Meta-analysis of genetic variants associated with human exceptional longevity

Introduction

In Sebastiani et al “Genetic signatures of exceptional longevity in humans” [1], we presented the results from a genome wide association study of exceptional longevity in 801 centenarians from the New England Centenarian Study (NECS, mean age at death 104 years) and 914 genetically matched controls. The study identified a group of 281 SNPs that, used jointly in a genetic risk model, had 60% sensitivity to discriminate between centenarians and healthy controls. The sensitivity of the model however increased with more extreme ages of the centenarians and reached 85% for subjects age>107 years. The 281 SNPs included rs2075650 in TOMM40/APOE that reached irrefutable genome-wide significance and replicated in an independent cohort of 253 nonagenarians and centenarians from the Elixir Pharmaceuticals Study of Extreme Longevity and 341 genetically matched controls. The other 280 SNPs were statistically significant with p-values ranging between 10-2 and 10-6 although their level of significance did not meet the stringent criterion for genome-wide significance of 5x10-8, thus raising the possibility that these associations could be false positives. We therefore set out to determine which of these 281 SNPs were associated with longevity in a meta-analysis that included the two original studies, in addition to a case control study of longevity with nonagenarians and centenarians from the Southern Italian Centenarian Study [2], and a case control study of nonagenarians and centenarians from the Long Life Family Study [3]. We also extended the meta-analysis to include genotype data of a subset of SNPs from the Japanese Centenarian Study [4].

Results

Table 1 lists the studies' characteristics. The ELIX, SICS, LLFS and JCS case-control studies were all smaller than the NECS and cases in the ELIX, SICS and LLFS were younger than the NECS. Controls in the LLFS were males who died by the age of 94 or females who died by the age of 95, and 85% of these controls are relatives of the cases (eg siblings who died at younger ages) from the same family of the cases, so that they provide the strongest type of genetic matching. Some of the controls in the NECS and ELIX studies were chosen from the Illumina repository of controls and their ages are unknown.

In the meta-analysis of additive genetic associations in NECS, ELIX, SICS and LLFS, 10 SNPs reached statistically significant association after Bonferroni correction (p-value < 0.05/280=0.00018). An additional 4 SNPs reached Bonferroni corrected statistical significance using a dominant model for the top-strand allele, and two SNPs reached Bonferroni corrected statistically significance using a recessive model for the top-strand allele (Table 2). The number of significant was much larger when a 5% and 6% false discovery rate corrections were used (Supplement Table 1). The Venn diagram in Figure 1 the number of significant SNPs from the meta-analysis of additive, dominant and recessive models, and 128 SNPs reached statistical significance with 6% false discovery rate. Note the substantial overlapping between the results with different genetic models. In fact, the 3 parameters of the additive, dominant and recessive models are functionally related and the tests are not independent. The full list of results for the meta-analysis of the 280 SNPs is in Supplement Table 1.

Only 19 SNPs in the set of 28 that reached FDR corrected significance had genotype data available in the JCS set, and meta-analysis of these 19 SNPs was extended to include the results for the JCS set Six SNPs (see Table 3) reached Bonferroni corrected significance (p-value< 0.0026=0.05/19) in the meta-analysis of results from additive models (4 SNPs), dominant models (1 SNP), and recessive models (1 SNP) and 4 of these 6 SNPs were not included in the list of 16 that reached Bonferroni corrected significance in the meta-analysis of NECS, ELIX, SICS and LLFS. Fourteen SNPs in 19 reached 6% FDR corrected statistical significance. Full details of the analysis that included the JCS set is in Supplement Table 2.

The list of 28 SNPs include both common and uncommon variants that are associated with increase odds for extreme longevity. Figure 2 plots of genetic effects for different genetic models versus the frequency of coded alleles and shows that both uncommon variants (allele frequency < 0.05 or > 0.95) and common variants are associated with increased odds for longevity. Common variants tend to have more modest effects than rare variants.

