When neuro-imaging first became widely available, people were surprised to find remarkably little evidence of focal brain damage in children with specific developmental disorders. As C. M. Leonard (1997)
remarked: “It was thought that MRI was a potential diagnostic tool—that there might be structural landmarks for each developmental disability…. There is now overwhelming evidence that children with learning disabilities do not have ‘holes in the brain’. No subsequent studies have found a one-to-one correlation between behavioral symptoms and MRI or postmortem pathology in learning disabilities” (p. 161).
As is discussed below, it is now widely accepted that inherited brain anomalies, rather than acquired brain damage, are a key factor in the aetiology of developmental dyslexia, SLI, and autistic disorder. We understand very little about how genetic variations can lead to abnormal brain development, but it is conceivable that, by affecting processes such as neuronal migration, neurotransmitter activity, programmed cell death, synaptic connectivity, or myelination, they result in a brain that functions in a nonoptimal fashion.
The bulk of the evidence for genetic aetiology of developmental disorders comes from behavioural studies adopting a genetically informative design. Such studies consider how far observed similarities between individuals (i.e., similarities in their phenotypes) are correlated with their genetic similarity (similarities in their genotypes). At the simplest level, one can do family studies to see whether a disorder is more common in first-degree relatives (parents and siblings) of an affected child than in first-degree relatives of unaffected children. Such studies are not watertight, however, because relatives typically share environments as well as genes. A more satisfactory design is the twin study, which capitalizes on the fact that there are two different processes that can lead to twinning: Splitting of a fertilized ovum will lead to genetically identical or monozygotic (MZ) twins, whereas fertilization of two ova around the same time will lead to fraternal or dizygotic (DZ) twins, whose genetic similarity is equivalent to that seen in other siblings. Although media attention focuses on studies of twins reared apart, one does not need to use such rare cases to gain useful information from a twin study. Twins growing up together will resemble one another, insofar as they are subject to many of the same environmental influences, including prenatal as well as postnatal factors. However, if similarity between two members of a twin pair is greater for MZ than for DZ twins, this points to a role of genes. When considering developmental disorders, we can assess the degree of concordance between twins (i.e., the proportion of affected cases with an affected co-twin) in both MZ and DZ twins, to get estimates of the relative roles of genetic and environmental factors in the aetiology. Studies adopting this approach have found high estimates of heritability for autism (Rutter, 2005
) and SLI (Bishop, 2002
). A more mixed picture has been seen for reading disability, though most studies have found sizeable genetic effects (Pennington & Olson, 2005
With the advent of fast-throughput genetic sequencers, a sense of optimism was generated that we would quickly move on to identify risk genes for developmental disorders. However, progress has been surprisingly slow. Identifying genes that are implicated in disorder is not a single-step procedure. Typically, one first uses a method known as linkage analysis to home in on a region that is likely to harbour relevant genes. Linkage analysis relies on the fact that the closer together two bits of DNA are, the greater the likelihood they will be inherited together. Some stretches of DNA do not contain genes and are highly variable from one individual to another. These can act as “markers”, allowing one to identify whether a portion of DNA has been inherited from the mother or the father. One starts with a set of siblings, both of whom have a disorder. Then for a pair of alleles at a given locus, one can work out whether zero, one, or two alleles in siblings are “identical by descent” (IBD)—that is, inherited from the same parent (see ). The observed IBD pattern is compared with the IBD pattern predicted by chance to identify stretches of DNA that are coinherited by affected individuals at above-expected rates. These DNA regions are likely to be close to genes important for the disorder (see Bishop, 2002
, for a fuller account). Once linkage analysis has identified a marker associated with disorder, more detailed analysis of genes in that chromosomal region can be done, to look for allelic differences between affected and unaffected people. For autism, where twin studies give exceptionally strong heritability estimates, several linkages have been reported, but these have proved difficult to replicate (Barnby & Monaco, 2003
). More success has been found with dyslexia (Fisher & DeFries, 2002
) and SLI (Newbury, Bishop, & Monaco, 2005
), but it is becoming increasingly clear that we are not going to find the
gene for any of these disorders: They are heterogeneous, and such genes as exist are likely to act in a probabilistic fashion, in interaction with other genetic and environmental factors.
