Review

Attention-deficit/hyperactivity disorder (ADHD) is a common neuropsychiatric disorder. The prevalence of the disorder in children is estimated at 5.3%, estimates for adults vary from 1 to 4% (Faraone et al. 2000b; Kessler et al. 2006; Kooij et al. 2005; Polanczyk and Rohde 2007). A large number of studies, including family, adoption and twin studies suggest that ADHD is among the most heritable of neuropsychiatric disorders: ADHD was found to segregate in families with different studies finding risk increases of two to eightfold for parents and siblings of affected children (reviewed by Faraone and Doyle 2001). Adoption studies showed that the biological parents of hyperactive children carry a higher risk for ADHD compared to adoptive parents (Cantwell 1975; Morrison and Stewart 1973), and that first-degree adoptive relatives of probands with ADHD have a lower disease risk than the first-degree biological relatives of non-adopted ADHD probands (Sprich et al. 2000). More than 20 twin studies have been published in the last 32 years, most reporting estimates of heritability for ADHD between 60 and nearly 100%, with a mean of 76% (Faraone et al. 2005; Haberstick et al. 2008; Heiser et al. 2006; McLoughlin et al. 2007; Schultz et al. 2006).

The genetic architecture of ADHD is not currently clear. A polygenic transmission model seems likely (Faraone and Doyle 2001; Morrison and Stewart 1974), although some authors have suggested more dominant gene effects (Acosta et al. 2004; Deutsch et al. 1990; Maher et al. 1999; Morrison and Stewart 1974). Since it is also known that environmental factors play a role in ADHD etiology (for review see Banerjee et al. 2007), the disorder seems best described as being of multifactorial origin.

Identifying the individual genes contributing to the genetic variance of ADHD has proven difficult. Two approaches have been available for this type of research until recently: hypothesis-driven candidate gene association studies and hypothesis-free genome-wide linkage analysis. Although hundreds of candidate gene studies have been reported (see for example Bobb et al. 2005; Khan and Faraone 2006), only a handful of associations have been replicated across studies, although none of these achieved genome-wide significance (e.g. Faraone et al. 2005; Li et al. 2006). These studies have mainly concentrated on genes involved in neurotransmission, particularly in the catecholaminergic systems involved in the response to ADHD medications. Clearly, the hypothesis-driven approach has been limited by our restricted knowledge regarding mechanisms involved in ADHD pathogenesis. Genome-wide, hypothesis-free linkage analysis has been performed in ADHD using qualitative and quantitative definitions of the disease phenotype (Arcos-Burgos et al. 2004; Asherson et al. 2008; Bakker et al. 2003; Hebebrand et al. 2006; Ogdie et al. 2003; Ogdie et al. 2004; Ogdie et al. 2006; Romanos et al. 2008; Zhou et al. 2008a), and recently also using ADHD endophenotypes at the level of neuropsychological functioning (Doyle et al. 2008; Rommelse et al. 2008c). Although linkage designs have been highly successful in the identification of genes for monogenic diseases, this has not been the case for polygenic or multifactorial disorders. The linkage studies in ADHD have identified a number of genetic loci (potentially) harboring genes for ADHD, but very little overlap was observed between studies, so far. A recent meta-analysis of ADHD linkage studies only confirmed one locus, on chromosome 16 (Zhou et al. 2008c).

Generally, the results of both candidate gene-based association studies and genome-wide linkage studies have not identified major genes for ADHD. In candidate gene studies, individual gene variants have only shown small effects, rarely reaching an odds ratio of 1.3 (e.g. Faraone et al. 2005). This strongly limits the power of genetic linkage designs in ADHD research, as these designs are suited for the identification of individual genetic variants that explain at least 10% of the genetic variance of a trait or disease (Risch and Merikangas 1996). A situation like this calls for a new analytic approach, one which combines the power to detect genetic variants of small effect size, like the association studies, with the possibility to perform hypothesis-free analyses of the entire genome. This has recently become feasible using the genome-wide association study (GWAS) methodology. In GWAS, 100,000 to more than 1,000,000 genetic variants (in this case single nucleotide polymorphisms, SNPs), distributed across the entire genome are genotyped using a microarray platform and (individually) tested for association with a trait or disorder. The selection of SNPs is mostly based on the distribution of linkage disequilibrium (LD) across the genome, as identified by the HapMap project (Frazer et al. 2007; The International HapMap Consortium 2005). This selection is sufficient to obtain information about most of the frequent genetic variation in the genome (Manolio et al. 2008; Pearson and Manolio 2008), especially for the newer platforms, which include around 1,000,000 or more SNPs. In the short time of their existence, GWAS have already led to the identification of more than 100 genetic variants significantly associated with about 40 different traits and diseases, including a number of psychiatric ones (e.g. Ferreira et al. 2008; O’Donovan et al. 2008; O’Donovan et al. 2009; Sklar et al. 2008). Although highly successful, design-wise (for an overview of GWAS approaches see Neale and Purcell 2008), the GWAS method has not been exploited to the fullest, given its high financial costs (sample sizes of several thousands to ten-thousands may be needed, see for a recent discussion of this matter Burton et al. 2008). Major progress in the field came through a number of GWAS initiatives funded by government and industry, like the Welcome Trust Case Control Consortium (WTCCC; 2007) and the Genetic Association Information Network (GAIN; Manolio et al. 2007). Among the 13 diseases that were analyzed in the first rounds of these two initiatives, the WTCCC included 1 and GAIN included 4 psychiatric disorders.

