A Bayesian network-based framework to uncover the causal effects of genes on complex traits based on GWAS data

Abstract

Abstract Deciphering the relationships between genes and complex traits could help us better understand the biological mechanisms leading to phenotypic variations and disease onset. Univariate gene-based analyses are widely used to characterize gene-phenotype relationships, but are subject to the influence of confounders. Furthermore, while some genes directly contribute to traits variations, others may exert their effects through other genes. How to quantify individual genes’ direct and indirect effects on complex traits remains an important yet challenging question. We presented a novel framework (BN-GWAS) to decipher the total and direct causal effects of individual genes using imputed gene expression data from GWAS and raw gene expression from GTEx. The study was partially motivated by the quest to differentiate “core” genes (genes with direct causal effect on the phenotype) from “peripheral” ones. Our proposed framework is based on a Bayesian network (BN) approach, which produces a directed graph showing the relationship between genes and the phenotype. The approach aims to uncover the overall causal structure, to examine the role of individual genes and quantify the direct and indirect effects by each gene. An important advantage and novelty of the proposed framework is that it allows gene expression and disease trait(s) to be evaluated in different samples, significantly improving the flexibility and applicability of the approach. It uses IDA and jointIDA incorporating a novel p-value-based regularization approach to quantify the causal effects (including total causal effects, direct causal effects, and medication effects) of genes. The proposed approach can be extended to decipher the joint causal network of 2 or more traits, and has high specificity and precision (a.k.a., positive predictive value), making it particularly useful for selecting genes for follow-up studies. We verified the feasibility and validity of the proposed framework by extensive simulations and applications to 52 traits across 14 tissues in the UK Biobank (UKBB). Split-half replication and stability selection analyses were performed to demonstrate the accuracy and efficiency of our proposed method to identify causally relevant genes. The identified (direct) causal genes were found to be significantly enriched for genes highlighted in the OpenTargets database, and the enrichment was stronger than achieved by conventional univariate gene-based tests. Encouragingly, many enriched pathways were supported by the literature, and some of the enriched drugs have been tested or used to treat patients in clinical practice. Our proposed framework provides a powerful way to prioritize genes with large direct or indirect causal effects and to estimate the ‘variable importance’ of such genes.