METEOR: a data-adaptive Mendelian randomization method for powerful detection of shared and specific exposures underlying multiple outcomes

Abstract Accurate identification of causal exposures for multimorbidity can benefit the co-prevention and co-management of multiple-related outcomes. This goal can be conceptually addressed within a multi-outcome Mendelian randomization (MR) framework. However, existing multi-outcome MR methods suffer from restrictions on format and availability of data inputs, fail to account for the potential sample overlap, rely on pre-selected independent instrumental variables (IVs), and are unable to account for horizontal pleiotropy. Here, we propose METEOR, a novel MR method that jointly models one exposure and multiple outcomes to identify both shared and outcome-specific causal exposures. METEOR accounts for sample overlap between exposure and outcomes, allows outcomes from different genome-wide association studies (GWAS) datasets, self-adaptively determines IVs from correlated single-nucleotide polymorphisms, and explicitly models horizontal pleiotropy. Using summary statistics, METEOR infers causal effects under a joint-likelihood framework with a scalable, sampling-based algorithm. Simulations show that METEOR presents well-calibrated $P$-values for both global and single-outcome tests, and achieves average power improvements of 55.33% and 56.50% over five existing MR methods in the global and single tests, respectively. In real data applications, METEOR produces the most accurate causal effect estimates in positive control analyses, reduces false positives by 18.75% in negative control analyses, and highlights that controlling BMI could benefit the co-management of multiple cardiovascular diseases (CVDs) and multiple gastrointestinal (GI) diseases, while controlling blood pressure could benefit the co-management of multimorbidity across CVDs and mental disorders (MDs), as well as across GI diseases and MDs.