Methylome-wide association study of whole blood DNA in the Norfolk Island isolate identifies robust loci associated with age

Introduction

Epigenetics is a rapidly developing area of research and refers to the heritable, but reversible, regulation of various genomic functions including gene expression. This provides mechanisms whereby an organism can dynamically respond to a change in its environment and modulate gene expression accordingly. As such, epigenetic mechanisms can have a profound effect on phenotype, including disease risk and progression. One such epigenetic mechanism is DNA methylation, with cytosine residue methylation at CpG dinucleotides well documented in many organisms. The investigation of an individual’s methylation pattern can reveal a lifetime record of environmental exposures as well as potential disease specific patterns.

Aging in humans has been associated with marked remodelling of the epigenetic architecture in terms of DNA methylation patterns [1–3]. To date these studies have identified age-related CpG associations in healthy populations [4,5], and age-related CpGs associated with disease susceptibility [6]. Studies have assessed the extremes of the age distribution by comparing the methylomes of centenarians with newborns [7,8]. One study proposed the idea of a methylation 'clock' for tissue specific aging [9], while another demonstrated the use of DNA methylation to predict all-cause mortality in later life [10]. This breadth of evidence highlights that age-associated methylation variants could be important in influencing age-related disease.

Most studies of methylation in relation to aging have focused on the use of whole blood DNA given its relative accessibility when compared to other cell and tissue types. The Infinium Human Methylation 450 Beadchip (450K Array) has been the most popular technology for conducting such methylome-wide association studies (MWAS). A large number of MWASs focusing on blood have been conducted and data for >4700 individual samples have been posited in the public repository called MARMAL-AID [11]. Interestingly, a recent meta-analysis of whole blood MWASs applied a new bioinformatics method to identify differentially methylated regions (DMRs) associated with aging in humans [12]. Briefly, CpG sites were identified based on their genomic features, defining groups of adjacent sites based on their density, and applying either a single site or region centric analysis. The authors suggested that this design should allow for better comparison across other studies of the similar design. This study resulted in a robust list of DMRs that vary across the lifespan and may also have potential importance in aging biology [12]. More recently, a review by Jones et al. [1], also provided a meta-analysis of age-related methylation papers and explored the overlap between the most significant CpG sites across 7 separate studies to identify 11 CpG sites commonly associated with age [1].

Most age-based MWAS conducted to date have been performed in unrelated cohorts collected from general populations. Whilst such studies can offer the advantage of large cohort sizes they can be negatively affected by confounding factors such as underlying genetic substructure and highly variable environmental influences. Genetically isolated populations may help overcome such issues due to having a reduced genomic and environmental diversity when compared to general populations [13]. The Norfolk Island (NI) population is a genetic isolate with a well-documented history [14]. Norfolk Island is geographically remote, located ~1600kms off the East Coast of Australia. The modern NI population was originally founded in the late 1780's on Pitcairn Island by 9 Bounty Mutineers and 6 Polynesian wives and in 1856 the founder descendants relocated to NI [15]. Given its remoteness, the population grew in almost complete isolation from mainland Australia. This, along with the small island size and strict immigration policies, has ensured that both genetic and environmental conditions shared by all NI individuals have remained fairly homogeneous. To date the isolate has been well characterised genetically and phenotypically as part of the Norfolk Island Health Study (NIHS) [16].

In an effort to further our understanding of age-related methylation we conducted a MWAS of whole blood DNA in a healthy cohort from the isolated population of NI. We applied a statistical algorithm called GLMnet [17], a machine-learning approach which conducts simultaneous analysis of CpGs by allowing mixing of ridge regression and Lasso (least absolute shrinkage and selection operator) in an elastic-net framework. We validated our top findings in the large cohort of publicly available methylation data from MARMAL-AID.

Results

Discussion

Here we report MWAS of peripheral blood DNA methylation in the NI genetic isolate. It is well established that genetic isolates offer advantage to the study of complex traits and disorders at a genomic level [13], but few studies have yet to explore this in the context of epigenetics [22,23]. Along with genetic homogeneity, one of the major benefits of studying an isolated population is the potential for reduced environmental effects, i.e. a shared environment (diet, weather, temperature, social structure). Using a cohort of healthy females from the NI population we identified numerous age-related methylation sites, many of which confirm already published research. Interestingly, several CpGs were uniquely identified in our study and validated in a large Illumina 450K public data set demonstrating the utility of population isolates in epigenetic association studies. Not only does this demonstrate the potential power of genetic isolates, but also the reliability and reproducibility of Illumina 450K DNA methylation data. Evidence for the reliability of Illumina 450K data has been presented in numerous studies, including our own [24], in particular accurate and precise estimates of methylation down to 2-4% detection levels [25–28] can be achieved. In this study in-silico validation was performed to confirm our findings using over 2300 individuals from 14 independent studies. This weight of evidence provides confidence that the findings in our study are valid and not due to any technical errors associated with the 450K arrays.

