Current research in the genetics of schizophrenia is guided primarily by the ‘common disease – common alleles' model (Reference ChakravartiChakravarti, 1999). This model originated from the hypothesis that the illness results from the cumulative impact of multiple common small-effect, genetic variants, interacting with environmental exposures to exceed a biological threshold (Reference Gottesman and ShieldsGottesman & Shields, 1982). The ‘common disease – common alleles' model for schizophrenia is heuristically appealing. The illness is relatively frequent and is found worldwide. Thus common susceptibility alleles shared across populations are plausible. The ‘common disease – common alleles' model has also been posited to explain the variable and inconsistent results of linkage studies devoted to finding genes of large effect responsible for schizophrenia, and the weak associations of various candidate genes with schizophrenia. Furthermore, mathematical modelling suggested that the observed decline in recurrence risk of disease with increased genetic distance from affected individuals is inconsistent with monogenic inheritance of large-effect alleles (Reference RischRisch, 1990). As pointed out by Risch, these models were based on the assumption that the illness was genetically homogeneous throughout the population for which the recurrence risks were calculated (Reference RischRisch, 1990). Increasing evidence suggests that schizophrenia is genetically heterogeneous (Reference Fanous and KendlerFanous & Kendler, 2005). If so, then recurrence risk data are also consistent with monogenic inheritance of large-effect alleles in a proportion of people with schizophrenia, with different alleles for different families.

We suggest that the ‘common disease – rare alleles' model explains many cases of schizophrenia. Our hypothesis is that many mutations predisposing to schizophrenia are highly penetrant and individually rare, even specific to single patients or families. In this model, different families harbour different mutations, either in the same gene or in different genes, but any one family carries only one or two mutations. Many different disease-associated mutations may occur in the same gene.

The ‘common disease – common alleles' and ‘common disease – rare alleles' models are not mutually exclusive (Reference Goldstein and ChikhiGoldstein & Chikhi, 2002). Rare severe mutations may occur in genes that also harbour more common variants with modest effects on disease risk. However, the two models have distinctly different implications for gene-finding strategies. Most current psychiatric genetic research is designed to identify common alleles or haplotypes associated with with increased risk of disease and shared by large numbers of patients compared with appropriate controls (Reference Merikangas and RischMerikangas & Risch, 2003). If many cases of schizophrenia stem from individually rare large-effect alleles, current approaches – even if executed perfectly – will fail to identify critical genes.

We argue that current observations from epidemiology and genetics of schizophrenia are consistent with the influence of a large number of individually rare deleterious mutations, many of which have occurred in the present or recent generations. Several features of schizophrenia support this view:

1. (a) Schizophrenia is familial, i.e. close relatives of affected persons are at increased risk of the illness. Occasional families are very severely affected (Reference Gottesman and ShieldsGottesman & Shields, 1982), but most patients have no close affected relative. Taken together, these observations are consistent with a subset of families harbouring high-penetrance, recently occurring alleles predisposing to schizophrenia, with different alleles present in different families. Unless caused by detectable chromosomal alterations (as in the case of DISC1), such alleles have heretofore been difficult to find because individual families are not sufficiently informative for single-family linkage analyses.

2. (b) Paternal age is consistently associated with increased risk of schizophrenia (Reference Brown, Schaefer and WyattBrown et al, 2002; Reference Dalman and AllebeckDalman & Allebeck, 2002; Reference Malaspina, Corcoran and FahimMalaspina et al, 2002; Reference Byrne, Agerbo and EwaldByrne et al, 2003; Reference El-Saadi, Pedersen and McNeilEl-Saadi et al, 2004; Reference Sipos, Rasmussen and HarrisonSipos et al, 2004; Reference Tsuchiya, Takagai and KawaiTsuchiya et al, 2005). Paternal age is also associated with increased rates of several types of de novo germ-line mutations (Reference CrowCrow, 2003).

