Evidence for SNP-SNP interaction identified through targeted sequencing of cleft case-parent trios

Abstract

Non-syndromic cleft lip with or without cleft palate (NSCL/P) is the most common craniofacial birth defect in humans, affecting 1 in 700 live births. This malformation has a complex etiology where multiple genes and several environmental factors influence risk. At least a dozen different genes have been confirmed to be associated with risk of NSCL/P in previous studies. However, all the known genetic risk factors cannot fully explain the observed heritability of NSCL/P, and several authors have suggested gene-gene (G×G) interaction may be important in the etiology of this complex and heterogeneous malformation. We tested for G×G interactions using common SNPs derived from targeted sequencing in 13 regions identified by previous studies spanning 6.3 MB of the genome in a study of 1,498 NSCL/P case-parent trios. We used the R-package trio to assess interactions between polymorphic markers in different genes, using a 1 degree of freedom (1df) test for screening, and a 4 degree of freedom (4df) test to assess statistical significance of epistatic interactions. To adjust for multiple comparisons, we performed permutation tests. The most significant interaction was observed between rs6029315 in MAFB and rs6681355 in IRF6 (4df p=3.8×10⁻⁸) in case-parent trios of European ancestry, which remained significant after correcting for multiple comparisons. However, no significant interaction was detected in trios of Asian ancestry.

Introduction

Oral clefts include three distinct anatomical malformations: cleft lip (CL), cleft palate (CP) and cleft lip and palate (CLP), and combined these represent the most common craniofacial birth defects in humans, affecting 1.7 per 1000 live births [Rahimov et al. 2012]. Since CL and CLP share similar epidemiologic distributions and develop during the same embryologic periods, CL and CLP are often grouped together as cleft lip with or without cleft palate (CL/P). The majority of oral cleft cases are “non-syndromic” because they occur as an isolated anomaly with no other structural abnormality or developmental disability in the child. Among all cases, approximately 70% of CL/P cases and 50% of CP cases are non-syndromic [Jugessur et al. 2009], while the remaining cases have another congenital anomaly or developmental delay representing some malformation syndrome. While the overall prevalence of oral clefts is high, this varies by type of cleft: the prevalence of CL/P ranges between 3.4 and 22.9 per 10,000 live births, while the prevalence of CP is 1.3–25.3 per 10,000 live births [Mossey and Castilla, 2003]. There are substantial differences in prevalence of CL/P across racial groups and populations: Asians and Native Americans have the highest rate of 2 per 1,000 live births, Caucasians have a prevalence around 1 per 1,000 and African populations have the lowest prevalence rate of 1 per 2,500 live births [Dixon et al. 2011; Mossey et al. 2009]. Gender is also related to risk to oral clefts, CL/P is more common in males with a 2:1 ratio of males: females, while CP is more frequent in females [Mossey et al. 2009; Matthews et al. 2015].

Oral clefts represent a complex and heterogeneous group of malformations where both genetic and environmental risk factors control risk [Leslie and Marazita, 2013]. Since 2009, genome-wide association studies (GWAS) using both case-control and case-parent trio designs have identified more than a dozen genes achieving genome-wide significance as influencing risk to oral clefts, and other possible genomic regions have been identified through genome-wide linkage studies using multiplex cleft families (i.e. those with more than one affected individual). As the list of putative causal genes expands, it is logical to ask if some of these genes may interact with one another in biological pathways that can be identified through tests for statistical interaction. Detecting such gene-gene (G×G) interactions on a large scale is a daunting challenge since the number of potential hypotheses and the incurred multiple comparisons burden limit the power to detect interaction even in studies with large sample sizes [Cordell, 2002]. Li et al. [2015] tested for G×G interaction involving the WNT signaling pathway using CL/P case-parent trios from the GWAS by Beaty et al. [2010] and identified an interaction between markers in WNT5B and MAFB among both Asian and European case-parent trios, as well as interactions between markers in WNT5A and IRF6 in Asian trios, and markers in the 8q24 region and WNT5B in European trios.

