A genome-wide scan of cleft lip triads identifies parent-of-origin interaction effects between ANK3 and maternal smoking, and between ARHGEF10 and alcohol consumption

Background: Although both genetic and environmental factors have been reported to influence the risk of isolated cleft lip with or without cleft palate (CL/P), the exact mechanisms behind CL/P are still largely unaccounted for. We recently developed new methods to identify parent-of-origin (PoO) interactions with environmental exposures (PoOxE) and now apply them to data from a genome-wide association study (GWAS) of families with children born with isolated CL/P. Methods: Genotypes from 1594 complete triads and 314 dyads (1908 nuclear families in total) with CL/P were available for the current analyses. Of these families, 1024 were Asian, 825 were European and 59 had other ancestries. After quality control, 341,191 SNPs remained from the original 569,244. The exposures were maternal cigarette smoking, use of alcohol, and use of vitamin supplements in the periconceptional period. Our new methodology detects if PoO effects are different across environmental strata and is implemented in the R-package Haplin. Results: Among Europeans, there was evidence of a PoOxSmoke effect for ANK3 with three SNPs (rs3793861, q=0.20, p=2.6e-6; rs7087489, q=0.20, p=3.1e-6; rs4310561, q=0.67, p=4.0e-5) and a PoOxAlcohol effect for ARHGEF10 with two SNPs (rs2294035, q=0.32, p=2.9e-6; rs4876274, q=0.76, p=1.3e-5). Conclusion: Our results indicate that the detected PoOxE effects have a plausible biological basis, and thus warrant replication in other independent cleft samples. Our demonstration of the feasibility of identifying complex interactions between relevant environmental exposures and PoO effects offers new avenues for future research aimed at unravelling the complex etiology of cleft lip defects.


Introduction
Cleft lip with or without cleft palate (CL/P) appears in approximately 3.4 to 22.9 per 10,000 live births 1 . Based on the severity of the cleft, patients undergo varying degrees of medical, dental, speech and psychosocial interventions over the first two decades of their lives, a long-term multidisciplinary treatment that not only imposes a heavy burden on patients and their families 2,3 , but also accounts for a substantial outlay in national healthcare budgets 4,5 .
Multiple genetic and environmental factors have been reported to influence the risk of CL/P, individually and through complex interactions in relevant biological pathways 6-10 . Major advances in high-throughput genotyping technologies, coupled with a boost in international collaborations, have led to substantial progress in gene-mapping for orofacial clefts, and the first wave of genome-wide association studies (GWAS) identified and replicated several key genes and loci associated with clefting [11][12][13][14][15][16] . Despite this success, the genetic variants identified so far collectively explain only a minor fraction of the total variance attributable to additive genetic effects, even though the heritability of CL/P among Europeans is more than 70% 17-20 . This has spurred renewed interest in investigating disease mechanisms other than fetal or maternal effects 21 . One example is parent-of-origin (PoO), where the effect of a particular allele in the offspring differs according to its parental origin [22][23][24] , and another is geneenvironment interaction (GxE), where fetal effects differ across strata of environmental exposures 25 . Identifying GxE effects may not only provide new insights into the causes of CL/P, but may also provide an opportunity to intervene on environmental risk factors alone, particularly in subgroups of the population that are genetically more susceptible to these environmental effects.
Recently, we went one step further and developed new methods for a genome-wide screening for PoO interactions with environmental exposures (i.e., PoOxE) in the case-parent triad setting 22 . We applied the new methodology, implemented in the R-package Haplin 26 , to isolated cleft palate only (CPO) 27 , using genotypes and exposure data from the largest published GWAS dataset on case-parent triads of orofacial clefts 11 . Epidemiological and embryological findings have previously shown that CL/P and CPO may have distinct etiologies. Therefore, we used the same GWAS dataset and methodology to perform a genome-wide scan for PoOxE effects in the larger sample of isolated CL/P.

Study participants
The study participants were mainly of Asian or European origin and were recruited as part of an international cleft collaboration 11 . Information was available on genotypes as well as maternal vitamin use, cigarette smoking and alcohol consumption in the periconceptional period (three months before and three months after pregnancy). The information on environmental exposures was based on interviews and questionnaires. More detailed characteristics of the study participants can be found in our recent work 28 . Table 1 shows the distribution of the CL/P families according to ethnicity, triad completeness and maternal exposure. There were 1908 families in the pooled sample (5424 individuals in total), which included all the participants. Of these, 825 families were in the European sample, 1024 families were in the Asian sample, and 59 families were in the sample consisting of other ethnicities (Table 1). We performed three main sets of analyses on the following samples: All participants (denoted as "pooled analysis"), only Asians ("Asian analysis"), and only Europeans ("European analysis"). The 59 families with other ethnicities were not analyzed as a group due to the small sample size, but they were included in the pooled analysis. In the pooled and European analyses, we examined all exposures. As cigarette smoking and alcohol consumption were rare among Asian mothers, we were only able to conduct PoOxVitamin analyses for this ethnicity.
Quality control for excluding single-nucleotide polymorphisms (SNPs) and individuals were conducted as described in Haaland et al. (2017) 27 . That is, we included SNPs with a missing call rate less than 5%, a minor allele frequency (MAF) greater than 5%, a p-value of less than 0.001 for the test for Hardy-Weinberg equilibrium presented by Wigginton et al. (2005) 29 , and a Mendelian error rate greater than 10%. Further, if two or more SNPs were in perfect linkage disequilibrium (r 2 =1) with each other, we only included one in the analyses. After applying these same criteria here, 341,191 were left for the current analyses from a total of 569,244 SNPs (Table 2).

