Assessing causal relationships in genomics: From Bradford-Hill criteria to complex gene-environment interactions and directed acyclic graphs

Abstract

Observational studies of human health and disease (basic, clinical and epidemiological) are vulnerable to methodological problems -such as selection bias and confounding- that make causal inferences problematic. Gene-disease associations are no exception, as they are commonly investigated using observational designs. A rich body of knowledge exists in medicine and epidemiology on the assessment of causal relationships involving personal and environmental causes of disease; it includes seminal causal criteria developed by Austin Bradford Hill and more recently applied directed acyclic graphs (DAGs). However, such knowledge has seldom been applied to assess causal relationships in clinical genetics and genomics, even in studies aimed at making inferences relevant for human health. Conversely, incorporating genetic causal knowledge into clinical and epidemiological causal reasoning is still a largely unexplored area.As the contribution of genetics to the understanding of disease aetiology becomes more important, causal assessment of genetic and genomic evidence becomes fundamental. The method we develop in this paper provides a simple and rigorous first step towards this goal. The present paper is an example of integrative research, i.e., research that integrates knowledge, data, methods, techniques, and reasoning from multiple disciplines, approaches and levels of analysis to generate knowledge that no discipline alone may achieve.

Figure 1

DAG demonstrating the ideas of confounding. A: U is an unobserved confounder for the association between X and Y and X is a cause of Y. B: U is an unobserved confounder for the association between X and Y but X is not a cause of Y. From purely observational data these two situations cannot be separated.

Figure 2

Three DAGs exhibiting the same conditional independence but with different causal interpretations.

Figure 3

DAG with a randomisation node R. R indicates whether X is randomised or allowed to arise naturally. A: U is a confounder. B: U is a mediator. Randomisation allows us to distinguish between these situations.

Figure 4

DAG showing all possible one way relationships for gene-environment interactions based on the observed variables.

Figure 5

Both DJ-1 gene and pesticide exposure need to be present to activate the interaction.

Figure 6

Pesticide has an effect but DJ-1 only has an effect if pesticide exposure is present.

Figure 7

DJ-1 has an effect but pesticide only has an effect if the gene mutation is present.

Figure 8

Both DJ-1 and the pesticide have an effect and there is a possible interaction in A but not in.

Figure 9

DAG representing the fruit-fly experiment where interventions were performed both on the genetic make-up and the pesticide exposure. The interaction can therefore be identified.