Rare coding variants of the adenosine A3 receptor are increased in autism: on the trail of the serotonin transporter regulome

To test for association of common alleles at ADORA3 in ASD, we selected SNPs to represent all common haplotypes (that is, ≥5%) across the transcriptional unit and flanking sequence taking into account linkage disequilibrium patterns. We used Haploview (http://​www.​broadinstitute.​org/​haploview) [37] to analyze CEPH (CEU; Caucasian) SNP data from the HapMap database (http://​www.​hapmap.​org), given that the vast majority of the sample was of European ancestry. The Haploview implementation of Tagger identified tag-SNPs to capture common alleles across ADORA3 at an r² ≥0.8 and minor allele frequency (MAF) ≥5%. TaqMan™ allelic discrimination assays for four SNPs that span 14.4 kb were obtained from Applied Biosystems (ABI, Foster City, CA, USA) as Assays-on-Demand (AoD) or designed as Assays-by-Design.

PCR amplification was conducted in a 5-μL volume in accordance with manufacturer’s recommendations. In brief, cycling conditions included an initial denaturation at 95°C for 7 min, followed by 50 cycles of 92°C for 15 s and 60°C for 1 min. Post-PCR allelic discrimination was conducted using an ABI 7900HT genotypes ABI Sequence Detection System software. Genotypes were checked for completeness (≥98%) and conformity to expectations under Hardy Weinberg Equilibrium (HWE). Other quality control procedures included inter- and intra-plate replicates and checks for Mendelian inconsistencies using PEDCHECK [38]. Families containing Mendelian inconsistencies were identified and excluded from additional analysis. Family-based allelic association testing was used to evaluate transmission of alleles in the autism families being studied. Single marker analysis was conducted using the family based association test (FBAT) [39]. FBAT analysis was conducted under the additive model, and significance was determined using the empirical variance (−e) option, as this provides a more conservative estimate of association, given the presence of multiplex families in the dataset. Statistical power required to detect a meaningful association was determined by power calculations using the Genetic Power Calculator (http://​pngu.​mgh.​harvard.​edu/​~purcell/​gpc/​) [40]. We assumed ASD as a discrete trait with a prevalence of 1/100 and a sample size of 1,000 trios, further assuming MAFs of 5%, 10%, and 30%, respectively, and a D´=0.8. Based on these calculations, we would have ≥80% power to detect odds ratios (ORs) of ≥1.59, ≥1.44, and ≥1.36 for (risk) allele frequencies of ≥5%, 10%, and 30%, respectively.

To permit association analysis within the major European subset of the family-based sample and for subsequent matching of cases and controls (see below), ancestry was determined using STRUCTURE [41] and multidimensional scaling (MDS) in PLINK [42] (http://​pngu.​mgh.​harvard.​edu/​~purcell/​plink) to analyze genome-wide parental (founder) genotype data [35, 43] from autism families. Genome-wide (GW) genotypes were derived from different platforms and in substantially different numbers. Many multiplex families were genotyped by the Autism Genome Project (AGP) in its Phase I linkage study using the Affymetrix 10 k SNP platform [44], and 10 k data were analyzed at that time using STRUCTURE. Other families had genotypes from Illumina 550 k [43] and/or 1 M SNP [35] arrays, and these were analyzed recently using MDS. Fortuitously, numerous families had both sets of genotypes and analyzed using both applications, and identical ancestry classifications provided confirmation that both STRUCTURE and MDS yielded consistent and robust assignments that also agreed with self-report information. A small number of families from the overall association sample (1) did not have GW genotypes, in which case self-report information determined classification or (2) had neither GW genotype data nor self-report information, in which case they were classified as being of ‘unknown’ ancestry.

