Fetal alcohol syndrome (FAS) is the most common preventable cause of mental retardation globally, and is a serious public health problem in South Africa 1. The range of prevalence rates reported in two different primary school cohorts from this community were 65.2–74.2 per 1 000 2 and 68.0–89.2 per 1000 1 respectively. This rate is alarmingly higher than the average observed for the developed world of 0.97 per 1000 live births 3.

The teratogenic effect of alcohol is well established and exposure to alcohol in utero is known to result in a widely variable phenotype. Fetal alcohol spectrum disorder (FASD) is an umbrella term used to describe the irreversible array of anomalies associated with in utero alcohol exposure 4. These anomalies include prenatal and postnatal growth retardation, central nervous system (CNS) dysfunction, characteristic craniofacial malformation and other organ abnormalities 567. The term FAS is a clinical description for children at the most severe end of the FASD spectrum, who display the full phenotype associated with in utero alcohol exposure.

Although alcohol consumption during pregnancy is the primary trigger for the presentation of FAS, the exact mechanisms for alcohol-induced teratogenic effects have not been elucidated. Research has shown that secondary factors, like genetic, epigenetic and environmental factors influence the outcome and severity of the disorder. Furthermore, a dose- and time-dependant relationship has been observed, where exposure to higher concentrations of alcohol at critical developmental stages resulted in more severe anomalies 8. An association between a variable genetic background and FAS development is primarily supported by the observation that FAS does not occur in all children exposed to alcohol during the prenatal period 9. This observation suggests that certain individuals may have a genetic predisposition to infliction of more severe damage by gestational alcohol consumption; and the varied phenotype observed in FASD may be a reflection of the varied susceptibility quotients in the genetic background of the individual. Streissguth and Dehaene 10 studied twin pairs with alcoholic mothers, and found the rate of concordance for FASD to be 100% for monozygotic twins, whereas digygotic twins showed only 64% concordance. Further support for the role of genetics in FAS development is obtained from animal model studies 11. Several studies in different mouse strains have shown variation in the extent and pattern of alcohol-induced malformation, as well as behavioural outcome 12131415. FAS can therefore be considered to be a multi-factorial or complex disease, suggesting that there are multiple genetic factors underlying susceptibility to FAS and the interactions between these factors as well as other factors are likely to be intricate.

To date, no FAS family linkage studies or genome wide association studies have been performed. Linkage studies require large family samples and this poses a significant challenge. Countries with the highest FAS rates are mostly resource-poor, possibly contributing to the reason why such studies have not yet been performed. Furthermore, linkage studies have not proven to be particularly successful in discovering the genetic causes of complex diseases, the critical factor being the generally weak genotype-phenotype association in multi-factorial disorders 16.

Few candidate gene association studies investigating the effect of specific genetic polymorphisms on the risk of FAS development have been published. These studies have generally focused on the alcohol dehydrogenase enzyme family members and conflicting results have been obtained. Stoler et al. 17 observed that the absence of the ADH1B*3 allele was protective for fetal outcome, in conflict with two other studies showing the presence of this allele to be protective 1819. The ADH1B*2 allele has been proposed to play a possible protective role, or to be a marker for protection in the South African mixed-ancestry population 20. However, the sample size for this association study was small, and results have not yet been replicated in other populations. Many other genes are likely to contribute towards the development of FAS and further investigation is required.

Candidate gene association studies remain the most practical and frequently employed approach in disease gene investigation for complex disorders. However, the main challenge when using this approach is to select suitable genes to test, especially for diseases with poorly understood aetiology. Recently, many computational candidate gene selection and -prioritization methods have been developed 2122232425262728293031. These tools aim to identify and prioritize putative disease genes by modelling specific characteristics of known disease genes, or by focusing on known disease features (such as gene expression profiles or phenotype). However, there is a vast quantity of information and data sources available currently, and it is expected that a tool with the flexibility to include a large array of data sources would positively aid disease gene discovery. The freely accessible tool Endeavour offers such an application 22. This tool is based on the premise of ranking unknown candidate genes according to their similarity with a known set of training genes. In the absence of a linked genetic region (which is the case with FAS), all genes in the genome must be included as a starting point for candidate gene selection, which is not feasible when using this approach.