3. (c) The illness is associated with decreased fertility (Reference NimgaonkarNimgaonkar, 1998; Reference Haukka, Suvisaari and LonnqvistHaukka et al, 2003). If this had been the case over long periods, then the frequencies of any ancient common alleles associated with schizophrenia would be reduced. An ongoing contribution of new and therefore individually rare risk alleles could explain the persistence of the disorder.

To explain our reasoning, we first describe recent findings in genomics and genetics of other complex human disorders. Then we show that the results of schizophrenia research are consistent with the existence of multiple individually rare alleles of large effect. Finally, we consider the implications of this model for future schizophrenia research.

THE HUMAN GENOME

The dynamics of the human genome are proving more complex than anticipated, revealing more mechanisms by which genetic changes may lead to human disease. Particularly relevant to our argument are the following observations (International Human Genome Sequencing Consortium, 2001a , b , 2004).

1. (a) Only about 2% of the genome consists of protein-coding genes. There are approximately 20 000 protein-coding genes, far fewer than the 80 000–100 000 hypothesised a decade ago. However, the human proteome is enormously complex. At most genes, variable transcription leads to multiple transcripts, and thus multiple proteins, derived from the same locus but with different amino acid sequences. Variable transcription is frequently tissue-specific. As a result, the consequences of a mutation may also be tissue-specific.

2. (b) Germ-line mutations occur more commonly than previously thought. Potentially deleterious new mutations may occur at a rate as high as three per zygote (Reference Eyre-Walker and KeightleyEyre-Walker & Keightley, 1999; Reference CrowCrow, 2000). Rates of occurrence of different classes of de novo mutations are differently influenced by parent of origin and parental age (Crow, Reference Crow2000, Reference Crow2003). The increased mutation rate associated with greater paternal age is particularly relevant, given that risk of schizophrenia is also associated with paternal age.

3. (c) Epigenetic alterations – stable changes in gene expression that do not depend on changes in DNA sequence (Reference Jaenisch and BirdJaenisch & Bird, 2003) –may play an important part in human disease, including psychiatric disorders. Recent intriguing observations of possible epigenetic effects related to development include phenotypic variability in monozygotic twins and imprinting effects on neurodevelopmental disorders (Reference Fraga, Ballestar and PazFraga (et al, 2005; Reference Wong, Gottesman and PetronisWong et al, 2005).

GENETIC HETEROGENEITY IN HUMAN DISEASE

The genetic heterogeneity of complex illnesses is the natural result of the origins of human genetic variation. The oldest human alleles originated in Africa millions of years before people first migrated out of Africa some 50 000 years ago (Reference Cavalli-Sforza, Menozzi and PiazzaCavalli-Sforza et al, 1994). These ancient variants are found in all human populations, are the most common worldwide, and account for approximately 95% of human variation. Yet the exponential growth of the human population has resulted in many new alleles, each individually rare and each specific to one population (or even one family). Most alleles are of this sort. Thus the paradox: most human variation is ancient and shared; most alleles are recent and individually rare. Given the size of the present human population and the rate of occurrence of new mutations, all mutations compatible with life have probably already occurred and will occur again. However, mutations with deleterious effects before or during the reproductive years will be less frequently transmitted to subsequent generations owing to their adverse impact on fertility or viability. Therefore, mutations with large effects on disease may be disproportionately of recent origin and individually very rare.

To the extent that any class of mutation – point mutations, copy number errors or abnormalities of chromosome number – occur spontaneously, they appear at similar rates in all human populations. All humans share the same basic genomic architecture, including the same genomic regions vulnerable to mutations. The incidence of schizophrenia does not appear to vary substantially across populations by virtue of genetic ancestry. This pattern is consistent with a disease due to multiple, independent de novo mutations that arise in many different vulnerable genes and genomic regions. Furthermore, environmental exposures with mutagenic consequences may lead to high rates of new mutations among exposed individuals. For example, environmental factors such as maternal starvation that are associated with disease (Reference Cannon, Kendell, Susser, Murray, Jones and SusserCannon (et al, 2003) may mediate their effects through de novo genetic or epigenetic mutations. We explore this theme in more detail below.

