Abstract
Phenotypic heterogeneity is a hallmark of complex traits, and genetic studies may focus on the trait as a whole or on individual subgroups. For example, in orofacial clefting (OFC), three subtypes – cleft lip (CL), cleft lip and palate (CLP), and cleft palate (CP) have been studied separately and in combination. It is more challenging, however, to dissect the genetic architecture and describe how a given locus may be contributing to distinct subtypes of a trait. We developed a framework for quantifying and interpreting evidence of subtype-specific or shared genetic effects in complex traits. We applied this technique to create a “cleft map” of the association of 30 genetic loci with three OFC subtypes. In addition to new associations, we found loci with subtype-specific effects (e.g., GRHL3 (CP), WNT5A (CLP)), as well as loci associated with two or all three subtypes. We cross-referenced these results with mouse craniofacial gene expression datasets, which identified promising candidate genes. However, we found no strong correlation between OFC subtypes and expression patterns. In aggregate, the cleft map revealed neither subtype-specific nor shared genetic effects operate in isolation in OFC architecture. Our approach can be easily applied to any complex trait with distinct phenotypic subgroups.
Introduction
Most complex human traits (defined as those with both genetic and non-genetic risk factors) exhibit some phenotypic heterogeneity and variable expression with potentially hundreds of significantly associated genetic risk factors showing strong evidence of association (i.e. achieving genome-wide significance in large studies). Determining the relevance of any particular genetic risk factor at the individual or family level. is a significant challenge. There are several methods to compare correlated quantitative phenotypes, including those using a multivariate regression framework and the correlation between multiple phenotypes in cohorts or samples of cases and controls. However, these methods estimate statistical correlation to measure phenotype relatedness and are not suitable for examining mutually exclusive disease subtypes. Further, when these methods make comparisons based on the entire genome, which lacks the specificity of a more localized approach and could obscure biologically meaningful relationships.
We propose a targeted approach based on summary statistics from genome-wide association to sort out phenotypic heterogeneity. To illustrate this approach, we apply it to orofacial clefts (OFCs), congenital birth defects affecting the face and oral cavity. OFCs are the most common human craniofacial birth defect and combined they occur in approximately 1 in 800 live births worldwide (Leslie & Marazita, 2013). Although there are many types of OFCs, the term is most commonly used to refer to clefts of the upper lip and/or palate. For our purposes, OFCs will refer to cleft lip (CL), cleft lip with cleft palate (CLP), or cleft palate (CP), the three most common types of OFCs. There is an additional combined category of CL with or without CP (CL/P), historically felt to be distinct from CP alone due to the separate embryological origins of the upper lip and secondary palate. Thus, within the CL/P group, CL and CLP have been considered variants of the same defect that only differed in severity (Marazita, 2012).
Notably, more recent extensive epidemiological, genetic, and biological data suggest a more complex relationship between CL, CLP, and CP with both common and unique etiologic factors. In population-based studies in Denmark (Grosen et al., 2010) and Norway (Sivertsen et al., 2008), the recurrence risk for siblings was not uniform. The recurrence risks are consistently highest within the same subtype—for example an individual with CLP is more likely to have a sibling with CLP rather than CL or CP—supporting the possibility of subtype-specific effects. Further, “between-subtype” recurrence risks for CL and CLP—for example an individual with CL having a sibling with CLP or vice versa—are lower than within-subtype risks, but are not equal, lending support for the hypothesis that genetic risks for CL and CLP may differ. The lowest recurrence risks were “between-subtype” risks involving CP, but were still higher than the baseline risk in the general population, suggesting some shared etiology between CL/P and CP. In the case of multiplex OFC families, the affected individuals in such families most commonly all have CL/P or have CP. Notably, there are also mixed multiplex families with both CL/P and CP present among relatives, most commonly seen with syndromic forms of OFCs.
