Common genetic variants, acting additively, are a major source of risk for autism

DNA samples from SSC and AGP family members genotyped on the Illumina Infinium^® 1Mv3 (duo) microarray or the Illumina Infinium^® 1Mv1 microarray were analyzed here. Specifically qualifying SSC samples were genotyped on the Illumina Infinium^® 1Mv3 (duo) microarray (71.8%) while most AGP samples were genotyped on the Illumina Infinium^® 1Mv1 microarray (98.7%). Both arrays genotype roughly 1,000,000 single nucleotide polymorphisms (SNPs) and the overlap between the SNP sets is almost perfect.

The SSC sample [19] includes >2,000 genotyped families. However, our analyses targeted a homogeneous subset of these data. First, we included only samples genotyped on an Illumina 1M array; families had to be ‘quads’ consisting of an unaffected mother and father, an affected proband and an unaffected sibling; and all members of a quad had to have complete genotypes (>95% completion rate). Only samples of European ancestry were included. European ancestry for the SSC families was determined using GemTools [22, 23] for all available SSC probands. To conduct the ancestry analysis we selected 5,156 SNPs with at least 99.9% calls for genotypes, had minor allele frequency MAF >0.05, and were at least 0.5 Mb apart. Individuals were clustered into nine ancestry groups based on four significant dimensions of ancestry. The central five clusters, which held a total of 1,686 families, were identified as being of European descent. The ancestry cluster information combined with complete genotype information yielded a total of 965 SSC families for the analysis.

The AGP Stage 1 dataset [16, 20] comprised 1,471 families, of which 1,141 were previously identified to be of European ancestry [20]. European ancestry was confirmed by analyses identical to those applied to the SSC families (see Additional file 1: Figure S1).

Probands for the SSC and AGP cohorts were diagnosed in a similar manner (for diagnostic protocol for SSC, see [19]; for AGP, see [16, 20]). All SSC parents were screened for Autism Spectrum Disorder by the Broad Autism Phenotype Questionnaire [24] (self-report) and the Social Reciprocity Scale - Adult Research Version [25] (informant report). Moreover, family history evaluation excluded first-, second-, or third-degree relatives who met diagnostic criteria for ASD or intellectual disability. For AGP families 46.2% were known to be multiplex, another 38.2% were identified as simplex on the basis of a family history indicating no known first- to third-degree relatives with ASD, and the remaining 15.6% were of unknown status. Note that most AGP parents were not systematically evaluated for ASD, unlike those from the SSC, and when AGP parents were systematically evaluated, the results were not used to screen out affected individuals and thus multiplex families. In addition, while all available SSC family members were genotyped, only parent-proband trios were genotyped for the AGP even when additional siblings were available.

Controls derived from a convenience sample, specifically 1,663 individuals from HealthABC [26]. Control samples were also genotyped on the Illumina Infinium^® 1Mv3 (duo) array, like most of the AGP data, providing excellent comparability with the case dataset. Moreover, we reasoned that ASD is sufficiently rare (approximately 1% [27]) that screened and unscreened controls would yield similar results.

To make heritability estimates comparable, we filtered all families and control subjects based on the following criteria: all were of European descent as determined by genetically-estimated ancestry (Additional file 1: Figure S1); genotypes for all family members met stringent quality control (QC) criteria; and control samples met identical QC criteria.

To evaluate the robustness of the results, 1,986 individuals of European descent from the Neurogenetics Research Consortium [32] (NGRC) were available through dbGap [33] and used as a second control sample. For the NGRC study, genotypes were produced using the Illumina Infinium^® Human Omni2.5 microarray. Therefore, to combine all four data sets, we performed QC on 444,200 SNPs genotyped on all platforms, yielding 391,425 SNPs for analyses.

To explore the impact of different cohorts and genotyping protocols on estimated heritability, we conducted a series of contrasts between SSC and AGP samples of the same relationship type – contrasting probands, mothers, fathers, and pseudo-controls – as well as HealthABC versus NGRC controls.

While 713,259 SNPs were used for primary analyses, they constitute a small fraction of the SNPs in the human genome. Hence the heritability presented could underestimate total heritability. On the other hand, because genotypes of SNPs in close proximity tend to be correlated due to linkage disequilibrium, it does not follow that the coverage of the genome by the SNPs used here estimate only a small fraction of the heritability. To determine the shortfall in “genomic coverage” and how it impacts estimates of heritability, we performed an experiment using data from the 1,000 Genomes project [34], under the assumption that coverage of common variants in the 1,000 Genomes data is perfect. Assessing all SNPs genotyped in our data, as well as subsets thereof, we estimated heritability of liability. Using the same subsets, but in 1,000 Genomes subjects, we estimated levels of genomic coverage. We can then relate estimated heritability to genomic coverage to develop a functional relationship between the two.

