Candidate-Gene Screening and Association Analysis at the Autism-Susceptibility Locus on Chromosome 16p: Evidence of Association at GRIN2A and ABAT

Abstract

Autism is a highly heritable neurodevelopmental disorder whose underlying genetic causes have yet to be identified. To date, there have been eight genome screens for autism, two of which identified a putative susceptibility locus on chromosome 16p. In the present study, 10 positional candidate genes that map to 16p11-13 were examined for coding variants: A2BP1, ABAT, BFAR, CREBBP, EMP2, GRIN2A, MRTF-B, SSTR5, TBX6, and UBN1. Screening of all coding and regulatory regions by denaturing high-performance liquid chromatography identified seven nonsynonymous changes. Five of these mutations were found to cosegregate with autism, but the mutations are not predicted to have deleterious effects on protein structure and are unlikely to represent significant etiological variants. Selected variants from candidate genes were genotyped in the entire International Molecular Genetics Study of Autism Consortium collection of 239 multiplex families and were tested for association with autism by use of the pedigree disequilibrium test. Additionally, genotype frequencies were compared between 239 unrelated affected individuals and 192 controls. Patterns of linkage disequilibrium were investigated, and the transmission of haplotypes across candidate genes was tested for association. Evidence of single-marker association was found for variants in ABAT, CREBBP, and GRIN2A. Within these genes, 12 single-nucleotide polymorphisms (SNPs) were subsequently genotyped in 91 autism trios (one affected individual and two unaffected parents), and the association was replicated within GRIN2A (Fisher's exact test, P<.0001). Logistic regression analysis of SNP data across GRIN2A and ABAT showed a trend toward haplotypic differences between cases and controls.

Introduction

Autism is a severe neurodevelopmental disorder that affects reciprocal communication and social interaction and is associated with repetitive and stereotyped behaviors. Recent epidemiological studies have revised the population prevalence of autism upward from 4 in 10,000 to 9–11 in 10,000 (Chakrabarti and Fombonne 2001; Fombonne 2003), and the overall rate of pervasive developmental disorders is ∼60 in 10,000 (Fombonne 2002). Autism is highly heritable and shows strong familial clustering, with a λ (sibling relative risk) of 75–100, although the mode of inheritance is complex (Bailey et al. 1995; Pickles et al. 1995). Autism is found in three-to-four times as many males as females, which suggests a possible sex-linked effect, but multivariate analysis of the phenotype in relatives does not support X-linked inheritance (Pickles et al. 2000). Moderate evidence of linkage to the X chromosome has been observed in two genome screens (Liu et al. 2001; Shao et al. 2002). A single mutation in each of the X-linked neuroligin genes NLGN3 and NLGN4 has also been reported, but the low frequency of the mutations makes it unlikely that they contribute significantly to autism etiology (Jamain et al. 2003; Stone et al. 2003; Vincent et al. 2004). Autism is associated with phenotypes of known etiologies in ∼6% of cases, including fragile-X syndrome, epilepsy, neurofibromatosis, and tuberous sclerosis (Fombonne 2002). The increased comorbidity of these disorders with autism may indicate a contribution to the autism phenotype by disruption of a common neurological—although not necessarily genetic—pathway.

Recently, the IMGSAC performed a whole-genome screen of 83 affected sib pairs (ASPs). Thirteen loci that displayed evidence of linkage were subsequently genotyped for 152 sib pairs (IMGSAC 2001). The three highest-scoring loci detected were on 2q21-33 at D2S2118, with a multipoint maximum LOD score (MLS) of 3.74; 7q22.2-31.3 at D7S477, with a multipoint MLS of 3.20; and 16p11-13 at D16S3102, with a multipoint MLS of 2.93. Detailed analysis of sex-specific and parent-of-origin effects at these loci has recently been reported (Lamb et al. 2005). The linkage to chromosome 16p was found in male-male ASPs (MLS of 2.48 for 145 male-male ASPs vs. MLS of 0 for 74 male-female or female-female ASPs) (Lamb et al. 2005). D16S3102 at the peak of linkage was physically mapped at 12.46 Mb in the May 2004 release of the Human Genome Sequence. Linkage overlapping the region on chromosome 16p, between 10 cM and 23 cM, has been reported by Philippe et al. (1999), with a multipoint MLS of 0.74 at 10.6 cM (n=51 multiplex families). Additionally, trisomy of 16p and duplications mapping to 16p11.2 and 16p12.2 have been reported for individuals with autism traits (Finelli et al. 2004).

To identify the autism-susceptibility gene at 16p11-13, positional functional candidate genes were selected, and their coding sequences were screened by denaturing high-performance liquid chromatography (DHPLC) for 48 individuals (from multiplex families) who contributed to linkage on 16p11-13. Variants detected by DHPLC were further characterized by sequencing, and missense variants were tested for cosegregation with autism. The candidate genes that were examined in the present study are described in table 1.

Table 1.

Candidate Genes Screened for Autism-Susceptibility Variants

Gene Symbol	Gene Name	Positiona	GenBankAccessionNumber	Genomic Region(kb)	mRNALength(bp)	No. of Exons
A2BP1	Ataxin 2–binding protein	6009132–7700843	NM_018723	420.449	2,279	16
ABAT	4-Aminobutyrate aminotransferase	8675992–8782887	BC031413	106.895	1,705	14
BFAR	Bifunctional apoptosis regulator	14634296–14670193	NM_016561	35.898	2,531	8
CREBBP	CREB-binding protein	3716569–3870712	NM_004380	154.144	8,694	32
EMP2	Epithelial membrane protein 2	10533475–10582040	NM_001424	48.566	690	4
GRIN2A	Glutamate receptor, ionotropic, NMDA 2A	9762922–10184112	NM_000833	421.191	6,293	14
MRTF-B	Myocardin-related transcription factor B	14235623–14268130	NM_014048	32.508	7,815	8
SSTR5	Somatostatin receptor 5	1068869–1069964	NM_001053	1.096	1,095	1
TBX6	T-box 6	30004582–30010709	NM_004608	6.128	2,468	9
UBN1	Ubinuclein 1	4842216–4872364	NM_016936	30.149	5,747	18

^aFrom the May 2004 release of the International Human Genome Sequencing Consortium.

In total, 38 SNPs were selected for genotyping in 239 multiplex families and were tested for association by use of the pedigree disequilibrium test (PDT). These included three SNPs in TSC2; TSC2 maps to 16p13 and encodes one of the genes that cause tuberous sclerosis, which has been reported to be associated with autism. Following the suggestion that there is a phenotypic overlap between attention-deficit/hyperactivity disorder (ADHD) and autism, four SNPs previously analyzed in a study of ADHD that found linkage to chromosome 16p were also genotyped (Smalley et al. 2002; Ogdie et al. 2003, 2004), as shown in table 2. All SNPs were also genotyped for 192 controls, and the allele and genotype frequencies were compared with those of 239 unrelated affected individuals. SNP data were analyzed for linkage disequilibrium (LD), association with autism, and haplotype transmission disequilibrium. Twelve SNPs that showed evidence of association were subsequently genotyped for 91 autism trios (one affected individual and two unaffected parents). Logistic regression analysis was performed for haplotypes from genes containing SNPs that showed evidence of association.