Figure 1 Schematic showing inheritance pattern for a small stretch of DNA. The grey region indicates an allelic variant associated with disorder. The region denoted by a, b, c, or d is a highly polymorphic noncoding region that can be used as a marker, because (more ...)
One reason why progress in the genetics of developmental disorders is slow may be because most genetic studies rely on textbook descriptions of disorders to define the phenotype. These typically specify in rather broad terms the impairments that have to be present to merit a diagnosis, together with exclusionary factors. So, for instance, SLI is diagnosed in a child whose language development lags significantly behind nonverbal development, and there is no obvious causal factor or other medical diagnosis such as hearing loss, neurological disease, or autistic disorder. This kind of diagnosis can be useful in identifying those who require special services, but it does not yield a homogenous group of children. On the one hand, it will include children with diverse difficulties in areas such as syntax, semantics, phonology, and pragmatics. On the other hand, it may exclude a child with significant language deficits because an arbitrary IQ cut-off is not met. Furthermore, although textbook definitions imply discrete disorders, the boundaries between conditions are often fuzzy—for example, SLI can be hard to distinguish from dyslexia (Bishop & Snowling, 2004
) and from autistic disorder (Bishop & Norbury, 2002
). There has been considerable interest in the idea that genetic discoveries may be more rapid if we use measures of underlying deficits to identify more coherent subtypes of disorder (Pennington, 1997
). To date, this approach has been most widely adopted with dyslexia, but progress has been somewhat erratic. Early suggestions that deficits in phonological and orthographic processing might be linked to different genes appear to have been false positives (Fisher & DeFries, 2002
); this kind of error is all too common in a field where one is conducting multiple statistical tests for linkage to many different loci. Nevertheless, there is some support for the idea that heritability may differ for different subtypes of dyslexia (Castles, Datta, Gayan, & Olson, 1999
). In the field of autism, family studies suggest that we may get clearer results if we distinguish subtypes of autism with a specific language profile (e.g., Shao et al., 2002
Identification of subtypes of disorder is one way forward, but it is not the complete solution. Rather than subdividing a disorder into ever more selective subgroups, we may sometimes need to do the opposite and consider adopting a broader conceptualization of the phenotype. The idea is not that we should simply lump together individuals with diverse clinical profiles, but rather that we should move away from a focus solely on clinical “disease” categories and develop instead more quantitative measures of underlying processes— so-called “endophenotypes” (Gottesman & Gould, 2003
The value of this approach was demonstrated to me when I conducted a twin study of SLI (Bishop, North, & Donlan, 1995
), in which we used conventional diagnostic criteria for SLI. Many twin pairs were categorized as discordant for SLI, but the “unaffected” twin often had clear evidence of language problems. In some cases, the mismatch between verbal and nonverbal skills was not great enough to merit a diagnosis of SLI, and in other cases, the child had a past history of SLI but did not show up as language-impaired on formal testing. This suggested that evidence for genetic influence on SLI would be stronger if we could use a measure of the endophenotype that captured the underlying deficits in such cases. Accordingly, we (Bishop, North, & Donlan, 1996
) gave a subset of this twin sample the children's nonword repetition test (CNRep: Gathercole, Willis, Baddeley, & Emslie, 1994
). The reason for choosing this measure was that Gathercole and Baddeley (1990)
had found that nonword repetition was strikingly deficient in SLI, with poor performance relating to the amount of material to be remembered, rather than discrimination or production of speech sounds. They proposed that nonword repetition was a sensitive test of the phonological loop component of working memory, a key process in normal language development that was implicated in vocabulary acquisition. Like Gathercole and Baddeley (1990)
, Bishop et al. (1996)
found major deficits in nonword repetition in children with SLI relative to control children, especially for longer nonwords. Furthermore, poor nonword repetition characterized many children who had a history of SLI, but who did not currently meet diagnostic criteria for this disorder. There were several MZ twin pairs who were concordant for poor nonword repetition, but discordant for a diagnosis of SLI. This suggested that nonword repetition might be a good behavioural marker for an underlying impairment, for which some children compensate.
