TY - JOUR
AU - Davey Smith,, George
AB - Mendelian randomization (MR),7 a genetic epidemiological approach, has made substantial inroads into our understanding of the causes, and consequences, of disease (1, 2). Conventionally, MR takes genetic variants associated with an exposure to estimate the causal effect on risk of disease. Availability of large-scale data and hypothesis-free genome-wide analyses has led to discovery of trait-associated genetic variants that surpass stringent thresholds for multiple testing correction, making genome-wide association study (GWAS) discoveries among the most reliable (i.e., among the least prone to false positives) in the scientific literature. A naïve criticism of GWAS is that the discoveries made therein have limited translational utility. Combining GWAS-identified single-nucleotide polymorphisms (SNPs) into genetic instruments under the MR approach, however, offers translational opportunities, yielding notable discoveries for public health and, in the longer-term, pharmaceutical development. More immediate translational opportunities include pharmacogenetics (3) and the emerging use of GWAS for disease prediction (4). In the case of MR, recent examples include identifying that genetic liability to diabetes causes erectile dysfunction (5) and that alcohol consumption may not protect from vascular disease (6). The elucidation of these would be seemingly unachievable under other study designs, e.g., the NIH-funded Moderate Alcohol and Cardiovascular Health trial (MACH15; NCT03169530) was abandoned because of perceived influence and bias from Big Alcohol (7). As with all scientific investigations, MR studies can be influenced by unique forms of bias, confounding, and analytical inaccuracies—what differentiates MR from conventional epidemiology is the availability of multiple sensitivity analyses (1, 8), enabling scientists to test many of the implicit assumptions. In this issue of Clinical Chemistry, Mohammadi-Shemirani and colleagues (9) have turned MR on its head. They propose taking a summary score of genetic variants robustly associated with estimated glomerular filtration rate (eGFR), a trait used to assess renal function, and relating this to a range of potential biomarkers. Taking 50 eGFR-associated SNPs in combination, they found an association with concentrations of the protein trefoil factor (TFF3). Importantly, cis-acting protein quantitative trait loci (pQTLs) did not demonstrate a causal effect of TFF3 on eGFR, suggesting that the biomarker identified was not a cause, but possibly a consequence of eGFR. A disease-associated genetic risk score (GRS) would be expected to associate with causes of the disease, consequences of the process leading to the disease, or the disease itself. For example, a GRS for lung cancer that included cholinergic receptor nicotinic alpha 5 subunit (CHRNA5)8 variants related to heaviness of smoking would be expected, therefore, to associate with smoking behavior. It is also possible that a GRS for a disease might identify biomarkers that are influenced by the process leading to disease or by the disease itself. In this case, if MR of the biomarker suggests that the biomarker does not cause disease (as in the study by Mohammadi-Shemirani and colleagues (9)), the biomarker may be a good candidate to investigate for disease prediction. What do biomarker associations of SNPs combined into a GRS that were originally identified for disease (or a precursor of disease) actually mean? We consider these to represent at least 7 scenarios (see Fig. 1), which we elucidate below and in the figure: note that we do not consider these to be exhaustive. Scenarios 1–4 represent “real” GRS-to-trait associations, whereas scenarios 5–7 represent potential artifacts owing to analytical method, study design, or diagnostic approach to disease. Potential scenarios of biomarker associations of a GRS for disease. Fig. 1. Open in new tabDownload slide The 7 scenarios are described in the main text. Purple cloud call-outs elucidate whether the biomarker is likely to have a role in disease prediction and/or whether it represents a potential therapeutic target for the treatment and/or prevention of disease. Fig. 1. Open in new