An R Package for Mendelian Randomization in the mediation setting • FLOWMR

This is the R package for FLOW-MR, a Bayesian framework for Mendelian Randomization in the mediation setting.

For details of the FLOW-MR framework, please refer to our manuscript.

Here we give a brief description of the method. Check tutorial for the complete workflow.

For the code reproducing the results in the manuscript, please visit link.

Hardware Requirements

The FLOWMR package requires a standard computer with sufficient RAM to support in-memory operations. A minimum of four CPU cores is recommended for parallel execution.

Operating System and Software Requirements

The package was developed and tested on macOS Sonoma 14.0.

It is expected to be compatible with both Windows and macOS. Before using FLOWMR, users should install the necessary dependencies from CRAN, as specified in the DESCRIPTION file.

Installation

To run the package, you might need to first install RCpp from CRAN

install.packages("RCpp")

Then the package can be installed in R from this Github repository:

library(devtools)
install_github("ZixuanWu1/FLOW-MR")

The installation should take no more than 5 minuts on a normal computer

Basic Usage

Recall in the Mediation setting, we have $\Gamma = \tilde{B} \cdot \Gamma + \alpha$ where $\Gamma$ is the matrix of true SNP effects, $\tilde{B}$ is an upper-triangular matrix with zero diagonalsand $\alpha$ is the horizontal pleiotropy. Here $\tilde{B}$ represents the matrix of direct effects. One might equivalently write this as $\Gamma = (I + {B}) \alpha$

In order to run MrMediation, we will need at least two inputs, Gamma_hat and Sd_hat. The Gamma_hat matrix is an observation of $\Gamma$ . Typically it is a $K \times P$ matrix where $K$ is the number of phenotypes and $P$ is the number of the measured genetic variants. The Sd_hat matrix is the matrix of the standard deviation of noise and is of same dimension as Gamma_hat.

To estimate the matrix $B$ , one can use the following command

result = BayesMediation(Gamma_hat, Sd_hat, inv = TRUE)

Here inv = TRUE implies we are estimating $\tilde{B}$ , as opposed to estimating ${B}$ when inv = FALSE.

One can also support the correlation among the measurement errors by setting cor to be the correlation matrix

result_cor = BayesMediation(Gamma_hat, Sd_hat, cor = cor_mat, inv = TRUE)

In addition, in order to compute the indirect effect of exposures on the outcome, one can simply use the following command

result = BayesMediation(Gamma_hat, Sd_hat, cor = cor_mat, indirect = T)

Sometimes we might see a warning that the algorithm might not have convergenced. This problem could be solved by using more iterations or changing the initialization method. One can also look at the traceplot of parameters for MCMC diagnosis. For instance, the following function gives the traceplot of the parameter of insterest of all the chains.

traceplot(result$raw, par = "B", ind = c(1,2))

(Here we are using $B[1,2]$ as an example.)