Here we present a step-wise strategy for testing for G×G interaction in 1,409 case-parent trios of European or Asian descent, using common variants identified from targeted sequencing of 13 recognized candidate genes/regions (8q24, ARHGAP29, BMP4, FGFR2, FOXE1, IRF6, MAFB, MSX1, NOG, NTN1, PAX7, PTCH1, VAX1, see Table I). We focused on common tagging SNPs available through the study described by Leslie et al. [2015]. Each of these 13 regions was previously shown to be associated with NSCL/P in either previous GWAS or genome-wide linkage studies. We limited our tests for interaction to common variants since low-frequency and rare variants do not provide sufficient power to detect epistasis with the number of trios available. To assure scalability and accuracy, we selected tagging SNPs, and carried out a two-step procedure to detect SNP-SNP interactions. We first used a very fast 1 degree of freedom (1df) Wald test in a simplified statistical model to generate a list of candidate SNP pairs. Since departures from the null in these 1df tests can also indicate violations of model specifications, we analyzed the candidate pairs using a 4df test based on a more general statistical model proposed by Cordell [2002] to comprehensibly assess epistatic SNP-SNP interactions. The overall significance in the context of multiple comparisons was then assessed using a permutation test.

Table I.

Candidate genes or regions sequenced in this study.

Gene	Targeted Region (GRCh37)	Total (kbp)	Number of common SNPs
IRF6	chr1:209837199-210468406	631.2	50
MAFB	chr20:38902646-39614513	711.9	111
ARHGAP29	chr1:94324660-95013109	688.4	102
8q24	chr8:129295896-130354946	1059.1	97
PAX7	chr1:18772300-19208054	435.8	172
VAX1	chr10:118421625-119167424	745.8	82
NTN1	chr17:8755114-9266060	510.9	175
NOG	chr17:54402837-54957390	554.6	73
FOXE1	chr9:100357692-100876841	519.1	52
MSX1	chr4:4825126-4901385	76.3	57
BMP4	ch14:54382690-54445053	62.4	12
FGFR2	chr10:123096374-123498771	402.4	42
PTCH1	chr9:98133647-98413162	279.5	48

Methods

Study Population

A total of 1,498 cleft case-parent trios were recruited from different sites in China, the Philippines, the United States and Europe (see Leslie et al. 2015 for a full description of the sample). These were used for targeted sequencing of 13 genes and regions considered to be prime candidates for containing genes or regulatory elements important in controlling risk to oral clefts (Table I). After quality control (QC), 1,409 case-parent trios remained available for analyses. We analyzed the case-parent trio data in two separate groups (shown in Table II): an Asian group which contained the Filipino and Chinese families (1,034 case-parent trios), and a smaller European group composed of European and European-American families (375 trios).

Table II.

Number of case-parent trios, by population, available for analysis after quality control.

Population	Country	Total Trios
Asian	China	401
	Philippines	633
Asian TOTAL		1,034
European	USA	266
	Denmark	9
	Hungary	65
	Spain	26
	Turkey	9
European TOTAL		375
TOTAL		1,409

Sequencing

Details of the sequencing protocol are available in Leslie et al. [2015]. In brief, 1µg of native genomic DNA was used to construct Illumina multiplexed libraries. Reads were mapped to the GRCh37-lite reference sequence using BWA [Li and Durbin, 2010]. Picard was used to merge alignments and mark duplicates, and Polymutt was used to perform germline and de novo variant calling. We used bam-readcount to identify and flag potential artifact variants [Leslie et al. 2015]. The single nucleotide variant calls were combined into a variant call file (VCF) file. All variants with a depth (DP) less than 7 or genotype quality (GQ) less than 20 were removed. Variants located within 75bp of indels or dinucleotide polymorphisms occurring in more than 5% of samples were included in analyses, but were flagged as potential artifacts.

