SRF: The 2009 schizophrenia GWAS paper (see SRF related news report) reported that common variation accounts for about one-third of total genetic risk for schizophrenia, and this was supported in a study by Wray and Visscher. Does this new study update that heritability estimate?
O'Donovan: No, we decided that there had been enough work on estimating heritability, and actually we did not revisit that particular aspect in this paper. However, Naomi Wray and a subgroup of the PGC are in the process of using the current data to see whether or not a larger dataset makes any impact on estimated inheritability. But it’s not really expected that that estimate will substantially change. Those estimates were reasonably robust to sample size, although the confidence interval around the estimate is dependent on sample size.
SRF: What’s the next numerical goal for the schizophrenia GWAS, and do you have a final sample size that you’re aiming for? How far along are you in that?
O'Donovan: This is a question that a number of us have been raising across the course of the last couple of years—the question of when you get to very diminishing returns—and at the moment I don’t suppose from our discussions within the PGC that anyone has got a very clear estimate of when the information accrued will plateau out. So the short-term goal would be to approximately double the key sample size by mid-2015, and I guess in doing that we would have a better picture of whether or not we’re still on the upward trajectory or whether or not discovery is leveling out. My suspicion is we'll still be on the upward trajectory. A second question is whether or not by discovering more you fill in more of the likely biology as well as just getting more associated variants.
SRF: What are the strategies for following up on the hits that you have?
O'Donovan: Well, for the common variant analysis, we’re really looking at enhancing sample size and using the chip that was developed in collaboration with the PGC. The primary movers in developing the chip were Ben Neal and Pamela Sklar. That’s quite an economical chip, and there are resources to genotype a few tens of thousands more cases. More individual groups are also signing up to join in the PGC endeavor, and I think through both of those routes we should be in a position to achieve the doubling of the sample size.
SRF: Do new and bigger chips come along that you then have to take into account, and if you do have a new chip that comes along, do you have to go back and genotype people you already genotyped with the previous chip?
O'Donovan: Well, the psych chip that is being used primarily for the common variants actually has fewer common variants than some of the chips that have already been used. It’s been designed with a balance of low cost but also extracting a high proportion of the common variation, and the chip has also had a lot of additional common variants added to it based upon even fairly weak association signals in the PGC data up until now. So I don’t think it’s necessary that we have to go back and re-genotype the earlier samples. There may be some advantages in genotyping the relatively small numbers of samples that were genotyped with very, very old chips. But the vast majority of the samples that have come in the PGC aren’t requiring going back to re-genotype.
SRF: Do you expect to find some number of false positives, and do you have any sense, even at this relatively high sample size, what that false positive rate might be? Are there any data from other areas that would help you understand if you had any false positives?
O'Donovan: Ultimately, the association tests are statistical tests, and in any statistical test there is a false-positive rate. At the genomewide-significant threshold of 5 x 10-8, maybe you would expect one study in 20 would get a false positive if that study has been well designed and appropriately controlled. However, if you look at the way previous strong findings are holding up in the bigger datasets, I think that what we would expect is that virtually all of the current findings are likely to be true positives, but you can’t exclude the possibility that a small number would be false positives.
SRF: And a strong finding is defined by P value?
O'Donovan: Well, it’s defined by surpassing the genomewide-significant threshold, providing you can be confident in the design and analysis. But it's worth casting an eye on sample size. You always have to ask if a finding is credible, given what we know about effect sizes and allele frequencies; in other words, is a sample size big enough to have a chance of having a true finding?
SRF: Is there a difference between something that’s 10-22 and 10-8 just over the threshold?
O'Donovan: Well, yes, the stronger the actual statistical evidence, the smaller the chance of it being a false positive. It’s a continuum.
SRF: All other things being equal, is a 10-22 hit a better candidate for follow-up than a 10-8 candidate?
O'Donovan: Kind of, but it's more complicated. A lot depends on what you mean by follow-up. If it's basic biological experiments or modeling, probably a better guide is whether the association is likely in a gene or miles away from any known gene. Is it tightly focused or spread across a large number of genes? If it's in a gene, is anything already known about it? Are there resources available like antibodies, and so on?