Table 2.

SNPs Genotyped in TSC2 and below ADHD Linkage Peak

Genomic Region anddbSNP Number	Positiona	Allele Variant
TSC2b:
rs2516740	2037111	G→T
rs2074968	2050572	G→C
rs1051771	2078585	G→C
ADHD linkage peakc:
rs153783	15509867	G→A
rs1125972	15514488	G→A
rs1065838	15535539	T→C
rs1107143	15594082	T→C

^aFrom the May 2004 release of the International Human Genome Sequencing Consortium.

^bThe distance between rs2516740 and rs2074968 is 13,461 bp; the distance between rs2074968 and rs1051771 is 28,013 bp.

^cThe distance between rs153783 and rs1125972 is 4,621 bp; the distance between rs1125972 and rs1065838 is 21,051 bp; the distance between rs1065838 and rs1107143 is 58,543 bp.

Description of Candidate Genes

GRIN2A (MIM 138253) and SSTR5 are cell-surface receptors. GRIN2A encodes one of four N-methyl-D-aspartate (NMDA) receptor 2 subunits that form an NMDA-receptor channel. NMDA receptors have been shown to be involved in long-term potentiation (LTP), an activity associated with learning and memory. Targeted disruption of the mouse homologue of GRIN2A, GluRɛ, reduces LTP and produces other behavioral deficits (Sakimura et al. 1995). Evidence of association between a GRIN2A exon 5 polymorphism and ADHD has also been reported recently (Turic et al. 2004). SSTR5 (MIM 182455) encodes a G-protein–coupled receptor for the inhibitory hormonal regulator somatostatin. There is evidence that somatostatin and its receptors are transiently expressed in developing neurons; they are thought to control cell migration. Lauritsen et al. (2003) have performed analysis of 12 polymorphisms in SSTR5 in 79 families with autism but found no evidence of association.

Three of the genes screened are involved in transcriptional regulation. TBX6 (MIM 602427) is a member of the T-box family that encodes transcription factors involved in the regulation of developmental processes and is expressed in discrete embryonic domains (Papapetrou et al. 1999). UBN1 encodes a ubiquitously expressed nuclear protein found to interact with EB1 and c-Jun, which, in turn, interact with cellular promoters and transcription factors. CREBBP (MIM 600140) encodes a large protein of 2,442 aa and binds specifically to the protein-kinase-A–phosphorylated form of the CREB protein (Chrivia et al. 1993). CREBBP can activate transcription through its C-terminus and acts to augment the ability of phosphorylated CREB to activate cAMP-responsive genes, interacting directly with transcription factor IIB. Mutations in CREBBP are responsible for Rubenstein-Taybi syndrome (RTS [MIM 180849]), which is characterized by mental retardation and broad thumbs and halluces. RTS has been associated with autism-like features in a minority of cases (Hellings et al. 2002).

EMP2 (MIM 602334) and BFAR are involved in cell-cycle control. EMP2 has high sequence similarity with peripheral myelin protein 22, suggesting that it may have a similar function in cell proliferation and cell-cell interactions. BFAR encodes a protein that inhibits Bax-induced and Fas-induced apoptosis (Itoh et al. 1993; Wei et al. 2001).

The remaining three candidate genes were chosen because of their expression patterns and potential connection with neurological disease. ABAT (MIM 137150) is responsible for the catabolism of gamma-aminobutyric acid (GABA) into succinic semialdehyde (De Biase et al. 1995). GABA is an important inhibitory neurotransmitter in the brain and is estimated to be present in 20%–50% of human synapses. GABA is involved directly and indirectly in the pathogenesis of many neurological diseases, including convulsions, which may be relevant, given that approximately one-third of individuals with autism develop epilepsy (Fombonne 2002). A2BP1 (MIM 605104) possesses a ribonucleoprotein motif, which is highly conserved among RNA-binding proteins. A2BP1 binds to the C-terminus of ataxin-2 and may contribute to the pathology of spinocerebellar ataxia type 2 (SCA2 [MIM 183090]) (Shibata et al. 2000). MRTF-B was identified from cDNA libraries of size-fractioned adult and fetal human brain (Nagase et al. 1999). MRTF-B is highly expressed in the amygdala, caudate nucleus, hippocampus, and fetal brain and was recently found to be involved in muscle differentiation (Selvaraj and Prywes 2003).

Material and Methods

IMGSAC Family Collection

The IMGSAC has collected family data in successive stages, and details of assessment protocols are described elsewhere (IMGSAC 2001). To summarize, after an initial screen for medical disorders and a preliminary assessment, parents were administered the Autism Diagnostic Interview–Revised (ADI-R) and the Vineland Adaptive Behavior Scales (Sparrow et al. 1984; Le Couteur et al. 1989). Probands were administered the Autism Diagnostic Observation Schedule (ADOS) (Lord et al. 2000), and psychometric data were obtained. Affected individuals also underwent a thorough medical examination so that those with tuberous sclerosis and neurofibromatosis could be excluded, and any morphological abnormalities were recorded. All participating clinical sites followed the same assessment protocol, and ADI-R and ADOS consensus-coding meetings were held regularly by clinicians. DNA was extracted from blood, buccal swabs, or cell lines by use of the Nucleon kit and standard techniques. Samples from one affected individual per multiplex family were tested for fragile-X syndrome, and, when possible, all affected individuals underwent karyotypic analysis (Lamb et al. 2005). Relevant ethical committees have reviewed this study.

The IMGSAC has collected a further 67 sibling pairs since the publication of the completed genome screen, for a total of 219 sibling pairs from 207 families. In addition, 32 families containing extended relative pairs have been collected. The male:female ratio of affected individuals is 3.9:1. One affected individual was selected at random from each multiplex family, to create a group of 239 affected unrelated individuals for case-control analysis. A singleton sample consisting of 91 autism trios with a clinical diagnosis of autism has also been collected. The singleton sample consists of 22 trios from Denmark, 37 trios from the Netherlands, and 32 trios from the United Kingdom. Analysis of linkage data from the IMGSAC sample on chromosomes 2,7, and 16 does not indicate that country of origin significantly affects either the magnitude or location of the linkage signal or the evidence of association in the complete IMGSAC sample (authors' unpublished data) (Bonora et al. 2005).

Controls

Random controls from the European Collection of Cell Cultures (ECACC) (n=192) were used for SNP typing. DNA was extracted at ECACC from lymphoblastoid cell lines derived from white randomly selected blood donors whose parents and grandparents were born in the United Kingdom or Ireland. The control samples were not characterized for autism phenotypes. Half of the controls were male and half female. Although the male:female ratio is different for cases and controls, only the presence of a very strong sex-specific allelic effect could reduce the ability to detect association. Linkage analysis based on the sex of APSs on chromosome 16 shows some evidence of a sex-specific effect, but it is unclear how this would affect the frequency of a disease-causing variant (Lamb et al. 2005).

