This case-control study included data on deliveries that had estimated due dates from 1995 to 1998 or 1999 to 2003. The study included live-born infants with isolated CLP (cases; N = 330 [149 from 1995 to 1998 and 181 from 1999 to 2003]) or without any structural malformation (controls; N = 367 [162 from 1995 to 1998 and 205 from 1999 to 2003]). Case information was abstracted from hospital reports and medical records following established procedures by the California Birth Defects Monitoring Program [Croen et al., 1991]. Each medical record was further reviewed by a medical geneticist (E.J.L.). Infants with trisomies were ineligible. Non-malformed controls were selected randomly to represent the population from which the cases were derived in selected counties and birth periods. This study, including the collection and use of archived newborn bloodspots, was approved by the California State Committee for the Protection of Human Subjects as well as Institutional Review Boards at Stanford University and the University of California, Berkeley. See Table I for breakdown of population demographics.

Mothers who delivered in the period 1999–2003 were eligible for interview if they were the biologic mother and carried the pregnancy of the study subject, they were not incarcerated, and their primary language was English or Spanish. Maternal interviews were conducted using a standardized, computer-based questionnaire, primarily by telephone, in English or Spanish, no earlier than 6 weeks after the infant’s estimated date of delivery. Information on a variety of exposures and conditions was solicited from women including age, race/ethnicity, educational level, and family history (first degree relative) of clefts. Queries specific to the periconceptional period, which was defined as 2 months before through two months after conception, included use of folic-acid containing vitamin supplements, diabetes (gestational, Type I, and Type II), and seizure medication use. The interview also included a modified version of the National Cancer Institute’s Health Habits and History Questionnaire, a well-known, semi-quantitative food frequency questionnaire with demonstrated reliability and validity [Block et al., 1990]. The food frequency questionnaire was modified to include ethnic foods appropriate to a diverse study population. This questionnaire provided information on dietary folate and energy intake. The median time between estimated date of delivery and interview completion was 11 months for cases and 9 months for controls.

Genomic DNA (gDNA) was available from newborn screening dried blood spots obtained from linkage efforts made by the California Birth Defects Monitoring Program. gDNA was extracted from dried blood spots and whole-genome amplified using QIAmp DNA Mini kits and Repli-G Midi amplification kits (Qiagen, Germantown, MD), respectively, as described previously [Marini et al., 2011]. The average yield of purified whole-genome amplified (WGA) DNA was 15 mg per individual.

SNPs for genotyping were identified from the following sources: (i) the sequencing of 31 folate pathway genes to identify risk factors for neural tube defects [Marini et al., 2011]. From that study, we chose all common alleles (MAF > 5%), all nonsynonymous changes (including singletons), and all variants that displayed a trend toward case-control association (P value <0.15). (ii) Additional nonsynonymous SNPs in these 31 genes present in public SNP databases (NCBI). (iii) SNPs previously published as having potential functional significance in folate-related genes. (iv) Haplotype tagging SNPs for a subset of folate pathway genes to enhance coverage beyond what was seen from exon sequencing [Marini et al., 2011].

Genotyping was performed primarily using TaqMan SNP genotyping assays (ThermoFisher, Foster City, CA) on the Fluidigm EP1 platform (Fluidigm, South San Francisco, CA) using a 96×96 dynamic array. Briefly, specific target amplification in each sample was performed in pooled groups of 96 primer pairs and then subject to individual TaqMan assays taking care to not include SNP sites within 200 bp of each other. All allele calls were checked manually following automated allele calling by the Fluidigm SNP Genotyping Analysis software. For multiple SNPs in close proximity and repeat polymorphisms that could not be genotyped by TaqMan (N = 25), genotyping was performed by Sanger sequencing of amplified PCR products. As quality assurance, genotyping assays were duplicated for a 10% subset of DNA samples. All genotyping was performed blinded to case and control status. After removing QC failures and mono-allelic SNPs, the final analytical list comprised 504 variants in 35 genes which are listed, along with attributes, in Table SI.

