Genome-wide association studies: how have they informed and misinformed us

In contrast to candidate-gene studies which focus on variants in genes with a known metabolic role, GWAS are not hypothesis driven. The advantage (but also challenge) of the hypothesis free approach used in GWAS is that it has the power to identify novel biological pathways associated with a phenotype or dietary component of interest. An example of this paradigm is the large research interest in uncovering the biological action of the fat mass and obesity associated (FTO) protein following the identification of the association of a SNP in the FTO gene with BMI in a 2007 GWAS study⁽ Reference Tung and Yeo ^{10 )}, with the authors Frayling et al. concluding that ‘FTO is a gene of unknown function in an unknown pathway’⁽ Reference Frayling, Timpson and Weedon ^{11 )}.

In GWAS, genetic variability is quantified in a group of cases v. matched controls, in order to establish disease associated variants. The standard SNP arrays typically have 300,000–2 × 10⁶ tagging SNP which are correlated with (derived from HapMap⁽ Reference Altshuler, Gibbs and Peltonen ^{12 )}) and provide information on 80–90% of common variation (frequency >5%), but far less for the low frequency variants (0·01–5%) and virtually none for the rare variants⁽ Reference Marian and Belmont ^{13 )}.

The first GWAS output was published in 2005 and identified a polymorphism in complement H to be linked with age-related macular disease⁽ Reference Klein, Zeiss and Chew ^{14 )}. This led the way to an explosion of activity, and to date, GWAS has identified over 1600 variants associated with 250 traits⁽ Reference Hindorff, MacArthur and Wise ^{15 )} and has had some success in identifying a large component of the genetic basis of particular phenotypes. For example, as reviewed by Manolio et al. for age-related macular degeneration five loci have been identified which collectively explain 50% of total heritability⁽ Reference Manolio, Collins and Cox ^{2 )}. In 2010, a meta-analysis of forty-six cohort studies confirmed ninety-five loci predictors of plasma lipids (total cholesterol, HDL-cholesterol, LDL-cholesterol and TAG) which explained 25–30% of the genetic component of these traits⁽ Reference Teslovich, Musunuru and Smith ^{16 )}. However, for the majority of polygenic traits, the identified variants only account for a much lower proportion of the total estimated heritability which varies from 20 to 80% depending on the phenotype of interest⁽ Reference Manolio, Collins and Cox ^{2 ,} Reference Lander ^{17 )}. For BMI and obesity, thirty-two individual loci have been identified and confirmed, but the effect size of each individual variant is small. The largest effect is evident for FTO, with each risk allele increasing BMI by on average 0·39 kg/m² and obesity risk by 1·20⁽ Reference Loos ^{18 )}. However, collectively the thirty-two loci explain only 1·45% of the phenotypic variation in BMI, equivalent to 2–4% of heritability⁽ Reference Speliotes, Willer and Berndt ^{19 )}.

There are likely to be a number of reasons why GWAS has resulted in only modest capture of heritability⁽ Reference Gibson ^{20 )}. Firstly, although the gene variant may have been identified by GWAS, it may not have emerged as significant or its effect size may be underestimated for a number of reasons which include:

1. 1. Use of stringent P value (typically P < 1 × 10⁻⁸) to compensate for multiple testing and eliminate false positive results. This may result in failure to detect many true signals and we may be ‘correcting away the hidden heritability’⁽ Reference Williams and Haines ^{21 )}. Alternative approaches to the use of strict P values, such as multi-stage confirmation of significant SNP in subsequent datasets⁽ Reference Easton, Pooley and Dunning ^{22 ,} Reference Consortium ^{23 )}, have led to the identification of further variants of interest;

2. 2. Imprecise phenotyping⁽ Reference Marian and Belmont ^{13 )}. For disease outcomes, patients in the ‘case group’ often present with a range of related conditions with variable genetic aetiology. For example, in the cardiovascular field, myocardial infarction, ischaemic heart disease and coronary stenosis are often pooled, although they have both common and separate aetiological components. For many outcomes, such as blood pressure, the precision of the measurement is problematic, while for others there is a large intra-individual variability such as pro-inflammatory cytokines and C-reactive protein, which means the trait is imprecisely captured.