Mutation Screening

Intron-exon boundaries were characterized in silico by use of National Center for Biotechnology Information LocusLink and University of California–Santa Cruz (UCSC) Golden Path (see UCSC Genome Browser Web site). The mRNA sequence was subjected to BLAST analysis, to determine or confirm genomic structure. Exon-specific primers were designed using Primer3. Large exons and promoter regions were covered by overlapping PCR products of ∼500 bp. PCRs were performed in a reaction mixture containing 40 ng of DNA and final concentrations of 2.5 mM KCl; 2.0, 2.5, or 3.0 mM MgCl₂; 0.2 mM primer; 0.2 mM dNTPs; 0.9 U AmpliTaq Gold polymerase; and 0.1 U PfuTurbo (Stratagene). For GC-rich products of 2.5%–10%, DMSO was used with a mix of 0.2 mM dATP, dTTP, and dCTP; 0.1 mM dGTP; and 0.1 mM 7-deaz-2′-deoxyguanosine-5′-triphosphate (Amersham Pharmacia Biotech). The PCR consisted of touchdown protocols with a maximum annealing temperature of 50°C–67.5°C (T1) and a minimum annealing temperature of 42.5°C–60°C (T2) (15 min at 95°C, followed by 14 cycles of 30 s at 95°C, 30 s at T1−0.5°C per cycle, and 30 s at 72°C, followed by 29 cycles of 30 s at 95°C, 30 s at T2, and 30 s at 72°C). PCR products were denatured and reannealed using the following protocol: 4 min at 95°C, 30 s at 95°C, and 42 cycles of 30 s at −1.6°C per cycle.

The Transgenomic WAVE DNA Fragment Analysis System was used to detect heterozygous individuals by DHPLC. Gradients were designed using WAVEMAKER Software (version 4.1) and spanned a temperature range of 51.5°C–70°C. Crude PCR products were loaded onto the DNASep Column, and elution profiles were analyzed for heteroduplex formation. Melting curves for fragments containing c⁷dGTP were experimentally determined.

Variant Identification

Common variants were identified by sequencing seven heteroduplex samples and a control sample with a homoduplex profile. Low-frequency variants were sequenced for all individuals showing the heteroduplex profile and for a control sample. Sequencing was performed with BigDye Terminator mix on ABI 377 machines, in accordance with standard protocols, and variants were analyzed with Sequence Analysis and Sequence Navigator.

SNP Typing

Mass-extension assay

Primers and mass-extension assays were designed for Sequenom SNP typing with SpectroDESIGNER. PCRs were performed in 10-μl reactions with 8 ng of DNA template and final concentrations of 2.5 mM KCl and MgCl₂, 200 μM dNTPs, 1 U of HotStarTaq Polymerase (Qiagen), and primer concentration of 0.15 mM. Multiplex PCRs were performed using up to four primer pairs per reaction; 45 cycles of 20 s at 95°C, 30 s at 56°C, and 1 min at 72°C were performed. Nonincorporated dNTPs were removed with shrimp alkaline phosphatase for 20 min at 37°C and for 5 min at 85°C. The mass-extension reaction was performed using MassEXTEND enzymes—thermosequenase, hME termination mixes, and hME extension primers; 55 cycles were performed for 5 s at 94°C, for 5 s at 52°C, and for 5 s at 72°C. Unincorporated ddNTPs and dNTPs were removed with SpectroCLEAN resin, and products were transferred to a 384 SpectroCHIP by use of the SpectroPOINT robot. The chip was read using the Bruker Biflex III Mass Spectrometer system. Sequenom Genotype Analyzer was used to manually check failed genotypes and to assign alleles, when possible. Data either were exported from SpectroTYPER using Report Generator or were directly streamed into the Wellcome Trust Centre for Human Genetics Laboratory Information Management System Integrated Genotyping System (IGS).

Restriction digest

PCR for ABAT exon 2 was performed in 10-μl reactions containing 24 ng of DNA and final concentrations of 2.5 mM KCl, 2.5 mM MgCl₂, 0.2 mM primer, 0.2 mM dNTPs, and 0.25 U AmpliTaq Gold polymerase. Touchdown PCR was performed with a maximum annealing temperature of 62.5°C. PCR product from each plate was tested on agarose and was incubated with 0.1 U of BsoB1 and 2 μl of buffer 2 (New England BioLabs), in a total volume of 20 μl at 37°C for 6 h. Digested products were visualized on a 3% agarose gel. Digests were genotyped blind to affection status and were entered into the IGS.

Statistical Analysis

Association analysis

The PDT was used to test for transmission disequilibrium (Martin et al. 2000) for the multiplex and singleton samples. This test is more powerful than the transmission/disequilibrium test (TDT) for the multiplex sample and analyzes informative branches of extended families with all available data. The PDT retains the within-family nature of the TDT and provides a valid test for association, even in the presence of population substructure. A test for the transmission disequilibrium of genotypes—rather than alleles—was also performed using the geno-PDT, an extension of the PDT (Martin et al. 2003a). Case-control association analysis was performed using related cases from the multiplex families and controls, as proposed by Slager and Schaid (2001). This test accounts for the biological relationship between cases and compares genotype frequencies between cases and controls by comparing the trend in proportions as the dosage of the risk allele increases. Genotype-frequency differences between unrelated affected cases from the multiplex and singleton samples and controls were analyzed using Fisher’s exact test for count data.

Haplotype analysis

Haplotypes were determined using PHASE v2.0.2 for groups of SNPs within candidate genes. Any genotypes that were not assigned phase with >90% certainty were disregarded. Because PHASE v2.02 reconstructs haplotypes from population genotype data, two rounds of haplotype reconstruction were performed. First, multiplex case and control data were analyzed together, and, second, singleton case and control data were analyzed together, with the result of small differences in the numbers of haplotypes constructed for controls. PHASE v2.0.2 was selected instead of GENEHUNTER 2.0 for determination of haplotypes, because PHASE v2.0.2 does not assume linkage equilibrium over short distances and can be used to generate haplotypes for individuals without parental genotypes. Haplotype analysis was performed using TRANSMIT for both multiplex and singleton data (Clayton and Jones 1999). TRANSMIT implements a TDT that analyzes the transmission of multilocus haplotypes and has good power when parental genotypes are available, as is the case in this sample (∼80%) (Martin et al. 2003b). For each haplotype, a χ² test with 1 df for excess transmission is produced. A global test for association is also produced with H-1 df (H=number of haplotypes for which information is available).

Logistic regression analysis

Logistic regression analysis of haplotype data was performed with the program R (R Project for Statistical Computing), as described by Wallenstein et al. (1998). Unrelated affected individuals from the multiplex and singleton families, identical to those used in the case-control analyses, were used in the logistic regression analysis. They impose an assumption of additivity on the haplotype analysis, thus providing predictor variables that include the number of copies (0, 1, or 2) of each haplotype and that yield a logistic regression for which the outcome is either case or control. The model is constructed so that each coefficient gives the log odds ratio for disease for an individual with a single copy of the relevant haplotype and another copy of the baseline haplotype, compared with an individual with two copies of a baseline haplotype. Models were fitted in which each gene (ABAT, CREBBP, GRIN2A, and TSC2) was tested alone and jointly. Genotypes for an individual (i.e., the pairs of haplotypes) are denoted as (x;y).