Taken together these observations imply a genetic predisposition for specific OFC subtypes, but there is limited genetic evidence supporting subtype-specific risk factors, especially for CL versus CLP. The primary focus of the OFC genetics literature has been on CL/P, where over 25 genetic risk loci have been identified to date from genome-wide studies, accounting for a modest portion (~30%) of the overall genetic variance for risk to CL/P (M. J. Dixon et al., 2011; Leslie et al., 2017; Leslie & Marazita, 2013; Ludwig et al., 2017; Yu et al., 2017). By contrast, only one locus has been identified for CP (Leslie, Liu, et al., 2016). To date, subtype-specific associations are limited to three loci: 13q31 near SPRY2 and GREM1 (15q13) associated specifically with CLP (Jia et al., 2015; Ludwig et al., 2016; Ludwig et al., 2012), and GRHL3 (1p36) associated with CP (Leslie, Liu, et al., 2016; Mangold et al., 2016). There is some evidence that markers near IRF6 (1q32) have a stronger effect on risk for CL than CLP, but this has not been consistently replicated (Leslie & Marazita, 2013; Rahimov et al., 2008).
Given the growing body of evidence suggesting the presence of subtype-specific signals and the broader knowledge base of shared signals, we hypothesize neither type of statistical signal operates in isolation to affect craniofacial development. Rather, it is the combination of shared risk loci and perhaps subtype-specific risk loci affecting an individual’s risk for OFC and their specific OFC subtype. In the current study, we sought to identify novel genetic risk variants for three specific OFC subtypes, CL, CLP, and CP, and examine all genetic risk loci for evidence of being specific to only one OFC subtype or of being shared between two or more subtypes of OFC.
Results/Discussion
Genome-wide meta-analysis of CL, CLP, and CP
First, we performed genome-wide meta-analyses for CL, CLP, and CP using imputed genotype data from the GENEVA and Pittsburgh Orofacial Cleft (POFC) consortia (Figure 1, Table S1). The GENEVA consortium used a family-based design and included 461 case-parent trios with CL, 1143 case-parent trios with CLP, and 451 case-parent trios with CP after removing individuals overlapping the two consortia. The POFC consortium included both a case-control arm and a case-parent trio arm, comprising 179 cases and 271 case-parent trios with CL, 644 cases and 1,048 case-parent trios with CLP, 78 cases and 165 case-parent trios with CP, plus 1,700 unaffected controls with no known family history of OFC drawn from the same populations as the unrelated cases. In the POFC case–control subgroup, we used logistic regression to test for association under an additive genetic model (Leslie, Carlson, et al., 2016). The two case-parent trio subgroups from POFC and GENEVA were analyzed separately using the allelic transmission disequilibrium test (TDT). The resulting effect estimates for the three analysis groups were combined using an inverse-variance weighted fixed-effects meta-analysis. This procedure was followed separately for genome-wide meta-analyses of CL and CLP (Figure S1); the results of the meta-analysis of CP was previously published (Leslie et al., 2017) and are also depicted in Figure S1.
From these three analyses by cleft type, a total of 1,231 SNPs across 29 loci demonstrating suggestive evidence of association (i.e. p<1.00×10−5) in any of the three analyses were selected for further follow-up (Figure 1, Table S2). In addition to the 19 recognized risk loci previously reported in GWAS of CL/P or CP separately (Leslie et al., 2017; Yu et al., 2017), this study serves as independent replication for 17q21.3 (near WNT9B and GOSR2). Nine additional risk loci, although not reaching formal genome-wide significance, were suggested. Three of these— on 1p36.3, 3p14.3, and 5q35.2—have obvious candidate genes previously implicated in craniofacial development or in human craniofacial anomalies. At 1p36.3, a balanced translocation disrupting the CAPZB gene was reported in a patient with micrognathia and CP; subsequent studies showed capzb(-/-) zebrafish mutants recapitulated these human phenotypes (Mukherjee et al., 2016). The 5q35.2 signal is adjacent to MSX2, a gene critical for human skull development and associated with craniosynostosis, parietal foramina, and orofacial clefting (Wilkie et al., 2000). Finally, the 3p14.3 locus had two independent signals within the same topologically-associated domain containing WNT5A, a gene in which mutations can cause mandibular hypoplasia in Robinow syndrome (Hosseini-Farahabadi et al., 2017; Person et al., 2010) and ERC2, encoding a synapse protein (Ohtsuka et al., 2002). As WNT5A is a stronger candidate gene for OFCs than ERC2, we represent the two signals as WNT5A “a” and WNT5A “b”.
