Epistasis and Genome-Wide Association Studies
Jim Manzi writes that it’s plausible that epistatic interactions are central to complex mental phenotypes, and that they might therefore prevent genome-wide association studies from achieving much success. In the comments to a response post by Razib, Jason Malloy does a pretty good job of showing that traits like IQ are primarily additive and that epistasis therefore won’t prevent successful GWA with good sample sizes. [UPDATE 06-25-2009: I've read more behavior genetics, and I'm not quite sure that Jason's view is correct. I think it's still an open question, actually.]
With all that in mind, some epistasis does exist, and it is worth uncovering. It will not be uncovered directly by genome-wide searches, though, because of multiple testing issues. Even a two-dimensional search overwhelms foreseeable sample sizes. However, a multi-step approach could work by breaking down the multiple dimensions into individual searches. Say that gene-A and gene-B only have an effect when they appear together. Thus, a GWA should pickup an effect from either gene-A or gene-B (whichever has a higher minor allele frequency, presumably), even if that effect is smaller than the overall effect of having both of them. Now, suppose we identify via GWA that gene-A is contributing to the phenotype. We could then do a second scan for interactions and identify gene-B.
Of course, scientists are not limited solely to association searches. They can also harness biological evidence of epistasis to help identify candidates. Because traits like IQ are primarily additive, epistasis is not the overwhelming bogeyman that it might first appear, and it should be possible to tackle in the years to come.
Reference:
Hirschhorn, J. N., & Daly, M. J. (2005). Genome-wide association studies for common diseases and complex traits. Nature Reviews Genetics, 6, 95-108.
Labels: Genetics
Ben:
Thanks for the comment and pointer. A few quick things:
1. I actually haven’t said that we know a GWAS can’t find the set of genes that determine some normal mental state, just that (i) we shouldn’t assume that this is possible unti it is accomplished, and that (ii) potential epistatic interactions are one important reason why this may not be feasible with conceivable case and control group sizes.
2. I agree that if we were to stipulate that gene-evironment interactions, bichemical process within the organism that interact with the genome to contribute to the phenotypic outcome and so forth were all negliglble, and further stipulate that only first-order interaction terms between genes were significant, that it would be plausible to accomplish the task with conceivably-sized case and control groups, given the power calculations.
3. I’ve made the point in several places that other techniques are available, and that even if they weren’t, we should be hesitant to assert confidently that some as-yet-unachieved scientific result is either sure to come, or on the other hand, is impossible.
Thanks for the reply. I’ll update the first sentence to better reflect your position.
2. I agree that if we were to stipulate that gene-evironment interactions, bichemical process within the organism that interact with the genome to contribute to the phenotypic outcome and so forth were all negliglble, and further stipulate that only first-order interaction terms between genes were significant, that it would be plausible to accomplish the task with conceivably-sized case and control groups, given the power calculations.
From what I’ve read gene-environment interactions are actually much more important than gene-gene interactions for IQ. However, I don’t see why we need to assume they’re negligible to assert that GWA’s will likely identify the genes for IQ. If a given gene interacts with the environment it would still presumably show up in GWA’s. Discovering the specific gene-environment interactions would be the difficulty not identification of the genes involved.
Ben:
Thanks – that’s very much appreciated.
Here’s the reason that this could be the case. If the interaction effects are multi-layered, high-order and complex, teasing out the isolated causal impact of any genetic, non-genetic biolgoical or environmental factor (whether a specifc gene or some interaction term) could easily overwhelm the number of indepent data points available in the GWAS. This is directly analogous to problems in economics, social science and business, in which such correlation (i.e., association) studies are notoriously misleading and inconclusive.
The evidence has not been gracious about the existence of important gene-environment interactions either:
M. Mcgue and T. Bouchard: “The existence of genotype-environment interaction (G x E, or differential sensitivity of genotypes to environments) for psychological characteristics, although intuitively plausible, has been difficult to demonstrate empirically. Attempts to identify G x E effects for personality (Bergeman et al 1988) and general cognitive ability (Capron & Duyme 1989) did not yield significant findings.”
I.D. Waldman: “There are a number of potential explanations for the dearth of genotype-environment interactions in human behavior genetic research on intelligence. The first is simply that genotype-environment interaction may not be important for individual differences in intelligence in most populations, given the ranges of intelligence, genotypes, and environments represented in these populations. Given the extant data, this may be the most plausible explanation …” (p 558)
both those refs are from the late 90′s. many studies supporting gXe have come out since then.
..
what do you think are the environmental factors (if any) that effect IQ in a modern society like the US?
The $2.7 billion National Children’s Study with a sample size of 100,000 and a life of 21 years is gearing up. Does anybody know if they are going to give IQ tests to the mothers, fathers, and children, or have IQ tests been yanked out of the study for reasons of political correctness. (We wouldn’t want the data being used to write another Bell Curve!)
Ben, at least some of those studies (the ones by Caspi’s group) have been criticized right here at GNXP:
http://www.gnxp.com/blog/2007/11/questioning-breastfeeding-iq-fads2.php
Steve,
From their research plan PDF:
To measure potential genetic and cognitive influences of parental IQ on the child, the Kaufman Brief Intelligence Test, Second Edition (KBIT-2) will be used. It has the advantage of including a non-verbal scale (in addition to a verbal one) which is relatively invulnerable to SES factors and language background. It yields scores similar to other intelligence tests, with a mean of 100 and a standard deviation of 15, making it possible to compare KBIT-2 scores with other measures of IQ. The Woodcock-Johnson-III Tests of Achievement Letter-Word Recognition subtest, which measures the individual?s word decoding skills, will also be administered.
The research plan also details the exhaustive genetic and environmental measures they’re planning on taking.
Matt,
yea, the breastfeeding one seems highly questionable. i’m thinking of doing a post on the current state of thinking regarding gXe for IQ. i don’t think it’s as simple as Malloy would have it. As this review article from 2007 discusses, it’s a relatively young and weakly researched area with little consensus. And there are some apparently strong supporting data-points for the idea of gXe in IQ.
Accumulating more data may not help for these questions. The central problem is that the tools that we have now (DZ/MZ comparisons, adoption studies) are all biased towards magnifying genetic factors.
DZ/MZ neglect epistasis and the fact that MZ twins spend much more time together than DZ twins (it is likely that common behavioural genes play a part in this, but we can’t know how much, which is precisely the question we’re trying to solve). Any adoption study in an even moderately civilised country is thwarted by the fact that adoptive parents are not randomly selected from the population at large, but are either taken from close relatives, or thoroughly screened by a grueling process of selection by social services – to a point, adoption studies can be said to control for environmental variance.
Separated twins were once thought to be the solution to thes problems, until people realised that “separated” children were often not really separated (Shalizi cites a case where “separated” children actually went to the same school and classroom). That, plus the adoption problem.
I do agree with Jason that separated twins can (in theory) tell us whether epistasis plays a significant role. I’m just not sure that the current data is reliable enough for this. In particular, I’d like to see anything that doesn’t come from the Bouchard group (independent replication FTW!)
Of course, I would welcome any attempt to drag me out of my ignorance and/or misperceptions!