There has been much interest and work investigating the genetic basis of autism [5]. Twin studies have shown that autism is a strongly inherited disorder [2,6,7], as monozygotic twins are concordant for this syndrome substantially more frequently than are dizygotic twins. For reasons that are not yet understood, autism affects boys about four times more often than girls. Non-genetic etiological factors are under careful consideration [8], given the controversy over changing estimates of the incidence of autism over the past decades [9].

Recently, we hypothesized that some forms of autism may be due to a disproportionate high level of excitation (or disproportionately weak inhibition) in neural circuits that mediate language and social behaviors [10]; related models have also been postulated [11,12]. A "noisy" (hyperexcitable, poorly functionally differentiated) cortex is inherently unstable, and susceptible to epilepsy, a common malady of autistic people [13]. Consistent with this model, a subset of patients with a mutation of the X-linked homeobox gene ARX have autistic features [14,15]. Analysis of the expression and function of murine Arx has revealed that this transcription factor controls proliferation of cortical progenitor cells and it regulates the development of the basal ganglia and the tangentially-migrating GABAergic neurons, derived from this region, that become cortical local circuit neurons [16]. Furthermore, Arx is expressed in mature cortical local circuit neurons [16,17]. Arx expression in GABAergic neurons in the basal ganglia and cortex could underlie the epilepsy and movement disorders in humans bearing Arx mutations.

The highly conserved Dlx1, 2, 5 & 6 homeobox transcription factors also control the development of the basal ganglia and cortical local circuit neurons [18-23]. This gene family regulates the expression of glutamic acid decarboxylase [24,25] (and Cobos and Rubenstein, unpublished). Furthermore, these genes control the expression of Arx in cells derived from basal ganglia progenitor domains [17]. Finally, mice lacking Dlx1 have defects in subsets of cortical local circuit neurons that lead to their apoptosis and subsequent onset of epilepsy [26]. Therefore, murine Dlx genes have a central role in controlling the development and function of forebrain GABAergic (inhibitory) neurons. The human DLX genes are organized as bigene clusters on chromosomes 2q31.1 (DLX1/2) and 7q21.3 (DLX5/6) and their expression is controlled by intra- and extragenic enhancers [27-30] These genes are close to loci that have been linked with autism (2q32, and 7q22-31) [31-34].

In a genome scan using 152 affected sibling pairs, a multipoint maximum lod score (MLS) of 3.74 was reported near D2S2188 [31]. This region on chromosome 2 has also been reported to be linked to autism in one study of 51 multiplex families [35], while other studies show linkage slightly more telomeric to this region [36,37]. Chromosome 7q, also an important region in linkage studies of autism [38], contains DLX5 and DLX6. The International Molecular Genetic Study of Autism Consortium (IMGSAC) reported a single locus MLS of 2.93 in 83 sibling pairs at D7S477, some 4.1 Mb telomeric to the DLX5/6 cluster on 7q22 [31]. When 152 sibling pairs were analyzed, a multipoint MLS of 3.20 was calculated. An earlier IMGSAC study reported linkage ~35 Mb telomeric to the DLX5/6 region [39].

The sample consisted of 161 autism probands and 58 non-autistic siblings obtained from AGRE, for a total of 2.08 Mb of sequence. The DNA variants discovered are summarized in Table ​Table1.1. The relative location of each variant within the DLX1/2 and DLX5/6 clusters is depicted in Figure ​Figure1.1. Variants are referred to by our simple name (e.g., DLX2 SNP-1) for the purpose of clarity within this manuscript, while technically accurate nomenclature [42,43] or dbSNP rs#, if available, are also referenced in Table ​Table1.1. Allele frequencies for these variants are detailed in Additional File 2. Thirty-three variants were observed, consisting of 31 single nucleotide polymorphisms (SNPs) and two insertion/deletion polymorphisms. Coding region variants consisted of three synonymous SNPs (one in DLX2, two in DLX5) and four non-synonymous SNPs (two in DLX2, two in DLX5). Additionally, an in-frame insertion of AGC in exon 1 of DLX2 leads to the insertion of an additional serine after serine 46, lengthening a poly-serine stretch from six to seven residues. DLX1 and DLX6 did not have any coding region variants in this study. One SNP each in DLX1 and DLX2, as well as two SNPs in DLX6, have been previously deposited in dbSNP. Interestingly, DNA variants were ~2.3 fold less common in the four DLX1/2 and DLX5/6 intergenic enhancers and DLX1/2 URE2. For 33 of 34 variants, for which corresponding chimpanzee genomic sequence was available, the major human allele matched the chimpanzee allele. For the single exception, in the brain specific DLX1/2 intergenic enhancer, the minor allele seen in our sample matches the allele seen in chimp, mouse, rat, and chicken. We observed none of three SNPs in DLX1 and all three SNPs in DLX2 observed in a recent resequencing effort in 48 autism probands [32].