GAIN has also been the setting for the first GWAS in ADHD, which was carried out in 958 Caucasian case–parent-trios collected as part of the International Multicentre ADHD Genetics (IMAGE) study in children (Brookes et al. 2006; Kuntsi et al. 2006). Using the genotyping data from this study, Neale et al. (2008a) reported a classic transmission disequilibrium analysis (TDT) of a categorically defined ADHD phenotype. In addition, Lasky-Su et al. (2008b) performed an analysis using quantitative measures of ADHD symptoms. In the meantime, a second set of genome-wide data has been reported by Lesch and coworkers. These researchers used pools of DNA from 343 ADHD-affected adults and 304 controls for their genome-wide association analysis of a categorical ADHD phenotype (Lesch et al. 2008). Following the first analyses, the GAIN/IMAGE ADHD dataset has since been used for a number of exploratory studies using additional phenotypes and designs, like the age of onset of ADHD symptoms (Lasky-Su et al. 2008a), conduct problems (Anney et al. 2008b), and even a genome-wide gene–environment interaction (G × E) study evaluating genetic variants moderating the effects of maternal expressed emotion on ADHD symptoms (and conduct problems; Sonuga-Barke et al. 2008). In the following paragraphs, we will discuss most of these studies and their most important findings. The two studies exploring conduct problems in the GAIN/IMAGE dataset (Anney et al. 2008b; Sonuga-Barke et al. 2008) are beyond the scope of this review.

As mentioned above, most of the GWAS data so far are derived from one sample of child–parent triads collected as part of the IMAGE study. IMAGE is a consortium of researchers from seven European countries and Israel. IMAGE has ascertained more than 1,400 families through a (preferentially combined subtype) ADHD-affected proband (Brookes et al. 2006; Kuntsi et al. 2006). The probands were required to be in the age range from 5 to 17 years. All families were of European Caucasian descent. In all families, both parents and at least one sibling, the latter not selected for phenotype, were required to take part in the study. IMAGE participants were extensively phenotyped for ADHD and comorbid disorders, using the Parental Account of Childhood Symptoms (PACS) (Taylor et al. 1986), administered by centrally trained investigators. In addition, rating scale measures were used, including the Long Version of Conners’ Parent and Teacher Rating Scales, as well as the parent and teacher version of the Strength and Difficulties Questionnaires (SDQ) (Conners 2003). A standardized algorithm was applied to PACS to derive each of the 18 DSM-IV ADHD items, and, taking into account a number of Conners items, a clinical diagnosis was made. Exclusion criteria included a low IQ (<70), autism, epilepsy, and brain or genetic disorders known to mimic ADHD symptoms. In GAIN, 958 family triads from IMAGE, including the proband and his/her parents were analyzed (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000016.v1.p1). DNA was isolated from blood by Rutgers Cell and DNA Repository (http://www.rucdr.org/). The samples were analyzed by Perlegen Sciences on a microarray specially designed for GAIN. The array investigates approximately 600,000 SNPs able to capture information of close to 100% of the common genetic variation in Caucasians (Neale et al. 2008a). Data cleaning of the genotyping results was performed through the National Centre for Biotechnology Information (NCBI), leaving 438,784 SNPs for analysis.

The first paper on the GAIN/IMAGE ADHD genome-wide dataset was based on the 909 complete family triads, for which full genotyping information was available (Neale et al. 2008a). Of these families, 790 contained male probands, 119 female probands. A total of 845 probands had a combined subtype ADHD diagnosis, the average number of symptoms was 16.1. The TDT design was chosen for this study which contained families recruited from all over Europe and from Israel, based on its robustness against population stratification. Earlier research by Neale et al. (2008b) had shown that differences indeed existed between the different genetic backgrounds of the IMAGE participants, especially with the Israeli families of European Caucasian descent. In addition, the family controlled design showed more stability in the presence of non-random genotyping error in GAIN (Anney et al. 2008a). No genome-wide significant results were observed (with the highest P values at 7.45E-06), although a power analysis presented in the paper suggests that the size of the dataset would have been sufficient to detect at least one genome-wide significant (P value 5.0E-08) finding for a genetic variant with a minor allele frequency (MAF) of 0.2 causing an odds ratio of 1.3. A table in the Neale et al. paper presents the highest-ranking 25 SNPs (based on P value) (see also Table 1 in this review).