3. 3. Control group of questionable quality⁽ Reference McCarthy, Abecasis and Cardon ^{24 )}. Often an individual in the control group does not have a clinical diagnosis of the primary outcome but is a registered patient for an alternative condition whose risk may also be impacted by the identified gene variants, therefore underestimating the effect size of the variant.

4. 4. True causal variants incompletely surveyed are not in full linkage disequilibrium with tagging SNP.

5. 5. A number of causal variants may exist in one locus, with only one tagging SNP chosen, which may result in an underestimation of the total heritability accounted for by that particular region.

Secondly, it is plausible and increasingly demonstrated that rarer single nucleotide changes, or structural variants such as copy number variations⁽ Reference Redon, Ishikawa and Fitch ^{25 ,} Reference Almal and Padh ^{26 )}, which are far more common than originally thought, could make a significant contribution to hidden variability (Fig. 2)⁽ Reference Almal and Padh ^{26 –} Reference Jakobsson, Scholz and Scheet ^{28 )}. Next generation sequencing is becoming increasingly feasible and affordable and is in more widespread use as a research tool⁽ Reference Manolio, Collins and Cox ^{2 ,} Reference Bras, Guerreiro and Hardy ^{29 ,} Reference Clarke, Zheng-Bradley and Smith ^{30 )}. This technology provides a complete map of an individual's genome, overcoming limitations associated with SNP tagging in GWAS and allowing the detection of less common variants. The 1000 Genomes Project, which will include far more than 1000 aims to capture all variants with <1% frequency and >0·1% in protein coding regions (exome)⁽ Reference Clarke, Zheng-Bradley and Smith ^{30 )}. It is hoped that this technology will detect much of the hidden heritability.

Fig. 2.

(colour online) Identification of genetic variants of various frequencies and effect sizes. (Adapted from⁽ Reference Manolio, Collins and Cox ^{2 ,} Reference McCarthy, Abecasis and Cardon ^{24 )}.) (1) Examples of Mendelian diseases include Huntington's disease, sickle cell anaemia and cystic fibrosis. Lifestyle, including diet composition often has a minimal effect on disease severity; (2) these variants are difficult to identify and given their rarity and small effect size their identification is not a priority; (3) currently a few of these disease-associated variants have been identified possibly due to their lack of representation on currently used genome-wide association studies (GWAS) arrays. Increased use of sequencing technologies and redesign of traditional arrays is predicted to substantially increase their detection rates; (4) one example of such a genotype is the association between the APOE4 allele and risk of Alzheimer's disease; (5) to date >95% of the identified common variants associated with disease have modest effects sizes 1·0–1·5, and explain only a small proportion of the total heritability of the phenotype.

Thirdly, the broad sense heritability model posits that once ‘unveiled’ neither common variants with modest impact nor rare alleles with high penetrance are likely to explain away missing heritability. It theorises that known genetic variation in the form of interactions, between allele pairs (dominance), between alleles in different genes (epistasis) and between genotype and environment (including diet composition) or physiological variables, explain a large proportion of inheritance.

Impact of physiological variables on genotype–phenotype associations

Currently, in genetic studies, populations are considered as single entities. It is becoming increasingly apparent that population genetic associations often under- or over-estimate the effect in subgroups. At this stage, it is too early to say what the contribution of differential penetrance in population subgroups to missing heritability is likely to be, but is likely to be significant.