Data processing

To evaluate the family relationship between members of the case-parent trios, we used BEAGLE’s fast-IBD to calculate identity by descent (IBD) between parents and their offspring. If a parent-child pair shared less than 40% of the targeted region, the trio was dropped from all analysis. To increase the power to detect G×G interaction, we only selected highly polymorphic variants with a minor allele frequency (MAF) larger than 20%. We also excluded all SNPs with missing genotype rate larger than 1%. We tested for Hardy-Weinberg equilibrium (HWE) in parents separately within Asian and European groups, and excluded SNPs that showed substantial departure from HWE (p<1×10⁻⁵). We used Haploview [Barrett et al., 2005] to choose tagging SNPs (defined as r² >0.8) within the Asian and European groups. Haplotype phasing required for the permutation tests was carried out using BEAGLE [Browning and Yu, 2009].

Screening step

To assure scalability, we implemented a screening step to generate candidate interactions among all pairwise tagging SNP combinations in the 13 genomic regions of interest (excluding pairs of SNPs from the same region), using a 1df Wald test for G×G interaction based on a conditional logistic regression model implemented in the open source Bioconductor package trio [Schwender et al. 2014]. As in the genotypic TDT [Schaid 1999], each trio is represented as the four possible Mendelian offspring given the parental genotypes, i.e. where the affected proband is the case and the other three possible genotypes serve as “pseudo-controls”. The conditional model (with grouping by family) assumes an additive mode of inheritance for both bi-allelic SNPs, and contains one parameter for interaction between the two SNPs. This model is considerably simpler than the more general modeling approach proposed by Cordell [2002], and allows for the rapid assessment of all pairwise interactions. The Cordell 4df LRT (see next section) requires numerical estimation of four and eight parameters respectively in conditional logistic regression models using an iterative procedure.

General model to assess interaction

Significant results based on the 1df Wald test can indicate epistasis, but also simply a violation of model specifications, in particular the interaction term itself. To avoid such biases, we selected the 500 most significant marker pairs from each of the $\left(\begin{array}{c}\hfill 13\hfill \\ \hfill 2\hfill \end{array}\right)=78$ pairwise combinations and comprehensibly assessed the potential SNP-SNP interactions using the model proposed by Cordell [2002]. Again representing each trio as the four possible offspring based on the parental genotypes (one affected “case” and three “pseudo-controls”), this model can for the i-th trio be written as

$$\mathit{\text{logit}}(p_i)={{\beta }_1}^{*}{X}_{Aai}+{{\beta }_2}^{*}{X}_{AAi}+{{\gamma }_1}^{*}{X}_{Bbi}+{{\gamma }_2}^{*}{X}_{BBi}+{i_{11}}^{*}{X_{AAi}}^{*}{X}_{BBi}+{i_{12}}^{*}{X_{AAi}}^{*}{X}_{Bbi}+{i_{21}}^{*}{X_{Aai}}^{*}{X}_{BBi}+{i_{22}}^{*}{X_{Aai}}^{*}{X}_{Bbi}$$

where X_Aa and X_AA are orthogonal representations of the genotype at locus A, and X_Bb and X_BB are orthogonal representations of the genotype at locus B (Cordell [2002]). The four parameters (i₁₁, i₁₂, i₂₁, i₂₂) encode the possible departure from additive genotype effects. To assess SNP-SNP interaction, we use a 4df LRT to compare the likelihood of the full model above to the likelihood of the reduced model without these epistatic parameters:

$$\mathit{\text{logit }}(p)={{\beta }_1}^{*}{X}_{Aa}+{{\beta }_2}^{*}{X}_{AA}+{{\gamma }_1}^{*}{X}_{Bb}+{{\gamma }_2}^{*}{X}_{BB}.$$

Permutation tests

We carried out permutation test to assess statistical significance in the light of multiple hypothesis testing. Specifically, we aimed to answer the following questions: (1) whether any region pair(s) showed evidence for interaction, i.e. is there any signal in the data? And (2) if so, which region/gene pairs and which SNPs within such a region pair interact. As our permutation data set, we created 1000 “shuffled” data sets using the phased parental haplotypes, to maintain the correlation structure between genotypes within a region. To create simulated genotypes for the children, we randomly chose one haplotype from each of the parents as the transmitted haplotype, yielding a simulated case and three simulated pseudo-controls for each trio.

