Association Between a High-Risk Autism Locus on 5p14 and Social Communication Spectrum Phenotypes in the General Population

ALSPAC is a population-based prospective birth cohort for which there has been extensive data collection on the health and development of children and their parents (20). All pregnant women in the Bristol, England, area with an expected delivery between April 1991 and December 1992 were approached for participation in the study. A total of 14,541 women enrolled in the study, and 13,988 children were alive at 1 year. Ethical approval for the study was obtained from the ALSPAC Law and Ethics Committee and the Local Research Ethics Committees.

Of 9,650 offspring for whom DNA was available in 2009, 9,100 were successfully genotyped, which included 7,862 singletons. Of those, information was available for 7,738 on ethnicity and maternal education (assessed during the antenatal period: below O-level [9.23%], O-level [52.39%], and above O-level [38.38%]; O-levels are U.K. school-leaving qualifications taken at age 16). A total of 7,313 genotyped ethnically homogeneous white European singletons (52.07% of them males) were selected for this study.

The ALSPAC cohort contains a small proportion of children with autism spectrum disorder who were identified from either National Health Service community pediatric records or Education Service databases for the region (21). Eighty-six children with autism spectrum disorder were identified by age 11, giving a prevalence of 62 per 10,000 children. rs4307059 genotype information was available for 41 unrelated white European children with autism spectrum disorder.

DNA was extracted as previously described (22). The genotyping of the DNA was performed by KBioscience (www.kbioscience.co.uk) using a competitive allele-specific polymerase chain reaction system (KASPar). A genotyping call rate of 94.2% was achieved, and all genotypes were distributed in adherence to Hardy-Weinberg equilibrium (exact p=0.56; N_CC=1,033 [14.13%], N_CT=3,463 [47.35%], N_TT=2,817 [38.52%]). rs4307059 alleles were coded with respect to the major T allele (frequency=62.42%) conferring risk for autism spectrum disorder (8).

Broader autism phenotypes that may be related to a diagnosis of autism spectrum disorder used by Wang and colleagues (8) (see part 1 of the data supplement that accompanies the online edition of this article) were chosen with respect to early language, (social) communication, verbal intelligence, social interaction, behavioral difficulties, and special educational needs. Instruments were selected to cover most comprehensively continuous traits between ages 3 and 12 for the majority of children in ALSPAC. By age 3, individuals with autism exhibit some abnormality in at least one of the three key areas (1, 23). For each selected instrument we studied all available repeated measurements within the specified time frame. None of these measures captured specifically the repetitive interests, behaviors, and activities spectrum (see part 2 of the online data supplement).

Early language patterns were screened because a delay in the acquisition of early language is the most common initial symptom of autism spectrum disorder recognized by parents (24). Mother-reported early childhood language and communication patterns were assessed with the MacArthur Toddler Communication Questionnaire (25) summary language score (at a mean age of 3.25 years).

Social communication skills were assessed with the mother-reported Social and Communication Disorders Checklist (SCDC) (19) (at mean ages of 7.7 and 10.8 years) and communicative impairments with the mother-reported Children's Communication Checklist (CCC) (26) (at a mean age of 9.7 years). The SCDC is a brief screening instrument of social reciprocity and verbal/ nonverbal communication (19), and the CCC captures important aspects of communication impairment of the broader autism phenotype in children (15).

Given the strong evidence for association between autism and cognitive functioning (3), verbal IQ was measured with the WISCIII (27) (at a mean age of 8.6 years). A short version of the test consisting of alternate items only (except the coding task) was applied by trained psychologists during the year 8 focus clinic examination.

Mother-reported social interaction was assessed using the sociability subscale of the Emotionality-Activity-Sociability Temperament Survey (EAS) (28) (at mean ages of 3.3, 4.8, and 5.8 years) and the peer problem subscale of the Strengths and Difficulties Questionnaire (SDQ) (29) (at mean ages of 4.0, 6.8, 8.2, 9.7, and 11.8 years). There is some evidence that EAS sociability scores predict the quality of interactions in preschool children and their ability to handle conflict (30), and most of the variation in social functioning, as assessed by more specialized instruments, such as the Autism Spectrum Screening Questionnaire (31), is likely to be reflected by the SDQ peer problem subscale. Child-reported friendship scores (at mean ages of 8.6 and 10.7 years) were derived from the Cambridge Hormones and Moods Project Friendship Questionnaire (HMP-FQ) (32) to rate children's happiness with their friendships.