Complex illnesses are almost universally characterised by allelic heterogeneity (multiple different mutations in the same gene leading to disease) and locus heterogeneity (mutations in multiple different genes leading to the same disease) (Reference Botstein and RischBotstein & Risch, 2003; Reference Goldstein, Cavalleri and AhmadiGoldstein et al, 2003). We propose that both are characteristic of schizophrenia. To understand the potential implications of genetic heterogeneity for schizophrenia, we briefly consider other complex disorders for which genes have been identified.

IMPLICATIONS OF GENETIC HETEROGENEITY FOR GENE DISCOVERY

All complex illnesses evaluated thus far are characterised by locus and allelic heterogeneity. Disease genes for these illnesses have been identified primarily by positional cloning in large kindreds. Although subsequent association studies confirmed their role, the original gene discoveries were dependent upon individual highly informative families. Such large informative kindreds are extremely rare in schizophrenia. Linkage studies of schizophrenia based on single gene models have not been successful at identifying causal mutations (Reference Owen, Williams and O'DonovanOwen et al, 2004). Because pedigrees with schizophrenia have not been large enough to be individually informative, studies generally pool data from different families. If many different genes were responsible for the illness in different families, pooling results across families would preclude identification of any of them.

Currently, most gene-discovery strategies for schizophrenia research – case–parent triad studies, candidate gene studies and haplotype association studies – are designed to identify alleles or haplotypes that appear more frequently among affected individuals than among appropriate controls (Reference Cannon, Kendell, Susser, Murray, Jones and SusserCannon et al, 2003). These designs are not robust to either allelic heterogeneity or locus heterogeneity. Sib-pair linkage analyses are designed to detect genomic regions consistently shared by affected siblings. Sib-pair analyses are robust to allelic heterogeneity but not to locus heterogeneity. For each of these designs, hundreds or thousands of rigorously diagnosed cases of unrelated affected and unaffected individuals are examined. If a substantial portion of schizophrenia stems from different individually rare alleles, increasing the number of cases also increases the number of different disease risk mutations represented among them. As a result, increasing sample size does not confer a corresponding increase in statistical power. In the most extreme scenario, in which every case results from a different mutation, an increase in sample size would not lead to any increase in the power to detect any one mutation. Consequently, even very large studies may fail to detect individually rare disease risk mutations.

In addition, genetic analyses that focus only on single nucleotide polymorphisms (SNPs), either individually or in haplotypes, rather than fully sequenced DNA, will inevitably miss rare disease alleles and thus fail to detect critical genes harbouring such alleles. Association and linkage studies generally assume that individuals sharing the same SNP-defined haplotype share the entire region, including any hypothetical embedded disease alleles. This assumption is reasonable for ancient alleles and nearly always true for related individuals for whom there is direct inheritance of the haplotype. However, this assumption is not reasonable for a study of unrelated individuals who carry disease alleles of recent origin. Rare recent mutations causing schizophrenia within the same haplotype will differ among unrelated individuals, diluting any association.

TWO CANDIDATE GENES FOR SCHIZOPHRENIA: DYSBINDIN AND DISC1

The familial nature of schizophrenia is well established. Large collaborative linkage studies have suggested multiple candidate chromosomal regions that may harbour genes associated with the illness. Regions best supported by genome-wide scans include 6p22–p24 (Reference Straub, Maclean and O'NeillStraub et al, 1995), 1q21–q22 (Reference Brustowicz, Hodgkinson and ChowBrustowicz (et al, 2000) and 13q32–q34 (Reference Blouin, Dombroski and NathBlouin (et al, 1998). Other regions with positive linkage findings include 1q42, 5q21–q33, 6q21–q25, 8p21–p22, 10p15–p11 and 22q11–q12 (Reference Owen, Williams and O'DonovanOwen (et al, 2004). These regions combined represent a substantial portion of the genome.

