tooluniverse-mendelian-randomization

What this skill does

# Mendelian Randomization (Causal Inference from Genetic Instruments)

**MR estimates the CAUSAL effect of an exposure on an outcome using genetic variants as instrumental variables.** Because alleles are randomized at conception, MR is largely robust to the confounding and reverse causation that bias observational associations. It is *not* a free lunch: the causal claim rests on three assumptions, and violating them (especially horizontal pleiotropy) silently biases the estimate.

**LOOK UP, DON'T GUESS:** never assert a causal MR estimate from memory. Genetic-instrument results are updated as new GWAS are published — always retrieve current evidence with `EpiGraphDB_get_mendelian_randomization`. Do not invent beta/p-values.

**Correlation ≠ causation, and genetic correlation ≠ causation.** A high genetic correlation (`rg`) means two traits share heritability — it does NOT establish a causal direction. Only MR (with valid instruments) speaks to causality. Report them as different kinds of evidence.

## The three instrumental-variable assumptions

| Assumption | Statement | How it fails | Check |
|---|---|---|---|
| **Relevance** | Instrument is robustly associated with the exposure | Weak instruments (low F-stat) → bias toward the confounded observational estimate | MOE score; instruments selected at GWAS significance |
| **Independence** | Instrument shares no common cause with the outcome | Population stratification, assortative mating | Ancestry-matched GWAS; report population |
| **Exclusion restriction** | Instrument affects the outcome ONLY through the exposure | **Horizontal pleiotropy** — the variant influences the outcome via another path | MR-Egger intercept ≈ 0; agreement across methods |

If you cannot speak to these, your causal claim is provisional. Say so.

## When to use

- "Does **[exposure]** causally affect **[outcome/disease]**?" — the core MR question.
- Triangulating an observational/epidemiological association ("BMI correlates with depression — is it causal?").
- Reverse-causation checks (bidirectional MR: does the outcome cause the exposure instead?).
- Prioritising drug targets / risk factors with genetic causal support.
- Distinguishing a causal driver from a shared-etiology bystander (MR vs genetic correlation).

This skill wraps the **IEU OpenGWAS / EpiGraphDB MR-EvE** ("MR Everything-vs-Everything") resource: a large matrix of pre-computed two-sample MR results between GWAS traits. It does **not** run a bespoke two-sample MR from raw summary statistics with your own instrument set — see *Limitations*.

## Anchor tools

| Tool | Purpose |
|---|---|
| `EpiGraphDB_search_opengwas` | Resolve a free-text trait to exact OpenGWAS study IDs + labels (DO THIS FIRST) |
| `EpiGraphDB_get_mendelian_randomization` | Pre-computed MR estimate(s) for an exposure→outcome trait pair (curated pairs; start here) |
| `OpenGWAS_get_mr_instruments` | Custom two-sample MR: fetch the exposure's clumped instruments + their harmonized outcome effects for *any* GWAS pair (needs a free `OPENGWAS_JWT`). Use when the pair isn't in MR-EvE |
| `EpiGraphDB_get_genetic_correlations` | `rg` between a trait and others (shared etiology, NOT causation). **Sparse** — see Step 4 caveat |
| `EpiGraphDB_get_drugs_for_trait` | Drugs targeting genes associated with a risk-factor trait (causal-target follow-up) |
| `gwas_search_associations` | GWAS Catalog associations, to inspect the instruments behind a trait |

## Workflow

### Step 1 — Resolve trait labels (avoid silent misses)
EpiGraphDB matches GWAS trait labels **exactly and case-sensitively**. Always resolve free text first:

```
EpiGraphDB_search_opengwas {"query": "coronary heart disease"}
# → returns ids like 'ieu-a-7' and the exact label 'Coronary heart disease'
```

Use the returned exact label (or a sentence-case form) in the MR call. The MR tool now retries sentence-case variants and returns a `metadata.note` when it falls back or finds nothing — **read that note**; an empty `mr_results` with a note means "labels didn't match", NOT "no causal effect".

