Combining evidence from Mendelian randomization and colocalization: Review and comparison of approaches

Abstract

Mendelian randomization and colocalization are two statistical approaches that can be applied to summarized data from genome-wide association studies (GWASs) to understand relationships between traits and diseases. However, despite similarities in scope, they are different in their objectives, implementation, and interpretation, in part because they were developed to serve different scientific communities. Mendelian randomization assesses whether genetic predictors of an exposure are associated with the outcome and interprets an association as evidence that the exposure has a causal effect on the outcome, whereas colocalization assesses whether two traits are affected by the same or distinct causal variants. When considering genetic variants in a single genetic region, both approaches can be performed. While a positive colocalization finding typically implies a non-zero Mendelian randomization estimate, the reverse is not generally true: there are several scenarios which would lead to a non-zero Mendelian randomization estimate but lack evidence for colocalization. These include the existence of distinct but correlated causal variants for the exposure and outcome, which would violate the Mendelian randomization assumptions, and a lack of strong associations with the outcome. As colocalization was developed in the GWAS tradition, typically evidence for colocalization is concluded only when there is strong evidence for associations with both traits. In contrast, a non-zero estimate from Mendelian randomization can be obtained despite only nominally significant genetic associations with the outcome at the locus. In this review, we discuss how the two approaches can provide complementary information on potential therapeutic targets.

Keywords: Causal inference; Genetic epidemiology; phenome-wide association study; post-GWAS investigations; shared heritability.

Figure 1

Schematic diagram illustrating analogy between Mendelian randomization and randomized trial Adapted from Hingorani and Humphries.

Figure 2

Schematic diagrams illustrating colocalization in five scenarios (A) Two traits with distinct causal variants in linkage disequilibrium. (B) Two unrelated traits with a shared causal variant. (C) Two traits with a shared causal variant where the first trait influences the second trait. (D and E) One shared causal variant and one distinct causal variant for trait 2. Scenarios (B) and (C) are examples of colocalization. For scenarios (D) and (E), there is colocalization at the shared variant, but not at the distinct variant. Colocalization is unable to distinguish between the scenarios in which trait 1 and trait 2 are causally unrelated (scenarios B and D), and in which trait 1 has a causal effect on trait 2 (scenarios C and E). Illustrative regional association plots for each scenario represent the negative log₁₀ p values for associations of variants with each trait (blue for trait 1, red for trait 2) plotted against chromosomal position.

Figure 3

Scatter plots of genetic associations with LDL cholesterol, coronary heart disease, and Alzheimer disease Genetic associations with LDL cholesterol (horizontal axis, standard deviation units) against genetic associations with (A) coronary heart disease and (B) Alzheimer disease (vertical axis, odds ratios) for 75 genetic variants associated with LDL cholesterol. Error bars represent 95% confidence intervals for the genetic associations; dashed line represents inverse-variance weighted estimate (dotted lines represent 95% confidence intervals for this estimate). In the right-hand plot, variants in the APOE gene region are marked with triangles.

Figure 4

Regional association plots for the PCSK9 gene region Genetic associations (negative log₁₀ p values) plotted against chromosome position (megabases, Mb) for variants around the PCSK9 gene region with LDL cholesterol, coronary heart disease risk, and Alzheimer disease risk. Note the well-defined peak around the lead variant for both LDL cholesterol and coronary heart disease (marked in red), and the absence of a well-defined peak around any lead variant for Alzheimer disease. Colocalization suggests that LDL cholesterol and coronary heart disease have a shared causal variant, which is this lead variant, and no evidence that there is a causal variant for Alzheimer disease at this locus. Figures were made using the karyoploteR package: http://bioconductor.org/packages/release/bioc/html/karyoploteR.html.

Figure 5

Regional association plots for the APOE gene region Genetic associations (negative log₁₀ p values) plotted against chromosome position (megabases, Mb) for variants around the APOE gene region with LDL cholesterol and Alzheimer disease risk. Note the well-defined peak around the lead variant for both traits (marked in green for LDL cholesterol, and blue for Alzheimer disease). However, in this case, colocalization suggests the peaks have distinct causal variants.