Subjects with ASD and their biological parents were recruited through the Korean Autism Research Consortium. Each child was initially screened for ASD by 2 board-certified child psychiatrists using the Diagnostic and Statistical Manual of Mental Disorders diagnostic criteria. To confirm the diagnosis, all subjects were evaluated on the basis of the Korean version of the Autism Diagnostic Observation Schedule (K-ADOS) and the Autism Diagnostic Interview-Revised (K-ADI-R). All subjects met the diagnostic criteria of ASD.^14,15 In addition, all subjects included in this study were not verbally fluent in simple conversational tasks (as expected at age >4 years), and the K-ADOS module 1 was applied. In the K-ADI-R, parents reported that their children showed word and/or phrase delay during development (>24 months and >33 months, respectively).

The psychometric properties of the probands were evaluated using the Korean Educational Developmental Institute-Wechsler Intelligence Scale for Children or the Korean version of the Vineland Social Maturity Scale (K-VSMS) depending on the capacity of the children.^16,17 We performed physical and neurological examinations, including electroencephalography (EEG) and chromosomal analyses, to reveal any physical or neurological conditions. Subjects who were diagnosed with neurofibromatosis, metabolic encephalopathy, organic brain diseases, fragile X syndrome, tuberous sclerosis, or those who were diagnosed with chromosomal abnormalities or other medical conditions that might be associated with ASD were excluded from the analysis. We received written informed consent from the parents. This study was approved by the institutional review boards of the institutions where the study was performed.

Genomic DNA was purified from whole blood samples using the FUJIFILM DNA Whole Blood Kit S and QuickGene-810. Concentration and purity analyses were performed for all samples using a NanoDrop ND-1000 spectrophotometer, and the integrity of the samples was tested by electrophoresis on a 1% agarose gel. The 260/280 optical density ratio of the samples had to be higher than 1.8 and the 260/230 ratio had to be higher than 2.0 for the samples to be included in the genotyping analyses. DNA aliquots (500 ng) were then prepared at a concentration of 50 ng/µL in a total volume of 10 µL.

After the samples were determined to be within the defined range, they were run on the Affymetrix Genome Wide 5.0, scanned, and analyzed. Array images were acquired using a GeneChip Scanner 7G with an autoloader that scanned each array. Raw DAT image files were generated using the GeneChip Operating System (GCOS) software. Each DAT image was processed by the GCOS software to generate a feature-extracted .CEL file.

All .CEL files were subjected to low-level quality control (QC) analysis using the Genotyping Console 2.1 software (Affymetrix) to determine their suitability for genotyping. This QC analysis included assessment of image quality to ensure that it was free of manufacturing or physical defects. Next, we examined the QC call rate (generated automatically when the .CEL files were imported into the Genotyping Console) of approximately 3,022 single-nucleotide polymorphisms (SNPs). The analysis of these SNPs has been reported to be sensitive to the DNA quality. This step included separate assessments of the QC call rates for SNPs that were examined in the NspI and StyI fragments. We only performed genotyping analysis on .CEL files with overall and fragment-specific QC call rates that exceeded 86%. Arrays that passed these criteria were subjected to Bayesian Robust Linear Modeling using Mahalanobis Distance genome-wide genotyping by using the Genotyping Console 2.1 at a confidence threshold score of 0.05. The mean value for the sample call rate was 98.1%. The IBS score was 1.57±0.55. Graphical representation of relationship errors (GRR), a graphical tool for verifying assumed relationships between the individuals in genetic studies, was used to detect common errors when using genotypes from many markers.¹⁸ SNPs were also subject to QC before analysis. To minimize genotyping errors, we excluded SNPs with a p value of <10^-4 from the calculation of the Hardy-Weinberg equilibrium and minor allele frequencies below 1% from the analysis when using PLINK 1.0.4 (http://pngu.mgh.harvard.edu/~purcell/plink/). After drawing Q-Q plots based on a call rate between 90% and 99%, we selected a call rate of 95% to control the marker quality. After application of the QC filters, 331095 out of 440094 SNPs remained.

We determined the Mendelian inheritance error and tested the family-based association for each individual polymorphism using the standard TDT method. In addition, we used MDR-PDT to detect epistasis on a genome-wide scale with 194 markers that had a p-value of <10^-3 in the TDT test.¹⁹ The false-discovery rate (FDR) is a method that considers the expected proportion of significant tests that are truly null. The FDR procedures proposed by Benjamini and Hochberg²⁰ were applied to adjust for multiple comparisons.

The average age of the 42 probands was 77.7±22.6 months (mean±SD; range, 49-149 months). All probands were males. The social quotient measured using the K-VSMS was 50.5±14.8 (range, 23-72 months). The mean IQ score, available only for 9 subjects because of the low level of functioning, was 46.2±12.2 (range, 31-65). The K-CARS score was 33.3±4.4 (range, 23-46). The average age at which the children spoke their first words, as reported by the parents in the K-ADI-R, was 35.9±21.6 months (range, 10-85 months). The age at which the children spoke their first phrase was 58.1±21.4 months (range, 13-108 months). Thirteen patients (30.6%) had EEG abnormalities suggesting a partial seizure or diffuse cerebral dysfunction, but none were diagnosed with a clinically significant partial seizure disorder.

Nine subjects (21.4%) had not yet spoken any significant meaningful single words, whereas another 13 subjects (30.6%) were not able to say developed 2-word phrases, including verbs, in spite of their ability to use 5 different single words every day. The mean communication domain score on K-ADOS was 6.0±1.5, and the score for the qualitative abnormalities in the communication domain on K-ADI-R was 14.4±3.1. All subjects obtained scores that exceeded the cut-offs. The subdomain score for the lack of/delay in spoken language and failure to compensate through gestures in the nonverbal subjects was 7.4±1.1, and the score for the lack of varied spontaneous make-believe or social imitative play was 5.7±0.7. All subjects showed significant disturbances in social interaction, repetitive behavior, and restricted interest domain based on the diagnostic algorithms of both K-ADOS and K-ADI-R.

The distribution of p values for the TDT is shown in Table 1. The distribution of p values examined in the discovery dataset was found to be closely matched to that expected for a null distribution, except at the extreme tail of low p values (Figure 1). The results of the family-based genome-wide association analyses are presented in Figure 2. The 30 most powerful properties according to the TDT results are shown in Table 2. The most statistically significant association before correction of multiple comparisons was found for an SNP (rs11212733) on chromosome 11q22.3 (p-value=9.76×10^-6).

In the MDR-PDT analysis, we detected best models from 1-locus to 4-locus. The results of the MDR-PDT are presented in Table 3 and Figure 3. In particular, rs7125479, which is also located in chromosome 11, was contained 1-locus, 2-locus, and 4-locus model. We also found that rs7125479 and rs11212733, the most significant SNPs in this study, were in linkage disequilibrium (r²=1.0) in the HapMap database. After correction for multiple comparisons using FDR, none of the SNPs remained significant.