Identification of independent signals
The 8q24 and IRF6 loci represent large genomic intervals and have previously shown multiple, statistically independent associations with various OFC phenotypes, although the independence of signals has only been formally tested for the 8q24 gene desert region (Leslie, Carlson, et al., 2016). We separated SNPs into multiple groups based on LD “clumps” calculated with PLINK software (Purcell et al., 2007). In doing so, we confirmed the presence of multiple independent signals in the IRF6 region and found evidence for a third signal at the 8q24 region (Figure S3). In total, three loci represented multiple independent signals—1q32 (IRF6), 8q24, and 3p14.3 (WNT5A)—thus the 29 associated loci comprised 34 independent signals.
Comparisons of CL, CLP, and CP
With 1,231 associated SNPs from the 34 independent signals revealed by our three meta-analyses, we used a heterogeneity Q-statistic (Schenker & Gentleman, 2001) to compare effects for each SNP among subgroups to determine if any of these loci showed evidence of subtype-specific or shared risk (Figure 1). Specifically, effects of CL were compared to those of CLP, and the effects of CLP compared to those of CP. These contrasts were selected based on the biological plausibility of shared genetic effects between clefts affecting the lip (CL and CLP) and clefts affecting the palate (CLP and CP). To aid in the identification of subtype-specific variants, the direction of association was calculated by the difference in absolute values of the log odds ratios (i.e. |log(ORCLP)| – |log(ORCL)|, |log(ORCLP)| – |log(ORCP)|).
We sought to visually represent these findings so the evidence of subtype-specific effects become clear. Rather than dichotomizing genetic effects as “subtype-specific” or “shared”, we wanted to represent both the statistical evidence for heterogeneity and the overall statistical evidence of association with cleft subtypes for each locus. To this end, we developed a graphical representation, hereafter referred to as “the cleft map” to describe the statistical effect of numerous genetic loci on the architecture of OFCs. On the cleft map, the position of any single SNP is determined by the sum of two vectors, each given by the –log10 p-value of the heterogeneity Q-statistic times the sign of the direction of the locus (Figure 1). The position of a SNP thus represents how heterogeneous its estimated effects were between two cleft subtypes; SNPs further from the origin demonstrate more statistical evidence of heterogeneity. The x-axis of the cleft map represents the CL vs. CLP comparison and the y-axis represents the CLP vs. CP comparison.
When all SNPs for a locus were plotted, they generally clustered in specific locations based on the cleft type(s) with which the locus is associated (Figure 2, Figure S4). For example, SNPs located near the origin of the axes were those that showed less evidence of cleft-subtype specificity. Such a pattern occurred for all 4 tested SNPs at the FOXE1 locus (Figure 2A), consistent with existing literature indicating associations with all OFC subtypes (Leslie et al., 2017; Moreno et al., 2009). In contrast, all 13 SNPs at the GRHL3 locus (Figure 2B) showed a significant CP-specific association (pCLP.CP<6.2×10−5) and are positioned along the y-axis in the lower half of the map. Similarly, all 126 SNPs in the IRF6 “a” locus (Figure 2C) showed a significant association with CL/P, as evidenced by significant differences in the CLP-CP comparison (pCLP.CP<6.9×10−5), but there was no evidence of difference between CL and CLP (pCLP.CL>0.26). The independent IRF6 “a” and IRF6 “b” regions contain rs2235371 (Zucchero et al., 2004) and rs642961, respectively. Previous studies suggested that rs642961 was preferentially associated with CL (Rahimov et al., 2008). Although IRF6 “b” SNPs show a quite complex relationship with different cleft types probably related to underlying differences in linkage disequilibrium (LD) (Figure S3), the location of the SNP cluster containing rs642961 in the upper-left quadrant of the cleft map supports a stronger effect on risk for CL compared to CLP (Figure 2D).
