Autism is a heterogeneous disorder characterized by impairments in social interaction, deficits in verbal and nonverbal communication, restricted interests and repetitive behaviors [1]. Autism comprises the severe end of a group of autism spectrum disorders (ASD) [2]. The prevalence of autism is estimated at 0.2%, with males being more likely than females to have a diagnosis of autism (ratio of approximately 4:1) [3].

There is compelling evidence from twin and family studies indicating a strong genetic component in autism. The average risk of recurrence of autism in siblings is approximately 10% [4] in families with one affected sibling, which is much higher than the prevalence in the general population, but much lower than would be expected for a single-gene disease [2]. Indeed, cases of autism have only rarely been associated with mutations in single genes or with chromosomal duplications or other abnormalities [5–9]. Furthermore, the inheritance pattern in most cases of autism is complex and not compatible with simple Mendelian inheritance [10, 11]. A series of common autism susceptibility genes has been identified in the literature, but taken individually, the risk associated with these genes remains modest. Although spontaneous mutations may help explain isolated cases of autism, the inheritance pattern of common variants may be at the root of autism in a large proportion of multiplex families.

Studies using simulated data have demonstrated that the joint analysis of such common low-risk variants has the potential to identify individuals at higher risk for disease, and could be useful for complex disease prediction [12–14]. In complex disease states such as type 2 diabetes [15–19], cancer [20, 21] and cardiovascular disease [22–25], the accumulation of multiple risk alleles markedly increases the risk of being affected, and compared with studying single nucleotide polymorphisms (SNPs) independently, it is better at identifying subgroups of individuals with significantly greater risk. Previous studies have shown associations between several genetic variants and autism, but many of these studies could not be replicated in different populations.

For our study, we selected single nucleotide polymorphisms (SNPs) within four genes that have been previously shown to be associated with autism. Original publications and replication studies have shown that SLC25A12 [26–28] and EN2 [29, 30] are associated with autism. PITX1 has been identified through a linkage study [31] using the physical identity by descent method [32]. The same method was used to identify ATP2B2 on chromosome 3 [33] in a region in which different autism linkage peaks and association with microsatellite markers have been reported [34–36]. We sought to demonstrate that joint analysis of common low-risk variants would have the potential to identify a subgroup of individuals at increased risk for developing autism. To assess the effect of combined analysis of multiple genetic variants in autism, we conducted a primary analysis on the first family sample using the four genes mentioned above. The aims of this analysis were to assess the strength of the association of the combination of these four genes with autism in a population of affected children, and to define a genetic scoring metric. We followed up with a second analysis to confirm our findings in an independent set of families, including a sibling diagnosed with autism.

None of the SNPs exhibited a departure from HWE, and allele frequencies were similar between samples (Table 1). All risk alleles were very common, with frequencies ranging from 0.4 (ATP2B2) to 0.9 (PITX), similar to those estimated in the literature [26, 27, 29–31, 33]. The distribution of case subjects in both samples was similar for each genetic score category (range 2 to 8), as was the distribution of pseudocontrol and control subjects. Control subjects were more often in the lower genetic score categories (≤ 4), and the case subjects were more often in the higher genetic score categories (> 6) (Figure 1).

Gene	SNP	Risk allele	AGRE sample	Seattle sample
PITX	rs6872664	C	0.89	0.90
ATP2B2	rs35678	T	0.45	0.41
SLC25A12	rs2292813	C	0.90	0.90
EN2	rs1861972	A	0.75	0.72

AGRE, Autism Genetic Resource Exchange; SNP, single nucleotide polymorphism.

The sensitivities and specificities for each genetic score were similar in the two samples, with overlapping CIs (Table 4). The resultant ROC curves for the genetic scores estimated in the two analyses (Figure 2) gave an AUC of 0.59 for both the AGRE (95% CI 0.55 to 0.64) and Seattle (95% CI 0.53 to 0.65) samples.

Genetic score	AGRE sample		Seattle sample
	Sensitivity (95% CI)	Specificity (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)
2*	1.00	0.00	1.00	0.00
3	0.99 (0.98 to 1.00)	0.02 (0.00 to 0.03)	0.99 (0.97 to 1.00)	0.01 (0.00 to 0.03)
4	0.94 (0.92 to 0.97)	0.06 (0.03 to 0.09)	0.95 (0.92 to 0.98)	0.04 (0.00 to 0.07)
5	0.81 (0.76 to 0.86)	0.26 (0.21 to 0.31)	0.81 (0.75 to 0.87)	0.26 (0.18 to 0.35)
6	0.50 (0.44 to 0.56)	0.67 (0.61 to 0.73)	0.50 (0.43 to 0.57)	0.65 (0.55 to 0.74)
7	0.19 (0.14 to 0.25)	0.91 (0.87 to 0.96)	0.17 (0.12 to 0.22)	0.93 (0.87 to 0.98)
8	0.09 (0.05 to 0.12)	0.98 (0.96 to 1.00)	0.08 (0.04 to 0.12)	0.98 (0.96 to 1.00)

AGRE, Autism Genetic Resource Exchange; CI, confidence interval.