For each of the 1,000 permutation data sets, we repeated our analytic procedure separately for each of the $\left(\begin{array}{c}\hfill 13\hfill \\ \hfill 2\hfill \end{array}\right)=78$ pairs of regions, for each region pair selecting the 500 most significant SNP pairs in the screening 1df Wald test, and evaluating these 500 candidates using the 4df Cordell LRT. For each permutation data set, we recorded the maximum test statistic from the LRT across all SNP-SNP interactions in the respective region pair. This yielded a set of 78×1,000 (maximum) test statistics representing results under the null for each of the 78 possible region pairs, across the 1000 permutations.

To assess whether any region pair(s) showed evidence for interaction, i.e. whether any signal is present in the data, we compared the maximum 4df test statistic observed for any SNP pair across all region pairs to the 1,000 corresponding values obtained in the same fashion from the permutation data. The p-value for this test was estimated by the fraction of permuted test statistics exceeding the observed. To assess which of the 78 region pairs (and which SNP pairs) show deviation from randomness, we compared for each region pair the maximum 4df test statistic observed for any SNP pair within the respective region pairs to the 1,000 corresponding values obtained in the same fashion from the permutation data. The p-value for each region pair was estimated by the fraction of permuted test statistics exceeding the observed. Significance under family-wise error rate protection for the 78 region pairs was assessed using a Bonferroni correction.

Due to the computational expense of our procedure, we considered the region-pair selection as a hypothesis-generating step, with the assessment of a particular SNP-SNP combination as the step requiring strict multiple-testing correction to control the family-wise error rate. Hence any SNP-SNP combination should have a permutation adjusted p-value less than 0.05/78=0.00064. Since 1000 permutations is not adequate to reach this level of significance, for any region pair that passed the first selection process, we generated an additional 1000 permuted data sets to further assess significance.

Results

We obtained 1,075 and 1,016 tag SNPs in European and Asian groups, respectively, after applying quality control filters to the common SNPs from the targeted sequencing data and selecting tagging markers. We only investigated interactions between different genes/regions, resulting in 78 different gene/region combinations for a total of 519,086 and 468,037 hypothesis tests among European and Asian trios, respectively. The most significant interaction based on the 4df LRT among Europeans was between rs6681355 in IRF6 and rs6029315 in MAFB (LRT=40.25, p=3.8×10⁻⁰⁸, Table III). Among the 1,000 permutations, the maximum test statistic across all SNP-SNP interactions in all 78 region pairs for Europeans exceeded this test statistic only 5 times (Figure 1A), yielding an empirical p-value of 0.005 with an upper bound for the 95% exact Binomial confidence interval of 0.0116 for this first hypothesis-generating stage. The maximum test statistic across all SNP-SNP interactions in the IRF6 / MAFB pair never exceeded the LRT of 40.25 (Figure 1B), yielding an empirical p-value < 0.001. The same was true in an additional independent 1,000 permutations, yielding a permutation p-value < 1/2,000 = 0.0005, and thus resulting in a p-value below the Bonferroni threshold of 0.05/78 = 0.00064.

Table III.

Top 10 most significant results from the 4 degree of freedom Likelihood Ratio Test for SNP-SNP interactions in 375 European case-parent trios (observed test statistic, nominal p-value, and nominal permutation p-value).

