G × E in child psychiatry and psychology: a broadening of the scope of enquiry as prompted by Munafò et al. (2014)

Rutter’s commentary provides us the opportunity to clarify some issues that he (and therefore, we suspect, others) may have misunderstood.

First, we do not dispute the existence of gene x environment (GxE) interactions. Rutter believes we are mounting a “one-sided attack on GxE in relation to life events”. This is simply not so. We have no argument against the existence of GxE in the aggregate. Everyone is familiar with the idea that when two people encounter the same environmental stressor the results can be very different, and it makes sense to attribute this variation in part to differences in genetic makeup. Twin and family data provide formal tests of this idea, and confirm its importance (Cadoret, Yates, Troughton, Woodworth, & Stewart, 1995; Cloninger, Sigvardsson, Bohman, & von Knorring, 1982; Kendler & KarkowskiShuman, 1997). There are also convincing data from other species for the existence of GxE (Caligari & Mather, 1975; Crabbe, Wahlsten, & Dudek, 1999; Henderson, 1976; Valdar et al., 2006). Indeed we fully concur with the three points Rutter makes at the end of his commentary as the main messages for practitioners. But those points have nothing to do with the intent of our article.

So what are we disputing? Even after the most cursory survey of the literature one cannot avoid the conclusion that the strength of evidence for the effects of GxE at a single locus (be it the serotonin transporter or the monoamine oxidase-A gene) is contested. We want to understand why there is still disagreement (after more than 10 years), and what could be done to resolve the issue. For clarity, from this point on we use the terms locus-specific GxE and genome-wide GxE. Our concern is that the results supporting the existence of locus-specific GxE for psychiatric phenotypes in human studies are not robust.

Second, Rutter’s commentary displays a common misunderstanding that the choice of statistical model can be determined in terms of biological plausibility. Our knowledge of biological mechanisms underlying multifactorial complex diseases does not allow us to choose one model over another to test specific hypotheses about underlying pathological processes. Indeed, the example we provide in Table 1 of our article demonstrates why evidence of GxE cannot be interpreted in terms of biological mechanisms. The fact that one can reach opposing conclusions from the same data by using different models does not mean, as Rutter’s commentary suggests, that it is invalid or impossible to compare such models. It simply demonstrates why inferences about biological mechanisms based on patterns of statistical interaction are inherently flawed. The pattern of interaction observed in single gene disorders such as phenylketonuria is an exception to this. However, contrary to Rutter’s interpretation, a biological interaction of this nature would show evidence of statistical interaction under any statistical model used (i.e., show model-independence).

The preference for additive models expressed by Kendler and Gardner (Kendler & Gardner, 2010), as cited in Rutter commentary, is based on the fact that evidence of greater than additive effects tells us that some individuals only developed the outcome because they were exposed to both G and E (i.e., that both factors co-participate in the same causal model of disease). Although this is confusingly described as ‘biological interaction’, it does not tell us anything about underlying biological mechanisms – a point emphasised by Kendler himself (Kendler & Gardner, 2010) – and in fact can also be inferred from lack of evidence of GxE under multiplicative models.

Third, running through the commentary is the assumption that genome-wide association studies (GWAS) of psychiatric disease have failed (“there are few successes of GWAS in the field of mental disorders, schizophrenia being a partial exception to this”), and that successful genetic mapping requires the inclusion of locus-specific GxE (particularly for studies of major depression). While it would certainly be interesting to include environmental measures in genetic analysis, we don’t think it is either necessary, or currently practical.

We now know that the genetic architecture of psychiatric disease is highly polygenic. For all psychiatric conditions that have been examined, including major depression, bipolar disorder, schizophrenia, autism, and attention deficit hyperactivity disorder (Lee et al., 2013), and Tourette syndrome and obsessional compulsive disorder (Davis et al., 2013), between one third and a half of the heritability is due to the combined direct effects of many segregating sequence variants distributed across the genome, each making a small individual contribution. This means that genome-wide significant effects can be detected by increasing sample size. The recent success with schizophrenia illustrates this well (Schizophrenia Working Group of the Psychiatric Genetics Consortium, in press), and in this respect progress is no different than for other non-psychiatric complex traits. Finding 8 loci contributing to the risk of hypertension required a sample size of 34,433, with follow up in a further 113,250 subjects (Newton-Cheh et al., 2009). Success does not require locus-specific GxE analysis – it just needs a larger sample size. Of course, it’s possible that the sample size needed will be very large indeed – 50,000, for example, for major depression (Flint & Kendler, 2014; Wray et al., 2012) – and that including locus-specific GxE might reduce the necessary sample size. However, genotyping costs continue to fall, as does the cost of collecting diagnostic data on a large scale (Perlis et al., 2012), while the costs of the in-depth phenotyping required to assess environmental effects have not declined. Unless locus-specific GxE effects are large enough that they allow a very substantial reduction in sample size (and so far there is no evidence that this is so), then it is more efficient to identify loci that robustly contribute a main effect before proceeding to test for locus specific GxE. In other words, we should focus efforts on detecting main effects.

In our view, attempts to resurrect candidate genes that have not provided robust evidence of a main effect by appealing to putative locus-specific GxE effects is a retrograde step. Our understanding of the genetic architecture of complex traits has moved on. In contrast to the chequered history of candidate gene studies, unbiased genome-wide approaches are revealing insights into the neurobiology of complex traits. As genome-wide association studies identify main effects of individual loci, these can then be interrogated further in order to understand the mechanisms they underpin, and to explore potential environmental effect modifiers.