LD mapping

Haplotype files were analyzed using HaploXT to generate an output file describing LD (Abecasis and Cookson 2000). Individual marker-by-marker contingency tables were produced and imported into the Graphical Overview of Linkage Disequilibrium (GOLD) program. Lewontin’s standardized disequilibrium coefficient D′ was used (Lewontin and Kojima 1960), because it accounts for allele frequency, although, like other measures of LD, it may be artificially inflated by rare haplotypes.

Error detection

Mendelian consistency of SNP genotype data was checked using PedCheck (O’Connell and Weeks 1998), and any inconsistent genotypes were removed. Data were prepared for statistical analysis by use of Mega2 (Mukhopadhyay et al. 1999; see the University of Pittsburgh Department of Human Genetics Web site). Genotypes flanking double recombinants, detected after running GENEHUNTER 2.0, were checked and disregarded if ambiguous. SIBMED (sib-pair mutation error detection) was also used to remove possible genotyping errors (Douglas et al. 2000).

Permutation of affection status: Fisher's exact test

An empirical experimentwise P value for association, to allow for multiple testing across all 28 SNPs, was obtained by random permutations of case and control status. For each permutation, the minimum P value from Fisher’s exact test across all SNPs was recorded. An empirical estimate of the experimentwise P value was then given by the proportion of permutations for which the minimum P value exceeded that obtained for the original sample. The same procedure was repeated for the singleton cases and controls.

Permutation of affection status: logistic regression analysis

To test the significance of the logistic haplotype regression analysis of the four genes, with consideration of multiple testing, PHASE v2.0.2 was used to generate “best pair” haplotypes for cases and controls; affection status was then permuted 10,000 times. An additive logistic regression model incorporating the joint haplotypic effects across ABAT, CREBBP, GRIN2A, and TSC2 was fitted to the permuted data, and P values were recorded. The significance of the joint model was estimated as the proportion of permutations in which the regression P value was smaller than that observed in the joint analysis of the original sample data.

Bioinformatic Analysis

Variants found within 20 bp of intron-exon boundaries were analyzed for possible effects on splice-site signal by use of sequence analysis tools available at the Berkeley Drosophila Genome Project Web site (Brunak et al. 1991; Kulp et al. 1997). Variants within exons were analyzed for possible changes to exon-splice enhancers by use of the ESEfinder interface (Cartegni et al. 2003). The program Pfam was used to determine the location of coding variants within known protein domains (Bateman et al. 2004). Homology searches by use of PolyPhen were performed to determine the extent of conservation at the position of missense mutations in known sequences (Sunyaev et al. 2001). PolyPhen uses empirical rules and structural information to characterize the effect of an amino acid substitution.

Results

Failure Rates and Error Checking

Of the 38 SNPs initially selected for genotyping, 10 were disregarded because of failing multiplex PCR, failing extension reaction, poor success rate, or a weak extension assay. Additionally, the A/G SNP causing the G56R substitution in ABAT was genotyped by restriction enzyme digestion because of spurious heterozygotes in the mass-extension assay. Of a total of 47,646 SNP genotyping assays performed, 44,647 genotypes were generated, a 93.7% success rate. Analysis with PedCheck led to the removal of 35 genotypes from the IMGSAC data (0.08% error rate). Most of these were due to nonconservative SNP genotyping allele calls from the Sequenom Genotype Analyzer. Checking for double recombinants resulted in the removal of an additional 26 genotypes. All markers were analyzed for Hardy-Weinberg equilibrium and were not found to deviate from expected allele frequencies (P<.01); a small number of markers with low heterozygosity did deviate from expected allele frequencies (P<.05), but this is likely to be due to chance. Finally, SIBMED was run with all SNP data, and microsatellite data were generated from the completion of the IMGSAC genome screen; an additional 14 low-probability genotypes were removed.

Mutation Screening

Seven missense variants were detected during mutation screening of the coding sequence, as described in table 3. Additionally, 108 noncoding/silent variants were found, as shown in table 4. Several missense variants were found in a single family; the T333I change in SSTR5 was found in an unaffected mother, and the N543K missense variant in MRTF-B was paternally transmitted to one of two affected sons. Bioinformatic analysis indicated that these variants do not occur in any known functional domain and that they were predicted to have a benign effect. However, there are few homologous sequences available for thorough conservation analysis of MRTF-B. Two missense variants were detected in BFAR. During screening, the first missense variant in exon 3, M140R, was found in six individuals and cosegregated with autism in these families, although three parents were found to be homozygous for arginine. The second missense variant in BFAR, V315I, was detected, during screening, in two individuals and cosegregated with autism in both families. Bioinformatic analysis of the missense variants in BFAR showed that they were likely to have only a benign effect on protein structure and that they do not occur at highly conserved residues. Two rare variants were found in UBN1. The A103N variant in exon 3 was detected in one family and was transmitted paternally to two affected offspring. This variant is in a region of low complexity without a strong domain identity. A variant in exon 9, Y435C, was detected in two individuals but cosegregated with autism in only one family. Although this variant was predicted to have a potentially harmful effect on protein function, no homologous sequences were available, and detailed comparative analysis was not possible.

Table 3.

Missense Mutations in Candidate Genes Identified by DHPLC

					Minor-allele Frequency in
Geneand Exona	cDNA Nucleotide	Wild-Type/Variant Allele	Amino Acid Substitution	No. ofHeterozygotesin ScreenedSamples (n=48)	Multiplex Cases (n=239)	Controls(n=192)
ABAT:
2	208	A/G	G56R	21	.457	.369
BFAR:
3	563	G/T	M140R	6	.332	.388
6	1096	A/G	V315I	2	.018	.077
MRTF-B:
5	1624	C/G	N543K	1	.024	.035
SSTR5:
1	92	C/T	T333I	1	NDb	NDb
UBN1:
3	422	A/C	A103N	1	.004	.000
9	1413	G/A	Y435C	2	.031	.027

^aSNPs genotyped in the entire IMGSAC sample and controls for all but exon 1.

^bND=no data were available.

Table 4.