Convergent Functional Genomics (CFG) is an approach used to identify and prioritize candidate genes, which relies on the cross-matching of animal model gene expression data with human genetic linkage data, as well as human tissue data and biological roles data 3233 This approach has many parallels to the approach described in this paper, as it prescribes a Bayesian-like methodology of reducing uncertainty through the combination of multiple independent lines of evidence, each by itself lacking sufficient power to confirm that a gene is a putative candidate gene, to produce a short list of high probability candidate genes 32. The approach of CFG relies principally on two lines of evidence – animal model data and human genetic linkage data. The approach we describe in this paper has the added advantage of allowing the inclusion of additional lines of evidence in the presence of limited expression studies in an animal model and the absence of FAS linkage studies.

Tiffin et al. 34 recently surveyed some of the methods for computational disease gene identification and concluded that using the methods in concert was more successful in prioritizing candidate genes for disease, than when each was used alone. This review additionally showed that using existing computational methods in concert highlighted potential candidates that are selected by a subset of methods and are missed by the other methods, depending on the type of data examined. This observation gives further evidence that the inclusion of more data sources may positively aid disease gene discovery.

In light of the current burden of FAS in many resource-poor communities and the inconclusive search for susceptibility genes, computational identification offers a novel and efficient approach to the identification of disease-causing genes. Initially we used the candidate gene selection method described by Tiffin et al. 31, but in the absence of a candidate genetic region, this method resulted in a large candidate gene list, as it relies on the selection of candidate disease genes only according to their expression profiles. This prompted us to devise a new prioritization method to rank genes from the candidate gene list for empirical investigation. The prioritization method described here is based on a simulation of a researcher's approach to selecting candidate disease genes. In this process, a variety of relevant database sources are mined for candidate genes that exhibit characteristics relevant to disease phenotype. Genes were prioritized based on binary evaluation, where genes were assessed using criteria pertinent to FAS to mine various database sources and to create criteria-specific gene lists. The validity of the binary prioritization method was assessed by investigating the protein-protein interactions, functional enrichment and common promoter element binding sites of the top-ranked genes.

According to the method described by Tiffin et al. 31, Dragon Disease Explorer (DDE) was used to extract eVOC anatomical terms from the body of literature, where after they were used to extract candidate genes from the Ensembl database. This method extracted a list of 10174 genes, a reduction of 70.3% from the original 34294 genes in the Ensembl database.

In order to select the most likely candidates from the initial candidate gene list, these genes were ranked according to the number of additional criteria (Table 1) they matched. The top-ranked genes (in ranked order) are shown in Table 2. FGFR1 was the top-ranked gene, present in 17 of the 29 criteria gene lists, followed by MSX1, present in 16 of the 29 criteria lists. FGFR2, FOXG1B and HOXA1 were present in 15 of the 29 criteria lists, followed by a group of 4, 17, 14 and 47 genes present in 14, 13, 12 and 11 criteria lists, respectively. This group of 87 genes was used as the prioritized candidate gene list for further analyses (see Additional file 1). This cut off (present in 11 of the 29 criteria lists) was used to select an appropriately-sized group of top-ranked genes.

Genes from the candidate gene lists that matched one or none of the criteria were considered to be unlikely candidates. Based on this premise, these 5055 genes (50%) from the candidate gene list were ranked as weak candidates. 87 Genes of the subset matching to no criteria were randomly selected for further analysis as a negative control set to assess the validity of the ranking method.

