Design and interpretation of eQTL-GWAS colocalisation studies: Lessons from a large-scale evaluation

Colocalisation analysis integrates GWAS and molecular QTL datasets to identify candidate effector genes. Even with a wide range of molecular QTLs, 40% or more of GWAS loci remain unexplained, leaving a "colocalisation gap". We systematically characterised two large-scale eQTL colocalisation studies, to describe the determinants of this gap and ultimately inform the selection and design of eQTL studies to close the gap. We analyse over 1.3 million colocalisation tests from Open Targets Genetics (OTG) and perform and analyse colocalisations from 14 immune-mediated disease (IMD) GWAS and 12 diverse immune cell eQTL studies, selected to cover a range of cellular granularities and sample sizes. We find that 50% of GWAS peaks in OTG and 34% in IMDs colocalised and were more likely to colocalise if they were located nearer to genes and had a more common lead variant. Colocalisation was also more likely to occur in disease relevant tissues. The lowest granularity immune cell eQTL studies had the largest sample sizes, the greatest eQTL discovery and produced the largest number of colocalisations, particularly for lower-frequency variants. However, while higher resolution eQTL studies detected fewer eQTLs, each of those eQTLs was more likely to colocalise with a GWAS peak, emphasising the importance of cell specific eQTLs. Indeed, over 50% of colocalisations were found in only one cell type. Overall, our results suggest that a diverse set of cells in different contexts, and large, high granularity studies will be needed to identify remaining colocalisations. In addition, we observed that 47% of GWAS peaks colocalised with multiple genes in OTG and 37% in IMDs. Through simulations, sensitivity analyses, and integration of enhancer-promoter capture data we find that multiple colocalisations likely represent coregulation. While disentangling causality from horizontal pleiotropy will ultimately require experimental perturbation, triangulation using different sources of observational data is likely to be necessary for gene prioritisation.