Figure 3 shows two networks of genes associated with Alzheimer's and coronary artery disease that included SNPs in the list of 281 reported in Sebastiani et al [1]. Nodes circled in blue in the Figure show the genes that include SNPs reaching statistical significance in the meta-analysis (20 of 38 genes associated with Alzheimer's disease, and 14 of 24 genes associated with coronary artery disease).

Discussion

After the publication of the article by Sebastiani et al [1] that presented the corrected results from an article published online in Science Express and then retracted because of technical errors [5], there was discussion about the validity of the 281 SNPs associated with exceptional longevity (http://blog.23andme.com/news/studies-of-extreme-longevity-extremely-challenging/) and confusion about the correct SNP list associated with exceptional longevity [6]. The main objective of this article was therefore to cast some light on the validity of the individual associations of the 281 longevity SNPs. The meta-analyses of 5 case control studies of exceptional longevity of Caucasians and Japanese identified 0 SNPs that reached Bonferroni corrected significance, and 128 that reached statistical significance with a 6% false discovery rate. Besides arbitrary thresholds for statistical significance, the concordance of genetic effects for many of the SNPs across the 5 different studies is also providing evidence of true associations (See Tables 2 and 3, and the supplementary tables for SNPs that reached significance with 5% FDR correction). The large number of replicated SNPs provides further support for the hypothesis that exceptional longevity is a multifactorial trait that is determined in part by an enrichment for common genetic variants imparting protective effects [1, 7, 8]. The rare existence of families that demonstrate clustering for extreme age also suggests the existence of rare variants as well [9].

The list of 128 SNPs include common and uncommon variants that are associated with increased odds for exceptional longevity (Figure 2). The common variants have modest effects and although they reach statistical significance when the multiple correction penalty is limited to 280 tests, they would fail to reach statistical significance at genome-wide level. This result is consistent with other genetic studies of aging and longevity that failed to discover SNPs reaching genome wide significance even with substantially larger sample sizes [10-13], although these variants can be predictive of lifespan [14]. Some of the variants in Tables 2 and 3 are associated with decreased odds for exceptional longevity. It is interesting to note that lack of deleterious variants is not a major feature of these genetic data and we and others indeed showed that centenarians do not appear to differ in the number of deleterious variants compared to healthy controls [1, 15, 16].

Some of the SNPs listed in Tables 2 and 3 are in genes previously associated with underlying mechanisms of aging and age related diseases, and Figure 3 shows an example of replicated genes previously associated with Alzheimer's disease and coronary artery disease. The best known longevity variant in Table 2 is rs2075650 in TOMM40 and linked to APOE. This result is consistent with those reported by other studies [6,11] and by the recent linkage analysis in [17]. The SNP rs2075650 was linked to the ∊4 allele of the APOE gene although the strength of linkage disequilibrium between rs2075650 and the SNPs rs7412 and rs429358 that define the 3 alleles of APOE seems to vary with ethnicity. The Leiden Longevity Study reported a r2=0.56 between rs2075650 and rs429358 in 8946 subjects [11], while in 4576 subjects of the LLFS the r2 between rs2075650 and rs429358 was only 0.27 [18]. The association between the SNP rs2075650 and the ∊4 allele is stronger at the genotype level and in LLFS subjects, 88% of carriers of the GG genotype carried at least one copy of the ∊4 allele, compared to 73% prevalence of the ∊4 allele in carriers of the AG genotype and 6% in carriers of the AA genotype. Interestingly, the effect of the G allele was more modest in the SICS set. This result is also consistent with the known variability of APOE alleles with ethnicities [19]. Unfortunately SNP data were not available for assessment of this association in the JCS set. Note that SNPs in the Alzheimer's disease associated genes CEACAM16, CSFR1, SORCS1, WWOX and DEFB1 (Figure 3) also reached Bonferroni corrected significance.