Nonmissense Variants in Candidate Genes Identified by DHPLC

			Minor-Allele Frequencya in
Gene, Location,and cDNANucleotide	Major/MinorAllele	No. ofHeterozygotesin ScreenedSamples(n=48)	Cases	Controls
A2BP1:
Intron 1:
730−6	+/CTT del	2	…	…
Intron 2:
804−47	A/C	24	…	…
804−48	A/T	24	…	…
Intron 3:
852−20	+/T del	23	…	…
Exon 5:
1067	C/G	2	…	…
1098	G/A	2	…	…
Intron 5:
1182−25	A/C	4	…	…
Exon 6b:
1271	A/G	1	.021	.003
Intron 7:
1379+109	A/G	1	…	…
Intron 11:
1673−17	−/CT ins	21	…	…
Intron 13:
183−19	A/G	2	…	…
Intron 15b:
1973+3	A/G	6	.004	0
3′ UTR:
2102+141	G/T	23	…	…
ABAT:
Intron 1b:
102+49	T/C	15	.391	.2952
Intron 2:
209+11	A/G	1	…	…
210−68	C/A	1	…	…
210−5	C/T	1	…	…
Exon 4b:
350	C/T	10	…	…
Intron 8:
582+14	−/CA ins	5	…	…
582+54	A/C	2	…	…
Exon 9:
685	C/T	1	…	…
Exon 12:
995	A/C	2	…	…
Intron 13:
1313−120	−/G ins	10	…	…
Intron 14:
1423+9	C/T	22	…	…
3′ UTR:
1544+71	A/G	1	…	…
BFAR:
Exon 1:
50	T/C	1	…	…
Intron 1:
89+105	G/A	4	…	…
Exon 2:
178	T/A	1	…	…
Intron 6:
937+84	−/GTG ins	1	…	…
3′ UTR:
1507+794	A/G	1	…	…
1507+795	G/C	1	…	…
CREBBP:
Exon 3:
258	G/A	1	…	…
Exon 4:
1037	T/C	3	…	…
Intron 3b,c:
1995−8585	T/C	NDd	.005	0
Intron 4:
1174−76	+/TTTG del	1	…	…
Intron 6:
1530−31	G/T	3	…	…
Intron 7b,c:
1730+215	G/A	NDd	.349	.363
1731−58	A/C	1	…	…
Exon 11:
2152	T/C	1	…	…
Intron 11b:
2311+78	C/T	3	.077	.007
Intron 13:
2358+59	T/C	1	…	…
Intron 17b:
3449+6163	C/T	23	.414	.522
Intron 20b,c:
3898+3049	C/T	NDd	.033	.015
Intron 22b:
4040−8	T/C	3	.014	.022
Exon 23:
4097	A/C	1	…	…
Intron 26b,c:
4480+42	G/A	11	.027	.015
Intron 29b:
4937−19	C/G	3	.028	.019
Exon 31:
5160	A/C	2	…	…
Exon 32:
7411	G/A	3	…	…
3′ UTR:
7525+26	−/C ins	2	…	…
EMP2:
Exon 2:
53	G/C	7	…	…
Exon 3:
172	C/T	15	…	…
205	G/A	23	…	…
245	C/T	23	…	…
3′ UTR:
Unidentified	Unidentified	13	…	…
GRIN2A:
Intron 1:
112−33	C/T	27	…	…
Intron 5:
1318−15	G/C	1	…	…
Exon 6b:
1576	T/C	21	.424	.597
Intron 7:
1808−29	G/C	1	…	…
Intron 8:
1964+31	−/GA ins	25	…	…
Intron 10b:
2316−35	G/C	25	.447	.381
Intron 11:
2316−21	−/GA ins	1	…	…
Exon 11:
2396	G/A	2	…	…
Exon 14:
3539	G/T	2	…	…
3′ UTRb:
4705+39	C/A	2	.005	.014
4105+1213	A/G	26	.298	.510
MRTF-B:
Exon 1:
90	G/C	1	…	…
Exon 2:
379	T/G	1	…	…
Intron 3:
1475−12	G/T	2	…	…
Intron 4:
1513+21	C/T	1	…	…
3′ UTRb:
2487+67	T/C	1	…	…
2487+561	T/C	1	…	…
2487+1110	G/C	1	…	…
2487+1441	T/A	1	…	…
2487+1713	A/C	1	…	…
2487+1849	G/T	1	…	…
TBX6:
5′ UTR:
120+6	A/G	3	…	…
Intron 7:
938−4	C/T	20	…	…
937+19	−/GAT ins	23	…	…
Exon 8:
1252	A/G	1	…	…
Intron 8:
1335+428	G/C	1	…	…
UBN1:
5′ UTR:
1−17	C/G	7	…	…
Intron 1:
75−109	TA/AG	3	…	…
75−63	T/C	3	…	…
75−21	A/G	6	…	…
74+26	T/C	3	…	…
75+40	T/G	25	…	…
74+68	G/T	2	…	…
Exon 3:
147	A/G	2	…	…
Intron 3:
454+22	A/G	2	…	…
Intron 4:
546+3	G/A	4	…	…
Exon 5:
599+5	C/T	11	…	…
Intron 6:
784+8	G/A	1	…	…
Intron 9:
1296+11	−/G ins	1	…	…
1296+39	+/TT del	1	…	…
1425+62	C/T	3	…	…
Intron 14:
1915+17	A/G	15	…	…
1915+53	A/G	2	…	…
Exon 15:
2743	T/G	7	…	…
2994	T/G	6	…	…
Exon 18:
3486	C/T	3	…	…
3519+1006	C/T	12	…	…
3′ UTR:
3519+157	T/C	7	…	…
3519+1008	−/C ins	1	…	…
3519+1889	C/T	4	…	…
3519+2302	C/G	1	…	…
Exon 18:
3519+2588	A/C	13	…	…
3519+2660	C/T	26	…	…

^aNo data were available unless values are shown. For cases, n=239; for controls, n=192.

^bSNPs genotyped in the entire IMGSAC sample and controls.

^cSNPs identified through dbSNP or Sequenom's RealSNP Array.

^dND=no data were available.

One missense variant was found at a high frequency in ABAT exon 2. However, the G56R change is distant from the enzyme’s catalytic site and is unlikely to hinder protein function. Sequence homology analysis also predicted that the variant is likely to be benign. All of the missense variants detected during screening, except that in SSTR5, were typed in the entire IMGSAC sample and 192 controls. No evidence of transmission disequilibrium was found, and no significant differences in case-control allele frequencies were observed for these missense variants, as shown in table 3. None of the variants detected by DHPLC within 20 bp of the intron/exon splice site were found to potentially affect splice signal.

Association Analysis and Logistic Regression Analysis for Candidate Genes

Twenty-eight SNPs were successfully genotyped in the IMGSAC families and controls, as shown in table 5. The family data were analyzed for allele and genotype transmission disequilibrium by use of the PDT. Fisher’s exact test for count data was used to compare genotypes observed in cases and controls. Additionally, multiplex genotype data were compared with controls by use of a trend test that accounts for relatedness between affected individuals. No significant evidence of association was found with these tests either for SNPs in A2BP1, BFAR, MRTF-B, TSC2, and UBN1 or for SNPs within the region of linkage with ADHD, as described in table 5. SNPs within ABAT, GRIN2A, TSC2, and CREBBP were analyzed for LD, to determine whether any haplotype blocks were present in the data. The SNPs within ABAT and GRIN2A were in strong LD with each other (D^′=0.667–1), whereas there was variability in LD between SNPs in TSC2 and CREBBP. Because of the manner in which most of the SNPs were identified and the comparatively large inter-SNP distances, it is unlikely that any underlying haplotype-block structure is fully described by these analyses. But, given the evidence of LD between SNPs in candidate genes, haplotype transmission analysis was performed. No evidence of haplotype transmission disequilibrium was detected, using TRANSMIT, in the IMGSAC families, although an exhaustive analysis of all possible SNP combinations was not performed. There was evidence of association with SNPs genotyped in ABAT, CREBBP, and GRIN2A, and further logistic regression analysis was performed, as described below. The logistic regression analysis performed does not rely on any LD between SNPs and assesses only the risk of being affected, given a particular genotype. Haplotypes are described as sequential SNP alleles across the candidate genes, with alleles labeled as in table 5 (e.g., haplotype h12 for ABAT describes allele 1 at rs1731017, followed by allele 2 at the SNP within exon 2 of the gene).