Initial screening for sequence variants at ADORA3 utilized whole blood derived DNA samples from 185 unrelated (predominantly Caucasian) cases (94% Caucasian; 5% African-American; 1% Hispanic). Non-clinical comparison samples were drawn from reference collections and consisted of lymphoblastoid cell lines DNA from 305 subjects: (1) 96 samples from the ‘Caucasian’ subset of the Coriell Human Genome Diversity Panel (http://​ccr.​coriell.​org/​sections/​Search/​Panel_​Detail.​aspx?​Ref=​HD100CAU&​PgId=​202); (2) 192 samples from the Human Random Control (HRC) collection corresponding to subjects of European ancestry recruited in the UK and obtained from Sigma-Aldrich (St. Louis, MO, USA; http://​www.​sigmaaldrich.​com/​life-science/​molecular-biology/​pcr/​human-genomic-dna.​html); and (3) 24 subjects from the neurologically-normal NINDS/Coriell African-American panel (http://​ccr.​coriell.​org/​Sections/​Search/​Panel_​Detail.​aspx?​Ref=​NDPT111&​PgId=​202). Thus, the comparison samples were 92% Caucasian and 8% African-American. This discovery sample of cases and controls provides 80% power to detect rare risk variants of 0.1%, 0.5%, 1% allele frequencies with ORs of 16.2, 5.25, and 3.71, respectively. Our a priori expectation is that functional ADORA3-containing, SERT-altering functional variants would confer large genetic effects, and thus we are reasonably powered under this assumption. However, we would be significantly underpowered if hypothetical rare variants conferred ORs <2.

PCR amplifying primers (Additional file 2: Table S2) were designed using Primer3 (http://​frodo.​wi.​mit.​edu/​primer3/​), and all potential amplicons were subjected to a BLAST-Like Alignment Tool search to ensure no significant matches existed elsewhere in the genome. Amplicons were optimized using a gradient of annealing temperatures, followed by agarose gel electrophoresis to determine ideal annealing temperature. PCR was then performed on the case (185) and control (305) screening samples. PCR reactions contained 7.1 nM of amplifying primers, 10 μL of 2× Mastermix, and 12 ng of DNA in a final volume of 20 μL. ADORA3 primers were designed for exons 1 to 2 and >500 bp of promoter sequence. PCR amplification conditions were 7 min at 95°C, 40 cycles of 15 s at 95°C, 12 s at the annealing temperature, and 60 s at 72°C, with a final extension at 72°C for 7 min. Following PCR, each product was examined by gel-electrophoresis, and once verified for expected size and specificity, excess primers and nucleotides were removed using Millipore 96 well filter plates. Samples were quantitated, subjected to Sanger sequencing, and data analyzed for variants using Sequencher v4.9 (Gene Codes Corporation, Ann Arbor, MI, USA). Specific variants were confirmed by independent PCR and sequencing, and for ASD cases with simultaneous determination of segregation by inclusion of other available family members, affected or unaffected.

Variants were assessed using multiple approaches. We noted whether or not they were previously documented in dbSNP or the 1000 Genomes project. In silico algorithms PolyPhen2 [45], SIFT [46], and SNAP [47] were used to provide a bioinformatic estimate of whether amino acid substitutions were likely to be ‘damaging’ or ‘not tolerated’ or benign. Cross-species conservation of the amino acid residue and surrounding sequence was also evaluated, along with available literature regarding structural features of the A3AR protein. Finally, to ask whether there might be evidence to support a global burden of rare (<1%) coding variants, the cohort allelic sums test (CAST) was used. CAST is a grouping method in which the number of individuals with one or more variants in a gene is compared between affected and unaffected individuals [48, 49]. Thus, we compared differences in rare allele counts in matched cases and controls in a 2 × 2 contingency table, and a Fisher’s Exact test was performed.

An independent replication sample for comparison to variation identified in the discovery phase corresponded to ASD cases and controls from the NIMH Repository. Cases and controls were pair-matched for ancestry, and samples used for whole exome sequence analysis as recently described [50]. SLC6A4 coding variants (described above) that affect SERT function and regulation are associated with not only ASD but also OCD, anxiety and mood disorders [17, 20, 21]. Therefore, case-control pairs were excluded if control subjects had any history of OCD or anxiety disorders. Sequence data at ADORA3 for a total of 339 case-control pairs were examined for ‘functional’ (that is, missense, nonsense, consensus splice site, and read-through) variants. Details regarding quality control procedures and read depth for case and control samples are detailed in the Supplemental Material of Neale et al. [50].

