Bayesian network-based Mendelian randomization for variant prioritization and phenotypic causal inference.

Area of Science:

• Genetics
• Causal Inference
• Statistical Genomics

Background:

• Mendelian randomization (MR) infers causality but faces challenges with genetic instrumental variables due to interactions, linkage, and pleiotropy.
• Existing MR methods struggle with complex genetic architectures and large-scale genomic datasets.

Purpose of the Study:

• To introduce Bayesian network-based Mendelian randomization (BNMR), a novel framework for robust causal inference using individual-level data.
• To address limitations in instrumental variable selection and pleiotropy in Mendelian randomization.

Main Methods:

• BNMR utilizes a random graph forest for Bayesian network structural learning to prioritize and select genetic variants.
• A shrinkage prior is incorporated within the Bayesian framework to achieve pleiotropy-robust effect estimation.
• The method is validated through simulations and applied to UK Biobank data.

Main Results:

• Simulations show BNMR effectively reduces false positives in variant selection and surpasses existing MR methods in accuracy and statistical power.
• Application to UK Biobank data identified causal links between hematological traits, blood pressures, and psychiatric disorders.
• BNMR demonstrates superior performance in handling complex genetic structures and large-scale genomic data.

Conclusions:

• BNMR offers a powerful and accurate approach for causal inference in genomics, overcoming key challenges of traditional Mendelian randomization.
• The framework's ability to handle complex genetic data facilitates real-world evidence studies and advances understanding of causal mechanisms.
• BNMR is a promising tool for uncovering complex biological relationships in large-scale genomic studies.