Candidate genes have been suggested in several of these regions, including dysbindin on 6p22, neuregulin on 8p22, G72 on 13q34, COMT on 22q11, RGS4 on 1q21 and GRM3 on 7q21 (see reviews by Reference Blouin, Dombroski and NathBlouin et al, 1998; Reference Harrison and WeinbergerHarrison & Weinberger, 2005). Each of these genes is biologically plausible (Reference Owen, Williams and O'DonovanOwen et al, 2004). However, for each candidate gene, both positive and negative associations have been reported with the same SNPs; strengths of effects are generally weak; the specific allele or haplotype associated with the illness varies across studies; and definitive causative mutations have not been identified.

To illustrate the implications of the common allele v. rare allele models, we will review two promising susceptibility genes, dysbindin (DTNBP1) and DISC1. Very different study designs revealed these genes, with correspondingly different results to date. Linkage, association and functional studies all support some role for dysbindin in schizophrenia. Several studies involving different populations have found positive associations of dysbindin alleles or haplotypes with schizophrenia (Reference Straub, Jiang and MacleanStraub et al, 2002; Reference Schwab, Knapp and MondabonSchwab et al, 2003; Reference Funke, Finn and PlocikFunke et al, 2004; Reference Kirov, Ivanov and WilliamsKirov et al, 2004; Reference Kohn, Danilovich and FilonKohn et al, 2004; Reference Numakawa, Yagasaki and IshimotoNumakawa et al, 2004; Reference Williams, Preece and MorrisWilliams et al, 2004; Reference Bray, Preece and WilliamsBray et al, 2005; Reference Gornick, Addington and SpornGornick et al, 2005). Dysbindin is widely expressed in brain, and appears to play a part in cognitive functioning and capacity (Reference Owen, Williams and O'DonovanOwen et al, 2004). Post-mortem studies suggest that brain levels of dysbindin may be reduced in individuals with schizophrenia (Reference Weickert, Straub and McClintockWeickert et al, 2004). However, no variant of dysbindin has been specifically linked to schizophrenia. Across different studies, the risk conferred by any dysbindin variant is small, with effect sizes of about 2.0 or less. Among positive association studies, the specific alleles associated with the disease differ. Moreover, an allele may be associated with increased disease risk in some studies and decreased risk in others (Reference Owen, Williams and O'DonovanOwen et al, 2004). In general, the variants of interest (defined by SNPs) are common and without known functional significance. An exception is SNP rs1047631, which has been associated with differences in the expression of dysbindin in brain (Reference Funke, Finn and PlocikFunke et al, 2004). However, the frequency of the haplotype with this SNP was similar between cases (45.6%) and controls (40.4%). Thus far, resequencing efforts have not revealed any coding sequence mutations in dysbindin among individuals with schizophrenia (Reference Liao and ChenLiao & Chen, 2004).

Therefore, at present the evidence supporting dysbindin is mixed. Variable associations with different alleles have been attributed to allelic heterogeneity; yet allelic heterogeneity refers to different disease-causing mutations in the same gene, not to the same allele reducing risk in some cases and increasing risk in others. There are at least three possible interpretations of these data. The most favoured in the literature is that dysbindin variants mediate disease risk as part of a complex interaction with other genes and environmental factors. This is possible, in principle, although difficult to test given the challenge of establishing the role of a mediating factor of small effect on a complex illness of unknown cause.

A second possibility is that relatively rare, as yet unidentified variants in the dysbindin locus are embedded in illness-associated haplotypes in some, but not all, cases (including those potentially located in non-coding regulatory regions). Such alleles could have substantial effects on the phenotype, but would be masked by studying only the common haplotype. The third possibility is that many, most, or all of the various positive associations with dysbindin are false positives. The number of different positive association studies (albeit with different variants) is taken as prima facie evidence that the gene must be involved with the disorder. However, dysbindin, like many genes involved with brain development, is large (> 140 kb). The dysbindin locus includes at least 363 SNPs (Reference Hinrichs, Karolchik and BaertschHinrichs et al, 2006), from which various candidates are selected for association studies. Incorporating linkage disequilibrium across the locus, many thousands of SNP and haplotype combinations appear in different populations. The potential for false positives is enormous. Unless negative and positive studies were published with equal frequency, this possibility is also nearly impossible to test.