Table 5.

Association Analysis of SNP Genotyping Data for Multiplex Families (n=239) and Singleton Families (n=91)^[Note]

		Percentage of Genotypes with Alleles									P for				Trend Testa Analysis
		1/1 in			1/2 in			2/2 in			Multiplex		Singleton			P for High-Risk Allele(s)			Fisher’s Exact Test of Sample
Genomic Region,Location, and dbSNP Number	Genotyping Success(%)	MultiplexCases	Singletons	Controls	MultiplexCases	Singletons	Controls	MultiplexCases	Singletons	Controls	Geno-PDT	Allelic PDT	Geno-PDT	Allelic PDT	No. of Sibling Pairs/Controls	1 and 2b	1c	2c	Multiplex	Singleton
A2BP1:
Exon 6	89.1	95.8	ND	99.4	4.2	ND	.6	0	ND	0			ND	ND	219/172				.0265	ND
Intron 15	54.3	0	ND	0	.8	ND	0	99.2	ND	100			ND	ND	103/113					ND
ABAT:
Intron 1:
rs1731017	76.8	37	47.3	49.7	42.5	33.8	40.8	20.4	18.9	9.5					142/147	.0061		.0113	.0089	.0112
Exon 2:	93.3	21	5.7	13.5	46.2	54	43.2	32.9	40.2	43.2					212/185		.0257		.0483
BFAR:
Exon 3	91.6	44.5	ND	37.5	43.6	ND	47.8	11.8	ND	14.7			ND	ND	209/184					ND
Exon 6	84.5	0	ND	.6	3.6	ND	4.2	96.4	ND	95.2			ND	ND	186/167					ND
CREBBP:
Intron 3:
rs3025702	97	0	ND	0	.9	ND	0	99.1	ND	100			ND	ND	244/187					ND
Intron 7:
rs130021	94.5	42.5	39.3	40.6	45.3	40.4	50.3	12.3	20.2	9.1					210/187					.03078
Intron 11:
rs130002	55.3	.6	0	0	2.5	1.1	1.4	96.9	98.9	98.6					127/74
Intron 17:
rs886528	89.2	17.1	14.6	27.2	45.5	48.3	41	37.4	37.1	31.8					166/173			.0103
Intron 20:
rs130025	90.5	0	0	0	6.5	2.3	2.9	93.5	97.7	97.1			.0196	.0196	219/172
Intron 22:
rs3025684	98.1	97.3	100	95.7	2.7	0	4.3	0	0	0	.036	.0074			225/186
Intron 26:
rs129967	92.8	0	0	0	5.4	0	2.9	94.6	100	97.1					227/175
Intron 29:
rs130008	94.9	0	1.1	0	5.6	4.5	3.7	94.4	94.3	96.3	.001	.0024			221/187
GRIN2A:
Exon 6	93.7	18.1	47.7	35.6	75.8	43.2	58.8	6.2	9.1	5.6					239/177	.0013		<.0001	.0003	.0486
Intron 10	96.3	20	54.7	14.5	75.7	37.2	81.2	4.3	8.1	4.3	.013				242/186			.0337		2.32×10−12
3′ UTR:	96.3	99.1	ND	97.3	.9	ND	2.7	0	ND	0			ND	ND	240/186					ND
rs1014531	75.5	8.9	19.7	26	78.5	49.3	50.4	12.6	31	23.6					218/127		.0214	<.0001	2.9×10-7	.0439
MRTF-B:
Exon 5	75.9	0	ND	0	4.7	ND	6.8	95.3	ND	93.2			ND	ND	155/177					ND
TSC2:
5′ UTR:
rs2516740	93.5	62.6	ND	63.5	31.3	ND	30.7	6.1	ND	5.8			ND	ND	219/189					ND
Intron 10:
rs2074968	93.7	17.4	ND	15	48.2	ND	52.8	34.4	ND	32.2			ND	ND	231/180					ND
3′ UTR:
rs1051771	82.1	0	ND	0	2.1	ND	6.8	97.9	ND	93.2		.0143	ND	ND	169/162	.0023	.0022		.0346	ND
UBN1:
Exon 3	83.8	100	ND	0	0	ND	0	0	ND	0			ND	ND	183/161					ND
Exon 9	96.8	93.9	ND	94.7	5.7	ND	5.3	.4	ND	0			ND	ND	233/187					ND
ADHD linkage peak:
14–15 Mbd:
rs153783	92.1	54.6	ND	51.6	37.7	ND	37.4	7.7	ND	11.1			ND	ND	202/190					ND
rs1125972	96.8	88.7	ND	86.6	10.4	ND	12.9	.9	ND	.5			ND	ND	240/186					ND
rs1065838	93.7	8.4	ND	8.4	40.2	ND	41.6	51.4	ND	50			ND	ND	201/190					ND
rs1107143	96.1	38.2	ND	39.2	46.5	ND	46.2	15.4	ND	14.5			ND	ND	237/186					ND

Note.— Results shown only when P<.05. ND=no data were available.

^aTrend test performed on multiplex sample.

^bAdditive allele effect.

^cDominant allele effect.

^dBelow linkage peak identified by Smalley et al. (2002).

ABAT

Both SNPs within ABAT showed some evidence of association in the multiplex sample, with the trend-test and case-control analysis. The strongest evidence of association was from the SNP within ABAT intron 1, which, when analyzed using the trend test, indicated that allele 2 was responsible for increasing the risk of autism (P<.001). Fisher's exact test for count data gave a P value of .0089, which indicates a significant difference in genotype distribution between cases and controls, as shown in table 5. This result was supported by the analysis of the same SNP in the singleton sample with use of Fisher's exact test (P=.0112), as shown in table 5. Logistic regression of SNPs within ABAT did not indicate that a particular haplotype significantly increased the risk of autism, although there is an overall difference in haplotype distribution between cases, from the multiplex and singleton samples, and controls (χ²=20.37; P=1.422×10^-4 for multiplex cases; χ²=6.18; P=.1031 for singleton cases), as shown in table 6. An additional SNP was also genotyped in exon 4 of ABAT but was excluded from the final analysis, despite supporting the results described above, because of a low genotyping success rate (65.9% in the multiplex sample; 73.9% in the singleton sample).

