Enhancing Mendelian Randomization Causal Inference through Pleiotropic Clustering in the Presence of Correlated Horizontal Pleiotropy

Abstract Correlated horizontal pleiotropy poses a formidable challenge in Mendelian randomization (MR). Existing methods, relying on sparse pleiotropic variants or zero modal pleiotropy assumption (ZEMPA), may face issues, especially in trait pairs with a high proportion of correlated horizontal pleiotropic variants. We propose an innovative Pleiotropic Clustering model for MR analysis (PCMR) to cluster instrument variables with horizontal or vertical pleiotropic effects. Our simulations showed that PCMR had superior performance in isolating causal effects and detecting correlated horizontal pleiotropy. Moreover, PCMR's causal evaluation extended the ZEMPA and can control false positives in correlated horizontal pleiotropy, even with a high proportion. Applied to real datasets, PCMR detected 15% (8 of 48) trait pairs significant with correlated pleiotropy in common diseases and further enhanced causal inference. Notably, PCMR elucidated why existing methods might overlook the established causal link between smoking behavior and type 2 diabetes. It also confirmed the considered non-causal relationship between high-density lipoprotein and coronary artery disease, and revealed potential non-causal relationships, such as from schizophrenia to major depressive disorder, offering new insights into complex disease mechanisms.