysis of the effects of APOE variants on extreme human longevity using genotype data of 28,297 subjects from seven studies of human longevity, including four studies of centenarians. By using three different genetic models and two definitions of longevity based on either a threshold model or age at death, we showed that ${\epsilon }_4$ is associated with decreased odds for EL, and an increased risk for death that persists even at the most extreme ages reached by less than 1% of the population. The analyses suggest that there is not much difference between the effects of genotypes ${\epsilon }_2{\epsilon }_{2 }$ and ${\epsilon }_2{\epsilon }_{3 }$ on living to or beyond the oldest one percentile of survival, and that carrying the ${\epsilon }_2{\epsilon }_2$ or ${\epsilon }_2{\epsilon }_3$ genotype is associated with significantly increased odds to reach EL, with decreased risk for death compared with carrying the genotype ${\epsilon }_3{\epsilon }_3$ but with only a modest reduction in risk for death beyond an age reached by less than 1% of the population and this risk reduction failed to reach statistical significance. Since the genotype group association restricted in the extreme survivors was powered to detect an HR of at least 0.83, we can exclude that the lack of statistically significant effect of this magnitude is due to the lack of power.

There is a rich body of evidence showing that the ${\epsilon }_4$ allele of APOE is associated with increased risk for Alzheimer’s disease and age-related cognitive decline and death (16,28,39,40), and our analysis supports the finding that this variant is associated with substantial reduction in the chance for EL. The role of the ${\epsilon }_2$ allele in longevity and health span however has been less clear (18,41). This allele is rare, and collecting sufficiently large samples for sufficiently powerful genetic association studies has been challenging. In addition, the prevalences of the three alleles vary by race and ethnicity (15,42) as well as by age group since carriers of the ${\epsilon }_4$ allele are at high risk for premature mortality. Because of this biodemographic selection, the ${\epsilon }_4$ allele tends to be less common in individuals who survive into older age, whereas the frequencies of the ${\epsilon }_2$ and ${\epsilon }_3$ alleles in older age groups may be larger than the frequencies in younger age groups independent of their relationship to longevity. Our analyses in this study show a clear positive effect of ${\epsilon }_2$ relative to ${\epsilon }_3$ on the odds for survival up to the oldest 1 percentile of survival that is independent of ${\epsilon }_4$ However, this effect does not continue with the same magnitude at more extreme ages.

In a recently published meta-analysis of four genome-wide association studies, we identified some uncommon variants that were associated with EL defined using a threshold model, and similar to the analysis presented here, the effect of these variants on the risk of death diminished at the most extreme ages (37). Using the same type of statistical analysis, we also showed that alleles of FOXO3 that have been associated with EL lose their effects when the analysis is restricted to survivors to extreme ages (oldest 1 percentile and older) (43). Genome-wide screening in other species has similarly shown that there exist gene variants that increase median lifespan without influencing maximum lifespan (44). Examples include SNPs in ADARB1 and ADARB2 and their orthologues in Caenorhabditis elegans (45). Therefore, it is possible that some genetic variants associated with living to the 5th percentile of survival, for example, do not increase the chance of achieving the 1st percentile of survival, and their overrepresentation in centenarians does not mean that they are causative variants for EL.

It is very important in studies of exceptional longevity to realize that just observing a variant or group of variants more frequently in an EL sample than in the general population does not necessarily mean that their influence on survival becomes stronger with even older ages. This observation makes intuitive sense because survival to the oldest 1 percentile (eg, age 97 for men and 100 for women) is 10 times more common than the oldest 0.1 percentile (eg, age 100 for men and 103 for women) and 100 times more common than the oldest 0.01 percentile (eg, age 104 for men and 107 for women). The extremely different odds of survival for these relatively tiny incremental changes in age likely mean that these age groups are different phenotypes resulting from different underlying genetic and nongenetic interactions and resultant biological mechanisms.

There are several implications resulting from these findings. Although confirming the association of the allele ${\epsilon }_2$ with survival to older age is important for understanding the biodemography of this gene variant, its value as a possible target for healthy aging therapeutics may be reduced by the lack of association with age at death at the extreme end of the survival curve. One of our goals in studying centenarians is to search for genetic factors that are associated with both EL and the delay or escape of morbidity and disability. As noted in the Introduction section, however, both compression of morbidity and disability are associated with extreme ages approaching the limit of human lifespan (2–4). The lack of effect on the risk for death at extreme ages may therefore imply that compounds targeting ${\epsilon }_2$ could produce less substantial effects on healthy aging, but this remains an open question.

