Evaluation of common genetic variants in 82 candidate genes as risk factors for neural tube defects

The primary sample set (320 NTD case families and 341 controls) was genotyped for 1517 tagging SNPs intended to capture common genetic variation in 82 candidate genes related to folate/vitamin B12 metabolism, transport of folate or vitamin B12, or transcriptional or developmental processes implicated in NTD mouse models (Table 1). Genotype data was successfully obtained for 1320 SNPs. Four tests of association were performed; two tests were to detect NTD case risk and two tests were to detect maternal risk for an NTD pregnancy. There were 203 SNPs in 54 genes that were significant (p < 0.05) by at least one test of association. A gene-based approach was used to select SNPs from fifteen genes to be genotyped in the secondary sample set. Five genes (mitochondrial folate transporter/carrier (MFTC), megalin (LRP2, low density lipoprotein receptor-related protein 2), DNA (cytosine-5-)-methyltransferase 3 beta (DNMT3B), phosphatidylethanolamine N-methyltransferase (PEMT), and euchromatic histone-lysine N-methyltransferase (EHMT1)) contained at least one SNP that was positive for both tests of a case effect or both tests of a maternal effect. An additional four genes (5-methyltetrahydrofolate-homocysteine methyltransferase 1 (MTR), cubilin (CUBN, intrinsic factor-cobalamin receptor), T (Brachyury homolog (mouse)), and AT rich interactive domain 1A (ARID1A)) contained more than five SNPs significant for any of the four tests of association. The remaining six genes (adenosine deaminase (ADA), ferritin, heavy polypeptide 1 (FTH1), cystathionase (CTH), peptidyl arginine deiminase, type IV (PADI4), low density lipoprotein receptor-related protein 6 (LRP6), and serine hydroxymethyltransferase 1 (SHMT1)) were selected based on a combination of factors, including the number of positive SNPs, their level of significance, and biological plausibility. Any SNP in these genes significant by any of the four tests of association (Table 2) was selected for genotyping in the secondary sample set (258 additional case families and 658 additional controls).

In the combined analyses each SNP was evaluated for contributing to NTD risk by twelve association tests: case and maternal effects were each evaluated using three case–control (or mother-control) tests and three family-based tests (see Methods). There were 68 SNPs in 30 genes that showed an association at the p < 0.01 level by any of the twelve tests (Table 3). Of these, twelve genes contained a single associated SNP. The remaining 56 SNPs were found in 18 genes. Not all associations were independent. Areas of interest were covered by tagSNPs as well as additional SNPs for physical coverage; as a result there were seven SNP pairs in this set with a linkage disequilibrium (LD) relationship above the threshold (r² ≥ 0.8) selected for tagging in this study (Figure 2, ARID1A_rs11247593 and ARID1A_rs11247594, CUBN_rs7070148 and CUBN_rs2273737, GART_rs2070388 and GART_rs4817580, MFTC_rs17803441 and MFTC_rs3134260, MTHFR_rs17367504 and MTHFR_rs17037425, MTR_rs10733117 and MTR_rs10925260, and PEMT_rs4646402 and PEMT rs1108579). Notably, for many genes, the associated SNPs occur in the same haplotype block (solid spine of LD based on D’ relationships), implying a single association signal for that gene (Figure 2, GART, COMT, MFTC, MTHFR, ARID1A, MTR, FOLH1, MTHFD1, PEMT, RAI1, ENOSF1). A single signal probably also accounts for the significant SNP pairs seen in PDGFRA (D’ = 0.65) and CDKN2A (D’ = 0.79). Genes exhibiting more than one independent SNP association include FTCD (D’ = 0.10), MTHFD1L (two separate haplotype blocks and two SNPs, D’ ≤ 0.57), CUBN (one haplotype block and a SNP, D’ ≤ 0.20), ALDH1A2 (D’ = 0.13) and ADA (a weak haplotype block and a SNP, D’ ≤ 0.19).

