Genome-wide linkage using the Social Responsiveness Scale in Utah autism pedigrees

Subjects were initially ascertained from 70 pedigrees that had at least two family members with an ASD. Six families did not have sufficient SRS phenotype information (defined as at least two family members with SRS scores) and were therefore dropped from this analysis. A total of 629 genotyped subjects from 64 families were used here, 181 of whom were defined as having either a strictly defined autistic disorder (AD) or a more broadly defined ASD. Table 1 gives the characteristics of these families which include 19 large extended pedigrees (6-9 generations), six families of moderate size (4-5 generations) and 39 smaller families (2-3 generations). The 19 extended pedigrees were identified or confirmed using the Utah Population Database (UPDB), a computerized genealogy database that contains family history information for over 6.5 million individuals who are, for the most part, descendants of the nineteenth century Utah pioneers http://​www.​hci.​utah.​edu/​groups/​ppr/​. Previous studies have shown low rates of inbreeding within the UPDB [25, 26]. Using the UPDB, we were able to identify many distant family relationships between the individuals with ASD that were not known to the subjects or their families. These relationships were kept confidential. This study has ongoing approval from the University of Utah Institutional Review Board (IRB). All adults participating in the research signed IRB-approved informed consent documents. All subjects under the age of 18 signed approved assent documents and their parents or guardians signed approved parental permission documents. Subjects in this study are independent from the Utah subjects who were contributed to the Autism Genome Project and studied by Liu et al. [27]. In addition, none of the families in the current analysis were included in the data set studied by Duvall et al. [20].

Families interested in participating were asked to give questionnaire consent, to give initial information regarding possible exclusion criteria and to complete the Social Communication Questionnaire (SCQ) [28]. The SCQ was developed as a parent report measure based on the Autism Diagnostic Interview-Revised [29]. A cutoff score of 15 on the SCQ has been shown to have good sensitivity (0.85) and specificity (0.75) for discriminating ASD from other diagnoses [30]. Subjects were contacted for possible inclusion if their SCQ score was above 15 or if they had a previous diagnosis of ASD. Subjects were excluded if they reported medical conditions known to be associated with autism (tuberous sclerosis, fragile X, neurofibromatosis, congenital rubella or phenylketonia) or evidence of brain injury. To the best of our knowledge, given currently available data, subjects did not have chromosomal anomalies that could confound our findings. If subjects were eligible for the study, they were asked to sign informed consent for DNA and additional assessments. When possible, all subjects with a suspected ASD were then given both the ADI-R and the Autism Diagnostic Observation Schedule-Generic [31] and study diagnoses were made using these assessments. For cases where assessments could not be obtained, the diagnoses were made according to Diagnostic and Statistical Manual of Mental Disorders (DSM)-IV criteria by a psychologist trained in autism assessment (JSM) using all available information (clinical records, other behavioural data, other questionnaire and interview information).

All available subjects were given the SRS. Completed questionnaires were obtained from a total of 523 family members who were genotyped. Completed questionnaires were obtained for 92.1% of index cases, 83.9% of siblings (affected and unaffected), 87.2% of fathers, 78.8% of mothers and 73.7% of more extended relatives (grandparents, great-grandparents). The child version of the SRS was completed for 205 subjects and the adult version was completed by 318 subjects. The child version was filled out by a parent for almost all cases (203/205; 99.02%). The adult version was usually filled out by a spouse (236/318; 74.21%) but parents were raters for adult siblings over age 18 (59/318; 18.6%). Other informants filled out the remainder of the adult SRS forms (23/318; 7.2%). No children were rated by a parent with ASD. However, after surveys were returned, five adult raters were found to have ASD. Scores from these five raters (fathers) on their spouses (mothers) were not used for subsequent analyses, leaving 518 subjects with scores. There were an additional four children whose parent was coded as a poor rater on the ADI for under-reporting of behaviours. While no data regarding rater quality were available for the SRS directly, the similarity of content across the ADI and SRS made ratings done by these parents was suspect. Scores for these subjects were therefore not used in the quantitative analyses of data. For qualitative SRS analysis, data on these subjects were not deleted because possible under-reporting on the SRS did not impact upon the SRS category.

Genotyping services were provided by the Center for Inherited Disease Research (CIDR) using the 6 K single nucleotide polymorphism (SNP) Illumina Linkage Panel 12, which includes 6090 SNP markers, with an average genetic coverage of 0.65 cM. There were 46 SNPs who were dropped due to poor performance, leaving 6044 SNPs. Data from 653 pedigree subjects passed stringent quality control criteria for initial inclusion in the study [32]. Of the 629 genotyped subjects in 64 families used in this study, 518 of had SRS data.

Eliminating linkage disequilibrium (LD) between markers in linkage studies has been strongly recommended, as false positive results can occur in the presence of LD [33‐36]. As these studies have shown, increases in type I error are most dramatic when LD is high and when there are missing genotypes for parents and other connecting relatives. Recommended thresholds of acceptable LD vary, but a pair-wise r² value of 0.05 between SNPs has now been supported with extensive simulation studies [34]. Therefore, prior to linkage analysis, we screened all SNPs for LD using the PLINK software package [37], which recursively removes SNPs in LD within a sliding window. We set a window size of 50 SNPs, shifted the window by five SNPs at each step, and used a variance inflation factor of 1.5, which is equivalent to a multiple correlation coefficient R² of 0.33 regressed simultaneously over all SNPs in the selected window. This R² considers not only the correlations between SNPs but also between linear combinations of SNPs [37] and corresponds in our data to a pair-wise r² of approximately 0.05. This screening for LD removed 1207 SNPs. Also, as part of the validation procedure, we removed 115 SNPs with a minor allele frequency less than 0.10 and 4 SNPs out of Hardy-Weinberg equilibrium. The total number of SNPs left after this phase was 4718.