Table 1 Findings from TDT-based GWAS by Neale et al. (2008a), including genes in or near which the association finding is present as well information regarding the role of the gene and its possible involvement in psychiatric disorders
Full size table

In a TDT design, the number of observed transmissions of a given allele of a SNP is compared to the expected frequency under the assumption of no association (Spielman and Ewens 1996). Performing the TDT, the authors noted a highly significant imbalance in the number of major and minor allele over-transmissions. Given the large numbers of SNPs analyzed, these numbers would have been expected to be similar. Removing additional SNPs that had failed quality control in one of the other two GWAS based on the Perlegen microarray improved this imbalance partly, suggesting that genotyping error or missingness is (at least partly) responsible for this effect. Therefore, the authors repeated their analysis including a correction factor. Comparison of the presented tables listing the top-25 SNPs before and after correction shows that this resulted in the loss of half of the listed SNPs and a general downward correction of P values. Although, as the authors state, it might be (too) early to explore the biological relevance of the strongest observed associations, it is a fact that the list contains a number of very interesting candidate genes for ADHD (Table 1), most notably those genes that also show association in the GWAS of quantitative ADHD-related phenotypes (Lasky-Su et al. 2008b), as indicated in Table 1. Among these is a SNP close to CNR1, which encodes the cannabinoid receptor 1, a gene that falls into a linkage region for ADHD from a primary analysis (Ogdie et al. 2004) and shows suggestive linkage in meta-analysis (Zhou et al. 2008c). CNR1 was also found associated with the disorder in candidate gene-based association studies (Lu et al. 2008; Ponce et al. 2003). Furthermore, association of CNR1 with alcohol and drug abuse and dependence has been observed (Zuo et al. 2007). Also of interest are the findings for the especially brain-expressed cytoskeleton-organizer DCLK1 and the extracellular matrix component SPOCK3, as well as the two potassium-channel regulators KCNIP1 and KCNIP4 (Table 1).

In an attempt to maximize the power of the GAIN/IMAGE family-based GWAS, Lasky-Su et al. (2008b) published a study focusing on quantitative ADHD phenotypes using the Family-based Association Test (FBAT) suite of programs. Based on the 18 DSM-IV symptoms of ADHD the authors constructed 3 quantitative phenotypes using (1) the number of hyperactive–impulsive symptoms, (2) the number of inattentive symptoms, and (3) the total number of symptoms. These phenotypes were each tested under three different genetic models (additive, dominant, recessive). Furthermore, three additional quantitative traits were constructed based on the Conners ADHD Rating Scales and the PACS using FBAT-PC methodology (Lange et al. 2004c). This approach constructs a slightly different phenotype for every SNP in order to maximize the heritability at this SNP. Association testing was subsequently limited to the set of ten SNPs per genotype with the highest power to detect association derived from a screening algorithm in the program PBAT (Lange et al. 2004b; van Steen and Lange 2005). In total, this study performed 18 GWAS, increasing the multiple testing burden; correction for multiple testing was limited to the single GWAS.

In addition to increasing power, using symptoms or rating scales as a basis for phenotype selection had the advantage that hyperactive–impulsive and inattentive symptoms could be assessed independently in the combined subtype ADHD of the GAIN/IMAGE participants. However, a prerequisite for this quantitative approach to be valid is that ADHD is a disorder at the extreme of a continuum observed in the population. Evidence suggesting that this is indeed true comes from twin studies showing strong heritability of quantitative measures of ADHD (Lasky-Su et al. 2008b, and references herein) as well as from a proband–sibling comparison in IMAGE (Chen et al. 2008; Thapar et al. 2006). Starting from the cleaned GAIN dataset, Lasky-Su restricted the analysis to 429,784 autosomal SNPs (as FBAT cannot handle X-linked markers) and 909 complete families. 87% of the probands were male, and the mean age of probands was 10.88 years. The average total number of symptoms was 16.11, the number of hyperactive–impulsive symptoms was 8.11 and the number of inattentive ones was 7.98.

Two genome-wide significant findings (corrected for single analysis) emerged from the formal tests, one for rs6565113 in an intron of CDH13 associated under an additive genetic model with the FBAT-PC-based phenotype derived from all ADHD symptoms, one for rs552655 in an intron of GFOD1 associated with the FBAT-PC phenotype for inattentive symptoms (dominant model). The CDH13 finding is especially interesting, since a SNP near this gene also is part of the top-25 in the GWAS by Neale et al. (2008a, b), though the two studies are not entirely independent, of course, and the gene is also found among the top-findings from the ADHD GWAS by Lesch et al. (2008, see below) and falls into the only significant ADHD linkage region identified in the recent meta-analysis (Zhou et al. 2008c). Furthermore, association with SNPs in CDH13 is one of the most consistent findings in genome-wide studies on a wide variety of phenotypes related to drug abuse and dependence (Uhl et al. 2008a, b), although there is no clear indication, which region of the gene is most important. Possibly, different regions or risk factors might cause different phenotypes. CDH13 codes for cadherin 13 (or T-cadherin), a member of a family of cell–cell adhesion proteins (Patel et al. 2003), in addition to being a regulator of neural cell growth. CDH13 shows a brain expression profile consistent with a role in ADHD, being expressed in regions showing volumetric reductions in patients with this disorder (Takeuchi et al. 2000; Takeuchi and Ohtsuki 2001; Valera et al. 2007). Very little is known about the second associated gene, GFOD1 and its product, glucose–fructose oxidoreductase-domain containing 1, except that it is expressed in brain and possibly plays a role in electron transport (Lasky-Su et al. 2008b).

