Using the MR-Base platform to investigate risk factors and drug targets for thousands of phenotypes

Principles of Mendelian randomization

Mendelian randomization is a method to assess the causal effect of an exposure on an outcome using an instrument, defined by one or more SNPs, as a proxy for the exposure. The SNPs are used as instrumental variables and must meet three conditions: (i) they must be associated with the exposure; (ii) they must only affect the outcome via the exposure; and (iii) there must be no factor that causes both the SNP and outcome. These conditions are known as the instrumental variable assumptions and are illustrated in Figure 1. SNPs are plausible instruments because they are determined at conception and generally cannot be subsequently affected by the environment⁴. If these assumptions hold, then Mendelian randomization effect estimates are unlikely to be due to confounding or reverse causation. However, Mendelian randomization is still subject to important limitations. For example, effect sizes may not be indicative of the effects of a clinical intervention later in life. This can occur for several reasons, such as cumulative exposure, where Mendelian randomization estimates may reflect the effect of lifelong exposure, or time-dependent exposure, where intervention outside of a critical period does not have an effect despite the Mendelian randomization estimates suggesting an effect exists. Mendelian randomization estimates may also be affected by issues such as horizontal pleiotropy, whereby the SNPs chosen to proxy the exposure may affect the outcome by pathways other than the exposure of interest leading to biased results^1,5–11.

Figure 1. Overview of the instrument assumptions.

Methodological advances mean that Mendelian randomization can be implemented using summary statistics from GWAS, without individual level data¹². In addition, the SNP-exposure associations and the SNP-outcome associations can be obtained from separate datasets, known as two-sample Mendelian randomization¹². This is in contrast to one-sample Mendelian randomization, where both the exposure and outcome are measured in all individuals from the same sample. As a result, two-sample Mendelian randomization can exploit much larger sample sizes and estimate effects with higher precision than is typically possible using any single sample. It also allows access to large case control studies of disease outcomes that may not have measured the exposure of interest. This can be particularly beneficial when considering expensive or difficult to measure phenotypes, such as DNA methylation, metabolomics, proteomics, and gene expression⁸.

Detailed discussion of the theory and interpretation of Mendelian randomization results, including the limitations of the method, can be found elsewhere^1,5–10. Key definitions used throughout the present discussion are given in Table 1.

Potential applications of MR-Base

MR-Base² is suitable for a broad range of applications and, consequently, is intended for use by a broad range of professionals. These include clinical and non-clinical researchers, public health specialists, policy makers and those in the pharmaceutical industry. Some of the key ways in which the platform can be used are summarised below and specific use cases are discussed in the Use cases section.

Implementation

MR-Base² can perform two-sample Mendelian randomization and provide summary statistics from a range of GWAS for this purpose. As highlighted previously, it can be accessed through a web application or as an R package. Data can either be accessed through the platform or be uploaded by the user, both of which are demonstrated in the following Use cases section. Data harmonization between the SNP-exposure associations and the SNP-outcome associations and Mendelian randomization are then performed according to options specified by the user using buttons on the web application or commands for the R package.

Operation

The web application can be accessed from any platform that allows the use of a java-script compatible graphical web browser. The R package can be accessed from any platform where R version 3.5 or later can be installed. Step-by-step instructions for the web application and code for the R package are available for each of the use cases (see Software availability)²². A generalized workflow for using the MR-Base web interface is provided in Box 3.

Use case 1: subjective wellbeing and cardiometabolic health

Our first use case demonstrates the rapid implementation of Mendelian randomization using MR-Base to investigate a risk factor proposed in the observational literature and enable transparency of the Mendelian randomization study. The specific workflow for this case study is provided, alongside the necessary code, data and results, on GitHub (see Software availability)²². It is based on work by Wootton et al. that used MR-Base to investigate the association between subjective wellbeing and 11 measures of cardiometabolic health²³. It has been reproduced based on information in the paper and, in particular, their code on GitHub. Studies with data on both subjective wellbeing and measures of cardiometabolic health are rare. Therefore, the authors chose to use two-sample Mendelian randomization so that they could use separate samples for their exposure (sample 1) and outcome (sample 2) phenotypes and use UK Biobank as a single-sample Mendelian randomization sensitivity analysis. In addition to this, the largest available GWAS of subjective wellbeing at the time of the study had identified just three SNPs to instrument this phenotype at the conventional genome-wide significance level and only one of these three SNPs replicated in an independent sample. Consequently, the authors used a lower p-value threshold of P < 5×10^-5 to increase the number of SNPs in their instrument. This potentially increases their instrument strength but may also increase their susceptibility to weak instrument bias or pleiotropy.

