Mendelian Randomization

The foundation of MR rests on selecting genetic variants that serve as proxies for the exposure. A valid instrument must satisfy three core assumptions:

• Relevance (IV1): The variant must be robustly associated with the exposure (e.g., an SNP in the PCSK9 gene reliably lowers LDL cholesterol).
• Independence (IV2): The variant must not be associated with confounders of the exposure-outcome relationship.
• Exclusion Restriction (IV3): The variant must affect the outcome only through the exposure, with no horizontal pleiotropy.

Genome-wide association studies (GWAS) provide the catalog of SNP-exposure associations that power MR analyses.

MR exploits Mendel's law of independent assortment, which states that alleles are randomly distributed at gamete formation, independent of environmental and behavioral confounders. This genetic randomization occurs at conception, decades before disease onset.

Because genotype is fixed at birth and unaffected by disease progression, MR is immune to reverse causation—the problem where disease status influences the exposure rather than vice versa. This temporal clarity is a decisive advantage over cross-sectional observational studies.

In two-sample MR, summary-level genetic associations with the exposure and outcome are extracted from separate, non-overlapping GWAS datasets. This design dramatically expands the scope of testable hypotheses.

Key advantages include:

• Leveraging large, publicly available GWAS summary statistics (e.g., UK Biobank, FinnGen).
• Avoiding the need for individual-level data with exposure, outcome, and genetics measured in the same cohort.
• Enabling rapid, high-throughput causal screening across thousands of molecular exposures (proteins, metabolites, transcripts) against hundreds of diseases.

Horizontal pleiotropy—where a genetic variant influences the outcome through pathways independent of the exposure—violates the exclusion restriction and biases MR estimates. A suite of robust methods has been developed to detect and correct for this:

• MR-Egger regression: Allows an intercept term to capture average directional pleiotropy, providing a bias-corrected causal estimate under the InSIDE (Instrument Strength Independent of Direct Effect) assumption.
• Weighted median estimator: Provides consistent estimates even when up to 50% of the weight comes from invalid instruments.
• MR-PRESSO: Detects and removes outlier SNPs that contribute disproportionately to heterogeneity.
• Contamination mixture method: Models the causal effect as a mixture distribution, identifying a cluster of valid instruments.

Cis-Mendelian randomization restricts genetic instruments to variants within or near the gene encoding a protein drug target. This localized approach mimics the specific action of a therapeutic agent.

Cis-MR is a cornerstone of drug target Mendelian randomization, used by pharmaceutical companies to:

• Predict the efficacy and safety of modulating a target protein before initiating costly clinical trials.
• Identify potential on-target adverse effects by testing the protein-disease relationship against a phenome-wide scan.
• Prioritize targets with genetic evidence, which are twice as likely to succeed in Phase II/III trials.

Example: Cis-MR using HMGCR variants (target of statins) correctly predicts LDL-cholesterol lowering and cardiovascular risk reduction.

Observational correlations cannot distinguish whether exposure causes outcome or outcome causes exposure. Bidirectional MR addresses this by performing two independent analyses:

1. Forward MR: Instruments for the exposure test its causal effect on the outcome.
2. Reverse MR: Instruments for the outcome test its causal effect on the exposure.

This framework is essential for untangling complex metabolic relationships. For instance, bidirectional MR has clarified that higher BMI causally increases type 2 diabetes risk, while the reverse pathway (diabetes causing obesity) shows minimal evidence. It also resolves whether biomarkers are causal mediators or merely reactive bystanders.

A variable used in causal inference to estimate the effect of an exposure on an outcome when unobserved confounding is present. For a valid IV, three core assumptions must hold: relevance (the IV is associated with the exposure), independence (the IV is not associated with confounders), and exclusion restriction (the IV affects the outcome only through the exposure). In Mendelian Randomization, single nucleotide polymorphisms (SNPs) serve as the instruments, leveraging the random assortment of alleles at conception to mimic a natural randomized controlled trial.

A hypothesis-free study design that scans the entire genome for single nucleotide polymorphisms (SNPs) statistically associated with a specific trait or disease. GWAS summary statistics provide the foundational exposure data for two-sample Mendelian Randomization. Key concepts include:

• Manhattan plot: Visualization of -log10(p-values) across chromosomes
• Linkage disequilibrium (LD): Non-random correlation of alleles at different loci, requiring careful clumping to select independent instruments
• Population stratification: Systematic ancestry differences between cases and controls that must be corrected to avoid spurious associations

A graphical model representing causal assumptions about the relationships between variables. Nodes represent variables, and directed edges (arrows) represent causal effects. DAGs are essential for:

• Identifying confounders: Variables that cause both exposure and outcome
• Detecting collider bias: Conditioning on a common effect of two variables can induce spurious associations
• Visualizing the IV assumptions: The DAG explicitly encodes that the genetic instrument affects the outcome only through the exposure, with no backdoor paths through confounders

The phenomenon where a single genetic variant influences multiple phenotypic traits through independent biological pathways. In Mendelian Randomization, pleiotropy is the primary threat to the exclusion restriction assumption. Two types are distinguished:

• Horizontal pleiotropy: The variant affects the outcome through a pathway independent of the exposure, directly violating the IV assumptions and biasing causal estimates
• Vertical pleiotropy: The variant affects the outcome through a mediator downstream of the exposure, which does not violate IV assumptions but requires careful interpretation

Methods like MR-Egger regression and the weighted median estimator are designed to be robust to horizontal pleiotropy.

A study design where the genetic variant-exposure associations and variant-outcome associations are estimated from two non-overlapping samples. This approach dramatically expands the scope of feasible analyses by leveraging publicly available GWAS summary statistics. Key considerations include:

• Sample overlap: Overlapping individuals between exposure and outcome GWAS can bias estimates toward the confounded observational association
• Harmonization: Ensuring the effect alleles are aligned across datasets, resolving palindromic SNPs with ambiguous strand orientation
• Population matching: Both samples should be drawn from similar ancestral populations to avoid bias from differing linkage disequilibrium patterns

A statistical method that assesses whether two traits share the same causal genetic variant at a given locus, distinguishing a shared causal mechanism from distinct variants in linkage disequilibrium. Colocalization strengthens MR findings by testing whether the genetic associations for exposure and outcome are driven by the same SNP. The approach computes posterior probabilities for competing hypotheses:

• H3: Two distinct causal variants (no colocalization)
• H4: A single shared causal variant (colocalization)

A high posterior probability for H4 provides evidence that the exposure-outcome relationship is causal rather than confounded by LD.