In addition to the formal testing of ten SNPs for each phenotype Lasky-Su and coworkers also report a more exploratory analysis of all SNPs in the dataset. There were 58 association findings with P values smaller than 10−5, of which 46 SNPs were unique. Some SNPs showed association under more than one genetic model and/or across phenotypes. Although none of the SNPs listed here are genome-wide significantly associated with ADHD, a high percentage of the SNPs is present in brain-expressed genes and/or in linkage regions for ADHD and other psychiatric disorders (see Table 2). A particularly interesting finding from this set includes the SNP in NOS1, encoding the neuronal form of nitric oxide synthase. NO acts as the second messenger of the N-methyl-d-aspartate receptor and interacts with both the dopaminergic as well as the serotonergic system in the human brain. NOS1 has been associated with impulsive and aggressive behavior and ADHD (Reif et al. 2009). Association has also been noted with Alzheimer’s disease (Galimberti et al. 2008) and schizophrenia, as well as with related neuropsychological performance (Reif et al. 2006).

Table 2 Association findings from quantitative GWAS by Lasky-Su et al. (2008b) with P values <10−5, including genes in or near which the association finding is present as well information regarding the role of the gene and its possible involvement in psychiatric disorders
Full size table

Lasky-Su and colleagues also reported P values for a selection of predefined (classic) ADHD candidate genes (Brookes et al. 2006). Most significant findings with P values in the range of 10−5 were observed for SLC9A9. This gene (which was with a total of 181 SNPs the largest gene analyzed) also contained the largest number of associations in terms of SNPs and phenotypes. Other findings at P values of 10−4 were for DDC and SNAP25. The SLC6A1, ADRB2, HTR1E, ADRA1A, DBH, BDNF, DRD2, TPH2, HTR2A, SLC6A2, PER1, CHRNA4, COMT and SYT1 genes showed association at 10−3. SLC9A9 encodes a sodium/hydrogen exchanger and may be a particularly viable candidate gene for ADHD. It was originally reported disrupted in a patient with ADHD (de Silva et al. 2003) and also shows mutations in patients with autism (Morrow et al. 2008) as well as having been found in multiple GWAS for addiction-related disorders (Uhl et al. 2008a, b).

A GWAS not based on the GAIN/IMAGE sample was published by Lesch et al. (2008). In their study they analyzed 343 German patients with persistent ADHD (mean age 32.9 years with a range from 18 to 65, 54.5% male, 61.5% combined subtype, 31.2% inattentive subtype, 7.3% hyperactive–impulsive subtype) and 304 controls (mean age 32.7, 51.3% male). Patients were diagnosed using the Structured Clinical Interview of DSM-IV axis 1 disorders (SCID-I; Jacob et al. 2007), supported by ratings from the Wender Utah Rating Scale and an ADHD diagnostic checklist applying DSM-IV criteria categorically and dimensionally (ADHD-DC) (Rosler et al. 2004). Comorbidity along the entire spectrum was analyzed using SCID-I and SCID-II, the latter assessing axis II disorders. Lifetime comorbidity with major depression was 44.7%, with substance use disorders it was 44.1% (of which 32.0% for cannabis use, 40.6% for alcohol use). Of the 304 controls only 54 were screened for psychiatric disorders prior to inclusion. Genotyping was on the Affymetrix GeneChip Human Mapping 500K sets. Importantly, instead of genotyping individual samples, pools were constructed of the 343 cases and 304 controls, respectively. For each pool nine technical replicates were analyzed. After data cleaning the authors ended up with allele frequencies for 504,219 SNPs, of which the autosomal SNPs were ranked based on the mean of rankings performed using ANOVA, rank sums and silhouette scores. Whereas the other ADHD GWAS reports relied strictly on statistical rankings to describe their results, Lesch and coworkers used additional criteria for further prioritization of SNPs for follow up, one being localization within or near genes (±100 kb flanking region), a second being presence in clusters of at least three SNPs located in regions of genes, a third being the comparison with a family-based association study performed in a dataset from a recently published linkage paper by the same group, or being present in a suggestive linkage region from this analysis in adult ADHD patients (Romanos et al. 2008). In addition, expression in brain, as well as earlier findings suggesting involvement in complex traits and neuropsychiatric disorders (especially substance use disorders) were taken into consideration.