In contrast, DISC1 and its associated non-coding antisense RNA DISC2 were originally identified by a balanced translocation involving chromosome 1q42 which segregated with schizophrenia (and other major psychiatric disorders) over four generations in a large Scottish kindred (Reference St Clair, Blackwood and MuirSt Clair et al, 1990; Millar et al, Reference Millar, Brown and Maule1998, Reference Millar, Wilson-Annan and Anderson2000). Sachs et al (Reference Sachs, Sawa and Holmes2005) found that a frameshift mutation that abnormally truncates DISC1 co-segregated with schizophrenia in three siblings. However, this mutation is also rarely found in controls with unknown diagnostic status (Reference Green, Norton and PeirceGreen et al, 2006). The gene PDE4B (phosphodiesterase 4B), which interacts with DISC1 in the neuronal cyclic adenosine monophosphase (AMP) pathway, was disrupted by a balanced translocation in two related individuals with chronic psychotic illnesses (Reference Millar, Pickard and MackieMillar et al, 2005). Finally, mice with a deletion variant that disrupts the DISC1 protein have working memory deficits (Reference Koike, Arguello and KvajoKoike et al, 2006). These findings suggest a role of rare large-effect mutations of DISC1 and of genes involved in DISC1 pathways in the development of schizophrenia.

Not surprisingly, DISC1 became the subject of association studies. Haplotypes of DISC1 were associated with schizophrenia and other mental illness in European and North American populations (Reference Ekelund, Hovatta and ParkerEkelund et al, 2001; Reference Hennah, Varilo and KestilaHennah et al, 2003; Reference Hodgkinson, Goldman and JaegerHodgkinson et al, 2004; Reference Callicott, Straub and PezawasCallicott et al, 2005; Reference Cannon, Hennah and van ErpCannon et al, 2005; Reference Hamshere, Bennett and WilliamsHamshere et al, 2005) but not in Japanese or Scottish populations (Reference Devon, Anderson and TeagueDevon et al, 2001; Reference Kockelkorn, Arai and MatsumotoKockelkorn et al, 2004; Reference Zhang, Tochigi and OhashiZhang et al, 2005). Within populations with positive associations, DISC1 haplotypes were also associated with putative endophenotypes, including neuroanatomical and/or neurocognitive profiles (Reference Hodgkinson, Goldman and JaegerHodgkinson et al, 2004; Reference Burdick, Hodgkinson and SzeszkoBurdick et al, 2005; Reference Cannon, Hennah and van ErpCannon et al, 2005). However, disease-risk haplotypes vary across populations and effect sizes are small. Therefore, although there is compelling evidence that rare large-effect mutations in DISC1 are associated with schizophrenia, it is not clear whether common polymorphisms play a part.

IMPLICATIONS FOR GENETIC RESEARCH IN SCHIZOPHRENIA

The ‘common disease – rare allele' model has important implications for gene-finding strategies. A current mantra in schizophrenia genetic research is the need for ever-larger sample sizes in order to obtain adequate statistical power to detect common small-effect variants (Reference Devon, Anderson and TeagueDevon et al, 2001). These designs are dependent upon the existence of disease-risk alleles that are shared across large numbers of unrelated cases. These strategies will be inadequate if schizophrenia in large part stems from individually rare disease-risk mutations in a large number of different genes.