Variants were treated as categorical variables, that is, homozygous wild-type as referent versus heterozygous or homozygous variant. Odds ratios and 95% confidence intervals were used to estimate relative risks (using SAS software version 9.4). For low frequency variants not amenable to OR calculation, we developed a simple metric (“Case-Control Difference”) which highlights potential risk alleles with the most differential occurrence between cases and controls. The Case-Control Difference (CCD) is calculated by the following formula: |x-y|/sqrt(x+y), where x and y equal the normalized allele counts in cases and controls, respectively. Thus, the CCD is a measure of the differential appearance of alleles in cases and controls (numerator) with a penalty for the occurrence of that allele in both populations (denominator). For example, a low frequency variant with five occurrences in cases and zero in controls would be an interesting risk candidate and receive a higher CCD score than an allele that occurs 10 times in cases and five times in controls. Although this tool was developed for low frequency variants, CCD scores can be quite high for higher-frequency alleles if there is a considerable case-control difference, though these alleles may be better highlighted by OR calculations (Table SII). The numerator is an absolute value so the metric reveals a measure where alleles show either increased or decreased risks. It should be emphasized that the CCD merely highlights differentially behaving alleles and does not ascribe statistical significance.

Gene-folate interactions were assessed for homozygous or heterozygous variants. A dichotomized composite variable, “combined folate intake,” was created by combining maternal periconceptional intake of folic acid-containing vitamin supplements with daily dietary intake of folate. Two categories were constructed corresponding to ≤ 25th percentile (to reflect low dietary folate intake) and >25th percentile, as determined from dietary folate intake levels among control mothers (312.67 μg). Combined folate intake was defined as low for women in the lowest quartile (≤ 25th percentile) who did not take supplemental folic acid in the periconceptional period. Not-low folate intake was defined as dietary folate intake above the lowest quartile (>25th percentile) regardless of maternal vitamin use. Interactions between each variant and the combined folate intake variable were investigated for effects on risk of CLP. An interaction with P-value <0.05 for the Wald test was considered statistically significant. Given that our study time period overlapped the introduction of folic acid fortification of grain products in the US (1998), we analyzed whether genetic results differed between the periods before and after fortification. For this analysis, we specified an index of 0 for before fortification (with estimated due dates from 1995 to 1998) and an index of 1 for after fortification (with estimated due dates from 1999 to 2003). The interaction between each SNP and this index was tested with the Wald test with P-value <0.05 as the statistically significant determinant.

Table II shows all significant associations for individual SNPs when heterozygous and homozygous (calculated separately as in Methods) within the entire population, as well as within non-Hispanic white and Hispanic white sub-groups. There were no variants statistically significantly associated (P-value <0.05) with CLP as both heterozygotes and homozygotes in the entire population (Table II). ALDH1L1_74483, a nonsynonymous I812V change, showed significant increased risk when heterozygous, but was too infrequently homozygous to estimate the OR. Of those alleles with frequencies high enough to estimate homozygous ORs, only GART_26477 showed the trend of increasing risk with variant allele number, but was significant only when homozygous. In addition, several variants displayed decreased risk trending with variant allele number: two tightly linked loci within the BHMT2 locus (BHMT2_8849, BHMT2_10283) significantly reduced risk when homozygous whereas MTHFD1_59019 was significant only as a heterozygote.

In non-Hispanic white infants, an intronic SNP in AHCYL2 (AHCYL2_165230) displayed significant risk when heterozygous (OR = 2.6; 95%CI 1.1–5.9) and homozygous (OR = 3.0; 95%CI 1.1–7.8). Interestingly, a second set of three tightly linked variants in the AHCYL2 gene (AHCYL2_165528, 173166, 203190) also displayed association with a significant reduction in risk seen only for homozygotes. In the CBS gene, the 68-bp insertion in intron 7 (c.844ins68, here called CBS_13779) and a second linked locus (CBS_13704) displayed significant risk when heterozygous, whereas a low-frequency intronic SNP in DHFR (DHFR_17874) displayed significantly reduced risk in heterozygotes. It should be emphasized, however, that the non-Hispanic white sub-population in the study was fairly small (Table I).