Amendments from Version 1
In this version we have considered the remarks of Dr. Evie Stergiakouli. We have addressed some limitations that were pointed out and generally edited the manuscript for clarity. Abstract: Removed reference to previously published work and added information about the methodology. Introduction: Rephrased sentence that overemphasized how intriguing the problem of missing heritability is in clefting. Methods: Clarified that 59 individuals of other ethnicities than Asian or European were included in the pooled sample and added a sentence about the American Statistical Association's negative attitudes towards using a fixed p-value as a threshold for statistical significance. Results: Revised titles of Table 3, Table 4 and Table 7 to include information about which analyses were conducted. Discussion: Added a paragraph about genomic imprinting as a possible mechanism for PoO effects, clarified that our findings should be treated with caution until they have been replicated, and addressed the limitation that we have not imputed genotypes. In general: Corrected minor errors and slightly altered text to improve flow and precision.

Statistical analysis
For statistical analysis, we used the statistical software Haplin 26 , which is written in the R statistical programming language 30 . Haplin is based on log-linear modeling in a maximum likelihood framework and is well-suited for the analysis of offspring-parent triads. Because Haplin uses the expectation-maximization (EM) algorithm to account for missing parental genotypes 26 , we were able to include the 314 case-parent dyads in the analyses beside the complete triads (Table 1). Haplin also uses the EM algorithm to reconstruct haplotypes, which enabled haplotype analyses for different combinations of SNPs in the genes that showed a plausible PoOxE effect.
A detailed description of the method for PoOxE analysis has been provided in our previous works 22,27,31 . Briefly, PoOxE effects were calculated as follows: 1) Calculate the relative risk (RR) for an allele inherited from the mother (RR mat ) and do the same for the father (RR pat ).
2) Calculate the relative risk ratio (RRR PoO =RR mat /RR pat ) between the RRs in (1). RRR PoO is thus an estimate of the parent-of-origin (PoO) effect.
Haplin uses a Wald test to test the null hypothesis of RRR PoOxE =1.
In order to control for multiple testing (one test for each of 341,191 SNPs), we obtained q-values using the false discovery rate (FDR) method described by Storey & Tibshirani (2003) 32 . Specifically, the q-values were calculated from the p-values with the R-function qvalue() 33 . A q-value of 0.2 corresponds to an FDR of 20%, which means that at least 80% of SNPs with a q-value less than 0.2 would be expected to be truly associated with the outcome. As in our previous work on isolated CPO 27 , we identified the top 20 SNPs for each of the analyses performed (see Results for details) and calculated relative risk ratios (RRRs) with 95% confidence intervals (95% CI). We paid more attention to a given gene if SNPs in that gene showed up multiple times in one set or across different sets of analyses. In accordance with recent recommendations by the American Statistical Association and others 34,35 , we did not consider a fixed p-value as a threshold for statistical significance.
To illustrate the general ability of the PoOxE analyses to detect true associations, power analyses for a wide range of PoOxE scenarios were performed using the Haplin function hapPowerAsymp(), as described in our recent works 22,36 .
We focused on the regions flanking SNPs in the most interesting genes and constructed regional plots based on R-scripts developed by the Diabetes Genetics Initiative of Broad Institute of Harvard and MIT, Lund University and Novartis Institutes of BioMedical Research 37 . Such plots capture the extent of linkage disequilibrium between a lead SNP and neighboring SNPs, while also providing information on recombination patterns and the position of genes.
R-scripts used to conduct the statistical analyses and create figures are available (see Software availability) 38 .

Bioinformatics analyses
To contextualize the findings, we searched for connections among a selection of genes in the STRING database 39 , as well as for enrichment of these genes in expression patterns using ExpressionAtlas 40 and BGee (R package BgeeDB_2.10.0) 41 . Further, using Hetionet (Ver.1.0) 42 , we searched for indirect links between the genes highlighted by our analyses, the exposures and the phenotype ("cleft lip"). Hetionet is a heterogeneous network of various relationships among various data types, such as interactions between genes, or regulation of gene expression between a drug and a gene. The data used in Hetionet were carefully curated from 29 publicly available databases. To simplify the query output, the number of relationships between any two of the input query nodes (i.e., exposure, cleft lip, and the genes) was set to at most two. The exact queries together with their output are available (see Software availability) 38 .