Behavioral difficulties were examined because social communicative deficits might be of prognostic significance with respect to behavioral adjustment at school (18). These were captured using mother-reported total behavioral difficulties as measured by the Revised Rutter Parent Scale for Preschool Children (R-RPS-PC) (33) during early life (at a mean age of 3.5 years), and later on by the SDQ (at mean ages of 4.0, 6.8, 8.2, 9.7, and 11.8 years). Children with special educational needs were identified using the Pupil Level Annual School Census (PLASC) for the 2003-2004 academic year (at a mean age of 11.8 years) supplied by the Department of Education for England, which has previously been used to identify autism cases in the ALSPAC sample (21). For each child attending a state school, the PLASC data provide information on the level of extra help being provided at school (“no special provision,” “school action,” “school action plus,” or “statement of special educational needs”; see part 3 of the online data supplement).

In total, 29 mostly continuous measures were subjected to genetic association analysis. For each phenotype, we determined whether it coded for high functionality (where higher scores indicate fewer problems or more functionality) or low functionality (where higher scores indicate more problems, less functionality, or fewer friends). Characteristics of the sample are summarized in Table 1, and trait intercorrelations are presented in Table S1 in the online data supplement.

Like other large longitudinal cohorts, ALSPAC has attrition over time. Although this has been linked to sociodemographic factors, the distribution of genotypes is generally unrelated to socioeconomic status (34). In line with this notion, there was no evidence that rs4307059 genotypes, examined for their association with social communication spectrum traits, were predictors of trait missingness (based on a logistic regression approach using all available 9,100 genotypes to predict missingness for the 29 selected phenotypes; minimum p value, 0.26; data not shown).

The association between rs4307059 and each of the selected phenotypes was investigated with single regression models. Normally distributed traits were analyzed with ordinary least squares regression assuming an additive genetic model. Skewed continuous traits were modeled as count data, and the error distribution was specified as quasi-Poisson to accommodate for over- or underdispersion of the data. Categorical outcomes were analyzed using multinomial regression. The genetic effect for the two latter models was assumed to be log-additive, which is consistent with previous research (8).

In all analyses, genetic effects were adjusted for maternal education, sex, and age at phenotypic assessment. Although unlikely to be a confounder, maternal education as a covariate may considerably reduce the amount of unexplained variance, especially with respect to language-related phenotypes (e.g., maternal education explained ≥6.5% variance [adjusted R²] in verbal IQ; data not shown). The statistical significance of genetic associations was assessed using F tests or likelihood ratio tests.

As our prior hypothesis assumed an underlying autism spectrum disorder QTL, implicating an association between decreased intellectual, social, and communicative functionality (and likewise more behavioral and social problems) and a higher load of the rs4307059 risk allele, the analysis was performed in one-sided or directed mode. The expected direction of the genetic effect is outlined for each phenotype in Table 2.

To account for trait interrelatedness, multiple testing, and random genotype dropout, all single-trait associations were adjusted using permutations. A Bonferroni correction would have been less suitable because it cannot account for phenotypic correlation between traits. Genotypes from all genotyped individuals were included in the permutation vector, irrespective of whether they had been phenotyped or not, in order to avoid bias due to genotype dropout because of phenotype missingness.

For the permutation analysis, we created 100,000 permuted data sets through random assignment of the genotype vector, thus preserving the correlation structure among traits. Empirical p values for each trait were obtained as the proportion of permutations for which the minimum two-sided p value for all traits was less than or equal to the trait p value observed in the original data set. A one-sided mode was applied by restricting the vector of permuted two-sided p values to those associations where the effect was consistent with the direction of the original effect.

Evidence for profile associations between social communication spectrum phenotypes and rs4307059 was sought using an adapted permutation approach. A factor analytic/principal components-based approach seemed less suitable for the following reasons. First, socioeconomic infiuences, such as maternal education, explain a high proportion of the phenotypic variance (especially for language-related traits; see above) and are likely to load high on the first principal components/orthogonal factors. Thus, these factors are less likely to represent genetic variation as genotypes are generally unrelated to socioeconomic infiuences (34). Because single SNP variation explains only a small proportion overall of the phenotypic variance, often ≤1% even for genome-wide significant signals (35), genetic variation may largely be represented by components or factors that are dominated by stochastic variation. Second, multiple genetic associations may result because of shared phenotypic variation, but also because of pleiotropic effects, that is, multiple phenotype effects per single gene. The choice of a permutation approach, however, allows combining the entire evidence from all traits into a joint signal, irrespective of missingness, phenotype interrelatedness, or distribution and without the need to cluster phenotypic variance. The empirical approach we used thus assessed the combined evidence for genetic association, which is consistent with an underlying QTL for autism spectrum disorder (directed mode). A combined statistic, as suggested by Fisher in 1932 (36), was generated aswhere k is the number of phenotypes and P is the two-sided p value. Empirical p values were obtained as the proportion of permutations in which the summary statistic was greater than or equal to the statistic in the original data set and each single effect was consistent with an underlying autism spectrum disorder QTL. Inconsistent p values were converted to 1. Thus, any inconsistent association in the original data set would not contribute to the original summary statistic but to the random variability during the permutation process.