tabDownload slide The 7 scenarios are described in the main text. Purple cloud call-outs elucidate whether the biomarker is likely to have a role in disease prediction and/or whether it represents a potential therapeutic target for the treatment and/or prevention of disease. First, the biomarker association of the GRS may represent a biomarker that causes disease. An example of this would be proprotein convertase subtilisin/kexin type 9 (PCSK9) genetic variants that associate with coronary heart disease (CHD) at GWAS significance, and which pinpoint LDL-cholesterol as being on the causal pathway to risk of CHD. A GRS-associated biomarker in this scenario may have predictive utility and could also represent a therapeutic target for disease prevention. Second, disease–GRS associations with a biomarker may represent a consequence of disease, akin to the conventional epidemiological phenomenon of “reverse causality.” For example, a GRS for CHD is associated with a higher likelihood of receiving statin therapy (analyzed in MR-Base (10) on December 8th, 2018). A naïve interpretation would be that statin therapy leads to a higher risk of CHD, but of course this is not the case: individuals with CHD are prescribed statin therapy as first-line treatment. Third, the biomarker association may arise as a result of vertical pleiotropy of a genetic variant or variants that associate with a biomarker (e.g., biomarker X in Fig. 1) that plays a causal role in disease, but where crucially the GRS-associated biomarker is not causally related to disease (biomarker X2 in Fig. 1). For example, genetic variation in interleukin 6 receptor (IL6R), which encodes the interleukin-6 (IL6) receptor (IL6R), is a cause of CHD. Altered activity of the IL6R pathway leads to perturbations in circulating concentrations of both IL6 (which binds to IL6R) and C-reactive protein (CRP; which IL6, acting through IL6R, modifies). A CHD GRS may identify CRP as associated in this situation as a result of such pleiotropy of the IL6R pathway—note that CRP is not a cause of CHD in this setting. In a similar way, biomarker associations owing to vertical pleiotropy of SNPs in the GRS representing pathways arising from disease would lead to a similar association (e.g., the association of a GRS with biomarker Z2 as a result of vertical pleiotropy through Z in Fig. 1). Fourth, a biomarker association from a GRS may represent horizontal pleiotropy (11) of ≥1 genetic variants that also associate with disease. In this situation, ≥1 genetic variants would associate with the identified biomarker, not from a pathway either leading to disease or arising from disease, but from another relationship altogether that is unrelated to disease. A biomarker identified in this scenario is unlikely to have any substantial role in prediction; it also does not represent a valid therapeutic target for disease prevention or treatment. Fifth, the association may arise owing to conditioning for a trait in the original GWAS (12). In using a GRS for type 2 diabetes (T2D), an inverse association with body mass index (BMI) may be identified, which is opposite to the well-recognized causal role of adiposity in T2D. However, conditioning on the original T2D GWAS for BMI would lead to selection of variants that, on average, increase liability to T2D but are associated with lower BMI (13). On regressing these SNPs onto phenotypes, an inverse association with BMI may be identified, purely as a result of the analytical model of the original T2D GWAS. Sixth, the association may be induced by selection biases affecting cases or controls. For example, if cases are identified from a screening program and controls are from the general population, selection biases into a screening program might identify associations of the GRS with socioeconomic factors such as years of education. Seventh, the biomarker leads to disease diagnosis, but the biomarker itself is not causal. For example, increased prostate-specific antigen (PSA) concentration is used to diagnose prostate cancer. Thus, a GRS of prostate cancer may associate with PSA even in the absence of a causal effect of PSA on prostate cancer. In the current article (9), the focus was on the application of a