Ethical statement
The individual institutional review boards of the members of the International Cleft Consortium provided ethical approval, which can be found in the online supplementary material of the original publication 43 . Written informed consent was provided by all participating families. Please refer to the dbGaP database for more information.

Results
For clarity, this section is structured as follows: We present the results of the PoOxE analyses of the pooled sample first (Table 3), followed by those of the European (Table 4) and Asian (Table 5) samples. We used the integrative database GeneCards and the gene-centric links therein to collate information on the genes in these tables. The 1000 Genomes browser was used to determine the chromosomal band location of a SNP. In the following sections, we focus on q-values, but all the corresponding p-values can also be found in Table 3-Table 5. Table 6 provides a reference for the full names of all the genes mentioned in Table 3-Table 5. Table 7 shows the results of the haplotype analyses of SNPs in the most interesting genes. Figure 1 and Figure 2 present visualizations of the results from the bioinformatics analyses, and regional plots for the most important regions from Table 3-Table 5 are shown in Figure 3 and Figure 4. Figure 5 illustrates power calculations to detect different PoOxE effects in single-SNP analyses under different parameters, such as different sample sizes and minor allele frequencies. Quantile-quantile (QQ) plots for each set of analyses are shown in Figure 6- Figure 8.

Pooled sample
All the top 20 SNPs in the PoOxAlcohol analysis had the same q-value of 0.99 and are therefore not considered here as they are probably false positives (Table 3). All the SNPs in the PoOxSmoke analysis had q-values of around 0.6. Even though these q-values are still quite large, they indicate that around 40% of the SNPs are potentially true PoOxE associations. Among the top 20 SNPs in the PoOxSmoke analysis, two are in the gene for 'Focadhesin' (FOCAD), two are in 'Platelet derived growth factor D' (PDGFD), two are in 'Ankyrin 3' (ANK3), and one is in 'Fraser syndrome 1' (FRAS1). Note that associations with PDGFD and ANK3 were also detected in the European analyses (see below). In the PoOxVitamin analysis, only three SNPs had q-values below 0.99, and none of the genes linked to these SNPs have previously been associated with orofacial clefts.

European sample
Among the SNPs with the lowest q-values in the PoOxAlcohol analysis, rs2294035 (q=0.32, p=2.9e-6) and rs4876274 (q=0.76, p=1.3e-5) are in 'Rho guanine nucleotide exchange factor 10' (ARHGEF10; GeneCards identifier [GCID]: GC08P001823) ( Table 4). The remaining SNPs had q-values above 0.76 and are not considered any further. ARHGEF10 has not previously been linked with orofacial clefts. In the PoOxSmoke analysis, three of the SNPs were in ANK3 (rs3793861: q=0.20, p=2.6e-6; rs7087489: q=0.20, p=3.1e-6; and rs4310561: q=0.67, p=4.0e-5). PoOxE effects in ANK3 were also detected in the analysis of the pooled sample above. To our knowledge, ANK3 has not previously been linked with orofacial clefts, and the same applies to SNP rs10763707 in 'Lysosome like 1' (LYZL1; GCID: GC10P029297), which had a q-value of 0.20. In the PoOxVitamin analysis, several of the SNPs shared the same q-value of 0.87 and are not considered any further. The top six SNPs had q-values of 0.44-0.65. Again, none of these genes appear to have any previous connections to clefting. For example, 'cAMP responsive element binding protein 5' (CREB5; GCID: GC07P028305) and its network of genes are involved in colorectal cancer 44 , while 'R-Spondin 3' (RSPO3; GCID: GC06P127118) is implicated in tumor development. That said, Park and co-workers reported that RSPO3 acts as an agonist in the canonical Wnt/β-catenin signaling 45 , a pathway known to be implicated in a wide range of developmental processes, including craniofacial development and homeostasis 46-49 . The 1000 Genomes browser was used to determine the chromosomal band location of a SNP. b If a SNP is located within a gene itself, the gene symbol is provided (the full names of the genes are provided in Table 6). SNPs located within 40 kb of a gene have the prefix '~', and those not located within a 40 kb-distance of a gene are denoted as NC (for 'not close'). Note that pseudogenes and non-coding RNAs are excluded. c Shared: Also featured in Table 4 or Table 5. SNP, single-nucleotide polymorphism; RRR, relative risk ratio; CI, confidence interval; NC, not close. a The 1000 Genomes browser was used to determine the chromosomal band location of a SNP. b If a SNP is located within a gene itself, the gene symbol is provided (the full names of the genes are provided in Table 6). SNPs located within 40 kb of a gene have the prefix '~', and those not located within a 40 kb-distance of a gene are denoted as NC (for 'not close'). Note that pseudogenes and non-coding RNAs are excluded. c Shared: Also featured in Table 3 or Table 5. SNP, single-nucleotide polymorphism; RRR, relative risk ratio; CI, confidence interval; NC, not close. a The 1000 Genomes browser was used to determine the chromosomal band location of a SNP. b If a SNP is located within a gene itself, the gene symbol is provided (the full names of the genes are provided in Table 6). SNPs located within 40 kb of a gene have the prefix '~', and those not located within a 40 kb-distance of a gene are denoted as NC (for 'not close'). Note that pseudogenes and non-coding RNAs (ncRNA) are excluded. There is no column for "shared" here, as none of these SNPs featured among those listed in Table 3 or Table 4. SNP, single-nucleotide polymorphism; RRR, relative risk ratio; CI, confidence interval; NC, not close.