First Gene	Second Gene	Marker 1	Marker 2	Test Statistic	p-value	Permutation p-value
IRF6	MAFB	rs6681355	rs6029315	40.25	3.83E-08	<0.0005*
MAFB	NTN1	rs6029421	rs8081873	32.17	1.76E-06	0.02
MAFB	PAX7	rs6029182	rs11584404	31.25	2.72E-06	0.02
NOG	NTN1	rs8074637	rs2315286	29.62	5.84E-06	0.04
ARHGAP29	MSX1	rs4147848	rs730575	29.01	7.80E-06	0.03
8q24	MAFB	rs6470670	rs3092775	28.05	1.22E-05	0.05
MSX1	NTN1	rs2968669	rs9892906	27.97	1.27E-05	0.05
MAFB	MSX1	rs6029145	rs6851263	26.75	2.23E-05	0.06
IRF6	PAX7	rs2484030	rs10907314	25.72	3.60E-05	0.10
8q24	FOXE1	rs72730212	rs16923269	25.35	4.27E-05	0.10

^*The permutation p-value for the IRF6/MAFB pair was based on 2,000 permutations to assess whether its value is below the Bonferroni threshold of 0.05/78 = 0.00064.

Figure 1.

Distributions of the maximum 4 degree-of-freedom likelihood ratio test statistics, across all pairwise SNP-SNP interactions, for 1,000 permutation case-parent trio data sets of European ancestry. (A) The maximum test statistic across all SNP-SNP interactions in all 78 region pairs. The largest observed test statistic (LRT=40.25) is indicated by the red vertical line. (B) The maximum test statistic across all SNP-SNP interactions in the IRF6 / MAFB pair. (C) The maximum test statistic across all SNP-SNP interactions in all region pairs except the IRF6 / MAFB pair. The largest observed test statistic (LRT=32.17) is indicated by the blue vertical line.

No other region pair however showed significant interaction after multiple comparisons correction in Europeans. Removing the IRF6 / MAFB pair from consideration in the permutation test yielded a p-value of 0.362 for the maximum LRT statistic across all SNP-SNP interactions in the remaining 77 region pairs for the Europeans (Figure 1C). With the exception of this IRF6 / MAFB pair, no permutation p-value was observed below 0.05/78 = 0.00064.

Although there were more case-parent trios of Asian than European descent, we found no indication of SNP-SNP interactions among the Asians. The most significant interaction based on the 4df LRT among Asians was between rs3761910 in ARHGAP29 and rs2149722 in PTCH1 (nominal p=7.7×10⁻⁰⁶, Table IV), which did not retain significance after correcting for multiple comparisons.

Table IV.

Top 10 most significant results from the 4 degree of freedom Likelihood Ratio Test for SNP-SNP interactions in 1,034 Asian case-parent trios (observed test statistic, nominal p-value, and nominal permutation p-value).

First Gene	Second Gene	Marker 1	Marker 2	Test Statistic	p-value	Permutation p-value
ARHGAP29	PTCH1	rs3761910	rs2149722	29.04	7.67E-06	0.03
NOG	NTN1	rs17821518	rs12452003	28.79	8.63E-06	0.04
MAFB	NTN1	rs13041631	rs72809908	28.16	1.16E-05	0.05
BMP4	FGFR2	rs2738265	rs2936861	27.41	1.64E-05	0.03
FGFR2	PAX7	rs12763463	rs11488726	26.23	2.84E-05	0.18
8q24	ARHGAP29	rs873232	rs3789398	25.82	3.44E-05	0.10
8q24	IRF6	rs10111530	rs6540559	24.94	5.16E-05	0.14
NOG	PAX7	rs227723	rs2236832	24.74	5.67E-05	0.22
8q24	NTN1	rs10956419	rs2429370	24.39	6.66E-05	0.26
MSX1	NOG	rs9291153	rs7222986	24.34	6.81E-05	0.07

Discussion

Compared to other large scale studies searching for evidence of gene-gene (G×G) interactions, our study implemented an efficient screening strategy to screen pairwise combinations of all highly polymorphic SNVs and focused more specific tests on the most promising pairs of markers. The 4df interaction model proposed by Cordell [2002] can detect a variety of epistatic interactions even if the individual markers (or the genes they tag) do not display detectable marginal gene effects. Moreover, to account for correlations between markers within a region due to LD between SNPs, we performed permutation testing which can control for multiple comparisons more effectively than a Bonferroni correction.