All analyses were carried out using the R software package (www.r-project.org).

Using quasi-Poisson, ordinary least squares, and multi-nomial regression models, we studied single associations between each of the 29 social communication spectrum phenotypes and rs4307059 in a sample of 7,313 ALSPAC children (see Table 2). A higher load of the autism spectrum disorder risk allele was related to more stereotyped conversation and decreased ability to understand pragmatic aspects of communication as measured by two highly related CCC scales (Spearman's rank-order correlation r_s=0.74, p<0.001; see Table S1 in the online data supplement). It should be noted that the stereotyped conversation scale is a subscale of the pragmatic communication composite scale (26). The expected log count of more functional communication scores decreased by 0.0069 per increase in T allele (95% CI=−0.010 to −0.003) for the stereotyped conversation subscale (CCC; at a mean age of 9.7 years). This was accompanied by a decrease in 0.60 (95% CI=−0.91 to −0.29) pragmatic communication scores per increase in T allele (CCC; at a mean age of 9.7 years), with the genetic variation explaining ∼0.26% of the total phenotypic variance (adjusted R²). An association between special educational needs (at a mean age of 11.8 years) and rs4307059 might be present as a trend.

Nominal associations were also observed between rs4307059 and other CCC scales (at a mean age of 9.7 years), social communication skills (SCDC, at a mean age of 10.8 years), behavioral difficulties (R-RPS-PC, at a mean age of 3.5 years; SDQ, at mean ages of 9.7 and 11.8 years), social interactions (HMP-FQ, at a mean age of 8.6 years), and cognitive functioning (WISC-III, at a mean age of 8.6 years), such that lower functionality, more problems, and fewer friends were associated with a greater frequency of the autism spectrum disorder risk allele (see Table 2).

The exclusion of 41 ALSPAC individuals with autism spectrum disorder did not attenuate the observed signals (data not shown). An association analysis restricted to these 41 individuals was not performed because of the small sample size. Multilevel sensitivity analysis (see part 4 and Table S2 in the online data supplement) on repeated measurements provided no evidence for variation in SNP effects over time. There was also no multilevel support for time-invariant SNP effects.

Closer examination of the observed cross-sectional single-trait associations revealed that for the majority of traits, the detected genetic effects were consistent with the presence of an underlying QTL for autism spectrum disorder (see Figure 1 and Table 2). An exception from this pattern was observed only for sociability (EAS) at a mean age of 4.8 years. Hence, we also investigated the combined phenotype association effect for rs4307059. Using an extended permutation approach, this provided empirical evidence for a joint association between rs4307059 and the phenotypic spectrum of traits (joint p=0.00066). Overall, the evidence for a joint signal was stronger than for any single-trait association alone. This combined signal was driven not only by highly associated single-trait associations but also by a phenotypic profile, as even after the exclusion of the CCC scales assessing stereotyped conversation and pragmatic aspects of communication we observed a joint association between rs4307059 and the remaining measures (joint p=0.0024). Communicative, cognitive, and social interactive components of the spectrum appear to play a major role in this association as further exclusion of information on total behavioral difficulties (R-RPS-PC, at a mean age of 3.5 years; SDQ, at mean ages of 4.0, 6.8, 8.2, 9.7, and 11.8 years) and special educational needs (at a mean age of 11.8 years) only marginally attenuated the effect (joint p=0.0048). None of the implicated phenotypes, including verbal IQ (WISC-III) and social communication skills (SCDC), showed evidence for association when adjusted for multiple testing as a single trait (see Table 2).

These findings must be interpreted within the context of potential limitations. First, although measures were derived from educational databases and mother and child reports, mother-reported data were overrepresented in the study and may have contributed to greater variance sharing. Second, the analysis included only white European children, and the reported genetic association therefore needs to be replicated in populations of different ethnic backgrounds. Third, instruments were chosen that focus on social communication spectrum pheno-types assessed between ages 3 and 12 for the majority of children in ALSPAC. However, they nonetheless represent a cohort-specific selection of measures, which may not be exhaustive. This is unlikely to affect the reported single-trait associations, but it may impinge on the total of the investigated phenotypic variance, which is reflected by the joint signal.