Although none of the findings were genome-wide significant, interesting candidate genes for ADHD are reported in this paper, too. Of the 80 top-ranked SNP findings, 30 fall into genes, of which nearly all are known to be expressed in brain (see Table 3). A second list, containing the top-10 of those genes with 3 or more SNPs among the 1,000 highest-ranking findings contains the genes KALRN (15 SNPs), ZNF354C (15 SNPs), WRNIP1 (14 SNPs), GRB10 (10 SNPs), DPP6 (7 SNPs), ARHGAP22 (11 SNPs), RAB38 (11 SNPs), FAT3 (8 SNPs), DA259379 (13 SNPs), NT5DC3 (20 SNPs). A third list summarizes all findings that show overlap with the linkage analysis in adult ADHD patients published earlier by the same group (Romanos et al. 2008), and in some instances also a family-based GWAS based on the 50 K linkage dataset (Table 3 in Lesch et al. 2008). Of the 30 genes listed there, 8 genes (CTNNA2, MOBP, MAP1B, REEP5, ASTN2, ATP2C2, CDH13 and ITGAE) overlap with the list of the 80 top-ranked SNP findings, as indicated in Table 3. The authors highlight the findings for ASTN2, CSMD2, ITGA11, CTNNA2 and CDH13 in their discussion. All of these genes have earlier shown association with substance use disorders and related phenotypes, which is interesting in the light of the frequent comorbidity with such disorders in (adult) ADHD (44.1% in the current study). The five genes encode proteins involved in cell adhesion and/or cell–cell communication in development and maturation of the brain. Also the UNC5B (Table 3) and KALRN genes fit into this category. The latter encodes a protein present in the postsynaptic density which regulates dendritic spine development and is important for the neuronal plasticity underlying memory and learning (Lesch et al. 2008; Penzes and Jones 2008).

Table 3 Top 30 single SNPs located in gene regions (including 100 kb of flanking sequences) ranked according to the mean rank calculated across three statistics as reported by Lesch et al. (2008) for association with persistent ADHD in adults
Full size table

Comparing this list with the findings from the other two GWAS, very little overlap is seen (which might be expected, given power limitations of the studies, see below), except for one finding on chromosome 16, which falls into CDH13. As indicated above, this gene was also among the top-findings from both GAIN/IMAGE GWAS studies, as well as many genome-wide studies of addiction related phenotypes, recently reviewed by Uhl et al. (2008a). An additional similarity to be noted is for two TLL genes, TLL1 found among the top-ranks of the analysis based in the categorical ADHD phenotype and showing nominal association in the quantitative analyses by Lasky-Su et al. (see Table V in Lasky-Su et al. 2008b), TLL2 present among the top-30 SNPs in gene regions in the analysis by Lesch et al. Both genes encode metalloproteases that cleave collagen and are expressed in multiple tissues including brain. A specific role in brain development or functioning has not been reported for either of these genes, though. Additional genes modulating the extracellular matrix are found in all three studies, including HAS3 (Lasky-Su et al. 2008b), SPOCK3 (Neale et al. 2008a), MAN2A2, GPC6 and MMP24 (Lesch et al. 2008). A third interesting overlap is for genes related to potassium-channel function, like the KCNC1 gene reported by Lesch et al. (2008), the KCNIP1 and KCNIP4 genes observed by Neale et al. (2008a)) and the DPP10 finding for total ADHD symptom count (Lasky-Su et al. 2008b).

An exploratory analysis based on the GAIN/IMAGE sample made use of the age of onset of hyperactive–impulsive and inattentive symptoms, as assessed with a question from the PACS interview, each, as an informative phenotype for a GWAS (Lasky-Su et al. 2008a). Earlier work of the authors had suggested that environmental exposure outside the womb were less likely to contribute significantly to the development of ADHD with a very early age of onset (Lasky-Su et al. 2007). The number of children included in the analysis was 930 with a mean age of 10.86 years, 810 of them being males. The average age at onset of ADHD symptoms was 2.67 years; for the hyperactive–impulsive symptoms this age was 2.77, for the inattentive symptoms it was 4.31 years. The study investigated 429,784 autosomal SNPs using the FBAT-logrank algorithm (Lange et al. 2004a) to identify genetic variants that could predispose children to an earlier age of onset of disease symptoms. Additive, dominant and recessive genetic models were tested. A total of 16 SNPs reached P values for association <10−5 including 14 unique findings (Table 4). There is no overlap with the results from the other GWAS. Although potentially due to chance it is remarkable to see that this list contains a number of genes that are involved in the response of the body to environmental exposures and/or are regulated by the environment, like the inflammation related ADAMTS2, the stress regulator MAP3K7 or the vitamin A-responsive NAV2 gene.

Table 4 Findings from genome-wide analysis of ADHD age of onset by Lasky-Su et al. (2008a), including genes in or near which the association finding is present
Full size table

An interesting new approach in GWAS is represented by the exploratory study by Sonuga-Barke et al. (2008). The author and his colleagues evaluated gene-environment interactions (G × E) in ADHD in a genome-wide analysis. Specifically, they investigated the effects of maternal warmth and criticism on ADHD severity in the GAIN/IMAGE dataset (also, a G × E analysis for conduct disorder symptoms is reported in the publication, but is not reviewed here). ADHD symptoms were derived from the PACS interview in IMAGE, maternal expressed emotion was assessed using the Camberwell Family Interview. Maternal warmth and criticism were coded, separately; warmth was assessed by tone of voice, spontaneity, sympathy and/or empathy toward the child. Criticism was coded from statements that criticized or found fault with the child based on the tone of voice and critical phrases (Sonuga-Barke et al. 2008). Statistical analysis made use of the FBAT-Interaction methodology (Vansteelandt et al. 2008). In total, 430,060 autosomal SNPs were included in the analysis, as were 909 patients (87% males) from complete trios. Mean age of the participants was 10.88 years, the average number of ADHD symptoms was 16.1, the mean overall measure for mother’s warmth was 1.43 (range from 0 to 3) and for maternal criticism it was 1.74 (range from 0 to 4). Additive, dominant and recessive genetic models were tested for the phenotype and the two environmental factors.

