Mendelian randomization for investigating causal roles of biomarkers in multifactorial health outcomes: a lesson from studies on liver biomarkers

Mendelian randomization that uses genetic variants as instrumental variables (IVs) is an example of the acquisition of cross-disciplinary analysis. Whereas methods for IV analysis in econometrics or for control of confounding in non-randomized studies were proposed several decades ago,¹ formal applications of Mendelian randomization were considered in early 2000s.¹ A simple search term (‘Mendelian randomization’ or ‘Mendelian randomisation’) in Medline PubMed shows 911 titles that are mainly investigating causal roles of a wide range of exposures (e.g. a certain biomarker) in multifactorial health outcomes. The most recent application of Mendelian randomization is to validate drug targets or investigate the performance (efficacy and safety) profiling of such targets in clinical trials.^2,3

Here the main concept is that a natural experiment (i.e. randomization at the time of conception) is assumed on the basis of Mendel’s laws, and therefore integrating observational-genetic evidence in Mendelian randomization can serve as an analogous to the evidence of clinical trials.^4–6 Mendelian randomization provides complementary evidence for which confounding or reverse causality is less likely;⁴ however, given underlying assumptions and limitations of Mendelian randomization, a single study is insufficient to support the potential causal or non-causal roles of exposures in outcomes.⁵ For instance, some recent studies by Noordam et al. and others investigated the causal associations between liver biomarkers and cardio-metabolic outcomes in different populations or settings (Figure 1).^7–13 Most studies have used at least two Mendelian randomization methods in a single or multiple IV analysis, and some used both individual-level and publicly available data [from genome-wide association studies (GWAS) consortia] in their report. I really agree with the notion that such efforts provide valuable insights into our understanding of the aetiology of multifactorial health outcomes like type 2 diabetes, and I think that it is the right time to formulate some guides or criteria to ease interpretation and the crafting of overall conclusions from the rapidly increasing Mendelian randomization studies.

Figure 1.

Mendelian randomization studies for investigating the causal associations between liver biomarkers and cardio-metabolic outcomes. The size of the circles indicates a minus log10 P-value for each causal estimate. Dark (red) colour corresponds to a positive association (P-value < 0.05), light (blue) colour an inverse association (P-value < 0.05) and the lighest (grey) colour a null association. T2D, type 2 diabetes; HbA1c, glycated haemaglobin A_1c; HOMA-B and HOMA-IR, Homeostatic Model Assessment (HOMA) of β-cell function and insulin resistance (IR); CVD, cardiovascular disease; SBP, systolic blood pressure; DBP, diastolic blood pressure; ALP, alkaline phosphatase; ALT, alanine aminotransferase; GGT, gamma-glutamyl transferase; 2SLS, two-stage least squares; gtx, Genetics ToolboX; MGMM, multiplicative generalized method of moments; IVM, inverse variance-weighted; ER, Egger regression; WM, weighted-median; MA, meta-analysis (colour online).

Following several efforts in risk prediction,^14,15 in 2015 the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) initiative developed a guide for the reporting of studies developing, validating or updating a prediction model.¹⁶ Although there are some similarities between implementing aetiological research and prediction research (like definitions of data source, outcomes and using regression models), Mendelian randomization studies use different methods (e.g. IV analysis).^4,5 Also contrary to prediction research, confounding, the measure of statistical significance and the causal relevance of exposures (not the strength of the effect size nor quantifying the added predictive value of a certain biomarker) are very important in Mendelian randomization studies.^14,16,17

Taken together, increasing cross-disciplinary collaborations to address potential methodological issues (like power and sample size calculations, model assumptions and a Bonferroni correction) provide an opportunity to better leverage large-scale GWAS data for identifying a valid IV for exposures and investigating causal estimation.^5,18 As the field and the analytical methods have been evolving, development of guides or criteria for designing and reporting a Mendelian randomization study helps improve the transparency and quality of evidence¹⁶ and prioritize replication studies where needed, and aids the readers to evaluate consistency (e.g. direction and statistical significance) in causal estimation and the quality of published studies.