We focused on the non-synonymous SNPs and the DLX2 serine insertion for further analysis given the plausibility of the functional significance of these amino acid changes. The DLX2 InDel-1 was observed on 3 autism chromosomes. DLX2 SNPs-2 and 6 and DLX5 SNP-6 were each seen on individual chromosomes, while DLX5 SNP-7 was observed on 4 chromosomes. These observations result in allele frequencies of 0.009, 0.003, 0.003, 0.003, and 0.012, respectively, in our autism sample. We compared the allele frequencies for each of these five observed DLX2 and DLX5 variants in the AGRE sample with the frequency of these DLX variants in a different population. For this purpose we sequenced 188 DNAs from the Coriell DNA Polymorphism Discovery Resource (PDR). This population is not phenotypically defined, but does represent a diverse array of populations for the detection of uncommon DNA polymorphisms. An additional two SNPs (SNPs 3 and 4) in DLX5 were observed in the PDR sample as part of this effort, but not in the autistic sample or the non-autistic siblings (Additional File 2). Additionally, we genotyped the four non-synonymous SNPs in 95 Caucasian and 95 African-American samples, also obtained from the Coriell Institute. No heterozygous genotypes were seen for SNPs 2 and 6 of DLX2. For SNPs 6 and 7 of DLX5, single heterozygous samples were seen, with both occurring in the Caucasian sample. Inspection of this data indicates that allele frequencies for the 4 protein-changing variants did not differ between autistic probands and their siblings, nor between autistic probands and the PDR and human variation panels.

In general, the variants were uncommon. Fifteen were observed only once in the autism probands. Ten were seen in 2–4 autistic individuals. Four of the variants were more common in the autism probands, with minor allele frequencies ranging between 3.4% and 46.5%.

Additional File 3 depicts the genotypes of non-synonymous DLX2 and DLX5 variants and the DLX5/6 intergenic enhancer variant in AGRE pedigrees in which they were detected. The segregation patterns for these variants do not clearly support the hypothesis that they are autism susceptibility variants. DLX2 InDel-1 and DLX2 SNPs-2 and DLX5 SNP-6 are found in affected and unaffected siblings within the same pedigrees. The minor variant for DLX2 SNP-6, which was also seen in the Caucasian panel, is introduced to one of two affected cousins by a father who is not related by blood to the second cousin. SNP-7 in DLX5, which predicts a substitution of Arginine for Serine at position 234, was seen in 3 pedigrees (Additional File 3, figure 1e), but also in the PDR and Caucasian panels. Four autism probands and one sibling diagnosed as broad-spectrum autism were heterozygous for this SNP. Two normal siblings and one autistic proband were homozygous for the wild type allele. It is transmitted to both affected siblings in two families, and to an affected sibling and unaffected sibling in a third. The rich phenotypic data collected by AGRE provided an opportunity to assess any relationship between phenotype and the variants described above. The average non-verbal IQ for all affected persons in the pedigrees in which the four non-synonymous SNPs occurred was 96.5. The two lowest scores, 68 and 85, occurred in two sibs (one with autism, the other with "not quite autism") who both were heterozygous for the DLX2 Serine insertion/deletion and both had early trigonocephaly which normalized. The most notable observation is with family AU0028, in which a heterozygous mother transmits the variant DLX5 SNP-7 to one of her two affected children. Both children have medical histories positive for generalized tonic-clonic seizures, although only the heterozygous child, required medication treatment. Both children also exhibited obsessive-compulsive disroder and attention deficity hyperactivity disorder (ADHD) symptoms, although only the heterozygous child received an ADHD diagnosis. The child heterozygous for the variant also exhibited failure to thrive and an early gait abnormality, which resolved. Given the discordance between genotypes of the affected siblings, it is hard to argue for the role of the DLX5 variant in these specific phenotypes.