The analysis did not produce any genome-wide significant findings, the highest P value reaching 9.08 × 10−7. Nineteen SNPs showed interaction association P values below 10−5 in the presence of main genetic effects (at P ≤ 10−5), of which five showed association under two genetic models (Table 5). Only two findings were related to maternal criticism. Most of the genes interacting with maternal warmth clustered in three distinct regions on chromosomes 6, 11 and 13. The regions on chromosomes 6 and 11 fall into genes, KIF6 and PIWIL4, respectively. The region of 58 kb on chromosome 13 includes 8 association findings for 4 unique SNPs, and although it does not contain genes, it shows extensive LD extending proximally to the SLC46A3 gene, the function of which is unknown. Interestingly, the KIF6 gene has repeatedly been shown to respond to environmental stimuli (Iakoubova et al. 2008b; Iakoubova et al. 2008a; Shiffman et al. 2008). In the referenced studies, a coding variant in KIF6 altered the response to statins in cardiovascular disease. The gene’s role in moderating the influence of mother’s warmth on ADHD severity will have to be investigated in further detail. An additional six interaction associations at P values of 10−4 were observed in the absence of genetic main effects (Table V in Sonuga-Barke et al. 2008). Interestingly, all but one finding were observed under a recessive genetic model, which strongly limited the number of informative families. Of these associations, only one was for mother’s warmth. This finding falls into the BCAS1 gene on chromosome 20, which is a potential oncogene with high expression in brain. Of the five association findings interacting with mother’s criticism, especially a finding on chromosome 2 is interesting, as it falls into the MTA3 gene, which has two additional SNPs associated with ADHD in the GWAS for quantitative phenotypes (Lasky-Su et al. 2008b). The authors also tested potential G × E effects for the list of ADHD candidate genes defined earlier (Brookes et al. 2006). Interaction association P values are most significant for DDC (mother’s warmth and criticism), HTR2A and SLC9A9 (warmth) at 10−4.

Table 5 Summary of the GxE interaction association P values <10−5 using total ADHD symptoms as the phenotype (Sonuga-Barke et al. 2008)
Full size table

The interpretation of the analytical strategy used in this study is difficult. First of all, the direction of causation is not clear: it could be that maternal expressed emotion influences ADHD severity in the child, on the other hand, a child with severe ADHD might also elicit more pronounced criticism and less warmth from parents. Furthermore, explaining the biological effects of the environmental moderators of ADHD severity is not straightforward. The authors discuss epigenetic effects (like DNA methylation) as possible mechanisms. This might indeed be supported by at least one of their findings: PIWIL4 plays a role in chromatin-modification, another epigenetic regulator of gene expression (Szyf et al. 2007, 2008).

Taking together all results from the GWAS performed in ADHD, the following is to be remarked: none of the findings so far show genome-wide significant association with ADHD according to the thresholds currently handled (Dudbridge and Gusnanto 2008; Pe’er et al. 2008; van Steen and Lange 2005). Nevertheless, the most high-ranking findings from the different studies contain interesting new candidates for further study. Little overlap is observed between studies, except for CDH13, which is found in three studies (Lasky-Su et al. 2008b; Lesch et al. 2008; Neale et al. 2008a). None of the ‘classic’ candidate genes for ADHD are found among the top-findings of either study, although two more recently reported genes, NOS1 and CNR1, are observed at least once. In the candidate gene analyses reported for the GAIN/IMAGE sample, most evidence is repeatedly found for involvement of SLC9A9 in ADHD. For all studies discussed here, overlap with results from linkage studies reported in ADHD and linkage as well as GWAS in other psychiatric disorders is multiple.

Looking at the processes potentially involved in the etiology of ADHD, little evidence (as yet) points to a direct involvement of neurotransmitters (at least the classic dopaminergic, noradrenergic and serotonergic pathways) or regulators of neurotransmission (except for DNM1). Some suggestions are found for involvement of ‘new’ neurotransmission and cell–cell communication systems, though, like glutamate, vasopressin, and TAFA proteins. Also, the potential involvement of potassium channel subunits and regulators definitely warrants further investigation. However, more basic processes also seem involved, like cell division, adhesion (especially via cadherin and integrin systems), neuronal migration, and neuronal plasticity, as well as related transcription, cell polarity and extracellular matrix regulation, and cytoskeletal remodeling processes. Our conclusion is consistent with and extends the conclusion drawn by Lesch and coworkers based on the findings from their GWAS, in which they suggest that neuronal spine formation and plasticity may underlie the pathophysiology of ADHD. Finding inflammatory genes (IL16 and ADAMTS2) is somewhat surprising, although inflammation is already known to be involved in different psychiatric disorders, like major depression and Alzheimer’s disease (Anisman 2009; Bazar et al. 2006; Garcia-Bueno et al. 2008; Serretti et al. 2007). The analysis of age of onset of ADHD revealed a number of environmentally regulated genes (Lasky-Su et al. 2008a), which may suggest a particularly important influence of the environment in this aspect of the disorder. However, all of these biological hypotheses must be tempered by the limited knowledge of the genetic architecture of ADHD at this point. Currently, we do not have sufficient information to draw strong conclusions about the relative impact of the biological pathways and genetic influences, aside from likely knowing that the genetic effects are modest in size.