SRF: The CLOZUK sample ascertained by clinicians rather than research diagnoses seems to have really helped get the sample size up. Will future samples also include clinically defined samples?
O'Donovan: Yes is the answer. The CLOZUK sample was quite instrumental in increasing sample size. I believe that the interim presentation by Stephen Ripke back in 2012, when CLOZUK was folded into the PGC and we could see the number of associations was really ramping up, was probably inspirational in getting other people to sign up to the PGC. And yes, some of the samples being genotyped at the moment will be clinical diagnoses like the CLOZUK sample itself. James Walters is the person who has led on this. We have already more than doubled the CLOZUK sample, and we are still collecting.
SRF: What are your thoughts on epistasis? You didn’t find evidence when you tested pairs in this study, but what about other kinds of interactions through three hits or variants that aren’t on this chip? It seems as though you could have unending scenarios of “what if?” Is that something that is important for the analysis that you’re currently doing with the GWAS, and is it important to think about epistasis for common variance in general in psychiatric disorders?
O'Donovan: Well, a fair statement is that as a consortium we're agnostic about epistasis. Certainly in the past I have been interested in it, and you know various members of the consortium are very strongly interested in it, though others are less so. Now, all we can say is that we don’t have evidence for its widespread operation. We don’t have any evidence for its operation at all in the current dataset. Having said that, as you'll be aware from your question, we performed a limited analysis using just the genomewide-significant hits but didn’t find anything notable. There are about four or five epistasis subprojects within the PGC taking different tacks. Some will include high-order interactions, but the results of those analyses are not yet in. So we're still in a position where it would be unwise for us to say, Yes, it’s widespread, or no, it’s not.
SRF: Among the issues raised in the discussion of the study on SRF (see SRF related news report) was the comment, from a drug development perspective, that ApoE was identified as an odds risk factor of 4 for Alzheimer's disease more than 20 years ago, yet there is no drug development program focused on ApoE, suggesting maybe there’s not much value in following up on GWAS hits.
O'Donovan: Well, I think you can’t always generalize from a single example, like the famous phrase, “One swallow does not make a summer.” We are open to the possibility—the hope—that identifying genomewide-significant associations may directly result in identification of targets. But it may very well be the case that it doesn’t. I think it’s a reasonable assumption that if we can gain traction on the biology of schizophrenia through common or rare variant analysis, that at least increases the probability that new treatments will be developed.
So are any of the things we identified direct targets? Well, the association already points in the vicinity of the dopamine D2 receptor; you can argue that that’s a kind of reverse proof of principle. If we didn’t know that D2 is a reasonable target for drugs, one might have eventually tested it based on our data, though I don’t know how drug companies operate.
But regardless of all that, it does seem to me a not very controversial statement that if we improve the biological understanding of etiology, we must increase the probability of being able to develop treatments.
SRF: Beyond the follow-up research, do you have a feeling that these results can have some impact on the explanations of schizophrenia in the public sphere, for example, questions about whether it has a biological basis? There are some very vocal people who reject the idea, who hold very much to a mind-brain dualism. Do you see these findings, in particular this paper, as being a watershed in some way, and how would you communicate what this paper means to those who don’t know a lot about biology and genetics?
O'Donovan: The people who tend to be very vocal in a very anti-biological way do not tend to be persuaded ever by data like these. I’ve been engaged in a few conversations now on a blog site called The Conversation. Now, admittedly, that’s only one vocal critic of biological psychiatry, but it’s pretty clear that studies like this do not influence people who do hold the most extreme views. What I hope is that work like this influences people whose minds are open to new evidence. That’s all one can hope for.
But cultural changes take sometimes generations. One study, no matter how big, no matter how many things it has found, does not change the culture. How do we communicate?
Even within psychiatry there are people who will not have it that psychiatric diagnosis exists in any valid way or that there is any point in diagnosis. That the biological route is misguided. And I think we try our best to explain that through conferences, user groups, blogs. But I think eventually if and when better treatments emerge directly from work of this nature, that’s probably what will change a lot of people’s minds. But there always will be people whose minds remain closed.
The people who do not accept that any treatments work at all—it’s pretty difficult to see them as people you really expect to change.