Asian sample
In the only analysis possible for this ethnic group (PoOxVitamin), all the SNPs had the same q-value of 0.86 (Table 5). They are thus most likely to be false positives and will not be considered any further.

Haplotype analysis of SNPs in ANK3 and ARHGEF10
We chose to focus here on the PoOxE effects detected with SNPs in ANK3 and ARHGEF10. As mentioned above, ANK3 showed up several times among the top PoOxSmoke hits both in the pooled and European analyses, and strong signals for SNPs in ARHGEF10 were detected twice in the European analysis of PoOxAlcohol. We conducted stratified analyses of the effect of the child's allele, GxE effects, PoO effects, and PoOxE effects for each SNP and haplotype (with haplotypes analyzed both in two-SNP and three-SNP combinations) in these two genes (Table 7). Specifically, we analyzed rs3793861, rs7087489 and rs4310561 in ANK3 that showed PoOxSmoke effects in the European sample, and rs2294035 and rs4876274 in ARHGEF10 that showed PoOxAlcohol effects in the same sample (Table 4). The results did not show any child effects or GxSmoke effects for single SNPs in ANK3. By contrast, the p-values were low for all three SNPs in the PoO analyses or PoOxSmoke analyses. This was also the case with the 't-a' allele in the two-SNP combination rs3793861-rs7087489 and the 'c-t-a' allele in the three-SNP-combination rs3793861-rs7087489-rs4310561. The other alleles were only associated with PoOxSmoke effects.
For ARHGEF10, we analyzed the two SNPs that showed PoOxAlcohol effects in the European sample (rs2294035 and rs4876274) but did not discover any effects in either single-SNP or haplotype analyses.

Bioinformatics analysis
Because of their low q-values, the genes appearing in the PoOxSmoke and PoOxAlcohol analyses in Table 4 were selected for further analyses using the STRING database, Expres-sionAtlas and BGee. However, none of the searches for direct links among the genes yielded any evidence to explain why those genes appeared in our results together.   Indirect relationships between ANK3 and nicotine dependence, and between ANK3 and cleft lip. The brown nodes represent diseases, blue nodes show genes/proteins, and green nodes represent organs (anatomy). Each arrow represents a specific relationship between nodes: "LOCALIZES_DiA" = disease was found to be localized in an anatomy (organ); "EXPRESSES_AuG", "UPREGULATES_ AuG", "DOWNREGULATES_AuG" mean that the gene is expressed, upregulated, or downregulated in the anatomy (organ), respectively; "INTERACTS_GiG" means that the two genes were found to interact with each other (physically, as proteins); "ASSOCIATES_DaG" means that the gene was found to be associated with the disease; "RESEMBLES_DrD" means that the two diseases were found to occur significantly more often together in MEDLINE articles than would be expected by chance alone. Note that in this setting, the term "disease" includes any adverse medical condition, like syndromes, mental disorders, congenital anomalies, and so on.

Figure 2. Indirect relationships between ARHGEF10 and alcohol dependence, and between ARHGEF10 and cleft lip.
The brown nodes represent diseases, blue nodes show genes/proteins, and green nodes represent organs (anatomy). Each arrow represents a specific relationship between nodes: "LOCALIZES_DiA" = disease was found to be localized in an anatomy (organ); "EXPRESSES_AuG", "UPREGULATES_AuG", "DOWNREGULATES_AuG" mean that the gene is expressed, upregulated, or downregulated in the anatomy (organ), respectively; "INTERACTS_GiG" means that the two genes were found to interact with each other (physically, as proteins); "ASSOCIATES_ DaG" means that the gene was found to be associated with the disease; "RESEMBLES_DrD" means that the two diseases were found to occur significantly more often together in MEDLINE articles than would be expected by chance alone. Note that in this setting, the term "disease" includes any adverse medical condition, like syndromes, mental disorders, congenital anomalies, and so on.
Regarding the indirect relationships, these are visualized in Figure 1 for relationships between ANK3 and cleft lip (Disease Ontology ID [DOID]: 9296), and, simultaneously, between ANK3 and nicotine dependence (DOID: 0050742). As Hetionet does not include information about smoking, we chose "nicotine dependence" as a proxy. ANK3 is connected to nicotine dependence through several nodes, two of which are particularly noteworthy. First, ANK3 has been reported to be strongly associated with attention-deficit/hyperactivity disorder (ADHD) 50 , and a connection between ADHD and nicotine dependence has been reported. The connectiom was calculated based on articles listed in MEDLINE, where this pair of conditions co-occured significantly more frequently than would be expected by chance 51,52 . The second path goes through the gene 'CRK Like Proto-Oncogene, Adaptor Protein' (CRKL). It interacts with ANK3 and is downregulated in nicotine dependence. Furthermore, ANK3 is expressed in the telencephalon (the most highly developed part of the forebrain), the embryo, and the head -all of which are all relevant to CL/P. Figure 2 shows the indirect relationships between cleft lip, ARH-GEF10 and alcohol dependence. Like nicotine dependence in the above analyses, alcohol dependence (DOID: 0050741) was used here as a proxy for maternal alcohol consumption. The only relationships found between ARHGEF10 and cleft lip are the expression of ARHGEF10 in the head and telencephalon. By contrast, there were twelve different relationships between ARHGEF10 and alcohol dependence. However, there were no shared paths connecting cleft lip and alcohol dependence via any of the 12 organs.

