Bio Causal Genomics Mendelian Randomization

Version Compatibility

Reference examples tested with: TwoSampleMR 0.6.0+, MendelianRandomization 0.10+, MR-PRESSO 1.0+, cause 1.2+, MVMR 0.4+, ieugwasr 1.0+, MRlap 0.0.3.2+, coloc 5.2+, mrclust 0.1+, lhcMR 0.0.1+, R 4.4+. Both TwoSampleMR 0.6.0 and ieugwasr 1.0 are the JWT-transition versions; older versions still expect deprecated OAuth.

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters
• CLI (plink, GCTA-GSMR): <tool> --version then <tool> --help

If code throws an error referencing a function that has moved (e.g. ieugwasr::ld_clump vs TwoSampleMR::clump_data) or an OAuth token failure, introspect the installed API and adapt the example rather than retrying.

Mendelian Randomization

"Test whether trait X causally affects trait Y from GWAS summary statistics" -> Use genetic variants as instrumental variables (IVs) that satisfy three assumptions (relevance, independence, exclusion restriction) to estimate beta_causal = beta_outcome / beta_exposure under the IV framework (Davey Smith & Ebrahim 2003 IJE 32:1; Burgess & Thompson 2021 Chapman & Hall/CRC, 2nd ed.). Tool choice is a decision about the regime (one-sample vs two-sample, sparse vs polygenic, drug-target vs polygenic exposure) and the pleiotropy model (balanced, directional InSIDE, correlated horizontal). Wrong tool inflates Type-I error or attenuates true effects in a direction predictable from the bias structure.

• R: TwoSampleMR::mr() orchestrates IVW + Egger + weighted median + weighted mode in one call
• R: MendelianRandomization::mr_ivw / mr_egger / mr_median / mr_mbe / mr_conmix per-method API (S4 objects; MR-RAPS is NOT in this package -- use TwoSampleMR::mr_raps() which wraps the GitHub mr.raps)
• R: MRPRESSO::mr_presso() global / outlier / distortion tests
• R: cause::cause() correlated horizontal pleiotropy mixture
• R: MVMR::strength_mvmr() + MVMR::ivw_mvmr() multivariable conditional-F + IVW

Statistical Model Taxonomy

Methodology evolves; benchmark consensus shifts every 2-3 years. Verify against the current Slob & Burgess 2020 Genet Epidemiol, Burgess 2023 Wellcome Open Res "Guidelines for performing Mendelian randomization" (v3+), and STROBE-MR 2021 reporting standards before locking a method as primary.

Decision Tree by Experimental Scenario

One-Sample vs Two-Sample Bias Direction

Operational rule: Whenever both GWAS came from UK Biobank (or any single biobank), treat the analysis as one-sample-equivalent and prefer MR-RAPS as primary. Treating it as "two-sample because separate GWAS files" is a common error and produces overestimates.

MRlap: unified correction for sample overlap + winner's curse + weak instruments

MRlap (Mounier & Kutalik 2023 Genet Epidemiol 47:314) jointly corrects three biases that previously required three separate tools: sample overlap, winner's curse, and weak-instrument bias. It builds on an LDSC scaffold (cross-trait LD-score regression intercept estimates the overlap-induced covariance) and reweights the IVW estimate against the analytical bias-correction formula.

remotes::install_github('n-mounier/MRlap')   # never on CRAN; bioconductor unsuitable
library(MRlap)

fit <- MRlap(
    exposure = gwas_X_df, exposure_name = 'BMI',
    outcome = gwas_Y_df, outcome_name = 'T2D',
    ld = 'eur_w_ld_chr/', hm3 = 'w_hm3.snplist',   # LDSC reference files
    MR_threshold = 5e-8, MR_pruning_dist = 500, MR_pruning_LD = 0.05
)
fit$MRcorrection$corrected_effect       # overlap + winner's curse + weak-IV corrected
fit$MRcorrection$corrected_effect_se
fit$LDSC$h2_exp                          # exposure heritability sanity check
fit$LDSC$int_crosstrait                  # cross-trait LDSC intercept; ~0 means no sample overlap

Decision rule -- prefer MRlap when: (a) any sample overlap is suspected, (b) only sumstats are available (no individual-level data for re-running GWAS on disjoint samples), (c) exposure discovery and outcome were both run inside the same biobank (UKB-on-UKB, FinnGen-on-FinnGen). MRlap returns NA / unstable estimates when h^2 < 0.05; in that regime, fall back to Burgess 2016 overlap-corrected IVW plus MR-RAPS for the weak-IV component.

Drug-Target / cis-MR Framework

cis-MR restricts instruments to the cis-regulatory window of the gene encoding the protein/transcript exposure (Schmidt 2020 Nat Commun 11:3255), relaxing the exclusion-restriction assumption because the protein product directly mediates the SNP's effect on the outcome. Operational core: extract cis-pQTL/cis-eQTL within +/-500 kb of the gene; clump at r2 < 0.1 (looser than polygenic MR to retain power within a narrow window); require colocalization PP.H4 >= 0.7; flag protein-altering variants (PAV) which can break SomaScan/Olink aptamer/antibody binding rather than reflect biology.

Full drug-target cis-MR workflow including UKB-PPP / deCODE / Fenland pQTL panels, PAV flagging, Olink vs SomaScan replication (~15-30% cross-platform disagreement), and the operational claim ladder lives in causal-genomics/proteome-mr-drug-target. Use that skill for any drug-target nomination.