In Hispanic infants, the strongest associations emanated from the ALDH1L1 gene and the BHMT/BHMT2/DMGDH locus. BHMT, BHMT2, and DMGDH lie adjacent within a 140 kb cluster on chromosome 5. Thus, several sets of variants exist that span all three genes and are in strong linkage disequilibrium effectively tagging the locus. BHMT2_13828, an intronic variant in one such set, is the only polymorphism that showed a significant OR when both heterozygous (OR = 0.6; 95%CI 0.4–0.9) and homozygous (OR = 0.5; 95%CI 0.3–0.8) with effect trending with allele number. Four other tightly linked variants from this locus (BHMT2_8849, BHMT2_10283; DMGDH_26164, DMGDH_26338) also showed significantly reduced risk when homozygous. The nonsynonymous I812 V variant of ALDH1L1 (ALDH1L1_74483) showed a very significant heterozygote risk (OR = 3.8; 95%CI 1.6–9.2) suggesting that the Hispanic sub-group drove the association seen in the overall population. Interestingly, two additional ALDH1L1 alleles, ALDH1L1_74427 (nonsynonymous, D793G) and ALDH1L1_77540 (intronic), which are tightly linked to each other but not to ALDH1L1_74483, were associated in Hispanics with significantly reduced risk when heterozygous.

Low-frequency variation, particularly in coding regions, has been implicated in risk for orofacial clefting [e.g., Leslie and Murray, 2013, Leslie et al., 2015]. Our study data contained a substantial number of variants with MAF <5% (N = 267; see Table SII) which often render OR calculations unstable. To highlight low frequency variants that may contribute to disease risk, we used a simple metric, the Case-Control Difference, which quantifies and rank orders variants with the most differential behavior in cases versus controls (see Methods). We applied this analysis to all variants with MAF <5% and the top 5% of rank-ordered lists for the entire population, as well as non-Hispanic and Hispanic sub-groups, are shown in Table III. Note that some variant alleles were frequent enough to allow OR calculations and this metric can further draw attention to variants previously observed with significant odds ratios (e.g., ALDH1L1_74483). The case-control difference also gives emphasis to low-frequency variants that narrowly miss statistical significance cut-offs. Of note in this group are DHFR_17874, significantly associated in non-Hispanic whites (Table II), but with a relatively strong case-control difference in the entire population, a new putative risk allele in BHMT (BHMT_17018; a synonymous Y284Y variant), and a variant within the 3^′-UTR of SHMT1 (SHMT1_35845). In addition, the case-control difference revealed three unlinked MTHFD1 alleles (MTHFD1_12347, 32373, 54248) among variants with the most differential occurrences.

In non-Hispanic whites, the case-control difference highlighted three tightly linked rare, nonsynonymous variants within the MTRR gene with notable differential behavior [MTRR_10208 (S257T), 21088 (R415C), 23290 (P450R)]. The metric also flagged an intronic variant in MAT1A (MAT1A_5181), which displays strong differential behavior in the entire population as well, and the significantly associated DHFR_17874 allele (see Table II).

In Hispanics, the highest scoring differential allele was ALDH1L1_74483, previously identified as a significant risk allele by OR (Table II). A second, rare variant within this gene (ALDH1L1_44924) was also among the most differentially occurring alleles. In addition, this analysis flagged four unlinked variants from homocysteine trans-sulfuration genes CBS and CTH as well as the MTHFD1 allele (MTHFD1_43530), which was significantly associated in the entire population.

We explored the potential interaction between maternal folate intake and each individual SNP. Thus, the population was stratified into “low” and “not-low” folate intake groups based on the combined data of supplementation and dietary intake (see Methods). Our a priori expectation was that increased risks for a particular variant would be substantially greater in the low folate group than in the not-low folate group. As shown in Table IV, there were a few associations that fit this pattern of risk with a P-value <0.05 for the Wald test of gene-folate interaction. In the entire population, low-folate specific risk was seen for an intronic variant of CBS (CBS_9548; OR 3.6, 95%CI 1.3–9.6) and two linked alleles of the cobalamin (vitamin B12) transporter TCN2: TCN2_9397 (intron) and the well-studied nonsynonymous SNP R259P (TCN2_9459; rs1801198), both with OR 2.8, 95%CI 1.2–6.3). Neither the CBS allele nor the TCN2 alleles were seen in the gene-only analyses of Table II. In fact, no alleles from either gene were significantly associated in the entire population. Such folate-responsive alleles might not be expected to be observed as significant in the entire population without folate intake stratification because the low-folate intake sub-group is relatively small (28 cases, 23 controls). The only significant odds ratio for increased risk with low-folate intake within race-ethnic sub-groups was CTH_4763 in Hispanics (OR 11.5, 95%CI 1.8–72.9), but the sample size was too small for meaningful conclusions.

