Mendelian Randomization (MR)

MR uses single nucleotide polymorphisms (SNPs) as proxies for modifiable exposures. A valid instrument must satisfy three core assumptions:

• Relevance (IV1): The genetic variant is robustly associated with the exposure (e.g., an eQTL for gene expression)
• Independence (IV2): The variant is not associated with confounders of the exposure-outcome relationship
• Exclusion Restriction (IV3): The variant affects the outcome only through the exposure, not via alternative pathways

Violation of these assumptions, particularly through horizontal pleiotropy, is the primary threat to MR validity.

The most widely implemented MR framework leverages summary-level data from independent GWAS consortia. This design extracts variant-exposure associations from one study and variant-outcome associations from another, enabling causal inference without individual-level data access.

Key advantages include:

• Massive statistical power by combining summary statistics from biobanks with hundreds of thousands of participants
• Privacy preservation as raw genotypes never leave their respective repositories
• Computational efficiency compared to individual-level regression

Tools like TwoSampleMR and MendelianRandomization R packages operationalize this approach.

Standard inverse-variance weighted (IVW) meta-analysis assumes no horizontal pleiotropy. Modern MR employs a suite of sensitivity analyses that relax this assumption:

• MR-Egger regression: Allows a non-zero intercept to detect and correct for directional pleiotropy, though at the cost of precision
• Weighted median estimator: Provides consistent causal estimates when at least 50% of the weight comes from valid instruments
• MR-PRESSO: Detects and removes outlier SNPs driving pleiotropic effects
• Contamination mixture method: Models the possibility that only a subset of variants are valid instruments

Triangulating evidence across these methods is standard practice.

Cis-Mendelian Randomization restricts genetic instruments to variants within or near the gene encoding a protein target, mimicking the specificity of a therapeutic intervention. This approach is foundational for drug target discovery and adverse event prediction.

Applications include:

• Predicting the on-target effects of PCSK9 inhibition on cardiovascular outcomes before phase III trials
• Identifying repurposing opportunities by linking genetically proxied protein levels to disease endpoints in pQTL studies
• Anticipating safety liabilities by testing genetically proxied target perturbation against a phenome-wide scan of health outcomes

This framework directly informs portfolio decisions in pharmaceutical R&D.

MR can disentangle complex causal networks through extended designs:

• Bidirectional MR: Tests the direction of causality between two correlated traits by performing MR in both directions, using instruments for each trait. This resolves whether trait A causes trait B, or vice versa
• Mediation MR (Two-Step MR): Decomposes the total causal effect of an exposure on an outcome into a direct effect and an indirect effect mediated through an intermediate molecular phenotype

For example, mediation MR can determine whether a genetic predisposition to obesity increases cardiovascular risk solely through elevated blood pressure, or through additional independent pathways.

A critical post-MR step to distinguish causality from linkage disequilibrium (LD) contamination. Colocalization tests whether the same causal variant drives both the exposure and outcome association signals.

Bayesian colocalization (e.g., coloc R package) computes posterior probabilities for competing hypotheses:

• H4: Both traits share a single causal variant (evidence for a causal mechanism)
• H3: Distinct causal variants for each trait (suggesting LD confounding, not causality)

A high posterior probability for H4 (>0.8) combined with a significant MR result provides strong evidence that the genetic association reflects a shared causal variant rather than residual confounding from correlated markers.