A method that employs an integrative literature- and data mining approach to select candidate genes was used to select candidate genes for FAS 31. This method extracted a gene list of 10174 genes. This list is relatively unspecific, and is likely to have a high false-positive rate. The most plausible explanation for the selection of such a large, ambiguous list is a lack of detailed information about the source of cDNA libraries, with the result that more general terms from higher up the ontology hierarchy are often used for annotation of the gene. This prompted us to devise a prioritization method to rank genes from the candidate gene list using many different data sources for laboratory investigation of individual candidate genes. A binary evaluation method was used to rank the candidate genes in the list, facilitating the selection of 87 top-ranked genes as the most likely candidate genes for further investigation.

Further analysis with available online tools such as DAVID and STRING highlighted protein-protein interaction, functional enrichment and probable biological significance among the top-ranked genes. STRING was used to investigate protein-protein interaction among the prioritized candidate genes, and highlighted a group of genes that interact (Figure 1 and Table 3). The candidate gene selection method described here focuses on gene annotation, and it is therefore possible that the top ranking genes are better annotated than low-ranked genes. Therefore the absence of protein-protein interaction among the low-ranked genes is not necessarily a reflection on level of interaction but may be related to the level of understanding of the gene and its function. It is accepted that the genes underlying complex disease (such as FAS) will be plentiful and the interactions between these factors are likely to be intricate. For this reason, STRING is a useful tool to highlight genes within the top-ranked gene list that interact and that may have a cooperative effect on disease outcome.

DAVID elucidates functional enrichment and biological significance within the top-ranked gene list, and highlighted the TGF-β and Mitogen-Activated Protein Kinase (MAPK) signalling pathways as primary candidate pathways for FAS development.

As a way of further assessing the list of 87 prioritized FAS candidate genes they were cross-matched against candidate genes for alcoholism, obtained using Convergent Functional Genomics 32. Although the two phenotypes are very different, one would expect some overlap in prioritized candidate genes since many of the mothers of FAS children suffer from alcoholism. The two prioritized candidate gene lists (87 genes for FAS and 65 for alcoholism) had only two high priority candidate genes in common – GNAS complex locus (GNAS) and high mobility group protein B1 (HMGB1). The remaining 63 candidate genes for alcoholism were also present in the initially selected list of 10174 genes, but were ranked below the arbitrary cut-off of 11/29 criteria used to select the highly prioritized candidate list for FAS.

Incorporating the set of alcoholism genes as a selection criterion into the binary evaluation method only added two more genes to the prioritized list. These were G1/S-specific cyclin-D1 (CCND1) and insulin-like growth factor I receptor (IGF1R). Both gene products contribute to cell proliferation and differentiation, and exhibit characteristics that also make them likely candidate genes for FAS. However, neither directly interacts in the two main prioritized pathways (TGF-β or MAPK signalling pathway). This comparison shows that these two related diseases (due to the involvement of alcohol in both) have potentially common genetic factors, but that they also exhibit diversity in terms of genetic susceptibility. This gene list was therefore not included in the final binary filtering analysis.

All TFBS that are found to be statistically significant for FAS are known to be involved in gene expression and regulation in the CNS, endocrine system or development. The AP-2 family of TF is crucial for neural gene expression and neuronal development 69; C/EBP is involved in neuronal signalling 70; the E2F family of TF is one of the key controllers of cell-cycle and has a known role in pathways controlling neuron death 71; ETF, the epidermal growth factor receptor-specific TF, is implicated in neuroblastoma 72; LEF1 is expressed in the nerve system of mammals 73; MAZ is involved in Hodgkin's disease and paraneoplastic cerebellar dysfunction 74 and during neuronal differentiation 75; MAZR is implicated in the development of mouse limb buds 76; MZF1 is involved in development 77 and implicated in the control of the BACE1 gene related to Alzheimer's disease 78; Pax-4 is involved in the endocrine system and development 79; Sp1 has multiple roles, but, for example, controls expression of Na+,K+-ATPase in neuronal cells 80; Spz1 is involved in cell-proliferation 81; TATA binding proteins are implicated in various processes involved in brain 82; the TFII-I transcription factor family is implicated in craniofacial development of humans and mice 83; VDR is associated with increased risk of schizophrenia 84; and ZF5 is implicated in neuroblastoma differentiation 85. These results support the prioritization of biologically relevant candidate disease genes.