Table 6.

Haplotype Logistic Regression Analysis

	No. of Multiplex
Gene andHaplotype	Casesa	Controlsb	Log RelativeRisk	SE	P (>|z|)	χ2	P	df
ABAT:						20.37	1.4226×10−4	3
h11c	18	7		.8226	.1246
h12	162	85	−.1536	.4397	.0727
h21	225	219	−.726	.4209	.0845
h22	15	3	.7212	.7466	.334
CREBBP:						16.52	.0355	8
h12122222c	7	3		1.3986	.242
h21121212	5	0	14.6899	649.2732	.982
h21221212	4	5	−.9564	.9743	.326
h22112222	4	2	−.1661	1.1172	.882
h22121212	71	85	−.9602	.7079	.175
h22121222	28	32	−.9925	.7698	.197
h22122212	118	93	−.7084	.7217	.326
h22122222	51	38	−.662	.732	.366
Otherd	10	2	.6144	1.0274	.558
GRIN2A:						62.05	1.7232×10-11	6
h1111c	34	42		.4716	.0103e
h1112	5	3	1.4334	.8411	.0883
h1122	168	73	.8251	.3143	.0086f
h2111	217	85	1.5479	.306	4.21×10-7f
h2121	2	20	−2.5906	.791	.001b
h2122	13	7	.4442	.5331	.4047
Otherd	5	2	1.3741	.9345	.1414
TSC2:						1.63	.8033	4
h112c	135	103		.22961	.22
h122	178	142	−.0364	.1733	.834
h212	13	6	.4427	.5116	.387
h222	45	33	.012	.2326	.959
Otherd	3	4	−.5625	.783	.472
	No. of Singleton
	Casesa	Controlsb	Log RelativeRisk	SE	P (>|z|)	χ2	P	df
ABAT:						6.18	.1031	3
h11c	2	7		1.6134	.2746
h12	55	106	.4878	.8411	.562
h21	120	244	.5369	.8129	.509
h22	7	3	2.122	1.0626	.0458
CREBBP:						9.77	.1346	6
h12122222c	4	2		1.5728	.422
h21221212	0	5	−16.7436	1,072.0759	.988
h22121212	35	86	−1.0608	.8039	.187
h22121222	12	33	−1.2422	.8471	.143
h22122212	49	96	−1.013	.8081	.21
h22122222	30	40	−.6101	.8348	.465
Otherd	2	6	−1.4596	1.1374	.199
GRIN2A:						37.05	5.8528×10-7	5
h1111c	22	42		.5113	.138
h1122	38	72	−.2564	.3701	.4885
h2111	72	83	.7756	.3248	.0169f
h2121	0	20	−17.5343	877.517	.9841
h2122	4	7	−.4984	.6939	.4726
Otherd	0	4	−16.6823	1,972.8826	.9933

^an=239.

^bn=192.

^cReferent haplotype.

^dHaplotypes combined when <5.

^eHaplotype has a protective effect P<.05.

^fHaplotype has a deleterious effect P<.05.

CREBBP

In the multiplex sample, association was found for CREBBP with two SNPs within introns 22 and 29, with use of the geno-PDT (P=.0363 and P=.001, respectively) and with use of the allelic PDT (P=.0074 and P=.0024, respectively). However, both these SNPs had a minor-allele frequency <6%, and the number of transmissions to affected individuals was small. There was some additional evidence of association from the trend test for the SNP in intron 17 (P=.0103), but these results were not replicated in the singleton sample. Conversely, SNPs within introns 7 and 20 showed evidence of association in the singleton sample (P=.03078 and P=.0196, respectively) but not in the multiplex sample. It should be noted that the minor-allele frequency was also low for the SNP in intron 20, and the P value was of marginal significance for the SNP in intron 7; it is likely these represent chance findings. These results are shown in table 5. There was no evidence from the logistic regression analysis to support the presence of a risk haplotype across CREBBP, as shown in table 6.

GRIN2A

One SNP in GRIN2A intron 10 showed some evidence of transmission disequilibrium with the geno-PDT (P=.013) and the trend test (P=.0337). This was supported by SNPs within exon 6 and the 3′ UTR that also showed evidence of association with the trend test and case-control analysis, as shown in table 5. Although the singleton analysis does appear to support these results, closer inspection shows that the transmission of different genotypes is responsible for the significant result for the SNP within exon 6—genotype 1/1 in the singleton sample compared with genotype 1/2 in the multiplex sample, as shown in table 5. A significant difference in the distribution of haplotypes, between cases and controls, was found using logistic regression analysis for both the multiplex and singleton samples (χ²=62.05; P=1.7232×10^-11) compared with controls (χ²=37.05; P=5.8528×10^-7). Haplotype h2111—genotypes x=(2,1,1,1), y=(1,1,1,1)—significantly increased the chances of being affected (P=4.21×10^-7), as did haplotype h1122—x=(1,1,2,2),y=(1,1,1,1)—although not as strongly (P=.0086), as shown in table 6. The effect of the h2111 haplotype was replicated in the singleton sample but at a less significant level (P=.0169). Additionally, haplotype h2121—genotypes x=(1,2,1,2), y=(1,1,1,1)—is found at a rate of 20/232 in controls, compared with 2/580 in multiplex cases, and appears to decrease the chances of being affected (P=.001).

Statistical significance of association and logistic regression analysis

A simple Bonferroni correction, to reduce the chance of type 1 error to 0.05 across all tests, gives a significance threshold of .00714, if we consider that seven tests were performed: geno-PDT, allelic PDT, trend test for high-risk alleles 1 and 2, trend test for high-risk allele 1, and trend test for high-risk allele 2, Fisher’s exact test, and logistic regression analysis. To reduce type 1 error to 0.001 across all tests, a Bonferroni significance threshold of .000142 is required. This represents a highly conservative correction, given that the PDTs and trend tests are correlated, although a number of the results in table 5 do exceed these significance levels.

Permutation tests, which provide more-accurate significance thresholds, were performed (see the “Material and Methods” section). For the single-point analysis with Fisher’s exact test across 28 SNPs, the minimum P value of 2.9×10^-7 for rs1014531 in GRIN2A in the multiplex case-control analysis was smaller than any obtained in 10,000 permutations, for an empirical P value of <10^-4. Similarly, for the joint logistic haplotype regression of all four genes, the observed χ² statistic of 102.9 with 34 df (nominal P=7.4×10^-9) was greater than for all 10,000 permutations; that is, an empirical P value of <10^-4.

Separate haplotype logistic analysis of the four genes indicates that GRIN2A has by far the most significant contribution. A test comparing the fit of GRIN2A alone with that of all four genes gave a χ² value of 47.44 with 25 df (P<.0043), which suggests that, whereas the GRIN2A haplotypes are the principal contributors, there may be some effect associated with the ABAT haplotypes.