We ranked SNPs by the lowest p-value for any test, accounting for relatedness to other highly significant SNPs. The nine genes exhibiting the 10 strongest association signals were: MFTC, CDKN2A, ADA, PEMT, CUBN, GART, ADA, DNMT3A, MTHFD1 and T (Brachyury) (Table 4). MFTC, ADA, PEMT and CUBN contained more than one significant SNP, and ADA showed evidence of two independent association signals.

SNPs in MFTC, PEMT and ADA account for seven of the top ten SNPs (Table 4). In MFTC, these two SNPs (rs17803441 and rs3134260) are essentially in perfect LD (D’ = 1.0, r² = 0.99). As expected, they yielded very similar evidence of NTD risk in a continuous model of logistic regression (rs17803441: OR = 1.61 [1.23-2.08], p = 0.0003, Risk Allele Frequency (RAF) = 0.07; rs3134260: OR = 1.56 [1.22-2.04], p = 0.0006, RAF = 0.07), as well as a recessive model of logistic regression (rs17803441: OR = 1.59 [1.19-2.08], p = 0.0013; rs3134260: OR = 1.54 [1.18-2.04], p = 0.0021). There are a total of five highly significant (p < 0.01) MFTC SNPs and they are all consistent with a case effect. MFTC rs10112450 is significant by the transmission disequilibrium test (TDT, GRR = 1.42, p = 0.0065, RAF = 0.79), and the remaining MFTC SNPs (rs1865855 and rs750606) show evidence of NTD risk to the case by logistic regression with a continuous or dominant model (p < 0.0091). Because all five of these SNPs fall in the same haplotype block (D’ > 0.96), it is likely there is a single variant in this region responsible for the association signals.

Three SNPs in PEMT were in the top ten SNPs, and two of these are in high linkage disequilibrium (rs1108579 and rs4646402; D’ = 1.0, r² = 0.809). These SNPs also yield very similar evidence of NTD risk by TDT (rs1108579: GRR = 1.47, p = 0.0006, RAF = 0.52; rs4646402: GRR = 1.43, p = 0.0009, RAF = 0.57). PEMT rs11656215 is less strongly linked to these SNPs (r² < 0.6), and shows association with NTD risk by TDT (GRR = 1.35, p = 0.0053, RAF = 0.49) and by log-linear analysis of a recessive model (GRR = 1.68 [1.24-2.28], p = 0.0008). A fourth highly significant SNP was found in the maternal analysis (rs16961845). It shows evidence of maternal risk in a recessive model of log-linear analysis (GRR = 1.92 [1.18-3.11], p = 0.0082), as well as recessive (p < 0.0048) and continuous (p < 0.0072) models of logistic regression. By r² measures of LD, this SNP is the least related to the other three SNPs (r² < 0.08), although all four of these SNPs fall in the same haplotype block (Figure 2).

Lastly, three SNPs in ADA were among the ten SNPs with the lowest p values. They show evidence of LD by D’ (≥0.82) but less so by r² (≤0.54). These three SNPs are all significantly associated with a maternal effect by logistic regression of a continuous model (rs2299686: OR = 1.30 [1.11-1.52], p = 0.0010, RAF = 0.45; rs427483: OR = 1.28 [1.10-1.52], p = 0.0018, RAF = 0.33; rs406383: OR = 1.33 [1.12-1.58], p = 0.0012, RAF = 0.25). Three other SNPs in ADA (rs6031682, rs452159 and rs6094017) also showed a highly significant association with a maternal effect in a continuous model of logistic regression (p < 0.0059). All six SNPs were highly significant (p < 0.01) for association with maternal risk in either a recessive model (rs406383, p = 0.0021) or dominant model (the other five SNPs, p ≤ 0.0063) of logistic regression. Additionally, ADA rs6031682 shows association with case risk in a continuous model of logistic regression (1.38 [1.10-1.73], p = 0.0052, RAF = 0.84). These six SNPs do not appear to be strongly linked; no haplotype blocks (solid spine of LD) larger than two SNPs were identified. It appears that rs6031682 is clearly outside of a degraded haplotype block consisting of the other five SNPs.