pre-disease GRS to identify novel biomarkers for prediction. As the authors note, the ideal biomarker for prediction is one that is independently and strongly associated with a disease and that makes meaningful improvements to disease prediction—i.e., it need not be causal. In this case, out of the 7 scenarios above, when might a biomarker be useful for prediction? Biomarker associations of a GRS arising from horizontal pleiotropy of a genetic instrument are unlikely to have clinical utility, which is also the case when the association is induced by adjusting for a trait in the discovery GWAS, or when there are selection biases present. The authors (9) tease apart potential sources of bias and, ultimately, are able to show that inclusion of TFF3 leads to modest improvements in disease prediction: the c-statistic increases from 0.59 to 0.60. Such measures of discrimination may be insensitive to small, but meaningful, increments in predictive utility (14). Comparatively, recent studies that incorporated genetic variants for prediction of CHD led to similar changes in area under the curve, from c-statistics of 0.67 for conventional risk factors to 0.70 with inclusion of 1.7 million genetic variants (4). When might combinations of measured phenotype and/or its genotypic variation be useful in prediction? When a biomarker itself is causally related to disease, phenotypic measurements are likely to be useful beyond genotype, because the measured trait will additionally capture environmental variation (as it is the combined genetic and environmental variation in the causal biomarker that ultimately leads to disease). In this same scenario, genotype will also be useful in addition to the measured phenotype as genotype represents a measure of lifelong exposure. When the biomarker is a consequence of disease (i.e., the association arises from reverse causality), measuring the trait is of importance as it gives a dynamic “read out” of disease status. Finally, when a biomarker is noncausal in disease, including genotype in the model may increase the predictive utility of the biomarker, through the principle of so-called “biomarker de-Mendelization,” by maximizing the nongenetic variation in the phenotype (15). With burgeoning technological advances in high-throughput phenotyping of omics down-stream of the invariant genome (notwithstanding CRISPR), unparalleled opportunities exist for repurposing and adapting the principles of MR to new approaches that may lead to novel prospects for the translation of GWAS discoveries. Such applications are expected to lead to advances in elucidating not only the causes and consequences of disease, but also, as in the current study, the potential for improvements in disease prediction. 7 Nonstandard abbreviations MR Mendelian randomization GWAS genome-wide association study SNP single nucleotide polymorphism eGFR estimated glomerular filtration rate TFF3 trefoil factor CHD coronary heart disease IL6 interleukin-6 IL6R interleukin-6 receptor CRP C-reactive protein T2D type 2 diabetes BMI body mass index PSA prostate-specific antigen. 8 Human Genes CHRNA5 cholinergic receptor nicotinic alpha 5 subunit PCSK9 proprotein convertase subtilisin/kexin type 9 IL6R interleukin 6 receptor. " (see article on page 427) " Author Contributions:All authors confirmed they have contributed to the intellectual content of this paper and have met the following 4 requirements: (a) significant contributions to the conception and design, acquisition of data, or analysis and interpretation of data; (b) drafting or revising the article for intellectual content; (c) final approval of the published article; and (d) agreement to be accountable for all aspects of the article thus ensuring that questions related to the accuracy or integrity of any part of the article are appropriately investigated and resolved. " Authors' Disclosures or Potential Conflicts of Interest:Upon manuscript submission, all authors completed the author disclosure form. Disclosures and/or potential conflicts of interest: " Employment or Leadership: None declared. " Consultant or Advisory Role: M.V. Holmes, Boehringer Ingelheim. " Stock Ownership: None declared. " Honoraria: None declared. " Research Funding: The Medical Research Council (MRC) and the University of Bristol fund the MRC Integrative Epidemiology Unit (MC_UU_00011/1, MC_UU_00011/3). M.V. Holmes works in a unit that receives funding from the UK Medical Research Council and is supported by a British Heart Foundation Intermediate Clinical Research Fellowship (FS/18/23/33512) and the National Institute for Health Research Oxford Biomedical Research Centre. " Expert Testimony: None declared. " Patents: None declared. References 1. Davies NM , Holmes MV, Davey Smith G. Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians . BMJ 2018 ; 362 : k601 . Google Scholar Crossref Search ADS PubMed WorldCat 2. Davey Smith G , Ebrahim S. ‘Mendelian randomization’: Can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol 2003 ; 32 : 1 – 22 . Google Scholar Crossref Search ADS PubMed WorldCat 3. Holmes MV , Davey Smith G. CETP inhibition and ADCY9 genotype: evidence of a qualitative pharmacogenetic interaction in cardiovascular disease? Preprint at https://www.biorxiv.org/content/10.1101/336875v1 ( 2018 ). 4. Inouye M , Abraham G, Nelson CP, Wood AM, Sweeting MJ, Dudbridge F. Genomic risk prediction of coronary artery disease in 480,000 adults: implications for primary prevention . J Am Coll Cardiol 2018 ; 72 : 1883 – 93 . Google Scholar Crossref Search ADS PubMed WorldCat 5. Bovijn J , Jackson L, Censin J, Chen CY, Laisk T, Laber S. GWAS identifies risk locus for erectile dysfunction and implicates hypothalamic neurobiology and diabetes in etiology . Am J Hum Genet 2019 ; 104 : 157 – 63 . Google Scholar Crossref Search ADS PubMed WorldCat 6. Holmes MV , Dale CE, Zuccolo L, Silverwood RJ, Guo Y, Ye Z. Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data . BMJ 2014 ; 349 : g4164 . Google Scholar Crossref Search ADS PubMed WorldCat 7. Dyer O . $100m alcohol study is cancelled amid pro-industry ”bias” . BMJ 2018 ; 361 : k2689 . Google Scholar Crossref Search ADS PubMed WorldCat 8. Hemani G , Bowden J, Davey Smith G. Evaluating the potential role of pleiotropy in Mendelian randomization studies . Hum Mol Genet 2018 ; 27 : R195 – R208 . Google Scholar Crossref Search ADS PubMed WorldCat 9. Mohammadi-Shemirani P , Sjaarda J, Gerstein HC, Treleaven DJ, Walsh JF, Mann JF. A Mendelian randomization-based approach to identify early and sensitive diagnostic biomarkers of disease . Clin Chem 2019 ; 65 : 427 – 36 . Google Scholar Crossref Search ADS PubMed WorldCat 10. Hemani G , Zheng J, Elsworth B, Wade KH, Haberland V, Baird D. The MR-base platform supports systematic causal inference across the human phenome . Elife 2018 ; 7 : e34408 . Google Scholar Crossref Search ADS PubMed WorldCat 11. Holmes MV , Ala-Korpela M, Davey Smith G. Mendelian randomization in cardiometabolic disease: challenges in evaluating causality . Nat Rev Cardiol 2017 ; 14 : 577 – 90 . Google Scholar Crossref Search ADS PubMed WorldCat 12. Holmes MV , Davey Smith G. Problems in interpreting and using GWAS of conditional phenotypes illustrated by ‘alcohol GWAS’ . Mol Psychiatry 2019 ; 24 : 167 – 8 . Google Scholar Crossref Search ADS PubMed WorldCat 13. Davey Smith G , Paternoster L, Relton C. When will Mendelian randomization become relevant for clinical practice and public health? JAMA 2017 ; 317 : 589 – 91 . Google Scholar Crossref Search ADS PubMed WorldCat 14. Cook NR . Statistical evaluation of prognostic versus diagnostic models: beyond the ROC curve . Clin Chem 2008 ; 54 : 17 – 23 . Google Scholar Crossref Search ADS PubMed WorldCat 15. Kjaergaard AD , Bojesen SE, Nordestgaard BG, Johansen JS, Davey Smith G. Biomarker de-Mendelization: principles, potentials and limitations of a strategy to improve biomarker prediction by reducing the component of variance explained by genotype . Preprint at https://www.biorxiv.org/content/10.1101/428276v1 ( 2018 ). © 2019 The American Association for Clinical Chemistry This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
TI - Can Mendelian Randomization Shift into Reverse Gear?
JF - Clinical Chemistry
DO - 10.1373/clinchem.2018.296806
DA - 2019-03-01
UR - https://www.deepdyve.com/lp/oxford-university-press/can-mendelian-randomization-shift-into-reverse-gear-N064mXRlfx
SP - 363
VL - 65
IS - 3
DP - DeepDyve
ER -