^aReference group.

In the Seattle sample, for which true controls were available, 17% of individuals with autism carried seven risk alleles (genetic score ≥ 7) compared with only 8% of children without autism (Figure 1). Low sensitivity of 0.17 (95% CI 0.12 to 0.22) was associated with a high specificity of 0.93 (95% CI 0.87 to 0.98), which translated into a 2.33-fold increased risk for autism (Table 4). In the Seattle sample, 8% of cases had a maximum genetic score of 8, compared with only 2% of unaffected siblings of individuals with autism. The specificities for an extreme genetic score of 8 reached 0.98 (95% CI 0.96 to 1.00), but this was due to reduction in sensitivity (0.08; 95% CI 0.04 to 0.12), leading to a 4.44 increased risk of autism.

We did not observe any association between the genetic score and IQ score in the whole case-control Seattle sample (P = 0.44), when adjusted for autism status and gender. This was also the case in affected (P = 0.63) and unaffected (P = 0.68) siblings adjusted for gender.

Complex pathologies such as autism may involve multiple common low-risk variants that confer relatively small predisposing effects. In the present study, we hypothesized that the combined analysis of such low-risk variants would more accurately identify a group of individuals at higher risk of developing the disease than an analysis of independent variants. To our knowledge, this is the first study report of its kind in autism.

Although many variants have been described in the literature as being associated with autism, few associations have been confirmed, possibly due to differences in methods and sample sets used, and to the weak effect of the individual variants. Previous studies have shown that variants in SLC25A12 [26–28], EN2 [29, 30], ATP2B2 [33] and PITX1 [31], have an association with autism. A few studies have physiologically linked these genes to autism as well.

Expression of SLC25A12 may be involved in the pathophysiology of autism [48] because postmortem samples of brain tissue from patients with autism have been found to have stronger expression of SLC25A12 than normal brain samples. The EN2 gene is specifically involved in patterning the region that gives rise to the cerebellum [49]. EN2 knockout mice exhibit neuroanatomical and behavioral abnormalities that partly resemble those in patients with autism [50]. PITX1 is a key regulator of hormonal genes in the pituitary-hypothalamic axis. Its putative involvement in autism is supported by evidence documenting abnormal levels of downstream hormones, such as adrenocorticotropic hormone (ACTH), beta-endorphin and cortisol in individuals with autism [51–53]. Deregulation of pro-opiomelanocortin and high levels of beta-endorphin in the morning, for example, have been shown to be involved in certain maladaptive behaviors, such as self-injurious behaviors, which are often seen in individuals with autism [54]. The ACTH-cortisol system, which also plays an important role in stress-related responses, is impaired in individuals with autism in whom lower cortisol levels and higher ACTH levels have been reported [44, 52]. Finally, the ATP2B2 gene codes for a major calcium pump expressed at particularly high levels in Purkinje neurons. Purkinje cells are involved in motor coordination, working memory and learning. Even though the exact correlation between Purkinje neurons and autism remains unclear, selective loss of Purkinje cells and cerebellar atrophies are the most consistently found neurological abnormalities in people diagnosed with autism [55].

Odds ratios ranging from 1.16 to 1.74 were observed in this study, and P values supported a nominal association for individual variants of EN2, ATP2B2 and PITX1, but not SLC25A12. However, because of its reported genetic association with autism [26, 27, 29] and the possible relationship with autism of its overexpression [48], it was included in the combined analysis. In the first stage of the combined assessment, a primary analysis was performed to evaluate the effect of the combination of variants as determined by the genetic score. In the second stage, a case-control analysis was performed to confirm the findings of the AGRE sample analysis. The distribution of genetic scores and ORs in this analysis was similar to that observed in the AGRE sample analysis, with an OR of 1.32. The ORs for a one-point increase in genetic score for the Caucasian subgroup in both the AGRE and Seattle sample analyses were also similar to those of the overall respective sample sets. We estimated that the OR for siblings who carry > seven risk alleles (approximately 12% of siblings), regardless of phenotype, was 3.15, compared with siblings who carry < four risk alleles (approximately 23% of siblings) in the Seattle sample, which used true sibling controls. A subgroup analysis did not yield any significant differences based on gender. Because of sample size, only the Caucasian subgroup could be analyzed. The ORs for a one-point increase in genetic score for the Caucasian subgroup in both the AGRE and Seattle sample analyses were also similar to those of the overall respective sample sets. We did not observe any association between IQ and genetic score, thus, we believe that the observed association is not an effect of the different IQ distributions between cases and controls.