Binary outcomes and non-collapsibility

MR with logistic-GWAS sumstats returns per-allele log-OR on the population-averaged (marginal) scale, NOT the conditional log-OR (Burgess 2017 Stat Methods Med Res 26:2333). For rare disease (prevalence < 10%), OR ~= RR ~= HR and the distinction is harmless. For common disease, OR diverges from RR/HR and the MR estimate cannot be back-converted to a conditional effect without strong assumptions; report as "per 1-SD increase in genetically-predicted X, OR for Y = ..." rather than implying an individual-level intervention effect.

Collider bias when conditioning on a collider variable (Coscia 2022 Eur J Epidemiol 37:671 formalizes this for stratified MR): case-only or disease-progression designs condition the sample on disease status, opening a collider path between any cause of disease and any cause of progression. MR within affected subsets without explicit adjustment for selection probability is fragile; weight by inverse probability of selection or restrict claims to the unconditioned population.

Per-Method Failure Modes

IVW under directional pleiotropy

Trigger: Several SNPs affect the outcome through pathways not via the exposure, in a consistent direction.

Mechanism: IVW is a weighted regression through origin; non-zero mean pleiotropy shifts the slope.

Symptom: Egger intercept p < 0.05 with non-zero estimate; IVW differs from weighted median; MR-PRESSO global test p < 0.05.

Fix: Use Egger (if I^2_GX >= 0.9 -- otherwise SIMEX-correct via the simex package applied to the Egger fit, treating se.exposure as measurement error in beta.exposure); cross-check with weighted median, MR-PRESSO, and CAUSE; report IVW only as one of a panel, never alone. The MendelianRandomization::mr_egger() function accepts distribution='normal' and reports the I.sq (I^2_GX) NOME diagnostic but applies no NOME/SIMEX correction to the estimate itself and does NOT expose a SIMEX wrapper.

Weak-instrument bias direction

Trigger: Mean per-instrument F-statistic < 10, or several individual F < 10.

Mechanism: Weak IVs amplify finite-sample correlation between IV-X and IV-Y errors; bias direction depends on overlap regime (see table above).

Symptom: Estimates shift markedly when removing the weakest instruments; one-sample MR estimates much larger than two-sample.

Fix: Compute F per instrument from the EXPOSURE GWAS, not the outcome; exclude F < 10; use MR-RAPS (handles weak IVs by design); for two-sample, also report unweighted IVW (less weak-IV-bias-inflated than weighted in some regimes).

Winner's curse at P~5e-8

Trigger: Discovery GWAS is the source of both instrument selection and effect-size estimates.

Mechanism: SNPs that just cross 5e-8 in discovery have over-estimated effect sizes (regression toward the mean in independent replication); MR uses inflated beta_X, biasing causal estimate.

Symptom: MR effect shrinks substantially when using effect sizes from an independent replication GWAS.

Fix: (1) Three-sample design (discovery / replication-for-instrument-effect / outcome) where feasible. (2) When sumstats-only: MRlap (Mounier 2023), MR-SimSS (sample-splitting from sumstats), or RIVW (Ma 2023 Ann Statist 51:211 -- rerandomized IVW) jointly correct winner's curse + weak IVs + overlap. (3) Jiang 2023 IJE 52:1209 empirical magnitude: variant-level inflation ~50-400% near the genome-wide-significance threshold, dropping to <25% when the minimum P <= 1e-13.

NOME violation invalidating Egger

Trigger: Running MR-Egger with I^2_GX < 0.9.

Mechanism: Egger assumes NO Measurement Error in exposure effect sizes (NOME); when violated, Egger slope is attenuated toward null with reciprocal bias on the intercept.

Symptom: mr_pleiotropy_test() Egger estimate disagrees with weighted median in magnitude but agrees in direction; Isq() function returns <0.9.

Fix: Compute Isq(beta_X, se_X) (Bowden 2016 IJE 45:1961); if <0.9, apply SIMEX correction via simex package or report Egger as exploratory only. The MendelianRandomization package's mr_egger() reports the I.sq (I^2_GX) NOME diagnostic but does not itself apply a NOME/SIMEX correction; SIMEX must be run separately.

Steiger filter false flag under unmeasured confounding

Trigger: Applying steiger_filtering() on traits with unmeasured shared confounders (e.g. SES).

Mechanism: Steiger compares variance explained in exposure vs outcome per SNP; an unmeasured confounder upstream of both produces SNPs that explain more variance in the outcome than the exposure, falsely flagging "reverse causation" (Lutz 2022 Genet Epidemiol 46:139).

Symptom: Many SNPs flagged as wrong-direction yet biology and prior MR support forward causation.

Fix: Treat Steiger as a heuristic, not gospel; cross-validate direction with bidirectional MR (forward + reverse with independent instrument sets); for known-confounder-rich domains (psychiatric traits, SES proxies) use LCV or LHC-MR instead, which jointly model confounders.

Palindromic SNP harmonization

Trigger: SNPs with alleles A/T or C/G near MAF 0.5.

Mechanism: Strand orientation is ambiguous for palindromic SNPs when allele frequencies are intermediate; flipping introduces sign errors that look like pleiotropy.

Symptom: harmonise_data() reports many palindromic SNPs dropped; remaining SNPs show heterogeneity from a handful.

Fix: Default action = 2 (infer from allele frequencies) drops MAF~0.5 palindromes; action = 3 drops ALL palindromes (most conservative); never use action = 1 (assumes forward strand) unless both GWAS are guaranteed to use the same strand convention. Document choice in methods.