Response to Mitchell and Porteus.

Abstract

We appreciate the opportunity to respond to the letter from Drs Mitchell and Porteus (1) regarding our paper from the Psychiatric GWAS Consortium describing a framework for interpreting genome-wide association studies (GWAS) for psychiatric disorders (2). The central assertion of the Mitchell and Porteus letter is that the causation of psychiatric disorders is mostly, perhaps entirely, under a multiple-rare variant model (MRV) as opposed to a common-disease/common-variant model (CDCV). Moreover, Mitchell and Porteus predict that any CDCV will explain only a very small fraction of the heritability of psychiatric disorders. Both points are inconsistent with the scientific evidence. There have been multiple reports in high-profile journals in the past year describing the identification of previously unknown common genetic variants for autism, bipolar disorder, and schizophrenia. Most of these findings have replicated in multiple samples with levels of significance that render false positives due to the play of chance unlikely (Table). Thus, the exclusivist MRV position advocated by Mitchell and Porteus is inconsistent with empirical data. Table Notable CDCV findings for three neuropsychiatric disorders. Mitchell and Porteus claimed that “common alleles will explain only a very small fraction of the heritability of psychiatric disorders”. In the Nature paper from the International Schizophrenia Consortium (7), a CDCV model whereby schizophrenia risk was determined by thousands of genetic variants was developed. This model then replicated in three independent schizophrenia samples (P-values 2×10-28, 5×10-11, and 0.008), two independent bipolar disorder samples (P-values 1×10-12 and 7×10-9), and, crucially, did not replicate in six non-psychiatric biomedical diseases (P-values ≥ 0.05). Simulation studies conservatively estimated that this CDCV model accounted for ∼30% of the variance in liability to schizophrenia (7), a sizeable proportion of its heritability (9). Notably, additional simulations showed that the pattern of empirical results were inconsistent with a MRV model (7). Thus, this assertion by Mitchell and Porteus is again at odds with the data. Mitchell and Porteus argue that CDCV findings will be biologically uninformative. Their argument is based on analogy (“haven't we learnt more about disease mechanism and potential routes to the treatment of Alzheimer's disease from rare variant examples than from … APOE*e4?”). We suggest that this argument by analogy is not compelling. (a) To provide context, the APOE*e4-Alzheimer's disease association is atypically large. In the NHGRI GWAS catalog (accessed 1 July 2009) (10), there are 339 associations with P-values < 5×10-8 from 104 studies, 274 SNPs, and 43 complex biomedical diseases: only 15 of these 339 associations were as large or larger than APOE*e4-Alzheimer's disease. An analogy to an atypical result diminishes relevance. (b) Mitchell and Porteus ignore analogies that do not support their position. As a sort of positive control, GWAS for Type 2 diabetes mellitus have CDCV examples that are the molecular targets of blockbuster drugs (e.g., HMGCR and PPARG, the targets of statins and thiazolidinediones). Examples from the human genetics literature where common variants have refocused research are ignored. Age-related macular degeneration is an example of how genetic findings clarified causality: deposition of complement in retinal drusen was thought to be a secondary phenomenon but identification of genetic variation in complement factor H (CHF) as a risk factor for age-related macular degeneration has clarified its role as a primary feature and established a new conceptual paradigm. (c) Even if Mitchell and Porteus' point about APOE*e4-Alzheimer's disease were true, it is unlikely to be a law of nature as they suggest. Specifically, they argue that Occam's razor and statistical probability make it more likely that the inheritance of one or a few risk genes by any individual case may be the more likely explanation for the majority of incidence. However, Francis Crick argued against the overzealous application of Occam's razor in biology (11) as a true mechanistic explanation may superficially appear complicated if considered apart from its evolutionary context. Moreover, the study of model organisms such as Drosophila shows that most phenotypes are the result of complicated genetic architectures: multiple genes, often showing pleotropy (thus likely associated with multiple traits) and epistasis, and even single mutation effects differing with genetic background and environment (12, 13). Explanations relying on single genes are unlikely to capture the fundamental complexity of most human complex traits. Genome-wide studies, in combination with system biology approaches, yield comprehensive information and are thus more useful to deal with the inherent complexity of organismal phenotypes. Mitchell and Porteus correctly characterize the identification of rare copy number variants (CNV) as successes of the GWAS era. However, a crucial issue goes unmentioned: many of these rare CNVs of strong effect are non-specific and increase risk for autism, mood disorders, schizophrenia, epilepsy, and mental retardation. It is possible that future work will identify a general mechanism by which such CNVs predispose to neurological and psychiatric illness (such as the convergence of CNV effects on a few physiological pathways) but their general biological utility remains to be proven, and the causal explanations could well prove to be more complex than Mitchell and Porteous seem to anticipate. Mitchell and Porteus characterize the liability threshold model as “statistical sleight of hand”. Briefly, we note that this model underlies most approaches to estimating heritability (14, 15), that estimates stemming from the liability threshold model are widely accepted, and that the origins of the liability threshold model predate the modern genetic era (16-19) rather than being “latched onto” to explain the failure of genome-wide linkage studies. Indeed, the liability threshold model has a strong resemblance to the CDCV model postulated in the ISC paper (7), and future work will determine whether this resemblance is coincidental or fundamental. We agree with Mitchell and Porteus that multiple methodological approaches are needed to disentangle the complexities of the roles of genes and environment in the etiology of psychiatric disorders. GWAS studies are just one approach. However, note that Mitchell and Porteus omit discussion of a crucial point: investigation hundreds of multiplex pedigrees dense with bipolar disorder and schizophrenia using genome-wide linkage approaches have failed to identify subsets of families where MRV segregate. This negative empirical finding detracts from their exclusivist MRV argument. Finally, we would like to return to the central point of the PGC paper where we sketched a framework for how genomic studies of neuropsychiatric disorders might eventuate. It now seems clear that, at least for autism and schizophrenia, the genetic part of their etiology is due to both common and rare genetic variation. To quote our paper: “… debates about the success or failure of GWAS are based on subjective interpretations deriving from a small fraction of the eventual data or upon beliefs about the unknown fundamental architecture of psychiatric disorders. Such predictions are not based on systematic, quantitative, and scientific examination of the evidence and are strongly influenced by pre-existing biases and the degree of dispositional “optimism” or “pessimism” of the commentator. We suggest that debates about the success or failure of GWAS in psychiatry ask the wrong question. The critical question is: What can integrated mega-analyses of all available GWAS data tell us about the etiology of these psychiatric disorders? It is not a matter of success, failure, optimism, or pessimism but rather systematic, dispassionate, and logical approaches to empirical data: in short, by application of the scientific method. These data must be interpreted within a logical and systematic framework whereby both positive and appropriately qualified negative findings are informative.” The framework for the causation of psychiatric disease is built on an increasingly solid foundation. There is now compelling and relatively strong evidence that many of the cardinal neuropsychiatric disorders are caused by both common and rare genetic variants.