henhaobio
Sign in

Nature GeneticsGenomicsPeer-reviewed journal article

Widespread pleiotropy confounds causal relationships between complex traits and diseases inferred from Mendelian randomization

Marie VerbanckiD, Chia‐Yen CheniD, Benjamin M. NealeiD, Ron DoiD

PDFDOIPreprint
In-depth full-text digest available · Jump to digest ↓

Abstract

A fundamental assumption in inferring causality of an exposure on complex disease using Mendelian randomization (MR) is that the genetic variant used as the instrumental variable cannot have pleiotropic effects. Violation of this no pleiotropy assumption can cause severe bias. Emerging evidence have supported a role for pleiotropy amongst disease-associated loci identified from GWA studies. However, the impact and extent of pleiotropy on MR is poorly understood. Here, we introduce a method called the Mendelian Randomization Pleiotropy RESidual Sum and Outlier (MR-PRESSO) test to detect and correct for pleiotropy in multi-instrument summary-level MR testing. We show using simulations that existing approaches are less sensitive to the detection of pleiotropy when it occurs in a subset of instrumental variables, as compared to MR-PRESSO. Next, we show that pleiotropy is widespread in MR, occurring in 41% amongst significant causal relationships (out of 4,250 MR tests total) from pairwise comparisons of 82 complex traits and diseases from summary level genome-wide association data. We demonstrate that pleiotropy causes distortion between-168% and 189% of the causal estimate in MR. Furthermore, pleiotropy induces false positive causal relationships-defined as those causal estimates that were no longer statistically significant in the pleiotropy corrected MR test but were previously significant in the naive MR test-in up to 10% of the MR tests using a P < 0.05 cutoff that is commonly used in MR studies. Finally, we show that MR-PRESSO can correct for distortion in the causal estimate in most cases. Our results demonstrate that pleiotropy is widespread and pervasive, and must be properly corrected for in order to maintain the validity of MR.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 1 (30 Jun 2017) · CC BY-NC-ND 4.0

MR-PRESSO detects pleiotropic outlier instruments in summary-level Mendelian randomization; across 82 traits it found pleiotropy in 41% of significant causal estimates, with distortions from −168% to 189%.

Mendelian randomization (MR) assumes each genetic instrument affects the outcome only through the exposure, yet many GWAS variants influence several traits, and how often this breaks MR had not been measured at scale. MR-PRESSO adds global, outlier and distortion tests that work inside inverse-variance-weighted (IVW) MR, unlike MR-Egger, which loses power, or the Q test, which cannot correct pleiotropy and showed inflated false positives.

Key results

  1. ShownPleiotropy is common among MR tests between complex traits: MR-PRESSO flags it often, while the MR-Egger intercept test flags it in none.

    4,250 MR tests across pairs of 82 traits, Bonferroni P = 1.17×10-5: global test significant in 22% (N = 922) vs 0% for the MR-Egger intercept; among the 191 significant IVW causal estimates, 41% (N = 78) showed pleiotropy, again with no MR-Egger signal. The two tests target different patterns (outlier vs average pleiotropy), so their hits barely overlap. · Evaluation of pleiotropy for complex traits and…; Detection and correction of pleiotropy using…

  2. SuggestedIn simulations MR-PRESSO detected pleiotropy confined to a small subset of instruments with good power and no type I error inflation, unlike the Q test (inflated) and the MR-Egger intercept (underpowered).

    10,000 simulations per scenario: MR-PRESSO and the Q test had power > 72% when more than 1% of variants were pleiotropic, and MR-PRESSO >75% power with 5% pleiotropic variants; in the scenario with no pleiotropy and βcausal = 0.05 the Q test's type I error was 39% vs 5% for MR-PRESSO; MR-Egger intercept power stayed < 13% in every scenario. Sample overlap had no appreciable effect. · Comparison of statistical properties of methods…; Discussion

  3. ShownRemoving pleiotropic outliers significantly changed about one in ten nominally significant causal estimates, in some cases by more than the size of the estimate.