With a few exceptions, the tight clustering of SNPs allowed us to simplify the map by representing each locus region centered around its SNP cluster (Table 1, Figure 3). In addition to the loci described above, several others demonstrated some evidence of cleft subtype specificity including MSX2 with CLP (Figure 2E, pCLP.CP<7.3×10−4, pCLP.CL<5.6×10−3), WNT5A “a” with CLP (Figure S4F, pCLP.CP<0.09, pCLP.CL<11×10−4), 9q21.32 with CL (Figure 2F, pCLP.CP=0.24, pCLP.CL=4.06×10−5), and 5p13.2 with CP (Figure S4K, pCLP.CP<2.2×10−4).
Expression analyses in associated loci
Above we have referred to each associated locus by either a plausible candidate gene (e.g., IRF6) based on the literature on OFCs and craniofacial development or as a genetic location (e.g., 8q24) for gene deserts or new loci. Although this comports with the standards of the field, recent work demonstrates that disease-associated variants can regulate distant genes, suggesting that the nearby genes prioritized by GWAS may not always be involved (Gupta et al.; Smemo et al.). In addition, multiple genes in a given region may be co-regulated and expressed in similar tissues or expressed in distinct compartments within the craniofacial complex (Attanasio et al., 2013). We, therefore, wanted to agnostically explore gene expression profiles of all genes contained in these regions to determine if their expression profiles could explain clustering of associated genes or provide mechanistic insights into the pathogenesis of OFC subtypes.
We first identified all of the genes located in the same topologically associated domains (TADs), because the associated SNPs may have regulatory functions and there is evidence that such SNPs are more likely to act upon genes located within their own TAD (Table S3) (Dekker et al.). We used published topological data from human embryonic stem cells because although there are no known craniofacial-specific TADs, boundaries are largely conserved across cell types (J. R. Dixon et al.). To prioritize highly expressed genes, we identified those genes with mouse homologs reported to be differentially expressed among several transcriptomic datasets from key periods, regions, and tissues in mouse facial development (Hooper et al.). Hooper et al. previously integrated these datasets with a weighted gene co-expression network analysis to generate 75 co-expression modules describing gene expression in the developing mouse face. We used these modules to annotate each gene in our list (Table S4). Of the 222 homologs in our 29 loci, 101 (45.44%) were present among these co-expression modules. Each region contained at least one gene assigned to a co-expression module, but only seven regions contained a single gene or only one gene with documented craniofacial expression (Figure S5). Overall, this is not surprising given these co-expression modules contained over 8,000 genes, or approximately 30% of the mouse genome.
To further refine the possible candidate genes, we turned to a complementary resource of gene expression profiles, SysFACE (Systems tool for craniofacial expression-based gene discovery) (Table S5) (Liu et al., 2017), which allows easy visualization of data from orofacial tissue microarrays or RNA-seq datasets for mandible, maxilla, frontonasal prominence, and palate, collected largely as part of the FaceBase consortium (Brinkley et al., 2016; Hochheiser et al., 2011). We used SysFACE to examine enriched craniofacial expression by comparing orofacial tissue data with embryonic whole body tissue (Lachke et al., 2012). For most loci, the SysFace analysis corroborated the co-expression modules or prioritized one gene. For example, at the PAX7 locus, although three genes (Pax7, Klhdc7a, and Aldh4a1) were found in craniofacial co-expression modules only Pax7 showed strong enrichment in SysFACE.