We detected evidence of a possible G×G interaction between markers in and around IRF6 and MAFB (rs6681355:rs6029315; empirical p<0.0005) in the 375 trios in the European group. This evidence of statistical interaction between SNPs in IRF6 and MAFB is especially interesting, because IRF6 is one of a few recognized NSCL/P loci showing consistent evidence of association with risk to NSCL/P across different populations, and some of its functional activities have been identified [Leslie and Marazita, 2013]. For instance, in humans mutations in IRF6 cause Van der Woude syndrome, the most common Mendelian syndrome involving oral clefts [Kondo et al. 2002], while common variants in regulatory elements confer risk of NSCL/P [Rahimov et al. 2008]. Studies in animal models have characterized IRF6 expression patterns (Knight et al. 2006) and elucidated the identities of several members of the regulatory network for IRF6 [Kousa and Schutte, 2015]. Less is known about MAFB, which resides on 20q12 and was first identified in a GWAS study [Beaty et al. 2010]. Expression studies in the mouse support some role for MAFB in palatal development [Beaty et al. 2010]. Because little else is known about the role of MAFB in craniofacial development, identifying a statistical interaction between MAFB and IRF6 is an important step. Interestingly, interactions between WNT5B and MAFB and between WNT5A and IRF6 have also been identified [Li et al. 2015], which could represent the beginnings of a new interaction network for palatal development.

We failed to detect significant G×G interaction in the Asian group of case-parent trios, despite its larger sample size. Many factors could limit our ability to detect G×G interaction between these same SNPs in this larger Asian group. Although we have a large sample of 1,034 Asian case-parent trios, markers can have different MAF across ancestral groups and some key genotypes might be under-represented in Asian populations, making it hard to fit the full interaction model for G×G interaction. In our study, we used tagging SNPs to reduce the number of multiple comparisons; however, relying on the most highly polymorphic tagging SNPs could also make it impossible to identify critical G×G interaction effects.

One of the limitations of our study was its modest sample size and low power to detect G×G interaction. Compared to detecting a marginal effect for any single SNP, detecting pairwise or 2-way G×G interaction requires much larger sample sizes, and it becomes hard to fit the 4 df G×G interaction model with its total of 8 parameters representing individual gene effects and their interactions. Even in large data sets with 2000–3000 individuals, it is difficult to detect epistasis for low frequency markers (i.e. those with MAF<0.1) [Emily et al., 2009]. Therefore our approach will only be powerful in detecting G×G interaction between highly polymorphic SNPs. Although we relied on an efficient two-stage screening strategy, the number of tests was still large. Another limitation of our study is that we only used parametric logistic regression models to test for G×G interaction. A major challenge of using traditional regression models to detect interaction is correctly specifying both the full and reduced models. Additionally, analyzing high-dimensional data, which often contains many potential interacting predictor variables, can lead to very sparse contingency tables with empty cells. Machine-learning or data-mining methods represent an alternative approach that do not rely solely a pre-specified parametric model.

Finally, the scope of our analysis was limited to targeted sequencing data on these 13 regions (see Table I) which were all previously shown by other studies to be associated with NSCL/P. Variants in other regions (i.e. those without significant marginal effects) could also be important in G×G interaction, but our study would have missed these completely. Interaction between SNPs in these regions and elements elsewhere in the genome will require more comprehensive genotyping or sequencing studies. Nonetheless, our evidence of significant G×G interaction between polymorphic markers in the IRF6 and MAFB genes in a group of case-parent trios of European ancestry is especially intriguing and should be explored more thoroughly.