We performed the experiment assessing “genomic coverage” as follows. We assumed genomic coverage of SNPs with MAF > 0.1 would be essentially complete for the 379 European samples analyzed by the 1,000 Genomes project. From these genomes we selected 50 1Mb regions in which at least 500 SNPs in the 1,000 Genomes samples had MAF > 0.10. Coverage of these regions by the 713,259 SNPs was calculated as a function of the number of other SNPs with MAF > 0.1 that were tagged by (correlated with) them; call the set of M = 713,259 SNPs “tagSNP”. The tagging evaluation was implemented using Hclust [35]. Forcing tagSNP to be in the set of selected tag SNPs from the region, Hclust evaluated how many more independent SNPs N were required to cover the region when the minimum linkage disequilibrium [36] r² amongst tags could be no less than X, where X = {0.5, 0.7, and 0.9}. Then, for each value of X, M/(M+N) estimates the coverage. Next we randomly sampled 50, 25 and 12.5% of the 713,259 SNPs (356,630, 178,315, and 89,158 SNPs respectively) five times and each time estimated coverage for these subsets.

The research described here is in compliance with the Helsinki Declaration, including appropriate informed consent or assent [16, 19, 20, 26, 32, 33].

Heritability of SSC probands, measured against HealthABC controls, was found to be 39.6% (Figure 1A, Table 1). SSC mothers, fathers and siblings, when contrasted to controls, yielded an estimated heritability approximately half that of probands (Figure 1A, Table 1), consistent with expected values from theoretical analyses of an additive model (Figure 1A). We also generate a “pseudo-control” from the alleles that parents did not transmit to their affected offspring by using the program Plink [37]. When these pseudo-controls were contrasted to the unrelated control sample they produce estimates roughly one-quarter of that identified in probands and close to the theoretical expectation, zero (Figure 1A), demonstrating that the probands received the majority of risk alleles carried by parents.

When heritability is estimated using AGP probands (Figure 1B, Table 1), the point estimates are larger than those from SSC (h²=55.2% versus 39.6%) although the 95% confidence intervals overlap. Moreover the decline in heritability for AGP parents relative to probands is 30% (55% for probands, 37% for parents), instead of the 50% seen for SSC, and heritability estimated from pseudo-controls is also higher (38%), consistent with parental values (Figure 1B, Table 1). These results suggest that AGP parents carry a greater load of additive risk variants than SSC parents and thus are, on average, closer to the threshold of being affected.

A major difference between the SSC and AGP samples was the ascertainment and assessment process. SSC parents were systematically screened on two instruments to ensure they did not meet criteria for a spectrum diagnosis. Most parents from AGP families were not evaluated in this way, and a small fraction of those parents met criteria for ASD [9, 16]. While not as systematic as the SSC phenotyping assessment, most AGP families did have available information about simplex versus multiplex status. Consequently, we were able to compare heritability of probands from AGP multiplex versus simplex families (Figure 1D, Table 2). The former was estimated at 65.5% by comparison to HealthABC, whereas probands for AGP simplex families it was 49.8%. Thus estimates of heritability for AGP simplex probands are somewhat closer to those from SSC probands (Figure 1C) than to estimates for AGP multiplex probands. Moreover, for multiplex families and the mixed set of AGP families (simplex/multiplex/unknown), both the observed and expected heritability for first-degree relatives was higher than that seen in simplex families (Figure 1). These results comport with the literature showing that unaffected relatives from multiplex families tend to exhibit more features of the broader autism phenotype than relatives in simplex families [38‐40] (see Additional file 2: Table S1 for estimates from combined simplex samples).

A curious observation from AGP multiplex families was that fathers generate larger heritability than mothers. We reasoned that this could be explained by three plausible hypotheses: (1) the confidence intervals of the paternal and maternal estimates overlap, so there is no true difference; (2) the load of risk variants is, in fact, greater for AGP fathers; or (3) fathers carry a larger number of both liability and protective alleles. The last of these requires some elaboration. Males are at much greater risk for ASD than females (4:1 or greater) and parents carry additive risk factors, yet AGP fathers and mothers are largely unaffected. It is possible, then, that the increased allele sharing in unaffected fathers is due to a greater proportion of protective alleles, with females being resilient for some other reason (for example, estrogen/testosterone balance) in the face of a similar degree of genetic risk.

