Phenome-scale causal network discovery with bidirectional mediated Mendelian randomization

Abstract

Inference of directed biological networks from observational genomics datasets is a crucial but notoriously difficult challenge. Modern population-scale biobanks, containing simultaneous measurements of traits, biomarkers, and genetic variation, offer an unprecedented opportunity to study biological networks. Mendelian randomization (MR) has received attention as a class of methods for inferring causal effects in observational data that uses genetic variants as instrumental variables, but MR methods rely on assumptions that limit their application to complex traits at the biobank-scale. Moreover, MR estimates the total effect of one trait on another, which may be mediated by other factors. Biobanks include measurements of many potential mediators, in principle enabling the conversion of MR estimates into direct effects representing a causal network. Here, we show that this can be accomplished by a flexible two stage procedure we call bidirectional mediated Mendelian randomization (bimmer). First, bimmer estimates the effect of every trait on every other. Next, bimmer finds a parsimonious network that explains these effects using direct and mediated causal paths. We introduce novel methods for both steps and show via extensive simulations that bimmer is able to learn causal network structures even in the presence of non-causal genetic correlation. We apply bimmer to 405 phenotypes from the UK biobank and demonstrate that learning the network structure is invaluable for interpreting the results of phenome-wide MR, while lending causal support to several recent observational studies. * bb2991@columbia.edu † dak2173@columbia.edu causal effects, MR methods must make strong assumptions that limit their ability to be applied at the biobank-scale. Perhaps the most controversial assumption is that the SNP only effects B through A (i.e. there is no horizontal pleiotropy). Recent methods such as Egger regression and the mode-based-estimator are able to relax this assumption, instead assuming there is no correlated pleiotropy or modal pleiotopy, respectively [11,12]. Another approach, the latent causal variable (LCV) model, is able to detect causality under arbitrarily-structured pleiotropy [13]. However, the quantity that LCV calculates is not interpretable as the causal effect size of A on B. Most MR studies also presuppose the direction of effect, specifying one phenotype as the outcome and the other as the exposure. Pre-specifying the effect direction is sound when the outcome is clearly biologically downstream of the exposure, but in some cases it is better to learn the direction of the effect from the data. Some researchers have instead used bidirectional MR [14,15], which tests for an effect in each direction, or gwas-pw [16], which infers the effect direction from the data. However, the utility of these approaches for complex traits, which might contain non-causal genetic correlation, is questionable [13].In mimicking a randomized controlled trial, MR estimates the total causal effect (TCE) of A on B [17]. This effec...

Simultaneous estimation of bi-directional causal effects and heritable confounding from GWAS summary statistics

“…The LCV approach [34] is a special case of our model, where the causal effects are not included in the model, but they estimate the confounder effect mixed with the causal effects to estimate a quantity of genetic causality proportion (GCP). In agreement with others [10,35] , we would not interpret non-zero GCP as evidence for causal effect. Moreover, in other simulation settings LCV has showed very low power to detect causal effects (by rejecting GCP=0) (Fig S15 in Howey et al [36] ).…”

MRSL: A phenome-wide causal discovery algorithm based on GWAS summary data

“…Our method can be easily implemented only using GWAS summary data, which is public available for the most phenotypes as the emergence of a large number GWAS studies with huge sample size. Published MR-based algorithm such as cGAUGE, requires the individual-level data and are thus not as easily available as the GWAS summary statistics, and is very time consuming [25]; BIMMER are implemented based on the complex inverse sparse regression and obtained an approximately estimation of DCE matrix, this require time roughly ࣩ(κ d 4 ) for d phenotypes [28]. In the result of simulation study 2-3, we found that the computing time of MRSL is only around 1/100 of BIMMER, and 1/1000 of cGAUGE, respectively.…”

“…Also, MRSL outputs the unbiased direct effect of each pair of variables. Moreover, MRSL can be applied into the structure with feedback loops between any two variables, because our main MR method IVW can powerfully deal with the case of bi-directional causal relationship between two variables [28,76]. Similar to MR analysis, GWAS summary data of d phenotypes should from the homogenous population.…”

“…This algorithm allows the unobserved confounders among all the variables and requires individual genetic and phenotypic data. A flexible two stage procedure called bidirectional mediated Mendelian randomization (BIMMER) can be used to infer sparse networks of direct causal effects (DCEs) from phenome-scale GWAS summary statistics [28]. However, this process is implemented by inverse sparse regression under the assumption that the DCE matrix is sparse.…”

Deep Mendelian Randomization: Investigating the Causal Knowledge of Genomic Deep Learning Models

“…In the future, we will aim to combine DeepMR with a causal network inference method such as our recent bimmer model [Brown and Knowles, 2020] to explicitly account for the influence of other assayed TFs on each pair. DeepMR would also benefit from accompanying tools for diagnosing when model-generated data deviates from or violates MR assumptions.…”