Several genome-wide association studies have associated tag SNPs in the cluster of genes PVRL2, TOMM40, APOE, and APOC1 in chromosome 19 to EL (40,46–48), and some investigators have hypothesized that these variants may be important for EL independently of the APOE alleles ${\epsilon }_2, {\epsilon }_3,{\epsilon }_4$ For example, Lu et al. used haplotype analysis of SNPs tagging this region to conjecture that PVRL2, TOMM40, in addition to APOE, may be important for EL (48). In a recent meta-analysis of four genome-wide association studies of EL (37), we identified a large cluster of SNPs in this region that was significantly associated with EL and, by examining their patterns of linkage disequilibrium, we identified two most significantly associated SNPs with EL (Supplementary Figure 6A). In this study, we analyzed the joint association of these two SNPs and also rs2075650 in TOMM40, with EL controlling for the effect of the APOE alleles using haplotype analysis (49). This analysis showed that the haplotypes that include the ${\epsilon }_3$ allele of APOE and the G allele of rs2075650 (haplotypes 10 and 19 shown in Supplementary Figure 6B) are associated with a 4% reduction of the odds for EL compared with haplotypes that include the ${\epsilon }_3$ allele of APOE and the A allele of rs2075650 (haplotypes 3, 13, and 15 shown in Supplementary Figure 6B). None of the SNPs rs6857, rs2075650, and rs769449 was associated with EL in carriers of haplotypes that include the ${\epsilon }_4$ allele of APOE. In addition, we observed virtually no variability of the SNPs rs6857, rs2075650, and rs769449 in the 1,660 haplotypes that include the ${\epsilon }_2$ allele of APOE. The analysis suggests that there may be some additional genetic signal in this locus that is detrimental to longevity, but did not detect genetic variants associated with increased chance for EL in addition to haplotypes including the ${\epsilon }_2$ allele of APOE.

Although all formal tests of heterogeneity failed to detect significant variability between study-specific ${\epsilon }_2$ effects, the forest plots given in Figures 1–3 and Supplementary Figures 3–5 suggest that there is some between-study variability of the association between ${\epsilon }_2$ and longevity. Both the discovery set and the HRS include a variety of European ethnicities, in which the frequency of APOE alleles can be different (15). In addition, the various studies in the meta-analyses include participants from different birth year cohorts, different continents, and exposures to different environments. Ethnicity, secular trends, and/or environmental risk factors could be modifiers of the genetic association of APOE alleles with EL (50), and larger data sets and different ways to analyze the data could discover significant interactions with APOE alleles that could suggest interventions to promote healthy aging and longevity.

A limitation of the analysis is that two of the data sets included in our study (Southern Italian Centenarian Study and HRS) did not have directly genotyped data for the APOE alleles. In New England Centenarian Study, we had genotyped and imputed data for the SNPs rs7412 and rs429358 that are used to define ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ and we studied the agreement between the two data sets using Cohen’s kappa measure of agreement. The analysis showed a reassuring 83% agreement, with most of the disagreement between ${\epsilon }_2{\epsilon }_3$ and ${\epsilon }_3{\epsilon }_3$ and ${\epsilon }_3{\epsilon }_4$ and ${\epsilon }_3{\epsilon }_3$

In conclusion, we assembled the largest study of APOE alleles ${\epsilon }_2$${\epsilon }_3$ and ${\epsilon }_4$ and extreme human longevity to date, and using a variety of analyses, we showed that ${\epsilon }_4$ is associated with a substantially decreased odds for EL and increased risk for death that persists even beyond ages reached by less than 1% of the population. We also showed that carrying one or more copies of the ${\epsilon }_2$ allele is associated with significantly increased odds to reach EL and with decreased risk for death compared with carrying the genotype ${\epsilon }_3{\epsilon }_3$ Although the magnitude of the ${\epsilon }_4$ effect is substantial (>80% increased risk for mortality), our analyses suggest that, on average, carrying the ${\epsilon }_2$ allele is associated with only a modest 6% reduction in risk for death that decreases to 2% at the extreme end of human lifespan and did not reach statistical significance. This moderated genetic effect of ${\epsilon }_2$ is very close to values reported in a much smaller study (19). The overall size of our study is large but the number of carriers of the ${\epsilon }_2{\epsilon }_2$ genotype is still relatively small (N = 159), and substantially larger sample sizes are needed to show a statistically significant association between the ${\epsilon }_2{\epsilon }_2$ genotype and EL of this magnitude. Larger power could also be gained by selecting long-lived individuals free of diseases; because consistently with the “expansion of morbidity” hypothesis, medical interventions that increase the population life expectancy may increase the number of people surviving to very old ages with disease (51). Interestingly, the variability of the results between studies suggests that it may be informative to examine ethnicity-specific associations between the ${\epsilon }_2$ allele and longevity and such analysis may point to modifiers of the genetic effect of APOE and discovery of interventions that promote healthy aging and longevity.

Materials and Methods

Study Populations

We aggregated data from three centenarian studies of European ethnicity (New England Centenarian Study, Southern Italian Centenarian Study, and Longevity Study) and a study of familial longevity (Long Life Family Study) to generate a discovery data set. The studies are described at length in the Supplementary Material and in reference (37) and have decided to share individual level data. We replicated the results in three additional studies. All study data are summarized in Table 1.

