Causal inference for multiple risk factors and diseases from genomics data

Abstract

AbstractIn high dimensional observational genotype-phenotype data, complex relationships and confounders make causal learning difficult. Here, we bridge a gap between genetic epidemiology and statistical causal inference, to demonstrate that graphical inference can fine-map trait-specific causal DNA variants and identify risk factors that are most likely to have a direct causal effect on a disease outcome. Our CI-GWAS approach learns a single graph representing the causal relationships among millions of DNA variants and 17 traits in less than 10 minutes on standard GPU architecture. We find over 100 trait-specific DNA variants that are exclusively exonic, with clear pathways from trait-specific “core genes” to each outcome. We separate pleiotropy from linkage to find evidence that PCSK9, LPA, and RP1-81D8.3 are pleiotropic for cardiovascular disease (CAD) with blood cholesterol, triglycerides, and low-density lipoprotiens respectively. CI-GWAS accounts for pleiotropy and selects waist-hip ratio, alcohol consumption and smoking as adjacent to CAD, with many other variables having complex paths linked through these risk factors. Our work facilitates extensive investigation of potential causal hypotheses across a wide-range of data.