Two genes identified in our study were DDO, containing a set of 5 CpG sites forming a robust DMR across the promoter region, and MYOF with a single CpG site in the gene. Methylation of specific CpG sites in these two genes is reported here for the first time to be associated with age in a healthy human cohort.

Previously, association of methylation at a single CpG DDO site (cg14956327) with age has been reported in a cohort of northern Europeans, however this study concentrated on older individuals at risk of metabolic syndrome [2]. In contrast our study subjects were selected to be as healthy as possible (never smoked, minimal metabolic risk factors), and were of an overall younger age (<47 years old). Our study initially identified four DDO promoter CpG sites as age-associated in this healthy female NI cohort. We then validated this by using publicly available data for 2316 blood samples from mixed sex subjects, confirming the association and identifying a fifth age-associated DDO CpG site also in the promoter region (see Supplementary Figure S2). As such our results confirm that the age-association identified across DDO promoter CpG sites is observed in both female and male individuals from all age ranges. Moreover, it is interesting to note that DDO was not identified in the majority of previous studies, which leads us to suggest that the machine learning method we have employed here at least offers additional value to existing analytical methods, and should be used to complement and enhance such approaches.

D-aspartate oxidase is an enzyme encoded by the DDO gene. The protein, a peroxisomal flavoprotein, catalyses the oxidative deamination of D-aspartate (D-Asp) and N-methyl D-aspartate (NMDA) [29]. The current body of literature detailing the role of DDO suggests that it is crucial to several metabolic processes, many of which involve the transfer of electrons and generation of reactive oxygen species [30]. Biologically the impact of these processes over time is known to result in an accumulation of these potentially damaging components, as in the situation of reactive oxygen species generated by the mitochondria. One theory is that the accumulation of these free-radicals over time is responsible for increased damage to important cellular components, which may contribute directly to the aging process. It has been reported that DDO expression is highest in the brain of both animal models and humans [31,32]; it is nearly absent during embryonic and perinatal development and progressively increases during adulthood. Recently, a murine study reported that DDO promoter demethylation enables postnatal DDO expression, and that constitutively suppressed DDO expression (in DDO knock-out mice) leads to increased extracellular D-Asp levels in the brain, resulting in precocious neuronal cell death triggered by excessive NMDA receptor stimulation [33]. This suggests a key role for DDO in preventing neurodegeneration during brain aging. Our results suggest this may also be the case in humans. We observed demethylation in blood and it continues throughout life, while in the Punzo et al. study, conducted using whole mouse brain, the demethylation appeared to plateau to a level of approximately 30% at 3 weeks post partum (around time of weaning).

We also identified DNA methylation at cg14060519 in MYOF as being associated with age in the NI cohort. Methylation of this CpG has not been previously associated with age, however another CpG in an exon of MYOF (cg14428166), which is ~99Kb away from cg14060519, has been previously associated with aging in a very large meta-analysis exploring interactions between age, gene expression (RNAseq), and methylation (450K) [19]. Of note, we were able to identify differential methylation of a previously unreported age-associated CpG site in MYOF in a relatively modest sample size from a genetic isolate with this observation validated strongly in public data (Figure 4). Further investigation of the genomic region around cg14060519 revealed no other CpG's within +/-2kb of this position. However, the intronic location of this CpG is within an annotated regulatory hotspot, featuring DNase I hypersensitivity, an H3K27Ac mark and ENCODE CHIP-seq data. MYOF codes for a protein called myoferlin [34]. Myoferlin is very similar structurally to dysferlin, both belong to the ferlin family of proteins. These are calcium-sensing, membrane-associated proteins which play an important role in muscle membrane repair and growth [35]. Mutations in these ferlin proteins can cause muscle weakness affecting both proximal and distal muscles. Myoferlin has been suggested as a candidate gene and potential modifier for muscular dystrophy [34], and is required for insulin-like growth factor response and muscle growth [36]. Furthermore, myoferlin has also been shown to be highly expressed in endothelial and vascular tissues where it has a role in membrane integrity via its regulation of vascular endothelial growth factor (VEGF) receptor-2 stability and signalling [37]. These data suggests a plausible biological relevance of a role for MYOF in the aging process.

While differentially methylated regions (multiple CpG sites) are most commonly reported, there is evidence that single-marker CpG sites, such as the one we identified in MYOF, can show associations with traits of interest, and indeed, functionality. Nile et al., (2008), reported that methylation status at a single CpG site in the promoter of IL6 affected the regulation of gene expression of IL6, potentially influencing rheumatoid arthritis [38]. In another study, by Fürst et al., (2012), a single differentially methylated CpG was identified to affect transcription of ESR1 [39]. More recently, Dick et al., (2014) identified a single CpG site from whole blood associated with BMI in obese individuals [40]. They also identified the same correlation at the same single CpG site in adipose tissue, providing an excellent example of a robust single marker association and its translation to a potential biomarker.