Regional plots
The regional plot for rs3793861 (Figure 3) shows that several SNPs in ANK3 that were not in linkage disequilibrium with   Setting the minor allele frequency to 0.2 while varying the number of unexposed and exposed triads (unexposed-exposed). Right panel: Setting the number of unexposed and exposed triads to 1100 and 500, respectively, while varying the minor allele frequency. In all analyses, the significance level was 0.05. We varied the maternal RR in exposed triads, so that RRR=RR mat (Exposed). The black curve is the same in both panels because of shared parameters. RRR, relative risk ratio; RR mat , relative risk for an allele inherited from the mother; MAF, minor allele frequency. rs3793861 had p-values in the range 10 -4 to 10 -3 , which lends support to either ANK3 itself or genes in its vicinity influencing the risk of clefting. However, we did not observe a similar pattern in the regional plot for rs2294035 (Figure 4). Figure 5 shows that the power does not increase appreciably when the minor allele frequency increases beyond 0.2. However, there is a lot to be gained by increasing the sample size from 500 unexposed and 300 exposed (European, smoke/alcohol) to 1400-600 (pooled, vitamin). Further, the RRRs in the plots are based on changing only the effect of the maternal allele in the exposed triads. The same RRRs could have been achieved in a number of ways, which complicates the interpretation of the RRRs in Table 3-Table 5. Still, if a strong effect is detected with a SNP in a gene, this strengthens the case for its contribution to clefting.