For biological processes with a known influence of sex, such as adiposity and plasma lipids, it is plausible to assume that the impact of genetic variation on these phenotypes may vary between sexes, with numerous demonstrations of this now available. In the Framingham Heart Offspring cohort, no variant showed genome wide significance for measures of obesity. However, sex dimorphism was evident with four polymorphisms in the lysophospholipase-like protein 1 (LYPLAL1) locus which encodes for a lipase/esterase in adipose tissue, having divergent effects in men and women⁽ Reference Benjamin, Suchindran and Pearce ^{31 )}. In a meta-analysis of the available GWAS, and using the waist:hip ratio as a measure of body fat topography, along with the LYPLAL1 signal, thirteen new loci for the waist:hip ratio were evident. Seven of these displayed dimorphism with a stronger effect in women⁽ Reference Heid, Jackson and Randall ^{32 )}. Using a candidate-gene approach we have reported a number of significant associations between common SNP and the postprandial lipaemic response, a CVD determinant of ever increasing prevalence⁽ Reference Jackson, Poppitt and Minihane ^{33 )}. For the leptin receptor (Gln223Arg, rs1137101) and APOA5 (−1131T > C, rs662799) variants, the effect of genotype was only evident in men⁽ Reference Olano-Martin, Abraham and Gill-Garrison ^{34 ,} Reference Jackson, Delgado-Lista and Gill ^{35 )}. For example, for leptin receptor, a 20% lower postprandial TAG response was evident in ArgArg v. GlnGln homozygotes with men and women combined, with a 35% difference in the men only group and no effect of genotype in women (Table 1)⁽ Reference Jackson, Delgado-Lista and Gill ^{35 )}.

Table 1.

Effect of the leptin receptor rs113701 genotype on postprandial lipaemia in UK adults (adapted from⁽ Reference Jackson, Delgado-Lista and Gill ^{35 )}) (Mean values with their standard errors)

* Values are group mean postprandial TAG area under the curves (mmol/l × 480 min) following consumption of test meals containing 49 g (0 min) and 29 g (330 min) total fat.

There is also evidence of racial/ethnic differences in the physiological impact of particular variants. Results from the Population Architecture Using Genomics and Epidemiology Consortium⁽ Reference Fesinmeyer, North and Ritchie ^{36 )} and a meta-analysis of forty-six individual GWAS⁽ Reference Teslovich, Musunuru and Smith ^{16 )}, which aimed to identify genome-wide signals for BMI and plasma lipids, respectively, showed a considerable overlap between associations in those of European and Asian ancestry, with more modest replication in more traditional populations such as African Americans and American Indians. Linkage disequilibrium patterns suggest that tagging SNP used in GWAS for Europeans may not adequately capture the genetic variation in other ethnic groups. Apparent ethnic differences in genotype × phenotype associations from GWAS may be in part attributable to differences in the habitual diets between populations.

Genome-wide association studies in the study of nutrigenetic interactions

Although GWAS methodology and modelling do not lend themselves well to the direct study of genotype × diet × disease associations, in part due to the fact that the sample size needed would be enormous⁽ Reference Thomas ^{37 )}, an increasing number of GWAS have nutrient status as their primary endpoint⁽ Reference Lemaitre, Tanaka and Tang ^{38 ,} Reference Wang, Zhang and Richards ^{39 )}. Using a genome-wide approach, Wang et al. identified three loci near the genes for cholesterol synthesis, hydroxylation and vitamin D transport as significant predictors of vitamin D status, which could potentially be used to set vitamin D intake recommendations⁽ Reference Wang, Zhang and Richards ^{39 )}. The output from GWAS has also informed the choice of variants for a more focused study of genotype × diet interaction in human epidemiology and intervention studies and in targeted replacement animal models. The identification of the FTO genotype by GWAS has led to a flurry of activity examining its impact on food intake, satiety and appetite and its interaction with macronutrient composition in determining BMI and risk of obesity⁽ Reference Tung and Yeo ^{10 ,} Reference McCaffery, Papandonatos and Peter ^{40 ,} Reference Razquin, Marti and Martinez ^{41 )}. In the RISCK randomised control trial, the impact of forty GWAS identified lipid associated SNP on the response of plasma lipids to a low-saturated fat diet were determined⁽ Reference Walker, Loos and Olson ^{42 )}. Relatively recent availability of GWAS data for cohorts for which participants have detailed dietary data, such as the Nurse's Health Study, Framingham Heart Studies and EPIC is likely to make a significant contribution to our nutrigenetic understanding in the near future.

