Author + information
- ↵*Reprint requests and correspondence:
Dr. Christopher Newton-Cheh, Center for Human Genetic Research, Cardiovascular Research Center, Harvard Medical School, Massachusetts General Hospital, 185 Cambridge Street, CPZN 5.242, Boston, Massachusetts 02114
Genetic association studies have recently identified many genetic variants that are common in the general population and associated with prevalent cardiovascular diseases and traits. The effects of such variants are often quite modest, with odds ratios of <1.5 for risk alleles. Individually, these variants typically explain little of the population variation in disease risk, although in aggregate they might explain as much as 5% to 15% of variation and thus could potentially have a role in clinical risk prediction. Variants with much larger effects have also been identified from family-based studies, but each is typically private to the family in which it arises, and hence very rare at the population level.
Even if the modest effect of a given variant turns out to be of no value for risk prediction, its implication of a gene previously unrecognized to influence a trait can open new avenues for basic investigation and potentially novel therapeutics. For example, the minor allele of the single nucleotide polymorphism (SNP) rs3846663 has been associated with a trivial 2.5-mg/dl higher low-density lipoprotein cholesterol (LDL-C) in a genome-wide association study (1). If one did not already know that 3-hydroxy-3-methylglutaryl-coenzyme A reductase is a rate-limiting enzyme in cholesterol synthesis (inhibited by statins), then finding this SNP in an intron of the HMGCRgene could allow identification of this “novel” target for cholesterol-lowering therapy.
There is yet another use to which genetic associations can be put, namely to rule in or out the involvement of a clinical trait as a causal factor in the development of disease. Observational studies can identify a factor associated with clinical outcomes but cannot establish whether the factor is causal. For example, the presence of yellow fingers (found in cigarette smokers) might be associated with lung cancer or increased LDL-C with ischemic heart disease. Abrogation of the effect with adjustment for potential environmental confounders such as cigarette use suggests a noncausal relationship, whereas robustness increases the likelihood of a causal connection. However, establishment of a causal relation generally requires an interventional trial, such as testing the ability of coloring fingers to reduce lung cancer risk or of administering statins to lower cardiovascular risk. Obviously, this is only possible for modifiable causal risk factors, and clinical trials are expensive, lengthy, and might prove hazardous to the individuals undergoing intervention. That we can now identify genetic variants that impact common, complex traits (e.g., LDL-C) and diseases (e.g., myocardial infarction) provides us with an alternate means to test for causality. Central to this concept, termed Mendelian randomization, is that genetic variants are randomly assigned at gametogenesis, following the laws of inheritance as described by Gregor Mendel, and hence are not influenced by other traits under study (2). Thus, we can use genotype as an instrumental variable, not subject to confounding, to elucidate the relationship between 2 associated variables.
First, the effect of genotype on a putative causal trait such as yellow fingers or LDL-C is estimated. Next, the effect of the putative causal trait on the outcome such as lung cancer or coronary heart disease (CHD) is estimated. With these 2 estimates in hand, one can calculate the expected risk of disease conferred by genotype under the model in which the intermediate factor lies in the causal pathway. Finally, the actual effect of genotype on the outcome variable is determined. If no association is observed in a sufficiently powered sample, it might then be concluded that a causal association between yellow fingers and lung cancer is unlikely. Contrast an imaginary SNP that causes yellow skin color to a SNP near genes encoding nicotinic acid receptors that influences tobacco use and risk of lung cancer (3).
Another example involves a recent report that examined the relationship of the G allele of rs7553007, a SNP in the CRPgene, which is reproducibly associated with higher C-reactive protein (CRP). In approximately 28,000 CHD cases compared with approximately 100,000 control subjects, the same allele was not associated with higher cardiovascular risk, arguing against the direct causal relationship between CRP and CHD that has previously been proposed (4). One concern of negative Mendelian randomization studies is the statistical power to exclude an effect; the large sample sizes examined in the CRP study allay such fears. However, such studies as the CRP study cannot exclude a weaker causal effect of CRP on ischemic heart disease, for which even larger studies would be needed.
In the same study, despite a strong association with higher CRP of the major A allele of SNP rs4420638 at the APOEgene, this allele was associated with lower CHD risk, a finding most consistent with the known strong association of the allele with decreased LDL-C (4,5). This example highlights one of the potential pitfalls of Mendelian randomization approaches: genetic variation might be associated with pleiotropic effects that can lead to the false inference that an intermediate factor is causal as well as to the false rejection of causality.
The Mendelian randomization approach has also been applied to several other risk factors for cardiovascular disease, including positive results for LDL-C (6,7), lipoprotein(a) [Lp(a)] (8,9), and natriuretic peptides (10) and negative results for fibrinogen (11), homocysteine (12), and high-density lipoprotein cholesterol (13,14). The latter 2 results mirror negative clinical trials of vitamin supplementation (15) and cholesteryl ester transfer protein inhibition (16), respectively.
