Inferring causal networks for genomic trios with high-dimensional confounding variables

Abstract

Causal structure inference under the principle of Mendelian randomization (PMR) enables the study of causal relationships between two molecular phenotypes (e.g. gene expression) by leveraging genetic variants as instrumental variables to impose constraints on edge direction. In practice, inference is often affected by numerous confounding variables, such as other genes, other processes and demographic variables, making it essential to robustly account for confounders. However, existing methods have limited ability to accommodate high-dimensional confounding. Here, we introduce Mendelian randomization for genomic networks (MRGN), a regression-based method under the PMR for inferring causal models in genomic trios from individual-level data while accounting for many confounders. MRGN extends beyond the canonical mediation model to distinguish among multiple causal relationships. By prioritizing conditional independence tests given the genetic variant and confounders, MRGN can detect causal relationships between phenotypes even when their marginal association is weak. Through extensive simulations, we demonstrate improved inference accuracy, robustness, and computational efficiency over existing approaches. We further apply MRGN to data from the Genotype-Tissue Expression Consortium to investigate how genetic variants may regulate gene expression in cis and trans.