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

Horizontal pleiotropy poses challenges in Mendelian randomization (MR) by biasing causal estimates, particularly when using numerous instrumental variables (IVs). We propose MRZ, a method that leverages IV-outcome association z-scores modeled as a mixture distribution to distinguish causal effects from IV-specific pleiotropy. MRZ classifies IVs into three categories: valid, uncorrelated pleiotropic, or correlated pleiotropic, enabling simultaneous estimation of causal and pleiotropic parameters through likelihood-based inference. Simulations demonstrate MRZ’s superior robustness and accuracy across diverse pleiotropic scenarios. When applied to investigate appendicular lean mass (ALM) and lipid traits, MRZ identifies significant negative causal effects of ALM on total cholesterol (TC) and low-density lipoprotein cholesterol (LDL-C), with no evidence of reverse causation. In contrast, existing methods frequently suggest bidirectional effects. Randomized controlled experiments in mice validate MRZ results, confirming ALM’s causal effects on TC/LDL-C but not vice versa. By addressing key limitations of conventional approaches vulnerable to horizontal pleiotropy, MRZ advances causal inference in complex trait architectures. The authors develop MRZ, a Mendelian randomization method that models variant-outcome z-scores as a mixture to separate causal effects from pleiotropy, and validate the causal findings in mouse experiments.