A new robust and accurate two-sample Mendelian randomization method with a large number of genetic variants

Horizontal pleiotropy can significantly confound causal estimates in Mendelian randomization (MR) studies, particularly when numerous instrumental variables (IVs) are employed. In this study we propose a novel statistical method, Mendelian Randomization analysis based on Z-scores (MRZ), to conduct robust and accurate MR analysis in the presence of pleiotropy. MRZ models the IV-outcome association z-score as a mixture distribution, separating the causal effect of the exposure on the outcome from pleiotropic effects specific to each IV. By classifying IVs into distinct categories (valid, uncorrelated pleiotropic, and correlated pleiotropic), MRZ constructs a likelihood function to estimate both causal and pleiotropic effects. Simulation studies demonstrate MRZs robustness, power, and accuracy in identifying causal effects under diverse pleiotropic scenarios and overlapped samples. In a bidirectional MR analysis of appendicular lean mass (ALM) and four lipid traits using both the UK Biobank (UKB)-internal datasets and the UKB-Global Lipids Genetics Consortium (GLGC) joint datasets, MRZ consistently identified a causal effect of ALM on total cholesterol (TC) and low-density lipoprotein cholesterol (LDL-C). Conversely, existing methods often detected mutual causal relationship between lipid traits and ALM, highlighting their susceptibility to confounding by horizontal pleiotropy. A randomized controlled experiment conducted in mice validated the absence of causal effect of TC on ALM, corroborating the MRZ findings and further emphasizing its resilience against pleiotropic biases.