Inferring a causal network among multiple traits is essential for unraveling complex biological relationships and informing interventions. Mendelian randomization (MR) has emerged as a powerful tool for causal inference, utilizing genetic variants as instrumental variables (IVs) to estimate causal effects. However, when the directions of causal relationships among traits are unknown, reconstructing the underlying causal network becomes challenging. In particular, the presence of cycles or feedback loops, which are common in biological systems, poses additional challenges for causal network inference, and remains largely under-studied with standard MR approaches and existing IV-based network inference methods. To address these issues, we introduce MR2G, a new statistical framework that enables robust inference of causal networks, including those with cycles, directly from GWAS summary statistics. MR2G is built on a formally defined recursive causal graph model that rigorously links direct causal effects to (univariable) MR estimands. It recovers a biologically interpretable causal network from pairwise MR effect estimates, while incorporating a network-informed IV screening strategy to reduce pleiotropic bias and improve robustness. Through realistic simulations, MR2G demonstrates superior accuracy and robustness in recovering complex causal structures, including those involving feedback loops. We apply MR2G to GWAS summary statistics for six complex diseases and nine cardiometabolic risk factors. MR2G not only recovers well-established causal pathways but also uncovers multiple feedback relationships, highlighting its utility in disentangling complex and biologically plausible causal networks from large-scale genetic data.
Lin et al. (Tue,) studied this question.