Our results support either the first or second hypotheses but are not consistent with the third. The first hypothesis is impossible to rule out given the limited sample size. For the second hypothesis, if AGP fathers were simply carrying greater risk, some of those additional risk alleles would be carried by the pseudo-controls and the heritability obtained from the contrast of probands and pseudo-controls should be substantially smaller than that observed from probands versus controls. Indeed the values are substantially smaller: 10.9% vs. 39.6% for SSC; 14.5% vs. 55.2% for all AGP; 0.0% vs. 49.8% for simplex AGP, and 27.1% vs. 65.5% for multiplex AGP. Finally, if (3) were true, then contrasting probands to pseudo-controls would produce substantial estimates of heritability because of the differentiation induced by protective alleles, but this is not observed.

If the additive variation for liability to ASD conforms to the traditional polygenic or infinitesimal model, then liability variants should be distributed at random over the genome. The implication is that if heritability were estimated for each chromosome, the resulting estimates should be correlated with the lengths of the chromosomes. On the other hand, if the heritability traced to a relatively small number of variants, even a few dozen, such a correlation would be unlikely. In fact, we observe significant correlation between per-chromosome heritability and chromosome length (Figure 2), both for simplex (r = 0.46, P value = 0.028) and multiplex (r= 0.54, P value = 0.0075) families.

In Figure 2 the deviation from prediction for chromosome X is surprising. For both multiplex and simplex families, heritability estimated from X is less than that predicted by its size. This is noteworthy because chromosome X has been cited as a possible source of sex-differential liability for ASD [41]. Our results suggest that common variants affecting liability do not cluster on chromosome X.

To evaluate the robustness of our results, we first contrasted the genotypes of SSC and AGP probands to a second large set of controls, 1,986 individuals from the Neurogenetics Research Consortium [32, 33]. These samples, genotyped on the Illumina Infinium^® Human Omni2.5, were filtered and subjected to QC in an identical fashion to the HealthABC control set. There was excellent agreement of heritability estimates for ASD from the two control samples (Tables 2 and 3) despite differences in ascertainment of the controls and the different genotyping platforms.

Next, the impact of different cohorts and genotyping platforms on estimates of heritability was explored by conducting a series of contrasts between SSC and AGP samples of the same relationship type: contrasting probands, mothers, fathers, and pseudo-controls. Note that most SSC samples were genotyped on the Illumina^® 1Mv3 (duo) microarray (71.8%) while most AGP samples were genotyped on the Illumina Infinium^® 1Mv1 microarray (98.7%). Contrasts between SSC and AGP samples of the same relationship type (Additional file 3: Table S2) produce estimates close to the difference between their control-based heritability. Indeed the estimates from direct contrasts were usually smaller than the difference of control-based heritability (for probands, 0.08 vs. 0.15 ≈ 0.552-0.396 from Table 1; for mothers, 0.11 vs. 0.17; for fathers, 0.19 vs. 0.17; and for pseudo-controls, 0.22 vs. 0.29). Thus these results are not consistent with effects attributable to genotyping platform or ascertainment beyond multiplex/simplex status. Implicit in these results is common genetic liability - SSC and AGP probands must share many liability variants despite their differences in ascertainment. Indeed when AGP multiplex probands are contrasted to SSC probands the resulting heritability is 0.23, quite similar to that expected by the difference in their estimated heritability (0.66 - 0.40 = 0.26); and when AGP simplex probands are contrasted to SSC probands, the resulting estimated heritability, 0.0, is below that of the difference in their estimated heritability (0.50 - 0.40 = 0.10). These results suggest that the difference between multiplex and simplex families is largely a matter of degree (see also [42]), namely the number of liability alleles carried by parents, rather than a fundamental difference in the genetic architecture [4, 43].