For the six new loci without any clear candidate genes (5p13.2, 6p22, 8q22.3, 9q21.32, 13q32.3), we used both datasets to identify likely candidate genes. At the 5p13.2 locus, Skp2 was present in the mesenchyme expression module and showed high SysFACE scores across multiple processes in the microarray data and RNA-seq expression in the palate. Other genes (Capsl and Slc1a3) were in the ectoderm module with enriched expression in the palate. Because this locus is associated only with CP, any or all of these genes could be relevant. At the 6p22 and 8q22.3 loci, multiple genes were found in both the Hooper and SysFACE datasets, but were expressed in different expression modules. Although FZD6 (8q22.3) was previously implicated in CLP by linkage in a large multiplex CLP family, as well as by craniofacial anomalies observed in a fzd6 morphant (Cvjetkovic et al., 2015), it is possible multiple genes contribute to the association signals at these loci. In sum, each newly associated locus contained one or more genes with craniofacial expression; detailed in vivo analyses will be required to pinpoint specific causal genes.
We performed an enrichment analysis for the set of genes present in these co-expression modules. These genes were enriched for broad biological processes such as “embryonic morphogenesis”, “epithelium development” and human and mouse phenotypes related to OFCs, including “oral cleft”, “perinatal lethality”, “abnormal craniofacial morphology”. Unfortunately, the broad terms from the enrichment analysis did not support more specific hypotheses about the pathogenesis of OFCs. The co-expression modules revealed a critical role for ectodermal genes in OFC pathogenesis, and fewer loci with mesenchymally-expressed genes. However, most genes were broadly expressed across multiple facial prominences, limiting our ability to hypothesize about any one mechanism for how these genes relate to OFC subtypes, but these tools will be useful for prioritizing genes for future association studies. A few genes, however, had very specific expression patterns worthy of further discussion. As one example, PAX7 expression was restricted to the frontonasal prominence, whereas other loci (i.e., SPRY2, MSX2) that clustered nearby in the cleft map showing stronger evidence of effects on risk for CLP were more broadly expressed in the maxilla and mandible. Similarly, the COL8A1 locus showed a stronger effect on CL than CLP, but was still very strongly expressed in the palate. Interestingly, Col8a1 expression was enriched early at E10.5 in the frontonasal and maxillary prominences, when lip fusion takes place. These patterns are consistent with the direction of the SNP effects at this locus where the same alleles conferred a protective effect (OR < 1) for CL and CLP scans, but a modest, (and not formally significant) risk effect (OR > 1) in the CP scan. Thus, our data argue both the early expression and the palatal expression are important. It is possible that the as-yet-unknown functional SNPs could promote palatal fusion and protect against a CP; alternatively, they could dysregulate COL8A1 promoting ectopic expression in the lip. Our current study cannot definitively answer these questions, but by demonstrating this locus may exert a stronger effect on risk to CL while the gene is expressed in the palate, it could motivate more targeted follow-up studies.
In our statistical analyses, we observed most of the loci previously identified in CL/P GWAS were positioned along the y-axis, indicating that there was no statistical difference between CL and CLP. This is not surprising, as a combined study of CL and CLP has the greatest power to identify loci with similar effects in both of these subgroups. One limitation of this study is that we did not capture several loci discovered in previous CL/P GWAS (e.g., RHPN2 (Leslie, Carlson, et al., 2016)) because no SNP showed p-values better than 10−5. As these loci were found only in the combined CL/P group, we would expect them to be positioned along the y-axis. Such SNPs show the statistical power of traditional analyses of CL/P, which have been successful. However, our study also demonstrates there are multiple loci with subtype-specific effects (e.g. WNT5A) or with stronger effects for CL than CLP (e.g. COL8A1) that are more difficult to detect in the combined analyses.