To look at the heritability of poor nonword repetition, we used a method devised by DeFries and Fulker (1985)
, which was designed to estimate heritability of poor scores on a quantitative dimension. The overall logic of DeFries–Fulker analysis is parallel to more traditional twin analytic methods. Essentially, one aims to estimate the relative contribution of three potential factors that can serve to make twins more or less similar to one another. Environmental factors specific to the individual, known technically as nonshared environment or e2g
, will lead to twin dissimilarity. Nonshared environment includes measurement error as well as longer term idiosyncratic influences on the child (e.g., disease that affects just one twin). If performance on nonword repetition were determined solely by e2g
then regardless of Twin A's score, we would predict that Twin B would score at the population mean. Now consider the hypothetical situation where performance on nonword repetition is determined solely by environmental factors shared by both twins—for example, the amount of language input they received from parents. Environmental factors that are common to both twins, or c2g
, make twins similar to one another, so, if such factors exert a large effect, when Twin A has a low score, we predict that Twin B will also have a low score. The effect of c2g
will be the same, regardless of whether twins are MZ or DZ. Finally, consider the situation where genes, or heritable factors (h2g
), are the only important influence on poor nonword repetition performance. Because MZ twins are genetically identical, if MZ Twin A has a low score, we predict that MZ Twin B will have an equally low score. However, DZ twins have, on average, 50% of segregating genes in common.2
For a group of DZ twins who are selected as having low scores, the prediction is that the mean score of their co-twins will regress half way to the population mean. In practice, any trait that we observe is likely to be influenced by h2g
, and e2g
, and our goal is to estimate the relative contribution of each of these in explaining the observed variance. illustrates how we can do this using DeFries–Fulker analysis. First we need to select a set of “probands”—that is, people with poor scores on our measure of interest (e.g., nonword repetition). To test the significance of the genetic term, we can use multiple regression to consider how well we can predict the scores of co-twins from the scores of probands. If the prediction is improved by including in the regression equation a term that represents the degree of genetic relationship between the twins (.5 for DZ and 1.0 for MZ), then this indicates that genes play a role in determining impairment. Bishop et al. (1996)
identified probands on the basis of poor scores on CNRep, regardless of whether they met diagnostic criteria for SLI. The estimate of heritability (h2g
) was close to 1.0, suggesting that genes play a major role in causing deficient phonological short-term memory (STM). These striking results from behavioural analysis have led to the inclusion of nonword repetition as a phenotypic index in molecular genetic studies. Strong linkage has been found to a site on chromosome 16 (SLI Consortium, 2002
), and candidate genes in this region are now being tested for association to disorder.
Figure 2 Illustration of DeFries–Fulker analysis. Data are transformed so that the population mean = 0 and the proband mean = 1. The effect of nonshared environment on impairment (e2g) is estimated as MZ proband mean − MZ co-twin mean. The effect (more ...)
The second example in this section concerns the rather discrepant findings concerning genetic influences on reading disability. Many twin studies have reported moderate to high estimates of heritability of reading disability, but there have been some notable exceptions, where environmental factors shared between twins have emerged as more important. One such case is a study by Bishop (2001b)
, who tested a general population sample of twin pairs and identified probands for DeFries–Fulker analysis by selecting those that scored more than 1 standard deviation below the mean on a test of nonword reading. Correlations between probands and their co-twins were high for both MZ and DZ twins, pointing to a major effect of shared environment (c2g
). Estimates of genetic influence (h2g
) were negligible. This was a surprising result, in the light of much higher heritability estimates for dyslexia obtained by other researchers (see Pennington & Olson, 2005
, for review). However, because a previous study with an SLI group had shown close overlap between nonword repetition and literacy skills, Bishop ran an augmented version of the DeFries–Fulker analysis in which nonword repetition was included in the regression equation. This gave a significant interaction, and it was evident that for children with impaired nonword repetition, reading disability was heritable, whereas for those with normal nonword repetition skills, it was largely environmentally determined. A similar pattern of results was obtained with a younger sample of 6-year-old twins who were at the first stages of learning to read (Bishop, Adams, & Norbury, 2004a
). High heritability estimates for literacy scores were found for children with poor nonword repetition, but not for those with average nonword repetition. This work has relevance for molecular genetic studies of dyslexia, because it implies that clearer genetic results will be seen if we focus attention on those poor readers with weak phonological short-term memory. Reliance on reading tests alone to identify probands will mean that we may include in our sample a substantial subset of children whose poor reading is strongly influenced by environmental risks.