CUBN was the only other gene with more than one SNP found in the ten most strongly associated signals. CUBN rs7070148 and CUBN rs2273737 were both significant for association with maternal risk in a continuous model of logistic regression (OR = 1.64 [1.22-2.17], p = 0.0010, RAF = 0.90 and OR = 1.54 [1.18-2.04], p = 0.0021, RAF = 0.89, respectively). These SNPs are highly linked (D’ = 0.935, r² = 0.823). There are three other highly significant (p < 0.01) SNPs in CUBN. Two other SNPs in the same haplotype block with CUBN rs7070148 and CUBN rs2273737 (rs1801222 and rs11254375, D’ < 0.76) also showed an association with maternal effects. A third SNP, CUBN rs11591606, is outside this block (D’ < 0.20) and is associated with maternal risk in a dominant model of logistic regression (OR = 4.15 [1.51-11.39], p = 0.0058, RAF = 0.17).

Correction for multiple testing was performed for three of the twelve tests of association. No adjusted p-value was found to remain significant, and no further correction was performed.

The recruitment of the Irish NTD families (cases and parents) and controls has been described [23, 34, 51, 52]. Briefly, the cohort includes 586 families with an NTD case; 442 of these families are full family triads (DNA from case, mother and father). For this study, 570 of the NTD families had sufficient DNA and were divided into two sets, one for primary analysis and one for secondary (combined) analysis. The primary and secondary sample sets were matched as closely as possible for the following parameters: the number of complete NTD triads, the proportion of NTD cases with spina bifida vs. other NTDs, and NTD case gender (Table 6).

The control population (n = 999) is a random sample drawn from 56,049 blood samples donated by women at their first prenatal visit to the three major maternity hospitals in Dublin (1986–1990). A subset of controls (n = 341) was randomly selected for the primary screen, and the remaining controls were used in the secondary set.

Written, informed consent was obtained from study subjects, their parents or their guardians. Archived control samples were anonymized prior to analysis. The study was approved by the Ethics Research Committee of the Health Research Board (Dublin, Ireland) and the Institutional Review Board at the National Human Genome Research Institute (Bethesda, MD, USA). Extraction of genomic DNA from blood samples and buccal swabs was performed using the QIAamp DNA Blood Mini Kit (Qiagen, Valencia, CA, USA).

Genes were chosen for study because they are involved in folate and vitamin B12 metabolism or other metabolic and signaling pathways implicated in the etiology of NTDs (Figure 1). Although previously published, MTHFR[52‐54], MTHFD1[23, 25], MTHFD1L[27], TCblR[34], and TP53[51] were reanalyzed in this study in order to compare the differing research strategy, which includes new genetic models of risk assessment. For each gene chosen, we evaluated the transcribed region of the gene and 10 kilobases (kb) upstream and downstream of the gene in an effort to include polymorphisms with potential proximal effects, such as promoter variants. In order to capture the common variation in each gene, SNPs genotyped by HapMap (Data Release 21, Phase II, Anon, 2003 [55]) were considered. A set of tagSNPs was identified using an algorithm based on optimizing for tagSNPs with maximal minor allele frequencies (MAFs) and an r² threshold of 0.8, while maximizing the MAF of the selected tagSNP. In addition to this set of tagSNPs, validated variants from dbSNP [26] were also selected for physical coverage so that spacing between SNPs would be less than 20 kb within D’ haplotype blocks and less than 5 kb between haplotype blocks. A total of 1517 SNPs were selected.

The selected 1517 SNPs were genotyped on the primary sample sets (320 NTD case families and 341 controls). Genotypes were generated by the Johns Hopkins University SNP Center (Baltimore, MD) using the Illumina GoldenGate assay (Illumina, San Diego, CA). Of the 1517 SNPs attempted, 1320 SNPs (87.2%) remained after filtering out low quality data (re-attempted on another platform, see below). The overall call rate for these 1320 SNPs was 98%. All but 150 of these SNPs had a call rate of >95%; the rest had an average call rate of 87.2% (±7.9%) and were re-genotyped on another platform (see below). Both the overall duplicate concordance rates and the Mendelian consistency rates were 99.99% for the 1320 accepted SNPs.