We also investigated whether the gene variant analyses were modified by the introduction of folic acid fortification of grain products in the US, that is, we analyzed whether genetic results differed between the periods before and after fortification. None of the observed P-values indicated that this term when added to the variant models produced statistically precise results (data not shown).

One of the most consistent observations in the literature regarding folate pathway-related genes and CLP risk is the frequently observed significant association of variants within the BHMT/BHMT2/DMGDH gene cluster on chromosome 5. These three genes lie in tandem within 140 kb at chromosomal region 5q14. Because of their close proximity, there are strong blocks of linkage disequilibrium that span these genes. Thus, a positive signal from one variant could, in theory, be tagging an allele anywhere in the region. Table V summarizes previous studies (including this one) in which variants in this region have been tested for CLP association. This table lists all previously published studies querying SNPs in this region, with a brief description of study population and those SNPs displaying significant associations. Quite convincingly, out of nine studies that use a variety of analytical techniques, seven (~ 80%) report at least one significant association as determined by the investigators. In this regard, this locus stands apart from most other folate genes interrogated in CLP association studies. Consistent with our results, Blanton et al. [2011] observed association primarily in Hispanics. It is noteworthy that several studies in Table V have observed significant associations with the maternal genotype at this locus. Given the correlation between maternal and fetal genotype, it is possible that fetal associations are a consequence of the fetal genotype simply serving as a proxy for the maternal genotype. It is currently unknown whether fetal or maternal genetic variation (at this or other loci) is more relevant for CLP risk.

While significant associations of common variants within this gene cluster are often observed, the risk attributed to the queried variants may be increased (odds ratios >1) or decreased (odds ratios <1) as reported in this study (see Table II and references in Table V). While the reason for this is not clear, this may reflect the contribution of underlying rare variants at this locus (which may occur on the haplotype of the major or minor allele of a nearby common variant) as has been previously suggested for clefts [Leslie and Murray, 2013] and as has been seen for regions highlighted by GWAS studies [Leslie et al., 2015].

There are further lines of evidence that lend support to this region as a genuine risk locus for CLP. BHMT/BHMT2 (Betaine-homocysteine methyltransferase) and DMGDH (Dimethyl-glycine dehydrogenase) function sequentially in a multi-step process converting choline to glycine. Measuring the relevant metabolites choline and betaine in maternal mid-pregnancy sera, Shaw et al. [2009] found altered metabolite levels in CLP case mothers, suggesting metabolic imbalances. Furthermore, increased intake of choline in the maternal diet has been associated with decreased CLP risk [Shaw et al., 2006]. In mouse models, BHMT knock-outs are not reported to increase the risk of clefts in offspring, though choline and one-carbon metabolism are perturbed [Teng et al., 2011]. Thus genetic studies, metabolite analysis and epidemiology converge at this locus for CLP risk.

We identified two low-frequency intronic variants within the DHFR gene with relatively strong differences in cases and controls (DHFR_17874, DHFR_26717). DHFR_17874 had a significant odds ratio in non-Hispanic whites (OR 0.2; 95%CI 0.1–0.8). Martinelli et al. [2014] also observed significant associations with reduced risk for CLP within DHFR in an Italian population, though one of these SNPs (DHFR_15483 here) did not show a significant association in the current study. Mechanistically, DHFR has been implicated in orofacial clefting from model system studies. Wahl et al. [2015] reported that methotrexate, a Dihydrofolate reductase inhibitor, causes craniofacial malformations and exacerbates clefting phenotypes in a Xenopus model. In addition, teratology studies in mice and rabbits show that maternal methotrexate exposure during early pregnancy results in orofacial clefting outcomes in offspring [Hyoun et al., 2012].