Discussion
The main aim of this paper was to identify genome-wide PoOxE effects in the larger sample of isolated CL/P, based on the same methodology and GWAS dataset we had previously used in a similar analysis of the smaller sample of isolated CPO 27 . As with the CPO study, the current analyses benefitted from being  based on the largest available GWAS dataset of case-parent triads of orofacial clefts to date. Moreover, data were available for two major ethnicities, European and Asian, which is useful in assessing the generalizability of the findings across different ethnic groups. In the current dataset, however, very few of the Asian mothers reported smoking cigarettes or consuming alcohol during the periconceptional period, thus preventing a comparison of PoOxE effects for these exposures across these two ethnic groups. This is a common impediment to GxE studies, where the number of exposed individuals needs to be large enough for a meaningful analysis 53 .
A possible mechanism for a PoO effect is genomic imprinting 21 . This occurs when DNA methylation in the germline causes the expression of alleles to be silenced depending on their parental origin. Maternal environmental exposures that affect methylation patterns may also affect paternally and maternally inherited alleles differently. Furthermore, a PoO effect that is not affected by an environmental exposure will not be detected in our PoOxE analysis. Hence, a detected PoOxE effect may have a better chance than a PoO or GxE effect of uncovering a true causal relationship involving genomic imprinting.
Relying on the q-values for assessing the false positive rate, our analyses detected possible PoOxSmoke effects with SNPs in LYZL1, ANK3, PDGFD, FOCAD and FRAS1, and possible PoOxAlcohol effects with two SNPs in ARHGEF10. Without formal validation in a comparable and independent replication cohort, it would be premature to accept these associations as true PoOxE effects. Not having previously been linked with orofacial clefts does not necessarily imply that the identified gene is not relevant for clefting. This applies to several genes in our analyses; for example, SHKBP1 and MAPK10 in the PoOxSmoke analysis of the European sample and TJP3 in the PoOxVitamin analysis of the pooled sample. The current study was primed to explore new hypotheses for disease mechanisms and to provide as many of the results as possible so that other researchers with access to similar GWAS datasets would be able to validate the findings presented here. To avoid being overly stringent, we thus presented all the results for the top 20 SNPs in Table 3-Table 5.
Despite an exhaustive literature search, we were unable to find any obvious evidence linking ANK3 and orofacial clefts. The Hetionet results confirmed this lack of a direct connection (Figure 1). ANK3 encodes a member of the Ankyrin family of proteins, whose function is to bind the integral membrane proteins to the spectrin-actin cytoskeleton. This is important for cell motility, activation, proliferation, contact and the maintenance of specialized membrane domains; cellular activities that are also relevant for the proper development of craniofacial structures. For example, Stankewich and colleagues 54 showed that the spectrin-ankyrin scaffold is important for cell migration, tissue patterning and organogenesis. Homozygous deletion of the gene encoding αII-spectrin in mice (Spna2) resulted in craniofacial, neural tube and cardiac anomalies, in addition to retarded intrauterine growth. Figure 3 indicates that several SNPs in ANK3 are potentially associated with clefting.
Like ANK3, ARHGEF10 has not previously been associated with orofacial clefts, and the Hetionet results are consistent with this observation (Figure 2). ARHGEF10 encodes a Rho guanine nucleotide exchange factor that may be involved in neural morphogenesis 55 . LYZL1 belongs to the family of lysozyme-like proteins that are implicated in sperm function and innate immunity 56 . According to GeneCards (GCID: GC09P020659), FOCAD encodes a tumor suppressor gene that is highly expressed in the brain. It has also been linked to Alzheimer's disease 57 . Furthermore, germline deletions in FOCAD are associated with polyposis and colorectal cancer 58 . Again, as with ANK3 and ARHGEF10 above, there do not seem to be any obvious connections between LYZL1 or FOCAD with clefting.
In contrast to the above genes, FRAS1 and several members of the platelet-derived growth factor (PDGF) gene family are known to be implicated in orofacial clefts. PDGFD is a member of the PDGF gene family and plays a central role in the PDFG receptor-alpha (PDGFR-α) signaling pathway. More specifically, disruption of Pdgf signaling results in clefting of the palate 59 . FRAS1 (GCID: GC04P078056) encodes an extracellular matrix protein that plays a critical role in epithelial-mesenchymal interactions during embryonic development 60 . Loss-of-function mutations in FRAS1 underlie Fraser syndrome, which is characterized by craniofacial, urogenital and respiratory system abnormalities 61 . Both of these genes are therefore worthy of further investigations in other isolated orofacial cleft cohorts.
A limitation of this study is that genotypes were not imputed.
To avoid Mendelian inconsistencies, the imputation procedure would have had to account for the full triads, as opposed to imputing each sample independently, which has not been done for our data. Instead we conducted post hoc haplotype analyses for combinations of top SNPs located close to each other. This is akin to imputation, in that such an analysis takes into account information about a whole area of DNA, instead of just one SNP. Table 7 shows that in the PoOxSmoke analysis of the two-SNP combination rs3793861-rs7087489 in ANK3, the p-value was slightly lower and the RRR slightly higher than in the corresponding analyses of each individual SNP. A similar pattern was observed in the PoOxAlcohol analyses of the rs2294035-rs4876274 haplotype in ARHGEF10. This indicates that the two-SNP combinations may be driving the effects observed with the individual SNPs.
The genes PDGFD, CSMD1 and RSUI detected here had previously showed up in a study focusing on identifing GxE effects in the same CL/P triads 28 . In that study, a possible GxVitamin effect was detected with PDGFD and RSUI, and a possible GxAlcohol effect was detected with CSMD1. In the current study, a PoOxSmoke effect was detected with PDGFD and PoOxVitamin effects were detected with RSUI and CSMD1. In other words, only RSUI had the same exposure (vitamin) across the studies. RSUI stands for 'Ras suppressor protein 1' and is localized to chromosome 10p13. Its protein product is found at cellextracellular matrix adhesion sites and has been reported to be involved in supressing v-Ras transformation in the Ras signal transduction pathway 62 . CSMD1 stands for 'CUB and Sushi multiple domains 1' and is localized to chromosome 8p23.2 (GCID: GC08M002953). It is involved in tumor suppression, as it has frequently been found to be deleted in many types of cancers 63,64 . Again, there does not seem to be any obvious connections to clefting.
None of the top SNPs identified in our previous study focusing on PoOxE effects in CPO triads overlapped with SNPs identified in this study of CL/P 27 . This is consistent with the observation that CPO and CL/P are etiologically distinct, so that the lead SNPs may be subtype-specific and differ between the two conditions. However, we detected associations with the 'cytochrome P450 family 4 subfamily F member 3' gene (CYP4F3 on chr 19p13.12) in the previous CPO analyses, and with the 'cytochrome P450 family 46 subfamily A member 1' gene (CYP46A1 on chr 14q32.2) in the present study. These two genes are members of the cytochrome P450 superfamily of enzymes that are primarily found in liver cells and whose function is to catalyze many reactions involved in the biotransformation of xeno-and endobiotics, and the biosynthesis of cholesterol and lipids, among others 65 . It is therefore not surprising that these genes would appear in an analysis focusing on smoking, alcohol and vitamin intake.
We searched Hetionet for indirect links between ANK3 or ARH-GEF10 and cleft lip, as well as between ANK3 or ARHGEF10 and nicotine or alcohol dependence, respectively. This approach has several limitations. First, using "nicotine dependence" and "alcohol dependence" in lieu of the actual smoking and alcohol consumption status may introduce some bias. Second, Hetionet is built from a curated set of database information, which means that not all the information, especially the newest, would be available. However, when interpreting our results, we used the source databases to make sure that the connections between the nodes are reliable.
To conclude, our search for an interaction between a PoO-effect and an environmental exposure for CL/P identified possible relationships between SNPs in ANK3 and maternal smoking, and SNPs in ARHGEF10 and maternal intake of alcohol. There is a possibility that these interactions have a biological basis, although without replication they remain speculative. Our demonstration of the feasibility of identifying complex interactions between relevant environmental exposures and PoO-effects opens new possibilities in the search for the genetic etiology of CL/P.