Statistical Analysis

We analyzed the effect of the ${\epsilon }_2$ allele using three different genetic models and two different statistical models that are briefly described. Full details are provided in the Supplementary Material.

Genetic models

In the genotypic model, we analyzed six genotypes of APOE alleles ${\epsilon }_2{\epsilon }_2, {\epsilon }_2{\epsilon }_3,{\epsilon }_2{\epsilon }_4,{\epsilon }_3{\epsilon }_3,{\epsilon }_3{\epsilon }_4,{\epsilon }_4{\epsilon }_4$ using ${\epsilon }_3{\epsilon }_3$ as referent group. In the genotype group model, we analyzed three genotype groups: $E2={\epsilon }_2{\epsilon }_2, {\epsilon }_2{\epsilon }_3$$E3={\epsilon }_3{\epsilon }_3$ and $E4={\epsilon }_2{\epsilon }_4,{\epsilon }_3{\epsilon }_4,{\epsilon }_4{\epsilon }_4$ using $E3$ as referent group. In the additive genetic model, we analyzed only carriers of the genotypes ${\epsilon }_2{\epsilon }_2, {\epsilon }_2{\epsilon }_3, {\epsilon }_3{\epsilon }_3$ that were coded as 2,1,0 to represent the number of copies of the ${\epsilon }_2$ allele. Each of these three different genetic models was used in statistical analyses aimed at detecting association of ${\epsilon }_2$ with EL defined by a threshold model (logistic regression) and age at death (survival analysis).

Logistic regression

We defined EL as living past the sex-specific age at which at most 1% of individuals from the 1900 birth year cohort survived (age 96 and older for men and 100 and older for women), whereas controls were defined either as individuals who died before reaching the threshold age or as alive random controls not selected for familial longevity. The analyses were adjusted by sex and, with the exception of the Danish and Japanese studies that are ethnically homogeneous, by the first four genome-wide principal components.

Survival analysis

To study the effect of ${\epsilon }_2$ on the age at death, we used Cox proportional hazards regression model of age at death, adjusted by sex and four genome-wide principal components in the discovery and HRS sets. Age was censored at last contact for alive individuals. The analysis in the Danish and Japanese sets was not adjusted for principal components given the ethnically homogeneous nature of those samples.

Meta-analysis

The results of the various data sets were meta-analyzed using inverse-variance weighting, in the package rmeta (R v3.4). Heterogeneity was tested using Cochran’s Q statistic that follows a ${\chi }_n^2$ with degrees of freedom equal to the number of studies −1.

Haplotype analysis

Haplotype analysis was conducted using the haplo.stats package in the R software (38). Associations of APOE and SNP haplotypes with EL were estimated using logistic regression. Haplotype-based conditional analysis estimated the allelic association between SNPs and EL in subsets of the data defined by APOE alleles. Both types of analyses were adjusted by sex and four genome-wide principal components.

Power calculations

We computed the number of events needed to detect an HR with (1-β)% power and α% type I error rate with the formula ${(z_{\frac{\alpha }{2}}+z_{\beta })}^2 /({\pi }_0{\pi }_1\text{log}{(HR)}^2)$ where $z_{\alpha }$ is the quantile of standard normal distribution associated with probability α and ${\pi }_0 {\pi }_1$ are the proportions of events in the two groups. Power calculations for the logistic regression models were conducted using Gpower software.

Funding

This work was supported by the National Institute on Aging (NIA cooperative agreements U01-AG023755 TP, U19-AG023122 TP, and R21AG056630 PS), the National Institute of General Medical Sciences (NIGMS) Interdisciplinary Training Grant for Biostatisticians Program (T32 GM74905), the William M. Wood Foundation (T.T.P.), and the Paulette and Marty Samowitz Family Foundation (T.T.P.). The Health and Retirement Study genetic data is sponsored by the NIA (grant numbers U01AG009740, RC2AG036495, and RC4AG039029) and was conducted by the University of Michigan. The Nathan Shock Centers of Excellence in the Basic Biology of Aging (P30AG038072) (N.B.), the Glenn Center for the Biology of Aging (Paul Glenn Foundation for Medical Research) (N.B.), NIH/NIA 1 R01AG044829 (PIs-Veghese/Barzilai), NIH/NIA1 R01 AG 042188-01 (Atzmon/Barzilai, NIH-1 R01 AG 046949 - 01 (Barzilai, PI). The Danish Twin Registry is supported by the Danish Agency for Science, Technology and Innovation, and the US National Institutes of Health (P01 AG08761). The Danish Aging Research Center is supported by the Velux Foundation. The Japanese Centenarian Study was supported by AMED Program for an Integrated Database of Clinical and Genomic Information, and Program for Initiative Research Projects, Keio University. Genotyping of APOE in the New England Centenarian Study was supported by the Clinical and Translational Science Institute, Boston University (1UL1TR001430).

Conflict of Interest

None reported.