Based on analyses of the primary sample sets, 93 SNPs were genotyped in the secondary sample set (250 NTD case families and 658 controls). Genotypes were generated by KASPar chemistry at Kbiosciences (Herts, UK). Three SNPs failed on this platform: rs127317149, rs7367859 and rs7096079. For the 90 successfully genotyped SNPs, the overall duplicate concordance rate was 99.81% and the overall Mendelian consistency rate was 99.94%.

SNPs that could not be assayed (n = 197) or returned low call rates (<95%, n = 150) via Illumina GoldenGate chemistry were re-genotyped in the entire sample set (570 NTD families and 999 controls) by detection of allele-specific primer extension using matrix-assisted laser desorption/ionization – time of flight (MALDI-TOF) mass spectrometry (Sequenom, San Diego, CA, USA). Genotyping data from SNPs with call rates above 95% were added to the final analyses (121/197 SNPs that failed and 106/150 SNPs that yielded low call rates). The overall duplicate concordance rate was 99.10% and the overall Mendelian consistency rate was 99.32% for these 227 SNPs.

To summarize, our final data set consisted of 1441 SNPs; 1320 high quality SNPs from the Illumina platform (93 of these SNPS were also typed in the secondary samples using the KASPar platform, and 106 of these SNPs were re-typed in the entire sample set using the Sequenom platform) plus 121 SNPs from the Sequenom platform.

D’ and r² measures of LD for SNPs of interest were estimated based on control genotypes using Haploview [56]. Haplotype blocks were based on D’ values using the Solid Spine of LD option in Haploview.

The design for this study involved two stages of genotyping. Rather than use the secondary samples for a replication study, joint analysis of the combined dataset was performed since it generally provides greater power to detect a genetic effect [57]. In the initial analysis, all SNPs successfully typed with the Illumina GoldenGate assay (n = 1320, including the 150 high quality SNPs with call rates of <95%) were analyzed in the primary sample set by four tests of association. Two tests were performed to evaluate case effects: 1) Logistic regression, a 1-degree of freedom (DOF) test of association between the affected status and number of risk alleles; and 2) the Spielman transmission disequilibrium test (TDT, [58]). Similarly, two tests were performed to evaluate maternal effects: 1) Logistic regression, a 1-DOF test of association between the maternal status and number of risk alleles; and 2) log-linear modeling with 2-DOF to test for effect of the maternal genotype.

SNPs were selected to be genotyped on all samples (570 NTD case families and 999 controls) if they met the following criteria: 1) SNPs of interest reaching a significance level in the primary analysis (n = 93, Table 2); or 2) failed SNPs (n = 121) and SNPs with low call rates (<95%, n = 106). Final analyses were performed on the entire dataset and consisted of twelve association tests. NTD case–control and NTD mother-control comparisons were performed using continuous, recessive and dominant coded models of logistic regression to generate odds ratios (ORs) and 95% confidence intervals (CI). Six family-based tests of association were also applied using log-linear models. The NTD triads (case, mother and father) were analyzed for case effects using recessive, dominant and linear (TDT) coding while direct maternal effects were analyzed using recessive, dominant and 2-DOF models.

To correct for multiple comparison, the final analysis used the complete information on the 1441 SNPs. Correction was performed by multivariate permutation (N = 9,999 random permutations) for three of the tests used in the initial analysis: case–control logistic regression, mother-control logistic regression and the TDT. This method accounts for any linkage disequilibrium between SNPs. Multivariate permuting of triads for the TDT was performed by treating the test as a one-sample test and permuting the risk allele [59]. Independent permutations were performed for the cases and controls or mothers and controls to adjust for multiple comparisons in the tests of logistic regression. The results were combined by Bonferroni adjustment to account for all comparisons (including all SNPs because of our multivariate permutation approach) while controlling the probability of any false positives at 5%. Since a SNP could only be found significant based on this combined analysis, our method provides type I error control regardless of the SNP selection process or number of SNPs selected for genotyping on all samples.