Underlying data
The GWAS data are available in the dbGaP database. Additional information regarding the inclusion/exclusion criteria of the study, the ethics statements, data variables, study history, publications, and other documentation related to the study is provided on the dbGaP website.
The dbGaP database at the National Center for Biotechnology Information, U.S. National Library of Science (NCBI/NLM) provides an extensive overview of the cleft dataset used in this study. Entering the dbGaP accession number phs000094.v1.p1 provides access to information regarding the variables, study documents, and datasets. For example, detailed information about the mother's exposure to alcohol, vitamins, and smoke is provided under the header "Variable Selection". Information on study questionnaires, institutional review boards and consent forms from each participating cohort can be found under the header "Documents".
Controlled-access data from dbGaP is available only through the dbGaP authorized access portal. There are separate procedures for accessing individual data, depending on whether the researcher is NIH-affiliated (intramural) or not (extramural). In addition, other restrictions apply; e.g., the principal investigator from the applying institution needs to be a permanently employed professor, senior scientist, or equivalent, to submit a data access request. A valid eRA Commons account for logging in to the dbGaP system is also mandatory.
identifying interactions between relevant environmental exposures and PoO effects. Totally, this paper was very well written; however, there are some concerns that should be addressed to improve the paper as detailed below: The GWAS data are available in the dbGaP database (the dbGaP accession number: phs000094.v1.p1). Since the imputed data were provided from dbGaP, it is more appropriate to use that data for interaction analysis to find more meaningful results in this study.
It seems the author only presented the process on quality control for SNPs, and the process on quality control for individuals should be added in the part of Methods.
In the present study, the author focused the q-value rather than p-value, which is more widely accepted as known. It is confused that the cut off of q-value chosen in each interaction analysis on different population is unclear.
It is unclear that the cutoff of q-value in each interaction analysis on different population. The author needs to determine a clear threshold for screening, which should be more strict.
There are a lot of tables in the article. It is recommended to add the forest plots of the interaction analysis in order to make the results more intuitive.
I am not sure if it is appropriate to state 'After quality control, 341,191 SNPs remained from the original 569,244.' in Methods section from abstract, because that seems to be part of the result.
All the population and analysis were based on nuclear families with cleft lip with/without cleft palate, it is proper to replace 'cleft lip' with 'cleft lip with/without cleft palate' in the title of this paper.

If applicable, is the statistical analysis and its interpretation appropriate? Partly
Are all the source data underlying the results available to ensure full reproducibility? Yes

Evie Stergiakouli
School of Oral and Dental Sciences, University of Bristol, Bristol, UK This is a very interesting paper focusing on identifying parent of origin interaction effects in CL/P using a genome-wide approach. The environmental factors selected are biologically plausible to be involved in cleft and the sample, although modest in size, involved full trios and individuals of non-European ancestry. The authors identify suggestive evidence of interaction for 3 SNPs in ANK3 with smoking and 10 SNPs for ARHGEF10 with alcohol consumption. This study provides proof of principle for their approach and the authors highlight the need for replication in independent samples. The authors follow the results with bioinformatics analysis and the paper is well written and detailed.
My main concern with the statistical analysis is with the use of directly genotyped SNPs in a genome-wide scan. Can the authors explain why there is no imputation of the directly genotyped SNPs? Imputation would have provided a much larger number of SNPs to work with. Also, what is the rational for not using the established genome-wide significance level (5*10-8)? If anything, the authors should have been more stringent than that since they effectively performed 3 genome-wide analyses (one for each environmental factor).
The abstract should be made for concise: The 59 individuals of other ancestries were not used for any analyses, so they should not be mentioned in the abstract. The methods section of the abstract does not contain any description of how the performed the analysis and relies heavily on their previous study.
I am not sure if the following statement should be called intriguing: 'Despite this success, the genetic variants identified so far collectively explain only a minor fraction of the total vari-ance attributable to additive genetic effects, which is intriguing considering the more than 70% heritability of CL/P among Europeans17-20'. This is not unusual, and it is the case for many other complex disorder and traits, so I would suggest rephrasing.
The table titles should include more information to make the paper easier to follow. For example, Table 3 should include that it is a PoOxE analysis.
Discussion: The authors need to highlight that none of their findings reached genome-wide significance level and given that they have performed more tests that in a standard genome-wide study, the evidence for association is weak. So, their findings require replication and/or meta-analysis in a larger sample. I am not sure I agree with their statement: 'Without formal validation in a comparable and independent replication cohort, it would be premature to dismiss the veracity of the remaining associations merely on replication cohort, it would be premature to dismiss the veracity of the remaining associations merely on the basis of their having too high q-values.' All of their findings need to be replicated and should be treated with caution until then.
Another point that I need to make is that I would have expected the discussion to be more about result interpretation and even trying to suggest possible mechanisms. At the moment the discussion is focused on detailed bioinformatics analysis which although interesting, would be more suited to the results section.