A subset of folate-dependent reactions plays a critical role in the synthesis of S-adenosylmethionine (AdoMet), which serves as the methyl group donor in DNA and protein methylation reactions. Homocysteine, which is the precursor of methionine (and, therefore, AdoMet) is also subsequently regenerated following AdoMet methyl transfer in a cyclic series of reactions referred to as the “methionine cycle.” The genes/enzymes of the methionine cycle have garnered interest in identifying risk for orofacial clefts because of some studies that report high plasma homocysteine levels in cleft case mothers [Wong et al., 1999; Kumari et al., 2013], though this does not appear to be the case in all populations queried [Shaw et al., 2009; Munger et al., 2011]. Nevertheless, it is noteworthy that the BHMT/BHMT2 reactions (whose genes are often implicated in cleft risk as discussed above) convert homocysteine to methionine, and thus can impact the methionine cycle. Furthermore, we have seen a strong representation of genes that impinge on the methionine cycle in non-Hispanic white cases: ACHYL2_165230 was the most significant risk allele in this group. A significant risk signal also emanated from the CBS (cystathionine-beta-synthase) 68bp insertion (c.844ins68; here CBS_13779) and a tightly linked allele. The case-control difference metric highlighted alleles in MTRR (Methionine Synthase Reductase; necessary for the folate-dependent conversion of homocysteine to methionine) and MAT1A (necessary for AdoMet synthesis). The 844ins68 allele of CBS has been tested in other populations with typically conflicting results. For example, Rubini et al.[2005] reported evidence of transmission distortion and imprinting for this allele in an Italian sample of 134 CLP triads, whereas Martinelli et al. [2011] found no association in a different Italian population. However, evidence of association for other alleles of CBS has been reported in multiple studies [Boyles et al., 2008, 2009; Blanton et al., 2011; Martinelli et al., 2011]. Finally, it is worth noting that the alleles/genes highlighted as conferring risk specifically with low maternal folate intake (Table IV) participate in the methionine cycle: CBS (CBS_9548) and the TCN2 cobalamin (vitamin B12) transporter (TCN2_9397, TCN2_9450). Vitamin B12 is essential for the Methionine synthase (MTR) reaction converting homocysteine to methionine and, thus, TCN2 activity impinges on homocysteine levels [Hsu et al., 2011].

Thus, these results re-focus emphasis on homocysteine clearance and the methionine cycle, yet the underlying mechanism is still unclear. Elevated homocysteine levels may be a specific marker for DNA methylation cycle aberration or could merely be indicative of altered folate metabolism in general. Wahl et al. [2015] have suggested that deficiencies in the folate-dependent synthesis of nucleotide precursors, rather than cellular methylation changes, underlie clefts in a Xenopus model. These defects lead to impairment of DNA synthesis, genomic instability and, ultimately, apoptosis.

We have identified alleles of MTHFD1 that are significantly associated with CLP in the entire population (MTHFD1_43530, 59019; Table II) as well as three additional, unlinked alleles highlighted by the case-control difference as some of the most differentially occurring in cases and controls (Table III). Similarly, alleles of MTHFD1 were also flagged by odds ratio and the case-control difference in Hispanics. To our knowledge, these variants have not been tested in previous studies, though other MTHFD1 variants have been tested for CLP risk. [Mostowska et al., 2006; Mills et al., 2008; Palmieri et al., 2008; Boyles et al., 2009; Blanton et al., 2011]. Mills et al. [2008] reported a significant association with the common, nonsynonymous R653Q variant and Blanton et al. [2011] reported significant associations in non-Hispanic white individuals for multiple MTHFD1 variants. It is noteworthy that, whereas homozygous deletions of MTHFD1 are embryonic lethal in mice, heterozygous mutations have negative impacts on purine synthesis (a folate-dependent process) and increased incidence of developmental defects in embryos including craniofacial malformations [Christensen et al., 2013]. MTHFD1 may therefore be considered a candidate risk gene.

The most significant risk allele in our study was ALDH1L1_74483, a nonsynonymous I812V alteration strongly associated in Hispanics. To the best of our knowledge, this polymorphism has not been previously tested for CLP association. Boyles et al. [2009] tested several ALDH1L1 variants, but none were significantly associated with CLP in their study. There is little phenotypic information surrounding ALDH1L1 mutation to implicate it in craniofacial development, thus the importance of ALDH1L1 variation in cleft risk remains to be determined.

This study has implicated several loci related to folate metabolism as important in conferring risk for isolated clefts. Although these findings are consistent with previous studies and with biological mechanisms, they were based on relatively small sample sizes. In particular, the power to detect gene-nutrient interaction effects was low. Therefore, replication of these findings in additional populations is warranted.