The causal relationship of LDL-C to cardiovascular risk is well-supported by the continuous, graded relationship of LDL-C with CHD risk in observational studies, the existence of Mendelian syndromes such as familial hypercholesterolemia with increased risk of CHD, and the consistent effect of pharmacologic LDL-C–lowering on CHD risk. If one still did not believe that LDL-C is a causal factor influencing CHD, a report in this issue of the Journalcould help convince the persistent skeptic (17).
Benn et al. (17) studied the amino acid altering R46L SNP (rs11591147) of the proprotein convertase subtilisin/kexin type 9 (PCSK9) gene. The L allele of this SNP has previously been associated with decreased LDL-C levels and decreased risk of CHD, as first reported by Cohen et al. (6) and later replicated by the Myocardial Infarction Genetics Consortium (18). Furthermore, rare gain-of-function variants in PCSK9have been found to cause familial hypercholesterolemia and increased risk of myocardial infarction. The minor 46L allele is relatively uncommon, with a frequency of approximately 1% to 3% in populations of European descent, but exerts a large effect on LDL-C, with a reduction of 9% to 16% in heterozygotes (0.35 to 0.57 mmol/l). Benn et al. (17) performed a Mendelian randomization study in 3 large samples from Denmark for a total sample size of 1,204 46L carriers and 45,699 noncarriers. They observed 0.43 mmol/l (13%) lower LDL-C in heterozygotes and an odds ratio (OR) of 0.70 (95% confidence interval: 0.58 to 0.86) for CHD, which is lower than the predicted OR of 0.95. Added to previous studies through meta-analysis, an overall reduction in LDL-C of 0.43 mmol/l (12%) and an OR of 0.72 (95% confidence interval: 0.62 to 0.84) for CHD were observed.
The stronger genetic effect compared with that predicted from each limb of a SNP-to-low-density lipoprotein (LDL) and LDL-to-CHD model could reflect, as suggested by Benn et al. (17) as well as others, the impact of lifelong exposure to genetically-determined lower LDL-C levels by contrast to the shorter time horizon of estimates from epidemiologic studies. In addition, imprecision in measurement and trait changes over time can lead to underestimation of each limb of the putative causal pathway and could result in a spurious mismatch between predicted and observed effects. Accordingly, Benn et al. (17) used regression dilution adjustment, on the basis of correlation of repeated measurements of LDL-C, to adjust the predicted effect.
As the authors point out, the discrepancy could also result from non-LDL mechanisms, in a fashion similar to the relationship of the APOESNP to both CRP and LDL-C. The very specification of the intermediate phenotype could also be an alternative explanation. LDL-C is typically (although not in the current report) measured in the fasting state to allow its more accurate estimation from total and high-density lipoprotein cholesterol measurements and to avoid acute effects of variability in dietary intake of fat and cholesterol. However, we spend only a portion of our lives in the fasting state, and it is presumably not only fasting LDL-C that contributes to atherosclerosis. If genetic variants are identical in their effect on fasting LDL-C but variable in their effect on nonfasting LDL-C, then effects on cardiovascular disease could be quite different than predicted on the basis of the fasting LDL-C effect. Could the stronger effect of the PCSK9variant on CHD then be an effect of its “day-long” exposure to altered LDL-C as well as the “lifelong” effect? Ultimately, refinement of phenotype to most closely capture the pathophysiology of human disease is a central aim of medicine.
Obviously, Mendelian randomization studies could never be the sole motivation for drug treatment. Clinical trials must always be the guide, given the previously described limitations and the possibility of off-target toxicity and other pleiotropic effects and the constant risk that our models of human disease are incomplete. Genetic studies can, however, be a powerful means to identify and prioritize potential therapeutic targets.
So what can we conclude at the moment regarding causal risk factors for CHD from such experiments? The present evidence indicates that, whereas CRP, homocysteine, and fibrinogen are secure as risk factors for CHD, they seem more like yellow fingers: noncausal but integrated markers of underlying vascular disease or inflammation (19). The current report by Benn et al. (17) and a wealth of prior observational and interventional studies support the role of LDL-C as a causal and modifiable risk factor for CHD. Evolving evidence supports a causal role for Lp(a), although studies of interventions that isolate Lp(a) effects are lacking. Interpretation of future Mendelian randomization studies will need to consider the strengths and weaknesses of the approach, but there is clearly much to be learned from human genetics.
Dr. Newton-Cheh is on the scientific advisory board for hypertension and heart failure at Merck.
↵* Editorials published in the Journal of the American College of Cardiologyreflect the views of the authors and do not necessarily represent the views of JACCor the American College of Cardiology.
- American College of Cardiology Foundation
- Davey Smith G.,
- Ebrahim S.
- Keavney B.,
- Danesh J.,
- Parish S.,
- et al.
- Lewis S.J.,
- Ebrahim S.,
- Davey Smith G.
- Nainggolan L.
- Benn M.,
- Nordestgaard B.G.,
- Grande P.,
- Schnohr P.,
- Tybjærg-Hansen A.
- Danesh J.,
- Pepys M.B.