To test whether common alleles at ADORA3 contribute to ASD risk, we genotyped a sample consisting of 958 autism families using four SNPs representing common haplotypes that span the ADORA3 locus. Primary FBAT analyses of genotype data were designed to test allelic transmissions along two axes of stratification: strict vs. spectrum diagnostic classifications and European vs. all ancestries, yielding four primary analyses. Analysis for each of the resulting four strata (strict-European and strict-all ancestries; spectrum-European and spectrum-all ancestries) showed no evidence to support common variant association at ADORA3 using data for the four SNPs tested (Additional file 3: Table S3). Analysis of haplotypes (>5%) across the locus that were captured by these SNPs also showed no significant association, consistent with single marker analysis (data not shown).

To identify potentially functional, risk variants in ADORA3, we screened all exons and the promoter region in 185 unrelated ASD cases and 305 non-clinical comparison samples by direct Sanger sequencing. Comparison samples were selected to ethnically match case samples for Caucasian and African-American subjects, which represented the vast majority of the case screening panel (see Methods). Multiple synonymous and non-synonymous variants were identified and are documented in Table 1. Three novel variants, two non-synonymous (ns) and one synonymous, were detected in cases (Leu90Val, Val171Ile, Cys194) but were not previously documented in dbSNP or 1000 Genomes (1 kG) [51]. A novel Ala195Thr substitution was identified in a single control sample. Experimental validation of variants by sequencing independent PCR products included parents and other siblings (when applicable). No de novo variants were detected, and thus all were inherited in ASD families. Bioinformatic algorithms PolyPhen2 [45], SIFT [46], and SNAP [47] that predict the likelihood of functional effect of coding variants on protein function were applied to rare (<1%) ns-variants identified. Val171Ile (detected only in cases) was predicted to be ‘damaging’ or ‘not tolerated’, however, Leu90Val along with Ala195Thr were predicted to be benign substitutions. We note that subsequent to our discovery of novel Leu90Val and Ile171Val variants, but during subsequent functional and follow-up genetic experiments, both variants emerged from 1 kG sequence data and were thus deposited into dbSNP, as reflected in Table 1.

We sought to determine whether there was evidence for an increase of rare ADORA3 coding variants in cases compared with controls. Both Leu90Val and Val171Ile were detected only in ASD cases and not controls (3/370 case chromosomes for each variant vs. 0/562 control chromosomes). A case-control comparison of these individual rare alleles does not reach significance given the small number of observations (Fisher’s Exact P=0.064 for Leu90Val and Val171Ile each). Given the inherent limitation in power to compare rates of single rare alleles, a gene-wide discovery model suggests that a better approach is to model all ‘functional’ (that is, non-synonymous, consensus splice site, and read-through) rare variants simultaneously for case-control comparisons. Therefore, we employed the Cohort Allelic Sums Test (CAST) to test for an overall increased burden of rare (<1%) nsSNPs in ADORA3 in ASD. We reasoned that any risk effect of higher frequency, common coding variants would be indexed by SNP and haplotype-based association studies described above. Allele counts obtained from sequence discovery in 185 cases and 305 controls were tabulated for Leu90Val, Val171Ile, and Ala195Thr, and a Fisher’s Exact test was conducted (Table 2). This analysis showed a significant burden effect (6/185 ASD individuals vs. 1/310 controls; P=0.013; OR=10.19, CI = 1.20-81.56). We note that, as described in Methods, we have low power to detect rare variants of only modest risk effect. Nevertheless, we observed a nominal increase in rare ‘functional’ variants in the discovery case sample vs. controls, and this prompted us to: (1) conduct functional studies of the novel variants found in ASD cases for effects on SERT; and (2) subsequently compare the putative increase of rare, ‘functional’ variants in ASD cases to data from exome sequence that became available from the NIH ARRA Autism Sequencing Consortium [50, 52]. Consistent with published data on accepted risk CNVs (for example, 16p11.2, 1q21.1, 22q11.2, 22q13.3, and so on) [53‐55], novel variants Leu90Val and Val171Ile did not always segregate to only (or all) affected individuals in a family, consistent with incomplete penetrance of these variants (Figure 1).