The SNP rs915179 in LMNA showed consistent effects in the NECS, ELIX and SICS sets but not the LLFS set, although the association reached Bonferroni corrected statistical significance in the meta-analysis of Caucasian studies. This result is consistent with recent findings in [20] and the data included in this analysis partially overlap with those included in [20]. The synonymous coding SNP rs4641 is in moderate LD with rs915179 and we noted that two supercentenarians (ages at death > 114 years) were both heterozygous carriers of this mutation [21] which was associated with survival advantage as well as increased chance for metabolic trait. One study suggested that this advantage is related to increased body mass index (BMI) [22] that in the very old could compensate the effects of sarcopenia and osteoporosis. Supporting an important role of BMI and longevity is the significant association of the SNP rs3906146 in LMX1B and rs9899404 in GIP (Supplement Table 1, FDR < 6%). The SNP rs3906146 is in moderate LD with the less common SNP rs867559 (r2=0.32, D'=1, based on CEPH from HapMap) that was found associated with high BMI and obesity in [23]. These results are in agreement with an important role of insulin and longevity [24,25]. Two SNPs in the WRN gene reached statistical significance in the meta-analysis of Caucasian studies (rs3024239 and rs1800392, Supplement Table 1, FDR < 6%). The association of rs1800392 became significant at 5% Bonferroni corrected significance when the meta-analysis included the JCS set (Supplement Table 2).

The G allele of SNP rs2596230 in RYR3 (Table 3) showed consistent recessive effects in all the studies (Meta OR=2.61). This SNP has been associated with age-related macular degeneration [26]. Note that Table 2 includes an additional SNP (rs216148 in CSF1R) that replicates in the ELIX set that has also been linked to age-related macular degeneration [27].

Limitations. Although this study represents the largest meta-analysis of exceptional longevity to date, the overall sample size is still relatively small to capture small genetic effects. The LLFS set is very large compared to the other studies, but most of the participants are still alive and younger and the LLFS will become much more powerful over time as these individuals achieve older ages. In addition, the median ages at death in the ELIX, SICS and LLFS sets were younger than the median age at death of 104 years in centenarians of the NECS and this difference in ages as well as birth cohort differences may have limited the replication rate. An additional limitation is that although we included studies of different ethnicities and race, the JCS samples were genotyped with a different genotyping platform and imputation failed, so that the majority of the SNPs in the list of 281 could not be tested and additional genotyping and sequencing are necessary. Additional replication studies will provide further evidence in support of these and perhaps other associations.

Material and methods

Conclusions

The original GWAS of exceptional longevity in centenarians from the NECS and genetically matched controls generated a substantial list of genetic candidates for human longevity that included both novel and well established longevity associated genes. The large number of SNPs that reached statistical significance in this analysis shows that many of these variants are robust and we expect to be able to replicate even larger numbers as studies grow in size, their subjects become older and hopefully many more studies will coordinate their efforts to use similar genotyping platforms so that a greater number of markers are comparable. Genetic variations associated with exceptional longevity that are noted in populations with very different genetic backgrounds may be particularly interesting and important to identify biologic pathways that influence exceptional longevity and processes as basic as rate of aging. Although some of the variants in Tables 2 and 3 are known to be associated with aging or age-related diseases, some of the replicated results point to novel genes that may influence aging and extreme lifespan and open new avenues of research on the genetic of extreme longevity.

The challenge is how to follow up these current results. Our multivariate modeling of exceptional longevity in [1] showed that different combinations of the 281 SNPs alleles determine different probabilities of survival to very old ages. Functional experiments that try to assess the individual effects of these variants with longevity may fail or be suboptimal because they would ignore other interacting genetic variants. A systems-based approach will be necessary to discover the synergistic and antagonistic effects of these many variants and their roles in extending lifespan and health-span.

Supplementary Materials

Supplementary Table 1

Supplementary Table 2