We propose an alternative strategy in selecting cases for study. Rare cases of schizophrenia with mutations that can be individually detected using current genomic technologies are extremely valuable. It is worthwhile devoting resources to finding them. It has been recognised for decades that such cases would include any large kindred with a number of well-diagnosed individuals or cases with identifiable genomic events of recent origin (e.g. balanced translocations). The emerging story of DISC1 highlights this strategy, since the gene was originally identified by a balanced translocation on chromosome 1q42 coinherited with schizophrenia (Reference Millar, Brown and MauleMillar (et al, 1998). Similar promising findings have been noted for specific genes in the 22q11 region (Reference Maynard, Haskell and LiebermanMaynard et al, 2002), stemming in part from the recognised association between deletions at 22q11 (i.e. the velo-cardio-facial syndrome) and schizophrenia.

Current genomic technology now enables the identification of an increasingly large number of classes of mutations. Heretofore, identifiable genomic events have been limited to chromosomal abnormalities such as translocations or deletions. However, as the resolution of genome-wide mutation screening technologies improves, smaller genomic events in informative cases or families can be detected. Once identified, a gene altered by a single chromosomal event becomes a candidate to be screened for other (typically smaller) mutations in other cases. Advanced resequencing technology allows for more rapid identification of mutations in candidate genes. The occurrence of multiple deleterious mutations among unrelated cases provides both biological evidence and epidemiological support for the causal role of the gene, using gene-based hypothesis testing strategies (Reference Chen, Sakoda and HsingChen et al, 2006).

Individuals who develop schizophrenia following a known environmental exposure are also potentially informative. Such exposures may focus gene discovery in two ways. First, genomic approaches (e.g. resequencing efforts) can focus on candidate genetic pathways relevant to the suspected exposure, screening for otherwise benign variants that are deleterious given the exposure. For example, associations between schizophrenia and in utero exposure to maternal starvation (Reference Susser, Neugebauer and HoekSusser (et al, 1996; Reference St Clair, Xu and WangSt Clair et al, 2005), and associations between schizophrenia and genes in the folate metabolic pathway (Reference Lewis, Zammit and GunnellLewis et al, 2005; Reference Picker and CoylePicker & Coyle, 2005) suggest that mutations in genes in the folate metabolism network could be linked to the illness. Second, the mutagenic effects of the environmental exposure can be evaluated. Following the same example, gestational folate deficiency may be mutagenic, in that it leads to an increase in the rate of mutation genome-wide (Reference McClellan, Susser and KingMcClellan et al, 2006). Among such cases, genome-wide mutational screening may detect de novo mutations. If disease-associated alleles are identified, other mutations within the same gene may confer some risk for other individuals with the same exposure. Furthermore, severe mutations in the same gene may lead to the disorder without the exposure.

As informative cases are identified, genomic technologies are needed for efficient screening for potential disease-associated mutations. Effective transcriptome and proteome-based tools are needed to characterise such variants. These technologies are under rapid development. It is already possible to detect deletions, duplications and other chromosomal aberrations of multiple kilobases anywhere in the genome, and the sensitivity of these methods for detecting smaller mutations is improving (Reference Sebat, Lakshmi and TrogeSebat et al, 2004; Reference Sharp, Locke and McGrathSharp et al, 2005). Resequencing tools are increasingly efficient, so large numbers can be screened for rare events in candidate genes first identified by rare, individually detectable events. User-friendly bioinformatics resources now exist to help characterise the structure and function of potential candidate genes. It is increasingly possible to characterise not only mutations in protein coding sequence but also mutations and epigenetic changes in regulatory regions, in non-coding RNA and in transposable elements.

To summarise, we propose that individually rare alleles with large effect, many of which are recent in origin, have a substantial role in causing schizophrenia. Current research designs that focus on collecting large samples of unrelated individuals for analysis of shared alleles or haplotypes are not suitable for detecting such disease alleles. In contrast, rare disease mutations may be revealed by studies of individuals and families that harbour informative genomic events, and by studies of exposed cohorts. A gene harbouring one mutation predisposing to schizophrenia is likely to harbour more than one, with frequencies ranging from relatively common to rare, and effects ranging from modest to severe. To see one is not to see them all.