Genetic drug target validation using Mendelian randomisation

Abstract

Mendelian randomisation (MR) analysis is an important tool to elucidate the causal relevance of environmental and biological risk factors for disease. However, causal inference is undermined if genetic variants used to instrument a risk factor also influence alternative disease-pathways (horizontal pleiotropy). Here we report how the 'no horizontal pleiotropy assumption' is strengthened when proteins are the risk factors of interest. Proteins are typically the proximal effectors of biological processes encoded in the genome. Moreover, proteins are the targets of most medicines, so MR studies of drug targets are becoming a fundamental tool in drug development. To enable such studies, we introduce a mathematical framework that contrasts MR analysis of proteins with that of risk factors located more distally in the causal chain from gene to disease. We illustrate key model decisions and introduce an analytical framework for maximising power and evaluating the robustness of analyses.

Fig. 1. Directed acyclic graphs of potential Mendelian randomisation pathways.

Nodes are presented in bold face, with G representing a genetic variant, P a protein drug target, X a biomarker, D the outcome, and U (potentially unmeasured) common causes of both P, X, D. Labelled paths represent the (causal) effects between nodes.

Fig. 2. Instrument selection related variation in the point estimates of drug target Mendelian randomisation studies on the lipid’s association with CHD.

Each estimate is based on randomly (500 iterations) selecting 4 SNPs out of 17 HMGCR, 30 PCSK9, 21 NPC1L1, 36 CETP candidate variants. Lipids data were used from the GLGC and linked to coronary heart disease data from CardiogramPlusC4D. estimates were grouped by the inclusion of instruments with worsted predicted functional or regulatory consequence; categories occurring less than five times were removed. Any pairwise LD was accounted for using the 1000 genomes ‘EUR’ reference panel and a generalised least squares method. The boxplots depict quartiles 1, 2 (median), and 3 as a box, with the whiskers presented as vertical bars and values ±1.5 times the interquartile range as dots.

Fig. 3. Mendelian randomisation estimates of the lipids weighted associations with CHD under increasingly liberal LD-clumping thresholds.

Lipids data were used from the GLGC, and linked to coronary heart disease data from CardiogramPlusC4D. Pairwise LD remaining after LD-clumping was accounted for using the 1000 genomes ‘EUR’ reference panel and a generalised least squares method. Estimates for PCSK9, HMGCR, and NPC1L1 are given per SD in LDL-C, CETP estimates per HDL-C reflecting the likely effectiveness pathway to CHD. The number of included variants is depicted above the x-axis. Estimates are given as OR with 95%CI (vertical error bars).

Fig. 4. Mendelian randomisation estimates of the lipids weighted associations with CHD stratified by functionally of the included variants.

Lipids data were used from the GLGC, and linked to coronary heart disease data from CardiogramPlusC4D. Pairwise LD remaining (after clumping on R-squared of 0.60) was accounted for using the 1000 genomes ‘EUR’ reference panel and a (GLS) generalised least squares method. Estimates for PCSK9, HMGCR, and NPC1L1 are given per SD in LDL-C, CETP estimates per HDL-C reflecting the likely effectiveness pathway to CHD. Estimates are given as OR with 95%CI (vertical error bars). Numerical details, including the number of variants used, are provided in Supplementary Tables 6–9.

Fig. 5. Mendelian randomisation estimates of protein level effects on CHD, with a grid of LD threshold.

Pairwise LD was accounted for using the 1000 genomes ‘EUR’ reference panel and a (GLS) generalised least squares method with or without Egger correction for possible horizontal pleiotropy. The number of included variants in the 1 mega base flanking region is depicted above the x-axis of the top panels. Estimates are given as OR with 95%CI (vertical error bars). The top panel depicts the variant to CHD or protein level effect for clumping threshold 0.5 for CETP (based on an Egger correct GLS model), and at 0.4 for PCSK9 using an IVW GLS model; 38 and 9 variants, respectively. Notice that the PCSK9 estimates were only available on the natural logarithmic scale.

Fig. 6. A pQTL based drug target MR phenome wide scan.

Results are presented as odds ratios with 95% confidence intervals (horizontal lines), positioned in the protein increasing direction. The total number of events and sample size are provided in the forest plot.

Fig. 7. A proposed drug target MR analysis framework.

The influence of LD and genetic region can be explored (and optimized) through simple grid-searching. Robustness of model choices in LD reference panel, the selection of functional or regulatory variants, and outlying or influential (high leverage) variants can be explored thorugh sensitivity analyses showcased here and in the supplementary analyses.