An important contribution of this work was the careful examination of the three large loci with multiple, independent signals based on LD. Of these, we found evidence for independent signals within WNT5A and IRF6 exerting potentially different effects between subtypes. In contrast, the three 8q24 signals were largely overlapping and associated with the combined phenotype CL/P (Figure S2 M-O); only the 8q24 “c” signal showed even marginal evidence of a stronger effect with CL than CLP. At the IRF6 locus, the “b” signal was tagged by rs642961, a SNP that disrupts the binding site of TFAP2α in the MCS9.7 enhancer (Rahimov et al., 2008). However, MCS9.7 activity did not completely recapitulate endogenous IRF6 activity, most notably in the medial-edge epithelium during palatal fusion, indicating the presence of some other enhancer (Fakhouri et al., 2012). Similarly, the 8q24 gene desert contains multiple craniofacial enhancers (Attanasio et al., 2013) which influence Myc expression (Uslu et al., 2014). Such enhancers are known to have restricted activity patterns and often act in a modular fashion to control gene expression by activating expression in different anatomical regions or at different points during development (Visel et al., 2009). These characteristics present a logical mechanism to drive phenotypic heterogeneity. Understanding the logic of these enhancers, the function of SNPs within them, and the relationship between risk alleles and disease subtypes will be critical for fully understanding the etiology of OFCs.
Our approach works well to describe SNPs and loci where the direction of the effect is the same among subtypes. However, when the odds ratios exhibit different effect directions, the log odds ratios may be close in value, and the sign representing the direction of the effect may fluctuate according to the subtype with a slightly larger effect at any given SNP. One can diagnose these cases by plotting estimated effects of the individual SNPs, resulting in two clusters of SNPs on opposite ends of the plot (as seen in Figure S4Z). Overall, plotting the centroid (as shown for the NOG locus (Figure S4AA)) most closely represents the overall picture of the locus—i.e. that there isn’t an association with one particular cleft type. However, if an allele truly increases risk for one cleft type and reduces risk for another, as was recently reported for NOG (Moreno Uribe et al., 2017), we may be obscuring biologically meaningful results. Similarly, there is a set of SNPs within IRF6 “c” (apparently independent of the IRF6 “a” and IRF6 “b” signals), with p-values in the CP meta-analysis of ~8×10−3 and whose minor alleles appear to increase risk for CP; these same alleles appear to decrease risk for CL/P. Although genetic association studies overwhelmingly support an association between common SNPs in IRF6 and risk of CL/P, dominant mutations in the gene cause Van der Woude syndrome, recognized as one of few syndromes where both CL/P and CP occur within the same family. With this new information about this locus, it may be time to revisit the idea that common SNPs within this locus could act as modifiers of which OFC phenotype appears in Van der Woude families (Leslie et al., 2013).
In summary, we developed a novel approach for dissecting the genetic contribution to phenotypic heterogeneity of OFCs. We identified novel genetic associations, some of which, like MSX2, showed evidence of subtype specificity, and would have been missed by traditional analyses where all subtypes are combined. Three loci, IRF6, WNT5A, and 8q24 represent multiple independent signals. Ours is the first study to formally confirm multiple signals from IRF6 locus, which had been suggested previously (Rahimov et al., 2008). Importantly, some of these signals showed different effects on the different OFC subtypes, adding an additional layer of complexity to the genetic architecture of this most common group of craniofacial malformatons. Finally, cross-referencing our results with gene expression data has generated new hypotheses about mechanisms by which OFC subtypes may occur. However, genetic studies alone are unlikely to completely elucidate the etiology of OFCs or other complex traits. Here, we focused only on fetal contributors to OFC risk, but the etiology of this disorder likely includes significant environmental components, epigenetic factors, parent-of-origin effects, stochastic processes, or additional genetic modifiers. With respect to OFCs, we have previously shown evidence of both common and rare (Carlson et al., 2017) genetic modifiers possibly distinguishing between CL from CLP. Our approach may serve to guide targeted tests for gene-gene interaction and risk score analyses to further disentangle the complex etiologic architecture of OFCs. Future extensions of this approach can incorporate these modifiers and interactions, and can examine other subtype definitions, other structural birth defects, or any other complex trait with phenotypic heterogeneity where subtypes can be delineated. Finally, we note our approach uses only GWAS or candidate gene study summary statistics, so information from multiple studies can more easily be leveraged to dissect the complex architecture of heterogeneous traits.