Abstracts related to FAS were obtained from the PubMed scientific literature database. In order to obtain all relevant literature, PubMed's automatic term mapping search of the literature might not be sufficient and a more robust search option of using Medical Subject Heading (MeSH) terms was implemented. Using this option also implies that all equivalent synonyms or lexical variants in English will be included in the search 86. Literature related to FAS was obtained using the following query: "(fetal alcohol syndrome [MH]) OR (fetal alcohol spectrum disorder* [tw])" Limits: only items with abstracts, English.

The online literature mining tool DDE 87 was used to extract eVOC ontology terms from the body of literature. The eVOC ontology is a controlled vocabulary used to describe the sample source of cDNA and SAGE libraries and labelled target cDNAs for microarray experiments. eVOC contains four major orthogonal ontologies – anatomical system, cell type, pathology and developmental stage 88. DDE provides summarized information from a body of submitted PubMed abstracts about frequency of occurrence of ontology terms within the text. This assists biologists in uncovering possible functional associations between disease and gene expression site. Following the method of Tiffin et al. 31, only eVOC anatomy terms were used to extract the initial candidate gene list. Cell type terms were used to populate criteria lists for the binary filtering approach (Table 1). Terms extracted matching to the developmental- and pathology ontologies were uninformative in this case (terms such as pathology or adult were extracted) and it was deemed that populating criteria lists using these terms would not contribute positively to the selectivity of the binary evaluation system. Therefore these terms were not further included.

The method previously described by Tiffin et al. 31 was used to extract candidate genes based on the information obtained from the literature mining. Figure 2 illustrates the process of literature- and data-mining used to select candidate genes. Briefly, this method ranks the extracted eVOC terms by calculating a ranking score for each associated eVOC term, according to the frequency of association and the frequency of annotation of the eVOC term. The four top-scoring eVOC terms were selected from the ranked list, and compared with eVOC terms annotated to genes within the Ensembl database (Ensembl v33, September 2005) to select candidates. The system allows for one mismatch, such that candidates selected are those annotated with at least three of the four top-scoring eVOC terms. This approach was tested by the authors on a subset of genes representative of those that might be selected by a linkage analysis study, and not the full complement of genes in the Ensembl database, as in the current study.

The integrated literature- and data-mining approach to identify candidate genes focuses exclusively on anatomical sites related to the disease of interest, and results in a large list of genes. In order to obtain a more focused assessment of the most likely candidates from this gene list, other criteria pertinent to FAS were investigated. Five main categories of criteria were used – cell type, biological process, homology, imprinted genes and phenotype simile. For each category there are multiple gene lists, each specified by one criterion (Table 1). The criteria-specific gene lists generated were compared to the candidate gene list (obtained from the integrated data- and literature-mining approach described above) to create a binary matrix. The binary evaluation was performed as follows: A gene in the candidate gene list was assigned a 1 when that gene was also present in the gene list obtained by a specific criterion. If the gene was absent from that list it was assigned a 0.

For each of the genes we calculated the final binary score, simply by summing all binary scores for each of the criteria used. Then we ranked all genes based on this score, with those having higher scores being higher in the rank list. Genes in the candidate list that were present in most criteria lists (i.e. those genes that obtaining the most 1-scores in the binary matrix) received the highest rank as candidates. This follows the premise that genes most commonly selected from additional independent sources possess characteristics that make them more promising candidates. Similarly, genes that were selected by only one or none of the additional criteria have a lower rank and are considered to be weak candidates.

A description of each category of criterion and the information used to assess the criteria are given below:

Protein-protein interactions, functional enrichment and common promoter element binding sites were investigated for the top-ranked genes (i.e. those with the highest binary score) to assess their biological significance as candidates for FAS. In comparison, the lowest-ranked genes were similarly evaluated to assess the validity of the ranking system in selecting biologically relevant genes from the original candidate gene list.