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

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.. CC-BY-NC-ND 4.0 International license peer-reviewed) is the author/funder. It is made available under aThe copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/157552 doi: bioRxiv preprint first posted online Epidemiological studies have established correlations among numerous exposures and complex diseases 1 . It can be challenging to draw causal inference from these studies due to reverse causation, confounding and/or other biases 2 .Mendelian randomization (MR) is a commonly-used human genetics approach that can infer causality of an exposure with complex disease 3,4 . MR presents a number of advantages over observational epidemiology, including controlling for non-heritable environmental confounders in such analyses and the ability to use genetic instruments to evaluate the impact of an exposure without necessitating the collection of that exposure in the outcome group. MR utilizes genetic variants as genetic instrumental variables (IVs) that are robustly associated with the exposure of interest and tests whether these variants results in a proportional effect on the outcome.With the discovery of thousands of trait-associated loci identified from genome-wide association (GWA) studies, multi-instrument MR methods have been developed to ...

Improving the accuracy of two-sample summary data Mendelian randomization: moving beyond the NOME assumption

“…Most importantly, we highlight the potential for serious type I error inflation of Cochran's Q statistic when using standard 1st order weights with weak instruments. In recent work, Verbank et al [30] also noted this same tendency and proposed a simulation-based alternative to 1st order weighting named 'MR-PRESSO'. Our simulations show that modified 2nd order weights can deliver much more accurate causal estimates and reliable tests for heterogeneity than either 1st or 2nd order weighting, and suggests that the computationally intensive MR-PRESSO approach is unnecessary.…”

“…Our simulations suggest that the exact implementation of modified 2nd order weights should be used when testing for the overall presence of heterogeneity (referred to as the 'global' test by Verbank et al [30]). However, they also suggest that the iterative implementation is preferable if looking at the individual contribution of each SNP to the Q statistic to decide on its status as an outlier.…”

Detecting genome-wide directional effects of transcription factor binding on polygenic disease risk

“…Regarding causality, this is because a consistent genome-wide directional effect of SNPs predicted to affect TF binding due to sequence change (across a large set of TF binding sites; see Table 1a) is less susceptible to pleiotropy, LD, and allelic heterogeneity 103,105 . The robustness of our method to these potential confounders is also greater than that of genetic correlation and Mendelian randomization 131 (MR) analyses, which can be confounded by reverse causality and pleiotropic effects [146][147][148] (and which would scale poorly because they would require TF ChIP-seq in many individuals for every TF/cell-type pair studied). The reason that our method is not confounded by reverse causality is that each of our annotations is produced in a cell population that is isogenic and therefore does not have variance in genetic liability for any trait.…”

Inferring causal relationships between phenotypes using summary statistics from genome-wide association studies

“…For example, the generalized summary data based Mendelian randomization analysis (GSMR) allows using the correlated genetic variants (Zhu et al 2017). The Mendelian Randomization Pleiotropy RESidual Sum and Outlier (MR-PRESSO) test can correct for distortion in the causal estimate in most cases by removal of offending instrumental variables that exhibit pleiotropy (Verbanck et al 2017). The latent causal variable (LCV) model can quantify the degree of partial genetic causality (O’Connor and Price 2017).…”