Methods
Contributing GWAS studies
Two consortia contributed to this study. The first, from the GENEVA consortium, used a family-based design and included 461 case-parent trios with cleft lip (CL), 1,143 case-parent trios with cleft lip and palate (CLP), and 451 case-parent trios with CP, respectively, from populations in Europe (Denmark and Norway), the United States, and Asia (Singapore, Taiwan, Philippines, Korea, and China). The specifics of this study have been described previously (Beaty et al., 2010; Leslie, Carlson, et al., 2016; Leslie, Liu, et al., 2016). Briefly, samples were genotyped for 589,945 SNPs on the Illumina Human610-Quadv.1_B BeadChip, genetic data were phased using SHAPEIT, and imputation was performed with IMPUTE2 software to the 1000 Genomes Phase 1 release (June 2011) reference panel. Genotype probabilities were converted to most-likely genotype calls with the GTOOL software (http://www.well.ox.ac.uk/~cfreeman/software/gwas/gtool.html).
The second consortium included samples from the Pittsburgh Orofacial Cleft (POFC) study, comprising 179 cases and 271 case-parent trios with CL, 644 cases and 1,048 trios with CLP, 78 cases and 165 trios with CP, and 1,700 unaffected controls with no history of craniofacial anomalies. Participants were recruited from 13 countries in North America (United States), Central or South America (Guatemala, Argentina, Colombia, Puerto Rico), Asia (China, Philippines), Europe (Denmark, Turkey, Spain), and Africa (Ethiopia, Nigeria). Additional details on recruitment, genotyping, and quality controls were previously described (Leslie, Carlson, et al., 2016; Leslie, Liu, et al., 2016). Briefly, these samples were genotyped for 539,473 SNPs on the Illumina HumanCore+Exome array. Data were phased with SHAPEIT2 and imputed using IMPUTE2 to the 1000 Genomes Phase 3 release (September 2014) reference panel. The most-likely genotypes (i.e. genotypes with the highest probability [Q]) were selected for statistical analysis only if the genotype with the highest probability was greater than 0.9.
A total of 412 individuals were in both the GENEVA OFC and POFC studies, so we excluded these participants from the GENEVA study for this analysis. Informed consent was obtained for all participants and all sites had both local IRB approval and approval at the University of Pittsburgh, the University of Iowa, or Johns Hopkins University. Individual level genotype and phenotype data for the GENEVA and POFC studies are available from dbGaP: phs000774.v1.p1 and phs000094.v1.p1.
Genome-wide meta-analyses for CL, CLP, and CP
We previously described our methods for quality control and meta-analysis of the GENEVA OFC and POFC studies (Leslie et al., 2017). GWAS was performed for CL, CLP, and CP separately on SNPs with minor allele frequencies (MAF) greater than 1% and not deviating from Hardy-Weinberg Equilibrium (HWE p > 0.0001) in the parents or control subjects. Tests of association using unrelated cases and controls matched for population of origin used logistic regression models, while case-parent trios were analyzed with or the allelic transmission disequilibrium test (TDT) implemented in PLINK. Within each OFC subtype, the resulting effects estimates were combined in an inverse-variance weighted fixed-effects meta-analysis. The combined estimate, a weighted log odds ratio, should follow a chi-squared distribution with two degrees of freedom under the null hypothesis of no association.
Comparisons of cleft types
From the three meta-analyses, SNPs demonstrating suggestive evidence of association (i.e. p < 1.00×10−5) in any scan were considered for further analysis. For each SNP, two heterogeneity Q-statistics were calculated to compare effects of CL to CLP, and CLP to CP (Schenker & Gentleman, 2001). The magnitude of the log odds ratios for these two cleft subtypes were also compared to indicate the direction of effect (i.e. which subtype showed a stronger effect). Together, the Q-statistics and directions of effect for the two comparisons (CL to CLP, CLP to CP) prescribe a point for each SNP on the cleft map. For each region, the centroid of these points was calculated to inform the overall effect of the locus.