To date, the majority of nutrigenetic information is derived from candidate-gene studies. Two of the most widely researched variants using this approach are APOE epsilon and MTHFR SNP, with approximately 6000 and 3000 associated published articles, respectively. However, despite extensive research focus there remains considerable uncertainty regarding the relative impact of these common genotypes on health and response to dietary change, which demonstrates the complexity of the interactions.

The renowned APOE epsilon genotype

As its name suggests apoE is an important modulator of many stages of lipoprotein metabolism and is the main lipid transporter in the central nervous system. Since its original identification, its pleiotropic nature has been realised with apoE now known to regulate immunity and inflammation, oxidative status and β-amyloid metabolism in the central nervous system. Two non-synonymous SNP in the APOE gene, result in three specific apoE protein isoforms namely apoE2, apoE3 and apoE4. The APOE genotype was originally described as a genetic contributor to CVD, with APOE4 carriers at increased risk. Over time, and with ever larger meta-analyses it has become apparent that at a population level the impact on CVD risk is marginal⁽ Reference Bennet, Di Angelantonio and Ye ^{43 ,} Reference Song, Stampfer and Liu ^{44 )}, and often does not emerge as a significant signal in GWAS (Table 2), although in individual population subgroups such as smokers, the APOE genotype remains a highly significant risk factor. In both the Northwick Park and Framingham Offspring cohorts, risk of CVD was about 2-fold higher in smokers who were wild-type E3/E3 v. E4 carriers (Table 3)⁽ Reference Humphries, Talmud and Hawe ^{46 ,} Reference Talmud, Stephens and Hawe ^{47 )}.

Table 2.

Meta-analysis of the impact of the APOE genotype on CVD and Alzheimer's disease risk

CAD, coronary artery disease; AD, Alzheimer's disease.

* E2 carriers include E2/E2 and E2/E3; E4 carriers include E3/E4 and E4/E4.

† OR (95% CI) with the wild-type E3/E3 genotype as the reference.

Table 3.

Impact of APOE genotype on CVD risk in smokers

* E2 carriers include E2/E2 and E2/E3; E4 carriers include E3/E4 and E4/E4.

† Hazard ratio (95% CI) with all genotypes combined, never smokers as reference.

‡ Hazard ratio (95% CI) with E3/E3 never smokers as reference.

More recently there has been much interest in the APOE genotype as a longevity gene. Although not fully consistent, the APOE4 allele has emerged as being associated with a shorter life-span⁽ Reference McKay, Silvestri and Chakravarthy ^{48 ,} Reference Murabito, Yuan and Lunetta ^{49 )}.

Perhaps the most consistent and consequential common genotype-disease association described to date is the impact of the APOE genotype on risk of age-related cognitive decline and Alzheimer's disease. As summarised in Table 2, the APOE3/E4 and APOE4/E4 individuals are at approximately 3–4- and 12–16-fold increased risk of Alzheimer's disease and have a much earlier age of onset⁽ Reference Bertram, McQueen and Mullin ^{45 )}. The clinical significance of the genotype is demonstrated by the fact that almost 50 and 10% of Alzheimer's disease patients are APOE3/E4 and APOE4/E4, whereas these genotype subgroups represent approximately 20 and 2%, respectively, of the general population⁽ Reference Minihane, Jofre-Monseny and Olano-Martin ^{50 ,} Reference Ward, Crean and Mercaldi ^{51 )}. Interestingly, when James D. Watson was presented with his genetic information, his being the first genome sequenced by next-generation sequence technologies, he elected not to know his APOE genotype⁽ Reference Wheeler, Srinivasan and Egholm ^{52 )}.