The availability of crystal structure for a human ligand-bound adenosine A2a receptor (A2aAR) [56] provides an important source of structure-function information that can inform predictions of a potential functional impact of the Leu90Val and Ile171Leu variants. Figure 2 depicts the structure of the A2aAR, modeling the positions of residues equivalent to Leu90 and Val171 in A3AR, in relation to bound adenosine. Our model predicts that the A3AR residues Leu90 and Val171 flank the adenosine-binding site, supporting a hypothesis that one or both of these variants may affect A3AR function. Leu90 and Val171 show consistent cross-species conservation in mammals, while Ala195 is less conserved (Figure 3). Functional studies therefore focused solely on the two variants identified in cases.

To test for potential functional effects caused by the Leu90Val and Val171Ile substitutions, we engineered human A3AR cDNA expression constructs to harbor either Leu90Val or Val171Ile variants for experimentation in a heterologous transfection system using CHO cells. In prior studies, we have found that CHO cells support both the investigation of A3AR regulation of SERT as well as studies of A3AR/SERT physical association [29, 30]. Western blot analysis of cell lysates using an anti-myc probe confirmed equivalent expression of myc-tagged wildtype and both variant A3AR constructs, Leu90Val [29] and Val171Ile (data not shown). We first asked whether either variant (G-protein coupled) receptors produced increased (or otherwise altered) levels of cGMP that might result in a downstream increase of SERT activity. The human A3AR and SERT constructs were co-transfected into CHO cells, and we assessed basal cGMP levels for both variants, and then following activation of receptors with the A3AR selective agonist IB-MECA. Figure 4A (at 0 min) demonstrates that CHO cells co-transfected with the Leu90Val variant display significantly elevated basal cGMP compared to wildtype A3AR co-transfected cells (L90V: 207.85% ± 45.71 vs. WT: 100.0% ± 0.02; P=0.015, n=3). Stimulation of the Leu90Val A3AR by IB-MECA (1 μM) revealed that the elevated cGMP production seen in the basal state persisted over the full time course and paralleled wildtype A3AR with both returning to their respective basal states after 40 min (Figure 4A and Additional file 4: Figure S1A; L90V: 163.3% ± 29.15 vs. WT: 100.0% ± 6.32; 1-tailed t-test P=0.049, n=3).

In contrast with Leu90Val, cells co-transfected with Val171Ile-A3AR/SERT displayed basal cGMP levels similar to those observed in wildtype A3AR co-transfected cells (V171I: 116.1% ± 16.3 vs. WT: 100.0% ± 9.2, n=3). Additionally, IB-MECA treatment in Val171Ile co-transfections failed to show any stimulation of cGMP production that was seen with wildtype A3AR. This inability of the Val171Ile receptor to be stimulated by IB-MECA appeared as a slightly diminished cGMP synthesis over the time course compared with wildtype, although this difference was not statistically significant (Additional file 4: Figure S1A; V171I: 90.14% ± 12.22 vs. WT: 100.0% ± 16.18; 1-tailed t-test P=0.303, n=3).

The IB-MECA induced stimulation of cGMP production in wildtype A3AR/SERT co-transfections displayed in Figure 4A paralleled the expected stimulation of SERT-mediated 5-HT uptake activity by wildtype A3AR (Figure 4B and [30]). Although the elevation in peak 5-HT uptake (that is, at 10 min) in Leu90Val A3AR/SERT co-transfections was not significantly greater than that for wildtype A3AR (Figure 4B; L90V: 148.5% ± 7.1 vs. WT: 140.3% ± 4.5; n=4), the IB-MECA-stimulated increase in 5-HT uptake activity mediated by the variant receptor persisted above that seen in cells transfected with the wildtype receptor throughout the time of our observations. Indeed, the significantly increased 5-HT uptake, seen at 40 min for Leu90Val A3AR transfected cells, the time by which wildtype A3AR-stimulated SERT activity had already returned to baseline levels (L90V: 131.3% ± 7.1 vs. WT: 108.3% ± 3.7; two-way ANOVA P=0.001, n=4) did not return to wildtype levels until the next time point measure, 20 min later (Figure 4B). These observations reflect an overall greater 5-HT uptake in Leu90Val-A3AR cells compared to wildtype A3AR (Additional file 4: Figure S1B; L90V: 112.2% ± 8.80 vs. WT: 100.0% ± 3.22; 1-tailed t-test P=0.020, n=4).