SysFACE: mouse orofacial transcriptome data analysis
Mouse orthologs of human candidate genes were analyzed for their absolute expression and enriched expression in orofacial tissue microarray datasets on the Affymetrix 430 2.0 platform (GeneChip Mouse Genome 430 2.0 Array) deposited in NCBI GEO (https://www.ncbi.nlm.nih.gov/geo/) and FaceBase (https://www.facebase.org) (Table S5). Enriched expression was estimated by comparing orofacial tissue data with whole embryonic body tissue (WB) reference control obtained from series (GSE32334) as previously described (Lachke et al., 2012). Datasets for mandible (mouse embryonic stages E10.0, E10.5, E11.0, E11.5, E12.0, E12.5), maxilla (E10.5, E11.0, E11.5, E12.0, E12.5,) and frontonasal (E10.5, E11.0, E11.5, E12.0, E12.5) were obtained from the series GSE7759. Palate datasets were obtained for mouse E13.5 (FaceBase series: FB00000468.01), E14.5 (FB00000474.01, GSE11400) and P0 (GSE31004). RNA-seq data on Illumina HiSeq2500 platform for mouse E14.5 posterior oral palate (FB00000768.01), anterior oral palate (FB00000769.01) and WB (whole body, unpublished) were also used. ‘R’ statistical environment (http://www.r-project.org/) was used to import raw microarray datafiles on Affymetrix 430 2.0 platform, followed by background correction and normalization using Affy package (Gautier et al., 2004) available at Bioconductor (www.bioconductor.org). Using a AffyBatch function, present/absent calls for probe sets were calculated and those with the highest median expression at significant p-values were collapsed into genes (Gautier et al., 2004). Differential gene expression (DEGs) and enrichment scores for all four orofacial tissues compared to WB were calculated using limma (Ritchie et al., 2015), and the detailed microarray workflow is described elsewhere (Anand et al., 2015). RNA-seq data on Illumina HiSeq2500 platform were first subjected to quality control analysis for reads by using FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/), and then subjected to sequence trimming and clipping using Trimmomatic (Bolger et al., 2014) with in house scripts (https://github.com/atulkakrana/preprocess.seq) (Mathioni et al., 2016). Reads were aligned against the Mus musculus reference genome using TopHat v2.0.9 (Trapnell et al., 2009) with recommended settings. Transcript assembly for measuring relative abundances was performed using Cufflinks v2.1.1 (Trapnell et al., 2009). After merging the assemblies by the function Cuffmerge, DEGs were identified using the Cuffdiff function (Trapnell et al., 2009). Statistically significant DEGs (comparison of orofacial datasest with WB) were identified using an in-house Python script.
Acknowledgments
We thank the study participants for their enthusiastic participation and acknowledge the local recruitment staff and collaborators for their tireless efforts that made this study.
Footnotes
Grant Numbers: This work was funded by grants from the National Institutes of Health: R00-DE025060 [EJL], R03-DE024776 [SAL, IS], X01-HG007485 [MLM, EF], R01-DE016148 [MLM., SMW], R01-DE009886 [MLM], R21-DE016930 [MLM], R01-DE014667 [LMM], R01-DE012472 [MLM], R01-DE011931 [JTH], R01-DE011948 [KC], R37-DE008559 (JCM, MLM) U01-DD000295 [GLW], K99-DE024571 [CJB], S21-MD001830 [CJB], U54-MD007587 [CJB], R01-DE014581 [TB], U01-DE018993 [TB]. Genotyping and data cleaning were provided via an NIH contract to the Johns Hopkins Center for Inherited Disease Research: HHSN268201200008I.