    Distortion test at P < 0.05: significant in almost 10% (N = 22) of 229 causal estimates, ranging from −168% to 189%; at the Bonferroni threshold 2.5% (N = 2 of 81), at −93% and 35%. Example: the IVW effect of body mass index (BMI) on C-reactive protein, 0.39 (P = 7.02 × 10-8), fell to 0.35 after removing one APOE-locus outlier (rs2075650), a 12% distortion the MR-Egger intercept missed (P = 0.92). · Pleiotropy confounds causal estimates in MR…

  4. ShownSome causal relationships significant in naive MR lose significance once pleiotropic outliers are removed, which the authors count as pleiotropy-induced false positives.

    Defining a false positive as an estimate significant in naive IVW but not after outlier correction: 10% (N = 24 of 229) of estimates at the nominal P < 0.05 threshold, and 1.2% (N = 1 of 81) at the Bonferroni threshold (P < 1.17 × 10-5). · Pleiotropy induces false positive causal…

  5. ShownOutlier removal and covariate adjustment cleared detectable pleiotropy in about half of affected tests, and established causal links survived correction.

    Of 922 tests with significant pleiotropy, MR-PRESSO outlier removal cleared the global-test signal in 46% (N = 422); Multi-Phenotype MR (MPMR) covariate adjustment cleared 22% (N = 20 of 93) singly and 34% (N = 22 of 42) jointly; together 47% (N = 438). After correction, LDL cholesterol on coronary artery disease (β 0.52) and BMI on type 2 diabetes (β 0.76) stayed significant. · Control of pleiotropic effects through outlier…; Causal relationships inferred from MR testing

Methods, data and reproducibility

MR-PRESSO compares each instrument's observed residual from a leave-one-out IVW regression with a null distribution simulated from the variants' standard errors (global and per-variant outlier tests), then tests whether removing outliers significantly changes the causal estimate. It was assessed in 10,000 simulations per scenario and applied to all 4,250 exposure-outcome pairs among 82 GWAS traits.

  • Public GWAS summary statistics for complex traits and diseases, lifted over to hg19 and imputed with ImpG · human · n = 82 traits and diseases
  • 1000 Genomes Phase 3 European reference panel used for summary-statistic imputation · human · n = N=503
  • Simulated multi-instrument MR data with two exposures and 1, 5 or 10% pleiotropic variants (directional or balanced), with and without sample overlap · n = 10,000 simulations per scenario
Methods & tools
MR-PRESSO, MR-Egger regression, Q test, IVW meta-analysis, MPMR, ImpG, UCSC Genome Browser LiftOver, NHGRI-EBI GWAS catalog, R
Reagents & environment
MR-PRESSO is implemented in R; no versions or package details are given.
Experimental conditions
Empirical P-values used 1,000,000 simulations for the global and outlier tests and 10,000 for the distortion test; the authors note this makes the method slower than alternatives. No hardware or run times are given.
Reproducibility gaps
No code or data availability statement appears in this text, and the 82 GWAS sources are listed only in supplementary material that is not included here.
Limitations (6)
  • In some tests the global test was significant but the outlier test could not pin down specific variants, possibly because a subset of variants each carry pleiotropy too weak to detect singly. · stated by the authors
  • Covariate adjustment with MPMR needs prior knowledge of the trait mediating the pleiotropy, and residual bias can remain if that trait is itself confounded with the outcome. · stated by the authors
  • Correction is incomplete: the strategies worked in 47% of the 922 pleiotropic tests, and estimates that stay pleiotropic should be read as biased. · stated by the authors
  • Because empirical P-values come from simulation, MR-PRESSO can run slower than competing methods. · stated by the authors
  • 'False positive' causal links are defined operationally as losing significance after outlier removal, not against a known truth, and they rest on small counts (24 of 229; 1 of 81). · noted in this summary
  • Reported denominators and thresholds vary across sections (4,250 vs >4,025 tests; 191 vs 81 significant estimates; Bonferroni 1.17×10-5 vs 1×10-5), which makes the proportions hard to trace. · noted in this summary

Read this if you run summary-statistic Mendelian randomization and need a way to detect, remove and quantify the impact of pleiotropic instruments.

Citations
8,596
FWCI
135.09
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISGrant '24Hu '22Kippersluis '17Ren '26Jiang '19Morrison '20
82% semantic matchGenetics
Andrew J. Grant, Stephen Burgess·The American Journal of Human Genetics·Jan 1, 2024· 40 citations