In the last 2 years, much progress has been made in identifying genes for psychiatric disorders through GWAS. From the studies so far it has become apparent that early power estimates have been far too optimistic and at least several thousands of samples will be needed for the design of an optimal study (Burton et al. 2008). Genetic heterogeneity appears even more extensive than previously expected, with relative risks due to single genetic variants in the range of 1.05–1.1 or even smaller (Ferreira et al. 2008; Neale et al. 2008a; O’Donovan et al. 2008; Sklar et al. 2008). ADHD is no exception to this, and so all of the ADHD GWAS performed to date are highly underpowered. In addition, Lesch and colleagues used a pooling design, which reduces power by about 70% and puts limitations to the genetic models that can be investigated (Lesch et al., 2008). Furthermore, using controls unscreened for the presence of psychiatric disorders can be expected to reduce power, though the effect of this is modest, given that the prevalence is at most 5% (McCarthy et al. 2008). The X-chromosome was not analyzed in most of the studies, due to a limitation of the FBAT-program. As a clear difference in prevalence of ADHD exists between genders, this might still be an interesting target to look at.

An important issue in psychiatric disorders in general and ADHD in particular is phenotypic heterogeneity. First, psychiatric diagnosis must rely on clinical symptoms that, without external validators, cannot be expected to reflect biologically relevant concepts that are closely linked to genes (Kuntsi et al. 2006; Thapar et al. 2006). Although the instruments used to define psychiatric GWAS phenotypes are generally chosen for high reliability, they have no proven biological validity and will therefore (at best) result in a more ‘noisy’ phenotype definition than most biological tests (Craddock et al. 2008). More importantly, however, the classical nosological approaches identify phenotypically distinct diseases, such as ADHD, substance abuse, affective disorder and schizophrenia, though these exhibit considerable overlap in symptomatology, clinical course, and long-term outcome. ADHD is hardly ever the only disorder a patient suffers from, comorbidity is the rule rather than the exception. For example, reported frequencies of depression and substance use disorders were also high in the GWAS sample used by Lesch et al. (2008). These factors severely hamper the power of genetic association studies and might explain why psychiatric genetics has had an extremely modest track record in pinpointing and characterizing susceptibility loci, even using the GWAS approaches.

That said, what are the perspectives for genome-wide association studies in ADHD? The results of the first studies—though most not genome-wide significant—are already very exciting and change the way in which we think about genetic factors involved in ADHD etiology. For now, GWAS holds the greatest promise for understanding the genetic architecture of ADHD! However, we need to improve the design of these studies by increasing sample size for more power and by further improving/extending phenotypic assessment to better reflect and partition the phenotypic complexity of the disorder. The former will only become possible through collaboration. As a field, ADHD has the good fortune of having several thousand samples collected world-wide, with collaborative efforts supported by the National Institute of Mental Health. Investigators with ADHD GWAS datasets are also participating in the Psychiatric GWAS Consortium (PGC), a confederation of 101 scientists from 11 countries and 48 institutions having GWAS datasets on ADHD, autism, bipolar disorder, schizophrenia and major depression (Psychiatric GWAS Consortium Steering Committee 2009). The PGC investigators have agreed to participate in coordinated mega-analyses both within and across disorders. Within 1 year, it is to be expected that a collaborative GWAS with more than 4,000 cases can be carried out via the PGC for ADHD. It is this spirit of collaboration that we believe will be needed to achieve the breakthroughs in genetic research that will ultimately help us discover new pathophysiological pathways and targets for treatment.

Improving GWAS design through extension of phenotypic assessment could be done by taking characteristics into consideration that we already know or suspect to play a role in genetic susceptibility. Given the limits of the nosological systems, subdividing ADHD according to subtype will probably not be sufficient, as there is already evidence of shared genetic factors between subtypes (Faraone et al. 2000a; Thapar et al. 2006). Taking into account comorbidity might deal with some of the limits of the phenotypes, as a number of studies have suggested that ADHD plus a specific comorbidity may constitute a distinct subtype of ADHD (Faraone et al. 2000b; Faraone 2000; Faraone and Doyle 2001; Fliers et al. 2009; Rasmussen et al. 2004; Smalley et al. 2000; Willcutt et al. 2000). A recent study in IMAGE can be used as an example that this affects association findings: Zhou et al. (2008b) showed that SLC6A3/DAT1 was only associated with ADHD in the absence of conduct disorder. Another example is gender; given the difference between sexes, certain genetic risk factors can be expected to affect males and females differently. That this affects association studies has already been shown in ADHD (Biederman et al. 2008; Guimaraes et al. 2007; Rommelse et al. 2008b) as well as other (related) psychiatric disorders (Craig et al. 2004; Diamond 2007; Haefner et al. 2008; Kendler 1998; Lang et al. 2008; Okamura et al. 2007; Rucci et al. 2009). An additional potentially important factor is age at diagnosis. It is well established that age influences cognitive performance, with children performing worse in neuropsychological tests than adolescents and adults (Addamo et al. 2007) on tasks that are relevant to ADHD (Rommelse et al. 2008a). We and others have presented evidence that age can modulate the association of SLC6A3/DAT1 with ADHD, with different haplotypes of the gene being associated with ADHD in childhood and adulthood (Barkley et al. 2006; Elia and Devoto 2007; Franke et al. 2008).