In striking contrast to the enhanced SERT activity observed with Val90 co-transfected cells, IB-MECA treatment of Ile171, compared with wildtype co-transfected cells failed to stimulate SERT activity (Figure 4C) across all agonist doses tested (for example, 1 μM IB-MECA Val171Ile: 101.8% ± 12.5 vs. WT: 136.3% ± 7.2; two-way ANOVA P=0.005; n=3-4).

With support from both initial Sanger discovery experiments and subsequent functional studies that potentially ASD-related, functional variation had been identified in ADORA3, we sought replication of the putative increase of such variants in ASD cases vs. controls. We examined whole exome sequence data from 339 cases and pair-matched controls of European ancestry generated by the NIH ARRA Autism Sequencing Consortium [50] for the presence of functional variants at ADORA3 (see methods for exclusion criteria). Nine non-synonymous and numerous synonymous variants were discovered in these pair-matched samples that were completely independent from the discovery sample. In total, nine of 12 subjects harboring non-synonymous variants were detected in cases, and the remaining three in controls (9/339 ASD individuals vs. 3/339 controls); although that relative increase was not statistically significant (Fisher’s Exact 1-tailed P=0.071; OR = 3.06, CI: 0.82-10.99; Table 3). Specifically, Leu90Val was found in one additional case and no controls, while Val171Ile was found in two additional cases but also one control. Of the remainder, four additional missense variants (Ile22Thr, Phe48Ser, Ala69Ser, and Leu294Thr) were found in one case each and no controls, and a single read-through mutation (*319Gln) predicted to add an additional 38 residues (>10% of the native protein) was found in two cases and zero controls. Two additional missense variants were detected in controls (Phe180Leu and Ala273Thr). Each of these variants passed stringent quality control thresholds and have read-depths that, based on empirical data, have an extremely high positive predictive value for being valid calls [50], although they have not been experimentally validated. Prediction of functional consequences using PolyPhen2 [45] indicated that two missense variants were forecast to be benign (Ala69Ser and Leu294Thr) and four missense to be damaging (Ile22Thr, Phe48Ser, Phe180Leu, and Ala273Thr). Combining both discovery and replication findings strengthens evidence for an increase of ‘functional’ variants in ASD, with counts of 15 of 524 cases compared with four of 644 controls harboring such variants (Fisher’s Exact 1-tailed, Bonferroni-corrected P=0.0025; OR = 4.72, CI: 1.56-14.30). Moreover, taking into account replication data, the enrichment of Leu90Val in ASD cases vs. controls becomes statistically significant (Fisher’s Exact 1-tailed P=0.040; 4/1,048 ASD vs. 0/1,288 control chromosomes). Although present in greater numbers in ASD cases, enrichment for Leu171Ile by itself did not reach significance (Fisher’s Exact 1-tailed P=0.068; 5/1,048 ASD vs. 1/1,288 controls).

We explored the possibility that dimensions of the ASD phenotype might stand out in affected carriers of Val90 and Ile171 or other ADORA3 variants. We examined domain and item-level data from the ADI-R to explore the possibility that certain ASD features might be more pronounced in ADORA3 rare variant-carriers. Such comparisons are inherently limited by a small number of carriers relative to hypothetical phenotypic effect sizes and different mechanistic effects on A3AR for a given variant. Our examination of available data showed no consistent pattern of elevated or diminished ADI-R domain scores in carriers vs. non-carriers (Additional file 5: Table S4). We found a similar lack of correlation upon examination of ADI-derived principal components analysis-derived scores [57]), as a complementary set of dimensional ASD traits (data not shown).

To determine whether rare ADORA3 variant carriers also harbored other known or likely risk factors, we examined data from published work describing CNVs in respective samples [58, 59]. Some inherited CNVs were identified in the six cases for which CNV data were available; for the remaining nine samples, CNV data were not available and several of these have not been subjected to array based genotyping or analysis. Most CNVs corresponded to regions also detected in controls, however, a single de novo duplication of approximately 211 kb (hg18: ChrX:153263157–153474401) was found in the male proband (NIMH ID: 217-14216-3470) harboring a Ala69Ser substitution in ADORA3.