Mendelian Randomization

Mendelian randomization (MR) is an analytical method in genetic epidemiology and human statistical genetics that uses inherited germline genetic variants—typically single-nucleotide polymorphisms (SNPs) identified in genome-wide association studies (GWAS)—as instrumental variables (IVs) to estimate the causal effect of a modifiable biological exposure X (such as plasma LDL cholesterol, C-reactive protein, body mass index, or circulating protein levels) on a clinical disease outcome Y (such as coronary artery disease, type 2 diabetes, or Alzheimer disease). First proposed conceptually by Martijn Katan (1986) to test whether low serum cholesterol causes cancer and formalized econometrically by George Davey Smith and Shah Ebrahim (2003), Mendelian randomization exploits Gregor Mendel's laws of segregation and independent assortment: because parental alleles are randomly allocated to offspring at meiosis and fixed at conception, germline genotypes are largely independent of postnatal socioeconomic, behavioral, and environmental confounders (U), and cannot be altered by subsequent disease onset (immunity to reverse causation).

By drawing a formal mathematical analogy between random allocation of alleles at conception and random treatment assignment in a randomized controlled trial (RCT), Mendelian randomization has become a cornerstone of human target validation—correctly predicting both the clinical triumph of PCSK9 and NPC1L1 inhibitors alongside statins for coronary heart disease and the clinical failure of CETP inhibitors (torcetrapib), C-reactive protein (CRP) lowering, and antioxidant vitamin supplementation.

In the directed acyclic graph (DAG) of Mendelian randomization, a genetic variant (or set of J independent variants) G_j serves as a valid instrumental variable for testing the causal effect β_XY of an exposure X on an outcome Y in the presence of unmeasured confounders U if and only if three core assumptions hold:

*Note on Vertical vs. Horizontal Pleiotropy: When a genetic variant alters gene expression → alters plasma LDL-C (X) → causes myocardial infarction (Y), this cascade is vertical (mediated) pleiotropy, which is the exact causal pathway MR is designed to measure. By contrast, when a variant independently alters both X and a second pathway Z that affects Y (G → X → Y and G → Z → Y), that bypass is horizontal pleiotropy, which violates IV3.*

In two-sample Mendelian randomization, summary association statistics for the SNP–exposure effect (β̂_Xj, standard error σ_Xj) are extracted from one large GWAS consortium (for example, the Global Lipids Genetics Consortium or UK Biobank proteomics), while SNP–outcome summary statistics (β̂_Yj, standard error σ_Yj) for the same LD-clumped independent instruments (r² < 0.001) are extracted from a second, non-overlapping case-control GWAS consortium (for example, CARDIoGRAMplusC4D):

1. Single-SNP Wald Ratio Estimator: For an individual genetic instrument G_j, because the total genetic effect on the outcome equals the product of its effect on the exposure times the causal effect of the exposure on the outcome (β_Yj = β_XY · β_Xj when horizontal pleiotropy α_j = 0), the causal estimate is the Wald ratio:

1. Inverse-Variance Weighted (IVW) Estimator (Burgess et al., 2013): Combining J independent genetic instruments by weighting each Wald ratio by its inverse variance is mathematically equivalent to a weighted linear regression of β̂_Yj on β̂_Xj forced through the origin (intercept = 0) with weights w_j = β̂_Xj² / σ_Yj²:

IVW achieves maximum statistical power when all J instruments are valid (or when horizontal pleiotropy α_j is balanced with mean zero), and instrument heterogeneity is audited via Cochran's Q statistic: Q = ∑ w_j (β̂_XY,j − β̂_IVW)² ∼ χ²_J−1.

1. MR-Egger Regression (Bowden et al., 2015): Removes the zero-intercept constraint of IVW, fitting the weighted regression β̂_Yj = α₀ + β̂_Egger β̂_Xj (with alleles oriented so all β̂_Xj > 0). Under the InSIDE assumption (Instrument Strength Independent of Direct Effect: Cov(α_j, β_Xj) = 0), a non-zero intercept (α₀ ≠ 0) detects directional horizontal pleiotropy, while the slope β̂_Egger provides a consistent causal estimate even if 100% of instruments have pleiotropic direct effects.
2. Weighted Median Estimator (Bowden et al., 2016): Computes the median of the empirical distribution of single-SNP Wald ratios β̂_XY,j weighted by normalized inverse-variance weights w_j / ∑w_k. Unlike IVW (which has a 0% breakdown point—a single invalid outlier SNP biases the estimate), the weighted median estimator remains statistically consistent as long as at least 50% of the total instrument weight comes from valid (non-pleiotropic) SNPs.