Given the remarkable similarities of heritability estimates obtained for either set of control samples (Tables 2 and 3), one might anticipate there would be little, if any, difference between these controls. When we contrasted these control samples, however, they produced a heritability of 26.5% (Additional file 3: Table S2). Mathematically, estimates of heritability arise from a high dimensional space of allele frequencies, phenotypes and their interrelationships. Therefore even if two controls groups evoke similar estimates of ASD heritability from the same sample of probands, the controls themselves need not be close in the multidimensional space of allele frequencies. What generates the differentiation between controls is unknown. It could arise from the different genotyping platforms or from differences in ascertainment. In light of this difference, the fact that both controls sets give rise to nearly identical estimates of heritability for all proband subsets is remarkable and suggests that the similarity amongst cases overwhelms differences between the controls.

There remains an unexplained feature of the results: estimates of heritability for pseudo-controls tend to be elevated over their theoretical values (Figure 1). Several genetic forces could be at play. The simulations to derive the distribution of liability in families also produce estimates for pseudo-controls. Those results show (Figure 1) that while the expected heritability for simplex families is zero, multiplex status raises the expected value to 20%. It is not unreasonable to assume that the simplex collections analyzed here contain families with unrealized multiplex potential, and that might be especially true for AGP families that had ascertainment criteria less stringent than those for SSC families.

A factor that will elevate the expected heritability in pseudo-controls is positive assortative mating (henceforth assortative mating). Assortative mating on phenotypes related to ASD liability has been previously reported [39]. When parents are genetically similar at liability loci and they bear affected offspring, their gametes will tend to be highly enriched for risk alleles, even those that are not transmitted to affected offspring. Simple simulations mimicking assortative mating show that it can exert an impact similar to the difference between simplex and multiplex status. When simplex probands had heritability of 50% (that is, simulation a in Methods), the expected heritability of pseudo-controls was 11.3% – versus 0% without assortative mating. When multiplex probands had heritability of 75%, the expected heritability of pseudo-controls was 42.8% – versus 20.2% without assortative mating. These simple experiments were not intended to cover the range of plausible scenarios for assortative mating relevant to ASD, which would be impossible, but rather to demonstrate the effect of such mating on the nature of pseudo-controls. Thus assortative mating could be an important and salient source of enrichment. Whether these forces explain all of the elevated heritability for pseudo-controls will require further data and analyses.

Because the set of SNPs used for primary analyses constitute a small fraction of the SNPs in the human genome, estimates of heritability (Figure 1) could be biased downward. Still, due to linkage disequilibrium, the degree of bias is not trivial to estimate. Therefore we performed an experiment to evaluate the shortfall in genomic coverage and how it impacts estimates of heritability. Results from the experiment are shown in Additional file 4: Figure S2, in which estimated heritability was plotted against estimated coverage. These results suggest that heritability estimates from probands, as shown in Figure 1, are good approximations. They represent only slight underestimates of what would be obtained had the entire genome been sampled.

In total our results demonstrate that a substantial portion of ASD liability arises from inherited variation acting additively. This pattern holds both for simplex and multiplex families, with the burden of liability greater in multiplex families, consistent with theoretical and empirical [38‐40] results. The modeling reported here does not differentiate between additive effects due to common versus rare variation. Nonetheless it is reasonable to assume that most of the estimated heritability traces to common variants because linkage disequilibrium between the common variants analyzed and rare liability variants should, on average, be small [44]. Thus the additive contribution of rare variants to ASD liability is likely underestimated. Imperfect coverage must also have an impact, but our analyses suggest its impact is not large (Additional file 4: Figure S2).

Our analyses cannot address other features of the genetic architecture of ASD, including non-additive genetic effects, which add to ASD’s broad-sense heritability [45], and de novo mutations. In addition, because they underestimate the impact of rare inherited variation, they differ from family-based estimates, such as from twin studies, that do capture these effects. Still our findings of substantial heritability are consistent with the majority of twin studies [1, 2] and are richer in some ways because the analytic technique [17, 18] used here provides a direct estimate of the proportion of liability attributable to additive genetic effects, whereas twin studies obtain their estimates by relying on assumptions that are approximations. For example, Zuk et al. [45] point out that non-additive genetic effects are almost surely a component of the genetic architecture of any trait, but these effects cannot be captured by twin designs. Yet for autism and other psychiatric disorders non-additive genetic effects could be an integral component [46‐48]. Twin designs also fail to capture other features, such as maternal effects [49] and de novo mutations, which are an important component of ASD genetic architecture [4‐11].

A recent ASD twin study [12] estimates 38% of ASD liability traces to additive genetic effects while 55% traces to common environment. Our point estimates would be close to theirs if ascertainment of their families was like that for SSC families, but not like that for AGP families. A substantial fraction of their dizygotic twins, however, are multiplex for ASD. Thus their point estimate for heritability from additive genetic effects is low relative to ours. If rare inherited variation contributes substantially to liability for ASD, this makes the 38% estimate seem lower still because twin studies should capture these effects whereas our estimates cannot.