Discussion

Seven missense variants were detected during the screening of 10 positional candidate genes on chromosome 16p11-13 for autism. Although a number of these variants cosegregated with autism, both their presence at similar frequencies in controls and the lack of evidence of transmission disequilibrium suggest that they are unlikely to make a significant contribution to the etiology of autism.

A large number of noncoding and silent mutations were detected, but these are not predicted to alter splice-site recognition sequences. However, bioinformatic analysis alone may be insufficient to detect the effects of intronic and silent variants on splicing and transcriptional efficiency, as recently reported for CFTR and SLC22A4 (Pagani et al. 2003; Tokuhiro et al. 2003), and it is possible that novel regulatory elements could be disrupted. A lack of missense mutations in coding sequence does not preclude the presence of harmful noncoding variants within a candidate gene. In fact, for a multigenic disorder such as autism, more subtle changes in gene regulation and expression are anticipated. This has been found for PAD14, a gene associated with rheumatoid arthritis (Suzuki et al. 2003). Suzuki et al. (2003) failed to detect any missense mutations in 48 affected individuals but found a functional SNP haplotype affecting mRNA stability.

The 28 SNPs genotyped in the IMGSAC families are unlikely to provide sufficient coverage of the genomic regions to conclusively exclude genes by association mapping. This was particularly true for genes such as A2BP1, which has a genomic region of 1,690 kb, from which only two SNPs with low heterozygosities were genotyped. Although a number of the SNPs genotyped in this study had low heterozygosities, it has been argued that they are useful for detecting association between disease variants with more-recent origins, despite reduced informativeness (Pritchard and Cox 2002). Additionally, even though most of the SNPs were successfully genotyped in 85% of individuals, the SNP-genotyping success rates had a range of 54.3%–98.1%. Although this failure rate could reduce confidence in the data used for analysis, Sequenom traces were manually inspected for markers with low success rates, and low-probability genotypes were removed when identified by software programs.

The association analysis performed in this study attempts to extract the maximum information from the data available and is, in some respects, exploratory. Family-based and case-control association analyses were performed for single-marker and multimarker haplotypes. Although this requires multiple testing of the data, it has been suggested that different approaches have varying power, depending on disease and marker-allele frequency. A family-based analysis benefits when parental genotypes are used as controls but may yield little informative data if marker heterozygosity is low and parental data is missing. Conversely, a case-control analysis may have more power at varying allele frequencies for the number of individuals genotyped but could provide an artifactual result because of population-specific differences between cases and controls (Risch and Teng 1998). It has also been suggested that the added information from haplotype analysis of SNPs has more power than single-marker analysis (Zaykin et al. 2002). A number of procedures have been suggested for correcting for multiple testing, including Bonferroni correction, but there is as yet little consensus as to an “ideal statistical framework” either for reporting P values or for analysis of SNP data (Roeder et al. 2005). We have presented raw P values uncorrected for multiple testing, but the significance of reported P values should be viewed with caution, given the number of statistical tests performed.

Empirical significance levels were determined by permutation of affection status for statistical tests involving cases and controls. These levels suggest that the results for the SNPs within GRIN2A for rs1014531 (P=2.9×10^-7) in the multiplex case-control analysis and for intron 10 (P=2.32×10^-12) in the singleton case-control analysis are significant results. Additionally, the logistic regression analysis suggests GRIN2A haplotypes are significant contributors to genetic risk for autism. Determining empirical P values for the family-based tests for association, such as the PDT, requires correlations between transmitted alleles within pedigrees to be preserved, and “[i]t is not clear whether a simple permutation procedure can be developed for this purpose in general pedigrees” (Martin et al. 2000, p. 153). Additionally, simple permutation of affection status is inappropriate for the trend test, because the correlation between alleles in related cases must be preserved (Slager and Schaid 2001). Interpreting the results from these statistical tests is challenging, but, on balance, the strongest evidence of association with autism comes principally from GRIN2A, with ABAT accounting for the remaining association.

Both SNPs genotyped in ABAT have genotypes associated with autism in the case-control analysis of the multiplex sample, with use of Fisher's exact test for count data (Fisher's exact test provides more power than the PDT, given the sample sizes). The result for the intron 1 SNP was also replicated in the singleton sample. The result is further supported by the logistic regression analysis, which indicates a difference in haplotype distribution between cases and controls. ABAT is a good positional functional candidate for autism, especially given its involvement in the catabolism of GABA, which is known to be involved in neurological disease (Wong et al. 2003). Other genes involved in this pathway have been studied elsewhere, including the GABA receptor subunit β3 (GABRB3) on 15q11-13, following reports of chromosomal abnormalities in autistic individuals. A number of studies have been published on the role of this pathway in autism (reviewed by Veenstra-VanderWeele and Cook [2004]).

Screening the coding sequence of ABAT revealed one high-frequency missense mutation, but this does not occur in a highly conserved position and was less strongly associated with autism, compared with the SNP within intron 1. It seems likely that any variants influencing the expression of ABAT in individuals with autism are located outside of the coding sequence within regulatory regions.

No missense variants were detected by DHPLC in GRIN2A. However, SNPs within GRIN2A gave strong evidence of association in the case-control analysis of the multiplex and singleton samples. There was also some evidence of transmission disequilibrium with the SNP in intron 10 in PDT results. Additionally, the logistic regression analysis found a significant difference in haplotype distribution between cases and controls. GRIN2A is an interesting candidate gene for autism, given its role in learning and memory and the evidence from mouse models of a complex behavioral phenotype (Bannerman et al. 2004). There is also evidence, from postmortem analysis, of expression differences in the glutamate system between cerebella of subjects with autism and of controls (Purcell et al. 2001). Variants within GRIN2A may also confer increased risk for ADHD, which indicates involvement in behavioral disorders (Turic et al. 2004).

The candidate-gene analysis presented here is substantial but not completely exhaustive. Candidate genes were screened for coding variants, and association analysis was performed, but additional SNPs with inter-SNP distances of 5–10 kb may be required to conclusively discount these genes having a role in autism susceptibility (Kruglyak 1999). Increasing the density of SNPs across candidate genes, on the basis of known patterns of LD, to capture the maximum information about haplotypic diversity is now increasingly feasible, as data become available from the International HapMap Project. Prior knowledge of the LD structure of a region will allow a more-intelligent and -efficient approach to association mapping in studies of candidate genes and perhaps will lead to less-tentative statistical analysis and more-definitive conclusions (Zondervan and Cardon 2004). A combination of strategies that include association mapping and complementary statistical and bioinformatic analysis will be needed to determine the significance of the results for GRIN2A and ABAT, as well as replication in independent samples. The design of subsequent studies should also include careful consideration of the control samples used for allele- and genotype-frequency comparisons, in terms of their ethnicity, sex, and independence. Chromosome 16p has been identified by a number of genome screens for autism and is likely to contain an autism-susceptibility variant. Although the 16p locus has been less consistently replicated than the chromosome 7 locus, it remains an important area for future research (Lamb et al. 2002). However, as with all multifactorial disorders, progressing from a broad linkage peak to identifiable functional genetic variants remains a significant challenge.