Given the location of DLX1/2 and Dlx5/6 in relation to autism linkage intervals, other groups have examined whether DNA variants in these genes are significantly associated with autism. The IMGSAC consortium conducted a sequence survey of DLX1 and DLX2 in 48 autistic probands [32], which identified three variants in DLX1 that were not detected in our much larger sample and three of the DLX2 variants seen in our sample. One explanation for the non-overlap in findings is the low allele frequency of most variants found in both studies. For example, our SNP-1 in DLX1 was in a region also sequenced by the IMGSAC group, but our allele frequency was 0.2%, making it highly unlikely that it would be detected by assaying 96 chromosomes [32]. The high frequency SNPs-1 and 4 occurred in both samples. Another reason may be the origin of samples. The AGRE samples are predominantly of U.S. origin [41], while the IMGSAC samples are from a variety of geographically diverse countries [39], raising the possibility of population specific variants. This is particularly the case with rare variants, which are less likely to be shared across populations. Finally, differences in coverage of the gene (coding and non-coding regions), depth of coverage (i.e., sample size), or variant detection technology (e.g., direct sequencing in this report, versus variant detection with denaturing high performance liquid chromatography followed by direct sequencing by Bacchelli et al.) may explain the discrepancy in variants described. For example, our DLX1 SNPs-2 and 3 lie outside of the region assayed by IMGSAC.

In another recent study in 99 AGRE pedigrees and 308 other pedigrees, two SNPs in DLX2 were investigated [44]. One of these, rs2228184, corresponding to our DLX2 SNP-4, a synonymous coding sequence variant, showed marginal association to autism. This common variant was equally common in autistic probands and their unaffected siblings in our study (Additional File 2). The second SNP in the study by Rabionet et al., which was not associated with autism, occurs outside of the region we sequenced. These published data and our own do not provide support for the possibility that common variation in the DLX loci is associated with autism.

Another study focusing on DLX6 reported the existence of a CAG repeat in exon 1 of DLX6 after assaying 90 Caucasian samples [45]. Although we sequenced the same region, we did not detect this variation. In any case, the uncommon nature of the DLX variants reported here, even in aggregate, are unlikely to provide the basis for any linkage signal in the DLX gene clusters on chromosomes 2 and 7.

A striking observation in our data was the prominent lack of sequence diversity in the five non-coding regions we investigated. In the 4,000 bp encompassing the four intergenic enhancers and DLX1/2 upstream regulatory element, we found only seven variants. However, given the sequence conservation of these functional elements [28], this result is not surprising. Indeed, three of the four intergenic elements are included in 481 genomic segments greater than 200 bp with 100% conservation of human sequence with mouse and rat [46]. In other words, 0.6% of known ultraconserved sequences can be found in the 0.001% of the genome representing the DLX1/2 and DLX5/6 clusters. These deeply conserved sequences that were not exonic were significantly enriched near genes involved with transcriptional regulation, and in particular, those with Homeobox domains (p < 10^-14) [46].

Of the variants identified, eight occurred in the coding regions of DLX2 and DLX5. The four variants that change the identity of an amino acid are non-conservative modifications: DLX2 SNP-2 (Glutamic acid to Lysine), DLX2 SNP-6 (Alanine to Threonine), DLX5 SNP-6 (Serine to Proline), and DLX5 SNP-7 (Serine to Arginine). In both DLX2 and DLX5, amino acid substitutions lie in conserved regions of the proteins. DLX2 SNP-2 is just N-terminal of the homeodomain in a region conserved among the human DLX2,3,5 subgroup. This DLX2 residue is conserved between human and chimp, dog, rat, and mouse, and fugu, while chicken, African clawed frog, and zebrafish contain the conservative Aspartic acid at the same position. The amino acids changed by DLX2 SNP-6, DLX5 SNP-6 and DLX5 SNP-7 are adjacent to a Proline-rich domain C-terminal to the homeodomain. The amino acids substituted by DLX5 SNP-6 and SNP-7 are invariant in 5 mammal species, chicken (except for the Serine changed to Arginine by SNP-7), and frog. DLX2 InDel-1 leads to the insertion of seventh Serine residue into a six residue polyserine tract within the conserved DLX2,3,5 DllA domain [21,47]. The functional significance of such a change is unknown.