Genomewide association studies [18, 50‐52] have detected only a handful of SNPs, all of small effect and none replicating reliably. Teaming this observation with our estimates of heritability (Figure 1) and the fact that these studies are underpowered to detect genetic variants of small effect size, but are otherwise well powered [15], we conclude there must be thousands of SNPs scattered across the genome with common liability alleles. Analyses of chromosome-specific heritability support this conclusion (Figure 2). Employing analyses like those proposed by Stahl et al. [53] could estimate this distribution of effects.

Because these loci have small effect, samples far larger than exist today will be required to identify a substantial fraction of them using standard genome-wide association methodology. Hence, for the immediate future, ample “missing heritability” for ASD will remain. Ingenious designs will be required in the near term [54] to identify SNPs affecting liability. In the longer term GWAS of a large number of ASD subjects, at least on the order of that performed for schizophrenia [55‐57], should be one of the priorities for the field of ASD genetics.

One way forward is to exploit shared liability across psychiatric disorders, taking advantage of larger samples [58] afforded by cross-disorder meta-analysis. There is now sound evidence for common variants affecting liability for schizophrenia [55‐57], including a study similar to ours [46]. Given the documented sharing of rare variants affecting risk for both disorders (for example [59]), it would not be surprising to find that some common variants affect liability to both schizophrenia and ASD.

The estimated heritability for schizophrenia using methods similar to ours is 23% [46]; for bipolar disorder and similar methods it is 40% [60]; and for major depression it is 32% [61]. None of these studies separate out simplex and multiplex families, so in that sense they are most comparable to the estimate obtained over all AGP families, 55%, although the representation of multiplex families in the AGP sample is likely larger than for the other samples. Regardless of the differences in simplex/multiplex representation, these estimates are stochastically similar, in view of their standard errors, emphasizing that common variants affect liability for most if not all psychiatric disorders. Moreover their impact appears to be similar in magnitude across disorders, as measured by heritability estimated from common variants.

That ASD shows the largest estimated heritability is notable and could reflect the fact that the sibling recurrence risk is, on average, higher for siblings of an ASD proband than for siblings of probands diagnosed with schizophrenia, bipolar disorder or major depression. Sibling recurrence risk is a ratio, defined as the probability of a sibling being affected, given that the proband is affected, divided by the prevalence of the disorder in the general population. Recent studies put this recurrence risk at almost 20 for ASD [62], whereas for schizophrenia it is 6 to 10 fold [63], for bipolar disorder it is 4 to 10 fold [64], and for major depression it is roughly twofold [64]. The larger heritability could also trace to differences among studies. It is possible that our estimates of heritability are inflated by unknown differences between our case and control samples, including ascertainment biases and genotype quality. Regarding the latter, we selected case and control samples genotyped on the same genotyping platform to minimize differences and we did not detect any large differences in allele frequencies, but we cannot rule out subtle differences in quality.

Regarding identification of common variants affecting liability, our results suggest that the contrast of case and pseudo-control genotypes, the “family-based” analysis, is not optimal. In many samples pseudo-controls carry a substantial burden of risk variants and their presence degrades the power of family-based analysis to detect risk SNPs (see also [30, 31]). Instead it appears that population-based controls contrasted with ASD cases would be a more powerful design [65], even after adjusting for ancestry [66]. In this regard it is intriguing that the earliest GWAS of ASD [50] used population-based controls to identify a single locus at 5p14.1, and this result has since garnered support from a functional study that reveals a plausible biological link to ASD liability [67].

The genetic architecture of ASD has numerous components: additive, non-additive and de novo genetic effects, as well as gene-gene and gene-environment interactions. The results shown here are relevant to only one of these components. Other components, such as de novo events, are also known to make a substantial contribution to liability [4‐11], while others remain to be thoroughly investigated [45]. Already analyses of rare variation of major effect has revealed a substantial number of genes affecting liability [8‐11, 68‐70]; it is reasonable to predict that common variants regulating expression of those ASD genes could also affect liability [71]. We hypothesize that the interplay of rare and common variants is critical not only to liability itself, but to the expression of ASD or other relevant psychiatric and developmental disorders. The dynamics of this interplay will likely be an important area for future autism research.