Abstract

We introduce the Sparse Instrumental Variable (SPIV) framework for distinguishing causal and non-causal explanations of associations between biomarkers and diseases. SPIV overcomes the key limitations of the classic instrumental variable (Mendelian Randomisation) methods for learning the direction of causality, by allowing for pleiotropic genotypic effects on disease outcomes. SPIV extends the recently introduced Likelihood-based Causality Model Selection approaches by jointly modelling the effects of multiple instruments and biomarkers in the presence of latent confounders and measurement noise. Where the biomarkers are gene expressions, SPIV can be used for fine mapping of quantitative trait loci detected in genetic linkage studies.

Our framework relies on sparse Bayesian linear modelling, which offers a rigorous approach to model comparison, and is particularly useful for addressing p>>n problems of genetic epidemiology. SPIV is inspired by automatic relevance determination and adaptive shrinkage methods, but is used for causal discovery and quantitative trait loci fine mapping rather than regression.

We demonstrate our framework by examining effects of gene transcript levels in the lung and liver on 40 quantitative traits in a sample of 260 mice from a heterogeneous stock. The full set of biomarkers consisted of over 47 500 transcripts and 100 000 genotypic instruments. We identify genes predictive of the considered traits, and show evidence of a direct causal relationship between Smurf2 and the CD4/CD8 ratio. We show that our approach has wide application in identification of biomarkers as possible targets for intervention, or as proxy endpoints for early-stage clinical trials.