Abstract
Mendelian randomization (MR) is widely used to evaluate causal effects of complex trait exposures on disease outcomes. Recently, MR has been increasingly applied to molecular traits, such as gene expression, to map risk genes. However, transcriptome-wide MR (TWMR) faces unique challenges. The number of available cis-QTLs as instrumental variables (IVs) is often limited, and horizontal pleiotropy is pervasive, violating core MR assumptions and compromising inference validity. We introduce FusioMR, a robust MR framework tailored for molecular trait exposures while also applicable to complex trait exposures. Our single-outcome model, FusioMR(s), incorporates gene-region-specific empirical priors informed by the number and strength of QTLs, linkage disequilibrium, and effect size consistency. It uses sampling-based inference to improve robustness when instruments are limited. Our multi-outcome model, FusioMR(m), is motivated by the observation that many complex diseases have correlated diseases, subtypes, or comorbidities, which could be affected by shared or correlated exposures. FusioMR(m) jointly analyzes two correlated outcomes, leveraging shared IVs and pleiotropic effects of shared/correlated exposures to improve estimation precision and power, particularly for underpowered outcomes. We applied FusioMR(s) to identify cell-type-specific gene expression traits associated with Alzheimer disease using single-cell eQTL and GWAS summary data. We applied FusioMR(m) to detect alternative polyadenylation events affecting atrial fibrillation and ischemic stroke, and to estimate the causal effect of low-density lipoprotein on ischemic stroke in South Asian populations by borrowing information from European ancestry data. These applications highlight the generalizability of FusioMR for both molecular and complex trait exposures.