If applicable, is the statistical analysis and its interpretation appropriate? Partly
Are all the source data underlying the results available to ensure full reproducibility? Yes

Are the conclusions drawn adequately supported by the results? Partly
No competing interests were disclosed.

Competing Interests:
Reviewer Expertise: Genetic epidemiology of childhood disorders I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard, however I have significant reservations, as outlined above.
Author Response 09 Jul 2019 , University of Bergen, Bergen, Norway

Øystein Ariansen Haaland
We would like to thank Dr. Stergiakouli for her insightful comments, and have addressed her concerns below. We have also made minor changes throughout the manuscript to correct grammatical errors or improve clarity. These are not addressed below, but are highlighted in the new version of the manuscript.

Comment:
My main concern with the statistical analysis is with the use of directly genotyped SNPs in a genome-wide scan. Can the authors explain why there is no imputation of the directly genotyped SNPs? Imputation would have provided a much larger number of SNPs to work with.

Reply:
Reply: Thank you for pointing this out.
Text changed in Abstract: "We recently developed new methods to identify parent-of-origin (PoO) interactions with environmental exposures (PoOxE) and now apply them to data from a genome-wide association study (GWAS) of families with children born with isolated CL/P." "The methodology applied in the analyses detects if PoO effects are different across environmental strata, and is implemented in the -package Haplin." R Comment: I am not sure if the following statement should be called intriguing: 'Despite this success, the genetic variants identified so far collectively explain only a minor fraction of the total variance attributable to additive genetic effects, which is intriguing considering the more than 70% heritability of CL/P among Europeans17-20'. This is not unusual, and it is the case for many other complex disorder and traits, so I would suggest rephrasing.
Reply: Thank you for pointing this out.
Text changed in Introduction: "Despite this success, the genetic variants identified so far collectively explain only a minor fraction of the total variance attributable to additive genetic effects, even though the heritability of CL/P among Europeans is more than 70% ." Comment: The table titles should include more information to make the paper easier to follow. For example, Table 3 should include that it is a PoOxE analysis.
Reply: Thank you for this comment.
Title changed in Table 3: "The top 20 single-nucleotide polymorphisms (SNPs) sorted by p-value in the pooled PoOxE analysis." Title changed in Table 4: "The top 20 single-nucleotide polymorphisms (SNPs) sorted by p-value in the European PoOxE analysis." Title changed in Table 7: "Stratified analyses of the top single-nucleotide polymorphisms (SNPs) and haplotypes in ANK3 (PoOxSmoke) and (PoOxAlcohol) " ARHGEF10 .
Comment: Discussion: The authors need to highlight that none of their findings reached genome-wide significance level and given that they have performed more tests that in a standard genome-wide [17][18][19][20] significance level and given that they have performed more tests that in a standard genome-wide study, the evidence for association is weak. So, their findings require replication and/or meta-analysis in a larger sample. I am not sure I agree with their statement: 'Without formal validation in a comparable and independent replication cohort, it would be premature to dismiss the veracity of the remaining associations merely on the basis of their having too high q-values.' All of their findings need to be replicated and should be treated with caution until then.
Reply: Thank you for pointing this out.
Text changed in Discussion: "Relying on the q-values for assessing the false positive rate, our analyses detected possible PoOxSmoke effects with SNPs in , and , and possible LYZL1 ANK3, PDGFD, FOCAD FRAS1 PoOxAlcohol effects with two SNPs in . Without formal validation in a comparable and ARHGEF10 independent replication cohort, it would be premature to accept these associations as true PoOxE effects." Comment: Another point that I need to make is that I would have expected the discussion to be more about result interpretation and even trying to suggest possible mechanisms. At the moment the discussion is focused on detailed bioinformatics analysis which although interesting, would be more suited to the results section.
Reply: Thank you for this suggestion.
Added to Discussion: "A possible mechanism for a PoO effect is genomic imprinting . This occurs when DNA methylation in the germline causes the expression of alleles to be silenced depending on their parental origin. Maternal environmental exposures that affect methylation patterns may also affect paternally and maternally inherited alleles differently. Furthermore, a PoO effect that is not affected by an environmental exposure will not be detected in our PoOxE analysis. Hence, a detected PoOxE effect may have a better chance than a PoO or GxE effect of uncovering a true causal relationship involving genomic imprinting." No competing interests were disclosed.

Competing Interests:
The benefits of publishing with F1000Research: Your article is published within days, with no editorial bias You can publish traditional articles, null/negative results, case reports, data notes and more 21