A more radical way of dealing with the shortcomings of the nosological system is to look deeper into the possibilities of using (ADHD) endophenotypes for GWAS. Endophenotypes, or intermediate phenotypes, represent heritable phenotypic constructs that are, presumably, more directly affected by genes than clinical symptoms or disease categories (Gottesman and Gould 2003; Szatmari et al. 2007a; Walters and Owen 2007). This should result in a stronger association between causal genes and the endophenotype than the clinical disease phenotype. The success of an endophenotype strategy requires either a higher heritability of the endophenotype compared to the disease phenotype (which is generally not generally observed, Szatmari et al. 2007a) or a reduced complexity of the genetic architecture of the endophenotype due to the involvement of less genes (see Fig. 1). Endophenotypes for ADHD have been proposed at different levels, e.g. those based on neuropsychological performance (Doyle et al. 2005; Rommelse 2008) and those based on neuroimaging (Castellanos and Tannock 2002), i.e. brain activity and structure. Recent genome-wide linkage studies of neuropsychological endophenotypes for ADHD indeed suggest that these measure offer increased power for finding loci compared to clinical phenotypes (Doyle et al. 2008; Rommelse et al. 2008c). A pooled GWAS on memory performance also suggests this (Papassotiropoulos et al. 2006). ‘Imaging genetics’ studies would be thought to be particularly powerful, requiring even smaller sample sizes (Dreher et al. 2008; van Haren et al. 2008). Studies using such ADHD endophenotypes for GWAS are currently ongoing.

Fig. 1
figure 1

Simplified schematic representation of the endophenotype concept in psychiatric genetics. Many genes (depicted in an oversimplified way as gene A to I) are involved in causing a disease symptom and the categorical disease phenotype. A reduced number of genes is involved in disease related endophenotypes, like ‘Function of a brain unit’ (Level 1), as assessed through performance on neuropsychological tests; even less genes play a role if an endophenotype at levels 2 or 3 is used for genetic association studies. On level 2, endophenotypes measured through (structural and functional) neuroimaging are shown, on level 3 cellular behavior, assessed through e.g. assays measuring migration, apoptosis or cell division rates are shown

Full size image

An additional aspect to be taken into consideration in GWAS is the potential role of G × E and gene–gene (G × G) interactions in the etiology of ADHD. As the exploratory analysis by Sonuga-Barke et al. (2008) shows, this type of analysis will help to further delineate mechanisms moderating disease severity and those leading to disease.

A last aspect to be discussed relates to the genetic models currently used in ADHD research. Up to now, we work under the assumption that the ‘common disease, common variant’ model will hold for ADHD. This model suggests that ADHD is caused by multiple variants that are common in the population and each contribute a small amount of disease risk. However, experience from GWAS in other multifactorial disorders now challenges this view and postulates the existence of rare genetic variants contributing to disease, that are not captured by GWAS (see also McCarthy et al. 2008 for a more elaborate discussion of this subject). This case of the ‘missing heritability’ might imply that private or low-frequency variants of large effect size exist and actually explain most of the heritability of a multifactorial disorder in individual patients. Copy number variants (CNVs) might contribute a part of this puzzle. CNVs include insertions, deletions and duplications and encompass relatively large genomic segments of 1 kb to several Mb in size (Redon et al. 2006; Sebat et al. 2004). Although they are a frequent class of polymorphism in the human genome, most CNVs are rare. Many reports indeed now suggest involvement of CNVs in psychiatric disorders (Lachman et al. 2007; Marshall et al. 2008; Sebat et al. 2007; Stefansson et al. 2002, 2008; Szatmari et al. 2007b; Ullmann et al. 2007; Vrijenhoek et al. 2008; Walsh et al. 2008; Weiss et al. 2008). The role of CNVs in ADHD has not been studied, yet.

In conclusion, the GWAS performed so far in ADHD, though far from conclusive, provide a first glimpse at genes for the disorder. Many more (much larger studies) will be needed. For this, collaboration between researchers as well as standardized protocols for phenotyping and DNA-collection will become increasingly important. Furthermore, understanding how the genes identified by GWAS contribute to ADHD risk, will require smart biological experiments facing the challenge of an organ as inaccessible and complex as the human brain.