Pharmaceutical genetics distinguishes sharply between two experimental designs of Mendelian randomization:

• Polygenic Mendelian Randomization: Uses dozens or hundreds of genome-wide significant SNPs scattered across the genome (trans and cis loci) to instrument a complex polygenic biomarker (such as BMI, HDL cholesterol, or educational attainment). Because distal variants act through diverse pathways, polygenic MR requires mandatory sensitivity triangulation across IVW, MR-Egger, Weighted Median, and MR-PRESSO outlier filtering.
• Drug-Target cis-Mendelian Randomization (cis-eQTL / cis-pQTL MR): Restricts genetic instruments strictly to variants located within or immediately flanking (±100–500 kb) the structural gene encoding a specific druggable protein target (for example, loss-of-function or missense variants in PCSK9, HMGCR, NPC1L1, IL6R, or TYK2), combined with Bayesian statistical colocalization (coloc PP_H4 > 0.80) to prove that the protein level (pQTL) and disease (GWAS) share the exact same underlying causal variant rather than two distinct variants in linkage disequilibrium.

Why did Mendelian Randomization show that raising HDL ("good") cholesterol does not prevent heart attacks, whereas lowering LDL ("bad") cholesterol does?

For decades, observational epidemiology showed a strong inverse correlation between plasma high-density lipoprotein cholesterol (HDL-C) and myocardial infarction, leading pharmaceutical companies to invest billions of dollars in CETP inhibitors (such as torcetrapib, dalcetrapib, and evacetrapib) designed to raise HDL-C. In a landmark 2012 Mendelian randomization study (Voight et al., The Lancet), researchers examined germline variants in endothelial lipase (LIPG Asn396Ser) and polygenic HDL-raising alleles that exclusively elevate plasma HDL-C without altering LDL-C or triglycerides:

• Carriers of lifelong genetically elevated HDL-C showed zero reduction in myocardial infarction risk (OR = 0.99 per 1 SD higher HDL-C, p = 0.85), proving that observational HDL-C associations were confounded by metabolic syndrome, physical activity, and triglyceride-rich remnant lipoproteins.
• Conversely, germline variants in LDLR, PCSK9, APOB, HMGCR (the target of statins), and NPC1L1 (the target of ezetimibe) that lower LDL-C (and apolipoprotein B particle number) reduced coronary heart disease risk by ~50%–54% per 1 mmol/L (38.7 mg/dL) lifelong reduction in LDL-C—more than double the ~22% reduction achieved per 1 mmol/L in a 5-year statin trial because germline alleles reduce cumulative arterial plaque burden across an entire lifetime.

Why did C-Reactive Protein (CRP) fail as a causal target in heart disease while the Interleukin-6 Receptor (IL-6R) succeeded?

Observational studies consistently show that elevated plasma C-reactive protein (CRP)—an acute-phase pentraxin synthesized by the liver in response to inflammatory cytokines—strongly predicts future myocardial infarction. However, cis-Mendelian randomization using variants inside the CRP gene locus itself showed zero causal effect of CRP on coronary heart disease (OR = 1.00): CRP is a downstream biomarker (spectator) of inflammation, not a causal driver of atherothrombosis. In contrast, cis-Mendelian randomization of the upstream IL6R (Interleukin-6 receptor) Asp358Ala variant—which impairs classical membrane IL-6R signaling and lowers downstream CRP—demonstrated a significant causal reduction in coronary heart disease risk, identifying the upstream NLRP3 → IL-1β → IL-6 signaling axis (validated clinically by the CANTOS trial of canakinumab) as the true causal therapeutic target.

1. Davey Smith, G., & Ebrahim, S. (2003). "'Mendelian randomization': Can genetic epidemiology contribute to understanding environmental determinants of disease?" International Journal of Epidemiology, 32(1), 1–22. https://doi.org/10.1093/ije/dyg070
2. Voight, B. F., Peloso, G. M., Orho-Melander, M., et al. (2012). "Plasma HDL cholesterol and risk of myocardial infarction: A mendelian randomisation study." The Lancet, 380(9841), 572–580. https://doi.org/10.1016/S0140-6736(12)60312-260312-2)
3. Bowden, J., Davey Smith, G., & Burgess, S. (2015). "Mendelian randomization with invalid instruments: Effect estimation and bias detection through Egger regression." International Journal of Epidemiology, 44(2), 512–525. https://doi.org/10.1093/ije/dyv080
4. Bowden, J., Davey Smith, G., Haycock, P. C., & Burgess, S. (2016). "Consistent estimation in Mendelian randomization with some invalid instruments using a weighted median estimator." Genetic Epidemiology, 40(4), 304–314. https://doi.org/10.1002/gepi.21965