It was straightforward to use two-sample Mendelian randomization with MR-Base in this study as it allowed the authors to consider multiple outcomes simultaneously and to use summary data that was already formatted for analysis. It also allowed specification of their preferred p-value threshold for the instrument. To make allowance for the non-independence of the selected SNPs, it was necessary to prune the SNPs to determine an independent set for the analysis. MR-Base allows users to select independent SNPs through a process known as ‘clumping’, which identifies independent signals by considering the linkage disequilibrium between SNPs. Linkage disequilibrium refers to the allelic association between groups of SNPs, which are typically located in a similar region of the genome. Failure to consider the linkage disequilibrium between SNPs can lead to overestimation of instrument strength and overly precise effect estimates. MR-Base overcomes this by picking the SNP from the group of SNPs in linkage disequilibrium that has the strongest evidence of association with the exposure for use in the Mendelian randomization analysis. In the Wootton et al. study, clumping reduces the subjective wellbeing instrument from 724 SNPs to 84 SNPs, highlighting the importance of this step in the analysis.

Use case 2: systolic blood pressure and coronary heart disease

The second use case demonstrates how MR-Base can be used to conduct sensitivity analyses and triangulate evidence. The specific workflow for this case study is provided, alongside the necessary code, data and results, on GitHub (see Software availability)²². Sample code for using the ‘TwoSampleMR’ R package based on this use case is provided in Box 4. It is based on work by Ference et al. that examined the effect of systolic blood pressure on coronary heart disease²⁴. Here, we recreate the Mendelian randomization component of the work. This can be triangulated with evidence from meta-analyses of prospective observational studies and randomized controlled trials. These meta-analyses found that lower systolic blood pressure reduced risk of coronary heart disease with odds ratios of 0.75 (95% CI: 0.71 to 0.78; p = 0.006) and 0.83 (95% CI: 0.76 to 0.90; p = 0.001), respectively. A full discussion of the triangulation element of this research is provided by Lawlor et al.¹⁶

# Load the TwoSampleMR package

library(TwoSampleMR)

# List the outcomes available in MR-Base

ao <- available_outcomes()

# Extract the instruments from the systolic blood pressure GWAS (ID: 'UKB-a:360')

exposure_dat <- extract_instruments(c('UKB-a:360'))

# Extract the outcome data from the coronary heart disease GWAS (ID: 7)

outcome_dat <- extract_outcome_data(exposure_dat$SNP, c('7'),
                                    proxies = 1, rsq = 0.8, align_alleles = 1,
                                    palindromes = 1, maf_threshold = 0.3)

# Harmonize the exposure and outcome data

dat <- harmonise_data(exposure_dat, outcome_dat, action = 2)

# Perform MR analysis

mr_results <- mr(dat)

In our results, as with the original paper, we were concerned about directional pleiotropy, which occurs when genetic variants affect the outcome independently of the exposure. This is because large GWAS, such as the GWAS of systolic blood pressure we use here, may identify SNPs of unknown function²⁵. To assess the effect of this upon our results, we can look at the MR-Egger regression intercept provided by default on the MR-Base web application and calculable using the TwoSampleMR package for R. The intercept provides an estimate of the magnitude of horizontal pleiotropy and in our case is -0.0087 (SE: 0.0059; p = 0.147). This suggests limited evidence for directional pleiotropy among our results. We also used several Mendelian randomization methods for our analysis as a further sensitivity analysis and found consistent results for the effect of increased systolic blood pressure on coronary heart disease, regardless of the method used (IVW - OR: 1.76, 95% CI: 1.48 to 2.10, p = 3.92e-10; MR Egger - OR: 2.64, 95% CI: 1.49 to 4.68, p = 1.09e-03; Weighted median – OR: 1.77, 95% CI: 1.52 to 2.05, p = 3.92e-10; Weighted mode – OR: 1.77, 95% CI: 1.26 to 2.49, p = 1.31e-3).

Use case 3: HMGCR and type 2 diabetes

Our final use case demonstrates how MR-Base can be used to replicate a study and appraise a potential pharmaceutical intervention. The specific workflow for this case study is provided, alongside the necessary code, data and results, on GitHub (see Software availability)²². It is based on research by Swerdlow et al. that investigated the effect of 3-hydroxy-3-methylglutaryl-CoA reductase (HMGCR), the target of statins, on risk of type 2 diabetes²⁶. This study used a single SNP as an instrument: rs17238484. To demonstrate the features of MR-Base, we have uploaded the data necessary to define the instrument for this analysis, instead of using data already within the platform. These data were extracted from the 2013 GWAS by the Global Lipids Genetics Consortium²⁷. Once uploaded, the column names must be mapped to those used by MR-Base before the analysis can be run. If you are using the R package, there are equivalent commands that perform data formatting (see this guide). Although units are not required for the analysis to run, it is important that the units of the instrument-exposure and instrument-outcome effects are known, as this determines the interpretation of the effect estimate obtained by Mendelian randomization.

If you use MR-Base, please cite the resource using reference³. We also ask that you cite and acknowledge the studies that contributed the data and methodology used in your analysis.

Underlying data

Source data for the use cases are available through the database integrated into the MR-Base platform. The database is large and contains data from over 20,000 GWAS; therefore, it is not possible to host this data on an external repository. The data can be accessed via the MR-Base web application or by using the TwoSampleMR package for R to interact with the application programming interface. Users are required to accept the data access agreement by logging in with a Google account before access to the data is granted.