The studies reviewed so far suggest that phonological short-term memory is an important component skill for language and literacy acquisition and has a strong genetic basis. Can we extend the argument further and use it as a phenotypic index in other disorders? Tager-Flusberg and Joseph (2003)
noted that nonword repetition, together with certain other language skills, is frequently, though not invariably, impaired in autistic disorder, and they suggested that the same genetic factors that led to SLI in one child might lead to autism in another. A study by Bishop et al. (2004b)
, however, gave a different picture. In this study, probands with autism and their first-degree relatives were given a test of nonword repetition. Many of the children with autism did poorly on this test, as predicted by Tager-Flusberg and Joseph (2003)
. However, scores of their first-degree relatives (i.e., parents and siblings, who share 50% of segregating genes) were unimpaired, indicating that in this population, the deficit was not heritable. This contrasts with the picture in SLI and dyslexia, where relatives of affected individuals do tend to have lower scores than control populations on nonword repetition (Bishop et al., 1996
; Raskind, Hsu, Berninger, Thomson, & Wijsman, 2000
). It would be of interest to compare the nature of errors made on nonword repetition by children with autism with that for children with SLI, as this might point to a different underlying cause of poor performance in the two groups and so help us devise a cleaner measure of an SLI endophenotype.
In concluding this section, it is important to sound a note of caution. We hope that by improving our measures of the phenotype, we may gain new insights into the genotype, but even if we restrict attention to the much simpler case of single-gene disorders, it is clear that the relationship between genotype and phenotype is not always straightforward. The impact of a gene can be influenced by the environment to which the organism is exposed, the genetic background against which it is expressed, and random stochastic processes (U. Wolf, 1997
). Environmental modulation of genetic effects is illustrated by the well-known example of phenylketonuria, where a genetic variant that usually has a detrimental effect on brain development can lead to a milder phenotype if rigorous dietary restraint is adopted (Smith, Beasley, & Ades, 1991
). Another example is Huntington's disease, which, as Spires and Hannan (2005)
put it, is often regarded as the “epitome of genetic determinism”: a notorious case where a single dominant allele causes a late-onset progressive neurological degeneration. This has been modelled in mice, where it has been found that the impact of the mutation can be modified by physical exercise early in life. In neurofibromatosis Type I, one can see affected relatives with the same mutation but vastly different cognitive sequelae, ranging from major mental impairment through to no detectable symptoms (Reiss & Denckla, 1996
)—with phenotypic variation probably influenced by interactions between different genes (U. Wolf, 1997
). And moving back to the disorders that are the focus of this paper, studies of twins make it clear that the way in which genetic risk for autism is manifest can be highly variable, even within a genetically identical MZ twin pair (Le Couteur et al., 1996
). This could be due to different environmental influences on the two twins in a pair, but it could also be the result of purely stochastic influences on morphogenesis (U. Wolf, 1997
In sum, I have given some examples from my own work showing how theoretically driven measures of underlying cognitive processes can be incorporated in a genetically informative design to help in the quest for risk genes for disorder. A major research agenda for neuropsychology is to derive better measures of such endophenotypes, rather than relying on surface manifestations of developmental disorders. It is hoped that by using quantitative measures of underlying cognitive processes rather than clinical syndromes we may find clearer relationships between phenotype and genotype. However, we must be aware that the cognitive phenotype is likely to be influenced by complex interactions between genes and environments, rather than deterministically by single genes.