The functional significance of the three synonymous SNPs (DLX2 SNP-4, DLX5 SNP-1 and DLX5 SNP-2) is uncertain, but cannot be summarily dismissed. For example, such "silent" variants can alter binding sites (exonic splice enhancers, ESE) for proteins involved in RNA splicing [48]. Using ESEfinder, a web-based application designed to analyze exonic sequences to identify potential ESEs responsive to the human SR proteins [49], each of these three variants alter the predicted strength or presence of recognition sites for one or more of several highly conserved and structurally related splicing factors termed Serine/Arginine-rich (SR) proteins (data not shown). For instance, the C to T substitution for DLX5 SNP-2 synonymous change abolishes binding sites for two of three SR proteins located in the region surrounding the SNP. The functional significance of this in silico observation is unknown, but highlights the potential importance of DNA variation that does not necessarily alter the primary structure or proteins. As is always the case with the analysis of rare variants, until the functionality of these variants is demonstrated, either through statistical differences in allele frequencies at the population level or through direct functional studies, these variants should not be considered disease mutations.

While the identification of variants that generate non-conserved amino acid changes in DLX2 and DLX5 in autistic people suggests that the DLX genes could contribute to autism susceptibility, there are limitations to our study. First the identified DLX variants identified here are rare variants that could be expected to naturally occur across the human population. While this is a possibility given the low likelihood that any random gene is an autism susceptibility locus [50], we believe that biological and genetic linkage data elevate the a priori probability that the DLX genes analyzed here may be autism genes. Furthermore, the nature of the amino acid changes suggests that they could alter the function of the DLX proteins, although direct demonstration of this is currently lacking. A second limitation is that the rarity of the DLX variants precludes the possibility that they account for a significant portion of the genetic susceptibility to autism. A further weakness is that our study did not allow the large-scale population-based case control design that would allow a better estimation of the probability that these variants contribute to causing autism. A third limitation involves the use of unaffected siblings in our mutation screen. These siblings were rigorously phenotyped, but not having clinical autism does not rule out the possibility that they have milder traits representing aspects of the autism phenotype. Thus, it is possible that variants shared by affected and "unaffected" siblings may be functionally significant. In terms of co-occurring medical conditions, we found no evidence for such in five families segregating non-synonymous amino acid variants. A fourth limitation is that the population sample we used to compare the allele frequencies of the autism variants with a "normal" population was not optimal. We used the DNA Polymorphism Discovery Resource sample, which is designed to mirror the sequence diversity of the human population. While this sample allowed us to examine a cross section of global genetic diversity, the use of this sample introduces two major limitations for the interpretation of our data. One is the lack of knowledge of phenotypes in the PDR and human variation samples raises the possibility that we may falsely conclude that a variant seen both in autistic probands and the Coriell samples is not involved with autism when in fact the Coriell samples with the variant unbeknownst to us may have autism or a related phenotype. The second involves the relative enrichment of the PDR sample for non-Caucasian samples when compared to our autism sample, which is predominantly Caucasian. This under-sampling of Caucasians (and consequent lack of power to detect rare variants) in the PDR sample may cause us to falsely attribute a rare variant as autism-related, when it may in reality be merely Caucasian-specific. This may indeed be the case for the two DLX5 non-synonymous variants, which were not seen in the PDR sample, but were seen in the Caucasian panel. Thus by including the Caucasian and African-American human variation panels, we have observed that the variants may be Caucasian-specific, albeit at low allele frequencies.

Despite the caveats described above, we suggest that the non-synonymous DLX SNPs could contribute to autism susceptibility for several reasons. Mice lacking DLX1 have epilepsy [26], a common feature in autistic patients. Heterozygosity of transcription factor mutations is well-known to cause human disease [51]. In mice the dosage of the DLX genes is known to be important in controlling the differentiation of forebrain GABAergic neurons and morphogenesis of craniofacial structures, including the middle and inner ear. Indeed, heterozygosity of DLX2 alters morphogenesis of the skull (Depew and Rubenstein, unpublished), although it is uncertain whether heterozygosity of a DLX gene alters brain function. Given recent evidence that the DLX5 locus is partially imprinted in humans [52], heterozygosity for DLX5 alleles could have profound ramifications. In our own pedigrees, we note that two of three pedigrees segregating the DLX5 non-synonymous SNP-7 show maternal transmission, while the single DLX5 SNP-6 pedigree showed paternal transmission (Additional File 3). Increases in Dlx5 expression have been found in mice lacking MECP2 (the Rett Syndrome gene), which are associated with alterations in long-range chromatin organization [53]. Therefore, several recent findings are increasing the likelihood that changes in DLX function/expression are involved in neuropsychiatric disorders.