DNA methylation is an epigenetic mechanism well known to be involved in the control of gene expression [41]. In this context we explored the association of age with gene expression in 330 NI individuals previously expression typed [18]. We observed no correlation between RNA level and age for either DDO or MYOF. Due to the location of the CpG sites identified in this study in regulatory and promoter regions we also explored associations of methylation and RNA expression for DDO and MYOF. We observed no significant associations between the methylation level of CpG sites and transcript expression in either gene. It may be that the relatively small sample size has impaired our ability to observe such correlations and as such future work will benefit from assaying a larger number of pedigree members. In addition, the Illumina HT12 array only has one probe assigned to DDO which presents a limitation in that a single probe does not account for other potential isoforms. Obtaining RNA-seq data in a larger sample size would allow exploration of multiple isoforms and their potential association with DNA methylation.

In this study, we highlight the promoter region of DDO as it falls within the classic ‘central dogma’ of an unmethylated promoter leading to gene expression while a methylated promoter results in a lack of gene expression. It is however likely that complex motif/barcode patterns of methylation across regions such as promoters influence gene expression, and that this is the result of clusters of correlated sites within these regions. The sites we report in DDO show strong pairwise correlation between each other (r=0.94-0.98 p=1x10-5-1x10-10), suggesting that any other CpG sites in close proximity likely show a similar methylation level. While DNA methylation is a potential modulator of gene expression, with hypermethylation of promoter regions strongly associated with decreased gene expression this is not always the case. Also, hypomethylation, of gene bodies has been associated with decreased mRNA expression [42] and it is increasingly realised that DNA methylation has effects on alternate splicing [43–45]. It is possible that alternate transcripts regulated by DNA methylation were not covered by the Illumina HT-12 arrays and therefore remain undetected in this study. In mouse brains demethylation of the DDO promoter correlated with increased DDO mRNA and treatment of neuronal cells with the DNA-methylating agent, azacitidine, caused a moderate increase in mRNA levels [33]. Here we have observed the same trend in demethylation in the DDO promoter in humans, but did not see the association with mRNA expression. One possible explanation is that while the methylation signal in whole blood is able to be identified in humans, the specific cell type expressing the mRNA may be masked by the heterogeneity of the cells within blood. Alternatively, the observation by Punzo et al., of increased DDO mRNA with promoter demethylation was reported in murine brains, and it may be that we have identified biomarkers in the cells that make up whole blood which tag methylation/expression interactions in other tissues, i.e. the brain. This type of biomarker identification in biologically relevant tissues is well documented by Dick et al., 2014. In this study an association with CpG sites within whole blood was observed with BMI [40]. When looking in several different tissues (adipose tissue, skin epithelium), the same association was observed in subcutaneous adipose tissue but not in the skin, suggesting that blood is potentially a biologically relevant source of biomarkers for BMI interaction in adipose [40].

In this study we demonstrate the ability of the NI population to detect age-associated methylation sites. Aside from the benefit of the NI genetic isolate as a discovery cohort, there are a number of additional potential reasons for this. Firstly, the majority of studies to date have examined specific age intervals, i.e. fetal/newborn, adulthood, old age, rather than encompassing the full age spectrum. In this study we obtained a large amount of public data ranging across ~100 years of human life, and utilized both the increase in power from the number of samples as well as the increase in phenotypic trait variation (age) to our advantage. Secondly, it is important to consider the methodology used to test for association. In this study we used a machine-learning approach implemented in the R package GLMnet [17]. There have been several MWAS publications that have used the GLMnet approach to identify CpG-DNA methylation:trait associations [9,46–48]. The GLMnet package provides benefits to the identification of associations in genomic data through the implementation of an elastic-net routine. The elastic-net framework brings together two established approaches, ridge regression and LASSO, and by applying specific tuning parameters is able to overcome limitations of either method. Here we demonstrate that the GLMnet method has identified both robust DMRs and single CpGs with potential biological relevance and statistical significance which clearly validate in a separate independent cohort of public data. Further validation of our approach is demonstrated by the very good and consistent overlap of both CpG sites and genes in our analysis with those from multiple previous studies and meta-analyses.

Conclusions

It is well established that isolated populations and large pedigrees have been beneficial for performing GWAS of complex traits. Here we have demonstrated the utility of isolated populations in the identification of methylation/epigenetic associations. Using an elastic-net framework we identified a panel of age associated DMRs in a sample of healthy NI females. This list is largely consistent with genes previously associated with aging. Interestingly our approach revealed a robust DMR in the promoter of DDO, a gene not previously reported in aging studies of healthy individuals. Our observations were validated in a large number of public blood samples, suggesting that this association exists in the methylomes of white blood cells, making it a potentially valuable biomarker for aging. Additionally, both DDO and MYOF are potentially relevant from a biological perspective to the mechanisms regulating aging. Further work is required to establish a role for these genes in aging, as well as additional modelling in other tissues to explore the potential of more systemic age-associated methylation profiles.

Materials and Methods

Supplementary Materials

Supplementary File

Supplementary Table 1

Supplementary Table 2