The sensitivity and specificity for each genetic score was similar in both analyses. The resultant ROC curves for the genetic score estimated in these analyses produced an AUC of 0.59, significantly different (P < 0.01) from a random prediction of the score (AUC = 0.50). These results may be considered a first step toward the goal of using common genetic variants at multiple loci to develop a multigene approach to predicting the risk of autism. Indeed, these data are in line with other multigene risk assessment models in other complex diseases, such as type 2 diabetes, with reported AUCs of approximately 0.6 [15, 16, 19, 56] using 3 to 18 common genetic variants. Results from these studies provide evidence that accumulation of multiple risk alleles markedly increases the risk of diabetes. However, compared with known nongenetic risk factors, such as body mass index, age and family history of the disease, combination of risk alleles provides only a relatively small increase in risk prediction [19, 56, 57]. In the case of autism, no early predictors exist to compare with genetic score. Nevertheless, there is significant interest in the early identification of infants at higher risk for autism because studies have shown that early intervention leads to significantly improved long-term outcome for the whole family [58, 59].

The AUC is not a clinically relevant way of summarizing predictive performances. Furthermore, the AUC is a poor metric for evaluating markers for disease diagnoses, screening or prognosis [60]. Using a threshold of six risk alleles allows identification of 18% of cases (19% and 17% in the AGRE and Seattle samples, respectively) while reducing the likelihood of false-positive results (around 8%). Importantly, these sensitivity and specificity values translate into a > 2-fold increased risk for autism. In such a disorder, defined entirely on the basis of child behavior, this can be considered an important first step toward improved risk prediction.

Several studies support a strong genetic contribution to the etiology of autism, including twin studies that estimated a heritability of > 90% [61, 62] and a family study that indicated a sibling recurrence risk ratio of 22 [63], which indicate that there is much potential for genetic prediction in autism. In the current study, the four SNPs used in the score are associated with a sibling recurrence risk ratio of 1.8. These results demonstrate that up to 10% of the variation in risk could be explained by these four genes.

Several strategies exist to identify relevant variants to be combined in a genetic score. These range from conservative approaches using only well-established factors [16, 56, 64, 65] to more liberal approaches, which include polymorphisms from candidate genes or from regions that have already been associated with the disease or that display functional relevance [20, 22, 66]. We selected a subset of four variants from four candidate genes already associated with autism. An additional strategy would have been to include genes identified in a genome-wide association study [43]. The approach described here could be applied to other established variants and be part of future studies. Only one recent paper reported genome-wide scan results, which identified a common genetic variant reaching genome-wide significance located on 5p14.1 in the intergenic region between the CDH10 and CDH9 genes [67]. The reported ORs for the associated variant were similar to those of the variants in this study. However, it would be interesting to add this variant into the risk score. In our study, we chose the risk allele count as the most commonly used method of constructing genetic scores from low-effect genetic variants [15, 16, 20, 24]. We used a recessive model to calculate the effect of ATP2B2, thereby strengthening our genetic score. Weighted scores have been proposed to increase discriminative power, but models that introduce information from variants with similar, small or moderate effect size only slightly increased the AUC [19, 55].

One limitation of our study was that we assumed an individual effect for each gene. Gene by gene and/or gene by environment interaction should be tested and could be introduced into the genetic score for a more accurate assessment of risk associated with a given genetic score. Another limitation was that the polymorphisms used are probably not causal variants but rather markers in linkage disequilibrium. Causal variants may have different allele distributions in siblings with and without autism, and could have a stronger effect on disease risk. This is not yet clearly established and needs to be clarified [68].

Surprisingly, for three of the four loci investigated in this study, the major allele was overtransmitted in cases versus controls. Although it is unclear whether this same effect may hold for other autism-related risk loci, the observation is interesting and is one that merits further consideration in the field.

In conclusion, the results of this study suggest that the joint analysis of common-risk variants will be more effective in identifying subgroups of individuals at risk than using single polymorphisms for risk assessment. The developed score could be improved by adding additional genes and environmental factors in the future.