### Step 2 — Run MR (exposure → outcome)
```
EpiGraphDB_get_mendelian_randomization {
  "exposure_trait": "LDL cholesterol",
  "outcome_trait":  "Coronary heart disease",
  "pval_threshold": 1e-5
}
```
Each row carries `beta` (causal effect estimate), `se`, `pval`, `method`, `moescore`, and the exposure/outcome IDs.

### Step 3 — Interpret (see tables below)
Direction, magnitude, instrument quality, and method agreement.

### Step 4 — Triangulate
1. **Bidirectional MR (primary triangulation)** — swap exposure and outcome to test reverse causation. A causal X→Y with no Y→X strengthens the claim; bidirectional signals suggest shared genetics or feedback. This is the reliable leg — lean on it.
2. **Multiple methods** — prefer pairs where IVW and a pleiotropy-robust method (MR-Egger, weighted median) agree in sign and significance.
3. **Genetic correlation (secondary, often empty)** — `EpiGraphDB_get_genetic_correlations` on the exposure. ⚠️ The `/genetic-cor` graph is **sparse**: it stores only strong edges (|rg| > 0.8), matches **exact, case-sensitive** labels distinct from OpenGWAS search labels, and **ignores** the `pval_threshold` argument. Common traits (e.g. 'Body mass index') return empty — that is a graph gap, **not** "no shared genetics." Read `metadata.note`; if empty, do NOT conclude absence — fall back to bidirectional MR. When it does return, high `rg` + significant MR = causal; high `rg` + null MR = shared etiology without a detectable causal path.

### Step 5 — Actionable follow-up (optional)
`EpiGraphDB_get_drugs_for_trait` surfaces drugs whose target genes drive a causal risk factor — a genetics-anchored repurposing hypothesis.

## Interpretation tables

### Causal effect (`beta`)
| Observation | Meaning |
|---|---|
| `beta > 0`, `pval` significant | Higher exposure causally **increases** the outcome (on the GWAS scale — often log-odds for a binary outcome) |
| `beta < 0`, `pval` significant | Higher exposure causally **decreases** the outcome |
| `pval` not significant | No detectable causal effect at the available instrument strength — **absence of evidence, not evidence of absence** |
| Effect on a binary outcome | `beta` is typically a log-odds-ratio; report `exp(beta)` as an odds ratio per SD/unit of exposure |

### Instrument quality (`moescore`, "Mixture of Experts")
| MOE | Confidence |
|---|---|
| > 0.9 | High-quality instrument selection — trust the estimate most |
| 0.6–0.9 | Moderate — corroborate with another exposure GWAS or method |
| < 0.6 | Weak — treat as hypothesis-generating only |

### Method (`method`)
| Method | Note |
|---|---|
| IVW (inverse-variance weighted) | Primary estimate; assumes no pleiotropy |
| MR-Egger | Allows directional pleiotropy; intercept ≠ 0 flags pleiotropy; lower power |
| Weighted median | Valid if ≥50% of instrument weight is from valid variants |
| Disagreement across methods | A red flag for pleiotropy — downgrade confidence |

## Limitations (state these honestly)

- **Two MR paths, different scopes.** `EpiGraphDB_get_mendelian_randomization` returns *pre-computed* MR-EvE estimates for curated trait pairs — fast, but limited to pairs IEU already ran. For a pair that isn't covered, or for custom instruments (your own p-value/clumping thresholds), use `OpenGWAS_get_mr_instruments` (needs a free `OPENGWAS_JWT`) to assemble harmonized exposure+outcome SNP data, then compute the IVW/MR-Egger estimate yourself (e.g. IVW = Σ(βx·βy/σy²)/Σ(βx²/σy²)) or hand the `mr_input` to the `TwoSampleMR` R package. Advanced sensitivity analyses (MR-PRESSO, Steiger, leave-one-out) still need `TwoSampleMR`.
- **Palindromic SNPs** (A/T, C/G) are not strand-resolved by `OpenGWAS_get_mr_instruments`; review or drop ambiguous ones before trusting the estimate.
- **Horizontal pleiotropy** is the dominant threat and cannot be fully excluded from a single estimate. Method agreement reduces but does not eliminate it.