We used the genetic map provided by CIDR based on the deCODE genetic map [38]. Base pair positions were obtained from the March 2006 human reference sequence (hg18) assembly. The SRS was first treated as a qualitative trait. Published cutoff scores [15] were used to delineate three groups: SRS unaffected (raw score below 54), SRS spectrum (raw score between 54 and 87) and SRS affected (raw score 87 or above). Subjects who were missing SRS scores were coded as unknown.

This qualitative phenotype was tested using MCLINK, a Markov chain Monte Carlo (MCMC) method that allows for multilocus linkage analysis on large extended pedigrees [39]. Using blocked Gibbs sampling, MCLINK generates inheritance matrices for haplotypes of the markers being analysed and estimates the log-likelihood function linkage statistics. This method results in the reconstruction of haplotype configuration for the entire pedigree structure, allowing for analysis of the full genetic information available in the pedigree. MCMC methods therefore maximize the utility of extended pedigree data, as genetic information increases with the number of meioses in a family [40]. Further, the multipoint analysis maximizes the genetic information available from the entire map of SNP genotypes. In multipoint analyses, genetic information increases with the density of the genotypes [41‐43] but is not decreased, with exclusion of markers in high linkage disequilibrium [42]. Internally, MCLINK runs the analysis five times to ensure a consistent solution. Results from MCLINK have shown a high degree of similarity to other MCMC linkage methods [44] and also to exact linkage methods and variance components linkage methods as applied to extended pedigrees [45]. MCLINK has been previously used to identify candidate genomic regions for a number of complex diseases in extended pedigrees [44, 46‐48]. Allele frequencies for the MCLINK analysis were estimated using all of the observed data.

We performed a non-parametric analysis and a general parametric model-based analysis using simple dominant and recessive model parameters that reproduced the reported population frequency of ASD [49]. For the recessive model, the disease allele frequency was set at 0.05; a penetrance of 0.01, 0.01, 0.5 was assigned to those with SRS scores in the spectrum category and 0.0014, 0.0014, 0.8 was assigned to those with SRS scores in the SRS affected category. For the dominant model, the disease allele frequency was set at 0.0025; penetrances were 0.01, 0.5, and 0.5 for those with SRS scores in the spectrum category and 0.0014, 0.8, 0.8 for those with SRS scores in the SRS affected category. Nonparametric methods provide an analytic approach independent of model assumptions. However, for the analysis of extended pedigrees, the parametric approach may be preferable. Parametric analysis provides assumptions about the genotype-phenotype relationship, simplifying the parameter space and allowing for more powerful and efficient analyses without increasing the false positive rate [22, 50, 51]. The multipoint heterogeneity LOD score (HLOD) allows for a proportion of unlinked pedigrees (1-alpha) and variation in the recombination fraction. The HLOD provided by MCLINK is robust to model misspecification and may reflect the true position of linkage peaks more accurately under conditions of appreciable heterogeneity [47, 50, 52]. HLOD scores have been shown to be more powerful than homogeneity LOD scores or model-free methods under these conditions [52, 53]. The HLOD has been shown to produce scores consistent with other published methods [47, 54]. P-values from the non-parametric qualitative analyses were converted to LOD scores for ease of interpretation and comparison with the parametric results. We present results in the commonly used framework of significance thresholds proposed by Lander and Kruglyak [55]: LOD score ≥ 1.88 (P = 0.0017) represents suggestive evidence and LOD score ≥ 3.30 (P = 0.000048) represents significant evidence. We recognize that our study design does not match perfectly with the design that generated these thresholds but they do provide reasonable benchmarks which enable us to organize and present our data. All P-values are theoretical, not empirical.

In order to analyse the SRS as a quantitative trait, we first applied a square root transform the SRS data to adjust for non-normality in the distribution, as has been done with previous SRS analyses [17]. The effects of age and gender were partialled out using regression. MCLINK was used to perform parametric quantitative trait analysis. In a quantitative trait analysis there is no concept of 'affected' or 'unaffected'; all individuals with a trait value are included in the analysis, other individuals are considered unknown. The MCLINK nonparametric analysis considers quantitative similarity among relatives and has been previously described [56]. For the parametric quantitative analysis, the penetrance vector for an individual is based on their observed trait value and three specified trait distributions for the three genotypes at the disease locus (no risk allele, one risk allele, two risk alleles). The three trait distributions are unknown, but can be estimated from the observed trait data. The SRS score cutoffs provide theoretical categories of unaffected, spectrum and affected subjects. Here we considered these three categories as a representing a hypothetical genetic model of: no risk allele, one risk allele and two risk alleles, and means and standard deviations of adjusted SRS scores were estimated. The three distributions for zero, one or two risk alleles were specified to be normal distributions with means of -1.552, 1.184, and 2.558 and standard deviations of 2.238, 1.581, and 1.356, respectively. The penetrance vector for an individual is simply the relative probabilities of the trait being observed from the three distributions. Only the relative probability between the three penetrance values is important as the linkage statistic is a likelihood ratio. For efficient calculation of parametric quantitative LOD scores, the continuous penetrance function is discretized into multiple classes (here, 40) and all individuals with trait values residing in a class are assigned penetrance vectors at the mid-point of the class. Classes were defined starting -3SD below the mean of the zero risk alleles distribution and adding equal increments to reach +3SD above the mean of the two risk alleles distribution.

In summary, we note that use of the SRS results in three important differences between this analysis and our previous affected-only analysis using clinical diagnosis on these families. First, this analysis includes information from an additional 337 clinically unaffected relatives. Second, for the qualitative analysis, affection status was defined using SRS data rather than clinical measures. Third, a quantitative analysis uses a full range of scores on autism-related traits in all measured pedigree members.