What Is Mendelian Randomization in Drug Discovery?

See why it matters for target choice

Most evidence linking a gene to a disease is association: the gene shows up more often in patients. Association cannot tell you direction. The gene might drive the disease, the disease might change the gene's activity, or a third factor might move both. Commit a programme to an association that runs the wrong way, and the target fails in the clinic for reasons no assay caught.

Mendelian randomization addresses this because the variant is fixed at conception, long before disease onset. Reverse causation cannot run backward in time, and most confounders are balanced across carriers and non-carriers. What remains is a causal estimate: if you perturb this target, does the disease move.

The payoff is measurable. Targets with human genetic support are roughly twice as likely to reach approval. Nelson and colleagues first showed this in 2015, finding that the share of drug mechanisms with direct genetic support rose from 2.0% at the preclinical stage to 8.2% among approved drugs. King and colleagues confirmed the roughly two-fold effect in 2019, and Minikel and colleagues refined it again in 2024. Genetic causal evidence is the closest thing target discovery has to a leading indicator of clinical success.

Limitations of Mandelian Randomization

Proper use of the method means naming its limits.

• Horizontal pleiotropy: the variant might reach the disease through some other route, not through the gene you are testing. The first guard is to use only variants that sit at the gene and control its own expression, so an effect on the disease almost has to run through that gene. Two further checks catch what slips past. One looks at the direction of effect and drops any variant that tracks the disease more closely than it tracks the gene, because that variant is probably acting somewhere else. The other compares the several variants used for one gene: if they tell a consistent story the estimate holds, and if they disagree more than chance allows, that disagreement is the tell-tale sign of a hidden second pathway, and the outliers are removed.
• Instrument strength: a weak genetic effect gives a noisy estimate.
• Lifelong versus acute: the method reflects a lifetime of small exposure, not a drug given for months. It speaks to whether a target is causal, less to dose or timing.
• Direction: separating cause from reverse cause takes care, and methods such as Steiger filtering help.
• Context: a genome-wide estimate does not tell you which tissue or cell type the effect runs through.

That last point matters most for target choice, and it is where Mendelian randomization on its own stops.

See how Disease Atlas uses Mendelian randomization

Disease Atlas by Euretos treats Mendelian randomization as its strongest evidence layer, not its only one. The platform ranks the protein-coding genome against each disease, and the causal layer built on Mendelian randomization sits at the top of that ranking because it answers the causal question directly. Around it sit complementary layers: perturbation responses, network context, single-cell expression, and a foundation model that integrates the signal. Candidates rise when these layers converge, independent of how often a gene appears in the literature.

In Disease Atlas the exposure is the gene's own expression. The instruments are expression quantitative trait loci (eQTLs), the common variants that tune how much a gene is transcribed, drawn from GTEx and eQTLGen, and the outcome is disease risk from large human cohorts such as UK Biobank and FinnGen. Because eQTLs are measured tissue by tissue, the causal estimate carries a tissue context with it, which is part of how the ranking resolves to the cell types where the disease operates. Where a gene has no strong eQTL instrument, the causal layer does not force an estimate. It leans on the perturbation evidence instead, and the scorecard shows which support a given rank rests on.

Two choices make the difference. First, every layer is built from primary biological data, genetic effect sizes, perturbation magnitudes, single-cell expression, rather than from claims extracted out of papers. Second, the result is resolved to the cell types where the disease operates, which is the context a bare genetic estimate cannot give you. The researcher moves from a causal signal to a ranked, cell-type-resolved, defensible shortlist in a single session.