The aetiological basis of this association is likely to be multi-factorial with APOE4 carriers having altered central nervous system lipid metabolism, vascular dysfunction, increased neuroinflammation and oxidative stress, β-amyloid deposition, synaptic dysfunction and impaired neurogenesis⁽ Reference Hauser, Narayanaswami and Ryan ^{53 )}. This is of wide interest in the quest for further establishment of the Alzheimer's disease pathological process and its treatment and prevention.

The often cited methylenetetrahydrofolate reductase C677T variant

MTHFR is an important enzyme in folate/homocysteine metabolism. It provides a clear demonstration of how genetic information could be used to provide targeted dietary advice in an at-risk population subgroup. A homozygous mutant genotype (TT, rs 1801133), which has a frequency of approximately 10% worldwide⁽ Reference Wilcken, Bamforth and Li ^{58 )}, is associated with reduced enzyme activity⁽ Reference Frosst, Blom and Milos ^{59 )}. Its subsequent impact on homocysteine concentrations, blood pressure and risk of diseases such as cancer and CVD, is variable and has been shown to be dependent on factors such as sex and ethnicity⁽ Reference Holmes, Newcombe and Hubacek ^{60 ,} Reference Xuan, Bai and Gao ^{61 )}. The penetrance of the genotype is also dependent on vitamin B status (folate, riboflavin, vitamin B₆ and vitamin B₁₂)⁽ Reference Holmes, Newcombe and Hubacek ^{60 ,} Reference McNulty, Pentieva and Hoey ^{62 )}. In two complementary intervention trials, Scott and colleagues elegantly demonstrated that riboflavin (a co-factor of MTHFR) lowered blood pressure in patients with the MTHFR TT genotype which is independent of background use of prescribed drugs⁽ Reference Horigan, McNulty and Ward ^{63 ,} Reference Wilson, Ward and McNulty ^{64 )}. Overall, there is considerable evidence to indicate that adequate vitamin B status is likely to abrogate the negative physiological impact of this genotype on disease risk.

Conclusion: the road ahead

The first draft of the majority (about 90%) of the sequence of the human genome was published in a Nature article entitled ‘Initial sequencing and analysis of the human genome’ a little over a decade ago (February 2001)⁽ Reference Lander, Linton and Birren ^{65 )} with the complete sequence (about 99·7%) available in 2004^{( 66 )}. At the time of availability such information was considered by many to be the panacea and one of the greatest ever medical achievements. Ten years on, many consider progress based on the human genome to be limited, but as reviewed by Eric Lander (who was lead author of the original 2001 publication), this may be a rather harsh assessment⁽ Reference Lander ^{17 )}. Of the 3000 Mendelian (monogenic) disorders whose genetic basis is known, the locus of the vast majority have been identified since 2001. Furthermore, GWAS and HapMap approaches have led to the identification of 1600 variants associated with 250 traits, which has established numerous novel pathological pathways and contributed to the development of novel therapies. Although sufficient to indicate the potential of the technology, there has been limited success in the use of genetic information for disease prediction and personalisation of therapeutics or preventative advice, with much of the estimated heritable component of disease risk and response to diet unaccounted for. Based on the available information, it appears that rather than being overestimated, the heritability is dark matter (i.e. it is real but we cannot see it yet), attributable to as yet undetected rare variants, or the underestimation of the impact of known variants. The wider use of sequencing will provide information on variants with a frequency of <5%. More detailed and precise characterisation of study participants in genetic studies and more sophisticated modelling will undoubtedly lead to the detection of variants of particular importance in population subgroups. Most of the benefits of genetics in public health remain to be realised and we undoubtedly have a long way to go. In the words of Churchill, it feels like the ‘end of the beginning’ rather than the ‘beginning of the end’.